From 36d22d82aa202bb199967e9512281e9a53db42c9 Mon Sep 17 00:00:00 2001 From: Daniel Baumann Date: Sun, 7 Apr 2024 21:33:14 +0200 Subject: Adding upstream version 115.7.0esr. Signed-off-by: Daniel Baumann --- third_party/aom/aom_dsp/add_noise.c | 73 + third_party/aom/aom_dsp/aom_convolve.c | 238 ++ third_party/aom/aom_dsp/aom_dsp.cmake | 356 +++ third_party/aom/aom_dsp/aom_dsp_common.h | 98 + third_party/aom/aom_dsp/aom_dsp_rtcd.c | 18 + third_party/aom/aom_dsp/aom_dsp_rtcd_defs.pl | 1575 +++++++++++++ third_party/aom/aom_dsp/aom_filter.h | 56 + third_party/aom/aom_dsp/aom_simd.h | 38 + third_party/aom/aom_dsp/aom_simd_inline.h | 21 + third_party/aom/aom_dsp/arm/blend_a64_mask_neon.c | 451 ++++ third_party/aom/aom_dsp/arm/fwd_txfm_neon.c | 222 ++ third_party/aom/aom_dsp/arm/intrapred_neon.c | 590 +++++ third_party/aom/aom_dsp/arm/loopfilter_neon.c | 928 ++++++++ third_party/aom/aom_dsp/arm/sad4d_neon.c | 226 ++ third_party/aom/aom_dsp/arm/sad_neon.c | 224 ++ third_party/aom/aom_dsp/arm/subpel_variance_neon.c | 131 ++ third_party/aom/aom_dsp/arm/subtract_neon.c | 81 + third_party/aom/aom_dsp/arm/variance_neon.c | 400 ++++ third_party/aom/aom_dsp/binary_codes_reader.c | 123 + third_party/aom/aom_dsp/binary_codes_reader.h | 47 + third_party/aom/aom_dsp/binary_codes_writer.c | 210 ++ third_party/aom/aom_dsp/binary_codes_writer.h | 68 + third_party/aom/aom_dsp/bitreader.h | 160 ++ third_party/aom/aom_dsp/bitreader_buffer.c | 67 + third_party/aom/aom_dsp/bitreader_buffer.h | 50 + third_party/aom/aom_dsp/bitwriter.h | 89 + third_party/aom/aom_dsp/bitwriter_buffer.c | 87 + third_party/aom/aom_dsp/bitwriter_buffer.h | 51 + third_party/aom/aom_dsp/blend.h | 45 + third_party/aom/aom_dsp/blend_a64_hmask.c | 69 + third_party/aom/aom_dsp/blend_a64_mask.c | 345 +++ third_party/aom/aom_dsp/blend_a64_vmask.c | 71 + third_party/aom/aom_dsp/buf_ans.c | 70 + third_party/aom/aom_dsp/buf_ans.h | 136 ++ third_party/aom/aom_dsp/daalaboolreader.c | 47 + third_party/aom/aom_dsp/daalaboolreader.h | 160 ++ third_party/aom/aom_dsp/daalaboolwriter.c | 31 + third_party/aom/aom_dsp/daalaboolwriter.h | 78 + third_party/aom/aom_dsp/entcode.c | 49 + third_party/aom/aom_dsp/entcode.h | 40 + third_party/aom/aom_dsp/entdec.c | 229 ++ third_party/aom/aom_dsp/entdec.h | 83 + third_party/aom/aom_dsp/entenc.c | 423 ++++ third_party/aom/aom_dsp/entenc.h | 85 + third_party/aom/aom_dsp/fastssim.c | 487 ++++ third_party/aom/aom_dsp/fft.c | 219 ++ third_party/aom/aom_dsp/fft_common.h | 1050 +++++++++ third_party/aom/aom_dsp/fwd_txfm.c | 103 + third_party/aom/aom_dsp/grain_synthesis.c | 1409 ++++++++++++ third_party/aom/aom_dsp/grain_synthesis.h | 122 + third_party/aom/aom_dsp/grain_table.c | 333 +++ third_party/aom/aom_dsp/grain_table.h | 102 + third_party/aom/aom_dsp/intrapred.c | 792 +++++++ third_party/aom/aom_dsp/intrapred_common.h | 47 + third_party/aom/aom_dsp/loopfilter.c | 925 ++++++++ third_party/aom/aom_dsp/mips/add_noise_msa.c | 61 + .../aom/aom_dsp/mips/aom_convolve8_horiz_msa.c | 694 ++++++ .../aom/aom_dsp/mips/aom_convolve8_vert_msa.c | 701 ++++++ .../aom/aom_dsp/mips/aom_convolve_copy_msa.c | 248 ++ third_party/aom/aom_dsp/mips/aom_convolve_msa.h | 79 + third_party/aom/aom_dsp/mips/common_dspr2.c | 31 + third_party/aom/aom_dsp/mips/common_dspr2.h | 51 + third_party/aom/aom_dsp/mips/convolve2_dspr2.c | 1031 +++++++++ .../aom/aom_dsp/mips/convolve2_horiz_dspr2.c | 681 ++++++ .../aom/aom_dsp/mips/convolve2_vert_dspr2.c | 237 ++ third_party/aom/aom_dsp/mips/convolve8_dspr2.c | 222 ++ .../aom/aom_dsp/mips/convolve8_horiz_dspr2.c | 879 ++++++++ .../aom/aom_dsp/mips/convolve8_vert_dspr2.c | 361 +++ .../aom/aom_dsp/mips/convolve_common_dspr2.h | 48 + third_party/aom/aom_dsp/mips/intrapred16_dspr2.c | 327 +++ third_party/aom/aom_dsp/mips/intrapred4_dspr2.c | 82 + third_party/aom/aom_dsp/mips/intrapred8_dspr2.c | 150 ++ third_party/aom/aom_dsp/mips/intrapred_msa.c | 550 +++++ third_party/aom/aom_dsp/mips/loopfilter_16_msa.c | 1488 ++++++++++++ third_party/aom/aom_dsp/mips/loopfilter_4_msa.c | 147 ++ third_party/aom/aom_dsp/mips/loopfilter_8_msa.c | 333 +++ .../aom/aom_dsp/mips/loopfilter_filters_dspr2.c | 328 +++ .../aom/aom_dsp/mips/loopfilter_filters_dspr2.h | 736 ++++++ .../aom/aom_dsp/mips/loopfilter_macros_dspr2.h | 437 ++++ .../aom/aom_dsp/mips/loopfilter_masks_dspr2.h | 357 +++ third_party/aom/aom_dsp/mips/loopfilter_mb_dspr2.c | 590 +++++ .../aom/aom_dsp/mips/loopfilter_mb_horiz_dspr2.c | 734 ++++++ .../aom/aom_dsp/mips/loopfilter_mb_vert_dspr2.c | 758 +++++++ third_party/aom/aom_dsp/mips/loopfilter_msa.h | 251 ++ third_party/aom/aom_dsp/mips/macros_msa.h | 2058 +++++++++++++++++ third_party/aom/aom_dsp/mips/sad_msa.c | 800 +++++++ .../aom/aom_dsp/mips/sub_pixel_variance_msa.c | 1792 +++++++++++++++ third_party/aom/aom_dsp/mips/subtract_msa.c | 266 +++ third_party/aom/aom_dsp/mips/variance_msa.c | 633 ++++++ third_party/aom/aom_dsp/noise_model.c | 1648 ++++++++++++++ third_party/aom/aom_dsp/noise_model.h | 323 +++ third_party/aom/aom_dsp/noise_util.c | 221 ++ third_party/aom/aom_dsp/noise_util.h | 68 + third_party/aom/aom_dsp/postproc.h | 26 + third_party/aom/aom_dsp/prob.h | 671 ++++++ third_party/aom/aom_dsp/psnr.c | 381 ++++ third_party/aom/aom_dsp/psnr.h | 79 + third_party/aom/aom_dsp/psnrhvs.c | 272 +++ third_party/aom/aom_dsp/quantize.c | 206 ++ third_party/aom/aom_dsp/quantize.h | 59 + third_party/aom/aom_dsp/sad.c | 304 +++ third_party/aom/aom_dsp/sad_av1.c | 248 ++ third_party/aom/aom_dsp/simd/v128_intrinsics.h | 344 +++ third_party/aom/aom_dsp/simd/v128_intrinsics_arm.h | 958 ++++++++ third_party/aom/aom_dsp/simd/v128_intrinsics_c.h | 888 ++++++++ third_party/aom/aom_dsp/simd/v128_intrinsics_x86.h | 656 ++++++ third_party/aom/aom_dsp/simd/v256_intrinsics.h | 376 +++ third_party/aom/aom_dsp/simd/v256_intrinsics_arm.h | 17 + third_party/aom/aom_dsp/simd/v256_intrinsics_c.h | 953 ++++++++ .../aom/aom_dsp/simd/v256_intrinsics_v128.h | 873 +++++++ third_party/aom/aom_dsp/simd/v256_intrinsics_x86.h | 750 ++++++ third_party/aom/aom_dsp/simd/v64_intrinsics.h | 232 ++ third_party/aom/aom_dsp/simd/v64_intrinsics_arm.h | 680 ++++++ third_party/aom/aom_dsp/simd/v64_intrinsics_c.h | 968 ++++++++ third_party/aom/aom_dsp/simd/v64_intrinsics_x86.h | 491 ++++ third_party/aom/aom_dsp/sse.c | 52 + third_party/aom/aom_dsp/ssim.c | 439 ++++ third_party/aom/aom_dsp/ssim.h | 87 + third_party/aom/aom_dsp/subtract.c | 53 + third_party/aom/aom_dsp/sum_squares.c | 40 + third_party/aom/aom_dsp/txfm_common.h | 91 + third_party/aom/aom_dsp/variance.c | 1579 +++++++++++++ third_party/aom/aom_dsp/variance.h | 130 ++ third_party/aom/aom_dsp/x86/aom_asm_stubs.c | 89 + .../aom/aom_dsp/x86/aom_convolve_copy_sse2.asm | 297 +++ .../aom/aom_dsp/x86/aom_high_subpixel_8t_sse2.asm | 613 +++++ .../x86/aom_high_subpixel_bilinear_sse2.asm | 338 +++ .../aom/aom_dsp/x86/aom_subpixel_8t_intrin_avx2.c | 1441 ++++++++++++ .../aom/aom_dsp/x86/aom_subpixel_8t_intrin_ssse3.c | 315 +++ .../aom/aom_dsp/x86/aom_subpixel_8t_sse2.asm | 615 +++++ .../aom/aom_dsp/x86/aom_subpixel_8t_ssse3.asm | 870 +++++++ .../aom/aom_dsp/x86/aom_subpixel_bilinear_sse2.asm | 295 +++ .../aom_dsp/x86/aom_subpixel_bilinear_ssse3.asm | 267 +++ third_party/aom/aom_dsp/x86/blend_a64_hmask_sse4.c | 34 + third_party/aom/aom_dsp/x86/blend_a64_mask_avx2.c | 900 ++++++++ third_party/aom/aom_dsp/x86/blend_a64_mask_sse4.c | 1109 +++++++++ third_party/aom/aom_dsp/x86/blend_a64_vmask_sse4.c | 283 +++ third_party/aom/aom_dsp/x86/blend_mask_sse4.h | 237 ++ third_party/aom/aom_dsp/x86/blend_sse4.h | 191 ++ third_party/aom/aom_dsp/x86/common_avx2.h | 147 ++ third_party/aom/aom_dsp/x86/convolve.h | 178 ++ third_party/aom/aom_dsp/x86/convolve_avx2.h | 199 ++ .../aom/aom_dsp/x86/convolve_common_intrin.h | 31 + third_party/aom/aom_dsp/x86/convolve_sse2.h | 121 + third_party/aom/aom_dsp/x86/convolve_sse4_1.h | 53 + third_party/aom/aom_dsp/x86/fft_avx2.c | 73 + third_party/aom/aom_dsp/x86/fft_sse2.c | 166 ++ third_party/aom/aom_dsp/x86/fwd_txfm_impl_sse2.h | 344 +++ third_party/aom/aom_dsp/x86/fwd_txfm_sse2.c | 69 + third_party/aom/aom_dsp/x86/fwd_txfm_sse2.h | 155 ++ .../aom/aom_dsp/x86/fwd_txfm_ssse3_x86_64.asm | 379 ++++ third_party/aom/aom_dsp/x86/highbd_convolve_avx2.c | 998 ++++++++ .../aom/aom_dsp/x86/highbd_convolve_ssse3.c | 251 ++ .../aom/aom_dsp/x86/highbd_intrapred_sse2.c | 984 ++++++++ .../aom/aom_dsp/x86/highbd_intrapred_sse2_asm.asm | 259 +++ .../aom/aom_dsp/x86/highbd_loopfilter_avx2.c | 66 + .../aom/aom_dsp/x86/highbd_loopfilter_sse2.c | 1697 ++++++++++++++ .../aom/aom_dsp/x86/highbd_quantize_intrin_avx2.c | 160 ++ .../aom/aom_dsp/x86/highbd_quantize_intrin_sse2.c | 148 ++ third_party/aom/aom_dsp/x86/highbd_sad4d_sse2.asm | 296 +++ third_party/aom/aom_dsp/x86/highbd_sad_sse2.asm | 374 +++ .../x86/highbd_subpel_variance_impl_sse2.asm | 1036 +++++++++ third_party/aom/aom_dsp/x86/highbd_subtract_sse2.c | 267 +++ third_party/aom/aom_dsp/x86/highbd_variance_avx2.c | 140 ++ .../aom/aom_dsp/x86/highbd_variance_impl_sse2.asm | 318 +++ third_party/aom/aom_dsp/x86/highbd_variance_sse2.c | 868 +++++++ third_party/aom/aom_dsp/x86/highbd_variance_sse4.c | 216 ++ third_party/aom/aom_dsp/x86/intrapred_avx2.c | 811 +++++++ third_party/aom/aom_dsp/x86/intrapred_sse2.c | 1430 ++++++++++++ third_party/aom/aom_dsp/x86/intrapred_sse2_asm.asm | 625 +++++ third_party/aom/aom_dsp/x86/intrapred_ssse3.c | 1692 ++++++++++++++ third_party/aom/aom_dsp/x86/inv_wht_sse2.asm | 107 + third_party/aom/aom_dsp/x86/jnt_sad_ssse3.c | 238 ++ third_party/aom/aom_dsp/x86/jnt_variance_ssse3.c | 192 ++ third_party/aom/aom_dsp/x86/loopfilter_sse2.c | 2385 ++++++++++++++++++++ third_party/aom/aom_dsp/x86/lpf_common_sse2.h | 215 ++ .../aom/aom_dsp/x86/masked_sad_intrin_avx2.c | 389 ++++ .../aom/aom_dsp/x86/masked_sad_intrin_ssse3.c | 402 ++++ .../aom/aom_dsp/x86/masked_sad_intrin_ssse3.h | 33 + .../aom/aom_dsp/x86/masked_variance_intrin_ssse3.c | 1065 +++++++++ .../aom/aom_dsp/x86/masked_variance_intrin_ssse3.h | 92 + third_party/aom/aom_dsp/x86/mem_sse2.h | 42 + third_party/aom/aom_dsp/x86/obmc_intrinsic_sse4.h | 58 + third_party/aom/aom_dsp/x86/obmc_intrinsic_ssse3.h | 54 + third_party/aom/aom_dsp/x86/obmc_sad_avx2.c | 270 +++ third_party/aom/aom_dsp/x86/obmc_sad_sse4.c | 268 +++ third_party/aom/aom_dsp/x86/obmc_variance_avx2.c | 190 ++ third_party/aom/aom_dsp/x86/obmc_variance_sse4.c | 380 ++++ .../aom/aom_dsp/x86/quantize_avx_x86_64.asm | 435 ++++ third_party/aom/aom_dsp/x86/quantize_sse2.c | 147 ++ .../aom/aom_dsp/x86/quantize_ssse3_x86_64.asm | 272 +++ third_party/aom/aom_dsp/x86/quantize_x86.h | 77 + third_party/aom/aom_dsp/x86/sad4d_avx2.c | 218 ++ third_party/aom/aom_dsp/x86/sad4d_sse2.asm | 257 +++ third_party/aom/aom_dsp/x86/sad_avx2.c | 189 ++ third_party/aom/aom_dsp/x86/sad_highbd_avx2.c | 1038 +++++++++ third_party/aom/aom_dsp/x86/sad_impl_avx2.c | 234 ++ third_party/aom/aom_dsp/x86/sad_sse2.asm | 353 +++ third_party/aom/aom_dsp/x86/sse_avx2.c | 250 ++ third_party/aom/aom_dsp/x86/sse_sse4.c | 241 ++ third_party/aom/aom_dsp/x86/ssim_opt_x86_64.asm | 222 ++ .../aom/aom_dsp/x86/subpel_variance_sse2.asm | 1481 ++++++++++++ third_party/aom/aom_dsp/x86/subtract_avx2.c | 108 + third_party/aom/aom_dsp/x86/subtract_sse2.asm | 146 ++ third_party/aom/aom_dsp/x86/sum_squares_avx2.c | 79 + third_party/aom/aom_dsp/x86/sum_squares_sse2.c | 203 ++ third_party/aom/aom_dsp/x86/sum_squares_sse2.h | 22 + third_party/aom/aom_dsp/x86/synonyms.h | 114 + third_party/aom/aom_dsp/x86/synonyms_avx2.h | 74 + third_party/aom/aom_dsp/x86/transpose_sse2.h | 420 ++++ third_party/aom/aom_dsp/x86/txfm_common_avx2.h | 199 ++ third_party/aom/aom_dsp/x86/txfm_common_sse2.h | 29 + third_party/aom/aom_dsp/x86/variance_avx2.c | 517 +++++ third_party/aom/aom_dsp/x86/variance_impl_avx2.c | 517 +++++ third_party/aom/aom_dsp/x86/variance_impl_ssse3.c | 129 ++ third_party/aom/aom_dsp/x86/variance_sse2.c | 806 +++++++ 216 files changed, 85134 insertions(+) create mode 100644 third_party/aom/aom_dsp/add_noise.c create mode 100644 third_party/aom/aom_dsp/aom_convolve.c create mode 100644 third_party/aom/aom_dsp/aom_dsp.cmake create mode 100644 third_party/aom/aom_dsp/aom_dsp_common.h create mode 100644 third_party/aom/aom_dsp/aom_dsp_rtcd.c create mode 100755 third_party/aom/aom_dsp/aom_dsp_rtcd_defs.pl create mode 100644 third_party/aom/aom_dsp/aom_filter.h create mode 100644 third_party/aom/aom_dsp/aom_simd.h create mode 100644 third_party/aom/aom_dsp/aom_simd_inline.h create mode 100644 third_party/aom/aom_dsp/arm/blend_a64_mask_neon.c create mode 100644 third_party/aom/aom_dsp/arm/fwd_txfm_neon.c create mode 100644 third_party/aom/aom_dsp/arm/intrapred_neon.c create mode 100644 third_party/aom/aom_dsp/arm/loopfilter_neon.c create mode 100644 third_party/aom/aom_dsp/arm/sad4d_neon.c create mode 100644 third_party/aom/aom_dsp/arm/sad_neon.c create mode 100644 third_party/aom/aom_dsp/arm/subpel_variance_neon.c create mode 100644 third_party/aom/aom_dsp/arm/subtract_neon.c create mode 100644 third_party/aom/aom_dsp/arm/variance_neon.c create mode 100644 third_party/aom/aom_dsp/binary_codes_reader.c create mode 100644 third_party/aom/aom_dsp/binary_codes_reader.h create mode 100644 third_party/aom/aom_dsp/binary_codes_writer.c create mode 100644 third_party/aom/aom_dsp/binary_codes_writer.h create mode 100644 third_party/aom/aom_dsp/bitreader.h create mode 100644 third_party/aom/aom_dsp/bitreader_buffer.c create mode 100644 third_party/aom/aom_dsp/bitreader_buffer.h create mode 100644 third_party/aom/aom_dsp/bitwriter.h create mode 100644 third_party/aom/aom_dsp/bitwriter_buffer.c create mode 100644 third_party/aom/aom_dsp/bitwriter_buffer.h create mode 100644 third_party/aom/aom_dsp/blend.h create mode 100644 third_party/aom/aom_dsp/blend_a64_hmask.c create mode 100644 third_party/aom/aom_dsp/blend_a64_mask.c create mode 100644 third_party/aom/aom_dsp/blend_a64_vmask.c create mode 100644 third_party/aom/aom_dsp/buf_ans.c create mode 100644 third_party/aom/aom_dsp/buf_ans.h create mode 100644 third_party/aom/aom_dsp/daalaboolreader.c create mode 100644 third_party/aom/aom_dsp/daalaboolreader.h create mode 100644 third_party/aom/aom_dsp/daalaboolwriter.c create mode 100644 third_party/aom/aom_dsp/daalaboolwriter.h create mode 100644 third_party/aom/aom_dsp/entcode.c create mode 100644 third_party/aom/aom_dsp/entcode.h create mode 100644 third_party/aom/aom_dsp/entdec.c create mode 100644 third_party/aom/aom_dsp/entdec.h create mode 100644 third_party/aom/aom_dsp/entenc.c create mode 100644 third_party/aom/aom_dsp/entenc.h create mode 100644 third_party/aom/aom_dsp/fastssim.c create mode 100644 third_party/aom/aom_dsp/fft.c create mode 100644 third_party/aom/aom_dsp/fft_common.h create mode 100644 third_party/aom/aom_dsp/fwd_txfm.c create mode 100644 third_party/aom/aom_dsp/grain_synthesis.c create mode 100644 third_party/aom/aom_dsp/grain_synthesis.h create mode 100644 third_party/aom/aom_dsp/grain_table.c create mode 100644 third_party/aom/aom_dsp/grain_table.h create mode 100644 third_party/aom/aom_dsp/intrapred.c create mode 100644 third_party/aom/aom_dsp/intrapred_common.h create mode 100644 third_party/aom/aom_dsp/loopfilter.c create mode 100644 third_party/aom/aom_dsp/mips/add_noise_msa.c create mode 100644 third_party/aom/aom_dsp/mips/aom_convolve8_horiz_msa.c create mode 100644 third_party/aom/aom_dsp/mips/aom_convolve8_vert_msa.c create mode 100644 third_party/aom/aom_dsp/mips/aom_convolve_copy_msa.c create mode 100644 third_party/aom/aom_dsp/mips/aom_convolve_msa.h create mode 100644 third_party/aom/aom_dsp/mips/common_dspr2.c create mode 100644 third_party/aom/aom_dsp/mips/common_dspr2.h create mode 100644 third_party/aom/aom_dsp/mips/convolve2_dspr2.c create mode 100644 third_party/aom/aom_dsp/mips/convolve2_horiz_dspr2.c create mode 100644 third_party/aom/aom_dsp/mips/convolve2_vert_dspr2.c create mode 100644 third_party/aom/aom_dsp/mips/convolve8_dspr2.c create mode 100644 third_party/aom/aom_dsp/mips/convolve8_horiz_dspr2.c create mode 100644 third_party/aom/aom_dsp/mips/convolve8_vert_dspr2.c create mode 100644 third_party/aom/aom_dsp/mips/convolve_common_dspr2.h create mode 100644 third_party/aom/aom_dsp/mips/intrapred16_dspr2.c create mode 100644 third_party/aom/aom_dsp/mips/intrapred4_dspr2.c create mode 100644 third_party/aom/aom_dsp/mips/intrapred8_dspr2.c create mode 100644 third_party/aom/aom_dsp/mips/intrapred_msa.c create mode 100644 third_party/aom/aom_dsp/mips/loopfilter_16_msa.c create mode 100644 third_party/aom/aom_dsp/mips/loopfilter_4_msa.c create mode 100644 third_party/aom/aom_dsp/mips/loopfilter_8_msa.c create mode 100644 third_party/aom/aom_dsp/mips/loopfilter_filters_dspr2.c create mode 100644 third_party/aom/aom_dsp/mips/loopfilter_filters_dspr2.h create mode 100644 third_party/aom/aom_dsp/mips/loopfilter_macros_dspr2.h create mode 100644 third_party/aom/aom_dsp/mips/loopfilter_masks_dspr2.h create mode 100644 third_party/aom/aom_dsp/mips/loopfilter_mb_dspr2.c create mode 100644 third_party/aom/aom_dsp/mips/loopfilter_mb_horiz_dspr2.c create mode 100644 third_party/aom/aom_dsp/mips/loopfilter_mb_vert_dspr2.c create mode 100644 third_party/aom/aom_dsp/mips/loopfilter_msa.h create mode 100644 third_party/aom/aom_dsp/mips/macros_msa.h create mode 100644 third_party/aom/aom_dsp/mips/sad_msa.c create mode 100644 third_party/aom/aom_dsp/mips/sub_pixel_variance_msa.c create mode 100644 third_party/aom/aom_dsp/mips/subtract_msa.c create mode 100644 third_party/aom/aom_dsp/mips/variance_msa.c create mode 100644 third_party/aom/aom_dsp/noise_model.c create mode 100644 third_party/aom/aom_dsp/noise_model.h create mode 100644 third_party/aom/aom_dsp/noise_util.c create mode 100644 third_party/aom/aom_dsp/noise_util.h create mode 100644 third_party/aom/aom_dsp/postproc.h create mode 100644 third_party/aom/aom_dsp/prob.h create mode 100644 third_party/aom/aom_dsp/psnr.c create mode 100644 third_party/aom/aom_dsp/psnr.h create mode 100644 third_party/aom/aom_dsp/psnrhvs.c create mode 100644 third_party/aom/aom_dsp/quantize.c create mode 100644 third_party/aom/aom_dsp/quantize.h create mode 100644 third_party/aom/aom_dsp/sad.c create mode 100644 third_party/aom/aom_dsp/sad_av1.c create mode 100644 third_party/aom/aom_dsp/simd/v128_intrinsics.h create mode 100644 third_party/aom/aom_dsp/simd/v128_intrinsics_arm.h create mode 100644 third_party/aom/aom_dsp/simd/v128_intrinsics_c.h create mode 100644 third_party/aom/aom_dsp/simd/v128_intrinsics_x86.h create mode 100644 third_party/aom/aom_dsp/simd/v256_intrinsics.h create mode 100644 third_party/aom/aom_dsp/simd/v256_intrinsics_arm.h create mode 100644 third_party/aom/aom_dsp/simd/v256_intrinsics_c.h create mode 100644 third_party/aom/aom_dsp/simd/v256_intrinsics_v128.h create mode 100644 third_party/aom/aom_dsp/simd/v256_intrinsics_x86.h create mode 100644 third_party/aom/aom_dsp/simd/v64_intrinsics.h create mode 100644 third_party/aom/aom_dsp/simd/v64_intrinsics_arm.h create mode 100644 third_party/aom/aom_dsp/simd/v64_intrinsics_c.h create mode 100644 third_party/aom/aom_dsp/simd/v64_intrinsics_x86.h create mode 100644 third_party/aom/aom_dsp/sse.c create mode 100644 third_party/aom/aom_dsp/ssim.c create mode 100644 third_party/aom/aom_dsp/ssim.h create mode 100644 third_party/aom/aom_dsp/subtract.c create mode 100644 third_party/aom/aom_dsp/sum_squares.c create mode 100644 third_party/aom/aom_dsp/txfm_common.h create mode 100644 third_party/aom/aom_dsp/variance.c create mode 100644 third_party/aom/aom_dsp/variance.h create mode 100644 third_party/aom/aom_dsp/x86/aom_asm_stubs.c create mode 100644 third_party/aom/aom_dsp/x86/aom_convolve_copy_sse2.asm create mode 100644 third_party/aom/aom_dsp/x86/aom_high_subpixel_8t_sse2.asm create mode 100644 third_party/aom/aom_dsp/x86/aom_high_subpixel_bilinear_sse2.asm create mode 100644 third_party/aom/aom_dsp/x86/aom_subpixel_8t_intrin_avx2.c create mode 100644 third_party/aom/aom_dsp/x86/aom_subpixel_8t_intrin_ssse3.c create mode 100644 third_party/aom/aom_dsp/x86/aom_subpixel_8t_sse2.asm create mode 100644 third_party/aom/aom_dsp/x86/aom_subpixel_8t_ssse3.asm create mode 100644 third_party/aom/aom_dsp/x86/aom_subpixel_bilinear_sse2.asm create mode 100644 third_party/aom/aom_dsp/x86/aom_subpixel_bilinear_ssse3.asm create mode 100644 third_party/aom/aom_dsp/x86/blend_a64_hmask_sse4.c create mode 100644 third_party/aom/aom_dsp/x86/blend_a64_mask_avx2.c create mode 100644 third_party/aom/aom_dsp/x86/blend_a64_mask_sse4.c create mode 100644 third_party/aom/aom_dsp/x86/blend_a64_vmask_sse4.c create mode 100644 third_party/aom/aom_dsp/x86/blend_mask_sse4.h create mode 100644 third_party/aom/aom_dsp/x86/blend_sse4.h create mode 100644 third_party/aom/aom_dsp/x86/common_avx2.h create mode 100644 third_party/aom/aom_dsp/x86/convolve.h create mode 100644 third_party/aom/aom_dsp/x86/convolve_avx2.h create mode 100644 third_party/aom/aom_dsp/x86/convolve_common_intrin.h create mode 100644 third_party/aom/aom_dsp/x86/convolve_sse2.h create mode 100644 third_party/aom/aom_dsp/x86/convolve_sse4_1.h create mode 100644 third_party/aom/aom_dsp/x86/fft_avx2.c create mode 100644 third_party/aom/aom_dsp/x86/fft_sse2.c create mode 100644 third_party/aom/aom_dsp/x86/fwd_txfm_impl_sse2.h create mode 100644 third_party/aom/aom_dsp/x86/fwd_txfm_sse2.c create mode 100644 third_party/aom/aom_dsp/x86/fwd_txfm_sse2.h create mode 100644 third_party/aom/aom_dsp/x86/fwd_txfm_ssse3_x86_64.asm create mode 100644 third_party/aom/aom_dsp/x86/highbd_convolve_avx2.c create mode 100644 third_party/aom/aom_dsp/x86/highbd_convolve_ssse3.c create mode 100644 third_party/aom/aom_dsp/x86/highbd_intrapred_sse2.c create mode 100644 third_party/aom/aom_dsp/x86/highbd_intrapred_sse2_asm.asm create mode 100644 third_party/aom/aom_dsp/x86/highbd_loopfilter_avx2.c create mode 100644 third_party/aom/aom_dsp/x86/highbd_loopfilter_sse2.c create mode 100644 third_party/aom/aom_dsp/x86/highbd_quantize_intrin_avx2.c create mode 100644 third_party/aom/aom_dsp/x86/highbd_quantize_intrin_sse2.c create mode 100644 third_party/aom/aom_dsp/x86/highbd_sad4d_sse2.asm create mode 100644 third_party/aom/aom_dsp/x86/highbd_sad_sse2.asm create mode 100644 third_party/aom/aom_dsp/x86/highbd_subpel_variance_impl_sse2.asm create mode 100644 third_party/aom/aom_dsp/x86/highbd_subtract_sse2.c create mode 100644 third_party/aom/aom_dsp/x86/highbd_variance_avx2.c create mode 100644 third_party/aom/aom_dsp/x86/highbd_variance_impl_sse2.asm create mode 100644 third_party/aom/aom_dsp/x86/highbd_variance_sse2.c create mode 100644 third_party/aom/aom_dsp/x86/highbd_variance_sse4.c create mode 100644 third_party/aom/aom_dsp/x86/intrapred_avx2.c create mode 100644 third_party/aom/aom_dsp/x86/intrapred_sse2.c create mode 100644 third_party/aom/aom_dsp/x86/intrapred_sse2_asm.asm create mode 100644 third_party/aom/aom_dsp/x86/intrapred_ssse3.c create mode 100644 third_party/aom/aom_dsp/x86/inv_wht_sse2.asm create mode 100644 third_party/aom/aom_dsp/x86/jnt_sad_ssse3.c create mode 100644 third_party/aom/aom_dsp/x86/jnt_variance_ssse3.c create mode 100644 third_party/aom/aom_dsp/x86/loopfilter_sse2.c create mode 100644 third_party/aom/aom_dsp/x86/lpf_common_sse2.h create mode 100644 third_party/aom/aom_dsp/x86/masked_sad_intrin_avx2.c create mode 100644 third_party/aom/aom_dsp/x86/masked_sad_intrin_ssse3.c create mode 100644 third_party/aom/aom_dsp/x86/masked_sad_intrin_ssse3.h create mode 100644 third_party/aom/aom_dsp/x86/masked_variance_intrin_ssse3.c create mode 100644 third_party/aom/aom_dsp/x86/masked_variance_intrin_ssse3.h create mode 100644 third_party/aom/aom_dsp/x86/mem_sse2.h create mode 100644 third_party/aom/aom_dsp/x86/obmc_intrinsic_sse4.h create mode 100644 third_party/aom/aom_dsp/x86/obmc_intrinsic_ssse3.h create mode 100644 third_party/aom/aom_dsp/x86/obmc_sad_avx2.c create mode 100644 third_party/aom/aom_dsp/x86/obmc_sad_sse4.c create mode 100644 third_party/aom/aom_dsp/x86/obmc_variance_avx2.c create mode 100644 third_party/aom/aom_dsp/x86/obmc_variance_sse4.c create mode 100644 third_party/aom/aom_dsp/x86/quantize_avx_x86_64.asm create mode 100644 third_party/aom/aom_dsp/x86/quantize_sse2.c create mode 100644 third_party/aom/aom_dsp/x86/quantize_ssse3_x86_64.asm create mode 100644 third_party/aom/aom_dsp/x86/quantize_x86.h create mode 100644 third_party/aom/aom_dsp/x86/sad4d_avx2.c create mode 100644 third_party/aom/aom_dsp/x86/sad4d_sse2.asm create mode 100644 third_party/aom/aom_dsp/x86/sad_avx2.c create mode 100644 third_party/aom/aom_dsp/x86/sad_highbd_avx2.c create mode 100644 third_party/aom/aom_dsp/x86/sad_impl_avx2.c create mode 100644 third_party/aom/aom_dsp/x86/sad_sse2.asm create mode 100644 third_party/aom/aom_dsp/x86/sse_avx2.c create mode 100644 third_party/aom/aom_dsp/x86/sse_sse4.c create mode 100644 third_party/aom/aom_dsp/x86/ssim_opt_x86_64.asm create mode 100644 third_party/aom/aom_dsp/x86/subpel_variance_sse2.asm create mode 100644 third_party/aom/aom_dsp/x86/subtract_avx2.c create mode 100644 third_party/aom/aom_dsp/x86/subtract_sse2.asm create mode 100644 third_party/aom/aom_dsp/x86/sum_squares_avx2.c create mode 100644 third_party/aom/aom_dsp/x86/sum_squares_sse2.c create mode 100644 third_party/aom/aom_dsp/x86/sum_squares_sse2.h create mode 100644 third_party/aom/aom_dsp/x86/synonyms.h create mode 100644 third_party/aom/aom_dsp/x86/synonyms_avx2.h create mode 100644 third_party/aom/aom_dsp/x86/transpose_sse2.h create mode 100644 third_party/aom/aom_dsp/x86/txfm_common_avx2.h create mode 100644 third_party/aom/aom_dsp/x86/txfm_common_sse2.h create mode 100644 third_party/aom/aom_dsp/x86/variance_avx2.c create mode 100644 third_party/aom/aom_dsp/x86/variance_impl_avx2.c create mode 100644 third_party/aom/aom_dsp/x86/variance_impl_ssse3.c create mode 100644 third_party/aom/aom_dsp/x86/variance_sse2.c (limited to 'third_party/aom/aom_dsp') diff --git a/third_party/aom/aom_dsp/add_noise.c b/third_party/aom/aom_dsp/add_noise.c new file mode 100644 index 0000000000..bfb3e7e009 --- /dev/null +++ b/third_party/aom/aom_dsp/add_noise.c @@ -0,0 +1,73 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include +#include + +#include "config/aom_config.h" +#include "config/aom_dsp_rtcd.h" + +#include "aom/aom_integer.h" +#include "aom_ports/mem.h" + +void aom_plane_add_noise_c(uint8_t *start, char *noise, char blackclamp[16], + char whiteclamp[16], char bothclamp[16], + unsigned int width, unsigned int height, int pitch) { + unsigned int i, j; + + for (i = 0; i < height; ++i) { + uint8_t *pos = start + i * pitch; + char *ref = (char *)(noise + (rand() & 0xff)); // NOLINT + + for (j = 0; j < width; ++j) { + int v = pos[j]; + + v = clamp(v - blackclamp[0], 0, 255); + v = clamp(v + bothclamp[0], 0, 255); + v = clamp(v - whiteclamp[0], 0, 255); + + pos[j] = v + ref[j]; + } + } +} + +static double gaussian(double sigma, double mu, double x) { + return 1 / (sigma * sqrt(2.0 * 3.14159265)) * + (exp(-(x - mu) * (x - mu) / (2 * sigma * sigma))); +} + +int aom_setup_noise(double sigma, int size, char *noise) { + char char_dist[256]; + int next = 0, i, j; + + // set up a 256 entry lookup that matches gaussian distribution + for (i = -32; i < 32; ++i) { + const int a_i = (int)(0.5 + 256 * gaussian(sigma, 0, i)); + if (a_i) { + for (j = 0; j < a_i; ++j) { + char_dist[next + j] = (char)i; + } + next = next + j; + } + } + + // Rounding error - might mean we have less than 256. + for (; next < 256; ++next) { + char_dist[next] = 0; + } + + for (i = 0; i < size; ++i) { + noise[i] = char_dist[rand() & 0xff]; // NOLINT + } + + // Returns the highest non 0 value used in distribution. + return -char_dist[0]; +} diff --git a/third_party/aom/aom_dsp/aom_convolve.c b/third_party/aom/aom_dsp/aom_convolve.c new file mode 100644 index 0000000000..4791826da2 --- /dev/null +++ b/third_party/aom/aom_dsp/aom_convolve.c @@ -0,0 +1,238 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include +#include + +#include "config/aom_config.h" +#include "config/aom_dsp_rtcd.h" + +#include "aom/aom_integer.h" +#include "aom_dsp/aom_dsp_common.h" +#include "aom_dsp/aom_filter.h" +#include "aom_ports/mem.h" + +static INLINE int horz_scalar_product(const uint8_t *a, const int16_t *b) { + int sum = 0; + for (int k = 0; k < SUBPEL_TAPS; ++k) sum += a[k] * b[k]; + return sum; +} + +static INLINE int vert_scalar_product(const uint8_t *a, ptrdiff_t a_stride, + const int16_t *b) { + int sum = 0; + for (int k = 0; k < SUBPEL_TAPS; ++k) sum += a[k * a_stride] * b[k]; + return sum; +} + +static void convolve_horiz(const uint8_t *src, ptrdiff_t src_stride, + uint8_t *dst, ptrdiff_t dst_stride, + const InterpKernel *x_filters, int x0_q4, + int x_step_q4, int w, int h) { + src -= SUBPEL_TAPS / 2 - 1; + for (int y = 0; y < h; ++y) { + int x_q4 = x0_q4; + for (int x = 0; x < w; ++x) { + const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS]; + const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK]; + const int sum = horz_scalar_product(src_x, x_filter); + dst[x] = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS)); + x_q4 += x_step_q4; + } + src += src_stride; + dst += dst_stride; + } +} + +static void convolve_vert(const uint8_t *src, ptrdiff_t src_stride, + uint8_t *dst, ptrdiff_t dst_stride, + const InterpKernel *y_filters, int y0_q4, + int y_step_q4, int w, int h) { + src -= src_stride * (SUBPEL_TAPS / 2 - 1); + + for (int x = 0; x < w; ++x) { + int y_q4 = y0_q4; + for (int y = 0; y < h; ++y) { + const unsigned char *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride]; + const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK]; + const int sum = vert_scalar_product(src_y, src_stride, y_filter); + dst[y * dst_stride] = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS)); + y_q4 += y_step_q4; + } + ++src; + ++dst; + } +} + +static const InterpKernel *get_filter_base(const int16_t *filter) { + // NOTE: This assumes that the filter table is 256-byte aligned. + // TODO(agrange) Modify to make independent of table alignment. + return (const InterpKernel *)(((intptr_t)filter) & ~((intptr_t)0xFF)); +} + +static int get_filter_offset(const int16_t *f, const InterpKernel *base) { + return (int)((const InterpKernel *)(intptr_t)f - base); +} + +void aom_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, + uint8_t *dst, ptrdiff_t dst_stride, + const int16_t *filter_x, int x_step_q4, + const int16_t *filter_y, int y_step_q4, int w, + int h) { + const InterpKernel *const filters_x = get_filter_base(filter_x); + const int x0_q4 = get_filter_offset(filter_x, filters_x); + + (void)filter_y; + (void)y_step_q4; + + convolve_horiz(src, src_stride, dst, dst_stride, filters_x, x0_q4, x_step_q4, + w, h); +} + +void aom_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, + uint8_t *dst, ptrdiff_t dst_stride, + const int16_t *filter_x, int x_step_q4, + const int16_t *filter_y, int y_step_q4, int w, + int h) { + const InterpKernel *const filters_y = get_filter_base(filter_y); + const int y0_q4 = get_filter_offset(filter_y, filters_y); + + (void)filter_x; + (void)x_step_q4; + + convolve_vert(src, src_stride, dst, dst_stride, filters_y, y0_q4, y_step_q4, + w, h); +} + +void aom_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, + ptrdiff_t dst_stride, const int16_t *filter_x, + int filter_x_stride, const int16_t *filter_y, + int filter_y_stride, int w, int h) { + int r; + + (void)filter_x; + (void)filter_x_stride; + (void)filter_y; + (void)filter_y_stride; + + for (r = h; r > 0; --r) { + memcpy(dst, src, w); + src += src_stride; + dst += dst_stride; + } +} + +static INLINE int highbd_vert_scalar_product(const uint16_t *a, + ptrdiff_t a_stride, + const int16_t *b) { + int sum = 0; + for (int k = 0; k < SUBPEL_TAPS; ++k) sum += a[k * a_stride] * b[k]; + return sum; +} + +static INLINE int highbd_horz_scalar_product(const uint16_t *a, + const int16_t *b) { + int sum = 0; + for (int k = 0; k < SUBPEL_TAPS; ++k) sum += a[k] * b[k]; + return sum; +} + +static void highbd_convolve_horiz(const uint8_t *src8, ptrdiff_t src_stride, + uint8_t *dst8, ptrdiff_t dst_stride, + const InterpKernel *x_filters, int x0_q4, + int x_step_q4, int w, int h, int bd) { + uint16_t *src = CONVERT_TO_SHORTPTR(src8); + uint16_t *dst = CONVERT_TO_SHORTPTR(dst8); + src -= SUBPEL_TAPS / 2 - 1; + for (int y = 0; y < h; ++y) { + int x_q4 = x0_q4; + for (int x = 0; x < w; ++x) { + const uint16_t *const src_x = &src[x_q4 >> SUBPEL_BITS]; + const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK]; + const int sum = highbd_horz_scalar_product(src_x, x_filter); + dst[x] = clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd); + x_q4 += x_step_q4; + } + src += src_stride; + dst += dst_stride; + } +} + +static void highbd_convolve_vert(const uint8_t *src8, ptrdiff_t src_stride, + uint8_t *dst8, ptrdiff_t dst_stride, + const InterpKernel *y_filters, int y0_q4, + int y_step_q4, int w, int h, int bd) { + uint16_t *src = CONVERT_TO_SHORTPTR(src8); + uint16_t *dst = CONVERT_TO_SHORTPTR(dst8); + src -= src_stride * (SUBPEL_TAPS / 2 - 1); + for (int x = 0; x < w; ++x) { + int y_q4 = y0_q4; + for (int y = 0; y < h; ++y) { + const uint16_t *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride]; + const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK]; + const int sum = highbd_vert_scalar_product(src_y, src_stride, y_filter); + dst[y * dst_stride] = + clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd); + y_q4 += y_step_q4; + } + ++src; + ++dst; + } +} + +void aom_highbd_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, + uint8_t *dst, ptrdiff_t dst_stride, + const int16_t *filter_x, int x_step_q4, + const int16_t *filter_y, int y_step_q4, int w, + int h, int bd) { + const InterpKernel *const filters_x = get_filter_base(filter_x); + const int x0_q4 = get_filter_offset(filter_x, filters_x); + (void)filter_y; + (void)y_step_q4; + + highbd_convolve_horiz(src, src_stride, dst, dst_stride, filters_x, x0_q4, + x_step_q4, w, h, bd); +} + +void aom_highbd_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, + uint8_t *dst, ptrdiff_t dst_stride, + const int16_t *filter_x, int x_step_q4, + const int16_t *filter_y, int y_step_q4, int w, + int h, int bd) { + const InterpKernel *const filters_y = get_filter_base(filter_y); + const int y0_q4 = get_filter_offset(filter_y, filters_y); + (void)filter_x; + (void)x_step_q4; + + highbd_convolve_vert(src, src_stride, dst, dst_stride, filters_y, y0_q4, + y_step_q4, w, h, bd); +} + +void aom_highbd_convolve_copy_c(const uint8_t *src8, ptrdiff_t src_stride, + uint8_t *dst8, ptrdiff_t dst_stride, + const int16_t *filter_x, int filter_x_stride, + const int16_t *filter_y, int filter_y_stride, + int w, int h, int bd) { + int r; + uint16_t *src = CONVERT_TO_SHORTPTR(src8); + uint16_t *dst = CONVERT_TO_SHORTPTR(dst8); + (void)filter_x; + (void)filter_y; + (void)filter_x_stride; + (void)filter_y_stride; + (void)bd; + + for (r = h; r > 0; --r) { + memcpy(dst, src, w * sizeof(uint16_t)); + src += src_stride; + dst += dst_stride; + } +} diff --git a/third_party/aom/aom_dsp/aom_dsp.cmake b/third_party/aom/aom_dsp/aom_dsp.cmake new file mode 100644 index 0000000000..11ff73756b --- /dev/null +++ b/third_party/aom/aom_dsp/aom_dsp.cmake @@ -0,0 +1,356 @@ +# +# Copyright (c) 2017, Alliance for Open Media. All rights reserved +# +# This source code is subject to the terms of the BSD 2 Clause License and the +# Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License was +# not distributed with this source code in the LICENSE file, you can obtain it +# at www.aomedia.org/license/software. If the Alliance for Open Media Patent +# License 1.0 was not distributed with this source code in the PATENTS file, you +# can obtain it at www.aomedia.org/license/patent. +# +if(AOM_AOM_DSP_AOM_DSP_CMAKE_) + return() +endif() # AOM_AOM_DSP_AOM_DSP_CMAKE_ +set(AOM_AOM_DSP_AOM_DSP_CMAKE_ 1) + +list(APPEND AOM_DSP_COMMON_SOURCES + "${AOM_ROOT}/aom_dsp/aom_convolve.c" + "${AOM_ROOT}/aom_dsp/aom_dsp_common.h" + "${AOM_ROOT}/aom_dsp/aom_filter.h" + "${AOM_ROOT}/aom_dsp/aom_simd.h" + "${AOM_ROOT}/aom_dsp/aom_simd_inline.h" + "${AOM_ROOT}/aom_dsp/bitreader_buffer.c" + "${AOM_ROOT}/aom_dsp/bitreader_buffer.h" + "${AOM_ROOT}/aom_dsp/bitwriter_buffer.c" + "${AOM_ROOT}/aom_dsp/bitwriter_buffer.h" + "${AOM_ROOT}/aom_dsp/blend.h" + "${AOM_ROOT}/aom_dsp/blend_a64_hmask.c" + "${AOM_ROOT}/aom_dsp/blend_a64_mask.c" + "${AOM_ROOT}/aom_dsp/blend_a64_vmask.c" + "${AOM_ROOT}/aom_dsp/entcode.c" + "${AOM_ROOT}/aom_dsp/entcode.h" + "${AOM_ROOT}/aom_dsp/fft.c" + "${AOM_ROOT}/aom_dsp/fft_common.h" + "${AOM_ROOT}/aom_dsp/intrapred.c" + "${AOM_ROOT}/aom_dsp/intrapred_common.h" + "${AOM_ROOT}/aom_dsp/loopfilter.c" + "${AOM_ROOT}/aom_dsp/prob.h" + "${AOM_ROOT}/aom_dsp/simd/v128_intrinsics.h" + "${AOM_ROOT}/aom_dsp/simd/v128_intrinsics_c.h" + "${AOM_ROOT}/aom_dsp/simd/v256_intrinsics.h" + "${AOM_ROOT}/aom_dsp/simd/v256_intrinsics_c.h" + "${AOM_ROOT}/aom_dsp/simd/v64_intrinsics.h" + "${AOM_ROOT}/aom_dsp/simd/v64_intrinsics_c.h" + "${AOM_ROOT}/aom_dsp/subtract.c" + "${AOM_ROOT}/aom_dsp/txfm_common.h" + "${AOM_ROOT}/aom_dsp/x86/convolve_common_intrin.h") + +list(APPEND AOM_DSP_COMMON_ASM_SSE2 + "${AOM_ROOT}/aom_dsp/x86/aom_convolve_copy_sse2.asm" + "${AOM_ROOT}/aom_dsp/x86/aom_high_subpixel_8t_sse2.asm" + "${AOM_ROOT}/aom_dsp/x86/aom_high_subpixel_bilinear_sse2.asm" + "${AOM_ROOT}/aom_dsp/x86/aom_subpixel_8t_sse2.asm" + "${AOM_ROOT}/aom_dsp/x86/aom_subpixel_bilinear_sse2.asm" + "${AOM_ROOT}/aom_dsp/x86/highbd_intrapred_sse2.asm" + "${AOM_ROOT}/aom_dsp/x86/intrapred_sse2.asm" + "${AOM_ROOT}/aom_dsp/x86/inv_wht_sse2.asm") + +list(APPEND AOM_DSP_COMMON_INTRIN_SSE2 + "${AOM_ROOT}/aom_dsp/x86/aom_asm_stubs.c" + "${AOM_ROOT}/aom_dsp/x86/convolve.h" + "${AOM_ROOT}/aom_dsp/x86/convolve_sse2.h" + "${AOM_ROOT}/aom_dsp/x86/fft_sse2.c" + "${AOM_ROOT}/aom_dsp/x86/highbd_intrapred_sse2.c" + "${AOM_ROOT}/aom_dsp/x86/highbd_loopfilter_sse2.c" + "${AOM_ROOT}/aom_dsp/x86/intrapred_sse2.c" + "${AOM_ROOT}/aom_dsp/x86/loopfilter_sse2.c" + "${AOM_ROOT}/aom_dsp/x86/lpf_common_sse2.h" + "${AOM_ROOT}/aom_dsp/x86/mem_sse2.h" + "${AOM_ROOT}/aom_dsp/x86/transpose_sse2.h" + "${AOM_ROOT}/aom_dsp/x86/txfm_common_sse2.h" + "${AOM_ROOT}/aom_dsp/x86/sum_squares_sse2.h") + +list(APPEND AOM_DSP_COMMON_ASM_SSSE3 + "${AOM_ROOT}/aom_dsp/x86/aom_subpixel_8t_ssse3.asm" + "${AOM_ROOT}/aom_dsp/x86/aom_subpixel_bilinear_ssse3.asm") + +list(APPEND AOM_DSP_COMMON_INTRIN_SSSE3 + "${AOM_ROOT}/aom_dsp/x86/aom_subpixel_8t_intrin_ssse3.c" + "${AOM_ROOT}/aom_dsp/x86/highbd_convolve_ssse3.c" + "${AOM_ROOT}/aom_dsp/x86/intrapred_ssse3.c") + +list(APPEND AOM_DSP_COMMON_INTRIN_SSE4_1 + "${AOM_ROOT}/aom_dsp/x86/blend_mask_sse4.h" + "${AOM_ROOT}/aom_dsp/x86/blend_a64_hmask_sse4.c" + "${AOM_ROOT}/aom_dsp/x86/blend_a64_mask_sse4.c" + "${AOM_ROOT}/aom_dsp/x86/blend_a64_vmask_sse4.c") + +list(APPEND AOM_DSP_COMMON_INTRIN_AVX2 + "${AOM_ROOT}/aom_dsp/x86/aom_subpixel_8t_intrin_avx2.c" + "${AOM_ROOT}/aom_dsp/x86/common_avx2.h" + "${AOM_ROOT}/aom_dsp/x86/txfm_common_avx2.h" + "${AOM_ROOT}/aom_dsp/x86/convolve_avx2.h" + "${AOM_ROOT}/aom_dsp/x86/fft_avx2.c" + "${AOM_ROOT}/aom_dsp/x86/highbd_convolve_avx2.c" + "${AOM_ROOT}/aom_dsp/x86/highbd_loopfilter_avx2.c" + "${AOM_ROOT}/aom_dsp/x86/intrapred_avx2.c" + "${AOM_ROOT}/aom_dsp/x86/blend_a64_mask_avx2.c") + +list(APPEND AOM_DSP_COMMON_INTRIN_NEON + "${AOM_ROOT}/aom_dsp/arm/fwd_txfm_neon.c" + "${AOM_ROOT}/aom_dsp/arm/loopfilter_neon.c" + "${AOM_ROOT}/aom_dsp/arm/intrapred_neon.c" + "${AOM_ROOT}/aom_dsp/arm/subtract_neon.c" + "${AOM_ROOT}/aom_dsp/arm/blend_a64_mask_neon.c") + +list(APPEND AOM_DSP_COMMON_INTRIN_DSPR2 + "${AOM_ROOT}/aom_dsp/mips/common_dspr2.c" + "${AOM_ROOT}/aom_dsp/mips/common_dspr2.h" + "${AOM_ROOT}/aom_dsp/mips/convolve2_dspr2.c" + "${AOM_ROOT}/aom_dsp/mips/convolve2_horiz_dspr2.c" + "${AOM_ROOT}/aom_dsp/mips/convolve2_vert_dspr2.c" + "${AOM_ROOT}/aom_dsp/mips/convolve8_dspr2.c" + "${AOM_ROOT}/aom_dsp/mips/convolve8_horiz_dspr2.c" + "${AOM_ROOT}/aom_dsp/mips/convolve8_vert_dspr2.c" + "${AOM_ROOT}/aom_dsp/mips/convolve_common_dspr2.h" + "${AOM_ROOT}/aom_dsp/mips/intrapred16_dspr2.c" + "${AOM_ROOT}/aom_dsp/mips/intrapred4_dspr2.c" + "${AOM_ROOT}/aom_dsp/mips/intrapred8_dspr2.c" + "${AOM_ROOT}/aom_dsp/mips/inv_txfm_dspr2.h") + +list(APPEND AOM_DSP_COMMON_INTRIN_MSA + "${AOM_ROOT}/aom_dsp/mips/aom_convolve8_horiz_msa.c" + "${AOM_ROOT}/aom_dsp/mips/aom_convolve8_vert_msa.c" + "${AOM_ROOT}/aom_dsp/mips/aom_convolve_copy_msa.c" + "${AOM_ROOT}/aom_dsp/mips/aom_convolve_msa.h" + "${AOM_ROOT}/aom_dsp/mips/intrapred_msa.c" + "${AOM_ROOT}/aom_dsp/mips/macros_msa.h") + +if(CONFIG_AV1_DECODER) + list(APPEND AOM_DSP_DECODER_SOURCES + "${AOM_ROOT}/aom_dsp/binary_codes_reader.c" + "${AOM_ROOT}/aom_dsp/binary_codes_reader.h" + "${AOM_ROOT}/aom_dsp/bitreader.h" + "${AOM_ROOT}/aom_dsp/daalaboolreader.c" + "${AOM_ROOT}/aom_dsp/daalaboolreader.h" + "${AOM_ROOT}/aom_dsp/entdec.c" "${AOM_ROOT}/aom_dsp/entdec.h" + "${AOM_ROOT}/aom_dsp/grain_synthesis.c" + "${AOM_ROOT}/aom_dsp/grain_synthesis.h") +endif() + +if(CONFIG_AV1_ENCODER) + list(APPEND AOM_DSP_ENCODER_SOURCES + "${AOM_ROOT}/aom_dsp/binary_codes_writer.c" + "${AOM_ROOT}/aom_dsp/binary_codes_writer.h" + "${AOM_ROOT}/aom_dsp/bitwriter.h" + "${AOM_ROOT}/aom_dsp/daalaboolwriter.c" + "${AOM_ROOT}/aom_dsp/daalaboolwriter.h" + "${AOM_ROOT}/aom_dsp/entenc.c" + "${AOM_ROOT}/aom_dsp/entenc.h" + "${AOM_ROOT}/aom_dsp/fwd_txfm.c" + "${AOM_ROOT}/aom_dsp/grain_table.c" + "${AOM_ROOT}/aom_dsp/grain_table.h" + "${AOM_ROOT}/aom_dsp/noise_model.c" + "${AOM_ROOT}/aom_dsp/noise_model.h" + "${AOM_ROOT}/aom_dsp/noise_util.c" + "${AOM_ROOT}/aom_dsp/noise_util.h" + "${AOM_ROOT}/aom_dsp/psnr.c" + "${AOM_ROOT}/aom_dsp/psnr.h" + "${AOM_ROOT}/aom_dsp/quantize.c" + "${AOM_ROOT}/aom_dsp/quantize.h" + "${AOM_ROOT}/aom_dsp/sad.c" + "${AOM_ROOT}/aom_dsp/sse.c" + "${AOM_ROOT}/aom_dsp/sad_av1.c" + "${AOM_ROOT}/aom_dsp/sum_squares.c" + "${AOM_ROOT}/aom_dsp/variance.c" + "${AOM_ROOT}/aom_dsp/variance.h") + + list(APPEND AOM_DSP_ENCODER_ASM_SSE2 + "${AOM_ROOT}/aom_dsp/x86/highbd_sad4d_sse2.asm" + "${AOM_ROOT}/aom_dsp/x86/highbd_sad_sse2.asm" + "${AOM_ROOT}/aom_dsp/x86/highbd_subpel_variance_impl_sse2.asm" + "${AOM_ROOT}/aom_dsp/x86/highbd_variance_impl_sse2.asm" + "${AOM_ROOT}/aom_dsp/x86/sad4d_sse2.asm" + "${AOM_ROOT}/aom_dsp/x86/sad_sse2.asm" + "${AOM_ROOT}/aom_dsp/x86/subpel_variance_sse2.asm" + "${AOM_ROOT}/aom_dsp/x86/subtract_sse2.asm") + + list(APPEND AOM_DSP_ENCODER_INTRIN_SSE2 + "${AOM_ROOT}/aom_dsp/x86/fwd_txfm_impl_sse2.h" + "${AOM_ROOT}/aom_dsp/x86/fwd_txfm_sse2.c" + "${AOM_ROOT}/aom_dsp/x86/fwd_txfm_sse2.h" + "${AOM_ROOT}/aom_dsp/x86/highbd_quantize_intrin_sse2.c" + "${AOM_ROOT}/aom_dsp/x86/highbd_subtract_sse2.c" + "${AOM_ROOT}/aom_dsp/x86/highbd_variance_sse2.c" + "${AOM_ROOT}/aom_dsp/x86/quantize_sse2.c" + "${AOM_ROOT}/aom_dsp/x86/quantize_x86.h" + "${AOM_ROOT}/aom_dsp/x86/sum_squares_sse2.c" + "${AOM_ROOT}/aom_dsp/x86/variance_sse2.c") + + list(APPEND AOM_DSP_ENCODER_ASM_SSSE3_X86_64 + "${AOM_ROOT}/aom_dsp/x86/fwd_txfm_ssse3_x86_64.asm" + "${AOM_ROOT}/aom_dsp/x86/ssim_opt_x86_64.asm") + + list(APPEND AOM_DSP_ENCODER_INTRIN_AVX2 + "${AOM_ROOT}/aom_dsp/x86/masked_sad_intrin_avx2.c" + "${AOM_ROOT}/aom_dsp/x86/subtract_avx2.c" + "${AOM_ROOT}/aom_dsp/x86/highbd_quantize_intrin_avx2.c" + "${AOM_ROOT}/aom_dsp/x86/sad4d_avx2.c" + "${AOM_ROOT}/aom_dsp/x86/sad_avx2.c" + "${AOM_ROOT}/aom_dsp/x86/sad_highbd_avx2.c" + "${AOM_ROOT}/aom_dsp/x86/sad_impl_avx2.c" + "${AOM_ROOT}/aom_dsp/x86/variance_avx2.c" + "${AOM_ROOT}/aom_dsp/x86/highbd_variance_avx2.c" + "${AOM_ROOT}/aom_dsp/x86/sse_avx2.c" + "${AOM_ROOT}/aom_dsp/x86/variance_impl_avx2.c" + "${AOM_ROOT}/aom_dsp/x86/obmc_sad_avx2.c" + "${AOM_ROOT}/aom_dsp/x86/obmc_variance_avx2.c" + "${AOM_ROOT}/aom_dsp/x86/sum_squares_avx2.c") + + list(APPEND AOM_DSP_ENCODER_ASM_SSSE3_X86_64 + "${AOM_ROOT}/aom_dsp/x86/quantize_ssse3_x86_64.asm") + + list(APPEND AOM_DSP_ENCODER_AVX_ASM_X86_64 + "${AOM_ROOT}/aom_dsp/x86/quantize_avx_x86_64.asm") + + list(APPEND AOM_DSP_ENCODER_INTRIN_SSSE3 + "${AOM_ROOT}/aom_dsp/x86/masked_sad_intrin_ssse3.h" + "${AOM_ROOT}/aom_dsp/x86/masked_sad_intrin_ssse3.c" + "${AOM_ROOT}/aom_dsp/x86/masked_variance_intrin_ssse3.h" + "${AOM_ROOT}/aom_dsp/x86/masked_variance_intrin_ssse3.c" + "${AOM_ROOT}/aom_dsp/x86/variance_impl_ssse3.c" + "${AOM_ROOT}/aom_dsp/x86/jnt_variance_ssse3.c" + "${AOM_ROOT}/aom_dsp/x86/jnt_sad_ssse3.c") + + list(APPEND AOM_DSP_ENCODER_INTRIN_SSE4_1 + "${AOM_ROOT}/aom_dsp/x86/highbd_variance_sse4.c" + "${AOM_ROOT}/aom_dsp/x86/sse_sse4.c" + "${AOM_ROOT}/aom_dsp/x86/obmc_sad_sse4.c" + "${AOM_ROOT}/aom_dsp/x86/obmc_variance_sse4.c") + + list(APPEND AOM_DSP_ENCODER_INTRIN_NEON + "${AOM_ROOT}/aom_dsp/arm/sad4d_neon.c" + "${AOM_ROOT}/aom_dsp/arm/sad_neon.c" + "${AOM_ROOT}/aom_dsp/arm/subpel_variance_neon.c" + "${AOM_ROOT}/aom_dsp/arm/variance_neon.c") + + list(APPEND AOM_DSP_ENCODER_INTRIN_MSA "${AOM_ROOT}/aom_dsp/mips/sad_msa.c" + "${AOM_ROOT}/aom_dsp/mips/subtract_msa.c" + "${AOM_ROOT}/aom_dsp/mips/variance_msa.c" + "${AOM_ROOT}/aom_dsp/mips/sub_pixel_variance_msa.c") + + if(CONFIG_INTERNAL_STATS) + list(APPEND AOM_DSP_ENCODER_SOURCES "${AOM_ROOT}/aom_dsp/fastssim.c" + "${AOM_ROOT}/aom_dsp/psnrhvs.c" "${AOM_ROOT}/aom_dsp/ssim.c" + "${AOM_ROOT}/aom_dsp/ssim.h") + endif() +endif() + +# Creates aom_dsp build targets. Must not be called until after libaom target +# has been created. +function(setup_aom_dsp_targets) + add_library(aom_dsp_common OBJECT ${AOM_DSP_COMMON_SOURCES}) + list(APPEND AOM_LIB_TARGETS aom_dsp_common) + create_dummy_source_file("aom_av1" "c" "dummy_source_file") + add_library(aom_dsp OBJECT "${dummy_source_file}") + target_sources(aom PRIVATE $) + list(APPEND AOM_LIB_TARGETS aom_dsp) + + # Not all generators support libraries consisting only of object files. Add a + # dummy source file to the aom_dsp target. + add_dummy_source_file_to_target("aom_dsp" "c") + + if(CONFIG_AV1_DECODER) + add_library(aom_dsp_decoder OBJECT ${AOM_DSP_DECODER_SOURCES}) + list(APPEND AOM_LIB_TARGETS aom_dsp_decoder) + target_sources(aom PRIVATE $) + endif() + + if(CONFIG_AV1_ENCODER) + add_library(aom_dsp_encoder OBJECT ${AOM_DSP_ENCODER_SOURCES}) + list(APPEND AOM_LIB_TARGETS aom_dsp_encoder) + target_sources(aom PRIVATE $) + endif() + + if(HAVE_SSE2) + add_asm_library("aom_dsp_common_sse2" "AOM_DSP_COMMON_ASM_SSE2" "aom") + add_intrinsics_object_library("-msse2" "sse2" "aom_dsp_common" + "AOM_DSP_COMMON_INTRIN_SSE2" "aom") + + if(CONFIG_AV1_ENCODER) + add_asm_library("aom_dsp_encoder_sse2" "AOM_DSP_ENCODER_ASM_SSE2" "aom") + add_intrinsics_object_library("-msse2" "sse2" "aom_dsp_encoder" + "AOM_DSP_ENCODER_INTRIN_SSE2" "aom") + endif() + endif() + + if(HAVE_SSSE3) + add_asm_library("aom_dsp_common_ssse3" "AOM_DSP_COMMON_ASM_SSSE3" "aom") + add_intrinsics_object_library("-mssse3" "ssse3" "aom_dsp_common" + "AOM_DSP_COMMON_INTRIN_SSSE3" "aom") + + if(CONFIG_AV1_ENCODER) + if("${AOM_TARGET_CPU}" STREQUAL "x86_64") + list(APPEND AOM_DSP_ENCODER_ASM_SSSE3 + ${AOM_DSP_ENCODER_ASM_SSSE3_X86_64}) + endif() + add_asm_library("aom_dsp_encoder_ssse3" "AOM_DSP_ENCODER_ASM_SSSE3" "aom") + add_intrinsics_object_library("-mssse3" "ssse3" "aom_dsp_encoder" + "AOM_DSP_ENCODER_INTRIN_SSSE3" "aom") + endif() + endif() + + if(HAVE_SSE4_1) + add_intrinsics_object_library("-msse4.1" "sse4_1" "aom_dsp_common" + "AOM_DSP_COMMON_INTRIN_SSE4_1" "aom") + if(CONFIG_AV1_ENCODER) + add_intrinsics_object_library("-msse4.1" "sse4_1" "aom_dsp_encoder" + "AOM_DSP_ENCODER_INTRIN_SSE4_1" "aom") + endif() + endif() + + if(HAVE_AVX AND "${AOM_TARGET_CPU}" STREQUAL "x86_64") + if(CONFIG_AV1_ENCODER) + add_asm_library("aom_dsp_encoder_avx" "AOM_DSP_ENCODER_AVX_ASM_X86_64" + "aom") + endif() + endif() + + if(HAVE_AVX2) + add_intrinsics_object_library("-mavx2" "avx2" "aom_dsp_common" + "AOM_DSP_COMMON_INTRIN_AVX2" "aom") + if(CONFIG_AV1_ENCODER) + add_intrinsics_object_library("-mavx2" "avx2" "aom_dsp_encoder" + "AOM_DSP_ENCODER_INTRIN_AVX2" "aom") + endif() + endif() + + if(HAVE_NEON) + add_intrinsics_object_library("${AOM_NEON_INTRIN_FLAG}" "neon" + "aom_dsp_common" "AOM_DSP_COMMON_INTRIN_NEON" + "aom") + if(CONFIG_AV1_ENCODER) + add_intrinsics_object_library("${AOM_NEON_INTRIN_FLAG}" "neon" + "aom_dsp_encoder" + "AOM_DSP_ENCODER_INTRIN_NEON" "aom") + endif() + endif() + + if(HAVE_DSPR2) + add_intrinsics_object_library("" "dspr2" "aom_dsp_common" + "AOM_DSP_COMMON_INTRIN_DSPR2" "aom") + endif() + + if(HAVE_MSA) + add_intrinsics_object_library("" "msa" "aom_dsp_common" + "AOM_DSP_COMMON_INTRIN_MSA" "aom") + if(CONFIG_AV1_ENCODER) + add_intrinsics_object_library("" "msa" "aom_dsp_encoder" + "AOM_DSP_ENCODER_INTRIN_MSA" "aom") + endif() + endif() + + # Pass the new lib targets up to the parent scope instance of + # $AOM_LIB_TARGETS. + set(AOM_LIB_TARGETS ${AOM_LIB_TARGETS} PARENT_SCOPE) +endfunction() diff --git a/third_party/aom/aom_dsp/aom_dsp_common.h b/third_party/aom/aom_dsp/aom_dsp_common.h new file mode 100644 index 0000000000..a185b23c8e --- /dev/null +++ b/third_party/aom/aom_dsp/aom_dsp_common.h @@ -0,0 +1,98 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_AOM_DSP_COMMON_H_ +#define AOM_AOM_DSP_AOM_DSP_COMMON_H_ + +#include "config/aom_config.h" + +#include "aom/aom_integer.h" +#include "aom_ports/mem.h" + +#ifdef __cplusplus +extern "C" { +#endif + +#ifndef MAX_SB_SIZE +#define MAX_SB_SIZE 128 +#endif // ndef MAX_SB_SIZE + +#define AOMMIN(x, y) (((x) < (y)) ? (x) : (y)) +#define AOMMAX(x, y) (((x) > (y)) ? (x) : (y)) + +#define IMPLIES(a, b) (!(a) || (b)) // Logical 'a implies b' (or 'a -> b') + +#define IS_POWER_OF_TWO(x) (((x) & ((x)-1)) == 0) + +/* Left shifting a negative value became undefined behavior in C99 (downgraded + from merely implementation-defined in C89). This should still compile to the + correct thing on any two's-complement machine, but avoid ubsan warnings.*/ +#define AOM_SIGNED_SHL(x, shift) ((x) * (((x)*0 + 1) << (shift))) + +// These can be used to give a hint about branch outcomes. +// This can have an effect, even if your target processor has a +// good branch predictor, as these hints can affect basic block +// ordering by the compiler. +#ifdef __GNUC__ +#define LIKELY(v) __builtin_expect(v, 1) +#define UNLIKELY(v) __builtin_expect(v, 0) +#else +#define LIKELY(v) (v) +#define UNLIKELY(v) (v) +#endif + +typedef uint8_t qm_val_t; +#define AOM_QM_BITS 5 + +// Note: +// tran_low_t is the datatype used for final transform coefficients. +// tran_high_t is the datatype used for intermediate transform stages. +typedef int64_t tran_high_t; +typedef int32_t tran_low_t; + +static INLINE uint8_t clip_pixel(int val) { + return (val > 255) ? 255 : (val < 0) ? 0 : val; +} + +static INLINE int clamp(int value, int low, int high) { + return value < low ? low : (value > high ? high : value); +} + +static INLINE int64_t clamp64(int64_t value, int64_t low, int64_t high) { + return value < low ? low : (value > high ? high : value); +} + +static INLINE double fclamp(double value, double low, double high) { + return value < low ? low : (value > high ? high : value); +} + +static INLINE uint16_t clip_pixel_highbd(int val, int bd) { + switch (bd) { + case 8: + default: return (uint16_t)clamp(val, 0, 255); + case 10: return (uint16_t)clamp(val, 0, 1023); + case 12: return (uint16_t)clamp(val, 0, 4095); + } +} + +// The result of this branchless code is equivalent to (value < 0 ? 0 : value) +// or max(0, value) and might be faster in some cases. +// Care should be taken since the behavior of right shifting signed type +// negative value is undefined by C standards and implementation defined, +static INLINE unsigned int negative_to_zero(int value) { + return value & ~(value >> (sizeof(value) * 8 - 1)); +} + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif // AOM_AOM_DSP_AOM_DSP_COMMON_H_ diff --git a/third_party/aom/aom_dsp/aom_dsp_rtcd.c b/third_party/aom/aom_dsp/aom_dsp_rtcd.c new file mode 100644 index 0000000000..1514bd64e6 --- /dev/null +++ b/third_party/aom/aom_dsp/aom_dsp_rtcd.c @@ -0,0 +1,18 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ +#include "config/aom_config.h" + +#define RTCD_C +#include "config/aom_dsp_rtcd.h" + +#include "aom_ports/aom_once.h" + +void aom_dsp_rtcd() { aom_once(setup_rtcd_internal); } diff --git a/third_party/aom/aom_dsp/aom_dsp_rtcd_defs.pl b/third_party/aom/aom_dsp/aom_dsp_rtcd_defs.pl new file mode 100755 index 0000000000..3c3326cdc3 --- /dev/null +++ b/third_party/aom/aom_dsp/aom_dsp_rtcd_defs.pl @@ -0,0 +1,1575 @@ +## +## Copyright (c) 2017, Alliance for Open Media. All rights reserved +## +## This source code is subject to the terms of the BSD 2 Clause License and +## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License +## was not distributed with this source code in the LICENSE file, you can +## obtain it at www.aomedia.org/license/software. If the Alliance for Open +## Media Patent License 1.0 was not distributed with this source code in the +## PATENTS file, you can obtain it at www.aomedia.org/license/patent. +## +sub aom_dsp_forward_decls() { +print <=4 && $h >=4 && ($w == 2*$h || $h == 2*$w)); + push @tx_sizes, [$w, $h] if ($w >=4 && $h >=4 && ($w == 4*$h || $h == 4*$w)); + } +} + +@pred_names = qw/dc dc_top dc_left dc_128 v h paeth smooth smooth_v smooth_h/; + +# +# Intra prediction +# + +foreach (@tx_sizes) { + ($w, $h) = @$_; + foreach $pred_name (@pred_names) { + add_proto "void", "aom_${pred_name}_predictor_${w}x${h}", + "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"; + add_proto "void", "aom_highbd_${pred_name}_predictor_${w}x${h}", + "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd"; + } +} + +specialize qw/aom_dc_top_predictor_4x4 msa neon sse2/; +specialize qw/aom_dc_top_predictor_4x8 sse2/; +specialize qw/aom_dc_top_predictor_4x16 sse2/; +specialize qw/aom_dc_top_predictor_8x4 sse2/; +specialize qw/aom_dc_top_predictor_8x8 neon msa sse2/; +specialize qw/aom_dc_top_predictor_8x16 sse2/; +specialize qw/aom_dc_top_predictor_8x32 sse2/; +specialize qw/aom_dc_top_predictor_16x4 sse2/; +specialize qw/aom_dc_top_predictor_16x8 sse2/; +specialize qw/aom_dc_top_predictor_16x16 neon msa sse2/; +specialize qw/aom_dc_top_predictor_16x32 sse2/; +specialize qw/aom_dc_top_predictor_16x64 sse2/; +specialize qw/aom_dc_top_predictor_32x8 sse2/; +specialize qw/aom_dc_top_predictor_32x16 sse2 avx2/; +specialize qw/aom_dc_top_predictor_32x32 msa neon sse2 avx2/; +specialize qw/aom_dc_top_predictor_32x64 sse2 avx2/; +specialize qw/aom_dc_top_predictor_64x64 sse2 avx2/; +specialize qw/aom_dc_top_predictor_64x32 sse2 avx2/; +specialize qw/aom_dc_top_predictor_64x16 sse2 avx2/; +specialize qw/aom_dc_left_predictor_4x4 msa neon sse2/; +specialize qw/aom_dc_left_predictor_4x8 sse2/; +specialize qw/aom_dc_left_predictor_4x16 sse2/; +specialize qw/aom_dc_left_predictor_8x4 sse2/; +specialize qw/aom_dc_left_predictor_8x8 neon msa sse2/; +specialize qw/aom_dc_left_predictor_8x16 sse2/; +specialize qw/aom_dc_left_predictor_8x32 sse2/; +specialize qw/aom_dc_left_predictor_16x4 sse2/; +specialize qw/aom_dc_left_predictor_16x8 sse2/; +specialize qw/aom_dc_left_predictor_16x16 neon msa sse2/; +specialize qw/aom_dc_left_predictor_16x32 sse2/; +specialize qw/aom_dc_left_predictor_16x64 sse2/; +specialize qw/aom_dc_left_predictor_32x8 sse2/; +specialize qw/aom_dc_left_predictor_32x16 sse2 avx2/; +specialize qw/aom_dc_left_predictor_32x32 msa neon sse2 avx2/; +specialize qw/aom_dc_left_predictor_32x64 sse2 avx2/; +specialize qw/aom_dc_left_predictor_64x64 sse2 avx2/; +specialize qw/aom_dc_left_predictor_64x32 sse2 avx2/; +specialize qw/aom_dc_left_predictor_64x16 sse2 avx2/; +specialize qw/aom_dc_128_predictor_4x4 msa neon sse2/; +specialize qw/aom_dc_128_predictor_4x8 sse2/; +specialize qw/aom_dc_128_predictor_4x16 sse2/; +specialize qw/aom_dc_128_predictor_8x4 sse2/; +specialize qw/aom_dc_128_predictor_8x8 neon msa sse2/; +specialize qw/aom_dc_128_predictor_8x16 sse2/; +specialize qw/aom_dc_128_predictor_8x32 sse2/; +specialize qw/aom_dc_128_predictor_16x4 sse2/; +specialize qw/aom_dc_128_predictor_16x8 sse2/; +specialize qw/aom_dc_128_predictor_16x16 neon msa sse2/; +specialize qw/aom_dc_128_predictor_16x32 sse2/; +specialize qw/aom_dc_128_predictor_16x64 sse2/; +specialize qw/aom_dc_128_predictor_32x8 sse2/; +specialize qw/aom_dc_128_predictor_32x16 sse2 avx2/; +specialize qw/aom_dc_128_predictor_32x32 msa neon sse2 avx2/; +specialize qw/aom_dc_128_predictor_32x64 sse2 avx2/; +specialize qw/aom_dc_128_predictor_64x64 sse2 avx2/; +specialize qw/aom_dc_128_predictor_64x32 sse2 avx2/; +specialize qw/aom_dc_128_predictor_64x16 sse2 avx2/; +specialize qw/aom_v_predictor_4x4 neon msa sse2/; +specialize qw/aom_v_predictor_4x8 sse2/; +specialize qw/aom_v_predictor_4x16 sse2/; +specialize qw/aom_v_predictor_8x4 sse2/; +specialize qw/aom_v_predictor_8x8 neon msa sse2/; +specialize qw/aom_v_predictor_8x16 sse2/; +specialize qw/aom_v_predictor_8x32 sse2/; +specialize qw/aom_v_predictor_16x4 sse2/; +specialize qw/aom_v_predictor_16x8 sse2/; +specialize qw/aom_v_predictor_16x16 neon msa sse2/; +specialize qw/aom_v_predictor_16x32 sse2/; +specialize qw/aom_v_predictor_16x64 sse2/; +specialize qw/aom_v_predictor_32x8 sse2/; +specialize qw/aom_v_predictor_32x16 sse2 avx2/; +specialize qw/aom_v_predictor_32x32 neon msa sse2 avx2/; +specialize qw/aom_v_predictor_32x64 sse2 avx2/; +specialize qw/aom_v_predictor_64x64 sse2 avx2/; +specialize qw/aom_v_predictor_64x32 sse2 avx2/; +specialize qw/aom_v_predictor_64x16 sse2 avx2/; +specialize qw/aom_h_predictor_4x8 sse2/; +specialize qw/aom_h_predictor_4x16 sse2/; +specialize qw/aom_h_predictor_4x4 neon dspr2 msa sse2/; +specialize qw/aom_h_predictor_8x4 sse2/; +specialize qw/aom_h_predictor_8x8 neon dspr2 msa sse2/; +specialize qw/aom_h_predictor_8x16 sse2/; +specialize qw/aom_h_predictor_8x32 sse2/; +specialize qw/aom_h_predictor_16x4 sse2/; +specialize qw/aom_h_predictor_16x8 sse2/; +specialize qw/aom_h_predictor_16x16 neon dspr2 msa sse2/; +specialize qw/aom_h_predictor_16x32 sse2/; +specialize qw/aom_h_predictor_16x64 sse2/; +specialize qw/aom_h_predictor_32x8 sse2/; +specialize qw/aom_h_predictor_32x16 sse2/; +specialize qw/aom_h_predictor_32x32 neon msa sse2 avx2/; +specialize qw/aom_h_predictor_32x64 sse2/; +specialize qw/aom_h_predictor_64x64 sse2/; +specialize qw/aom_h_predictor_64x32 sse2/; +specialize qw/aom_h_predictor_64x16 sse2/; +specialize qw/aom_paeth_predictor_4x4 ssse3/; +specialize qw/aom_paeth_predictor_4x8 ssse3/; +specialize qw/aom_paeth_predictor_4x16 ssse3/; +specialize qw/aom_paeth_predictor_8x4 ssse3/; +specialize qw/aom_paeth_predictor_8x8 ssse3/; +specialize qw/aom_paeth_predictor_8x16 ssse3/; +specialize qw/aom_paeth_predictor_8x32 ssse3/; +specialize qw/aom_paeth_predictor_16x4 ssse3/; +specialize qw/aom_paeth_predictor_16x8 ssse3 avx2/; +specialize qw/aom_paeth_predictor_16x16 ssse3 avx2/; +specialize qw/aom_paeth_predictor_16x32 ssse3 avx2/; +specialize qw/aom_paeth_predictor_16x64 ssse3 avx2/; +specialize qw/aom_paeth_predictor_32x8 ssse3/; +specialize qw/aom_paeth_predictor_32x16 ssse3 avx2/; +specialize qw/aom_paeth_predictor_32x32 ssse3 avx2/; +specialize qw/aom_paeth_predictor_32x64 ssse3 avx2/; +specialize qw/aom_paeth_predictor_64x32 ssse3 avx2/; +specialize qw/aom_paeth_predictor_64x64 ssse3 avx2/; +specialize qw/aom_paeth_predictor_64x16 ssse3 avx2/; +specialize qw/aom_paeth_predictor_16x8 ssse3/; +specialize qw/aom_paeth_predictor_16x16 ssse3/; +specialize qw/aom_paeth_predictor_16x32 ssse3/; +specialize qw/aom_paeth_predictor_32x16 ssse3/; +specialize qw/aom_paeth_predictor_32x32 ssse3/; +specialize qw/aom_smooth_predictor_4x4 ssse3/; +specialize qw/aom_smooth_predictor_4x8 ssse3/; +specialize qw/aom_smooth_predictor_4x16 ssse3/; +specialize qw/aom_smooth_predictor_8x4 ssse3/; +specialize qw/aom_smooth_predictor_8x8 ssse3/; +specialize qw/aom_smooth_predictor_8x16 ssse3/; +specialize qw/aom_smooth_predictor_8x32 ssse3/; +specialize qw/aom_smooth_predictor_16x4 ssse3/; +specialize qw/aom_smooth_predictor_16x8 ssse3/; +specialize qw/aom_smooth_predictor_16x16 ssse3/; +specialize qw/aom_smooth_predictor_16x32 ssse3/; +specialize qw/aom_smooth_predictor_16x64 ssse3/; +specialize qw/aom_smooth_predictor_32x8 ssse3/; +specialize qw/aom_smooth_predictor_32x16 ssse3/; +specialize qw/aom_smooth_predictor_32x32 ssse3/; +specialize qw/aom_smooth_predictor_32x64 ssse3/; +specialize qw/aom_smooth_predictor_64x64 ssse3/; +specialize qw/aom_smooth_predictor_64x32 ssse3/; +specialize qw/aom_smooth_predictor_64x16 ssse3/; + +specialize qw/aom_smooth_v_predictor_4x4 ssse3/; +specialize qw/aom_smooth_v_predictor_4x8 ssse3/; +specialize qw/aom_smooth_v_predictor_4x16 ssse3/; +specialize qw/aom_smooth_v_predictor_8x4 ssse3/; +specialize qw/aom_smooth_v_predictor_8x8 ssse3/; +specialize qw/aom_smooth_v_predictor_8x16 ssse3/; +specialize qw/aom_smooth_v_predictor_8x32 ssse3/; +specialize qw/aom_smooth_v_predictor_16x4 ssse3/; +specialize qw/aom_smooth_v_predictor_16x8 ssse3/; +specialize qw/aom_smooth_v_predictor_16x16 ssse3/; +specialize qw/aom_smooth_v_predictor_16x32 ssse3/; +specialize qw/aom_smooth_v_predictor_16x64 ssse3/; +specialize qw/aom_smooth_v_predictor_32x8 ssse3/; +specialize qw/aom_smooth_v_predictor_32x16 ssse3/; +specialize qw/aom_smooth_v_predictor_32x32 ssse3/; +specialize qw/aom_smooth_v_predictor_32x64 ssse3/; +specialize qw/aom_smooth_v_predictor_64x64 ssse3/; +specialize qw/aom_smooth_v_predictor_64x32 ssse3/; +specialize qw/aom_smooth_v_predictor_64x16 ssse3/; + +specialize qw/aom_smooth_h_predictor_4x4 ssse3/; +specialize qw/aom_smooth_h_predictor_4x8 ssse3/; +specialize qw/aom_smooth_h_predictor_4x16 ssse3/; +specialize qw/aom_smooth_h_predictor_8x4 ssse3/; +specialize qw/aom_smooth_h_predictor_8x8 ssse3/; +specialize qw/aom_smooth_h_predictor_8x16 ssse3/; +specialize qw/aom_smooth_h_predictor_8x32 ssse3/; +specialize qw/aom_smooth_h_predictor_16x4 ssse3/; +specialize qw/aom_smooth_h_predictor_16x8 ssse3/; +specialize qw/aom_smooth_h_predictor_16x16 ssse3/; +specialize qw/aom_smooth_h_predictor_16x32 ssse3/; +specialize qw/aom_smooth_h_predictor_16x64 ssse3/; +specialize qw/aom_smooth_h_predictor_32x8 ssse3/; +specialize qw/aom_smooth_h_predictor_32x16 ssse3/; +specialize qw/aom_smooth_h_predictor_32x32 ssse3/; +specialize qw/aom_smooth_h_predictor_32x64 ssse3/; +specialize qw/aom_smooth_h_predictor_64x64 ssse3/; +specialize qw/aom_smooth_h_predictor_64x32 ssse3/; +specialize qw/aom_smooth_h_predictor_64x16 ssse3/; + +# TODO(yunqingwang): optimize rectangular DC_PRED to replace division +# by multiply and shift. +specialize qw/aom_dc_predictor_4x4 dspr2 msa neon sse2/; +specialize qw/aom_dc_predictor_4x8 sse2/; +specialize qw/aom_dc_predictor_4x16 sse2/; +specialize qw/aom_dc_predictor_8x4 sse2/; +specialize qw/aom_dc_predictor_8x8 dspr2 neon msa sse2/; +specialize qw/aom_dc_predictor_8x16 sse2/; +specialize qw/aom_dc_predictor_8x32 sse2/; +specialize qw/aom_dc_predictor_16x4 sse2/; +specialize qw/aom_dc_predictor_16x8 sse2/; +specialize qw/aom_dc_predictor_16x16 dspr2 neon msa sse2/; +specialize qw/aom_dc_predictor_16x32 sse2/; +specialize qw/aom_dc_predictor_16x64 sse2/; +specialize qw/aom_dc_predictor_32x8 sse2/; +specialize qw/aom_dc_predictor_32x16 sse2 avx2/; +specialize qw/aom_dc_predictor_32x32 msa neon sse2 avx2/; +specialize qw/aom_dc_predictor_32x64 sse2 avx2/; +specialize qw/aom_dc_predictor_64x64 sse2 avx2/; +specialize qw/aom_dc_predictor_64x32 sse2 avx2/; +specialize qw/aom_dc_predictor_64x16 sse2 avx2/; + + specialize qw/aom_highbd_v_predictor_4x4 sse2/; + specialize qw/aom_highbd_v_predictor_4x8 sse2/; + specialize qw/aom_highbd_v_predictor_8x4 sse2/; + specialize qw/aom_highbd_v_predictor_8x8 sse2/; + specialize qw/aom_highbd_v_predictor_8x16 sse2/; + specialize qw/aom_highbd_v_predictor_16x8 sse2/; + specialize qw/aom_highbd_v_predictor_16x16 sse2/; + specialize qw/aom_highbd_v_predictor_16x32 sse2/; + specialize qw/aom_highbd_v_predictor_32x16 sse2/; + specialize qw/aom_highbd_v_predictor_32x32 sse2/; + + # TODO(yunqingwang): optimize rectangular DC_PRED to replace division + # by multiply and shift. + specialize qw/aom_highbd_dc_predictor_4x4 sse2 neon/; + specialize qw/aom_highbd_dc_predictor_4x8 sse2/; + specialize qw/aom_highbd_dc_predictor_8x4 sse2/;; + specialize qw/aom_highbd_dc_predictor_8x8 sse2 neon/;; + specialize qw/aom_highbd_dc_predictor_8x16 sse2/;; + specialize qw/aom_highbd_dc_predictor_16x8 sse2/; + specialize qw/aom_highbd_dc_predictor_16x16 sse2 neon/; + specialize qw/aom_highbd_dc_predictor_16x32 sse2/; + specialize qw/aom_highbd_dc_predictor_32x16 sse2/; + specialize qw/aom_highbd_dc_predictor_32x32 sse2 neon/; + specialize qw/aom_highbd_dc_predictor_64x64 neon/; + + specialize qw/aom_highbd_h_predictor_4x4 sse2/; + specialize qw/aom_highbd_h_predictor_4x8 sse2/; + specialize qw/aom_highbd_h_predictor_8x4 sse2/; + specialize qw/aom_highbd_h_predictor_8x8 sse2/; + specialize qw/aom_highbd_h_predictor_8x16 sse2/; + specialize qw/aom_highbd_h_predictor_16x8 sse2/; + specialize qw/aom_highbd_h_predictor_16x16 sse2/; + specialize qw/aom_highbd_h_predictor_16x32 sse2/; + specialize qw/aom_highbd_h_predictor_32x16 sse2/; + specialize qw/aom_highbd_h_predictor_32x32 sse2/; + specialize qw/aom_highbd_dc_left_predictor_4x4 sse2/; + specialize qw/aom_highbd_dc_top_predictor_4x4 sse2/; + specialize qw/aom_highbd_dc_128_predictor_4x4 sse2/; + specialize qw/aom_highbd_dc_left_predictor_4x8 sse2/; + specialize qw/aom_highbd_dc_top_predictor_4x8 sse2/; + specialize qw/aom_highbd_dc_128_predictor_4x8 sse2/; + specialize qw/aom_highbd_dc_left_predictor_8x4 sse2/; + specialize qw/aom_highbd_dc_top_predictor_8x4 sse2/; + specialize qw/aom_highbd_dc_128_predictor_8x4 sse2/; + specialize qw/aom_highbd_dc_left_predictor_8x8 sse2/; + specialize qw/aom_highbd_dc_top_predictor_8x8 sse2/; + specialize qw/aom_highbd_dc_128_predictor_8x8 sse2/; + specialize qw/aom_highbd_dc_left_predictor_8x16 sse2/; + specialize qw/aom_highbd_dc_top_predictor_8x16 sse2/; + specialize qw/aom_highbd_dc_128_predictor_8x16 sse2/; + specialize qw/aom_highbd_dc_left_predictor_16x8 sse2/; + specialize qw/aom_highbd_dc_top_predictor_16x8 sse2/; + specialize qw/aom_highbd_dc_128_predictor_16x8 sse2/; + specialize qw/aom_highbd_dc_left_predictor_16x16 sse2/; + specialize qw/aom_highbd_dc_top_predictor_16x16 sse2/; + specialize qw/aom_highbd_dc_128_predictor_16x16 sse2/; + specialize qw/aom_highbd_dc_left_predictor_16x32 sse2/; + specialize qw/aom_highbd_dc_top_predictor_16x32 sse2/; + specialize qw/aom_highbd_dc_128_predictor_16x32 sse2/; + specialize qw/aom_highbd_dc_left_predictor_32x16 sse2/; + specialize qw/aom_highbd_dc_top_predictor_32x16 sse2/; + specialize qw/aom_highbd_dc_128_predictor_32x16 sse2/; + specialize qw/aom_highbd_dc_left_predictor_32x32 sse2/; + specialize qw/aom_highbd_dc_top_predictor_32x32 sse2/; + specialize qw/aom_highbd_dc_128_predictor_32x32 sse2/; + +# +# Sub Pixel Filters +# +add_proto qw/void aom_convolve_copy/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h"; +add_proto qw/void aom_convolve8_horiz/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h"; +add_proto qw/void aom_convolve8_vert/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h"; + +specialize qw/aom_convolve_copy sse2 /; +specialize qw/aom_convolve8_horiz sse2 ssse3/, "$avx2_ssse3"; +specialize qw/aom_convolve8_vert sse2 ssse3/, "$avx2_ssse3"; + +add_proto qw/void aom_highbd_convolve_copy/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps"; +specialize qw/aom_highbd_convolve_copy sse2 avx2/; + +add_proto qw/void aom_highbd_convolve8_horiz/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps"; +specialize qw/aom_highbd_convolve8_horiz avx2/, "$sse2_x86_64"; + +add_proto qw/void aom_highbd_convolve8_vert/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps"; +specialize qw/aom_highbd_convolve8_vert avx2/, "$sse2_x86_64"; + +# +# Loopfilter +# +add_proto qw/void aom_lpf_vertical_14/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh"; +specialize qw/aom_lpf_vertical_14 sse2 neon/; + +add_proto qw/void aom_lpf_vertical_14_dual/, "uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1"; +specialize qw/aom_lpf_vertical_14_dual sse2/; + +add_proto qw/void aom_lpf_vertical_6/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh"; +specialize qw/aom_lpf_vertical_6 sse2 neon/; + +add_proto qw/void aom_lpf_vertical_8/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh"; +specialize qw/aom_lpf_vertical_8 sse2 neon/; + +add_proto qw/void aom_lpf_vertical_8_dual/, "uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1"; +specialize qw/aom_lpf_vertical_8_dual sse2/; + +add_proto qw/void aom_lpf_vertical_4/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh"; +specialize qw/aom_lpf_vertical_4 sse2 neon/; + +add_proto qw/void aom_lpf_vertical_4_dual/, "uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1"; +specialize qw/aom_lpf_vertical_4_dual sse2/; + +add_proto qw/void aom_lpf_horizontal_14/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh"; +specialize qw/aom_lpf_horizontal_14 sse2 neon/; + +add_proto qw/void aom_lpf_horizontal_14_dual/, "uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1"; +specialize qw/aom_lpf_horizontal_14_dual sse2/; + +add_proto qw/void aom_lpf_horizontal_6/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh"; +specialize qw/aom_lpf_horizontal_6 sse2 neon/; + +add_proto qw/void aom_lpf_horizontal_6_dual/, "uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1"; +specialize qw/aom_lpf_horizontal_6_dual sse2/; + +add_proto qw/void aom_lpf_horizontal_8/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh"; +specialize qw/aom_lpf_horizontal_8 sse2 neon/; + +add_proto qw/void aom_lpf_horizontal_8_dual/, "uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1"; +specialize qw/aom_lpf_horizontal_8_dual sse2/; + +add_proto qw/void aom_lpf_horizontal_4/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh"; +specialize qw/aom_lpf_horizontal_4 sse2 neon/; + +add_proto qw/void aom_lpf_horizontal_4_dual/, "uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1"; +specialize qw/aom_lpf_horizontal_4_dual sse2/; + +add_proto qw/void aom_highbd_lpf_vertical_14/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd"; +specialize qw/aom_highbd_lpf_vertical_14 sse2/; + +add_proto qw/void aom_highbd_lpf_vertical_14_dual/, "uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd"; +specialize qw/aom_highbd_lpf_vertical_14_dual sse2 avx2/; + +add_proto qw/void aom_highbd_lpf_vertical_8/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd"; +specialize qw/aom_highbd_lpf_vertical_8 sse2/; + +add_proto qw/void aom_highbd_lpf_vertical_6/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd"; +specialize qw/aom_highbd_lpf_vertical_6 sse2/; + +add_proto qw/void aom_lpf_vertical_6_dual/, "uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1"; +specialize qw/aom_lpf_vertical_6_dual sse2/; + +add_proto qw/void aom_highbd_lpf_vertical_6_dual/, "uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd"; +specialize qw/aom_highbd_lpf_vertical_6_dual sse2/; + +add_proto qw/void aom_highbd_lpf_vertical_8_dual/, "uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd"; +specialize qw/aom_highbd_lpf_vertical_8_dual sse2 avx2/; + +add_proto qw/void aom_highbd_lpf_vertical_4/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd"; +specialize qw/aom_highbd_lpf_vertical_4 sse2/; + +add_proto qw/void aom_highbd_lpf_vertical_4_dual/, "uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd"; +specialize qw/aom_highbd_lpf_vertical_4_dual sse2 avx2/; + +add_proto qw/void aom_highbd_lpf_horizontal_14/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd"; +specialize qw/aom_highbd_lpf_horizontal_14 sse2/; + +add_proto qw/void aom_highbd_lpf_horizontal_14_dual/, "uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limt1, const uint8_t *thresh1,int bd"; +specialize qw/aom_highbd_lpf_horizontal_14_dual sse2 avx2/; + +add_proto qw/void aom_highbd_lpf_horizontal_6/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd"; +specialize qw/aom_highbd_lpf_horizontal_6 sse2/; + +add_proto qw/void aom_highbd_lpf_horizontal_6_dual/, "uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd"; +specialize qw/aom_highbd_lpf_horizontal_6_dual sse2/; + +add_proto qw/void aom_highbd_lpf_horizontal_8/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd"; +specialize qw/aom_highbd_lpf_horizontal_8 sse2/; + +add_proto qw/void aom_highbd_lpf_horizontal_8_dual/, "uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd"; +specialize qw/aom_highbd_lpf_horizontal_8_dual sse2 avx2/; + +add_proto qw/void aom_highbd_lpf_horizontal_4/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd"; +specialize qw/aom_highbd_lpf_horizontal_4 sse2/; + +add_proto qw/void aom_highbd_lpf_horizontal_4_dual/, "uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd"; +specialize qw/aom_highbd_lpf_horizontal_4_dual sse2 avx2/; + +# Helper functions. +add_proto qw/void av1_round_shift_array/, "int32_t *arr, int size, int bit"; +specialize "av1_round_shift_array", qw/sse4_1 neon/; + +# +# Encoder functions. +# + +# +# Forward transform +# +if (aom_config("CONFIG_AV1_ENCODER") eq "yes"){ + add_proto qw/void aom_fdct8x8/, "const int16_t *input, tran_low_t *output, int stride"; + specialize qw/aom_fdct8x8 sse2/, "$ssse3_x86_64"; + + # High bit depth + add_proto qw/void aom_highbd_fdct8x8/, "const int16_t *input, tran_low_t *output, int stride"; + specialize qw/aom_highbd_fdct8x8 sse2/; + + # FFT/IFFT (float) only used for denoising (and noise power spectral density estimation) + add_proto qw/void aom_fft2x2_float/, "const float *input, float *temp, float *output"; + + add_proto qw/void aom_fft4x4_float/, "const float *input, float *temp, float *output"; + specialize qw/aom_fft4x4_float sse2/; + + add_proto qw/void aom_fft8x8_float/, "const float *input, float *temp, float *output"; + specialize qw/aom_fft8x8_float avx2 sse2/; + + add_proto qw/void aom_fft16x16_float/, "const float *input, float *temp, float *output"; + specialize qw/aom_fft16x16_float avx2 sse2/; + + add_proto qw/void aom_fft32x32_float/, "const float *input, float *temp, float *output"; + specialize qw/aom_fft32x32_float avx2 sse2/; + + add_proto qw/void aom_ifft2x2_float/, "const float *input, float *temp, float *output"; + + add_proto qw/void aom_ifft4x4_float/, "const float *input, float *temp, float *output"; + specialize qw/aom_ifft4x4_float sse2/; + + add_proto qw/void aom_ifft8x8_float/, "const float *input, float *temp, float *output"; + specialize qw/aom_ifft8x8_float avx2 sse2/; + + add_proto qw/void aom_ifft16x16_float/, "const float *input, float *temp, float *output"; + specialize qw/aom_ifft16x16_float avx2 sse2/; + + add_proto qw/void aom_ifft32x32_float/, "const float *input, float *temp, float *output"; + specialize qw/aom_ifft32x32_float avx2 sse2/; +} # CONFIG_AV1_ENCODER + +# +# Quantization +# +if (aom_config("CONFIG_AV1_ENCODER") eq "yes") { + add_proto qw/void aom_quantize_b/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan"; + specialize qw/aom_quantize_b sse2/, "$ssse3_x86_64", "$avx_x86_64"; + + add_proto qw/void aom_quantize_b_32x32/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan"; + specialize qw/aom_quantize_b_32x32/, "$ssse3_x86_64", "$avx_x86_64"; + + add_proto qw/void aom_quantize_b_64x64/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan"; +} # CONFIG_AV1_ENCODER + +if (aom_config("CONFIG_AV1_ENCODER") eq "yes") { + add_proto qw/void aom_highbd_quantize_b/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan"; + specialize qw/aom_highbd_quantize_b sse2 avx2/; + + add_proto qw/void aom_highbd_quantize_b_32x32/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan"; + specialize qw/aom_highbd_quantize_b_32x32 sse2/; + + add_proto qw/void aom_highbd_quantize_b_64x64/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan"; + +} # CONFIG_AV1_ENCODER + +# +# Alpha blending with mask +# +add_proto qw/void aom_lowbd_blend_a64_d16_mask/, "uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subx, int suby, ConvolveParams *conv_params"; +specialize qw/aom_lowbd_blend_a64_d16_mask sse4_1 avx2 neon/; +add_proto qw/void aom_highbd_blend_a64_d16_mask/, "uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subx, int suby, ConvolveParams *conv_params, const int bd"; +add_proto qw/void aom_blend_a64_mask/, "uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subx, int suby"; +add_proto qw/void aom_blend_a64_hmask/, "uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h"; +add_proto qw/void aom_blend_a64_vmask/, "uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h"; +specialize "aom_blend_a64_mask", qw/sse4_1 avx2/; +specialize "aom_blend_a64_hmask", qw/sse4_1/; +specialize "aom_blend_a64_vmask", qw/sse4_1/; + +add_proto qw/void aom_highbd_blend_a64_mask/, "uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subx, int suby, int bd"; +add_proto qw/void aom_highbd_blend_a64_hmask/, "uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h, int bd"; +add_proto qw/void aom_highbd_blend_a64_vmask/, "uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h, int bd"; +specialize "aom_highbd_blend_a64_mask", qw/sse4_1/; +specialize "aom_highbd_blend_a64_hmask", qw/sse4_1/; +specialize "aom_highbd_blend_a64_vmask", qw/sse4_1/; + +if (aom_config("CONFIG_AV1_ENCODER") eq "yes") { + # + # Block subtraction + # + add_proto qw/void aom_subtract_block/, "int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride"; + specialize qw/aom_subtract_block neon msa sse2 avx2/; + + add_proto qw/void aom_highbd_subtract_block/, "int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride, int bd"; + specialize qw/aom_highbd_subtract_block sse2/; + + add_proto qw/int64_t/, "aom_sse", "const uint8_t *a, int a_stride, const uint8_t *b,int b_stride, int width, int height"; + specialize qw/aom_sse sse4_1 avx2/; + + add_proto qw/int64_t/, "aom_highbd_sse", "const uint8_t *a8, int a_stride, const uint8_t *b8,int b_stride, int width, int height"; + specialize qw/aom_highbd_sse sse4_1 avx2/; + + if (aom_config("CONFIG_AV1_ENCODER") eq "yes") { + # + # Sum of Squares + # + add_proto qw/uint64_t aom_sum_squares_2d_i16/, "const int16_t *src, int stride, int width, int height"; + specialize qw/aom_sum_squares_2d_i16 sse2 avx2/; + + add_proto qw/uint64_t aom_sum_squares_i16/, "const int16_t *src, uint32_t N"; + specialize qw/aom_sum_squares_i16 sse2/; + + } + + + # + # Single block SAD / Single block Avg SAD + # + foreach (@block_sizes) { + ($w, $h) = @$_; + add_proto qw/unsigned int/, "aom_sad${w}x${h}", "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride"; + add_proto qw/unsigned int/, "aom_sad${w}x${h}_avg", "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred"; + add_proto qw/unsigned int/, "aom_jnt_sad${w}x${h}_avg", "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const JNT_COMP_PARAMS *jcp_param"; + } + + specialize qw/aom_sad128x128 avx2 sse2/; + specialize qw/aom_sad128x64 avx2 sse2/; + specialize qw/aom_sad64x128 avx2 sse2/; + specialize qw/aom_sad64x64 avx2 neon msa sse2/; + specialize qw/aom_sad64x32 avx2 msa sse2/; + specialize qw/aom_sad32x64 avx2 msa sse2/; + specialize qw/aom_sad32x32 avx2 neon msa sse2/; + specialize qw/aom_sad32x16 avx2 msa sse2/; + specialize qw/aom_sad16x32 msa sse2/; + specialize qw/aom_sad16x16 neon msa sse2/; + specialize qw/aom_sad16x8 neon msa sse2/; + specialize qw/aom_sad8x16 neon msa sse2/; + specialize qw/aom_sad8x8 neon msa sse2/; + specialize qw/aom_sad8x4 msa sse2/; + specialize qw/aom_sad4x8 msa sse2/; + specialize qw/aom_sad4x4 neon msa sse2/; + + specialize qw/aom_sad128x128_avg avx2 sse2/; + specialize qw/aom_sad128x64_avg avx2 sse2/; + specialize qw/aom_sad64x128_avg avx2 sse2/; + specialize qw/aom_sad64x64_avg avx2 msa sse2/; + specialize qw/aom_sad64x32_avg avx2 msa sse2/; + specialize qw/aom_sad32x64_avg avx2 msa sse2/; + specialize qw/aom_sad32x32_avg avx2 msa sse2/; + specialize qw/aom_sad32x16_avg avx2 msa sse2/; + specialize qw/aom_sad16x32_avg msa sse2/; + specialize qw/aom_sad16x16_avg msa sse2/; + specialize qw/aom_sad16x8_avg msa sse2/; + specialize qw/aom_sad8x16_avg msa sse2/; + specialize qw/aom_sad8x8_avg msa sse2/; + specialize qw/aom_sad8x4_avg msa sse2/; + specialize qw/aom_sad4x8_avg msa sse2/; + specialize qw/aom_sad4x4_avg msa sse2/; + + specialize qw/aom_sad4x16 sse2/; + specialize qw/aom_sad16x4 sse2/; + specialize qw/aom_sad8x32 sse2/; + specialize qw/aom_sad32x8 sse2/; + specialize qw/aom_sad16x64 sse2/; + specialize qw/aom_sad64x16 sse2/; + + specialize qw/aom_sad4x16_avg sse2/; + specialize qw/aom_sad16x4_avg sse2/; + specialize qw/aom_sad8x32_avg sse2/; + specialize qw/aom_sad32x8_avg sse2/; + specialize qw/aom_sad16x64_avg sse2/; + specialize qw/aom_sad64x16_avg sse2/; + + specialize qw/aom_jnt_sad128x128_avg ssse3/; + specialize qw/aom_jnt_sad128x64_avg ssse3/; + specialize qw/aom_jnt_sad64x128_avg ssse3/; + specialize qw/aom_jnt_sad64x64_avg ssse3/; + specialize qw/aom_jnt_sad64x32_avg ssse3/; + specialize qw/aom_jnt_sad32x64_avg ssse3/; + specialize qw/aom_jnt_sad32x32_avg ssse3/; + specialize qw/aom_jnt_sad32x16_avg ssse3/; + specialize qw/aom_jnt_sad16x32_avg ssse3/; + specialize qw/aom_jnt_sad16x16_avg ssse3/; + specialize qw/aom_jnt_sad16x8_avg ssse3/; + specialize qw/aom_jnt_sad8x16_avg ssse3/; + specialize qw/aom_jnt_sad8x8_avg ssse3/; + specialize qw/aom_jnt_sad8x4_avg ssse3/; + specialize qw/aom_jnt_sad4x8_avg ssse3/; + specialize qw/aom_jnt_sad4x4_avg ssse3/; + + specialize qw/aom_jnt_sad4x16_avg ssse3/; + specialize qw/aom_jnt_sad16x4_avg ssse3/; + specialize qw/aom_jnt_sad8x32_avg ssse3/; + specialize qw/aom_jnt_sad32x8_avg ssse3/; + specialize qw/aom_jnt_sad16x64_avg ssse3/; + specialize qw/aom_jnt_sad64x16_avg ssse3/; + + add_proto qw/unsigned int/, "aom_sad4xh", "const uint8_t *a, int a_stride, const uint8_t *b, int b_stride, int width, int height"; + add_proto qw/unsigned int/, "aom_sad8xh", "const uint8_t *a, int a_stride, const uint8_t *b, int b_stride, int width, int height"; + add_proto qw/unsigned int/, "aom_sad16xh", "const uint8_t *a, int a_stride, const uint8_t *b, int b_stride, int width, int height"; + add_proto qw/unsigned int/, "aom_sad32xh", "const uint8_t *a, int a_stride, const uint8_t *b, int b_stride, int width, int height"; + add_proto qw/unsigned int/, "aom_sad64xh", "const uint8_t *a, int a_stride, const uint8_t *b, int b_stride, int width, int height"; + add_proto qw/unsigned int/, "aom_sad128xh", "const uint8_t *a, int a_stride, const uint8_t *b, int b_stride, int width, int height"; + + specialize qw/aom_sad4xh sse2/; + specialize qw/aom_sad8xh sse2/; + specialize qw/aom_sad16xh sse2/; + specialize qw/aom_sad32xh sse2/; + specialize qw/aom_sad64xh sse2/; + specialize qw/aom_sad128xh sse2/; + + + foreach (@block_sizes) { + ($w, $h) = @$_; + add_proto qw/unsigned int/, "aom_highbd_sad${w}x${h}", "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride"; + add_proto qw/unsigned int/, "aom_highbd_sad${w}x${h}_avg", "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred"; + if ($w != 128 && $h != 128 && $w != 4) { + specialize "aom_highbd_sad${w}x${h}", qw/sse2/; + specialize "aom_highbd_sad${w}x${h}_avg", qw/sse2/; + } + add_proto qw/unsigned int/, "aom_highbd_jnt_sad${w}x${h}_avg", "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const JNT_COMP_PARAMS* jcp_param"; + } + specialize qw/aom_highbd_sad128x128 avx2/; + specialize qw/aom_highbd_sad128x64 avx2/; + specialize qw/aom_highbd_sad64x128 avx2/; + specialize qw/aom_highbd_sad64x64 avx2 sse2/; + specialize qw/aom_highbd_sad64x32 avx2 sse2/; + specialize qw/aom_highbd_sad32x64 avx2 sse2/; + specialize qw/aom_highbd_sad32x32 avx2 sse2/; + specialize qw/aom_highbd_sad32x16 avx2 sse2/; + specialize qw/aom_highbd_sad16x32 avx2 sse2/; + specialize qw/aom_highbd_sad16x16 avx2 sse2/; + specialize qw/aom_highbd_sad16x8 avx2 sse2/; + specialize qw/aom_highbd_sad8x4 sse2/; + + specialize qw/aom_highbd_sad128x128_avg avx2/; + specialize qw/aom_highbd_sad128x64_avg avx2/; + specialize qw/aom_highbd_sad64x128_avg avx2/; + specialize qw/aom_highbd_sad64x64_avg avx2 sse2/; + specialize qw/aom_highbd_sad64x32_avg avx2 sse2/; + specialize qw/aom_highbd_sad32x64_avg avx2 sse2/; + specialize qw/aom_highbd_sad32x32_avg avx2 sse2/; + specialize qw/aom_highbd_sad32x16_avg avx2 sse2/; + specialize qw/aom_highbd_sad16x32_avg avx2 sse2/; + specialize qw/aom_highbd_sad16x16_avg avx2 sse2/; + specialize qw/aom_highbd_sad16x8_avg avx2 sse2/; + specialize qw/aom_highbd_sad8x4_avg sse2/; + + specialize qw/aom_highbd_sad16x4 sse2/; + specialize qw/aom_highbd_sad8x32 sse2/; + specialize qw/aom_highbd_sad32x8 sse2/; + specialize qw/aom_highbd_sad16x64 sse2/; + specialize qw/aom_highbd_sad64x16 sse2/; + + specialize qw/aom_highbd_sad16x4_avg sse2/; + specialize qw/aom_highbd_sad8x32_avg sse2/; + specialize qw/aom_highbd_sad32x8_avg sse2/; + specialize qw/aom_highbd_sad16x64_avg sse2/; + specialize qw/aom_highbd_sad64x16_avg sse2/; + + # + # Masked SAD + # + foreach (@block_sizes) { + ($w, $h) = @$_; + add_proto qw/unsigned int/, "aom_masked_sad${w}x${h}", "const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask"; + specialize "aom_masked_sad${w}x${h}", qw/ssse3 avx2/; + } + + + foreach (@block_sizes) { + ($w, $h) = @$_; + add_proto qw/unsigned int/, "aom_highbd_masked_sad${w}x${h}", "const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask"; + specialize "aom_highbd_masked_sad${w}x${h}", qw/ssse3 avx2/; + } + + + # + # OBMC SAD + # + foreach (@block_sizes) { + ($w, $h) = @$_; + add_proto qw/unsigned int/, "aom_obmc_sad${w}x${h}", "const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask"; + if (! (($w == 128 && $h == 32) || ($w == 32 && $h == 128))) { + specialize "aom_obmc_sad${w}x${h}", qw/sse4_1 avx2/; + } + } + + + foreach (@block_sizes) { + ($w, $h) = @$_; + add_proto qw/unsigned int/, "aom_highbd_obmc_sad${w}x${h}", "const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask"; + if (! (($w == 128 && $h == 32) || ($w == 32 && $h == 128))) { + specialize "aom_highbd_obmc_sad${w}x${h}", qw/sse4_1 avx2/; + } + } + + + # + # Multi-block SAD, comparing a reference to N independent blocks + # + foreach (@block_sizes) { + ($w, $h) = @$_; + add_proto qw/void/, "aom_sad${w}x${h}x4d", "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array"; + } + + specialize qw/aom_sad128x128x4d avx2 sse2/; + specialize qw/aom_sad128x64x4d avx2 sse2/; + specialize qw/aom_sad64x128x4d avx2 sse2/; + specialize qw/aom_sad64x64x4d avx2 neon msa sse2/; + specialize qw/aom_sad64x32x4d avx2 msa sse2/; + specialize qw/aom_sad32x64x4d avx2 msa sse2/; + specialize qw/aom_sad32x32x4d avx2 neon msa sse2/; + specialize qw/aom_sad32x16x4d msa sse2/; + specialize qw/aom_sad16x32x4d msa sse2/; + specialize qw/aom_sad16x16x4d neon msa sse2/; + specialize qw/aom_sad16x8x4d msa sse2/; + specialize qw/aom_sad8x16x4d msa sse2/; + specialize qw/aom_sad8x8x4d msa sse2/; + specialize qw/aom_sad8x4x4d msa sse2/; + specialize qw/aom_sad4x8x4d msa sse2/; + specialize qw/aom_sad4x4x4d msa sse2/; + + specialize qw/aom_sad4x16x4d sse2/; + specialize qw/aom_sad16x4x4d sse2/; + specialize qw/aom_sad8x32x4d sse2/; + specialize qw/aom_sad32x8x4d sse2/; + specialize qw/aom_sad16x64x4d sse2/; + specialize qw/aom_sad64x16x4d sse2/; + + # + # Multi-block SAD, comparing a reference to N independent blocks + # + foreach (@block_sizes) { + ($w, $h) = @$_; + add_proto qw/void/, "aom_highbd_sad${w}x${h}x4d", "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array"; + if ($w != 128 && $h != 128) { + specialize "aom_highbd_sad${w}x${h}x4d", qw/sse2/; + } + } + specialize qw/aom_highbd_sad128x128x4d avx2/; + specialize qw/aom_highbd_sad128x64x4d avx2/; + specialize qw/aom_highbd_sad64x128x4d avx2/; + specialize qw/aom_highbd_sad64x64x4d sse2 avx2/; + specialize qw/aom_highbd_sad64x32x4d sse2 avx2/; + specialize qw/aom_highbd_sad32x64x4d sse2 avx2/; + specialize qw/aom_highbd_sad32x32x4d sse2 avx2/; + specialize qw/aom_highbd_sad32x16x4d sse2 avx2/; + specialize qw/aom_highbd_sad16x32x4d sse2 avx2/; + specialize qw/aom_highbd_sad16x16x4d sse2 avx2/; + specialize qw/aom_highbd_sad16x8x4d sse2 avx2/; + specialize qw/aom_highbd_sad8x16x4d sse2/; + specialize qw/aom_highbd_sad8x8x4d sse2/; + specialize qw/aom_highbd_sad8x4x4d sse2/; + specialize qw/aom_highbd_sad4x8x4d sse2/; + specialize qw/aom_highbd_sad4x4x4d sse2/; + + specialize qw/aom_highbd_sad4x16x4d sse2/; + specialize qw/aom_highbd_sad16x4x4d sse2/; + specialize qw/aom_highbd_sad8x32x4d sse2/; + specialize qw/aom_highbd_sad32x8x4d sse2/; + specialize qw/aom_highbd_sad16x64x4d sse2/; + specialize qw/aom_highbd_sad64x16x4d sse2/; + + # + # Structured Similarity (SSIM) + # + if (aom_config("CONFIG_INTERNAL_STATS") eq "yes") { + add_proto qw/void aom_ssim_parms_8x8/, "const uint8_t *s, int sp, const uint8_t *r, int rp, uint32_t *sum_s, uint32_t *sum_r, uint32_t *sum_sq_s, uint32_t *sum_sq_r, uint32_t *sum_sxr"; + specialize qw/aom_ssim_parms_8x8/, "$sse2_x86_64"; + + add_proto qw/void aom_ssim_parms_16x16/, "const uint8_t *s, int sp, const uint8_t *r, int rp, uint32_t *sum_s, uint32_t *sum_r, uint32_t *sum_sq_s, uint32_t *sum_sq_r, uint32_t *sum_sxr"; + specialize qw/aom_ssim_parms_16x16/, "$sse2_x86_64"; + + add_proto qw/void aom_highbd_ssim_parms_8x8/, "const uint16_t *s, int sp, const uint16_t *r, int rp, uint32_t *sum_s, uint32_t *sum_r, uint32_t *sum_sq_s, uint32_t *sum_sq_r, uint32_t *sum_sxr"; + + } +} # CONFIG_AV1_ENCODER + +if (aom_config("CONFIG_AV1_ENCODER") eq "yes") { + + # + # Specialty Variance + # + add_proto qw/void aom_get16x16var/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum"; + + add_proto qw/void aom_get8x8var/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum"; + + specialize qw/aom_get16x16var neon msa/; + specialize qw/aom_get8x8var neon msa/; + + + add_proto qw/unsigned int aom_mse16x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int recon_stride, unsigned int *sse"; + add_proto qw/unsigned int aom_mse16x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int recon_stride, unsigned int *sse"; + add_proto qw/unsigned int aom_mse8x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int recon_stride, unsigned int *sse"; + add_proto qw/unsigned int aom_mse8x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int recon_stride, unsigned int *sse"; + + specialize qw/aom_mse16x16 sse2 avx2 neon msa/; + specialize qw/aom_mse16x8 sse2 msa/; + specialize qw/aom_mse8x16 sse2 msa/; + specialize qw/aom_mse8x8 sse2 msa/; + + foreach $bd (8, 10, 12) { + add_proto qw/void/, "aom_highbd_${bd}_get16x16var", "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum"; + add_proto qw/void/, "aom_highbd_${bd}_get8x8var", "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum"; + + add_proto qw/unsigned int/, "aom_highbd_${bd}_mse16x16", "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int recon_stride, unsigned int *sse"; + add_proto qw/unsigned int/, "aom_highbd_${bd}_mse16x8", "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int recon_stride, unsigned int *sse"; + add_proto qw/unsigned int/, "aom_highbd_${bd}_mse8x16", "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int recon_stride, unsigned int *sse"; + add_proto qw/unsigned int/, "aom_highbd_${bd}_mse8x8", "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int recon_stride, unsigned int *sse"; + + specialize "aom_highbd_${bd}_mse16x16", qw/sse2/; + specialize "aom_highbd_${bd}_mse8x8", qw/sse2/; + } + + + # + # + # + add_proto qw/void aom_upsampled_pred/, "MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col, + const MV *const mv, uint8_t *comp_pred, int width, int height, int subpel_x_q3, + int subpel_y_q3, const uint8_t *ref, int ref_stride, int subpel_search"; + specialize qw/aom_upsampled_pred sse2/; + + add_proto qw/void aom_comp_avg_upsampled_pred/, "MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col, + const MV *const mv, uint8_t *comp_pred, const uint8_t *pred, int width, + int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref, + int ref_stride, int subpel_search"; + specialize qw/aom_comp_avg_upsampled_pred sse2/; + + add_proto qw/void aom_jnt_comp_avg_upsampled_pred/, "MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col, + const MV *const mv, uint8_t *comp_pred, const uint8_t *pred, int width, + int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref, + int ref_stride, const JNT_COMP_PARAMS *jcp_param, int subpel_search"; + specialize qw/aom_jnt_comp_avg_upsampled_pred ssse3/; + + add_proto qw/void aom_comp_mask_upsampled_pred/, "MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col, + const MV *const mv, uint8_t *comp_pred, const uint8_t *pred, int width, + int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref, + int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask, + int subpel_search"; + specialize qw/aom_comp_mask_upsampled_pred sse2/; + + + add_proto qw/void aom_highbd_upsampled_pred/, "MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col, + const MV *const mv, uint8_t *comp_pred8, int width, int height, int subpel_x_q3, + int subpel_y_q3, const uint8_t *ref8, int ref_stride, int bd, int subpel_search"; + specialize qw/aom_highbd_upsampled_pred sse2/; + + add_proto qw/void aom_highbd_comp_avg_upsampled_pred/, "MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col, + const MV *const mv, uint8_t *comp_pred8, const uint8_t *pred8, int width, + int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref8, int ref_stride, int bd, int subpel_search"; + specialize qw/aom_highbd_comp_avg_upsampled_pred sse2/; + + add_proto qw/void aom_highbd_jnt_comp_avg_upsampled_pred/, "MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col, + const MV *const mv, uint8_t *comp_pred8, const uint8_t *pred8, int width, + int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref8, + int ref_stride, int bd, const JNT_COMP_PARAMS *jcp_param, int subpel_search"; + specialize qw/aom_highbd_jnt_comp_avg_upsampled_pred sse2/; + + + # + # + # + add_proto qw/unsigned int aom_get_mb_ss/, "const int16_t *"; + add_proto qw/unsigned int aom_get4x4sse_cs/, "const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int ref_stride"; + + specialize qw/aom_get_mb_ss sse2 msa/; + specialize qw/aom_get4x4sse_cs neon msa/; + + # + # Variance / Subpixel Variance / Subpixel Avg Variance + # + add_proto qw/unsigned int/, "aom_variance2x2", "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"; + + add_proto qw/unsigned int/, "aom_variance2x4", "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"; + + add_proto qw/unsigned int/, "aom_variance4x2", "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"; + + foreach (@block_sizes) { + ($w, $h) = @$_; + add_proto qw/unsigned int/, "aom_variance${w}x${h}", "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"; + add_proto qw/uint32_t/, "aom_sub_pixel_variance${w}x${h}", "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse"; + add_proto qw/uint32_t/, "aom_sub_pixel_avg_variance${w}x${h}", "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred"; + add_proto qw/uint32_t/, "aom_jnt_sub_pixel_avg_variance${w}x${h}", "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const JNT_COMP_PARAMS *jcp_param"; + } + specialize qw/aom_variance128x128 sse2 avx2 /; + specialize qw/aom_variance128x64 sse2 avx2 /; + specialize qw/aom_variance64x128 sse2 avx2 /; + specialize qw/aom_variance64x64 sse2 avx2 neon msa/; + specialize qw/aom_variance64x32 sse2 avx2 neon msa/; + specialize qw/aom_variance32x64 sse2 avx2 neon msa/; + specialize qw/aom_variance32x32 sse2 avx2 neon msa/; + specialize qw/aom_variance32x16 sse2 avx2 msa/; + specialize qw/aom_variance16x32 sse2 avx2 msa/; + specialize qw/aom_variance16x16 sse2 avx2 neon msa/; + specialize qw/aom_variance16x8 sse2 avx2 neon msa/; + specialize qw/aom_variance8x16 sse2 neon msa/; + specialize qw/aom_variance8x8 sse2 neon msa/; + specialize qw/aom_variance8x4 sse2 msa/; + specialize qw/aom_variance4x8 sse2 msa/; + specialize qw/aom_variance4x4 sse2 msa/; + + specialize qw/aom_sub_pixel_variance128x128 avx2 sse2 ssse3/; + specialize qw/aom_sub_pixel_variance128x64 avx2 sse2 ssse3/; + specialize qw/aom_sub_pixel_variance64x128 avx2 sse2 ssse3/; + specialize qw/aom_sub_pixel_variance64x64 avx2 neon msa sse2 ssse3/; + specialize qw/aom_sub_pixel_variance64x32 avx2 msa sse2 ssse3/; + specialize qw/aom_sub_pixel_variance32x64 avx2 msa sse2 ssse3/; + specialize qw/aom_sub_pixel_variance32x32 avx2 neon msa sse2 ssse3/; + specialize qw/aom_sub_pixel_variance32x16 avx2 msa sse2 ssse3/; + specialize qw/aom_sub_pixel_variance16x32 msa sse2 ssse3/; + specialize qw/aom_sub_pixel_variance16x16 neon msa sse2 ssse3/; + specialize qw/aom_sub_pixel_variance16x8 msa sse2 ssse3/; + specialize qw/aom_sub_pixel_variance8x16 msa sse2 ssse3/; + specialize qw/aom_sub_pixel_variance8x8 neon msa sse2 ssse3/; + specialize qw/aom_sub_pixel_variance8x4 msa sse2 ssse3/; + specialize qw/aom_sub_pixel_variance4x8 msa sse2 ssse3/; + specialize qw/aom_sub_pixel_variance4x4 msa sse2 ssse3/; + + specialize qw/aom_sub_pixel_avg_variance128x128 avx2 sse2 ssse3/; + specialize qw/aom_sub_pixel_avg_variance128x64 avx2 sse2 ssse3/; + specialize qw/aom_sub_pixel_avg_variance64x128 avx2 sse2 ssse3/; + specialize qw/aom_sub_pixel_avg_variance64x64 avx2 msa sse2 ssse3/; + specialize qw/aom_sub_pixel_avg_variance64x32 avx2 msa sse2 ssse3/; + specialize qw/aom_sub_pixel_avg_variance32x64 avx2 msa sse2 ssse3/; + specialize qw/aom_sub_pixel_avg_variance32x32 avx2 msa sse2 ssse3/; + specialize qw/aom_sub_pixel_avg_variance32x16 avx2 msa sse2 ssse3/; + specialize qw/aom_sub_pixel_avg_variance16x32 msa sse2 ssse3/; + specialize qw/aom_sub_pixel_avg_variance16x16 msa sse2 ssse3/; + specialize qw/aom_sub_pixel_avg_variance16x8 msa sse2 ssse3/; + specialize qw/aom_sub_pixel_avg_variance8x16 msa sse2 ssse3/; + specialize qw/aom_sub_pixel_avg_variance8x8 msa sse2 ssse3/; + specialize qw/aom_sub_pixel_avg_variance8x4 msa sse2 ssse3/; + specialize qw/aom_sub_pixel_avg_variance4x8 msa sse2 ssse3/; + specialize qw/aom_sub_pixel_avg_variance4x4 msa sse2 ssse3/; + + specialize qw/aom_variance4x16 sse2/; + specialize qw/aom_variance16x4 sse2 avx2/; + specialize qw/aom_variance8x32 sse2/; + specialize qw/aom_variance32x8 sse2 avx2/; + specialize qw/aom_variance16x64 sse2 avx2/; + specialize qw/aom_variance64x16 sse2 avx2/; + specialize qw/aom_sub_pixel_variance4x16 sse2 ssse3/; + specialize qw/aom_sub_pixel_variance16x4 sse2 ssse3/; + specialize qw/aom_sub_pixel_variance8x32 sse2 ssse3/; + specialize qw/aom_sub_pixel_variance32x8 sse2 ssse3/; + specialize qw/aom_sub_pixel_variance16x64 sse2 ssse3/; + specialize qw/aom_sub_pixel_variance64x16 sse2 ssse3/; + specialize qw/aom_sub_pixel_avg_variance4x16 sse2 ssse3/; + specialize qw/aom_sub_pixel_avg_variance16x4 sse2 ssse3/; + specialize qw/aom_sub_pixel_avg_variance8x32 sse2 ssse3/; + specialize qw/aom_sub_pixel_avg_variance32x8 sse2 ssse3/; + specialize qw/aom_sub_pixel_avg_variance16x64 sse2 ssse3/; + specialize qw/aom_sub_pixel_avg_variance64x16 sse2 ssse3/; + + specialize qw/aom_jnt_sub_pixel_avg_variance64x64 ssse3/; + specialize qw/aom_jnt_sub_pixel_avg_variance64x32 ssse3/; + specialize qw/aom_jnt_sub_pixel_avg_variance32x64 ssse3/; + specialize qw/aom_jnt_sub_pixel_avg_variance32x32 ssse3/; + specialize qw/aom_jnt_sub_pixel_avg_variance32x16 ssse3/; + specialize qw/aom_jnt_sub_pixel_avg_variance16x32 ssse3/; + specialize qw/aom_jnt_sub_pixel_avg_variance16x16 ssse3/; + specialize qw/aom_jnt_sub_pixel_avg_variance16x8 ssse3/; + specialize qw/aom_jnt_sub_pixel_avg_variance8x16 ssse3/; + specialize qw/aom_jnt_sub_pixel_avg_variance8x8 ssse3/; + specialize qw/aom_jnt_sub_pixel_avg_variance8x4 ssse3/; + specialize qw/aom_jnt_sub_pixel_avg_variance4x8 ssse3/; + specialize qw/aom_jnt_sub_pixel_avg_variance4x4 ssse3/; + + specialize qw/aom_jnt_sub_pixel_avg_variance4x16 ssse3/; + specialize qw/aom_jnt_sub_pixel_avg_variance16x4 ssse3/; + specialize qw/aom_jnt_sub_pixel_avg_variance8x32 ssse3/; + specialize qw/aom_jnt_sub_pixel_avg_variance32x8 ssse3/; + specialize qw/aom_jnt_sub_pixel_avg_variance16x64 ssse3/; + specialize qw/aom_jnt_sub_pixel_avg_variance64x16 ssse3/; + + specialize qw/aom_jnt_sub_pixel_avg_variance128x128 ssse3/; + specialize qw/aom_jnt_sub_pixel_avg_variance128x64 ssse3/; + specialize qw/aom_jnt_sub_pixel_avg_variance64x128 ssse3/; + + + foreach $bd (8, 10, 12) { + add_proto qw/unsigned int/, "aom_highbd_${bd}_variance2x2", "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"; + + add_proto qw/unsigned int/, "aom_highbd_${bd}_variance2x4", "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"; + + add_proto qw/unsigned int/, "aom_highbd_${bd}_variance4x2", "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"; + + foreach (@block_sizes) { + ($w, $h) = @$_; + add_proto qw/unsigned int/, "aom_highbd_${bd}_variance${w}x${h}", "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse"; + add_proto qw/uint32_t/, "aom_highbd_${bd}_sub_pixel_variance${w}x${h}", "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse"; + add_proto qw/uint32_t/, "aom_highbd_${bd}_sub_pixel_avg_variance${w}x${h}", "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred"; + if ($w != 128 && $h != 128 && $w != 4 && $h != 4) { + specialize "aom_highbd_${bd}_variance${w}x${h}", "sse2"; + } + # TODO(david.barker): When ext-partition-types is enabled, we currently + # don't have vectorized 4x16 highbd variance functions + if ($w == 4 && $h == 4) { + specialize "aom_highbd_${bd}_variance${w}x${h}", "sse4_1"; + } + if ($w != 128 && $h != 128 && $w != 4) { + specialize "aom_highbd_${bd}_sub_pixel_variance${w}x${h}", qw/sse2/; + specialize "aom_highbd_${bd}_sub_pixel_avg_variance${w}x${h}", qw/sse2/; + } + if ($w == 4 && $h == 4) { + specialize "aom_highbd_${bd}_sub_pixel_variance${w}x${h}", "sse4_1"; + specialize "aom_highbd_${bd}_sub_pixel_avg_variance${w}x${h}", "sse4_1"; + } + + add_proto qw/uint32_t/, "aom_highbd_${bd}_jnt_sub_pixel_avg_variance${w}x${h}", "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const JNT_COMP_PARAMS* jcp_param"; + } + } + + # + # Masked Variance / Masked Subpixel Variance + # + foreach (@block_sizes) { + ($w, $h) = @$_; + add_proto qw/unsigned int/, "aom_masked_sub_pixel_variance${w}x${h}", "const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse"; + specialize "aom_masked_sub_pixel_variance${w}x${h}", qw/ssse3/; + } + + + foreach $bd ("_8_", "_10_", "_12_") { + foreach (@block_sizes) { + ($w, $h) = @$_; + add_proto qw/unsigned int/, "aom_highbd${bd}masked_sub_pixel_variance${w}x${h}", "const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse"; + specialize "aom_highbd${bd}masked_sub_pixel_variance${w}x${h}", qw/ssse3/; + } + } + + + # + # OBMC Variance / OBMC Subpixel Variance + # + foreach (@block_sizes) { + ($w, $h) = @$_; + add_proto qw/unsigned int/, "aom_obmc_variance${w}x${h}", "const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse"; + add_proto qw/unsigned int/, "aom_obmc_sub_pixel_variance${w}x${h}", "const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse"; + specialize "aom_obmc_variance${w}x${h}", qw/sse4_1 avx2/; + specialize "aom_obmc_sub_pixel_variance${w}x${h}", q/sse4_1/; + } + + + foreach $bd ("_", "_10_", "_12_") { + foreach (@block_sizes) { + ($w, $h) = @$_; + add_proto qw/unsigned int/, "aom_highbd${bd}obmc_variance${w}x${h}", "const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse"; + add_proto qw/unsigned int/, "aom_highbd${bd}obmc_sub_pixel_variance${w}x${h}", "const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse"; + specialize "aom_highbd${bd}obmc_variance${w}x${h}", qw/sse4_1/; + } + } + + + add_proto qw/uint32_t aom_sub_pixel_avg_variance64x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred"; + specialize qw/aom_sub_pixel_avg_variance64x64 avx2 msa sse2 ssse3/; + + add_proto qw/uint32_t aom_sub_pixel_avg_variance64x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred"; + specialize qw/aom_sub_pixel_avg_variance64x32 msa sse2 ssse3/; + + add_proto qw/uint32_t aom_sub_pixel_avg_variance32x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred"; + specialize qw/aom_sub_pixel_avg_variance32x64 msa sse2 ssse3/; + + add_proto qw/uint32_t aom_sub_pixel_avg_variance32x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred"; + specialize qw/aom_sub_pixel_avg_variance32x32 avx2 msa sse2 ssse3/; + + add_proto qw/uint32_t aom_sub_pixel_avg_variance32x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred"; + specialize qw/aom_sub_pixel_avg_variance32x16 msa sse2 ssse3/; + + add_proto qw/uint32_t aom_sub_pixel_avg_variance16x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred"; + specialize qw/aom_sub_pixel_avg_variance16x32 msa sse2 ssse3/; + + add_proto qw/uint32_t aom_sub_pixel_avg_variance16x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred"; + specialize qw/aom_sub_pixel_avg_variance16x16 msa sse2 ssse3/; + + add_proto qw/uint32_t aom_sub_pixel_avg_variance16x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred"; + specialize qw/aom_sub_pixel_avg_variance16x8 msa sse2 ssse3/; + + add_proto qw/uint32_t aom_sub_pixel_avg_variance8x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred"; + specialize qw/aom_sub_pixel_avg_variance8x16 msa sse2 ssse3/; + + add_proto qw/uint32_t aom_sub_pixel_avg_variance8x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred"; + specialize qw/aom_sub_pixel_avg_variance8x8 msa sse2 ssse3/; + + add_proto qw/uint32_t aom_sub_pixel_avg_variance8x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred"; + specialize qw/aom_sub_pixel_avg_variance8x4 msa sse2 ssse3/; + + add_proto qw/uint32_t aom_sub_pixel_avg_variance4x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred"; + specialize qw/aom_sub_pixel_avg_variance4x8 msa sse2 ssse3/; + + add_proto qw/uint32_t aom_sub_pixel_avg_variance4x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred"; + specialize qw/aom_sub_pixel_avg_variance4x4 msa sse2 ssse3/; + + # + # Comp Avg + # + add_proto qw/void aom_comp_avg_pred/, "uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride"; + + add_proto qw/void aom_jnt_comp_avg_pred/, "uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const JNT_COMP_PARAMS *jcp_param"; + specialize qw/aom_jnt_comp_avg_pred ssse3/; + + add_proto qw/unsigned int aom_highbd_12_variance128x128/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"; + specialize qw/aom_highbd_12_variance128x128 sse2/; + + add_proto qw/unsigned int aom_highbd_12_variance128x64/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"; + specialize qw/aom_highbd_12_variance128x64 sse2/; + + add_proto qw/unsigned int aom_highbd_12_variance64x128/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"; + specialize qw/aom_highbd_12_variance64x128 sse2/; + + add_proto qw/unsigned int aom_highbd_12_variance64x64/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"; + specialize qw/aom_highbd_12_variance64x64 sse2/; + + add_proto qw/unsigned int aom_highbd_12_variance64x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"; + specialize qw/aom_highbd_12_variance64x32 sse2/; + + add_proto qw/unsigned int aom_highbd_12_variance32x64/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"; + specialize qw/aom_highbd_12_variance32x64 sse2/; + + add_proto qw/unsigned int aom_highbd_12_variance32x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"; + specialize qw/aom_highbd_12_variance32x32 sse2/; + + add_proto qw/unsigned int aom_highbd_12_variance32x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"; + specialize qw/aom_highbd_12_variance32x16 sse2/; + + add_proto qw/unsigned int aom_highbd_12_variance16x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"; + specialize qw/aom_highbd_12_variance16x32 sse2/; + + add_proto qw/unsigned int aom_highbd_12_variance16x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"; + specialize qw/aom_highbd_12_variance16x16 sse2/; + + add_proto qw/unsigned int aom_highbd_12_variance16x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"; + specialize qw/aom_highbd_12_variance16x8 sse2/; + + add_proto qw/unsigned int aom_highbd_12_variance8x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"; + specialize qw/aom_highbd_12_variance8x16 sse2/; + + add_proto qw/unsigned int aom_highbd_12_variance8x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"; + specialize qw/aom_highbd_12_variance8x8 sse2/; + + add_proto qw/unsigned int aom_highbd_12_variance8x4/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"; + add_proto qw/unsigned int aom_highbd_12_variance4x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"; + add_proto qw/unsigned int aom_highbd_12_variance4x4/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"; + + add_proto qw/unsigned int aom_highbd_10_variance128x128/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"; + specialize qw/aom_highbd_10_variance128x128 sse2 avx2/; + + add_proto qw/unsigned int aom_highbd_10_variance128x64/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"; + specialize qw/aom_highbd_10_variance128x64 sse2 avx2/; + + add_proto qw/unsigned int aom_highbd_10_variance64x128/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"; + specialize qw/aom_highbd_10_variance64x128 sse2 avx2/; + + add_proto qw/unsigned int aom_highbd_10_variance64x64/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"; + specialize qw/aom_highbd_10_variance64x64 sse2 avx2/; + + add_proto qw/unsigned int aom_highbd_10_variance64x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"; + specialize qw/aom_highbd_10_variance64x32 sse2 avx2/; + + add_proto qw/unsigned int aom_highbd_10_variance32x64/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"; + specialize qw/aom_highbd_10_variance32x64 sse2 avx2/; + + add_proto qw/unsigned int aom_highbd_10_variance32x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"; + specialize qw/aom_highbd_10_variance32x32 sse2 avx2/; + + add_proto qw/unsigned int aom_highbd_10_variance32x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"; + specialize qw/aom_highbd_10_variance32x16 sse2 avx2/; + + add_proto qw/unsigned int aom_highbd_10_variance16x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"; + specialize qw/aom_highbd_10_variance16x32 sse2 avx2/; + + add_proto qw/unsigned int aom_highbd_10_variance16x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"; + specialize qw/aom_highbd_10_variance16x16 sse2 avx2/; + + add_proto qw/unsigned int aom_highbd_10_variance16x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"; + specialize qw/aom_highbd_10_variance16x8 sse2 avx2/; + + add_proto qw/unsigned int aom_highbd_10_variance8x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"; + specialize qw/aom_highbd_10_variance8x16 sse2 avx2/; + + add_proto qw/unsigned int aom_highbd_10_variance8x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"; + specialize qw/aom_highbd_10_variance8x8 sse2 avx2/; + + add_proto qw/unsigned int aom_highbd_10_variance8x4/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"; + add_proto qw/unsigned int aom_highbd_10_variance4x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"; + add_proto qw/unsigned int aom_highbd_10_variance4x4/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"; + + add_proto qw/unsigned int aom_highbd_8_variance128x128/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"; + specialize qw/aom_highbd_8_variance128x128 sse2/; + + add_proto qw/unsigned int aom_highbd_8_variance128x64/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"; + specialize qw/aom_highbd_8_variance128x64 sse2/; + + add_proto qw/unsigned int aom_highbd_8_variance64x128/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"; + specialize qw/aom_highbd_8_variance64x128 sse2/; + + add_proto qw/unsigned int aom_highbd_8_variance64x64/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"; + specialize qw/aom_highbd_8_variance64x64 sse2/; + + add_proto qw/unsigned int aom_highbd_8_variance64x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"; + specialize qw/aom_highbd_8_variance64x32 sse2/; + + add_proto qw/unsigned int aom_highbd_8_variance32x64/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"; + specialize qw/aom_highbd_8_variance32x64 sse2/; + + add_proto qw/unsigned int aom_highbd_8_variance32x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"; + specialize qw/aom_highbd_8_variance32x32 sse2/; + + add_proto qw/unsigned int aom_highbd_8_variance32x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"; + specialize qw/aom_highbd_8_variance32x16 sse2/; + + add_proto qw/unsigned int aom_highbd_8_variance16x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"; + specialize qw/aom_highbd_8_variance16x32 sse2/; + + add_proto qw/unsigned int aom_highbd_8_variance16x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"; + specialize qw/aom_highbd_8_variance16x16 sse2/; + + add_proto qw/unsigned int aom_highbd_8_variance16x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"; + specialize qw/aom_highbd_8_variance16x8 sse2/; + + add_proto qw/unsigned int aom_highbd_8_variance8x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"; + specialize qw/aom_highbd_8_variance8x16 sse2/; + + add_proto qw/unsigned int aom_highbd_8_variance8x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"; + specialize qw/aom_highbd_8_variance8x8 sse2/; + + add_proto qw/unsigned int aom_highbd_8_variance8x4/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"; + add_proto qw/unsigned int aom_highbd_8_variance4x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"; + add_proto qw/unsigned int aom_highbd_8_variance4x4/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"; + + add_proto qw/void aom_highbd_8_get16x16var/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum"; + add_proto qw/void aom_highbd_8_get8x8var/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum"; + + add_proto qw/void aom_highbd_10_get16x16var/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum"; + add_proto qw/void aom_highbd_10_get8x8var/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum"; + + add_proto qw/void aom_highbd_12_get16x16var/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum"; + add_proto qw/void aom_highbd_12_get8x8var/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum"; + + add_proto qw/unsigned int aom_highbd_8_mse16x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int recon_stride, unsigned int *sse"; + specialize qw/aom_highbd_8_mse16x16 sse2/; + + add_proto qw/unsigned int aom_highbd_8_mse16x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int recon_stride, unsigned int *sse"; + add_proto qw/unsigned int aom_highbd_8_mse8x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int recon_stride, unsigned int *sse"; + add_proto qw/unsigned int aom_highbd_8_mse8x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int recon_stride, unsigned int *sse"; + specialize qw/aom_highbd_8_mse8x8 sse2/; + + add_proto qw/unsigned int aom_highbd_10_mse16x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int recon_stride, unsigned int *sse"; + specialize qw/aom_highbd_10_mse16x16 sse2/; + + add_proto qw/unsigned int aom_highbd_10_mse16x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int recon_stride, unsigned int *sse"; + add_proto qw/unsigned int aom_highbd_10_mse8x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int recon_stride, unsigned int *sse"; + add_proto qw/unsigned int aom_highbd_10_mse8x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int recon_stride, unsigned int *sse"; + specialize qw/aom_highbd_10_mse8x8 sse2/; + + add_proto qw/unsigned int aom_highbd_12_mse16x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int recon_stride, unsigned int *sse"; + specialize qw/aom_highbd_12_mse16x16 sse2/; + + add_proto qw/unsigned int aom_highbd_12_mse16x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int recon_stride, unsigned int *sse"; + add_proto qw/unsigned int aom_highbd_12_mse8x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int recon_stride, unsigned int *sse"; + add_proto qw/unsigned int aom_highbd_12_mse8x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int recon_stride, unsigned int *sse"; + specialize qw/aom_highbd_12_mse8x8 sse2/; + + add_proto qw/void aom_highbd_comp_avg_pred/, "uint8_t *comp_pred8, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride"; + + add_proto qw/void aom_highbd_jnt_comp_avg_pred/, "uint8_t *comp_pred8, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride, const JNT_COMP_PARAMS *jcp_param"; + specialize qw/aom_highbd_jnt_comp_avg_pred sse2/; + + # + # Subpixel Variance + # + add_proto qw/uint32_t aom_highbd_12_sub_pixel_variance64x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse"; + specialize qw/aom_highbd_12_sub_pixel_variance64x64 sse2/; + + add_proto qw/uint32_t aom_highbd_12_sub_pixel_variance64x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse"; + specialize qw/aom_highbd_12_sub_pixel_variance64x32 sse2/; + + add_proto qw/uint32_t aom_highbd_12_sub_pixel_variance32x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse"; + specialize qw/aom_highbd_12_sub_pixel_variance32x64 sse2/; + + add_proto qw/uint32_t aom_highbd_12_sub_pixel_variance32x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse"; + specialize qw/aom_highbd_12_sub_pixel_variance32x32 sse2/; + + add_proto qw/uint32_t aom_highbd_12_sub_pixel_variance32x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse"; + specialize qw/aom_highbd_12_sub_pixel_variance32x16 sse2/; + + add_proto qw/uint32_t aom_highbd_12_sub_pixel_variance16x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse"; + specialize qw/aom_highbd_12_sub_pixel_variance16x32 sse2/; + + add_proto qw/uint32_t aom_highbd_12_sub_pixel_variance16x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse"; + specialize qw/aom_highbd_12_sub_pixel_variance16x16 sse2/; + + add_proto qw/uint32_t aom_highbd_12_sub_pixel_variance16x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse"; + specialize qw/aom_highbd_12_sub_pixel_variance16x8 sse2/; + + add_proto qw/uint32_t aom_highbd_12_sub_pixel_variance8x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse"; + specialize qw/aom_highbd_12_sub_pixel_variance8x16 sse2/; + + add_proto qw/uint32_t aom_highbd_12_sub_pixel_variance8x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse"; + specialize qw/aom_highbd_12_sub_pixel_variance8x8 sse2/; + + add_proto qw/uint32_t aom_highbd_12_sub_pixel_variance8x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse"; + specialize qw/aom_highbd_12_sub_pixel_variance8x4 sse2/; + + add_proto qw/uint32_t aom_highbd_12_sub_pixel_variance4x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse"; + add_proto qw/uint32_t aom_highbd_12_sub_pixel_variance4x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse"; + + add_proto qw/uint32_t aom_highbd_10_sub_pixel_variance64x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse"; + specialize qw/aom_highbd_10_sub_pixel_variance64x64 sse2/; + + add_proto qw/uint32_t aom_highbd_10_sub_pixel_variance64x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse"; + specialize qw/aom_highbd_10_sub_pixel_variance64x32 sse2/; + + add_proto qw/uint32_t aom_highbd_10_sub_pixel_variance32x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse"; + specialize qw/aom_highbd_10_sub_pixel_variance32x64 sse2/; + + add_proto qw/uint32_t aom_highbd_10_sub_pixel_variance32x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse"; + specialize qw/aom_highbd_10_sub_pixel_variance32x32 sse2/; + + add_proto qw/uint32_t aom_highbd_10_sub_pixel_variance32x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse"; + specialize qw/aom_highbd_10_sub_pixel_variance32x16 sse2/; + + add_proto qw/uint32_t aom_highbd_10_sub_pixel_variance16x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse"; + specialize qw/aom_highbd_10_sub_pixel_variance16x32 sse2/; + + add_proto qw/uint32_t aom_highbd_10_sub_pixel_variance16x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse"; + specialize qw/aom_highbd_10_sub_pixel_variance16x16 sse2/; + + add_proto qw/uint32_t aom_highbd_10_sub_pixel_variance16x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse"; + specialize qw/aom_highbd_10_sub_pixel_variance16x8 sse2/; + + add_proto qw/uint32_t aom_highbd_10_sub_pixel_variance8x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse"; + specialize qw/aom_highbd_10_sub_pixel_variance8x16 sse2/; + + add_proto qw/uint32_t aom_highbd_10_sub_pixel_variance8x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse"; + specialize qw/aom_highbd_10_sub_pixel_variance8x8 sse2/; + + add_proto qw/uint32_t aom_highbd_10_sub_pixel_variance8x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse"; + specialize qw/aom_highbd_10_sub_pixel_variance8x4 sse2/; + + add_proto qw/uint32_t aom_highbd_10_sub_pixel_variance4x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse"; + add_proto qw/uint32_t aom_highbd_10_sub_pixel_variance4x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse"; + + add_proto qw/uint32_t aom_highbd_8_sub_pixel_variance64x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse"; + specialize qw/aom_highbd_8_sub_pixel_variance64x64 sse2/; + + add_proto qw/uint32_t aom_highbd_8_sub_pixel_variance64x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse"; + specialize qw/aom_highbd_8_sub_pixel_variance64x32 sse2/; + + add_proto qw/uint32_t aom_highbd_8_sub_pixel_variance32x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse"; + specialize qw/aom_highbd_8_sub_pixel_variance32x64 sse2/; + + add_proto qw/uint32_t aom_highbd_8_sub_pixel_variance32x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse"; + specialize qw/aom_highbd_8_sub_pixel_variance32x32 sse2/; + + add_proto qw/uint32_t aom_highbd_8_sub_pixel_variance32x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse"; + specialize qw/aom_highbd_8_sub_pixel_variance32x16 sse2/; + + add_proto qw/uint32_t aom_highbd_8_sub_pixel_variance16x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse"; + specialize qw/aom_highbd_8_sub_pixel_variance16x32 sse2/; + + add_proto qw/uint32_t aom_highbd_8_sub_pixel_variance16x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse"; + specialize qw/aom_highbd_8_sub_pixel_variance16x16 sse2/; + + add_proto qw/uint32_t aom_highbd_8_sub_pixel_variance16x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse"; + specialize qw/aom_highbd_8_sub_pixel_variance16x8 sse2/; + + add_proto qw/uint32_t aom_highbd_8_sub_pixel_variance8x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse"; + specialize qw/aom_highbd_8_sub_pixel_variance8x16 sse2/; + + add_proto qw/uint32_t aom_highbd_8_sub_pixel_variance8x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse"; + specialize qw/aom_highbd_8_sub_pixel_variance8x8 sse2/; + + add_proto qw/uint32_t aom_highbd_8_sub_pixel_variance8x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse"; + specialize qw/aom_highbd_8_sub_pixel_variance8x4 sse2/; + + add_proto qw/uint32_t aom_highbd_8_sub_pixel_variance4x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse"; + add_proto qw/uint32_t aom_highbd_8_sub_pixel_variance4x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse"; + + add_proto qw/uint32_t aom_highbd_12_sub_pixel_avg_variance64x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred"; + specialize qw/aom_highbd_12_sub_pixel_avg_variance64x64 sse2/; + + add_proto qw/uint32_t aom_highbd_12_sub_pixel_avg_variance64x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred"; + specialize qw/aom_highbd_12_sub_pixel_avg_variance64x32 sse2/; + + add_proto qw/uint32_t aom_highbd_12_sub_pixel_avg_variance32x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred"; + specialize qw/aom_highbd_12_sub_pixel_avg_variance32x64 sse2/; + + add_proto qw/uint32_t aom_highbd_12_sub_pixel_avg_variance32x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred"; + specialize qw/aom_highbd_12_sub_pixel_avg_variance32x32 sse2/; + + add_proto qw/uint32_t aom_highbd_12_sub_pixel_avg_variance32x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred"; + specialize qw/aom_highbd_12_sub_pixel_avg_variance32x16 sse2/; + + add_proto qw/uint32_t aom_highbd_12_sub_pixel_avg_variance16x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred"; + specialize qw/aom_highbd_12_sub_pixel_avg_variance16x32 sse2/; + + add_proto qw/uint32_t aom_highbd_12_sub_pixel_avg_variance16x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred"; + specialize qw/aom_highbd_12_sub_pixel_avg_variance16x16 sse2/; + + add_proto qw/uint32_t aom_highbd_12_sub_pixel_avg_variance16x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred"; + specialize qw/aom_highbd_12_sub_pixel_avg_variance16x8 sse2/; + + add_proto qw/uint32_t aom_highbd_12_sub_pixel_avg_variance8x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred"; + specialize qw/aom_highbd_12_sub_pixel_avg_variance8x16 sse2/; + + add_proto qw/uint32_t aom_highbd_12_sub_pixel_avg_variance8x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred"; + specialize qw/aom_highbd_12_sub_pixel_avg_variance8x8 sse2/; + + add_proto qw/uint32_t aom_highbd_12_sub_pixel_avg_variance8x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred"; + specialize qw/aom_highbd_12_sub_pixel_avg_variance8x4 sse2/; + + add_proto qw/uint32_t aom_highbd_12_sub_pixel_avg_variance4x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred"; + add_proto qw/uint32_t aom_highbd_12_sub_pixel_avg_variance4x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred"; + + add_proto qw/uint32_t aom_highbd_10_sub_pixel_avg_variance64x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred"; + specialize qw/aom_highbd_10_sub_pixel_avg_variance64x64 sse2/; + + add_proto qw/uint32_t aom_highbd_10_sub_pixel_avg_variance64x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred"; + specialize qw/aom_highbd_10_sub_pixel_avg_variance64x32 sse2/; + + add_proto qw/uint32_t aom_highbd_10_sub_pixel_avg_variance32x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred"; + specialize qw/aom_highbd_10_sub_pixel_avg_variance32x64 sse2/; + + add_proto qw/uint32_t aom_highbd_10_sub_pixel_avg_variance32x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred"; + specialize qw/aom_highbd_10_sub_pixel_avg_variance32x32 sse2/; + + add_proto qw/uint32_t aom_highbd_10_sub_pixel_avg_variance32x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred"; + specialize qw/aom_highbd_10_sub_pixel_avg_variance32x16 sse2/; + + add_proto qw/uint32_t aom_highbd_10_sub_pixel_avg_variance16x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred"; + specialize qw/aom_highbd_10_sub_pixel_avg_variance16x32 sse2/; + + add_proto qw/uint32_t aom_highbd_10_sub_pixel_avg_variance16x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred"; + specialize qw/aom_highbd_10_sub_pixel_avg_variance16x16 sse2/; + + add_proto qw/uint32_t aom_highbd_10_sub_pixel_avg_variance16x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred"; + specialize qw/aom_highbd_10_sub_pixel_avg_variance16x8 sse2/; + + add_proto qw/uint32_t aom_highbd_10_sub_pixel_avg_variance8x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred"; + specialize qw/aom_highbd_10_sub_pixel_avg_variance8x16 sse2/; + + add_proto qw/uint32_t aom_highbd_10_sub_pixel_avg_variance8x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred"; + specialize qw/aom_highbd_10_sub_pixel_avg_variance8x8 sse2/; + + add_proto qw/uint32_t aom_highbd_10_sub_pixel_avg_variance8x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred"; + specialize qw/aom_highbd_10_sub_pixel_avg_variance8x4 sse2/; + + add_proto qw/uint32_t aom_highbd_10_sub_pixel_avg_variance4x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred"; + add_proto qw/uint32_t aom_highbd_10_sub_pixel_avg_variance4x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred"; + + add_proto qw/uint32_t aom_highbd_8_sub_pixel_avg_variance64x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred"; + specialize qw/aom_highbd_8_sub_pixel_avg_variance64x64 sse2/; + + add_proto qw/uint32_t aom_highbd_8_sub_pixel_avg_variance64x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred"; + specialize qw/aom_highbd_8_sub_pixel_avg_variance64x32 sse2/; + + add_proto qw/uint32_t aom_highbd_8_sub_pixel_avg_variance32x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred"; + specialize qw/aom_highbd_8_sub_pixel_avg_variance32x64 sse2/; + + add_proto qw/uint32_t aom_highbd_8_sub_pixel_avg_variance32x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred"; + specialize qw/aom_highbd_8_sub_pixel_avg_variance32x32 sse2/; + + add_proto qw/uint32_t aom_highbd_8_sub_pixel_avg_variance32x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred"; + specialize qw/aom_highbd_8_sub_pixel_avg_variance32x16 sse2/; + + add_proto qw/uint32_t aom_highbd_8_sub_pixel_avg_variance16x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred"; + specialize qw/aom_highbd_8_sub_pixel_avg_variance16x32 sse2/; + + add_proto qw/uint32_t aom_highbd_8_sub_pixel_avg_variance16x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred"; + specialize qw/aom_highbd_8_sub_pixel_avg_variance16x16 sse2/; + + add_proto qw/uint32_t aom_highbd_8_sub_pixel_avg_variance16x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred"; + specialize qw/aom_highbd_8_sub_pixel_avg_variance16x8 sse2/; + + add_proto qw/uint32_t aom_highbd_8_sub_pixel_avg_variance8x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred"; + specialize qw/aom_highbd_8_sub_pixel_avg_variance8x16 sse2/; + + add_proto qw/uint32_t aom_highbd_8_sub_pixel_avg_variance8x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred"; + specialize qw/aom_highbd_8_sub_pixel_avg_variance8x8 sse2/; + + add_proto qw/uint32_t aom_highbd_8_sub_pixel_avg_variance8x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred"; + specialize qw/aom_highbd_8_sub_pixel_avg_variance8x4 sse2/; + + add_proto qw/uint32_t aom_highbd_8_sub_pixel_avg_variance4x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred"; + add_proto qw/uint32_t aom_highbd_8_sub_pixel_avg_variance4x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred"; + + + + add_proto qw/void aom_comp_mask_pred/, "uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask"; + specialize qw/aom_comp_mask_pred ssse3 avx2/; + + add_proto qw/void aom_highbd_comp_mask_pred/, "uint8_t *comp_pred, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask"; + specialize qw/aom_highbd_comp_mask_pred sse2 avx2/; + +} # CONFIG_AV1_ENCODER + +1; diff --git a/third_party/aom/aom_dsp/aom_filter.h b/third_party/aom/aom_dsp/aom_filter.h new file mode 100644 index 0000000000..00686ac388 --- /dev/null +++ b/third_party/aom/aom_dsp/aom_filter.h @@ -0,0 +1,56 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_AOM_FILTER_H_ +#define AOM_AOM_DSP_AOM_FILTER_H_ + +#include "aom/aom_integer.h" + +#ifdef __cplusplus +extern "C" { +#endif + +#define FILTER_BITS 7 + +#define SUBPEL_BITS 4 +#define SUBPEL_MASK ((1 << SUBPEL_BITS) - 1) +#define SUBPEL_SHIFTS (1 << SUBPEL_BITS) +#define SUBPEL_TAPS 8 + +#define SCALE_SUBPEL_BITS 10 +#define SCALE_SUBPEL_SHIFTS (1 << SCALE_SUBPEL_BITS) +#define SCALE_SUBPEL_MASK (SCALE_SUBPEL_SHIFTS - 1) +#define SCALE_EXTRA_BITS (SCALE_SUBPEL_BITS - SUBPEL_BITS) +#define SCALE_EXTRA_OFF ((1 << SCALE_EXTRA_BITS) / 2) + +#define RS_SUBPEL_BITS 6 +#define RS_SUBPEL_MASK ((1 << RS_SUBPEL_BITS) - 1) +#define RS_SCALE_SUBPEL_BITS 14 +#define RS_SCALE_SUBPEL_MASK ((1 << RS_SCALE_SUBPEL_BITS) - 1) +#define RS_SCALE_EXTRA_BITS (RS_SCALE_SUBPEL_BITS - RS_SUBPEL_BITS) +#define RS_SCALE_EXTRA_OFF (1 << (RS_SCALE_EXTRA_BITS - 1)) + +typedef int16_t InterpKernel[SUBPEL_TAPS]; + +#define BIL_SUBPEL_BITS 3 +#define BIL_SUBPEL_SHIFTS (1 << BIL_SUBPEL_BITS) + +// 2 tap bilinear filters +static const uint8_t bilinear_filters_2t[BIL_SUBPEL_SHIFTS][2] = { + { 128, 0 }, { 112, 16 }, { 96, 32 }, { 80, 48 }, + { 64, 64 }, { 48, 80 }, { 32, 96 }, { 16, 112 }, +}; + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif // AOM_AOM_DSP_AOM_FILTER_H_ diff --git a/third_party/aom/aom_dsp/aom_simd.h b/third_party/aom/aom_dsp/aom_simd.h new file mode 100644 index 0000000000..ab950ca55b --- /dev/null +++ b/third_party/aom/aom_dsp/aom_simd.h @@ -0,0 +1,38 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_AOM_SIMD_H_ +#define AOM_AOM_DSP_AOM_SIMD_H_ + +#include + +#if defined(_WIN32) +#include +#endif + +#include "config/aom_config.h" + +#include "aom_dsp/aom_simd_inline.h" + +#define SIMD_CHECK 1 // Sanity checks in C equivalents + +#if HAVE_NEON +#include "simd/v256_intrinsics_arm.h" +// VS compiling for 32 bit targets does not support vector types in +// structs as arguments, which makes the v256 type of the intrinsics +// hard to support, so optimizations for this target are disabled. +#elif HAVE_SSE2 && (defined(_WIN64) || !defined(_MSC_VER) || defined(__clang__)) +#include "simd/v256_intrinsics_x86.h" +#else +#include "simd/v256_intrinsics.h" +#endif + +#endif // AOM_AOM_DSP_AOM_SIMD_H_ diff --git a/third_party/aom/aom_dsp/aom_simd_inline.h b/third_party/aom/aom_dsp/aom_simd_inline.h new file mode 100644 index 0000000000..eb333f6f66 --- /dev/null +++ b/third_party/aom/aom_dsp/aom_simd_inline.h @@ -0,0 +1,21 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_AOM_SIMD_INLINE_H_ +#define AOM_AOM_DSP_AOM_SIMD_INLINE_H_ + +#include "aom/aom_integer.h" + +#ifndef SIMD_INLINE +#define SIMD_INLINE static AOM_FORCE_INLINE +#endif + +#endif // AOM_AOM_DSP_AOM_SIMD_INLINE_H_ diff --git a/third_party/aom/aom_dsp/arm/blend_a64_mask_neon.c b/third_party/aom/aom_dsp/arm/blend_a64_mask_neon.c new file mode 100644 index 0000000000..e7f08a5fdb --- /dev/null +++ b/third_party/aom/aom_dsp/arm/blend_a64_mask_neon.c @@ -0,0 +1,451 @@ +/* + * Copyright (c) 2018, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include +#include + +#include "aom/aom_integer.h" +#include "aom_dsp/aom_dsp_common.h" +#include "aom_dsp/blend.h" +#include "aom_ports/mem.h" +#include "av1/common/arm/mem_neon.h" +#include "config/aom_dsp_rtcd.h" + +static INLINE void blend8x1(int16x8_t mask, int16x8_t src_0, int16x8_t src_1, + const int16x8_t v_maxval, int16x8_t *res) { + int32x4_t im_res_low, im_res_high; + const int16x8_t max_minus_mask = vsubq_s16(v_maxval, mask); + + im_res_low = vmull_s16(vget_low_s16(mask), vget_low_s16(src_0)); + im_res_low = + vmlal_s16(im_res_low, vget_low_s16(max_minus_mask), vget_low_s16(src_1)); + + im_res_high = vmull_s16(vget_high_s16(mask), vget_high_s16(src_0)); + im_res_high = vmlal_s16(im_res_high, vget_high_s16(max_minus_mask), + vget_high_s16(src_1)); + + *res = vcombine_s16(vshrn_n_s32(im_res_low, AOM_BLEND_A64_ROUND_BITS), + vshrn_n_s32(im_res_high, AOM_BLEND_A64_ROUND_BITS)); +} + +static INLINE void blend_8x4(uint8_t *dst, uint32_t dst_stride, + const CONV_BUF_TYPE *src0, uint32_t src0_stride, + const CONV_BUF_TYPE *src1, uint32_t src1_stride, + int16x8_t mask0, int16x8_t mask1, int16x8_t mask2, + int16x8_t mask3, const int16x8_t v_maxval, + const uint16x8_t vec_round_offset, + const int16x8_t vec_round_bits) { + int16x8_t src0_0, src0_1, src0_2, src0_3; + int16x8_t src1_0, src1_1, src1_2, src1_3; + int16x8_t im_res_0, im_res_1, im_res_2, im_res_3; + + load_s16_8x4((int16_t *)src0, (int32_t)src0_stride, &src0_0, &src0_1, &src0_2, + &src0_3); + load_s16_8x4((int16_t *)src1, (int32_t)src1_stride, &src1_0, &src1_1, &src1_2, + &src1_3); + + blend8x1(mask0, src0_0, src1_0, v_maxval, &im_res_0); + blend8x1(mask1, src0_1, src1_1, v_maxval, &im_res_1); + blend8x1(mask2, src0_2, src1_2, v_maxval, &im_res_2); + blend8x1(mask3, src0_3, src1_3, v_maxval, &im_res_3); + + uint16x8_t im_res1_0 = + vqsubq_u16(vreinterpretq_u16_s16(im_res_0), vec_round_offset); + uint16x8_t im_res1_1 = + vqsubq_u16(vreinterpretq_u16_s16(im_res_1), vec_round_offset); + uint16x8_t im_res1_2 = + vqsubq_u16(vreinterpretq_u16_s16(im_res_2), vec_round_offset); + uint16x8_t im_res1_3 = + vqsubq_u16(vreinterpretq_u16_s16(im_res_3), vec_round_offset); + + im_res_0 = vshlq_s16(vreinterpretq_s16_u16(im_res1_0), vec_round_bits); + im_res_1 = vshlq_s16(vreinterpretq_s16_u16(im_res1_1), vec_round_bits); + im_res_2 = vshlq_s16(vreinterpretq_s16_u16(im_res1_2), vec_round_bits); + im_res_3 = vshlq_s16(vreinterpretq_s16_u16(im_res1_3), vec_round_bits); + + vst1_u8((dst + 0 * dst_stride), vqmovun_s16(im_res_0)); + vst1_u8((dst + 1 * dst_stride), vqmovun_s16(im_res_1)); + vst1_u8((dst + 2 * dst_stride), vqmovun_s16(im_res_2)); + vst1_u8((dst + 3 * dst_stride), vqmovun_s16(im_res_3)); +} + +static INLINE void blend_4x4(uint8_t *dst, uint32_t dst_stride, + const CONV_BUF_TYPE *src0, uint32_t src0_stride, + const CONV_BUF_TYPE *src1, uint32_t src1_stride, + int16x4_t mask0, int16x4_t mask1, int16x4_t mask2, + int16x4_t mask3, const int16x8_t v_maxval, + const uint16x8_t vec_round_offset, + const int16x8_t vec_round_bits) { + int16x8_t src0_0, src0_1; + int16x8_t src1_0, src1_1; + uint64x2_t tu0 = vdupq_n_u64(0), tu1 = vdupq_n_u64(0), tu2 = vdupq_n_u64(0), + tu3 = vdupq_n_u64(0); + int16x8_t mask0_1, mask2_3; + int16x8_t res0, res1; + + load_unaligned_u16_4x4(src0, src0_stride, &tu0, &tu1); + load_unaligned_u16_4x4(src1, src1_stride, &tu2, &tu3); + + src0_0 = vreinterpretq_s16_u64(tu0); + src0_1 = vreinterpretq_s16_u64(tu1); + + src1_0 = vreinterpretq_s16_u64(tu2); + src1_1 = vreinterpretq_s16_u64(tu3); + + mask0_1 = vcombine_s16(mask0, mask1); + mask2_3 = vcombine_s16(mask2, mask3); + + blend8x1(mask0_1, src0_0, src1_0, v_maxval, &res0); + blend8x1(mask2_3, src0_1, src1_1, v_maxval, &res1); + + uint16x8_t im_res_0 = + vqsubq_u16(vreinterpretq_u16_s16(res0), vec_round_offset); + uint16x8_t im_res_1 = + vqsubq_u16(vreinterpretq_u16_s16(res1), vec_round_offset); + + src0_0 = vshlq_s16(vreinterpretq_s16_u16(im_res_0), vec_round_bits); + src0_1 = vshlq_s16(vreinterpretq_s16_u16(im_res_1), vec_round_bits); + + uint8x8_t res_0 = vqmovun_s16(src0_0); + uint8x8_t res_1 = vqmovun_s16(src0_1); + + vst1_lane_u32((uint32_t *)(dst + 0 * dst_stride), vreinterpret_u32_u8(res_0), + 0); + vst1_lane_u32((uint32_t *)(dst + 1 * dst_stride), vreinterpret_u32_u8(res_0), + 1); + vst1_lane_u32((uint32_t *)(dst + 2 * dst_stride), vreinterpret_u32_u8(res_1), + 0); + vst1_lane_u32((uint32_t *)(dst + 3 * dst_stride), vreinterpret_u32_u8(res_1), + 1); +} + +void aom_lowbd_blend_a64_d16_mask_neon( + uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, + uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, + ConvolveParams *conv_params) { + int i = 0; + const int bd = 8; + int w_tmp = w; + const uint8_t *mask_tmp = mask; + const CONV_BUF_TYPE *src0_tmp = src0; + const CONV_BUF_TYPE *src1_tmp = src1; + uint8_t *dst_tmp = dst; + + const int offset_bits = bd + 2 * FILTER_BITS - conv_params->round_0; + const int round_offset = (1 << (offset_bits - conv_params->round_1)) + + (1 << (offset_bits - conv_params->round_1 - 1)); + const int round_bits = + 2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1; + + assert(IMPLIES((void *)src0 == dst, src0_stride == dst_stride)); + assert(IMPLIES((void *)src1 == dst, src1_stride == dst_stride)); + + assert(h >= 4); + assert(w >= 4); + assert(IS_POWER_OF_TWO(h)); + assert(IS_POWER_OF_TWO(w)); + + uint8x8_t s0, s1, s2, s3; + uint32x2_t tu0 = vdup_n_u32(0), tu1 = vdup_n_u32(0), tu2 = vdup_n_u32(0), + tu3 = vdup_n_u32(0); + uint8x16_t t0, t1, t2, t3, t4, t5, t6, t7; + int16x8_t mask0, mask1, mask2, mask3; + int16x8_t mask4, mask5, mask6, mask7; + int32x4_t m0_32, m1_32, m2_32, m3_32; + int32x4_t m4_32, m5_32, m6_32, m7_32; + uint8x8_t mask0_l, mask1_l, mask2_l, mask3_l; + uint8x8_t mask4_l, mask5_l, mask6_l, mask7_l; + int16x4_t mask0_low, mask1_low, mask2_low, mask3_low; + const uint16x4_t vec_zero = vdup_n_u16(0); + const uint16_t offset = round_offset - (1 << (round_bits - 1)); + const int16x8_t v_maxval = vdupq_n_s16(AOM_BLEND_A64_MAX_ALPHA); + const int16x8_t vec_round_bits = vdupq_n_s16(-round_bits); + const uint16x8_t vec_offset = vdupq_n_u16(offset); + + if (subw == 0 && subh == 0) { + if (w_tmp > 7) { + do { + w_tmp = w; + do { + load_u8_8x4(mask_tmp, mask_stride, &s0, &s1, &s2, &s3); + + mask0 = vmovl_s8(vreinterpret_s8_u8(s0)); + mask1 = vmovl_s8(vreinterpret_s8_u8(s1)); + mask2 = vmovl_s8(vreinterpret_s8_u8(s2)); + mask3 = vmovl_s8(vreinterpret_s8_u8(s3)); + + blend_8x4(dst_tmp, dst_stride, src0_tmp, src0_stride, src1_tmp, + src1_stride, mask0, mask1, mask2, mask3, v_maxval, + vec_offset, vec_round_bits); + + w_tmp -= 8; + mask_tmp += 8; + dst_tmp += 8; + src0_tmp += 8; + src1_tmp += 8; + } while (w_tmp > 7); + i += 4; + mask_tmp += (4 * mask_stride) - w; + dst_tmp += (4 * dst_stride) - w; + src0_tmp += (4 * src0_stride) - w; + src1_tmp += (4 * src1_stride) - w; + } while (i < h); + } else { + do { + load_unaligned_u8_4x4(mask_tmp, mask_stride, &tu0, &tu1); + + mask0 = vreinterpretq_s16_u16(vmovl_u8(vreinterpret_u8_u32(tu0))); + mask1 = vreinterpretq_s16_u16(vmovl_u8(vreinterpret_u8_u32(tu1))); + + mask0_low = vget_low_s16(mask0); + mask1_low = vget_high_s16(mask0); + mask2_low = vget_low_s16(mask1); + mask3_low = vget_high_s16(mask1); + + blend_4x4(dst_tmp, dst_stride, src0_tmp, src0_stride, src1_tmp, + src1_stride, mask0_low, mask1_low, mask2_low, mask3_low, + v_maxval, vec_offset, vec_round_bits); + + i += 4; + mask_tmp += (4 * mask_stride); + dst_tmp += (4 * dst_stride); + src0_tmp += (4 * src0_stride); + src1_tmp += (4 * src1_stride); + } while (i < h); + } + } else if (subw == 1 && subh == 1) { + if (w_tmp > 7) { + do { + w_tmp = w; + do { + load_u8_16x8(mask_tmp, mask_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6, + &t7); + + mask0 = + vreinterpretq_s16_u16(vaddl_u8(vget_low_u8(t0), vget_low_u8(t1))); + mask1 = + vreinterpretq_s16_u16(vaddl_u8(vget_low_u8(t2), vget_low_u8(t3))); + mask2 = + vreinterpretq_s16_u16(vaddl_u8(vget_low_u8(t4), vget_low_u8(t5))); + mask3 = + vreinterpretq_s16_u16(vaddl_u8(vget_low_u8(t6), vget_low_u8(t7))); + + mask4 = vreinterpretq_s16_u16( + vaddl_u8(vget_high_u8(t0), vget_high_u8(t1))); + mask5 = vreinterpretq_s16_u16( + vaddl_u8(vget_high_u8(t2), vget_high_u8(t3))); + mask6 = vreinterpretq_s16_u16( + vaddl_u8(vget_high_u8(t4), vget_high_u8(t5))); + mask7 = vreinterpretq_s16_u16( + vaddl_u8(vget_high_u8(t6), vget_high_u8(t7))); + + m0_32 = vpaddlq_s16(mask0); + m1_32 = vpaddlq_s16(mask1); + m2_32 = vpaddlq_s16(mask2); + m3_32 = vpaddlq_s16(mask3); + + m4_32 = vpaddlq_s16(mask4); + m5_32 = vpaddlq_s16(mask5); + m6_32 = vpaddlq_s16(mask6); + m7_32 = vpaddlq_s16(mask7); + + mask0 = + vcombine_s16(vqrshrn_n_s32(m0_32, 2), vqrshrn_n_s32(m4_32, 2)); + mask1 = + vcombine_s16(vqrshrn_n_s32(m1_32, 2), vqrshrn_n_s32(m5_32, 2)); + mask2 = + vcombine_s16(vqrshrn_n_s32(m2_32, 2), vqrshrn_n_s32(m6_32, 2)); + mask3 = + vcombine_s16(vqrshrn_n_s32(m3_32, 2), vqrshrn_n_s32(m7_32, 2)); + + blend_8x4(dst_tmp, dst_stride, src0_tmp, src0_stride, src1_tmp, + src1_stride, mask0, mask1, mask2, mask3, v_maxval, + vec_offset, vec_round_bits); + + w_tmp -= 8; + mask_tmp += 16; + dst_tmp += 8; + src0_tmp += 8; + src1_tmp += 8; + } while (w_tmp > 7); + i += 4; + mask_tmp += (8 * mask_stride) - (2 * w); + dst_tmp += (4 * dst_stride) - w; + src0_tmp += (4 * src0_stride) - w; + src1_tmp += (4 * src1_stride) - w; + } while (i < h); + } else { + do { + load_u8_8x8(mask_tmp, mask_stride, &mask0_l, &mask1_l, &mask2_l, + &mask3_l, &mask4_l, &mask5_l, &mask6_l, &mask7_l); + + mask0 = vreinterpretq_s16_u16(vaddl_u8(mask0_l, mask1_l)); + mask1 = vreinterpretq_s16_u16(vaddl_u8(mask2_l, mask3_l)); + mask2 = vreinterpretq_s16_u16(vaddl_u8(mask4_l, mask5_l)); + mask3 = vreinterpretq_s16_u16(vaddl_u8(mask6_l, mask7_l)); + + m0_32 = vpaddlq_s16(mask0); + m1_32 = vpaddlq_s16(mask1); + m2_32 = vpaddlq_s16(mask2); + m3_32 = vpaddlq_s16(mask3); + + mask0_low = vqrshrn_n_s32(m0_32, 2); + mask1_low = vqrshrn_n_s32(m1_32, 2); + mask2_low = vqrshrn_n_s32(m2_32, 2); + mask3_low = vqrshrn_n_s32(m3_32, 2); + + blend_4x4(dst_tmp, dst_stride, src0_tmp, src0_stride, src1_tmp, + src1_stride, mask0_low, mask1_low, mask2_low, mask3_low, + v_maxval, vec_offset, vec_round_bits); + + i += 4; + mask_tmp += (8 * mask_stride); + dst_tmp += (4 * dst_stride); + src0_tmp += (4 * src0_stride); + src1_tmp += (4 * src1_stride); + } while (i < h); + } + } else if (subw == 1 && subh == 0) { + if (w_tmp > 7) { + do { + w_tmp = w; + do { + load_u8_16x4(mask_tmp, mask_stride, &t0, &t1, &t2, &t3); + + mask0 = vreinterpretq_s16_u16(vcombine_u16( + vpaddl_u8(vget_low_u8(t0)), vpaddl_u8(vget_high_u8(t0)))); + mask1 = vreinterpretq_s16_u16(vcombine_u16( + vpaddl_u8(vget_low_u8(t1)), vpaddl_u8(vget_high_u8(t1)))); + mask2 = vreinterpretq_s16_u16(vcombine_u16( + vpaddl_u8(vget_low_u8(t2)), vpaddl_u8(vget_high_u8(t2)))); + mask3 = vreinterpretq_s16_u16(vcombine_u16( + vpaddl_u8(vget_low_u8(t3)), vpaddl_u8(vget_high_u8(t3)))); + + mask0 = vmovl_s8(vqrshrn_n_s16(mask0, 1)); + mask1 = vmovl_s8(vqrshrn_n_s16(mask1, 1)); + mask2 = vmovl_s8(vqrshrn_n_s16(mask2, 1)); + mask3 = vmovl_s8(vqrshrn_n_s16(mask3, 1)); + + blend_8x4(dst_tmp, dst_stride, src0_tmp, src0_stride, src1_tmp, + src1_stride, mask0, mask1, mask2, mask3, v_maxval, + vec_offset, vec_round_bits); + w_tmp -= 8; + mask_tmp += 16; + dst_tmp += 8; + src0_tmp += 8; + src1_tmp += 8; + } while (w_tmp > 7); + i += 4; + mask_tmp += (4 * mask_stride) - (2 * w); + dst_tmp += (4 * dst_stride) - w; + src0_tmp += (4 * src0_stride) - w; + src1_tmp += (4 * src1_stride) - w; + } while (i < h); + } else { + do { + load_u8_8x4(mask_tmp, mask_stride, &mask0_l, &mask1_l, &mask2_l, + &mask3_l); + + mask0 = + vreinterpretq_s16_u16(vcombine_u16(vpaddl_u8(mask0_l), vec_zero)); + mask1 = + vreinterpretq_s16_u16(vcombine_u16(vpaddl_u8(mask1_l), vec_zero)); + mask2 = + vreinterpretq_s16_u16(vcombine_u16(vpaddl_u8(mask2_l), vec_zero)); + mask3 = + vreinterpretq_s16_u16(vcombine_u16(vpaddl_u8(mask3_l), vec_zero)); + + mask0_low = vget_low_s16(vmovl_s8(vqrshrn_n_s16(mask0, 1))); + mask1_low = vget_low_s16(vmovl_s8(vqrshrn_n_s16(mask1, 1))); + mask2_low = vget_low_s16(vmovl_s8(vqrshrn_n_s16(mask2, 1))); + mask3_low = vget_low_s16(vmovl_s8(vqrshrn_n_s16(mask3, 1))); + + blend_4x4(dst_tmp, dst_stride, src0_tmp, src0_stride, src1_tmp, + src1_stride, mask0_low, mask1_low, mask2_low, mask3_low, + v_maxval, vec_offset, vec_round_bits); + + i += 4; + mask_tmp += (4 * mask_stride); + dst_tmp += (4 * dst_stride); + src0_tmp += (4 * src0_stride); + src1_tmp += (4 * src1_stride); + } while (i < h); + } + } else { + if (w_tmp > 7) { + do { + w_tmp = w; + do { + load_u8_8x8(mask_tmp, mask_stride, &mask0_l, &mask1_l, &mask2_l, + &mask3_l, &mask4_l, &mask5_l, &mask6_l, &mask7_l); + + mask0 = vreinterpretq_s16_u16(vaddl_u8(mask0_l, mask1_l)); + mask1 = vreinterpretq_s16_u16(vaddl_u8(mask2_l, mask3_l)); + mask2 = vreinterpretq_s16_u16(vaddl_u8(mask4_l, mask5_l)); + mask3 = vreinterpretq_s16_u16(vaddl_u8(mask6_l, mask7_l)); + + mask0 = vmovl_s8(vqrshrn_n_s16(mask0, 1)); + mask1 = vmovl_s8(vqrshrn_n_s16(mask1, 1)); + mask2 = vmovl_s8(vqrshrn_n_s16(mask2, 1)); + mask3 = vmovl_s8(vqrshrn_n_s16(mask3, 1)); + + blend_8x4(dst_tmp, dst_stride, src0_tmp, src0_stride, src1_tmp, + src1_stride, mask0, mask1, mask2, mask3, v_maxval, + vec_offset, vec_round_bits); + + w_tmp -= 8; + mask_tmp += 8; + dst_tmp += 8; + src0_tmp += 8; + src1_tmp += 8; + } while (w_tmp > 7); + i += 4; + mask_tmp += (8 * mask_stride) - w; + dst_tmp += (4 * dst_stride) - w; + src0_tmp += (4 * src0_stride) - w; + src1_tmp += (4 * src1_stride) - w; + } while (i < h); + } else { + do { + load_unaligned_u8_4x4(mask_tmp, 2 * mask_stride, &tu0, &tu1); + load_unaligned_u8_4x4(mask_tmp + mask_stride, 2 * mask_stride, &tu2, + &tu3); + + s0 = vreinterpret_u8_u32(tu0); + s1 = vreinterpret_u8_u32(tu1); + s2 = vreinterpret_u8_u32(tu2); + s3 = vreinterpret_u8_u32(tu3); + + mask0 = vreinterpretq_s16_u16(vaddl_u8(s0, s2)); + mask1 = vreinterpretq_s16_u16(vaddl_u8(s1, s3)); + + mask0 = vmovl_s8(vqrshrn_n_s16(mask0, 1)); + mask1 = vmovl_s8(vqrshrn_n_s16(mask1, 1)); + + mask0_low = vget_low_s16(mask0); + mask1_low = vget_high_s16(mask0); + mask2_low = vget_low_s16(mask1); + mask3_low = vget_high_s16(mask1); + + blend_4x4(dst_tmp, dst_stride, src0_tmp, src0_stride, src1_tmp, + src1_stride, mask0_low, mask1_low, mask2_low, mask3_low, + v_maxval, vec_offset, vec_round_bits); + + i += 4; + mask_tmp += (8 * mask_stride); + dst_tmp += (4 * dst_stride); + src0_tmp += (4 * src0_stride); + src1_tmp += (4 * src1_stride); + } while (i < h); + } + } +} diff --git a/third_party/aom/aom_dsp/arm/fwd_txfm_neon.c b/third_party/aom/aom_dsp/arm/fwd_txfm_neon.c new file mode 100644 index 0000000000..e4300c9920 --- /dev/null +++ b/third_party/aom/aom_dsp/arm/fwd_txfm_neon.c @@ -0,0 +1,222 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include + +#include "config/aom_config.h" + +#include "aom_dsp/txfm_common.h" + +void aom_fdct8x8_neon(const int16_t *input, int16_t *final_output, int stride) { + int i; + // stage 1 + int16x8_t input_0 = vshlq_n_s16(vld1q_s16(&input[0 * stride]), 2); + int16x8_t input_1 = vshlq_n_s16(vld1q_s16(&input[1 * stride]), 2); + int16x8_t input_2 = vshlq_n_s16(vld1q_s16(&input[2 * stride]), 2); + int16x8_t input_3 = vshlq_n_s16(vld1q_s16(&input[3 * stride]), 2); + int16x8_t input_4 = vshlq_n_s16(vld1q_s16(&input[4 * stride]), 2); + int16x8_t input_5 = vshlq_n_s16(vld1q_s16(&input[5 * stride]), 2); + int16x8_t input_6 = vshlq_n_s16(vld1q_s16(&input[6 * stride]), 2); + int16x8_t input_7 = vshlq_n_s16(vld1q_s16(&input[7 * stride]), 2); + for (i = 0; i < 2; ++i) { + int16x8_t out_0, out_1, out_2, out_3, out_4, out_5, out_6, out_7; + const int16x8_t v_s0 = vaddq_s16(input_0, input_7); + const int16x8_t v_s1 = vaddq_s16(input_1, input_6); + const int16x8_t v_s2 = vaddq_s16(input_2, input_5); + const int16x8_t v_s3 = vaddq_s16(input_3, input_4); + const int16x8_t v_s4 = vsubq_s16(input_3, input_4); + const int16x8_t v_s5 = vsubq_s16(input_2, input_5); + const int16x8_t v_s6 = vsubq_s16(input_1, input_6); + const int16x8_t v_s7 = vsubq_s16(input_0, input_7); + // fdct4(step, step); + int16x8_t v_x0 = vaddq_s16(v_s0, v_s3); + int16x8_t v_x1 = vaddq_s16(v_s1, v_s2); + int16x8_t v_x2 = vsubq_s16(v_s1, v_s2); + int16x8_t v_x3 = vsubq_s16(v_s0, v_s3); + // fdct4(step, step); + int32x4_t v_t0_lo = vaddl_s16(vget_low_s16(v_x0), vget_low_s16(v_x1)); + int32x4_t v_t0_hi = vaddl_s16(vget_high_s16(v_x0), vget_high_s16(v_x1)); + int32x4_t v_t1_lo = vsubl_s16(vget_low_s16(v_x0), vget_low_s16(v_x1)); + int32x4_t v_t1_hi = vsubl_s16(vget_high_s16(v_x0), vget_high_s16(v_x1)); + int32x4_t v_t2_lo = vmull_n_s16(vget_low_s16(v_x2), (int16_t)cospi_24_64); + int32x4_t v_t2_hi = vmull_n_s16(vget_high_s16(v_x2), (int16_t)cospi_24_64); + int32x4_t v_t3_lo = vmull_n_s16(vget_low_s16(v_x3), (int16_t)cospi_24_64); + int32x4_t v_t3_hi = vmull_n_s16(vget_high_s16(v_x3), (int16_t)cospi_24_64); + v_t2_lo = vmlal_n_s16(v_t2_lo, vget_low_s16(v_x3), (int16_t)cospi_8_64); + v_t2_hi = vmlal_n_s16(v_t2_hi, vget_high_s16(v_x3), (int16_t)cospi_8_64); + v_t3_lo = vmlsl_n_s16(v_t3_lo, vget_low_s16(v_x2), (int16_t)cospi_8_64); + v_t3_hi = vmlsl_n_s16(v_t3_hi, vget_high_s16(v_x2), (int16_t)cospi_8_64); + v_t0_lo = vmulq_n_s32(v_t0_lo, (int32_t)cospi_16_64); + v_t0_hi = vmulq_n_s32(v_t0_hi, (int32_t)cospi_16_64); + v_t1_lo = vmulq_n_s32(v_t1_lo, (int32_t)cospi_16_64); + v_t1_hi = vmulq_n_s32(v_t1_hi, (int32_t)cospi_16_64); + { + const int16x4_t a = vrshrn_n_s32(v_t0_lo, DCT_CONST_BITS); + const int16x4_t b = vrshrn_n_s32(v_t0_hi, DCT_CONST_BITS); + const int16x4_t c = vrshrn_n_s32(v_t1_lo, DCT_CONST_BITS); + const int16x4_t d = vrshrn_n_s32(v_t1_hi, DCT_CONST_BITS); + const int16x4_t e = vrshrn_n_s32(v_t2_lo, DCT_CONST_BITS); + const int16x4_t f = vrshrn_n_s32(v_t2_hi, DCT_CONST_BITS); + const int16x4_t g = vrshrn_n_s32(v_t3_lo, DCT_CONST_BITS); + const int16x4_t h = vrshrn_n_s32(v_t3_hi, DCT_CONST_BITS); + out_0 = vcombine_s16(a, c); // 00 01 02 03 40 41 42 43 + out_2 = vcombine_s16(e, g); // 20 21 22 23 60 61 62 63 + out_4 = vcombine_s16(b, d); // 04 05 06 07 44 45 46 47 + out_6 = vcombine_s16(f, h); // 24 25 26 27 64 65 66 67 + } + // Stage 2 + v_x0 = vsubq_s16(v_s6, v_s5); + v_x1 = vaddq_s16(v_s6, v_s5); + v_t0_lo = vmull_n_s16(vget_low_s16(v_x0), (int16_t)cospi_16_64); + v_t0_hi = vmull_n_s16(vget_high_s16(v_x0), (int16_t)cospi_16_64); + v_t1_lo = vmull_n_s16(vget_low_s16(v_x1), (int16_t)cospi_16_64); + v_t1_hi = vmull_n_s16(vget_high_s16(v_x1), (int16_t)cospi_16_64); + { + const int16x4_t a = vrshrn_n_s32(v_t0_lo, DCT_CONST_BITS); + const int16x4_t b = vrshrn_n_s32(v_t0_hi, DCT_CONST_BITS); + const int16x4_t c = vrshrn_n_s32(v_t1_lo, DCT_CONST_BITS); + const int16x4_t d = vrshrn_n_s32(v_t1_hi, DCT_CONST_BITS); + const int16x8_t ab = vcombine_s16(a, b); + const int16x8_t cd = vcombine_s16(c, d); + // Stage 3 + v_x0 = vaddq_s16(v_s4, ab); + v_x1 = vsubq_s16(v_s4, ab); + v_x2 = vsubq_s16(v_s7, cd); + v_x3 = vaddq_s16(v_s7, cd); + } + // Stage 4 + v_t0_lo = vmull_n_s16(vget_low_s16(v_x3), (int16_t)cospi_4_64); + v_t0_hi = vmull_n_s16(vget_high_s16(v_x3), (int16_t)cospi_4_64); + v_t0_lo = vmlal_n_s16(v_t0_lo, vget_low_s16(v_x0), (int16_t)cospi_28_64); + v_t0_hi = vmlal_n_s16(v_t0_hi, vget_high_s16(v_x0), (int16_t)cospi_28_64); + v_t1_lo = vmull_n_s16(vget_low_s16(v_x1), (int16_t)cospi_12_64); + v_t1_hi = vmull_n_s16(vget_high_s16(v_x1), (int16_t)cospi_12_64); + v_t1_lo = vmlal_n_s16(v_t1_lo, vget_low_s16(v_x2), (int16_t)cospi_20_64); + v_t1_hi = vmlal_n_s16(v_t1_hi, vget_high_s16(v_x2), (int16_t)cospi_20_64); + v_t2_lo = vmull_n_s16(vget_low_s16(v_x2), (int16_t)cospi_12_64); + v_t2_hi = vmull_n_s16(vget_high_s16(v_x2), (int16_t)cospi_12_64); + v_t2_lo = vmlsl_n_s16(v_t2_lo, vget_low_s16(v_x1), (int16_t)cospi_20_64); + v_t2_hi = vmlsl_n_s16(v_t2_hi, vget_high_s16(v_x1), (int16_t)cospi_20_64); + v_t3_lo = vmull_n_s16(vget_low_s16(v_x3), (int16_t)cospi_28_64); + v_t3_hi = vmull_n_s16(vget_high_s16(v_x3), (int16_t)cospi_28_64); + v_t3_lo = vmlsl_n_s16(v_t3_lo, vget_low_s16(v_x0), (int16_t)cospi_4_64); + v_t3_hi = vmlsl_n_s16(v_t3_hi, vget_high_s16(v_x0), (int16_t)cospi_4_64); + { + const int16x4_t a = vrshrn_n_s32(v_t0_lo, DCT_CONST_BITS); + const int16x4_t b = vrshrn_n_s32(v_t0_hi, DCT_CONST_BITS); + const int16x4_t c = vrshrn_n_s32(v_t1_lo, DCT_CONST_BITS); + const int16x4_t d = vrshrn_n_s32(v_t1_hi, DCT_CONST_BITS); + const int16x4_t e = vrshrn_n_s32(v_t2_lo, DCT_CONST_BITS); + const int16x4_t f = vrshrn_n_s32(v_t2_hi, DCT_CONST_BITS); + const int16x4_t g = vrshrn_n_s32(v_t3_lo, DCT_CONST_BITS); + const int16x4_t h = vrshrn_n_s32(v_t3_hi, DCT_CONST_BITS); + out_1 = vcombine_s16(a, c); // 10 11 12 13 50 51 52 53 + out_3 = vcombine_s16(e, g); // 30 31 32 33 70 71 72 73 + out_5 = vcombine_s16(b, d); // 14 15 16 17 54 55 56 57 + out_7 = vcombine_s16(f, h); // 34 35 36 37 74 75 76 77 + } + // transpose 8x8 + { + // 00 01 02 03 40 41 42 43 + // 10 11 12 13 50 51 52 53 + // 20 21 22 23 60 61 62 63 + // 30 31 32 33 70 71 72 73 + // 04 05 06 07 44 45 46 47 + // 14 15 16 17 54 55 56 57 + // 24 25 26 27 64 65 66 67 + // 34 35 36 37 74 75 76 77 + const int32x4x2_t r02_s32 = + vtrnq_s32(vreinterpretq_s32_s16(out_0), vreinterpretq_s32_s16(out_2)); + const int32x4x2_t r13_s32 = + vtrnq_s32(vreinterpretq_s32_s16(out_1), vreinterpretq_s32_s16(out_3)); + const int32x4x2_t r46_s32 = + vtrnq_s32(vreinterpretq_s32_s16(out_4), vreinterpretq_s32_s16(out_6)); + const int32x4x2_t r57_s32 = + vtrnq_s32(vreinterpretq_s32_s16(out_5), vreinterpretq_s32_s16(out_7)); + const int16x8x2_t r01_s16 = + vtrnq_s16(vreinterpretq_s16_s32(r02_s32.val[0]), + vreinterpretq_s16_s32(r13_s32.val[0])); + const int16x8x2_t r23_s16 = + vtrnq_s16(vreinterpretq_s16_s32(r02_s32.val[1]), + vreinterpretq_s16_s32(r13_s32.val[1])); + const int16x8x2_t r45_s16 = + vtrnq_s16(vreinterpretq_s16_s32(r46_s32.val[0]), + vreinterpretq_s16_s32(r57_s32.val[0])); + const int16x8x2_t r67_s16 = + vtrnq_s16(vreinterpretq_s16_s32(r46_s32.val[1]), + vreinterpretq_s16_s32(r57_s32.val[1])); + input_0 = r01_s16.val[0]; + input_1 = r01_s16.val[1]; + input_2 = r23_s16.val[0]; + input_3 = r23_s16.val[1]; + input_4 = r45_s16.val[0]; + input_5 = r45_s16.val[1]; + input_6 = r67_s16.val[0]; + input_7 = r67_s16.val[1]; + // 00 10 20 30 40 50 60 70 + // 01 11 21 31 41 51 61 71 + // 02 12 22 32 42 52 62 72 + // 03 13 23 33 43 53 63 73 + // 04 14 24 34 44 54 64 74 + // 05 15 25 35 45 55 65 75 + // 06 16 26 36 46 56 66 76 + // 07 17 27 37 47 57 67 77 + } + } // for + { + // from aom_dct_sse2.c + // Post-condition (division by two) + // division of two 16 bits signed numbers using shifts + // n / 2 = (n - (n >> 15)) >> 1 + const int16x8_t sign_in0 = vshrq_n_s16(input_0, 15); + const int16x8_t sign_in1 = vshrq_n_s16(input_1, 15); + const int16x8_t sign_in2 = vshrq_n_s16(input_2, 15); + const int16x8_t sign_in3 = vshrq_n_s16(input_3, 15); + const int16x8_t sign_in4 = vshrq_n_s16(input_4, 15); + const int16x8_t sign_in5 = vshrq_n_s16(input_5, 15); + const int16x8_t sign_in6 = vshrq_n_s16(input_6, 15); + const int16x8_t sign_in7 = vshrq_n_s16(input_7, 15); + input_0 = vhsubq_s16(input_0, sign_in0); + input_1 = vhsubq_s16(input_1, sign_in1); + input_2 = vhsubq_s16(input_2, sign_in2); + input_3 = vhsubq_s16(input_3, sign_in3); + input_4 = vhsubq_s16(input_4, sign_in4); + input_5 = vhsubq_s16(input_5, sign_in5); + input_6 = vhsubq_s16(input_6, sign_in6); + input_7 = vhsubq_s16(input_7, sign_in7); + // store results + vst1q_s16(&final_output[0 * 8], input_0); + vst1q_s16(&final_output[1 * 8], input_1); + vst1q_s16(&final_output[2 * 8], input_2); + vst1q_s16(&final_output[3 * 8], input_3); + vst1q_s16(&final_output[4 * 8], input_4); + vst1q_s16(&final_output[5 * 8], input_5); + vst1q_s16(&final_output[6 * 8], input_6); + vst1q_s16(&final_output[7 * 8], input_7); + } +} + +void aom_fdct8x8_1_neon(const int16_t *input, int16_t *output, int stride) { + int r; + int16x8_t sum = vld1q_s16(&input[0]); + for (r = 1; r < 8; ++r) { + const int16x8_t input_00 = vld1q_s16(&input[r * stride]); + sum = vaddq_s16(sum, input_00); + } + { + const int32x4_t a = vpaddlq_s16(sum); + const int64x2_t b = vpaddlq_s32(a); + const int32x2_t c = vadd_s32(vreinterpret_s32_s64(vget_low_s64(b)), + vreinterpret_s32_s64(vget_high_s64(b))); + output[0] = vget_lane_s16(vreinterpret_s16_s32(c), 0); + output[1] = 0; + } +} diff --git a/third_party/aom/aom_dsp/arm/intrapred_neon.c b/third_party/aom/aom_dsp/arm/intrapred_neon.c new file mode 100644 index 0000000000..c85b1e9100 --- /dev/null +++ b/third_party/aom/aom_dsp/arm/intrapred_neon.c @@ -0,0 +1,590 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include + +#include "config/aom_config.h" +#include "config/aom_dsp_rtcd.h" + +#include "aom/aom_integer.h" + +//------------------------------------------------------------------------------ +// DC 4x4 + +// 'do_above' and 'do_left' facilitate branch removal when inlined. +static INLINE void dc_4x4(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, + const uint8_t *left, int do_above, int do_left) { + uint16x8_t sum_top; + uint16x8_t sum_left; + uint8x8_t dc0; + + if (do_above) { + const uint8x8_t A = vld1_u8(above); // top row + const uint16x4_t p0 = vpaddl_u8(A); // cascading summation of the top + const uint16x4_t p1 = vpadd_u16(p0, p0); + sum_top = vcombine_u16(p1, p1); + } + + if (do_left) { + const uint8x8_t L = vld1_u8(left); // left border + const uint16x4_t p0 = vpaddl_u8(L); // cascading summation of the left + const uint16x4_t p1 = vpadd_u16(p0, p0); + sum_left = vcombine_u16(p1, p1); + } + + if (do_above && do_left) { + const uint16x8_t sum = vaddq_u16(sum_left, sum_top); + dc0 = vrshrn_n_u16(sum, 3); + } else if (do_above) { + dc0 = vrshrn_n_u16(sum_top, 2); + } else if (do_left) { + dc0 = vrshrn_n_u16(sum_left, 2); + } else { + dc0 = vdup_n_u8(0x80); + } + + { + const uint8x8_t dc = vdup_lane_u8(dc0, 0); + int i; + for (i = 0; i < 4; ++i) { + vst1_lane_u32((uint32_t *)(dst + i * stride), vreinterpret_u32_u8(dc), 0); + } + } +} + +void aom_dc_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + dc_4x4(dst, stride, above, left, 1, 1); +} + +void aom_dc_left_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + (void)above; + dc_4x4(dst, stride, NULL, left, 0, 1); +} + +void aom_dc_top_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + (void)left; + dc_4x4(dst, stride, above, NULL, 1, 0); +} + +void aom_dc_128_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + (void)above; + (void)left; + dc_4x4(dst, stride, NULL, NULL, 0, 0); +} + +//------------------------------------------------------------------------------ +// DC 8x8 + +// 'do_above' and 'do_left' facilitate branch removal when inlined. +static INLINE void dc_8x8(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, + const uint8_t *left, int do_above, int do_left) { + uint16x8_t sum_top; + uint16x8_t sum_left; + uint8x8_t dc0; + + if (do_above) { + const uint8x8_t A = vld1_u8(above); // top row + const uint16x4_t p0 = vpaddl_u8(A); // cascading summation of the top + const uint16x4_t p1 = vpadd_u16(p0, p0); + const uint16x4_t p2 = vpadd_u16(p1, p1); + sum_top = vcombine_u16(p2, p2); + } + + if (do_left) { + const uint8x8_t L = vld1_u8(left); // left border + const uint16x4_t p0 = vpaddl_u8(L); // cascading summation of the left + const uint16x4_t p1 = vpadd_u16(p0, p0); + const uint16x4_t p2 = vpadd_u16(p1, p1); + sum_left = vcombine_u16(p2, p2); + } + + if (do_above && do_left) { + const uint16x8_t sum = vaddq_u16(sum_left, sum_top); + dc0 = vrshrn_n_u16(sum, 4); + } else if (do_above) { + dc0 = vrshrn_n_u16(sum_top, 3); + } else if (do_left) { + dc0 = vrshrn_n_u16(sum_left, 3); + } else { + dc0 = vdup_n_u8(0x80); + } + + { + const uint8x8_t dc = vdup_lane_u8(dc0, 0); + int i; + for (i = 0; i < 8; ++i) { + vst1_u32((uint32_t *)(dst + i * stride), vreinterpret_u32_u8(dc)); + } + } +} + +void aom_dc_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + dc_8x8(dst, stride, above, left, 1, 1); +} + +void aom_dc_left_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + (void)above; + dc_8x8(dst, stride, NULL, left, 0, 1); +} + +void aom_dc_top_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + (void)left; + dc_8x8(dst, stride, above, NULL, 1, 0); +} + +void aom_dc_128_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + (void)above; + (void)left; + dc_8x8(dst, stride, NULL, NULL, 0, 0); +} + +//------------------------------------------------------------------------------ +// DC 16x16 + +// 'do_above' and 'do_left' facilitate branch removal when inlined. +static INLINE void dc_16x16(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left, + int do_above, int do_left) { + uint16x8_t sum_top; + uint16x8_t sum_left; + uint8x8_t dc0; + + if (do_above) { + const uint8x16_t A = vld1q_u8(above); // top row + const uint16x8_t p0 = vpaddlq_u8(A); // cascading summation of the top + const uint16x4_t p1 = vadd_u16(vget_low_u16(p0), vget_high_u16(p0)); + const uint16x4_t p2 = vpadd_u16(p1, p1); + const uint16x4_t p3 = vpadd_u16(p2, p2); + sum_top = vcombine_u16(p3, p3); + } + + if (do_left) { + const uint8x16_t L = vld1q_u8(left); // left row + const uint16x8_t p0 = vpaddlq_u8(L); // cascading summation of the left + const uint16x4_t p1 = vadd_u16(vget_low_u16(p0), vget_high_u16(p0)); + const uint16x4_t p2 = vpadd_u16(p1, p1); + const uint16x4_t p3 = vpadd_u16(p2, p2); + sum_left = vcombine_u16(p3, p3); + } + + if (do_above && do_left) { + const uint16x8_t sum = vaddq_u16(sum_left, sum_top); + dc0 = vrshrn_n_u16(sum, 5); + } else if (do_above) { + dc0 = vrshrn_n_u16(sum_top, 4); + } else if (do_left) { + dc0 = vrshrn_n_u16(sum_left, 4); + } else { + dc0 = vdup_n_u8(0x80); + } + + { + const uint8x16_t dc = vdupq_lane_u8(dc0, 0); + int i; + for (i = 0; i < 16; ++i) { + vst1q_u8(dst + i * stride, dc); + } + } +} + +void aom_dc_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + dc_16x16(dst, stride, above, left, 1, 1); +} + +void aom_dc_left_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + (void)above; + dc_16x16(dst, stride, NULL, left, 0, 1); +} + +void aom_dc_top_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + (void)left; + dc_16x16(dst, stride, above, NULL, 1, 0); +} + +void aom_dc_128_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + (void)above; + (void)left; + dc_16x16(dst, stride, NULL, NULL, 0, 0); +} + +//------------------------------------------------------------------------------ +// DC 32x32 + +// 'do_above' and 'do_left' facilitate branch removal when inlined. +static INLINE void dc_32x32(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left, + int do_above, int do_left) { + uint16x8_t sum_top; + uint16x8_t sum_left; + uint8x8_t dc0; + + if (do_above) { + const uint8x16_t A0 = vld1q_u8(above); // top row + const uint8x16_t A1 = vld1q_u8(above + 16); + const uint16x8_t p0 = vpaddlq_u8(A0); // cascading summation of the top + const uint16x8_t p1 = vpaddlq_u8(A1); + const uint16x8_t p2 = vaddq_u16(p0, p1); + const uint16x4_t p3 = vadd_u16(vget_low_u16(p2), vget_high_u16(p2)); + const uint16x4_t p4 = vpadd_u16(p3, p3); + const uint16x4_t p5 = vpadd_u16(p4, p4); + sum_top = vcombine_u16(p5, p5); + } + + if (do_left) { + const uint8x16_t L0 = vld1q_u8(left); // left row + const uint8x16_t L1 = vld1q_u8(left + 16); + const uint16x8_t p0 = vpaddlq_u8(L0); // cascading summation of the left + const uint16x8_t p1 = vpaddlq_u8(L1); + const uint16x8_t p2 = vaddq_u16(p0, p1); + const uint16x4_t p3 = vadd_u16(vget_low_u16(p2), vget_high_u16(p2)); + const uint16x4_t p4 = vpadd_u16(p3, p3); + const uint16x4_t p5 = vpadd_u16(p4, p4); + sum_left = vcombine_u16(p5, p5); + } + + if (do_above && do_left) { + const uint16x8_t sum = vaddq_u16(sum_left, sum_top); + dc0 = vrshrn_n_u16(sum, 6); + } else if (do_above) { + dc0 = vrshrn_n_u16(sum_top, 5); + } else if (do_left) { + dc0 = vrshrn_n_u16(sum_left, 5); + } else { + dc0 = vdup_n_u8(0x80); + } + + { + const uint8x16_t dc = vdupq_lane_u8(dc0, 0); + int i; + for (i = 0; i < 32; ++i) { + vst1q_u8(dst + i * stride, dc); + vst1q_u8(dst + i * stride + 16, dc); + } + } +} + +void aom_dc_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + dc_32x32(dst, stride, above, left, 1, 1); +} + +void aom_dc_left_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + (void)above; + dc_32x32(dst, stride, NULL, left, 0, 1); +} + +void aom_dc_top_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + (void)left; + dc_32x32(dst, stride, above, NULL, 1, 0); +} + +void aom_dc_128_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + (void)above; + (void)left; + dc_32x32(dst, stride, NULL, NULL, 0, 0); +} + +// ----------------------------------------------------------------------------- + +void aom_d135_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + const uint8x8_t XABCD_u8 = vld1_u8(above - 1); + const uint64x1_t XABCD = vreinterpret_u64_u8(XABCD_u8); + const uint64x1_t ____XABC = vshl_n_u64(XABCD, 32); + const uint32x2_t zero = vdup_n_u32(0); + const uint32x2_t IJKL = vld1_lane_u32((const uint32_t *)left, zero, 0); + const uint8x8_t IJKL_u8 = vreinterpret_u8_u32(IJKL); + const uint64x1_t LKJI____ = vreinterpret_u64_u8(vrev32_u8(IJKL_u8)); + const uint64x1_t LKJIXABC = vorr_u64(LKJI____, ____XABC); + const uint8x8_t KJIXABC_ = vreinterpret_u8_u64(vshr_n_u64(LKJIXABC, 8)); + const uint8x8_t JIXABC__ = vreinterpret_u8_u64(vshr_n_u64(LKJIXABC, 16)); + const uint8_t D = vget_lane_u8(XABCD_u8, 4); + const uint8x8_t JIXABCD_ = vset_lane_u8(D, JIXABC__, 6); + const uint8x8_t LKJIXABC_u8 = vreinterpret_u8_u64(LKJIXABC); + const uint8x8_t avg1 = vhadd_u8(JIXABCD_, LKJIXABC_u8); + const uint8x8_t avg2 = vrhadd_u8(avg1, KJIXABC_); + const uint64x1_t avg2_u64 = vreinterpret_u64_u8(avg2); + const uint32x2_t r3 = vreinterpret_u32_u8(avg2); + const uint32x2_t r2 = vreinterpret_u32_u64(vshr_n_u64(avg2_u64, 8)); + const uint32x2_t r1 = vreinterpret_u32_u64(vshr_n_u64(avg2_u64, 16)); + const uint32x2_t r0 = vreinterpret_u32_u64(vshr_n_u64(avg2_u64, 24)); + vst1_lane_u32((uint32_t *)(dst + 0 * stride), r0, 0); + vst1_lane_u32((uint32_t *)(dst + 1 * stride), r1, 0); + vst1_lane_u32((uint32_t *)(dst + 2 * stride), r2, 0); + vst1_lane_u32((uint32_t *)(dst + 3 * stride), r3, 0); +} + +void aom_v_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + int i; + uint32x2_t d0u32 = vdup_n_u32(0); + (void)left; + + d0u32 = vld1_lane_u32((const uint32_t *)above, d0u32, 0); + for (i = 0; i < 4; i++, dst += stride) + vst1_lane_u32((uint32_t *)dst, d0u32, 0); +} + +void aom_v_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + int i; + uint8x8_t d0u8 = vdup_n_u8(0); + (void)left; + + d0u8 = vld1_u8(above); + for (i = 0; i < 8; i++, dst += stride) vst1_u8(dst, d0u8); +} + +void aom_v_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + int i; + uint8x16_t q0u8 = vdupq_n_u8(0); + (void)left; + + q0u8 = vld1q_u8(above); + for (i = 0; i < 16; i++, dst += stride) vst1q_u8(dst, q0u8); +} + +void aom_v_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + int i; + uint8x16_t q0u8 = vdupq_n_u8(0); + uint8x16_t q1u8 = vdupq_n_u8(0); + (void)left; + + q0u8 = vld1q_u8(above); + q1u8 = vld1q_u8(above + 16); + for (i = 0; i < 32; i++, dst += stride) { + vst1q_u8(dst, q0u8); + vst1q_u8(dst + 16, q1u8); + } +} + +void aom_h_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + uint8x8_t d0u8 = vdup_n_u8(0); + uint32x2_t d1u32 = vdup_n_u32(0); + (void)above; + + d1u32 = vld1_lane_u32((const uint32_t *)left, d1u32, 0); + + d0u8 = vdup_lane_u8(vreinterpret_u8_u32(d1u32), 0); + vst1_lane_u32((uint32_t *)dst, vreinterpret_u32_u8(d0u8), 0); + dst += stride; + d0u8 = vdup_lane_u8(vreinterpret_u8_u32(d1u32), 1); + vst1_lane_u32((uint32_t *)dst, vreinterpret_u32_u8(d0u8), 0); + dst += stride; + d0u8 = vdup_lane_u8(vreinterpret_u8_u32(d1u32), 2); + vst1_lane_u32((uint32_t *)dst, vreinterpret_u32_u8(d0u8), 0); + dst += stride; + d0u8 = vdup_lane_u8(vreinterpret_u8_u32(d1u32), 3); + vst1_lane_u32((uint32_t *)dst, vreinterpret_u32_u8(d0u8), 0); +} + +void aom_h_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + uint8x8_t d0u8 = vdup_n_u8(0); + uint64x1_t d1u64 = vdup_n_u64(0); + (void)above; + + d1u64 = vld1_u64((const uint64_t *)left); + + d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 0); + vst1_u8(dst, d0u8); + dst += stride; + d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 1); + vst1_u8(dst, d0u8); + dst += stride; + d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 2); + vst1_u8(dst, d0u8); + dst += stride; + d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 3); + vst1_u8(dst, d0u8); + dst += stride; + d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 4); + vst1_u8(dst, d0u8); + dst += stride; + d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 5); + vst1_u8(dst, d0u8); + dst += stride; + d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 6); + vst1_u8(dst, d0u8); + dst += stride; + d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 7); + vst1_u8(dst, d0u8); +} + +void aom_h_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + int j; + uint8x8_t d2u8 = vdup_n_u8(0); + uint8x16_t q0u8 = vdupq_n_u8(0); + uint8x16_t q1u8 = vdupq_n_u8(0); + (void)above; + + q1u8 = vld1q_u8(left); + d2u8 = vget_low_u8(q1u8); + for (j = 0; j < 2; j++, d2u8 = vget_high_u8(q1u8)) { + q0u8 = vdupq_lane_u8(d2u8, 0); + vst1q_u8(dst, q0u8); + dst += stride; + q0u8 = vdupq_lane_u8(d2u8, 1); + vst1q_u8(dst, q0u8); + dst += stride; + q0u8 = vdupq_lane_u8(d2u8, 2); + vst1q_u8(dst, q0u8); + dst += stride; + q0u8 = vdupq_lane_u8(d2u8, 3); + vst1q_u8(dst, q0u8); + dst += stride; + q0u8 = vdupq_lane_u8(d2u8, 4); + vst1q_u8(dst, q0u8); + dst += stride; + q0u8 = vdupq_lane_u8(d2u8, 5); + vst1q_u8(dst, q0u8); + dst += stride; + q0u8 = vdupq_lane_u8(d2u8, 6); + vst1q_u8(dst, q0u8); + dst += stride; + q0u8 = vdupq_lane_u8(d2u8, 7); + vst1q_u8(dst, q0u8); + dst += stride; + } +} + +void aom_h_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + int j, k; + uint8x8_t d2u8 = vdup_n_u8(0); + uint8x16_t q0u8 = vdupq_n_u8(0); + uint8x16_t q1u8 = vdupq_n_u8(0); + (void)above; + + for (k = 0; k < 2; k++, left += 16) { + q1u8 = vld1q_u8(left); + d2u8 = vget_low_u8(q1u8); + for (j = 0; j < 2; j++, d2u8 = vget_high_u8(q1u8)) { + q0u8 = vdupq_lane_u8(d2u8, 0); + vst1q_u8(dst, q0u8); + vst1q_u8(dst + 16, q0u8); + dst += stride; + q0u8 = vdupq_lane_u8(d2u8, 1); + vst1q_u8(dst, q0u8); + vst1q_u8(dst + 16, q0u8); + dst += stride; + q0u8 = vdupq_lane_u8(d2u8, 2); + vst1q_u8(dst, q0u8); + vst1q_u8(dst + 16, q0u8); + dst += stride; + q0u8 = vdupq_lane_u8(d2u8, 3); + vst1q_u8(dst, q0u8); + vst1q_u8(dst + 16, q0u8); + dst += stride; + q0u8 = vdupq_lane_u8(d2u8, 4); + vst1q_u8(dst, q0u8); + vst1q_u8(dst + 16, q0u8); + dst += stride; + q0u8 = vdupq_lane_u8(d2u8, 5); + vst1q_u8(dst, q0u8); + vst1q_u8(dst + 16, q0u8); + dst += stride; + q0u8 = vdupq_lane_u8(d2u8, 6); + vst1q_u8(dst, q0u8); + vst1q_u8(dst + 16, q0u8); + dst += stride; + q0u8 = vdupq_lane_u8(d2u8, 7); + vst1q_u8(dst, q0u8); + vst1q_u8(dst + 16, q0u8); + dst += stride; + } + } +} + +static INLINE void highbd_dc_predictor(uint16_t *dst, ptrdiff_t stride, int bw, + const uint16_t *above, + const uint16_t *left) { + assert(bw >= 4); + assert(IS_POWER_OF_TWO(bw)); + int expected_dc, sum = 0; + const int count = bw * 2; + uint32x4_t sum_q = vdupq_n_u32(0); + uint32x2_t sum_d; + uint16_t *dst_1; + if (bw >= 8) { + for (int i = 0; i < bw; i += 8) { + sum_q = vpadalq_u16(sum_q, vld1q_u16(above)); + sum_q = vpadalq_u16(sum_q, vld1q_u16(left)); + above += 8; + left += 8; + } + sum_d = vadd_u32(vget_low_u32(sum_q), vget_high_u32(sum_q)); + sum = vget_lane_s32(vreinterpret_s32_u64(vpaddl_u32(sum_d)), 0); + expected_dc = (sum + (count >> 1)) / count; + const uint16x8_t dc = vdupq_n_u16((uint16_t)expected_dc); + for (int r = 0; r < bw; r++) { + dst_1 = dst; + for (int i = 0; i < bw; i += 8) { + vst1q_u16(dst_1, dc); + dst_1 += 8; + } + dst += stride; + } + } else { // 4x4 + sum_q = vaddl_u16(vld1_u16(above), vld1_u16(left)); + sum_d = vadd_u32(vget_low_u32(sum_q), vget_high_u32(sum_q)); + sum = vget_lane_s32(vreinterpret_s32_u64(vpaddl_u32(sum_d)), 0); + expected_dc = (sum + (count >> 1)) / count; + const uint16x4_t dc = vdup_n_u16((uint16_t)expected_dc); + for (int r = 0; r < bw; r++) { + vst1_u16(dst, dc); + dst += stride; + } + } +} + +#define intra_pred_highbd_sized_neon(type, width) \ + void aom_highbd_##type##_predictor_##width##x##width##_neon( \ + uint16_t *dst, ptrdiff_t stride, const uint16_t *above, \ + const uint16_t *left, int bd) { \ + (void)bd; \ + highbd_##type##_predictor(dst, stride, width, above, left); \ + } + +#define intra_pred_square(type) \ + intra_pred_highbd_sized_neon(type, 4); \ + intra_pred_highbd_sized_neon(type, 8); \ + intra_pred_highbd_sized_neon(type, 16); \ + intra_pred_highbd_sized_neon(type, 32); \ + intra_pred_highbd_sized_neon(type, 64); + +intra_pred_square(dc); +#undef intra_pred_square diff --git a/third_party/aom/aom_dsp/arm/loopfilter_neon.c b/third_party/aom/aom_dsp/arm/loopfilter_neon.c new file mode 100644 index 0000000000..bdc67626d6 --- /dev/null +++ b/third_party/aom/aom_dsp/arm/loopfilter_neon.c @@ -0,0 +1,928 @@ +/* + * Copyright (c) 2018, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include + +#include "config/aom_dsp_rtcd.h" +#include "config/aom_config.h" + +#include "aom/aom_integer.h" +#include "av1/common/arm/mem_neon.h" +#include "av1/common/arm/transpose_neon.h" + +static INLINE uint8x8_t lpf_mask(uint8x8_t p3q3, uint8x8_t p2q2, uint8x8_t p1q1, + uint8x8_t p0q0, const uint8_t blimit, + const uint8_t limit) { + // Calculate mask values for four samples + uint32x2x2_t p0q0_p1q1; + uint16x8_t temp_16x8; + uint16x4_t temp0_16x4, temp1_16x4; + uint8x8_t mask_8x8, temp_8x8; + const uint8x8_t limit_8x8 = vdup_n_u8(limit); + const uint16x4_t blimit_16x4 = vdup_n_u16((uint16_t)blimit); + + mask_8x8 = vabd_u8(p3q3, p2q2); + mask_8x8 = vmax_u8(mask_8x8, vabd_u8(p2q2, p1q1)); + mask_8x8 = vmax_u8(mask_8x8, vabd_u8(p1q1, p0q0)); + mask_8x8 = vcle_u8(mask_8x8, limit_8x8); + + temp_8x8 = vreinterpret_u8_u32(vrev64_u32(vreinterpret_u32_u8(mask_8x8))); + mask_8x8 = vand_u8(mask_8x8, temp_8x8); + + p0q0_p1q1 = vtrn_u32(vreinterpret_u32_u8(p0q0), vreinterpret_u32_u8(p1q1)); + temp_8x8 = vabd_u8(vreinterpret_u8_u32(p0q0_p1q1.val[0]), + vreinterpret_u8_u32(p0q0_p1q1.val[1])); + temp_16x8 = vmovl_u8(temp_8x8); + temp0_16x4 = vshl_n_u16(vget_low_u16(temp_16x8), 1); + temp1_16x4 = vshr_n_u16(vget_high_u16(temp_16x8), 1); + temp0_16x4 = vadd_u16(temp0_16x4, temp1_16x4); + temp0_16x4 = vcle_u16(temp0_16x4, blimit_16x4); + temp_8x8 = vmovn_u16(vcombine_u16(temp0_16x4, temp0_16x4)); + + mask_8x8 = vand_u8(mask_8x8, temp_8x8); + + return mask_8x8; +} + +static INLINE uint8x8_t lpf_mask2(uint8x8_t p1q1, uint8x8_t p0q0, + const uint8_t blimit, const uint8_t limit) { + uint32x2x2_t p0q0_p1q1; + uint16x8_t temp_16x8; + uint16x4_t temp0_16x4, temp1_16x4; + const uint16x4_t blimit_16x4 = vdup_n_u16(blimit); + const uint8x8_t limit_8x8 = vdup_n_u8(limit); + uint8x8_t mask_8x8, temp_8x8; + + mask_8x8 = vabd_u8(p1q1, p0q0); + mask_8x8 = vcle_u8(mask_8x8, limit_8x8); + + temp_8x8 = vreinterpret_u8_u32(vrev64_u32(vreinterpret_u32_u8(mask_8x8))); + mask_8x8 = vand_u8(mask_8x8, temp_8x8); + + p0q0_p1q1 = vtrn_u32(vreinterpret_u32_u8(p0q0), vreinterpret_u32_u8(p1q1)); + temp_8x8 = vabd_u8(vreinterpret_u8_u32(p0q0_p1q1.val[0]), + vreinterpret_u8_u32(p0q0_p1q1.val[1])); + temp_16x8 = vmovl_u8(temp_8x8); + temp0_16x4 = vshl_n_u16(vget_low_u16(temp_16x8), 1); + temp1_16x4 = vshr_n_u16(vget_high_u16(temp_16x8), 1); + temp0_16x4 = vadd_u16(temp0_16x4, temp1_16x4); + temp0_16x4 = vcle_u16(temp0_16x4, blimit_16x4); + temp_8x8 = vmovn_u16(vcombine_u16(temp0_16x4, temp0_16x4)); + + mask_8x8 = vand_u8(mask_8x8, temp_8x8); + + return mask_8x8; +} + +static INLINE uint8x8_t lpf_flat_mask4(uint8x8_t p3q3, uint8x8_t p2q2, + uint8x8_t p1q1, uint8x8_t p0q0) { + const uint8x8_t thresh_8x8 = vdup_n_u8(1); // for bd==8 threshold is always 1 + uint8x8_t flat_8x8, temp_8x8; + + flat_8x8 = vabd_u8(p1q1, p0q0); + flat_8x8 = vmax_u8(flat_8x8, vabd_u8(p2q2, p0q0)); + flat_8x8 = vmax_u8(flat_8x8, vabd_u8(p3q3, p0q0)); + flat_8x8 = vcle_u8(flat_8x8, thresh_8x8); + + temp_8x8 = vreinterpret_u8_u32(vrev64_u32(vreinterpret_u32_u8(flat_8x8))); + flat_8x8 = vand_u8(flat_8x8, temp_8x8); + + return flat_8x8; +} + +static INLINE uint8x8_t lpf_flat_mask3(uint8x8_t p2q2, uint8x8_t p1q1, + uint8x8_t p0q0) { + const uint8x8_t thresh_8x8 = vdup_n_u8(1); // for bd==8 threshold is always 1 + uint8x8_t flat_8x8, temp_8x8; + + flat_8x8 = vabd_u8(p1q1, p0q0); + flat_8x8 = vmax_u8(flat_8x8, vabd_u8(p2q2, p0q0)); + flat_8x8 = vcle_u8(flat_8x8, thresh_8x8); + + temp_8x8 = vreinterpret_u8_u32(vrev64_u32(vreinterpret_u32_u8(flat_8x8))); + flat_8x8 = vand_u8(flat_8x8, temp_8x8); + + return flat_8x8; +} + +static INLINE uint8x8_t lpf_mask3_chroma(uint8x8_t p2q2, uint8x8_t p1q1, + uint8x8_t p0q0, const uint8_t blimit, + const uint8_t limit) { + // Calculate mask3 values for four samples + uint32x2x2_t p0q0_p1q1; + uint16x8_t temp_16x8; + uint16x4_t temp0_16x4, temp1_16x4; + uint8x8_t mask_8x8, temp_8x8; + const uint8x8_t limit_8x8 = vdup_n_u8(limit); + const uint16x4_t blimit_16x4 = vdup_n_u16((uint16_t)blimit); + + mask_8x8 = vabd_u8(p2q2, p1q1); + mask_8x8 = vmax_u8(mask_8x8, vabd_u8(p1q1, p0q0)); + mask_8x8 = vcle_u8(mask_8x8, limit_8x8); + + temp_8x8 = vreinterpret_u8_u32(vrev64_u32(vreinterpret_u32_u8(mask_8x8))); + mask_8x8 = vand_u8(mask_8x8, temp_8x8); + + p0q0_p1q1 = vtrn_u32(vreinterpret_u32_u8(p0q0), vreinterpret_u32_u8(p1q1)); + temp_8x8 = vabd_u8(vreinterpret_u8_u32(p0q0_p1q1.val[0]), + vreinterpret_u8_u32(p0q0_p1q1.val[1])); + temp_16x8 = vmovl_u8(temp_8x8); + temp0_16x4 = vshl_n_u16(vget_low_u16(temp_16x8), 1); + temp1_16x4 = vshr_n_u16(vget_high_u16(temp_16x8), 1); + temp0_16x4 = vadd_u16(temp0_16x4, temp1_16x4); + temp0_16x4 = vcle_u16(temp0_16x4, blimit_16x4); + temp_8x8 = vmovn_u16(vcombine_u16(temp0_16x4, temp0_16x4)); + + mask_8x8 = vand_u8(mask_8x8, temp_8x8); + + return mask_8x8; +} + +static void lpf_14_neon(uint8x8_t *p6q6, uint8x8_t *p5q5, uint8x8_t *p4q4, + uint8x8_t *p3q3, uint8x8_t *p2q2, uint8x8_t *p1q1, + uint8x8_t *p0q0, const uint8_t blimit, + const uint8_t limit, const uint8_t thresh) { + uint16x8_t out; + uint8x8_t out_f14_pq0, out_f14_pq1, out_f14_pq2, out_f14_pq3, out_f14_pq4, + out_f14_pq5; + uint8x8_t out_f7_pq0, out_f7_pq1, out_f7_pq2; + uint8x8_t out_f4_pq0, out_f4_pq1; + uint8x8_t mask_8x8, flat_8x8, flat2_8x8; + uint8x8_t q0p0, q1p1, q2p2; + + // Calculate filter masks + mask_8x8 = lpf_mask(*p3q3, *p2q2, *p1q1, *p0q0, blimit, limit); + flat_8x8 = lpf_flat_mask4(*p3q3, *p2q2, *p1q1, *p0q0); + flat2_8x8 = lpf_flat_mask4(*p6q6, *p5q5, *p4q4, *p0q0); + { + // filter 4 + int32x2x2_t ps0_qs0, ps1_qs1; + int16x8_t filter_s16; + const uint8x8_t thresh_f4 = vdup_n_u8(thresh); + uint8x8_t temp0_8x8, temp1_8x8; + int8x8_t ps0_s8, ps1_s8, qs0_s8, qs1_s8, temp_s8; + int8x8_t op0, oq0, op1, oq1; + int8x8_t pq_s0, pq_s1; + int8x8_t filter_s8, filter1_s8, filter2_s8; + int8x8_t hev_8x8; + const int8x8_t sign_mask = vdup_n_s8(0x80); + const int8x8_t val_4 = vdup_n_s8(4); + const int8x8_t val_3 = vdup_n_s8(3); + + pq_s0 = veor_s8(vreinterpret_s8_u8(*p0q0), sign_mask); + pq_s1 = veor_s8(vreinterpret_s8_u8(*p1q1), sign_mask); + + ps0_qs0 = vtrn_s32(vreinterpret_s32_s8(pq_s0), vreinterpret_s32_s8(pq_s0)); + ps1_qs1 = vtrn_s32(vreinterpret_s32_s8(pq_s1), vreinterpret_s32_s8(pq_s1)); + ps0_s8 = vreinterpret_s8_s32(ps0_qs0.val[0]); + qs0_s8 = vreinterpret_s8_s32(ps0_qs0.val[1]); + ps1_s8 = vreinterpret_s8_s32(ps1_qs1.val[0]); + qs1_s8 = vreinterpret_s8_s32(ps1_qs1.val[1]); + + // hev_mask + temp0_8x8 = vcgt_u8(vabd_u8(*p0q0, *p1q1), thresh_f4); + temp1_8x8 = vreinterpret_u8_u32(vrev64_u32(vreinterpret_u32_u8(temp0_8x8))); + hev_8x8 = vreinterpret_s8_u8(vorr_u8(temp0_8x8, temp1_8x8)); + + // add outer taps if we have high edge variance + filter_s8 = vqsub_s8(ps1_s8, qs1_s8); + filter_s8 = vand_s8(filter_s8, hev_8x8); + + // inner taps + temp_s8 = vqsub_s8(qs0_s8, ps0_s8); + filter_s16 = vmovl_s8(filter_s8); + filter_s16 = vmlal_s8(filter_s16, temp_s8, val_3); + filter_s8 = vqmovn_s16(filter_s16); + filter_s8 = vand_s8(filter_s8, vreinterpret_s8_u8(mask_8x8)); + + filter1_s8 = vqadd_s8(filter_s8, val_4); + filter2_s8 = vqadd_s8(filter_s8, val_3); + filter1_s8 = vshr_n_s8(filter1_s8, 3); + filter2_s8 = vshr_n_s8(filter2_s8, 3); + + oq0 = veor_s8(vqsub_s8(qs0_s8, filter1_s8), sign_mask); + op0 = veor_s8(vqadd_s8(ps0_s8, filter2_s8), sign_mask); + + hev_8x8 = vmvn_s8(hev_8x8); + filter_s8 = vrshr_n_s8(filter1_s8, 1); + filter_s8 = vand_s8(filter_s8, hev_8x8); + + oq1 = veor_s8(vqsub_s8(qs1_s8, filter_s8), sign_mask); + op1 = veor_s8(vqadd_s8(ps1_s8, filter_s8), sign_mask); + + out_f4_pq0 = vreinterpret_u8_s8(vext_s8(op0, oq0, 4)); + out_f4_pq1 = vreinterpret_u8_s8(vext_s8(op1, oq1, 4)); + } + // reverse p and q + q0p0 = vreinterpret_u8_u32(vrev64_u32(vreinterpret_u32_u8(*p0q0))); + q1p1 = vreinterpret_u8_u32(vrev64_u32(vreinterpret_u32_u8(*p1q1))); + q2p2 = vreinterpret_u8_u32(vrev64_u32(vreinterpret_u32_u8(*p2q2))); + { + // filter 8 + uint16x8_t out_pq0, out_pq1, out_pq2; + out = vaddl_u8(*p3q3, *p2q2); + out = vaddw_u8(out, *p1q1); + out = vaddw_u8(out, *p0q0); + + out = vaddw_u8(out, q0p0); + out_pq1 = vaddw_u8(out, *p3q3); + out_pq2 = vaddw_u8(out_pq1, *p3q3); + out_pq2 = vaddw_u8(out_pq2, *p2q2); + out_pq1 = vaddw_u8(out_pq1, *p1q1); + out_pq1 = vaddw_u8(out_pq1, q1p1); + + out_pq0 = vaddw_u8(out, *p0q0); + out_pq0 = vaddw_u8(out_pq0, q1p1); + out_pq0 = vaddw_u8(out_pq0, q2p2); + + out_f7_pq0 = vrshrn_n_u16(out_pq0, 3); + out_f7_pq1 = vrshrn_n_u16(out_pq1, 3); + out_f7_pq2 = vrshrn_n_u16(out_pq2, 3); + } + { + // filter 14 + uint16x8_t out_pq0, out_pq1, out_pq2, out_pq3, out_pq4, out_pq5; + uint16x8_t p6q6_2, p6q6_temp, qp_sum; + uint8x8_t qp_rev; + + out = vaddw_u8(out, *p4q4); + out = vaddw_u8(out, *p5q5); + out = vaddw_u8(out, *p6q6); + + out_pq5 = vaddw_u8(out, *p4q4); + out_pq4 = vaddw_u8(out_pq5, *p3q3); + out_pq3 = vaddw_u8(out_pq4, *p2q2); + + out_pq5 = vaddw_u8(out_pq5, *p5q5); + out_pq4 = vaddw_u8(out_pq4, *p5q5); + + out_pq0 = vaddw_u8(out, *p1q1); + out_pq1 = vaddw_u8(out_pq0, *p2q2); + out_pq2 = vaddw_u8(out_pq1, *p3q3); + + out_pq0 = vaddw_u8(out_pq0, *p0q0); + out_pq1 = vaddw_u8(out_pq1, *p0q0); + + out_pq1 = vaddw_u8(out_pq1, *p6q6); + p6q6_2 = vaddl_u8(*p6q6, *p6q6); + out_pq2 = vaddq_u16(out_pq2, p6q6_2); + p6q6_temp = vaddw_u8(p6q6_2, *p6q6); + out_pq3 = vaddq_u16(out_pq3, p6q6_temp); + p6q6_temp = vaddw_u8(p6q6_temp, *p6q6); + out_pq4 = vaddq_u16(out_pq4, p6q6_temp); + p6q6_temp = vaddq_u16(p6q6_temp, p6q6_2); + out_pq5 = vaddq_u16(out_pq5, p6q6_temp); + + out_pq4 = vaddw_u8(out_pq4, q1p1); + + qp_sum = vaddl_u8(q2p2, q1p1); + out_pq3 = vaddq_u16(out_pq3, qp_sum); + + qp_rev = vreinterpret_u8_u32(vrev64_u32(vreinterpret_u32_u8(*p3q3))); + qp_sum = vaddw_u8(qp_sum, qp_rev); + out_pq2 = vaddq_u16(out_pq2, qp_sum); + + qp_rev = vreinterpret_u8_u32(vrev64_u32(vreinterpret_u32_u8(*p4q4))); + qp_sum = vaddw_u8(qp_sum, qp_rev); + out_pq1 = vaddq_u16(out_pq1, qp_sum); + + qp_rev = vreinterpret_u8_u32(vrev64_u32(vreinterpret_u32_u8(*p5q5))); + qp_sum = vaddw_u8(qp_sum, qp_rev); + out_pq0 = vaddq_u16(out_pq0, qp_sum); + + out_pq0 = vaddw_u8(out_pq0, q0p0); + + out_f14_pq0 = vrshrn_n_u16(out_pq0, 4); + out_f14_pq1 = vrshrn_n_u16(out_pq1, 4); + out_f14_pq2 = vrshrn_n_u16(out_pq2, 4); + out_f14_pq3 = vrshrn_n_u16(out_pq3, 4); + out_f14_pq4 = vrshrn_n_u16(out_pq4, 4); + out_f14_pq5 = vrshrn_n_u16(out_pq5, 4); + } + { + uint8x8_t filter4_cond, filter8_cond, filter14_cond; + filter8_cond = vand_u8(flat_8x8, mask_8x8); + filter4_cond = vmvn_u8(filter8_cond); + filter14_cond = vand_u8(filter8_cond, flat2_8x8); + + // filter4 outputs + *p0q0 = vbsl_u8(filter4_cond, out_f4_pq0, *p0q0); + *p1q1 = vbsl_u8(filter4_cond, out_f4_pq1, *p1q1); + + // filter8 outputs + *p0q0 = vbsl_u8(filter8_cond, out_f7_pq0, *p0q0); + *p1q1 = vbsl_u8(filter8_cond, out_f7_pq1, *p1q1); + *p2q2 = vbsl_u8(filter8_cond, out_f7_pq2, *p2q2); + + // filter14 outputs + *p0q0 = vbsl_u8(filter14_cond, out_f14_pq0, *p0q0); + *p1q1 = vbsl_u8(filter14_cond, out_f14_pq1, *p1q1); + *p2q2 = vbsl_u8(filter14_cond, out_f14_pq2, *p2q2); + *p3q3 = vbsl_u8(filter14_cond, out_f14_pq3, *p3q3); + *p4q4 = vbsl_u8(filter14_cond, out_f14_pq4, *p4q4); + *p5q5 = vbsl_u8(filter14_cond, out_f14_pq5, *p5q5); + } +} + +static void lpf_8_neon(uint8x8_t *p3q3, uint8x8_t *p2q2, uint8x8_t *p1q1, + uint8x8_t *p0q0, const uint8_t blimit, + const uint8_t limit, const uint8_t thresh) { + uint16x8_t out; + uint8x8_t out_f7_pq0, out_f7_pq1, out_f7_pq2; + uint8x8_t out_f4_pq0, out_f4_pq1; + uint8x8_t mask_8x8, flat_8x8; + + // Calculate filter masks + mask_8x8 = lpf_mask(*p3q3, *p2q2, *p1q1, *p0q0, blimit, limit); + flat_8x8 = lpf_flat_mask4(*p3q3, *p2q2, *p1q1, *p0q0); + { + // filter 4 + int32x2x2_t ps0_qs0, ps1_qs1; + int16x8_t filter_s16; + const uint8x8_t thresh_f4 = vdup_n_u8(thresh); + uint8x8_t temp0_8x8, temp1_8x8; + int8x8_t ps0_s8, ps1_s8, qs0_s8, qs1_s8, temp_s8; + int8x8_t op0, oq0, op1, oq1; + int8x8_t pq_s0, pq_s1; + int8x8_t filter_s8, filter1_s8, filter2_s8; + int8x8_t hev_8x8; + const int8x8_t sign_mask = vdup_n_s8(0x80); + const int8x8_t val_4 = vdup_n_s8(4); + const int8x8_t val_3 = vdup_n_s8(3); + + pq_s0 = veor_s8(vreinterpret_s8_u8(*p0q0), sign_mask); + pq_s1 = veor_s8(vreinterpret_s8_u8(*p1q1), sign_mask); + + ps0_qs0 = vtrn_s32(vreinterpret_s32_s8(pq_s0), vreinterpret_s32_s8(pq_s0)); + ps1_qs1 = vtrn_s32(vreinterpret_s32_s8(pq_s1), vreinterpret_s32_s8(pq_s1)); + ps0_s8 = vreinterpret_s8_s32(ps0_qs0.val[0]); + qs0_s8 = vreinterpret_s8_s32(ps0_qs0.val[1]); + ps1_s8 = vreinterpret_s8_s32(ps1_qs1.val[0]); + qs1_s8 = vreinterpret_s8_s32(ps1_qs1.val[1]); + + // hev_mask + temp0_8x8 = vcgt_u8(vabd_u8(*p0q0, *p1q1), thresh_f4); + temp1_8x8 = vreinterpret_u8_u32(vrev64_u32(vreinterpret_u32_u8(temp0_8x8))); + hev_8x8 = vreinterpret_s8_u8(vorr_u8(temp0_8x8, temp1_8x8)); + + // add outer taps if we have high edge variance + filter_s8 = vqsub_s8(ps1_s8, qs1_s8); + filter_s8 = vand_s8(filter_s8, hev_8x8); + + // inner taps + temp_s8 = vqsub_s8(qs0_s8, ps0_s8); + filter_s16 = vmovl_s8(filter_s8); + filter_s16 = vmlal_s8(filter_s16, temp_s8, val_3); + filter_s8 = vqmovn_s16(filter_s16); + filter_s8 = vand_s8(filter_s8, vreinterpret_s8_u8(mask_8x8)); + + filter1_s8 = vqadd_s8(filter_s8, val_4); + filter2_s8 = vqadd_s8(filter_s8, val_3); + filter1_s8 = vshr_n_s8(filter1_s8, 3); + filter2_s8 = vshr_n_s8(filter2_s8, 3); + + oq0 = veor_s8(vqsub_s8(qs0_s8, filter1_s8), sign_mask); + op0 = veor_s8(vqadd_s8(ps0_s8, filter2_s8), sign_mask); + + hev_8x8 = vmvn_s8(hev_8x8); + filter_s8 = vrshr_n_s8(filter1_s8, 1); + filter_s8 = vand_s8(filter_s8, hev_8x8); + + oq1 = veor_s8(vqsub_s8(qs1_s8, filter_s8), sign_mask); + op1 = veor_s8(vqadd_s8(ps1_s8, filter_s8), sign_mask); + + out_f4_pq0 = vreinterpret_u8_s8(vext_s8(op0, oq0, 4)); + out_f4_pq1 = vreinterpret_u8_s8(vext_s8(op1, oq1, 4)); + } + { + // filter 8 + uint16x8_t out_pq0, out_pq1, out_pq2; + uint8x8_t q0p0, q1p1, q2p2; + + out = vaddl_u8(*p3q3, *p2q2); + out = vaddw_u8(out, *p1q1); + out = vaddw_u8(out, *p0q0); + + // reverse p and q + q0p0 = vreinterpret_u8_u32(vrev64_u32(vreinterpret_u32_u8(*p0q0))); + q1p1 = vreinterpret_u8_u32(vrev64_u32(vreinterpret_u32_u8(*p1q1))); + q2p2 = vreinterpret_u8_u32(vrev64_u32(vreinterpret_u32_u8(*p2q2))); + + out = vaddw_u8(out, q0p0); + out_pq1 = vaddw_u8(out, *p3q3); + out_pq2 = vaddw_u8(out_pq1, *p3q3); + out_pq2 = vaddw_u8(out_pq2, *p2q2); + out_pq1 = vaddw_u8(out_pq1, *p1q1); + out_pq1 = vaddw_u8(out_pq1, q1p1); + + out_pq0 = vaddw_u8(out, *p0q0); + out_pq0 = vaddw_u8(out_pq0, q1p1); + out_pq0 = vaddw_u8(out_pq0, q2p2); + + out_f7_pq0 = vrshrn_n_u16(out_pq0, 3); + out_f7_pq1 = vrshrn_n_u16(out_pq1, 3); + out_f7_pq2 = vrshrn_n_u16(out_pq2, 3); + } + { + uint8x8_t filter4_cond, filter8_cond; + filter8_cond = vand_u8(flat_8x8, mask_8x8); + filter4_cond = vmvn_u8(filter8_cond); + + // filter4 outputs + *p0q0 = vbsl_u8(filter4_cond, out_f4_pq0, *p0q0); + *p1q1 = vbsl_u8(filter4_cond, out_f4_pq1, *p1q1); + + // filter8 outputs + *p0q0 = vbsl_u8(filter8_cond, out_f7_pq0, *p0q0); + *p1q1 = vbsl_u8(filter8_cond, out_f7_pq1, *p1q1); + *p2q2 = vbsl_u8(filter8_cond, out_f7_pq2, *p2q2); + } +} + +static void lpf_6_neon(uint8x8_t *p2q2, uint8x8_t *p1q1, uint8x8_t *p0q0, + const uint8_t blimit, const uint8_t limit, + const uint8_t thresh) { + uint16x8_t out; + uint8x8_t out_f6_pq0, out_f6_pq1; + uint8x8_t out_f4_pq0, out_f4_pq1; + uint8x8_t mask_8x8, flat_8x8; + + // Calculate filter masks + mask_8x8 = lpf_mask3_chroma(*p2q2, *p1q1, *p0q0, blimit, limit); + flat_8x8 = lpf_flat_mask3(*p2q2, *p1q1, *p0q0); + { + // filter 4 + int32x2x2_t ps0_qs0, ps1_qs1; + int16x8_t filter_s16; + const uint8x8_t thresh_f4 = vdup_n_u8(thresh); + uint8x8_t temp0_8x8, temp1_8x8; + int8x8_t ps0_s8, ps1_s8, qs0_s8, qs1_s8, temp_s8; + int8x8_t op0, oq0, op1, oq1; + int8x8_t pq_s0, pq_s1; + int8x8_t filter_s8, filter1_s8, filter2_s8; + int8x8_t hev_8x8; + const int8x8_t sign_mask = vdup_n_s8(0x80); + const int8x8_t val_4 = vdup_n_s8(4); + const int8x8_t val_3 = vdup_n_s8(3); + + pq_s0 = veor_s8(vreinterpret_s8_u8(*p0q0), sign_mask); + pq_s1 = veor_s8(vreinterpret_s8_u8(*p1q1), sign_mask); + + ps0_qs0 = vtrn_s32(vreinterpret_s32_s8(pq_s0), vreinterpret_s32_s8(pq_s0)); + ps1_qs1 = vtrn_s32(vreinterpret_s32_s8(pq_s1), vreinterpret_s32_s8(pq_s1)); + ps0_s8 = vreinterpret_s8_s32(ps0_qs0.val[0]); + qs0_s8 = vreinterpret_s8_s32(ps0_qs0.val[1]); + ps1_s8 = vreinterpret_s8_s32(ps1_qs1.val[0]); + qs1_s8 = vreinterpret_s8_s32(ps1_qs1.val[1]); + + // hev_mask + temp0_8x8 = vcgt_u8(vabd_u8(*p0q0, *p1q1), thresh_f4); + temp1_8x8 = vreinterpret_u8_u32(vrev64_u32(vreinterpret_u32_u8(temp0_8x8))); + hev_8x8 = vreinterpret_s8_u8(vorr_u8(temp0_8x8, temp1_8x8)); + + // add outer taps if we have high edge variance + filter_s8 = vqsub_s8(ps1_s8, qs1_s8); + filter_s8 = vand_s8(filter_s8, hev_8x8); + + // inner taps + temp_s8 = vqsub_s8(qs0_s8, ps0_s8); + filter_s16 = vmovl_s8(filter_s8); + filter_s16 = vmlal_s8(filter_s16, temp_s8, val_3); + filter_s8 = vqmovn_s16(filter_s16); + filter_s8 = vand_s8(filter_s8, vreinterpret_s8_u8(mask_8x8)); + + filter1_s8 = vqadd_s8(filter_s8, val_4); + filter2_s8 = vqadd_s8(filter_s8, val_3); + filter1_s8 = vshr_n_s8(filter1_s8, 3); + filter2_s8 = vshr_n_s8(filter2_s8, 3); + + oq0 = veor_s8(vqsub_s8(qs0_s8, filter1_s8), sign_mask); + op0 = veor_s8(vqadd_s8(ps0_s8, filter2_s8), sign_mask); + + filter_s8 = vrshr_n_s8(filter1_s8, 1); + filter_s8 = vbic_s8(filter_s8, hev_8x8); + + oq1 = veor_s8(vqsub_s8(qs1_s8, filter_s8), sign_mask); + op1 = veor_s8(vqadd_s8(ps1_s8, filter_s8), sign_mask); + + out_f4_pq0 = vreinterpret_u8_s8(vext_s8(op0, oq0, 4)); + out_f4_pq1 = vreinterpret_u8_s8(vext_s8(op1, oq1, 4)); + } + { + // filter 6 + uint16x8_t out_pq0, out_pq1; + uint8x8_t pq_rev; + + out = vaddl_u8(*p0q0, *p1q1); + out = vaddq_u16(out, out); + out = vaddw_u8(out, *p2q2); + + pq_rev = vreinterpret_u8_u32(vrev64_u32(vreinterpret_u32_u8(*p0q0))); + out = vaddw_u8(out, pq_rev); + + out_pq0 = vaddw_u8(out, pq_rev); + pq_rev = vreinterpret_u8_u32(vrev64_u32(vreinterpret_u32_u8(*p1q1))); + out_pq0 = vaddw_u8(out_pq0, pq_rev); + + out_pq1 = vaddw_u8(out, *p2q2); + out_pq1 = vaddw_u8(out_pq1, *p2q2); + + out_f6_pq0 = vrshrn_n_u16(out_pq0, 3); + out_f6_pq1 = vrshrn_n_u16(out_pq1, 3); + } + { + uint8x8_t filter4_cond, filter6_cond; + filter6_cond = vand_u8(flat_8x8, mask_8x8); + filter4_cond = vmvn_u8(filter6_cond); + + // filter4 outputs + *p0q0 = vbsl_u8(filter4_cond, out_f4_pq0, *p0q0); + *p1q1 = vbsl_u8(filter4_cond, out_f4_pq1, *p1q1); + + // filter6 outputs + *p0q0 = vbsl_u8(filter6_cond, out_f6_pq0, *p0q0); + *p1q1 = vbsl_u8(filter6_cond, out_f6_pq1, *p1q1); + } +} + +static void lpf_4_neon(uint8x8_t *p1q1, uint8x8_t *p0q0, const uint8_t blimit, + const uint8_t limit, const uint8_t thresh) { + int32x2x2_t ps0_qs0, ps1_qs1; + int16x8_t filter_s16; + const uint8x8_t thresh_f4 = vdup_n_u8(thresh); + uint8x8_t mask_8x8, temp0_8x8, temp1_8x8; + int8x8_t ps0_s8, ps1_s8, qs0_s8, qs1_s8, temp_s8; + int8x8_t op0, oq0, op1, oq1; + int8x8_t pq_s0, pq_s1; + int8x8_t filter_s8, filter1_s8, filter2_s8; + int8x8_t hev_8x8; + const int8x8_t sign_mask = vdup_n_s8(0x80); + const int8x8_t val_4 = vdup_n_s8(4); + const int8x8_t val_3 = vdup_n_s8(3); + + // Calculate filter mask + mask_8x8 = lpf_mask2(*p1q1, *p0q0, blimit, limit); + + pq_s0 = veor_s8(vreinterpret_s8_u8(*p0q0), sign_mask); + pq_s1 = veor_s8(vreinterpret_s8_u8(*p1q1), sign_mask); + + ps0_qs0 = vtrn_s32(vreinterpret_s32_s8(pq_s0), vreinterpret_s32_s8(pq_s0)); + ps1_qs1 = vtrn_s32(vreinterpret_s32_s8(pq_s1), vreinterpret_s32_s8(pq_s1)); + ps0_s8 = vreinterpret_s8_s32(ps0_qs0.val[0]); + qs0_s8 = vreinterpret_s8_s32(ps0_qs0.val[1]); + ps1_s8 = vreinterpret_s8_s32(ps1_qs1.val[0]); + qs1_s8 = vreinterpret_s8_s32(ps1_qs1.val[1]); + + // hev_mask + temp0_8x8 = vcgt_u8(vabd_u8(*p0q0, *p1q1), thresh_f4); + temp1_8x8 = vreinterpret_u8_u32(vrev64_u32(vreinterpret_u32_u8(temp0_8x8))); + hev_8x8 = vreinterpret_s8_u8(vorr_u8(temp0_8x8, temp1_8x8)); + + // add outer taps if we have high edge variance + filter_s8 = vqsub_s8(ps1_s8, qs1_s8); + filter_s8 = vand_s8(filter_s8, hev_8x8); + + // inner taps + temp_s8 = vqsub_s8(qs0_s8, ps0_s8); + filter_s16 = vmovl_s8(filter_s8); + filter_s16 = vmlal_s8(filter_s16, temp_s8, val_3); + filter_s8 = vqmovn_s16(filter_s16); + filter_s8 = vand_s8(filter_s8, vreinterpret_s8_u8(mask_8x8)); + + filter1_s8 = vqadd_s8(filter_s8, val_4); + filter2_s8 = vqadd_s8(filter_s8, val_3); + filter1_s8 = vshr_n_s8(filter1_s8, 3); + filter2_s8 = vshr_n_s8(filter2_s8, 3); + + oq0 = veor_s8(vqsub_s8(qs0_s8, filter1_s8), sign_mask); + op0 = veor_s8(vqadd_s8(ps0_s8, filter2_s8), sign_mask); + + filter_s8 = vrshr_n_s8(filter1_s8, 1); + filter_s8 = vbic_s8(filter_s8, hev_8x8); + + oq1 = veor_s8(vqsub_s8(qs1_s8, filter_s8), sign_mask); + op1 = veor_s8(vqadd_s8(ps1_s8, filter_s8), sign_mask); + + *p0q0 = vreinterpret_u8_s8(vext_s8(op0, oq0, 4)); + *p1q1 = vreinterpret_u8_s8(vext_s8(op1, oq1, 4)); +} + +void aom_lpf_vertical_14_neon(uint8_t *src, int stride, const uint8_t *blimit, + const uint8_t *limit, const uint8_t *thresh) { + uint8x16_t row0, row1, row2, row3; + uint8x8_t pxp3, p6p2, p5p1, p4p0; + uint8x8_t q0q4, q1q5, q2q6, q3qy; + uint32x2x2_t p6q6_p2q2, p5q5_p1q1, p4q4_p0q0, pxqx_p3q3; + uint32x2_t pq_rev; + uint8x8_t p0q0, p1q1, p2q2, p3q3, p4q4, p5q5, p6q6; + + // row0: x p6 p5 p4 p3 p2 p1 p0 | q0 q1 q2 q3 q4 q5 q6 y + // row1: x p6 p5 p4 p3 p2 p1 p0 | q0 q1 q2 q3 q4 q5 q6 y + // row2: x p6 p5 p4 p3 p2 p1 p0 | q0 q1 q2 q3 q4 q5 q6 y + // row3: x p6 p5 p4 p3 p2 p1 p0 | q0 q1 q2 q3 q4 q5 q6 y + load_u8_8x16(src - 8, stride, &row0, &row1, &row2, &row3); + + pxp3 = vget_low_u8(row0); + p6p2 = vget_low_u8(row1); + p5p1 = vget_low_u8(row2); + p4p0 = vget_low_u8(row3); + transpose_u8_8x4(&pxp3, &p6p2, &p5p1, &p4p0); + + q0q4 = vget_high_u8(row0); + q1q5 = vget_high_u8(row1); + q2q6 = vget_high_u8(row2); + q3qy = vget_high_u8(row3); + transpose_u8_8x4(&q0q4, &q1q5, &q2q6, &q3qy); + + pq_rev = vrev64_u32(vreinterpret_u32_u8(q3qy)); + pxqx_p3q3 = vtrn_u32(vreinterpret_u32_u8(pxp3), pq_rev); + + pq_rev = vrev64_u32(vreinterpret_u32_u8(q1q5)); + p5q5_p1q1 = vtrn_u32(vreinterpret_u32_u8(p5p1), pq_rev); + + pq_rev = vrev64_u32(vreinterpret_u32_u8(q0q4)); + p4q4_p0q0 = vtrn_u32(vreinterpret_u32_u8(p4p0), pq_rev); + + pq_rev = vrev64_u32(vreinterpret_u32_u8(q2q6)); + p6q6_p2q2 = vtrn_u32(vreinterpret_u32_u8(p6p2), pq_rev); + + p0q0 = vreinterpret_u8_u32(p4q4_p0q0.val[1]); + p1q1 = vreinterpret_u8_u32(p5q5_p1q1.val[1]); + p2q2 = vreinterpret_u8_u32(p6q6_p2q2.val[1]); + p3q3 = vreinterpret_u8_u32(pxqx_p3q3.val[1]); + p4q4 = vreinterpret_u8_u32(p4q4_p0q0.val[0]); + p5q5 = vreinterpret_u8_u32(p5q5_p1q1.val[0]); + p6q6 = vreinterpret_u8_u32(p6q6_p2q2.val[0]); + + lpf_14_neon(&p6q6, &p5q5, &p4q4, &p3q3, &p2q2, &p1q1, &p0q0, *blimit, *limit, + *thresh); + + pxqx_p3q3 = vtrn_u32(pxqx_p3q3.val[0], vreinterpret_u32_u8(p3q3)); + p5q5_p1q1 = vtrn_u32(vreinterpret_u32_u8(p5q5), vreinterpret_u32_u8(p1q1)); + p4q4_p0q0 = vtrn_u32(vreinterpret_u32_u8(p4q4), vreinterpret_u32_u8(p0q0)); + p6q6_p2q2 = vtrn_u32(vreinterpret_u32_u8(p6q6), vreinterpret_u32_u8(p2q2)); + + pxqx_p3q3.val[1] = vrev64_u32(pxqx_p3q3.val[1]); + p5q5_p1q1.val[1] = vrev64_u32(p5q5_p1q1.val[1]); + p4q4_p0q0.val[1] = vrev64_u32(p4q4_p0q0.val[1]); + p6q6_p2q2.val[1] = vrev64_u32(p6q6_p2q2.val[1]); + + q0q4 = vreinterpret_u8_u32(p4q4_p0q0.val[1]); + q1q5 = vreinterpret_u8_u32(p5q5_p1q1.val[1]); + q2q6 = vreinterpret_u8_u32(p6q6_p2q2.val[1]); + q3qy = vreinterpret_u8_u32(pxqx_p3q3.val[1]); + transpose_u8_8x4(&q0q4, &q1q5, &q2q6, &q3qy); + + pxp3 = vreinterpret_u8_u32(pxqx_p3q3.val[0]); + p6p2 = vreinterpret_u8_u32(p6q6_p2q2.val[0]); + p5p1 = vreinterpret_u8_u32(p5q5_p1q1.val[0]); + p4p0 = vreinterpret_u8_u32(p4q4_p0q0.val[0]); + transpose_u8_8x4(&pxp3, &p6p2, &p5p1, &p4p0); + + row0 = vcombine_u8(pxp3, q0q4); + row1 = vcombine_u8(p6p2, q1q5); + row2 = vcombine_u8(p5p1, q2q6); + row3 = vcombine_u8(p4p0, q3qy); + + store_u8_8x16(src - 8, stride, row0, row1, row2, row3); +} + +void aom_lpf_vertical_8_neon(uint8_t *src, int stride, const uint8_t *blimit, + const uint8_t *limit, const uint8_t *thresh) { + uint32x2x2_t p2q2_p1q1, p3q3_p0q0; + uint32x2_t pq_rev; + uint8x8_t p3q0, p2q1, p1q2, p0q3; + uint8x8_t p0q0, p1q1, p2q2, p3q3; + + // row0: p3 p2 p1 p0 | q0 q1 q2 q3 + // row1: p3 p2 p1 p0 | q0 q1 q2 q3 + // row2: p3 p2 p1 p0 | q0 q1 q2 q3 + // row3: p3 p2 p1 p0 | q0 q1 q2 q3 + load_u8_8x4(src - 4, stride, &p3q0, &p2q1, &p1q2, &p0q3); + + transpose_u8_8x4(&p3q0, &p2q1, &p1q2, &p0q3); + + pq_rev = vrev64_u32(vreinterpret_u32_u8(p0q3)); + p3q3_p0q0 = vtrn_u32(vreinterpret_u32_u8(p3q0), pq_rev); + + pq_rev = vrev64_u32(vreinterpret_u32_u8(p1q2)); + p2q2_p1q1 = vtrn_u32(vreinterpret_u32_u8(p2q1), pq_rev); + + p0q0 = vreinterpret_u8_u32(vrev64_u32(p3q3_p0q0.val[1])); + p1q1 = vreinterpret_u8_u32(vrev64_u32(p2q2_p1q1.val[1])); + p2q2 = vreinterpret_u8_u32(p2q2_p1q1.val[0]); + p3q3 = vreinterpret_u8_u32(p3q3_p0q0.val[0]); + + lpf_8_neon(&p3q3, &p2q2, &p1q1, &p0q0, *blimit, *limit, *thresh); + + pq_rev = vrev64_u32(vreinterpret_u32_u8(p0q0)); + p3q3_p0q0 = vtrn_u32(vreinterpret_u32_u8(p3q3), pq_rev); + + pq_rev = vrev64_u32(vreinterpret_u32_u8(p1q1)); + p2q2_p1q1 = vtrn_u32(vreinterpret_u32_u8(p2q2), pq_rev); + + p0q3 = vreinterpret_u8_u32(vrev64_u32(p3q3_p0q0.val[1])); + p1q2 = vreinterpret_u8_u32(vrev64_u32(p2q2_p1q1.val[1])); + p2q1 = vreinterpret_u8_u32(p2q2_p1q1.val[0]); + p3q0 = vreinterpret_u8_u32(p3q3_p0q0.val[0]); + transpose_u8_8x4(&p3q0, &p2q1, &p1q2, &p0q3); + + store_u8_8x4(src - 4, stride, p3q0, p2q1, p1q2, p0q3); +} + +void aom_lpf_vertical_6_neon(uint8_t *src, int stride, const uint8_t *blimit, + const uint8_t *limit, const uint8_t *thresh) { + uint32x2x2_t p2q2_p1q1, pxqy_p0q0; + uint32x2_t pq_rev; + uint8x8_t pxq0, p2q1, p1q2, p0qy; + uint8x8_t p0q0, p1q1, p2q2, pxqy; + + // row0: px p2 p1 p0 | q0 q1 q2 qy + // row1: px p2 p1 p0 | q0 q1 q2 qy + // row2: px p2 p1 p0 | q0 q1 q2 qy + // row3: px p2 p1 p0 | q0 q1 q2 qy + load_u8_8x4(src - 4, stride, &pxq0, &p2q1, &p1q2, &p0qy); + + transpose_u8_8x4(&pxq0, &p2q1, &p1q2, &p0qy); + + pq_rev = vrev64_u32(vreinterpret_u32_u8(p0qy)); + pxqy_p0q0 = vtrn_u32(vreinterpret_u32_u8(pxq0), pq_rev); + + pq_rev = vrev64_u32(vreinterpret_u32_u8(p1q2)); + p2q2_p1q1 = vtrn_u32(vreinterpret_u32_u8(p2q1), pq_rev); + + p0q0 = vreinterpret_u8_u32(vrev64_u32(pxqy_p0q0.val[1])); + p1q1 = vreinterpret_u8_u32(vrev64_u32(p2q2_p1q1.val[1])); + p2q2 = vreinterpret_u8_u32(p2q2_p1q1.val[0]); + pxqy = vreinterpret_u8_u32(pxqy_p0q0.val[0]); + + lpf_6_neon(&p2q2, &p1q1, &p0q0, *blimit, *limit, *thresh); + + pq_rev = vrev64_u32(vreinterpret_u32_u8(p0q0)); + pxqy_p0q0 = vtrn_u32(vreinterpret_u32_u8(pxqy), pq_rev); + + pq_rev = vrev64_u32(vreinterpret_u32_u8(p1q1)); + p2q2_p1q1 = vtrn_u32(vreinterpret_u32_u8(p2q2), pq_rev); + + p0qy = vreinterpret_u8_u32(vrev64_u32(pxqy_p0q0.val[1])); + p1q2 = vreinterpret_u8_u32(vrev64_u32(p2q2_p1q1.val[1])); + p2q1 = vreinterpret_u8_u32(p2q2_p1q1.val[0]); + pxq0 = vreinterpret_u8_u32(pxqy_p0q0.val[0]); + transpose_u8_8x4(&pxq0, &p2q1, &p1q2, &p0qy); + + store_u8_8x4(src - 4, stride, pxq0, p2q1, p1q2, p0qy); +} + +void aom_lpf_vertical_4_neon(uint8_t *src, int stride, const uint8_t *blimit, + const uint8_t *limit, const uint8_t *thresh) { + uint32x2x2_t p1q0_p0q1, p1q1_p0q0, p1p0_q1q0; + uint32x2_t pq_rev; + uint8x8_t UNINITIALIZED_IS_SAFE(p1p0), q0q1, p0q0, p1q1; + + // row0: p1 p0 | q0 q1 + // row1: p1 p0 | q0 q1 + // row2: p1 p0 | q0 q1 + // row3: p1 p0 | q0 q1 + load_u8_4x1(src - 2, &p1p0, 0); + load_u8_4x1((src - 2) + 1 * stride, &p1p0, 1); + load_u8_4x1((src - 2) + 2 * stride, &q0q1, 0); + load_u8_4x1((src - 2) + 3 * stride, &q0q1, 1); + + transpose_u8_4x4(&p1p0, &q0q1); + + p1q0_p0q1 = vtrn_u32(vreinterpret_u32_u8(p1p0), vreinterpret_u32_u8(q0q1)); + + pq_rev = vrev64_u32(p1q0_p0q1.val[1]); + p1q1_p0q0 = vtrn_u32(p1q0_p0q1.val[0], pq_rev); + + p1q1 = vreinterpret_u8_u32(p1q1_p0q0.val[0]); + p0q0 = vreinterpret_u8_u32(p1q1_p0q0.val[1]); + + lpf_4_neon(&p1q1, &p0q0, *blimit, *limit, *thresh); + + p1p0_q1q0 = vtrn_u32(vreinterpret_u32_u8(p1q1), vreinterpret_u32_u8(p0q0)); + + p1p0 = vreinterpret_u8_u32(p1p0_q1q0.val[0]); + q0q1 = vreinterpret_u8_u32(vrev64_u32(p1p0_q1q0.val[1])); + + transpose_u8_4x4(&p1p0, &q0q1); + + store_u8_4x1(src - 2, p1p0, 0); + store_u8_4x1((src - 2) + 1 * stride, q0q1, 0); + store_u8_4x1((src - 2) + 2 * stride, p1p0, 1); + store_u8_4x1((src - 2) + 3 * stride, q0q1, 1); +} + +void aom_lpf_horizontal_14_neon(uint8_t *src, int stride, const uint8_t *blimit, + const uint8_t *limit, const uint8_t *thresh) { + uint8x8_t p0q0, p1q1, p2q2, p3q3, p4q4, p5q5, UNINITIALIZED_IS_SAFE(p6q6); + + load_u8_4x1(src - 7 * stride, &p6q6, 0); + load_u8_4x1(src - 6 * stride, &p5q5, 0); + load_u8_4x1(src - 5 * stride, &p4q4, 0); + load_u8_4x1(src - 4 * stride, &p3q3, 0); + load_u8_4x1(src - 3 * stride, &p2q2, 0); + load_u8_4x1(src - 2 * stride, &p1q1, 0); + load_u8_4x1(src - 1 * stride, &p0q0, 0); + load_u8_4x1(src + 0 * stride, &p0q0, 1); + load_u8_4x1(src + 1 * stride, &p1q1, 1); + load_u8_4x1(src + 2 * stride, &p2q2, 1); + load_u8_4x1(src + 3 * stride, &p3q3, 1); + load_u8_4x1(src + 4 * stride, &p4q4, 1); + load_u8_4x1(src + 5 * stride, &p5q5, 1); + load_u8_4x1(src + 6 * stride, &p6q6, 1); + + lpf_14_neon(&p6q6, &p5q5, &p4q4, &p3q3, &p2q2, &p1q1, &p0q0, *blimit, *limit, + *thresh); + + store_u8_4x1(src - 6 * stride, p5q5, 0); + store_u8_4x1(src - 5 * stride, p4q4, 0); + store_u8_4x1(src - 4 * stride, p3q3, 0); + store_u8_4x1(src - 3 * stride, p2q2, 0); + store_u8_4x1(src - 2 * stride, p1q1, 0); + store_u8_4x1(src - 1 * stride, p0q0, 0); + store_u8_4x1(src + 0 * stride, p0q0, 1); + store_u8_4x1(src + 1 * stride, p1q1, 1); + store_u8_4x1(src + 2 * stride, p2q2, 1); + store_u8_4x1(src + 3 * stride, p3q3, 1); + store_u8_4x1(src + 4 * stride, p4q4, 1); + store_u8_4x1(src + 5 * stride, p5q5, 1); +} + +void aom_lpf_horizontal_8_neon(uint8_t *src, int stride, const uint8_t *blimit, + const uint8_t *limit, const uint8_t *thresh) { + uint8x8_t p0q0, p1q1, p2q2, p3q3; + + p3q3 = vreinterpret_u8_u32(vld1_dup_u32((uint32_t *)(src - 4 * stride))); + p2q2 = vreinterpret_u8_u32(vld1_dup_u32((uint32_t *)(src - 3 * stride))); + p1q1 = vreinterpret_u8_u32(vld1_dup_u32((uint32_t *)(src - 2 * stride))); + p0q0 = vreinterpret_u8_u32(vld1_dup_u32((uint32_t *)(src - 1 * stride))); + p0q0 = vreinterpret_u8_u32(vld1_lane_u32((uint32_t *)(src + 0 * stride), + vreinterpret_u32_u8(p0q0), 1)); + p1q1 = vreinterpret_u8_u32(vld1_lane_u32((uint32_t *)(src + 1 * stride), + vreinterpret_u32_u8(p1q1), 1)); + p2q2 = vreinterpret_u8_u32(vld1_lane_u32((uint32_t *)(src + 2 * stride), + vreinterpret_u32_u8(p2q2), 1)); + p3q3 = vreinterpret_u8_u32(vld1_lane_u32((uint32_t *)(src + 3 * stride), + vreinterpret_u32_u8(p3q3), 1)); + + lpf_8_neon(&p3q3, &p2q2, &p1q1, &p0q0, *blimit, *limit, *thresh); + + vst1_lane_u32((uint32_t *)(src - 4 * stride), vreinterpret_u32_u8(p3q3), 0); + vst1_lane_u32((uint32_t *)(src - 3 * stride), vreinterpret_u32_u8(p2q2), 0); + vst1_lane_u32((uint32_t *)(src - 2 * stride), vreinterpret_u32_u8(p1q1), 0); + vst1_lane_u32((uint32_t *)(src - 1 * stride), vreinterpret_u32_u8(p0q0), 0); + vst1_lane_u32((uint32_t *)(src + 0 * stride), vreinterpret_u32_u8(p0q0), 1); + vst1_lane_u32((uint32_t *)(src + 1 * stride), vreinterpret_u32_u8(p1q1), 1); + vst1_lane_u32((uint32_t *)(src + 2 * stride), vreinterpret_u32_u8(p2q2), 1); + vst1_lane_u32((uint32_t *)(src + 3 * stride), vreinterpret_u32_u8(p3q3), 1); +} + +void aom_lpf_horizontal_6_neon(uint8_t *src, int stride, const uint8_t *blimit, + const uint8_t *limit, const uint8_t *thresh) { + uint8x8_t p0q0, p1q1, p2q2; + + p2q2 = vreinterpret_u8_u32(vld1_dup_u32((uint32_t *)(src - 3 * stride))); + p1q1 = vreinterpret_u8_u32(vld1_dup_u32((uint32_t *)(src - 2 * stride))); + p0q0 = vreinterpret_u8_u32(vld1_dup_u32((uint32_t *)(src - 1 * stride))); + p0q0 = vreinterpret_u8_u32(vld1_lane_u32((uint32_t *)(src + 0 * stride), + vreinterpret_u32_u8(p0q0), 1)); + p1q1 = vreinterpret_u8_u32(vld1_lane_u32((uint32_t *)(src + 1 * stride), + vreinterpret_u32_u8(p1q1), 1)); + p2q2 = vreinterpret_u8_u32(vld1_lane_u32((uint32_t *)(src + 2 * stride), + vreinterpret_u32_u8(p2q2), 1)); + + lpf_6_neon(&p2q2, &p1q1, &p0q0, *blimit, *limit, *thresh); + + vst1_lane_u32((uint32_t *)(src - 3 * stride), vreinterpret_u32_u8(p2q2), 0); + vst1_lane_u32((uint32_t *)(src - 2 * stride), vreinterpret_u32_u8(p1q1), 0); + vst1_lane_u32((uint32_t *)(src - 1 * stride), vreinterpret_u32_u8(p0q0), 0); + vst1_lane_u32((uint32_t *)(src + 0 * stride), vreinterpret_u32_u8(p0q0), 1); + vst1_lane_u32((uint32_t *)(src + 1 * stride), vreinterpret_u32_u8(p1q1), 1); + vst1_lane_u32((uint32_t *)(src + 2 * stride), vreinterpret_u32_u8(p2q2), 1); +} + +void aom_lpf_horizontal_4_neon(uint8_t *src, int stride, const uint8_t *blimit, + const uint8_t *limit, const uint8_t *thresh) { + uint8x8_t p0q0, UNINITIALIZED_IS_SAFE(p1q1); + + load_u8_4x1(src - 2 * stride, &p1q1, 0); + load_u8_4x1(src - 1 * stride, &p0q0, 0); + load_u8_4x1(src + 0 * stride, &p0q0, 1); + load_u8_4x1(src + 1 * stride, &p1q1, 1); + + lpf_4_neon(&p1q1, &p0q0, *blimit, *limit, *thresh); + + store_u8_4x1(src - 2 * stride, p1q1, 0); + store_u8_4x1(src - 1 * stride, p0q0, 0); + store_u8_4x1(src + 0 * stride, p0q0, 1); + store_u8_4x1(src + 1 * stride, p1q1, 1); +} diff --git a/third_party/aom/aom_dsp/arm/sad4d_neon.c b/third_party/aom/aom_dsp/arm/sad4d_neon.c new file mode 100644 index 0000000000..606950ab25 --- /dev/null +++ b/third_party/aom/aom_dsp/arm/sad4d_neon.c @@ -0,0 +1,226 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include + +#include "config/aom_config.h" +#include "config/aom_dsp_rtcd.h" + +#include "aom/aom_integer.h" + +static INLINE unsigned int horizontal_long_add_16x8(const uint16x8_t vec_lo, + const uint16x8_t vec_hi) { + const uint32x4_t vec_l_lo = + vaddl_u16(vget_low_u16(vec_lo), vget_high_u16(vec_lo)); + const uint32x4_t vec_l_hi = + vaddl_u16(vget_low_u16(vec_hi), vget_high_u16(vec_hi)); + const uint32x4_t a = vaddq_u32(vec_l_lo, vec_l_hi); + const uint64x2_t b = vpaddlq_u32(a); + const uint32x2_t c = vadd_u32(vreinterpret_u32_u64(vget_low_u64(b)), + vreinterpret_u32_u64(vget_high_u64(b))); + return vget_lane_u32(c, 0); +} + +// Calculate the absolute difference of 64 bytes from vec_src_00, vec_src_16, +// vec_src_32, vec_src_48 and ref. Accumulate partial sums in vec_sum_ref_lo +// and vec_sum_ref_hi. +static void sad_neon_64(const uint8x16_t vec_src_00, + const uint8x16_t vec_src_16, + const uint8x16_t vec_src_32, + const uint8x16_t vec_src_48, const uint8_t *ref, + uint16x8_t *vec_sum_ref_lo, + uint16x8_t *vec_sum_ref_hi) { + const uint8x16_t vec_ref_00 = vld1q_u8(ref); + const uint8x16_t vec_ref_16 = vld1q_u8(ref + 16); + const uint8x16_t vec_ref_32 = vld1q_u8(ref + 32); + const uint8x16_t vec_ref_48 = vld1q_u8(ref + 48); + + *vec_sum_ref_lo = vabal_u8(*vec_sum_ref_lo, vget_low_u8(vec_src_00), + vget_low_u8(vec_ref_00)); + *vec_sum_ref_hi = vabal_u8(*vec_sum_ref_hi, vget_high_u8(vec_src_00), + vget_high_u8(vec_ref_00)); + *vec_sum_ref_lo = vabal_u8(*vec_sum_ref_lo, vget_low_u8(vec_src_16), + vget_low_u8(vec_ref_16)); + *vec_sum_ref_hi = vabal_u8(*vec_sum_ref_hi, vget_high_u8(vec_src_16), + vget_high_u8(vec_ref_16)); + *vec_sum_ref_lo = vabal_u8(*vec_sum_ref_lo, vget_low_u8(vec_src_32), + vget_low_u8(vec_ref_32)); + *vec_sum_ref_hi = vabal_u8(*vec_sum_ref_hi, vget_high_u8(vec_src_32), + vget_high_u8(vec_ref_32)); + *vec_sum_ref_lo = vabal_u8(*vec_sum_ref_lo, vget_low_u8(vec_src_48), + vget_low_u8(vec_ref_48)); + *vec_sum_ref_hi = vabal_u8(*vec_sum_ref_hi, vget_high_u8(vec_src_48), + vget_high_u8(vec_ref_48)); +} + +// Calculate the absolute difference of 32 bytes from vec_src_00, vec_src_16, +// and ref. Accumulate partial sums in vec_sum_ref_lo and vec_sum_ref_hi. +static void sad_neon_32(const uint8x16_t vec_src_00, + const uint8x16_t vec_src_16, const uint8_t *ref, + uint16x8_t *vec_sum_ref_lo, + uint16x8_t *vec_sum_ref_hi) { + const uint8x16_t vec_ref_00 = vld1q_u8(ref); + const uint8x16_t vec_ref_16 = vld1q_u8(ref + 16); + + *vec_sum_ref_lo = vabal_u8(*vec_sum_ref_lo, vget_low_u8(vec_src_00), + vget_low_u8(vec_ref_00)); + *vec_sum_ref_hi = vabal_u8(*vec_sum_ref_hi, vget_high_u8(vec_src_00), + vget_high_u8(vec_ref_00)); + *vec_sum_ref_lo = vabal_u8(*vec_sum_ref_lo, vget_low_u8(vec_src_16), + vget_low_u8(vec_ref_16)); + *vec_sum_ref_hi = vabal_u8(*vec_sum_ref_hi, vget_high_u8(vec_src_16), + vget_high_u8(vec_ref_16)); +} + +void aom_sad64x64x4d_neon(const uint8_t *src, int src_stride, + const uint8_t *const ref[4], int ref_stride, + uint32_t *res) { + int i; + uint16x8_t vec_sum_ref0_lo = vdupq_n_u16(0); + uint16x8_t vec_sum_ref0_hi = vdupq_n_u16(0); + uint16x8_t vec_sum_ref1_lo = vdupq_n_u16(0); + uint16x8_t vec_sum_ref1_hi = vdupq_n_u16(0); + uint16x8_t vec_sum_ref2_lo = vdupq_n_u16(0); + uint16x8_t vec_sum_ref2_hi = vdupq_n_u16(0); + uint16x8_t vec_sum_ref3_lo = vdupq_n_u16(0); + uint16x8_t vec_sum_ref3_hi = vdupq_n_u16(0); + const uint8_t *ref0, *ref1, *ref2, *ref3; + ref0 = ref[0]; + ref1 = ref[1]; + ref2 = ref[2]; + ref3 = ref[3]; + + for (i = 0; i < 64; ++i) { + const uint8x16_t vec_src_00 = vld1q_u8(src); + const uint8x16_t vec_src_16 = vld1q_u8(src + 16); + const uint8x16_t vec_src_32 = vld1q_u8(src + 32); + const uint8x16_t vec_src_48 = vld1q_u8(src + 48); + + sad_neon_64(vec_src_00, vec_src_16, vec_src_32, vec_src_48, ref0, + &vec_sum_ref0_lo, &vec_sum_ref0_hi); + sad_neon_64(vec_src_00, vec_src_16, vec_src_32, vec_src_48, ref1, + &vec_sum_ref1_lo, &vec_sum_ref1_hi); + sad_neon_64(vec_src_00, vec_src_16, vec_src_32, vec_src_48, ref2, + &vec_sum_ref2_lo, &vec_sum_ref2_hi); + sad_neon_64(vec_src_00, vec_src_16, vec_src_32, vec_src_48, ref3, + &vec_sum_ref3_lo, &vec_sum_ref3_hi); + + src += src_stride; + ref0 += ref_stride; + ref1 += ref_stride; + ref2 += ref_stride; + ref3 += ref_stride; + } + + res[0] = horizontal_long_add_16x8(vec_sum_ref0_lo, vec_sum_ref0_hi); + res[1] = horizontal_long_add_16x8(vec_sum_ref1_lo, vec_sum_ref1_hi); + res[2] = horizontal_long_add_16x8(vec_sum_ref2_lo, vec_sum_ref2_hi); + res[3] = horizontal_long_add_16x8(vec_sum_ref3_lo, vec_sum_ref3_hi); +} + +void aom_sad32x32x4d_neon(const uint8_t *src, int src_stride, + const uint8_t *const ref[4], int ref_stride, + uint32_t *res) { + int i; + uint16x8_t vec_sum_ref0_lo = vdupq_n_u16(0); + uint16x8_t vec_sum_ref0_hi = vdupq_n_u16(0); + uint16x8_t vec_sum_ref1_lo = vdupq_n_u16(0); + uint16x8_t vec_sum_ref1_hi = vdupq_n_u16(0); + uint16x8_t vec_sum_ref2_lo = vdupq_n_u16(0); + uint16x8_t vec_sum_ref2_hi = vdupq_n_u16(0); + uint16x8_t vec_sum_ref3_lo = vdupq_n_u16(0); + uint16x8_t vec_sum_ref3_hi = vdupq_n_u16(0); + const uint8_t *ref0, *ref1, *ref2, *ref3; + ref0 = ref[0]; + ref1 = ref[1]; + ref2 = ref[2]; + ref3 = ref[3]; + + for (i = 0; i < 32; ++i) { + const uint8x16_t vec_src_00 = vld1q_u8(src); + const uint8x16_t vec_src_16 = vld1q_u8(src + 16); + + sad_neon_32(vec_src_00, vec_src_16, ref0, &vec_sum_ref0_lo, + &vec_sum_ref0_hi); + sad_neon_32(vec_src_00, vec_src_16, ref1, &vec_sum_ref1_lo, + &vec_sum_ref1_hi); + sad_neon_32(vec_src_00, vec_src_16, ref2, &vec_sum_ref2_lo, + &vec_sum_ref2_hi); + sad_neon_32(vec_src_00, vec_src_16, ref3, &vec_sum_ref3_lo, + &vec_sum_ref3_hi); + + src += src_stride; + ref0 += ref_stride; + ref1 += ref_stride; + ref2 += ref_stride; + ref3 += ref_stride; + } + + res[0] = horizontal_long_add_16x8(vec_sum_ref0_lo, vec_sum_ref0_hi); + res[1] = horizontal_long_add_16x8(vec_sum_ref1_lo, vec_sum_ref1_hi); + res[2] = horizontal_long_add_16x8(vec_sum_ref2_lo, vec_sum_ref2_hi); + res[3] = horizontal_long_add_16x8(vec_sum_ref3_lo, vec_sum_ref3_hi); +} + +void aom_sad16x16x4d_neon(const uint8_t *src, int src_stride, + const uint8_t *const ref[4], int ref_stride, + uint32_t *res) { + int i; + uint16x8_t vec_sum_ref0_lo = vdupq_n_u16(0); + uint16x8_t vec_sum_ref0_hi = vdupq_n_u16(0); + uint16x8_t vec_sum_ref1_lo = vdupq_n_u16(0); + uint16x8_t vec_sum_ref1_hi = vdupq_n_u16(0); + uint16x8_t vec_sum_ref2_lo = vdupq_n_u16(0); + uint16x8_t vec_sum_ref2_hi = vdupq_n_u16(0); + uint16x8_t vec_sum_ref3_lo = vdupq_n_u16(0); + uint16x8_t vec_sum_ref3_hi = vdupq_n_u16(0); + const uint8_t *ref0, *ref1, *ref2, *ref3; + ref0 = ref[0]; + ref1 = ref[1]; + ref2 = ref[2]; + ref3 = ref[3]; + + for (i = 0; i < 16; ++i) { + const uint8x16_t vec_src = vld1q_u8(src); + const uint8x16_t vec_ref0 = vld1q_u8(ref0); + const uint8x16_t vec_ref1 = vld1q_u8(ref1); + const uint8x16_t vec_ref2 = vld1q_u8(ref2); + const uint8x16_t vec_ref3 = vld1q_u8(ref3); + + vec_sum_ref0_lo = + vabal_u8(vec_sum_ref0_lo, vget_low_u8(vec_src), vget_low_u8(vec_ref0)); + vec_sum_ref0_hi = vabal_u8(vec_sum_ref0_hi, vget_high_u8(vec_src), + vget_high_u8(vec_ref0)); + vec_sum_ref1_lo = + vabal_u8(vec_sum_ref1_lo, vget_low_u8(vec_src), vget_low_u8(vec_ref1)); + vec_sum_ref1_hi = vabal_u8(vec_sum_ref1_hi, vget_high_u8(vec_src), + vget_high_u8(vec_ref1)); + vec_sum_ref2_lo = + vabal_u8(vec_sum_ref2_lo, vget_low_u8(vec_src), vget_low_u8(vec_ref2)); + vec_sum_ref2_hi = vabal_u8(vec_sum_ref2_hi, vget_high_u8(vec_src), + vget_high_u8(vec_ref2)); + vec_sum_ref3_lo = + vabal_u8(vec_sum_ref3_lo, vget_low_u8(vec_src), vget_low_u8(vec_ref3)); + vec_sum_ref3_hi = vabal_u8(vec_sum_ref3_hi, vget_high_u8(vec_src), + vget_high_u8(vec_ref3)); + + src += src_stride; + ref0 += ref_stride; + ref1 += ref_stride; + ref2 += ref_stride; + ref3 += ref_stride; + } + + res[0] = horizontal_long_add_16x8(vec_sum_ref0_lo, vec_sum_ref0_hi); + res[1] = horizontal_long_add_16x8(vec_sum_ref1_lo, vec_sum_ref1_hi); + res[2] = horizontal_long_add_16x8(vec_sum_ref2_lo, vec_sum_ref2_hi); + res[3] = horizontal_long_add_16x8(vec_sum_ref3_lo, vec_sum_ref3_hi); +} diff --git a/third_party/aom/aom_dsp/arm/sad_neon.c b/third_party/aom/aom_dsp/arm/sad_neon.c new file mode 100644 index 0000000000..a39de91d60 --- /dev/null +++ b/third_party/aom/aom_dsp/arm/sad_neon.c @@ -0,0 +1,224 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include + +#include "config/aom_config.h" + +#include "aom/aom_integer.h" + +unsigned int aom_sad8x16_neon(unsigned char *src_ptr, int src_stride, + unsigned char *ref_ptr, int ref_stride) { + uint8x8_t d0, d8; + uint16x8_t q12; + uint32x4_t q1; + uint64x2_t q3; + uint32x2_t d5; + int i; + + d0 = vld1_u8(src_ptr); + src_ptr += src_stride; + d8 = vld1_u8(ref_ptr); + ref_ptr += ref_stride; + q12 = vabdl_u8(d0, d8); + + for (i = 0; i < 15; i++) { + d0 = vld1_u8(src_ptr); + src_ptr += src_stride; + d8 = vld1_u8(ref_ptr); + ref_ptr += ref_stride; + q12 = vabal_u8(q12, d0, d8); + } + + q1 = vpaddlq_u16(q12); + q3 = vpaddlq_u32(q1); + d5 = vadd_u32(vreinterpret_u32_u64(vget_low_u64(q3)), + vreinterpret_u32_u64(vget_high_u64(q3))); + + return vget_lane_u32(d5, 0); +} + +unsigned int aom_sad4x4_neon(unsigned char *src_ptr, int src_stride, + unsigned char *ref_ptr, int ref_stride) { + uint8x8_t d0, d8; + uint16x8_t q12; + uint32x2_t d1; + uint64x1_t d3; + int i; + + d0 = vld1_u8(src_ptr); + src_ptr += src_stride; + d8 = vld1_u8(ref_ptr); + ref_ptr += ref_stride; + q12 = vabdl_u8(d0, d8); + + for (i = 0; i < 3; i++) { + d0 = vld1_u8(src_ptr); + src_ptr += src_stride; + d8 = vld1_u8(ref_ptr); + ref_ptr += ref_stride; + q12 = vabal_u8(q12, d0, d8); + } + + d1 = vpaddl_u16(vget_low_u16(q12)); + d3 = vpaddl_u32(d1); + + return vget_lane_u32(vreinterpret_u32_u64(d3), 0); +} + +unsigned int aom_sad16x8_neon(unsigned char *src_ptr, int src_stride, + unsigned char *ref_ptr, int ref_stride) { + uint8x16_t q0, q4; + uint16x8_t q12, q13; + uint32x4_t q1; + uint64x2_t q3; + uint32x2_t d5; + int i; + + q0 = vld1q_u8(src_ptr); + src_ptr += src_stride; + q4 = vld1q_u8(ref_ptr); + ref_ptr += ref_stride; + q12 = vabdl_u8(vget_low_u8(q0), vget_low_u8(q4)); + q13 = vabdl_u8(vget_high_u8(q0), vget_high_u8(q4)); + + for (i = 0; i < 7; i++) { + q0 = vld1q_u8(src_ptr); + src_ptr += src_stride; + q4 = vld1q_u8(ref_ptr); + ref_ptr += ref_stride; + q12 = vabal_u8(q12, vget_low_u8(q0), vget_low_u8(q4)); + q13 = vabal_u8(q13, vget_high_u8(q0), vget_high_u8(q4)); + } + + q12 = vaddq_u16(q12, q13); + q1 = vpaddlq_u16(q12); + q3 = vpaddlq_u32(q1); + d5 = vadd_u32(vreinterpret_u32_u64(vget_low_u64(q3)), + vreinterpret_u32_u64(vget_high_u64(q3))); + + return vget_lane_u32(d5, 0); +} + +static INLINE unsigned int horizontal_long_add_16x8(const uint16x8_t vec_lo, + const uint16x8_t vec_hi) { + const uint32x4_t vec_l_lo = + vaddl_u16(vget_low_u16(vec_lo), vget_high_u16(vec_lo)); + const uint32x4_t vec_l_hi = + vaddl_u16(vget_low_u16(vec_hi), vget_high_u16(vec_hi)); + const uint32x4_t a = vaddq_u32(vec_l_lo, vec_l_hi); + const uint64x2_t b = vpaddlq_u32(a); + const uint32x2_t c = vadd_u32(vreinterpret_u32_u64(vget_low_u64(b)), + vreinterpret_u32_u64(vget_high_u64(b))); + return vget_lane_u32(c, 0); +} +static INLINE unsigned int horizontal_add_16x8(const uint16x8_t vec_16x8) { + const uint32x4_t a = vpaddlq_u16(vec_16x8); + const uint64x2_t b = vpaddlq_u32(a); + const uint32x2_t c = vadd_u32(vreinterpret_u32_u64(vget_low_u64(b)), + vreinterpret_u32_u64(vget_high_u64(b))); + return vget_lane_u32(c, 0); +} + +unsigned int aom_sad64x64_neon(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride) { + int i; + uint16x8_t vec_accum_lo = vdupq_n_u16(0); + uint16x8_t vec_accum_hi = vdupq_n_u16(0); + for (i = 0; i < 64; ++i) { + const uint8x16_t vec_src_00 = vld1q_u8(src); + const uint8x16_t vec_src_16 = vld1q_u8(src + 16); + const uint8x16_t vec_src_32 = vld1q_u8(src + 32); + const uint8x16_t vec_src_48 = vld1q_u8(src + 48); + const uint8x16_t vec_ref_00 = vld1q_u8(ref); + const uint8x16_t vec_ref_16 = vld1q_u8(ref + 16); + const uint8x16_t vec_ref_32 = vld1q_u8(ref + 32); + const uint8x16_t vec_ref_48 = vld1q_u8(ref + 48); + src += src_stride; + ref += ref_stride; + vec_accum_lo = vabal_u8(vec_accum_lo, vget_low_u8(vec_src_00), + vget_low_u8(vec_ref_00)); + vec_accum_hi = vabal_u8(vec_accum_hi, vget_high_u8(vec_src_00), + vget_high_u8(vec_ref_00)); + vec_accum_lo = vabal_u8(vec_accum_lo, vget_low_u8(vec_src_16), + vget_low_u8(vec_ref_16)); + vec_accum_hi = vabal_u8(vec_accum_hi, vget_high_u8(vec_src_16), + vget_high_u8(vec_ref_16)); + vec_accum_lo = vabal_u8(vec_accum_lo, vget_low_u8(vec_src_32), + vget_low_u8(vec_ref_32)); + vec_accum_hi = vabal_u8(vec_accum_hi, vget_high_u8(vec_src_32), + vget_high_u8(vec_ref_32)); + vec_accum_lo = vabal_u8(vec_accum_lo, vget_low_u8(vec_src_48), + vget_low_u8(vec_ref_48)); + vec_accum_hi = vabal_u8(vec_accum_hi, vget_high_u8(vec_src_48), + vget_high_u8(vec_ref_48)); + } + return horizontal_long_add_16x8(vec_accum_lo, vec_accum_hi); +} + +unsigned int aom_sad32x32_neon(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride) { + int i; + uint16x8_t vec_accum_lo = vdupq_n_u16(0); + uint16x8_t vec_accum_hi = vdupq_n_u16(0); + + for (i = 0; i < 32; ++i) { + const uint8x16_t vec_src_00 = vld1q_u8(src); + const uint8x16_t vec_src_16 = vld1q_u8(src + 16); + const uint8x16_t vec_ref_00 = vld1q_u8(ref); + const uint8x16_t vec_ref_16 = vld1q_u8(ref + 16); + src += src_stride; + ref += ref_stride; + vec_accum_lo = vabal_u8(vec_accum_lo, vget_low_u8(vec_src_00), + vget_low_u8(vec_ref_00)); + vec_accum_hi = vabal_u8(vec_accum_hi, vget_high_u8(vec_src_00), + vget_high_u8(vec_ref_00)); + vec_accum_lo = vabal_u8(vec_accum_lo, vget_low_u8(vec_src_16), + vget_low_u8(vec_ref_16)); + vec_accum_hi = vabal_u8(vec_accum_hi, vget_high_u8(vec_src_16), + vget_high_u8(vec_ref_16)); + } + return horizontal_add_16x8(vaddq_u16(vec_accum_lo, vec_accum_hi)); +} + +unsigned int aom_sad16x16_neon(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride) { + int i; + uint16x8_t vec_accum_lo = vdupq_n_u16(0); + uint16x8_t vec_accum_hi = vdupq_n_u16(0); + + for (i = 0; i < 16; ++i) { + const uint8x16_t vec_src = vld1q_u8(src); + const uint8x16_t vec_ref = vld1q_u8(ref); + src += src_stride; + ref += ref_stride; + vec_accum_lo = + vabal_u8(vec_accum_lo, vget_low_u8(vec_src), vget_low_u8(vec_ref)); + vec_accum_hi = + vabal_u8(vec_accum_hi, vget_high_u8(vec_src), vget_high_u8(vec_ref)); + } + return horizontal_add_16x8(vaddq_u16(vec_accum_lo, vec_accum_hi)); +} + +unsigned int aom_sad8x8_neon(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride) { + int i; + uint16x8_t vec_accum = vdupq_n_u16(0); + + for (i = 0; i < 8; ++i) { + const uint8x8_t vec_src = vld1_u8(src); + const uint8x8_t vec_ref = vld1_u8(ref); + src += src_stride; + ref += ref_stride; + vec_accum = vabal_u8(vec_accum, vec_src, vec_ref); + } + return horizontal_add_16x8(vec_accum); +} diff --git a/third_party/aom/aom_dsp/arm/subpel_variance_neon.c b/third_party/aom/aom_dsp/arm/subpel_variance_neon.c new file mode 100644 index 0000000000..cf618eee77 --- /dev/null +++ b/third_party/aom/aom_dsp/arm/subpel_variance_neon.c @@ -0,0 +1,131 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include + +#include "config/aom_dsp_rtcd.h" +#include "config/aom_config.h" + +#include "aom_ports/mem.h" +#include "aom/aom_integer.h" + +#include "aom_dsp/aom_filter.h" +#include "aom_dsp/variance.h" + +static void var_filter_block2d_bil_w8(const uint8_t *src_ptr, + uint8_t *output_ptr, + unsigned int src_pixels_per_line, + int pixel_step, + unsigned int output_height, + unsigned int output_width, + const uint8_t *filter) { + const uint8x8_t f0 = vmov_n_u8(filter[0]); + const uint8x8_t f1 = vmov_n_u8(filter[1]); + unsigned int i; + for (i = 0; i < output_height; ++i) { + const uint8x8_t src_0 = vld1_u8(&src_ptr[0]); + const uint8x8_t src_1 = vld1_u8(&src_ptr[pixel_step]); + const uint16x8_t a = vmull_u8(src_0, f0); + const uint16x8_t b = vmlal_u8(a, src_1, f1); + const uint8x8_t out = vrshrn_n_u16(b, FILTER_BITS); + vst1_u8(&output_ptr[0], out); + // Next row... + src_ptr += src_pixels_per_line; + output_ptr += output_width; + } +} + +static void var_filter_block2d_bil_w16(const uint8_t *src_ptr, + uint8_t *output_ptr, + unsigned int src_pixels_per_line, + int pixel_step, + unsigned int output_height, + unsigned int output_width, + const uint8_t *filter) { + const uint8x8_t f0 = vmov_n_u8(filter[0]); + const uint8x8_t f1 = vmov_n_u8(filter[1]); + unsigned int i, j; + for (i = 0; i < output_height; ++i) { + for (j = 0; j < output_width; j += 16) { + const uint8x16_t src_0 = vld1q_u8(&src_ptr[j]); + const uint8x16_t src_1 = vld1q_u8(&src_ptr[j + pixel_step]); + const uint16x8_t a = vmull_u8(vget_low_u8(src_0), f0); + const uint16x8_t b = vmlal_u8(a, vget_low_u8(src_1), f1); + const uint8x8_t out_lo = vrshrn_n_u16(b, FILTER_BITS); + const uint16x8_t c = vmull_u8(vget_high_u8(src_0), f0); + const uint16x8_t d = vmlal_u8(c, vget_high_u8(src_1), f1); + const uint8x8_t out_hi = vrshrn_n_u16(d, FILTER_BITS); + vst1q_u8(&output_ptr[j], vcombine_u8(out_lo, out_hi)); + } + // Next row... + src_ptr += src_pixels_per_line; + output_ptr += output_width; + } +} + +unsigned int aom_sub_pixel_variance8x8_neon(const uint8_t *src, int src_stride, + int xoffset, int yoffset, + const uint8_t *dst, int dst_stride, + unsigned int *sse) { + DECLARE_ALIGNED(16, uint8_t, temp2[8 * 8]); + DECLARE_ALIGNED(16, uint8_t, fdata3[9 * 8]); + + var_filter_block2d_bil_w8(src, fdata3, src_stride, 1, 9, 8, + bilinear_filters_2t[xoffset]); + var_filter_block2d_bil_w8(fdata3, temp2, 8, 8, 8, 8, + bilinear_filters_2t[yoffset]); + return aom_variance8x8_neon(temp2, 8, dst, dst_stride, sse); +} + +unsigned int aom_sub_pixel_variance16x16_neon(const uint8_t *src, + int src_stride, int xoffset, + int yoffset, const uint8_t *dst, + int dst_stride, + unsigned int *sse) { + DECLARE_ALIGNED(16, uint8_t, temp2[16 * 16]); + DECLARE_ALIGNED(16, uint8_t, fdata3[17 * 16]); + + var_filter_block2d_bil_w16(src, fdata3, src_stride, 1, 17, 16, + bilinear_filters_2t[xoffset]); + var_filter_block2d_bil_w16(fdata3, temp2, 16, 16, 16, 16, + bilinear_filters_2t[yoffset]); + return aom_variance16x16_neon(temp2, 16, dst, dst_stride, sse); +} + +unsigned int aom_sub_pixel_variance32x32_neon(const uint8_t *src, + int src_stride, int xoffset, + int yoffset, const uint8_t *dst, + int dst_stride, + unsigned int *sse) { + DECLARE_ALIGNED(16, uint8_t, temp2[32 * 32]); + DECLARE_ALIGNED(16, uint8_t, fdata3[33 * 32]); + + var_filter_block2d_bil_w16(src, fdata3, src_stride, 1, 33, 32, + bilinear_filters_2t[xoffset]); + var_filter_block2d_bil_w16(fdata3, temp2, 32, 32, 32, 32, + bilinear_filters_2t[yoffset]); + return aom_variance32x32_neon(temp2, 32, dst, dst_stride, sse); +} + +unsigned int aom_sub_pixel_variance64x64_neon(const uint8_t *src, + int src_stride, int xoffset, + int yoffset, const uint8_t *dst, + int dst_stride, + unsigned int *sse) { + DECLARE_ALIGNED(16, uint8_t, temp2[64 * 64]); + DECLARE_ALIGNED(16, uint8_t, fdata3[65 * 64]); + + var_filter_block2d_bil_w16(src, fdata3, src_stride, 1, 65, 64, + bilinear_filters_2t[xoffset]); + var_filter_block2d_bil_w16(fdata3, temp2, 64, 64, 64, 64, + bilinear_filters_2t[yoffset]); + return aom_variance64x64_neon(temp2, 64, dst, dst_stride, sse); +} diff --git a/third_party/aom/aom_dsp/arm/subtract_neon.c b/third_party/aom/aom_dsp/arm/subtract_neon.c new file mode 100644 index 0000000000..28f5ace8e1 --- /dev/null +++ b/third_party/aom/aom_dsp/arm/subtract_neon.c @@ -0,0 +1,81 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include + +#include "config/aom_config.h" + +#include "aom/aom_integer.h" + +void aom_subtract_block_neon(int rows, int cols, int16_t *diff, + ptrdiff_t diff_stride, const uint8_t *src, + ptrdiff_t src_stride, const uint8_t *pred, + ptrdiff_t pred_stride) { + int r, c; + + if (cols > 16) { + for (r = 0; r < rows; ++r) { + for (c = 0; c < cols; c += 32) { + const uint8x16_t v_src_00 = vld1q_u8(&src[c + 0]); + const uint8x16_t v_src_16 = vld1q_u8(&src[c + 16]); + const uint8x16_t v_pred_00 = vld1q_u8(&pred[c + 0]); + const uint8x16_t v_pred_16 = vld1q_u8(&pred[c + 16]); + const uint16x8_t v_diff_lo_00 = + vsubl_u8(vget_low_u8(v_src_00), vget_low_u8(v_pred_00)); + const uint16x8_t v_diff_hi_00 = + vsubl_u8(vget_high_u8(v_src_00), vget_high_u8(v_pred_00)); + const uint16x8_t v_diff_lo_16 = + vsubl_u8(vget_low_u8(v_src_16), vget_low_u8(v_pred_16)); + const uint16x8_t v_diff_hi_16 = + vsubl_u8(vget_high_u8(v_src_16), vget_high_u8(v_pred_16)); + vst1q_s16(&diff[c + 0], vreinterpretq_s16_u16(v_diff_lo_00)); + vst1q_s16(&diff[c + 8], vreinterpretq_s16_u16(v_diff_hi_00)); + vst1q_s16(&diff[c + 16], vreinterpretq_s16_u16(v_diff_lo_16)); + vst1q_s16(&diff[c + 24], vreinterpretq_s16_u16(v_diff_hi_16)); + } + diff += diff_stride; + pred += pred_stride; + src += src_stride; + } + } else if (cols > 8) { + for (r = 0; r < rows; ++r) { + const uint8x16_t v_src = vld1q_u8(&src[0]); + const uint8x16_t v_pred = vld1q_u8(&pred[0]); + const uint16x8_t v_diff_lo = + vsubl_u8(vget_low_u8(v_src), vget_low_u8(v_pred)); + const uint16x8_t v_diff_hi = + vsubl_u8(vget_high_u8(v_src), vget_high_u8(v_pred)); + vst1q_s16(&diff[0], vreinterpretq_s16_u16(v_diff_lo)); + vst1q_s16(&diff[8], vreinterpretq_s16_u16(v_diff_hi)); + diff += diff_stride; + pred += pred_stride; + src += src_stride; + } + } else if (cols > 4) { + for (r = 0; r < rows; ++r) { + const uint8x8_t v_src = vld1_u8(&src[0]); + const uint8x8_t v_pred = vld1_u8(&pred[0]); + const uint16x8_t v_diff = vsubl_u8(v_src, v_pred); + vst1q_s16(&diff[0], vreinterpretq_s16_u16(v_diff)); + diff += diff_stride; + pred += pred_stride; + src += src_stride; + } + } else { + for (r = 0; r < rows; ++r) { + for (c = 0; c < cols; ++c) diff[c] = src[c] - pred[c]; + + diff += diff_stride; + pred += pred_stride; + src += src_stride; + } + } +} diff --git a/third_party/aom/aom_dsp/arm/variance_neon.c b/third_party/aom/aom_dsp/arm/variance_neon.c new file mode 100644 index 0000000000..74385a6010 --- /dev/null +++ b/third_party/aom/aom_dsp/arm/variance_neon.c @@ -0,0 +1,400 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include + +#include "config/aom_dsp_rtcd.h" +#include "config/aom_config.h" + +#include "aom/aom_integer.h" +#include "aom_ports/mem.h" + +static INLINE int horizontal_add_s16x8(const int16x8_t v_16x8) { + const int32x4_t a = vpaddlq_s16(v_16x8); + const int64x2_t b = vpaddlq_s32(a); + const int32x2_t c = vadd_s32(vreinterpret_s32_s64(vget_low_s64(b)), + vreinterpret_s32_s64(vget_high_s64(b))); + return vget_lane_s32(c, 0); +} + +static INLINE int horizontal_add_s32x4(const int32x4_t v_32x4) { + const int64x2_t b = vpaddlq_s32(v_32x4); + const int32x2_t c = vadd_s32(vreinterpret_s32_s64(vget_low_s64(b)), + vreinterpret_s32_s64(vget_high_s64(b))); + return vget_lane_s32(c, 0); +} + +// w * h must be less than 2048 or local variable v_sum may overflow. +static void variance_neon_w8(const uint8_t *a, int a_stride, const uint8_t *b, + int b_stride, int w, int h, uint32_t *sse, + int *sum) { + int i, j; + int16x8_t v_sum = vdupq_n_s16(0); + int32x4_t v_sse_lo = vdupq_n_s32(0); + int32x4_t v_sse_hi = vdupq_n_s32(0); + + for (i = 0; i < h; ++i) { + for (j = 0; j < w; j += 8) { + const uint8x8_t v_a = vld1_u8(&a[j]); + const uint8x8_t v_b = vld1_u8(&b[j]); + const uint16x8_t v_diff = vsubl_u8(v_a, v_b); + const int16x8_t sv_diff = vreinterpretq_s16_u16(v_diff); + v_sum = vaddq_s16(v_sum, sv_diff); + v_sse_lo = + vmlal_s16(v_sse_lo, vget_low_s16(sv_diff), vget_low_s16(sv_diff)); + v_sse_hi = + vmlal_s16(v_sse_hi, vget_high_s16(sv_diff), vget_high_s16(sv_diff)); + } + a += a_stride; + b += b_stride; + } + + *sum = horizontal_add_s16x8(v_sum); + *sse = (unsigned int)horizontal_add_s32x4(vaddq_s32(v_sse_lo, v_sse_hi)); +} + +void aom_get8x8var_neon(const uint8_t *a, int a_stride, const uint8_t *b, + int b_stride, unsigned int *sse, int *sum) { + variance_neon_w8(a, a_stride, b, b_stride, 8, 8, sse, sum); +} + +void aom_get16x16var_neon(const uint8_t *a, int a_stride, const uint8_t *b, + int b_stride, unsigned int *sse, int *sum) { + variance_neon_w8(a, a_stride, b, b_stride, 16, 16, sse, sum); +} + +unsigned int aom_variance8x8_neon(const uint8_t *a, int a_stride, + const uint8_t *b, int b_stride, + unsigned int *sse) { + int sum; + variance_neon_w8(a, a_stride, b, b_stride, 8, 8, sse, &sum); + return *sse - ((sum * sum) >> 6); +} + +unsigned int aom_variance16x16_neon(const uint8_t *a, int a_stride, + const uint8_t *b, int b_stride, + unsigned int *sse) { + int sum; + variance_neon_w8(a, a_stride, b, b_stride, 16, 16, sse, &sum); + return *sse - (((unsigned int)((int64_t)sum * sum)) >> 8); +} + +unsigned int aom_variance32x32_neon(const uint8_t *a, int a_stride, + const uint8_t *b, int b_stride, + unsigned int *sse) { + int sum; + variance_neon_w8(a, a_stride, b, b_stride, 32, 32, sse, &sum); + return *sse - (unsigned int)(((int64_t)sum * sum) >> 10); +} + +unsigned int aom_variance32x64_neon(const uint8_t *a, int a_stride, + const uint8_t *b, int b_stride, + unsigned int *sse) { + int sum1, sum2; + uint32_t sse1, sse2; + variance_neon_w8(a, a_stride, b, b_stride, 32, 32, &sse1, &sum1); + variance_neon_w8(a + (32 * a_stride), a_stride, b + (32 * b_stride), b_stride, + 32, 32, &sse2, &sum2); + *sse = sse1 + sse2; + sum1 += sum2; + return *sse - (unsigned int)(((int64_t)sum1 * sum1) >> 11); +} + +unsigned int aom_variance64x32_neon(const uint8_t *a, int a_stride, + const uint8_t *b, int b_stride, + unsigned int *sse) { + int sum1, sum2; + uint32_t sse1, sse2; + variance_neon_w8(a, a_stride, b, b_stride, 64, 16, &sse1, &sum1); + variance_neon_w8(a + (16 * a_stride), a_stride, b + (16 * b_stride), b_stride, + 64, 16, &sse2, &sum2); + *sse = sse1 + sse2; + sum1 += sum2; + return *sse - (unsigned int)(((int64_t)sum1 * sum1) >> 11); +} + +unsigned int aom_variance64x64_neon(const uint8_t *a, int a_stride, + const uint8_t *b, int b_stride, + unsigned int *sse) { + int sum1, sum2; + uint32_t sse1, sse2; + + variance_neon_w8(a, a_stride, b, b_stride, 64, 16, &sse1, &sum1); + variance_neon_w8(a + (16 * a_stride), a_stride, b + (16 * b_stride), b_stride, + 64, 16, &sse2, &sum2); + sse1 += sse2; + sum1 += sum2; + + variance_neon_w8(a + (16 * 2 * a_stride), a_stride, b + (16 * 2 * b_stride), + b_stride, 64, 16, &sse2, &sum2); + sse1 += sse2; + sum1 += sum2; + + variance_neon_w8(a + (16 * 3 * a_stride), a_stride, b + (16 * 3 * b_stride), + b_stride, 64, 16, &sse2, &sum2); + *sse = sse1 + sse2; + sum1 += sum2; + return *sse - (unsigned int)(((int64_t)sum1 * sum1) >> 12); +} + +unsigned int aom_variance16x8_neon(const unsigned char *src_ptr, + int source_stride, + const unsigned char *ref_ptr, + int recon_stride, unsigned int *sse) { + int i; + int16x4_t d22s16, d23s16, d24s16, d25s16, d26s16, d27s16, d28s16, d29s16; + uint32x2_t d0u32, d10u32; + int64x1_t d0s64, d1s64; + uint8x16_t q0u8, q1u8, q2u8, q3u8; + uint16x8_t q11u16, q12u16, q13u16, q14u16; + int32x4_t q8s32, q9s32, q10s32; + int64x2_t q0s64, q1s64, q5s64; + + q8s32 = vdupq_n_s32(0); + q9s32 = vdupq_n_s32(0); + q10s32 = vdupq_n_s32(0); + + for (i = 0; i < 4; i++) { + q0u8 = vld1q_u8(src_ptr); + src_ptr += source_stride; + q1u8 = vld1q_u8(src_ptr); + src_ptr += source_stride; + __builtin_prefetch(src_ptr); + + q2u8 = vld1q_u8(ref_ptr); + ref_ptr += recon_stride; + q3u8 = vld1q_u8(ref_ptr); + ref_ptr += recon_stride; + __builtin_prefetch(ref_ptr); + + q11u16 = vsubl_u8(vget_low_u8(q0u8), vget_low_u8(q2u8)); + q12u16 = vsubl_u8(vget_high_u8(q0u8), vget_high_u8(q2u8)); + q13u16 = vsubl_u8(vget_low_u8(q1u8), vget_low_u8(q3u8)); + q14u16 = vsubl_u8(vget_high_u8(q1u8), vget_high_u8(q3u8)); + + d22s16 = vreinterpret_s16_u16(vget_low_u16(q11u16)); + d23s16 = vreinterpret_s16_u16(vget_high_u16(q11u16)); + q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q11u16)); + q9s32 = vmlal_s16(q9s32, d22s16, d22s16); + q10s32 = vmlal_s16(q10s32, d23s16, d23s16); + + d24s16 = vreinterpret_s16_u16(vget_low_u16(q12u16)); + d25s16 = vreinterpret_s16_u16(vget_high_u16(q12u16)); + q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q12u16)); + q9s32 = vmlal_s16(q9s32, d24s16, d24s16); + q10s32 = vmlal_s16(q10s32, d25s16, d25s16); + + d26s16 = vreinterpret_s16_u16(vget_low_u16(q13u16)); + d27s16 = vreinterpret_s16_u16(vget_high_u16(q13u16)); + q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q13u16)); + q9s32 = vmlal_s16(q9s32, d26s16, d26s16); + q10s32 = vmlal_s16(q10s32, d27s16, d27s16); + + d28s16 = vreinterpret_s16_u16(vget_low_u16(q14u16)); + d29s16 = vreinterpret_s16_u16(vget_high_u16(q14u16)); + q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q14u16)); + q9s32 = vmlal_s16(q9s32, d28s16, d28s16); + q10s32 = vmlal_s16(q10s32, d29s16, d29s16); + } + + q10s32 = vaddq_s32(q10s32, q9s32); + q0s64 = vpaddlq_s32(q8s32); + q1s64 = vpaddlq_s32(q10s32); + + d0s64 = vadd_s64(vget_low_s64(q0s64), vget_high_s64(q0s64)); + d1s64 = vadd_s64(vget_low_s64(q1s64), vget_high_s64(q1s64)); + + q5s64 = vmull_s32(vreinterpret_s32_s64(d0s64), vreinterpret_s32_s64(d0s64)); + vst1_lane_u32((uint32_t *)sse, vreinterpret_u32_s64(d1s64), 0); + + d10u32 = vshr_n_u32(vreinterpret_u32_s64(vget_low_s64(q5s64)), 7); + d0u32 = vsub_u32(vreinterpret_u32_s64(d1s64), d10u32); + + return vget_lane_u32(d0u32, 0); +} + +unsigned int aom_variance8x16_neon(const unsigned char *src_ptr, + int source_stride, + const unsigned char *ref_ptr, + int recon_stride, unsigned int *sse) { + int i; + uint8x8_t d0u8, d2u8, d4u8, d6u8; + int16x4_t d22s16, d23s16, d24s16, d25s16; + uint32x2_t d0u32, d10u32; + int64x1_t d0s64, d1s64; + uint16x8_t q11u16, q12u16; + int32x4_t q8s32, q9s32, q10s32; + int64x2_t q0s64, q1s64, q5s64; + + q8s32 = vdupq_n_s32(0); + q9s32 = vdupq_n_s32(0); + q10s32 = vdupq_n_s32(0); + + for (i = 0; i < 8; i++) { + d0u8 = vld1_u8(src_ptr); + src_ptr += source_stride; + d2u8 = vld1_u8(src_ptr); + src_ptr += source_stride; + __builtin_prefetch(src_ptr); + + d4u8 = vld1_u8(ref_ptr); + ref_ptr += recon_stride; + d6u8 = vld1_u8(ref_ptr); + ref_ptr += recon_stride; + __builtin_prefetch(ref_ptr); + + q11u16 = vsubl_u8(d0u8, d4u8); + q12u16 = vsubl_u8(d2u8, d6u8); + + d22s16 = vreinterpret_s16_u16(vget_low_u16(q11u16)); + d23s16 = vreinterpret_s16_u16(vget_high_u16(q11u16)); + q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q11u16)); + q9s32 = vmlal_s16(q9s32, d22s16, d22s16); + q10s32 = vmlal_s16(q10s32, d23s16, d23s16); + + d24s16 = vreinterpret_s16_u16(vget_low_u16(q12u16)); + d25s16 = vreinterpret_s16_u16(vget_high_u16(q12u16)); + q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q12u16)); + q9s32 = vmlal_s16(q9s32, d24s16, d24s16); + q10s32 = vmlal_s16(q10s32, d25s16, d25s16); + } + + q10s32 = vaddq_s32(q10s32, q9s32); + q0s64 = vpaddlq_s32(q8s32); + q1s64 = vpaddlq_s32(q10s32); + + d0s64 = vadd_s64(vget_low_s64(q0s64), vget_high_s64(q0s64)); + d1s64 = vadd_s64(vget_low_s64(q1s64), vget_high_s64(q1s64)); + + q5s64 = vmull_s32(vreinterpret_s32_s64(d0s64), vreinterpret_s32_s64(d0s64)); + vst1_lane_u32((uint32_t *)sse, vreinterpret_u32_s64(d1s64), 0); + + d10u32 = vshr_n_u32(vreinterpret_u32_s64(vget_low_s64(q5s64)), 7); + d0u32 = vsub_u32(vreinterpret_u32_s64(d1s64), d10u32); + + return vget_lane_u32(d0u32, 0); +} + +unsigned int aom_mse16x16_neon(const unsigned char *src_ptr, int source_stride, + const unsigned char *ref_ptr, int recon_stride, + unsigned int *sse) { + int i; + int16x4_t d22s16, d23s16, d24s16, d25s16, d26s16, d27s16, d28s16, d29s16; + int64x1_t d0s64; + uint8x16_t q0u8, q1u8, q2u8, q3u8; + int32x4_t q7s32, q8s32, q9s32, q10s32; + uint16x8_t q11u16, q12u16, q13u16, q14u16; + int64x2_t q1s64; + + q7s32 = vdupq_n_s32(0); + q8s32 = vdupq_n_s32(0); + q9s32 = vdupq_n_s32(0); + q10s32 = vdupq_n_s32(0); + + for (i = 0; i < 8; i++) { // mse16x16_neon_loop + q0u8 = vld1q_u8(src_ptr); + src_ptr += source_stride; + q1u8 = vld1q_u8(src_ptr); + src_ptr += source_stride; + q2u8 = vld1q_u8(ref_ptr); + ref_ptr += recon_stride; + q3u8 = vld1q_u8(ref_ptr); + ref_ptr += recon_stride; + + q11u16 = vsubl_u8(vget_low_u8(q0u8), vget_low_u8(q2u8)); + q12u16 = vsubl_u8(vget_high_u8(q0u8), vget_high_u8(q2u8)); + q13u16 = vsubl_u8(vget_low_u8(q1u8), vget_low_u8(q3u8)); + q14u16 = vsubl_u8(vget_high_u8(q1u8), vget_high_u8(q3u8)); + + d22s16 = vreinterpret_s16_u16(vget_low_u16(q11u16)); + d23s16 = vreinterpret_s16_u16(vget_high_u16(q11u16)); + q7s32 = vmlal_s16(q7s32, d22s16, d22s16); + q8s32 = vmlal_s16(q8s32, d23s16, d23s16); + + d24s16 = vreinterpret_s16_u16(vget_low_u16(q12u16)); + d25s16 = vreinterpret_s16_u16(vget_high_u16(q12u16)); + q9s32 = vmlal_s16(q9s32, d24s16, d24s16); + q10s32 = vmlal_s16(q10s32, d25s16, d25s16); + + d26s16 = vreinterpret_s16_u16(vget_low_u16(q13u16)); + d27s16 = vreinterpret_s16_u16(vget_high_u16(q13u16)); + q7s32 = vmlal_s16(q7s32, d26s16, d26s16); + q8s32 = vmlal_s16(q8s32, d27s16, d27s16); + + d28s16 = vreinterpret_s16_u16(vget_low_u16(q14u16)); + d29s16 = vreinterpret_s16_u16(vget_high_u16(q14u16)); + q9s32 = vmlal_s16(q9s32, d28s16, d28s16); + q10s32 = vmlal_s16(q10s32, d29s16, d29s16); + } + + q7s32 = vaddq_s32(q7s32, q8s32); + q9s32 = vaddq_s32(q9s32, q10s32); + q10s32 = vaddq_s32(q7s32, q9s32); + + q1s64 = vpaddlq_s32(q10s32); + d0s64 = vadd_s64(vget_low_s64(q1s64), vget_high_s64(q1s64)); + + vst1_lane_u32((uint32_t *)sse, vreinterpret_u32_s64(d0s64), 0); + return vget_lane_u32(vreinterpret_u32_s64(d0s64), 0); +} + +unsigned int aom_get4x4sse_cs_neon(const unsigned char *src_ptr, + int source_stride, + const unsigned char *ref_ptr, + int recon_stride) { + int16x4_t d22s16, d24s16, d26s16, d28s16; + int64x1_t d0s64; + uint8x8_t d0u8, d1u8, d2u8, d3u8, d4u8, d5u8, d6u8, d7u8; + int32x4_t q7s32, q8s32, q9s32, q10s32; + uint16x8_t q11u16, q12u16, q13u16, q14u16; + int64x2_t q1s64; + + d0u8 = vld1_u8(src_ptr); + src_ptr += source_stride; + d4u8 = vld1_u8(ref_ptr); + ref_ptr += recon_stride; + d1u8 = vld1_u8(src_ptr); + src_ptr += source_stride; + d5u8 = vld1_u8(ref_ptr); + ref_ptr += recon_stride; + d2u8 = vld1_u8(src_ptr); + src_ptr += source_stride; + d6u8 = vld1_u8(ref_ptr); + ref_ptr += recon_stride; + d3u8 = vld1_u8(src_ptr); + src_ptr += source_stride; + d7u8 = vld1_u8(ref_ptr); + ref_ptr += recon_stride; + + q11u16 = vsubl_u8(d0u8, d4u8); + q12u16 = vsubl_u8(d1u8, d5u8); + q13u16 = vsubl_u8(d2u8, d6u8); + q14u16 = vsubl_u8(d3u8, d7u8); + + d22s16 = vget_low_s16(vreinterpretq_s16_u16(q11u16)); + d24s16 = vget_low_s16(vreinterpretq_s16_u16(q12u16)); + d26s16 = vget_low_s16(vreinterpretq_s16_u16(q13u16)); + d28s16 = vget_low_s16(vreinterpretq_s16_u16(q14u16)); + + q7s32 = vmull_s16(d22s16, d22s16); + q8s32 = vmull_s16(d24s16, d24s16); + q9s32 = vmull_s16(d26s16, d26s16); + q10s32 = vmull_s16(d28s16, d28s16); + + q7s32 = vaddq_s32(q7s32, q8s32); + q9s32 = vaddq_s32(q9s32, q10s32); + q9s32 = vaddq_s32(q7s32, q9s32); + + q1s64 = vpaddlq_s32(q9s32); + d0s64 = vadd_s64(vget_low_s64(q1s64), vget_high_s64(q1s64)); + + return vget_lane_u32(vreinterpret_u32_s64(d0s64), 0); +} diff --git a/third_party/aom/aom_dsp/binary_codes_reader.c b/third_party/aom/aom_dsp/binary_codes_reader.c new file mode 100644 index 0000000000..01088010a1 --- /dev/null +++ b/third_party/aom/aom_dsp/binary_codes_reader.c @@ -0,0 +1,123 @@ +/* + * Copyright (c) 2017, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include "aom_dsp/binary_codes_reader.h" + +#include "av1/common/common.h" + +// Inverse recenters a non-negative literal v around a reference r +static uint16_t inv_recenter_nonneg(uint16_t r, uint16_t v) { + if (v > (r << 1)) + return v; + else if ((v & 1) == 0) + return (v >> 1) + r; + else + return r - ((v + 1) >> 1); +} + +// Inverse recenters a non-negative literal v in [0, n-1] around a +// reference r also in [0, n-1] +static uint16_t inv_recenter_finite_nonneg(uint16_t n, uint16_t r, uint16_t v) { + if ((r << 1) <= n) { + return inv_recenter_nonneg(r, v); + } else { + return n - 1 - inv_recenter_nonneg(n - 1 - r, v); + } +} + +uint16_t aom_read_primitive_quniform_(aom_reader *r, + uint16_t n ACCT_STR_PARAM) { + if (n <= 1) return 0; + const int l = get_msb(n) + 1; + const int m = (1 << l) - n; + const int v = aom_read_literal(r, l - 1, ACCT_STR_NAME); + return v < m ? v : (v << 1) - m + aom_read_bit(r, ACCT_STR_NAME); +} + +static uint16_t aom_rb_read_primitive_quniform(struct aom_read_bit_buffer *rb, + uint16_t n) { + if (n <= 1) return 0; + const int l = get_msb(n) + 1; + const int m = (1 << l) - n; + const int v = aom_rb_read_literal(rb, l - 1); + return v < m ? v : (v << 1) - m + aom_rb_read_bit(rb); +} + +// Decode finite subexponential code that for a symbol v in [0, n-1] with +// parameter k +uint16_t aom_read_primitive_subexpfin_(aom_reader *r, uint16_t n, + uint16_t k ACCT_STR_PARAM) { + int i = 0; + int mk = 0; + + while (1) { + int b = (i ? k + i - 1 : k); + int a = (1 << b); + + if (n <= mk + 3 * a) { + return aom_read_primitive_quniform(r, n - mk, ACCT_STR_NAME) + mk; + } + + if (!aom_read_bit(r, ACCT_STR_NAME)) { + return aom_read_literal(r, b, ACCT_STR_NAME) + mk; + } + + i = i + 1; + mk += a; + } + + assert(0); + return 0; +} + +static uint16_t aom_rb_read_primitive_subexpfin(struct aom_read_bit_buffer *rb, + uint16_t n, uint16_t k) { + int i = 0; + int mk = 0; + + while (1) { + int b = (i ? k + i - 1 : k); + int a = (1 << b); + + if (n <= mk + 3 * a) { + return aom_rb_read_primitive_quniform(rb, n - mk) + mk; + } + + if (!aom_rb_read_bit(rb)) { + return aom_rb_read_literal(rb, b) + mk; + } + + i = i + 1; + mk += a; + } + + assert(0); + return 0; +} + +uint16_t aom_read_primitive_refsubexpfin_(aom_reader *r, uint16_t n, uint16_t k, + uint16_t ref ACCT_STR_PARAM) { + return inv_recenter_finite_nonneg( + n, ref, aom_read_primitive_subexpfin(r, n, k, ACCT_STR_NAME)); +} + +static uint16_t aom_rb_read_primitive_refsubexpfin( + struct aom_read_bit_buffer *rb, uint16_t n, uint16_t k, uint16_t ref) { + return inv_recenter_finite_nonneg(n, ref, + aom_rb_read_primitive_subexpfin(rb, n, k)); +} + +int16_t aom_rb_read_signed_primitive_refsubexpfin( + struct aom_read_bit_buffer *rb, uint16_t n, uint16_t k, int16_t ref) { + ref += n - 1; + const uint16_t scaled_n = (n << 1) - 1; + return aom_rb_read_primitive_refsubexpfin(rb, scaled_n, k, ref) - n + 1; +} diff --git a/third_party/aom/aom_dsp/binary_codes_reader.h b/third_party/aom/aom_dsp/binary_codes_reader.h new file mode 100644 index 0000000000..364a67469a --- /dev/null +++ b/third_party/aom/aom_dsp/binary_codes_reader.h @@ -0,0 +1,47 @@ +/* + * Copyright (c) 2017, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_BINARY_CODES_READER_H_ +#define AOM_AOM_DSP_BINARY_CODES_READER_H_ + +#ifdef __cplusplus +extern "C" { +#endif + +#include + +#include "config/aom_config.h" + +#include "aom/aom_integer.h" +#include "aom_dsp/bitreader.h" +#include "aom_dsp/bitreader_buffer.h" + +#define aom_read_primitive_quniform(r, n, ACCT_STR_NAME) \ + aom_read_primitive_quniform_(r, n ACCT_STR_ARG(ACCT_STR_NAME)) +#define aom_read_primitive_subexpfin(r, n, k, ACCT_STR_NAME) \ + aom_read_primitive_subexpfin_(r, n, k ACCT_STR_ARG(ACCT_STR_NAME)) +#define aom_read_primitive_refsubexpfin(r, n, k, ref, ACCT_STR_NAME) \ + aom_read_primitive_refsubexpfin_(r, n, k, ref ACCT_STR_ARG(ACCT_STR_NAME)) + +uint16_t aom_read_primitive_quniform_(aom_reader *r, uint16_t n ACCT_STR_PARAM); +uint16_t aom_read_primitive_subexpfin_(aom_reader *r, uint16_t n, + uint16_t k ACCT_STR_PARAM); +uint16_t aom_read_primitive_refsubexpfin_(aom_reader *r, uint16_t n, uint16_t k, + uint16_t ref ACCT_STR_PARAM); + +int16_t aom_rb_read_signed_primitive_refsubexpfin( + struct aom_read_bit_buffer *rb, uint16_t n, uint16_t k, int16_t ref); + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif // AOM_AOM_DSP_BINARY_CODES_READER_H_ diff --git a/third_party/aom/aom_dsp/binary_codes_writer.c b/third_party/aom/aom_dsp/binary_codes_writer.c new file mode 100644 index 0000000000..ee7a9f567e --- /dev/null +++ b/third_party/aom/aom_dsp/binary_codes_writer.c @@ -0,0 +1,210 @@ +/* + * Copyright (c) 2017, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include "aom_dsp/bitwriter.h" +#include "aom_dsp/binary_codes_writer.h" + +#include "av1/common/common.h" + +// Recenters a non-negative literal v around a reference r +static uint16_t recenter_nonneg(uint16_t r, uint16_t v) { + if (v > (r << 1)) + return v; + else if (v >= r) + return ((v - r) << 1); + else + return ((r - v) << 1) - 1; +} + +// Recenters a non-negative literal v in [0, n-1] around a +// reference r also in [0, n-1] +static uint16_t recenter_finite_nonneg(uint16_t n, uint16_t r, uint16_t v) { + if ((r << 1) <= n) { + return recenter_nonneg(r, v); + } else { + return recenter_nonneg(n - 1 - r, n - 1 - v); + } +} + +// Codes a symbol v in [-2^mag_bits, 2^mag_bits]. +// mag_bits is number of bits for magnitude. The alphabet is of size +// 2 * 2^mag_bits + 1, symmetric around 0, where one bit is used to +// indicate 0 or non-zero, mag_bits bits are used to indicate magnitide +// and 1 more bit for the sign if non-zero. +void aom_write_primitive_symmetric(aom_writer *w, int16_t v, + unsigned int abs_bits) { + if (v == 0) { + aom_write_bit(w, 0); + } else { + const int x = abs(v); + const int s = v < 0; + aom_write_bit(w, 1); + aom_write_bit(w, s); + aom_write_literal(w, x - 1, abs_bits); + } +} + +int aom_count_primitive_symmetric(int16_t v, unsigned int abs_bits) { + return (v == 0 ? 1 : abs_bits + 2); +} + +// Encodes a value v in [0, n-1] quasi-uniformly +void aom_write_primitive_quniform(aom_writer *w, uint16_t n, uint16_t v) { + if (n <= 1) return; + const int l = get_msb(n) + 1; + const int m = (1 << l) - n; + if (v < m) { + aom_write_literal(w, v, l - 1); + } else { + aom_write_literal(w, m + ((v - m) >> 1), l - 1); + aom_write_bit(w, (v - m) & 1); + } +} + +static void aom_wb_write_primitive_quniform(struct aom_write_bit_buffer *wb, + uint16_t n, uint16_t v) { + if (n <= 1) return; + const int l = get_msb(n) + 1; + const int m = (1 << l) - n; + if (v < m) { + aom_wb_write_literal(wb, v, l - 1); + } else { + aom_wb_write_literal(wb, m + ((v - m) >> 1), l - 1); + aom_wb_write_bit(wb, (v - m) & 1); + } +} + +int aom_count_primitive_quniform(uint16_t n, uint16_t v) { + if (n <= 1) return 0; + const int l = get_msb(n) + 1; + const int m = (1 << l) - n; + return v < m ? l - 1 : l; +} + +// Finite subexponential code that codes a symbol v in [0, n-1] with parameter k +void aom_write_primitive_subexpfin(aom_writer *w, uint16_t n, uint16_t k, + uint16_t v) { + int i = 0; + int mk = 0; + while (1) { + int b = (i ? k + i - 1 : k); + int a = (1 << b); + if (n <= mk + 3 * a) { + aom_write_primitive_quniform(w, n - mk, v - mk); + break; + } else { + int t = (v >= mk + a); + aom_write_bit(w, t); + if (t) { + i = i + 1; + mk += a; + } else { + aom_write_literal(w, v - mk, b); + break; + } + } + } +} + +static void aom_wb_write_primitive_subexpfin(struct aom_write_bit_buffer *wb, + uint16_t n, uint16_t k, + uint16_t v) { + int i = 0; + int mk = 0; + while (1) { + int b = (i ? k + i - 1 : k); + int a = (1 << b); + if (n <= mk + 3 * a) { + aom_wb_write_primitive_quniform(wb, n - mk, v - mk); + break; + } else { + int t = (v >= mk + a); + aom_wb_write_bit(wb, t); + if (t) { + i = i + 1; + mk += a; + } else { + aom_wb_write_literal(wb, v - mk, b); + break; + } + } + } +} + +int aom_count_primitive_subexpfin(uint16_t n, uint16_t k, uint16_t v) { + int count = 0; + int i = 0; + int mk = 0; + while (1) { + int b = (i ? k + i - 1 : k); + int a = (1 << b); + if (n <= mk + 3 * a) { + count += aom_count_primitive_quniform(n - mk, v - mk); + break; + } else { + int t = (v >= mk + a); + count++; + if (t) { + i = i + 1; + mk += a; + } else { + count += b; + break; + } + } + } + return count; +} + +// Finite subexponential code that codes a symbol v in [0, n-1] with parameter k +// based on a reference ref also in [0, n-1]. +// Recenters symbol around r first and then uses a finite subexponential code. +void aom_write_primitive_refsubexpfin(aom_writer *w, uint16_t n, uint16_t k, + uint16_t ref, uint16_t v) { + aom_write_primitive_subexpfin(w, n, k, recenter_finite_nonneg(n, ref, v)); +} + +static void aom_wb_write_primitive_refsubexpfin(struct aom_write_bit_buffer *wb, + uint16_t n, uint16_t k, + uint16_t ref, uint16_t v) { + aom_wb_write_primitive_subexpfin(wb, n, k, recenter_finite_nonneg(n, ref, v)); +} + +void aom_write_signed_primitive_refsubexpfin(aom_writer *w, uint16_t n, + uint16_t k, int16_t ref, + int16_t v) { + ref += n - 1; + v += n - 1; + const uint16_t scaled_n = (n << 1) - 1; + aom_write_primitive_refsubexpfin(w, scaled_n, k, ref, v); +} + +void aom_wb_write_signed_primitive_refsubexpfin(struct aom_write_bit_buffer *wb, + uint16_t n, uint16_t k, + int16_t ref, int16_t v) { + ref += n - 1; + v += n - 1; + const uint16_t scaled_n = (n << 1) - 1; + aom_wb_write_primitive_refsubexpfin(wb, scaled_n, k, ref, v); +} + +int aom_count_primitive_refsubexpfin(uint16_t n, uint16_t k, uint16_t ref, + uint16_t v) { + return aom_count_primitive_subexpfin(n, k, recenter_finite_nonneg(n, ref, v)); +} + +int aom_count_signed_primitive_refsubexpfin(uint16_t n, uint16_t k, int16_t ref, + int16_t v) { + ref += n - 1; + v += n - 1; + const uint16_t scaled_n = (n << 1) - 1; + return aom_count_primitive_refsubexpfin(scaled_n, k, ref, v); +} diff --git a/third_party/aom/aom_dsp/binary_codes_writer.h b/third_party/aom/aom_dsp/binary_codes_writer.h new file mode 100644 index 0000000000..c360e0e298 --- /dev/null +++ b/third_party/aom/aom_dsp/binary_codes_writer.h @@ -0,0 +1,68 @@ +/* + * Copyright (c) 2017, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_BINARY_CODES_WRITER_H_ +#define AOM_AOM_DSP_BINARY_CODES_WRITER_H_ + +#ifdef __cplusplus +extern "C" { +#endif + +#include +#include "config/aom_config.h" + +#include "aom/aom_integer.h" +#include "aom_dsp/bitwriter.h" +#include "aom_dsp/bitwriter_buffer.h" + +// Codes a symbol v in [-2^mag_bits, 2^mag_bits] +// mag_bits is number of bits for magnitude. The alphabet is of size +// 2 * 2^mag_bits + 1, symmetric around 0, where one bit is used to +// indicate 0 or non-zero, mag_bits bits are used to indicate magnitide +// and 1 more bit for the sign if non-zero. +void aom_write_primitive_symmetric(aom_writer *w, int16_t v, + unsigned int mag_bits); + +// Encodes a value v in [0, n-1] quasi-uniformly +void aom_write_primitive_quniform(aom_writer *w, uint16_t n, uint16_t v); + +// Finite subexponential code that codes a symbol v in [0, n-1] with parameter k +void aom_write_primitive_subexpfin(aom_writer *w, uint16_t n, uint16_t k, + uint16_t v); + +// Finite subexponential code that codes a symbol v in [0, n-1] with parameter k +// based on a reference ref also in [0, n-1]. +void aom_write_primitive_refsubexpfin(aom_writer *w, uint16_t n, uint16_t k, + uint16_t ref, uint16_t v); + +// Finite subexponential code that codes a symbol v in [-(n-1), n-1] with +// parameter k based on a reference ref also in [-(n-1), n-1]. +void aom_write_signed_primitive_refsubexpfin(aom_writer *w, uint16_t n, + uint16_t k, int16_t ref, + int16_t v); + +void aom_wb_write_signed_primitive_refsubexpfin(struct aom_write_bit_buffer *wb, + uint16_t n, uint16_t k, + int16_t ref, int16_t v); + +// Functions that counts bits for the above primitives +int aom_count_primitive_symmetric(int16_t v, unsigned int mag_bits); +int aom_count_primitive_quniform(uint16_t n, uint16_t v); +int aom_count_primitive_subexpfin(uint16_t n, uint16_t k, uint16_t v); +int aom_count_primitive_refsubexpfin(uint16_t n, uint16_t k, uint16_t ref, + uint16_t v); +int aom_count_signed_primitive_refsubexpfin(uint16_t n, uint16_t k, int16_t ref, + int16_t v); +#ifdef __cplusplus +} // extern "C" +#endif + +#endif // AOM_AOM_DSP_BINARY_CODES_WRITER_H_ diff --git a/third_party/aom/aom_dsp/bitreader.h b/third_party/aom/aom_dsp/bitreader.h new file mode 100644 index 0000000000..7c0efcc785 --- /dev/null +++ b/third_party/aom/aom_dsp/bitreader.h @@ -0,0 +1,160 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_BITREADER_H_ +#define AOM_AOM_DSP_BITREADER_H_ + +#include +#include + +#include "config/aom_config.h" + +#include "aom/aomdx.h" +#include "aom/aom_integer.h" +#include "aom_dsp/daalaboolreader.h" +#include "aom_dsp/prob.h" +#include "av1/common/odintrin.h" + +#if CONFIG_ACCOUNTING +#include "av1/decoder/accounting.h" +#define ACCT_STR_NAME acct_str +#define ACCT_STR_PARAM , const char *ACCT_STR_NAME +#define ACCT_STR_ARG(s) , s +#else +#define ACCT_STR_PARAM +#define ACCT_STR_ARG(s) +#endif + +#define aom_read(r, prob, ACCT_STR_NAME) \ + aom_read_(r, prob ACCT_STR_ARG(ACCT_STR_NAME)) +#define aom_read_bit(r, ACCT_STR_NAME) \ + aom_read_bit_(r ACCT_STR_ARG(ACCT_STR_NAME)) +#define aom_read_tree(r, tree, probs, ACCT_STR_NAME) \ + aom_read_tree_(r, tree, probs ACCT_STR_ARG(ACCT_STR_NAME)) +#define aom_read_literal(r, bits, ACCT_STR_NAME) \ + aom_read_literal_(r, bits ACCT_STR_ARG(ACCT_STR_NAME)) +#define aom_read_cdf(r, cdf, nsymbs, ACCT_STR_NAME) \ + aom_read_cdf_(r, cdf, nsymbs ACCT_STR_ARG(ACCT_STR_NAME)) +#define aom_read_symbol(r, cdf, nsymbs, ACCT_STR_NAME) \ + aom_read_symbol_(r, cdf, nsymbs ACCT_STR_ARG(ACCT_STR_NAME)) + +#ifdef __cplusplus +extern "C" { +#endif + +typedef struct daala_reader aom_reader; + +static INLINE int aom_reader_init(aom_reader *r, const uint8_t *buffer, + size_t size) { + return aom_daala_reader_init(r, buffer, (int)size); +} + +static INLINE const uint8_t *aom_reader_find_begin(aom_reader *r) { + return aom_daala_reader_find_begin(r); +} + +static INLINE const uint8_t *aom_reader_find_end(aom_reader *r) { + return aom_daala_reader_find_end(r); +} + +static INLINE int aom_reader_has_error(aom_reader *r) { + return aom_daala_reader_has_error(r); +} + +// Returns true if the bit reader has tried to decode more data from the buffer +// than was actually provided. +static INLINE int aom_reader_has_overflowed(const aom_reader *r) { + return aom_daala_reader_has_overflowed(r); +} + +// Returns the position in the bit reader in bits. +static INLINE uint32_t aom_reader_tell(const aom_reader *r) { + return aom_daala_reader_tell(r); +} + +// Returns the position in the bit reader in 1/8th bits. +static INLINE uint32_t aom_reader_tell_frac(const aom_reader *r) { + return aom_daala_reader_tell_frac(r); +} + +#if CONFIG_ACCOUNTING +static INLINE void aom_process_accounting(const aom_reader *r ACCT_STR_PARAM) { + if (r->accounting != NULL) { + uint32_t tell_frac; + tell_frac = aom_reader_tell_frac(r); + aom_accounting_record(r->accounting, ACCT_STR_NAME, + tell_frac - r->accounting->last_tell_frac); + r->accounting->last_tell_frac = tell_frac; + } +} + +static INLINE void aom_update_symb_counts(const aom_reader *r, int is_binary) { + if (r->accounting != NULL) { + r->accounting->syms.num_multi_syms += !is_binary; + r->accounting->syms.num_binary_syms += !!is_binary; + } +} +#endif + +static INLINE int aom_read_(aom_reader *r, int prob ACCT_STR_PARAM) { + int ret; + ret = aom_daala_read(r, prob); +#if CONFIG_ACCOUNTING + if (ACCT_STR_NAME) aom_process_accounting(r, ACCT_STR_NAME); + aom_update_symb_counts(r, 1); +#endif + return ret; +} + +static INLINE int aom_read_bit_(aom_reader *r ACCT_STR_PARAM) { + int ret; + ret = aom_read(r, 128, NULL); // aom_prob_half +#if CONFIG_ACCOUNTING + if (ACCT_STR_NAME) aom_process_accounting(r, ACCT_STR_NAME); +#endif + return ret; +} + +static INLINE int aom_read_literal_(aom_reader *r, int bits ACCT_STR_PARAM) { + int literal = 0, bit; + + for (bit = bits - 1; bit >= 0; bit--) literal |= aom_read_bit(r, NULL) << bit; +#if CONFIG_ACCOUNTING + if (ACCT_STR_NAME) aom_process_accounting(r, ACCT_STR_NAME); +#endif + return literal; +} + +static INLINE int aom_read_cdf_(aom_reader *r, const aom_cdf_prob *cdf, + int nsymbs ACCT_STR_PARAM) { + int ret; + ret = daala_read_symbol(r, cdf, nsymbs); + +#if CONFIG_ACCOUNTING + if (ACCT_STR_NAME) aom_process_accounting(r, ACCT_STR_NAME); + aom_update_symb_counts(r, (nsymbs == 2)); +#endif + return ret; +} + +static INLINE int aom_read_symbol_(aom_reader *r, aom_cdf_prob *cdf, + int nsymbs ACCT_STR_PARAM) { + int ret; + ret = aom_read_cdf(r, cdf, nsymbs, ACCT_STR_NAME); + if (r->allow_update_cdf) update_cdf(cdf, ret, nsymbs); + return ret; +} + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif // AOM_AOM_DSP_BITREADER_H_ diff --git a/third_party/aom/aom_dsp/bitreader_buffer.c b/third_party/aom/aom_dsp/bitreader_buffer.c new file mode 100644 index 0000000000..b532117849 --- /dev/null +++ b/third_party/aom/aom_dsp/bitreader_buffer.c @@ -0,0 +1,67 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include + +#include "config/aom_config.h" + +#include "aom_dsp/bitreader_buffer.h" + +size_t aom_rb_bytes_read(const struct aom_read_bit_buffer *rb) { + return (rb->bit_offset + 7) >> 3; +} + +int aom_rb_read_bit(struct aom_read_bit_buffer *rb) { + const uint32_t off = rb->bit_offset; + const uint32_t p = off >> 3; + const int q = 7 - (int)(off & 0x7); + if (rb->bit_buffer + p < rb->bit_buffer_end) { + const int bit = (rb->bit_buffer[p] >> q) & 1; + rb->bit_offset = off + 1; + return bit; + } else { + if (rb->error_handler) rb->error_handler(rb->error_handler_data); + return 0; + } +} + +int aom_rb_read_literal(struct aom_read_bit_buffer *rb, int bits) { + assert(bits <= 31); + int value = 0, bit; + for (bit = bits - 1; bit >= 0; bit--) value |= aom_rb_read_bit(rb) << bit; + return value; +} + +uint32_t aom_rb_read_unsigned_literal(struct aom_read_bit_buffer *rb, + int bits) { + assert(bits <= 32); + uint32_t value = 0; + int bit; + for (bit = bits - 1; bit >= 0; bit--) + value |= (uint32_t)aom_rb_read_bit(rb) << bit; + return value; +} + +int aom_rb_read_inv_signed_literal(struct aom_read_bit_buffer *rb, int bits) { + const int nbits = sizeof(unsigned) * 8 - bits - 1; + const unsigned value = (unsigned)aom_rb_read_literal(rb, bits + 1) << nbits; + return ((int)value) >> nbits; +} + +uint32_t aom_rb_read_uvlc(struct aom_read_bit_buffer *rb) { + int leading_zeros = 0; + while (!aom_rb_read_bit(rb)) ++leading_zeros; + // Maximum 32 bits. + if (leading_zeros >= 32) return UINT32_MAX; + const uint32_t base = (1u << leading_zeros) - 1; + const uint32_t value = aom_rb_read_literal(rb, leading_zeros); + return base + value; +} diff --git a/third_party/aom/aom_dsp/bitreader_buffer.h b/third_party/aom/aom_dsp/bitreader_buffer.h new file mode 100644 index 0000000000..725ca1ea2b --- /dev/null +++ b/third_party/aom/aom_dsp/bitreader_buffer.h @@ -0,0 +1,50 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_BITREADER_BUFFER_H_ +#define AOM_AOM_DSP_BITREADER_BUFFER_H_ + +#include + +#include "aom/aom_integer.h" + +#ifdef __cplusplus +extern "C" { +#endif + +typedef void (*aom_rb_error_handler)(void *data); + +struct aom_read_bit_buffer { + const uint8_t *bit_buffer; + const uint8_t *bit_buffer_end; + uint32_t bit_offset; + + void *error_handler_data; + aom_rb_error_handler error_handler; +}; + +size_t aom_rb_bytes_read(const struct aom_read_bit_buffer *rb); + +int aom_rb_read_bit(struct aom_read_bit_buffer *rb); + +int aom_rb_read_literal(struct aom_read_bit_buffer *rb, int bits); + +uint32_t aom_rb_read_unsigned_literal(struct aom_read_bit_buffer *rb, int bits); + +int aom_rb_read_inv_signed_literal(struct aom_read_bit_buffer *rb, int bits); + +uint32_t aom_rb_read_uvlc(struct aom_read_bit_buffer *rb); + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif // AOM_AOM_DSP_BITREADER_BUFFER_H_ diff --git a/third_party/aom/aom_dsp/bitwriter.h b/third_party/aom/aom_dsp/bitwriter.h new file mode 100644 index 0000000000..b5ecc2382b --- /dev/null +++ b/third_party/aom/aom_dsp/bitwriter.h @@ -0,0 +1,89 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_BITWRITER_H_ +#define AOM_AOM_DSP_BITWRITER_H_ + +#include + +#include "config/aom_config.h" + +#include "aom_dsp/daalaboolwriter.h" +#include "aom_dsp/prob.h" + +#if CONFIG_RD_DEBUG +#include "av1/common/blockd.h" +#include "av1/encoder/cost.h" +#endif + +#ifdef __cplusplus +extern "C" { +#endif + +typedef struct daala_writer aom_writer; + +typedef struct TOKEN_STATS { + int cost; +#if CONFIG_RD_DEBUG + int txb_coeff_cost_map[TXB_COEFF_COST_MAP_SIZE][TXB_COEFF_COST_MAP_SIZE]; +#endif +} TOKEN_STATS; + +static INLINE void init_token_stats(TOKEN_STATS *token_stats) { +#if CONFIG_RD_DEBUG + int r, c; + for (r = 0; r < TXB_COEFF_COST_MAP_SIZE; ++r) { + for (c = 0; c < TXB_COEFF_COST_MAP_SIZE; ++c) { + token_stats->txb_coeff_cost_map[r][c] = 0; + } + } +#endif + token_stats->cost = 0; +} + +static INLINE void aom_start_encode(aom_writer *bc, uint8_t *buffer) { + aom_daala_start_encode(bc, buffer); +} + +static INLINE int aom_stop_encode(aom_writer *bc) { + return aom_daala_stop_encode(bc); +} + +static INLINE void aom_write(aom_writer *br, int bit, int probability) { + aom_daala_write(br, bit, probability); +} + +static INLINE void aom_write_bit(aom_writer *w, int bit) { + aom_write(w, bit, 128); // aom_prob_half +} + +static INLINE void aom_write_literal(aom_writer *w, int data, int bits) { + int bit; + + for (bit = bits - 1; bit >= 0; bit--) aom_write_bit(w, 1 & (data >> bit)); +} + +static INLINE void aom_write_cdf(aom_writer *w, int symb, + const aom_cdf_prob *cdf, int nsymbs) { + daala_write_symbol(w, symb, cdf, nsymbs); +} + +static INLINE void aom_write_symbol(aom_writer *w, int symb, aom_cdf_prob *cdf, + int nsymbs) { + aom_write_cdf(w, symb, cdf, nsymbs); + if (w->allow_update_cdf) update_cdf(cdf, symb, nsymbs); +} + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif // AOM_AOM_DSP_BITWRITER_H_ diff --git a/third_party/aom/aom_dsp/bitwriter_buffer.c b/third_party/aom/aom_dsp/bitwriter_buffer.c new file mode 100644 index 0000000000..596246deb2 --- /dev/null +++ b/third_party/aom/aom_dsp/bitwriter_buffer.c @@ -0,0 +1,87 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include +#include +#include + +#include "config/aom_config.h" + +#include "aom_dsp/bitwriter_buffer.h" + +int aom_wb_is_byte_aligned(const struct aom_write_bit_buffer *wb) { + return (wb->bit_offset % CHAR_BIT == 0); +} + +uint32_t aom_wb_bytes_written(const struct aom_write_bit_buffer *wb) { + return wb->bit_offset / CHAR_BIT + (wb->bit_offset % CHAR_BIT > 0); +} + +void aom_wb_write_bit(struct aom_write_bit_buffer *wb, int bit) { + const int off = (int)wb->bit_offset; + const int p = off / CHAR_BIT; + const int q = CHAR_BIT - 1 - off % CHAR_BIT; + if (q == CHAR_BIT - 1) { + // Zero next char and write bit + wb->bit_buffer[p] = bit << q; + } else { + wb->bit_buffer[p] &= ~(1 << q); + wb->bit_buffer[p] |= bit << q; + } + wb->bit_offset = off + 1; +} + +void aom_wb_overwrite_bit(struct aom_write_bit_buffer *wb, int bit) { + // Do not zero bytes but overwrite exisiting values + const int off = (int)wb->bit_offset; + const int p = off / CHAR_BIT; + const int q = CHAR_BIT - 1 - off % CHAR_BIT; + wb->bit_buffer[p] &= ~(1 << q); + wb->bit_buffer[p] |= bit << q; + wb->bit_offset = off + 1; +} + +void aom_wb_write_literal(struct aom_write_bit_buffer *wb, int data, int bits) { + assert(bits <= 31); + int bit; + for (bit = bits - 1; bit >= 0; bit--) aom_wb_write_bit(wb, (data >> bit) & 1); +} + +void aom_wb_write_unsigned_literal(struct aom_write_bit_buffer *wb, + uint32_t data, int bits) { + assert(bits <= 32); + int bit; + for (bit = bits - 1; bit >= 0; bit--) aom_wb_write_bit(wb, (data >> bit) & 1); +} + +void aom_wb_overwrite_literal(struct aom_write_bit_buffer *wb, int data, + int bits) { + int bit; + for (bit = bits - 1; bit >= 0; bit--) + aom_wb_overwrite_bit(wb, (data >> bit) & 1); +} + +void aom_wb_write_inv_signed_literal(struct aom_write_bit_buffer *wb, int data, + int bits) { + aom_wb_write_literal(wb, data, bits + 1); +} + +void aom_wb_write_uvlc(struct aom_write_bit_buffer *wb, uint32_t v) { + int64_t shift_val = ++v; + int leading_zeroes = 1; + + assert(shift_val > 0); + + while (shift_val >>= 1) leading_zeroes += 2; + + aom_wb_write_literal(wb, 0, leading_zeroes >> 1); + aom_wb_write_unsigned_literal(wb, v, (leading_zeroes + 1) >> 1); +} diff --git a/third_party/aom/aom_dsp/bitwriter_buffer.h b/third_party/aom/aom_dsp/bitwriter_buffer.h new file mode 100644 index 0000000000..d0311284f2 --- /dev/null +++ b/third_party/aom/aom_dsp/bitwriter_buffer.h @@ -0,0 +1,51 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_BITWRITER_BUFFER_H_ +#define AOM_AOM_DSP_BITWRITER_BUFFER_H_ + +#include "aom/aom_integer.h" + +#ifdef __cplusplus +extern "C" { +#endif + +struct aom_write_bit_buffer { + uint8_t *bit_buffer; + uint32_t bit_offset; +}; + +int aom_wb_is_byte_aligned(const struct aom_write_bit_buffer *wb); + +uint32_t aom_wb_bytes_written(const struct aom_write_bit_buffer *wb); + +void aom_wb_write_bit(struct aom_write_bit_buffer *wb, int bit); + +void aom_wb_overwrite_bit(struct aom_write_bit_buffer *wb, int bit); + +void aom_wb_write_literal(struct aom_write_bit_buffer *wb, int data, int bits); + +void aom_wb_write_unsigned_literal(struct aom_write_bit_buffer *wb, + uint32_t data, int bits); + +void aom_wb_overwrite_literal(struct aom_write_bit_buffer *wb, int data, + int bits); + +void aom_wb_write_inv_signed_literal(struct aom_write_bit_buffer *wb, int data, + int bits); + +void aom_wb_write_uvlc(struct aom_write_bit_buffer *wb, uint32_t v); + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif // AOM_AOM_DSP_BITWRITER_BUFFER_H_ diff --git a/third_party/aom/aom_dsp/blend.h b/third_party/aom/aom_dsp/blend.h new file mode 100644 index 0000000000..fd87dc1810 --- /dev/null +++ b/third_party/aom/aom_dsp/blend.h @@ -0,0 +1,45 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_BLEND_H_ +#define AOM_AOM_DSP_BLEND_H_ + +#include "aom_ports/mem.h" + +// Various blending functions and macros. +// See also the aom_blend_* functions in aom_dsp_rtcd.h + +// Alpha blending with alpha values from the range [0, 64], where 64 +// means use the first input and 0 means use the second input. + +#define AOM_BLEND_A64_ROUND_BITS 6 +#define AOM_BLEND_A64_MAX_ALPHA (1 << AOM_BLEND_A64_ROUND_BITS) // 64 + +#define AOM_BLEND_A64(a, v0, v1) \ + ROUND_POWER_OF_TWO((a) * (v0) + (AOM_BLEND_A64_MAX_ALPHA - (a)) * (v1), \ + AOM_BLEND_A64_ROUND_BITS) + +// Alpha blending with alpha values from the range [0, 256], where 256 +// means use the first input and 0 means use the second input. +#define AOM_BLEND_A256_ROUND_BITS 8 +#define AOM_BLEND_A256_MAX_ALPHA (1 << AOM_BLEND_A256_ROUND_BITS) // 256 + +#define AOM_BLEND_A256(a, v0, v1) \ + ROUND_POWER_OF_TWO((a) * (v0) + (AOM_BLEND_A256_MAX_ALPHA - (a)) * (v1), \ + AOM_BLEND_A256_ROUND_BITS) + +// Blending by averaging. +#define AOM_BLEND_AVG(v0, v1) ROUND_POWER_OF_TWO((v0) + (v1), 1) + +#define DIFF_FACTOR_LOG2 4 +#define DIFF_FACTOR (1 << DIFF_FACTOR_LOG2) + +#endif // AOM_AOM_DSP_BLEND_H_ diff --git a/third_party/aom/aom_dsp/blend_a64_hmask.c b/third_party/aom/aom_dsp/blend_a64_hmask.c new file mode 100644 index 0000000000..0554b43d1a --- /dev/null +++ b/third_party/aom/aom_dsp/blend_a64_hmask.c @@ -0,0 +1,69 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include + +#include "aom/aom_integer.h" +#include "aom_ports/mem.h" +#include "aom_dsp/aom_dsp_common.h" +#include "aom_dsp/blend.h" + +#include "config/aom_dsp_rtcd.h" + +void aom_blend_a64_hmask_c(uint8_t *dst, uint32_t dst_stride, + const uint8_t *src0, uint32_t src0_stride, + const uint8_t *src1, uint32_t src1_stride, + const uint8_t *mask, int w, int h) { + int i, j; + + assert(IMPLIES(src0 == dst, src0_stride == dst_stride)); + assert(IMPLIES(src1 == dst, src1_stride == dst_stride)); + + assert(h >= 1); + assert(w >= 1); + assert(IS_POWER_OF_TWO(h)); + assert(IS_POWER_OF_TWO(w)); + + for (i = 0; i < h; ++i) { + for (j = 0; j < w; ++j) { + dst[i * dst_stride + j] = AOM_BLEND_A64( + mask[j], src0[i * src0_stride + j], src1[i * src1_stride + j]); + } + } +} + +void aom_highbd_blend_a64_hmask_c(uint8_t *dst_8, uint32_t dst_stride, + const uint8_t *src0_8, uint32_t src0_stride, + const uint8_t *src1_8, uint32_t src1_stride, + const uint8_t *mask, int w, int h, int bd) { + int i, j; + uint16_t *dst = CONVERT_TO_SHORTPTR(dst_8); + const uint16_t *src0 = CONVERT_TO_SHORTPTR(src0_8); + const uint16_t *src1 = CONVERT_TO_SHORTPTR(src1_8); + (void)bd; + + assert(IMPLIES(src0 == dst, src0_stride == dst_stride)); + assert(IMPLIES(src1 == dst, src1_stride == dst_stride)); + + assert(h >= 1); + assert(w >= 1); + assert(IS_POWER_OF_TWO(h)); + assert(IS_POWER_OF_TWO(w)); + + assert(bd == 8 || bd == 10 || bd == 12); + + for (i = 0; i < h; ++i) { + for (j = 0; j < w; ++j) { + dst[i * dst_stride + j] = AOM_BLEND_A64( + mask[j], src0[i * src0_stride + j], src1[i * src1_stride + j]); + } + } +} diff --git a/third_party/aom/aom_dsp/blend_a64_mask.c b/third_party/aom/aom_dsp/blend_a64_mask.c new file mode 100644 index 0000000000..992cc5c0c4 --- /dev/null +++ b/third_party/aom/aom_dsp/blend_a64_mask.c @@ -0,0 +1,345 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include + +#include "aom/aom_integer.h" +#include "aom_ports/mem.h" +#include "aom_dsp/blend.h" +#include "aom_dsp/aom_dsp_common.h" + +#include "config/aom_dsp_rtcd.h" + +// Blending with alpha mask. Mask values come from the range [0, 64], +// as described for AOM_BLEND_A64 in aom_dsp/blend.h. src0 or src1 can +// be the same as dst, or dst can be different from both sources. + +// NOTE(david.barker): The input and output of aom_blend_a64_d32_mask_c() are +// in a higher intermediate precision, and will later be rounded down to pixel +// precision. +// Thus, in order to avoid double-rounding, we want to use normal right shifts +// within this function, not ROUND_POWER_OF_TWO. +// This works because of the identity: +// ROUND_POWER_OF_TWO(x >> y, z) == ROUND_POWER_OF_TWO(x, y+z) +// +// In contrast, the output of the non-d32 functions will not be further rounded, +// so we *should* use ROUND_POWER_OF_TWO there. + +void aom_lowbd_blend_a64_d16_mask_c( + uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, + uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, + ConvolveParams *conv_params) { + int i, j; + const int bd = 8; + const int offset_bits = bd + 2 * FILTER_BITS - conv_params->round_0; + const int round_offset = (1 << (offset_bits - conv_params->round_1)) + + (1 << (offset_bits - conv_params->round_1 - 1)); + const int round_bits = + 2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1; + + assert(IMPLIES((void *)src0 == dst, src0_stride == dst_stride)); + assert(IMPLIES((void *)src1 == dst, src1_stride == dst_stride)); + + assert(h >= 4); + assert(w >= 4); + assert(IS_POWER_OF_TWO(h)); + assert(IS_POWER_OF_TWO(w)); + + if (subw == 0 && subh == 0) { + for (i = 0; i < h; ++i) { + for (j = 0; j < w; ++j) { + int32_t res; + const int m = mask[i * mask_stride + j]; + res = ((m * (int32_t)src0[i * src0_stride + j] + + (AOM_BLEND_A64_MAX_ALPHA - m) * + (int32_t)src1[i * src1_stride + j]) >> + AOM_BLEND_A64_ROUND_BITS); + res -= round_offset; + dst[i * dst_stride + j] = + clip_pixel(ROUND_POWER_OF_TWO(res, round_bits)); + } + } + } else if (subw == 1 && subh == 1) { + for (i = 0; i < h; ++i) { + for (j = 0; j < w; ++j) { + int32_t res; + const int m = ROUND_POWER_OF_TWO( + mask[(2 * i) * mask_stride + (2 * j)] + + mask[(2 * i + 1) * mask_stride + (2 * j)] + + mask[(2 * i) * mask_stride + (2 * j + 1)] + + mask[(2 * i + 1) * mask_stride + (2 * j + 1)], + 2); + res = ((m * (int32_t)src0[i * src0_stride + j] + + (AOM_BLEND_A64_MAX_ALPHA - m) * + (int32_t)src1[i * src1_stride + j]) >> + AOM_BLEND_A64_ROUND_BITS); + res -= round_offset; + dst[i * dst_stride + j] = + clip_pixel(ROUND_POWER_OF_TWO(res, round_bits)); + } + } + } else if (subw == 1 && subh == 0) { + for (i = 0; i < h; ++i) { + for (j = 0; j < w; ++j) { + int32_t res; + const int m = AOM_BLEND_AVG(mask[i * mask_stride + (2 * j)], + mask[i * mask_stride + (2 * j + 1)]); + res = ((m * (int32_t)src0[i * src0_stride + j] + + (AOM_BLEND_A64_MAX_ALPHA - m) * + (int32_t)src1[i * src1_stride + j]) >> + AOM_BLEND_A64_ROUND_BITS); + res -= round_offset; + dst[i * dst_stride + j] = + clip_pixel(ROUND_POWER_OF_TWO(res, round_bits)); + } + } + } else { + for (i = 0; i < h; ++i) { + for (j = 0; j < w; ++j) { + int32_t res; + const int m = AOM_BLEND_AVG(mask[(2 * i) * mask_stride + j], + mask[(2 * i + 1) * mask_stride + j]); + res = ((int32_t)(m * (int32_t)src0[i * src0_stride + j] + + (AOM_BLEND_A64_MAX_ALPHA - m) * + (int32_t)src1[i * src1_stride + j]) >> + AOM_BLEND_A64_ROUND_BITS); + res -= round_offset; + dst[i * dst_stride + j] = + clip_pixel(ROUND_POWER_OF_TWO(res, round_bits)); + } + } + } +} + +void aom_highbd_blend_a64_d16_mask_c( + uint8_t *dst_8, uint32_t dst_stride, const CONV_BUF_TYPE *src0, + uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, + ConvolveParams *conv_params, const int bd) { + const int offset_bits = bd + 2 * FILTER_BITS - conv_params->round_0; + const int round_offset = (1 << (offset_bits - conv_params->round_1)) + + (1 << (offset_bits - conv_params->round_1 - 1)); + const int round_bits = + 2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1; + uint16_t *dst = CONVERT_TO_SHORTPTR(dst_8); + + assert(IMPLIES(src0 == dst, src0_stride == dst_stride)); + assert(IMPLIES(src1 == dst, src1_stride == dst_stride)); + + assert(h >= 1); + assert(w >= 1); + assert(IS_POWER_OF_TWO(h)); + assert(IS_POWER_OF_TWO(w)); + + // excerpt from clip_pixel_highbd() + // set saturation_value to (1 << bd) - 1 + unsigned int saturation_value; + switch (bd) { + case 8: + default: saturation_value = 255; break; + case 10: saturation_value = 1023; break; + case 12: saturation_value = 4095; break; + } + + if (subw == 0 && subh == 0) { + for (int i = 0; i < h; ++i) { + for (int j = 0; j < w; ++j) { + int32_t res; + const int m = mask[j]; + res = ((m * src0[j] + (AOM_BLEND_A64_MAX_ALPHA - m) * src1[j]) >> + AOM_BLEND_A64_ROUND_BITS); + res -= round_offset; + unsigned int v = negative_to_zero(ROUND_POWER_OF_TWO(res, round_bits)); + dst[j] = AOMMIN(v, saturation_value); + } + mask += mask_stride; + src0 += src0_stride; + src1 += src1_stride; + dst += dst_stride; + } + } else if (subw == 1 && subh == 1) { + for (int i = 0; i < h; ++i) { + for (int j = 0; j < w; ++j) { + int32_t res; + const int m = ROUND_POWER_OF_TWO( + mask[2 * j] + mask[mask_stride + 2 * j] + mask[2 * j + 1] + + mask[mask_stride + 2 * j + 1], + 2); + res = (m * src0[j] + (AOM_BLEND_A64_MAX_ALPHA - m) * src1[j]) >> + AOM_BLEND_A64_ROUND_BITS; + res -= round_offset; + unsigned int v = negative_to_zero(ROUND_POWER_OF_TWO(res, round_bits)); + dst[j] = AOMMIN(v, saturation_value); + } + mask += 2 * mask_stride; + src0 += src0_stride; + src1 += src1_stride; + dst += dst_stride; + } + } else if (subw == 1 && subh == 0) { + for (int i = 0; i < h; ++i) { + for (int j = 0; j < w; ++j) { + int32_t res; + const int m = AOM_BLEND_AVG(mask[2 * j], mask[2 * j + 1]); + res = (m * src0[j] + (AOM_BLEND_A64_MAX_ALPHA - m) * src1[j]) >> + AOM_BLEND_A64_ROUND_BITS; + res -= round_offset; + unsigned int v = negative_to_zero(ROUND_POWER_OF_TWO(res, round_bits)); + dst[j] = AOMMIN(v, saturation_value); + } + mask += mask_stride; + src0 += src0_stride; + src1 += src1_stride; + dst += dst_stride; + } + } else { + for (int i = 0; i < h; ++i) { + for (int j = 0; j < w; ++j) { + int32_t res; + const int m = AOM_BLEND_AVG(mask[j], mask[mask_stride + j]); + res = (m * src0[j] + (AOM_BLEND_A64_MAX_ALPHA - m) * src1[j]) >> + AOM_BLEND_A64_ROUND_BITS; + res -= round_offset; + unsigned int v = negative_to_zero(ROUND_POWER_OF_TWO(res, round_bits)); + dst[j] = AOMMIN(v, saturation_value); + } + mask += 2 * mask_stride; + src0 += src0_stride; + src1 += src1_stride; + dst += dst_stride; + } + } +} + +// Blending with alpha mask. Mask values come from the range [0, 64], +// as described for AOM_BLEND_A64 in aom_dsp/blend.h. src0 or src1 can +// be the same as dst, or dst can be different from both sources. + +void aom_blend_a64_mask_c(uint8_t *dst, uint32_t dst_stride, + const uint8_t *src0, uint32_t src0_stride, + const uint8_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int w, + int h, int subw, int subh) { + int i, j; + + assert(IMPLIES(src0 == dst, src0_stride == dst_stride)); + assert(IMPLIES(src1 == dst, src1_stride == dst_stride)); + + assert(h >= 1); + assert(w >= 1); + assert(IS_POWER_OF_TWO(h)); + assert(IS_POWER_OF_TWO(w)); + + if (subw == 0 && subh == 0) { + for (i = 0; i < h; ++i) { + for (j = 0; j < w; ++j) { + const int m = mask[i * mask_stride + j]; + dst[i * dst_stride + j] = AOM_BLEND_A64(m, src0[i * src0_stride + j], + src1[i * src1_stride + j]); + } + } + } else if (subw == 1 && subh == 1) { + for (i = 0; i < h; ++i) { + for (j = 0; j < w; ++j) { + const int m = ROUND_POWER_OF_TWO( + mask[(2 * i) * mask_stride + (2 * j)] + + mask[(2 * i + 1) * mask_stride + (2 * j)] + + mask[(2 * i) * mask_stride + (2 * j + 1)] + + mask[(2 * i + 1) * mask_stride + (2 * j + 1)], + 2); + dst[i * dst_stride + j] = AOM_BLEND_A64(m, src0[i * src0_stride + j], + src1[i * src1_stride + j]); + } + } + } else if (subw == 1 && subh == 0) { + for (i = 0; i < h; ++i) { + for (j = 0; j < w; ++j) { + const int m = AOM_BLEND_AVG(mask[i * mask_stride + (2 * j)], + mask[i * mask_stride + (2 * j + 1)]); + dst[i * dst_stride + j] = AOM_BLEND_A64(m, src0[i * src0_stride + j], + src1[i * src1_stride + j]); + } + } + } else { + for (i = 0; i < h; ++i) { + for (j = 0; j < w; ++j) { + const int m = AOM_BLEND_AVG(mask[(2 * i) * mask_stride + j], + mask[(2 * i + 1) * mask_stride + j]); + dst[i * dst_stride + j] = AOM_BLEND_A64(m, src0[i * src0_stride + j], + src1[i * src1_stride + j]); + } + } + } +} + +void aom_highbd_blend_a64_mask_c(uint8_t *dst_8, uint32_t dst_stride, + const uint8_t *src0_8, uint32_t src0_stride, + const uint8_t *src1_8, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, + int w, int h, int subw, int subh, int bd) { + int i, j; + uint16_t *dst = CONVERT_TO_SHORTPTR(dst_8); + const uint16_t *src0 = CONVERT_TO_SHORTPTR(src0_8); + const uint16_t *src1 = CONVERT_TO_SHORTPTR(src1_8); + (void)bd; + + assert(IMPLIES(src0 == dst, src0_stride == dst_stride)); + assert(IMPLIES(src1 == dst, src1_stride == dst_stride)); + + assert(h >= 1); + assert(w >= 1); + assert(IS_POWER_OF_TWO(h)); + assert(IS_POWER_OF_TWO(w)); + + assert(bd == 8 || bd == 10 || bd == 12); + + if (subw == 0 && subh == 0) { + for (i = 0; i < h; ++i) { + for (j = 0; j < w; ++j) { + const int m = mask[i * mask_stride + j]; + dst[i * dst_stride + j] = AOM_BLEND_A64(m, src0[i * src0_stride + j], + src1[i * src1_stride + j]); + } + } + } else if (subw == 1 && subh == 1) { + for (i = 0; i < h; ++i) { + for (j = 0; j < w; ++j) { + const int m = ROUND_POWER_OF_TWO( + mask[(2 * i) * mask_stride + (2 * j)] + + mask[(2 * i + 1) * mask_stride + (2 * j)] + + mask[(2 * i) * mask_stride + (2 * j + 1)] + + mask[(2 * i + 1) * mask_stride + (2 * j + 1)], + 2); + dst[i * dst_stride + j] = AOM_BLEND_A64(m, src0[i * src0_stride + j], + src1[i * src1_stride + j]); + } + } + } else if (subw == 1 && subh == 0) { + for (i = 0; i < h; ++i) { + for (j = 0; j < w; ++j) { + const int m = AOM_BLEND_AVG(mask[i * mask_stride + (2 * j)], + mask[i * mask_stride + (2 * j + 1)]); + dst[i * dst_stride + j] = AOM_BLEND_A64(m, src0[i * src0_stride + j], + src1[i * src1_stride + j]); + } + } + } else { + for (i = 0; i < h; ++i) { + for (j = 0; j < w; ++j) { + const int m = AOM_BLEND_AVG(mask[(2 * i) * mask_stride + j], + mask[(2 * i + 1) * mask_stride + j]); + dst[i * dst_stride + j] = AOM_BLEND_A64(m, src0[i * src0_stride + j], + src1[i * src1_stride + j]); + } + } + } +} diff --git a/third_party/aom/aom_dsp/blend_a64_vmask.c b/third_party/aom/aom_dsp/blend_a64_vmask.c new file mode 100644 index 0000000000..4f222e17f0 --- /dev/null +++ b/third_party/aom/aom_dsp/blend_a64_vmask.c @@ -0,0 +1,71 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include + +#include "aom/aom_integer.h" +#include "aom_ports/mem.h" +#include "aom_dsp/aom_dsp_common.h" +#include "aom_dsp/blend.h" + +#include "config/aom_dsp_rtcd.h" + +void aom_blend_a64_vmask_c(uint8_t *dst, uint32_t dst_stride, + const uint8_t *src0, uint32_t src0_stride, + const uint8_t *src1, uint32_t src1_stride, + const uint8_t *mask, int w, int h) { + int i, j; + + assert(IMPLIES(src0 == dst, src0_stride == dst_stride)); + assert(IMPLIES(src1 == dst, src1_stride == dst_stride)); + + assert(h >= 1); + assert(w >= 1); + assert(IS_POWER_OF_TWO(h)); + assert(IS_POWER_OF_TWO(w)); + + for (i = 0; i < h; ++i) { + const int m = mask[i]; + for (j = 0; j < w; ++j) { + dst[i * dst_stride + j] = AOM_BLEND_A64(m, src0[i * src0_stride + j], + src1[i * src1_stride + j]); + } + } +} + +void aom_highbd_blend_a64_vmask_c(uint8_t *dst_8, uint32_t dst_stride, + const uint8_t *src0_8, uint32_t src0_stride, + const uint8_t *src1_8, uint32_t src1_stride, + const uint8_t *mask, int w, int h, int bd) { + int i, j; + uint16_t *dst = CONVERT_TO_SHORTPTR(dst_8); + const uint16_t *src0 = CONVERT_TO_SHORTPTR(src0_8); + const uint16_t *src1 = CONVERT_TO_SHORTPTR(src1_8); + (void)bd; + + assert(IMPLIES(src0 == dst, src0_stride == dst_stride)); + assert(IMPLIES(src1 == dst, src1_stride == dst_stride)); + + assert(h >= 1); + assert(w >= 1); + assert(IS_POWER_OF_TWO(h)); + assert(IS_POWER_OF_TWO(w)); + + assert(bd == 8 || bd == 10 || bd == 12); + + for (i = 0; i < h; ++i) { + const int m = mask[i]; + for (j = 0; j < w; ++j) { + dst[i * dst_stride + j] = AOM_BLEND_A64(m, src0[i * src0_stride + j], + src1[i * src1_stride + j]); + } + } +} diff --git a/third_party/aom/aom_dsp/buf_ans.c b/third_party/aom/aom_dsp/buf_ans.c new file mode 100644 index 0000000000..f7703dffc6 --- /dev/null +++ b/third_party/aom/aom_dsp/buf_ans.c @@ -0,0 +1,70 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include + +#include "aom_dsp/buf_ans.h" +#include "aom_mem/aom_mem.h" +#include "aom/internal/aom_codec_internal.h" + +void aom_buf_ans_alloc(struct BufAnsCoder *c, + struct aom_internal_error_info *error) { + c->error = error; + assert(c->size > 1); + AOM_CHECK_MEM_ERROR(error, c->buf, aom_malloc(c->size * sizeof(*c->buf))); + // Initialize to overfull to trigger the assert in write. + c->offset = c->size + 1; +} + +void aom_buf_ans_free(struct BufAnsCoder *c) { + aom_free(c->buf); + c->buf = NULL; + c->size = 0; +} + +#if !ANS_MAX_SYMBOLS +void aom_buf_ans_grow(struct BufAnsCoder *c) { + struct buffered_ans_symbol *new_buf = NULL; + int new_size = c->size * 2; + AOM_CHECK_MEM_ERROR(c->error, new_buf, + aom_malloc(new_size * sizeof(*new_buf))); + memcpy(new_buf, c->buf, c->size * sizeof(*c->buf)); + aom_free(c->buf); + c->buf = new_buf; + c->size = new_size; +} +#endif + +void aom_buf_ans_flush(struct BufAnsCoder *const c) { + int offset; +#if ANS_MAX_SYMBOLS + if (c->offset == 0) return; +#endif + assert(c->offset > 0); + offset = c->offset - 1; + // Code the first symbol such that it brings the state to the smallest normal + // state from an initial state that would have been a subnormal/refill state. + if (c->buf[offset].method == ANS_METHOD_RANS) { + c->ans.state += c->buf[offset].val_start; + } else { + c->ans.state += c->buf[offset].val_start ? c->buf[offset].prob : 0; + } + for (offset = offset - 1; offset >= 0; --offset) { + if (c->buf[offset].method == ANS_METHOD_RANS) { + rans_write(&c->ans, c->buf[offset].val_start, c->buf[offset].prob); + } else { + rabs_write(&c->ans, (uint8_t)c->buf[offset].val_start, + (AnsP8)c->buf[offset].prob); + } + } + c->offset = 0; + c->output_bytes += ans_write_end(&c->ans); +} diff --git a/third_party/aom/aom_dsp/buf_ans.h b/third_party/aom/aom_dsp/buf_ans.h new file mode 100644 index 0000000000..985fcdf9e5 --- /dev/null +++ b/third_party/aom/aom_dsp/buf_ans.h @@ -0,0 +1,136 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_BUF_ANS_H_ +#define AOM_AOM_DSP_BUF_ANS_H_ +// Buffered forward ANS writer. +// Symbols are written to the writer in forward (decode) order and serialized +// backwards due to ANS's stack like behavior. + +#include +#include "config/aom_config.h" + +#include "aom/aom_integer.h" +#include "aom_dsp/ans.h" +#include "aom_dsp/answriter.h" + +#ifdef __cplusplus +extern "C" { +#endif // __cplusplus + +#define ANS_METHOD_RABS 0 +#define ANS_METHOD_RANS 1 + +struct buffered_ans_symbol { + unsigned int method : 1; // one of ANS_METHOD_RABS or ANS_METHOD_RANS + // TODO(aconverse): Should be possible to write this in terms of start for ABS + unsigned int val_start : RANS_PROB_BITS; // Boolean value for ABS + // start in symbol cycle for Rans + unsigned int prob : RANS_PROB_BITS; // Probability of this symbol +}; + +struct BufAnsCoder { + struct aom_internal_error_info *error; + struct buffered_ans_symbol *buf; + struct AnsCoder ans; + int size; + int offset; + int output_bytes; +#if ANS_MAX_SYMBOLS + int window_size; +#endif + int pos; // Dummy variable to store the output buffer after closing + uint8_t allow_update_cdf; +}; + +// Allocate a buffered ANS coder to store size symbols. +// When ANS_MAX_SYMBOLS is turned on, the size is the fixed size of each ANS +// partition. +// When ANS_MAX_SYMBOLS is turned off, size is merely an initial hint and the +// buffer will grow on demand +void aom_buf_ans_alloc(struct BufAnsCoder *c, + struct aom_internal_error_info *error); + +void aom_buf_ans_free(struct BufAnsCoder *c); + +#if !ANS_MAX_SYMBOLS +void aom_buf_ans_grow(struct BufAnsCoder *c); +#endif + +void aom_buf_ans_flush(struct BufAnsCoder *const c); + +static INLINE void buf_ans_write_init(struct BufAnsCoder *const c, + uint8_t *const output_buffer) { + c->offset = 0; + c->output_bytes = 0; + ans_write_init(&c->ans, output_buffer); +} + +static INLINE void buf_rabs_write(struct BufAnsCoder *const c, uint8_t val, + AnsP8 prob) { + assert(c->offset <= c->size); +#if !ANS_MAX_SYMBOLS + if (c->offset == c->size) { + aom_buf_ans_grow(c); + } +#endif + c->buf[c->offset].method = ANS_METHOD_RABS; + c->buf[c->offset].val_start = val; + c->buf[c->offset].prob = prob; + ++c->offset; +#if ANS_MAX_SYMBOLS + if (c->offset == c->size) aom_buf_ans_flush(c); +#endif +} + +// Buffer one symbol for encoding using rANS. +// cum_prob: The cumulative probability before this symbol (the offset of +// the symbol in the symbol cycle) +// prob: The probability of this symbol (l_s from the paper) +// RANS_PRECISION takes the place of m from the paper. +static INLINE void buf_rans_write(struct BufAnsCoder *const c, + aom_cdf_prob cum_prob, aom_cdf_prob prob) { + assert(c->offset <= c->size); +#if !ANS_MAX_SYMBOLS + if (c->offset == c->size) { + aom_buf_ans_grow(c); + } +#endif + c->buf[c->offset].method = ANS_METHOD_RANS; + c->buf[c->offset].val_start = cum_prob; + c->buf[c->offset].prob = prob; + ++c->offset; +#if ANS_MAX_SYMBOLS + if (c->offset == c->size) aom_buf_ans_flush(c); +#endif +} + +static INLINE void buf_rabs_write_bit(struct BufAnsCoder *c, int bit) { + buf_rabs_write(c, bit, 128); +} + +static INLINE void buf_rabs_write_literal(struct BufAnsCoder *c, int literal, + int bits) { + int bit; + + assert(bits < 31); + for (bit = bits - 1; bit >= 0; bit--) + buf_rabs_write_bit(c, 1 & (literal >> bit)); +} + +static INLINE int buf_ans_write_end(struct BufAnsCoder *const c) { + assert(c->offset == 0); + return c->output_bytes; +} +#ifdef __cplusplus +} // extern "C" +#endif // __cplusplus +#endif // AOM_AOM_DSP_BUF_ANS_H_ diff --git a/third_party/aom/aom_dsp/daalaboolreader.c b/third_party/aom/aom_dsp/daalaboolreader.c new file mode 100644 index 0000000000..6c2259f233 --- /dev/null +++ b/third_party/aom/aom_dsp/daalaboolreader.c @@ -0,0 +1,47 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include "aom_dsp/daalaboolreader.h" + +int aom_daala_reader_init(daala_reader *r, const uint8_t *buffer, int size) { + if (size && !buffer) { + return 1; + } + r->buffer_end = buffer + size; + r->buffer = buffer; + od_ec_dec_init(&r->ec, buffer, size); +#if CONFIG_ACCOUNTING + r->accounting = NULL; +#endif + return 0; +} + +const uint8_t *aom_daala_reader_find_begin(daala_reader *r) { + return r->buffer; +} + +const uint8_t *aom_daala_reader_find_end(daala_reader *r) { + return r->buffer_end; +} + +uint32_t aom_daala_reader_tell(const daala_reader *r) { + return od_ec_dec_tell(&r->ec); +} + +uint32_t aom_daala_reader_tell_frac(const daala_reader *r) { + return od_ec_dec_tell_frac(&r->ec); +} + +int aom_daala_reader_has_overflowed(const daala_reader *r) { + const uint32_t tell_bits = aom_daala_reader_tell(r); + const uint32_t tell_bytes = (tell_bits + 7) >> 3; + return ((ptrdiff_t)tell_bytes > r->buffer_end - r->buffer); +} diff --git a/third_party/aom/aom_dsp/daalaboolreader.h b/third_party/aom/aom_dsp/daalaboolreader.h new file mode 100644 index 0000000000..ba78f916d7 --- /dev/null +++ b/third_party/aom/aom_dsp/daalaboolreader.h @@ -0,0 +1,160 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_DAALABOOLREADER_H_ +#define AOM_AOM_DSP_DAALABOOLREADER_H_ + +#include "aom/aom_integer.h" +#include "aom_dsp/entdec.h" +#include "aom_dsp/prob.h" +#if CONFIG_ACCOUNTING +#include "av1/decoder/accounting.h" +#endif +#if CONFIG_BITSTREAM_DEBUG +#include +#include "aom_util/debug_util.h" +#endif // CONFIG_BITSTREAM_DEBUG + +#ifdef __cplusplus +extern "C" { +#endif + +struct daala_reader { + const uint8_t *buffer; + const uint8_t *buffer_end; + od_ec_dec ec; +#if CONFIG_ACCOUNTING + Accounting *accounting; +#endif + uint8_t allow_update_cdf; +}; + +typedef struct daala_reader daala_reader; + +int aom_daala_reader_init(daala_reader *r, const uint8_t *buffer, int size); +const uint8_t *aom_daala_reader_find_begin(daala_reader *r); +const uint8_t *aom_daala_reader_find_end(daala_reader *r); +uint32_t aom_daala_reader_tell(const daala_reader *r); +uint32_t aom_daala_reader_tell_frac(const daala_reader *r); +// Returns true if the reader has tried to decode more data from the buffer +// than was actually provided. +int aom_daala_reader_has_overflowed(const daala_reader *r); + +static INLINE int aom_daala_read(daala_reader *r, int prob) { + int bit; + int p = (0x7FFFFF - (prob << 15) + prob) >> 8; +#if CONFIG_BITSTREAM_DEBUG +/*{ + const int queue_r = bitstream_queue_get_read(); + const int frame_idx = bitstream_queue_get_frame_read(); + if (frame_idx == 0 && queue_r == 0) { + fprintf(stderr, "\n *** bitstream queue at frame_idx_r %d queue_r %d\n", + frame_idx, queue_r); + } +}*/ +#endif + + bit = od_ec_decode_bool_q15(&r->ec, p); + +#if CONFIG_BITSTREAM_DEBUG + { + int i; + int ref_bit, ref_nsymbs; + aom_cdf_prob ref_cdf[16]; + const int queue_r = bitstream_queue_get_read(); + const int frame_idx = bitstream_queue_get_frame_read(); + bitstream_queue_pop(&ref_bit, ref_cdf, &ref_nsymbs); + if (ref_nsymbs != 2) { + fprintf(stderr, + "\n *** [bit] nsymbs error, frame_idx_r %d nsymbs %d ref_nsymbs " + "%d queue_r %d\n", + frame_idx, 2, ref_nsymbs, queue_r); + assert(0); + } + if ((ref_nsymbs != 2) || (ref_cdf[0] != (aom_cdf_prob)p) || + (ref_cdf[1] != 32767)) { + fprintf(stderr, + "\n *** [bit] cdf error, frame_idx_r %d cdf {%d, %d} ref_cdf {%d", + frame_idx, p, 32767, ref_cdf[0]); + for (i = 1; i < ref_nsymbs; ++i) fprintf(stderr, ", %d", ref_cdf[i]); + fprintf(stderr, "} queue_r %d\n", queue_r); + assert(0); + } + if (bit != ref_bit) { + fprintf(stderr, + "\n *** [bit] symb error, frame_idx_r %d symb %d ref_symb %d " + "queue_r %d\n", + frame_idx, bit, ref_bit, queue_r); + assert(0); + } + } +#endif + + return bit; +} + +static INLINE int aom_daala_reader_has_error(daala_reader *r) { + return r->ec.error; +} + +static INLINE int daala_read_symbol(daala_reader *r, const aom_cdf_prob *cdf, + int nsymbs) { + int symb; + assert(cdf != NULL); + symb = od_ec_decode_cdf_q15(&r->ec, cdf, nsymbs); + +#if CONFIG_BITSTREAM_DEBUG + { + int i; + int cdf_error = 0; + int ref_symb, ref_nsymbs; + aom_cdf_prob ref_cdf[16]; + const int queue_r = bitstream_queue_get_read(); + const int frame_idx = bitstream_queue_get_frame_read(); + bitstream_queue_pop(&ref_symb, ref_cdf, &ref_nsymbs); + if (nsymbs != ref_nsymbs) { + fprintf(stderr, + "\n *** nsymbs error, frame_idx_r %d nsymbs %d ref_nsymbs %d " + "queue_r %d\n", + frame_idx, nsymbs, ref_nsymbs, queue_r); + cdf_error = 0; + assert(0); + } else { + for (i = 0; i < nsymbs; ++i) + if (cdf[i] != ref_cdf[i]) cdf_error = 1; + } + if (cdf_error) { + fprintf(stderr, "\n *** cdf error, frame_idx_r %d cdf {%d", frame_idx, + cdf[0]); + for (i = 1; i < nsymbs; ++i) fprintf(stderr, ", %d", cdf[i]); + fprintf(stderr, "} ref_cdf {%d", ref_cdf[0]); + for (i = 1; i < ref_nsymbs; ++i) fprintf(stderr, ", %d", ref_cdf[i]); + fprintf(stderr, "} queue_r %d\n", queue_r); + assert(0); + } + if (symb != ref_symb) { + fprintf( + stderr, + "\n *** symb error, frame_idx_r %d symb %d ref_symb %d queue_r %d\n", + frame_idx, symb, ref_symb, queue_r); + assert(0); + } + } +#endif + + return symb; +} + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif // AOM_AOM_DSP_DAALABOOLREADER_H_ diff --git a/third_party/aom/aom_dsp/daalaboolwriter.c b/third_party/aom/aom_dsp/daalaboolwriter.c new file mode 100644 index 0000000000..b24ffbf3fe --- /dev/null +++ b/third_party/aom/aom_dsp/daalaboolwriter.c @@ -0,0 +1,31 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include +#include "aom_dsp/daalaboolwriter.h" + +void aom_daala_start_encode(daala_writer *br, uint8_t *source) { + br->buffer = source; + br->pos = 0; + od_ec_enc_init(&br->ec, 62025); +} + +int aom_daala_stop_encode(daala_writer *br) { + int nb_bits; + uint32_t daala_bytes; + unsigned char *daala_data; + daala_data = od_ec_enc_done(&br->ec, &daala_bytes); + nb_bits = od_ec_enc_tell(&br->ec); + memcpy(br->buffer, daala_data, daala_bytes); + br->pos = daala_bytes; + od_ec_enc_clear(&br->ec); + return nb_bits; +} diff --git a/third_party/aom/aom_dsp/daalaboolwriter.h b/third_party/aom/aom_dsp/daalaboolwriter.h new file mode 100644 index 0000000000..3848877ced --- /dev/null +++ b/third_party/aom/aom_dsp/daalaboolwriter.h @@ -0,0 +1,78 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_DAALABOOLWRITER_H_ +#define AOM_AOM_DSP_DAALABOOLWRITER_H_ + +#include + +#include "aom_dsp/entenc.h" +#include "aom_dsp/prob.h" +#if CONFIG_BITSTREAM_DEBUG +#include "aom_util/debug_util.h" +#endif // CONFIG_BITSTREAM_DEBUG + +#ifdef __cplusplus +extern "C" { +#endif + +struct daala_writer { + unsigned int pos; + uint8_t *buffer; + od_ec_enc ec; + uint8_t allow_update_cdf; +}; + +typedef struct daala_writer daala_writer; + +void aom_daala_start_encode(daala_writer *w, uint8_t *buffer); +int aom_daala_stop_encode(daala_writer *w); + +static INLINE void aom_daala_write(daala_writer *w, int bit, int prob) { + int p = (0x7FFFFF - (prob << 15) + prob) >> 8; +#if CONFIG_BITSTREAM_DEBUG + aom_cdf_prob cdf[2] = { (aom_cdf_prob)p, 32767 }; + /*int queue_r = 0; + int frame_idx_r = 0; + int queue_w = bitstream_queue_get_write(); + int frame_idx_w = bitstream_queue_get_frame_write(); + if (frame_idx_w == frame_idx_r && queue_w == queue_r) { + fprintf(stderr, "\n *** bitstream queue at frame_idx_w %d queue_w %d\n", + frame_idx_w, queue_w); + }*/ + bitstream_queue_push(bit, cdf, 2); +#endif + + od_ec_encode_bool_q15(&w->ec, bit, p); +} + +static INLINE void daala_write_symbol(daala_writer *w, int symb, + const aom_cdf_prob *cdf, int nsymbs) { +#if CONFIG_BITSTREAM_DEBUG + /*int queue_r = 0; + int frame_idx_r = 0; + int queue_w = bitstream_queue_get_write(); + int frame_idx_w = bitstream_queue_get_frame_write(); + if (frame_idx_w == frame_idx_r && queue_w == queue_r) { + fprintf(stderr, "\n *** bitstream queue at frame_idx_w %d queue_w %d\n", + frame_idx_w, queue_w); + }*/ + bitstream_queue_push(symb, cdf, nsymbs); +#endif + + od_ec_encode_cdf_q15(&w->ec, symb, cdf, nsymbs); +} + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif // AOM_AOM_DSP_DAALABOOLWRITER_H_ diff --git a/third_party/aom/aom_dsp/entcode.c b/third_party/aom/aom_dsp/entcode.c new file mode 100644 index 0000000000..aad96c6fc6 --- /dev/null +++ b/third_party/aom/aom_dsp/entcode.c @@ -0,0 +1,49 @@ +/* + * Copyright (c) 2001-2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include "aom_dsp/entcode.h" + +/*Given the current total integer number of bits used and the current value of + rng, computes the fraction number of bits used to OD_BITRES precision. + This is used by od_ec_enc_tell_frac() and od_ec_dec_tell_frac(). + nbits_total: The number of whole bits currently used, i.e., the value + returned by od_ec_enc_tell() or od_ec_dec_tell(). + rng: The current value of rng from either the encoder or decoder state. + Return: The number of bits scaled by 2**OD_BITRES. + This will always be slightly larger than the exact value (e.g., all + rounding error is in the positive direction).*/ +uint32_t od_ec_tell_frac(uint32_t nbits_total, uint32_t rng) { + uint32_t nbits; + int l; + int i; + /*To handle the non-integral number of bits still left in the encoder/decoder + state, we compute the worst-case number of bits of val that must be + encoded to ensure that the value is inside the range for any possible + subsequent bits. + The computation here is independent of val itself (the decoder does not + even track that value), even though the real number of bits used after + od_ec_enc_done() may be 1 smaller if rng is a power of two and the + corresponding trailing bits of val are all zeros. + If we did try to track that special case, then coding a value with a + probability of 1/(1 << n) might sometimes appear to use more than n bits. + This may help explain the surprising result that a newly initialized + encoder or decoder claims to have used 1 bit.*/ + nbits = nbits_total << OD_BITRES; + l = 0; + for (i = OD_BITRES; i-- > 0;) { + int b; + rng = rng * rng >> 15; + b = (int)(rng >> 16); + l = l << 1 | b; + rng >>= b; + } + return nbits - l; +} diff --git a/third_party/aom/aom_dsp/entcode.h b/third_party/aom/aom_dsp/entcode.h new file mode 100644 index 0000000000..7ba2b1c392 --- /dev/null +++ b/third_party/aom/aom_dsp/entcode.h @@ -0,0 +1,40 @@ +/* + * Copyright (c) 2001-2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_ENTCODE_H_ +#define AOM_AOM_DSP_ENTCODE_H_ + +#include +#include +#include "av1/common/odintrin.h" +#include "aom_dsp/prob.h" + +#define EC_PROB_SHIFT 6 +#define EC_MIN_PROB 4 // must be <= (1< 1/8th bits.*/ +#define OD_BITRES (3) + +#define OD_ICDF AOM_ICDF + +/*See entcode.c for further documentation.*/ + +OD_WARN_UNUSED_RESULT uint32_t od_ec_tell_frac(uint32_t nbits_total, + uint32_t rng); + +#endif // AOM_AOM_DSP_ENTCODE_H_ diff --git a/third_party/aom/aom_dsp/entdec.c b/third_party/aom/aom_dsp/entdec.c new file mode 100644 index 0000000000..d1764c47b1 --- /dev/null +++ b/third_party/aom/aom_dsp/entdec.c @@ -0,0 +1,229 @@ +/* + * Copyright (c) 2001-2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include +#include "aom_dsp/entdec.h" +#include "aom_dsp/prob.h" + +/*A range decoder. + This is an entropy decoder based upon \cite{Mar79}, which is itself a + rediscovery of the FIFO arithmetic code introduced by \cite{Pas76}. + It is very similar to arithmetic encoding, except that encoding is done with + digits in any base, instead of with bits, and so it is faster when using + larger bases (i.e.: a byte). + The author claims an average waste of $\frac{1}{2}\log_b(2b)$ bits, where $b$ + is the base, longer than the theoretical optimum, but to my knowledge there + is no published justification for this claim. + This only seems true when using near-infinite precision arithmetic so that + the process is carried out with no rounding errors. + + An excellent description of implementation details is available at + http://www.arturocampos.com/ac_range.html + A recent work \cite{MNW98} which proposes several changes to arithmetic + encoding for efficiency actually re-discovers many of the principles + behind range encoding, and presents a good theoretical analysis of them. + + End of stream is handled by writing out the smallest number of bits that + ensures that the stream will be correctly decoded regardless of the value of + any subsequent bits. + od_ec_dec_tell() can be used to determine how many bits were needed to decode + all the symbols thus far; other data can be packed in the remaining bits of + the input buffer. + @PHDTHESIS{Pas76, + author="Richard Clark Pasco", + title="Source coding algorithms for fast data compression", + school="Dept. of Electrical Engineering, Stanford University", + address="Stanford, CA", + month=May, + year=1976, + URL="http://www.richpasco.org/scaffdc.pdf" + } + @INPROCEEDINGS{Mar79, + author="Martin, G.N.N.", + title="Range encoding: an algorithm for removing redundancy from a digitised + message", + booktitle="Video & Data Recording Conference", + year=1979, + address="Southampton", + month=Jul, + URL="http://www.compressconsult.com/rangecoder/rngcod.pdf.gz" + } + @ARTICLE{MNW98, + author="Alistair Moffat and Radford Neal and Ian H. Witten", + title="Arithmetic Coding Revisited", + journal="{ACM} Transactions on Information Systems", + year=1998, + volume=16, + number=3, + pages="256--294", + month=Jul, + URL="http://researchcommons.waikato.ac.nz/bitstream/handle/10289/78/content.pdf" + }*/ + +/*This is meant to be a large, positive constant that can still be efficiently + loaded as an immediate (on platforms like ARM, for example). + Even relatively modest values like 100 would work fine.*/ +#define OD_EC_LOTS_OF_BITS (0x4000) + +/*The return value of od_ec_dec_tell does not change across an od_ec_dec_refill + call.*/ +static void od_ec_dec_refill(od_ec_dec *dec) { + int s; + od_ec_window dif; + int16_t cnt; + const unsigned char *bptr; + const unsigned char *end; + dif = dec->dif; + cnt = dec->cnt; + bptr = dec->bptr; + end = dec->end; + s = OD_EC_WINDOW_SIZE - 9 - (cnt + 15); + for (; s >= 0 && bptr < end; s -= 8, bptr++) { + assert(s <= OD_EC_WINDOW_SIZE - 8); + dif ^= (od_ec_window)bptr[0] << s; + cnt += 8; + } + if (bptr >= end) { + dec->tell_offs += OD_EC_LOTS_OF_BITS - cnt; + cnt = OD_EC_LOTS_OF_BITS; + } + dec->dif = dif; + dec->cnt = cnt; + dec->bptr = bptr; +} + +/*Takes updated dif and range values, renormalizes them so that + 32768 <= rng < 65536 (reading more bytes from the stream into dif if + necessary), and stores them back in the decoder context. + dif: The new value of dif. + rng: The new value of the range. + ret: The value to return. + Return: ret. + This allows the compiler to jump to this function via a tail-call.*/ +static int od_ec_dec_normalize(od_ec_dec *dec, od_ec_window dif, unsigned rng, + int ret) { + int d; + assert(rng <= 65535U); + // The number of leading zeros in the 16-bit binary representation of rng. + d = 16 - OD_ILOG_NZ(rng); + dec->cnt -= d; + /*This is equivalent to shifting in 1's instead of 0's.*/ + dec->dif = ((dif + 1) << d) - 1; + dec->rng = rng << d; + if (dec->cnt < 0) od_ec_dec_refill(dec); + return ret; +} + +/*Initializes the decoder. + buf: The input buffer to use. + Return: 0 on success, or a negative value on error.*/ +void od_ec_dec_init(od_ec_dec *dec, const unsigned char *buf, + uint32_t storage) { + dec->buf = buf; + dec->tell_offs = 10 - (OD_EC_WINDOW_SIZE - 8); + dec->end = buf + storage; + dec->bptr = buf; + dec->dif = ((od_ec_window)1 << (OD_EC_WINDOW_SIZE - 1)) - 1; + dec->rng = 0x8000; + dec->cnt = -15; + dec->error = 0; + od_ec_dec_refill(dec); +} + +/*Decode a single binary value. + f: The probability that the bit is one, scaled by 32768. + Return: The value decoded (0 or 1).*/ +int od_ec_decode_bool_q15(od_ec_dec *dec, unsigned f) { + od_ec_window dif; + od_ec_window vw; + unsigned r; + unsigned r_new; + unsigned v; + int ret; + assert(0 < f); + assert(f < 32768U); + dif = dec->dif; + r = dec->rng; + assert(dif >> (OD_EC_WINDOW_SIZE - 16) < r); + assert(32768U <= r); + v = ((r >> 8) * (uint32_t)(f >> EC_PROB_SHIFT) >> (7 - EC_PROB_SHIFT)); + v += EC_MIN_PROB; + vw = (od_ec_window)v << (OD_EC_WINDOW_SIZE - 16); + ret = 1; + r_new = v; + if (dif >= vw) { + r_new = r - v; + dif -= vw; + ret = 0; + } + return od_ec_dec_normalize(dec, dif, r_new, ret); +} + +/*Decodes a symbol given an inverse cumulative distribution function (CDF) + table in Q15. + icdf: CDF_PROB_TOP minus the CDF, such that symbol s falls in the range + [s > 0 ? (CDF_PROB_TOP - icdf[s - 1]) : 0, CDF_PROB_TOP - icdf[s]). + The values must be monotonically non-increasing, and icdf[nsyms - 1] + must be 0. + nsyms: The number of symbols in the alphabet. + This should be at most 16. + Return: The decoded symbol s.*/ +int od_ec_decode_cdf_q15(od_ec_dec *dec, const uint16_t *icdf, int nsyms) { + od_ec_window dif; + unsigned r; + unsigned c; + unsigned u; + unsigned v; + int ret; + (void)nsyms; + dif = dec->dif; + r = dec->rng; + const int N = nsyms - 1; + + assert(dif >> (OD_EC_WINDOW_SIZE - 16) < r); + assert(icdf[nsyms - 1] == OD_ICDF(CDF_PROB_TOP)); + assert(32768U <= r); + assert(7 - EC_PROB_SHIFT - CDF_SHIFT >= 0); + c = (unsigned)(dif >> (OD_EC_WINDOW_SIZE - 16)); + v = r; + ret = -1; + do { + u = v; + v = ((r >> 8) * (uint32_t)(icdf[++ret] >> EC_PROB_SHIFT) >> + (7 - EC_PROB_SHIFT - CDF_SHIFT)); + v += EC_MIN_PROB * (N - ret); + } while (c < v); + assert(v < u); + assert(u <= r); + r = u - v; + dif -= (od_ec_window)v << (OD_EC_WINDOW_SIZE - 16); + return od_ec_dec_normalize(dec, dif, r, ret); +} + +/*Returns the number of bits "used" by the decoded symbols so far. + This same number can be computed in either the encoder or the decoder, and is + suitable for making coding decisions. + Return: The number of bits. + This will always be slightly larger than the exact value (e.g., all + rounding error is in the positive direction).*/ +int od_ec_dec_tell(const od_ec_dec *dec) { + return (int)((dec->bptr - dec->buf) * 8 - dec->cnt + dec->tell_offs); +} + +/*Returns the number of bits "used" by the decoded symbols so far. + This same number can be computed in either the encoder or the decoder, and is + suitable for making coding decisions. + Return: The number of bits scaled by 2**OD_BITRES. + This will always be slightly larger than the exact value (e.g., all + rounding error is in the positive direction).*/ +uint32_t od_ec_dec_tell_frac(const od_ec_dec *dec) { + return od_ec_tell_frac(od_ec_dec_tell(dec), dec->rng); +} diff --git a/third_party/aom/aom_dsp/entdec.h b/third_party/aom/aom_dsp/entdec.h new file mode 100644 index 0000000000..283bf18315 --- /dev/null +++ b/third_party/aom/aom_dsp/entdec.h @@ -0,0 +1,83 @@ +/* + * Copyright (c) 2001-2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_ENTDEC_H_ +#define AOM_AOM_DSP_ENTDEC_H_ +#include +#include "aom_dsp/entcode.h" + +#ifdef __cplusplus +extern "C" { +#endif + +typedef struct od_ec_dec od_ec_dec; + +#if defined(OD_ACCOUNTING) && OD_ACCOUNTING +#define OD_ACC_STR , char *acc_str +#define od_ec_dec_bits(dec, ftb, str) od_ec_dec_bits_(dec, ftb, str) +#else +#define OD_ACC_STR +#define od_ec_dec_bits(dec, ftb, str) od_ec_dec_bits_(dec, ftb) +#endif + +/*The entropy decoder context.*/ +struct od_ec_dec { + /*The start of the current input buffer.*/ + const unsigned char *buf; + /*An offset used to keep track of tell after reaching the end of the stream. + This is constant throughout most of the decoding process, but becomes + important once we hit the end of the buffer and stop incrementing pointers + (and instead pretend cnt has lots of bits).*/ + int32_t tell_offs; + /*The end of the current input buffer.*/ + const unsigned char *end; + /*The read pointer for the entropy-coded bits.*/ + const unsigned char *bptr; + /*The difference between the high end of the current range, (low + rng), and + the coded value, minus 1. + This stores up to OD_EC_WINDOW_SIZE bits of that difference, but the + decoder only uses the top 16 bits of the window to decode the next symbol. + As we shift up during renormalization, if we don't have enough bits left in + the window to fill the top 16, we'll read in more bits of the coded + value.*/ + od_ec_window dif; + /*The number of values in the current range.*/ + uint16_t rng; + /*The number of bits of data in the current value.*/ + int16_t cnt; + /*Nonzero if an error occurred.*/ + int error; +}; + +/*See entdec.c for further documentation.*/ + +void od_ec_dec_init(od_ec_dec *dec, const unsigned char *buf, uint32_t storage) + OD_ARG_NONNULL(1) OD_ARG_NONNULL(2); + +OD_WARN_UNUSED_RESULT int od_ec_decode_bool_q15(od_ec_dec *dec, unsigned f) + OD_ARG_NONNULL(1); +OD_WARN_UNUSED_RESULT int od_ec_decode_cdf_q15(od_ec_dec *dec, + const uint16_t *cdf, int nsyms) + OD_ARG_NONNULL(1) OD_ARG_NONNULL(2); + +OD_WARN_UNUSED_RESULT uint32_t od_ec_dec_bits_(od_ec_dec *dec, unsigned ftb) + OD_ARG_NONNULL(1); + +OD_WARN_UNUSED_RESULT int od_ec_dec_tell(const od_ec_dec *dec) + OD_ARG_NONNULL(1); +OD_WARN_UNUSED_RESULT uint32_t od_ec_dec_tell_frac(const od_ec_dec *dec) + OD_ARG_NONNULL(1); + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif // AOM_AOM_DSP_ENTDEC_H_ diff --git a/third_party/aom/aom_dsp/entenc.c b/third_party/aom/aom_dsp/entenc.c new file mode 100644 index 0000000000..a61da263cb --- /dev/null +++ b/third_party/aom/aom_dsp/entenc.c @@ -0,0 +1,423 @@ +/* + * Copyright (c) 2001-2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include +#include +#include +#include +#include "aom_dsp/entenc.h" +#include "aom_dsp/prob.h" + +#if OD_MEASURE_EC_OVERHEAD +#if !defined(M_LOG2E) +#define M_LOG2E (1.4426950408889634073599246810019) +#endif +#define OD_LOG2(x) (M_LOG2E * log(x)) +#endif // OD_MEASURE_EC_OVERHEAD + +/*A range encoder. + See entdec.c and the references for implementation details \cite{Mar79,MNW98}. + + @INPROCEEDINGS{Mar79, + author="Martin, G.N.N.", + title="Range encoding: an algorithm for removing redundancy from a digitised + message", + booktitle="Video \& Data Recording Conference", + year=1979, + address="Southampton", + month=Jul, + URL="http://www.compressconsult.com/rangecoder/rngcod.pdf.gz" + } + @ARTICLE{MNW98, + author="Alistair Moffat and Radford Neal and Ian H. Witten", + title="Arithmetic Coding Revisited", + journal="{ACM} Transactions on Information Systems", + year=1998, + volume=16, + number=3, + pages="256--294", + month=Jul, + URL="http://researchcommons.waikato.ac.nz/bitstream/handle/10289/78/content.pdf" + }*/ + +/*Takes updated low and range values, renormalizes them so that + 32768 <= rng < 65536 (flushing bytes from low to the pre-carry buffer if + necessary), and stores them back in the encoder context. + low: The new value of low. + rng: The new value of the range.*/ +static void od_ec_enc_normalize(od_ec_enc *enc, od_ec_window low, + unsigned rng) { + int d; + int c; + int s; + c = enc->cnt; + assert(rng <= 65535U); + // The number of leading zeros in the 16-bit binary representation of rng. + d = 16 - OD_ILOG_NZ(rng); + s = c + d; + /*TODO: Right now we flush every time we have at least one byte available. + Instead we should use an od_ec_window and flush right before we're about to + shift bits off the end of the window. + For a 32-bit window this is about the same amount of work, but for a 64-bit + window it should be a fair win.*/ + if (s >= 0) { + uint16_t *buf; + uint32_t storage; + uint32_t offs; + unsigned m; + buf = enc->precarry_buf; + storage = enc->precarry_storage; + offs = enc->offs; + if (offs + 2 > storage) { + storage = 2 * storage + 2; + buf = (uint16_t *)realloc(buf, sizeof(*buf) * storage); + if (buf == NULL) { + enc->error = -1; + enc->offs = 0; + return; + } + enc->precarry_buf = buf; + enc->precarry_storage = storage; + } + c += 16; + m = (1 << c) - 1; + if (s >= 8) { + assert(offs < storage); + buf[offs++] = (uint16_t)(low >> c); + low &= m; + c -= 8; + m >>= 8; + } + assert(offs < storage); + buf[offs++] = (uint16_t)(low >> c); + s = c + d - 24; + low &= m; + enc->offs = offs; + } + enc->low = low << d; + enc->rng = rng << d; + enc->cnt = s; +} + +/*Initializes the encoder. + size: The initial size of the buffer, in bytes.*/ +void od_ec_enc_init(od_ec_enc *enc, uint32_t size) { + od_ec_enc_reset(enc); + enc->buf = (unsigned char *)malloc(sizeof(*enc->buf) * size); + enc->storage = size; + if (size > 0 && enc->buf == NULL) { + enc->storage = 0; + enc->error = -1; + } + enc->precarry_buf = (uint16_t *)malloc(sizeof(*enc->precarry_buf) * size); + enc->precarry_storage = size; + if (size > 0 && enc->precarry_buf == NULL) { + enc->precarry_storage = 0; + enc->error = -1; + } +} + +/*Reinitializes the encoder.*/ +void od_ec_enc_reset(od_ec_enc *enc) { + enc->offs = 0; + enc->low = 0; + enc->rng = 0x8000; + /*This is initialized to -9 so that it crosses zero after we've accumulated + one byte + one carry bit.*/ + enc->cnt = -9; + enc->error = 0; +#if OD_MEASURE_EC_OVERHEAD + enc->entropy = 0; + enc->nb_symbols = 0; +#endif +} + +/*Frees the buffers used by the encoder.*/ +void od_ec_enc_clear(od_ec_enc *enc) { + free(enc->precarry_buf); + free(enc->buf); +} + +/*Encodes a symbol given its frequency in Q15. + fl: CDF_PROB_TOP minus the cumulative frequency of all symbols that come + before the + one to be encoded. + fh: CDF_PROB_TOP minus the cumulative frequency of all symbols up to and + including + the one to be encoded.*/ +static void od_ec_encode_q15(od_ec_enc *enc, unsigned fl, unsigned fh, int s, + int nsyms) { + od_ec_window l; + unsigned r; + unsigned u; + unsigned v; + l = enc->low; + r = enc->rng; + assert(32768U <= r); + assert(fh <= fl); + assert(fl <= 32768U); + assert(7 - EC_PROB_SHIFT - CDF_SHIFT >= 0); + const int N = nsyms - 1; + if (fl < CDF_PROB_TOP) { + u = ((r >> 8) * (uint32_t)(fl >> EC_PROB_SHIFT) >> + (7 - EC_PROB_SHIFT - CDF_SHIFT)) + + EC_MIN_PROB * (N - (s - 1)); + v = ((r >> 8) * (uint32_t)(fh >> EC_PROB_SHIFT) >> + (7 - EC_PROB_SHIFT - CDF_SHIFT)) + + EC_MIN_PROB * (N - (s + 0)); + l += r - u; + r = u - v; + } else { + r -= ((r >> 8) * (uint32_t)(fh >> EC_PROB_SHIFT) >> + (7 - EC_PROB_SHIFT - CDF_SHIFT)) + + EC_MIN_PROB * (N - (s + 0)); + } + od_ec_enc_normalize(enc, l, r); +#if OD_MEASURE_EC_OVERHEAD + enc->entropy -= OD_LOG2((double)(OD_ICDF(fh) - OD_ICDF(fl)) / CDF_PROB_TOP.); + enc->nb_symbols++; +#endif +} + +/*Encode a single binary value. + val: The value to encode (0 or 1). + f: The probability that the val is one, scaled by 32768.*/ +void od_ec_encode_bool_q15(od_ec_enc *enc, int val, unsigned f) { + od_ec_window l; + unsigned r; + unsigned v; + assert(0 < f); + assert(f < 32768U); + l = enc->low; + r = enc->rng; + assert(32768U <= r); + v = ((r >> 8) * (uint32_t)(f >> EC_PROB_SHIFT) >> (7 - EC_PROB_SHIFT)); + v += EC_MIN_PROB; + if (val) l += r - v; + r = val ? v : r - v; + od_ec_enc_normalize(enc, l, r); +#if OD_MEASURE_EC_OVERHEAD + enc->entropy -= OD_LOG2((double)(val ? f : (32768 - f)) / 32768.); + enc->nb_symbols++; +#endif +} + +/*Encodes a symbol given a cumulative distribution function (CDF) table in Q15. + s: The index of the symbol to encode. + icdf: 32768 minus the CDF, such that symbol s falls in the range + [s > 0 ? (32768 - icdf[s - 1]) : 0, 32768 - icdf[s]). + The values must be monotonically decreasing, and icdf[nsyms - 1] must + be 0. + nsyms: The number of symbols in the alphabet. + This should be at most 16.*/ +void od_ec_encode_cdf_q15(od_ec_enc *enc, int s, const uint16_t *icdf, + int nsyms) { + (void)nsyms; + assert(s >= 0); + assert(s < nsyms); + assert(icdf[nsyms - 1] == OD_ICDF(CDF_PROB_TOP)); + od_ec_encode_q15(enc, s > 0 ? icdf[s - 1] : OD_ICDF(0), icdf[s], s, nsyms); +} + +/*Overwrites a few bits at the very start of an existing stream, after they + have already been encoded. + This makes it possible to have a few flags up front, where it is easy for + decoders to access them without parsing the whole stream, even if their + values are not determined until late in the encoding process, without having + to buffer all the intermediate symbols in the encoder. + In order for this to work, at least nbits bits must have already been encoded + using probabilities that are an exact power of two. + The encoder can verify the number of encoded bits is sufficient, but cannot + check this latter condition. + val: The bits to encode (in the least nbits significant bits). + They will be decoded in order from most-significant to least. + nbits: The number of bits to overwrite. + This must be no more than 8.*/ +void od_ec_enc_patch_initial_bits(od_ec_enc *enc, unsigned val, int nbits) { + int shift; + unsigned mask; + assert(nbits >= 0); + assert(nbits <= 8); + assert(val < 1U << nbits); + shift = 8 - nbits; + mask = ((1U << nbits) - 1) << shift; + if (enc->offs > 0) { + /*The first byte has been finalized.*/ + enc->precarry_buf[0] = + (uint16_t)((enc->precarry_buf[0] & ~mask) | val << shift); + } else if (9 + enc->cnt + (enc->rng == 0x8000) > nbits) { + /*The first byte has yet to be output.*/ + enc->low = (enc->low & ~((od_ec_window)mask << (16 + enc->cnt))) | + (od_ec_window)val << (16 + enc->cnt + shift); + } else { + /*The encoder hasn't even encoded _nbits of data yet.*/ + enc->error = -1; + } +} + +#if OD_MEASURE_EC_OVERHEAD +#include +#endif + +/*Indicates that there are no more symbols to encode. + All remaining output bytes are flushed to the output buffer. + od_ec_enc_reset() should be called before using the encoder again. + bytes: Returns the size of the encoded data in the returned buffer. + Return: A pointer to the start of the final buffer, or NULL if there was an + encoding error.*/ +unsigned char *od_ec_enc_done(od_ec_enc *enc, uint32_t *nbytes) { + unsigned char *out; + uint32_t storage; + uint16_t *buf; + uint32_t offs; + od_ec_window m; + od_ec_window e; + od_ec_window l; + int c; + int s; + if (enc->error) return NULL; +#if OD_MEASURE_EC_OVERHEAD + { + uint32_t tell; + /* Don't count the 1 bit we lose to raw bits as overhead. */ + tell = od_ec_enc_tell(enc) - 1; + fprintf(stderr, "overhead: %f%%\n", + 100 * (tell - enc->entropy) / enc->entropy); + fprintf(stderr, "efficiency: %f bits/symbol\n", + (double)tell / enc->nb_symbols); + } +#endif + /*We output the minimum number of bits that ensures that the symbols encoded + thus far will be decoded correctly regardless of the bits that follow.*/ + l = enc->low; + c = enc->cnt; + s = 10; + m = 0x3FFF; + e = ((l + m) & ~m) | (m + 1); + s += c; + offs = enc->offs; + buf = enc->precarry_buf; + if (s > 0) { + unsigned n; + storage = enc->precarry_storage; + if (offs + ((s + 7) >> 3) > storage) { + storage = storage * 2 + ((s + 7) >> 3); + buf = (uint16_t *)realloc(buf, sizeof(*buf) * storage); + if (buf == NULL) { + enc->error = -1; + return NULL; + } + enc->precarry_buf = buf; + enc->precarry_storage = storage; + } + n = (1 << (c + 16)) - 1; + do { + assert(offs < storage); + buf[offs++] = (uint16_t)(e >> (c + 16)); + e &= n; + s -= 8; + c -= 8; + n >>= 8; + } while (s > 0); + } + /*Make sure there's enough room for the entropy-coded bits.*/ + out = enc->buf; + storage = enc->storage; + c = OD_MAXI((s + 7) >> 3, 0); + if (offs + c > storage) { + storage = offs + c; + out = (unsigned char *)realloc(out, sizeof(*out) * storage); + if (out == NULL) { + enc->error = -1; + return NULL; + } + enc->buf = out; + enc->storage = storage; + } + *nbytes = offs; + /*Perform carry propagation.*/ + assert(offs <= storage); + out = out + storage - offs; + c = 0; + while (offs > 0) { + offs--; + c = buf[offs] + c; + out[offs] = (unsigned char)c; + c >>= 8; + } + /*Note: Unless there's an allocation error, if you keep encoding into the + current buffer and call this function again later, everything will work + just fine (you won't get a new packet out, but you will get a single + buffer with the new data appended to the old). + However, this function is O(N) where N is the amount of data coded so far, + so calling it more than once for a given packet is a bad idea.*/ + return out; +} + +/*Returns the number of bits "used" by the encoded symbols so far. + This same number can be computed in either the encoder or the decoder, and is + suitable for making coding decisions. + Warning: The value returned by this function can decrease compared to an + earlier call, even after encoding more data, if there is an encoding error + (i.e., a failure to allocate enough space for the output buffer). + Return: The number of bits. + This will always be slightly larger than the exact value (e.g., all + rounding error is in the positive direction).*/ +int od_ec_enc_tell(const od_ec_enc *enc) { + /*The 10 here counteracts the offset of -9 baked into cnt, and adds 1 extra + bit, which we reserve for terminating the stream.*/ + return (enc->cnt + 10) + enc->offs * 8; +} + +/*Returns the number of bits "used" by the encoded symbols so far. + This same number can be computed in either the encoder or the decoder, and is + suitable for making coding decisions. + Warning: The value returned by this function can decrease compared to an + earlier call, even after encoding more data, if there is an encoding error + (i.e., a failure to allocate enough space for the output buffer). + Return: The number of bits scaled by 2**OD_BITRES. + This will always be slightly larger than the exact value (e.g., all + rounding error is in the positive direction).*/ +uint32_t od_ec_enc_tell_frac(const od_ec_enc *enc) { + return od_ec_tell_frac(od_ec_enc_tell(enc), enc->rng); +} + +/*Saves a entropy coder checkpoint to dst. + This allows an encoder to reverse a series of entropy coder + decisions if it decides that the information would have been + better coded some other way.*/ +void od_ec_enc_checkpoint(od_ec_enc *dst, const od_ec_enc *src) { + OD_COPY(dst, src, 1); +} + +/*Restores an entropy coder checkpoint saved by od_ec_enc_checkpoint. + This can only be used to restore from checkpoints earlier in the target + state's history: you can not switch backwards and forwards or otherwise + switch to a state which isn't a casual ancestor of the current state. + Restore is also incompatible with patching the initial bits, as the + changes will remain in the restored version.*/ +void od_ec_enc_rollback(od_ec_enc *dst, const od_ec_enc *src) { + unsigned char *buf; + uint32_t storage; + uint16_t *precarry_buf; + uint32_t precarry_storage; + assert(dst->storage >= src->storage); + assert(dst->precarry_storage >= src->precarry_storage); + buf = dst->buf; + storage = dst->storage; + precarry_buf = dst->precarry_buf; + precarry_storage = dst->precarry_storage; + OD_COPY(dst, src, 1); + dst->buf = buf; + dst->storage = storage; + dst->precarry_buf = precarry_buf; + dst->precarry_storage = precarry_storage; +} diff --git a/third_party/aom/aom_dsp/entenc.h b/third_party/aom/aom_dsp/entenc.h new file mode 100644 index 0000000000..3551d42508 --- /dev/null +++ b/third_party/aom/aom_dsp/entenc.h @@ -0,0 +1,85 @@ +/* + * Copyright (c) 2001-2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_ENTENC_H_ +#define AOM_AOM_DSP_ENTENC_H_ +#include +#include "aom_dsp/entcode.h" + +#ifdef __cplusplus +extern "C" { +#endif + +typedef struct od_ec_enc od_ec_enc; + +#define OD_MEASURE_EC_OVERHEAD (0) + +/*The entropy encoder context.*/ +struct od_ec_enc { + /*Buffered output. + This contains only the raw bits until the final call to od_ec_enc_done(), + where all the arithmetic-coded data gets prepended to it.*/ + unsigned char *buf; + /*The size of the buffer.*/ + uint32_t storage; + /*A buffer for output bytes with their associated carry flags.*/ + uint16_t *precarry_buf; + /*The size of the pre-carry buffer.*/ + uint32_t precarry_storage; + /*The offset at which the next entropy-coded byte will be written.*/ + uint32_t offs; + /*The low end of the current range.*/ + od_ec_window low; + /*The number of values in the current range.*/ + uint16_t rng; + /*The number of bits of data in the current value.*/ + int16_t cnt; + /*Nonzero if an error occurred.*/ + int error; +#if OD_MEASURE_EC_OVERHEAD + double entropy; + int nb_symbols; +#endif +}; + +/*See entenc.c for further documentation.*/ + +void od_ec_enc_init(od_ec_enc *enc, uint32_t size) OD_ARG_NONNULL(1); +void od_ec_enc_reset(od_ec_enc *enc) OD_ARG_NONNULL(1); +void od_ec_enc_clear(od_ec_enc *enc) OD_ARG_NONNULL(1); + +void od_ec_encode_bool_q15(od_ec_enc *enc, int val, unsigned f_q15) + OD_ARG_NONNULL(1); +void od_ec_encode_cdf_q15(od_ec_enc *enc, int s, const uint16_t *cdf, int nsyms) + OD_ARG_NONNULL(1) OD_ARG_NONNULL(3); + +void od_ec_enc_bits(od_ec_enc *enc, uint32_t fl, unsigned ftb) + OD_ARG_NONNULL(1); + +void od_ec_enc_patch_initial_bits(od_ec_enc *enc, unsigned val, int nbits) + OD_ARG_NONNULL(1); +OD_WARN_UNUSED_RESULT unsigned char *od_ec_enc_done(od_ec_enc *enc, + uint32_t *nbytes) + OD_ARG_NONNULL(1) OD_ARG_NONNULL(2); + +OD_WARN_UNUSED_RESULT int od_ec_enc_tell(const od_ec_enc *enc) + OD_ARG_NONNULL(1); +OD_WARN_UNUSED_RESULT uint32_t od_ec_enc_tell_frac(const od_ec_enc *enc) + OD_ARG_NONNULL(1); + +void od_ec_enc_checkpoint(od_ec_enc *dst, const od_ec_enc *src); +void od_ec_enc_rollback(od_ec_enc *dst, const od_ec_enc *src); + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif // AOM_AOM_DSP_ENTENC_H_ diff --git a/third_party/aom/aom_dsp/fastssim.c b/third_party/aom/aom_dsp/fastssim.c new file mode 100644 index 0000000000..3804519b31 --- /dev/null +++ b/third_party/aom/aom_dsp/fastssim.c @@ -0,0 +1,487 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + * + * This code was originally written by: Nathan E. Egge, at the Daala + * project. + */ +#include +#include +#include +#include + +#include "config/aom_config.h" +#include "config/aom_dsp_rtcd.h" + +#include "aom_dsp/ssim.h" +#include "aom_ports/system_state.h" + +typedef struct fs_level fs_level; +typedef struct fs_ctx fs_ctx; + +#define SSIM_C1 (255 * 255 * 0.01 * 0.01) +#define SSIM_C2 (255 * 255 * 0.03 * 0.03) +#define SSIM_C1_10 (1023 * 1023 * 0.01 * 0.01) +#define SSIM_C1_12 (4095 * 4095 * 0.01 * 0.01) +#define SSIM_C2_10 (1023 * 1023 * 0.03 * 0.03) +#define SSIM_C2_12 (4095 * 4095 * 0.03 * 0.03) + +#define FS_MINI(_a, _b) ((_a) < (_b) ? (_a) : (_b)) +#define FS_MAXI(_a, _b) ((_a) > (_b) ? (_a) : (_b)) + +struct fs_level { + uint32_t *im1; + uint32_t *im2; + double *ssim; + int w; + int h; +}; + +struct fs_ctx { + fs_level *level; + int nlevels; + unsigned *col_buf; +}; + +static void fs_ctx_init(fs_ctx *_ctx, int _w, int _h, int _nlevels) { + unsigned char *data; + size_t data_size; + int lw; + int lh; + int l; + lw = (_w + 1) >> 1; + lh = (_h + 1) >> 1; + data_size = + _nlevels * sizeof(fs_level) + 2 * (lw + 8) * 8 * sizeof(*_ctx->col_buf); + for (l = 0; l < _nlevels; l++) { + size_t im_size; + size_t level_size; + im_size = lw * (size_t)lh; + level_size = 2 * im_size * sizeof(*_ctx->level[l].im1); + level_size += sizeof(*_ctx->level[l].ssim) - 1; + level_size /= sizeof(*_ctx->level[l].ssim); + level_size += im_size; + level_size *= sizeof(*_ctx->level[l].ssim); + data_size += level_size; + lw = (lw + 1) >> 1; + lh = (lh + 1) >> 1; + } + data = (unsigned char *)malloc(data_size); + _ctx->level = (fs_level *)data; + _ctx->nlevels = _nlevels; + data += _nlevels * sizeof(*_ctx->level); + lw = (_w + 1) >> 1; + lh = (_h + 1) >> 1; + for (l = 0; l < _nlevels; l++) { + size_t im_size; + size_t level_size; + _ctx->level[l].w = lw; + _ctx->level[l].h = lh; + im_size = lw * (size_t)lh; + level_size = 2 * im_size * sizeof(*_ctx->level[l].im1); + level_size += sizeof(*_ctx->level[l].ssim) - 1; + level_size /= sizeof(*_ctx->level[l].ssim); + level_size *= sizeof(*_ctx->level[l].ssim); + _ctx->level[l].im1 = (uint32_t *)data; + _ctx->level[l].im2 = _ctx->level[l].im1 + im_size; + data += level_size; + _ctx->level[l].ssim = (double *)data; + data += im_size * sizeof(*_ctx->level[l].ssim); + lw = (lw + 1) >> 1; + lh = (lh + 1) >> 1; + } + _ctx->col_buf = (unsigned *)data; +} + +static void fs_ctx_clear(fs_ctx *_ctx) { free(_ctx->level); } + +static void fs_downsample_level(fs_ctx *_ctx, int _l) { + const uint32_t *src1; + const uint32_t *src2; + uint32_t *dst1; + uint32_t *dst2; + int w2; + int h2; + int w; + int h; + int i; + int j; + w = _ctx->level[_l].w; + h = _ctx->level[_l].h; + dst1 = _ctx->level[_l].im1; + dst2 = _ctx->level[_l].im2; + w2 = _ctx->level[_l - 1].w; + h2 = _ctx->level[_l - 1].h; + src1 = _ctx->level[_l - 1].im1; + src2 = _ctx->level[_l - 1].im2; + for (j = 0; j < h; j++) { + int j0offs; + int j1offs; + j0offs = 2 * j * w2; + j1offs = FS_MINI(2 * j + 1, h2) * w2; + for (i = 0; i < w; i++) { + int i0; + int i1; + i0 = 2 * i; + i1 = FS_MINI(i0 + 1, w2); + dst1[j * w + i] = src1[j0offs + i0] + src1[j0offs + i1] + + src1[j1offs + i0] + src1[j1offs + i1]; + dst2[j * w + i] = src2[j0offs + i0] + src2[j0offs + i1] + + src2[j1offs + i0] + src2[j1offs + i1]; + } + } +} + +static void fs_downsample_level0(fs_ctx *_ctx, const uint8_t *_src1, + int _s1ystride, const uint8_t *_src2, + int _s2ystride, int _w, int _h, uint32_t shift, + int buf_is_hbd) { + uint32_t *dst1; + uint32_t *dst2; + int w; + int h; + int i; + int j; + w = _ctx->level[0].w; + h = _ctx->level[0].h; + dst1 = _ctx->level[0].im1; + dst2 = _ctx->level[0].im2; + for (j = 0; j < h; j++) { + int j0; + int j1; + j0 = 2 * j; + j1 = FS_MINI(j0 + 1, _h); + for (i = 0; i < w; i++) { + int i0; + int i1; + i0 = 2 * i; + i1 = FS_MINI(i0 + 1, _w); + if (!buf_is_hbd) { + dst1[j * w + i] = + _src1[j0 * _s1ystride + i0] + _src1[j0 * _s1ystride + i1] + + _src1[j1 * _s1ystride + i0] + _src1[j1 * _s1ystride + i1]; + dst2[j * w + i] = + _src2[j0 * _s2ystride + i0] + _src2[j0 * _s2ystride + i1] + + _src2[j1 * _s2ystride + i0] + _src2[j1 * _s2ystride + i1]; + } else { + uint16_t *src1s = CONVERT_TO_SHORTPTR(_src1); + uint16_t *src2s = CONVERT_TO_SHORTPTR(_src2); + dst1[j * w + i] = (src1s[j0 * _s1ystride + i0] >> shift) + + (src1s[j0 * _s1ystride + i1] >> shift) + + (src1s[j1 * _s1ystride + i0] >> shift) + + (src1s[j1 * _s1ystride + i1] >> shift); + dst2[j * w + i] = (src2s[j0 * _s2ystride + i0] >> shift) + + (src2s[j0 * _s2ystride + i1] >> shift) + + (src2s[j1 * _s2ystride + i0] >> shift) + + (src2s[j1 * _s2ystride + i1] >> shift); + } + } + } +} + +static void fs_apply_luminance(fs_ctx *_ctx, int _l, int bit_depth) { + unsigned *col_sums_x; + unsigned *col_sums_y; + uint32_t *im1; + uint32_t *im2; + double *ssim; + double c1; + int w; + int h; + int j0offs; + int j1offs; + int i; + int j; + double ssim_c1 = SSIM_C1; + + if (bit_depth == 10) ssim_c1 = SSIM_C1_10; + if (bit_depth == 12) ssim_c1 = SSIM_C1_12; + + w = _ctx->level[_l].w; + h = _ctx->level[_l].h; + col_sums_x = _ctx->col_buf; + col_sums_y = col_sums_x + w; + im1 = _ctx->level[_l].im1; + im2 = _ctx->level[_l].im2; + for (i = 0; i < w; i++) col_sums_x[i] = 5 * im1[i]; + for (i = 0; i < w; i++) col_sums_y[i] = 5 * im2[i]; + for (j = 1; j < 4; j++) { + j1offs = FS_MINI(j, h - 1) * w; + for (i = 0; i < w; i++) col_sums_x[i] += im1[j1offs + i]; + for (i = 0; i < w; i++) col_sums_y[i] += im2[j1offs + i]; + } + ssim = _ctx->level[_l].ssim; + c1 = (double)(ssim_c1 * 4096 * (1 << 4 * _l)); + for (j = 0; j < h; j++) { + unsigned mux; + unsigned muy; + int i0; + int i1; + mux = 5 * col_sums_x[0]; + muy = 5 * col_sums_y[0]; + for (i = 1; i < 4; i++) { + i1 = FS_MINI(i, w - 1); + mux += col_sums_x[i1]; + muy += col_sums_y[i1]; + } + for (i = 0; i < w; i++) { + ssim[j * w + i] *= (2 * mux * (double)muy + c1) / + (mux * (double)mux + muy * (double)muy + c1); + if (i + 1 < w) { + i0 = FS_MAXI(0, i - 4); + i1 = FS_MINI(i + 4, w - 1); + mux += col_sums_x[i1] - col_sums_x[i0]; + muy += col_sums_x[i1] - col_sums_x[i0]; + } + } + if (j + 1 < h) { + j0offs = FS_MAXI(0, j - 4) * w; + for (i = 0; i < w; i++) col_sums_x[i] -= im1[j0offs + i]; + for (i = 0; i < w; i++) col_sums_y[i] -= im2[j0offs + i]; + j1offs = FS_MINI(j + 4, h - 1) * w; + for (i = 0; i < w; i++) col_sums_x[i] += im1[j1offs + i]; + for (i = 0; i < w; i++) col_sums_y[i] += im2[j1offs + i]; + } + } +} + +#define FS_COL_SET(_col, _joffs, _ioffs) \ + do { \ + unsigned gx; \ + unsigned gy; \ + gx = gx_buf[((j + (_joffs)) & 7) * stride + i + (_ioffs)]; \ + gy = gy_buf[((j + (_joffs)) & 7) * stride + i + (_ioffs)]; \ + col_sums_gx2[(_col)] = gx * (double)gx; \ + col_sums_gy2[(_col)] = gy * (double)gy; \ + col_sums_gxgy[(_col)] = gx * (double)gy; \ + } while (0) + +#define FS_COL_ADD(_col, _joffs, _ioffs) \ + do { \ + unsigned gx; \ + unsigned gy; \ + gx = gx_buf[((j + (_joffs)) & 7) * stride + i + (_ioffs)]; \ + gy = gy_buf[((j + (_joffs)) & 7) * stride + i + (_ioffs)]; \ + col_sums_gx2[(_col)] += gx * (double)gx; \ + col_sums_gy2[(_col)] += gy * (double)gy; \ + col_sums_gxgy[(_col)] += gx * (double)gy; \ + } while (0) + +#define FS_COL_SUB(_col, _joffs, _ioffs) \ + do { \ + unsigned gx; \ + unsigned gy; \ + gx = gx_buf[((j + (_joffs)) & 7) * stride + i + (_ioffs)]; \ + gy = gy_buf[((j + (_joffs)) & 7) * stride + i + (_ioffs)]; \ + col_sums_gx2[(_col)] -= gx * (double)gx; \ + col_sums_gy2[(_col)] -= gy * (double)gy; \ + col_sums_gxgy[(_col)] -= gx * (double)gy; \ + } while (0) + +#define FS_COL_COPY(_col1, _col2) \ + do { \ + col_sums_gx2[(_col1)] = col_sums_gx2[(_col2)]; \ + col_sums_gy2[(_col1)] = col_sums_gy2[(_col2)]; \ + col_sums_gxgy[(_col1)] = col_sums_gxgy[(_col2)]; \ + } while (0) + +#define FS_COL_HALVE(_col1, _col2) \ + do { \ + col_sums_gx2[(_col1)] = col_sums_gx2[(_col2)] * 0.5; \ + col_sums_gy2[(_col1)] = col_sums_gy2[(_col2)] * 0.5; \ + col_sums_gxgy[(_col1)] = col_sums_gxgy[(_col2)] * 0.5; \ + } while (0) + +#define FS_COL_DOUBLE(_col1, _col2) \ + do { \ + col_sums_gx2[(_col1)] = col_sums_gx2[(_col2)] * 2; \ + col_sums_gy2[(_col1)] = col_sums_gy2[(_col2)] * 2; \ + col_sums_gxgy[(_col1)] = col_sums_gxgy[(_col2)] * 2; \ + } while (0) + +static void fs_calc_structure(fs_ctx *_ctx, int _l, int bit_depth) { + uint32_t *im1; + uint32_t *im2; + unsigned *gx_buf; + unsigned *gy_buf; + double *ssim; + double col_sums_gx2[8]; + double col_sums_gy2[8]; + double col_sums_gxgy[8]; + double c2; + int stride; + int w; + int h; + int i; + int j; + double ssim_c2 = SSIM_C2; + if (bit_depth == 10) ssim_c2 = SSIM_C2_10; + if (bit_depth == 12) ssim_c2 = SSIM_C2_12; + + w = _ctx->level[_l].w; + h = _ctx->level[_l].h; + im1 = _ctx->level[_l].im1; + im2 = _ctx->level[_l].im2; + ssim = _ctx->level[_l].ssim; + gx_buf = _ctx->col_buf; + stride = w + 8; + gy_buf = gx_buf + 8 * stride; + memset(gx_buf, 0, 2 * 8 * stride * sizeof(*gx_buf)); + c2 = ssim_c2 * (1 << 4 * _l) * 16 * 104; + for (j = 0; j < h + 4; j++) { + if (j < h - 1) { + for (i = 0; i < w - 1; i++) { + unsigned g1; + unsigned g2; + unsigned gx; + unsigned gy; + g1 = abs((int)im1[(j + 1) * w + i + 1] - (int)im1[j * w + i]); + g2 = abs((int)im1[(j + 1) * w + i] - (int)im1[j * w + i + 1]); + gx = 4 * FS_MAXI(g1, g2) + FS_MINI(g1, g2); + g1 = abs((int)im2[(j + 1) * w + i + 1] - (int)im2[j * w + i]); + g2 = abs((int)im2[(j + 1) * w + i] - (int)im2[j * w + i + 1]); + gy = 4 * FS_MAXI(g1, g2) + FS_MINI(g1, g2); + gx_buf[(j & 7) * stride + i + 4] = gx; + gy_buf[(j & 7) * stride + i + 4] = gy; + } + } else { + memset(gx_buf + (j & 7) * stride, 0, stride * sizeof(*gx_buf)); + memset(gy_buf + (j & 7) * stride, 0, stride * sizeof(*gy_buf)); + } + if (j >= 4) { + int k; + col_sums_gx2[3] = col_sums_gx2[2] = col_sums_gx2[1] = col_sums_gx2[0] = 0; + col_sums_gy2[3] = col_sums_gy2[2] = col_sums_gy2[1] = col_sums_gy2[0] = 0; + col_sums_gxgy[3] = col_sums_gxgy[2] = col_sums_gxgy[1] = + col_sums_gxgy[0] = 0; + for (i = 4; i < 8; i++) { + FS_COL_SET(i, -1, 0); + FS_COL_ADD(i, 0, 0); + for (k = 1; k < 8 - i; k++) { + FS_COL_DOUBLE(i, i); + FS_COL_ADD(i, -k - 1, 0); + FS_COL_ADD(i, k, 0); + } + } + for (i = 0; i < w; i++) { + double mugx2; + double mugy2; + double mugxgy; + mugx2 = col_sums_gx2[0]; + for (k = 1; k < 8; k++) mugx2 += col_sums_gx2[k]; + mugy2 = col_sums_gy2[0]; + for (k = 1; k < 8; k++) mugy2 += col_sums_gy2[k]; + mugxgy = col_sums_gxgy[0]; + for (k = 1; k < 8; k++) mugxgy += col_sums_gxgy[k]; + ssim[(j - 4) * w + i] = (2 * mugxgy + c2) / (mugx2 + mugy2 + c2); + if (i + 1 < w) { + FS_COL_SET(0, -1, 1); + FS_COL_ADD(0, 0, 1); + FS_COL_SUB(2, -3, 2); + FS_COL_SUB(2, 2, 2); + FS_COL_HALVE(1, 2); + FS_COL_SUB(3, -4, 3); + FS_COL_SUB(3, 3, 3); + FS_COL_HALVE(2, 3); + FS_COL_COPY(3, 4); + FS_COL_DOUBLE(4, 5); + FS_COL_ADD(4, -4, 5); + FS_COL_ADD(4, 3, 5); + FS_COL_DOUBLE(5, 6); + FS_COL_ADD(5, -3, 6); + FS_COL_ADD(5, 2, 6); + FS_COL_DOUBLE(6, 7); + FS_COL_ADD(6, -2, 7); + FS_COL_ADD(6, 1, 7); + FS_COL_SET(7, -1, 8); + FS_COL_ADD(7, 0, 8); + } + } + } + } +} + +#define FS_NLEVELS (4) + +/*These weights were derived from the default weights found in Wang's original + Matlab implementation: {0.0448, 0.2856, 0.2363, 0.1333}. + We drop the finest scale and renormalize the rest to sum to 1.*/ + +static const double FS_WEIGHTS[FS_NLEVELS] = { + 0.2989654541015625, 0.3141326904296875, 0.2473602294921875, 0.1395416259765625 +}; + +static double fs_average(fs_ctx *_ctx, int _l) { + double *ssim; + double ret; + int w; + int h; + int i; + int j; + w = _ctx->level[_l].w; + h = _ctx->level[_l].h; + ssim = _ctx->level[_l].ssim; + ret = 0; + for (j = 0; j < h; j++) + for (i = 0; i < w; i++) ret += ssim[j * w + i]; + return pow(ret / (w * h), FS_WEIGHTS[_l]); +} + +static double convert_ssim_db(double _ssim, double _weight) { + assert(_weight >= _ssim); + if ((_weight - _ssim) < 1e-10) return MAX_SSIM_DB; + return 10 * (log10(_weight) - log10(_weight - _ssim)); +} + +static double calc_ssim(const uint8_t *_src, int _systride, const uint8_t *_dst, + int _dystride, int _w, int _h, uint32_t _bd, + uint32_t _shift, int buf_is_hbd) { + fs_ctx ctx; + double ret; + int l; + ret = 1; + fs_ctx_init(&ctx, _w, _h, FS_NLEVELS); + fs_downsample_level0(&ctx, _src, _systride, _dst, _dystride, _w, _h, _shift, + buf_is_hbd); + for (l = 0; l < FS_NLEVELS - 1; l++) { + fs_calc_structure(&ctx, l, _bd); + ret *= fs_average(&ctx, l); + fs_downsample_level(&ctx, l + 1); + } + fs_calc_structure(&ctx, l, _bd); + fs_apply_luminance(&ctx, l, _bd); + ret *= fs_average(&ctx, l); + fs_ctx_clear(&ctx); + return ret; +} + +double aom_calc_fastssim(const YV12_BUFFER_CONFIG *source, + const YV12_BUFFER_CONFIG *dest, double *ssim_y, + double *ssim_u, double *ssim_v, uint32_t bd, + uint32_t in_bd) { + double ssimv; + uint32_t bd_shift = 0; + aom_clear_system_state(); + assert(bd >= in_bd); + assert(source->flags == dest->flags); + int buf_is_hbd = source->flags & YV12_FLAG_HIGHBITDEPTH; + bd_shift = bd - in_bd; + + *ssim_y = calc_ssim(source->y_buffer, source->y_stride, dest->y_buffer, + dest->y_stride, source->y_crop_width, + source->y_crop_height, in_bd, bd_shift, buf_is_hbd); + *ssim_u = calc_ssim(source->u_buffer, source->uv_stride, dest->u_buffer, + dest->uv_stride, source->uv_crop_width, + source->uv_crop_height, in_bd, bd_shift, buf_is_hbd); + *ssim_v = calc_ssim(source->v_buffer, source->uv_stride, dest->v_buffer, + dest->uv_stride, source->uv_crop_width, + source->uv_crop_height, in_bd, bd_shift, buf_is_hbd); + ssimv = (*ssim_y) * .8 + .1 * ((*ssim_u) + (*ssim_v)); + return convert_ssim_db(ssimv, 1.0); +} diff --git a/third_party/aom/aom_dsp/fft.c b/third_party/aom/aom_dsp/fft.c new file mode 100644 index 0000000000..0ba71cfb34 --- /dev/null +++ b/third_party/aom/aom_dsp/fft.c @@ -0,0 +1,219 @@ +/* + * Copyright (c) 2018, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include "aom_dsp/aom_dsp_common.h" +#include "aom_dsp/fft_common.h" + +static INLINE void simple_transpose(const float *A, float *B, int n) { + for (int y = 0; y < n; y++) { + for (int x = 0; x < n; x++) { + B[y * n + x] = A[x * n + y]; + } + } +} + +// The 1d transform is real to complex and packs the complex results in +// a way to take advantage of conjugate symmetry (e.g., the n/2 + 1 real +// components, followed by the n/2 - 1 imaginary components). After the +// transform is done on the rows, the first n/2 + 1 columns are real, and +// the remaining are the imaginary components. After the transform on the +// columns, the region of [0, n/2]x[0, n/2] contains the real part of +// fft of the real columns. The real part of the 2d fft also includes the +// imaginary part of transformed imaginary columns. This function assembles +// the correct outputs while putting the real and imaginary components +// next to each other. +static INLINE void unpack_2d_output(const float *col_fft, float *output, + int n) { + for (int y = 0; y <= n / 2; ++y) { + const int y2 = y + n / 2; + const int y_extra = y2 > n / 2 && y2 < n; + + for (int x = 0; x <= n / 2; ++x) { + const int x2 = x + n / 2; + const int x_extra = x2 > n / 2 && x2 < n; + output[2 * (y * n + x)] = + col_fft[y * n + x] - (x_extra && y_extra ? col_fft[y2 * n + x2] : 0); + output[2 * (y * n + x) + 1] = (y_extra ? col_fft[y2 * n + x] : 0) + + (x_extra ? col_fft[y * n + x2] : 0); + if (y_extra) { + output[2 * ((n - y) * n + x)] = + col_fft[y * n + x] + + (x_extra && y_extra ? col_fft[y2 * n + x2] : 0); + output[2 * ((n - y) * n + x) + 1] = + -(y_extra ? col_fft[y2 * n + x] : 0) + + (x_extra ? col_fft[y * n + x2] : 0); + } + } + } +} + +void aom_fft_2d_gen(const float *input, float *temp, float *output, int n, + aom_fft_1d_func_t tform, aom_fft_transpose_func_t transpose, + aom_fft_unpack_func_t unpack, int vec_size) { + for (int x = 0; x < n; x += vec_size) { + tform(input + x, output + x, n); + } + transpose(output, temp, n); + + for (int x = 0; x < n; x += vec_size) { + tform(temp + x, output + x, n); + } + transpose(output, temp, n); + + unpack(temp, output, n); +} + +static INLINE void store_float(float *output, float input) { *output = input; } +static INLINE float add_float(float a, float b) { return a + b; } +static INLINE float sub_float(float a, float b) { return a - b; } +static INLINE float mul_float(float a, float b) { return a * b; } + +GEN_FFT_2(void, float, float, float, *, store_float); +GEN_FFT_4(void, float, float, float, *, store_float, (float), add_float, + sub_float); +GEN_FFT_8(void, float, float, float, *, store_float, (float), add_float, + sub_float, mul_float); +GEN_FFT_16(void, float, float, float, *, store_float, (float), add_float, + sub_float, mul_float); +GEN_FFT_32(void, float, float, float, *, store_float, (float), add_float, + sub_float, mul_float); + +void aom_fft2x2_float_c(const float *input, float *temp, float *output) { + aom_fft_2d_gen(input, temp, output, 2, aom_fft1d_2_float, simple_transpose, + unpack_2d_output, 1); +} + +void aom_fft4x4_float_c(const float *input, float *temp, float *output) { + aom_fft_2d_gen(input, temp, output, 4, aom_fft1d_4_float, simple_transpose, + unpack_2d_output, 1); +} + +void aom_fft8x8_float_c(const float *input, float *temp, float *output) { + aom_fft_2d_gen(input, temp, output, 8, aom_fft1d_8_float, simple_transpose, + unpack_2d_output, 1); +} + +void aom_fft16x16_float_c(const float *input, float *temp, float *output) { + aom_fft_2d_gen(input, temp, output, 16, aom_fft1d_16_float, simple_transpose, + unpack_2d_output, 1); +} + +void aom_fft32x32_float_c(const float *input, float *temp, float *output) { + aom_fft_2d_gen(input, temp, output, 32, aom_fft1d_32_float, simple_transpose, + unpack_2d_output, 1); +} + +void aom_ifft_2d_gen(const float *input, float *temp, float *output, int n, + aom_fft_1d_func_t fft_single, aom_fft_1d_func_t fft_multi, + aom_fft_1d_func_t ifft_multi, + aom_fft_transpose_func_t transpose, int vec_size) { + // Column 0 and n/2 have conjugate symmetry, so we can directly do the ifft + // and get real outputs. + for (int y = 0; y <= n / 2; ++y) { + output[y * n] = input[2 * y * n]; + output[y * n + 1] = input[2 * (y * n + n / 2)]; + } + for (int y = n / 2 + 1; y < n; ++y) { + output[y * n] = input[2 * (y - n / 2) * n + 1]; + output[y * n + 1] = input[2 * ((y - n / 2) * n + n / 2) + 1]; + } + + for (int i = 0; i < 2; i += vec_size) { + ifft_multi(output + i, temp + i, n); + } + + // For the other columns, since we don't have a full ifft for complex inputs + // we have to split them into the real and imaginary counterparts. + // Pack the real component, then the imaginary components. + for (int y = 0; y < n; ++y) { + for (int x = 1; x < n / 2; ++x) { + output[y * n + (x + 1)] = input[2 * (y * n + x)]; + } + for (int x = 1; x < n / 2; ++x) { + output[y * n + (x + n / 2)] = input[2 * (y * n + x) + 1]; + } + } + for (int y = 2; y < vec_size; y++) { + fft_single(output + y, temp + y, n); + } + // This is the part that can be sped up with SIMD + for (int y = AOMMAX(2, vec_size); y < n; y += vec_size) { + fft_multi(output + y, temp + y, n); + } + + // Put the 0 and n/2 th results in the correct place. + for (int x = 0; x < n; ++x) { + output[x] = temp[x * n]; + output[(n / 2) * n + x] = temp[x * n + 1]; + } + // This rearranges and transposes. + for (int y = 1; y < n / 2; ++y) { + // Fill in the real columns + for (int x = 0; x <= n / 2; ++x) { + output[x + y * n] = + temp[(y + 1) + x * n] + + ((x > 0 && x < n / 2) ? temp[(y + n / 2) + (x + n / 2) * n] : 0); + } + for (int x = n / 2 + 1; x < n; ++x) { + output[x + y * n] = temp[(y + 1) + (n - x) * n] - + temp[(y + n / 2) + ((n - x) + n / 2) * n]; + } + // Fill in the imag columns + for (int x = 0; x <= n / 2; ++x) { + output[x + (y + n / 2) * n] = + temp[(y + n / 2) + x * n] - + ((x > 0 && x < n / 2) ? temp[(y + 1) + (x + n / 2) * n] : 0); + } + for (int x = n / 2 + 1; x < n; ++x) { + output[x + (y + n / 2) * n] = temp[(y + 1) + ((n - x) + n / 2) * n] + + temp[(y + n / 2) + (n - x) * n]; + } + } + for (int y = 0; y < n; y += vec_size) { + ifft_multi(output + y, temp + y, n); + } + transpose(temp, output, n); +} + +GEN_IFFT_2(void, float, float, float, *, store_float); +GEN_IFFT_4(void, float, float, float, *, store_float, (float), add_float, + sub_float); +GEN_IFFT_8(void, float, float, float, *, store_float, (float), add_float, + sub_float, mul_float); +GEN_IFFT_16(void, float, float, float, *, store_float, (float), add_float, + sub_float, mul_float); +GEN_IFFT_32(void, float, float, float, *, store_float, (float), add_float, + sub_float, mul_float); + +void aom_ifft2x2_float_c(const float *input, float *temp, float *output) { + aom_ifft_2d_gen(input, temp, output, 2, aom_fft1d_2_float, aom_fft1d_2_float, + aom_ifft1d_2_float, simple_transpose, 1); +} + +void aom_ifft4x4_float_c(const float *input, float *temp, float *output) { + aom_ifft_2d_gen(input, temp, output, 4, aom_fft1d_4_float, aom_fft1d_4_float, + aom_ifft1d_4_float, simple_transpose, 1); +} + +void aom_ifft8x8_float_c(const float *input, float *temp, float *output) { + aom_ifft_2d_gen(input, temp, output, 8, aom_fft1d_8_float, aom_fft1d_8_float, + aom_ifft1d_8_float, simple_transpose, 1); +} + +void aom_ifft16x16_float_c(const float *input, float *temp, float *output) { + aom_ifft_2d_gen(input, temp, output, 16, aom_fft1d_16_float, + aom_fft1d_16_float, aom_ifft1d_16_float, simple_transpose, 1); +} + +void aom_ifft32x32_float_c(const float *input, float *temp, float *output) { + aom_ifft_2d_gen(input, temp, output, 32, aom_fft1d_32_float, + aom_fft1d_32_float, aom_ifft1d_32_float, simple_transpose, 1); +} diff --git a/third_party/aom/aom_dsp/fft_common.h b/third_party/aom/aom_dsp/fft_common.h new file mode 100644 index 0000000000..5137331aec --- /dev/null +++ b/third_party/aom/aom_dsp/fft_common.h @@ -0,0 +1,1050 @@ +/* + * Copyright (c) 2018, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_FFT_COMMON_H_ +#define AOM_AOM_DSP_FFT_COMMON_H_ + +#ifdef __cplusplus +extern "C" { +#endif + +/*!\brief A function pointer for computing 1d fft and ifft. + * + * The function will point to an implementation for a specific transform size, + * and may perform the transforms using vectorized instructions. + * + * For a non-vectorized forward transforms of size n, the input and output + * buffers will be size n. The output takes advantage of conjugate symmetry and + * packs the results as: [r_0, r_1, ..., r_{n/2}, i_1, ..., i_{n/2-1}], where + * (r_{j}, i_{j}) is the complex output for index j. + * + * An inverse transform will assume that the complex "input" is packed + * similarly. Its output will be real. + * + * Non-vectorized transforms (e.g., on a single row) would use a stride = 1. + * + * Vectorized implementations are parallelized along the columns so that the fft + * can be performed on multiple columns at a time. In such cases the data block + * for input and output is typically square (n x n) and the stride will + * correspond to the spacing between rows. At minimum, the input size must be + * n x simd_vector_length. + * + * \param[in] input Input buffer. See above for size restrictions. + * \param[out] output Output buffer. See above for size restrictions. + * \param[in] stride The spacing in number of elements between rows + * (or elements) + */ +typedef void (*aom_fft_1d_func_t)(const float *input, float *output, + int stride); + +// Declare some of the forward non-vectorized transforms which are used in some +// of the vectorized implementations +void aom_fft1d_4_float(const float *input, float *output, int stride); +void aom_fft1d_8_float(const float *input, float *output, int stride); +void aom_fft1d_16_float(const float *input, float *output, int stride); +void aom_fft1d_32_float(const float *input, float *output, int stride); + +/**\!brief Function pointer for transposing a matrix of floats. + * + * \param[in] input Input buffer (size n x n) + * \param[out] output Output buffer (size n x n) + * \param[in] n Extent of one dimension of the square matrix. + */ +typedef void (*aom_fft_transpose_func_t)(const float *input, float *output, + int n); + +/**\!brief Function pointer for re-arranging intermediate 2d transform results. + * + * After re-arrangement, the real and imaginary components will be packed + * tightly next to each other. + * + * \param[in] input Input buffer (size n x n) + * \param[out] output Output buffer (size 2 x n x n) + * \param[in] n Extent of one dimension of the square matrix. + */ +typedef void (*aom_fft_unpack_func_t)(const float *input, float *output, int n); + +/*!\brief Performs a 2d fft with the given functions. + * + * This generator function allows for multiple different implementations of 2d + * fft with different vector operations, without having to redefine the main + * body multiple times. + * + * \param[in] input Input buffer to run the transform on (size n x n) + * \param[out] temp Working buffer for computing the transform (size n x n) + * \param[out] output Output buffer (size 2 x n x n) + * \param[in] tform Forward transform function + * \param[in] transpose Transpose function (for n x n matrix) + * \param[in] unpack Unpack function used to massage outputs to correct form + * \param[in] vec_size Vector size (the transform is done vec_size units at + * a time) + */ +void aom_fft_2d_gen(const float *input, float *temp, float *output, int n, + aom_fft_1d_func_t tform, aom_fft_transpose_func_t transpose, + aom_fft_unpack_func_t unpack, int vec_size); + +/*!\brief Perform a 2d inverse fft with the given helper functions + * + * \param[in] input Input buffer to run the transform on (size 2 x n x n) + * \param[out] temp Working buffer for computations (size 2 x n x n) + * \param[out] output Output buffer (size n x n) + * \param[in] fft_single Forward transform function (non vectorized) + * \param[in] fft_multi Forward transform function (vectorized) + * \param[in] ifft_multi Inverse transform function (vectorized) + * \param[in] transpose Transpose function (for n x n matrix) + * \param[in] vec_size Vector size (the transform is done vec_size + * units at a time) + */ +void aom_ifft_2d_gen(const float *input, float *temp, float *output, int n, + aom_fft_1d_func_t fft_single, aom_fft_1d_func_t fft_multi, + aom_fft_1d_func_t ifft_multi, + aom_fft_transpose_func_t transpose, int vec_size); +#ifdef __cplusplus +} +#endif + +// The macros below define 1D fft/ifft for different data types and for +// different simd vector intrinsic types. + +#define GEN_FFT_2(ret, suffix, T, T_VEC, load, store) \ + ret aom_fft1d_2_##suffix(const T *input, T *output, int stride) { \ + const T_VEC i0 = load(input + 0 * stride); \ + const T_VEC i1 = load(input + 1 * stride); \ + store(output + 0 * stride, i0 + i1); \ + store(output + 1 * stride, i0 - i1); \ + } + +#define GEN_FFT_4(ret, suffix, T, T_VEC, load, store, constant, add, sub) \ + ret aom_fft1d_4_##suffix(const T *input, T *output, int stride) { \ + const T_VEC kWeight0 = constant(0.0f); \ + const T_VEC i0 = load(input + 0 * stride); \ + const T_VEC i1 = load(input + 1 * stride); \ + const T_VEC i2 = load(input + 2 * stride); \ + const T_VEC i3 = load(input + 3 * stride); \ + const T_VEC w0 = add(i0, i2); \ + const T_VEC w1 = sub(i0, i2); \ + const T_VEC w2 = add(i1, i3); \ + const T_VEC w3 = sub(i1, i3); \ + store(output + 0 * stride, add(w0, w2)); \ + store(output + 1 * stride, w1); \ + store(output + 2 * stride, sub(w0, w2)); \ + store(output + 3 * stride, sub(kWeight0, w3)); \ + } + +#define GEN_FFT_8(ret, suffix, T, T_VEC, load, store, constant, add, sub, mul) \ + ret aom_fft1d_8_##suffix(const T *input, T *output, int stride) { \ + const T_VEC kWeight0 = constant(0.0f); \ + const T_VEC kWeight2 = constant(0.707107f); \ + const T_VEC i0 = load(input + 0 * stride); \ + const T_VEC i1 = load(input + 1 * stride); \ + const T_VEC i2 = load(input + 2 * stride); \ + const T_VEC i3 = load(input + 3 * stride); \ + const T_VEC i4 = load(input + 4 * stride); \ + const T_VEC i5 = load(input + 5 * stride); \ + const T_VEC i6 = load(input + 6 * stride); \ + const T_VEC i7 = load(input + 7 * stride); \ + const T_VEC w0 = add(i0, i4); \ + const T_VEC w1 = sub(i0, i4); \ + const T_VEC w2 = add(i2, i6); \ + const T_VEC w3 = sub(i2, i6); \ + const T_VEC w4 = add(w0, w2); \ + const T_VEC w5 = sub(w0, w2); \ + const T_VEC w7 = add(i1, i5); \ + const T_VEC w8 = sub(i1, i5); \ + const T_VEC w9 = add(i3, i7); \ + const T_VEC w10 = sub(i3, i7); \ + const T_VEC w11 = add(w7, w9); \ + const T_VEC w12 = sub(w7, w9); \ + store(output + 0 * stride, add(w4, w11)); \ + store(output + 1 * stride, add(w1, mul(kWeight2, sub(w8, w10)))); \ + store(output + 2 * stride, w5); \ + store(output + 3 * stride, sub(w1, mul(kWeight2, sub(w8, w10)))); \ + store(output + 4 * stride, sub(w4, w11)); \ + store(output + 5 * stride, \ + sub(sub(kWeight0, w3), mul(kWeight2, add(w10, w8)))); \ + store(output + 6 * stride, sub(kWeight0, w12)); \ + store(output + 7 * stride, sub(w3, mul(kWeight2, add(w10, w8)))); \ + } + +#define GEN_FFT_16(ret, suffix, T, T_VEC, load, store, constant, add, sub, \ + mul) \ + ret aom_fft1d_16_##suffix(const T *input, T *output, int stride) { \ + const T_VEC kWeight0 = constant(0.0f); \ + const T_VEC kWeight2 = constant(0.707107f); \ + const T_VEC kWeight3 = constant(0.92388f); \ + const T_VEC kWeight4 = constant(0.382683f); \ + const T_VEC i0 = load(input + 0 * stride); \ + const T_VEC i1 = load(input + 1 * stride); \ + const T_VEC i2 = load(input + 2 * stride); \ + const T_VEC i3 = load(input + 3 * stride); \ + const T_VEC i4 = load(input + 4 * stride); \ + const T_VEC i5 = load(input + 5 * stride); \ + const T_VEC i6 = load(input + 6 * stride); \ + const T_VEC i7 = load(input + 7 * stride); \ + const T_VEC i8 = load(input + 8 * stride); \ + const T_VEC i9 = load(input + 9 * stride); \ + const T_VEC i10 = load(input + 10 * stride); \ + const T_VEC i11 = load(input + 11 * stride); \ + const T_VEC i12 = load(input + 12 * stride); \ + const T_VEC i13 = load(input + 13 * stride); \ + const T_VEC i14 = load(input + 14 * stride); \ + const T_VEC i15 = load(input + 15 * stride); \ + const T_VEC w0 = add(i0, i8); \ + const T_VEC w1 = sub(i0, i8); \ + const T_VEC w2 = add(i4, i12); \ + const T_VEC w3 = sub(i4, i12); \ + const T_VEC w4 = add(w0, w2); \ + const T_VEC w5 = sub(w0, w2); \ + const T_VEC w7 = add(i2, i10); \ + const T_VEC w8 = sub(i2, i10); \ + const T_VEC w9 = add(i6, i14); \ + const T_VEC w10 = sub(i6, i14); \ + const T_VEC w11 = add(w7, w9); \ + const T_VEC w12 = sub(w7, w9); \ + const T_VEC w14 = add(w4, w11); \ + const T_VEC w15 = sub(w4, w11); \ + const T_VEC w16[2] = { add(w1, mul(kWeight2, sub(w8, w10))), \ + sub(sub(kWeight0, w3), \ + mul(kWeight2, add(w10, w8))) }; \ + const T_VEC w18[2] = { sub(w1, mul(kWeight2, sub(w8, w10))), \ + sub(w3, mul(kWeight2, add(w10, w8))) }; \ + const T_VEC w19 = add(i1, i9); \ + const T_VEC w20 = sub(i1, i9); \ + const T_VEC w21 = add(i5, i13); \ + const T_VEC w22 = sub(i5, i13); \ + const T_VEC w23 = add(w19, w21); \ + const T_VEC w24 = sub(w19, w21); \ + const T_VEC w26 = add(i3, i11); \ + const T_VEC w27 = sub(i3, i11); \ + const T_VEC w28 = add(i7, i15); \ + const T_VEC w29 = sub(i7, i15); \ + const T_VEC w30 = add(w26, w28); \ + const T_VEC w31 = sub(w26, w28); \ + const T_VEC w33 = add(w23, w30); \ + const T_VEC w34 = sub(w23, w30); \ + const T_VEC w35[2] = { add(w20, mul(kWeight2, sub(w27, w29))), \ + sub(sub(kWeight0, w22), \ + mul(kWeight2, add(w29, w27))) }; \ + const T_VEC w37[2] = { sub(w20, mul(kWeight2, sub(w27, w29))), \ + sub(w22, mul(kWeight2, add(w29, w27))) }; \ + store(output + 0 * stride, add(w14, w33)); \ + store(output + 1 * stride, \ + add(w16[0], add(mul(kWeight3, w35[0]), mul(kWeight4, w35[1])))); \ + store(output + 2 * stride, add(w5, mul(kWeight2, sub(w24, w31)))); \ + store(output + 3 * stride, \ + add(w18[0], add(mul(kWeight4, w37[0]), mul(kWeight3, w37[1])))); \ + store(output + 4 * stride, w15); \ + store(output + 5 * stride, \ + add(w18[0], sub(sub(kWeight0, mul(kWeight4, w37[0])), \ + mul(kWeight3, w37[1])))); \ + store(output + 6 * stride, sub(w5, mul(kWeight2, sub(w24, w31)))); \ + store(output + 7 * stride, \ + add(w16[0], sub(sub(kWeight0, mul(kWeight3, w35[0])), \ + mul(kWeight4, w35[1])))); \ + store(output + 8 * stride, sub(w14, w33)); \ + store(output + 9 * stride, \ + add(w16[1], sub(mul(kWeight3, w35[1]), mul(kWeight4, w35[0])))); \ + store(output + 10 * stride, \ + sub(sub(kWeight0, w12), mul(kWeight2, add(w31, w24)))); \ + store(output + 11 * stride, \ + add(w18[1], sub(mul(kWeight4, w37[1]), mul(kWeight3, w37[0])))); \ + store(output + 12 * stride, sub(kWeight0, w34)); \ + store(output + 13 * stride, \ + sub(sub(kWeight0, w18[1]), \ + sub(mul(kWeight3, w37[0]), mul(kWeight4, w37[1])))); \ + store(output + 14 * stride, sub(w12, mul(kWeight2, add(w31, w24)))); \ + store(output + 15 * stride, \ + sub(sub(kWeight0, w16[1]), \ + sub(mul(kWeight4, w35[0]), mul(kWeight3, w35[1])))); \ + } + +#define GEN_FFT_32(ret, suffix, T, T_VEC, load, store, constant, add, sub, \ + mul) \ + ret aom_fft1d_32_##suffix(const T *input, T *output, int stride) { \ + const T_VEC kWeight0 = constant(0.0f); \ + const T_VEC kWeight2 = constant(0.707107f); \ + const T_VEC kWeight3 = constant(0.92388f); \ + const T_VEC kWeight4 = constant(0.382683f); \ + const T_VEC kWeight5 = constant(0.980785f); \ + const T_VEC kWeight6 = constant(0.19509f); \ + const T_VEC kWeight7 = constant(0.83147f); \ + const T_VEC kWeight8 = constant(0.55557f); \ + const T_VEC i0 = load(input + 0 * stride); \ + const T_VEC i1 = load(input + 1 * stride); \ + const T_VEC i2 = load(input + 2 * stride); \ + const T_VEC i3 = load(input + 3 * stride); \ + const T_VEC i4 = load(input + 4 * stride); \ + const T_VEC i5 = load(input + 5 * stride); \ + const T_VEC i6 = load(input + 6 * stride); \ + const T_VEC i7 = load(input + 7 * stride); \ + const T_VEC i8 = load(input + 8 * stride); \ + const T_VEC i9 = load(input + 9 * stride); \ + const T_VEC i10 = load(input + 10 * stride); \ + const T_VEC i11 = load(input + 11 * stride); \ + const T_VEC i12 = load(input + 12 * stride); \ + const T_VEC i13 = load(input + 13 * stride); \ + const T_VEC i14 = load(input + 14 * stride); \ + const T_VEC i15 = load(input + 15 * stride); \ + const T_VEC i16 = load(input + 16 * stride); \ + const T_VEC i17 = load(input + 17 * stride); \ + const T_VEC i18 = load(input + 18 * stride); \ + const T_VEC i19 = load(input + 19 * stride); \ + const T_VEC i20 = load(input + 20 * stride); \ + const T_VEC i21 = load(input + 21 * stride); \ + const T_VEC i22 = load(input + 22 * stride); \ + const T_VEC i23 = load(input + 23 * stride); \ + const T_VEC i24 = load(input + 24 * stride); \ + const T_VEC i25 = load(input + 25 * stride); \ + const T_VEC i26 = load(input + 26 * stride); \ + const T_VEC i27 = load(input + 27 * stride); \ + const T_VEC i28 = load(input + 28 * stride); \ + const T_VEC i29 = load(input + 29 * stride); \ + const T_VEC i30 = load(input + 30 * stride); \ + const T_VEC i31 = load(input + 31 * stride); \ + const T_VEC w0 = add(i0, i16); \ + const T_VEC w1 = sub(i0, i16); \ + const T_VEC w2 = add(i8, i24); \ + const T_VEC w3 = sub(i8, i24); \ + const T_VEC w4 = add(w0, w2); \ + const T_VEC w5 = sub(w0, w2); \ + const T_VEC w7 = add(i4, i20); \ + const T_VEC w8 = sub(i4, i20); \ + const T_VEC w9 = add(i12, i28); \ + const T_VEC w10 = sub(i12, i28); \ + const T_VEC w11 = add(w7, w9); \ + const T_VEC w12 = sub(w7, w9); \ + const T_VEC w14 = add(w4, w11); \ + const T_VEC w15 = sub(w4, w11); \ + const T_VEC w16[2] = { add(w1, mul(kWeight2, sub(w8, w10))), \ + sub(sub(kWeight0, w3), \ + mul(kWeight2, add(w10, w8))) }; \ + const T_VEC w18[2] = { sub(w1, mul(kWeight2, sub(w8, w10))), \ + sub(w3, mul(kWeight2, add(w10, w8))) }; \ + const T_VEC w19 = add(i2, i18); \ + const T_VEC w20 = sub(i2, i18); \ + const T_VEC w21 = add(i10, i26); \ + const T_VEC w22 = sub(i10, i26); \ + const T_VEC w23 = add(w19, w21); \ + const T_VEC w24 = sub(w19, w21); \ + const T_VEC w26 = add(i6, i22); \ + const T_VEC w27 = sub(i6, i22); \ + const T_VEC w28 = add(i14, i30); \ + const T_VEC w29 = sub(i14, i30); \ + const T_VEC w30 = add(w26, w28); \ + const T_VEC w31 = sub(w26, w28); \ + const T_VEC w33 = add(w23, w30); \ + const T_VEC w34 = sub(w23, w30); \ + const T_VEC w35[2] = { add(w20, mul(kWeight2, sub(w27, w29))), \ + sub(sub(kWeight0, w22), \ + mul(kWeight2, add(w29, w27))) }; \ + const T_VEC w37[2] = { sub(w20, mul(kWeight2, sub(w27, w29))), \ + sub(w22, mul(kWeight2, add(w29, w27))) }; \ + const T_VEC w38 = add(w14, w33); \ + const T_VEC w39 = sub(w14, w33); \ + const T_VEC w40[2] = { \ + add(w16[0], add(mul(kWeight3, w35[0]), mul(kWeight4, w35[1]))), \ + add(w16[1], sub(mul(kWeight3, w35[1]), mul(kWeight4, w35[0]))) \ + }; \ + const T_VEC w41[2] = { add(w5, mul(kWeight2, sub(w24, w31))), \ + sub(sub(kWeight0, w12), \ + mul(kWeight2, add(w31, w24))) }; \ + const T_VEC w42[2] = { \ + add(w18[0], add(mul(kWeight4, w37[0]), mul(kWeight3, w37[1]))), \ + add(w18[1], sub(mul(kWeight4, w37[1]), mul(kWeight3, w37[0]))) \ + }; \ + const T_VEC w44[2] = { \ + add(w18[0], \ + sub(sub(kWeight0, mul(kWeight4, w37[0])), mul(kWeight3, w37[1]))), \ + sub(sub(kWeight0, w18[1]), \ + sub(mul(kWeight3, w37[0]), mul(kWeight4, w37[1]))) \ + }; \ + const T_VEC w45[2] = { sub(w5, mul(kWeight2, sub(w24, w31))), \ + sub(w12, mul(kWeight2, add(w31, w24))) }; \ + const T_VEC w46[2] = { \ + add(w16[0], \ + sub(sub(kWeight0, mul(kWeight3, w35[0])), mul(kWeight4, w35[1]))), \ + sub(sub(kWeight0, w16[1]), \ + sub(mul(kWeight4, w35[0]), mul(kWeight3, w35[1]))) \ + }; \ + const T_VEC w47 = add(i1, i17); \ + const T_VEC w48 = sub(i1, i17); \ + const T_VEC w49 = add(i9, i25); \ + const T_VEC w50 = sub(i9, i25); \ + const T_VEC w51 = add(w47, w49); \ + const T_VEC w52 = sub(w47, w49); \ + const T_VEC w54 = add(i5, i21); \ + const T_VEC w55 = sub(i5, i21); \ + const T_VEC w56 = add(i13, i29); \ + const T_VEC w57 = sub(i13, i29); \ + const T_VEC w58 = add(w54, w56); \ + const T_VEC w59 = sub(w54, w56); \ + const T_VEC w61 = add(w51, w58); \ + const T_VEC w62 = sub(w51, w58); \ + const T_VEC w63[2] = { add(w48, mul(kWeight2, sub(w55, w57))), \ + sub(sub(kWeight0, w50), \ + mul(kWeight2, add(w57, w55))) }; \ + const T_VEC w65[2] = { sub(w48, mul(kWeight2, sub(w55, w57))), \ + sub(w50, mul(kWeight2, add(w57, w55))) }; \ + const T_VEC w66 = add(i3, i19); \ + const T_VEC w67 = sub(i3, i19); \ + const T_VEC w68 = add(i11, i27); \ + const T_VEC w69 = sub(i11, i27); \ + const T_VEC w70 = add(w66, w68); \ + const T_VEC w71 = sub(w66, w68); \ + const T_VEC w73 = add(i7, i23); \ + const T_VEC w74 = sub(i7, i23); \ + const T_VEC w75 = add(i15, i31); \ + const T_VEC w76 = sub(i15, i31); \ + const T_VEC w77 = add(w73, w75); \ + const T_VEC w78 = sub(w73, w75); \ + const T_VEC w80 = add(w70, w77); \ + const T_VEC w81 = sub(w70, w77); \ + const T_VEC w82[2] = { add(w67, mul(kWeight2, sub(w74, w76))), \ + sub(sub(kWeight0, w69), \ + mul(kWeight2, add(w76, w74))) }; \ + const T_VEC w84[2] = { sub(w67, mul(kWeight2, sub(w74, w76))), \ + sub(w69, mul(kWeight2, add(w76, w74))) }; \ + const T_VEC w85 = add(w61, w80); \ + const T_VEC w86 = sub(w61, w80); \ + const T_VEC w87[2] = { \ + add(w63[0], add(mul(kWeight3, w82[0]), mul(kWeight4, w82[1]))), \ + add(w63[1], sub(mul(kWeight3, w82[1]), mul(kWeight4, w82[0]))) \ + }; \ + const T_VEC w88[2] = { add(w52, mul(kWeight2, sub(w71, w78))), \ + sub(sub(kWeight0, w59), \ + mul(kWeight2, add(w78, w71))) }; \ + const T_VEC w89[2] = { \ + add(w65[0], add(mul(kWeight4, w84[0]), mul(kWeight3, w84[1]))), \ + add(w65[1], sub(mul(kWeight4, w84[1]), mul(kWeight3, w84[0]))) \ + }; \ + const T_VEC w91[2] = { \ + add(w65[0], \ + sub(sub(kWeight0, mul(kWeight4, w84[0])), mul(kWeight3, w84[1]))), \ + sub(sub(kWeight0, w65[1]), \ + sub(mul(kWeight3, w84[0]), mul(kWeight4, w84[1]))) \ + }; \ + const T_VEC w92[2] = { sub(w52, mul(kWeight2, sub(w71, w78))), \ + sub(w59, mul(kWeight2, add(w78, w71))) }; \ + const T_VEC w93[2] = { \ + add(w63[0], \ + sub(sub(kWeight0, mul(kWeight3, w82[0])), mul(kWeight4, w82[1]))), \ + sub(sub(kWeight0, w63[1]), \ + sub(mul(kWeight4, w82[0]), mul(kWeight3, w82[1]))) \ + }; \ + store(output + 0 * stride, add(w38, w85)); \ + store(output + 1 * stride, \ + add(w40[0], add(mul(kWeight5, w87[0]), mul(kWeight6, w87[1])))); \ + store(output + 2 * stride, \ + add(w41[0], add(mul(kWeight3, w88[0]), mul(kWeight4, w88[1])))); \ + store(output + 3 * stride, \ + add(w42[0], add(mul(kWeight7, w89[0]), mul(kWeight8, w89[1])))); \ + store(output + 4 * stride, add(w15, mul(kWeight2, sub(w62, w81)))); \ + store(output + 5 * stride, \ + add(w44[0], add(mul(kWeight8, w91[0]), mul(kWeight7, w91[1])))); \ + store(output + 6 * stride, \ + add(w45[0], add(mul(kWeight4, w92[0]), mul(kWeight3, w92[1])))); \ + store(output + 7 * stride, \ + add(w46[0], add(mul(kWeight6, w93[0]), mul(kWeight5, w93[1])))); \ + store(output + 8 * stride, w39); \ + store(output + 9 * stride, \ + add(w46[0], sub(sub(kWeight0, mul(kWeight6, w93[0])), \ + mul(kWeight5, w93[1])))); \ + store(output + 10 * stride, \ + add(w45[0], sub(sub(kWeight0, mul(kWeight4, w92[0])), \ + mul(kWeight3, w92[1])))); \ + store(output + 11 * stride, \ + add(w44[0], sub(sub(kWeight0, mul(kWeight8, w91[0])), \ + mul(kWeight7, w91[1])))); \ + store(output + 12 * stride, sub(w15, mul(kWeight2, sub(w62, w81)))); \ + store(output + 13 * stride, \ + add(w42[0], sub(sub(kWeight0, mul(kWeight7, w89[0])), \ + mul(kWeight8, w89[1])))); \ + store(output + 14 * stride, \ + add(w41[0], sub(sub(kWeight0, mul(kWeight3, w88[0])), \ + mul(kWeight4, w88[1])))); \ + store(output + 15 * stride, \ + add(w40[0], sub(sub(kWeight0, mul(kWeight5, w87[0])), \ + mul(kWeight6, w87[1])))); \ + store(output + 16 * stride, sub(w38, w85)); \ + store(output + 17 * stride, \ + add(w40[1], sub(mul(kWeight5, w87[1]), mul(kWeight6, w87[0])))); \ + store(output + 18 * stride, \ + add(w41[1], sub(mul(kWeight3, w88[1]), mul(kWeight4, w88[0])))); \ + store(output + 19 * stride, \ + add(w42[1], sub(mul(kWeight7, w89[1]), mul(kWeight8, w89[0])))); \ + store(output + 20 * stride, \ + sub(sub(kWeight0, w34), mul(kWeight2, add(w81, w62)))); \ + store(output + 21 * stride, \ + add(w44[1], sub(mul(kWeight8, w91[1]), mul(kWeight7, w91[0])))); \ + store(output + 22 * stride, \ + add(w45[1], sub(mul(kWeight4, w92[1]), mul(kWeight3, w92[0])))); \ + store(output + 23 * stride, \ + add(w46[1], sub(mul(kWeight6, w93[1]), mul(kWeight5, w93[0])))); \ + store(output + 24 * stride, sub(kWeight0, w86)); \ + store(output + 25 * stride, \ + sub(sub(kWeight0, w46[1]), \ + sub(mul(kWeight5, w93[0]), mul(kWeight6, w93[1])))); \ + store(output + 26 * stride, \ + sub(sub(kWeight0, w45[1]), \ + sub(mul(kWeight3, w92[0]), mul(kWeight4, w92[1])))); \ + store(output + 27 * stride, \ + sub(sub(kWeight0, w44[1]), \ + sub(mul(kWeight7, w91[0]), mul(kWeight8, w91[1])))); \ + store(output + 28 * stride, sub(w34, mul(kWeight2, add(w81, w62)))); \ + store(output + 29 * stride, \ + sub(sub(kWeight0, w42[1]), \ + sub(mul(kWeight8, w89[0]), mul(kWeight7, w89[1])))); \ + store(output + 30 * stride, \ + sub(sub(kWeight0, w41[1]), \ + sub(mul(kWeight4, w88[0]), mul(kWeight3, w88[1])))); \ + store(output + 31 * stride, \ + sub(sub(kWeight0, w40[1]), \ + sub(mul(kWeight6, w87[0]), mul(kWeight5, w87[1])))); \ + } + +#define GEN_IFFT_2(ret, suffix, T, T_VEC, load, store) \ + ret aom_ifft1d_2_##suffix(const T *input, T *output, int stride) { \ + const T_VEC i0 = load(input + 0 * stride); \ + const T_VEC i1 = load(input + 1 * stride); \ + store(output + 0 * stride, i0 + i1); \ + store(output + 1 * stride, i0 - i1); \ + } + +#define GEN_IFFT_4(ret, suffix, T, T_VEC, load, store, constant, add, sub) \ + ret aom_ifft1d_4_##suffix(const T *input, T *output, int stride) { \ + const T_VEC kWeight0 = constant(0.0f); \ + const T_VEC i0 = load(input + 0 * stride); \ + const T_VEC i1 = load(input + 1 * stride); \ + const T_VEC i2 = load(input + 2 * stride); \ + const T_VEC i3 = load(input + 3 * stride); \ + const T_VEC w2 = add(i0, i2); \ + const T_VEC w3 = sub(i0, i2); \ + const T_VEC w4[2] = { add(i1, i1), sub(i3, i3) }; \ + const T_VEC w5[2] = { sub(i1, i1), sub(sub(kWeight0, i3), i3) }; \ + store(output + 0 * stride, add(w2, w4[0])); \ + store(output + 1 * stride, add(w3, w5[1])); \ + store(output + 2 * stride, sub(w2, w4[0])); \ + store(output + 3 * stride, sub(w3, w5[1])); \ + } + +#define GEN_IFFT_8(ret, suffix, T, T_VEC, load, store, constant, add, sub, \ + mul) \ + ret aom_ifft1d_8_##suffix(const T *input, T *output, int stride) { \ + const T_VEC kWeight0 = constant(0.0f); \ + const T_VEC kWeight2 = constant(0.707107f); \ + const T_VEC i0 = load(input + 0 * stride); \ + const T_VEC i1 = load(input + 1 * stride); \ + const T_VEC i2 = load(input + 2 * stride); \ + const T_VEC i3 = load(input + 3 * stride); \ + const T_VEC i4 = load(input + 4 * stride); \ + const T_VEC i5 = load(input + 5 * stride); \ + const T_VEC i6 = load(input + 6 * stride); \ + const T_VEC i7 = load(input + 7 * stride); \ + const T_VEC w6 = add(i0, i4); \ + const T_VEC w7 = sub(i0, i4); \ + const T_VEC w8[2] = { add(i2, i2), sub(i6, i6) }; \ + const T_VEC w9[2] = { sub(i2, i2), sub(sub(kWeight0, i6), i6) }; \ + const T_VEC w10[2] = { add(w6, w8[0]), w8[1] }; \ + const T_VEC w11[2] = { sub(w6, w8[0]), sub(kWeight0, w8[1]) }; \ + const T_VEC w12[2] = { add(w7, w9[1]), sub(kWeight0, w9[0]) }; \ + const T_VEC w13[2] = { sub(w7, w9[1]), w9[0] }; \ + const T_VEC w14[2] = { add(i1, i3), sub(i7, i5) }; \ + const T_VEC w15[2] = { sub(i1, i3), sub(sub(kWeight0, i5), i7) }; \ + const T_VEC w16[2] = { add(i3, i1), sub(i5, i7) }; \ + const T_VEC w17[2] = { sub(i3, i1), sub(sub(kWeight0, i7), i5) }; \ + const T_VEC w18[2] = { add(w14[0], w16[0]), add(w14[1], w16[1]) }; \ + const T_VEC w19[2] = { sub(w14[0], w16[0]), sub(w14[1], w16[1]) }; \ + const T_VEC w20[2] = { add(w15[0], w17[1]), sub(w15[1], w17[0]) }; \ + const T_VEC w21[2] = { sub(w15[0], w17[1]), add(w15[1], w17[0]) }; \ + store(output + 0 * stride, add(w10[0], w18[0])); \ + store(output + 1 * stride, \ + add(w12[0], mul(kWeight2, add(w20[0], w20[1])))); \ + store(output + 2 * stride, add(w11[0], w19[1])); \ + store(output + 3 * stride, \ + sub(w13[0], mul(kWeight2, sub(w21[0], w21[1])))); \ + store(output + 4 * stride, sub(w10[0], w18[0])); \ + store(output + 5 * stride, \ + add(w12[0], sub(sub(kWeight0, mul(kWeight2, w20[0])), \ + mul(kWeight2, w20[1])))); \ + store(output + 6 * stride, sub(w11[0], w19[1])); \ + store(output + 7 * stride, \ + add(w13[0], mul(kWeight2, sub(w21[0], w21[1])))); \ + } + +#define GEN_IFFT_16(ret, suffix, T, T_VEC, load, store, constant, add, sub, \ + mul) \ + ret aom_ifft1d_16_##suffix(const T *input, T *output, int stride) { \ + const T_VEC kWeight0 = constant(0.0f); \ + const T_VEC kWeight2 = constant(0.707107f); \ + const T_VEC kWeight3 = constant(0.92388f); \ + const T_VEC kWeight4 = constant(0.382683f); \ + const T_VEC i0 = load(input + 0 * stride); \ + const T_VEC i1 = load(input + 1 * stride); \ + const T_VEC i2 = load(input + 2 * stride); \ + const T_VEC i3 = load(input + 3 * stride); \ + const T_VEC i4 = load(input + 4 * stride); \ + const T_VEC i5 = load(input + 5 * stride); \ + const T_VEC i6 = load(input + 6 * stride); \ + const T_VEC i7 = load(input + 7 * stride); \ + const T_VEC i8 = load(input + 8 * stride); \ + const T_VEC i9 = load(input + 9 * stride); \ + const T_VEC i10 = load(input + 10 * stride); \ + const T_VEC i11 = load(input + 11 * stride); \ + const T_VEC i12 = load(input + 12 * stride); \ + const T_VEC i13 = load(input + 13 * stride); \ + const T_VEC i14 = load(input + 14 * stride); \ + const T_VEC i15 = load(input + 15 * stride); \ + const T_VEC w14 = add(i0, i8); \ + const T_VEC w15 = sub(i0, i8); \ + const T_VEC w16[2] = { add(i4, i4), sub(i12, i12) }; \ + const T_VEC w17[2] = { sub(i4, i4), sub(sub(kWeight0, i12), i12) }; \ + const T_VEC w18[2] = { add(w14, w16[0]), w16[1] }; \ + const T_VEC w19[2] = { sub(w14, w16[0]), sub(kWeight0, w16[1]) }; \ + const T_VEC w20[2] = { add(w15, w17[1]), sub(kWeight0, w17[0]) }; \ + const T_VEC w21[2] = { sub(w15, w17[1]), w17[0] }; \ + const T_VEC w22[2] = { add(i2, i6), sub(i14, i10) }; \ + const T_VEC w23[2] = { sub(i2, i6), sub(sub(kWeight0, i10), i14) }; \ + const T_VEC w24[2] = { add(i6, i2), sub(i10, i14) }; \ + const T_VEC w25[2] = { sub(i6, i2), sub(sub(kWeight0, i14), i10) }; \ + const T_VEC w26[2] = { add(w22[0], w24[0]), add(w22[1], w24[1]) }; \ + const T_VEC w27[2] = { sub(w22[0], w24[0]), sub(w22[1], w24[1]) }; \ + const T_VEC w28[2] = { add(w23[0], w25[1]), sub(w23[1], w25[0]) }; \ + const T_VEC w29[2] = { sub(w23[0], w25[1]), add(w23[1], w25[0]) }; \ + const T_VEC w30[2] = { add(w18[0], w26[0]), add(w18[1], w26[1]) }; \ + const T_VEC w31[2] = { sub(w18[0], w26[0]), sub(w18[1], w26[1]) }; \ + const T_VEC w32[2] = { add(w20[0], mul(kWeight2, add(w28[0], w28[1]))), \ + add(w20[1], mul(kWeight2, sub(w28[1], w28[0]))) }; \ + const T_VEC w33[2] = { add(w20[0], \ + sub(sub(kWeight0, mul(kWeight2, w28[0])), \ + mul(kWeight2, w28[1]))), \ + add(w20[1], mul(kWeight2, sub(w28[0], w28[1]))) }; \ + const T_VEC w34[2] = { add(w19[0], w27[1]), sub(w19[1], w27[0]) }; \ + const T_VEC w35[2] = { sub(w19[0], w27[1]), add(w19[1], w27[0]) }; \ + const T_VEC w36[2] = { sub(w21[0], mul(kWeight2, sub(w29[0], w29[1]))), \ + sub(w21[1], mul(kWeight2, add(w29[1], w29[0]))) }; \ + const T_VEC w37[2] = { add(w21[0], mul(kWeight2, sub(w29[0], w29[1]))), \ + add(w21[1], mul(kWeight2, add(w29[1], w29[0]))) }; \ + const T_VEC w38[2] = { add(i1, i7), sub(i15, i9) }; \ + const T_VEC w39[2] = { sub(i1, i7), sub(sub(kWeight0, i9), i15) }; \ + const T_VEC w40[2] = { add(i5, i3), sub(i11, i13) }; \ + const T_VEC w41[2] = { sub(i5, i3), sub(sub(kWeight0, i13), i11) }; \ + const T_VEC w42[2] = { add(w38[0], w40[0]), add(w38[1], w40[1]) }; \ + const T_VEC w43[2] = { sub(w38[0], w40[0]), sub(w38[1], w40[1]) }; \ + const T_VEC w44[2] = { add(w39[0], w41[1]), sub(w39[1], w41[0]) }; \ + const T_VEC w45[2] = { sub(w39[0], w41[1]), add(w39[1], w41[0]) }; \ + const T_VEC w46[2] = { add(i3, i5), sub(i13, i11) }; \ + const T_VEC w47[2] = { sub(i3, i5), sub(sub(kWeight0, i11), i13) }; \ + const T_VEC w48[2] = { add(i7, i1), sub(i9, i15) }; \ + const T_VEC w49[2] = { sub(i7, i1), sub(sub(kWeight0, i15), i9) }; \ + const T_VEC w50[2] = { add(w46[0], w48[0]), add(w46[1], w48[1]) }; \ + const T_VEC w51[2] = { sub(w46[0], w48[0]), sub(w46[1], w48[1]) }; \ + const T_VEC w52[2] = { add(w47[0], w49[1]), sub(w47[1], w49[0]) }; \ + const T_VEC w53[2] = { sub(w47[0], w49[1]), add(w47[1], w49[0]) }; \ + const T_VEC w54[2] = { add(w42[0], w50[0]), add(w42[1], w50[1]) }; \ + const T_VEC w55[2] = { sub(w42[0], w50[0]), sub(w42[1], w50[1]) }; \ + const T_VEC w56[2] = { add(w44[0], mul(kWeight2, add(w52[0], w52[1]))), \ + add(w44[1], mul(kWeight2, sub(w52[1], w52[0]))) }; \ + const T_VEC w57[2] = { add(w44[0], \ + sub(sub(kWeight0, mul(kWeight2, w52[0])), \ + mul(kWeight2, w52[1]))), \ + add(w44[1], mul(kWeight2, sub(w52[0], w52[1]))) }; \ + const T_VEC w58[2] = { add(w43[0], w51[1]), sub(w43[1], w51[0]) }; \ + const T_VEC w59[2] = { sub(w43[0], w51[1]), add(w43[1], w51[0]) }; \ + const T_VEC w60[2] = { sub(w45[0], mul(kWeight2, sub(w53[0], w53[1]))), \ + sub(w45[1], mul(kWeight2, add(w53[1], w53[0]))) }; \ + const T_VEC w61[2] = { add(w45[0], mul(kWeight2, sub(w53[0], w53[1]))), \ + add(w45[1], mul(kWeight2, add(w53[1], w53[0]))) }; \ + store(output + 0 * stride, add(w30[0], w54[0])); \ + store(output + 1 * stride, \ + add(w32[0], add(mul(kWeight3, w56[0]), mul(kWeight4, w56[1])))); \ + store(output + 2 * stride, \ + add(w34[0], mul(kWeight2, add(w58[0], w58[1])))); \ + store(output + 3 * stride, \ + add(w36[0], add(mul(kWeight4, w60[0]), mul(kWeight3, w60[1])))); \ + store(output + 4 * stride, add(w31[0], w55[1])); \ + store(output + 5 * stride, \ + sub(w33[0], sub(mul(kWeight4, w57[0]), mul(kWeight3, w57[1])))); \ + store(output + 6 * stride, \ + sub(w35[0], mul(kWeight2, sub(w59[0], w59[1])))); \ + store(output + 7 * stride, \ + sub(w37[0], sub(mul(kWeight3, w61[0]), mul(kWeight4, w61[1])))); \ + store(output + 8 * stride, sub(w30[0], w54[0])); \ + store(output + 9 * stride, \ + add(w32[0], sub(sub(kWeight0, mul(kWeight3, w56[0])), \ + mul(kWeight4, w56[1])))); \ + store(output + 10 * stride, \ + add(w34[0], sub(sub(kWeight0, mul(kWeight2, w58[0])), \ + mul(kWeight2, w58[1])))); \ + store(output + 11 * stride, \ + add(w36[0], sub(sub(kWeight0, mul(kWeight4, w60[0])), \ + mul(kWeight3, w60[1])))); \ + store(output + 12 * stride, sub(w31[0], w55[1])); \ + store(output + 13 * stride, \ + add(w33[0], sub(mul(kWeight4, w57[0]), mul(kWeight3, w57[1])))); \ + store(output + 14 * stride, \ + add(w35[0], mul(kWeight2, sub(w59[0], w59[1])))); \ + store(output + 15 * stride, \ + add(w37[0], sub(mul(kWeight3, w61[0]), mul(kWeight4, w61[1])))); \ + } +#define GEN_IFFT_32(ret, suffix, T, T_VEC, load, store, constant, add, sub, \ + mul) \ + ret aom_ifft1d_32_##suffix(const T *input, T *output, int stride) { \ + const T_VEC kWeight0 = constant(0.0f); \ + const T_VEC kWeight2 = constant(0.707107f); \ + const T_VEC kWeight3 = constant(0.92388f); \ + const T_VEC kWeight4 = constant(0.382683f); \ + const T_VEC kWeight5 = constant(0.980785f); \ + const T_VEC kWeight6 = constant(0.19509f); \ + const T_VEC kWeight7 = constant(0.83147f); \ + const T_VEC kWeight8 = constant(0.55557f); \ + const T_VEC i0 = load(input + 0 * stride); \ + const T_VEC i1 = load(input + 1 * stride); \ + const T_VEC i2 = load(input + 2 * stride); \ + const T_VEC i3 = load(input + 3 * stride); \ + const T_VEC i4 = load(input + 4 * stride); \ + const T_VEC i5 = load(input + 5 * stride); \ + const T_VEC i6 = load(input + 6 * stride); \ + const T_VEC i7 = load(input + 7 * stride); \ + const T_VEC i8 = load(input + 8 * stride); \ + const T_VEC i9 = load(input + 9 * stride); \ + const T_VEC i10 = load(input + 10 * stride); \ + const T_VEC i11 = load(input + 11 * stride); \ + const T_VEC i12 = load(input + 12 * stride); \ + const T_VEC i13 = load(input + 13 * stride); \ + const T_VEC i14 = load(input + 14 * stride); \ + const T_VEC i15 = load(input + 15 * stride); \ + const T_VEC i16 = load(input + 16 * stride); \ + const T_VEC i17 = load(input + 17 * stride); \ + const T_VEC i18 = load(input + 18 * stride); \ + const T_VEC i19 = load(input + 19 * stride); \ + const T_VEC i20 = load(input + 20 * stride); \ + const T_VEC i21 = load(input + 21 * stride); \ + const T_VEC i22 = load(input + 22 * stride); \ + const T_VEC i23 = load(input + 23 * stride); \ + const T_VEC i24 = load(input + 24 * stride); \ + const T_VEC i25 = load(input + 25 * stride); \ + const T_VEC i26 = load(input + 26 * stride); \ + const T_VEC i27 = load(input + 27 * stride); \ + const T_VEC i28 = load(input + 28 * stride); \ + const T_VEC i29 = load(input + 29 * stride); \ + const T_VEC i30 = load(input + 30 * stride); \ + const T_VEC i31 = load(input + 31 * stride); \ + const T_VEC w30 = add(i0, i16); \ + const T_VEC w31 = sub(i0, i16); \ + const T_VEC w32[2] = { add(i8, i8), sub(i24, i24) }; \ + const T_VEC w33[2] = { sub(i8, i8), sub(sub(kWeight0, i24), i24) }; \ + const T_VEC w34[2] = { add(w30, w32[0]), w32[1] }; \ + const T_VEC w35[2] = { sub(w30, w32[0]), sub(kWeight0, w32[1]) }; \ + const T_VEC w36[2] = { add(w31, w33[1]), sub(kWeight0, w33[0]) }; \ + const T_VEC w37[2] = { sub(w31, w33[1]), w33[0] }; \ + const T_VEC w38[2] = { add(i4, i12), sub(i28, i20) }; \ + const T_VEC w39[2] = { sub(i4, i12), sub(sub(kWeight0, i20), i28) }; \ + const T_VEC w40[2] = { add(i12, i4), sub(i20, i28) }; \ + const T_VEC w41[2] = { sub(i12, i4), sub(sub(kWeight0, i28), i20) }; \ + const T_VEC w42[2] = { add(w38[0], w40[0]), add(w38[1], w40[1]) }; \ + const T_VEC w43[2] = { sub(w38[0], w40[0]), sub(w38[1], w40[1]) }; \ + const T_VEC w44[2] = { add(w39[0], w41[1]), sub(w39[1], w41[0]) }; \ + const T_VEC w45[2] = { sub(w39[0], w41[1]), add(w39[1], w41[0]) }; \ + const T_VEC w46[2] = { add(w34[0], w42[0]), add(w34[1], w42[1]) }; \ + const T_VEC w47[2] = { sub(w34[0], w42[0]), sub(w34[1], w42[1]) }; \ + const T_VEC w48[2] = { add(w36[0], mul(kWeight2, add(w44[0], w44[1]))), \ + add(w36[1], mul(kWeight2, sub(w44[1], w44[0]))) }; \ + const T_VEC w49[2] = { add(w36[0], \ + sub(sub(kWeight0, mul(kWeight2, w44[0])), \ + mul(kWeight2, w44[1]))), \ + add(w36[1], mul(kWeight2, sub(w44[0], w44[1]))) }; \ + const T_VEC w50[2] = { add(w35[0], w43[1]), sub(w35[1], w43[0]) }; \ + const T_VEC w51[2] = { sub(w35[0], w43[1]), add(w35[1], w43[0]) }; \ + const T_VEC w52[2] = { sub(w37[0], mul(kWeight2, sub(w45[0], w45[1]))), \ + sub(w37[1], mul(kWeight2, add(w45[1], w45[0]))) }; \ + const T_VEC w53[2] = { add(w37[0], mul(kWeight2, sub(w45[0], w45[1]))), \ + add(w37[1], mul(kWeight2, add(w45[1], w45[0]))) }; \ + const T_VEC w54[2] = { add(i2, i14), sub(i30, i18) }; \ + const T_VEC w55[2] = { sub(i2, i14), sub(sub(kWeight0, i18), i30) }; \ + const T_VEC w56[2] = { add(i10, i6), sub(i22, i26) }; \ + const T_VEC w57[2] = { sub(i10, i6), sub(sub(kWeight0, i26), i22) }; \ + const T_VEC w58[2] = { add(w54[0], w56[0]), add(w54[1], w56[1]) }; \ + const T_VEC w59[2] = { sub(w54[0], w56[0]), sub(w54[1], w56[1]) }; \ + const T_VEC w60[2] = { add(w55[0], w57[1]), sub(w55[1], w57[0]) }; \ + const T_VEC w61[2] = { sub(w55[0], w57[1]), add(w55[1], w57[0]) }; \ + const T_VEC w62[2] = { add(i6, i10), sub(i26, i22) }; \ + const T_VEC w63[2] = { sub(i6, i10), sub(sub(kWeight0, i22), i26) }; \ + const T_VEC w64[2] = { add(i14, i2), sub(i18, i30) }; \ + const T_VEC w65[2] = { sub(i14, i2), sub(sub(kWeight0, i30), i18) }; \ + const T_VEC w66[2] = { add(w62[0], w64[0]), add(w62[1], w64[1]) }; \ + const T_VEC w67[2] = { sub(w62[0], w64[0]), sub(w62[1], w64[1]) }; \ + const T_VEC w68[2] = { add(w63[0], w65[1]), sub(w63[1], w65[0]) }; \ + const T_VEC w69[2] = { sub(w63[0], w65[1]), add(w63[1], w65[0]) }; \ + const T_VEC w70[2] = { add(w58[0], w66[0]), add(w58[1], w66[1]) }; \ + const T_VEC w71[2] = { sub(w58[0], w66[0]), sub(w58[1], w66[1]) }; \ + const T_VEC w72[2] = { add(w60[0], mul(kWeight2, add(w68[0], w68[1]))), \ + add(w60[1], mul(kWeight2, sub(w68[1], w68[0]))) }; \ + const T_VEC w73[2] = { add(w60[0], \ + sub(sub(kWeight0, mul(kWeight2, w68[0])), \ + mul(kWeight2, w68[1]))), \ + add(w60[1], mul(kWeight2, sub(w68[0], w68[1]))) }; \ + const T_VEC w74[2] = { add(w59[0], w67[1]), sub(w59[1], w67[0]) }; \ + const T_VEC w75[2] = { sub(w59[0], w67[1]), add(w59[1], w67[0]) }; \ + const T_VEC w76[2] = { sub(w61[0], mul(kWeight2, sub(w69[0], w69[1]))), \ + sub(w61[1], mul(kWeight2, add(w69[1], w69[0]))) }; \ + const T_VEC w77[2] = { add(w61[0], mul(kWeight2, sub(w69[0], w69[1]))), \ + add(w61[1], mul(kWeight2, add(w69[1], w69[0]))) }; \ + const T_VEC w78[2] = { add(w46[0], w70[0]), add(w46[1], w70[1]) }; \ + const T_VEC w79[2] = { sub(w46[0], w70[0]), sub(w46[1], w70[1]) }; \ + const T_VEC w80[2] = { \ + add(w48[0], add(mul(kWeight3, w72[0]), mul(kWeight4, w72[1]))), \ + add(w48[1], sub(mul(kWeight3, w72[1]), mul(kWeight4, w72[0]))) \ + }; \ + const T_VEC w81[2] = { \ + add(w48[0], \ + sub(sub(kWeight0, mul(kWeight3, w72[0])), mul(kWeight4, w72[1]))), \ + add(w48[1], sub(mul(kWeight4, w72[0]), mul(kWeight3, w72[1]))) \ + }; \ + const T_VEC w82[2] = { add(w50[0], mul(kWeight2, add(w74[0], w74[1]))), \ + add(w50[1], mul(kWeight2, sub(w74[1], w74[0]))) }; \ + const T_VEC w83[2] = { add(w50[0], \ + sub(sub(kWeight0, mul(kWeight2, w74[0])), \ + mul(kWeight2, w74[1]))), \ + add(w50[1], mul(kWeight2, sub(w74[0], w74[1]))) }; \ + const T_VEC w84[2] = { \ + add(w52[0], add(mul(kWeight4, w76[0]), mul(kWeight3, w76[1]))), \ + add(w52[1], sub(mul(kWeight4, w76[1]), mul(kWeight3, w76[0]))) \ + }; \ + const T_VEC w85[2] = { \ + add(w52[0], \ + sub(sub(kWeight0, mul(kWeight4, w76[0])), mul(kWeight3, w76[1]))), \ + add(w52[1], sub(mul(kWeight3, w76[0]), mul(kWeight4, w76[1]))) \ + }; \ + const T_VEC w86[2] = { add(w47[0], w71[1]), sub(w47[1], w71[0]) }; \ + const T_VEC w87[2] = { sub(w47[0], w71[1]), add(w47[1], w71[0]) }; \ + const T_VEC w88[2] = { \ + sub(w49[0], sub(mul(kWeight4, w73[0]), mul(kWeight3, w73[1]))), \ + add(w49[1], \ + sub(sub(kWeight0, mul(kWeight4, w73[1])), mul(kWeight3, w73[0]))) \ + }; \ + const T_VEC w89[2] = { \ + add(w49[0], sub(mul(kWeight4, w73[0]), mul(kWeight3, w73[1]))), \ + add(w49[1], add(mul(kWeight4, w73[1]), mul(kWeight3, w73[0]))) \ + }; \ + const T_VEC w90[2] = { sub(w51[0], mul(kWeight2, sub(w75[0], w75[1]))), \ + sub(w51[1], mul(kWeight2, add(w75[1], w75[0]))) }; \ + const T_VEC w91[2] = { add(w51[0], mul(kWeight2, sub(w75[0], w75[1]))), \ + add(w51[1], mul(kWeight2, add(w75[1], w75[0]))) }; \ + const T_VEC w92[2] = { \ + sub(w53[0], sub(mul(kWeight3, w77[0]), mul(kWeight4, w77[1]))), \ + add(w53[1], \ + sub(sub(kWeight0, mul(kWeight3, w77[1])), mul(kWeight4, w77[0]))) \ + }; \ + const T_VEC w93[2] = { \ + add(w53[0], sub(mul(kWeight3, w77[0]), mul(kWeight4, w77[1]))), \ + add(w53[1], add(mul(kWeight3, w77[1]), mul(kWeight4, w77[0]))) \ + }; \ + const T_VEC w94[2] = { add(i1, i15), sub(i31, i17) }; \ + const T_VEC w95[2] = { sub(i1, i15), sub(sub(kWeight0, i17), i31) }; \ + const T_VEC w96[2] = { add(i9, i7), sub(i23, i25) }; \ + const T_VEC w97[2] = { sub(i9, i7), sub(sub(kWeight0, i25), i23) }; \ + const T_VEC w98[2] = { add(w94[0], w96[0]), add(w94[1], w96[1]) }; \ + const T_VEC w99[2] = { sub(w94[0], w96[0]), sub(w94[1], w96[1]) }; \ + const T_VEC w100[2] = { add(w95[0], w97[1]), sub(w95[1], w97[0]) }; \ + const T_VEC w101[2] = { sub(w95[0], w97[1]), add(w95[1], w97[0]) }; \ + const T_VEC w102[2] = { add(i5, i11), sub(i27, i21) }; \ + const T_VEC w103[2] = { sub(i5, i11), sub(sub(kWeight0, i21), i27) }; \ + const T_VEC w104[2] = { add(i13, i3), sub(i19, i29) }; \ + const T_VEC w105[2] = { sub(i13, i3), sub(sub(kWeight0, i29), i19) }; \ + const T_VEC w106[2] = { add(w102[0], w104[0]), add(w102[1], w104[1]) }; \ + const T_VEC w107[2] = { sub(w102[0], w104[0]), sub(w102[1], w104[1]) }; \ + const T_VEC w108[2] = { add(w103[0], w105[1]), sub(w103[1], w105[0]) }; \ + const T_VEC w109[2] = { sub(w103[0], w105[1]), add(w103[1], w105[0]) }; \ + const T_VEC w110[2] = { add(w98[0], w106[0]), add(w98[1], w106[1]) }; \ + const T_VEC w111[2] = { sub(w98[0], w106[0]), sub(w98[1], w106[1]) }; \ + const T_VEC w112[2] = { \ + add(w100[0], mul(kWeight2, add(w108[0], w108[1]))), \ + add(w100[1], mul(kWeight2, sub(w108[1], w108[0]))) \ + }; \ + const T_VEC w113[2] = { \ + add(w100[0], \ + sub(sub(kWeight0, mul(kWeight2, w108[0])), mul(kWeight2, w108[1]))), \ + add(w100[1], mul(kWeight2, sub(w108[0], w108[1]))) \ + }; \ + const T_VEC w114[2] = { add(w99[0], w107[1]), sub(w99[1], w107[0]) }; \ + const T_VEC w115[2] = { sub(w99[0], w107[1]), add(w99[1], w107[0]) }; \ + const T_VEC w116[2] = { \ + sub(w101[0], mul(kWeight2, sub(w109[0], w109[1]))), \ + sub(w101[1], mul(kWeight2, add(w109[1], w109[0]))) \ + }; \ + const T_VEC w117[2] = { \ + add(w101[0], mul(kWeight2, sub(w109[0], w109[1]))), \ + add(w101[1], mul(kWeight2, add(w109[1], w109[0]))) \ + }; \ + const T_VEC w118[2] = { add(i3, i13), sub(i29, i19) }; \ + const T_VEC w119[2] = { sub(i3, i13), sub(sub(kWeight0, i19), i29) }; \ + const T_VEC w120[2] = { add(i11, i5), sub(i21, i27) }; \ + const T_VEC w121[2] = { sub(i11, i5), sub(sub(kWeight0, i27), i21) }; \ + const T_VEC w122[2] = { add(w118[0], w120[0]), add(w118[1], w120[1]) }; \ + const T_VEC w123[2] = { sub(w118[0], w120[0]), sub(w118[1], w120[1]) }; \ + const T_VEC w124[2] = { add(w119[0], w121[1]), sub(w119[1], w121[0]) }; \ + const T_VEC w125[2] = { sub(w119[0], w121[1]), add(w119[1], w121[0]) }; \ + const T_VEC w126[2] = { add(i7, i9), sub(i25, i23) }; \ + const T_VEC w127[2] = { sub(i7, i9), sub(sub(kWeight0, i23), i25) }; \ + const T_VEC w128[2] = { add(i15, i1), sub(i17, i31) }; \ + const T_VEC w129[2] = { sub(i15, i1), sub(sub(kWeight0, i31), i17) }; \ + const T_VEC w130[2] = { add(w126[0], w128[0]), add(w126[1], w128[1]) }; \ + const T_VEC w131[2] = { sub(w126[0], w128[0]), sub(w126[1], w128[1]) }; \ + const T_VEC w132[2] = { add(w127[0], w129[1]), sub(w127[1], w129[0]) }; \ + const T_VEC w133[2] = { sub(w127[0], w129[1]), add(w127[1], w129[0]) }; \ + const T_VEC w134[2] = { add(w122[0], w130[0]), add(w122[1], w130[1]) }; \ + const T_VEC w135[2] = { sub(w122[0], w130[0]), sub(w122[1], w130[1]) }; \ + const T_VEC w136[2] = { \ + add(w124[0], mul(kWeight2, add(w132[0], w132[1]))), \ + add(w124[1], mul(kWeight2, sub(w132[1], w132[0]))) \ + }; \ + const T_VEC w137[2] = { \ + add(w124[0], \ + sub(sub(kWeight0, mul(kWeight2, w132[0])), mul(kWeight2, w132[1]))), \ + add(w124[1], mul(kWeight2, sub(w132[0], w132[1]))) \ + }; \ + const T_VEC w138[2] = { add(w123[0], w131[1]), sub(w123[1], w131[0]) }; \ + const T_VEC w139[2] = { sub(w123[0], w131[1]), add(w123[1], w131[0]) }; \ + const T_VEC w140[2] = { \ + sub(w125[0], mul(kWeight2, sub(w133[0], w133[1]))), \ + sub(w125[1], mul(kWeight2, add(w133[1], w133[0]))) \ + }; \ + const T_VEC w141[2] = { \ + add(w125[0], mul(kWeight2, sub(w133[0], w133[1]))), \ + add(w125[1], mul(kWeight2, add(w133[1], w133[0]))) \ + }; \ + const T_VEC w142[2] = { add(w110[0], w134[0]), add(w110[1], w134[1]) }; \ + const T_VEC w143[2] = { sub(w110[0], w134[0]), sub(w110[1], w134[1]) }; \ + const T_VEC w144[2] = { \ + add(w112[0], add(mul(kWeight3, w136[0]), mul(kWeight4, w136[1]))), \ + add(w112[1], sub(mul(kWeight3, w136[1]), mul(kWeight4, w136[0]))) \ + }; \ + const T_VEC w145[2] = { \ + add(w112[0], \ + sub(sub(kWeight0, mul(kWeight3, w136[0])), mul(kWeight4, w136[1]))), \ + add(w112[1], sub(mul(kWeight4, w136[0]), mul(kWeight3, w136[1]))) \ + }; \ + const T_VEC w146[2] = { \ + add(w114[0], mul(kWeight2, add(w138[0], w138[1]))), \ + add(w114[1], mul(kWeight2, sub(w138[1], w138[0]))) \ + }; \ + const T_VEC w147[2] = { \ + add(w114[0], \ + sub(sub(kWeight0, mul(kWeight2, w138[0])), mul(kWeight2, w138[1]))), \ + add(w114[1], mul(kWeight2, sub(w138[0], w138[1]))) \ + }; \ + const T_VEC w148[2] = { \ + add(w116[0], add(mul(kWeight4, w140[0]), mul(kWeight3, w140[1]))), \ + add(w116[1], sub(mul(kWeight4, w140[1]), mul(kWeight3, w140[0]))) \ + }; \ + const T_VEC w149[2] = { \ + add(w116[0], \ + sub(sub(kWeight0, mul(kWeight4, w140[0])), mul(kWeight3, w140[1]))), \ + add(w116[1], sub(mul(kWeight3, w140[0]), mul(kWeight4, w140[1]))) \ + }; \ + const T_VEC w150[2] = { add(w111[0], w135[1]), sub(w111[1], w135[0]) }; \ + const T_VEC w151[2] = { sub(w111[0], w135[1]), add(w111[1], w135[0]) }; \ + const T_VEC w152[2] = { \ + sub(w113[0], sub(mul(kWeight4, w137[0]), mul(kWeight3, w137[1]))), \ + add(w113[1], \ + sub(sub(kWeight0, mul(kWeight4, w137[1])), mul(kWeight3, w137[0]))) \ + }; \ + const T_VEC w153[2] = { \ + add(w113[0], sub(mul(kWeight4, w137[0]), mul(kWeight3, w137[1]))), \ + add(w113[1], add(mul(kWeight4, w137[1]), mul(kWeight3, w137[0]))) \ + }; \ + const T_VEC w154[2] = { \ + sub(w115[0], mul(kWeight2, sub(w139[0], w139[1]))), \ + sub(w115[1], mul(kWeight2, add(w139[1], w139[0]))) \ + }; \ + const T_VEC w155[2] = { \ + add(w115[0], mul(kWeight2, sub(w139[0], w139[1]))), \ + add(w115[1], mul(kWeight2, add(w139[1], w139[0]))) \ + }; \ + const T_VEC w156[2] = { \ + sub(w117[0], sub(mul(kWeight3, w141[0]), mul(kWeight4, w141[1]))), \ + add(w117[1], \ + sub(sub(kWeight0, mul(kWeight3, w141[1])), mul(kWeight4, w141[0]))) \ + }; \ + const T_VEC w157[2] = { \ + add(w117[0], sub(mul(kWeight3, w141[0]), mul(kWeight4, w141[1]))), \ + add(w117[1], add(mul(kWeight3, w141[1]), mul(kWeight4, w141[0]))) \ + }; \ + store(output + 0 * stride, add(w78[0], w142[0])); \ + store(output + 1 * stride, \ + add(w80[0], add(mul(kWeight5, w144[0]), mul(kWeight6, w144[1])))); \ + store(output + 2 * stride, \ + add(w82[0], add(mul(kWeight3, w146[0]), mul(kWeight4, w146[1])))); \ + store(output + 3 * stride, \ + add(w84[0], add(mul(kWeight7, w148[0]), mul(kWeight8, w148[1])))); \ + store(output + 4 * stride, \ + add(w86[0], mul(kWeight2, add(w150[0], w150[1])))); \ + store(output + 5 * stride, \ + add(w88[0], add(mul(kWeight8, w152[0]), mul(kWeight7, w152[1])))); \ + store(output + 6 * stride, \ + add(w90[0], add(mul(kWeight4, w154[0]), mul(kWeight3, w154[1])))); \ + store(output + 7 * stride, \ + add(w92[0], add(mul(kWeight6, w156[0]), mul(kWeight5, w156[1])))); \ + store(output + 8 * stride, add(w79[0], w143[1])); \ + store(output + 9 * stride, \ + sub(w81[0], sub(mul(kWeight6, w145[0]), mul(kWeight5, w145[1])))); \ + store(output + 10 * stride, \ + sub(w83[0], sub(mul(kWeight4, w147[0]), mul(kWeight3, w147[1])))); \ + store(output + 11 * stride, \ + sub(w85[0], sub(mul(kWeight8, w149[0]), mul(kWeight7, w149[1])))); \ + store(output + 12 * stride, \ + sub(w87[0], mul(kWeight2, sub(w151[0], w151[1])))); \ + store(output + 13 * stride, \ + sub(w89[0], sub(mul(kWeight7, w153[0]), mul(kWeight8, w153[1])))); \ + store(output + 14 * stride, \ + sub(w91[0], sub(mul(kWeight3, w155[0]), mul(kWeight4, w155[1])))); \ + store(output + 15 * stride, \ + sub(w93[0], sub(mul(kWeight5, w157[0]), mul(kWeight6, w157[1])))); \ + store(output + 16 * stride, sub(w78[0], w142[0])); \ + store(output + 17 * stride, \ + add(w80[0], sub(sub(kWeight0, mul(kWeight5, w144[0])), \ + mul(kWeight6, w144[1])))); \ + store(output + 18 * stride, \ + add(w82[0], sub(sub(kWeight0, mul(kWeight3, w146[0])), \ + mul(kWeight4, w146[1])))); \ + store(output + 19 * stride, \ + add(w84[0], sub(sub(kWeight0, mul(kWeight7, w148[0])), \ + mul(kWeight8, w148[1])))); \ + store(output + 20 * stride, \ + add(w86[0], sub(sub(kWeight0, mul(kWeight2, w150[0])), \ + mul(kWeight2, w150[1])))); \ + store(output + 21 * stride, \ + add(w88[0], sub(sub(kWeight0, mul(kWeight8, w152[0])), \ + mul(kWeight7, w152[1])))); \ + store(output + 22 * stride, \ + add(w90[0], sub(sub(kWeight0, mul(kWeight4, w154[0])), \ + mul(kWeight3, w154[1])))); \ + store(output + 23 * stride, \ + add(w92[0], sub(sub(kWeight0, mul(kWeight6, w156[0])), \ + mul(kWeight5, w156[1])))); \ + store(output + 24 * stride, sub(w79[0], w143[1])); \ + store(output + 25 * stride, \ + add(w81[0], sub(mul(kWeight6, w145[0]), mul(kWeight5, w145[1])))); \ + store(output + 26 * stride, \ + add(w83[0], sub(mul(kWeight4, w147[0]), mul(kWeight3, w147[1])))); \ + store(output + 27 * stride, \ + add(w85[0], sub(mul(kWeight8, w149[0]), mul(kWeight7, w149[1])))); \ + store(output + 28 * stride, \ + add(w87[0], mul(kWeight2, sub(w151[0], w151[1])))); \ + store(output + 29 * stride, \ + add(w89[0], sub(mul(kWeight7, w153[0]), mul(kWeight8, w153[1])))); \ + store(output + 30 * stride, \ + add(w91[0], sub(mul(kWeight3, w155[0]), mul(kWeight4, w155[1])))); \ + store(output + 31 * stride, \ + add(w93[0], sub(mul(kWeight5, w157[0]), mul(kWeight6, w157[1])))); \ + } + +#endif // AOM_AOM_DSP_FFT_COMMON_H_ diff --git a/third_party/aom/aom_dsp/fwd_txfm.c b/third_party/aom/aom_dsp/fwd_txfm.c new file mode 100644 index 0000000000..e50f951c17 --- /dev/null +++ b/third_party/aom/aom_dsp/fwd_txfm.c @@ -0,0 +1,103 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include +#include "aom_dsp/txfm_common.h" +#include "config/aom_dsp_rtcd.h" + +void aom_fdct8x8_c(const int16_t *input, tran_low_t *final_output, int stride) { + int i, j; + tran_low_t intermediate[64]; + int pass; + tran_low_t *output = intermediate; + const tran_low_t *in = NULL; + + // Transform columns + for (pass = 0; pass < 2; ++pass) { + tran_high_t s0, s1, s2, s3, s4, s5, s6, s7; // canbe16 + tran_high_t t0, t1, t2, t3; // needs32 + tran_high_t x0, x1, x2, x3; // canbe16 + + for (i = 0; i < 8; i++) { + // stage 1 + if (pass == 0) { + s0 = (input[0 * stride] + input[7 * stride]) * 4; + s1 = (input[1 * stride] + input[6 * stride]) * 4; + s2 = (input[2 * stride] + input[5 * stride]) * 4; + s3 = (input[3 * stride] + input[4 * stride]) * 4; + s4 = (input[3 * stride] - input[4 * stride]) * 4; + s5 = (input[2 * stride] - input[5 * stride]) * 4; + s6 = (input[1 * stride] - input[6 * stride]) * 4; + s7 = (input[0 * stride] - input[7 * stride]) * 4; + ++input; + } else { + s0 = in[0 * 8] + in[7 * 8]; + s1 = in[1 * 8] + in[6 * 8]; + s2 = in[2 * 8] + in[5 * 8]; + s3 = in[3 * 8] + in[4 * 8]; + s4 = in[3 * 8] - in[4 * 8]; + s5 = in[2 * 8] - in[5 * 8]; + s6 = in[1 * 8] - in[6 * 8]; + s7 = in[0 * 8] - in[7 * 8]; + ++in; + } + + // fdct4(step, step); + x0 = s0 + s3; + x1 = s1 + s2; + x2 = s1 - s2; + x3 = s0 - s3; + t0 = (x0 + x1) * cospi_16_64; + t1 = (x0 - x1) * cospi_16_64; + t2 = x2 * cospi_24_64 + x3 * cospi_8_64; + t3 = -x2 * cospi_8_64 + x3 * cospi_24_64; + output[0] = (tran_low_t)fdct_round_shift(t0); + output[2] = (tran_low_t)fdct_round_shift(t2); + output[4] = (tran_low_t)fdct_round_shift(t1); + output[6] = (tran_low_t)fdct_round_shift(t3); + + // Stage 2 + t0 = (s6 - s5) * cospi_16_64; + t1 = (s6 + s5) * cospi_16_64; + t2 = fdct_round_shift(t0); + t3 = fdct_round_shift(t1); + + // Stage 3 + x0 = s4 + t2; + x1 = s4 - t2; + x2 = s7 - t3; + x3 = s7 + t3; + + // Stage 4 + t0 = x0 * cospi_28_64 + x3 * cospi_4_64; + t1 = x1 * cospi_12_64 + x2 * cospi_20_64; + t2 = x2 * cospi_12_64 + x1 * -cospi_20_64; + t3 = x3 * cospi_28_64 + x0 * -cospi_4_64; + output[1] = (tran_low_t)fdct_round_shift(t0); + output[3] = (tran_low_t)fdct_round_shift(t2); + output[5] = (tran_low_t)fdct_round_shift(t1); + output[7] = (tran_low_t)fdct_round_shift(t3); + output += 8; + } + in = intermediate; + output = final_output; + } + + // Rows + for (i = 0; i < 8; ++i) { + for (j = 0; j < 8; ++j) final_output[j + i * 8] /= 2; + } +} + +void aom_highbd_fdct8x8_c(const int16_t *input, tran_low_t *final_output, + int stride) { + aom_fdct8x8_c(input, final_output, stride); +} diff --git a/third_party/aom/aom_dsp/grain_synthesis.c b/third_party/aom/aom_dsp/grain_synthesis.c new file mode 100644 index 0000000000..b96e1c3190 --- /dev/null +++ b/third_party/aom/aom_dsp/grain_synthesis.c @@ -0,0 +1,1409 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +/*!\file + * \brief Describes film grain parameters and film grain synthesis + * + */ + +#include +#include +#include +#include +#include "aom_dsp/grain_synthesis.h" +#include "aom_mem/aom_mem.h" + +// Samples with Gaussian distribution in the range of [-2048, 2047] (12 bits) +// with zero mean and standard deviation of about 512. +// should be divided by 4 for 10-bit range and 16 for 8-bit range. +static const int gaussian_sequence[2048] = { + 56, 568, -180, 172, 124, -84, 172, -64, -900, 24, 820, + 224, 1248, 996, 272, -8, -916, -388, -732, -104, -188, 800, + 112, -652, -320, -376, 140, -252, 492, -168, 44, -788, 588, + -584, 500, -228, 12, 680, 272, -476, 972, -100, 652, 368, + 432, -196, -720, -192, 1000, -332, 652, -136, -552, -604, -4, + 192, -220, -136, 1000, -52, 372, -96, -624, 124, -24, 396, + 540, -12, -104, 640, 464, 244, -208, -84, 368, -528, -740, + 248, -968, -848, 608, 376, -60, -292, -40, -156, 252, -292, + 248, 224, -280, 400, -244, 244, -60, 76, -80, 212, 532, + 340, 128, -36, 824, -352, -60, -264, -96, -612, 416, -704, + 220, -204, 640, -160, 1220, -408, 900, 336, 20, -336, -96, + -792, 304, 48, -28, -1232, -1172, -448, 104, -292, -520, 244, + 60, -948, 0, -708, 268, 108, 356, -548, 488, -344, -136, + 488, -196, -224, 656, -236, -1128, 60, 4, 140, 276, -676, + -376, 168, -108, 464, 8, 564, 64, 240, 308, -300, -400, + -456, -136, 56, 120, -408, -116, 436, 504, -232, 328, 844, + -164, -84, 784, -168, 232, -224, 348, -376, 128, 568, 96, + -1244, -288, 276, 848, 832, -360, 656, 464, -384, -332, -356, + 728, -388, 160, -192, 468, 296, 224, 140, -776, -100, 280, + 4, 196, 44, -36, -648, 932, 16, 1428, 28, 528, 808, + 772, 20, 268, 88, -332, -284, 124, -384, -448, 208, -228, + -1044, -328, 660, 380, -148, -300, 588, 240, 540, 28, 136, + -88, -436, 256, 296, -1000, 1400, 0, -48, 1056, -136, 264, + -528, -1108, 632, -484, -592, -344, 796, 124, -668, -768, 388, + 1296, -232, -188, -200, -288, -4, 308, 100, -168, 256, -500, + 204, -508, 648, -136, 372, -272, -120, -1004, -552, -548, -384, + 548, -296, 428, -108, -8, -912, -324, -224, -88, -112, -220, + -100, 996, -796, 548, 360, -216, 180, 428, -200, -212, 148, + 96, 148, 284, 216, -412, -320, 120, -300, -384, -604, -572, + -332, -8, -180, -176, 696, 116, -88, 628, 76, 44, -516, + 240, -208, -40, 100, -592, 344, -308, -452, -228, 20, 916, + -1752, -136, -340, -804, 140, 40, 512, 340, 248, 184, -492, + 896, -156, 932, -628, 328, -688, -448, -616, -752, -100, 560, + -1020, 180, -800, -64, 76, 576, 1068, 396, 660, 552, -108, + -28, 320, -628, 312, -92, -92, -472, 268, 16, 560, 516, + -672, -52, 492, -100, 260, 384, 284, 292, 304, -148, 88, + -152, 1012, 1064, -228, 164, -376, -684, 592, -392, 156, 196, + -524, -64, -884, 160, -176, 636, 648, 404, -396, -436, 864, + 424, -728, 988, -604, 904, -592, 296, -224, 536, -176, -920, + 436, -48, 1176, -884, 416, -776, -824, -884, 524, -548, -564, + -68, -164, -96, 692, 364, -692, -1012, -68, 260, -480, 876, + -1116, 452, -332, -352, 892, -1088, 1220, -676, 12, -292, 244, + 496, 372, -32, 280, 200, 112, -440, -96, 24, -644, -184, + 56, -432, 224, -980, 272, -260, 144, -436, 420, 356, 364, + -528, 76, 172, -744, -368, 404, -752, -416, 684, -688, 72, + 540, 416, 92, 444, 480, -72, -1416, 164, -1172, -68, 24, + 424, 264, 1040, 128, -912, -524, -356, 64, 876, -12, 4, + -88, 532, 272, -524, 320, 276, -508, 940, 24, -400, -120, + 756, 60, 236, -412, 100, 376, -484, 400, -100, -740, -108, + -260, 328, -268, 224, -200, -416, 184, -604, -564, -20, 296, + 60, 892, -888, 60, 164, 68, -760, 216, -296, 904, -336, + -28, 404, -356, -568, -208, -1480, -512, 296, 328, -360, -164, + -1560, -776, 1156, -428, 164, -504, -112, 120, -216, -148, -264, + 308, 32, 64, -72, 72, 116, 176, -64, -272, 460, -536, + -784, -280, 348, 108, -752, -132, 524, -540, -776, 116, -296, + -1196, -288, -560, 1040, -472, 116, -848, -1116, 116, 636, 696, + 284, -176, 1016, 204, -864, -648, -248, 356, 972, -584, -204, + 264, 880, 528, -24, -184, 116, 448, -144, 828, 524, 212, + -212, 52, 12, 200, 268, -488, -404, -880, 824, -672, -40, + 908, -248, 500, 716, -576, 492, -576, 16, 720, -108, 384, + 124, 344, 280, 576, -500, 252, 104, -308, 196, -188, -8, + 1268, 296, 1032, -1196, 436, 316, 372, -432, -200, -660, 704, + -224, 596, -132, 268, 32, -452, 884, 104, -1008, 424, -1348, + -280, 4, -1168, 368, 476, 696, 300, -8, 24, 180, -592, + -196, 388, 304, 500, 724, -160, 244, -84, 272, -256, -420, + 320, 208, -144, -156, 156, 364, 452, 28, 540, 316, 220, + -644, -248, 464, 72, 360, 32, -388, 496, -680, -48, 208, + -116, -408, 60, -604, -392, 548, -840, 784, -460, 656, -544, + -388, -264, 908, -800, -628, -612, -568, 572, -220, 164, 288, + -16, -308, 308, -112, -636, -760, 280, -668, 432, 364, 240, + -196, 604, 340, 384, 196, 592, -44, -500, 432, -580, -132, + 636, -76, 392, 4, -412, 540, 508, 328, -356, -36, 16, + -220, -64, -248, -60, 24, -192, 368, 1040, 92, -24, -1044, + -32, 40, 104, 148, 192, -136, -520, 56, -816, -224, 732, + 392, 356, 212, -80, -424, -1008, -324, 588, -1496, 576, 460, + -816, -848, 56, -580, -92, -1372, -112, -496, 200, 364, 52, + -140, 48, -48, -60, 84, 72, 40, 132, -356, -268, -104, + -284, -404, 732, -520, 164, -304, -540, 120, 328, -76, -460, + 756, 388, 588, 236, -436, -72, -176, -404, -316, -148, 716, + -604, 404, -72, -88, -888, -68, 944, 88, -220, -344, 960, + 472, 460, -232, 704, 120, 832, -228, 692, -508, 132, -476, + 844, -748, -364, -44, 1116, -1104, -1056, 76, 428, 552, -692, + 60, 356, 96, -384, -188, -612, -576, 736, 508, 892, 352, + -1132, 504, -24, -352, 324, 332, -600, -312, 292, 508, -144, + -8, 484, 48, 284, -260, -240, 256, -100, -292, -204, -44, + 472, -204, 908, -188, -1000, -256, 92, 1164, -392, 564, 356, + 652, -28, -884, 256, 484, -192, 760, -176, 376, -524, -452, + -436, 860, -736, 212, 124, 504, -476, 468, 76, -472, 552, + -692, -944, -620, 740, -240, 400, 132, 20, 192, -196, 264, + -668, -1012, -60, 296, -316, -828, 76, -156, 284, -768, -448, + -832, 148, 248, 652, 616, 1236, 288, -328, -400, -124, 588, + 220, 520, -696, 1032, 768, -740, -92, -272, 296, 448, -464, + 412, -200, 392, 440, -200, 264, -152, -260, 320, 1032, 216, + 320, -8, -64, 156, -1016, 1084, 1172, 536, 484, -432, 132, + 372, -52, -256, 84, 116, -352, 48, 116, 304, -384, 412, + 924, -300, 528, 628, 180, 648, 44, -980, -220, 1320, 48, + 332, 748, 524, -268, -720, 540, -276, 564, -344, -208, -196, + 436, 896, 88, -392, 132, 80, -964, -288, 568, 56, -48, + -456, 888, 8, 552, -156, -292, 948, 288, 128, -716, -292, + 1192, -152, 876, 352, -600, -260, -812, -468, -28, -120, -32, + -44, 1284, 496, 192, 464, 312, -76, -516, -380, -456, -1012, + -48, 308, -156, 36, 492, -156, -808, 188, 1652, 68, -120, + -116, 316, 160, -140, 352, 808, -416, 592, 316, -480, 56, + 528, -204, -568, 372, -232, 752, -344, 744, -4, 324, -416, + -600, 768, 268, -248, -88, -132, -420, -432, 80, -288, 404, + -316, -1216, -588, 520, -108, 92, -320, 368, -480, -216, -92, + 1688, -300, 180, 1020, -176, 820, -68, -228, -260, 436, -904, + 20, 40, -508, 440, -736, 312, 332, 204, 760, -372, 728, + 96, -20, -632, -520, -560, 336, 1076, -64, -532, 776, 584, + 192, 396, -728, -520, 276, -188, 80, -52, -612, -252, -48, + 648, 212, -688, 228, -52, -260, 428, -412, -272, -404, 180, + 816, -796, 48, 152, 484, -88, -216, 988, 696, 188, -528, + 648, -116, -180, 316, 476, 12, -564, 96, 476, -252, -364, + -376, -392, 556, -256, -576, 260, -352, 120, -16, -136, -260, + -492, 72, 556, 660, 580, 616, 772, 436, 424, -32, -324, + -1268, 416, -324, -80, 920, 160, 228, 724, 32, -516, 64, + 384, 68, -128, 136, 240, 248, -204, -68, 252, -932, -120, + -480, -628, -84, 192, 852, -404, -288, -132, 204, 100, 168, + -68, -196, -868, 460, 1080, 380, -80, 244, 0, 484, -888, + 64, 184, 352, 600, 460, 164, 604, -196, 320, -64, 588, + -184, 228, 12, 372, 48, -848, -344, 224, 208, -200, 484, + 128, -20, 272, -468, -840, 384, 256, -720, -520, -464, -580, + 112, -120, 644, -356, -208, -608, -528, 704, 560, -424, 392, + 828, 40, 84, 200, -152, 0, -144, 584, 280, -120, 80, + -556, -972, -196, -472, 724, 80, 168, -32, 88, 160, -688, + 0, 160, 356, 372, -776, 740, -128, 676, -248, -480, 4, + -364, 96, 544, 232, -1032, 956, 236, 356, 20, -40, 300, + 24, -676, -596, 132, 1120, -104, 532, -1096, 568, 648, 444, + 508, 380, 188, -376, -604, 1488, 424, 24, 756, -220, -192, + 716, 120, 920, 688, 168, 44, -460, 568, 284, 1144, 1160, + 600, 424, 888, 656, -356, -320, 220, 316, -176, -724, -188, + -816, -628, -348, -228, -380, 1012, -452, -660, 736, 928, 404, + -696, -72, -268, -892, 128, 184, -344, -780, 360, 336, 400, + 344, 428, 548, -112, 136, -228, -216, -820, -516, 340, 92, + -136, 116, -300, 376, -244, 100, -316, -520, -284, -12, 824, + 164, -548, -180, -128, 116, -924, -828, 268, -368, -580, 620, + 192, 160, 0, -1676, 1068, 424, -56, -360, 468, -156, 720, + 288, -528, 556, -364, 548, -148, 504, 316, 152, -648, -620, + -684, -24, -376, -384, -108, -920, -1032, 768, 180, -264, -508, + -1268, -260, -60, 300, -240, 988, 724, -376, -576, -212, -736, + 556, 192, 1092, -620, -880, 376, -56, -4, -216, -32, 836, + 268, 396, 1332, 864, -600, 100, 56, -412, -92, 356, 180, + 884, -468, -436, 292, -388, -804, -704, -840, 368, -348, 140, + -724, 1536, 940, 372, 112, -372, 436, -480, 1136, 296, -32, + -228, 132, -48, -220, 868, -1016, -60, -1044, -464, 328, 916, + 244, 12, -736, -296, 360, 468, -376, -108, -92, 788, 368, + -56, 544, 400, -672, -420, 728, 16, 320, 44, -284, -380, + -796, 488, 132, 204, -596, -372, 88, -152, -908, -636, -572, + -624, -116, -692, -200, -56, 276, -88, 484, -324, 948, 864, + 1000, -456, -184, -276, 292, -296, 156, 676, 320, 160, 908, + -84, -1236, -288, -116, 260, -372, -644, 732, -756, -96, 84, + 344, -520, 348, -688, 240, -84, 216, -1044, -136, -676, -396, + -1500, 960, -40, 176, 168, 1516, 420, -504, -344, -364, -360, + 1216, -940, -380, -212, 252, -660, -708, 484, -444, -152, 928, + -120, 1112, 476, -260, 560, -148, -344, 108, -196, 228, -288, + 504, 560, -328, -88, 288, -1008, 460, -228, 468, -836, -196, + 76, 388, 232, 412, -1168, -716, -644, 756, -172, -356, -504, + 116, 432, 528, 48, 476, -168, -608, 448, 160, -532, -272, + 28, -676, -12, 828, 980, 456, 520, 104, -104, 256, -344, + -4, -28, -368, -52, -524, -572, -556, -200, 768, 1124, -208, + -512, 176, 232, 248, -148, -888, 604, -600, -304, 804, -156, + -212, 488, -192, -804, -256, 368, -360, -916, -328, 228, -240, + -448, -472, 856, -556, -364, 572, -12, -156, -368, -340, 432, + 252, -752, -152, 288, 268, -580, -848, -592, 108, -76, 244, + 312, -716, 592, -80, 436, 360, 4, -248, 160, 516, 584, + 732, 44, -468, -280, -292, -156, -588, 28, 308, 912, 24, + 124, 156, 180, -252, 944, -924, -772, -520, -428, -624, 300, + -212, -1144, 32, -724, 800, -1128, -212, -1288, -848, 180, -416, + 440, 192, -576, -792, -76, -1080, 80, -532, -352, -132, 380, + -820, 148, 1112, 128, 164, 456, 700, -924, 144, -668, -384, + 648, -832, 508, 552, -52, -100, -656, 208, -568, 748, -88, + 680, 232, 300, 192, -408, -1012, -152, -252, -268, 272, -876, + -664, -648, -332, -136, 16, 12, 1152, -28, 332, -536, 320, + -672, -460, -316, 532, -260, 228, -40, 1052, -816, 180, 88, + -496, -556, -672, -368, 428, 92, 356, 404, -408, 252, 196, + -176, -556, 792, 268, 32, 372, 40, 96, -332, 328, 120, + 372, -900, -40, 472, -264, -592, 952, 128, 656, 112, 664, + -232, 420, 4, -344, -464, 556, 244, -416, -32, 252, 0, + -412, 188, -696, 508, -476, 324, -1096, 656, -312, 560, 264, + -136, 304, 160, -64, -580, 248, 336, -720, 560, -348, -288, + -276, -196, -500, 852, -544, -236, -1128, -992, -776, 116, 56, + 52, 860, 884, 212, -12, 168, 1020, 512, -552, 924, -148, + 716, 188, 164, -340, -520, -184, 880, -152, -680, -208, -1156, + -300, -528, -472, 364, 100, -744, -1056, -32, 540, 280, 144, + -676, -32, -232, -280, -224, 96, 568, -76, 172, 148, 148, + 104, 32, -296, -32, 788, -80, 32, -16, 280, 288, 944, + 428, -484 +}; + +static const int gauss_bits = 11; + +static int luma_subblock_size_y = 32; +static int luma_subblock_size_x = 32; + +static int chroma_subblock_size_y = 16; +static int chroma_subblock_size_x = 16; + +static const int min_luma_legal_range = 16; +static const int max_luma_legal_range = 235; + +static const int min_chroma_legal_range = 16; +static const int max_chroma_legal_range = 240; + +static int scaling_lut_y[256]; +static int scaling_lut_cb[256]; +static int scaling_lut_cr[256]; + +static int grain_center; +static int grain_min; +static int grain_max; + +static uint16_t random_register = 0; // random number generator register + +static void init_arrays(const aom_film_grain_t *params, int luma_stride, + int chroma_stride, int ***pred_pos_luma_p, + int ***pred_pos_chroma_p, int **luma_grain_block, + int **cb_grain_block, int **cr_grain_block, + int **y_line_buf, int **cb_line_buf, int **cr_line_buf, + int **y_col_buf, int **cb_col_buf, int **cr_col_buf, + int luma_grain_samples, int chroma_grain_samples, + int chroma_subsamp_y, int chroma_subsamp_x) { + memset(scaling_lut_y, 0, sizeof(*scaling_lut_y) * 256); + memset(scaling_lut_cb, 0, sizeof(*scaling_lut_cb) * 256); + memset(scaling_lut_cr, 0, sizeof(*scaling_lut_cr) * 256); + + int num_pos_luma = 2 * params->ar_coeff_lag * (params->ar_coeff_lag + 1); + int num_pos_chroma = num_pos_luma; + if (params->num_y_points > 0) ++num_pos_chroma; + + int **pred_pos_luma; + int **pred_pos_chroma; + + pred_pos_luma = (int **)aom_malloc(sizeof(*pred_pos_luma) * num_pos_luma); + + for (int row = 0; row < num_pos_luma; row++) { + pred_pos_luma[row] = (int *)aom_malloc(sizeof(**pred_pos_luma) * 3); + } + + pred_pos_chroma = + (int **)aom_malloc(sizeof(*pred_pos_chroma) * num_pos_chroma); + + for (int row = 0; row < num_pos_chroma; row++) { + pred_pos_chroma[row] = (int *)aom_malloc(sizeof(**pred_pos_chroma) * 3); + } + + int pos_ar_index = 0; + + for (int row = -params->ar_coeff_lag; row < 0; row++) { + for (int col = -params->ar_coeff_lag; col < params->ar_coeff_lag + 1; + col++) { + pred_pos_luma[pos_ar_index][0] = row; + pred_pos_luma[pos_ar_index][1] = col; + pred_pos_luma[pos_ar_index][2] = 0; + + pred_pos_chroma[pos_ar_index][0] = row; + pred_pos_chroma[pos_ar_index][1] = col; + pred_pos_chroma[pos_ar_index][2] = 0; + ++pos_ar_index; + } + } + + for (int col = -params->ar_coeff_lag; col < 0; col++) { + pred_pos_luma[pos_ar_index][0] = 0; + pred_pos_luma[pos_ar_index][1] = col; + pred_pos_luma[pos_ar_index][2] = 0; + + pred_pos_chroma[pos_ar_index][0] = 0; + pred_pos_chroma[pos_ar_index][1] = col; + pred_pos_chroma[pos_ar_index][2] = 0; + + ++pos_ar_index; + } + + if (params->num_y_points > 0) { + pred_pos_chroma[pos_ar_index][0] = 0; + pred_pos_chroma[pos_ar_index][1] = 0; + pred_pos_chroma[pos_ar_index][2] = 1; + } + + *pred_pos_luma_p = pred_pos_luma; + *pred_pos_chroma_p = pred_pos_chroma; + + *y_line_buf = (int *)aom_malloc(sizeof(**y_line_buf) * luma_stride * 2); + *cb_line_buf = (int *)aom_malloc(sizeof(**cb_line_buf) * chroma_stride * + (2 >> chroma_subsamp_y)); + *cr_line_buf = (int *)aom_malloc(sizeof(**cr_line_buf) * chroma_stride * + (2 >> chroma_subsamp_y)); + + *y_col_buf = + (int *)aom_malloc(sizeof(**y_col_buf) * (luma_subblock_size_y + 2) * 2); + *cb_col_buf = + (int *)aom_malloc(sizeof(**cb_col_buf) * + (chroma_subblock_size_y + (2 >> chroma_subsamp_y)) * + (2 >> chroma_subsamp_x)); + *cr_col_buf = + (int *)aom_malloc(sizeof(**cr_col_buf) * + (chroma_subblock_size_y + (2 >> chroma_subsamp_y)) * + (2 >> chroma_subsamp_x)); + + *luma_grain_block = + (int *)aom_malloc(sizeof(**luma_grain_block) * luma_grain_samples); + *cb_grain_block = + (int *)aom_malloc(sizeof(**cb_grain_block) * chroma_grain_samples); + *cr_grain_block = + (int *)aom_malloc(sizeof(**cr_grain_block) * chroma_grain_samples); +} + +static void dealloc_arrays(const aom_film_grain_t *params, int ***pred_pos_luma, + int ***pred_pos_chroma, int **luma_grain_block, + int **cb_grain_block, int **cr_grain_block, + int **y_line_buf, int **cb_line_buf, + int **cr_line_buf, int **y_col_buf, int **cb_col_buf, + int **cr_col_buf) { + int num_pos_luma = 2 * params->ar_coeff_lag * (params->ar_coeff_lag + 1); + int num_pos_chroma = num_pos_luma; + if (params->num_y_points > 0) ++num_pos_chroma; + + for (int row = 0; row < num_pos_luma; row++) { + aom_free((*pred_pos_luma)[row]); + } + aom_free(*pred_pos_luma); + + for (int row = 0; row < num_pos_chroma; row++) { + aom_free((*pred_pos_chroma)[row]); + } + aom_free((*pred_pos_chroma)); + + aom_free(*y_line_buf); + + aom_free(*cb_line_buf); + + aom_free(*cr_line_buf); + + aom_free(*y_col_buf); + + aom_free(*cb_col_buf); + + aom_free(*cr_col_buf); + + aom_free(*luma_grain_block); + + aom_free(*cb_grain_block); + + aom_free(*cr_grain_block); +} + +// get a number between 0 and 2^bits - 1 +static INLINE int get_random_number(int bits) { + uint16_t bit; + bit = ((random_register >> 0) ^ (random_register >> 1) ^ + (random_register >> 3) ^ (random_register >> 12)) & + 1; + random_register = (random_register >> 1) | (bit << 15); + return (random_register >> (16 - bits)) & ((1 << bits) - 1); +} + +static void init_random_generator(int luma_line, uint16_t seed) { + // same for the picture + + uint16_t msb = (seed >> 8) & 255; + uint16_t lsb = seed & 255; + + random_register = (msb << 8) + lsb; + + // changes for each row + int luma_num = luma_line >> 5; + + random_register ^= ((luma_num * 37 + 178) & 255) << 8; + random_register ^= ((luma_num * 173 + 105) & 255); +} + +// Return 0 for success, -1 for failure +static int generate_luma_grain_block( + const aom_film_grain_t *params, int **pred_pos_luma, int *luma_grain_block, + int luma_block_size_y, int luma_block_size_x, int luma_grain_stride, + int left_pad, int top_pad, int right_pad, int bottom_pad) { + if (params->num_y_points == 0) { + memset(luma_grain_block, 0, + sizeof(*luma_grain_block) * luma_block_size_y * luma_grain_stride); + return 0; + } + + int bit_depth = params->bit_depth; + int gauss_sec_shift = 12 - bit_depth + params->grain_scale_shift; + + int num_pos_luma = 2 * params->ar_coeff_lag * (params->ar_coeff_lag + 1); + int rounding_offset = (1 << (params->ar_coeff_shift - 1)); + + for (int i = 0; i < luma_block_size_y; i++) + for (int j = 0; j < luma_block_size_x; j++) + luma_grain_block[i * luma_grain_stride + j] = + (gaussian_sequence[get_random_number(gauss_bits)] + + ((1 << gauss_sec_shift) >> 1)) >> + gauss_sec_shift; + + for (int i = top_pad; i < luma_block_size_y - bottom_pad; i++) + for (int j = left_pad; j < luma_block_size_x - right_pad; j++) { + int wsum = 0; + for (int pos = 0; pos < num_pos_luma; pos++) { + wsum = wsum + params->ar_coeffs_y[pos] * + luma_grain_block[(i + pred_pos_luma[pos][0]) * + luma_grain_stride + + j + pred_pos_luma[pos][1]]; + } + luma_grain_block[i * luma_grain_stride + j] = + clamp(luma_grain_block[i * luma_grain_stride + j] + + ((wsum + rounding_offset) >> params->ar_coeff_shift), + grain_min, grain_max); + } + return 0; +} + +// Return 0 for success, -1 for failure +static int generate_chroma_grain_blocks( + const aom_film_grain_t *params, + // int** pred_pos_luma, + int **pred_pos_chroma, int *luma_grain_block, int *cb_grain_block, + int *cr_grain_block, int luma_grain_stride, int chroma_block_size_y, + int chroma_block_size_x, int chroma_grain_stride, int left_pad, int top_pad, + int right_pad, int bottom_pad, int chroma_subsamp_y, int chroma_subsamp_x) { + int bit_depth = params->bit_depth; + int gauss_sec_shift = 12 - bit_depth + params->grain_scale_shift; + + int num_pos_chroma = 2 * params->ar_coeff_lag * (params->ar_coeff_lag + 1); + if (params->num_y_points > 0) ++num_pos_chroma; + int rounding_offset = (1 << (params->ar_coeff_shift - 1)); + int chroma_grain_block_size = chroma_block_size_y * chroma_grain_stride; + + if (params->num_cb_points || params->chroma_scaling_from_luma) { + init_random_generator(7 << 5, params->random_seed); + + for (int i = 0; i < chroma_block_size_y; i++) + for (int j = 0; j < chroma_block_size_x; j++) + cb_grain_block[i * chroma_grain_stride + j] = + (gaussian_sequence[get_random_number(gauss_bits)] + + ((1 << gauss_sec_shift) >> 1)) >> + gauss_sec_shift; + } else { + memset(cb_grain_block, 0, + sizeof(*cb_grain_block) * chroma_grain_block_size); + } + + if (params->num_cr_points || params->chroma_scaling_from_luma) { + init_random_generator(11 << 5, params->random_seed); + + for (int i = 0; i < chroma_block_size_y; i++) + for (int j = 0; j < chroma_block_size_x; j++) + cr_grain_block[i * chroma_grain_stride + j] = + (gaussian_sequence[get_random_number(gauss_bits)] + + ((1 << gauss_sec_shift) >> 1)) >> + gauss_sec_shift; + } else { + memset(cr_grain_block, 0, + sizeof(*cr_grain_block) * chroma_grain_block_size); + } + + for (int i = top_pad; i < chroma_block_size_y - bottom_pad; i++) + for (int j = left_pad; j < chroma_block_size_x - right_pad; j++) { + int wsum_cb = 0; + int wsum_cr = 0; + for (int pos = 0; pos < num_pos_chroma; pos++) { + if (pred_pos_chroma[pos][2] == 0) { + wsum_cb = wsum_cb + params->ar_coeffs_cb[pos] * + cb_grain_block[(i + pred_pos_chroma[pos][0]) * + chroma_grain_stride + + j + pred_pos_chroma[pos][1]]; + wsum_cr = wsum_cr + params->ar_coeffs_cr[pos] * + cr_grain_block[(i + pred_pos_chroma[pos][0]) * + chroma_grain_stride + + j + pred_pos_chroma[pos][1]]; + } else if (pred_pos_chroma[pos][2] == 1) { + int av_luma = 0; + int luma_coord_y = ((i - top_pad) << chroma_subsamp_y) + top_pad; + int luma_coord_x = ((j - left_pad) << chroma_subsamp_x) + left_pad; + + for (int k = luma_coord_y; k < luma_coord_y + chroma_subsamp_y + 1; + k++) + for (int l = luma_coord_x; l < luma_coord_x + chroma_subsamp_x + 1; + l++) + av_luma += luma_grain_block[k * luma_grain_stride + l]; + + av_luma = + (av_luma + ((1 << (chroma_subsamp_y + chroma_subsamp_x)) >> 1)) >> + (chroma_subsamp_y + chroma_subsamp_x); + + wsum_cb = wsum_cb + params->ar_coeffs_cb[pos] * av_luma; + wsum_cr = wsum_cr + params->ar_coeffs_cr[pos] * av_luma; + } else { + fprintf( + stderr, + "Grain synthesis: prediction between two chroma components is " + "not supported!"); + return -1; + } + } + if (params->num_cb_points || params->chroma_scaling_from_luma) + cb_grain_block[i * chroma_grain_stride + j] = + clamp(cb_grain_block[i * chroma_grain_stride + j] + + ((wsum_cb + rounding_offset) >> params->ar_coeff_shift), + grain_min, grain_max); + if (params->num_cr_points || params->chroma_scaling_from_luma) + cr_grain_block[i * chroma_grain_stride + j] = + clamp(cr_grain_block[i * chroma_grain_stride + j] + + ((wsum_cr + rounding_offset) >> params->ar_coeff_shift), + grain_min, grain_max); + } + return 0; +} + +static void init_scaling_function(const int scaling_points[][2], int num_points, + int scaling_lut[]) { + if (num_points == 0) return; + + for (int i = 0; i < scaling_points[0][0]; i++) + scaling_lut[i] = scaling_points[0][1]; + + for (int point = 0; point < num_points - 1; point++) { + int delta_y = scaling_points[point + 1][1] - scaling_points[point][1]; + int delta_x = scaling_points[point + 1][0] - scaling_points[point][0]; + + int64_t delta = delta_y * ((65536 + (delta_x >> 1)) / delta_x); + + for (int x = 0; x < delta_x; x++) { + scaling_lut[scaling_points[point][0] + x] = + scaling_points[point][1] + (int)((x * delta + 32768) >> 16); + } + } + + for (int i = scaling_points[num_points - 1][0]; i < 256; i++) + scaling_lut[i] = scaling_points[num_points - 1][1]; +} + +// function that extracts samples from a LUT (and interpolates intemediate +// frames for 10- and 12-bit video) +static int scale_LUT(int *scaling_lut, int index, int bit_depth) { + int x = index >> (bit_depth - 8); + + if (!(bit_depth - 8) || x == 255) + return scaling_lut[x]; + else + return scaling_lut[x] + (((scaling_lut[x + 1] - scaling_lut[x]) * + (index & ((1 << (bit_depth - 8)) - 1)) + + (1 << (bit_depth - 9))) >> + (bit_depth - 8)); +} + +static void add_noise_to_block(const aom_film_grain_t *params, uint8_t *luma, + uint8_t *cb, uint8_t *cr, int luma_stride, + int chroma_stride, int *luma_grain, + int *cb_grain, int *cr_grain, + int luma_grain_stride, int chroma_grain_stride, + int half_luma_height, int half_luma_width, + int bit_depth, int chroma_subsamp_y, + int chroma_subsamp_x, int mc_identity) { + int cb_mult = params->cb_mult - 128; // fixed scale + int cb_luma_mult = params->cb_luma_mult - 128; // fixed scale + int cb_offset = params->cb_offset - 256; + + int cr_mult = params->cr_mult - 128; // fixed scale + int cr_luma_mult = params->cr_luma_mult - 128; // fixed scale + int cr_offset = params->cr_offset - 256; + + int rounding_offset = (1 << (params->scaling_shift - 1)); + + int apply_y = params->num_y_points > 0 ? 1 : 0; + int apply_cb = + (params->num_cb_points > 0 || params->chroma_scaling_from_luma) ? 1 : 0; + int apply_cr = + (params->num_cr_points > 0 || params->chroma_scaling_from_luma) ? 1 : 0; + + if (params->chroma_scaling_from_luma) { + cb_mult = 0; // fixed scale + cb_luma_mult = 64; // fixed scale + cb_offset = 0; + + cr_mult = 0; // fixed scale + cr_luma_mult = 64; // fixed scale + cr_offset = 0; + } + + int min_luma, max_luma, min_chroma, max_chroma; + + if (params->clip_to_restricted_range) { + min_luma = min_luma_legal_range; + max_luma = max_luma_legal_range; + + if (mc_identity) { + min_chroma = min_luma_legal_range; + max_chroma = max_luma_legal_range; + } else { + min_chroma = min_chroma_legal_range; + max_chroma = max_chroma_legal_range; + } + } else { + min_luma = min_chroma = 0; + max_luma = max_chroma = 255; + } + + for (int i = 0; i < (half_luma_height << (1 - chroma_subsamp_y)); i++) { + for (int j = 0; j < (half_luma_width << (1 - chroma_subsamp_x)); j++) { + int average_luma = 0; + if (chroma_subsamp_x) { + average_luma = (luma[(i << chroma_subsamp_y) * luma_stride + + (j << chroma_subsamp_x)] + + luma[(i << chroma_subsamp_y) * luma_stride + + (j << chroma_subsamp_x) + 1] + + 1) >> + 1; + } else { + average_luma = luma[(i << chroma_subsamp_y) * luma_stride + j]; + } + + if (apply_cb) { + cb[i * chroma_stride + j] = clamp( + cb[i * chroma_stride + j] + + ((scale_LUT(scaling_lut_cb, + clamp(((average_luma * cb_luma_mult + + cb_mult * cb[i * chroma_stride + j]) >> + 6) + + cb_offset, + 0, (256 << (bit_depth - 8)) - 1), + 8) * + cb_grain[i * chroma_grain_stride + j] + + rounding_offset) >> + params->scaling_shift), + min_chroma, max_chroma); + } + + if (apply_cr) { + cr[i * chroma_stride + j] = clamp( + cr[i * chroma_stride + j] + + ((scale_LUT(scaling_lut_cr, + clamp(((average_luma * cr_luma_mult + + cr_mult * cr[i * chroma_stride + j]) >> + 6) + + cr_offset, + 0, (256 << (bit_depth - 8)) - 1), + 8) * + cr_grain[i * chroma_grain_stride + j] + + rounding_offset) >> + params->scaling_shift), + min_chroma, max_chroma); + } + } + } + + if (apply_y) { + for (int i = 0; i < (half_luma_height << 1); i++) { + for (int j = 0; j < (half_luma_width << 1); j++) { + luma[i * luma_stride + j] = + clamp(luma[i * luma_stride + j] + + ((scale_LUT(scaling_lut_y, luma[i * luma_stride + j], 8) * + luma_grain[i * luma_grain_stride + j] + + rounding_offset) >> + params->scaling_shift), + min_luma, max_luma); + } + } + } +} + +static void add_noise_to_block_hbd( + const aom_film_grain_t *params, uint16_t *luma, uint16_t *cb, uint16_t *cr, + int luma_stride, int chroma_stride, int *luma_grain, int *cb_grain, + int *cr_grain, int luma_grain_stride, int chroma_grain_stride, + int half_luma_height, int half_luma_width, int bit_depth, + int chroma_subsamp_y, int chroma_subsamp_x, int mc_identity) { + int cb_mult = params->cb_mult - 128; // fixed scale + int cb_luma_mult = params->cb_luma_mult - 128; // fixed scale + // offset value depends on the bit depth + int cb_offset = (params->cb_offset << (bit_depth - 8)) - (1 << bit_depth); + + int cr_mult = params->cr_mult - 128; // fixed scale + int cr_luma_mult = params->cr_luma_mult - 128; // fixed scale + // offset value depends on the bit depth + int cr_offset = (params->cr_offset << (bit_depth - 8)) - (1 << bit_depth); + + int rounding_offset = (1 << (params->scaling_shift - 1)); + + int apply_y = params->num_y_points > 0 ? 1 : 0; + int apply_cb = + (params->num_cb_points > 0 || params->chroma_scaling_from_luma) > 0 ? 1 + : 0; + int apply_cr = + (params->num_cr_points > 0 || params->chroma_scaling_from_luma) > 0 ? 1 + : 0; + + if (params->chroma_scaling_from_luma) { + cb_mult = 0; // fixed scale + cb_luma_mult = 64; // fixed scale + cb_offset = 0; + + cr_mult = 0; // fixed scale + cr_luma_mult = 64; // fixed scale + cr_offset = 0; + } + + int min_luma, max_luma, min_chroma, max_chroma; + + if (params->clip_to_restricted_range) { + min_luma = min_luma_legal_range << (bit_depth - 8); + max_luma = max_luma_legal_range << (bit_depth - 8); + + if (mc_identity) { + min_chroma = min_luma_legal_range << (bit_depth - 8); + max_chroma = max_luma_legal_range << (bit_depth - 8); + } else { + min_chroma = min_chroma_legal_range << (bit_depth - 8); + max_chroma = max_chroma_legal_range << (bit_depth - 8); + } + } else { + min_luma = min_chroma = 0; + max_luma = max_chroma = (256 << (bit_depth - 8)) - 1; + } + + for (int i = 0; i < (half_luma_height << (1 - chroma_subsamp_y)); i++) { + for (int j = 0; j < (half_luma_width << (1 - chroma_subsamp_x)); j++) { + int average_luma = 0; + if (chroma_subsamp_x) { + average_luma = (luma[(i << chroma_subsamp_y) * luma_stride + + (j << chroma_subsamp_x)] + + luma[(i << chroma_subsamp_y) * luma_stride + + (j << chroma_subsamp_x) + 1] + + 1) >> + 1; + } else { + average_luma = luma[(i << chroma_subsamp_y) * luma_stride + j]; + } + + if (apply_cb) { + cb[i * chroma_stride + j] = clamp( + cb[i * chroma_stride + j] + + ((scale_LUT(scaling_lut_cb, + clamp(((average_luma * cb_luma_mult + + cb_mult * cb[i * chroma_stride + j]) >> + 6) + + cb_offset, + 0, (256 << (bit_depth - 8)) - 1), + bit_depth) * + cb_grain[i * chroma_grain_stride + j] + + rounding_offset) >> + params->scaling_shift), + min_chroma, max_chroma); + } + if (apply_cr) { + cr[i * chroma_stride + j] = clamp( + cr[i * chroma_stride + j] + + ((scale_LUT(scaling_lut_cr, + clamp(((average_luma * cr_luma_mult + + cr_mult * cr[i * chroma_stride + j]) >> + 6) + + cr_offset, + 0, (256 << (bit_depth - 8)) - 1), + bit_depth) * + cr_grain[i * chroma_grain_stride + j] + + rounding_offset) >> + params->scaling_shift), + min_chroma, max_chroma); + } + } + } + + if (apply_y) { + for (int i = 0; i < (half_luma_height << 1); i++) { + for (int j = 0; j < (half_luma_width << 1); j++) { + luma[i * luma_stride + j] = + clamp(luma[i * luma_stride + j] + + ((scale_LUT(scaling_lut_y, luma[i * luma_stride + j], + bit_depth) * + luma_grain[i * luma_grain_stride + j] + + rounding_offset) >> + params->scaling_shift), + min_luma, max_luma); + } + } + } +} + +static void copy_rect(uint8_t *src, int src_stride, uint8_t *dst, + int dst_stride, int width, int height, + int use_high_bit_depth) { + int hbd_coeff = use_high_bit_depth ? 2 : 1; + while (height) { + memcpy(dst, src, width * sizeof(uint8_t) * hbd_coeff); + src += src_stride; + dst += dst_stride; + --height; + } + return; +} + +static void copy_area(int *src, int src_stride, int *dst, int dst_stride, + int width, int height) { + while (height) { + memcpy(dst, src, width * sizeof(*src)); + src += src_stride; + dst += dst_stride; + --height; + } + return; +} + +static void extend_even(uint8_t *dst, int dst_stride, int width, int height, + int use_high_bit_depth) { + if ((width & 1) == 0 && (height & 1) == 0) return; + if (use_high_bit_depth) { + uint16_t *dst16 = (uint16_t *)dst; + int dst16_stride = dst_stride / 2; + if (width & 1) { + for (int i = 0; i < height; ++i) + dst16[i * dst16_stride + width] = dst16[i * dst16_stride + width - 1]; + } + width = (width + 1) & (~1); + if (height & 1) { + memcpy(&dst16[height * dst16_stride], &dst16[(height - 1) * dst16_stride], + sizeof(*dst16) * width); + } + } else { + if (width & 1) { + for (int i = 0; i < height; ++i) + dst[i * dst_stride + width] = dst[i * dst_stride + width - 1]; + } + width = (width + 1) & (~1); + if (height & 1) { + memcpy(&dst[height * dst_stride], &dst[(height - 1) * dst_stride], + sizeof(*dst) * width); + } + } +} + +static void ver_boundary_overlap(int *left_block, int left_stride, + int *right_block, int right_stride, + int *dst_block, int dst_stride, int width, + int height) { + if (width == 1) { + while (height) { + *dst_block = clamp((*left_block * 23 + *right_block * 22 + 16) >> 5, + grain_min, grain_max); + left_block += left_stride; + right_block += right_stride; + dst_block += dst_stride; + --height; + } + return; + } else if (width == 2) { + while (height) { + dst_block[0] = clamp((27 * left_block[0] + 17 * right_block[0] + 16) >> 5, + grain_min, grain_max); + dst_block[1] = clamp((17 * left_block[1] + 27 * right_block[1] + 16) >> 5, + grain_min, grain_max); + left_block += left_stride; + right_block += right_stride; + dst_block += dst_stride; + --height; + } + return; + } +} + +static void hor_boundary_overlap(int *top_block, int top_stride, + int *bottom_block, int bottom_stride, + int *dst_block, int dst_stride, int width, + int height) { + if (height == 1) { + while (width) { + *dst_block = clamp((*top_block * 23 + *bottom_block * 22 + 16) >> 5, + grain_min, grain_max); + ++top_block; + ++bottom_block; + ++dst_block; + --width; + } + return; + } else if (height == 2) { + while (width) { + dst_block[0] = clamp((27 * top_block[0] + 17 * bottom_block[0] + 16) >> 5, + grain_min, grain_max); + dst_block[dst_stride] = clamp((17 * top_block[top_stride] + + 27 * bottom_block[bottom_stride] + 16) >> + 5, + grain_min, grain_max); + ++top_block; + ++bottom_block; + ++dst_block; + --width; + } + return; + } +} + +int av1_add_film_grain(const aom_film_grain_t *params, const aom_image_t *src, + aom_image_t *dst) { + uint8_t *luma, *cb, *cr; + int height, width, luma_stride, chroma_stride; + int use_high_bit_depth = 0; + int chroma_subsamp_x = 0; + int chroma_subsamp_y = 0; + int mc_identity = src->mc == AOM_CICP_MC_IDENTITY ? 1 : 0; + + switch (src->fmt) { + case AOM_IMG_FMT_AOMI420: + case AOM_IMG_FMT_I420: + use_high_bit_depth = 0; + chroma_subsamp_x = 1; + chroma_subsamp_y = 1; + break; + case AOM_IMG_FMT_I42016: + use_high_bit_depth = 1; + chroma_subsamp_x = 1; + chroma_subsamp_y = 1; + break; + // case AOM_IMG_FMT_444A: + case AOM_IMG_FMT_I444: + use_high_bit_depth = 0; + chroma_subsamp_x = 0; + chroma_subsamp_y = 0; + break; + case AOM_IMG_FMT_I44416: + use_high_bit_depth = 1; + chroma_subsamp_x = 0; + chroma_subsamp_y = 0; + break; + case AOM_IMG_FMT_I422: + use_high_bit_depth = 0; + chroma_subsamp_x = 1; + chroma_subsamp_y = 0; + break; + case AOM_IMG_FMT_I42216: + use_high_bit_depth = 1; + chroma_subsamp_x = 1; + chroma_subsamp_y = 0; + break; + default: // unknown input format + fprintf(stderr, "Film grain error: input format is not supported!"); + return -1; + } + + assert(params->bit_depth == src->bit_depth); + + dst->fmt = src->fmt; + dst->bit_depth = src->bit_depth; + + dst->r_w = src->r_w; + dst->r_h = src->r_h; + dst->d_w = src->d_w; + dst->d_h = src->d_h; + + dst->cp = src->cp; + dst->tc = src->tc; + dst->mc = src->mc; + + dst->monochrome = src->monochrome; + dst->csp = src->csp; + dst->range = src->range; + + dst->x_chroma_shift = src->x_chroma_shift; + dst->y_chroma_shift = src->y_chroma_shift; + + dst->temporal_id = src->temporal_id; + dst->spatial_id = src->spatial_id; + + width = src->d_w % 2 ? src->d_w + 1 : src->d_w; + height = src->d_h % 2 ? src->d_h + 1 : src->d_h; + + copy_rect(src->planes[AOM_PLANE_Y], src->stride[AOM_PLANE_Y], + dst->planes[AOM_PLANE_Y], dst->stride[AOM_PLANE_Y], src->d_w, + src->d_h, use_high_bit_depth); + // Note that dst is already assumed to be aligned to even. + extend_even(dst->planes[AOM_PLANE_Y], dst->stride[AOM_PLANE_Y], src->d_w, + src->d_h, use_high_bit_depth); + + if (!src->monochrome) { + copy_rect(src->planes[AOM_PLANE_U], src->stride[AOM_PLANE_U], + dst->planes[AOM_PLANE_U], dst->stride[AOM_PLANE_U], + width >> chroma_subsamp_x, height >> chroma_subsamp_y, + use_high_bit_depth); + + copy_rect(src->planes[AOM_PLANE_V], src->stride[AOM_PLANE_V], + dst->planes[AOM_PLANE_V], dst->stride[AOM_PLANE_V], + width >> chroma_subsamp_x, height >> chroma_subsamp_y, + use_high_bit_depth); + } + + luma = dst->planes[AOM_PLANE_Y]; + cb = dst->planes[AOM_PLANE_U]; + cr = dst->planes[AOM_PLANE_V]; + + // luma and chroma strides in samples + luma_stride = dst->stride[AOM_PLANE_Y] >> use_high_bit_depth; + chroma_stride = dst->stride[AOM_PLANE_U] >> use_high_bit_depth; + + return av1_add_film_grain_run( + params, luma, cb, cr, height, width, luma_stride, chroma_stride, + use_high_bit_depth, chroma_subsamp_y, chroma_subsamp_x, mc_identity); +} + +int av1_add_film_grain_run(const aom_film_grain_t *params, uint8_t *luma, + uint8_t *cb, uint8_t *cr, int height, int width, + int luma_stride, int chroma_stride, + int use_high_bit_depth, int chroma_subsamp_y, + int chroma_subsamp_x, int mc_identity) { + int **pred_pos_luma; + int **pred_pos_chroma; + int *luma_grain_block; + int *cb_grain_block; + int *cr_grain_block; + + int *y_line_buf; + int *cb_line_buf; + int *cr_line_buf; + + int *y_col_buf; + int *cb_col_buf; + int *cr_col_buf; + + random_register = params->random_seed; + + int left_pad = 3; + int right_pad = 3; // padding to offset for AR coefficients + int top_pad = 3; + int bottom_pad = 0; + + int ar_padding = 3; // maximum lag used for stabilization of AR coefficients + + luma_subblock_size_y = 32; + luma_subblock_size_x = 32; + + chroma_subblock_size_y = luma_subblock_size_y >> chroma_subsamp_y; + chroma_subblock_size_x = luma_subblock_size_x >> chroma_subsamp_x; + + // Initial padding is only needed for generation of + // film grain templates (to stabilize the AR process) + // Only a 64x64 luma and 32x32 chroma part of a template + // is used later for adding grain, padding can be discarded + + int luma_block_size_y = + top_pad + 2 * ar_padding + luma_subblock_size_y * 2 + bottom_pad; + int luma_block_size_x = left_pad + 2 * ar_padding + luma_subblock_size_x * 2 + + 2 * ar_padding + right_pad; + + int chroma_block_size_y = top_pad + (2 >> chroma_subsamp_y) * ar_padding + + chroma_subblock_size_y * 2 + bottom_pad; + int chroma_block_size_x = left_pad + (2 >> chroma_subsamp_x) * ar_padding + + chroma_subblock_size_x * 2 + + (2 >> chroma_subsamp_x) * ar_padding + right_pad; + + int luma_grain_stride = luma_block_size_x; + int chroma_grain_stride = chroma_block_size_x; + + int overlap = params->overlap_flag; + int bit_depth = params->bit_depth; + + grain_center = 128 << (bit_depth - 8); + grain_min = 0 - grain_center; + grain_max = (256 << (bit_depth - 8)) - 1 - grain_center; + + init_arrays(params, luma_stride, chroma_stride, &pred_pos_luma, + &pred_pos_chroma, &luma_grain_block, &cb_grain_block, + &cr_grain_block, &y_line_buf, &cb_line_buf, &cr_line_buf, + &y_col_buf, &cb_col_buf, &cr_col_buf, + luma_block_size_y * luma_block_size_x, + chroma_block_size_y * chroma_block_size_x, chroma_subsamp_y, + chroma_subsamp_x); + + if (generate_luma_grain_block(params, pred_pos_luma, luma_grain_block, + luma_block_size_y, luma_block_size_x, + luma_grain_stride, left_pad, top_pad, right_pad, + bottom_pad)) + return -1; + + if (generate_chroma_grain_blocks( + params, + // pred_pos_luma, + pred_pos_chroma, luma_grain_block, cb_grain_block, cr_grain_block, + luma_grain_stride, chroma_block_size_y, chroma_block_size_x, + chroma_grain_stride, left_pad, top_pad, right_pad, bottom_pad, + chroma_subsamp_y, chroma_subsamp_x)) + return -1; + + init_scaling_function(params->scaling_points_y, params->num_y_points, + scaling_lut_y); + + if (params->chroma_scaling_from_luma) { + memcpy(scaling_lut_cb, scaling_lut_y, sizeof(*scaling_lut_y) * 256); + memcpy(scaling_lut_cr, scaling_lut_y, sizeof(*scaling_lut_y) * 256); + } else { + init_scaling_function(params->scaling_points_cb, params->num_cb_points, + scaling_lut_cb); + init_scaling_function(params->scaling_points_cr, params->num_cr_points, + scaling_lut_cr); + } + for (int y = 0; y < height / 2; y += (luma_subblock_size_y >> 1)) { + init_random_generator(y * 2, params->random_seed); + + for (int x = 0; x < width / 2; x += (luma_subblock_size_x >> 1)) { + int offset_y = get_random_number(8); + int offset_x = (offset_y >> 4) & 15; + offset_y &= 15; + + int luma_offset_y = left_pad + 2 * ar_padding + (offset_y << 1); + int luma_offset_x = top_pad + 2 * ar_padding + (offset_x << 1); + + int chroma_offset_y = top_pad + (2 >> chroma_subsamp_y) * ar_padding + + offset_y * (2 >> chroma_subsamp_y); + int chroma_offset_x = left_pad + (2 >> chroma_subsamp_x) * ar_padding + + offset_x * (2 >> chroma_subsamp_x); + + if (overlap && x) { + ver_boundary_overlap( + y_col_buf, 2, + luma_grain_block + luma_offset_y * luma_grain_stride + + luma_offset_x, + luma_grain_stride, y_col_buf, 2, 2, + AOMMIN(luma_subblock_size_y + 2, height - (y << 1))); + + ver_boundary_overlap( + cb_col_buf, 2 >> chroma_subsamp_x, + cb_grain_block + chroma_offset_y * chroma_grain_stride + + chroma_offset_x, + chroma_grain_stride, cb_col_buf, 2 >> chroma_subsamp_x, + 2 >> chroma_subsamp_x, + AOMMIN(chroma_subblock_size_y + (2 >> chroma_subsamp_y), + (height - (y << 1)) >> chroma_subsamp_y)); + + ver_boundary_overlap( + cr_col_buf, 2 >> chroma_subsamp_x, + cr_grain_block + chroma_offset_y * chroma_grain_stride + + chroma_offset_x, + chroma_grain_stride, cr_col_buf, 2 >> chroma_subsamp_x, + 2 >> chroma_subsamp_x, + AOMMIN(chroma_subblock_size_y + (2 >> chroma_subsamp_y), + (height - (y << 1)) >> chroma_subsamp_y)); + + int i = y ? 1 : 0; + + if (use_high_bit_depth) { + add_noise_to_block_hbd( + params, + (uint16_t *)luma + ((y + i) << 1) * luma_stride + (x << 1), + (uint16_t *)cb + + ((y + i) << (1 - chroma_subsamp_y)) * chroma_stride + + (x << (1 - chroma_subsamp_x)), + (uint16_t *)cr + + ((y + i) << (1 - chroma_subsamp_y)) * chroma_stride + + (x << (1 - chroma_subsamp_x)), + luma_stride, chroma_stride, y_col_buf + i * 4, + cb_col_buf + i * (2 - chroma_subsamp_y) * (2 - chroma_subsamp_x), + cr_col_buf + i * (2 - chroma_subsamp_y) * (2 - chroma_subsamp_x), + 2, (2 - chroma_subsamp_x), + AOMMIN(luma_subblock_size_y >> 1, height / 2 - y) - i, 1, + bit_depth, chroma_subsamp_y, chroma_subsamp_x, mc_identity); + } else { + add_noise_to_block( + params, luma + ((y + i) << 1) * luma_stride + (x << 1), + cb + ((y + i) << (1 - chroma_subsamp_y)) * chroma_stride + + (x << (1 - chroma_subsamp_x)), + cr + ((y + i) << (1 - chroma_subsamp_y)) * chroma_stride + + (x << (1 - chroma_subsamp_x)), + luma_stride, chroma_stride, y_col_buf + i * 4, + cb_col_buf + i * (2 - chroma_subsamp_y) * (2 - chroma_subsamp_x), + cr_col_buf + i * (2 - chroma_subsamp_y) * (2 - chroma_subsamp_x), + 2, (2 - chroma_subsamp_x), + AOMMIN(luma_subblock_size_y >> 1, height / 2 - y) - i, 1, + bit_depth, chroma_subsamp_y, chroma_subsamp_x, mc_identity); + } + } + + if (overlap && y) { + if (x) { + hor_boundary_overlap(y_line_buf + (x << 1), luma_stride, y_col_buf, 2, + y_line_buf + (x << 1), luma_stride, 2, 2); + + hor_boundary_overlap(cb_line_buf + x * (2 >> chroma_subsamp_x), + chroma_stride, cb_col_buf, 2 >> chroma_subsamp_x, + cb_line_buf + x * (2 >> chroma_subsamp_x), + chroma_stride, 2 >> chroma_subsamp_x, + 2 >> chroma_subsamp_y); + + hor_boundary_overlap(cr_line_buf + x * (2 >> chroma_subsamp_x), + chroma_stride, cr_col_buf, 2 >> chroma_subsamp_x, + cr_line_buf + x * (2 >> chroma_subsamp_x), + chroma_stride, 2 >> chroma_subsamp_x, + 2 >> chroma_subsamp_y); + } + + hor_boundary_overlap( + y_line_buf + ((x ? x + 1 : 0) << 1), luma_stride, + luma_grain_block + luma_offset_y * luma_grain_stride + + luma_offset_x + (x ? 2 : 0), + luma_grain_stride, y_line_buf + ((x ? x + 1 : 0) << 1), luma_stride, + AOMMIN(luma_subblock_size_x - ((x ? 1 : 0) << 1), + width - ((x ? x + 1 : 0) << 1)), + 2); + + hor_boundary_overlap( + cb_line_buf + ((x ? x + 1 : 0) << (1 - chroma_subsamp_x)), + chroma_stride, + cb_grain_block + chroma_offset_y * chroma_grain_stride + + chroma_offset_x + ((x ? 1 : 0) << (1 - chroma_subsamp_x)), + chroma_grain_stride, + cb_line_buf + ((x ? x + 1 : 0) << (1 - chroma_subsamp_x)), + chroma_stride, + AOMMIN(chroma_subblock_size_x - + ((x ? 1 : 0) << (1 - chroma_subsamp_x)), + (width - ((x ? x + 1 : 0) << 1)) >> chroma_subsamp_x), + 2 >> chroma_subsamp_y); + + hor_boundary_overlap( + cr_line_buf + ((x ? x + 1 : 0) << (1 - chroma_subsamp_x)), + chroma_stride, + cr_grain_block + chroma_offset_y * chroma_grain_stride + + chroma_offset_x + ((x ? 1 : 0) << (1 - chroma_subsamp_x)), + chroma_grain_stride, + cr_line_buf + ((x ? x + 1 : 0) << (1 - chroma_subsamp_x)), + chroma_stride, + AOMMIN(chroma_subblock_size_x - + ((x ? 1 : 0) << (1 - chroma_subsamp_x)), + (width - ((x ? x + 1 : 0) << 1)) >> chroma_subsamp_x), + 2 >> chroma_subsamp_y); + + if (use_high_bit_depth) { + add_noise_to_block_hbd( + params, (uint16_t *)luma + (y << 1) * luma_stride + (x << 1), + (uint16_t *)cb + (y << (1 - chroma_subsamp_y)) * chroma_stride + + (x << ((1 - chroma_subsamp_x))), + (uint16_t *)cr + (y << (1 - chroma_subsamp_y)) * chroma_stride + + (x << ((1 - chroma_subsamp_x))), + luma_stride, chroma_stride, y_line_buf + (x << 1), + cb_line_buf + (x << (1 - chroma_subsamp_x)), + cr_line_buf + (x << (1 - chroma_subsamp_x)), luma_stride, + chroma_stride, 1, + AOMMIN(luma_subblock_size_x >> 1, width / 2 - x), bit_depth, + chroma_subsamp_y, chroma_subsamp_x, mc_identity); + } else { + add_noise_to_block( + params, luma + (y << 1) * luma_stride + (x << 1), + cb + (y << (1 - chroma_subsamp_y)) * chroma_stride + + (x << ((1 - chroma_subsamp_x))), + cr + (y << (1 - chroma_subsamp_y)) * chroma_stride + + (x << ((1 - chroma_subsamp_x))), + luma_stride, chroma_stride, y_line_buf + (x << 1), + cb_line_buf + (x << (1 - chroma_subsamp_x)), + cr_line_buf + (x << (1 - chroma_subsamp_x)), luma_stride, + chroma_stride, 1, + AOMMIN(luma_subblock_size_x >> 1, width / 2 - x), bit_depth, + chroma_subsamp_y, chroma_subsamp_x, mc_identity); + } + } + + int i = overlap && y ? 1 : 0; + int j = overlap && x ? 1 : 0; + + if (use_high_bit_depth) { + add_noise_to_block_hbd( + params, + (uint16_t *)luma + ((y + i) << 1) * luma_stride + ((x + j) << 1), + (uint16_t *)cb + + ((y + i) << (1 - chroma_subsamp_y)) * chroma_stride + + ((x + j) << (1 - chroma_subsamp_x)), + (uint16_t *)cr + + ((y + i) << (1 - chroma_subsamp_y)) * chroma_stride + + ((x + j) << (1 - chroma_subsamp_x)), + luma_stride, chroma_stride, + luma_grain_block + (luma_offset_y + (i << 1)) * luma_grain_stride + + luma_offset_x + (j << 1), + cb_grain_block + + (chroma_offset_y + (i << (1 - chroma_subsamp_y))) * + chroma_grain_stride + + chroma_offset_x + (j << (1 - chroma_subsamp_x)), + cr_grain_block + + (chroma_offset_y + (i << (1 - chroma_subsamp_y))) * + chroma_grain_stride + + chroma_offset_x + (j << (1 - chroma_subsamp_x)), + luma_grain_stride, chroma_grain_stride, + AOMMIN(luma_subblock_size_y >> 1, height / 2 - y) - i, + AOMMIN(luma_subblock_size_x >> 1, width / 2 - x) - j, bit_depth, + chroma_subsamp_y, chroma_subsamp_x, mc_identity); + } else { + add_noise_to_block( + params, luma + ((y + i) << 1) * luma_stride + ((x + j) << 1), + cb + ((y + i) << (1 - chroma_subsamp_y)) * chroma_stride + + ((x + j) << (1 - chroma_subsamp_x)), + cr + ((y + i) << (1 - chroma_subsamp_y)) * chroma_stride + + ((x + j) << (1 - chroma_subsamp_x)), + luma_stride, chroma_stride, + luma_grain_block + (luma_offset_y + (i << 1)) * luma_grain_stride + + luma_offset_x + (j << 1), + cb_grain_block + + (chroma_offset_y + (i << (1 - chroma_subsamp_y))) * + chroma_grain_stride + + chroma_offset_x + (j << (1 - chroma_subsamp_x)), + cr_grain_block + + (chroma_offset_y + (i << (1 - chroma_subsamp_y))) * + chroma_grain_stride + + chroma_offset_x + (j << (1 - chroma_subsamp_x)), + luma_grain_stride, chroma_grain_stride, + AOMMIN(luma_subblock_size_y >> 1, height / 2 - y) - i, + AOMMIN(luma_subblock_size_x >> 1, width / 2 - x) - j, bit_depth, + chroma_subsamp_y, chroma_subsamp_x, mc_identity); + } + + if (overlap) { + if (x) { + // Copy overlapped column bufer to line buffer + copy_area(y_col_buf + (luma_subblock_size_y << 1), 2, + y_line_buf + (x << 1), luma_stride, 2, 2); + + copy_area( + cb_col_buf + (chroma_subblock_size_y << (1 - chroma_subsamp_x)), + 2 >> chroma_subsamp_x, + cb_line_buf + (x << (1 - chroma_subsamp_x)), chroma_stride, + 2 >> chroma_subsamp_x, 2 >> chroma_subsamp_y); + + copy_area( + cr_col_buf + (chroma_subblock_size_y << (1 - chroma_subsamp_x)), + 2 >> chroma_subsamp_x, + cr_line_buf + (x << (1 - chroma_subsamp_x)), chroma_stride, + 2 >> chroma_subsamp_x, 2 >> chroma_subsamp_y); + } + + // Copy grain to the line buffer for overlap with a bottom block + copy_area( + luma_grain_block + + (luma_offset_y + luma_subblock_size_y) * luma_grain_stride + + luma_offset_x + ((x ? 2 : 0)), + luma_grain_stride, y_line_buf + ((x ? x + 1 : 0) << 1), luma_stride, + AOMMIN(luma_subblock_size_x, width - (x << 1)) - (x ? 2 : 0), 2); + + copy_area(cb_grain_block + + (chroma_offset_y + chroma_subblock_size_y) * + chroma_grain_stride + + chroma_offset_x + (x ? 2 >> chroma_subsamp_x : 0), + chroma_grain_stride, + cb_line_buf + ((x ? x + 1 : 0) << (1 - chroma_subsamp_x)), + chroma_stride, + AOMMIN(chroma_subblock_size_x, + ((width - (x << 1)) >> chroma_subsamp_x)) - + (x ? 2 >> chroma_subsamp_x : 0), + 2 >> chroma_subsamp_y); + + copy_area(cr_grain_block + + (chroma_offset_y + chroma_subblock_size_y) * + chroma_grain_stride + + chroma_offset_x + (x ? 2 >> chroma_subsamp_x : 0), + chroma_grain_stride, + cr_line_buf + ((x ? x + 1 : 0) << (1 - chroma_subsamp_x)), + chroma_stride, + AOMMIN(chroma_subblock_size_x, + ((width - (x << 1)) >> chroma_subsamp_x)) - + (x ? 2 >> chroma_subsamp_x : 0), + 2 >> chroma_subsamp_y); + + // Copy grain to the column buffer for overlap with the next block to + // the right + + copy_area(luma_grain_block + luma_offset_y * luma_grain_stride + + luma_offset_x + luma_subblock_size_x, + luma_grain_stride, y_col_buf, 2, 2, + AOMMIN(luma_subblock_size_y + 2, height - (y << 1))); + + copy_area(cb_grain_block + chroma_offset_y * chroma_grain_stride + + chroma_offset_x + chroma_subblock_size_x, + chroma_grain_stride, cb_col_buf, 2 >> chroma_subsamp_x, + 2 >> chroma_subsamp_x, + AOMMIN(chroma_subblock_size_y + (2 >> chroma_subsamp_y), + (height - (y << 1)) >> chroma_subsamp_y)); + + copy_area(cr_grain_block + chroma_offset_y * chroma_grain_stride + + chroma_offset_x + chroma_subblock_size_x, + chroma_grain_stride, cr_col_buf, 2 >> chroma_subsamp_x, + 2 >> chroma_subsamp_x, + AOMMIN(chroma_subblock_size_y + (2 >> chroma_subsamp_y), + (height - (y << 1)) >> chroma_subsamp_y)); + } + } + } + + dealloc_arrays(params, &pred_pos_luma, &pred_pos_chroma, &luma_grain_block, + &cb_grain_block, &cr_grain_block, &y_line_buf, &cb_line_buf, + &cr_line_buf, &y_col_buf, &cb_col_buf, &cr_col_buf); + return 0; +} diff --git a/third_party/aom/aom_dsp/grain_synthesis.h b/third_party/aom/aom_dsp/grain_synthesis.h new file mode 100644 index 0000000000..7aee6f6f44 --- /dev/null +++ b/third_party/aom/aom_dsp/grain_synthesis.h @@ -0,0 +1,122 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +/*!\file + * \brief Describes film grain parameters and film grain synthesis + * + */ +#ifndef AOM_AOM_DSP_GRAIN_SYNTHESIS_H_ +#define AOM_AOM_DSP_GRAIN_SYNTHESIS_H_ + +#ifdef __cplusplus +extern "C" { +#endif + +#include "aom_dsp/aom_dsp_common.h" +#include "aom/aom_image.h" + +/*!\brief Structure containing film grain synthesis parameters for a frame + * + * This structure contains input parameters for film grain synthesis + */ +typedef struct { + int apply_grain; + + int update_parameters; + + // 8 bit values + int scaling_points_y[14][2]; + int num_y_points; // value: 0..14 + + // 8 bit values + int scaling_points_cb[10][2]; + int num_cb_points; // value: 0..10 + + // 8 bit values + int scaling_points_cr[10][2]; + int num_cr_points; // value: 0..10 + + int scaling_shift; // values : 8..11 + + int ar_coeff_lag; // values: 0..3 + + // 8 bit values + int ar_coeffs_y[24]; + int ar_coeffs_cb[25]; + int ar_coeffs_cr[25]; + + // Shift value: AR coeffs range + // 6: [-2, 2) + // 7: [-1, 1) + // 8: [-0.5, 0.5) + // 9: [-0.25, 0.25) + int ar_coeff_shift; // values : 6..9 + + int cb_mult; // 8 bits + int cb_luma_mult; // 8 bits + int cb_offset; // 9 bits + + int cr_mult; // 8 bits + int cr_luma_mult; // 8 bits + int cr_offset; // 9 bits + + int overlap_flag; + + int clip_to_restricted_range; + + unsigned int bit_depth; // video bit depth + + int chroma_scaling_from_luma; + + int grain_scale_shift; + + uint16_t random_seed; +} aom_film_grain_t; + +/*!\brief Add film grain + * + * Add film grain to an image + * + * Returns 0 for success, -1 for failure + * + * \param[in] grain_params Grain parameters + * \param[in] luma luma plane + * \param[in] cb cb plane + * \param[in] cr cr plane + * \param[in] height luma plane height + * \param[in] width luma plane width + * \param[in] luma_stride luma plane stride + * \param[in] chroma_stride chroma plane stride + */ +int av1_add_film_grain_run(const aom_film_grain_t *grain_params, uint8_t *luma, + uint8_t *cb, uint8_t *cr, int height, int width, + int luma_stride, int chroma_stride, + int use_high_bit_depth, int chroma_subsamp_y, + int chroma_subsamp_x, int mc_identity); + +/*!\brief Add film grain + * + * Add film grain to an image + * + * Returns 0 for success, -1 for failure + * + * \param[in] grain_params Grain parameters + * \param[in] src Source image + * \param[out] dst Resulting image with grain + */ +int av1_add_film_grain(const aom_film_grain_t *grain_params, + const aom_image_t *src, aom_image_t *dst); + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif // AOM_AOM_DSP_GRAIN_SYNTHESIS_H_ diff --git a/third_party/aom/aom_dsp/grain_table.c b/third_party/aom/aom_dsp/grain_table.c new file mode 100644 index 0000000000..0d6a73f558 --- /dev/null +++ b/third_party/aom/aom_dsp/grain_table.c @@ -0,0 +1,333 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +/*!\file + * \brief This file has the implementation details of the grain table. + * + * The file format is an ascii representation for readability and + * editability. Array parameters are separated from the non-array + * parameters and prefixed with a few characters to make for easy + * localization with a parameter set. Each entry is prefixed with "E" + * and the other parameters are only specified if "update-parms" is + * non-zero. + * + * filmgrn1 + * E + * p ... + * sY ... + * sCb ... + * sCr ... + * cY .... + * cCb .... + * cCr .... + * E ... + */ +#include +#include +#include "aom_dsp/aom_dsp_common.h" +#include "aom_dsp/grain_table.h" +#include "aom_mem/aom_mem.h" + +static const char kFileMagic[8] = "filmgrn1"; + +static void grain_table_entry_read(FILE *file, + struct aom_internal_error_info *error_info, + aom_film_grain_table_entry_t *entry) { + aom_film_grain_t *pars = &entry->params; + int num_read = + fscanf(file, "E %" PRId64 " %" PRId64 " %d %hd %d\n", &entry->start_time, + &entry->end_time, &pars->apply_grain, &pars->random_seed, + &pars->update_parameters); + if (num_read == 0 && feof(file)) return; + if (num_read != 5) { + aom_internal_error(error_info, AOM_CODEC_ERROR, + "Unable to read entry header. Read %d != 5", num_read); + return; + } + if (pars->update_parameters) { + num_read = fscanf(file, "p %d %d %d %d %d %d %d %d %d %d %d %d\n", + &pars->ar_coeff_lag, &pars->ar_coeff_shift, + &pars->grain_scale_shift, &pars->scaling_shift, + &pars->chroma_scaling_from_luma, &pars->overlap_flag, + &pars->cb_mult, &pars->cb_luma_mult, &pars->cb_offset, + &pars->cr_mult, &pars->cr_luma_mult, &pars->cr_offset); + if (num_read != 12) { + aom_internal_error(error_info, AOM_CODEC_ERROR, + "Unable to read entry params. Read %d != 12", + num_read); + return; + } + if (!fscanf(file, "\tsY %d ", &pars->num_y_points)) { + aom_internal_error(error_info, AOM_CODEC_ERROR, + "Unable to read num y points"); + return; + } + for (int i = 0; i < pars->num_y_points; ++i) { + if (2 != fscanf(file, "%d %d", &pars->scaling_points_y[i][0], + &pars->scaling_points_y[i][1])) { + aom_internal_error(error_info, AOM_CODEC_ERROR, + "Unable to read y scaling points"); + return; + } + } + if (!fscanf(file, "\n\tsCb %d", &pars->num_cb_points)) { + aom_internal_error(error_info, AOM_CODEC_ERROR, + "Unable to read num cb points"); + return; + } + for (int i = 0; i < pars->num_cb_points; ++i) { + if (2 != fscanf(file, "%d %d", &pars->scaling_points_cb[i][0], + &pars->scaling_points_cb[i][1])) { + aom_internal_error(error_info, AOM_CODEC_ERROR, + "Unable to read cb scaling points"); + return; + } + } + if (!fscanf(file, "\n\tsCr %d", &pars->num_cr_points)) { + aom_internal_error(error_info, AOM_CODEC_ERROR, + "Unable to read num cr points"); + return; + } + for (int i = 0; i < pars->num_cr_points; ++i) { + if (2 != fscanf(file, "%d %d", &pars->scaling_points_cr[i][0], + &pars->scaling_points_cr[i][1])) { + aom_internal_error(error_info, AOM_CODEC_ERROR, + "Unable to read cr scaling points"); + return; + } + } + + fscanf(file, "\n\tcY"); + const int n = 2 * pars->ar_coeff_lag * (pars->ar_coeff_lag + 1); + for (int i = 0; i < n; ++i) { + if (1 != fscanf(file, "%d", &pars->ar_coeffs_y[i])) { + aom_internal_error(error_info, AOM_CODEC_ERROR, + "Unable to read Y coeffs"); + return; + } + } + fscanf(file, "\n\tcCb"); + for (int i = 0; i <= n; ++i) { + if (1 != fscanf(file, "%d", &pars->ar_coeffs_cb[i])) { + aom_internal_error(error_info, AOM_CODEC_ERROR, + "Unable to read Cb coeffs"); + return; + } + } + fscanf(file, "\n\tcCr"); + for (int i = 0; i <= n; ++i) { + if (1 != fscanf(file, "%d", &pars->ar_coeffs_cr[i])) { + aom_internal_error(error_info, AOM_CODEC_ERROR, + "Unable to read Cr coeffs"); + return; + } + } + fscanf(file, "\n"); + } +} + +void grain_table_entry_write(FILE *file, aom_film_grain_table_entry_t *entry) { + const aom_film_grain_t *pars = &entry->params; + fprintf(file, "E %" PRId64 " %" PRId64 " %d %d %d\n", entry->start_time, + entry->end_time, pars->apply_grain, pars->random_seed, + pars->update_parameters); + if (pars->update_parameters) { + fprintf(file, "\tp %d %d %d %d %d %d %d %d %d %d %d %d\n", + pars->ar_coeff_lag, pars->ar_coeff_shift, pars->grain_scale_shift, + pars->scaling_shift, pars->chroma_scaling_from_luma, + pars->overlap_flag, pars->cb_mult, pars->cb_luma_mult, + pars->cb_offset, pars->cr_mult, pars->cr_luma_mult, + pars->cr_offset); + fprintf(file, "\tsY %d ", pars->num_y_points); + for (int i = 0; i < pars->num_y_points; ++i) { + fprintf(file, " %d %d", pars->scaling_points_y[i][0], + pars->scaling_points_y[i][1]); + } + fprintf(file, "\n\tsCb %d", pars->num_cb_points); + for (int i = 0; i < pars->num_cb_points; ++i) { + fprintf(file, " %d %d", pars->scaling_points_cb[i][0], + pars->scaling_points_cb[i][1]); + } + fprintf(file, "\n\tsCr %d", pars->num_cr_points); + for (int i = 0; i < pars->num_cr_points; ++i) { + fprintf(file, " %d %d", pars->scaling_points_cr[i][0], + pars->scaling_points_cr[i][1]); + } + fprintf(file, "\n\tcY"); + const int n = 2 * pars->ar_coeff_lag * (pars->ar_coeff_lag + 1); + for (int i = 0; i < n; ++i) { + fprintf(file, " %d", pars->ar_coeffs_y[i]); + } + fprintf(file, "\n\tcCb"); + for (int i = 0; i <= n; ++i) { + fprintf(file, " %d", pars->ar_coeffs_cb[i]); + } + fprintf(file, "\n\tcCr"); + for (int i = 0; i <= n; ++i) { + fprintf(file, " %d", pars->ar_coeffs_cr[i]); + } + fprintf(file, "\n"); + } +} + +void aom_film_grain_table_append(aom_film_grain_table_t *t, int64_t time_stamp, + int64_t end_time, + const aom_film_grain_t *grain) { + if (!t->tail || memcmp(grain, &t->tail->params, sizeof(*grain))) { + aom_film_grain_table_entry_t *new_tail = aom_malloc(sizeof(*new_tail)); + memset(new_tail, 0, sizeof(*new_tail)); + if (t->tail) t->tail->next = new_tail; + if (!t->head) t->head = new_tail; + t->tail = new_tail; + + new_tail->start_time = time_stamp; + new_tail->end_time = end_time; + new_tail->params = *grain; + } else { + t->tail->end_time = AOMMAX(t->tail->end_time, end_time); + t->tail->start_time = AOMMIN(t->tail->start_time, time_stamp); + } +} + +int aom_film_grain_table_lookup(aom_film_grain_table_t *t, int64_t time_stamp, + int64_t end_time, int erase, + aom_film_grain_t *grain) { + aom_film_grain_table_entry_t *entry = t->head; + aom_film_grain_table_entry_t *prev_entry = 0; + int16_t random_seed = grain ? grain->random_seed : 0; + if (grain) memset(grain, 0, sizeof(*grain)); + + while (entry) { + aom_film_grain_table_entry_t *next = entry->next; + if (time_stamp >= entry->start_time && time_stamp < entry->end_time) { + if (grain) { + *grain = entry->params; + if (time_stamp != 0) grain->random_seed = random_seed; + } + if (!erase) return 1; + + const int64_t entry_end_time = entry->end_time; + if (time_stamp <= entry->start_time && end_time >= entry->end_time) { + if (t->tail == entry) t->tail = prev_entry; + if (prev_entry) { + prev_entry->next = entry->next; + } else { + t->head = entry->next; + } + aom_free(entry); + } else if (time_stamp <= entry->start_time && + end_time < entry->end_time) { + entry->start_time = end_time; + } else if (time_stamp > entry->start_time && + end_time >= entry->end_time) { + entry->end_time = time_stamp; + } else { + aom_film_grain_table_entry_t *new_entry = + aom_malloc(sizeof(*new_entry)); + new_entry->next = entry->next; + new_entry->start_time = end_time; + new_entry->end_time = entry->end_time; + new_entry->params = entry->params; + entry->next = new_entry; + entry->end_time = time_stamp; + if (t->tail == entry) t->tail = new_entry; + } + // If segments aren't aligned, delete from the beggining of subsequent + // segments + if (end_time > entry_end_time) { + aom_film_grain_table_lookup(t, entry->end_time, end_time, 1, 0); + } + return 1; + } + prev_entry = entry; + entry = next; + } + return 0; +} + +aom_codec_err_t aom_film_grain_table_read( + aom_film_grain_table_t *t, const char *filename, + struct aom_internal_error_info *error_info) { + FILE *file = fopen(filename, "rb"); + if (!file) { + aom_internal_error(error_info, AOM_CODEC_ERROR, "Unable to open %s", + filename); + return error_info->error_code; + } + error_info->error_code = AOM_CODEC_OK; + + // Read in one extra character as there should be white space after + // the header. + char magic[9]; + if (!fread(magic, 9, 1, file) || memcmp(magic, kFileMagic, 8)) { + aom_internal_error(error_info, AOM_CODEC_ERROR, + "Unable to read (or invalid) file magic"); + fclose(file); + return error_info->error_code; + } + + aom_film_grain_table_entry_t *prev_entry = 0; + while (!feof(file)) { + aom_film_grain_table_entry_t *entry = aom_malloc(sizeof(*entry)); + memset(entry, 0, sizeof(*entry)); + grain_table_entry_read(file, error_info, entry); + entry->next = 0; + + if (prev_entry) prev_entry->next = entry; + if (!t->head) t->head = entry; + t->tail = entry; + prev_entry = entry; + + if (error_info->error_code != AOM_CODEC_OK) break; + } + + fclose(file); + return error_info->error_code; +} + +aom_codec_err_t aom_film_grain_table_write( + const aom_film_grain_table_t *t, const char *filename, + struct aom_internal_error_info *error_info) { + error_info->error_code = AOM_CODEC_OK; + + FILE *file = fopen(filename, "wb"); + if (!file) { + aom_internal_error(error_info, AOM_CODEC_ERROR, "Unable to open file %s", + filename); + return error_info->error_code; + } + + if (!fwrite(kFileMagic, 8, 1, file)) { + aom_internal_error(error_info, AOM_CODEC_ERROR, + "Unable to write file magic"); + fclose(file); + return error_info->error_code; + } + + fprintf(file, "\n"); + aom_film_grain_table_entry_t *entry = t->head; + while (entry) { + grain_table_entry_write(file, entry); + entry = entry->next; + } + fclose(file); + return error_info->error_code; +} + +void aom_film_grain_table_free(aom_film_grain_table_t *t) { + aom_film_grain_table_entry_t *entry = t->head; + while (entry) { + aom_film_grain_table_entry_t *next = entry->next; + aom_free(entry); + entry = next; + } + memset(t, 0, sizeof(*t)); +} diff --git a/third_party/aom/aom_dsp/grain_table.h b/third_party/aom/aom_dsp/grain_table.h new file mode 100644 index 0000000000..a8ac50730e --- /dev/null +++ b/third_party/aom/aom_dsp/grain_table.h @@ -0,0 +1,102 @@ +/* + * Copyright (c) 2018, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +/*!\file + * \brief A table mapping from time to corresponding film grain parameters. + * + * In order to apply grain synthesis in the decoder, the film grain parameters + * need to be signalled in the encoder. The film grain parameters are time + * varying, and for two-pass encoding (and denoiser implementation flexibility) + * it is common to denoise the video and do parameter estimation before encoding + * the denoised video. + * + * The film grain table is used to provide this flexibility and is used as a + * parameter that is passed to the encoder. + * + * Further, if regraining is to be done in say a single pass mode, or in two + * pass within the encoder (before frames are added to the lookahead buffer), + * this data structure can be used to keep track of on-the-fly estimated grain + * parameters, that are then extracted from the table before the encoded frame + * is written. + */ +#ifndef AOM_AOM_DSP_GRAIN_TABLE_H_ +#define AOM_AOM_DSP_GRAIN_TABLE_H_ + +#ifdef __cplusplus +extern "C" { +#endif + +#include "aom_dsp/grain_synthesis.h" +#include "aom/internal/aom_codec_internal.h" + +typedef struct aom_film_grain_table_entry_t { + aom_film_grain_t params; + int64_t start_time; + int64_t end_time; + struct aom_film_grain_table_entry_t *next; +} aom_film_grain_table_entry_t; + +typedef struct { + aom_film_grain_table_entry_t *head; + aom_film_grain_table_entry_t *tail; +} aom_film_grain_table_t; + +/*!\brief Add a mapping from [time_stamp, end_time) to the given grain + * parameters + * + * \param[in/out] table The grain table + * \param[in] time_stamp The start time stamp + * \param[in] end_stamp The end time_stamp + * \param[in] grain The grain parameters + */ +void aom_film_grain_table_append(aom_film_grain_table_t *table, + int64_t time_stamp, int64_t end_time, + const aom_film_grain_t *grain); + +/*!\brief Look-up (and optionally erase) the grain parameters for the given time + * + * \param[in] table The grain table + * \param[in] time_stamp The start time stamp + * \param[in] end_stamp The end time_stamp + * \param[in] erase Whether the time segment can be deleted + * \param[out] grain The output grain parameters + */ +int aom_film_grain_table_lookup(aom_film_grain_table_t *t, int64_t time_stamp, + int64_t end_time, int erase, + aom_film_grain_t *grain); + +/*!\brief Reads the grain table from a file. + * + * \param[out] table The grain table + * \param[in] filename The file to read from + * \param[in] error_info Error info for tracking errors + */ +aom_codec_err_t aom_film_grain_table_read( + aom_film_grain_table_t *table, const char *filename, + struct aom_internal_error_info *error_info); + +/*!\brief Writes the grain table from a file. + * + * \param[out] table The grain table + * \param[in] filename The file to read from + * \param[in] error_info Error info for tracking errors + */ +aom_codec_err_t aom_film_grain_table_write( + const aom_film_grain_table_t *t, const char *filename, + struct aom_internal_error_info *error_info); + +void aom_film_grain_table_free(aom_film_grain_table_t *t); + +#ifdef __cplusplus +} +#endif + +#endif // AOM_AOM_DSP_GRAIN_TABLE_H_ diff --git a/third_party/aom/aom_dsp/intrapred.c b/third_party/aom/aom_dsp/intrapred.c new file mode 100644 index 0000000000..c6aa6b207b --- /dev/null +++ b/third_party/aom/aom_dsp/intrapred.c @@ -0,0 +1,792 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include +#include + +#include "config/aom_config.h" +#include "config/aom_dsp_rtcd.h" + +#include "aom_dsp/aom_dsp_common.h" +#include "aom_dsp/intrapred_common.h" +#include "aom_mem/aom_mem.h" +#include "aom_ports/bitops.h" + +static INLINE void v_predictor(uint8_t *dst, ptrdiff_t stride, int bw, int bh, + const uint8_t *above, const uint8_t *left) { + int r; + (void)left; + + for (r = 0; r < bh; r++) { + memcpy(dst, above, bw); + dst += stride; + } +} + +static INLINE void h_predictor(uint8_t *dst, ptrdiff_t stride, int bw, int bh, + const uint8_t *above, const uint8_t *left) { + int r; + (void)above; + + for (r = 0; r < bh; r++) { + memset(dst, left[r], bw); + dst += stride; + } +} + +static INLINE int abs_diff(int a, int b) { return (a > b) ? a - b : b - a; } + +static INLINE uint16_t paeth_predictor_single(uint16_t left, uint16_t top, + uint16_t top_left) { + const int base = top + left - top_left; + const int p_left = abs_diff(base, left); + const int p_top = abs_diff(base, top); + const int p_top_left = abs_diff(base, top_left); + + // Return nearest to base of left, top and top_left. + return (p_left <= p_top && p_left <= p_top_left) + ? left + : (p_top <= p_top_left) ? top : top_left; +} + +static INLINE void paeth_predictor(uint8_t *dst, ptrdiff_t stride, int bw, + int bh, const uint8_t *above, + const uint8_t *left) { + int r, c; + const uint8_t ytop_left = above[-1]; + + for (r = 0; r < bh; r++) { + for (c = 0; c < bw; c++) + dst[c] = (uint8_t)paeth_predictor_single(left[r], above[c], ytop_left); + dst += stride; + } +} + +// Some basic checks on weights for smooth predictor. +#define sm_weights_sanity_checks(weights_w, weights_h, weights_scale, \ + pred_scale) \ + assert(weights_w[0] < weights_scale); \ + assert(weights_h[0] < weights_scale); \ + assert(weights_scale - weights_w[bw - 1] < weights_scale); \ + assert(weights_scale - weights_h[bh - 1] < weights_scale); \ + assert(pred_scale < 31) // ensures no overflow when calculating predictor. + +#define divide_round(value, bits) (((value) + (1 << ((bits)-1))) >> (bits)) + +static INLINE void smooth_predictor(uint8_t *dst, ptrdiff_t stride, int bw, + int bh, const uint8_t *above, + const uint8_t *left) { + const uint8_t below_pred = left[bh - 1]; // estimated by bottom-left pixel + const uint8_t right_pred = above[bw - 1]; // estimated by top-right pixel + const uint8_t *const sm_weights_w = sm_weight_arrays + bw; + const uint8_t *const sm_weights_h = sm_weight_arrays + bh; + // scale = 2 * 2^sm_weight_log2_scale + const int log2_scale = 1 + sm_weight_log2_scale; + const uint16_t scale = (1 << sm_weight_log2_scale); + sm_weights_sanity_checks(sm_weights_w, sm_weights_h, scale, + log2_scale + sizeof(*dst)); + int r; + for (r = 0; r < bh; ++r) { + int c; + for (c = 0; c < bw; ++c) { + const uint8_t pixels[] = { above[c], below_pred, left[r], right_pred }; + const uint8_t weights[] = { sm_weights_h[r], scale - sm_weights_h[r], + sm_weights_w[c], scale - sm_weights_w[c] }; + uint32_t this_pred = 0; + int i; + assert(scale >= sm_weights_h[r] && scale >= sm_weights_w[c]); + for (i = 0; i < 4; ++i) { + this_pred += weights[i] * pixels[i]; + } + dst[c] = divide_round(this_pred, log2_scale); + } + dst += stride; + } +} + +static INLINE void smooth_v_predictor(uint8_t *dst, ptrdiff_t stride, int bw, + int bh, const uint8_t *above, + const uint8_t *left) { + const uint8_t below_pred = left[bh - 1]; // estimated by bottom-left pixel + const uint8_t *const sm_weights = sm_weight_arrays + bh; + // scale = 2^sm_weight_log2_scale + const int log2_scale = sm_weight_log2_scale; + const uint16_t scale = (1 << sm_weight_log2_scale); + sm_weights_sanity_checks(sm_weights, sm_weights, scale, + log2_scale + sizeof(*dst)); + + int r; + for (r = 0; r < bh; r++) { + int c; + for (c = 0; c < bw; ++c) { + const uint8_t pixels[] = { above[c], below_pred }; + const uint8_t weights[] = { sm_weights[r], scale - sm_weights[r] }; + uint32_t this_pred = 0; + assert(scale >= sm_weights[r]); + int i; + for (i = 0; i < 2; ++i) { + this_pred += weights[i] * pixels[i]; + } + dst[c] = divide_round(this_pred, log2_scale); + } + dst += stride; + } +} + +static INLINE void smooth_h_predictor(uint8_t *dst, ptrdiff_t stride, int bw, + int bh, const uint8_t *above, + const uint8_t *left) { + const uint8_t right_pred = above[bw - 1]; // estimated by top-right pixel + const uint8_t *const sm_weights = sm_weight_arrays + bw; + // scale = 2^sm_weight_log2_scale + const int log2_scale = sm_weight_log2_scale; + const uint16_t scale = (1 << sm_weight_log2_scale); + sm_weights_sanity_checks(sm_weights, sm_weights, scale, + log2_scale + sizeof(*dst)); + + int r; + for (r = 0; r < bh; r++) { + int c; + for (c = 0; c < bw; ++c) { + const uint8_t pixels[] = { left[r], right_pred }; + const uint8_t weights[] = { sm_weights[c], scale - sm_weights[c] }; + uint32_t this_pred = 0; + assert(scale >= sm_weights[c]); + int i; + for (i = 0; i < 2; ++i) { + this_pred += weights[i] * pixels[i]; + } + dst[c] = divide_round(this_pred, log2_scale); + } + dst += stride; + } +} + +static INLINE void dc_128_predictor(uint8_t *dst, ptrdiff_t stride, int bw, + int bh, const uint8_t *above, + const uint8_t *left) { + int r; + (void)above; + (void)left; + + for (r = 0; r < bh; r++) { + memset(dst, 128, bw); + dst += stride; + } +} + +static INLINE void dc_left_predictor(uint8_t *dst, ptrdiff_t stride, int bw, + int bh, const uint8_t *above, + const uint8_t *left) { + int i, r, expected_dc, sum = 0; + (void)above; + + for (i = 0; i < bh; i++) sum += left[i]; + expected_dc = (sum + (bh >> 1)) / bh; + + for (r = 0; r < bh; r++) { + memset(dst, expected_dc, bw); + dst += stride; + } +} + +static INLINE void dc_top_predictor(uint8_t *dst, ptrdiff_t stride, int bw, + int bh, const uint8_t *above, + const uint8_t *left) { + int i, r, expected_dc, sum = 0; + (void)left; + + for (i = 0; i < bw; i++) sum += above[i]; + expected_dc = (sum + (bw >> 1)) / bw; + + for (r = 0; r < bh; r++) { + memset(dst, expected_dc, bw); + dst += stride; + } +} + +static INLINE void dc_predictor(uint8_t *dst, ptrdiff_t stride, int bw, int bh, + const uint8_t *above, const uint8_t *left) { + int i, r, expected_dc, sum = 0; + const int count = bw + bh; + + for (i = 0; i < bw; i++) { + sum += above[i]; + } + for (i = 0; i < bh; i++) { + sum += left[i]; + } + + expected_dc = (sum + (count >> 1)) / count; + + for (r = 0; r < bh; r++) { + memset(dst, expected_dc, bw); + dst += stride; + } +} + +static INLINE int divide_using_multiply_shift(int num, int shift1, + int multiplier, int shift2) { + const int interm = num >> shift1; + return interm * multiplier >> shift2; +} + + // The constants (multiplier and shifts) for a given block size are obtained + // as follows: + // - Let sum_w_h = block width + block height. + // - Shift 'sum_w_h' right until we reach an odd number. Let the number of + // shifts for that block size be called 'shift1' (see the parameter in + // dc_predictor_rect() function), and let the odd number be 'd'. [d has only 2 + // possible values: d = 3 for a 1:2 rect block and d = 5 for a 1:4 rect + // block]. + // - Find multipliers for (i) dividing by 3, and (ii) dividing by 5, + // using the "Algorithm 1" in: + // http://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=1467632 + // by ensuring that m + n = 16 (in that algorithm). This ensures that our 2nd + // shift will be 16, regardless of the block size. + + // Note: For low bitdepth, assembly code may be optimized by using smaller + // constants for smaller block sizes, where the range of the 'sum' is + // restricted to fewer bits. + +#define DC_MULTIPLIER_1X2 0x5556 +#define DC_MULTIPLIER_1X4 0x3334 + +#define DC_SHIFT2 16 + +static INLINE void dc_predictor_rect(uint8_t *dst, ptrdiff_t stride, int bw, + int bh, const uint8_t *above, + const uint8_t *left, int shift1, + int multiplier) { + int sum = 0; + + for (int i = 0; i < bw; i++) { + sum += above[i]; + } + for (int i = 0; i < bh; i++) { + sum += left[i]; + } + + const int expected_dc = divide_using_multiply_shift( + sum + ((bw + bh) >> 1), shift1, multiplier, DC_SHIFT2); + assert(expected_dc < (1 << 8)); + + for (int r = 0; r < bh; r++) { + memset(dst, expected_dc, bw); + dst += stride; + } +} + +#undef DC_SHIFT2 + +void aom_dc_predictor_4x8_c(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + dc_predictor_rect(dst, stride, 4, 8, above, left, 2, DC_MULTIPLIER_1X2); +} + +void aom_dc_predictor_8x4_c(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + dc_predictor_rect(dst, stride, 8, 4, above, left, 2, DC_MULTIPLIER_1X2); +} + +void aom_dc_predictor_4x16_c(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + dc_predictor_rect(dst, stride, 4, 16, above, left, 2, DC_MULTIPLIER_1X4); +} + +void aom_dc_predictor_16x4_c(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + dc_predictor_rect(dst, stride, 16, 4, above, left, 2, DC_MULTIPLIER_1X4); +} + +void aom_dc_predictor_8x16_c(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + dc_predictor_rect(dst, stride, 8, 16, above, left, 3, DC_MULTIPLIER_1X2); +} + +void aom_dc_predictor_16x8_c(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + dc_predictor_rect(dst, stride, 16, 8, above, left, 3, DC_MULTIPLIER_1X2); +} + +void aom_dc_predictor_8x32_c(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + dc_predictor_rect(dst, stride, 8, 32, above, left, 3, DC_MULTIPLIER_1X4); +} + +void aom_dc_predictor_32x8_c(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + dc_predictor_rect(dst, stride, 32, 8, above, left, 3, DC_MULTIPLIER_1X4); +} + +void aom_dc_predictor_16x32_c(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + dc_predictor_rect(dst, stride, 16, 32, above, left, 4, DC_MULTIPLIER_1X2); +} + +void aom_dc_predictor_32x16_c(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + dc_predictor_rect(dst, stride, 32, 16, above, left, 4, DC_MULTIPLIER_1X2); +} + +void aom_dc_predictor_16x64_c(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + dc_predictor_rect(dst, stride, 16, 64, above, left, 4, DC_MULTIPLIER_1X4); +} + +void aom_dc_predictor_64x16_c(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + dc_predictor_rect(dst, stride, 64, 16, above, left, 4, DC_MULTIPLIER_1X4); +} + +void aom_dc_predictor_32x64_c(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + dc_predictor_rect(dst, stride, 32, 64, above, left, 5, DC_MULTIPLIER_1X2); +} + +void aom_dc_predictor_64x32_c(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + dc_predictor_rect(dst, stride, 64, 32, above, left, 5, DC_MULTIPLIER_1X2); +} + +#undef DC_MULTIPLIER_1X2 +#undef DC_MULTIPLIER_1X4 + +static INLINE void highbd_v_predictor(uint16_t *dst, ptrdiff_t stride, int bw, + int bh, const uint16_t *above, + const uint16_t *left, int bd) { + int r; + (void)left; + (void)bd; + for (r = 0; r < bh; r++) { + memcpy(dst, above, bw * sizeof(uint16_t)); + dst += stride; + } +} + +static INLINE void highbd_h_predictor(uint16_t *dst, ptrdiff_t stride, int bw, + int bh, const uint16_t *above, + const uint16_t *left, int bd) { + int r; + (void)above; + (void)bd; + for (r = 0; r < bh; r++) { + aom_memset16(dst, left[r], bw); + dst += stride; + } +} + +static INLINE void highbd_paeth_predictor(uint16_t *dst, ptrdiff_t stride, + int bw, int bh, const uint16_t *above, + const uint16_t *left, int bd) { + int r, c; + const uint16_t ytop_left = above[-1]; + (void)bd; + + for (r = 0; r < bh; r++) { + for (c = 0; c < bw; c++) + dst[c] = paeth_predictor_single(left[r], above[c], ytop_left); + dst += stride; + } +} + +static INLINE void highbd_smooth_predictor(uint16_t *dst, ptrdiff_t stride, + int bw, int bh, + const uint16_t *above, + const uint16_t *left, int bd) { + (void)bd; + const uint16_t below_pred = left[bh - 1]; // estimated by bottom-left pixel + const uint16_t right_pred = above[bw - 1]; // estimated by top-right pixel + const uint8_t *const sm_weights_w = sm_weight_arrays + bw; + const uint8_t *const sm_weights_h = sm_weight_arrays + bh; + // scale = 2 * 2^sm_weight_log2_scale + const int log2_scale = 1 + sm_weight_log2_scale; + const uint16_t scale = (1 << sm_weight_log2_scale); + sm_weights_sanity_checks(sm_weights_w, sm_weights_h, scale, + log2_scale + sizeof(*dst)); + int r; + for (r = 0; r < bh; ++r) { + int c; + for (c = 0; c < bw; ++c) { + const uint16_t pixels[] = { above[c], below_pred, left[r], right_pred }; + const uint8_t weights[] = { sm_weights_h[r], scale - sm_weights_h[r], + sm_weights_w[c], scale - sm_weights_w[c] }; + uint32_t this_pred = 0; + int i; + assert(scale >= sm_weights_h[r] && scale >= sm_weights_w[c]); + for (i = 0; i < 4; ++i) { + this_pred += weights[i] * pixels[i]; + } + dst[c] = divide_round(this_pred, log2_scale); + } + dst += stride; + } +} + +static INLINE void highbd_smooth_v_predictor(uint16_t *dst, ptrdiff_t stride, + int bw, int bh, + const uint16_t *above, + const uint16_t *left, int bd) { + (void)bd; + const uint16_t below_pred = left[bh - 1]; // estimated by bottom-left pixel + const uint8_t *const sm_weights = sm_weight_arrays + bh; + // scale = 2^sm_weight_log2_scale + const int log2_scale = sm_weight_log2_scale; + const uint16_t scale = (1 << sm_weight_log2_scale); + sm_weights_sanity_checks(sm_weights, sm_weights, scale, + log2_scale + sizeof(*dst)); + + int r; + for (r = 0; r < bh; r++) { + int c; + for (c = 0; c < bw; ++c) { + const uint16_t pixels[] = { above[c], below_pred }; + const uint8_t weights[] = { sm_weights[r], scale - sm_weights[r] }; + uint32_t this_pred = 0; + assert(scale >= sm_weights[r]); + int i; + for (i = 0; i < 2; ++i) { + this_pred += weights[i] * pixels[i]; + } + dst[c] = divide_round(this_pred, log2_scale); + } + dst += stride; + } +} + +static INLINE void highbd_smooth_h_predictor(uint16_t *dst, ptrdiff_t stride, + int bw, int bh, + const uint16_t *above, + const uint16_t *left, int bd) { + (void)bd; + const uint16_t right_pred = above[bw - 1]; // estimated by top-right pixel + const uint8_t *const sm_weights = sm_weight_arrays + bw; + // scale = 2^sm_weight_log2_scale + const int log2_scale = sm_weight_log2_scale; + const uint16_t scale = (1 << sm_weight_log2_scale); + sm_weights_sanity_checks(sm_weights, sm_weights, scale, + log2_scale + sizeof(*dst)); + + int r; + for (r = 0; r < bh; r++) { + int c; + for (c = 0; c < bw; ++c) { + const uint16_t pixels[] = { left[r], right_pred }; + const uint8_t weights[] = { sm_weights[c], scale - sm_weights[c] }; + uint32_t this_pred = 0; + assert(scale >= sm_weights[c]); + int i; + for (i = 0; i < 2; ++i) { + this_pred += weights[i] * pixels[i]; + } + dst[c] = divide_round(this_pred, log2_scale); + } + dst += stride; + } +} + +static INLINE void highbd_dc_128_predictor(uint16_t *dst, ptrdiff_t stride, + int bw, int bh, + const uint16_t *above, + const uint16_t *left, int bd) { + int r; + (void)above; + (void)left; + + for (r = 0; r < bh; r++) { + aom_memset16(dst, 128 << (bd - 8), bw); + dst += stride; + } +} + +static INLINE void highbd_dc_left_predictor(uint16_t *dst, ptrdiff_t stride, + int bw, int bh, + const uint16_t *above, + const uint16_t *left, int bd) { + int i, r, expected_dc, sum = 0; + (void)above; + (void)bd; + + for (i = 0; i < bh; i++) sum += left[i]; + expected_dc = (sum + (bh >> 1)) / bh; + + for (r = 0; r < bh; r++) { + aom_memset16(dst, expected_dc, bw); + dst += stride; + } +} + +static INLINE void highbd_dc_top_predictor(uint16_t *dst, ptrdiff_t stride, + int bw, int bh, + const uint16_t *above, + const uint16_t *left, int bd) { + int i, r, expected_dc, sum = 0; + (void)left; + (void)bd; + + for (i = 0; i < bw; i++) sum += above[i]; + expected_dc = (sum + (bw >> 1)) / bw; + + for (r = 0; r < bh; r++) { + aom_memset16(dst, expected_dc, bw); + dst += stride; + } +} + +static INLINE void highbd_dc_predictor(uint16_t *dst, ptrdiff_t stride, int bw, + int bh, const uint16_t *above, + const uint16_t *left, int bd) { + int i, r, expected_dc, sum = 0; + const int count = bw + bh; + (void)bd; + + for (i = 0; i < bw; i++) { + sum += above[i]; + } + for (i = 0; i < bh; i++) { + sum += left[i]; + } + + expected_dc = (sum + (count >> 1)) / count; + + for (r = 0; r < bh; r++) { + aom_memset16(dst, expected_dc, bw); + dst += stride; + } +} + +// Obtained similarly as DC_MULTIPLIER_1X2 and DC_MULTIPLIER_1X4 above, but +// assume 2nd shift of 17 bits instead of 16. +// Note: Strictly speaking, 2nd shift needs to be 17 only when: +// - bit depth == 12, and +// - bw + bh is divisible by 5 (as opposed to divisible by 3). +// All other cases can use half the multipliers with a shift of 16 instead. +// This special optimization can be used when writing assembly code. +#define HIGHBD_DC_MULTIPLIER_1X2 0xAAAB +// Note: This constant is odd, but a smaller even constant (0x199a) with the +// appropriate shift should work for neon in 8/10-bit. +#define HIGHBD_DC_MULTIPLIER_1X4 0x6667 + +#define HIGHBD_DC_SHIFT2 17 + +static INLINE void highbd_dc_predictor_rect(uint16_t *dst, ptrdiff_t stride, + int bw, int bh, + const uint16_t *above, + const uint16_t *left, int bd, + int shift1, uint32_t multiplier) { + int sum = 0; + (void)bd; + + for (int i = 0; i < bw; i++) { + sum += above[i]; + } + for (int i = 0; i < bh; i++) { + sum += left[i]; + } + + const int expected_dc = divide_using_multiply_shift( + sum + ((bw + bh) >> 1), shift1, multiplier, HIGHBD_DC_SHIFT2); + assert(expected_dc < (1 << bd)); + + for (int r = 0; r < bh; r++) { + aom_memset16(dst, expected_dc, bw); + dst += stride; + } +} + +#undef HIGHBD_DC_SHIFT2 + +void aom_highbd_dc_predictor_4x8_c(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, const uint16_t *left, + int bd) { + highbd_dc_predictor_rect(dst, stride, 4, 8, above, left, bd, 2, + HIGHBD_DC_MULTIPLIER_1X2); +} + +void aom_highbd_dc_predictor_8x4_c(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, const uint16_t *left, + int bd) { + highbd_dc_predictor_rect(dst, stride, 8, 4, above, left, bd, 2, + HIGHBD_DC_MULTIPLIER_1X2); +} + +void aom_highbd_dc_predictor_4x16_c(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, const uint16_t *left, + int bd) { + highbd_dc_predictor_rect(dst, stride, 4, 16, above, left, bd, 2, + HIGHBD_DC_MULTIPLIER_1X4); +} + +void aom_highbd_dc_predictor_16x4_c(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, const uint16_t *left, + int bd) { + highbd_dc_predictor_rect(dst, stride, 16, 4, above, left, bd, 2, + HIGHBD_DC_MULTIPLIER_1X4); +} + +void aom_highbd_dc_predictor_8x16_c(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, const uint16_t *left, + int bd) { + highbd_dc_predictor_rect(dst, stride, 8, 16, above, left, bd, 3, + HIGHBD_DC_MULTIPLIER_1X2); +} + +void aom_highbd_dc_predictor_16x8_c(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, const uint16_t *left, + int bd) { + highbd_dc_predictor_rect(dst, stride, 16, 8, above, left, bd, 3, + HIGHBD_DC_MULTIPLIER_1X2); +} + +void aom_highbd_dc_predictor_8x32_c(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, const uint16_t *left, + int bd) { + highbd_dc_predictor_rect(dst, stride, 8, 32, above, left, bd, 3, + HIGHBD_DC_MULTIPLIER_1X4); +} + +void aom_highbd_dc_predictor_32x8_c(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, const uint16_t *left, + int bd) { + highbd_dc_predictor_rect(dst, stride, 32, 8, above, left, bd, 3, + HIGHBD_DC_MULTIPLIER_1X4); +} + +void aom_highbd_dc_predictor_16x32_c(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, + const uint16_t *left, int bd) { + highbd_dc_predictor_rect(dst, stride, 16, 32, above, left, bd, 4, + HIGHBD_DC_MULTIPLIER_1X2); +} + +void aom_highbd_dc_predictor_32x16_c(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, + const uint16_t *left, int bd) { + highbd_dc_predictor_rect(dst, stride, 32, 16, above, left, bd, 4, + HIGHBD_DC_MULTIPLIER_1X2); +} + +void aom_highbd_dc_predictor_16x64_c(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, + const uint16_t *left, int bd) { + highbd_dc_predictor_rect(dst, stride, 16, 64, above, left, bd, 4, + HIGHBD_DC_MULTIPLIER_1X4); +} + +void aom_highbd_dc_predictor_64x16_c(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, + const uint16_t *left, int bd) { + highbd_dc_predictor_rect(dst, stride, 64, 16, above, left, bd, 4, + HIGHBD_DC_MULTIPLIER_1X4); +} + +void aom_highbd_dc_predictor_32x64_c(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, + const uint16_t *left, int bd) { + highbd_dc_predictor_rect(dst, stride, 32, 64, above, left, bd, 5, + HIGHBD_DC_MULTIPLIER_1X2); +} + +void aom_highbd_dc_predictor_64x32_c(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, + const uint16_t *left, int bd) { + highbd_dc_predictor_rect(dst, stride, 64, 32, above, left, bd, 5, + HIGHBD_DC_MULTIPLIER_1X2); +} + +#undef HIGHBD_DC_MULTIPLIER_1X2 +#undef HIGHBD_DC_MULTIPLIER_1X4 + +// This serves as a wrapper function, so that all the prediction functions +// can be unified and accessed as a pointer array. Note that the boundary +// above and left are not necessarily used all the time. +#define intra_pred_sized(type, width, height) \ + void aom_##type##_predictor_##width##x##height##_c( \ + uint8_t *dst, ptrdiff_t stride, const uint8_t *above, \ + const uint8_t *left) { \ + type##_predictor(dst, stride, width, height, above, left); \ + } + +#define intra_pred_highbd_sized(type, width, height) \ + void aom_highbd_##type##_predictor_##width##x##height##_c( \ + uint16_t *dst, ptrdiff_t stride, const uint16_t *above, \ + const uint16_t *left, int bd) { \ + highbd_##type##_predictor(dst, stride, width, height, above, left, bd); \ + } + +/* clang-format off */ +#define intra_pred_rectangular(type) \ + intra_pred_sized(type, 4, 8) \ + intra_pred_sized(type, 8, 4) \ + intra_pred_sized(type, 8, 16) \ + intra_pred_sized(type, 16, 8) \ + intra_pred_sized(type, 16, 32) \ + intra_pred_sized(type, 32, 16) \ + intra_pred_sized(type, 32, 64) \ + intra_pred_sized(type, 64, 32) \ + intra_pred_sized(type, 4, 16) \ + intra_pred_sized(type, 16, 4) \ + intra_pred_sized(type, 8, 32) \ + intra_pred_sized(type, 32, 8) \ + intra_pred_sized(type, 16, 64) \ + intra_pred_sized(type, 64, 16) \ + intra_pred_highbd_sized(type, 4, 8) \ + intra_pred_highbd_sized(type, 8, 4) \ + intra_pred_highbd_sized(type, 8, 16) \ + intra_pred_highbd_sized(type, 16, 8) \ + intra_pred_highbd_sized(type, 16, 32) \ + intra_pred_highbd_sized(type, 32, 16) \ + intra_pred_highbd_sized(type, 32, 64) \ + intra_pred_highbd_sized(type, 64, 32) \ + intra_pred_highbd_sized(type, 4, 16) \ + intra_pred_highbd_sized(type, 16, 4) \ + intra_pred_highbd_sized(type, 8, 32) \ + intra_pred_highbd_sized(type, 32, 8) \ + intra_pred_highbd_sized(type, 16, 64) \ + intra_pred_highbd_sized(type, 64, 16) +#define intra_pred_above_4x4(type) \ + intra_pred_sized(type, 8, 8) \ + intra_pred_sized(type, 16, 16) \ + intra_pred_sized(type, 32, 32) \ + intra_pred_sized(type, 64, 64) \ + intra_pred_highbd_sized(type, 4, 4) \ + intra_pred_highbd_sized(type, 8, 8) \ + intra_pred_highbd_sized(type, 16, 16) \ + intra_pred_highbd_sized(type, 32, 32) \ + intra_pred_highbd_sized(type, 64, 64) \ + intra_pred_rectangular(type) +#define intra_pred_allsizes(type) \ + intra_pred_sized(type, 4, 4) \ + intra_pred_above_4x4(type) +#define intra_pred_square(type) \ + intra_pred_sized(type, 4, 4) \ + intra_pred_sized(type, 8, 8) \ + intra_pred_sized(type, 16, 16) \ + intra_pred_sized(type, 32, 32) \ + intra_pred_sized(type, 64, 64) \ + intra_pred_highbd_sized(type, 4, 4) \ + intra_pred_highbd_sized(type, 8, 8) \ + intra_pred_highbd_sized(type, 16, 16) \ + intra_pred_highbd_sized(type, 32, 32) \ + intra_pred_highbd_sized(type, 64, 64) + +intra_pred_allsizes(v) +intra_pred_allsizes(h) +intra_pred_allsizes(smooth) +intra_pred_allsizes(smooth_v) +intra_pred_allsizes(smooth_h) +intra_pred_allsizes(paeth) +intra_pred_allsizes(dc_128) +intra_pred_allsizes(dc_left) +intra_pred_allsizes(dc_top) +intra_pred_square(dc) +/* clang-format on */ +#undef intra_pred_allsizes diff --git a/third_party/aom/aom_dsp/intrapred_common.h b/third_party/aom/aom_dsp/intrapred_common.h new file mode 100644 index 0000000000..3ec62a86ef --- /dev/null +++ b/third_party/aom/aom_dsp/intrapred_common.h @@ -0,0 +1,47 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_INTRAPRED_COMMON_H_ +#define AOM_AOM_DSP_INTRAPRED_COMMON_H_ + +#include "config/aom_config.h" + +// Weights are quadratic from '1' to '1 / block_size', scaled by +// 2^sm_weight_log2_scale. +static const int sm_weight_log2_scale = 8; + +// max(block_size_wide[BLOCK_LARGEST], block_size_high[BLOCK_LARGEST]) +#define MAX_BLOCK_DIM 64 + +/* clang-format off */ +static const uint8_t sm_weight_arrays[2 * MAX_BLOCK_DIM] = { + // Unused, because we always offset by bs, which is at least 2. + 0, 0, + // bs = 2 + 255, 128, + // bs = 4 + 255, 149, 85, 64, + // bs = 8 + 255, 197, 146, 105, 73, 50, 37, 32, + // bs = 16 + 255, 225, 196, 170, 145, 123, 102, 84, 68, 54, 43, 33, 26, 20, 17, 16, + // bs = 32 + 255, 240, 225, 210, 196, 182, 169, 157, 145, 133, 122, 111, 101, 92, 83, 74, + 66, 59, 52, 45, 39, 34, 29, 25, 21, 17, 14, 12, 10, 9, 8, 8, + // bs = 64 + 255, 248, 240, 233, 225, 218, 210, 203, 196, 189, 182, 176, 169, 163, 156, + 150, 144, 138, 133, 127, 121, 116, 111, 106, 101, 96, 91, 86, 82, 77, 73, 69, + 65, 61, 57, 54, 50, 47, 44, 41, 38, 35, 32, 29, 27, 25, 22, 20, 18, 16, 15, + 13, 12, 10, 9, 8, 7, 6, 6, 5, 5, 4, 4, 4, +}; +/* clang-format on */ + +#endif // AOM_AOM_DSP_INTRAPRED_COMMON_H_ diff --git a/third_party/aom/aom_dsp/loopfilter.c b/third_party/aom/aom_dsp/loopfilter.c new file mode 100644 index 0000000000..a3f261824f --- /dev/null +++ b/third_party/aom/aom_dsp/loopfilter.c @@ -0,0 +1,925 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include + +#include "config/aom_config.h" +#include "config/aom_dsp_rtcd.h" + +#include "aom_dsp/aom_dsp_common.h" +#include "aom_ports/mem.h" + +static INLINE int8_t signed_char_clamp(int t) { + return (int8_t)clamp(t, -128, 127); +} + +static INLINE int16_t signed_char_clamp_high(int t, int bd) { + switch (bd) { + case 10: return (int16_t)clamp(t, -128 * 4, 128 * 4 - 1); + case 12: return (int16_t)clamp(t, -128 * 16, 128 * 16 - 1); + case 8: + default: return (int16_t)clamp(t, -128, 128 - 1); + } +} + +// should we apply any filter at all: 11111111 yes, 00000000 no +static INLINE int8_t filter_mask2(uint8_t limit, uint8_t blimit, uint8_t p1, + uint8_t p0, uint8_t q0, uint8_t q1) { + int8_t mask = 0; + mask |= (abs(p1 - p0) > limit) * -1; + mask |= (abs(q1 - q0) > limit) * -1; + mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit) * -1; + return ~mask; +} + +static INLINE int8_t filter_mask(uint8_t limit, uint8_t blimit, uint8_t p3, + uint8_t p2, uint8_t p1, uint8_t p0, uint8_t q0, + uint8_t q1, uint8_t q2, uint8_t q3) { + int8_t mask = 0; + mask |= (abs(p3 - p2) > limit) * -1; + mask |= (abs(p2 - p1) > limit) * -1; + mask |= (abs(p1 - p0) > limit) * -1; + mask |= (abs(q1 - q0) > limit) * -1; + mask |= (abs(q2 - q1) > limit) * -1; + mask |= (abs(q3 - q2) > limit) * -1; + mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit) * -1; + return ~mask; +} + +static INLINE int8_t filter_mask3_chroma(uint8_t limit, uint8_t blimit, + uint8_t p2, uint8_t p1, uint8_t p0, + uint8_t q0, uint8_t q1, uint8_t q2) { + int8_t mask = 0; + mask |= (abs(p2 - p1) > limit) * -1; + mask |= (abs(p1 - p0) > limit) * -1; + mask |= (abs(q1 - q0) > limit) * -1; + mask |= (abs(q2 - q1) > limit) * -1; + mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit) * -1; + return ~mask; +} + +static INLINE int8_t flat_mask3_chroma(uint8_t thresh, uint8_t p2, uint8_t p1, + uint8_t p0, uint8_t q0, uint8_t q1, + uint8_t q2) { + int8_t mask = 0; + mask |= (abs(p1 - p0) > thresh) * -1; + mask |= (abs(q1 - q0) > thresh) * -1; + mask |= (abs(p2 - p0) > thresh) * -1; + mask |= (abs(q2 - q0) > thresh) * -1; + return ~mask; +} + +static INLINE int8_t flat_mask4(uint8_t thresh, uint8_t p3, uint8_t p2, + uint8_t p1, uint8_t p0, uint8_t q0, uint8_t q1, + uint8_t q2, uint8_t q3) { + int8_t mask = 0; + mask |= (abs(p1 - p0) > thresh) * -1; + mask |= (abs(q1 - q0) > thresh) * -1; + mask |= (abs(p2 - p0) > thresh) * -1; + mask |= (abs(q2 - q0) > thresh) * -1; + mask |= (abs(p3 - p0) > thresh) * -1; + mask |= (abs(q3 - q0) > thresh) * -1; + return ~mask; +} + +// is there high edge variance internal edge: 11111111 yes, 00000000 no +static INLINE int8_t hev_mask(uint8_t thresh, uint8_t p1, uint8_t p0, + uint8_t q0, uint8_t q1) { + int8_t hev = 0; + hev |= (abs(p1 - p0) > thresh) * -1; + hev |= (abs(q1 - q0) > thresh) * -1; + return hev; +} + +static INLINE void filter4(int8_t mask, uint8_t thresh, uint8_t *op1, + uint8_t *op0, uint8_t *oq0, uint8_t *oq1) { + int8_t filter1, filter2; + + const int8_t ps1 = (int8_t)*op1 ^ 0x80; + const int8_t ps0 = (int8_t)*op0 ^ 0x80; + const int8_t qs0 = (int8_t)*oq0 ^ 0x80; + const int8_t qs1 = (int8_t)*oq1 ^ 0x80; + const uint8_t hev = hev_mask(thresh, *op1, *op0, *oq0, *oq1); + + // add outer taps if we have high edge variance + int8_t filter = signed_char_clamp(ps1 - qs1) & hev; + + // inner taps + filter = signed_char_clamp(filter + 3 * (qs0 - ps0)) & mask; + + // save bottom 3 bits so that we round one side +4 and the other +3 + // if it equals 4 we'll set to adjust by -1 to account for the fact + // we'd round 3 the other way + filter1 = signed_char_clamp(filter + 4) >> 3; + filter2 = signed_char_clamp(filter + 3) >> 3; + + *oq0 = signed_char_clamp(qs0 - filter1) ^ 0x80; + *op0 = signed_char_clamp(ps0 + filter2) ^ 0x80; + + // outer tap adjustments + filter = ROUND_POWER_OF_TWO(filter1, 1) & ~hev; + + *oq1 = signed_char_clamp(qs1 - filter) ^ 0x80; + *op1 = signed_char_clamp(ps1 + filter) ^ 0x80; +} + +void aom_lpf_horizontal_4_c(uint8_t *s, int p /* pitch */, + const uint8_t *blimit, const uint8_t *limit, + const uint8_t *thresh) { + int i; + int count = 4; + + // loop filter designed to work using chars so that we can make maximum use + // of 8 bit simd instructions. + for (i = 0; i < count; ++i) { + const uint8_t p1 = s[-2 * p], p0 = s[-p]; + const uint8_t q0 = s[0 * p], q1 = s[1 * p]; + const int8_t mask = filter_mask2(*limit, *blimit, p1, p0, q0, q1); + filter4(mask, *thresh, s - 2 * p, s - 1 * p, s, s + 1 * p); + ++s; + } +} + +void aom_lpf_horizontal_4_dual_c(uint8_t *s, int p, const uint8_t *blimit0, + const uint8_t *limit0, const uint8_t *thresh0, + const uint8_t *blimit1, const uint8_t *limit1, + const uint8_t *thresh1) { + aom_lpf_horizontal_4_c(s, p, blimit0, limit0, thresh0); + aom_lpf_horizontal_4_c(s + 4, p, blimit1, limit1, thresh1); +} + +void aom_lpf_vertical_4_c(uint8_t *s, int pitch, const uint8_t *blimit, + const uint8_t *limit, const uint8_t *thresh) { + int i; + int count = 4; + + // loop filter designed to work using chars so that we can make maximum use + // of 8 bit simd instructions. + for (i = 0; i < count; ++i) { + const uint8_t p1 = s[-2], p0 = s[-1]; + const uint8_t q0 = s[0], q1 = s[1]; + const int8_t mask = filter_mask2(*limit, *blimit, p1, p0, q0, q1); + filter4(mask, *thresh, s - 2, s - 1, s, s + 1); + s += pitch; + } +} + +void aom_lpf_vertical_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, + const uint8_t *limit0, const uint8_t *thresh0, + const uint8_t *blimit1, const uint8_t *limit1, + const uint8_t *thresh1) { + aom_lpf_vertical_4_c(s, pitch, blimit0, limit0, thresh0); + aom_lpf_vertical_4_c(s + 4 * pitch, pitch, blimit1, limit1, thresh1); +} + +static INLINE void filter6(int8_t mask, uint8_t thresh, int8_t flat, + uint8_t *op2, uint8_t *op1, uint8_t *op0, + uint8_t *oq0, uint8_t *oq1, uint8_t *oq2) { + if (flat && mask) { + const uint8_t p2 = *op2, p1 = *op1, p0 = *op0; + const uint8_t q0 = *oq0, q1 = *oq1, q2 = *oq2; + + // 5-tap filter [1, 2, 2, 2, 1] + *op1 = ROUND_POWER_OF_TWO(p2 * 3 + p1 * 2 + p0 * 2 + q0, 3); + *op0 = ROUND_POWER_OF_TWO(p2 + p1 * 2 + p0 * 2 + q0 * 2 + q1, 3); + *oq0 = ROUND_POWER_OF_TWO(p1 + p0 * 2 + q0 * 2 + q1 * 2 + q2, 3); + *oq1 = ROUND_POWER_OF_TWO(p0 + q0 * 2 + q1 * 2 + q2 * 3, 3); + } else { + filter4(mask, thresh, op1, op0, oq0, oq1); + } +} + +static INLINE void filter8(int8_t mask, uint8_t thresh, int8_t flat, + uint8_t *op3, uint8_t *op2, uint8_t *op1, + uint8_t *op0, uint8_t *oq0, uint8_t *oq1, + uint8_t *oq2, uint8_t *oq3) { + if (flat && mask) { + const uint8_t p3 = *op3, p2 = *op2, p1 = *op1, p0 = *op0; + const uint8_t q0 = *oq0, q1 = *oq1, q2 = *oq2, q3 = *oq3; + + // 7-tap filter [1, 1, 1, 2, 1, 1, 1] + *op2 = ROUND_POWER_OF_TWO(p3 + p3 + p3 + 2 * p2 + p1 + p0 + q0, 3); + *op1 = ROUND_POWER_OF_TWO(p3 + p3 + p2 + 2 * p1 + p0 + q0 + q1, 3); + *op0 = ROUND_POWER_OF_TWO(p3 + p2 + p1 + 2 * p0 + q0 + q1 + q2, 3); + *oq0 = ROUND_POWER_OF_TWO(p2 + p1 + p0 + 2 * q0 + q1 + q2 + q3, 3); + *oq1 = ROUND_POWER_OF_TWO(p1 + p0 + q0 + 2 * q1 + q2 + q3 + q3, 3); + *oq2 = ROUND_POWER_OF_TWO(p0 + q0 + q1 + 2 * q2 + q3 + q3 + q3, 3); + } else { + filter4(mask, thresh, op1, op0, oq0, oq1); + } +} + +void aom_lpf_horizontal_6_c(uint8_t *s, int p, const uint8_t *blimit, + const uint8_t *limit, const uint8_t *thresh) { + int i; + int count = 4; + + // loop filter designed to work using chars so that we can make maximum use + // of 8 bit simd instructions. + for (i = 0; i < count; ++i) { + const uint8_t p2 = s[-3 * p], p1 = s[-2 * p], p0 = s[-p]; + const uint8_t q0 = s[0 * p], q1 = s[1 * p], q2 = s[2 * p]; + + const int8_t mask = + filter_mask3_chroma(*limit, *blimit, p2, p1, p0, q0, q1, q2); + const int8_t flat = flat_mask3_chroma(1, p2, p1, p0, q0, q1, q2); + filter6(mask, *thresh, flat, s - 3 * p, s - 2 * p, s - 1 * p, s, s + 1 * p, + s + 2 * p); + ++s; + } +} + +void aom_lpf_horizontal_6_dual_c(uint8_t *s, int p, const uint8_t *blimit0, + const uint8_t *limit0, const uint8_t *thresh0, + const uint8_t *blimit1, const uint8_t *limit1, + const uint8_t *thresh1) { + aom_lpf_horizontal_6_c(s, p, blimit0, limit0, thresh0); + aom_lpf_horizontal_6_c(s + 4, p, blimit1, limit1, thresh1); +} + +void aom_lpf_horizontal_8_c(uint8_t *s, int p, const uint8_t *blimit, + const uint8_t *limit, const uint8_t *thresh) { + int i; + int count = 4; + + // loop filter designed to work using chars so that we can make maximum use + // of 8 bit simd instructions. + for (i = 0; i < count; ++i) { + const uint8_t p3 = s[-4 * p], p2 = s[-3 * p], p1 = s[-2 * p], p0 = s[-p]; + const uint8_t q0 = s[0 * p], q1 = s[1 * p], q2 = s[2 * p], q3 = s[3 * p]; + + const int8_t mask = + filter_mask(*limit, *blimit, p3, p2, p1, p0, q0, q1, q2, q3); + const int8_t flat = flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3); + filter8(mask, *thresh, flat, s - 4 * p, s - 3 * p, s - 2 * p, s - 1 * p, s, + s + 1 * p, s + 2 * p, s + 3 * p); + ++s; + } +} + +void aom_lpf_horizontal_8_dual_c(uint8_t *s, int p, const uint8_t *blimit0, + const uint8_t *limit0, const uint8_t *thresh0, + const uint8_t *blimit1, const uint8_t *limit1, + const uint8_t *thresh1) { + aom_lpf_horizontal_8_c(s, p, blimit0, limit0, thresh0); + aom_lpf_horizontal_8_c(s + 4, p, blimit1, limit1, thresh1); +} + +void aom_lpf_vertical_6_c(uint8_t *s, int pitch, const uint8_t *blimit, + const uint8_t *limit, const uint8_t *thresh) { + int i; + int count = 4; + + for (i = 0; i < count; ++i) { + const uint8_t p2 = s[-3], p1 = s[-2], p0 = s[-1]; + const uint8_t q0 = s[0], q1 = s[1], q2 = s[2]; + const int8_t mask = + filter_mask3_chroma(*limit, *blimit, p2, p1, p0, q0, q1, q2); + const int8_t flat = flat_mask3_chroma(1, p2, p1, p0, q0, q1, q2); + filter6(mask, *thresh, flat, s - 3, s - 2, s - 1, s, s + 1, s + 2); + s += pitch; + } +} + +void aom_lpf_vertical_6_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, + const uint8_t *limit0, const uint8_t *thresh0, + const uint8_t *blimit1, const uint8_t *limit1, + const uint8_t *thresh1) { + aom_lpf_vertical_6_c(s, pitch, blimit0, limit0, thresh0); + aom_lpf_vertical_6_c(s + 4 * pitch, pitch, blimit1, limit1, thresh1); +} + +void aom_lpf_vertical_8_c(uint8_t *s, int pitch, const uint8_t *blimit, + const uint8_t *limit, const uint8_t *thresh) { + int i; + int count = 4; + + for (i = 0; i < count; ++i) { + const uint8_t p3 = s[-4], p2 = s[-3], p1 = s[-2], p0 = s[-1]; + const uint8_t q0 = s[0], q1 = s[1], q2 = s[2], q3 = s[3]; + const int8_t mask = + filter_mask(*limit, *blimit, p3, p2, p1, p0, q0, q1, q2, q3); + const int8_t flat = flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3); + filter8(mask, *thresh, flat, s - 4, s - 3, s - 2, s - 1, s, s + 1, s + 2, + s + 3); + s += pitch; + } +} + +void aom_lpf_vertical_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, + const uint8_t *limit0, const uint8_t *thresh0, + const uint8_t *blimit1, const uint8_t *limit1, + const uint8_t *thresh1) { + aom_lpf_vertical_8_c(s, pitch, blimit0, limit0, thresh0); + aom_lpf_vertical_8_c(s + 4 * pitch, pitch, blimit1, limit1, thresh1); +} + +static INLINE void filter14(int8_t mask, uint8_t thresh, int8_t flat, + int8_t flat2, uint8_t *op6, uint8_t *op5, + uint8_t *op4, uint8_t *op3, uint8_t *op2, + uint8_t *op1, uint8_t *op0, uint8_t *oq0, + uint8_t *oq1, uint8_t *oq2, uint8_t *oq3, + uint8_t *oq4, uint8_t *oq5, uint8_t *oq6) { + if (flat2 && flat && mask) { + const uint8_t p6 = *op6, p5 = *op5, p4 = *op4, p3 = *op3, p2 = *op2, + p1 = *op1, p0 = *op0; + const uint8_t q0 = *oq0, q1 = *oq1, q2 = *oq2, q3 = *oq3, q4 = *oq4, + q5 = *oq5, q6 = *oq6; + + // 13-tap filter [1, 1, 1, 1, 1, 2, 2, 2, 1, 1, 1, 1, 1] + *op5 = ROUND_POWER_OF_TWO(p6 * 7 + p5 * 2 + p4 * 2 + p3 + p2 + p1 + p0 + q0, + 4); + *op4 = ROUND_POWER_OF_TWO( + p6 * 5 + p5 * 2 + p4 * 2 + p3 * 2 + p2 + p1 + p0 + q0 + q1, 4); + *op3 = ROUND_POWER_OF_TWO( + p6 * 4 + p5 + p4 * 2 + p3 * 2 + p2 * 2 + p1 + p0 + q0 + q1 + q2, 4); + *op2 = ROUND_POWER_OF_TWO( + p6 * 3 + p5 + p4 + p3 * 2 + p2 * 2 + p1 * 2 + p0 + q0 + q1 + q2 + q3, + 4); + *op1 = ROUND_POWER_OF_TWO(p6 * 2 + p5 + p4 + p3 + p2 * 2 + p1 * 2 + p0 * 2 + + q0 + q1 + q2 + q3 + q4, + 4); + *op0 = ROUND_POWER_OF_TWO(p6 + p5 + p4 + p3 + p2 + p1 * 2 + p0 * 2 + + q0 * 2 + q1 + q2 + q3 + q4 + q5, + 4); + *oq0 = ROUND_POWER_OF_TWO(p5 + p4 + p3 + p2 + p1 + p0 * 2 + q0 * 2 + + q1 * 2 + q2 + q3 + q4 + q5 + q6, + 4); + *oq1 = ROUND_POWER_OF_TWO(p4 + p3 + p2 + p1 + p0 + q0 * 2 + q1 * 2 + + q2 * 2 + q3 + q4 + q5 + q6 * 2, + 4); + *oq2 = ROUND_POWER_OF_TWO( + p3 + p2 + p1 + p0 + q0 + q1 * 2 + q2 * 2 + q3 * 2 + q4 + q5 + q6 * 3, + 4); + *oq3 = ROUND_POWER_OF_TWO( + p2 + p1 + p0 + q0 + q1 + q2 * 2 + q3 * 2 + q4 * 2 + q5 + q6 * 4, 4); + *oq4 = ROUND_POWER_OF_TWO( + p1 + p0 + q0 + q1 + q2 + q3 * 2 + q4 * 2 + q5 * 2 + q6 * 5, 4); + *oq5 = ROUND_POWER_OF_TWO(p0 + q0 + q1 + q2 + q3 + q4 * 2 + q5 * 2 + q6 * 7, + 4); + } else { + filter8(mask, thresh, flat, op3, op2, op1, op0, oq0, oq1, oq2, oq3); + } +} + +static void mb_lpf_horizontal_edge_w(uint8_t *s, int p, const uint8_t *blimit, + const uint8_t *limit, + const uint8_t *thresh, int count) { + int i; + int step = 4; + + // loop filter designed to work using chars so that we can make maximum use + // of 8 bit simd instructions. + for (i = 0; i < step * count; ++i) { + const uint8_t p6 = s[-7 * p], p5 = s[-6 * p], p4 = s[-5 * p], + p3 = s[-4 * p], p2 = s[-3 * p], p1 = s[-2 * p], p0 = s[-p]; + const uint8_t q0 = s[0 * p], q1 = s[1 * p], q2 = s[2 * p], q3 = s[3 * p], + q4 = s[4 * p], q5 = s[5 * p], q6 = s[6 * p]; + const int8_t mask = + filter_mask(*limit, *blimit, p3, p2, p1, p0, q0, q1, q2, q3); + const int8_t flat = flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3); + const int8_t flat2 = flat_mask4(1, p6, p5, p4, p0, q0, q4, q5, q6); + + filter14(mask, *thresh, flat, flat2, s - 7 * p, s - 6 * p, s - 5 * p, + s - 4 * p, s - 3 * p, s - 2 * p, s - 1 * p, s, s + 1 * p, + s + 2 * p, s + 3 * p, s + 4 * p, s + 5 * p, s + 6 * p); + ++s; + } +} + +void aom_lpf_horizontal_14_c(uint8_t *s, int p, const uint8_t *blimit, + const uint8_t *limit, const uint8_t *thresh) { + mb_lpf_horizontal_edge_w(s, p, blimit, limit, thresh, 1); +} + +void aom_lpf_horizontal_14_dual_c(uint8_t *s, int p, const uint8_t *blimit0, + const uint8_t *limit0, const uint8_t *thresh0, + const uint8_t *blimit1, const uint8_t *limit1, + const uint8_t *thresh1) { + mb_lpf_horizontal_edge_w(s, p, blimit0, limit0, thresh0, 1); + mb_lpf_horizontal_edge_w(s + 4, p, blimit1, limit1, thresh1, 1); +} + +static void mb_lpf_vertical_edge_w(uint8_t *s, int p, const uint8_t *blimit, + const uint8_t *limit, const uint8_t *thresh, + int count) { + int i; + + for (i = 0; i < count; ++i) { + const uint8_t p6 = s[-7], p5 = s[-6], p4 = s[-5], p3 = s[-4], p2 = s[-3], + p1 = s[-2], p0 = s[-1]; + const uint8_t q0 = s[0], q1 = s[1], q2 = s[2], q3 = s[3], q4 = s[4], + q5 = s[5], q6 = s[6]; + const int8_t mask = + filter_mask(*limit, *blimit, p3, p2, p1, p0, q0, q1, q2, q3); + const int8_t flat = flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3); + const int8_t flat2 = flat_mask4(1, p6, p5, p4, p0, q0, q4, q5, q6); + + filter14(mask, *thresh, flat, flat2, s - 7, s - 6, s - 5, s - 4, s - 3, + s - 2, s - 1, s, s + 1, s + 2, s + 3, s + 4, s + 5, s + 6); + s += p; + } +} + +void aom_lpf_vertical_14_c(uint8_t *s, int p, const uint8_t *blimit, + const uint8_t *limit, const uint8_t *thresh) { + mb_lpf_vertical_edge_w(s, p, blimit, limit, thresh, 4); +} + +void aom_lpf_vertical_14_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, + const uint8_t *limit0, const uint8_t *thresh0, + const uint8_t *blimit1, const uint8_t *limit1, + const uint8_t *thresh1) { + mb_lpf_vertical_edge_w(s, pitch, blimit0, limit0, thresh0, 4); + mb_lpf_vertical_edge_w(s + 4 * pitch, pitch, blimit1, limit1, thresh1, 4); +} + +// Should we apply any filter at all: 11111111 yes, 00000000 no ? +static INLINE int8_t highbd_filter_mask2(uint8_t limit, uint8_t blimit, + uint16_t p1, uint16_t p0, uint16_t q0, + uint16_t q1, int bd) { + int8_t mask = 0; + int16_t limit16 = (uint16_t)limit << (bd - 8); + int16_t blimit16 = (uint16_t)blimit << (bd - 8); + mask |= (abs(p1 - p0) > limit16) * -1; + mask |= (abs(q1 - q0) > limit16) * -1; + mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit16) * -1; + return ~mask; +} + +// Should we apply any filter at all: 11111111 yes, 00000000 no ? +static INLINE int8_t highbd_filter_mask(uint8_t limit, uint8_t blimit, + uint16_t p3, uint16_t p2, uint16_t p1, + uint16_t p0, uint16_t q0, uint16_t q1, + uint16_t q2, uint16_t q3, int bd) { + int8_t mask = 0; + int16_t limit16 = (uint16_t)limit << (bd - 8); + int16_t blimit16 = (uint16_t)blimit << (bd - 8); + mask |= (abs(p3 - p2) > limit16) * -1; + mask |= (abs(p2 - p1) > limit16) * -1; + mask |= (abs(p1 - p0) > limit16) * -1; + mask |= (abs(q1 - q0) > limit16) * -1; + mask |= (abs(q2 - q1) > limit16) * -1; + mask |= (abs(q3 - q2) > limit16) * -1; + mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit16) * -1; + return ~mask; +} + +static INLINE int8_t highbd_filter_mask3_chroma(uint8_t limit, uint8_t blimit, + uint16_t p2, uint16_t p1, + uint16_t p0, uint16_t q0, + uint16_t q1, uint16_t q2, + int bd) { + int8_t mask = 0; + int16_t limit16 = (uint16_t)limit << (bd - 8); + int16_t blimit16 = (uint16_t)blimit << (bd - 8); + mask |= (abs(p2 - p1) > limit16) * -1; + mask |= (abs(p1 - p0) > limit16) * -1; + mask |= (abs(q1 - q0) > limit16) * -1; + mask |= (abs(q2 - q1) > limit16) * -1; + mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit16) * -1; + return ~mask; +} + +static INLINE int8_t highbd_flat_mask3_chroma(uint8_t thresh, uint16_t p2, + uint16_t p1, uint16_t p0, + uint16_t q0, uint16_t q1, + uint16_t q2, int bd) { + int8_t mask = 0; + int16_t thresh16 = (uint16_t)thresh << (bd - 8); + mask |= (abs(p1 - p0) > thresh16) * -1; + mask |= (abs(q1 - q0) > thresh16) * -1; + mask |= (abs(p2 - p0) > thresh16) * -1; + mask |= (abs(q2 - q0) > thresh16) * -1; + return ~mask; +} + +static INLINE int8_t highbd_flat_mask4(uint8_t thresh, uint16_t p3, uint16_t p2, + uint16_t p1, uint16_t p0, uint16_t q0, + uint16_t q1, uint16_t q2, uint16_t q3, + int bd) { + int8_t mask = 0; + int16_t thresh16 = (uint16_t)thresh << (bd - 8); + mask |= (abs(p1 - p0) > thresh16) * -1; + mask |= (abs(q1 - q0) > thresh16) * -1; + mask |= (abs(p2 - p0) > thresh16) * -1; + mask |= (abs(q2 - q0) > thresh16) * -1; + mask |= (abs(p3 - p0) > thresh16) * -1; + mask |= (abs(q3 - q0) > thresh16) * -1; + return ~mask; +} + +// Is there high edge variance internal edge: +// 11111111_11111111 yes, 00000000_00000000 no ? +static INLINE int16_t highbd_hev_mask(uint8_t thresh, uint16_t p1, uint16_t p0, + uint16_t q0, uint16_t q1, int bd) { + int16_t hev = 0; + int16_t thresh16 = (uint16_t)thresh << (bd - 8); + hev |= (abs(p1 - p0) > thresh16) * -1; + hev |= (abs(q1 - q0) > thresh16) * -1; + return hev; +} + +static INLINE void highbd_filter4(int8_t mask, uint8_t thresh, uint16_t *op1, + uint16_t *op0, uint16_t *oq0, uint16_t *oq1, + int bd) { + int16_t filter1, filter2; + // ^0x80 equivalent to subtracting 0x80 from the values to turn them + // into -128 to +127 instead of 0 to 255. + int shift = bd - 8; + const int16_t ps1 = (int16_t)*op1 - (0x80 << shift); + const int16_t ps0 = (int16_t)*op0 - (0x80 << shift); + const int16_t qs0 = (int16_t)*oq0 - (0x80 << shift); + const int16_t qs1 = (int16_t)*oq1 - (0x80 << shift); + const uint16_t hev = highbd_hev_mask(thresh, *op1, *op0, *oq0, *oq1, bd); + + // Add outer taps if we have high edge variance. + int16_t filter = signed_char_clamp_high(ps1 - qs1, bd) & hev; + + // Inner taps. + filter = signed_char_clamp_high(filter + 3 * (qs0 - ps0), bd) & mask; + + // Save bottom 3 bits so that we round one side +4 and the other +3 + // if it equals 4 we'll set to adjust by -1 to account for the fact + // we'd round 3 the other way. + filter1 = signed_char_clamp_high(filter + 4, bd) >> 3; + filter2 = signed_char_clamp_high(filter + 3, bd) >> 3; + + *oq0 = signed_char_clamp_high(qs0 - filter1, bd) + (0x80 << shift); + *op0 = signed_char_clamp_high(ps0 + filter2, bd) + (0x80 << shift); + + // Outer tap adjustments. + filter = ROUND_POWER_OF_TWO(filter1, 1) & ~hev; + + *oq1 = signed_char_clamp_high(qs1 - filter, bd) + (0x80 << shift); + *op1 = signed_char_clamp_high(ps1 + filter, bd) + (0x80 << shift); +} + +void aom_highbd_lpf_horizontal_4_c(uint16_t *s, int p /* pitch */, + const uint8_t *blimit, const uint8_t *limit, + const uint8_t *thresh, int bd) { + int i; + int count = 4; + + // loop filter designed to work using chars so that we can make maximum use + // of 8 bit simd instructions. + for (i = 0; i < count; ++i) { + const uint16_t p1 = s[-2 * p]; + const uint16_t p0 = s[-p]; + const uint16_t q0 = s[0 * p]; + const uint16_t q1 = s[1 * p]; + const int8_t mask = + highbd_filter_mask2(*limit, *blimit, p1, p0, q0, q1, bd); + highbd_filter4(mask, *thresh, s - 2 * p, s - 1 * p, s, s + 1 * p, bd); + ++s; + } +} + +void aom_highbd_lpf_horizontal_4_dual_c( + uint16_t *s, int p, const uint8_t *blimit0, const uint8_t *limit0, + const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, + const uint8_t *thresh1, int bd) { + aom_highbd_lpf_horizontal_4_c(s, p, blimit0, limit0, thresh0, bd); + aom_highbd_lpf_horizontal_4_c(s + 4, p, blimit1, limit1, thresh1, bd); +} + +void aom_highbd_lpf_vertical_4_c(uint16_t *s, int pitch, const uint8_t *blimit, + const uint8_t *limit, const uint8_t *thresh, + int bd) { + int i; + int count = 4; + + // loop filter designed to work using chars so that we can make maximum use + // of 8 bit simd instructions. + for (i = 0; i < count; ++i) { + const uint16_t p1 = s[-2], p0 = s[-1]; + const uint16_t q0 = s[0], q1 = s[1]; + const int8_t mask = + highbd_filter_mask2(*limit, *blimit, p1, p0, q0, q1, bd); + highbd_filter4(mask, *thresh, s - 2, s - 1, s, s + 1, bd); + s += pitch; + } +} + +void aom_highbd_lpf_vertical_4_dual_c( + uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, + const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, + const uint8_t *thresh1, int bd) { + aom_highbd_lpf_vertical_4_c(s, pitch, blimit0, limit0, thresh0, bd); + aom_highbd_lpf_vertical_4_c(s + 4 * pitch, pitch, blimit1, limit1, thresh1, + bd); +} + +static INLINE void highbd_filter6(int8_t mask, uint8_t thresh, int8_t flat, + uint16_t *op2, uint16_t *op1, uint16_t *op0, + uint16_t *oq0, uint16_t *oq1, uint16_t *oq2, + int bd) { + if (flat && mask) { + const uint16_t p2 = *op2, p1 = *op1, p0 = *op0; + const uint16_t q0 = *oq0, q1 = *oq1, q2 = *oq2; + + // 5-tap filter [1, 2, 2, 2, 1] + *op1 = ROUND_POWER_OF_TWO(p2 * 3 + p1 * 2 + p0 * 2 + q0, 3); + *op0 = ROUND_POWER_OF_TWO(p2 + p1 * 2 + p0 * 2 + q0 * 2 + q1, 3); + *oq0 = ROUND_POWER_OF_TWO(p1 + p0 * 2 + q0 * 2 + q1 * 2 + q2, 3); + *oq1 = ROUND_POWER_OF_TWO(p0 + q0 * 2 + q1 * 2 + q2 * 3, 3); + } else { + highbd_filter4(mask, thresh, op1, op0, oq0, oq1, bd); + } +} + +static INLINE void highbd_filter8(int8_t mask, uint8_t thresh, int8_t flat, + uint16_t *op3, uint16_t *op2, uint16_t *op1, + uint16_t *op0, uint16_t *oq0, uint16_t *oq1, + uint16_t *oq2, uint16_t *oq3, int bd) { + if (flat && mask) { + const uint16_t p3 = *op3, p2 = *op2, p1 = *op1, p0 = *op0; + const uint16_t q0 = *oq0, q1 = *oq1, q2 = *oq2, q3 = *oq3; + + // 7-tap filter [1, 1, 1, 2, 1, 1, 1] + *op2 = ROUND_POWER_OF_TWO(p3 + p3 + p3 + 2 * p2 + p1 + p0 + q0, 3); + *op1 = ROUND_POWER_OF_TWO(p3 + p3 + p2 + 2 * p1 + p0 + q0 + q1, 3); + *op0 = ROUND_POWER_OF_TWO(p3 + p2 + p1 + 2 * p0 + q0 + q1 + q2, 3); + *oq0 = ROUND_POWER_OF_TWO(p2 + p1 + p0 + 2 * q0 + q1 + q2 + q3, 3); + *oq1 = ROUND_POWER_OF_TWO(p1 + p0 + q0 + 2 * q1 + q2 + q3 + q3, 3); + *oq2 = ROUND_POWER_OF_TWO(p0 + q0 + q1 + 2 * q2 + q3 + q3 + q3, 3); + } else { + highbd_filter4(mask, thresh, op1, op0, oq0, oq1, bd); + } +} + +void aom_highbd_lpf_horizontal_8_c(uint16_t *s, int p, const uint8_t *blimit, + const uint8_t *limit, const uint8_t *thresh, + int bd) { + int i; + int count = 4; + + // loop filter designed to work using chars so that we can make maximum use + // of 8 bit simd instructions. + for (i = 0; i < count; ++i) { + const uint16_t p3 = s[-4 * p], p2 = s[-3 * p], p1 = s[-2 * p], p0 = s[-p]; + const uint16_t q0 = s[0 * p], q1 = s[1 * p], q2 = s[2 * p], q3 = s[3 * p]; + + const int8_t mask = + highbd_filter_mask(*limit, *blimit, p3, p2, p1, p0, q0, q1, q2, q3, bd); + const int8_t flat = + highbd_flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3, bd); + highbd_filter8(mask, *thresh, flat, s - 4 * p, s - 3 * p, s - 2 * p, + s - 1 * p, s, s + 1 * p, s + 2 * p, s + 3 * p, bd); + ++s; + } +} + +void aom_highbd_lpf_horizontal_6_c(uint16_t *s, int p, const uint8_t *blimit, + const uint8_t *limit, const uint8_t *thresh, + int bd) { + int i; + int count = 4; + + // loop filter designed to work using chars so that we can make maximum use + // of 8 bit simd instructions. + for (i = 0; i < count; ++i) { + const uint16_t p2 = s[-3 * p], p1 = s[-2 * p], p0 = s[-p]; + const uint16_t q0 = s[0 * p], q1 = s[1 * p], q2 = s[2 * p]; + + const int8_t mask = + highbd_filter_mask3_chroma(*limit, *blimit, p2, p1, p0, q0, q1, q2, bd); + const int8_t flat = highbd_flat_mask3_chroma(1, p2, p1, p0, q0, q1, q2, bd); + highbd_filter6(mask, *thresh, flat, s - 3 * p, s - 2 * p, s - 1 * p, s, + s + 1 * p, s + 2 * p, bd); + ++s; + } +} + +void aom_highbd_lpf_horizontal_6_dual_c( + uint16_t *s, int p, const uint8_t *blimit0, const uint8_t *limit0, + const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, + const uint8_t *thresh1, int bd) { + aom_highbd_lpf_horizontal_6_c(s, p, blimit0, limit0, thresh0, bd); + aom_highbd_lpf_horizontal_6_c(s + 4, p, blimit1, limit1, thresh1, bd); +} + +void aom_highbd_lpf_horizontal_8_dual_c( + uint16_t *s, int p, const uint8_t *blimit0, const uint8_t *limit0, + const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, + const uint8_t *thresh1, int bd) { + aom_highbd_lpf_horizontal_8_c(s, p, blimit0, limit0, thresh0, bd); + aom_highbd_lpf_horizontal_8_c(s + 4, p, blimit1, limit1, thresh1, bd); +} + +void aom_highbd_lpf_vertical_6_c(uint16_t *s, int pitch, const uint8_t *blimit, + const uint8_t *limit, const uint8_t *thresh, + int bd) { + int i; + int count = 4; + + for (i = 0; i < count; ++i) { + const uint16_t p2 = s[-3], p1 = s[-2], p0 = s[-1]; + const uint16_t q0 = s[0], q1 = s[1], q2 = s[2]; + const int8_t mask = + highbd_filter_mask3_chroma(*limit, *blimit, p2, p1, p0, q0, q1, q2, bd); + const int8_t flat = highbd_flat_mask3_chroma(1, p2, p1, p0, q0, q1, q2, bd); + highbd_filter6(mask, *thresh, flat, s - 3, s - 2, s - 1, s, s + 1, s + 2, + bd); + s += pitch; + } +} + +void aom_highbd_lpf_vertical_6_dual_c( + uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, + const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, + const uint8_t *thresh1, int bd) { + aom_highbd_lpf_vertical_6_c(s, pitch, blimit0, limit0, thresh0, bd); + aom_highbd_lpf_vertical_6_c(s + 4 * pitch, pitch, blimit1, limit1, thresh1, + bd); +} + +void aom_highbd_lpf_vertical_8_c(uint16_t *s, int pitch, const uint8_t *blimit, + const uint8_t *limit, const uint8_t *thresh, + int bd) { + int i; + int count = 4; + + for (i = 0; i < count; ++i) { + const uint16_t p3 = s[-4], p2 = s[-3], p1 = s[-2], p0 = s[-1]; + const uint16_t q0 = s[0], q1 = s[1], q2 = s[2], q3 = s[3]; + const int8_t mask = + highbd_filter_mask(*limit, *blimit, p3, p2, p1, p0, q0, q1, q2, q3, bd); + const int8_t flat = + highbd_flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3, bd); + highbd_filter8(mask, *thresh, flat, s - 4, s - 3, s - 2, s - 1, s, s + 1, + s + 2, s + 3, bd); + s += pitch; + } +} + +void aom_highbd_lpf_vertical_8_dual_c( + uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, + const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, + const uint8_t *thresh1, int bd) { + aom_highbd_lpf_vertical_8_c(s, pitch, blimit0, limit0, thresh0, bd); + aom_highbd_lpf_vertical_8_c(s + 4 * pitch, pitch, blimit1, limit1, thresh1, + bd); +} + +static INLINE void highbd_filter14(int8_t mask, uint8_t thresh, int8_t flat, + int8_t flat2, uint16_t *op6, uint16_t *op5, + uint16_t *op4, uint16_t *op3, uint16_t *op2, + uint16_t *op1, uint16_t *op0, uint16_t *oq0, + uint16_t *oq1, uint16_t *oq2, uint16_t *oq3, + uint16_t *oq4, uint16_t *oq5, uint16_t *oq6, + int bd) { + if (flat2 && flat && mask) { + const uint16_t p6 = *op6; + const uint16_t p5 = *op5; + const uint16_t p4 = *op4; + const uint16_t p3 = *op3; + const uint16_t p2 = *op2; + const uint16_t p1 = *op1; + const uint16_t p0 = *op0; + const uint16_t q0 = *oq0; + const uint16_t q1 = *oq1; + const uint16_t q2 = *oq2; + const uint16_t q3 = *oq3; + const uint16_t q4 = *oq4; + const uint16_t q5 = *oq5; + const uint16_t q6 = *oq6; + + // 13-tap filter [1, 1, 1, 1, 1, 2, 2, 2, 1, 1, 1, 1, 1] + *op5 = ROUND_POWER_OF_TWO(p6 * 7 + p5 * 2 + p4 * 2 + p3 + p2 + p1 + p0 + q0, + 4); + *op4 = ROUND_POWER_OF_TWO( + p6 * 5 + p5 * 2 + p4 * 2 + p3 * 2 + p2 + p1 + p0 + q0 + q1, 4); + *op3 = ROUND_POWER_OF_TWO( + p6 * 4 + p5 + p4 * 2 + p3 * 2 + p2 * 2 + p1 + p0 + q0 + q1 + q2, 4); + *op2 = ROUND_POWER_OF_TWO( + p6 * 3 + p5 + p4 + p3 * 2 + p2 * 2 + p1 * 2 + p0 + q0 + q1 + q2 + q3, + 4); + *op1 = ROUND_POWER_OF_TWO(p6 * 2 + p5 + p4 + p3 + p2 * 2 + p1 * 2 + p0 * 2 + + q0 + q1 + q2 + q3 + q4, + 4); + *op0 = ROUND_POWER_OF_TWO(p6 + p5 + p4 + p3 + p2 + p1 * 2 + p0 * 2 + + q0 * 2 + q1 + q2 + q3 + q4 + q5, + 4); + *oq0 = ROUND_POWER_OF_TWO(p5 + p4 + p3 + p2 + p1 + p0 * 2 + q0 * 2 + + q1 * 2 + q2 + q3 + q4 + q5 + q6, + 4); + *oq1 = ROUND_POWER_OF_TWO(p4 + p3 + p2 + p1 + p0 + q0 * 2 + q1 * 2 + + q2 * 2 + q3 + q4 + q5 + q6 * 2, + 4); + *oq2 = ROUND_POWER_OF_TWO( + p3 + p2 + p1 + p0 + q0 + q1 * 2 + q2 * 2 + q3 * 2 + q4 + q5 + q6 * 3, + 4); + *oq3 = ROUND_POWER_OF_TWO( + p2 + p1 + p0 + q0 + q1 + q2 * 2 + q3 * 2 + q4 * 2 + q5 + q6 * 4, 4); + *oq4 = ROUND_POWER_OF_TWO( + p1 + p0 + q0 + q1 + q2 + q3 * 2 + q4 * 2 + q5 * 2 + q6 * 5, 4); + *oq5 = ROUND_POWER_OF_TWO(p0 + q0 + q1 + q2 + q3 + q4 * 2 + q5 * 2 + q6 * 7, + 4); + } else { + highbd_filter8(mask, thresh, flat, op3, op2, op1, op0, oq0, oq1, oq2, oq3, + bd); + } +} + +static void highbd_mb_lpf_horizontal_edge_w(uint16_t *s, int p, + const uint8_t *blimit, + const uint8_t *limit, + const uint8_t *thresh, int count, + int bd) { + int i; + int step = 4; + + // loop filter designed to work using chars so that we can make maximum use + // of 8 bit simd instructions. + for (i = 0; i < step * count; ++i) { + const uint16_t p3 = s[-4 * p]; + const uint16_t p2 = s[-3 * p]; + const uint16_t p1 = s[-2 * p]; + const uint16_t p0 = s[-p]; + const uint16_t q0 = s[0 * p]; + const uint16_t q1 = s[1 * p]; + const uint16_t q2 = s[2 * p]; + const uint16_t q3 = s[3 * p]; + const int8_t mask = + highbd_filter_mask(*limit, *blimit, p3, p2, p1, p0, q0, q1, q2, q3, bd); + const int8_t flat = + highbd_flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3, bd); + + const int8_t flat2 = + highbd_flat_mask4(1, s[-7 * p], s[-6 * p], s[-5 * p], p0, q0, s[4 * p], + s[5 * p], s[6 * p], bd); + + highbd_filter14(mask, *thresh, flat, flat2, s - 7 * p, s - 6 * p, s - 5 * p, + s - 4 * p, s - 3 * p, s - 2 * p, s - 1 * p, s, s + 1 * p, + s + 2 * p, s + 3 * p, s + 4 * p, s + 5 * p, s + 6 * p, bd); + ++s; + } +} + +void aom_highbd_lpf_horizontal_14_c(uint16_t *s, int p, const uint8_t *blimit, + const uint8_t *limit, const uint8_t *thresh, + int bd) { + highbd_mb_lpf_horizontal_edge_w(s, p, blimit, limit, thresh, 1, bd); +} + +void aom_highbd_lpf_horizontal_14_dual_c( + uint16_t *s, int p, const uint8_t *blimit0, const uint8_t *limit0, + const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, + const uint8_t *thresh1, int bd) { + highbd_mb_lpf_horizontal_edge_w(s, p, blimit0, limit0, thresh0, 1, bd); + highbd_mb_lpf_horizontal_edge_w(s + 4, p, blimit1, limit1, thresh1, 1, bd); +} + +static void highbd_mb_lpf_vertical_edge_w(uint16_t *s, int p, + const uint8_t *blimit, + const uint8_t *limit, + const uint8_t *thresh, int count, + int bd) { + int i; + + for (i = 0; i < count; ++i) { + const uint16_t p3 = s[-4]; + const uint16_t p2 = s[-3]; + const uint16_t p1 = s[-2]; + const uint16_t p0 = s[-1]; + const uint16_t q0 = s[0]; + const uint16_t q1 = s[1]; + const uint16_t q2 = s[2]; + const uint16_t q3 = s[3]; + const int8_t mask = + highbd_filter_mask(*limit, *blimit, p3, p2, p1, p0, q0, q1, q2, q3, bd); + const int8_t flat = + highbd_flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3, bd); + const int8_t flat2 = + highbd_flat_mask4(1, s[-7], s[-6], s[-5], p0, q0, s[4], s[5], s[6], bd); + + highbd_filter14(mask, *thresh, flat, flat2, s - 7, s - 6, s - 5, s - 4, + s - 3, s - 2, s - 1, s, s + 1, s + 2, s + 3, s + 4, s + 5, + s + 6, bd); + s += p; + } +} + +void aom_highbd_lpf_vertical_14_c(uint16_t *s, int p, const uint8_t *blimit, + const uint8_t *limit, const uint8_t *thresh, + int bd) { + highbd_mb_lpf_vertical_edge_w(s, p, blimit, limit, thresh, 4, bd); +} + +void aom_highbd_lpf_vertical_14_dual_c( + uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, + const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, + const uint8_t *thresh1, int bd) { + highbd_mb_lpf_vertical_edge_w(s, pitch, blimit0, limit0, thresh0, 4, bd); + highbd_mb_lpf_vertical_edge_w(s + 4 * pitch, pitch, blimit1, limit1, thresh1, + 4, bd); +} diff --git a/third_party/aom/aom_dsp/mips/add_noise_msa.c b/third_party/aom/aom_dsp/mips/add_noise_msa.c new file mode 100644 index 0000000000..96d04cff0c --- /dev/null +++ b/third_party/aom/aom_dsp/mips/add_noise_msa.c @@ -0,0 +1,61 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include + +#include "aom_dsp/mips/macros_msa.h" + +void aom_plane_add_noise_msa(uint8_t *start_ptr, char *noise, + char blackclamp[16], char whiteclamp[16], + char bothclamp[16], uint32_t width, + uint32_t height, int32_t pitch) { + uint32_t i, j; + + for (i = 0; i < height / 2; ++i) { + uint8_t *pos0_ptr = start_ptr + (2 * i) * pitch; + int8_t *ref0_ptr = (int8_t *)(noise + (rand() & 0xff)); + uint8_t *pos1_ptr = start_ptr + (2 * i + 1) * pitch; + int8_t *ref1_ptr = (int8_t *)(noise + (rand() & 0xff)); + for (j = width / 16; j--;) { + v16i8 temp00_s, temp01_s; + v16u8 temp00, temp01, black_clamp, white_clamp; + v16u8 pos0, ref0, pos1, ref1; + v16i8 const127 = __msa_ldi_b(127); + + pos0 = LD_UB(pos0_ptr); + ref0 = LD_UB(ref0_ptr); + pos1 = LD_UB(pos1_ptr); + ref1 = LD_UB(ref1_ptr); + black_clamp = (v16u8)__msa_fill_b(blackclamp[0]); + white_clamp = (v16u8)__msa_fill_b(whiteclamp[0]); + temp00 = (pos0 < black_clamp); + pos0 = __msa_bmnz_v(pos0, black_clamp, temp00); + temp01 = (pos1 < black_clamp); + pos1 = __msa_bmnz_v(pos1, black_clamp, temp01); + XORI_B2_128_UB(pos0, pos1); + temp00_s = __msa_adds_s_b((v16i8)white_clamp, const127); + temp00 = (v16u8)(temp00_s < pos0); + pos0 = (v16u8)__msa_bmnz_v((v16u8)pos0, (v16u8)temp00_s, temp00); + temp01_s = __msa_adds_s_b((v16i8)white_clamp, const127); + temp01 = (temp01_s < pos1); + pos1 = (v16u8)__msa_bmnz_v((v16u8)pos1, (v16u8)temp01_s, temp01); + XORI_B2_128_UB(pos0, pos1); + pos0 += ref0; + ST_UB(pos0, pos0_ptr); + pos1 += ref1; + ST_UB(pos1, pos1_ptr); + pos0_ptr += 16; + pos1_ptr += 16; + ref0_ptr += 16; + ref1_ptr += 16; + } + } +} diff --git a/third_party/aom/aom_dsp/mips/aom_convolve8_horiz_msa.c b/third_party/aom/aom_dsp/mips/aom_convolve8_horiz_msa.c new file mode 100644 index 0000000000..363fad308d --- /dev/null +++ b/third_party/aom/aom_dsp/mips/aom_convolve8_horiz_msa.c @@ -0,0 +1,694 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include + +#include "config/aom_dsp_rtcd.h" + +#include "aom_dsp/mips/aom_convolve_msa.h" + +static void common_hz_8t_4x4_msa(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter) { + v16u8 mask0, mask1, mask2, mask3, out; + v16i8 src0, src1, src2, src3, filt0, filt1, filt2, filt3; + v8i16 filt, out0, out1; + + mask0 = LD_UB(&mc_filt_mask_arr[16]); + src -= 3; + + /* rearranging filter */ + filt = LD_SH(filter); + SPLATI_H4_SB(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3); + + mask1 = mask0 + 2; + mask2 = mask0 + 4; + mask3 = mask0 + 6; + + LD_SB4(src, src_stride, src0, src1, src2, src3); + XORI_B4_128_SB(src0, src1, src2, src3); + HORIZ_8TAP_4WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2, mask3, + filt0, filt1, filt2, filt3, out0, out1); + SRARI_H2_SH(out0, out1, FILTER_BITS); + SAT_SH2_SH(out0, out1, 7); + out = PCKEV_XORI128_UB(out0, out1); + ST4x4_UB(out, out, 0, 1, 2, 3, dst, dst_stride); +} + +static void common_hz_8t_4x8_msa(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter) { + v16i8 filt0, filt1, filt2, filt3; + v16i8 src0, src1, src2, src3; + v16u8 mask0, mask1, mask2, mask3, out; + v8i16 filt, out0, out1, out2, out3; + + mask0 = LD_UB(&mc_filt_mask_arr[16]); + src -= 3; + + /* rearranging filter */ + filt = LD_SH(filter); + SPLATI_H4_SB(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3); + + mask1 = mask0 + 2; + mask2 = mask0 + 4; + mask3 = mask0 + 6; + + LD_SB4(src, src_stride, src0, src1, src2, src3); + XORI_B4_128_SB(src0, src1, src2, src3); + src += (4 * src_stride); + HORIZ_8TAP_4WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2, mask3, + filt0, filt1, filt2, filt3, out0, out1); + LD_SB4(src, src_stride, src0, src1, src2, src3); + XORI_B4_128_SB(src0, src1, src2, src3); + HORIZ_8TAP_4WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2, mask3, + filt0, filt1, filt2, filt3, out2, out3); + SRARI_H4_SH(out0, out1, out2, out3, FILTER_BITS); + SAT_SH4_SH(out0, out1, out2, out3, 7); + out = PCKEV_XORI128_UB(out0, out1); + ST4x4_UB(out, out, 0, 1, 2, 3, dst, dst_stride); + dst += (4 * dst_stride); + out = PCKEV_XORI128_UB(out2, out3); + ST4x4_UB(out, out, 0, 1, 2, 3, dst, dst_stride); +} + +static void common_hz_8t_4w_msa(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter, int32_t height) { + if (4 == height) { + common_hz_8t_4x4_msa(src, src_stride, dst, dst_stride, filter); + } else if (8 == height) { + common_hz_8t_4x8_msa(src, src_stride, dst, dst_stride, filter); + } +} + +static void common_hz_8t_8x4_msa(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter) { + v16i8 src0, src1, src2, src3, filt0, filt1, filt2, filt3; + v16u8 mask0, mask1, mask2, mask3, tmp0, tmp1; + v8i16 filt, out0, out1, out2, out3; + + mask0 = LD_UB(&mc_filt_mask_arr[0]); + src -= 3; + + /* rearranging filter */ + filt = LD_SH(filter); + SPLATI_H4_SB(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3); + + mask1 = mask0 + 2; + mask2 = mask0 + 4; + mask3 = mask0 + 6; + + LD_SB4(src, src_stride, src0, src1, src2, src3); + XORI_B4_128_SB(src0, src1, src2, src3); + HORIZ_8TAP_8WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2, mask3, + filt0, filt1, filt2, filt3, out0, out1, out2, + out3); + SRARI_H4_SH(out0, out1, out2, out3, FILTER_BITS); + SAT_SH4_SH(out0, out1, out2, out3, 7); + tmp0 = PCKEV_XORI128_UB(out0, out1); + tmp1 = PCKEV_XORI128_UB(out2, out3); + ST8x4_UB(tmp0, tmp1, dst, dst_stride); +} + +static void common_hz_8t_8x8mult_msa(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter, int32_t height) { + uint32_t loop_cnt; + v16i8 src0, src1, src2, src3, filt0, filt1, filt2, filt3; + v16u8 mask0, mask1, mask2, mask3, tmp0, tmp1; + v8i16 filt, out0, out1, out2, out3; + + mask0 = LD_UB(&mc_filt_mask_arr[0]); + src -= 3; + + /* rearranging filter */ + filt = LD_SH(filter); + SPLATI_H4_SB(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3); + + mask1 = mask0 + 2; + mask2 = mask0 + 4; + mask3 = mask0 + 6; + + for (loop_cnt = (height >> 2); loop_cnt--;) { + LD_SB4(src, src_stride, src0, src1, src2, src3); + XORI_B4_128_SB(src0, src1, src2, src3); + src += (4 * src_stride); + HORIZ_8TAP_8WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2, + mask3, filt0, filt1, filt2, filt3, out0, out1, + out2, out3); + SRARI_H4_SH(out0, out1, out2, out3, FILTER_BITS); + SAT_SH4_SH(out0, out1, out2, out3, 7); + tmp0 = PCKEV_XORI128_UB(out0, out1); + tmp1 = PCKEV_XORI128_UB(out2, out3); + ST8x4_UB(tmp0, tmp1, dst, dst_stride); + dst += (4 * dst_stride); + } +} + +static void common_hz_8t_8w_msa(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter, int32_t height) { + if (4 == height) { + common_hz_8t_8x4_msa(src, src_stride, dst, dst_stride, filter); + } else { + common_hz_8t_8x8mult_msa(src, src_stride, dst, dst_stride, filter, height); + } +} + +static void common_hz_8t_16w_msa(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter, int32_t height) { + uint32_t loop_cnt; + v16i8 src0, src1, src2, src3, filt0, filt1, filt2, filt3; + v16u8 mask0, mask1, mask2, mask3, out; + v8i16 filt, out0, out1, out2, out3; + + mask0 = LD_UB(&mc_filt_mask_arr[0]); + src -= 3; + + /* rearranging filter */ + filt = LD_SH(filter); + SPLATI_H4_SB(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3); + + mask1 = mask0 + 2; + mask2 = mask0 + 4; + mask3 = mask0 + 6; + + for (loop_cnt = (height >> 1); loop_cnt--;) { + LD_SB2(src, src_stride, src0, src2); + LD_SB2(src + 8, src_stride, src1, src3); + XORI_B4_128_SB(src0, src1, src2, src3); + src += (2 * src_stride); + HORIZ_8TAP_8WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2, + mask3, filt0, filt1, filt2, filt3, out0, out1, + out2, out3); + SRARI_H4_SH(out0, out1, out2, out3, FILTER_BITS); + SAT_SH4_SH(out0, out1, out2, out3, 7); + out = PCKEV_XORI128_UB(out0, out1); + ST_UB(out, dst); + dst += dst_stride; + out = PCKEV_XORI128_UB(out2, out3); + ST_UB(out, dst); + dst += dst_stride; + } +} + +static void common_hz_8t_32w_msa(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter, int32_t height) { + uint32_t loop_cnt; + v16i8 src0, src1, src2, src3, filt0, filt1, filt2, filt3; + v16u8 mask0, mask1, mask2, mask3, out; + v8i16 filt, out0, out1, out2, out3; + + mask0 = LD_UB(&mc_filt_mask_arr[0]); + src -= 3; + + /* rearranging filter */ + filt = LD_SH(filter); + SPLATI_H4_SB(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3); + + mask1 = mask0 + 2; + mask2 = mask0 + 4; + mask3 = mask0 + 6; + + for (loop_cnt = (height >> 1); loop_cnt--;) { + src0 = LD_SB(src); + src2 = LD_SB(src + 16); + src3 = LD_SB(src + 24); + src1 = __msa_sldi_b(src2, src0, 8); + src += src_stride; + XORI_B4_128_SB(src0, src1, src2, src3); + HORIZ_8TAP_8WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2, + mask3, filt0, filt1, filt2, filt3, out0, out1, + out2, out3); + SRARI_H4_SH(out0, out1, out2, out3, FILTER_BITS); + SAT_SH4_SH(out0, out1, out2, out3, 7); + + src0 = LD_SB(src); + src2 = LD_SB(src + 16); + src3 = LD_SB(src + 24); + src1 = __msa_sldi_b(src2, src0, 8); + src += src_stride; + + out = PCKEV_XORI128_UB(out0, out1); + ST_UB(out, dst); + out = PCKEV_XORI128_UB(out2, out3); + ST_UB(out, dst + 16); + dst += dst_stride; + + XORI_B4_128_SB(src0, src1, src2, src3); + HORIZ_8TAP_8WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2, + mask3, filt0, filt1, filt2, filt3, out0, out1, + out2, out3); + SRARI_H4_SH(out0, out1, out2, out3, FILTER_BITS); + SAT_SH4_SH(out0, out1, out2, out3, 7); + out = PCKEV_XORI128_UB(out0, out1); + ST_UB(out, dst); + out = PCKEV_XORI128_UB(out2, out3); + ST_UB(out, dst + 16); + dst += dst_stride; + } +} + +static void common_hz_8t_64w_msa(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter, int32_t height) { + int32_t loop_cnt; + v16i8 src0, src1, src2, src3, filt0, filt1, filt2, filt3; + v16u8 mask0, mask1, mask2, mask3, out; + v8i16 filt, out0, out1, out2, out3; + + mask0 = LD_UB(&mc_filt_mask_arr[0]); + src -= 3; + + /* rearranging filter */ + filt = LD_SH(filter); + SPLATI_H4_SB(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3); + + mask1 = mask0 + 2; + mask2 = mask0 + 4; + mask3 = mask0 + 6; + + for (loop_cnt = height; loop_cnt--;) { + src0 = LD_SB(src); + src2 = LD_SB(src + 16); + src3 = LD_SB(src + 24); + src1 = __msa_sldi_b(src2, src0, 8); + + XORI_B4_128_SB(src0, src1, src2, src3); + HORIZ_8TAP_8WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2, + mask3, filt0, filt1, filt2, filt3, out0, out1, + out2, out3); + SRARI_H4_SH(out0, out1, out2, out3, FILTER_BITS); + SAT_SH4_SH(out0, out1, out2, out3, 7); + out = PCKEV_XORI128_UB(out0, out1); + ST_UB(out, dst); + out = PCKEV_XORI128_UB(out2, out3); + ST_UB(out, dst + 16); + + src0 = LD_SB(src + 32); + src2 = LD_SB(src + 48); + src3 = LD_SB(src + 56); + src1 = __msa_sldi_b(src2, src0, 8); + src += src_stride; + + XORI_B4_128_SB(src0, src1, src2, src3); + HORIZ_8TAP_8WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2, + mask3, filt0, filt1, filt2, filt3, out0, out1, + out2, out3); + SRARI_H4_SH(out0, out1, out2, out3, FILTER_BITS); + SAT_SH4_SH(out0, out1, out2, out3, 7); + out = PCKEV_XORI128_UB(out0, out1); + ST_UB(out, dst + 32); + out = PCKEV_XORI128_UB(out2, out3); + ST_UB(out, dst + 48); + dst += dst_stride; + } +} + +static void common_hz_2t_4x4_msa(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter) { + v16i8 src0, src1, src2, src3, mask; + v16u8 filt0, vec0, vec1, res0, res1; + v8u16 vec2, vec3, filt; + + mask = LD_SB(&mc_filt_mask_arr[16]); + + /* rearranging filter */ + filt = LD_UH(filter); + filt0 = (v16u8)__msa_splati_h((v8i16)filt, 0); + + LD_SB4(src, src_stride, src0, src1, src2, src3); + VSHF_B2_UB(src0, src1, src2, src3, mask, mask, vec0, vec1); + DOTP_UB2_UH(vec0, vec1, filt0, filt0, vec2, vec3); + SRARI_H2_UH(vec2, vec3, FILTER_BITS); + PCKEV_B2_UB(vec2, vec2, vec3, vec3, res0, res1); + ST4x4_UB(res0, res1, 0, 1, 0, 1, dst, dst_stride); +} + +static void common_hz_2t_4x8_msa(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter) { + v16u8 vec0, vec1, vec2, vec3, filt0; + v16i8 src0, src1, src2, src3, src4, src5, src6, src7, mask; + v16i8 res0, res1, res2, res3; + v8u16 vec4, vec5, vec6, vec7, filt; + + mask = LD_SB(&mc_filt_mask_arr[16]); + + /* rearranging filter */ + filt = LD_UH(filter); + filt0 = (v16u8)__msa_splati_h((v8i16)filt, 0); + + LD_SB8(src, src_stride, src0, src1, src2, src3, src4, src5, src6, src7); + VSHF_B2_UB(src0, src1, src2, src3, mask, mask, vec0, vec1); + VSHF_B2_UB(src4, src5, src6, src7, mask, mask, vec2, vec3); + DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, vec4, vec5, + vec6, vec7); + SRARI_H4_UH(vec4, vec5, vec6, vec7, FILTER_BITS); + PCKEV_B4_SB(vec4, vec4, vec5, vec5, vec6, vec6, vec7, vec7, res0, res1, res2, + res3); + ST4x4_UB(res0, res1, 0, 1, 0, 1, dst, dst_stride); + dst += (4 * dst_stride); + ST4x4_UB(res2, res3, 0, 1, 0, 1, dst, dst_stride); +} + +static void common_hz_2t_4w_msa(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter, int32_t height) { + if (4 == height) { + common_hz_2t_4x4_msa(src, src_stride, dst, dst_stride, filter); + } else if (8 == height) { + common_hz_2t_4x8_msa(src, src_stride, dst, dst_stride, filter); + } +} + +static void common_hz_2t_8x4_msa(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter) { + v16u8 filt0; + v16i8 src0, src1, src2, src3, mask; + v8u16 vec0, vec1, vec2, vec3, filt; + + mask = LD_SB(&mc_filt_mask_arr[0]); + + /* rearranging filter */ + filt = LD_UH(filter); + filt0 = (v16u8)__msa_splati_h((v8i16)filt, 0); + + LD_SB4(src, src_stride, src0, src1, src2, src3); + VSHF_B2_UH(src0, src0, src1, src1, mask, mask, vec0, vec1); + VSHF_B2_UH(src2, src2, src3, src3, mask, mask, vec2, vec3); + DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, vec0, vec1, + vec2, vec3); + SRARI_H4_UH(vec0, vec1, vec2, vec3, FILTER_BITS); + PCKEV_B2_SB(vec1, vec0, vec3, vec2, src0, src1); + ST8x4_UB(src0, src1, dst, dst_stride); +} + +static void common_hz_2t_8x8mult_msa(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter, int32_t height) { + v16u8 filt0; + v16i8 src0, src1, src2, src3, mask, out0, out1; + v8u16 vec0, vec1, vec2, vec3, filt; + + mask = LD_SB(&mc_filt_mask_arr[0]); + + /* rearranging filter */ + filt = LD_UH(filter); + filt0 = (v16u8)__msa_splati_h((v8i16)filt, 0); + + LD_SB4(src, src_stride, src0, src1, src2, src3); + src += (4 * src_stride); + + VSHF_B2_UH(src0, src0, src1, src1, mask, mask, vec0, vec1); + VSHF_B2_UH(src2, src2, src3, src3, mask, mask, vec2, vec3); + DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, vec0, vec1, + vec2, vec3); + SRARI_H4_UH(vec0, vec1, vec2, vec3, FILTER_BITS); + + LD_SB4(src, src_stride, src0, src1, src2, src3); + src += (4 * src_stride); + + PCKEV_B2_SB(vec1, vec0, vec3, vec2, out0, out1); + ST8x4_UB(out0, out1, dst, dst_stride); + dst += (4 * dst_stride); + + VSHF_B2_UH(src0, src0, src1, src1, mask, mask, vec0, vec1); + VSHF_B2_UH(src2, src2, src3, src3, mask, mask, vec2, vec3); + DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, vec0, vec1, + vec2, vec3); + SRARI_H4_UH(vec0, vec1, vec2, vec3, FILTER_BITS); + PCKEV_B2_SB(vec1, vec0, vec3, vec2, out0, out1); + ST8x4_UB(out0, out1, dst, dst_stride); + dst += (4 * dst_stride); + + if (16 == height) { + LD_SB4(src, src_stride, src0, src1, src2, src3); + src += (4 * src_stride); + + VSHF_B2_UH(src0, src0, src1, src1, mask, mask, vec0, vec1); + VSHF_B2_UH(src2, src2, src3, src3, mask, mask, vec2, vec3); + DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, vec0, vec1, + vec2, vec3); + SRARI_H4_UH(vec0, vec1, vec2, vec3, FILTER_BITS); + LD_SB4(src, src_stride, src0, src1, src2, src3); + src += (4 * src_stride); + + PCKEV_B2_SB(vec1, vec0, vec3, vec2, out0, out1); + ST8x4_UB(out0, out1, dst, dst_stride); + + VSHF_B2_UH(src0, src0, src1, src1, mask, mask, vec0, vec1); + VSHF_B2_UH(src2, src2, src3, src3, mask, mask, vec2, vec3); + DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, vec0, vec1, + vec2, vec3); + SRARI_H4_UH(vec0, vec1, vec2, vec3, FILTER_BITS); + PCKEV_B2_SB(vec1, vec0, vec3, vec2, out0, out1); + ST8x4_UB(out0, out1, dst + 4 * dst_stride, dst_stride); + } +} + +static void common_hz_2t_8w_msa(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter, int32_t height) { + if (4 == height) { + common_hz_2t_8x4_msa(src, src_stride, dst, dst_stride, filter); + } else { + common_hz_2t_8x8mult_msa(src, src_stride, dst, dst_stride, filter, height); + } +} + +static void common_hz_2t_16w_msa(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter, int32_t height) { + uint32_t loop_cnt; + v16i8 src0, src1, src2, src3, src4, src5, src6, src7, mask; + v16u8 filt0, vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7; + v8u16 out0, out1, out2, out3, out4, out5, out6, out7, filt; + + mask = LD_SB(&mc_filt_mask_arr[0]); + + loop_cnt = (height >> 2) - 1; + + /* rearranging filter */ + filt = LD_UH(filter); + filt0 = (v16u8)__msa_splati_h((v8i16)filt, 0); + + LD_SB4(src, src_stride, src0, src2, src4, src6); + LD_SB4(src + 8, src_stride, src1, src3, src5, src7); + src += (4 * src_stride); + + VSHF_B2_UB(src0, src0, src1, src1, mask, mask, vec0, vec1); + VSHF_B2_UB(src2, src2, src3, src3, mask, mask, vec2, vec3); + VSHF_B2_UB(src4, src4, src5, src5, mask, mask, vec4, vec5); + VSHF_B2_UB(src6, src6, src7, src7, mask, mask, vec6, vec7); + DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, out0, out1, + out2, out3); + DOTP_UB4_UH(vec4, vec5, vec6, vec7, filt0, filt0, filt0, filt0, out4, out5, + out6, out7); + SRARI_H4_UH(out0, out1, out2, out3, FILTER_BITS); + SRARI_H4_UH(out4, out5, out6, out7, FILTER_BITS); + PCKEV_ST_SB(out0, out1, dst); + dst += dst_stride; + PCKEV_ST_SB(out2, out3, dst); + dst += dst_stride; + PCKEV_ST_SB(out4, out5, dst); + dst += dst_stride; + PCKEV_ST_SB(out6, out7, dst); + dst += dst_stride; + + for (; loop_cnt--;) { + LD_SB4(src, src_stride, src0, src2, src4, src6); + LD_SB4(src + 8, src_stride, src1, src3, src5, src7); + src += (4 * src_stride); + + VSHF_B2_UB(src0, src0, src1, src1, mask, mask, vec0, vec1); + VSHF_B2_UB(src2, src2, src3, src3, mask, mask, vec2, vec3); + VSHF_B2_UB(src4, src4, src5, src5, mask, mask, vec4, vec5); + VSHF_B2_UB(src6, src6, src7, src7, mask, mask, vec6, vec7); + DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, out0, out1, + out2, out3); + DOTP_UB4_UH(vec4, vec5, vec6, vec7, filt0, filt0, filt0, filt0, out4, out5, + out6, out7); + SRARI_H4_UH(out0, out1, out2, out3, FILTER_BITS); + SRARI_H4_UH(out4, out5, out6, out7, FILTER_BITS); + PCKEV_ST_SB(out0, out1, dst); + dst += dst_stride; + PCKEV_ST_SB(out2, out3, dst); + dst += dst_stride; + PCKEV_ST_SB(out4, out5, dst); + dst += dst_stride; + PCKEV_ST_SB(out6, out7, dst); + dst += dst_stride; + } +} + +static void common_hz_2t_32w_msa(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter, int32_t height) { + uint32_t loop_cnt; + v16i8 src0, src1, src2, src3, src4, src5, src6, src7, mask; + v16u8 filt0, vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7; + v8u16 out0, out1, out2, out3, out4, out5, out6, out7, filt; + + mask = LD_SB(&mc_filt_mask_arr[0]); + + /* rearranging filter */ + filt = LD_UH(filter); + filt0 = (v16u8)__msa_splati_h((v8i16)filt, 0); + + for (loop_cnt = height >> 1; loop_cnt--;) { + src0 = LD_SB(src); + src2 = LD_SB(src + 16); + src3 = LD_SB(src + 24); + src1 = __msa_sldi_b(src2, src0, 8); + src += src_stride; + src4 = LD_SB(src); + src6 = LD_SB(src + 16); + src7 = LD_SB(src + 24); + src5 = __msa_sldi_b(src6, src4, 8); + src += src_stride; + + VSHF_B2_UB(src0, src0, src1, src1, mask, mask, vec0, vec1); + VSHF_B2_UB(src2, src2, src3, src3, mask, mask, vec2, vec3); + VSHF_B2_UB(src4, src4, src5, src5, mask, mask, vec4, vec5); + VSHF_B2_UB(src6, src6, src7, src7, mask, mask, vec6, vec7); + DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, out0, out1, + out2, out3); + DOTP_UB4_UH(vec4, vec5, vec6, vec7, filt0, filt0, filt0, filt0, out4, out5, + out6, out7); + SRARI_H4_UH(out0, out1, out2, out3, FILTER_BITS); + SRARI_H4_UH(out4, out5, out6, out7, FILTER_BITS); + PCKEV_ST_SB(out0, out1, dst); + PCKEV_ST_SB(out2, out3, dst + 16); + dst += dst_stride; + PCKEV_ST_SB(out4, out5, dst); + PCKEV_ST_SB(out6, out7, dst + 16); + dst += dst_stride; + } +} + +static void common_hz_2t_64w_msa(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter, int32_t height) { + uint32_t loop_cnt; + v16i8 src0, src1, src2, src3, src4, src5, src6, src7, mask; + v16u8 filt0, vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7; + v8u16 out0, out1, out2, out3, out4, out5, out6, out7, filt; + + mask = LD_SB(&mc_filt_mask_arr[0]); + + /* rearranging filter */ + filt = LD_UH(filter); + filt0 = (v16u8)__msa_splati_h((v8i16)filt, 0); + + for (loop_cnt = height; loop_cnt--;) { + src0 = LD_SB(src); + src2 = LD_SB(src + 16); + src4 = LD_SB(src + 32); + src6 = LD_SB(src + 48); + src7 = LD_SB(src + 56); + SLDI_B3_SB(src2, src4, src6, src0, src2, src4, src1, src3, src5, 8); + src += src_stride; + + VSHF_B2_UB(src0, src0, src1, src1, mask, mask, vec0, vec1); + VSHF_B2_UB(src2, src2, src3, src3, mask, mask, vec2, vec3); + VSHF_B2_UB(src4, src4, src5, src5, mask, mask, vec4, vec5); + VSHF_B2_UB(src6, src6, src7, src7, mask, mask, vec6, vec7); + DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, out0, out1, + out2, out3); + DOTP_UB4_UH(vec4, vec5, vec6, vec7, filt0, filt0, filt0, filt0, out4, out5, + out6, out7); + SRARI_H4_UH(out0, out1, out2, out3, FILTER_BITS); + SRARI_H4_UH(out4, out5, out6, out7, FILTER_BITS); + PCKEV_ST_SB(out0, out1, dst); + PCKEV_ST_SB(out2, out3, dst + 16); + PCKEV_ST_SB(out4, out5, dst + 32); + PCKEV_ST_SB(out6, out7, dst + 48); + dst += dst_stride; + } +} + +void aom_convolve8_horiz_msa(const uint8_t *src, ptrdiff_t src_stride, + uint8_t *dst, ptrdiff_t dst_stride, + const int16_t *filter_x, int x_step_q4, + const int16_t *filter_y, int y_step_q4, int w, + int h) { + int8_t cnt, filt_hor[8]; + + assert(x_step_q4 == 16); + assert(((const int32_t *)filter_x)[1] != 0x800000); + + for (cnt = 0; cnt < 8; ++cnt) { + filt_hor[cnt] = filter_x[cnt]; + } + + if (((const int32_t *)filter_x)[0] == 0) { + switch (w) { + case 4: + common_hz_2t_4w_msa(src, (int32_t)src_stride, dst, (int32_t)dst_stride, + &filt_hor[3], h); + break; + case 8: + common_hz_2t_8w_msa(src, (int32_t)src_stride, dst, (int32_t)dst_stride, + &filt_hor[3], h); + break; + case 16: + common_hz_2t_16w_msa(src, (int32_t)src_stride, dst, (int32_t)dst_stride, + &filt_hor[3], h); + break; + case 32: + common_hz_2t_32w_msa(src, (int32_t)src_stride, dst, (int32_t)dst_stride, + &filt_hor[3], h); + break; + case 64: + common_hz_2t_64w_msa(src, (int32_t)src_stride, dst, (int32_t)dst_stride, + &filt_hor[3], h); + break; + default: + aom_convolve8_horiz_c(src, src_stride, dst, dst_stride, filter_x, + x_step_q4, filter_y, y_step_q4, w, h); + break; + } + } else { + switch (w) { + case 4: + common_hz_8t_4w_msa(src, (int32_t)src_stride, dst, (int32_t)dst_stride, + filt_hor, h); + break; + case 8: + common_hz_8t_8w_msa(src, (int32_t)src_stride, dst, (int32_t)dst_stride, + filt_hor, h); + break; + case 16: + common_hz_8t_16w_msa(src, (int32_t)src_stride, dst, (int32_t)dst_stride, + filt_hor, h); + break; + case 32: + common_hz_8t_32w_msa(src, (int32_t)src_stride, dst, (int32_t)dst_stride, + filt_hor, h); + break; + case 64: + common_hz_8t_64w_msa(src, (int32_t)src_stride, dst, (int32_t)dst_stride, + filt_hor, h); + break; + default: + aom_convolve8_horiz_c(src, src_stride, dst, dst_stride, filter_x, + x_step_q4, filter_y, y_step_q4, w, h); + break; + } + } +} diff --git a/third_party/aom/aom_dsp/mips/aom_convolve8_vert_msa.c b/third_party/aom/aom_dsp/mips/aom_convolve8_vert_msa.c new file mode 100644 index 0000000000..aa962b41f7 --- /dev/null +++ b/third_party/aom/aom_dsp/mips/aom_convolve8_vert_msa.c @@ -0,0 +1,701 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include + +#include "config/aom_dsp_rtcd.h" + +#include "aom_dsp/mips/aom_convolve_msa.h" + +static void common_vt_8t_4w_msa(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter, int32_t height) { + uint32_t loop_cnt; + v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10; + v16i8 src10_r, src32_r, src54_r, src76_r, src98_r, src21_r, src43_r; + v16i8 src65_r, src87_r, src109_r, src2110, src4332, src6554, src8776; + v16i8 src10998, filt0, filt1, filt2, filt3; + v16u8 out; + v8i16 filt, out10, out32; + + src -= (3 * src_stride); + + filt = LD_SH(filter); + SPLATI_H4_SB(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3); + + LD_SB7(src, src_stride, src0, src1, src2, src3, src4, src5, src6); + src += (7 * src_stride); + + ILVR_B4_SB(src1, src0, src3, src2, src5, src4, src2, src1, src10_r, src32_r, + src54_r, src21_r); + ILVR_B2_SB(src4, src3, src6, src5, src43_r, src65_r); + ILVR_D3_SB(src21_r, src10_r, src43_r, src32_r, src65_r, src54_r, src2110, + src4332, src6554); + XORI_B3_128_SB(src2110, src4332, src6554); + + for (loop_cnt = (height >> 2); loop_cnt--;) { + LD_SB4(src, src_stride, src7, src8, src9, src10); + src += (4 * src_stride); + + ILVR_B4_SB(src7, src6, src8, src7, src9, src8, src10, src9, src76_r, + src87_r, src98_r, src109_r); + ILVR_D2_SB(src87_r, src76_r, src109_r, src98_r, src8776, src10998); + XORI_B2_128_SB(src8776, src10998); + out10 = FILT_8TAP_DPADD_S_H(src2110, src4332, src6554, src8776, filt0, + filt1, filt2, filt3); + out32 = FILT_8TAP_DPADD_S_H(src4332, src6554, src8776, src10998, filt0, + filt1, filt2, filt3); + SRARI_H2_SH(out10, out32, FILTER_BITS); + SAT_SH2_SH(out10, out32, 7); + out = PCKEV_XORI128_UB(out10, out32); + ST4x4_UB(out, out, 0, 1, 2, 3, dst, dst_stride); + dst += (4 * dst_stride); + + src2110 = src6554; + src4332 = src8776; + src6554 = src10998; + src6 = src10; + } +} + +static void common_vt_8t_8w_msa(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter, int32_t height) { + uint32_t loop_cnt; + v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10; + v16i8 src10_r, src32_r, src54_r, src76_r, src98_r, src21_r, src43_r; + v16i8 src65_r, src87_r, src109_r, filt0, filt1, filt2, filt3; + v16u8 tmp0, tmp1; + v8i16 filt, out0_r, out1_r, out2_r, out3_r; + + src -= (3 * src_stride); + + filt = LD_SH(filter); + SPLATI_H4_SB(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3); + + LD_SB7(src, src_stride, src0, src1, src2, src3, src4, src5, src6); + XORI_B7_128_SB(src0, src1, src2, src3, src4, src5, src6); + src += (7 * src_stride); + ILVR_B4_SB(src1, src0, src3, src2, src5, src4, src2, src1, src10_r, src32_r, + src54_r, src21_r); + ILVR_B2_SB(src4, src3, src6, src5, src43_r, src65_r); + + for (loop_cnt = (height >> 2); loop_cnt--;) { + LD_SB4(src, src_stride, src7, src8, src9, src10); + XORI_B4_128_SB(src7, src8, src9, src10); + src += (4 * src_stride); + + ILVR_B4_SB(src7, src6, src8, src7, src9, src8, src10, src9, src76_r, + src87_r, src98_r, src109_r); + out0_r = FILT_8TAP_DPADD_S_H(src10_r, src32_r, src54_r, src76_r, filt0, + filt1, filt2, filt3); + out1_r = FILT_8TAP_DPADD_S_H(src21_r, src43_r, src65_r, src87_r, filt0, + filt1, filt2, filt3); + out2_r = FILT_8TAP_DPADD_S_H(src32_r, src54_r, src76_r, src98_r, filt0, + filt1, filt2, filt3); + out3_r = FILT_8TAP_DPADD_S_H(src43_r, src65_r, src87_r, src109_r, filt0, + filt1, filt2, filt3); + SRARI_H4_SH(out0_r, out1_r, out2_r, out3_r, FILTER_BITS); + SAT_SH4_SH(out0_r, out1_r, out2_r, out3_r, 7); + tmp0 = PCKEV_XORI128_UB(out0_r, out1_r); + tmp1 = PCKEV_XORI128_UB(out2_r, out3_r); + ST8x4_UB(tmp0, tmp1, dst, dst_stride); + dst += (4 * dst_stride); + + src10_r = src54_r; + src32_r = src76_r; + src54_r = src98_r; + src21_r = src65_r; + src43_r = src87_r; + src65_r = src109_r; + src6 = src10; + } +} + +static void common_vt_8t_16w_msa(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter, int32_t height) { + uint32_t loop_cnt; + v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10; + v16i8 filt0, filt1, filt2, filt3; + v16i8 src10_r, src32_r, src54_r, src76_r, src98_r, src21_r, src43_r; + v16i8 src65_r, src87_r, src109_r, src10_l, src32_l, src54_l, src76_l; + v16i8 src98_l, src21_l, src43_l, src65_l, src87_l, src109_l; + v16u8 tmp0, tmp1, tmp2, tmp3; + v8i16 filt, out0_r, out1_r, out2_r, out3_r, out0_l, out1_l, out2_l, out3_l; + + src -= (3 * src_stride); + + filt = LD_SH(filter); + SPLATI_H4_SB(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3); + + LD_SB7(src, src_stride, src0, src1, src2, src3, src4, src5, src6); + XORI_B7_128_SB(src0, src1, src2, src3, src4, src5, src6); + src += (7 * src_stride); + ILVR_B4_SB(src1, src0, src3, src2, src5, src4, src2, src1, src10_r, src32_r, + src54_r, src21_r); + ILVR_B2_SB(src4, src3, src6, src5, src43_r, src65_r); + ILVL_B4_SB(src1, src0, src3, src2, src5, src4, src2, src1, src10_l, src32_l, + src54_l, src21_l); + ILVL_B2_SB(src4, src3, src6, src5, src43_l, src65_l); + + for (loop_cnt = (height >> 2); loop_cnt--;) { + LD_SB4(src, src_stride, src7, src8, src9, src10); + XORI_B4_128_SB(src7, src8, src9, src10); + src += (4 * src_stride); + + ILVR_B4_SB(src7, src6, src8, src7, src9, src8, src10, src9, src76_r, + src87_r, src98_r, src109_r); + ILVL_B4_SB(src7, src6, src8, src7, src9, src8, src10, src9, src76_l, + src87_l, src98_l, src109_l); + out0_r = FILT_8TAP_DPADD_S_H(src10_r, src32_r, src54_r, src76_r, filt0, + filt1, filt2, filt3); + out1_r = FILT_8TAP_DPADD_S_H(src21_r, src43_r, src65_r, src87_r, filt0, + filt1, filt2, filt3); + out2_r = FILT_8TAP_DPADD_S_H(src32_r, src54_r, src76_r, src98_r, filt0, + filt1, filt2, filt3); + out3_r = FILT_8TAP_DPADD_S_H(src43_r, src65_r, src87_r, src109_r, filt0, + filt1, filt2, filt3); + out0_l = FILT_8TAP_DPADD_S_H(src10_l, src32_l, src54_l, src76_l, filt0, + filt1, filt2, filt3); + out1_l = FILT_8TAP_DPADD_S_H(src21_l, src43_l, src65_l, src87_l, filt0, + filt1, filt2, filt3); + out2_l = FILT_8TAP_DPADD_S_H(src32_l, src54_l, src76_l, src98_l, filt0, + filt1, filt2, filt3); + out3_l = FILT_8TAP_DPADD_S_H(src43_l, src65_l, src87_l, src109_l, filt0, + filt1, filt2, filt3); + SRARI_H4_SH(out0_r, out1_r, out2_r, out3_r, FILTER_BITS); + SRARI_H4_SH(out0_l, out1_l, out2_l, out3_l, FILTER_BITS); + SAT_SH4_SH(out0_r, out1_r, out2_r, out3_r, 7); + SAT_SH4_SH(out0_l, out1_l, out2_l, out3_l, 7); + PCKEV_B4_UB(out0_l, out0_r, out1_l, out1_r, out2_l, out2_r, out3_l, out3_r, + tmp0, tmp1, tmp2, tmp3); + XORI_B4_128_UB(tmp0, tmp1, tmp2, tmp3); + ST_UB4(tmp0, tmp1, tmp2, tmp3, dst, dst_stride); + dst += (4 * dst_stride); + + src10_r = src54_r; + src32_r = src76_r; + src54_r = src98_r; + src21_r = src65_r; + src43_r = src87_r; + src65_r = src109_r; + src10_l = src54_l; + src32_l = src76_l; + src54_l = src98_l; + src21_l = src65_l; + src43_l = src87_l; + src65_l = src109_l; + src6 = src10; + } +} + +static void common_vt_8t_16w_mult_msa(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter, int32_t height, + int32_t width) { + const uint8_t *src_tmp; + uint8_t *dst_tmp; + uint32_t loop_cnt, cnt; + v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10; + v16i8 filt0, filt1, filt2, filt3; + v16i8 src10_r, src32_r, src54_r, src76_r, src98_r, src21_r, src43_r; + v16i8 src65_r, src87_r, src109_r, src10_l, src32_l, src54_l, src76_l; + v16i8 src98_l, src21_l, src43_l, src65_l, src87_l, src109_l; + v16u8 tmp0, tmp1, tmp2, tmp3; + v8i16 filt, out0_r, out1_r, out2_r, out3_r, out0_l, out1_l, out2_l, out3_l; + + src -= (3 * src_stride); + + filt = LD_SH(filter); + SPLATI_H4_SB(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3); + + for (cnt = (width >> 4); cnt--;) { + src_tmp = src; + dst_tmp = dst; + + LD_SB7(src_tmp, src_stride, src0, src1, src2, src3, src4, src5, src6); + XORI_B7_128_SB(src0, src1, src2, src3, src4, src5, src6); + src_tmp += (7 * src_stride); + ILVR_B4_SB(src1, src0, src3, src2, src5, src4, src2, src1, src10_r, src32_r, + src54_r, src21_r); + ILVR_B2_SB(src4, src3, src6, src5, src43_r, src65_r); + ILVL_B4_SB(src1, src0, src3, src2, src5, src4, src2, src1, src10_l, src32_l, + src54_l, src21_l); + ILVL_B2_SB(src4, src3, src6, src5, src43_l, src65_l); + + for (loop_cnt = (height >> 2); loop_cnt--;) { + LD_SB4(src_tmp, src_stride, src7, src8, src9, src10); + XORI_B4_128_SB(src7, src8, src9, src10); + src_tmp += (4 * src_stride); + ILVR_B4_SB(src7, src6, src8, src7, src9, src8, src10, src9, src76_r, + src87_r, src98_r, src109_r); + ILVL_B4_SB(src7, src6, src8, src7, src9, src8, src10, src9, src76_l, + src87_l, src98_l, src109_l); + out0_r = FILT_8TAP_DPADD_S_H(src10_r, src32_r, src54_r, src76_r, filt0, + filt1, filt2, filt3); + out1_r = FILT_8TAP_DPADD_S_H(src21_r, src43_r, src65_r, src87_r, filt0, + filt1, filt2, filt3); + out2_r = FILT_8TAP_DPADD_S_H(src32_r, src54_r, src76_r, src98_r, filt0, + filt1, filt2, filt3); + out3_r = FILT_8TAP_DPADD_S_H(src43_r, src65_r, src87_r, src109_r, filt0, + filt1, filt2, filt3); + out0_l = FILT_8TAP_DPADD_S_H(src10_l, src32_l, src54_l, src76_l, filt0, + filt1, filt2, filt3); + out1_l = FILT_8TAP_DPADD_S_H(src21_l, src43_l, src65_l, src87_l, filt0, + filt1, filt2, filt3); + out2_l = FILT_8TAP_DPADD_S_H(src32_l, src54_l, src76_l, src98_l, filt0, + filt1, filt2, filt3); + out3_l = FILT_8TAP_DPADD_S_H(src43_l, src65_l, src87_l, src109_l, filt0, + filt1, filt2, filt3); + SRARI_H4_SH(out0_r, out1_r, out2_r, out3_r, FILTER_BITS); + SRARI_H4_SH(out0_l, out1_l, out2_l, out3_l, FILTER_BITS); + SAT_SH4_SH(out0_r, out1_r, out2_r, out3_r, 7); + SAT_SH4_SH(out0_l, out1_l, out2_l, out3_l, 7); + PCKEV_B4_UB(out0_l, out0_r, out1_l, out1_r, out2_l, out2_r, out3_l, + out3_r, tmp0, tmp1, tmp2, tmp3); + XORI_B4_128_UB(tmp0, tmp1, tmp2, tmp3); + ST_UB4(tmp0, tmp1, tmp2, tmp3, dst_tmp, dst_stride); + dst_tmp += (4 * dst_stride); + + src10_r = src54_r; + src32_r = src76_r; + src54_r = src98_r; + src21_r = src65_r; + src43_r = src87_r; + src65_r = src109_r; + src10_l = src54_l; + src32_l = src76_l; + src54_l = src98_l; + src21_l = src65_l; + src43_l = src87_l; + src65_l = src109_l; + src6 = src10; + } + + src += 16; + dst += 16; + } +} + +static void common_vt_8t_32w_msa(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter, int32_t height) { + common_vt_8t_16w_mult_msa(src, src_stride, dst, dst_stride, filter, height, + 32); +} + +static void common_vt_8t_64w_msa(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter, int32_t height) { + common_vt_8t_16w_mult_msa(src, src_stride, dst, dst_stride, filter, height, + 64); +} + +static void common_vt_2t_4x4_msa(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter) { + v16i8 src0, src1, src2, src3, src4; + v16i8 src10_r, src32_r, src21_r, src43_r, src2110, src4332; + v16u8 filt0; + v8i16 filt; + v8u16 tmp0, tmp1; + + filt = LD_SH(filter); + filt0 = (v16u8)__msa_splati_h(filt, 0); + + LD_SB5(src, src_stride, src0, src1, src2, src3, src4); + src += (5 * src_stride); + + ILVR_B4_SB(src1, src0, src2, src1, src3, src2, src4, src3, src10_r, src21_r, + src32_r, src43_r); + ILVR_D2_SB(src21_r, src10_r, src43_r, src32_r, src2110, src4332); + DOTP_UB2_UH(src2110, src4332, filt0, filt0, tmp0, tmp1); + SRARI_H2_UH(tmp0, tmp1, FILTER_BITS); + src2110 = __msa_pckev_b((v16i8)tmp1, (v16i8)tmp0); + ST4x4_UB(src2110, src2110, 0, 1, 2, 3, dst, dst_stride); +} + +static void common_vt_2t_4x8_msa(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter) { + v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8; + v16i8 src10_r, src32_r, src54_r, src76_r, src21_r, src43_r; + v16i8 src65_r, src87_r, src2110, src4332, src6554, src8776; + v8u16 tmp0, tmp1, tmp2, tmp3; + v16u8 filt0; + v8i16 filt; + + filt = LD_SH(filter); + filt0 = (v16u8)__msa_splati_h(filt, 0); + + LD_SB8(src, src_stride, src0, src1, src2, src3, src4, src5, src6, src7); + src += (8 * src_stride); + + src8 = LD_SB(src); + src += src_stride; + + ILVR_B4_SB(src1, src0, src2, src1, src3, src2, src4, src3, src10_r, src21_r, + src32_r, src43_r); + ILVR_B4_SB(src5, src4, src6, src5, src7, src6, src8, src7, src54_r, src65_r, + src76_r, src87_r); + ILVR_D4_SB(src21_r, src10_r, src43_r, src32_r, src65_r, src54_r, src87_r, + src76_r, src2110, src4332, src6554, src8776); + DOTP_UB4_UH(src2110, src4332, src6554, src8776, filt0, filt0, filt0, filt0, + tmp0, tmp1, tmp2, tmp3); + SRARI_H4_UH(tmp0, tmp1, tmp2, tmp3, FILTER_BITS); + PCKEV_B2_SB(tmp1, tmp0, tmp3, tmp2, src2110, src4332); + ST4x4_UB(src2110, src2110, 0, 1, 2, 3, dst, dst_stride); + ST4x4_UB(src4332, src4332, 0, 1, 2, 3, dst + 4 * dst_stride, dst_stride); +} + +static void common_vt_2t_4w_msa(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter, int32_t height) { + if (4 == height) { + common_vt_2t_4x4_msa(src, src_stride, dst, dst_stride, filter); + } else if (8 == height) { + common_vt_2t_4x8_msa(src, src_stride, dst, dst_stride, filter); + } +} + +static void common_vt_2t_8x4_msa(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter) { + v16u8 src0, src1, src2, src3, src4, vec0, vec1, vec2, vec3, filt0; + v16i8 out0, out1; + v8u16 tmp0, tmp1, tmp2, tmp3; + v8i16 filt; + + /* rearranging filter_y */ + filt = LD_SH(filter); + filt0 = (v16u8)__msa_splati_h(filt, 0); + + LD_UB5(src, src_stride, src0, src1, src2, src3, src4); + ILVR_B2_UB(src1, src0, src2, src1, vec0, vec1); + ILVR_B2_UB(src3, src2, src4, src3, vec2, vec3); + DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, tmp0, tmp1, + tmp2, tmp3); + SRARI_H4_UH(tmp0, tmp1, tmp2, tmp3, FILTER_BITS); + PCKEV_B2_SB(tmp1, tmp0, tmp3, tmp2, out0, out1); + ST8x4_UB(out0, out1, dst, dst_stride); +} + +static void common_vt_2t_8x8mult_msa(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter, int32_t height) { + uint32_t loop_cnt; + v16u8 src0, src1, src2, src3, src4, src5, src6, src7, src8; + v16u8 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, filt0; + v16i8 out0, out1; + v8u16 tmp0, tmp1, tmp2, tmp3; + v8i16 filt; + + /* rearranging filter_y */ + filt = LD_SH(filter); + filt0 = (v16u8)__msa_splati_h(filt, 0); + + src0 = LD_UB(src); + src += src_stride; + + for (loop_cnt = (height >> 3); loop_cnt--;) { + LD_UB8(src, src_stride, src1, src2, src3, src4, src5, src6, src7, src8); + src += (8 * src_stride); + + ILVR_B4_UB(src1, src0, src2, src1, src3, src2, src4, src3, vec0, vec1, vec2, + vec3); + ILVR_B4_UB(src5, src4, src6, src5, src7, src6, src8, src7, vec4, vec5, vec6, + vec7); + DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, tmp0, tmp1, + tmp2, tmp3); + SRARI_H4_UH(tmp0, tmp1, tmp2, tmp3, FILTER_BITS); + PCKEV_B2_SB(tmp1, tmp0, tmp3, tmp2, out0, out1); + ST8x4_UB(out0, out1, dst, dst_stride); + dst += (4 * dst_stride); + + DOTP_UB4_UH(vec4, vec5, vec6, vec7, filt0, filt0, filt0, filt0, tmp0, tmp1, + tmp2, tmp3); + SRARI_H4_UH(tmp0, tmp1, tmp2, tmp3, FILTER_BITS); + PCKEV_B2_SB(tmp1, tmp0, tmp3, tmp2, out0, out1); + ST8x4_UB(out0, out1, dst, dst_stride); + dst += (4 * dst_stride); + + src0 = src8; + } +} + +static void common_vt_2t_8w_msa(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter, int32_t height) { + if (4 == height) { + common_vt_2t_8x4_msa(src, src_stride, dst, dst_stride, filter); + } else { + common_vt_2t_8x8mult_msa(src, src_stride, dst, dst_stride, filter, height); + } +} + +static void common_vt_2t_16w_msa(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter, int32_t height) { + uint32_t loop_cnt; + v16u8 src0, src1, src2, src3, src4; + v16u8 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, filt0; + v8u16 tmp0, tmp1, tmp2, tmp3; + v8i16 filt; + + /* rearranging filter_y */ + filt = LD_SH(filter); + filt0 = (v16u8)__msa_splati_h(filt, 0); + + src0 = LD_UB(src); + src += src_stride; + + for (loop_cnt = (height >> 2); loop_cnt--;) { + LD_UB4(src, src_stride, src1, src2, src3, src4); + src += (4 * src_stride); + + ILVR_B2_UB(src1, src0, src2, src1, vec0, vec2); + ILVL_B2_UB(src1, src0, src2, src1, vec1, vec3); + DOTP_UB2_UH(vec0, vec1, filt0, filt0, tmp0, tmp1); + SRARI_H2_UH(tmp0, tmp1, FILTER_BITS); + PCKEV_ST_SB(tmp0, tmp1, dst); + dst += dst_stride; + + ILVR_B2_UB(src3, src2, src4, src3, vec4, vec6); + ILVL_B2_UB(src3, src2, src4, src3, vec5, vec7); + DOTP_UB2_UH(vec2, vec3, filt0, filt0, tmp2, tmp3); + SRARI_H2_UH(tmp2, tmp3, FILTER_BITS); + PCKEV_ST_SB(tmp2, tmp3, dst); + dst += dst_stride; + + DOTP_UB2_UH(vec4, vec5, filt0, filt0, tmp0, tmp1); + SRARI_H2_UH(tmp0, tmp1, FILTER_BITS); + PCKEV_ST_SB(tmp0, tmp1, dst); + dst += dst_stride; + + DOTP_UB2_UH(vec6, vec7, filt0, filt0, tmp2, tmp3); + SRARI_H2_UH(tmp2, tmp3, FILTER_BITS); + PCKEV_ST_SB(tmp2, tmp3, dst); + dst += dst_stride; + + src0 = src4; + } +} + +static void common_vt_2t_32w_msa(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter, int32_t height) { + uint32_t loop_cnt; + v16u8 src0, src1, src2, src3, src4, src5, src6, src7, src8, src9; + v16u8 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, filt0; + v8u16 tmp0, tmp1, tmp2, tmp3; + v8i16 filt; + + /* rearranging filter_y */ + filt = LD_SH(filter); + filt0 = (v16u8)__msa_splati_h(filt, 0); + + src0 = LD_UB(src); + src5 = LD_UB(src + 16); + src += src_stride; + + for (loop_cnt = (height >> 2); loop_cnt--;) { + LD_UB4(src, src_stride, src1, src2, src3, src4); + ILVR_B2_UB(src1, src0, src2, src1, vec0, vec2); + ILVL_B2_UB(src1, src0, src2, src1, vec1, vec3); + + LD_UB4(src + 16, src_stride, src6, src7, src8, src9); + src += (4 * src_stride); + + DOTP_UB2_UH(vec0, vec1, filt0, filt0, tmp0, tmp1); + SRARI_H2_UH(tmp0, tmp1, FILTER_BITS); + PCKEV_ST_SB(tmp0, tmp1, dst); + DOTP_UB2_UH(vec2, vec3, filt0, filt0, tmp2, tmp3); + SRARI_H2_UH(tmp2, tmp3, FILTER_BITS); + PCKEV_ST_SB(tmp2, tmp3, dst + dst_stride); + + ILVR_B2_UB(src3, src2, src4, src3, vec4, vec6); + ILVL_B2_UB(src3, src2, src4, src3, vec5, vec7); + DOTP_UB2_UH(vec4, vec5, filt0, filt0, tmp0, tmp1); + SRARI_H2_UH(tmp0, tmp1, FILTER_BITS); + PCKEV_ST_SB(tmp0, tmp1, dst + 2 * dst_stride); + + DOTP_UB2_UH(vec6, vec7, filt0, filt0, tmp2, tmp3); + SRARI_H2_UH(tmp2, tmp3, FILTER_BITS); + PCKEV_ST_SB(tmp2, tmp3, dst + 3 * dst_stride); + + ILVR_B2_UB(src6, src5, src7, src6, vec0, vec2); + ILVL_B2_UB(src6, src5, src7, src6, vec1, vec3); + DOTP_UB2_UH(vec0, vec1, filt0, filt0, tmp0, tmp1); + SRARI_H2_UH(tmp0, tmp1, FILTER_BITS); + PCKEV_ST_SB(tmp0, tmp1, dst + 16); + + DOTP_UB2_UH(vec2, vec3, filt0, filt0, tmp2, tmp3); + SRARI_H2_UH(tmp2, tmp3, FILTER_BITS); + PCKEV_ST_SB(tmp2, tmp3, dst + 16 + dst_stride); + + ILVR_B2_UB(src8, src7, src9, src8, vec4, vec6); + ILVL_B2_UB(src8, src7, src9, src8, vec5, vec7); + DOTP_UB2_UH(vec4, vec5, filt0, filt0, tmp0, tmp1); + SRARI_H2_UH(tmp0, tmp1, FILTER_BITS); + PCKEV_ST_SB(tmp0, tmp1, dst + 16 + 2 * dst_stride); + + DOTP_UB2_UH(vec6, vec7, filt0, filt0, tmp2, tmp3); + SRARI_H2_UH(tmp2, tmp3, FILTER_BITS); + PCKEV_ST_SB(tmp2, tmp3, dst + 16 + 3 * dst_stride); + dst += (4 * dst_stride); + + src0 = src4; + src5 = src9; + } +} + +static void common_vt_2t_64w_msa(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter, int32_t height) { + uint32_t loop_cnt; + v16u8 src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10; + v16u8 src11, vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, filt0; + v8u16 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7; + v8i16 filt; + + /* rearranging filter_y */ + filt = LD_SH(filter); + filt0 = (v16u8)__msa_splati_h(filt, 0); + + LD_UB4(src, 16, src0, src3, src6, src9); + src += src_stride; + + for (loop_cnt = (height >> 1); loop_cnt--;) { + LD_UB2(src, src_stride, src1, src2); + LD_UB2(src + 16, src_stride, src4, src5); + LD_UB2(src + 32, src_stride, src7, src8); + LD_UB2(src + 48, src_stride, src10, src11); + src += (2 * src_stride); + + ILVR_B2_UB(src1, src0, src2, src1, vec0, vec2); + ILVL_B2_UB(src1, src0, src2, src1, vec1, vec3); + DOTP_UB2_UH(vec0, vec1, filt0, filt0, tmp0, tmp1); + SRARI_H2_UH(tmp0, tmp1, FILTER_BITS); + PCKEV_ST_SB(tmp0, tmp1, dst); + + DOTP_UB2_UH(vec2, vec3, filt0, filt0, tmp2, tmp3); + SRARI_H2_UH(tmp2, tmp3, FILTER_BITS); + PCKEV_ST_SB(tmp2, tmp3, dst + dst_stride); + + ILVR_B2_UB(src4, src3, src5, src4, vec4, vec6); + ILVL_B2_UB(src4, src3, src5, src4, vec5, vec7); + DOTP_UB2_UH(vec4, vec5, filt0, filt0, tmp4, tmp5); + SRARI_H2_UH(tmp4, tmp5, FILTER_BITS); + PCKEV_ST_SB(tmp4, tmp5, dst + 16); + + DOTP_UB2_UH(vec6, vec7, filt0, filt0, tmp6, tmp7); + SRARI_H2_UH(tmp6, tmp7, FILTER_BITS); + PCKEV_ST_SB(tmp6, tmp7, dst + 16 + dst_stride); + + ILVR_B2_UB(src7, src6, src8, src7, vec0, vec2); + ILVL_B2_UB(src7, src6, src8, src7, vec1, vec3); + DOTP_UB2_UH(vec0, vec1, filt0, filt0, tmp0, tmp1); + SRARI_H2_UH(tmp0, tmp1, FILTER_BITS); + PCKEV_ST_SB(tmp0, tmp1, dst + 32); + + DOTP_UB2_UH(vec2, vec3, filt0, filt0, tmp2, tmp3); + SRARI_H2_UH(tmp2, tmp3, FILTER_BITS); + PCKEV_ST_SB(tmp2, tmp3, dst + 32 + dst_stride); + + ILVR_B2_UB(src10, src9, src11, src10, vec4, vec6); + ILVL_B2_UB(src10, src9, src11, src10, vec5, vec7); + DOTP_UB2_UH(vec4, vec5, filt0, filt0, tmp4, tmp5); + SRARI_H2_UH(tmp4, tmp5, FILTER_BITS); + PCKEV_ST_SB(tmp4, tmp5, dst + 48); + + DOTP_UB2_UH(vec6, vec7, filt0, filt0, tmp6, tmp7); + SRARI_H2_UH(tmp6, tmp7, FILTER_BITS); + PCKEV_ST_SB(tmp6, tmp7, dst + 48 + dst_stride); + dst += (2 * dst_stride); + + src0 = src2; + src3 = src5; + src6 = src8; + src9 = src11; + } +} + +void aom_convolve8_vert_msa(const uint8_t *src, ptrdiff_t src_stride, + uint8_t *dst, ptrdiff_t dst_stride, + const int16_t *filter_x, int x_step_q4, + const int16_t *filter_y, int y_step_q4, int w, + int h) { + int8_t cnt, filt_ver[8]; + + assert(y_step_q4 == 16); + assert(((const int32_t *)filter_y)[1] != 0x800000); + + for (cnt = 8; cnt--;) { + filt_ver[cnt] = filter_y[cnt]; + } + + if (((const int32_t *)filter_y)[0] == 0) { + switch (w) { + case 4: + common_vt_2t_4w_msa(src, (int32_t)src_stride, dst, (int32_t)dst_stride, + &filt_ver[3], h); + break; + case 8: + common_vt_2t_8w_msa(src, (int32_t)src_stride, dst, (int32_t)dst_stride, + &filt_ver[3], h); + break; + case 16: + common_vt_2t_16w_msa(src, (int32_t)src_stride, dst, (int32_t)dst_stride, + &filt_ver[3], h); + break; + case 32: + common_vt_2t_32w_msa(src, (int32_t)src_stride, dst, (int32_t)dst_stride, + &filt_ver[3], h); + break; + case 64: + common_vt_2t_64w_msa(src, (int32_t)src_stride, dst, (int32_t)dst_stride, + &filt_ver[3], h); + break; + default: + aom_convolve8_vert_c(src, src_stride, dst, dst_stride, filter_x, + x_step_q4, filter_y, y_step_q4, w, h); + break; + } + } else { + switch (w) { + case 4: + common_vt_8t_4w_msa(src, (int32_t)src_stride, dst, (int32_t)dst_stride, + filt_ver, h); + break; + case 8: + common_vt_8t_8w_msa(src, (int32_t)src_stride, dst, (int32_t)dst_stride, + filt_ver, h); + break; + case 16: + common_vt_8t_16w_msa(src, (int32_t)src_stride, dst, (int32_t)dst_stride, + filt_ver, h); + break; + case 32: + common_vt_8t_32w_msa(src, (int32_t)src_stride, dst, (int32_t)dst_stride, + filt_ver, h); + break; + case 64: + common_vt_8t_64w_msa(src, (int32_t)src_stride, dst, (int32_t)dst_stride, + filt_ver, h); + break; + default: + aom_convolve8_vert_c(src, src_stride, dst, dst_stride, filter_x, + x_step_q4, filter_y, y_step_q4, w, h); + break; + } + } +} diff --git a/third_party/aom/aom_dsp/mips/aom_convolve_copy_msa.c b/third_party/aom/aom_dsp/mips/aom_convolve_copy_msa.c new file mode 100644 index 0000000000..f7f116f4da --- /dev/null +++ b/third_party/aom/aom_dsp/mips/aom_convolve_copy_msa.c @@ -0,0 +1,248 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include +#include "aom_dsp/mips/macros_msa.h" + +static void copy_width8_msa(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, int32_t height) { + int32_t cnt; + uint64_t out0, out1, out2, out3, out4, out5, out6, out7; + v16u8 src0, src1, src2, src3, src4, src5, src6, src7; + + if (0 == height % 12) { + for (cnt = (height / 12); cnt--;) { + LD_UB8(src, src_stride, src0, src1, src2, src3, src4, src5, src6, src7); + src += (8 * src_stride); + + out0 = __msa_copy_u_d((v2i64)src0, 0); + out1 = __msa_copy_u_d((v2i64)src1, 0); + out2 = __msa_copy_u_d((v2i64)src2, 0); + out3 = __msa_copy_u_d((v2i64)src3, 0); + out4 = __msa_copy_u_d((v2i64)src4, 0); + out5 = __msa_copy_u_d((v2i64)src5, 0); + out6 = __msa_copy_u_d((v2i64)src6, 0); + out7 = __msa_copy_u_d((v2i64)src7, 0); + + SD4(out0, out1, out2, out3, dst, dst_stride); + dst += (4 * dst_stride); + SD4(out4, out5, out6, out7, dst, dst_stride); + dst += (4 * dst_stride); + + LD_UB4(src, src_stride, src0, src1, src2, src3); + src += (4 * src_stride); + + out0 = __msa_copy_u_d((v2i64)src0, 0); + out1 = __msa_copy_u_d((v2i64)src1, 0); + out2 = __msa_copy_u_d((v2i64)src2, 0); + out3 = __msa_copy_u_d((v2i64)src3, 0); + SD4(out0, out1, out2, out3, dst, dst_stride); + dst += (4 * dst_stride); + } + } else if (0 == height % 8) { + for (cnt = height >> 3; cnt--;) { + LD_UB8(src, src_stride, src0, src1, src2, src3, src4, src5, src6, src7); + src += (8 * src_stride); + + out0 = __msa_copy_u_d((v2i64)src0, 0); + out1 = __msa_copy_u_d((v2i64)src1, 0); + out2 = __msa_copy_u_d((v2i64)src2, 0); + out3 = __msa_copy_u_d((v2i64)src3, 0); + out4 = __msa_copy_u_d((v2i64)src4, 0); + out5 = __msa_copy_u_d((v2i64)src5, 0); + out6 = __msa_copy_u_d((v2i64)src6, 0); + out7 = __msa_copy_u_d((v2i64)src7, 0); + + SD4(out0, out1, out2, out3, dst, dst_stride); + dst += (4 * dst_stride); + SD4(out4, out5, out6, out7, dst, dst_stride); + dst += (4 * dst_stride); + } + } else if (0 == height % 4) { + for (cnt = (height / 4); cnt--;) { + LD_UB4(src, src_stride, src0, src1, src2, src3); + src += (4 * src_stride); + out0 = __msa_copy_u_d((v2i64)src0, 0); + out1 = __msa_copy_u_d((v2i64)src1, 0); + out2 = __msa_copy_u_d((v2i64)src2, 0); + out3 = __msa_copy_u_d((v2i64)src3, 0); + + SD4(out0, out1, out2, out3, dst, dst_stride); + dst += (4 * dst_stride); + } + } else if (0 == height % 2) { + for (cnt = (height / 2); cnt--;) { + LD_UB2(src, src_stride, src0, src1); + src += (2 * src_stride); + out0 = __msa_copy_u_d((v2i64)src0, 0); + out1 = __msa_copy_u_d((v2i64)src1, 0); + + SD(out0, dst); + dst += dst_stride; + SD(out1, dst); + dst += dst_stride; + } + } +} + +static void copy_16multx8mult_msa(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int32_t height, int32_t width) { + int32_t cnt, loop_cnt; + const uint8_t *src_tmp; + uint8_t *dst_tmp; + v16u8 src0, src1, src2, src3, src4, src5, src6, src7; + + for (cnt = (width >> 4); cnt--;) { + src_tmp = src; + dst_tmp = dst; + + for (loop_cnt = (height >> 3); loop_cnt--;) { + LD_UB8(src_tmp, src_stride, src0, src1, src2, src3, src4, src5, src6, + src7); + src_tmp += (8 * src_stride); + + ST_UB8(src0, src1, src2, src3, src4, src5, src6, src7, dst_tmp, + dst_stride); + dst_tmp += (8 * dst_stride); + } + + src += 16; + dst += 16; + } +} + +static void copy_width16_msa(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, int32_t height) { + int32_t cnt; + v16u8 src0, src1, src2, src3, src4, src5, src6, src7; + + if (0 == height % 12) { + for (cnt = (height / 12); cnt--;) { + LD_UB8(src, src_stride, src0, src1, src2, src3, src4, src5, src6, src7); + src += (8 * src_stride); + ST_UB8(src0, src1, src2, src3, src4, src5, src6, src7, dst, dst_stride); + dst += (8 * dst_stride); + + LD_UB4(src, src_stride, src0, src1, src2, src3); + src += (4 * src_stride); + ST_UB4(src0, src1, src2, src3, dst, dst_stride); + dst += (4 * dst_stride); + } + } else if (0 == height % 8) { + copy_16multx8mult_msa(src, src_stride, dst, dst_stride, height, 16); + } else if (0 == height % 4) { + for (cnt = (height >> 2); cnt--;) { + LD_UB4(src, src_stride, src0, src1, src2, src3); + src += (4 * src_stride); + + ST_UB4(src0, src1, src2, src3, dst, dst_stride); + dst += (4 * dst_stride); + } + } +} + +static void copy_width32_msa(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, int32_t height) { + int32_t cnt; + v16u8 src0, src1, src2, src3, src4, src5, src6, src7; + + if (0 == height % 12) { + for (cnt = (height / 12); cnt--;) { + LD_UB4(src, src_stride, src0, src1, src2, src3); + LD_UB4(src + 16, src_stride, src4, src5, src6, src7); + src += (4 * src_stride); + ST_UB4(src0, src1, src2, src3, dst, dst_stride); + ST_UB4(src4, src5, src6, src7, dst + 16, dst_stride); + dst += (4 * dst_stride); + + LD_UB4(src, src_stride, src0, src1, src2, src3); + LD_UB4(src + 16, src_stride, src4, src5, src6, src7); + src += (4 * src_stride); + ST_UB4(src0, src1, src2, src3, dst, dst_stride); + ST_UB4(src4, src5, src6, src7, dst + 16, dst_stride); + dst += (4 * dst_stride); + + LD_UB4(src, src_stride, src0, src1, src2, src3); + LD_UB4(src + 16, src_stride, src4, src5, src6, src7); + src += (4 * src_stride); + ST_UB4(src0, src1, src2, src3, dst, dst_stride); + ST_UB4(src4, src5, src6, src7, dst + 16, dst_stride); + dst += (4 * dst_stride); + } + } else if (0 == height % 8) { + copy_16multx8mult_msa(src, src_stride, dst, dst_stride, height, 32); + } else if (0 == height % 4) { + for (cnt = (height >> 2); cnt--;) { + LD_UB4(src, src_stride, src0, src1, src2, src3); + LD_UB4(src + 16, src_stride, src4, src5, src6, src7); + src += (4 * src_stride); + ST_UB4(src0, src1, src2, src3, dst, dst_stride); + ST_UB4(src4, src5, src6, src7, dst + 16, dst_stride); + dst += (4 * dst_stride); + } + } +} + +static void copy_width64_msa(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, int32_t height) { + copy_16multx8mult_msa(src, src_stride, dst, dst_stride, height, 64); +} + +void aom_convolve_copy_msa(const uint8_t *src, ptrdiff_t src_stride, + uint8_t *dst, ptrdiff_t dst_stride, + const int16_t *filter_x, int32_t filter_x_stride, + const int16_t *filter_y, int32_t filter_y_stride, + int32_t w, int32_t h) { + (void)filter_x; + (void)filter_y; + (void)filter_x_stride; + (void)filter_y_stride; + + switch (w) { + case 4: { + uint32_t cnt, tmp; + /* 1 word storage */ + for (cnt = h; cnt--;) { + tmp = LW(src); + SW(tmp, dst); + src += src_stride; + dst += dst_stride; + } + break; + } + case 8: { + copy_width8_msa(src, src_stride, dst, dst_stride, h); + break; + } + case 16: { + copy_width16_msa(src, src_stride, dst, dst_stride, h); + break; + } + case 32: { + copy_width32_msa(src, src_stride, dst, dst_stride, h); + break; + } + case 64: { + copy_width64_msa(src, src_stride, dst, dst_stride, h); + break; + } + default: { + uint32_t cnt; + for (cnt = h; cnt--;) { + memcpy(dst, src, w); + src += src_stride; + dst += dst_stride; + } + break; + } + } +} diff --git a/third_party/aom/aom_dsp/mips/aom_convolve_msa.h b/third_party/aom/aom_dsp/mips/aom_convolve_msa.h new file mode 100644 index 0000000000..852415c201 --- /dev/null +++ b/third_party/aom/aom_dsp/mips/aom_convolve_msa.h @@ -0,0 +1,79 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_MIPS_AOM_CONVOLVE_MSA_H_ +#define AOM_AOM_DSP_MIPS_AOM_CONVOLVE_MSA_H_ + +#include "aom_dsp/mips/macros_msa.h" +#include "aom_dsp/aom_filter.h" + +extern const uint8_t mc_filt_mask_arr[16 * 3]; + +#define FILT_8TAP_DPADD_S_H(vec0, vec1, vec2, vec3, filt0, filt1, filt2, \ + filt3) \ + ({ \ + v8i16 tmp_dpadd_0, tmp_dpadd_1; \ + \ + tmp_dpadd_0 = __msa_dotp_s_h((v16i8)vec0, (v16i8)filt0); \ + tmp_dpadd_0 = __msa_dpadd_s_h(tmp_dpadd_0, (v16i8)vec1, (v16i8)filt1); \ + tmp_dpadd_1 = __msa_dotp_s_h((v16i8)vec2, (v16i8)filt2); \ + tmp_dpadd_1 = __msa_dpadd_s_h(tmp_dpadd_1, (v16i8)vec3, (v16i8)filt3); \ + tmp_dpadd_0 = __msa_adds_s_h(tmp_dpadd_0, tmp_dpadd_1); \ + \ + tmp_dpadd_0; \ + }) + +#define HORIZ_8TAP_4WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, \ + mask2, mask3, filt0, filt1, filt2, filt3, \ + out0, out1) \ + { \ + v16i8 vec0_m, vec1_m, vec2_m, vec3_m, vec4_m, vec5_m, vec6_m, vec7_m; \ + v8i16 res0_m, res1_m, res2_m, res3_m; \ + \ + VSHF_B2_SB(src0, src1, src2, src3, mask0, mask0, vec0_m, vec1_m); \ + DOTP_SB2_SH(vec0_m, vec1_m, filt0, filt0, res0_m, res1_m); \ + VSHF_B2_SB(src0, src1, src2, src3, mask1, mask1, vec2_m, vec3_m); \ + DPADD_SB2_SH(vec2_m, vec3_m, filt1, filt1, res0_m, res1_m); \ + VSHF_B2_SB(src0, src1, src2, src3, mask2, mask2, vec4_m, vec5_m); \ + DOTP_SB2_SH(vec4_m, vec5_m, filt2, filt2, res2_m, res3_m); \ + VSHF_B2_SB(src0, src1, src2, src3, mask3, mask3, vec6_m, vec7_m); \ + DPADD_SB2_SH(vec6_m, vec7_m, filt3, filt3, res2_m, res3_m); \ + ADDS_SH2_SH(res0_m, res2_m, res1_m, res3_m, out0, out1); \ + } + +#define HORIZ_8TAP_8WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, \ + mask2, mask3, filt0, filt1, filt2, filt3, \ + out0, out1, out2, out3) \ + { \ + v16i8 vec0_m, vec1_m, vec2_m, vec3_m, vec4_m, vec5_m, vec6_m, vec7_m; \ + v8i16 res0_m, res1_m, res2_m, res3_m, res4_m, res5_m, res6_m, res7_m; \ + \ + VSHF_B2_SB(src0, src0, src1, src1, mask0, mask0, vec0_m, vec1_m); \ + VSHF_B2_SB(src2, src2, src3, src3, mask0, mask0, vec2_m, vec3_m); \ + DOTP_SB4_SH(vec0_m, vec1_m, vec2_m, vec3_m, filt0, filt0, filt0, filt0, \ + res0_m, res1_m, res2_m, res3_m); \ + VSHF_B2_SB(src0, src0, src1, src1, mask2, mask2, vec0_m, vec1_m); \ + VSHF_B2_SB(src2, src2, src3, src3, mask2, mask2, vec2_m, vec3_m); \ + DOTP_SB4_SH(vec0_m, vec1_m, vec2_m, vec3_m, filt2, filt2, filt2, filt2, \ + res4_m, res5_m, res6_m, res7_m); \ + VSHF_B2_SB(src0, src0, src1, src1, mask1, mask1, vec4_m, vec5_m); \ + VSHF_B2_SB(src2, src2, src3, src3, mask1, mask1, vec6_m, vec7_m); \ + DPADD_SB4_SH(vec4_m, vec5_m, vec6_m, vec7_m, filt1, filt1, filt1, filt1, \ + res0_m, res1_m, res2_m, res3_m); \ + VSHF_B2_SB(src0, src0, src1, src1, mask3, mask3, vec4_m, vec5_m); \ + VSHF_B2_SB(src2, src2, src3, src3, mask3, mask3, vec6_m, vec7_m); \ + DPADD_SB4_SH(vec4_m, vec5_m, vec6_m, vec7_m, filt3, filt3, filt3, filt3, \ + res4_m, res5_m, res6_m, res7_m); \ + ADDS_SH4_SH(res0_m, res4_m, res1_m, res5_m, res2_m, res6_m, res3_m, \ + res7_m, out0, out1, out2, out3); \ + } + +#endif // AOM_AOM_DSP_MIPS_AOM_CONVOLVE_MSA_H_ diff --git a/third_party/aom/aom_dsp/mips/common_dspr2.c b/third_party/aom/aom_dsp/mips/common_dspr2.c new file mode 100644 index 0000000000..00ab75dc31 --- /dev/null +++ b/third_party/aom/aom_dsp/mips/common_dspr2.c @@ -0,0 +1,31 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include "aom_dsp/mips/common_dspr2.h" + +#if HAVE_DSPR2 +uint8_t aom_ff_cropTbl_a[256 + 2 * CROP_WIDTH]; +uint8_t *aom_ff_cropTbl; + +void aom_dsputil_static_init(void) { + int i; + + for (i = 0; i < 256; i++) aom_ff_cropTbl_a[i + CROP_WIDTH] = i; + + for (i = 0; i < CROP_WIDTH; i++) { + aom_ff_cropTbl_a[i] = 0; + aom_ff_cropTbl_a[i + CROP_WIDTH + 256] = 255; + } + + aom_ff_cropTbl = &aom_ff_cropTbl_a[CROP_WIDTH]; +} + +#endif diff --git a/third_party/aom/aom_dsp/mips/common_dspr2.h b/third_party/aom/aom_dsp/mips/common_dspr2.h new file mode 100644 index 0000000000..c42188d62d --- /dev/null +++ b/third_party/aom/aom_dsp/mips/common_dspr2.h @@ -0,0 +1,51 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_MIPS_COMMON_DSPR2_H_ +#define AOM_AOM_DSP_MIPS_COMMON_DSPR2_H_ + +#include + +#include "config/aom_config.h" + +#include "aom/aom_integer.h" + +#ifdef __cplusplus +extern "C" { +#endif +#if HAVE_DSPR2 +#define CROP_WIDTH 512 + +extern uint8_t *aom_ff_cropTbl; // From "aom_dsp/mips/intrapred4_dspr2.c" + +static INLINE void prefetch_load(const unsigned char *src) { + __asm__ __volatile__("pref 0, 0(%[src]) \n\t" : : [src] "r"(src)); +} + +/* prefetch data for store */ +static INLINE void prefetch_store(unsigned char *dst) { + __asm__ __volatile__("pref 1, 0(%[dst]) \n\t" : : [dst] "r"(dst)); +} + +static INLINE void prefetch_load_streamed(const unsigned char *src) { + __asm__ __volatile__("pref 4, 0(%[src]) \n\t" : : [src] "r"(src)); +} + +/* prefetch data for store */ +static INLINE void prefetch_store_streamed(unsigned char *dst) { + __asm__ __volatile__("pref 5, 0(%[dst]) \n\t" : : [dst] "r"(dst)); +} +#endif // #if HAVE_DSPR2 +#ifdef __cplusplus +} // extern "C" +#endif + +#endif // AOM_AOM_DSP_MIPS_COMMON_DSPR2_H_ diff --git a/third_party/aom/aom_dsp/mips/convolve2_dspr2.c b/third_party/aom/aom_dsp/mips/convolve2_dspr2.c new file mode 100644 index 0000000000..08bf1ab30f --- /dev/null +++ b/third_party/aom/aom_dsp/mips/convolve2_dspr2.c @@ -0,0 +1,1031 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include +#include + +#include "config/aom_dsp_rtcd.h" + +#include "aom_dsp/mips/convolve_common_dspr2.h" +#include "aom_dsp/aom_dsp_common.h" +#include "aom_dsp/aom_filter.h" +#include "aom_ports/mem.h" + +#if HAVE_DSPR2 +static void convolve_bi_horiz_4_transposed_dspr2( + const uint8_t *src, int32_t src_stride, uint8_t *dst, int32_t dst_stride, + const int16_t *filter_x0, int32_t h) { + int32_t y; + uint8_t *cm = aom_ff_cropTbl; + uint8_t *dst_ptr; + int32_t Temp1, Temp2; + uint32_t vector4a = 64; + uint32_t tp1, tp2; + uint32_t p1, p2; + const int16_t *filter = &filter_x0[3]; + uint32_t filter45; + + filter45 = ((const int32_t *)filter)[0]; + + for (y = h; y--;) { + dst_ptr = dst; + /* prefetch data to cache memory */ + prefetch_load(src + src_stride); + prefetch_load(src + src_stride + 32); + + __asm__ __volatile__( + "ulw %[tp1], 0(%[src]) \n\t" + "ulw %[tp2], 4(%[src]) \n\t" + + /* even 1. pixel */ + "mtlo %[vector4a], $ac3 \n\t" + "mthi $zero, $ac3 \n\t" + "preceu.ph.qbr %[p1], %[tp1] \n\t" + "preceu.ph.qbl %[p2], %[tp1] \n\t" + "dpa.w.ph $ac3, %[p1], %[filter45] \n\t" + "extp %[Temp1], $ac3, 31 \n\t" + + /* even 2. pixel */ + "mtlo %[vector4a], $ac2 \n\t" + "mthi $zero, $ac2 \n\t" + "balign %[tp2], %[tp1], 3 \n\t" + "dpa.w.ph $ac2, %[p2], %[filter45] \n\t" + "extp %[Temp2], $ac2, 31 \n\t" + + /* odd 1. pixel */ + "lbux %[tp1], %[Temp1](%[cm]) \n\t" + "mtlo %[vector4a], $ac3 \n\t" + "mthi $zero, $ac3 \n\t" + "preceu.ph.qbr %[p1], %[tp2] \n\t" + "preceu.ph.qbl %[p2], %[tp2] \n\t" + "dpa.w.ph $ac3, %[p1], %[filter45] \n\t" + "extp %[Temp1], $ac3, 31 \n\t" + + /* odd 2. pixel */ + "lbux %[tp2], %[Temp2](%[cm]) \n\t" + "mtlo %[vector4a], $ac2 \n\t" + "mthi $zero, $ac2 \n\t" + "dpa.w.ph $ac2, %[p2], %[filter45] \n\t" + "extp %[Temp2], $ac2, 31 \n\t" + + /* clamp */ + "lbux %[p1], %[Temp1](%[cm]) \n\t" + "lbux %[p2], %[Temp2](%[cm]) \n\t" + + /* store bytes */ + "sb %[tp1], 0(%[dst_ptr]) \n\t" + "addu %[dst_ptr], %[dst_ptr], %[dst_stride] \n\t" + + "sb %[p1], 0(%[dst_ptr]) \n\t" + "addu %[dst_ptr], %[dst_ptr], %[dst_stride] \n\t" + + "sb %[tp2], 0(%[dst_ptr]) \n\t" + "addu %[dst_ptr], %[dst_ptr], %[dst_stride] \n\t" + + "sb %[p2], 0(%[dst_ptr]) \n\t" + "addu %[dst_ptr], %[dst_ptr], %[dst_stride] \n\t" + + : [tp1] "=&r"(tp1), [tp2] "=&r"(tp2), [p1] "=&r"(p1), [p2] "=&r"(p2), + [Temp1] "=&r"(Temp1), [Temp2] "=&r"(Temp2), [dst_ptr] "+r"(dst_ptr) + : [filter45] "r"(filter45), [vector4a] "r"(vector4a), [cm] "r"(cm), + [src] "r"(src), [dst_stride] "r"(dst_stride)); + + /* Next row... */ + src += src_stride; + dst += 1; + } +} + +static void convolve_bi_horiz_8_transposed_dspr2( + const uint8_t *src, int32_t src_stride, uint8_t *dst, int32_t dst_stride, + const int16_t *filter_x0, int32_t h) { + int32_t y; + uint8_t *cm = aom_ff_cropTbl; + uint8_t *dst_ptr; + uint32_t vector4a = 64; + int32_t Temp1, Temp2, Temp3; + uint32_t tp1, tp2, tp3; + uint32_t p1, p2, p3, p4; + uint8_t *odd_dst; + uint32_t dst_pitch_2 = (dst_stride << 1); + const int16_t *filter = &filter_x0[3]; + uint32_t filter45; + + filter45 = ((const int32_t *)filter)[0]; + + for (y = h; y--;) { + /* prefetch data to cache memory */ + prefetch_load(src + src_stride); + prefetch_load(src + src_stride + 32); + + dst_ptr = dst; + odd_dst = (dst_ptr + dst_stride); + + __asm__ __volatile__( + "ulw %[tp1], 0(%[src]) \n\t" + "ulw %[tp2], 4(%[src]) \n\t" + + /* even 1. pixel */ + "mtlo %[vector4a], $ac3 \n\t" + "mthi $zero, $ac3 \n\t" + "mtlo %[vector4a], $ac2 \n\t" + "mthi $zero, $ac2 \n\t" + "preceu.ph.qbr %[p1], %[tp1] \n\t" + "preceu.ph.qbl %[p2], %[tp1] \n\t" + "preceu.ph.qbr %[p3], %[tp2] \n\t" + "preceu.ph.qbl %[p4], %[tp2] \n\t" + "ulw %[tp3], 8(%[src]) \n\t" + "dpa.w.ph $ac3, %[p1], %[filter45] \n\t" + "extp %[Temp1], $ac3, 31 \n\t" + + /* even 2. pixel */ + "dpa.w.ph $ac2, %[p2], %[filter45] \n\t" + "extp %[Temp3], $ac2, 31 \n\t" + + /* even 3. pixel */ + "lbux %[Temp2], %[Temp1](%[cm]) \n\t" + "mtlo %[vector4a], $ac1 \n\t" + "mthi $zero, $ac1 \n\t" + "balign %[tp3], %[tp2], 3 \n\t" + "balign %[tp2], %[tp1], 3 \n\t" + "dpa.w.ph $ac1, %[p3], %[filter45] \n\t" + "lbux %[tp1], %[Temp3](%[cm]) \n\t" + "extp %[p3], $ac1, 31 \n\t" + + /* even 4. pixel */ + "mtlo %[vector4a], $ac2 \n\t" + "mthi $zero, $ac2 \n\t" + "mtlo %[vector4a], $ac3 \n\t" + "mthi $zero, $ac3 \n\t" + "sb %[Temp2], 0(%[dst_ptr]) \n\t" + "addu %[dst_ptr], %[dst_ptr], %[dst_pitch_2] \n\t" + "sb %[tp1], 0(%[dst_ptr]) \n\t" + "addu %[dst_ptr], %[dst_ptr], %[dst_pitch_2] \n\t" + + "dpa.w.ph $ac2, %[p4], %[filter45] \n\t" + "extp %[Temp3], $ac2, 31 \n\t" + + "lbux %[Temp1], %[p3](%[cm]) " + "\n\t" + + /* odd 1. pixel */ + "mtlo %[vector4a], $ac1 \n\t" + "mthi $zero, $ac1 \n\t" + "preceu.ph.qbr %[p1], %[tp2] \n\t" + "preceu.ph.qbl %[p2], %[tp2] \n\t" + "preceu.ph.qbr %[p3], %[tp3] \n\t" + "preceu.ph.qbl %[p4], %[tp3] \n\t" + "sb %[Temp1], 0(%[dst_ptr]) \n\t" + "addu %[dst_ptr], %[dst_ptr], %[dst_pitch_2] \n\t" + + "dpa.w.ph $ac3, %[p1], %[filter45] \n\t" + "extp %[Temp2], $ac3, 31 \n\t" + + /* odd 2. pixel */ + "lbux %[tp1], %[Temp3](%[cm]) \n\t" + "mtlo %[vector4a], $ac3 \n\t" + "mthi $zero, $ac3 \n\t" + "mtlo %[vector4a], $ac2 \n\t" + "mthi $zero, $ac2 \n\t" + "dpa.w.ph $ac1, %[p2], %[filter45] \n\t" + "sb %[tp1], 0(%[dst_ptr]) \n\t" + "addu %[dst_ptr], %[dst_ptr], %[dst_pitch_2] \n\t" + "extp %[Temp3], $ac1, 31 \n\t" + + /* odd 3. pixel */ + "lbux %[tp3], %[Temp2](%[cm]) \n\t" + "dpa.w.ph $ac3, %[p3], %[filter45] \n\t" + "extp %[Temp2], $ac3, 31 \n\t" + + /* odd 4. pixel */ + "sb %[tp3], 0(%[odd_dst]) \n\t" + "addu %[odd_dst], %[odd_dst], %[dst_pitch_2] \n\t" + "dpa.w.ph $ac2, %[p4], %[filter45] \n\t" + "extp %[Temp1], $ac2, 31 \n\t" + + /* clamp */ + "lbux %[p4], %[Temp3](%[cm]) \n\t" + "lbux %[p2], %[Temp2](%[cm]) \n\t" + "lbux %[p1], %[Temp1](%[cm]) \n\t" + + /* store bytes */ + "sb %[p4], 0(%[odd_dst]) \n\t" + "addu %[odd_dst], %[odd_dst], %[dst_pitch_2] \n\t" + + "sb %[p2], 0(%[odd_dst]) \n\t" + "addu %[odd_dst], %[odd_dst], %[dst_pitch_2] \n\t" + + "sb %[p1], 0(%[odd_dst]) \n\t" + + : [tp1] "=&r"(tp1), [tp2] "=&r"(tp2), [tp3] "=&r"(tp3), [p1] "=&r"(p1), + [p2] "=&r"(p2), [p3] "=&r"(p3), [p4] "=&r"(p4), [Temp1] "=&r"(Temp1), + [Temp2] "=&r"(Temp2), [Temp3] "=&r"(Temp3), [dst_ptr] "+r"(dst_ptr), + [odd_dst] "+r"(odd_dst) + : [filter45] "r"(filter45), [vector4a] "r"(vector4a), [cm] "r"(cm), + [src] "r"(src), [dst_pitch_2] "r"(dst_pitch_2)); + + /* Next row... */ + src += src_stride; + dst += 1; + } +} + +static void convolve_bi_horiz_16_transposed_dspr2( + const uint8_t *src_ptr, int32_t src_stride, uint8_t *dst_ptr, + int32_t dst_stride, const int16_t *filter_x0, int32_t h, int32_t count) { + int32_t c, y; + const uint8_t *src; + uint8_t *dst; + uint8_t *cm = aom_ff_cropTbl; + uint32_t vector_64 = 64; + int32_t Temp1, Temp2, Temp3; + uint32_t qload1, qload2; + uint32_t p1, p2, p3, p4, p5; + uint32_t st1, st2, st3; + uint32_t dst_pitch_2 = (dst_stride << 1); + uint8_t *odd_dst; + const int16_t *filter = &filter_x0[3]; + uint32_t filter45; + + filter45 = ((const int32_t *)filter)[0]; + + for (y = h; y--;) { + /* prefetch data to cache memory */ + prefetch_load(src_ptr + src_stride); + prefetch_load(src_ptr + src_stride + 32); + + src = src_ptr; + dst = dst_ptr; + + odd_dst = (dst + dst_stride); + + for (c = 0; c < count; c++) { + __asm__ __volatile__( + "ulw %[qload1], 0(%[src]) " + "\n\t" + "ulw %[qload2], 4(%[src]) " + "\n\t" + + /* even 1. pixel */ + "mtlo %[vector_64], $ac1 " + "\n\t" /* even 1 */ + "mthi $zero, $ac1 " + "\n\t" + "mtlo %[vector_64], $ac2 " + "\n\t" /* even 2 */ + "mthi $zero, $ac2 " + "\n\t" + "preceu.ph.qbr %[p1], %[qload1] " + "\n\t" + "preceu.ph.qbl %[p2], %[qload1] " + "\n\t" + "preceu.ph.qbr %[p3], %[qload2] " + "\n\t" + "preceu.ph.qbl %[p4], %[qload2] " + "\n\t" + "ulw %[qload1], 8(%[src]) " + "\n\t" + "dpa.w.ph $ac1, %[p1], %[filter45] " + "\n\t" /* even 1 */ + "extp %[Temp1], $ac1, 31 " + "\n\t" /* even 1 */ + + /* even 2. pixel */ + "mtlo %[vector_64], $ac3 " + "\n\t" /* even 3 */ + "mthi $zero, $ac3 " + "\n\t" + "preceu.ph.qbr %[p1], %[qload1] " + "\n\t" + "preceu.ph.qbl %[p5], %[qload1] " + "\n\t" + "ulw %[qload2], 12(%[src]) " + "\n\t" + "dpa.w.ph $ac2, %[p2], %[filter45] " + "\n\t" /* even 1 */ + "lbux %[st1], %[Temp1](%[cm]) " + "\n\t" /* even 1 */ + "extp %[Temp2], $ac2, 31 " + "\n\t" /* even 1 */ + + /* even 3. pixel */ + "mtlo %[vector_64], $ac1 " + "\n\t" /* even 4 */ + "mthi $zero, $ac1 " + "\n\t" + "preceu.ph.qbr %[p2], %[qload2] " + "\n\t" + "sb %[st1], 0(%[dst]) " + "\n\t" /* even 1 */ + "addu %[dst], %[dst], %[dst_pitch_2] " + " \n\t" + "dpa.w.ph $ac3, %[p3], %[filter45] " + "\n\t" /* even 3 */ + "extp %[Temp3], $ac3, 31 " + "\n\t" /* even 3 */ + "lbux %[st2], %[Temp2](%[cm]) " + "\n\t" /* even 1 */ + + /* even 4. pixel */ + "mtlo %[vector_64], $ac2 " + "\n\t" /* even 5 */ + "mthi $zero, $ac2 " + "\n\t" + "preceu.ph.qbl %[p3], %[qload2] " + "\n\t" + "sb %[st2], 0(%[dst]) " + "\n\t" /* even 2 */ + "addu %[dst], %[dst], %[dst_pitch_2] " + "\n\t" + "dpa.w.ph $ac1, %[p4], %[filter45] " + "\n\t" /* even 4 */ + "extp %[Temp1], $ac1, 31 " + "\n\t" /* even 4 */ + "lbux %[st3], %[Temp3](%[cm]) " + "\n\t" /* even 3 */ + + /* even 5. pixel */ + "mtlo %[vector_64], $ac3 " + "\n\t" /* even 6 */ + "mthi $zero, $ac3 " + "\n\t" + "sb %[st3], 0(%[dst]) " + "\n\t" /* even 3 */ + "addu %[dst], %[dst], %[dst_pitch_2] " + "\n\t" + "dpa.w.ph $ac2, %[p1], %[filter45] " + "\n\t" /* even 5 */ + "extp %[Temp2], $ac2, 31 " + "\n\t" /* even 5 */ + "lbux %[st1], %[Temp1](%[cm]) " + "\n\t" /* even 4 */ + + /* even 6. pixel */ + "mtlo %[vector_64], $ac1 " + "\n\t" /* even 7 */ + "mthi $zero, $ac1 " + "\n\t" + "sb %[st1], 0(%[dst]) " + "\n\t" /* even 4 */ + "addu %[dst], %[dst], %[dst_pitch_2] " + "\n\t" + "ulw %[qload1], 20(%[src]) " + "\n\t" + "dpa.w.ph $ac3, %[p5], %[filter45] " + "\n\t" /* even 6 */ + "extp %[Temp3], $ac3, 31 " + "\n\t" /* even 6 */ + "lbux %[st2], %[Temp2](%[cm]) " + "\n\t" /* even 5 */ + + /* even 7. pixel */ + "mtlo %[vector_64], $ac2 " + "\n\t" /* even 8 */ + "mthi $zero, $ac2 " + "\n\t" + "preceu.ph.qbr %[p5], %[qload1] " + "\n\t" + "sb %[st2], 0(%[dst]) " + "\n\t" /* even 5 */ + "addu %[dst], %[dst], %[dst_pitch_2] " + "\n\t" + "dpa.w.ph $ac1, %[p2], %[filter45] " + "\n\t" /* even 7 */ + "extp %[Temp1], $ac1, 31 " + "\n\t" /* even 7 */ + "lbux %[st3], %[Temp3](%[cm]) " + "\n\t" /* even 6 */ + + /* even 8. pixel */ + "mtlo %[vector_64], $ac3 " + "\n\t" /* odd 1 */ + "mthi $zero, $ac3 " + "\n\t" + "dpa.w.ph $ac2, %[p3], %[filter45] " + "\n\t" /* even 8 */ + "sb %[st3], 0(%[dst]) " + "\n\t" /* even 6 */ + "addu %[dst], %[dst], %[dst_pitch_2] " + "\n\t" + "extp %[Temp2], $ac2, 31 " + "\n\t" /* even 8 */ + "lbux %[st1], %[Temp1](%[cm]) " + "\n\t" /* even 7 */ + + /* ODD pixels */ + "ulw %[qload1], 1(%[src]) " + "\n\t" + "ulw %[qload2], 5(%[src]) " + "\n\t" + + /* odd 1. pixel */ + "mtlo %[vector_64], $ac1 " + "\n\t" /* odd 2 */ + "mthi $zero, $ac1 " + "\n\t" + "preceu.ph.qbr %[p1], %[qload1] " + "\n\t" + "preceu.ph.qbl %[p2], %[qload1] " + "\n\t" + "preceu.ph.qbr %[p3], %[qload2] " + "\n\t" + "preceu.ph.qbl %[p4], %[qload2] " + "\n\t" + "sb %[st1], 0(%[dst]) " + "\n\t" /* even 7 */ + "addu %[dst], %[dst], %[dst_pitch_2] " + "\n\t" + "ulw %[qload2], 9(%[src]) " + "\n\t" + "dpa.w.ph $ac3, %[p1], %[filter45] " + "\n\t" /* odd 1 */ + "extp %[Temp3], $ac3, 31 " + "\n\t" /* odd 1 */ + "lbux %[st2], %[Temp2](%[cm]) " + "\n\t" /* even 8 */ + + /* odd 2. pixel */ + "mtlo %[vector_64], $ac2 " + "\n\t" /* odd 3 */ + "mthi $zero, $ac2 " + "\n\t" + "preceu.ph.qbr %[p1], %[qload2] " + "\n\t" + "preceu.ph.qbl %[p5], %[qload2] " + "\n\t" + "sb %[st2], 0(%[dst]) " + "\n\t" /* even 8 */ + "ulw %[qload1], 13(%[src]) " + "\n\t" + "dpa.w.ph $ac1, %[p2], %[filter45] " + "\n\t" /* odd 2 */ + "extp %[Temp1], $ac1, 31 " + "\n\t" /* odd 2 */ + "lbux %[st3], %[Temp3](%[cm]) " + "\n\t" /* odd 1 */ + + /* odd 3. pixel */ + "mtlo %[vector_64], $ac3 " + "\n\t" /* odd 4 */ + "mthi $zero, $ac3 " + "\n\t" + "preceu.ph.qbr %[p2], %[qload1] " + "\n\t" + "sb %[st3], 0(%[odd_dst]) " + "\n\t" /* odd 1 */ + "addu %[odd_dst], %[odd_dst], %[dst_pitch_2] " + "\n\t" + "dpa.w.ph $ac2, %[p3], %[filter45] " + "\n\t" /* odd 3 */ + "extp %[Temp2], $ac2, 31 " + "\n\t" /* odd 3 */ + "lbux %[st1], %[Temp1](%[cm]) " + "\n\t" /* odd 2 */ + + /* odd 4. pixel */ + "mtlo %[vector_64], $ac1 " + "\n\t" /* odd 5 */ + "mthi $zero, $ac1 " + "\n\t" + "preceu.ph.qbl %[p3], %[qload1] " + "\n\t" + "sb %[st1], 0(%[odd_dst]) " + "\n\t" /* odd 2 */ + "addu %[odd_dst], %[odd_dst], %[dst_pitch_2] " + "\n\t" + "dpa.w.ph $ac3, %[p4], %[filter45] " + "\n\t" /* odd 4 */ + "extp %[Temp3], $ac3, 31 " + "\n\t" /* odd 4 */ + "lbux %[st2], %[Temp2](%[cm]) " + "\n\t" /* odd 3 */ + + /* odd 5. pixel */ + "mtlo %[vector_64], $ac2 " + "\n\t" /* odd 6 */ + "mthi $zero, $ac2 " + "\n\t" + "sb %[st2], 0(%[odd_dst]) " + "\n\t" /* odd 3 */ + "addu %[odd_dst], %[odd_dst], %[dst_pitch_2] " + "\n\t" + "dpa.w.ph $ac1, %[p1], %[filter45] " + "\n\t" /* odd 5 */ + "extp %[Temp1], $ac1, 31 " + "\n\t" /* odd 5 */ + "lbux %[st3], %[Temp3](%[cm]) " + "\n\t" /* odd 4 */ + + /* odd 6. pixel */ + "mtlo %[vector_64], $ac3 " + "\n\t" /* odd 7 */ + "mthi $zero, $ac3 " + "\n\t" + "sb %[st3], 0(%[odd_dst]) " + "\n\t" /* odd 4 */ + "addu %[odd_dst], %[odd_dst], %[dst_pitch_2] " + "\n\t" + "ulw %[qload1], 21(%[src]) " + "\n\t" + "dpa.w.ph $ac2, %[p5], %[filter45] " + "\n\t" /* odd 6 */ + "extp %[Temp2], $ac2, 31 " + "\n\t" /* odd 6 */ + "lbux %[st1], %[Temp1](%[cm]) " + "\n\t" /* odd 5 */ + + /* odd 7. pixel */ + "mtlo %[vector_64], $ac1 " + "\n\t" /* odd 8 */ + "mthi $zero, $ac1 " + "\n\t" + "preceu.ph.qbr %[p5], %[qload1] " + "\n\t" + "sb %[st1], 0(%[odd_dst]) " + "\n\t" /* odd 5 */ + "addu %[odd_dst], %[odd_dst], %[dst_pitch_2] " + "\n\t" + "dpa.w.ph $ac3, %[p2], %[filter45] " + "\n\t" /* odd 7 */ + "extp %[Temp3], $ac3, 31 " + "\n\t" /* odd 7 */ + + /* odd 8. pixel */ + "dpa.w.ph $ac1, %[p3], %[filter45] " + "\n\t" /* odd 8 */ + "extp %[Temp1], $ac1, 31 " + "\n\t" /* odd 8 */ + + "lbux %[st2], %[Temp2](%[cm]) " + "\n\t" /* odd 6 */ + "lbux %[st3], %[Temp3](%[cm]) " + "\n\t" /* odd 7 */ + "lbux %[st1], %[Temp1](%[cm]) " + "\n\t" /* odd 8 */ + + "sb %[st2], 0(%[odd_dst]) " + "\n\t" /* odd 6 */ + "addu %[odd_dst], %[odd_dst], %[dst_pitch_2] " + "\n\t" + + "sb %[st3], 0(%[odd_dst]) " + "\n\t" /* odd 7 */ + "addu %[odd_dst], %[odd_dst], %[dst_pitch_2] " + "\n\t" + + "sb %[st1], 0(%[odd_dst]) " + "\n\t" /* odd 8 */ + + : [qload1] "=&r"(qload1), [qload2] "=&r"(qload2), [p5] "=&r"(p5), + [st1] "=&r"(st1), [st2] "=&r"(st2), [st3] "=&r"(st3), + [p1] "=&r"(p1), [p2] "=&r"(p2), [p3] "=&r"(p3), [p4] "=&r"(p4), + [Temp1] "=&r"(Temp1), [Temp2] "=&r"(Temp2), [Temp3] "=&r"(Temp3), + [dst] "+r"(dst), [odd_dst] "+r"(odd_dst) + : [filter45] "r"(filter45), [vector_64] "r"(vector_64), [cm] "r"(cm), + [src] "r"(src), [dst_pitch_2] "r"(dst_pitch_2)); + + src += 16; + dst = (dst_ptr + ((c + 1) * 16 * dst_stride)); + odd_dst = (dst + dst_stride); + } + + /* Next row... */ + src_ptr += src_stride; + dst_ptr += 1; + } +} + +static void convolve_bi_horiz_64_transposed_dspr2( + const uint8_t *src_ptr, int32_t src_stride, uint8_t *dst_ptr, + int32_t dst_stride, const int16_t *filter_x0, int32_t h) { + int32_t c, y; + const uint8_t *src; + uint8_t *dst; + uint8_t *cm = aom_ff_cropTbl; + uint32_t vector_64 = 64; + int32_t Temp1, Temp2, Temp3; + uint32_t qload1, qload2; + uint32_t p1, p2, p3, p4, p5; + uint32_t st1, st2, st3; + uint32_t dst_pitch_2 = (dst_stride << 1); + uint8_t *odd_dst; + const int16_t *filter = &filter_x0[3]; + uint32_t filter45; + + filter45 = ((const int32_t *)filter)[0]; + + for (y = h; y--;) { + /* prefetch data to cache memory */ + prefetch_load(src_ptr + src_stride); + prefetch_load(src_ptr + src_stride + 32); + prefetch_load(src_ptr + src_stride + 64); + + src = src_ptr; + dst = dst_ptr; + + odd_dst = (dst + dst_stride); + + for (c = 0; c < 4; c++) { + __asm__ __volatile__( + "ulw %[qload1], 0(%[src]) " + "\n\t" + "ulw %[qload2], 4(%[src]) " + "\n\t" + + /* even 1. pixel */ + "mtlo %[vector_64], $ac1 " + "\n\t" /* even 1 */ + "mthi $zero, $ac1 " + "\n\t" + "mtlo %[vector_64], $ac2 " + "\n\t" /* even 2 */ + "mthi $zero, $ac2 " + "\n\t" + "preceu.ph.qbr %[p1], %[qload1] " + "\n\t" + "preceu.ph.qbl %[p2], %[qload1] " + "\n\t" + "preceu.ph.qbr %[p3], %[qload2] " + "\n\t" + "preceu.ph.qbl %[p4], %[qload2] " + "\n\t" + "ulw %[qload1], 8(%[src]) " + "\n\t" + "dpa.w.ph $ac1, %[p1], %[filter45] " + "\n\t" /* even 1 */ + "extp %[Temp1], $ac1, 31 " + "\n\t" /* even 1 */ + + /* even 2. pixel */ + "mtlo %[vector_64], $ac3 " + "\n\t" /* even 3 */ + "mthi $zero, $ac3 " + "\n\t" + "preceu.ph.qbr %[p1], %[qload1] " + "\n\t" + "preceu.ph.qbl %[p5], %[qload1] " + "\n\t" + "ulw %[qload2], 12(%[src]) " + "\n\t" + "dpa.w.ph $ac2, %[p2], %[filter45] " + "\n\t" /* even 1 */ + "lbux %[st1], %[Temp1](%[cm]) " + "\n\t" /* even 1 */ + "extp %[Temp2], $ac2, 31 " + "\n\t" /* even 1 */ + + /* even 3. pixel */ + "mtlo %[vector_64], $ac1 " + "\n\t" /* even 4 */ + "mthi $zero, $ac1 " + "\n\t" + "preceu.ph.qbr %[p2], %[qload2] " + "\n\t" + "sb %[st1], 0(%[dst]) " + "\n\t" /* even 1 */ + "addu %[dst], %[dst], %[dst_pitch_2] " + " \n\t" + "dpa.w.ph $ac3, %[p3], %[filter45] " + "\n\t" /* even 3 */ + "extp %[Temp3], $ac3, 31 " + "\n\t" /* even 3 */ + "lbux %[st2], %[Temp2](%[cm]) " + "\n\t" /* even 1 */ + + /* even 4. pixel */ + "mtlo %[vector_64], $ac2 " + "\n\t" /* even 5 */ + "mthi $zero, $ac2 " + "\n\t" + "preceu.ph.qbl %[p3], %[qload2] " + "\n\t" + "sb %[st2], 0(%[dst]) " + "\n\t" /* even 2 */ + "addu %[dst], %[dst], %[dst_pitch_2] " + "\n\t" + "dpa.w.ph $ac1, %[p4], %[filter45] " + "\n\t" /* even 4 */ + "extp %[Temp1], $ac1, 31 " + "\n\t" /* even 4 */ + "lbux %[st3], %[Temp3](%[cm]) " + "\n\t" /* even 3 */ + + /* even 5. pixel */ + "mtlo %[vector_64], $ac3 " + "\n\t" /* even 6 */ + "mthi $zero, $ac3 " + "\n\t" + "sb %[st3], 0(%[dst]) " + "\n\t" /* even 3 */ + "addu %[dst], %[dst], %[dst_pitch_2] " + "\n\t" + "dpa.w.ph $ac2, %[p1], %[filter45] " + "\n\t" /* even 5 */ + "extp %[Temp2], $ac2, 31 " + "\n\t" /* even 5 */ + "lbux %[st1], %[Temp1](%[cm]) " + "\n\t" /* even 4 */ + + /* even 6. pixel */ + "mtlo %[vector_64], $ac1 " + "\n\t" /* even 7 */ + "mthi $zero, $ac1 " + "\n\t" + "sb %[st1], 0(%[dst]) " + "\n\t" /* even 4 */ + "addu %[dst], %[dst], %[dst_pitch_2] " + "\n\t" + "ulw %[qload1], 20(%[src]) " + "\n\t" + "dpa.w.ph $ac3, %[p5], %[filter45] " + "\n\t" /* even 6 */ + "extp %[Temp3], $ac3, 31 " + "\n\t" /* even 6 */ + "lbux %[st2], %[Temp2](%[cm]) " + "\n\t" /* even 5 */ + + /* even 7. pixel */ + "mtlo %[vector_64], $ac2 " + "\n\t" /* even 8 */ + "mthi $zero, $ac2 " + "\n\t" + "preceu.ph.qbr %[p5], %[qload1] " + "\n\t" + "sb %[st2], 0(%[dst]) " + "\n\t" /* even 5 */ + "addu %[dst], %[dst], %[dst_pitch_2] " + "\n\t" + "dpa.w.ph $ac1, %[p2], %[filter45] " + "\n\t" /* even 7 */ + "extp %[Temp1], $ac1, 31 " + "\n\t" /* even 7 */ + "lbux %[st3], %[Temp3](%[cm]) " + "\n\t" /* even 6 */ + + /* even 8. pixel */ + "mtlo %[vector_64], $ac3 " + "\n\t" /* odd 1 */ + "mthi $zero, $ac3 " + "\n\t" + "dpa.w.ph $ac2, %[p3], %[filter45] " + "\n\t" /* even 8 */ + "sb %[st3], 0(%[dst]) " + "\n\t" /* even 6 */ + "addu %[dst], %[dst], %[dst_pitch_2] " + "\n\t" + "extp %[Temp2], $ac2, 31 " + "\n\t" /* even 8 */ + "lbux %[st1], %[Temp1](%[cm]) " + "\n\t" /* even 7 */ + + /* ODD pixels */ + "ulw %[qload1], 1(%[src]) " + "\n\t" + "ulw %[qload2], 5(%[src]) " + "\n\t" + + /* odd 1. pixel */ + "mtlo %[vector_64], $ac1 " + "\n\t" /* odd 2 */ + "mthi $zero, $ac1 " + "\n\t" + "preceu.ph.qbr %[p1], %[qload1] " + "\n\t" + "preceu.ph.qbl %[p2], %[qload1] " + "\n\t" + "preceu.ph.qbr %[p3], %[qload2] " + "\n\t" + "preceu.ph.qbl %[p4], %[qload2] " + "\n\t" + "sb %[st1], 0(%[dst]) " + "\n\t" /* even 7 */ + "addu %[dst], %[dst], %[dst_pitch_2] " + "\n\t" + "ulw %[qload2], 9(%[src]) " + "\n\t" + "dpa.w.ph $ac3, %[p1], %[filter45] " + "\n\t" /* odd 1 */ + "extp %[Temp3], $ac3, 31 " + "\n\t" /* odd 1 */ + "lbux %[st2], %[Temp2](%[cm]) " + "\n\t" /* even 8 */ + + /* odd 2. pixel */ + "mtlo %[vector_64], $ac2 " + "\n\t" /* odd 3 */ + "mthi $zero, $ac2 " + "\n\t" + "preceu.ph.qbr %[p1], %[qload2] " + "\n\t" + "preceu.ph.qbl %[p5], %[qload2] " + "\n\t" + "sb %[st2], 0(%[dst]) " + "\n\t" /* even 8 */ + "ulw %[qload1], 13(%[src]) " + "\n\t" + "dpa.w.ph $ac1, %[p2], %[filter45] " + "\n\t" /* odd 2 */ + "extp %[Temp1], $ac1, 31 " + "\n\t" /* odd 2 */ + "lbux %[st3], %[Temp3](%[cm]) " + "\n\t" /* odd 1 */ + + /* odd 3. pixel */ + "mtlo %[vector_64], $ac3 " + "\n\t" /* odd 4 */ + "mthi $zero, $ac3 " + "\n\t" + "preceu.ph.qbr %[p2], %[qload1] " + "\n\t" + "sb %[st3], 0(%[odd_dst]) " + "\n\t" /* odd 1 */ + "addu %[odd_dst], %[odd_dst], %[dst_pitch_2] " + "\n\t" + "dpa.w.ph $ac2, %[p3], %[filter45] " + "\n\t" /* odd 3 */ + "extp %[Temp2], $ac2, 31 " + "\n\t" /* odd 3 */ + "lbux %[st1], %[Temp1](%[cm]) " + "\n\t" /* odd 2 */ + + /* odd 4. pixel */ + "mtlo %[vector_64], $ac1 " + "\n\t" /* odd 5 */ + "mthi $zero, $ac1 " + "\n\t" + "preceu.ph.qbl %[p3], %[qload1] " + "\n\t" + "sb %[st1], 0(%[odd_dst]) " + "\n\t" /* odd 2 */ + "addu %[odd_dst], %[odd_dst], %[dst_pitch_2] " + "\n\t" + "dpa.w.ph $ac3, %[p4], %[filter45] " + "\n\t" /* odd 4 */ + "extp %[Temp3], $ac3, 31 " + "\n\t" /* odd 4 */ + "lbux %[st2], %[Temp2](%[cm]) " + "\n\t" /* odd 3 */ + + /* odd 5. pixel */ + "mtlo %[vector_64], $ac2 " + "\n\t" /* odd 6 */ + "mthi $zero, $ac2 " + "\n\t" + "sb %[st2], 0(%[odd_dst]) " + "\n\t" /* odd 3 */ + "addu %[odd_dst], %[odd_dst], %[dst_pitch_2] " + "\n\t" + "dpa.w.ph $ac1, %[p1], %[filter45] " + "\n\t" /* odd 5 */ + "extp %[Temp1], $ac1, 31 " + "\n\t" /* odd 5 */ + "lbux %[st3], %[Temp3](%[cm]) " + "\n\t" /* odd 4 */ + + /* odd 6. pixel */ + "mtlo %[vector_64], $ac3 " + "\n\t" /* odd 7 */ + "mthi $zero, $ac3 " + "\n\t" + "sb %[st3], 0(%[odd_dst]) " + "\n\t" /* odd 4 */ + "addu %[odd_dst], %[odd_dst], %[dst_pitch_2] " + "\n\t" + "ulw %[qload1], 21(%[src]) " + "\n\t" + "dpa.w.ph $ac2, %[p5], %[filter45] " + "\n\t" /* odd 6 */ + "extp %[Temp2], $ac2, 31 " + "\n\t" /* odd 6 */ + "lbux %[st1], %[Temp1](%[cm]) " + "\n\t" /* odd 5 */ + + /* odd 7. pixel */ + "mtlo %[vector_64], $ac1 " + "\n\t" /* odd 8 */ + "mthi $zero, $ac1 " + "\n\t" + "preceu.ph.qbr %[p5], %[qload1] " + "\n\t" + "sb %[st1], 0(%[odd_dst]) " + "\n\t" /* odd 5 */ + "addu %[odd_dst], %[odd_dst], %[dst_pitch_2] " + "\n\t" + "dpa.w.ph $ac3, %[p2], %[filter45] " + "\n\t" /* odd 7 */ + "extp %[Temp3], $ac3, 31 " + "\n\t" /* odd 7 */ + + /* odd 8. pixel */ + "dpa.w.ph $ac1, %[p3], %[filter45] " + "\n\t" /* odd 8 */ + "extp %[Temp1], $ac1, 31 " + "\n\t" /* odd 8 */ + + "lbux %[st2], %[Temp2](%[cm]) " + "\n\t" /* odd 6 */ + "lbux %[st3], %[Temp3](%[cm]) " + "\n\t" /* odd 7 */ + "lbux %[st1], %[Temp1](%[cm]) " + "\n\t" /* odd 8 */ + + "sb %[st2], 0(%[odd_dst]) " + "\n\t" /* odd 6 */ + "addu %[odd_dst], %[odd_dst], %[dst_pitch_2] " + "\n\t" + + "sb %[st3], 0(%[odd_dst]) " + "\n\t" /* odd 7 */ + "addu %[odd_dst], %[odd_dst], %[dst_pitch_2] " + "\n\t" + + "sb %[st1], 0(%[odd_dst]) " + "\n\t" /* odd 8 */ + + : [qload1] "=&r"(qload1), [qload2] "=&r"(qload2), [p5] "=&r"(p5), + [st1] "=&r"(st1), [st2] "=&r"(st2), [st3] "=&r"(st3), + [p1] "=&r"(p1), [p2] "=&r"(p2), [p3] "=&r"(p3), [p4] "=&r"(p4), + [Temp1] "=&r"(Temp1), [Temp2] "=&r"(Temp2), [Temp3] "=&r"(Temp3), + [dst] "+r"(dst), [odd_dst] "+r"(odd_dst) + : [filter45] "r"(filter45), [vector_64] "r"(vector_64), [cm] "r"(cm), + [src] "r"(src), [dst_pitch_2] "r"(dst_pitch_2)); + + src += 16; + dst = (dst_ptr + ((c + 1) * 16 * dst_stride)); + odd_dst = (dst + dst_stride); + } + + /* Next row... */ + src_ptr += src_stride; + dst_ptr += 1; + } +} + +void convolve_bi_horiz_transposed(const uint8_t *src, ptrdiff_t src_stride, + uint8_t *dst, ptrdiff_t dst_stride, + const int16_t *filter, int w, int h) { + int x, y; + + for (y = 0; y < h; ++y) { + for (x = 0; x < w; ++x) { + int sum = 0; + + sum += src[x] * filter[3]; + sum += src[x + 1] * filter[4]; + + dst[x * dst_stride] = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS)); + } + + src += src_stride; + dst += 1; + } +} + +void aom_convolve2_dspr2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, + ptrdiff_t dst_stride, const int16_t *filter, int w, + int h) { + uint32_t pos = 38; + + /* bit positon for extract from acc */ + __asm__ __volatile__("wrdsp %[pos], 1 \n\t" + : + : [pos] "r"(pos)); + + /* prefetch data to cache memory */ + prefetch_load(src); + prefetch_load(src + 32); + + switch (w) { + case 4: + convolve_bi_horiz_4_transposed_dspr2(src, src_stride, dst, dst_stride, + filter, h); + break; + case 8: + convolve_bi_horiz_8_transposed_dspr2(src, src_stride, dst, dst_stride, + filter, h); + break; + case 16: + case 32: + convolve_bi_horiz_16_transposed_dspr2(src, src_stride, dst, dst_stride, + filter, h, (w / 16)); + break; + case 64: + prefetch_load(src + 32); + convolve_bi_horiz_64_transposed_dspr2(src, src_stride, dst, dst_stride, + filter, h); + break; + default: + convolve_bi_horiz_transposed(src, src_stride, dst, dst_stride, filter, w, + h); + break; + } +} +#endif diff --git a/third_party/aom/aom_dsp/mips/convolve2_horiz_dspr2.c b/third_party/aom/aom_dsp/mips/convolve2_horiz_dspr2.c new file mode 100644 index 0000000000..097da73ca0 --- /dev/null +++ b/third_party/aom/aom_dsp/mips/convolve2_horiz_dspr2.c @@ -0,0 +1,681 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include +#include + +#include "config/aom_dsp_rtcd.h" + +#include "aom_dsp/mips/convolve_common_dspr2.h" +#include "aom_dsp/aom_dsp_common.h" +#include "aom_ports/mem.h" + +#if HAVE_DSPR2 +static void convolve_bi_horiz_4_dspr2(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + const int16_t *filter_x0, int32_t h) { + int32_t y; + uint8_t *cm = aom_ff_cropTbl; + int32_t Temp1, Temp2, Temp3, Temp4; + uint32_t vector4a = 64; + uint32_t tp1, tp2; + uint32_t p1, p2; + const int16_t *filter = &filter_x0[3]; + uint32_t filter45; + + filter45 = ((const int32_t *)filter)[0]; + + for (y = h; y--;) { + /* prefetch data to cache memory */ + prefetch_load(src + src_stride); + prefetch_load(src + src_stride + 32); + prefetch_store(dst + dst_stride); + + __asm__ __volatile__( + "ulw %[tp1], 0(%[src]) \n\t" + "ulw %[tp2], 4(%[src]) \n\t" + + /* even 1. pixel */ + "mtlo %[vector4a], $ac3 \n\t" + "mthi $zero, $ac3 \n\t" + "preceu.ph.qbr %[p1], %[tp1] \n\t" + "preceu.ph.qbl %[p2], %[tp1] \n\t" + "dpa.w.ph $ac3, %[p1], %[filter45] \n\t" + "extp %[Temp1], $ac3, 31 \n\t" + + /* even 2. pixel */ + "mtlo %[vector4a], $ac2 \n\t" + "mthi $zero, $ac2 \n\t" + "balign %[tp2], %[tp1], 3 \n\t" + "dpa.w.ph $ac2, %[p2], %[filter45] \n\t" + "extp %[Temp3], $ac2, 31 \n\t" + + /* odd 1. pixel */ + "lbux %[tp1], %[Temp1](%[cm]) \n\t" + "mtlo %[vector4a], $ac3 \n\t" + "mthi $zero, $ac3 \n\t" + "preceu.ph.qbr %[p1], %[tp2] \n\t" + "preceu.ph.qbl %[p2], %[tp2] \n\t" + "dpa.w.ph $ac3, %[p1], %[filter45] \n\t" + "extp %[Temp2], $ac3, 31 \n\t" + + /* odd 2. pixel */ + "lbux %[tp2], %[Temp3](%[cm]) \n\t" + "mtlo %[vector4a], $ac2 \n\t" + "mthi $zero, $ac2 \n\t" + "dpa.w.ph $ac2, %[p2], %[filter45] \n\t" + "extp %[Temp4], $ac2, 31 \n\t" + + /* clamp */ + "lbux %[p1], %[Temp2](%[cm]) \n\t" + "lbux %[p2], %[Temp4](%[cm]) \n\t" + + /* store bytes */ + "sb %[tp1], 0(%[dst]) \n\t" + "sb %[p1], 1(%[dst]) \n\t" + "sb %[tp2], 2(%[dst]) \n\t" + "sb %[p2], 3(%[dst]) \n\t" + + : [tp1] "=&r"(tp1), [tp2] "=&r"(tp2), [p1] "=&r"(p1), [p2] "=&r"(p2), + [Temp1] "=&r"(Temp1), [Temp2] "=&r"(Temp2), [Temp3] "=&r"(Temp3), + [Temp4] "=&r"(Temp4) + : [filter45] "r"(filter45), [vector4a] "r"(vector4a), [cm] "r"(cm), + [dst] "r"(dst), [src] "r"(src)); + + /* Next row... */ + src += src_stride; + dst += dst_stride; + } +} + +static void convolve_bi_horiz_8_dspr2(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + const int16_t *filter_x0, int32_t h) { + int32_t y; + uint8_t *cm = aom_ff_cropTbl; + uint32_t vector4a = 64; + int32_t Temp1, Temp2, Temp3; + uint32_t tp1, tp2, tp3; + uint32_t p1, p2, p3, p4; + uint32_t st0, st1; + const int16_t *filter = &filter_x0[3]; + uint32_t filter45; + + filter45 = ((const int32_t *)filter)[0]; + + for (y = h; y--;) { + /* prefetch data to cache memory */ + prefetch_load(src + src_stride); + prefetch_load(src + src_stride + 32); + prefetch_store(dst + dst_stride); + + __asm__ __volatile__( + "ulw %[tp1], 0(%[src]) \n\t" + "ulw %[tp2], 4(%[src]) \n\t" + + /* even 1. pixel */ + "mtlo %[vector4a], $ac3 \n\t" + "mthi $zero, $ac3 \n\t" + "mtlo %[vector4a], $ac2 \n\t" + "mthi $zero, $ac2 \n\t" + "preceu.ph.qbr %[p1], %[tp1] \n\t" + "preceu.ph.qbl %[p2], %[tp1] \n\t" + "preceu.ph.qbr %[p3], %[tp2] \n\t" + "preceu.ph.qbl %[p4], %[tp2] \n\t" + "ulw %[tp3], 8(%[src]) \n\t" + "dpa.w.ph $ac3, %[p1], %[filter45] \n\t" + "extp %[Temp1], $ac3, 31 \n\t" + + /* even 2. pixel */ + "dpa.w.ph $ac2, %[p2], %[filter45] \n\t" + "extp %[Temp3], $ac2, 31 \n\t" + + /* even 3. pixel */ + "lbux %[st0], %[Temp1](%[cm]) \n\t" + "mtlo %[vector4a], $ac1 \n\t" + "mthi $zero, $ac1 \n\t" + "dpa.w.ph $ac1, %[p3], %[filter45] \n\t" + "extp %[Temp1], $ac1, 31 \n\t" + + /* even 4. pixel */ + "mtlo %[vector4a], $ac2 \n\t" + "mthi $zero, $ac2 \n\t" + "mtlo %[vector4a], $ac3 \n\t" + "mthi $zero, $ac3 \n\t" + "sb %[st0], 0(%[dst]) \n\t" + "lbux %[st1], %[Temp3](%[cm]) \n\t" + + "balign %[tp3], %[tp2], 3 \n\t" + "balign %[tp2], %[tp1], 3 \n\t" + + "dpa.w.ph $ac2, %[p4], %[filter45] \n\t" + "extp %[Temp3], $ac2, 31 \n\t" + + "lbux %[st0], %[Temp1](%[cm]) \n\t" + + /* odd 1. pixel */ + "mtlo %[vector4a], $ac1 \n\t" + "mthi $zero, $ac1 \n\t" + "sb %[st1], 2(%[dst]) \n\t" + "preceu.ph.qbr %[p1], %[tp2] \n\t" + "preceu.ph.qbl %[p2], %[tp2] \n\t" + "preceu.ph.qbr %[p3], %[tp3] \n\t" + "preceu.ph.qbl %[p4], %[tp3] \n\t" + "sb %[st0], 4(%[dst]) \n\t" + "dpa.w.ph $ac3, %[p1], %[filter45] \n\t" + "extp %[Temp2], $ac3, 31 \n\t" + + /* odd 2. pixel */ + "mtlo %[vector4a], $ac3 \n\t" + "mthi $zero, $ac3 \n\t" + "mtlo %[vector4a], $ac2 \n\t" + "mthi $zero, $ac2 \n\t" + "lbux %[st0], %[Temp3](%[cm]) \n\t" + "dpa.w.ph $ac1, %[p2], %[filter45] \n\t" + "extp %[Temp3], $ac1, 31 \n\t" + + /* odd 3. pixel */ + "lbux %[st1], %[Temp2](%[cm]) \n\t" + "dpa.w.ph $ac3, %[p3], %[filter45] \n\t" + "extp %[Temp2], $ac3, 31 \n\t" + + /* odd 4. pixel */ + "sb %[st1], 1(%[dst]) \n\t" + "sb %[st0], 6(%[dst]) \n\t" + "dpa.w.ph $ac2, %[p4], %[filter45] \n\t" + "extp %[Temp1], $ac2, 31 \n\t" + + /* clamp */ + "lbux %[p4], %[Temp3](%[cm]) \n\t" + "lbux %[p2], %[Temp2](%[cm]) \n\t" + "lbux %[p1], %[Temp1](%[cm]) \n\t" + + /* store bytes */ + "sb %[p4], 3(%[dst]) \n\t" + "sb %[p2], 5(%[dst]) \n\t" + "sb %[p1], 7(%[dst]) \n\t" + + : [tp1] "=&r"(tp1), [tp2] "=&r"(tp2), [tp3] "=&r"(tp3), + [st0] "=&r"(st0), [st1] "=&r"(st1), [p1] "=&r"(p1), [p2] "=&r"(p2), + [p3] "=&r"(p3), [p4] "=&r"(p4), [Temp1] "=&r"(Temp1), + [Temp2] "=&r"(Temp2), [Temp3] "=&r"(Temp3) + : [filter45] "r"(filter45), [vector4a] "r"(vector4a), [cm] "r"(cm), + [dst] "r"(dst), [src] "r"(src)); + + /* Next row... */ + src += src_stride; + dst += dst_stride; + } +} + +static void convolve_bi_horiz_16_dspr2(const uint8_t *src_ptr, + int32_t src_stride, uint8_t *dst_ptr, + int32_t dst_stride, + const int16_t *filter_x0, int32_t h, + int32_t count) { + int32_t y, c; + const uint8_t *src; + uint8_t *dst; + uint8_t *cm = aom_ff_cropTbl; + uint32_t vector_64 = 64; + int32_t Temp1, Temp2, Temp3; + uint32_t qload1, qload2, qload3; + uint32_t p1, p2, p3, p4, p5; + uint32_t st1, st2, st3; + const int16_t *filter = &filter_x0[3]; + uint32_t filter45; + + filter45 = ((const int32_t *)filter)[0]; + + for (y = h; y--;) { + src = src_ptr; + dst = dst_ptr; + + /* prefetch data to cache memory */ + prefetch_load(src_ptr + src_stride); + prefetch_load(src_ptr + src_stride + 32); + prefetch_store(dst_ptr + dst_stride); + + for (c = 0; c < count; c++) { + __asm__ __volatile__( + "ulw %[qload1], 0(%[src]) \n\t" + "ulw %[qload2], 4(%[src]) \n\t" + + /* even 1. pixel */ + "mtlo %[vector_64], $ac1 \n\t" /* even 1 */ + "mthi $zero, $ac1 \n\t" + "mtlo %[vector_64], $ac2 \n\t" /* even 2 */ + "mthi $zero, $ac2 \n\t" + "preceu.ph.qbr %[p1], %[qload1] \n\t" + "preceu.ph.qbl %[p2], %[qload1] \n\t" + "preceu.ph.qbr %[p3], %[qload2] \n\t" + "preceu.ph.qbl %[p4], %[qload2] \n\t" + "ulw %[qload3], 8(%[src]) \n\t" + "dpa.w.ph $ac1, %[p1], %[filter45] \n\t" /* even 1 */ + "extp %[Temp1], $ac1, 31 \n\t" /* even 1 */ + + /* even 2. pixel */ + "mtlo %[vector_64], $ac3 \n\t" /* even 3 */ + "mthi $zero, $ac3 \n\t" + "preceu.ph.qbr %[p1], %[qload3] \n\t" + "preceu.ph.qbl %[p5], %[qload3] \n\t" + "ulw %[qload1], 12(%[src]) \n\t" + "dpa.w.ph $ac2, %[p2], %[filter45] \n\t" /* even 1 */ + "extp %[Temp2], $ac2, 31 \n\t" /* even 1 */ + "lbux %[st1], %[Temp1](%[cm]) \n\t" /* even 1 */ + + /* even 3. pixel */ + "mtlo %[vector_64], $ac1 \n\t" /* even 4 */ + "mthi $zero, $ac1 \n\t" + "preceu.ph.qbr %[p2], %[qload1] \n\t" + "sb %[st1], 0(%[dst]) \n\t" /* even 1 */ + "dpa.w.ph $ac3, %[p3], %[filter45] \n\t" /* even 3 */ + "extp %[Temp3], $ac3, 31 \n\t" /* even 3 */ + "lbux %[st2], %[Temp2](%[cm]) \n\t" /* even 1 */ + + /* even 4. pixel */ + "mtlo %[vector_64], $ac2 \n\t" /* even 5 */ + "mthi $zero, $ac2 \n\t" + "preceu.ph.qbl %[p3], %[qload1] \n\t" + "sb %[st2], 2(%[dst]) \n\t" /* even 1 */ + "dpa.w.ph $ac1, %[p4], %[filter45] \n\t" /* even 4 */ + "extp %[Temp1], $ac1, 31 \n\t" /* even 4 */ + "lbux %[st3], %[Temp3](%[cm]) \n\t" /* even 3 */ + + /* even 5. pixel */ + "mtlo %[vector_64], $ac3 \n\t" /* even 6 */ + "mthi $zero, $ac3 \n\t" + "sb %[st3], 4(%[dst]) \n\t" /* even 3 */ + "dpa.w.ph $ac2, %[p1], %[filter45] \n\t" /* even 5 */ + "extp %[Temp2], $ac2, 31 \n\t" /* even 5 */ + "lbux %[st1], %[Temp1](%[cm]) \n\t" /* even 4 */ + + /* even 6. pixel */ + "mtlo %[vector_64], $ac1 \n\t" /* even 7 */ + "mthi $zero, $ac1 \n\t" + "sb %[st1], 6(%[dst]) \n\t" /* even 4 */ + "dpa.w.ph $ac3, %[p5], %[filter45] \n\t" /* even 6 */ + "extp %[Temp3], $ac3, 31 \n\t" /* even 6 */ + "lbux %[st2], %[Temp2](%[cm]) \n\t" /* even 5 */ + + /* even 7. pixel */ + "mtlo %[vector_64], $ac2 \n\t" /* even 8 */ + "mthi $zero, $ac2 \n\t" + "sb %[st2], 8(%[dst]) \n\t" /* even 5 */ + "dpa.w.ph $ac1, %[p2], %[filter45] \n\t" /* even 7 */ + "extp %[Temp1], $ac1, 31 \n\t" /* even 7 */ + "lbux %[st3], %[Temp3](%[cm]) \n\t" /* even 6 */ + + /* even 8. pixel */ + "mtlo %[vector_64], $ac3 \n\t" /* odd 1 */ + "mthi $zero, $ac3 \n\t" + "dpa.w.ph $ac2, %[p3], %[filter45] \n\t" /* even 8 */ + "sb %[st3], 10(%[dst]) \n\t" /* even 6 */ + "extp %[Temp2], $ac2, 31 \n\t" /* even 8 */ + "lbux %[st1], %[Temp1](%[cm]) \n\t" /* even 7 */ + + /* ODD pixels */ + "ulw %[qload1], 1(%[src]) \n\t" + "ulw %[qload2], 5(%[src]) \n\t" + + /* odd 1. pixel */ + "mtlo %[vector_64], $ac1 \n\t" /* odd 2 */ + "mthi $zero, $ac1 \n\t" + "preceu.ph.qbr %[p1], %[qload1] \n\t" + "preceu.ph.qbl %[p2], %[qload1] \n\t" + "preceu.ph.qbr %[p3], %[qload2] \n\t" + "preceu.ph.qbl %[p4], %[qload2] \n\t" + "sb %[st1], 12(%[dst]) \n\t" /* even 7 */ + "ulw %[qload3], 9(%[src]) \n\t" + "dpa.w.ph $ac3, %[p1], %[filter45] \n\t" /* odd 1 */ + "extp %[Temp3], $ac3, 31 \n\t" /* odd 1 */ + "lbux %[st2], %[Temp2](%[cm]) \n\t" /* even 8 */ + + /* odd 2. pixel */ + "mtlo %[vector_64], $ac2 \n\t" /* odd 3 */ + "mthi $zero, $ac2 \n\t" + "preceu.ph.qbr %[p1], %[qload3] \n\t" + "preceu.ph.qbl %[p5], %[qload3] \n\t" + "sb %[st2], 14(%[dst]) \n\t" /* even 8 */ + "ulw %[qload1], 13(%[src]) \n\t" + "dpa.w.ph $ac1, %[p2], %[filter45] \n\t" /* odd 2 */ + "extp %[Temp1], $ac1, 31 \n\t" /* odd 2 */ + "lbux %[st3], %[Temp3](%[cm]) \n\t" /* odd 1 */ + + /* odd 3. pixel */ + "mtlo %[vector_64], $ac3 \n\t" /* odd 4 */ + "mthi $zero, $ac3 \n\t" + "preceu.ph.qbr %[p2], %[qload1] \n\t" + "sb %[st3], 1(%[dst]) \n\t" /* odd 1 */ + "dpa.w.ph $ac2, %[p3], %[filter45] \n\t" /* odd 3 */ + "extp %[Temp2], $ac2, 31 \n\t" /* odd 3 */ + "lbux %[st1], %[Temp1](%[cm]) \n\t" /* odd 2 */ + + /* odd 4. pixel */ + "mtlo %[vector_64], $ac1 \n\t" /* odd 5 */ + "mthi $zero, $ac1 \n\t" + "preceu.ph.qbl %[p3], %[qload1] \n\t" + "sb %[st1], 3(%[dst]) \n\t" /* odd 2 */ + "dpa.w.ph $ac3, %[p4], %[filter45] \n\t" /* odd 4 */ + "extp %[Temp3], $ac3, 31 \n\t" /* odd 4 */ + "lbux %[st2], %[Temp2](%[cm]) \n\t" /* odd 3 */ + + /* odd 5. pixel */ + "mtlo %[vector_64], $ac2 \n\t" /* odd 6 */ + "mthi $zero, $ac2 \n\t" + "sb %[st2], 5(%[dst]) \n\t" /* odd 3 */ + "dpa.w.ph $ac1, %[p1], %[filter45] \n\t" /* odd 5 */ + "extp %[Temp1], $ac1, 31 \n\t" /* odd 5 */ + "lbux %[st3], %[Temp3](%[cm]) \n\t" /* odd 4 */ + + /* odd 6. pixel */ + "mtlo %[vector_64], $ac3 \n\t" /* odd 7 */ + "mthi $zero, $ac3 \n\t" + "sb %[st3], 7(%[dst]) \n\t" /* odd 4 */ + "dpa.w.ph $ac2, %[p5], %[filter45] \n\t" /* odd 6 */ + "extp %[Temp2], $ac2, 31 \n\t" /* odd 6 */ + "lbux %[st1], %[Temp1](%[cm]) \n\t" /* odd 5 */ + + /* odd 7. pixel */ + "mtlo %[vector_64], $ac1 \n\t" /* odd 8 */ + "mthi $zero, $ac1 \n\t" + "sb %[st1], 9(%[dst]) \n\t" /* odd 5 */ + "dpa.w.ph $ac3, %[p2], %[filter45] \n\t" /* odd 7 */ + "extp %[Temp3], $ac3, 31 \n\t" /* odd 7 */ + + /* odd 8. pixel */ + "dpa.w.ph $ac1, %[p3], %[filter45] \n\t" /* odd 8 */ + "extp %[Temp1], $ac1, 31 \n\t" /* odd 8 */ + + "lbux %[st2], %[Temp2](%[cm]) \n\t" /* odd 6 */ + "lbux %[st3], %[Temp3](%[cm]) \n\t" /* odd 7 */ + "lbux %[st1], %[Temp1](%[cm]) \n\t" /* odd 8 */ + + "sb %[st2], 11(%[dst]) \n\t" /* odd 6 */ + "sb %[st3], 13(%[dst]) \n\t" /* odd 7 */ + "sb %[st1], 15(%[dst]) \n\t" /* odd 8 */ + + : [qload1] "=&r"(qload1), [qload2] "=&r"(qload2), + [qload3] "=&r"(qload3), [st1] "=&r"(st1), [st2] "=&r"(st2), + [st3] "=&r"(st3), [p1] "=&r"(p1), [p2] "=&r"(p2), [p3] "=&r"(p3), + [p4] "=&r"(p4), [p5] "=&r"(p5), [Temp1] "=&r"(Temp1), + [Temp2] "=&r"(Temp2), [Temp3] "=&r"(Temp3) + : [filter45] "r"(filter45), [vector_64] "r"(vector_64), [cm] "r"(cm), + [dst] "r"(dst), [src] "r"(src)); + + src += 16; + dst += 16; + } + + /* Next row... */ + src_ptr += src_stride; + dst_ptr += dst_stride; + } +} + +static void convolve_bi_horiz_64_dspr2(const uint8_t *src_ptr, + int32_t src_stride, uint8_t *dst_ptr, + int32_t dst_stride, + const int16_t *filter_x0, int32_t h) { + int32_t y, c; + const uint8_t *src; + uint8_t *dst; + uint8_t *cm = aom_ff_cropTbl; + uint32_t vector_64 = 64; + int32_t Temp1, Temp2, Temp3; + uint32_t qload1, qload2, qload3; + uint32_t p1, p2, p3, p4, p5; + uint32_t st1, st2, st3; + const int16_t *filter = &filter_x0[3]; + uint32_t filter45; + + filter45 = ((const int32_t *)filter)[0]; + + for (y = h; y--;) { + src = src_ptr; + dst = dst_ptr; + + /* prefetch data to cache memory */ + prefetch_load(src_ptr + src_stride); + prefetch_load(src_ptr + src_stride + 32); + prefetch_load(src_ptr + src_stride + 64); + prefetch_store(dst_ptr + dst_stride); + prefetch_store(dst_ptr + dst_stride + 32); + + for (c = 0; c < 4; c++) { + __asm__ __volatile__( + "ulw %[qload1], 0(%[src]) \n\t" + "ulw %[qload2], 4(%[src]) \n\t" + + /* even 1. pixel */ + "mtlo %[vector_64], $ac1 \n\t" /* even 1 */ + "mthi $zero, $ac1 \n\t" + "mtlo %[vector_64], $ac2 \n\t" /* even 2 */ + "mthi $zero, $ac2 \n\t" + "preceu.ph.qbr %[p1], %[qload1] \n\t" + "preceu.ph.qbl %[p2], %[qload1] \n\t" + "preceu.ph.qbr %[p3], %[qload2] \n\t" + "preceu.ph.qbl %[p4], %[qload2] \n\t" + "ulw %[qload3], 8(%[src]) \n\t" + "dpa.w.ph $ac1, %[p1], %[filter45] \n\t" /* even 1 */ + "extp %[Temp1], $ac1, 31 \n\t" /* even 1 */ + + /* even 2. pixel */ + "mtlo %[vector_64], $ac3 \n\t" /* even 3 */ + "mthi $zero, $ac3 \n\t" + "preceu.ph.qbr %[p1], %[qload3] \n\t" + "preceu.ph.qbl %[p5], %[qload3] \n\t" + "ulw %[qload1], 12(%[src]) \n\t" + "dpa.w.ph $ac2, %[p2], %[filter45] \n\t" /* even 1 */ + "extp %[Temp2], $ac2, 31 \n\t" /* even 1 */ + "lbux %[st1], %[Temp1](%[cm]) \n\t" /* even 1 */ + + /* even 3. pixel */ + "mtlo %[vector_64], $ac1 \n\t" /* even 4 */ + "mthi $zero, $ac1 \n\t" + "preceu.ph.qbr %[p2], %[qload1] \n\t" + "sb %[st1], 0(%[dst]) \n\t" /* even 1 */ + "dpa.w.ph $ac3, %[p3], %[filter45] \n\t" /* even 3 */ + "extp %[Temp3], $ac3, 31 \n\t" /* even 3 */ + "lbux %[st2], %[Temp2](%[cm]) \n\t" /* even 1 */ + + /* even 4. pixel */ + "mtlo %[vector_64], $ac2 \n\t" /* even 5 */ + "mthi $zero, $ac2 \n\t" + "preceu.ph.qbl %[p3], %[qload1] \n\t" + "sb %[st2], 2(%[dst]) \n\t" /* even 1 */ + "dpa.w.ph $ac1, %[p4], %[filter45] \n\t" /* even 4 */ + "extp %[Temp1], $ac1, 31 \n\t" /* even 4 */ + "lbux %[st3], %[Temp3](%[cm]) \n\t" /* even 3 */ + + /* even 5. pixel */ + "mtlo %[vector_64], $ac3 \n\t" /* even 6 */ + "mthi $zero, $ac3 \n\t" + "sb %[st3], 4(%[dst]) \n\t" /* even 3 */ + "dpa.w.ph $ac2, %[p1], %[filter45] \n\t" /* even 5 */ + "extp %[Temp2], $ac2, 31 \n\t" /* even 5 */ + "lbux %[st1], %[Temp1](%[cm]) \n\t" /* even 4 */ + + /* even 6. pixel */ + "mtlo %[vector_64], $ac1 \n\t" /* even 7 */ + "mthi $zero, $ac1 \n\t" + "sb %[st1], 6(%[dst]) \n\t" /* even 4 */ + "dpa.w.ph $ac3, %[p5], %[filter45] \n\t" /* even 6 */ + "extp %[Temp3], $ac3, 31 \n\t" /* even 6 */ + "lbux %[st2], %[Temp2](%[cm]) \n\t" /* even 5 */ + + /* even 7. pixel */ + "mtlo %[vector_64], $ac2 \n\t" /* even 8 */ + "mthi $zero, $ac2 \n\t" + "sb %[st2], 8(%[dst]) \n\t" /* even 5 */ + "dpa.w.ph $ac1, %[p2], %[filter45] \n\t" /* even 7 */ + "extp %[Temp1], $ac1, 31 \n\t" /* even 7 */ + "lbux %[st3], %[Temp3](%[cm]) \n\t" /* even 6 */ + + /* even 8. pixel */ + "mtlo %[vector_64], $ac3 \n\t" /* odd 1 */ + "mthi $zero, $ac3 \n\t" + "dpa.w.ph $ac2, %[p3], %[filter45] \n\t" /* even 8 */ + "sb %[st3], 10(%[dst]) \n\t" /* even 6 */ + "extp %[Temp2], $ac2, 31 \n\t" /* even 8 */ + "lbux %[st1], %[Temp1](%[cm]) \n\t" /* even 7 */ + + /* ODD pixels */ + "ulw %[qload1], 1(%[src]) \n\t" + "ulw %[qload2], 5(%[src]) \n\t" + + /* odd 1. pixel */ + "mtlo %[vector_64], $ac1 \n\t" /* odd 2 */ + "mthi $zero, $ac1 \n\t" + "preceu.ph.qbr %[p1], %[qload1] \n\t" + "preceu.ph.qbl %[p2], %[qload1] \n\t" + "preceu.ph.qbr %[p3], %[qload2] \n\t" + "preceu.ph.qbl %[p4], %[qload2] \n\t" + "sb %[st1], 12(%[dst]) \n\t" /* even 7 */ + "ulw %[qload3], 9(%[src]) \n\t" + "dpa.w.ph $ac3, %[p1], %[filter45] \n\t" /* odd 1 */ + "extp %[Temp3], $ac3, 31 \n\t" /* odd 1 */ + "lbux %[st2], %[Temp2](%[cm]) \n\t" /* even 8 */ + + /* odd 2. pixel */ + "mtlo %[vector_64], $ac2 \n\t" /* odd 3 */ + "mthi $zero, $ac2 \n\t" + "preceu.ph.qbr %[p1], %[qload3] \n\t" + "preceu.ph.qbl %[p5], %[qload3] \n\t" + "sb %[st2], 14(%[dst]) \n\t" /* even 8 */ + "ulw %[qload1], 13(%[src]) \n\t" + "dpa.w.ph $ac1, %[p2], %[filter45] \n\t" /* odd 2 */ + "extp %[Temp1], $ac1, 31 \n\t" /* odd 2 */ + "lbux %[st3], %[Temp3](%[cm]) \n\t" /* odd 1 */ + + /* odd 3. pixel */ + "mtlo %[vector_64], $ac3 \n\t" /* odd 4 */ + "mthi $zero, $ac3 \n\t" + "preceu.ph.qbr %[p2], %[qload1] \n\t" + "sb %[st3], 1(%[dst]) \n\t" /* odd 1 */ + "dpa.w.ph $ac2, %[p3], %[filter45] \n\t" /* odd 3 */ + "extp %[Temp2], $ac2, 31 \n\t" /* odd 3 */ + "lbux %[st1], %[Temp1](%[cm]) \n\t" /* odd 2 */ + + /* odd 4. pixel */ + "mtlo %[vector_64], $ac1 \n\t" /* odd 5 */ + "mthi $zero, $ac1 \n\t" + "preceu.ph.qbl %[p3], %[qload1] \n\t" + "sb %[st1], 3(%[dst]) \n\t" /* odd 2 */ + "dpa.w.ph $ac3, %[p4], %[filter45] \n\t" /* odd 4 */ + "extp %[Temp3], $ac3, 31 \n\t" /* odd 4 */ + "lbux %[st2], %[Temp2](%[cm]) \n\t" /* odd 3 */ + + /* odd 5. pixel */ + "mtlo %[vector_64], $ac2 \n\t" /* odd 6 */ + "mthi $zero, $ac2 \n\t" + "sb %[st2], 5(%[dst]) \n\t" /* odd 3 */ + "dpa.w.ph $ac1, %[p1], %[filter45] \n\t" /* odd 5 */ + "extp %[Temp1], $ac1, 31 \n\t" /* odd 5 */ + "lbux %[st3], %[Temp3](%[cm]) \n\t" /* odd 4 */ + + /* odd 6. pixel */ + "mtlo %[vector_64], $ac3 \n\t" /* odd 7 */ + "mthi $zero, $ac3 \n\t" + "sb %[st3], 7(%[dst]) \n\t" /* odd 4 */ + "dpa.w.ph $ac2, %[p5], %[filter45] \n\t" /* odd 6 */ + "extp %[Temp2], $ac2, 31 \n\t" /* odd 6 */ + "lbux %[st1], %[Temp1](%[cm]) \n\t" /* odd 5 */ + + /* odd 7. pixel */ + "mtlo %[vector_64], $ac1 \n\t" /* odd 8 */ + "mthi $zero, $ac1 \n\t" + "sb %[st1], 9(%[dst]) \n\t" /* odd 5 */ + "dpa.w.ph $ac3, %[p2], %[filter45] \n\t" /* odd 7 */ + "extp %[Temp3], $ac3, 31 \n\t" /* odd 7 */ + + /* odd 8. pixel */ + "dpa.w.ph $ac1, %[p3], %[filter45] \n\t" /* odd 8 */ + "extp %[Temp1], $ac1, 31 \n\t" /* odd 8 */ + + "lbux %[st2], %[Temp2](%[cm]) \n\t" /* odd 6 */ + "lbux %[st3], %[Temp3](%[cm]) \n\t" /* odd 7 */ + "lbux %[st1], %[Temp1](%[cm]) \n\t" /* odd 8 */ + + "sb %[st2], 11(%[dst]) \n\t" /* odd 6 */ + "sb %[st3], 13(%[dst]) \n\t" /* odd 7 */ + "sb %[st1], 15(%[dst]) \n\t" /* odd 8 */ + + : [qload1] "=&r"(qload1), [qload2] "=&r"(qload2), + [qload3] "=&r"(qload3), [st1] "=&r"(st1), [st2] "=&r"(st2), + [st3] "=&r"(st3), [p1] "=&r"(p1), [p2] "=&r"(p2), [p3] "=&r"(p3), + [p4] "=&r"(p4), [p5] "=&r"(p5), [Temp1] "=&r"(Temp1), + [Temp2] "=&r"(Temp2), [Temp3] "=&r"(Temp3) + : [filter45] "r"(filter45), [vector_64] "r"(vector_64), [cm] "r"(cm), + [dst] "r"(dst), [src] "r"(src)); + + src += 16; + dst += 16; + } + + /* Next row... */ + src_ptr += src_stride; + dst_ptr += dst_stride; + } +} + +void aom_convolve2_horiz_dspr2(const uint8_t *src, ptrdiff_t src_stride, + uint8_t *dst, ptrdiff_t dst_stride, + const int16_t *filter_x, int x_step_q4, + const int16_t *filter_y, int y_step_q4, int w, + int h) { + uint32_t pos = 38; + + assert(x_step_q4 == 16); + + prefetch_load((const uint8_t *)filter_x); + + /* bit positon for extract from acc */ + __asm__ __volatile__("wrdsp %[pos], 1 \n\t" + : + : [pos] "r"(pos)); + + /* prefetch data to cache memory */ + prefetch_load(src); + prefetch_load(src + 32); + prefetch_store(dst); + + switch (w) { + case 4: + convolve_bi_horiz_4_dspr2(src, (int32_t)src_stride, dst, + (int32_t)dst_stride, filter_x, (int32_t)h); + break; + case 8: + convolve_bi_horiz_8_dspr2(src, (int32_t)src_stride, dst, + (int32_t)dst_stride, filter_x, (int32_t)h); + break; + case 16: + convolve_bi_horiz_16_dspr2(src, (int32_t)src_stride, dst, + (int32_t)dst_stride, filter_x, (int32_t)h, 1); + break; + case 32: + convolve_bi_horiz_16_dspr2(src, (int32_t)src_stride, dst, + (int32_t)dst_stride, filter_x, (int32_t)h, 2); + break; + case 64: + prefetch_load(src + 64); + prefetch_store(dst + 32); + + convolve_bi_horiz_64_dspr2(src, (int32_t)src_stride, dst, + (int32_t)dst_stride, filter_x, (int32_t)h); + break; + default: + aom_convolve8_horiz_c(src, src_stride, dst, dst_stride, filter_x, + x_step_q4, filter_y, y_step_q4, w, h); + break; + } +} +#endif diff --git a/third_party/aom/aom_dsp/mips/convolve2_vert_dspr2.c b/third_party/aom/aom_dsp/mips/convolve2_vert_dspr2.c new file mode 100644 index 0000000000..40abfd89eb --- /dev/null +++ b/third_party/aom/aom_dsp/mips/convolve2_vert_dspr2.c @@ -0,0 +1,237 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include +#include + +#include "config/aom_dsp_rtcd.h" + +#include "aom_dsp/mips/convolve_common_dspr2.h" +#include "aom_dsp/aom_dsp_common.h" +#include "aom_ports/mem.h" + +#if HAVE_DSPR2 +static void convolve_bi_vert_4_dspr2(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + const int16_t *filter_y, int32_t w, + int32_t h) { + int32_t x, y; + const uint8_t *src_ptr; + uint8_t *dst_ptr; + uint8_t *cm = aom_ff_cropTbl; + uint32_t vector4a = 64; + uint32_t load1, load2; + uint32_t p1, p2; + uint32_t scratch1; + uint32_t store1, store2; + int32_t Temp1, Temp2; + const int16_t *filter = &filter_y[3]; + uint32_t filter45; + + filter45 = ((const int32_t *)filter)[0]; + + for (y = h; y--;) { + /* prefetch data to cache memory */ + prefetch_store(dst + dst_stride); + + for (x = 0; x < w; x += 4) { + src_ptr = src + x; + dst_ptr = dst + x; + + __asm__ __volatile__( + "ulw %[load1], 0(%[src_ptr]) \n\t" + "add %[src_ptr], %[src_ptr], %[src_stride] \n\t" + "ulw %[load2], 0(%[src_ptr]) \n\t" + + "mtlo %[vector4a], $ac0 \n\t" + "mtlo %[vector4a], $ac1 \n\t" + "mtlo %[vector4a], $ac2 \n\t" + "mtlo %[vector4a], $ac3 \n\t" + "mthi $zero, $ac0 \n\t" + "mthi $zero, $ac1 \n\t" + "mthi $zero, $ac2 \n\t" + "mthi $zero, $ac3 \n\t" + + "preceu.ph.qbr %[scratch1], %[load1] \n\t" + "preceu.ph.qbr %[p1], %[load2] \n\t" + + "precrq.ph.w %[p2], %[p1], %[scratch1] \n\t" /* pixel 2 */ + "append %[p1], %[scratch1], 16 \n\t" /* pixel 1 */ + + "dpa.w.ph $ac0, %[p1], %[filter45] \n\t" + "dpa.w.ph $ac1, %[p2], %[filter45] \n\t" + + "preceu.ph.qbl %[scratch1], %[load1] \n\t" + "preceu.ph.qbl %[p1], %[load2] \n\t" + + "precrq.ph.w %[p2], %[p1], %[scratch1] \n\t" /* pixel 2 */ + "append %[p1], %[scratch1], 16 \n\t" /* pixel 1 */ + + "dpa.w.ph $ac2, %[p1], %[filter45] \n\t" + "dpa.w.ph $ac3, %[p2], %[filter45] \n\t" + + "extp %[Temp1], $ac0, 31 \n\t" + "extp %[Temp2], $ac1, 31 \n\t" + + "lbux %[store1], %[Temp1](%[cm]) \n\t" + "extp %[Temp1], $ac2, 31 \n\t" + + "lbux %[store2], %[Temp2](%[cm]) \n\t" + "extp %[Temp2], $ac3, 31 \n\t" + + "sb %[store1], 0(%[dst_ptr]) \n\t" + "sb %[store2], 1(%[dst_ptr]) \n\t" + + "lbux %[store1], %[Temp1](%[cm]) \n\t" + "lbux %[store2], %[Temp2](%[cm]) \n\t" + + "sb %[store1], 2(%[dst_ptr]) \n\t" + "sb %[store2], 3(%[dst_ptr]) \n\t" + + : [load1] "=&r"(load1), [load2] "=&r"(load2), [p1] "=&r"(p1), + [p2] "=&r"(p2), [scratch1] "=&r"(scratch1), [Temp1] "=&r"(Temp1), + [Temp2] "=&r"(Temp2), [store1] "=&r"(store1), + [store2] "=&r"(store2), [src_ptr] "+r"(src_ptr) + : [filter45] "r"(filter45), [vector4a] "r"(vector4a), + [src_stride] "r"(src_stride), [cm] "r"(cm), [dst_ptr] "r"(dst_ptr)); + } + + /* Next row... */ + src += src_stride; + dst += dst_stride; + } +} + +static void convolve_bi_vert_64_dspr2(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + const int16_t *filter_y, int32_t h) { + int32_t x, y; + const uint8_t *src_ptr; + uint8_t *dst_ptr; + uint8_t *cm = aom_ff_cropTbl; + uint32_t vector4a = 64; + uint32_t load1, load2; + uint32_t p1, p2; + uint32_t scratch1; + uint32_t store1, store2; + int32_t Temp1, Temp2; + const int16_t *filter = &filter_y[3]; + uint32_t filter45; + + filter45 = ((const int32_t *)filter)[0]; + + for (y = h; y--;) { + /* prefetch data to cache memory */ + prefetch_store(dst + dst_stride); + + for (x = 0; x < 64; x += 4) { + src_ptr = src + x; + dst_ptr = dst + x; + + __asm__ __volatile__( + "ulw %[load1], 0(%[src_ptr]) \n\t" + "add %[src_ptr], %[src_ptr], %[src_stride] \n\t" + "ulw %[load2], 0(%[src_ptr]) \n\t" + + "mtlo %[vector4a], $ac0 \n\t" + "mtlo %[vector4a], $ac1 \n\t" + "mtlo %[vector4a], $ac2 \n\t" + "mtlo %[vector4a], $ac3 \n\t" + "mthi $zero, $ac0 \n\t" + "mthi $zero, $ac1 \n\t" + "mthi $zero, $ac2 \n\t" + "mthi $zero, $ac3 \n\t" + + "preceu.ph.qbr %[scratch1], %[load1] \n\t" + "preceu.ph.qbr %[p1], %[load2] \n\t" + + "precrq.ph.w %[p2], %[p1], %[scratch1] \n\t" /* pixel 2 */ + "append %[p1], %[scratch1], 16 \n\t" /* pixel 1 */ + + "dpa.w.ph $ac0, %[p1], %[filter45] \n\t" + "dpa.w.ph $ac1, %[p2], %[filter45] \n\t" + + "preceu.ph.qbl %[scratch1], %[load1] \n\t" + "preceu.ph.qbl %[p1], %[load2] \n\t" + + "precrq.ph.w %[p2], %[p1], %[scratch1] \n\t" /* pixel 2 */ + "append %[p1], %[scratch1], 16 \n\t" /* pixel 1 */ + + "dpa.w.ph $ac2, %[p1], %[filter45] \n\t" + "dpa.w.ph $ac3, %[p2], %[filter45] \n\t" + + "extp %[Temp1], $ac0, 31 \n\t" + "extp %[Temp2], $ac1, 31 \n\t" + + "lbux %[store1], %[Temp1](%[cm]) \n\t" + "extp %[Temp1], $ac2, 31 \n\t" + + "lbux %[store2], %[Temp2](%[cm]) \n\t" + "extp %[Temp2], $ac3, 31 \n\t" + + "sb %[store1], 0(%[dst_ptr]) \n\t" + "sb %[store2], 1(%[dst_ptr]) \n\t" + + "lbux %[store1], %[Temp1](%[cm]) \n\t" + "lbux %[store2], %[Temp2](%[cm]) \n\t" + + "sb %[store1], 2(%[dst_ptr]) \n\t" + "sb %[store2], 3(%[dst_ptr]) \n\t" + + : [load1] "=&r"(load1), [load2] "=&r"(load2), [p1] "=&r"(p1), + [p2] "=&r"(p2), [scratch1] "=&r"(scratch1), [Temp1] "=&r"(Temp1), + [Temp2] "=&r"(Temp2), [store1] "=&r"(store1), + [store2] "=&r"(store2), [src_ptr] "+r"(src_ptr) + : [filter45] "r"(filter45), [vector4a] "r"(vector4a), + [src_stride] "r"(src_stride), [cm] "r"(cm), [dst_ptr] "r"(dst_ptr)); + } + + /* Next row... */ + src += src_stride; + dst += dst_stride; + } +} + +void aom_convolve2_vert_dspr2(const uint8_t *src, ptrdiff_t src_stride, + uint8_t *dst, ptrdiff_t dst_stride, + const int16_t *filter_x, int x_step_q4, + const int16_t *filter_y, int y_step_q4, int w, + int h) { + uint32_t pos = 38; + + assert(y_step_q4 == 16); + + /* bit positon for extract from acc */ + __asm__ __volatile__("wrdsp %[pos], 1 \n\t" + : + : [pos] "r"(pos)); + + prefetch_store(dst); + + switch (w) { + case 4: + case 8: + case 16: + case 32: + convolve_bi_vert_4_dspr2(src, src_stride, dst, dst_stride, filter_y, w, + h); + break; + case 64: + prefetch_store(dst + 32); + convolve_bi_vert_64_dspr2(src, src_stride, dst, dst_stride, filter_y, h); + break; + default: + aom_convolve8_vert_c(src, src_stride, dst, dst_stride, filter_x, + x_step_q4, filter_y, y_step_q4, w, h); + break; + } +} +#endif diff --git a/third_party/aom/aom_dsp/mips/convolve8_dspr2.c b/third_party/aom/aom_dsp/mips/convolve8_dspr2.c new file mode 100644 index 0000000000..af54b42647 --- /dev/null +++ b/third_party/aom/aom_dsp/mips/convolve8_dspr2.c @@ -0,0 +1,222 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include +#include + +#include "config/aom_dsp_rtcd.h" + +#include "aom_dsp/mips/convolve_common_dspr2.h" +#include "aom_dsp/aom_dsp_common.h" +#include "aom_dsp/aom_filter.h" +#include "aom_ports/mem.h" + +#if HAVE_DSPR2 +void aom_convolve_copy_dspr2(const uint8_t *src, ptrdiff_t src_stride, + uint8_t *dst, ptrdiff_t dst_stride, + const int16_t *filter_x, int filter_x_stride, + const int16_t *filter_y, int filter_y_stride, + int w, int h) { + int x, y; + + (void)filter_x; + (void)filter_x_stride; + (void)filter_y; + (void)filter_y_stride; + + /* prefetch data to cache memory */ + prefetch_load(src); + prefetch_load(src + 32); + prefetch_store(dst); + + switch (w) { + case 4: { + uint32_t tp1; + + /* 1 word storage */ + for (y = h; y--;) { + prefetch_load(src + src_stride); + prefetch_load(src + src_stride + 32); + prefetch_store(dst + dst_stride); + + __asm__ __volatile__( + "ulw %[tp1], (%[src]) \n\t" + "sw %[tp1], (%[dst]) \n\t" /* store */ + + : [tp1] "=&r"(tp1) + : [src] "r"(src), [dst] "r"(dst)); + + src += src_stride; + dst += dst_stride; + } + } break; + case 8: { + uint32_t tp1, tp2; + + /* 2 word storage */ + for (y = h; y--;) { + prefetch_load(src + src_stride); + prefetch_load(src + src_stride + 32); + prefetch_store(dst + dst_stride); + + __asm__ __volatile__( + "ulw %[tp1], 0(%[src]) \n\t" + "ulw %[tp2], 4(%[src]) \n\t" + "sw %[tp1], 0(%[dst]) \n\t" /* store */ + "sw %[tp2], 4(%[dst]) \n\t" /* store */ + + : [tp1] "=&r"(tp1), [tp2] "=&r"(tp2) + : [src] "r"(src), [dst] "r"(dst)); + + src += src_stride; + dst += dst_stride; + } + } break; + case 16: { + uint32_t tp1, tp2, tp3, tp4; + + /* 4 word storage */ + for (y = h; y--;) { + prefetch_load(src + src_stride); + prefetch_load(src + src_stride + 32); + prefetch_store(dst + dst_stride); + + __asm__ __volatile__( + "ulw %[tp1], 0(%[src]) \n\t" + "ulw %[tp2], 4(%[src]) \n\t" + "ulw %[tp3], 8(%[src]) \n\t" + "ulw %[tp4], 12(%[src]) \n\t" + + "sw %[tp1], 0(%[dst]) \n\t" /* store */ + "sw %[tp2], 4(%[dst]) \n\t" /* store */ + "sw %[tp3], 8(%[dst]) \n\t" /* store */ + "sw %[tp4], 12(%[dst]) \n\t" /* store */ + + : [tp1] "=&r"(tp1), [tp2] "=&r"(tp2), [tp3] "=&r"(tp3), + [tp4] "=&r"(tp4) + : [src] "r"(src), [dst] "r"(dst)); + + src += src_stride; + dst += dst_stride; + } + } break; + case 32: { + uint32_t tp1, tp2, tp3, tp4; + uint32_t tp5, tp6, tp7, tp8; + + /* 8 word storage */ + for (y = h; y--;) { + prefetch_load(src + src_stride); + prefetch_load(src + src_stride + 32); + prefetch_store(dst + dst_stride); + + __asm__ __volatile__( + "ulw %[tp1], 0(%[src]) \n\t" + "ulw %[tp2], 4(%[src]) \n\t" + "ulw %[tp3], 8(%[src]) \n\t" + "ulw %[tp4], 12(%[src]) \n\t" + "ulw %[tp5], 16(%[src]) \n\t" + "ulw %[tp6], 20(%[src]) \n\t" + "ulw %[tp7], 24(%[src]) \n\t" + "ulw %[tp8], 28(%[src]) \n\t" + + "sw %[tp1], 0(%[dst]) \n\t" /* store */ + "sw %[tp2], 4(%[dst]) \n\t" /* store */ + "sw %[tp3], 8(%[dst]) \n\t" /* store */ + "sw %[tp4], 12(%[dst]) \n\t" /* store */ + "sw %[tp5], 16(%[dst]) \n\t" /* store */ + "sw %[tp6], 20(%[dst]) \n\t" /* store */ + "sw %[tp7], 24(%[dst]) \n\t" /* store */ + "sw %[tp8], 28(%[dst]) \n\t" /* store */ + + : [tp1] "=&r"(tp1), [tp2] "=&r"(tp2), [tp3] "=&r"(tp3), + [tp4] "=&r"(tp4), [tp5] "=&r"(tp5), [tp6] "=&r"(tp6), + [tp7] "=&r"(tp7), [tp8] "=&r"(tp8) + : [src] "r"(src), [dst] "r"(dst)); + + src += src_stride; + dst += dst_stride; + } + } break; + case 64: { + uint32_t tp1, tp2, tp3, tp4; + uint32_t tp5, tp6, tp7, tp8; + + prefetch_load(src + 64); + prefetch_store(dst + 32); + + /* 16 word storage */ + for (y = h; y--;) { + prefetch_load(src + src_stride); + prefetch_load(src + src_stride + 32); + prefetch_load(src + src_stride + 64); + prefetch_store(dst + dst_stride); + prefetch_store(dst + dst_stride + 32); + + __asm__ __volatile__( + "ulw %[tp1], 0(%[src]) \n\t" + "ulw %[tp2], 4(%[src]) \n\t" + "ulw %[tp3], 8(%[src]) \n\t" + "ulw %[tp4], 12(%[src]) \n\t" + "ulw %[tp5], 16(%[src]) \n\t" + "ulw %[tp6], 20(%[src]) \n\t" + "ulw %[tp7], 24(%[src]) \n\t" + "ulw %[tp8], 28(%[src]) \n\t" + + "sw %[tp1], 0(%[dst]) \n\t" /* store */ + "sw %[tp2], 4(%[dst]) \n\t" /* store */ + "sw %[tp3], 8(%[dst]) \n\t" /* store */ + "sw %[tp4], 12(%[dst]) \n\t" /* store */ + "sw %[tp5], 16(%[dst]) \n\t" /* store */ + "sw %[tp6], 20(%[dst]) \n\t" /* store */ + "sw %[tp7], 24(%[dst]) \n\t" /* store */ + "sw %[tp8], 28(%[dst]) \n\t" /* store */ + + "ulw %[tp1], 32(%[src]) \n\t" + "ulw %[tp2], 36(%[src]) \n\t" + "ulw %[tp3], 40(%[src]) \n\t" + "ulw %[tp4], 44(%[src]) \n\t" + "ulw %[tp5], 48(%[src]) \n\t" + "ulw %[tp6], 52(%[src]) \n\t" + "ulw %[tp7], 56(%[src]) \n\t" + "ulw %[tp8], 60(%[src]) \n\t" + + "sw %[tp1], 32(%[dst]) \n\t" /* store */ + "sw %[tp2], 36(%[dst]) \n\t" /* store */ + "sw %[tp3], 40(%[dst]) \n\t" /* store */ + "sw %[tp4], 44(%[dst]) \n\t" /* store */ + "sw %[tp5], 48(%[dst]) \n\t" /* store */ + "sw %[tp6], 52(%[dst]) \n\t" /* store */ + "sw %[tp7], 56(%[dst]) \n\t" /* store */ + "sw %[tp8], 60(%[dst]) \n\t" /* store */ + + : [tp1] "=&r"(tp1), [tp2] "=&r"(tp2), [tp3] "=&r"(tp3), + [tp4] "=&r"(tp4), [tp5] "=&r"(tp5), [tp6] "=&r"(tp6), + [tp7] "=&r"(tp7), [tp8] "=&r"(tp8) + : [src] "r"(src), [dst] "r"(dst)); + + src += src_stride; + dst += dst_stride; + } + } break; + default: + for (y = h; y--;) { + for (x = 0; x < w; ++x) { + dst[x] = src[x]; + } + + src += src_stride; + dst += dst_stride; + } + break; + } +} +#endif diff --git a/third_party/aom/aom_dsp/mips/convolve8_horiz_dspr2.c b/third_party/aom/aom_dsp/mips/convolve8_horiz_dspr2.c new file mode 100644 index 0000000000..f9c6879abd --- /dev/null +++ b/third_party/aom/aom_dsp/mips/convolve8_horiz_dspr2.c @@ -0,0 +1,879 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include +#include + +#include "config/aom_dsp_rtcd.h" + +#include "aom_dsp/mips/convolve_common_dspr2.h" +#include "aom_dsp/aom_dsp_common.h" +#include "aom_dsp/aom_filter.h" +#include "aom_ports/mem.h" + +#if HAVE_DSPR2 +static void convolve_horiz_4_dspr2(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + const int16_t *filter_x0, int32_t h) { + int32_t y; + uint8_t *cm = aom_ff_cropTbl; + int32_t vector1b, vector2b, vector3b, vector4b; + int32_t Temp1, Temp2, Temp3, Temp4; + uint32_t vector4a = 64; + uint32_t tp1, tp2; + uint32_t p1, p2, p3, p4; + uint32_t n1, n2, n3, n4; + uint32_t tn1, tn2; + + vector1b = ((const int32_t *)filter_x0)[0]; + vector2b = ((const int32_t *)filter_x0)[1]; + vector3b = ((const int32_t *)filter_x0)[2]; + vector4b = ((const int32_t *)filter_x0)[3]; + + for (y = h; y--;) { + /* prefetch data to cache memory */ + prefetch_load(src + src_stride); + prefetch_load(src + src_stride + 32); + prefetch_store(dst + dst_stride); + + __asm__ __volatile__( + "ulw %[tp1], 0(%[src]) \n\t" + "ulw %[tp2], 4(%[src]) \n\t" + + /* even 1. pixel */ + "mtlo %[vector4a], $ac3 \n\t" + "mthi $zero, $ac3 \n\t" + "preceu.ph.qbr %[p1], %[tp1] \n\t" + "preceu.ph.qbl %[p2], %[tp1] \n\t" + "preceu.ph.qbr %[p3], %[tp2] \n\t" + "preceu.ph.qbl %[p4], %[tp2] \n\t" + "dpa.w.ph $ac3, %[p1], %[vector1b] \n\t" + "dpa.w.ph $ac3, %[p2], %[vector2b] \n\t" + "dpa.w.ph $ac3, %[p3], %[vector3b] \n\t" + "ulw %[tn2], 8(%[src]) \n\t" + "dpa.w.ph $ac3, %[p4], %[vector4b] \n\t" + "extp %[Temp1], $ac3, 31 \n\t" + + /* even 2. pixel */ + "mtlo %[vector4a], $ac2 \n\t" + "mthi $zero, $ac2 \n\t" + "preceu.ph.qbr %[p1], %[tn2] \n\t" + "balign %[tn1], %[tn2], 3 \n\t" + "balign %[tn2], %[tp2], 3 \n\t" + "balign %[tp2], %[tp1], 3 \n\t" + "dpa.w.ph $ac2, %[p2], %[vector1b] \n\t" + "dpa.w.ph $ac2, %[p3], %[vector2b] \n\t" + "dpa.w.ph $ac2, %[p4], %[vector3b] \n\t" + "dpa.w.ph $ac2, %[p1], %[vector4b] \n\t" + "extp %[Temp3], $ac2, 31 \n\t" + + /* odd 1. pixel */ + "lbux %[tp1], %[Temp1](%[cm]) \n\t" + "mtlo %[vector4a], $ac3 \n\t" + "mthi $zero, $ac3 \n\t" + "preceu.ph.qbr %[n1], %[tp2] \n\t" + "preceu.ph.qbl %[n2], %[tp2] \n\t" + "preceu.ph.qbr %[n3], %[tn2] \n\t" + "preceu.ph.qbl %[n4], %[tn2] \n\t" + "dpa.w.ph $ac3, %[n1], %[vector1b] \n\t" + "dpa.w.ph $ac3, %[n2], %[vector2b] \n\t" + "dpa.w.ph $ac3, %[n3], %[vector3b] \n\t" + "dpa.w.ph $ac3, %[n4], %[vector4b] \n\t" + "extp %[Temp2], $ac3, 31 \n\t" + + /* odd 2. pixel */ + "lbux %[tp2], %[Temp3](%[cm]) \n\t" + "mtlo %[vector4a], $ac2 \n\t" + "mthi $zero, $ac2 \n\t" + "preceu.ph.qbr %[n1], %[tn1] \n\t" + "dpa.w.ph $ac2, %[n2], %[vector1b] \n\t" + "dpa.w.ph $ac2, %[n3], %[vector2b] \n\t" + "dpa.w.ph $ac2, %[n4], %[vector3b] \n\t" + "dpa.w.ph $ac2, %[n1], %[vector4b] \n\t" + "extp %[Temp4], $ac2, 31 \n\t" + + /* clamp */ + "lbux %[tn1], %[Temp2](%[cm]) \n\t" + "lbux %[n2], %[Temp4](%[cm]) \n\t" + + /* store bytes */ + "sb %[tp1], 0(%[dst]) \n\t" + "sb %[tn1], 1(%[dst]) \n\t" + "sb %[tp2], 2(%[dst]) \n\t" + "sb %[n2], 3(%[dst]) \n\t" + + : [tp1] "=&r"(tp1), [tp2] "=&r"(tp2), [tn1] "=&r"(tn1), + [tn2] "=&r"(tn2), [p1] "=&r"(p1), [p2] "=&r"(p2), [p3] "=&r"(p3), + [p4] "=&r"(p4), [n1] "=&r"(n1), [n2] "=&r"(n2), [n3] "=&r"(n3), + [n4] "=&r"(n4), [Temp1] "=&r"(Temp1), [Temp2] "=&r"(Temp2), + [Temp3] "=&r"(Temp3), [Temp4] "=&r"(Temp4) + : [vector1b] "r"(vector1b), [vector2b] "r"(vector2b), + [vector3b] "r"(vector3b), [vector4b] "r"(vector4b), + [vector4a] "r"(vector4a), [cm] "r"(cm), [dst] "r"(dst), + [src] "r"(src)); + + /* Next row... */ + src += src_stride; + dst += dst_stride; + } +} + +static void convolve_horiz_8_dspr2(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + const int16_t *filter_x0, int32_t h) { + int32_t y; + uint8_t *cm = aom_ff_cropTbl; + uint32_t vector4a = 64; + int32_t vector1b, vector2b, vector3b, vector4b; + int32_t Temp1, Temp2, Temp3; + uint32_t tp1, tp2; + uint32_t p1, p2, p3, p4, n1; + uint32_t tn1, tn2, tn3; + uint32_t st0, st1; + + vector1b = ((const int32_t *)filter_x0)[0]; + vector2b = ((const int32_t *)filter_x0)[1]; + vector3b = ((const int32_t *)filter_x0)[2]; + vector4b = ((const int32_t *)filter_x0)[3]; + + for (y = h; y--;) { + /* prefetch data to cache memory */ + prefetch_load(src + src_stride); + prefetch_load(src + src_stride + 32); + prefetch_store(dst + dst_stride); + + __asm__ __volatile__( + "ulw %[tp1], 0(%[src]) \n\t" + "ulw %[tp2], 4(%[src]) \n\t" + + /* even 1. pixel */ + "mtlo %[vector4a], $ac3 \n\t" + "mthi $zero, $ac3 \n\t" + "mtlo %[vector4a], $ac2 \n\t" + "mthi $zero, $ac2 \n\t" + "preceu.ph.qbr %[p1], %[tp1] \n\t" + "preceu.ph.qbl %[p2], %[tp1] \n\t" + "preceu.ph.qbr %[p3], %[tp2] \n\t" + "preceu.ph.qbl %[p4], %[tp2] \n\t" + "ulw %[tn2], 8(%[src]) \n\t" + "dpa.w.ph $ac3, %[p1], %[vector1b] \n\t" + "dpa.w.ph $ac3, %[p2], %[vector2b] \n\t" + "dpa.w.ph $ac3, %[p3], %[vector3b] \n\t" + "dpa.w.ph $ac3, %[p4], %[vector4b] \n\t" + "extp %[Temp1], $ac3, 31 \n\t" + + /* even 2. pixel */ + "preceu.ph.qbr %[p1], %[tn2] \n\t" + "preceu.ph.qbl %[n1], %[tn2] \n\t" + "ulw %[tn1], 12(%[src]) \n\t" + "dpa.w.ph $ac2, %[p2], %[vector1b] \n\t" + "dpa.w.ph $ac2, %[p3], %[vector2b] \n\t" + "dpa.w.ph $ac2, %[p4], %[vector3b] \n\t" + "dpa.w.ph $ac2, %[p1], %[vector4b] \n\t" + "extp %[Temp3], $ac2, 31 \n\t" + + /* even 3. pixel */ + "lbux %[st0], %[Temp1](%[cm]) \n\t" + "mtlo %[vector4a], $ac1 \n\t" + "mthi $zero, $ac1 \n\t" + "preceu.ph.qbr %[p2], %[tn1] \n\t" + "dpa.w.ph $ac1, %[p3], %[vector1b] \n\t" + "dpa.w.ph $ac1, %[p4], %[vector2b] \n\t" + "dpa.w.ph $ac1, %[p1], %[vector3b] \n\t" + "dpa.w.ph $ac1, %[n1], %[vector4b] \n\t" + "extp %[Temp1], $ac1, 31 \n\t" + + /* even 4. pixel */ + "mtlo %[vector4a], $ac2 \n\t" + "mthi $zero, $ac2 \n\t" + "mtlo %[vector4a], $ac3 \n\t" + "mthi $zero, $ac3 \n\t" + "sb %[st0], 0(%[dst]) \n\t" + "lbux %[st1], %[Temp3](%[cm]) \n\t" + + "balign %[tn3], %[tn1], 3 \n\t" + "balign %[tn1], %[tn2], 3 \n\t" + "balign %[tn2], %[tp2], 3 \n\t" + "balign %[tp2], %[tp1], 3 \n\t" + + "dpa.w.ph $ac2, %[p4], %[vector1b] \n\t" + "dpa.w.ph $ac2, %[p1], %[vector2b] \n\t" + "dpa.w.ph $ac2, %[n1], %[vector3b] \n\t" + "dpa.w.ph $ac2, %[p2], %[vector4b] \n\t" + "extp %[Temp3], $ac2, 31 \n\t" + + "lbux %[st0], %[Temp1](%[cm]) \n\t" + + /* odd 1. pixel */ + "mtlo %[vector4a], $ac1 \n\t" + "mthi $zero, $ac1 \n\t" + "sb %[st1], 2(%[dst]) \n\t" + "preceu.ph.qbr %[p1], %[tp2] \n\t" + "preceu.ph.qbl %[p2], %[tp2] \n\t" + "preceu.ph.qbr %[p3], %[tn2] \n\t" + "preceu.ph.qbl %[p4], %[tn2] \n\t" + "sb %[st0], 4(%[dst]) \n\t" + "dpa.w.ph $ac3, %[p1], %[vector1b] \n\t" + "dpa.w.ph $ac3, %[p2], %[vector2b] \n\t" + "dpa.w.ph $ac3, %[p3], %[vector3b] \n\t" + "dpa.w.ph $ac3, %[p4], %[vector4b] \n\t" + "extp %[Temp2], $ac3, 31 \n\t" + + /* odd 2. pixel */ + "mtlo %[vector4a], $ac3 \n\t" + "mthi $zero, $ac3 \n\t" + "mtlo %[vector4a], $ac2 \n\t" + "mthi $zero, $ac2 \n\t" + "preceu.ph.qbr %[p1], %[tn1] \n\t" + "preceu.ph.qbl %[n1], %[tn1] \n\t" + "lbux %[st0], %[Temp3](%[cm]) \n\t" + "dpa.w.ph $ac1, %[p2], %[vector1b] \n\t" + "dpa.w.ph $ac1, %[p3], %[vector2b] \n\t" + "dpa.w.ph $ac1, %[p4], %[vector3b] \n\t" + "dpa.w.ph $ac1, %[p1], %[vector4b] \n\t" + "extp %[Temp3], $ac1, 31 \n\t" + + /* odd 3. pixel */ + "lbux %[st1], %[Temp2](%[cm]) \n\t" + "preceu.ph.qbr %[p2], %[tn3] \n\t" + "dpa.w.ph $ac3, %[p3], %[vector1b] \n\t" + "dpa.w.ph $ac3, %[p4], %[vector2b] \n\t" + "dpa.w.ph $ac3, %[p1], %[vector3b] \n\t" + "dpa.w.ph $ac3, %[n1], %[vector4b] \n\t" + "extp %[Temp2], $ac3, 31 \n\t" + + /* odd 4. pixel */ + "sb %[st1], 1(%[dst]) \n\t" + "sb %[st0], 6(%[dst]) \n\t" + "dpa.w.ph $ac2, %[p4], %[vector1b] \n\t" + "dpa.w.ph $ac2, %[p1], %[vector2b] \n\t" + "dpa.w.ph $ac2, %[n1], %[vector3b] \n\t" + "dpa.w.ph $ac2, %[p2], %[vector4b] \n\t" + "extp %[Temp1], $ac2, 31 \n\t" + + /* clamp */ + "lbux %[p4], %[Temp3](%[cm]) \n\t" + "lbux %[p2], %[Temp2](%[cm]) \n\t" + "lbux %[n1], %[Temp1](%[cm]) \n\t" + + /* store bytes */ + "sb %[p4], 3(%[dst]) \n\t" + "sb %[p2], 5(%[dst]) \n\t" + "sb %[n1], 7(%[dst]) \n\t" + + : [tp1] "=&r"(tp1), [tp2] "=&r"(tp2), [tn1] "=&r"(tn1), + [tn2] "=&r"(tn2), [tn3] "=&r"(tn3), [st0] "=&r"(st0), + [st1] "=&r"(st1), [p1] "=&r"(p1), [p2] "=&r"(p2), [p3] "=&r"(p3), + [p4] "=&r"(p4), [n1] "=&r"(n1), [Temp1] "=&r"(Temp1), + [Temp2] "=&r"(Temp2), [Temp3] "=&r"(Temp3) + : [vector1b] "r"(vector1b), [vector2b] "r"(vector2b), + [vector3b] "r"(vector3b), [vector4b] "r"(vector4b), + [vector4a] "r"(vector4a), [cm] "r"(cm), [dst] "r"(dst), + [src] "r"(src)); + + /* Next row... */ + src += src_stride; + dst += dst_stride; + } +} + +static void convolve_horiz_16_dspr2(const uint8_t *src_ptr, int32_t src_stride, + uint8_t *dst_ptr, int32_t dst_stride, + const int16_t *filter_x0, int32_t h, + int32_t count) { + int32_t y, c; + const uint8_t *src; + uint8_t *dst; + uint8_t *cm = aom_ff_cropTbl; + uint32_t vector_64 = 64; + int32_t filter12, filter34, filter56, filter78; + int32_t Temp1, Temp2, Temp3; + uint32_t qload1, qload2, qload3; + uint32_t p1, p2, p3, p4, p5; + uint32_t st1, st2, st3; + + filter12 = ((const int32_t *)filter_x0)[0]; + filter34 = ((const int32_t *)filter_x0)[1]; + filter56 = ((const int32_t *)filter_x0)[2]; + filter78 = ((const int32_t *)filter_x0)[3]; + + for (y = h; y--;) { + src = src_ptr; + dst = dst_ptr; + + /* prefetch data to cache memory */ + prefetch_load(src_ptr + src_stride); + prefetch_load(src_ptr + src_stride + 32); + prefetch_store(dst_ptr + dst_stride); + + for (c = 0; c < count; c++) { + __asm__ __volatile__( + "ulw %[qload1], 0(%[src]) \n\t" + "ulw %[qload2], 4(%[src]) \n\t" + + /* even 1. pixel */ + "mtlo %[vector_64], $ac1 \n\t" /* even 1 */ + "mthi $zero, $ac1 \n\t" + "mtlo %[vector_64], $ac2 \n\t" /* even 2 */ + "mthi $zero, $ac2 \n\t" + "preceu.ph.qbr %[p1], %[qload1] \n\t" + "preceu.ph.qbl %[p2], %[qload1] \n\t" + "preceu.ph.qbr %[p3], %[qload2] \n\t" + "preceu.ph.qbl %[p4], %[qload2] \n\t" + "ulw %[qload3], 8(%[src]) \n\t" + "dpa.w.ph $ac1, %[p1], %[filter12] \n\t" /* even 1 */ + "dpa.w.ph $ac1, %[p2], %[filter34] \n\t" /* even 1 */ + "dpa.w.ph $ac1, %[p3], %[filter56] \n\t" /* even 1 */ + "dpa.w.ph $ac1, %[p4], %[filter78] \n\t" /* even 1 */ + "extp %[Temp1], $ac1, 31 \n\t" /* even 1 */ + + /* even 2. pixel */ + "mtlo %[vector_64], $ac3 \n\t" /* even 3 */ + "mthi $zero, $ac3 \n\t" + "preceu.ph.qbr %[p1], %[qload3] \n\t" + "preceu.ph.qbl %[p5], %[qload3] \n\t" + "ulw %[qload1], 12(%[src]) \n\t" + "dpa.w.ph $ac2, %[p2], %[filter12] \n\t" /* even 1 */ + "dpa.w.ph $ac2, %[p3], %[filter34] \n\t" /* even 1 */ + "dpa.w.ph $ac2, %[p4], %[filter56] \n\t" /* even 1 */ + "dpa.w.ph $ac2, %[p1], %[filter78] \n\t" /* even 1 */ + "extp %[Temp2], $ac2, 31 \n\t" /* even 1 */ + "lbux %[st1], %[Temp1](%[cm]) \n\t" /* even 1 */ + + /* even 3. pixel */ + "mtlo %[vector_64], $ac1 \n\t" /* even 4 */ + "mthi $zero, $ac1 \n\t" + "preceu.ph.qbr %[p2], %[qload1] \n\t" + "sb %[st1], 0(%[dst]) \n\t" /* even 1 */ + "dpa.w.ph $ac3, %[p3], %[filter12] \n\t" /* even 3 */ + "dpa.w.ph $ac3, %[p4], %[filter34] \n\t" /* even 3 */ + "dpa.w.ph $ac3, %[p1], %[filter56] \n\t" /* even 3 */ + "dpa.w.ph $ac3, %[p5], %[filter78] \n\t" /* even 3 */ + "extp %[Temp3], $ac3, 31 \n\t" /* even 3 */ + "lbux %[st2], %[Temp2](%[cm]) \n\t" /* even 1 */ + + /* even 4. pixel */ + "mtlo %[vector_64], $ac2 \n\t" /* even 5 */ + "mthi $zero, $ac2 \n\t" + "preceu.ph.qbl %[p3], %[qload1] \n\t" + "sb %[st2], 2(%[dst]) \n\t" /* even 1 */ + "ulw %[qload2], 16(%[src]) \n\t" + "dpa.w.ph $ac1, %[p4], %[filter12] \n\t" /* even 4 */ + "dpa.w.ph $ac1, %[p1], %[filter34] \n\t" /* even 4 */ + "dpa.w.ph $ac1, %[p5], %[filter56] \n\t" /* even 4 */ + "dpa.w.ph $ac1, %[p2], %[filter78] \n\t" /* even 4 */ + "extp %[Temp1], $ac1, 31 \n\t" /* even 4 */ + "lbux %[st3], %[Temp3](%[cm]) \n\t" /* even 3 */ + + /* even 5. pixel */ + "mtlo %[vector_64], $ac3 \n\t" /* even 6 */ + "mthi $zero, $ac3 \n\t" + "preceu.ph.qbr %[p4], %[qload2] \n\t" + "sb %[st3], 4(%[dst]) \n\t" /* even 3 */ + "dpa.w.ph $ac2, %[p1], %[filter12] \n\t" /* even 5 */ + "dpa.w.ph $ac2, %[p5], %[filter34] \n\t" /* even 5 */ + "dpa.w.ph $ac2, %[p2], %[filter56] \n\t" /* even 5 */ + "dpa.w.ph $ac2, %[p3], %[filter78] \n\t" /* even 5 */ + "extp %[Temp2], $ac2, 31 \n\t" /* even 5 */ + "lbux %[st1], %[Temp1](%[cm]) \n\t" /* even 4 */ + + /* even 6. pixel */ + "mtlo %[vector_64], $ac1 \n\t" /* even 7 */ + "mthi $zero, $ac1 \n\t" + "preceu.ph.qbl %[p1], %[qload2] \n\t" + "sb %[st1], 6(%[dst]) \n\t" /* even 4 */ + "ulw %[qload3], 20(%[src]) \n\t" + "dpa.w.ph $ac3, %[p5], %[filter12] \n\t" /* even 6 */ + "dpa.w.ph $ac3, %[p2], %[filter34] \n\t" /* even 6 */ + "dpa.w.ph $ac3, %[p3], %[filter56] \n\t" /* even 6 */ + "dpa.w.ph $ac3, %[p4], %[filter78] \n\t" /* even 6 */ + "extp %[Temp3], $ac3, 31 \n\t" /* even 6 */ + "lbux %[st2], %[Temp2](%[cm]) \n\t" /* even 5 */ + + /* even 7. pixel */ + "mtlo %[vector_64], $ac2 \n\t" /* even 8 */ + "mthi $zero, $ac2 \n\t" + "preceu.ph.qbr %[p5], %[qload3] \n\t" + "sb %[st2], 8(%[dst]) \n\t" /* even 5 */ + "dpa.w.ph $ac1, %[p2], %[filter12] \n\t" /* even 7 */ + "dpa.w.ph $ac1, %[p3], %[filter34] \n\t" /* even 7 */ + "dpa.w.ph $ac1, %[p4], %[filter56] \n\t" /* even 7 */ + "dpa.w.ph $ac1, %[p1], %[filter78] \n\t" /* even 7 */ + "extp %[Temp1], $ac1, 31 \n\t" /* even 7 */ + "lbux %[st3], %[Temp3](%[cm]) \n\t" /* even 6 */ + + /* even 8. pixel */ + "mtlo %[vector_64], $ac3 \n\t" /* odd 1 */ + "mthi $zero, $ac3 \n\t" + "dpa.w.ph $ac2, %[p3], %[filter12] \n\t" /* even 8 */ + "dpa.w.ph $ac2, %[p4], %[filter34] \n\t" /* even 8 */ + "sb %[st3], 10(%[dst]) \n\t" /* even 6 */ + "dpa.w.ph $ac2, %[p1], %[filter56] \n\t" /* even 8 */ + "dpa.w.ph $ac2, %[p5], %[filter78] \n\t" /* even 8 */ + "extp %[Temp2], $ac2, 31 \n\t" /* even 8 */ + "lbux %[st1], %[Temp1](%[cm]) \n\t" /* even 7 */ + + /* ODD pixels */ + "ulw %[qload1], 1(%[src]) \n\t" + "ulw %[qload2], 5(%[src]) \n\t" + + /* odd 1. pixel */ + "mtlo %[vector_64], $ac1 \n\t" /* odd 2 */ + "mthi $zero, $ac1 \n\t" + "preceu.ph.qbr %[p1], %[qload1] \n\t" + "preceu.ph.qbl %[p2], %[qload1] \n\t" + "preceu.ph.qbr %[p3], %[qload2] \n\t" + "preceu.ph.qbl %[p4], %[qload2] \n\t" + "sb %[st1], 12(%[dst]) \n\t" /* even 7 */ + "ulw %[qload3], 9(%[src]) \n\t" + "dpa.w.ph $ac3, %[p1], %[filter12] \n\t" /* odd 1 */ + "dpa.w.ph $ac3, %[p2], %[filter34] \n\t" /* odd 1 */ + "dpa.w.ph $ac3, %[p3], %[filter56] \n\t" /* odd 1 */ + "dpa.w.ph $ac3, %[p4], %[filter78] \n\t" /* odd 1 */ + "extp %[Temp3], $ac3, 31 \n\t" /* odd 1 */ + "lbux %[st2], %[Temp2](%[cm]) \n\t" /* even 8 */ + + /* odd 2. pixel */ + "mtlo %[vector_64], $ac2 \n\t" /* odd 3 */ + "mthi $zero, $ac2 \n\t" + "preceu.ph.qbr %[p1], %[qload3] \n\t" + "preceu.ph.qbl %[p5], %[qload3] \n\t" + "sb %[st2], 14(%[dst]) \n\t" /* even 8 */ + "ulw %[qload1], 13(%[src]) \n\t" + "dpa.w.ph $ac1, %[p2], %[filter12] \n\t" /* odd 2 */ + "dpa.w.ph $ac1, %[p3], %[filter34] \n\t" /* odd 2 */ + "dpa.w.ph $ac1, %[p4], %[filter56] \n\t" /* odd 2 */ + "dpa.w.ph $ac1, %[p1], %[filter78] \n\t" /* odd 2 */ + "extp %[Temp1], $ac1, 31 \n\t" /* odd 2 */ + "lbux %[st3], %[Temp3](%[cm]) \n\t" /* odd 1 */ + + /* odd 3. pixel */ + "mtlo %[vector_64], $ac3 \n\t" /* odd 4 */ + "mthi $zero, $ac3 \n\t" + "preceu.ph.qbr %[p2], %[qload1] \n\t" + "sb %[st3], 1(%[dst]) \n\t" /* odd 1 */ + "dpa.w.ph $ac2, %[p3], %[filter12] \n\t" /* odd 3 */ + "dpa.w.ph $ac2, %[p4], %[filter34] \n\t" /* odd 3 */ + "dpa.w.ph $ac2, %[p1], %[filter56] \n\t" /* odd 3 */ + "dpa.w.ph $ac2, %[p5], %[filter78] \n\t" /* odd 3 */ + "extp %[Temp2], $ac2, 31 \n\t" /* odd 3 */ + "lbux %[st1], %[Temp1](%[cm]) \n\t" /* odd 2 */ + + /* odd 4. pixel */ + "mtlo %[vector_64], $ac1 \n\t" /* odd 5 */ + "mthi $zero, $ac1 \n\t" + "preceu.ph.qbl %[p3], %[qload1] \n\t" + "sb %[st1], 3(%[dst]) \n\t" /* odd 2 */ + "ulw %[qload2], 17(%[src]) \n\t" + "dpa.w.ph $ac3, %[p4], %[filter12] \n\t" /* odd 4 */ + "dpa.w.ph $ac3, %[p1], %[filter34] \n\t" /* odd 4 */ + "dpa.w.ph $ac3, %[p5], %[filter56] \n\t" /* odd 4 */ + "dpa.w.ph $ac3, %[p2], %[filter78] \n\t" /* odd 4 */ + "extp %[Temp3], $ac3, 31 \n\t" /* odd 4 */ + "lbux %[st2], %[Temp2](%[cm]) \n\t" /* odd 3 */ + + /* odd 5. pixel */ + "mtlo %[vector_64], $ac2 \n\t" /* odd 6 */ + "mthi $zero, $ac2 \n\t" + "preceu.ph.qbr %[p4], %[qload2] \n\t" + "sb %[st2], 5(%[dst]) \n\t" /* odd 3 */ + "dpa.w.ph $ac1, %[p1], %[filter12] \n\t" /* odd 5 */ + "dpa.w.ph $ac1, %[p5], %[filter34] \n\t" /* odd 5 */ + "dpa.w.ph $ac1, %[p2], %[filter56] \n\t" /* odd 5 */ + "dpa.w.ph $ac1, %[p3], %[filter78] \n\t" /* odd 5 */ + "extp %[Temp1], $ac1, 31 \n\t" /* odd 5 */ + "lbux %[st3], %[Temp3](%[cm]) \n\t" /* odd 4 */ + + /* odd 6. pixel */ + "mtlo %[vector_64], $ac3 \n\t" /* odd 7 */ + "mthi $zero, $ac3 \n\t" + "preceu.ph.qbl %[p1], %[qload2] \n\t" + "sb %[st3], 7(%[dst]) \n\t" /* odd 4 */ + "ulw %[qload3], 21(%[src]) \n\t" + "dpa.w.ph $ac2, %[p5], %[filter12] \n\t" /* odd 6 */ + "dpa.w.ph $ac2, %[p2], %[filter34] \n\t" /* odd 6 */ + "dpa.w.ph $ac2, %[p3], %[filter56] \n\t" /* odd 6 */ + "dpa.w.ph $ac2, %[p4], %[filter78] \n\t" /* odd 6 */ + "extp %[Temp2], $ac2, 31 \n\t" /* odd 6 */ + "lbux %[st1], %[Temp1](%[cm]) \n\t" /* odd 5 */ + + /* odd 7. pixel */ + "mtlo %[vector_64], $ac1 \n\t" /* odd 8 */ + "mthi $zero, $ac1 \n\t" + "preceu.ph.qbr %[p5], %[qload3] \n\t" + "sb %[st1], 9(%[dst]) \n\t" /* odd 5 */ + "dpa.w.ph $ac3, %[p2], %[filter12] \n\t" /* odd 7 */ + "dpa.w.ph $ac3, %[p3], %[filter34] \n\t" /* odd 7 */ + "dpa.w.ph $ac3, %[p4], %[filter56] \n\t" /* odd 7 */ + "dpa.w.ph $ac3, %[p1], %[filter78] \n\t" /* odd 7 */ + "extp %[Temp3], $ac3, 31 \n\t" /* odd 7 */ + + /* odd 8. pixel */ + "dpa.w.ph $ac1, %[p3], %[filter12] \n\t" /* odd 8 */ + "dpa.w.ph $ac1, %[p4], %[filter34] \n\t" /* odd 8 */ + "dpa.w.ph $ac1, %[p1], %[filter56] \n\t" /* odd 8 */ + "dpa.w.ph $ac1, %[p5], %[filter78] \n\t" /* odd 8 */ + "extp %[Temp1], $ac1, 31 \n\t" /* odd 8 */ + + "lbux %[st2], %[Temp2](%[cm]) \n\t" /* odd 6 */ + "lbux %[st3], %[Temp3](%[cm]) \n\t" /* odd 7 */ + "lbux %[st1], %[Temp1](%[cm]) \n\t" /* odd 8 */ + + "sb %[st2], 11(%[dst]) \n\t" /* odd 6 */ + "sb %[st3], 13(%[dst]) \n\t" /* odd 7 */ + "sb %[st1], 15(%[dst]) \n\t" /* odd 8 */ + + : [qload1] "=&r"(qload1), [qload2] "=&r"(qload2), + [qload3] "=&r"(qload3), [st1] "=&r"(st1), [st2] "=&r"(st2), + [st3] "=&r"(st3), [p1] "=&r"(p1), [p2] "=&r"(p2), [p3] "=&r"(p3), + [p4] "=&r"(p4), [p5] "=&r"(p5), [Temp1] "=&r"(Temp1), + [Temp2] "=&r"(Temp2), [Temp3] "=&r"(Temp3) + : [filter12] "r"(filter12), [filter34] "r"(filter34), + [filter56] "r"(filter56), [filter78] "r"(filter78), + [vector_64] "r"(vector_64), [cm] "r"(cm), [dst] "r"(dst), + [src] "r"(src)); + + src += 16; + dst += 16; + } + + /* Next row... */ + src_ptr += src_stride; + dst_ptr += dst_stride; + } +} + +static void convolve_horiz_64_dspr2(const uint8_t *src_ptr, int32_t src_stride, + uint8_t *dst_ptr, int32_t dst_stride, + const int16_t *filter_x0, int32_t h) { + int32_t y, c; + const uint8_t *src; + uint8_t *dst; + uint8_t *cm = aom_ff_cropTbl; + uint32_t vector_64 = 64; + int32_t filter12, filter34, filter56, filter78; + int32_t Temp1, Temp2, Temp3; + uint32_t qload1, qload2, qload3; + uint32_t p1, p2, p3, p4, p5; + uint32_t st1, st2, st3; + + filter12 = ((const int32_t *)filter_x0)[0]; + filter34 = ((const int32_t *)filter_x0)[1]; + filter56 = ((const int32_t *)filter_x0)[2]; + filter78 = ((const int32_t *)filter_x0)[3]; + + for (y = h; y--;) { + src = src_ptr; + dst = dst_ptr; + + /* prefetch data to cache memory */ + prefetch_load(src_ptr + src_stride); + prefetch_load(src_ptr + src_stride + 32); + prefetch_load(src_ptr + src_stride + 64); + prefetch_store(dst_ptr + dst_stride); + prefetch_store(dst_ptr + dst_stride + 32); + + for (c = 0; c < 4; c++) { + __asm__ __volatile__( + "ulw %[qload1], 0(%[src]) \n\t" + "ulw %[qload2], 4(%[src]) \n\t" + + /* even 1. pixel */ + "mtlo %[vector_64], $ac1 \n\t" /* even 1 */ + "mthi $zero, $ac1 \n\t" + "mtlo %[vector_64], $ac2 \n\t" /* even 2 */ + "mthi $zero, $ac2 \n\t" + "preceu.ph.qbr %[p1], %[qload1] \n\t" + "preceu.ph.qbl %[p2], %[qload1] \n\t" + "preceu.ph.qbr %[p3], %[qload2] \n\t" + "preceu.ph.qbl %[p4], %[qload2] \n\t" + "ulw %[qload3], 8(%[src]) \n\t" + "dpa.w.ph $ac1, %[p1], %[filter12] \n\t" /* even 1 */ + "dpa.w.ph $ac1, %[p2], %[filter34] \n\t" /* even 1 */ + "dpa.w.ph $ac1, %[p3], %[filter56] \n\t" /* even 1 */ + "dpa.w.ph $ac1, %[p4], %[filter78] \n\t" /* even 1 */ + "extp %[Temp1], $ac1, 31 \n\t" /* even 1 */ + + /* even 2. pixel */ + "mtlo %[vector_64], $ac3 \n\t" /* even 3 */ + "mthi $zero, $ac3 \n\t" + "preceu.ph.qbr %[p1], %[qload3] \n\t" + "preceu.ph.qbl %[p5], %[qload3] \n\t" + "ulw %[qload1], 12(%[src]) \n\t" + "dpa.w.ph $ac2, %[p2], %[filter12] \n\t" /* even 1 */ + "dpa.w.ph $ac2, %[p3], %[filter34] \n\t" /* even 1 */ + "dpa.w.ph $ac2, %[p4], %[filter56] \n\t" /* even 1 */ + "dpa.w.ph $ac2, %[p1], %[filter78] \n\t" /* even 1 */ + "extp %[Temp2], $ac2, 31 \n\t" /* even 1 */ + "lbux %[st1], %[Temp1](%[cm]) \n\t" /* even 1 */ + + /* even 3. pixel */ + "mtlo %[vector_64], $ac1 \n\t" /* even 4 */ + "mthi $zero, $ac1 \n\t" + "preceu.ph.qbr %[p2], %[qload1] \n\t" + "sb %[st1], 0(%[dst]) \n\t" /* even 1 */ + "dpa.w.ph $ac3, %[p3], %[filter12] \n\t" /* even 3 */ + "dpa.w.ph $ac3, %[p4], %[filter34] \n\t" /* even 3 */ + "dpa.w.ph $ac3, %[p1], %[filter56] \n\t" /* even 3 */ + "dpa.w.ph $ac3, %[p5], %[filter78] \n\t" /* even 3 */ + "extp %[Temp3], $ac3, 31 \n\t" /* even 3 */ + "lbux %[st2], %[Temp2](%[cm]) \n\t" /* even 1 */ + + /* even 4. pixel */ + "mtlo %[vector_64], $ac2 \n\t" /* even 5 */ + "mthi $zero, $ac2 \n\t" + "preceu.ph.qbl %[p3], %[qload1] \n\t" + "sb %[st2], 2(%[dst]) \n\t" /* even 1 */ + "ulw %[qload2], 16(%[src]) \n\t" + "dpa.w.ph $ac1, %[p4], %[filter12] \n\t" /* even 4 */ + "dpa.w.ph $ac1, %[p1], %[filter34] \n\t" /* even 4 */ + "dpa.w.ph $ac1, %[p5], %[filter56] \n\t" /* even 4 */ + "dpa.w.ph $ac1, %[p2], %[filter78] \n\t" /* even 4 */ + "extp %[Temp1], $ac1, 31 \n\t" /* even 4 */ + "lbux %[st3], %[Temp3](%[cm]) \n\t" /* even 3 */ + + /* even 5. pixel */ + "mtlo %[vector_64], $ac3 \n\t" /* even 6 */ + "mthi $zero, $ac3 \n\t" + "preceu.ph.qbr %[p4], %[qload2] \n\t" + "sb %[st3], 4(%[dst]) \n\t" /* even 3 */ + "dpa.w.ph $ac2, %[p1], %[filter12] \n\t" /* even 5 */ + "dpa.w.ph $ac2, %[p5], %[filter34] \n\t" /* even 5 */ + "dpa.w.ph $ac2, %[p2], %[filter56] \n\t" /* even 5 */ + "dpa.w.ph $ac2, %[p3], %[filter78] \n\t" /* even 5 */ + "extp %[Temp2], $ac2, 31 \n\t" /* even 5 */ + "lbux %[st1], %[Temp1](%[cm]) \n\t" /* even 4 */ + + /* even 6. pixel */ + "mtlo %[vector_64], $ac1 \n\t" /* even 7 */ + "mthi $zero, $ac1 \n\t" + "preceu.ph.qbl %[p1], %[qload2] \n\t" + "sb %[st1], 6(%[dst]) \n\t" /* even 4 */ + "ulw %[qload3], 20(%[src]) \n\t" + "dpa.w.ph $ac3, %[p5], %[filter12] \n\t" /* even 6 */ + "dpa.w.ph $ac3, %[p2], %[filter34] \n\t" /* even 6 */ + "dpa.w.ph $ac3, %[p3], %[filter56] \n\t" /* even 6 */ + "dpa.w.ph $ac3, %[p4], %[filter78] \n\t" /* even 6 */ + "extp %[Temp3], $ac3, 31 \n\t" /* even 6 */ + "lbux %[st2], %[Temp2](%[cm]) \n\t" /* even 5 */ + + /* even 7. pixel */ + "mtlo %[vector_64], $ac2 \n\t" /* even 8 */ + "mthi $zero, $ac2 \n\t" + "preceu.ph.qbr %[p5], %[qload3] \n\t" + "sb %[st2], 8(%[dst]) \n\t" /* even 5 */ + "dpa.w.ph $ac1, %[p2], %[filter12] \n\t" /* even 7 */ + "dpa.w.ph $ac1, %[p3], %[filter34] \n\t" /* even 7 */ + "dpa.w.ph $ac1, %[p4], %[filter56] \n\t" /* even 7 */ + "dpa.w.ph $ac1, %[p1], %[filter78] \n\t" /* even 7 */ + "extp %[Temp1], $ac1, 31 \n\t" /* even 7 */ + "lbux %[st3], %[Temp3](%[cm]) \n\t" /* even 6 */ + + /* even 8. pixel */ + "mtlo %[vector_64], $ac3 \n\t" /* odd 1 */ + "mthi $zero, $ac3 \n\t" + "dpa.w.ph $ac2, %[p3], %[filter12] \n\t" /* even 8 */ + "dpa.w.ph $ac2, %[p4], %[filter34] \n\t" /* even 8 */ + "sb %[st3], 10(%[dst]) \n\t" /* even 6 */ + "dpa.w.ph $ac2, %[p1], %[filter56] \n\t" /* even 8 */ + "dpa.w.ph $ac2, %[p5], %[filter78] \n\t" /* even 8 */ + "extp %[Temp2], $ac2, 31 \n\t" /* even 8 */ + "lbux %[st1], %[Temp1](%[cm]) \n\t" /* even 7 */ + + /* ODD pixels */ + "ulw %[qload1], 1(%[src]) \n\t" + "ulw %[qload2], 5(%[src]) \n\t" + + /* odd 1. pixel */ + "mtlo %[vector_64], $ac1 \n\t" /* odd 2 */ + "mthi $zero, $ac1 \n\t" + "preceu.ph.qbr %[p1], %[qload1] \n\t" + "preceu.ph.qbl %[p2], %[qload1] \n\t" + "preceu.ph.qbr %[p3], %[qload2] \n\t" + "preceu.ph.qbl %[p4], %[qload2] \n\t" + "sb %[st1], 12(%[dst]) \n\t" /* even 7 */ + "ulw %[qload3], 9(%[src]) \n\t" + "dpa.w.ph $ac3, %[p1], %[filter12] \n\t" /* odd 1 */ + "dpa.w.ph $ac3, %[p2], %[filter34] \n\t" /* odd 1 */ + "dpa.w.ph $ac3, %[p3], %[filter56] \n\t" /* odd 1 */ + "dpa.w.ph $ac3, %[p4], %[filter78] \n\t" /* odd 1 */ + "extp %[Temp3], $ac3, 31 \n\t" /* odd 1 */ + "lbux %[st2], %[Temp2](%[cm]) \n\t" /* even 8 */ + + /* odd 2. pixel */ + "mtlo %[vector_64], $ac2 \n\t" /* odd 3 */ + "mthi $zero, $ac2 \n\t" + "preceu.ph.qbr %[p1], %[qload3] \n\t" + "preceu.ph.qbl %[p5], %[qload3] \n\t" + "sb %[st2], 14(%[dst]) \n\t" /* even 8 */ + "ulw %[qload1], 13(%[src]) \n\t" + "dpa.w.ph $ac1, %[p2], %[filter12] \n\t" /* odd 2 */ + "dpa.w.ph $ac1, %[p3], %[filter34] \n\t" /* odd 2 */ + "dpa.w.ph $ac1, %[p4], %[filter56] \n\t" /* odd 2 */ + "dpa.w.ph $ac1, %[p1], %[filter78] \n\t" /* odd 2 */ + "extp %[Temp1], $ac1, 31 \n\t" /* odd 2 */ + "lbux %[st3], %[Temp3](%[cm]) \n\t" /* odd 1 */ + + /* odd 3. pixel */ + "mtlo %[vector_64], $ac3 \n\t" /* odd 4 */ + "mthi $zero, $ac3 \n\t" + "preceu.ph.qbr %[p2], %[qload1] \n\t" + "sb %[st3], 1(%[dst]) \n\t" /* odd 1 */ + "dpa.w.ph $ac2, %[p3], %[filter12] \n\t" /* odd 3 */ + "dpa.w.ph $ac2, %[p4], %[filter34] \n\t" /* odd 3 */ + "dpa.w.ph $ac2, %[p1], %[filter56] \n\t" /* odd 3 */ + "dpa.w.ph $ac2, %[p5], %[filter78] \n\t" /* odd 3 */ + "extp %[Temp2], $ac2, 31 \n\t" /* odd 3 */ + "lbux %[st1], %[Temp1](%[cm]) \n\t" /* odd 2 */ + + /* odd 4. pixel */ + "mtlo %[vector_64], $ac1 \n\t" /* odd 5 */ + "mthi $zero, $ac1 \n\t" + "preceu.ph.qbl %[p3], %[qload1] \n\t" + "sb %[st1], 3(%[dst]) \n\t" /* odd 2 */ + "ulw %[qload2], 17(%[src]) \n\t" + "dpa.w.ph $ac3, %[p4], %[filter12] \n\t" /* odd 4 */ + "dpa.w.ph $ac3, %[p1], %[filter34] \n\t" /* odd 4 */ + "dpa.w.ph $ac3, %[p5], %[filter56] \n\t" /* odd 4 */ + "dpa.w.ph $ac3, %[p2], %[filter78] \n\t" /* odd 4 */ + "extp %[Temp3], $ac3, 31 \n\t" /* odd 4 */ + "lbux %[st2], %[Temp2](%[cm]) \n\t" /* odd 3 */ + + /* odd 5. pixel */ + "mtlo %[vector_64], $ac2 \n\t" /* odd 6 */ + "mthi $zero, $ac2 \n\t" + "preceu.ph.qbr %[p4], %[qload2] \n\t" + "sb %[st2], 5(%[dst]) \n\t" /* odd 3 */ + "dpa.w.ph $ac1, %[p1], %[filter12] \n\t" /* odd 5 */ + "dpa.w.ph $ac1, %[p5], %[filter34] \n\t" /* odd 5 */ + "dpa.w.ph $ac1, %[p2], %[filter56] \n\t" /* odd 5 */ + "dpa.w.ph $ac1, %[p3], %[filter78] \n\t" /* odd 5 */ + "extp %[Temp1], $ac1, 31 \n\t" /* odd 5 */ + "lbux %[st3], %[Temp3](%[cm]) \n\t" /* odd 4 */ + + /* odd 6. pixel */ + "mtlo %[vector_64], $ac3 \n\t" /* odd 7 */ + "mthi $zero, $ac3 \n\t" + "preceu.ph.qbl %[p1], %[qload2] \n\t" + "sb %[st3], 7(%[dst]) \n\t" /* odd 4 */ + "ulw %[qload3], 21(%[src]) \n\t" + "dpa.w.ph $ac2, %[p5], %[filter12] \n\t" /* odd 6 */ + "dpa.w.ph $ac2, %[p2], %[filter34] \n\t" /* odd 6 */ + "dpa.w.ph $ac2, %[p3], %[filter56] \n\t" /* odd 6 */ + "dpa.w.ph $ac2, %[p4], %[filter78] \n\t" /* odd 6 */ + "extp %[Temp2], $ac2, 31 \n\t" /* odd 6 */ + "lbux %[st1], %[Temp1](%[cm]) \n\t" /* odd 5 */ + + /* odd 7. pixel */ + "mtlo %[vector_64], $ac1 \n\t" /* odd 8 */ + "mthi $zero, $ac1 \n\t" + "preceu.ph.qbr %[p5], %[qload3] \n\t" + "sb %[st1], 9(%[dst]) \n\t" /* odd 5 */ + "dpa.w.ph $ac3, %[p2], %[filter12] \n\t" /* odd 7 */ + "dpa.w.ph $ac3, %[p3], %[filter34] \n\t" /* odd 7 */ + "dpa.w.ph $ac3, %[p4], %[filter56] \n\t" /* odd 7 */ + "dpa.w.ph $ac3, %[p1], %[filter78] \n\t" /* odd 7 */ + "extp %[Temp3], $ac3, 31 \n\t" /* odd 7 */ + + /* odd 8. pixel */ + "dpa.w.ph $ac1, %[p3], %[filter12] \n\t" /* odd 8 */ + "dpa.w.ph $ac1, %[p4], %[filter34] \n\t" /* odd 8 */ + "dpa.w.ph $ac1, %[p1], %[filter56] \n\t" /* odd 8 */ + "dpa.w.ph $ac1, %[p5], %[filter78] \n\t" /* odd 8 */ + "extp %[Temp1], $ac1, 31 \n\t" /* odd 8 */ + + "lbux %[st2], %[Temp2](%[cm]) \n\t" /* odd 6 */ + "lbux %[st3], %[Temp3](%[cm]) \n\t" /* odd 7 */ + "lbux %[st1], %[Temp1](%[cm]) \n\t" /* odd 8 */ + + "sb %[st2], 11(%[dst]) \n\t" /* odd 6 */ + "sb %[st3], 13(%[dst]) \n\t" /* odd 7 */ + "sb %[st1], 15(%[dst]) \n\t" /* odd 8 */ + + : [qload1] "=&r"(qload1), [qload2] "=&r"(qload2), + [qload3] "=&r"(qload3), [st1] "=&r"(st1), [st2] "=&r"(st2), + [st3] "=&r"(st3), [p1] "=&r"(p1), [p2] "=&r"(p2), [p3] "=&r"(p3), + [p4] "=&r"(p4), [p5] "=&r"(p5), [Temp1] "=&r"(Temp1), + [Temp2] "=&r"(Temp2), [Temp3] "=&r"(Temp3) + : [filter12] "r"(filter12), [filter34] "r"(filter34), + [filter56] "r"(filter56), [filter78] "r"(filter78), + [vector_64] "r"(vector_64), [cm] "r"(cm), [dst] "r"(dst), + [src] "r"(src)); + + src += 16; + dst += 16; + } + + /* Next row... */ + src_ptr += src_stride; + dst_ptr += dst_stride; + } +} + +void aom_convolve8_horiz_dspr2(const uint8_t *src, ptrdiff_t src_stride, + uint8_t *dst, ptrdiff_t dst_stride, + const int16_t *filter_x, int x_step_q4, + const int16_t *filter_y, int y_step_q4, int w, + int h) { + assert(x_step_q4 == 16); + assert(((const int32_t *)filter_x)[1] != 0x800000); + + if (((const int32_t *)filter_x)[0] == 0) { + aom_convolve2_horiz_dspr2(src, src_stride, dst, dst_stride, filter_x, + x_step_q4, filter_y, y_step_q4, w, h); + } else { + uint32_t pos = 38; + + prefetch_load((const uint8_t *)filter_x); + src -= 3; + + /* bit positon for extract from acc */ + __asm__ __volatile__("wrdsp %[pos], 1 \n\t" + : + : [pos] "r"(pos)); + + /* prefetch data to cache memory */ + prefetch_load(src); + prefetch_load(src + 32); + prefetch_store(dst); + + switch (w) { + case 4: + convolve_horiz_4_dspr2(src, (int32_t)src_stride, dst, + (int32_t)dst_stride, filter_x, (int32_t)h); + break; + case 8: + convolve_horiz_8_dspr2(src, (int32_t)src_stride, dst, + (int32_t)dst_stride, filter_x, (int32_t)h); + break; + case 16: + convolve_horiz_16_dspr2(src, (int32_t)src_stride, dst, + (int32_t)dst_stride, filter_x, (int32_t)h, 1); + break; + case 32: + convolve_horiz_16_dspr2(src, (int32_t)src_stride, dst, + (int32_t)dst_stride, filter_x, (int32_t)h, 2); + break; + case 64: + prefetch_load(src + 64); + prefetch_store(dst + 32); + + convolve_horiz_64_dspr2(src, (int32_t)src_stride, dst, + (int32_t)dst_stride, filter_x, (int32_t)h); + break; + default: + aom_convolve8_horiz_c(src + 3, src_stride, dst, dst_stride, filter_x, + x_step_q4, filter_y, y_step_q4, w, h); + break; + } + } +} +#endif diff --git a/third_party/aom/aom_dsp/mips/convolve8_vert_dspr2.c b/third_party/aom/aom_dsp/mips/convolve8_vert_dspr2.c new file mode 100644 index 0000000000..201e664279 --- /dev/null +++ b/third_party/aom/aom_dsp/mips/convolve8_vert_dspr2.c @@ -0,0 +1,361 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include +#include + +#include "config/aom_dsp_rtcd.h" + +#include "aom_dsp/mips/convolve_common_dspr2.h" +#include "aom_dsp/aom_dsp_common.h" +#include "aom_dsp/aom_filter.h" +#include "aom_ports/mem.h" + +#if HAVE_DSPR2 +static void convolve_vert_4_dspr2(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + const int16_t *filter_y, int32_t w, + int32_t h) { + int32_t x, y; + const uint8_t *src_ptr; + uint8_t *dst_ptr; + uint8_t *cm = aom_ff_cropTbl; + uint32_t vector4a = 64; + uint32_t load1, load2, load3, load4; + uint32_t p1, p2; + uint32_t n1, n2; + uint32_t scratch1, scratch2; + uint32_t store1, store2; + int32_t vector1b, vector2b, vector3b, vector4b; + int32_t Temp1, Temp2; + + vector1b = ((const int32_t *)filter_y)[0]; + vector2b = ((const int32_t *)filter_y)[1]; + vector3b = ((const int32_t *)filter_y)[2]; + vector4b = ((const int32_t *)filter_y)[3]; + + src -= 3 * src_stride; + + for (y = h; y--;) { + /* prefetch data to cache memory */ + prefetch_store(dst + dst_stride); + + for (x = 0; x < w; x += 4) { + src_ptr = src + x; + dst_ptr = dst + x; + + __asm__ __volatile__( + "ulw %[load1], 0(%[src_ptr]) \n\t" + "add %[src_ptr], %[src_ptr], %[src_stride] \n\t" + "ulw %[load2], 0(%[src_ptr]) \n\t" + "add %[src_ptr], %[src_ptr], %[src_stride] \n\t" + "ulw %[load3], 0(%[src_ptr]) \n\t" + "add %[src_ptr], %[src_ptr], %[src_stride] \n\t" + "ulw %[load4], 0(%[src_ptr]) \n\t" + + "mtlo %[vector4a], $ac0 \n\t" + "mtlo %[vector4a], $ac1 \n\t" + "mtlo %[vector4a], $ac2 \n\t" + "mtlo %[vector4a], $ac3 \n\t" + "mthi $zero, $ac0 \n\t" + "mthi $zero, $ac1 \n\t" + "mthi $zero, $ac2 \n\t" + "mthi $zero, $ac3 \n\t" + + "preceu.ph.qbr %[scratch1], %[load1] \n\t" + "preceu.ph.qbr %[p1], %[load2] \n\t" + "precrq.ph.w %[n1], %[p1], %[scratch1] \n\t" /* pixel 2 */ + "append %[p1], %[scratch1], 16 \n\t" /* pixel 1 */ + "preceu.ph.qbr %[scratch2], %[load3] \n\t" + "preceu.ph.qbr %[p2], %[load4] \n\t" + "precrq.ph.w %[n2], %[p2], %[scratch2] \n\t" /* pixel 2 */ + "append %[p2], %[scratch2], 16 \n\t" /* pixel 1 */ + + "dpa.w.ph $ac0, %[p1], %[vector1b] \n\t" + "dpa.w.ph $ac0, %[p2], %[vector2b] \n\t" + "dpa.w.ph $ac1, %[n1], %[vector1b] \n\t" + "dpa.w.ph $ac1, %[n2], %[vector2b] \n\t" + + "preceu.ph.qbl %[scratch1], %[load1] \n\t" + "preceu.ph.qbl %[p1], %[load2] \n\t" + "precrq.ph.w %[n1], %[p1], %[scratch1] \n\t" /* pixel 2 */ + "append %[p1], %[scratch1], 16 \n\t" /* pixel 1 */ + "preceu.ph.qbl %[scratch2], %[load3] \n\t" + "preceu.ph.qbl %[p2], %[load4] \n\t" + "precrq.ph.w %[n2], %[p2], %[scratch2] \n\t" /* pixel 2 */ + "append %[p2], %[scratch2], 16 \n\t" /* pixel 1 */ + + "dpa.w.ph $ac2, %[p1], %[vector1b] \n\t" + "dpa.w.ph $ac2, %[p2], %[vector2b] \n\t" + "dpa.w.ph $ac3, %[n1], %[vector1b] \n\t" + "dpa.w.ph $ac3, %[n2], %[vector2b] \n\t" + + "add %[src_ptr], %[src_ptr], %[src_stride] \n\t" + "ulw %[load1], 0(%[src_ptr]) \n\t" + "add %[src_ptr], %[src_ptr], %[src_stride] \n\t" + "ulw %[load2], 0(%[src_ptr]) \n\t" + "add %[src_ptr], %[src_ptr], %[src_stride] \n\t" + "ulw %[load3], 0(%[src_ptr]) \n\t" + "add %[src_ptr], %[src_ptr], %[src_stride] \n\t" + "ulw %[load4], 0(%[src_ptr]) \n\t" + + "preceu.ph.qbr %[scratch1], %[load1] \n\t" + "preceu.ph.qbr %[p1], %[load2] \n\t" + "precrq.ph.w %[n1], %[p1], %[scratch1] \n\t" /* pixel 2 */ + "append %[p1], %[scratch1], 16 \n\t" /* pixel 1 */ + "preceu.ph.qbr %[scratch2], %[load3] \n\t" + "preceu.ph.qbr %[p2], %[load4] \n\t" + "precrq.ph.w %[n2], %[p2], %[scratch2] \n\t" /* pixel 2 */ + "append %[p2], %[scratch2], 16 \n\t" /* pixel 1 */ + + "dpa.w.ph $ac0, %[p1], %[vector3b] \n\t" + "dpa.w.ph $ac0, %[p2], %[vector4b] \n\t" + "extp %[Temp1], $ac0, 31 \n\t" + "dpa.w.ph $ac1, %[n1], %[vector3b] \n\t" + "dpa.w.ph $ac1, %[n2], %[vector4b] \n\t" + "extp %[Temp2], $ac1, 31 \n\t" + + "preceu.ph.qbl %[scratch1], %[load1] \n\t" + "preceu.ph.qbl %[p1], %[load2] \n\t" + "precrq.ph.w %[n1], %[p1], %[scratch1] \n\t" /* pixel 2 */ + "append %[p1], %[scratch1], 16 \n\t" /* pixel 1 */ + "preceu.ph.qbl %[scratch2], %[load3] \n\t" + "preceu.ph.qbl %[p2], %[load4] \n\t" + "precrq.ph.w %[n2], %[p2], %[scratch2] \n\t" /* pixel 2 */ + "append %[p2], %[scratch2], 16 \n\t" /* pixel 1 */ + + "lbux %[store1], %[Temp1](%[cm]) \n\t" + "dpa.w.ph $ac2, %[p1], %[vector3b] \n\t" + "dpa.w.ph $ac2, %[p2], %[vector4b] \n\t" + "extp %[Temp1], $ac2, 31 \n\t" + + "lbux %[store2], %[Temp2](%[cm]) \n\t" + "dpa.w.ph $ac3, %[n1], %[vector3b] \n\t" + "dpa.w.ph $ac3, %[n2], %[vector4b] \n\t" + "extp %[Temp2], $ac3, 31 \n\t" + + "sb %[store1], 0(%[dst_ptr]) \n\t" + "sb %[store2], 1(%[dst_ptr]) \n\t" + + "lbux %[store1], %[Temp1](%[cm]) \n\t" + "lbux %[store2], %[Temp2](%[cm]) \n\t" + + "sb %[store1], 2(%[dst_ptr]) \n\t" + "sb %[store2], 3(%[dst_ptr]) \n\t" + + : [load1] "=&r"(load1), [load2] "=&r"(load2), [load3] "=&r"(load3), + [load4] "=&r"(load4), [p1] "=&r"(p1), [p2] "=&r"(p2), + [n1] "=&r"(n1), [n2] "=&r"(n2), [scratch1] "=&r"(scratch1), + [scratch2] "=&r"(scratch2), [Temp1] "=&r"(Temp1), + [Temp2] "=&r"(Temp2), [store1] "=&r"(store1), + [store2] "=&r"(store2), [src_ptr] "+r"(src_ptr) + : [vector1b] "r"(vector1b), [vector2b] "r"(vector2b), + [vector3b] "r"(vector3b), [vector4b] "r"(vector4b), + [vector4a] "r"(vector4a), [src_stride] "r"(src_stride), + [cm] "r"(cm), [dst_ptr] "r"(dst_ptr)); + } + + /* Next row... */ + src += src_stride; + dst += dst_stride; + } +} + +static void convolve_vert_64_dspr2(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + const int16_t *filter_y, int32_t h) { + int32_t x, y; + const uint8_t *src_ptr; + uint8_t *dst_ptr; + uint8_t *cm = aom_ff_cropTbl; + uint32_t vector4a = 64; + uint32_t load1, load2, load3, load4; + uint32_t p1, p2; + uint32_t n1, n2; + uint32_t scratch1, scratch2; + uint32_t store1, store2; + int32_t vector1b, vector2b, vector3b, vector4b; + int32_t Temp1, Temp2; + + vector1b = ((const int32_t *)filter_y)[0]; + vector2b = ((const int32_t *)filter_y)[1]; + vector3b = ((const int32_t *)filter_y)[2]; + vector4b = ((const int32_t *)filter_y)[3]; + + src -= 3 * src_stride; + + for (y = h; y--;) { + /* prefetch data to cache memory */ + prefetch_store(dst + dst_stride); + prefetch_store(dst + dst_stride + 32); + + for (x = 0; x < 64; x += 4) { + src_ptr = src + x; + dst_ptr = dst + x; + + __asm__ __volatile__( + "ulw %[load1], 0(%[src_ptr]) \n\t" + "add %[src_ptr], %[src_ptr], %[src_stride] \n\t" + "ulw %[load2], 0(%[src_ptr]) \n\t" + "add %[src_ptr], %[src_ptr], %[src_stride] \n\t" + "ulw %[load3], 0(%[src_ptr]) \n\t" + "add %[src_ptr], %[src_ptr], %[src_stride] \n\t" + "ulw %[load4], 0(%[src_ptr]) \n\t" + + "mtlo %[vector4a], $ac0 \n\t" + "mtlo %[vector4a], $ac1 \n\t" + "mtlo %[vector4a], $ac2 \n\t" + "mtlo %[vector4a], $ac3 \n\t" + "mthi $zero, $ac0 \n\t" + "mthi $zero, $ac1 \n\t" + "mthi $zero, $ac2 \n\t" + "mthi $zero, $ac3 \n\t" + + "preceu.ph.qbr %[scratch1], %[load1] \n\t" + "preceu.ph.qbr %[p1], %[load2] \n\t" + "precrq.ph.w %[n1], %[p1], %[scratch1] \n\t" /* pixel 2 */ + "append %[p1], %[scratch1], 16 \n\t" /* pixel 1 */ + "preceu.ph.qbr %[scratch2], %[load3] \n\t" + "preceu.ph.qbr %[p2], %[load4] \n\t" + "precrq.ph.w %[n2], %[p2], %[scratch2] \n\t" /* pixel 2 */ + "append %[p2], %[scratch2], 16 \n\t" /* pixel 1 */ + + "dpa.w.ph $ac0, %[p1], %[vector1b] \n\t" + "dpa.w.ph $ac0, %[p2], %[vector2b] \n\t" + "dpa.w.ph $ac1, %[n1], %[vector1b] \n\t" + "dpa.w.ph $ac1, %[n2], %[vector2b] \n\t" + + "preceu.ph.qbl %[scratch1], %[load1] \n\t" + "preceu.ph.qbl %[p1], %[load2] \n\t" + "precrq.ph.w %[n1], %[p1], %[scratch1] \n\t" /* pixel 2 */ + "append %[p1], %[scratch1], 16 \n\t" /* pixel 1 */ + "preceu.ph.qbl %[scratch2], %[load3] \n\t" + "preceu.ph.qbl %[p2], %[load4] \n\t" + "precrq.ph.w %[n2], %[p2], %[scratch2] \n\t" /* pixel 2 */ + "append %[p2], %[scratch2], 16 \n\t" /* pixel 1 */ + + "dpa.w.ph $ac2, %[p1], %[vector1b] \n\t" + "dpa.w.ph $ac2, %[p2], %[vector2b] \n\t" + "dpa.w.ph $ac3, %[n1], %[vector1b] \n\t" + "dpa.w.ph $ac3, %[n2], %[vector2b] \n\t" + + "add %[src_ptr], %[src_ptr], %[src_stride] \n\t" + "ulw %[load1], 0(%[src_ptr]) \n\t" + "add %[src_ptr], %[src_ptr], %[src_stride] \n\t" + "ulw %[load2], 0(%[src_ptr]) \n\t" + "add %[src_ptr], %[src_ptr], %[src_stride] \n\t" + "ulw %[load3], 0(%[src_ptr]) \n\t" + "add %[src_ptr], %[src_ptr], %[src_stride] \n\t" + "ulw %[load4], 0(%[src_ptr]) \n\t" + + "preceu.ph.qbr %[scratch1], %[load1] \n\t" + "preceu.ph.qbr %[p1], %[load2] \n\t" + "precrq.ph.w %[n1], %[p1], %[scratch1] \n\t" /* pixel 2 */ + "append %[p1], %[scratch1], 16 \n\t" /* pixel 1 */ + "preceu.ph.qbr %[scratch2], %[load3] \n\t" + "preceu.ph.qbr %[p2], %[load4] \n\t" + "precrq.ph.w %[n2], %[p2], %[scratch2] \n\t" /* pixel 2 */ + "append %[p2], %[scratch2], 16 \n\t" /* pixel 1 */ + + "dpa.w.ph $ac0, %[p1], %[vector3b] \n\t" + "dpa.w.ph $ac0, %[p2], %[vector4b] \n\t" + "extp %[Temp1], $ac0, 31 \n\t" + "dpa.w.ph $ac1, %[n1], %[vector3b] \n\t" + "dpa.w.ph $ac1, %[n2], %[vector4b] \n\t" + "extp %[Temp2], $ac1, 31 \n\t" + + "preceu.ph.qbl %[scratch1], %[load1] \n\t" + "preceu.ph.qbl %[p1], %[load2] \n\t" + "precrq.ph.w %[n1], %[p1], %[scratch1] \n\t" /* pixel 2 */ + "append %[p1], %[scratch1], 16 \n\t" /* pixel 1 */ + "preceu.ph.qbl %[scratch2], %[load3] \n\t" + "preceu.ph.qbl %[p2], %[load4] \n\t" + "precrq.ph.w %[n2], %[p2], %[scratch2] \n\t" /* pixel 2 */ + "append %[p2], %[scratch2], 16 \n\t" /* pixel 1 */ + + "lbux %[store1], %[Temp1](%[cm]) \n\t" + "dpa.w.ph $ac2, %[p1], %[vector3b] \n\t" + "dpa.w.ph $ac2, %[p2], %[vector4b] \n\t" + "extp %[Temp1], $ac2, 31 \n\t" + + "lbux %[store2], %[Temp2](%[cm]) \n\t" + "dpa.w.ph $ac3, %[n1], %[vector3b] \n\t" + "dpa.w.ph $ac3, %[n2], %[vector4b] \n\t" + "extp %[Temp2], $ac3, 31 \n\t" + + "sb %[store1], 0(%[dst_ptr]) \n\t" + "sb %[store2], 1(%[dst_ptr]) \n\t" + + "lbux %[store1], %[Temp1](%[cm]) \n\t" + "lbux %[store2], %[Temp2](%[cm]) \n\t" + + "sb %[store1], 2(%[dst_ptr]) \n\t" + "sb %[store2], 3(%[dst_ptr]) \n\t" + + : [load1] "=&r"(load1), [load2] "=&r"(load2), [load3] "=&r"(load3), + [load4] "=&r"(load4), [p1] "=&r"(p1), [p2] "=&r"(p2), + [n1] "=&r"(n1), [n2] "=&r"(n2), [scratch1] "=&r"(scratch1), + [scratch2] "=&r"(scratch2), [Temp1] "=&r"(Temp1), + [Temp2] "=&r"(Temp2), [store1] "=&r"(store1), + [store2] "=&r"(store2), [src_ptr] "+r"(src_ptr) + : [vector1b] "r"(vector1b), [vector2b] "r"(vector2b), + [vector3b] "r"(vector3b), [vector4b] "r"(vector4b), + [vector4a] "r"(vector4a), [src_stride] "r"(src_stride), + [cm] "r"(cm), [dst_ptr] "r"(dst_ptr)); + } + + /* Next row... */ + src += src_stride; + dst += dst_stride; + } +} + +void aom_convolve8_vert_dspr2(const uint8_t *src, ptrdiff_t src_stride, + uint8_t *dst, ptrdiff_t dst_stride, + const int16_t *filter_x, int x_step_q4, + const int16_t *filter_y, int y_step_q4, int w, + int h) { + assert(y_step_q4 == 16); + assert(((const int32_t *)filter_y)[1] != 0x800000); + + if (((const int32_t *)filter_y)[0] == 0) { + aom_convolve2_vert_dspr2(src, src_stride, dst, dst_stride, filter_x, + x_step_q4, filter_y, y_step_q4, w, h); + } else { + uint32_t pos = 38; + + /* bit positon for extract from acc */ + __asm__ __volatile__("wrdsp %[pos], 1 \n\t" + : + : [pos] "r"(pos)); + + prefetch_store(dst); + + switch (w) { + case 4: + case 8: + case 16: + case 32: + convolve_vert_4_dspr2(src, src_stride, dst, dst_stride, filter_y, w, h); + break; + case 64: + prefetch_store(dst + 32); + convolve_vert_64_dspr2(src, src_stride, dst, dst_stride, filter_y, h); + break; + default: + aom_convolve8_vert_c(src, src_stride, dst, dst_stride, filter_x, + x_step_q4, filter_y, y_step_q4, w, h); + break; + } + } +} + +#endif diff --git a/third_party/aom/aom_dsp/mips/convolve_common_dspr2.h b/third_party/aom/aom_dsp/mips/convolve_common_dspr2.h new file mode 100644 index 0000000000..e5d48a884d --- /dev/null +++ b/third_party/aom/aom_dsp/mips/convolve_common_dspr2.h @@ -0,0 +1,48 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_MIPS_CONVOLVE_COMMON_DSPR2_H_ +#define AOM_AOM_DSP_MIPS_CONVOLVE_COMMON_DSPR2_H_ + +#include + +#include "config/aom_config.h" + +#include "aom/aom_integer.h" +#include "aom_dsp/mips/common_dspr2.h" + +#ifdef __cplusplus +extern "C" { +#endif + +#if HAVE_DSPR2 +void aom_convolve2_horiz_dspr2(const uint8_t *src, ptrdiff_t src_stride, + uint8_t *dst, ptrdiff_t dst_stride, + const int16_t *filter_x, int x_step_q4, + const int16_t *filter_y, int y_step_q4, int w, + int h); + +void aom_convolve2_dspr2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, + ptrdiff_t dst_stride, const int16_t *filter, int w, + int h); + +void aom_convolve2_vert_dspr2(const uint8_t *src, ptrdiff_t src_stride, + uint8_t *dst, ptrdiff_t dst_stride, + const int16_t *filter_x, int x_step_q4, + const int16_t *filter_y, int y_step_q4, int w, + int h); + +#endif // #if HAVE_DSPR2 +#ifdef __cplusplus +} // extern "C" +#endif + +#endif // AOM_AOM_DSP_MIPS_CONVOLVE_COMMON_DSPR2_H_ diff --git a/third_party/aom/aom_dsp/mips/intrapred16_dspr2.c b/third_party/aom/aom_dsp/mips/intrapred16_dspr2.c new file mode 100644 index 0000000000..7c221ae899 --- /dev/null +++ b/third_party/aom/aom_dsp/mips/intrapred16_dspr2.c @@ -0,0 +1,327 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include "aom_dsp/mips/common_dspr2.h" + +#if HAVE_DSPR2 +void aom_h_predictor_16x16_dspr2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + int32_t tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7, tmp8; + int32_t tmp9, tmp10, tmp11, tmp12, tmp13, tmp14, tmp15, tmp16; + + (void)above; + + __asm__ __volatile__( + "lb %[tmp1], (%[left]) \n\t" + "lb %[tmp2], 1(%[left]) \n\t" + "lb %[tmp3], 2(%[left]) \n\t" + "lb %[tmp4], 3(%[left]) \n\t" + "lb %[tmp5], 4(%[left]) \n\t" + "lb %[tmp6], 5(%[left]) \n\t" + "lb %[tmp7], 6(%[left]) \n\t" + "lb %[tmp8], 7(%[left]) \n\t" + "lb %[tmp9], 8(%[left]) \n\t" + "lb %[tmp10], 9(%[left]) \n\t" + "lb %[tmp11], 10(%[left]) \n\t" + "lb %[tmp12], 11(%[left]) \n\t" + "lb %[tmp13], 12(%[left]) \n\t" + "lb %[tmp14], 13(%[left]) \n\t" + "lb %[tmp15], 14(%[left]) \n\t" + "lb %[tmp16], 15(%[left]) \n\t" + + "replv.qb %[tmp1], %[tmp1] \n\t" + "replv.qb %[tmp2], %[tmp2] \n\t" + "replv.qb %[tmp3], %[tmp3] \n\t" + "replv.qb %[tmp4], %[tmp4] \n\t" + "replv.qb %[tmp5], %[tmp5] \n\t" + "replv.qb %[tmp6], %[tmp6] \n\t" + "replv.qb %[tmp7], %[tmp7] \n\t" + "replv.qb %[tmp8], %[tmp8] \n\t" + "replv.qb %[tmp9], %[tmp9] \n\t" + "replv.qb %[tmp10], %[tmp10] \n\t" + "replv.qb %[tmp11], %[tmp11] \n\t" + "replv.qb %[tmp12], %[tmp12] \n\t" + "replv.qb %[tmp13], %[tmp13] \n\t" + "replv.qb %[tmp14], %[tmp14] \n\t" + "replv.qb %[tmp15], %[tmp15] \n\t" + "replv.qb %[tmp16], %[tmp16] \n\t" + + "sw %[tmp1], (%[dst]) \n\t" + "sw %[tmp1], 4(%[dst]) \n\t" + "sw %[tmp1], 8(%[dst]) \n\t" + "sw %[tmp1], 12(%[dst]) \n\t" + + "add %[dst], %[dst], %[stride] \n\t" + "sw %[tmp2], (%[dst]) \n\t" + "sw %[tmp2], 4(%[dst]) \n\t" + "sw %[tmp2], 8(%[dst]) \n\t" + "sw %[tmp2], 12(%[dst]) \n\t" + + "add %[dst], %[dst], %[stride] \n\t" + "sw %[tmp3], (%[dst]) \n\t" + "sw %[tmp3], 4(%[dst]) \n\t" + "sw %[tmp3], 8(%[dst]) \n\t" + "sw %[tmp3], 12(%[dst]) \n\t" + + "add %[dst], %[dst], %[stride] \n\t" + "sw %[tmp4], (%[dst]) \n\t" + "sw %[tmp4], 4(%[dst]) \n\t" + "sw %[tmp4], 8(%[dst]) \n\t" + "sw %[tmp4], 12(%[dst]) \n\t" + + "add %[dst], %[dst], %[stride] \n\t" + "sw %[tmp5], (%[dst]) \n\t" + "sw %[tmp5], 4(%[dst]) \n\t" + "sw %[tmp5], 8(%[dst]) \n\t" + "sw %[tmp5], 12(%[dst]) \n\t" + + "add %[dst], %[dst], %[stride] \n\t" + "sw %[tmp6], (%[dst]) \n\t" + "sw %[tmp6], 4(%[dst]) \n\t" + "sw %[tmp6], 8(%[dst]) \n\t" + "sw %[tmp6], 12(%[dst]) \n\t" + + "add %[dst], %[dst], %[stride] \n\t" + "sw %[tmp7], (%[dst]) \n\t" + "sw %[tmp7], 4(%[dst]) \n\t" + "sw %[tmp7], 8(%[dst]) \n\t" + "sw %[tmp7], 12(%[dst]) \n\t" + + "add %[dst], %[dst], %[stride] \n\t" + "sw %[tmp8], (%[dst]) \n\t" + "sw %[tmp8], 4(%[dst]) \n\t" + "sw %[tmp8], 8(%[dst]) \n\t" + "sw %[tmp8], 12(%[dst]) \n\t" + + "add %[dst], %[dst], %[stride] \n\t" + "sw %[tmp9], (%[dst]) \n\t" + "sw %[tmp9], 4(%[dst]) \n\t" + "sw %[tmp9], 8(%[dst]) \n\t" + "sw %[tmp9], 12(%[dst]) \n\t" + + "add %[dst], %[dst], %[stride] \n\t" + "sw %[tmp10], (%[dst]) \n\t" + "sw %[tmp10], 4(%[dst]) \n\t" + "sw %[tmp10], 8(%[dst]) \n\t" + "sw %[tmp10], 12(%[dst]) \n\t" + + "add %[dst], %[dst], %[stride] \n\t" + "sw %[tmp11], (%[dst]) \n\t" + "sw %[tmp11], 4(%[dst]) \n\t" + "sw %[tmp11], 8(%[dst]) \n\t" + "sw %[tmp11], 12(%[dst]) \n\t" + + "add %[dst], %[dst], %[stride] \n\t" + "sw %[tmp12], (%[dst]) \n\t" + "sw %[tmp12], 4(%[dst]) \n\t" + "sw %[tmp12], 8(%[dst]) \n\t" + "sw %[tmp12], 12(%[dst]) \n\t" + + "add %[dst], %[dst], %[stride] \n\t" + "sw %[tmp13], (%[dst]) \n\t" + "sw %[tmp13], 4(%[dst]) \n\t" + "sw %[tmp13], 8(%[dst]) \n\t" + "sw %[tmp13], 12(%[dst]) \n\t" + + "add %[dst], %[dst], %[stride] \n\t" + "sw %[tmp14], (%[dst]) \n\t" + "sw %[tmp14], 4(%[dst]) \n\t" + "sw %[tmp14], 8(%[dst]) \n\t" + "sw %[tmp14], 12(%[dst]) \n\t" + + "add %[dst], %[dst], %[stride] \n\t" + "sw %[tmp15], (%[dst]) \n\t" + "sw %[tmp15], 4(%[dst]) \n\t" + "sw %[tmp15], 8(%[dst]) \n\t" + "sw %[tmp15], 12(%[dst]) \n\t" + + "add %[dst], %[dst], %[stride] \n\t" + "sw %[tmp16], (%[dst]) \n\t" + "sw %[tmp16], 4(%[dst]) \n\t" + "sw %[tmp16], 8(%[dst]) \n\t" + "sw %[tmp16], 12(%[dst]) \n\t" + + : [tmp1] "=&r"(tmp1), [tmp2] "=&r"(tmp2), [tmp3] "=&r"(tmp3), + [tmp4] "=&r"(tmp4), [tmp5] "=&r"(tmp5), [tmp7] "=&r"(tmp7), + [tmp6] "=&r"(tmp6), [tmp8] "=&r"(tmp8), [tmp9] "=&r"(tmp9), + [tmp10] "=&r"(tmp10), [tmp11] "=&r"(tmp11), [tmp12] "=&r"(tmp12), + [tmp13] "=&r"(tmp13), [tmp14] "=&r"(tmp14), [tmp15] "=&r"(tmp15), + [tmp16] "=&r"(tmp16) + : [left] "r"(left), [dst] "r"(dst), [stride] "r"(stride)); +} + +void aom_dc_predictor_16x16_dspr2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + int32_t expected_dc; + int32_t average; + int32_t tmp, above1, above_l1, above_r1, left1, left_r1, left_l1; + int32_t above2, left2; + + __asm__ __volatile__( + "lw %[above1], (%[above]) \n\t" + "lw %[above2], 4(%[above]) \n\t" + "lw %[left1], (%[left]) \n\t" + "lw %[left2], 4(%[left]) \n\t" + + "preceu.ph.qbl %[above_l1], %[above1] \n\t" + "preceu.ph.qbr %[above_r1], %[above1] \n\t" + "preceu.ph.qbl %[left_l1], %[left1] \n\t" + "preceu.ph.qbr %[left_r1], %[left1] \n\t" + + "addu.ph %[average], %[above_r1], %[above_l1] \n\t" + "addu.ph %[average], %[average], %[left_l1] \n\t" + "addu.ph %[average], %[average], %[left_r1] \n\t" + + "preceu.ph.qbl %[above_l1], %[above2] \n\t" + "preceu.ph.qbr %[above_r1], %[above2] \n\t" + "preceu.ph.qbl %[left_l1], %[left2] \n\t" + "preceu.ph.qbr %[left_r1], %[left2] \n\t" + + "addu.ph %[average], %[average], %[above_l1] \n\t" + "addu.ph %[average], %[average], %[above_r1] \n\t" + "addu.ph %[average], %[average], %[left_l1] \n\t" + "addu.ph %[average], %[average], %[left_r1] \n\t" + + "lw %[above1], 8(%[above]) \n\t" + "lw %[above2], 12(%[above]) \n\t" + "lw %[left1], 8(%[left]) \n\t" + "lw %[left2], 12(%[left]) \n\t" + + "preceu.ph.qbl %[above_l1], %[above1] \n\t" + "preceu.ph.qbr %[above_r1], %[above1] \n\t" + "preceu.ph.qbl %[left_l1], %[left1] \n\t" + "preceu.ph.qbr %[left_r1], %[left1] \n\t" + + "addu.ph %[average], %[average], %[above_l1] \n\t" + "addu.ph %[average], %[average], %[above_r1] \n\t" + "addu.ph %[average], %[average], %[left_l1] \n\t" + "addu.ph %[average], %[average], %[left_r1] \n\t" + + "preceu.ph.qbl %[above_l1], %[above2] \n\t" + "preceu.ph.qbr %[above_r1], %[above2] \n\t" + "preceu.ph.qbl %[left_l1], %[left2] \n\t" + "preceu.ph.qbr %[left_r1], %[left2] \n\t" + + "addu.ph %[average], %[average], %[above_l1] \n\t" + "addu.ph %[average], %[average], %[above_r1] \n\t" + "addu.ph %[average], %[average], %[left_l1] \n\t" + "addu.ph %[average], %[average], %[left_r1] \n\t" + + "addiu %[average], %[average], 16 \n\t" + "srl %[tmp], %[average], 16 \n\t" + "addu.ph %[average], %[tmp], %[average] \n\t" + "srl %[expected_dc], %[average], 5 \n\t" + "replv.qb %[expected_dc], %[expected_dc] \n\t" + + "sw %[expected_dc], (%[dst]) \n\t" + "sw %[expected_dc], 4(%[dst]) \n\t" + "sw %[expected_dc], 8(%[dst]) \n\t" + "sw %[expected_dc], 12(%[dst]) \n\t" + + "add %[dst], %[dst], %[stride] \n\t" + "sw %[expected_dc], (%[dst]) \n\t" + "sw %[expected_dc], 4(%[dst]) \n\t" + "sw %[expected_dc], 8(%[dst]) \n\t" + "sw %[expected_dc], 12(%[dst]) \n\t" + + "add %[dst], %[dst], %[stride] \n\t" + "sw %[expected_dc], (%[dst]) \n\t" + "sw %[expected_dc], 4(%[dst]) \n\t" + "sw %[expected_dc], 8(%[dst]) \n\t" + "sw %[expected_dc], 12(%[dst]) \n\t" + + "add %[dst], %[dst], %[stride] \n\t" + "sw %[expected_dc], (%[dst]) \n\t" + "sw %[expected_dc], 4(%[dst]) \n\t" + "sw %[expected_dc], 8(%[dst]) \n\t" + "sw %[expected_dc], 12(%[dst]) \n\t" + + "add %[dst], %[dst], %[stride] \n\t" + "sw %[expected_dc], (%[dst]) \n\t" + "sw %[expected_dc], 4(%[dst]) \n\t" + "sw %[expected_dc], 8(%[dst]) \n\t" + "sw %[expected_dc], 12(%[dst]) \n\t" + + "add %[dst], %[dst], %[stride] \n\t" + "sw %[expected_dc], (%[dst]) \n\t" + "sw %[expected_dc], 4(%[dst]) \n\t" + "sw %[expected_dc], 8(%[dst]) \n\t" + "sw %[expected_dc], 12(%[dst]) \n\t" + + "add %[dst], %[dst], %[stride] \n\t" + "sw %[expected_dc], (%[dst]) \n\t" + "sw %[expected_dc], 4(%[dst]) \n\t" + "sw %[expected_dc], 8(%[dst]) \n\t" + "sw %[expected_dc], 12(%[dst]) \n\t" + + "add %[dst], %[dst], %[stride] \n\t" + "sw %[expected_dc], (%[dst]) \n\t" + "sw %[expected_dc], 4(%[dst]) \n\t" + "sw %[expected_dc], 8(%[dst]) \n\t" + "sw %[expected_dc], 12(%[dst]) \n\t" + + "add %[dst], %[dst], %[stride] \n\t" + "sw %[expected_dc], (%[dst]) \n\t" + "sw %[expected_dc], 4(%[dst]) \n\t" + "sw %[expected_dc], 8(%[dst]) \n\t" + "sw %[expected_dc], 12(%[dst]) \n\t" + + "add %[dst], %[dst], %[stride] \n\t" + "sw %[expected_dc], (%[dst]) \n\t" + "sw %[expected_dc], 4(%[dst]) \n\t" + "sw %[expected_dc], 8(%[dst]) \n\t" + "sw %[expected_dc], 12(%[dst]) \n\t" + + "add %[dst], %[dst], %[stride] \n\t" + "sw %[expected_dc], (%[dst]) \n\t" + "sw %[expected_dc], 4(%[dst]) \n\t" + "sw %[expected_dc], 8(%[dst]) \n\t" + "sw %[expected_dc], 12(%[dst]) \n\t" + + "add %[dst], %[dst], %[stride] \n\t" + "sw %[expected_dc], (%[dst]) \n\t" + "sw %[expected_dc], 4(%[dst]) \n\t" + "sw %[expected_dc], 8(%[dst]) \n\t" + "sw %[expected_dc], 12(%[dst]) \n\t" + + "add %[dst], %[dst], %[stride] \n\t" + "sw %[expected_dc], (%[dst]) \n\t" + "sw %[expected_dc], 4(%[dst]) \n\t" + "sw %[expected_dc], 8(%[dst]) \n\t" + "sw %[expected_dc], 12(%[dst]) \n\t" + + "add %[dst], %[dst], %[stride] \n\t" + "sw %[expected_dc], (%[dst]) \n\t" + "sw %[expected_dc], 4(%[dst]) \n\t" + "sw %[expected_dc], 8(%[dst]) \n\t" + "sw %[expected_dc], 12(%[dst]) \n\t" + + "add %[dst], %[dst], %[stride] \n\t" + "sw %[expected_dc], (%[dst]) \n\t" + "sw %[expected_dc], 4(%[dst]) \n\t" + "sw %[expected_dc], 8(%[dst]) \n\t" + "sw %[expected_dc], 12(%[dst]) \n\t" + + "add %[dst], %[dst], %[stride] \n\t" + "sw %[expected_dc], (%[dst]) \n\t" + "sw %[expected_dc], 4(%[dst]) \n\t" + "sw %[expected_dc], 8(%[dst]) \n\t" + "sw %[expected_dc], 12(%[dst]) \n\t" + + : [left1] "=&r"(left1), [above1] "=&r"(above1), [left_l1] "=&r"(left_l1), + [above_l1] "=&r"(above_l1), [left_r1] "=&r"(left_r1), + [above_r1] "=&r"(above_r1), [above2] "=&r"(above2), + [left2] "=&r"(left2), [average] "=&r"(average), [tmp] "=&r"(tmp), + [expected_dc] "=&r"(expected_dc) + : [above] "r"(above), [left] "r"(left), [dst] "r"(dst), + [stride] "r"(stride)); +} +#endif // #if HAVE_DSPR2 diff --git a/third_party/aom/aom_dsp/mips/intrapred4_dspr2.c b/third_party/aom/aom_dsp/mips/intrapred4_dspr2.c new file mode 100644 index 0000000000..0a21979c74 --- /dev/null +++ b/third_party/aom/aom_dsp/mips/intrapred4_dspr2.c @@ -0,0 +1,82 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include "aom_dsp/mips/common_dspr2.h" + +#if HAVE_DSPR2 +void aom_h_predictor_4x4_dspr2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + int32_t tmp1, tmp2, tmp3, tmp4; + (void)above; + + __asm__ __volatile__( + "lb %[tmp1], (%[left]) \n\t" + "lb %[tmp2], 1(%[left]) \n\t" + "lb %[tmp3], 2(%[left]) \n\t" + "lb %[tmp4], 3(%[left]) \n\t" + "replv.qb %[tmp1], %[tmp1] \n\t" + "replv.qb %[tmp2], %[tmp2] \n\t" + "replv.qb %[tmp3], %[tmp3] \n\t" + "replv.qb %[tmp4], %[tmp4] \n\t" + "sw %[tmp1], (%[dst]) \n\t" + "add %[dst], %[dst], %[stride] \n\t" + "sw %[tmp2], (%[dst]) \n\t" + "add %[dst], %[dst], %[stride] \n\t" + "sw %[tmp3], (%[dst]) \n\t" + "add %[dst], %[dst], %[stride] \n\t" + "sw %[tmp4], (%[dst]) \n\t" + + : [tmp1] "=&r"(tmp1), [tmp2] "=&r"(tmp2), [tmp3] "=&r"(tmp3), + [tmp4] "=&r"(tmp4) + : [left] "r"(left), [dst] "r"(dst), [stride] "r"(stride)); +} + +void aom_dc_predictor_4x4_dspr2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + int32_t expected_dc; + int32_t average; + int32_t tmp, above_c, above_l, above_r, left_c, left_r, left_l; + + __asm__ __volatile__( + "lw %[above_c], (%[above]) \n\t" + "lw %[left_c], (%[left]) \n\t" + + "preceu.ph.qbl %[above_l], %[above_c] \n\t" + "preceu.ph.qbr %[above_r], %[above_c] \n\t" + "preceu.ph.qbl %[left_l], %[left_c] \n\t" + "preceu.ph.qbr %[left_r], %[left_c] \n\t" + + "addu.ph %[average], %[above_r], %[above_l] \n\t" + "addu.ph %[average], %[average], %[left_l] \n\t" + "addu.ph %[average], %[average], %[left_r] \n\t" + "addiu %[average], %[average], 4 \n\t" + "srl %[tmp], %[average], 16 \n\t" + "addu.ph %[average], %[tmp], %[average] \n\t" + "srl %[expected_dc], %[average], 3 \n\t" + "replv.qb %[expected_dc], %[expected_dc] \n\t" + + "sw %[expected_dc], (%[dst]) \n\t" + "add %[dst], %[dst], %[stride] \n\t" + "sw %[expected_dc], (%[dst]) \n\t" + "add %[dst], %[dst], %[stride] \n\t" + "sw %[expected_dc], (%[dst]) \n\t" + "add %[dst], %[dst], %[stride] \n\t" + "sw %[expected_dc], (%[dst]) \n\t" + + : [above_c] "=&r"(above_c), [above_l] "=&r"(above_l), + [above_r] "=&r"(above_r), [left_c] "=&r"(left_c), + [left_l] "=&r"(left_l), [left_r] "=&r"(left_r), + [average] "=&r"(average), [tmp] "=&r"(tmp), + [expected_dc] "=&r"(expected_dc) + : [above] "r"(above), [left] "r"(left), [dst] "r"(dst), + [stride] "r"(stride)); +} +#endif // #if HAVE_DSPR2 diff --git a/third_party/aom/aom_dsp/mips/intrapred8_dspr2.c b/third_party/aom/aom_dsp/mips/intrapred8_dspr2.c new file mode 100644 index 0000000000..d42a77c802 --- /dev/null +++ b/third_party/aom/aom_dsp/mips/intrapred8_dspr2.c @@ -0,0 +1,150 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include "aom_dsp/mips/common_dspr2.h" + +#if HAVE_DSPR2 +void aom_h_predictor_8x8_dspr2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + int32_t tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7, tmp8; + (void)above; + + __asm__ __volatile__( + "lb %[tmp1], (%[left]) \n\t" + "lb %[tmp2], 1(%[left]) \n\t" + "lb %[tmp3], 2(%[left]) \n\t" + "lb %[tmp4], 3(%[left]) \n\t" + "lb %[tmp5], 4(%[left]) \n\t" + "lb %[tmp6], 5(%[left]) \n\t" + "lb %[tmp7], 6(%[left]) \n\t" + "lb %[tmp8], 7(%[left]) \n\t" + + "replv.qb %[tmp1], %[tmp1] \n\t" + "replv.qb %[tmp2], %[tmp2] \n\t" + "replv.qb %[tmp3], %[tmp3] \n\t" + "replv.qb %[tmp4], %[tmp4] \n\t" + "replv.qb %[tmp5], %[tmp5] \n\t" + "replv.qb %[tmp6], %[tmp6] \n\t" + "replv.qb %[tmp7], %[tmp7] \n\t" + "replv.qb %[tmp8], %[tmp8] \n\t" + + "sw %[tmp1], (%[dst]) \n\t" + "sw %[tmp1], 4(%[dst]) \n\t" + "add %[dst], %[dst], %[stride] \n\t" + "sw %[tmp2], (%[dst]) \n\t" + "sw %[tmp2], 4(%[dst]) \n\t" + "add %[dst], %[dst], %[stride] \n\t" + "sw %[tmp3], (%[dst]) \n\t" + "sw %[tmp3], 4(%[dst]) \n\t" + "add %[dst], %[dst], %[stride] \n\t" + "sw %[tmp4], (%[dst]) \n\t" + "sw %[tmp4], 4(%[dst]) \n\t" + "add %[dst], %[dst], %[stride] \n\t" + "sw %[tmp5], (%[dst]) \n\t" + "sw %[tmp5], 4(%[dst]) \n\t" + "add %[dst], %[dst], %[stride] \n\t" + "sw %[tmp6], (%[dst]) \n\t" + "sw %[tmp6], 4(%[dst]) \n\t" + "add %[dst], %[dst], %[stride] \n\t" + "sw %[tmp7], (%[dst]) \n\t" + "sw %[tmp7], 4(%[dst]) \n\t" + "add %[dst], %[dst], %[stride] \n\t" + "sw %[tmp8], (%[dst]) \n\t" + "sw %[tmp8], 4(%[dst]) \n\t" + + : [tmp1] "=&r"(tmp1), [tmp2] "=&r"(tmp2), [tmp3] "=&r"(tmp3), + [tmp4] "=&r"(tmp4), [tmp5] "=&r"(tmp5), [tmp7] "=&r"(tmp7), + [tmp6] "=&r"(tmp6), [tmp8] "=&r"(tmp8) + : [left] "r"(left), [dst] "r"(dst), [stride] "r"(stride)); +} + +void aom_dc_predictor_8x8_dspr2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + int32_t expected_dc; + int32_t average; + int32_t tmp, above1, above_l1, above_r1, left1, left_r1, left_l1; + int32_t above2, above_l2, above_r2, left2, left_r2, left_l2; + + __asm__ __volatile__( + "lw %[above1], (%[above]) \n\t" + "lw %[above2], 4(%[above]) \n\t" + "lw %[left1], (%[left]) \n\t" + "lw %[left2], 4(%[left]) \n\t" + + "preceu.ph.qbl %[above_l1], %[above1] \n\t" + "preceu.ph.qbr %[above_r1], %[above1] \n\t" + "preceu.ph.qbl %[left_l1], %[left1] \n\t" + "preceu.ph.qbr %[left_r1], %[left1] \n\t" + + "preceu.ph.qbl %[above_l2], %[above2] \n\t" + "preceu.ph.qbr %[above_r2], %[above2] \n\t" + "preceu.ph.qbl %[left_l2], %[left2] \n\t" + "preceu.ph.qbr %[left_r2], %[left2] \n\t" + + "addu.ph %[average], %[above_r1], %[above_l1] \n\t" + "addu.ph %[average], %[average], %[left_l1] \n\t" + "addu.ph %[average], %[average], %[left_r1] \n\t" + + "addu.ph %[average], %[average], %[above_l2] \n\t" + "addu.ph %[average], %[average], %[above_r2] \n\t" + "addu.ph %[average], %[average], %[left_l2] \n\t" + "addu.ph %[average], %[average], %[left_r2] \n\t" + + "addiu %[average], %[average], 8 \n\t" + + "srl %[tmp], %[average], 16 \n\t" + "addu.ph %[average], %[tmp], %[average] \n\t" + "srl %[expected_dc], %[average], 4 \n\t" + "replv.qb %[expected_dc], %[expected_dc] \n\t" + + "sw %[expected_dc], (%[dst]) \n\t" + "sw %[expected_dc], 4(%[dst]) \n\t" + + "add %[dst], %[dst], %[stride] \n\t" + "sw %[expected_dc], (%[dst]) \n\t" + "sw %[expected_dc], 4(%[dst]) \n\t" + + "add %[dst], %[dst], %[stride] \n\t" + "sw %[expected_dc], (%[dst]) \n\t" + "sw %[expected_dc], 4(%[dst]) \n\t" + + "add %[dst], %[dst], %[stride] \n\t" + "sw %[expected_dc], (%[dst]) \n\t" + "sw %[expected_dc], 4(%[dst]) \n\t" + + "add %[dst], %[dst], %[stride] \n\t" + "sw %[expected_dc], (%[dst]) \n\t" + "sw %[expected_dc], 4(%[dst]) \n\t" + + "add %[dst], %[dst], %[stride] \n\t" + "sw %[expected_dc], (%[dst]) \n\t" + "sw %[expected_dc], 4(%[dst]) \n\t" + + "add %[dst], %[dst], %[stride] \n\t" + "sw %[expected_dc], (%[dst]) \n\t" + "sw %[expected_dc], 4(%[dst]) \n\t" + + "add %[dst], %[dst], %[stride] \n\t" + "sw %[expected_dc], (%[dst]) \n\t" + "sw %[expected_dc], 4(%[dst]) \n\t" + + : [above1] "=&r"(above1), [above_l1] "=&r"(above_l1), + [above_r1] "=&r"(above_r1), [left1] "=&r"(left1), + [left_l1] "=&r"(left_l1), [left_r1] "=&r"(left_r1), + [above2] "=&r"(above2), [above_l2] "=&r"(above_l2), + [above_r2] "=&r"(above_r2), [left2] "=&r"(left2), + [left_l2] "=&r"(left_l2), [left_r2] "=&r"(left_r2), + [average] "=&r"(average), [tmp] "=&r"(tmp), + [expected_dc] "=&r"(expected_dc) + : [above] "r"(above), [left] "r"(left), [dst] "r"(dst), + [stride] "r"(stride)); +} +#endif // #if HAVE_DSPR2 diff --git a/third_party/aom/aom_dsp/mips/intrapred_msa.c b/third_party/aom/aom_dsp/mips/intrapred_msa.c new file mode 100644 index 0000000000..9f25cc1ca0 --- /dev/null +++ b/third_party/aom/aom_dsp/mips/intrapred_msa.c @@ -0,0 +1,550 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include "config/aom_dsp_rtcd.h" + +#include "aom_dsp/mips/macros_msa.h" + +#define IPRED_SUBS_UH2_UH(in0, in1, out0, out1) \ + { \ + out0 = __msa_subs_u_h(out0, in0); \ + out1 = __msa_subs_u_h(out1, in1); \ + } + +static void intra_predict_vert_4x4_msa(const uint8_t *src, uint8_t *dst, + int32_t dst_stride) { + uint32_t src_data; + + src_data = LW(src); + + SW4(src_data, src_data, src_data, src_data, dst, dst_stride); +} + +static void intra_predict_vert_8x8_msa(const uint8_t *src, uint8_t *dst, + int32_t dst_stride) { + uint32_t row; + uint32_t src_data1, src_data2; + + src_data1 = LW(src); + src_data2 = LW(src + 4); + + for (row = 8; row--;) { + SW(src_data1, dst); + SW(src_data2, (dst + 4)); + dst += dst_stride; + } +} + +static void intra_predict_vert_16x16_msa(const uint8_t *src, uint8_t *dst, + int32_t dst_stride) { + uint32_t row; + v16u8 src0; + + src0 = LD_UB(src); + + for (row = 16; row--;) { + ST_UB(src0, dst); + dst += dst_stride; + } +} + +static void intra_predict_vert_32x32_msa(const uint8_t *src, uint8_t *dst, + int32_t dst_stride) { + uint32_t row; + v16u8 src1, src2; + + src1 = LD_UB(src); + src2 = LD_UB(src + 16); + + for (row = 32; row--;) { + ST_UB2(src1, src2, dst, 16); + dst += dst_stride; + } +} + +static void intra_predict_horiz_4x4_msa(const uint8_t *src, uint8_t *dst, + int32_t dst_stride) { + uint32_t out0, out1, out2, out3; + + out0 = src[0] * 0x01010101; + out1 = src[1] * 0x01010101; + out2 = src[2] * 0x01010101; + out3 = src[3] * 0x01010101; + + SW4(out0, out1, out2, out3, dst, dst_stride); +} + +static void intra_predict_horiz_8x8_msa(const uint8_t *src, uint8_t *dst, + int32_t dst_stride) { + uint64_t out0, out1, out2, out3, out4, out5, out6, out7; + + out0 = src[0] * 0x0101010101010101ull; + out1 = src[1] * 0x0101010101010101ull; + out2 = src[2] * 0x0101010101010101ull; + out3 = src[3] * 0x0101010101010101ull; + out4 = src[4] * 0x0101010101010101ull; + out5 = src[5] * 0x0101010101010101ull; + out6 = src[6] * 0x0101010101010101ull; + out7 = src[7] * 0x0101010101010101ull; + + SD4(out0, out1, out2, out3, dst, dst_stride); + dst += (4 * dst_stride); + SD4(out4, out5, out6, out7, dst, dst_stride); +} + +static void intra_predict_horiz_16x16_msa(const uint8_t *src, uint8_t *dst, + int32_t dst_stride) { + uint32_t row; + uint8_t inp0, inp1, inp2, inp3; + v16u8 src0, src1, src2, src3; + + for (row = 4; row--;) { + inp0 = src[0]; + inp1 = src[1]; + inp2 = src[2]; + inp3 = src[3]; + src += 4; + + src0 = (v16u8)__msa_fill_b(inp0); + src1 = (v16u8)__msa_fill_b(inp1); + src2 = (v16u8)__msa_fill_b(inp2); + src3 = (v16u8)__msa_fill_b(inp3); + + ST_UB4(src0, src1, src2, src3, dst, dst_stride); + dst += (4 * dst_stride); + } +} + +static void intra_predict_horiz_32x32_msa(const uint8_t *src, uint8_t *dst, + int32_t dst_stride) { + uint32_t row; + uint8_t inp0, inp1, inp2, inp3; + v16u8 src0, src1, src2, src3; + + for (row = 8; row--;) { + inp0 = src[0]; + inp1 = src[1]; + inp2 = src[2]; + inp3 = src[3]; + src += 4; + + src0 = (v16u8)__msa_fill_b(inp0); + src1 = (v16u8)__msa_fill_b(inp1); + src2 = (v16u8)__msa_fill_b(inp2); + src3 = (v16u8)__msa_fill_b(inp3); + + ST_UB2(src0, src0, dst, 16); + dst += dst_stride; + ST_UB2(src1, src1, dst, 16); + dst += dst_stride; + ST_UB2(src2, src2, dst, 16); + dst += dst_stride; + ST_UB2(src3, src3, dst, 16); + dst += dst_stride; + } +} + +static void intra_predict_dc_4x4_msa(const uint8_t *src_top, + const uint8_t *src_left, uint8_t *dst, + int32_t dst_stride) { + uint32_t val0, val1; + v16i8 store, src = { 0 }; + v8u16 sum_h; + v4u32 sum_w; + v2u64 sum_d; + + val0 = LW(src_top); + val1 = LW(src_left); + INSERT_W2_SB(val0, val1, src); + sum_h = __msa_hadd_u_h((v16u8)src, (v16u8)src); + sum_w = __msa_hadd_u_w(sum_h, sum_h); + sum_d = __msa_hadd_u_d(sum_w, sum_w); + sum_w = (v4u32)__msa_srari_w((v4i32)sum_d, 3); + store = __msa_splati_b((v16i8)sum_w, 0); + val0 = __msa_copy_u_w((v4i32)store, 0); + + SW4(val0, val0, val0, val0, dst, dst_stride); +} + +static void intra_predict_dc_tl_4x4_msa(const uint8_t *src, uint8_t *dst, + int32_t dst_stride) { + uint32_t val0; + v16i8 store, data = { 0 }; + v8u16 sum_h; + v4u32 sum_w; + + val0 = LW(src); + data = (v16i8)__msa_insert_w((v4i32)data, 0, val0); + sum_h = __msa_hadd_u_h((v16u8)data, (v16u8)data); + sum_w = __msa_hadd_u_w(sum_h, sum_h); + sum_w = (v4u32)__msa_srari_w((v4i32)sum_w, 2); + store = __msa_splati_b((v16i8)sum_w, 0); + val0 = __msa_copy_u_w((v4i32)store, 0); + + SW4(val0, val0, val0, val0, dst, dst_stride); +} + +static void intra_predict_128dc_4x4_msa(uint8_t *dst, int32_t dst_stride) { + uint32_t out; + const v16i8 store = __msa_ldi_b(128); + + out = __msa_copy_u_w((v4i32)store, 0); + + SW4(out, out, out, out, dst, dst_stride); +} + +static void intra_predict_dc_8x8_msa(const uint8_t *src_top, + const uint8_t *src_left, uint8_t *dst, + int32_t dst_stride) { + uint64_t val0, val1; + v16i8 store; + v16u8 src = { 0 }; + v8u16 sum_h; + v4u32 sum_w; + v2u64 sum_d; + + val0 = LD(src_top); + val1 = LD(src_left); + INSERT_D2_UB(val0, val1, src); + sum_h = __msa_hadd_u_h(src, src); + sum_w = __msa_hadd_u_w(sum_h, sum_h); + sum_d = __msa_hadd_u_d(sum_w, sum_w); + sum_w = (v4u32)__msa_pckev_w((v4i32)sum_d, (v4i32)sum_d); + sum_d = __msa_hadd_u_d(sum_w, sum_w); + sum_w = (v4u32)__msa_srari_w((v4i32)sum_d, 4); + store = __msa_splati_b((v16i8)sum_w, 0); + val0 = __msa_copy_u_d((v2i64)store, 0); + + SD4(val0, val0, val0, val0, dst, dst_stride); + dst += (4 * dst_stride); + SD4(val0, val0, val0, val0, dst, dst_stride); +} + +static void intra_predict_dc_tl_8x8_msa(const uint8_t *src, uint8_t *dst, + int32_t dst_stride) { + uint64_t val0; + v16i8 store; + v16u8 data = { 0 }; + v8u16 sum_h; + v4u32 sum_w; + v2u64 sum_d; + + val0 = LD(src); + data = (v16u8)__msa_insert_d((v2i64)data, 0, val0); + sum_h = __msa_hadd_u_h(data, data); + sum_w = __msa_hadd_u_w(sum_h, sum_h); + sum_d = __msa_hadd_u_d(sum_w, sum_w); + sum_w = (v4u32)__msa_srari_w((v4i32)sum_d, 3); + store = __msa_splati_b((v16i8)sum_w, 0); + val0 = __msa_copy_u_d((v2i64)store, 0); + + SD4(val0, val0, val0, val0, dst, dst_stride); + dst += (4 * dst_stride); + SD4(val0, val0, val0, val0, dst, dst_stride); +} + +static void intra_predict_128dc_8x8_msa(uint8_t *dst, int32_t dst_stride) { + uint64_t out; + const v16i8 store = __msa_ldi_b(128); + + out = __msa_copy_u_d((v2i64)store, 0); + + SD4(out, out, out, out, dst, dst_stride); + dst += (4 * dst_stride); + SD4(out, out, out, out, dst, dst_stride); +} + +static void intra_predict_dc_16x16_msa(const uint8_t *src_top, + const uint8_t *src_left, uint8_t *dst, + int32_t dst_stride) { + v16u8 top, left, out; + v8u16 sum_h, sum_top, sum_left; + v4u32 sum_w; + v2u64 sum_d; + + top = LD_UB(src_top); + left = LD_UB(src_left); + HADD_UB2_UH(top, left, sum_top, sum_left); + sum_h = sum_top + sum_left; + sum_w = __msa_hadd_u_w(sum_h, sum_h); + sum_d = __msa_hadd_u_d(sum_w, sum_w); + sum_w = (v4u32)__msa_pckev_w((v4i32)sum_d, (v4i32)sum_d); + sum_d = __msa_hadd_u_d(sum_w, sum_w); + sum_w = (v4u32)__msa_srari_w((v4i32)sum_d, 5); + out = (v16u8)__msa_splati_b((v16i8)sum_w, 0); + + ST_UB8(out, out, out, out, out, out, out, out, dst, dst_stride); + dst += (8 * dst_stride); + ST_UB8(out, out, out, out, out, out, out, out, dst, dst_stride); +} + +static void intra_predict_dc_tl_16x16_msa(const uint8_t *src, uint8_t *dst, + int32_t dst_stride) { + v16u8 data, out; + v8u16 sum_h; + v4u32 sum_w; + v2u64 sum_d; + + data = LD_UB(src); + sum_h = __msa_hadd_u_h(data, data); + sum_w = __msa_hadd_u_w(sum_h, sum_h); + sum_d = __msa_hadd_u_d(sum_w, sum_w); + sum_w = (v4u32)__msa_pckev_w((v4i32)sum_d, (v4i32)sum_d); + sum_d = __msa_hadd_u_d(sum_w, sum_w); + sum_w = (v4u32)__msa_srari_w((v4i32)sum_d, 4); + out = (v16u8)__msa_splati_b((v16i8)sum_w, 0); + + ST_UB8(out, out, out, out, out, out, out, out, dst, dst_stride); + dst += (8 * dst_stride); + ST_UB8(out, out, out, out, out, out, out, out, dst, dst_stride); +} + +static void intra_predict_128dc_16x16_msa(uint8_t *dst, int32_t dst_stride) { + const v16u8 out = (v16u8)__msa_ldi_b(128); + + ST_UB8(out, out, out, out, out, out, out, out, dst, dst_stride); + dst += (8 * dst_stride); + ST_UB8(out, out, out, out, out, out, out, out, dst, dst_stride); +} + +static void intra_predict_dc_32x32_msa(const uint8_t *src_top, + const uint8_t *src_left, uint8_t *dst, + int32_t dst_stride) { + uint32_t row; + v16u8 top0, top1, left0, left1, out; + v8u16 sum_h, sum_top0, sum_top1, sum_left0, sum_left1; + v4u32 sum_w; + v2u64 sum_d; + + LD_UB2(src_top, 16, top0, top1); + LD_UB2(src_left, 16, left0, left1); + HADD_UB2_UH(top0, top1, sum_top0, sum_top1); + HADD_UB2_UH(left0, left1, sum_left0, sum_left1); + sum_h = sum_top0 + sum_top1; + sum_h += sum_left0 + sum_left1; + sum_w = __msa_hadd_u_w(sum_h, sum_h); + sum_d = __msa_hadd_u_d(sum_w, sum_w); + sum_w = (v4u32)__msa_pckev_w((v4i32)sum_d, (v4i32)sum_d); + sum_d = __msa_hadd_u_d(sum_w, sum_w); + sum_w = (v4u32)__msa_srari_w((v4i32)sum_d, 6); + out = (v16u8)__msa_splati_b((v16i8)sum_w, 0); + + for (row = 16; row--;) { + ST_UB2(out, out, dst, 16); + dst += dst_stride; + ST_UB2(out, out, dst, 16); + dst += dst_stride; + } +} + +static void intra_predict_dc_tl_32x32_msa(const uint8_t *src, uint8_t *dst, + int32_t dst_stride) { + uint32_t row; + v16u8 data0, data1, out; + v8u16 sum_h, sum_data0, sum_data1; + v4u32 sum_w; + v2u64 sum_d; + + LD_UB2(src, 16, data0, data1); + HADD_UB2_UH(data0, data1, sum_data0, sum_data1); + sum_h = sum_data0 + sum_data1; + sum_w = __msa_hadd_u_w(sum_h, sum_h); + sum_d = __msa_hadd_u_d(sum_w, sum_w); + sum_w = (v4u32)__msa_pckev_w((v4i32)sum_d, (v4i32)sum_d); + sum_d = __msa_hadd_u_d(sum_w, sum_w); + sum_w = (v4u32)__msa_srari_w((v4i32)sum_d, 5); + out = (v16u8)__msa_splati_b((v16i8)sum_w, 0); + + for (row = 16; row--;) { + ST_UB2(out, out, dst, 16); + dst += dst_stride; + ST_UB2(out, out, dst, 16); + dst += dst_stride; + } +} + +static void intra_predict_128dc_32x32_msa(uint8_t *dst, int32_t dst_stride) { + uint32_t row; + const v16u8 out = (v16u8)__msa_ldi_b(128); + + for (row = 16; row--;) { + ST_UB2(out, out, dst, 16); + dst += dst_stride; + ST_UB2(out, out, dst, 16); + dst += dst_stride; + } +} + +void aom_v_predictor_4x4_msa(uint8_t *dst, ptrdiff_t y_stride, + const uint8_t *above, const uint8_t *left) { + (void)left; + + intra_predict_vert_4x4_msa(above, dst, y_stride); +} + +void aom_v_predictor_8x8_msa(uint8_t *dst, ptrdiff_t y_stride, + const uint8_t *above, const uint8_t *left) { + (void)left; + + intra_predict_vert_8x8_msa(above, dst, y_stride); +} + +void aom_v_predictor_16x16_msa(uint8_t *dst, ptrdiff_t y_stride, + const uint8_t *above, const uint8_t *left) { + (void)left; + + intra_predict_vert_16x16_msa(above, dst, y_stride); +} + +void aom_v_predictor_32x32_msa(uint8_t *dst, ptrdiff_t y_stride, + const uint8_t *above, const uint8_t *left) { + (void)left; + + intra_predict_vert_32x32_msa(above, dst, y_stride); +} + +void aom_h_predictor_4x4_msa(uint8_t *dst, ptrdiff_t y_stride, + const uint8_t *above, const uint8_t *left) { + (void)above; + + intra_predict_horiz_4x4_msa(left, dst, y_stride); +} + +void aom_h_predictor_8x8_msa(uint8_t *dst, ptrdiff_t y_stride, + const uint8_t *above, const uint8_t *left) { + (void)above; + + intra_predict_horiz_8x8_msa(left, dst, y_stride); +} + +void aom_h_predictor_16x16_msa(uint8_t *dst, ptrdiff_t y_stride, + const uint8_t *above, const uint8_t *left) { + (void)above; + + intra_predict_horiz_16x16_msa(left, dst, y_stride); +} + +void aom_h_predictor_32x32_msa(uint8_t *dst, ptrdiff_t y_stride, + const uint8_t *above, const uint8_t *left) { + (void)above; + + intra_predict_horiz_32x32_msa(left, dst, y_stride); +} + +void aom_dc_predictor_4x4_msa(uint8_t *dst, ptrdiff_t y_stride, + const uint8_t *above, const uint8_t *left) { + intra_predict_dc_4x4_msa(above, left, dst, y_stride); +} + +void aom_dc_predictor_8x8_msa(uint8_t *dst, ptrdiff_t y_stride, + const uint8_t *above, const uint8_t *left) { + intra_predict_dc_8x8_msa(above, left, dst, y_stride); +} + +void aom_dc_predictor_16x16_msa(uint8_t *dst, ptrdiff_t y_stride, + const uint8_t *above, const uint8_t *left) { + intra_predict_dc_16x16_msa(above, left, dst, y_stride); +} + +void aom_dc_predictor_32x32_msa(uint8_t *dst, ptrdiff_t y_stride, + const uint8_t *above, const uint8_t *left) { + intra_predict_dc_32x32_msa(above, left, dst, y_stride); +} + +void aom_dc_top_predictor_4x4_msa(uint8_t *dst, ptrdiff_t y_stride, + const uint8_t *above, const uint8_t *left) { + (void)left; + + intra_predict_dc_tl_4x4_msa(above, dst, y_stride); +} + +void aom_dc_top_predictor_8x8_msa(uint8_t *dst, ptrdiff_t y_stride, + const uint8_t *above, const uint8_t *left) { + (void)left; + + intra_predict_dc_tl_8x8_msa(above, dst, y_stride); +} + +void aom_dc_top_predictor_16x16_msa(uint8_t *dst, ptrdiff_t y_stride, + const uint8_t *above, const uint8_t *left) { + (void)left; + + intra_predict_dc_tl_16x16_msa(above, dst, y_stride); +} + +void aom_dc_top_predictor_32x32_msa(uint8_t *dst, ptrdiff_t y_stride, + const uint8_t *above, const uint8_t *left) { + (void)left; + + intra_predict_dc_tl_32x32_msa(above, dst, y_stride); +} + +void aom_dc_left_predictor_4x4_msa(uint8_t *dst, ptrdiff_t y_stride, + const uint8_t *above, const uint8_t *left) { + (void)above; + + intra_predict_dc_tl_4x4_msa(left, dst, y_stride); +} + +void aom_dc_left_predictor_8x8_msa(uint8_t *dst, ptrdiff_t y_stride, + const uint8_t *above, const uint8_t *left) { + (void)above; + + intra_predict_dc_tl_8x8_msa(left, dst, y_stride); +} + +void aom_dc_left_predictor_16x16_msa(uint8_t *dst, ptrdiff_t y_stride, + const uint8_t *above, + const uint8_t *left) { + (void)above; + + intra_predict_dc_tl_16x16_msa(left, dst, y_stride); +} + +void aom_dc_left_predictor_32x32_msa(uint8_t *dst, ptrdiff_t y_stride, + const uint8_t *above, + const uint8_t *left) { + (void)above; + + intra_predict_dc_tl_32x32_msa(left, dst, y_stride); +} + +void aom_dc_128_predictor_4x4_msa(uint8_t *dst, ptrdiff_t y_stride, + const uint8_t *above, const uint8_t *left) { + (void)above; + (void)left; + + intra_predict_128dc_4x4_msa(dst, y_stride); +} + +void aom_dc_128_predictor_8x8_msa(uint8_t *dst, ptrdiff_t y_stride, + const uint8_t *above, const uint8_t *left) { + (void)above; + (void)left; + + intra_predict_128dc_8x8_msa(dst, y_stride); +} + +void aom_dc_128_predictor_16x16_msa(uint8_t *dst, ptrdiff_t y_stride, + const uint8_t *above, const uint8_t *left) { + (void)above; + (void)left; + + intra_predict_128dc_16x16_msa(dst, y_stride); +} + +void aom_dc_128_predictor_32x32_msa(uint8_t *dst, ptrdiff_t y_stride, + const uint8_t *above, const uint8_t *left) { + (void)above; + (void)left; + + intra_predict_128dc_32x32_msa(dst, y_stride); +} diff --git a/third_party/aom/aom_dsp/mips/loopfilter_16_msa.c b/third_party/aom/aom_dsp/mips/loopfilter_16_msa.c new file mode 100644 index 0000000000..38a10e9b22 --- /dev/null +++ b/third_party/aom/aom_dsp/mips/loopfilter_16_msa.c @@ -0,0 +1,1488 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include "aom_ports/mem.h" +#include "aom_dsp/mips/loopfilter_msa.h" + +int32_t aom_hz_lpf_t4_and_t8_16w(uint8_t *src, int32_t pitch, uint8_t *filter48, + const uint8_t *b_limit_ptr, + const uint8_t *limit_ptr, + const uint8_t *thresh_ptr) { + v16u8 p3, p2, p1, p0, q3, q2, q1, q0; + v16u8 p2_out, p1_out, p0_out, q0_out, q1_out, q2_out; + v16u8 flat, mask, hev, thresh, b_limit, limit; + v8u16 p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r; + v8u16 p3_l, p2_l, p1_l, p0_l, q0_l, q1_l, q2_l, q3_l; + v8i16 p2_filt8_r, p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r; + v8i16 p2_filt8_l, p1_filt8_l, p0_filt8_l, q0_filt8_l, q1_filt8_l, q2_filt8_l; + v16u8 zero = { 0 }; + + /* load vector elements */ + LD_UB8(src - (4 * pitch), pitch, p3, p2, p1, p0, q0, q1, q2, q3); + + thresh = (v16u8)__msa_fill_b(*thresh_ptr); + b_limit = (v16u8)__msa_fill_b(*b_limit_ptr); + limit = (v16u8)__msa_fill_b(*limit_ptr); + + /* mask and hev */ + LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh, hev, + mask, flat); + AOM_FLAT4(p3, p2, p0, q0, q2, q3, flat); + AOM_LPF_FILTER4_4W(p1, p0, q0, q1, mask, hev, p1_out, p0_out, q0_out, q1_out); + + if (__msa_test_bz_v(flat)) { + ST_UB4(p1_out, p0_out, q0_out, q1_out, (src - 2 * pitch), pitch); + + return 1; + } else { + ILVR_B8_UH(zero, p3, zero, p2, zero, p1, zero, p0, zero, q0, zero, q1, zero, + q2, zero, q3, p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r); + AOM_FILTER8(p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r, p2_filt8_r, + p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r); + + ILVL_B4_UH(zero, p3, zero, p2, zero, p1, zero, p0, p3_l, p2_l, p1_l, p0_l); + ILVL_B4_UH(zero, q0, zero, q1, zero, q2, zero, q3, q0_l, q1_l, q2_l, q3_l); + AOM_FILTER8(p3_l, p2_l, p1_l, p0_l, q0_l, q1_l, q2_l, q3_l, p2_filt8_l, + p1_filt8_l, p0_filt8_l, q0_filt8_l, q1_filt8_l, q2_filt8_l); + + /* convert 16 bit output data into 8 bit */ + PCKEV_B4_SH(p2_filt8_l, p2_filt8_r, p1_filt8_l, p1_filt8_r, p0_filt8_l, + p0_filt8_r, q0_filt8_l, q0_filt8_r, p2_filt8_r, p1_filt8_r, + p0_filt8_r, q0_filt8_r); + PCKEV_B2_SH(q1_filt8_l, q1_filt8_r, q2_filt8_l, q2_filt8_r, q1_filt8_r, + q2_filt8_r); + + /* store pixel values */ + p2_out = __msa_bmnz_v(p2, (v16u8)p2_filt8_r, flat); + p1_out = __msa_bmnz_v(p1_out, (v16u8)p1_filt8_r, flat); + p0_out = __msa_bmnz_v(p0_out, (v16u8)p0_filt8_r, flat); + q0_out = __msa_bmnz_v(q0_out, (v16u8)q0_filt8_r, flat); + q1_out = __msa_bmnz_v(q1_out, (v16u8)q1_filt8_r, flat); + q2_out = __msa_bmnz_v(q2, (v16u8)q2_filt8_r, flat); + + ST_UB4(p2_out, p1_out, p0_out, q0_out, filter48, 16); + filter48 += (4 * 16); + ST_UB2(q1_out, q2_out, filter48, 16); + filter48 += (2 * 16); + ST_UB(flat, filter48); + + return 0; + } +} + +void aom_hz_lpf_t16_16w(uint8_t *src, int32_t pitch, uint8_t *filter48) { + v16u8 flat, flat2, filter8; + v16i8 zero = { 0 }; + v16u8 p7, p6, p5, p4, p3, p2, p1, p0, q0, q1, q2, q3, q4, q5, q6, q7; + v8u16 p7_r_in, p6_r_in, p5_r_in, p4_r_in, p3_r_in, p2_r_in, p1_r_in, p0_r_in; + v8u16 q7_r_in, q6_r_in, q5_r_in, q4_r_in, q3_r_in, q2_r_in, q1_r_in, q0_r_in; + v8u16 p7_l_in, p6_l_in, p5_l_in, p4_l_in, p3_l_in, p2_l_in, p1_l_in, p0_l_in; + v8u16 q7_l_in, q6_l_in, q5_l_in, q4_l_in, q3_l_in, q2_l_in, q1_l_in, q0_l_in; + v8u16 tmp0_r, tmp1_r, tmp0_l, tmp1_l; + v8i16 l_out, r_out; + + flat = LD_UB(filter48 + 96); + + LD_UB8((src - 8 * pitch), pitch, p7, p6, p5, p4, p3, p2, p1, p0); + LD_UB8(src, pitch, q0, q1, q2, q3, q4, q5, q6, q7); + AOM_FLAT5(p7, p6, p5, p4, p0, q0, q4, q5, q6, q7, flat, flat2); + + if (__msa_test_bz_v(flat2)) { + LD_UB4(filter48, 16, p2, p1, p0, q0); + LD_UB2(filter48 + 4 * 16, 16, q1, q2); + + src -= 3 * pitch; + ST_UB4(p2, p1, p0, q0, src, pitch); + src += (4 * pitch); + ST_UB2(q1, q2, src, pitch); + } else { + src -= 7 * pitch; + + ILVR_B8_UH(zero, p7, zero, p6, zero, p5, zero, p4, zero, p3, zero, p2, zero, + p1, zero, p0, p7_r_in, p6_r_in, p5_r_in, p4_r_in, p3_r_in, + p2_r_in, p1_r_in, p0_r_in); + + q0_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q0); + + tmp0_r = p7_r_in << 3; + tmp0_r -= p7_r_in; + tmp0_r += p6_r_in; + tmp0_r += q0_r_in; + tmp1_r = p6_r_in + p5_r_in; + tmp1_r += p4_r_in; + tmp1_r += p3_r_in; + tmp1_r += p2_r_in; + tmp1_r += p1_r_in; + tmp1_r += p0_r_in; + tmp1_r += tmp0_r; + r_out = __msa_srari_h((v8i16)tmp1_r, 4); + + ILVL_B4_UH(zero, p7, zero, p6, zero, p5, zero, p4, p7_l_in, p6_l_in, + p5_l_in, p4_l_in); + ILVL_B4_UH(zero, p3, zero, p2, zero, p1, zero, p0, p3_l_in, p2_l_in, + p1_l_in, p0_l_in); + q0_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q0); + + tmp0_l = p7_l_in << 3; + tmp0_l -= p7_l_in; + tmp0_l += p6_l_in; + tmp0_l += q0_l_in; + tmp1_l = p6_l_in + p5_l_in; + tmp1_l += p4_l_in; + tmp1_l += p3_l_in; + tmp1_l += p2_l_in; + tmp1_l += p1_l_in; + tmp1_l += p0_l_in; + tmp1_l += tmp0_l; + l_out = __msa_srari_h((v8i16)tmp1_l, 4); + + r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out); + p6 = __msa_bmnz_v(p6, (v16u8)r_out, flat2); + ST_UB(p6, src); + src += pitch; + + /* p5 */ + q1_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q1); + tmp0_r = p5_r_in - p6_r_in; + tmp0_r += q1_r_in; + tmp0_r -= p7_r_in; + tmp1_r += tmp0_r; + r_out = __msa_srari_h((v8i16)tmp1_r, 4); + + q1_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q1); + tmp0_l = p5_l_in - p6_l_in; + tmp0_l += q1_l_in; + tmp0_l -= p7_l_in; + tmp1_l += tmp0_l; + l_out = __msa_srari_h((v8i16)tmp1_l, 4); + + r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out); + p5 = __msa_bmnz_v(p5, (v16u8)r_out, flat2); + ST_UB(p5, src); + src += pitch; + + /* p4 */ + q2_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q2); + tmp0_r = p4_r_in - p5_r_in; + tmp0_r += q2_r_in; + tmp0_r -= p7_r_in; + tmp1_r += tmp0_r; + r_out = (v8i16)__msa_srari_h((v8i16)tmp1_r, 4); + + q2_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q2); + tmp0_l = p4_l_in - p5_l_in; + tmp0_l += q2_l_in; + tmp0_l -= p7_l_in; + tmp1_l += tmp0_l; + l_out = __msa_srari_h((v8i16)tmp1_l, 4); + + r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out); + p4 = __msa_bmnz_v(p4, (v16u8)r_out, flat2); + ST_UB(p4, src); + src += pitch; + + /* p3 */ + q3_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q3); + tmp0_r = p3_r_in - p4_r_in; + tmp0_r += q3_r_in; + tmp0_r -= p7_r_in; + tmp1_r += tmp0_r; + r_out = __msa_srari_h((v8i16)tmp1_r, 4); + + q3_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q3); + tmp0_l = p3_l_in - p4_l_in; + tmp0_l += q3_l_in; + tmp0_l -= p7_l_in; + tmp1_l += tmp0_l; + l_out = __msa_srari_h((v8i16)tmp1_l, 4); + + r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out); + p3 = __msa_bmnz_v(p3, (v16u8)r_out, flat2); + ST_UB(p3, src); + src += pitch; + + /* p2 */ + q4_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q4); + filter8 = LD_UB(filter48); + tmp0_r = p2_r_in - p3_r_in; + tmp0_r += q4_r_in; + tmp0_r -= p7_r_in; + tmp1_r += tmp0_r; + r_out = __msa_srari_h((v8i16)tmp1_r, 4); + + q4_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q4); + tmp0_l = p2_l_in - p3_l_in; + tmp0_l += q4_l_in; + tmp0_l -= p7_l_in; + tmp1_l += tmp0_l; + l_out = __msa_srari_h((v8i16)tmp1_l, 4); + + r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out); + filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2); + ST_UB(filter8, src); + src += pitch; + + /* p1 */ + q5_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q5); + filter8 = LD_UB(filter48 + 16); + tmp0_r = p1_r_in - p2_r_in; + tmp0_r += q5_r_in; + tmp0_r -= p7_r_in; + tmp1_r += tmp0_r; + r_out = __msa_srari_h((v8i16)tmp1_r, 4); + + q5_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q5); + tmp0_l = p1_l_in - p2_l_in; + tmp0_l += q5_l_in; + tmp0_l -= p7_l_in; + tmp1_l += tmp0_l; + l_out = __msa_srari_h((v8i16)tmp1_l, 4); + + r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out); + filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2); + ST_UB(filter8, src); + src += pitch; + + /* p0 */ + q6_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q6); + filter8 = LD_UB(filter48 + 32); + tmp0_r = p0_r_in - p1_r_in; + tmp0_r += q6_r_in; + tmp0_r -= p7_r_in; + tmp1_r += tmp0_r; + r_out = __msa_srari_h((v8i16)tmp1_r, 4); + + q6_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q6); + tmp0_l = p0_l_in - p1_l_in; + tmp0_l += q6_l_in; + tmp0_l -= p7_l_in; + tmp1_l += tmp0_l; + l_out = __msa_srari_h((v8i16)tmp1_l, 4); + + r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out); + filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2); + ST_UB(filter8, src); + src += pitch; + + /* q0 */ + q7_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q7); + filter8 = LD_UB(filter48 + 48); + tmp0_r = q7_r_in - p0_r_in; + tmp0_r += q0_r_in; + tmp0_r -= p7_r_in; + tmp1_r += tmp0_r; + r_out = __msa_srari_h((v8i16)tmp1_r, 4); + + q7_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q7); + tmp0_l = q7_l_in - p0_l_in; + tmp0_l += q0_l_in; + tmp0_l -= p7_l_in; + tmp1_l += tmp0_l; + l_out = __msa_srari_h((v8i16)tmp1_l, 4); + + r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out); + filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2); + ST_UB(filter8, src); + src += pitch; + + /* q1 */ + filter8 = LD_UB(filter48 + 64); + tmp0_r = q7_r_in - q0_r_in; + tmp0_r += q1_r_in; + tmp0_r -= p6_r_in; + tmp1_r += tmp0_r; + r_out = __msa_srari_h((v8i16)tmp1_r, 4); + + tmp0_l = q7_l_in - q0_l_in; + tmp0_l += q1_l_in; + tmp0_l -= p6_l_in; + tmp1_l += tmp0_l; + l_out = __msa_srari_h((v8i16)tmp1_l, 4); + + r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out); + filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2); + ST_UB(filter8, src); + src += pitch; + + /* q2 */ + filter8 = LD_UB(filter48 + 80); + tmp0_r = q7_r_in - q1_r_in; + tmp0_r += q2_r_in; + tmp0_r -= p5_r_in; + tmp1_r += tmp0_r; + r_out = __msa_srari_h((v8i16)tmp1_r, 4); + + tmp0_l = q7_l_in - q1_l_in; + tmp0_l += q2_l_in; + tmp0_l -= p5_l_in; + tmp1_l += tmp0_l; + l_out = __msa_srari_h((v8i16)tmp1_l, 4); + + r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out); + filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2); + ST_UB(filter8, src); + src += pitch; + + /* q3 */ + tmp0_r = q7_r_in - q2_r_in; + tmp0_r += q3_r_in; + tmp0_r -= p4_r_in; + tmp1_r += tmp0_r; + r_out = __msa_srari_h((v8i16)tmp1_r, 4); + + tmp0_l = q7_l_in - q2_l_in; + tmp0_l += q3_l_in; + tmp0_l -= p4_l_in; + tmp1_l += tmp0_l; + l_out = __msa_srari_h((v8i16)tmp1_l, 4); + + r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out); + q3 = __msa_bmnz_v(q3, (v16u8)r_out, flat2); + ST_UB(q3, src); + src += pitch; + + /* q4 */ + tmp0_r = q7_r_in - q3_r_in; + tmp0_r += q4_r_in; + tmp0_r -= p3_r_in; + tmp1_r += tmp0_r; + r_out = __msa_srari_h((v8i16)tmp1_r, 4); + + tmp0_l = q7_l_in - q3_l_in; + tmp0_l += q4_l_in; + tmp0_l -= p3_l_in; + tmp1_l += tmp0_l; + l_out = __msa_srari_h((v8i16)tmp1_l, 4); + + r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out); + q4 = __msa_bmnz_v(q4, (v16u8)r_out, flat2); + ST_UB(q4, src); + src += pitch; + + /* q5 */ + tmp0_r = q7_r_in - q4_r_in; + tmp0_r += q5_r_in; + tmp0_r -= p2_r_in; + tmp1_r += tmp0_r; + r_out = __msa_srari_h((v8i16)tmp1_r, 4); + + tmp0_l = q7_l_in - q4_l_in; + tmp0_l += q5_l_in; + tmp0_l -= p2_l_in; + tmp1_l += tmp0_l; + l_out = __msa_srari_h((v8i16)tmp1_l, 4); + + r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out); + q5 = __msa_bmnz_v(q5, (v16u8)r_out, flat2); + ST_UB(q5, src); + src += pitch; + + /* q6 */ + tmp0_r = q7_r_in - q5_r_in; + tmp0_r += q6_r_in; + tmp0_r -= p1_r_in; + tmp1_r += tmp0_r; + r_out = __msa_srari_h((v8i16)tmp1_r, 4); + + tmp0_l = q7_l_in - q5_l_in; + tmp0_l += q6_l_in; + tmp0_l -= p1_l_in; + tmp1_l += tmp0_l; + l_out = __msa_srari_h((v8i16)tmp1_l, 4); + + r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out); + q6 = __msa_bmnz_v(q6, (v16u8)r_out, flat2); + ST_UB(q6, src); + } +} + +static void mb_lpf_horizontal_edge_dual(uint8_t *src, int32_t pitch, + const uint8_t *b_limit_ptr, + const uint8_t *limit_ptr, + const uint8_t *thresh_ptr, + int32_t count) { + DECLARE_ALIGNED(32, uint8_t, filter48[16 * 8]); + uint8_t early_exit = 0; + + (void)count; + + early_exit = aom_hz_lpf_t4_and_t8_16w(src, pitch, &filter48[0], b_limit_ptr, + limit_ptr, thresh_ptr); + + if (0 == early_exit) { + aom_hz_lpf_t16_16w(src, pitch, filter48); + } +} + +static void mb_lpf_horizontal_edge(uint8_t *src, int32_t pitch, + const uint8_t *b_limit_ptr, + const uint8_t *limit_ptr, + const uint8_t *thresh_ptr, int32_t count) { + if (1 == count) { + uint64_t p2_d, p1_d, p0_d, q0_d, q1_d, q2_d; + uint64_t dword0, dword1; + v16u8 flat2, mask, hev, flat, thresh, b_limit, limit; + v16u8 p3, p2, p1, p0, q3, q2, q1, q0, p7, p6, p5, p4, q4, q5, q6, q7; + v16u8 p2_out, p1_out, p0_out, q0_out, q1_out, q2_out; + v16u8 p0_filter16, p1_filter16; + v8i16 p2_filter8, p1_filter8, p0_filter8; + v8i16 q0_filter8, q1_filter8, q2_filter8; + v8u16 p7_r, p6_r, p5_r, p4_r, q7_r, q6_r, q5_r, q4_r; + v8u16 p3_r, p2_r, p1_r, p0_r, q3_r, q2_r, q1_r, q0_r; + v16i8 zero = { 0 }; + v8u16 tmp0, tmp1, tmp2; + + /* load vector elements */ + LD_UB8((src - 4 * pitch), pitch, p3, p2, p1, p0, q0, q1, q2, q3); + + thresh = (v16u8)__msa_fill_b(*thresh_ptr); + b_limit = (v16u8)__msa_fill_b(*b_limit_ptr); + limit = (v16u8)__msa_fill_b(*limit_ptr); + + LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh, hev, + mask, flat); + AOM_FLAT4(p3, p2, p0, q0, q2, q3, flat); + AOM_LPF_FILTER4_8W(p1, p0, q0, q1, mask, hev, p1_out, p0_out, q0_out, + q1_out); + + flat = (v16u8)__msa_ilvr_d((v2i64)zero, (v2i64)flat); + + if (__msa_test_bz_v(flat)) { + p1_d = __msa_copy_u_d((v2i64)p1_out, 0); + p0_d = __msa_copy_u_d((v2i64)p0_out, 0); + q0_d = __msa_copy_u_d((v2i64)q0_out, 0); + q1_d = __msa_copy_u_d((v2i64)q1_out, 0); + SD4(p1_d, p0_d, q0_d, q1_d, src - 2 * pitch, pitch); + } else { + /* convert 8 bit input data into 16 bit */ + ILVR_B8_UH(zero, p3, zero, p2, zero, p1, zero, p0, zero, q0, zero, q1, + zero, q2, zero, q3, p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, + q3_r); + AOM_FILTER8(p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r, p2_filter8, + p1_filter8, p0_filter8, q0_filter8, q1_filter8, q2_filter8); + + /* convert 16 bit output data into 8 bit */ + PCKEV_B4_SH(zero, p2_filter8, zero, p1_filter8, zero, p0_filter8, zero, + q0_filter8, p2_filter8, p1_filter8, p0_filter8, q0_filter8); + PCKEV_B2_SH(zero, q1_filter8, zero, q2_filter8, q1_filter8, q2_filter8); + + /* store pixel values */ + p2_out = __msa_bmnz_v(p2, (v16u8)p2_filter8, flat); + p1_out = __msa_bmnz_v(p1_out, (v16u8)p1_filter8, flat); + p0_out = __msa_bmnz_v(p0_out, (v16u8)p0_filter8, flat); + q0_out = __msa_bmnz_v(q0_out, (v16u8)q0_filter8, flat); + q1_out = __msa_bmnz_v(q1_out, (v16u8)q1_filter8, flat); + q2_out = __msa_bmnz_v(q2, (v16u8)q2_filter8, flat); + + /* load 16 vector elements */ + LD_UB4((src - 8 * pitch), pitch, p7, p6, p5, p4); + LD_UB4(src + (4 * pitch), pitch, q4, q5, q6, q7); + + AOM_FLAT5(p7, p6, p5, p4, p0, q0, q4, q5, q6, q7, flat, flat2); + + if (__msa_test_bz_v(flat2)) { + p2_d = __msa_copy_u_d((v2i64)p2_out, 0); + p1_d = __msa_copy_u_d((v2i64)p1_out, 0); + p0_d = __msa_copy_u_d((v2i64)p0_out, 0); + q0_d = __msa_copy_u_d((v2i64)q0_out, 0); + q1_d = __msa_copy_u_d((v2i64)q1_out, 0); + q2_d = __msa_copy_u_d((v2i64)q2_out, 0); + + SD4(p2_d, p1_d, p0_d, q0_d, src - 3 * pitch, pitch); + SD(q1_d, src + pitch); + SD(q2_d, src + 2 * pitch); + } else { + /* LSB(right) 8 pixel operation */ + ILVR_B8_UH(zero, p7, zero, p6, zero, p5, zero, p4, zero, q4, zero, q5, + zero, q6, zero, q7, p7_r, p6_r, p5_r, p4_r, q4_r, q5_r, q6_r, + q7_r); + + tmp0 = p7_r << 3; + tmp0 -= p7_r; + tmp0 += p6_r; + tmp0 += q0_r; + + src -= 7 * pitch; + + /* calculation of p6 and p5 */ + tmp1 = p6_r + p5_r + p4_r + p3_r; + tmp1 += (p2_r + p1_r + p0_r); + tmp1 += tmp0; + p0_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4); + tmp0 = p5_r - p6_r + q1_r - p7_r; + tmp1 += tmp0; + p1_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4); + PCKEV_B2_UB(zero, p0_filter16, zero, p1_filter16, p0_filter16, + p1_filter16); + p0_filter16 = __msa_bmnz_v(p6, p0_filter16, flat2); + p1_filter16 = __msa_bmnz_v(p5, p1_filter16, flat2); + dword0 = __msa_copy_u_d((v2i64)p0_filter16, 0); + dword1 = __msa_copy_u_d((v2i64)p1_filter16, 0); + SD(dword0, src); + src += pitch; + SD(dword1, src); + src += pitch; + + /* calculation of p4 and p3 */ + tmp0 = p4_r - p5_r + q2_r - p7_r; + tmp2 = p3_r - p4_r + q3_r - p7_r; + tmp1 += tmp0; + p0_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4); + tmp1 += tmp2; + p1_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4); + PCKEV_B2_UB(zero, p0_filter16, zero, p1_filter16, p0_filter16, + p1_filter16); + p0_filter16 = __msa_bmnz_v(p4, p0_filter16, flat2); + p1_filter16 = __msa_bmnz_v(p3, p1_filter16, flat2); + dword0 = __msa_copy_u_d((v2i64)p0_filter16, 0); + dword1 = __msa_copy_u_d((v2i64)p1_filter16, 0); + SD(dword0, src); + src += pitch; + SD(dword1, src); + src += pitch; + + /* calculation of p2 and p1 */ + tmp0 = p2_r - p3_r + q4_r - p7_r; + tmp2 = p1_r - p2_r + q5_r - p7_r; + tmp1 += tmp0; + p0_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4); + tmp1 += tmp2; + p1_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4); + PCKEV_B2_UB(zero, p0_filter16, zero, p1_filter16, p0_filter16, + p1_filter16); + p0_filter16 = __msa_bmnz_v(p2_out, p0_filter16, flat2); + p1_filter16 = __msa_bmnz_v(p1_out, p1_filter16, flat2); + dword0 = __msa_copy_u_d((v2i64)p0_filter16, 0); + dword1 = __msa_copy_u_d((v2i64)p1_filter16, 0); + SD(dword0, src); + src += pitch; + SD(dword1, src); + src += pitch; + + /* calculation of p0 and q0 */ + tmp0 = (p0_r - p1_r) + (q6_r - p7_r); + tmp2 = (q7_r - p0_r) + (q0_r - p7_r); + tmp1 += tmp0; + p0_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4); + tmp1 += tmp2; + p1_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4); + PCKEV_B2_UB(zero, p0_filter16, zero, p1_filter16, p0_filter16, + p1_filter16); + p0_filter16 = __msa_bmnz_v(p0_out, p0_filter16, flat2); + p1_filter16 = __msa_bmnz_v(q0_out, p1_filter16, flat2); + dword0 = __msa_copy_u_d((v2i64)p0_filter16, 0); + dword1 = __msa_copy_u_d((v2i64)p1_filter16, 0); + SD(dword0, src); + src += pitch; + SD(dword1, src); + src += pitch; + + /* calculation of q1 and q2 */ + tmp0 = q7_r - q0_r + q1_r - p6_r; + tmp2 = q7_r - q1_r + q2_r - p5_r; + tmp1 += tmp0; + p0_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4); + tmp1 += tmp2; + p1_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4); + PCKEV_B2_UB(zero, p0_filter16, zero, p1_filter16, p0_filter16, + p1_filter16); + p0_filter16 = __msa_bmnz_v(q1_out, p0_filter16, flat2); + p1_filter16 = __msa_bmnz_v(q2_out, p1_filter16, flat2); + dword0 = __msa_copy_u_d((v2i64)p0_filter16, 0); + dword1 = __msa_copy_u_d((v2i64)p1_filter16, 0); + SD(dword0, src); + src += pitch; + SD(dword1, src); + src += pitch; + + /* calculation of q3 and q4 */ + tmp0 = (q7_r - q2_r) + (q3_r - p4_r); + tmp2 = (q7_r - q3_r) + (q4_r - p3_r); + tmp1 += tmp0; + p0_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4); + tmp1 += tmp2; + p1_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4); + PCKEV_B2_UB(zero, p0_filter16, zero, p1_filter16, p0_filter16, + p1_filter16); + p0_filter16 = __msa_bmnz_v(q3, p0_filter16, flat2); + p1_filter16 = __msa_bmnz_v(q4, p1_filter16, flat2); + dword0 = __msa_copy_u_d((v2i64)p0_filter16, 0); + dword1 = __msa_copy_u_d((v2i64)p1_filter16, 0); + SD(dword0, src); + src += pitch; + SD(dword1, src); + src += pitch; + + /* calculation of q5 and q6 */ + tmp0 = (q7_r - q4_r) + (q5_r - p2_r); + tmp2 = (q7_r - q5_r) + (q6_r - p1_r); + tmp1 += tmp0; + p0_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4); + tmp1 += tmp2; + p1_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4); + PCKEV_B2_UB(zero, p0_filter16, zero, p1_filter16, p0_filter16, + p1_filter16); + p0_filter16 = __msa_bmnz_v(q5, p0_filter16, flat2); + p1_filter16 = __msa_bmnz_v(q6, p1_filter16, flat2); + dword0 = __msa_copy_u_d((v2i64)p0_filter16, 0); + dword1 = __msa_copy_u_d((v2i64)p1_filter16, 0); + SD(dword0, src); + src += pitch; + SD(dword1, src); + } + } + } else { + mb_lpf_horizontal_edge_dual(src, pitch, b_limit_ptr, limit_ptr, thresh_ptr, + count); + } +} + +void aom_lpf_horizontal_16_msa(uint8_t *src, int32_t pitch, + const uint8_t *b_limit_ptr, + const uint8_t *limit_ptr, + const uint8_t *thresh_ptr) { + mb_lpf_horizontal_edge(src, pitch, b_limit_ptr, limit_ptr, thresh_ptr, 1); +} + +void aom_lpf_horizontal_16_dual_msa(uint8_t *src, int32_t pitch, + const uint8_t *b_limit_ptr, + const uint8_t *limit_ptr, + const uint8_t *thresh_ptr) { + mb_lpf_horizontal_edge(src, pitch, b_limit_ptr, limit_ptr, thresh_ptr, 2); +} + +static void transpose_16x8_to_8x16(uint8_t *input, int32_t in_pitch, + uint8_t *output, int32_t out_pitch) { + v16u8 p7_org, p6_org, p5_org, p4_org, p3_org, p2_org, p1_org, p0_org; + v16i8 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7; + v16u8 p7, p6, p5, p4, p3, p2, p1, p0, q0, q1, q2, q3, q4, q5, q6, q7; + + LD_UB8(input, in_pitch, p7_org, p6_org, p5_org, p4_org, p3_org, p2_org, + p1_org, p0_org); + /* 8x8 transpose */ + TRANSPOSE8x8_UB_UB(p7_org, p6_org, p5_org, p4_org, p3_org, p2_org, p1_org, + p0_org, p7, p6, p5, p4, p3, p2, p1, p0); + /* 8x8 transpose */ + ILVL_B4_SB(p5_org, p7_org, p4_org, p6_org, p1_org, p3_org, p0_org, p2_org, + tmp0, tmp1, tmp2, tmp3); + ILVR_B2_SB(tmp1, tmp0, tmp3, tmp2, tmp4, tmp6); + ILVL_B2_SB(tmp1, tmp0, tmp3, tmp2, tmp5, tmp7); + ILVR_W2_UB(tmp6, tmp4, tmp7, tmp5, q0, q4); + ILVL_W2_UB(tmp6, tmp4, tmp7, tmp5, q2, q6); + SLDI_B4_0_UB(q0, q2, q4, q6, q1, q3, q5, q7, 8); + + ST_UB8(p7, p6, p5, p4, p3, p2, p1, p0, output, out_pitch); + output += (8 * out_pitch); + ST_UB8(q0, q1, q2, q3, q4, q5, q6, q7, output, out_pitch); +} + +static void transpose_8x16_to_16x8(uint8_t *input, int32_t in_pitch, + uint8_t *output, int32_t out_pitch) { + v16u8 p7_o, p6_o, p5_o, p4_o, p3_o, p2_o, p1_o, p0_o; + v16u8 p7, p6, p5, p4, p3, p2, p1, p0, q0, q1, q2, q3, q4, q5, q6, q7; + + LD_UB8(input, in_pitch, p7, p6, p5, p4, p3, p2, p1, p0); + LD_UB8(input + (8 * in_pitch), in_pitch, q0, q1, q2, q3, q4, q5, q6, q7); + TRANSPOSE16x8_UB_UB(p7, p6, p5, p4, p3, p2, p1, p0, q0, q1, q2, q3, q4, q5, + q6, q7, p7_o, p6_o, p5_o, p4_o, p3_o, p2_o, p1_o, p0_o); + ST_UB8(p7_o, p6_o, p5_o, p4_o, p3_o, p2_o, p1_o, p0_o, output, out_pitch); +} + +static void transpose_16x16(uint8_t *input, int32_t in_pitch, uint8_t *output, + int32_t out_pitch) { + v16u8 row0, row1, row2, row3, row4, row5, row6, row7; + v16u8 row8, row9, row10, row11, row12, row13, row14, row15; + v16u8 p7, p6, p5, p4, p3, p2, p1, p0, q0, q1, q2, q3, q4, q5, q6, q7; + v8i16 tmp0, tmp1, tmp4, tmp5, tmp6, tmp7; + v4i32 tmp2, tmp3; + + LD_UB8(input, in_pitch, row0, row1, row2, row3, row4, row5, row6, row7); + input += (8 * in_pitch); + LD_UB8(input, in_pitch, row8, row9, row10, row11, row12, row13, row14, row15); + + TRANSPOSE16x8_UB_UB(row0, row1, row2, row3, row4, row5, row6, row7, row8, + row9, row10, row11, row12, row13, row14, row15, p7, p6, + p5, p4, p3, p2, p1, p0); + + /* transpose 16x8 matrix into 8x16 */ + /* total 8 intermediate register and 32 instructions */ + q7 = (v16u8)__msa_ilvod_d((v2i64)row8, (v2i64)row0); + q6 = (v16u8)__msa_ilvod_d((v2i64)row9, (v2i64)row1); + q5 = (v16u8)__msa_ilvod_d((v2i64)row10, (v2i64)row2); + q4 = (v16u8)__msa_ilvod_d((v2i64)row11, (v2i64)row3); + q3 = (v16u8)__msa_ilvod_d((v2i64)row12, (v2i64)row4); + q2 = (v16u8)__msa_ilvod_d((v2i64)row13, (v2i64)row5); + q1 = (v16u8)__msa_ilvod_d((v2i64)row14, (v2i64)row6); + q0 = (v16u8)__msa_ilvod_d((v2i64)row15, (v2i64)row7); + + ILVEV_B2_SH(q7, q6, q5, q4, tmp0, tmp1); + tmp4 = (v8i16)__msa_ilvod_b((v16i8)q6, (v16i8)q7); + tmp5 = (v8i16)__msa_ilvod_b((v16i8)q4, (v16i8)q5); + + ILVEV_B2_UB(q3, q2, q1, q0, q5, q7); + tmp6 = (v8i16)__msa_ilvod_b((v16i8)q2, (v16i8)q3); + tmp7 = (v8i16)__msa_ilvod_b((v16i8)q0, (v16i8)q1); + + ILVEV_H2_SW(tmp0, tmp1, q5, q7, tmp2, tmp3); + q0 = (v16u8)__msa_ilvev_w(tmp3, tmp2); + q4 = (v16u8)__msa_ilvod_w(tmp3, tmp2); + + tmp2 = (v4i32)__msa_ilvod_h(tmp1, tmp0); + tmp3 = (v4i32)__msa_ilvod_h((v8i16)q7, (v8i16)q5); + q2 = (v16u8)__msa_ilvev_w(tmp3, tmp2); + q6 = (v16u8)__msa_ilvod_w(tmp3, tmp2); + + ILVEV_H2_SW(tmp4, tmp5, tmp6, tmp7, tmp2, tmp3); + q1 = (v16u8)__msa_ilvev_w(tmp3, tmp2); + q5 = (v16u8)__msa_ilvod_w(tmp3, tmp2); + + tmp2 = (v4i32)__msa_ilvod_h(tmp5, tmp4); + tmp3 = (v4i32)__msa_ilvod_h(tmp7, tmp6); + q3 = (v16u8)__msa_ilvev_w(tmp3, tmp2); + q7 = (v16u8)__msa_ilvod_w(tmp3, tmp2); + + ST_UB8(p7, p6, p5, p4, p3, p2, p1, p0, output, out_pitch); + output += (8 * out_pitch); + ST_UB8(q0, q1, q2, q3, q4, q5, q6, q7, output, out_pitch); +} + +int32_t aom_vt_lpf_t4_and_t8_8w(uint8_t *src, uint8_t *filter48, + uint8_t *src_org, int32_t pitch_org, + const uint8_t *b_limit_ptr, + const uint8_t *limit_ptr, + const uint8_t *thresh_ptr) { + v16u8 p3, p2, p1, p0, q3, q2, q1, q0; + v16u8 p2_out, p1_out, p0_out, q0_out, q1_out, q2_out; + v16u8 flat, mask, hev, thresh, b_limit, limit; + v8u16 p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r; + v8i16 p2_filt8_r, p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r; + v16i8 zero = { 0 }; + v8i16 vec0, vec1, vec2, vec3; + + /* load vector elements */ + LD_UB8(src - (4 * 16), 16, p3, p2, p1, p0, q0, q1, q2, q3); + + thresh = (v16u8)__msa_fill_b(*thresh_ptr); + b_limit = (v16u8)__msa_fill_b(*b_limit_ptr); + limit = (v16u8)__msa_fill_b(*limit_ptr); + + /* mask and hev */ + LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh, hev, + mask, flat); + /* flat4 */ + AOM_FLAT4(p3, p2, p0, q0, q2, q3, flat); + /* filter4 */ + AOM_LPF_FILTER4_8W(p1, p0, q0, q1, mask, hev, p1_out, p0_out, q0_out, q1_out); + + flat = (v16u8)__msa_ilvr_d((v2i64)zero, (v2i64)flat); + + if (__msa_test_bz_v(flat)) { + ILVR_B2_SH(p0_out, p1_out, q1_out, q0_out, vec0, vec1); + ILVRL_H2_SH(vec1, vec0, vec2, vec3); + ST4x8_UB(vec2, vec3, (src_org - 2), pitch_org); + return 1; + } else { + ILVR_B8_UH(zero, p3, zero, p2, zero, p1, zero, p0, zero, q0, zero, q1, zero, + q2, zero, q3, p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r); + AOM_FILTER8(p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r, p2_filt8_r, + p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r); + + /* convert 16 bit output data into 8 bit */ + p2_r = (v8u16)__msa_pckev_b((v16i8)p2_filt8_r, (v16i8)p2_filt8_r); + p1_r = (v8u16)__msa_pckev_b((v16i8)p1_filt8_r, (v16i8)p1_filt8_r); + p0_r = (v8u16)__msa_pckev_b((v16i8)p0_filt8_r, (v16i8)p0_filt8_r); + q0_r = (v8u16)__msa_pckev_b((v16i8)q0_filt8_r, (v16i8)q0_filt8_r); + q1_r = (v8u16)__msa_pckev_b((v16i8)q1_filt8_r, (v16i8)q1_filt8_r); + q2_r = (v8u16)__msa_pckev_b((v16i8)q2_filt8_r, (v16i8)q2_filt8_r); + + /* store pixel values */ + p2_out = __msa_bmnz_v(p2, (v16u8)p2_r, flat); + p1_out = __msa_bmnz_v(p1_out, (v16u8)p1_r, flat); + p0_out = __msa_bmnz_v(p0_out, (v16u8)p0_r, flat); + q0_out = __msa_bmnz_v(q0_out, (v16u8)q0_r, flat); + q1_out = __msa_bmnz_v(q1_out, (v16u8)q1_r, flat); + q2_out = __msa_bmnz_v(q2, (v16u8)q2_r, flat); + + ST_UB4(p2_out, p1_out, p0_out, q0_out, filter48, 16); + filter48 += (4 * 16); + ST_UB2(q1_out, q2_out, filter48, 16); + filter48 += (2 * 16); + ST_UB(flat, filter48); + + return 0; + } +} + +int32_t aom_vt_lpf_t16_8w(uint8_t *src, uint8_t *src_org, int32_t pitch, + uint8_t *filter48) { + v16i8 zero = { 0 }; + v16u8 filter8, flat, flat2; + v16u8 p7, p6, p5, p4, p3, p2, p1, p0, q0, q1, q2, q3, q4, q5, q6, q7; + v8u16 p7_r_in, p6_r_in, p5_r_in, p4_r_in, p3_r_in, p2_r_in, p1_r_in, p0_r_in; + v8u16 q7_r_in, q6_r_in, q5_r_in, q4_r_in, q3_r_in, q2_r_in, q1_r_in, q0_r_in; + v8u16 tmp0_r, tmp1_r; + v8i16 r_out; + + flat = LD_UB(filter48 + 6 * 16); + + LD_UB8((src - 8 * 16), 16, p7, p6, p5, p4, p3, p2, p1, p0); + LD_UB8(src, 16, q0, q1, q2, q3, q4, q5, q6, q7); + + AOM_FLAT5(p7, p6, p5, p4, p0, q0, q4, q5, q6, q7, flat, flat2); + + if (__msa_test_bz_v(flat2)) { + v8i16 vec0, vec1, vec2, vec3, vec4; + + LD_UB4(filter48, 16, p2, p1, p0, q0); + LD_UB2(filter48 + 4 * 16, 16, q1, q2); + + ILVR_B2_SH(p1, p2, q0, p0, vec0, vec1); + ILVRL_H2_SH(vec1, vec0, vec3, vec4); + vec2 = (v8i16)__msa_ilvr_b((v16i8)q2, (v16i8)q1); + + src_org -= 3; + ST4x4_UB(vec3, vec3, 0, 1, 2, 3, src_org, pitch); + ST2x4_UB(vec2, 0, (src_org + 4), pitch); + src_org += (4 * pitch); + ST4x4_UB(vec4, vec4, 0, 1, 2, 3, src_org, pitch); + ST2x4_UB(vec2, 4, (src_org + 4), pitch); + + return 1; + } else { + src -= 7 * 16; + + ILVR_B8_UH(zero, p7, zero, p6, zero, p5, zero, p4, zero, p3, zero, p2, zero, + p1, zero, p0, p7_r_in, p6_r_in, p5_r_in, p4_r_in, p3_r_in, + p2_r_in, p1_r_in, p0_r_in); + q0_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q0); + + tmp0_r = p7_r_in << 3; + tmp0_r -= p7_r_in; + tmp0_r += p6_r_in; + tmp0_r += q0_r_in; + tmp1_r = p6_r_in + p5_r_in; + tmp1_r += p4_r_in; + tmp1_r += p3_r_in; + tmp1_r += p2_r_in; + tmp1_r += p1_r_in; + tmp1_r += p0_r_in; + tmp1_r += tmp0_r; + + r_out = __msa_srari_h((v8i16)tmp1_r, 4); + r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out); + p6 = __msa_bmnz_v(p6, (v16u8)r_out, flat2); + ST8x1_UB(p6, src); + src += 16; + + /* p5 */ + q1_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q1); + tmp0_r = p5_r_in - p6_r_in; + tmp0_r += q1_r_in; + tmp0_r -= p7_r_in; + tmp1_r += tmp0_r; + r_out = __msa_srari_h((v8i16)tmp1_r, 4); + r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out); + p5 = __msa_bmnz_v(p5, (v16u8)r_out, flat2); + ST8x1_UB(p5, src); + src += 16; + + /* p4 */ + q2_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q2); + tmp0_r = p4_r_in - p5_r_in; + tmp0_r += q2_r_in; + tmp0_r -= p7_r_in; + tmp1_r += tmp0_r; + r_out = __msa_srari_h((v8i16)tmp1_r, 4); + r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out); + p4 = __msa_bmnz_v(p4, (v16u8)r_out, flat2); + ST8x1_UB(p4, src); + src += 16; + + /* p3 */ + q3_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q3); + tmp0_r = p3_r_in - p4_r_in; + tmp0_r += q3_r_in; + tmp0_r -= p7_r_in; + tmp1_r += tmp0_r; + r_out = __msa_srari_h((v8i16)tmp1_r, 4); + r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out); + p3 = __msa_bmnz_v(p3, (v16u8)r_out, flat2); + ST8x1_UB(p3, src); + src += 16; + + /* p2 */ + q4_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q4); + filter8 = LD_UB(filter48); + tmp0_r = p2_r_in - p3_r_in; + tmp0_r += q4_r_in; + tmp0_r -= p7_r_in; + tmp1_r += tmp0_r; + r_out = __msa_srari_h((v8i16)tmp1_r, 4); + r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out); + filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2); + ST8x1_UB(filter8, src); + src += 16; + + /* p1 */ + q5_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q5); + filter8 = LD_UB(filter48 + 16); + tmp0_r = p1_r_in - p2_r_in; + tmp0_r += q5_r_in; + tmp0_r -= p7_r_in; + tmp1_r += tmp0_r; + r_out = __msa_srari_h((v8i16)tmp1_r, 4); + r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out); + filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2); + ST8x1_UB(filter8, src); + src += 16; + + /* p0 */ + q6_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q6); + filter8 = LD_UB(filter48 + 32); + tmp0_r = p0_r_in - p1_r_in; + tmp0_r += q6_r_in; + tmp0_r -= p7_r_in; + tmp1_r += tmp0_r; + r_out = __msa_srari_h((v8i16)tmp1_r, 4); + r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out); + filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2); + ST8x1_UB(filter8, src); + src += 16; + + /* q0 */ + q7_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q7); + filter8 = LD_UB(filter48 + 48); + tmp0_r = q7_r_in - p0_r_in; + tmp0_r += q0_r_in; + tmp0_r -= p7_r_in; + tmp1_r += tmp0_r; + r_out = __msa_srari_h((v8i16)tmp1_r, 4); + r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out); + filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2); + ST8x1_UB(filter8, src); + src += 16; + + /* q1 */ + filter8 = LD_UB(filter48 + 64); + tmp0_r = q7_r_in - q0_r_in; + tmp0_r += q1_r_in; + tmp0_r -= p6_r_in; + tmp1_r += tmp0_r; + r_out = __msa_srari_h((v8i16)tmp1_r, 4); + r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out); + filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2); + ST8x1_UB(filter8, src); + src += 16; + + /* q2 */ + filter8 = LD_UB(filter48 + 80); + tmp0_r = q7_r_in - q1_r_in; + tmp0_r += q2_r_in; + tmp0_r -= p5_r_in; + tmp1_r += tmp0_r; + r_out = __msa_srari_h((v8i16)tmp1_r, 4); + r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out); + filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2); + ST8x1_UB(filter8, src); + src += 16; + + /* q3 */ + tmp0_r = q7_r_in - q2_r_in; + tmp0_r += q3_r_in; + tmp0_r -= p4_r_in; + tmp1_r += tmp0_r; + r_out = __msa_srari_h((v8i16)tmp1_r, 4); + r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out); + q3 = __msa_bmnz_v(q3, (v16u8)r_out, flat2); + ST8x1_UB(q3, src); + src += 16; + + /* q4 */ + tmp0_r = q7_r_in - q3_r_in; + tmp0_r += q4_r_in; + tmp0_r -= p3_r_in; + tmp1_r += tmp0_r; + r_out = __msa_srari_h((v8i16)tmp1_r, 4); + r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out); + q4 = __msa_bmnz_v(q4, (v16u8)r_out, flat2); + ST8x1_UB(q4, src); + src += 16; + + /* q5 */ + tmp0_r = q7_r_in - q4_r_in; + tmp0_r += q5_r_in; + tmp0_r -= p2_r_in; + tmp1_r += tmp0_r; + r_out = __msa_srari_h((v8i16)tmp1_r, 4); + r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out); + q5 = __msa_bmnz_v(q5, (v16u8)r_out, flat2); + ST8x1_UB(q5, src); + src += 16; + + /* q6 */ + tmp0_r = q7_r_in - q5_r_in; + tmp0_r += q6_r_in; + tmp0_r -= p1_r_in; + tmp1_r += tmp0_r; + r_out = __msa_srari_h((v8i16)tmp1_r, 4); + r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out); + q6 = __msa_bmnz_v(q6, (v16u8)r_out, flat2); + ST8x1_UB(q6, src); + + return 0; + } +} + +void aom_lpf_vertical_16_msa(uint8_t *src, int32_t pitch, + const uint8_t *b_limit_ptr, + const uint8_t *limit_ptr, + const uint8_t *thresh_ptr) { + uint8_t early_exit = 0; + DECLARE_ALIGNED(32, uint8_t, transposed_input[16 * 24]); + uint8_t *filter48 = &transposed_input[16 * 16]; + + transpose_16x8_to_8x16(src - 8, pitch, transposed_input, 16); + + early_exit = + aom_vt_lpf_t4_and_t8_8w((transposed_input + 16 * 8), &filter48[0], src, + pitch, b_limit_ptr, limit_ptr, thresh_ptr); + + if (0 == early_exit) { + early_exit = aom_vt_lpf_t16_8w((transposed_input + 16 * 8), src, pitch, + &filter48[0]); + + if (0 == early_exit) { + transpose_8x16_to_16x8(transposed_input, 16, src - 8, pitch); + } + } +} + +int32_t aom_vt_lpf_t4_and_t8_16w(uint8_t *src, uint8_t *filter48, + uint8_t *src_org, int32_t pitch, + const uint8_t *b_limit_ptr, + const uint8_t *limit_ptr, + const uint8_t *thresh_ptr) { + v16u8 p3, p2, p1, p0, q3, q2, q1, q0; + v16u8 p2_out, p1_out, p0_out, q0_out, q1_out, q2_out; + v16u8 flat, mask, hev, thresh, b_limit, limit; + v8u16 p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r; + v8u16 p3_l, p2_l, p1_l, p0_l, q0_l, q1_l, q2_l, q3_l; + v8i16 p2_filt8_r, p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r; + v8i16 p2_filt8_l, p1_filt8_l, p0_filt8_l, q0_filt8_l, q1_filt8_l, q2_filt8_l; + v16i8 zero = { 0 }; + v8i16 vec0, vec1, vec2, vec3, vec4, vec5; + + /* load vector elements */ + LD_UB8(src - (4 * 16), 16, p3, p2, p1, p0, q0, q1, q2, q3); + + thresh = (v16u8)__msa_fill_b(*thresh_ptr); + b_limit = (v16u8)__msa_fill_b(*b_limit_ptr); + limit = (v16u8)__msa_fill_b(*limit_ptr); + + /* mask and hev */ + LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh, hev, + mask, flat); + /* flat4 */ + AOM_FLAT4(p3, p2, p0, q0, q2, q3, flat); + /* filter4 */ + AOM_LPF_FILTER4_4W(p1, p0, q0, q1, mask, hev, p1_out, p0_out, q0_out, q1_out); + + if (__msa_test_bz_v(flat)) { + ILVR_B2_SH(p0_out, p1_out, q1_out, q0_out, vec0, vec1); + ILVRL_H2_SH(vec1, vec0, vec2, vec3); + ILVL_B2_SH(p0_out, p1_out, q1_out, q0_out, vec0, vec1); + ILVRL_H2_SH(vec1, vec0, vec4, vec5); + + src_org -= 2; + ST4x8_UB(vec2, vec3, src_org, pitch); + src_org += 8 * pitch; + ST4x8_UB(vec4, vec5, src_org, pitch); + + return 1; + } else { + ILVR_B8_UH(zero, p3, zero, p2, zero, p1, zero, p0, zero, q0, zero, q1, zero, + q2, zero, q3, p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r); + AOM_FILTER8(p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r, p2_filt8_r, + p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r); + ILVL_B4_UH(zero, p3, zero, p2, zero, p1, zero, p0, p3_l, p2_l, p1_l, p0_l); + ILVL_B4_UH(zero, q0, zero, q1, zero, q2, zero, q3, q0_l, q1_l, q2_l, q3_l); + AOM_FILTER8(p3_l, p2_l, p1_l, p0_l, q0_l, q1_l, q2_l, q3_l, p2_filt8_l, + p1_filt8_l, p0_filt8_l, q0_filt8_l, q1_filt8_l, q2_filt8_l); + + /* convert 16 bit output data into 8 bit */ + PCKEV_B4_SH(p2_filt8_l, p2_filt8_r, p1_filt8_l, p1_filt8_r, p0_filt8_l, + p0_filt8_r, q0_filt8_l, q0_filt8_r, p2_filt8_r, p1_filt8_r, + p0_filt8_r, q0_filt8_r); + PCKEV_B2_SH(q1_filt8_l, q1_filt8_r, q2_filt8_l, q2_filt8_r, q1_filt8_r, + q2_filt8_r); + + /* store pixel values */ + p2_out = __msa_bmnz_v(p2, (v16u8)p2_filt8_r, flat); + p1_out = __msa_bmnz_v(p1_out, (v16u8)p1_filt8_r, flat); + p0_out = __msa_bmnz_v(p0_out, (v16u8)p0_filt8_r, flat); + q0_out = __msa_bmnz_v(q0_out, (v16u8)q0_filt8_r, flat); + q1_out = __msa_bmnz_v(q1_out, (v16u8)q1_filt8_r, flat); + q2_out = __msa_bmnz_v(q2, (v16u8)q2_filt8_r, flat); + + ST_UB4(p2_out, p1_out, p0_out, q0_out, filter48, 16); + filter48 += (4 * 16); + ST_UB2(q1_out, q2_out, filter48, 16); + filter48 += (2 * 16); + ST_UB(flat, filter48); + + return 0; + } +} + +int32_t aom_vt_lpf_t16_16w(uint8_t *src, uint8_t *src_org, int32_t pitch, + uint8_t *filter48) { + v16u8 flat, flat2, filter8; + v16i8 zero = { 0 }; + v16u8 p7, p6, p5, p4, p3, p2, p1, p0, q0, q1, q2, q3, q4, q5, q6, q7; + v8u16 p7_r_in, p6_r_in, p5_r_in, p4_r_in, p3_r_in, p2_r_in, p1_r_in, p0_r_in; + v8u16 q7_r_in, q6_r_in, q5_r_in, q4_r_in, q3_r_in, q2_r_in, q1_r_in, q0_r_in; + v8u16 p7_l_in, p6_l_in, p5_l_in, p4_l_in, p3_l_in, p2_l_in, p1_l_in, p0_l_in; + v8u16 q7_l_in, q6_l_in, q5_l_in, q4_l_in, q3_l_in, q2_l_in, q1_l_in, q0_l_in; + v8u16 tmp0_r, tmp1_r, tmp0_l, tmp1_l; + v8i16 l_out, r_out; + + flat = LD_UB(filter48 + 6 * 16); + + LD_UB8((src - 8 * 16), 16, p7, p6, p5, p4, p3, p2, p1, p0); + LD_UB8(src, 16, q0, q1, q2, q3, q4, q5, q6, q7); + + AOM_FLAT5(p7, p6, p5, p4, p0, q0, q4, q5, q6, q7, flat, flat2); + + if (__msa_test_bz_v(flat2)) { + v8i16 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7; + + LD_UB4(filter48, 16, p2, p1, p0, q0); + LD_UB2(filter48 + 4 * 16, 16, q1, q2); + + ILVR_B2_SH(p1, p2, q0, p0, vec0, vec1); + ILVRL_H2_SH(vec1, vec0, vec3, vec4); + ILVL_B2_SH(p1, p2, q0, p0, vec0, vec1); + ILVRL_H2_SH(vec1, vec0, vec6, vec7); + ILVRL_B2_SH(q2, q1, vec2, vec5); + + src_org -= 3; + ST4x4_UB(vec3, vec3, 0, 1, 2, 3, src_org, pitch); + ST2x4_UB(vec2, 0, (src_org + 4), pitch); + src_org += (4 * pitch); + ST4x4_UB(vec4, vec4, 0, 1, 2, 3, src_org, pitch); + ST2x4_UB(vec2, 4, (src_org + 4), pitch); + src_org += (4 * pitch); + ST4x4_UB(vec6, vec6, 0, 1, 2, 3, src_org, pitch); + ST2x4_UB(vec5, 0, (src_org + 4), pitch); + src_org += (4 * pitch); + ST4x4_UB(vec7, vec7, 0, 1, 2, 3, src_org, pitch); + ST2x4_UB(vec5, 4, (src_org + 4), pitch); + + return 1; + } else { + src -= 7 * 16; + + ILVR_B8_UH(zero, p7, zero, p6, zero, p5, zero, p4, zero, p3, zero, p2, zero, + p1, zero, p0, p7_r_in, p6_r_in, p5_r_in, p4_r_in, p3_r_in, + p2_r_in, p1_r_in, p0_r_in); + q0_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q0); + + tmp0_r = p7_r_in << 3; + tmp0_r -= p7_r_in; + tmp0_r += p6_r_in; + tmp0_r += q0_r_in; + tmp1_r = p6_r_in + p5_r_in; + tmp1_r += p4_r_in; + tmp1_r += p3_r_in; + tmp1_r += p2_r_in; + tmp1_r += p1_r_in; + tmp1_r += p0_r_in; + tmp1_r += tmp0_r; + r_out = __msa_srari_h((v8i16)tmp1_r, 4); + + ILVL_B4_UH(zero, p7, zero, p6, zero, p5, zero, p4, p7_l_in, p6_l_in, + p5_l_in, p4_l_in); + ILVL_B4_UH(zero, p3, zero, p2, zero, p1, zero, p0, p3_l_in, p2_l_in, + p1_l_in, p0_l_in); + q0_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q0); + + tmp0_l = p7_l_in << 3; + tmp0_l -= p7_l_in; + tmp0_l += p6_l_in; + tmp0_l += q0_l_in; + tmp1_l = p6_l_in + p5_l_in; + tmp1_l += p4_l_in; + tmp1_l += p3_l_in; + tmp1_l += p2_l_in; + tmp1_l += p1_l_in; + tmp1_l += p0_l_in; + tmp1_l += tmp0_l; + l_out = __msa_srari_h((v8i16)tmp1_l, 4); + + r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out); + p6 = __msa_bmnz_v(p6, (v16u8)r_out, flat2); + ST_UB(p6, src); + src += 16; + + /* p5 */ + q1_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q1); + tmp0_r = p5_r_in - p6_r_in; + tmp0_r += q1_r_in; + tmp0_r -= p7_r_in; + tmp1_r += tmp0_r; + r_out = __msa_srari_h((v8i16)tmp1_r, 4); + q1_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q1); + tmp0_l = p5_l_in - p6_l_in; + tmp0_l += q1_l_in; + tmp0_l -= p7_l_in; + tmp1_l += tmp0_l; + l_out = __msa_srari_h((v8i16)tmp1_l, 4); + r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out); + p5 = __msa_bmnz_v(p5, (v16u8)r_out, flat2); + ST_UB(p5, src); + src += 16; + + /* p4 */ + q2_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q2); + tmp0_r = p4_r_in - p5_r_in; + tmp0_r += q2_r_in; + tmp0_r -= p7_r_in; + tmp1_r += tmp0_r; + r_out = __msa_srari_h((v8i16)tmp1_r, 4); + q2_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q2); + tmp0_l = p4_l_in - p5_l_in; + tmp0_l += q2_l_in; + tmp0_l -= p7_l_in; + tmp1_l += tmp0_l; + l_out = __msa_srari_h((v8i16)tmp1_l, 4); + r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out); + p4 = __msa_bmnz_v(p4, (v16u8)r_out, flat2); + ST_UB(p4, src); + src += 16; + + /* p3 */ + q3_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q3); + tmp0_r = p3_r_in - p4_r_in; + tmp0_r += q3_r_in; + tmp0_r -= p7_r_in; + tmp1_r += tmp0_r; + r_out = __msa_srari_h((v8i16)tmp1_r, 4); + q3_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q3); + tmp0_l = p3_l_in - p4_l_in; + tmp0_l += q3_l_in; + tmp0_l -= p7_l_in; + tmp1_l += tmp0_l; + l_out = __msa_srari_h((v8i16)tmp1_l, 4); + r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out); + p3 = __msa_bmnz_v(p3, (v16u8)r_out, flat2); + ST_UB(p3, src); + src += 16; + + /* p2 */ + q4_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q4); + filter8 = LD_UB(filter48); + tmp0_r = p2_r_in - p3_r_in; + tmp0_r += q4_r_in; + tmp0_r -= p7_r_in; + tmp1_r += tmp0_r; + r_out = __msa_srari_h((v8i16)tmp1_r, 4); + q4_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q4); + tmp0_l = p2_l_in - p3_l_in; + tmp0_l += q4_l_in; + tmp0_l -= p7_l_in; + tmp1_l += tmp0_l; + l_out = __msa_srari_h((v8i16)tmp1_l, 4); + r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out); + filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2); + ST_UB(filter8, src); + src += 16; + + /* p1 */ + q5_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q5); + filter8 = LD_UB(filter48 + 16); + tmp0_r = p1_r_in - p2_r_in; + tmp0_r += q5_r_in; + tmp0_r -= p7_r_in; + tmp1_r += tmp0_r; + r_out = __msa_srari_h((v8i16)tmp1_r, 4); + q5_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q5); + tmp0_l = p1_l_in - p2_l_in; + tmp0_l += q5_l_in; + tmp0_l -= p7_l_in; + tmp1_l += tmp0_l; + l_out = __msa_srari_h((v8i16)(tmp1_l), 4); + r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out); + filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2); + ST_UB(filter8, src); + src += 16; + + /* p0 */ + q6_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q6); + filter8 = LD_UB(filter48 + 32); + tmp0_r = p0_r_in - p1_r_in; + tmp0_r += q6_r_in; + tmp0_r -= p7_r_in; + tmp1_r += tmp0_r; + r_out = __msa_srari_h((v8i16)tmp1_r, 4); + q6_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q6); + tmp0_l = p0_l_in - p1_l_in; + tmp0_l += q6_l_in; + tmp0_l -= p7_l_in; + tmp1_l += tmp0_l; + l_out = __msa_srari_h((v8i16)tmp1_l, 4); + r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out); + filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2); + ST_UB(filter8, src); + src += 16; + + /* q0 */ + q7_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q7); + filter8 = LD_UB(filter48 + 48); + tmp0_r = q7_r_in - p0_r_in; + tmp0_r += q0_r_in; + tmp0_r -= p7_r_in; + tmp1_r += tmp0_r; + r_out = __msa_srari_h((v8i16)tmp1_r, 4); + q7_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q7); + tmp0_l = q7_l_in - p0_l_in; + tmp0_l += q0_l_in; + tmp0_l -= p7_l_in; + tmp1_l += tmp0_l; + l_out = __msa_srari_h((v8i16)tmp1_l, 4); + r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out); + filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2); + ST_UB(filter8, src); + src += 16; + + /* q1 */ + filter8 = LD_UB(filter48 + 64); + tmp0_r = q7_r_in - q0_r_in; + tmp0_r += q1_r_in; + tmp0_r -= p6_r_in; + tmp1_r += tmp0_r; + r_out = __msa_srari_h((v8i16)tmp1_r, 4); + tmp0_l = q7_l_in - q0_l_in; + tmp0_l += q1_l_in; + tmp0_l -= p6_l_in; + tmp1_l += tmp0_l; + l_out = __msa_srari_h((v8i16)tmp1_l, 4); + r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out); + filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2); + ST_UB(filter8, src); + src += 16; + + /* q2 */ + filter8 = LD_UB(filter48 + 80); + tmp0_r = q7_r_in - q1_r_in; + tmp0_r += q2_r_in; + tmp0_r -= p5_r_in; + tmp1_r += tmp0_r; + r_out = __msa_srari_h((v8i16)tmp1_r, 4); + tmp0_l = q7_l_in - q1_l_in; + tmp0_l += q2_l_in; + tmp0_l -= p5_l_in; + tmp1_l += tmp0_l; + l_out = __msa_srari_h((v8i16)tmp1_l, 4); + r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out); + filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2); + ST_UB(filter8, src); + src += 16; + + /* q3 */ + tmp0_r = q7_r_in - q2_r_in; + tmp0_r += q3_r_in; + tmp0_r -= p4_r_in; + tmp1_r += tmp0_r; + r_out = __msa_srari_h((v8i16)tmp1_r, 4); + tmp0_l = q7_l_in - q2_l_in; + tmp0_l += q3_l_in; + tmp0_l -= p4_l_in; + tmp1_l += tmp0_l; + l_out = __msa_srari_h((v8i16)tmp1_l, 4); + r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out); + q3 = __msa_bmnz_v(q3, (v16u8)r_out, flat2); + ST_UB(q3, src); + src += 16; + + /* q4 */ + tmp0_r = q7_r_in - q3_r_in; + tmp0_r += q4_r_in; + tmp0_r -= p3_r_in; + tmp1_r += tmp0_r; + r_out = __msa_srari_h((v8i16)tmp1_r, 4); + tmp0_l = q7_l_in - q3_l_in; + tmp0_l += q4_l_in; + tmp0_l -= p3_l_in; + tmp1_l += tmp0_l; + l_out = __msa_srari_h((v8i16)tmp1_l, 4); + r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out); + q4 = __msa_bmnz_v(q4, (v16u8)r_out, flat2); + ST_UB(q4, src); + src += 16; + + /* q5 */ + tmp0_r = q7_r_in - q4_r_in; + tmp0_r += q5_r_in; + tmp0_r -= p2_r_in; + tmp1_r += tmp0_r; + r_out = __msa_srari_h((v8i16)tmp1_r, 4); + tmp0_l = q7_l_in - q4_l_in; + tmp0_l += q5_l_in; + tmp0_l -= p2_l_in; + tmp1_l += tmp0_l; + l_out = __msa_srari_h((v8i16)tmp1_l, 4); + r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out); + q5 = __msa_bmnz_v(q5, (v16u8)r_out, flat2); + ST_UB(q5, src); + src += 16; + + /* q6 */ + tmp0_r = q7_r_in - q5_r_in; + tmp0_r += q6_r_in; + tmp0_r -= p1_r_in; + tmp1_r += tmp0_r; + r_out = __msa_srari_h((v8i16)tmp1_r, 4); + tmp0_l = q7_l_in - q5_l_in; + tmp0_l += q6_l_in; + tmp0_l -= p1_l_in; + tmp1_l += tmp0_l; + l_out = __msa_srari_h((v8i16)tmp1_l, 4); + r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out); + q6 = __msa_bmnz_v(q6, (v16u8)r_out, flat2); + ST_UB(q6, src); + + return 0; + } +} + +void aom_lpf_vertical_16_dual_msa(uint8_t *src, int32_t pitch, + const uint8_t *b_limit_ptr, + const uint8_t *limit_ptr, + const uint8_t *thresh_ptr) { + uint8_t early_exit = 0; + DECLARE_ALIGNED(32, uint8_t, transposed_input[16 * 24]); + uint8_t *filter48 = &transposed_input[16 * 16]; + + transpose_16x16((src - 8), pitch, &transposed_input[0], 16); + + early_exit = + aom_vt_lpf_t4_and_t8_16w((transposed_input + 16 * 8), &filter48[0], src, + pitch, b_limit_ptr, limit_ptr, thresh_ptr); + + if (0 == early_exit) { + early_exit = aom_vt_lpf_t16_16w((transposed_input + 16 * 8), src, pitch, + &filter48[0]); + + if (0 == early_exit) { + transpose_16x16(transposed_input, 16, (src - 8), pitch); + } + } +} diff --git a/third_party/aom/aom_dsp/mips/loopfilter_4_msa.c b/third_party/aom/aom_dsp/mips/loopfilter_4_msa.c new file mode 100644 index 0000000000..dc0a977645 --- /dev/null +++ b/third_party/aom/aom_dsp/mips/loopfilter_4_msa.c @@ -0,0 +1,147 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include "aom_dsp/mips/loopfilter_msa.h" + +void aom_lpf_horizontal_4_msa(uint8_t *src, int32_t pitch, + const uint8_t *b_limit_ptr, + const uint8_t *limit_ptr, + const uint8_t *thresh_ptr) { + uint64_t p1_d, p0_d, q0_d, q1_d; + v16u8 mask, hev, flat, thresh, b_limit, limit; + v16u8 p3, p2, p1, p0, q3, q2, q1, q0, p1_out, p0_out, q0_out, q1_out; + + /* load vector elements */ + LD_UB8((src - 4 * pitch), pitch, p3, p2, p1, p0, q0, q1, q2, q3); + + thresh = (v16u8)__msa_fill_b(*thresh_ptr); + b_limit = (v16u8)__msa_fill_b(*b_limit_ptr); + limit = (v16u8)__msa_fill_b(*limit_ptr); + + LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh, hev, + mask, flat); + AOM_LPF_FILTER4_8W(p1, p0, q0, q1, mask, hev, p1_out, p0_out, q0_out, q1_out); + + p1_d = __msa_copy_u_d((v2i64)p1_out, 0); + p0_d = __msa_copy_u_d((v2i64)p0_out, 0); + q0_d = __msa_copy_u_d((v2i64)q0_out, 0); + q1_d = __msa_copy_u_d((v2i64)q1_out, 0); + SD4(p1_d, p0_d, q0_d, q1_d, (src - 2 * pitch), pitch); +} + +void aom_lpf_horizontal_4_dual_msa(uint8_t *src, int32_t pitch, + const uint8_t *b_limit0_ptr, + const uint8_t *limit0_ptr, + const uint8_t *thresh0_ptr, + const uint8_t *b_limit1_ptr, + const uint8_t *limit1_ptr, + const uint8_t *thresh1_ptr) { + v16u8 mask, hev, flat, thresh0, b_limit0, limit0, thresh1, b_limit1, limit1; + v16u8 p3, p2, p1, p0, q3, q2, q1, q0; + + /* load vector elements */ + LD_UB8((src - 4 * pitch), pitch, p3, p2, p1, p0, q0, q1, q2, q3); + + thresh0 = (v16u8)__msa_fill_b(*thresh0_ptr); + thresh1 = (v16u8)__msa_fill_b(*thresh1_ptr); + thresh0 = (v16u8)__msa_ilvr_d((v2i64)thresh1, (v2i64)thresh0); + + b_limit0 = (v16u8)__msa_fill_b(*b_limit0_ptr); + b_limit1 = (v16u8)__msa_fill_b(*b_limit1_ptr); + b_limit0 = (v16u8)__msa_ilvr_d((v2i64)b_limit1, (v2i64)b_limit0); + + limit0 = (v16u8)__msa_fill_b(*limit0_ptr); + limit1 = (v16u8)__msa_fill_b(*limit1_ptr); + limit0 = (v16u8)__msa_ilvr_d((v2i64)limit1, (v2i64)limit0); + + LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit0, b_limit0, thresh0, hev, + mask, flat); + AOM_LPF_FILTER4_4W(p1, p0, q0, q1, mask, hev, p1, p0, q0, q1); + + ST_UB4(p1, p0, q0, q1, (src - 2 * pitch), pitch); +} + +void aom_lpf_vertical_4_msa(uint8_t *src, int32_t pitch, + const uint8_t *b_limit_ptr, + const uint8_t *limit_ptr, + const uint8_t *thresh_ptr) { + v16u8 mask, hev, flat, limit, thresh, b_limit; + v16u8 p3, p2, p1, p0, q3, q2, q1, q0; + v8i16 vec0, vec1, vec2, vec3; + + LD_UB8((src - 4), pitch, p3, p2, p1, p0, q0, q1, q2, q3); + + thresh = (v16u8)__msa_fill_b(*thresh_ptr); + b_limit = (v16u8)__msa_fill_b(*b_limit_ptr); + limit = (v16u8)__msa_fill_b(*limit_ptr); + + TRANSPOSE8x8_UB_UB(p3, p2, p1, p0, q0, q1, q2, q3, p3, p2, p1, p0, q0, q1, q2, + q3); + LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh, hev, + mask, flat); + AOM_LPF_FILTER4_8W(p1, p0, q0, q1, mask, hev, p1, p0, q0, q1); + ILVR_B2_SH(p0, p1, q1, q0, vec0, vec1); + ILVRL_H2_SH(vec1, vec0, vec2, vec3); + + src -= 2; + ST4x4_UB(vec2, vec2, 0, 1, 2, 3, src, pitch); + src += 4 * pitch; + ST4x4_UB(vec3, vec3, 0, 1, 2, 3, src, pitch); +} + +void aom_lpf_vertical_4_dual_msa(uint8_t *src, int32_t pitch, + const uint8_t *b_limit0_ptr, + const uint8_t *limit0_ptr, + const uint8_t *thresh0_ptr, + const uint8_t *b_limit1_ptr, + const uint8_t *limit1_ptr, + const uint8_t *thresh1_ptr) { + v16u8 mask, hev, flat; + v16u8 thresh0, b_limit0, limit0, thresh1, b_limit1, limit1; + v16u8 p3, p2, p1, p0, q3, q2, q1, q0; + v16u8 row0, row1, row2, row3, row4, row5, row6, row7; + v16u8 row8, row9, row10, row11, row12, row13, row14, row15; + v8i16 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5; + + LD_UB8(src - 4, pitch, row0, row1, row2, row3, row4, row5, row6, row7); + LD_UB8(src - 4 + (8 * pitch), pitch, row8, row9, row10, row11, row12, row13, + row14, row15); + + TRANSPOSE16x8_UB_UB(row0, row1, row2, row3, row4, row5, row6, row7, row8, + row9, row10, row11, row12, row13, row14, row15, p3, p2, + p1, p0, q0, q1, q2, q3); + + thresh0 = (v16u8)__msa_fill_b(*thresh0_ptr); + thresh1 = (v16u8)__msa_fill_b(*thresh1_ptr); + thresh0 = (v16u8)__msa_ilvr_d((v2i64)thresh1, (v2i64)thresh0); + + b_limit0 = (v16u8)__msa_fill_b(*b_limit0_ptr); + b_limit1 = (v16u8)__msa_fill_b(*b_limit1_ptr); + b_limit0 = (v16u8)__msa_ilvr_d((v2i64)b_limit1, (v2i64)b_limit0); + + limit0 = (v16u8)__msa_fill_b(*limit0_ptr); + limit1 = (v16u8)__msa_fill_b(*limit1_ptr); + limit0 = (v16u8)__msa_ilvr_d((v2i64)limit1, (v2i64)limit0); + + LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit0, b_limit0, thresh0, hev, + mask, flat); + AOM_LPF_FILTER4_4W(p1, p0, q0, q1, mask, hev, p1, p0, q0, q1); + ILVR_B2_SH(p0, p1, q1, q0, tmp0, tmp1); + ILVRL_H2_SH(tmp1, tmp0, tmp2, tmp3); + ILVL_B2_SH(p0, p1, q1, q0, tmp0, tmp1); + ILVRL_H2_SH(tmp1, tmp0, tmp4, tmp5); + + src -= 2; + + ST4x8_UB(tmp2, tmp3, src, pitch); + src += (8 * pitch); + ST4x8_UB(tmp4, tmp5, src, pitch); +} diff --git a/third_party/aom/aom_dsp/mips/loopfilter_8_msa.c b/third_party/aom/aom_dsp/mips/loopfilter_8_msa.c new file mode 100644 index 0000000000..dc203e79cf --- /dev/null +++ b/third_party/aom/aom_dsp/mips/loopfilter_8_msa.c @@ -0,0 +1,333 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include "aom_dsp/mips/loopfilter_msa.h" + +void aom_lpf_horizontal_8_msa(uint8_t *src, int32_t pitch, + const uint8_t *b_limit_ptr, + const uint8_t *limit_ptr, + const uint8_t *thresh_ptr) { + uint64_t p2_d, p1_d, p0_d, q0_d, q1_d, q2_d; + v16u8 mask, hev, flat, thresh, b_limit, limit; + v16u8 p3, p2, p1, p0, q3, q2, q1, q0; + v16u8 p2_out, p1_out, p0_out, q0_out, q1_out, q2_out; + v8i16 p2_filter8, p1_filter8, p0_filter8, q0_filter8, q1_filter8, q2_filter8; + v8u16 p3_r, p2_r, p1_r, p0_r, q3_r, q2_r, q1_r, q0_r; + v16i8 zero = { 0 }; + + /* load vector elements */ + LD_UB8((src - 4 * pitch), pitch, p3, p2, p1, p0, q0, q1, q2, q3); + + thresh = (v16u8)__msa_fill_b(*thresh_ptr); + b_limit = (v16u8)__msa_fill_b(*b_limit_ptr); + limit = (v16u8)__msa_fill_b(*limit_ptr); + + LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh, hev, + mask, flat); + AOM_FLAT4(p3, p2, p0, q0, q2, q3, flat); + AOM_LPF_FILTER4_8W(p1, p0, q0, q1, mask, hev, p1_out, p0_out, q0_out, q1_out); + + flat = (v16u8)__msa_ilvr_d((v2i64)zero, (v2i64)flat); + + if (__msa_test_bz_v(flat)) { + p1_d = __msa_copy_u_d((v2i64)p1_out, 0); + p0_d = __msa_copy_u_d((v2i64)p0_out, 0); + q0_d = __msa_copy_u_d((v2i64)q0_out, 0); + q1_d = __msa_copy_u_d((v2i64)q1_out, 0); + SD4(p1_d, p0_d, q0_d, q1_d, (src - 2 * pitch), pitch); + } else { + ILVR_B8_UH(zero, p3, zero, p2, zero, p1, zero, p0, zero, q0, zero, q1, zero, + q2, zero, q3, p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r); + AOM_FILTER8(p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r, p2_filter8, + p1_filter8, p0_filter8, q0_filter8, q1_filter8, q2_filter8); + + /* convert 16 bit output data into 8 bit */ + PCKEV_B4_SH(zero, p2_filter8, zero, p1_filter8, zero, p0_filter8, zero, + q0_filter8, p2_filter8, p1_filter8, p0_filter8, q0_filter8); + PCKEV_B2_SH(zero, q1_filter8, zero, q2_filter8, q1_filter8, q2_filter8); + + /* store pixel values */ + p2_out = __msa_bmnz_v(p2, (v16u8)p2_filter8, flat); + p1_out = __msa_bmnz_v(p1_out, (v16u8)p1_filter8, flat); + p0_out = __msa_bmnz_v(p0_out, (v16u8)p0_filter8, flat); + q0_out = __msa_bmnz_v(q0_out, (v16u8)q0_filter8, flat); + q1_out = __msa_bmnz_v(q1_out, (v16u8)q1_filter8, flat); + q2_out = __msa_bmnz_v(q2, (v16u8)q2_filter8, flat); + + p2_d = __msa_copy_u_d((v2i64)p2_out, 0); + p1_d = __msa_copy_u_d((v2i64)p1_out, 0); + p0_d = __msa_copy_u_d((v2i64)p0_out, 0); + q0_d = __msa_copy_u_d((v2i64)q0_out, 0); + q1_d = __msa_copy_u_d((v2i64)q1_out, 0); + q2_d = __msa_copy_u_d((v2i64)q2_out, 0); + + src -= 3 * pitch; + + SD4(p2_d, p1_d, p0_d, q0_d, src, pitch); + src += (4 * pitch); + SD(q1_d, src); + src += pitch; + SD(q2_d, src); + } +} + +void aom_lpf_horizontal_8_dual_msa( + uint8_t *src, int32_t pitch, const uint8_t *b_limit0, const uint8_t *limit0, + const uint8_t *thresh0, const uint8_t *b_limit1, const uint8_t *limit1, + const uint8_t *thresh1) { + v16u8 p3, p2, p1, p0, q3, q2, q1, q0; + v16u8 p2_out, p1_out, p0_out, q0_out, q1_out, q2_out; + v16u8 flat, mask, hev, tmp, thresh, b_limit, limit; + v8u16 p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r; + v8u16 p3_l, p2_l, p1_l, p0_l, q0_l, q1_l, q2_l, q3_l; + v8i16 p2_filt8_r, p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r; + v8i16 p2_filt8_l, p1_filt8_l, p0_filt8_l, q0_filt8_l, q1_filt8_l, q2_filt8_l; + v16u8 zero = { 0 }; + + /* load vector elements */ + LD_UB8(src - (4 * pitch), pitch, p3, p2, p1, p0, q0, q1, q2, q3); + + thresh = (v16u8)__msa_fill_b(*thresh0); + tmp = (v16u8)__msa_fill_b(*thresh1); + thresh = (v16u8)__msa_ilvr_d((v2i64)tmp, (v2i64)thresh); + + b_limit = (v16u8)__msa_fill_b(*b_limit0); + tmp = (v16u8)__msa_fill_b(*b_limit1); + b_limit = (v16u8)__msa_ilvr_d((v2i64)tmp, (v2i64)b_limit); + + limit = (v16u8)__msa_fill_b(*limit0); + tmp = (v16u8)__msa_fill_b(*limit1); + limit = (v16u8)__msa_ilvr_d((v2i64)tmp, (v2i64)limit); + + /* mask and hev */ + LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh, hev, + mask, flat); + AOM_FLAT4(p3, p2, p0, q0, q2, q3, flat); + AOM_LPF_FILTER4_4W(p1, p0, q0, q1, mask, hev, p1_out, p0_out, q0_out, q1_out); + + if (__msa_test_bz_v(flat)) { + ST_UB4(p1_out, p0_out, q0_out, q1_out, (src - 2 * pitch), pitch); + } else { + ILVR_B8_UH(zero, p3, zero, p2, zero, p1, zero, p0, zero, q0, zero, q1, zero, + q2, zero, q3, p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r); + AOM_FILTER8(p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r, p2_filt8_r, + p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r); + + ILVL_B4_UH(zero, p3, zero, p2, zero, p1, zero, p0, p3_l, p2_l, p1_l, p0_l); + ILVL_B4_UH(zero, q0, zero, q1, zero, q2, zero, q3, q0_l, q1_l, q2_l, q3_l); + AOM_FILTER8(p3_l, p2_l, p1_l, p0_l, q0_l, q1_l, q2_l, q3_l, p2_filt8_l, + p1_filt8_l, p0_filt8_l, q0_filt8_l, q1_filt8_l, q2_filt8_l); + + /* convert 16 bit output data into 8 bit */ + PCKEV_B4_SH(p2_filt8_l, p2_filt8_r, p1_filt8_l, p1_filt8_r, p0_filt8_l, + p0_filt8_r, q0_filt8_l, q0_filt8_r, p2_filt8_r, p1_filt8_r, + p0_filt8_r, q0_filt8_r); + PCKEV_B2_SH(q1_filt8_l, q1_filt8_r, q2_filt8_l, q2_filt8_r, q1_filt8_r, + q2_filt8_r); + + /* store pixel values */ + p2_out = __msa_bmnz_v(p2, (v16u8)p2_filt8_r, flat); + p1_out = __msa_bmnz_v(p1_out, (v16u8)p1_filt8_r, flat); + p0_out = __msa_bmnz_v(p0_out, (v16u8)p0_filt8_r, flat); + q0_out = __msa_bmnz_v(q0_out, (v16u8)q0_filt8_r, flat); + q1_out = __msa_bmnz_v(q1_out, (v16u8)q1_filt8_r, flat); + q2_out = __msa_bmnz_v(q2, (v16u8)q2_filt8_r, flat); + + src -= 3 * pitch; + + ST_UB4(p2_out, p1_out, p0_out, q0_out, src, pitch); + src += (4 * pitch); + ST_UB2(q1_out, q2_out, src, pitch); + src += (2 * pitch); + } +} + +void aom_lpf_vertical_8_msa(uint8_t *src, int32_t pitch, + const uint8_t *b_limit_ptr, + const uint8_t *limit_ptr, + const uint8_t *thresh_ptr) { + v16u8 p3, p2, p1, p0, q3, q2, q1, q0; + v16u8 p1_out, p0_out, q0_out, q1_out; + v16u8 flat, mask, hev, thresh, b_limit, limit; + v8u16 p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r; + v8i16 p2_filt8_r, p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r; + v16u8 zero = { 0 }; + v8i16 vec0, vec1, vec2, vec3, vec4; + + /* load vector elements */ + LD_UB8(src - 4, pitch, p3, p2, p1, p0, q0, q1, q2, q3); + + TRANSPOSE8x8_UB_UB(p3, p2, p1, p0, q0, q1, q2, q3, p3, p2, p1, p0, q0, q1, q2, + q3); + + thresh = (v16u8)__msa_fill_b(*thresh_ptr); + b_limit = (v16u8)__msa_fill_b(*b_limit_ptr); + limit = (v16u8)__msa_fill_b(*limit_ptr); + + /* mask and hev */ + LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh, hev, + mask, flat); + /* flat4 */ + AOM_FLAT4(p3, p2, p0, q0, q2, q3, flat); + /* filter4 */ + AOM_LPF_FILTER4_8W(p1, p0, q0, q1, mask, hev, p1_out, p0_out, q0_out, q1_out); + + flat = (v16u8)__msa_ilvr_d((v2i64)zero, (v2i64)flat); + + if (__msa_test_bz_v(flat)) { + /* Store 4 pixels p1-_q1 */ + ILVR_B2_SH(p0_out, p1_out, q1_out, q0_out, vec0, vec1); + ILVRL_H2_SH(vec1, vec0, vec2, vec3); + + src -= 2; + ST4x4_UB(vec2, vec2, 0, 1, 2, 3, src, pitch); + src += 4 * pitch; + ST4x4_UB(vec3, vec3, 0, 1, 2, 3, src, pitch); + } else { + ILVR_B8_UH(zero, p3, zero, p2, zero, p1, zero, p0, zero, q0, zero, q1, zero, + q2, zero, q3, p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r); + AOM_FILTER8(p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r, p2_filt8_r, + p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r); + /* convert 16 bit output data into 8 bit */ + PCKEV_B4_SH(p2_filt8_r, p2_filt8_r, p1_filt8_r, p1_filt8_r, p0_filt8_r, + p0_filt8_r, q0_filt8_r, q0_filt8_r, p2_filt8_r, p1_filt8_r, + p0_filt8_r, q0_filt8_r); + PCKEV_B2_SH(q1_filt8_r, q1_filt8_r, q2_filt8_r, q2_filt8_r, q1_filt8_r, + q2_filt8_r); + + /* store pixel values */ + p2 = __msa_bmnz_v(p2, (v16u8)p2_filt8_r, flat); + p1 = __msa_bmnz_v(p1_out, (v16u8)p1_filt8_r, flat); + p0 = __msa_bmnz_v(p0_out, (v16u8)p0_filt8_r, flat); + q0 = __msa_bmnz_v(q0_out, (v16u8)q0_filt8_r, flat); + q1 = __msa_bmnz_v(q1_out, (v16u8)q1_filt8_r, flat); + q2 = __msa_bmnz_v(q2, (v16u8)q2_filt8_r, flat); + + /* Store 6 pixels p2-_q2 */ + ILVR_B2_SH(p1, p2, q0, p0, vec0, vec1); + ILVRL_H2_SH(vec1, vec0, vec2, vec3); + vec4 = (v8i16)__msa_ilvr_b((v16i8)q2, (v16i8)q1); + + src -= 3; + ST4x4_UB(vec2, vec2, 0, 1, 2, 3, src, pitch); + ST2x4_UB(vec4, 0, src + 4, pitch); + src += (4 * pitch); + ST4x4_UB(vec3, vec3, 0, 1, 2, 3, src, pitch); + ST2x4_UB(vec4, 4, src + 4, pitch); + } +} + +void aom_lpf_vertical_8_dual_msa(uint8_t *src, int32_t pitch, + const uint8_t *b_limit0, const uint8_t *limit0, + const uint8_t *thresh0, + const uint8_t *b_limit1, const uint8_t *limit1, + const uint8_t *thresh1) { + uint8_t *temp_src; + v16u8 p3, p2, p1, p0, q3, q2, q1, q0; + v16u8 p1_out, p0_out, q0_out, q1_out; + v16u8 flat, mask, hev, thresh, b_limit, limit; + v16u8 row4, row5, row6, row7, row12, row13, row14, row15; + v8u16 p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r; + v8u16 p3_l, p2_l, p1_l, p0_l, q0_l, q1_l, q2_l, q3_l; + v8i16 p2_filt8_r, p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r; + v8i16 p2_filt8_l, p1_filt8_l, p0_filt8_l, q0_filt8_l, q1_filt8_l, q2_filt8_l; + v16u8 zero = { 0 }; + v8i16 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7; + + temp_src = src - 4; + + LD_UB8(temp_src, pitch, p0, p1, p2, p3, row4, row5, row6, row7); + temp_src += (8 * pitch); + LD_UB8(temp_src, pitch, q3, q2, q1, q0, row12, row13, row14, row15); + + /* transpose 16x8 matrix into 8x16 */ + TRANSPOSE16x8_UB_UB(p0, p1, p2, p3, row4, row5, row6, row7, q3, q2, q1, q0, + row12, row13, row14, row15, p3, p2, p1, p0, q0, q1, q2, + q3); + + thresh = (v16u8)__msa_fill_b(*thresh0); + vec0 = (v8i16)__msa_fill_b(*thresh1); + thresh = (v16u8)__msa_ilvr_d((v2i64)vec0, (v2i64)thresh); + + b_limit = (v16u8)__msa_fill_b(*b_limit0); + vec0 = (v8i16)__msa_fill_b(*b_limit1); + b_limit = (v16u8)__msa_ilvr_d((v2i64)vec0, (v2i64)b_limit); + + limit = (v16u8)__msa_fill_b(*limit0); + vec0 = (v8i16)__msa_fill_b(*limit1); + limit = (v16u8)__msa_ilvr_d((v2i64)vec0, (v2i64)limit); + + /* mask and hev */ + LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh, hev, + mask, flat); + /* flat4 */ + AOM_FLAT4(p3, p2, p0, q0, q2, q3, flat); + /* filter4 */ + AOM_LPF_FILTER4_4W(p1, p0, q0, q1, mask, hev, p1_out, p0_out, q0_out, q1_out); + + if (__msa_test_bz_v(flat)) { + ILVR_B2_SH(p0_out, p1_out, q1_out, q0_out, vec0, vec1); + ILVRL_H2_SH(vec1, vec0, vec2, vec3); + ILVL_B2_SH(p0_out, p1_out, q1_out, q0_out, vec0, vec1); + ILVRL_H2_SH(vec1, vec0, vec4, vec5); + + src -= 2; + ST4x8_UB(vec2, vec3, src, pitch); + src += 8 * pitch; + ST4x8_UB(vec4, vec5, src, pitch); + } else { + ILVR_B8_UH(zero, p3, zero, p2, zero, p1, zero, p0, zero, q0, zero, q1, zero, + q2, zero, q3, p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r); + AOM_FILTER8(p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r, p2_filt8_r, + p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r); + + ILVL_B4_UH(zero, p3, zero, p2, zero, p1, zero, p0, p3_l, p2_l, p1_l, p0_l); + ILVL_B4_UH(zero, q0, zero, q1, zero, q2, zero, q3, q0_l, q1_l, q2_l, q3_l); + + /* filter8 */ + AOM_FILTER8(p3_l, p2_l, p1_l, p0_l, q0_l, q1_l, q2_l, q3_l, p2_filt8_l, + p1_filt8_l, p0_filt8_l, q0_filt8_l, q1_filt8_l, q2_filt8_l); + + /* convert 16 bit output data into 8 bit */ + PCKEV_B4_SH(p2_filt8_l, p2_filt8_r, p1_filt8_l, p1_filt8_r, p0_filt8_l, + p0_filt8_r, q0_filt8_l, q0_filt8_r, p2_filt8_r, p1_filt8_r, + p0_filt8_r, q0_filt8_r); + PCKEV_B2_SH(q1_filt8_l, q1_filt8_r, q2_filt8_l, q2_filt8_r, q1_filt8_r, + q2_filt8_r); + + /* store pixel values */ + p2 = __msa_bmnz_v(p2, (v16u8)p2_filt8_r, flat); + p1 = __msa_bmnz_v(p1_out, (v16u8)p1_filt8_r, flat); + p0 = __msa_bmnz_v(p0_out, (v16u8)p0_filt8_r, flat); + q0 = __msa_bmnz_v(q0_out, (v16u8)q0_filt8_r, flat); + q1 = __msa_bmnz_v(q1_out, (v16u8)q1_filt8_r, flat); + q2 = __msa_bmnz_v(q2, (v16u8)q2_filt8_r, flat); + + ILVR_B2_SH(p1, p2, q0, p0, vec0, vec1); + ILVRL_H2_SH(vec1, vec0, vec3, vec4); + ILVL_B2_SH(p1, p2, q0, p0, vec0, vec1); + ILVRL_H2_SH(vec1, vec0, vec6, vec7); + ILVRL_B2_SH(q2, q1, vec2, vec5); + + src -= 3; + ST4x4_UB(vec3, vec3, 0, 1, 2, 3, src, pitch); + ST2x4_UB(vec2, 0, src + 4, pitch); + src += (4 * pitch); + ST4x4_UB(vec4, vec4, 0, 1, 2, 3, src, pitch); + ST2x4_UB(vec2, 4, src + 4, pitch); + src += (4 * pitch); + ST4x4_UB(vec6, vec6, 0, 1, 2, 3, src, pitch); + ST2x4_UB(vec5, 0, src + 4, pitch); + src += (4 * pitch); + ST4x4_UB(vec7, vec7, 0, 1, 2, 3, src, pitch); + ST2x4_UB(vec5, 4, src + 4, pitch); + } +} diff --git a/third_party/aom/aom_dsp/mips/loopfilter_filters_dspr2.c b/third_party/aom/aom_dsp/mips/loopfilter_filters_dspr2.c new file mode 100644 index 0000000000..8c41278beb --- /dev/null +++ b/third_party/aom/aom_dsp/mips/loopfilter_filters_dspr2.c @@ -0,0 +1,328 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include + +#include "config/aom_dsp_rtcd.h" + +#include "aom/aom_integer.h" +#include "aom_dsp/mips/common_dspr2.h" +#include "aom_dsp/mips/loopfilter_filters_dspr2.h" +#include "aom_dsp/mips/loopfilter_macros_dspr2.h" +#include "aom_dsp/mips/loopfilter_masks_dspr2.h" +#include "aom_mem/aom_mem.h" + +#if HAVE_DSPR2 +void aom_lpf_horizontal_4_dspr2(unsigned char *s, int pitch, + const uint8_t *blimit, const uint8_t *limit, + const uint8_t *thresh) { + uint8_t i; + uint32_t mask; + uint32_t hev; + uint32_t pm1, p0, p1, p2, p3, p4, p5, p6; + uint8_t *sm1, *s0, *s1, *s2, *s3, *s4, *s5, *s6; + uint32_t thresh_vec, flimit_vec, limit_vec; + uint32_t uflimit, ulimit, uthresh; + + uflimit = *blimit; + ulimit = *limit; + uthresh = *thresh; + + /* create quad-byte */ + __asm__ __volatile__( + "replv.qb %[thresh_vec], %[uthresh] \n\t" + "replv.qb %[flimit_vec], %[uflimit] \n\t" + "replv.qb %[limit_vec], %[ulimit] \n\t" + + : [thresh_vec] "=&r"(thresh_vec), [flimit_vec] "=&r"(flimit_vec), + [limit_vec] "=r"(limit_vec) + : [uthresh] "r"(uthresh), [uflimit] "r"(uflimit), [ulimit] "r"(ulimit)); + + /* prefetch data for store */ + prefetch_store(s); + + /* loop filter designed to work using chars so that we can make maximum use + of 8 bit simd instructions. */ + for (i = 0; i < 2; i++) { + sm1 = s - (pitch << 2); + s0 = sm1 + pitch; + s1 = s0 + pitch; + s2 = s - pitch; + s3 = s; + s4 = s + pitch; + s5 = s4 + pitch; + s6 = s5 + pitch; + + __asm__ __volatile__( + "lw %[p1], (%[s1]) \n\t" + "lw %[p2], (%[s2]) \n\t" + "lw %[p3], (%[s3]) \n\t" + "lw %[p4], (%[s4]) \n\t" + + : [p1] "=&r"(p1), [p2] "=&r"(p2), [p3] "=&r"(p3), [p4] "=&r"(p4) + : [s1] "r"(s1), [s2] "r"(s2), [s3] "r"(s3), [s4] "r"(s4)); + + /* if (p1 - p4 == 0) and (p2 - p3 == 0) + mask will be zero and filtering is not needed */ + if (!(((p1 - p4) == 0) && ((p2 - p3) == 0))) { + __asm__ __volatile__( + "lw %[pm1], (%[sm1]) \n\t" + "lw %[p0], (%[s0]) \n\t" + "lw %[p5], (%[s5]) \n\t" + "lw %[p6], (%[s6]) \n\t" + + : [pm1] "=&r"(pm1), [p0] "=&r"(p0), [p5] "=&r"(p5), [p6] "=&r"(p6) + : [sm1] "r"(sm1), [s0] "r"(s0), [s5] "r"(s5), [s6] "r"(s6)); + + filter_hev_mask_dspr2(limit_vec, flimit_vec, p1, p2, pm1, p0, p3, p4, p5, + p6, thresh_vec, &hev, &mask); + + /* if mask == 0 do filtering is not needed */ + if (mask) { + /* filtering */ + filter_dspr2(mask, hev, &p1, &p2, &p3, &p4); + + __asm__ __volatile__( + "sw %[p1], (%[s1]) \n\t" + "sw %[p2], (%[s2]) \n\t" + "sw %[p3], (%[s3]) \n\t" + "sw %[p4], (%[s4]) \n\t" + + : + : [p1] "r"(p1), [p2] "r"(p2), [p3] "r"(p3), [p4] "r"(p4), + [s1] "r"(s1), [s2] "r"(s2), [s3] "r"(s3), [s4] "r"(s4)); + } + } + + s = s + 4; + } +} + +void aom_lpf_vertical_4_dspr2(unsigned char *s, int pitch, + const uint8_t *blimit, const uint8_t *limit, + const uint8_t *thresh) { + uint8_t i; + uint32_t mask, hev; + uint32_t pm1, p0, p1, p2, p3, p4, p5, p6; + uint8_t *s1, *s2, *s3, *s4; + uint32_t prim1, prim2, sec3, sec4, prim3, prim4; + uint32_t thresh_vec, flimit_vec, limit_vec; + uint32_t uflimit, ulimit, uthresh; + + uflimit = *blimit; + ulimit = *limit; + uthresh = *thresh; + + /* create quad-byte */ + __asm__ __volatile__( + "replv.qb %[thresh_vec], %[uthresh] \n\t" + "replv.qb %[flimit_vec], %[uflimit] \n\t" + "replv.qb %[limit_vec], %[ulimit] \n\t" + + : [thresh_vec] "=&r"(thresh_vec), [flimit_vec] "=&r"(flimit_vec), + [limit_vec] "=r"(limit_vec) + : [uthresh] "r"(uthresh), [uflimit] "r"(uflimit), [ulimit] "r"(ulimit)); + + /* prefetch data for store */ + prefetch_store(s + pitch); + + for (i = 0; i < 2; i++) { + s1 = s; + s2 = s + pitch; + s3 = s2 + pitch; + s4 = s3 + pitch; + s = s4 + pitch; + + /* load quad-byte vectors + * memory is 4 byte aligned + */ + p2 = *((uint32_t *)(s1 - 4)); + p6 = *((uint32_t *)(s1)); + p1 = *((uint32_t *)(s2 - 4)); + p5 = *((uint32_t *)(s2)); + p0 = *((uint32_t *)(s3 - 4)); + p4 = *((uint32_t *)(s3)); + pm1 = *((uint32_t *)(s4 - 4)); + p3 = *((uint32_t *)(s4)); + + /* transpose pm1, p0, p1, p2 */ + __asm__ __volatile__( + "precrq.qb.ph %[prim1], %[p2], %[p1] \n\t" + "precr.qb.ph %[prim2], %[p2], %[p1] \n\t" + "precrq.qb.ph %[prim3], %[p0], %[pm1] \n\t" + "precr.qb.ph %[prim4], %[p0], %[pm1] \n\t" + + "precrq.qb.ph %[p1], %[prim1], %[prim2] \n\t" + "precr.qb.ph %[pm1], %[prim1], %[prim2] \n\t" + "precrq.qb.ph %[sec3], %[prim3], %[prim4] \n\t" + "precr.qb.ph %[sec4], %[prim3], %[prim4] \n\t" + + "precrq.ph.w %[p2], %[p1], %[sec3] \n\t" + "precrq.ph.w %[p0], %[pm1], %[sec4] \n\t" + "append %[p1], %[sec3], 16 \n\t" + "append %[pm1], %[sec4], 16 \n\t" + + : [prim1] "=&r"(prim1), [prim2] "=&r"(prim2), [prim3] "=&r"(prim3), + [prim4] "=&r"(prim4), [p2] "+r"(p2), [p1] "+r"(p1), [p0] "+r"(p0), + [pm1] "+r"(pm1), [sec3] "=&r"(sec3), [sec4] "=&r"(sec4) + :); + + /* transpose p3, p4, p5, p6 */ + __asm__ __volatile__( + "precrq.qb.ph %[prim1], %[p6], %[p5] \n\t" + "precr.qb.ph %[prim2], %[p6], %[p5] \n\t" + "precrq.qb.ph %[prim3], %[p4], %[p3] \n\t" + "precr.qb.ph %[prim4], %[p4], %[p3] \n\t" + + "precrq.qb.ph %[p5], %[prim1], %[prim2] \n\t" + "precr.qb.ph %[p3], %[prim1], %[prim2] \n\t" + "precrq.qb.ph %[sec3], %[prim3], %[prim4] \n\t" + "precr.qb.ph %[sec4], %[prim3], %[prim4] \n\t" + + "precrq.ph.w %[p6], %[p5], %[sec3] \n\t" + "precrq.ph.w %[p4], %[p3], %[sec4] \n\t" + "append %[p5], %[sec3], 16 \n\t" + "append %[p3], %[sec4], 16 \n\t" + + : [prim1] "=&r"(prim1), [prim2] "=&r"(prim2), [prim3] "=&r"(prim3), + [prim4] "=&r"(prim4), [p6] "+r"(p6), [p5] "+r"(p5), [p4] "+r"(p4), + [p3] "+r"(p3), [sec3] "=&r"(sec3), [sec4] "=&r"(sec4) + :); + + /* if (p1 - p4 == 0) and (p2 - p3 == 0) + * mask will be zero and filtering is not needed + */ + if (!(((p1 - p4) == 0) && ((p2 - p3) == 0))) { + filter_hev_mask_dspr2(limit_vec, flimit_vec, p1, p2, pm1, p0, p3, p4, p5, + p6, thresh_vec, &hev, &mask); + + /* if mask == 0 do filtering is not needed */ + if (mask) { + /* filtering */ + filter_dspr2(mask, hev, &p1, &p2, &p3, &p4); + + /* unpack processed 4x4 neighborhood + * don't use transpose on output data + * because memory isn't aligned + */ + __asm__ __volatile__( + "sb %[p4], 1(%[s4]) \n\t" + "sb %[p3], 0(%[s4]) \n\t" + "sb %[p2], -1(%[s4]) \n\t" + "sb %[p1], -2(%[s4]) \n\t" + + : + : [p4] "r"(p4), [p3] "r"(p3), [p2] "r"(p2), [p1] "r"(p1), + [s4] "r"(s4)); + + __asm__ __volatile__( + "srl %[p4], %[p4], 8 \n\t" + "srl %[p3], %[p3], 8 \n\t" + "srl %[p2], %[p2], 8 \n\t" + "srl %[p1], %[p1], 8 \n\t" + + : [p4] "+r"(p4), [p3] "+r"(p3), [p2] "+r"(p2), [p1] "+r"(p1) + :); + + __asm__ __volatile__( + "sb %[p4], 1(%[s3]) \n\t" + "sb %[p3], 0(%[s3]) \n\t" + "sb %[p2], -1(%[s3]) \n\t" + "sb %[p1], -2(%[s3]) \n\t" + + : [p1] "+r"(p1) + : [p4] "r"(p4), [p3] "r"(p3), [p2] "r"(p2), [s3] "r"(s3)); + + __asm__ __volatile__( + "srl %[p4], %[p4], 8 \n\t" + "srl %[p3], %[p3], 8 \n\t" + "srl %[p2], %[p2], 8 \n\t" + "srl %[p1], %[p1], 8 \n\t" + + : [p4] "+r"(p4), [p3] "+r"(p3), [p2] "+r"(p2), [p1] "+r"(p1) + :); + + __asm__ __volatile__( + "sb %[p4], 1(%[s2]) \n\t" + "sb %[p3], 0(%[s2]) \n\t" + "sb %[p2], -1(%[s2]) \n\t" + "sb %[p1], -2(%[s2]) \n\t" + + : + : [p4] "r"(p4), [p3] "r"(p3), [p2] "r"(p2), [p1] "r"(p1), + [s2] "r"(s2)); + + __asm__ __volatile__( + "srl %[p4], %[p4], 8 \n\t" + "srl %[p3], %[p3], 8 \n\t" + "srl %[p2], %[p2], 8 \n\t" + "srl %[p1], %[p1], 8 \n\t" + + : [p4] "+r"(p4), [p3] "+r"(p3), [p2] "+r"(p2), [p1] "+r"(p1) + :); + + __asm__ __volatile__( + "sb %[p4], 1(%[s1]) \n\t" + "sb %[p3], 0(%[s1]) \n\t" + "sb %[p2], -1(%[s1]) \n\t" + "sb %[p1], -2(%[s1]) \n\t" + + : + : [p4] "r"(p4), [p3] "r"(p3), [p2] "r"(p2), [p1] "r"(p1), + [s1] "r"(s1)); + } + } + } +} + +void aom_lpf_horizontal_4_dual_dspr2( + uint8_t *s, int p /* pitch */, const uint8_t *blimit0, + const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, + const uint8_t *limit1, const uint8_t *thresh1) { + aom_lpf_horizontal_4_dspr2(s, p, blimit0, limit0, thresh0); + aom_lpf_horizontal_4_dspr2(s + 8, p, blimit1, limit1, thresh1); +} + +void aom_lpf_horizontal_8_dual_dspr2( + uint8_t *s, int p /* pitch */, const uint8_t *blimit0, + const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, + const uint8_t *limit1, const uint8_t *thresh1) { + aom_lpf_horizontal_8_dspr2(s, p, blimit0, limit0, thresh0); + aom_lpf_horizontal_8_dspr2(s + 8, p, blimit1, limit1, thresh1); +} + +void aom_lpf_vertical_4_dual_dspr2(uint8_t *s, int p, const uint8_t *blimit0, + const uint8_t *limit0, + const uint8_t *thresh0, + const uint8_t *blimit1, + const uint8_t *limit1, + const uint8_t *thresh1) { + aom_lpf_vertical_4_dspr2(s, p, blimit0, limit0, thresh0); + aom_lpf_vertical_4_dspr2(s + 8 * p, p, blimit1, limit1, thresh1); +} + +void aom_lpf_vertical_8_dual_dspr2(uint8_t *s, int p, const uint8_t *blimit0, + const uint8_t *limit0, + const uint8_t *thresh0, + const uint8_t *blimit1, + const uint8_t *limit1, + const uint8_t *thresh1) { + aom_lpf_vertical_8_dspr2(s, p, blimit0, limit0, thresh0); + aom_lpf_vertical_8_dspr2(s + 8 * p, p, blimit1, limit1, thresh1); +} + +void aom_lpf_vertical_16_dual_dspr2(uint8_t *s, int p, const uint8_t *blimit, + const uint8_t *limit, + const uint8_t *thresh) { + aom_lpf_vertical_16_dspr2(s, p, blimit, limit, thresh); + aom_lpf_vertical_16_dspr2(s + 8 * p, p, blimit, limit, thresh); +} +#endif // #if HAVE_DSPR2 diff --git a/third_party/aom/aom_dsp/mips/loopfilter_filters_dspr2.h b/third_party/aom/aom_dsp/mips/loopfilter_filters_dspr2.h new file mode 100644 index 0000000000..28f0dc35a3 --- /dev/null +++ b/third_party/aom/aom_dsp/mips/loopfilter_filters_dspr2.h @@ -0,0 +1,736 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_MIPS_LOOPFILTER_FILTERS_DSPR2_H_ +#define AOM_AOM_DSP_MIPS_LOOPFILTER_FILTERS_DSPR2_H_ + +#include + +#include "config/aom_dsp_rtcd.h" + +#include "aom/aom_integer.h" +#include "aom_mem/aom_mem.h" +#include "aom_ports/mem.h" + +#ifdef __cplusplus +extern "C" { +#endif + +#if HAVE_DSPR2 +/* inputs & outputs are quad-byte vectors */ +static INLINE void filter_dspr2(uint32_t mask, uint32_t hev, uint32_t *ps1, + uint32_t *ps0, uint32_t *qs0, uint32_t *qs1) { + int32_t aom_filter_l, aom_filter_r; + int32_t Filter1_l, Filter1_r, Filter2_l, Filter2_r; + int32_t subr_r, subr_l; + uint32_t t1, t2, HWM, t3; + uint32_t hev_l, hev_r, mask_l, mask_r, invhev_l, invhev_r; + int32_t vps1, vps0, vqs0, vqs1; + int32_t vps1_l, vps1_r, vps0_l, vps0_r, vqs0_l, vqs0_r, vqs1_l, vqs1_r; + uint32_t N128; + + N128 = 0x80808080; + t1 = 0x03000300; + t2 = 0x04000400; + t3 = 0x01000100; + HWM = 0xFF00FF00; + + vps0 = (*ps0) ^ N128; + vps1 = (*ps1) ^ N128; + vqs0 = (*qs0) ^ N128; + vqs1 = (*qs1) ^ N128; + + /* use halfword pairs instead quad-bytes because of accuracy */ + vps0_l = vps0 & HWM; + vps0_r = vps0 << 8; + vps0_r = vps0_r & HWM; + + vps1_l = vps1 & HWM; + vps1_r = vps1 << 8; + vps1_r = vps1_r & HWM; + + vqs0_l = vqs0 & HWM; + vqs0_r = vqs0 << 8; + vqs0_r = vqs0_r & HWM; + + vqs1_l = vqs1 & HWM; + vqs1_r = vqs1 << 8; + vqs1_r = vqs1_r & HWM; + + mask_l = mask & HWM; + mask_r = mask << 8; + mask_r = mask_r & HWM; + + hev_l = hev & HWM; + hev_r = hev << 8; + hev_r = hev_r & HWM; + + __asm__ __volatile__( + /* aom_filter = aom_signed_char_clamp(ps1 - qs1); */ + "subq_s.ph %[aom_filter_l], %[vps1_l], %[vqs1_l] \n\t" + "subq_s.ph %[aom_filter_r], %[vps1_r], %[vqs1_r] \n\t" + + /* qs0 - ps0 */ + "subq_s.ph %[subr_l], %[vqs0_l], %[vps0_l] \n\t" + "subq_s.ph %[subr_r], %[vqs0_r], %[vps0_r] \n\t" + + /* aom_filter &= hev; */ + "and %[aom_filter_l], %[aom_filter_l], %[hev_l] \n\t" + "and %[aom_filter_r], %[aom_filter_r], %[hev_r] \n\t" + + /* aom_filter = aom_signed_char_clamp(aom_filter + 3 * (qs0 - ps0)); */ + "addq_s.ph %[aom_filter_l], %[aom_filter_l], %[subr_l] \n\t" + "addq_s.ph %[aom_filter_r], %[aom_filter_r], %[subr_r] \n\t" + "xor %[invhev_l], %[hev_l], %[HWM] \n\t" + "addq_s.ph %[aom_filter_l], %[aom_filter_l], %[subr_l] \n\t" + "addq_s.ph %[aom_filter_r], %[aom_filter_r], %[subr_r] \n\t" + "xor %[invhev_r], %[hev_r], %[HWM] \n\t" + "addq_s.ph %[aom_filter_l], %[aom_filter_l], %[subr_l] \n\t" + "addq_s.ph %[aom_filter_r], %[aom_filter_r], %[subr_r] \n\t" + + /* aom_filter &= mask; */ + "and %[aom_filter_l], %[aom_filter_l], %[mask_l] \n\t" + "and %[aom_filter_r], %[aom_filter_r], %[mask_r] \n\t" + + : [aom_filter_l] "=&r"(aom_filter_l), [aom_filter_r] "=&r"(aom_filter_r), + [subr_l] "=&r"(subr_l), [subr_r] "=&r"(subr_r), + [invhev_l] "=&r"(invhev_l), [invhev_r] "=&r"(invhev_r) + : [vps0_l] "r"(vps0_l), [vps0_r] "r"(vps0_r), [vps1_l] "r"(vps1_l), + [vps1_r] "r"(vps1_r), [vqs0_l] "r"(vqs0_l), [vqs0_r] "r"(vqs0_r), + [vqs1_l] "r"(vqs1_l), [vqs1_r] "r"(vqs1_r), [mask_l] "r"(mask_l), + [mask_r] "r"(mask_r), [hev_l] "r"(hev_l), [hev_r] "r"(hev_r), + [HWM] "r"(HWM)); + + /* save bottom 3 bits so that we round one side +4 and the other +3 */ + __asm__ __volatile__( + /* Filter2 = aom_signed_char_clamp(aom_filter + 3) >>= 3; */ + "addq_s.ph %[Filter1_l], %[aom_filter_l], %[t2] \n\t" + "addq_s.ph %[Filter1_r], %[aom_filter_r], %[t2] \n\t" + + /* Filter1 = aom_signed_char_clamp(aom_filter + 4) >>= 3; */ + "addq_s.ph %[Filter2_l], %[aom_filter_l], %[t1] \n\t" + "addq_s.ph %[Filter2_r], %[aom_filter_r], %[t1] \n\t" + "shra.ph %[Filter1_r], %[Filter1_r], 3 \n\t" + "shra.ph %[Filter1_l], %[Filter1_l], 3 \n\t" + + "shra.ph %[Filter2_l], %[Filter2_l], 3 \n\t" + "shra.ph %[Filter2_r], %[Filter2_r], 3 \n\t" + + "and %[Filter1_l], %[Filter1_l], %[HWM] \n\t" + "and %[Filter1_r], %[Filter1_r], %[HWM] \n\t" + + /* vps0 = aom_signed_char_clamp(ps0 + Filter2); */ + "addq_s.ph %[vps0_l], %[vps0_l], %[Filter2_l] \n\t" + "addq_s.ph %[vps0_r], %[vps0_r], %[Filter2_r] \n\t" + + /* vqs0 = aom_signed_char_clamp(qs0 - Filter1); */ + "subq_s.ph %[vqs0_l], %[vqs0_l], %[Filter1_l] \n\t" + "subq_s.ph %[vqs0_r], %[vqs0_r], %[Filter1_r] \n\t" + + : [Filter1_l] "=&r"(Filter1_l), [Filter1_r] "=&r"(Filter1_r), + [Filter2_l] "=&r"(Filter2_l), [Filter2_r] "=&r"(Filter2_r), + [vps0_l] "+r"(vps0_l), [vps0_r] "+r"(vps0_r), [vqs0_l] "+r"(vqs0_l), + [vqs0_r] "+r"(vqs0_r) + : [t1] "r"(t1), [t2] "r"(t2), [HWM] "r"(HWM), + [aom_filter_l] "r"(aom_filter_l), [aom_filter_r] "r"(aom_filter_r)); + + __asm__ __volatile__( + /* (aom_filter += 1) >>= 1 */ + "addqh.ph %[Filter1_l], %[Filter1_l], %[t3] \n\t" + "addqh.ph %[Filter1_r], %[Filter1_r], %[t3] \n\t" + + /* aom_filter &= ~hev; */ + "and %[Filter1_l], %[Filter1_l], %[invhev_l] \n\t" + "and %[Filter1_r], %[Filter1_r], %[invhev_r] \n\t" + + /* vps1 = aom_signed_char_clamp(ps1 + aom_filter); */ + "addq_s.ph %[vps1_l], %[vps1_l], %[Filter1_l] \n\t" + "addq_s.ph %[vps1_r], %[vps1_r], %[Filter1_r] \n\t" + + /* vqs1 = aom_signed_char_clamp(qs1 - aom_filter); */ + "subq_s.ph %[vqs1_l], %[vqs1_l], %[Filter1_l] \n\t" + "subq_s.ph %[vqs1_r], %[vqs1_r], %[Filter1_r] \n\t" + + : [Filter1_l] "+r"(Filter1_l), [Filter1_r] "+r"(Filter1_r), + [vps1_l] "+r"(vps1_l), [vps1_r] "+r"(vps1_r), [vqs1_l] "+r"(vqs1_l), + [vqs1_r] "+r"(vqs1_r) + : [t3] "r"(t3), [invhev_l] "r"(invhev_l), [invhev_r] "r"(invhev_r)); + + /* Create quad-bytes from halfword pairs */ + vqs0_l = vqs0_l & HWM; + vqs1_l = vqs1_l & HWM; + vps0_l = vps0_l & HWM; + vps1_l = vps1_l & HWM; + + __asm__ __volatile__( + "shrl.ph %[vqs0_r], %[vqs0_r], 8 \n\t" + "shrl.ph %[vps0_r], %[vps0_r], 8 \n\t" + "shrl.ph %[vqs1_r], %[vqs1_r], 8 \n\t" + "shrl.ph %[vps1_r], %[vps1_r], 8 \n\t" + + : [vps1_r] "+r"(vps1_r), [vqs1_r] "+r"(vqs1_r), [vps0_r] "+r"(vps0_r), + [vqs0_r] "+r"(vqs0_r) + :); + + vqs0 = vqs0_l | vqs0_r; + vqs1 = vqs1_l | vqs1_r; + vps0 = vps0_l | vps0_r; + vps1 = vps1_l | vps1_r; + + *ps0 = vps0 ^ N128; + *ps1 = vps1 ^ N128; + *qs0 = vqs0 ^ N128; + *qs1 = vqs1 ^ N128; +} + +static INLINE void filter1_dspr2(uint32_t mask, uint32_t hev, uint32_t ps1, + uint32_t ps0, uint32_t qs0, uint32_t qs1, + uint32_t *p1_f0, uint32_t *p0_f0, + uint32_t *q0_f0, uint32_t *q1_f0) { + int32_t aom_filter_l, aom_filter_r; + int32_t Filter1_l, Filter1_r, Filter2_l, Filter2_r; + int32_t subr_r, subr_l; + uint32_t t1, t2, HWM, t3; + uint32_t hev_l, hev_r, mask_l, mask_r, invhev_l, invhev_r; + int32_t vps1, vps0, vqs0, vqs1; + int32_t vps1_l, vps1_r, vps0_l, vps0_r, vqs0_l, vqs0_r, vqs1_l, vqs1_r; + uint32_t N128; + + N128 = 0x80808080; + t1 = 0x03000300; + t2 = 0x04000400; + t3 = 0x01000100; + HWM = 0xFF00FF00; + + vps0 = (ps0) ^ N128; + vps1 = (ps1) ^ N128; + vqs0 = (qs0) ^ N128; + vqs1 = (qs1) ^ N128; + + /* use halfword pairs instead quad-bytes because of accuracy */ + vps0_l = vps0 & HWM; + vps0_r = vps0 << 8; + vps0_r = vps0_r & HWM; + + vps1_l = vps1 & HWM; + vps1_r = vps1 << 8; + vps1_r = vps1_r & HWM; + + vqs0_l = vqs0 & HWM; + vqs0_r = vqs0 << 8; + vqs0_r = vqs0_r & HWM; + + vqs1_l = vqs1 & HWM; + vqs1_r = vqs1 << 8; + vqs1_r = vqs1_r & HWM; + + mask_l = mask & HWM; + mask_r = mask << 8; + mask_r = mask_r & HWM; + + hev_l = hev & HWM; + hev_r = hev << 8; + hev_r = hev_r & HWM; + + __asm__ __volatile__( + /* aom_filter = aom_signed_char_clamp(ps1 - qs1); */ + "subq_s.ph %[aom_filter_l], %[vps1_l], %[vqs1_l] \n\t" + "subq_s.ph %[aom_filter_r], %[vps1_r], %[vqs1_r] \n\t" + + /* qs0 - ps0 */ + "subq_s.ph %[subr_l], %[vqs0_l], %[vps0_l] \n\t" + "subq_s.ph %[subr_r], %[vqs0_r], %[vps0_r] \n\t" + + /* aom_filter &= hev; */ + "and %[aom_filter_l], %[aom_filter_l], %[hev_l] \n\t" + "and %[aom_filter_r], %[aom_filter_r], %[hev_r] \n\t" + + /* aom_filter = aom_signed_char_clamp(aom_filter + 3 * (qs0 - ps0)); */ + "addq_s.ph %[aom_filter_l], %[aom_filter_l], %[subr_l] \n\t" + "addq_s.ph %[aom_filter_r], %[aom_filter_r], %[subr_r] \n\t" + "xor %[invhev_l], %[hev_l], %[HWM] \n\t" + "addq_s.ph %[aom_filter_l], %[aom_filter_l], %[subr_l] \n\t" + "addq_s.ph %[aom_filter_r], %[aom_filter_r], %[subr_r] \n\t" + "xor %[invhev_r], %[hev_r], %[HWM] \n\t" + "addq_s.ph %[aom_filter_l], %[aom_filter_l], %[subr_l] \n\t" + "addq_s.ph %[aom_filter_r], %[aom_filter_r], %[subr_r] \n\t" + + /* aom_filter &= mask; */ + "and %[aom_filter_l], %[aom_filter_l], %[mask_l] \n\t" + "and %[aom_filter_r], %[aom_filter_r], %[mask_r] \n\t" + + : [aom_filter_l] "=&r"(aom_filter_l), [aom_filter_r] "=&r"(aom_filter_r), + [subr_l] "=&r"(subr_l), [subr_r] "=&r"(subr_r), + [invhev_l] "=&r"(invhev_l), [invhev_r] "=&r"(invhev_r) + : [vps0_l] "r"(vps0_l), [vps0_r] "r"(vps0_r), [vps1_l] "r"(vps1_l), + [vps1_r] "r"(vps1_r), [vqs0_l] "r"(vqs0_l), [vqs0_r] "r"(vqs0_r), + [vqs1_l] "r"(vqs1_l), [vqs1_r] "r"(vqs1_r), [mask_l] "r"(mask_l), + [mask_r] "r"(mask_r), [hev_l] "r"(hev_l), [hev_r] "r"(hev_r), + [HWM] "r"(HWM)); + + /* save bottom 3 bits so that we round one side +4 and the other +3 */ + __asm__ __volatile__( + /* Filter2 = aom_signed_char_clamp(aom_filter + 3) >>= 3; */ + "addq_s.ph %[Filter1_l], %[aom_filter_l], %[t2] \n\t" + "addq_s.ph %[Filter1_r], %[aom_filter_r], %[t2] \n\t" + + /* Filter1 = aom_signed_char_clamp(aom_filter + 4) >>= 3; */ + "addq_s.ph %[Filter2_l], %[aom_filter_l], %[t1] \n\t" + "addq_s.ph %[Filter2_r], %[aom_filter_r], %[t1] \n\t" + "shra.ph %[Filter1_r], %[Filter1_r], 3 \n\t" + "shra.ph %[Filter1_l], %[Filter1_l], 3 \n\t" + + "shra.ph %[Filter2_l], %[Filter2_l], 3 \n\t" + "shra.ph %[Filter2_r], %[Filter2_r], 3 \n\t" + + "and %[Filter1_l], %[Filter1_l], %[HWM] \n\t" + "and %[Filter1_r], %[Filter1_r], %[HWM] \n\t" + + /* vps0 = aom_signed_char_clamp(ps0 + Filter2); */ + "addq_s.ph %[vps0_l], %[vps0_l], %[Filter2_l] \n\t" + "addq_s.ph %[vps0_r], %[vps0_r], %[Filter2_r] \n\t" + + /* vqs0 = aom_signed_char_clamp(qs0 - Filter1); */ + "subq_s.ph %[vqs0_l], %[vqs0_l], %[Filter1_l] \n\t" + "subq_s.ph %[vqs0_r], %[vqs0_r], %[Filter1_r] \n\t" + + : [Filter1_l] "=&r"(Filter1_l), [Filter1_r] "=&r"(Filter1_r), + [Filter2_l] "=&r"(Filter2_l), [Filter2_r] "=&r"(Filter2_r), + [vps0_l] "+r"(vps0_l), [vps0_r] "+r"(vps0_r), [vqs0_l] "+r"(vqs0_l), + [vqs0_r] "+r"(vqs0_r) + : [t1] "r"(t1), [t2] "r"(t2), [HWM] "r"(HWM), + [aom_filter_l] "r"(aom_filter_l), [aom_filter_r] "r"(aom_filter_r)); + + __asm__ __volatile__( + /* (aom_filter += 1) >>= 1 */ + "addqh.ph %[Filter1_l], %[Filter1_l], %[t3] \n\t" + "addqh.ph %[Filter1_r], %[Filter1_r], %[t3] \n\t" + + /* aom_filter &= ~hev; */ + "and %[Filter1_l], %[Filter1_l], %[invhev_l] \n\t" + "and %[Filter1_r], %[Filter1_r], %[invhev_r] \n\t" + + /* vps1 = aom_signed_char_clamp(ps1 + aom_filter); */ + "addq_s.ph %[vps1_l], %[vps1_l], %[Filter1_l] \n\t" + "addq_s.ph %[vps1_r], %[vps1_r], %[Filter1_r] \n\t" + + /* vqs1 = aom_signed_char_clamp(qs1 - aom_filter); */ + "subq_s.ph %[vqs1_l], %[vqs1_l], %[Filter1_l] \n\t" + "subq_s.ph %[vqs1_r], %[vqs1_r], %[Filter1_r] \n\t" + + : [Filter1_l] "+r"(Filter1_l), [Filter1_r] "+r"(Filter1_r), + [vps1_l] "+r"(vps1_l), [vps1_r] "+r"(vps1_r), [vqs1_l] "+r"(vqs1_l), + [vqs1_r] "+r"(vqs1_r) + : [t3] "r"(t3), [invhev_l] "r"(invhev_l), [invhev_r] "r"(invhev_r)); + + /* Create quad-bytes from halfword pairs */ + vqs0_l = vqs0_l & HWM; + vqs1_l = vqs1_l & HWM; + vps0_l = vps0_l & HWM; + vps1_l = vps1_l & HWM; + + __asm__ __volatile__( + "shrl.ph %[vqs0_r], %[vqs0_r], 8 \n\t" + "shrl.ph %[vps0_r], %[vps0_r], 8 \n\t" + "shrl.ph %[vqs1_r], %[vqs1_r], 8 \n\t" + "shrl.ph %[vps1_r], %[vps1_r], 8 \n\t" + + : [vps1_r] "+r"(vps1_r), [vqs1_r] "+r"(vqs1_r), [vps0_r] "+r"(vps0_r), + [vqs0_r] "+r"(vqs0_r) + :); + + vqs0 = vqs0_l | vqs0_r; + vqs1 = vqs1_l | vqs1_r; + vps0 = vps0_l | vps0_r; + vps1 = vps1_l | vps1_r; + + *p0_f0 = vps0 ^ N128; + *p1_f0 = vps1 ^ N128; + *q0_f0 = vqs0 ^ N128; + *q1_f0 = vqs1 ^ N128; +} + +static INLINE void mbfilter_dspr2(uint32_t *op3, uint32_t *op2, uint32_t *op1, + uint32_t *op0, uint32_t *oq0, uint32_t *oq1, + uint32_t *oq2, uint32_t *oq3) { + /* use a 7 tap filter [1, 1, 1, 2, 1, 1, 1] for flat line */ + const uint32_t p3 = *op3, p2 = *op2, p1 = *op1, p0 = *op0; + const uint32_t q0 = *oq0, q1 = *oq1, q2 = *oq2, q3 = *oq3; + uint32_t res_op2, res_op1, res_op0; + uint32_t res_oq0, res_oq1, res_oq2; + uint32_t tmp; + uint32_t add_p210_q012; + uint32_t u32Four = 0x00040004; + + /* *op2 = ROUND_POWER_OF_TWO(p3 + p3 + p3 + p2 + p2 + p1 + p0 + q0, 3) 1 */ + /* *op1 = ROUND_POWER_OF_TWO(p3 + p3 + p2 + p1 + p1 + p0 + q0 + q1, 3) 2 */ + /* *op0 = ROUND_POWER_OF_TWO(p3 + p2 + p1 + p0 + p0 + q0 + q1 + q2, 3) 3 */ + /* *oq0 = ROUND_POWER_OF_TWO(p2 + p1 + p0 + q0 + q0 + q1 + q2 + q3, 3) 4 */ + /* *oq1 = ROUND_POWER_OF_TWO(p1 + p0 + q0 + q1 + q1 + q2 + q3 + q3, 3) 5 */ + /* *oq2 = ROUND_POWER_OF_TWO(p0 + q0 + q1 + q2 + q2 + q3 + q3 + q3, 3) 6 */ + + __asm__ __volatile__( + "addu.ph %[add_p210_q012], %[p2], %[p1] \n\t" + "addu.ph %[add_p210_q012], %[add_p210_q012], %[p0] \n\t" + "addu.ph %[add_p210_q012], %[add_p210_q012], %[q0] \n\t" + "addu.ph %[add_p210_q012], %[add_p210_q012], %[q1] \n\t" + "addu.ph %[add_p210_q012], %[add_p210_q012], %[q2] \n\t" + "addu.ph %[add_p210_q012], %[add_p210_q012], %[u32Four] \n\t" + + "shll.ph %[tmp], %[p3], 1 \n\t" + "addu.ph %[res_op2], %[tmp], %[p3] \n\t" + "addu.ph %[res_op1], %[p3], %[p3] \n\t" + "addu.ph %[res_op2], %[res_op2], %[p2] \n\t" + "addu.ph %[res_op1], %[res_op1], %[p1] \n\t" + "addu.ph %[res_op2], %[res_op2], %[add_p210_q012] \n\t" + "addu.ph %[res_op1], %[res_op1], %[add_p210_q012] \n\t" + "subu.ph %[res_op2], %[res_op2], %[q1] \n\t" + "subu.ph %[res_op1], %[res_op1], %[q2] \n\t" + "subu.ph %[res_op2], %[res_op2], %[q2] \n\t" + "shrl.ph %[res_op1], %[res_op1], 3 \n\t" + "shrl.ph %[res_op2], %[res_op2], 3 \n\t" + "addu.ph %[res_op0], %[p3], %[p0] \n\t" + "addu.ph %[res_oq0], %[q0], %[q3] \n\t" + "addu.ph %[res_op0], %[res_op0], %[add_p210_q012] \n\t" + "addu.ph %[res_oq0], %[res_oq0], %[add_p210_q012] \n\t" + "addu.ph %[res_oq1], %[q3], %[q3] \n\t" + "shll.ph %[tmp], %[q3], 1 \n\t" + "addu.ph %[res_oq1], %[res_oq1], %[q1] \n\t" + "addu.ph %[res_oq2], %[tmp], %[q3] \n\t" + "addu.ph %[res_oq1], %[res_oq1], %[add_p210_q012] \n\t" + "addu.ph %[res_oq2], %[res_oq2], %[add_p210_q012] \n\t" + "subu.ph %[res_oq1], %[res_oq1], %[p2] \n\t" + "addu.ph %[res_oq2], %[res_oq2], %[q2] \n\t" + "shrl.ph %[res_oq1], %[res_oq1], 3 \n\t" + "subu.ph %[res_oq2], %[res_oq2], %[p2] \n\t" + "shrl.ph %[res_oq0], %[res_oq0], 3 \n\t" + "subu.ph %[res_oq2], %[res_oq2], %[p1] \n\t" + "shrl.ph %[res_op0], %[res_op0], 3 \n\t" + "shrl.ph %[res_oq2], %[res_oq2], 3 \n\t" + + : [add_p210_q012] "=&r"(add_p210_q012), [tmp] "=&r"(tmp), + [res_op2] "=&r"(res_op2), [res_op1] "=&r"(res_op1), + [res_op0] "=&r"(res_op0), [res_oq0] "=&r"(res_oq0), + [res_oq1] "=&r"(res_oq1), [res_oq2] "=&r"(res_oq2) + : [p0] "r"(p0), [q0] "r"(q0), [p1] "r"(p1), [q1] "r"(q1), [p2] "r"(p2), + [q2] "r"(q2), [p3] "r"(p3), [q3] "r"(q3), [u32Four] "r"(u32Four)); + + *op2 = res_op2; + *op1 = res_op1; + *op0 = res_op0; + *oq0 = res_oq0; + *oq1 = res_oq1; + *oq2 = res_oq2; +} + +static INLINE void mbfilter1_dspr2(uint32_t p3, uint32_t p2, uint32_t p1, + uint32_t p0, uint32_t q0, uint32_t q1, + uint32_t q2, uint32_t q3, uint32_t *op2_f1, + uint32_t *op1_f1, uint32_t *op0_f1, + uint32_t *oq0_f1, uint32_t *oq1_f1, + uint32_t *oq2_f1) { + /* use a 7 tap filter [1, 1, 1, 2, 1, 1, 1] for flat line */ + uint32_t res_op2, res_op1, res_op0; + uint32_t res_oq0, res_oq1, res_oq2; + uint32_t tmp; + uint32_t add_p210_q012; + uint32_t u32Four = 0x00040004; + + /* *op2 = ROUND_POWER_OF_TWO(p3 + p3 + p3 + p2 + p2 + p1 + p0 + q0, 3) 1 */ + /* *op1 = ROUND_POWER_OF_TWO(p3 + p3 + p2 + p1 + p1 + p0 + q0 + q1, 3) 2 */ + /* *op0 = ROUND_POWER_OF_TWO(p3 + p2 + p1 + p0 + p0 + q0 + q1 + q2, 3) 3 */ + /* *oq0 = ROUND_POWER_OF_TWO(p2 + p1 + p0 + q0 + q0 + q1 + q2 + q3, 3) 4 */ + /* *oq1 = ROUND_POWER_OF_TWO(p1 + p0 + q0 + q1 + q1 + q2 + q3 + q3, 3) 5 */ + /* *oq2 = ROUND_POWER_OF_TWO(p0 + q0 + q1 + q2 + q2 + q3 + q3 + q3, 3) 6 */ + + __asm__ __volatile__( + "addu.ph %[add_p210_q012], %[p2], %[p1] \n\t" + "addu.ph %[add_p210_q012], %[add_p210_q012], %[p0] \n\t" + "addu.ph %[add_p210_q012], %[add_p210_q012], %[q0] \n\t" + "addu.ph %[add_p210_q012], %[add_p210_q012], %[q1] \n\t" + "addu.ph %[add_p210_q012], %[add_p210_q012], %[q2] \n\t" + "addu.ph %[add_p210_q012], %[add_p210_q012], %[u32Four] \n\t" + + "shll.ph %[tmp], %[p3], 1 \n\t" + "addu.ph %[res_op2], %[tmp], %[p3] \n\t" + "addu.ph %[res_op1], %[p3], %[p3] \n\t" + "addu.ph %[res_op2], %[res_op2], %[p2] \n\t" + "addu.ph %[res_op1], %[res_op1], %[p1] \n\t" + "addu.ph %[res_op2], %[res_op2], %[add_p210_q012] \n\t" + "addu.ph %[res_op1], %[res_op1], %[add_p210_q012] \n\t" + "subu.ph %[res_op2], %[res_op2], %[q1] \n\t" + "subu.ph %[res_op1], %[res_op1], %[q2] \n\t" + "subu.ph %[res_op2], %[res_op2], %[q2] \n\t" + "shrl.ph %[res_op1], %[res_op1], 3 \n\t" + "shrl.ph %[res_op2], %[res_op2], 3 \n\t" + "addu.ph %[res_op0], %[p3], %[p0] \n\t" + "addu.ph %[res_oq0], %[q0], %[q3] \n\t" + "addu.ph %[res_op0], %[res_op0], %[add_p210_q012] \n\t" + "addu.ph %[res_oq0], %[res_oq0], %[add_p210_q012] \n\t" + "addu.ph %[res_oq1], %[q3], %[q3] \n\t" + "shll.ph %[tmp], %[q3], 1 \n\t" + "addu.ph %[res_oq1], %[res_oq1], %[q1] \n\t" + "addu.ph %[res_oq2], %[tmp], %[q3] \n\t" + "addu.ph %[res_oq1], %[res_oq1], %[add_p210_q012] \n\t" + "addu.ph %[res_oq2], %[res_oq2], %[add_p210_q012] \n\t" + "subu.ph %[res_oq1], %[res_oq1], %[p2] \n\t" + "addu.ph %[res_oq2], %[res_oq2], %[q2] \n\t" + "shrl.ph %[res_oq1], %[res_oq1], 3 \n\t" + "subu.ph %[res_oq2], %[res_oq2], %[p2] \n\t" + "shrl.ph %[res_oq0], %[res_oq0], 3 \n\t" + "subu.ph %[res_oq2], %[res_oq2], %[p1] \n\t" + "shrl.ph %[res_op0], %[res_op0], 3 \n\t" + "shrl.ph %[res_oq2], %[res_oq2], 3 \n\t" + + : [add_p210_q012] "=&r"(add_p210_q012), [tmp] "=&r"(tmp), + [res_op2] "=&r"(res_op2), [res_op1] "=&r"(res_op1), + [res_op0] "=&r"(res_op0), [res_oq0] "=&r"(res_oq0), + [res_oq1] "=&r"(res_oq1), [res_oq2] "=&r"(res_oq2) + : [p0] "r"(p0), [q0] "r"(q0), [p1] "r"(p1), [q1] "r"(q1), [p2] "r"(p2), + [q2] "r"(q2), [p3] "r"(p3), [q3] "r"(q3), [u32Four] "r"(u32Four)); + + *op2_f1 = res_op2; + *op1_f1 = res_op1; + *op0_f1 = res_op0; + *oq0_f1 = res_oq0; + *oq1_f1 = res_oq1; + *oq2_f1 = res_oq2; +} + +static INLINE void wide_mbfilter_dspr2( + uint32_t *op7, uint32_t *op6, uint32_t *op5, uint32_t *op4, uint32_t *op3, + uint32_t *op2, uint32_t *op1, uint32_t *op0, uint32_t *oq0, uint32_t *oq1, + uint32_t *oq2, uint32_t *oq3, uint32_t *oq4, uint32_t *oq5, uint32_t *oq6, + uint32_t *oq7) { + const uint32_t p7 = *op7, p6 = *op6, p5 = *op5, p4 = *op4; + const uint32_t p3 = *op3, p2 = *op2, p1 = *op1, p0 = *op0; + const uint32_t q0 = *oq0, q1 = *oq1, q2 = *oq2, q3 = *oq3; + const uint32_t q4 = *oq4, q5 = *oq5, q6 = *oq6, q7 = *oq7; + uint32_t res_op6, res_op5, res_op4, res_op3, res_op2, res_op1, res_op0; + uint32_t res_oq0, res_oq1, res_oq2, res_oq3, res_oq4, res_oq5, res_oq6; + uint32_t tmp; + uint32_t add_p6toq6; + uint32_t u32Eight = 0x00080008; + + __asm__ __volatile__( + /* addition of p6,p5,p4,p3,p2,p1,p0,q0,q1,q2,q3,q4,q5,q6 + which is used most of the time */ + "addu.ph %[add_p6toq6], %[p6], %[p5] \n\t" + "addu.ph %[add_p6toq6], %[add_p6toq6], %[p4] \n\t" + "addu.ph %[add_p6toq6], %[add_p6toq6], %[p3] \n\t" + "addu.ph %[add_p6toq6], %[add_p6toq6], %[p2] \n\t" + "addu.ph %[add_p6toq6], %[add_p6toq6], %[p1] \n\t" + "addu.ph %[add_p6toq6], %[add_p6toq6], %[p0] \n\t" + "addu.ph %[add_p6toq6], %[add_p6toq6], %[q0] \n\t" + "addu.ph %[add_p6toq6], %[add_p6toq6], %[q1] \n\t" + "addu.ph %[add_p6toq6], %[add_p6toq6], %[q2] \n\t" + "addu.ph %[add_p6toq6], %[add_p6toq6], %[q3] \n\t" + "addu.ph %[add_p6toq6], %[add_p6toq6], %[q4] \n\t" + "addu.ph %[add_p6toq6], %[add_p6toq6], %[q5] \n\t" + "addu.ph %[add_p6toq6], %[add_p6toq6], %[q6] \n\t" + "addu.ph %[add_p6toq6], %[add_p6toq6], %[u32Eight] \n\t" + + : [add_p6toq6] "=&r"(add_p6toq6) + : [p6] "r"(p6), [p5] "r"(p5), [p4] "r"(p4), [p3] "r"(p3), [p2] "r"(p2), + [p1] "r"(p1), [p0] "r"(p0), [q0] "r"(q0), [q1] "r"(q1), [q2] "r"(q2), + [q3] "r"(q3), [q4] "r"(q4), [q5] "r"(q5), [q6] "r"(q6), + [u32Eight] "r"(u32Eight)); + + __asm__ __volatile__( + /* *op6 = ROUND_POWER_OF_TWO(p7 * 7 + p6 * 2 + p5 + p4 + + p3 + p2 + p1 + p0 + q0, 4) */ + "shll.ph %[tmp], %[p7], 3 \n\t" + "subu.ph %[res_op6], %[tmp], %[p7] \n\t" + "addu.ph %[res_op6], %[res_op6], %[p6] \n\t" + "addu.ph %[res_op6], %[res_op6], %[add_p6toq6] \n\t" + "subu.ph %[res_op6], %[res_op6], %[q1] \n\t" + "subu.ph %[res_op6], %[res_op6], %[q2] \n\t" + "subu.ph %[res_op6], %[res_op6], %[q3] \n\t" + "subu.ph %[res_op6], %[res_op6], %[q4] \n\t" + "subu.ph %[res_op6], %[res_op6], %[q5] \n\t" + "subu.ph %[res_op6], %[res_op6], %[q6] \n\t" + "shrl.ph %[res_op6], %[res_op6], 4 \n\t" + + /* *op5 = ROUND_POWER_OF_TWO(p7 * 6 + p6 + p5 * 2 + p4 + p3 + + p2 + p1 + p0 + q0 + q1, 4) */ + "shll.ph %[tmp], %[p7], 2 \n\t" + "addu.ph %[res_op5], %[tmp], %[p7] \n\t" + "addu.ph %[res_op5], %[res_op5], %[p7] \n\t" + "addu.ph %[res_op5], %[res_op5], %[p5] \n\t" + "addu.ph %[res_op5], %[res_op5], %[add_p6toq6] \n\t" + "subu.ph %[res_op5], %[res_op5], %[q2] \n\t" + "subu.ph %[res_op5], %[res_op5], %[q3] \n\t" + "subu.ph %[res_op5], %[res_op5], %[q4] \n\t" + "subu.ph %[res_op5], %[res_op5], %[q5] \n\t" + "subu.ph %[res_op5], %[res_op5], %[q6] \n\t" + "shrl.ph %[res_op5], %[res_op5], 4 \n\t" + + /* *op4 = ROUND_POWER_OF_TWO(p7 * 5 + p6 + p5 + p4 * 2 + p3 + p2 + + p1 + p0 + q0 + q1 + q2, 4) */ + "shll.ph %[tmp], %[p7], 2 \n\t" + "addu.ph %[res_op4], %[tmp], %[p7] \n\t" + "addu.ph %[res_op4], %[res_op4], %[p4] \n\t" + "addu.ph %[res_op4], %[res_op4], %[add_p6toq6] \n\t" + "subu.ph %[res_op4], %[res_op4], %[q3] \n\t" + "subu.ph %[res_op4], %[res_op4], %[q4] \n\t" + "subu.ph %[res_op4], %[res_op4], %[q5] \n\t" + "subu.ph %[res_op4], %[res_op4], %[q6] \n\t" + "shrl.ph %[res_op4], %[res_op4], 4 \n\t" + + /* *op3 = ROUND_POWER_OF_TWO(p7 * 4 + p6 + p5 + p4 + p3 * 2 + p2 + + p1 + p0 + q0 + q1 + q2 + q3, 4) */ + "shll.ph %[tmp], %[p7], 2 \n\t" + "addu.ph %[res_op3], %[tmp], %[p3] \n\t" + "addu.ph %[res_op3], %[res_op3], %[add_p6toq6] \n\t" + "subu.ph %[res_op3], %[res_op3], %[q4] \n\t" + "subu.ph %[res_op3], %[res_op3], %[q5] \n\t" + "subu.ph %[res_op3], %[res_op3], %[q6] \n\t" + "shrl.ph %[res_op3], %[res_op3], 4 \n\t" + + /* *op2 = ROUND_POWER_OF_TWO(p7 * 3 + p6 + p5 + p4 + p3 + p2 * 2 + p1 + + p0 + q0 + q1 + q2 + q3 + q4, 4) */ + "shll.ph %[tmp], %[p7], 1 \n\t" + "addu.ph %[res_op2], %[tmp], %[p7] \n\t" + "addu.ph %[res_op2], %[res_op2], %[p2] \n\t" + "addu.ph %[res_op2], %[res_op2], %[add_p6toq6] \n\t" + "subu.ph %[res_op2], %[res_op2], %[q5] \n\t" + "subu.ph %[res_op2], %[res_op2], %[q6] \n\t" + "shrl.ph %[res_op2], %[res_op2], 4 \n\t" + + /* *op1 = ROUND_POWER_OF_TWO(p7 * 2 + p6 + p5 + p4 + p3 + p2 + p1 * 2 + + p0 + q0 + q1 + q2 + q3 + q4 + q5, 4); */ + "shll.ph %[tmp], %[p7], 1 \n\t" + "addu.ph %[res_op1], %[tmp], %[p1] \n\t" + "addu.ph %[res_op1], %[res_op1], %[add_p6toq6] \n\t" + "subu.ph %[res_op1], %[res_op1], %[q6] \n\t" + "shrl.ph %[res_op1], %[res_op1], 4 \n\t" + + /* *op0 = ROUND_POWER_OF_TWO(p7 + p6 + p5 + p4 + p3 + p2 + p1 + p0 * 2 + + q0 + q1 + q2 + q3 + q4 + q5 + q6, 4) */ + "addu.ph %[res_op0], %[p7], %[p0] \n\t" + "addu.ph %[res_op0], %[res_op0], %[add_p6toq6] \n\t" + "shrl.ph %[res_op0], %[res_op0], 4 \n\t" + + : [res_op6] "=&r"(res_op6), [res_op5] "=&r"(res_op5), + [res_op4] "=&r"(res_op4), [res_op3] "=&r"(res_op3), + [res_op2] "=&r"(res_op2), [res_op1] "=&r"(res_op1), + [res_op0] "=&r"(res_op0), [tmp] "=&r"(tmp) + : [p7] "r"(p7), [p6] "r"(p6), [p5] "r"(p5), [p4] "r"(p4), [p3] "r"(p3), + [p2] "r"(p2), [p1] "r"(p1), [p0] "r"(p0), [q2] "r"(q2), [q1] "r"(q1), + [q3] "r"(q3), [q4] "r"(q4), [q5] "r"(q5), [q6] "r"(q6), + [add_p6toq6] "r"(add_p6toq6)); + + *op6 = res_op6; + *op5 = res_op5; + *op4 = res_op4; + *op3 = res_op3; + *op2 = res_op2; + *op1 = res_op1; + *op0 = res_op0; + + __asm__ __volatile__( + /* *oq0 = ROUND_POWER_OF_TWO(p6 + p5 + p4 + p3 + p2 + p1 + p0 + q0 * 2 + + q1 + q2 + q3 + q4 + q5 + q6 + q7, 4); */ + "addu.ph %[res_oq0], %[q7], %[q0] \n\t" + "addu.ph %[res_oq0], %[res_oq0], %[add_p6toq6] \n\t" + "shrl.ph %[res_oq0], %[res_oq0], 4 \n\t" + + /* *oq1 = ROUND_POWER_OF_TWO(p5 + p4 + p3 + p2 + p1 + p0 + q0 + q1 * 2 + + q2 + q3 + q4 + q5 + q6 + q7 * 2, 4) */ + "shll.ph %[tmp], %[q7], 1 \n\t" + "addu.ph %[res_oq1], %[tmp], %[q1] \n\t" + "addu.ph %[res_oq1], %[res_oq1], %[add_p6toq6] \n\t" + "subu.ph %[res_oq1], %[res_oq1], %[p6] \n\t" + "shrl.ph %[res_oq1], %[res_oq1], 4 \n\t" + + /* *oq2 = ROUND_POWER_OF_TWO(p4 + p3 + p2 + p1 + p0 + q0 + q1 + q2 * 2 + + q3 + q4 + q5 + q6 + q7 * 3, 4) */ + "shll.ph %[tmp], %[q7], 1 \n\t" + "addu.ph %[res_oq2], %[tmp], %[q7] \n\t" + "addu.ph %[res_oq2], %[res_oq2], %[q2] \n\t" + "addu.ph %[res_oq2], %[res_oq2], %[add_p6toq6] \n\t" + "subu.ph %[res_oq2], %[res_oq2], %[p5] \n\t" + "subu.ph %[res_oq2], %[res_oq2], %[p6] \n\t" + "shrl.ph %[res_oq2], %[res_oq2], 4 \n\t" + + /* *oq3 = ROUND_POWER_OF_TWO(p3 + p2 + p1 + p0 + q0 + q1 + q2 + + q3 * 2 + q4 + q5 + q6 + q7 * 4, 4) */ + "shll.ph %[tmp], %[q7], 2 \n\t" + "addu.ph %[res_oq3], %[tmp], %[q3] \n\t" + "addu.ph %[res_oq3], %[res_oq3], %[add_p6toq6] \n\t" + "subu.ph %[res_oq3], %[res_oq3], %[p4] \n\t" + "subu.ph %[res_oq3], %[res_oq3], %[p5] \n\t" + "subu.ph %[res_oq3], %[res_oq3], %[p6] \n\t" + "shrl.ph %[res_oq3], %[res_oq3], 4 \n\t" + + /* *oq4 = ROUND_POWER_OF_TWO(p2 + p1 + p0 + q0 + q1 + q2 + q3 + + q4 * 2 + q5 + q6 + q7 * 5, 4) */ + "shll.ph %[tmp], %[q7], 2 \n\t" + "addu.ph %[res_oq4], %[tmp], %[q7] \n\t" + "addu.ph %[res_oq4], %[res_oq4], %[q4] \n\t" + "addu.ph %[res_oq4], %[res_oq4], %[add_p6toq6] \n\t" + "subu.ph %[res_oq4], %[res_oq4], %[p3] \n\t" + "subu.ph %[res_oq4], %[res_oq4], %[p4] \n\t" + "subu.ph %[res_oq4], %[res_oq4], %[p5] \n\t" + "subu.ph %[res_oq4], %[res_oq4], %[p6] \n\t" + "shrl.ph %[res_oq4], %[res_oq4], 4 \n\t" + + /* *oq5 = ROUND_POWER_OF_TWO(p1 + p0 + q0 + q1 + q2 + q3 + q4 + + q5 * 2 + q6 + q7 * 6, 4) */ + "shll.ph %[tmp], %[q7], 2 \n\t" + "addu.ph %[res_oq5], %[tmp], %[q7] \n\t" + "addu.ph %[res_oq5], %[res_oq5], %[q7] \n\t" + "addu.ph %[res_oq5], %[res_oq5], %[q5] \n\t" + "addu.ph %[res_oq5], %[res_oq5], %[add_p6toq6] \n\t" + "subu.ph %[res_oq5], %[res_oq5], %[p2] \n\t" + "subu.ph %[res_oq5], %[res_oq5], %[p3] \n\t" + "subu.ph %[res_oq5], %[res_oq5], %[p4] \n\t" + "subu.ph %[res_oq5], %[res_oq5], %[p5] \n\t" + "subu.ph %[res_oq5], %[res_oq5], %[p6] \n\t" + "shrl.ph %[res_oq5], %[res_oq5], 4 \n\t" + + /* *oq6 = ROUND_POWER_OF_TWO(p0 + q0 + q1 + q2 + q3 + + q4 + q5 + q6 * 2 + q7 * 7, 4) */ + "shll.ph %[tmp], %[q7], 3 \n\t" + "subu.ph %[res_oq6], %[tmp], %[q7] \n\t" + "addu.ph %[res_oq6], %[res_oq6], %[q6] \n\t" + "addu.ph %[res_oq6], %[res_oq6], %[add_p6toq6] \n\t" + "subu.ph %[res_oq6], %[res_oq6], %[p1] \n\t" + "subu.ph %[res_oq6], %[res_oq6], %[p2] \n\t" + "subu.ph %[res_oq6], %[res_oq6], %[p3] \n\t" + "subu.ph %[res_oq6], %[res_oq6], %[p4] \n\t" + "subu.ph %[res_oq6], %[res_oq6], %[p5] \n\t" + "subu.ph %[res_oq6], %[res_oq6], %[p6] \n\t" + "shrl.ph %[res_oq6], %[res_oq6], 4 \n\t" + + : [res_oq6] "=&r"(res_oq6), [res_oq5] "=&r"(res_oq5), + [res_oq4] "=&r"(res_oq4), [res_oq3] "=&r"(res_oq3), + [res_oq2] "=&r"(res_oq2), [res_oq1] "=&r"(res_oq1), + [res_oq0] "=&r"(res_oq0), [tmp] "=&r"(tmp) + : [q7] "r"(q7), [q6] "r"(q6), [q5] "r"(q5), [q4] "r"(q4), [q3] "r"(q3), + [q2] "r"(q2), [q1] "r"(q1), [q0] "r"(q0), [p1] "r"(p1), [p2] "r"(p2), + [p3] "r"(p3), [p4] "r"(p4), [p5] "r"(p5), [p6] "r"(p6), + [add_p6toq6] "r"(add_p6toq6)); + + *oq0 = res_oq0; + *oq1 = res_oq1; + *oq2 = res_oq2; + *oq3 = res_oq3; + *oq4 = res_oq4; + *oq5 = res_oq5; + *oq6 = res_oq6; +} +#endif // #if HAVE_DSPR2 +#ifdef __cplusplus +} // extern "C" +#endif + +#endif // AOM_AOM_DSP_MIPS_LOOPFILTER_FILTERS_DSPR2_H_ diff --git a/third_party/aom/aom_dsp/mips/loopfilter_macros_dspr2.h b/third_party/aom/aom_dsp/mips/loopfilter_macros_dspr2.h new file mode 100644 index 0000000000..62295d69dd --- /dev/null +++ b/third_party/aom/aom_dsp/mips/loopfilter_macros_dspr2.h @@ -0,0 +1,437 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_MIPS_LOOPFILTER_MACROS_DSPR2_H_ +#define AOM_AOM_DSP_MIPS_LOOPFILTER_MACROS_DSPR2_H_ + +#include + +#include "config/aom_dsp_rtcd.h" + +#include "aom/aom_integer.h" +#include "aom_mem/aom_mem.h" + +#ifdef __cplusplus +extern "C" { +#endif + +#if HAVE_DSPR2 +#define STORE_F0() \ + { \ + __asm__ __volatile__( \ + "sb %[q1_f0], 1(%[s4]) \n\t" \ + "sb %[q0_f0], 0(%[s4]) \n\t" \ + "sb %[p0_f0], -1(%[s4]) \n\t" \ + "sb %[p1_f0], -2(%[s4]) \n\t" \ + \ + : \ + : [q1_f0] "r"(q1_f0), [q0_f0] "r"(q0_f0), [p0_f0] "r"(p0_f0), \ + [p1_f0] "r"(p1_f0), [s4] "r"(s4)); \ + \ + __asm__ __volatile__( \ + "srl %[q1_f0], %[q1_f0], 8 \n\t" \ + "srl %[q0_f0], %[q0_f0], 8 \n\t" \ + "srl %[p0_f0], %[p0_f0], 8 \n\t" \ + "srl %[p1_f0], %[p1_f0], 8 \n\t" \ + \ + : [q1_f0] "+r"(q1_f0), [q0_f0] "+r"(q0_f0), [p0_f0] "+r"(p0_f0), \ + [p1_f0] "+r"(p1_f0) \ + :); \ + \ + __asm__ __volatile__( \ + "sb %[q1_f0], 1(%[s3]) \n\t" \ + "sb %[q0_f0], 0(%[s3]) \n\t" \ + "sb %[p0_f0], -1(%[s3]) \n\t" \ + "sb %[p1_f0], -2(%[s3]) \n\t" \ + \ + : [p1_f0] "+r"(p1_f0) \ + : [q1_f0] "r"(q1_f0), [q0_f0] "r"(q0_f0), [s3] "r"(s3), \ + [p0_f0] "r"(p0_f0)); \ + \ + __asm__ __volatile__( \ + "srl %[q1_f0], %[q1_f0], 8 \n\t" \ + "srl %[q0_f0], %[q0_f0], 8 \n\t" \ + "srl %[p0_f0], %[p0_f0], 8 \n\t" \ + "srl %[p1_f0], %[p1_f0], 8 \n\t" \ + \ + : [q1_f0] "+r"(q1_f0), [q0_f0] "+r"(q0_f0), [p0_f0] "+r"(p0_f0), \ + [p1_f0] "+r"(p1_f0) \ + :); \ + \ + __asm__ __volatile__( \ + "sb %[q1_f0], 1(%[s2]) \n\t" \ + "sb %[q0_f0], 0(%[s2]) \n\t" \ + "sb %[p0_f0], -1(%[s2]) \n\t" \ + "sb %[p1_f0], -2(%[s2]) \n\t" \ + \ + : \ + : [q1_f0] "r"(q1_f0), [q0_f0] "r"(q0_f0), [p0_f0] "r"(p0_f0), \ + [p1_f0] "r"(p1_f0), [s2] "r"(s2)); \ + \ + __asm__ __volatile__( \ + "srl %[q1_f0], %[q1_f0], 8 \n\t" \ + "srl %[q0_f0], %[q0_f0], 8 \n\t" \ + "srl %[p0_f0], %[p0_f0], 8 \n\t" \ + "srl %[p1_f0], %[p1_f0], 8 \n\t" \ + \ + : [q1_f0] "+r"(q1_f0), [q0_f0] "+r"(q0_f0), [p0_f0] "+r"(p0_f0), \ + [p1_f0] "+r"(p1_f0) \ + :); \ + \ + __asm__ __volatile__( \ + "sb %[q1_f0], 1(%[s1]) \n\t" \ + "sb %[q0_f0], 0(%[s1]) \n\t" \ + "sb %[p0_f0], -1(%[s1]) \n\t" \ + "sb %[p1_f0], -2(%[s1]) \n\t" \ + \ + : \ + : [q1_f0] "r"(q1_f0), [q0_f0] "r"(q0_f0), [p0_f0] "r"(p0_f0), \ + [p1_f0] "r"(p1_f0), [s1] "r"(s1)); \ + } + +#define STORE_F1() \ + { \ + __asm__ __volatile__( \ + "sb %[q2_r], 2(%[s4]) \n\t" \ + "sb %[q1_r], 1(%[s4]) \n\t" \ + "sb %[q0_r], 0(%[s4]) \n\t" \ + "sb %[p0_r], -1(%[s4]) \n\t" \ + "sb %[p1_r], -2(%[s4]) \n\t" \ + "sb %[p2_r], -3(%[s4]) \n\t" \ + \ + : \ + : [q2_r] "r"(q2_r), [q1_r] "r"(q1_r), [q0_r] "r"(q0_r), \ + [p0_r] "r"(p0_r), [p1_r] "r"(p1_r), [p2_r] "r"(p2_r), [s4] "r"(s4)); \ + \ + __asm__ __volatile__( \ + "srl %[q2_r], %[q2_r], 16 \n\t" \ + "srl %[q1_r], %[q1_r], 16 \n\t" \ + "srl %[q0_r], %[q0_r], 16 \n\t" \ + "srl %[p0_r], %[p0_r], 16 \n\t" \ + "srl %[p1_r], %[p1_r], 16 \n\t" \ + "srl %[p2_r], %[p2_r], 16 \n\t" \ + \ + : [q2_r] "+r"(q2_r), [q1_r] "+r"(q1_r), [q0_r] "+r"(q0_r), \ + [p0_r] "+r"(p0_r), [p1_r] "+r"(p1_r), [p2_r] "+r"(p2_r) \ + :); \ + \ + __asm__ __volatile__( \ + "sb %[q2_r], 2(%[s3]) \n\t" \ + "sb %[q1_r], 1(%[s3]) \n\t" \ + "sb %[q0_r], 0(%[s3]) \n\t" \ + "sb %[p0_r], -1(%[s3]) \n\t" \ + "sb %[p1_r], -2(%[s3]) \n\t" \ + "sb %[p2_r], -3(%[s3]) \n\t" \ + \ + : \ + : [q2_r] "r"(q2_r), [q1_r] "r"(q1_r), [q0_r] "r"(q0_r), \ + [p0_r] "r"(p0_r), [p1_r] "r"(p1_r), [p2_r] "r"(p2_r), [s3] "r"(s3)); \ + \ + __asm__ __volatile__( \ + "sb %[q2_l], 2(%[s2]) \n\t" \ + "sb %[q1_l], 1(%[s2]) \n\t" \ + "sb %[q0_l], 0(%[s2]) \n\t" \ + "sb %[p0_l], -1(%[s2]) \n\t" \ + "sb %[p1_l], -2(%[s2]) \n\t" \ + "sb %[p2_l], -3(%[s2]) \n\t" \ + \ + : \ + : [q2_l] "r"(q2_l), [q1_l] "r"(q1_l), [q0_l] "r"(q0_l), \ + [p0_l] "r"(p0_l), [p1_l] "r"(p1_l), [p2_l] "r"(p2_l), [s2] "r"(s2)); \ + \ + __asm__ __volatile__( \ + "srl %[q2_l], %[q2_l], 16 \n\t" \ + "srl %[q1_l], %[q1_l], 16 \n\t" \ + "srl %[q0_l], %[q0_l], 16 \n\t" \ + "srl %[p0_l], %[p0_l], 16 \n\t" \ + "srl %[p1_l], %[p1_l], 16 \n\t" \ + "srl %[p2_l], %[p2_l], 16 \n\t" \ + \ + : [q2_l] "+r"(q2_l), [q1_l] "+r"(q1_l), [q0_l] "+r"(q0_l), \ + [p0_l] "+r"(p0_l), [p1_l] "+r"(p1_l), [p2_l] "+r"(p2_l) \ + :); \ + \ + __asm__ __volatile__( \ + "sb %[q2_l], 2(%[s1]) \n\t" \ + "sb %[q1_l], 1(%[s1]) \n\t" \ + "sb %[q0_l], 0(%[s1]) \n\t" \ + "sb %[p0_l], -1(%[s1]) \n\t" \ + "sb %[p1_l], -2(%[s1]) \n\t" \ + "sb %[p2_l], -3(%[s1]) \n\t" \ + \ + : \ + : [q2_l] "r"(q2_l), [q1_l] "r"(q1_l), [q0_l] "r"(q0_l), \ + [p0_l] "r"(p0_l), [p1_l] "r"(p1_l), [p2_l] "r"(p2_l), [s1] "r"(s1)); \ + } + +#define STORE_F2() \ + { \ + __asm__ __volatile__( \ + "sb %[q6_r], 6(%[s4]) \n\t" \ + "sb %[q5_r], 5(%[s4]) \n\t" \ + "sb %[q4_r], 4(%[s4]) \n\t" \ + "sb %[q3_r], 3(%[s4]) \n\t" \ + "sb %[q2_r], 2(%[s4]) \n\t" \ + "sb %[q1_r], 1(%[s4]) \n\t" \ + "sb %[q0_r], 0(%[s4]) \n\t" \ + "sb %[p0_r], -1(%[s4]) \n\t" \ + "sb %[p1_r], -2(%[s4]) \n\t" \ + "sb %[p2_r], -3(%[s4]) \n\t" \ + "sb %[p3_r], -4(%[s4]) \n\t" \ + "sb %[p4_r], -5(%[s4]) \n\t" \ + "sb %[p5_r], -6(%[s4]) \n\t" \ + "sb %[p6_r], -7(%[s4]) \n\t" \ + \ + : \ + : [q6_r] "r"(q6_r), [q5_r] "r"(q5_r), [q4_r] "r"(q4_r), \ + [q3_r] "r"(q3_r), [q2_r] "r"(q2_r), [q1_r] "r"(q1_r), \ + [q0_r] "r"(q0_r), [p0_r] "r"(p0_r), [p1_r] "r"(p1_r), \ + [p2_r] "r"(p2_r), [p3_r] "r"(p3_r), [p4_r] "r"(p4_r), \ + [p5_r] "r"(p5_r), [p6_r] "r"(p6_r), [s4] "r"(s4)); \ + \ + __asm__ __volatile__( \ + "srl %[q6_r], %[q6_r], 16 \n\t" \ + "srl %[q5_r], %[q5_r], 16 \n\t" \ + "srl %[q4_r], %[q4_r], 16 \n\t" \ + "srl %[q3_r], %[q3_r], 16 \n\t" \ + "srl %[q2_r], %[q2_r], 16 \n\t" \ + "srl %[q1_r], %[q1_r], 16 \n\t" \ + "srl %[q0_r], %[q0_r], 16 \n\t" \ + "srl %[p0_r], %[p0_r], 16 \n\t" \ + "srl %[p1_r], %[p1_r], 16 \n\t" \ + "srl %[p2_r], %[p2_r], 16 \n\t" \ + "srl %[p3_r], %[p3_r], 16 \n\t" \ + "srl %[p4_r], %[p4_r], 16 \n\t" \ + "srl %[p5_r], %[p5_r], 16 \n\t" \ + "srl %[p6_r], %[p6_r], 16 \n\t" \ + \ + : [q6_r] "+r"(q6_r), [q5_r] "+r"(q5_r), [q4_r] "+r"(q4_r), \ + [q3_r] "+r"(q3_r), [q2_r] "+r"(q2_r), [q1_r] "+r"(q1_r), \ + [q0_r] "+r"(q0_r), [p0_r] "+r"(p0_r), [p1_r] "+r"(p1_r), \ + [p2_r] "+r"(p2_r), [p3_r] "+r"(p3_r), [p4_r] "+r"(p4_r), \ + [p5_r] "+r"(p5_r), [p6_r] "+r"(p6_r) \ + :); \ + \ + __asm__ __volatile__( \ + "sb %[q6_r], 6(%[s3]) \n\t" \ + "sb %[q5_r], 5(%[s3]) \n\t" \ + "sb %[q4_r], 4(%[s3]) \n\t" \ + "sb %[q3_r], 3(%[s3]) \n\t" \ + "sb %[q2_r], 2(%[s3]) \n\t" \ + "sb %[q1_r], 1(%[s3]) \n\t" \ + "sb %[q0_r], 0(%[s3]) \n\t" \ + "sb %[p0_r], -1(%[s3]) \n\t" \ + "sb %[p1_r], -2(%[s3]) \n\t" \ + "sb %[p2_r], -3(%[s3]) \n\t" \ + "sb %[p3_r], -4(%[s3]) \n\t" \ + "sb %[p4_r], -5(%[s3]) \n\t" \ + "sb %[p5_r], -6(%[s3]) \n\t" \ + "sb %[p6_r], -7(%[s3]) \n\t" \ + \ + : \ + : [q6_r] "r"(q6_r), [q5_r] "r"(q5_r), [q4_r] "r"(q4_r), \ + [q3_r] "r"(q3_r), [q2_r] "r"(q2_r), [q1_r] "r"(q1_r), \ + [q0_r] "r"(q0_r), [p0_r] "r"(p0_r), [p1_r] "r"(p1_r), \ + [p2_r] "r"(p2_r), [p3_r] "r"(p3_r), [p4_r] "r"(p4_r), \ + [p5_r] "r"(p5_r), [p6_r] "r"(p6_r), [s3] "r"(s3)); \ + \ + __asm__ __volatile__( \ + "sb %[q6_l], 6(%[s2]) \n\t" \ + "sb %[q5_l], 5(%[s2]) \n\t" \ + "sb %[q4_l], 4(%[s2]) \n\t" \ + "sb %[q3_l], 3(%[s2]) \n\t" \ + "sb %[q2_l], 2(%[s2]) \n\t" \ + "sb %[q1_l], 1(%[s2]) \n\t" \ + "sb %[q0_l], 0(%[s2]) \n\t" \ + "sb %[p0_l], -1(%[s2]) \n\t" \ + "sb %[p1_l], -2(%[s2]) \n\t" \ + "sb %[p2_l], -3(%[s2]) \n\t" \ + "sb %[p3_l], -4(%[s2]) \n\t" \ + "sb %[p4_l], -5(%[s2]) \n\t" \ + "sb %[p5_l], -6(%[s2]) \n\t" \ + "sb %[p6_l], -7(%[s2]) \n\t" \ + \ + : \ + : [q6_l] "r"(q6_l), [q5_l] "r"(q5_l), [q4_l] "r"(q4_l), \ + [q3_l] "r"(q3_l), [q2_l] "r"(q2_l), [q1_l] "r"(q1_l), \ + [q0_l] "r"(q0_l), [p0_l] "r"(p0_l), [p1_l] "r"(p1_l), \ + [p2_l] "r"(p2_l), [p3_l] "r"(p3_l), [p4_l] "r"(p4_l), \ + [p5_l] "r"(p5_l), [p6_l] "r"(p6_l), [s2] "r"(s2)); \ + \ + __asm__ __volatile__( \ + "srl %[q6_l], %[q6_l], 16 \n\t" \ + "srl %[q5_l], %[q5_l], 16 \n\t" \ + "srl %[q4_l], %[q4_l], 16 \n\t" \ + "srl %[q3_l], %[q3_l], 16 \n\t" \ + "srl %[q2_l], %[q2_l], 16 \n\t" \ + "srl %[q1_l], %[q1_l], 16 \n\t" \ + "srl %[q0_l], %[q0_l], 16 \n\t" \ + "srl %[p0_l], %[p0_l], 16 \n\t" \ + "srl %[p1_l], %[p1_l], 16 \n\t" \ + "srl %[p2_l], %[p2_l], 16 \n\t" \ + "srl %[p3_l], %[p3_l], 16 \n\t" \ + "srl %[p4_l], %[p4_l], 16 \n\t" \ + "srl %[p5_l], %[p5_l], 16 \n\t" \ + "srl %[p6_l], %[p6_l], 16 \n\t" \ + \ + : [q6_l] "+r"(q6_l), [q5_l] "+r"(q5_l), [q4_l] "+r"(q4_l), \ + [q3_l] "+r"(q3_l), [q2_l] "+r"(q2_l), [q1_l] "+r"(q1_l), \ + [q0_l] "+r"(q0_l), [p0_l] "+r"(p0_l), [p1_l] "+r"(p1_l), \ + [p2_l] "+r"(p2_l), [p3_l] "+r"(p3_l), [p4_l] "+r"(p4_l), \ + [p5_l] "+r"(p5_l), [p6_l] "+r"(p6_l) \ + :); \ + \ + __asm__ __volatile__( \ + "sb %[q6_l], 6(%[s1]) \n\t" \ + "sb %[q5_l], 5(%[s1]) \n\t" \ + "sb %[q4_l], 4(%[s1]) \n\t" \ + "sb %[q3_l], 3(%[s1]) \n\t" \ + "sb %[q2_l], 2(%[s1]) \n\t" \ + "sb %[q1_l], 1(%[s1]) \n\t" \ + "sb %[q0_l], 0(%[s1]) \n\t" \ + "sb %[p0_l], -1(%[s1]) \n\t" \ + "sb %[p1_l], -2(%[s1]) \n\t" \ + "sb %[p2_l], -3(%[s1]) \n\t" \ + "sb %[p3_l], -4(%[s1]) \n\t" \ + "sb %[p4_l], -5(%[s1]) \n\t" \ + "sb %[p5_l], -6(%[s1]) \n\t" \ + "sb %[p6_l], -7(%[s1]) \n\t" \ + \ + : \ + : [q6_l] "r"(q6_l), [q5_l] "r"(q5_l), [q4_l] "r"(q4_l), \ + [q3_l] "r"(q3_l), [q2_l] "r"(q2_l), [q1_l] "r"(q1_l), \ + [q0_l] "r"(q0_l), [p0_l] "r"(p0_l), [p1_l] "r"(p1_l), \ + [p2_l] "r"(p2_l), [p3_l] "r"(p3_l), [p4_l] "r"(p4_l), \ + [p5_l] "r"(p5_l), [p6_l] "r"(p6_l), [s1] "r"(s1)); \ + } + +#define PACK_LEFT_0TO3() \ + { \ + __asm__ __volatile__( \ + "preceu.ph.qbl %[p3_l], %[p3] \n\t" \ + "preceu.ph.qbl %[p2_l], %[p2] \n\t" \ + "preceu.ph.qbl %[p1_l], %[p1] \n\t" \ + "preceu.ph.qbl %[p0_l], %[p0] \n\t" \ + "preceu.ph.qbl %[q0_l], %[q0] \n\t" \ + "preceu.ph.qbl %[q1_l], %[q1] \n\t" \ + "preceu.ph.qbl %[q2_l], %[q2] \n\t" \ + "preceu.ph.qbl %[q3_l], %[q3] \n\t" \ + \ + : [p3_l] "=&r"(p3_l), [p2_l] "=&r"(p2_l), [p1_l] "=&r"(p1_l), \ + [p0_l] "=&r"(p0_l), [q0_l] "=&r"(q0_l), [q1_l] "=&r"(q1_l), \ + [q2_l] "=&r"(q2_l), [q3_l] "=&r"(q3_l) \ + : [p3] "r"(p3), [p2] "r"(p2), [p1] "r"(p1), [p0] "r"(p0), \ + [q0] "r"(q0), [q1] "r"(q1), [q2] "r"(q2), [q3] "r"(q3)); \ + } + +#define PACK_LEFT_4TO7() \ + { \ + __asm__ __volatile__( \ + "preceu.ph.qbl %[p7_l], %[p7] \n\t" \ + "preceu.ph.qbl %[p6_l], %[p6] \n\t" \ + "preceu.ph.qbl %[p5_l], %[p5] \n\t" \ + "preceu.ph.qbl %[p4_l], %[p4] \n\t" \ + "preceu.ph.qbl %[q4_l], %[q4] \n\t" \ + "preceu.ph.qbl %[q5_l], %[q5] \n\t" \ + "preceu.ph.qbl %[q6_l], %[q6] \n\t" \ + "preceu.ph.qbl %[q7_l], %[q7] \n\t" \ + \ + : [p7_l] "=&r"(p7_l), [p6_l] "=&r"(p6_l), [p5_l] "=&r"(p5_l), \ + [p4_l] "=&r"(p4_l), [q4_l] "=&r"(q4_l), [q5_l] "=&r"(q5_l), \ + [q6_l] "=&r"(q6_l), [q7_l] "=&r"(q7_l) \ + : [p7] "r"(p7), [p6] "r"(p6), [p5] "r"(p5), [p4] "r"(p4), \ + [q4] "r"(q4), [q5] "r"(q5), [q6] "r"(q6), [q7] "r"(q7)); \ + } + +#define PACK_RIGHT_0TO3() \ + { \ + __asm__ __volatile__( \ + "preceu.ph.qbr %[p3_r], %[p3] \n\t" \ + "preceu.ph.qbr %[p2_r], %[p2] \n\t" \ + "preceu.ph.qbr %[p1_r], %[p1] \n\t" \ + "preceu.ph.qbr %[p0_r], %[p0] \n\t" \ + "preceu.ph.qbr %[q0_r], %[q0] \n\t" \ + "preceu.ph.qbr %[q1_r], %[q1] \n\t" \ + "preceu.ph.qbr %[q2_r], %[q2] \n\t" \ + "preceu.ph.qbr %[q3_r], %[q3] \n\t" \ + \ + : [p3_r] "=&r"(p3_r), [p2_r] "=&r"(p2_r), [p1_r] "=&r"(p1_r), \ + [p0_r] "=&r"(p0_r), [q0_r] "=&r"(q0_r), [q1_r] "=&r"(q1_r), \ + [q2_r] "=&r"(q2_r), [q3_r] "=&r"(q3_r) \ + : [p3] "r"(p3), [p2] "r"(p2), [p1] "r"(p1), [p0] "r"(p0), \ + [q0] "r"(q0), [q1] "r"(q1), [q2] "r"(q2), [q3] "r"(q3)); \ + } + +#define PACK_RIGHT_4TO7() \ + { \ + __asm__ __volatile__( \ + "preceu.ph.qbr %[p7_r], %[p7] \n\t" \ + "preceu.ph.qbr %[p6_r], %[p6] \n\t" \ + "preceu.ph.qbr %[p5_r], %[p5] \n\t" \ + "preceu.ph.qbr %[p4_r], %[p4] \n\t" \ + "preceu.ph.qbr %[q4_r], %[q4] \n\t" \ + "preceu.ph.qbr %[q5_r], %[q5] \n\t" \ + "preceu.ph.qbr %[q6_r], %[q6] \n\t" \ + "preceu.ph.qbr %[q7_r], %[q7] \n\t" \ + \ + : [p7_r] "=&r"(p7_r), [p6_r] "=&r"(p6_r), [p5_r] "=&r"(p5_r), \ + [p4_r] "=&r"(p4_r), [q4_r] "=&r"(q4_r), [q5_r] "=&r"(q5_r), \ + [q6_r] "=&r"(q6_r), [q7_r] "=&r"(q7_r) \ + : [p7] "r"(p7), [p6] "r"(p6), [p5] "r"(p5), [p4] "r"(p4), \ + [q4] "r"(q4), [q5] "r"(q5), [q6] "r"(q6), [q7] "r"(q7)); \ + } + +#define COMBINE_LEFT_RIGHT_0TO2() \ + { \ + __asm__ __volatile__( \ + "precr.qb.ph %[p2], %[p2_l], %[p2_r] \n\t" \ + "precr.qb.ph %[p1], %[p1_l], %[p1_r] \n\t" \ + "precr.qb.ph %[p0], %[p0_l], %[p0_r] \n\t" \ + "precr.qb.ph %[q0], %[q0_l], %[q0_r] \n\t" \ + "precr.qb.ph %[q1], %[q1_l], %[q1_r] \n\t" \ + "precr.qb.ph %[q2], %[q2_l], %[q2_r] \n\t" \ + \ + : [p2] "=&r"(p2), [p1] "=&r"(p1), [p0] "=&r"(p0), [q0] "=&r"(q0), \ + [q1] "=&r"(q1), [q2] "=&r"(q2) \ + : [p2_l] "r"(p2_l), [p2_r] "r"(p2_r), [p1_l] "r"(p1_l), \ + [p1_r] "r"(p1_r), [p0_l] "r"(p0_l), [p0_r] "r"(p0_r), \ + [q0_l] "r"(q0_l), [q0_r] "r"(q0_r), [q1_l] "r"(q1_l), \ + [q1_r] "r"(q1_r), [q2_l] "r"(q2_l), [q2_r] "r"(q2_r)); \ + } + +#define COMBINE_LEFT_RIGHT_3TO6() \ + { \ + __asm__ __volatile__( \ + "precr.qb.ph %[p6], %[p6_l], %[p6_r] \n\t" \ + "precr.qb.ph %[p5], %[p5_l], %[p5_r] \n\t" \ + "precr.qb.ph %[p4], %[p4_l], %[p4_r] \n\t" \ + "precr.qb.ph %[p3], %[p3_l], %[p3_r] \n\t" \ + "precr.qb.ph %[q3], %[q3_l], %[q3_r] \n\t" \ + "precr.qb.ph %[q4], %[q4_l], %[q4_r] \n\t" \ + "precr.qb.ph %[q5], %[q5_l], %[q5_r] \n\t" \ + "precr.qb.ph %[q6], %[q6_l], %[q6_r] \n\t" \ + \ + : [p6] "=&r"(p6), [p5] "=&r"(p5), [p4] "=&r"(p4), [p3] "=&r"(p3), \ + [q3] "=&r"(q3), [q4] "=&r"(q4), [q5] "=&r"(q5), [q6] "=&r"(q6) \ + : [p6_l] "r"(p6_l), [p5_l] "r"(p5_l), [p4_l] "r"(p4_l), \ + [p3_l] "r"(p3_l), [p6_r] "r"(p6_r), [p5_r] "r"(p5_r), \ + [p4_r] "r"(p4_r), [p3_r] "r"(p3_r), [q3_l] "r"(q3_l), \ + [q4_l] "r"(q4_l), [q5_l] "r"(q5_l), [q6_l] "r"(q6_l), \ + [q3_r] "r"(q3_r), [q4_r] "r"(q4_r), [q5_r] "r"(q5_r), \ + [q6_r] "r"(q6_r)); \ + } + +#endif // #if HAVE_DSPR2 +#ifdef __cplusplus +} // extern "C" +#endif + +#endif // AOM_AOM_DSP_MIPS_LOOPFILTER_MACROS_DSPR2_H_ diff --git a/third_party/aom/aom_dsp/mips/loopfilter_masks_dspr2.h b/third_party/aom/aom_dsp/mips/loopfilter_masks_dspr2.h new file mode 100644 index 0000000000..a0f57f386a --- /dev/null +++ b/third_party/aom/aom_dsp/mips/loopfilter_masks_dspr2.h @@ -0,0 +1,357 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_MIPS_LOOPFILTER_MASKS_DSPR2_H_ +#define AOM_AOM_DSP_MIPS_LOOPFILTER_MASKS_DSPR2_H_ + +#include + +#include "config/aom_dsp_rtcd.h" + +#include "aom/aom_integer.h" +#include "aom_mem/aom_mem.h" + +#ifdef __cplusplus +extern "C" { +#endif + +#if HAVE_DSPR2 +/* processing 4 pixels at the same time + * compute hev and mask in the same function */ +static INLINE void filter_hev_mask_dspr2(uint32_t limit, uint32_t flimit, + uint32_t p1, uint32_t p0, uint32_t p3, + uint32_t p2, uint32_t q0, uint32_t q1, + uint32_t q2, uint32_t q3, + uint32_t thresh, uint32_t *hev, + uint32_t *mask) { + uint32_t c, r, r3, r_k; + uint32_t s1, s2, s3; + uint32_t ones = 0xFFFFFFFF; + uint32_t hev1; + + __asm__ __volatile__( + /* mask |= (abs(p3 - p2) > limit) */ + "subu_s.qb %[c], %[p3], %[p2] \n\t" + "subu_s.qb %[r_k], %[p2], %[p3] \n\t" + "or %[r_k], %[r_k], %[c] \n\t" + "cmpgu.lt.qb %[c], %[limit], %[r_k] \n\t" + "or %[r], $0, %[c] \n\t" + + /* mask |= (abs(p2 - p1) > limit) */ + "subu_s.qb %[c], %[p2], %[p1] \n\t" + "subu_s.qb %[r_k], %[p1], %[p2] \n\t" + "or %[r_k], %[r_k], %[c] \n\t" + "cmpgu.lt.qb %[c], %[limit], %[r_k] \n\t" + "or %[r], %[r], %[c] \n\t" + + /* mask |= (abs(p1 - p0) > limit) + * hev |= (abs(p1 - p0) > thresh) + */ + "subu_s.qb %[c], %[p1], %[p0] \n\t" + "subu_s.qb %[r_k], %[p0], %[p1] \n\t" + "or %[r_k], %[r_k], %[c] \n\t" + "cmpgu.lt.qb %[c], %[thresh], %[r_k] \n\t" + "or %[r3], $0, %[c] \n\t" + "cmpgu.lt.qb %[c], %[limit], %[r_k] \n\t" + "or %[r], %[r], %[c] \n\t" + + /* mask |= (abs(q1 - q0) > limit) + * hev |= (abs(q1 - q0) > thresh) + */ + "subu_s.qb %[c], %[q1], %[q0] \n\t" + "subu_s.qb %[r_k], %[q0], %[q1] \n\t" + "or %[r_k], %[r_k], %[c] \n\t" + "cmpgu.lt.qb %[c], %[thresh], %[r_k] \n\t" + "or %[r3], %[r3], %[c] \n\t" + "cmpgu.lt.qb %[c], %[limit], %[r_k] \n\t" + "or %[r], %[r], %[c] \n\t" + + /* mask |= (abs(q2 - q1) > limit) */ + "subu_s.qb %[c], %[q2], %[q1] \n\t" + "subu_s.qb %[r_k], %[q1], %[q2] \n\t" + "or %[r_k], %[r_k], %[c] \n\t" + "cmpgu.lt.qb %[c], %[limit], %[r_k] \n\t" + "or %[r], %[r], %[c] \n\t" + "sll %[r3], %[r3], 24 \n\t" + + /* mask |= (abs(q3 - q2) > limit) */ + "subu_s.qb %[c], %[q3], %[q2] \n\t" + "subu_s.qb %[r_k], %[q2], %[q3] \n\t" + "or %[r_k], %[r_k], %[c] \n\t" + "cmpgu.lt.qb %[c], %[limit], %[r_k] \n\t" + "or %[r], %[r], %[c] \n\t" + + : [c] "=&r"(c), [r_k] "=&r"(r_k), [r] "=&r"(r), [r3] "=&r"(r3) + : [limit] "r"(limit), [p3] "r"(p3), [p2] "r"(p2), [p1] "r"(p1), + [p0] "r"(p0), [q1] "r"(q1), [q0] "r"(q0), [q2] "r"(q2), [q3] "r"(q3), + [thresh] "r"(thresh)); + + __asm__ __volatile__( + /* abs(p0 - q0) */ + "subu_s.qb %[c], %[p0], %[q0] \n\t" + "subu_s.qb %[r_k], %[q0], %[p0] \n\t" + "wrdsp %[r3] \n\t" + "or %[s1], %[r_k], %[c] \n\t" + + /* abs(p1 - q1) */ + "subu_s.qb %[c], %[p1], %[q1] \n\t" + "addu_s.qb %[s3], %[s1], %[s1] \n\t" + "pick.qb %[hev1], %[ones], $0 \n\t" + "subu_s.qb %[r_k], %[q1], %[p1] \n\t" + "or %[s2], %[r_k], %[c] \n\t" + + /* abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > flimit * 2 + limit */ + "shrl.qb %[s2], %[s2], 1 \n\t" + "addu_s.qb %[s1], %[s2], %[s3] \n\t" + "cmpgu.lt.qb %[c], %[flimit], %[s1] \n\t" + "or %[r], %[r], %[c] \n\t" + "sll %[r], %[r], 24 \n\t" + + "wrdsp %[r] \n\t" + "pick.qb %[s2], $0, %[ones] \n\t" + + : [c] "=&r"(c), [r_k] "=&r"(r_k), [s1] "=&r"(s1), [hev1] "=&r"(hev1), + [s2] "=&r"(s2), [r] "+r"(r), [s3] "=&r"(s3) + : [p0] "r"(p0), [q0] "r"(q0), [p1] "r"(p1), [r3] "r"(r3), [q1] "r"(q1), + [ones] "r"(ones), [flimit] "r"(flimit)); + + *hev = hev1; + *mask = s2; +} + +static INLINE void filter_hev_mask_flatmask4_dspr2( + uint32_t limit, uint32_t flimit, uint32_t thresh, uint32_t p1, uint32_t p0, + uint32_t p3, uint32_t p2, uint32_t q0, uint32_t q1, uint32_t q2, + uint32_t q3, uint32_t *hev, uint32_t *mask, uint32_t *flat) { + uint32_t c, r, r3, r_k, r_flat; + uint32_t s1, s2, s3; + uint32_t ones = 0xFFFFFFFF; + uint32_t flat_thresh = 0x01010101; + uint32_t hev1; + uint32_t flat1; + + __asm__ __volatile__( + /* mask |= (abs(p3 - p2) > limit) */ + "subu_s.qb %[c], %[p3], %[p2] \n\t" + "subu_s.qb %[r_k], %[p2], %[p3] \n\t" + "or %[r_k], %[r_k], %[c] \n\t" + "cmpgu.lt.qb %[c], %[limit], %[r_k] \n\t" + "or %[r], $0, %[c] \n\t" + + /* mask |= (abs(p2 - p1) > limit) */ + "subu_s.qb %[c], %[p2], %[p1] \n\t" + "subu_s.qb %[r_k], %[p1], %[p2] \n\t" + "or %[r_k], %[r_k], %[c] \n\t" + "cmpgu.lt.qb %[c], %[limit], %[r_k] \n\t" + "or %[r], %[r], %[c] \n\t" + + /* mask |= (abs(p1 - p0) > limit) + * hev |= (abs(p1 - p0) > thresh) + * flat |= (abs(p1 - p0) > thresh) + */ + "subu_s.qb %[c], %[p1], %[p0] \n\t" + "subu_s.qb %[r_k], %[p0], %[p1] \n\t" + "or %[r_k], %[r_k], %[c] \n\t" + "cmpgu.lt.qb %[c], %[thresh], %[r_k] \n\t" + "or %[r3], $0, %[c] \n\t" + "cmpgu.lt.qb %[c], %[limit], %[r_k] \n\t" + "or %[r], %[r], %[c] \n\t" + "cmpgu.lt.qb %[c], %[flat_thresh], %[r_k] \n\t" + "or %[r_flat], $0, %[c] \n\t" + + /* mask |= (abs(q1 - q0) > limit) + * hev |= (abs(q1 - q0) > thresh) + * flat |= (abs(q1 - q0) > thresh) + */ + "subu_s.qb %[c], %[q1], %[q0] \n\t" + "subu_s.qb %[r_k], %[q0], %[q1] \n\t" + "or %[r_k], %[r_k], %[c] \n\t" + "cmpgu.lt.qb %[c], %[thresh], %[r_k] \n\t" + "or %[r3], %[r3], %[c] \n\t" + "cmpgu.lt.qb %[c], %[limit], %[r_k] \n\t" + "or %[r], %[r], %[c] \n\t" + "cmpgu.lt.qb %[c], %[flat_thresh], %[r_k] \n\t" + "or %[r_flat], %[r_flat], %[c] \n\t" + + /* flat |= (abs(p0 - p2) > thresh) */ + "subu_s.qb %[c], %[p0], %[p2] \n\t" + "subu_s.qb %[r_k], %[p2], %[p0] \n\t" + "or %[r_k], %[r_k], %[c] \n\t" + "cmpgu.lt.qb %[c], %[flat_thresh], %[r_k] \n\t" + "or %[r_flat], %[r_flat], %[c] \n\t" + + /* flat |= (abs(q0 - q2) > thresh) */ + "subu_s.qb %[c], %[q0], %[q2] \n\t" + "subu_s.qb %[r_k], %[q2], %[q0] \n\t" + "or %[r_k], %[r_k], %[c] \n\t" + "cmpgu.lt.qb %[c], %[flat_thresh], %[r_k] \n\t" + "or %[r_flat], %[r_flat], %[c] \n\t" + + /* flat |= (abs(p3 - p0) > thresh) */ + "subu_s.qb %[c], %[p3], %[p0] \n\t" + "subu_s.qb %[r_k], %[p0], %[p3] \n\t" + "or %[r_k], %[r_k], %[c] \n\t" + "cmpgu.lt.qb %[c], %[flat_thresh], %[r_k] \n\t" + "or %[r_flat], %[r_flat], %[c] \n\t" + + /* flat |= (abs(q3 - q0) > thresh) */ + "subu_s.qb %[c], %[q3], %[q0] \n\t" + "subu_s.qb %[r_k], %[q0], %[q3] \n\t" + "or %[r_k], %[r_k], %[c] \n\t" + "cmpgu.lt.qb %[c], %[flat_thresh], %[r_k] \n\t" + "or %[r_flat], %[r_flat], %[c] \n\t" + "sll %[r_flat], %[r_flat], 24 \n\t" + /* look at stall here */ + "wrdsp %[r_flat] \n\t" + "pick.qb %[flat1], $0, %[ones] \n\t" + + /* mask |= (abs(q2 - q1) > limit) */ + "subu_s.qb %[c], %[q2], %[q1] \n\t" + "subu_s.qb %[r_k], %[q1], %[q2] \n\t" + "or %[r_k], %[r_k], %[c] \n\t" + "cmpgu.lt.qb %[c], %[limit], %[r_k] \n\t" + "or %[r], %[r], %[c] \n\t" + "sll %[r3], %[r3], 24 \n\t" + + /* mask |= (abs(q3 - q2) > limit) */ + "subu_s.qb %[c], %[q3], %[q2] \n\t" + "subu_s.qb %[r_k], %[q2], %[q3] \n\t" + "or %[r_k], %[r_k], %[c] \n\t" + "cmpgu.lt.qb %[c], %[limit], %[r_k] \n\t" + "or %[r], %[r], %[c] \n\t" + + : [c] "=&r"(c), [r_k] "=&r"(r_k), [r] "=&r"(r), [r3] "=&r"(r3), + [r_flat] "=&r"(r_flat), [flat1] "=&r"(flat1) + : [limit] "r"(limit), [p3] "r"(p3), [p2] "r"(p2), [p1] "r"(p1), + [p0] "r"(p0), [q1] "r"(q1), [q0] "r"(q0), [q2] "r"(q2), [q3] "r"(q3), + [thresh] "r"(thresh), [flat_thresh] "r"(flat_thresh), [ones] "r"(ones)); + + __asm__ __volatile__( + /* abs(p0 - q0) */ + "subu_s.qb %[c], %[p0], %[q0] \n\t" + "subu_s.qb %[r_k], %[q0], %[p0] \n\t" + "wrdsp %[r3] \n\t" + "or %[s1], %[r_k], %[c] \n\t" + + /* abs(p1 - q1) */ + "subu_s.qb %[c], %[p1], %[q1] \n\t" + "addu_s.qb %[s3], %[s1], %[s1] \n\t" + "pick.qb %[hev1], %[ones], $0 \n\t" + "subu_s.qb %[r_k], %[q1], %[p1] \n\t" + "or %[s2], %[r_k], %[c] \n\t" + + /* abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > flimit * 2 + limit */ + "shrl.qb %[s2], %[s2], 1 \n\t" + "addu_s.qb %[s1], %[s2], %[s3] \n\t" + "cmpgu.lt.qb %[c], %[flimit], %[s1] \n\t" + "or %[r], %[r], %[c] \n\t" + "sll %[r], %[r], 24 \n\t" + + "wrdsp %[r] \n\t" + "pick.qb %[s2], $0, %[ones] \n\t" + + : [c] "=&r"(c), [r_k] "=&r"(r_k), [s1] "=&r"(s1), [hev1] "=&r"(hev1), + [s2] "=&r"(s2), [r] "+r"(r), [s3] "=&r"(s3) + : [p0] "r"(p0), [q0] "r"(q0), [p1] "r"(p1), [r3] "r"(r3), [q1] "r"(q1), + [ones] "r"(ones), [flimit] "r"(flimit)); + + *hev = hev1; + *mask = s2; + *flat = flat1; +} + +static INLINE void flatmask5(uint32_t p4, uint32_t p3, uint32_t p2, uint32_t p1, + uint32_t p0, uint32_t q0, uint32_t q1, uint32_t q2, + uint32_t q3, uint32_t q4, uint32_t *flat2) { + uint32_t c, r, r_k, r_flat; + uint32_t ones = 0xFFFFFFFF; + uint32_t flat_thresh = 0x01010101; + uint32_t flat1, flat3; + + __asm__ __volatile__( + /* flat |= (abs(p4 - p0) > thresh) */ + "subu_s.qb %[c], %[p4], %[p0] \n\t" + "subu_s.qb %[r_k], %[p0], %[p4] \n\t" + "or %[r_k], %[r_k], %[c] \n\t" + "cmpgu.lt.qb %[c], %[flat_thresh], %[r_k] \n\t" + "or %[r], $0, %[c] \n\t" + + /* flat |= (abs(q4 - q0) > thresh) */ + "subu_s.qb %[c], %[q4], %[q0] \n\t" + "subu_s.qb %[r_k], %[q0], %[q4] \n\t" + "or %[r_k], %[r_k], %[c] \n\t" + "cmpgu.lt.qb %[c], %[flat_thresh], %[r_k] \n\t" + "or %[r], %[r], %[c] \n\t" + "sll %[r], %[r], 24 \n\t" + "wrdsp %[r] \n\t" + "pick.qb %[flat3], $0, %[ones] \n\t" + + /* flat |= (abs(p1 - p0) > thresh) */ + "subu_s.qb %[c], %[p1], %[p0] \n\t" + "subu_s.qb %[r_k], %[p0], %[p1] \n\t" + "or %[r_k], %[r_k], %[c] \n\t" + "cmpgu.lt.qb %[c], %[flat_thresh], %[r_k] \n\t" + "or %[r_flat], $0, %[c] \n\t" + + /* flat |= (abs(q1 - q0) > thresh) */ + "subu_s.qb %[c], %[q1], %[q0] \n\t" + "subu_s.qb %[r_k], %[q0], %[q1] \n\t" + "or %[r_k], %[r_k], %[c] \n\t" + "cmpgu.lt.qb %[c], %[flat_thresh], %[r_k] \n\t" + "or %[r_flat], %[r_flat], %[c] \n\t" + + /* flat |= (abs(p0 - p2) > thresh) */ + "subu_s.qb %[c], %[p0], %[p2] \n\t" + "subu_s.qb %[r_k], %[p2], %[p0] \n\t" + "or %[r_k], %[r_k], %[c] \n\t" + "cmpgu.lt.qb %[c], %[flat_thresh], %[r_k] \n\t" + "or %[r_flat], %[r_flat], %[c] \n\t" + + /* flat |= (abs(q0 - q2) > thresh) */ + "subu_s.qb %[c], %[q0], %[q2] \n\t" + "subu_s.qb %[r_k], %[q2], %[q0] \n\t" + "or %[r_k], %[r_k], %[c] \n\t" + "cmpgu.lt.qb %[c], %[flat_thresh], %[r_k] \n\t" + "or %[r_flat], %[r_flat], %[c] \n\t" + + /* flat |= (abs(p3 - p0) > thresh) */ + "subu_s.qb %[c], %[p3], %[p0] \n\t" + "subu_s.qb %[r_k], %[p0], %[p3] \n\t" + "or %[r_k], %[r_k], %[c] \n\t" + "cmpgu.lt.qb %[c], %[flat_thresh], %[r_k] \n\t" + "or %[r_flat], %[r_flat], %[c] \n\t" + + /* flat |= (abs(q3 - q0) > thresh) */ + "subu_s.qb %[c], %[q3], %[q0] \n\t" + "subu_s.qb %[r_k], %[q0], %[q3] \n\t" + "or %[r_k], %[r_k], %[c] \n\t" + "cmpgu.lt.qb %[c], %[flat_thresh], %[r_k] \n\t" + "or %[r_flat], %[r_flat], %[c] \n\t" + "sll %[r_flat], %[r_flat], 24 \n\t" + "wrdsp %[r_flat] \n\t" + "pick.qb %[flat1], $0, %[ones] \n\t" + /* flat & flatmask4(thresh, p3, p2, p1, p0, q0, q1, q2, q3) */ + "and %[flat1], %[flat3], %[flat1] \n\t" + + : [c] "=&r"(c), [r_k] "=&r"(r_k), [r] "=&r"(r), [r_flat] "=&r"(r_flat), + [flat1] "=&r"(flat1), [flat3] "=&r"(flat3) + : [p4] "r"(p4), [p3] "r"(p3), [p2] "r"(p2), [p1] "r"(p1), [p0] "r"(p0), + [q0] "r"(q0), [q1] "r"(q1), [q2] "r"(q2), [q3] "r"(q3), [q4] "r"(q4), + [flat_thresh] "r"(flat_thresh), [ones] "r"(ones)); + + *flat2 = flat1; +} +#endif // #if HAVE_DSPR2 +#ifdef __cplusplus +} // extern "C" +#endif + +#endif // AOM_AOM_DSP_MIPS_LOOPFILTER_MASKS_DSPR2_H_ diff --git a/third_party/aom/aom_dsp/mips/loopfilter_mb_dspr2.c b/third_party/aom/aom_dsp/mips/loopfilter_mb_dspr2.c new file mode 100644 index 0000000000..b67ccfe9d8 --- /dev/null +++ b/third_party/aom/aom_dsp/mips/loopfilter_mb_dspr2.c @@ -0,0 +1,590 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include + +#include "config/aom_dsp_rtcd.h" + +#include "aom/aom_integer.h" +#include "aom_dsp/mips/common_dspr2.h" +#include "aom_dsp/mips/loopfilter_filters_dspr2.h" +#include "aom_dsp/mips/loopfilter_macros_dspr2.h" +#include "aom_dsp/mips/loopfilter_masks_dspr2.h" +#include "aom_mem/aom_mem.h" + +#if HAVE_DSPR2 +void aom_lpf_horizontal_8_dspr2(unsigned char *s, int pitch, + const uint8_t *blimit, const uint8_t *limit, + const uint8_t *thresh) { + uint32_t mask; + uint32_t hev, flat; + uint8_t i; + uint8_t *sp3, *sp2, *sp1, *sp0, *sq0, *sq1, *sq2, *sq3; + uint32_t thresh_vec, flimit_vec, limit_vec; + uint32_t uflimit, ulimit, uthresh; + uint32_t p1_f0, p0_f0, q0_f0, q1_f0; + uint32_t p3, p2, p1, p0, q0, q1, q2, q3; + uint32_t p0_l, p1_l, p2_l, p3_l, q0_l, q1_l, q2_l, q3_l; + uint32_t p0_r, p1_r, p2_r, p3_r, q0_r, q1_r, q2_r, q3_r; + + uflimit = *blimit; + ulimit = *limit; + uthresh = *thresh; + + /* create quad-byte */ + __asm__ __volatile__( + "replv.qb %[thresh_vec], %[uthresh] \n\t" + "replv.qb %[flimit_vec], %[uflimit] \n\t" + "replv.qb %[limit_vec], %[ulimit] \n\t" + + : [thresh_vec] "=&r"(thresh_vec), [flimit_vec] "=&r"(flimit_vec), + [limit_vec] "=r"(limit_vec) + : [uthresh] "r"(uthresh), [uflimit] "r"(uflimit), [ulimit] "r"(ulimit)); + + /* prefetch data for store */ + prefetch_store(s); + + for (i = 0; i < 2; i++) { + sp3 = s - (pitch << 2); + sp2 = sp3 + pitch; + sp1 = sp2 + pitch; + sp0 = sp1 + pitch; + sq0 = s; + sq1 = s + pitch; + sq2 = sq1 + pitch; + sq3 = sq2 + pitch; + + __asm__ __volatile__( + "lw %[p3], (%[sp3]) \n\t" + "lw %[p2], (%[sp2]) \n\t" + "lw %[p1], (%[sp1]) \n\t" + "lw %[p0], (%[sp0]) \n\t" + "lw %[q0], (%[sq0]) \n\t" + "lw %[q1], (%[sq1]) \n\t" + "lw %[q2], (%[sq2]) \n\t" + "lw %[q3], (%[sq3]) \n\t" + + : [p3] "=&r"(p3), [p2] "=&r"(p2), [p1] "=&r"(p1), [p0] "=&r"(p0), + [q3] "=&r"(q3), [q2] "=&r"(q2), [q1] "=&r"(q1), [q0] "=&r"(q0) + : [sp3] "r"(sp3), [sp2] "r"(sp2), [sp1] "r"(sp1), [sp0] "r"(sp0), + [sq3] "r"(sq3), [sq2] "r"(sq2), [sq1] "r"(sq1), [sq0] "r"(sq0)); + + filter_hev_mask_flatmask4_dspr2(limit_vec, flimit_vec, thresh_vec, p1, p0, + p3, p2, q0, q1, q2, q3, &hev, &mask, &flat); + + if ((flat == 0) && (mask != 0)) { + filter1_dspr2(mask, hev, p1, p0, q0, q1, &p1_f0, &p0_f0, &q0_f0, &q1_f0); + + __asm__ __volatile__( + "sw %[p1_f0], (%[sp1]) \n\t" + "sw %[p0_f0], (%[sp0]) \n\t" + "sw %[q0_f0], (%[sq0]) \n\t" + "sw %[q1_f0], (%[sq1]) \n\t" + + : + : [p1_f0] "r"(p1_f0), [p0_f0] "r"(p0_f0), [q0_f0] "r"(q0_f0), + [q1_f0] "r"(q1_f0), [sp1] "r"(sp1), [sp0] "r"(sp0), [sq0] "r"(sq0), + [sq1] "r"(sq1)); + } else if ((mask & flat) == 0xFFFFFFFF) { + /* left 2 element operation */ + PACK_LEFT_0TO3() + mbfilter_dspr2(&p3_l, &p2_l, &p1_l, &p0_l, &q0_l, &q1_l, &q2_l, &q3_l); + + /* right 2 element operation */ + PACK_RIGHT_0TO3() + mbfilter_dspr2(&p3_r, &p2_r, &p1_r, &p0_r, &q0_r, &q1_r, &q2_r, &q3_r); + + COMBINE_LEFT_RIGHT_0TO2() + + __asm__ __volatile__( + "sw %[p2], (%[sp2]) \n\t" + "sw %[p1], (%[sp1]) \n\t" + "sw %[p0], (%[sp0]) \n\t" + "sw %[q0], (%[sq0]) \n\t" + "sw %[q1], (%[sq1]) \n\t" + "sw %[q2], (%[sq2]) \n\t" + + : + : [p2] "r"(p2), [p1] "r"(p1), [p0] "r"(p0), [q0] "r"(q0), + [q1] "r"(q1), [q2] "r"(q2), [sp2] "r"(sp2), [sp1] "r"(sp1), + [sp0] "r"(sp0), [sq0] "r"(sq0), [sq1] "r"(sq1), [sq2] "r"(sq2)); + } else if ((flat != 0) && (mask != 0)) { + /* filtering */ + filter1_dspr2(mask, hev, p1, p0, q0, q1, &p1_f0, &p0_f0, &q0_f0, &q1_f0); + + /* left 2 element operation */ + PACK_LEFT_0TO3() + mbfilter_dspr2(&p3_l, &p2_l, &p1_l, &p0_l, &q0_l, &q1_l, &q2_l, &q3_l); + + /* right 2 element operation */ + PACK_RIGHT_0TO3() + mbfilter_dspr2(&p3_r, &p2_r, &p1_r, &p0_r, &q0_r, &q1_r, &q2_r, &q3_r); + + if (mask & flat & 0x000000FF) { + __asm__ __volatile__( + "sb %[p2_r], (%[sp2]) \n\t" + "sb %[p1_r], (%[sp1]) \n\t" + "sb %[p0_r], (%[sp0]) \n\t" + "sb %[q0_r], (%[sq0]) \n\t" + "sb %[q1_r], (%[sq1]) \n\t" + "sb %[q2_r], (%[sq2]) \n\t" + + : + : [p2_r] "r"(p2_r), [p1_r] "r"(p1_r), [p0_r] "r"(p0_r), + [q0_r] "r"(q0_r), [q1_r] "r"(q1_r), [q2_r] "r"(q2_r), + [sp2] "r"(sp2), [sp1] "r"(sp1), [sp0] "r"(sp0), [sq0] "r"(sq0), + [sq1] "r"(sq1), [sq2] "r"(sq2)); + } else if (mask & 0x000000FF) { + __asm__ __volatile__( + "sb %[p1_f0], (%[sp1]) \n\t" + "sb %[p0_f0], (%[sp0]) \n\t" + "sb %[q0_f0], (%[sq0]) \n\t" + "sb %[q1_f0], (%[sq1]) \n\t" + + : + : [p1_f0] "r"(p1_f0), [p0_f0] "r"(p0_f0), [q0_f0] "r"(q0_f0), + [q1_f0] "r"(q1_f0), [sp1] "r"(sp1), [sp0] "r"(sp0), + [sq0] "r"(sq0), [sq1] "r"(sq1)); + } + + __asm__ __volatile__( + "srl %[p2_r], %[p2_r], 16 \n\t" + "srl %[p1_r], %[p1_r], 16 \n\t" + "srl %[p0_r], %[p0_r], 16 \n\t" + "srl %[q0_r], %[q0_r], 16 \n\t" + "srl %[q1_r], %[q1_r], 16 \n\t" + "srl %[q2_r], %[q2_r], 16 \n\t" + "srl %[p1_f0], %[p1_f0], 8 \n\t" + "srl %[p0_f0], %[p0_f0], 8 \n\t" + "srl %[q0_f0], %[q0_f0], 8 \n\t" + "srl %[q1_f0], %[q1_f0], 8 \n\t" + + : [p2_r] "+r"(p2_r), [p1_r] "+r"(p1_r), [p0_r] "+r"(p0_r), + [q0_r] "+r"(q0_r), [q1_r] "+r"(q1_r), [q2_r] "+r"(q2_r), + [p1_f0] "+r"(p1_f0), [p0_f0] "+r"(p0_f0), [q0_f0] "+r"(q0_f0), + [q1_f0] "+r"(q1_f0) + :); + + if (mask & flat & 0x0000FF00) { + __asm__ __volatile__( + "sb %[p2_r], +1(%[sp2]) \n\t" + "sb %[p1_r], +1(%[sp1]) \n\t" + "sb %[p0_r], +1(%[sp0]) \n\t" + "sb %[q0_r], +1(%[sq0]) \n\t" + "sb %[q1_r], +1(%[sq1]) \n\t" + "sb %[q2_r], +1(%[sq2]) \n\t" + + : + : [p2_r] "r"(p2_r), [p1_r] "r"(p1_r), [p0_r] "r"(p0_r), + [q0_r] "r"(q0_r), [q1_r] "r"(q1_r), [q2_r] "r"(q2_r), + [sp2] "r"(sp2), [sp1] "r"(sp1), [sp0] "r"(sp0), [sq0] "r"(sq0), + [sq1] "r"(sq1), [sq2] "r"(sq2)); + } else if (mask & 0x0000FF00) { + __asm__ __volatile__( + "sb %[p1_f0], +1(%[sp1]) \n\t" + "sb %[p0_f0], +1(%[sp0]) \n\t" + "sb %[q0_f0], +1(%[sq0]) \n\t" + "sb %[q1_f0], +1(%[sq1]) \n\t" + + : + : [p1_f0] "r"(p1_f0), [p0_f0] "r"(p0_f0), [q0_f0] "r"(q0_f0), + [q1_f0] "r"(q1_f0), [sp1] "r"(sp1), [sp0] "r"(sp0), + [sq0] "r"(sq0), [sq1] "r"(sq1)); + } + + __asm__ __volatile__( + "srl %[p1_f0], %[p1_f0], 8 \n\t" + "srl %[p0_f0], %[p0_f0], 8 \n\t" + "srl %[q0_f0], %[q0_f0], 8 \n\t" + "srl %[q1_f0], %[q1_f0], 8 \n\t" + + : [p2] "+r"(p2), [p1] "+r"(p1), [p0] "+r"(p0), [q0] "+r"(q0), + [q1] "+r"(q1), [q2] "+r"(q2), [p1_f0] "+r"(p1_f0), + [p0_f0] "+r"(p0_f0), [q0_f0] "+r"(q0_f0), [q1_f0] "+r"(q1_f0) + :); + + if (mask & flat & 0x00FF0000) { + __asm__ __volatile__( + "sb %[p2_l], +2(%[sp2]) \n\t" + "sb %[p1_l], +2(%[sp1]) \n\t" + "sb %[p0_l], +2(%[sp0]) \n\t" + "sb %[q0_l], +2(%[sq0]) \n\t" + "sb %[q1_l], +2(%[sq1]) \n\t" + "sb %[q2_l], +2(%[sq2]) \n\t" + + : + : [p2_l] "r"(p2_l), [p1_l] "r"(p1_l), [p0_l] "r"(p0_l), + [q0_l] "r"(q0_l), [q1_l] "r"(q1_l), [q2_l] "r"(q2_l), + [sp2] "r"(sp2), [sp1] "r"(sp1), [sp0] "r"(sp0), [sq0] "r"(sq0), + [sq1] "r"(sq1), [sq2] "r"(sq2)); + } else if (mask & 0x00FF0000) { + __asm__ __volatile__( + "sb %[p1_f0], +2(%[sp1]) \n\t" + "sb %[p0_f0], +2(%[sp0]) \n\t" + "sb %[q0_f0], +2(%[sq0]) \n\t" + "sb %[q1_f0], +2(%[sq1]) \n\t" + + : + : [p1_f0] "r"(p1_f0), [p0_f0] "r"(p0_f0), [q0_f0] "r"(q0_f0), + [q1_f0] "r"(q1_f0), [sp1] "r"(sp1), [sp0] "r"(sp0), + [sq0] "r"(sq0), [sq1] "r"(sq1)); + } + + __asm__ __volatile__( + "srl %[p2_l], %[p2_l], 16 \n\t" + "srl %[p1_l], %[p1_l], 16 \n\t" + "srl %[p0_l], %[p0_l], 16 \n\t" + "srl %[q0_l], %[q0_l], 16 \n\t" + "srl %[q1_l], %[q1_l], 16 \n\t" + "srl %[q2_l], %[q2_l], 16 \n\t" + "srl %[p1_f0], %[p1_f0], 8 \n\t" + "srl %[p0_f0], %[p0_f0], 8 \n\t" + "srl %[q0_f0], %[q0_f0], 8 \n\t" + "srl %[q1_f0], %[q1_f0], 8 \n\t" + + : [p2_l] "+r"(p2_l), [p1_l] "+r"(p1_l), [p0_l] "+r"(p0_l), + [q0_l] "+r"(q0_l), [q1_l] "+r"(q1_l), [q2_l] "+r"(q2_l), + [p1_f0] "+r"(p1_f0), [p0_f0] "+r"(p0_f0), [q0_f0] "+r"(q0_f0), + [q1_f0] "+r"(q1_f0) + :); + + if (mask & flat & 0xFF000000) { + __asm__ __volatile__( + "sb %[p2_l], +3(%[sp2]) \n\t" + "sb %[p1_l], +3(%[sp1]) \n\t" + "sb %[p0_l], +3(%[sp0]) \n\t" + "sb %[q0_l], +3(%[sq0]) \n\t" + "sb %[q1_l], +3(%[sq1]) \n\t" + "sb %[q2_l], +3(%[sq2]) \n\t" + + : + : [p2_l] "r"(p2_l), [p1_l] "r"(p1_l), [p0_l] "r"(p0_l), + [q0_l] "r"(q0_l), [q1_l] "r"(q1_l), [q2_l] "r"(q2_l), + [sp2] "r"(sp2), [sp1] "r"(sp1), [sp0] "r"(sp0), [sq0] "r"(sq0), + [sq1] "r"(sq1), [sq2] "r"(sq2)); + } else if (mask & 0xFF000000) { + __asm__ __volatile__( + "sb %[p1_f0], +3(%[sp1]) \n\t" + "sb %[p0_f0], +3(%[sp0]) \n\t" + "sb %[q0_f0], +3(%[sq0]) \n\t" + "sb %[q1_f0], +3(%[sq1]) \n\t" + + : + : [p1_f0] "r"(p1_f0), [p0_f0] "r"(p0_f0), [q0_f0] "r"(q0_f0), + [q1_f0] "r"(q1_f0), [sp1] "r"(sp1), [sp0] "r"(sp0), + [sq0] "r"(sq0), [sq1] "r"(sq1)); + } + } + + s = s + 4; + } +} + +void aom_lpf_vertical_8_dspr2(unsigned char *s, int pitch, + const uint8_t *blimit, const uint8_t *limit, + const uint8_t *thresh) { + uint8_t i; + uint32_t mask, hev, flat; + uint8_t *s1, *s2, *s3, *s4; + uint32_t prim1, prim2, sec3, sec4, prim3, prim4; + uint32_t thresh_vec, flimit_vec, limit_vec; + uint32_t uflimit, ulimit, uthresh; + uint32_t p3, p2, p1, p0, q3, q2, q1, q0; + uint32_t p1_f0, p0_f0, q0_f0, q1_f0; + uint32_t p0_l, p1_l, p2_l, p3_l, q0_l, q1_l, q2_l, q3_l; + uint32_t p0_r, p1_r, p2_r, p3_r, q0_r, q1_r, q2_r, q3_r; + + uflimit = *blimit; + ulimit = *limit; + uthresh = *thresh; + + /* create quad-byte */ + __asm__ __volatile__( + "replv.qb %[thresh_vec], %[uthresh] \n\t" + "replv.qb %[flimit_vec], %[uflimit] \n\t" + "replv.qb %[limit_vec], %[ulimit] \n\t" + + : [thresh_vec] "=&r"(thresh_vec), [flimit_vec] "=&r"(flimit_vec), + [limit_vec] "=r"(limit_vec) + : [uthresh] "r"(uthresh), [uflimit] "r"(uflimit), [ulimit] "r"(ulimit)); + + prefetch_store(s + pitch); + + for (i = 0; i < 2; i++) { + s1 = s; + s2 = s + pitch; + s3 = s2 + pitch; + s4 = s3 + pitch; + s = s4 + pitch; + + __asm__ __volatile__( + "lw %[p0], -4(%[s1]) \n\t" + "lw %[p1], -4(%[s2]) \n\t" + "lw %[p2], -4(%[s3]) \n\t" + "lw %[p3], -4(%[s4]) \n\t" + "lw %[q3], (%[s1]) \n\t" + "lw %[q2], (%[s2]) \n\t" + "lw %[q1], (%[s3]) \n\t" + "lw %[q0], (%[s4]) \n\t" + + : [p3] "=&r"(p3), [p2] "=&r"(p2), [p1] "=&r"(p1), [p0] "=&r"(p0), + [q0] "=&r"(q0), [q1] "=&r"(q1), [q2] "=&r"(q2), [q3] "=&r"(q3) + : [s1] "r"(s1), [s2] "r"(s2), [s3] "r"(s3), [s4] "r"(s4)); + + /* transpose p3, p2, p1, p0 + original (when loaded from memory) + register -4 -3 -2 -1 + p0 p0_0 p0_1 p0_2 p0_3 + p1 p1_0 p1_1 p1_2 p1_3 + p2 p2_0 p2_1 p2_2 p2_3 + p3 p3_0 p3_1 p3_2 p3_3 + + after transpose + register + p0 p3_3 p2_3 p1_3 p0_3 + p1 p3_2 p2_2 p1_2 p0_2 + p2 p3_1 p2_1 p1_1 p0_1 + p3 p3_0 p2_0 p1_0 p0_0 + */ + __asm__ __volatile__( + "precrq.qb.ph %[prim1], %[p0], %[p1] \n\t" + "precr.qb.ph %[prim2], %[p0], %[p1] \n\t" + "precrq.qb.ph %[prim3], %[p2], %[p3] \n\t" + "precr.qb.ph %[prim4], %[p2], %[p3] \n\t" + + "precrq.qb.ph %[p1], %[prim1], %[prim2] \n\t" + "precr.qb.ph %[p3], %[prim1], %[prim2] \n\t" + "precrq.qb.ph %[sec3], %[prim3], %[prim4] \n\t" + "precr.qb.ph %[sec4], %[prim3], %[prim4] \n\t" + + "precrq.ph.w %[p0], %[p1], %[sec3] \n\t" + "precrq.ph.w %[p2], %[p3], %[sec4] \n\t" + "append %[p1], %[sec3], 16 \n\t" + "append %[p3], %[sec4], 16 \n\t" + + : [prim1] "=&r"(prim1), [prim2] "=&r"(prim2), [prim3] "=&r"(prim3), + [prim4] "=&r"(prim4), [p0] "+r"(p0), [p1] "+r"(p1), [p2] "+r"(p2), + [p3] "+r"(p3), [sec3] "=&r"(sec3), [sec4] "=&r"(sec4) + :); + + /* transpose q0, q1, q2, q3 + original (when loaded from memory) + register +1 +2 +3 +4 + q3 q3_0 q3_1 q3_2 q3_3 + q2 q2_0 q2_1 q2_2 q2_3 + q1 q1_0 q1_1 q1_2 q1_3 + q0 q0_0 q0_1 q0_2 q0_3 + + after transpose + register + q3 q0_3 q1_3 q2_3 q3_3 + q2 q0_2 q1_2 q2_2 q3_2 + q1 q0_1 q1_1 q2_1 q3_1 + q0 q0_0 q1_0 q2_0 q3_0 + */ + __asm__ __volatile__( + "precrq.qb.ph %[prim1], %[q3], %[q2] \n\t" + "precr.qb.ph %[prim2], %[q3], %[q2] \n\t" + "precrq.qb.ph %[prim3], %[q1], %[q0] \n\t" + "precr.qb.ph %[prim4], %[q1], %[q0] \n\t" + + "precrq.qb.ph %[q2], %[prim1], %[prim2] \n\t" + "precr.qb.ph %[q0], %[prim1], %[prim2] \n\t" + "precrq.qb.ph %[sec3], %[prim3], %[prim4] \n\t" + "precr.qb.ph %[sec4], %[prim3], %[prim4] \n\t" + + "precrq.ph.w %[q3], %[q2], %[sec3] \n\t" + "precrq.ph.w %[q1], %[q0], %[sec4] \n\t" + "append %[q2], %[sec3], 16 \n\t" + "append %[q0], %[sec4], 16 \n\t" + + : [prim1] "=&r"(prim1), [prim2] "=&r"(prim2), [prim3] "=&r"(prim3), + [prim4] "=&r"(prim4), [q3] "+r"(q3), [q2] "+r"(q2), [q1] "+r"(q1), + [q0] "+r"(q0), [sec3] "=&r"(sec3), [sec4] "=&r"(sec4) + :); + + filter_hev_mask_flatmask4_dspr2(limit_vec, flimit_vec, thresh_vec, p1, p0, + p3, p2, q0, q1, q2, q3, &hev, &mask, &flat); + + if ((flat == 0) && (mask != 0)) { + filter1_dspr2(mask, hev, p1, p0, q0, q1, &p1_f0, &p0_f0, &q0_f0, &q1_f0); + STORE_F0() + } else if ((mask & flat) == 0xFFFFFFFF) { + /* left 2 element operation */ + PACK_LEFT_0TO3() + mbfilter_dspr2(&p3_l, &p2_l, &p1_l, &p0_l, &q0_l, &q1_l, &q2_l, &q3_l); + + /* right 2 element operation */ + PACK_RIGHT_0TO3() + mbfilter_dspr2(&p3_r, &p2_r, &p1_r, &p0_r, &q0_r, &q1_r, &q2_r, &q3_r); + + STORE_F1() + } else if ((flat != 0) && (mask != 0)) { + filter1_dspr2(mask, hev, p1, p0, q0, q1, &p1_f0, &p0_f0, &q0_f0, &q1_f0); + + /* left 2 element operation */ + PACK_LEFT_0TO3() + mbfilter_dspr2(&p3_l, &p2_l, &p1_l, &p0_l, &q0_l, &q1_l, &q2_l, &q3_l); + + /* right 2 element operation */ + PACK_RIGHT_0TO3() + mbfilter_dspr2(&p3_r, &p2_r, &p1_r, &p0_r, &q0_r, &q1_r, &q2_r, &q3_r); + + if (mask & flat & 0x000000FF) { + __asm__ __volatile__( + "sb %[p2_r], -3(%[s4]) \n\t" + "sb %[p1_r], -2(%[s4]) \n\t" + "sb %[p0_r], -1(%[s4]) \n\t" + "sb %[q0_r], (%[s4]) \n\t" + "sb %[q1_r], +1(%[s4]) \n\t" + "sb %[q2_r], +2(%[s4]) \n\t" + + : + : [p2_r] "r"(p2_r), [p1_r] "r"(p1_r), [p0_r] "r"(p0_r), + [q0_r] "r"(q0_r), [q1_r] "r"(q1_r), [q2_r] "r"(q2_r), + [s4] "r"(s4)); + } else if (mask & 0x000000FF) { + __asm__ __volatile__( + "sb %[p1_f0], -2(%[s4]) \n\t" + "sb %[p0_f0], -1(%[s4]) \n\t" + "sb %[q0_f0], (%[s4]) \n\t" + "sb %[q1_f0], +1(%[s4]) \n\t" + + : + : [p1_f0] "r"(p1_f0), [p0_f0] "r"(p0_f0), [q0_f0] "r"(q0_f0), + [q1_f0] "r"(q1_f0), [s4] "r"(s4)); + } + + __asm__ __volatile__( + "srl %[p2_r], %[p2_r], 16 \n\t" + "srl %[p1_r], %[p1_r], 16 \n\t" + "srl %[p0_r], %[p0_r], 16 \n\t" + "srl %[q0_r], %[q0_r], 16 \n\t" + "srl %[q1_r], %[q1_r], 16 \n\t" + "srl %[q2_r], %[q2_r], 16 \n\t" + "srl %[p1_f0], %[p1_f0], 8 \n\t" + "srl %[p0_f0], %[p0_f0], 8 \n\t" + "srl %[q0_f0], %[q0_f0], 8 \n\t" + "srl %[q1_f0], %[q1_f0], 8 \n\t" + + : [p2_r] "+r"(p2_r), [p1_r] "+r"(p1_r), [p0_r] "+r"(p0_r), + [q0_r] "+r"(q0_r), [q1_r] "+r"(q1_r), [q2_r] "+r"(q2_r), + [p1_f0] "+r"(p1_f0), [p0_f0] "+r"(p0_f0), [q0_f0] "+r"(q0_f0), + [q1_f0] "+r"(q1_f0) + :); + + if (mask & flat & 0x0000FF00) { + __asm__ __volatile__( + "sb %[p2_r], -3(%[s3]) \n\t" + "sb %[p1_r], -2(%[s3]) \n\t" + "sb %[p0_r], -1(%[s3]) \n\t" + "sb %[q0_r], (%[s3]) \n\t" + "sb %[q1_r], +1(%[s3]) \n\t" + "sb %[q2_r], +2(%[s3]) \n\t" + + : + : [p2_r] "r"(p2_r), [p1_r] "r"(p1_r), [p0_r] "r"(p0_r), + [q0_r] "r"(q0_r), [q1_r] "r"(q1_r), [q2_r] "r"(q2_r), + [s3] "r"(s3)); + } else if (mask & 0x0000FF00) { + __asm__ __volatile__( + "sb %[p1_f0], -2(%[s3]) \n\t" + "sb %[p0_f0], -1(%[s3]) \n\t" + "sb %[q0_f0], (%[s3]) \n\t" + "sb %[q1_f0], +1(%[s3]) \n\t" + + : + : [p1_f0] "r"(p1_f0), [p0_f0] "r"(p0_f0), [q0_f0] "r"(q0_f0), + [q1_f0] "r"(q1_f0), [s3] "r"(s3)); + } + + __asm__ __volatile__( + "srl %[p1_f0], %[p1_f0], 8 \n\t" + "srl %[p0_f0], %[p0_f0], 8 \n\t" + "srl %[q0_f0], %[q0_f0], 8 \n\t" + "srl %[q1_f0], %[q1_f0], 8 \n\t" + + : [p2] "+r"(p2), [p1] "+r"(p1), [p0] "+r"(p0), [q0] "+r"(q0), + [q1] "+r"(q1), [q2] "+r"(q2), [p1_f0] "+r"(p1_f0), + [p0_f0] "+r"(p0_f0), [q0_f0] "+r"(q0_f0), [q1_f0] "+r"(q1_f0) + :); + + if (mask & flat & 0x00FF0000) { + __asm__ __volatile__( + "sb %[p2_l], -3(%[s2]) \n\t" + "sb %[p1_l], -2(%[s2]) \n\t" + "sb %[p0_l], -1(%[s2]) \n\t" + "sb %[q0_l], (%[s2]) \n\t" + "sb %[q1_l], +1(%[s2]) \n\t" + "sb %[q2_l], +2(%[s2]) \n\t" + + : + : [p2_l] "r"(p2_l), [p1_l] "r"(p1_l), [p0_l] "r"(p0_l), + [q0_l] "r"(q0_l), [q1_l] "r"(q1_l), [q2_l] "r"(q2_l), + [s2] "r"(s2)); + } else if (mask & 0x00FF0000) { + __asm__ __volatile__( + "sb %[p1_f0], -2(%[s2]) \n\t" + "sb %[p0_f0], -1(%[s2]) \n\t" + "sb %[q0_f0], (%[s2]) \n\t" + "sb %[q1_f0], +1(%[s2]) \n\t" + + : + : [p1_f0] "r"(p1_f0), [p0_f0] "r"(p0_f0), [q0_f0] "r"(q0_f0), + [q1_f0] "r"(q1_f0), [s2] "r"(s2)); + } + + __asm__ __volatile__( + "srl %[p2_l], %[p2_l], 16 \n\t" + "srl %[p1_l], %[p1_l], 16 \n\t" + "srl %[p0_l], %[p0_l], 16 \n\t" + "srl %[q0_l], %[q0_l], 16 \n\t" + "srl %[q1_l], %[q1_l], 16 \n\t" + "srl %[q2_l], %[q2_l], 16 \n\t" + "srl %[p1_f0], %[p1_f0], 8 \n\t" + "srl %[p0_f0], %[p0_f0], 8 \n\t" + "srl %[q0_f0], %[q0_f0], 8 \n\t" + "srl %[q1_f0], %[q1_f0], 8 \n\t" + + : [p2_l] "+r"(p2_l), [p1_l] "+r"(p1_l), [p0_l] "+r"(p0_l), + [q0_l] "+r"(q0_l), [q1_l] "+r"(q1_l), [q2_l] "+r"(q2_l), + [p1_f0] "+r"(p1_f0), [p0_f0] "+r"(p0_f0), [q0_f0] "+r"(q0_f0), + [q1_f0] "+r"(q1_f0) + :); + + if (mask & flat & 0xFF000000) { + __asm__ __volatile__( + "sb %[p2_l], -3(%[s1]) \n\t" + "sb %[p1_l], -2(%[s1]) \n\t" + "sb %[p0_l], -1(%[s1]) \n\t" + "sb %[q0_l], (%[s1]) \n\t" + "sb %[q1_l], +1(%[s1]) \n\t" + "sb %[q2_l], +2(%[s1]) \n\t" + + : + : [p2_l] "r"(p2_l), [p1_l] "r"(p1_l), [p0_l] "r"(p0_l), + [q0_l] "r"(q0_l), [q1_l] "r"(q1_l), [q2_l] "r"(q2_l), + [s1] "r"(s1)); + } else if (mask & 0xFF000000) { + __asm__ __volatile__( + "sb %[p1_f0], -2(%[s1]) \n\t" + "sb %[p0_f0], -1(%[s1]) \n\t" + "sb %[q0_f0], (%[s1]) \n\t" + "sb %[q1_f0], +1(%[s1]) \n\t" + + : + : [p1_f0] "r"(p1_f0), [p0_f0] "r"(p0_f0), [q0_f0] "r"(q0_f0), + [q1_f0] "r"(q1_f0), [s1] "r"(s1)); + } + } + } +} +#endif // #if HAVE_DSPR2 diff --git a/third_party/aom/aom_dsp/mips/loopfilter_mb_horiz_dspr2.c b/third_party/aom/aom_dsp/mips/loopfilter_mb_horiz_dspr2.c new file mode 100644 index 0000000000..34733e42ef --- /dev/null +++ b/third_party/aom/aom_dsp/mips/loopfilter_mb_horiz_dspr2.c @@ -0,0 +1,734 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include + +#include "config/aom_dsp_rtcd.h" + +#include "aom/aom_integer.h" +#include "aom_dsp/mips/common_dspr2.h" +#include "aom_dsp/mips/loopfilter_filters_dspr2.h" +#include "aom_dsp/mips/loopfilter_macros_dspr2.h" +#include "aom_dsp/mips/loopfilter_masks_dspr2.h" +#include "aom_mem/aom_mem.h" + +#if HAVE_DSPR2 +static void mb_lpf_horizontal_edge(unsigned char *s, int pitch, + const uint8_t *blimit, const uint8_t *limit, + const uint8_t *thresh, int count) { + uint32_t mask; + uint32_t hev, flat, flat2; + uint8_t i; + uint8_t *sp7, *sp6, *sp5, *sp4, *sp3, *sp2, *sp1, *sp0; + uint8_t *sq0, *sq1, *sq2, *sq3, *sq4, *sq5, *sq6, *sq7; + uint32_t thresh_vec, flimit_vec, limit_vec; + uint32_t uflimit, ulimit, uthresh; + uint32_t p7, p6, p5, p4, p3, p2, p1, p0, q0, q1, q2, q3, q4, q5, q6, q7; + uint32_t p1_f0, p0_f0, q0_f0, q1_f0; + uint32_t p7_l, p6_l, p5_l, p4_l, p3_l, p2_l, p1_l, p0_l; + uint32_t q0_l, q1_l, q2_l, q3_l, q4_l, q5_l, q6_l, q7_l; + uint32_t p7_r, p6_r, p5_r, p4_r, p3_r, p2_r, p1_r, p0_r; + uint32_t q0_r, q1_r, q2_r, q3_r, q4_r, q5_r, q6_r, q7_r; + uint32_t p2_l_f1, p1_l_f1, p0_l_f1, p2_r_f1, p1_r_f1, p0_r_f1; + uint32_t q0_l_f1, q1_l_f1, q2_l_f1, q0_r_f1, q1_r_f1, q2_r_f1; + + uflimit = *blimit; + ulimit = *limit; + uthresh = *thresh; + + /* create quad-byte */ + __asm__ __volatile__( + "replv.qb %[thresh_vec], %[uthresh] \n\t" + "replv.qb %[flimit_vec], %[uflimit] \n\t" + "replv.qb %[limit_vec], %[ulimit] \n\t" + + : [thresh_vec] "=&r"(thresh_vec), [flimit_vec] "=&r"(flimit_vec), + [limit_vec] "=r"(limit_vec) + : [uthresh] "r"(uthresh), [uflimit] "r"(uflimit), [ulimit] "r"(ulimit)); + + /* prefetch data for store */ + prefetch_store(s); + + for (i = 0; i < (2 * count); i++) { + sp7 = s - (pitch << 3); + sp6 = sp7 + pitch; + sp5 = sp6 + pitch; + sp4 = sp5 + pitch; + sp3 = sp4 + pitch; + sp2 = sp3 + pitch; + sp1 = sp2 + pitch; + sp0 = sp1 + pitch; + sq0 = s; + sq1 = s + pitch; + sq2 = sq1 + pitch; + sq3 = sq2 + pitch; + sq4 = sq3 + pitch; + sq5 = sq4 + pitch; + sq6 = sq5 + pitch; + sq7 = sq6 + pitch; + + __asm__ __volatile__( + "lw %[p7], (%[sp7]) \n\t" + "lw %[p6], (%[sp6]) \n\t" + "lw %[p5], (%[sp5]) \n\t" + "lw %[p4], (%[sp4]) \n\t" + "lw %[p3], (%[sp3]) \n\t" + "lw %[p2], (%[sp2]) \n\t" + "lw %[p1], (%[sp1]) \n\t" + "lw %[p0], (%[sp0]) \n\t" + + : [p3] "=&r"(p3), [p2] "=&r"(p2), [p1] "=&r"(p1), [p0] "=&r"(p0), + [p7] "=&r"(p7), [p6] "=&r"(p6), [p5] "=&r"(p5), [p4] "=&r"(p4) + : [sp3] "r"(sp3), [sp2] "r"(sp2), [sp1] "r"(sp1), [sp0] "r"(sp0), + [sp4] "r"(sp4), [sp5] "r"(sp5), [sp6] "r"(sp6), [sp7] "r"(sp7)); + + __asm__ __volatile__( + "lw %[q0], (%[sq0]) \n\t" + "lw %[q1], (%[sq1]) \n\t" + "lw %[q2], (%[sq2]) \n\t" + "lw %[q3], (%[sq3]) \n\t" + "lw %[q4], (%[sq4]) \n\t" + "lw %[q5], (%[sq5]) \n\t" + "lw %[q6], (%[sq6]) \n\t" + "lw %[q7], (%[sq7]) \n\t" + + : [q3] "=&r"(q3), [q2] "=&r"(q2), [q1] "=&r"(q1), [q0] "=&r"(q0), + [q7] "=&r"(q7), [q6] "=&r"(q6), [q5] "=&r"(q5), [q4] "=&r"(q4) + : [sq3] "r"(sq3), [sq2] "r"(sq2), [sq1] "r"(sq1), [sq0] "r"(sq0), + [sq4] "r"(sq4), [sq5] "r"(sq5), [sq6] "r"(sq6), [sq7] "r"(sq7)); + + filter_hev_mask_flatmask4_dspr2(limit_vec, flimit_vec, thresh_vec, p1, p0, + p3, p2, q0, q1, q2, q3, &hev, &mask, &flat); + + flatmask5(p7, p6, p5, p4, p0, q0, q4, q5, q6, q7, &flat2); + + /* f0 */ + if (((flat2 == 0) && (flat == 0) && (mask != 0)) || + ((flat2 != 0) && (flat == 0) && (mask != 0))) { + filter1_dspr2(mask, hev, p1, p0, q0, q1, &p1_f0, &p0_f0, &q0_f0, &q1_f0); + + __asm__ __volatile__( + "sw %[p1_f0], (%[sp1]) \n\t" + "sw %[p0_f0], (%[sp0]) \n\t" + "sw %[q0_f0], (%[sq0]) \n\t" + "sw %[q1_f0], (%[sq1]) \n\t" + + : + : [p1_f0] "r"(p1_f0), [p0_f0] "r"(p0_f0), [q0_f0] "r"(q0_f0), + [q1_f0] "r"(q1_f0), [sp1] "r"(sp1), [sp0] "r"(sp0), [sq0] "r"(sq0), + [sq1] "r"(sq1)); + } else if ((flat2 == 0XFFFFFFFF) && (flat == 0xFFFFFFFF) && + (mask == 0xFFFFFFFF)) { + /* f2 */ + PACK_LEFT_0TO3() + PACK_LEFT_4TO7() + wide_mbfilter_dspr2(&p7_l, &p6_l, &p5_l, &p4_l, &p3_l, &p2_l, &p1_l, + &p0_l, &q0_l, &q1_l, &q2_l, &q3_l, &q4_l, &q5_l, + &q6_l, &q7_l); + + PACK_RIGHT_0TO3() + PACK_RIGHT_4TO7() + wide_mbfilter_dspr2(&p7_r, &p6_r, &p5_r, &p4_r, &p3_r, &p2_r, &p1_r, + &p0_r, &q0_r, &q1_r, &q2_r, &q3_r, &q4_r, &q5_r, + &q6_r, &q7_r); + + COMBINE_LEFT_RIGHT_0TO2() + COMBINE_LEFT_RIGHT_3TO6() + + __asm__ __volatile__( + "sw %[p6], (%[sp6]) \n\t" + "sw %[p5], (%[sp5]) \n\t" + "sw %[p4], (%[sp4]) \n\t" + "sw %[p3], (%[sp3]) \n\t" + "sw %[p2], (%[sp2]) \n\t" + "sw %[p1], (%[sp1]) \n\t" + "sw %[p0], (%[sp0]) \n\t" + + : + : [p6] "r"(p6), [p5] "r"(p5), [p4] "r"(p4), [p3] "r"(p3), + [p2] "r"(p2), [p1] "r"(p1), [p0] "r"(p0), [sp6] "r"(sp6), + [sp5] "r"(sp5), [sp4] "r"(sp4), [sp3] "r"(sp3), [sp2] "r"(sp2), + [sp1] "r"(sp1), [sp0] "r"(sp0)); + + __asm__ __volatile__( + "sw %[q6], (%[sq6]) \n\t" + "sw %[q5], (%[sq5]) \n\t" + "sw %[q4], (%[sq4]) \n\t" + "sw %[q3], (%[sq3]) \n\t" + "sw %[q2], (%[sq2]) \n\t" + "sw %[q1], (%[sq1]) \n\t" + "sw %[q0], (%[sq0]) \n\t" + + : + : [q6] "r"(q6), [q5] "r"(q5), [q4] "r"(q4), [q3] "r"(q3), + [q2] "r"(q2), [q1] "r"(q1), [q0] "r"(q0), [sq6] "r"(sq6), + [sq5] "r"(sq5), [sq4] "r"(sq4), [sq3] "r"(sq3), [sq2] "r"(sq2), + [sq1] "r"(sq1), [sq0] "r"(sq0)); + } else if ((flat2 == 0) && (flat == 0xFFFFFFFF) && (mask == 0xFFFFFFFF)) { + /* f1 */ + /* left 2 element operation */ + PACK_LEFT_0TO3() + mbfilter_dspr2(&p3_l, &p2_l, &p1_l, &p0_l, &q0_l, &q1_l, &q2_l, &q3_l); + + /* right 2 element operation */ + PACK_RIGHT_0TO3() + mbfilter_dspr2(&p3_r, &p2_r, &p1_r, &p0_r, &q0_r, &q1_r, &q2_r, &q3_r); + + COMBINE_LEFT_RIGHT_0TO2() + + __asm__ __volatile__( + "sw %[p2], (%[sp2]) \n\t" + "sw %[p1], (%[sp1]) \n\t" + "sw %[p0], (%[sp0]) \n\t" + "sw %[q0], (%[sq0]) \n\t" + "sw %[q1], (%[sq1]) \n\t" + "sw %[q2], (%[sq2]) \n\t" + + : + : [p2] "r"(p2), [p1] "r"(p1), [p0] "r"(p0), [q0] "r"(q0), + [q1] "r"(q1), [q2] "r"(q2), [sp2] "r"(sp2), [sp1] "r"(sp1), + [sp0] "r"(sp0), [sq0] "r"(sq0), [sq1] "r"(sq1), [sq2] "r"(sq2)); + } else if ((flat2 == 0) && (flat != 0) && (mask != 0)) { + /* f0+f1 */ + filter1_dspr2(mask, hev, p1, p0, q0, q1, &p1_f0, &p0_f0, &q0_f0, &q1_f0); + + /* left 2 element operation */ + PACK_LEFT_0TO3() + mbfilter_dspr2(&p3_l, &p2_l, &p1_l, &p0_l, &q0_l, &q1_l, &q2_l, &q3_l); + + /* right 2 element operation */ + PACK_RIGHT_0TO3() + mbfilter_dspr2(&p3_r, &p2_r, &p1_r, &p0_r, &q0_r, &q1_r, &q2_r, &q3_r); + + if (mask & flat & 0x000000FF) { + __asm__ __volatile__( + "sb %[p2_r], (%[sp2]) \n\t" + "sb %[p1_r], (%[sp1]) \n\t" + "sb %[p0_r], (%[sp0]) \n\t" + "sb %[q0_r], (%[sq0]) \n\t" + "sb %[q1_r], (%[sq1]) \n\t" + "sb %[q2_r], (%[sq2]) \n\t" + + : + : [p2_r] "r"(p2_r), [p1_r] "r"(p1_r), [p0_r] "r"(p0_r), + [q0_r] "r"(q0_r), [q1_r] "r"(q1_r), [q2_r] "r"(q2_r), + [sp2] "r"(sp2), [sp1] "r"(sp1), [sp0] "r"(sp0), [sq0] "r"(sq0), + [sq1] "r"(sq1), [sq2] "r"(sq2)); + } else if (mask & 0x000000FF) { + __asm__ __volatile__( + "sb %[p1_f0], (%[sp1]) \n\t" + "sb %[p0_f0], (%[sp0]) \n\t" + "sb %[q0_f0], (%[sq0]) \n\t" + "sb %[q1_f0], (%[sq1]) \n\t" + + : + : [p1_f0] "r"(p1_f0), [p0_f0] "r"(p0_f0), [q0_f0] "r"(q0_f0), + [q1_f0] "r"(q1_f0), [sp1] "r"(sp1), [sp0] "r"(sp0), + [sq0] "r"(sq0), [sq1] "r"(sq1)); + } + + __asm__ __volatile__( + "srl %[p2_r], %[p2_r], 16 \n\t" + "srl %[p1_r], %[p1_r], 16 \n\t" + "srl %[p0_r], %[p0_r], 16 \n\t" + "srl %[q0_r], %[q0_r], 16 \n\t" + "srl %[q1_r], %[q1_r], 16 \n\t" + "srl %[q2_r], %[q2_r], 16 \n\t" + "srl %[p1_f0], %[p1_f0], 8 \n\t" + "srl %[p0_f0], %[p0_f0], 8 \n\t" + "srl %[q0_f0], %[q0_f0], 8 \n\t" + "srl %[q1_f0], %[q1_f0], 8 \n\t" + + : [p2_r] "+r"(p2_r), [p1_r] "+r"(p1_r), [p0_r] "+r"(p0_r), + [q0_r] "+r"(q0_r), [q1_r] "+r"(q1_r), [q2_r] "+r"(q2_r), + [p1_f0] "+r"(p1_f0), [p0_f0] "+r"(p0_f0), [q0_f0] "+r"(q0_f0), + [q1_f0] "+r"(q1_f0) + :); + + if (mask & flat & 0x0000FF00) { + __asm__ __volatile__( + "sb %[p2_r], +1(%[sp2]) \n\t" + "sb %[p1_r], +1(%[sp1]) \n\t" + "sb %[p0_r], +1(%[sp0]) \n\t" + "sb %[q0_r], +1(%[sq0]) \n\t" + "sb %[q1_r], +1(%[sq1]) \n\t" + "sb %[q2_r], +1(%[sq2]) \n\t" + + : + : [p2_r] "r"(p2_r), [p1_r] "r"(p1_r), [p0_r] "r"(p0_r), + [q0_r] "r"(q0_r), [q1_r] "r"(q1_r), [q2_r] "r"(q2_r), + [sp2] "r"(sp2), [sp1] "r"(sp1), [sp0] "r"(sp0), [sq0] "r"(sq0), + [sq1] "r"(sq1), [sq2] "r"(sq2)); + } else if (mask & 0x0000FF00) { + __asm__ __volatile__( + "sb %[p1_f0], +1(%[sp1]) \n\t" + "sb %[p0_f0], +1(%[sp0]) \n\t" + "sb %[q0_f0], +1(%[sq0]) \n\t" + "sb %[q1_f0], +1(%[sq1]) \n\t" + + : + : [p1_f0] "r"(p1_f0), [p0_f0] "r"(p0_f0), [q0_f0] "r"(q0_f0), + [q1_f0] "r"(q1_f0), [sp1] "r"(sp1), [sp0] "r"(sp0), + [sq0] "r"(sq0), [sq1] "r"(sq1)); + } + + __asm__ __volatile__( + "srl %[p1_f0], %[p1_f0], 8 \n\t" + "srl %[p0_f0], %[p0_f0], 8 \n\t" + "srl %[q0_f0], %[q0_f0], 8 \n\t" + "srl %[q1_f0], %[q1_f0], 8 \n\t" + + : [p1_f0] "+r"(p1_f0), [p0_f0] "+r"(p0_f0), [q0_f0] "+r"(q0_f0), + [q1_f0] "+r"(q1_f0) + :); + + if (mask & flat & 0x00FF0000) { + __asm__ __volatile__( + "sb %[p2_l], +2(%[sp2]) \n\t" + "sb %[p1_l], +2(%[sp1]) \n\t" + "sb %[p0_l], +2(%[sp0]) \n\t" + "sb %[q0_l], +2(%[sq0]) \n\t" + "sb %[q1_l], +2(%[sq1]) \n\t" + "sb %[q2_l], +2(%[sq2]) \n\t" + + : + : [p2_l] "r"(p2_l), [p1_l] "r"(p1_l), [p0_l] "r"(p0_l), + [q0_l] "r"(q0_l), [q1_l] "r"(q1_l), [q2_l] "r"(q2_l), + [sp2] "r"(sp2), [sp1] "r"(sp1), [sp0] "r"(sp0), [sq0] "r"(sq0), + [sq1] "r"(sq1), [sq2] "r"(sq2)); + } else if (mask & 0x00FF0000) { + __asm__ __volatile__( + "sb %[p1_f0], +2(%[sp1]) \n\t" + "sb %[p0_f0], +2(%[sp0]) \n\t" + "sb %[q0_f0], +2(%[sq0]) \n\t" + "sb %[q1_f0], +2(%[sq1]) \n\t" + + : + : [p1_f0] "r"(p1_f0), [p0_f0] "r"(p0_f0), [q0_f0] "r"(q0_f0), + [q1_f0] "r"(q1_f0), [sp1] "r"(sp1), [sp0] "r"(sp0), + [sq0] "r"(sq0), [sq1] "r"(sq1)); + } + + __asm__ __volatile__( + "srl %[p2_l], %[p2_l], 16 \n\t" + "srl %[p1_l], %[p1_l], 16 \n\t" + "srl %[p0_l], %[p0_l], 16 \n\t" + "srl %[q0_l], %[q0_l], 16 \n\t" + "srl %[q1_l], %[q1_l], 16 \n\t" + "srl %[q2_l], %[q2_l], 16 \n\t" + "srl %[p1_f0], %[p1_f0], 8 \n\t" + "srl %[p0_f0], %[p0_f0], 8 \n\t" + "srl %[q0_f0], %[q0_f0], 8 \n\t" + "srl %[q1_f0], %[q1_f0], 8 \n\t" + + : [p2_l] "+r"(p2_l), [p1_l] "+r"(p1_l), [p0_l] "+r"(p0_l), + [q0_l] "+r"(q0_l), [q1_l] "+r"(q1_l), [q2_l] "+r"(q2_l), + [p1_f0] "+r"(p1_f0), [p0_f0] "+r"(p0_f0), [q0_f0] "+r"(q0_f0), + [q1_f0] "+r"(q1_f0) + :); + + if (mask & flat & 0xFF000000) { + __asm__ __volatile__( + "sb %[p2_l], +3(%[sp2]) \n\t" + "sb %[p1_l], +3(%[sp1]) \n\t" + "sb %[p0_l], +3(%[sp0]) \n\t" + "sb %[q0_l], +3(%[sq0]) \n\t" + "sb %[q1_l], +3(%[sq1]) \n\t" + "sb %[q2_l], +3(%[sq2]) \n\t" + + : + : [p2_l] "r"(p2_l), [p1_l] "r"(p1_l), [p0_l] "r"(p0_l), + [q0_l] "r"(q0_l), [q1_l] "r"(q1_l), [q2_l] "r"(q2_l), + [sp2] "r"(sp2), [sp1] "r"(sp1), [sp0] "r"(sp0), [sq0] "r"(sq0), + [sq1] "r"(sq1), [sq2] "r"(sq2)); + } else if (mask & 0xFF000000) { + __asm__ __volatile__( + "sb %[p1_f0], +3(%[sp1]) \n\t" + "sb %[p0_f0], +3(%[sp0]) \n\t" + "sb %[q0_f0], +3(%[sq0]) \n\t" + "sb %[q1_f0], +3(%[sq1]) \n\t" + + : + : [p1_f0] "r"(p1_f0), [p0_f0] "r"(p0_f0), [q0_f0] "r"(q0_f0), + [q1_f0] "r"(q1_f0), [sp1] "r"(sp1), [sp0] "r"(sp0), + [sq0] "r"(sq0), [sq1] "r"(sq1)); + } + } else if ((flat2 != 0) && (flat != 0) && (mask != 0)) { + /* f0 + f1 + f2 */ + /* f0 function */ + filter1_dspr2(mask, hev, p1, p0, q0, q1, &p1_f0, &p0_f0, &q0_f0, &q1_f0); + + /* f1 function */ + /* left 2 element operation */ + PACK_LEFT_0TO3() + mbfilter1_dspr2(p3_l, p2_l, p1_l, p0_l, q0_l, q1_l, q2_l, q3_l, &p2_l_f1, + &p1_l_f1, &p0_l_f1, &q0_l_f1, &q1_l_f1, &q2_l_f1); + + /* right 2 element operation */ + PACK_RIGHT_0TO3() + mbfilter1_dspr2(p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r, &p2_r_f1, + &p1_r_f1, &p0_r_f1, &q0_r_f1, &q1_r_f1, &q2_r_f1); + + /* f2 function */ + PACK_LEFT_4TO7() + wide_mbfilter_dspr2(&p7_l, &p6_l, &p5_l, &p4_l, &p3_l, &p2_l, &p1_l, + &p0_l, &q0_l, &q1_l, &q2_l, &q3_l, &q4_l, &q5_l, + &q6_l, &q7_l); + + PACK_RIGHT_4TO7() + wide_mbfilter_dspr2(&p7_r, &p6_r, &p5_r, &p4_r, &p3_r, &p2_r, &p1_r, + &p0_r, &q0_r, &q1_r, &q2_r, &q3_r, &q4_r, &q5_r, + &q6_r, &q7_r); + + if (mask & flat & flat2 & 0x000000FF) { + __asm__ __volatile__( + "sb %[p6_r], (%[sp6]) \n\t" + "sb %[p5_r], (%[sp5]) \n\t" + "sb %[p4_r], (%[sp4]) \n\t" + "sb %[p3_r], (%[sp3]) \n\t" + "sb %[p2_r], (%[sp2]) \n\t" + "sb %[p1_r], (%[sp1]) \n\t" + "sb %[p0_r], (%[sp0]) \n\t" + + : + : [p6_r] "r"(p6_r), [p5_r] "r"(p5_r), [p4_r] "r"(p4_r), + [p3_r] "r"(p3_r), [p2_r] "r"(p2_r), [p1_r] "r"(p1_r), + [sp6] "r"(sp6), [sp5] "r"(sp5), [sp4] "r"(sp4), [sp3] "r"(sp3), + [sp2] "r"(sp2), [sp1] "r"(sp1), [p0_r] "r"(p0_r), [sp0] "r"(sp0)); + + __asm__ __volatile__( + "sb %[q0_r], (%[sq0]) \n\t" + "sb %[q1_r], (%[sq1]) \n\t" + "sb %[q2_r], (%[sq2]) \n\t" + "sb %[q3_r], (%[sq3]) \n\t" + "sb %[q4_r], (%[sq4]) \n\t" + "sb %[q5_r], (%[sq5]) \n\t" + "sb %[q6_r], (%[sq6]) \n\t" + + : + : [q0_r] "r"(q0_r), [q1_r] "r"(q1_r), [q2_r] "r"(q2_r), + [q3_r] "r"(q3_r), [q4_r] "r"(q4_r), [q5_r] "r"(q5_r), + [q6_r] "r"(q6_r), [sq0] "r"(sq0), [sq1] "r"(sq1), [sq2] "r"(sq2), + [sq3] "r"(sq3), [sq4] "r"(sq4), [sq5] "r"(sq5), [sq6] "r"(sq6)); + } else if (mask & flat & 0x000000FF) { + __asm__ __volatile__( + "sb %[p2_r_f1], (%[sp2]) \n\t" + "sb %[p1_r_f1], (%[sp1]) \n\t" + "sb %[p0_r_f1], (%[sp0]) \n\t" + "sb %[q0_r_f1], (%[sq0]) \n\t" + "sb %[q1_r_f1], (%[sq1]) \n\t" + "sb %[q2_r_f1], (%[sq2]) \n\t" + + : + : [p2_r_f1] "r"(p2_r_f1), [p1_r_f1] "r"(p1_r_f1), + [p0_r_f1] "r"(p0_r_f1), [q0_r_f1] "r"(q0_r_f1), + [q1_r_f1] "r"(q1_r_f1), [q2_r_f1] "r"(q2_r_f1), [sp2] "r"(sp2), + [sp1] "r"(sp1), [sp0] "r"(sp0), [sq0] "r"(sq0), [sq1] "r"(sq1), + [sq2] "r"(sq2)); + } else if (mask & 0x000000FF) { + __asm__ __volatile__( + "sb %[p1_f0], (%[sp1]) \n\t" + "sb %[p0_f0], (%[sp0]) \n\t" + "sb %[q0_f0], (%[sq0]) \n\t" + "sb %[q1_f0], (%[sq1]) \n\t" + + : + : [p1_f0] "r"(p1_f0), [p0_f0] "r"(p0_f0), [q0_f0] "r"(q0_f0), + [q1_f0] "r"(q1_f0), [sp1] "r"(sp1), [sp0] "r"(sp0), + [sq0] "r"(sq0), [sq1] "r"(sq1)); + } + + __asm__ __volatile__( + "srl %[p6_r], %[p6_r], 16 \n\t" + "srl %[p5_r], %[p5_r], 16 \n\t" + "srl %[p4_r], %[p4_r], 16 \n\t" + "srl %[p3_r], %[p3_r], 16 \n\t" + "srl %[p2_r], %[p2_r], 16 \n\t" + "srl %[p1_r], %[p1_r], 16 \n\t" + "srl %[p0_r], %[p0_r], 16 \n\t" + "srl %[q0_r], %[q0_r], 16 \n\t" + "srl %[q1_r], %[q1_r], 16 \n\t" + "srl %[q2_r], %[q2_r], 16 \n\t" + "srl %[q3_r], %[q3_r], 16 \n\t" + "srl %[q4_r], %[q4_r], 16 \n\t" + "srl %[q5_r], %[q5_r], 16 \n\t" + "srl %[q6_r], %[q6_r], 16 \n\t" + + : [q0_r] "+r"(q0_r), [q1_r] "+r"(q1_r), [q2_r] "+r"(q2_r), + [q3_r] "+r"(q3_r), [q4_r] "+r"(q4_r), [q5_r] "+r"(q5_r), + [p6_r] "+r"(p6_r), [p5_r] "+r"(p5_r), [p4_r] "+r"(p4_r), + [p3_r] "+r"(p3_r), [p2_r] "+r"(p2_r), [p1_r] "+r"(p1_r), + [q6_r] "+r"(q6_r), [p0_r] "+r"(p0_r) + :); + + __asm__ __volatile__( + "srl %[p2_r_f1], %[p2_r_f1], 16 \n\t" + "srl %[p1_r_f1], %[p1_r_f1], 16 \n\t" + "srl %[p0_r_f1], %[p0_r_f1], 16 \n\t" + "srl %[q0_r_f1], %[q0_r_f1], 16 \n\t" + "srl %[q1_r_f1], %[q1_r_f1], 16 \n\t" + "srl %[q2_r_f1], %[q2_r_f1], 16 \n\t" + "srl %[p1_f0], %[p1_f0], 8 \n\t" + "srl %[p0_f0], %[p0_f0], 8 \n\t" + "srl %[q0_f0], %[q0_f0], 8 \n\t" + "srl %[q1_f0], %[q1_f0], 8 \n\t" + + : [p2_r_f1] "+r"(p2_r_f1), [p1_r_f1] "+r"(p1_r_f1), + [p0_r_f1] "+r"(p0_r_f1), [q0_r_f1] "+r"(q0_r_f1), + [q1_r_f1] "+r"(q1_r_f1), [q2_r_f1] "+r"(q2_r_f1), + [p1_f0] "+r"(p1_f0), [p0_f0] "+r"(p0_f0), [q0_f0] "+r"(q0_f0), + [q1_f0] "+r"(q1_f0) + :); + + if (mask & flat & flat2 & 0x0000FF00) { + __asm__ __volatile__( + "sb %[p6_r], +1(%[sp6]) \n\t" + "sb %[p5_r], +1(%[sp5]) \n\t" + "sb %[p4_r], +1(%[sp4]) \n\t" + "sb %[p3_r], +1(%[sp3]) \n\t" + "sb %[p2_r], +1(%[sp2]) \n\t" + "sb %[p1_r], +1(%[sp1]) \n\t" + "sb %[p0_r], +1(%[sp0]) \n\t" + + : + : [p6_r] "r"(p6_r), [p5_r] "r"(p5_r), [p4_r] "r"(p4_r), + [p3_r] "r"(p3_r), [p2_r] "r"(p2_r), [p1_r] "r"(p1_r), + [p0_r] "r"(p0_r), [sp6] "r"(sp6), [sp5] "r"(sp5), [sp4] "r"(sp4), + [sp3] "r"(sp3), [sp2] "r"(sp2), [sp1] "r"(sp1), [sp0] "r"(sp0)); + + __asm__ __volatile__( + "sb %[q0_r], +1(%[sq0]) \n\t" + "sb %[q1_r], +1(%[sq1]) \n\t" + "sb %[q2_r], +1(%[sq2]) \n\t" + "sb %[q3_r], +1(%[sq3]) \n\t" + "sb %[q4_r], +1(%[sq4]) \n\t" + "sb %[q5_r], +1(%[sq5]) \n\t" + "sb %[q6_r], +1(%[sq6]) \n\t" + + : + : [q0_r] "r"(q0_r), [q1_r] "r"(q1_r), [q2_r] "r"(q2_r), + [q3_r] "r"(q3_r), [q4_r] "r"(q4_r), [q5_r] "r"(q5_r), + [q6_r] "r"(q6_r), [sq0] "r"(sq0), [sq1] "r"(sq1), [sq2] "r"(sq2), + [sq3] "r"(sq3), [sq4] "r"(sq4), [sq5] "r"(sq5), [sq6] "r"(sq6)); + } else if (mask & flat & 0x0000FF00) { + __asm__ __volatile__( + "sb %[p2_r_f1], +1(%[sp2]) \n\t" + "sb %[p1_r_f1], +1(%[sp1]) \n\t" + "sb %[p0_r_f1], +1(%[sp0]) \n\t" + "sb %[q0_r_f1], +1(%[sq0]) \n\t" + "sb %[q1_r_f1], +1(%[sq1]) \n\t" + "sb %[q2_r_f1], +1(%[sq2]) \n\t" + + : + : [p2_r_f1] "r"(p2_r_f1), [p1_r_f1] "r"(p1_r_f1), + [p0_r_f1] "r"(p0_r_f1), [q0_r_f1] "r"(q0_r_f1), + [q1_r_f1] "r"(q1_r_f1), [q2_r_f1] "r"(q2_r_f1), [sp2] "r"(sp2), + [sp1] "r"(sp1), [sp0] "r"(sp0), [sq0] "r"(sq0), [sq1] "r"(sq1), + [sq2] "r"(sq2)); + } else if (mask & 0x0000FF00) { + __asm__ __volatile__( + "sb %[p1_f0], +1(%[sp1]) \n\t" + "sb %[p0_f0], +1(%[sp0]) \n\t" + "sb %[q0_f0], +1(%[sq0]) \n\t" + "sb %[q1_f0], +1(%[sq1]) \n\t" + + : + : [p1_f0] "r"(p1_f0), [p0_f0] "r"(p0_f0), [q0_f0] "r"(q0_f0), + [q1_f0] "r"(q1_f0), [sp1] "r"(sp1), [sp0] "r"(sp0), + [sq0] "r"(sq0), [sq1] "r"(sq1)); + } + + __asm__ __volatile__( + "srl %[p1_f0], %[p1_f0], 8 \n\t" + "srl %[p0_f0], %[p0_f0], 8 \n\t" + "srl %[q0_f0], %[q0_f0], 8 \n\t" + "srl %[q1_f0], %[q1_f0], 8 \n\t" + + : [p1_f0] "+r"(p1_f0), [p0_f0] "+r"(p0_f0), [q0_f0] "+r"(q0_f0), + [q1_f0] "+r"(q1_f0) + :); + + if (mask & flat & flat2 & 0x00FF0000) { + __asm__ __volatile__( + "sb %[p6_l], +2(%[sp6]) \n\t" + "sb %[p5_l], +2(%[sp5]) \n\t" + "sb %[p4_l], +2(%[sp4]) \n\t" + "sb %[p3_l], +2(%[sp3]) \n\t" + "sb %[p2_l], +2(%[sp2]) \n\t" + "sb %[p1_l], +2(%[sp1]) \n\t" + "sb %[p0_l], +2(%[sp0]) \n\t" + + : + : [p6_l] "r"(p6_l), [p5_l] "r"(p5_l), [p4_l] "r"(p4_l), + [p3_l] "r"(p3_l), [p2_l] "r"(p2_l), [p1_l] "r"(p1_l), + [p0_l] "r"(p0_l), [sp6] "r"(sp6), [sp5] "r"(sp5), [sp4] "r"(sp4), + [sp3] "r"(sp3), [sp2] "r"(sp2), [sp1] "r"(sp1), [sp0] "r"(sp0)); + + __asm__ __volatile__( + "sb %[q0_l], +2(%[sq0]) \n\t" + "sb %[q1_l], +2(%[sq1]) \n\t" + "sb %[q2_l], +2(%[sq2]) \n\t" + "sb %[q3_l], +2(%[sq3]) \n\t" + "sb %[q4_l], +2(%[sq4]) \n\t" + "sb %[q5_l], +2(%[sq5]) \n\t" + "sb %[q6_l], +2(%[sq6]) \n\t" + + : + : [q0_l] "r"(q0_l), [q1_l] "r"(q1_l), [q2_l] "r"(q2_l), + [q3_l] "r"(q3_l), [q4_l] "r"(q4_l), [q5_l] "r"(q5_l), + [q6_l] "r"(q6_l), [sq0] "r"(sq0), [sq1] "r"(sq1), [sq2] "r"(sq2), + [sq3] "r"(sq3), [sq4] "r"(sq4), [sq5] "r"(sq5), [sq6] "r"(sq6)); + } else if (mask & flat & 0x00FF0000) { + __asm__ __volatile__( + "sb %[p2_l_f1], +2(%[sp2]) \n\t" + "sb %[p1_l_f1], +2(%[sp1]) \n\t" + "sb %[p0_l_f1], +2(%[sp0]) \n\t" + "sb %[q0_l_f1], +2(%[sq0]) \n\t" + "sb %[q1_l_f1], +2(%[sq1]) \n\t" + "sb %[q2_l_f1], +2(%[sq2]) \n\t" + + : + : [p2_l_f1] "r"(p2_l_f1), [p1_l_f1] "r"(p1_l_f1), + [p0_l_f1] "r"(p0_l_f1), [q0_l_f1] "r"(q0_l_f1), + [q1_l_f1] "r"(q1_l_f1), [q2_l_f1] "r"(q2_l_f1), [sp2] "r"(sp2), + [sp1] "r"(sp1), [sp0] "r"(sp0), [sq0] "r"(sq0), [sq1] "r"(sq1), + [sq2] "r"(sq2)); + } else if (mask & 0x00FF0000) { + __asm__ __volatile__( + "sb %[p1_f0], +2(%[sp1]) \n\t" + "sb %[p0_f0], +2(%[sp0]) \n\t" + "sb %[q0_f0], +2(%[sq0]) \n\t" + "sb %[q1_f0], +2(%[sq1]) \n\t" + + : + : [p1_f0] "r"(p1_f0), [p0_f0] "r"(p0_f0), [q0_f0] "r"(q0_f0), + [q1_f0] "r"(q1_f0), [sp1] "r"(sp1), [sp0] "r"(sp0), + [sq0] "r"(sq0), [sq1] "r"(sq1)); + } + + __asm__ __volatile__( + "srl %[p6_l], %[p6_l], 16 \n\t" + "srl %[p5_l], %[p5_l], 16 \n\t" + "srl %[p4_l], %[p4_l], 16 \n\t" + "srl %[p3_l], %[p3_l], 16 \n\t" + "srl %[p2_l], %[p2_l], 16 \n\t" + "srl %[p1_l], %[p1_l], 16 \n\t" + "srl %[p0_l], %[p0_l], 16 \n\t" + "srl %[q0_l], %[q0_l], 16 \n\t" + "srl %[q1_l], %[q1_l], 16 \n\t" + "srl %[q2_l], %[q2_l], 16 \n\t" + "srl %[q3_l], %[q3_l], 16 \n\t" + "srl %[q4_l], %[q4_l], 16 \n\t" + "srl %[q5_l], %[q5_l], 16 \n\t" + "srl %[q6_l], %[q6_l], 16 \n\t" + + : [q0_l] "+r"(q0_l), [q1_l] "+r"(q1_l), [q2_l] "+r"(q2_l), + [q3_l] "+r"(q3_l), [q4_l] "+r"(q4_l), [q5_l] "+r"(q5_l), + [q6_l] "+r"(q6_l), [p6_l] "+r"(p6_l), [p5_l] "+r"(p5_l), + [p4_l] "+r"(p4_l), [p3_l] "+r"(p3_l), [p2_l] "+r"(p2_l), + [p1_l] "+r"(p1_l), [p0_l] "+r"(p0_l) + :); + + __asm__ __volatile__( + "srl %[p2_l_f1], %[p2_l_f1], 16 \n\t" + "srl %[p1_l_f1], %[p1_l_f1], 16 \n\t" + "srl %[p0_l_f1], %[p0_l_f1], 16 \n\t" + "srl %[q0_l_f1], %[q0_l_f1], 16 \n\t" + "srl %[q1_l_f1], %[q1_l_f1], 16 \n\t" + "srl %[q2_l_f1], %[q2_l_f1], 16 \n\t" + "srl %[p1_f0], %[p1_f0], 8 \n\t" + "srl %[p0_f0], %[p0_f0], 8 \n\t" + "srl %[q0_f0], %[q0_f0], 8 \n\t" + "srl %[q1_f0], %[q1_f0], 8 \n\t" + + : [p2_l_f1] "+r"(p2_l_f1), [p1_l_f1] "+r"(p1_l_f1), + [p0_l_f1] "+r"(p0_l_f1), [q0_l_f1] "+r"(q0_l_f1), + [q1_l_f1] "+r"(q1_l_f1), [q2_l_f1] "+r"(q2_l_f1), + [p1_f0] "+r"(p1_f0), [p0_f0] "+r"(p0_f0), [q0_f0] "+r"(q0_f0), + [q1_f0] "+r"(q1_f0) + :); + + if (mask & flat & flat2 & 0xFF000000) { + __asm__ __volatile__( + "sb %[p6_l], +3(%[sp6]) \n\t" + "sb %[p5_l], +3(%[sp5]) \n\t" + "sb %[p4_l], +3(%[sp4]) \n\t" + "sb %[p3_l], +3(%[sp3]) \n\t" + "sb %[p2_l], +3(%[sp2]) \n\t" + "sb %[p1_l], +3(%[sp1]) \n\t" + "sb %[p0_l], +3(%[sp0]) \n\t" + + : + : [p6_l] "r"(p6_l), [p5_l] "r"(p5_l), [p4_l] "r"(p4_l), + [p3_l] "r"(p3_l), [p2_l] "r"(p2_l), [p1_l] "r"(p1_l), + [p0_l] "r"(p0_l), [sp6] "r"(sp6), [sp5] "r"(sp5), [sp4] "r"(sp4), + [sp3] "r"(sp3), [sp2] "r"(sp2), [sp1] "r"(sp1), [sp0] "r"(sp0)); + + __asm__ __volatile__( + "sb %[q0_l], +3(%[sq0]) \n\t" + "sb %[q1_l], +3(%[sq1]) \n\t" + "sb %[q2_l], +3(%[sq2]) \n\t" + "sb %[q3_l], +3(%[sq3]) \n\t" + "sb %[q4_l], +3(%[sq4]) \n\t" + "sb %[q5_l], +3(%[sq5]) \n\t" + "sb %[q6_l], +3(%[sq6]) \n\t" + + : + : [q0_l] "r"(q0_l), [q1_l] "r"(q1_l), [q2_l] "r"(q2_l), + [q3_l] "r"(q3_l), [q4_l] "r"(q4_l), [q5_l] "r"(q5_l), + [sq0] "r"(sq0), [sq1] "r"(sq1), [sq2] "r"(sq2), [sq3] "r"(sq3), + [sq4] "r"(sq4), [sq5] "r"(sq5), [q6_l] "r"(q6_l), [sq6] "r"(sq6)); + } else if (mask & flat & 0xFF000000) { + __asm__ __volatile__( + "sb %[p2_l_f1], +3(%[sp2]) \n\t" + "sb %[p1_l_f1], +3(%[sp1]) \n\t" + "sb %[p0_l_f1], +3(%[sp0]) \n\t" + "sb %[q0_l_f1], +3(%[sq0]) \n\t" + "sb %[q1_l_f1], +3(%[sq1]) \n\t" + "sb %[q2_l_f1], +3(%[sq2]) \n\t" + + : + : [p2_l_f1] "r"(p2_l_f1), [p1_l_f1] "r"(p1_l_f1), + [p0_l_f1] "r"(p0_l_f1), [q0_l_f1] "r"(q0_l_f1), + [q1_l_f1] "r"(q1_l_f1), [q2_l_f1] "r"(q2_l_f1), [sp2] "r"(sp2), + [sp1] "r"(sp1), [sp0] "r"(sp0), [sq0] "r"(sq0), [sq1] "r"(sq1), + [sq2] "r"(sq2)); + } else if (mask & 0xFF000000) { + __asm__ __volatile__( + "sb %[p1_f0], +3(%[sp1]) \n\t" + "sb %[p0_f0], +3(%[sp0]) \n\t" + "sb %[q0_f0], +3(%[sq0]) \n\t" + "sb %[q1_f0], +3(%[sq1]) \n\t" + + : + : [p1_f0] "r"(p1_f0), [p0_f0] "r"(p0_f0), [q0_f0] "r"(q0_f0), + [q1_f0] "r"(q1_f0), [sp1] "r"(sp1), [sp0] "r"(sp0), + [sq0] "r"(sq0), [sq1] "r"(sq1)); + } + } + + s = s + 4; + } +} + +void aom_lpf_horizontal_16_dspr2(unsigned char *s, int pitch, + const uint8_t *blimit, const uint8_t *limit, + const uint8_t *thresh) { + mb_lpf_horizontal_edge(s, pitch, blimit, limit, thresh, 1); +} + +void aom_lpf_horizontal_16_dual_dspr2(unsigned char *s, int pitch, + const uint8_t *blimit, + const uint8_t *limit, + const uint8_t *thresh) { + mb_lpf_horizontal_edge(s, pitch, blimit, limit, thresh, 2); +} +#endif // #if HAVE_DSPR2 diff --git a/third_party/aom/aom_dsp/mips/loopfilter_mb_vert_dspr2.c b/third_party/aom/aom_dsp/mips/loopfilter_mb_vert_dspr2.c new file mode 100644 index 0000000000..3d3f1ec971 --- /dev/null +++ b/third_party/aom/aom_dsp/mips/loopfilter_mb_vert_dspr2.c @@ -0,0 +1,758 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include + +#include "config/aom_dsp_rtcd.h" + +#include "aom/aom_integer.h" +#include "aom_dsp/mips/common_dspr2.h" +#include "aom_dsp/mips/loopfilter_filters_dspr2.h" +#include "aom_dsp/mips/loopfilter_macros_dspr2.h" +#include "aom_dsp/mips/loopfilter_masks_dspr2.h" +#include "aom_mem/aom_mem.h" + +#if HAVE_DSPR2 +void aom_lpf_vertical_16_dspr2(uint8_t *s, int pitch, const uint8_t *blimit, + const uint8_t *limit, const uint8_t *thresh) { + uint8_t i; + uint32_t mask, hev, flat, flat2; + uint8_t *s1, *s2, *s3, *s4; + uint32_t prim1, prim2, sec3, sec4, prim3, prim4; + uint32_t thresh_vec, flimit_vec, limit_vec; + uint32_t uflimit, ulimit, uthresh; + uint32_t p7, p6, p5, p4, p3, p2, p1, p0, q0, q1, q2, q3, q4, q5, q6, q7; + uint32_t p1_f0, p0_f0, q0_f0, q1_f0; + uint32_t p7_l, p6_l, p5_l, p4_l, p3_l, p2_l, p1_l, p0_l; + uint32_t q0_l, q1_l, q2_l, q3_l, q4_l, q5_l, q6_l, q7_l; + uint32_t p7_r, p6_r, p5_r, p4_r, p3_r, p2_r, p1_r, p0_r; + uint32_t q0_r, q1_r, q2_r, q3_r, q4_r, q5_r, q6_r, q7_r; + uint32_t p2_l_f1, p1_l_f1, p0_l_f1, p2_r_f1, p1_r_f1, p0_r_f1; + uint32_t q0_l_f1, q1_l_f1, q2_l_f1, q0_r_f1, q1_r_f1, q2_r_f1; + + uflimit = *blimit; + ulimit = *limit; + uthresh = *thresh; + + /* create quad-byte */ + __asm__ __volatile__( + "replv.qb %[thresh_vec], %[uthresh] \n\t" + "replv.qb %[flimit_vec], %[uflimit] \n\t" + "replv.qb %[limit_vec], %[ulimit] \n\t" + + : [thresh_vec] "=&r"(thresh_vec), [flimit_vec] "=&r"(flimit_vec), + [limit_vec] "=r"(limit_vec) + : [uthresh] "r"(uthresh), [uflimit] "r"(uflimit), [ulimit] "r"(ulimit)); + + prefetch_store(s + pitch); + + for (i = 0; i < 2; i++) { + s1 = s; + s2 = s + pitch; + s3 = s2 + pitch; + s4 = s3 + pitch; + s = s4 + pitch; + + __asm__ __volatile__( + "lw %[p0], -4(%[s1]) \n\t" + "lw %[p1], -4(%[s2]) \n\t" + "lw %[p2], -4(%[s3]) \n\t" + "lw %[p3], -4(%[s4]) \n\t" + "lw %[p4], -8(%[s1]) \n\t" + "lw %[p5], -8(%[s2]) \n\t" + "lw %[p6], -8(%[s3]) \n\t" + "lw %[p7], -8(%[s4]) \n\t" + + : [p3] "=&r"(p3), [p2] "=&r"(p2), [p1] "=&r"(p1), [p0] "=&r"(p0), + [p7] "=&r"(p7), [p6] "=&r"(p6), [p5] "=&r"(p5), [p4] "=&r"(p4) + : [s1] "r"(s1), [s2] "r"(s2), [s3] "r"(s3), [s4] "r"(s4)); + + __asm__ __volatile__( + "lw %[q3], (%[s1]) \n\t" + "lw %[q2], (%[s2]) \n\t" + "lw %[q1], (%[s3]) \n\t" + "lw %[q0], (%[s4]) \n\t" + "lw %[q7], +4(%[s1]) \n\t" + "lw %[q6], +4(%[s2]) \n\t" + "lw %[q5], +4(%[s3]) \n\t" + "lw %[q4], +4(%[s4]) \n\t" + + : [q3] "=&r"(q3), [q2] "=&r"(q2), [q1] "=&r"(q1), [q0] "=&r"(q0), + [q7] "=&r"(q7), [q6] "=&r"(q6), [q5] "=&r"(q5), [q4] "=&r"(q4) + : [s1] "r"(s1), [s2] "r"(s2), [s3] "r"(s3), [s4] "r"(s4)); + + /* transpose p3, p2, p1, p0 + original (when loaded from memory) + register -4 -3 -2 -1 + p0 p0_0 p0_1 p0_2 p0_3 + p1 p1_0 p1_1 p1_2 p1_3 + p2 p2_0 p2_1 p2_2 p2_3 + p3 p3_0 p3_1 p3_2 p3_3 + + after transpose + register + p0 p3_3 p2_3 p1_3 p0_3 + p1 p3_2 p2_2 p1_2 p0_2 + p2 p3_1 p2_1 p1_1 p0_1 + p3 p3_0 p2_0 p1_0 p0_0 + */ + __asm__ __volatile__( + "precrq.qb.ph %[prim1], %[p0], %[p1] \n\t" + "precr.qb.ph %[prim2], %[p0], %[p1] \n\t" + "precrq.qb.ph %[prim3], %[p2], %[p3] \n\t" + "precr.qb.ph %[prim4], %[p2], %[p3] \n\t" + + "precrq.qb.ph %[p1], %[prim1], %[prim2] \n\t" + "precr.qb.ph %[p3], %[prim1], %[prim2] \n\t" + "precrq.qb.ph %[sec3], %[prim3], %[prim4] \n\t" + "precr.qb.ph %[sec4], %[prim3], %[prim4] \n\t" + + "precrq.ph.w %[p0], %[p1], %[sec3] \n\t" + "precrq.ph.w %[p2], %[p3], %[sec4] \n\t" + "append %[p1], %[sec3], 16 \n\t" + "append %[p3], %[sec4], 16 \n\t" + + : [prim1] "=&r"(prim1), [prim2] "=&r"(prim2), [prim3] "=&r"(prim3), + [prim4] "=&r"(prim4), [p0] "+r"(p0), [p1] "+r"(p1), [p2] "+r"(p2), + [p3] "+r"(p3), [sec3] "=&r"(sec3), [sec4] "=&r"(sec4) + :); + + /* transpose q0, q1, q2, q3 + original (when loaded from memory) + register +1 +2 +3 +4 + q3 q3_0 q3_1 q3_2 q3_3 + q2 q2_0 q2_1 q2_2 q2_3 + q1 q1_0 q1_1 q1_2 q1_3 + q0 q0_0 q0_1 q0_2 q0_3 + + after transpose + register + q3 q0_3 q1_3 q2_3 q3_3 + q2 q0_2 q1_2 q2_2 q3_2 + q1 q0_1 q1_1 q2_1 q3_1 + q0 q0_0 q1_0 q2_0 q3_0 + */ + __asm__ __volatile__( + "precrq.qb.ph %[prim1], %[q3], %[q2] \n\t" + "precr.qb.ph %[prim2], %[q3], %[q2] \n\t" + "precrq.qb.ph %[prim3], %[q1], %[q0] \n\t" + "precr.qb.ph %[prim4], %[q1], %[q0] \n\t" + + "precrq.qb.ph %[q2], %[prim1], %[prim2] \n\t" + "precr.qb.ph %[q0], %[prim1], %[prim2] \n\t" + "precrq.qb.ph %[sec3], %[prim3], %[prim4] \n\t" + "precr.qb.ph %[sec4], %[prim3], %[prim4] \n\t" + + "precrq.ph.w %[q3], %[q2], %[sec3] \n\t" + "precrq.ph.w %[q1], %[q0], %[sec4] \n\t" + "append %[q2], %[sec3], 16 \n\t" + "append %[q0], %[sec4], 16 \n\t" + + : [prim1] "=&r"(prim1), [prim2] "=&r"(prim2), [prim3] "=&r"(prim3), + [prim4] "=&r"(prim4), [q3] "+r"(q3), [q2] "+r"(q2), [q1] "+r"(q1), + [q0] "+r"(q0), [sec3] "=&r"(sec3), [sec4] "=&r"(sec4) + :); + + /* transpose p7, p6, p5, p4 + original (when loaded from memory) + register -8 -7 -6 -5 + p4 p4_0 p4_1 p4_2 p4_3 + p5 p5_0 p5_1 p5_2 p5_3 + p6 p6_0 p6_1 p6_2 p6_3 + p7 p7_0 p7_1 p7_2 p7_3 + + after transpose + register + p4 p7_3 p6_3 p5_3 p4_3 + p5 p7_2 p6_2 p5_2 p4_2 + p6 p7_1 p6_1 p5_1 p4_1 + p7 p7_0 p6_0 p5_0 p4_0 + */ + __asm__ __volatile__( + "precrq.qb.ph %[prim1], %[p4], %[p5] \n\t" + "precr.qb.ph %[prim2], %[p4], %[p5] \n\t" + "precrq.qb.ph %[prim3], %[p6], %[p7] \n\t" + "precr.qb.ph %[prim4], %[p6], %[p7] \n\t" + + "precrq.qb.ph %[p5], %[prim1], %[prim2] \n\t" + "precr.qb.ph %[p7], %[prim1], %[prim2] \n\t" + "precrq.qb.ph %[sec3], %[prim3], %[prim4] \n\t" + "precr.qb.ph %[sec4], %[prim3], %[prim4] \n\t" + + "precrq.ph.w %[p4], %[p5], %[sec3] \n\t" + "precrq.ph.w %[p6], %[p7], %[sec4] \n\t" + "append %[p5], %[sec3], 16 \n\t" + "append %[p7], %[sec4], 16 \n\t" + + : [prim1] "=&r"(prim1), [prim2] "=&r"(prim2), [prim3] "=&r"(prim3), + [prim4] "=&r"(prim4), [p4] "+r"(p4), [p5] "+r"(p5), [p6] "+r"(p6), + [p7] "+r"(p7), [sec3] "=&r"(sec3), [sec4] "=&r"(sec4) + :); + + /* transpose q4, q5, q6, q7 + original (when loaded from memory) + register +5 +6 +7 +8 + q7 q7_0 q7_1 q7_2 q7_3 + q6 q6_0 q6_1 q6_2 q6_3 + q5 q5_0 q5_1 q5_2 q5_3 + q4 q4_0 q4_1 q4_2 q4_3 + + after transpose + register + q7 q4_3 q5_3 q26_3 q7_3 + q6 q4_2 q5_2 q26_2 q7_2 + q5 q4_1 q5_1 q26_1 q7_1 + q4 q4_0 q5_0 q26_0 q7_0 + */ + __asm__ __volatile__( + "precrq.qb.ph %[prim1], %[q7], %[q6] \n\t" + "precr.qb.ph %[prim2], %[q7], %[q6] \n\t" + "precrq.qb.ph %[prim3], %[q5], %[q4] \n\t" + "precr.qb.ph %[prim4], %[q5], %[q4] \n\t" + + "precrq.qb.ph %[q6], %[prim1], %[prim2] \n\t" + "precr.qb.ph %[q4], %[prim1], %[prim2] \n\t" + "precrq.qb.ph %[sec3], %[prim3], %[prim4] \n\t" + "precr.qb.ph %[sec4], %[prim3], %[prim4] \n\t" + + "precrq.ph.w %[q7], %[q6], %[sec3] \n\t" + "precrq.ph.w %[q5], %[q4], %[sec4] \n\t" + "append %[q6], %[sec3], 16 \n\t" + "append %[q4], %[sec4], 16 \n\t" + + : [prim1] "=&r"(prim1), [prim2] "=&r"(prim2), [prim3] "=&r"(prim3), + [prim4] "=&r"(prim4), [q7] "+r"(q7), [q6] "+r"(q6), [q5] "+r"(q5), + [q4] "+r"(q4), [sec3] "=&r"(sec3), [sec4] "=&r"(sec4) + :); + + filter_hev_mask_flatmask4_dspr2(limit_vec, flimit_vec, thresh_vec, p1, p0, + p3, p2, q0, q1, q2, q3, &hev, &mask, &flat); + + flatmask5(p7, p6, p5, p4, p0, q0, q4, q5, q6, q7, &flat2); + + /* f0 */ + if (((flat2 == 0) && (flat == 0) && (mask != 0)) || + ((flat2 != 0) && (flat == 0) && (mask != 0))) { + filter1_dspr2(mask, hev, p1, p0, q0, q1, &p1_f0, &p0_f0, &q0_f0, &q1_f0); + STORE_F0() + } else if ((flat2 == 0XFFFFFFFF) && (flat == 0xFFFFFFFF) && + (mask == 0xFFFFFFFF)) { + /* f2 */ + PACK_LEFT_0TO3() + PACK_LEFT_4TO7() + wide_mbfilter_dspr2(&p7_l, &p6_l, &p5_l, &p4_l, &p3_l, &p2_l, &p1_l, + &p0_l, &q0_l, &q1_l, &q2_l, &q3_l, &q4_l, &q5_l, + &q6_l, &q7_l); + + PACK_RIGHT_0TO3() + PACK_RIGHT_4TO7() + wide_mbfilter_dspr2(&p7_r, &p6_r, &p5_r, &p4_r, &p3_r, &p2_r, &p1_r, + &p0_r, &q0_r, &q1_r, &q2_r, &q3_r, &q4_r, &q5_r, + &q6_r, &q7_r); + + STORE_F2() + } else if ((flat2 == 0) && (flat == 0xFFFFFFFF) && (mask == 0xFFFFFFFF)) { + /* f1 */ + PACK_LEFT_0TO3() + mbfilter_dspr2(&p3_l, &p2_l, &p1_l, &p0_l, &q0_l, &q1_l, &q2_l, &q3_l); + + PACK_RIGHT_0TO3() + mbfilter_dspr2(&p3_r, &p2_r, &p1_r, &p0_r, &q0_r, &q1_r, &q2_r, &q3_r); + + STORE_F1() + } else if ((flat2 == 0) && (flat != 0) && (mask != 0)) { + /* f0 + f1 */ + filter1_dspr2(mask, hev, p1, p0, q0, q1, &p1_f0, &p0_f0, &q0_f0, &q1_f0); + + /* left 2 element operation */ + PACK_LEFT_0TO3() + mbfilter_dspr2(&p3_l, &p2_l, &p1_l, &p0_l, &q0_l, &q1_l, &q2_l, &q3_l); + + /* right 2 element operation */ + PACK_RIGHT_0TO3() + mbfilter_dspr2(&p3_r, &p2_r, &p1_r, &p0_r, &q0_r, &q1_r, &q2_r, &q3_r); + + if (mask & flat & 0x000000FF) { + __asm__ __volatile__( + "sb %[p2_r], -3(%[s4]) \n\t" + "sb %[p1_r], -2(%[s4]) \n\t" + "sb %[p0_r], -1(%[s4]) \n\t" + "sb %[q0_r], (%[s4]) \n\t" + "sb %[q1_r], +1(%[s4]) \n\t" + "sb %[q2_r], +2(%[s4]) \n\t" + + : + : [p2_r] "r"(p2_r), [p1_r] "r"(p1_r), [p0_r] "r"(p0_r), + [q0_r] "r"(q0_r), [q1_r] "r"(q1_r), [q2_r] "r"(q2_r), + [s4] "r"(s4)); + } else if (mask & 0x000000FF) { + __asm__ __volatile__( + "sb %[p1_f0], -2(%[s4]) \n\t" + "sb %[p0_f0], -1(%[s4]) \n\t" + "sb %[q0_f0], (%[s4]) \n\t" + "sb %[q1_f0], +1(%[s4]) \n\t" + + : + : [p1_f0] "r"(p1_f0), [p0_f0] "r"(p0_f0), [q0_f0] "r"(q0_f0), + [q1_f0] "r"(q1_f0), [s4] "r"(s4)); + } + + __asm__ __volatile__( + "srl %[p2_r], %[p2_r], 16 \n\t" + "srl %[p1_r], %[p1_r], 16 \n\t" + "srl %[p0_r], %[p0_r], 16 \n\t" + "srl %[q0_r], %[q0_r], 16 \n\t" + "srl %[q1_r], %[q1_r], 16 \n\t" + "srl %[q2_r], %[q2_r], 16 \n\t" + "srl %[p1_f0], %[p1_f0], 8 \n\t" + "srl %[p0_f0], %[p0_f0], 8 \n\t" + "srl %[q0_f0], %[q0_f0], 8 \n\t" + "srl %[q1_f0], %[q1_f0], 8 \n\t" + + : [p2_r] "+r"(p2_r), [p1_r] "+r"(p1_r), [p0_r] "+r"(p0_r), + [q0_r] "+r"(q0_r), [q1_r] "+r"(q1_r), [q2_r] "+r"(q2_r), + [p1_f0] "+r"(p1_f0), [p0_f0] "+r"(p0_f0), [q0_f0] "+r"(q0_f0), + [q1_f0] "+r"(q1_f0) + :); + + if (mask & flat & 0x0000FF00) { + __asm__ __volatile__( + "sb %[p2_r], -3(%[s3]) \n\t" + "sb %[p1_r], -2(%[s3]) \n\t" + "sb %[p0_r], -1(%[s3]) \n\t" + "sb %[q0_r], (%[s3]) \n\t" + "sb %[q1_r], +1(%[s3]) \n\t" + "sb %[q2_r], +2(%[s3]) \n\t" + + : + : [p2_r] "r"(p2_r), [p1_r] "r"(p1_r), [p0_r] "r"(p0_r), + [q0_r] "r"(q0_r), [q1_r] "r"(q1_r), [q2_r] "r"(q2_r), + [s3] "r"(s3)); + } else if (mask & 0x0000FF00) { + __asm__ __volatile__( + "sb %[p1_f0], -2(%[s3]) \n\t" + "sb %[p0_f0], -1(%[s3]) \n\t" + "sb %[q0_f0], (%[s3]) \n\t" + "sb %[q1_f0], +1(%[s3]) \n\t" + + : + : [p1_f0] "r"(p1_f0), [p0_f0] "r"(p0_f0), [q0_f0] "r"(q0_f0), + [q1_f0] "r"(q1_f0), [s3] "r"(s3)); + } + + __asm__ __volatile__( + "srl %[p1_f0], %[p1_f0], 8 \n\t" + "srl %[p0_f0], %[p0_f0], 8 \n\t" + "srl %[q0_f0], %[q0_f0], 8 \n\t" + "srl %[q1_f0], %[q1_f0], 8 \n\t" + + : [p1_f0] "+r"(p1_f0), [p0_f0] "+r"(p0_f0), [q0_f0] "+r"(q0_f0), + [q1_f0] "+r"(q1_f0) + :); + + if (mask & flat & 0x00FF0000) { + __asm__ __volatile__( + "sb %[p2_l], -3(%[s2]) \n\t" + "sb %[p1_l], -2(%[s2]) \n\t" + "sb %[p0_l], -1(%[s2]) \n\t" + "sb %[q0_l], (%[s2]) \n\t" + "sb %[q1_l], +1(%[s2]) \n\t" + "sb %[q2_l], +2(%[s2]) \n\t" + + : + : [p2_l] "r"(p2_l), [p1_l] "r"(p1_l), [p0_l] "r"(p0_l), + [q0_l] "r"(q0_l), [q1_l] "r"(q1_l), [q2_l] "r"(q2_l), + [s2] "r"(s2)); + } else if (mask & 0x00FF0000) { + __asm__ __volatile__( + "sb %[p1_f0], -2(%[s2]) \n\t" + "sb %[p0_f0], -1(%[s2]) \n\t" + "sb %[q0_f0], (%[s2]) \n\t" + "sb %[q1_f0], +1(%[s2]) \n\t" + + : + : [p1_f0] "r"(p1_f0), [p0_f0] "r"(p0_f0), [q0_f0] "r"(q0_f0), + [q1_f0] "r"(q1_f0), [s2] "r"(s2)); + } + + __asm__ __volatile__( + "srl %[p2_l], %[p2_l], 16 \n\t" + "srl %[p1_l], %[p1_l], 16 \n\t" + "srl %[p0_l], %[p0_l], 16 \n\t" + "srl %[q0_l], %[q0_l], 16 \n\t" + "srl %[q1_l], %[q1_l], 16 \n\t" + "srl %[q2_l], %[q2_l], 16 \n\t" + "srl %[p1_f0], %[p1_f0], 8 \n\t" + "srl %[p0_f0], %[p0_f0], 8 \n\t" + "srl %[q0_f0], %[q0_f0], 8 \n\t" + "srl %[q1_f0], %[q1_f0], 8 \n\t" + + : [p2_l] "+r"(p2_l), [p1_l] "+r"(p1_l), [p0_l] "+r"(p0_l), + [q0_l] "+r"(q0_l), [q1_l] "+r"(q1_l), [q2_l] "+r"(q2_l), + [p1_f0] "+r"(p1_f0), [p0_f0] "+r"(p0_f0), [q0_f0] "+r"(q0_f0), + [q1_f0] "+r"(q1_f0) + :); + + if (mask & flat & 0xFF000000) { + __asm__ __volatile__( + "sb %[p2_l], -3(%[s1]) \n\t" + "sb %[p1_l], -2(%[s1]) \n\t" + "sb %[p0_l], -1(%[s1]) \n\t" + "sb %[q0_l], (%[s1]) \n\t" + "sb %[q1_l], +1(%[s1]) \n\t" + "sb %[q2_l], +2(%[s1]) \n\t" + + : + : [p2_l] "r"(p2_l), [p1_l] "r"(p1_l), [p0_l] "r"(p0_l), + [q0_l] "r"(q0_l), [q1_l] "r"(q1_l), [q2_l] "r"(q2_l), + [s1] "r"(s1)); + } else if (mask & 0xFF000000) { + __asm__ __volatile__( + "sb %[p1_f0], -2(%[s1]) \n\t" + "sb %[p0_f0], -1(%[s1]) \n\t" + "sb %[q0_f0], (%[s1]) \n\t" + "sb %[q1_f0], +1(%[s1]) \n\t" + + : + : [p1_f0] "r"(p1_f0), [p0_f0] "r"(p0_f0), [q0_f0] "r"(q0_f0), + [q1_f0] "r"(q1_f0), [s1] "r"(s1)); + } + } else if ((flat2 != 0) && (flat != 0) && (mask != 0)) { + /* f0+f1+f2 */ + filter1_dspr2(mask, hev, p1, p0, q0, q1, &p1_f0, &p0_f0, &q0_f0, &q1_f0); + + PACK_LEFT_0TO3() + mbfilter1_dspr2(p3_l, p2_l, p1_l, p0_l, q0_l, q1_l, q2_l, q3_l, &p2_l_f1, + &p1_l_f1, &p0_l_f1, &q0_l_f1, &q1_l_f1, &q2_l_f1); + + PACK_RIGHT_0TO3() + mbfilter1_dspr2(p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r, &p2_r_f1, + &p1_r_f1, &p0_r_f1, &q0_r_f1, &q1_r_f1, &q2_r_f1); + + PACK_LEFT_4TO7() + wide_mbfilter_dspr2(&p7_l, &p6_l, &p5_l, &p4_l, &p3_l, &p2_l, &p1_l, + &p0_l, &q0_l, &q1_l, &q2_l, &q3_l, &q4_l, &q5_l, + &q6_l, &q7_l); + + PACK_RIGHT_4TO7() + wide_mbfilter_dspr2(&p7_r, &p6_r, &p5_r, &p4_r, &p3_r, &p2_r, &p1_r, + &p0_r, &q0_r, &q1_r, &q2_r, &q3_r, &q4_r, &q5_r, + &q6_r, &q7_r); + + if (mask & flat & flat2 & 0x000000FF) { + __asm__ __volatile__( + "sb %[p6_r], -7(%[s4]) \n\t" + "sb %[p5_r], -6(%[s4]) \n\t" + "sb %[p4_r], -5(%[s4]) \n\t" + "sb %[p3_r], -4(%[s4]) \n\t" + "sb %[p2_r], -3(%[s4]) \n\t" + "sb %[p1_r], -2(%[s4]) \n\t" + "sb %[p0_r], -1(%[s4]) \n\t" + + : + : [p6_r] "r"(p6_r), [p5_r] "r"(p5_r), [p4_r] "r"(p4_r), + [p3_r] "r"(p3_r), [p2_r] "r"(p2_r), [p1_r] "r"(p1_r), + [p0_r] "r"(p0_r), [s4] "r"(s4)); + + __asm__ __volatile__( + "sb %[q0_r], (%[s4]) \n\t" + "sb %[q1_r], +1(%[s4]) \n\t" + "sb %[q2_r], +2(%[s4]) \n\t" + "sb %[q3_r], +3(%[s4]) \n\t" + "sb %[q4_r], +4(%[s4]) \n\t" + "sb %[q5_r], +5(%[s4]) \n\t" + "sb %[q6_r], +6(%[s4]) \n\t" + + : + : [q0_r] "r"(q0_r), [q1_r] "r"(q1_r), [q2_r] "r"(q2_r), + [q3_r] "r"(q3_r), [q4_r] "r"(q4_r), [q5_r] "r"(q5_r), + [q6_r] "r"(q6_r), [s4] "r"(s4)); + } else if (mask & flat & 0x000000FF) { + __asm__ __volatile__( + "sb %[p2_r_f1], -3(%[s4]) \n\t" + "sb %[p1_r_f1], -2(%[s4]) \n\t" + "sb %[p0_r_f1], -1(%[s4]) \n\t" + "sb %[q0_r_f1], (%[s4]) \n\t" + "sb %[q1_r_f1], +1(%[s4]) \n\t" + "sb %[q2_r_f1], +2(%[s4]) \n\t" + + : + : [p2_r_f1] "r"(p2_r_f1), [p1_r_f1] "r"(p1_r_f1), + [p0_r_f1] "r"(p0_r_f1), [q0_r_f1] "r"(q0_r_f1), + [q1_r_f1] "r"(q1_r_f1), [q2_r_f1] "r"(q2_r_f1), [s4] "r"(s4)); + } else if (mask & 0x000000FF) { + __asm__ __volatile__( + "sb %[p1_f0], -2(%[s4]) \n\t" + "sb %[p0_f0], -1(%[s4]) \n\t" + "sb %[q0_f0], (%[s4]) \n\t" + "sb %[q1_f0], +1(%[s4]) \n\t" + + : + : [p1_f0] "r"(p1_f0), [p0_f0] "r"(p0_f0), [q0_f0] "r"(q0_f0), + [q1_f0] "r"(q1_f0), [s4] "r"(s4)); + } + + __asm__ __volatile__( + "srl %[p6_r], %[p6_r], 16 \n\t" + "srl %[p5_r], %[p5_r], 16 \n\t" + "srl %[p4_r], %[p4_r], 16 \n\t" + "srl %[p3_r], %[p3_r], 16 \n\t" + "srl %[p2_r], %[p2_r], 16 \n\t" + "srl %[p1_r], %[p1_r], 16 \n\t" + "srl %[p0_r], %[p0_r], 16 \n\t" + "srl %[q0_r], %[q0_r], 16 \n\t" + "srl %[q1_r], %[q1_r], 16 \n\t" + "srl %[q2_r], %[q2_r], 16 \n\t" + "srl %[q3_r], %[q3_r], 16 \n\t" + "srl %[q4_r], %[q4_r], 16 \n\t" + "srl %[q5_r], %[q5_r], 16 \n\t" + "srl %[q6_r], %[q6_r], 16 \n\t" + + : [q0_r] "+r"(q0_r), [q1_r] "+r"(q1_r), [q2_r] "+r"(q2_r), + [q3_r] "+r"(q3_r), [q4_r] "+r"(q4_r), [q5_r] "+r"(q5_r), + [q6_r] "+r"(q6_r), [p6_r] "+r"(p6_r), [p5_r] "+r"(p5_r), + [p4_r] "+r"(p4_r), [p3_r] "+r"(p3_r), [p2_r] "+r"(p2_r), + [p1_r] "+r"(p1_r), [p0_r] "+r"(p0_r) + :); + + __asm__ __volatile__( + "srl %[p2_r_f1], %[p2_r_f1], 16 \n\t" + "srl %[p1_r_f1], %[p1_r_f1], 16 \n\t" + "srl %[p0_r_f1], %[p0_r_f1], 16 \n\t" + "srl %[q0_r_f1], %[q0_r_f1], 16 \n\t" + "srl %[q1_r_f1], %[q1_r_f1], 16 \n\t" + "srl %[q2_r_f1], %[q2_r_f1], 16 \n\t" + "srl %[p1_f0], %[p1_f0], 8 \n\t" + "srl %[p0_f0], %[p0_f0], 8 \n\t" + "srl %[q0_f0], %[q0_f0], 8 \n\t" + "srl %[q1_f0], %[q1_f0], 8 \n\t" + + : [p2_r_f1] "+r"(p2_r_f1), [p1_r_f1] "+r"(p1_r_f1), + [p0_r_f1] "+r"(p0_r_f1), [q0_r_f1] "+r"(q0_r_f1), + [q1_r_f1] "+r"(q1_r_f1), [q2_r_f1] "+r"(q2_r_f1), + [p1_f0] "+r"(p1_f0), [p0_f0] "+r"(p0_f0), [q0_f0] "+r"(q0_f0), + [q1_f0] "+r"(q1_f0) + :); + + if (mask & flat & flat2 & 0x0000FF00) { + __asm__ __volatile__( + "sb %[p6_r], -7(%[s3]) \n\t" + "sb %[p5_r], -6(%[s3]) \n\t" + "sb %[p4_r], -5(%[s3]) \n\t" + "sb %[p3_r], -4(%[s3]) \n\t" + "sb %[p2_r], -3(%[s3]) \n\t" + "sb %[p1_r], -2(%[s3]) \n\t" + "sb %[p0_r], -1(%[s3]) \n\t" + + : + : [p6_r] "r"(p6_r), [p5_r] "r"(p5_r), [p4_r] "r"(p4_r), + [p3_r] "r"(p3_r), [p2_r] "r"(p2_r), [p1_r] "r"(p1_r), + [p0_r] "r"(p0_r), [s3] "r"(s3)); + + __asm__ __volatile__( + "sb %[q0_r], (%[s3]) \n\t" + "sb %[q1_r], +1(%[s3]) \n\t" + "sb %[q2_r], +2(%[s3]) \n\t" + "sb %[q3_r], +3(%[s3]) \n\t" + "sb %[q4_r], +4(%[s3]) \n\t" + "sb %[q5_r], +5(%[s3]) \n\t" + "sb %[q6_r], +6(%[s3]) \n\t" + + : + : [q0_r] "r"(q0_r), [q1_r] "r"(q1_r), [q2_r] "r"(q2_r), + [q3_r] "r"(q3_r), [q4_r] "r"(q4_r), [q5_r] "r"(q5_r), + [q6_r] "r"(q6_r), [s3] "r"(s3)); + } else if (mask & flat & 0x0000FF00) { + __asm__ __volatile__( + "sb %[p2_r_f1], -3(%[s3]) \n\t" + "sb %[p1_r_f1], -2(%[s3]) \n\t" + "sb %[p0_r_f1], -1(%[s3]) \n\t" + "sb %[q0_r_f1], (%[s3]) \n\t" + "sb %[q1_r_f1], +1(%[s3]) \n\t" + "sb %[q2_r_f1], +2(%[s3]) \n\t" + + : + : [p2_r_f1] "r"(p2_r_f1), [p1_r_f1] "r"(p1_r_f1), + [p0_r_f1] "r"(p0_r_f1), [q0_r_f1] "r"(q0_r_f1), + [q1_r_f1] "r"(q1_r_f1), [q2_r_f1] "r"(q2_r_f1), [s3] "r"(s3)); + } else if (mask & 0x0000FF00) { + __asm__ __volatile__( + "sb %[p1_f0], -2(%[s3]) \n\t" + "sb %[p0_f0], -1(%[s3]) \n\t" + "sb %[q0_f0], (%[s3]) \n\t" + "sb %[q1_f0], +1(%[s3]) \n\t" + + : + : [p1_f0] "r"(p1_f0), [p0_f0] "r"(p0_f0), [q0_f0] "r"(q0_f0), + [q1_f0] "r"(q1_f0), [s3] "r"(s3)); + } + + __asm__ __volatile__( + "srl %[p1_f0], %[p1_f0], 8 \n\t" + "srl %[p0_f0], %[p0_f0], 8 \n\t" + "srl %[q0_f0], %[q0_f0], 8 \n\t" + "srl %[q1_f0], %[q1_f0], 8 \n\t" + + : [p1_f0] "+r"(p1_f0), [p0_f0] "+r"(p0_f0), [q0_f0] "+r"(q0_f0), + [q1_f0] "+r"(q1_f0) + :); + + if (mask & flat & flat2 & 0x00FF0000) { + __asm__ __volatile__( + "sb %[p6_l], -7(%[s2]) \n\t" + "sb %[p5_l], -6(%[s2]) \n\t" + "sb %[p4_l], -5(%[s2]) \n\t" + "sb %[p3_l], -4(%[s2]) \n\t" + "sb %[p2_l], -3(%[s2]) \n\t" + "sb %[p1_l], -2(%[s2]) \n\t" + "sb %[p0_l], -1(%[s2]) \n\t" + + : + : [p6_l] "r"(p6_l), [p5_l] "r"(p5_l), [p4_l] "r"(p4_l), + [p3_l] "r"(p3_l), [p2_l] "r"(p2_l), [p1_l] "r"(p1_l), + [p0_l] "r"(p0_l), [s2] "r"(s2)); + + __asm__ __volatile__( + "sb %[q0_l], (%[s2]) \n\t" + "sb %[q1_l], +1(%[s2]) \n\t" + "sb %[q2_l], +2(%[s2]) \n\t" + "sb %[q3_l], +3(%[s2]) \n\t" + "sb %[q4_l], +4(%[s2]) \n\t" + "sb %[q5_l], +5(%[s2]) \n\t" + "sb %[q6_l], +6(%[s2]) \n\t" + + : + : [q0_l] "r"(q0_l), [q1_l] "r"(q1_l), [q2_l] "r"(q2_l), + [q3_l] "r"(q3_l), [q4_l] "r"(q4_l), [q5_l] "r"(q5_l), + [q6_l] "r"(q6_l), [s2] "r"(s2)); + } else if (mask & flat & 0x00FF0000) { + __asm__ __volatile__( + "sb %[p2_l_f1], -3(%[s2]) \n\t" + "sb %[p1_l_f1], -2(%[s2]) \n\t" + "sb %[p0_l_f1], -1(%[s2]) \n\t" + "sb %[q0_l_f1], (%[s2]) \n\t" + "sb %[q1_l_f1], +1(%[s2]) \n\t" + "sb %[q2_l_f1], +2(%[s2]) \n\t" + + : + : [p2_l_f1] "r"(p2_l_f1), [p1_l_f1] "r"(p1_l_f1), + [p0_l_f1] "r"(p0_l_f1), [q0_l_f1] "r"(q0_l_f1), + [q1_l_f1] "r"(q1_l_f1), [q2_l_f1] "r"(q2_l_f1), [s2] "r"(s2)); + } else if (mask & 0x00FF0000) { + __asm__ __volatile__( + "sb %[p1_f0], -2(%[s2]) \n\t" + "sb %[p0_f0], -1(%[s2]) \n\t" + "sb %[q0_f0], (%[s2]) \n\t" + "sb %[q1_f0], +1(%[s2]) \n\t" + + : + : [p1_f0] "r"(p1_f0), [p0_f0] "r"(p0_f0), [q0_f0] "r"(q0_f0), + [q1_f0] "r"(q1_f0), [s2] "r"(s2)); + } + + __asm__ __volatile__( + "srl %[p6_l], %[p6_l], 16 \n\t" + "srl %[p5_l], %[p5_l], 16 \n\t" + "srl %[p4_l], %[p4_l], 16 \n\t" + "srl %[p3_l], %[p3_l], 16 \n\t" + "srl %[p2_l], %[p2_l], 16 \n\t" + "srl %[p1_l], %[p1_l], 16 \n\t" + "srl %[p0_l], %[p0_l], 16 \n\t" + "srl %[q0_l], %[q0_l], 16 \n\t" + "srl %[q1_l], %[q1_l], 16 \n\t" + "srl %[q2_l], %[q2_l], 16 \n\t" + "srl %[q3_l], %[q3_l], 16 \n\t" + "srl %[q4_l], %[q4_l], 16 \n\t" + "srl %[q5_l], %[q5_l], 16 \n\t" + "srl %[q6_l], %[q6_l], 16 \n\t" + + : [q0_l] "+r"(q0_l), [q1_l] "+r"(q1_l), [q2_l] "+r"(q2_l), + [q3_l] "+r"(q3_l), [q4_l] "+r"(q4_l), [q5_l] "+r"(q5_l), + [q6_l] "+r"(q6_l), [p6_l] "+r"(p6_l), [p5_l] "+r"(p5_l), + [p4_l] "+r"(p4_l), [p3_l] "+r"(p3_l), [p2_l] "+r"(p2_l), + [p1_l] "+r"(p1_l), [p0_l] "+r"(p0_l) + :); + + __asm__ __volatile__( + "srl %[p2_l_f1], %[p2_l_f1], 16 \n\t" + "srl %[p1_l_f1], %[p1_l_f1], 16 \n\t" + "srl %[p0_l_f1], %[p0_l_f1], 16 \n\t" + "srl %[q0_l_f1], %[q0_l_f1], 16 \n\t" + "srl %[q1_l_f1], %[q1_l_f1], 16 \n\t" + "srl %[q2_l_f1], %[q2_l_f1], 16 \n\t" + "srl %[p1_f0], %[p1_f0], 8 \n\t" + "srl %[p0_f0], %[p0_f0], 8 \n\t" + "srl %[q0_f0], %[q0_f0], 8 \n\t" + "srl %[q1_f0], %[q1_f0], 8 \n\t" + + : [p2_l_f1] "+r"(p2_l_f1), [p1_l_f1] "+r"(p1_l_f1), + [p0_l_f1] "+r"(p0_l_f1), [q0_l_f1] "+r"(q0_l_f1), + [q1_l_f1] "+r"(q1_l_f1), [q2_l_f1] "+r"(q2_l_f1), + [p1_f0] "+r"(p1_f0), [p0_f0] "+r"(p0_f0), [q0_f0] "+r"(q0_f0), + [q1_f0] "+r"(q1_f0) + :); + + if (mask & flat & flat2 & 0xFF000000) { + __asm__ __volatile__( + "sb %[p6_l], -7(%[s1]) \n\t" + "sb %[p5_l], -6(%[s1]) \n\t" + "sb %[p4_l], -5(%[s1]) \n\t" + "sb %[p3_l], -4(%[s1]) \n\t" + "sb %[p2_l], -3(%[s1]) \n\t" + "sb %[p1_l], -2(%[s1]) \n\t" + "sb %[p0_l], -1(%[s1]) \n\t" + + : + : [p6_l] "r"(p6_l), [p5_l] "r"(p5_l), [p4_l] "r"(p4_l), + [p3_l] "r"(p3_l), [p2_l] "r"(p2_l), [p1_l] "r"(p1_l), + [p0_l] "r"(p0_l), [s1] "r"(s1)); + + __asm__ __volatile__( + "sb %[q0_l], (%[s1]) \n\t" + "sb %[q1_l], 1(%[s1]) \n\t" + "sb %[q2_l], 2(%[s1]) \n\t" + "sb %[q3_l], 3(%[s1]) \n\t" + "sb %[q4_l], 4(%[s1]) \n\t" + "sb %[q5_l], 5(%[s1]) \n\t" + "sb %[q6_l], 6(%[s1]) \n\t" + + : + : [q0_l] "r"(q0_l), [q1_l] "r"(q1_l), [q2_l] "r"(q2_l), + [q3_l] "r"(q3_l), [q4_l] "r"(q4_l), [q5_l] "r"(q5_l), + [q6_l] "r"(q6_l), [s1] "r"(s1)); + } else if (mask & flat & 0xFF000000) { + __asm__ __volatile__( + "sb %[p2_l_f1], -3(%[s1]) \n\t" + "sb %[p1_l_f1], -2(%[s1]) \n\t" + "sb %[p0_l_f1], -1(%[s1]) \n\t" + "sb %[q0_l_f1], (%[s1]) \n\t" + "sb %[q1_l_f1], +1(%[s1]) \n\t" + "sb %[q2_l_f1], +2(%[s1]) \n\t" + + : + : [p2_l_f1] "r"(p2_l_f1), [p1_l_f1] "r"(p1_l_f1), + [p0_l_f1] "r"(p0_l_f1), [q0_l_f1] "r"(q0_l_f1), + [q1_l_f1] "r"(q1_l_f1), [q2_l_f1] "r"(q2_l_f1), [s1] "r"(s1)); + } else if (mask & 0xFF000000) { + __asm__ __volatile__( + "sb %[p1_f0], -2(%[s1]) \n\t" + "sb %[p0_f0], -1(%[s1]) \n\t" + "sb %[q0_f0], (%[s1]) \n\t" + "sb %[q1_f0], +1(%[s1]) \n\t" + + : + : [p1_f0] "r"(p1_f0), [p0_f0] "r"(p0_f0), [q0_f0] "r"(q0_f0), + [q1_f0] "r"(q1_f0), [s1] "r"(s1)); + } + } + } +} +#endif // #if HAVE_DSPR2 diff --git a/third_party/aom/aom_dsp/mips/loopfilter_msa.h b/third_party/aom/aom_dsp/mips/loopfilter_msa.h new file mode 100644 index 0000000000..54b0bb4bd9 --- /dev/null +++ b/third_party/aom/aom_dsp/mips/loopfilter_msa.h @@ -0,0 +1,251 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_MIPS_LOOPFILTER_MSA_H_ +#define AOM_AOM_DSP_MIPS_LOOPFILTER_MSA_H_ + +#include "aom_dsp/mips/macros_msa.h" + +#define AOM_LPF_FILTER4_8W(p1_in, p0_in, q0_in, q1_in, mask_in, hev_in, \ + p1_out, p0_out, q0_out, q1_out) \ + { \ + v16i8 p1_m, p0_m, q0_m, q1_m, q0_sub_p0, filt_sign; \ + v16i8 filt, filt1, filt2, cnst4b, cnst3b; \ + v8i16 q0_sub_p0_r, filt_r, cnst3h; \ + \ + p1_m = (v16i8)__msa_xori_b(p1_in, 0x80); \ + p0_m = (v16i8)__msa_xori_b(p0_in, 0x80); \ + q0_m = (v16i8)__msa_xori_b(q0_in, 0x80); \ + q1_m = (v16i8)__msa_xori_b(q1_in, 0x80); \ + \ + filt = __msa_subs_s_b(p1_m, q1_m); \ + filt = filt & (v16i8)hev_in; \ + q0_sub_p0 = q0_m - p0_m; \ + filt_sign = __msa_clti_s_b(filt, 0); \ + \ + cnst3h = __msa_ldi_h(3); \ + q0_sub_p0_r = (v8i16)__msa_ilvr_b(q0_sub_p0, q0_sub_p0); \ + q0_sub_p0_r = __msa_dotp_s_h((v16i8)q0_sub_p0_r, (v16i8)cnst3h); \ + filt_r = (v8i16)__msa_ilvr_b(filt_sign, filt); \ + filt_r += q0_sub_p0_r; \ + filt_r = __msa_sat_s_h(filt_r, 7); \ + \ + /* combine left and right part */ \ + filt = __msa_pckev_b((v16i8)filt_r, (v16i8)filt_r); \ + \ + filt = filt & (v16i8)mask_in; \ + cnst4b = __msa_ldi_b(4); \ + filt1 = __msa_adds_s_b(filt, cnst4b); \ + filt1 >>= 3; \ + \ + cnst3b = __msa_ldi_b(3); \ + filt2 = __msa_adds_s_b(filt, cnst3b); \ + filt2 >>= 3; \ + \ + q0_m = __msa_subs_s_b(q0_m, filt1); \ + q0_out = __msa_xori_b((v16u8)q0_m, 0x80); \ + p0_m = __msa_adds_s_b(p0_m, filt2); \ + p0_out = __msa_xori_b((v16u8)p0_m, 0x80); \ + \ + filt = __msa_srari_b(filt1, 1); \ + hev_in = __msa_xori_b((v16u8)hev_in, 0xff); \ + filt = filt & (v16i8)hev_in; \ + \ + q1_m = __msa_subs_s_b(q1_m, filt); \ + q1_out = __msa_xori_b((v16u8)q1_m, 0x80); \ + p1_m = __msa_adds_s_b(p1_m, filt); \ + p1_out = __msa_xori_b((v16u8)p1_m, 0x80); \ + } + +#define AOM_LPF_FILTER4_4W(p1_in, p0_in, q0_in, q1_in, mask_in, hev_in, \ + p1_out, p0_out, q0_out, q1_out) \ + { \ + v16i8 p1_m, p0_m, q0_m, q1_m, q0_sub_p0, filt_sign; \ + v16i8 filt, filt1, filt2, cnst4b, cnst3b; \ + v8i16 q0_sub_p0_r, q0_sub_p0_l, filt_l, filt_r, cnst3h; \ + \ + p1_m = (v16i8)__msa_xori_b(p1_in, 0x80); \ + p0_m = (v16i8)__msa_xori_b(p0_in, 0x80); \ + q0_m = (v16i8)__msa_xori_b(q0_in, 0x80); \ + q1_m = (v16i8)__msa_xori_b(q1_in, 0x80); \ + \ + filt = __msa_subs_s_b(p1_m, q1_m); \ + \ + filt = filt & (v16i8)hev_in; \ + \ + q0_sub_p0 = q0_m - p0_m; \ + filt_sign = __msa_clti_s_b(filt, 0); \ + \ + cnst3h = __msa_ldi_h(3); \ + q0_sub_p0_r = (v8i16)__msa_ilvr_b(q0_sub_p0, q0_sub_p0); \ + q0_sub_p0_r = __msa_dotp_s_h((v16i8)q0_sub_p0_r, (v16i8)cnst3h); \ + filt_r = (v8i16)__msa_ilvr_b(filt_sign, filt); \ + filt_r += q0_sub_p0_r; \ + filt_r = __msa_sat_s_h(filt_r, 7); \ + \ + q0_sub_p0_l = (v8i16)__msa_ilvl_b(q0_sub_p0, q0_sub_p0); \ + q0_sub_p0_l = __msa_dotp_s_h((v16i8)q0_sub_p0_l, (v16i8)cnst3h); \ + filt_l = (v8i16)__msa_ilvl_b(filt_sign, filt); \ + filt_l += q0_sub_p0_l; \ + filt_l = __msa_sat_s_h(filt_l, 7); \ + \ + filt = __msa_pckev_b((v16i8)filt_l, (v16i8)filt_r); \ + filt = filt & (v16i8)mask_in; \ + \ + cnst4b = __msa_ldi_b(4); \ + filt1 = __msa_adds_s_b(filt, cnst4b); \ + filt1 >>= 3; \ + \ + cnst3b = __msa_ldi_b(3); \ + filt2 = __msa_adds_s_b(filt, cnst3b); \ + filt2 >>= 3; \ + \ + q0_m = __msa_subs_s_b(q0_m, filt1); \ + q0_out = __msa_xori_b((v16u8)q0_m, 0x80); \ + p0_m = __msa_adds_s_b(p0_m, filt2); \ + p0_out = __msa_xori_b((v16u8)p0_m, 0x80); \ + \ + filt = __msa_srari_b(filt1, 1); \ + hev_in = __msa_xori_b((v16u8)hev_in, 0xff); \ + filt = filt & (v16i8)hev_in; \ + \ + q1_m = __msa_subs_s_b(q1_m, filt); \ + q1_out = __msa_xori_b((v16u8)q1_m, 0x80); \ + p1_m = __msa_adds_s_b(p1_m, filt); \ + p1_out = __msa_xori_b((v16u8)p1_m, 0x80); \ + } + +#define AOM_FLAT4(p3_in, p2_in, p0_in, q0_in, q2_in, q3_in, flat_out) \ + { \ + v16u8 tmp_flat4, p2_a_sub_p0, q2_a_sub_q0, p3_a_sub_p0, q3_a_sub_q0; \ + v16u8 zero_in = { 0 }; \ + \ + tmp_flat4 = __msa_ori_b(zero_in, 1); \ + p2_a_sub_p0 = __msa_asub_u_b(p2_in, p0_in); \ + q2_a_sub_q0 = __msa_asub_u_b(q2_in, q0_in); \ + p3_a_sub_p0 = __msa_asub_u_b(p3_in, p0_in); \ + q3_a_sub_q0 = __msa_asub_u_b(q3_in, q0_in); \ + \ + p2_a_sub_p0 = __msa_max_u_b(p2_a_sub_p0, q2_a_sub_q0); \ + flat_out = __msa_max_u_b(p2_a_sub_p0, flat_out); \ + p3_a_sub_p0 = __msa_max_u_b(p3_a_sub_p0, q3_a_sub_q0); \ + flat_out = __msa_max_u_b(p3_a_sub_p0, flat_out); \ + \ + flat_out = (tmp_flat4 < (v16u8)flat_out); \ + flat_out = __msa_xori_b(flat_out, 0xff); \ + flat_out = flat_out & (mask); \ + } + +#define AOM_FLAT5(p7_in, p6_in, p5_in, p4_in, p0_in, q0_in, q4_in, q5_in, \ + q6_in, q7_in, flat_in, flat2_out) \ + { \ + v16u8 tmp_flat5, zero_in = { 0 }; \ + v16u8 p4_a_sub_p0, q4_a_sub_q0, p5_a_sub_p0, q5_a_sub_q0; \ + v16u8 p6_a_sub_p0, q6_a_sub_q0, p7_a_sub_p0, q7_a_sub_q0; \ + \ + tmp_flat5 = __msa_ori_b(zero_in, 1); \ + p4_a_sub_p0 = __msa_asub_u_b(p4_in, p0_in); \ + q4_a_sub_q0 = __msa_asub_u_b(q4_in, q0_in); \ + p5_a_sub_p0 = __msa_asub_u_b(p5_in, p0_in); \ + q5_a_sub_q0 = __msa_asub_u_b(q5_in, q0_in); \ + p6_a_sub_p0 = __msa_asub_u_b(p6_in, p0_in); \ + q6_a_sub_q0 = __msa_asub_u_b(q6_in, q0_in); \ + p7_a_sub_p0 = __msa_asub_u_b(p7_in, p0_in); \ + q7_a_sub_q0 = __msa_asub_u_b(q7_in, q0_in); \ + \ + p4_a_sub_p0 = __msa_max_u_b(p4_a_sub_p0, q4_a_sub_q0); \ + flat2_out = __msa_max_u_b(p5_a_sub_p0, q5_a_sub_q0); \ + flat2_out = __msa_max_u_b(p4_a_sub_p0, flat2_out); \ + p6_a_sub_p0 = __msa_max_u_b(p6_a_sub_p0, q6_a_sub_q0); \ + flat2_out = __msa_max_u_b(p6_a_sub_p0, flat2_out); \ + p7_a_sub_p0 = __msa_max_u_b(p7_a_sub_p0, q7_a_sub_q0); \ + flat2_out = __msa_max_u_b(p7_a_sub_p0, flat2_out); \ + \ + flat2_out = (tmp_flat5 < (v16u8)flat2_out); \ + flat2_out = __msa_xori_b(flat2_out, 0xff); \ + flat2_out = flat2_out & flat_in; \ + } + +#define AOM_FILTER8(p3_in, p2_in, p1_in, p0_in, q0_in, q1_in, q2_in, q3_in, \ + p2_filt8_out, p1_filt8_out, p0_filt8_out, q0_filt8_out, \ + q1_filt8_out, q2_filt8_out) \ + { \ + v8u16 tmp_filt8_0, tmp_filt8_1, tmp_filt8_2; \ + \ + tmp_filt8_2 = p2_in + p1_in + p0_in; \ + tmp_filt8_0 = p3_in << 1; \ + \ + tmp_filt8_0 = tmp_filt8_0 + tmp_filt8_2 + q0_in; \ + tmp_filt8_1 = tmp_filt8_0 + p3_in + p2_in; \ + p2_filt8_out = (v8i16)__msa_srari_h((v8i16)tmp_filt8_1, 3); \ + \ + tmp_filt8_1 = tmp_filt8_0 + p1_in + q1_in; \ + p1_filt8_out = (v8i16)__msa_srari_h((v8i16)tmp_filt8_1, 3); \ + \ + tmp_filt8_1 = q2_in + q1_in + q0_in; \ + tmp_filt8_2 = tmp_filt8_2 + tmp_filt8_1; \ + tmp_filt8_0 = tmp_filt8_2 + (p0_in); \ + tmp_filt8_0 = tmp_filt8_0 + (p3_in); \ + p0_filt8_out = (v8i16)__msa_srari_h((v8i16)tmp_filt8_0, 3); \ + \ + tmp_filt8_0 = q2_in + q3_in; \ + tmp_filt8_0 = p0_in + tmp_filt8_1 + tmp_filt8_0; \ + tmp_filt8_1 = q3_in + q3_in; \ + tmp_filt8_1 = tmp_filt8_1 + tmp_filt8_0; \ + q2_filt8_out = (v8i16)__msa_srari_h((v8i16)tmp_filt8_1, 3); \ + \ + tmp_filt8_0 = tmp_filt8_2 + q3_in; \ + tmp_filt8_1 = tmp_filt8_0 + q0_in; \ + q0_filt8_out = (v8i16)__msa_srari_h((v8i16)tmp_filt8_1, 3); \ + \ + tmp_filt8_1 = tmp_filt8_0 - p2_in; \ + tmp_filt8_0 = q1_in + q3_in; \ + tmp_filt8_1 = tmp_filt8_0 + tmp_filt8_1; \ + q1_filt8_out = (v8i16)__msa_srari_h((v8i16)tmp_filt8_1, 3); \ + } + +#define LPF_MASK_HEV(p3_in, p2_in, p1_in, p0_in, q0_in, q1_in, q2_in, q3_in, \ + limit_in, b_limit_in, thresh_in, hev_out, mask_out, \ + flat_out) \ + { \ + v16u8 p3_asub_p2_m, p2_asub_p1_m, p1_asub_p0_m, q1_asub_q0_m; \ + v16u8 p1_asub_q1_m, p0_asub_q0_m, q3_asub_q2_m, q2_asub_q1_m; \ + \ + /* absolute subtraction of pixel values */ \ + p3_asub_p2_m = __msa_asub_u_b(p3_in, p2_in); \ + p2_asub_p1_m = __msa_asub_u_b(p2_in, p1_in); \ + p1_asub_p0_m = __msa_asub_u_b(p1_in, p0_in); \ + q1_asub_q0_m = __msa_asub_u_b(q1_in, q0_in); \ + q2_asub_q1_m = __msa_asub_u_b(q2_in, q1_in); \ + q3_asub_q2_m = __msa_asub_u_b(q3_in, q2_in); \ + p0_asub_q0_m = __msa_asub_u_b(p0_in, q0_in); \ + p1_asub_q1_m = __msa_asub_u_b(p1_in, q1_in); \ + \ + /* calculation of hev */ \ + flat_out = __msa_max_u_b(p1_asub_p0_m, q1_asub_q0_m); \ + hev_out = thresh_in < (v16u8)flat_out; \ + \ + /* calculation of mask */ \ + p0_asub_q0_m = __msa_adds_u_b(p0_asub_q0_m, p0_asub_q0_m); \ + p1_asub_q1_m >>= 1; \ + p0_asub_q0_m = __msa_adds_u_b(p0_asub_q0_m, p1_asub_q1_m); \ + \ + mask_out = b_limit_in < p0_asub_q0_m; \ + mask_out = __msa_max_u_b(flat_out, mask_out); \ + p3_asub_p2_m = __msa_max_u_b(p3_asub_p2_m, p2_asub_p1_m); \ + mask_out = __msa_max_u_b(p3_asub_p2_m, mask_out); \ + q2_asub_q1_m = __msa_max_u_b(q2_asub_q1_m, q3_asub_q2_m); \ + mask_out = __msa_max_u_b(q2_asub_q1_m, mask_out); \ + \ + mask_out = limit_in < (v16u8)mask_out; \ + mask_out = __msa_xori_b(mask_out, 0xff); \ + } +#endif // AOM_AOM_DSP_MIPS_LOOPFILTER_MSA_H_ diff --git a/third_party/aom/aom_dsp/mips/macros_msa.h b/third_party/aom/aom_dsp/mips/macros_msa.h new file mode 100644 index 0000000000..9bfc27147b --- /dev/null +++ b/third_party/aom/aom_dsp/mips/macros_msa.h @@ -0,0 +1,2058 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_MIPS_MACROS_MSA_H_ +#define AOM_AOM_DSP_MIPS_MACROS_MSA_H_ + +#include + +#include "config/aom_config.h" + +#include "aom/aom_integer.h" + +#define LD_B(RTYPE, psrc) *((const RTYPE *)(psrc)) +#define LD_UB(...) LD_B(v16u8, __VA_ARGS__) +#define LD_SB(...) LD_B(v16i8, __VA_ARGS__) + +#define LD_H(RTYPE, psrc) *((const RTYPE *)(psrc)) +#define LD_UH(...) LD_H(v8u16, __VA_ARGS__) +#define LD_SH(...) LD_H(v8i16, __VA_ARGS__) + +#define LD_W(RTYPE, psrc) *((const RTYPE *)(psrc)) +#define LD_SW(...) LD_W(v4i32, __VA_ARGS__) + +#define ST_B(RTYPE, in, pdst) *((RTYPE *)(pdst)) = (in) +#define ST_UB(...) ST_B(v16u8, __VA_ARGS__) +#define ST_SB(...) ST_B(v16i8, __VA_ARGS__) + +#define ST_H(RTYPE, in, pdst) *((RTYPE *)(pdst)) = (in) +#define ST_SH(...) ST_H(v8i16, __VA_ARGS__) + +#define ST_W(RTYPE, in, pdst) *((RTYPE *)(pdst)) = (in) +#define ST_SW(...) ST_W(v4i32, __VA_ARGS__) + +#if (__mips_isa_rev >= 6) +#define LH(psrc) \ + ({ \ + const uint8_t *psrc_m = (const uint8_t *)(psrc); \ + uint16_t val_m; \ + \ + __asm__ __volatile__("lh %[val_m], %[psrc_m] \n\t" \ + \ + : [val_m] "=r"(val_m) \ + : [psrc_m] "m"(*psrc_m)); \ + \ + val_m; \ + }) + +#define LW(psrc) \ + ({ \ + const uint8_t *psrc_m = (const uint8_t *)(psrc); \ + uint32_t val_m; \ + \ + __asm__ __volatile__("lw %[val_m], %[psrc_m] \n\t" \ + \ + : [val_m] "=r"(val_m) \ + : [psrc_m] "m"(*psrc_m)); \ + \ + val_m; \ + }) + +#if (__mips == 64) +#define LD(psrc) \ + ({ \ + const uint8_t *psrc_m = (const uint8_t *)(psrc); \ + uint64_t val_m = 0; \ + \ + __asm__ __volatile__("ld %[val_m], %[psrc_m] \n\t" \ + \ + : [val_m] "=r"(val_m) \ + : [psrc_m] "m"(*psrc_m)); \ + \ + val_m; \ + }) +#else // !(__mips == 64) +#define LD(psrc) \ + ({ \ + const uint8_t *psrc_m = (const uint8_t *)(psrc); \ + uint32_t val0_m, val1_m; \ + uint64_t val_m = 0; \ + \ + val0_m = LW(psrc_m); \ + val1_m = LW(psrc_m + 4); \ + \ + val_m = (uint64_t)(val1_m); \ + val_m = (uint64_t)((val_m << 32) & 0xFFFFFFFF00000000); \ + val_m = (uint64_t)(val_m | (uint64_t)val0_m); \ + \ + val_m; \ + }) +#endif // (__mips == 64) + +#define SH(val, pdst) \ + { \ + uint8_t *pdst_m = (uint8_t *)(pdst); \ + const uint16_t val_m = (val); \ + \ + __asm__ __volatile__("sh %[val_m], %[pdst_m] \n\t" \ + \ + : [pdst_m] "=m"(*pdst_m) \ + : [val_m] "r"(val_m)); \ + } + +#define SW(val, pdst) \ + { \ + uint8_t *pdst_m = (uint8_t *)(pdst); \ + const uint32_t val_m = (val); \ + \ + __asm__ __volatile__("sw %[val_m], %[pdst_m] \n\t" \ + \ + : [pdst_m] "=m"(*pdst_m) \ + : [val_m] "r"(val_m)); \ + } + +#define SD(val, pdst) \ + { \ + uint8_t *pdst_m = (uint8_t *)(pdst); \ + const uint64_t val_m = (val); \ + \ + __asm__ __volatile__("sd %[val_m], %[pdst_m] \n\t" \ + \ + : [pdst_m] "=m"(*pdst_m) \ + : [val_m] "r"(val_m)); \ + } +#else // !(__mips_isa_rev >= 6) +#define LH(psrc) \ + ({ \ + const uint8_t *psrc_m = (const uint8_t *)(psrc); \ + uint16_t val_m; \ + \ + __asm__ __volatile__("ulh %[val_m], %[psrc_m] \n\t" \ + \ + : [val_m] "=r"(val_m) \ + : [psrc_m] "m"(*psrc_m)); \ + \ + val_m; \ + }) + +#define LW(psrc) \ + ({ \ + const uint8_t *psrc_m = (const uint8_t *)(psrc); \ + uint32_t val_m; \ + \ + __asm__ __volatile__("ulw %[val_m], %[psrc_m] \n\t" \ + \ + : [val_m] "=r"(val_m) \ + : [psrc_m] "m"(*psrc_m)); \ + \ + val_m; \ + }) + +#if (__mips == 64) +#define LD(psrc) \ + ({ \ + const uint8_t *psrc_m = (const uint8_t *)(psrc); \ + uint64_t val_m = 0; \ + \ + __asm__ __volatile__("uld %[val_m], %[psrc_m] \n\t" \ + \ + : [val_m] "=r"(val_m) \ + : [psrc_m] "m"(*psrc_m)); \ + \ + val_m; \ + }) +#else // !(__mips == 64) +#define LD(psrc) \ + ({ \ + const uint8_t *psrc_m1 = (const uint8_t *)(psrc); \ + uint32_t val0_m, val1_m; \ + uint64_t val_m_combined = 0; \ + \ + val0_m = LW(psrc_m1); \ + val1_m = LW(psrc_m1 + 4); \ + \ + val_m_combined = (uint64_t)(val1_m); \ + val_m_combined = (uint64_t)((val_m_combined << 32) & 0xFFFFFFFF00000000); \ + val_m_combined = (uint64_t)(val_m_combined | (uint64_t)val0_m); \ + \ + val_m_combined; \ + }) +#endif // (__mips == 64) + +#define SH(val, pdst) \ + { \ + uint8_t *pdst_m = (uint8_t *)(pdst); \ + const uint16_t val_m = (val); \ + \ + __asm__ __volatile__("ush %[val_m], %[pdst_m] \n\t" \ + \ + : [pdst_m] "=m"(*pdst_m) \ + : [val_m] "r"(val_m)); \ + } + +#define SW(val, pdst) \ + { \ + uint8_t *pdst_m = (uint8_t *)(pdst); \ + const uint32_t val_m = (val); \ + \ + __asm__ __volatile__("usw %[val_m], %[pdst_m] \n\t" \ + \ + : [pdst_m] "=m"(*pdst_m) \ + : [val_m] "r"(val_m)); \ + } + +#define SD(val, pdst) \ + { \ + uint8_t *pdst_m1 = (uint8_t *)(pdst); \ + uint32_t val0_m, val1_m; \ + \ + val0_m = (uint32_t)((val)&0x00000000FFFFFFFF); \ + val1_m = (uint32_t)(((val) >> 32) & 0x00000000FFFFFFFF); \ + \ + SW(val0_m, pdst_m1); \ + SW(val1_m, pdst_m1 + 4); \ + } +#endif // (__mips_isa_rev >= 6) + +/* Description : Load 4 words with stride + Arguments : Inputs - psrc, stride + Outputs - out0, out1, out2, out3 + Details : Load word in 'out0' from (psrc) + Load word in 'out1' from (psrc + stride) + Load word in 'out2' from (psrc + 2 * stride) + Load word in 'out3' from (psrc + 3 * stride) +*/ +#define LW4(psrc, stride, out0, out1, out2, out3) \ + { \ + out0 = LW((psrc)); \ + out1 = LW((psrc) + stride); \ + out2 = LW((psrc) + 2 * stride); \ + out3 = LW((psrc) + 3 * stride); \ + } + +/* Description : Load double words with stride + Arguments : Inputs - psrc, stride + Outputs - out0, out1 + Details : Load double word in 'out0' from (psrc) + Load double word in 'out1' from (psrc + stride) +*/ +#define LD2(psrc, stride, out0, out1) \ + { \ + out0 = LD((psrc)); \ + out1 = LD((psrc) + stride); \ + } +#define LD4(psrc, stride, out0, out1, out2, out3) \ + { \ + LD2((psrc), stride, out0, out1); \ + LD2((psrc) + 2 * stride, stride, out2, out3); \ + } + +/* Description : Store 4 words with stride + Arguments : Inputs - in0, in1, in2, in3, pdst, stride + Details : Store word from 'in0' to (pdst) + Store word from 'in1' to (pdst + stride) + Store word from 'in2' to (pdst + 2 * stride) + Store word from 'in3' to (pdst + 3 * stride) +*/ +#define SW4(in0, in1, in2, in3, pdst, stride) \ + { \ + SW(in0, (pdst)) \ + SW(in1, (pdst) + stride); \ + SW(in2, (pdst) + 2 * stride); \ + SW(in3, (pdst) + 3 * stride); \ + } + +/* Description : Store 4 double words with stride + Arguments : Inputs - in0, in1, in2, in3, pdst, stride + Details : Store double word from 'in0' to (pdst) + Store double word from 'in1' to (pdst + stride) + Store double word from 'in2' to (pdst + 2 * stride) + Store double word from 'in3' to (pdst + 3 * stride) +*/ +#define SD4(in0, in1, in2, in3, pdst, stride) \ + { \ + SD(in0, (pdst)) \ + SD(in1, (pdst) + stride); \ + SD(in2, (pdst) + 2 * stride); \ + SD(in3, (pdst) + 3 * stride); \ + } + +/* Description : Load vectors with 16 byte elements with stride + Arguments : Inputs - psrc, stride + Outputs - out0, out1 + Return Type - as per RTYPE + Details : Load 16 byte elements in 'out0' from (psrc) + Load 16 byte elements in 'out1' from (psrc + stride) +*/ +#define LD_B2(RTYPE, psrc, stride, out0, out1) \ + { \ + out0 = LD_B(RTYPE, (psrc)); \ + out1 = LD_B(RTYPE, (psrc) + stride); \ + } +#define LD_UB2(...) LD_B2(v16u8, __VA_ARGS__) +#define LD_SB2(...) LD_B2(v16i8, __VA_ARGS__) + +#define LD_B3(RTYPE, psrc, stride, out0, out1, out2) \ + { \ + LD_B2(RTYPE, (psrc), stride, out0, out1); \ + out2 = LD_B(RTYPE, (psrc) + 2 * stride); \ + } +#define LD_UB3(...) LD_B3(v16u8, __VA_ARGS__) + +#define LD_B4(RTYPE, psrc, stride, out0, out1, out2, out3) \ + { \ + LD_B2(RTYPE, (psrc), stride, out0, out1); \ + LD_B2(RTYPE, (psrc) + 2 * stride, stride, out2, out3); \ + } +#define LD_UB4(...) LD_B4(v16u8, __VA_ARGS__) +#define LD_SB4(...) LD_B4(v16i8, __VA_ARGS__) + +#define LD_B5(RTYPE, psrc, stride, out0, out1, out2, out3, out4) \ + { \ + LD_B4(RTYPE, (psrc), stride, out0, out1, out2, out3); \ + out4 = LD_B(RTYPE, (psrc) + 4 * stride); \ + } +#define LD_UB5(...) LD_B5(v16u8, __VA_ARGS__) +#define LD_SB5(...) LD_B5(v16i8, __VA_ARGS__) + +#define LD_B7(RTYPE, psrc, stride, out0, out1, out2, out3, out4, out5, out6) \ + { \ + LD_B5(RTYPE, (psrc), stride, out0, out1, out2, out3, out4); \ + LD_B2(RTYPE, (psrc) + 5 * stride, stride, out5, out6); \ + } +#define LD_SB7(...) LD_B7(v16i8, __VA_ARGS__) + +#define LD_B8(RTYPE, psrc, stride, out0, out1, out2, out3, out4, out5, out6, \ + out7) \ + { \ + LD_B4(RTYPE, (psrc), stride, out0, out1, out2, out3); \ + LD_B4(RTYPE, (psrc) + 4 * stride, stride, out4, out5, out6, out7); \ + } +#define LD_UB8(...) LD_B8(v16u8, __VA_ARGS__) +#define LD_SB8(...) LD_B8(v16i8, __VA_ARGS__) + +/* Description : Load vectors with 8 halfword elements with stride + Arguments : Inputs - psrc, stride + Outputs - out0, out1 + Details : Load 8 halfword elements in 'out0' from (psrc) + Load 8 halfword elements in 'out1' from (psrc + stride) +*/ +#define LD_H2(RTYPE, psrc, stride, out0, out1) \ + { \ + out0 = LD_H(RTYPE, (psrc)); \ + out1 = LD_H(RTYPE, (psrc) + (stride)); \ + } +#define LD_SH2(...) LD_H2(v8i16, __VA_ARGS__) + +#define LD_H4(RTYPE, psrc, stride, out0, out1, out2, out3) \ + { \ + LD_H2(RTYPE, (psrc), stride, out0, out1); \ + LD_H2(RTYPE, (psrc) + 2 * stride, stride, out2, out3); \ + } +#define LD_SH4(...) LD_H4(v8i16, __VA_ARGS__) + +#define LD_H8(RTYPE, psrc, stride, out0, out1, out2, out3, out4, out5, out6, \ + out7) \ + { \ + LD_H4(RTYPE, (psrc), stride, out0, out1, out2, out3); \ + LD_H4(RTYPE, (psrc) + 4 * stride, stride, out4, out5, out6, out7); \ + } +#define LD_SH8(...) LD_H8(v8i16, __VA_ARGS__) + +#define LD_H16(RTYPE, psrc, stride, out0, out1, out2, out3, out4, out5, out6, \ + out7, out8, out9, out10, out11, out12, out13, out14, out15) \ + { \ + LD_H8(RTYPE, (psrc), stride, out0, out1, out2, out3, out4, out5, out6, \ + out7); \ + LD_H8(RTYPE, (psrc) + 8 * stride, stride, out8, out9, out10, out11, out12, \ + out13, out14, out15); \ + } +#define LD_SH16(...) LD_H16(v8i16, __VA_ARGS__) + +/* Description : Load 4x4 block of signed halfword elements from 1D source + data into 4 vectors (Each vector with 4 signed halfwords) + Arguments : Input - psrc + Outputs - out0, out1, out2, out3 +*/ +#define LD4x4_SH(psrc, out0, out1, out2, out3) \ + { \ + out0 = LD_SH(psrc); \ + out2 = LD_SH(psrc + 8); \ + out1 = (v8i16)__msa_ilvl_d((v2i64)out0, (v2i64)out0); \ + out3 = (v8i16)__msa_ilvl_d((v2i64)out2, (v2i64)out2); \ + } + +/* Description : Load 2 vectors of signed word elements with stride + Arguments : Inputs - psrc, stride + Outputs - out0, out1 + Return Type - signed word +*/ +#define LD_SW2(psrc, stride, out0, out1) \ + { \ + out0 = LD_SW((psrc)); \ + out1 = LD_SW((psrc) + stride); \ + } + +/* Description : Store vectors of 16 byte elements with stride + Arguments : Inputs - in0, in1, pdst, stride + Details : Store 16 byte elements from 'in0' to (pdst) + Store 16 byte elements from 'in1' to (pdst + stride) +*/ +#define ST_B2(RTYPE, in0, in1, pdst, stride) \ + { \ + ST_B(RTYPE, in0, (pdst)); \ + ST_B(RTYPE, in1, (pdst) + stride); \ + } +#define ST_UB2(...) ST_B2(v16u8, __VA_ARGS__) + +#define ST_B4(RTYPE, in0, in1, in2, in3, pdst, stride) \ + { \ + ST_B2(RTYPE, in0, in1, (pdst), stride); \ + ST_B2(RTYPE, in2, in3, (pdst) + 2 * stride, stride); \ + } +#define ST_UB4(...) ST_B4(v16u8, __VA_ARGS__) + +#define ST_B8(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7, pdst, stride) \ + { \ + ST_B4(RTYPE, in0, in1, in2, in3, pdst, stride); \ + ST_B4(RTYPE, in4, in5, in6, in7, (pdst) + 4 * stride, stride); \ + } +#define ST_UB8(...) ST_B8(v16u8, __VA_ARGS__) + +/* Description : Store vectors of 8 halfword elements with stride + Arguments : Inputs - in0, in1, pdst, stride + Details : Store 8 halfword elements from 'in0' to (pdst) + Store 8 halfword elements from 'in1' to (pdst + stride) +*/ +#define ST_H2(RTYPE, in0, in1, pdst, stride) \ + { \ + ST_H(RTYPE, in0, (pdst)); \ + ST_H(RTYPE, in1, (pdst) + stride); \ + } +#define ST_SH2(...) ST_H2(v8i16, __VA_ARGS__) + +#define ST_H4(RTYPE, in0, in1, in2, in3, pdst, stride) \ + { \ + ST_H2(RTYPE, in0, in1, (pdst), stride); \ + ST_H2(RTYPE, in2, in3, (pdst) + 2 * stride, stride); \ + } +#define ST_SH4(...) ST_H4(v8i16, __VA_ARGS__) + +#define ST_H8(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7, pdst, stride) \ + { \ + ST_H4(RTYPE, in0, in1, in2, in3, (pdst), stride); \ + ST_H4(RTYPE, in4, in5, in6, in7, (pdst) + 4 * stride, stride); \ + } +#define ST_SH8(...) ST_H8(v8i16, __VA_ARGS__) + +/* Description : Store vectors of word elements with stride + Arguments : Inputs - in0, in1, pdst, stride + Details : Store 4 word elements from 'in0' to (pdst) + Store 4 word elements from 'in1' to (pdst + stride) +*/ +#define ST_SW2(in0, in1, pdst, stride) \ + { \ + ST_SW(in0, (pdst)); \ + ST_SW(in1, (pdst) + stride); \ + } + +/* Description : Store 2x4 byte block to destination memory from input vector + Arguments : Inputs - in, stidx, pdst, stride + Details : Index 'stidx' halfword element from 'in' vector is copied to + the GP register and stored to (pdst) + Index 'stidx+1' halfword element from 'in' vector is copied to + the GP register and stored to (pdst + stride) + Index 'stidx+2' halfword element from 'in' vector is copied to + the GP register and stored to (pdst + 2 * stride) + Index 'stidx+3' halfword element from 'in' vector is copied to + the GP register and stored to (pdst + 3 * stride) +*/ +#define ST2x4_UB(in, stidx, pdst, stride) \ + { \ + uint16_t out0_m, out1_m, out2_m, out3_m; \ + uint8_t *pblk_2x4_m = (uint8_t *)(pdst); \ + \ + out0_m = __msa_copy_u_h((v8i16)in, (stidx)); \ + out1_m = __msa_copy_u_h((v8i16)in, (stidx + 1)); \ + out2_m = __msa_copy_u_h((v8i16)in, (stidx + 2)); \ + out3_m = __msa_copy_u_h((v8i16)in, (stidx + 3)); \ + \ + SH(out0_m, pblk_2x4_m); \ + SH(out1_m, pblk_2x4_m + stride); \ + SH(out2_m, pblk_2x4_m + 2 * stride); \ + SH(out3_m, pblk_2x4_m + 3 * stride); \ + } + +/* Description : Store 4x2 byte block to destination memory from input vector + Arguments : Inputs - in, pdst, stride + Details : Index 0 word element from 'in' vector is copied to the GP + register and stored to (pdst) + Index 1 word element from 'in' vector is copied to the GP + register and stored to (pdst + stride) +*/ +#define ST4x2_UB(in, pdst, stride) \ + { \ + uint32_t out0_m, out1_m; \ + uint8_t *pblk_4x2_m = (uint8_t *)(pdst); \ + \ + out0_m = __msa_copy_u_w((v4i32)in, 0); \ + out1_m = __msa_copy_u_w((v4i32)in, 1); \ + \ + SW(out0_m, pblk_4x2_m); \ + SW(out1_m, pblk_4x2_m + stride); \ + } + +/* Description : Store 4x4 byte block to destination memory from input vector + Arguments : Inputs - in0, in1, pdst, stride + Details : 'Idx0' word element from input vector 'in0' is copied to the + GP register and stored to (pdst) + 'Idx1' word element from input vector 'in0' is copied to the + GP register and stored to (pdst + stride) + 'Idx2' word element from input vector 'in0' is copied to the + GP register and stored to (pdst + 2 * stride) + 'Idx3' word element from input vector 'in0' is copied to the + GP register and stored to (pdst + 3 * stride) +*/ +#define ST4x4_UB(in0, in1, idx0, idx1, idx2, idx3, pdst, stride) \ + { \ + uint32_t out0_m, out1_m, out2_m, out3_m; \ + uint8_t *pblk_4x4_m = (uint8_t *)(pdst); \ + \ + out0_m = __msa_copy_u_w((v4i32)in0, idx0); \ + out1_m = __msa_copy_u_w((v4i32)in0, idx1); \ + out2_m = __msa_copy_u_w((v4i32)in1, idx2); \ + out3_m = __msa_copy_u_w((v4i32)in1, idx3); \ + \ + SW4(out0_m, out1_m, out2_m, out3_m, pblk_4x4_m, stride); \ + } +#define ST4x8_UB(in0, in1, pdst, stride) \ + { \ + uint8_t *pblk_4x8 = (uint8_t *)(pdst); \ + \ + ST4x4_UB(in0, in0, 0, 1, 2, 3, pblk_4x8, stride); \ + ST4x4_UB(in1, in1, 0, 1, 2, 3, pblk_4x8 + 4 * stride, stride); \ + } + +/* Description : Store 8x1 byte block to destination memory from input vector + Arguments : Inputs - in, pdst + Details : Index 0 double word element from 'in' vector is copied to the + GP register and stored to (pdst) +*/ +#define ST8x1_UB(in, pdst) \ + { \ + uint64_t out0_m; \ + \ + out0_m = __msa_copy_u_d((v2i64)in, 0); \ + SD(out0_m, pdst); \ + } + +/* Description : Store 8x2 byte block to destination memory from input vector + Arguments : Inputs - in, pdst, stride + Details : Index 0 double word element from 'in' vector is copied to the + GP register and stored to (pdst) + Index 1 double word element from 'in' vector is copied to the + GP register and stored to (pdst + stride) +*/ +#define ST8x2_UB(in, pdst, stride) \ + { \ + uint64_t out0_m, out1_m; \ + uint8_t *pblk_8x2_m = (uint8_t *)(pdst); \ + \ + out0_m = __msa_copy_u_d((v2i64)in, 0); \ + out1_m = __msa_copy_u_d((v2i64)in, 1); \ + \ + SD(out0_m, pblk_8x2_m); \ + SD(out1_m, pblk_8x2_m + stride); \ + } + +/* Description : Store 8x4 byte block to destination memory from input + vectors + Arguments : Inputs - in0, in1, pdst, stride + Details : Index 0 double word element from 'in0' vector is copied to the + GP register and stored to (pdst) + Index 1 double word element from 'in0' vector is copied to the + GP register and stored to (pdst + stride) + Index 0 double word element from 'in1' vector is copied to the + GP register and stored to (pdst + 2 * stride) + Index 1 double word element from 'in1' vector is copied to the + GP register and stored to (pdst + 3 * stride) +*/ +#define ST8x4_UB(in0, in1, pdst, stride) \ + { \ + uint64_t out0_m, out1_m, out2_m, out3_m; \ + uint8_t *pblk_8x4_m = (uint8_t *)(pdst); \ + \ + out0_m = __msa_copy_u_d((v2i64)in0, 0); \ + out1_m = __msa_copy_u_d((v2i64)in0, 1); \ + out2_m = __msa_copy_u_d((v2i64)in1, 0); \ + out3_m = __msa_copy_u_d((v2i64)in1, 1); \ + \ + SD4(out0_m, out1_m, out2_m, out3_m, pblk_8x4_m, stride); \ + } + +/* Description : average with rounding (in0 + in1 + 1) / 2. + Arguments : Inputs - in0, in1, in2, in3, + Outputs - out0, out1 + Return Type - as per RTYPE + Details : Each unsigned byte element from 'in0' vector is added with + each unsigned byte element from 'in1' vector. Then the average + with rounding is calculated and written to 'out0' +*/ +#define AVER_UB2(RTYPE, in0, in1, in2, in3, out0, out1) \ + { \ + out0 = (RTYPE)__msa_aver_u_b((v16u8)in0, (v16u8)in1); \ + out1 = (RTYPE)__msa_aver_u_b((v16u8)in2, (v16u8)in3); \ + } +#define AVER_UB2_UB(...) AVER_UB2(v16u8, __VA_ARGS__) + +#define AVER_UB4(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, \ + out2, out3) \ + { \ + AVER_UB2(RTYPE, in0, in1, in2, in3, out0, out1) \ + AVER_UB2(RTYPE, in4, in5, in6, in7, out2, out3) \ + } +#define AVER_UB4_UB(...) AVER_UB4(v16u8, __VA_ARGS__) + +/* Description : Immediate number of elements to slide with zero + Arguments : Inputs - in0, in1, slide_val + Outputs - out0, out1 + Return Type - as per RTYPE + Details : Byte elements from 'zero_m' vector are slid into 'in0' by + value specified in the 'slide_val' +*/ +#define SLDI_B2_0(RTYPE, in0, in1, out0, out1, slide_val) \ + { \ + v16i8 zero_m = { 0 }; \ + out0 = (RTYPE)__msa_sldi_b((v16i8)zero_m, (v16i8)in0, slide_val); \ + out1 = (RTYPE)__msa_sldi_b((v16i8)zero_m, (v16i8)in1, slide_val); \ + } +#define SLDI_B2_0_SW(...) SLDI_B2_0(v4i32, __VA_ARGS__) + +#define SLDI_B4_0(RTYPE, in0, in1, in2, in3, out0, out1, out2, out3, \ + slide_val) \ + { \ + SLDI_B2_0(RTYPE, in0, in1, out0, out1, slide_val); \ + SLDI_B2_0(RTYPE, in2, in3, out2, out3, slide_val); \ + } +#define SLDI_B4_0_UB(...) SLDI_B4_0(v16u8, __VA_ARGS__) + +/* Description : Immediate number of elements to slide + Arguments : Inputs - in0_0, in0_1, in1_0, in1_1, slide_val + Outputs - out0, out1 + Return Type - as per RTYPE + Details : Byte elements from 'in0_0' vector are slid into 'in1_0' by + value specified in the 'slide_val' +*/ +#define SLDI_B2(RTYPE, in0_0, in0_1, in1_0, in1_1, out0, out1, slide_val) \ + { \ + out0 = (RTYPE)__msa_sldi_b((v16i8)in0_0, (v16i8)in1_0, slide_val); \ + out1 = (RTYPE)__msa_sldi_b((v16i8)in0_1, (v16i8)in1_1, slide_val); \ + } +#define SLDI_B2_UB(...) SLDI_B2(v16u8, __VA_ARGS__) +#define SLDI_B2_SH(...) SLDI_B2(v8i16, __VA_ARGS__) + +#define SLDI_B3(RTYPE, in0_0, in0_1, in0_2, in1_0, in1_1, in1_2, out0, out1, \ + out2, slide_val) \ + { \ + SLDI_B2(RTYPE, in0_0, in0_1, in1_0, in1_1, out0, out1, slide_val) \ + out2 = (RTYPE)__msa_sldi_b((v16i8)in0_2, (v16i8)in1_2, slide_val); \ + } +#define SLDI_B3_SB(...) SLDI_B3(v16i8, __VA_ARGS__) +#define SLDI_B3_UH(...) SLDI_B3(v8u16, __VA_ARGS__) + +/* Description : Shuffle byte vector elements as per mask vector + Arguments : Inputs - in0, in1, in2, in3, mask0, mask1 + Outputs - out0, out1 + Return Type - as per RTYPE + Details : Byte elements from 'in0' & 'in1' are copied selectively to + 'out0' as per control vector 'mask0' +*/ +#define VSHF_B2(RTYPE, in0, in1, in2, in3, mask0, mask1, out0, out1) \ + { \ + out0 = (RTYPE)__msa_vshf_b((v16i8)mask0, (v16i8)in1, (v16i8)in0); \ + out1 = (RTYPE)__msa_vshf_b((v16i8)mask1, (v16i8)in3, (v16i8)in2); \ + } +#define VSHF_B2_UB(...) VSHF_B2(v16u8, __VA_ARGS__) +#define VSHF_B2_SB(...) VSHF_B2(v16i8, __VA_ARGS__) +#define VSHF_B2_UH(...) VSHF_B2(v8u16, __VA_ARGS__) + +#define VSHF_B4(RTYPE, in0, in1, mask0, mask1, mask2, mask3, out0, out1, out2, \ + out3) \ + { \ + VSHF_B2(RTYPE, in0, in1, in0, in1, mask0, mask1, out0, out1); \ + VSHF_B2(RTYPE, in0, in1, in0, in1, mask2, mask3, out2, out3); \ + } +#define VSHF_B4_SB(...) VSHF_B4(v16i8, __VA_ARGS__) +#define VSHF_B4_SH(...) VSHF_B4(v8i16, __VA_ARGS__) + +/* Description : Dot product of byte vector elements + Arguments : Inputs - mult0, mult1, cnst0, cnst1 + Outputs - out0, out1 + Return Type - as per RTYPE + Details : Unsigned byte elements from 'mult0' are multiplied with + unsigned byte elements from 'cnst0' producing a result + twice the size of input i.e. unsigned halfword. + The multiplication result of adjacent odd-even elements + are added together and written to the 'out0' vector +*/ +#define DOTP_UB2(RTYPE, mult0, mult1, cnst0, cnst1, out0, out1) \ + { \ + out0 = (RTYPE)__msa_dotp_u_h((v16u8)mult0, (v16u8)cnst0); \ + out1 = (RTYPE)__msa_dotp_u_h((v16u8)mult1, (v16u8)cnst1); \ + } +#define DOTP_UB2_UH(...) DOTP_UB2(v8u16, __VA_ARGS__) + +#define DOTP_UB4(RTYPE, mult0, mult1, mult2, mult3, cnst0, cnst1, cnst2, \ + cnst3, out0, out1, out2, out3) \ + { \ + DOTP_UB2(RTYPE, mult0, mult1, cnst0, cnst1, out0, out1); \ + DOTP_UB2(RTYPE, mult2, mult3, cnst2, cnst3, out2, out3); \ + } +#define DOTP_UB4_UH(...) DOTP_UB4(v8u16, __VA_ARGS__) + +/* Description : Dot product of byte vector elements + Arguments : Inputs - mult0, mult1, cnst0, cnst1 + Outputs - out0, out1 + Return Type - as per RTYPE + Details : Signed byte elements from 'mult0' are multiplied with + signed byte elements from 'cnst0' producing a result + twice the size of input i.e. signed halfword. + The multiplication result of adjacent odd-even elements + are added together and written to the 'out0' vector +*/ +#define DOTP_SB2(RTYPE, mult0, mult1, cnst0, cnst1, out0, out1) \ + { \ + out0 = (RTYPE)__msa_dotp_s_h((v16i8)mult0, (v16i8)cnst0); \ + out1 = (RTYPE)__msa_dotp_s_h((v16i8)mult1, (v16i8)cnst1); \ + } +#define DOTP_SB2_SH(...) DOTP_SB2(v8i16, __VA_ARGS__) + +#define DOTP_SB4(RTYPE, mult0, mult1, mult2, mult3, cnst0, cnst1, cnst2, \ + cnst3, out0, out1, out2, out3) \ + { \ + DOTP_SB2(RTYPE, mult0, mult1, cnst0, cnst1, out0, out1); \ + DOTP_SB2(RTYPE, mult2, mult3, cnst2, cnst3, out2, out3); \ + } +#define DOTP_SB4_SH(...) DOTP_SB4(v8i16, __VA_ARGS__) + +/* Description : Dot product of halfword vector elements + Arguments : Inputs - mult0, mult1, cnst0, cnst1 + Outputs - out0, out1 + Return Type - as per RTYPE + Details : Signed halfword elements from 'mult0' are multiplied with + signed halfword elements from 'cnst0' producing a result + twice the size of input i.e. signed word. + The multiplication result of adjacent odd-even elements + are added together and written to the 'out0' vector +*/ +#define DOTP_SH2(RTYPE, mult0, mult1, cnst0, cnst1, out0, out1) \ + { \ + out0 = (RTYPE)__msa_dotp_s_w((v8i16)mult0, (v8i16)cnst0); \ + out1 = (RTYPE)__msa_dotp_s_w((v8i16)mult1, (v8i16)cnst1); \ + } +#define DOTP_SH2_SW(...) DOTP_SH2(v4i32, __VA_ARGS__) + +#define DOTP_SH4(RTYPE, mult0, mult1, mult2, mult3, cnst0, cnst1, cnst2, \ + cnst3, out0, out1, out2, out3) \ + { \ + DOTP_SH2(RTYPE, mult0, mult1, cnst0, cnst1, out0, out1); \ + DOTP_SH2(RTYPE, mult2, mult3, cnst2, cnst3, out2, out3); \ + } +#define DOTP_SH4_SW(...) DOTP_SH4(v4i32, __VA_ARGS__) + +/* Description : Dot product of word vector elements + Arguments : Inputs - mult0, mult1, cnst0, cnst1 + Outputs - out0, out1 + Return Type - as per RTYPE + Details : Signed word elements from 'mult0' are multiplied with + signed word elements from 'cnst0' producing a result + twice the size of input i.e. signed double word. + The multiplication result of adjacent odd-even elements + are added together and written to the 'out0' vector +*/ +#define DOTP_SW2(RTYPE, mult0, mult1, cnst0, cnst1, out0, out1) \ + { \ + out0 = (RTYPE)__msa_dotp_s_d((v4i32)mult0, (v4i32)cnst0); \ + out1 = (RTYPE)__msa_dotp_s_d((v4i32)mult1, (v4i32)cnst1); \ + } +#define DOTP_SW2_SD(...) DOTP_SW2(v2i64, __VA_ARGS__) + +/* Description : Dot product & addition of byte vector elements + Arguments : Inputs - mult0, mult1, cnst0, cnst1 + Outputs - out0, out1 + Return Type - as per RTYPE + Details : Signed byte elements from 'mult0' are multiplied with + signed byte elements from 'cnst0' producing a result + twice the size of input i.e. signed halfword. + The multiplication result of adjacent odd-even elements + are added to the 'out0' vector +*/ +#define DPADD_SB2(RTYPE, mult0, mult1, cnst0, cnst1, out0, out1) \ + { \ + out0 = (RTYPE)__msa_dpadd_s_h((v8i16)out0, (v16i8)mult0, (v16i8)cnst0); \ + out1 = (RTYPE)__msa_dpadd_s_h((v8i16)out1, (v16i8)mult1, (v16i8)cnst1); \ + } +#define DPADD_SB2_SH(...) DPADD_SB2(v8i16, __VA_ARGS__) + +#define DPADD_SB4(RTYPE, mult0, mult1, mult2, mult3, cnst0, cnst1, cnst2, \ + cnst3, out0, out1, out2, out3) \ + { \ + DPADD_SB2(RTYPE, mult0, mult1, cnst0, cnst1, out0, out1); \ + DPADD_SB2(RTYPE, mult2, mult3, cnst2, cnst3, out2, out3); \ + } +#define DPADD_SB4_SH(...) DPADD_SB4(v8i16, __VA_ARGS__) + +/* Description : Dot product & addition of halfword vector elements + Arguments : Inputs - mult0, mult1, cnst0, cnst1 + Outputs - out0, out1 + Return Type - as per RTYPE + Details : Signed halfword elements from 'mult0' are multiplied with + signed halfword elements from 'cnst0' producing a result + twice the size of input i.e. signed word. + The multiplication result of adjacent odd-even elements + are added to the 'out0' vector +*/ +#define DPADD_SH2(RTYPE, mult0, mult1, cnst0, cnst1, out0, out1) \ + { \ + out0 = (RTYPE)__msa_dpadd_s_w((v4i32)out0, (v8i16)mult0, (v8i16)cnst0); \ + out1 = (RTYPE)__msa_dpadd_s_w((v4i32)out1, (v8i16)mult1, (v8i16)cnst1); \ + } +#define DPADD_SH2_SW(...) DPADD_SH2(v4i32, __VA_ARGS__) + +/* Description : Dot product & addition of double word vector elements + Arguments : Inputs - mult0, mult1 + Outputs - out0, out1 + Return Type - as per RTYPE + Details : Each signed word element from 'mult0' is multiplied with itself + producing an intermediate result twice the size of input + i.e. signed double word + The multiplication result of adjacent odd-even elements + are added to the 'out0' vector +*/ +#define DPADD_SD2(RTYPE, mult0, mult1, out0, out1) \ + { \ + out0 = (RTYPE)__msa_dpadd_s_d((v2i64)out0, (v4i32)mult0, (v4i32)mult0); \ + out1 = (RTYPE)__msa_dpadd_s_d((v2i64)out1, (v4i32)mult1, (v4i32)mult1); \ + } +#define DPADD_SD2_SD(...) DPADD_SD2(v2i64, __VA_ARGS__) + +/* Description : Minimum values between unsigned elements of + either vector are copied to the output vector + Arguments : Inputs - in0, in1, min_vec + Outputs - in place operation + Return Type - as per RTYPE + Details : Minimum of unsigned halfword element values from 'in0' and + 'min_vec' are written to output vector 'in0' +*/ +#define MIN_UH2(RTYPE, in0, in1, min_vec) \ + { \ + in0 = (RTYPE)__msa_min_u_h((v8u16)in0, min_vec); \ + in1 = (RTYPE)__msa_min_u_h((v8u16)in1, min_vec); \ + } +#define MIN_UH2_UH(...) MIN_UH2(v8u16, __VA_ARGS__) + +#define MIN_UH4(RTYPE, in0, in1, in2, in3, min_vec) \ + { \ + MIN_UH2(RTYPE, in0, in1, min_vec); \ + MIN_UH2(RTYPE, in2, in3, min_vec); \ + } +#define MIN_UH4_UH(...) MIN_UH4(v8u16, __VA_ARGS__) + +/* Description : Clips all signed halfword elements of input vector + between 0 & 255 + Arguments : Input - in + Output - out_m + Return Type - signed halfword +*/ +#define CLIP_SH_0_255(in) \ + ({ \ + v8i16 max_m = __msa_ldi_h(255); \ + v8i16 out_m; \ + \ + out_m = __msa_maxi_s_h((v8i16)in, 0); \ + out_m = __msa_min_s_h((v8i16)max_m, (v8i16)out_m); \ + out_m; \ + }) +#define CLIP_SH2_0_255(in0, in1) \ + { \ + in0 = CLIP_SH_0_255(in0); \ + in1 = CLIP_SH_0_255(in1); \ + } +#define CLIP_SH4_0_255(in0, in1, in2, in3) \ + { \ + CLIP_SH2_0_255(in0, in1); \ + CLIP_SH2_0_255(in2, in3); \ + } + +/* Description : Horizontal addition of 4 signed word elements of input vector + Arguments : Input - in (signed word vector) + Output - sum_m (i32 sum) + Return Type - signed word (GP) + Details : 4 signed word elements of 'in' vector are added together and + the resulting integer sum is returned +*/ +#define HADD_SW_S32(in) \ + ({ \ + v2i64 res0_m, res1_m; \ + int32_t sum_m; \ + \ + res0_m = __msa_hadd_s_d((v4i32)in, (v4i32)in); \ + res1_m = __msa_splati_d(res0_m, 1); \ + res0_m = res0_m + res1_m; \ + sum_m = __msa_copy_s_w((v4i32)res0_m, 0); \ + sum_m; \ + }) + +/* Description : Horizontal addition of 8 unsigned halfword elements + Arguments : Inputs - in (unsigned halfword vector) + Outputs - sum_m (u32 sum) + Return Type - unsigned word + Details : 8 unsigned halfword elements of input vector are added + together and the resulting integer sum is returned +*/ +#define HADD_UH_U32(in) \ + ({ \ + v4u32 res_m; \ + v2u64 res0_m, res1_m; \ + uint32_t sum_m; \ + \ + res_m = __msa_hadd_u_w((v8u16)in, (v8u16)in); \ + res0_m = __msa_hadd_u_d(res_m, res_m); \ + res1_m = (v2u64)__msa_splati_d((v2i64)res0_m, 1); \ + res0_m = res0_m + res1_m; \ + sum_m = __msa_copy_u_w((v4i32)res0_m, 0); \ + sum_m; \ + }) + +/* Description : Horizontal addition of unsigned byte vector elements + Arguments : Inputs - in0, in1 + Outputs - out0, out1 + Return Type - as per RTYPE + Details : Each unsigned odd byte element from 'in0' is added to + even unsigned byte element from 'in0' (pairwise) and the + halfword result is written to 'out0' +*/ +#define HADD_UB2(RTYPE, in0, in1, out0, out1) \ + { \ + out0 = (RTYPE)__msa_hadd_u_h((v16u8)in0, (v16u8)in0); \ + out1 = (RTYPE)__msa_hadd_u_h((v16u8)in1, (v16u8)in1); \ + } +#define HADD_UB2_UH(...) HADD_UB2(v8u16, __VA_ARGS__) + +#define HADD_UB4(RTYPE, in0, in1, in2, in3, out0, out1, out2, out3) \ + { \ + HADD_UB2(RTYPE, in0, in1, out0, out1); \ + HADD_UB2(RTYPE, in2, in3, out2, out3); \ + } +#define HADD_UB4_UH(...) HADD_UB4(v8u16, __VA_ARGS__) + +/* Description : Horizontal subtraction of unsigned byte vector elements + Arguments : Inputs - in0, in1 + Outputs - out0, out1 + Return Type - as per RTYPE + Details : Each unsigned odd byte element from 'in0' is subtracted from + even unsigned byte element from 'in0' (pairwise) and the + halfword result is written to 'out0' +*/ +#define HSUB_UB2(RTYPE, in0, in1, out0, out1) \ + { \ + out0 = (RTYPE)__msa_hsub_u_h((v16u8)in0, (v16u8)in0); \ + out1 = (RTYPE)__msa_hsub_u_h((v16u8)in1, (v16u8)in1); \ + } +#define HSUB_UB2_SH(...) HSUB_UB2(v8i16, __VA_ARGS__) + +/* Description : SAD (Sum of Absolute Difference) + Arguments : Inputs - in0, in1, ref0, ref1 + Outputs - sad_m (halfword vector) + Return Type - unsigned halfword + Details : Absolute difference of all the byte elements from 'in0' with + 'ref0' is calculated and preserved in 'diff0'. Then even-odd + pairs are added together to generate 8 halfword results. +*/ +#define SAD_UB2_UH(in0, in1, ref0, ref1) \ + ({ \ + v16u8 diff0_m, diff1_m; \ + v8u16 sad_m = { 0 }; \ + \ + diff0_m = __msa_asub_u_b((v16u8)in0, (v16u8)ref0); \ + diff1_m = __msa_asub_u_b((v16u8)in1, (v16u8)ref1); \ + \ + sad_m += __msa_hadd_u_h((v16u8)diff0_m, (v16u8)diff0_m); \ + sad_m += __msa_hadd_u_h((v16u8)diff1_m, (v16u8)diff1_m); \ + \ + sad_m; \ + }) + +/* Description : Horizontal subtraction of signed halfword vector elements + Arguments : Inputs - in0, in1 + Outputs - out0, out1 + Return Type - as per RTYPE + Details : Each signed odd halfword element from 'in0' is subtracted from + even signed halfword element from 'in0' (pairwise) and the + word result is written to 'out0' +*/ +#define HSUB_UH2(RTYPE, in0, in1, out0, out1) \ + { \ + out0 = (RTYPE)__msa_hsub_s_w((v8i16)in0, (v8i16)in0); \ + out1 = (RTYPE)__msa_hsub_s_w((v8i16)in1, (v8i16)in1); \ + } +#define HSUB_UH2_SW(...) HSUB_UH2(v4i32, __VA_ARGS__) + +/* Description : Set element n input vector to GPR value + Arguments : Inputs - in0, in1, in2, in3 + Output - out + Return Type - as per RTYPE + Details : Set element 0 in vector 'out' to value specified in 'in0' +*/ +#define INSERT_W2(RTYPE, in0, in1, out) \ + { \ + out = (RTYPE)__msa_insert_w((v4i32)out, 0, in0); \ + out = (RTYPE)__msa_insert_w((v4i32)out, 1, in1); \ + } +#define INSERT_W2_SB(...) INSERT_W2(v16i8, __VA_ARGS__) + +#define INSERT_W4(RTYPE, in0, in1, in2, in3, out) \ + { \ + out = (RTYPE)__msa_insert_w((v4i32)out, 0, in0); \ + out = (RTYPE)__msa_insert_w((v4i32)out, 1, in1); \ + out = (RTYPE)__msa_insert_w((v4i32)out, 2, in2); \ + out = (RTYPE)__msa_insert_w((v4i32)out, 3, in3); \ + } +#define INSERT_W4_UB(...) INSERT_W4(v16u8, __VA_ARGS__) +#define INSERT_W4_SB(...) INSERT_W4(v16i8, __VA_ARGS__) + +#define INSERT_D2(RTYPE, in0, in1, out) \ + { \ + out = (RTYPE)__msa_insert_d((v2i64)out, 0, in0); \ + out = (RTYPE)__msa_insert_d((v2i64)out, 1, in1); \ + } +#define INSERT_D2_UB(...) INSERT_D2(v16u8, __VA_ARGS__) +#define INSERT_D2_SB(...) INSERT_D2(v16i8, __VA_ARGS__) + +/* Description : Interleave even byte elements from vectors + Arguments : Inputs - in0, in1, in2, in3 + Outputs - out0, out1 + Return Type - as per RTYPE + Details : Even byte elements of 'in0' and 'in1' are interleaved + and written to 'out0' +*/ +#define ILVEV_B2(RTYPE, in0, in1, in2, in3, out0, out1) \ + { \ + out0 = (RTYPE)__msa_ilvev_b((v16i8)in1, (v16i8)in0); \ + out1 = (RTYPE)__msa_ilvev_b((v16i8)in3, (v16i8)in2); \ + } +#define ILVEV_B2_UB(...) ILVEV_B2(v16u8, __VA_ARGS__) +#define ILVEV_B2_SH(...) ILVEV_B2(v8i16, __VA_ARGS__) + +/* Description : Interleave even halfword elements from vectors + Arguments : Inputs - in0, in1, in2, in3 + Outputs - out0, out1 + Return Type - as per RTYPE + Details : Even halfword elements of 'in0' and 'in1' are interleaved + and written to 'out0' +*/ +#define ILVEV_H2(RTYPE, in0, in1, in2, in3, out0, out1) \ + { \ + out0 = (RTYPE)__msa_ilvev_h((v8i16)in1, (v8i16)in0); \ + out1 = (RTYPE)__msa_ilvev_h((v8i16)in3, (v8i16)in2); \ + } +#define ILVEV_H2_UB(...) ILVEV_H2(v16u8, __VA_ARGS__) +#define ILVEV_H2_SH(...) ILVEV_H2(v8i16, __VA_ARGS__) +#define ILVEV_H2_SW(...) ILVEV_H2(v4i32, __VA_ARGS__) + +/* Description : Interleave even word elements from vectors + Arguments : Inputs - in0, in1, in2, in3 + Outputs - out0, out1 + Return Type - as per RTYPE + Details : Even word elements of 'in0' and 'in1' are interleaved + and written to 'out0' +*/ +#define ILVEV_W2(RTYPE, in0, in1, in2, in3, out0, out1) \ + { \ + out0 = (RTYPE)__msa_ilvev_w((v4i32)in1, (v4i32)in0); \ + out1 = (RTYPE)__msa_ilvev_w((v4i32)in3, (v4i32)in2); \ + } +#define ILVEV_W2_SB(...) ILVEV_W2(v16i8, __VA_ARGS__) + +/* Description : Interleave even double word elements from vectors + Arguments : Inputs - in0, in1, in2, in3 + Outputs - out0, out1 + Return Type - as per RTYPE + Details : Even double word elements of 'in0' and 'in1' are interleaved + and written to 'out0' +*/ +#define ILVEV_D2(RTYPE, in0, in1, in2, in3, out0, out1) \ + { \ + out0 = (RTYPE)__msa_ilvev_d((v2i64)in1, (v2i64)in0); \ + out1 = (RTYPE)__msa_ilvev_d((v2i64)in3, (v2i64)in2); \ + } +#define ILVEV_D2_UB(...) ILVEV_D2(v16u8, __VA_ARGS__) + +/* Description : Interleave left half of byte elements from vectors + Arguments : Inputs - in0, in1, in2, in3 + Outputs - out0, out1 + Return Type - as per RTYPE + Details : Left half of byte elements of 'in0' and 'in1' are interleaved + and written to 'out0'. +*/ +#define ILVL_B2(RTYPE, in0, in1, in2, in3, out0, out1) \ + { \ + out0 = (RTYPE)__msa_ilvl_b((v16i8)in0, (v16i8)in1); \ + out1 = (RTYPE)__msa_ilvl_b((v16i8)in2, (v16i8)in3); \ + } +#define ILVL_B2_UB(...) ILVL_B2(v16u8, __VA_ARGS__) +#define ILVL_B2_SB(...) ILVL_B2(v16i8, __VA_ARGS__) +#define ILVL_B2_UH(...) ILVL_B2(v8u16, __VA_ARGS__) +#define ILVL_B2_SH(...) ILVL_B2(v8i16, __VA_ARGS__) + +#define ILVL_B4(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, \ + out2, out3) \ + { \ + ILVL_B2(RTYPE, in0, in1, in2, in3, out0, out1); \ + ILVL_B2(RTYPE, in4, in5, in6, in7, out2, out3); \ + } +#define ILVL_B4_SB(...) ILVL_B4(v16i8, __VA_ARGS__) +#define ILVL_B4_SH(...) ILVL_B4(v8i16, __VA_ARGS__) +#define ILVL_B4_UH(...) ILVL_B4(v8u16, __VA_ARGS__) + +/* Description : Interleave left half of halfword elements from vectors + Arguments : Inputs - in0, in1, in2, in3 + Outputs - out0, out1 + Return Type - as per RTYPE + Details : Left half of halfword elements of 'in0' and 'in1' are + interleaved and written to 'out0'. +*/ +#define ILVL_H2(RTYPE, in0, in1, in2, in3, out0, out1) \ + { \ + out0 = (RTYPE)__msa_ilvl_h((v8i16)in0, (v8i16)in1); \ + out1 = (RTYPE)__msa_ilvl_h((v8i16)in2, (v8i16)in3); \ + } +#define ILVL_H2_SH(...) ILVL_H2(v8i16, __VA_ARGS__) +#define ILVL_H2_SW(...) ILVL_H2(v4i32, __VA_ARGS__) + +/* Description : Interleave left half of word elements from vectors + Arguments : Inputs - in0, in1, in2, in3 + Outputs - out0, out1 + Return Type - as per RTYPE + Details : Left half of word elements of 'in0' and 'in1' are interleaved + and written to 'out0'. +*/ +#define ILVL_W2(RTYPE, in0, in1, in2, in3, out0, out1) \ + { \ + out0 = (RTYPE)__msa_ilvl_w((v4i32)in0, (v4i32)in1); \ + out1 = (RTYPE)__msa_ilvl_w((v4i32)in2, (v4i32)in3); \ + } +#define ILVL_W2_UB(...) ILVL_W2(v16u8, __VA_ARGS__) +#define ILVL_W2_SH(...) ILVL_W2(v8i16, __VA_ARGS__) + +/* Description : Interleave right half of byte elements from vectors + Arguments : Inputs - in0, in1, in2, in3 + Outputs - out0, out1 + Return Type - as per RTYPE + Details : Right half of byte elements of 'in0' and 'in1' are interleaved + and written to out0. +*/ +#define ILVR_B2(RTYPE, in0, in1, in2, in3, out0, out1) \ + { \ + out0 = (RTYPE)__msa_ilvr_b((v16i8)in0, (v16i8)in1); \ + out1 = (RTYPE)__msa_ilvr_b((v16i8)in2, (v16i8)in3); \ + } +#define ILVR_B2_UB(...) ILVR_B2(v16u8, __VA_ARGS__) +#define ILVR_B2_SB(...) ILVR_B2(v16i8, __VA_ARGS__) +#define ILVR_B2_UH(...) ILVR_B2(v8u16, __VA_ARGS__) +#define ILVR_B2_SH(...) ILVR_B2(v8i16, __VA_ARGS__) + +#define ILVR_B4(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, \ + out2, out3) \ + { \ + ILVR_B2(RTYPE, in0, in1, in2, in3, out0, out1); \ + ILVR_B2(RTYPE, in4, in5, in6, in7, out2, out3); \ + } +#define ILVR_B4_UB(...) ILVR_B4(v16u8, __VA_ARGS__) +#define ILVR_B4_SB(...) ILVR_B4(v16i8, __VA_ARGS__) +#define ILVR_B4_UH(...) ILVR_B4(v8u16, __VA_ARGS__) +#define ILVR_B4_SH(...) ILVR_B4(v8i16, __VA_ARGS__) + +#define ILVR_B8(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7, in8, in9, in10, \ + in11, in12, in13, in14, in15, out0, out1, out2, out3, out4, \ + out5, out6, out7) \ + { \ + ILVR_B4(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, out2, \ + out3); \ + ILVR_B4(RTYPE, in8, in9, in10, in11, in12, in13, in14, in15, out4, out5, \ + out6, out7); \ + } +#define ILVR_B8_UH(...) ILVR_B8(v8u16, __VA_ARGS__) + +/* Description : Interleave right half of halfword elements from vectors + Arguments : Inputs - in0, in1, in2, in3 + Outputs - out0, out1 + Return Type - as per RTYPE + Details : Right half of halfword elements of 'in0' and 'in1' are + interleaved and written to 'out0'. +*/ +#define ILVR_H2(RTYPE, in0, in1, in2, in3, out0, out1) \ + { \ + out0 = (RTYPE)__msa_ilvr_h((v8i16)in0, (v8i16)in1); \ + out1 = (RTYPE)__msa_ilvr_h((v8i16)in2, (v8i16)in3); \ + } +#define ILVR_H2_SH(...) ILVR_H2(v8i16, __VA_ARGS__) +#define ILVR_H2_SW(...) ILVR_H2(v4i32, __VA_ARGS__) + +#define ILVR_H4(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, \ + out2, out3) \ + { \ + ILVR_H2(RTYPE, in0, in1, in2, in3, out0, out1); \ + ILVR_H2(RTYPE, in4, in5, in6, in7, out2, out3); \ + } +#define ILVR_H4_SH(...) ILVR_H4(v8i16, __VA_ARGS__) + +#define ILVR_W2(RTYPE, in0, in1, in2, in3, out0, out1) \ + { \ + out0 = (RTYPE)__msa_ilvr_w((v4i32)in0, (v4i32)in1); \ + out1 = (RTYPE)__msa_ilvr_w((v4i32)in2, (v4i32)in3); \ + } +#define ILVR_W2_UB(...) ILVR_W2(v16u8, __VA_ARGS__) +#define ILVR_W2_SH(...) ILVR_W2(v8i16, __VA_ARGS__) + +#define ILVR_W4(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, \ + out2, out3) \ + { \ + ILVR_W2(RTYPE, in0, in1, in2, in3, out0, out1); \ + ILVR_W2(RTYPE, in4, in5, in6, in7, out2, out3); \ + } +#define ILVR_W4_UB(...) ILVR_W4(v16u8, __VA_ARGS__) + +/* Description : Interleave right half of double word elements from vectors + Arguments : Inputs - in0, in1, in2, in3 + Outputs - out0, out1 + Return Type - as per RTYPE + Details : Right half of double word elements of 'in0' and 'in1' are + interleaved and written to 'out0'. +*/ +#define ILVR_D2(RTYPE, in0, in1, in2, in3, out0, out1) \ + { \ + out0 = (RTYPE)__msa_ilvr_d((v2i64)(in0), (v2i64)(in1)); \ + out1 = (RTYPE)__msa_ilvr_d((v2i64)(in2), (v2i64)(in3)); \ + } +#define ILVR_D2_UB(...) ILVR_D2(v16u8, __VA_ARGS__) +#define ILVR_D2_SB(...) ILVR_D2(v16i8, __VA_ARGS__) +#define ILVR_D2_SH(...) ILVR_D2(v8i16, __VA_ARGS__) + +#define ILVR_D3(RTYPE, in0, in1, in2, in3, in4, in5, out0, out1, out2) \ + { \ + ILVR_D2(RTYPE, in0, in1, in2, in3, out0, out1); \ + out2 = (RTYPE)__msa_ilvr_d((v2i64)(in4), (v2i64)(in5)); \ + } +#define ILVR_D3_SB(...) ILVR_D3(v16i8, __VA_ARGS__) + +#define ILVR_D4(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, \ + out2, out3) \ + { \ + ILVR_D2(RTYPE, in0, in1, in2, in3, out0, out1); \ + ILVR_D2(RTYPE, in4, in5, in6, in7, out2, out3); \ + } +#define ILVR_D4_SB(...) ILVR_D4(v16i8, __VA_ARGS__) +#define ILVR_D4_UB(...) ILVR_D4(v16u8, __VA_ARGS__) + +/* Description : Interleave both left and right half of input vectors + Arguments : Inputs - in0, in1 + Outputs - out0, out1 + Return Type - as per RTYPE + Details : Right half of byte elements from 'in0' and 'in1' are + interleaved and written to 'out0' +*/ +#define ILVRL_B2(RTYPE, in0, in1, out0, out1) \ + { \ + out0 = (RTYPE)__msa_ilvr_b((v16i8)in0, (v16i8)in1); \ + out1 = (RTYPE)__msa_ilvl_b((v16i8)in0, (v16i8)in1); \ + } +#define ILVRL_B2_UB(...) ILVRL_B2(v16u8, __VA_ARGS__) +#define ILVRL_B2_SB(...) ILVRL_B2(v16i8, __VA_ARGS__) +#define ILVRL_B2_UH(...) ILVRL_B2(v8u16, __VA_ARGS__) +#define ILVRL_B2_SH(...) ILVRL_B2(v8i16, __VA_ARGS__) + +#define ILVRL_H2(RTYPE, in0, in1, out0, out1) \ + { \ + out0 = (RTYPE)__msa_ilvr_h((v8i16)in0, (v8i16)in1); \ + out1 = (RTYPE)__msa_ilvl_h((v8i16)in0, (v8i16)in1); \ + } +#define ILVRL_H2_SH(...) ILVRL_H2(v8i16, __VA_ARGS__) +#define ILVRL_H2_SW(...) ILVRL_H2(v4i32, __VA_ARGS__) + +#define ILVRL_W2(RTYPE, in0, in1, out0, out1) \ + { \ + out0 = (RTYPE)__msa_ilvr_w((v4i32)in0, (v4i32)in1); \ + out1 = (RTYPE)__msa_ilvl_w((v4i32)in0, (v4i32)in1); \ + } +#define ILVRL_W2_UB(...) ILVRL_W2(v16u8, __VA_ARGS__) +#define ILVRL_W2_SH(...) ILVRL_W2(v8i16, __VA_ARGS__) +#define ILVRL_W2_SW(...) ILVRL_W2(v4i32, __VA_ARGS__) + +/* Description : Saturate the halfword element values to the max + unsigned value of (sat_val + 1) bits + The element data width remains unchanged + Arguments : Inputs - in0, in1, sat_val + Outputs - in place operation + Return Type - as per RTYPE + Details : Each unsigned halfword element from 'in0' is saturated to the + value generated with (sat_val + 1) bit range. + The results are written in place +*/ +#define SAT_UH2(RTYPE, in0, in1, sat_val) \ + { \ + in0 = (RTYPE)__msa_sat_u_h((v8u16)in0, sat_val); \ + in1 = (RTYPE)__msa_sat_u_h((v8u16)in1, sat_val); \ + } +#define SAT_UH2_UH(...) SAT_UH2(v8u16, __VA_ARGS__) + +#define SAT_UH4(RTYPE, in0, in1, in2, in3, sat_val) \ + { \ + SAT_UH2(RTYPE, in0, in1, sat_val); \ + SAT_UH2(RTYPE, in2, in3, sat_val) \ + } +#define SAT_UH4_UH(...) SAT_UH4(v8u16, __VA_ARGS__) + +/* Description : Saturate the halfword element values to the max + unsigned value of (sat_val + 1) bits + The element data width remains unchanged + Arguments : Inputs - in0, in1, sat_val + Outputs - in place operation + Return Type - as per RTYPE + Details : Each unsigned halfword element from 'in0' is saturated to the + value generated with (sat_val + 1) bit range + The results are written in place +*/ +#define SAT_SH2(RTYPE, in0, in1, sat_val) \ + { \ + in0 = (RTYPE)__msa_sat_s_h((v8i16)in0, sat_val); \ + in1 = (RTYPE)__msa_sat_s_h((v8i16)in1, sat_val); \ + } +#define SAT_SH2_SH(...) SAT_SH2(v8i16, __VA_ARGS__) + +#define SAT_SH4(RTYPE, in0, in1, in2, in3, sat_val) \ + { \ + SAT_SH2(RTYPE, in0, in1, sat_val); \ + SAT_SH2(RTYPE, in2, in3, sat_val); \ + } +#define SAT_SH4_SH(...) SAT_SH4(v8i16, __VA_ARGS__) + +/* Description : Indexed halfword element values are replicated to all + elements in output vector + Arguments : Inputs - in, idx0, idx1 + Outputs - out0, out1 + Return Type - as per RTYPE + Details : 'idx0' element value from 'in' vector is replicated to all + elements in 'out0' vector + Valid index range for halfword operation is 0-7 +*/ +#define SPLATI_H2(RTYPE, in, idx0, idx1, out0, out1) \ + { \ + out0 = (RTYPE)__msa_splati_h((v8i16)in, idx0); \ + out1 = (RTYPE)__msa_splati_h((v8i16)in, idx1); \ + } +#define SPLATI_H2_SH(...) SPLATI_H2(v8i16, __VA_ARGS__) + +#define SPLATI_H4(RTYPE, in, idx0, idx1, idx2, idx3, out0, out1, out2, out3) \ + { \ + SPLATI_H2(RTYPE, in, idx0, idx1, out0, out1); \ + SPLATI_H2(RTYPE, in, idx2, idx3, out2, out3); \ + } +#define SPLATI_H4_SB(...) SPLATI_H4(v16i8, __VA_ARGS__) +#define SPLATI_H4_SH(...) SPLATI_H4(v8i16, __VA_ARGS__) + +/* Description : Pack even byte elements of vector pairs + Arguments : Inputs - in0, in1, in2, in3 + Outputs - out0, out1 + Return Type - as per RTYPE + Details : Even byte elements of 'in0' are copied to the left half of + 'out0' & even byte elements of 'in1' are copied to the right + half of 'out0'. +*/ +#define PCKEV_B2(RTYPE, in0, in1, in2, in3, out0, out1) \ + { \ + out0 = (RTYPE)__msa_pckev_b((v16i8)in0, (v16i8)in1); \ + out1 = (RTYPE)__msa_pckev_b((v16i8)in2, (v16i8)in3); \ + } +#define PCKEV_B2_SB(...) PCKEV_B2(v16i8, __VA_ARGS__) +#define PCKEV_B2_UB(...) PCKEV_B2(v16u8, __VA_ARGS__) +#define PCKEV_B2_SH(...) PCKEV_B2(v8i16, __VA_ARGS__) + +#define PCKEV_B4(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, \ + out2, out3) \ + { \ + PCKEV_B2(RTYPE, in0, in1, in2, in3, out0, out1); \ + PCKEV_B2(RTYPE, in4, in5, in6, in7, out2, out3); \ + } +#define PCKEV_B4_SB(...) PCKEV_B4(v16i8, __VA_ARGS__) +#define PCKEV_B4_UB(...) PCKEV_B4(v16u8, __VA_ARGS__) +#define PCKEV_B4_SH(...) PCKEV_B4(v8i16, __VA_ARGS__) + +/* Description : Pack even halfword elements of vector pairs + Arguments : Inputs - in0, in1, in2, in3 + Outputs - out0, out1 + Return Type - as per RTYPE + Details : Even halfword elements of 'in0' are copied to the left half of + 'out0' & even halfword elements of 'in1' are copied to the + right half of 'out0'. +*/ +#define PCKEV_H2(RTYPE, in0, in1, in2, in3, out0, out1) \ + { \ + out0 = (RTYPE)__msa_pckev_h((v8i16)in0, (v8i16)in1); \ + out1 = (RTYPE)__msa_pckev_h((v8i16)in2, (v8i16)in3); \ + } +#define PCKEV_H2_SH(...) PCKEV_H2(v8i16, __VA_ARGS__) +#define PCKEV_H2_SW(...) PCKEV_H2(v4i32, __VA_ARGS__) + +#define PCKEV_H4(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, \ + out2, out3) \ + { \ + PCKEV_H2(RTYPE, in0, in1, in2, in3, out0, out1); \ + PCKEV_H2(RTYPE, in4, in5, in6, in7, out2, out3); \ + } +#define PCKEV_H4_SH(...) PCKEV_H4(v8i16, __VA_ARGS__) + +/* Description : Pack even double word elements of vector pairs + Arguments : Inputs - in0, in1, in2, in3 + Outputs - out0, out1 + Return Type - as per RTYPE + Details : Even double elements of 'in0' are copied to the left half of + 'out0' & even double elements of 'in1' are copied to the right + half of 'out0'. +*/ +#define PCKEV_D2(RTYPE, in0, in1, in2, in3, out0, out1) \ + { \ + out0 = (RTYPE)__msa_pckev_d((v2i64)in0, (v2i64)in1); \ + out1 = (RTYPE)__msa_pckev_d((v2i64)in2, (v2i64)in3); \ + } +#define PCKEV_D2_UB(...) PCKEV_D2(v16u8, __VA_ARGS__) +#define PCKEV_D2_SH(...) PCKEV_D2(v8i16, __VA_ARGS__) + +#define PCKEV_D4(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, \ + out2, out3) \ + { \ + PCKEV_D2(RTYPE, in0, in1, in2, in3, out0, out1); \ + PCKEV_D2(RTYPE, in4, in5, in6, in7, out2, out3); \ + } +#define PCKEV_D4_UB(...) PCKEV_D4(v16u8, __VA_ARGS__) + +/* Description : Each byte element is logically xor'ed with immediate 128 + Arguments : Inputs - in0, in1 + Outputs - in place operation + Return Type - as per RTYPE + Details : Each unsigned byte element from input vector 'in0' is + logically xor'ed with 128 and the result is stored in-place. +*/ +#define XORI_B2_128(RTYPE, in0, in1) \ + { \ + in0 = (RTYPE)__msa_xori_b((v16u8)in0, 128); \ + in1 = (RTYPE)__msa_xori_b((v16u8)in1, 128); \ + } +#define XORI_B2_128_UB(...) XORI_B2_128(v16u8, __VA_ARGS__) +#define XORI_B2_128_SB(...) XORI_B2_128(v16i8, __VA_ARGS__) + +#define XORI_B3_128(RTYPE, in0, in1, in2) \ + { \ + XORI_B2_128(RTYPE, in0, in1); \ + in2 = (RTYPE)__msa_xori_b((v16u8)in2, 128); \ + } +#define XORI_B3_128_SB(...) XORI_B3_128(v16i8, __VA_ARGS__) + +#define XORI_B4_128(RTYPE, in0, in1, in2, in3) \ + { \ + XORI_B2_128(RTYPE, in0, in1); \ + XORI_B2_128(RTYPE, in2, in3); \ + } +#define XORI_B4_128_UB(...) XORI_B4_128(v16u8, __VA_ARGS__) +#define XORI_B4_128_SB(...) XORI_B4_128(v16i8, __VA_ARGS__) + +#define XORI_B7_128(RTYPE, in0, in1, in2, in3, in4, in5, in6) \ + { \ + XORI_B4_128(RTYPE, in0, in1, in2, in3); \ + XORI_B3_128(RTYPE, in4, in5, in6); \ + } +#define XORI_B7_128_SB(...) XORI_B7_128(v16i8, __VA_ARGS__) + +/* Description : Average of signed halfword elements -> (a + b) / 2 + Arguments : Inputs - in0, in1, in2, in3, in4, in5, in6, in7 + Outputs - out0, out1, out2, out3 + Return Type - as per RTYPE + Details : Each signed halfword element from 'in0' is added to each + signed halfword element of 'in1' with full precision resulting + in one extra bit in the result. The result is then divided by + 2 and written to 'out0' +*/ +#define AVE_SH4(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, \ + out2, out3) \ + { \ + out0 = (RTYPE)__msa_ave_s_h((v8i16)in0, (v8i16)in1); \ + out1 = (RTYPE)__msa_ave_s_h((v8i16)in2, (v8i16)in3); \ + out2 = (RTYPE)__msa_ave_s_h((v8i16)in4, (v8i16)in5); \ + out3 = (RTYPE)__msa_ave_s_h((v8i16)in6, (v8i16)in7); \ + } +#define AVE_SH4_SH(...) AVE_SH4(v8i16, __VA_ARGS__) + +/* Description : Addition of signed halfword elements and signed saturation + Arguments : Inputs - in0, in1, in2, in3 + Outputs - out0, out1 + Return Type - as per RTYPE + Details : Signed halfword elements from 'in0' are added to signed + halfword elements of 'in1'. The result is then signed saturated + between halfword data type range +*/ +#define ADDS_SH2(RTYPE, in0, in1, in2, in3, out0, out1) \ + { \ + out0 = (RTYPE)__msa_adds_s_h((v8i16)in0, (v8i16)in1); \ + out1 = (RTYPE)__msa_adds_s_h((v8i16)in2, (v8i16)in3); \ + } +#define ADDS_SH2_SH(...) ADDS_SH2(v8i16, __VA_ARGS__) + +#define ADDS_SH4(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, \ + out2, out3) \ + { \ + ADDS_SH2(RTYPE, in0, in1, in2, in3, out0, out1); \ + ADDS_SH2(RTYPE, in4, in5, in6, in7, out2, out3); \ + } +#define ADDS_SH4_SH(...) ADDS_SH4(v8i16, __VA_ARGS__) + +/* Description : Shift left all elements of vector (generic for all data types) + Arguments : Inputs - in0, in1, in2, in3, shift + Outputs - in place operation + Return Type - as per input vector RTYPE + Details : Each element of vector 'in0' is left shifted by 'shift' and + the result is written in-place. +*/ +#define SLLI_4V(in0, in1, in2, in3, shift) \ + { \ + in0 = in0 << shift; \ + in1 = in1 << shift; \ + in2 = in2 << shift; \ + in3 = in3 << shift; \ + } + +/* Description : Arithmetic shift right all elements of vector + (generic for all data types) + Arguments : Inputs - in0, in1, in2, in3, shift + Outputs - in place operation + Return Type - as per input vector RTYPE + Details : Each element of vector 'in0' is right shifted by 'shift' and + the result is written in-place. 'shift' is a GP variable. +*/ +#define SRA_4V(in0, in1, in2, in3, shift) \ + { \ + in0 = in0 >> shift; \ + in1 = in1 >> shift; \ + in2 = in2 >> shift; \ + in3 = in3 >> shift; \ + } + +/* Description : Shift right arithmetic rounded words + Arguments : Inputs - in0, in1, shift + Outputs - in place operation + Return Type - as per RTYPE + Details : Each element of vector 'in0' is shifted right arithmetically by + the number of bits in the corresponding element in the vector + 'shift'. The last discarded bit is added to shifted value for + rounding and the result is written in-place. + 'shift' is a vector. +*/ +#define SRAR_W2(RTYPE, in0, in1, shift) \ + { \ + in0 = (RTYPE)__msa_srar_w((v4i32)in0, (v4i32)shift); \ + in1 = (RTYPE)__msa_srar_w((v4i32)in1, (v4i32)shift); \ + } + +#define SRAR_W4(RTYPE, in0, in1, in2, in3, shift) \ + { \ + SRAR_W2(RTYPE, in0, in1, shift) \ + SRAR_W2(RTYPE, in2, in3, shift) \ + } +#define SRAR_W4_SW(...) SRAR_W4(v4i32, __VA_ARGS__) + +/* Description : Shift right arithmetic rounded (immediate) + Arguments : Inputs - in0, in1, shift + Outputs - in place operation + Return Type - as per RTYPE + Details : Each element of vector 'in0' is shifted right arithmetically by + the value in 'shift'. The last discarded bit is added to the + shifted value for rounding and the result is written in-place. + 'shift' is an immediate value. +*/ +#define SRARI_H2(RTYPE, in0, in1, shift) \ + { \ + in0 = (RTYPE)__msa_srari_h((v8i16)in0, shift); \ + in1 = (RTYPE)__msa_srari_h((v8i16)in1, shift); \ + } +#define SRARI_H2_UH(...) SRARI_H2(v8u16, __VA_ARGS__) +#define SRARI_H2_SH(...) SRARI_H2(v8i16, __VA_ARGS__) + +#define SRARI_H4(RTYPE, in0, in1, in2, in3, shift) \ + { \ + SRARI_H2(RTYPE, in0, in1, shift); \ + SRARI_H2(RTYPE, in2, in3, shift); \ + } +#define SRARI_H4_UH(...) SRARI_H4(v8u16, __VA_ARGS__) +#define SRARI_H4_SH(...) SRARI_H4(v8i16, __VA_ARGS__) + +#define SRARI_W2(RTYPE, in0, in1, shift) \ + { \ + in0 = (RTYPE)__msa_srari_w((v4i32)in0, shift); \ + in1 = (RTYPE)__msa_srari_w((v4i32)in1, shift); \ + } +#define SRARI_W2_SW(...) SRARI_W2(v4i32, __VA_ARGS__) + +#define SRARI_W4(RTYPE, in0, in1, in2, in3, shift) \ + { \ + SRARI_W2(RTYPE, in0, in1, shift); \ + SRARI_W2(RTYPE, in2, in3, shift); \ + } +#define SRARI_W4_SW(...) SRARI_W4(v4i32, __VA_ARGS__) + +/* Description : Logical shift right all elements of vector (immediate) + Arguments : Inputs - in0, in1, in2, in3, shift + Outputs - out0, out1, out2, out3 + Return Type - as per RTYPE + Details : Each element of vector 'in0' is right shifted by 'shift' and + the result is written in-place. 'shift' is an immediate value. +*/ +#define SRLI_H4(RTYPE, in0, in1, in2, in3, out0, out1, out2, out3, shift) \ + { \ + out0 = (RTYPE)__msa_srli_h((v8i16)in0, shift); \ + out1 = (RTYPE)__msa_srli_h((v8i16)in1, shift); \ + out2 = (RTYPE)__msa_srli_h((v8i16)in2, shift); \ + out3 = (RTYPE)__msa_srli_h((v8i16)in3, shift); \ + } +#define SRLI_H4_SH(...) SRLI_H4(v8i16, __VA_ARGS__) + +/* Description : Multiplication of pairs of vectors + Arguments : Inputs - in0, in1, in2, in3 + Outputs - out0, out1 + Details : Each element from 'in0' is multiplied with elements from 'in1' + and the result is written to 'out0' +*/ +#define MUL2(in0, in1, in2, in3, out0, out1) \ + { \ + out0 = in0 * in1; \ + out1 = in2 * in3; \ + } +#define MUL4(in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, out2, out3) \ + { \ + MUL2(in0, in1, in2, in3, out0, out1); \ + MUL2(in4, in5, in6, in7, out2, out3); \ + } + +/* Description : Addition of 2 pairs of vectors + Arguments : Inputs - in0, in1, in2, in3 + Outputs - out0, out1 + Details : Each element in 'in0' is added to 'in1' and result is written + to 'out0'. +*/ +#define ADD2(in0, in1, in2, in3, out0, out1) \ + { \ + out0 = in0 + in1; \ + out1 = in2 + in3; \ + } +#define ADD4(in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, out2, out3) \ + { \ + ADD2(in0, in1, in2, in3, out0, out1); \ + ADD2(in4, in5, in6, in7, out2, out3); \ + } + +/* Description : Subtraction of 2 pairs of vectors + Arguments : Inputs - in0, in1, in2, in3 + Outputs - out0, out1 + Details : Each element in 'in1' is subtracted from 'in0' and result is + written to 'out0'. +*/ +#define SUB2(in0, in1, in2, in3, out0, out1) \ + { \ + out0 = in0 - in1; \ + out1 = in2 - in3; \ + } +#define SUB4(in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, out2, out3) \ + { \ + out0 = in0 - in1; \ + out1 = in2 - in3; \ + out2 = in4 - in5; \ + out3 = in6 - in7; \ + } + +/* Description : Sign extend halfword elements from right half of the vector + Arguments : Input - in (halfword vector) + Output - out (sign extended word vector) + Return Type - signed word + Details : Sign bit of halfword elements from input vector 'in' is + extracted and interleaved with same vector 'in0' to generate + 4 word elements keeping sign intact +*/ +#define UNPCK_R_SH_SW(in, out) \ + { \ + v8i16 sign_m; \ + \ + sign_m = __msa_clti_s_h((v8i16)in, 0); \ + out = (v4i32)__msa_ilvr_h(sign_m, (v8i16)in); \ + } + +/* Description : Zero extend unsigned byte elements to halfword elements + Arguments : Input - in (unsigned byte vector) + Outputs - out0, out1 (unsigned halfword vectors) + Return Type - signed halfword + Details : Zero extended right half of vector is returned in 'out0' + Zero extended left half of vector is returned in 'out1' +*/ +#define UNPCK_UB_SH(in, out0, out1) \ + { \ + v16i8 zero_m = { 0 }; \ + \ + ILVRL_B2_SH(zero_m, in, out0, out1); \ + } + +/* Description : Sign extend halfword elements from input vector and return + the result in pair of vectors + Arguments : Input - in (halfword vector) + Outputs - out0, out1 (sign extended word vectors) + Return Type - signed word + Details : Sign bit of halfword elements from input vector 'in' is + extracted and interleaved right with same vector 'in0' to + generate 4 signed word elements in 'out0' + Then interleaved left with same vector 'in0' to + generate 4 signed word elements in 'out1' +*/ +#define UNPCK_SH_SW(in, out0, out1) \ + { \ + v8i16 tmp_m; \ + \ + tmp_m = __msa_clti_s_h((v8i16)in, 0); \ + ILVRL_H2_SW(tmp_m, in, out0, out1); \ + } + +/* Description : Butterfly of 4 input vectors + Arguments : Inputs - in0, in1, in2, in3 + Outputs - out0, out1, out2, out3 + Details : Butterfly operation +*/ +#define BUTTERFLY_4(in0, in1, in2, in3, out0, out1, out2, out3) \ + { \ + out0 = in0 + in3; \ + out1 = in1 + in2; \ + \ + out2 = in1 - in2; \ + out3 = in0 - in3; \ + } + +/* Description : Butterfly of 8 input vectors + Arguments : Inputs - in0 ... in7 + Outputs - out0 .. out7 + Details : Butterfly operation +*/ +#define BUTTERFLY_8(in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, out2, \ + out3, out4, out5, out6, out7) \ + { \ + out0 = in0 + in7; \ + out1 = in1 + in6; \ + out2 = in2 + in5; \ + out3 = in3 + in4; \ + \ + out4 = in3 - in4; \ + out5 = in2 - in5; \ + out6 = in1 - in6; \ + out7 = in0 - in7; \ + } + +/* Description : Butterfly of 16 input vectors + Arguments : Inputs - in0 ... in15 + Outputs - out0 .. out15 + Details : Butterfly operation +*/ +#define BUTTERFLY_16(in0, in1, in2, in3, in4, in5, in6, in7, in8, in9, in10, \ + in11, in12, in13, in14, in15, out0, out1, out2, out3, \ + out4, out5, out6, out7, out8, out9, out10, out11, out12, \ + out13, out14, out15) \ + { \ + out0 = in0 + in15; \ + out1 = in1 + in14; \ + out2 = in2 + in13; \ + out3 = in3 + in12; \ + out4 = in4 + in11; \ + out5 = in5 + in10; \ + out6 = in6 + in9; \ + out7 = in7 + in8; \ + \ + out8 = in7 - in8; \ + out9 = in6 - in9; \ + out10 = in5 - in10; \ + out11 = in4 - in11; \ + out12 = in3 - in12; \ + out13 = in2 - in13; \ + out14 = in1 - in14; \ + out15 = in0 - in15; \ + } + +/* Description : Transpose input 8x8 byte block + Arguments : Inputs - in0, in1, in2, in3, in4, in5, in6, in7 + Outputs - out0, out1, out2, out3, out4, out5, out6, out7 + Return Type - as per RTYPE +*/ +#define TRANSPOSE8x8_UB(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7, out0, \ + out1, out2, out3, out4, out5, out6, out7) \ + { \ + v16i8 tmp0_m, tmp1_m, tmp2_m, tmp3_m; \ + v16i8 tmp4_m, tmp5_m, tmp6_m, tmp7_m; \ + \ + ILVR_B4_SB(in2, in0, in3, in1, in6, in4, in7, in5, tmp0_m, tmp1_m, tmp2_m, \ + tmp3_m); \ + ILVRL_B2_SB(tmp1_m, tmp0_m, tmp4_m, tmp5_m); \ + ILVRL_B2_SB(tmp3_m, tmp2_m, tmp6_m, tmp7_m); \ + ILVRL_W2(RTYPE, tmp6_m, tmp4_m, out0, out2); \ + ILVRL_W2(RTYPE, tmp7_m, tmp5_m, out4, out6); \ + SLDI_B2_0(RTYPE, out0, out2, out1, out3, 8); \ + SLDI_B2_0(RTYPE, out4, out6, out5, out7, 8); \ + } +#define TRANSPOSE8x8_UB_UB(...) TRANSPOSE8x8_UB(v16u8, __VA_ARGS__) + +/* Description : Transpose 16x8 block into 8x16 with byte elements in vectors + Arguments : Inputs - in0, in1, in2, in3, in4, in5, in6, in7, + in8, in9, in10, in11, in12, in13, in14, in15 + Outputs - out0, out1, out2, out3, out4, out5, out6, out7 + Return Type - unsigned byte +*/ +#define TRANSPOSE16x8_UB_UB(in0, in1, in2, in3, in4, in5, in6, in7, in8, in9, \ + in10, in11, in12, in13, in14, in15, out0, out1, \ + out2, out3, out4, out5, out6, out7) \ + { \ + v16u8 tmp0_m, tmp1_m, tmp2_m, tmp3_m; \ + v16u8 tmp4_m, tmp5_m, tmp6_m, tmp7_m; \ + \ + ILVEV_D2_UB(in0, in8, in1, in9, out7, out6); \ + ILVEV_D2_UB(in2, in10, in3, in11, out5, out4); \ + ILVEV_D2_UB(in4, in12, in5, in13, out3, out2); \ + ILVEV_D2_UB(in6, in14, in7, in15, out1, out0); \ + \ + tmp0_m = (v16u8)__msa_ilvev_b((v16i8)out6, (v16i8)out7); \ + tmp4_m = (v16u8)__msa_ilvod_b((v16i8)out6, (v16i8)out7); \ + tmp1_m = (v16u8)__msa_ilvev_b((v16i8)out4, (v16i8)out5); \ + tmp5_m = (v16u8)__msa_ilvod_b((v16i8)out4, (v16i8)out5); \ + out5 = (v16u8)__msa_ilvev_b((v16i8)out2, (v16i8)out3); \ + tmp6_m = (v16u8)__msa_ilvod_b((v16i8)out2, (v16i8)out3); \ + out7 = (v16u8)__msa_ilvev_b((v16i8)out0, (v16i8)out1); \ + tmp7_m = (v16u8)__msa_ilvod_b((v16i8)out0, (v16i8)out1); \ + \ + ILVEV_H2_UB(tmp0_m, tmp1_m, out5, out7, tmp2_m, tmp3_m); \ + out0 = (v16u8)__msa_ilvev_w((v4i32)tmp3_m, (v4i32)tmp2_m); \ + out4 = (v16u8)__msa_ilvod_w((v4i32)tmp3_m, (v4i32)tmp2_m); \ + \ + tmp2_m = (v16u8)__msa_ilvod_h((v8i16)tmp1_m, (v8i16)tmp0_m); \ + tmp3_m = (v16u8)__msa_ilvod_h((v8i16)out7, (v8i16)out5); \ + out2 = (v16u8)__msa_ilvev_w((v4i32)tmp3_m, (v4i32)tmp2_m); \ + out6 = (v16u8)__msa_ilvod_w((v4i32)tmp3_m, (v4i32)tmp2_m); \ + \ + ILVEV_H2_UB(tmp4_m, tmp5_m, tmp6_m, tmp7_m, tmp2_m, tmp3_m); \ + out1 = (v16u8)__msa_ilvev_w((v4i32)tmp3_m, (v4i32)tmp2_m); \ + out5 = (v16u8)__msa_ilvod_w((v4i32)tmp3_m, (v4i32)tmp2_m); \ + \ + tmp2_m = (v16u8)__msa_ilvod_h((v8i16)tmp5_m, (v8i16)tmp4_m); \ + tmp2_m = (v16u8)__msa_ilvod_h((v8i16)tmp5_m, (v8i16)tmp4_m); \ + tmp3_m = (v16u8)__msa_ilvod_h((v8i16)tmp7_m, (v8i16)tmp6_m); \ + tmp3_m = (v16u8)__msa_ilvod_h((v8i16)tmp7_m, (v8i16)tmp6_m); \ + out3 = (v16u8)__msa_ilvev_w((v4i32)tmp3_m, (v4i32)tmp2_m); \ + out7 = (v16u8)__msa_ilvod_w((v4i32)tmp3_m, (v4i32)tmp2_m); \ + } + +/* Description : Transpose 4x4 block with half word elements in vectors + Arguments : Inputs - in0, in1, in2, in3 + Outputs - out0, out1, out2, out3 + Return Type - signed halfword +*/ +#define TRANSPOSE4x4_SH_SH(in0, in1, in2, in3, out0, out1, out2, out3) \ + { \ + v8i16 s0_m, s1_m; \ + \ + ILVR_H2_SH(in1, in0, in3, in2, s0_m, s1_m); \ + ILVRL_W2_SH(s1_m, s0_m, out0, out2); \ + out1 = (v8i16)__msa_ilvl_d((v2i64)out0, (v2i64)out0); \ + out3 = (v8i16)__msa_ilvl_d((v2i64)out0, (v2i64)out2); \ + } + +/* Description : Transpose 4x8 block with half word elements in vectors + Arguments : Inputs - in0, in1, in2, in3, in4, in5, in6, in7 + Outputs - out0, out1, out2, out3, out4, out5, out6, out7 + Return Type - signed halfword +*/ +#define TRANSPOSE4X8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, \ + out2, out3, out4, out5, out6, out7) \ + { \ + v8i16 tmp0_m, tmp1_m, tmp2_m, tmp3_m; \ + v8i16 tmp0_n, tmp1_n, tmp2_n, tmp3_n; \ + v8i16 zero_m = { 0 }; \ + \ + ILVR_H4_SH(in1, in0, in3, in2, in5, in4, in7, in6, tmp0_n, tmp1_n, tmp2_n, \ + tmp3_n); \ + ILVRL_W2_SH(tmp1_n, tmp0_n, tmp0_m, tmp2_m); \ + ILVRL_W2_SH(tmp3_n, tmp2_n, tmp1_m, tmp3_m); \ + \ + out0 = (v8i16)__msa_ilvr_d((v2i64)tmp1_m, (v2i64)tmp0_m); \ + out1 = (v8i16)__msa_ilvl_d((v2i64)tmp1_m, (v2i64)tmp0_m); \ + out2 = (v8i16)__msa_ilvr_d((v2i64)tmp3_m, (v2i64)tmp2_m); \ + out3 = (v8i16)__msa_ilvl_d((v2i64)tmp3_m, (v2i64)tmp2_m); \ + \ + out4 = zero_m; \ + out5 = zero_m; \ + out6 = zero_m; \ + out7 = zero_m; \ + } + +/* Description : Transpose 8x4 block with half word elements in vectors + Arguments : Inputs - in0, in1, in2, in3, in4, in5, in6, in7 + Outputs - out0, out1, out2, out3, out4, out5, out6, out7 + Return Type - signed halfword +*/ +#define TRANSPOSE8X4_SH_SH(in0, in1, in2, in3, out0, out1, out2, out3) \ + { \ + v8i16 tmp0_m, tmp1_m, tmp2_m, tmp3_m; \ + \ + ILVR_H2_SH(in1, in0, in3, in2, tmp0_m, tmp1_m); \ + ILVL_H2_SH(in1, in0, in3, in2, tmp2_m, tmp3_m); \ + ILVR_W2_SH(tmp1_m, tmp0_m, tmp3_m, tmp2_m, out0, out2); \ + ILVL_W2_SH(tmp1_m, tmp0_m, tmp3_m, tmp2_m, out1, out3); \ + } + +/* Description : Transpose 8x8 block with half word elements in vectors + Arguments : Inputs - in0, in1, in2, in3, in4, in5, in6, in7 + Outputs - out0, out1, out2, out3, out4, out5, out6, out7 + Return Type - as per RTYPE +*/ +#define TRANSPOSE8x8_H(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7, out0, \ + out1, out2, out3, out4, out5, out6, out7) \ + { \ + v8i16 s0_m, s1_m; \ + v8i16 tmp0_m, tmp1_m, tmp2_m, tmp3_m; \ + v8i16 tmp4_m, tmp5_m, tmp6_m, tmp7_m; \ + \ + ILVR_H2_SH(in6, in4, in7, in5, s0_m, s1_m); \ + ILVRL_H2_SH(s1_m, s0_m, tmp0_m, tmp1_m); \ + ILVL_H2_SH(in6, in4, in7, in5, s0_m, s1_m); \ + ILVRL_H2_SH(s1_m, s0_m, tmp2_m, tmp3_m); \ + ILVR_H2_SH(in2, in0, in3, in1, s0_m, s1_m); \ + ILVRL_H2_SH(s1_m, s0_m, tmp4_m, tmp5_m); \ + ILVL_H2_SH(in2, in0, in3, in1, s0_m, s1_m); \ + ILVRL_H2_SH(s1_m, s0_m, tmp6_m, tmp7_m); \ + PCKEV_D4(RTYPE, tmp0_m, tmp4_m, tmp1_m, tmp5_m, tmp2_m, tmp6_m, tmp3_m, \ + tmp7_m, out0, out2, out4, out6); \ + out1 = (RTYPE)__msa_pckod_d((v2i64)tmp0_m, (v2i64)tmp4_m); \ + out3 = (RTYPE)__msa_pckod_d((v2i64)tmp1_m, (v2i64)tmp5_m); \ + out5 = (RTYPE)__msa_pckod_d((v2i64)tmp2_m, (v2i64)tmp6_m); \ + out7 = (RTYPE)__msa_pckod_d((v2i64)tmp3_m, (v2i64)tmp7_m); \ + } +#define TRANSPOSE8x8_SH_SH(...) TRANSPOSE8x8_H(v8i16, __VA_ARGS__) + +/* Description : Transpose 4x4 block with word elements in vectors + Arguments : Inputs - in0, in1, in2, in3 + Outputs - out0, out1, out2, out3 + Return Type - signed word +*/ +#define TRANSPOSE4x4_SW_SW(in0, in1, in2, in3, out0, out1, out2, out3) \ + { \ + v4i32 s0_m, s1_m, s2_m, s3_m; \ + \ + ILVRL_W2_SW(in1, in0, s0_m, s1_m); \ + ILVRL_W2_SW(in3, in2, s2_m, s3_m); \ + \ + out0 = (v4i32)__msa_ilvr_d((v2i64)s2_m, (v2i64)s0_m); \ + out1 = (v4i32)__msa_ilvl_d((v2i64)s2_m, (v2i64)s0_m); \ + out2 = (v4i32)__msa_ilvr_d((v2i64)s3_m, (v2i64)s1_m); \ + out3 = (v4i32)__msa_ilvl_d((v2i64)s3_m, (v2i64)s1_m); \ + } + +/* Description : Add block 4x4 + Arguments : Inputs - in0, in1, in2, in3, pdst, stride + Details : Least significant 4 bytes from each input vector are added to + the destination bytes, clipped between 0-255 and stored. +*/ +#define ADDBLK_ST4x4_UB(in0, in1, in2, in3, pdst, stride) \ + { \ + uint32_t src0_m, src1_m, src2_m, src3_m; \ + v8i16 inp0_m, inp1_m, res0_m, res1_m; \ + v16i8 dst0_m = { 0 }; \ + v16i8 dst1_m = { 0 }; \ + v16i8 zero_m = { 0 }; \ + \ + ILVR_D2_SH(in1, in0, in3, in2, inp0_m, inp1_m) \ + LW4(pdst, stride, src0_m, src1_m, src2_m, src3_m); \ + INSERT_W2_SB(src0_m, src1_m, dst0_m); \ + INSERT_W2_SB(src2_m, src3_m, dst1_m); \ + ILVR_B2_SH(zero_m, dst0_m, zero_m, dst1_m, res0_m, res1_m); \ + ADD2(res0_m, inp0_m, res1_m, inp1_m, res0_m, res1_m); \ + CLIP_SH2_0_255(res0_m, res1_m); \ + PCKEV_B2_SB(res0_m, res0_m, res1_m, res1_m, dst0_m, dst1_m); \ + ST4x4_UB(dst0_m, dst1_m, 0, 1, 0, 1, pdst, stride); \ + } + +/* Description : Pack even elements of input vectors & xor with 128 + Arguments : Inputs - in0, in1 + Output - out_m + Return Type - unsigned byte + Details : Signed byte even elements from 'in0' and 'in1' are packed + together in one vector and the resulting vector is xor'ed with + 128 to shift the range from signed to unsigned byte +*/ +#define PCKEV_XORI128_UB(in0, in1) \ + ({ \ + v16u8 out_m; \ + \ + out_m = (v16u8)__msa_pckev_b((v16i8)in1, (v16i8)in0); \ + out_m = (v16u8)__msa_xori_b((v16u8)out_m, 128); \ + out_m; \ + }) + +/* Description : Converts inputs to unsigned bytes, interleave, average & store + as 8x4 unsigned byte block + Arguments : Inputs - in0, in1, in2, in3, dst0, dst1, dst2, dst3, + pdst, stride +*/ +#define CONVERT_UB_AVG_ST8x4_UB(in0, in1, in2, in3, dst0, dst1, dst2, dst3, \ + pdst, stride) \ + { \ + v16u8 tmp0_m, tmp1_m, tmp2_m, tmp3_m; \ + \ + tmp0_m = PCKEV_XORI128_UB(in0, in1); \ + tmp1_m = PCKEV_XORI128_UB(in2, in3); \ + ILVR_D2_UB(dst1, dst0, dst3, dst2, tmp2_m, tmp3_m); \ + AVER_UB2_UB(tmp0_m, tmp2_m, tmp1_m, tmp3_m, tmp0_m, tmp1_m); \ + ST8x4_UB(tmp0_m, tmp1_m, pdst, stride); \ + } + +/* Description : Pack even byte elements and store byte vector in destination + memory + Arguments : Inputs - in0, in1, pdst +*/ +#define PCKEV_ST_SB(in0, in1, pdst) \ + { \ + v16i8 tmp_m; \ + \ + tmp_m = __msa_pckev_b((v16i8)in1, (v16i8)in0); \ + ST_SB(tmp_m, (pdst)); \ + } + +/* Description : Horizontal 2 tap filter kernel code + Arguments : Inputs - in0, in1, mask, coeff, shift +*/ +#define HORIZ_2TAP_FILT_UH(in0, in1, mask, coeff, shift) \ + ({ \ + v16i8 tmp0_m; \ + v8u16 tmp1_m; \ + \ + tmp0_m = __msa_vshf_b((v16i8)mask, (v16i8)in1, (v16i8)in0); \ + tmp1_m = __msa_dotp_u_h((v16u8)tmp0_m, (v16u8)coeff); \ + tmp1_m = (v8u16)__msa_srari_h((v8i16)tmp1_m, shift); \ + \ + tmp1_m; \ + }) +#endif // AOM_AOM_DSP_MIPS_MACROS_MSA_H_ diff --git a/third_party/aom/aom_dsp/mips/sad_msa.c b/third_party/aom/aom_dsp/mips/sad_msa.c new file mode 100644 index 0000000000..58cdd80d99 --- /dev/null +++ b/third_party/aom/aom_dsp/mips/sad_msa.c @@ -0,0 +1,800 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include "config/aom_dsp_rtcd.h" + +#include "aom_dsp/mips/macros_msa.h" + +#define SAD_INSVE_W4(RTYPE, in0, in1, in2, in3, out) \ + { \ + out = (RTYPE)__msa_insve_w((v4i32)out, 0, (v4i32)in0); \ + out = (RTYPE)__msa_insve_w((v4i32)out, 1, (v4i32)in1); \ + out = (RTYPE)__msa_insve_w((v4i32)out, 2, (v4i32)in2); \ + out = (RTYPE)__msa_insve_w((v4i32)out, 3, (v4i32)in3); \ + } +#define SAD_INSVE_W4_UB(...) SAD_INSVE_W4(v16u8, __VA_ARGS__) + +static uint32_t sad_4width_msa(const uint8_t *src_ptr, int32_t src_stride, + const uint8_t *ref_ptr, int32_t ref_stride, + int32_t height) { + int32_t ht_cnt; + uint32_t src0, src1, src2, src3, ref0, ref1, ref2, ref3; + v16u8 src = { 0 }; + v16u8 ref = { 0 }; + v16u8 diff; + v8u16 sad = { 0 }; + + for (ht_cnt = (height >> 2); ht_cnt--;) { + LW4(src_ptr, src_stride, src0, src1, src2, src3); + src_ptr += (4 * src_stride); + LW4(ref_ptr, ref_stride, ref0, ref1, ref2, ref3); + ref_ptr += (4 * ref_stride); + + INSERT_W4_UB(src0, src1, src2, src3, src); + INSERT_W4_UB(ref0, ref1, ref2, ref3, ref); + + diff = __msa_asub_u_b(src, ref); + sad += __msa_hadd_u_h(diff, diff); + } + + return HADD_UH_U32(sad); +} + +static uint32_t sad_8width_msa(const uint8_t *src, int32_t src_stride, + const uint8_t *ref, int32_t ref_stride, + int32_t height) { + int32_t ht_cnt; + v16u8 src0, src1, src2, src3, ref0, ref1, ref2, ref3; + v8u16 sad = { 0 }; + + for (ht_cnt = (height >> 2); ht_cnt--;) { + LD_UB4(src, src_stride, src0, src1, src2, src3); + src += (4 * src_stride); + LD_UB4(ref, ref_stride, ref0, ref1, ref2, ref3); + ref += (4 * ref_stride); + + PCKEV_D4_UB(src1, src0, src3, src2, ref1, ref0, ref3, ref2, src0, src1, + ref0, ref1); + sad += SAD_UB2_UH(src0, src1, ref0, ref1); + } + + return HADD_UH_U32(sad); +} + +static uint32_t sad_16width_msa(const uint8_t *src, int32_t src_stride, + const uint8_t *ref, int32_t ref_stride, + int32_t height) { + int32_t ht_cnt; + v16u8 src0, src1, ref0, ref1; + v8u16 sad = { 0 }; + + for (ht_cnt = (height >> 2); ht_cnt--;) { + LD_UB2(src, src_stride, src0, src1); + src += (2 * src_stride); + LD_UB2(ref, ref_stride, ref0, ref1); + ref += (2 * ref_stride); + sad += SAD_UB2_UH(src0, src1, ref0, ref1); + + LD_UB2(src, src_stride, src0, src1); + src += (2 * src_stride); + LD_UB2(ref, ref_stride, ref0, ref1); + ref += (2 * ref_stride); + sad += SAD_UB2_UH(src0, src1, ref0, ref1); + } + + return HADD_UH_U32(sad); +} + +static uint32_t sad_32width_msa(const uint8_t *src, int32_t src_stride, + const uint8_t *ref, int32_t ref_stride, + int32_t height) { + int32_t ht_cnt; + v16u8 src0, src1, ref0, ref1; + v8u16 sad = { 0 }; + + for (ht_cnt = (height >> 2); ht_cnt--;) { + LD_UB2(src, 16, src0, src1); + src += src_stride; + LD_UB2(ref, 16, ref0, ref1); + ref += ref_stride; + sad += SAD_UB2_UH(src0, src1, ref0, ref1); + + LD_UB2(src, 16, src0, src1); + src += src_stride; + LD_UB2(ref, 16, ref0, ref1); + ref += ref_stride; + sad += SAD_UB2_UH(src0, src1, ref0, ref1); + + LD_UB2(src, 16, src0, src1); + src += src_stride; + LD_UB2(ref, 16, ref0, ref1); + ref += ref_stride; + sad += SAD_UB2_UH(src0, src1, ref0, ref1); + + LD_UB2(src, 16, src0, src1); + src += src_stride; + LD_UB2(ref, 16, ref0, ref1); + ref += ref_stride; + sad += SAD_UB2_UH(src0, src1, ref0, ref1); + } + + return HADD_UH_U32(sad); +} + +static uint32_t sad_64width_msa(const uint8_t *src, int32_t src_stride, + const uint8_t *ref, int32_t ref_stride, + int32_t height) { + int32_t ht_cnt; + uint32_t sad = 0; + v16u8 src0, src1, src2, src3; + v16u8 ref0, ref1, ref2, ref3; + v8u16 sad0 = { 0 }; + v8u16 sad1 = { 0 }; + + for (ht_cnt = (height >> 1); ht_cnt--;) { + LD_UB4(src, 16, src0, src1, src2, src3); + src += src_stride; + LD_UB4(ref, 16, ref0, ref1, ref2, ref3); + ref += ref_stride; + sad0 += SAD_UB2_UH(src0, src1, ref0, ref1); + sad1 += SAD_UB2_UH(src2, src3, ref2, ref3); + + LD_UB4(src, 16, src0, src1, src2, src3); + src += src_stride; + LD_UB4(ref, 16, ref0, ref1, ref2, ref3); + ref += ref_stride; + sad0 += SAD_UB2_UH(src0, src1, ref0, ref1); + sad1 += SAD_UB2_UH(src2, src3, ref2, ref3); + } + + sad = HADD_UH_U32(sad0); + sad += HADD_UH_U32(sad1); + + return sad; +} + +static void sad_4width_x4d_msa(const uint8_t *src_ptr, int32_t src_stride, + const uint8_t *const aref_ptr[], + int32_t ref_stride, int32_t height, + uint32_t *sad_array) { + const uint8_t *ref0_ptr, *ref1_ptr, *ref2_ptr, *ref3_ptr; + int32_t ht_cnt; + uint32_t src0, src1, src2, src3; + uint32_t ref0, ref1, ref2, ref3; + v16u8 src = { 0 }; + v16u8 ref = { 0 }; + v16u8 diff; + v8u16 sad0 = { 0 }; + v8u16 sad1 = { 0 }; + v8u16 sad2 = { 0 }; + v8u16 sad3 = { 0 }; + + ref0_ptr = aref_ptr[0]; + ref1_ptr = aref_ptr[1]; + ref2_ptr = aref_ptr[2]; + ref3_ptr = aref_ptr[3]; + + for (ht_cnt = (height >> 2); ht_cnt--;) { + LW4(src_ptr, src_stride, src0, src1, src2, src3); + INSERT_W4_UB(src0, src1, src2, src3, src); + src_ptr += (4 * src_stride); + + LW4(ref0_ptr, ref_stride, ref0, ref1, ref2, ref3); + INSERT_W4_UB(ref0, ref1, ref2, ref3, ref); + ref0_ptr += (4 * ref_stride); + + diff = __msa_asub_u_b(src, ref); + sad0 += __msa_hadd_u_h(diff, diff); + + LW4(ref1_ptr, ref_stride, ref0, ref1, ref2, ref3); + INSERT_W4_UB(ref0, ref1, ref2, ref3, ref); + ref1_ptr += (4 * ref_stride); + + diff = __msa_asub_u_b(src, ref); + sad1 += __msa_hadd_u_h(diff, diff); + + LW4(ref2_ptr, ref_stride, ref0, ref1, ref2, ref3); + INSERT_W4_UB(ref0, ref1, ref2, ref3, ref); + ref2_ptr += (4 * ref_stride); + + diff = __msa_asub_u_b(src, ref); + sad2 += __msa_hadd_u_h(diff, diff); + + LW4(ref3_ptr, ref_stride, ref0, ref1, ref2, ref3); + INSERT_W4_UB(ref0, ref1, ref2, ref3, ref); + ref3_ptr += (4 * ref_stride); + + diff = __msa_asub_u_b(src, ref); + sad3 += __msa_hadd_u_h(diff, diff); + } + + sad_array[0] = HADD_UH_U32(sad0); + sad_array[1] = HADD_UH_U32(sad1); + sad_array[2] = HADD_UH_U32(sad2); + sad_array[3] = HADD_UH_U32(sad3); +} + +static void sad_8width_x4d_msa(const uint8_t *src_ptr, int32_t src_stride, + const uint8_t *const aref_ptr[], + int32_t ref_stride, int32_t height, + uint32_t *sad_array) { + int32_t ht_cnt; + const uint8_t *ref0_ptr, *ref1_ptr, *ref2_ptr, *ref3_ptr; + v16u8 src0, src1, src2, src3; + v16u8 ref0, ref1, ref2, ref3, ref4, ref5, ref6, ref7; + v16u8 ref8, ref9, ref10, ref11, ref12, ref13, ref14, ref15; + v8u16 sad0 = { 0 }; + v8u16 sad1 = { 0 }; + v8u16 sad2 = { 0 }; + v8u16 sad3 = { 0 }; + + ref0_ptr = aref_ptr[0]; + ref1_ptr = aref_ptr[1]; + ref2_ptr = aref_ptr[2]; + ref3_ptr = aref_ptr[3]; + + for (ht_cnt = (height >> 2); ht_cnt--;) { + LD_UB4(src_ptr, src_stride, src0, src1, src2, src3); + src_ptr += (4 * src_stride); + LD_UB4(ref0_ptr, ref_stride, ref0, ref1, ref2, ref3); + ref0_ptr += (4 * ref_stride); + LD_UB4(ref1_ptr, ref_stride, ref4, ref5, ref6, ref7); + ref1_ptr += (4 * ref_stride); + LD_UB4(ref2_ptr, ref_stride, ref8, ref9, ref10, ref11); + ref2_ptr += (4 * ref_stride); + LD_UB4(ref3_ptr, ref_stride, ref12, ref13, ref14, ref15); + ref3_ptr += (4 * ref_stride); + + PCKEV_D2_UB(src1, src0, src3, src2, src0, src1); + PCKEV_D2_UB(ref1, ref0, ref3, ref2, ref0, ref1); + sad0 += SAD_UB2_UH(src0, src1, ref0, ref1); + + PCKEV_D2_UB(ref5, ref4, ref7, ref6, ref0, ref1); + sad1 += SAD_UB2_UH(src0, src1, ref0, ref1); + + PCKEV_D2_UB(ref9, ref8, ref11, ref10, ref0, ref1); + sad2 += SAD_UB2_UH(src0, src1, ref0, ref1); + + PCKEV_D2_UB(ref13, ref12, ref15, ref14, ref0, ref1); + sad3 += SAD_UB2_UH(src0, src1, ref0, ref1); + } + + sad_array[0] = HADD_UH_U32(sad0); + sad_array[1] = HADD_UH_U32(sad1); + sad_array[2] = HADD_UH_U32(sad2); + sad_array[3] = HADD_UH_U32(sad3); +} + +static void sad_16width_x4d_msa(const uint8_t *src_ptr, int32_t src_stride, + const uint8_t *const aref_ptr[], + int32_t ref_stride, int32_t height, + uint32_t *sad_array) { + int32_t ht_cnt; + const uint8_t *ref0_ptr, *ref1_ptr, *ref2_ptr, *ref3_ptr; + v16u8 src, ref0, ref1, ref2, ref3, diff; + v8u16 sad0 = { 0 }; + v8u16 sad1 = { 0 }; + v8u16 sad2 = { 0 }; + v8u16 sad3 = { 0 }; + + ref0_ptr = aref_ptr[0]; + ref1_ptr = aref_ptr[1]; + ref2_ptr = aref_ptr[2]; + ref3_ptr = aref_ptr[3]; + + for (ht_cnt = (height >> 1); ht_cnt--;) { + src = LD_UB(src_ptr); + src_ptr += src_stride; + ref0 = LD_UB(ref0_ptr); + ref0_ptr += ref_stride; + ref1 = LD_UB(ref1_ptr); + ref1_ptr += ref_stride; + ref2 = LD_UB(ref2_ptr); + ref2_ptr += ref_stride; + ref3 = LD_UB(ref3_ptr); + ref3_ptr += ref_stride; + + diff = __msa_asub_u_b(src, ref0); + sad0 += __msa_hadd_u_h(diff, diff); + diff = __msa_asub_u_b(src, ref1); + sad1 += __msa_hadd_u_h(diff, diff); + diff = __msa_asub_u_b(src, ref2); + sad2 += __msa_hadd_u_h(diff, diff); + diff = __msa_asub_u_b(src, ref3); + sad3 += __msa_hadd_u_h(diff, diff); + + src = LD_UB(src_ptr); + src_ptr += src_stride; + ref0 = LD_UB(ref0_ptr); + ref0_ptr += ref_stride; + ref1 = LD_UB(ref1_ptr); + ref1_ptr += ref_stride; + ref2 = LD_UB(ref2_ptr); + ref2_ptr += ref_stride; + ref3 = LD_UB(ref3_ptr); + ref3_ptr += ref_stride; + + diff = __msa_asub_u_b(src, ref0); + sad0 += __msa_hadd_u_h(diff, diff); + diff = __msa_asub_u_b(src, ref1); + sad1 += __msa_hadd_u_h(diff, diff); + diff = __msa_asub_u_b(src, ref2); + sad2 += __msa_hadd_u_h(diff, diff); + diff = __msa_asub_u_b(src, ref3); + sad3 += __msa_hadd_u_h(diff, diff); + } + + sad_array[0] = HADD_UH_U32(sad0); + sad_array[1] = HADD_UH_U32(sad1); + sad_array[2] = HADD_UH_U32(sad2); + sad_array[3] = HADD_UH_U32(sad3); +} + +static void sad_32width_x4d_msa(const uint8_t *src, int32_t src_stride, + const uint8_t *const aref_ptr[], + int32_t ref_stride, int32_t height, + uint32_t *sad_array) { + const uint8_t *ref0_ptr, *ref1_ptr, *ref2_ptr, *ref3_ptr; + int32_t ht_cnt; + v16u8 src0, src1, ref0, ref1; + v8u16 sad0 = { 0 }; + v8u16 sad1 = { 0 }; + v8u16 sad2 = { 0 }; + v8u16 sad3 = { 0 }; + + ref0_ptr = aref_ptr[0]; + ref1_ptr = aref_ptr[1]; + ref2_ptr = aref_ptr[2]; + ref3_ptr = aref_ptr[3]; + + for (ht_cnt = height; ht_cnt--;) { + LD_UB2(src, 16, src0, src1); + src += src_stride; + + LD_UB2(ref0_ptr, 16, ref0, ref1); + ref0_ptr += ref_stride; + sad0 += SAD_UB2_UH(src0, src1, ref0, ref1); + + LD_UB2(ref1_ptr, 16, ref0, ref1); + ref1_ptr += ref_stride; + sad1 += SAD_UB2_UH(src0, src1, ref0, ref1); + + LD_UB2(ref2_ptr, 16, ref0, ref1); + ref2_ptr += ref_stride; + sad2 += SAD_UB2_UH(src0, src1, ref0, ref1); + + LD_UB2(ref3_ptr, 16, ref0, ref1); + ref3_ptr += ref_stride; + sad3 += SAD_UB2_UH(src0, src1, ref0, ref1); + } + + sad_array[0] = HADD_UH_U32(sad0); + sad_array[1] = HADD_UH_U32(sad1); + sad_array[2] = HADD_UH_U32(sad2); + sad_array[3] = HADD_UH_U32(sad3); +} + +static void sad_64width_x4d_msa(const uint8_t *src, int32_t src_stride, + const uint8_t *const aref_ptr[], + int32_t ref_stride, int32_t height, + uint32_t *sad_array) { + const uint8_t *ref0_ptr, *ref1_ptr, *ref2_ptr, *ref3_ptr; + int32_t ht_cnt; + v16u8 src0, src1, src2, src3; + v16u8 ref0, ref1, ref2, ref3; + v8u16 sad0_0 = { 0 }; + v8u16 sad0_1 = { 0 }; + v8u16 sad1_0 = { 0 }; + v8u16 sad1_1 = { 0 }; + v8u16 sad2_0 = { 0 }; + v8u16 sad2_1 = { 0 }; + v8u16 sad3_0 = { 0 }; + v8u16 sad3_1 = { 0 }; + + ref0_ptr = aref_ptr[0]; + ref1_ptr = aref_ptr[1]; + ref2_ptr = aref_ptr[2]; + ref3_ptr = aref_ptr[3]; + + for (ht_cnt = height; ht_cnt--;) { + LD_UB4(src, 16, src0, src1, src2, src3); + src += src_stride; + + LD_UB4(ref0_ptr, 16, ref0, ref1, ref2, ref3); + ref0_ptr += ref_stride; + sad0_0 += SAD_UB2_UH(src0, src1, ref0, ref1); + sad0_1 += SAD_UB2_UH(src2, src3, ref2, ref3); + + LD_UB4(ref1_ptr, 16, ref0, ref1, ref2, ref3); + ref1_ptr += ref_stride; + sad1_0 += SAD_UB2_UH(src0, src1, ref0, ref1); + sad1_1 += SAD_UB2_UH(src2, src3, ref2, ref3); + + LD_UB4(ref2_ptr, 16, ref0, ref1, ref2, ref3); + ref2_ptr += ref_stride; + sad2_0 += SAD_UB2_UH(src0, src1, ref0, ref1); + sad2_1 += SAD_UB2_UH(src2, src3, ref2, ref3); + + LD_UB4(ref3_ptr, 16, ref0, ref1, ref2, ref3); + ref3_ptr += ref_stride; + sad3_0 += SAD_UB2_UH(src0, src1, ref0, ref1); + sad3_1 += SAD_UB2_UH(src2, src3, ref2, ref3); + } + + sad_array[0] = HADD_UH_U32(sad0_0); + sad_array[0] += HADD_UH_U32(sad0_1); + sad_array[1] = HADD_UH_U32(sad1_0); + sad_array[1] += HADD_UH_U32(sad1_1); + sad_array[2] = HADD_UH_U32(sad2_0); + sad_array[2] += HADD_UH_U32(sad2_1); + sad_array[3] = HADD_UH_U32(sad3_0); + sad_array[3] += HADD_UH_U32(sad3_1); +} + +static uint32_t avgsad_4width_msa(const uint8_t *src_ptr, int32_t src_stride, + const uint8_t *ref_ptr, int32_t ref_stride, + int32_t height, const uint8_t *sec_pred) { + int32_t ht_cnt; + uint32_t src0, src1, src2, src3, ref0, ref1, ref2, ref3; + v16u8 src = { 0 }; + v16u8 ref = { 0 }; + v16u8 diff, pred, comp; + v8u16 sad = { 0 }; + + for (ht_cnt = (height >> 2); ht_cnt--;) { + LW4(src_ptr, src_stride, src0, src1, src2, src3); + src_ptr += (4 * src_stride); + LW4(ref_ptr, ref_stride, ref0, ref1, ref2, ref3); + ref_ptr += (4 * ref_stride); + pred = LD_UB(sec_pred); + sec_pred += 16; + + INSERT_W4_UB(src0, src1, src2, src3, src); + INSERT_W4_UB(ref0, ref1, ref2, ref3, ref); + + comp = __msa_aver_u_b(pred, ref); + diff = __msa_asub_u_b(src, comp); + sad += __msa_hadd_u_h(diff, diff); + } + + return HADD_UH_U32(sad); +} + +static uint32_t avgsad_8width_msa(const uint8_t *src, int32_t src_stride, + const uint8_t *ref, int32_t ref_stride, + int32_t height, const uint8_t *sec_pred) { + int32_t ht_cnt; + v16u8 src0, src1, src2, src3, ref0, ref1, ref2, ref3; + v16u8 diff0, diff1, pred0, pred1; + v8u16 sad = { 0 }; + + for (ht_cnt = (height >> 2); ht_cnt--;) { + LD_UB4(src, src_stride, src0, src1, src2, src3); + src += (4 * src_stride); + LD_UB4(ref, ref_stride, ref0, ref1, ref2, ref3); + ref += (4 * ref_stride); + LD_UB2(sec_pred, 16, pred0, pred1); + sec_pred += 32; + PCKEV_D4_UB(src1, src0, src3, src2, ref1, ref0, ref3, ref2, src0, src1, + ref0, ref1); + AVER_UB2_UB(pred0, ref0, pred1, ref1, diff0, diff1); + sad += SAD_UB2_UH(src0, src1, diff0, diff1); + } + + return HADD_UH_U32(sad); +} + +static uint32_t avgsad_16width_msa(const uint8_t *src, int32_t src_stride, + const uint8_t *ref, int32_t ref_stride, + int32_t height, const uint8_t *sec_pred) { + int32_t ht_cnt; + v16u8 src0, src1, src2, src3, ref0, ref1, ref2, ref3; + v16u8 pred0, pred1, pred2, pred3, comp0, comp1; + v8u16 sad = { 0 }; + + for (ht_cnt = (height >> 3); ht_cnt--;) { + LD_UB4(src, src_stride, src0, src1, src2, src3); + src += (4 * src_stride); + LD_UB4(ref, ref_stride, ref0, ref1, ref2, ref3); + ref += (4 * ref_stride); + LD_UB4(sec_pred, 16, pred0, pred1, pred2, pred3); + sec_pred += (4 * 16); + AVER_UB2_UB(pred0, ref0, pred1, ref1, comp0, comp1); + sad += SAD_UB2_UH(src0, src1, comp0, comp1); + AVER_UB2_UB(pred2, ref2, pred3, ref3, comp0, comp1); + sad += SAD_UB2_UH(src2, src3, comp0, comp1); + + LD_UB4(src, src_stride, src0, src1, src2, src3); + src += (4 * src_stride); + LD_UB4(ref, ref_stride, ref0, ref1, ref2, ref3); + ref += (4 * ref_stride); + LD_UB4(sec_pred, 16, pred0, pred1, pred2, pred3); + sec_pred += (4 * 16); + AVER_UB2_UB(pred0, ref0, pred1, ref1, comp0, comp1); + sad += SAD_UB2_UH(src0, src1, comp0, comp1); + AVER_UB2_UB(pred2, ref2, pred3, ref3, comp0, comp1); + sad += SAD_UB2_UH(src2, src3, comp0, comp1); + } + + return HADD_UH_U32(sad); +} + +static uint32_t avgsad_32width_msa(const uint8_t *src, int32_t src_stride, + const uint8_t *ref, int32_t ref_stride, + int32_t height, const uint8_t *sec_pred) { + int32_t ht_cnt; + v16u8 src0, src1, src2, src3, src4, src5, src6, src7; + v16u8 ref0, ref1, ref2, ref3, ref4, ref5, ref6, ref7; + v16u8 pred0, pred1, pred2, pred3, pred4, pred5, pred6, pred7; + v16u8 comp0, comp1; + v8u16 sad = { 0 }; + + for (ht_cnt = (height >> 2); ht_cnt--;) { + LD_UB4(src, src_stride, src0, src2, src4, src6); + LD_UB4(src + 16, src_stride, src1, src3, src5, src7); + src += (4 * src_stride); + + LD_UB4(ref, ref_stride, ref0, ref2, ref4, ref6); + LD_UB4(ref + 16, ref_stride, ref1, ref3, ref5, ref7); + ref += (4 * ref_stride); + + LD_UB4(sec_pred, 32, pred0, pred2, pred4, pred6); + LD_UB4(sec_pred + 16, 32, pred1, pred3, pred5, pred7); + sec_pred += (4 * 32); + + AVER_UB2_UB(pred0, ref0, pred1, ref1, comp0, comp1); + sad += SAD_UB2_UH(src0, src1, comp0, comp1); + AVER_UB2_UB(pred2, ref2, pred3, ref3, comp0, comp1); + sad += SAD_UB2_UH(src2, src3, comp0, comp1); + AVER_UB2_UB(pred4, ref4, pred5, ref5, comp0, comp1); + sad += SAD_UB2_UH(src4, src5, comp0, comp1); + AVER_UB2_UB(pred6, ref6, pred7, ref7, comp0, comp1); + sad += SAD_UB2_UH(src6, src7, comp0, comp1); + } + + return HADD_UH_U32(sad); +} + +static uint32_t avgsad_64width_msa(const uint8_t *src, int32_t src_stride, + const uint8_t *ref, int32_t ref_stride, + int32_t height, const uint8_t *sec_pred) { + int32_t ht_cnt; + v16u8 src0, src1, src2, src3; + v16u8 ref0, ref1, ref2, ref3; + v16u8 comp0, comp1, comp2, comp3; + v16u8 pred0, pred1, pred2, pred3; + v8u16 sad0 = { 0 }; + v8u16 sad1 = { 0 }; + v4u32 sad; + + for (ht_cnt = (height >> 2); ht_cnt--;) { + LD_UB4(src, 16, src0, src1, src2, src3); + src += src_stride; + LD_UB4(ref, 16, ref0, ref1, ref2, ref3); + ref += ref_stride; + LD_UB4(sec_pred, 16, pred0, pred1, pred2, pred3); + sec_pred += 64; + AVER_UB4_UB(pred0, ref0, pred1, ref1, pred2, ref2, pred3, ref3, comp0, + comp1, comp2, comp3); + sad0 += SAD_UB2_UH(src0, src1, comp0, comp1); + sad1 += SAD_UB2_UH(src2, src3, comp2, comp3); + + LD_UB4(src, 16, src0, src1, src2, src3); + src += src_stride; + LD_UB4(ref, 16, ref0, ref1, ref2, ref3); + ref += ref_stride; + LD_UB4(sec_pred, 16, pred0, pred1, pred2, pred3); + sec_pred += 64; + AVER_UB4_UB(pred0, ref0, pred1, ref1, pred2, ref2, pred3, ref3, comp0, + comp1, comp2, comp3); + sad0 += SAD_UB2_UH(src0, src1, comp0, comp1); + sad1 += SAD_UB2_UH(src2, src3, comp2, comp3); + + LD_UB4(src, 16, src0, src1, src2, src3); + src += src_stride; + LD_UB4(ref, 16, ref0, ref1, ref2, ref3); + ref += ref_stride; + LD_UB4(sec_pred, 16, pred0, pred1, pred2, pred3); + sec_pred += 64; + AVER_UB4_UB(pred0, ref0, pred1, ref1, pred2, ref2, pred3, ref3, comp0, + comp1, comp2, comp3); + sad0 += SAD_UB2_UH(src0, src1, comp0, comp1); + sad1 += SAD_UB2_UH(src2, src3, comp2, comp3); + + LD_UB4(src, 16, src0, src1, src2, src3); + src += src_stride; + LD_UB4(ref, 16, ref0, ref1, ref2, ref3); + ref += ref_stride; + LD_UB4(sec_pred, 16, pred0, pred1, pred2, pred3); + sec_pred += 64; + AVER_UB4_UB(pred0, ref0, pred1, ref1, pred2, ref2, pred3, ref3, comp0, + comp1, comp2, comp3); + sad0 += SAD_UB2_UH(src0, src1, comp0, comp1); + sad1 += SAD_UB2_UH(src2, src3, comp2, comp3); + } + + sad = __msa_hadd_u_w(sad0, sad0); + sad += __msa_hadd_u_w(sad1, sad1); + + return HADD_SW_S32(sad); +} + +#define AOM_SAD_4xHEIGHT_MSA(height) \ + uint32_t aom_sad4x##height##_msa(const uint8_t *src, int32_t src_stride, \ + const uint8_t *ref, int32_t ref_stride) { \ + return sad_4width_msa(src, src_stride, ref, ref_stride, height); \ + } + +#define AOM_SAD_8xHEIGHT_MSA(height) \ + uint32_t aom_sad8x##height##_msa(const uint8_t *src, int32_t src_stride, \ + const uint8_t *ref, int32_t ref_stride) { \ + return sad_8width_msa(src, src_stride, ref, ref_stride, height); \ + } + +#define AOM_SAD_16xHEIGHT_MSA(height) \ + uint32_t aom_sad16x##height##_msa(const uint8_t *src, int32_t src_stride, \ + const uint8_t *ref, int32_t ref_stride) { \ + return sad_16width_msa(src, src_stride, ref, ref_stride, height); \ + } + +#define AOM_SAD_32xHEIGHT_MSA(height) \ + uint32_t aom_sad32x##height##_msa(const uint8_t *src, int32_t src_stride, \ + const uint8_t *ref, int32_t ref_stride) { \ + return sad_32width_msa(src, src_stride, ref, ref_stride, height); \ + } + +#define AOM_SAD_64xHEIGHT_MSA(height) \ + uint32_t aom_sad64x##height##_msa(const uint8_t *src, int32_t src_stride, \ + const uint8_t *ref, int32_t ref_stride) { \ + return sad_64width_msa(src, src_stride, ref, ref_stride, height); \ + } + +#define AOM_SAD_4xHEIGHTx4D_MSA(height) \ + void aom_sad4x##height##x4d_msa(const uint8_t *src, int32_t src_stride, \ + const uint8_t *const refs[], \ + int32_t ref_stride, uint32_t *sads) { \ + sad_4width_x4d_msa(src, src_stride, refs, ref_stride, height, sads); \ + } + +#define AOM_SAD_8xHEIGHTx4D_MSA(height) \ + void aom_sad8x##height##x4d_msa(const uint8_t *src, int32_t src_stride, \ + const uint8_t *const refs[], \ + int32_t ref_stride, uint32_t *sads) { \ + sad_8width_x4d_msa(src, src_stride, refs, ref_stride, height, sads); \ + } + +#define AOM_SAD_16xHEIGHTx4D_MSA(height) \ + void aom_sad16x##height##x4d_msa(const uint8_t *src, int32_t src_stride, \ + const uint8_t *const refs[], \ + int32_t ref_stride, uint32_t *sads) { \ + sad_16width_x4d_msa(src, src_stride, refs, ref_stride, height, sads); \ + } + +#define AOM_SAD_32xHEIGHTx4D_MSA(height) \ + void aom_sad32x##height##x4d_msa(const uint8_t *src, int32_t src_stride, \ + const uint8_t *const refs[], \ + int32_t ref_stride, uint32_t *sads) { \ + sad_32width_x4d_msa(src, src_stride, refs, ref_stride, height, sads); \ + } + +#define AOM_SAD_64xHEIGHTx4D_MSA(height) \ + void aom_sad64x##height##x4d_msa(const uint8_t *src, int32_t src_stride, \ + const uint8_t *const refs[], \ + int32_t ref_stride, uint32_t *sads) { \ + sad_64width_x4d_msa(src, src_stride, refs, ref_stride, height, sads); \ + } + +#define AOM_AVGSAD_4xHEIGHT_MSA(height) \ + uint32_t aom_sad4x##height##_avg_msa(const uint8_t *src, int32_t src_stride, \ + const uint8_t *ref, int32_t ref_stride, \ + const uint8_t *second_pred) { \ + return avgsad_4width_msa(src, src_stride, ref, ref_stride, height, \ + second_pred); \ + } + +#define AOM_AVGSAD_8xHEIGHT_MSA(height) \ + uint32_t aom_sad8x##height##_avg_msa(const uint8_t *src, int32_t src_stride, \ + const uint8_t *ref, int32_t ref_stride, \ + const uint8_t *second_pred) { \ + return avgsad_8width_msa(src, src_stride, ref, ref_stride, height, \ + second_pred); \ + } + +#define AOM_AVGSAD_16xHEIGHT_MSA(height) \ + uint32_t aom_sad16x##height##_avg_msa( \ + const uint8_t *src, int32_t src_stride, const uint8_t *ref, \ + int32_t ref_stride, const uint8_t *second_pred) { \ + return avgsad_16width_msa(src, src_stride, ref, ref_stride, height, \ + second_pred); \ + } + +#define AOM_AVGSAD_32xHEIGHT_MSA(height) \ + uint32_t aom_sad32x##height##_avg_msa( \ + const uint8_t *src, int32_t src_stride, const uint8_t *ref, \ + int32_t ref_stride, const uint8_t *second_pred) { \ + return avgsad_32width_msa(src, src_stride, ref, ref_stride, height, \ + second_pred); \ + } + +#define AOM_AVGSAD_64xHEIGHT_MSA(height) \ + uint32_t aom_sad64x##height##_avg_msa( \ + const uint8_t *src, int32_t src_stride, const uint8_t *ref, \ + int32_t ref_stride, const uint8_t *second_pred) { \ + return avgsad_64width_msa(src, src_stride, ref, ref_stride, height, \ + second_pred); \ + } + +/* clang-format off */ +// 64x64 +AOM_SAD_64xHEIGHT_MSA(64) +AOM_SAD_64xHEIGHTx4D_MSA(64) +AOM_AVGSAD_64xHEIGHT_MSA(64) + +// 64x32 +AOM_SAD_64xHEIGHT_MSA(32) +AOM_SAD_64xHEIGHTx4D_MSA(32) +AOM_AVGSAD_64xHEIGHT_MSA(32) + +// 32x64 +AOM_SAD_32xHEIGHT_MSA(64) +AOM_SAD_32xHEIGHTx4D_MSA(64) +AOM_AVGSAD_32xHEIGHT_MSA(64) + +// 32x32 +AOM_SAD_32xHEIGHT_MSA(32) +AOM_SAD_32xHEIGHTx4D_MSA(32) +AOM_AVGSAD_32xHEIGHT_MSA(32) + +// 32x16 +AOM_SAD_32xHEIGHT_MSA(16) +AOM_SAD_32xHEIGHTx4D_MSA(16) +AOM_AVGSAD_32xHEIGHT_MSA(16) + +// 16x32 +AOM_SAD_16xHEIGHT_MSA(32) +AOM_SAD_16xHEIGHTx4D_MSA(32) +AOM_AVGSAD_16xHEIGHT_MSA(32) + +// 16x16 +AOM_SAD_16xHEIGHT_MSA(16) +AOM_SAD_16xHEIGHTx4D_MSA(16) +AOM_AVGSAD_16xHEIGHT_MSA(16) + +// 16x8 +AOM_SAD_16xHEIGHT_MSA(8) +AOM_SAD_16xHEIGHTx4D_MSA(8) +AOM_AVGSAD_16xHEIGHT_MSA(8) + +// 8x16 +AOM_SAD_8xHEIGHT_MSA(16) +AOM_SAD_8xHEIGHTx4D_MSA(16) +AOM_AVGSAD_8xHEIGHT_MSA(16) + +// 8x8 +AOM_SAD_8xHEIGHT_MSA(8) +AOM_SAD_8xHEIGHTx4D_MSA(8) +AOM_AVGSAD_8xHEIGHT_MSA(8) + +// 8x4 +AOM_SAD_8xHEIGHT_MSA(4) +AOM_SAD_8xHEIGHTx4D_MSA(4) +AOM_AVGSAD_8xHEIGHT_MSA(4) + +// 4x8 +AOM_SAD_4xHEIGHT_MSA(8) +AOM_SAD_4xHEIGHTx4D_MSA(8) +AOM_AVGSAD_4xHEIGHT_MSA(8) + +// 4x4 +AOM_SAD_4xHEIGHT_MSA(4) +AOM_SAD_4xHEIGHTx4D_MSA(4) +AOM_AVGSAD_4xHEIGHT_MSA(4) + /* clang-format on */ diff --git a/third_party/aom/aom_dsp/mips/sub_pixel_variance_msa.c b/third_party/aom/aom_dsp/mips/sub_pixel_variance_msa.c new file mode 100644 index 0000000000..810b6efaaa --- /dev/null +++ b/third_party/aom/aom_dsp/mips/sub_pixel_variance_msa.c @@ -0,0 +1,1792 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include "config/aom_dsp_rtcd.h" + +#include "aom_ports/mem.h" +#include "aom_dsp/mips/macros_msa.h" +#include "aom_dsp/aom_filter.h" +#include "aom_dsp/variance.h" + +#define CALC_MSE_AVG_B(src, ref, var, sub) \ + { \ + v16u8 src_l0_m, src_l1_m; \ + v8i16 res_l0_m, res_l1_m; \ + \ + ILVRL_B2_UB(src, ref, src_l0_m, src_l1_m); \ + HSUB_UB2_SH(src_l0_m, src_l1_m, res_l0_m, res_l1_m); \ + DPADD_SH2_SW(res_l0_m, res_l1_m, res_l0_m, res_l1_m, var, var); \ + \ + sub += res_l0_m + res_l1_m; \ + } + +#define VARIANCE_WxH(sse, diff, shift) sse - (((uint32_t)diff * diff) >> shift) + +#define VARIANCE_LARGE_WxH(sse, diff, shift) \ + sse - (((int64_t)diff * diff) >> shift) + +static uint32_t avg_sse_diff_4width_msa(const uint8_t *src_ptr, + int32_t src_stride, + const uint8_t *ref_ptr, + int32_t ref_stride, + const uint8_t *sec_pred, int32_t height, + int32_t *diff) { + int32_t ht_cnt; + uint32_t src0, src1, src2, src3; + uint32_t ref0, ref1, ref2, ref3; + v16u8 pred, src = { 0 }; + v16u8 ref = { 0 }; + v8i16 avg = { 0 }; + v4i32 vec, var = { 0 }; + + for (ht_cnt = (height >> 2); ht_cnt--;) { + pred = LD_UB(sec_pred); + sec_pred += 16; + LW4(src_ptr, src_stride, src0, src1, src2, src3); + src_ptr += (4 * src_stride); + LW4(ref_ptr, ref_stride, ref0, ref1, ref2, ref3); + ref_ptr += (4 * ref_stride); + + INSERT_W4_UB(src0, src1, src2, src3, src); + INSERT_W4_UB(ref0, ref1, ref2, ref3, ref); + + src = __msa_aver_u_b(src, pred); + CALC_MSE_AVG_B(src, ref, var, avg); + } + + vec = __msa_hadd_s_w(avg, avg); + *diff = HADD_SW_S32(vec); + + return HADD_SW_S32(var); +} + +static uint32_t avg_sse_diff_8width_msa(const uint8_t *src_ptr, + int32_t src_stride, + const uint8_t *ref_ptr, + int32_t ref_stride, + const uint8_t *sec_pred, int32_t height, + int32_t *diff) { + int32_t ht_cnt; + v16u8 src0, src1, src2, src3; + v16u8 ref0, ref1, ref2, ref3; + v16u8 pred0, pred1; + v8i16 avg = { 0 }; + v4i32 vec, var = { 0 }; + + for (ht_cnt = (height >> 2); ht_cnt--;) { + LD_UB2(sec_pred, 16, pred0, pred1); + sec_pred += 32; + LD_UB4(src_ptr, src_stride, src0, src1, src2, src3); + src_ptr += (4 * src_stride); + LD_UB4(ref_ptr, ref_stride, ref0, ref1, ref2, ref3); + ref_ptr += (4 * ref_stride); + + PCKEV_D4_UB(src1, src0, src3, src2, ref1, ref0, ref3, ref2, src0, src1, + ref0, ref1); + AVER_UB2_UB(src0, pred0, src1, pred1, src0, src1); + CALC_MSE_AVG_B(src0, ref0, var, avg); + CALC_MSE_AVG_B(src1, ref1, var, avg); + } + + vec = __msa_hadd_s_w(avg, avg); + *diff = HADD_SW_S32(vec); + + return HADD_SW_S32(var); +} + +static uint32_t avg_sse_diff_16width_msa(const uint8_t *src_ptr, + int32_t src_stride, + const uint8_t *ref_ptr, + int32_t ref_stride, + const uint8_t *sec_pred, + int32_t height, int32_t *diff) { + int32_t ht_cnt; + v16u8 src, ref, pred; + v8i16 avg = { 0 }; + v4i32 vec, var = { 0 }; + + for (ht_cnt = (height >> 2); ht_cnt--;) { + pred = LD_UB(sec_pred); + sec_pred += 16; + src = LD_UB(src_ptr); + src_ptr += src_stride; + ref = LD_UB(ref_ptr); + ref_ptr += ref_stride; + src = __msa_aver_u_b(src, pred); + CALC_MSE_AVG_B(src, ref, var, avg); + + pred = LD_UB(sec_pred); + sec_pred += 16; + src = LD_UB(src_ptr); + src_ptr += src_stride; + ref = LD_UB(ref_ptr); + ref_ptr += ref_stride; + src = __msa_aver_u_b(src, pred); + CALC_MSE_AVG_B(src, ref, var, avg); + + pred = LD_UB(sec_pred); + sec_pred += 16; + src = LD_UB(src_ptr); + src_ptr += src_stride; + ref = LD_UB(ref_ptr); + ref_ptr += ref_stride; + src = __msa_aver_u_b(src, pred); + CALC_MSE_AVG_B(src, ref, var, avg); + + pred = LD_UB(sec_pred); + sec_pred += 16; + src = LD_UB(src_ptr); + src_ptr += src_stride; + ref = LD_UB(ref_ptr); + ref_ptr += ref_stride; + src = __msa_aver_u_b(src, pred); + CALC_MSE_AVG_B(src, ref, var, avg); + } + + vec = __msa_hadd_s_w(avg, avg); + *diff = HADD_SW_S32(vec); + + return HADD_SW_S32(var); +} + +static uint32_t avg_sse_diff_32width_msa(const uint8_t *src_ptr, + int32_t src_stride, + const uint8_t *ref_ptr, + int32_t ref_stride, + const uint8_t *sec_pred, + int32_t height, int32_t *diff) { + int32_t ht_cnt; + v16u8 src0, src1, ref0, ref1, pred0, pred1; + v8i16 avg = { 0 }; + v4i32 vec, var = { 0 }; + + for (ht_cnt = (height >> 2); ht_cnt--;) { + LD_UB2(sec_pred, 16, pred0, pred1); + sec_pred += 32; + LD_UB2(src_ptr, 16, src0, src1); + src_ptr += src_stride; + LD_UB2(ref_ptr, 16, ref0, ref1); + ref_ptr += ref_stride; + AVER_UB2_UB(src0, pred0, src1, pred1, src0, src1); + CALC_MSE_AVG_B(src0, ref0, var, avg); + CALC_MSE_AVG_B(src1, ref1, var, avg); + + LD_UB2(sec_pred, 16, pred0, pred1); + sec_pred += 32; + LD_UB2(src_ptr, 16, src0, src1); + src_ptr += src_stride; + LD_UB2(ref_ptr, 16, ref0, ref1); + ref_ptr += ref_stride; + AVER_UB2_UB(src0, pred0, src1, pred1, src0, src1); + CALC_MSE_AVG_B(src0, ref0, var, avg); + CALC_MSE_AVG_B(src1, ref1, var, avg); + + LD_UB2(sec_pred, 16, pred0, pred1); + sec_pred += 32; + LD_UB2(src_ptr, 16, src0, src1); + src_ptr += src_stride; + LD_UB2(ref_ptr, 16, ref0, ref1); + ref_ptr += ref_stride; + AVER_UB2_UB(src0, pred0, src1, pred1, src0, src1); + CALC_MSE_AVG_B(src0, ref0, var, avg); + CALC_MSE_AVG_B(src1, ref1, var, avg); + + LD_UB2(sec_pred, 16, pred0, pred1); + sec_pred += 32; + LD_UB2(src_ptr, 16, src0, src1); + src_ptr += src_stride; + LD_UB2(ref_ptr, 16, ref0, ref1); + ref_ptr += ref_stride; + AVER_UB2_UB(src0, pred0, src1, pred1, src0, src1); + CALC_MSE_AVG_B(src0, ref0, var, avg); + CALC_MSE_AVG_B(src1, ref1, var, avg); + } + + vec = __msa_hadd_s_w(avg, avg); + *diff = HADD_SW_S32(vec); + + return HADD_SW_S32(var); +} + +static uint32_t avg_sse_diff_32x64_msa(const uint8_t *src_ptr, + int32_t src_stride, + const uint8_t *ref_ptr, + int32_t ref_stride, + const uint8_t *sec_pred, int32_t *diff) { + int32_t ht_cnt; + v16u8 src0, src1, ref0, ref1, pred0, pred1; + v8i16 avg0 = { 0 }; + v8i16 avg1 = { 0 }; + v4i32 vec, var = { 0 }; + + for (ht_cnt = 16; ht_cnt--;) { + LD_UB2(sec_pred, 16, pred0, pred1); + sec_pred += 32; + LD_UB2(src_ptr, 16, src0, src1); + src_ptr += src_stride; + LD_UB2(ref_ptr, 16, ref0, ref1); + ref_ptr += ref_stride; + AVER_UB2_UB(src0, pred0, src1, pred1, src0, src1); + CALC_MSE_AVG_B(src0, ref0, var, avg0); + CALC_MSE_AVG_B(src1, ref1, var, avg1); + + LD_UB2(sec_pred, 16, pred0, pred1); + sec_pred += 32; + LD_UB2(src_ptr, 16, src0, src1); + src_ptr += src_stride; + LD_UB2(ref_ptr, 16, ref0, ref1); + ref_ptr += ref_stride; + AVER_UB2_UB(src0, pred0, src1, pred1, src0, src1); + CALC_MSE_AVG_B(src0, ref0, var, avg0); + CALC_MSE_AVG_B(src1, ref1, var, avg1); + + LD_UB2(sec_pred, 16, pred0, pred1); + sec_pred += 32; + LD_UB2(src_ptr, 16, src0, src1); + src_ptr += src_stride; + LD_UB2(ref_ptr, 16, ref0, ref1); + ref_ptr += ref_stride; + AVER_UB2_UB(src0, pred0, src1, pred1, src0, src1); + CALC_MSE_AVG_B(src0, ref0, var, avg0); + CALC_MSE_AVG_B(src1, ref1, var, avg1); + + LD_UB2(sec_pred, 16, pred0, pred1); + sec_pred += 32; + LD_UB2(src_ptr, 16, src0, src1); + src_ptr += src_stride; + LD_UB2(ref_ptr, 16, ref0, ref1); + ref_ptr += ref_stride; + AVER_UB2_UB(src0, pred0, src1, pred1, src0, src1); + CALC_MSE_AVG_B(src0, ref0, var, avg0); + CALC_MSE_AVG_B(src1, ref1, var, avg1); + } + + vec = __msa_hadd_s_w(avg0, avg0); + vec += __msa_hadd_s_w(avg1, avg1); + *diff = HADD_SW_S32(vec); + + return HADD_SW_S32(var); +} + +static uint32_t avg_sse_diff_64x32_msa(const uint8_t *src_ptr, + int32_t src_stride, + const uint8_t *ref_ptr, + int32_t ref_stride, + const uint8_t *sec_pred, int32_t *diff) { + int32_t ht_cnt; + v16u8 src0, src1, src2, src3; + v16u8 ref0, ref1, ref2, ref3; + v16u8 pred0, pred1, pred2, pred3; + v8i16 avg0 = { 0 }; + v8i16 avg1 = { 0 }; + v4i32 vec, var = { 0 }; + + for (ht_cnt = 16; ht_cnt--;) { + LD_UB4(sec_pred, 16, pred0, pred1, pred2, pred3); + sec_pred += 64; + LD_UB4(src_ptr, 16, src0, src1, src2, src3); + src_ptr += src_stride; + LD_UB4(ref_ptr, 16, ref0, ref1, ref2, ref3); + ref_ptr += ref_stride; + AVER_UB4_UB(src0, pred0, src1, pred1, src2, pred2, src3, pred3, src0, src1, + src2, src3); + CALC_MSE_AVG_B(src0, ref0, var, avg0); + CALC_MSE_AVG_B(src2, ref2, var, avg0); + CALC_MSE_AVG_B(src1, ref1, var, avg1); + CALC_MSE_AVG_B(src3, ref3, var, avg1); + + LD_UB4(sec_pred, 16, pred0, pred1, pred2, pred3); + sec_pred += 64; + LD_UB4(src_ptr, 16, src0, src1, src2, src3); + src_ptr += src_stride; + LD_UB4(ref_ptr, 16, ref0, ref1, ref2, ref3); + ref_ptr += ref_stride; + AVER_UB4_UB(src0, pred0, src1, pred1, src2, pred2, src3, pred3, src0, src1, + src2, src3); + CALC_MSE_AVG_B(src0, ref0, var, avg0); + CALC_MSE_AVG_B(src2, ref2, var, avg0); + CALC_MSE_AVG_B(src1, ref1, var, avg1); + CALC_MSE_AVG_B(src3, ref3, var, avg1); + } + + vec = __msa_hadd_s_w(avg0, avg0); + vec += __msa_hadd_s_w(avg1, avg1); + + *diff = HADD_SW_S32(vec); + + return HADD_SW_S32(var); +} + +static uint32_t avg_sse_diff_64x64_msa(const uint8_t *src_ptr, + int32_t src_stride, + const uint8_t *ref_ptr, + int32_t ref_stride, + const uint8_t *sec_pred, int32_t *diff) { + int32_t ht_cnt; + v16u8 src0, src1, src2, src3; + v16u8 ref0, ref1, ref2, ref3; + v16u8 pred0, pred1, pred2, pred3; + v8i16 avg0 = { 0 }; + v8i16 avg1 = { 0 }; + v8i16 avg2 = { 0 }; + v8i16 avg3 = { 0 }; + v4i32 vec, var = { 0 }; + + for (ht_cnt = 32; ht_cnt--;) { + LD_UB4(sec_pred, 16, pred0, pred1, pred2, pred3); + sec_pred += 64; + LD_UB4(src_ptr, 16, src0, src1, src2, src3); + src_ptr += src_stride; + LD_UB4(ref_ptr, 16, ref0, ref1, ref2, ref3); + ref_ptr += ref_stride; + AVER_UB4_UB(src0, pred0, src1, pred1, src2, pred2, src3, pred3, src0, src1, + src2, src3); + CALC_MSE_AVG_B(src0, ref0, var, avg0); + CALC_MSE_AVG_B(src1, ref1, var, avg1); + CALC_MSE_AVG_B(src2, ref2, var, avg2); + CALC_MSE_AVG_B(src3, ref3, var, avg3); + + LD_UB4(sec_pred, 16, pred0, pred1, pred2, pred3); + sec_pred += 64; + LD_UB4(src_ptr, 16, src0, src1, src2, src3); + src_ptr += src_stride; + LD_UB4(ref_ptr, 16, ref0, ref1, ref2, ref3); + ref_ptr += ref_stride; + AVER_UB4_UB(src0, pred0, src1, pred1, src2, pred2, src3, pred3, src0, src1, + src2, src3); + CALC_MSE_AVG_B(src0, ref0, var, avg0); + CALC_MSE_AVG_B(src1, ref1, var, avg1); + CALC_MSE_AVG_B(src2, ref2, var, avg2); + CALC_MSE_AVG_B(src3, ref3, var, avg3); + } + + vec = __msa_hadd_s_w(avg0, avg0); + vec += __msa_hadd_s_w(avg1, avg1); + vec += __msa_hadd_s_w(avg2, avg2); + vec += __msa_hadd_s_w(avg3, avg3); + *diff = HADD_SW_S32(vec); + + return HADD_SW_S32(var); +} + +static uint32_t sub_pixel_sse_diff_4width_h_msa( + const uint8_t *src, int32_t src_stride, const uint8_t *dst, + int32_t dst_stride, const uint8_t *filter, int32_t height, int32_t *diff) { + int16_t filtval; + uint32_t loop_cnt; + uint32_t ref0, ref1, ref2, ref3; + v16u8 filt0, ref = { 0 }; + v16i8 src0, src1, src2, src3; + v16i8 mask = { 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8 }; + v8u16 vec0, vec1, vec2, vec3; + v8i16 avg = { 0 }; + v4i32 vec, var = { 0 }; + + filtval = LH(filter); + filt0 = (v16u8)__msa_fill_h(filtval); + + for (loop_cnt = (height >> 2); loop_cnt--;) { + LD_SB4(src, src_stride, src0, src1, src2, src3); + src += (4 * src_stride); + LW4(dst, dst_stride, ref0, ref1, ref2, ref3); + dst += (4 * dst_stride); + INSERT_W4_UB(ref0, ref1, ref2, ref3, ref); + VSHF_B2_UH(src0, src0, src1, src1, mask, mask, vec0, vec1); + VSHF_B2_UH(src2, src2, src3, src3, mask, mask, vec2, vec3); + DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, vec0, vec1, + vec2, vec3); + SRARI_H4_UH(vec0, vec1, vec2, vec3, FILTER_BITS); + PCKEV_B4_SB(vec0, vec0, vec1, vec1, vec2, vec2, vec3, vec3, src0, src1, + src2, src3); + ILVEV_W2_SB(src0, src1, src2, src3, src0, src2); + src0 = (v16i8)__msa_ilvev_d((v2i64)src2, (v2i64)src0); + CALC_MSE_AVG_B(src0, ref, var, avg); + } + + vec = __msa_hadd_s_w(avg, avg); + *diff = HADD_SW_S32(vec); + + return HADD_SW_S32(var); +} + +static uint32_t sub_pixel_sse_diff_8width_h_msa( + const uint8_t *src, int32_t src_stride, const uint8_t *dst, + int32_t dst_stride, const uint8_t *filter, int32_t height, int32_t *diff) { + int16_t filtval; + uint32_t loop_cnt; + v16u8 filt0, out, ref0, ref1, ref2, ref3; + v16i8 src0, src1, src2, src3; + v16i8 mask = { 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8 }; + v8u16 vec0, vec1, vec2, vec3; + v8i16 avg = { 0 }; + v4i32 vec, var = { 0 }; + + filtval = LH(filter); + filt0 = (v16u8)__msa_fill_h(filtval); + + for (loop_cnt = (height >> 2); loop_cnt--;) { + LD_SB4(src, src_stride, src0, src1, src2, src3); + src += (4 * src_stride); + LD_UB4(dst, dst_stride, ref0, ref1, ref2, ref3); + dst += (4 * dst_stride); + + PCKEV_D2_UB(ref1, ref0, ref3, ref2, ref0, ref1); + VSHF_B2_UH(src0, src0, src1, src1, mask, mask, vec0, vec1); + VSHF_B2_UH(src2, src2, src3, src3, mask, mask, vec2, vec3); + DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, vec0, vec1, + vec2, vec3); + SRARI_H4_UH(vec0, vec1, vec2, vec3, FILTER_BITS); + PCKEV_B4_SB(vec0, vec0, vec1, vec1, vec2, vec2, vec3, vec3, src0, src1, + src2, src3); + out = (v16u8)__msa_ilvev_d((v2i64)src1, (v2i64)src0); + CALC_MSE_AVG_B(out, ref0, var, avg); + out = (v16u8)__msa_ilvev_d((v2i64)src3, (v2i64)src2); + CALC_MSE_AVG_B(out, ref1, var, avg); + } + + vec = __msa_hadd_s_w(avg, avg); + *diff = HADD_SW_S32(vec); + + return HADD_SW_S32(var); +} + +static uint32_t sub_pixel_sse_diff_16width_h_msa( + const uint8_t *src, int32_t src_stride, const uint8_t *dst, + int32_t dst_stride, const uint8_t *filter, int32_t height, int32_t *diff) { + int16_t filtval; + uint32_t loop_cnt; + v16i8 src0, src1, src2, src3, src4, src5, src6, src7; + v16i8 mask = { 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8 }; + v16u8 dst0, dst1, dst2, dst3, filt0; + v8u16 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7; + v8u16 out0, out1, out2, out3, out4, out5, out6, out7; + v8i16 avg = { 0 }; + v4i32 vec, var = { 0 }; + + filtval = LH(filter); + filt0 = (v16u8)__msa_fill_h(filtval); + + for (loop_cnt = (height >> 2); loop_cnt--;) { + LD_SB4(src, src_stride, src0, src2, src4, src6); + LD_SB4(src + 8, src_stride, src1, src3, src5, src7); + src += (4 * src_stride); + LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3); + dst += (4 * dst_stride); + + VSHF_B2_UH(src0, src0, src1, src1, mask, mask, vec0, vec1); + VSHF_B2_UH(src2, src2, src3, src3, mask, mask, vec2, vec3); + VSHF_B2_UH(src4, src4, src5, src5, mask, mask, vec4, vec5); + VSHF_B2_UH(src6, src6, src7, src7, mask, mask, vec6, vec7); + DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, out0, out1, + out2, out3); + DOTP_UB4_UH(vec4, vec5, vec6, vec7, filt0, filt0, filt0, filt0, out4, out5, + out6, out7); + SRARI_H4_UH(out0, out1, out2, out3, FILTER_BITS); + SRARI_H4_UH(out4, out5, out6, out7, FILTER_BITS); + PCKEV_B4_SB(out1, out0, out3, out2, out5, out4, out7, out6, src0, src1, + src2, src3); + CALC_MSE_AVG_B(src0, dst0, var, avg); + CALC_MSE_AVG_B(src1, dst1, var, avg); + CALC_MSE_AVG_B(src2, dst2, var, avg); + CALC_MSE_AVG_B(src3, dst3, var, avg); + } + + vec = __msa_hadd_s_w(avg, avg); + *diff = HADD_SW_S32(vec); + + return HADD_SW_S32(var); +} + +static uint32_t sub_pixel_sse_diff_32width_h_msa( + const uint8_t *src, int32_t src_stride, const uint8_t *dst, + int32_t dst_stride, const uint8_t *filter, int32_t height, int32_t *diff) { + uint32_t loop_cnt, sse = 0; + int32_t diff0[2]; + + for (loop_cnt = 0; loop_cnt < 2; ++loop_cnt) { + sse += sub_pixel_sse_diff_16width_h_msa(src, src_stride, dst, dst_stride, + filter, height, &diff0[loop_cnt]); + src += 16; + dst += 16; + } + + *diff = diff0[0] + diff0[1]; + + return sse; +} + +static uint32_t sub_pixel_sse_diff_64width_h_msa( + const uint8_t *src, int32_t src_stride, const uint8_t *dst, + int32_t dst_stride, const uint8_t *filter, int32_t height, int32_t *diff) { + uint32_t loop_cnt, sse = 0; + int32_t diff0[4]; + + for (loop_cnt = 0; loop_cnt < 4; ++loop_cnt) { + sse += sub_pixel_sse_diff_16width_h_msa(src, src_stride, dst, dst_stride, + filter, height, &diff0[loop_cnt]); + src += 16; + dst += 16; + } + + *diff = diff0[0] + diff0[1] + diff0[2] + diff0[3]; + + return sse; +} + +static uint32_t sub_pixel_sse_diff_4width_v_msa( + const uint8_t *src, int32_t src_stride, const uint8_t *dst, + int32_t dst_stride, const uint8_t *filter, int32_t height, int32_t *diff) { + int16_t filtval; + uint32_t loop_cnt; + uint32_t ref0, ref1, ref2, ref3; + v16u8 src0, src1, src2, src3, src4, out; + v16u8 src10_r, src32_r, src21_r, src43_r; + v16u8 ref = { 0 }; + v16u8 src2110, src4332; + v16u8 filt0; + v8i16 avg = { 0 }; + v4i32 vec, var = { 0 }; + v8u16 tmp0, tmp1; + + filtval = LH(filter); + filt0 = (v16u8)__msa_fill_h(filtval); + + src0 = LD_UB(src); + src += src_stride; + + for (loop_cnt = (height >> 2); loop_cnt--;) { + LD_UB4(src, src_stride, src1, src2, src3, src4); + src += (4 * src_stride); + LW4(dst, dst_stride, ref0, ref1, ref2, ref3); + dst += (4 * dst_stride); + + INSERT_W4_UB(ref0, ref1, ref2, ref3, ref); + ILVR_B4_UB(src1, src0, src2, src1, src3, src2, src4, src3, src10_r, src21_r, + src32_r, src43_r); + ILVR_D2_UB(src21_r, src10_r, src43_r, src32_r, src2110, src4332); + DOTP_UB2_UH(src2110, src4332, filt0, filt0, tmp0, tmp1); + SRARI_H2_UH(tmp0, tmp1, FILTER_BITS); + out = (v16u8)__msa_pckev_b((v16i8)tmp1, (v16i8)tmp0); + CALC_MSE_AVG_B(out, ref, var, avg); + src0 = src4; + } + + vec = __msa_hadd_s_w(avg, avg); + *diff = HADD_SW_S32(vec); + + return HADD_SW_S32(var); +} + +static uint32_t sub_pixel_sse_diff_8width_v_msa( + const uint8_t *src, int32_t src_stride, const uint8_t *dst, + int32_t dst_stride, const uint8_t *filter, int32_t height, int32_t *diff) { + int16_t filtval; + uint32_t loop_cnt; + v16u8 src0, src1, src2, src3, src4; + v16u8 ref0, ref1, ref2, ref3; + v8u16 vec0, vec1, vec2, vec3; + v8u16 tmp0, tmp1, tmp2, tmp3; + v16u8 filt0; + v8i16 avg = { 0 }; + v4i32 vec, var = { 0 }; + + filtval = LH(filter); + filt0 = (v16u8)__msa_fill_h(filtval); + + src0 = LD_UB(src); + src += src_stride; + + for (loop_cnt = (height >> 2); loop_cnt--;) { + LD_UB4(src, src_stride, src1, src2, src3, src4); + src += (4 * src_stride); + LD_UB4(dst, dst_stride, ref0, ref1, ref2, ref3); + dst += (4 * dst_stride); + + PCKEV_D2_UB(ref1, ref0, ref3, ref2, ref0, ref1); + ILVR_B4_UH(src1, src0, src2, src1, src3, src2, src4, src3, vec0, vec1, vec2, + vec3); + DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, tmp0, tmp1, + tmp2, tmp3); + SRARI_H4_UH(tmp0, tmp1, tmp2, tmp3, FILTER_BITS); + PCKEV_B2_UB(tmp1, tmp0, tmp3, tmp2, src0, src1); + CALC_MSE_AVG_B(src0, ref0, var, avg); + CALC_MSE_AVG_B(src1, ref1, var, avg); + src0 = src4; + } + + vec = __msa_hadd_s_w(avg, avg); + *diff = HADD_SW_S32(vec); + + return HADD_SW_S32(var); +} + +static uint32_t sub_pixel_sse_diff_16width_v_msa( + const uint8_t *src, int32_t src_stride, const uint8_t *dst, + int32_t dst_stride, const uint8_t *filter, int32_t height, int32_t *diff) { + int16_t filtval; + uint32_t loop_cnt; + v16u8 ref0, ref1, ref2, ref3; + v16u8 src0, src1, src2, src3, src4; + v16u8 out0, out1, out2, out3; + v16u8 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7; + v8u16 tmp0, tmp1, tmp2, tmp3; + v16u8 filt0; + v8i16 avg = { 0 }; + v4i32 vec, var = { 0 }; + + filtval = LH(filter); + filt0 = (v16u8)__msa_fill_h(filtval); + + src0 = LD_UB(src); + src += src_stride; + + for (loop_cnt = (height >> 2); loop_cnt--;) { + LD_UB4(src, src_stride, src1, src2, src3, src4); + src += (4 * src_stride); + LD_UB4(dst, dst_stride, ref0, ref1, ref2, ref3); + dst += (4 * dst_stride); + + ILVR_B2_UB(src1, src0, src2, src1, vec0, vec2); + ILVL_B2_UB(src1, src0, src2, src1, vec1, vec3); + DOTP_UB2_UH(vec0, vec1, filt0, filt0, tmp0, tmp1); + SRARI_H2_UH(tmp0, tmp1, FILTER_BITS); + out0 = (v16u8)__msa_pckev_b((v16i8)tmp1, (v16i8)tmp0); + + ILVR_B2_UB(src3, src2, src4, src3, vec4, vec6); + ILVL_B2_UB(src3, src2, src4, src3, vec5, vec7); + DOTP_UB2_UH(vec2, vec3, filt0, filt0, tmp2, tmp3); + SRARI_H2_UH(tmp2, tmp3, FILTER_BITS); + out1 = (v16u8)__msa_pckev_b((v16i8)tmp3, (v16i8)tmp2); + + DOTP_UB2_UH(vec4, vec5, filt0, filt0, tmp0, tmp1); + SRARI_H2_UH(tmp0, tmp1, FILTER_BITS); + out2 = (v16u8)__msa_pckev_b((v16i8)tmp1, (v16i8)tmp0); + DOTP_UB2_UH(vec6, vec7, filt0, filt0, tmp2, tmp3); + SRARI_H2_UH(tmp2, tmp3, FILTER_BITS); + out3 = (v16u8)__msa_pckev_b((v16i8)tmp3, (v16i8)tmp2); + + src0 = src4; + + CALC_MSE_AVG_B(out0, ref0, var, avg); + CALC_MSE_AVG_B(out1, ref1, var, avg); + CALC_MSE_AVG_B(out2, ref2, var, avg); + CALC_MSE_AVG_B(out3, ref3, var, avg); + } + + vec = __msa_hadd_s_w(avg, avg); + *diff = HADD_SW_S32(vec); + + return HADD_SW_S32(var); +} + +static uint32_t sub_pixel_sse_diff_32width_v_msa( + const uint8_t *src, int32_t src_stride, const uint8_t *dst, + int32_t dst_stride, const uint8_t *filter, int32_t height, int32_t *diff) { + uint32_t loop_cnt, sse = 0; + int32_t diff0[2]; + + for (loop_cnt = 0; loop_cnt < 2; ++loop_cnt) { + sse += sub_pixel_sse_diff_16width_v_msa(src, src_stride, dst, dst_stride, + filter, height, &diff0[loop_cnt]); + src += 16; + dst += 16; + } + + *diff = diff0[0] + diff0[1]; + + return sse; +} + +static uint32_t sub_pixel_sse_diff_64width_v_msa( + const uint8_t *src, int32_t src_stride, const uint8_t *dst, + int32_t dst_stride, const uint8_t *filter, int32_t height, int32_t *diff) { + uint32_t loop_cnt, sse = 0; + int32_t diff0[4]; + + for (loop_cnt = 0; loop_cnt < 4; ++loop_cnt) { + sse += sub_pixel_sse_diff_16width_v_msa(src, src_stride, dst, dst_stride, + filter, height, &diff0[loop_cnt]); + src += 16; + dst += 16; + } + + *diff = diff0[0] + diff0[1] + diff0[2] + diff0[3]; + + return sse; +} + +static uint32_t sub_pixel_sse_diff_4width_hv_msa( + const uint8_t *src, int32_t src_stride, const uint8_t *dst, + int32_t dst_stride, const uint8_t *filter_horiz, const uint8_t *filter_vert, + int32_t height, int32_t *diff) { + int16_t filtval; + uint32_t loop_cnt; + uint32_t ref0, ref1, ref2, ref3; + v16u8 src0, src1, src2, src3, src4; + v16u8 out, ref = { 0 }; + v16u8 filt_vt, filt_hz, vec0, vec1; + v16u8 mask = { 0, 1, 1, 2, 2, 3, 3, 4, 16, 17, 17, 18, 18, 19, 19, 20 }; + v8u16 hz_out0, hz_out1, hz_out2, hz_out3, hz_out4; + v8u16 tmp0, tmp1; + v8i16 avg = { 0 }; + v4i32 vec, var = { 0 }; + + filtval = LH(filter_horiz); + filt_hz = (v16u8)__msa_fill_h(filtval); + filtval = LH(filter_vert); + filt_vt = (v16u8)__msa_fill_h(filtval); + + src0 = LD_UB(src); + src += src_stride; + + for (loop_cnt = (height >> 2); loop_cnt--;) { + LD_UB4(src, src_stride, src1, src2, src3, src4); + src += (4 * src_stride); + LW4(dst, dst_stride, ref0, ref1, ref2, ref3); + dst += (4 * dst_stride); + INSERT_W4_UB(ref0, ref1, ref2, ref3, ref); + hz_out0 = HORIZ_2TAP_FILT_UH(src0, src1, mask, filt_hz, FILTER_BITS); + hz_out2 = HORIZ_2TAP_FILT_UH(src2, src3, mask, filt_hz, FILTER_BITS); + hz_out4 = HORIZ_2TAP_FILT_UH(src4, src4, mask, filt_hz, FILTER_BITS); + hz_out1 = (v8u16)__msa_sldi_b((v16i8)hz_out2, (v16i8)hz_out0, 8); + hz_out3 = (v8u16)__msa_pckod_d((v2i64)hz_out4, (v2i64)hz_out2); + ILVEV_B2_UB(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1); + DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp0, tmp1); + SRARI_H2_UH(tmp0, tmp1, FILTER_BITS); + out = (v16u8)__msa_pckev_b((v16i8)tmp1, (v16i8)tmp0); + CALC_MSE_AVG_B(out, ref, var, avg); + src0 = src4; + } + + vec = __msa_hadd_s_w(avg, avg); + *diff = HADD_SW_S32(vec); + + return HADD_SW_S32(var); +} + +static uint32_t sub_pixel_sse_diff_8width_hv_msa( + const uint8_t *src, int32_t src_stride, const uint8_t *dst, + int32_t dst_stride, const uint8_t *filter_horiz, const uint8_t *filter_vert, + int32_t height, int32_t *diff) { + int16_t filtval; + uint32_t loop_cnt; + v16u8 ref0, ref1, ref2, ref3; + v16u8 src0, src1, src2, src3, src4; + v16u8 out0, out1; + v16u8 mask = { 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8 }; + v8u16 hz_out0, hz_out1; + v8u16 tmp0, tmp1, tmp2, tmp3; + v16u8 filt_vt, filt_hz, vec0; + v8i16 avg = { 0 }; + v4i32 vec, var = { 0 }; + + filtval = LH(filter_horiz); + filt_hz = (v16u8)__msa_fill_h(filtval); + filtval = LH(filter_vert); + filt_vt = (v16u8)__msa_fill_h(filtval); + + src0 = LD_UB(src); + src += src_stride; + hz_out0 = HORIZ_2TAP_FILT_UH(src0, src0, mask, filt_hz, FILTER_BITS); + + for (loop_cnt = (height >> 2); loop_cnt--;) { + LD_UB4(src, src_stride, src1, src2, src3, src4); + src += (4 * src_stride); + LD_UB4(dst, dst_stride, ref0, ref1, ref2, ref3); + dst += (4 * dst_stride); + + PCKEV_D2_UB(ref1, ref0, ref3, ref2, ref0, ref1); + hz_out1 = HORIZ_2TAP_FILT_UH(src1, src1, mask, filt_hz, FILTER_BITS); + vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out1, (v16i8)hz_out0); + tmp0 = __msa_dotp_u_h(vec0, filt_vt); + hz_out0 = HORIZ_2TAP_FILT_UH(src2, src2, mask, filt_hz, FILTER_BITS); + vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out0, (v16i8)hz_out1); + tmp1 = __msa_dotp_u_h(vec0, filt_vt); + SRARI_H2_UH(tmp0, tmp1, FILTER_BITS); + hz_out1 = HORIZ_2TAP_FILT_UH(src3, src3, mask, filt_hz, FILTER_BITS); + vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out1, (v16i8)hz_out0); + tmp2 = __msa_dotp_u_h(vec0, filt_vt); + hz_out0 = HORIZ_2TAP_FILT_UH(src4, src4, mask, filt_hz, FILTER_BITS); + vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out0, (v16i8)hz_out1); + tmp3 = __msa_dotp_u_h(vec0, filt_vt); + SRARI_H2_UH(tmp2, tmp3, FILTER_BITS); + PCKEV_B2_UB(tmp1, tmp0, tmp3, tmp2, out0, out1); + CALC_MSE_AVG_B(out0, ref0, var, avg); + CALC_MSE_AVG_B(out1, ref1, var, avg); + } + + vec = __msa_hadd_s_w(avg, avg); + *diff = HADD_SW_S32(vec); + + return HADD_SW_S32(var); +} + +static uint32_t sub_pixel_sse_diff_16width_hv_msa( + const uint8_t *src, int32_t src_stride, const uint8_t *dst, + int32_t dst_stride, const uint8_t *filter_horiz, const uint8_t *filter_vert, + int32_t height, int32_t *diff) { + int16_t filtval; + uint32_t loop_cnt; + v16u8 src0, src1, src2, src3, src4, src5, src6, src7; + v16u8 ref0, ref1, ref2, ref3; + v16u8 filt_hz, filt_vt, vec0, vec1; + v16u8 mask = { 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8 }; + v8u16 hz_out0, hz_out1, hz_out2, hz_out3; + v8u16 tmp0, tmp1; + v8i16 avg = { 0 }; + v4i32 vec, var = { 0 }; + + filtval = LH(filter_horiz); + filt_hz = (v16u8)__msa_fill_h(filtval); + filtval = LH(filter_vert); + filt_vt = (v16u8)__msa_fill_h(filtval); + + LD_UB2(src, 8, src0, src1); + src += src_stride; + + hz_out0 = HORIZ_2TAP_FILT_UH(src0, src0, mask, filt_hz, FILTER_BITS); + hz_out2 = HORIZ_2TAP_FILT_UH(src1, src1, mask, filt_hz, FILTER_BITS); + + for (loop_cnt = (height >> 2); loop_cnt--;) { + LD_UB4(src, src_stride, src0, src2, src4, src6); + LD_UB4(src + 8, src_stride, src1, src3, src5, src7); + src += (4 * src_stride); + LD_UB4(dst, dst_stride, ref0, ref1, ref2, ref3); + dst += (4 * dst_stride); + + hz_out1 = HORIZ_2TAP_FILT_UH(src0, src0, mask, filt_hz, FILTER_BITS); + hz_out3 = HORIZ_2TAP_FILT_UH(src1, src1, mask, filt_hz, FILTER_BITS); + ILVEV_B2_UB(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1); + DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp0, tmp1); + SRARI_H2_UH(tmp0, tmp1, FILTER_BITS); + src0 = (v16u8)__msa_pckev_b((v16i8)tmp1, (v16i8)tmp0); + + hz_out0 = HORIZ_2TAP_FILT_UH(src2, src2, mask, filt_hz, FILTER_BITS); + hz_out2 = HORIZ_2TAP_FILT_UH(src3, src3, mask, filt_hz, FILTER_BITS); + ILVEV_B2_UB(hz_out1, hz_out0, hz_out3, hz_out2, vec0, vec1); + DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp0, tmp1); + SRARI_H2_UH(tmp0, tmp1, FILTER_BITS); + src1 = (v16u8)__msa_pckev_b((v16i8)tmp1, (v16i8)tmp0); + + hz_out1 = HORIZ_2TAP_FILT_UH(src4, src4, mask, filt_hz, FILTER_BITS); + hz_out3 = HORIZ_2TAP_FILT_UH(src5, src5, mask, filt_hz, FILTER_BITS); + ILVEV_B2_UB(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1); + DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp0, tmp1); + SRARI_H2_UH(tmp0, tmp1, FILTER_BITS); + src2 = (v16u8)__msa_pckev_b((v16i8)tmp1, (v16i8)tmp0); + + hz_out0 = HORIZ_2TAP_FILT_UH(src6, src6, mask, filt_hz, FILTER_BITS); + hz_out2 = HORIZ_2TAP_FILT_UH(src7, src7, mask, filt_hz, FILTER_BITS); + ILVEV_B2_UB(hz_out1, hz_out0, hz_out3, hz_out2, vec0, vec1); + DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp0, tmp1); + SRARI_H2_UH(tmp0, tmp1, FILTER_BITS); + src3 = (v16u8)__msa_pckev_b((v16i8)tmp1, (v16i8)tmp0); + + CALC_MSE_AVG_B(src0, ref0, var, avg); + CALC_MSE_AVG_B(src1, ref1, var, avg); + CALC_MSE_AVG_B(src2, ref2, var, avg); + CALC_MSE_AVG_B(src3, ref3, var, avg); + } + + vec = __msa_hadd_s_w(avg, avg); + *diff = HADD_SW_S32(vec); + + return HADD_SW_S32(var); +} + +static uint32_t sub_pixel_sse_diff_32width_hv_msa( + const uint8_t *src, int32_t src_stride, const uint8_t *dst, + int32_t dst_stride, const uint8_t *filter_horiz, const uint8_t *filter_vert, + int32_t height, int32_t *diff) { + uint32_t loop_cnt, sse = 0; + int32_t diff0[2]; + + for (loop_cnt = 0; loop_cnt < 2; ++loop_cnt) { + sse += sub_pixel_sse_diff_16width_hv_msa(src, src_stride, dst, dst_stride, + filter_horiz, filter_vert, height, + &diff0[loop_cnt]); + src += 16; + dst += 16; + } + + *diff = diff0[0] + diff0[1]; + + return sse; +} + +static uint32_t sub_pixel_sse_diff_64width_hv_msa( + const uint8_t *src, int32_t src_stride, const uint8_t *dst, + int32_t dst_stride, const uint8_t *filter_horiz, const uint8_t *filter_vert, + int32_t height, int32_t *diff) { + uint32_t loop_cnt, sse = 0; + int32_t diff0[4]; + + for (loop_cnt = 0; loop_cnt < 4; ++loop_cnt) { + sse += sub_pixel_sse_diff_16width_hv_msa(src, src_stride, dst, dst_stride, + filter_horiz, filter_vert, height, + &diff0[loop_cnt]); + src += 16; + dst += 16; + } + + *diff = diff0[0] + diff0[1] + diff0[2] + diff0[3]; + + return sse; +} + +static uint32_t sub_pixel_avg_sse_diff_4width_h_msa( + const uint8_t *src, int32_t src_stride, const uint8_t *dst, + int32_t dst_stride, const uint8_t *sec_pred, const uint8_t *filter, + int32_t height, int32_t *diff) { + int16_t filtval; + uint32_t loop_cnt; + uint32_t ref0, ref1, ref2, ref3; + v16u8 out, pred, filt0, ref = { 0 }; + v16i8 src0, src1, src2, src3; + v16i8 mask = { 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8 }; + v8u16 vec0, vec1, vec2, vec3; + v8i16 avg = { 0 }; + v4i32 vec, var = { 0 }; + + filtval = LH(filter); + filt0 = (v16u8)__msa_fill_h(filtval); + + for (loop_cnt = (height >> 2); loop_cnt--;) { + LD_SB4(src, src_stride, src0, src1, src2, src3); + src += (4 * src_stride); + pred = LD_UB(sec_pred); + sec_pred += 16; + LW4(dst, dst_stride, ref0, ref1, ref2, ref3); + dst += (4 * dst_stride); + + INSERT_W4_UB(ref0, ref1, ref2, ref3, ref); + VSHF_B2_UH(src0, src0, src1, src1, mask, mask, vec0, vec1); + VSHF_B2_UH(src2, src2, src3, src3, mask, mask, vec2, vec3); + DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, vec0, vec1, + vec2, vec3); + SRARI_H4_UH(vec0, vec1, vec2, vec3, FILTER_BITS); + PCKEV_B4_SB(vec0, vec0, vec1, vec1, vec2, vec2, vec3, vec3, src0, src1, + src2, src3); + ILVEV_W2_SB(src0, src1, src2, src3, src0, src2); + out = (v16u8)__msa_ilvev_d((v2i64)src2, (v2i64)src0); + out = __msa_aver_u_b(out, pred); + CALC_MSE_AVG_B(out, ref, var, avg); + } + + vec = __msa_hadd_s_w(avg, avg); + *diff = HADD_SW_S32(vec); + + return HADD_SW_S32(var); +} + +static uint32_t sub_pixel_avg_sse_diff_8width_h_msa( + const uint8_t *src, int32_t src_stride, const uint8_t *dst, + int32_t dst_stride, const uint8_t *sec_pred, const uint8_t *filter, + int32_t height, int32_t *diff) { + int16_t filtval; + uint32_t loop_cnt; + v16u8 out, pred, filt0; + v16u8 ref0, ref1, ref2, ref3; + v16i8 src0, src1, src2, src3; + v16i8 mask = { 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8 }; + v8u16 vec0, vec1, vec2, vec3; + v8i16 avg = { 0 }; + v4i32 vec, var = { 0 }; + + filtval = LH(filter); + filt0 = (v16u8)__msa_fill_h(filtval); + + for (loop_cnt = (height >> 2); loop_cnt--;) { + LD_SB4(src, src_stride, src0, src1, src2, src3); + src += (4 * src_stride); + LD_UB4(dst, dst_stride, ref0, ref1, ref2, ref3); + dst += (4 * dst_stride); + + PCKEV_D2_UB(ref1, ref0, ref3, ref2, ref0, ref1); + VSHF_B2_UH(src0, src0, src1, src1, mask, mask, vec0, vec1); + VSHF_B2_UH(src2, src2, src3, src3, mask, mask, vec2, vec3); + DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, vec0, vec1, + vec2, vec3); + SRARI_H4_UH(vec0, vec1, vec2, vec3, FILTER_BITS); + PCKEV_B4_SB(vec0, vec0, vec1, vec1, vec2, vec2, vec3, vec3, src0, src1, + src2, src3); + out = (v16u8)__msa_ilvev_d((v2i64)src1, (v2i64)src0); + + pred = LD_UB(sec_pred); + sec_pred += 16; + out = __msa_aver_u_b(out, pred); + CALC_MSE_AVG_B(out, ref0, var, avg); + out = (v16u8)__msa_ilvev_d((v2i64)src3, (v2i64)src2); + pred = LD_UB(sec_pred); + sec_pred += 16; + out = __msa_aver_u_b(out, pred); + CALC_MSE_AVG_B(out, ref1, var, avg); + } + + vec = __msa_hadd_s_w(avg, avg); + *diff = HADD_SW_S32(vec); + + return HADD_SW_S32(var); +} + +static uint32_t subpel_avg_ssediff_16w_h_msa( + const uint8_t *src, int32_t src_stride, const uint8_t *dst, + int32_t dst_stride, const uint8_t *sec_pred, const uint8_t *filter, + int32_t height, int32_t *diff, int32_t width) { + int16_t filtval; + uint32_t loop_cnt; + v16i8 src0, src1, src2, src3, src4, src5, src6, src7; + v16i8 mask = { 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8 }; + v16u8 dst0, dst1, dst2, dst3; + v16u8 tmp0, tmp1, tmp2, tmp3; + v16u8 pred0, pred1, pred2, pred3, filt0; + v8u16 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7; + v8u16 out0, out1, out2, out3, out4, out5, out6, out7; + v8i16 avg = { 0 }; + v4i32 vec, var = { 0 }; + + filtval = LH(filter); + filt0 = (v16u8)__msa_fill_h(filtval); + + for (loop_cnt = (height >> 2); loop_cnt--;) { + LD_SB4(src, src_stride, src0, src2, src4, src6); + LD_SB4(src + 8, src_stride, src1, src3, src5, src7); + src += (4 * src_stride); + LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3); + dst += (4 * dst_stride); + LD_UB4(sec_pred, width, pred0, pred1, pred2, pred3); + sec_pred += (4 * width); + + VSHF_B2_UH(src0, src0, src1, src1, mask, mask, vec0, vec1); + VSHF_B2_UH(src2, src2, src3, src3, mask, mask, vec2, vec3); + VSHF_B2_UH(src4, src4, src5, src5, mask, mask, vec4, vec5); + VSHF_B2_UH(src6, src6, src7, src7, mask, mask, vec6, vec7); + DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, out0, out1, + out2, out3); + DOTP_UB4_UH(vec4, vec5, vec6, vec7, filt0, filt0, filt0, filt0, out4, out5, + out6, out7); + SRARI_H4_UH(out0, out1, out2, out3, FILTER_BITS); + SRARI_H4_UH(out4, out5, out6, out7, FILTER_BITS); + PCKEV_B4_UB(out1, out0, out3, out2, out5, out4, out7, out6, tmp0, tmp1, + tmp2, tmp3); + AVER_UB4_UB(tmp0, pred0, tmp1, pred1, tmp2, pred2, tmp3, pred3, tmp0, tmp1, + tmp2, tmp3); + + CALC_MSE_AVG_B(tmp0, dst0, var, avg); + CALC_MSE_AVG_B(tmp1, dst1, var, avg); + CALC_MSE_AVG_B(tmp2, dst2, var, avg); + CALC_MSE_AVG_B(tmp3, dst3, var, avg); + } + + vec = __msa_hadd_s_w(avg, avg); + *diff = HADD_SW_S32(vec); + + return HADD_SW_S32(var); +} + +static uint32_t sub_pixel_avg_sse_diff_16width_h_msa( + const uint8_t *src, int32_t src_stride, const uint8_t *dst, + int32_t dst_stride, const uint8_t *sec_pred, const uint8_t *filter, + int32_t height, int32_t *diff) { + return subpel_avg_ssediff_16w_h_msa(src, src_stride, dst, dst_stride, + sec_pred, filter, height, diff, 16); +} + +static uint32_t sub_pixel_avg_sse_diff_32width_h_msa( + const uint8_t *src, int32_t src_stride, const uint8_t *dst, + int32_t dst_stride, const uint8_t *sec_pred, const uint8_t *filter, + int32_t height, int32_t *diff) { + uint32_t loop_cnt, sse = 0; + int32_t diff0[2]; + + for (loop_cnt = 0; loop_cnt < 2; ++loop_cnt) { + sse += + subpel_avg_ssediff_16w_h_msa(src, src_stride, dst, dst_stride, sec_pred, + filter, height, &diff0[loop_cnt], 32); + src += 16; + dst += 16; + sec_pred += 16; + } + + *diff = diff0[0] + diff0[1]; + + return sse; +} + +static uint32_t sub_pixel_avg_sse_diff_64width_h_msa( + const uint8_t *src, int32_t src_stride, const uint8_t *dst, + int32_t dst_stride, const uint8_t *sec_pred, const uint8_t *filter, + int32_t height, int32_t *diff) { + uint32_t loop_cnt, sse = 0; + int32_t diff0[4]; + + for (loop_cnt = 0; loop_cnt < 4; ++loop_cnt) { + sse += + subpel_avg_ssediff_16w_h_msa(src, src_stride, dst, dst_stride, sec_pred, + filter, height, &diff0[loop_cnt], 64); + src += 16; + dst += 16; + sec_pred += 16; + } + + *diff = diff0[0] + diff0[1] + diff0[2] + diff0[3]; + + return sse; +} + +static uint32_t sub_pixel_avg_sse_diff_4width_v_msa( + const uint8_t *src, int32_t src_stride, const uint8_t *dst, + int32_t dst_stride, const uint8_t *sec_pred, const uint8_t *filter, + int32_t height, int32_t *diff) { + int16_t filtval; + uint32_t loop_cnt; + uint32_t ref0, ref1, ref2, ref3; + v16u8 src0, src1, src2, src3, src4; + v16u8 src10_r, src32_r, src21_r, src43_r; + v16u8 out, pred, ref = { 0 }; + v16u8 src2110, src4332, filt0; + v8i16 avg = { 0 }; + v4i32 vec, var = { 0 }; + v8u16 tmp0, tmp1; + + filtval = LH(filter); + filt0 = (v16u8)__msa_fill_h(filtval); + + src0 = LD_UB(src); + src += src_stride; + + for (loop_cnt = (height >> 2); loop_cnt--;) { + LD_UB4(src, src_stride, src1, src2, src3, src4); + src += (4 * src_stride); + pred = LD_UB(sec_pred); + sec_pred += 16; + LW4(dst, dst_stride, ref0, ref1, ref2, ref3); + dst += (4 * dst_stride); + + INSERT_W4_UB(ref0, ref1, ref2, ref3, ref); + ILVR_B4_UB(src1, src0, src2, src1, src3, src2, src4, src3, src10_r, src21_r, + src32_r, src43_r); + ILVR_D2_UB(src21_r, src10_r, src43_r, src32_r, src2110, src4332); + DOTP_UB2_UH(src2110, src4332, filt0, filt0, tmp0, tmp1); + SRARI_H2_UH(tmp0, tmp1, FILTER_BITS); + + out = (v16u8)__msa_pckev_b((v16i8)tmp1, (v16i8)tmp0); + out = __msa_aver_u_b(out, pred); + CALC_MSE_AVG_B(out, ref, var, avg); + src0 = src4; + } + + vec = __msa_hadd_s_w(avg, avg); + *diff = HADD_SW_S32(vec); + + return HADD_SW_S32(var); +} + +static uint32_t sub_pixel_avg_sse_diff_8width_v_msa( + const uint8_t *src, int32_t src_stride, const uint8_t *dst, + int32_t dst_stride, const uint8_t *sec_pred, const uint8_t *filter, + int32_t height, int32_t *diff) { + int16_t filtval; + uint32_t loop_cnt; + v16u8 src0, src1, src2, src3, src4; + v16u8 ref0, ref1, ref2, ref3; + v16u8 pred0, pred1, filt0; + v8u16 vec0, vec1, vec2, vec3; + v8u16 tmp0, tmp1, tmp2, tmp3; + v8i16 avg = { 0 }; + v4i32 vec, var = { 0 }; + + filtval = LH(filter); + filt0 = (v16u8)__msa_fill_h(filtval); + + src0 = LD_UB(src); + src += src_stride; + + for (loop_cnt = (height >> 2); loop_cnt--;) { + LD_UB4(src, src_stride, src1, src2, src3, src4); + src += (4 * src_stride); + LD_UB2(sec_pred, 16, pred0, pred1); + sec_pred += 32; + LD_UB4(dst, dst_stride, ref0, ref1, ref2, ref3); + dst += (4 * dst_stride); + PCKEV_D2_UB(ref1, ref0, ref3, ref2, ref0, ref1); + ILVR_B4_UH(src1, src0, src2, src1, src3, src2, src4, src3, vec0, vec1, vec2, + vec3); + DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, tmp0, tmp1, + tmp2, tmp3); + SRARI_H4_UH(tmp0, tmp1, tmp2, tmp3, FILTER_BITS); + PCKEV_B2_UB(tmp1, tmp0, tmp3, tmp2, src0, src1); + AVER_UB2_UB(src0, pred0, src1, pred1, src0, src1); + CALC_MSE_AVG_B(src0, ref0, var, avg); + CALC_MSE_AVG_B(src1, ref1, var, avg); + + src0 = src4; + } + + vec = __msa_hadd_s_w(avg, avg); + *diff = HADD_SW_S32(vec); + + return HADD_SW_S32(var); +} + +static uint32_t subpel_avg_ssediff_16w_v_msa( + const uint8_t *src, int32_t src_stride, const uint8_t *dst, + int32_t dst_stride, const uint8_t *sec_pred, const uint8_t *filter, + int32_t height, int32_t *diff, int32_t width) { + int16_t filtval; + uint32_t loop_cnt; + v16u8 ref0, ref1, ref2, ref3; + v16u8 pred0, pred1, pred2, pred3; + v16u8 src0, src1, src2, src3, src4; + v16u8 out0, out1, out2, out3, filt0; + v8u16 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7; + v8u16 tmp0, tmp1, tmp2, tmp3; + v8i16 avg = { 0 }; + v4i32 vec, var = { 0 }; + + filtval = LH(filter); + filt0 = (v16u8)__msa_fill_h(filtval); + + src0 = LD_UB(src); + src += src_stride; + + for (loop_cnt = (height >> 2); loop_cnt--;) { + LD_UB4(src, src_stride, src1, src2, src3, src4); + src += (4 * src_stride); + LD_UB4(sec_pred, width, pred0, pred1, pred2, pred3); + sec_pred += (4 * width); + + ILVR_B2_UH(src1, src0, src2, src1, vec0, vec2); + ILVL_B2_UH(src1, src0, src2, src1, vec1, vec3); + DOTP_UB2_UH(vec0, vec1, filt0, filt0, tmp0, tmp1); + SRARI_H2_UH(tmp0, tmp1, FILTER_BITS); + out0 = (v16u8)__msa_pckev_b((v16i8)tmp1, (v16i8)tmp0); + + ILVR_B2_UH(src3, src2, src4, src3, vec4, vec6); + ILVL_B2_UH(src3, src2, src4, src3, vec5, vec7); + DOTP_UB2_UH(vec2, vec3, filt0, filt0, tmp2, tmp3); + SRARI_H2_UH(tmp2, tmp3, FILTER_BITS); + out1 = (v16u8)__msa_pckev_b((v16i8)tmp3, (v16i8)tmp2); + + DOTP_UB2_UH(vec4, vec5, filt0, filt0, tmp0, tmp1); + SRARI_H2_UH(tmp0, tmp1, FILTER_BITS); + out2 = (v16u8)__msa_pckev_b((v16i8)tmp1, (v16i8)tmp0); + + DOTP_UB2_UH(vec6, vec7, filt0, filt0, tmp2, tmp3); + SRARI_H2_UH(tmp2, tmp3, FILTER_BITS); + out3 = (v16u8)__msa_pckev_b((v16i8)tmp3, (v16i8)tmp2); + + src0 = src4; + LD_UB4(dst, dst_stride, ref0, ref1, ref2, ref3); + dst += (4 * dst_stride); + + AVER_UB4_UB(out0, pred0, out1, pred1, out2, pred2, out3, pred3, out0, out1, + out2, out3); + + CALC_MSE_AVG_B(out0, ref0, var, avg); + CALC_MSE_AVG_B(out1, ref1, var, avg); + CALC_MSE_AVG_B(out2, ref2, var, avg); + CALC_MSE_AVG_B(out3, ref3, var, avg); + } + + vec = __msa_hadd_s_w(avg, avg); + *diff = HADD_SW_S32(vec); + + return HADD_SW_S32(var); +} + +static uint32_t sub_pixel_avg_sse_diff_16width_v_msa( + const uint8_t *src, int32_t src_stride, const uint8_t *dst, + int32_t dst_stride, const uint8_t *sec_pred, const uint8_t *filter, + int32_t height, int32_t *diff) { + return subpel_avg_ssediff_16w_v_msa(src, src_stride, dst, dst_stride, + sec_pred, filter, height, diff, 16); +} + +static uint32_t sub_pixel_avg_sse_diff_32width_v_msa( + const uint8_t *src, int32_t src_stride, const uint8_t *dst, + int32_t dst_stride, const uint8_t *sec_pred, const uint8_t *filter, + int32_t height, int32_t *diff) { + uint32_t loop_cnt, sse = 0; + int32_t diff0[2]; + + for (loop_cnt = 0; loop_cnt < 2; ++loop_cnt) { + sse += + subpel_avg_ssediff_16w_v_msa(src, src_stride, dst, dst_stride, sec_pred, + filter, height, &diff0[loop_cnt], 32); + src += 16; + dst += 16; + sec_pred += 16; + } + + *diff = diff0[0] + diff0[1]; + + return sse; +} + +static uint32_t sub_pixel_avg_sse_diff_64width_v_msa( + const uint8_t *src, int32_t src_stride, const uint8_t *dst, + int32_t dst_stride, const uint8_t *sec_pred, const uint8_t *filter, + int32_t height, int32_t *diff) { + uint32_t loop_cnt, sse = 0; + int32_t diff0[4]; + + for (loop_cnt = 0; loop_cnt < 4; ++loop_cnt) { + sse += + subpel_avg_ssediff_16w_v_msa(src, src_stride, dst, dst_stride, sec_pred, + filter, height, &diff0[loop_cnt], 64); + src += 16; + dst += 16; + sec_pred += 16; + } + + *diff = diff0[0] + diff0[1] + diff0[2] + diff0[3]; + + return sse; +} + +static uint32_t sub_pixel_avg_sse_diff_4width_hv_msa( + const uint8_t *src, int32_t src_stride, const uint8_t *dst, + int32_t dst_stride, const uint8_t *sec_pred, const uint8_t *filter_horiz, + const uint8_t *filter_vert, int32_t height, int32_t *diff) { + int16_t filtval; + uint32_t loop_cnt; + uint32_t ref0, ref1, ref2, ref3; + v16u8 src0, src1, src2, src3, src4; + v16u8 mask = { 0, 1, 1, 2, 2, 3, 3, 4, 16, 17, 17, 18, 18, 19, 19, 20 }; + v16u8 filt_hz, filt_vt, vec0, vec1; + v16u8 out, pred, ref = { 0 }; + v8u16 hz_out0, hz_out1, hz_out2, hz_out3, hz_out4, tmp0, tmp1; + v8i16 avg = { 0 }; + v4i32 vec, var = { 0 }; + + filtval = LH(filter_horiz); + filt_hz = (v16u8)__msa_fill_h(filtval); + filtval = LH(filter_vert); + filt_vt = (v16u8)__msa_fill_h(filtval); + + src0 = LD_UB(src); + src += src_stride; + + for (loop_cnt = (height >> 2); loop_cnt--;) { + LD_UB4(src, src_stride, src1, src2, src3, src4); + src += (4 * src_stride); + pred = LD_UB(sec_pred); + sec_pred += 16; + LW4(dst, dst_stride, ref0, ref1, ref2, ref3); + dst += (4 * dst_stride); + INSERT_W4_UB(ref0, ref1, ref2, ref3, ref); + hz_out0 = HORIZ_2TAP_FILT_UH(src0, src1, mask, filt_hz, FILTER_BITS); + hz_out2 = HORIZ_2TAP_FILT_UH(src2, src3, mask, filt_hz, FILTER_BITS); + hz_out4 = HORIZ_2TAP_FILT_UH(src4, src4, mask, filt_hz, FILTER_BITS); + hz_out1 = (v8u16)__msa_sldi_b((v16i8)hz_out2, (v16i8)hz_out0, 8); + hz_out3 = (v8u16)__msa_pckod_d((v2i64)hz_out4, (v2i64)hz_out2); + ILVEV_B2_UB(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1); + DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp0, tmp1); + SRARI_H2_UH(tmp0, tmp1, FILTER_BITS); + out = (v16u8)__msa_pckev_b((v16i8)tmp1, (v16i8)tmp0); + out = __msa_aver_u_b(out, pred); + CALC_MSE_AVG_B(out, ref, var, avg); + src0 = src4; + } + + vec = __msa_hadd_s_w(avg, avg); + *diff = HADD_SW_S32(vec); + + return HADD_SW_S32(var); +} + +static uint32_t sub_pixel_avg_sse_diff_8width_hv_msa( + const uint8_t *src, int32_t src_stride, const uint8_t *dst, + int32_t dst_stride, const uint8_t *sec_pred, const uint8_t *filter_horiz, + const uint8_t *filter_vert, int32_t height, int32_t *diff) { + int16_t filtval; + uint32_t loop_cnt; + v16u8 ref0, ref1, ref2, ref3; + v16u8 src0, src1, src2, src3, src4; + v16u8 pred0, pred1, out0, out1; + v16u8 filt_hz, filt_vt, vec0; + v16u8 mask = { 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8 }; + v8u16 hz_out0, hz_out1, tmp0, tmp1, tmp2, tmp3; + v8i16 avg = { 0 }; + v4i32 vec, var = { 0 }; + + filtval = LH(filter_horiz); + filt_hz = (v16u8)__msa_fill_h(filtval); + filtval = LH(filter_vert); + filt_vt = (v16u8)__msa_fill_h(filtval); + + src0 = LD_UB(src); + src += src_stride; + hz_out0 = HORIZ_2TAP_FILT_UH(src0, src0, mask, filt_hz, FILTER_BITS); + + for (loop_cnt = (height >> 2); loop_cnt--;) { + LD_UB4(src, src_stride, src1, src2, src3, src4); + src += (4 * src_stride); + LD_UB2(sec_pred, 16, pred0, pred1); + sec_pred += 32; + LD_UB4(dst, dst_stride, ref0, ref1, ref2, ref3); + dst += (4 * dst_stride); + + PCKEV_D2_UB(ref1, ref0, ref3, ref2, ref0, ref1); + hz_out1 = HORIZ_2TAP_FILT_UH(src1, src1, mask, filt_hz, FILTER_BITS); + + vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out1, (v16i8)hz_out0); + tmp0 = __msa_dotp_u_h(vec0, filt_vt); + hz_out0 = HORIZ_2TAP_FILT_UH(src2, src2, mask, filt_hz, FILTER_BITS); + + vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out0, (v16i8)hz_out1); + tmp1 = __msa_dotp_u_h(vec0, filt_vt); + SRARI_H2_UH(tmp0, tmp1, FILTER_BITS); + hz_out1 = HORIZ_2TAP_FILT_UH(src3, src3, mask, filt_hz, FILTER_BITS); + + vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out1, (v16i8)hz_out0); + tmp2 = __msa_dotp_u_h(vec0, filt_vt); + hz_out0 = HORIZ_2TAP_FILT_UH(src4, src4, mask, filt_hz, FILTER_BITS); + + vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out0, (v16i8)hz_out1); + tmp3 = __msa_dotp_u_h(vec0, filt_vt); + + SRARI_H2_UH(tmp2, tmp3, FILTER_BITS); + PCKEV_B2_UB(tmp1, tmp0, tmp3, tmp2, out0, out1); + AVER_UB2_UB(out0, pred0, out1, pred1, out0, out1); + + CALC_MSE_AVG_B(out0, ref0, var, avg); + CALC_MSE_AVG_B(out1, ref1, var, avg); + } + + vec = __msa_hadd_s_w(avg, avg); + *diff = HADD_SW_S32(vec); + + return HADD_SW_S32(var); +} + +static uint32_t subpel_avg_ssediff_16w_hv_msa( + const uint8_t *src, int32_t src_stride, const uint8_t *dst, + int32_t dst_stride, const uint8_t *sec_pred, const uint8_t *filter_horiz, + const uint8_t *filter_vert, int32_t height, int32_t *diff, int32_t width) { + int16_t filtval; + uint32_t loop_cnt; + v16u8 src0, src1, src2, src3, src4, src5, src6, src7; + v16u8 ref0, ref1, ref2, ref3; + v16u8 pred0, pred1, pred2, pred3; + v16u8 out0, out1, out2, out3; + v16u8 filt_hz, filt_vt, vec0, vec1; + v16u8 mask = { 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8 }; + v8u16 hz_out0, hz_out1, hz_out2, hz_out3, tmp0, tmp1; + v8i16 avg = { 0 }; + v4i32 vec, var = { 0 }; + + filtval = LH(filter_horiz); + filt_hz = (v16u8)__msa_fill_h(filtval); + filtval = LH(filter_vert); + filt_vt = (v16u8)__msa_fill_h(filtval); + + LD_UB2(src, 8, src0, src1); + src += src_stride; + + hz_out0 = HORIZ_2TAP_FILT_UH(src0, src0, mask, filt_hz, FILTER_BITS); + hz_out2 = HORIZ_2TAP_FILT_UH(src1, src1, mask, filt_hz, FILTER_BITS); + + for (loop_cnt = (height >> 2); loop_cnt--;) { + LD_UB4(src, src_stride, src0, src2, src4, src6); + LD_UB4(src + 8, src_stride, src1, src3, src5, src7); + src += (4 * src_stride); + LD_UB4(sec_pred, width, pred0, pred1, pred2, pred3); + sec_pred += (4 * width); + + hz_out1 = HORIZ_2TAP_FILT_UH(src0, src0, mask, filt_hz, FILTER_BITS); + hz_out3 = HORIZ_2TAP_FILT_UH(src1, src1, mask, filt_hz, FILTER_BITS); + ILVEV_B2_UB(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1); + DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp0, tmp1); + SRARI_H2_UH(tmp0, tmp1, FILTER_BITS); + out0 = (v16u8)__msa_pckev_b((v16i8)tmp1, (v16i8)tmp0); + + hz_out0 = HORIZ_2TAP_FILT_UH(src2, src2, mask, filt_hz, FILTER_BITS); + hz_out2 = HORIZ_2TAP_FILT_UH(src3, src3, mask, filt_hz, FILTER_BITS); + ILVEV_B2_UB(hz_out1, hz_out0, hz_out3, hz_out2, vec0, vec1); + DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp0, tmp1); + SRARI_H2_UH(tmp0, tmp1, FILTER_BITS); + out1 = (v16u8)__msa_pckev_b((v16i8)tmp1, (v16i8)tmp0); + + hz_out1 = HORIZ_2TAP_FILT_UH(src4, src4, mask, filt_hz, FILTER_BITS); + hz_out3 = HORIZ_2TAP_FILT_UH(src5, src5, mask, filt_hz, FILTER_BITS); + ILVEV_B2_UB(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1); + DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp0, tmp1); + SRARI_H2_UH(tmp0, tmp1, FILTER_BITS); + out2 = (v16u8)__msa_pckev_b((v16i8)tmp1, (v16i8)tmp0); + + hz_out0 = HORIZ_2TAP_FILT_UH(src6, src6, mask, filt_hz, FILTER_BITS); + hz_out2 = HORIZ_2TAP_FILT_UH(src7, src7, mask, filt_hz, FILTER_BITS); + ILVEV_B2_UB(hz_out1, hz_out0, hz_out3, hz_out2, vec0, vec1); + DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp0, tmp1); + SRARI_H2_UH(tmp0, tmp1, FILTER_BITS); + out3 = (v16u8)__msa_pckev_b((v16i8)tmp1, (v16i8)tmp0); + + LD_UB4(dst, dst_stride, ref0, ref1, ref2, ref3); + dst += (4 * dst_stride); + + AVER_UB4_UB(out0, pred0, out1, pred1, out2, pred2, out3, pred3, out0, out1, + out2, out3); + + CALC_MSE_AVG_B(out0, ref0, var, avg); + CALC_MSE_AVG_B(out1, ref1, var, avg); + CALC_MSE_AVG_B(out2, ref2, var, avg); + CALC_MSE_AVG_B(out3, ref3, var, avg); + } + + vec = __msa_hadd_s_w(avg, avg); + *diff = HADD_SW_S32(vec); + + return HADD_SW_S32(var); +} + +static uint32_t sub_pixel_avg_sse_diff_16width_hv_msa( + const uint8_t *src, int32_t src_stride, const uint8_t *dst, + int32_t dst_stride, const uint8_t *sec_pred, const uint8_t *filter_horiz, + const uint8_t *filter_vert, int32_t height, int32_t *diff) { + return subpel_avg_ssediff_16w_hv_msa(src, src_stride, dst, dst_stride, + sec_pred, filter_horiz, filter_vert, + height, diff, 16); +} + +static uint32_t sub_pixel_avg_sse_diff_32width_hv_msa( + const uint8_t *src, int32_t src_stride, const uint8_t *dst, + int32_t dst_stride, const uint8_t *sec_pred, const uint8_t *filter_horiz, + const uint8_t *filter_vert, int32_t height, int32_t *diff) { + uint32_t loop_cnt, sse = 0; + int32_t diff0[2]; + + for (loop_cnt = 0; loop_cnt < 2; ++loop_cnt) { + sse += subpel_avg_ssediff_16w_hv_msa(src, src_stride, dst, dst_stride, + sec_pred, filter_horiz, filter_vert, + height, &diff0[loop_cnt], 32); + src += 16; + dst += 16; + sec_pred += 16; + } + + *diff = diff0[0] + diff0[1]; + + return sse; +} + +static uint32_t sub_pixel_avg_sse_diff_64width_hv_msa( + const uint8_t *src, int32_t src_stride, const uint8_t *dst, + int32_t dst_stride, const uint8_t *sec_pred, const uint8_t *filter_horiz, + const uint8_t *filter_vert, int32_t height, int32_t *diff) { + uint32_t loop_cnt, sse = 0; + int32_t diff0[4]; + + for (loop_cnt = 0; loop_cnt < 4; ++loop_cnt) { + sse += subpel_avg_ssediff_16w_hv_msa(src, src_stride, dst, dst_stride, + sec_pred, filter_horiz, filter_vert, + height, &diff0[loop_cnt], 64); + src += 16; + dst += 16; + sec_pred += 16; + } + + *diff = diff0[0] + diff0[1] + diff0[2] + diff0[3]; + + return sse; +} + +#define VARIANCE_4Wx4H(sse, diff) VARIANCE_WxH(sse, diff, 4); +#define VARIANCE_4Wx8H(sse, diff) VARIANCE_WxH(sse, diff, 5); +#define VARIANCE_8Wx4H(sse, diff) VARIANCE_WxH(sse, diff, 5); +#define VARIANCE_8Wx8H(sse, diff) VARIANCE_WxH(sse, diff, 6); +#define VARIANCE_8Wx16H(sse, diff) VARIANCE_WxH(sse, diff, 7); +#define VARIANCE_16Wx8H(sse, diff) VARIANCE_WxH(sse, diff, 7); +#define VARIANCE_16Wx16H(sse, diff) VARIANCE_WxH(sse, diff, 8); + +#define VARIANCE_16Wx32H(sse, diff) VARIANCE_LARGE_WxH(sse, diff, 9); +#define VARIANCE_32Wx16H(sse, diff) VARIANCE_LARGE_WxH(sse, diff, 9); +#define VARIANCE_32Wx32H(sse, diff) VARIANCE_LARGE_WxH(sse, diff, 10); +#define VARIANCE_32Wx64H(sse, diff) VARIANCE_LARGE_WxH(sse, diff, 11); +#define VARIANCE_64Wx32H(sse, diff) VARIANCE_LARGE_WxH(sse, diff, 11); +#define VARIANCE_64Wx64H(sse, diff) VARIANCE_LARGE_WxH(sse, diff, 12); + +#define AOM_SUB_PIXEL_VARIANCE_WDXHT_MSA(wd, ht) \ + uint32_t aom_sub_pixel_variance##wd##x##ht##_msa( \ + const uint8_t *src, int32_t src_stride, int32_t xoffset, \ + int32_t yoffset, const uint8_t *ref, int32_t ref_stride, \ + uint32_t *sse) { \ + int32_t diff; \ + uint32_t var; \ + const uint8_t *h_filter = bilinear_filters_2t[xoffset]; \ + const uint8_t *v_filter = bilinear_filters_2t[yoffset]; \ + \ + if (yoffset) { \ + if (xoffset) { \ + *sse = sub_pixel_sse_diff_##wd##width_hv_msa( \ + src, src_stride, ref, ref_stride, h_filter, v_filter, ht, &diff); \ + } else { \ + *sse = sub_pixel_sse_diff_##wd##width_v_msa( \ + src, src_stride, ref, ref_stride, v_filter, ht, &diff); \ + } \ + \ + var = VARIANCE_##wd##Wx##ht##H(*sse, diff); \ + } else { \ + if (xoffset) { \ + *sse = sub_pixel_sse_diff_##wd##width_h_msa( \ + src, src_stride, ref, ref_stride, h_filter, ht, &diff); \ + \ + var = VARIANCE_##wd##Wx##ht##H(*sse, diff); \ + } else { \ + var = aom_variance##wd##x##ht##_msa(src, src_stride, ref, ref_stride, \ + sse); \ + } \ + } \ + \ + return var; \ + } + +/* clang-format off */ +AOM_SUB_PIXEL_VARIANCE_WDXHT_MSA(4, 4) +AOM_SUB_PIXEL_VARIANCE_WDXHT_MSA(4, 8) + +AOM_SUB_PIXEL_VARIANCE_WDXHT_MSA(8, 4) +AOM_SUB_PIXEL_VARIANCE_WDXHT_MSA(8, 8) +AOM_SUB_PIXEL_VARIANCE_WDXHT_MSA(8, 16) + +AOM_SUB_PIXEL_VARIANCE_WDXHT_MSA(16, 8) +AOM_SUB_PIXEL_VARIANCE_WDXHT_MSA(16, 16) +AOM_SUB_PIXEL_VARIANCE_WDXHT_MSA(16, 32) + +AOM_SUB_PIXEL_VARIANCE_WDXHT_MSA(32, 16) +AOM_SUB_PIXEL_VARIANCE_WDXHT_MSA(32, 32) +AOM_SUB_PIXEL_VARIANCE_WDXHT_MSA(32, 64) + +AOM_SUB_PIXEL_VARIANCE_WDXHT_MSA(64, 32) +AOM_SUB_PIXEL_VARIANCE_WDXHT_MSA(64, 64) +/* clang-format on */ + +#define AOM_SUB_PIXEL_AVG_VARIANCE_WDXHT_MSA(wd, ht) \ + uint32_t aom_sub_pixel_avg_variance##wd##x##ht##_msa( \ + const uint8_t *src_ptr, int32_t src_stride, int32_t xoffset, \ + int32_t yoffset, const uint8_t *ref_ptr, int32_t ref_stride, \ + uint32_t *sse, const uint8_t *sec_pred) { \ + int32_t diff; \ + const uint8_t *h_filter = bilinear_filters_2t[xoffset]; \ + const uint8_t *v_filter = bilinear_filters_2t[yoffset]; \ + \ + if (yoffset) { \ + if (xoffset) { \ + *sse = sub_pixel_avg_sse_diff_##wd##width_hv_msa( \ + src_ptr, src_stride, ref_ptr, ref_stride, sec_pred, h_filter, \ + v_filter, ht, &diff); \ + } else { \ + *sse = sub_pixel_avg_sse_diff_##wd##width_v_msa( \ + src_ptr, src_stride, ref_ptr, ref_stride, sec_pred, v_filter, ht, \ + &diff); \ + } \ + } else { \ + if (xoffset) { \ + *sse = sub_pixel_avg_sse_diff_##wd##width_h_msa( \ + src_ptr, src_stride, ref_ptr, ref_stride, sec_pred, h_filter, ht, \ + &diff); \ + } else { \ + *sse = avg_sse_diff_##wd##width_msa(src_ptr, src_stride, ref_ptr, \ + ref_stride, sec_pred, ht, &diff); \ + } \ + } \ + \ + return VARIANCE_##wd##Wx##ht##H(*sse, diff); \ + } + +/* clang-format off */ +AOM_SUB_PIXEL_AVG_VARIANCE_WDXHT_MSA(4, 4) +AOM_SUB_PIXEL_AVG_VARIANCE_WDXHT_MSA(4, 8) + +AOM_SUB_PIXEL_AVG_VARIANCE_WDXHT_MSA(8, 4) +AOM_SUB_PIXEL_AVG_VARIANCE_WDXHT_MSA(8, 8) +AOM_SUB_PIXEL_AVG_VARIANCE_WDXHT_MSA(8, 16) + +AOM_SUB_PIXEL_AVG_VARIANCE_WDXHT_MSA(16, 8) +AOM_SUB_PIXEL_AVG_VARIANCE_WDXHT_MSA(16, 16) +AOM_SUB_PIXEL_AVG_VARIANCE_WDXHT_MSA(16, 32) + +AOM_SUB_PIXEL_AVG_VARIANCE_WDXHT_MSA(32, 16) +AOM_SUB_PIXEL_AVG_VARIANCE_WDXHT_MSA(32, 32) +/* clang-format on */ + +uint32_t aom_sub_pixel_avg_variance32x64_msa(const uint8_t *src_ptr, + int32_t src_stride, + int32_t xoffset, int32_t yoffset, + const uint8_t *ref_ptr, + int32_t ref_stride, uint32_t *sse, + const uint8_t *sec_pred) { + int32_t diff; + const uint8_t *h_filter = bilinear_filters_2t[xoffset]; + const uint8_t *v_filter = bilinear_filters_2t[yoffset]; + + if (yoffset) { + if (xoffset) { + *sse = sub_pixel_avg_sse_diff_32width_hv_msa( + src_ptr, src_stride, ref_ptr, ref_stride, sec_pred, h_filter, + v_filter, 64, &diff); + } else { + *sse = sub_pixel_avg_sse_diff_32width_v_msa(src_ptr, src_stride, ref_ptr, + ref_stride, sec_pred, + v_filter, 64, &diff); + } + } else { + if (xoffset) { + *sse = sub_pixel_avg_sse_diff_32width_h_msa(src_ptr, src_stride, ref_ptr, + ref_stride, sec_pred, + h_filter, 64, &diff); + } else { + *sse = avg_sse_diff_32x64_msa(src_ptr, src_stride, ref_ptr, ref_stride, + sec_pred, &diff); + } + } + + return VARIANCE_32Wx64H(*sse, diff); +} + +#define AOM_SUB_PIXEL_AVG_VARIANCE64XHEIGHT_MSA(ht) \ + uint32_t aom_sub_pixel_avg_variance64x##ht##_msa( \ + const uint8_t *src_ptr, int32_t src_stride, int32_t xoffset, \ + int32_t yoffset, const uint8_t *ref_ptr, int32_t ref_stride, \ + uint32_t *sse, const uint8_t *sec_pred) { \ + int32_t diff; \ + const uint8_t *h_filter = bilinear_filters_2t[xoffset]; \ + const uint8_t *v_filter = bilinear_filters_2t[yoffset]; \ + \ + if (yoffset) { \ + if (xoffset) { \ + *sse = sub_pixel_avg_sse_diff_64width_hv_msa( \ + src_ptr, src_stride, ref_ptr, ref_stride, sec_pred, h_filter, \ + v_filter, ht, &diff); \ + } else { \ + *sse = sub_pixel_avg_sse_diff_64width_v_msa( \ + src_ptr, src_stride, ref_ptr, ref_stride, sec_pred, v_filter, ht, \ + &diff); \ + } \ + } else { \ + if (xoffset) { \ + *sse = sub_pixel_avg_sse_diff_64width_h_msa( \ + src_ptr, src_stride, ref_ptr, ref_stride, sec_pred, h_filter, ht, \ + &diff); \ + } else { \ + *sse = avg_sse_diff_64x##ht##_msa(src_ptr, src_stride, ref_ptr, \ + ref_stride, sec_pred, &diff); \ + } \ + } \ + \ + return VARIANCE_64Wx##ht##H(*sse, diff); \ + } + +/* clang-format off */ +AOM_SUB_PIXEL_AVG_VARIANCE64XHEIGHT_MSA(32) +AOM_SUB_PIXEL_AVG_VARIANCE64XHEIGHT_MSA(64) +/* clang-format on */ diff --git a/third_party/aom/aom_dsp/mips/subtract_msa.c b/third_party/aom/aom_dsp/mips/subtract_msa.c new file mode 100644 index 0000000000..bfed773ac8 --- /dev/null +++ b/third_party/aom/aom_dsp/mips/subtract_msa.c @@ -0,0 +1,266 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include "config/aom_dsp_rtcd.h" + +#include "aom_dsp/mips/macros_msa.h" + +static void sub_blk_4x4_msa(const uint8_t *src_ptr, int32_t src_stride, + const uint8_t *pred_ptr, int32_t pred_stride, + int16_t *diff_ptr, int32_t diff_stride) { + uint32_t src0, src1, src2, src3; + uint32_t pred0, pred1, pred2, pred3; + v16i8 src = { 0 }; + v16i8 pred = { 0 }; + v16u8 src_l0, src_l1; + v8i16 diff0, diff1; + + LW4(src_ptr, src_stride, src0, src1, src2, src3); + LW4(pred_ptr, pred_stride, pred0, pred1, pred2, pred3); + INSERT_W4_SB(src0, src1, src2, src3, src); + INSERT_W4_SB(pred0, pred1, pred2, pred3, pred); + ILVRL_B2_UB(src, pred, src_l0, src_l1); + HSUB_UB2_SH(src_l0, src_l1, diff0, diff1); + ST8x4_UB(diff0, diff1, diff_ptr, (2 * diff_stride)); +} + +static void sub_blk_8x8_msa(const uint8_t *src_ptr, int32_t src_stride, + const uint8_t *pred_ptr, int32_t pred_stride, + int16_t *diff_ptr, int32_t diff_stride) { + uint32_t loop_cnt; + uint64_t src0, src1, pred0, pred1; + v16i8 src = { 0 }; + v16i8 pred = { 0 }; + v16u8 src_l0, src_l1; + v8i16 diff0, diff1; + + for (loop_cnt = 4; loop_cnt--;) { + LD2(src_ptr, src_stride, src0, src1); + src_ptr += (2 * src_stride); + LD2(pred_ptr, pred_stride, pred0, pred1); + pred_ptr += (2 * pred_stride); + + INSERT_D2_SB(src0, src1, src); + INSERT_D2_SB(pred0, pred1, pred); + ILVRL_B2_UB(src, pred, src_l0, src_l1); + HSUB_UB2_SH(src_l0, src_l1, diff0, diff1); + ST_SH2(diff0, diff1, diff_ptr, diff_stride); + diff_ptr += (2 * diff_stride); + } +} + +static void sub_blk_16x16_msa(const uint8_t *src, int32_t src_stride, + const uint8_t *pred, int32_t pred_stride, + int16_t *diff, int32_t diff_stride) { + int8_t count; + v16i8 src0, src1, src2, src3, src4, src5, src6, src7; + v16i8 pred0, pred1, pred2, pred3, pred4, pred5, pred6, pred7; + v16u8 src_l0, src_l1; + v8i16 diff0, diff1; + + for (count = 2; count--;) { + LD_SB8(src, src_stride, src0, src1, src2, src3, src4, src5, src6, src7); + src += (8 * src_stride); + + LD_SB8(pred, pred_stride, pred0, pred1, pred2, pred3, pred4, pred5, pred6, + pred7); + pred += (8 * pred_stride); + + ILVRL_B2_UB(src0, pred0, src_l0, src_l1); + HSUB_UB2_SH(src_l0, src_l1, diff0, diff1); + ST_SH2(diff0, diff1, diff, 8); + diff += diff_stride; + + ILVRL_B2_UB(src1, pred1, src_l0, src_l1); + HSUB_UB2_SH(src_l0, src_l1, diff0, diff1); + ST_SH2(diff0, diff1, diff, 8); + diff += diff_stride; + + ILVRL_B2_UB(src2, pred2, src_l0, src_l1); + HSUB_UB2_SH(src_l0, src_l1, diff0, diff1); + ST_SH2(diff0, diff1, diff, 8); + diff += diff_stride; + + ILVRL_B2_UB(src3, pred3, src_l0, src_l1); + HSUB_UB2_SH(src_l0, src_l1, diff0, diff1); + ST_SH2(diff0, diff1, diff, 8); + diff += diff_stride; + + ILVRL_B2_UB(src4, pred4, src_l0, src_l1); + HSUB_UB2_SH(src_l0, src_l1, diff0, diff1); + ST_SH2(diff0, diff1, diff, 8); + diff += diff_stride; + + ILVRL_B2_UB(src5, pred5, src_l0, src_l1); + HSUB_UB2_SH(src_l0, src_l1, diff0, diff1); + ST_SH2(diff0, diff1, diff, 8); + diff += diff_stride; + + ILVRL_B2_UB(src6, pred6, src_l0, src_l1); + HSUB_UB2_SH(src_l0, src_l1, diff0, diff1); + ST_SH2(diff0, diff1, diff, 8); + diff += diff_stride; + + ILVRL_B2_UB(src7, pred7, src_l0, src_l1); + HSUB_UB2_SH(src_l0, src_l1, diff0, diff1); + ST_SH2(diff0, diff1, diff, 8); + diff += diff_stride; + } +} + +static void sub_blk_32x32_msa(const uint8_t *src, int32_t src_stride, + const uint8_t *pred, int32_t pred_stride, + int16_t *diff, int32_t diff_stride) { + uint32_t loop_cnt; + v16i8 src0, src1, src2, src3, src4, src5, src6, src7; + v16i8 pred0, pred1, pred2, pred3, pred4, pred5, pred6, pred7; + v16u8 src_l0, src_l1; + v8i16 diff0, diff1; + + for (loop_cnt = 8; loop_cnt--;) { + LD_SB2(src, 16, src0, src1); + src += src_stride; + LD_SB2(src, 16, src2, src3); + src += src_stride; + LD_SB2(src, 16, src4, src5); + src += src_stride; + LD_SB2(src, 16, src6, src7); + src += src_stride; + + LD_SB2(pred, 16, pred0, pred1); + pred += pred_stride; + LD_SB2(pred, 16, pred2, pred3); + pred += pred_stride; + LD_SB2(pred, 16, pred4, pred5); + pred += pred_stride; + LD_SB2(pred, 16, pred6, pred7); + pred += pred_stride; + + ILVRL_B2_UB(src0, pred0, src_l0, src_l1); + HSUB_UB2_SH(src_l0, src_l1, diff0, diff1); + ST_SH2(diff0, diff1, diff, 8); + ILVRL_B2_UB(src1, pred1, src_l0, src_l1); + HSUB_UB2_SH(src_l0, src_l1, diff0, diff1); + ST_SH2(diff0, diff1, diff + 16, 8); + diff += diff_stride; + + ILVRL_B2_UB(src2, pred2, src_l0, src_l1); + HSUB_UB2_SH(src_l0, src_l1, diff0, diff1); + ST_SH2(diff0, diff1, diff, 8); + ILVRL_B2_UB(src3, pred3, src_l0, src_l1); + HSUB_UB2_SH(src_l0, src_l1, diff0, diff1); + ST_SH2(diff0, diff1, diff + 16, 8); + diff += diff_stride; + + ILVRL_B2_UB(src4, pred4, src_l0, src_l1); + HSUB_UB2_SH(src_l0, src_l1, diff0, diff1); + ST_SH2(diff0, diff1, diff, 8); + ILVRL_B2_UB(src5, pred5, src_l0, src_l1); + HSUB_UB2_SH(src_l0, src_l1, diff0, diff1); + ST_SH2(diff0, diff1, diff + 16, 8); + diff += diff_stride; + + ILVRL_B2_UB(src6, pred6, src_l0, src_l1); + HSUB_UB2_SH(src_l0, src_l1, diff0, diff1); + ST_SH2(diff0, diff1, diff, 8); + ILVRL_B2_UB(src7, pred7, src_l0, src_l1); + HSUB_UB2_SH(src_l0, src_l1, diff0, diff1); + ST_SH2(diff0, diff1, diff + 16, 8); + diff += diff_stride; + } +} + +static void sub_blk_64x64_msa(const uint8_t *src, int32_t src_stride, + const uint8_t *pred, int32_t pred_stride, + int16_t *diff, int32_t diff_stride) { + uint32_t loop_cnt; + v16i8 src0, src1, src2, src3, src4, src5, src6, src7; + v16i8 pred0, pred1, pred2, pred3, pred4, pred5, pred6, pred7; + v16u8 src_l0, src_l1; + v8i16 diff0, diff1; + + for (loop_cnt = 32; loop_cnt--;) { + LD_SB4(src, 16, src0, src1, src2, src3); + src += src_stride; + LD_SB4(src, 16, src4, src5, src6, src7); + src += src_stride; + + LD_SB4(pred, 16, pred0, pred1, pred2, pred3); + pred += pred_stride; + LD_SB4(pred, 16, pred4, pred5, pred6, pred7); + pred += pred_stride; + + ILVRL_B2_UB(src0, pred0, src_l0, src_l1); + HSUB_UB2_SH(src_l0, src_l1, diff0, diff1); + ST_SH2(diff0, diff1, diff, 8); + ILVRL_B2_UB(src1, pred1, src_l0, src_l1); + HSUB_UB2_SH(src_l0, src_l1, diff0, diff1); + ST_SH2(diff0, diff1, diff + 16, 8); + ILVRL_B2_UB(src2, pred2, src_l0, src_l1); + HSUB_UB2_SH(src_l0, src_l1, diff0, diff1); + ST_SH2(diff0, diff1, diff + 32, 8); + ILVRL_B2_UB(src3, pred3, src_l0, src_l1); + HSUB_UB2_SH(src_l0, src_l1, diff0, diff1); + ST_SH2(diff0, diff1, diff + 48, 8); + diff += diff_stride; + + ILVRL_B2_UB(src4, pred4, src_l0, src_l1); + HSUB_UB2_SH(src_l0, src_l1, diff0, diff1); + ST_SH2(diff0, diff1, diff, 8); + ILVRL_B2_UB(src5, pred5, src_l0, src_l1); + HSUB_UB2_SH(src_l0, src_l1, diff0, diff1); + ST_SH2(diff0, diff1, diff + 16, 8); + ILVRL_B2_UB(src6, pred6, src_l0, src_l1); + HSUB_UB2_SH(src_l0, src_l1, diff0, diff1); + ST_SH2(diff0, diff1, diff + 32, 8); + ILVRL_B2_UB(src7, pred7, src_l0, src_l1); + HSUB_UB2_SH(src_l0, src_l1, diff0, diff1); + ST_SH2(diff0, diff1, diff + 48, 8); + diff += diff_stride; + } +} + +void aom_subtract_block_msa(int32_t rows, int32_t cols, int16_t *diff_ptr, + ptrdiff_t diff_stride, const uint8_t *src_ptr, + ptrdiff_t src_stride, const uint8_t *pred_ptr, + ptrdiff_t pred_stride) { + if (rows == cols) { + switch (rows) { + case 4: + sub_blk_4x4_msa(src_ptr, src_stride, pred_ptr, pred_stride, diff_ptr, + diff_stride); + break; + case 8: + sub_blk_8x8_msa(src_ptr, src_stride, pred_ptr, pred_stride, diff_ptr, + diff_stride); + break; + case 16: + sub_blk_16x16_msa(src_ptr, src_stride, pred_ptr, pred_stride, diff_ptr, + diff_stride); + break; + case 32: + sub_blk_32x32_msa(src_ptr, src_stride, pred_ptr, pred_stride, diff_ptr, + diff_stride); + break; + case 64: + sub_blk_64x64_msa(src_ptr, src_stride, pred_ptr, pred_stride, diff_ptr, + diff_stride); + break; + default: + aom_subtract_block_c(rows, cols, diff_ptr, diff_stride, src_ptr, + src_stride, pred_ptr, pred_stride); + break; + } + } else { + aom_subtract_block_c(rows, cols, diff_ptr, diff_stride, src_ptr, src_stride, + pred_ptr, pred_stride); + } +} diff --git a/third_party/aom/aom_dsp/mips/variance_msa.c b/third_party/aom/aom_dsp/mips/variance_msa.c new file mode 100644 index 0000000000..065c09ac55 --- /dev/null +++ b/third_party/aom/aom_dsp/mips/variance_msa.c @@ -0,0 +1,633 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include "config/aom_dsp_rtcd.h" + +#include "aom_dsp/mips/macros_msa.h" + +#define CALC_MSE_B(src, ref, var) \ + { \ + v16u8 src_l0_m, src_l1_m; \ + v8i16 res_l0_m, res_l1_m; \ + \ + ILVRL_B2_UB(src, ref, src_l0_m, src_l1_m); \ + HSUB_UB2_SH(src_l0_m, src_l1_m, res_l0_m, res_l1_m); \ + DPADD_SH2_SW(res_l0_m, res_l1_m, res_l0_m, res_l1_m, var, var); \ + } + +#define CALC_MSE_AVG_B(src, ref, var, sub) \ + { \ + v16u8 src_l0_m, src_l1_m; \ + v8i16 res_l0_m, res_l1_m; \ + \ + ILVRL_B2_UB(src, ref, src_l0_m, src_l1_m); \ + HSUB_UB2_SH(src_l0_m, src_l1_m, res_l0_m, res_l1_m); \ + DPADD_SH2_SW(res_l0_m, res_l1_m, res_l0_m, res_l1_m, var, var); \ + \ + sub += res_l0_m + res_l1_m; \ + } + +#define VARIANCE_WxH(sse, diff, shift) sse - (((uint32_t)diff * diff) >> shift) + +#define VARIANCE_LARGE_WxH(sse, diff, shift) \ + sse - (((int64_t)diff * diff) >> shift) + +static uint32_t sse_diff_4width_msa(const uint8_t *src_ptr, int32_t src_stride, + const uint8_t *ref_ptr, int32_t ref_stride, + int32_t height, int32_t *diff) { + uint32_t src0, src1, src2, src3; + uint32_t ref0, ref1, ref2, ref3; + int32_t ht_cnt; + v16u8 src = { 0 }; + v16u8 ref = { 0 }; + v8i16 avg = { 0 }; + v4i32 vec, var = { 0 }; + + for (ht_cnt = (height >> 2); ht_cnt--;) { + LW4(src_ptr, src_stride, src0, src1, src2, src3); + src_ptr += (4 * src_stride); + LW4(ref_ptr, ref_stride, ref0, ref1, ref2, ref3); + ref_ptr += (4 * ref_stride); + + INSERT_W4_UB(src0, src1, src2, src3, src); + INSERT_W4_UB(ref0, ref1, ref2, ref3, ref); + CALC_MSE_AVG_B(src, ref, var, avg); + } + + vec = __msa_hadd_s_w(avg, avg); + *diff = HADD_SW_S32(vec); + + return HADD_SW_S32(var); +} + +static uint32_t sse_diff_8width_msa(const uint8_t *src_ptr, int32_t src_stride, + const uint8_t *ref_ptr, int32_t ref_stride, + int32_t height, int32_t *diff) { + int32_t ht_cnt; + v16u8 src0, src1, src2, src3; + v16u8 ref0, ref1, ref2, ref3; + v8i16 avg = { 0 }; + v4i32 vec, var = { 0 }; + + for (ht_cnt = (height >> 2); ht_cnt--;) { + LD_UB4(src_ptr, src_stride, src0, src1, src2, src3); + src_ptr += (4 * src_stride); + LD_UB4(ref_ptr, ref_stride, ref0, ref1, ref2, ref3); + ref_ptr += (4 * ref_stride); + + PCKEV_D4_UB(src1, src0, src3, src2, ref1, ref0, ref3, ref2, src0, src1, + ref0, ref1); + CALC_MSE_AVG_B(src0, ref0, var, avg); + CALC_MSE_AVG_B(src1, ref1, var, avg); + } + + vec = __msa_hadd_s_w(avg, avg); + *diff = HADD_SW_S32(vec); + + return HADD_SW_S32(var); +} + +static uint32_t sse_diff_16width_msa(const uint8_t *src_ptr, int32_t src_stride, + const uint8_t *ref_ptr, int32_t ref_stride, + int32_t height, int32_t *diff) { + int32_t ht_cnt; + v16u8 src, ref; + v8i16 avg = { 0 }; + v4i32 vec, var = { 0 }; + + for (ht_cnt = (height >> 2); ht_cnt--;) { + src = LD_UB(src_ptr); + src_ptr += src_stride; + ref = LD_UB(ref_ptr); + ref_ptr += ref_stride; + CALC_MSE_AVG_B(src, ref, var, avg); + + src = LD_UB(src_ptr); + src_ptr += src_stride; + ref = LD_UB(ref_ptr); + ref_ptr += ref_stride; + CALC_MSE_AVG_B(src, ref, var, avg); + + src = LD_UB(src_ptr); + src_ptr += src_stride; + ref = LD_UB(ref_ptr); + ref_ptr += ref_stride; + CALC_MSE_AVG_B(src, ref, var, avg); + + src = LD_UB(src_ptr); + src_ptr += src_stride; + ref = LD_UB(ref_ptr); + ref_ptr += ref_stride; + CALC_MSE_AVG_B(src, ref, var, avg); + } + + vec = __msa_hadd_s_w(avg, avg); + *diff = HADD_SW_S32(vec); + + return HADD_SW_S32(var); +} + +static uint32_t sse_diff_32width_msa(const uint8_t *src_ptr, int32_t src_stride, + const uint8_t *ref_ptr, int32_t ref_stride, + int32_t height, int32_t *diff) { + int32_t ht_cnt; + v16u8 src0, src1, ref0, ref1; + v8i16 avg = { 0 }; + v4i32 vec, var = { 0 }; + + for (ht_cnt = (height >> 2); ht_cnt--;) { + LD_UB2(src_ptr, 16, src0, src1); + src_ptr += src_stride; + LD_UB2(ref_ptr, 16, ref0, ref1); + ref_ptr += ref_stride; + CALC_MSE_AVG_B(src0, ref0, var, avg); + CALC_MSE_AVG_B(src1, ref1, var, avg); + + LD_UB2(src_ptr, 16, src0, src1); + src_ptr += src_stride; + LD_UB2(ref_ptr, 16, ref0, ref1); + ref_ptr += ref_stride; + CALC_MSE_AVG_B(src0, ref0, var, avg); + CALC_MSE_AVG_B(src1, ref1, var, avg); + + LD_UB2(src_ptr, 16, src0, src1); + src_ptr += src_stride; + LD_UB2(ref_ptr, 16, ref0, ref1); + ref_ptr += ref_stride; + CALC_MSE_AVG_B(src0, ref0, var, avg); + CALC_MSE_AVG_B(src1, ref1, var, avg); + + LD_UB2(src_ptr, 16, src0, src1); + src_ptr += src_stride; + LD_UB2(ref_ptr, 16, ref0, ref1); + ref_ptr += ref_stride; + CALC_MSE_AVG_B(src0, ref0, var, avg); + CALC_MSE_AVG_B(src1, ref1, var, avg); + } + + vec = __msa_hadd_s_w(avg, avg); + *diff = HADD_SW_S32(vec); + + return HADD_SW_S32(var); +} + +static uint32_t sse_diff_32x64_msa(const uint8_t *src_ptr, int32_t src_stride, + const uint8_t *ref_ptr, int32_t ref_stride, + int32_t *diff) { + int32_t ht_cnt; + v16u8 src0, src1, ref0, ref1; + v8i16 avg0 = { 0 }; + v8i16 avg1 = { 0 }; + v4i32 vec, var = { 0 }; + + for (ht_cnt = 16; ht_cnt--;) { + LD_UB2(src_ptr, 16, src0, src1); + src_ptr += src_stride; + LD_UB2(ref_ptr, 16, ref0, ref1); + ref_ptr += ref_stride; + CALC_MSE_AVG_B(src0, ref0, var, avg0); + CALC_MSE_AVG_B(src1, ref1, var, avg1); + + LD_UB2(src_ptr, 16, src0, src1); + src_ptr += src_stride; + LD_UB2(ref_ptr, 16, ref0, ref1); + ref_ptr += ref_stride; + CALC_MSE_AVG_B(src0, ref0, var, avg0); + CALC_MSE_AVG_B(src1, ref1, var, avg1); + + LD_UB2(src_ptr, 16, src0, src1); + src_ptr += src_stride; + LD_UB2(ref_ptr, 16, ref0, ref1); + ref_ptr += ref_stride; + CALC_MSE_AVG_B(src0, ref0, var, avg0); + CALC_MSE_AVG_B(src1, ref1, var, avg1); + + LD_UB2(src_ptr, 16, src0, src1); + src_ptr += src_stride; + LD_UB2(ref_ptr, 16, ref0, ref1); + ref_ptr += ref_stride; + CALC_MSE_AVG_B(src0, ref0, var, avg0); + CALC_MSE_AVG_B(src1, ref1, var, avg1); + } + + vec = __msa_hadd_s_w(avg0, avg0); + vec += __msa_hadd_s_w(avg1, avg1); + *diff = HADD_SW_S32(vec); + + return HADD_SW_S32(var); +} + +static uint32_t sse_diff_64x32_msa(const uint8_t *src_ptr, int32_t src_stride, + const uint8_t *ref_ptr, int32_t ref_stride, + int32_t *diff) { + int32_t ht_cnt; + v16u8 src0, src1, src2, src3; + v16u8 ref0, ref1, ref2, ref3; + v8i16 avg0 = { 0 }; + v8i16 avg1 = { 0 }; + v4i32 vec, var = { 0 }; + + for (ht_cnt = 16; ht_cnt--;) { + LD_UB4(src_ptr, 16, src0, src1, src2, src3); + src_ptr += src_stride; + LD_UB4(ref_ptr, 16, ref0, ref1, ref2, ref3); + ref_ptr += ref_stride; + CALC_MSE_AVG_B(src0, ref0, var, avg0); + CALC_MSE_AVG_B(src2, ref2, var, avg0); + CALC_MSE_AVG_B(src1, ref1, var, avg1); + CALC_MSE_AVG_B(src3, ref3, var, avg1); + + LD_UB4(src_ptr, 16, src0, src1, src2, src3); + src_ptr += src_stride; + LD_UB4(ref_ptr, 16, ref0, ref1, ref2, ref3); + ref_ptr += ref_stride; + CALC_MSE_AVG_B(src0, ref0, var, avg0); + CALC_MSE_AVG_B(src2, ref2, var, avg0); + CALC_MSE_AVG_B(src1, ref1, var, avg1); + CALC_MSE_AVG_B(src3, ref3, var, avg1); + } + + vec = __msa_hadd_s_w(avg0, avg0); + vec += __msa_hadd_s_w(avg1, avg1); + *diff = HADD_SW_S32(vec); + + return HADD_SW_S32(var); +} + +static uint32_t sse_diff_64x64_msa(const uint8_t *src_ptr, int32_t src_stride, + const uint8_t *ref_ptr, int32_t ref_stride, + int32_t *diff) { + int32_t ht_cnt; + v16u8 src0, src1, src2, src3; + v16u8 ref0, ref1, ref2, ref3; + v8i16 avg0 = { 0 }; + v8i16 avg1 = { 0 }; + v8i16 avg2 = { 0 }; + v8i16 avg3 = { 0 }; + v4i32 vec, var = { 0 }; + + for (ht_cnt = 32; ht_cnt--;) { + LD_UB4(src_ptr, 16, src0, src1, src2, src3); + src_ptr += src_stride; + LD_UB4(ref_ptr, 16, ref0, ref1, ref2, ref3); + ref_ptr += ref_stride; + + CALC_MSE_AVG_B(src0, ref0, var, avg0); + CALC_MSE_AVG_B(src1, ref1, var, avg1); + CALC_MSE_AVG_B(src2, ref2, var, avg2); + CALC_MSE_AVG_B(src3, ref3, var, avg3); + LD_UB4(src_ptr, 16, src0, src1, src2, src3); + src_ptr += src_stride; + LD_UB4(ref_ptr, 16, ref0, ref1, ref2, ref3); + ref_ptr += ref_stride; + CALC_MSE_AVG_B(src0, ref0, var, avg0); + CALC_MSE_AVG_B(src1, ref1, var, avg1); + CALC_MSE_AVG_B(src2, ref2, var, avg2); + CALC_MSE_AVG_B(src3, ref3, var, avg3); + } + + vec = __msa_hadd_s_w(avg0, avg0); + vec += __msa_hadd_s_w(avg1, avg1); + vec += __msa_hadd_s_w(avg2, avg2); + vec += __msa_hadd_s_w(avg3, avg3); + *diff = HADD_SW_S32(vec); + + return HADD_SW_S32(var); +} + +static uint32_t get_mb_ss_msa(const int16_t *src) { + uint32_t sum, cnt; + v8i16 src0, src1, src2, src3; + v4i32 src0_l, src1_l, src2_l, src3_l; + v4i32 src0_r, src1_r, src2_r, src3_r; + v2i64 sq_src_l = { 0 }; + v2i64 sq_src_r = { 0 }; + + for (cnt = 8; cnt--;) { + LD_SH4(src, 8, src0, src1, src2, src3); + src += 4 * 8; + + UNPCK_SH_SW(src0, src0_l, src0_r); + UNPCK_SH_SW(src1, src1_l, src1_r); + UNPCK_SH_SW(src2, src2_l, src2_r); + UNPCK_SH_SW(src3, src3_l, src3_r); + + DPADD_SD2_SD(src0_l, src0_r, sq_src_l, sq_src_r); + DPADD_SD2_SD(src1_l, src1_r, sq_src_l, sq_src_r); + DPADD_SD2_SD(src2_l, src2_r, sq_src_l, sq_src_r); + DPADD_SD2_SD(src3_l, src3_r, sq_src_l, sq_src_r); + } + + sq_src_l += __msa_splati_d(sq_src_l, 1); + sq_src_r += __msa_splati_d(sq_src_r, 1); + + sum = __msa_copy_s_d(sq_src_l, 0); + sum += __msa_copy_s_d(sq_src_r, 0); + + return sum; +} + +static uint32_t sse_4width_msa(const uint8_t *src_ptr, int32_t src_stride, + const uint8_t *ref_ptr, int32_t ref_stride, + int32_t height) { + int32_t ht_cnt; + uint32_t src0, src1, src2, src3; + uint32_t ref0, ref1, ref2, ref3; + v16u8 src = { 0 }; + v16u8 ref = { 0 }; + v4i32 var = { 0 }; + + for (ht_cnt = (height >> 2); ht_cnt--;) { + LW4(src_ptr, src_stride, src0, src1, src2, src3); + src_ptr += (4 * src_stride); + LW4(ref_ptr, ref_stride, ref0, ref1, ref2, ref3); + ref_ptr += (4 * ref_stride); + + INSERT_W4_UB(src0, src1, src2, src3, src); + INSERT_W4_UB(ref0, ref1, ref2, ref3, ref); + CALC_MSE_B(src, ref, var); + } + + return HADD_SW_S32(var); +} + +static uint32_t sse_8width_msa(const uint8_t *src_ptr, int32_t src_stride, + const uint8_t *ref_ptr, int32_t ref_stride, + int32_t height) { + int32_t ht_cnt; + v16u8 src0, src1, src2, src3; + v16u8 ref0, ref1, ref2, ref3; + v4i32 var = { 0 }; + + for (ht_cnt = (height >> 2); ht_cnt--;) { + LD_UB4(src_ptr, src_stride, src0, src1, src2, src3); + src_ptr += (4 * src_stride); + LD_UB4(ref_ptr, ref_stride, ref0, ref1, ref2, ref3); + ref_ptr += (4 * ref_stride); + + PCKEV_D4_UB(src1, src0, src3, src2, ref1, ref0, ref3, ref2, src0, src1, + ref0, ref1); + CALC_MSE_B(src0, ref0, var); + CALC_MSE_B(src1, ref1, var); + } + + return HADD_SW_S32(var); +} + +static uint32_t sse_16width_msa(const uint8_t *src_ptr, int32_t src_stride, + const uint8_t *ref_ptr, int32_t ref_stride, + int32_t height) { + int32_t ht_cnt; + v16u8 src, ref; + v4i32 var = { 0 }; + + for (ht_cnt = (height >> 2); ht_cnt--;) { + src = LD_UB(src_ptr); + src_ptr += src_stride; + ref = LD_UB(ref_ptr); + ref_ptr += ref_stride; + CALC_MSE_B(src, ref, var); + + src = LD_UB(src_ptr); + src_ptr += src_stride; + ref = LD_UB(ref_ptr); + ref_ptr += ref_stride; + CALC_MSE_B(src, ref, var); + + src = LD_UB(src_ptr); + src_ptr += src_stride; + ref = LD_UB(ref_ptr); + ref_ptr += ref_stride; + CALC_MSE_B(src, ref, var); + + src = LD_UB(src_ptr); + src_ptr += src_stride; + ref = LD_UB(ref_ptr); + ref_ptr += ref_stride; + CALC_MSE_B(src, ref, var); + } + + return HADD_SW_S32(var); +} + +static uint32_t sse_32width_msa(const uint8_t *src_ptr, int32_t src_stride, + const uint8_t *ref_ptr, int32_t ref_stride, + int32_t height) { + int32_t ht_cnt; + v16u8 src0, src1, ref0, ref1; + v4i32 var = { 0 }; + + for (ht_cnt = (height >> 2); ht_cnt--;) { + LD_UB2(src_ptr, 16, src0, src1); + src_ptr += src_stride; + LD_UB2(ref_ptr, 16, ref0, ref1); + ref_ptr += ref_stride; + CALC_MSE_B(src0, ref0, var); + CALC_MSE_B(src1, ref1, var); + + LD_UB2(src_ptr, 16, src0, src1); + src_ptr += src_stride; + LD_UB2(ref_ptr, 16, ref0, ref1); + ref_ptr += ref_stride; + CALC_MSE_B(src0, ref0, var); + CALC_MSE_B(src1, ref1, var); + + LD_UB2(src_ptr, 16, src0, src1); + src_ptr += src_stride; + LD_UB2(ref_ptr, 16, ref0, ref1); + ref_ptr += ref_stride; + CALC_MSE_B(src0, ref0, var); + CALC_MSE_B(src1, ref1, var); + + LD_UB2(src_ptr, 16, src0, src1); + src_ptr += src_stride; + LD_UB2(ref_ptr, 16, ref0, ref1); + ref_ptr += ref_stride; + CALC_MSE_B(src0, ref0, var); + CALC_MSE_B(src1, ref1, var); + } + + return HADD_SW_S32(var); +} + +static uint32_t sse_64width_msa(const uint8_t *src_ptr, int32_t src_stride, + const uint8_t *ref_ptr, int32_t ref_stride, + int32_t height) { + int32_t ht_cnt; + v16u8 src0, src1, src2, src3; + v16u8 ref0, ref1, ref2, ref3; + v4i32 var = { 0 }; + + for (ht_cnt = height >> 1; ht_cnt--;) { + LD_UB4(src_ptr, 16, src0, src1, src2, src3); + src_ptr += src_stride; + LD_UB4(ref_ptr, 16, ref0, ref1, ref2, ref3); + ref_ptr += ref_stride; + CALC_MSE_B(src0, ref0, var); + CALC_MSE_B(src2, ref2, var); + CALC_MSE_B(src1, ref1, var); + CALC_MSE_B(src3, ref3, var); + + LD_UB4(src_ptr, 16, src0, src1, src2, src3); + src_ptr += src_stride; + LD_UB4(ref_ptr, 16, ref0, ref1, ref2, ref3); + ref_ptr += ref_stride; + CALC_MSE_B(src0, ref0, var); + CALC_MSE_B(src2, ref2, var); + CALC_MSE_B(src1, ref1, var); + CALC_MSE_B(src3, ref3, var); + } + + return HADD_SW_S32(var); +} + +uint32_t aom_get4x4sse_cs_msa(const uint8_t *src_ptr, int32_t src_stride, + const uint8_t *ref_ptr, int32_t ref_stride) { + uint32_t err = 0; + uint32_t src0, src1, src2, src3; + uint32_t ref0, ref1, ref2, ref3; + v16i8 src = { 0 }; + v16i8 ref = { 0 }; + v16u8 src_vec0, src_vec1; + v8i16 diff0, diff1; + v4i32 err0 = { 0 }; + v4i32 err1 = { 0 }; + + LW4(src_ptr, src_stride, src0, src1, src2, src3); + LW4(ref_ptr, ref_stride, ref0, ref1, ref2, ref3); + INSERT_W4_SB(src0, src1, src2, src3, src); + INSERT_W4_SB(ref0, ref1, ref2, ref3, ref); + ILVRL_B2_UB(src, ref, src_vec0, src_vec1); + HSUB_UB2_SH(src_vec0, src_vec1, diff0, diff1); + DPADD_SH2_SW(diff0, diff1, diff0, diff1, err0, err1); + err = HADD_SW_S32(err0); + err += HADD_SW_S32(err1); + + return err; +} + +#define VARIANCE_4Wx4H(sse, diff) VARIANCE_WxH(sse, diff, 4); +#define VARIANCE_4Wx8H(sse, diff) VARIANCE_WxH(sse, diff, 5); +#define VARIANCE_8Wx4H(sse, diff) VARIANCE_WxH(sse, diff, 5); +#define VARIANCE_8Wx8H(sse, diff) VARIANCE_WxH(sse, diff, 6); +#define VARIANCE_8Wx16H(sse, diff) VARIANCE_WxH(sse, diff, 7); +#define VARIANCE_16Wx8H(sse, diff) VARIANCE_WxH(sse, diff, 7); +#define VARIANCE_16Wx16H(sse, diff) VARIANCE_WxH(sse, diff, 8); + +#define VARIANCE_16Wx32H(sse, diff) VARIANCE_LARGE_WxH(sse, diff, 9); +#define VARIANCE_32Wx16H(sse, diff) VARIANCE_LARGE_WxH(sse, diff, 9); +#define VARIANCE_32Wx32H(sse, diff) VARIANCE_LARGE_WxH(sse, diff, 10); +#define VARIANCE_32Wx64H(sse, diff) VARIANCE_LARGE_WxH(sse, diff, 11); +#define VARIANCE_64Wx32H(sse, diff) VARIANCE_LARGE_WxH(sse, diff, 11); +#define VARIANCE_64Wx64H(sse, diff) VARIANCE_LARGE_WxH(sse, diff, 12); + +#define AOM_VARIANCE_WDXHT_MSA(wd, ht) \ + uint32_t aom_variance##wd##x##ht##_msa( \ + const uint8_t *src, int32_t src_stride, const uint8_t *ref, \ + int32_t ref_stride, uint32_t *sse) { \ + int32_t diff; \ + \ + *sse = \ + sse_diff_##wd##width_msa(src, src_stride, ref, ref_stride, ht, &diff); \ + \ + return VARIANCE_##wd##Wx##ht##H(*sse, diff); \ + } + +/* clang-format off */ +AOM_VARIANCE_WDXHT_MSA(4, 4) +AOM_VARIANCE_WDXHT_MSA(4, 8) + +AOM_VARIANCE_WDXHT_MSA(8, 4) +AOM_VARIANCE_WDXHT_MSA(8, 8) +AOM_VARIANCE_WDXHT_MSA(8, 16) + +AOM_VARIANCE_WDXHT_MSA(16, 8) +AOM_VARIANCE_WDXHT_MSA(16, 16) +AOM_VARIANCE_WDXHT_MSA(16, 32) + +AOM_VARIANCE_WDXHT_MSA(32, 16) +AOM_VARIANCE_WDXHT_MSA(32, 32) +/* clang-format on */ + +uint32_t aom_variance32x64_msa(const uint8_t *src, int32_t src_stride, + const uint8_t *ref, int32_t ref_stride, + uint32_t *sse) { + int32_t diff; + + *sse = sse_diff_32x64_msa(src, src_stride, ref, ref_stride, &diff); + + return VARIANCE_32Wx64H(*sse, diff); +} + +uint32_t aom_variance64x32_msa(const uint8_t *src, int32_t src_stride, + const uint8_t *ref, int32_t ref_stride, + uint32_t *sse) { + int32_t diff; + + *sse = sse_diff_64x32_msa(src, src_stride, ref, ref_stride, &diff); + + return VARIANCE_64Wx32H(*sse, diff); +} + +uint32_t aom_variance64x64_msa(const uint8_t *src, int32_t src_stride, + const uint8_t *ref, int32_t ref_stride, + uint32_t *sse) { + int32_t diff; + + *sse = sse_diff_64x64_msa(src, src_stride, ref, ref_stride, &diff); + + return VARIANCE_64Wx64H(*sse, diff); +} + +uint32_t aom_mse8x8_msa(const uint8_t *src, int32_t src_stride, + const uint8_t *ref, int32_t ref_stride, uint32_t *sse) { + *sse = sse_8width_msa(src, src_stride, ref, ref_stride, 8); + + return *sse; +} + +uint32_t aom_mse8x16_msa(const uint8_t *src, int32_t src_stride, + const uint8_t *ref, int32_t ref_stride, + uint32_t *sse) { + *sse = sse_8width_msa(src, src_stride, ref, ref_stride, 16); + + return *sse; +} + +uint32_t aom_mse16x8_msa(const uint8_t *src, int32_t src_stride, + const uint8_t *ref, int32_t ref_stride, + uint32_t *sse) { + *sse = sse_16width_msa(src, src_stride, ref, ref_stride, 8); + + return *sse; +} + +uint32_t aom_mse16x16_msa(const uint8_t *src, int32_t src_stride, + const uint8_t *ref, int32_t ref_stride, + uint32_t *sse) { + *sse = sse_16width_msa(src, src_stride, ref, ref_stride, 16); + + return *sse; +} + +void aom_get8x8var_msa(const uint8_t *src, int32_t src_stride, + const uint8_t *ref, int32_t ref_stride, uint32_t *sse, + int32_t *sum) { + *sse = sse_diff_8width_msa(src, src_stride, ref, ref_stride, 8, sum); +} + +void aom_get16x16var_msa(const uint8_t *src, int32_t src_stride, + const uint8_t *ref, int32_t ref_stride, uint32_t *sse, + int32_t *sum) { + *sse = sse_diff_16width_msa(src, src_stride, ref, ref_stride, 16, sum); +} + +uint32_t aom_get_mb_ss_msa(const int16_t *src) { return get_mb_ss_msa(src); } diff --git a/third_party/aom/aom_dsp/noise_model.c b/third_party/aom/aom_dsp/noise_model.c new file mode 100644 index 0000000000..2faee85063 --- /dev/null +++ b/third_party/aom/aom_dsp/noise_model.c @@ -0,0 +1,1648 @@ +/* + * Copyright (c) 2017, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include +#include +#include +#include + +#include "aom_dsp/aom_dsp_common.h" +#include "aom_dsp/noise_model.h" +#include "aom_dsp/noise_util.h" +#include "aom_mem/aom_mem.h" +#include "av1/common/common.h" +#include "av1/encoder/mathutils.h" + +#define kLowPolyNumParams 3 + +static const int kMaxLag = 4; + +// Defines a function that can be used to obtain the mean of a block for the +// provided data type (uint8_t, or uint16_t) +#define GET_BLOCK_MEAN(INT_TYPE, suffix) \ + static double get_block_mean_##suffix(const INT_TYPE *data, int w, int h, \ + int stride, int x_o, int y_o, \ + int block_size) { \ + const int max_h = AOMMIN(h - y_o, block_size); \ + const int max_w = AOMMIN(w - x_o, block_size); \ + double block_mean = 0; \ + for (int y = 0; y < max_h; ++y) { \ + for (int x = 0; x < max_w; ++x) { \ + block_mean += data[(y_o + y) * stride + x_o + x]; \ + } \ + } \ + return block_mean / (max_w * max_h); \ + } + +GET_BLOCK_MEAN(uint8_t, lowbd); +GET_BLOCK_MEAN(uint16_t, highbd); + +static INLINE double get_block_mean(const uint8_t *data, int w, int h, + int stride, int x_o, int y_o, + int block_size, int use_highbd) { + if (use_highbd) + return get_block_mean_highbd((const uint16_t *)data, w, h, stride, x_o, y_o, + block_size); + return get_block_mean_lowbd(data, w, h, stride, x_o, y_o, block_size); +} + +// Defines a function that can be used to obtain the variance of a block +// for the provided data type (uint8_t, or uint16_t) +#define GET_NOISE_VAR(INT_TYPE, suffix) \ + static double get_noise_var_##suffix( \ + const INT_TYPE *data, const INT_TYPE *denoised, int stride, int w, \ + int h, int x_o, int y_o, int block_size_x, int block_size_y) { \ + const int max_h = AOMMIN(h - y_o, block_size_y); \ + const int max_w = AOMMIN(w - x_o, block_size_x); \ + double noise_var = 0; \ + double noise_mean = 0; \ + for (int y = 0; y < max_h; ++y) { \ + for (int x = 0; x < max_w; ++x) { \ + double noise = (double)data[(y_o + y) * stride + x_o + x] - \ + denoised[(y_o + y) * stride + x_o + x]; \ + noise_mean += noise; \ + noise_var += noise * noise; \ + } \ + } \ + noise_mean /= (max_w * max_h); \ + return noise_var / (max_w * max_h) - noise_mean * noise_mean; \ + } + +GET_NOISE_VAR(uint8_t, lowbd); +GET_NOISE_VAR(uint16_t, highbd); + +static INLINE double get_noise_var(const uint8_t *data, const uint8_t *denoised, + int w, int h, int stride, int x_o, int y_o, + int block_size_x, int block_size_y, + int use_highbd) { + if (use_highbd) + return get_noise_var_highbd((const uint16_t *)data, + (const uint16_t *)denoised, w, h, stride, x_o, + y_o, block_size_x, block_size_y); + return get_noise_var_lowbd(data, denoised, w, h, stride, x_o, y_o, + block_size_x, block_size_y); +} + +static void equation_system_clear(aom_equation_system_t *eqns) { + const int n = eqns->n; + memset(eqns->A, 0, sizeof(*eqns->A) * n * n); + memset(eqns->x, 0, sizeof(*eqns->x) * n); + memset(eqns->b, 0, sizeof(*eqns->b) * n); +} + +static void equation_system_copy(aom_equation_system_t *dst, + const aom_equation_system_t *src) { + const int n = dst->n; + memcpy(dst->A, src->A, sizeof(*dst->A) * n * n); + memcpy(dst->x, src->x, sizeof(*dst->x) * n); + memcpy(dst->b, src->b, sizeof(*dst->b) * n); +} + +static int equation_system_init(aom_equation_system_t *eqns, int n) { + eqns->A = (double *)aom_malloc(sizeof(*eqns->A) * n * n); + eqns->b = (double *)aom_malloc(sizeof(*eqns->b) * n); + eqns->x = (double *)aom_malloc(sizeof(*eqns->x) * n); + eqns->n = n; + if (!eqns->A || !eqns->b || !eqns->x) { + fprintf(stderr, "Failed to allocate system of equations of size %d\n", n); + aom_free(eqns->A); + aom_free(eqns->b); + aom_free(eqns->x); + memset(eqns, 0, sizeof(*eqns)); + return 0; + } + equation_system_clear(eqns); + return 1; +} + +static int equation_system_solve(aom_equation_system_t *eqns) { + const int n = eqns->n; + double *b = (double *)aom_malloc(sizeof(*b) * n); + double *A = (double *)aom_malloc(sizeof(*A) * n * n); + int ret = 0; + if (A == NULL || b == NULL) { + fprintf(stderr, "Unable to allocate temp values of size %dx%d\n", n, n); + aom_free(b); + aom_free(A); + return 0; + } + memcpy(A, eqns->A, sizeof(*eqns->A) * n * n); + memcpy(b, eqns->b, sizeof(*eqns->b) * n); + ret = linsolve(n, A, eqns->n, b, eqns->x); + aom_free(b); + aom_free(A); + + if (ret == 0) { + return 0; + } + return 1; +} + +static void equation_system_add(aom_equation_system_t *dest, + aom_equation_system_t *src) { + const int n = dest->n; + int i, j; + for (i = 0; i < n; ++i) { + for (j = 0; j < n; ++j) { + dest->A[i * n + j] += src->A[i * n + j]; + } + dest->b[i] += src->b[i]; + } +} + +static void equation_system_free(aom_equation_system_t *eqns) { + if (!eqns) return; + aom_free(eqns->A); + aom_free(eqns->b); + aom_free(eqns->x); + memset(eqns, 0, sizeof(*eqns)); +} + +static void noise_strength_solver_clear(aom_noise_strength_solver_t *solver) { + equation_system_clear(&solver->eqns); + solver->num_equations = 0; + solver->total = 0; +} + +static void noise_strength_solver_add(aom_noise_strength_solver_t *dest, + aom_noise_strength_solver_t *src) { + equation_system_add(&dest->eqns, &src->eqns); + dest->num_equations += src->num_equations; + dest->total += src->total; +} + +// Return the number of coefficients required for the given parameters +static int num_coeffs(const aom_noise_model_params_t params) { + const int n = 2 * params.lag + 1; + switch (params.shape) { + case AOM_NOISE_SHAPE_DIAMOND: return params.lag * (params.lag + 1); + case AOM_NOISE_SHAPE_SQUARE: return (n * n) / 2; + } + return 0; +} + +static int noise_state_init(aom_noise_state_t *state, int n, int bit_depth) { + const int kNumBins = 20; + if (!equation_system_init(&state->eqns, n)) { + fprintf(stderr, "Failed initialization noise state with size %d\n", n); + return 0; + } + state->ar_gain = 1.0; + state->num_observations = 0; + return aom_noise_strength_solver_init(&state->strength_solver, kNumBins, + bit_depth); +} + +static void set_chroma_coefficient_fallback_soln(aom_equation_system_t *eqns) { + const double kTolerance = 1e-6; + const int last = eqns->n - 1; + // Set all of the AR coefficients to zero, but try to solve for correlation + // with the luma channel + memset(eqns->x, 0, sizeof(*eqns->x) * eqns->n); + if (fabs(eqns->A[last * eqns->n + last]) > kTolerance) { + eqns->x[last] = eqns->b[last] / eqns->A[last * eqns->n + last]; + } +} + +int aom_noise_strength_lut_init(aom_noise_strength_lut_t *lut, int num_points) { + if (!lut) return 0; + lut->points = (double(*)[2])aom_malloc(num_points * sizeof(*lut->points)); + if (!lut->points) return 0; + lut->num_points = num_points; + memset(lut->points, 0, sizeof(*lut->points) * num_points); + return 1; +} + +void aom_noise_strength_lut_free(aom_noise_strength_lut_t *lut) { + if (!lut) return; + aom_free(lut->points); + memset(lut, 0, sizeof(*lut)); +} + +double aom_noise_strength_lut_eval(const aom_noise_strength_lut_t *lut, + double x) { + int i = 0; + // Constant extrapolation for x < x_0. + if (x < lut->points[0][0]) return lut->points[0][1]; + for (i = 0; i < lut->num_points - 1; ++i) { + if (x >= lut->points[i][0] && x <= lut->points[i + 1][0]) { + const double a = + (x - lut->points[i][0]) / (lut->points[i + 1][0] - lut->points[i][0]); + return lut->points[i + 1][1] * a + lut->points[i][1] * (1.0 - a); + } + } + // Constant extrapolation for x > x_{n-1} + return lut->points[lut->num_points - 1][1]; +} + +static double noise_strength_solver_get_bin_index( + const aom_noise_strength_solver_t *solver, double value) { + const double val = + fclamp(value, solver->min_intensity, solver->max_intensity); + const double range = solver->max_intensity - solver->min_intensity; + return (solver->num_bins - 1) * (val - solver->min_intensity) / range; +} + +static double noise_strength_solver_get_value( + const aom_noise_strength_solver_t *solver, double x) { + const double bin = noise_strength_solver_get_bin_index(solver, x); + const int bin_i0 = (int)floor(bin); + const int bin_i1 = AOMMIN(solver->num_bins - 1, bin_i0 + 1); + const double a = bin - bin_i0; + return (1.0 - a) * solver->eqns.x[bin_i0] + a * solver->eqns.x[bin_i1]; +} + +void aom_noise_strength_solver_add_measurement( + aom_noise_strength_solver_t *solver, double block_mean, double noise_std) { + const double bin = noise_strength_solver_get_bin_index(solver, block_mean); + const int bin_i0 = (int)floor(bin); + const int bin_i1 = AOMMIN(solver->num_bins - 1, bin_i0 + 1); + const double a = bin - bin_i0; + const int n = solver->num_bins; + solver->eqns.A[bin_i0 * n + bin_i0] += (1.0 - a) * (1.0 - a); + solver->eqns.A[bin_i1 * n + bin_i0] += a * (1.0 - a); + solver->eqns.A[bin_i1 * n + bin_i1] += a * a; + solver->eqns.A[bin_i0 * n + bin_i1] += a * (1.0 - a); + solver->eqns.b[bin_i0] += (1.0 - a) * noise_std; + solver->eqns.b[bin_i1] += a * noise_std; + solver->total += noise_std; + solver->num_equations++; +} + +int aom_noise_strength_solver_solve(aom_noise_strength_solver_t *solver) { + // Add regularization proportional to the number of constraints + const int n = solver->num_bins; + const double kAlpha = 2.0 * (double)(solver->num_equations) / n; + int result = 0; + double mean = 0; + + // Do this in a non-destructive manner so it is not confusing to the caller + double *old_A = solver->eqns.A; + double *A = (double *)aom_malloc(sizeof(*A) * n * n); + if (!A) { + fprintf(stderr, "Unable to allocate copy of A\n"); + return 0; + } + memcpy(A, old_A, sizeof(*A) * n * n); + + for (int i = 0; i < n; ++i) { + const int i_lo = AOMMAX(0, i - 1); + const int i_hi = AOMMIN(n - 1, i + 1); + A[i * n + i_lo] -= kAlpha; + A[i * n + i] += 2 * kAlpha; + A[i * n + i_hi] -= kAlpha; + } + + // Small regularization to give average noise strength + mean = solver->total / solver->num_equations; + for (int i = 0; i < n; ++i) { + A[i * n + i] += 1.0 / 8192.; + solver->eqns.b[i] += mean / 8192.; + } + solver->eqns.A = A; + result = equation_system_solve(&solver->eqns); + solver->eqns.A = old_A; + + aom_free(A); + return result; +} + +int aom_noise_strength_solver_init(aom_noise_strength_solver_t *solver, + int num_bins, int bit_depth) { + if (!solver) return 0; + memset(solver, 0, sizeof(*solver)); + solver->num_bins = num_bins; + solver->min_intensity = 0; + solver->max_intensity = (1 << bit_depth) - 1; + solver->total = 0; + solver->num_equations = 0; + return equation_system_init(&solver->eqns, num_bins); +} + +void aom_noise_strength_solver_free(aom_noise_strength_solver_t *solver) { + if (!solver) return; + equation_system_free(&solver->eqns); +} + +double aom_noise_strength_solver_get_center( + const aom_noise_strength_solver_t *solver, int i) { + const double range = solver->max_intensity - solver->min_intensity; + const int n = solver->num_bins; + return ((double)i) / (n - 1) * range + solver->min_intensity; +} + +// Computes the residual if a point were to be removed from the lut. This is +// calculated as the area between the output of the solver and the line segment +// that would be formed between [x_{i - 1}, x_{i + 1}). +static void update_piecewise_linear_residual( + const aom_noise_strength_solver_t *solver, + const aom_noise_strength_lut_t *lut, double *residual, int start, int end) { + const double dx = 255. / solver->num_bins; + for (int i = AOMMAX(start, 1); i < AOMMIN(end, lut->num_points - 1); ++i) { + const int lower = AOMMAX(0, (int)floor(noise_strength_solver_get_bin_index( + solver, lut->points[i - 1][0]))); + const int upper = AOMMIN(solver->num_bins - 1, + (int)ceil(noise_strength_solver_get_bin_index( + solver, lut->points[i + 1][0]))); + double r = 0; + for (int j = lower; j <= upper; ++j) { + const double x = aom_noise_strength_solver_get_center(solver, j); + if (x < lut->points[i - 1][0]) continue; + if (x >= lut->points[i + 1][0]) continue; + const double y = solver->eqns.x[j]; + const double a = (x - lut->points[i - 1][0]) / + (lut->points[i + 1][0] - lut->points[i - 1][0]); + const double estimate_y = + lut->points[i - 1][1] * (1.0 - a) + lut->points[i + 1][1] * a; + r += fabs(y - estimate_y); + } + residual[i] = r * dx; + } +} + +int aom_noise_strength_solver_fit_piecewise( + const aom_noise_strength_solver_t *solver, int max_output_points, + aom_noise_strength_lut_t *lut) { + // The tolerance is normalized to be give consistent results between + // different bit-depths. + const double kTolerance = solver->max_intensity * 0.00625 / 255.0; + if (!aom_noise_strength_lut_init(lut, solver->num_bins)) { + fprintf(stderr, "Failed to init lut\n"); + return 0; + } + for (int i = 0; i < solver->num_bins; ++i) { + lut->points[i][0] = aom_noise_strength_solver_get_center(solver, i); + lut->points[i][1] = solver->eqns.x[i]; + } + if (max_output_points < 0) { + max_output_points = solver->num_bins; + } + + double *residual = aom_malloc(solver->num_bins * sizeof(*residual)); + memset(residual, 0, sizeof(*residual) * solver->num_bins); + + update_piecewise_linear_residual(solver, lut, residual, 0, solver->num_bins); + + // Greedily remove points if there are too many or if it doesn't hurt local + // approximation (never remove the end points) + while (lut->num_points > 2) { + int min_index = 1; + for (int j = 1; j < lut->num_points - 1; ++j) { + if (residual[j] < residual[min_index]) { + min_index = j; + } + } + const double dx = + lut->points[min_index + 1][0] - lut->points[min_index - 1][0]; + const double avg_residual = residual[min_index] / dx; + if (lut->num_points <= max_output_points && avg_residual > kTolerance) { + break; + } + + const int num_remaining = lut->num_points - min_index - 1; + memmove(lut->points + min_index, lut->points + min_index + 1, + sizeof(lut->points[0]) * num_remaining); + lut->num_points--; + + update_piecewise_linear_residual(solver, lut, residual, min_index - 1, + min_index + 1); + } + aom_free(residual); + return 1; +} + +int aom_flat_block_finder_init(aom_flat_block_finder_t *block_finder, + int block_size, int bit_depth, int use_highbd) { + const int n = block_size * block_size; + aom_equation_system_t eqns; + double *AtA_inv = 0; + double *A = 0; + int x = 0, y = 0, i = 0, j = 0; + if (!equation_system_init(&eqns, kLowPolyNumParams)) { + fprintf(stderr, "Failed to init equation system for block_size=%d\n", + block_size); + return 0; + } + + AtA_inv = (double *)aom_malloc(kLowPolyNumParams * kLowPolyNumParams * + sizeof(*AtA_inv)); + A = (double *)aom_malloc(kLowPolyNumParams * n * sizeof(*A)); + if (AtA_inv == NULL || A == NULL) { + fprintf(stderr, "Failed to alloc A or AtA_inv for block_size=%d\n", + block_size); + aom_free(AtA_inv); + aom_free(A); + equation_system_free(&eqns); + return 0; + } + + block_finder->A = A; + block_finder->AtA_inv = AtA_inv; + block_finder->block_size = block_size; + block_finder->normalization = (1 << bit_depth) - 1; + block_finder->use_highbd = use_highbd; + + for (y = 0; y < block_size; ++y) { + const double yd = ((double)y - block_size / 2.) / (block_size / 2.); + for (x = 0; x < block_size; ++x) { + const double xd = ((double)x - block_size / 2.) / (block_size / 2.); + const double coords[3] = { yd, xd, 1 }; + const int row = y * block_size + x; + A[kLowPolyNumParams * row + 0] = yd; + A[kLowPolyNumParams * row + 1] = xd; + A[kLowPolyNumParams * row + 2] = 1; + + for (i = 0; i < kLowPolyNumParams; ++i) { + for (j = 0; j < kLowPolyNumParams; ++j) { + eqns.A[kLowPolyNumParams * i + j] += coords[i] * coords[j]; + } + } + } + } + + // Lazy inverse using existing equation solver. + for (i = 0; i < kLowPolyNumParams; ++i) { + memset(eqns.b, 0, sizeof(*eqns.b) * kLowPolyNumParams); + eqns.b[i] = 1; + equation_system_solve(&eqns); + + for (j = 0; j < kLowPolyNumParams; ++j) { + AtA_inv[j * kLowPolyNumParams + i] = eqns.x[j]; + } + } + equation_system_free(&eqns); + return 1; +} + +void aom_flat_block_finder_free(aom_flat_block_finder_t *block_finder) { + if (!block_finder) return; + aom_free(block_finder->A); + aom_free(block_finder->AtA_inv); + memset(block_finder, 0, sizeof(*block_finder)); +} + +void aom_flat_block_finder_extract_block( + const aom_flat_block_finder_t *block_finder, const uint8_t *const data, + int w, int h, int stride, int offsx, int offsy, double *plane, + double *block) { + const int block_size = block_finder->block_size; + const int n = block_size * block_size; + const double *A = block_finder->A; + const double *AtA_inv = block_finder->AtA_inv; + double plane_coords[kLowPolyNumParams]; + double AtA_inv_b[kLowPolyNumParams]; + int xi, yi, i; + + if (block_finder->use_highbd) { + const uint16_t *const data16 = (const uint16_t *const)data; + for (yi = 0; yi < block_size; ++yi) { + const int y = clamp(offsy + yi, 0, h - 1); + for (xi = 0; xi < block_size; ++xi) { + const int x = clamp(offsx + xi, 0, w - 1); + block[yi * block_size + xi] = + ((double)data16[y * stride + x]) / block_finder->normalization; + } + } + } else { + for (yi = 0; yi < block_size; ++yi) { + const int y = clamp(offsy + yi, 0, h - 1); + for (xi = 0; xi < block_size; ++xi) { + const int x = clamp(offsx + xi, 0, w - 1); + block[yi * block_size + xi] = + ((double)data[y * stride + x]) / block_finder->normalization; + } + } + } + multiply_mat(block, A, AtA_inv_b, 1, n, kLowPolyNumParams); + multiply_mat(AtA_inv, AtA_inv_b, plane_coords, kLowPolyNumParams, + kLowPolyNumParams, 1); + multiply_mat(A, plane_coords, plane, n, kLowPolyNumParams, 1); + + for (i = 0; i < n; ++i) { + block[i] -= plane[i]; + } +} + +typedef struct { + int index; + float score; +} index_and_score_t; + +static int compare_scores(const void *a, const void *b) { + const float diff = + ((index_and_score_t *)a)->score - ((index_and_score_t *)b)->score; + if (diff < 0) + return -1; + else if (diff > 0) + return 1; + return 0; +} + +int aom_flat_block_finder_run(const aom_flat_block_finder_t *block_finder, + const uint8_t *const data, int w, int h, + int stride, uint8_t *flat_blocks) { + // The gradient-based features used in this code are based on: + // A. Kokaram, D. Kelly, H. Denman and A. Crawford, "Measuring noise + // correlation for improved video denoising," 2012 19th, ICIP. + // The thresholds are more lenient to allow for correct grain modeling + // if extreme cases. + const int block_size = block_finder->block_size; + const int n = block_size * block_size; + const double kTraceThreshold = 0.15 / (32 * 32); + const double kRatioThreshold = 1.25; + const double kNormThreshold = 0.08 / (32 * 32); + const double kVarThreshold = 0.005 / (double)n; + const int num_blocks_w = (w + block_size - 1) / block_size; + const int num_blocks_h = (h + block_size - 1) / block_size; + int num_flat = 0; + int bx = 0, by = 0; + double *plane = (double *)aom_malloc(n * sizeof(*plane)); + double *block = (double *)aom_malloc(n * sizeof(*block)); + index_and_score_t *scores = (index_and_score_t *)aom_malloc( + num_blocks_w * num_blocks_h * sizeof(*scores)); + if (plane == NULL || block == NULL || scores == NULL) { + fprintf(stderr, "Failed to allocate memory for block of size %d\n", n); + aom_free(plane); + aom_free(block); + aom_free(scores); + return -1; + } + +#ifdef NOISE_MODEL_LOG_SCORE + fprintf(stderr, "score = ["); +#endif + for (by = 0; by < num_blocks_h; ++by) { + for (bx = 0; bx < num_blocks_w; ++bx) { + // Compute gradient covariance matrix. + double Gxx = 0, Gxy = 0, Gyy = 0; + double var = 0; + double mean = 0; + int xi, yi; + aom_flat_block_finder_extract_block(block_finder, data, w, h, stride, + bx * block_size, by * block_size, + plane, block); + + for (yi = 1; yi < block_size - 1; ++yi) { + for (xi = 1; xi < block_size - 1; ++xi) { + const double gx = (block[yi * block_size + xi + 1] - + block[yi * block_size + xi - 1]) / + 2; + const double gy = (block[yi * block_size + xi + block_size] - + block[yi * block_size + xi - block_size]) / + 2; + Gxx += gx * gx; + Gxy += gx * gy; + Gyy += gy * gy; + + mean += block[yi * block_size + xi]; + var += block[yi * block_size + xi] * block[yi * block_size + xi]; + } + } + mean /= (block_size - 2) * (block_size - 2); + + // Normalize gradients by block_size. + Gxx /= ((block_size - 2) * (block_size - 2)); + Gxy /= ((block_size - 2) * (block_size - 2)); + Gyy /= ((block_size - 2) * (block_size - 2)); + var = var / ((block_size - 2) * (block_size - 2)) - mean * mean; + + { + const double trace = Gxx + Gyy; + const double det = Gxx * Gyy - Gxy * Gxy; + const double e1 = (trace + sqrt(trace * trace - 4 * det)) / 2.; + const double e2 = (trace - sqrt(trace * trace - 4 * det)) / 2.; + const double norm = e1; // Spectral norm + const double ratio = (e1 / AOMMAX(e2, 1e-6)); + const int is_flat = (trace < kTraceThreshold) && + (ratio < kRatioThreshold) && + (norm < kNormThreshold) && (var > kVarThreshold); + // The following weights are used to combine the above features to give + // a sigmoid score for flatness. If the input was normalized to [0,100] + // the magnitude of these values would be close to 1 (e.g., weights + // corresponding to variance would be a factor of 10000x smaller). + // The weights are given in the following order: + // [{var}, {ratio}, {trace}, {norm}, offset] + // with one of the most discriminative being simply the variance. + const double weights[5] = { -6682, -0.2056, 13087, -12434, 2.5694 }; + const float score = + (float)(1.0 / (1 + exp(-(weights[0] * var + weights[1] * ratio + + weights[2] * trace + weights[3] * norm + + weights[4])))); + flat_blocks[by * num_blocks_w + bx] = is_flat ? 255 : 0; + scores[by * num_blocks_w + bx].score = var > kVarThreshold ? score : 0; + scores[by * num_blocks_w + bx].index = by * num_blocks_w + bx; +#ifdef NOISE_MODEL_LOG_SCORE + fprintf(stderr, "%g %g %g %g %g %d ", score, var, ratio, trace, norm, + is_flat); +#endif + num_flat += is_flat; + } + } +#ifdef NOISE_MODEL_LOG_SCORE + fprintf(stderr, "\n"); +#endif + } +#ifdef NOISE_MODEL_LOG_SCORE + fprintf(stderr, "];\n"); +#endif + // Find the top-scored blocks (most likely to be flat) and set the flat blocks + // be the union of the thresholded results and the top 10th percentile of the + // scored results. + qsort(scores, num_blocks_w * num_blocks_h, sizeof(*scores), &compare_scores); + const int top_nth_percentile = num_blocks_w * num_blocks_h * 90 / 100; + const float score_threshold = scores[top_nth_percentile].score; + for (int i = 0; i < num_blocks_w * num_blocks_h; ++i) { + if (scores[i].score >= score_threshold) { + num_flat += flat_blocks[scores[i].index] == 0; + flat_blocks[scores[i].index] |= 1; + } + } + aom_free(block); + aom_free(plane); + aom_free(scores); + return num_flat; +} + +int aom_noise_model_init(aom_noise_model_t *model, + const aom_noise_model_params_t params) { + const int n = num_coeffs(params); + const int lag = params.lag; + const int bit_depth = params.bit_depth; + int x = 0, y = 0, i = 0, c = 0; + + memset(model, 0, sizeof(*model)); + if (params.lag < 1) { + fprintf(stderr, "Invalid noise param: lag = %d must be >= 1\n", params.lag); + return 0; + } + if (params.lag > kMaxLag) { + fprintf(stderr, "Invalid noise param: lag = %d must be <= %d\n", params.lag, + kMaxLag); + return 0; + } + + memcpy(&model->params, ¶ms, sizeof(params)); + for (c = 0; c < 3; ++c) { + if (!noise_state_init(&model->combined_state[c], n + (c > 0), bit_depth)) { + fprintf(stderr, "Failed to allocate noise state for channel %d\n", c); + aom_noise_model_free(model); + return 0; + } + if (!noise_state_init(&model->latest_state[c], n + (c > 0), bit_depth)) { + fprintf(stderr, "Failed to allocate noise state for channel %d\n", c); + aom_noise_model_free(model); + return 0; + } + } + model->n = n; + model->coords = (int(*)[2])aom_malloc(sizeof(*model->coords) * n); + + for (y = -lag; y <= 0; ++y) { + const int max_x = y == 0 ? -1 : lag; + for (x = -lag; x <= max_x; ++x) { + switch (params.shape) { + case AOM_NOISE_SHAPE_DIAMOND: + if (abs(x) <= y + lag) { + model->coords[i][0] = x; + model->coords[i][1] = y; + ++i; + } + break; + case AOM_NOISE_SHAPE_SQUARE: + model->coords[i][0] = x; + model->coords[i][1] = y; + ++i; + break; + default: + fprintf(stderr, "Invalid shape\n"); + aom_noise_model_free(model); + return 0; + } + } + } + assert(i == n); + return 1; +} + +void aom_noise_model_free(aom_noise_model_t *model) { + int c = 0; + if (!model) return; + + aom_free(model->coords); + for (c = 0; c < 3; ++c) { + equation_system_free(&model->latest_state[c].eqns); + equation_system_free(&model->combined_state[c].eqns); + + equation_system_free(&model->latest_state[c].strength_solver.eqns); + equation_system_free(&model->combined_state[c].strength_solver.eqns); + } + memset(model, 0, sizeof(*model)); +} + +// Extracts the neighborhood defined by coords around point (x, y) from +// the difference between the data and denoised images. Also extracts the +// entry (possibly downsampled) for (x, y) in the alt_data (e.g., luma). +#define EXTRACT_AR_ROW(INT_TYPE, suffix) \ + static double extract_ar_row_##suffix( \ + int(*coords)[2], int num_coords, const INT_TYPE *const data, \ + const INT_TYPE *const denoised, int stride, int sub_log2[2], \ + const INT_TYPE *const alt_data, const INT_TYPE *const alt_denoised, \ + int alt_stride, int x, int y, double *buffer) { \ + for (int i = 0; i < num_coords; ++i) { \ + const int x_i = x + coords[i][0], y_i = y + coords[i][1]; \ + buffer[i] = \ + (double)data[y_i * stride + x_i] - denoised[y_i * stride + x_i]; \ + } \ + const double val = \ + (double)data[y * stride + x] - denoised[y * stride + x]; \ + \ + if (alt_data && alt_denoised) { \ + double avg_data = 0, avg_denoised = 0; \ + int num_samples = 0; \ + for (int dy_i = 0; dy_i < (1 << sub_log2[1]); dy_i++) { \ + const int y_up = (y << sub_log2[1]) + dy_i; \ + for (int dx_i = 0; dx_i < (1 << sub_log2[0]); dx_i++) { \ + const int x_up = (x << sub_log2[0]) + dx_i; \ + avg_data += alt_data[y_up * alt_stride + x_up]; \ + avg_denoised += alt_denoised[y_up * alt_stride + x_up]; \ + num_samples++; \ + } \ + } \ + buffer[num_coords] = (avg_data - avg_denoised) / num_samples; \ + } \ + return val; \ + } + +EXTRACT_AR_ROW(uint8_t, lowbd); +EXTRACT_AR_ROW(uint16_t, highbd); + +static int add_block_observations( + aom_noise_model_t *noise_model, int c, const uint8_t *const data, + const uint8_t *const denoised, int w, int h, int stride, int sub_log2[2], + const uint8_t *const alt_data, const uint8_t *const alt_denoised, + int alt_stride, const uint8_t *const flat_blocks, int block_size, + int num_blocks_w, int num_blocks_h) { + const int lag = noise_model->params.lag; + const int num_coords = noise_model->n; + const double normalization = (1 << noise_model->params.bit_depth) - 1; + double *A = noise_model->latest_state[c].eqns.A; + double *b = noise_model->latest_state[c].eqns.b; + double *buffer = (double *)aom_malloc(sizeof(*buffer) * (num_coords + 1)); + const int n = noise_model->latest_state[c].eqns.n; + + if (!buffer) { + fprintf(stderr, "Unable to allocate buffer of size %d\n", num_coords + 1); + return 0; + } + for (int by = 0; by < num_blocks_h; ++by) { + const int y_o = by * (block_size >> sub_log2[1]); + for (int bx = 0; bx < num_blocks_w; ++bx) { + const int x_o = bx * (block_size >> sub_log2[0]); + if (!flat_blocks[by * num_blocks_w + bx]) { + continue; + } + int y_start = + (by > 0 && flat_blocks[(by - 1) * num_blocks_w + bx]) ? 0 : lag; + int x_start = + (bx > 0 && flat_blocks[by * num_blocks_w + bx - 1]) ? 0 : lag; + int y_end = AOMMIN((h >> sub_log2[1]) - by * (block_size >> sub_log2[1]), + block_size >> sub_log2[1]); + int x_end = AOMMIN( + (w >> sub_log2[0]) - bx * (block_size >> sub_log2[0]) - lag, + (bx + 1 < num_blocks_w && flat_blocks[by * num_blocks_w + bx + 1]) + ? (block_size >> sub_log2[0]) + : ((block_size >> sub_log2[0]) - lag)); + for (int y = y_start; y < y_end; ++y) { + for (int x = x_start; x < x_end; ++x) { + const double val = + noise_model->params.use_highbd + ? extract_ar_row_highbd(noise_model->coords, num_coords, + (const uint16_t *const)data, + (const uint16_t *const)denoised, + stride, sub_log2, + (const uint16_t *const)alt_data, + (const uint16_t *const)alt_denoised, + alt_stride, x + x_o, y + y_o, buffer) + : extract_ar_row_lowbd(noise_model->coords, num_coords, data, + denoised, stride, sub_log2, alt_data, + alt_denoised, alt_stride, x + x_o, + y + y_o, buffer); + for (int i = 0; i < n; ++i) { + for (int j = 0; j < n; ++j) { + A[i * n + j] += + (buffer[i] * buffer[j]) / (normalization * normalization); + } + b[i] += (buffer[i] * val) / (normalization * normalization); + } + noise_model->latest_state[c].num_observations++; + } + } + } + } + aom_free(buffer); + return 1; +} + +static void add_noise_std_observations( + aom_noise_model_t *noise_model, int c, const double *coeffs, + const uint8_t *const data, const uint8_t *const denoised, int w, int h, + int stride, int sub_log2[2], const uint8_t *const alt_data, int alt_stride, + const uint8_t *const flat_blocks, int block_size, int num_blocks_w, + int num_blocks_h) { + const int num_coords = noise_model->n; + aom_noise_strength_solver_t *noise_strength_solver = + &noise_model->latest_state[c].strength_solver; + + const aom_noise_strength_solver_t *noise_strength_luma = + &noise_model->latest_state[0].strength_solver; + const double luma_gain = noise_model->latest_state[0].ar_gain; + const double noise_gain = noise_model->latest_state[c].ar_gain; + for (int by = 0; by < num_blocks_h; ++by) { + const int y_o = by * (block_size >> sub_log2[1]); + for (int bx = 0; bx < num_blocks_w; ++bx) { + const int x_o = bx * (block_size >> sub_log2[0]); + if (!flat_blocks[by * num_blocks_w + bx]) { + continue; + } + const int num_samples_h = + AOMMIN((h >> sub_log2[1]) - by * (block_size >> sub_log2[1]), + block_size >> sub_log2[1]); + const int num_samples_w = + AOMMIN((w >> sub_log2[0]) - bx * (block_size >> sub_log2[0]), + (block_size >> sub_log2[0])); + // Make sure that we have a reasonable amount of samples to consider the + // block + if (num_samples_w * num_samples_h > block_size) { + const double block_mean = get_block_mean( + alt_data ? alt_data : data, w, h, alt_data ? alt_stride : stride, + x_o << sub_log2[0], y_o << sub_log2[1], block_size, + noise_model->params.use_highbd); + const double noise_var = get_noise_var( + data, denoised, stride, w >> sub_log2[0], h >> sub_log2[1], x_o, + y_o, block_size >> sub_log2[0], block_size >> sub_log2[1], + noise_model->params.use_highbd); + // We want to remove the part of the noise that came from being + // correlated with luma. Note that the noise solver for luma must + // have already been run. + const double luma_strength = + c > 0 ? luma_gain * noise_strength_solver_get_value( + noise_strength_luma, block_mean) + : 0; + const double corr = c > 0 ? coeffs[num_coords] : 0; + // Chroma noise: + // N(0, noise_var) = N(0, uncorr_var) + corr * N(0, luma_strength^2) + // The uncorrelated component: + // uncorr_var = noise_var - (corr * luma_strength)^2 + // But don't allow fully correlated noise (hence the max), since the + // synthesis cannot model it. + const double uncorr_std = sqrt( + AOMMAX(noise_var / 16, noise_var - pow(corr * luma_strength, 2))); + // After we've removed correlation with luma, undo the gain that will + // come from running the IIR filter. + const double adjusted_strength = uncorr_std / noise_gain; + aom_noise_strength_solver_add_measurement( + noise_strength_solver, block_mean, adjusted_strength); + } + } + } +} + +// Return true if the noise estimate appears to be different from the combined +// (multi-frame) estimate. The difference is measured by checking whether the +// AR coefficients have diverged (using a threshold on normalized cross +// correlation), or whether the noise strength has changed. +static int is_noise_model_different(aom_noise_model_t *const noise_model) { + // These thresholds are kind of arbitrary and will likely need further tuning + // (or exported as parameters). The threshold on noise strength is a weighted + // difference between the noise strength histograms + const double kCoeffThreshold = 0.9; + const double kStrengthThreshold = + 0.005 * (1 << (noise_model->params.bit_depth - 8)); + for (int c = 0; c < 1; ++c) { + const double corr = + aom_normalized_cross_correlation(noise_model->latest_state[c].eqns.x, + noise_model->combined_state[c].eqns.x, + noise_model->combined_state[c].eqns.n); + if (corr < kCoeffThreshold) return 1; + + const double dx = + 1.0 / noise_model->latest_state[c].strength_solver.num_bins; + + const aom_equation_system_t *latest_eqns = + &noise_model->latest_state[c].strength_solver.eqns; + const aom_equation_system_t *combined_eqns = + &noise_model->combined_state[c].strength_solver.eqns; + double diff = 0; + double total_weight = 0; + for (int j = 0; j < latest_eqns->n; ++j) { + double weight = 0; + for (int i = 0; i < latest_eqns->n; ++i) { + weight += latest_eqns->A[i * latest_eqns->n + j]; + } + weight = sqrt(weight); + diff += weight * fabs(latest_eqns->x[j] - combined_eqns->x[j]); + total_weight += weight; + } + if (diff * dx / total_weight > kStrengthThreshold) return 1; + } + return 0; +} + +static int ar_equation_system_solve(aom_noise_state_t *state, int is_chroma) { + const int ret = equation_system_solve(&state->eqns); + state->ar_gain = 1.0; + if (!ret) return ret; + + // Update the AR gain from the equation system as it will be used to fit + // the noise strength as a function of intensity. In the Yule-Walker + // equations, the diagonal should be the variance of the correlated noise. + // In the case of the least squares estimate, there will be some variability + // in the diagonal. So use the mean of the diagonal as the estimate of + // overall variance (this works for least squares or Yule-Walker formulation). + double var = 0; + const int n = state->eqns.n; + for (int i = 0; i < (state->eqns.n - is_chroma); ++i) { + var += state->eqns.A[i * n + i] / state->num_observations; + } + var /= (n - is_chroma); + + // Keep track of E(Y^2) = + E(X^2) + // In the case that we are using chroma and have an estimate of correlation + // with luma we adjust that estimate slightly to remove the correlated bits by + // subtracting out the last column of a scaled by our correlation estimate + // from b. E(y^2) = + double sum_covar = 0; + for (int i = 0; i < state->eqns.n - is_chroma; ++i) { + double bi = state->eqns.b[i]; + if (is_chroma) { + bi -= state->eqns.A[i * n + (n - 1)] * state->eqns.x[n - 1]; + } + sum_covar += (bi * state->eqns.x[i]) / state->num_observations; + } + // Now, get an estimate of the variance of uncorrelated noise signal and use + // it to determine the gain of the AR filter. + const double noise_var = AOMMAX(var - sum_covar, 1e-6); + state->ar_gain = AOMMAX(1, sqrt(AOMMAX(var / noise_var, 1e-6))); + return ret; +} + +aom_noise_status_t aom_noise_model_update( + aom_noise_model_t *const noise_model, const uint8_t *const data[3], + const uint8_t *const denoised[3], int w, int h, int stride[3], + int chroma_sub_log2[2], const uint8_t *const flat_blocks, int block_size) { + const int num_blocks_w = (w + block_size - 1) / block_size; + const int num_blocks_h = (h + block_size - 1) / block_size; + int y_model_different = 0; + int num_blocks = 0; + int i = 0, channel = 0; + + if (block_size <= 1) { + fprintf(stderr, "block_size = %d must be > 1\n", block_size); + return AOM_NOISE_STATUS_INVALID_ARGUMENT; + } + + if (block_size < noise_model->params.lag * 2 + 1) { + fprintf(stderr, "block_size = %d must be >= %d\n", block_size, + noise_model->params.lag * 2 + 1); + return AOM_NOISE_STATUS_INVALID_ARGUMENT; + } + + // Clear the latest equation system + for (i = 0; i < 3; ++i) { + equation_system_clear(&noise_model->latest_state[i].eqns); + noise_model->latest_state[i].num_observations = 0; + noise_strength_solver_clear(&noise_model->latest_state[i].strength_solver); + } + + // Check that we have enough flat blocks + for (i = 0; i < num_blocks_h * num_blocks_w; ++i) { + if (flat_blocks[i]) { + num_blocks++; + } + } + + if (num_blocks <= 1) { + fprintf(stderr, "Not enough flat blocks to update noise estimate\n"); + return AOM_NOISE_STATUS_INSUFFICIENT_FLAT_BLOCKS; + } + + for (channel = 0; channel < 3; ++channel) { + int no_subsampling[2] = { 0, 0 }; + const uint8_t *alt_data = channel > 0 ? data[0] : 0; + const uint8_t *alt_denoised = channel > 0 ? denoised[0] : 0; + int *sub = channel > 0 ? chroma_sub_log2 : no_subsampling; + const int is_chroma = channel != 0; + if (!data[channel] || !denoised[channel]) break; + if (!add_block_observations(noise_model, channel, data[channel], + denoised[channel], w, h, stride[channel], sub, + alt_data, alt_denoised, stride[0], flat_blocks, + block_size, num_blocks_w, num_blocks_h)) { + fprintf(stderr, "Adding block observation failed\n"); + return AOM_NOISE_STATUS_INTERNAL_ERROR; + } + + if (!ar_equation_system_solve(&noise_model->latest_state[channel], + is_chroma)) { + if (is_chroma) { + set_chroma_coefficient_fallback_soln( + &noise_model->latest_state[channel].eqns); + } else { + fprintf(stderr, "Solving latest noise equation system failed %d!\n", + channel); + return AOM_NOISE_STATUS_INTERNAL_ERROR; + } + } + + add_noise_std_observations( + noise_model, channel, noise_model->latest_state[channel].eqns.x, + data[channel], denoised[channel], w, h, stride[channel], sub, alt_data, + stride[0], flat_blocks, block_size, num_blocks_w, num_blocks_h); + + if (!aom_noise_strength_solver_solve( + &noise_model->latest_state[channel].strength_solver)) { + fprintf(stderr, "Solving latest noise strength failed!\n"); + return AOM_NOISE_STATUS_INTERNAL_ERROR; + } + + // Check noise characteristics and return if error. + if (channel == 0 && + noise_model->combined_state[channel].strength_solver.num_equations > + 0 && + is_noise_model_different(noise_model)) { + y_model_different = 1; + } + + // Don't update the combined stats if the y model is different. + if (y_model_different) continue; + + noise_model->combined_state[channel].num_observations += + noise_model->latest_state[channel].num_observations; + equation_system_add(&noise_model->combined_state[channel].eqns, + &noise_model->latest_state[channel].eqns); + if (!ar_equation_system_solve(&noise_model->combined_state[channel], + is_chroma)) { + if (is_chroma) { + set_chroma_coefficient_fallback_soln( + &noise_model->combined_state[channel].eqns); + } else { + fprintf(stderr, "Solving combined noise equation system failed %d!\n", + channel); + return AOM_NOISE_STATUS_INTERNAL_ERROR; + } + } + + noise_strength_solver_add( + &noise_model->combined_state[channel].strength_solver, + &noise_model->latest_state[channel].strength_solver); + + if (!aom_noise_strength_solver_solve( + &noise_model->combined_state[channel].strength_solver)) { + fprintf(stderr, "Solving combined noise strength failed!\n"); + return AOM_NOISE_STATUS_INTERNAL_ERROR; + } + } + + return y_model_different ? AOM_NOISE_STATUS_DIFFERENT_NOISE_TYPE + : AOM_NOISE_STATUS_OK; +} + +void aom_noise_model_save_latest(aom_noise_model_t *noise_model) { + for (int c = 0; c < 3; c++) { + equation_system_copy(&noise_model->combined_state[c].eqns, + &noise_model->latest_state[c].eqns); + equation_system_copy(&noise_model->combined_state[c].strength_solver.eqns, + &noise_model->latest_state[c].strength_solver.eqns); + noise_model->combined_state[c].strength_solver.num_equations = + noise_model->latest_state[c].strength_solver.num_equations; + noise_model->combined_state[c].num_observations = + noise_model->latest_state[c].num_observations; + noise_model->combined_state[c].ar_gain = + noise_model->latest_state[c].ar_gain; + } +} + +int aom_noise_model_get_grain_parameters(aom_noise_model_t *const noise_model, + aom_film_grain_t *film_grain) { + if (noise_model->params.lag > 3) { + fprintf(stderr, "params.lag = %d > 3\n", noise_model->params.lag); + return 0; + } + uint16_t random_seed = film_grain->random_seed; + memset(film_grain, 0, sizeof(*film_grain)); + film_grain->random_seed = random_seed; + + film_grain->apply_grain = 1; + film_grain->update_parameters = 1; + + film_grain->ar_coeff_lag = noise_model->params.lag; + + // Convert the scaling functions to 8 bit values + aom_noise_strength_lut_t scaling_points[3]; + aom_noise_strength_solver_fit_piecewise( + &noise_model->combined_state[0].strength_solver, 14, scaling_points + 0); + aom_noise_strength_solver_fit_piecewise( + &noise_model->combined_state[1].strength_solver, 10, scaling_points + 1); + aom_noise_strength_solver_fit_piecewise( + &noise_model->combined_state[2].strength_solver, 10, scaling_points + 2); + + // Both the domain and the range of the scaling functions in the film_grain + // are normalized to 8-bit (e.g., they are implicitly scaled during grain + // synthesis). + const double strength_divisor = 1 << (noise_model->params.bit_depth - 8); + double max_scaling_value = 1e-4; + for (int c = 0; c < 3; ++c) { + for (int i = 0; i < scaling_points[c].num_points; ++i) { + scaling_points[c].points[i][0] = + AOMMIN(255, scaling_points[c].points[i][0] / strength_divisor); + scaling_points[c].points[i][1] = + AOMMIN(255, scaling_points[c].points[i][1] / strength_divisor); + max_scaling_value = + AOMMAX(scaling_points[c].points[i][1], max_scaling_value); + } + } + + // Scaling_shift values are in the range [8,11] + const int max_scaling_value_log2 = + clamp((int)floor(log2(max_scaling_value) + 1), 2, 5); + film_grain->scaling_shift = 5 + (8 - max_scaling_value_log2); + + const double scale_factor = 1 << (8 - max_scaling_value_log2); + film_grain->num_y_points = scaling_points[0].num_points; + film_grain->num_cb_points = scaling_points[1].num_points; + film_grain->num_cr_points = scaling_points[2].num_points; + + int(*film_grain_scaling[3])[2] = { + film_grain->scaling_points_y, + film_grain->scaling_points_cb, + film_grain->scaling_points_cr, + }; + for (int c = 0; c < 3; c++) { + for (int i = 0; i < scaling_points[c].num_points; ++i) { + film_grain_scaling[c][i][0] = (int)(scaling_points[c].points[i][0] + 0.5); + film_grain_scaling[c][i][1] = clamp( + (int)(scale_factor * scaling_points[c].points[i][1] + 0.5), 0, 255); + } + } + aom_noise_strength_lut_free(scaling_points + 0); + aom_noise_strength_lut_free(scaling_points + 1); + aom_noise_strength_lut_free(scaling_points + 2); + + // Convert the ar_coeffs into 8-bit values + const int n_coeff = noise_model->combined_state[0].eqns.n; + double max_coeff = 1e-4, min_coeff = -1e-4; + double y_corr[2] = { 0, 0 }; + double avg_luma_strength = 0; + for (int c = 0; c < 3; c++) { + aom_equation_system_t *eqns = &noise_model->combined_state[c].eqns; + for (int i = 0; i < n_coeff; ++i) { + max_coeff = AOMMAX(max_coeff, eqns->x[i]); + min_coeff = AOMMIN(min_coeff, eqns->x[i]); + } + // Since the correlation between luma/chroma was computed in an already + // scaled space, we adjust it in the un-scaled space. + aom_noise_strength_solver_t *solver = + &noise_model->combined_state[c].strength_solver; + // Compute a weighted average of the strength for the channel. + double average_strength = 0, total_weight = 0; + for (int i = 0; i < solver->eqns.n; ++i) { + double w = 0; + for (int j = 0; j < solver->eqns.n; ++j) { + w += solver->eqns.A[i * solver->eqns.n + j]; + } + w = sqrt(w); + average_strength += solver->eqns.x[i] * w; + total_weight += w; + } + if (total_weight == 0) + average_strength = 1; + else + average_strength /= total_weight; + if (c == 0) { + avg_luma_strength = average_strength; + } else { + y_corr[c - 1] = avg_luma_strength * eqns->x[n_coeff] / average_strength; + max_coeff = AOMMAX(max_coeff, y_corr[c - 1]); + min_coeff = AOMMIN(min_coeff, y_corr[c - 1]); + } + } + // Shift value: AR coeffs range (values 6-9) + // 6: [-2, 2), 7: [-1, 1), 8: [-0.5, 0.5), 9: [-0.25, 0.25) + film_grain->ar_coeff_shift = + clamp(7 - (int)AOMMAX(1 + floor(log2(max_coeff)), ceil(log2(-min_coeff))), + 6, 9); + double scale_ar_coeff = 1 << film_grain->ar_coeff_shift; + int *ar_coeffs[3] = { + film_grain->ar_coeffs_y, + film_grain->ar_coeffs_cb, + film_grain->ar_coeffs_cr, + }; + for (int c = 0; c < 3; ++c) { + aom_equation_system_t *eqns = &noise_model->combined_state[c].eqns; + for (int i = 0; i < n_coeff; ++i) { + ar_coeffs[c][i] = + clamp((int)round(scale_ar_coeff * eqns->x[i]), -128, 127); + } + if (c > 0) { + ar_coeffs[c][n_coeff] = + clamp((int)round(scale_ar_coeff * y_corr[c - 1]), -128, 127); + } + } + + // At the moment, the noise modeling code assumes that the chroma scaling + // functions are a function of luma. + film_grain->cb_mult = 128; // 8 bits + film_grain->cb_luma_mult = 192; // 8 bits + film_grain->cb_offset = 256; // 9 bits + + film_grain->cr_mult = 128; // 8 bits + film_grain->cr_luma_mult = 192; // 8 bits + film_grain->cr_offset = 256; // 9 bits + + film_grain->chroma_scaling_from_luma = 0; + film_grain->grain_scale_shift = 0; + film_grain->overlap_flag = 1; + return 1; +} + +static void pointwise_multiply(const float *a, float *b, int n) { + for (int i = 0; i < n; ++i) { + b[i] *= a[i]; + } +} + +static float *get_half_cos_window(int block_size) { + float *window_function = + (float *)aom_malloc(block_size * block_size * sizeof(*window_function)); + for (int y = 0; y < block_size; ++y) { + const double cos_yd = cos((.5 + y) * PI / block_size - PI / 2); + for (int x = 0; x < block_size; ++x) { + const double cos_xd = cos((.5 + x) * PI / block_size - PI / 2); + window_function[y * block_size + x] = (float)(cos_yd * cos_xd); + } + } + return window_function; +} + +#define DITHER_AND_QUANTIZE(INT_TYPE, suffix) \ + static void dither_and_quantize_##suffix( \ + float *result, int result_stride, INT_TYPE *denoised, int w, int h, \ + int stride, int chroma_sub_w, int chroma_sub_h, int block_size, \ + float block_normalization) { \ + for (int y = 0; y < (h >> chroma_sub_h); ++y) { \ + for (int x = 0; x < (w >> chroma_sub_w); ++x) { \ + const int result_idx = \ + (y + (block_size >> chroma_sub_h)) * result_stride + x + \ + (block_size >> chroma_sub_w); \ + INT_TYPE new_val = (INT_TYPE)AOMMIN( \ + AOMMAX(result[result_idx] * block_normalization + 0.5f, 0), \ + block_normalization); \ + const float err = \ + -(((float)new_val) / block_normalization - result[result_idx]); \ + denoised[y * stride + x] = new_val; \ + if (x + 1 < (w >> chroma_sub_w)) { \ + result[result_idx + 1] += err * 7.0f / 16.0f; \ + } \ + if (y + 1 < (h >> chroma_sub_h)) { \ + if (x > 0) { \ + result[result_idx + result_stride - 1] += err * 3.0f / 16.0f; \ + } \ + result[result_idx + result_stride] += err * 5.0f / 16.0f; \ + if (x + 1 < (w >> chroma_sub_w)) { \ + result[result_idx + result_stride + 1] += err * 1.0f / 16.0f; \ + } \ + } \ + } \ + } \ + } + +DITHER_AND_QUANTIZE(uint8_t, lowbd); +DITHER_AND_QUANTIZE(uint16_t, highbd); + +int aom_wiener_denoise_2d(const uint8_t *const data[3], uint8_t *denoised[3], + int w, int h, int stride[3], int chroma_sub[2], + float *noise_psd[3], int block_size, int bit_depth, + int use_highbd) { + float *plane = NULL, *block = NULL, *window_full = NULL, + *window_chroma = NULL; + double *block_d = NULL, *plane_d = NULL; + struct aom_noise_tx_t *tx_full = NULL; + struct aom_noise_tx_t *tx_chroma = NULL; + const int num_blocks_w = (w + block_size - 1) / block_size; + const int num_blocks_h = (h + block_size - 1) / block_size; + const int result_stride = (num_blocks_w + 2) * block_size; + const int result_height = (num_blocks_h + 2) * block_size; + float *result = NULL; + int init_success = 1; + aom_flat_block_finder_t block_finder_full; + aom_flat_block_finder_t block_finder_chroma; + const float kBlockNormalization = (float)((1 << bit_depth) - 1); + if (chroma_sub[0] != chroma_sub[1]) { + fprintf(stderr, + "aom_wiener_denoise_2d doesn't handle different chroma " + "subsampling"); + return 0; + } + init_success &= aom_flat_block_finder_init(&block_finder_full, block_size, + bit_depth, use_highbd); + result = (float *)aom_malloc((num_blocks_h + 2) * block_size * result_stride * + sizeof(*result)); + plane = (float *)aom_malloc(block_size * block_size * sizeof(*plane)); + block = + (float *)aom_memalign(32, 2 * block_size * block_size * sizeof(*block)); + block_d = (double *)aom_malloc(block_size * block_size * sizeof(*block_d)); + plane_d = (double *)aom_malloc(block_size * block_size * sizeof(*plane_d)); + window_full = get_half_cos_window(block_size); + tx_full = aom_noise_tx_malloc(block_size); + + if (chroma_sub[0] != 0) { + init_success &= aom_flat_block_finder_init(&block_finder_chroma, + block_size >> chroma_sub[0], + bit_depth, use_highbd); + window_chroma = get_half_cos_window(block_size >> chroma_sub[0]); + tx_chroma = aom_noise_tx_malloc(block_size >> chroma_sub[0]); + } else { + window_chroma = window_full; + tx_chroma = tx_full; + } + + init_success &= (tx_full != NULL) && (tx_chroma != NULL) && (plane != NULL) && + (plane_d != NULL) && (block != NULL) && (block_d != NULL) && + (window_full != NULL) && (window_chroma != NULL) && + (result != NULL); + for (int c = init_success ? 0 : 3; c < 3; ++c) { + float *window_function = c == 0 ? window_full : window_chroma; + aom_flat_block_finder_t *block_finder = &block_finder_full; + const int chroma_sub_h = c > 0 ? chroma_sub[1] : 0; + const int chroma_sub_w = c > 0 ? chroma_sub[0] : 0; + struct aom_noise_tx_t *tx = + (c > 0 && chroma_sub[0] > 0) ? tx_chroma : tx_full; + if (!data[c] || !denoised[c]) continue; + if (c > 0 && chroma_sub[0] != 0) { + block_finder = &block_finder_chroma; + } + memset(result, 0, sizeof(*result) * result_stride * result_height); + // Do overlapped block processing (half overlapped). The block rows can + // easily be done in parallel + for (int offsy = 0; offsy < (block_size >> chroma_sub_h); + offsy += (block_size >> chroma_sub_h) / 2) { + for (int offsx = 0; offsx < (block_size >> chroma_sub_w); + offsx += (block_size >> chroma_sub_w) / 2) { + // Pad the boundary when processing each block-set. + for (int by = -1; by < num_blocks_h; ++by) { + for (int bx = -1; bx < num_blocks_w; ++bx) { + const int pixels_per_block = + (block_size >> chroma_sub_w) * (block_size >> chroma_sub_h); + aom_flat_block_finder_extract_block( + block_finder, data[c], w >> chroma_sub_w, h >> chroma_sub_h, + stride[c], bx * (block_size >> chroma_sub_w) + offsx, + by * (block_size >> chroma_sub_h) + offsy, plane_d, block_d); + for (int j = 0; j < pixels_per_block; ++j) { + block[j] = (float)block_d[j]; + plane[j] = (float)plane_d[j]; + } + pointwise_multiply(window_function, block, pixels_per_block); + aom_noise_tx_forward(tx, block); + aom_noise_tx_filter(tx, noise_psd[c]); + aom_noise_tx_inverse(tx, block); + + // Apply window function to the plane approximation (we will apply + // it to the sum of plane + block when composing the results). + pointwise_multiply(window_function, plane, pixels_per_block); + + for (int y = 0; y < (block_size >> chroma_sub_h); ++y) { + const int y_result = + y + (by + 1) * (block_size >> chroma_sub_h) + offsy; + for (int x = 0; x < (block_size >> chroma_sub_w); ++x) { + const int x_result = + x + (bx + 1) * (block_size >> chroma_sub_w) + offsx; + result[y_result * result_stride + x_result] += + (block[y * (block_size >> chroma_sub_w) + x] + + plane[y * (block_size >> chroma_sub_w) + x]) * + window_function[y * (block_size >> chroma_sub_w) + x]; + } + } + } + } + } + } + if (use_highbd) { + dither_and_quantize_highbd(result, result_stride, (uint16_t *)denoised[c], + w, h, stride[c], chroma_sub_w, chroma_sub_h, + block_size, kBlockNormalization); + } else { + dither_and_quantize_lowbd(result, result_stride, denoised[c], w, h, + stride[c], chroma_sub_w, chroma_sub_h, + block_size, kBlockNormalization); + } + } + aom_free(result); + aom_free(plane); + aom_free(block); + aom_free(plane_d); + aom_free(block_d); + aom_free(window_full); + + aom_noise_tx_free(tx_full); + + aom_flat_block_finder_free(&block_finder_full); + if (chroma_sub[0] != 0) { + aom_flat_block_finder_free(&block_finder_chroma); + aom_free(window_chroma); + aom_noise_tx_free(tx_chroma); + } + return init_success; +} + +struct aom_denoise_and_model_t { + int block_size; + int bit_depth; + float noise_level; + + // Size of current denoised buffer and flat_block buffer + int width; + int height; + int y_stride; + int uv_stride; + int num_blocks_w; + int num_blocks_h; + + // Buffers for image and noise_psd allocated on the fly + float *noise_psd[3]; + uint8_t *denoised[3]; + uint8_t *flat_blocks; + + aom_flat_block_finder_t flat_block_finder; + aom_noise_model_t noise_model; +}; + +struct aom_denoise_and_model_t *aom_denoise_and_model_alloc(int bit_depth, + int block_size, + float noise_level) { + struct aom_denoise_and_model_t *ctx = + (struct aom_denoise_and_model_t *)aom_malloc( + sizeof(struct aom_denoise_and_model_t)); + if (!ctx) { + fprintf(stderr, "Unable to allocate denoise_and_model struct\n"); + return NULL; + } + memset(ctx, 0, sizeof(*ctx)); + + ctx->block_size = block_size; + ctx->noise_level = noise_level; + ctx->bit_depth = bit_depth; + + ctx->noise_psd[0] = + aom_malloc(sizeof(*ctx->noise_psd[0]) * block_size * block_size); + ctx->noise_psd[1] = + aom_malloc(sizeof(*ctx->noise_psd[1]) * block_size * block_size); + ctx->noise_psd[2] = + aom_malloc(sizeof(*ctx->noise_psd[2]) * block_size * block_size); + if (!ctx->noise_psd[0] || !ctx->noise_psd[1] || !ctx->noise_psd[2]) { + fprintf(stderr, "Unable to allocate noise PSD buffers\n"); + aom_denoise_and_model_free(ctx); + return NULL; + } + return ctx; +} + +void aom_denoise_and_model_free(struct aom_denoise_and_model_t *ctx) { + aom_free(ctx->flat_blocks); + for (int i = 0; i < 3; ++i) { + aom_free(ctx->denoised[i]); + aom_free(ctx->noise_psd[i]); + } + aom_noise_model_free(&ctx->noise_model); + aom_flat_block_finder_free(&ctx->flat_block_finder); + aom_free(ctx); +} + +static int denoise_and_model_realloc_if_necessary( + struct aom_denoise_and_model_t *ctx, YV12_BUFFER_CONFIG *sd) { + if (ctx->width == sd->y_width && ctx->height == sd->y_height && + ctx->y_stride == sd->y_stride && ctx->uv_stride == sd->uv_stride) + return 1; + const int use_highbd = (sd->flags & YV12_FLAG_HIGHBITDEPTH) != 0; + const int block_size = ctx->block_size; + + ctx->width = sd->y_width; + ctx->height = sd->y_height; + ctx->y_stride = sd->y_stride; + ctx->uv_stride = sd->uv_stride; + + for (int i = 0; i < 3; ++i) { + aom_free(ctx->denoised[i]); + ctx->denoised[i] = NULL; + } + aom_free(ctx->flat_blocks); + ctx->flat_blocks = NULL; + + ctx->denoised[0] = aom_malloc((sd->y_stride * sd->y_height) << use_highbd); + ctx->denoised[1] = aom_malloc((sd->uv_stride * sd->uv_height) << use_highbd); + ctx->denoised[2] = aom_malloc((sd->uv_stride * sd->uv_height) << use_highbd); + if (!ctx->denoised[0] || !ctx->denoised[1] || !ctx->denoised[2]) { + fprintf(stderr, "Unable to allocate denoise buffers\n"); + return 0; + } + ctx->num_blocks_w = (sd->y_width + ctx->block_size - 1) / ctx->block_size; + ctx->num_blocks_h = (sd->y_height + ctx->block_size - 1) / ctx->block_size; + ctx->flat_blocks = aom_malloc(ctx->num_blocks_w * ctx->num_blocks_h); + + aom_flat_block_finder_free(&ctx->flat_block_finder); + if (!aom_flat_block_finder_init(&ctx->flat_block_finder, ctx->block_size, + ctx->bit_depth, use_highbd)) { + fprintf(stderr, "Unable to init flat block finder\n"); + return 0; + } + + const aom_noise_model_params_t params = { AOM_NOISE_SHAPE_SQUARE, 3, + ctx->bit_depth, use_highbd }; + aom_noise_model_free(&ctx->noise_model); + if (!aom_noise_model_init(&ctx->noise_model, params)) { + fprintf(stderr, "Unable to init noise model\n"); + return 0; + } + + // Simply use a flat PSD (although we could use the flat blocks to estimate + // PSD) those to estimate an actual noise PSD) + const float y_noise_level = + aom_noise_psd_get_default_value(ctx->block_size, ctx->noise_level); + const float uv_noise_level = aom_noise_psd_get_default_value( + ctx->block_size >> sd->subsampling_x, ctx->noise_level); + for (int i = 0; i < block_size * block_size; ++i) { + ctx->noise_psd[0][i] = y_noise_level; + ctx->noise_psd[1][i] = ctx->noise_psd[2][i] = uv_noise_level; + } + return 1; +} + +int aom_denoise_and_model_run(struct aom_denoise_and_model_t *ctx, + YV12_BUFFER_CONFIG *sd, + aom_film_grain_t *film_grain) { + const int block_size = ctx->block_size; + const int use_highbd = (sd->flags & YV12_FLAG_HIGHBITDEPTH) != 0; + uint8_t *raw_data[3] = { + use_highbd ? (uint8_t *)CONVERT_TO_SHORTPTR(sd->y_buffer) : sd->y_buffer, + use_highbd ? (uint8_t *)CONVERT_TO_SHORTPTR(sd->u_buffer) : sd->u_buffer, + use_highbd ? (uint8_t *)CONVERT_TO_SHORTPTR(sd->v_buffer) : sd->v_buffer, + }; + const uint8_t *const data[3] = { raw_data[0], raw_data[1], raw_data[2] }; + int strides[3] = { sd->y_stride, sd->uv_stride, sd->uv_stride }; + int chroma_sub_log2[2] = { sd->subsampling_x, sd->subsampling_y }; + + if (!denoise_and_model_realloc_if_necessary(ctx, sd)) { + fprintf(stderr, "Unable to realloc buffers\n"); + return 0; + } + + aom_flat_block_finder_run(&ctx->flat_block_finder, data[0], sd->y_width, + sd->y_height, strides[0], ctx->flat_blocks); + + if (!aom_wiener_denoise_2d(data, ctx->denoised, sd->y_width, sd->y_height, + strides, chroma_sub_log2, ctx->noise_psd, + block_size, ctx->bit_depth, use_highbd)) { + fprintf(stderr, "Unable to denoise image\n"); + return 0; + } + + const aom_noise_status_t status = aom_noise_model_update( + &ctx->noise_model, data, (const uint8_t *const *)ctx->denoised, + sd->y_width, sd->y_height, strides, chroma_sub_log2, ctx->flat_blocks, + block_size); + int have_noise_estimate = 0; + if (status == AOM_NOISE_STATUS_OK) { + have_noise_estimate = 1; + } else if (status == AOM_NOISE_STATUS_DIFFERENT_NOISE_TYPE) { + aom_noise_model_save_latest(&ctx->noise_model); + have_noise_estimate = 1; + } else { + // Unable to update noise model; proceed if we have a previous estimate. + have_noise_estimate = + (ctx->noise_model.combined_state[0].strength_solver.num_equations > 0); + } + + film_grain->apply_grain = 0; + if (have_noise_estimate) { + if (!aom_noise_model_get_grain_parameters(&ctx->noise_model, film_grain)) { + fprintf(stderr, "Unable to get grain parameters.\n"); + return 0; + } + if (!film_grain->random_seed) { + film_grain->random_seed = 7391; + } + memcpy(raw_data[0], ctx->denoised[0], + (strides[0] * sd->y_height) << use_highbd); + memcpy(raw_data[1], ctx->denoised[1], + (strides[1] * sd->uv_height) << use_highbd); + memcpy(raw_data[2], ctx->denoised[2], + (strides[2] * sd->uv_height) << use_highbd); + } + return 1; +} diff --git a/third_party/aom/aom_dsp/noise_model.h b/third_party/aom/aom_dsp/noise_model.h new file mode 100644 index 0000000000..049d5be151 --- /dev/null +++ b/third_party/aom/aom_dsp/noise_model.h @@ -0,0 +1,323 @@ +/* + * Copyright (c) 2017, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_NOISE_MODEL_H_ +#define AOM_AOM_DSP_NOISE_MODEL_H_ + +#ifdef __cplusplus +extern "C" { +#endif // __cplusplus + +#include +#include "aom_dsp/grain_synthesis.h" +#include "aom_scale/yv12config.h" + +/*!\brief Wrapper of data required to represent linear system of eqns and soln. + */ +typedef struct { + double *A; + double *b; + double *x; + int n; +} aom_equation_system_t; + +/*!\brief Representation of a piecewise linear curve + * + * Holds n points as (x, y) pairs, that store the curve. + */ +typedef struct { + double (*points)[2]; + int num_points; +} aom_noise_strength_lut_t; + +/*!\brief Init the noise strength lut with the given number of points*/ +int aom_noise_strength_lut_init(aom_noise_strength_lut_t *lut, int num_points); + +/*!\brief Frees the noise strength lut. */ +void aom_noise_strength_lut_free(aom_noise_strength_lut_t *lut); + +/*!\brief Evaluate the lut at the point x. + * + * \param[in] lut The lut data. + * \param[in] x The coordinate to evaluate the lut. + */ +double aom_noise_strength_lut_eval(const aom_noise_strength_lut_t *lut, + double x); + +/*!\brief Helper struct to model noise strength as a function of intensity. + * + * Internally, this structure holds a representation of a linear system + * of equations that models noise strength (standard deviation) as a + * function of intensity. The mapping is initially stored using a + * piecewise representation with evenly spaced bins that cover the entire + * domain from [min_intensity, max_intensity]. Each observation (x,y) gives a + * constraint of the form: + * y_{i} (1 - a) + y_{i+1} a = y + * where y_{i} is the value of bin i and x_{i} <= x <= x_{i+1} and + * a = x/(x_{i+1} - x{i}). The equation system holds the corresponding + * normal equations. + * + * As there may be missing data, the solution is regularized to get a + * complete set of values for the bins. A reduced representation after + * solving can be obtained by getting the corresponding noise_strength_lut_t. + */ +typedef struct { + aom_equation_system_t eqns; + double min_intensity; + double max_intensity; + int num_bins; + int num_equations; + double total; +} aom_noise_strength_solver_t; + +/*!\brief Initializes the noise solver with the given number of bins. + * + * Returns 0 if initialization fails. + * + * \param[in] solver The noise solver to be initialized. + * \param[in] num_bins Number of bins to use in the internal representation. + * \param[in] bit_depth The bit depth used to derive {min,max}_intensity. + */ +int aom_noise_strength_solver_init(aom_noise_strength_solver_t *solver, + int num_bins, int bit_depth); +void aom_noise_strength_solver_free(aom_noise_strength_solver_t *solver); + +/*!\brief Gets the x coordinate of bin i. + * + * \param[in] i The bin whose coordinate to query. + */ +double aom_noise_strength_solver_get_center( + const aom_noise_strength_solver_t *solver, int i); + +/*!\brief Add an observation of the block mean intensity to its noise strength. + * + * \param[in] block_mean The average block intensity, + * \param[in] noise_std The observed noise strength. + */ +void aom_noise_strength_solver_add_measurement( + aom_noise_strength_solver_t *solver, double block_mean, double noise_std); + +/*!\brief Solves the current set of equations for the noise strength. */ +int aom_noise_strength_solver_solve(aom_noise_strength_solver_t *solver); + +/*!\brief Fits a reduced piecewise linear lut to the internal solution + * + * \param[in] max_num_points The maximum number of output points + * \param[out] lut The output piecewise linear lut. + */ +int aom_noise_strength_solver_fit_piecewise( + const aom_noise_strength_solver_t *solver, int max_num_points, + aom_noise_strength_lut_t *lut); + +/*!\brief Helper for holding precomputed data for finding flat blocks. + * + * Internally a block is modeled with a low-order polynomial model. A + * planar model would be a bunch of equations like: + * <[y_i x_i 1], [a_1, a_2, a_3]> = b_i + * for each point in the block. The system matrix A with row i as [y_i x_i 1] + * is maintained as is the inverse, inv(A'*A), so that the plane parameters + * can be fit for each block. + */ +typedef struct { + double *AtA_inv; + double *A; + int num_params; // The number of parameters used for internal low-order model + int block_size; // The block size the finder was initialized with + double normalization; // Normalization factor (1 / (2^(bit_depth) - 1)) + int use_highbd; // Whether input data should be interpreted as uint16 +} aom_flat_block_finder_t; + +/*!\brief Init the block_finder with the given block size, bit_depth */ +int aom_flat_block_finder_init(aom_flat_block_finder_t *block_finder, + int block_size, int bit_depth, int use_highbd); +void aom_flat_block_finder_free(aom_flat_block_finder_t *block_finder); + +/*!\brief Helper to extract a block and low order "planar" model. */ +void aom_flat_block_finder_extract_block( + const aom_flat_block_finder_t *block_finder, const uint8_t *const data, + int w, int h, int stride, int offsx, int offsy, double *plane, + double *block); + +/*!\brief Runs the flat block finder on the input data. + * + * Find flat blocks in the input image data. Returns a map of + * flat_blocks, where the value of flat_blocks map will be non-zero + * when a block is determined to be flat. A higher value indicates a bigger + * confidence in the decision. + */ +int aom_flat_block_finder_run(const aom_flat_block_finder_t *block_finder, + const uint8_t *const data, int w, int h, + int stride, uint8_t *flat_blocks); + +// The noise shape indicates the allowed coefficients in the AR model. +typedef enum { + AOM_NOISE_SHAPE_DIAMOND = 0, + AOM_NOISE_SHAPE_SQUARE = 1 +} aom_noise_shape; + +// The parameters of the noise model include the shape type, lag, the +// bit depth of the input images provided, and whether the input images +// will be using uint16 (or uint8) representation. +typedef struct { + aom_noise_shape shape; + int lag; + int bit_depth; + int use_highbd; +} aom_noise_model_params_t; + +/*!\brief State of a noise model estimate for a single channel. + * + * This contains a system of equations that can be used to solve + * for the auto-regressive coefficients as well as a noise strength + * solver that can be used to model noise strength as a function of + * intensity. + */ +typedef struct { + aom_equation_system_t eqns; + aom_noise_strength_solver_t strength_solver; + int num_observations; // The number of observations in the eqn system + double ar_gain; // The gain of the current AR filter +} aom_noise_state_t; + +/*!\brief Complete model of noise for a planar video + * + * This includes a noise model for the latest frame and an aggregated + * estimate over all previous frames that had similar parameters. + */ +typedef struct { + aom_noise_model_params_t params; + aom_noise_state_t combined_state[3]; // Combined state per channel + aom_noise_state_t latest_state[3]; // Latest state per channel + int (*coords)[2]; // Offsets (x,y) of the coefficient samples + int n; // Number of parameters (size of coords) + int bit_depth; +} aom_noise_model_t; + +/*!\brief Result of a noise model update. */ +typedef enum { + AOM_NOISE_STATUS_OK = 0, + AOM_NOISE_STATUS_INVALID_ARGUMENT, + AOM_NOISE_STATUS_INSUFFICIENT_FLAT_BLOCKS, + AOM_NOISE_STATUS_DIFFERENT_NOISE_TYPE, + AOM_NOISE_STATUS_INTERNAL_ERROR, +} aom_noise_status_t; + +/*!\brief Initializes a noise model with the given parameters. + * + * Returns 0 on failure. + */ +int aom_noise_model_init(aom_noise_model_t *model, + const aom_noise_model_params_t params); +void aom_noise_model_free(aom_noise_model_t *model); + +/*!\brief Updates the noise model with a new frame observation. + * + * Updates the noise model with measurements from the given input frame and a + * denoised variant of it. Noise is sampled from flat blocks using the flat + * block map. + * + * Returns a noise_status indicating if the update was successful. If the + * Update was successful, the combined_state is updated with measurements from + * the provided frame. If status is OK or DIFFERENT_NOISE_TYPE, the latest noise + * state will be updated with measurements from the provided frame. + * + * \param[in,out] noise_model The noise model to be updated + * \param[in] data Raw frame data + * \param[in] denoised Denoised frame data. + * \param[in] w Frame width + * \param[in] h Frame height + * \param[in] strides Stride of the planes + * \param[in] chroma_sub_log2 Chroma subsampling for planes != 0. + * \param[in] flat_blocks A map to blocks that have been determined flat + * \param[in] block_size The size of blocks. + */ +aom_noise_status_t aom_noise_model_update( + aom_noise_model_t *const noise_model, const uint8_t *const data[3], + const uint8_t *const denoised[3], int w, int h, int strides[3], + int chroma_sub_log2[2], const uint8_t *const flat_blocks, int block_size); + +/*\brief Save the "latest" estimate into the "combined" estimate. + * + * This is meant to be called when the noise modeling detected a change + * in parameters (or for example, if a user wanted to reset estimation at + * a shot boundary). + */ +void aom_noise_model_save_latest(aom_noise_model_t *noise_model); + +/*!\brief Converts the noise_model parameters to the corresponding + * grain_parameters. + * + * The noise structs in this file are suitable for estimation (e.g., using + * floats), but the grain parameters in the bitstream are quantized. This + * function does the conversion by selecting the correct quantization levels. + */ +int aom_noise_model_get_grain_parameters(aom_noise_model_t *const noise_model, + aom_film_grain_t *film_grain); + +/*!\brief Perform a Wiener filter denoising in 2D using the provided noise psd. + * + * \param[in] data Raw frame data + * \param[out] denoised Denoised frame data + * \param[in] w Frame width + * \param[in] h Frame height + * \param[in] stride Stride of the planes + * \param[in] chroma_sub_log2 Chroma subsampling for planes != 0. + * \param[in] noise_psd The power spectral density of the noise + * \param[in] block_size The size of blocks + * \param[in] bit_depth Bit depth of the image + * \param[in] use_highbd If true, uint8 pointers are interpreted as + * uint16 and stride is measured in uint16. + * This must be true when bit_depth >= 10. + */ +int aom_wiener_denoise_2d(const uint8_t *const data[3], uint8_t *denoised[3], + int w, int h, int stride[3], int chroma_sub_log2[2], + float *noise_psd[3], int block_size, int bit_depth, + int use_highbd); + +struct aom_denoise_and_model_t; + +/*!\brief Denoise the buffer and model the residual noise. + * + * This is meant to be called sequentially on input frames. The input buffer + * is denoised and the residual noise is modelled. The current noise estimate + * is populated in film_grain. Returns true on success. The grain.apply_grain + * parameter will be true when the input buffer was successfully denoised and + * grain was modelled. Returns false on error. + * + * \param[in] ctx Struct allocated with aom_denoise_and_model_alloc + * that holds some buffers for denoising and the current + * noise estimate. + * \param[in/out] buf The raw input buffer to be denoised. + * \param[out] grain Output film grain parameters + */ +int aom_denoise_and_model_run(struct aom_denoise_and_model_t *ctx, + YV12_BUFFER_CONFIG *buf, aom_film_grain_t *grain); + +/*!\brief Allocates a context that can be used for denoising and noise modeling. + * + * \param[in] bit_depth Bit depth of buffers this will be run on. + * \param[in] block_size Block size for noise modeling and flat block + * estimation + * \param[in] noise_level The noise_level (2.5 for moderate noise, and 5 for + * higher levels of noise) + */ +struct aom_denoise_and_model_t *aom_denoise_and_model_alloc(int bit_depth, + int block_size, + float noise_level); + +/*!\brief Frees the denoise context allocated with aom_denoise_and_model_alloc + */ +void aom_denoise_and_model_free(struct aom_denoise_and_model_t *denoise_model); + +#ifdef __cplusplus +} // extern "C" +#endif // __cplusplus +#endif // AOM_AOM_DSP_NOISE_MODEL_H_ diff --git a/third_party/aom/aom_dsp/noise_util.c b/third_party/aom/aom_dsp/noise_util.c new file mode 100644 index 0000000000..87e8e9fecc --- /dev/null +++ b/third_party/aom/aom_dsp/noise_util.c @@ -0,0 +1,221 @@ +/* + * Copyright (c) 2017, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include + +#include +#include +#include + +#include "aom_dsp/noise_util.h" +#include "aom_dsp/fft_common.h" +#include "aom_mem/aom_mem.h" +#include "config/aom_dsp_rtcd.h" + +float aom_noise_psd_get_default_value(int block_size, float factor) { + return (factor * factor / 10000) * block_size * block_size / 8; +} + +// Internal representation of noise transform. It keeps track of the +// transformed data and a temporary working buffer to use during the +// transform. +struct aom_noise_tx_t { + float *tx_block; + float *temp; + int block_size; + void (*fft)(const float *, float *, float *); + void (*ifft)(const float *, float *, float *); +}; + +struct aom_noise_tx_t *aom_noise_tx_malloc(int block_size) { + struct aom_noise_tx_t *noise_tx = + (struct aom_noise_tx_t *)aom_malloc(sizeof(struct aom_noise_tx_t)); + if (!noise_tx) return NULL; + memset(noise_tx, 0, sizeof(*noise_tx)); + switch (block_size) { + case 2: + noise_tx->fft = aom_fft2x2_float; + noise_tx->ifft = aom_ifft2x2_float; + break; + case 4: + noise_tx->fft = aom_fft4x4_float; + noise_tx->ifft = aom_ifft4x4_float; + break; + case 8: + noise_tx->fft = aom_fft8x8_float; + noise_tx->ifft = aom_ifft8x8_float; + break; + case 16: + noise_tx->fft = aom_fft16x16_float; + noise_tx->ifft = aom_ifft16x16_float; + break; + case 32: + noise_tx->fft = aom_fft32x32_float; + noise_tx->ifft = aom_ifft32x32_float; + break; + default: + aom_free(noise_tx); + fprintf(stderr, "Unsupported block size %d\n", block_size); + return NULL; + } + noise_tx->block_size = block_size; + noise_tx->tx_block = (float *)aom_memalign( + 32, 2 * sizeof(*noise_tx->tx_block) * block_size * block_size); + noise_tx->temp = (float *)aom_memalign( + 32, 2 * sizeof(*noise_tx->temp) * block_size * block_size); + if (!noise_tx->tx_block || !noise_tx->temp) { + aom_noise_tx_free(noise_tx); + return NULL; + } + // Clear the buffers up front. Some outputs of the forward transform are + // real only (the imaginary component will never be touched) + memset(noise_tx->tx_block, 0, + 2 * sizeof(*noise_tx->tx_block) * block_size * block_size); + memset(noise_tx->temp, 0, + 2 * sizeof(*noise_tx->temp) * block_size * block_size); + return noise_tx; +} + +void aom_noise_tx_forward(struct aom_noise_tx_t *noise_tx, const float *data) { + noise_tx->fft(data, noise_tx->temp, noise_tx->tx_block); +} + +void aom_noise_tx_filter(struct aom_noise_tx_t *noise_tx, const float *psd) { + const int block_size = noise_tx->block_size; + const float kBeta = 1.1f; + const float kEps = 1e-6f; + for (int y = 0; y < block_size; ++y) { + for (int x = 0; x < block_size; ++x) { + int i = y * block_size + x; + float *c = noise_tx->tx_block + 2 * i; + const float p = c[0] * c[0] + c[1] * c[1]; + if (p > kBeta * psd[i] && p > 1e-6) { + noise_tx->tx_block[2 * i + 0] *= (p - psd[i]) / AOMMAX(p, kEps); + noise_tx->tx_block[2 * i + 1] *= (p - psd[i]) / AOMMAX(p, kEps); + } else { + noise_tx->tx_block[2 * i + 0] *= (kBeta - 1.0f) / kBeta; + noise_tx->tx_block[2 * i + 1] *= (kBeta - 1.0f) / kBeta; + } + } + } +} + +void aom_noise_tx_inverse(struct aom_noise_tx_t *noise_tx, float *data) { + const int n = noise_tx->block_size * noise_tx->block_size; + noise_tx->ifft(noise_tx->tx_block, noise_tx->temp, data); + for (int i = 0; i < n; ++i) { + data[i] /= n; + } +} + +void aom_noise_tx_add_energy(const struct aom_noise_tx_t *noise_tx, + float *psd) { + const int block_size = noise_tx->block_size; + for (int yb = 0; yb < block_size; ++yb) { + for (int xb = 0; xb <= block_size / 2; ++xb) { + float *c = noise_tx->tx_block + 2 * (yb * block_size + xb); + psd[yb * block_size + xb] += c[0] * c[0] + c[1] * c[1]; + } + } +} + +void aom_noise_tx_free(struct aom_noise_tx_t *noise_tx) { + if (!noise_tx) return; + aom_free(noise_tx->tx_block); + aom_free(noise_tx->temp); + aom_free(noise_tx); +} + +double aom_normalized_cross_correlation(const double *a, const double *b, + int n) { + double c = 0; + double a_len = 0; + double b_len = 0; + for (int i = 0; i < n; ++i) { + a_len += a[i] * a[i]; + b_len += b[i] * b[i]; + c += a[i] * b[i]; + } + return c / (sqrt(a_len) * sqrt(b_len)); +} + +int aom_noise_data_validate(const double *data, int w, int h) { + const double kVarianceThreshold = 2; + const double kMeanThreshold = 2; + + int x = 0, y = 0; + int ret_value = 1; + double var = 0, mean = 0; + double *mean_x, *mean_y, *var_x, *var_y; + + // Check that noise variance is not increasing in x or y + // and that the data is zero mean. + mean_x = (double *)aom_malloc(sizeof(*mean_x) * w); + var_x = (double *)aom_malloc(sizeof(*var_x) * w); + mean_y = (double *)aom_malloc(sizeof(*mean_x) * h); + var_y = (double *)aom_malloc(sizeof(*var_y) * h); + + memset(mean_x, 0, sizeof(*mean_x) * w); + memset(var_x, 0, sizeof(*var_x) * w); + memset(mean_y, 0, sizeof(*mean_y) * h); + memset(var_y, 0, sizeof(*var_y) * h); + + for (y = 0; y < h; ++y) { + for (x = 0; x < w; ++x) { + const double d = data[y * w + x]; + var_x[x] += d * d; + var_y[y] += d * d; + mean_x[x] += d; + mean_y[y] += d; + var += d * d; + mean += d; + } + } + mean /= (w * h); + var = var / (w * h) - mean * mean; + + for (y = 0; y < h; ++y) { + mean_y[y] /= h; + var_y[y] = var_y[y] / h - mean_y[y] * mean_y[y]; + if (fabs(var_y[y] - var) >= kVarianceThreshold) { + fprintf(stderr, "Variance distance too large %f %f\n", var_y[y], var); + ret_value = 0; + break; + } + if (fabs(mean_y[y] - mean) >= kMeanThreshold) { + fprintf(stderr, "Mean distance too large %f %f\n", mean_y[y], mean); + ret_value = 0; + break; + } + } + + for (x = 0; x < w; ++x) { + mean_x[x] /= w; + var_x[x] = var_x[x] / w - mean_x[x] * mean_x[x]; + if (fabs(var_x[x] - var) >= kVarianceThreshold) { + fprintf(stderr, "Variance distance too large %f %f\n", var_x[x], var); + ret_value = 0; + break; + } + if (fabs(mean_x[x] - mean) >= kMeanThreshold) { + fprintf(stderr, "Mean distance too large %f %f\n", mean_x[x], mean); + ret_value = 0; + break; + } + } + + aom_free(mean_x); + aom_free(mean_y); + aom_free(var_x); + aom_free(var_y); + + return ret_value; +} diff --git a/third_party/aom/aom_dsp/noise_util.h b/third_party/aom/aom_dsp/noise_util.h new file mode 100644 index 0000000000..2284a171a4 --- /dev/null +++ b/third_party/aom/aom_dsp/noise_util.h @@ -0,0 +1,68 @@ +/* + * Copyright (c) 2017, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_NOISE_UTIL_H_ +#define AOM_AOM_DSP_NOISE_UTIL_H_ + +#ifdef __cplusplus +extern "C" { +#endif // __cplusplus + +// aom_noise_tx_t is an abstraction of a transform that is used for denoising. +// It is meant to be lightweight and does hold the transformed data (as +// the user should not be manipulating the transformed data directly). +struct aom_noise_tx_t; + +// Allocates and returns a aom_noise_tx_t useful for denoising the given +// block_size. The resulting aom_noise_tx_t should be free'd with +// aom_noise_tx_free. +struct aom_noise_tx_t *aom_noise_tx_malloc(int block_size); +void aom_noise_tx_free(struct aom_noise_tx_t *aom_noise_tx); + +// Transforms the internal data and holds it in the aom_noise_tx's internal +// buffer. For compatibility with existing SIMD implementations, "data" must +// be 32-byte aligned. +void aom_noise_tx_forward(struct aom_noise_tx_t *aom_noise_tx, + const float *data); + +// Filters aom_noise_tx's internal data using the provided noise power spectral +// density. The PSD must be at least block_size * block_size and should be +// populated with a constant or via estimates taken from +// aom_noise_tx_add_energy. +void aom_noise_tx_filter(struct aom_noise_tx_t *aom_noise_tx, const float *psd); + +// Performs an inverse transform using the internal transform data. +// For compatibility with existing SIMD implementations, "data" must be 32-byte +// aligned. +void aom_noise_tx_inverse(struct aom_noise_tx_t *aom_noise_tx, float *data); + +// Aggregates the power of the buffered transform data into the psd buffer. +void aom_noise_tx_add_energy(const struct aom_noise_tx_t *aom_noise_tx, + float *psd); + +// Returns a default value suitable for denosing a transform of the given +// block_size. The noise "factor" determines the strength of the noise to +// be removed. A value of about 2.5 can be used for moderate denoising, +// where a value of 5.0 can be used for a high level of denoising. +float aom_noise_psd_get_default_value(int block_size, float factor); + +// Computes normalized cross correlation of two vectors a and b of length n. +double aom_normalized_cross_correlation(const double *a, const double *b, + int n); + +// Validates the correlated noise in the data buffer of size (w, h). +int aom_noise_data_validate(const double *data, int w, int h); + +#ifdef __cplusplus +} // extern "C" +#endif // __cplusplus + +#endif // AOM_AOM_DSP_NOISE_UTIL_H_ diff --git a/third_party/aom/aom_dsp/postproc.h b/third_party/aom/aom_dsp/postproc.h new file mode 100644 index 0000000000..f3d87f2642 --- /dev/null +++ b/third_party/aom/aom_dsp/postproc.h @@ -0,0 +1,26 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_POSTPROC_H_ +#define AOM_AOM_DSP_POSTPROC_H_ + +#ifdef __cplusplus +extern "C" { +#endif + +// Fills a noise buffer with gaussian noise strength determined by sigma. +int aom_setup_noise(double sigma, int size, char *noise); + +#ifdef __cplusplus +} +#endif + +#endif // AOM_AOM_DSP_POSTPROC_H_ diff --git a/third_party/aom/aom_dsp/prob.h b/third_party/aom/aom_dsp/prob.h new file mode 100644 index 0000000000..d003a986e6 --- /dev/null +++ b/third_party/aom/aom_dsp/prob.h @@ -0,0 +1,671 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_PROB_H_ +#define AOM_AOM_DSP_PROB_H_ + +#include +#include + +#include "config/aom_config.h" + +#include "aom_dsp/aom_dsp_common.h" +#include "aom_dsp/entcode.h" +#include "aom_ports/bitops.h" +#include "aom_ports/mem.h" + +#ifdef __cplusplus +extern "C" { +#endif + +// TODO(negge): Rename this aom_prob once we remove vpxbool. +typedef uint16_t aom_cdf_prob; + +#define CDF_SIZE(x) ((x) + 1) +#define CDF_PROB_BITS 15 +#define CDF_PROB_TOP (1 << CDF_PROB_BITS) +#define CDF_INIT_TOP 32768 +#define CDF_SHIFT (15 - CDF_PROB_BITS) +/*The value stored in an iCDF is CDF_PROB_TOP minus the actual cumulative + probability (an "inverse" CDF). + This function converts from one representation to the other (and is its own + inverse).*/ +#define AOM_ICDF(x) (CDF_PROB_TOP - (x)) + +#if CDF_SHIFT == 0 + +#define AOM_CDF2(a0) AOM_ICDF(a0), AOM_ICDF(CDF_PROB_TOP), 0 +#define AOM_CDF3(a0, a1) AOM_ICDF(a0), AOM_ICDF(a1), AOM_ICDF(CDF_PROB_TOP), 0 +#define AOM_CDF4(a0, a1, a2) \ + AOM_ICDF(a0), AOM_ICDF(a1), AOM_ICDF(a2), AOM_ICDF(CDF_PROB_TOP), 0 +#define AOM_CDF5(a0, a1, a2, a3) \ + AOM_ICDF(a0) \ + , AOM_ICDF(a1), AOM_ICDF(a2), AOM_ICDF(a3), AOM_ICDF(CDF_PROB_TOP), 0 +#define AOM_CDF6(a0, a1, a2, a3, a4) \ + AOM_ICDF(a0) \ + , AOM_ICDF(a1), AOM_ICDF(a2), AOM_ICDF(a3), AOM_ICDF(a4), \ + AOM_ICDF(CDF_PROB_TOP), 0 +#define AOM_CDF7(a0, a1, a2, a3, a4, a5) \ + AOM_ICDF(a0) \ + , AOM_ICDF(a1), AOM_ICDF(a2), AOM_ICDF(a3), AOM_ICDF(a4), AOM_ICDF(a5), \ + AOM_ICDF(CDF_PROB_TOP), 0 +#define AOM_CDF8(a0, a1, a2, a3, a4, a5, a6) \ + AOM_ICDF(a0) \ + , AOM_ICDF(a1), AOM_ICDF(a2), AOM_ICDF(a3), AOM_ICDF(a4), AOM_ICDF(a5), \ + AOM_ICDF(a6), AOM_ICDF(CDF_PROB_TOP), 0 +#define AOM_CDF9(a0, a1, a2, a3, a4, a5, a6, a7) \ + AOM_ICDF(a0) \ + , AOM_ICDF(a1), AOM_ICDF(a2), AOM_ICDF(a3), AOM_ICDF(a4), AOM_ICDF(a5), \ + AOM_ICDF(a6), AOM_ICDF(a7), AOM_ICDF(CDF_PROB_TOP), 0 +#define AOM_CDF10(a0, a1, a2, a3, a4, a5, a6, a7, a8) \ + AOM_ICDF(a0) \ + , AOM_ICDF(a1), AOM_ICDF(a2), AOM_ICDF(a3), AOM_ICDF(a4), AOM_ICDF(a5), \ + AOM_ICDF(a6), AOM_ICDF(a7), AOM_ICDF(a8), AOM_ICDF(CDF_PROB_TOP), 0 +#define AOM_CDF11(a0, a1, a2, a3, a4, a5, a6, a7, a8, a9) \ + AOM_ICDF(a0) \ + , AOM_ICDF(a1), AOM_ICDF(a2), AOM_ICDF(a3), AOM_ICDF(a4), AOM_ICDF(a5), \ + AOM_ICDF(a6), AOM_ICDF(a7), AOM_ICDF(a8), AOM_ICDF(a9), \ + AOM_ICDF(CDF_PROB_TOP), 0 +#define AOM_CDF12(a0, a1, a2, a3, a4, a5, a6, a7, a8, a9, a10) \ + AOM_ICDF(a0) \ + , AOM_ICDF(a1), AOM_ICDF(a2), AOM_ICDF(a3), AOM_ICDF(a4), AOM_ICDF(a5), \ + AOM_ICDF(a6), AOM_ICDF(a7), AOM_ICDF(a8), AOM_ICDF(a9), AOM_ICDF(a10), \ + AOM_ICDF(CDF_PROB_TOP), 0 +#define AOM_CDF13(a0, a1, a2, a3, a4, a5, a6, a7, a8, a9, a10, a11) \ + AOM_ICDF(a0) \ + , AOM_ICDF(a1), AOM_ICDF(a2), AOM_ICDF(a3), AOM_ICDF(a4), AOM_ICDF(a5), \ + AOM_ICDF(a6), AOM_ICDF(a7), AOM_ICDF(a8), AOM_ICDF(a9), AOM_ICDF(a10), \ + AOM_ICDF(a11), AOM_ICDF(CDF_PROB_TOP), 0 +#define AOM_CDF14(a0, a1, a2, a3, a4, a5, a6, a7, a8, a9, a10, a11, a12) \ + AOM_ICDF(a0) \ + , AOM_ICDF(a1), AOM_ICDF(a2), AOM_ICDF(a3), AOM_ICDF(a4), AOM_ICDF(a5), \ + AOM_ICDF(a6), AOM_ICDF(a7), AOM_ICDF(a8), AOM_ICDF(a9), AOM_ICDF(a10), \ + AOM_ICDF(a11), AOM_ICDF(a12), AOM_ICDF(CDF_PROB_TOP), 0 +#define AOM_CDF15(a0, a1, a2, a3, a4, a5, a6, a7, a8, a9, a10, a11, a12, a13) \ + AOM_ICDF(a0) \ + , AOM_ICDF(a1), AOM_ICDF(a2), AOM_ICDF(a3), AOM_ICDF(a4), AOM_ICDF(a5), \ + AOM_ICDF(a6), AOM_ICDF(a7), AOM_ICDF(a8), AOM_ICDF(a9), AOM_ICDF(a10), \ + AOM_ICDF(a11), AOM_ICDF(a12), AOM_ICDF(a13), AOM_ICDF(CDF_PROB_TOP), 0 +#define AOM_CDF16(a0, a1, a2, a3, a4, a5, a6, a7, a8, a9, a10, a11, a12, a13, \ + a14) \ + AOM_ICDF(a0) \ + , AOM_ICDF(a1), AOM_ICDF(a2), AOM_ICDF(a3), AOM_ICDF(a4), AOM_ICDF(a5), \ + AOM_ICDF(a6), AOM_ICDF(a7), AOM_ICDF(a8), AOM_ICDF(a9), AOM_ICDF(a10), \ + AOM_ICDF(a11), AOM_ICDF(a12), AOM_ICDF(a13), AOM_ICDF(a14), \ + AOM_ICDF(CDF_PROB_TOP), 0 + +#else +#define AOM_CDF2(a0) \ + AOM_ICDF((((a0)-1) * ((CDF_INIT_TOP >> CDF_SHIFT) - 2) + \ + ((CDF_INIT_TOP - 2) >> 1)) / \ + ((CDF_INIT_TOP - 2)) + \ + 1) \ + , AOM_ICDF(CDF_PROB_TOP), 0 +#define AOM_CDF3(a0, a1) \ + AOM_ICDF((((a0)-1) * ((CDF_INIT_TOP >> CDF_SHIFT) - 3) + \ + ((CDF_INIT_TOP - 3) >> 1)) / \ + ((CDF_INIT_TOP - 3)) + \ + 1) \ + , \ + AOM_ICDF((((a1)-2) * ((CDF_INIT_TOP >> CDF_SHIFT) - 3) + \ + ((CDF_INIT_TOP - 3) >> 1)) / \ + ((CDF_INIT_TOP - 3)) + \ + 2), \ + AOM_ICDF(CDF_PROB_TOP), 0 +#define AOM_CDF4(a0, a1, a2) \ + AOM_ICDF((((a0)-1) * ((CDF_INIT_TOP >> CDF_SHIFT) - 4) + \ + ((CDF_INIT_TOP - 4) >> 1)) / \ + ((CDF_INIT_TOP - 4)) + \ + 1) \ + , \ + AOM_ICDF((((a1)-2) * ((CDF_INIT_TOP >> CDF_SHIFT) - 4) + \ + ((CDF_INIT_TOP - 4) >> 1)) / \ + ((CDF_INIT_TOP - 4)) + \ + 2), \ + AOM_ICDF((((a2)-3) * ((CDF_INIT_TOP >> CDF_SHIFT) - 4) + \ + ((CDF_INIT_TOP - 4) >> 1)) / \ + ((CDF_INIT_TOP - 4)) + \ + 3), \ + AOM_ICDF(CDF_PROB_TOP), 0 +#define AOM_CDF5(a0, a1, a2, a3) \ + AOM_ICDF((((a0)-1) * ((CDF_INIT_TOP >> CDF_SHIFT) - 5) + \ + ((CDF_INIT_TOP - 5) >> 1)) / \ + ((CDF_INIT_TOP - 5)) + \ + 1) \ + , \ + AOM_ICDF((((a1)-2) * ((CDF_INIT_TOP >> CDF_SHIFT) - 5) + \ + ((CDF_INIT_TOP - 5) >> 1)) / \ + ((CDF_INIT_TOP - 5)) + \ + 2), \ + AOM_ICDF((((a2)-3) * ((CDF_INIT_TOP >> CDF_SHIFT) - 5) + \ + ((CDF_INIT_TOP - 5) >> 1)) / \ + ((CDF_INIT_TOP - 5)) + \ + 3), \ + AOM_ICDF((((a3)-4) * ((CDF_INIT_TOP >> CDF_SHIFT) - 5) + \ + ((CDF_INIT_TOP - 5) >> 1)) / \ + ((CDF_INIT_TOP - 5)) + \ + 4), \ + AOM_ICDF(CDF_PROB_TOP), 0 +#define AOM_CDF6(a0, a1, a2, a3, a4) \ + AOM_ICDF((((a0)-1) * ((CDF_INIT_TOP >> CDF_SHIFT) - 6) + \ + ((CDF_INIT_TOP - 6) >> 1)) / \ + ((CDF_INIT_TOP - 6)) + \ + 1) \ + , \ + AOM_ICDF((((a1)-2) * ((CDF_INIT_TOP >> CDF_SHIFT) - 6) + \ + ((CDF_INIT_TOP - 6) >> 1)) / \ + ((CDF_INIT_TOP - 6)) + \ + 2), \ + AOM_ICDF((((a2)-3) * ((CDF_INIT_TOP >> CDF_SHIFT) - 6) + \ + ((CDF_INIT_TOP - 6) >> 1)) / \ + ((CDF_INIT_TOP - 6)) + \ + 3), \ + AOM_ICDF((((a3)-4) * ((CDF_INIT_TOP >> CDF_SHIFT) - 6) + \ + ((CDF_INIT_TOP - 6) >> 1)) / \ + ((CDF_INIT_TOP - 6)) + \ + 4), \ + AOM_ICDF((((a4)-5) * ((CDF_INIT_TOP >> CDF_SHIFT) - 6) + \ + ((CDF_INIT_TOP - 6) >> 1)) / \ + ((CDF_INIT_TOP - 6)) + \ + 5), \ + AOM_ICDF(CDF_PROB_TOP), 0 +#define AOM_CDF7(a0, a1, a2, a3, a4, a5) \ + AOM_ICDF((((a0)-1) * ((CDF_INIT_TOP >> CDF_SHIFT) - 7) + \ + ((CDF_INIT_TOP - 7) >> 1)) / \ + ((CDF_INIT_TOP - 7)) + \ + 1) \ + , \ + AOM_ICDF((((a1)-2) * ((CDF_INIT_TOP >> CDF_SHIFT) - 7) + \ + ((CDF_INIT_TOP - 7) >> 1)) / \ + ((CDF_INIT_TOP - 7)) + \ + 2), \ + AOM_ICDF((((a2)-3) * ((CDF_INIT_TOP >> CDF_SHIFT) - 7) + \ + ((CDF_INIT_TOP - 7) >> 1)) / \ + ((CDF_INIT_TOP - 7)) + \ + 3), \ + AOM_ICDF((((a3)-4) * ((CDF_INIT_TOP >> CDF_SHIFT) - 7) + \ + ((CDF_INIT_TOP - 7) >> 1)) / \ + ((CDF_INIT_TOP - 7)) + \ + 4), \ + AOM_ICDF((((a4)-5) * ((CDF_INIT_TOP >> CDF_SHIFT) - 7) + \ + ((CDF_INIT_TOP - 7) >> 1)) / \ + ((CDF_INIT_TOP - 7)) + \ + 5), \ + AOM_ICDF((((a5)-6) * ((CDF_INIT_TOP >> CDF_SHIFT) - 7) + \ + ((CDF_INIT_TOP - 7) >> 1)) / \ + ((CDF_INIT_TOP - 7)) + \ + 6), \ + AOM_ICDF(CDF_PROB_TOP), 0 +#define AOM_CDF8(a0, a1, a2, a3, a4, a5, a6) \ + AOM_ICDF((((a0)-1) * ((CDF_INIT_TOP >> CDF_SHIFT) - 8) + \ + ((CDF_INIT_TOP - 8) >> 1)) / \ + ((CDF_INIT_TOP - 8)) + \ + 1) \ + , \ + AOM_ICDF((((a1)-2) * ((CDF_INIT_TOP >> CDF_SHIFT) - 8) + \ + ((CDF_INIT_TOP - 8) >> 1)) / \ + ((CDF_INIT_TOP - 8)) + \ + 2), \ + AOM_ICDF((((a2)-3) * ((CDF_INIT_TOP >> CDF_SHIFT) - 8) + \ + ((CDF_INIT_TOP - 8) >> 1)) / \ + ((CDF_INIT_TOP - 8)) + \ + 3), \ + AOM_ICDF((((a3)-4) * ((CDF_INIT_TOP >> CDF_SHIFT) - 8) + \ + ((CDF_INIT_TOP - 8) >> 1)) / \ + ((CDF_INIT_TOP - 8)) + \ + 4), \ + AOM_ICDF((((a4)-5) * ((CDF_INIT_TOP >> CDF_SHIFT) - 8) + \ + ((CDF_INIT_TOP - 8) >> 1)) / \ + ((CDF_INIT_TOP - 8)) + \ + 5), \ + AOM_ICDF((((a5)-6) * ((CDF_INIT_TOP >> CDF_SHIFT) - 8) + \ + ((CDF_INIT_TOP - 8) >> 1)) / \ + ((CDF_INIT_TOP - 8)) + \ + 6), \ + AOM_ICDF((((a6)-7) * ((CDF_INIT_TOP >> CDF_SHIFT) - 8) + \ + ((CDF_INIT_TOP - 8) >> 1)) / \ + ((CDF_INIT_TOP - 8)) + \ + 7), \ + AOM_ICDF(CDF_PROB_TOP), 0 +#define AOM_CDF9(a0, a1, a2, a3, a4, a5, a6, a7) \ + AOM_ICDF((((a0)-1) * ((CDF_INIT_TOP >> CDF_SHIFT) - 9) + \ + ((CDF_INIT_TOP - 9) >> 1)) / \ + ((CDF_INIT_TOP - 9)) + \ + 1) \ + , \ + AOM_ICDF((((a1)-2) * ((CDF_INIT_TOP >> CDF_SHIFT) - 9) + \ + ((CDF_INIT_TOP - 9) >> 1)) / \ + ((CDF_INIT_TOP - 9)) + \ + 2), \ + AOM_ICDF((((a2)-3) * ((CDF_INIT_TOP >> CDF_SHIFT) - 9) + \ + ((CDF_INIT_TOP - 9) >> 1)) / \ + ((CDF_INIT_TOP - 9)) + \ + 3), \ + AOM_ICDF((((a3)-4) * ((CDF_INIT_TOP >> CDF_SHIFT) - 9) + \ + ((CDF_INIT_TOP - 9) >> 1)) / \ + ((CDF_INIT_TOP - 9)) + \ + 4), \ + AOM_ICDF((((a4)-5) * ((CDF_INIT_TOP >> CDF_SHIFT) - 9) + \ + ((CDF_INIT_TOP - 9) >> 1)) / \ + ((CDF_INIT_TOP - 9)) + \ + 5), \ + AOM_ICDF((((a5)-6) * ((CDF_INIT_TOP >> CDF_SHIFT) - 9) + \ + ((CDF_INIT_TOP - 9) >> 1)) / \ + ((CDF_INIT_TOP - 9)) + \ + 6), \ + AOM_ICDF((((a6)-7) * ((CDF_INIT_TOP >> CDF_SHIFT) - 9) + \ + ((CDF_INIT_TOP - 9) >> 1)) / \ + ((CDF_INIT_TOP - 9)) + \ + 7), \ + AOM_ICDF((((a7)-8) * ((CDF_INIT_TOP >> CDF_SHIFT) - 9) + \ + ((CDF_INIT_TOP - 9) >> 1)) / \ + ((CDF_INIT_TOP - 9)) + \ + 8), \ + AOM_ICDF(CDF_PROB_TOP), 0 +#define AOM_CDF10(a0, a1, a2, a3, a4, a5, a6, a7, a8) \ + AOM_ICDF((((a0)-1) * ((CDF_INIT_TOP >> CDF_SHIFT) - 10) + \ + ((CDF_INIT_TOP - 10) >> 1)) / \ + ((CDF_INIT_TOP - 10)) + \ + 1) \ + , \ + AOM_ICDF((((a1)-2) * ((CDF_INIT_TOP >> CDF_SHIFT) - 10) + \ + ((CDF_INIT_TOP - 10) >> 1)) / \ + ((CDF_INIT_TOP - 10)) + \ + 2), \ + AOM_ICDF((((a2)-3) * ((CDF_INIT_TOP >> CDF_SHIFT) - 10) + \ + ((CDF_INIT_TOP - 10) >> 1)) / \ + ((CDF_INIT_TOP - 10)) + \ + 3), \ + AOM_ICDF((((a3)-4) * ((CDF_INIT_TOP >> CDF_SHIFT) - 10) + \ + ((CDF_INIT_TOP - 10) >> 1)) / \ + ((CDF_INIT_TOP - 10)) + \ + 4), \ + AOM_ICDF((((a4)-5) * ((CDF_INIT_TOP >> CDF_SHIFT) - 10) + \ + ((CDF_INIT_TOP - 10) >> 1)) / \ + ((CDF_INIT_TOP - 10)) + \ + 5), \ + AOM_ICDF((((a5)-6) * ((CDF_INIT_TOP >> CDF_SHIFT) - 10) + \ + ((CDF_INIT_TOP - 10) >> 1)) / \ + ((CDF_INIT_TOP - 10)) + \ + 6), \ + AOM_ICDF((((a6)-7) * ((CDF_INIT_TOP >> CDF_SHIFT) - 10) + \ + ((CDF_INIT_TOP - 10) >> 1)) / \ + ((CDF_INIT_TOP - 10)) + \ + 7), \ + AOM_ICDF((((a7)-8) * ((CDF_INIT_TOP >> CDF_SHIFT) - 10) + \ + ((CDF_INIT_TOP - 10) >> 1)) / \ + ((CDF_INIT_TOP - 10)) + \ + 8), \ + AOM_ICDF((((a8)-9) * ((CDF_INIT_TOP >> CDF_SHIFT) - 10) + \ + ((CDF_INIT_TOP - 10) >> 1)) / \ + ((CDF_INIT_TOP - 10)) + \ + 9), \ + AOM_ICDF(CDF_PROB_TOP), 0 +#define AOM_CDF11(a0, a1, a2, a3, a4, a5, a6, a7, a8, a9) \ + AOM_ICDF((((a0)-1) * ((CDF_INIT_TOP >> CDF_SHIFT) - 11) + \ + ((CDF_INIT_TOP - 11) >> 1)) / \ + ((CDF_INIT_TOP - 11)) + \ + 1) \ + , \ + AOM_ICDF((((a1)-2) * ((CDF_INIT_TOP >> CDF_SHIFT) - 11) + \ + ((CDF_INIT_TOP - 11) >> 1)) / \ + ((CDF_INIT_TOP - 11)) + \ + 2), \ + AOM_ICDF((((a2)-3) * ((CDF_INIT_TOP >> CDF_SHIFT) - 11) + \ + ((CDF_INIT_TOP - 11) >> 1)) / \ + ((CDF_INIT_TOP - 11)) + \ + 3), \ + AOM_ICDF((((a3)-4) * ((CDF_INIT_TOP >> CDF_SHIFT) - 11) + \ + ((CDF_INIT_TOP - 11) >> 1)) / \ + ((CDF_INIT_TOP - 11)) + \ + 4), \ + AOM_ICDF((((a4)-5) * ((CDF_INIT_TOP >> CDF_SHIFT) - 11) + \ + ((CDF_INIT_TOP - 11) >> 1)) / \ + ((CDF_INIT_TOP - 11)) + \ + 5), \ + AOM_ICDF((((a5)-6) * ((CDF_INIT_TOP >> CDF_SHIFT) - 11) + \ + ((CDF_INIT_TOP - 11) >> 1)) / \ + ((CDF_INIT_TOP - 11)) + \ + 6), \ + AOM_ICDF((((a6)-7) * ((CDF_INIT_TOP >> CDF_SHIFT) - 11) + \ + ((CDF_INIT_TOP - 11) >> 1)) / \ + ((CDF_INIT_TOP - 11)) + \ + 7), \ + AOM_ICDF((((a7)-8) * ((CDF_INIT_TOP >> CDF_SHIFT) - 11) + \ + ((CDF_INIT_TOP - 11) >> 1)) / \ + ((CDF_INIT_TOP - 11)) + \ + 8), \ + AOM_ICDF((((a8)-9) * ((CDF_INIT_TOP >> CDF_SHIFT) - 11) + \ + ((CDF_INIT_TOP - 11) >> 1)) / \ + ((CDF_INIT_TOP - 11)) + \ + 9), \ + AOM_ICDF((((a9)-10) * ((CDF_INIT_TOP >> CDF_SHIFT) - 11) + \ + ((CDF_INIT_TOP - 11) >> 1)) / \ + ((CDF_INIT_TOP - 11)) + \ + 10), \ + AOM_ICDF(CDF_PROB_TOP), 0 +#define AOM_CDF12(a0, a1, a2, a3, a4, a5, a6, a7, a8, a9, a10) \ + AOM_ICDF((((a0)-1) * ((CDF_INIT_TOP >> CDF_SHIFT) - 12) + \ + ((CDF_INIT_TOP - 12) >> 1)) / \ + ((CDF_INIT_TOP - 12)) + \ + 1) \ + , \ + AOM_ICDF((((a1)-2) * ((CDF_INIT_TOP >> CDF_SHIFT) - 12) + \ + ((CDF_INIT_TOP - 12) >> 1)) / \ + ((CDF_INIT_TOP - 12)) + \ + 2), \ + AOM_ICDF((((a2)-3) * ((CDF_INIT_TOP >> CDF_SHIFT) - 12) + \ + ((CDF_INIT_TOP - 12) >> 1)) / \ + ((CDF_INIT_TOP - 12)) + \ + 3), \ + AOM_ICDF((((a3)-4) * ((CDF_INIT_TOP >> CDF_SHIFT) - 12) + \ + ((CDF_INIT_TOP - 12) >> 1)) / \ + ((CDF_INIT_TOP - 12)) + \ + 4), \ + AOM_ICDF((((a4)-5) * ((CDF_INIT_TOP >> CDF_SHIFT) - 12) + \ + ((CDF_INIT_TOP - 12) >> 1)) / \ + ((CDF_INIT_TOP - 12)) + \ + 5), \ + AOM_ICDF((((a5)-6) * ((CDF_INIT_TOP >> CDF_SHIFT) - 12) + \ + ((CDF_INIT_TOP - 12) >> 1)) / \ + ((CDF_INIT_TOP - 12)) + \ + 6), \ + AOM_ICDF((((a6)-7) * ((CDF_INIT_TOP >> CDF_SHIFT) - 12) + \ + ((CDF_INIT_TOP - 12) >> 1)) / \ + ((CDF_INIT_TOP - 12)) + \ + 7), \ + AOM_ICDF((((a7)-8) * ((CDF_INIT_TOP >> CDF_SHIFT) - 12) + \ + ((CDF_INIT_TOP - 12) >> 1)) / \ + ((CDF_INIT_TOP - 12)) + \ + 8), \ + AOM_ICDF((((a8)-9) * ((CDF_INIT_TOP >> CDF_SHIFT) - 12) + \ + ((CDF_INIT_TOP - 12) >> 1)) / \ + ((CDF_INIT_TOP - 12)) + \ + 9), \ + AOM_ICDF((((a9)-10) * ((CDF_INIT_TOP >> CDF_SHIFT) - 12) + \ + ((CDF_INIT_TOP - 12) >> 1)) / \ + ((CDF_INIT_TOP - 12)) + \ + 10), \ + AOM_ICDF((((a10)-11) * ((CDF_INIT_TOP >> CDF_SHIFT) - 12) + \ + ((CDF_INIT_TOP - 12) >> 1)) / \ + ((CDF_INIT_TOP - 12)) + \ + 11), \ + AOM_ICDF(CDF_PROB_TOP), 0 +#define AOM_CDF13(a0, a1, a2, a3, a4, a5, a6, a7, a8, a9, a10, a11) \ + AOM_ICDF((((a0)-1) * ((CDF_INIT_TOP >> CDF_SHIFT) - 13) + \ + ((CDF_INIT_TOP - 13) >> 1)) / \ + ((CDF_INIT_TOP - 13)) + \ + 1) \ + , \ + AOM_ICDF((((a1)-2) * ((CDF_INIT_TOP >> CDF_SHIFT) - 13) + \ + ((CDF_INIT_TOP - 13) >> 1)) / \ + ((CDF_INIT_TOP - 13)) + \ + 2), \ + AOM_ICDF((((a2)-3) * ((CDF_INIT_TOP >> CDF_SHIFT) - 13) + \ + ((CDF_INIT_TOP - 13) >> 1)) / \ + ((CDF_INIT_TOP - 13)) + \ + 3), \ + AOM_ICDF((((a3)-4) * ((CDF_INIT_TOP >> CDF_SHIFT) - 13) + \ + ((CDF_INIT_TOP - 13) >> 1)) / \ + ((CDF_INIT_TOP - 13)) + \ + 4), \ + AOM_ICDF((((a4)-5) * ((CDF_INIT_TOP >> CDF_SHIFT) - 13) + \ + ((CDF_INIT_TOP - 13) >> 1)) / \ + ((CDF_INIT_TOP - 13)) + \ + 5), \ + AOM_ICDF((((a5)-6) * ((CDF_INIT_TOP >> CDF_SHIFT) - 13) + \ + ((CDF_INIT_TOP - 13) >> 1)) / \ + ((CDF_INIT_TOP - 13)) + \ + 6), \ + AOM_ICDF((((a6)-7) * ((CDF_INIT_TOP >> CDF_SHIFT) - 13) + \ + ((CDF_INIT_TOP - 13) >> 1)) / \ + ((CDF_INIT_TOP - 13)) + \ + 7), \ + AOM_ICDF((((a7)-8) * ((CDF_INIT_TOP >> CDF_SHIFT) - 13) + \ + ((CDF_INIT_TOP - 13) >> 1)) / \ + ((CDF_INIT_TOP - 13)) + \ + 8), \ + AOM_ICDF((((a8)-9) * ((CDF_INIT_TOP >> CDF_SHIFT) - 13) + \ + ((CDF_INIT_TOP - 13) >> 1)) / \ + ((CDF_INIT_TOP - 13)) + \ + 9), \ + AOM_ICDF((((a9)-10) * ((CDF_INIT_TOP >> CDF_SHIFT) - 13) + \ + ((CDF_INIT_TOP - 13) >> 1)) / \ + ((CDF_INIT_TOP - 13)) + \ + 10), \ + AOM_ICDF((((a10)-11) * ((CDF_INIT_TOP >> CDF_SHIFT) - 13) + \ + ((CDF_INIT_TOP - 13) >> 1)) / \ + ((CDF_INIT_TOP - 13)) + \ + 11), \ + AOM_ICDF((((a11)-12) * ((CDF_INIT_TOP >> CDF_SHIFT) - 13) + \ + ((CDF_INIT_TOP - 13) >> 1)) / \ + ((CDF_INIT_TOP - 13)) + \ + 12), \ + AOM_ICDF(CDF_PROB_TOP), 0 +#define AOM_CDF14(a0, a1, a2, a3, a4, a5, a6, a7, a8, a9, a10, a11, a12) \ + AOM_ICDF((((a0)-1) * ((CDF_INIT_TOP >> CDF_SHIFT) - 14) + \ + ((CDF_INIT_TOP - 14) >> 1)) / \ + ((CDF_INIT_TOP - 14)) + \ + 1) \ + , \ + AOM_ICDF((((a1)-2) * ((CDF_INIT_TOP >> CDF_SHIFT) - 14) + \ + ((CDF_INIT_TOP - 14) >> 1)) / \ + ((CDF_INIT_TOP - 14)) + \ + 2), \ + AOM_ICDF((((a2)-3) * ((CDF_INIT_TOP >> CDF_SHIFT) - 14) + \ + ((CDF_INIT_TOP - 14) >> 1)) / \ + ((CDF_INIT_TOP - 14)) + \ + 3), \ + AOM_ICDF((((a3)-4) * ((CDF_INIT_TOP >> CDF_SHIFT) - 14) + \ + ((CDF_INIT_TOP - 14) >> 1)) / \ + ((CDF_INIT_TOP - 14)) + \ + 4), \ + AOM_ICDF((((a4)-5) * ((CDF_INIT_TOP >> CDF_SHIFT) - 14) + \ + ((CDF_INIT_TOP - 14) >> 1)) / \ + ((CDF_INIT_TOP - 14)) + \ + 5), \ + AOM_ICDF((((a5)-6) * ((CDF_INIT_TOP >> CDF_SHIFT) - 14) + \ + ((CDF_INIT_TOP - 14) >> 1)) / \ + ((CDF_INIT_TOP - 14)) + \ + 6), \ + AOM_ICDF((((a6)-7) * ((CDF_INIT_TOP >> CDF_SHIFT) - 14) + \ + ((CDF_INIT_TOP - 14) >> 1)) / \ + ((CDF_INIT_TOP - 14)) + \ + 7), \ + AOM_ICDF((((a7)-8) * ((CDF_INIT_TOP >> CDF_SHIFT) - 14) + \ + ((CDF_INIT_TOP - 14) >> 1)) / \ + ((CDF_INIT_TOP - 14)) + \ + 8), \ + AOM_ICDF((((a8)-9) * ((CDF_INIT_TOP >> CDF_SHIFT) - 14) + \ + ((CDF_INIT_TOP - 14) >> 1)) / \ + ((CDF_INIT_TOP - 14)) + \ + 9), \ + AOM_ICDF((((a9)-10) * ((CDF_INIT_TOP >> CDF_SHIFT) - 14) + \ + ((CDF_INIT_TOP - 14) >> 1)) / \ + ((CDF_INIT_TOP - 14)) + \ + 10), \ + AOM_ICDF((((a10)-11) * ((CDF_INIT_TOP >> CDF_SHIFT) - 14) + \ + ((CDF_INIT_TOP - 14) >> 1)) / \ + ((CDF_INIT_TOP - 14)) + \ + 11), \ + AOM_ICDF((((a11)-12) * ((CDF_INIT_TOP >> CDF_SHIFT) - 14) + \ + ((CDF_INIT_TOP - 14) >> 1)) / \ + ((CDF_INIT_TOP - 14)) + \ + 12), \ + AOM_ICDF((((a12)-13) * ((CDF_INIT_TOP >> CDF_SHIFT) - 14) + \ + ((CDF_INIT_TOP - 14) >> 1)) / \ + ((CDF_INIT_TOP - 14)) + \ + 13), \ + AOM_ICDF(CDF_PROB_TOP), 0 +#define AOM_CDF15(a0, a1, a2, a3, a4, a5, a6, a7, a8, a9, a10, a11, a12, a13) \ + AOM_ICDF((((a0)-1) * ((CDF_INIT_TOP >> CDF_SHIFT) - 15) + \ + ((CDF_INIT_TOP - 15) >> 1)) / \ + ((CDF_INIT_TOP - 15)) + \ + 1) \ + , \ + AOM_ICDF((((a1)-2) * ((CDF_INIT_TOP >> CDF_SHIFT) - 15) + \ + ((CDF_INIT_TOP - 15) >> 1)) / \ + ((CDF_INIT_TOP - 15)) + \ + 2), \ + AOM_ICDF((((a2)-3) * ((CDF_INIT_TOP >> CDF_SHIFT) - 15) + \ + ((CDF_INIT_TOP - 15) >> 1)) / \ + ((CDF_INIT_TOP - 15)) + \ + 3), \ + AOM_ICDF((((a3)-4) * ((CDF_INIT_TOP >> CDF_SHIFT) - 15) + \ + ((CDF_INIT_TOP - 15) >> 1)) / \ + ((CDF_INIT_TOP - 15)) + \ + 4), \ + AOM_ICDF((((a4)-5) * ((CDF_INIT_TOP >> CDF_SHIFT) - 15) + \ + ((CDF_INIT_TOP - 15) >> 1)) / \ + ((CDF_INIT_TOP - 15)) + \ + 5), \ + AOM_ICDF((((a5)-6) * ((CDF_INIT_TOP >> CDF_SHIFT) - 15) + \ + ((CDF_INIT_TOP - 15) >> 1)) / \ + ((CDF_INIT_TOP - 15)) + \ + 6), \ + AOM_ICDF((((a6)-7) * ((CDF_INIT_TOP >> CDF_SHIFT) - 15) + \ + ((CDF_INIT_TOP - 15) >> 1)) / \ + ((CDF_INIT_TOP - 15)) + \ + 7), \ + AOM_ICDF((((a7)-8) * ((CDF_INIT_TOP >> CDF_SHIFT) - 15) + \ + ((CDF_INIT_TOP - 15) >> 1)) / \ + ((CDF_INIT_TOP - 15)) + \ + 8), \ + AOM_ICDF((((a8)-9) * ((CDF_INIT_TOP >> CDF_SHIFT) - 15) + \ + ((CDF_INIT_TOP - 15) >> 1)) / \ + ((CDF_INIT_TOP - 15)) + \ + 9), \ + AOM_ICDF((((a9)-10) * ((CDF_INIT_TOP >> CDF_SHIFT) - 15) + \ + ((CDF_INIT_TOP - 15) >> 1)) / \ + ((CDF_INIT_TOP - 15)) + \ + 10), \ + AOM_ICDF((((a10)-11) * ((CDF_INIT_TOP >> CDF_SHIFT) - 15) + \ + ((CDF_INIT_TOP - 15) >> 1)) / \ + ((CDF_INIT_TOP - 15)) + \ + 11), \ + AOM_ICDF((((a11)-12) * ((CDF_INIT_TOP >> CDF_SHIFT) - 15) + \ + ((CDF_INIT_TOP - 15) >> 1)) / \ + ((CDF_INIT_TOP - 15)) + \ + 12), \ + AOM_ICDF((((a12)-13) * ((CDF_INIT_TOP >> CDF_SHIFT) - 15) + \ + ((CDF_INIT_TOP - 15) >> 1)) / \ + ((CDF_INIT_TOP - 15)) + \ + 13), \ + AOM_ICDF((((a13)-14) * ((CDF_INIT_TOP >> CDF_SHIFT) - 15) + \ + ((CDF_INIT_TOP - 15) >> 1)) / \ + ((CDF_INIT_TOP - 15)) + \ + 14), \ + AOM_ICDF(CDF_PROB_TOP), 0 +#define AOM_CDF16(a0, a1, a2, a3, a4, a5, a6, a7, a8, a9, a10, a11, a12, a13, \ + a14) \ + AOM_ICDF((((a0)-1) * ((CDF_INIT_TOP >> CDF_SHIFT) - 16) + \ + ((CDF_INIT_TOP - 16) >> 1)) / \ + ((CDF_INIT_TOP - 16)) + \ + 1) \ + , \ + AOM_ICDF((((a1)-2) * ((CDF_INIT_TOP >> CDF_SHIFT) - 16) + \ + ((CDF_INIT_TOP - 16) >> 1)) / \ + ((CDF_INIT_TOP - 16)) + \ + 2), \ + AOM_ICDF((((a2)-3) * ((CDF_INIT_TOP >> CDF_SHIFT) - 16) + \ + ((CDF_INIT_TOP - 16) >> 1)) / \ + ((CDF_INIT_TOP - 16)) + \ + 3), \ + AOM_ICDF((((a3)-4) * ((CDF_INIT_TOP >> CDF_SHIFT) - 16) + \ + ((CDF_INIT_TOP - 16) >> 1)) / \ + ((CDF_INIT_TOP - 16)) + \ + 4), \ + AOM_ICDF((((a4)-5) * ((CDF_INIT_TOP >> CDF_SHIFT) - 16) + \ + ((CDF_INIT_TOP - 16) >> 1)) / \ + ((CDF_INIT_TOP - 16)) + \ + 5), \ + AOM_ICDF((((a5)-6) * ((CDF_INIT_TOP >> CDF_SHIFT) - 16) + \ + ((CDF_INIT_TOP - 16) >> 1)) / \ + ((CDF_INIT_TOP - 16)) + \ + 6), \ + AOM_ICDF((((a6)-7) * ((CDF_INIT_TOP >> CDF_SHIFT) - 16) + \ + ((CDF_INIT_TOP - 16) >> 1)) / \ + ((CDF_INIT_TOP - 16)) + \ + 7), \ + AOM_ICDF((((a7)-8) * ((CDF_INIT_TOP >> CDF_SHIFT) - 16) + \ + ((CDF_INIT_TOP - 16) >> 1)) / \ + ((CDF_INIT_TOP - 16)) + \ + 8), \ + AOM_ICDF((((a8)-9) * ((CDF_INIT_TOP >> CDF_SHIFT) - 16) + \ + ((CDF_INIT_TOP - 16) >> 1)) / \ + ((CDF_INIT_TOP - 16)) + \ + 9), \ + AOM_ICDF((((a9)-10) * ((CDF_INIT_TOP >> CDF_SHIFT) - 16) + \ + ((CDF_INIT_TOP - 16) >> 1)) / \ + ((CDF_INIT_TOP - 16)) + \ + 10), \ + AOM_ICDF((((a10)-11) * ((CDF_INIT_TOP >> CDF_SHIFT) - 16) + \ + ((CDF_INIT_TOP - 16) >> 1)) / \ + ((CDF_INIT_TOP - 16)) + \ + 11), \ + AOM_ICDF((((a11)-12) * ((CDF_INIT_TOP >> CDF_SHIFT) - 16) + \ + ((CDF_INIT_TOP - 16) >> 1)) / \ + ((CDF_INIT_TOP - 16)) + \ + 12), \ + AOM_ICDF((((a12)-13) * ((CDF_INIT_TOP >> CDF_SHIFT) - 16) + \ + ((CDF_INIT_TOP - 16) >> 1)) / \ + ((CDF_INIT_TOP - 16)) + \ + 13), \ + AOM_ICDF((((a13)-14) * ((CDF_INIT_TOP >> CDF_SHIFT) - 16) + \ + ((CDF_INIT_TOP - 16) >> 1)) / \ + ((CDF_INIT_TOP - 16)) + \ + 14), \ + AOM_ICDF((((a14)-15) * ((CDF_INIT_TOP >> CDF_SHIFT) - 16) + \ + ((CDF_INIT_TOP - 16) >> 1)) / \ + ((CDF_INIT_TOP - 16)) + \ + 15), \ + AOM_ICDF(CDF_PROB_TOP), 0 + +#endif + +static INLINE uint8_t get_prob(unsigned int num, unsigned int den) { + assert(den != 0); + { + const int p = (int)(((uint64_t)num * 256 + (den >> 1)) / den); + // (p > 255) ? 255 : (p < 1) ? 1 : p; + const int clipped_prob = p | ((255 - p) >> 23) | (p == 0); + return (uint8_t)clipped_prob; + } +} + +static INLINE void update_cdf(aom_cdf_prob *cdf, int val, int nsymbs) { + int rate; + int i, tmp; + + static const int nsymbs2speed[17] = { 0, 0, 1, 1, 2, 2, 2, 2, 2, + 2, 2, 2, 2, 2, 2, 2, 2 }; + assert(nsymbs < 17); + rate = 3 + (cdf[nsymbs] > 15) + (cdf[nsymbs] > 31) + + nsymbs2speed[nsymbs]; // + get_msb(nsymbs); + tmp = AOM_ICDF(0); + + // Single loop (faster) + for (i = 0; i < nsymbs - 1; ++i) { + tmp = (i == val) ? 0 : tmp; + if (tmp < cdf[i]) { + cdf[i] -= ((cdf[i] - tmp) >> rate); + } else { + cdf[i] += ((tmp - cdf[i]) >> rate); + } + } + cdf[nsymbs] += (cdf[nsymbs] < 32); +} + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif // AOM_AOM_DSP_PROB_H_ diff --git a/third_party/aom/aom_dsp/psnr.c b/third_party/aom/aom_dsp/psnr.c new file mode 100644 index 0000000000..50f376a4a1 --- /dev/null +++ b/third_party/aom/aom_dsp/psnr.c @@ -0,0 +1,381 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include +#include + +#include "config/aom_dsp_rtcd.h" + +#include "aom_dsp/psnr.h" +#include "aom_scale/yv12config.h" + +double aom_sse_to_psnr(double samples, double peak, double sse) { + if (sse > 0.0) { + const double psnr = 10.0 * log10(samples * peak * peak / sse); + return psnr > MAX_PSNR ? MAX_PSNR : psnr; + } else { + return MAX_PSNR; + } +} + +/* TODO(yaowu): The block_variance calls the unoptimized versions of variance() + * and highbd_8_variance(). It should not. + */ +static void encoder_variance(const uint8_t *a, int a_stride, const uint8_t *b, + int b_stride, int w, int h, unsigned int *sse, + int *sum) { + int i, j; + + *sum = 0; + *sse = 0; + + for (i = 0; i < h; i++) { + for (j = 0; j < w; j++) { + const int diff = a[j] - b[j]; + *sum += diff; + *sse += diff * diff; + } + + a += a_stride; + b += b_stride; + } +} + +static void encoder_highbd_variance64(const uint8_t *a8, int a_stride, + const uint8_t *b8, int b_stride, int w, + int h, uint64_t *sse, int64_t *sum) { + const uint16_t *a = CONVERT_TO_SHORTPTR(a8); + const uint16_t *b = CONVERT_TO_SHORTPTR(b8); + int64_t tsum = 0; + uint64_t tsse = 0; + for (int i = 0; i < h; ++i) { + int32_t lsum = 0; + for (int j = 0; j < w; ++j) { + const int diff = a[j] - b[j]; + lsum += diff; + tsse += (uint32_t)(diff * diff); + } + tsum += lsum; + a += a_stride; + b += b_stride; + } + *sum = tsum; + *sse = tsse; +} + +static void encoder_highbd_8_variance(const uint8_t *a8, int a_stride, + const uint8_t *b8, int b_stride, int w, + int h, unsigned int *sse, int *sum) { + uint64_t sse_long = 0; + int64_t sum_long = 0; + encoder_highbd_variance64(a8, a_stride, b8, b_stride, w, h, &sse_long, + &sum_long); + *sse = (unsigned int)sse_long; + *sum = (int)sum_long; +} + +static int64_t get_sse(const uint8_t *a, int a_stride, const uint8_t *b, + int b_stride, int width, int height) { + const int dw = width % 16; + const int dh = height % 16; + int64_t total_sse = 0; + unsigned int sse = 0; + int sum = 0; + int x, y; + + if (dw > 0) { + encoder_variance(&a[width - dw], a_stride, &b[width - dw], b_stride, dw, + height, &sse, &sum); + total_sse += sse; + } + + if (dh > 0) { + encoder_variance(&a[(height - dh) * a_stride], a_stride, + &b[(height - dh) * b_stride], b_stride, width - dw, dh, + &sse, &sum); + total_sse += sse; + } + + for (y = 0; y < height / 16; ++y) { + const uint8_t *pa = a; + const uint8_t *pb = b; + for (x = 0; x < width / 16; ++x) { + aom_mse16x16(pa, a_stride, pb, b_stride, &sse); + total_sse += sse; + + pa += 16; + pb += 16; + } + + a += 16 * a_stride; + b += 16 * b_stride; + } + + return total_sse; +} + +static int64_t highbd_get_sse_shift(const uint8_t *a8, int a_stride, + const uint8_t *b8, int b_stride, int width, + int height, unsigned int input_shift) { + const uint16_t *a = CONVERT_TO_SHORTPTR(a8); + const uint16_t *b = CONVERT_TO_SHORTPTR(b8); + int64_t total_sse = 0; + int x, y; + for (y = 0; y < height; ++y) { + for (x = 0; x < width; ++x) { + int64_t diff; + diff = (a[x] >> input_shift) - (b[x] >> input_shift); + total_sse += diff * diff; + } + a += a_stride; + b += b_stride; + } + return total_sse; +} + +static int64_t highbd_get_sse(const uint8_t *a, int a_stride, const uint8_t *b, + int b_stride, int width, int height) { + int64_t total_sse = 0; + int x, y; + const int dw = width % 16; + const int dh = height % 16; + unsigned int sse = 0; + int sum = 0; + if (dw > 0) { + encoder_highbd_8_variance(&a[width - dw], a_stride, &b[width - dw], + b_stride, dw, height, &sse, &sum); + total_sse += sse; + } + if (dh > 0) { + encoder_highbd_8_variance(&a[(height - dh) * a_stride], a_stride, + &b[(height - dh) * b_stride], b_stride, + width - dw, dh, &sse, &sum); + total_sse += sse; + } + for (y = 0; y < height / 16; ++y) { + const uint8_t *pa = a; + const uint8_t *pb = b; + for (x = 0; x < width / 16; ++x) { + aom_highbd_8_mse16x16(pa, a_stride, pb, b_stride, &sse); + total_sse += sse; + pa += 16; + pb += 16; + } + a += 16 * a_stride; + b += 16 * b_stride; + } + return total_sse; +} + +int64_t aom_get_y_sse_part(const YV12_BUFFER_CONFIG *a, + const YV12_BUFFER_CONFIG *b, int hstart, int width, + int vstart, int height) { + return get_sse(a->y_buffer + vstart * a->y_stride + hstart, a->y_stride, + b->y_buffer + vstart * b->y_stride + hstart, b->y_stride, + width, height); +} + +int64_t aom_get_y_sse(const YV12_BUFFER_CONFIG *a, + const YV12_BUFFER_CONFIG *b) { + assert(a->y_crop_width == b->y_crop_width); + assert(a->y_crop_height == b->y_crop_height); + + return get_sse(a->y_buffer, a->y_stride, b->y_buffer, b->y_stride, + a->y_crop_width, a->y_crop_height); +} + +int64_t aom_get_u_sse_part(const YV12_BUFFER_CONFIG *a, + const YV12_BUFFER_CONFIG *b, int hstart, int width, + int vstart, int height) { + return get_sse(a->u_buffer + vstart * a->uv_stride + hstart, a->uv_stride, + b->u_buffer + vstart * b->uv_stride + hstart, b->uv_stride, + width, height); +} + +int64_t aom_get_u_sse(const YV12_BUFFER_CONFIG *a, + const YV12_BUFFER_CONFIG *b) { + assert(a->uv_crop_width == b->uv_crop_width); + assert(a->uv_crop_height == b->uv_crop_height); + + return get_sse(a->u_buffer, a->uv_stride, b->u_buffer, b->uv_stride, + a->uv_crop_width, a->uv_crop_height); +} + +int64_t aom_get_v_sse_part(const YV12_BUFFER_CONFIG *a, + const YV12_BUFFER_CONFIG *b, int hstart, int width, + int vstart, int height) { + return get_sse(a->v_buffer + vstart * a->uv_stride + hstart, a->uv_stride, + b->v_buffer + vstart * b->uv_stride + hstart, b->uv_stride, + width, height); +} + +int64_t aom_get_v_sse(const YV12_BUFFER_CONFIG *a, + const YV12_BUFFER_CONFIG *b) { + assert(a->uv_crop_width == b->uv_crop_width); + assert(a->uv_crop_height == b->uv_crop_height); + + return get_sse(a->v_buffer, a->uv_stride, b->v_buffer, b->uv_stride, + a->uv_crop_width, a->uv_crop_height); +} + +int64_t aom_highbd_get_y_sse_part(const YV12_BUFFER_CONFIG *a, + const YV12_BUFFER_CONFIG *b, int hstart, + int width, int vstart, int height) { + return highbd_get_sse( + a->y_buffer + vstart * a->y_stride + hstart, a->y_stride, + b->y_buffer + vstart * b->y_stride + hstart, b->y_stride, width, height); +} + +int64_t aom_highbd_get_y_sse(const YV12_BUFFER_CONFIG *a, + const YV12_BUFFER_CONFIG *b) { + assert(a->y_crop_width == b->y_crop_width); + assert(a->y_crop_height == b->y_crop_height); + assert((a->flags & YV12_FLAG_HIGHBITDEPTH) != 0); + assert((b->flags & YV12_FLAG_HIGHBITDEPTH) != 0); + + return highbd_get_sse(a->y_buffer, a->y_stride, b->y_buffer, b->y_stride, + a->y_crop_width, a->y_crop_height); +} + +int64_t aom_highbd_get_u_sse_part(const YV12_BUFFER_CONFIG *a, + const YV12_BUFFER_CONFIG *b, int hstart, + int width, int vstart, int height) { + return highbd_get_sse(a->u_buffer + vstart * a->uv_stride + hstart, + a->uv_stride, + b->u_buffer + vstart * b->uv_stride + hstart, + b->uv_stride, width, height); +} + +int64_t aom_highbd_get_u_sse(const YV12_BUFFER_CONFIG *a, + const YV12_BUFFER_CONFIG *b) { + assert(a->uv_crop_width == b->uv_crop_width); + assert(a->uv_crop_height == b->uv_crop_height); + assert((a->flags & YV12_FLAG_HIGHBITDEPTH) != 0); + assert((b->flags & YV12_FLAG_HIGHBITDEPTH) != 0); + + return highbd_get_sse(a->u_buffer, a->uv_stride, b->u_buffer, b->uv_stride, + a->uv_crop_width, a->uv_crop_height); +} + +int64_t aom_highbd_get_v_sse_part(const YV12_BUFFER_CONFIG *a, + const YV12_BUFFER_CONFIG *b, int hstart, + int width, int vstart, int height) { + return highbd_get_sse(a->v_buffer + vstart * a->uv_stride + hstart, + a->uv_stride, + b->v_buffer + vstart * b->uv_stride + hstart, + b->uv_stride, width, height); +} + +int64_t aom_highbd_get_v_sse(const YV12_BUFFER_CONFIG *a, + const YV12_BUFFER_CONFIG *b) { + assert(a->uv_crop_width == b->uv_crop_width); + assert(a->uv_crop_height == b->uv_crop_height); + assert((a->flags & YV12_FLAG_HIGHBITDEPTH) != 0); + assert((b->flags & YV12_FLAG_HIGHBITDEPTH) != 0); + + return highbd_get_sse(a->v_buffer, a->uv_stride, b->v_buffer, b->uv_stride, + a->uv_crop_width, a->uv_crop_height); +} + +int64_t aom_get_sse_plane(const YV12_BUFFER_CONFIG *a, + const YV12_BUFFER_CONFIG *b, int plane, int highbd) { + if (highbd) { + switch (plane) { + case 0: return aom_highbd_get_y_sse(a, b); + case 1: return aom_highbd_get_u_sse(a, b); + case 2: return aom_highbd_get_v_sse(a, b); + default: assert(plane >= 0 && plane <= 2); return 0; + } + } + switch (plane) { + case 0: return aom_get_y_sse(a, b); + case 1: return aom_get_u_sse(a, b); + case 2: return aom_get_v_sse(a, b); + default: assert(plane >= 0 && plane <= 2); return 0; + } +} + +void aom_calc_highbd_psnr(const YV12_BUFFER_CONFIG *a, + const YV12_BUFFER_CONFIG *b, PSNR_STATS *psnr, + uint32_t bit_depth, uint32_t in_bit_depth) { + const int widths[3] = { a->y_crop_width, a->uv_crop_width, a->uv_crop_width }; + const int heights[3] = { a->y_crop_height, a->uv_crop_height, + a->uv_crop_height }; + const int a_strides[3] = { a->y_stride, a->uv_stride, a->uv_stride }; + const int b_strides[3] = { b->y_stride, b->uv_stride, b->uv_stride }; + int i; + uint64_t total_sse = 0; + uint32_t total_samples = 0; + const double peak = (double)((1 << in_bit_depth) - 1); + const unsigned int input_shift = bit_depth - in_bit_depth; + + for (i = 0; i < 3; ++i) { + const int w = widths[i]; + const int h = heights[i]; + const uint32_t samples = w * h; + uint64_t sse; + if (a->flags & YV12_FLAG_HIGHBITDEPTH) { + if (input_shift) { + sse = highbd_get_sse_shift(a->buffers[i], a_strides[i], b->buffers[i], + b_strides[i], w, h, input_shift); + } else { + sse = highbd_get_sse(a->buffers[i], a_strides[i], b->buffers[i], + b_strides[i], w, h); + } + } else { + sse = get_sse(a->buffers[i], a_strides[i], b->buffers[i], b_strides[i], w, + h); + } + psnr->sse[1 + i] = sse; + psnr->samples[1 + i] = samples; + psnr->psnr[1 + i] = aom_sse_to_psnr(samples, peak, (double)sse); + + total_sse += sse; + total_samples += samples; + } + + psnr->sse[0] = total_sse; + psnr->samples[0] = total_samples; + psnr->psnr[0] = + aom_sse_to_psnr((double)total_samples, peak, (double)total_sse); +} + +void aom_calc_psnr(const YV12_BUFFER_CONFIG *a, const YV12_BUFFER_CONFIG *b, + PSNR_STATS *psnr) { + static const double peak = 255.0; + const int widths[3] = { a->y_crop_width, a->uv_crop_width, a->uv_crop_width }; + const int heights[3] = { a->y_crop_height, a->uv_crop_height, + a->uv_crop_height }; + const int a_strides[3] = { a->y_stride, a->uv_stride, a->uv_stride }; + const int b_strides[3] = { b->y_stride, b->uv_stride, b->uv_stride }; + int i; + uint64_t total_sse = 0; + uint32_t total_samples = 0; + + for (i = 0; i < 3; ++i) { + const int w = widths[i]; + const int h = heights[i]; + const uint32_t samples = w * h; + const uint64_t sse = + get_sse(a->buffers[i], a_strides[i], b->buffers[i], b_strides[i], w, h); + psnr->sse[1 + i] = sse; + psnr->samples[1 + i] = samples; + psnr->psnr[1 + i] = aom_sse_to_psnr(samples, peak, (double)sse); + + total_sse += sse; + total_samples += samples; + } + + psnr->sse[0] = total_sse; + psnr->samples[0] = total_samples; + psnr->psnr[0] = + aom_sse_to_psnr((double)total_samples, peak, (double)total_sse); +} diff --git a/third_party/aom/aom_dsp/psnr.h b/third_party/aom/aom_dsp/psnr.h new file mode 100644 index 0000000000..58e4e71ee7 --- /dev/null +++ b/third_party/aom/aom_dsp/psnr.h @@ -0,0 +1,79 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_PSNR_H_ +#define AOM_AOM_DSP_PSNR_H_ + +#include "aom_scale/yv12config.h" + +#define MAX_PSNR 100.0 + +#ifdef __cplusplus +extern "C" { +#endif + +typedef struct { + double psnr[4]; // total/y/u/v + uint64_t sse[4]; // total/y/u/v + uint32_t samples[4]; // total/y/u/v +} PSNR_STATS; + +/*!\brief Converts SSE to PSNR + * + * Converts sum of squared errros (SSE) to peak signal-to-noise ratio (PNSR). + * + * \param[in] samples Number of samples + * \param[in] peak Max sample value + * \param[in] sse Sum of squared errors + */ +double aom_sse_to_psnr(double samples, double peak, double sse); +int64_t aom_get_y_sse_part(const YV12_BUFFER_CONFIG *a, + const YV12_BUFFER_CONFIG *b, int hstart, int width, + int vstart, int height); +int64_t aom_get_y_sse(const YV12_BUFFER_CONFIG *a, const YV12_BUFFER_CONFIG *b); +int64_t aom_get_u_sse_part(const YV12_BUFFER_CONFIG *a, + const YV12_BUFFER_CONFIG *b, int hstart, int width, + int vstart, int height); +int64_t aom_get_u_sse(const YV12_BUFFER_CONFIG *a, const YV12_BUFFER_CONFIG *b); +int64_t aom_get_v_sse_part(const YV12_BUFFER_CONFIG *a, + const YV12_BUFFER_CONFIG *b, int hstart, int width, + int vstart, int height); +int64_t aom_get_v_sse(const YV12_BUFFER_CONFIG *a, const YV12_BUFFER_CONFIG *b); +int64_t aom_get_sse_plane(const YV12_BUFFER_CONFIG *a, + const YV12_BUFFER_CONFIG *b, int plane, int highbd); +int64_t aom_highbd_get_y_sse_part(const YV12_BUFFER_CONFIG *a, + const YV12_BUFFER_CONFIG *b, int hstart, + int width, int vstart, int height); +int64_t aom_highbd_get_y_sse(const YV12_BUFFER_CONFIG *a, + const YV12_BUFFER_CONFIG *b); +int64_t aom_highbd_get_u_sse_part(const YV12_BUFFER_CONFIG *a, + const YV12_BUFFER_CONFIG *b, int hstart, + int width, int vstart, int height); +int64_t aom_highbd_get_u_sse(const YV12_BUFFER_CONFIG *a, + const YV12_BUFFER_CONFIG *b); +int64_t aom_highbd_get_v_sse_part(const YV12_BUFFER_CONFIG *a, + const YV12_BUFFER_CONFIG *b, int hstart, + int width, int vstart, int height); +int64_t aom_highbd_get_v_sse(const YV12_BUFFER_CONFIG *a, + const YV12_BUFFER_CONFIG *b); +void aom_calc_highbd_psnr(const YV12_BUFFER_CONFIG *a, + const YV12_BUFFER_CONFIG *b, PSNR_STATS *psnr, + unsigned int bit_depth, unsigned int in_bit_depth); +void aom_calc_psnr(const YV12_BUFFER_CONFIG *a, const YV12_BUFFER_CONFIG *b, + PSNR_STATS *psnr); + +double aom_psnrhvs(const YV12_BUFFER_CONFIG *source, + const YV12_BUFFER_CONFIG *dest, double *phvs_y, + double *phvs_u, double *phvs_v, uint32_t bd, uint32_t in_bd); +#ifdef __cplusplus +} // extern "C" +#endif +#endif // AOM_AOM_DSP_PSNR_H_ diff --git a/third_party/aom/aom_dsp/psnrhvs.c b/third_party/aom/aom_dsp/psnrhvs.c new file mode 100644 index 0000000000..30fe21d9ce --- /dev/null +++ b/third_party/aom/aom_dsp/psnrhvs.c @@ -0,0 +1,272 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + * + * This code was originally written by: Gregory Maxwell, at the Daala + * project. + */ + +#include +#include +#include +#include + +#include "config/aom_config.h" +#include "config/aom_dsp_rtcd.h" + +#include "aom_dsp/psnr.h" +#include "aom_dsp/ssim.h" +#include "aom_ports/system_state.h" + +static void od_bin_fdct8x8(tran_low_t *y, int ystride, const int16_t *x, + int xstride) { + int i, j; + (void)xstride; + aom_fdct8x8(x, y, ystride); + for (i = 0; i < 8; i++) + for (j = 0; j < 8; j++) + *(y + ystride * i + j) = (*(y + ystride * i + j) + 4) >> 3; +} + +static void hbd_od_bin_fdct8x8(tran_low_t *y, int ystride, const int16_t *x, + int xstride) { + int i, j; + (void)xstride; + aom_highbd_fdct8x8(x, y, ystride); + for (i = 0; i < 8; i++) + for (j = 0; j < 8; j++) + *(y + ystride * i + j) = (*(y + ystride * i + j) + 4) >> 3; +} + +/* Normalized inverse quantization matrix for 8x8 DCT at the point of + * transparency. This is not the JPEG based matrix from the paper, + this one gives a slightly higher MOS agreement.*/ +static const double csf_y[8][8] = { + { 1.6193873005, 2.2901594831, 2.08509755623, 1.48366094411, 1.00227514334, + 0.678296995242, 0.466224900598, 0.3265091542 }, + { 2.2901594831, 1.94321815382, 2.04793073064, 1.68731108984, 1.2305666963, + 0.868920337363, 0.61280991668, 0.436405793551 }, + { 2.08509755623, 2.04793073064, 1.34329019223, 1.09205635862, 0.875748795257, + 0.670882927016, 0.501731932449, 0.372504254596 }, + { 1.48366094411, 1.68731108984, 1.09205635862, 0.772819797575, 0.605636379554, + 0.48309405692, 0.380429446972, 0.295774038565 }, + { 1.00227514334, 1.2305666963, 0.875748795257, 0.605636379554, 0.448996256676, + 0.352889268808, 0.283006984131, 0.226951348204 }, + { 0.678296995242, 0.868920337363, 0.670882927016, 0.48309405692, + 0.352889268808, 0.27032073436, 0.215017739696, 0.17408067321 }, + { 0.466224900598, 0.61280991668, 0.501731932449, 0.380429446972, + 0.283006984131, 0.215017739696, 0.168869545842, 0.136153931001 }, + { 0.3265091542, 0.436405793551, 0.372504254596, 0.295774038565, + 0.226951348204, 0.17408067321, 0.136153931001, 0.109083846276 } +}; +static const double csf_cb420[8][8] = { + { 1.91113096927, 2.46074210438, 1.18284184739, 1.14982565193, 1.05017074788, + 0.898018824055, 0.74725392039, 0.615105596242 }, + { 2.46074210438, 1.58529308355, 1.21363250036, 1.38190029285, 1.33100189972, + 1.17428548929, 0.996404342439, 0.830890433625 }, + { 1.18284184739, 1.21363250036, 0.978712413627, 1.02624506078, 1.03145147362, + 0.960060382087, 0.849823426169, 0.731221236837 }, + { 1.14982565193, 1.38190029285, 1.02624506078, 0.861317501629, 0.801821139099, + 0.751437590932, 0.685398513368, 0.608694761374 }, + { 1.05017074788, 1.33100189972, 1.03145147362, 0.801821139099, 0.676555426187, + 0.605503172737, 0.55002013668, 0.495804539034 }, + { 0.898018824055, 1.17428548929, 0.960060382087, 0.751437590932, + 0.605503172737, 0.514674450957, 0.454353482512, 0.407050308965 }, + { 0.74725392039, 0.996404342439, 0.849823426169, 0.685398513368, + 0.55002013668, 0.454353482512, 0.389234902883, 0.342353999733 }, + { 0.615105596242, 0.830890433625, 0.731221236837, 0.608694761374, + 0.495804539034, 0.407050308965, 0.342353999733, 0.295530605237 } +}; +static const double csf_cr420[8][8] = { + { 2.03871978502, 2.62502345193, 1.26180942886, 1.11019789803, 1.01397751469, + 0.867069376285, 0.721500455585, 0.593906509971 }, + { 2.62502345193, 1.69112867013, 1.17180569821, 1.3342742857, 1.28513006198, + 1.13381474809, 0.962064122248, 0.802254508198 }, + { 1.26180942886, 1.17180569821, 0.944981930573, 0.990876405848, + 0.995903384143, 0.926972725286, 0.820534991409, 0.706020324706 }, + { 1.11019789803, 1.3342742857, 0.990876405848, 0.831632933426, 0.77418706195, + 0.725539939514, 0.661776842059, 0.587716619023 }, + { 1.01397751469, 1.28513006198, 0.995903384143, 0.77418706195, 0.653238524286, + 0.584635025748, 0.531064164893, 0.478717061273 }, + { 0.867069376285, 1.13381474809, 0.926972725286, 0.725539939514, + 0.584635025748, 0.496936637883, 0.438694579826, 0.393021669543 }, + { 0.721500455585, 0.962064122248, 0.820534991409, 0.661776842059, + 0.531064164893, 0.438694579826, 0.375820256136, 0.330555063063 }, + { 0.593906509971, 0.802254508198, 0.706020324706, 0.587716619023, + 0.478717061273, 0.393021669543, 0.330555063063, 0.285345396658 } +}; + +static double convert_score_db(double _score, double _weight, int bit_depth) { + int16_t pix_max = 255; + assert(_score * _weight >= 0.0); + if (bit_depth == 10) + pix_max = 1023; + else if (bit_depth == 12) + pix_max = 4095; + + if (_weight * _score < pix_max * pix_max * 1e-10) return MAX_PSNR; + return 10 * (log10(pix_max * pix_max) - log10(_weight * _score)); +} + +static double calc_psnrhvs(const unsigned char *src, int _systride, + const unsigned char *dst, int _dystride, double _par, + int _w, int _h, int _step, const double _csf[8][8], + uint32_t _shift, int buf_is_hbd) { + double ret; + const uint8_t *_src8 = src; + const uint8_t *_dst8 = dst; + const uint16_t *_src16 = CONVERT_TO_SHORTPTR(src); + const uint16_t *_dst16 = CONVERT_TO_SHORTPTR(dst); + DECLARE_ALIGNED(16, int16_t, dct_s[8 * 8]); + DECLARE_ALIGNED(16, int16_t, dct_d[8 * 8]); + DECLARE_ALIGNED(16, tran_low_t, dct_s_coef[8 * 8]); + DECLARE_ALIGNED(16, tran_low_t, dct_d_coef[8 * 8]); + double mask[8][8]; + int pixels; + int x; + int y; + (void)_par; + ret = pixels = 0; + /*In the PSNR-HVS-M paper[1] the authors describe the construction of + their masking table as "we have used the quantization table for the + color component Y of JPEG [6] that has been also obtained on the + basis of CSF. Note that the values in quantization table JPEG have + been normalized and then squared." Their CSF matrix (from PSNR-HVS) + was also constructed from the JPEG matrices. I can not find any obvious + scheme of normalizing to produce their table, but if I multiply their + CSF by 0.38857 and square the result I get their masking table. + I have no idea where this constant comes from, but deviating from it + too greatly hurts MOS agreement. + + [1] Nikolay Ponomarenko, Flavia Silvestri, Karen Egiazarian, Marco Carli, + Jaakko Astola, Vladimir Lukin, "On between-coefficient contrast masking + of DCT basis functions", CD-ROM Proceedings of the Third + International Workshop on Video Processing and Quality Metrics for Consumer + Electronics VPQM-07, Scottsdale, Arizona, USA, 25-26 January, 2007, 4 p.*/ + for (x = 0; x < 8; x++) + for (y = 0; y < 8; y++) + mask[x][y] = + (_csf[x][y] * 0.3885746225901003) * (_csf[x][y] * 0.3885746225901003); + for (y = 0; y < _h - 7; y += _step) { + for (x = 0; x < _w - 7; x += _step) { + int i; + int j; + double s_means[4]; + double d_means[4]; + double s_vars[4]; + double d_vars[4]; + double s_gmean = 0; + double d_gmean = 0; + double s_gvar = 0; + double d_gvar = 0; + double s_mask = 0; + double d_mask = 0; + for (i = 0; i < 4; i++) + s_means[i] = d_means[i] = s_vars[i] = d_vars[i] = 0; + for (i = 0; i < 8; i++) { + for (j = 0; j < 8; j++) { + int sub = ((i & 12) >> 2) + ((j & 12) >> 1); + if (!buf_is_hbd) { + dct_s[i * 8 + j] = _src8[(y + i) * _systride + (j + x)]; + dct_d[i * 8 + j] = _dst8[(y + i) * _dystride + (j + x)]; + } else { + dct_s[i * 8 + j] = _src16[(y + i) * _systride + (j + x)] >> _shift; + dct_d[i * 8 + j] = _dst16[(y + i) * _dystride + (j + x)] >> _shift; + } + s_gmean += dct_s[i * 8 + j]; + d_gmean += dct_d[i * 8 + j]; + s_means[sub] += dct_s[i * 8 + j]; + d_means[sub] += dct_d[i * 8 + j]; + } + } + s_gmean /= 64.f; + d_gmean /= 64.f; + for (i = 0; i < 4; i++) s_means[i] /= 16.f; + for (i = 0; i < 4; i++) d_means[i] /= 16.f; + for (i = 0; i < 8; i++) { + for (j = 0; j < 8; j++) { + int sub = ((i & 12) >> 2) + ((j & 12) >> 1); + s_gvar += (dct_s[i * 8 + j] - s_gmean) * (dct_s[i * 8 + j] - s_gmean); + d_gvar += (dct_d[i * 8 + j] - d_gmean) * (dct_d[i * 8 + j] - d_gmean); + s_vars[sub] += (dct_s[i * 8 + j] - s_means[sub]) * + (dct_s[i * 8 + j] - s_means[sub]); + d_vars[sub] += (dct_d[i * 8 + j] - d_means[sub]) * + (dct_d[i * 8 + j] - d_means[sub]); + } + } + s_gvar *= 1 / 63.f * 64; + d_gvar *= 1 / 63.f * 64; + for (i = 0; i < 4; i++) s_vars[i] *= 1 / 15.f * 16; + for (i = 0; i < 4; i++) d_vars[i] *= 1 / 15.f * 16; + if (s_gvar > 0) + s_gvar = (s_vars[0] + s_vars[1] + s_vars[2] + s_vars[3]) / s_gvar; + if (d_gvar > 0) + d_gvar = (d_vars[0] + d_vars[1] + d_vars[2] + d_vars[3]) / d_gvar; + if (!buf_is_hbd) { + od_bin_fdct8x8(dct_s_coef, 8, dct_s, 8); + od_bin_fdct8x8(dct_d_coef, 8, dct_d, 8); + } else { + hbd_od_bin_fdct8x8(dct_s_coef, 8, dct_s, 8); + hbd_od_bin_fdct8x8(dct_d_coef, 8, dct_d, 8); + } + for (i = 0; i < 8; i++) + for (j = (i == 0); j < 8; j++) + s_mask += dct_s_coef[i * 8 + j] * dct_s_coef[i * 8 + j] * mask[i][j]; + for (i = 0; i < 8; i++) + for (j = (i == 0); j < 8; j++) + d_mask += dct_d_coef[i * 8 + j] * dct_d_coef[i * 8 + j] * mask[i][j]; + s_mask = sqrt(s_mask * s_gvar) / 32.f; + d_mask = sqrt(d_mask * d_gvar) / 32.f; + if (d_mask > s_mask) s_mask = d_mask; + for (i = 0; i < 8; i++) { + for (j = 0; j < 8; j++) { + double err; + err = fabs((double)(dct_s_coef[i * 8 + j] - dct_d_coef[i * 8 + j])); + if (i != 0 || j != 0) + err = err < s_mask / mask[i][j] ? 0 : err - s_mask / mask[i][j]; + ret += (err * _csf[i][j]) * (err * _csf[i][j]); + pixels++; + } + } + } + } + if (pixels <= 0) return 0; + ret /= pixels; + return ret; +} + +double aom_psnrhvs(const YV12_BUFFER_CONFIG *src, const YV12_BUFFER_CONFIG *dst, + double *y_psnrhvs, double *u_psnrhvs, double *v_psnrhvs, + uint32_t bd, uint32_t in_bd) { + double psnrhvs; + const double par = 1.0; + const int step = 7; + uint32_t bd_shift = 0; + aom_clear_system_state(); + assert(bd == 8 || bd == 10 || bd == 12); + assert(bd >= in_bd); + assert(src->flags == dst->flags); + const int buf_is_hbd = src->flags & YV12_FLAG_HIGHBITDEPTH; + + bd_shift = bd - in_bd; + + *y_psnrhvs = calc_psnrhvs( + src->y_buffer, src->y_stride, dst->y_buffer, dst->y_stride, par, + src->y_crop_width, src->y_crop_height, step, csf_y, bd_shift, buf_is_hbd); + *u_psnrhvs = + calc_psnrhvs(src->u_buffer, src->uv_stride, dst->u_buffer, dst->uv_stride, + par, src->uv_crop_width, src->uv_crop_height, step, + csf_cb420, bd_shift, buf_is_hbd); + *v_psnrhvs = + calc_psnrhvs(src->v_buffer, src->uv_stride, dst->v_buffer, dst->uv_stride, + par, src->uv_crop_width, src->uv_crop_height, step, + csf_cr420, bd_shift, buf_is_hbd); + psnrhvs = (*y_psnrhvs) * .8 + .1 * ((*u_psnrhvs) + (*v_psnrhvs)); + return convert_score_db(psnrhvs, 1.0, in_bd); +} diff --git a/third_party/aom/aom_dsp/quantize.c b/third_party/aom/aom_dsp/quantize.c new file mode 100644 index 0000000000..62dbd86a92 --- /dev/null +++ b/third_party/aom/aom_dsp/quantize.c @@ -0,0 +1,206 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include "aom_dsp/quantize.h" +#include "aom_mem/aom_mem.h" + +void quantize_b_helper_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, + const int16_t *zbin_ptr, const int16_t *round_ptr, + const int16_t *quant_ptr, + const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, + tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, + uint16_t *eob_ptr, const int16_t *scan, + const int16_t *iscan, const qm_val_t *qm_ptr, + const qm_val_t *iqm_ptr, const int log_scale) { + const int zbins[2] = { ROUND_POWER_OF_TWO(zbin_ptr[0], log_scale), + ROUND_POWER_OF_TWO(zbin_ptr[1], log_scale) }; + const int nzbins[2] = { zbins[0] * -1, zbins[1] * -1 }; + int i, non_zero_count = (int)n_coeffs, eob = -1; + (void)iscan; + + memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr)); + memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr)); + + // Pre-scan pass + for (i = (int)n_coeffs - 1; i >= 0; i--) { + const int rc = scan[i]; + const qm_val_t wt = qm_ptr != NULL ? qm_ptr[rc] : (1 << AOM_QM_BITS); + const int coeff = coeff_ptr[rc] * wt; + + if (coeff < (zbins[rc != 0] * (1 << AOM_QM_BITS)) && + coeff > (nzbins[rc != 0] * (1 << AOM_QM_BITS))) + non_zero_count--; + else + break; + } + + // Quantization pass: All coefficients with index >= zero_flag are + // skippable. Note: zero_flag can be zero. + for (i = 0; i < non_zero_count; i++) { + const int rc = scan[i]; + const int coeff = coeff_ptr[rc]; + const int coeff_sign = (coeff >> 31); + const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign; + int tmp32; + + const qm_val_t wt = qm_ptr != NULL ? qm_ptr[rc] : (1 << AOM_QM_BITS); + if (abs_coeff * wt >= (zbins[rc != 0] << AOM_QM_BITS)) { + int64_t tmp = + clamp(abs_coeff + ROUND_POWER_OF_TWO(round_ptr[rc != 0], log_scale), + INT16_MIN, INT16_MAX); + tmp *= wt; + tmp32 = (int)(((((tmp * quant_ptr[rc != 0]) >> 16) + tmp) * + quant_shift_ptr[rc != 0]) >> + (16 - log_scale + AOM_QM_BITS)); // quantization + qcoeff_ptr[rc] = (tmp32 ^ coeff_sign) - coeff_sign; + const int iwt = iqm_ptr != NULL ? iqm_ptr[rc] : (1 << AOM_QM_BITS); + const int dequant = + (dequant_ptr[rc != 0] * iwt + (1 << (AOM_QM_BITS - 1))) >> + AOM_QM_BITS; + const tran_low_t abs_dqcoeff = (tmp32 * dequant) >> log_scale; + dqcoeff_ptr[rc] = (tran_low_t)((abs_dqcoeff ^ coeff_sign) - coeff_sign); + + if (tmp32) eob = i; + } + } + *eob_ptr = eob + 1; +} + +void highbd_quantize_b_helper_c( + const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, + const int16_t *round_ptr, const int16_t *quant_ptr, + const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, + tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, + const int16_t *scan, const int16_t *iscan, const qm_val_t *qm_ptr, + const qm_val_t *iqm_ptr, const int log_scale) { + int i, eob = -1; + const int zbins[2] = { ROUND_POWER_OF_TWO(zbin_ptr[0], log_scale), + ROUND_POWER_OF_TWO(zbin_ptr[1], log_scale) }; + const int nzbins[2] = { zbins[0] * -1, zbins[1] * -1 }; + int dequant; + int idx_arr[4096]; + (void)iscan; + int idx = 0; + + memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr)); + memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr)); + + // Pre-scan pass + for (i = 0; i < n_coeffs; i++) { + const int rc = scan[i]; + const qm_val_t wt = qm_ptr != NULL ? qm_ptr[rc] : (1 << AOM_QM_BITS); + const int coeff = coeff_ptr[rc] * wt; + + // If the coefficient is out of the base ZBIN range, keep it for + // quantization. + if (coeff >= (zbins[rc != 0] * (1 << AOM_QM_BITS)) || + coeff <= (nzbins[rc != 0] * (1 << AOM_QM_BITS))) + idx_arr[idx++] = i; + } + + // Quantization pass: only process the coefficients selected in + // pre-scan pass. Note: idx can be zero. + for (i = 0; i < idx; i++) { + const int rc = scan[idx_arr[i]]; + const int coeff = coeff_ptr[rc]; + const int coeff_sign = (coeff >> 31); + const qm_val_t wt = qm_ptr != NULL ? qm_ptr[rc] : (1 << AOM_QM_BITS); + const qm_val_t iwt = iqm_ptr != NULL ? iqm_ptr[rc] : (1 << AOM_QM_BITS); + const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign; + const int64_t tmp1 = + abs_coeff + ROUND_POWER_OF_TWO(round_ptr[rc != 0], log_scale); + const int64_t tmpw = tmp1 * wt; + const int64_t tmp2 = ((tmpw * quant_ptr[rc != 0]) >> 16) + tmpw; + const int abs_qcoeff = (int)((tmp2 * quant_shift_ptr[rc != 0]) >> + (16 - log_scale + AOM_QM_BITS)); + qcoeff_ptr[rc] = (tran_low_t)((abs_qcoeff ^ coeff_sign) - coeff_sign); + dequant = + (dequant_ptr[rc != 0] * iwt + (1 << (AOM_QM_BITS - 1))) >> AOM_QM_BITS; + const tran_low_t abs_dqcoeff = (abs_qcoeff * dequant) >> log_scale; + dqcoeff_ptr[rc] = (tran_low_t)((abs_dqcoeff ^ coeff_sign) - coeff_sign); + if (abs_qcoeff) eob = idx_arr[i]; + } + *eob_ptr = eob + 1; +} + +/* These functions should only be called when quantisation matrices + are not used. */ +void aom_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, + const int16_t *zbin_ptr, const int16_t *round_ptr, + const int16_t *quant_ptr, const int16_t *quant_shift_ptr, + tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, + const int16_t *dequant_ptr, uint16_t *eob_ptr, + const int16_t *scan, const int16_t *iscan) { + quantize_b_helper_c(coeff_ptr, n_coeffs, zbin_ptr, round_ptr, quant_ptr, + quant_shift_ptr, qcoeff_ptr, dqcoeff_ptr, dequant_ptr, + eob_ptr, scan, iscan, NULL, NULL, 0); +} + +void aom_quantize_b_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, + const int16_t *zbin_ptr, const int16_t *round_ptr, + const int16_t *quant_ptr, + const int16_t *quant_shift_ptr, + tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, + const int16_t *dequant_ptr, uint16_t *eob_ptr, + const int16_t *scan, const int16_t *iscan) { + quantize_b_helper_c(coeff_ptr, n_coeffs, zbin_ptr, round_ptr, quant_ptr, + quant_shift_ptr, qcoeff_ptr, dqcoeff_ptr, dequant_ptr, + eob_ptr, scan, iscan, NULL, NULL, 1); +} + +void aom_quantize_b_64x64_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, + const int16_t *zbin_ptr, const int16_t *round_ptr, + const int16_t *quant_ptr, + const int16_t *quant_shift_ptr, + tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, + const int16_t *dequant_ptr, uint16_t *eob_ptr, + const int16_t *scan, const int16_t *iscan) { + quantize_b_helper_c(coeff_ptr, n_coeffs, zbin_ptr, round_ptr, quant_ptr, + quant_shift_ptr, qcoeff_ptr, dqcoeff_ptr, dequant_ptr, + eob_ptr, scan, iscan, NULL, NULL, 2); +} + +void aom_highbd_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, + const int16_t *zbin_ptr, const int16_t *round_ptr, + const int16_t *quant_ptr, + const int16_t *quant_shift_ptr, + tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, + const int16_t *dequant_ptr, uint16_t *eob_ptr, + const int16_t *scan, const int16_t *iscan) { + highbd_quantize_b_helper_c(coeff_ptr, n_coeffs, zbin_ptr, round_ptr, + quant_ptr, quant_shift_ptr, qcoeff_ptr, + dqcoeff_ptr, dequant_ptr, eob_ptr, scan, iscan, + NULL, NULL, 0); +} + +void aom_highbd_quantize_b_32x32_c( + const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, + const int16_t *round_ptr, const int16_t *quant_ptr, + const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, + tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, + const int16_t *scan, const int16_t *iscan) { + highbd_quantize_b_helper_c(coeff_ptr, n_coeffs, zbin_ptr, round_ptr, + quant_ptr, quant_shift_ptr, qcoeff_ptr, + dqcoeff_ptr, dequant_ptr, eob_ptr, scan, iscan, + NULL, NULL, 1); +} + +void aom_highbd_quantize_b_64x64_c( + const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, + const int16_t *round_ptr, const int16_t *quant_ptr, + const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, + tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, + const int16_t *scan, const int16_t *iscan) { + highbd_quantize_b_helper_c(coeff_ptr, n_coeffs, zbin_ptr, round_ptr, + quant_ptr, quant_shift_ptr, qcoeff_ptr, + dqcoeff_ptr, dequant_ptr, eob_ptr, scan, iscan, + NULL, NULL, 2); +} diff --git a/third_party/aom/aom_dsp/quantize.h b/third_party/aom/aom_dsp/quantize.h new file mode 100644 index 0000000000..c55ab234e7 --- /dev/null +++ b/third_party/aom/aom_dsp/quantize.h @@ -0,0 +1,59 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_QUANTIZE_H_ +#define AOM_AOM_DSP_QUANTIZE_H_ + +#include "config/aom_config.h" + +#include "aom_dsp/aom_dsp_common.h" + +#ifdef __cplusplus +extern "C" { +#endif + +void quantize_b_helper_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, + const int16_t *zbin_ptr, const int16_t *round_ptr, + const int16_t *quant_ptr, + const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, + tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, + uint16_t *eob_ptr, const int16_t *scan, + const int16_t *iscan, const qm_val_t *qm_ptr, + const qm_val_t *iqm_ptr, const int log_scale); + +void aom_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, + const int16_t *zbin_ptr, const int16_t *round_ptr, + const int16_t *quant_ptr, const int16_t *quant_shift_ptr, + tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, + const int16_t *dequant_ptr, uint16_t *eob_ptr, + const int16_t *scan, const int16_t *iscan); + +void highbd_quantize_b_helper_c( + const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, + const int16_t *round_ptr, const int16_t *quant_ptr, + const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, + tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, + const int16_t *scan, const int16_t *iscan, const qm_val_t *qm_ptr, + const qm_val_t *iqm_ptr, const int log_scale); + +void aom_highbd_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, + const int16_t *zbin_ptr, const int16_t *round_ptr, + const int16_t *quant_ptr, + const int16_t *quant_shift_ptr, + tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, + const int16_t *dequant_ptr, uint16_t *eob_ptr, + const int16_t *scan, const int16_t *iscan); + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif // AOM_AOM_DSP_QUANTIZE_H_ diff --git a/third_party/aom/aom_dsp/sad.c b/third_party/aom/aom_dsp/sad.c new file mode 100644 index 0000000000..1e24df4a56 --- /dev/null +++ b/third_party/aom/aom_dsp/sad.c @@ -0,0 +1,304 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include + +#include "config/aom_config.h" +#include "config/aom_dsp_rtcd.h" + +#include "aom/aom_integer.h" +#include "aom_ports/mem.h" +#include "aom_dsp/blend.h" + +/* Sum the difference between every corresponding element of the buffers. */ +static INLINE unsigned int sad(const uint8_t *a, int a_stride, const uint8_t *b, + int b_stride, int width, int height) { + int y, x; + unsigned int sad = 0; + + for (y = 0; y < height; y++) { + for (x = 0; x < width; x++) sad += abs(a[x] - b[x]); + + a += a_stride; + b += b_stride; + } + return sad; +} + +#define sadMxh(m) \ + unsigned int aom_sad##m##xh_c(const uint8_t *a, int a_stride, \ + const uint8_t *b, int b_stride, int width, \ + int height) { \ + return sad(a, a_stride, b, b_stride, width, height); \ + } + +#define sadMxN(m, n) \ + unsigned int aom_sad##m##x##n##_c(const uint8_t *src, int src_stride, \ + const uint8_t *ref, int ref_stride) { \ + return sad(src, src_stride, ref, ref_stride, m, n); \ + } \ + unsigned int aom_sad##m##x##n##_avg_c(const uint8_t *src, int src_stride, \ + const uint8_t *ref, int ref_stride, \ + const uint8_t *second_pred) { \ + uint8_t comp_pred[m * n]; \ + aom_comp_avg_pred(comp_pred, second_pred, m, n, ref, ref_stride); \ + return sad(src, src_stride, comp_pred, m, m, n); \ + } \ + unsigned int aom_jnt_sad##m##x##n##_avg_c( \ + const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, \ + const uint8_t *second_pred, const JNT_COMP_PARAMS *jcp_param) { \ + uint8_t comp_pred[m * n]; \ + aom_jnt_comp_avg_pred_c(comp_pred, second_pred, m, n, ref, ref_stride, \ + jcp_param); \ + return sad(src, src_stride, comp_pred, m, m, n); \ + } + +// Calculate sad against 4 reference locations and store each in sad_array +#define sadMxNx4D(m, n) \ + void aom_sad##m##x##n##x4d_c(const uint8_t *src, int src_stride, \ + const uint8_t *const ref_array[], \ + int ref_stride, uint32_t *sad_array) { \ + int i; \ + for (i = 0; i < 4; ++i) \ + sad_array[i] = \ + aom_sad##m##x##n##_c(src, src_stride, ref_array[i], ref_stride); \ + } + +/* clang-format off */ +// 128x128 +sadMxN(128, 128) +sadMxNx4D(128, 128) + +// 128x64 +sadMxN(128, 64) +sadMxNx4D(128, 64) + +// 64x128 +sadMxN(64, 128) +sadMxNx4D(64, 128) + +// 64x64 +sadMxN(64, 64) +sadMxNx4D(64, 64) + +// 64x32 +sadMxN(64, 32) +sadMxNx4D(64, 32) + +// 32x64 +sadMxN(32, 64) +sadMxNx4D(32, 64) + +// 32x32 +sadMxN(32, 32) +sadMxNx4D(32, 32) + +// 32x16 +sadMxN(32, 16) +sadMxNx4D(32, 16) + +// 16x32 +sadMxN(16, 32) +sadMxNx4D(16, 32) + +// 16x16 +sadMxN(16, 16) +sadMxNx4D(16, 16) + +// 16x8 +sadMxN(16, 8) +sadMxNx4D(16, 8) + +// 8x16 +sadMxN(8, 16) +sadMxNx4D(8, 16) + +// 8x8 +sadMxN(8, 8) +sadMxNx4D(8, 8) + +// 8x4 +sadMxN(8, 4) +sadMxNx4D(8, 4) + +// 4x8 +sadMxN(4, 8) +sadMxNx4D(4, 8) + +// 4x4 +sadMxN(4, 4) +sadMxNx4D(4, 4) + +sadMxh(128); +sadMxh(64); +sadMxh(32); +sadMxh(16); +sadMxh(8); +sadMxh(4); + +sadMxN(4, 16) +sadMxNx4D(4, 16) +sadMxN(16, 4) +sadMxNx4D(16, 4) +sadMxN(8, 32) +sadMxNx4D(8, 32) +sadMxN(32, 8) +sadMxNx4D(32, 8) +sadMxN(16, 64) +sadMxNx4D(16, 64) +sadMxN(64, 16) +sadMxNx4D(64, 16) + + /* clang-format on */ + + static INLINE + unsigned int highbd_sad(const uint8_t *a8, int a_stride, const uint8_t *b8, + int b_stride, int width, int height) { + int y, x; + unsigned int sad = 0; + const uint16_t *a = CONVERT_TO_SHORTPTR(a8); + const uint16_t *b = CONVERT_TO_SHORTPTR(b8); + for (y = 0; y < height; y++) { + for (x = 0; x < width; x++) sad += abs(a[x] - b[x]); + + a += a_stride; + b += b_stride; + } + return sad; +} + +static INLINE unsigned int highbd_sadb(const uint8_t *a8, int a_stride, + const uint16_t *b, int b_stride, + int width, int height) { + int y, x; + unsigned int sad = 0; + const uint16_t *a = CONVERT_TO_SHORTPTR(a8); + for (y = 0; y < height; y++) { + for (x = 0; x < width; x++) sad += abs(a[x] - b[x]); + + a += a_stride; + b += b_stride; + } + return sad; +} + +#define highbd_sadMxN(m, n) \ + unsigned int aom_highbd_sad##m##x##n##_c(const uint8_t *src, int src_stride, \ + const uint8_t *ref, \ + int ref_stride) { \ + return highbd_sad(src, src_stride, ref, ref_stride, m, n); \ + } \ + unsigned int aom_highbd_sad##m##x##n##_avg_c( \ + const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, \ + const uint8_t *second_pred) { \ + uint16_t comp_pred[m * n]; \ + aom_highbd_comp_avg_pred(CONVERT_TO_BYTEPTR(comp_pred), second_pred, m, n, \ + ref, ref_stride); \ + return highbd_sadb(src, src_stride, comp_pred, m, m, n); \ + } \ + unsigned int aom_highbd_jnt_sad##m##x##n##_avg_c( \ + const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, \ + const uint8_t *second_pred, const JNT_COMP_PARAMS *jcp_param) { \ + uint16_t comp_pred[m * n]; \ + aom_highbd_jnt_comp_avg_pred(CONVERT_TO_BYTEPTR(comp_pred), second_pred, \ + m, n, ref, ref_stride, jcp_param); \ + return highbd_sadb(src, src_stride, comp_pred, m, m, n); \ + } + +#define highbd_sadMxNx4D(m, n) \ + void aom_highbd_sad##m##x##n##x4d_c(const uint8_t *src, int src_stride, \ + const uint8_t *const ref_array[], \ + int ref_stride, uint32_t *sad_array) { \ + int i; \ + for (i = 0; i < 4; ++i) { \ + sad_array[i] = aom_highbd_sad##m##x##n##_c(src, src_stride, \ + ref_array[i], ref_stride); \ + } \ + } + +/* clang-format off */ +// 128x128 +highbd_sadMxN(128, 128) +highbd_sadMxNx4D(128, 128) + +// 128x64 +highbd_sadMxN(128, 64) +highbd_sadMxNx4D(128, 64) + +// 64x128 +highbd_sadMxN(64, 128) +highbd_sadMxNx4D(64, 128) + +// 64x64 +highbd_sadMxN(64, 64) +highbd_sadMxNx4D(64, 64) + +// 64x32 +highbd_sadMxN(64, 32) +highbd_sadMxNx4D(64, 32) + +// 32x64 +highbd_sadMxN(32, 64) +highbd_sadMxNx4D(32, 64) + +// 32x32 +highbd_sadMxN(32, 32) +highbd_sadMxNx4D(32, 32) + +// 32x16 +highbd_sadMxN(32, 16) +highbd_sadMxNx4D(32, 16) + +// 16x32 +highbd_sadMxN(16, 32) +highbd_sadMxNx4D(16, 32) + +// 16x16 +highbd_sadMxN(16, 16) +highbd_sadMxNx4D(16, 16) + +// 16x8 +highbd_sadMxN(16, 8) +highbd_sadMxNx4D(16, 8) + +// 8x16 +highbd_sadMxN(8, 16) +highbd_sadMxNx4D(8, 16) + +// 8x8 +highbd_sadMxN(8, 8) +highbd_sadMxNx4D(8, 8) + +// 8x4 +highbd_sadMxN(8, 4) +highbd_sadMxNx4D(8, 4) + +// 4x8 +highbd_sadMxN(4, 8) +highbd_sadMxNx4D(4, 8) + +// 4x4 +highbd_sadMxN(4, 4) +highbd_sadMxNx4D(4, 4) + +highbd_sadMxN(4, 16) +highbd_sadMxNx4D(4, 16) +highbd_sadMxN(16, 4) +highbd_sadMxNx4D(16, 4) +highbd_sadMxN(8, 32) +highbd_sadMxNx4D(8, 32) +highbd_sadMxN(32, 8) +highbd_sadMxNx4D(32, 8) +highbd_sadMxN(16, 64) +highbd_sadMxNx4D(16, 64) +highbd_sadMxN(64, 16) +highbd_sadMxNx4D(64, 16) + /* clang-format on */ diff --git a/third_party/aom/aom_dsp/sad_av1.c b/third_party/aom/aom_dsp/sad_av1.c new file mode 100644 index 0000000000..c176001d64 --- /dev/null +++ b/third_party/aom/aom_dsp/sad_av1.c @@ -0,0 +1,248 @@ +/* + * Copyright (c) 2017, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include + +#include "config/aom_config.h" +#include "config/aom_dsp_rtcd.h" + +#include "aom/aom_integer.h" +#include "aom_ports/mem.h" +#include "aom_dsp/blend.h" + +static INLINE unsigned int masked_sad(const uint8_t *src, int src_stride, + const uint8_t *a, int a_stride, + const uint8_t *b, int b_stride, + const uint8_t *m, int m_stride, int width, + int height) { + int y, x; + unsigned int sad = 0; + for (y = 0; y < height; y++) { + for (x = 0; x < width; x++) { + const int16_t pred = AOM_BLEND_A64(m[x], a[x], b[x]); + sad += abs(pred - src[x]); + } + src += src_stride; + a += a_stride; + b += b_stride; + m += m_stride; + } + sad = (sad + 31) >> 6; + return sad; +} + +#define MASKSADMxN(m, n) \ + unsigned int aom_masked_sad##m##x##n##_c( \ + const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, \ + const uint8_t *second_pred, const uint8_t *msk, int msk_stride, \ + int invert_mask) { \ + if (!invert_mask) \ + return masked_sad(src, src_stride, ref, ref_stride, second_pred, m, msk, \ + msk_stride, m, n); \ + else \ + return masked_sad(src, src_stride, second_pred, m, ref, ref_stride, msk, \ + msk_stride, m, n); \ + } + +/* clang-format off */ +MASKSADMxN(128, 128) +MASKSADMxN(128, 64) +MASKSADMxN(64, 128) +MASKSADMxN(64, 64) +MASKSADMxN(64, 32) +MASKSADMxN(32, 64) +MASKSADMxN(32, 32) +MASKSADMxN(32, 16) +MASKSADMxN(16, 32) +MASKSADMxN(16, 16) +MASKSADMxN(16, 8) +MASKSADMxN(8, 16) +MASKSADMxN(8, 8) +MASKSADMxN(8, 4) +MASKSADMxN(4, 8) +MASKSADMxN(4, 4) +MASKSADMxN(4, 16) +MASKSADMxN(16, 4) +MASKSADMxN(8, 32) +MASKSADMxN(32, 8) +MASKSADMxN(16, 64) +MASKSADMxN(64, 16) + + /* clang-format on */ + + static INLINE + unsigned int highbd_masked_sad(const uint8_t *src8, int src_stride, + const uint8_t *a8, int a_stride, + const uint8_t *b8, int b_stride, + const uint8_t *m, int m_stride, int width, + int height) { + int y, x; + unsigned int sad = 0; + const uint16_t *src = CONVERT_TO_SHORTPTR(src8); + const uint16_t *a = CONVERT_TO_SHORTPTR(a8); + const uint16_t *b = CONVERT_TO_SHORTPTR(b8); + + for (y = 0; y < height; y++) { + for (x = 0; x < width; x++) { + const uint16_t pred = AOM_BLEND_A64(m[x], a[x], b[x]); + sad += abs(pred - src[x]); + } + + src += src_stride; + a += a_stride; + b += b_stride; + m += m_stride; + } + sad = (sad + 31) >> 6; + + return sad; +} + +#define HIGHBD_MASKSADMXN(m, n) \ + unsigned int aom_highbd_masked_sad##m##x##n##_c( \ + const uint8_t *src8, int src_stride, const uint8_t *ref8, \ + int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, \ + int msk_stride, int invert_mask) { \ + if (!invert_mask) \ + return highbd_masked_sad(src8, src_stride, ref8, ref_stride, \ + second_pred8, m, msk, msk_stride, m, n); \ + else \ + return highbd_masked_sad(src8, src_stride, second_pred8, m, ref8, \ + ref_stride, msk, msk_stride, m, n); \ + } + +HIGHBD_MASKSADMXN(128, 128) +HIGHBD_MASKSADMXN(128, 64) +HIGHBD_MASKSADMXN(64, 128) +HIGHBD_MASKSADMXN(64, 64) +HIGHBD_MASKSADMXN(64, 32) +HIGHBD_MASKSADMXN(32, 64) +HIGHBD_MASKSADMXN(32, 32) +HIGHBD_MASKSADMXN(32, 16) +HIGHBD_MASKSADMXN(16, 32) +HIGHBD_MASKSADMXN(16, 16) +HIGHBD_MASKSADMXN(16, 8) +HIGHBD_MASKSADMXN(8, 16) +HIGHBD_MASKSADMXN(8, 8) +HIGHBD_MASKSADMXN(8, 4) +HIGHBD_MASKSADMXN(4, 8) +HIGHBD_MASKSADMXN(4, 4) +HIGHBD_MASKSADMXN(4, 16) +HIGHBD_MASKSADMXN(16, 4) +HIGHBD_MASKSADMXN(8, 32) +HIGHBD_MASKSADMXN(32, 8) +HIGHBD_MASKSADMXN(16, 64) +HIGHBD_MASKSADMXN(64, 16) + +// pre: predictor being evaluated +// wsrc: target weighted prediction (has been *4096 to keep precision) +// mask: 2d weights (scaled by 4096) +static INLINE unsigned int obmc_sad(const uint8_t *pre, int pre_stride, + const int32_t *wsrc, const int32_t *mask, + int width, int height) { + int y, x; + unsigned int sad = 0; + + for (y = 0; y < height; y++) { + for (x = 0; x < width; x++) + sad += ROUND_POWER_OF_TWO(abs(wsrc[x] - pre[x] * mask[x]), 12); + + pre += pre_stride; + wsrc += width; + mask += width; + } + + return sad; +} + +#define OBMCSADMxN(m, n) \ + unsigned int aom_obmc_sad##m##x##n##_c(const uint8_t *ref, int ref_stride, \ + const int32_t *wsrc, \ + const int32_t *mask) { \ + return obmc_sad(ref, ref_stride, wsrc, mask, m, n); \ + } + +/* clang-format off */ +OBMCSADMxN(128, 128) +OBMCSADMxN(128, 64) +OBMCSADMxN(64, 128) +OBMCSADMxN(64, 64) +OBMCSADMxN(64, 32) +OBMCSADMxN(32, 64) +OBMCSADMxN(32, 32) +OBMCSADMxN(32, 16) +OBMCSADMxN(16, 32) +OBMCSADMxN(16, 16) +OBMCSADMxN(16, 8) +OBMCSADMxN(8, 16) +OBMCSADMxN(8, 8) +OBMCSADMxN(8, 4) +OBMCSADMxN(4, 8) +OBMCSADMxN(4, 4) +OBMCSADMxN(4, 16) +OBMCSADMxN(16, 4) +OBMCSADMxN(8, 32) +OBMCSADMxN(32, 8) +OBMCSADMxN(16, 64) +OBMCSADMxN(64, 16) + /* clang-format on */ + + static INLINE + unsigned int highbd_obmc_sad(const uint8_t *pre8, int pre_stride, + const int32_t *wsrc, const int32_t *mask, + int width, int height) { + int y, x; + unsigned int sad = 0; + const uint16_t *pre = CONVERT_TO_SHORTPTR(pre8); + + for (y = 0; y < height; y++) { + for (x = 0; x < width; x++) + sad += ROUND_POWER_OF_TWO(abs(wsrc[x] - pre[x] * mask[x]), 12); + + pre += pre_stride; + wsrc += width; + mask += width; + } + + return sad; +} + +#define HIGHBD_OBMCSADMXN(m, n) \ + unsigned int aom_highbd_obmc_sad##m##x##n##_c( \ + const uint8_t *ref, int ref_stride, const int32_t *wsrc, \ + const int32_t *mask) { \ + return highbd_obmc_sad(ref, ref_stride, wsrc, mask, m, n); \ + } + +/* clang-format off */ +HIGHBD_OBMCSADMXN(128, 128) +HIGHBD_OBMCSADMXN(128, 64) +HIGHBD_OBMCSADMXN(64, 128) +HIGHBD_OBMCSADMXN(64, 64) +HIGHBD_OBMCSADMXN(64, 32) +HIGHBD_OBMCSADMXN(32, 64) +HIGHBD_OBMCSADMXN(32, 32) +HIGHBD_OBMCSADMXN(32, 16) +HIGHBD_OBMCSADMXN(16, 32) +HIGHBD_OBMCSADMXN(16, 16) +HIGHBD_OBMCSADMXN(16, 8) +HIGHBD_OBMCSADMXN(8, 16) +HIGHBD_OBMCSADMXN(8, 8) +HIGHBD_OBMCSADMXN(8, 4) +HIGHBD_OBMCSADMXN(4, 8) +HIGHBD_OBMCSADMXN(4, 4) +HIGHBD_OBMCSADMXN(4, 16) +HIGHBD_OBMCSADMXN(16, 4) +HIGHBD_OBMCSADMXN(8, 32) +HIGHBD_OBMCSADMXN(32, 8) +HIGHBD_OBMCSADMXN(16, 64) +HIGHBD_OBMCSADMXN(64, 16) +/* clang-format on */ diff --git a/third_party/aom/aom_dsp/simd/v128_intrinsics.h b/third_party/aom/aom_dsp/simd/v128_intrinsics.h new file mode 100644 index 0000000000..01dbb8fd27 --- /dev/null +++ b/third_party/aom/aom_dsp/simd/v128_intrinsics.h @@ -0,0 +1,344 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_SIMD_V128_INTRINSICS_H_ +#define AOM_AOM_DSP_SIMD_V128_INTRINSICS_H_ + +#include +#include +#include + +#include "aom_dsp/simd/v128_intrinsics_c.h" +#include "aom_dsp/simd/v64_intrinsics.h" + +/* Fallback to plain, unoptimised C. */ + +typedef c_v128 v128; + +SIMD_INLINE uint32_t v128_low_u32(v128 a) { return c_v128_low_u32(a); } +SIMD_INLINE v64 v128_low_v64(v128 a) { return c_v128_low_v64(a); } +SIMD_INLINE v64 v128_high_v64(v128 a) { return c_v128_high_v64(a); } +SIMD_INLINE v128 v128_from_64(uint64_t hi, uint64_t lo) { + return c_v128_from_64(hi, lo); +} +SIMD_INLINE v128 v128_from_v64(v64 hi, v64 lo) { + return c_v128_from_v64(hi, lo); +} +SIMD_INLINE v128 v128_from_32(uint32_t a, uint32_t b, uint32_t c, uint32_t d) { + return c_v128_from_32(a, b, c, d); +} + +SIMD_INLINE v128 v128_load_unaligned(const void *p) { + return c_v128_load_unaligned(p); +} +SIMD_INLINE v128 v128_load_aligned(const void *p) { + return c_v128_load_aligned(p); +} + +SIMD_INLINE void v128_store_unaligned(void *p, v128 a) { + c_v128_store_unaligned(p, a); +} +SIMD_INLINE void v128_store_aligned(void *p, v128 a) { + c_v128_store_aligned(p, a); +} + +SIMD_INLINE v128 v128_align(v128 a, v128 b, unsigned int c) { + return c_v128_align(a, b, c); +} + +SIMD_INLINE v128 v128_zero() { return c_v128_zero(); } +SIMD_INLINE v128 v128_dup_8(uint8_t x) { return c_v128_dup_8(x); } +SIMD_INLINE v128 v128_dup_16(uint16_t x) { return c_v128_dup_16(x); } +SIMD_INLINE v128 v128_dup_32(uint32_t x) { return c_v128_dup_32(x); } +SIMD_INLINE v128 v128_dup_64(uint64_t x) { return c_v128_dup_64(x); } + +typedef uint32_t sad128_internal; +SIMD_INLINE sad128_internal v128_sad_u8_init() { return c_v128_sad_u8_init(); } +SIMD_INLINE sad128_internal v128_sad_u8(sad128_internal s, v128 a, v128 b) { + return c_v128_sad_u8(s, a, b); +} +SIMD_INLINE uint32_t v128_sad_u8_sum(sad128_internal s) { + return c_v128_sad_u8_sum(s); +} +typedef uint32_t ssd128_internal; +SIMD_INLINE ssd128_internal v128_ssd_u8_init() { return c_v128_ssd_u8_init(); } +SIMD_INLINE ssd128_internal v128_ssd_u8(ssd128_internal s, v128 a, v128 b) { + return c_v128_ssd_u8(s, a, b); +} +SIMD_INLINE uint32_t v128_ssd_u8_sum(ssd128_internal s) { + return c_v128_ssd_u8_sum(s); +} +SIMD_INLINE int64_t v128_dotp_su8(v128 a, v128 b) { + return c_v128_dotp_su8(a, b); +} +SIMD_INLINE int64_t v128_dotp_s16(v128 a, v128 b) { + return c_v128_dotp_s16(a, b); +} +SIMD_INLINE int64_t v128_dotp_s32(v128 a, v128 b) { + return c_v128_dotp_s32(a, b); +} +SIMD_INLINE uint64_t v128_hadd_u8(v128 a) { return c_v128_hadd_u8(a); } + +SIMD_INLINE v128 v128_or(v128 a, v128 b) { return c_v128_or(a, b); } +SIMD_INLINE v128 v128_xor(v128 a, v128 b) { return c_v128_xor(a, b); } +SIMD_INLINE v128 v128_and(v128 a, v128 b) { return c_v128_and(a, b); } +SIMD_INLINE v128 v128_andn(v128 a, v128 b) { return c_v128_andn(a, b); } + +SIMD_INLINE v128 v128_add_8(v128 a, v128 b) { return c_v128_add_8(a, b); } +SIMD_INLINE v128 v128_add_16(v128 a, v128 b) { return c_v128_add_16(a, b); } +SIMD_INLINE v128 v128_sadd_u8(v128 a, v128 b) { return c_v128_sadd_u8(a, b); } +SIMD_INLINE v128 v128_sadd_s8(v128 a, v128 b) { return c_v128_sadd_s8(a, b); } +SIMD_INLINE v128 v128_sadd_s16(v128 a, v128 b) { return c_v128_sadd_s16(a, b); } +SIMD_INLINE v128 v128_add_32(v128 a, v128 b) { return c_v128_add_32(a, b); } +SIMD_INLINE v128 v128_add_64(v128 a, v128 b) { return c_v128_add_64(a, b); } +SIMD_INLINE v128 v128_padd_u8(v128 a) { return c_v128_padd_u8(a); } +SIMD_INLINE v128 v128_padd_s16(v128 a) { return c_v128_padd_s16(a); } +SIMD_INLINE v128 v128_sub_8(v128 a, v128 b) { return c_v128_sub_8(a, b); } +SIMD_INLINE v128 v128_ssub_u8(v128 a, v128 b) { return c_v128_ssub_u8(a, b); } +SIMD_INLINE v128 v128_ssub_s8(v128 a, v128 b) { return c_v128_ssub_s8(a, b); } +SIMD_INLINE v128 v128_sub_16(v128 a, v128 b) { return c_v128_sub_16(a, b); } +SIMD_INLINE v128 v128_ssub_s16(v128 a, v128 b) { return c_v128_ssub_s16(a, b); } +SIMD_INLINE v128 v128_ssub_u16(v128 a, v128 b) { return c_v128_ssub_u16(a, b); } +SIMD_INLINE v128 v128_sub_32(v128 a, v128 b) { return c_v128_sub_32(a, b); } +SIMD_INLINE v128 v128_sub_64(v128 a, v128 b) { return c_v128_sub_64(a, b); } +SIMD_INLINE v128 v128_abs_s16(v128 a) { return c_v128_abs_s16(a); } +SIMD_INLINE v128 v128_abs_s8(v128 a) { return c_v128_abs_s8(a); } + +SIMD_INLINE v128 v128_mul_s16(v64 a, v64 b) { return c_v128_mul_s16(a, b); } +SIMD_INLINE v128 v128_mullo_s16(v128 a, v128 b) { + return c_v128_mullo_s16(a, b); +} +SIMD_INLINE v128 v128_mulhi_s16(v128 a, v128 b) { + return c_v128_mulhi_s16(a, b); +} +SIMD_INLINE v128 v128_mullo_s32(v128 a, v128 b) { + return c_v128_mullo_s32(a, b); +} +SIMD_INLINE v128 v128_madd_s16(v128 a, v128 b) { return c_v128_madd_s16(a, b); } +SIMD_INLINE v128 v128_madd_us8(v128 a, v128 b) { return c_v128_madd_us8(a, b); } + +SIMD_INLINE uint32_t v128_movemask_8(v128 a) { return c_v128_movemask_8(a); } +SIMD_INLINE v128 v128_blend_8(v128 a, v128 b, v128 c) { + return c_v128_blend_8(a, b, c); +} + +SIMD_INLINE v128 v128_avg_u8(v128 a, v128 b) { return c_v128_avg_u8(a, b); } +SIMD_INLINE v128 v128_rdavg_u8(v128 a, v128 b) { return c_v128_rdavg_u8(a, b); } +SIMD_INLINE v128 v128_rdavg_u16(v128 a, v128 b) { + return c_v128_rdavg_u16(a, b); +} +SIMD_INLINE v128 v128_avg_u16(v128 a, v128 b) { return c_v128_avg_u16(a, b); } +SIMD_INLINE v128 v128_min_u8(v128 a, v128 b) { return c_v128_min_u8(a, b); } +SIMD_INLINE v128 v128_max_u8(v128 a, v128 b) { return c_v128_max_u8(a, b); } +SIMD_INLINE v128 v128_min_s8(v128 a, v128 b) { return c_v128_min_s8(a, b); } +SIMD_INLINE v128 v128_max_s8(v128 a, v128 b) { return c_v128_max_s8(a, b); } +SIMD_INLINE v128 v128_min_s16(v128 a, v128 b) { return c_v128_min_s16(a, b); } +SIMD_INLINE v128 v128_max_s16(v128 a, v128 b) { return c_v128_max_s16(a, b); } +SIMD_INLINE v128 v128_min_s32(v128 a, v128 b) { return c_v128_min_s32(a, b); } +SIMD_INLINE v128 v128_max_s32(v128 a, v128 b) { return c_v128_max_s32(a, b); } + +SIMD_INLINE v128 v128_ziplo_8(v128 a, v128 b) { return c_v128_ziplo_8(a, b); } +SIMD_INLINE v128 v128_ziphi_8(v128 a, v128 b) { return c_v128_ziphi_8(a, b); } +SIMD_INLINE v128 v128_ziplo_16(v128 a, v128 b) { return c_v128_ziplo_16(a, b); } +SIMD_INLINE v128 v128_ziphi_16(v128 a, v128 b) { return c_v128_ziphi_16(a, b); } +SIMD_INLINE v128 v128_ziplo_32(v128 a, v128 b) { return c_v128_ziplo_32(a, b); } +SIMD_INLINE v128 v128_ziphi_32(v128 a, v128 b) { return c_v128_ziphi_32(a, b); } +SIMD_INLINE v128 v128_ziplo_64(v128 a, v128 b) { return c_v128_ziplo_64(a, b); } +SIMD_INLINE v128 v128_ziphi_64(v128 a, v128 b) { return c_v128_ziphi_64(a, b); } +SIMD_INLINE v128 v128_zip_8(v64 a, v64 b) { return c_v128_zip_8(a, b); } +SIMD_INLINE v128 v128_zip_16(v64 a, v64 b) { return c_v128_zip_16(a, b); } +SIMD_INLINE v128 v128_zip_32(v64 a, v64 b) { return c_v128_zip_32(a, b); } +SIMD_INLINE v128 v128_unziplo_8(v128 a, v128 b) { + return c_v128_unziplo_8(a, b); +} +SIMD_INLINE v128 v128_unziphi_8(v128 a, v128 b) { + return c_v128_unziphi_8(a, b); +} +SIMD_INLINE v128 v128_unziplo_16(v128 a, v128 b) { + return c_v128_unziplo_16(a, b); +} +SIMD_INLINE v128 v128_unziphi_16(v128 a, v128 b) { + return c_v128_unziphi_16(a, b); +} +SIMD_INLINE v128 v128_unziplo_32(v128 a, v128 b) { + return c_v128_unziplo_32(a, b); +} +SIMD_INLINE v128 v128_unziphi_32(v128 a, v128 b) { + return c_v128_unziphi_32(a, b); +} +SIMD_INLINE v128 v128_unpack_u8_s16(v64 a) { return c_v128_unpack_u8_s16(a); } +SIMD_INLINE v128 v128_unpacklo_u8_s16(v128 a) { + return c_v128_unpacklo_u8_s16(a); +} +SIMD_INLINE v128 v128_unpackhi_u8_s16(v128 a) { + return c_v128_unpackhi_u8_s16(a); +} +SIMD_INLINE v128 v128_unpack_s8_s16(v64 a) { return c_v128_unpack_s8_s16(a); } +SIMD_INLINE v128 v128_unpacklo_s8_s16(v128 a) { + return c_v128_unpacklo_s8_s16(a); +} +SIMD_INLINE v128 v128_unpackhi_s8_s16(v128 a) { + return c_v128_unpackhi_s8_s16(a); +} +SIMD_INLINE v128 v128_pack_s32_s16(v128 a, v128 b) { + return c_v128_pack_s32_s16(a, b); +} +SIMD_INLINE v128 v128_pack_s32_u16(v128 a, v128 b) { + return c_v128_pack_s32_u16(a, b); +} +SIMD_INLINE v128 v128_pack_s16_u8(v128 a, v128 b) { + return c_v128_pack_s16_u8(a, b); +} +SIMD_INLINE v128 v128_pack_s16_s8(v128 a, v128 b) { + return c_v128_pack_s16_s8(a, b); +} +SIMD_INLINE v128 v128_unpack_u16_s32(v64 a) { return c_v128_unpack_u16_s32(a); } +SIMD_INLINE v128 v128_unpack_s16_s32(v64 a) { return c_v128_unpack_s16_s32(a); } +SIMD_INLINE v128 v128_unpacklo_u16_s32(v128 a) { + return c_v128_unpacklo_u16_s32(a); +} +SIMD_INLINE v128 v128_unpacklo_s16_s32(v128 a) { + return c_v128_unpacklo_s16_s32(a); +} +SIMD_INLINE v128 v128_unpackhi_u16_s32(v128 a) { + return c_v128_unpackhi_u16_s32(a); +} +SIMD_INLINE v128 v128_unpackhi_s16_s32(v128 a) { + return c_v128_unpackhi_s16_s32(a); +} +SIMD_INLINE v128 v128_shuffle_8(v128 a, v128 pattern) { + return c_v128_shuffle_8(a, pattern); +} + +SIMD_INLINE v128 v128_cmpgt_s8(v128 a, v128 b) { return c_v128_cmpgt_s8(a, b); } +SIMD_INLINE v128 v128_cmplt_s8(v128 a, v128 b) { return c_v128_cmplt_s8(a, b); } +SIMD_INLINE v128 v128_cmpeq_8(v128 a, v128 b) { return c_v128_cmpeq_8(a, b); } +SIMD_INLINE v128 v128_cmpgt_s16(v128 a, v128 b) { + return c_v128_cmpgt_s16(a, b); +} +SIMD_INLINE v128 v128_cmplt_s16(v128 a, v128 b) { + return c_v128_cmplt_s16(a, b); +} +SIMD_INLINE v128 v128_cmpeq_16(v128 a, v128 b) { return c_v128_cmpeq_16(a, b); } + +SIMD_INLINE v128 v128_cmpgt_s32(v128 a, v128 b) { + return c_v128_cmpgt_s32(a, b); +} +SIMD_INLINE v128 v128_cmplt_s32(v128 a, v128 b) { + return c_v128_cmplt_s32(a, b); +} +SIMD_INLINE v128 v128_cmpeq_32(v128 a, v128 b) { return c_v128_cmpeq_32(a, b); } + +SIMD_INLINE v128 v128_shl_8(v128 a, unsigned int c) { + return c_v128_shl_8(a, c); +} +SIMD_INLINE v128 v128_shr_u8(v128 a, unsigned int c) { + return c_v128_shr_u8(a, c); +} +SIMD_INLINE v128 v128_shr_s8(v128 a, unsigned int c) { + return c_v128_shr_s8(a, c); +} +SIMD_INLINE v128 v128_shl_16(v128 a, unsigned int c) { + return c_v128_shl_16(a, c); +} +SIMD_INLINE v128 v128_shr_u16(v128 a, unsigned int c) { + return c_v128_shr_u16(a, c); +} +SIMD_INLINE v128 v128_shr_s16(v128 a, unsigned int c) { + return c_v128_shr_s16(a, c); +} +SIMD_INLINE v128 v128_shl_32(v128 a, unsigned int c) { + return c_v128_shl_32(a, c); +} +SIMD_INLINE v128 v128_shr_u32(v128 a, unsigned int c) { + return c_v128_shr_u32(a, c); +} +SIMD_INLINE v128 v128_shr_s32(v128 a, unsigned int c) { + return c_v128_shr_s32(a, c); +} +SIMD_INLINE v128 v128_shl_64(v128 a, unsigned int c) { + return c_v128_shl_64(a, c); +} +SIMD_INLINE v128 v128_shr_u64(v128 a, unsigned int c) { + return c_v128_shr_u64(a, c); +} +SIMD_INLINE v128 v128_shr_s64(v128 a, unsigned int c) { + return c_v128_shr_s64(a, c); +} + +SIMD_INLINE v128 v128_shr_n_byte(v128 a, unsigned int n) { + return c_v128_shr_n_byte(a, n); +} +SIMD_INLINE v128 v128_shl_n_byte(v128 a, unsigned int n) { + return c_v128_shl_n_byte(a, n); +} +SIMD_INLINE v128 v128_shl_n_8(v128 a, unsigned int n) { + return c_v128_shl_n_8(a, n); +} +SIMD_INLINE v128 v128_shl_n_16(v128 a, unsigned int n) { + return c_v128_shl_n_16(a, n); +} +SIMD_INLINE v128 v128_shl_n_32(v128 a, unsigned int n) { + return c_v128_shl_n_32(a, n); +} +SIMD_INLINE v128 v128_shl_n_64(v128 a, unsigned int n) { + return c_v128_shl_n_64(a, n); +} +SIMD_INLINE v128 v128_shr_n_u8(v128 a, unsigned int n) { + return c_v128_shr_n_u8(a, n); +} +SIMD_INLINE v128 v128_shr_n_u16(v128 a, unsigned int n) { + return c_v128_shr_n_u16(a, n); +} +SIMD_INLINE v128 v128_shr_n_u32(v128 a, unsigned int n) { + return c_v128_shr_n_u32(a, n); +} +SIMD_INLINE v128 v128_shr_n_u64(v128 a, unsigned int n) { + return c_v128_shr_n_u64(a, n); +} +SIMD_INLINE v128 v128_shr_n_s8(v128 a, unsigned int n) { + return c_v128_shr_n_s8(a, n); +} +SIMD_INLINE v128 v128_shr_n_s16(v128 a, unsigned int n) { + return c_v128_shr_n_s16(a, n); +} +SIMD_INLINE v128 v128_shr_n_s32(v128 a, unsigned int n) { + return c_v128_shr_n_s32(a, n); +} +SIMD_INLINE v128 v128_shr_n_s64(v128 a, unsigned int n) { + return c_v128_shr_n_s64(a, n); +} + +typedef uint32_t sad128_internal_u16; +SIMD_INLINE sad128_internal_u16 v128_sad_u16_init() { + return c_v128_sad_u16_init(); +} +SIMD_INLINE sad128_internal_u16 v128_sad_u16(sad128_internal_u16 s, v128 a, + v128 b) { + return c_v128_sad_u16(s, a, b); +} +SIMD_INLINE uint32_t v128_sad_u16_sum(sad128_internal_u16 s) { + return c_v128_sad_u16_sum(s); +} + +typedef uint64_t ssd128_internal_s16; +SIMD_INLINE ssd128_internal_s16 v128_ssd_s16_init() { + return c_v128_ssd_s16_init(); +} +SIMD_INLINE ssd128_internal_s16 v128_ssd_s16(ssd128_internal_s16 s, v128 a, + v128 b) { + return c_v128_ssd_s16(s, a, b); +} +SIMD_INLINE uint64_t v128_ssd_s16_sum(ssd128_internal_s16 s) { + return c_v128_ssd_s16_sum(s); +} + +#endif // AOM_AOM_DSP_SIMD_V128_INTRINSICS_H_ diff --git a/third_party/aom/aom_dsp/simd/v128_intrinsics_arm.h b/third_party/aom/aom_dsp/simd/v128_intrinsics_arm.h new file mode 100644 index 0000000000..3c669d5792 --- /dev/null +++ b/third_party/aom/aom_dsp/simd/v128_intrinsics_arm.h @@ -0,0 +1,958 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_SIMD_V128_INTRINSICS_ARM_H_ +#define AOM_AOM_DSP_SIMD_V128_INTRINSICS_ARM_H_ + +#include + +#include "aom_dsp/simd/v64_intrinsics_arm.h" + +typedef int64x2_t v128; + +SIMD_INLINE uint32_t v128_low_u32(v128 a) { + return v64_low_u32(vget_low_s64(a)); +} + +SIMD_INLINE v64 v128_low_v64(v128 a) { return vget_low_s64(a); } + +SIMD_INLINE v64 v128_high_v64(v128 a) { return vget_high_s64(a); } + +SIMD_INLINE v128 v128_from_v64(v64 a, v64 b) { return vcombine_s64(b, a); } + +SIMD_INLINE v128 v128_from_64(uint64_t a, uint64_t b) { + return vcombine_s64((int64x1_t)b, (int64x1_t)a); +} + +SIMD_INLINE v128 v128_from_32(uint32_t a, uint32_t b, uint32_t c, uint32_t d) { + return vcombine_s64(v64_from_32(c, d), v64_from_32(a, b)); +} + +SIMD_INLINE v128 v128_load_aligned(const void *p) { + return vreinterpretq_s64_u8(vld1q_u8((const uint8_t *)p)); +} + +SIMD_INLINE v128 v128_load_unaligned(const void *p) { + return v128_load_aligned(p); +} + +SIMD_INLINE void v128_store_aligned(void *p, v128 r) { + vst1q_u8((uint8_t *)p, vreinterpretq_u8_s64(r)); +} + +SIMD_INLINE void v128_store_unaligned(void *p, v128 r) { + vst1q_u8((uint8_t *)p, vreinterpretq_u8_s64(r)); +} + +SIMD_INLINE v128 v128_align(v128 a, v128 b, unsigned int c) { +// The following functions require an immediate. +// Some compilers will check this during optimisation, others wont. +#if defined(__OPTIMIZE__) && __OPTIMIZE__ && !defined(__clang__) + return c ? vreinterpretq_s64_s8( + vextq_s8(vreinterpretq_s8_s64(b), vreinterpretq_s8_s64(a), c)) + : b; +#else + return c < 8 ? v128_from_v64(v64_align(v128_low_v64(a), v128_high_v64(b), c), + v64_align(v128_high_v64(b), v128_low_v64(b), c)) + : v128_from_v64( + v64_align(v128_high_v64(a), v128_low_v64(a), c - 8), + v64_align(v128_low_v64(a), v128_high_v64(b), c - 8)); +#endif +} + +SIMD_INLINE v128 v128_zero() { return vreinterpretq_s64_u8(vdupq_n_u8(0)); } + +SIMD_INLINE v128 v128_ones() { return vreinterpretq_s64_u8(vdupq_n_u8(-1)); } + +SIMD_INLINE v128 v128_dup_8(uint8_t x) { + return vreinterpretq_s64_u8(vdupq_n_u8(x)); +} + +SIMD_INLINE v128 v128_dup_16(uint16_t x) { + return vreinterpretq_s64_u16(vdupq_n_u16(x)); +} + +SIMD_INLINE v128 v128_dup_32(uint32_t x) { + return vreinterpretq_s64_u32(vdupq_n_u32(x)); +} + +SIMD_INLINE v128 v128_dup_64(uint64_t x) { + return vreinterpretq_s64_u64(vdupq_n_u64(x)); +} + +SIMD_INLINE int64_t v128_dotp_su8(v128 a, v128 b) { + int16x8_t t1 = vmulq_s16( + vmovl_s8(vreinterpret_s8_s64(vget_low_s64(a))), + vreinterpretq_s16_u16(vmovl_u8(vreinterpret_u8_s64(vget_low_s64(b))))); + int16x8_t t2 = vmulq_s16( + vmovl_s8(vreinterpret_s8_s64(vget_high_s64(a))), + vreinterpretq_s16_u16(vmovl_u8(vreinterpret_u8_s64(vget_high_s64(b))))); +#if defined(__aarch64__) + return vaddlvq_s16(t1) + vaddlvq_s16(t2); +#else + int64x2_t t = vpaddlq_s32(vaddq_s32(vpaddlq_s16(t1), vpaddlq_s16(t2))); + return (int64_t)vget_high_s64(t) + (int64_t)vget_low_s64(t); +#endif +} + +SIMD_INLINE int64_t v128_dotp_s16(v128 a, v128 b) { + return v64_dotp_s16(vget_high_s64(a), vget_high_s64(b)) + + v64_dotp_s16(vget_low_s64(a), vget_low_s64(b)); +} + +SIMD_INLINE int64_t v128_dotp_s32(v128 a, v128 b) { + int64x2_t t = vpaddlq_s32( + vmulq_s32(vreinterpretq_s32_s64(a), vreinterpretq_s32_s64(b))); + return (int64_t)vget_high_s64(t) + (int64_t)vget_low_s64(t); +} + +SIMD_INLINE uint64_t v128_hadd_u8(v128 x) { +#if defined(__aarch64__) + return vaddlvq_u8(vreinterpretq_u8_s64(x)); +#else + uint64x2_t t = vpaddlq_u32(vpaddlq_u16(vpaddlq_u8(vreinterpretq_u8_s64(x)))); + return vget_lane_s32( + vreinterpret_s32_u64(vadd_u64(vget_high_u64(t), vget_low_u64(t))), 0); +#endif +} + +SIMD_INLINE v128 v128_padd_s16(v128 a) { + return vreinterpretq_s64_s32(vpaddlq_s16(vreinterpretq_s16_s64(a))); +} + +SIMD_INLINE v128 v128_padd_u8(v128 a) { + return vreinterpretq_s64_u16(vpaddlq_u8(vreinterpretq_u8_s64(a))); +} + +typedef struct { + sad64_internal hi, lo; +} sad128_internal; + +SIMD_INLINE sad128_internal v128_sad_u8_init() { + sad128_internal s; + s.hi = s.lo = vdupq_n_u16(0); + return s; +} + +/* Implementation dependent return value. Result must be finalised with + v128_sad_u8_sum(). + The result for more than 32 v128_sad_u8() calls is undefined. */ +SIMD_INLINE sad128_internal v128_sad_u8(sad128_internal s, v128 a, v128 b) { + sad128_internal r; + r.hi = v64_sad_u8(s.hi, vget_high_s64(a), vget_high_s64(b)); + r.lo = v64_sad_u8(s.lo, vget_low_s64(a), vget_low_s64(b)); + return r; +} + +SIMD_INLINE uint32_t v128_sad_u8_sum(sad128_internal s) { +#if defined(__aarch64__) + return vaddlvq_u16(s.hi) + vaddlvq_u16(s.lo); +#else + uint64x2_t t = vpaddlq_u32(vpaddlq_u16(vaddq_u16(s.hi, s.lo))); + return (uint32_t)(uint64_t)(vget_high_u64(t) + vget_low_u64(t)); +#endif +} + +typedef struct { + ssd64_internal hi, lo; +} ssd128_internal; + +SIMD_INLINE ssd128_internal v128_ssd_u8_init() { + ssd128_internal s; + s.hi = s.lo = v64_ssd_u8_init(); + return s; +} + +/* Implementation dependent return value. Result must be finalised with + * v128_ssd_u8_sum(). */ +SIMD_INLINE ssd128_internal v128_ssd_u8(ssd128_internal s, v128 a, v128 b) { + ssd128_internal r; + r.hi = v64_ssd_u8(s.hi, vget_high_s64(a), vget_high_s64(b)); + r.lo = v64_ssd_u8(s.lo, vget_low_s64(a), vget_low_s64(b)); + return r; +} + +SIMD_INLINE uint32_t v128_ssd_u8_sum(ssd128_internal s) { + return (uint32_t)(v64_ssd_u8_sum(s.hi) + v64_ssd_u8_sum(s.lo)); +} + +SIMD_INLINE v128 v128_or(v128 x, v128 y) { return vorrq_s64(x, y); } + +SIMD_INLINE v128 v128_xor(v128 x, v128 y) { return veorq_s64(x, y); } + +SIMD_INLINE v128 v128_and(v128 x, v128 y) { return vandq_s64(x, y); } + +SIMD_INLINE v128 v128_andn(v128 x, v128 y) { return vbicq_s64(x, y); } + +SIMD_INLINE v128 v128_add_8(v128 x, v128 y) { + return vreinterpretq_s64_u8( + vaddq_u8(vreinterpretq_u8_s64(x), vreinterpretq_u8_s64(y))); +} + +SIMD_INLINE v128 v128_sadd_u8(v128 x, v128 y) { + return vreinterpretq_s64_u8( + vqaddq_u8(vreinterpretq_u8_s64(x), vreinterpretq_u8_s64(y))); +} + +SIMD_INLINE v128 v128_sadd_s8(v128 x, v128 y) { + return vreinterpretq_s64_s8( + vqaddq_s8(vreinterpretq_s8_s64(x), vreinterpretq_s8_s64(y))); +} + +SIMD_INLINE v128 v128_add_16(v128 x, v128 y) { + return vreinterpretq_s64_s16( + vaddq_s16(vreinterpretq_s16_s64(x), vreinterpretq_s16_s64(y))); +} + +SIMD_INLINE v128 v128_sadd_s16(v128 x, v128 y) { + return vreinterpretq_s64_s16( + vqaddq_s16(vreinterpretq_s16_s64(x), vreinterpretq_s16_s64(y))); +} + +SIMD_INLINE v128 v128_add_32(v128 x, v128 y) { + return vreinterpretq_s64_u32( + vaddq_u32(vreinterpretq_u32_s64(x), vreinterpretq_u32_s64(y))); +} + +SIMD_INLINE v128 v128_add_64(v128 x, v128 y) { + return vreinterpretq_s64_u64( + vaddq_u64(vreinterpretq_u64_s64(x), vreinterpretq_u64_s64(y))); +} + +SIMD_INLINE v128 v128_sub_8(v128 x, v128 y) { + return vreinterpretq_s64_u8( + vsubq_u8(vreinterpretq_u8_s64(x), vreinterpretq_u8_s64(y))); +} + +SIMD_INLINE v128 v128_sub_16(v128 x, v128 y) { + return vreinterpretq_s64_s16( + vsubq_s16(vreinterpretq_s16_s64(x), vreinterpretq_s16_s64(y))); +} + +SIMD_INLINE v128 v128_ssub_s16(v128 x, v128 y) { + return vreinterpretq_s64_s16( + vqsubq_s16(vreinterpretq_s16_s64(x), vreinterpretq_s16_s64(y))); +} + +SIMD_INLINE v128 v128_ssub_u16(v128 x, v128 y) { + return vreinterpretq_s64_u16( + vqsubq_u16(vreinterpretq_u16_s64(x), vreinterpretq_u16_s64(y))); +} + +SIMD_INLINE v128 v128_ssub_u8(v128 x, v128 y) { + return vreinterpretq_s64_u8( + vqsubq_u8(vreinterpretq_u8_s64(x), vreinterpretq_u8_s64(y))); +} + +SIMD_INLINE v128 v128_ssub_s8(v128 x, v128 y) { + return vreinterpretq_s64_s8( + vqsubq_s8(vreinterpretq_s8_s64(x), vreinterpretq_s8_s64(y))); +} + +SIMD_INLINE v128 v128_sub_32(v128 x, v128 y) { + return vreinterpretq_s64_s32( + vsubq_s32(vreinterpretq_s32_s64(x), vreinterpretq_s32_s64(y))); +} + +SIMD_INLINE v128 v128_sub_64(v128 x, v128 y) { return vsubq_s64(x, y); } + +SIMD_INLINE v128 v128_abs_s16(v128 x) { + return vreinterpretq_s64_s16(vabsq_s16(vreinterpretq_s16_s64(x))); +} + +SIMD_INLINE v128 v128_abs_s8(v128 x) { + return vreinterpretq_s64_s8(vabsq_s8(vreinterpretq_s8_s64(x))); +} + +SIMD_INLINE v128 v128_mul_s16(v64 a, v64 b) { + return vreinterpretq_s64_s32( + vmull_s16(vreinterpret_s16_s64(a), vreinterpret_s16_s64(b))); +} + +SIMD_INLINE v128 v128_mullo_s16(v128 a, v128 b) { + return vreinterpretq_s64_s16( + vmulq_s16(vreinterpretq_s16_s64(a), vreinterpretq_s16_s64(b))); +} + +SIMD_INLINE v128 v128_mulhi_s16(v128 a, v128 b) { +#if defined(__aarch64__) + return vreinterpretq_s64_s16(vuzp2q_s16( + vreinterpretq_s16_s32(vmull_s16(vreinterpret_s16_s64(vget_low_s64(a)), + vreinterpret_s16_s64(vget_low_s64(b)))), + vreinterpretq_s16_s32( + vmull_high_s16(vreinterpretq_s16_s64(a), vreinterpretq_s16_s64(b))))); +#else + return v128_from_v64(v64_mulhi_s16(vget_high_s64(a), vget_high_s64(b)), + v64_mulhi_s16(vget_low_s64(a), vget_low_s64(b))); +#endif +} + +SIMD_INLINE v128 v128_mullo_s32(v128 a, v128 b) { + return vreinterpretq_s64_s32( + vmulq_s32(vreinterpretq_s32_s64(a), vreinterpretq_s32_s64(b))); +} + +SIMD_INLINE v128 v128_madd_s16(v128 a, v128 b) { +#if defined(__aarch64__) + int32x4_t t1 = vmull_s16(vreinterpret_s16_s64(vget_low_s64(a)), + vreinterpret_s16_s64(vget_low_s64(b))); + int32x4_t t2 = + vmull_high_s16(vreinterpretq_s16_s64(a), vreinterpretq_s16_s64(b)); + return vreinterpretq_s64_s32(vpaddq_s32(t1, t2)); +#else + return v128_from_v64(v64_madd_s16(vget_high_s64(a), vget_high_s64(b)), + v64_madd_s16(vget_low_s64(a), vget_low_s64(b))); +#endif +} + +SIMD_INLINE v128 v128_madd_us8(v128 a, v128 b) { +#if defined(__aarch64__) + int16x8_t t1 = vmulq_s16( + vreinterpretq_s16_u16(vmovl_u8(vreinterpret_u8_s64(vget_low_s64(a)))), + vmovl_s8(vreinterpret_s8_s64(vget_low_s64(b)))); + int16x8_t t2 = vmulq_s16( + vreinterpretq_s16_u16(vmovl_u8(vreinterpret_u8_s64(vget_high_s64(a)))), + vmovl_s8(vreinterpret_s8_s64(vget_high_s64(b)))); + return vreinterpretq_s64_s16( + vqaddq_s16(vuzp1q_s16(t1, t2), vuzp2q_s16(t1, t2))); +#else + return v128_from_v64(v64_madd_us8(vget_high_s64(a), vget_high_s64(b)), + v64_madd_us8(vget_low_s64(a), vget_low_s64(b))); +#endif +} + +SIMD_INLINE v128 v128_avg_u8(v128 x, v128 y) { + return vreinterpretq_s64_u8( + vrhaddq_u8(vreinterpretq_u8_s64(x), vreinterpretq_u8_s64(y))); +} + +SIMD_INLINE v128 v128_rdavg_u8(v128 x, v128 y) { + return vreinterpretq_s64_u8( + vhaddq_u8(vreinterpretq_u8_s64(x), vreinterpretq_u8_s64(y))); +} + +SIMD_INLINE v128 v128_rdavg_u16(v128 x, v128 y) { + return vreinterpretq_s64_u16( + vhaddq_u16(vreinterpretq_u16_s64(x), vreinterpretq_u16_s64(y))); +} + +SIMD_INLINE v128 v128_avg_u16(v128 x, v128 y) { + return vreinterpretq_s64_u16( + vrhaddq_u16(vreinterpretq_u16_s64(x), vreinterpretq_u16_s64(y))); +} + +SIMD_INLINE v128 v128_min_u8(v128 x, v128 y) { + return vreinterpretq_s64_u8( + vminq_u8(vreinterpretq_u8_s64(x), vreinterpretq_u8_s64(y))); +} + +SIMD_INLINE v128 v128_max_u8(v128 x, v128 y) { + return vreinterpretq_s64_u8( + vmaxq_u8(vreinterpretq_u8_s64(x), vreinterpretq_u8_s64(y))); +} + +SIMD_INLINE v128 v128_min_s8(v128 x, v128 y) { + return vreinterpretq_s64_s8( + vminq_s8(vreinterpretq_s8_s64(x), vreinterpretq_s8_s64(y))); +} + +SIMD_INLINE uint32_t v128_movemask_8(v128 a) { + a = vreinterpretq_s64_u8(vcltq_s8(vreinterpretq_s8_s64(a), vdupq_n_s8(0))); +#if defined(__aarch64__) + uint8x16_t m = + vandq_u8(vreinterpretq_u8_s64(a), + vreinterpretq_u8_u64(vdupq_n_u64(0x8040201008040201ULL))); + return vaddv_u8(vget_low_u8(m)) + (vaddv_u8(vget_high_u8(m)) << 8); +#else + uint64x2_t m = vpaddlq_u32(vpaddlq_u16(vpaddlq_u8( + vandq_u8(vreinterpretq_u8_s64(a), + vreinterpretq_u8_u64(vdupq_n_u64(0x8040201008040201ULL)))))); + return v64_low_u32( + v64_ziplo_8(v128_high_v64((v128)m), v128_low_v64((v128)m))); +#endif +} + +SIMD_INLINE v128 v128_blend_8(v128 a, v128 b, v128 c) { + c = vreinterpretq_s64_u8(vcltq_s8(vreinterpretq_s8_s64(c), vdupq_n_s8(0))); + return v128_or(v128_and(b, c), v128_andn(a, c)); +} + +SIMD_INLINE v128 v128_max_s8(v128 x, v128 y) { + return vreinterpretq_s64_s8( + vmaxq_s8(vreinterpretq_s8_s64(x), vreinterpretq_s8_s64(y))); +} + +SIMD_INLINE v128 v128_min_s16(v128 x, v128 y) { + return vreinterpretq_s64_s16( + vminq_s16(vreinterpretq_s16_s64(x), vreinterpretq_s16_s64(y))); +} + +SIMD_INLINE v128 v128_max_s16(v128 x, v128 y) { + return vreinterpretq_s64_s16( + vmaxq_s16(vreinterpretq_s16_s64(x), vreinterpretq_s16_s64(y))); +} + +SIMD_INLINE v128 v128_min_s32(v128 x, v128 y) { + return vreinterpretq_s64_s32( + vminq_s32(vreinterpretq_s32_s64(x), vreinterpretq_s32_s64(y))); +} + +SIMD_INLINE v128 v128_max_s32(v128 x, v128 y) { + return vreinterpretq_s64_s32( + vmaxq_s32(vreinterpretq_s32_s64(x), vreinterpretq_s32_s64(y))); +} + +SIMD_INLINE v128 v128_ziplo_8(v128 x, v128 y) { +#if defined(__aarch64__) + return vreinterpretq_s64_u8( + vzip1q_u8(vreinterpretq_u8_s64(y), vreinterpretq_u8_s64(x))); +#else + uint8x16x2_t r = vzipq_u8(vreinterpretq_u8_s64(y), vreinterpretq_u8_s64(x)); + return vreinterpretq_s64_u8(r.val[0]); +#endif +} + +SIMD_INLINE v128 v128_ziphi_8(v128 x, v128 y) { +#if defined(__aarch64__) + return vreinterpretq_s64_u8( + vzip2q_u8(vreinterpretq_u8_s64(y), vreinterpretq_u8_s64(x))); +#else + uint8x16x2_t r = vzipq_u8(vreinterpretq_u8_s64(y), vreinterpretq_u8_s64(x)); + return vreinterpretq_s64_u8(r.val[1]); +#endif +} + +SIMD_INLINE v128 v128_zip_8(v64 x, v64 y) { + uint8x8x2_t r = vzip_u8(vreinterpret_u8_s64(y), vreinterpret_u8_s64(x)); + return vreinterpretq_s64_u8(vcombine_u8(r.val[0], r.val[1])); +} + +SIMD_INLINE v128 v128_ziplo_16(v128 x, v128 y) { +#if defined(__aarch64__) + return vreinterpretq_s64_u16( + vzip1q_u16(vreinterpretq_u16_s64(y), vreinterpretq_u16_s64(x))); +#else + int16x8x2_t r = vzipq_s16(vreinterpretq_s16_s64(y), vreinterpretq_s16_s64(x)); + return vreinterpretq_s64_s16(r.val[0]); +#endif +} + +SIMD_INLINE v128 v128_ziphi_16(v128 x, v128 y) { +#if defined(__aarch64__) + return vreinterpretq_s64_u16( + vzip2q_u16(vreinterpretq_u16_s64(y), vreinterpretq_u16_s64(x))); +#else + int16x8x2_t r = vzipq_s16(vreinterpretq_s16_s64(y), vreinterpretq_s16_s64(x)); + return vreinterpretq_s64_s16(r.val[1]); +#endif +} + +SIMD_INLINE v128 v128_zip_16(v64 x, v64 y) { + uint16x4x2_t r = vzip_u16(vreinterpret_u16_s64(y), vreinterpret_u16_s64(x)); + return vreinterpretq_s64_u16(vcombine_u16(r.val[0], r.val[1])); +} + +SIMD_INLINE v128 v128_ziplo_32(v128 x, v128 y) { +#if defined(__aarch64__) + return vreinterpretq_s64_u32( + vzip1q_u32(vreinterpretq_u32_s64(y), vreinterpretq_u32_s64(x))); +#else + int32x4x2_t r = vzipq_s32(vreinterpretq_s32_s64(y), vreinterpretq_s32_s64(x)); + return vreinterpretq_s64_s32(r.val[0]); +#endif +} + +SIMD_INLINE v128 v128_ziphi_32(v128 x, v128 y) { +#if defined(__aarch64__) + return vreinterpretq_s64_u32( + vzip2q_u32(vreinterpretq_u32_s64(y), vreinterpretq_u32_s64(x))); +#else + int32x4x2_t r = vzipq_s32(vreinterpretq_s32_s64(y), vreinterpretq_s32_s64(x)); + return vreinterpretq_s64_s32(r.val[1]); +#endif +} + +SIMD_INLINE v128 v128_zip_32(v64 x, v64 y) { + uint32x2x2_t r = vzip_u32(vreinterpret_u32_s64(y), vreinterpret_u32_s64(x)); + return vreinterpretq_s64_u32(vcombine_u32(r.val[0], r.val[1])); +} + +SIMD_INLINE v128 v128_ziplo_64(v128 a, v128 b) { + return v128_from_v64(vget_low_s64((int64x2_t)a), vget_low_s64((int64x2_t)b)); +} + +SIMD_INLINE v128 v128_ziphi_64(v128 a, v128 b) { + return v128_from_v64(vget_high_s64((int64x2_t)a), + vget_high_s64((int64x2_t)b)); +} + +SIMD_INLINE v128 v128_unziplo_8(v128 x, v128 y) { +#if defined(__aarch64__) + return vreinterpretq_s64_u8( + vuzp1q_u8(vreinterpretq_u8_s64(y), vreinterpretq_u8_s64(x))); +#else + uint8x16x2_t r = vuzpq_u8(vreinterpretq_u8_s64(y), vreinterpretq_u8_s64(x)); + return vreinterpretq_s64_u8(r.val[0]); +#endif +} + +SIMD_INLINE v128 v128_unziphi_8(v128 x, v128 y) { +#if defined(__aarch64__) + return vreinterpretq_s64_u8( + vuzp2q_u8(vreinterpretq_u8_s64(y), vreinterpretq_u8_s64(x))); +#else + uint8x16x2_t r = vuzpq_u8(vreinterpretq_u8_s64(y), vreinterpretq_u8_s64(x)); + return vreinterpretq_s64_u8(r.val[1]); +#endif +} + +SIMD_INLINE v128 v128_unziplo_16(v128 x, v128 y) { +#if defined(__aarch64__) + return vreinterpretq_s64_u16( + vuzp1q_u16(vreinterpretq_u16_s64(y), vreinterpretq_u16_s64(x))); +#else + uint16x8x2_t r = + vuzpq_u16(vreinterpretq_u16_s64(y), vreinterpretq_u16_s64(x)); + return vreinterpretq_s64_u16(r.val[0]); +#endif +} + +SIMD_INLINE v128 v128_unziphi_16(v128 x, v128 y) { +#if defined(__aarch64__) + return vreinterpretq_s64_u16( + vuzp2q_u16(vreinterpretq_u16_s64(y), vreinterpretq_u16_s64(x))); +#else + uint16x8x2_t r = + vuzpq_u16(vreinterpretq_u16_s64(y), vreinterpretq_u16_s64(x)); + return vreinterpretq_s64_u16(r.val[1]); +#endif +} + +SIMD_INLINE v128 v128_unziplo_32(v128 x, v128 y) { +#if defined(__aarch64__) + return vreinterpretq_s64_u32( + vuzp1q_u32(vreinterpretq_u32_s64(y), vreinterpretq_u32_s64(x))); +#else + uint32x4x2_t r = + vuzpq_u32(vreinterpretq_u32_s64(y), vreinterpretq_u32_s64(x)); + return vreinterpretq_s64_u32(r.val[0]); +#endif +} + +SIMD_INLINE v128 v128_unziphi_32(v128 x, v128 y) { +#if defined(__aarch64__) + return vreinterpretq_s64_u32( + vuzp2q_u32(vreinterpretq_u32_s64(y), vreinterpretq_u32_s64(x))); +#else + uint32x4x2_t r = + vuzpq_u32(vreinterpretq_u32_s64(y), vreinterpretq_u32_s64(x)); + return vreinterpretq_s64_u32(r.val[1]); +#endif +} + +SIMD_INLINE v128 v128_unpack_u8_s16(v64 a) { + return vreinterpretq_s64_u16(vmovl_u8(vreinterpret_u8_s64(a))); +} + +SIMD_INLINE v128 v128_unpacklo_u8_s16(v128 a) { + return vreinterpretq_s64_u16(vmovl_u8(vreinterpret_u8_s64(vget_low_s64(a)))); +} + +SIMD_INLINE v128 v128_unpackhi_u8_s16(v128 a) { + return vreinterpretq_s64_u16(vmovl_u8(vreinterpret_u8_s64(vget_high_s64(a)))); +} + +SIMD_INLINE v128 v128_unpack_s8_s16(v64 a) { + return vreinterpretq_s64_s16(vmovl_s8(vreinterpret_s8_s64(a))); +} + +SIMD_INLINE v128 v128_unpacklo_s8_s16(v128 a) { + return vreinterpretq_s64_s16(vmovl_s8(vreinterpret_s8_s64(vget_low_s64(a)))); +} + +SIMD_INLINE v128 v128_unpackhi_s8_s16(v128 a) { + return vreinterpretq_s64_s16(vmovl_s8(vreinterpret_s8_s64(vget_high_s64(a)))); +} + +SIMD_INLINE v128 v128_pack_s32_s16(v128 a, v128 b) { + return v128_from_v64( + vreinterpret_s64_s16(vqmovn_s32(vreinterpretq_s32_s64(a))), + vreinterpret_s64_s16(vqmovn_s32(vreinterpretq_s32_s64(b)))); +} + +SIMD_INLINE v128 v128_pack_s32_u16(v128 a, v128 b) { + return v128_from_v64( + vreinterpret_s64_u16(vqmovun_s32(vreinterpretq_s32_s64(a))), + vreinterpret_s64_u16(vqmovun_s32(vreinterpretq_s32_s64(b)))); +} + +SIMD_INLINE v128 v128_pack_s16_u8(v128 a, v128 b) { + return v128_from_v64( + vreinterpret_s64_u8(vqmovun_s16(vreinterpretq_s16_s64(a))), + vreinterpret_s64_u8(vqmovun_s16(vreinterpretq_s16_s64(b)))); +} + +SIMD_INLINE v128 v128_pack_s16_s8(v128 a, v128 b) { + return v128_from_v64( + vreinterpret_s64_s8(vqmovn_s16(vreinterpretq_s16_s64(a))), + vreinterpret_s64_s8(vqmovn_s16(vreinterpretq_s16_s64(b)))); +} + +SIMD_INLINE v128 v128_unpack_u16_s32(v64 a) { + return vreinterpretq_s64_u32(vmovl_u16(vreinterpret_u16_s64(a))); +} + +SIMD_INLINE v128 v128_unpack_s16_s32(v64 a) { + return vreinterpretq_s64_s32(vmovl_s16(vreinterpret_s16_s64(a))); +} + +SIMD_INLINE v128 v128_unpacklo_u16_s32(v128 a) { + return vreinterpretq_s64_u32( + vmovl_u16(vreinterpret_u16_s64(vget_low_s64(a)))); +} + +SIMD_INLINE v128 v128_unpacklo_s16_s32(v128 a) { + return vreinterpretq_s64_s32( + vmovl_s16(vreinterpret_s16_s64(vget_low_s64(a)))); +} + +SIMD_INLINE v128 v128_unpackhi_u16_s32(v128 a) { + return vreinterpretq_s64_u32( + vmovl_u16(vreinterpret_u16_s64(vget_high_s64(a)))); +} + +SIMD_INLINE v128 v128_unpackhi_s16_s32(v128 a) { + return vreinterpretq_s64_s32( + vmovl_s16(vreinterpret_s16_s64(vget_high_s64(a)))); +} + +SIMD_INLINE v128 v128_shuffle_8(v128 x, v128 pattern) { +#if defined(__aarch64__) + return vreinterpretq_s64_u8( + vqtbl1q_u8(vreinterpretq_u8_s64(x), vreinterpretq_u8_s64(pattern))); +#else + uint8x8x2_t p = { { vget_low_u8(vreinterpretq_u8_s64(x)), + vget_high_u8(vreinterpretq_u8_s64(x)) } }; + return v128_from_64((uint64_t)vreinterpret_s64_u8(vtbl2_u8( + p, vreinterpret_u8_s64(vget_high_s64(pattern)))), + (uint64_t)vreinterpret_s64_u8(vtbl2_u8( + p, vreinterpret_u8_s64(vget_low_s64(pattern))))); +#endif +} + +SIMD_INLINE v128 v128_cmpgt_s8(v128 x, v128 y) { + return vreinterpretq_s64_u8( + vcgtq_s8(vreinterpretq_s8_s64(x), vreinterpretq_s8_s64(y))); +} + +SIMD_INLINE v128 v128_cmplt_s8(v128 x, v128 y) { + return vreinterpretq_s64_u8( + vcltq_s8(vreinterpretq_s8_s64(x), vreinterpretq_s8_s64(y))); +} + +SIMD_INLINE v128 v128_cmpeq_8(v128 x, v128 y) { + return vreinterpretq_s64_u8( + vceqq_u8(vreinterpretq_u8_s64(x), vreinterpretq_u8_s64(y))); +} + +SIMD_INLINE v128 v128_cmpgt_s16(v128 x, v128 y) { + return vreinterpretq_s64_u16( + vcgtq_s16(vreinterpretq_s16_s64(x), vreinterpretq_s16_s64(y))); +} + +SIMD_INLINE v128 v128_cmplt_s16(v128 x, v128 y) { + return vreinterpretq_s64_u16( + vcltq_s16(vreinterpretq_s16_s64(x), vreinterpretq_s16_s64(y))); +} + +SIMD_INLINE v128 v128_cmpeq_16(v128 x, v128 y) { + return vreinterpretq_s64_u16( + vceqq_s16(vreinterpretq_s16_s64(x), vreinterpretq_s16_s64(y))); +} + +SIMD_INLINE v128 v128_cmpgt_s32(v128 x, v128 y) { + return vreinterpretq_s64_u32( + vcgtq_s32(vreinterpretq_s32_s64(x), vreinterpretq_s32_s64(y))); +} + +SIMD_INLINE v128 v128_cmplt_s32(v128 x, v128 y) { + return vreinterpretq_s64_u32( + vcltq_s32(vreinterpretq_s32_s64(x), vreinterpretq_s32_s64(y))); +} + +SIMD_INLINE v128 v128_cmpeq_32(v128 x, v128 y) { + return vreinterpretq_s64_u32( + vceqq_s32(vreinterpretq_s32_s64(x), vreinterpretq_s32_s64(y))); +} + +SIMD_INLINE v128 v128_shl_8(v128 a, unsigned int c) { + return (c > 7) ? v128_zero() + : vreinterpretq_s64_u8( + vshlq_u8(vreinterpretq_u8_s64(a), vdupq_n_s8(c))); +} + +SIMD_INLINE v128 v128_shr_u8(v128 a, unsigned int c) { + return (c > 7) ? v128_zero() + : vreinterpretq_s64_u8( + vshlq_u8(vreinterpretq_u8_s64(a), vdupq_n_s8(-c))); +} + +SIMD_INLINE v128 v128_shr_s8(v128 a, unsigned int c) { + return (c > 7) ? v128_ones() + : vreinterpretq_s64_s8( + vshlq_s8(vreinterpretq_s8_s64(a), vdupq_n_s8(-c))); +} + +SIMD_INLINE v128 v128_shl_16(v128 a, unsigned int c) { + return (c > 15) ? v128_zero() + : vreinterpretq_s64_u16( + vshlq_u16(vreinterpretq_u16_s64(a), vdupq_n_s16(c))); +} + +SIMD_INLINE v128 v128_shr_u16(v128 a, unsigned int c) { + return (c > 15) ? v128_zero() + : vreinterpretq_s64_u16( + vshlq_u16(vreinterpretq_u16_s64(a), vdupq_n_s16(-c))); +} + +SIMD_INLINE v128 v128_shr_s16(v128 a, unsigned int c) { + return (c > 15) ? v128_ones() + : vreinterpretq_s64_s16( + vshlq_s16(vreinterpretq_s16_s64(a), vdupq_n_s16(-c))); +} + +SIMD_INLINE v128 v128_shl_32(v128 a, unsigned int c) { + return (c > 31) ? v128_zero() + : vreinterpretq_s64_u32( + vshlq_u32(vreinterpretq_u32_s64(a), vdupq_n_s32(c))); +} + +SIMD_INLINE v128 v128_shr_u32(v128 a, unsigned int c) { + return (c > 31) ? v128_zero() + : vreinterpretq_s64_u32( + vshlq_u32(vreinterpretq_u32_s64(a), vdupq_n_s32(-c))); +} + +SIMD_INLINE v128 v128_shr_s32(v128 a, unsigned int c) { + return (c > 31) ? v128_ones() + : vreinterpretq_s64_s32( + vshlq_s32(vreinterpretq_s32_s64(a), vdupq_n_s32(-c))); +} + +SIMD_INLINE v128 v128_shl_64(v128 a, unsigned int c) { + return (c > 63) ? v128_zero() + : vreinterpretq_s64_u64( + vshlq_u64(vreinterpretq_u64_s64(a), vdupq_n_s64(c))); +} + +SIMD_INLINE v128 v128_shr_u64(v128 a, unsigned int c) { + return (c > 63) ? v128_zero() + : vreinterpretq_s64_u64( + vshlq_u64(vreinterpretq_u64_s64(a), vdupq_n_s64(-c))); +} + +SIMD_INLINE v128 v128_shr_s64(v128 a, unsigned int c) { + return (c > 63) ? v128_ones() : vshlq_s64(a, vdupq_n_s64(-c)); +} + +#if defined(__OPTIMIZE__) && __OPTIMIZE__ && !defined(__clang__) + +SIMD_INLINE v128 v128_shl_n_byte(v128 a, unsigned int n) { + return n < 8 + ? v128_from_64( + (uint64_t)vorr_u64( + vshl_n_u64(vreinterpret_u64_s64(vget_high_s64(a)), + n * 8), + vshr_n_u64(vreinterpret_u64_s64(vget_low_s64(a)), + (8 - n) * 8)), + (uint64_t)vshl_n_u64(vreinterpret_u64_s64(vget_low_s64(a)), + n * 8)) + : (n == 8 ? v128_from_64( + (uint64_t)vreinterpret_u64_s64(vget_low_s64(a)), 0) + : v128_from_64((uint64_t)vshl_n_u64( + vreinterpret_u64_s64(vget_low_s64(a)), + (n - 8) * 8), + 0)); +} + +SIMD_INLINE v128 v128_shr_n_byte(v128 a, unsigned int n) { + return n < 8 + ? v128_from_64( + (uint64_t)vshr_n_u64(vreinterpret_u64_s64(vget_high_s64(a)), + n * 8), + (uint64_t)vorr_u64( + vshr_n_u64(vreinterpret_u64_s64(vget_low_s64(a)), n * 8), + vshl_n_u64(vreinterpret_u64_s64(vget_high_s64(a)), + (8 - n) * 8))) + : (n == 8 ? v128_from_64(0, (uint64_t)vreinterpret_u64_s64( + vget_high_s64(a))) + : v128_from_64( + 0, (uint64_t)vshr_n_u64( + vreinterpret_u64_s64(vget_high_s64(a)), + (n - 8) * 8))); +} + +SIMD_INLINE v128 v128_shl_n_8(v128 a, unsigned int c) { + return vreinterpretq_s64_u8(vshlq_n_u8(vreinterpretq_u8_s64(a), c)); +} + +SIMD_INLINE v128 v128_shr_n_u8(v128 a, unsigned int c) { + return vreinterpretq_s64_u8(vshrq_n_u8(vreinterpretq_u8_s64(a), c)); +} + +SIMD_INLINE v128 v128_shr_n_s8(v128 a, unsigned int c) { + return vreinterpretq_s64_s8(vshrq_n_s8(vreinterpretq_s8_s64(a), c)); +} + +SIMD_INLINE v128 v128_shl_n_16(v128 a, unsigned int c) { + return vreinterpretq_s64_u16(vshlq_n_u16(vreinterpretq_u16_s64(a), c)); +} + +SIMD_INLINE v128 v128_shr_n_u16(v128 a, unsigned int c) { + return vreinterpretq_s64_u16(vshrq_n_u16(vreinterpretq_u16_s64(a), c)); +} + +SIMD_INLINE v128 v128_shr_n_s16(v128 a, unsigned int c) { + return vreinterpretq_s64_s16(vshrq_n_s16(vreinterpretq_s16_s64(a), c)); +} + +SIMD_INLINE v128 v128_shl_n_32(v128 a, unsigned int c) { + return vreinterpretq_s64_u32(vshlq_n_u32(vreinterpretq_u32_s64(a), c)); +} + +SIMD_INLINE v128 v128_shr_n_u32(v128 a, unsigned int c) { + return vreinterpretq_s64_u32(vshrq_n_u32(vreinterpretq_u32_s64(a), c)); +} + +SIMD_INLINE v128 v128_shr_n_s32(v128 a, unsigned int c) { + return vreinterpretq_s64_s32(vshrq_n_s32(vreinterpretq_s32_s64(a), c)); +} + +SIMD_INLINE v128 v128_shl_n_64(v128 a, unsigned int c) { + return vreinterpretq_s64_u64(vshlq_n_u64(vreinterpretq_u64_s64(a), c)); +} + +SIMD_INLINE v128 v128_shr_n_u64(v128 a, unsigned int c) { + return vreinterpretq_s64_u64(vshrq_n_u64(vreinterpretq_u64_s64(a), c)); +} + +SIMD_INLINE v128 v128_shr_n_s64(v128 a, unsigned int c) { + return vshrq_n_s64(a, c); +} + +#else + +SIMD_INLINE v128 v128_shl_n_byte(v128 a, unsigned int n) { + if (n < 8) + return v128_from_v64(v64_or(v64_shl_n_byte(v128_high_v64(a), n), + v64_shr_n_byte(v128_low_v64(a), 8 - n)), + v64_shl_n_byte(v128_low_v64(a), n)); + else + return v128_from_v64(v64_shl_n_byte(v128_low_v64(a), n - 8), v64_zero()); +} + +SIMD_INLINE v128 v128_shr_n_byte(v128 a, unsigned int n) { + if (n < 8) + return v128_from_v64(v64_shr_n_byte(v128_high_v64(a), n), + v64_or(v64_shr_n_byte(v128_low_v64(a), n), + v64_shl_n_byte(v128_high_v64(a), 8 - n))); + else + return v128_from_v64(v64_zero(), v64_shr_n_byte(v128_high_v64(a), n - 8)); +} + +SIMD_INLINE v128 v128_shl_n_8(v128 a, unsigned int c) { + return v128_shl_8(a, c); +} + +SIMD_INLINE v128 v128_shr_n_u8(v128 a, unsigned int c) { + return v128_shr_u8(a, c); +} + +SIMD_INLINE v128 v128_shr_n_s8(v128 a, unsigned int c) { + return v128_shr_s8(a, c); +} + +SIMD_INLINE v128 v128_shl_n_16(v128 a, unsigned int c) { + return v128_shl_16(a, c); +} + +SIMD_INLINE v128 v128_shr_n_u16(v128 a, unsigned int c) { + return v128_shr_u16(a, c); +} + +SIMD_INLINE v128 v128_shr_n_s16(v128 a, unsigned int c) { + return v128_shr_s16(a, c); +} + +SIMD_INLINE v128 v128_shl_n_32(v128 a, unsigned int c) { + return v128_shl_32(a, c); +} + +SIMD_INLINE v128 v128_shr_n_u32(v128 a, unsigned int c) { + return v128_shr_u32(a, c); +} + +SIMD_INLINE v128 v128_shr_n_s32(v128 a, unsigned int c) { + return v128_shr_s32(a, c); +} + +SIMD_INLINE v128 v128_shl_n_64(v128 a, unsigned int c) { + return v128_shl_64(a, c); +} + +SIMD_INLINE v128 v128_shr_n_u64(v128 a, unsigned int c) { + return v128_shr_u64(a, c); +} + +SIMD_INLINE v128 v128_shr_n_s64(v128 a, unsigned int c) { + return v128_shr_s64(a, c); +} + +#endif + +typedef uint32x4_t sad128_internal_u16; + +SIMD_INLINE sad128_internal_u16 v128_sad_u16_init() { return vdupq_n_u32(0); } + +/* Implementation dependent return value. Result must be finalised with + * v128_sad_u16_sum(). */ +SIMD_INLINE sad128_internal_u16 v128_sad_u16(sad128_internal_u16 s, v128 a, + v128 b) { + return vaddq_u32( + s, vpaddlq_u16(vsubq_u16( + vmaxq_u16(vreinterpretq_u16_s64(a), vreinterpretq_u16_s64(b)), + vminq_u16(vreinterpretq_u16_s64(a), vreinterpretq_u16_s64(b))))); +} + +SIMD_INLINE uint32_t v128_sad_u16_sum(sad128_internal_u16 s) { + uint64x2_t t = vpaddlq_u32(s); + return (uint32_t)(uint64_t)vget_high_u64(t) + + (uint32_t)(uint64_t)vget_low_u64(t); +} + +typedef v128 ssd128_internal_s16; +SIMD_INLINE ssd128_internal_s16 v128_ssd_s16_init() { return v128_zero(); } + +/* Implementation dependent return value. Result must be finalised with + * v128_ssd_s16_sum(). */ +SIMD_INLINE ssd128_internal_s16 v128_ssd_s16(ssd128_internal_s16 s, v128 a, + v128 b) { + v128 d = v128_sub_16(a, b); + d = v128_madd_s16(d, d); + return v128_add_64( + s, vreinterpretq_s64_u64(vpaddlq_u32(vreinterpretq_u32_s64(d)))); +} + +SIMD_INLINE uint64_t v128_ssd_s16_sum(ssd128_internal_s16 s) { + return v64_u64(v128_low_v64(s)) + v64_u64(v128_high_v64(s)); +} + +#endif // AOM_AOM_DSP_SIMD_V128_INTRINSICS_ARM_H_ diff --git a/third_party/aom/aom_dsp/simd/v128_intrinsics_c.h b/third_party/aom/aom_dsp/simd/v128_intrinsics_c.h new file mode 100644 index 0000000000..bbe9a9d281 --- /dev/null +++ b/third_party/aom/aom_dsp/simd/v128_intrinsics_c.h @@ -0,0 +1,888 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_SIMD_V128_INTRINSICS_C_H_ +#define AOM_AOM_DSP_SIMD_V128_INTRINSICS_C_H_ + +#include +#include + +#include "config/aom_config.h" + +#include "aom_dsp/simd/v64_intrinsics_c.h" + +typedef union { + uint8_t u8[16]; + uint16_t u16[8]; + uint32_t u32[4]; + uint64_t u64[2]; + int8_t s8[16]; + int16_t s16[8]; + int32_t s32[4]; + int64_t s64[2]; + c_v64 v64[2]; +} c_v128; + +SIMD_INLINE uint32_t c_v128_low_u32(c_v128 a) { return a.u32[0]; } + +SIMD_INLINE c_v64 c_v128_low_v64(c_v128 a) { return a.v64[0]; } + +SIMD_INLINE c_v64 c_v128_high_v64(c_v128 a) { return a.v64[1]; } + +SIMD_INLINE c_v128 c_v128_from_64(uint64_t hi, uint64_t lo) { + c_v128 t; + t.u64[1] = hi; + t.u64[0] = lo; + return t; +} + +SIMD_INLINE c_v128 c_v128_from_v64(c_v64 hi, c_v64 lo) { + c_v128 t; + t.v64[1] = hi; + t.v64[0] = lo; + return t; +} + +SIMD_INLINE c_v128 c_v128_from_32(uint32_t a, uint32_t b, uint32_t c, + uint32_t d) { + c_v128 t; + t.u32[3] = a; + t.u32[2] = b; + t.u32[1] = c; + t.u32[0] = d; + return t; +} + +SIMD_INLINE c_v128 c_v128_load_unaligned(const void *p) { + c_v128 t; + uint8_t *pp = (uint8_t *)p; + uint8_t *q = (uint8_t *)&t; + int c; + for (c = 0; c < 16; c++) q[c] = pp[c]; + return t; +} + +SIMD_INLINE c_v128 c_v128_load_aligned(const void *p) { + if (SIMD_CHECK && (uintptr_t)p & 15) { + fprintf(stderr, "Error: unaligned v128 load at %p\n", p); + abort(); + } + return c_v128_load_unaligned(p); +} + +SIMD_INLINE void c_v128_store_unaligned(void *p, c_v128 a) { + uint8_t *pp = (uint8_t *)p; + uint8_t *q = (uint8_t *)&a; + int c; + for (c = 0; c < 16; c++) pp[c] = q[c]; +} + +SIMD_INLINE void c_v128_store_aligned(void *p, c_v128 a) { + if (SIMD_CHECK && (uintptr_t)p & 15) { + fprintf(stderr, "Error: unaligned v128 store at %p\n", p); + abort(); + } + c_v128_store_unaligned(p, a); +} + +SIMD_INLINE c_v128 c_v128_zero() { + c_v128 t; + t.u64[1] = t.u64[0] = 0; + return t; +} + +SIMD_INLINE c_v128 c_v128_dup_8(uint8_t x) { + c_v128 t; + t.v64[1] = t.v64[0] = c_v64_dup_8(x); + return t; +} + +SIMD_INLINE c_v128 c_v128_dup_16(uint16_t x) { + c_v128 t; + t.v64[1] = t.v64[0] = c_v64_dup_16(x); + return t; +} + +SIMD_INLINE c_v128 c_v128_dup_32(uint32_t x) { + c_v128 t; + t.v64[1] = t.v64[0] = c_v64_dup_32(x); + return t; +} + +SIMD_INLINE c_v128 c_v128_dup_64(uint64_t x) { + c_v128 t; + t.u64[1] = t.u64[0] = x; + return t; +} + +SIMD_INLINE int64_t c_v128_dotp_su8(c_v128 a, c_v128 b) { + return c_v64_dotp_su8(a.v64[1], b.v64[1]) + + c_v64_dotp_su8(a.v64[0], b.v64[0]); +} + +SIMD_INLINE int64_t c_v128_dotp_s16(c_v128 a, c_v128 b) { + return c_v64_dotp_s16(a.v64[1], b.v64[1]) + + c_v64_dotp_s16(a.v64[0], b.v64[0]); +} + +SIMD_INLINE int64_t c_v128_dotp_s32(c_v128 a, c_v128 b) { + // 32 bit products, 64 bit sum + return (int64_t)(int32_t)((int64_t)a.s32[3] * b.s32[3]) + + (int64_t)(int32_t)((int64_t)a.s32[2] * b.s32[2]) + + (int64_t)(int32_t)((int64_t)a.s32[1] * b.s32[1]) + + (int64_t)(int32_t)((int64_t)a.s32[0] * b.s32[0]); +} + +SIMD_INLINE uint64_t c_v128_hadd_u8(c_v128 a) { + return c_v64_hadd_u8(a.v64[1]) + c_v64_hadd_u8(a.v64[0]); +} + +typedef uint32_t c_sad128_internal; + +SIMD_INLINE c_sad128_internal c_v128_sad_u8_init() { return 0; } + +/* Implementation dependent return value. Result must be finalised with + v128_sad_u8_sum(). + The result for more than 32 v128_sad_u8() calls is undefined. */ +SIMD_INLINE c_sad128_internal c_v128_sad_u8(c_sad128_internal s, c_v128 a, + c_v128 b) { + int c; + for (c = 0; c < 16; c++) + s += a.u8[c] > b.u8[c] ? a.u8[c] - b.u8[c] : b.u8[c] - a.u8[c]; + return s; +} + +SIMD_INLINE uint32_t c_v128_sad_u8_sum(c_sad128_internal s) { return s; } + +typedef uint32_t c_ssd128_internal; + +SIMD_INLINE c_ssd128_internal c_v128_ssd_u8_init() { return 0; } + +/* Implementation dependent return value. Result must be finalised with + * v128_ssd_u8_sum(). */ +SIMD_INLINE c_ssd128_internal c_v128_ssd_u8(c_ssd128_internal s, c_v128 a, + c_v128 b) { + int c; + for (c = 0; c < 16; c++) s += (a.u8[c] - b.u8[c]) * (a.u8[c] - b.u8[c]); + return s; +} + +SIMD_INLINE uint32_t c_v128_ssd_u8_sum(c_ssd128_internal s) { return s; } + +SIMD_INLINE c_v128 c_v128_or(c_v128 a, c_v128 b) { + return c_v128_from_v64(c_v64_or(a.v64[1], b.v64[1]), + c_v64_or(a.v64[0], b.v64[0])); +} + +SIMD_INLINE c_v128 c_v128_xor(c_v128 a, c_v128 b) { + return c_v128_from_v64(c_v64_xor(a.v64[1], b.v64[1]), + c_v64_xor(a.v64[0], b.v64[0])); +} + +SIMD_INLINE c_v128 c_v128_and(c_v128 a, c_v128 b) { + return c_v128_from_v64(c_v64_and(a.v64[1], b.v64[1]), + c_v64_and(a.v64[0], b.v64[0])); +} + +SIMD_INLINE c_v128 c_v128_andn(c_v128 a, c_v128 b) { + return c_v128_from_v64(c_v64_andn(a.v64[1], b.v64[1]), + c_v64_andn(a.v64[0], b.v64[0])); +} + +SIMD_INLINE c_v128 c_v128_add_8(c_v128 a, c_v128 b) { + return c_v128_from_v64(c_v64_add_8(a.v64[1], b.v64[1]), + c_v64_add_8(a.v64[0], b.v64[0])); +} + +SIMD_INLINE c_v128 c_v128_add_16(c_v128 a, c_v128 b) { + return c_v128_from_v64(c_v64_add_16(a.v64[1], b.v64[1]), + c_v64_add_16(a.v64[0], b.v64[0])); +} + +SIMD_INLINE c_v128 c_v128_sadd_u8(c_v128 a, c_v128 b) { + return c_v128_from_v64(c_v64_sadd_u8(a.v64[1], b.v64[1]), + c_v64_sadd_u8(a.v64[0], b.v64[0])); +} + +SIMD_INLINE c_v128 c_v128_sadd_s8(c_v128 a, c_v128 b) { + return c_v128_from_v64(c_v64_sadd_s8(a.v64[1], b.v64[1]), + c_v64_sadd_s8(a.v64[0], b.v64[0])); +} + +SIMD_INLINE c_v128 c_v128_sadd_s16(c_v128 a, c_v128 b) { + return c_v128_from_v64(c_v64_sadd_s16(a.v64[1], b.v64[1]), + c_v64_sadd_s16(a.v64[0], b.v64[0])); +} + +SIMD_INLINE c_v128 c_v128_add_32(c_v128 a, c_v128 b) { + return c_v128_from_v64(c_v64_add_32(a.v64[1], b.v64[1]), + c_v64_add_32(a.v64[0], b.v64[0])); +} + +SIMD_INLINE c_v128 c_v128_add_64(c_v128 a, c_v128 b) { + // Two complement overflow (silences sanitizers) + return c_v128_from_64( + a.v64[1].u64 > ~b.v64[1].u64 ? a.v64[1].u64 - ~b.v64[1].u64 - 1 + : a.v64[1].u64 + b.v64[1].u64, + a.v64[0].u64 > ~b.v64[0].u64 ? a.v64[0].u64 - ~b.v64[0].u64 - 1 + : a.v64[0].u64 + b.v64[0].u64); +} + +SIMD_INLINE c_v128 c_v128_padd_s16(c_v128 a) { + c_v128 t; + t.s32[0] = (int32_t)a.s16[0] + (int32_t)a.s16[1]; + t.s32[1] = (int32_t)a.s16[2] + (int32_t)a.s16[3]; + t.s32[2] = (int32_t)a.s16[4] + (int32_t)a.s16[5]; + t.s32[3] = (int32_t)a.s16[6] + (int32_t)a.s16[7]; + return t; +} + +SIMD_INLINE c_v128 c_v128_padd_u8(c_v128 a) { + c_v128 t; + t.u16[0] = (uint16_t)a.u8[0] + (uint16_t)a.u8[1]; + t.u16[1] = (uint16_t)a.u8[2] + (uint16_t)a.u8[3]; + t.u16[2] = (uint16_t)a.u8[4] + (uint16_t)a.u8[5]; + t.u16[3] = (uint16_t)a.u8[6] + (uint16_t)a.u8[7]; + t.u16[4] = (uint16_t)a.u8[8] + (uint16_t)a.u8[9]; + t.u16[5] = (uint16_t)a.u8[10] + (uint16_t)a.u8[11]; + t.u16[6] = (uint16_t)a.u8[12] + (uint16_t)a.u8[13]; + t.u16[7] = (uint16_t)a.u8[14] + (uint16_t)a.u8[15]; + return t; +} + +SIMD_INLINE c_v128 c_v128_sub_8(c_v128 a, c_v128 b) { + return c_v128_from_v64(c_v64_sub_8(a.v64[1], b.v64[1]), + c_v64_sub_8(a.v64[0], b.v64[0])); +} + +SIMD_INLINE c_v128 c_v128_ssub_u8(c_v128 a, c_v128 b) { + return c_v128_from_v64(c_v64_ssub_u8(a.v64[1], b.v64[1]), + c_v64_ssub_u8(a.v64[0], b.v64[0])); +} + +SIMD_INLINE c_v128 c_v128_ssub_s8(c_v128 a, c_v128 b) { + return c_v128_from_v64(c_v64_ssub_s8(a.v64[1], b.v64[1]), + c_v64_ssub_s8(a.v64[0], b.v64[0])); +} + +SIMD_INLINE c_v128 c_v128_sub_16(c_v128 a, c_v128 b) { + return c_v128_from_v64(c_v64_sub_16(a.v64[1], b.v64[1]), + c_v64_sub_16(a.v64[0], b.v64[0])); +} + +SIMD_INLINE c_v128 c_v128_ssub_s16(c_v128 a, c_v128 b) { + return c_v128_from_v64(c_v64_ssub_s16(a.v64[1], b.v64[1]), + c_v64_ssub_s16(a.v64[0], b.v64[0])); +} + +SIMD_INLINE c_v128 c_v128_ssub_u16(c_v128 a, c_v128 b) { + return c_v128_from_v64(c_v64_ssub_u16(a.v64[1], b.v64[1]), + c_v64_ssub_u16(a.v64[0], b.v64[0])); +} + +SIMD_INLINE c_v128 c_v128_sub_32(c_v128 a, c_v128 b) { + return c_v128_from_v64(c_v64_sub_32(a.v64[1], b.v64[1]), + c_v64_sub_32(a.v64[0], b.v64[0])); +} + +SIMD_INLINE c_v128 c_v128_sub_64(c_v128 a, c_v128 b) { + // Two complement underflow (silences sanitizers) + return c_v128_from_64( + a.v64[1].u64 < b.v64[1].u64 ? a.v64[1].u64 + ~b.v64[1].u64 + 1 + : a.v64[1].u64 - b.v64[1].u64, + a.v64[0].u64 < b.v64[0].u64 ? a.v64[0].u64 + ~b.v64[0].u64 + 1 + : a.v64[0].u64 - b.v64[0].u64); +} + +SIMD_INLINE c_v128 c_v128_abs_s16(c_v128 a) { + return c_v128_from_v64(c_v64_abs_s16(a.v64[1]), c_v64_abs_s16(a.v64[0])); +} + +SIMD_INLINE c_v128 c_v128_abs_s8(c_v128 a) { + return c_v128_from_v64(c_v64_abs_s8(a.v64[1]), c_v64_abs_s8(a.v64[0])); +} + +SIMD_INLINE c_v128 c_v128_mul_s16(c_v64 a, c_v64 b) { + c_v64 lo_bits = c_v64_mullo_s16(a, b); + c_v64 hi_bits = c_v64_mulhi_s16(a, b); + return c_v128_from_v64(c_v64_ziphi_16(hi_bits, lo_bits), + c_v64_ziplo_16(hi_bits, lo_bits)); +} + +SIMD_INLINE c_v128 c_v128_mullo_s16(c_v128 a, c_v128 b) { + return c_v128_from_v64(c_v64_mullo_s16(a.v64[1], b.v64[1]), + c_v64_mullo_s16(a.v64[0], b.v64[0])); +} + +SIMD_INLINE c_v128 c_v128_mulhi_s16(c_v128 a, c_v128 b) { + return c_v128_from_v64(c_v64_mulhi_s16(a.v64[1], b.v64[1]), + c_v64_mulhi_s16(a.v64[0], b.v64[0])); +} + +SIMD_INLINE c_v128 c_v128_mullo_s32(c_v128 a, c_v128 b) { + return c_v128_from_v64(c_v64_mullo_s32(a.v64[1], b.v64[1]), + c_v64_mullo_s32(a.v64[0], b.v64[0])); +} + +SIMD_INLINE c_v128 c_v128_madd_s16(c_v128 a, c_v128 b) { + return c_v128_from_v64(c_v64_madd_s16(a.v64[1], b.v64[1]), + c_v64_madd_s16(a.v64[0], b.v64[0])); +} + +SIMD_INLINE c_v128 c_v128_madd_us8(c_v128 a, c_v128 b) { + return c_v128_from_v64(c_v64_madd_us8(a.v64[1], b.v64[1]), + c_v64_madd_us8(a.v64[0], b.v64[0])); +} + +SIMD_INLINE c_v128 c_v128_avg_u8(c_v128 a, c_v128 b) { + return c_v128_from_v64(c_v64_avg_u8(a.v64[1], b.v64[1]), + c_v64_avg_u8(a.v64[0], b.v64[0])); +} + +SIMD_INLINE c_v128 c_v128_rdavg_u8(c_v128 a, c_v128 b) { + return c_v128_from_v64(c_v64_rdavg_u8(a.v64[1], b.v64[1]), + c_v64_rdavg_u8(a.v64[0], b.v64[0])); +} + +SIMD_INLINE c_v128 c_v128_rdavg_u16(c_v128 a, c_v128 b) { + return c_v128_from_v64(c_v64_rdavg_u16(a.v64[1], b.v64[1]), + c_v64_rdavg_u16(a.v64[0], b.v64[0])); +} + +SIMD_INLINE c_v128 c_v128_avg_u16(c_v128 a, c_v128 b) { + return c_v128_from_v64(c_v64_avg_u16(a.v64[1], b.v64[1]), + c_v64_avg_u16(a.v64[0], b.v64[0])); +} + +SIMD_INLINE c_v128 c_v128_min_u8(c_v128 a, c_v128 b) { + return c_v128_from_v64(c_v64_min_u8(a.v64[1], b.v64[1]), + c_v64_min_u8(a.v64[0], b.v64[0])); +} + +SIMD_INLINE c_v128 c_v128_max_u8(c_v128 a, c_v128 b) { + return c_v128_from_v64(c_v64_max_u8(a.v64[1], b.v64[1]), + c_v64_max_u8(a.v64[0], b.v64[0])); +} + +SIMD_INLINE c_v128 c_v128_min_s8(c_v128 a, c_v128 b) { + return c_v128_from_v64(c_v64_min_s8(a.v64[1], b.v64[1]), + c_v64_min_s8(a.v64[0], b.v64[0])); +} + +SIMD_INLINE uint32_t c_v128_movemask_8(c_v128 a) { + return ((a.s8[15] < 0) << 15) | ((a.s8[14] < 0) << 14) | + ((a.s8[13] < 0) << 13) | ((a.s8[12] < 0) << 12) | + ((a.s8[11] < 0) << 11) | ((a.s8[10] < 0) << 10) | + ((a.s8[9] < 0) << 9) | ((a.s8[8] < 0) << 8) | ((a.s8[7] < 0) << 7) | + ((a.s8[6] < 0) << 6) | ((a.s8[5] < 0) << 5) | ((a.s8[4] < 0) << 4) | + ((a.s8[3] < 0) << 3) | ((a.s8[2] < 0) << 2) | ((a.s8[1] < 0) << 1) | + ((a.s8[0] < 0) << 0); +} + +SIMD_INLINE c_v128 c_v128_blend_8(c_v128 a, c_v128 b, c_v128 c) { + c_v128 t; + for (int i = 0; i < 16; i++) t.u8[i] = c.s8[i] < 0 ? b.u8[i] : a.u8[i]; + return t; +} + +SIMD_INLINE c_v128 c_v128_max_s8(c_v128 a, c_v128 b) { + return c_v128_from_v64(c_v64_max_s8(a.v64[1], b.v64[1]), + c_v64_max_s8(a.v64[0], b.v64[0])); +} + +SIMD_INLINE c_v128 c_v128_min_s16(c_v128 a, c_v128 b) { + return c_v128_from_v64(c_v64_min_s16(a.v64[1], b.v64[1]), + c_v64_min_s16(a.v64[0], b.v64[0])); +} + +SIMD_INLINE c_v128 c_v128_max_s16(c_v128 a, c_v128 b) { + return c_v128_from_v64(c_v64_max_s16(a.v64[1], b.v64[1]), + c_v64_max_s16(a.v64[0], b.v64[0])); +} + +SIMD_INLINE c_v128 c_v128_max_s32(c_v128 a, c_v128 b) { + c_v128 t; + int c; + for (c = 0; c < 4; c++) t.s32[c] = a.s32[c] > b.s32[c] ? a.s32[c] : b.s32[c]; + return t; +} + +SIMD_INLINE c_v128 c_v128_min_s32(c_v128 a, c_v128 b) { + c_v128 t; + int c; + for (c = 0; c < 4; c++) t.s32[c] = a.s32[c] > b.s32[c] ? b.s32[c] : a.s32[c]; + return t; +} + +SIMD_INLINE c_v128 c_v128_ziplo_8(c_v128 a, c_v128 b) { + return c_v128_from_v64(c_v64_ziphi_8(a.v64[0], b.v64[0]), + c_v64_ziplo_8(a.v64[0], b.v64[0])); +} + +SIMD_INLINE c_v128 c_v128_ziphi_8(c_v128 a, c_v128 b) { + return c_v128_from_v64(c_v64_ziphi_8(a.v64[1], b.v64[1]), + c_v64_ziplo_8(a.v64[1], b.v64[1])); +} + +SIMD_INLINE c_v128 c_v128_ziplo_16(c_v128 a, c_v128 b) { + return c_v128_from_v64(c_v64_ziphi_16(a.v64[0], b.v64[0]), + c_v64_ziplo_16(a.v64[0], b.v64[0])); +} + +SIMD_INLINE c_v128 c_v128_ziphi_16(c_v128 a, c_v128 b) { + return c_v128_from_v64(c_v64_ziphi_16(a.v64[1], b.v64[1]), + c_v64_ziplo_16(a.v64[1], b.v64[1])); +} + +SIMD_INLINE c_v128 c_v128_ziplo_32(c_v128 a, c_v128 b) { + return c_v128_from_v64(c_v64_ziphi_32(a.v64[0], b.v64[0]), + c_v64_ziplo_32(a.v64[0], b.v64[0])); +} + +SIMD_INLINE c_v128 c_v128_ziphi_32(c_v128 a, c_v128 b) { + return c_v128_from_v64(c_v64_ziphi_32(a.v64[1], b.v64[1]), + c_v64_ziplo_32(a.v64[1], b.v64[1])); +} + +SIMD_INLINE c_v128 c_v128_ziplo_64(c_v128 a, c_v128 b) { + return c_v128_from_v64(a.v64[0], b.v64[0]); +} + +SIMD_INLINE c_v128 c_v128_ziphi_64(c_v128 a, c_v128 b) { + return c_v128_from_v64(a.v64[1], b.v64[1]); +} + +SIMD_INLINE c_v128 c_v128_zip_8(c_v64 a, c_v64 b) { + return c_v128_from_v64(c_v64_ziphi_8(a, b), c_v64_ziplo_8(a, b)); +} + +SIMD_INLINE c_v128 c_v128_zip_16(c_v64 a, c_v64 b) { + return c_v128_from_v64(c_v64_ziphi_16(a, b), c_v64_ziplo_16(a, b)); +} + +SIMD_INLINE c_v128 c_v128_zip_32(c_v64 a, c_v64 b) { + return c_v128_from_v64(c_v64_ziphi_32(a, b), c_v64_ziplo_32(a, b)); +} + +SIMD_INLINE c_v128 _c_v128_unzip_8(c_v128 a, c_v128 b, int mode) { + c_v128 t; + if (mode) { + t.u8[15] = b.u8[15]; + t.u8[14] = b.u8[13]; + t.u8[13] = b.u8[11]; + t.u8[12] = b.u8[9]; + t.u8[11] = b.u8[7]; + t.u8[10] = b.u8[5]; + t.u8[9] = b.u8[3]; + t.u8[8] = b.u8[1]; + t.u8[7] = a.u8[15]; + t.u8[6] = a.u8[13]; + t.u8[5] = a.u8[11]; + t.u8[4] = a.u8[9]; + t.u8[3] = a.u8[7]; + t.u8[2] = a.u8[5]; + t.u8[1] = a.u8[3]; + t.u8[0] = a.u8[1]; + } else { + t.u8[15] = a.u8[14]; + t.u8[14] = a.u8[12]; + t.u8[13] = a.u8[10]; + t.u8[12] = a.u8[8]; + t.u8[11] = a.u8[6]; + t.u8[10] = a.u8[4]; + t.u8[9] = a.u8[2]; + t.u8[8] = a.u8[0]; + t.u8[7] = b.u8[14]; + t.u8[6] = b.u8[12]; + t.u8[5] = b.u8[10]; + t.u8[4] = b.u8[8]; + t.u8[3] = b.u8[6]; + t.u8[2] = b.u8[4]; + t.u8[1] = b.u8[2]; + t.u8[0] = b.u8[0]; + } + return t; +} + +SIMD_INLINE c_v128 c_v128_unziplo_8(c_v128 a, c_v128 b) { + return CONFIG_BIG_ENDIAN ? _c_v128_unzip_8(a, b, 1) + : _c_v128_unzip_8(a, b, 0); +} + +SIMD_INLINE c_v128 c_v128_unziphi_8(c_v128 a, c_v128 b) { + return CONFIG_BIG_ENDIAN ? _c_v128_unzip_8(b, a, 0) + : _c_v128_unzip_8(b, a, 1); +} + +SIMD_INLINE c_v128 _c_v128_unzip_16(c_v128 a, c_v128 b, int mode) { + c_v128 t; + if (mode) { + t.u16[7] = b.u16[7]; + t.u16[6] = b.u16[5]; + t.u16[5] = b.u16[3]; + t.u16[4] = b.u16[1]; + t.u16[3] = a.u16[7]; + t.u16[2] = a.u16[5]; + t.u16[1] = a.u16[3]; + t.u16[0] = a.u16[1]; + } else { + t.u16[7] = a.u16[6]; + t.u16[6] = a.u16[4]; + t.u16[5] = a.u16[2]; + t.u16[4] = a.u16[0]; + t.u16[3] = b.u16[6]; + t.u16[2] = b.u16[4]; + t.u16[1] = b.u16[2]; + t.u16[0] = b.u16[0]; + } + return t; +} + +SIMD_INLINE c_v128 c_v128_unziplo_16(c_v128 a, c_v128 b) { + return CONFIG_BIG_ENDIAN ? _c_v128_unzip_16(a, b, 1) + : _c_v128_unzip_16(a, b, 0); +} + +SIMD_INLINE c_v128 c_v128_unziphi_16(c_v128 a, c_v128 b) { + return CONFIG_BIG_ENDIAN ? _c_v128_unzip_16(b, a, 0) + : _c_v128_unzip_16(b, a, 1); +} + +SIMD_INLINE c_v128 _c_v128_unzip_32(c_v128 a, c_v128 b, int mode) { + c_v128 t; + if (mode) { + t.u32[3] = b.u32[3]; + t.u32[2] = b.u32[1]; + t.u32[1] = a.u32[3]; + t.u32[0] = a.u32[1]; + } else { + t.u32[3] = a.u32[2]; + t.u32[2] = a.u32[0]; + t.u32[1] = b.u32[2]; + t.u32[0] = b.u32[0]; + } + return t; +} + +SIMD_INLINE c_v128 c_v128_unziplo_32(c_v128 a, c_v128 b) { + return CONFIG_BIG_ENDIAN ? _c_v128_unzip_32(a, b, 1) + : _c_v128_unzip_32(a, b, 0); +} + +SIMD_INLINE c_v128 c_v128_unziphi_32(c_v128 a, c_v128 b) { + return CONFIG_BIG_ENDIAN ? _c_v128_unzip_32(b, a, 0) + : _c_v128_unzip_32(b, a, 1); +} + +SIMD_INLINE c_v128 c_v128_unpack_u8_s16(c_v64 a) { + return c_v128_from_v64(c_v64_unpackhi_u8_s16(a), c_v64_unpacklo_u8_s16(a)); +} + +SIMD_INLINE c_v128 c_v128_unpacklo_u8_s16(c_v128 a) { + return c_v128_from_v64(c_v64_unpackhi_u8_s16(a.v64[0]), + c_v64_unpacklo_u8_s16(a.v64[0])); +} + +SIMD_INLINE c_v128 c_v128_unpackhi_u8_s16(c_v128 a) { + return c_v128_from_v64(c_v64_unpackhi_u8_s16(a.v64[1]), + c_v64_unpacklo_u8_s16(a.v64[1])); +} + +SIMD_INLINE c_v128 c_v128_unpack_s8_s16(c_v64 a) { + return c_v128_from_v64(c_v64_unpackhi_s8_s16(a), c_v64_unpacklo_s8_s16(a)); +} + +SIMD_INLINE c_v128 c_v128_unpacklo_s8_s16(c_v128 a) { + return c_v128_from_v64(c_v64_unpackhi_s8_s16(a.v64[0]), + c_v64_unpacklo_s8_s16(a.v64[0])); +} + +SIMD_INLINE c_v128 c_v128_unpackhi_s8_s16(c_v128 a) { + return c_v128_from_v64(c_v64_unpackhi_s8_s16(a.v64[1]), + c_v64_unpacklo_s8_s16(a.v64[1])); +} + +SIMD_INLINE c_v128 c_v128_pack_s32_s16(c_v128 a, c_v128 b) { + return c_v128_from_v64(c_v64_pack_s32_s16(a.v64[1], a.v64[0]), + c_v64_pack_s32_s16(b.v64[1], b.v64[0])); +} + +SIMD_INLINE c_v128 c_v128_pack_s32_u16(c_v128 a, c_v128 b) { + return c_v128_from_v64(c_v64_pack_s32_u16(a.v64[1], a.v64[0]), + c_v64_pack_s32_u16(b.v64[1], b.v64[0])); +} + +SIMD_INLINE c_v128 c_v128_pack_s16_u8(c_v128 a, c_v128 b) { + return c_v128_from_v64(c_v64_pack_s16_u8(a.v64[1], a.v64[0]), + c_v64_pack_s16_u8(b.v64[1], b.v64[0])); +} + +SIMD_INLINE c_v128 c_v128_pack_s16_s8(c_v128 a, c_v128 b) { + return c_v128_from_v64(c_v64_pack_s16_s8(a.v64[1], a.v64[0]), + c_v64_pack_s16_s8(b.v64[1], b.v64[0])); +} + +SIMD_INLINE c_v128 c_v128_unpack_u16_s32(c_v64 a) { + return c_v128_from_v64(c_v64_unpackhi_u16_s32(a), c_v64_unpacklo_u16_s32(a)); +} + +SIMD_INLINE c_v128 c_v128_unpack_s16_s32(c_v64 a) { + return c_v128_from_v64(c_v64_unpackhi_s16_s32(a), c_v64_unpacklo_s16_s32(a)); +} + +SIMD_INLINE c_v128 c_v128_unpacklo_u16_s32(c_v128 a) { + return c_v128_from_v64(c_v64_unpackhi_u16_s32(a.v64[0]), + c_v64_unpacklo_u16_s32(a.v64[0])); +} + +SIMD_INLINE c_v128 c_v128_unpacklo_s16_s32(c_v128 a) { + return c_v128_from_v64(c_v64_unpackhi_s16_s32(a.v64[0]), + c_v64_unpacklo_s16_s32(a.v64[0])); +} + +SIMD_INLINE c_v128 c_v128_unpackhi_u16_s32(c_v128 a) { + return c_v128_from_v64(c_v64_unpackhi_u16_s32(a.v64[1]), + c_v64_unpacklo_u16_s32(a.v64[1])); +} + +SIMD_INLINE c_v128 c_v128_unpackhi_s16_s32(c_v128 a) { + return c_v128_from_v64(c_v64_unpackhi_s16_s32(a.v64[1]), + c_v64_unpacklo_s16_s32(a.v64[1])); +} + +SIMD_INLINE c_v128 c_v128_shuffle_8(c_v128 a, c_v128 pattern) { + c_v128 t; + int c; + for (c = 0; c < 16; c++) + t.u8[c] = a.u8[CONFIG_BIG_ENDIAN ? 15 - (pattern.u8[c] & 15) + : pattern.u8[c] & 15]; + + return t; +} + +SIMD_INLINE c_v128 c_v128_cmpgt_s8(c_v128 a, c_v128 b) { + return c_v128_from_v64(c_v64_cmpgt_s8(a.v64[1], b.v64[1]), + c_v64_cmpgt_s8(a.v64[0], b.v64[0])); +} + +SIMD_INLINE c_v128 c_v128_cmplt_s8(c_v128 a, c_v128 b) { + return c_v128_from_v64(c_v64_cmplt_s8(a.v64[1], b.v64[1]), + c_v64_cmplt_s8(a.v64[0], b.v64[0])); +} + +SIMD_INLINE c_v128 c_v128_cmpeq_8(c_v128 a, c_v128 b) { + return c_v128_from_v64(c_v64_cmpeq_8(a.v64[1], b.v64[1]), + c_v64_cmpeq_8(a.v64[0], b.v64[0])); +} + +SIMD_INLINE c_v128 c_v128_cmpgt_s16(c_v128 a, c_v128 b) { + return c_v128_from_v64(c_v64_cmpgt_s16(a.v64[1], b.v64[1]), + c_v64_cmpgt_s16(a.v64[0], b.v64[0])); +} + +SIMD_INLINE c_v128 c_v128_cmplt_s16(c_v128 a, c_v128 b) { + return c_v128_from_v64(c_v64_cmplt_s16(a.v64[1], b.v64[1]), + c_v64_cmplt_s16(a.v64[0], b.v64[0])); +} + +SIMD_INLINE c_v128 c_v128_cmpeq_16(c_v128 a, c_v128 b) { + return c_v128_from_v64(c_v64_cmpeq_16(a.v64[1], b.v64[1]), + c_v64_cmpeq_16(a.v64[0], b.v64[0])); +} + +SIMD_INLINE c_v128 c_v128_cmpgt_s32(c_v128 a, c_v128 b) { + c_v128 t; + int c; + for (c = 0; c < 4; c++) t.s32[c] = -(a.s32[c] > b.s32[c]); + return t; +} + +SIMD_INLINE c_v128 c_v128_cmplt_s32(c_v128 a, c_v128 b) { + c_v128 t; + int c; + for (c = 0; c < 4; c++) t.s32[c] = -(a.s32[c] < b.s32[c]); + return t; +} + +SIMD_INLINE c_v128 c_v128_cmpeq_32(c_v128 a, c_v128 b) { + c_v128 t; + int c; + for (c = 0; c < 4; c++) t.s32[c] = -(a.s32[c] == b.s32[c]); + return t; +} + +SIMD_INLINE c_v128 c_v128_shl_n_byte(c_v128 a, const unsigned int n) { + if (n < 8) + return c_v128_from_v64(c_v64_or(c_v64_shl_n_byte(a.v64[1], n), + c_v64_shr_n_byte(a.v64[0], 8 - n)), + c_v64_shl_n_byte(a.v64[0], n)); + else + return c_v128_from_v64(c_v64_shl_n_byte(a.v64[0], n - 8), c_v64_zero()); +} + +SIMD_INLINE c_v128 c_v128_shr_n_byte(c_v128 a, const unsigned int n) { + if (n < 8) + return c_v128_from_v64(c_v64_shr_n_byte(a.v64[1], n), + c_v64_or(c_v64_shr_n_byte(a.v64[0], n), + c_v64_shl_n_byte(a.v64[1], 8 - n))); + else + return c_v128_from_v64(c_v64_zero(), c_v64_shr_n_byte(a.v64[1], n - 8)); +} + +SIMD_INLINE c_v128 c_v128_align(c_v128 a, c_v128 b, const unsigned int c) { + if (SIMD_CHECK && c > 15) { + fprintf(stderr, "Error: undefined alignment %d\n", c); + abort(); + } + return c ? c_v128_or(c_v128_shr_n_byte(b, c), c_v128_shl_n_byte(a, 16 - c)) + : b; +} + +SIMD_INLINE c_v128 c_v128_shl_8(c_v128 a, const unsigned int c) { + return c_v128_from_v64(c_v64_shl_8(a.v64[1], c), c_v64_shl_8(a.v64[0], c)); +} + +SIMD_INLINE c_v128 c_v128_shr_u8(c_v128 a, const unsigned int c) { + return c_v128_from_v64(c_v64_shr_u8(a.v64[1], c), c_v64_shr_u8(a.v64[0], c)); +} + +SIMD_INLINE c_v128 c_v128_shr_s8(c_v128 a, const unsigned int c) { + return c_v128_from_v64(c_v64_shr_s8(a.v64[1], c), c_v64_shr_s8(a.v64[0], c)); +} + +SIMD_INLINE c_v128 c_v128_shl_16(c_v128 a, const unsigned int c) { + return c_v128_from_v64(c_v64_shl_16(a.v64[1], c), c_v64_shl_16(a.v64[0], c)); +} + +SIMD_INLINE c_v128 c_v128_shr_u16(c_v128 a, const unsigned int c) { + return c_v128_from_v64(c_v64_shr_u16(a.v64[1], c), + c_v64_shr_u16(a.v64[0], c)); +} + +SIMD_INLINE c_v128 c_v128_shr_s16(c_v128 a, const unsigned int c) { + return c_v128_from_v64(c_v64_shr_s16(a.v64[1], c), + c_v64_shr_s16(a.v64[0], c)); +} + +SIMD_INLINE c_v128 c_v128_shl_32(c_v128 a, const unsigned int c) { + return c_v128_from_v64(c_v64_shl_32(a.v64[1], c), c_v64_shl_32(a.v64[0], c)); +} + +SIMD_INLINE c_v128 c_v128_shr_u32(c_v128 a, const unsigned int c) { + return c_v128_from_v64(c_v64_shr_u32(a.v64[1], c), + c_v64_shr_u32(a.v64[0], c)); +} + +SIMD_INLINE c_v128 c_v128_shr_s32(c_v128 a, const unsigned int c) { + return c_v128_from_v64(c_v64_shr_s32(a.v64[1], c), + c_v64_shr_s32(a.v64[0], c)); +} + +SIMD_INLINE c_v128 c_v128_shl_64(c_v128 a, const unsigned int c) { + a.v64[1].u64 <<= c; + a.v64[0].u64 <<= c; + return c_v128_from_v64(a.v64[1], a.v64[0]); +} + +SIMD_INLINE c_v128 c_v128_shr_u64(c_v128 a, const unsigned int c) { + a.v64[1].u64 >>= c; + a.v64[0].u64 >>= c; + return c_v128_from_v64(a.v64[1], a.v64[0]); +} + +SIMD_INLINE c_v128 c_v128_shr_s64(c_v128 a, const unsigned int c) { + a.v64[1].s64 >>= c; + a.v64[0].s64 >>= c; + return c_v128_from_v64(a.v64[1], a.v64[0]); +} + +SIMD_INLINE c_v128 c_v128_shl_n_8(c_v128 a, const unsigned int n) { + return c_v128_shl_8(a, n); +} + +SIMD_INLINE c_v128 c_v128_shl_n_16(c_v128 a, const unsigned int n) { + return c_v128_shl_16(a, n); +} + +SIMD_INLINE c_v128 c_v128_shl_n_32(c_v128 a, const unsigned int n) { + return c_v128_shl_32(a, n); +} + +SIMD_INLINE c_v128 c_v128_shl_n_64(c_v128 a, const unsigned int n) { + return c_v128_shl_64(a, n); +} + +SIMD_INLINE c_v128 c_v128_shr_n_u8(c_v128 a, const unsigned int n) { + return c_v128_shr_u8(a, n); +} + +SIMD_INLINE c_v128 c_v128_shr_n_u16(c_v128 a, const unsigned int n) { + return c_v128_shr_u16(a, n); +} + +SIMD_INLINE c_v128 c_v128_shr_n_u32(c_v128 a, const unsigned int n) { + return c_v128_shr_u32(a, n); +} + +SIMD_INLINE c_v128 c_v128_shr_n_u64(c_v128 a, const unsigned int n) { + return c_v128_shr_u64(a, n); +} + +SIMD_INLINE c_v128 c_v128_shr_n_s8(c_v128 a, const unsigned int n) { + return c_v128_shr_s8(a, n); +} + +SIMD_INLINE c_v128 c_v128_shr_n_s16(c_v128 a, const unsigned int n) { + return c_v128_shr_s16(a, n); +} + +SIMD_INLINE c_v128 c_v128_shr_n_s32(c_v128 a, const unsigned int n) { + return c_v128_shr_s32(a, n); +} + +SIMD_INLINE c_v128 c_v128_shr_n_s64(c_v128 a, const unsigned int n) { + return c_v128_shr_s64(a, n); +} + +typedef uint32_t c_sad128_internal_u16; + +SIMD_INLINE c_sad128_internal_u16 c_v128_sad_u16_init() { return 0; } + +/* Implementation dependent return value. Result must be finalised with + * v128_sad_u16_sum(). */ +SIMD_INLINE c_sad128_internal_u16 c_v128_sad_u16(c_sad128_internal_u16 s, + c_v128 a, c_v128 b) { + int c; + for (c = 0; c < 8; c++) + s += a.u16[c] > b.u16[c] ? a.u16[c] - b.u16[c] : b.u16[c] - a.u16[c]; + return s; +} + +SIMD_INLINE uint32_t c_v128_sad_u16_sum(c_sad128_internal_u16 s) { return s; } + +typedef uint64_t c_ssd128_internal_s16; + +SIMD_INLINE c_ssd128_internal_s16 c_v128_ssd_s16_init() { return 0; } + +/* Implementation dependent return value. Result must be finalised with + * v128_ssd_s16_sum(). */ +SIMD_INLINE c_ssd128_internal_s16 c_v128_ssd_s16(c_ssd128_internal_s16 s, + c_v128 a, c_v128 b) { + int c; + for (c = 0; c < 8; c++) + s += (int32_t)(int16_t)(a.s16[c] - b.s16[c]) * + (int32_t)(int16_t)(a.s16[c] - b.s16[c]); + return s; +} + +SIMD_INLINE uint64_t c_v128_ssd_s16_sum(c_ssd128_internal_s16 s) { return s; } + +#endif // AOM_AOM_DSP_SIMD_V128_INTRINSICS_C_H_ diff --git a/third_party/aom/aom_dsp/simd/v128_intrinsics_x86.h b/third_party/aom/aom_dsp/simd/v128_intrinsics_x86.h new file mode 100644 index 0000000000..6c7241ff4f --- /dev/null +++ b/third_party/aom/aom_dsp/simd/v128_intrinsics_x86.h @@ -0,0 +1,656 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_SIMD_V128_INTRINSICS_X86_H_ +#define AOM_AOM_DSP_SIMD_V128_INTRINSICS_X86_H_ + +#include +#include "aom_dsp/simd/v64_intrinsics_x86.h" + +typedef __m128i v128; + +SIMD_INLINE uint32_t v128_low_u32(v128 a) { + return (uint32_t)_mm_cvtsi128_si32(a); +} + +SIMD_INLINE v64 v128_low_v64(v128 a) { + return _mm_unpacklo_epi64(a, v64_zero()); +} + +SIMD_INLINE v64 v128_high_v64(v128 a) { return _mm_srli_si128(a, 8); } + +SIMD_INLINE v128 v128_from_v64(v64 a, v64 b) { + return _mm_unpacklo_epi64(b, a); +} + +SIMD_INLINE v128 v128_from_64(uint64_t a, uint64_t b) { + return v128_from_v64(v64_from_64(a), v64_from_64(b)); +} + +SIMD_INLINE v128 v128_from_32(uint32_t a, uint32_t b, uint32_t c, uint32_t d) { + return _mm_set_epi32(a, b, c, d); +} + +SIMD_INLINE v128 v128_load_aligned(const void *p) { + return _mm_load_si128((__m128i *)p); +} + +SIMD_INLINE v128 v128_load_unaligned(const void *p) { +#if defined(__SSSE3__) + return (__m128i)_mm_lddqu_si128((__m128i *)p); +#else + return _mm_loadu_si128((__m128i *)p); +#endif +} + +SIMD_INLINE void v128_store_aligned(void *p, v128 a) { + _mm_store_si128((__m128i *)p, a); +} + +SIMD_INLINE void v128_store_unaligned(void *p, v128 a) { + _mm_storeu_si128((__m128i *)p, a); +} + +// The following function requires an immediate. +// Some compilers will check this during optimisation, others wont. +#if defined(__OPTIMIZE__) && __OPTIMIZE__ && !defined(__clang__) +#if defined(__SSSE3__) +SIMD_INLINE v128 v128_align(v128 a, v128 b, const unsigned int c) { + return c ? _mm_alignr_epi8(a, b, c) : b; +} +#else +#define v128_align(a, b, c) \ + ((c) ? _mm_or_si128(_mm_srli_si128(b, c), _mm_slli_si128(a, 16 - (c))) : (b)) +#endif +#else +#if defined(__SSSE3__) +#define v128_align(a, b, c) ((c) ? _mm_alignr_epi8(a, b, (uint8_t)(c)) : (b)) +#else +#define v128_align(a, b, c) \ + ((c) ? _mm_or_si128(_mm_srli_si128(b, c), _mm_slli_si128(a, 16 - (c))) : (b)) +#endif +#endif + +SIMD_INLINE v128 v128_zero() { return _mm_setzero_si128(); } + +SIMD_INLINE v128 v128_dup_8(uint8_t x) { return _mm_set1_epi8(x); } + +SIMD_INLINE v128 v128_dup_16(uint16_t x) { return _mm_set1_epi16(x); } + +SIMD_INLINE v128 v128_dup_32(uint32_t x) { return _mm_set1_epi32(x); } + +SIMD_INLINE v128 v128_dup_64(uint64_t x) { + // _mm_set_pi64x and _mm_cvtsi64x_si64 missing in some compilers + return _mm_set_epi32(x >> 32, (uint32_t)x, x >> 32, (uint32_t)x); +} + +SIMD_INLINE v128 v128_add_8(v128 a, v128 b) { return _mm_add_epi8(a, b); } + +SIMD_INLINE v128 v128_add_16(v128 a, v128 b) { return _mm_add_epi16(a, b); } + +SIMD_INLINE v128 v128_sadd_u8(v128 a, v128 b) { return _mm_adds_epu8(a, b); } + +SIMD_INLINE v128 v128_sadd_s8(v128 a, v128 b) { return _mm_adds_epi8(a, b); } + +SIMD_INLINE v128 v128_sadd_s16(v128 a, v128 b) { return _mm_adds_epi16(a, b); } + +SIMD_INLINE v128 v128_add_32(v128 a, v128 b) { return _mm_add_epi32(a, b); } + +SIMD_INLINE v128 v128_add_64(v128 a, v128 b) { return _mm_add_epi64(a, b); } + +SIMD_INLINE v128 v128_padd_s16(v128 a) { + return _mm_madd_epi16(a, _mm_set1_epi16(1)); +} + +SIMD_INLINE v128 v128_sub_8(v128 a, v128 b) { return _mm_sub_epi8(a, b); } + +SIMD_INLINE v128 v128_ssub_u8(v128 a, v128 b) { return _mm_subs_epu8(a, b); } + +SIMD_INLINE v128 v128_ssub_s8(v128 a, v128 b) { return _mm_subs_epi8(a, b); } + +SIMD_INLINE v128 v128_sub_16(v128 a, v128 b) { return _mm_sub_epi16(a, b); } + +SIMD_INLINE v128 v128_ssub_s16(v128 a, v128 b) { return _mm_subs_epi16(a, b); } + +SIMD_INLINE v128 v128_ssub_u16(v128 a, v128 b) { return _mm_subs_epu16(a, b); } + +SIMD_INLINE v128 v128_sub_32(v128 a, v128 b) { return _mm_sub_epi32(a, b); } + +SIMD_INLINE v128 v128_sub_64(v128 a, v128 b) { return _mm_sub_epi64(a, b); } + +SIMD_INLINE v128 v128_abs_s16(v128 a) { +#if defined(__SSSE3__) + return _mm_abs_epi16(a); +#else + return _mm_max_epi16(a, _mm_sub_epi16(_mm_setzero_si128(), a)); +#endif +} + +SIMD_INLINE v128 v128_abs_s8(v128 a) { +#if defined(__SSSE3__) + return _mm_abs_epi8(a); +#else + v128 sign = _mm_cmplt_epi8(a, _mm_setzero_si128()); + return _mm_xor_si128(sign, _mm_add_epi8(a, sign)); +#endif +} + +SIMD_INLINE v128 v128_ziplo_8(v128 a, v128 b) { + return _mm_unpacklo_epi8(b, a); +} + +SIMD_INLINE v128 v128_ziphi_8(v128 a, v128 b) { + return _mm_unpackhi_epi8(b, a); +} + +SIMD_INLINE v128 v128_ziplo_16(v128 a, v128 b) { + return _mm_unpacklo_epi16(b, a); +} + +SIMD_INLINE v128 v128_ziphi_16(v128 a, v128 b) { + return _mm_unpackhi_epi16(b, a); +} + +SIMD_INLINE v128 v128_ziplo_32(v128 a, v128 b) { + return _mm_unpacklo_epi32(b, a); +} + +SIMD_INLINE v128 v128_ziphi_32(v128 a, v128 b) { + return _mm_unpackhi_epi32(b, a); +} + +SIMD_INLINE v128 v128_ziplo_64(v128 a, v128 b) { + return _mm_unpacklo_epi64(b, a); +} + +SIMD_INLINE v128 v128_ziphi_64(v128 a, v128 b) { + return _mm_unpackhi_epi64(b, a); +} + +SIMD_INLINE v128 v128_zip_8(v64 a, v64 b) { return _mm_unpacklo_epi8(b, a); } + +SIMD_INLINE v128 v128_zip_16(v64 a, v64 b) { return _mm_unpacklo_epi16(b, a); } + +SIMD_INLINE v128 v128_zip_32(v64 a, v64 b) { return _mm_unpacklo_epi32(b, a); } + +SIMD_INLINE v128 v128_unziphi_8(v128 a, v128 b) { + return _mm_packs_epi16(_mm_srai_epi16(b, 8), _mm_srai_epi16(a, 8)); +} + +SIMD_INLINE v128 v128_unziplo_8(v128 a, v128 b) { +#if defined(__SSSE3__) +#ifdef __x86_64__ + v128 order = _mm_cvtsi64_si128(0x0e0c0a0806040200LL); +#else + v128 order = _mm_set_epi32(0, 0, 0x0e0c0a08, 0x06040200); +#endif + return _mm_unpacklo_epi64(_mm_shuffle_epi8(b, order), + _mm_shuffle_epi8(a, order)); +#else + return v128_unziphi_8(_mm_slli_si128(a, 1), _mm_slli_si128(b, 1)); +#endif +} + +SIMD_INLINE v128 v128_unziphi_16(v128 a, v128 b) { + return _mm_packs_epi32(_mm_srai_epi32(b, 16), _mm_srai_epi32(a, 16)); +} + +SIMD_INLINE v128 v128_unziplo_16(v128 a, v128 b) { +#if defined(__SSSE3__) +#ifdef __x86_64__ + v128 order = _mm_cvtsi64_si128(0x0d0c090805040100LL); +#else + v128 order = _mm_set_epi32(0, 0, 0x0d0c0908, 0x05040100); +#endif + return _mm_unpacklo_epi64(_mm_shuffle_epi8(b, order), + _mm_shuffle_epi8(a, order)); +#else + return v128_unziphi_16(_mm_slli_si128(a, 2), _mm_slli_si128(b, 2)); +#endif +} + +SIMD_INLINE v128 v128_unziphi_32(v128 a, v128 b) { + return _mm_castps_si128(_mm_shuffle_ps( + _mm_castsi128_ps(b), _mm_castsi128_ps(a), _MM_SHUFFLE(3, 1, 3, 1))); +} + +SIMD_INLINE v128 v128_unziplo_32(v128 a, v128 b) { + return _mm_castps_si128(_mm_shuffle_ps( + _mm_castsi128_ps(b), _mm_castsi128_ps(a), _MM_SHUFFLE(2, 0, 2, 0))); +} + +SIMD_INLINE v128 v128_unpack_u8_s16(v64 a) { + return _mm_unpacklo_epi8(a, _mm_setzero_si128()); +} + +SIMD_INLINE v128 v128_unpacklo_u8_s16(v128 a) { + return _mm_unpacklo_epi8(a, _mm_setzero_si128()); +} + +SIMD_INLINE v128 v128_unpackhi_u8_s16(v128 a) { + return _mm_unpackhi_epi8(a, _mm_setzero_si128()); +} + +SIMD_INLINE v128 v128_unpack_s8_s16(v64 a) { + return _mm_srai_epi16(_mm_unpacklo_epi8(a, a), 8); +} + +SIMD_INLINE v128 v128_unpacklo_s8_s16(v128 a) { + return _mm_srai_epi16(_mm_unpacklo_epi8(a, a), 8); +} + +SIMD_INLINE v128 v128_unpackhi_s8_s16(v128 a) { + return _mm_srai_epi16(_mm_unpackhi_epi8(a, a), 8); +} + +SIMD_INLINE v128 v128_pack_s32_s16(v128 a, v128 b) { + return _mm_packs_epi32(b, a); +} + +SIMD_INLINE v128 v128_pack_s32_u16(v128 a, v128 b) { +#if defined(__SSE4_1__) + return _mm_packus_epi32(b, a); +#else + return v128_from_v64(v64_pack_s32_u16(v128_high_v64(a), v128_low_v64(a)), + v64_pack_s32_u16(v128_high_v64(b), v128_low_v64(b))); +#endif +} + +SIMD_INLINE v128 v128_pack_s16_u8(v128 a, v128 b) { + return _mm_packus_epi16(b, a); +} + +SIMD_INLINE v128 v128_pack_s16_s8(v128 a, v128 b) { + return _mm_packs_epi16(b, a); +} + +SIMD_INLINE v128 v128_unpack_u16_s32(v64 a) { + return _mm_unpacklo_epi16(a, _mm_setzero_si128()); +} + +SIMD_INLINE v128 v128_unpack_s16_s32(v64 a) { + return _mm_srai_epi32(_mm_unpacklo_epi16(a, a), 16); +} + +SIMD_INLINE v128 v128_unpacklo_u16_s32(v128 a) { + return _mm_unpacklo_epi16(a, _mm_setzero_si128()); +} + +SIMD_INLINE v128 v128_unpacklo_s16_s32(v128 a) { + return _mm_srai_epi32(_mm_unpacklo_epi16(a, a), 16); +} + +SIMD_INLINE v128 v128_unpackhi_u16_s32(v128 a) { + return _mm_unpackhi_epi16(a, _mm_setzero_si128()); +} + +SIMD_INLINE v128 v128_unpackhi_s16_s32(v128 a) { + return _mm_srai_epi32(_mm_unpackhi_epi16(a, a), 16); +} + +SIMD_INLINE v128 v128_shuffle_8(v128 x, v128 pattern) { +#if defined(__SSSE3__) + return _mm_shuffle_epi8(x, pattern); +#else + v128 output; + unsigned char *input = (unsigned char *)&x; + unsigned char *index = (unsigned char *)&pattern; + char *selected = (char *)&output; + int counter; + + for (counter = 0; counter < 16; counter++) { + selected[counter] = input[index[counter] & 15]; + } + + return output; +#endif +} + +SIMD_INLINE int64_t v128_dotp_su8(v128 a, v128 b) { + v128 t1 = _mm_madd_epi16(v128_unpackhi_s8_s16(a), v128_unpackhi_u8_s16(b)); + v128 t2 = _mm_madd_epi16(v128_unpacklo_s8_s16(a), v128_unpacklo_u8_s16(b)); + v128 t = v128_add_32(t1, t2); + t = v128_add_32(t, _mm_srli_si128(t, 8)); + t = v128_add_32(t, _mm_srli_si128(t, 4)); + return (int32_t)v128_low_u32(t); +} + +SIMD_INLINE int64_t v128_dotp_s16(v128 a, v128 b) { + v128 r = _mm_madd_epi16(a, b); +#if defined(__SSE4_1__) && defined(__x86_64__) + v128 c = _mm_add_epi64(_mm_cvtepi32_epi64(r), + _mm_cvtepi32_epi64(_mm_srli_si128(r, 8))); + return _mm_cvtsi128_si64(_mm_add_epi64(c, _mm_srli_si128(c, 8))); +#else + return (int64_t)_mm_cvtsi128_si32(r) + + (int64_t)_mm_cvtsi128_si32(_mm_srli_si128(r, 4)) + + (int64_t)_mm_cvtsi128_si32(_mm_srli_si128(r, 8)) + + (int64_t)_mm_cvtsi128_si32(_mm_srli_si128(r, 12)); +#endif +} + +SIMD_INLINE uint64_t v128_hadd_u8(v128 a) { + v128 t = _mm_sad_epu8(a, _mm_setzero_si128()); + return v64_low_u32(v128_low_v64(t)) + v64_low_u32(v128_high_v64(t)); +} + +typedef v128 sad128_internal; + +SIMD_INLINE sad128_internal v128_sad_u8_init() { return _mm_setzero_si128(); } + +/* Implementation dependent return value. Result must be finalised with + v128_sad_sum(). + The result for more than 32 v128_sad_u8() calls is undefined. */ +SIMD_INLINE sad128_internal v128_sad_u8(sad128_internal s, v128 a, v128 b) { + return _mm_add_epi64(s, _mm_sad_epu8(a, b)); +} + +SIMD_INLINE uint32_t v128_sad_u8_sum(sad128_internal s) { + return v128_low_u32(_mm_add_epi32(s, _mm_unpackhi_epi64(s, s))); +} + +typedef int32_t ssd128_internal; + +SIMD_INLINE ssd128_internal v128_ssd_u8_init() { return 0; } + +/* Implementation dependent return value. Result must be finalised with + * v128_ssd_sum(). */ +SIMD_INLINE ssd128_internal v128_ssd_u8(ssd128_internal s, v128 a, v128 b) { + v128 z = _mm_setzero_si128(); + v128 l = _mm_sub_epi16(_mm_unpacklo_epi8(a, z), _mm_unpacklo_epi8(b, z)); + v128 h = _mm_sub_epi16(_mm_unpackhi_epi8(a, z), _mm_unpackhi_epi8(b, z)); + v128 rl = _mm_madd_epi16(l, l); + v128 rh = _mm_madd_epi16(h, h); + v128 r = _mm_add_epi32(rl, rh); + r = _mm_add_epi32(r, _mm_srli_si128(r, 8)); + r = _mm_add_epi32(r, _mm_srli_si128(r, 4)); + return s + _mm_cvtsi128_si32(r); +} + +SIMD_INLINE int32_t v128_ssd_u8_sum(ssd128_internal s) { return s; } + +SIMD_INLINE v128 v128_or(v128 a, v128 b) { return _mm_or_si128(a, b); } + +SIMD_INLINE v128 v128_xor(v128 a, v128 b) { return _mm_xor_si128(a, b); } + +SIMD_INLINE v128 v128_and(v128 a, v128 b) { return _mm_and_si128(a, b); } + +SIMD_INLINE v128 v128_andn(v128 a, v128 b) { return _mm_andnot_si128(b, a); } + +SIMD_INLINE v128 v128_mul_s16(v64 a, v64 b) { + v64 lo_bits = v64_mullo_s16(a, b); + v64 hi_bits = v64_mulhi_s16(a, b); + return v128_from_v64(v64_ziphi_16(hi_bits, lo_bits), + v64_ziplo_16(hi_bits, lo_bits)); +} + +SIMD_INLINE v128 v128_mullo_s16(v128 a, v128 b) { + return _mm_mullo_epi16(a, b); +} + +SIMD_INLINE v128 v128_mulhi_s16(v128 a, v128 b) { + return _mm_mulhi_epi16(a, b); +} + +SIMD_INLINE v128 v128_mullo_s32(v128 a, v128 b) { +#if defined(__SSE4_1__) + return _mm_mullo_epi32(a, b); +#else + return _mm_unpacklo_epi32( + _mm_shuffle_epi32(_mm_mul_epu32(a, b), 8), + _mm_shuffle_epi32( + _mm_mul_epu32(_mm_srli_si128(a, 4), _mm_srli_si128(b, 4)), 8)); +#endif +} + +SIMD_INLINE int64_t v128_dotp_s32(v128 a, v128 b) { + v128 r = v128_mullo_s32(a, b); + return (int64_t)_mm_cvtsi128_si32(r) + + (int64_t)_mm_cvtsi128_si32(_mm_srli_si128(r, 4)) + + (int64_t)_mm_cvtsi128_si32(_mm_srli_si128(r, 8)) + + (int64_t)_mm_cvtsi128_si32(_mm_srli_si128(r, 12)); +} + +SIMD_INLINE v128 v128_madd_s16(v128 a, v128 b) { return _mm_madd_epi16(a, b); } + +SIMD_INLINE v128 v128_madd_us8(v128 a, v128 b) { +#if defined(__SSSE3__) + return _mm_maddubs_epi16(a, b); +#else + return _mm_packs_epi32( + _mm_madd_epi16(_mm_unpacklo_epi8(a, _mm_setzero_si128()), + _mm_srai_epi16(_mm_unpacklo_epi8(b, b), 8)), + _mm_madd_epi16(_mm_unpackhi_epi8(a, _mm_setzero_si128()), + _mm_srai_epi16(_mm_unpackhi_epi8(b, b), 8))); +#endif +} + +SIMD_INLINE v128 v128_padd_u8(v128 a) { + return v128_madd_us8(a, _mm_set1_epi8(1)); +} + +SIMD_INLINE v128 v128_avg_u8(v128 a, v128 b) { return _mm_avg_epu8(a, b); } + +SIMD_INLINE v128 v128_rdavg_u8(v128 a, v128 b) { + return _mm_sub_epi8(_mm_avg_epu8(a, b), + _mm_and_si128(_mm_xor_si128(a, b), v128_dup_8(1))); +} + +SIMD_INLINE v128 v128_rdavg_u16(v128 a, v128 b) { + return _mm_sub_epi16(_mm_avg_epu16(a, b), + _mm_and_si128(_mm_xor_si128(a, b), v128_dup_16(1))); +} + +SIMD_INLINE v128 v128_avg_u16(v128 a, v128 b) { return _mm_avg_epu16(a, b); } + +SIMD_INLINE v128 v128_min_u8(v128 a, v128 b) { return _mm_min_epu8(a, b); } + +SIMD_INLINE v128 v128_max_u8(v128 a, v128 b) { return _mm_max_epu8(a, b); } + +SIMD_INLINE v128 v128_min_s8(v128 a, v128 b) { +#if defined(__SSE4_1__) + return _mm_min_epi8(a, b); +#else + v128 mask = _mm_cmplt_epi8(a, b); + return _mm_or_si128(_mm_andnot_si128(mask, b), _mm_and_si128(mask, a)); +#endif +} + +SIMD_INLINE uint32_t v128_movemask_8(v128 a) { return _mm_movemask_epi8(a); } + +SIMD_INLINE v128 v128_blend_8(v128 a, v128 b, v128 c) { +#if defined(__SSE4_1__) + return _mm_blendv_epi8(a, b, c); +#else + c = _mm_cmplt_epi8(c, v128_zero()); + return v128_or(v128_and(b, c), v128_andn(a, c)); +#endif +} + +SIMD_INLINE v128 v128_max_s8(v128 a, v128 b) { +#if defined(__SSE4_1__) + return _mm_max_epi8(a, b); +#else + v128 mask = _mm_cmplt_epi8(b, a); + return _mm_or_si128(_mm_andnot_si128(mask, b), _mm_and_si128(mask, a)); +#endif +} + +SIMD_INLINE v128 v128_min_s16(v128 a, v128 b) { return _mm_min_epi16(a, b); } + +SIMD_INLINE v128 v128_max_s16(v128 a, v128 b) { return _mm_max_epi16(a, b); } + +SIMD_INLINE v128 v128_min_s32(v128 a, v128 b) { +#if defined(__SSE4_1__) + return _mm_min_epi32(a, b); +#else + v128 mask = _mm_cmplt_epi32(a, b); + return _mm_or_si128(_mm_andnot_si128(mask, b), _mm_and_si128(mask, a)); +#endif +} + +SIMD_INLINE v128 v128_max_s32(v128 a, v128 b) { +#if defined(__SSE4_1__) + return _mm_max_epi32(a, b); +#else + v128 mask = _mm_cmplt_epi32(b, a); + return _mm_or_si128(_mm_andnot_si128(mask, b), _mm_and_si128(mask, a)); +#endif +} + +SIMD_INLINE v128 v128_cmpgt_s8(v128 a, v128 b) { return _mm_cmpgt_epi8(a, b); } + +SIMD_INLINE v128 v128_cmplt_s8(v128 a, v128 b) { return _mm_cmplt_epi8(a, b); } + +SIMD_INLINE v128 v128_cmpeq_8(v128 a, v128 b) { return _mm_cmpeq_epi8(a, b); } + +SIMD_INLINE v128 v128_cmpgt_s16(v128 a, v128 b) { + return _mm_cmpgt_epi16(a, b); +} + +SIMD_INLINE v128 v128_cmplt_s16(v128 a, v128 b) { + return _mm_cmplt_epi16(a, b); +} + +SIMD_INLINE v128 v128_cmpeq_32(v128 a, v128 b) { return _mm_cmpeq_epi32(a, b); } + +SIMD_INLINE v128 v128_cmpgt_s32(v128 a, v128 b) { + return _mm_cmpgt_epi32(a, b); +} + +SIMD_INLINE v128 v128_cmplt_s32(v128 a, v128 b) { + return _mm_cmplt_epi32(a, b); +} + +SIMD_INLINE v128 v128_cmpeq_16(v128 a, v128 b) { return _mm_cmpeq_epi16(a, b); } + +SIMD_INLINE v128 v128_shl_8(v128 a, unsigned int c) { + return _mm_and_si128(_mm_set1_epi8((uint8_t)(0xff << c)), + _mm_sll_epi16(a, _mm_cvtsi32_si128(c))); +} + +SIMD_INLINE v128 v128_shr_u8(v128 a, unsigned int c) { + return _mm_and_si128(_mm_set1_epi8(0xff >> c), + _mm_srl_epi16(a, _mm_cvtsi32_si128(c))); +} + +SIMD_INLINE v128 v128_shr_s8(v128 a, unsigned int c) { + __m128i x = _mm_cvtsi32_si128(c + 8); + return _mm_packs_epi16(_mm_sra_epi16(_mm_unpacklo_epi8(a, a), x), + _mm_sra_epi16(_mm_unpackhi_epi8(a, a), x)); +} + +SIMD_INLINE v128 v128_shl_16(v128 a, unsigned int c) { + return _mm_sll_epi16(a, _mm_cvtsi32_si128(c)); +} + +SIMD_INLINE v128 v128_shr_u16(v128 a, unsigned int c) { + return _mm_srl_epi16(a, _mm_cvtsi32_si128(c)); +} + +SIMD_INLINE v128 v128_shr_s16(v128 a, unsigned int c) { + return _mm_sra_epi16(a, _mm_cvtsi32_si128(c)); +} + +SIMD_INLINE v128 v128_shl_32(v128 a, unsigned int c) { + return _mm_sll_epi32(a, _mm_cvtsi32_si128(c)); +} + +SIMD_INLINE v128 v128_shr_u32(v128 a, unsigned int c) { + return _mm_srl_epi32(a, _mm_cvtsi32_si128(c)); +} + +SIMD_INLINE v128 v128_shr_s32(v128 a, unsigned int c) { + return _mm_sra_epi32(a, _mm_cvtsi32_si128(c)); +} + +SIMD_INLINE v128 v128_shl_64(v128 a, unsigned int c) { + return _mm_sll_epi64(a, _mm_cvtsi32_si128(c)); +} + +SIMD_INLINE v128 v128_shr_u64(v128 a, unsigned int c) { + return _mm_srl_epi64(a, _mm_cvtsi32_si128(c)); +} + +SIMD_INLINE v128 v128_shr_s64(v128 a, unsigned int c) { + // _mm_sra_epi64 is missing in gcc? + return v128_from_64((int64_t)v64_u64(v128_high_v64(a)) >> c, + (int64_t)v64_u64(v128_low_v64(a)) >> c); + // return _mm_sra_epi64(a, _mm_cvtsi32_si128(c)); +} + +/* These intrinsics require immediate values, so we must use #defines + to enforce that. */ +#define v128_shl_n_byte(a, c) _mm_slli_si128(a, (c)&127) +#define v128_shr_n_byte(a, c) _mm_srli_si128(a, (c)&127) +#define v128_shl_n_8(a, c) \ + _mm_and_si128(_mm_set1_epi8((uint8_t)(0xff << (c))), _mm_slli_epi16(a, c)) +#define v128_shr_n_u8(a, c) \ + _mm_and_si128(_mm_set1_epi8(0xff >> (c)), _mm_srli_epi16(a, c)) +#define v128_shr_n_s8(a, c) \ + _mm_packs_epi16(_mm_srai_epi16(_mm_unpacklo_epi8(a, a), (c) + 8), \ + _mm_srai_epi16(_mm_unpackhi_epi8(a, a), (c) + 8)) +#define v128_shl_n_16(a, c) _mm_slli_epi16(a, c) +#define v128_shr_n_u16(a, c) _mm_srli_epi16(a, c) +#define v128_shr_n_s16(a, c) _mm_srai_epi16(a, c) +#define v128_shl_n_32(a, c) _mm_slli_epi32(a, c) +#define v128_shr_n_u32(a, c) _mm_srli_epi32(a, c) +#define v128_shr_n_s32(a, c) _mm_srai_epi32(a, c) +#define v128_shl_n_64(a, c) _mm_slli_epi64(a, c) +#define v128_shr_n_u64(a, c) _mm_srli_epi64(a, c) +#define v128_shr_n_s64(a, c) \ + v128_shr_s64(a, c) // _mm_srai_epi64 missing in gcc? + +typedef v128 sad128_internal_u16; + +SIMD_INLINE sad128_internal_u16 v128_sad_u16_init() { return v128_zero(); } + +/* Implementation dependent return value. Result must be finalised with + * v128_sad_u16_sum(). */ +SIMD_INLINE sad128_internal_u16 v128_sad_u16(sad128_internal_u16 s, v128 a, + v128 b) { +#if defined(__SSE4_1__) + v128 t = v128_sub_16(_mm_max_epu16(a, b), _mm_min_epu16(a, b)); +#else + v128 t = v128_cmplt_s16(v128_xor(a, v128_dup_16(32768)), + v128_xor(b, v128_dup_16(32768))); + t = v128_sub_16(v128_or(v128_and(b, t), v128_andn(a, t)), + v128_or(v128_and(a, t), v128_andn(b, t))); +#endif + return v128_add_32( + s, v128_add_32(v128_unpackhi_u16_s32(t), v128_unpacklo_u16_s32(t))); +} + +SIMD_INLINE uint32_t v128_sad_u16_sum(sad128_internal_u16 s) { + return v128_low_u32(s) + v128_low_u32(v128_shr_n_byte(s, 4)) + + v128_low_u32(v128_shr_n_byte(s, 8)) + + v128_low_u32(v128_shr_n_byte(s, 12)); +} + +typedef v128 ssd128_internal_s16; + +SIMD_INLINE ssd128_internal_s16 v128_ssd_s16_init() { return v128_zero(); } + +/* Implementation dependent return value. Result must be finalised with + * v128_ssd_s16_sum(). */ +SIMD_INLINE ssd128_internal_s16 v128_ssd_s16(ssd128_internal_s16 s, v128 a, + v128 b) { + v128 d = v128_sub_16(a, b); + d = v128_madd_s16(d, d); + return v128_add_64(s, v128_add_64(_mm_unpackhi_epi32(d, v128_zero()), + _mm_unpacklo_epi32(d, v128_zero()))); +} + +SIMD_INLINE uint64_t v128_ssd_s16_sum(ssd128_internal_s16 s) { + return v64_u64(v128_low_v64(s)) + v64_u64(v128_high_v64(s)); +} + +#endif // AOM_AOM_DSP_SIMD_V128_INTRINSICS_X86_H_ diff --git a/third_party/aom/aom_dsp/simd/v256_intrinsics.h b/third_party/aom/aom_dsp/simd/v256_intrinsics.h new file mode 100644 index 0000000000..cb99d35b71 --- /dev/null +++ b/third_party/aom/aom_dsp/simd/v256_intrinsics.h @@ -0,0 +1,376 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_SIMD_V256_INTRINSICS_H_ +#define AOM_AOM_DSP_SIMD_V256_INTRINSICS_H_ + +#include +#include +#include + +#include "aom_dsp/simd/v256_intrinsics_c.h" +#include "aom_dsp/simd/v128_intrinsics.h" +#include "aom_dsp/simd/v64_intrinsics.h" + +/* Fallback to plain, unoptimised C. */ + +typedef c_v256 v256; + +SIMD_INLINE uint32_t v256_low_u32(v256 a) { return c_v256_low_u32(a); } +SIMD_INLINE v64 v256_low_v64(v256 a) { return c_v256_low_v64(a); } +SIMD_INLINE uint64_t v256_low_u64(v256 a) { return c_v256_low_u64(a); } +SIMD_INLINE v128 v256_low_v128(v256 a) { return c_v256_low_v128(a); } +SIMD_INLINE v128 v256_high_v128(v256 a) { return c_v256_high_v128(a); } +SIMD_INLINE v256 v256_from_v128(v128 hi, v128 lo) { + return c_v256_from_v128(hi, lo); +} +SIMD_INLINE v256 v256_from_64(uint64_t a, uint64_t b, uint64_t c, uint64_t d) { + return c_v256_from_64(a, b, c, d); +} +SIMD_INLINE v256 v256_from_v64(v64 a, v64 b, v64 c, v64 d) { + return c_v256_from_v64(a, b, c, d); +} + +SIMD_INLINE v256 v256_load_unaligned(const void *p) { + return c_v256_load_unaligned(p); +} +SIMD_INLINE v256 v256_load_aligned(const void *p) { + return c_v256_load_aligned(p); +} + +SIMD_INLINE void v256_store_unaligned(void *p, v256 a) { + c_v256_store_unaligned(p, a); +} +SIMD_INLINE void v256_store_aligned(void *p, v256 a) { + c_v256_store_aligned(p, a); +} + +SIMD_INLINE v256 v256_align(v256 a, v256 b, unsigned int c) { + return c_v256_align(a, b, c); +} + +SIMD_INLINE v256 v256_zero() { return c_v256_zero(); } +SIMD_INLINE v256 v256_dup_8(uint8_t x) { return c_v256_dup_8(x); } +SIMD_INLINE v256 v256_dup_16(uint16_t x) { return c_v256_dup_16(x); } +SIMD_INLINE v256 v256_dup_32(uint32_t x) { return c_v256_dup_32(x); } +SIMD_INLINE v256 v256_dup_64(uint64_t x) { return c_v256_dup_64(x); } + +typedef uint32_t sad256_internal; +SIMD_INLINE sad256_internal v256_sad_u8_init() { return c_v256_sad_u8_init(); } +SIMD_INLINE sad256_internal v256_sad_u8(sad256_internal s, v256 a, v256 b) { + return c_v256_sad_u8(s, a, b); +} +SIMD_INLINE uint32_t v256_sad_u8_sum(sad256_internal s) { + return c_v256_sad_u8_sum(s); +} +typedef uint32_t ssd256_internal; +SIMD_INLINE ssd256_internal v256_ssd_u8_init() { return c_v256_ssd_u8_init(); } +SIMD_INLINE ssd256_internal v256_ssd_u8(ssd256_internal s, v256 a, v256 b) { + return c_v256_ssd_u8(s, a, b); +} +SIMD_INLINE uint32_t v256_ssd_u8_sum(ssd256_internal s) { + return c_v256_ssd_u8_sum(s); +} + +SIMD_INLINE int64_t v256_dotp_su8(v256 a, v256 b) { + return c_v256_dotp_su8(a, b); +} +SIMD_INLINE int64_t v256_dotp_s16(v256 a, v256 b) { + return c_v256_dotp_s16(a, b); +} +SIMD_INLINE int64_t v256_dotp_s32(v256 a, v256 b) { + return c_v256_dotp_s32(a, b); +} +SIMD_INLINE uint64_t v256_hadd_u8(v256 a) { return c_v256_hadd_u8(a); } + +SIMD_INLINE v256 v256_or(v256 a, v256 b) { return c_v256_or(a, b); } +SIMD_INLINE v256 v256_xor(v256 a, v256 b) { return c_v256_xor(a, b); } +SIMD_INLINE v256 v256_and(v256 a, v256 b) { return c_v256_and(a, b); } +SIMD_INLINE v256 v256_andn(v256 a, v256 b) { return c_v256_andn(a, b); } + +SIMD_INLINE v256 v256_add_8(v256 a, v256 b) { return c_v256_add_8(a, b); } +SIMD_INLINE v256 v256_add_16(v256 a, v256 b) { return c_v256_add_16(a, b); } +SIMD_INLINE v256 v256_sadd_s8(v256 a, v256 b) { return c_v256_sadd_s8(a, b); } +SIMD_INLINE v256 v256_sadd_u8(v256 a, v256 b) { return c_v256_sadd_u8(a, b); } +SIMD_INLINE v256 v256_sadd_s16(v256 a, v256 b) { return c_v256_sadd_s16(a, b); } +SIMD_INLINE v256 v256_add_32(v256 a, v256 b) { return c_v256_add_32(a, b); } +SIMD_INLINE v256 v256_add_64(v256 a, v256 b) { return c_v256_add_64(a, b); } +SIMD_INLINE v256 v256_sub_64(v256 a, v256 b) { return c_v256_sub_64(a, b); } +SIMD_INLINE v256 v256_padd_u8(v256 a) { return c_v256_padd_u8(a); } +SIMD_INLINE v256 v256_padd_s16(v256 a) { return c_v256_padd_s16(a); } +SIMD_INLINE v256 v256_sub_8(v256 a, v256 b) { return c_v256_sub_8(a, b); } +SIMD_INLINE v256 v256_ssub_u8(v256 a, v256 b) { return c_v256_ssub_u8(a, b); } +SIMD_INLINE v256 v256_ssub_s8(v256 a, v256 b) { return c_v256_ssub_s8(a, b); } +SIMD_INLINE v256 v256_sub_16(v256 a, v256 b) { return c_v256_sub_16(a, b); } +SIMD_INLINE v256 v256_ssub_s16(v256 a, v256 b) { return c_v256_ssub_s16(a, b); } +SIMD_INLINE v256 v256_ssub_u16(v256 a, v256 b) { return c_v256_ssub_u16(a, b); } +SIMD_INLINE v256 v256_sub_32(v256 a, v256 b) { return c_v256_sub_32(a, b); } +SIMD_INLINE v256 v256_abs_s16(v256 a) { return c_v256_abs_s16(a); } +SIMD_INLINE v256 v256_abs_s8(v256 a) { return c_v256_abs_s8(a); } + +SIMD_INLINE v256 v256_mul_s16(v128 a, v128 b) { return c_v256_mul_s16(a, b); } +SIMD_INLINE v256 v256_mullo_s16(v256 a, v256 b) { + return c_v256_mullo_s16(a, b); +} +SIMD_INLINE v256 v256_mulhi_s16(v256 a, v256 b) { + return c_v256_mulhi_s16(a, b); +} +SIMD_INLINE v256 v256_mullo_s32(v256 a, v256 b) { + return c_v256_mullo_s32(a, b); +} +SIMD_INLINE v256 v256_madd_s16(v256 a, v256 b) { return c_v256_madd_s16(a, b); } +SIMD_INLINE v256 v256_madd_us8(v256 a, v256 b) { return c_v256_madd_us8(a, b); } + +SIMD_INLINE uint32_t v256_movemask_8(v256 a) { return c_v256_movemask_8(a); } +SIMD_INLINE v256 v256_blend_8(v256 a, v256 b, v256 c) { + return c_v256_blend_8(a, b, c); +} + +SIMD_INLINE v256 v256_avg_u8(v256 a, v256 b) { return c_v256_avg_u8(a, b); } +SIMD_INLINE v256 v256_rdavg_u8(v256 a, v256 b) { return c_v256_rdavg_u8(a, b); } +SIMD_INLINE v256 v256_rdavg_u16(v256 a, v256 b) { + return c_v256_rdavg_u16(a, b); +} +SIMD_INLINE v256 v256_avg_u16(v256 a, v256 b) { return c_v256_avg_u16(a, b); } +SIMD_INLINE v256 v256_min_u8(v256 a, v256 b) { return c_v256_min_u8(a, b); } +SIMD_INLINE v256 v256_max_u8(v256 a, v256 b) { return c_v256_max_u8(a, b); } +SIMD_INLINE v256 v256_min_s8(v256 a, v256 b) { return c_v256_min_s8(a, b); } +SIMD_INLINE v256 v256_max_s8(v256 a, v256 b) { return c_v256_max_s8(a, b); } +SIMD_INLINE v256 v256_min_s16(v256 a, v256 b) { return c_v256_min_s16(a, b); } +SIMD_INLINE v256 v256_max_s16(v256 a, v256 b) { return c_v256_max_s16(a, b); } +SIMD_INLINE v256 v256_min_s32(v256 a, v256 b) { return c_v256_min_s32(a, b); } +SIMD_INLINE v256 v256_max_s32(v256 a, v256 b) { return c_v256_max_s32(a, b); } + +SIMD_INLINE v256 v256_ziplo_8(v256 a, v256 b) { return c_v256_ziplo_8(a, b); } +SIMD_INLINE v256 v256_ziphi_8(v256 a, v256 b) { return c_v256_ziphi_8(a, b); } +SIMD_INLINE v256 v256_ziplo_16(v256 a, v256 b) { return c_v256_ziplo_16(a, b); } +SIMD_INLINE v256 v256_ziphi_16(v256 a, v256 b) { return c_v256_ziphi_16(a, b); } +SIMD_INLINE v256 v256_ziplo_32(v256 a, v256 b) { return c_v256_ziplo_32(a, b); } +SIMD_INLINE v256 v256_ziphi_32(v256 a, v256 b) { return c_v256_ziphi_32(a, b); } +SIMD_INLINE v256 v256_ziplo_64(v256 a, v256 b) { return c_v256_ziplo_64(a, b); } +SIMD_INLINE v256 v256_ziphi_64(v256 a, v256 b) { return c_v256_ziphi_64(a, b); } +SIMD_INLINE v256 v256_ziplo_128(v256 a, v256 b) { + return c_v256_ziplo_128(a, b); +} +SIMD_INLINE v256 v256_ziphi_128(v256 a, v256 b) { + return c_v256_ziphi_128(a, b); +} +SIMD_INLINE v256 v256_zip_8(v128 a, v128 b) { return c_v256_zip_8(a, b); } +SIMD_INLINE v256 v256_zip_16(v128 a, v128 b) { return c_v256_zip_16(a, b); } +SIMD_INLINE v256 v256_zip_32(v128 a, v128 b) { return c_v256_zip_32(a, b); } +SIMD_INLINE v256 v256_unziplo_8(v256 a, v256 b) { + return c_v256_unziplo_8(a, b); +} +SIMD_INLINE v256 v256_unziphi_8(v256 a, v256 b) { + return c_v256_unziphi_8(a, b); +} +SIMD_INLINE v256 v256_unziplo_16(v256 a, v256 b) { + return c_v256_unziplo_16(a, b); +} +SIMD_INLINE v256 v256_unziphi_16(v256 a, v256 b) { + return c_v256_unziphi_16(a, b); +} +SIMD_INLINE v256 v256_unziplo_32(v256 a, v256 b) { + return c_v256_unziplo_32(a, b); +} +SIMD_INLINE v256 v256_unziphi_32(v256 a, v256 b) { + return c_v256_unziphi_32(a, b); +} +SIMD_INLINE v256 v256_unziplo_64(v256 a, v256 b) { + return c_v256_unziplo_64(a, b); +} +SIMD_INLINE v256 v256_unziphi_64(v256 a, v256 b) { + return c_v256_unziphi_64(a, b); +} +SIMD_INLINE v256 v256_unpack_u8_s16(v128 a) { return c_v256_unpack_u8_s16(a); } +SIMD_INLINE v256 v256_unpacklo_u8_s16(v256 a) { + return c_v256_unpacklo_u8_s16(a); +} +SIMD_INLINE v256 v256_unpackhi_u8_s16(v256 a) { + return c_v256_unpackhi_u8_s16(a); +} +SIMD_INLINE v256 v256_unpack_s8_s16(v128 a) { return c_v256_unpack_s8_s16(a); } +SIMD_INLINE v256 v256_unpacklo_s8_s16(v256 a) { + return c_v256_unpacklo_s8_s16(a); +} +SIMD_INLINE v256 v256_unpackhi_s8_s16(v256 a) { + return c_v256_unpackhi_s8_s16(a); +} +SIMD_INLINE v256 v256_pack_s32_s16(v256 a, v256 b) { + return c_v256_pack_s32_s16(a, b); +} +SIMD_INLINE v256 v256_pack_s32_u16(v256 a, v256 b) { + return c_v256_pack_s32_u16(a, b); +} +SIMD_INLINE v256 v256_pack_s16_u8(v256 a, v256 b) { + return c_v256_pack_s16_u8(a, b); +} +SIMD_INLINE v256 v256_pack_s16_s8(v256 a, v256 b) { + return c_v256_pack_s16_s8(a, b); +} +SIMD_INLINE v256 v256_unpack_u16_s32(v128 a) { + return c_v256_unpack_u16_s32(a); +} +SIMD_INLINE v256 v256_unpack_s16_s32(v128 a) { + return c_v256_unpack_s16_s32(a); +} +SIMD_INLINE v256 v256_unpacklo_u16_s32(v256 a) { + return c_v256_unpacklo_u16_s32(a); +} +SIMD_INLINE v256 v256_unpacklo_s16_s32(v256 a) { + return c_v256_unpacklo_s16_s32(a); +} +SIMD_INLINE v256 v256_unpackhi_u16_s32(v256 a) { + return c_v256_unpackhi_u16_s32(a); +} +SIMD_INLINE v256 v256_unpackhi_s16_s32(v256 a) { + return c_v256_unpackhi_s16_s32(a); +} +SIMD_INLINE v256 v256_shuffle_8(v256 a, v256 pattern) { + return c_v256_shuffle_8(a, pattern); +} +SIMD_INLINE v256 v256_wideshuffle_8(v256 a, v256 b, v256 pattern) { + return c_v256_wideshuffle_8(a, b, pattern); +} +SIMD_INLINE v256 v256_pshuffle_8(v256 a, v256 pattern) { + return c_v256_pshuffle_8(a, pattern); +} + +SIMD_INLINE v256 v256_cmpgt_s8(v256 a, v256 b) { return c_v256_cmpgt_s8(a, b); } +SIMD_INLINE v256 v256_cmplt_s8(v256 a, v256 b) { return c_v256_cmplt_s8(a, b); } +SIMD_INLINE v256 v256_cmpeq_8(v256 a, v256 b) { return c_v256_cmpeq_8(a, b); } +SIMD_INLINE v256 v256_cmpgt_s16(v256 a, v256 b) { + return c_v256_cmpgt_s16(a, b); +} +SIMD_INLINE v256 v256_cmplt_s16(v256 a, v256 b) { + return c_v256_cmplt_s16(a, b); +} +SIMD_INLINE v256 v256_cmpeq_16(v256 a, v256 b) { return c_v256_cmpeq_16(a, b); } +SIMD_INLINE v256 v256_cmpeq_32(v256 a, v256 b) { return c_v256_cmpeq_32(a, b); } + +SIMD_INLINE v256 v256_cmpgt_s32(v256 a, v256 b) { + return c_v256_cmpgt_s32(a, b); +} +SIMD_INLINE v256 v256_cmplt_s32(v256 a, v256 b) { + return c_v256_cmplt_s32(a, b); +} +SIMD_INLINE v256 v256_shl_8(v256 a, unsigned int c) { + return c_v256_shl_8(a, c); +} +SIMD_INLINE v256 v256_shr_u8(v256 a, unsigned int c) { + return c_v256_shr_u8(a, c); +} +SIMD_INLINE v256 v256_shr_s8(v256 a, unsigned int c) { + return c_v256_shr_s8(a, c); +} +SIMD_INLINE v256 v256_shl_16(v256 a, unsigned int c) { + return c_v256_shl_16(a, c); +} +SIMD_INLINE v256 v256_shr_u16(v256 a, unsigned int c) { + return c_v256_shr_u16(a, c); +} +SIMD_INLINE v256 v256_shr_s16(v256 a, unsigned int c) { + return c_v256_shr_s16(a, c); +} +SIMD_INLINE v256 v256_shl_32(v256 a, unsigned int c) { + return c_v256_shl_32(a, c); +} +SIMD_INLINE v256 v256_shr_u32(v256 a, unsigned int c) { + return c_v256_shr_u32(a, c); +} +SIMD_INLINE v256 v256_shr_s32(v256 a, unsigned int c) { + return c_v256_shr_s32(a, c); +} +SIMD_INLINE v256 v256_shl_64(v256 a, unsigned int c) { + return c_v256_shl_64(a, c); +} +SIMD_INLINE v256 v256_shr_u64(v256 a, unsigned int c) { + return c_v256_shr_u64(a, c); +} +SIMD_INLINE v256 v256_shr_s64(v256 a, unsigned int c) { + return c_v256_shr_s64(a, c); +} + +SIMD_INLINE v256 v256_shr_n_byte(v256 a, unsigned int n) { + return c_v256_shr_n_byte(a, n); +} +SIMD_INLINE v256 v256_shl_n_byte(v256 a, unsigned int n) { + return c_v256_shl_n_byte(a, n); +} +SIMD_INLINE v256 v256_shl_n_8(v256 a, unsigned int n) { + return c_v256_shl_n_8(a, n); +} +SIMD_INLINE v256 v256_shl_n_16(v256 a, unsigned int n) { + return c_v256_shl_n_16(a, n); +} +SIMD_INLINE v256 v256_shl_n_32(v256 a, unsigned int n) { + return c_v256_shl_n_32(a, n); +} +SIMD_INLINE v256 v256_shl_n_64(v256 a, unsigned int n) { + return c_v256_shl_n_64(a, n); +} +SIMD_INLINE v256 v256_shr_n_u8(v256 a, unsigned int n) { + return c_v256_shr_n_u8(a, n); +} +SIMD_INLINE v256 v256_shr_n_u16(v256 a, unsigned int n) { + return c_v256_shr_n_u16(a, n); +} +SIMD_INLINE v256 v256_shr_n_u32(v256 a, unsigned int n) { + return c_v256_shr_n_u32(a, n); +} +SIMD_INLINE v256 v256_shr_n_u64(v256 a, unsigned int n) { + return c_v256_shr_n_u64(a, n); +} +SIMD_INLINE v256 v256_shr_n_s8(v256 a, unsigned int n) { + return c_v256_shr_n_s8(a, n); +} +SIMD_INLINE v256 v256_shr_n_s16(v256 a, unsigned int n) { + return c_v256_shr_n_s16(a, n); +} +SIMD_INLINE v256 v256_shr_n_s32(v256 a, unsigned int n) { + return c_v256_shr_n_s32(a, n); +} +SIMD_INLINE v256 v256_shr_n_s64(v256 a, unsigned int n) { + return c_v256_shr_n_s64(a, n); +} + +SIMD_INLINE v256 v256_shr_n_word(v256 a, unsigned int n) { + return c_v256_shr_n_word(a, n); +} +SIMD_INLINE v256 v256_shl_n_word(v256 a, unsigned int n) { + return c_v256_shl_n_word(a, n); +} + +typedef uint32_t sad256_internal_u16; +SIMD_INLINE sad256_internal_u16 v256_sad_u16_init() { + return c_v256_sad_u16_init(); +} +SIMD_INLINE sad256_internal_u16 v256_sad_u16(sad256_internal_u16 s, v256 a, + v256 b) { + return c_v256_sad_u16(s, a, b); +} +SIMD_INLINE uint32_t v256_sad_u16_sum(sad256_internal_u16 s) { + return c_v256_sad_u16_sum(s); +} + +typedef uint64_t ssd256_internal_s16; +SIMD_INLINE ssd256_internal_s16 v256_ssd_s16_init() { + return c_v256_ssd_s16_init(); +} +SIMD_INLINE ssd256_internal_s16 v256_ssd_s16(ssd256_internal_s16 s, v256 a, + v256 b) { + return c_v256_ssd_s16(s, a, b); +} +SIMD_INLINE uint64_t v256_ssd_s16_sum(ssd256_internal_s16 s) { + return c_v256_ssd_s16_sum(s); +} + +#endif // AOM_AOM_DSP_SIMD_V256_INTRINSICS_H_ diff --git a/third_party/aom/aom_dsp/simd/v256_intrinsics_arm.h b/third_party/aom/aom_dsp/simd/v256_intrinsics_arm.h new file mode 100644 index 0000000000..bd86ea1721 --- /dev/null +++ b/third_party/aom/aom_dsp/simd/v256_intrinsics_arm.h @@ -0,0 +1,17 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_SIMD_V256_INTRINSICS_ARM_H_ +#define AOM_AOM_DSP_SIMD_V256_INTRINSICS_ARM_H_ + +#include "aom_dsp/simd/v256_intrinsics_v128.h" + +#endif // AOM_AOM_DSP_SIMD_V256_INTRINSICS_ARM_H_ diff --git a/third_party/aom/aom_dsp/simd/v256_intrinsics_c.h b/third_party/aom/aom_dsp/simd/v256_intrinsics_c.h new file mode 100644 index 0000000000..a1c08e95ad --- /dev/null +++ b/third_party/aom/aom_dsp/simd/v256_intrinsics_c.h @@ -0,0 +1,953 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_SIMD_V256_INTRINSICS_C_H_ +#define AOM_AOM_DSP_SIMD_V256_INTRINSICS_C_H_ + +#include +#include + +#include "config/aom_config.h" + +#include "aom_dsp/simd/v128_intrinsics_c.h" + +typedef union { + uint8_t u8[32]; + uint16_t u16[16]; + uint32_t u32[8]; + uint64_t u64[4]; + int8_t s8[32]; + int16_t s16[16]; + int32_t s32[8]; + int64_t s64[4]; + c_v64 v64[4]; + c_v128 v128[2]; +} c_v256; + +SIMD_INLINE uint32_t c_v256_low_u32(c_v256 a) { return a.u32[0]; } + +SIMD_INLINE c_v64 c_v256_low_v64(c_v256 a) { return a.v64[0]; } + +SIMD_INLINE uint64_t c_v256_low_u64(c_v256 a) { return a.u64[0]; } + +SIMD_INLINE c_v128 c_v256_low_v128(c_v256 a) { return a.v128[0]; } + +SIMD_INLINE c_v128 c_v256_high_v128(c_v256 a) { return a.v128[1]; } + +SIMD_INLINE c_v256 c_v256_from_v128(c_v128 hi, c_v128 lo) { + c_v256 t; + t.v128[1] = hi; + t.v128[0] = lo; + return t; +} + +SIMD_INLINE c_v256 c_v256_from_64(uint64_t a, uint64_t b, uint64_t c, + uint64_t d) { + c_v256 t; + t.u64[3] = a; + t.u64[2] = b; + t.u64[1] = c; + t.u64[0] = d; + return t; +} + +SIMD_INLINE c_v256 c_v256_from_v64(c_v64 a, c_v64 b, c_v64 c, c_v64 d) { + c_v256 t; + t.u64[3] = a.u64; + t.u64[2] = b.u64; + t.u64[1] = c.u64; + t.u64[0] = d.u64; + return t; +} + +SIMD_INLINE c_v256 c_v256_load_unaligned(const void *p) { + c_v256 t; + uint8_t *pp = (uint8_t *)p; + uint8_t *q = (uint8_t *)&t; + int c; + for (c = 0; c < 32; c++) q[c] = pp[c]; + return t; +} + +SIMD_INLINE c_v256 c_v256_load_aligned(const void *p) { + if (SIMD_CHECK && (uintptr_t)p & 31) { + fprintf(stderr, "Error: unaligned v256 load at %p\n", p); + abort(); + } + return c_v256_load_unaligned(p); +} + +SIMD_INLINE void c_v256_store_unaligned(void *p, c_v256 a) { + uint8_t *pp = (uint8_t *)p; + uint8_t *q = (uint8_t *)&a; + int c; + for (c = 0; c < 32; c++) pp[c] = q[c]; +} + +SIMD_INLINE void c_v256_store_aligned(void *p, c_v256 a) { + if (SIMD_CHECK && (uintptr_t)p & 31) { + fprintf(stderr, "Error: unaligned v256 store at %p\n", p); + abort(); + } + c_v256_store_unaligned(p, a); +} + +SIMD_INLINE c_v256 c_v256_zero() { + c_v256 t; + t.u64[3] = t.u64[2] = t.u64[1] = t.u64[0] = 0; + return t; +} + +SIMD_INLINE c_v256 c_v256_dup_8(uint8_t x) { + c_v256 t; + t.v64[3] = t.v64[2] = t.v64[1] = t.v64[0] = c_v64_dup_8(x); + return t; +} + +SIMD_INLINE c_v256 c_v256_dup_16(uint16_t x) { + c_v256 t; + t.v64[3] = t.v64[2] = t.v64[1] = t.v64[0] = c_v64_dup_16(x); + return t; +} + +SIMD_INLINE c_v256 c_v256_dup_32(uint32_t x) { + c_v256 t; + t.v64[3] = t.v64[2] = t.v64[1] = t.v64[0] = c_v64_dup_32(x); + return t; +} + +SIMD_INLINE c_v256 c_v256_dup_64(uint64_t x) { + c_v256 t; + t.u64[3] = t.u64[2] = t.u64[1] = t.u64[0] = x; + return t; +} + +SIMD_INLINE int64_t c_v256_dotp_su8(c_v256 a, c_v256 b) { + return c_v128_dotp_su8(a.v128[1], b.v128[1]) + + c_v128_dotp_su8(a.v128[0], b.v128[0]); +} + +SIMD_INLINE int64_t c_v256_dotp_s16(c_v256 a, c_v256 b) { + return c_v128_dotp_s16(a.v128[1], b.v128[1]) + + c_v128_dotp_s16(a.v128[0], b.v128[0]); +} + +SIMD_INLINE int64_t c_v256_dotp_s32(c_v256 a, c_v256 b) { + return c_v128_dotp_s32(a.v128[1], b.v128[1]) + + c_v128_dotp_s32(a.v128[0], b.v128[0]); +} + +SIMD_INLINE uint64_t c_v256_hadd_u8(c_v256 a) { + return c_v128_hadd_u8(a.v128[1]) + c_v128_hadd_u8(a.v128[0]); +} + +typedef uint32_t c_sad256_internal; + +SIMD_INLINE c_sad256_internal c_v256_sad_u8_init() { return 0; } + +/* Implementation dependent return value. Result must be finalised with + v256_sad_u8_sum(). + The result for more than 16 v256_sad_u8() calls is undefined. */ +SIMD_INLINE c_sad256_internal c_v256_sad_u8(c_sad256_internal s, c_v256 a, + c_v256 b) { + int c; + for (c = 0; c < 32; c++) + s += a.u8[c] > b.u8[c] ? a.u8[c] - b.u8[c] : b.u8[c] - a.u8[c]; + return s; +} + +SIMD_INLINE uint32_t c_v256_sad_u8_sum(c_sad256_internal s) { return s; } + +typedef uint32_t c_ssd256_internal; + +SIMD_INLINE c_ssd256_internal c_v256_ssd_u8_init() { return 0; } + +/* Implementation dependent return value. Result must be finalised with + * v256_ssd_u8_sum(). */ +SIMD_INLINE c_ssd256_internal c_v256_ssd_u8(c_ssd256_internal s, c_v256 a, + c_v256 b) { + int c; + for (c = 0; c < 32; c++) s += (a.u8[c] - b.u8[c]) * (a.u8[c] - b.u8[c]); + return s; +} + +SIMD_INLINE uint32_t c_v256_ssd_u8_sum(c_ssd256_internal s) { return s; } + +SIMD_INLINE c_v256 c_v256_or(c_v256 a, c_v256 b) { + return c_v256_from_v128(c_v128_or(a.v128[1], b.v128[1]), + c_v128_or(a.v128[0], b.v128[0])); +} + +SIMD_INLINE c_v256 c_v256_xor(c_v256 a, c_v256 b) { + return c_v256_from_v128(c_v128_xor(a.v128[1], b.v128[1]), + c_v128_xor(a.v128[0], b.v128[0])); +} + +SIMD_INLINE c_v256 c_v256_and(c_v256 a, c_v256 b) { + return c_v256_from_v128(c_v128_and(a.v128[1], b.v128[1]), + c_v128_and(a.v128[0], b.v128[0])); +} + +SIMD_INLINE c_v256 c_v256_andn(c_v256 a, c_v256 b) { + return c_v256_from_v128(c_v128_andn(a.v128[1], b.v128[1]), + c_v128_andn(a.v128[0], b.v128[0])); +} + +SIMD_INLINE c_v256 c_v256_add_8(c_v256 a, c_v256 b) { + return c_v256_from_v128(c_v128_add_8(a.v128[1], b.v128[1]), + c_v128_add_8(a.v128[0], b.v128[0])); +} + +SIMD_INLINE c_v256 c_v256_add_16(c_v256 a, c_v256 b) { + return c_v256_from_v128(c_v128_add_16(a.v128[1], b.v128[1]), + c_v128_add_16(a.v128[0], b.v128[0])); +} + +SIMD_INLINE c_v256 c_v256_sadd_s8(c_v256 a, c_v256 b) { + return c_v256_from_v128(c_v128_sadd_s8(a.v128[1], b.v128[1]), + c_v128_sadd_s8(a.v128[0], b.v128[0])); +} + +SIMD_INLINE c_v256 c_v256_sadd_u8(c_v256 a, c_v256 b) { + return c_v256_from_v128(c_v128_sadd_u8(a.v128[1], b.v128[1]), + c_v128_sadd_u8(a.v128[0], b.v128[0])); +} + +SIMD_INLINE c_v256 c_v256_sadd_s16(c_v256 a, c_v256 b) { + return c_v256_from_v128(c_v128_sadd_s16(a.v128[1], b.v128[1]), + c_v128_sadd_s16(a.v128[0], b.v128[0])); +} + +SIMD_INLINE c_v256 c_v256_add_32(c_v256 a, c_v256 b) { + return c_v256_from_v128(c_v128_add_32(a.v128[1], b.v128[1]), + c_v128_add_32(a.v128[0], b.v128[0])); +} + +SIMD_INLINE c_v256 c_v256_add_64(c_v256 a, c_v256 b) { + return c_v256_from_v128(c_v128_add_64(a.v128[1], b.v128[1]), + c_v128_add_64(a.v128[0], b.v128[0])); +} + +SIMD_INLINE c_v256 c_v256_sub_64(c_v256 a, c_v256 b) { + return c_v256_from_v128(c_v128_sub_64(a.v128[1], b.v128[1]), + c_v128_sub_64(a.v128[0], b.v128[0])); +} + +SIMD_INLINE c_v256 c_v256_padd_u8(c_v256 a) { + c_v256 t; + for (int i = 0; i < 16; i++) + t.u16[i] = (uint16_t)a.u8[i * 2] + (uint16_t)a.u8[i * 2 + 1]; + return t; +} + +SIMD_INLINE c_v256 c_v256_padd_s16(c_v256 a) { + c_v256 t; + t.s32[0] = (int32_t)a.s16[0] + (int32_t)a.s16[1]; + t.s32[1] = (int32_t)a.s16[2] + (int32_t)a.s16[3]; + t.s32[2] = (int32_t)a.s16[4] + (int32_t)a.s16[5]; + t.s32[3] = (int32_t)a.s16[6] + (int32_t)a.s16[7]; + t.s32[4] = (int32_t)a.s16[8] + (int32_t)a.s16[9]; + t.s32[5] = (int32_t)a.s16[10] + (int32_t)a.s16[11]; + t.s32[6] = (int32_t)a.s16[12] + (int32_t)a.s16[13]; + t.s32[7] = (int32_t)a.s16[14] + (int32_t)a.s16[15]; + return t; +} + +SIMD_INLINE c_v256 c_v256_sub_8(c_v256 a, c_v256 b) { + return c_v256_from_v128(c_v128_sub_8(a.v128[1], b.v128[1]), + c_v128_sub_8(a.v128[0], b.v128[0])); +} + +SIMD_INLINE c_v256 c_v256_ssub_u8(c_v256 a, c_v256 b) { + return c_v256_from_v128(c_v128_ssub_u8(a.v128[1], b.v128[1]), + c_v128_ssub_u8(a.v128[0], b.v128[0])); +} + +SIMD_INLINE c_v256 c_v256_ssub_s8(c_v256 a, c_v256 b) { + return c_v256_from_v128(c_v128_ssub_s8(a.v128[1], b.v128[1]), + c_v128_ssub_s8(a.v128[0], b.v128[0])); +} + +SIMD_INLINE c_v256 c_v256_sub_16(c_v256 a, c_v256 b) { + return c_v256_from_v128(c_v128_sub_16(a.v128[1], b.v128[1]), + c_v128_sub_16(a.v128[0], b.v128[0])); +} + +SIMD_INLINE c_v256 c_v256_ssub_s16(c_v256 a, c_v256 b) { + return c_v256_from_v128(c_v128_ssub_s16(a.v128[1], b.v128[1]), + c_v128_ssub_s16(a.v128[0], b.v128[0])); +} + +SIMD_INLINE c_v256 c_v256_ssub_u16(c_v256 a, c_v256 b) { + return c_v256_from_v128(c_v128_ssub_u16(a.v128[1], b.v128[1]), + c_v128_ssub_u16(a.v128[0], b.v128[0])); +} + +SIMD_INLINE c_v256 c_v256_sub_32(c_v256 a, c_v256 b) { + return c_v256_from_v128(c_v128_sub_32(a.v128[1], b.v128[1]), + c_v128_sub_32(a.v128[0], b.v128[0])); +} + +SIMD_INLINE c_v256 c_v256_abs_s16(c_v256 a) { + return c_v256_from_v128(c_v128_abs_s16(a.v128[1]), c_v128_abs_s16(a.v128[0])); +} + +SIMD_INLINE c_v256 c_v256_abs_s8(c_v256 a) { + return c_v256_from_v128(c_v128_abs_s8(a.v128[1]), c_v128_abs_s8(a.v128[0])); +} + +SIMD_INLINE c_v256 c_v256_mul_s16(c_v128 a, c_v128 b) { + c_v128 lo_bits = c_v128_mullo_s16(a, b); + c_v128 hi_bits = c_v128_mulhi_s16(a, b); + return c_v256_from_v128(c_v128_ziphi_16(hi_bits, lo_bits), + c_v128_ziplo_16(hi_bits, lo_bits)); +} + +SIMD_INLINE c_v256 c_v256_mullo_s16(c_v256 a, c_v256 b) { + return c_v256_from_v128(c_v128_mullo_s16(a.v128[1], b.v128[1]), + c_v128_mullo_s16(a.v128[0], b.v128[0])); +} + +SIMD_INLINE c_v256 c_v256_mulhi_s16(c_v256 a, c_v256 b) { + return c_v256_from_v128(c_v128_mulhi_s16(a.v128[1], b.v128[1]), + c_v128_mulhi_s16(a.v128[0], b.v128[0])); +} + +SIMD_INLINE c_v256 c_v256_mullo_s32(c_v256 a, c_v256 b) { + return c_v256_from_v128(c_v128_mullo_s32(a.v128[1], b.v128[1]), + c_v128_mullo_s32(a.v128[0], b.v128[0])); +} + +SIMD_INLINE c_v256 c_v256_madd_s16(c_v256 a, c_v256 b) { + return c_v256_from_v128(c_v128_madd_s16(a.v128[1], b.v128[1]), + c_v128_madd_s16(a.v128[0], b.v128[0])); +} + +SIMD_INLINE c_v256 c_v256_madd_us8(c_v256 a, c_v256 b) { + return c_v256_from_v128(c_v128_madd_us8(a.v128[1], b.v128[1]), + c_v128_madd_us8(a.v128[0], b.v128[0])); +} + +SIMD_INLINE c_v256 c_v256_avg_u8(c_v256 a, c_v256 b) { + return c_v256_from_v128(c_v128_avg_u8(a.v128[1], b.v128[1]), + c_v128_avg_u8(a.v128[0], b.v128[0])); +} + +SIMD_INLINE c_v256 c_v256_rdavg_u8(c_v256 a, c_v256 b) { + return c_v256_from_v128(c_v128_rdavg_u8(a.v128[1], b.v128[1]), + c_v128_rdavg_u8(a.v128[0], b.v128[0])); +} + +SIMD_INLINE c_v256 c_v256_rdavg_u16(c_v256 a, c_v256 b) { + return c_v256_from_v128(c_v128_rdavg_u16(a.v128[1], b.v128[1]), + c_v128_rdavg_u16(a.v128[0], b.v128[0])); +} + +SIMD_INLINE c_v256 c_v256_avg_u16(c_v256 a, c_v256 b) { + return c_v256_from_v128(c_v128_avg_u16(a.v128[1], b.v128[1]), + c_v128_avg_u16(a.v128[0], b.v128[0])); +} + +SIMD_INLINE c_v256 c_v256_min_u8(c_v256 a, c_v256 b) { + return c_v256_from_v128(c_v128_min_u8(a.v128[1], b.v128[1]), + c_v128_min_u8(a.v128[0], b.v128[0])); +} + +SIMD_INLINE c_v256 c_v256_max_u8(c_v256 a, c_v256 b) { + return c_v256_from_v128(c_v128_max_u8(a.v128[1], b.v128[1]), + c_v128_max_u8(a.v128[0], b.v128[0])); +} + +SIMD_INLINE c_v256 c_v256_min_s8(c_v256 a, c_v256 b) { + return c_v256_from_v128(c_v128_min_s8(a.v128[1], b.v128[1]), + c_v128_min_s8(a.v128[0], b.v128[0])); +} + +SIMD_INLINE uint32_t c_v256_movemask_8(c_v256 a) { + return ((a.s8[31] < 0) << 31) | ((a.s8[30] < 0) << 30) | + ((a.s8[29] < 0) << 29) | ((a.s8[28] < 0) << 28) | + ((a.s8[27] < 0) << 27) | ((a.s8[26] < 0) << 26) | + ((a.s8[25] < 0) << 25) | ((a.s8[24] < 0) << 24) | + ((a.s8[23] < 0) << 23) | ((a.s8[22] < 0) << 22) | + ((a.s8[21] < 0) << 21) | ((a.s8[20] < 0) << 20) | + ((a.s8[19] < 0) << 19) | ((a.s8[18] < 0) << 18) | + ((a.s8[17] < 0) << 17) | ((a.s8[16] < 0) << 16) | + ((a.s8[15] < 0) << 15) | ((a.s8[14] < 0) << 14) | + ((a.s8[13] < 0) << 13) | ((a.s8[12] < 0) << 12) | + ((a.s8[11] < 0) << 11) | ((a.s8[10] < 0) << 10) | + ((a.s8[9] < 0) << 9) | ((a.s8[8] < 0) << 8) | ((a.s8[7] < 0) << 7) | + ((a.s8[6] < 0) << 6) | ((a.s8[5] < 0) << 5) | ((a.s8[4] < 0) << 4) | + ((a.s8[3] < 0) << 3) | ((a.s8[2] < 0) << 2) | ((a.s8[1] < 0) << 1) | + ((a.s8[0] < 0) << 0); +} + +SIMD_INLINE c_v256 c_v256_blend_8(c_v256 a, c_v256 b, c_v256 c) { + c_v256 t; + for (int i = 0; i < 32; i++) t.u8[i] = c.s8[i] < 0 ? b.u8[i] : a.u8[i]; + return t; +} + +SIMD_INLINE c_v256 c_v256_max_s8(c_v256 a, c_v256 b) { + return c_v256_from_v128(c_v128_max_s8(a.v128[1], b.v128[1]), + c_v128_max_s8(a.v128[0], b.v128[0])); +} + +SIMD_INLINE c_v256 c_v256_min_s16(c_v256 a, c_v256 b) { + return c_v256_from_v128(c_v128_min_s16(a.v128[1], b.v128[1]), + c_v128_min_s16(a.v128[0], b.v128[0])); +} + +SIMD_INLINE c_v256 c_v256_max_s16(c_v256 a, c_v256 b) { + return c_v256_from_v128(c_v128_max_s16(a.v128[1], b.v128[1]), + c_v128_max_s16(a.v128[0], b.v128[0])); +} + +SIMD_INLINE c_v256 c_v256_min_s32(c_v256 a, c_v256 b) { + return c_v256_from_v128(c_v128_min_s32(a.v128[1], b.v128[1]), + c_v128_min_s32(a.v128[0], b.v128[0])); +} + +SIMD_INLINE c_v256 c_v256_max_s32(c_v256 a, c_v256 b) { + return c_v256_from_v128(c_v128_max_s32(a.v128[1], b.v128[1]), + c_v128_max_s32(a.v128[0], b.v128[0])); +} + +SIMD_INLINE c_v256 c_v256_ziplo_8(c_v256 a, c_v256 b) { + return c_v256_from_v128(c_v128_ziphi_8(a.v128[0], b.v128[0]), + c_v128_ziplo_8(a.v128[0], b.v128[0])); +} + +SIMD_INLINE c_v256 c_v256_ziphi_8(c_v256 a, c_v256 b) { + return c_v256_from_v128(c_v128_ziphi_8(a.v128[1], b.v128[1]), + c_v128_ziplo_8(a.v128[1], b.v128[1])); +} + +SIMD_INLINE c_v256 c_v256_ziplo_16(c_v256 a, c_v256 b) { + return c_v256_from_v128(c_v128_ziphi_16(a.v128[0], b.v128[0]), + c_v128_ziplo_16(a.v128[0], b.v128[0])); +} + +SIMD_INLINE c_v256 c_v256_ziphi_16(c_v256 a, c_v256 b) { + return c_v256_from_v128(c_v128_ziphi_16(a.v128[1], b.v128[1]), + c_v128_ziplo_16(a.v128[1], b.v128[1])); +} + +SIMD_INLINE c_v256 c_v256_ziplo_32(c_v256 a, c_v256 b) { + return c_v256_from_v128(c_v128_ziphi_32(a.v128[0], b.v128[0]), + c_v128_ziplo_32(a.v128[0], b.v128[0])); +} + +SIMD_INLINE c_v256 c_v256_ziphi_32(c_v256 a, c_v256 b) { + return c_v256_from_v128(c_v128_ziphi_32(a.v128[1], b.v128[1]), + c_v128_ziplo_32(a.v128[1], b.v128[1])); +} + +SIMD_INLINE c_v256 c_v256_ziplo_64(c_v256 a, c_v256 b) { + return c_v256_from_v128(c_v128_ziphi_64(a.v128[0], b.v128[0]), + c_v128_ziplo_64(a.v128[0], b.v128[0])); +} + +SIMD_INLINE c_v256 c_v256_ziphi_64(c_v256 a, c_v256 b) { + return c_v256_from_v128(c_v128_ziphi_64(a.v128[1], b.v128[1]), + c_v128_ziplo_64(a.v128[1], b.v128[1])); +} + +SIMD_INLINE c_v256 c_v256_ziplo_128(c_v256 a, c_v256 b) { + return c_v256_from_v128(a.v128[0], b.v128[0]); +} + +SIMD_INLINE c_v256 c_v256_ziphi_128(c_v256 a, c_v256 b) { + return c_v256_from_v128(a.v128[1], b.v128[1]); +} + +SIMD_INLINE c_v256 c_v256_zip_8(c_v128 a, c_v128 b) { + return c_v256_from_v128(c_v128_ziphi_8(a, b), c_v128_ziplo_8(a, b)); +} + +SIMD_INLINE c_v256 c_v256_zip_16(c_v128 a, c_v128 b) { + return c_v256_from_v128(c_v128_ziphi_16(a, b), c_v128_ziplo_16(a, b)); +} + +SIMD_INLINE c_v256 c_v256_zip_32(c_v128 a, c_v128 b) { + return c_v256_from_v128(c_v128_ziphi_32(a, b), c_v128_ziplo_32(a, b)); +} + +SIMD_INLINE c_v256 _c_v256_unzip_8(c_v256 a, c_v256 b, int mode) { + c_v256 t; + int i; + if (mode) { + for (i = 0; i < 16; i++) { + t.u8[i] = a.u8[i * 2 + 1]; + t.u8[i + 16] = b.u8[i * 2 + 1]; + } + } else { + for (i = 0; i < 16; i++) { + t.u8[i] = b.u8[i * 2]; + t.u8[i + 16] = a.u8[i * 2]; + } + } + return t; +} + +SIMD_INLINE c_v256 c_v256_unziplo_8(c_v256 a, c_v256 b) { + return CONFIG_BIG_ENDIAN ? _c_v256_unzip_8(a, b, 1) + : _c_v256_unzip_8(a, b, 0); +} + +SIMD_INLINE c_v256 c_v256_unziphi_8(c_v256 a, c_v256 b) { + return CONFIG_BIG_ENDIAN ? _c_v256_unzip_8(b, a, 0) + : _c_v256_unzip_8(b, a, 1); +} + +SIMD_INLINE c_v256 _c_v256_unzip_16(c_v256 a, c_v256 b, int mode) { + c_v256 t; + int i; + if (mode) { + for (i = 0; i < 8; i++) { + t.u16[i] = a.u16[i * 2 + 1]; + t.u16[i + 8] = b.u16[i * 2 + 1]; + } + } else { + for (i = 0; i < 8; i++) { + t.u16[i] = b.u16[i * 2]; + t.u16[i + 8] = a.u16[i * 2]; + } + } + return t; +} + +SIMD_INLINE c_v256 c_v256_unziplo_16(c_v256 a, c_v256 b) { + return CONFIG_BIG_ENDIAN ? _c_v256_unzip_16(a, b, 1) + : _c_v256_unzip_16(a, b, 0); +} + +SIMD_INLINE c_v256 c_v256_unziphi_16(c_v256 a, c_v256 b) { + return CONFIG_BIG_ENDIAN ? _c_v256_unzip_16(b, a, 0) + : _c_v256_unzip_16(b, a, 1); +} + +SIMD_INLINE c_v256 _c_v256_unzip_32(c_v256 a, c_v256 b, int mode) { + c_v256 t; + if (mode) { + t.u32[7] = b.u32[7]; + t.u32[6] = b.u32[5]; + t.u32[5] = b.u32[3]; + t.u32[4] = b.u32[1]; + t.u32[3] = a.u32[7]; + t.u32[2] = a.u32[5]; + t.u32[1] = a.u32[3]; + t.u32[0] = a.u32[1]; + } else { + t.u32[7] = a.u32[6]; + t.u32[6] = a.u32[4]; + t.u32[5] = a.u32[2]; + t.u32[4] = a.u32[0]; + t.u32[3] = b.u32[6]; + t.u32[2] = b.u32[4]; + t.u32[1] = b.u32[2]; + t.u32[0] = b.u32[0]; + } + return t; +} + +SIMD_INLINE c_v256 c_v256_unziplo_32(c_v256 a, c_v256 b) { + return CONFIG_BIG_ENDIAN ? _c_v256_unzip_32(a, b, 1) + : _c_v256_unzip_32(a, b, 0); +} + +SIMD_INLINE c_v256 c_v256_unziphi_32(c_v256 a, c_v256 b) { + return CONFIG_BIG_ENDIAN ? _c_v256_unzip_32(b, a, 0) + : _c_v256_unzip_32(b, a, 1); +} + +SIMD_INLINE c_v256 _c_v256_unzip_64(c_v256 a, c_v256 b, int mode) { + c_v256 t; + if (mode) { + t.u64[3] = b.u64[3]; + t.u64[2] = b.u64[1]; + t.u64[1] = a.u64[3]; + t.u64[0] = a.u64[1]; + } else { + t.u64[3] = a.u64[2]; + t.u64[2] = a.u64[0]; + t.u64[1] = b.u64[2]; + t.u64[0] = b.u64[0]; + } + return t; +} + +SIMD_INLINE c_v256 c_v256_unziplo_64(c_v256 a, c_v256 b) { + return CONFIG_BIG_ENDIAN ? _c_v256_unzip_64(a, b, 1) + : _c_v256_unzip_64(a, b, 0); +} + +SIMD_INLINE c_v256 c_v256_unziphi_64(c_v256 a, c_v256 b) { + return CONFIG_BIG_ENDIAN ? _c_v256_unzip_64(b, a, 0) + : _c_v256_unzip_64(b, a, 1); +} + +SIMD_INLINE c_v256 c_v256_unpack_u8_s16(c_v128 a) { + return c_v256_from_v128(c_v128_unpackhi_u8_s16(a), c_v128_unpacklo_u8_s16(a)); +} + +SIMD_INLINE c_v256 c_v256_unpacklo_u8_s16(c_v256 a) { + return c_v256_from_v128(c_v128_unpackhi_u8_s16(a.v128[0]), + c_v128_unpacklo_u8_s16(a.v128[0])); +} + +SIMD_INLINE c_v256 c_v256_unpackhi_u8_s16(c_v256 a) { + return c_v256_from_v128(c_v128_unpackhi_u8_s16(a.v128[1]), + c_v128_unpacklo_u8_s16(a.v128[1])); +} + +SIMD_INLINE c_v256 c_v256_unpack_s8_s16(c_v128 a) { + return c_v256_from_v128(c_v128_unpackhi_s8_s16(a), c_v128_unpacklo_s8_s16(a)); +} + +SIMD_INLINE c_v256 c_v256_unpacklo_s8_s16(c_v256 a) { + return c_v256_from_v128(c_v128_unpackhi_s8_s16(a.v128[0]), + c_v128_unpacklo_s8_s16(a.v128[0])); +} + +SIMD_INLINE c_v256 c_v256_unpackhi_s8_s16(c_v256 a) { + return c_v256_from_v128(c_v128_unpackhi_s8_s16(a.v128[1]), + c_v128_unpacklo_s8_s16(a.v128[1])); +} + +SIMD_INLINE c_v256 c_v256_pack_s32_s16(c_v256 a, c_v256 b) { + return c_v256_from_v128(c_v128_pack_s32_s16(a.v128[1], a.v128[0]), + c_v128_pack_s32_s16(b.v128[1], b.v128[0])); +} + +SIMD_INLINE c_v256 c_v256_pack_s32_u16(c_v256 a, c_v256 b) { + return c_v256_from_v128(c_v128_pack_s32_u16(a.v128[1], a.v128[0]), + c_v128_pack_s32_u16(b.v128[1], b.v128[0])); +} + +SIMD_INLINE c_v256 c_v256_pack_s16_u8(c_v256 a, c_v256 b) { + return c_v256_from_v128(c_v128_pack_s16_u8(a.v128[1], a.v128[0]), + c_v128_pack_s16_u8(b.v128[1], b.v128[0])); +} + +SIMD_INLINE c_v256 c_v256_pack_s16_s8(c_v256 a, c_v256 b) { + return c_v256_from_v128(c_v128_pack_s16_s8(a.v128[1], a.v128[0]), + c_v128_pack_s16_s8(b.v128[1], b.v128[0])); +} + +SIMD_INLINE c_v256 c_v256_unpack_u16_s32(c_v128 a) { + return c_v256_from_v128(c_v128_unpackhi_u16_s32(a), + c_v128_unpacklo_u16_s32(a)); +} + +SIMD_INLINE c_v256 c_v256_unpack_s16_s32(c_v128 a) { + return c_v256_from_v128(c_v128_unpackhi_s16_s32(a), + c_v128_unpacklo_s16_s32(a)); +} + +SIMD_INLINE c_v256 c_v256_unpacklo_u16_s32(c_v256 a) { + return c_v256_from_v128(c_v128_unpackhi_u16_s32(a.v128[0]), + c_v128_unpacklo_u16_s32(a.v128[0])); +} + +SIMD_INLINE c_v256 c_v256_unpacklo_s16_s32(c_v256 a) { + return c_v256_from_v128(c_v128_unpackhi_s16_s32(a.v128[0]), + c_v128_unpacklo_s16_s32(a.v128[0])); +} + +SIMD_INLINE c_v256 c_v256_unpackhi_u16_s32(c_v256 a) { + return c_v256_from_v128(c_v128_unpackhi_u16_s32(a.v128[1]), + c_v128_unpacklo_u16_s32(a.v128[1])); +} + +SIMD_INLINE c_v256 c_v256_unpackhi_s16_s32(c_v256 a) { + return c_v256_from_v128(c_v128_unpackhi_s16_s32(a.v128[1]), + c_v128_unpacklo_s16_s32(a.v128[1])); +} + +SIMD_INLINE c_v256 c_v256_shuffle_8(c_v256 a, c_v256 pattern) { + c_v256 t; + int c; + for (c = 0; c < 32; c++) + t.u8[c] = a.u8[CONFIG_BIG_ENDIAN ? 31 - (pattern.u8[c] & 31) + : pattern.u8[c] & 31]; + + return t; +} + +SIMD_INLINE c_v256 c_v256_wideshuffle_8(c_v256 a, c_v256 b, c_v256 pattern) { + c_v256 t; + int c; + for (c = 0; c < 32; c++) + t.u8[c] = (pattern.u8[c] < 32 + ? b.u8 + : a.u8)[CONFIG_BIG_ENDIAN ? 31 - (pattern.u8[c] & 31) + : pattern.u8[c] & 31]; + return t; +} + +// Pairwise / dual-lane shuffle: shuffle two 128 bit lates. +SIMD_INLINE c_v256 c_v256_pshuffle_8(c_v256 a, c_v256 pattern) { + return c_v256_from_v128( + c_v128_shuffle_8(c_v256_high_v128(a), c_v256_high_v128(pattern)), + c_v128_shuffle_8(c_v256_low_v128(a), c_v256_low_v128(pattern))); +} + +SIMD_INLINE c_v256 c_v256_cmpgt_s8(c_v256 a, c_v256 b) { + return c_v256_from_v128(c_v128_cmpgt_s8(a.v128[1], b.v128[1]), + c_v128_cmpgt_s8(a.v128[0], b.v128[0])); +} + +SIMD_INLINE c_v256 c_v256_cmplt_s8(c_v256 a, c_v256 b) { + return c_v256_from_v128(c_v128_cmplt_s8(a.v128[1], b.v128[1]), + c_v128_cmplt_s8(a.v128[0], b.v128[0])); +} + +SIMD_INLINE c_v256 c_v256_cmpeq_8(c_v256 a, c_v256 b) { + return c_v256_from_v128(c_v128_cmpeq_8(a.v128[1], b.v128[1]), + c_v128_cmpeq_8(a.v128[0], b.v128[0])); +} + +SIMD_INLINE c_v256 c_v256_cmpgt_s16(c_v256 a, c_v256 b) { + return c_v256_from_v128(c_v128_cmpgt_s16(a.v128[1], b.v128[1]), + c_v128_cmpgt_s16(a.v128[0], b.v128[0])); +} + +SIMD_INLINE c_v256 c_v256_cmplt_s16(c_v256 a, c_v256 b) { + return c_v256_from_v128(c_v128_cmplt_s16(a.v128[1], b.v128[1]), + c_v128_cmplt_s16(a.v128[0], b.v128[0])); +} + +SIMD_INLINE c_v256 c_v256_cmpeq_16(c_v256 a, c_v256 b) { + return c_v256_from_v128(c_v128_cmpeq_16(a.v128[1], b.v128[1]), + c_v128_cmpeq_16(a.v128[0], b.v128[0])); +} + +SIMD_INLINE c_v256 c_v256_cmpgt_s32(c_v256 a, c_v256 b) { + return c_v256_from_v128(c_v128_cmpgt_s32(a.v128[1], b.v128[1]), + c_v128_cmpgt_s32(a.v128[0], b.v128[0])); +} + +SIMD_INLINE c_v256 c_v256_cmplt_s32(c_v256 a, c_v256 b) { + return c_v256_from_v128(c_v128_cmplt_s32(a.v128[1], b.v128[1]), + c_v128_cmplt_s32(a.v128[0], b.v128[0])); +} + +SIMD_INLINE c_v256 c_v256_cmpeq_32(c_v256 a, c_v256 b) { + return c_v256_from_v128(c_v128_cmpeq_32(a.v128[1], b.v128[1]), + c_v128_cmpeq_32(a.v128[0], b.v128[0])); +} + +SIMD_INLINE c_v256 c_v256_shl_n_byte(c_v256 a, unsigned int n) { + if (n < 16) + return c_v256_from_v128(c_v128_or(c_v128_shl_n_byte(a.v128[1], n), + c_v128_shr_n_byte(a.v128[0], 16 - n)), + c_v128_shl_n_byte(a.v128[0], n)); + else if (n > 16) + return c_v256_from_v128(c_v128_shl_n_byte(a.v128[0], n - 16), + c_v128_zero()); + else + return c_v256_from_v128(c_v256_low_v128(a), c_v128_zero()); +} + +SIMD_INLINE c_v256 c_v256_shr_n_byte(c_v256 a, unsigned int n) { + if (n < 16) + return c_v256_from_v128(c_v128_shr_n_byte(a.v128[1], n), + c_v128_or(c_v128_shr_n_byte(a.v128[0], n), + c_v128_shl_n_byte(a.v128[1], 16 - n))); + else if (n > 16) + return c_v256_from_v128(c_v128_zero(), + c_v128_shr_n_byte(a.v128[1], n - 16)); + else + return c_v256_from_v128(c_v128_zero(), c_v256_high_v128(a)); +} + +SIMD_INLINE c_v256 c_v256_align(c_v256 a, c_v256 b, unsigned int c) { + if (SIMD_CHECK && c > 31) { + fprintf(stderr, "Error: undefined alignment %d\n", c); + abort(); + } + return c ? c_v256_or(c_v256_shr_n_byte(b, c), c_v256_shl_n_byte(a, 32 - c)) + : b; +} + +SIMD_INLINE c_v256 c_v256_shl_8(c_v256 a, unsigned int c) { + return c_v256_from_v128(c_v128_shl_8(a.v128[1], c), + c_v128_shl_8(a.v128[0], c)); +} + +SIMD_INLINE c_v256 c_v256_shr_u8(c_v256 a, unsigned int c) { + return c_v256_from_v128(c_v128_shr_u8(a.v128[1], c), + c_v128_shr_u8(a.v128[0], c)); +} + +SIMD_INLINE c_v256 c_v256_shr_s8(c_v256 a, unsigned int c) { + return c_v256_from_v128(c_v128_shr_s8(a.v128[1], c), + c_v128_shr_s8(a.v128[0], c)); +} + +SIMD_INLINE c_v256 c_v256_shl_16(c_v256 a, unsigned int c) { + return c_v256_from_v128(c_v128_shl_16(a.v128[1], c), + c_v128_shl_16(a.v128[0], c)); +} + +SIMD_INLINE c_v256 c_v256_shr_u16(c_v256 a, unsigned int c) { + return c_v256_from_v128(c_v128_shr_u16(a.v128[1], c), + c_v128_shr_u16(a.v128[0], c)); +} + +SIMD_INLINE c_v256 c_v256_shr_s16(c_v256 a, unsigned int c) { + return c_v256_from_v128(c_v128_shr_s16(a.v128[1], c), + c_v128_shr_s16(a.v128[0], c)); +} + +SIMD_INLINE c_v256 c_v256_shl_32(c_v256 a, unsigned int c) { + return c_v256_from_v128(c_v128_shl_32(a.v128[1], c), + c_v128_shl_32(a.v128[0], c)); +} + +SIMD_INLINE c_v256 c_v256_shr_u32(c_v256 a, unsigned int c) { + return c_v256_from_v128(c_v128_shr_u32(a.v128[1], c), + c_v128_shr_u32(a.v128[0], c)); +} + +SIMD_INLINE c_v256 c_v256_shr_s32(c_v256 a, unsigned int c) { + return c_v256_from_v128(c_v128_shr_s32(a.v128[1], c), + c_v128_shr_s32(a.v128[0], c)); +} + +SIMD_INLINE c_v256 c_v256_shr_s64(c_v256 a, unsigned int n) { + c_v256 t; + if (SIMD_CHECK && n > 63) { + fprintf(stderr, "Error: undefined s64 shift right %d\n", n); + abort(); + } + t.s64[3] = a.s64[3] >> n; + t.s64[2] = a.s64[2] >> n; + t.s64[1] = a.s64[1] >> n; + t.s64[0] = a.s64[0] >> n; + return t; +} + +SIMD_INLINE c_v256 c_v256_shr_u64(c_v256 a, unsigned int n) { + c_v256 t; + if (SIMD_CHECK && n > 63) { + fprintf(stderr, "Error: undefined s64 shift right %d\n", n); + abort(); + } + t.u64[3] = a.u64[3] >> n; + t.u64[2] = a.u64[2] >> n; + t.u64[1] = a.u64[1] >> n; + t.u64[0] = a.u64[0] >> n; + return t; +} + +SIMD_INLINE c_v256 c_v256_shl_64(c_v256 a, unsigned int n) { + c_v256 t; + if (SIMD_CHECK && n > 63) { + fprintf(stderr, "Error: undefined s64 shift right %d\n", n); + abort(); + } + t.u64[3] = a.u64[3] << n; + t.u64[2] = a.u64[2] << n; + t.u64[1] = a.u64[1] << n; + t.u64[0] = a.u64[0] << n; + return t; +} + +SIMD_INLINE c_v256 c_v256_shl_n_8(c_v256 a, unsigned int n) { + return c_v256_shl_8(a, n); +} + +SIMD_INLINE c_v256 c_v256_shl_n_16(c_v256 a, unsigned int n) { + return c_v256_shl_16(a, n); +} + +SIMD_INLINE c_v256 c_v256_shl_n_32(c_v256 a, unsigned int n) { + return c_v256_shl_32(a, n); +} + +SIMD_INLINE c_v256 c_v256_shl_n_64(c_v256 a, unsigned int n) { + return c_v256_shl_64(a, n); +} + +SIMD_INLINE c_v256 c_v256_shr_n_u8(c_v256 a, unsigned int n) { + return c_v256_shr_u8(a, n); +} + +SIMD_INLINE c_v256 c_v256_shr_n_u16(c_v256 a, unsigned int n) { + return c_v256_shr_u16(a, n); +} + +SIMD_INLINE c_v256 c_v256_shr_n_u32(c_v256 a, unsigned int n) { + return c_v256_shr_u32(a, n); +} + +SIMD_INLINE c_v256 c_v256_shr_n_u64(c_v256 a, unsigned int n) { + return c_v256_shr_u64(a, n); +} + +SIMD_INLINE c_v256 c_v256_shr_n_s8(c_v256 a, unsigned int n) { + return c_v256_shr_s8(a, n); +} + +SIMD_INLINE c_v256 c_v256_shr_n_s16(c_v256 a, unsigned int n) { + return c_v256_shr_s16(a, n); +} + +SIMD_INLINE c_v256 c_v256_shr_n_s32(c_v256 a, unsigned int n) { + return c_v256_shr_s32(a, n); +} + +SIMD_INLINE c_v256 c_v256_shr_n_s64(c_v256 a, unsigned int n) { + return c_v256_shr_s64(a, n); +} + +SIMD_INLINE c_v256 c_v256_shr_n_word(c_v256 a, const unsigned int n) { + return c_v256_shr_n_byte(a, 2 * n); +} +SIMD_INLINE c_v256 c_v256_shl_n_word(c_v256 a, const unsigned int n) { + return c_v256_shl_n_byte(a, 2 * n); +} + +typedef uint32_t c_sad256_internal_u16; + +SIMD_INLINE c_sad256_internal_u16 c_v256_sad_u16_init() { return 0; } + +/* Implementation dependent return value. Result must be finalised with + v256_sad_u16_sum(). */ +SIMD_INLINE c_sad256_internal_u16 c_v256_sad_u16(c_sad256_internal_u16 s, + c_v256 a, c_v256 b) { + int c; + for (c = 0; c < 16; c++) + s += a.u16[c] > b.u16[c] ? a.u16[c] - b.u16[c] : b.u16[c] - a.u16[c]; + return s; +} + +SIMD_INLINE uint32_t c_v256_sad_u16_sum(c_sad256_internal_u16 s) { return s; } + +typedef uint64_t c_ssd256_internal_s16; + +SIMD_INLINE c_ssd256_internal_s16 c_v256_ssd_s16_init() { return 0; } + +/* Implementation dependent return value. Result must be finalised with + * v256_ssd_s16_sum(). */ +SIMD_INLINE c_ssd256_internal_s16 c_v256_ssd_s16(c_ssd256_internal_s16 s, + c_v256 a, c_v256 b) { + int c; + for (c = 0; c < 16; c++) + s += (int32_t)(int16_t)(a.s16[c] - b.s16[c]) * + (int32_t)(int16_t)(a.s16[c] - b.s16[c]); + return s; +} + +SIMD_INLINE uint64_t c_v256_ssd_s16_sum(c_ssd256_internal_s16 s) { return s; } + +#endif // AOM_AOM_DSP_SIMD_V256_INTRINSICS_C_H_ diff --git a/third_party/aom/aom_dsp/simd/v256_intrinsics_v128.h b/third_party/aom/aom_dsp/simd/v256_intrinsics_v128.h new file mode 100644 index 0000000000..d5b7905eff --- /dev/null +++ b/third_party/aom/aom_dsp/simd/v256_intrinsics_v128.h @@ -0,0 +1,873 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_SIMD_V256_INTRINSICS_V128_H_ +#define AOM_AOM_DSP_SIMD_V256_INTRINSICS_V128_H_ + +#if HAVE_NEON +#include "aom_dsp/simd/v128_intrinsics_arm.h" +#elif HAVE_SSE2 +#include "aom_dsp/simd/v128_intrinsics_x86.h" +#else +#include "aom_dsp/simd/v128_intrinsics.h" +#endif + +#if HAVE_NEON +typedef int64x2x2_t v256; +#else +typedef struct { + v128 val[2]; +} v256; +#endif + +SIMD_INLINE uint32_t v256_low_u32(v256 a) { return v128_low_u32(a.val[0]); } + +SIMD_INLINE v64 v256_low_v64(v256 a) { return v128_low_v64(a.val[0]); } + +SIMD_INLINE uint64_t v256_low_u64(v256 a) { return v64_u64(v256_low_v64(a)); } + +SIMD_INLINE v128 v256_low_v128(v256 a) { return a.val[0]; } + +SIMD_INLINE v128 v256_high_v128(v256 a) { return a.val[1]; } + +SIMD_INLINE v256 v256_from_v128(v128 hi, v128 lo) { + v256 t; + t.val[1] = hi; + t.val[0] = lo; + return t; +} + +SIMD_INLINE v256 v256_from_64(uint64_t a, uint64_t b, uint64_t c, uint64_t d) { + return v256_from_v128(v128_from_64(a, b), v128_from_64(c, d)); +} + +SIMD_INLINE v256 v256_from_v64(v64 a, v64 b, v64 c, v64 d) { + return v256_from_v128(v128_from_v64(a, b), v128_from_v64(c, d)); +} + +SIMD_INLINE v256 v256_load_unaligned(const void *p) { + return v256_from_v128(v128_load_unaligned((uint8_t *)p + 16), + v128_load_unaligned(p)); +} + +SIMD_INLINE v256 v256_load_aligned(const void *p) { + return v256_from_v128(v128_load_aligned((uint8_t *)p + 16), + v128_load_aligned(p)); +} + +SIMD_INLINE void v256_store_unaligned(void *p, v256 a) { + v128_store_unaligned(p, a.val[0]); + v128_store_unaligned((uint8_t *)p + 16, a.val[1]); +} + +SIMD_INLINE void v256_store_aligned(void *p, v256 a) { + v128_store_aligned(p, a.val[0]); + v128_store_aligned((uint8_t *)p + 16, a.val[1]); +} + +SIMD_INLINE v256 v256_zero() { + return v256_from_v128(v128_zero(), v128_zero()); +} + +SIMD_INLINE v256 v256_dup_8(uint8_t x) { + v128 t = v128_dup_8(x); + return v256_from_v128(t, t); +} + +SIMD_INLINE v256 v256_dup_16(uint16_t x) { + v128 t = v128_dup_16(x); + return v256_from_v128(t, t); +} + +SIMD_INLINE v256 v256_dup_32(uint32_t x) { + v128 t = v128_dup_32(x); + return v256_from_v128(t, t); +} + +SIMD_INLINE v256 v256_dup_64(uint64_t x) { + v128 t = v128_dup_64(x); + return v256_from_v128(t, t); +} + +SIMD_INLINE int64_t v256_dotp_su8(v256 a, v256 b) { + return v128_dotp_su8(a.val[1], b.val[1]) + v128_dotp_su8(a.val[0], b.val[0]); +} + +SIMD_INLINE int64_t v256_dotp_s16(v256 a, v256 b) { + return v128_dotp_s16(a.val[1], b.val[1]) + v128_dotp_s16(a.val[0], b.val[0]); +} + +SIMD_INLINE int64_t v256_dotp_s32(v256 a, v256 b) { + return v128_dotp_s32(a.val[1], b.val[1]) + v128_dotp_s32(a.val[0], b.val[0]); +} + +SIMD_INLINE uint64_t v256_hadd_u8(v256 a) { + return v128_hadd_u8(a.val[1]) + v128_hadd_u8(a.val[0]); +} + +typedef struct { + sad128_internal val[2]; +} sad256_internal; + +SIMD_INLINE sad256_internal v256_sad_u8_init() { + sad256_internal t; + t.val[1] = v128_sad_u8_init(); + t.val[0] = v128_sad_u8_init(); + return t; +} + +/* Implementation dependent return value. Result must be finalised with + v256_sad_u8_sum(). + The result for more than 16 v256_sad_u8() calls is undefined. */ +SIMD_INLINE sad256_internal v256_sad_u8(sad256_internal s, v256 a, v256 b) { + sad256_internal t; + t.val[1] = v128_sad_u8(s.val[1], a.val[1], b.val[1]); + t.val[0] = v128_sad_u8(s.val[0], a.val[0], b.val[0]); + return t; +} + +SIMD_INLINE uint32_t v256_sad_u8_sum(sad256_internal s) { + return v128_sad_u8_sum(s.val[1]) + v128_sad_u8_sum(s.val[0]); +} + +typedef struct { + ssd128_internal val[2]; +} ssd256_internal; + +SIMD_INLINE ssd256_internal v256_ssd_u8_init() { + ssd256_internal t; + t.val[1] = v128_ssd_u8_init(); + t.val[0] = v128_ssd_u8_init(); + return t; +} + +/* Implementation dependent return value. Result must be finalised with + * v256_ssd_u8_sum(). */ +SIMD_INLINE ssd256_internal v256_ssd_u8(ssd256_internal s, v256 a, v256 b) { + ssd256_internal t; + t.val[1] = v128_ssd_u8(s.val[1], a.val[1], b.val[1]); + t.val[0] = v128_ssd_u8(s.val[0], a.val[0], b.val[0]); + return t; +} + +SIMD_INLINE uint32_t v256_ssd_u8_sum(ssd256_internal s) { + return v128_ssd_u8_sum(s.val[1]) + v128_ssd_u8_sum(s.val[0]); +} + +SIMD_INLINE v256 v256_or(v256 a, v256 b) { + return v256_from_v128(v128_or(a.val[1], b.val[1]), + v128_or(a.val[0], b.val[0])); +} + +SIMD_INLINE v256 v256_xor(v256 a, v256 b) { + return v256_from_v128(v128_xor(a.val[1], b.val[1]), + v128_xor(a.val[0], b.val[0])); +} + +SIMD_INLINE v256 v256_and(v256 a, v256 b) { + return v256_from_v128(v128_and(a.val[1], b.val[1]), + v128_and(a.val[0], b.val[0])); +} + +SIMD_INLINE v256 v256_andn(v256 a, v256 b) { + return v256_from_v128(v128_andn(a.val[1], b.val[1]), + v128_andn(a.val[0], b.val[0])); +} + +SIMD_INLINE v256 v256_add_8(v256 a, v256 b) { + return v256_from_v128(v128_add_8(a.val[1], b.val[1]), + v128_add_8(a.val[0], b.val[0])); +} + +SIMD_INLINE v256 v256_add_16(v256 a, v256 b) { + return v256_from_v128(v128_add_16(a.val[1], b.val[1]), + v128_add_16(a.val[0], b.val[0])); +} + +SIMD_INLINE v256 v256_sadd_s8(v256 a, v256 b) { + return v256_from_v128(v128_sadd_s8(a.val[1], b.val[1]), + v128_sadd_s8(a.val[0], b.val[0])); +} + +SIMD_INLINE v256 v256_sadd_u8(v256 a, v256 b) { + return v256_from_v128(v128_sadd_u8(a.val[1], b.val[1]), + v128_sadd_u8(a.val[0], b.val[0])); +} + +SIMD_INLINE v256 v256_sadd_s16(v256 a, v256 b) { + return v256_from_v128(v128_sadd_s16(a.val[1], b.val[1]), + v128_sadd_s16(a.val[0], b.val[0])); +} + +SIMD_INLINE v256 v256_add_32(v256 a, v256 b) { + return v256_from_v128(v128_add_32(a.val[1], b.val[1]), + v128_add_32(a.val[0], b.val[0])); +} + +SIMD_INLINE v256 v256_add_64(v256 a, v256 b) { + return v256_from_v128(v128_add_64(a.val[1], b.val[1]), + v128_add_64(a.val[0], b.val[0])); +} + +SIMD_INLINE v256 v256_padd_u8(v256 a) { + return v256_from_v128(v128_padd_u8(a.val[1]), v128_padd_u8(a.val[0])); +} + +SIMD_INLINE v256 v256_padd_s16(v256 a) { + return v256_from_v128(v128_padd_s16(a.val[1]), v128_padd_s16(a.val[0])); +} + +SIMD_INLINE v256 v256_sub_8(v256 a, v256 b) { + return v256_from_v128(v128_sub_8(a.val[1], b.val[1]), + v128_sub_8(a.val[0], b.val[0])); +} + +SIMD_INLINE v256 v256_ssub_u8(v256 a, v256 b) { + return v256_from_v128(v128_ssub_u8(a.val[1], b.val[1]), + v128_ssub_u8(a.val[0], b.val[0])); +} + +SIMD_INLINE v256 v256_ssub_s8(v256 a, v256 b) { + return v256_from_v128(v128_ssub_s8(a.val[1], b.val[1]), + v128_ssub_s8(a.val[0], b.val[0])); +} + +SIMD_INLINE v256 v256_sub_16(v256 a, v256 b) { + return v256_from_v128(v128_sub_16(a.val[1], b.val[1]), + v128_sub_16(a.val[0], b.val[0])); +} + +SIMD_INLINE v256 v256_ssub_s16(v256 a, v256 b) { + return v256_from_v128(v128_ssub_s16(a.val[1], b.val[1]), + v128_ssub_s16(a.val[0], b.val[0])); +} + +SIMD_INLINE v256 v256_ssub_u16(v256 a, v256 b) { + return v256_from_v128(v128_ssub_u16(a.val[1], b.val[1]), + v128_ssub_u16(a.val[0], b.val[0])); +} + +SIMD_INLINE v256 v256_sub_32(v256 a, v256 b) { + return v256_from_v128(v128_sub_32(a.val[1], b.val[1]), + v128_sub_32(a.val[0], b.val[0])); +} + +SIMD_INLINE v256 v256_sub_64(v256 a, v256 b) { + return v256_from_v128(v128_sub_64(a.val[1], b.val[1]), + v128_sub_64(a.val[0], b.val[0])); +} + +SIMD_INLINE v256 v256_abs_s16(v256 a) { + return v256_from_v128(v128_abs_s16(a.val[1]), v128_abs_s16(a.val[0])); +} + +SIMD_INLINE v256 v256_abs_s8(v256 a) { + return v256_from_v128(v128_abs_s8(a.val[1]), v128_abs_s8(a.val[0])); +} + +SIMD_INLINE v256 v256_mul_s16(v128 a, v128 b) { + v128 lo_bits = v128_mullo_s16(a, b); + v128 hi_bits = v128_mulhi_s16(a, b); + return v256_from_v128(v128_ziphi_16(hi_bits, lo_bits), + v128_ziplo_16(hi_bits, lo_bits)); +} + +SIMD_INLINE v256 v256_mullo_s16(v256 a, v256 b) { + return v256_from_v128(v128_mullo_s16(a.val[1], b.val[1]), + v128_mullo_s16(a.val[0], b.val[0])); +} + +SIMD_INLINE v256 v256_mulhi_s16(v256 a, v256 b) { + return v256_from_v128(v128_mulhi_s16(a.val[1], b.val[1]), + v128_mulhi_s16(a.val[0], b.val[0])); +} + +SIMD_INLINE v256 v256_mullo_s32(v256 a, v256 b) { + return v256_from_v128(v128_mullo_s32(a.val[1], b.val[1]), + v128_mullo_s32(a.val[0], b.val[0])); +} + +SIMD_INLINE v256 v256_madd_s16(v256 a, v256 b) { + return v256_from_v128(v128_madd_s16(a.val[1], b.val[1]), + v128_madd_s16(a.val[0], b.val[0])); +} + +SIMD_INLINE v256 v256_madd_us8(v256 a, v256 b) { + return v256_from_v128(v128_madd_us8(a.val[1], b.val[1]), + v128_madd_us8(a.val[0], b.val[0])); +} + +SIMD_INLINE v256 v256_avg_u8(v256 a, v256 b) { + return v256_from_v128(v128_avg_u8(a.val[1], b.val[1]), + v128_avg_u8(a.val[0], b.val[0])); +} + +SIMD_INLINE v256 v256_rdavg_u8(v256 a, v256 b) { + return v256_from_v128(v128_rdavg_u8(a.val[1], b.val[1]), + v128_rdavg_u8(a.val[0], b.val[0])); +} + +SIMD_INLINE v256 v256_rdavg_u16(v256 a, v256 b) { + return v256_from_v128(v128_rdavg_u16(a.val[1], b.val[1]), + v128_rdavg_u16(a.val[0], b.val[0])); +} + +SIMD_INLINE v256 v256_avg_u16(v256 a, v256 b) { + return v256_from_v128(v128_avg_u16(a.val[1], b.val[1]), + v128_avg_u16(a.val[0], b.val[0])); +} + +SIMD_INLINE v256 v256_min_u8(v256 a, v256 b) { + return v256_from_v128(v128_min_u8(a.val[1], b.val[1]), + v128_min_u8(a.val[0], b.val[0])); +} + +SIMD_INLINE v256 v256_max_u8(v256 a, v256 b) { + return v256_from_v128(v128_max_u8(a.val[1], b.val[1]), + v128_max_u8(a.val[0], b.val[0])); +} + +SIMD_INLINE v256 v256_min_s8(v256 a, v256 b) { + return v256_from_v128(v128_min_s8(a.val[1], b.val[1]), + v128_min_s8(a.val[0], b.val[0])); +} + +SIMD_INLINE uint32_t v256_movemask_8(v256 a) { + return (v128_movemask_8(v256_high_v128(a)) << 16) | + v128_movemask_8(v256_low_v128(a)); +} + +SIMD_INLINE v256 v256_blend_8(v256 a, v256 b, v256 c) { + return v256_from_v128(v128_blend_8(a.val[1], b.val[1], c.val[1]), + v128_blend_8(a.val[0], b.val[0], c.val[0])); +} + +SIMD_INLINE v256 v256_max_s8(v256 a, v256 b) { + return v256_from_v128(v128_max_s8(a.val[1], b.val[1]), + v128_max_s8(a.val[0], b.val[0])); +} + +SIMD_INLINE v256 v256_min_s16(v256 a, v256 b) { + return v256_from_v128(v128_min_s16(a.val[1], b.val[1]), + v128_min_s16(a.val[0], b.val[0])); +} + +SIMD_INLINE v256 v256_max_s16(v256 a, v256 b) { + return v256_from_v128(v128_max_s16(a.val[1], b.val[1]), + v128_max_s16(a.val[0], b.val[0])); +} + +SIMD_INLINE v256 v256_min_s32(v256 a, v256 b) { + return v256_from_v128(v128_min_s32(a.val[1], b.val[1]), + v128_min_s32(a.val[0], b.val[0])); +} + +SIMD_INLINE v256 v256_max_s32(v256 a, v256 b) { + return v256_from_v128(v128_max_s32(a.val[1], b.val[1]), + v128_max_s32(a.val[0], b.val[0])); +} + +SIMD_INLINE v256 v256_ziplo_8(v256 a, v256 b) { + return v256_from_v128(v128_ziphi_8(a.val[0], b.val[0]), + v128_ziplo_8(a.val[0], b.val[0])); +} + +SIMD_INLINE v256 v256_ziphi_8(v256 a, v256 b) { + return v256_from_v128(v128_ziphi_8(a.val[1], b.val[1]), + v128_ziplo_8(a.val[1], b.val[1])); +} + +SIMD_INLINE v256 v256_ziplo_16(v256 a, v256 b) { + return v256_from_v128(v128_ziphi_16(a.val[0], b.val[0]), + v128_ziplo_16(a.val[0], b.val[0])); +} + +SIMD_INLINE v256 v256_ziphi_16(v256 a, v256 b) { + return v256_from_v128(v128_ziphi_16(a.val[1], b.val[1]), + v128_ziplo_16(a.val[1], b.val[1])); +} + +SIMD_INLINE v256 v256_ziplo_32(v256 a, v256 b) { + return v256_from_v128(v128_ziphi_32(a.val[0], b.val[0]), + v128_ziplo_32(a.val[0], b.val[0])); +} + +SIMD_INLINE v256 v256_ziphi_32(v256 a, v256 b) { + return v256_from_v128(v128_ziphi_32(a.val[1], b.val[1]), + v128_ziplo_32(a.val[1], b.val[1])); +} + +SIMD_INLINE v256 v256_ziplo_64(v256 a, v256 b) { + return v256_from_v128(v128_ziphi_64(a.val[0], b.val[0]), + v128_ziplo_64(a.val[0], b.val[0])); +} + +SIMD_INLINE v256 v256_ziphi_64(v256 a, v256 b) { + return v256_from_v128(v128_ziphi_64(a.val[1], b.val[1]), + v128_ziplo_64(a.val[1], b.val[1])); +} + +SIMD_INLINE v256 v256_ziplo_128(v256 a, v256 b) { + return v256_from_v128(a.val[0], b.val[0]); +} + +SIMD_INLINE v256 v256_ziphi_128(v256 a, v256 b) { + return v256_from_v128(a.val[1], b.val[1]); +} + +SIMD_INLINE v256 v256_zip_8(v128 a, v128 b) { + return v256_from_v128(v128_ziphi_8(a, b), v128_ziplo_8(a, b)); +} + +SIMD_INLINE v256 v256_zip_16(v128 a, v128 b) { + return v256_from_v128(v128_ziphi_16(a, b), v128_ziplo_16(a, b)); +} + +SIMD_INLINE v256 v256_zip_32(v128 a, v128 b) { + return v256_from_v128(v128_ziphi_32(a, b), v128_ziplo_32(a, b)); +} + +SIMD_INLINE v256 v256_unziplo_8(v256 a, v256 b) { + return v256_from_v128(v128_unziplo_8(a.val[1], a.val[0]), + v128_unziplo_8(b.val[1], b.val[0])); +} + +SIMD_INLINE v256 v256_unziphi_8(v256 a, v256 b) { + return v256_from_v128(v128_unziphi_8(a.val[1], a.val[0]), + v128_unziphi_8(b.val[1], b.val[0])); +} + +SIMD_INLINE v256 v256_unziplo_16(v256 a, v256 b) { + return v256_from_v128(v128_unziplo_16(a.val[1], a.val[0]), + v128_unziplo_16(b.val[1], b.val[0])); +} + +SIMD_INLINE v256 v256_unziphi_16(v256 a, v256 b) { + return v256_from_v128(v128_unziphi_16(a.val[1], a.val[0]), + v128_unziphi_16(b.val[1], b.val[0])); +} + +SIMD_INLINE v256 v256_unziplo_32(v256 a, v256 b) { + return v256_from_v128(v128_unziplo_32(a.val[1], a.val[0]), + v128_unziplo_32(b.val[1], b.val[0])); +} + +SIMD_INLINE v256 v256_unziphi_32(v256 a, v256 b) { + return v256_from_v128(v128_unziphi_32(a.val[1], a.val[0]), + v128_unziphi_32(b.val[1], b.val[0])); +} + +SIMD_INLINE v256 v256_unziplo_64(v256 a, v256 b) { +#if HAVE_SSE2 + return v256_from_v128( + _mm_castpd_si128(_mm_shuffle_pd(_mm_castsi128_pd(a.val[0]), + _mm_castsi128_pd(a.val[1]), 0)), + _mm_castpd_si128(_mm_shuffle_pd(_mm_castsi128_pd(b.val[0]), + _mm_castsi128_pd(b.val[1]), 0))); +#else + return v256_from_v64(v128_low_v64(a.val[1]), v128_low_v64(a.val[0]), + v128_low_v64(b.val[1]), v128_low_v64(b.val[0])); +#endif +} + +SIMD_INLINE v256 v256_unziphi_64(v256 a, v256 b) { +#if HAVE_SSE2 + return v256_from_v128( + _mm_castpd_si128(_mm_shuffle_pd(_mm_castsi128_pd(a.val[0]), + _mm_castsi128_pd(a.val[1]), 3)), + _mm_castpd_si128(_mm_shuffle_pd(_mm_castsi128_pd(b.val[0]), + _mm_castsi128_pd(b.val[1]), 3))); +#else + return v256_from_v64(v128_high_v64(a.val[1]), v128_high_v64(a.val[0]), + v128_high_v64(b.val[1]), v128_high_v64(b.val[0])); +#endif +} + +SIMD_INLINE v256 v256_unpack_u8_s16(v128 a) { + return v256_from_v128(v128_unpackhi_u8_s16(a), v128_unpacklo_u8_s16(a)); +} + +SIMD_INLINE v256 v256_unpacklo_u8_s16(v256 a) { + return v256_from_v128(v128_unpackhi_u8_s16(a.val[0]), + v128_unpacklo_u8_s16(a.val[0])); +} + +SIMD_INLINE v256 v256_unpackhi_u8_s16(v256 a) { + return v256_from_v128(v128_unpackhi_u8_s16(a.val[1]), + v128_unpacklo_u8_s16(a.val[1])); +} + +SIMD_INLINE v256 v256_unpack_s8_s16(v128 a) { + return v256_from_v128(v128_unpackhi_s8_s16(a), v128_unpacklo_s8_s16(a)); +} + +SIMD_INLINE v256 v256_unpacklo_s8_s16(v256 a) { + return v256_from_v128(v128_unpackhi_s8_s16(a.val[0]), + v128_unpacklo_s8_s16(a.val[0])); +} + +SIMD_INLINE v256 v256_unpackhi_s8_s16(v256 a) { + return v256_from_v128(v128_unpackhi_s8_s16(a.val[1]), + v128_unpacklo_s8_s16(a.val[1])); +} + +SIMD_INLINE v256 v256_pack_s32_s16(v256 a, v256 b) { + return v256_from_v128(v128_pack_s32_s16(a.val[1], a.val[0]), + v128_pack_s32_s16(b.val[1], b.val[0])); +} + +SIMD_INLINE v256 v256_pack_s32_u16(v256 a, v256 b) { + return v256_from_v128(v128_pack_s32_u16(a.val[1], a.val[0]), + v128_pack_s32_u16(b.val[1], b.val[0])); +} + +SIMD_INLINE v256 v256_pack_s16_u8(v256 a, v256 b) { + return v256_from_v128(v128_pack_s16_u8(a.val[1], a.val[0]), + v128_pack_s16_u8(b.val[1], b.val[0])); +} + +SIMD_INLINE v256 v256_pack_s16_s8(v256 a, v256 b) { + return v256_from_v128(v128_pack_s16_s8(a.val[1], a.val[0]), + v128_pack_s16_s8(b.val[1], b.val[0])); +} + +SIMD_INLINE v256 v256_unpack_u16_s32(v128 a) { + return v256_from_v128(v128_unpackhi_u16_s32(a), v128_unpacklo_u16_s32(a)); +} + +SIMD_INLINE v256 v256_unpack_s16_s32(v128 a) { + return v256_from_v128(v128_unpackhi_s16_s32(a), v128_unpacklo_s16_s32(a)); +} + +SIMD_INLINE v256 v256_unpacklo_u16_s32(v256 a) { + return v256_from_v128(v128_unpackhi_u16_s32(a.val[0]), + v128_unpacklo_u16_s32(a.val[0])); +} + +SIMD_INLINE v256 v256_unpacklo_s16_s32(v256 a) { + return v256_from_v128(v128_unpackhi_s16_s32(a.val[0]), + v128_unpacklo_s16_s32(a.val[0])); +} + +SIMD_INLINE v256 v256_unpackhi_u16_s32(v256 a) { + return v256_from_v128(v128_unpackhi_u16_s32(a.val[1]), + v128_unpacklo_u16_s32(a.val[1])); +} + +SIMD_INLINE v256 v256_unpackhi_s16_s32(v256 a) { + return v256_from_v128(v128_unpackhi_s16_s32(a.val[1]), + v128_unpacklo_s16_s32(a.val[1])); +} + +SIMD_INLINE v256 v256_cmpgt_s8(v256 a, v256 b) { + return v256_from_v128(v128_cmpgt_s8(a.val[1], b.val[1]), + v128_cmpgt_s8(a.val[0], b.val[0])); +} + +SIMD_INLINE v256 v256_cmplt_s8(v256 a, v256 b) { + return v256_from_v128(v128_cmplt_s8(a.val[1], b.val[1]), + v128_cmplt_s8(a.val[0], b.val[0])); +} + +SIMD_INLINE v256 v256_cmpeq_8(v256 a, v256 b) { + return v256_from_v128(v128_cmpeq_8(a.val[1], b.val[1]), + v128_cmpeq_8(a.val[0], b.val[0])); +} + +SIMD_INLINE v256 v256_cmpgt_s16(v256 a, v256 b) { + return v256_from_v128(v128_cmpgt_s16(a.val[1], b.val[1]), + v128_cmpgt_s16(a.val[0], b.val[0])); +} + +SIMD_INLINE v256 v256_cmplt_s16(v256 a, v256 b) { + return v256_from_v128(v128_cmplt_s16(a.val[1], b.val[1]), + v128_cmplt_s16(a.val[0], b.val[0])); +} + +SIMD_INLINE v256 v256_cmpeq_16(v256 a, v256 b) { + return v256_from_v128(v128_cmpeq_16(a.val[1], b.val[1]), + v128_cmpeq_16(a.val[0], b.val[0])); +} + +SIMD_INLINE v256 v256_cmpgt_s32(v256 a, v256 b) { + return v256_from_v128(v128_cmpgt_s32(a.val[1], b.val[1]), + v128_cmpgt_s32(a.val[0], b.val[0])); +} + +SIMD_INLINE v256 v256_cmplt_s32(v256 a, v256 b) { + return v256_from_v128(v128_cmplt_s32(a.val[1], b.val[1]), + v128_cmplt_s32(a.val[0], b.val[0])); +} + +SIMD_INLINE v256 v256_cmpeq_32(v256 a, v256 b) { + return v256_from_v128(v128_cmpeq_32(a.val[1], b.val[1]), + v128_cmpeq_32(a.val[0], b.val[0])); +} + +SIMD_INLINE v256 v256_shuffle_8(v256 x, v256 pattern) { +#if HAVE_NEON +#if defined(__aarch64__) + uint8x16x2_t p = { { vreinterpretq_u8_s64(x.val[0]), + vreinterpretq_u8_s64(x.val[1]) } }; + return v256_from_v128( + vreinterpretq_s64_u8(vqtbl2q_u8(p, vreinterpretq_u8_s64(pattern.val[1]))), + vreinterpretq_s64_u8( + vqtbl2q_u8(p, vreinterpretq_u8_s64(pattern.val[0])))); +#else + uint8x8x4_t p = { { vget_low_u8(vreinterpretq_u8_s64(x.val[0])), + vget_high_u8(vreinterpretq_u8_s64(x.val[0])), + vget_low_u8(vreinterpretq_u8_s64(x.val[1])), + vget_high_u8(vreinterpretq_u8_s64(x.val[1])) } }; + return v256_from_64( + (uint64_t)vreinterpret_s64_u8( + vtbl4_u8(p, vreinterpret_u8_s64(vget_high_s64(pattern.val[1])))), + (uint64_t)vreinterpret_s64_u8( + vtbl4_u8(p, vreinterpret_u8_s64(vget_low_s64(pattern.val[1])))), + (uint64_t)vreinterpret_s64_u8( + vtbl4_u8(p, vreinterpret_u8_s64(vget_high_s64(pattern.val[0])))), + (uint64_t)vreinterpret_s64_u8( + vtbl4_u8(p, vreinterpret_u8_s64(vget_low_s64(pattern.val[0]))))); +#endif +#else + v128 c16 = v128_dup_8(16); + v128 maskhi = v128_cmplt_s8(pattern.val[1], c16); + v128 masklo = v128_cmplt_s8(pattern.val[0], c16); + return v256_from_v128( + v128_blend_8(v128_shuffle_8(x.val[1], v128_sub_8(pattern.val[1], c16)), + v128_shuffle_8(x.val[0], pattern.val[1]), maskhi), + v128_blend_8(v128_shuffle_8(x.val[1], v128_sub_8(pattern.val[0], c16)), + v128_shuffle_8(x.val[0], pattern.val[0]), masklo)); +#endif +} + +SIMD_INLINE v256 v256_wideshuffle_8(v256 x, v256 y, v256 pattern) { +#if HAVE_NEON +#if defined(__aarch64__) + uint8x16x4_t p = { { + vreinterpretq_u8_s64(y.val[0]), + vreinterpretq_u8_s64(y.val[1]), + vreinterpretq_u8_s64(x.val[0]), + vreinterpretq_u8_s64(x.val[1]), + } }; + return v256_from_v128( + vreinterpretq_s64_u8(vqtbl4q_u8(p, vreinterpretq_u8_s64(pattern.val[1]))), + vreinterpretq_s64_u8( + vqtbl4q_u8(p, vreinterpretq_u8_s64(pattern.val[0])))); +#else + v256 c32 = v256_dup_8(32); + v256 p32 = v256_sub_8(pattern, c32); + uint8x8x4_t p = { { vget_low_u8(vreinterpretq_u8_s64(x.val[0])), + vget_high_u8(vreinterpretq_u8_s64(x.val[0])), + vget_low_u8(vreinterpretq_u8_s64(x.val[1])), + vget_high_u8(vreinterpretq_u8_s64(x.val[1])) } }; + uint8x8x4_t q = { { vget_low_u8(vreinterpretq_u8_s64(y.val[0])), + vget_high_u8(vreinterpretq_u8_s64(y.val[0])), + vget_low_u8(vreinterpretq_u8_s64(y.val[1])), + vget_high_u8(vreinterpretq_u8_s64(y.val[1])) } }; + v256 r1 = + v256_from_64((uint64_t)vreinterpret_s64_u8(vtbl4_u8( + p, vreinterpret_u8_s64(vget_high_s64(p32.val[1])))), + (uint64_t)vreinterpret_s64_u8(vtbl4_u8( + p, vreinterpret_u8_s64(vget_low_s64(p32.val[1])))), + (uint64_t)vreinterpret_s64_u8(vtbl4_u8( + p, vreinterpret_u8_s64(vget_high_s64(p32.val[0])))), + (uint64_t)vreinterpret_s64_u8(vtbl4_u8( + p, vreinterpret_u8_s64(vget_low_s64(p32.val[0]))))); + v256 r2 = + v256_from_64((uint64_t)vreinterpret_s64_u8(vtbl4_u8( + q, vreinterpret_u8_s64(vget_high_s64(pattern.val[1])))), + (uint64_t)vreinterpret_s64_u8(vtbl4_u8( + q, vreinterpret_u8_s64(vget_low_s64(pattern.val[1])))), + (uint64_t)vreinterpret_s64_u8(vtbl4_u8( + q, vreinterpret_u8_s64(vget_high_s64(pattern.val[0])))), + (uint64_t)vreinterpret_s64_u8(vtbl4_u8( + q, vreinterpret_u8_s64(vget_low_s64(pattern.val[0]))))); + return v256_blend_8(r1, r2, v256_cmplt_s8(pattern, c32)); +#endif +#else + v128 c16 = v128_dup_8(16); + v128 c32 = v128_dup_8(32); + v128 c48 = v128_dup_8(48); + v128 maskhi16 = v128_cmpgt_s8(c16, pattern.val[1]); + v128 masklo16 = v128_cmpgt_s8(c16, pattern.val[0]); + v128 maskhi48 = v128_cmpgt_s8(c48, pattern.val[1]); + v128 masklo48 = v128_cmpgt_s8(c48, pattern.val[0]); + v256 r1 = v256_from_v128( + v128_blend_8(v128_shuffle_8(x.val[1], v128_sub_8(pattern.val[1], c48)), + v128_shuffle_8(x.val[0], v128_sub_8(pattern.val[1], c32)), + maskhi48), + v128_blend_8(v128_shuffle_8(x.val[1], v128_sub_8(pattern.val[0], c48)), + v128_shuffle_8(x.val[0], v128_sub_8(pattern.val[0], c32)), + masklo48)); + v256 r2 = v256_from_v128( + v128_blend_8(v128_shuffle_8(y.val[1], v128_sub_8(pattern.val[1], c16)), + v128_shuffle_8(y.val[0], pattern.val[1]), maskhi16), + v128_blend_8(v128_shuffle_8(y.val[1], v128_sub_8(pattern.val[0], c16)), + v128_shuffle_8(y.val[0], pattern.val[0]), masklo16)); + return v256_blend_8(r1, r2, v256_cmpgt_s8(v256_from_v128(c32, c32), pattern)); +#endif +} + +SIMD_INLINE v256 v256_pshuffle_8(v256 a, v256 pattern) { + return v256_from_v128( + v128_shuffle_8(v256_high_v128(a), v256_high_v128(pattern)), + v128_shuffle_8(v256_low_v128(a), v256_low_v128(pattern))); +} + +SIMD_INLINE v256 v256_shl_8(v256 a, const unsigned int c) { + return v256_from_v128(v128_shl_8(a.val[1], c), v128_shl_8(a.val[0], c)); +} + +SIMD_INLINE v256 v256_shr_u8(v256 a, const unsigned int c) { + return v256_from_v128(v128_shr_u8(a.val[1], c), v128_shr_u8(a.val[0], c)); +} + +SIMD_INLINE v256 v256_shr_s8(v256 a, const unsigned int c) { + return v256_from_v128(v128_shr_s8(a.val[1], c), v128_shr_s8(a.val[0], c)); +} + +SIMD_INLINE v256 v256_shl_16(v256 a, const unsigned int c) { + return v256_from_v128(v128_shl_16(a.val[1], c), v128_shl_16(a.val[0], c)); +} + +SIMD_INLINE v256 v256_shr_u16(v256 a, const unsigned int c) { + return v256_from_v128(v128_shr_u16(a.val[1], c), v128_shr_u16(a.val[0], c)); +} + +SIMD_INLINE v256 v256_shr_s16(v256 a, const unsigned int c) { + return v256_from_v128(v128_shr_s16(a.val[1], c), v128_shr_s16(a.val[0], c)); +} + +SIMD_INLINE v256 v256_shl_32(v256 a, const unsigned int c) { + return v256_from_v128(v128_shl_32(a.val[1], c), v128_shl_32(a.val[0], c)); +} + +SIMD_INLINE v256 v256_shr_u32(v256 a, const unsigned int c) { + return v256_from_v128(v128_shr_u32(a.val[1], c), v128_shr_u32(a.val[0], c)); +} + +SIMD_INLINE v256 v256_shr_s32(v256 a, const unsigned int c) { + return v256_from_v128(v128_shr_s32(a.val[1], c), v128_shr_s32(a.val[0], c)); +} + +SIMD_INLINE v256 v256_shl_64(v256 a, const unsigned int c) { + return v256_from_v128(v128_shl_64(a.val[1], c), v128_shl_64(a.val[0], c)); +} + +SIMD_INLINE v256 v256_shr_u64(v256 a, const unsigned int c) { + return v256_from_v128(v128_shr_u64(a.val[1], c), v128_shr_u64(a.val[0], c)); +} + +SIMD_INLINE v256 v256_shr_s64(v256 a, const unsigned int c) { + return v256_from_v128(v128_shr_s64(a.val[1], c), v128_shr_s64(a.val[0], c)); +} + +/* These intrinsics require immediate values, so we must use #defines + to enforce that. */ +#define v256_shl_n_byte(a, n) \ + ((n) < 16 ? v256_from_v128(v128_or(v128_shl_n_byte(a.val[1], n), \ + v128_shr_n_byte(a.val[0], 16 - (n))), \ + v128_shl_n_byte(a.val[0], (n))) \ + : v256_from_v128( \ + (n) > 16 ? v128_shl_n_byte(a.val[0], (n)-16) : a.val[0], \ + v128_zero())) + +#define v256_shr_n_byte(a, n) \ + ((n) < 16 ? v256_from_v128(v128_shr_n_byte(a.val[1], n), \ + v128_or(v128_shr_n_byte(a.val[0], n), \ + v128_shl_n_byte(a.val[1], 16 - (n)))) \ + : v256_from_v128( \ + v128_zero(), \ + (n) > 16 ? v128_shr_n_byte(a.val[1], (n)-16) : a.val[1])) + +#define v256_align(a, b, c) \ + ((c) ? v256_or(v256_shr_n_byte(b, c), v256_shl_n_byte(a, 32 - (c))) : b) + +#define v256_shl_n_8(a, n) \ + v256_from_v128(v128_shl_n_8(a.val[1], n), v128_shl_n_8(a.val[0], n)) +#define v256_shl_n_16(a, n) \ + v256_from_v128(v128_shl_n_16(a.val[1], n), v128_shl_n_16(a.val[0], n)) +#define v256_shl_n_32(a, n) \ + v256_from_v128(v128_shl_n_32(a.val[1], n), v128_shl_n_32(a.val[0], n)) +#define v256_shl_n_64(a, n) \ + v256_from_v128(v128_shl_n_64(a.val[1], n), v128_shl_n_64(a.val[0], n)) +#define v256_shr_n_u8(a, n) \ + v256_from_v128(v128_shr_n_u8(a.val[1], n), v128_shr_n_u8(a.val[0], n)) +#define v256_shr_n_u16(a, n) \ + v256_from_v128(v128_shr_n_u16(a.val[1], n), v128_shr_n_u16(a.val[0], n)) +#define v256_shr_n_u32(a, n) \ + v256_from_v128(v128_shr_n_u32(a.val[1], n), v128_shr_n_u32(a.val[0], n)) +#define v256_shr_n_u64(a, n) \ + v256_from_v128(v128_shr_n_u64(a.val[1], n), v128_shr_n_u64(a.val[0], n)) +#define v256_shr_n_s8(a, n) \ + v256_from_v128(v128_shr_n_s8(a.val[1], n), v128_shr_n_s8(a.val[0], n)) +#define v256_shr_n_s16(a, n) \ + v256_from_v128(v128_shr_n_s16(a.val[1], n), v128_shr_n_s16(a.val[0], n)) +#define v256_shr_n_s32(a, n) \ + v256_from_v128(v128_shr_n_s32(a.val[1], n), v128_shr_n_s32(a.val[0], n)) +#define v256_shr_n_s64(a, n) \ + v256_from_v128(v128_shr_n_s64(a.val[1], n), v128_shr_n_s64(a.val[0], n)) + +#define v256_shr_n_word(a, n) v256_shr_n_byte(a, 2 * (n)) +#define v256_shl_n_word(a, n) v256_shl_n_byte(a, 2 * (n)) + +typedef struct { + sad128_internal_u16 val[2]; +} sad256_internal_u16; + +SIMD_INLINE sad256_internal_u16 v256_sad_u16_init() { + sad256_internal_u16 t; + t.val[1] = v128_sad_u16_init(); + t.val[0] = v128_sad_u16_init(); + return t; +} + +/* Implementation dependent return value. Result must be finalised with + v256_sad_u16_sum(). + The result for more than 16 v256_sad_u16() calls is undefined. */ +SIMD_INLINE sad256_internal_u16 v256_sad_u16(sad256_internal_u16 s, v256 a, + v256 b) { + sad256_internal_u16 t; + t.val[1] = v128_sad_u16(s.val[1], a.val[1], b.val[1]); + t.val[0] = v128_sad_u16(s.val[0], a.val[0], b.val[0]); + return t; +} + +SIMD_INLINE uint32_t v256_sad_u16_sum(sad256_internal_u16 s) { + return v128_sad_u16_sum(s.val[1]) + v128_sad_u16_sum(s.val[0]); +} + +typedef struct { + ssd128_internal_s16 val[2]; +} ssd256_internal_s16; + +SIMD_INLINE ssd256_internal_s16 v256_ssd_s16_init() { + ssd256_internal_s16 t; + t.val[1] = v128_ssd_s16_init(); + t.val[0] = v128_ssd_s16_init(); + return t; +} + +/* Implementation dependent return value. Result must be finalised with + * v256_ssd_s16_sum(). */ +SIMD_INLINE ssd256_internal_s16 v256_ssd_s16(ssd256_internal_s16 s, v256 a, + v256 b) { + ssd256_internal_s16 t; + t.val[1] = v128_ssd_s16(s.val[1], a.val[1], b.val[1]); + t.val[0] = v128_ssd_s16(s.val[0], a.val[0], b.val[0]); + return t; +} + +SIMD_INLINE uint64_t v256_ssd_s16_sum(ssd256_internal_s16 s) { + return v128_ssd_s16_sum(s.val[1]) + v128_ssd_s16_sum(s.val[0]); +} + +#endif // AOM_AOM_DSP_SIMD_V256_INTRINSICS_V128_H_ diff --git a/third_party/aom/aom_dsp/simd/v256_intrinsics_x86.h b/third_party/aom/aom_dsp/simd/v256_intrinsics_x86.h new file mode 100644 index 0000000000..44594bc41c --- /dev/null +++ b/third_party/aom/aom_dsp/simd/v256_intrinsics_x86.h @@ -0,0 +1,750 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_SIMD_V256_INTRINSICS_X86_H_ +#define AOM_AOM_DSP_SIMD_V256_INTRINSICS_X86_H_ + +#if !defined(__AVX2__) + +#include "aom_dsp/simd/v256_intrinsics_v128.h" + +#else + +// The _m256i type seems to cause problems for g++'s mangling prior to +// version 5, but adding -fabi-version=0 fixes this. +#if !defined(__clang__) && defined(__GNUC__) && __GNUC__ < 5 && \ + defined(__AVX2__) && defined(__cplusplus) +#pragma GCC optimize "-fabi-version=0" +#endif + +#include + +#include "aom_dsp/simd/v128_intrinsics_x86.h" + +typedef __m256i v256; + +SIMD_INLINE uint32_t v256_low_u32(v256 a) { + return (uint32_t)_mm_cvtsi128_si32(_mm256_extracti128_si256(a, 0)); +} + +SIMD_INLINE v64 v256_low_v64(v256 a) { + return _mm_unpacklo_epi64(_mm256_extracti128_si256(a, 0), v64_zero()); +} + +SIMD_INLINE uint64_t v256_low_u64(v256 a) { return v64_u64(v256_low_v64(a)); } + +SIMD_INLINE v128 v256_low_v128(v256 a) { return _mm256_castsi256_si128(a); } + +SIMD_INLINE v128 v256_high_v128(v256 a) { + return _mm256_extracti128_si256(a, 1); +} + +SIMD_INLINE v256 v256_from_v128(v128 a, v128 b) { + // gcc seems to be missing _mm256_set_m128i() + return _mm256_inserti128_si256(_mm256_castsi128_si256(b), a, 1); +} + +SIMD_INLINE v256 v256_from_v64(v64 a, v64 b, v64 c, v64 d) { + return v256_from_v128(v128_from_v64(a, b), v128_from_v64(c, d)); +} + +SIMD_INLINE v256 v256_from_64(uint64_t a, uint64_t b, uint64_t c, uint64_t d) { + return v256_from_v128(v128_from_64(a, b), v128_from_64(c, d)); +} + +SIMD_INLINE v256 v256_load_aligned(const void *p) { + return _mm256_load_si256((const __m256i *)p); +} + +SIMD_INLINE v256 v256_load_unaligned(const void *p) { + return _mm256_loadu_si256((const __m256i *)p); +} + +SIMD_INLINE void v256_store_aligned(void *p, v256 a) { + _mm256_store_si256((__m256i *)p, a); +} + +SIMD_INLINE void v256_store_unaligned(void *p, v256 a) { + _mm256_storeu_si256((__m256i *)p, a); +} + +SIMD_INLINE v256 v256_zero() { return _mm256_setzero_si256(); } + +SIMD_INLINE v256 v256_dup_8(uint8_t x) { return _mm256_set1_epi8(x); } + +SIMD_INLINE v256 v256_dup_16(uint16_t x) { return _mm256_set1_epi16(x); } + +SIMD_INLINE v256 v256_dup_32(uint32_t x) { return _mm256_set1_epi32(x); } + +SIMD_INLINE v256 v256_dup_64(uint64_t x) { return _mm256_set1_epi64x(x); } + +SIMD_INLINE v256 v256_add_8(v256 a, v256 b) { return _mm256_add_epi8(a, b); } + +SIMD_INLINE v256 v256_add_16(v256 a, v256 b) { return _mm256_add_epi16(a, b); } + +SIMD_INLINE v256 v256_sadd_u8(v256 a, v256 b) { return _mm256_adds_epu8(a, b); } + +SIMD_INLINE v256 v256_sadd_s8(v256 a, v256 b) { return _mm256_adds_epi8(a, b); } + +SIMD_INLINE v256 v256_sadd_s16(v256 a, v256 b) { + return _mm256_adds_epi16(a, b); +} + +SIMD_INLINE v256 v256_add_32(v256 a, v256 b) { return _mm256_add_epi32(a, b); } + +SIMD_INLINE v256 v256_add_64(v256 a, v256 b) { return _mm256_add_epi64(a, b); } + +SIMD_INLINE v256 v256_padd_u8(v256 a) { + return _mm256_maddubs_epi16(a, _mm256_set1_epi8(1)); +} + +SIMD_INLINE v256 v256_padd_s16(v256 a) { + return _mm256_madd_epi16(a, _mm256_set1_epi16(1)); +} + +SIMD_INLINE v256 v256_sub_8(v256 a, v256 b) { return _mm256_sub_epi8(a, b); } + +SIMD_INLINE v256 v256_ssub_u8(v256 a, v256 b) { return _mm256_subs_epu8(a, b); } + +SIMD_INLINE v256 v256_ssub_s8(v256 a, v256 b) { return _mm256_subs_epi8(a, b); } + +SIMD_INLINE v256 v256_sub_16(v256 a, v256 b) { return _mm256_sub_epi16(a, b); } + +SIMD_INLINE v256 v256_ssub_s16(v256 a, v256 b) { + return _mm256_subs_epi16(a, b); +} + +SIMD_INLINE v256 v256_ssub_u16(v256 a, v256 b) { + return _mm256_subs_epu16(a, b); +} + +SIMD_INLINE v256 v256_sub_32(v256 a, v256 b) { return _mm256_sub_epi32(a, b); } + +SIMD_INLINE v256 v256_sub_64(v256 a, v256 b) { return _mm256_sub_epi64(a, b); } + +SIMD_INLINE v256 v256_abs_s16(v256 a) { return _mm256_abs_epi16(a); } + +SIMD_INLINE v256 v256_abs_s8(v256 a) { return _mm256_abs_epi8(a); } + +// AVX doesn't have the direct intrinsics to zip/unzip 8, 16, 32 bit +// lanes of lower or upper halves of a 256bit vector because the +// unpack/pack intrinsics operate on the 256 bit input vector as 2 +// independent 128 bit vectors. +SIMD_INLINE v256 v256_ziplo_8(v256 a, v256 b) { + return _mm256_unpacklo_epi8( + _mm256_permute4x64_epi64(b, _MM_SHUFFLE(3, 1, 2, 0)), + _mm256_permute4x64_epi64(a, _MM_SHUFFLE(3, 1, 2, 0))); +} + +SIMD_INLINE v256 v256_ziphi_8(v256 a, v256 b) { + return _mm256_unpackhi_epi8( + _mm256_permute4x64_epi64(b, _MM_SHUFFLE(3, 1, 2, 0)), + _mm256_permute4x64_epi64(a, _MM_SHUFFLE(3, 1, 2, 0))); +} + +SIMD_INLINE v256 v256_ziplo_16(v256 a, v256 b) { + return _mm256_unpacklo_epi16( + _mm256_permute4x64_epi64(b, _MM_SHUFFLE(3, 1, 2, 0)), + _mm256_permute4x64_epi64(a, _MM_SHUFFLE(3, 1, 2, 0))); +} + +SIMD_INLINE v256 v256_ziphi_16(v256 a, v256 b) { + return _mm256_unpackhi_epi16( + _mm256_permute4x64_epi64(b, _MM_SHUFFLE(3, 1, 2, 0)), + _mm256_permute4x64_epi64(a, _MM_SHUFFLE(3, 1, 2, 0))); +} + +SIMD_INLINE v256 v256_ziplo_32(v256 a, v256 b) { + return _mm256_unpacklo_epi32( + _mm256_permute4x64_epi64(b, _MM_SHUFFLE(3, 1, 2, 0)), + _mm256_permute4x64_epi64(a, _MM_SHUFFLE(3, 1, 2, 0))); +} + +SIMD_INLINE v256 v256_ziphi_32(v256 a, v256 b) { + return _mm256_unpackhi_epi32( + _mm256_permute4x64_epi64(b, _MM_SHUFFLE(3, 1, 2, 0)), + _mm256_permute4x64_epi64(a, _MM_SHUFFLE(3, 1, 2, 0))); +} + +SIMD_INLINE v256 v256_ziplo_64(v256 a, v256 b) { + return _mm256_unpacklo_epi64( + _mm256_permute4x64_epi64(b, _MM_SHUFFLE(3, 1, 2, 0)), + _mm256_permute4x64_epi64(a, _MM_SHUFFLE(3, 1, 2, 0))); +} + +SIMD_INLINE v256 v256_ziphi_64(v256 a, v256 b) { + return _mm256_unpackhi_epi64( + _mm256_permute4x64_epi64(b, _MM_SHUFFLE(3, 1, 2, 0)), + _mm256_permute4x64_epi64(a, _MM_SHUFFLE(3, 1, 2, 0))); +} + +SIMD_INLINE v256 v256_ziplo_128(v256 a, v256 b) { + return v256_from_v128(v256_low_v128(a), v256_low_v128(b)); +} + +SIMD_INLINE v256 v256_ziphi_128(v256 a, v256 b) { + return v256_from_v128(v256_high_v128(a), v256_high_v128(b)); +} + +SIMD_INLINE v256 v256_zip_8(v128 a, v128 b) { + return v256_from_v128(v128_ziphi_8(a, b), v128_ziplo_8(a, b)); +} + +SIMD_INLINE v256 v256_zip_16(v128 a, v128 b) { + return v256_from_v128(v128_ziphi_16(a, b), v128_ziplo_16(a, b)); +} + +SIMD_INLINE v256 v256_zip_32(v128 a, v128 b) { + return v256_from_v128(v128_ziphi_32(a, b), v128_ziplo_32(a, b)); +} + +SIMD_INLINE v256 v256_unziphi_8(v256 a, v256 b) { + return _mm256_permute4x64_epi64( + _mm256_packs_epi16(_mm256_srai_epi16(b, 8), _mm256_srai_epi16(a, 8)), + _MM_SHUFFLE(3, 1, 2, 0)); +} + +SIMD_INLINE v256 v256_unziplo_8(v256 a, v256 b) { + return v256_unziphi_8(_mm256_slli_si256(a, 1), _mm256_slli_si256(b, 1)); +} + +SIMD_INLINE v256 v256_unziphi_16(v256 a, v256 b) { + return _mm256_permute4x64_epi64( + _mm256_packs_epi32(_mm256_srai_epi32(b, 16), _mm256_srai_epi32(a, 16)), + _MM_SHUFFLE(3, 1, 2, 0)); +} + +SIMD_INLINE v256 v256_unziplo_16(v256 a, v256 b) { + return v256_unziphi_16(_mm256_slli_si256(a, 2), _mm256_slli_si256(b, 2)); +} + +SIMD_INLINE v256 v256_unziphi_32(v256 a, v256 b) { + return _mm256_permute4x64_epi64( + _mm256_castps_si256(_mm256_shuffle_ps(_mm256_castsi256_ps(b), + _mm256_castsi256_ps(a), + _MM_SHUFFLE(3, 1, 3, 1))), + _MM_SHUFFLE(3, 1, 2, 0)); +} + +SIMD_INLINE v256 v256_unziplo_32(v256 a, v256 b) { + return _mm256_permute4x64_epi64( + _mm256_castps_si256(_mm256_shuffle_ps(_mm256_castsi256_ps(b), + _mm256_castsi256_ps(a), + _MM_SHUFFLE(2, 0, 2, 0))), + _MM_SHUFFLE(3, 1, 2, 0)); +} + +SIMD_INLINE v256 v256_unziphi_64(v256 a, v256 b) { + return _mm256_permute4x64_epi64( + _mm256_castpd_si256(_mm256_shuffle_pd(_mm256_castsi256_pd(b), + _mm256_castsi256_pd(a), 15)), + _MM_SHUFFLE(3, 1, 2, 0)); +} + +SIMD_INLINE v256 v256_unziplo_64(v256 a, v256 b) { + return _mm256_permute4x64_epi64( + _mm256_castpd_si256( + _mm256_shuffle_pd(_mm256_castsi256_pd(b), _mm256_castsi256_pd(a), 0)), + _MM_SHUFFLE(3, 1, 2, 0)); +} + +SIMD_INLINE v256 v256_unpack_u8_s16(v128 a) { + return v256_from_v128(v128_unpackhi_u8_s16(a), v128_unpacklo_u8_s16(a)); +} + +SIMD_INLINE v256 v256_unpacklo_u8_s16(v256 a) { + return _mm256_unpacklo_epi8( + _mm256_permute4x64_epi64(a, _MM_SHUFFLE(3, 1, 2, 0)), + _mm256_setzero_si256()); +} + +SIMD_INLINE v256 v256_unpackhi_u8_s16(v256 a) { + return _mm256_unpackhi_epi8( + _mm256_permute4x64_epi64(a, _MM_SHUFFLE(3, 1, 2, 0)), + _mm256_setzero_si256()); +} + +SIMD_INLINE v256 v256_unpack_s8_s16(v128 a) { + return v256_from_v128(v128_unpackhi_s8_s16(a), v128_unpacklo_s8_s16(a)); +} + +SIMD_INLINE v256 v256_unpacklo_s8_s16(v256 a) { + return _mm256_srai_epi16( + _mm256_unpacklo_epi8( + a, _mm256_permute4x64_epi64(a, _MM_SHUFFLE(3, 1, 2, 0))), + 8); +} + +SIMD_INLINE v256 v256_unpackhi_s8_s16(v256 a) { + return _mm256_srai_epi16( + _mm256_unpackhi_epi8( + a, _mm256_permute4x64_epi64(a, _MM_SHUFFLE(3, 1, 2, 0))), + 8); +} + +SIMD_INLINE v256 v256_pack_s32_s16(v256 a, v256 b) { + return _mm256_permute4x64_epi64(_mm256_packs_epi32(b, a), + _MM_SHUFFLE(3, 1, 2, 0)); +} + +SIMD_INLINE v256 v256_pack_s32_u16(v256 a, v256 b) { + return _mm256_permute4x64_epi64(_mm256_packus_epi32(b, a), + _MM_SHUFFLE(3, 1, 2, 0)); +} + +SIMD_INLINE v256 v256_pack_s16_u8(v256 a, v256 b) { + return _mm256_permute4x64_epi64(_mm256_packus_epi16(b, a), + _MM_SHUFFLE(3, 1, 2, 0)); +} + +SIMD_INLINE v256 v256_pack_s16_s8(v256 a, v256 b) { + return _mm256_permute4x64_epi64(_mm256_packs_epi16(b, a), + _MM_SHUFFLE(3, 1, 2, 0)); +} + +SIMD_INLINE v256 v256_unpack_u16_s32(v128 a) { + return v256_from_v128(v128_unpackhi_u16_s32(a), v128_unpacklo_u16_s32(a)); +} + +SIMD_INLINE v256 v256_unpack_s16_s32(v128 a) { + return v256_from_v128(v128_unpackhi_s16_s32(a), v128_unpacklo_s16_s32(a)); +} + +SIMD_INLINE v256 v256_unpacklo_u16_s32(v256 a) { + return _mm256_unpacklo_epi16( + _mm256_permute4x64_epi64(a, _MM_SHUFFLE(3, 1, 2, 0)), + _mm256_setzero_si256()); +} + +SIMD_INLINE v256 v256_unpacklo_s16_s32(v256 a) { + return _mm256_srai_epi32( + _mm256_unpacklo_epi16( + a, _mm256_permute4x64_epi64(a, _MM_SHUFFLE(3, 1, 2, 0))), + 16); +} + +SIMD_INLINE v256 v256_unpackhi_u16_s32(v256 a) { + return _mm256_unpackhi_epi16( + _mm256_permute4x64_epi64(a, _MM_SHUFFLE(3, 1, 2, 0)), + _mm256_setzero_si256()); +} + +SIMD_INLINE v256 v256_unpackhi_s16_s32(v256 a) { + return _mm256_srai_epi32( + _mm256_unpackhi_epi16( + a, _mm256_permute4x64_epi64(a, _MM_SHUFFLE(3, 1, 2, 0))), + 16); +} + +SIMD_INLINE v256 v256_shuffle_8(v256 a, v256 pattern) { + return _mm256_blendv_epi8( + _mm256_shuffle_epi8( + _mm256_permute2x128_si256(a, a, _MM_SHUFFLE(0, 1, 0, 1)), pattern), + _mm256_shuffle_epi8( + _mm256_permute2x128_si256(a, a, _MM_SHUFFLE(0, 0, 0, 0)), pattern), + _mm256_cmpgt_epi8(v256_dup_8(16), pattern)); +} + +SIMD_INLINE v256 v256_wideshuffle_8(v256 a, v256 b, v256 pattern) { + v256 c32 = v256_dup_8(32); + v256 p32 = v256_sub_8(pattern, c32); + v256 r1 = _mm256_blendv_epi8( + _mm256_shuffle_epi8( + _mm256_permute2x128_si256(a, b, _MM_SHUFFLE(0, 1, 0, 1)), p32), + _mm256_shuffle_epi8( + _mm256_permute2x128_si256(a, b, _MM_SHUFFLE(0, 0, 0, 0)), p32), + _mm256_cmpgt_epi8(v256_dup_8(48), pattern)); + v256 r2 = _mm256_blendv_epi8( + _mm256_shuffle_epi8( + _mm256_permute2x128_si256(a, b, _MM_SHUFFLE(0, 3, 0, 3)), pattern), + _mm256_shuffle_epi8( + _mm256_permute2x128_si256(a, b, _MM_SHUFFLE(0, 2, 0, 2)), pattern), + _mm256_cmpgt_epi8(v256_dup_8(16), pattern)); + return _mm256_blendv_epi8(r1, r2, _mm256_cmpgt_epi8(c32, pattern)); +} + +SIMD_INLINE v256 v256_pshuffle_8(v256 a, v256 pattern) { + return _mm256_shuffle_epi8(a, pattern); +} + +SIMD_INLINE int64_t v256_dotp_su8(v256 a, v256 b) { + v256 t1 = _mm256_madd_epi16(v256_unpackhi_s8_s16(a), v256_unpackhi_u8_s16(b)); + v256 t2 = _mm256_madd_epi16(v256_unpacklo_s8_s16(a), v256_unpacklo_u8_s16(b)); + t1 = _mm256_add_epi32(t1, t2); + v128 t = _mm_add_epi32(_mm256_extracti128_si256(t1, 0), + _mm256_extracti128_si256(t1, 1)); + t = _mm_add_epi32(t, _mm_srli_si128(t, 8)); + t = _mm_add_epi32(t, _mm_srli_si128(t, 4)); + return (int32_t)v128_low_u32(t); +} + +SIMD_INLINE int64_t v256_dotp_s16(v256 a, v256 b) { + v256 r = _mm256_madd_epi16(a, b); +#if defined(__x86_64__) + v128 t; + r = _mm256_add_epi64(_mm256_cvtepi32_epi64(v256_high_v128(r)), + _mm256_cvtepi32_epi64(v256_low_v128(r))); + t = v256_low_v128(_mm256_add_epi64( + r, _mm256_permute2x128_si256(r, r, _MM_SHUFFLE(2, 0, 0, 1)))); + return _mm_cvtsi128_si64(_mm_add_epi64(t, _mm_srli_si128(t, 8))); +#else + v128 l = v256_low_v128(r); + v128 h = v256_high_v128(r); + return (int64_t)_mm_cvtsi128_si32(l) + + (int64_t)_mm_cvtsi128_si32(_mm_srli_si128(l, 4)) + + (int64_t)_mm_cvtsi128_si32(_mm_srli_si128(l, 8)) + + (int64_t)_mm_cvtsi128_si32(_mm_srli_si128(l, 12)) + + (int64_t)_mm_cvtsi128_si32(h) + + (int64_t)_mm_cvtsi128_si32(_mm_srli_si128(h, 4)) + + (int64_t)_mm_cvtsi128_si32(_mm_srli_si128(h, 8)) + + (int64_t)_mm_cvtsi128_si32(_mm_srli_si128(h, 12)); +#endif +} + +SIMD_INLINE int64_t v256_dotp_s32(v256 a, v256 b) { + v256 r = _mm256_mullo_epi32(a, b); +#if defined(__x86_64__) + v128 t; + r = _mm256_add_epi64(_mm256_cvtepi32_epi64(v256_high_v128(r)), + _mm256_cvtepi32_epi64(v256_low_v128(r))); + t = v256_low_v128(_mm256_add_epi64( + r, _mm256_permute2x128_si256(r, r, _MM_SHUFFLE(2, 0, 0, 1)))); + return _mm_cvtsi128_si64(_mm_add_epi64(t, _mm_srli_si128(t, 8))); +#else + v128 l = v256_low_v128(r); + v128 h = v256_high_v128(r); + return (int64_t)_mm_cvtsi128_si32(l) + + (int64_t)_mm_cvtsi128_si32(_mm_srli_si128(l, 4)) + + (int64_t)_mm_cvtsi128_si32(_mm_srli_si128(l, 8)) + + (int64_t)_mm_cvtsi128_si32(_mm_srli_si128(l, 12)) + + (int64_t)_mm_cvtsi128_si32(h) + + (int64_t)_mm_cvtsi128_si32(_mm_srli_si128(h, 4)) + + (int64_t)_mm_cvtsi128_si32(_mm_srli_si128(h, 8)) + + (int64_t)_mm_cvtsi128_si32(_mm_srli_si128(h, 12)); +#endif +} + +SIMD_INLINE uint64_t v256_hadd_u8(v256 a) { + v256 t = _mm256_sad_epu8(a, _mm256_setzero_si256()); + v128 lo = v256_low_v128(t); + v128 hi = v256_high_v128(t); + lo = v128_add_32(lo, hi); + return v64_low_u32(v128_low_v64(lo)) + v128_low_u32(v128_high_v64(lo)); +} + +typedef v256 sad256_internal; + +SIMD_INLINE sad256_internal v256_sad_u8_init() { + return _mm256_setzero_si256(); +} + +/* Implementation dependent return value. Result must be finalised with + v256_sad_u8_sum(). + The result for more than 32 v256_sad_u8() calls is undefined. */ +SIMD_INLINE sad256_internal v256_sad_u8(sad256_internal s, v256 a, v256 b) { + return _mm256_add_epi64(s, _mm256_sad_epu8(a, b)); +} + +SIMD_INLINE uint32_t v256_sad_u8_sum(sad256_internal s) { + v256 t = _mm256_add_epi32(s, _mm256_unpackhi_epi64(s, s)); + return v128_low_u32(_mm_add_epi32(v256_high_v128(t), v256_low_v128(t))); +} + +typedef v256 ssd256_internal; + +SIMD_INLINE ssd256_internal v256_ssd_u8_init() { + return _mm256_setzero_si256(); +} + +/* Implementation dependent return value. Result must be finalised with + * v256_ssd_u8_sum(). */ +SIMD_INLINE ssd256_internal v256_ssd_u8(ssd256_internal s, v256 a, v256 b) { + v256 l = _mm256_sub_epi16(_mm256_unpacklo_epi8(a, _mm256_setzero_si256()), + _mm256_unpacklo_epi8(b, _mm256_setzero_si256())); + v256 h = _mm256_sub_epi16(_mm256_unpackhi_epi8(a, _mm256_setzero_si256()), + _mm256_unpackhi_epi8(b, _mm256_setzero_si256())); + v256 rl = _mm256_madd_epi16(l, l); + v256 rh = _mm256_madd_epi16(h, h); + v128 c = _mm_cvtsi32_si128(32); + rl = _mm256_add_epi32(rl, _mm256_srli_si256(rl, 8)); + rl = _mm256_add_epi32(rl, _mm256_srli_si256(rl, 4)); + rh = _mm256_add_epi32(rh, _mm256_srli_si256(rh, 8)); + rh = _mm256_add_epi32(rh, _mm256_srli_si256(rh, 4)); + return _mm256_add_epi64( + s, + _mm256_srl_epi64(_mm256_sll_epi64(_mm256_unpacklo_epi64(rl, rh), c), c)); +} + +SIMD_INLINE uint32_t v256_ssd_u8_sum(ssd256_internal s) { + v256 t = _mm256_add_epi32(s, _mm256_unpackhi_epi64(s, s)); + return v128_low_u32(_mm_add_epi32(v256_high_v128(t), v256_low_v128(t))); +} + +SIMD_INLINE v256 v256_or(v256 a, v256 b) { return _mm256_or_si256(a, b); } + +SIMD_INLINE v256 v256_xor(v256 a, v256 b) { return _mm256_xor_si256(a, b); } + +SIMD_INLINE v256 v256_and(v256 a, v256 b) { return _mm256_and_si256(a, b); } + +SIMD_INLINE v256 v256_andn(v256 a, v256 b) { return _mm256_andnot_si256(b, a); } + +SIMD_INLINE v256 v256_mul_s16(v64 a, v64 b) { + v128 lo_bits = v128_mullo_s16(a, b); + v128 hi_bits = v128_mulhi_s16(a, b); + return v256_from_v128(v128_ziphi_16(hi_bits, lo_bits), + v128_ziplo_16(hi_bits, lo_bits)); +} + +SIMD_INLINE v256 v256_mullo_s16(v256 a, v256 b) { + return _mm256_mullo_epi16(a, b); +} + +SIMD_INLINE v256 v256_mulhi_s16(v256 a, v256 b) { + return _mm256_mulhi_epi16(a, b); +} + +SIMD_INLINE v256 v256_mullo_s32(v256 a, v256 b) { + return _mm256_mullo_epi32(a, b); +} + +SIMD_INLINE v256 v256_madd_s16(v256 a, v256 b) { + return _mm256_madd_epi16(a, b); +} + +SIMD_INLINE v256 v256_madd_us8(v256 a, v256 b) { + return _mm256_maddubs_epi16(a, b); +} + +SIMD_INLINE v256 v256_avg_u8(v256 a, v256 b) { return _mm256_avg_epu8(a, b); } + +SIMD_INLINE v256 v256_rdavg_u8(v256 a, v256 b) { + return _mm256_sub_epi8( + _mm256_avg_epu8(a, b), + _mm256_and_si256(_mm256_xor_si256(a, b), v256_dup_8(1))); +} + +SIMD_INLINE v256 v256_rdavg_u16(v256 a, v256 b) { + return _mm256_sub_epi16( + _mm256_avg_epu16(a, b), + _mm256_and_si256(_mm256_xor_si256(a, b), v256_dup_16(1))); +} + +SIMD_INLINE v256 v256_avg_u16(v256 a, v256 b) { return _mm256_avg_epu16(a, b); } + +SIMD_INLINE v256 v256_min_u8(v256 a, v256 b) { return _mm256_min_epu8(a, b); } + +SIMD_INLINE v256 v256_max_u8(v256 a, v256 b) { return _mm256_max_epu8(a, b); } + +SIMD_INLINE v256 v256_min_s8(v256 a, v256 b) { return _mm256_min_epi8(a, b); } + +SIMD_INLINE uint32_t v256_movemask_8(v256 a) { return _mm256_movemask_epi8(a); } + +SIMD_INLINE v256 v256_blend_8(v256 a, v256 b, v256 c) { + return _mm256_blendv_epi8(a, b, c); +} + +SIMD_INLINE v256 v256_max_s8(v256 a, v256 b) { return _mm256_max_epi8(a, b); } + +SIMD_INLINE v256 v256_min_s16(v256 a, v256 b) { return _mm256_min_epi16(a, b); } + +SIMD_INLINE v256 v256_max_s16(v256 a, v256 b) { return _mm256_max_epi16(a, b); } + +SIMD_INLINE v256 v256_min_s32(v256 a, v256 b) { return _mm256_min_epi32(a, b); } + +SIMD_INLINE v256 v256_max_s32(v256 a, v256 b) { return _mm256_max_epi32(a, b); } + +SIMD_INLINE v256 v256_cmpgt_s8(v256 a, v256 b) { + return _mm256_cmpgt_epi8(a, b); +} + +SIMD_INLINE v256 v256_cmplt_s8(v256 a, v256 b) { + return _mm256_cmpgt_epi8(b, a); +} + +SIMD_INLINE v256 v256_cmpeq_8(v256 a, v256 b) { + return _mm256_cmpeq_epi8(a, b); +} + +SIMD_INLINE v256 v256_cmpgt_s16(v256 a, v256 b) { + return _mm256_cmpgt_epi16(a, b); +} + +SIMD_INLINE v256 v256_cmplt_s16(v256 a, v256 b) { + return _mm256_cmpgt_epi16(b, a); +} + +SIMD_INLINE v256 v256_cmpeq_16(v256 a, v256 b) { + return _mm256_cmpeq_epi16(a, b); +} + +SIMD_INLINE v256 v256_cmpgt_s32(v256 a, v256 b) { + return _mm256_cmpgt_epi32(a, b); +} + +SIMD_INLINE v256 v256_cmplt_s32(v256 a, v256 b) { + return _mm256_cmpgt_epi32(b, a); +} + +SIMD_INLINE v256 v256_cmpeq_32(v256 a, v256 b) { + return _mm256_cmpeq_epi32(a, b); +} + +SIMD_INLINE v256 v256_shl_8(v256 a, unsigned int c) { + return _mm256_and_si256(_mm256_set1_epi8((uint8_t)(0xff << c)), + _mm256_sll_epi16(a, _mm_cvtsi32_si128(c))); +} + +SIMD_INLINE v256 v256_shr_u8(v256 a, unsigned int c) { + return _mm256_and_si256(_mm256_set1_epi8(0xff >> c), + _mm256_srl_epi16(a, _mm_cvtsi32_si128(c))); +} + +SIMD_INLINE v256 v256_shr_s8(v256 a, unsigned int c) { + __m128i x = _mm_cvtsi32_si128(c + 8); + return _mm256_packs_epi16(_mm256_sra_epi16(_mm256_unpacklo_epi8(a, a), x), + _mm256_sra_epi16(_mm256_unpackhi_epi8(a, a), x)); +} + +SIMD_INLINE v256 v256_shl_16(v256 a, unsigned int c) { + return _mm256_sll_epi16(a, _mm_cvtsi32_si128(c)); +} + +SIMD_INLINE v256 v256_shr_u16(v256 a, unsigned int c) { + return _mm256_srl_epi16(a, _mm_cvtsi32_si128(c)); +} + +SIMD_INLINE v256 v256_shr_s16(v256 a, unsigned int c) { + return _mm256_sra_epi16(a, _mm_cvtsi32_si128(c)); +} + +SIMD_INLINE v256 v256_shl_32(v256 a, unsigned int c) { + return _mm256_sll_epi32(a, _mm_cvtsi32_si128(c)); +} + +SIMD_INLINE v256 v256_shr_u32(v256 a, unsigned int c) { + return _mm256_srl_epi32(a, _mm_cvtsi32_si128(c)); +} + +SIMD_INLINE v256 v256_shr_s32(v256 a, unsigned int c) { + return _mm256_sra_epi32(a, _mm_cvtsi32_si128(c)); +} + +SIMD_INLINE v256 v256_shl_64(v256 a, unsigned int c) { + return _mm256_sll_epi64(a, _mm_cvtsi32_si128(c)); +} + +SIMD_INLINE v256 v256_shr_u64(v256 a, unsigned int c) { + return _mm256_srl_epi64(a, _mm_cvtsi32_si128(c)); +} + +SIMD_INLINE v256 v256_shr_s64(v256 a, unsigned int c) { +#if defined(__AVX512F__) + return _mm256_sra_epi64(a, _mm_cvtsi32_si128(c)); +#else + return v256_from_v128(v128_shr_s64(v256_high_v128(a), c), + v128_shr_s64(v256_low_v128(a), c)); +#endif +} + +/* These intrinsics require immediate values, so we must use #defines + to enforce that. */ +// _mm256_slli_si256 works on 128 bit lanes and can't be used +#define v256_shl_n_byte(a, n) \ + ((n) < 16 ? v256_from_v128( \ + v128_align(v256_high_v128(a), v256_low_v128(a), 16 - (n)), \ + v128_shl_n_byte(v256_low_v128(a), n)) \ + : _mm256_inserti128_si256( \ + _mm256_setzero_si256(), \ + v128_shl_n_byte(v256_low_v128(a), (n)-16), 1)) + +// _mm256_srli_si256 works on 128 bit lanes and can't be used +#define v256_shr_n_byte(a, n) \ + ((n) < 16 \ + ? _mm256_alignr_epi8( \ + _mm256_permute2x128_si256(a, a, _MM_SHUFFLE(2, 0, 0, 1)), a, n) \ + : _mm256_inserti128_si256( \ + _mm256_setzero_si256(), \ + v128_align(v256_high_v128(a), v256_high_v128(a), n), 0)) + +// _mm256_alignr_epi8 works on two 128 bit lanes and can't be used +#define v256_align(a, b, c) \ + ((c) ? v256_or(v256_shr_n_byte(b, c), v256_shl_n_byte(a, 32 - c)) : b) + +#define v256_shl_n_8(a, c) \ + _mm256_and_si256(_mm256_set1_epi8((uint8_t)(0xff << (c))), \ + _mm256_slli_epi16(a, c)) +#define v256_shr_n_u8(a, c) \ + _mm256_and_si256(_mm256_set1_epi8(0xff >> (c)), _mm256_srli_epi16(a, c)) +#define v256_shr_n_s8(a, c) \ + _mm256_packs_epi16(_mm256_srai_epi16(_mm256_unpacklo_epi8(a, a), (c) + 8), \ + _mm256_srai_epi16(_mm256_unpackhi_epi8(a, a), (c) + 8)) +#define v256_shl_n_16(a, c) _mm256_slli_epi16(a, c) +#define v256_shr_n_u16(a, c) _mm256_srli_epi16(a, c) +#define v256_shr_n_s16(a, c) _mm256_srai_epi16(a, c) +#define v256_shl_n_32(a, c) _mm256_slli_epi32(a, c) +#define v256_shr_n_u32(a, c) _mm256_srli_epi32(a, c) +#define v256_shr_n_s32(a, c) _mm256_srai_epi32(a, c) +#define v256_shl_n_64(a, c) _mm256_slli_epi64(a, c) +#define v256_shr_n_u64(a, c) _mm256_srli_epi64(a, c) +#define v256_shr_n_s64(a, c) \ + v256_shr_s64((a), (c)) // _mm256_srai_epi64 broken in gcc? +#define v256_shr_n_word(a, n) v256_shr_n_byte(a, 2 * (n)) +#define v256_shl_n_word(a, n) v256_shl_n_byte(a, 2 * (n)) + +typedef v256 sad256_internal_u16; + +SIMD_INLINE sad256_internal_u16 v256_sad_u16_init() { return v256_zero(); } + +/* Implementation dependent return value. Result must be finalised with + * v256_sad_u16_sum(). */ +SIMD_INLINE sad256_internal_u16 v256_sad_u16(sad256_internal_u16 s, v256 a, + v256 b) { +#if defined(__SSE4_1__) + v256 t = v256_sub_16(_mm256_max_epu16(a, b), _mm256_min_epu16(a, b)); +#else + v256 t = v256_cmplt_s16(v256_xor(a, v256_dup_16(32768)), + v256_xor(b, v256_dup_16(32768))); + t = v256_sub_16(v256_or(v256_and(b, t), v256_andn(a, t)), + v256_or(v256_and(a, t), v256_andn(b, t))); +#endif + return v256_add_32( + s, v256_add_32(v256_unpackhi_u16_s32(t), v256_unpacklo_u16_s32(t))); +} + +SIMD_INLINE uint32_t v256_sad_u16_sum(sad256_internal_u16 s) { + v128 t = v128_add_32(v256_high_v128(s), v256_low_v128(s)); + return v128_low_u32(t) + v128_low_u32(v128_shr_n_byte(t, 4)) + + v128_low_u32(v128_shr_n_byte(t, 8)) + + v128_low_u32(v128_shr_n_byte(t, 12)); +} + +typedef v256 ssd256_internal_s16; + +SIMD_INLINE ssd256_internal_s16 v256_ssd_s16_init() { return v256_zero(); } + +/* Implementation dependent return value. Result must be finalised with + * v256_ssd_s16_sum(). */ +SIMD_INLINE ssd256_internal_s16 v256_ssd_s16(ssd256_internal_s16 s, v256 a, + v256 b) { + v256 d = v256_sub_16(a, b); + d = v256_madd_s16(d, d); + return v256_add_64(s, v256_add_64(_mm256_unpackhi_epi32(d, v256_zero()), + _mm256_unpacklo_epi32(d, v256_zero()))); +} + +SIMD_INLINE uint64_t v256_ssd_s16_sum(ssd256_internal_s16 s) { + v128 t = v128_add_64(v256_high_v128(s), v256_low_v128(s)); + return v64_u64(v128_low_v64(t)) + v64_u64(v128_high_v64(t)); +} + +#endif + +#endif // AOM_AOM_DSP_SIMD_V256_INTRINSICS_X86_H_ diff --git a/third_party/aom/aom_dsp/simd/v64_intrinsics.h b/third_party/aom/aom_dsp/simd/v64_intrinsics.h new file mode 100644 index 0000000000..afc55428d8 --- /dev/null +++ b/third_party/aom/aom_dsp/simd/v64_intrinsics.h @@ -0,0 +1,232 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_SIMD_V64_INTRINSICS_H_ +#define AOM_AOM_DSP_SIMD_V64_INTRINSICS_H_ + +#include +#include + +#include "aom_dsp/simd/v64_intrinsics_c.h" + +/* Fallback to plain, unoptimised C. */ + +typedef c_v64 v64; + +SIMD_INLINE uint32_t v64_low_u32(v64 a) { return c_v64_low_u32(a); } +SIMD_INLINE uint32_t v64_high_u32(v64 a) { return c_v64_high_u32(a); } +SIMD_INLINE int32_t v64_low_s32(v64 a) { return c_v64_low_s32(a); } +SIMD_INLINE int32_t v64_high_s32(v64 a) { return c_v64_high_s32(a); } +SIMD_INLINE v64 v64_from_32(uint32_t x, uint32_t y) { + return c_v64_from_32(x, y); +} +SIMD_INLINE v64 v64_from_64(uint64_t x) { return c_v64_from_64(x); } +SIMD_INLINE uint64_t v64_u64(v64 x) { return c_v64_u64(x); } +SIMD_INLINE v64 v64_from_16(uint16_t a, uint16_t b, uint16_t c, uint16_t d) { + return c_v64_from_16(a, b, c, d); +} + +SIMD_INLINE uint32_t u32_load_unaligned(const void *p) { + return c_u32_load_unaligned(p); +} +SIMD_INLINE uint32_t u32_load_aligned(const void *p) { + return c_u32_load_aligned(p); +} +SIMD_INLINE void u32_store_unaligned(void *p, uint32_t a) { + c_u32_store_unaligned(p, a); +} +SIMD_INLINE void u32_store_aligned(void *p, uint32_t a) { + c_u32_store_aligned(p, a); +} + +SIMD_INLINE v64 v64_load_unaligned(const void *p) { + return c_v64_load_unaligned(p); +} +SIMD_INLINE v64 v64_load_aligned(const void *p) { + return c_v64_load_aligned(p); +} + +SIMD_INLINE void v64_store_unaligned(void *p, v64 a) { + c_v64_store_unaligned(p, a); +} +SIMD_INLINE void v64_store_aligned(void *p, v64 a) { + c_v64_store_aligned(p, a); +} + +SIMD_INLINE v64 v64_align(v64 a, v64 b, unsigned int c) { + return c_v64_align(a, b, c); +} + +SIMD_INLINE v64 v64_zero() { return c_v64_zero(); } +SIMD_INLINE v64 v64_dup_8(uint8_t x) { return c_v64_dup_8(x); } +SIMD_INLINE v64 v64_dup_16(uint16_t x) { return c_v64_dup_16(x); } +SIMD_INLINE v64 v64_dup_32(uint32_t x) { return c_v64_dup_32(x); } + +SIMD_INLINE v64 v64_add_8(v64 a, v64 b) { return c_v64_add_8(a, b); } +SIMD_INLINE v64 v64_add_16(v64 a, v64 b) { return c_v64_add_16(a, b); } +SIMD_INLINE v64 v64_sadd_u8(v64 a, v64 b) { return c_v64_sadd_u8(a, b); } +SIMD_INLINE v64 v64_sadd_s8(v64 a, v64 b) { return c_v64_sadd_s8(a, b); } +SIMD_INLINE v64 v64_sadd_s16(v64 a, v64 b) { return c_v64_sadd_s16(a, b); } +SIMD_INLINE v64 v64_add_32(v64 a, v64 b) { return c_v64_add_32(a, b); } +SIMD_INLINE v64 v64_sub_8(v64 a, v64 b) { return c_v64_sub_8(a, b); } +SIMD_INLINE v64 v64_ssub_u8(v64 a, v64 b) { return c_v64_ssub_u8(a, b); } +SIMD_INLINE v64 v64_ssub_s8(v64 a, v64 b) { return c_v64_ssub_s8(a, b); } +SIMD_INLINE v64 v64_sub_16(v64 a, v64 b) { return c_v64_sub_16(a, b); } +SIMD_INLINE v64 v64_ssub_s16(v64 a, v64 b) { return c_v64_ssub_s16(a, b); } +SIMD_INLINE v64 v64_ssub_u16(v64 a, v64 b) { return c_v64_ssub_u16(a, b); } +SIMD_INLINE v64 v64_sub_32(v64 a, v64 b) { return c_v64_sub_32(a, b); } +SIMD_INLINE v64 v64_abs_s16(v64 a) { return c_v64_abs_s16(a); } +SIMD_INLINE v64 v64_abs_s8(v64 a) { return c_v64_abs_s8(a); } + +SIMD_INLINE v64 v64_ziplo_8(v64 a, v64 b) { return c_v64_ziplo_8(a, b); } +SIMD_INLINE v64 v64_ziphi_8(v64 a, v64 b) { return c_v64_ziphi_8(a, b); } +SIMD_INLINE v64 v64_ziplo_16(v64 a, v64 b) { return c_v64_ziplo_16(a, b); } +SIMD_INLINE v64 v64_ziphi_16(v64 a, v64 b) { return c_v64_ziphi_16(a, b); } +SIMD_INLINE v64 v64_ziplo_32(v64 a, v64 b) { return c_v64_ziplo_32(a, b); } +SIMD_INLINE v64 v64_ziphi_32(v64 a, v64 b) { return c_v64_ziphi_32(a, b); } +SIMD_INLINE v64 v64_unziplo_8(v64 a, v64 b) { return c_v64_unziplo_8(a, b); } +SIMD_INLINE v64 v64_unziphi_8(v64 a, v64 b) { return c_v64_unziphi_8(a, b); } +SIMD_INLINE v64 v64_unziplo_16(v64 a, v64 b) { return c_v64_unziplo_16(a, b); } +SIMD_INLINE v64 v64_unziphi_16(v64 a, v64 b) { return c_v64_unziphi_16(a, b); } +SIMD_INLINE v64 v64_unpacklo_u8_s16(v64 a) { return c_v64_unpacklo_u8_s16(a); } +SIMD_INLINE v64 v64_unpackhi_u8_s16(v64 a) { return c_v64_unpackhi_u8_s16(a); } +SIMD_INLINE v64 v64_unpacklo_s8_s16(v64 a) { return c_v64_unpacklo_s8_s16(a); } +SIMD_INLINE v64 v64_unpackhi_s8_s16(v64 a) { return c_v64_unpackhi_s8_s16(a); } +SIMD_INLINE v64 v64_pack_s32_s16(v64 a, v64 b) { + return c_v64_pack_s32_s16(a, b); +} +SIMD_INLINE v64 v64_pack_s32_u16(v64 a, v64 b) { + return c_v64_pack_s32_u16(a, b); +} +SIMD_INLINE v64 v64_pack_s16_u8(v64 a, v64 b) { + return c_v64_pack_s16_u8(a, b); +} +SIMD_INLINE v64 v64_pack_s16_s8(v64 a, v64 b) { + return c_v64_pack_s16_s8(a, b); +} +SIMD_INLINE v64 v64_unpacklo_u16_s32(v64 a) { + return c_v64_unpacklo_u16_s32(a); +} +SIMD_INLINE v64 v64_unpacklo_s16_s32(v64 a) { + return c_v64_unpacklo_s16_s32(a); +} +SIMD_INLINE v64 v64_unpackhi_u16_s32(v64 a) { + return c_v64_unpackhi_u16_s32(a); +} +SIMD_INLINE v64 v64_unpackhi_s16_s32(v64 a) { + return c_v64_unpackhi_s16_s32(a); +} +SIMD_INLINE v64 v64_shuffle_8(v64 a, v64 pattern) { + return c_v64_shuffle_8(a, pattern); +} + +typedef uint32_t sad64_internal; +SIMD_INLINE sad64_internal v64_sad_u8_init() { return c_v64_sad_u8_init(); } +SIMD_INLINE sad64_internal v64_sad_u8(sad64_internal s, v64 a, v64 b) { + return c_v64_sad_u8(s, a, b); +} +SIMD_INLINE uint32_t v64_sad_u8_sum(sad64_internal s) { + return c_v64_sad_u8_sum(s); +} +typedef uint32_t ssd64_internal; +SIMD_INLINE ssd64_internal v64_ssd_u8_init() { return c_v64_ssd_u8_init(); } +SIMD_INLINE ssd64_internal v64_ssd_u8(ssd64_internal s, v64 a, v64 b) { + return c_v64_ssd_u8(s, a, b); +} +SIMD_INLINE uint32_t v64_ssd_u8_sum(ssd64_internal s) { + return c_v64_ssd_u8_sum(s); +} +SIMD_INLINE int64_t v64_dotp_su8(v64 a, v64 b) { return c_v64_dotp_su8(a, b); } +SIMD_INLINE int64_t v64_dotp_s16(v64 a, v64 b) { return c_v64_dotp_s16(a, b); } +SIMD_INLINE uint64_t v64_hadd_u8(v64 a) { return c_v64_hadd_u8(a); } +SIMD_INLINE int64_t v64_hadd_s16(v64 a) { return c_v64_hadd_s16(a); } + +SIMD_INLINE v64 v64_or(v64 a, v64 b) { return c_v64_or(a, b); } +SIMD_INLINE v64 v64_xor(v64 a, v64 b) { return c_v64_xor(a, b); } +SIMD_INLINE v64 v64_and(v64 a, v64 b) { return c_v64_and(a, b); } +SIMD_INLINE v64 v64_andn(v64 a, v64 b) { return c_v64_andn(a, b); } + +SIMD_INLINE v64 v64_mullo_s16(v64 a, v64 b) { return c_v64_mullo_s16(a, b); } +SIMD_INLINE v64 v64_mulhi_s16(v64 a, v64 b) { return c_v64_mulhi_s16(a, b); } +SIMD_INLINE v64 v64_mullo_s32(v64 a, v64 b) { return c_v64_mullo_s32(a, b); } +SIMD_INLINE v64 v64_madd_s16(v64 a, v64 b) { return c_v64_madd_s16(a, b); } +SIMD_INLINE v64 v64_madd_us8(v64 a, v64 b) { return c_v64_madd_us8(a, b); } + +SIMD_INLINE v64 v64_avg_u8(v64 a, v64 b) { return c_v64_avg_u8(a, b); } +SIMD_INLINE v64 v64_rdavg_u8(v64 a, v64 b) { return c_v64_rdavg_u8(a, b); } +SIMD_INLINE v64 v64_rdavg_u16(v64 a, v64 b) { return c_v64_rdavg_u16(a, b); } +SIMD_INLINE v64 v64_avg_u16(v64 a, v64 b) { return c_v64_avg_u16(a, b); } +SIMD_INLINE v64 v64_min_u8(v64 a, v64 b) { return c_v64_min_u8(a, b); } +SIMD_INLINE v64 v64_max_u8(v64 a, v64 b) { return c_v64_max_u8(a, b); } +SIMD_INLINE v64 v64_min_s8(v64 a, v64 b) { return c_v64_min_s8(a, b); } +SIMD_INLINE v64 v64_max_s8(v64 a, v64 b) { return c_v64_max_s8(a, b); } +SIMD_INLINE v64 v64_min_s16(v64 a, v64 b) { return c_v64_min_s16(a, b); } +SIMD_INLINE v64 v64_max_s16(v64 a, v64 b) { return c_v64_max_s16(a, b); } + +SIMD_INLINE v64 v64_cmpgt_s8(v64 a, v64 b) { return c_v64_cmpgt_s8(a, b); } +SIMD_INLINE v64 v64_cmplt_s8(v64 a, v64 b) { return c_v64_cmplt_s8(a, b); } +SIMD_INLINE v64 v64_cmpeq_8(v64 a, v64 b) { return c_v64_cmpeq_8(a, b); } +SIMD_INLINE v64 v64_cmpgt_s16(v64 a, v64 b) { return c_v64_cmpgt_s16(a, b); } +SIMD_INLINE v64 v64_cmplt_s16(v64 a, v64 b) { return c_v64_cmplt_s16(a, b); } +SIMD_INLINE v64 v64_cmpeq_16(v64 a, v64 b) { return c_v64_cmpeq_16(a, b); } + +SIMD_INLINE v64 v64_shl_8(v64 a, unsigned int n) { return c_v64_shl_8(a, n); } +SIMD_INLINE v64 v64_shr_u8(v64 a, unsigned int n) { return c_v64_shr_u8(a, n); } +SIMD_INLINE v64 v64_shr_s8(v64 a, unsigned int n) { return c_v64_shr_s8(a, n); } +SIMD_INLINE v64 v64_shl_16(v64 a, unsigned int n) { return c_v64_shl_16(a, n); } +SIMD_INLINE v64 v64_shr_u16(v64 a, unsigned int n) { + return c_v64_shr_u16(a, n); +} +SIMD_INLINE v64 v64_shr_s16(v64 a, unsigned int n) { + return c_v64_shr_s16(a, n); +} +SIMD_INLINE v64 v64_shl_32(v64 a, unsigned int n) { return c_v64_shl_32(a, n); } +SIMD_INLINE v64 v64_shr_u32(v64 a, unsigned int n) { + return c_v64_shr_u32(a, n); +} +SIMD_INLINE v64 v64_shr_s32(v64 a, unsigned int n) { + return c_v64_shr_s32(a, n); +} +SIMD_INLINE v64 v64_shr_n_byte(v64 a, unsigned int n) { + return c_v64_shr_n_byte(a, n); +} +SIMD_INLINE v64 v64_shl_n_byte(v64 a, unsigned int n) { + return c_v64_shl_n_byte(a, n); +} +SIMD_INLINE v64 v64_shl_n_8(v64 a, unsigned int c) { + return c_v64_shl_n_8(a, c); +} +SIMD_INLINE v64 v64_shr_n_u8(v64 a, unsigned int c) { + return c_v64_shr_n_u8(a, c); +} +SIMD_INLINE v64 v64_shr_n_s8(v64 a, unsigned int c) { + return c_v64_shr_n_s8(a, c); +} +SIMD_INLINE v64 v64_shl_n_16(v64 a, unsigned int c) { + return c_v64_shl_n_16(a, c); +} +SIMD_INLINE v64 v64_shr_n_u16(v64 a, unsigned int c) { + return c_v64_shr_n_u16(a, c); +} +SIMD_INLINE v64 v64_shr_n_s16(v64 a, unsigned int c) { + return c_v64_shr_n_s16(a, c); +} +SIMD_INLINE v64 v64_shl_n_32(v64 a, unsigned int c) { + return c_v64_shl_n_32(a, c); +} +SIMD_INLINE v64 v64_shr_n_u32(v64 a, unsigned int c) { + return c_v64_shr_n_u32(a, c); +} +SIMD_INLINE v64 v64_shr_n_s32(v64 a, unsigned int c) { + return c_v64_shr_n_s32(a, c); +} + +#endif // AOM_AOM_DSP_SIMD_V64_INTRINSICS_H_ diff --git a/third_party/aom/aom_dsp/simd/v64_intrinsics_arm.h b/third_party/aom/aom_dsp/simd/v64_intrinsics_arm.h new file mode 100644 index 0000000000..8f39ad6e81 --- /dev/null +++ b/third_party/aom/aom_dsp/simd/v64_intrinsics_arm.h @@ -0,0 +1,680 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_SIMD_V64_INTRINSICS_ARM_H_ +#define AOM_AOM_DSP_SIMD_V64_INTRINSICS_ARM_H_ + +#include + +#include "aom_dsp/simd/v64_intrinsics_arm.h" +#include "aom_ports/arm.h" + +#ifdef AOM_INCOMPATIBLE_GCC +#error Incompatible gcc +#endif + +typedef int64x1_t v64; + +SIMD_INLINE uint32_t v64_low_u32(v64 a) { + return vget_lane_u32(vreinterpret_u32_s64(a), 0); +} + +SIMD_INLINE uint32_t v64_high_u32(v64 a) { + return vget_lane_u32(vreinterpret_u32_s64(a), 1); +} + +SIMD_INLINE int32_t v64_low_s32(v64 a) { + return vget_lane_s32(vreinterpret_s32_s64(a), 0); +} + +SIMD_INLINE int32_t v64_high_s32(v64 a) { + return vget_lane_s32(vreinterpret_s32_s64(a), 1); +} + +SIMD_INLINE v64 v64_from_16(uint16_t a, uint16_t b, uint16_t c, uint16_t d) { + return vcreate_s64((uint64_t)a << 48 | (uint64_t)b << 32 | (uint64_t)c << 16 | + d); +} + +SIMD_INLINE v64 v64_from_32(uint32_t x, uint32_t y) { + return vcreate_s64((uint64_t)x << 32 | y); +} + +SIMD_INLINE v64 v64_from_64(uint64_t x) { return vcreate_s64(x); } + +SIMD_INLINE uint64_t v64_u64(v64 x) { return (uint64_t)x; } + +SIMD_INLINE uint32_t u32_load_aligned(const void *p) { + return *((uint32_t *)p); +} + +SIMD_INLINE uint32_t u32_load_unaligned(const void *p) { + return vget_lane_u32(vreinterpret_u32_u8(vld1_u8((const uint8_t *)p)), 0); +} + +SIMD_INLINE void u32_store_aligned(void *p, uint32_t a) { + *((uint32_t *)p) = a; +} + +SIMD_INLINE void u32_store_unaligned(void *p, uint32_t a) { +#if defined(__clang__) + vst1_lane_u32((uint32_t *)p, vreinterpret_u32_s64((uint64x1_t)(uint64_t)a), + 0); +#elif defined(__CC_ARM) + *(__packed uint32_t *)p) = a; +#elif defined(__GNUC__) + *((__attribute((packed)) uint32_t *)p) = a; +#else + vst1_lane_u32((uint32_t *)p, vreinterpret_u32_s64((uint64x1_t)(uint64_t)a), + 0); +#endif +} + +SIMD_INLINE v64 v64_load_aligned(const void *p) { + return vreinterpret_s64_u8(vld1_u8((const uint8_t *)p)); +} + +SIMD_INLINE v64 v64_load_unaligned(const void *p) { + return v64_load_aligned(p); +} + +SIMD_INLINE void v64_store_aligned(void *p, v64 r) { + vst1_u8((uint8_t *)p, vreinterpret_u8_s64(r)); +} + +SIMD_INLINE void v64_store_unaligned(void *p, v64 r) { + vst1_u8((uint8_t *)p, vreinterpret_u8_s64(r)); +} + +// The following function requires an immediate. +// Some compilers will check this if it's optimising, others wont. +SIMD_INLINE v64 v64_align(v64 a, v64 b, unsigned int c) { +#if defined(__OPTIMIZE__) && __OPTIMIZE__ && !defined(__clang__) + return c ? vreinterpret_s64_s8( + vext_s8(vreinterpret_s8_s64(b), vreinterpret_s8_s64(a), c)) + : b; +#else + return c ? v64_from_64(((uint64_t)b >> c * 8) | ((uint64_t)a << (8 - c) * 8)) + : b; +#endif +} + +SIMD_INLINE v64 v64_zero() { return vreinterpret_s64_u8(vdup_n_u8(0)); } + +SIMD_INLINE v64 v64_dup_8(uint8_t x) { + return vreinterpret_s64_u8(vdup_n_u8(x)); +} + +SIMD_INLINE v64 v64_dup_16(uint16_t x) { + return vreinterpret_s64_u16(vdup_n_u16(x)); +} + +SIMD_INLINE v64 v64_dup_32(uint32_t x) { + return vreinterpret_s64_u32(vdup_n_u32(x)); +} + +SIMD_INLINE int64_t v64_dotp_su8(v64 x, v64 y) { + int16x8_t t = + vmulq_s16(vmovl_s8(vreinterpret_s8_s64(x)), + vreinterpretq_s16_u16(vmovl_u8(vreinterpret_u8_s64(y)))); +#if defined(__aarch64__) + return vaddlvq_s16(t); +#else + int64x2_t r = vpaddlq_s32(vpaddlq_s16(t)); + return (int64_t)vadd_s64(vget_high_s64(r), vget_low_s64(r)); +#endif +} + +SIMD_INLINE int64_t v64_dotp_s16(v64 x, v64 y) { +#if defined(__aarch64__) + return vaddlvq_s32( + vmull_s16(vreinterpret_s16_s64(x), vreinterpret_s16_s64(y))); +#else + int64x2_t r = + vpaddlq_s32(vmull_s16(vreinterpret_s16_s64(x), vreinterpret_s16_s64(y))); + return (int64_t)(vget_high_s64(r) + vget_low_s64(r)); +#endif +} + +SIMD_INLINE uint64_t v64_hadd_u8(v64 x) { +#if defined(__aarch64__) + return vaddlv_u8(vreinterpret_u8_s64(x)); +#else + return (uint64_t)vpaddl_u32(vpaddl_u16(vpaddl_u8(vreinterpret_u8_s64(x)))); +#endif +} + +SIMD_INLINE int64_t v64_hadd_s16(v64 a) { + return (int64_t)vpaddl_s32(vpaddl_s16(vreinterpret_s16_s64(a))); +} + +typedef uint16x8_t sad64_internal; + +SIMD_INLINE sad64_internal v64_sad_u8_init() { return vdupq_n_u16(0); } + +// Implementation dependent return value. Result must be finalised with +// v64_sad_u8_sum(). +SIMD_INLINE sad64_internal v64_sad_u8(sad64_internal s, v64 a, v64 b) { + return vabal_u8(s, vreinterpret_u8_s64(a), vreinterpret_u8_s64(b)); +} + +SIMD_INLINE uint32_t v64_sad_u8_sum(sad64_internal s) { +#if defined(__aarch64__) + return vaddlvq_u16(s); +#else + uint64x2_t r = vpaddlq_u32(vpaddlq_u16(s)); + return (uint32_t)(uint64_t)(vget_high_u64(r) + vget_low_u64(r)); +#endif +} + +typedef uint32x4_t ssd64_internal; + +SIMD_INLINE ssd64_internal v64_ssd_u8_init() { return vdupq_n_u32(0); } + +// Implementation dependent return value. Result must be finalised with +// v64_ssd_u8_sum(). +SIMD_INLINE ssd64_internal v64_ssd_u8(ssd64_internal s, v64 a, v64 b) { + uint8x8_t t = vabd_u8(vreinterpret_u8_s64(a), vreinterpret_u8_s64(b)); + return vaddq_u32(s, vpaddlq_u16(vmull_u8(t, t))); +} + +SIMD_INLINE uint32_t v64_ssd_u8_sum(ssd64_internal s) { +#if defined(__aarch64__) + return vaddvq_u32(s); +#else + uint64x2_t t = vpaddlq_u32(s); + return vget_lane_u32( + vreinterpret_u32_u64(vadd_u64(vget_high_u64(t), vget_low_u64(t))), 0); +#endif +} + +SIMD_INLINE v64 v64_or(v64 x, v64 y) { return vorr_s64(x, y); } + +SIMD_INLINE v64 v64_xor(v64 x, v64 y) { return veor_s64(x, y); } + +SIMD_INLINE v64 v64_and(v64 x, v64 y) { return vand_s64(x, y); } + +SIMD_INLINE v64 v64_andn(v64 x, v64 y) { return vbic_s64(x, y); } + +SIMD_INLINE v64 v64_add_8(v64 x, v64 y) { + return vreinterpret_s64_u8( + vadd_u8(vreinterpret_u8_s64(x), vreinterpret_u8_s64(y))); +} + +SIMD_INLINE v64 v64_sadd_u8(v64 x, v64 y) { + return vreinterpret_s64_u8( + vqadd_u8(vreinterpret_u8_s64(x), vreinterpret_u8_s64(y))); +} + +SIMD_INLINE v64 v64_sadd_s8(v64 x, v64 y) { + return vreinterpret_s64_s8( + vqadd_s8(vreinterpret_s8_s64(x), vreinterpret_s8_s64(y))); +} + +SIMD_INLINE v64 v64_add_16(v64 x, v64 y) { + return vreinterpret_s64_s16( + vadd_s16(vreinterpret_s16_s64(x), vreinterpret_s16_s64(y))); +} + +SIMD_INLINE v64 v64_sadd_s16(v64 x, v64 y) { + return vreinterpret_s64_s16( + vqadd_s16(vreinterpret_s16_s64(x), vreinterpret_s16_s64(y))); +} + +SIMD_INLINE v64 v64_add_32(v64 x, v64 y) { + return vreinterpret_s64_u32( + vadd_u32(vreinterpret_u32_s64(x), vreinterpret_u32_s64(y))); +} + +SIMD_INLINE v64 v64_sub_8(v64 x, v64 y) { + return vreinterpret_s64_u8( + vsub_u8(vreinterpret_u8_s64(x), vreinterpret_u8_s64(y))); +} + +SIMD_INLINE v64 v64_sub_16(v64 x, v64 y) { + return vreinterpret_s64_s16( + vsub_s16(vreinterpret_s16_s64(x), vreinterpret_s16_s64(y))); +} + +SIMD_INLINE v64 v64_ssub_s16(v64 x, v64 y) { + return vreinterpret_s64_s16( + vqsub_s16(vreinterpret_s16_s64(x), vreinterpret_s16_s64(y))); +} + +SIMD_INLINE v64 v64_ssub_u16(v64 x, v64 y) { + return vreinterpret_s64_u16( + vqsub_u16(vreinterpret_u16_s64(x), vreinterpret_u16_s64(y))); +} + +SIMD_INLINE v64 v64_ssub_u8(v64 x, v64 y) { + return vreinterpret_s64_u8( + vqsub_u8(vreinterpret_u8_s64(x), vreinterpret_u8_s64(y))); +} + +SIMD_INLINE v64 v64_ssub_s8(v64 x, v64 y) { + return vreinterpret_s64_s8( + vqsub_s8(vreinterpret_s8_s64(x), vreinterpret_s8_s64(y))); +} + +SIMD_INLINE v64 v64_sub_32(v64 x, v64 y) { + return vreinterpret_s64_s32( + vsub_s32(vreinterpret_s32_s64(x), vreinterpret_s32_s64(y))); +} + +SIMD_INLINE v64 v64_abs_s16(v64 x) { + return vreinterpret_s64_s16(vabs_s16(vreinterpret_s16_s64(x))); +} + +SIMD_INLINE v64 v64_abs_s8(v64 x) { + return vreinterpret_s64_s8(vabs_s8(vreinterpret_s8_s64(x))); +} + +SIMD_INLINE v64 v64_mullo_s16(v64 x, v64 y) { + return vreinterpret_s64_s16( + vmul_s16(vreinterpret_s16_s64(x), vreinterpret_s16_s64(y))); +} + +SIMD_INLINE v64 v64_mulhi_s16(v64 x, v64 y) { +#if defined(__aarch64__) + int16x8_t t = vreinterpretq_s16_s32( + vmull_s16(vreinterpret_s16_s64(x), vreinterpret_s16_s64(y))); + return vget_low_s64(vreinterpretq_s64_s16(vuzp2q_s16(t, t))); +#else + return vreinterpret_s64_s16(vmovn_s32(vshrq_n_s32( + vmull_s16(vreinterpret_s16_s64(x), vreinterpret_s16_s64(y)), 16))); +#endif +} + +SIMD_INLINE v64 v64_mullo_s32(v64 x, v64 y) { + return vreinterpret_s64_s32( + vmul_s32(vreinterpret_s32_s64(x), vreinterpret_s32_s64(y))); +} + +SIMD_INLINE v64 v64_madd_s16(v64 x, v64 y) { + int32x4_t t = vmull_s16(vreinterpret_s16_s64(x), vreinterpret_s16_s64(y)); + return vreinterpret_s64_s32( + vpadd_s32(vreinterpret_s32_s64(vget_low_s64(vreinterpretq_s64_s32(t))), + vreinterpret_s32_s64(vget_high_s64(vreinterpretq_s64_s32(t))))); +} + +SIMD_INLINE v64 v64_madd_us8(v64 x, v64 y) { + int16x8_t t = + vmulq_s16(vreinterpretq_s16_u16(vmovl_u8(vreinterpret_u8_s64(x))), + vmovl_s8(vreinterpret_s8_s64(y))); + return vreinterpret_s64_s16(vqmovn_s32(vpaddlq_s16(t))); +} + +SIMD_INLINE v64 v64_avg_u8(v64 x, v64 y) { + return vreinterpret_s64_u8( + vrhadd_u8(vreinterpret_u8_s64(x), vreinterpret_u8_s64(y))); +} + +SIMD_INLINE v64 v64_rdavg_u8(v64 x, v64 y) { + return vreinterpret_s64_u8( + vhadd_u8(vreinterpret_u8_s64(x), vreinterpret_u8_s64(y))); +} + +SIMD_INLINE v64 v64_rdavg_u16(v64 x, v64 y) { + return vreinterpret_s64_u16( + vhadd_u16(vreinterpret_u16_s64(x), vreinterpret_u16_s64(y))); +} + +SIMD_INLINE v64 v64_avg_u16(v64 x, v64 y) { + return vreinterpret_s64_u16( + vrhadd_u16(vreinterpret_u16_s64(x), vreinterpret_u16_s64(y))); +} + +SIMD_INLINE v64 v64_max_u8(v64 x, v64 y) { + return vreinterpret_s64_u8( + vmax_u8(vreinterpret_u8_s64(x), vreinterpret_u8_s64(y))); +} + +SIMD_INLINE v64 v64_min_u8(v64 x, v64 y) { + return vreinterpret_s64_u8( + vmin_u8(vreinterpret_u8_s64(x), vreinterpret_u8_s64(y))); +} + +SIMD_INLINE v64 v64_max_s8(v64 x, v64 y) { + return vreinterpret_s64_s8( + vmax_s8(vreinterpret_s8_s64(x), vreinterpret_s8_s64(y))); +} + +SIMD_INLINE v64 v64_min_s8(v64 x, v64 y) { + return vreinterpret_s64_s8( + vmin_s8(vreinterpret_s8_s64(x), vreinterpret_s8_s64(y))); +} + +SIMD_INLINE v64 v64_max_s16(v64 x, v64 y) { + return vreinterpret_s64_s16( + vmax_s16(vreinterpret_s16_s64(x), vreinterpret_s16_s64(y))); +} + +SIMD_INLINE v64 v64_min_s16(v64 x, v64 y) { + return vreinterpret_s64_s16( + vmin_s16(vreinterpret_s16_s64(x), vreinterpret_s16_s64(y))); +} + +SIMD_INLINE v64 v64_ziplo_8(v64 x, v64 y) { +#if defined(__aarch64__) + return vreinterpret_s64_u8( + vzip1_u8(vreinterpret_u8_s64(y), vreinterpret_u8_s64(x))); +#else + uint8x8x2_t r = vzip_u8(vreinterpret_u8_s64(y), vreinterpret_u8_s64(x)); + return vreinterpret_s64_u8(r.val[0]); +#endif +} + +SIMD_INLINE v64 v64_ziphi_8(v64 x, v64 y) { +#if defined(__aarch64__) + return vreinterpret_s64_u8( + vzip2_u8(vreinterpret_u8_s64(y), vreinterpret_u8_s64(x))); +#else + uint8x8x2_t r = vzip_u8(vreinterpret_u8_s64(y), vreinterpret_u8_s64(x)); + return vreinterpret_s64_u8(r.val[1]); +#endif +} + +SIMD_INLINE v64 v64_ziplo_16(v64 x, v64 y) { +#if defined(__aarch64__) + return vreinterpret_s64_u16( + vzip1_u16(vreinterpret_u16_s64(y), vreinterpret_u16_s64(x))); +#else + int16x4x2_t r = vzip_s16(vreinterpret_s16_s64(y), vreinterpret_s16_s64(x)); + return vreinterpret_s64_s16(r.val[0]); +#endif +} + +SIMD_INLINE v64 v64_ziphi_16(v64 x, v64 y) { +#if defined(__aarch64__) + return vreinterpret_s64_u16( + vzip2_u16(vreinterpret_u16_s64(y), vreinterpret_u16_s64(x))); +#else + int16x4x2_t r = vzip_s16(vreinterpret_s16_s64(y), vreinterpret_s16_s64(x)); + return vreinterpret_s64_s16(r.val[1]); +#endif +} + +SIMD_INLINE v64 v64_ziplo_32(v64 x, v64 y) { +#if defined(__aarch64__) + return vreinterpret_s64_u32( + vzip1_u32(vreinterpret_u32_s64(y), vreinterpret_u32_s64(x))); +#else + int32x2x2_t r = vzip_s32(vreinterpret_s32_s64(y), vreinterpret_s32_s64(x)); + return vreinterpret_s64_s32(r.val[0]); +#endif +} + +SIMD_INLINE v64 v64_ziphi_32(v64 x, v64 y) { +#if defined(__aarch64__) + return vreinterpret_s64_u32( + vzip2_u32(vreinterpret_u32_s64(y), vreinterpret_u32_s64(x))); +#else + int32x2x2_t r = vzip_s32(vreinterpret_s32_s64(y), vreinterpret_s32_s64(x)); + return vreinterpret_s64_s32(r.val[1]); +#endif +} + +SIMD_INLINE v64 v64_unpacklo_u8_s16(v64 a) { + return vreinterpret_s64_u16(vget_low_u16(vmovl_u8(vreinterpret_u8_s64(a)))); +} + +SIMD_INLINE v64 v64_unpackhi_u8_s16(v64 a) { + return vreinterpret_s64_u16(vget_high_u16(vmovl_u8(vreinterpret_u8_s64(a)))); +} + +SIMD_INLINE v64 v64_unpacklo_s8_s16(v64 a) { + return vreinterpret_s64_s16(vget_low_s16(vmovl_s8(vreinterpret_s8_s64(a)))); +} + +SIMD_INLINE v64 v64_unpackhi_s8_s16(v64 a) { + return vreinterpret_s64_s16(vget_high_s16(vmovl_s8(vreinterpret_s8_s64(a)))); +} + +SIMD_INLINE v64 v64_pack_s32_s16(v64 x, v64 y) { + return vreinterpret_s64_s16(vqmovn_s32( + vcombine_s32(vreinterpret_s32_s64(y), vreinterpret_s32_s64(x)))); +} + +SIMD_INLINE v64 v64_pack_s32_u16(v64 x, v64 y) { + return vreinterpret_s64_u16(vqmovun_s32( + vcombine_s32(vreinterpret_s32_s64(y), vreinterpret_s32_s64(x)))); +} + +SIMD_INLINE v64 v64_pack_s16_u8(v64 x, v64 y) { + return vreinterpret_s64_u8(vqmovun_s16(vreinterpretq_s16_s32( + vcombine_s32(vreinterpret_s32_s64(y), vreinterpret_s32_s64(x))))); +} + +SIMD_INLINE v64 v64_pack_s16_s8(v64 x, v64 y) { + return vreinterpret_s64_s8(vqmovn_s16(vreinterpretq_s16_s32( + vcombine_s32(vreinterpret_s32_s64(y), vreinterpret_s32_s64(x))))); +} + +SIMD_INLINE v64 v64_unziplo_8(v64 x, v64 y) { +#if defined(__aarch64__) + return vreinterpret_s64_u8( + vuzp1_u8(vreinterpret_u8_s64(y), vreinterpret_u8_s64(x))); +#else + uint8x8x2_t r = vuzp_u8(vreinterpret_u8_s64(y), vreinterpret_u8_s64(x)); + return vreinterpret_s64_u8(r.val[0]); +#endif +} + +SIMD_INLINE v64 v64_unziphi_8(v64 x, v64 y) { +#if defined(__aarch64__) + return vreinterpret_s64_u8( + vuzp2_u8(vreinterpret_u8_s64(y), vreinterpret_u8_s64(x))); +#else + uint8x8x2_t r = vuzp_u8(vreinterpret_u8_s64(y), vreinterpret_u8_s64(x)); + return vreinterpret_s64_u8(r.val[1]); +#endif +} + +SIMD_INLINE v64 v64_unziplo_16(v64 x, v64 y) { +#if defined(__aarch64__) + return vreinterpret_s64_u16( + vuzp1_u16(vreinterpret_u16_s64(y), vreinterpret_u16_s64(x))); +#else + uint16x4x2_t r = vuzp_u16(vreinterpret_u16_s64(y), vreinterpret_u16_s64(x)); + return vreinterpret_s64_u16(r.val[0]); +#endif +} + +SIMD_INLINE v64 v64_unziphi_16(v64 x, v64 y) { +#if defined(__aarch64__) + return vreinterpret_s64_u16( + vuzp2_u16(vreinterpret_u16_s64(y), vreinterpret_u16_s64(x))); +#else + uint16x4x2_t r = vuzp_u16(vreinterpret_u16_s64(y), vreinterpret_u16_s64(x)); + return vreinterpret_s64_u16(r.val[1]); +#endif +} + +SIMD_INLINE v64 v64_unpacklo_s16_s32(v64 x) { + return vreinterpret_s64_s32(vget_low_s32(vmovl_s16(vreinterpret_s16_s64(x)))); +} + +SIMD_INLINE v64 v64_unpacklo_u16_s32(v64 x) { + return vreinterpret_s64_u32(vget_low_u32(vmovl_u16(vreinterpret_u16_s64(x)))); +} + +SIMD_INLINE v64 v64_unpackhi_s16_s32(v64 x) { + return vreinterpret_s64_s32( + vget_high_s32(vmovl_s16(vreinterpret_s16_s64(x)))); +} + +SIMD_INLINE v64 v64_unpackhi_u16_s32(v64 x) { + return vreinterpret_s64_u32( + vget_high_u32(vmovl_u16(vreinterpret_u16_s64(x)))); +} + +SIMD_INLINE v64 v64_shuffle_8(v64 x, v64 pattern) { + return vreinterpret_s64_u8( + vtbl1_u8(vreinterpret_u8_s64(x), vreinterpret_u8_s64(pattern))); +} + +SIMD_INLINE v64 v64_cmpgt_s8(v64 x, v64 y) { + return vreinterpret_s64_u8( + vcgt_s8(vreinterpret_s8_s64(x), vreinterpret_s8_s64(y))); +} + +SIMD_INLINE v64 v64_cmplt_s8(v64 x, v64 y) { + return vreinterpret_s64_u8( + vclt_s8(vreinterpret_s8_s64(x), vreinterpret_s8_s64(y))); +} + +SIMD_INLINE v64 v64_cmpeq_8(v64 x, v64 y) { + return vreinterpret_s64_u8( + vceq_u8(vreinterpret_u8_s64(x), vreinterpret_u8_s64(y))); +} + +SIMD_INLINE v64 v64_cmpgt_s16(v64 x, v64 y) { + return vreinterpret_s64_u16( + vcgt_s16(vreinterpret_s16_s64(x), vreinterpret_s16_s64(y))); +} + +SIMD_INLINE v64 v64_cmplt_s16(v64 x, v64 y) { + return vreinterpret_s64_u16( + vclt_s16(vreinterpret_s16_s64(x), vreinterpret_s16_s64(y))); +} + +SIMD_INLINE v64 v64_cmpeq_16(v64 x, v64 y) { + return vreinterpret_s64_u16( + vceq_s16(vreinterpret_s16_s64(x), vreinterpret_s16_s64(y))); +} + +SIMD_INLINE v64 v64_shl_8(v64 a, unsigned int c) { + return vreinterpret_s64_u8(vshl_u8(vreinterpret_u8_s64(a), vdup_n_s8(c))); +} + +SIMD_INLINE v64 v64_shr_u8(v64 a, unsigned int c) { + return vreinterpret_s64_u8(vshl_u8(vreinterpret_u8_s64(a), vdup_n_s8(-c))); +} + +SIMD_INLINE v64 v64_shr_s8(v64 a, unsigned int c) { + return vreinterpret_s64_s8(vshl_s8(vreinterpret_s8_s64(a), vdup_n_s8(-c))); +} + +SIMD_INLINE v64 v64_shl_16(v64 a, unsigned int c) { + return vreinterpret_s64_u16(vshl_u16(vreinterpret_u16_s64(a), vdup_n_s16(c))); +} + +SIMD_INLINE v64 v64_shr_u16(v64 a, unsigned int c) { + return vreinterpret_s64_u16( + vshl_u16(vreinterpret_u16_s64(a), vdup_n_s16(-(int)c))); +} + +SIMD_INLINE v64 v64_shr_s16(v64 a, unsigned int c) { + return vreinterpret_s64_s16( + vshl_s16(vreinterpret_s16_s64(a), vdup_n_s16(-(int)c))); +} + +SIMD_INLINE v64 v64_shl_32(v64 a, unsigned int c) { + return vreinterpret_s64_u32(vshl_u32(vreinterpret_u32_s64(a), vdup_n_s32(c))); +} + +SIMD_INLINE v64 v64_shr_u32(v64 a, unsigned int c) { + return vreinterpret_s64_u32( + vshl_u32(vreinterpret_u32_s64(a), vdup_n_s32(-(int)c))); +} + +SIMD_INLINE v64 v64_shr_s32(v64 a, unsigned int c) { + return vreinterpret_s64_s32( + vshl_s32(vreinterpret_s32_s64(a), vdup_n_s32(-(int)c))); +} + +// The following functions require an immediate. +// Some compilers will check this during optimisation, others wont. +#if defined(__OPTIMIZE__) && __OPTIMIZE__ && !defined(__clang__) + +SIMD_INLINE v64 v64_shl_n_byte(v64 a, unsigned int c) { + return vshl_n_s64(a, c * 8); +} + +SIMD_INLINE v64 v64_shr_n_byte(v64 a, unsigned int c) { + return c ? (v64)vshr_n_u64(vreinterpret_u64_s64(a), c * 8) : a; +} + +SIMD_INLINE v64 v64_shl_n_8(v64 a, unsigned int c) { + return vreinterpret_s64_u8(vshl_n_u8(vreinterpret_u8_s64(a), c)); +} + +SIMD_INLINE v64 v64_shr_n_u8(v64 a, unsigned int c) { + return vreinterpret_s64_u8(vshr_n_u8(vreinterpret_u8_s64(a), c)); +} + +SIMD_INLINE v64 v64_shr_n_s8(v64 a, unsigned int c) { + return vreinterpret_s64_s8(vshr_n_s8(vreinterpret_s8_s64(a), c)); +} + +SIMD_INLINE v64 v64_shl_n_16(v64 a, unsigned int c) { + return vreinterpret_s64_u16(vshl_n_u16(vreinterpret_u16_s64(a), c)); +} + +SIMD_INLINE v64 v64_shr_n_u16(v64 a, unsigned int c) { + return vreinterpret_s64_u16(vshr_n_u16(vreinterpret_u16_s64(a), c)); +} + +SIMD_INLINE v64 v64_shr_n_s16(v64 a, unsigned int c) { + return vreinterpret_s64_s16(vshr_n_s16(vreinterpret_s16_s64(a), c)); +} + +SIMD_INLINE v64 v64_shl_n_32(v64 a, unsigned int c) { + return vreinterpret_s64_u32(vshl_n_u32(vreinterpret_u32_s64(a), c)); +} + +SIMD_INLINE v64 v64_shr_n_u32(v64 a, unsigned int c) { + return vreinterpret_s64_u32(vshr_n_u32(vreinterpret_u32_s64(a), c)); +} + +SIMD_INLINE v64 v64_shr_n_s32(v64 a, unsigned int c) { + return vreinterpret_s64_s32(vshr_n_s32(vreinterpret_s32_s64(a), c)); +} + +#else + +SIMD_INLINE v64 v64_shl_n_byte(v64 a, unsigned int c) { + return v64_from_64(v64_u64(a) << c * 8); +} + +SIMD_INLINE v64 v64_shr_n_byte(v64 a, unsigned int c) { + return v64_from_64(v64_u64(a) >> c * 8); +} + +SIMD_INLINE v64 v64_shl_n_8(v64 a, unsigned int c) { return v64_shl_8(a, c); } + +SIMD_INLINE v64 v64_shr_n_u8(v64 a, unsigned int c) { return v64_shr_u8(a, c); } + +SIMD_INLINE v64 v64_shr_n_s8(v64 a, unsigned int c) { return v64_shr_s8(a, c); } + +SIMD_INLINE v64 v64_shl_n_16(v64 a, unsigned int c) { return v64_shl_16(a, c); } + +SIMD_INLINE v64 v64_shr_n_u16(v64 a, unsigned int c) { + return v64_shr_u16(a, c); +} + +SIMD_INLINE v64 v64_shr_n_s16(v64 a, unsigned int c) { + return v64_shr_s16(a, c); +} + +SIMD_INLINE v64 v64_shl_n_32(v64 a, unsigned int c) { return v64_shl_32(a, c); } + +SIMD_INLINE v64 v64_shr_n_u32(v64 a, unsigned int c) { + return v64_shr_u32(a, c); +} + +SIMD_INLINE v64 v64_shr_n_s32(v64 a, unsigned int c) { + return v64_shr_s32(a, c); +} + +#endif + +#endif // AOM_AOM_DSP_SIMD_V64_INTRINSICS_ARM_H_ diff --git a/third_party/aom/aom_dsp/simd/v64_intrinsics_c.h b/third_party/aom/aom_dsp/simd/v64_intrinsics_c.h new file mode 100644 index 0000000000..028d68c4f7 --- /dev/null +++ b/third_party/aom/aom_dsp/simd/v64_intrinsics_c.h @@ -0,0 +1,968 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_SIMD_V64_INTRINSICS_C_H_ +#define AOM_AOM_DSP_SIMD_V64_INTRINSICS_C_H_ + +/* Note: This implements the intrinsics in plain, unoptimised C. + Intended for reference, porting or debugging. */ + +#include +#include + +#include "config/aom_config.h" + +typedef union { + uint8_t u8[8]; + uint16_t u16[4]; + uint32_t u32[2]; + uint64_t u64; + int8_t s8[8]; + int16_t s16[4]; + int32_t s32[2]; + int64_t s64; +} c_v64; + +SIMD_INLINE uint32_t c_v64_low_u32(c_v64 a) { + return a.u32[!!CONFIG_BIG_ENDIAN]; +} + +SIMD_INLINE uint32_t c_v64_high_u32(c_v64 a) { + return a.u32[!CONFIG_BIG_ENDIAN]; +} + +SIMD_INLINE int32_t c_v64_low_s32(c_v64 a) { + return a.s32[!!CONFIG_BIG_ENDIAN]; +} + +SIMD_INLINE int32_t c_v64_high_s32(c_v64 a) { + return a.s32[!CONFIG_BIG_ENDIAN]; +} + +SIMD_INLINE c_v64 c_v64_from_32(uint32_t x, uint32_t y) { + c_v64 t; + t.u32[!CONFIG_BIG_ENDIAN] = x; + t.u32[!!CONFIG_BIG_ENDIAN] = y; + return t; +} + +SIMD_INLINE c_v64 c_v64_from_64(uint64_t x) { + c_v64 t; + t.u64 = x; + return t; +} + +SIMD_INLINE uint64_t c_v64_u64(c_v64 x) { return x.u64; } + +SIMD_INLINE c_v64 c_v64_from_16(uint16_t a, uint16_t b, uint16_t c, + uint16_t d) { + c_v64 t; + if (CONFIG_BIG_ENDIAN) { + t.u16[0] = a; + t.u16[1] = b; + t.u16[2] = c; + t.u16[3] = d; + } else { + t.u16[3] = a; + t.u16[2] = b; + t.u16[1] = c; + t.u16[0] = d; + } + return t; +} + +SIMD_INLINE uint32_t c_u32_load_unaligned(const void *p) { + uint32_t t; + uint8_t *pp = (uint8_t *)p; + uint8_t *q = (uint8_t *)&t; + int c; + for (c = 0; c < 4; c++) q[c] = pp[c]; + return t; +} + +SIMD_INLINE void c_u32_store_unaligned(void *p, uint32_t a) { + uint8_t *pp = (uint8_t *)p; + uint8_t *q = (uint8_t *)&a; + int c; + for (c = 0; c < 4; c++) pp[c] = q[c]; +} + +SIMD_INLINE uint32_t c_u32_load_aligned(const void *p) { + if (SIMD_CHECK && (uintptr_t)p & 3) { + fprintf(stderr, "Error: Unaligned u32 load at %p\n", p); + abort(); + } + return c_u32_load_unaligned(p); +} + +SIMD_INLINE void c_u32_store_aligned(void *p, uint32_t a) { + if (SIMD_CHECK && (uintptr_t)p & 3) { + fprintf(stderr, "Error: Unaligned u32 store at %p\n", p); + abort(); + } + c_u32_store_unaligned(p, a); +} + +SIMD_INLINE c_v64 c_v64_load_unaligned(const void *p) { + c_v64 t; + uint8_t *pp = (uint8_t *)p; + uint8_t *q = (uint8_t *)&t; + int c; + for (c = 0; c < 8; c++) q[c] = pp[c]; + return t; +} + +SIMD_INLINE c_v64 c_v64_load_aligned(const void *p) { + if (SIMD_CHECK && (uintptr_t)p & 7) { + fprintf(stderr, "Error: Unaligned c_v64 load at %p\n", p); + abort(); + } + return c_v64_load_unaligned(p); +} + +SIMD_INLINE void c_v64_store_unaligned(void *p, c_v64 a) { + uint8_t *q = (uint8_t *)p; + uint8_t *r = (uint8_t *)&a; + int c; + for (c = 0; c < 8; c++) q[c] = r[c]; +} + +SIMD_INLINE void c_v64_store_aligned(void *p, c_v64 a) { + if (SIMD_CHECK && (uintptr_t)p & 7) { + fprintf(stderr, "Error: Unaligned c_v64 store at %p\n", p); + abort(); + } + c_v64_store_unaligned(p, a); +} + +SIMD_INLINE c_v64 c_v64_zero() { + c_v64 t; + t.u64 = 0; + return t; +} + +SIMD_INLINE c_v64 c_v64_dup_8(uint8_t x) { + c_v64 t; + t.u8[0] = t.u8[1] = t.u8[2] = t.u8[3] = t.u8[4] = t.u8[5] = t.u8[6] = + t.u8[7] = x; + return t; +} + +SIMD_INLINE c_v64 c_v64_dup_16(uint16_t x) { + c_v64 t; + t.u16[0] = t.u16[1] = t.u16[2] = t.u16[3] = x; + return t; +} + +SIMD_INLINE c_v64 c_v64_dup_32(uint32_t x) { + c_v64 t; + t.u32[0] = t.u32[1] = x; + return t; +} + +SIMD_INLINE c_v64 c_v64_add_8(c_v64 a, c_v64 b) { + c_v64 t; + int c; + for (c = 0; c < 8; c++) t.u8[c] = a.u8[c] + b.u8[c]; + return t; +} + +SIMD_INLINE c_v64 c_v64_add_16(c_v64 a, c_v64 b) { + c_v64 t; + int c; + for (c = 0; c < 4; c++) t.u16[c] = a.u16[c] + b.u16[c]; + return t; +} + +SIMD_INLINE c_v64 c_v64_sadd_u8(c_v64 a, c_v64 b) { + c_v64 t; + int c; + for (c = 0; c < 8; c++) + t.u8[c] = (int16_t)a.u8[c] + (int16_t)b.u8[c] > 255 + ? 255 + : (int16_t)a.u8[c] + (int16_t)b.u8[c] < 0 + ? 0 + : (int16_t)a.u8[c] + (int16_t)b.u8[c]; + return t; +} + +SIMD_INLINE c_v64 c_v64_sadd_s8(c_v64 a, c_v64 b) { + c_v64 t; + int c; + for (c = 0; c < 8; c++) + t.s8[c] = (int16_t)a.s8[c] + (int16_t)b.s8[c] > 127 + ? 127 + : (int16_t)a.s8[c] + (int16_t)b.s8[c] < -128 + ? -128 + : (int16_t)a.s8[c] + (int16_t)b.s8[c]; + return t; +} + +SIMD_INLINE c_v64 c_v64_sadd_s16(c_v64 a, c_v64 b) { + c_v64 t; + int c; + for (c = 0; c < 4; c++) + t.s16[c] = (int32_t)a.s16[c] + (int32_t)b.s16[c] > 32767 + ? 32767 + : (int32_t)a.s16[c] + (int32_t)b.s16[c] < -32768 + ? -32768 + : (int32_t)a.s16[c] + (int32_t)b.s16[c]; + return t; +} + +SIMD_INLINE c_v64 c_v64_add_32(c_v64 a, c_v64 b) { + c_v64 t; + t.u32[0] = (uint32_t)((uint64_t)a.u32[0] + b.u32[0]); + t.u32[1] = (uint32_t)((uint64_t)a.u32[1] + b.u32[1]); + return t; +} + +SIMD_INLINE c_v64 c_v64_sub_8(c_v64 a, c_v64 b) { + c_v64 t; + int c; + for (c = 0; c < 8; c++) t.u8[c] = a.u8[c] - b.u8[c]; + return t; +} + +SIMD_INLINE c_v64 c_v64_ssub_u8(c_v64 a, c_v64 b) { + c_v64 t; + int c; + for (c = 0; c < 8; c++) t.u8[c] = a.u8[c] < b.u8[c] ? 0 : a.u8[c] - b.u8[c]; + return t; +} + +SIMD_INLINE c_v64 c_v64_ssub_s8(c_v64 a, c_v64 b) { + c_v64 t; + int c; + for (c = 0; c < 8; c++) { + int16_t d = (int16_t)a.s8[c] - (int16_t)b.s8[c]; + t.s8[c] = d > 127 ? 127 : (d < -128 ? -128 : d); + } + return t; +} + +SIMD_INLINE c_v64 c_v64_sub_16(c_v64 a, c_v64 b) { + c_v64 t; + int c; + for (c = 0; c < 4; c++) t.u16[c] = a.u16[c] - b.u16[c]; + return t; +} + +SIMD_INLINE c_v64 c_v64_ssub_s16(c_v64 a, c_v64 b) { + c_v64 t; + int c; + for (c = 0; c < 4; c++) + t.s16[c] = (int32_t)a.s16[c] - (int32_t)b.s16[c] < -32768 + ? -32768 + : (int32_t)a.s16[c] - (int32_t)b.s16[c] > 32767 + ? 32767 + : (int32_t)a.s16[c] - (int32_t)b.s16[c]; + return t; +} + +SIMD_INLINE c_v64 c_v64_ssub_u16(c_v64 a, c_v64 b) { + c_v64 t; + int c; + for (c = 0; c < 4; c++) + t.u16[c] = + (int32_t)a.u16[c] - (int32_t)b.u16[c] < 0 ? 0 : a.u16[c] - b.u16[c]; + return t; +} + +SIMD_INLINE c_v64 c_v64_sub_32(c_v64 a, c_v64 b) { + c_v64 t; + t.u32[0] = (uint32_t)((int64_t)a.u32[0] - b.u32[0]); + t.u32[1] = (uint32_t)((int64_t)a.u32[1] - b.u32[1]); + return t; +} + +SIMD_INLINE c_v64 c_v64_abs_s16(c_v64 a) { + c_v64 t; + int c; + for (c = 0; c < 4; c++) + t.u16[c] = (int16_t)a.u16[c] > 0 ? a.u16[c] : -a.u16[c]; + return t; +} + +SIMD_INLINE c_v64 c_v64_abs_s8(c_v64 a) { + c_v64 t; + int c; + for (c = 0; c < 8; c++) t.u8[c] = (int8_t)a.u8[c] > 0 ? a.u8[c] : -a.u8[c]; + return t; +} + +SIMD_INLINE c_v64 _c_v64_zip_8(c_v64 a, c_v64 b, int mode) { + c_v64 t; + if (mode) { + t.u8[7] = a.u8[7]; + t.u8[6] = b.u8[7]; + t.u8[5] = a.u8[6]; + t.u8[4] = b.u8[6]; + t.u8[3] = a.u8[5]; + t.u8[2] = b.u8[5]; + t.u8[1] = a.u8[4]; + t.u8[0] = b.u8[4]; + } else { + t.u8[7] = a.u8[3]; + t.u8[6] = b.u8[3]; + t.u8[5] = a.u8[2]; + t.u8[4] = b.u8[2]; + t.u8[3] = a.u8[1]; + t.u8[2] = b.u8[1]; + t.u8[1] = a.u8[0]; + t.u8[0] = b.u8[0]; + } + return t; +} + +SIMD_INLINE c_v64 c_v64_ziplo_8(c_v64 a, c_v64 b) { + return CONFIG_BIG_ENDIAN ? _c_v64_zip_8(b, a, 1) : _c_v64_zip_8(a, b, 0); +} + +SIMD_INLINE c_v64 c_v64_ziphi_8(c_v64 a, c_v64 b) { + return CONFIG_BIG_ENDIAN ? _c_v64_zip_8(b, a, 0) : _c_v64_zip_8(a, b, 1); +} + +SIMD_INLINE c_v64 _c_v64_zip_16(c_v64 a, c_v64 b, int mode) { + c_v64 t; + if (mode) { + t.u16[3] = a.u16[3]; + t.u16[2] = b.u16[3]; + t.u16[1] = a.u16[2]; + t.u16[0] = b.u16[2]; + } else { + t.u16[3] = a.u16[1]; + t.u16[2] = b.u16[1]; + t.u16[1] = a.u16[0]; + t.u16[0] = b.u16[0]; + } + return t; +} + +SIMD_INLINE c_v64 c_v64_ziplo_16(c_v64 a, c_v64 b) { + return CONFIG_BIG_ENDIAN ? _c_v64_zip_16(b, a, 1) : _c_v64_zip_16(a, b, 0); +} + +SIMD_INLINE c_v64 c_v64_ziphi_16(c_v64 a, c_v64 b) { + return CONFIG_BIG_ENDIAN ? _c_v64_zip_16(b, a, 0) : _c_v64_zip_16(a, b, 1); +} + +SIMD_INLINE c_v64 _c_v64_zip_32(c_v64 a, c_v64 b, int mode) { + c_v64 t; + if (mode) { + t.u32[1] = a.u32[1]; + t.u32[0] = b.u32[1]; + } else { + t.u32[1] = a.u32[0]; + t.u32[0] = b.u32[0]; + } + return t; +} + +SIMD_INLINE c_v64 c_v64_ziplo_32(c_v64 a, c_v64 b) { + return CONFIG_BIG_ENDIAN ? _c_v64_zip_32(b, a, 1) : _c_v64_zip_32(a, b, 0); +} + +SIMD_INLINE c_v64 c_v64_ziphi_32(c_v64 a, c_v64 b) { + return CONFIG_BIG_ENDIAN ? _c_v64_zip_32(b, a, 0) : _c_v64_zip_32(a, b, 1); +} + +SIMD_INLINE c_v64 _c_v64_unzip_8(c_v64 a, c_v64 b, int mode) { + c_v64 t; + if (mode) { + t.u8[7] = b.u8[7]; + t.u8[6] = b.u8[5]; + t.u8[5] = b.u8[3]; + t.u8[4] = b.u8[1]; + t.u8[3] = a.u8[7]; + t.u8[2] = a.u8[5]; + t.u8[1] = a.u8[3]; + t.u8[0] = a.u8[1]; + } else { + t.u8[7] = a.u8[6]; + t.u8[6] = a.u8[4]; + t.u8[5] = a.u8[2]; + t.u8[4] = a.u8[0]; + t.u8[3] = b.u8[6]; + t.u8[2] = b.u8[4]; + t.u8[1] = b.u8[2]; + t.u8[0] = b.u8[0]; + } + return t; +} + +SIMD_INLINE c_v64 c_v64_unziplo_8(c_v64 a, c_v64 b) { + return CONFIG_BIG_ENDIAN ? _c_v64_unzip_8(a, b, 1) : _c_v64_unzip_8(a, b, 0); +} + +SIMD_INLINE c_v64 c_v64_unziphi_8(c_v64 a, c_v64 b) { + return CONFIG_BIG_ENDIAN ? _c_v64_unzip_8(b, a, 0) : _c_v64_unzip_8(b, a, 1); +} + +SIMD_INLINE c_v64 _c_v64_unzip_16(c_v64 a, c_v64 b, int mode) { + c_v64 t; + if (mode) { + t.u16[3] = b.u16[3]; + t.u16[2] = b.u16[1]; + t.u16[1] = a.u16[3]; + t.u16[0] = a.u16[1]; + } else { + t.u16[3] = a.u16[2]; + t.u16[2] = a.u16[0]; + t.u16[1] = b.u16[2]; + t.u16[0] = b.u16[0]; + } + return t; +} + +SIMD_INLINE c_v64 c_v64_unziplo_16(c_v64 a, c_v64 b) { + return CONFIG_BIG_ENDIAN ? _c_v64_unzip_16(a, b, 1) + : _c_v64_unzip_16(a, b, 0); +} + +SIMD_INLINE c_v64 c_v64_unziphi_16(c_v64 a, c_v64 b) { + return CONFIG_BIG_ENDIAN ? _c_v64_unzip_16(b, a, 0) + : _c_v64_unzip_16(b, a, 1); +} + +SIMD_INLINE c_v64 c_v64_unpacklo_u8_s16(c_v64 a) { + c_v64 t; + int endian = !!CONFIG_BIG_ENDIAN * 4; + t.s16[3] = (int16_t)a.u8[3 + endian]; + t.s16[2] = (int16_t)a.u8[2 + endian]; + t.s16[1] = (int16_t)a.u8[1 + endian]; + t.s16[0] = (int16_t)a.u8[0 + endian]; + return t; +} + +SIMD_INLINE c_v64 c_v64_unpackhi_u8_s16(c_v64 a) { + c_v64 t; + int endian = !!CONFIG_BIG_ENDIAN * 4; + t.s16[3] = (int16_t)a.u8[7 - endian]; + t.s16[2] = (int16_t)a.u8[6 - endian]; + t.s16[1] = (int16_t)a.u8[5 - endian]; + t.s16[0] = (int16_t)a.u8[4 - endian]; + return t; +} + +SIMD_INLINE c_v64 c_v64_unpacklo_s8_s16(c_v64 a) { + c_v64 t; + int endian = !!CONFIG_BIG_ENDIAN * 4; + t.s16[3] = (int16_t)a.s8[3 + endian]; + t.s16[2] = (int16_t)a.s8[2 + endian]; + t.s16[1] = (int16_t)a.s8[1 + endian]; + t.s16[0] = (int16_t)a.s8[0 + endian]; + return t; +} + +SIMD_INLINE c_v64 c_v64_unpackhi_s8_s16(c_v64 a) { + c_v64 t; + int endian = !!CONFIG_BIG_ENDIAN * 4; + t.s16[3] = (int16_t)a.s8[7 - endian]; + t.s16[2] = (int16_t)a.s8[6 - endian]; + t.s16[1] = (int16_t)a.s8[5 - endian]; + t.s16[0] = (int16_t)a.s8[4 - endian]; + return t; +} + +SIMD_INLINE c_v64 c_v64_pack_s32_s16(c_v64 a, c_v64 b) { + c_v64 t; + if (CONFIG_BIG_ENDIAN) { + c_v64 u = a; + a = b; + b = u; + } + t.s16[3] = a.s32[1] > 32767 ? 32767 : a.s32[1] < -32768 ? -32768 : a.s32[1]; + t.s16[2] = a.s32[0] > 32767 ? 32767 : a.s32[0] < -32768 ? -32768 : a.s32[0]; + t.s16[1] = b.s32[1] > 32767 ? 32767 : b.s32[1] < -32768 ? -32768 : b.s32[1]; + t.s16[0] = b.s32[0] > 32767 ? 32767 : b.s32[0] < -32768 ? -32768 : b.s32[0]; + return t; +} + +SIMD_INLINE c_v64 c_v64_pack_s32_u16(c_v64 a, c_v64 b) { + c_v64 t; + if (CONFIG_BIG_ENDIAN) { + c_v64 u = a; + a = b; + b = u; + } + t.u16[3] = a.s32[1] > 65535 ? 65535 : a.s32[1] < 0 ? 0 : a.s32[1]; + t.u16[2] = a.s32[0] > 65535 ? 65535 : a.s32[0] < 0 ? 0 : a.s32[0]; + t.u16[1] = b.s32[1] > 65535 ? 65535 : b.s32[1] < 0 ? 0 : b.s32[1]; + t.u16[0] = b.s32[0] > 65535 ? 65535 : b.s32[0] < 0 ? 0 : b.s32[0]; + return t; +} + +SIMD_INLINE c_v64 c_v64_pack_s16_u8(c_v64 a, c_v64 b) { + c_v64 t; + if (CONFIG_BIG_ENDIAN) { + c_v64 u = a; + a = b; + b = u; + } + t.u8[7] = a.s16[3] > 255 ? 255 : a.s16[3] < 0 ? 0 : a.s16[3]; + t.u8[6] = a.s16[2] > 255 ? 255 : a.s16[2] < 0 ? 0 : a.s16[2]; + t.u8[5] = a.s16[1] > 255 ? 255 : a.s16[1] < 0 ? 0 : a.s16[1]; + t.u8[4] = a.s16[0] > 255 ? 255 : a.s16[0] < 0 ? 0 : a.s16[0]; + t.u8[3] = b.s16[3] > 255 ? 255 : b.s16[3] < 0 ? 0 : b.s16[3]; + t.u8[2] = b.s16[2] > 255 ? 255 : b.s16[2] < 0 ? 0 : b.s16[2]; + t.u8[1] = b.s16[1] > 255 ? 255 : b.s16[1] < 0 ? 0 : b.s16[1]; + t.u8[0] = b.s16[0] > 255 ? 255 : b.s16[0] < 0 ? 0 : b.s16[0]; + return t; +} + +SIMD_INLINE c_v64 c_v64_pack_s16_s8(c_v64 a, c_v64 b) { + c_v64 t; + if (CONFIG_BIG_ENDIAN) { + c_v64 u = a; + a = b; + b = u; + } + t.u8[7] = a.s16[3] > 127 ? 127 : a.s16[3] < -128 ? 128 : a.s16[3]; + t.u8[6] = a.s16[2] > 127 ? 127 : a.s16[2] < -128 ? 128 : a.s16[2]; + t.u8[5] = a.s16[1] > 127 ? 127 : a.s16[1] < -128 ? 128 : a.s16[1]; + t.u8[4] = a.s16[0] > 127 ? 127 : a.s16[0] < -128 ? 128 : a.s16[0]; + t.u8[3] = b.s16[3] > 127 ? 127 : b.s16[3] < -128 ? 128 : b.s16[3]; + t.u8[2] = b.s16[2] > 127 ? 127 : b.s16[2] < -128 ? 128 : b.s16[2]; + t.u8[1] = b.s16[1] > 127 ? 127 : b.s16[1] < -128 ? 128 : b.s16[1]; + t.u8[0] = b.s16[0] > 127 ? 127 : b.s16[0] < -128 ? 128 : b.s16[0]; + return t; +} + +SIMD_INLINE c_v64 c_v64_unpacklo_u16_s32(c_v64 a) { + c_v64 t; + t.s32[1] = a.u16[1 + !!CONFIG_BIG_ENDIAN * 2]; + t.s32[0] = a.u16[0 + !!CONFIG_BIG_ENDIAN * 2]; + return t; +} + +SIMD_INLINE c_v64 c_v64_unpacklo_s16_s32(c_v64 a) { + c_v64 t; + t.s32[1] = a.s16[1 + !!CONFIG_BIG_ENDIAN * 2]; + t.s32[0] = a.s16[0 + !!CONFIG_BIG_ENDIAN * 2]; + return t; +} + +SIMD_INLINE c_v64 c_v64_unpackhi_u16_s32(c_v64 a) { + c_v64 t; + t.s32[1] = a.u16[3 - !!CONFIG_BIG_ENDIAN * 2]; + t.s32[0] = a.u16[2 - !!CONFIG_BIG_ENDIAN * 2]; + return t; +} + +SIMD_INLINE c_v64 c_v64_unpackhi_s16_s32(c_v64 a) { + c_v64 t; + t.s32[1] = a.s16[3 - !!CONFIG_BIG_ENDIAN * 2]; + t.s32[0] = a.s16[2 - !!CONFIG_BIG_ENDIAN * 2]; + return t; +} + +SIMD_INLINE c_v64 c_v64_shuffle_8(c_v64 a, c_v64 pattern) { + c_v64 t; + int c; + for (c = 0; c < 8; c++) { + if (SIMD_CHECK && (pattern.u8[c] & ~7)) { + fprintf(stderr, "Error: Undefined v64_shuffle_8 index %d/%d\n", + pattern.u8[c], c); + abort(); + } + t.u8[c] = + a.u8[CONFIG_BIG_ENDIAN ? 7 - (pattern.u8[c] & 7) : pattern.u8[c] & 7]; + } + return t; +} + +SIMD_INLINE int64_t c_v64_dotp_su8(c_v64 a, c_v64 b) { + return a.s8[7] * b.u8[7] + a.s8[6] * b.u8[6] + a.s8[5] * b.u8[5] + + a.s8[4] * b.u8[4] + a.s8[3] * b.u8[3] + a.s8[2] * b.u8[2] + + a.s8[1] * b.u8[1] + a.s8[0] * b.u8[0]; +} + +SIMD_INLINE int64_t c_v64_dotp_s16(c_v64 a, c_v64 b) { + return (int64_t)(a.s16[3] * b.s16[3] + a.s16[2] * b.s16[2]) + + (int64_t)(a.s16[1] * b.s16[1] + a.s16[0] * b.s16[0]); +} + +SIMD_INLINE uint64_t c_v64_hadd_u8(c_v64 a) { + return a.u8[7] + a.u8[6] + a.u8[5] + a.u8[4] + a.u8[3] + a.u8[2] + a.u8[1] + + a.u8[0]; +} + +SIMD_INLINE int64_t c_v64_hadd_s16(c_v64 a) { + return a.s16[3] + a.s16[2] + a.s16[1] + a.s16[0]; +} + +typedef uint32_t c_sad64_internal; + +/* Implementation dependent return value. Result must be finalised with + v64_sad_u8_sum(). + The result for more than 32 v64_sad_u8() calls is undefined. */ +SIMD_INLINE c_sad64_internal c_v64_sad_u8_init() { return 0; } + +SIMD_INLINE c_sad64_internal c_v64_sad_u8(c_sad64_internal s, c_v64 a, + c_v64 b) { + int c; + for (c = 0; c < 8; c++) + s += a.u8[c] > b.u8[c] ? a.u8[c] - b.u8[c] : b.u8[c] - a.u8[c]; + return s; +} + +SIMD_INLINE uint32_t c_v64_sad_u8_sum(c_sad64_internal s) { return s; } + +typedef uint32_t c_ssd64_internal; + +/* Implementation dependent return value. Result must be finalised with + * v64_ssd_u8_sum(). */ +SIMD_INLINE c_ssd64_internal c_v64_ssd_u8_init() { return 0; } + +SIMD_INLINE c_ssd64_internal c_v64_ssd_u8(c_ssd64_internal s, c_v64 a, + c_v64 b) { + int c; + for (c = 0; c < 8; c++) s += (a.u8[c] - b.u8[c]) * (a.u8[c] - b.u8[c]); + return s; +} + +SIMD_INLINE uint32_t c_v64_ssd_u8_sum(c_ssd64_internal s) { return s; } + +SIMD_INLINE c_v64 c_v64_or(c_v64 a, c_v64 b) { + c_v64 t; + t.u64 = a.u64 | b.u64; + return t; +} + +SIMD_INLINE c_v64 c_v64_xor(c_v64 a, c_v64 b) { + c_v64 t; + t.u64 = a.u64 ^ b.u64; + return t; +} + +SIMD_INLINE c_v64 c_v64_and(c_v64 a, c_v64 b) { + c_v64 t; + t.u64 = a.u64 & b.u64; + return t; +} + +SIMD_INLINE c_v64 c_v64_andn(c_v64 a, c_v64 b) { + c_v64 t; + t.u64 = a.u64 & ~b.u64; + return t; +} + +SIMD_INLINE c_v64 c_v64_mullo_s16(c_v64 a, c_v64 b) { + c_v64 t; + int c; + for (c = 0; c < 4; c++) t.s16[c] = (int16_t)(a.s16[c] * b.s16[c]); + return t; +} + +SIMD_INLINE c_v64 c_v64_mulhi_s16(c_v64 a, c_v64 b) { + c_v64 t; + int c; + for (c = 0; c < 4; c++) t.s16[c] = (a.s16[c] * b.s16[c]) >> 16; + return t; +} + +SIMD_INLINE c_v64 c_v64_mullo_s32(c_v64 a, c_v64 b) { + c_v64 t; + t.s32[0] = (int32_t)((int64_t)a.s32[0] * b.s32[0]); + t.s32[1] = (int32_t)((int64_t)a.s32[1] * b.s32[1]); + return t; +} + +SIMD_INLINE c_v64 c_v64_madd_s16(c_v64 a, c_v64 b) { + c_v64 t; + t.s32[0] = a.s16[0] * b.s16[0] + a.s16[1] * b.s16[1]; + t.s32[1] = a.s16[2] * b.s16[2] + a.s16[3] * b.s16[3]; + return t; +} + +SIMD_INLINE c_v64 c_v64_madd_us8(c_v64 a, c_v64 b) { + c_v64 t; + int32_t u; + u = a.u8[0] * b.s8[0] + a.u8[1] * b.s8[1]; + t.s16[0] = u > 32767 ? 32767 : u < -32768 ? -32768 : u; + u = a.u8[2] * b.s8[2] + a.u8[3] * b.s8[3]; + t.s16[1] = u > 32767 ? 32767 : u < -32768 ? -32768 : u; + u = a.u8[4] * b.s8[4] + a.u8[5] * b.s8[5]; + t.s16[2] = u > 32767 ? 32767 : u < -32768 ? -32768 : u; + u = a.u8[6] * b.s8[6] + a.u8[7] * b.s8[7]; + t.s16[3] = u > 32767 ? 32767 : u < -32768 ? -32768 : u; + return t; +} + +SIMD_INLINE c_v64 c_v64_avg_u8(c_v64 a, c_v64 b) { + c_v64 t; + int c; + for (c = 0; c < 8; c++) t.u8[c] = (a.u8[c] + b.u8[c] + 1) >> 1; + return t; +} + +SIMD_INLINE c_v64 c_v64_rdavg_u8(c_v64 a, c_v64 b) { + c_v64 t; + int c; + for (c = 0; c < 8; c++) t.u8[c] = (a.u8[c] + b.u8[c]) >> 1; + return t; +} + +SIMD_INLINE c_v64 c_v64_rdavg_u16(c_v64 a, c_v64 b) { + c_v64 t; + int c; + for (c = 0; c < 4; c++) t.u16[c] = (a.u16[c] + b.u16[c]) >> 1; + return t; +} + +SIMD_INLINE c_v64 c_v64_avg_u16(c_v64 a, c_v64 b) { + c_v64 t; + int c; + for (c = 0; c < 4; c++) t.u16[c] = (a.u16[c] + b.u16[c] + 1) >> 1; + return t; +} + +SIMD_INLINE c_v64 c_v64_min_u8(c_v64 a, c_v64 b) { + c_v64 t; + int c; + for (c = 0; c < 8; c++) t.u8[c] = a.u8[c] > b.u8[c] ? b.u8[c] : a.u8[c]; + return t; +} + +SIMD_INLINE c_v64 c_v64_max_u8(c_v64 a, c_v64 b) { + c_v64 t; + int c; + for (c = 0; c < 8; c++) t.u8[c] = a.u8[c] > b.u8[c] ? a.u8[c] : b.u8[c]; + return t; +} + +SIMD_INLINE c_v64 c_v64_min_s8(c_v64 a, c_v64 b) { + c_v64 t; + int c; + for (c = 0; c < 8; c++) t.s8[c] = a.s8[c] > b.s8[c] ? b.s8[c] : a.s8[c]; + return t; +} + +SIMD_INLINE c_v64 c_v64_max_s8(c_v64 a, c_v64 b) { + c_v64 t; + int c; + for (c = 0; c < 8; c++) t.s8[c] = a.s8[c] > b.s8[c] ? a.s8[c] : b.s8[c]; + return t; +} + +SIMD_INLINE c_v64 c_v64_min_s16(c_v64 a, c_v64 b) { + c_v64 t; + int c; + for (c = 0; c < 4; c++) t.s16[c] = a.s16[c] > b.s16[c] ? b.s16[c] : a.s16[c]; + return t; +} + +SIMD_INLINE c_v64 c_v64_max_s16(c_v64 a, c_v64 b) { + c_v64 t; + int c; + for (c = 0; c < 4; c++) t.s16[c] = a.s16[c] > b.s16[c] ? a.s16[c] : b.s16[c]; + return t; +} + +SIMD_INLINE c_v64 c_v64_cmpgt_s8(c_v64 a, c_v64 b) { + c_v64 t; + int c; + for (c = 0; c < 8; c++) t.s8[c] = -(a.s8[c] > b.s8[c]); + return t; +} + +SIMD_INLINE c_v64 c_v64_cmplt_s8(c_v64 a, c_v64 b) { + c_v64 t; + int c; + for (c = 0; c < 8; c++) t.s8[c] = -(a.s8[c] < b.s8[c]); + return t; +} + +SIMD_INLINE c_v64 c_v64_cmpeq_8(c_v64 a, c_v64 b) { + c_v64 t; + int c; + for (c = 0; c < 8; c++) t.s8[c] = -(a.u8[c] == b.u8[c]); + return t; +} + +SIMD_INLINE c_v64 c_v64_cmpgt_s16(c_v64 a, c_v64 b) { + c_v64 t; + int c; + for (c = 0; c < 4; c++) t.s16[c] = -(a.s16[c] > b.s16[c]); + return t; +} + +SIMD_INLINE c_v64 c_v64_cmplt_s16(c_v64 a, c_v64 b) { + c_v64 t; + int c; + for (c = 0; c < 4; c++) t.s16[c] = -(a.s16[c] < b.s16[c]); + return t; +} + +SIMD_INLINE c_v64 c_v64_cmpeq_16(c_v64 a, c_v64 b) { + c_v64 t; + int c; + for (c = 0; c < 4; c++) t.s16[c] = -(a.u16[c] == b.u16[c]); + return t; +} + +SIMD_INLINE c_v64 c_v64_shl_8(c_v64 a, unsigned int n) { + c_v64 t; + int c; + if (SIMD_CHECK && n > 7) { + fprintf(stderr, "Error: Undefined u8 shift left %d\n", n); + abort(); + } + for (c = 0; c < 8; c++) t.s8[c] = a.u8[c] << n; + return t; +} + +SIMD_INLINE c_v64 c_v64_shr_u8(c_v64 a, unsigned int n) { + c_v64 t; + int c; + if (SIMD_CHECK && n > 7) { + fprintf(stderr, "Error: Undefined u8 shift right %d\n", n); + abort(); + } + for (c = 0; c < 8; c++) t.u8[c] = a.u8[c] >> n; + return t; +} + +SIMD_INLINE c_v64 c_v64_shr_s8(c_v64 a, unsigned int n) { + c_v64 t; + int c; + if (SIMD_CHECK && n > 7) { + fprintf(stderr, "Error: Undefined s8 shift right %d\n", n); + abort(); + } + for (c = 0; c < 8; c++) t.s8[c] = a.s8[c] >> n; + return t; +} + +SIMD_INLINE c_v64 c_v64_shl_16(c_v64 a, unsigned int n) { + c_v64 t; + int c; + if (SIMD_CHECK && n > 15) { + fprintf(stderr, "Error: Undefined u16 shift left %d\n", n); + abort(); + } + for (c = 0; c < 4; c++) t.u16[c] = a.u16[c] << n; + return t; +} + +SIMD_INLINE c_v64 c_v64_shr_u16(c_v64 a, unsigned int n) { + c_v64 t; + int c; + if (SIMD_CHECK && n > 15) { + fprintf(stderr, "Error: Undefined u16 shift right %d\n", n); + abort(); + } + for (c = 0; c < 4; c++) t.u16[c] = a.u16[c] >> n; + return t; +} + +SIMD_INLINE c_v64 c_v64_shr_s16(c_v64 a, unsigned int n) { + c_v64 t; + int c; + if (SIMD_CHECK && n > 15) { + fprintf(stderr, "Error: undefined s16 shift right %d\n", n); + abort(); + } + for (c = 0; c < 4; c++) t.s16[c] = a.s16[c] >> n; + return t; +} + +SIMD_INLINE c_v64 c_v64_shl_32(c_v64 a, unsigned int n) { + c_v64 t; + if (SIMD_CHECK && n > 31) { + fprintf(stderr, "Error: undefined u32 shift left %d\n", n); + abort(); + } + t.u32[1] = a.u32[1] << n; + t.u32[0] = a.u32[0] << n; + return t; +} + +SIMD_INLINE c_v64 c_v64_shr_u32(c_v64 a, unsigned int n) { + c_v64 t; + if (SIMD_CHECK && n > 31) { + fprintf(stderr, "Error: undefined u32 shift right %d\n", n); + abort(); + } + t.u32[1] = a.u32[1] >> n; + t.u32[0] = a.u32[0] >> n; + return t; +} + +SIMD_INLINE c_v64 c_v64_shr_s32(c_v64 a, unsigned int n) { + c_v64 t; + if (SIMD_CHECK && n > 31) { + fprintf(stderr, "Error: undefined s32 shift right %d\n", n); + abort(); + } + t.s32[1] = a.s32[1] >> n; + t.s32[0] = a.s32[0] >> n; + return t; +} + +SIMD_INLINE c_v64 c_v64_shr_n_byte(c_v64 x, unsigned int i) { + c_v64 t; + t.u64 = x.u64 >> i * 8; + return t; +} + +SIMD_INLINE c_v64 c_v64_shl_n_byte(c_v64 x, unsigned int i) { + c_v64 t; + t.u64 = x.u64 << i * 8; + return t; +} + +SIMD_INLINE c_v64 c_v64_align(c_v64 a, c_v64 b, unsigned int c) { + if (SIMD_CHECK && c > 7) { + fprintf(stderr, "Error: undefined alignment %d\n", c); + abort(); + } + return c ? c_v64_or(c_v64_shr_n_byte(b, c), c_v64_shl_n_byte(a, 8 - c)) : b; +} + +SIMD_INLINE c_v64 c_v64_shl_n_8(c_v64 a, unsigned int c) { + return c_v64_shl_8(a, c); +} + +SIMD_INLINE c_v64 c_v64_shr_n_u8(c_v64 a, unsigned int c) { + return c_v64_shr_u8(a, c); +} + +SIMD_INLINE c_v64 c_v64_shr_n_s8(c_v64 a, unsigned int c) { + return c_v64_shr_s8(a, c); +} + +SIMD_INLINE c_v64 c_v64_shl_n_16(c_v64 a, unsigned int c) { + return c_v64_shl_16(a, c); +} + +SIMD_INLINE c_v64 c_v64_shr_n_u16(c_v64 a, unsigned int c) { + return c_v64_shr_u16(a, c); +} + +SIMD_INLINE c_v64 c_v64_shr_n_s16(c_v64 a, unsigned int c) { + return c_v64_shr_s16(a, c); +} + +SIMD_INLINE c_v64 c_v64_shl_n_32(c_v64 a, unsigned int c) { + return c_v64_shl_32(a, c); +} + +SIMD_INLINE c_v64 c_v64_shr_n_u32(c_v64 a, unsigned int c) { + return c_v64_shr_u32(a, c); +} + +SIMD_INLINE c_v64 c_v64_shr_n_s32(c_v64 a, unsigned int c) { + return c_v64_shr_s32(a, c); +} + +#endif // AOM_AOM_DSP_SIMD_V64_INTRINSICS_C_H_ diff --git a/third_party/aom/aom_dsp/simd/v64_intrinsics_x86.h b/third_party/aom/aom_dsp/simd/v64_intrinsics_x86.h new file mode 100644 index 0000000000..5f9a57b379 --- /dev/null +++ b/third_party/aom/aom_dsp/simd/v64_intrinsics_x86.h @@ -0,0 +1,491 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_SIMD_V64_INTRINSICS_X86_H_ +#define AOM_AOM_DSP_SIMD_V64_INTRINSICS_X86_H_ + +#include +#if defined(__SSSE3__) +#include +#endif +#if defined(__SSE4_1__) +#include +#endif + +typedef __m128i v64; + +SIMD_INLINE uint32_t v64_low_u32(v64 a) { + return (uint32_t)_mm_cvtsi128_si32(a); +} + +SIMD_INLINE uint32_t v64_high_u32(v64 a) { + return (uint32_t)_mm_cvtsi128_si32(_mm_srli_si128(a, 4)); +} + +SIMD_INLINE int32_t v64_low_s32(v64 a) { return (int32_t)_mm_cvtsi128_si32(a); } + +SIMD_INLINE int32_t v64_high_s32(v64 a) { + return (int32_t)_mm_cvtsi128_si32(_mm_srli_si128(a, 4)); +} + +SIMD_INLINE v64 v64_from_16(uint16_t a, uint16_t b, uint16_t c, uint16_t d) { + return _mm_packs_epi32( + _mm_set_epi32((int16_t)a, (int16_t)b, (int16_t)c, (int16_t)d), + _mm_setzero_si128()); +} + +SIMD_INLINE v64 v64_from_32(uint32_t x, uint32_t y) { + return _mm_set_epi32(0, 0, x, y); +} + +SIMD_INLINE v64 v64_from_64(uint64_t x) { +#ifdef __x86_64__ + return _mm_cvtsi64_si128(x); +#else + return _mm_set_epi32(0, 0, x >> 32, (uint32_t)x); +#endif +} + +SIMD_INLINE uint64_t v64_u64(v64 x) { + return (uint64_t)v64_low_u32(x) | ((uint64_t)v64_high_u32(x) << 32); +} + +SIMD_INLINE uint32_t u32_load_aligned(const void *p) { + return *((uint32_t *)p); +} + +SIMD_INLINE uint32_t u32_load_unaligned(const void *p) { + return *((uint32_t *)p); +} + +SIMD_INLINE void u32_store_aligned(void *p, uint32_t a) { + *((uint32_t *)p) = a; +} + +SIMD_INLINE void u32_store_unaligned(void *p, uint32_t a) { + *((uint32_t *)p) = a; +} + +SIMD_INLINE v64 v64_load_aligned(const void *p) { + return _mm_loadl_epi64((__m128i *)p); +} + +SIMD_INLINE v64 v64_load_unaligned(const void *p) { + return _mm_loadl_epi64((__m128i *)p); +} + +SIMD_INLINE void v64_store_aligned(void *p, v64 a) { + _mm_storel_epi64((__m128i *)p, a); +} + +SIMD_INLINE void v64_store_unaligned(void *p, v64 a) { + _mm_storel_epi64((__m128i *)p, a); +} + +#if defined(__OPTIMIZE__) && __OPTIMIZE__ && !defined(__clang__) +#define v64_align(a, b, c) \ + ((c) ? _mm_srli_si128(_mm_unpacklo_epi64(b, a), (c)) : b) +#else +#define v64_align(a, b, c) \ + ((c) ? v64_from_64((v64_u64(b) >> (c)*8) | (v64_u64(a) << (8 - (c)) * 8)) \ + : (b)) +#endif + +SIMD_INLINE v64 v64_zero() { return _mm_setzero_si128(); } + +SIMD_INLINE v64 v64_dup_8(uint8_t x) { return _mm_set1_epi8(x); } + +SIMD_INLINE v64 v64_dup_16(uint16_t x) { return _mm_set1_epi16(x); } + +SIMD_INLINE v64 v64_dup_32(uint32_t x) { return _mm_set1_epi32(x); } + +SIMD_INLINE v64 v64_add_8(v64 a, v64 b) { return _mm_add_epi8(a, b); } + +SIMD_INLINE v64 v64_add_16(v64 a, v64 b) { return _mm_add_epi16(a, b); } + +SIMD_INLINE v64 v64_sadd_u8(v64 a, v64 b) { return _mm_adds_epu8(a, b); } + +SIMD_INLINE v64 v64_sadd_s8(v64 a, v64 b) { return _mm_adds_epi8(a, b); } + +SIMD_INLINE v64 v64_sadd_s16(v64 a, v64 b) { return _mm_adds_epi16(a, b); } + +SIMD_INLINE v64 v64_add_32(v64 a, v64 b) { return _mm_add_epi32(a, b); } + +SIMD_INLINE v64 v64_sub_8(v64 a, v64 b) { return _mm_sub_epi8(a, b); } + +SIMD_INLINE v64 v64_ssub_u8(v64 a, v64 b) { return _mm_subs_epu8(a, b); } + +SIMD_INLINE v64 v64_ssub_s8(v64 a, v64 b) { return _mm_subs_epi8(a, b); } + +SIMD_INLINE v64 v64_sub_16(v64 a, v64 b) { return _mm_sub_epi16(a, b); } + +SIMD_INLINE v64 v64_ssub_s16(v64 a, v64 b) { return _mm_subs_epi16(a, b); } + +SIMD_INLINE v64 v64_ssub_u16(v64 a, v64 b) { return _mm_subs_epu16(a, b); } + +SIMD_INLINE v64 v64_sub_32(v64 a, v64 b) { return _mm_sub_epi32(a, b); } + +SIMD_INLINE v64 v64_abs_s16(v64 a) { +#if defined(__SSSE3__) + return _mm_abs_epi16(a); +#else + return _mm_max_epi16(a, _mm_sub_epi16(_mm_setzero_si128(), a)); +#endif +} + +SIMD_INLINE v64 v64_abs_s8(v64 a) { +#if defined(__SSSE3__) + return _mm_abs_epi8(a); +#else + v64 sign = _mm_cmplt_epi8(a, _mm_setzero_si128()); + return _mm_xor_si128(sign, _mm_add_epi8(a, sign)); +#endif +} + +SIMD_INLINE v64 v64_ziplo_8(v64 a, v64 b) { return _mm_unpacklo_epi8(b, a); } + +SIMD_INLINE v64 v64_ziphi_8(v64 a, v64 b) { + return _mm_srli_si128(_mm_unpacklo_epi8(b, a), 8); +} + +SIMD_INLINE v64 v64_ziplo_16(v64 a, v64 b) { return _mm_unpacklo_epi16(b, a); } + +SIMD_INLINE v64 v64_ziphi_16(v64 a, v64 b) { + return _mm_srli_si128(_mm_unpacklo_epi16(b, a), 8); +} + +SIMD_INLINE v64 v64_ziplo_32(v64 a, v64 b) { return _mm_unpacklo_epi32(b, a); } + +SIMD_INLINE v64 v64_ziphi_32(v64 a, v64 b) { + return _mm_srli_si128(_mm_unpacklo_epi32(b, a), 8); +} + +SIMD_INLINE v64 v64_pack_s32_s16(v64 a, v64 b) { + __m128i t = _mm_unpacklo_epi64(b, a); + return _mm_packs_epi32(t, t); +} + +SIMD_INLINE v64 v64_pack_s32_u16(v64 a, v64 b) { +#if defined(__SSE4_1__) + __m128i t = _mm_unpacklo_epi64(b, a); + return _mm_packus_epi32(t, t); +#else + int32_t ah = v64_high_u32(a); + int32_t al = v64_low_u32(a); + int32_t bh = v64_high_u32(b); + int32_t bl = v64_low_u32(b); + return v64_from_16(ah > 65535 ? 65535 : ah < 0 ? 0 : ah, + al > 65535 ? 65535 : al < 0 ? 0 : al, + bh > 65535 ? 65535 : bh < 0 ? 0 : bh, + bl > 65535 ? 65535 : bl < 0 ? 0 : bl); +#endif +} + +SIMD_INLINE v64 v64_pack_s16_u8(v64 a, v64 b) { + __m128i t = _mm_unpacklo_epi64(b, a); + return _mm_packus_epi16(t, t); +} + +SIMD_INLINE v64 v64_pack_s16_s8(v64 a, v64 b) { + __m128i t = _mm_unpacklo_epi64(b, a); + return _mm_packs_epi16(t, t); +} + +SIMD_INLINE v64 v64_unziphi_8(v64 a, v64 b) { +#if defined(__SSSE3__) + return _mm_shuffle_epi8(_mm_unpacklo_epi64(b, a), + v64_from_64(0x0f0d0b0907050301LL)); +#else + return _mm_packus_epi16( + _mm_unpacklo_epi64(_mm_srli_epi16(b, 8), _mm_srli_epi16(a, 8)), + _mm_setzero_si128()); +#endif +} + +SIMD_INLINE v64 v64_unziplo_8(v64 a, v64 b) { +#if defined(__SSSE3__) + return _mm_shuffle_epi8(_mm_unpacklo_epi64(b, a), + v64_from_64(0x0e0c0a0806040200LL)); +#else + return v64_unziphi_8(_mm_slli_si128(a, 1), _mm_slli_si128(b, 1)); +#endif +} + +SIMD_INLINE v64 v64_unziphi_16(v64 a, v64 b) { +#if defined(__SSSE3__) + return _mm_shuffle_epi8(_mm_unpacklo_epi64(b, a), + v64_from_64(0x0f0e0b0a07060302LL)); +#else + return _mm_packs_epi32( + _mm_unpacklo_epi64(_mm_srai_epi32(b, 16), _mm_srai_epi32(a, 16)), + _mm_setzero_si128()); +#endif +} + +SIMD_INLINE v64 v64_unziplo_16(v64 a, v64 b) { +#if defined(__SSSE3__) + return _mm_shuffle_epi8(_mm_unpacklo_epi64(b, a), + v64_from_64(0x0d0c090805040100LL)); +#else + return v64_unziphi_16(_mm_slli_si128(a, 2), _mm_slli_si128(b, 2)); +#endif +} + +SIMD_INLINE v64 v64_unpacklo_u8_s16(v64 a) { + return _mm_unpacklo_epi8(a, _mm_setzero_si128()); +} + +SIMD_INLINE v64 v64_unpackhi_u8_s16(v64 a) { + return _mm_srli_si128(_mm_unpacklo_epi8(a, _mm_setzero_si128()), 8); +} + +SIMD_INLINE v64 v64_unpacklo_s8_s16(v64 a) { + return _mm_srai_epi16(_mm_unpacklo_epi8(a, a), 8); +} + +SIMD_INLINE v64 v64_unpackhi_s8_s16(v64 a) { + return _mm_srli_si128(_mm_srai_epi16(_mm_unpacklo_epi8(a, a), 8), 8); +} + +SIMD_INLINE v64 v64_unpacklo_u16_s32(v64 a) { + return _mm_unpacklo_epi16(a, _mm_setzero_si128()); +} + +SIMD_INLINE v64 v64_unpacklo_s16_s32(v64 a) { + return _mm_srai_epi32(_mm_unpacklo_epi16(_mm_setzero_si128(), a), 16); +} + +SIMD_INLINE v64 v64_unpackhi_u16_s32(v64 a) { + return _mm_srli_si128(_mm_unpacklo_epi16(a, _mm_setzero_si128()), 8); +} + +SIMD_INLINE v64 v64_unpackhi_s16_s32(v64 a) { + return _mm_srli_si128( + _mm_srai_epi32(_mm_unpacklo_epi16(_mm_setzero_si128(), a), 16), 8); +} + +SIMD_INLINE v64 v64_shuffle_8(v64 x, v64 pattern) { +#if defined(__SSSE3__) + return _mm_shuffle_epi8(x, pattern); +#else + v64 output; + unsigned char *input = (unsigned char *)&x; + unsigned char *index = (unsigned char *)&pattern; + char *selected = (char *)&output; + int counter; + + for (counter = 0; counter < 8; counter++) { + selected[counter] = input[index[counter]]; + } + + return output; +#endif +} + +SIMD_INLINE int64_t v64_dotp_su8(v64 a, v64 b) { + __m128i t = _mm_madd_epi16(_mm_srai_epi16(_mm_unpacklo_epi8(a, a), 8), + _mm_unpacklo_epi8(b, _mm_setzero_si128())); + t = _mm_add_epi32(t, _mm_srli_si128(t, 8)); + t = _mm_add_epi32(t, _mm_srli_si128(t, 4)); + return (int32_t)v64_low_u32(t); +} + +SIMD_INLINE int64_t v64_dotp_s16(v64 a, v64 b) { + __m128i r = _mm_madd_epi16(a, b); +#if defined(__SSE4_1__) && defined(__x86_64__) + __m128i x = _mm_cvtepi32_epi64(r); + return _mm_cvtsi128_si64(_mm_add_epi64(x, _mm_srli_si128(x, 8))); +#else + return (int64_t)_mm_cvtsi128_si32(_mm_srli_si128(r, 4)) + + (int64_t)_mm_cvtsi128_si32(r); +#endif +} + +SIMD_INLINE uint64_t v64_hadd_u8(v64 a) { + return v64_low_u32(_mm_sad_epu8(a, _mm_setzero_si128())); +} + +SIMD_INLINE int64_t v64_hadd_s16(v64 a) { + return v64_dotp_s16(a, v64_dup_16(1)); +} + +typedef v64 sad64_internal; + +SIMD_INLINE sad64_internal v64_sad_u8_init() { return _mm_setzero_si128(); } + +/* Implementation dependent return value. Result must be finalised with + v64_sad_u8_sum(). + The result for more than 32 v64_sad_u8() calls is undefined. */ +SIMD_INLINE sad64_internal v64_sad_u8(sad64_internal s, v64 a, v64 b) { + return _mm_add_epi64(s, _mm_sad_epu8(a, b)); +} + +SIMD_INLINE uint32_t v64_sad_u8_sum(sad64_internal s) { return v64_low_u32(s); } + +typedef v64 ssd64_internal; + +SIMD_INLINE ssd64_internal v64_ssd_u8_init() { return _mm_setzero_si128(); } + +/* Implementation dependent return value. Result must be finalised with + * v64_ssd_u8_sum(). */ +SIMD_INLINE ssd64_internal v64_ssd_u8(ssd64_internal s, v64 a, v64 b) { + v64 l = v64_sub_16(v64_ziplo_8(v64_zero(), a), v64_ziplo_8(v64_zero(), b)); + v64 h = v64_sub_16(v64_ziphi_8(v64_zero(), a), v64_ziphi_8(v64_zero(), b)); + v64 r = v64_add_32(_mm_madd_epi16(l, l), _mm_madd_epi16(h, h)); + return _mm_add_epi64( + s, v64_ziplo_32(v64_zero(), _mm_add_epi32(r, _mm_srli_si128(r, 4)))); +} + +SIMD_INLINE uint32_t v64_ssd_u8_sum(sad64_internal s) { return v64_low_u32(s); } + +SIMD_INLINE v64 v64_or(v64 a, v64 b) { return _mm_or_si128(a, b); } + +SIMD_INLINE v64 v64_xor(v64 a, v64 b) { return _mm_xor_si128(a, b); } + +SIMD_INLINE v64 v64_and(v64 a, v64 b) { return _mm_and_si128(a, b); } + +SIMD_INLINE v64 v64_andn(v64 a, v64 b) { return _mm_andnot_si128(b, a); } + +SIMD_INLINE v64 v64_mullo_s16(v64 a, v64 b) { return _mm_mullo_epi16(a, b); } + +SIMD_INLINE v64 v64_mulhi_s16(v64 a, v64 b) { return _mm_mulhi_epi16(a, b); } + +SIMD_INLINE v64 v64_mullo_s32(v64 a, v64 b) { +#if defined(__SSE4_1__) + return _mm_mullo_epi32(a, b); +#else + return _mm_unpacklo_epi32( + _mm_mul_epu32(a, b), + _mm_mul_epu32(_mm_srli_si128(a, 4), _mm_srli_si128(b, 4))); +#endif +} + +SIMD_INLINE v64 v64_madd_s16(v64 a, v64 b) { return _mm_madd_epi16(a, b); } + +SIMD_INLINE v64 v64_madd_us8(v64 a, v64 b) { +#if defined(__SSSE3__) + return _mm_maddubs_epi16(a, b); +#else + __m128i t = _mm_madd_epi16(_mm_unpacklo_epi8(a, _mm_setzero_si128()), + _mm_srai_epi16(_mm_unpacklo_epi8(b, b), 8)); + return _mm_packs_epi32(t, t); +#endif +} + +SIMD_INLINE v64 v64_avg_u8(v64 a, v64 b) { return _mm_avg_epu8(a, b); } + +SIMD_INLINE v64 v64_rdavg_u8(v64 a, v64 b) { + return _mm_sub_epi8(_mm_avg_epu8(a, b), + _mm_and_si128(_mm_xor_si128(a, b), v64_dup_8(1))); +} + +SIMD_INLINE v64 v64_rdavg_u16(v64 a, v64 b) { + return _mm_sub_epi16(_mm_avg_epu16(a, b), + _mm_and_si128(_mm_xor_si128(a, b), v64_dup_16(1))); +} + +SIMD_INLINE v64 v64_avg_u16(v64 a, v64 b) { return _mm_avg_epu16(a, b); } + +SIMD_INLINE v64 v64_min_u8(v64 a, v64 b) { return _mm_min_epu8(a, b); } + +SIMD_INLINE v64 v64_max_u8(v64 a, v64 b) { return _mm_max_epu8(a, b); } + +SIMD_INLINE v64 v64_min_s8(v64 a, v64 b) { +#if defined(__SSE4_1__) + return _mm_min_epi8(a, b); +#else + v64 mask = _mm_cmplt_epi8(a, b); + return _mm_or_si128(_mm_andnot_si128(mask, b), _mm_and_si128(mask, a)); +#endif +} + +SIMD_INLINE v64 v64_max_s8(v64 a, v64 b) { +#if defined(__SSE4_1__) + return _mm_max_epi8(a, b); +#else + v64 mask = _mm_cmplt_epi8(b, a); + return _mm_or_si128(_mm_andnot_si128(mask, b), _mm_and_si128(mask, a)); +#endif +} + +SIMD_INLINE v64 v64_min_s16(v64 a, v64 b) { return _mm_min_epi16(a, b); } + +SIMD_INLINE v64 v64_max_s16(v64 a, v64 b) { return _mm_max_epi16(a, b); } + +SIMD_INLINE v64 v64_cmpgt_s8(v64 a, v64 b) { return _mm_cmpgt_epi8(a, b); } + +SIMD_INLINE v64 v64_cmplt_s8(v64 a, v64 b) { return _mm_cmplt_epi8(a, b); } + +SIMD_INLINE v64 v64_cmpeq_8(v64 a, v64 b) { return _mm_cmpeq_epi8(a, b); } + +SIMD_INLINE v64 v64_cmpgt_s16(v64 a, v64 b) { return _mm_cmpgt_epi16(a, b); } + +SIMD_INLINE v64 v64_cmplt_s16(v64 a, v64 b) { return _mm_cmplt_epi16(a, b); } + +SIMD_INLINE v64 v64_cmpeq_16(v64 a, v64 b) { return _mm_cmpeq_epi16(a, b); } + +SIMD_INLINE v64 v64_shl_8(v64 a, unsigned int c) { + return _mm_and_si128(_mm_set1_epi8((uint8_t)(0xff << c)), + _mm_sll_epi16(a, _mm_cvtsi32_si128(c))); +} + +SIMD_INLINE v64 v64_shr_u8(v64 a, unsigned int c) { + return _mm_and_si128(_mm_set1_epi8(0xff >> c), + _mm_srl_epi16(a, _mm_cvtsi32_si128(c))); +} + +SIMD_INLINE v64 v64_shr_s8(v64 a, unsigned int c) { + return _mm_packs_epi16( + _mm_sra_epi16(_mm_unpacklo_epi8(a, a), _mm_cvtsi32_si128(c + 8)), a); +} + +SIMD_INLINE v64 v64_shl_16(v64 a, unsigned int c) { + return _mm_sll_epi16(a, _mm_cvtsi32_si128(c)); +} + +SIMD_INLINE v64 v64_shr_u16(v64 a, unsigned int c) { + return _mm_srl_epi16(a, _mm_cvtsi32_si128(c)); +} + +SIMD_INLINE v64 v64_shr_s16(v64 a, unsigned int c) { + return _mm_sra_epi16(a, _mm_cvtsi32_si128(c)); +} + +SIMD_INLINE v64 v64_shl_32(v64 a, unsigned int c) { + return _mm_sll_epi32(a, _mm_cvtsi32_si128(c)); +} + +SIMD_INLINE v64 v64_shr_u32(v64 a, unsigned int c) { + return _mm_srl_epi32(a, _mm_cvtsi32_si128(c)); +} + +SIMD_INLINE v64 v64_shr_s32(v64 a, unsigned int c) { + return _mm_sra_epi32(a, _mm_cvtsi32_si128(c)); +} + +/* These intrinsics require immediate values, so we must use #defines + to enforce that. */ +#define v64_shl_n_byte(a, c) _mm_slli_si128(a, c) +#define v64_shr_n_byte(a, c) _mm_srli_si128(_mm_unpacklo_epi64(a, a), c + 8) +#define v64_shl_n_8(a, c) \ + _mm_and_si128(_mm_set1_epi8((uint8_t)(0xff << (c))), _mm_slli_epi16(a, c)) +#define v64_shr_n_u8(a, c) \ + _mm_and_si128(_mm_set1_epi8(0xff >> (c)), _mm_srli_epi16(a, c)) +#define v64_shr_n_s8(a, c) \ + _mm_packs_epi16(_mm_srai_epi16(_mm_unpacklo_epi8(a, a), (c) + 8), a) +#define v64_shl_n_16(a, c) _mm_slli_epi16(a, c) +#define v64_shr_n_u16(a, c) _mm_srli_epi16(a, c) +#define v64_shr_n_s16(a, c) _mm_srai_epi16(a, c) +#define v64_shl_n_32(a, c) _mm_slli_epi32(a, c) +#define v64_shr_n_u32(a, c) _mm_srli_epi32(a, c) +#define v64_shr_n_s32(a, c) _mm_srai_epi32(a, c) + +#endif // AOM_AOM_DSP_SIMD_V64_INTRINSICS_X86_H_ diff --git a/third_party/aom/aom_dsp/sse.c b/third_party/aom/aom_dsp/sse.c new file mode 100644 index 0000000000..2493948079 --- /dev/null +++ b/third_party/aom/aom_dsp/sse.c @@ -0,0 +1,52 @@ +/* + * Copyright (c) 2018, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +/* Sum the difference between every corresponding element of the buffers. */ + +#include "config/aom_config.h" +#include "config/aom_dsp_rtcd.h" + +#include "aom/aom_integer.h" + +int64_t aom_sse_c(const uint8_t *a, int a_stride, const uint8_t *b, + int b_stride, int width, int height) { + int y, x; + int64_t sse = 0; + + for (y = 0; y < height; y++) { + for (x = 0; x < width; x++) { + const int32_t diff = abs(a[x] - b[x]); + sse += diff * diff; + } + + a += a_stride; + b += b_stride; + } + return sse; +} + +int64_t aom_highbd_sse_c(const uint8_t *a8, int a_stride, const uint8_t *b8, + int b_stride, int width, int height) { + int y, x; + int64_t sse = 0; + uint16_t *a = CONVERT_TO_SHORTPTR(a8); + uint16_t *b = CONVERT_TO_SHORTPTR(b8); + for (y = 0; y < height; y++) { + for (x = 0; x < width; x++) { + const int32_t diff = (int32_t)(a[x]) - (int32_t)(b[x]); + sse += diff * diff; + } + + a += a_stride; + b += b_stride; + } + return sse; +} diff --git a/third_party/aom/aom_dsp/ssim.c b/third_party/aom/aom_dsp/ssim.c new file mode 100644 index 0000000000..681770ba97 --- /dev/null +++ b/third_party/aom/aom_dsp/ssim.c @@ -0,0 +1,439 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include +#include + +#include "config/aom_dsp_rtcd.h" + +#include "aom_dsp/ssim.h" +#include "aom_ports/mem.h" +#include "aom_ports/system_state.h" + +void aom_ssim_parms_16x16_c(const uint8_t *s, int sp, const uint8_t *r, int rp, + uint32_t *sum_s, uint32_t *sum_r, + uint32_t *sum_sq_s, uint32_t *sum_sq_r, + uint32_t *sum_sxr) { + int i, j; + for (i = 0; i < 16; i++, s += sp, r += rp) { + for (j = 0; j < 16; j++) { + *sum_s += s[j]; + *sum_r += r[j]; + *sum_sq_s += s[j] * s[j]; + *sum_sq_r += r[j] * r[j]; + *sum_sxr += s[j] * r[j]; + } + } +} + +void aom_ssim_parms_8x8_c(const uint8_t *s, int sp, const uint8_t *r, int rp, + uint32_t *sum_s, uint32_t *sum_r, uint32_t *sum_sq_s, + uint32_t *sum_sq_r, uint32_t *sum_sxr) { + int i, j; + for (i = 0; i < 8; i++, s += sp, r += rp) { + for (j = 0; j < 8; j++) { + *sum_s += s[j]; + *sum_r += r[j]; + *sum_sq_s += s[j] * s[j]; + *sum_sq_r += r[j] * r[j]; + *sum_sxr += s[j] * r[j]; + } + } +} + +void aom_highbd_ssim_parms_8x8_c(const uint16_t *s, int sp, const uint16_t *r, + int rp, uint32_t *sum_s, uint32_t *sum_r, + uint32_t *sum_sq_s, uint32_t *sum_sq_r, + uint32_t *sum_sxr) { + int i, j; + for (i = 0; i < 8; i++, s += sp, r += rp) { + for (j = 0; j < 8; j++) { + *sum_s += s[j]; + *sum_r += r[j]; + *sum_sq_s += s[j] * s[j]; + *sum_sq_r += r[j] * r[j]; + *sum_sxr += s[j] * r[j]; + } + } +} + +static const int64_t cc1 = 26634; // (64^2*(.01*255)^2 +static const int64_t cc2 = 239708; // (64^2*(.03*255)^2 +static const int64_t cc1_10 = 428658; // (64^2*(.01*1023)^2 +static const int64_t cc2_10 = 3857925; // (64^2*(.03*1023)^2 +static const int64_t cc1_12 = 6868593; // (64^2*(.01*4095)^2 +static const int64_t cc2_12 = 61817334; // (64^2*(.03*4095)^2 + +static double similarity(uint32_t sum_s, uint32_t sum_r, uint32_t sum_sq_s, + uint32_t sum_sq_r, uint32_t sum_sxr, int count, + uint32_t bd) { + int64_t ssim_n, ssim_d; + int64_t c1, c2; + if (bd == 8) { + // scale the constants by number of pixels + c1 = (cc1 * count * count) >> 12; + c2 = (cc2 * count * count) >> 12; + } else if (bd == 10) { + c1 = (cc1_10 * count * count) >> 12; + c2 = (cc2_10 * count * count) >> 12; + } else if (bd == 12) { + c1 = (cc1_12 * count * count) >> 12; + c2 = (cc2_12 * count * count) >> 12; + } else { + c1 = c2 = 0; + assert(0); + } + + ssim_n = (2 * sum_s * sum_r + c1) * + ((int64_t)2 * count * sum_sxr - (int64_t)2 * sum_s * sum_r + c2); + + ssim_d = (sum_s * sum_s + sum_r * sum_r + c1) * + ((int64_t)count * sum_sq_s - (int64_t)sum_s * sum_s + + (int64_t)count * sum_sq_r - (int64_t)sum_r * sum_r + c2); + + return ssim_n * 1.0 / ssim_d; +} + +static double ssim_8x8(const uint8_t *s, int sp, const uint8_t *r, int rp) { + uint32_t sum_s = 0, sum_r = 0, sum_sq_s = 0, sum_sq_r = 0, sum_sxr = 0; + aom_ssim_parms_8x8(s, sp, r, rp, &sum_s, &sum_r, &sum_sq_s, &sum_sq_r, + &sum_sxr); + return similarity(sum_s, sum_r, sum_sq_s, sum_sq_r, sum_sxr, 64, 8); +} + +static double highbd_ssim_8x8(const uint16_t *s, int sp, const uint16_t *r, + int rp, uint32_t bd, uint32_t shift) { + uint32_t sum_s = 0, sum_r = 0, sum_sq_s = 0, sum_sq_r = 0, sum_sxr = 0; + aom_highbd_ssim_parms_8x8(s, sp, r, rp, &sum_s, &sum_r, &sum_sq_s, &sum_sq_r, + &sum_sxr); + return similarity(sum_s >> shift, sum_r >> shift, sum_sq_s >> (2 * shift), + sum_sq_r >> (2 * shift), sum_sxr >> (2 * shift), 64, bd); +} + +// We are using a 8x8 moving window with starting location of each 8x8 window +// on the 4x4 pixel grid. Such arrangement allows the windows to overlap +// block boundaries to penalize blocking artifacts. +static double aom_ssim2(const uint8_t *img1, const uint8_t *img2, + int stride_img1, int stride_img2, int width, + int height) { + int i, j; + int samples = 0; + double ssim_total = 0; + + // sample point start with each 4x4 location + for (i = 0; i <= height - 8; + i += 4, img1 += stride_img1 * 4, img2 += stride_img2 * 4) { + for (j = 0; j <= width - 8; j += 4) { + double v = ssim_8x8(img1 + j, stride_img1, img2 + j, stride_img2); + ssim_total += v; + samples++; + } + } + ssim_total /= samples; + return ssim_total; +} + +static double aom_highbd_ssim2(const uint8_t *img1, const uint8_t *img2, + int stride_img1, int stride_img2, int width, + int height, uint32_t bd, uint32_t shift) { + int i, j; + int samples = 0; + double ssim_total = 0; + + // sample point start with each 4x4 location + for (i = 0; i <= height - 8; + i += 4, img1 += stride_img1 * 4, img2 += stride_img2 * 4) { + for (j = 0; j <= width - 8; j += 4) { + double v = highbd_ssim_8x8(CONVERT_TO_SHORTPTR(img1 + j), stride_img1, + CONVERT_TO_SHORTPTR(img2 + j), stride_img2, bd, + shift); + ssim_total += v; + samples++; + } + } + ssim_total /= samples; + return ssim_total; +} + +double aom_calc_ssim(const YV12_BUFFER_CONFIG *source, + const YV12_BUFFER_CONFIG *dest, double *weight) { + double abc[3]; + for (int i = 0; i < 3; ++i) { + const int is_uv = i > 0; + abc[i] = aom_ssim2(source->buffers[i], dest->buffers[i], + source->strides[is_uv], dest->strides[is_uv], + source->crop_widths[is_uv], source->crop_heights[is_uv]); + } + + *weight = 1; + return abc[0] * .8 + .1 * (abc[1] + abc[2]); +} + +// traditional ssim as per: http://en.wikipedia.org/wiki/Structural_similarity +// +// Re working out the math -> +// +// ssim(x,y) = (2*mean(x)*mean(y) + c1)*(2*cov(x,y)+c2) / +// ((mean(x)^2+mean(y)^2+c1)*(var(x)+var(y)+c2)) +// +// mean(x) = sum(x) / n +// +// cov(x,y) = (n*sum(xi*yi)-sum(x)*sum(y))/(n*n) +// +// var(x) = (n*sum(xi*xi)-sum(xi)*sum(xi))/(n*n) +// +// ssim(x,y) = +// (2*sum(x)*sum(y)/(n*n) + c1)*(2*(n*sum(xi*yi)-sum(x)*sum(y))/(n*n)+c2) / +// (((sum(x)*sum(x)+sum(y)*sum(y))/(n*n) +c1) * +// ((n*sum(xi*xi) - sum(xi)*sum(xi))/(n*n)+ +// (n*sum(yi*yi) - sum(yi)*sum(yi))/(n*n)+c2))) +// +// factoring out n*n +// +// ssim(x,y) = +// (2*sum(x)*sum(y) + n*n*c1)*(2*(n*sum(xi*yi)-sum(x)*sum(y))+n*n*c2) / +// (((sum(x)*sum(x)+sum(y)*sum(y)) + n*n*c1) * +// (n*sum(xi*xi)-sum(xi)*sum(xi)+n*sum(yi*yi)-sum(yi)*sum(yi)+n*n*c2)) +// +// Replace c1 with n*n * c1 for the final step that leads to this code: +// The final step scales by 12 bits so we don't lose precision in the constants. + +static double ssimv_similarity(const Ssimv *sv, int64_t n) { + // Scale the constants by number of pixels. + const int64_t c1 = (cc1 * n * n) >> 12; + const int64_t c2 = (cc2 * n * n) >> 12; + + const double l = 1.0 * (2 * sv->sum_s * sv->sum_r + c1) / + (sv->sum_s * sv->sum_s + sv->sum_r * sv->sum_r + c1); + + // Since these variables are unsigned sums, convert to double so + // math is done in double arithmetic. + const double v = (2.0 * n * sv->sum_sxr - 2 * sv->sum_s * sv->sum_r + c2) / + (n * sv->sum_sq_s - sv->sum_s * sv->sum_s + + n * sv->sum_sq_r - sv->sum_r * sv->sum_r + c2); + + return l * v; +} + +// The first term of the ssim metric is a luminance factor. +// +// (2*mean(x)*mean(y) + c1)/ (mean(x)^2+mean(y)^2+c1) +// +// This luminance factor is super sensitive to the dark side of luminance +// values and completely insensitive on the white side. check out 2 sets +// (1,3) and (250,252) the term gives ( 2*1*3/(1+9) = .60 +// 2*250*252/ (250^2+252^2) => .99999997 +// +// As a result in this tweaked version of the calculation in which the +// luminance is taken as percentage off from peak possible. +// +// 255 * 255 - (sum_s - sum_r) / count * (sum_s - sum_r) / count +// +static double ssimv_similarity2(const Ssimv *sv, int64_t n) { + // Scale the constants by number of pixels. + const int64_t c1 = (cc1 * n * n) >> 12; + const int64_t c2 = (cc2 * n * n) >> 12; + + const double mean_diff = (1.0 * sv->sum_s - sv->sum_r) / n; + const double l = (255 * 255 - mean_diff * mean_diff + c1) / (255 * 255 + c1); + + // Since these variables are unsigned, sums convert to double so + // math is done in double arithmetic. + const double v = (2.0 * n * sv->sum_sxr - 2 * sv->sum_s * sv->sum_r + c2) / + (n * sv->sum_sq_s - sv->sum_s * sv->sum_s + + n * sv->sum_sq_r - sv->sum_r * sv->sum_r + c2); + + return l * v; +} +static void ssimv_parms(uint8_t *img1, int img1_pitch, uint8_t *img2, + int img2_pitch, Ssimv *sv) { + aom_ssim_parms_8x8(img1, img1_pitch, img2, img2_pitch, &sv->sum_s, &sv->sum_r, + &sv->sum_sq_s, &sv->sum_sq_r, &sv->sum_sxr); +} + +double aom_get_ssim_metrics(uint8_t *img1, int img1_pitch, uint8_t *img2, + int img2_pitch, int width, int height, Ssimv *sv2, + Metrics *m, int do_inconsistency) { + double dssim_total = 0; + double ssim_total = 0; + double ssim2_total = 0; + double inconsistency_total = 0; + int i, j; + int c = 0; + double norm; + double old_ssim_total = 0; + aom_clear_system_state(); + // We can sample points as frequently as we like start with 1 per 4x4. + for (i = 0; i < height; + i += 4, img1 += img1_pitch * 4, img2 += img2_pitch * 4) { + for (j = 0; j < width; j += 4, ++c) { + Ssimv sv = { 0, 0, 0, 0, 0, 0 }; + double ssim; + double ssim2; + double dssim; + uint32_t var_new; + uint32_t var_old; + uint32_t mean_new; + uint32_t mean_old; + double ssim_new; + double ssim_old; + + // Not sure there's a great way to handle the edge pixels + // in ssim when using a window. Seems biased against edge pixels + // however you handle this. This uses only samples that are + // fully in the frame. + if (j + 8 <= width && i + 8 <= height) { + ssimv_parms(img1 + j, img1_pitch, img2 + j, img2_pitch, &sv); + } + + ssim = ssimv_similarity(&sv, 64); + ssim2 = ssimv_similarity2(&sv, 64); + + sv.ssim = ssim2; + + // dssim is calculated to use as an actual error metric and + // is scaled up to the same range as sum square error. + // Since we are subsampling every 16th point maybe this should be + // *16 ? + dssim = 255 * 255 * (1 - ssim2) / 2; + + // Here I introduce a new error metric: consistency-weighted + // SSIM-inconsistency. This metric isolates frames where the + // SSIM 'suddenly' changes, e.g. if one frame in every 8 is much + // sharper or blurrier than the others. Higher values indicate a + // temporally inconsistent SSIM. There are two ideas at work: + // + // 1) 'SSIM-inconsistency': the total inconsistency value + // reflects how much SSIM values are changing between this + // source / reference frame pair and the previous pair. + // + // 2) 'consistency-weighted': weights de-emphasize areas in the + // frame where the scene content has changed. Changes in scene + // content are detected via changes in local variance and local + // mean. + // + // Thus the overall measure reflects how inconsistent the SSIM + // values are, over consistent regions of the frame. + // + // The metric has three terms: + // + // term 1 -> uses change in scene Variance to weight error score + // 2 * var(Fi)*var(Fi-1) / (var(Fi)^2+var(Fi-1)^2) + // larger changes from one frame to the next mean we care + // less about consistency. + // + // term 2 -> uses change in local scene luminance to weight error + // 2 * avg(Fi)*avg(Fi-1) / (avg(Fi)^2+avg(Fi-1)^2) + // larger changes from one frame to the next mean we care + // less about consistency. + // + // term3 -> measures inconsistency in ssim scores between frames + // 1 - ( 2 * ssim(Fi)*ssim(Fi-1)/(ssim(Fi)^2+sssim(Fi-1)^2). + // + // This term compares the ssim score for the same location in 2 + // subsequent frames. + var_new = sv.sum_sq_s - sv.sum_s * sv.sum_s / 64; + var_old = sv2[c].sum_sq_s - sv2[c].sum_s * sv2[c].sum_s / 64; + mean_new = sv.sum_s; + mean_old = sv2[c].sum_s; + ssim_new = sv.ssim; + ssim_old = sv2[c].ssim; + + if (do_inconsistency) { + // We do the metric once for every 4x4 block in the image. Since + // we are scaling the error to SSE for use in a psnr calculation + // 1.0 = 4x4x255x255 the worst error we can possibly have. + static const double kScaling = 4. * 4 * 255 * 255; + + // The constants have to be non 0 to avoid potential divide by 0 + // issues other than that they affect kind of a weighting between + // the terms. No testing of what the right terms should be has been + // done. + static const double c1 = 1, c2 = 1, c3 = 1; + + // This measures how much consistent variance is in two consecutive + // source frames. 1.0 means they have exactly the same variance. + const double variance_term = + (2.0 * var_old * var_new + c1) / + (1.0 * var_old * var_old + 1.0 * var_new * var_new + c1); + + // This measures how consistent the local mean are between two + // consecutive frames. 1.0 means they have exactly the same mean. + const double mean_term = + (2.0 * mean_old * mean_new + c2) / + (1.0 * mean_old * mean_old + 1.0 * mean_new * mean_new + c2); + + // This measures how consistent the ssims of two + // consecutive frames is. 1.0 means they are exactly the same. + double ssim_term = + pow((2.0 * ssim_old * ssim_new + c3) / + (ssim_old * ssim_old + ssim_new * ssim_new + c3), + 5); + + double this_inconsistency; + + // Floating point math sometimes makes this > 1 by a tiny bit. + // We want the metric to scale between 0 and 1.0 so we can convert + // it to an snr scaled value. + if (ssim_term > 1) ssim_term = 1; + + // This converts the consistency metric to an inconsistency metric + // ( so we can scale it like psnr to something like sum square error. + // The reason for the variance and mean terms is the assumption that + // if there are big changes in the source we shouldn't penalize + // inconsistency in ssim scores a bit less as it will be less visible + // to the user. + this_inconsistency = (1 - ssim_term) * variance_term * mean_term; + + this_inconsistency *= kScaling; + inconsistency_total += this_inconsistency; + } + sv2[c] = sv; + ssim_total += ssim; + ssim2_total += ssim2; + dssim_total += dssim; + + old_ssim_total += ssim_old; + } + old_ssim_total += 0; + } + + norm = 1. / (width / 4) / (height / 4); + ssim_total *= norm; + ssim2_total *= norm; + m->ssim2 = ssim2_total; + m->ssim = ssim_total; + if (old_ssim_total == 0) inconsistency_total = 0; + + m->ssimc = inconsistency_total; + + m->dssim = dssim_total; + return inconsistency_total; +} + +double aom_highbd_calc_ssim(const YV12_BUFFER_CONFIG *source, + const YV12_BUFFER_CONFIG *dest, double *weight, + uint32_t bd, uint32_t in_bd) { + assert(bd >= in_bd); + const uint32_t shift = bd - in_bd; + + double abc[3]; + for (int i = 0; i < 3; ++i) { + const int is_uv = i > 0; + abc[i] = aom_highbd_ssim2(source->buffers[i], dest->buffers[i], + source->strides[is_uv], dest->strides[is_uv], + source->crop_widths[is_uv], + source->crop_heights[is_uv], in_bd, shift); + } + + *weight = 1; + return abc[0] * .8 + .1 * (abc[1] + abc[2]); +} diff --git a/third_party/aom/aom_dsp/ssim.h b/third_party/aom/aom_dsp/ssim.h new file mode 100644 index 0000000000..55038f4c22 --- /dev/null +++ b/third_party/aom/aom_dsp/ssim.h @@ -0,0 +1,87 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_SSIM_H_ +#define AOM_AOM_DSP_SSIM_H_ + +#define MAX_SSIM_DB 100.0; + +#ifdef __cplusplus +extern "C" { +#endif + +#include "config/aom_config.h" + +#include "aom_scale/yv12config.h" + +// metrics used for calculating ssim, ssim2, dssim, and ssimc +typedef struct { + // source sum ( over 8x8 region ) + uint32_t sum_s; + + // reference sum (over 8x8 region ) + uint32_t sum_r; + + // source sum squared ( over 8x8 region ) + uint32_t sum_sq_s; + + // reference sum squared (over 8x8 region ) + uint32_t sum_sq_r; + + // sum of source times reference (over 8x8 region) + uint32_t sum_sxr; + + // calculated ssim score between source and reference + double ssim; +} Ssimv; + +// metrics collected on a frame basis +typedef struct { + // ssim consistency error metric ( see code for explanation ) + double ssimc; + + // standard ssim + double ssim; + + // revised ssim ( see code for explanation) + double ssim2; + + // ssim restated as an error metric like sse + double dssim; + + // dssim converted to decibels + double dssimd; + + // ssimc converted to decibels + double ssimcd; +} Metrics; + +double aom_get_ssim_metrics(uint8_t *img1, int img1_pitch, uint8_t *img2, + int img2_pitch, int width, int height, Ssimv *sv2, + Metrics *m, int do_inconsistency); + +double aom_calc_ssim(const YV12_BUFFER_CONFIG *source, + const YV12_BUFFER_CONFIG *dest, double *weight); + +double aom_calc_fastssim(const YV12_BUFFER_CONFIG *source, + const YV12_BUFFER_CONFIG *dest, double *ssim_y, + double *ssim_u, double *ssim_v, uint32_t bd, + uint32_t in_bd); + +double aom_highbd_calc_ssim(const YV12_BUFFER_CONFIG *source, + const YV12_BUFFER_CONFIG *dest, double *weight, + uint32_t bd, uint32_t in_bd); + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif // AOM_AOM_DSP_SSIM_H_ diff --git a/third_party/aom/aom_dsp/subtract.c b/third_party/aom/aom_dsp/subtract.c new file mode 100644 index 0000000000..2f6da96e59 --- /dev/null +++ b/third_party/aom/aom_dsp/subtract.c @@ -0,0 +1,53 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include + +#include "config/aom_config.h" +#include "config/aom_dsp_rtcd.h" + +#include "aom/aom_integer.h" +#include "aom_ports/mem.h" + +void aom_subtract_block_c(int rows, int cols, int16_t *diff, + ptrdiff_t diff_stride, const uint8_t *src, + ptrdiff_t src_stride, const uint8_t *pred, + ptrdiff_t pred_stride) { + int r, c; + + for (r = 0; r < rows; r++) { + for (c = 0; c < cols; c++) diff[c] = src[c] - pred[c]; + + diff += diff_stride; + pred += pred_stride; + src += src_stride; + } +} + +void aom_highbd_subtract_block_c(int rows, int cols, int16_t *diff, + ptrdiff_t diff_stride, const uint8_t *src8, + ptrdiff_t src_stride, const uint8_t *pred8, + ptrdiff_t pred_stride, int bd) { + int r, c; + uint16_t *src = CONVERT_TO_SHORTPTR(src8); + uint16_t *pred = CONVERT_TO_SHORTPTR(pred8); + (void)bd; + + for (r = 0; r < rows; r++) { + for (c = 0; c < cols; c++) { + diff[c] = src[c] - pred[c]; + } + + diff += diff_stride; + pred += pred_stride; + src += src_stride; + } +} diff --git a/third_party/aom/aom_dsp/sum_squares.c b/third_party/aom/aom_dsp/sum_squares.c new file mode 100644 index 0000000000..44ec41f2ed --- /dev/null +++ b/third_party/aom/aom_dsp/sum_squares.c @@ -0,0 +1,40 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include + +#include "config/aom_dsp_rtcd.h" + +uint64_t aom_sum_squares_2d_i16_c(const int16_t *src, int src_stride, int width, + int height) { + int r, c; + uint64_t ss = 0; + + for (r = 0; r < height; r++) { + for (c = 0; c < width; c++) { + const int16_t v = src[c]; + ss += v * v; + } + src += src_stride; + } + + return ss; +} + +uint64_t aom_sum_squares_i16_c(const int16_t *src, uint32_t n) { + uint64_t ss = 0; + do { + const int16_t v = *src++; + ss += v * v; + } while (--n); + + return ss; +} diff --git a/third_party/aom/aom_dsp/txfm_common.h b/third_party/aom/aom_dsp/txfm_common.h new file mode 100644 index 0000000000..f98242840c --- /dev/null +++ b/third_party/aom/aom_dsp/txfm_common.h @@ -0,0 +1,91 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_TXFM_COMMON_H_ +#define AOM_AOM_DSP_TXFM_COMMON_H_ + +#include "aom_dsp/aom_dsp_common.h" +#include "av1/common/enums.h" + +// Constants and Macros used by all idct/dct functions +#define DCT_CONST_BITS 14 +#define DCT_CONST_ROUNDING (1 << (DCT_CONST_BITS - 1)) + +#define UNIT_QUANT_SHIFT 2 +#define UNIT_QUANT_FACTOR (1 << UNIT_QUANT_SHIFT) + +typedef struct txfm_param { + // for both forward and inverse transforms + TX_TYPE tx_type; + TX_SIZE tx_size; + int lossless; + int bd; + // are the pixel buffers octets or shorts? This should collapse to + // bd==8 implies !is_hbd, but that's not certain right now. + int is_hbd; + TxSetType tx_set_type; + // for inverse transforms only + int eob; +} TxfmParam; + +// Constants: +// for (int i = 1; i< 32; ++i) +// printf("static const int cospi_%d_64 = %.0f;\n", i, +// round(16384 * cos(i*M_PI/64))); +// Note: sin(k*Pi/64) = cos((32-k)*Pi/64) +static const tran_high_t cospi_1_64 = 16364; +static const tran_high_t cospi_2_64 = 16305; +static const tran_high_t cospi_3_64 = 16207; +static const tran_high_t cospi_4_64 = 16069; +static const tran_high_t cospi_5_64 = 15893; +static const tran_high_t cospi_6_64 = 15679; +static const tran_high_t cospi_7_64 = 15426; +static const tran_high_t cospi_8_64 = 15137; +static const tran_high_t cospi_9_64 = 14811; +static const tran_high_t cospi_10_64 = 14449; +static const tran_high_t cospi_11_64 = 14053; +static const tran_high_t cospi_12_64 = 13623; +static const tran_high_t cospi_13_64 = 13160; +static const tran_high_t cospi_14_64 = 12665; +static const tran_high_t cospi_15_64 = 12140; +static const tran_high_t cospi_16_64 = 11585; +static const tran_high_t cospi_17_64 = 11003; +static const tran_high_t cospi_18_64 = 10394; +static const tran_high_t cospi_19_64 = 9760; +static const tran_high_t cospi_20_64 = 9102; +static const tran_high_t cospi_21_64 = 8423; +static const tran_high_t cospi_22_64 = 7723; +static const tran_high_t cospi_23_64 = 7005; +static const tran_high_t cospi_24_64 = 6270; +static const tran_high_t cospi_25_64 = 5520; +static const tran_high_t cospi_26_64 = 4756; +static const tran_high_t cospi_27_64 = 3981; +static const tran_high_t cospi_28_64 = 3196; +static const tran_high_t cospi_29_64 = 2404; +static const tran_high_t cospi_30_64 = 1606; +static const tran_high_t cospi_31_64 = 804; + +// 16384 * sqrt(2) * sin(kPi/9) * 2 / 3 +static const tran_high_t sinpi_1_9 = 5283; +static const tran_high_t sinpi_2_9 = 9929; +static const tran_high_t sinpi_3_9 = 13377; +static const tran_high_t sinpi_4_9 = 15212; + +// 16384 * sqrt(2) +static const tran_high_t Sqrt2 = 23170; +static const tran_high_t InvSqrt2 = 11585; + +static INLINE tran_high_t fdct_round_shift(tran_high_t input) { + tran_high_t rv = ROUND_POWER_OF_TWO(input, DCT_CONST_BITS); + return rv; +} + +#endif // AOM_AOM_DSP_TXFM_COMMON_H_ diff --git a/third_party/aom/aom_dsp/variance.c b/third_party/aom/aom_dsp/variance.c new file mode 100644 index 0000000000..23b7153098 --- /dev/null +++ b/third_party/aom/aom_dsp/variance.c @@ -0,0 +1,1579 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ +#include +#include +#include + +#include "config/aom_config.h" +#include "config/aom_dsp_rtcd.h" +#include "config/av1_rtcd.h" + +#include "aom/aom_integer.h" +#include "aom_ports/mem.h" + +#include "aom_dsp/aom_filter.h" +#include "aom_dsp/blend.h" +#include "aom_dsp/variance.h" + +#include "av1/common/filter.h" +#include "av1/common/onyxc_int.h" +#include "av1/common/reconinter.h" + +uint32_t aom_get4x4sse_cs_c(const uint8_t *a, int a_stride, const uint8_t *b, + int b_stride) { + int distortion = 0; + int r, c; + + for (r = 0; r < 4; ++r) { + for (c = 0; c < 4; ++c) { + int diff = a[c] - b[c]; + distortion += diff * diff; + } + + a += a_stride; + b += b_stride; + } + + return distortion; +} + +uint32_t aom_get_mb_ss_c(const int16_t *a) { + unsigned int i, sum = 0; + + for (i = 0; i < 256; ++i) { + sum += a[i] * a[i]; + } + + return sum; +} + +static void variance(const uint8_t *a, int a_stride, const uint8_t *b, + int b_stride, int w, int h, uint32_t *sse, int *sum) { + int i, j; + + *sum = 0; + *sse = 0; + + for (i = 0; i < h; ++i) { + for (j = 0; j < w; ++j) { + const int diff = a[j] - b[j]; + *sum += diff; + *sse += diff * diff; + } + + a += a_stride; + b += b_stride; + } +} + +uint32_t aom_sse_odd_size(const uint8_t *a, int a_stride, const uint8_t *b, + int b_stride, int w, int h) { + uint32_t sse; + int sum; + variance(a, a_stride, b, b_stride, w, h, &sse, &sum); + return sse; +} + +// Applies a 1-D 2-tap bilinear filter to the source block in either horizontal +// or vertical direction to produce the filtered output block. Used to implement +// the first-pass of 2-D separable filter. +// +// Produces int16_t output to retain precision for the next pass. Two filter +// taps should sum to FILTER_WEIGHT. pixel_step defines whether the filter is +// applied horizontally (pixel_step = 1) or vertically (pixel_step = stride). +// It defines the offset required to move from one input to the next. +void aom_var_filter_block2d_bil_first_pass_c(const uint8_t *a, uint16_t *b, + unsigned int src_pixels_per_line, + unsigned int pixel_step, + unsigned int output_height, + unsigned int output_width, + const uint8_t *filter) { + unsigned int i, j; + + for (i = 0; i < output_height; ++i) { + for (j = 0; j < output_width; ++j) { + b[j] = ROUND_POWER_OF_TWO( + (int)a[0] * filter[0] + (int)a[pixel_step] * filter[1], FILTER_BITS); + + ++a; + } + + a += src_pixels_per_line - output_width; + b += output_width; + } +} + +// Applies a 1-D 2-tap bilinear filter to the source block in either horizontal +// or vertical direction to produce the filtered output block. Used to implement +// the second-pass of 2-D separable filter. +// +// Requires 16-bit input as produced by filter_block2d_bil_first_pass. Two +// filter taps should sum to FILTER_WEIGHT. pixel_step defines whether the +// filter is applied horizontally (pixel_step = 1) or vertically +// (pixel_step = stride). It defines the offset required to move from one input +// to the next. Output is 8-bit. +void aom_var_filter_block2d_bil_second_pass_c(const uint16_t *a, uint8_t *b, + unsigned int src_pixels_per_line, + unsigned int pixel_step, + unsigned int output_height, + unsigned int output_width, + const uint8_t *filter) { + unsigned int i, j; + + for (i = 0; i < output_height; ++i) { + for (j = 0; j < output_width; ++j) { + b[j] = ROUND_POWER_OF_TWO( + (int)a[0] * filter[0] + (int)a[pixel_step] * filter[1], FILTER_BITS); + ++a; + } + + a += src_pixels_per_line - output_width; + b += output_width; + } +} + +#define VAR(W, H) \ + uint32_t aom_variance##W##x##H##_c(const uint8_t *a, int a_stride, \ + const uint8_t *b, int b_stride, \ + uint32_t *sse) { \ + int sum; \ + variance(a, a_stride, b, b_stride, W, H, sse, &sum); \ + return *sse - (uint32_t)(((int64_t)sum * sum) / (W * H)); \ + } + +#define SUBPIX_VAR(W, H) \ + uint32_t aom_sub_pixel_variance##W##x##H##_c( \ + const uint8_t *a, int a_stride, int xoffset, int yoffset, \ + const uint8_t *b, int b_stride, uint32_t *sse) { \ + uint16_t fdata3[(H + 1) * W]; \ + uint8_t temp2[H * W]; \ + \ + aom_var_filter_block2d_bil_first_pass_c(a, fdata3, a_stride, 1, H + 1, W, \ + bilinear_filters_2t[xoffset]); \ + aom_var_filter_block2d_bil_second_pass_c(fdata3, temp2, W, W, H, W, \ + bilinear_filters_2t[yoffset]); \ + \ + return aom_variance##W##x##H##_c(temp2, W, b, b_stride, sse); \ + } + +#define SUBPIX_AVG_VAR(W, H) \ + uint32_t aom_sub_pixel_avg_variance##W##x##H##_c( \ + const uint8_t *a, int a_stride, int xoffset, int yoffset, \ + const uint8_t *b, int b_stride, uint32_t *sse, \ + const uint8_t *second_pred) { \ + uint16_t fdata3[(H + 1) * W]; \ + uint8_t temp2[H * W]; \ + DECLARE_ALIGNED(16, uint8_t, temp3[H * W]); \ + \ + aom_var_filter_block2d_bil_first_pass_c(a, fdata3, a_stride, 1, H + 1, W, \ + bilinear_filters_2t[xoffset]); \ + aom_var_filter_block2d_bil_second_pass_c(fdata3, temp2, W, W, H, W, \ + bilinear_filters_2t[yoffset]); \ + \ + aom_comp_avg_pred(temp3, second_pred, W, H, temp2, W); \ + \ + return aom_variance##W##x##H##_c(temp3, W, b, b_stride, sse); \ + } \ + uint32_t aom_jnt_sub_pixel_avg_variance##W##x##H##_c( \ + const uint8_t *a, int a_stride, int xoffset, int yoffset, \ + const uint8_t *b, int b_stride, uint32_t *sse, \ + const uint8_t *second_pred, const JNT_COMP_PARAMS *jcp_param) { \ + uint16_t fdata3[(H + 1) * W]; \ + uint8_t temp2[H * W]; \ + DECLARE_ALIGNED(16, uint8_t, temp3[H * W]); \ + \ + aom_var_filter_block2d_bil_first_pass_c(a, fdata3, a_stride, 1, H + 1, W, \ + bilinear_filters_2t[xoffset]); \ + aom_var_filter_block2d_bil_second_pass_c(fdata3, temp2, W, W, H, W, \ + bilinear_filters_2t[yoffset]); \ + \ + aom_jnt_comp_avg_pred(temp3, second_pred, W, H, temp2, W, jcp_param); \ + \ + return aom_variance##W##x##H(temp3, W, b, b_stride, sse); \ + } + +/* Identical to the variance call except it takes an additional parameter, sum, + * and returns that value using pass-by-reference instead of returning + * sse - sum^2 / w*h + */ +#define GET_VAR(W, H) \ + void aom_get##W##x##H##var_c(const uint8_t *a, int a_stride, \ + const uint8_t *b, int b_stride, uint32_t *sse, \ + int *sum) { \ + variance(a, a_stride, b, b_stride, W, H, sse, sum); \ + } + +/* Identical to the variance call except it does not calculate the + * sse - sum^2 / w*h and returns sse in addtion to modifying the passed in + * variable. + */ +#define MSE(W, H) \ + uint32_t aom_mse##W##x##H##_c(const uint8_t *a, int a_stride, \ + const uint8_t *b, int b_stride, \ + uint32_t *sse) { \ + int sum; \ + variance(a, a_stride, b, b_stride, W, H, sse, &sum); \ + return *sse; \ + } + +/* All three forms of the variance are available in the same sizes. */ +#define VARIANCES(W, H) \ + VAR(W, H) \ + SUBPIX_VAR(W, H) \ + SUBPIX_AVG_VAR(W, H) + +VARIANCES(128, 128) +VARIANCES(128, 64) +VARIANCES(64, 128) +VARIANCES(64, 64) +VARIANCES(64, 32) +VARIANCES(32, 64) +VARIANCES(32, 32) +VARIANCES(32, 16) +VARIANCES(16, 32) +VARIANCES(16, 16) +VARIANCES(16, 8) +VARIANCES(8, 16) +VARIANCES(8, 8) +VARIANCES(8, 4) +VARIANCES(4, 8) +VARIANCES(4, 4) +VARIANCES(4, 2) +VARIANCES(2, 4) +VARIANCES(2, 2) +VARIANCES(4, 16) +VARIANCES(16, 4) +VARIANCES(8, 32) +VARIANCES(32, 8) +VARIANCES(16, 64) +VARIANCES(64, 16) + +GET_VAR(16, 16) +GET_VAR(8, 8) + +MSE(16, 16) +MSE(16, 8) +MSE(8, 16) +MSE(8, 8) + +void aom_comp_avg_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, + int height, const uint8_t *ref, int ref_stride) { + int i, j; + + for (i = 0; i < height; ++i) { + for (j = 0; j < width; ++j) { + const int tmp = pred[j] + ref[j]; + comp_pred[j] = ROUND_POWER_OF_TWO(tmp, 1); + } + comp_pred += width; + pred += width; + ref += ref_stride; + } +} + +// Get pred block from up-sampled reference. +void aom_upsampled_pred_c(MACROBLOCKD *xd, const AV1_COMMON *const cm, + int mi_row, int mi_col, const MV *const mv, + uint8_t *comp_pred, int width, int height, + int subpel_x_q3, int subpel_y_q3, const uint8_t *ref, + int ref_stride, int subpel_search) { + // expect xd == NULL only in tests + if (xd != NULL) { + const MB_MODE_INFO *mi = xd->mi[0]; + const int ref_num = 0; + const int is_intrabc = is_intrabc_block(mi); + const struct scale_factors *const sf = + is_intrabc ? &cm->sf_identity : &xd->block_refs[ref_num]->sf; + const int is_scaled = av1_is_scaled(sf); + + if (is_scaled) { + // Note: This is mostly a copy from the >=8X8 case in + // build_inter_predictors() function, with some small tweaks. + + // Some assumptions. + const int plane = 0; + + // Get pre-requisites. + const struct macroblockd_plane *const pd = &xd->plane[plane]; + const int ssx = pd->subsampling_x; + const int ssy = pd->subsampling_y; + assert(ssx == 0 && ssy == 0); + const struct buf_2d *const dst_buf = &pd->dst; + const struct buf_2d *const pre_buf = + is_intrabc ? dst_buf : &pd->pre[ref_num]; + const int mi_x = mi_col * MI_SIZE; + const int mi_y = mi_row * MI_SIZE; + + // Calculate subpel_x/y and x/y_step. + const int row_start = 0; // Because ss_y is 0. + const int col_start = 0; // Because ss_x is 0. + const int pre_x = (mi_x + MI_SIZE * col_start) >> ssx; + const int pre_y = (mi_y + MI_SIZE * row_start) >> ssy; + int orig_pos_y = pre_y << SUBPEL_BITS; + orig_pos_y += mv->row * (1 << (1 - ssy)); + int orig_pos_x = pre_x << SUBPEL_BITS; + orig_pos_x += mv->col * (1 << (1 - ssx)); + int pos_y = sf->scale_value_y(orig_pos_y, sf); + int pos_x = sf->scale_value_x(orig_pos_x, sf); + pos_x += SCALE_EXTRA_OFF; + pos_y += SCALE_EXTRA_OFF; + + const int top = -AOM_LEFT_TOP_MARGIN_SCALED(ssy); + const int left = -AOM_LEFT_TOP_MARGIN_SCALED(ssx); + const int bottom = (pre_buf->height + AOM_INTERP_EXTEND) + << SCALE_SUBPEL_BITS; + const int right = (pre_buf->width + AOM_INTERP_EXTEND) + << SCALE_SUBPEL_BITS; + pos_y = clamp(pos_y, top, bottom); + pos_x = clamp(pos_x, left, right); + + const uint8_t *const pre = + pre_buf->buf0 + (pos_y >> SCALE_SUBPEL_BITS) * pre_buf->stride + + (pos_x >> SCALE_SUBPEL_BITS); + + const SubpelParams subpel_params = { sf->x_step_q4, sf->y_step_q4, + pos_x & SCALE_SUBPEL_MASK, + pos_y & SCALE_SUBPEL_MASK }; + + // Get warp types. + const WarpedMotionParams *const wm = + &xd->global_motion[mi->ref_frame[ref_num]]; + const int is_global = is_global_mv_block(mi, wm->wmtype); + WarpTypesAllowed warp_types; + warp_types.global_warp_allowed = is_global; + warp_types.local_warp_allowed = mi->motion_mode == WARPED_CAUSAL; + + // Get convolve parameters. + ConvolveParams conv_params = get_conv_params(0, plane, xd->bd); + const InterpFilters filters = + av1_broadcast_interp_filter(EIGHTTAP_REGULAR); + + // Get the inter predictor. + const int build_for_obmc = 0; + av1_make_inter_predictor(pre, pre_buf->stride, comp_pred, width, + &subpel_params, sf, width, height, &conv_params, + filters, &warp_types, mi_x >> pd->subsampling_x, + mi_y >> pd->subsampling_y, plane, ref_num, mi, + build_for_obmc, xd, cm->allow_warped_motion); + + return; + } + } + + const InterpFilterParams *filter = + (subpel_search == 1) + ? av1_get_4tap_interp_filter_params(EIGHTTAP_REGULAR) + : av1_get_interp_filter_params_with_block_size(EIGHTTAP_REGULAR, 8); + + if (!subpel_x_q3 && !subpel_y_q3) { + for (int i = 0; i < height; i++) { + memcpy(comp_pred, ref, width * sizeof(*comp_pred)); + comp_pred += width; + ref += ref_stride; + } + } else if (!subpel_y_q3) { + const int16_t *const kernel = + av1_get_interp_filter_subpel_kernel(filter, subpel_x_q3 << 1); + aom_convolve8_horiz_c(ref, ref_stride, comp_pred, width, kernel, 16, NULL, + -1, width, height); + } else if (!subpel_x_q3) { + const int16_t *const kernel = + av1_get_interp_filter_subpel_kernel(filter, subpel_y_q3 << 1); + aom_convolve8_vert_c(ref, ref_stride, comp_pred, width, NULL, -1, kernel, + 16, width, height); + } else { + DECLARE_ALIGNED(16, uint8_t, + temp[((MAX_SB_SIZE * 2 + 16) + 16) * MAX_SB_SIZE]); + const int16_t *const kernel_x = + av1_get_interp_filter_subpel_kernel(filter, subpel_x_q3 << 1); + const int16_t *const kernel_y = + av1_get_interp_filter_subpel_kernel(filter, subpel_y_q3 << 1); + const int intermediate_height = + (((height - 1) * 8 + subpel_y_q3) >> 3) + filter->taps; + assert(intermediate_height <= (MAX_SB_SIZE * 2 + 16) + 16); + aom_convolve8_horiz_c(ref - ref_stride * ((filter->taps >> 1) - 1), + ref_stride, temp, MAX_SB_SIZE, kernel_x, 16, NULL, -1, + width, intermediate_height); + aom_convolve8_vert_c(temp + MAX_SB_SIZE * ((filter->taps >> 1) - 1), + MAX_SB_SIZE, comp_pred, width, NULL, -1, kernel_y, 16, + width, height); + } +} + +void aom_comp_avg_upsampled_pred_c(MACROBLOCKD *xd, const AV1_COMMON *const cm, + int mi_row, int mi_col, const MV *const mv, + uint8_t *comp_pred, const uint8_t *pred, + int width, int height, int subpel_x_q3, + int subpel_y_q3, const uint8_t *ref, + int ref_stride, int subpel_search) { + int i, j; + + aom_upsampled_pred(xd, cm, mi_row, mi_col, mv, comp_pred, width, height, + subpel_x_q3, subpel_y_q3, ref, ref_stride, subpel_search); + for (i = 0; i < height; i++) { + for (j = 0; j < width; j++) { + comp_pred[j] = ROUND_POWER_OF_TWO(comp_pred[j] + pred[j], 1); + } + comp_pred += width; + pred += width; + } +} + +void aom_jnt_comp_avg_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, + int height, const uint8_t *ref, int ref_stride, + const JNT_COMP_PARAMS *jcp_param) { + int i, j; + const int fwd_offset = jcp_param->fwd_offset; + const int bck_offset = jcp_param->bck_offset; + + for (i = 0; i < height; ++i) { + for (j = 0; j < width; ++j) { + int tmp = pred[j] * bck_offset + ref[j] * fwd_offset; + tmp = ROUND_POWER_OF_TWO(tmp, DIST_PRECISION_BITS); + comp_pred[j] = (uint8_t)tmp; + } + comp_pred += width; + pred += width; + ref += ref_stride; + } +} + +void aom_jnt_comp_avg_upsampled_pred_c( + MACROBLOCKD *xd, const AV1_COMMON *const cm, int mi_row, int mi_col, + const MV *const mv, uint8_t *comp_pred, const uint8_t *pred, int width, + int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref, + int ref_stride, const JNT_COMP_PARAMS *jcp_param, int subpel_search) { + int i, j; + const int fwd_offset = jcp_param->fwd_offset; + const int bck_offset = jcp_param->bck_offset; + + aom_upsampled_pred_c(xd, cm, mi_row, mi_col, mv, comp_pred, width, height, + subpel_x_q3, subpel_y_q3, ref, ref_stride, + subpel_search); + + for (i = 0; i < height; i++) { + for (j = 0; j < width; j++) { + int tmp = pred[j] * bck_offset + comp_pred[j] * fwd_offset; + tmp = ROUND_POWER_OF_TWO(tmp, DIST_PRECISION_BITS); + comp_pred[j] = (uint8_t)tmp; + } + comp_pred += width; + pred += width; + } +} + +static void highbd_variance64(const uint8_t *a8, int a_stride, + const uint8_t *b8, int b_stride, int w, int h, + uint64_t *sse, int64_t *sum) { + const uint16_t *a = CONVERT_TO_SHORTPTR(a8); + const uint16_t *b = CONVERT_TO_SHORTPTR(b8); + int64_t tsum = 0; + uint64_t tsse = 0; + for (int i = 0; i < h; ++i) { + int32_t lsum = 0; + for (int j = 0; j < w; ++j) { + const int diff = a[j] - b[j]; + lsum += diff; + tsse += (uint32_t)(diff * diff); + } + tsum += lsum; + a += a_stride; + b += b_stride; + } + *sum = tsum; + *sse = tsse; +} + +uint64_t aom_highbd_sse_odd_size(const uint8_t *a, int a_stride, + const uint8_t *b, int b_stride, int w, int h) { + uint64_t sse; + int64_t sum; + highbd_variance64(a, a_stride, b, b_stride, w, h, &sse, &sum); + return sse; +} + +static void highbd_8_variance(const uint8_t *a8, int a_stride, + const uint8_t *b8, int b_stride, int w, int h, + uint32_t *sse, int *sum) { + uint64_t sse_long = 0; + int64_t sum_long = 0; + highbd_variance64(a8, a_stride, b8, b_stride, w, h, &sse_long, &sum_long); + *sse = (uint32_t)sse_long; + *sum = (int)sum_long; +} + +static void highbd_10_variance(const uint8_t *a8, int a_stride, + const uint8_t *b8, int b_stride, int w, int h, + uint32_t *sse, int *sum) { + uint64_t sse_long = 0; + int64_t sum_long = 0; + highbd_variance64(a8, a_stride, b8, b_stride, w, h, &sse_long, &sum_long); + *sse = (uint32_t)ROUND_POWER_OF_TWO(sse_long, 4); + *sum = (int)ROUND_POWER_OF_TWO(sum_long, 2); +} + +static void highbd_12_variance(const uint8_t *a8, int a_stride, + const uint8_t *b8, int b_stride, int w, int h, + uint32_t *sse, int *sum) { + uint64_t sse_long = 0; + int64_t sum_long = 0; + highbd_variance64(a8, a_stride, b8, b_stride, w, h, &sse_long, &sum_long); + *sse = (uint32_t)ROUND_POWER_OF_TWO(sse_long, 8); + *sum = (int)ROUND_POWER_OF_TWO(sum_long, 4); +} + +#define HIGHBD_VAR(W, H) \ + uint32_t aom_highbd_8_variance##W##x##H##_c(const uint8_t *a, int a_stride, \ + const uint8_t *b, int b_stride, \ + uint32_t *sse) { \ + int sum; \ + highbd_8_variance(a, a_stride, b, b_stride, W, H, sse, &sum); \ + return *sse - (uint32_t)(((int64_t)sum * sum) / (W * H)); \ + } \ + \ + uint32_t aom_highbd_10_variance##W##x##H##_c(const uint8_t *a, int a_stride, \ + const uint8_t *b, int b_stride, \ + uint32_t *sse) { \ + int sum; \ + int64_t var; \ + highbd_10_variance(a, a_stride, b, b_stride, W, H, sse, &sum); \ + var = (int64_t)(*sse) - (((int64_t)sum * sum) / (W * H)); \ + return (var >= 0) ? (uint32_t)var : 0; \ + } \ + \ + uint32_t aom_highbd_12_variance##W##x##H##_c(const uint8_t *a, int a_stride, \ + const uint8_t *b, int b_stride, \ + uint32_t *sse) { \ + int sum; \ + int64_t var; \ + highbd_12_variance(a, a_stride, b, b_stride, W, H, sse, &sum); \ + var = (int64_t)(*sse) - (((int64_t)sum * sum) / (W * H)); \ + return (var >= 0) ? (uint32_t)var : 0; \ + } + +#define HIGHBD_GET_VAR(S) \ + void aom_highbd_8_get##S##x##S##var_c(const uint8_t *src, int src_stride, \ + const uint8_t *ref, int ref_stride, \ + uint32_t *sse, int *sum) { \ + highbd_8_variance(src, src_stride, ref, ref_stride, S, S, sse, sum); \ + } \ + \ + void aom_highbd_10_get##S##x##S##var_c(const uint8_t *src, int src_stride, \ + const uint8_t *ref, int ref_stride, \ + uint32_t *sse, int *sum) { \ + highbd_10_variance(src, src_stride, ref, ref_stride, S, S, sse, sum); \ + } \ + \ + void aom_highbd_12_get##S##x##S##var_c(const uint8_t *src, int src_stride, \ + const uint8_t *ref, int ref_stride, \ + uint32_t *sse, int *sum) { \ + highbd_12_variance(src, src_stride, ref, ref_stride, S, S, sse, sum); \ + } + +#define HIGHBD_MSE(W, H) \ + uint32_t aom_highbd_8_mse##W##x##H##_c(const uint8_t *src, int src_stride, \ + const uint8_t *ref, int ref_stride, \ + uint32_t *sse) { \ + int sum; \ + highbd_8_variance(src, src_stride, ref, ref_stride, W, H, sse, &sum); \ + return *sse; \ + } \ + \ + uint32_t aom_highbd_10_mse##W##x##H##_c(const uint8_t *src, int src_stride, \ + const uint8_t *ref, int ref_stride, \ + uint32_t *sse) { \ + int sum; \ + highbd_10_variance(src, src_stride, ref, ref_stride, W, H, sse, &sum); \ + return *sse; \ + } \ + \ + uint32_t aom_highbd_12_mse##W##x##H##_c(const uint8_t *src, int src_stride, \ + const uint8_t *ref, int ref_stride, \ + uint32_t *sse) { \ + int sum; \ + highbd_12_variance(src, src_stride, ref, ref_stride, W, H, sse, &sum); \ + return *sse; \ + } + +void aom_highbd_var_filter_block2d_bil_first_pass( + const uint8_t *src_ptr8, uint16_t *output_ptr, + unsigned int src_pixels_per_line, int pixel_step, + unsigned int output_height, unsigned int output_width, + const uint8_t *filter) { + unsigned int i, j; + uint16_t *src_ptr = CONVERT_TO_SHORTPTR(src_ptr8); + for (i = 0; i < output_height; ++i) { + for (j = 0; j < output_width; ++j) { + output_ptr[j] = ROUND_POWER_OF_TWO( + (int)src_ptr[0] * filter[0] + (int)src_ptr[pixel_step] * filter[1], + FILTER_BITS); + + ++src_ptr; + } + + // Next row... + src_ptr += src_pixels_per_line - output_width; + output_ptr += output_width; + } +} + +void aom_highbd_var_filter_block2d_bil_second_pass( + const uint16_t *src_ptr, uint16_t *output_ptr, + unsigned int src_pixels_per_line, unsigned int pixel_step, + unsigned int output_height, unsigned int output_width, + const uint8_t *filter) { + unsigned int i, j; + + for (i = 0; i < output_height; ++i) { + for (j = 0; j < output_width; ++j) { + output_ptr[j] = ROUND_POWER_OF_TWO( + (int)src_ptr[0] * filter[0] + (int)src_ptr[pixel_step] * filter[1], + FILTER_BITS); + ++src_ptr; + } + + src_ptr += src_pixels_per_line - output_width; + output_ptr += output_width; + } +} + +#define HIGHBD_SUBPIX_VAR(W, H) \ + uint32_t aom_highbd_8_sub_pixel_variance##W##x##H##_c( \ + const uint8_t *src, int src_stride, int xoffset, int yoffset, \ + const uint8_t *dst, int dst_stride, uint32_t *sse) { \ + uint16_t fdata3[(H + 1) * W]; \ + uint16_t temp2[H * W]; \ + \ + aom_highbd_var_filter_block2d_bil_first_pass( \ + src, fdata3, src_stride, 1, H + 1, W, bilinear_filters_2t[xoffset]); \ + aom_highbd_var_filter_block2d_bil_second_pass( \ + fdata3, temp2, W, W, H, W, bilinear_filters_2t[yoffset]); \ + \ + return aom_highbd_8_variance##W##x##H##_c(CONVERT_TO_BYTEPTR(temp2), W, \ + dst, dst_stride, sse); \ + } \ + \ + uint32_t aom_highbd_10_sub_pixel_variance##W##x##H##_c( \ + const uint8_t *src, int src_stride, int xoffset, int yoffset, \ + const uint8_t *dst, int dst_stride, uint32_t *sse) { \ + uint16_t fdata3[(H + 1) * W]; \ + uint16_t temp2[H * W]; \ + \ + aom_highbd_var_filter_block2d_bil_first_pass( \ + src, fdata3, src_stride, 1, H + 1, W, bilinear_filters_2t[xoffset]); \ + aom_highbd_var_filter_block2d_bil_second_pass( \ + fdata3, temp2, W, W, H, W, bilinear_filters_2t[yoffset]); \ + \ + return aom_highbd_10_variance##W##x##H##_c(CONVERT_TO_BYTEPTR(temp2), W, \ + dst, dst_stride, sse); \ + } \ + \ + uint32_t aom_highbd_12_sub_pixel_variance##W##x##H##_c( \ + const uint8_t *src, int src_stride, int xoffset, int yoffset, \ + const uint8_t *dst, int dst_stride, uint32_t *sse) { \ + uint16_t fdata3[(H + 1) * W]; \ + uint16_t temp2[H * W]; \ + \ + aom_highbd_var_filter_block2d_bil_first_pass( \ + src, fdata3, src_stride, 1, H + 1, W, bilinear_filters_2t[xoffset]); \ + aom_highbd_var_filter_block2d_bil_second_pass( \ + fdata3, temp2, W, W, H, W, bilinear_filters_2t[yoffset]); \ + \ + return aom_highbd_12_variance##W##x##H##_c(CONVERT_TO_BYTEPTR(temp2), W, \ + dst, dst_stride, sse); \ + } + +#define HIGHBD_SUBPIX_AVG_VAR(W, H) \ + uint32_t aom_highbd_8_sub_pixel_avg_variance##W##x##H##_c( \ + const uint8_t *src, int src_stride, int xoffset, int yoffset, \ + const uint8_t *dst, int dst_stride, uint32_t *sse, \ + const uint8_t *second_pred) { \ + uint16_t fdata3[(H + 1) * W]; \ + uint16_t temp2[H * W]; \ + DECLARE_ALIGNED(16, uint16_t, temp3[H * W]); \ + \ + aom_highbd_var_filter_block2d_bil_first_pass( \ + src, fdata3, src_stride, 1, H + 1, W, bilinear_filters_2t[xoffset]); \ + aom_highbd_var_filter_block2d_bil_second_pass( \ + fdata3, temp2, W, W, H, W, bilinear_filters_2t[yoffset]); \ + \ + aom_highbd_comp_avg_pred_c(CONVERT_TO_BYTEPTR(temp3), second_pred, W, H, \ + CONVERT_TO_BYTEPTR(temp2), W); \ + \ + return aom_highbd_8_variance##W##x##H##_c(CONVERT_TO_BYTEPTR(temp3), W, \ + dst, dst_stride, sse); \ + } \ + \ + uint32_t aom_highbd_10_sub_pixel_avg_variance##W##x##H##_c( \ + const uint8_t *src, int src_stride, int xoffset, int yoffset, \ + const uint8_t *dst, int dst_stride, uint32_t *sse, \ + const uint8_t *second_pred) { \ + uint16_t fdata3[(H + 1) * W]; \ + uint16_t temp2[H * W]; \ + DECLARE_ALIGNED(16, uint16_t, temp3[H * W]); \ + \ + aom_highbd_var_filter_block2d_bil_first_pass( \ + src, fdata3, src_stride, 1, H + 1, W, bilinear_filters_2t[xoffset]); \ + aom_highbd_var_filter_block2d_bil_second_pass( \ + fdata3, temp2, W, W, H, W, bilinear_filters_2t[yoffset]); \ + \ + aom_highbd_comp_avg_pred_c(CONVERT_TO_BYTEPTR(temp3), second_pred, W, H, \ + CONVERT_TO_BYTEPTR(temp2), W); \ + \ + return aom_highbd_10_variance##W##x##H##_c(CONVERT_TO_BYTEPTR(temp3), W, \ + dst, dst_stride, sse); \ + } \ + \ + uint32_t aom_highbd_12_sub_pixel_avg_variance##W##x##H##_c( \ + const uint8_t *src, int src_stride, int xoffset, int yoffset, \ + const uint8_t *dst, int dst_stride, uint32_t *sse, \ + const uint8_t *second_pred) { \ + uint16_t fdata3[(H + 1) * W]; \ + uint16_t temp2[H * W]; \ + DECLARE_ALIGNED(16, uint16_t, temp3[H * W]); \ + \ + aom_highbd_var_filter_block2d_bil_first_pass( \ + src, fdata3, src_stride, 1, H + 1, W, bilinear_filters_2t[xoffset]); \ + aom_highbd_var_filter_block2d_bil_second_pass( \ + fdata3, temp2, W, W, H, W, bilinear_filters_2t[yoffset]); \ + \ + aom_highbd_comp_avg_pred_c(CONVERT_TO_BYTEPTR(temp3), second_pred, W, H, \ + CONVERT_TO_BYTEPTR(temp2), W); \ + \ + return aom_highbd_12_variance##W##x##H##_c(CONVERT_TO_BYTEPTR(temp3), W, \ + dst, dst_stride, sse); \ + } \ + \ + uint32_t aom_highbd_8_jnt_sub_pixel_avg_variance##W##x##H##_c( \ + const uint8_t *src, int src_stride, int xoffset, int yoffset, \ + const uint8_t *dst, int dst_stride, uint32_t *sse, \ + const uint8_t *second_pred, const JNT_COMP_PARAMS *jcp_param) { \ + uint16_t fdata3[(H + 1) * W]; \ + uint16_t temp2[H * W]; \ + DECLARE_ALIGNED(16, uint16_t, temp3[H * W]); \ + \ + aom_highbd_var_filter_block2d_bil_first_pass( \ + src, fdata3, src_stride, 1, H + 1, W, bilinear_filters_2t[xoffset]); \ + aom_highbd_var_filter_block2d_bil_second_pass( \ + fdata3, temp2, W, W, H, W, bilinear_filters_2t[yoffset]); \ + \ + aom_highbd_jnt_comp_avg_pred(CONVERT_TO_BYTEPTR(temp3), second_pred, W, H, \ + CONVERT_TO_BYTEPTR(temp2), W, jcp_param); \ + \ + return aom_highbd_8_variance##W##x##H(CONVERT_TO_BYTEPTR(temp3), W, dst, \ + dst_stride, sse); \ + } \ + \ + uint32_t aom_highbd_10_jnt_sub_pixel_avg_variance##W##x##H##_c( \ + const uint8_t *src, int src_stride, int xoffset, int yoffset, \ + const uint8_t *dst, int dst_stride, uint32_t *sse, \ + const uint8_t *second_pred, const JNT_COMP_PARAMS *jcp_param) { \ + uint16_t fdata3[(H + 1) * W]; \ + uint16_t temp2[H * W]; \ + DECLARE_ALIGNED(16, uint16_t, temp3[H * W]); \ + \ + aom_highbd_var_filter_block2d_bil_first_pass( \ + src, fdata3, src_stride, 1, H + 1, W, bilinear_filters_2t[xoffset]); \ + aom_highbd_var_filter_block2d_bil_second_pass( \ + fdata3, temp2, W, W, H, W, bilinear_filters_2t[yoffset]); \ + \ + aom_highbd_jnt_comp_avg_pred(CONVERT_TO_BYTEPTR(temp3), second_pred, W, H, \ + CONVERT_TO_BYTEPTR(temp2), W, jcp_param); \ + \ + return aom_highbd_10_variance##W##x##H(CONVERT_TO_BYTEPTR(temp3), W, dst, \ + dst_stride, sse); \ + } \ + \ + uint32_t aom_highbd_12_jnt_sub_pixel_avg_variance##W##x##H##_c( \ + const uint8_t *src, int src_stride, int xoffset, int yoffset, \ + const uint8_t *dst, int dst_stride, uint32_t *sse, \ + const uint8_t *second_pred, const JNT_COMP_PARAMS *jcp_param) { \ + uint16_t fdata3[(H + 1) * W]; \ + uint16_t temp2[H * W]; \ + DECLARE_ALIGNED(16, uint16_t, temp3[H * W]); \ + \ + aom_highbd_var_filter_block2d_bil_first_pass( \ + src, fdata3, src_stride, 1, H + 1, W, bilinear_filters_2t[xoffset]); \ + aom_highbd_var_filter_block2d_bil_second_pass( \ + fdata3, temp2, W, W, H, W, bilinear_filters_2t[yoffset]); \ + \ + aom_highbd_jnt_comp_avg_pred(CONVERT_TO_BYTEPTR(temp3), second_pred, W, H, \ + CONVERT_TO_BYTEPTR(temp2), W, jcp_param); \ + \ + return aom_highbd_12_variance##W##x##H(CONVERT_TO_BYTEPTR(temp3), W, dst, \ + dst_stride, sse); \ + } + +/* All three forms of the variance are available in the same sizes. */ +#define HIGHBD_VARIANCES(W, H) \ + HIGHBD_VAR(W, H) \ + HIGHBD_SUBPIX_VAR(W, H) \ + HIGHBD_SUBPIX_AVG_VAR(W, H) + +HIGHBD_VARIANCES(128, 128) +HIGHBD_VARIANCES(128, 64) +HIGHBD_VARIANCES(64, 128) +HIGHBD_VARIANCES(64, 64) +HIGHBD_VARIANCES(64, 32) +HIGHBD_VARIANCES(32, 64) +HIGHBD_VARIANCES(32, 32) +HIGHBD_VARIANCES(32, 16) +HIGHBD_VARIANCES(16, 32) +HIGHBD_VARIANCES(16, 16) +HIGHBD_VARIANCES(16, 8) +HIGHBD_VARIANCES(8, 16) +HIGHBD_VARIANCES(8, 8) +HIGHBD_VARIANCES(8, 4) +HIGHBD_VARIANCES(4, 8) +HIGHBD_VARIANCES(4, 4) +HIGHBD_VARIANCES(4, 2) +HIGHBD_VARIANCES(2, 4) +HIGHBD_VARIANCES(2, 2) +HIGHBD_VARIANCES(4, 16) +HIGHBD_VARIANCES(16, 4) +HIGHBD_VARIANCES(8, 32) +HIGHBD_VARIANCES(32, 8) +HIGHBD_VARIANCES(16, 64) +HIGHBD_VARIANCES(64, 16) + +HIGHBD_GET_VAR(8) +HIGHBD_GET_VAR(16) + +HIGHBD_MSE(16, 16) +HIGHBD_MSE(16, 8) +HIGHBD_MSE(8, 16) +HIGHBD_MSE(8, 8) + +void aom_highbd_comp_avg_pred_c(uint8_t *comp_pred8, const uint8_t *pred8, + int width, int height, const uint8_t *ref8, + int ref_stride) { + int i, j; + uint16_t *pred = CONVERT_TO_SHORTPTR(pred8); + uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); + uint16_t *comp_pred = CONVERT_TO_SHORTPTR(comp_pred8); + for (i = 0; i < height; ++i) { + for (j = 0; j < width; ++j) { + const int tmp = pred[j] + ref[j]; + comp_pred[j] = ROUND_POWER_OF_TWO(tmp, 1); + } + comp_pred += width; + pred += width; + ref += ref_stride; + } +} + +void aom_highbd_upsampled_pred_c(MACROBLOCKD *xd, + const struct AV1Common *const cm, int mi_row, + int mi_col, const MV *const mv, + uint8_t *comp_pred8, int width, int height, + int subpel_x_q3, int subpel_y_q3, + const uint8_t *ref8, int ref_stride, int bd, + int subpel_search) { + // expect xd == NULL only in tests + if (xd != NULL) { + const MB_MODE_INFO *mi = xd->mi[0]; + const int ref_num = 0; + const int is_intrabc = is_intrabc_block(mi); + const struct scale_factors *const sf = + is_intrabc ? &cm->sf_identity : &xd->block_refs[ref_num]->sf; + const int is_scaled = av1_is_scaled(sf); + + if (is_scaled) { + // Note: This is mostly a copy from the >=8X8 case in + // build_inter_predictors() function, with some small tweaks. + // Some assumptions. + const int plane = 0; + + // Get pre-requisites. + const struct macroblockd_plane *const pd = &xd->plane[plane]; + const int ssx = pd->subsampling_x; + const int ssy = pd->subsampling_y; + assert(ssx == 0 && ssy == 0); + const struct buf_2d *const dst_buf = &pd->dst; + const struct buf_2d *const pre_buf = + is_intrabc ? dst_buf : &pd->pre[ref_num]; + const int mi_x = mi_col * MI_SIZE; + const int mi_y = mi_row * MI_SIZE; + + // Calculate subpel_x/y and x/y_step. + const int row_start = 0; // Because ss_y is 0. + const int col_start = 0; // Because ss_x is 0. + const int pre_x = (mi_x + MI_SIZE * col_start) >> ssx; + const int pre_y = (mi_y + MI_SIZE * row_start) >> ssy; + int orig_pos_y = pre_y << SUBPEL_BITS; + orig_pos_y += mv->row * (1 << (1 - ssy)); + int orig_pos_x = pre_x << SUBPEL_BITS; + orig_pos_x += mv->col * (1 << (1 - ssx)); + int pos_y = sf->scale_value_y(orig_pos_y, sf); + int pos_x = sf->scale_value_x(orig_pos_x, sf); + pos_x += SCALE_EXTRA_OFF; + pos_y += SCALE_EXTRA_OFF; + + const int top = -AOM_LEFT_TOP_MARGIN_SCALED(ssy); + const int left = -AOM_LEFT_TOP_MARGIN_SCALED(ssx); + const int bottom = (pre_buf->height + AOM_INTERP_EXTEND) + << SCALE_SUBPEL_BITS; + const int right = (pre_buf->width + AOM_INTERP_EXTEND) + << SCALE_SUBPEL_BITS; + pos_y = clamp(pos_y, top, bottom); + pos_x = clamp(pos_x, left, right); + + const uint8_t *const pre = + pre_buf->buf0 + (pos_y >> SCALE_SUBPEL_BITS) * pre_buf->stride + + (pos_x >> SCALE_SUBPEL_BITS); + + const SubpelParams subpel_params = { sf->x_step_q4, sf->y_step_q4, + pos_x & SCALE_SUBPEL_MASK, + pos_y & SCALE_SUBPEL_MASK }; + + // Get warp types. + const WarpedMotionParams *const wm = + &xd->global_motion[mi->ref_frame[ref_num]]; + const int is_global = is_global_mv_block(mi, wm->wmtype); + WarpTypesAllowed warp_types; + warp_types.global_warp_allowed = is_global; + warp_types.local_warp_allowed = mi->motion_mode == WARPED_CAUSAL; + + // Get convolve parameters. + ConvolveParams conv_params = get_conv_params(0, plane, xd->bd); + const InterpFilters filters = + av1_broadcast_interp_filter(EIGHTTAP_REGULAR); + + // Get the inter predictor. + const int build_for_obmc = 0; + av1_make_inter_predictor(pre, pre_buf->stride, comp_pred8, width, + &subpel_params, sf, width, height, &conv_params, + filters, &warp_types, mi_x >> pd->subsampling_x, + mi_y >> pd->subsampling_y, plane, ref_num, mi, + build_for_obmc, xd, cm->allow_warped_motion); + + return; + } + } + + const InterpFilterParams *filter = + (subpel_search == 1) + ? av1_get_4tap_interp_filter_params(EIGHTTAP_REGULAR) + : av1_get_interp_filter_params_with_block_size(EIGHTTAP_REGULAR, 8); + + if (!subpel_x_q3 && !subpel_y_q3) { + const uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); + uint16_t *comp_pred = CONVERT_TO_SHORTPTR(comp_pred8); + for (int i = 0; i < height; i++) { + memcpy(comp_pred, ref, width * sizeof(*comp_pred)); + comp_pred += width; + ref += ref_stride; + } + } else if (!subpel_y_q3) { + const int16_t *const kernel = + av1_get_interp_filter_subpel_kernel(filter, subpel_x_q3 << 1); + aom_highbd_convolve8_horiz(ref8, ref_stride, comp_pred8, width, kernel, 16, + NULL, -1, width, height, bd); + } else if (!subpel_x_q3) { + const int16_t *const kernel = + av1_get_interp_filter_subpel_kernel(filter, subpel_y_q3 << 1); + aom_highbd_convolve8_vert(ref8, ref_stride, comp_pred8, width, NULL, -1, + kernel, 16, width, height, bd); + } else { + DECLARE_ALIGNED(16, uint16_t, + temp[((MAX_SB_SIZE + 16) + 16) * MAX_SB_SIZE]); + const int16_t *const kernel_x = + av1_get_interp_filter_subpel_kernel(filter, subpel_x_q3 << 1); + const int16_t *const kernel_y = + av1_get_interp_filter_subpel_kernel(filter, subpel_y_q3 << 1); + const int intermediate_height = + (((height - 1) * 8 + subpel_y_q3) >> 3) + filter->taps; + assert(intermediate_height <= (MAX_SB_SIZE * 2 + 16) + 16); + aom_highbd_convolve8_horiz(ref8 - ref_stride * ((filter->taps >> 1) - 1), + ref_stride, CONVERT_TO_BYTEPTR(temp), + MAX_SB_SIZE, kernel_x, 16, NULL, -1, width, + intermediate_height, bd); + aom_highbd_convolve8_vert( + CONVERT_TO_BYTEPTR(temp + MAX_SB_SIZE * ((filter->taps >> 1) - 1)), + MAX_SB_SIZE, comp_pred8, width, NULL, -1, kernel_y, 16, width, height, + bd); + } +} + +void aom_highbd_comp_avg_upsampled_pred_c( + MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col, + const MV *const mv, uint8_t *comp_pred8, const uint8_t *pred8, int width, + int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref8, + int ref_stride, int bd, int subpel_search) { + int i, j; + + const uint16_t *pred = CONVERT_TO_SHORTPTR(pred8); + uint16_t *comp_pred = CONVERT_TO_SHORTPTR(comp_pred8); + aom_highbd_upsampled_pred(xd, cm, mi_row, mi_col, mv, comp_pred8, width, + height, subpel_x_q3, subpel_y_q3, ref8, ref_stride, + bd, subpel_search); + for (i = 0; i < height; ++i) { + for (j = 0; j < width; ++j) { + comp_pred[j] = ROUND_POWER_OF_TWO(pred[j] + comp_pred[j], 1); + } + comp_pred += width; + pred += width; + } +} + +void aom_highbd_jnt_comp_avg_pred_c(uint8_t *comp_pred8, const uint8_t *pred8, + int width, int height, const uint8_t *ref8, + int ref_stride, + const JNT_COMP_PARAMS *jcp_param) { + int i, j; + const int fwd_offset = jcp_param->fwd_offset; + const int bck_offset = jcp_param->bck_offset; + uint16_t *pred = CONVERT_TO_SHORTPTR(pred8); + uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); + uint16_t *comp_pred = CONVERT_TO_SHORTPTR(comp_pred8); + + for (i = 0; i < height; ++i) { + for (j = 0; j < width; ++j) { + int tmp = pred[j] * bck_offset + ref[j] * fwd_offset; + tmp = ROUND_POWER_OF_TWO(tmp, DIST_PRECISION_BITS); + comp_pred[j] = (uint16_t)tmp; + } + comp_pred += width; + pred += width; + ref += ref_stride; + } +} + +void aom_highbd_jnt_comp_avg_upsampled_pred_c( + MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col, + const MV *const mv, uint8_t *comp_pred8, const uint8_t *pred8, int width, + int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref8, + int ref_stride, int bd, const JNT_COMP_PARAMS *jcp_param, + int subpel_search) { + int i, j; + const int fwd_offset = jcp_param->fwd_offset; + const int bck_offset = jcp_param->bck_offset; + const uint16_t *pred = CONVERT_TO_SHORTPTR(pred8); + uint16_t *comp_pred = CONVERT_TO_SHORTPTR(comp_pred8); + aom_highbd_upsampled_pred(xd, cm, mi_row, mi_col, mv, comp_pred8, width, + height, subpel_x_q3, subpel_y_q3, ref8, ref_stride, + bd, subpel_search); + + for (i = 0; i < height; i++) { + for (j = 0; j < width; j++) { + int tmp = pred[j] * bck_offset + comp_pred[j] * fwd_offset; + tmp = ROUND_POWER_OF_TWO(tmp, DIST_PRECISION_BITS); + comp_pred[j] = (uint16_t)tmp; + } + comp_pred += width; + pred += width; + } +} + +void aom_comp_mask_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, + int height, const uint8_t *ref, int ref_stride, + const uint8_t *mask, int mask_stride, + int invert_mask) { + int i, j; + const uint8_t *src0 = invert_mask ? pred : ref; + const uint8_t *src1 = invert_mask ? ref : pred; + const int stride0 = invert_mask ? width : ref_stride; + const int stride1 = invert_mask ? ref_stride : width; + for (i = 0; i < height; ++i) { + for (j = 0; j < width; ++j) { + comp_pred[j] = AOM_BLEND_A64(mask[j], src0[j], src1[j]); + } + comp_pred += width; + src0 += stride0; + src1 += stride1; + mask += mask_stride; + } +} + +void aom_comp_mask_upsampled_pred_c(MACROBLOCKD *xd, const AV1_COMMON *const cm, + int mi_row, int mi_col, const MV *const mv, + uint8_t *comp_pred, const uint8_t *pred, + int width, int height, int subpel_x_q3, + int subpel_y_q3, const uint8_t *ref, + int ref_stride, const uint8_t *mask, + int mask_stride, int invert_mask, + int subpel_search) { + if (subpel_x_q3 | subpel_y_q3) { + aom_upsampled_pred_c(xd, cm, mi_row, mi_col, mv, comp_pred, width, height, + subpel_x_q3, subpel_y_q3, ref, ref_stride, + subpel_search); + ref = comp_pred; + ref_stride = width; + } + aom_comp_mask_pred_c(comp_pred, pred, width, height, ref, ref_stride, mask, + mask_stride, invert_mask); +} + +#define MASK_SUBPIX_VAR(W, H) \ + unsigned int aom_masked_sub_pixel_variance##W##x##H##_c( \ + const uint8_t *src, int src_stride, int xoffset, int yoffset, \ + const uint8_t *ref, int ref_stride, const uint8_t *second_pred, \ + const uint8_t *msk, int msk_stride, int invert_mask, \ + unsigned int *sse) { \ + uint16_t fdata3[(H + 1) * W]; \ + uint8_t temp2[H * W]; \ + DECLARE_ALIGNED(16, uint8_t, temp3[H * W]); \ + \ + aom_var_filter_block2d_bil_first_pass_c(src, fdata3, src_stride, 1, H + 1, \ + W, bilinear_filters_2t[xoffset]); \ + aom_var_filter_block2d_bil_second_pass_c(fdata3, temp2, W, W, H, W, \ + bilinear_filters_2t[yoffset]); \ + \ + aom_comp_mask_pred_c(temp3, second_pred, W, H, temp2, W, msk, msk_stride, \ + invert_mask); \ + return aom_variance##W##x##H##_c(temp3, W, ref, ref_stride, sse); \ + } + +MASK_SUBPIX_VAR(4, 4) +MASK_SUBPIX_VAR(4, 8) +MASK_SUBPIX_VAR(8, 4) +MASK_SUBPIX_VAR(8, 8) +MASK_SUBPIX_VAR(8, 16) +MASK_SUBPIX_VAR(16, 8) +MASK_SUBPIX_VAR(16, 16) +MASK_SUBPIX_VAR(16, 32) +MASK_SUBPIX_VAR(32, 16) +MASK_SUBPIX_VAR(32, 32) +MASK_SUBPIX_VAR(32, 64) +MASK_SUBPIX_VAR(64, 32) +MASK_SUBPIX_VAR(64, 64) +MASK_SUBPIX_VAR(64, 128) +MASK_SUBPIX_VAR(128, 64) +MASK_SUBPIX_VAR(128, 128) +MASK_SUBPIX_VAR(4, 16) +MASK_SUBPIX_VAR(16, 4) +MASK_SUBPIX_VAR(8, 32) +MASK_SUBPIX_VAR(32, 8) +MASK_SUBPIX_VAR(16, 64) +MASK_SUBPIX_VAR(64, 16) + +void aom_highbd_comp_mask_pred_c(uint8_t *comp_pred8, const uint8_t *pred8, + int width, int height, const uint8_t *ref8, + int ref_stride, const uint8_t *mask, + int mask_stride, int invert_mask) { + int i, j; + uint16_t *pred = CONVERT_TO_SHORTPTR(pred8); + uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); + uint16_t *comp_pred = CONVERT_TO_SHORTPTR(comp_pred8); + for (i = 0; i < height; ++i) { + for (j = 0; j < width; ++j) { + if (!invert_mask) + comp_pred[j] = AOM_BLEND_A64(mask[j], ref[j], pred[j]); + else + comp_pred[j] = AOM_BLEND_A64(mask[j], pred[j], ref[j]); + } + comp_pred += width; + pred += width; + ref += ref_stride; + mask += mask_stride; + } +} + +void aom_highbd_comp_mask_upsampled_pred( + MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col, + const MV *const mv, uint8_t *comp_pred8, const uint8_t *pred8, int width, + int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref8, + int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask, + int bd, int subpel_search) { + aom_highbd_upsampled_pred(xd, cm, mi_row, mi_col, mv, comp_pred8, width, + height, subpel_x_q3, subpel_y_q3, ref8, ref_stride, + bd, subpel_search); + aom_highbd_comp_mask_pred(comp_pred8, pred8, width, height, comp_pred8, width, + mask, mask_stride, invert_mask); +} + +#define HIGHBD_MASK_SUBPIX_VAR(W, H) \ + unsigned int aom_highbd_8_masked_sub_pixel_variance##W##x##H##_c( \ + const uint8_t *src, int src_stride, int xoffset, int yoffset, \ + const uint8_t *ref, int ref_stride, const uint8_t *second_pred, \ + const uint8_t *msk, int msk_stride, int invert_mask, \ + unsigned int *sse) { \ + uint16_t fdata3[(H + 1) * W]; \ + uint16_t temp2[H * W]; \ + DECLARE_ALIGNED(16, uint16_t, temp3[H * W]); \ + \ + aom_highbd_var_filter_block2d_bil_first_pass( \ + src, fdata3, src_stride, 1, H + 1, W, bilinear_filters_2t[xoffset]); \ + aom_highbd_var_filter_block2d_bil_second_pass( \ + fdata3, temp2, W, W, H, W, bilinear_filters_2t[yoffset]); \ + \ + aom_highbd_comp_mask_pred_c(CONVERT_TO_BYTEPTR(temp3), second_pred, W, H, \ + CONVERT_TO_BYTEPTR(temp2), W, msk, msk_stride, \ + invert_mask); \ + \ + return aom_highbd_8_variance##W##x##H##_c(CONVERT_TO_BYTEPTR(temp3), W, \ + ref, ref_stride, sse); \ + } \ + \ + unsigned int aom_highbd_10_masked_sub_pixel_variance##W##x##H##_c( \ + const uint8_t *src, int src_stride, int xoffset, int yoffset, \ + const uint8_t *ref, int ref_stride, const uint8_t *second_pred, \ + const uint8_t *msk, int msk_stride, int invert_mask, \ + unsigned int *sse) { \ + uint16_t fdata3[(H + 1) * W]; \ + uint16_t temp2[H * W]; \ + DECLARE_ALIGNED(16, uint16_t, temp3[H * W]); \ + \ + aom_highbd_var_filter_block2d_bil_first_pass( \ + src, fdata3, src_stride, 1, H + 1, W, bilinear_filters_2t[xoffset]); \ + aom_highbd_var_filter_block2d_bil_second_pass( \ + fdata3, temp2, W, W, H, W, bilinear_filters_2t[yoffset]); \ + \ + aom_highbd_comp_mask_pred_c(CONVERT_TO_BYTEPTR(temp3), second_pred, W, H, \ + CONVERT_TO_BYTEPTR(temp2), W, msk, msk_stride, \ + invert_mask); \ + \ + return aom_highbd_10_variance##W##x##H##_c(CONVERT_TO_BYTEPTR(temp3), W, \ + ref, ref_stride, sse); \ + } \ + \ + unsigned int aom_highbd_12_masked_sub_pixel_variance##W##x##H##_c( \ + const uint8_t *src, int src_stride, int xoffset, int yoffset, \ + const uint8_t *ref, int ref_stride, const uint8_t *second_pred, \ + const uint8_t *msk, int msk_stride, int invert_mask, \ + unsigned int *sse) { \ + uint16_t fdata3[(H + 1) * W]; \ + uint16_t temp2[H * W]; \ + DECLARE_ALIGNED(16, uint16_t, temp3[H * W]); \ + \ + aom_highbd_var_filter_block2d_bil_first_pass( \ + src, fdata3, src_stride, 1, H + 1, W, bilinear_filters_2t[xoffset]); \ + aom_highbd_var_filter_block2d_bil_second_pass( \ + fdata3, temp2, W, W, H, W, bilinear_filters_2t[yoffset]); \ + \ + aom_highbd_comp_mask_pred_c(CONVERT_TO_BYTEPTR(temp3), second_pred, W, H, \ + CONVERT_TO_BYTEPTR(temp2), W, msk, msk_stride, \ + invert_mask); \ + \ + return aom_highbd_12_variance##W##x##H##_c(CONVERT_TO_BYTEPTR(temp3), W, \ + ref, ref_stride, sse); \ + } + +HIGHBD_MASK_SUBPIX_VAR(4, 4) +HIGHBD_MASK_SUBPIX_VAR(4, 8) +HIGHBD_MASK_SUBPIX_VAR(8, 4) +HIGHBD_MASK_SUBPIX_VAR(8, 8) +HIGHBD_MASK_SUBPIX_VAR(8, 16) +HIGHBD_MASK_SUBPIX_VAR(16, 8) +HIGHBD_MASK_SUBPIX_VAR(16, 16) +HIGHBD_MASK_SUBPIX_VAR(16, 32) +HIGHBD_MASK_SUBPIX_VAR(32, 16) +HIGHBD_MASK_SUBPIX_VAR(32, 32) +HIGHBD_MASK_SUBPIX_VAR(32, 64) +HIGHBD_MASK_SUBPIX_VAR(64, 32) +HIGHBD_MASK_SUBPIX_VAR(64, 64) +HIGHBD_MASK_SUBPIX_VAR(64, 128) +HIGHBD_MASK_SUBPIX_VAR(128, 64) +HIGHBD_MASK_SUBPIX_VAR(128, 128) +HIGHBD_MASK_SUBPIX_VAR(4, 16) +HIGHBD_MASK_SUBPIX_VAR(16, 4) +HIGHBD_MASK_SUBPIX_VAR(8, 32) +HIGHBD_MASK_SUBPIX_VAR(32, 8) +HIGHBD_MASK_SUBPIX_VAR(16, 64) +HIGHBD_MASK_SUBPIX_VAR(64, 16) + +static INLINE void obmc_variance(const uint8_t *pre, int pre_stride, + const int32_t *wsrc, const int32_t *mask, + int w, int h, unsigned int *sse, int *sum) { + int i, j; + + *sse = 0; + *sum = 0; + + for (i = 0; i < h; i++) { + for (j = 0; j < w; j++) { + int diff = ROUND_POWER_OF_TWO_SIGNED(wsrc[j] - pre[j] * mask[j], 12); + *sum += diff; + *sse += diff * diff; + } + + pre += pre_stride; + wsrc += w; + mask += w; + } +} + +#define OBMC_VAR(W, H) \ + unsigned int aom_obmc_variance##W##x##H##_c( \ + const uint8_t *pre, int pre_stride, const int32_t *wsrc, \ + const int32_t *mask, unsigned int *sse) { \ + int sum; \ + obmc_variance(pre, pre_stride, wsrc, mask, W, H, sse, &sum); \ + return *sse - (unsigned int)(((int64_t)sum * sum) / (W * H)); \ + } + +#define OBMC_SUBPIX_VAR(W, H) \ + unsigned int aom_obmc_sub_pixel_variance##W##x##H##_c( \ + const uint8_t *pre, int pre_stride, int xoffset, int yoffset, \ + const int32_t *wsrc, const int32_t *mask, unsigned int *sse) { \ + uint16_t fdata3[(H + 1) * W]; \ + uint8_t temp2[H * W]; \ + \ + aom_var_filter_block2d_bil_first_pass_c(pre, fdata3, pre_stride, 1, H + 1, \ + W, bilinear_filters_2t[xoffset]); \ + aom_var_filter_block2d_bil_second_pass_c(fdata3, temp2, W, W, H, W, \ + bilinear_filters_2t[yoffset]); \ + \ + return aom_obmc_variance##W##x##H##_c(temp2, W, wsrc, mask, sse); \ + } + +OBMC_VAR(4, 4) +OBMC_SUBPIX_VAR(4, 4) + +OBMC_VAR(4, 8) +OBMC_SUBPIX_VAR(4, 8) + +OBMC_VAR(8, 4) +OBMC_SUBPIX_VAR(8, 4) + +OBMC_VAR(8, 8) +OBMC_SUBPIX_VAR(8, 8) + +OBMC_VAR(8, 16) +OBMC_SUBPIX_VAR(8, 16) + +OBMC_VAR(16, 8) +OBMC_SUBPIX_VAR(16, 8) + +OBMC_VAR(16, 16) +OBMC_SUBPIX_VAR(16, 16) + +OBMC_VAR(16, 32) +OBMC_SUBPIX_VAR(16, 32) + +OBMC_VAR(32, 16) +OBMC_SUBPIX_VAR(32, 16) + +OBMC_VAR(32, 32) +OBMC_SUBPIX_VAR(32, 32) + +OBMC_VAR(32, 64) +OBMC_SUBPIX_VAR(32, 64) + +OBMC_VAR(64, 32) +OBMC_SUBPIX_VAR(64, 32) + +OBMC_VAR(64, 64) +OBMC_SUBPIX_VAR(64, 64) + +OBMC_VAR(64, 128) +OBMC_SUBPIX_VAR(64, 128) + +OBMC_VAR(128, 64) +OBMC_SUBPIX_VAR(128, 64) + +OBMC_VAR(128, 128) +OBMC_SUBPIX_VAR(128, 128) + +OBMC_VAR(4, 16) +OBMC_SUBPIX_VAR(4, 16) +OBMC_VAR(16, 4) +OBMC_SUBPIX_VAR(16, 4) +OBMC_VAR(8, 32) +OBMC_SUBPIX_VAR(8, 32) +OBMC_VAR(32, 8) +OBMC_SUBPIX_VAR(32, 8) +OBMC_VAR(16, 64) +OBMC_SUBPIX_VAR(16, 64) +OBMC_VAR(64, 16) +OBMC_SUBPIX_VAR(64, 16) + +static INLINE void highbd_obmc_variance64(const uint8_t *pre8, int pre_stride, + const int32_t *wsrc, + const int32_t *mask, int w, int h, + uint64_t *sse, int64_t *sum) { + int i, j; + uint16_t *pre = CONVERT_TO_SHORTPTR(pre8); + + *sse = 0; + *sum = 0; + + for (i = 0; i < h; i++) { + for (j = 0; j < w; j++) { + int diff = ROUND_POWER_OF_TWO_SIGNED(wsrc[j] - pre[j] * mask[j], 12); + *sum += diff; + *sse += diff * diff; + } + + pre += pre_stride; + wsrc += w; + mask += w; + } +} + +static INLINE void highbd_obmc_variance(const uint8_t *pre8, int pre_stride, + const int32_t *wsrc, + const int32_t *mask, int w, int h, + unsigned int *sse, int *sum) { + int64_t sum64; + uint64_t sse64; + highbd_obmc_variance64(pre8, pre_stride, wsrc, mask, w, h, &sse64, &sum64); + *sum = (int)sum64; + *sse = (unsigned int)sse64; +} + +static INLINE void highbd_10_obmc_variance(const uint8_t *pre8, int pre_stride, + const int32_t *wsrc, + const int32_t *mask, int w, int h, + unsigned int *sse, int *sum) { + int64_t sum64; + uint64_t sse64; + highbd_obmc_variance64(pre8, pre_stride, wsrc, mask, w, h, &sse64, &sum64); + *sum = (int)ROUND_POWER_OF_TWO(sum64, 2); + *sse = (unsigned int)ROUND_POWER_OF_TWO(sse64, 4); +} + +static INLINE void highbd_12_obmc_variance(const uint8_t *pre8, int pre_stride, + const int32_t *wsrc, + const int32_t *mask, int w, int h, + unsigned int *sse, int *sum) { + int64_t sum64; + uint64_t sse64; + highbd_obmc_variance64(pre8, pre_stride, wsrc, mask, w, h, &sse64, &sum64); + *sum = (int)ROUND_POWER_OF_TWO(sum64, 4); + *sse = (unsigned int)ROUND_POWER_OF_TWO(sse64, 8); +} + +#define HIGHBD_OBMC_VAR(W, H) \ + unsigned int aom_highbd_obmc_variance##W##x##H##_c( \ + const uint8_t *pre, int pre_stride, const int32_t *wsrc, \ + const int32_t *mask, unsigned int *sse) { \ + int sum; \ + highbd_obmc_variance(pre, pre_stride, wsrc, mask, W, H, sse, &sum); \ + return *sse - (unsigned int)(((int64_t)sum * sum) / (W * H)); \ + } \ + \ + unsigned int aom_highbd_10_obmc_variance##W##x##H##_c( \ + const uint8_t *pre, int pre_stride, const int32_t *wsrc, \ + const int32_t *mask, unsigned int *sse) { \ + int sum; \ + int64_t var; \ + highbd_10_obmc_variance(pre, pre_stride, wsrc, mask, W, H, sse, &sum); \ + var = (int64_t)(*sse) - (((int64_t)sum * sum) / (W * H)); \ + return (var >= 0) ? (uint32_t)var : 0; \ + } \ + \ + unsigned int aom_highbd_12_obmc_variance##W##x##H##_c( \ + const uint8_t *pre, int pre_stride, const int32_t *wsrc, \ + const int32_t *mask, unsigned int *sse) { \ + int sum; \ + int64_t var; \ + highbd_12_obmc_variance(pre, pre_stride, wsrc, mask, W, H, sse, &sum); \ + var = (int64_t)(*sse) - (((int64_t)sum * sum) / (W * H)); \ + return (var >= 0) ? (uint32_t)var : 0; \ + } + +#define HIGHBD_OBMC_SUBPIX_VAR(W, H) \ + unsigned int aom_highbd_obmc_sub_pixel_variance##W##x##H##_c( \ + const uint8_t *pre, int pre_stride, int xoffset, int yoffset, \ + const int32_t *wsrc, const int32_t *mask, unsigned int *sse) { \ + uint16_t fdata3[(H + 1) * W]; \ + uint16_t temp2[H * W]; \ + \ + aom_highbd_var_filter_block2d_bil_first_pass( \ + pre, fdata3, pre_stride, 1, H + 1, W, bilinear_filters_2t[xoffset]); \ + aom_highbd_var_filter_block2d_bil_second_pass( \ + fdata3, temp2, W, W, H, W, bilinear_filters_2t[yoffset]); \ + \ + return aom_highbd_obmc_variance##W##x##H##_c(CONVERT_TO_BYTEPTR(temp2), W, \ + wsrc, mask, sse); \ + } \ + \ + unsigned int aom_highbd_10_obmc_sub_pixel_variance##W##x##H##_c( \ + const uint8_t *pre, int pre_stride, int xoffset, int yoffset, \ + const int32_t *wsrc, const int32_t *mask, unsigned int *sse) { \ + uint16_t fdata3[(H + 1) * W]; \ + uint16_t temp2[H * W]; \ + \ + aom_highbd_var_filter_block2d_bil_first_pass( \ + pre, fdata3, pre_stride, 1, H + 1, W, bilinear_filters_2t[xoffset]); \ + aom_highbd_var_filter_block2d_bil_second_pass( \ + fdata3, temp2, W, W, H, W, bilinear_filters_2t[yoffset]); \ + \ + return aom_highbd_10_obmc_variance##W##x##H##_c(CONVERT_TO_BYTEPTR(temp2), \ + W, wsrc, mask, sse); \ + } \ + \ + unsigned int aom_highbd_12_obmc_sub_pixel_variance##W##x##H##_c( \ + const uint8_t *pre, int pre_stride, int xoffset, int yoffset, \ + const int32_t *wsrc, const int32_t *mask, unsigned int *sse) { \ + uint16_t fdata3[(H + 1) * W]; \ + uint16_t temp2[H * W]; \ + \ + aom_highbd_var_filter_block2d_bil_first_pass( \ + pre, fdata3, pre_stride, 1, H + 1, W, bilinear_filters_2t[xoffset]); \ + aom_highbd_var_filter_block2d_bil_second_pass( \ + fdata3, temp2, W, W, H, W, bilinear_filters_2t[yoffset]); \ + \ + return aom_highbd_12_obmc_variance##W##x##H##_c(CONVERT_TO_BYTEPTR(temp2), \ + W, wsrc, mask, sse); \ + } + +HIGHBD_OBMC_VAR(4, 4) +HIGHBD_OBMC_SUBPIX_VAR(4, 4) + +HIGHBD_OBMC_VAR(4, 8) +HIGHBD_OBMC_SUBPIX_VAR(4, 8) + +HIGHBD_OBMC_VAR(8, 4) +HIGHBD_OBMC_SUBPIX_VAR(8, 4) + +HIGHBD_OBMC_VAR(8, 8) +HIGHBD_OBMC_SUBPIX_VAR(8, 8) + +HIGHBD_OBMC_VAR(8, 16) +HIGHBD_OBMC_SUBPIX_VAR(8, 16) + +HIGHBD_OBMC_VAR(16, 8) +HIGHBD_OBMC_SUBPIX_VAR(16, 8) + +HIGHBD_OBMC_VAR(16, 16) +HIGHBD_OBMC_SUBPIX_VAR(16, 16) + +HIGHBD_OBMC_VAR(16, 32) +HIGHBD_OBMC_SUBPIX_VAR(16, 32) + +HIGHBD_OBMC_VAR(32, 16) +HIGHBD_OBMC_SUBPIX_VAR(32, 16) + +HIGHBD_OBMC_VAR(32, 32) +HIGHBD_OBMC_SUBPIX_VAR(32, 32) + +HIGHBD_OBMC_VAR(32, 64) +HIGHBD_OBMC_SUBPIX_VAR(32, 64) + +HIGHBD_OBMC_VAR(64, 32) +HIGHBD_OBMC_SUBPIX_VAR(64, 32) + +HIGHBD_OBMC_VAR(64, 64) +HIGHBD_OBMC_SUBPIX_VAR(64, 64) + +HIGHBD_OBMC_VAR(64, 128) +HIGHBD_OBMC_SUBPIX_VAR(64, 128) + +HIGHBD_OBMC_VAR(128, 64) +HIGHBD_OBMC_SUBPIX_VAR(128, 64) + +HIGHBD_OBMC_VAR(128, 128) +HIGHBD_OBMC_SUBPIX_VAR(128, 128) + +HIGHBD_OBMC_VAR(4, 16) +HIGHBD_OBMC_SUBPIX_VAR(4, 16) +HIGHBD_OBMC_VAR(16, 4) +HIGHBD_OBMC_SUBPIX_VAR(16, 4) +HIGHBD_OBMC_VAR(8, 32) +HIGHBD_OBMC_SUBPIX_VAR(8, 32) +HIGHBD_OBMC_VAR(32, 8) +HIGHBD_OBMC_SUBPIX_VAR(32, 8) +HIGHBD_OBMC_VAR(16, 64) +HIGHBD_OBMC_SUBPIX_VAR(16, 64) +HIGHBD_OBMC_VAR(64, 16) +HIGHBD_OBMC_SUBPIX_VAR(64, 16) diff --git a/third_party/aom/aom_dsp/variance.h b/third_party/aom/aom_dsp/variance.h new file mode 100644 index 0000000000..362da29d33 --- /dev/null +++ b/third_party/aom/aom_dsp/variance.h @@ -0,0 +1,130 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_VARIANCE_H_ +#define AOM_AOM_DSP_VARIANCE_H_ + +#include "config/aom_config.h" + +#include "aom/aom_integer.h" + +#ifdef __cplusplus +extern "C" { +#endif + +#define FILTER_BITS 7 +#define FILTER_WEIGHT 128 + +typedef unsigned int (*aom_sad_fn_t)(const uint8_t *a, int a_stride, + const uint8_t *b, int b_stride); + +typedef unsigned int (*aom_sad_avg_fn_t)(const uint8_t *a, int a_stride, + const uint8_t *b, int b_stride, + const uint8_t *second_pred); + +typedef void (*aom_copy32xn_fn_t)(const uint8_t *a, int a_stride, uint8_t *b, + int b_stride, int n); + +typedef void (*aom_sad_multi_d_fn_t)(const uint8_t *a, int a_stride, + const uint8_t *const b_array[], + int b_stride, unsigned int *sad_array); + +typedef unsigned int (*aom_variance_fn_t)(const uint8_t *a, int a_stride, + const uint8_t *b, int b_stride, + unsigned int *sse); + +typedef unsigned int (*aom_subpixvariance_fn_t)(const uint8_t *a, int a_stride, + int xoffset, int yoffset, + const uint8_t *b, int b_stride, + unsigned int *sse); + +typedef unsigned int (*aom_subp_avg_variance_fn_t)( + const uint8_t *a, int a_stride, int xoffset, int yoffset, const uint8_t *b, + int b_stride, unsigned int *sse, const uint8_t *second_pred); + +typedef unsigned int (*aom_jnt_sad_avg_fn_t)(const uint8_t *a, int a_stride, + const uint8_t *b, int b_stride, + const uint8_t *second_pred, + const JNT_COMP_PARAMS *jcp_param); + +typedef unsigned int (*aom_jnt_subp_avg_variance_fn_t)( + const uint8_t *a, int a_stride, int xoffset, int yoffset, const uint8_t *b, + int b_stride, unsigned int *sse, const uint8_t *second_pred, + const JNT_COMP_PARAMS *jcp_param); + +typedef unsigned int (*aom_masked_sad_fn_t)(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride, + const uint8_t *second_pred, + const uint8_t *msk, int msk_stride, + int invert_mask); +typedef unsigned int (*aom_masked_subpixvariance_fn_t)( + const uint8_t *src, int src_stride, int xoffset, int yoffset, + const uint8_t *ref, int ref_stride, const uint8_t *second_pred, + const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse); + +void aom_highbd_comp_mask_upsampled_pred( + MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col, + const MV *const mv, uint8_t *comp_pred8, const uint8_t *pred8, int width, + int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref8, + int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask, + int bd, int subpel_search); + +typedef unsigned int (*aom_obmc_sad_fn_t)(const uint8_t *pred, int pred_stride, + const int32_t *wsrc, + const int32_t *msk); +typedef unsigned int (*aom_obmc_variance_fn_t)(const uint8_t *pred, + int pred_stride, + const int32_t *wsrc, + const int32_t *msk, + unsigned int *sse); +typedef unsigned int (*aom_obmc_subpixvariance_fn_t)( + const uint8_t *pred, int pred_stride, int xoffset, int yoffset, + const int32_t *wsrc, const int32_t *msk, unsigned int *sse); + +typedef struct aom_variance_vtable { + aom_sad_fn_t sdf; + aom_sad_avg_fn_t sdaf; + aom_variance_fn_t vf; + aom_subpixvariance_fn_t svf; + aom_subp_avg_variance_fn_t svaf; + aom_sad_multi_d_fn_t sdx4df; + aom_masked_sad_fn_t msdf; + aom_masked_subpixvariance_fn_t msvf; + aom_obmc_sad_fn_t osdf; + aom_obmc_variance_fn_t ovf; + aom_obmc_subpixvariance_fn_t osvf; + aom_jnt_sad_avg_fn_t jsdaf; + aom_jnt_subp_avg_variance_fn_t jsvaf; +} aom_variance_fn_ptr_t; + +void aom_highbd_var_filter_block2d_bil_first_pass( + const uint8_t *src_ptr8, uint16_t *output_ptr, + unsigned int src_pixels_per_line, int pixel_step, + unsigned int output_height, unsigned int output_width, + const uint8_t *filter); + +void aom_highbd_var_filter_block2d_bil_second_pass( + const uint16_t *src_ptr, uint16_t *output_ptr, + unsigned int src_pixels_per_line, unsigned int pixel_step, + unsigned int output_height, unsigned int output_width, + const uint8_t *filter); + +uint32_t aom_sse_odd_size(const uint8_t *a, int a_stride, const uint8_t *b, + int b_stride, int w, int h); + +uint64_t aom_highbd_sse_odd_size(const uint8_t *a, int a_stride, + const uint8_t *b, int b_stride, int w, int h); + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif // AOM_AOM_DSP_VARIANCE_H_ diff --git a/third_party/aom/aom_dsp/x86/aom_asm_stubs.c b/third_party/aom/aom_dsp/x86/aom_asm_stubs.c new file mode 100644 index 0000000000..5f5bf5f14e --- /dev/null +++ b/third_party/aom/aom_dsp/x86/aom_asm_stubs.c @@ -0,0 +1,89 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include "config/aom_config.h" +#include "config/aom_dsp_rtcd.h" + +#include "aom_dsp/x86/convolve.h" + +#if HAVE_SSE2 +filter8_1dfunction aom_filter_block1d16_v8_sse2; +filter8_1dfunction aom_filter_block1d16_h8_sse2; +filter8_1dfunction aom_filter_block1d8_v8_sse2; +filter8_1dfunction aom_filter_block1d8_h8_sse2; +filter8_1dfunction aom_filter_block1d4_v8_sse2; +filter8_1dfunction aom_filter_block1d4_h8_sse2; + +#define aom_filter_block1d16_h4_sse2 aom_filter_block1d16_h8_sse2 +#define aom_filter_block1d16_v4_sse2 aom_filter_block1d16_v8_sse2 +#define aom_filter_block1d8_h4_sse2 aom_filter_block1d8_h8_sse2 +#define aom_filter_block1d8_v4_sse2 aom_filter_block1d8_v8_sse2 +#define aom_filter_block1d4_h4_sse2 aom_filter_block1d4_h8_sse2 +#define aom_filter_block1d4_v4_sse2 aom_filter_block1d4_v8_sse2 + +filter8_1dfunction aom_filter_block1d16_v2_sse2; +filter8_1dfunction aom_filter_block1d16_h2_sse2; +filter8_1dfunction aom_filter_block1d8_v2_sse2; +filter8_1dfunction aom_filter_block1d8_h2_sse2; +filter8_1dfunction aom_filter_block1d4_v2_sse2; +filter8_1dfunction aom_filter_block1d4_h2_sse2; + +// void aom_convolve8_horiz_sse2(const uint8_t *src, ptrdiff_t src_stride, +// uint8_t *dst, ptrdiff_t dst_stride, +// const int16_t *filter_x, int x_step_q4, +// const int16_t *filter_y, int y_step_q4, +// int w, int h); +// void aom_convolve8_vert_sse2(const uint8_t *src, ptrdiff_t src_stride, +// uint8_t *dst, ptrdiff_t dst_stride, +// const int16_t *filter_x, int x_step_q4, +// const int16_t *filter_y, int y_step_q4, +// int w, int h); +FUN_CONV_1D(horiz, x_step_q4, filter_x, h, src, , sse2); +FUN_CONV_1D(vert, y_step_q4, filter_y, v, src - src_stride * 3, , sse2); + +#if ARCH_X86_64 +highbd_filter8_1dfunction aom_highbd_filter_block1d16_v8_sse2; +highbd_filter8_1dfunction aom_highbd_filter_block1d16_h8_sse2; +highbd_filter8_1dfunction aom_highbd_filter_block1d8_v8_sse2; +highbd_filter8_1dfunction aom_highbd_filter_block1d8_h8_sse2; +highbd_filter8_1dfunction aom_highbd_filter_block1d4_v8_sse2; +highbd_filter8_1dfunction aom_highbd_filter_block1d4_h8_sse2; + +highbd_filter8_1dfunction aom_highbd_filter_block1d16_v2_sse2; +highbd_filter8_1dfunction aom_highbd_filter_block1d16_h2_sse2; +highbd_filter8_1dfunction aom_highbd_filter_block1d8_v2_sse2; +highbd_filter8_1dfunction aom_highbd_filter_block1d8_h2_sse2; +highbd_filter8_1dfunction aom_highbd_filter_block1d4_v2_sse2; +highbd_filter8_1dfunction aom_highbd_filter_block1d4_h2_sse2; + +// void aom_highbd_convolve8_horiz_sse2(const uint8_t *src, +// ptrdiff_t src_stride, +// uint8_t *dst, +// ptrdiff_t dst_stride, +// const int16_t *filter_x, +// int x_step_q4, +// const int16_t *filter_y, +// int y_step_q4, +// int w, int h, int bd); +// void aom_highbd_convolve8_vert_sse2(const uint8_t *src, +// ptrdiff_t src_stride, +// uint8_t *dst, +// ptrdiff_t dst_stride, +// const int16_t *filter_x, +// int x_step_q4, +// const int16_t *filter_y, +// int y_step_q4, +// int w, int h, int bd); +HIGH_FUN_CONV_1D(horiz, x_step_q4, filter_x, h, src, , sse2); +HIGH_FUN_CONV_1D(vert, y_step_q4, filter_y, v, src - src_stride * 3, , sse2); + +#endif // ARCH_X86_64 +#endif // HAVE_SSE2 diff --git a/third_party/aom/aom_dsp/x86/aom_convolve_copy_sse2.asm b/third_party/aom/aom_dsp/x86/aom_convolve_copy_sse2.asm new file mode 100644 index 0000000000..7283c32b8c --- /dev/null +++ b/third_party/aom/aom_dsp/x86/aom_convolve_copy_sse2.asm @@ -0,0 +1,297 @@ +; +; Copyright (c) 2016, Alliance for Open Media. All rights reserved +; +; This source code is subject to the terms of the BSD 2 Clause License and +; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License +; was not distributed with this source code in the LICENSE file, you can +; obtain it at www.aomedia.org/license/software. If the Alliance for Open +; Media Patent License 1.0 was not distributed with this source code in the +; PATENTS file, you can obtain it at www.aomedia.org/license/patent. +; + +; + +%include "third_party/x86inc/x86inc.asm" + +SECTION .text + +%macro convolve_fn 1-2 +%ifidn %1, avg +%define AUX_XMM_REGS 4 +%else +%define AUX_XMM_REGS 0 +%endif +%ifidn %2, highbd +%define pavg pavgw +cglobal %2_convolve_%1, 4, 7, 4+AUX_XMM_REGS, src, src_stride, \ + dst, dst_stride, \ + fx, fxs, fy, fys, w, h, bd +%else +%define pavg pavgb +cglobal convolve_%1, 4, 7, 4+AUX_XMM_REGS, src, src_stride, \ + dst, dst_stride, \ + fx, fxs, fy, fys, w, h +%endif + mov r4d, dword wm +%ifidn %2, highbd + shl r4d, 1 + shl srcq, 1 + shl src_strideq, 1 + shl dstq, 1 + shl dst_strideq, 1 +%else + cmp r4d, 4 + je .w4 +%endif + cmp r4d, 8 + je .w8 + cmp r4d, 16 + je .w16 + cmp r4d, 32 + je .w32 + + cmp r4d, 64 + je .w64 +%ifidn %2, highbd + cmp r4d, 128 + je .w128 + +.w256: + mov r4d, dword hm +.loop256: + movu m0, [srcq] + movu m1, [srcq+16] + movu m2, [srcq+32] + movu m3, [srcq+48] +%ifidn %1, avg + pavg m0, [dstq] + pavg m1, [dstq+16] + pavg m2, [dstq+32] + pavg m3, [dstq+48] +%endif + mova [dstq ], m0 + mova [dstq+16], m1 + mova [dstq+32], m2 + mova [dstq+48], m3 + movu m0, [srcq+64] + movu m1, [srcq+80] + movu m2, [srcq+96] + movu m3, [srcq+112] +%ifidn %1, avg + pavg m0, [dstq+64] + pavg m1, [dstq+80] + pavg m2, [dstq+96] + pavg m3, [dstq+112] +%endif + mova [dstq+64], m0 + mova [dstq+80], m1 + mova [dstq+96], m2 + mova [dstq+112], m3 + movu m0, [srcq+128] + movu m1, [srcq+128+16] + movu m2, [srcq+128+32] + movu m3, [srcq+128+48] +%ifidn %1, avg + pavg m0, [dstq+128] + pavg m1, [dstq+128+16] + pavg m2, [dstq+128+32] + pavg m3, [dstq+128+48] +%endif + mova [dstq+128 ], m0 + mova [dstq+128+16], m1 + mova [dstq+128+32], m2 + mova [dstq+128+48], m3 + movu m0, [srcq+128+64] + movu m1, [srcq+128+80] + movu m2, [srcq+128+96] + movu m3, [srcq+128+112] + add srcq, src_strideq +%ifidn %1, avg + pavg m0, [dstq+128+64] + pavg m1, [dstq+128+80] + pavg m2, [dstq+128+96] + pavg m3, [dstq+128+112] +%endif + mova [dstq+128+64], m0 + mova [dstq+128+80], m1 + mova [dstq+128+96], m2 + mova [dstq+128+112], m3 + add dstq, dst_strideq + sub r4d, 1 + jnz .loop256 + RET +%endif + +.w128: + mov r4d, dword hm +.loop128: + movu m0, [srcq] + movu m1, [srcq+16] + movu m2, [srcq+32] + movu m3, [srcq+48] +%ifidn %1, avg + pavg m0, [dstq] + pavg m1, [dstq+16] + pavg m2, [dstq+32] + pavg m3, [dstq+48] +%endif + mova [dstq ], m0 + mova [dstq+16], m1 + mova [dstq+32], m2 + mova [dstq+48], m3 + movu m0, [srcq+64] + movu m1, [srcq+80] + movu m2, [srcq+96] + movu m3, [srcq+112] + add srcq, src_strideq +%ifidn %1, avg + pavg m0, [dstq+64] + pavg m1, [dstq+80] + pavg m2, [dstq+96] + pavg m3, [dstq+112] +%endif + mova [dstq+64], m0 + mova [dstq+80], m1 + mova [dstq+96], m2 + mova [dstq+112], m3 + add dstq, dst_strideq + sub r4d, 1 + jnz .loop128 + RET + +.w64: + mov r4d, dword hm +.loop64: + movu m0, [srcq] + movu m1, [srcq+16] + movu m2, [srcq+32] + movu m3, [srcq+48] + add srcq, src_strideq +%ifidn %1, avg + pavg m0, [dstq] + pavg m1, [dstq+16] + pavg m2, [dstq+32] + pavg m3, [dstq+48] +%endif + mova [dstq ], m0 + mova [dstq+16], m1 + mova [dstq+32], m2 + mova [dstq+48], m3 + add dstq, dst_strideq + sub r4d, 1 + jnz .loop64 + RET + +.w32: + mov r4d, dword hm +.loop32: + movu m0, [srcq] + movu m1, [srcq+16] + movu m2, [srcq+src_strideq] + movu m3, [srcq+src_strideq+16] + lea srcq, [srcq+src_strideq*2] +%ifidn %1, avg + pavg m0, [dstq] + pavg m1, [dstq +16] + pavg m2, [dstq+dst_strideq] + pavg m3, [dstq+dst_strideq+16] +%endif + mova [dstq ], m0 + mova [dstq +16], m1 + mova [dstq+dst_strideq ], m2 + mova [dstq+dst_strideq+16], m3 + lea dstq, [dstq+dst_strideq*2] + sub r4d, 2 + jnz .loop32 + RET + +.w16: + mov r4d, dword hm + lea r5q, [src_strideq*3] + lea r6q, [dst_strideq*3] +.loop16: + movu m0, [srcq] + movu m1, [srcq+src_strideq] + movu m2, [srcq+src_strideq*2] + movu m3, [srcq+r5q] + lea srcq, [srcq+src_strideq*4] +%ifidn %1, avg + pavg m0, [dstq] + pavg m1, [dstq+dst_strideq] + pavg m2, [dstq+dst_strideq*2] + pavg m3, [dstq+r6q] +%endif + mova [dstq ], m0 + mova [dstq+dst_strideq ], m1 + mova [dstq+dst_strideq*2], m2 + mova [dstq+r6q ], m3 + lea dstq, [dstq+dst_strideq*4] + sub r4d, 4 + jnz .loop16 + RET + +.w8: + mov r4d, dword hm + lea r5q, [src_strideq*3] + lea r6q, [dst_strideq*3] +.loop8: + movh m0, [srcq] + movh m1, [srcq+src_strideq] + movh m2, [srcq+src_strideq*2] + movh m3, [srcq+r5q] + lea srcq, [srcq+src_strideq*4] +%ifidn %1, avg + movh m4, [dstq] + movh m5, [dstq+dst_strideq] + movh m6, [dstq+dst_strideq*2] + movh m7, [dstq+r6q] + pavg m0, m4 + pavg m1, m5 + pavg m2, m6 + pavg m3, m7 +%endif + movh [dstq ], m0 + movh [dstq+dst_strideq ], m1 + movh [dstq+dst_strideq*2], m2 + movh [dstq+r6q ], m3 + lea dstq, [dstq+dst_strideq*4] + sub r4d, 4 + jnz .loop8 + RET + +%ifnidn %2, highbd +.w4: + mov r4d, dword hm + lea r5q, [src_strideq*3] + lea r6q, [dst_strideq*3] +.loop4: + movd m0, [srcq] + movd m1, [srcq+src_strideq] + movd m2, [srcq+src_strideq*2] + movd m3, [srcq+r5q] + lea srcq, [srcq+src_strideq*4] +%ifidn %1, avg + movd m4, [dstq] + movd m5, [dstq+dst_strideq] + movd m6, [dstq+dst_strideq*2] + movd m7, [dstq+r6q] + pavg m0, m4 + pavg m1, m5 + pavg m2, m6 + pavg m3, m7 +%endif + movd [dstq ], m0 + movd [dstq+dst_strideq ], m1 + movd [dstq+dst_strideq*2], m2 + movd [dstq+r6q ], m3 + lea dstq, [dstq+dst_strideq*4] + sub r4d, 4 + jnz .loop4 + RET +%endif +%endmacro + +INIT_XMM sse2 +convolve_fn copy +convolve_fn avg +convolve_fn copy, highbd diff --git a/third_party/aom/aom_dsp/x86/aom_high_subpixel_8t_sse2.asm b/third_party/aom/aom_dsp/x86/aom_high_subpixel_8t_sse2.asm new file mode 100644 index 0000000000..b6f040791a --- /dev/null +++ b/third_party/aom/aom_dsp/x86/aom_high_subpixel_8t_sse2.asm @@ -0,0 +1,613 @@ +; +; Copyright (c) 2016, Alliance for Open Media. All rights reserved +; +; This source code is subject to the terms of the BSD 2 Clause License and +; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License +; was not distributed with this source code in the LICENSE file, you can +; obtain it at www.aomedia.org/license/software. If the Alliance for Open +; Media Patent License 1.0 was not distributed with this source code in the +; PATENTS file, you can obtain it at www.aomedia.org/license/patent. +; + +; + + +%include "aom_ports/x86_abi_support.asm" + +;Note: tap3 and tap4 have to be applied and added after other taps to avoid +;overflow. + +%macro HIGH_GET_FILTERS_4 0 + mov rdx, arg(5) ;filter ptr + mov rcx, 0x00000040 + + movdqa xmm7, [rdx] ;load filters + pshuflw xmm0, xmm7, 0b ;k0 + pshuflw xmm1, xmm7, 01010101b ;k1 + pshuflw xmm2, xmm7, 10101010b ;k2 + pshuflw xmm3, xmm7, 11111111b ;k3 + psrldq xmm7, 8 + pshuflw xmm4, xmm7, 0b ;k4 + pshuflw xmm5, xmm7, 01010101b ;k5 + pshuflw xmm6, xmm7, 10101010b ;k6 + pshuflw xmm7, xmm7, 11111111b ;k7 + + punpcklwd xmm0, xmm6 + punpcklwd xmm2, xmm5 + punpcklwd xmm3, xmm4 + punpcklwd xmm1, xmm7 + + movdqa k0k6, xmm0 + movdqa k2k5, xmm2 + movdqa k3k4, xmm3 + movdqa k1k7, xmm1 + + movq xmm6, rcx + pshufd xmm6, xmm6, 0 + movdqa krd, xmm6 + + ;Compute max and min values of a pixel + mov rdx, 0x00010001 + movsxd rcx, DWORD PTR arg(6) ;bps + movq xmm0, rdx + movq xmm1, rcx + pshufd xmm0, xmm0, 0b + movdqa xmm2, xmm0 + psllw xmm0, xmm1 + psubw xmm0, xmm2 + pxor xmm1, xmm1 + movdqa max, xmm0 ;max value (for clamping) + movdqa min, xmm1 ;min value (for clamping) + +%endm + +%macro HIGH_APPLY_FILTER_4 1 + punpcklwd xmm0, xmm6 ;two row in one register + punpcklwd xmm1, xmm7 + punpcklwd xmm2, xmm5 + punpcklwd xmm3, xmm4 + + pmaddwd xmm0, k0k6 ;multiply the filter factors + pmaddwd xmm1, k1k7 + pmaddwd xmm2, k2k5 + pmaddwd xmm3, k3k4 + + paddd xmm0, xmm1 ;sum + paddd xmm0, xmm2 + paddd xmm0, xmm3 + + paddd xmm0, krd ;rounding + psrad xmm0, 7 ;shift + packssdw xmm0, xmm0 ;pack to word + + ;clamp the values + pminsw xmm0, max + pmaxsw xmm0, min + +%if %1 + movq xmm1, [rdi] + pavgw xmm0, xmm1 +%endif + movq [rdi], xmm0 +%endm + +%macro HIGH_GET_FILTERS 0 + mov rdx, arg(5) ;filter ptr + mov rsi, arg(0) ;src_ptr + mov rdi, arg(2) ;output_ptr + mov rcx, 0x00000040 + + movdqa xmm7, [rdx] ;load filters + pshuflw xmm0, xmm7, 0b ;k0 + pshuflw xmm1, xmm7, 01010101b ;k1 + pshuflw xmm2, xmm7, 10101010b ;k2 + pshuflw xmm3, xmm7, 11111111b ;k3 + pshufhw xmm4, xmm7, 0b ;k4 + pshufhw xmm5, xmm7, 01010101b ;k5 + pshufhw xmm6, xmm7, 10101010b ;k6 + pshufhw xmm7, xmm7, 11111111b ;k7 + punpcklqdq xmm2, xmm2 + punpcklqdq xmm3, xmm3 + punpcklwd xmm0, xmm1 + punpckhwd xmm6, xmm7 + punpckhwd xmm2, xmm5 + punpckhwd xmm3, xmm4 + + movdqa k0k1, xmm0 ;store filter factors on stack + movdqa k6k7, xmm6 + movdqa k2k5, xmm2 + movdqa k3k4, xmm3 + + movq xmm6, rcx + pshufd xmm6, xmm6, 0 + movdqa krd, xmm6 ;rounding + + ;Compute max and min values of a pixel + mov rdx, 0x00010001 + movsxd rcx, DWORD PTR arg(6) ;bps + movq xmm0, rdx + movq xmm1, rcx + pshufd xmm0, xmm0, 0b + movdqa xmm2, xmm0 + psllw xmm0, xmm1 + psubw xmm0, xmm2 + pxor xmm1, xmm1 + movdqa max, xmm0 ;max value (for clamping) + movdqa min, xmm1 ;min value (for clamping) +%endm + +%macro LOAD_VERT_8 1 + movdqu xmm0, [rsi + %1] ;0 + movdqu xmm1, [rsi + rax + %1] ;1 + movdqu xmm6, [rsi + rdx * 2 + %1] ;6 + lea rsi, [rsi + rax] + movdqu xmm7, [rsi + rdx * 2 + %1] ;7 + movdqu xmm2, [rsi + rax + %1] ;2 + movdqu xmm3, [rsi + rax * 2 + %1] ;3 + movdqu xmm4, [rsi + rdx + %1] ;4 + movdqu xmm5, [rsi + rax * 4 + %1] ;5 +%endm + +%macro HIGH_APPLY_FILTER_8 2 + movdqu temp, xmm4 + movdqa xmm4, xmm0 + punpcklwd xmm0, xmm1 + punpckhwd xmm4, xmm1 + movdqa xmm1, xmm6 + punpcklwd xmm6, xmm7 + punpckhwd xmm1, xmm7 + movdqa xmm7, xmm2 + punpcklwd xmm2, xmm5 + punpckhwd xmm7, xmm5 + + movdqu xmm5, temp + movdqu temp, xmm4 + movdqa xmm4, xmm3 + punpcklwd xmm3, xmm5 + punpckhwd xmm4, xmm5 + movdqu xmm5, temp + + pmaddwd xmm0, k0k1 + pmaddwd xmm5, k0k1 + pmaddwd xmm6, k6k7 + pmaddwd xmm1, k6k7 + pmaddwd xmm2, k2k5 + pmaddwd xmm7, k2k5 + pmaddwd xmm3, k3k4 + pmaddwd xmm4, k3k4 + + paddd xmm0, xmm6 + paddd xmm0, xmm2 + paddd xmm0, xmm3 + paddd xmm5, xmm1 + paddd xmm5, xmm7 + paddd xmm5, xmm4 + + paddd xmm0, krd ;rounding + paddd xmm5, krd + psrad xmm0, 7 ;shift + psrad xmm5, 7 + packssdw xmm0, xmm5 ;pack back to word + + ;clamp the values + pminsw xmm0, max + pmaxsw xmm0, min + +%if %1 + movdqu xmm1, [rdi + %2] + pavgw xmm0, xmm1 +%endif + movdqu [rdi + %2], xmm0 +%endm + +SECTION .text + +;void aom_filter_block1d4_v8_sse2 +;( +; unsigned char *src_ptr, +; unsigned int src_pitch, +; unsigned char *output_ptr, +; unsigned int out_pitch, +; unsigned int output_height, +; short *filter +;) +global sym(aom_highbd_filter_block1d4_v8_sse2) PRIVATE +sym(aom_highbd_filter_block1d4_v8_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 7 + SAVE_XMM 7 + push rsi + push rdi + push rbx + ; end prolog + + ALIGN_STACK 16, rax + sub rsp, 16 * 7 + %define k0k6 [rsp + 16 * 0] + %define k2k5 [rsp + 16 * 1] + %define k3k4 [rsp + 16 * 2] + %define k1k7 [rsp + 16 * 3] + %define krd [rsp + 16 * 4] + %define max [rsp + 16 * 5] + %define min [rsp + 16 * 6] + + HIGH_GET_FILTERS_4 + + mov rsi, arg(0) ;src_ptr + mov rdi, arg(2) ;output_ptr + + movsxd rax, DWORD PTR arg(1) ;pixels_per_line + movsxd rbx, DWORD PTR arg(3) ;out_pitch + lea rax, [rax + rax] ;bytes per line + lea rbx, [rbx + rbx] + lea rdx, [rax + rax * 2] + movsxd rcx, DWORD PTR arg(4) ;output_height + +.loop: + movq xmm0, [rsi] ;load src: row 0 + movq xmm1, [rsi + rax] ;1 + movq xmm6, [rsi + rdx * 2] ;6 + lea rsi, [rsi + rax] + movq xmm7, [rsi + rdx * 2] ;7 + movq xmm2, [rsi + rax] ;2 + movq xmm3, [rsi + rax * 2] ;3 + movq xmm4, [rsi + rdx] ;4 + movq xmm5, [rsi + rax * 4] ;5 + + HIGH_APPLY_FILTER_4 0 + + lea rdi, [rdi + rbx] + dec rcx + jnz .loop + + add rsp, 16 * 7 + pop rsp + pop rbx + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +;void aom_filter_block1d8_v8_sse2 +;( +; unsigned char *src_ptr, +; unsigned int src_pitch, +; unsigned char *output_ptr, +; unsigned int out_pitch, +; unsigned int output_height, +; short *filter +;) +global sym(aom_highbd_filter_block1d8_v8_sse2) PRIVATE +sym(aom_highbd_filter_block1d8_v8_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 7 + SAVE_XMM 7 + push rsi + push rdi + push rbx + ; end prolog + + ALIGN_STACK 16, rax + sub rsp, 16 * 8 + %define k0k1 [rsp + 16 * 0] + %define k6k7 [rsp + 16 * 1] + %define k2k5 [rsp + 16 * 2] + %define k3k4 [rsp + 16 * 3] + %define krd [rsp + 16 * 4] + %define temp [rsp + 16 * 5] + %define max [rsp + 16 * 6] + %define min [rsp + 16 * 7] + + HIGH_GET_FILTERS + + movsxd rax, DWORD PTR arg(1) ;pixels_per_line + movsxd rbx, DWORD PTR arg(3) ;out_pitch + lea rax, [rax + rax] ;bytes per line + lea rbx, [rbx + rbx] + lea rdx, [rax + rax * 2] + movsxd rcx, DWORD PTR arg(4) ;output_height + +.loop: + LOAD_VERT_8 0 + HIGH_APPLY_FILTER_8 0, 0 + + lea rdi, [rdi + rbx] + dec rcx + jnz .loop + + add rsp, 16 * 8 + pop rsp + pop rbx + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +;void aom_filter_block1d16_v8_sse2 +;( +; unsigned char *src_ptr, +; unsigned int src_pitch, +; unsigned char *output_ptr, +; unsigned int out_pitch, +; unsigned int output_height, +; short *filter +;) +global sym(aom_highbd_filter_block1d16_v8_sse2) PRIVATE +sym(aom_highbd_filter_block1d16_v8_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 7 + SAVE_XMM 7 + push rsi + push rdi + push rbx + ; end prolog + + ALIGN_STACK 16, rax + sub rsp, 16 * 8 + %define k0k1 [rsp + 16 * 0] + %define k6k7 [rsp + 16 * 1] + %define k2k5 [rsp + 16 * 2] + %define k3k4 [rsp + 16 * 3] + %define krd [rsp + 16 * 4] + %define temp [rsp + 16 * 5] + %define max [rsp + 16 * 6] + %define min [rsp + 16 * 7] + + HIGH_GET_FILTERS + + movsxd rax, DWORD PTR arg(1) ;pixels_per_line + movsxd rbx, DWORD PTR arg(3) ;out_pitch + lea rax, [rax + rax] ;bytes per line + lea rbx, [rbx + rbx] + lea rdx, [rax + rax * 2] + movsxd rcx, DWORD PTR arg(4) ;output_height + +.loop: + LOAD_VERT_8 0 + HIGH_APPLY_FILTER_8 0, 0 + sub rsi, rax + + LOAD_VERT_8 16 + HIGH_APPLY_FILTER_8 0, 16 + add rdi, rbx + + dec rcx + jnz .loop + + add rsp, 16 * 8 + pop rsp + pop rbx + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +;void aom_filter_block1d4_h8_sse2 +;( +; unsigned char *src_ptr, +; unsigned int src_pixels_per_line, +; unsigned char *output_ptr, +; unsigned int output_pitch, +; unsigned int output_height, +; short *filter +;) +global sym(aom_highbd_filter_block1d4_h8_sse2) PRIVATE +sym(aom_highbd_filter_block1d4_h8_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 7 + SAVE_XMM 7 + push rsi + push rdi + ; end prolog + + ALIGN_STACK 16, rax + sub rsp, 16 * 7 + %define k0k6 [rsp + 16 * 0] + %define k2k5 [rsp + 16 * 1] + %define k3k4 [rsp + 16 * 2] + %define k1k7 [rsp + 16 * 3] + %define krd [rsp + 16 * 4] + %define max [rsp + 16 * 5] + %define min [rsp + 16 * 6] + + HIGH_GET_FILTERS_4 + + mov rsi, arg(0) ;src_ptr + mov rdi, arg(2) ;output_ptr + + movsxd rax, DWORD PTR arg(1) ;pixels_per_line + movsxd rdx, DWORD PTR arg(3) ;out_pitch + lea rax, [rax + rax] ;bytes per line + lea rdx, [rdx + rdx] + movsxd rcx, DWORD PTR arg(4) ;output_height + +.loop: + movdqu xmm0, [rsi - 6] ;load src + movdqu xmm4, [rsi + 2] + movdqa xmm1, xmm0 + movdqa xmm6, xmm4 + movdqa xmm7, xmm4 + movdqa xmm2, xmm0 + movdqa xmm3, xmm0 + movdqa xmm5, xmm4 + + psrldq xmm1, 2 + psrldq xmm6, 4 + psrldq xmm7, 6 + psrldq xmm2, 4 + psrldq xmm3, 6 + psrldq xmm5, 2 + + HIGH_APPLY_FILTER_4 0 + + lea rsi, [rsi + rax] + lea rdi, [rdi + rdx] + dec rcx + jnz .loop + + add rsp, 16 * 7 + pop rsp + + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +;void aom_filter_block1d8_h8_sse2 +;( +; unsigned char *src_ptr, +; unsigned int src_pixels_per_line, +; unsigned char *output_ptr, +; unsigned int output_pitch, +; unsigned int output_height, +; short *filter +;) +global sym(aom_highbd_filter_block1d8_h8_sse2) PRIVATE +sym(aom_highbd_filter_block1d8_h8_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 7 + SAVE_XMM 7 + push rsi + push rdi + ; end prolog + + ALIGN_STACK 16, rax + sub rsp, 16 * 8 + %define k0k1 [rsp + 16 * 0] + %define k6k7 [rsp + 16 * 1] + %define k2k5 [rsp + 16 * 2] + %define k3k4 [rsp + 16 * 3] + %define krd [rsp + 16 * 4] + %define temp [rsp + 16 * 5] + %define max [rsp + 16 * 6] + %define min [rsp + 16 * 7] + + HIGH_GET_FILTERS + + movsxd rax, DWORD PTR arg(1) ;pixels_per_line + movsxd rdx, DWORD PTR arg(3) ;out_pitch + lea rax, [rax + rax] ;bytes per line + lea rdx, [rdx + rdx] + movsxd rcx, DWORD PTR arg(4) ;output_height + +.loop: + movdqu xmm0, [rsi - 6] ;load src + movdqu xmm1, [rsi - 4] + movdqu xmm2, [rsi - 2] + movdqu xmm3, [rsi] + movdqu xmm4, [rsi + 2] + movdqu xmm5, [rsi + 4] + movdqu xmm6, [rsi + 6] + movdqu xmm7, [rsi + 8] + + HIGH_APPLY_FILTER_8 0, 0 + + lea rsi, [rsi + rax] + lea rdi, [rdi + rdx] + dec rcx + jnz .loop + + add rsp, 16 * 8 + pop rsp + + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +;void aom_filter_block1d16_h8_sse2 +;( +; unsigned char *src_ptr, +; unsigned int src_pixels_per_line, +; unsigned char *output_ptr, +; unsigned int output_pitch, +; unsigned int output_height, +; short *filter +;) +global sym(aom_highbd_filter_block1d16_h8_sse2) PRIVATE +sym(aom_highbd_filter_block1d16_h8_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 7 + SAVE_XMM 7 + push rsi + push rdi + ; end prolog + + ALIGN_STACK 16, rax + sub rsp, 16 * 8 + %define k0k1 [rsp + 16 * 0] + %define k6k7 [rsp + 16 * 1] + %define k2k5 [rsp + 16 * 2] + %define k3k4 [rsp + 16 * 3] + %define krd [rsp + 16 * 4] + %define temp [rsp + 16 * 5] + %define max [rsp + 16 * 6] + %define min [rsp + 16 * 7] + + HIGH_GET_FILTERS + + movsxd rax, DWORD PTR arg(1) ;pixels_per_line + movsxd rdx, DWORD PTR arg(3) ;out_pitch + lea rax, [rax + rax] ;bytes per line + lea rdx, [rdx + rdx] + movsxd rcx, DWORD PTR arg(4) ;output_height + +.loop: + movdqu xmm0, [rsi - 6] ;load src + movdqu xmm1, [rsi - 4] + movdqu xmm2, [rsi - 2] + movdqu xmm3, [rsi] + movdqu xmm4, [rsi + 2] + movdqu xmm5, [rsi + 4] + movdqu xmm6, [rsi + 6] + movdqu xmm7, [rsi + 8] + + HIGH_APPLY_FILTER_8 0, 0 + + movdqu xmm0, [rsi + 10] ;load src + movdqu xmm1, [rsi + 12] + movdqu xmm2, [rsi + 14] + movdqu xmm3, [rsi + 16] + movdqu xmm4, [rsi + 18] + movdqu xmm5, [rsi + 20] + movdqu xmm6, [rsi + 22] + movdqu xmm7, [rsi + 24] + + HIGH_APPLY_FILTER_8 0, 16 + + lea rsi, [rsi + rax] + lea rdi, [rdi + rdx] + dec rcx + jnz .loop + + add rsp, 16 * 8 + pop rsp + + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret diff --git a/third_party/aom/aom_dsp/x86/aom_high_subpixel_bilinear_sse2.asm b/third_party/aom/aom_dsp/x86/aom_high_subpixel_bilinear_sse2.asm new file mode 100644 index 0000000000..7b3fe6419a --- /dev/null +++ b/third_party/aom/aom_dsp/x86/aom_high_subpixel_bilinear_sse2.asm @@ -0,0 +1,338 @@ +; +; Copyright (c) 2016, Alliance for Open Media. All rights reserved +; +; This source code is subject to the terms of the BSD 2 Clause License and +; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License +; was not distributed with this source code in the LICENSE file, you can +; obtain it at www.aomedia.org/license/software. If the Alliance for Open +; Media Patent License 1.0 was not distributed with this source code in the +; PATENTS file, you can obtain it at www.aomedia.org/license/patent. +; + +; + +%include "aom_ports/x86_abi_support.asm" + +%macro HIGH_GET_PARAM_4 0 + mov rdx, arg(5) ;filter ptr + mov rsi, arg(0) ;src_ptr + mov rdi, arg(2) ;output_ptr + mov rcx, 0x00000040 + + movdqa xmm3, [rdx] ;load filters + pshuflw xmm4, xmm3, 11111111b ;k3 + psrldq xmm3, 8 + pshuflw xmm3, xmm3, 0b ;k4 + punpcklwd xmm4, xmm3 ;k3k4 + + movq xmm3, rcx ;rounding + pshufd xmm3, xmm3, 0 + + mov rdx, 0x00010001 + movsxd rcx, DWORD PTR arg(6) ;bps + movq xmm5, rdx + movq xmm2, rcx + pshufd xmm5, xmm5, 0b + movdqa xmm1, xmm5 + psllw xmm5, xmm2 + psubw xmm5, xmm1 ;max value (for clamping) + pxor xmm2, xmm2 ;min value (for clamping) + + movsxd rax, DWORD PTR arg(1) ;pixels_per_line + movsxd rdx, DWORD PTR arg(3) ;out_pitch + movsxd rcx, DWORD PTR arg(4) ;output_height +%endm + +%macro HIGH_APPLY_FILTER_4 1 + + punpcklwd xmm0, xmm1 ;two row in one register + pmaddwd xmm0, xmm4 ;multiply the filter factors + + paddd xmm0, xmm3 ;rounding + psrad xmm0, 7 ;shift + packssdw xmm0, xmm0 ;pack to word + + ;clamp the values + pminsw xmm0, xmm5 + pmaxsw xmm0, xmm2 + +%if %1 + movq xmm1, [rdi] + pavgw xmm0, xmm1 +%endif + + movq [rdi], xmm0 + lea rsi, [rsi + 2*rax] + lea rdi, [rdi + 2*rdx] + dec rcx +%endm + +%if ARCH_X86_64 +%macro HIGH_GET_PARAM 0 + mov rdx, arg(5) ;filter ptr + mov rsi, arg(0) ;src_ptr + mov rdi, arg(2) ;output_ptr + mov rcx, 0x00000040 + + movdqa xmm6, [rdx] ;load filters + + pshuflw xmm7, xmm6, 11111111b ;k3 + pshufhw xmm6, xmm6, 0b ;k4 + psrldq xmm6, 8 + punpcklwd xmm7, xmm6 ;k3k4k3k4k3k4k3k4 + + movq xmm4, rcx ;rounding + pshufd xmm4, xmm4, 0 + + mov rdx, 0x00010001 + movsxd rcx, DWORD PTR arg(6) ;bps + movq xmm8, rdx + movq xmm5, rcx + pshufd xmm8, xmm8, 0b + movdqa xmm1, xmm8 + psllw xmm8, xmm5 + psubw xmm8, xmm1 ;max value (for clamping) + pxor xmm5, xmm5 ;min value (for clamping) + + movsxd rax, DWORD PTR arg(1) ;pixels_per_line + movsxd rdx, DWORD PTR arg(3) ;out_pitch + movsxd rcx, DWORD PTR arg(4) ;output_height +%endm + +%macro HIGH_APPLY_FILTER_8 1 + movdqa xmm6, xmm0 + punpckhwd xmm6, xmm1 + punpcklwd xmm0, xmm1 + pmaddwd xmm6, xmm7 + pmaddwd xmm0, xmm7 + + paddd xmm6, xmm4 ;rounding + paddd xmm0, xmm4 ;rounding + psrad xmm6, 7 ;shift + psrad xmm0, 7 ;shift + packssdw xmm0, xmm6 ;pack back to word + + ;clamp the values + pminsw xmm0, xmm8 + pmaxsw xmm0, xmm5 + +%if %1 + movdqu xmm1, [rdi] + pavgw xmm0, xmm1 +%endif + movdqu [rdi], xmm0 ;store the result + + lea rsi, [rsi + 2*rax] + lea rdi, [rdi + 2*rdx] + dec rcx +%endm + +%macro HIGH_APPLY_FILTER_16 1 + movdqa xmm9, xmm0 + movdqa xmm6, xmm2 + punpckhwd xmm9, xmm1 + punpckhwd xmm6, xmm3 + punpcklwd xmm0, xmm1 + punpcklwd xmm2, xmm3 + + pmaddwd xmm9, xmm7 + pmaddwd xmm6, xmm7 + pmaddwd xmm0, xmm7 + pmaddwd xmm2, xmm7 + + paddd xmm9, xmm4 ;rounding + paddd xmm6, xmm4 + paddd xmm0, xmm4 + paddd xmm2, xmm4 + + psrad xmm9, 7 ;shift + psrad xmm6, 7 + psrad xmm0, 7 + psrad xmm2, 7 + + packssdw xmm0, xmm9 ;pack back to word + packssdw xmm2, xmm6 ;pack back to word + + ;clamp the values + pminsw xmm0, xmm8 + pmaxsw xmm0, xmm5 + pminsw xmm2, xmm8 + pmaxsw xmm2, xmm5 + +%if %1 + movdqu xmm1, [rdi] + movdqu xmm3, [rdi + 16] + pavgw xmm0, xmm1 + pavgw xmm2, xmm3 +%endif + movdqu [rdi], xmm0 ;store the result + movdqu [rdi + 16], xmm2 ;store the result + + lea rsi, [rsi + 2*rax] + lea rdi, [rdi + 2*rdx] + dec rcx +%endm +%endif + +SECTION .text + +global sym(aom_highbd_filter_block1d4_v2_sse2) PRIVATE +sym(aom_highbd_filter_block1d4_v2_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 7 + push rsi + push rdi + ; end prolog + + HIGH_GET_PARAM_4 +.loop: + movq xmm0, [rsi] ;load src + movq xmm1, [rsi + 2*rax] + + HIGH_APPLY_FILTER_4 0 + jnz .loop + + ; begin epilog + pop rdi + pop rsi + UNSHADOW_ARGS + pop rbp + ret + +%if ARCH_X86_64 +global sym(aom_highbd_filter_block1d8_v2_sse2) PRIVATE +sym(aom_highbd_filter_block1d8_v2_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 7 + SAVE_XMM 8 + push rsi + push rdi + ; end prolog + + HIGH_GET_PARAM +.loop: + movdqu xmm0, [rsi] ;0 + movdqu xmm1, [rsi + 2*rax] ;1 + + HIGH_APPLY_FILTER_8 0 + jnz .loop + + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +global sym(aom_highbd_filter_block1d16_v2_sse2) PRIVATE +sym(aom_highbd_filter_block1d16_v2_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 7 + SAVE_XMM 9 + push rsi + push rdi + ; end prolog + + HIGH_GET_PARAM +.loop: + movdqu xmm0, [rsi] ;0 + movdqu xmm2, [rsi + 16] + movdqu xmm1, [rsi + 2*rax] ;1 + movdqu xmm3, [rsi + 2*rax + 16] + + HIGH_APPLY_FILTER_16 0 + jnz .loop + + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret +%endif + +global sym(aom_highbd_filter_block1d4_h2_sse2) PRIVATE +sym(aom_highbd_filter_block1d4_h2_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 7 + push rsi + push rdi + ; end prolog + + HIGH_GET_PARAM_4 +.loop: + movdqu xmm0, [rsi] ;load src + movdqa xmm1, xmm0 + psrldq xmm1, 2 + + HIGH_APPLY_FILTER_4 0 + jnz .loop + + ; begin epilog + pop rdi + pop rsi + UNSHADOW_ARGS + pop rbp + ret + +%if ARCH_X86_64 +global sym(aom_highbd_filter_block1d8_h2_sse2) PRIVATE +sym(aom_highbd_filter_block1d8_h2_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 7 + SAVE_XMM 8 + push rsi + push rdi + ; end prolog + + HIGH_GET_PARAM +.loop: + movdqu xmm0, [rsi] ;load src + movdqu xmm1, [rsi + 2] + + HIGH_APPLY_FILTER_8 0 + jnz .loop + + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +global sym(aom_highbd_filter_block1d16_h2_sse2) PRIVATE +sym(aom_highbd_filter_block1d16_h2_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 7 + SAVE_XMM 9 + push rsi + push rdi + ; end prolog + + HIGH_GET_PARAM +.loop: + movdqu xmm0, [rsi] ;load src + movdqu xmm1, [rsi + 2] + movdqu xmm2, [rsi + 16] + movdqu xmm3, [rsi + 18] + + HIGH_APPLY_FILTER_16 0 + jnz .loop + + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret +%endif diff --git a/third_party/aom/aom_dsp/x86/aom_subpixel_8t_intrin_avx2.c b/third_party/aom/aom_dsp/x86/aom_subpixel_8t_intrin_avx2.c new file mode 100644 index 0000000000..94b5da1714 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/aom_subpixel_8t_intrin_avx2.c @@ -0,0 +1,1441 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include + +#include "config/aom_dsp_rtcd.h" + +#include "aom_dsp/x86/convolve.h" +#include "aom_dsp/x86/convolve_avx2.h" +#include "aom_ports/mem.h" + +#if defined(__clang__) +#if (__clang_major__ > 0 && __clang_major__ < 3) || \ + (__clang_major__ == 3 && __clang_minor__ <= 3) || \ + (defined(__APPLE__) && defined(__apple_build_version__) && \ + ((__clang_major__ == 4 && __clang_minor__ <= 2) || \ + (__clang_major__ == 5 && __clang_minor__ == 0))) +#define MM256_BROADCASTSI128_SI256(x) \ + _mm_broadcastsi128_si256((__m128i const *)&(x)) +#else // clang > 3.3, and not 5.0 on macosx. +#define MM256_BROADCASTSI128_SI256(x) _mm256_broadcastsi128_si256(x) +#endif // clang <= 3.3 +#elif defined(__GNUC__) +#if __GNUC__ < 4 || (__GNUC__ == 4 && __GNUC_MINOR__ <= 6) +#define MM256_BROADCASTSI128_SI256(x) \ + _mm_broadcastsi128_si256((__m128i const *)&(x)) +#elif __GNUC__ == 4 && __GNUC_MINOR__ == 7 +#define MM256_BROADCASTSI128_SI256(x) _mm_broadcastsi128_si256(x) +#else // gcc > 4.7 +#define MM256_BROADCASTSI128_SI256(x) _mm256_broadcastsi128_si256(x) +#endif // gcc <= 4.6 +#else // !(gcc || clang) +#define MM256_BROADCASTSI128_SI256(x) _mm256_broadcastsi128_si256(x) +#endif // __clang__ + +static INLINE void xx_storeu2_epi32(const uint8_t *output_ptr, + const ptrdiff_t stride, const __m256i *a) { + *((uint32_t *)(output_ptr)) = _mm_cvtsi128_si32(_mm256_castsi256_si128(*a)); + *((uint32_t *)(output_ptr + stride)) = + _mm_cvtsi128_si32(_mm256_extracti128_si256(*a, 1)); +} + +static INLINE __m256i xx_loadu2_epi64(const void *hi, const void *lo) { + __m256i a = _mm256_castsi128_si256(_mm_loadl_epi64((const __m128i *)(lo))); + a = _mm256_inserti128_si256(a, _mm_loadl_epi64((const __m128i *)(hi)), 1); + return a; +} + +static INLINE void xx_storeu2_epi64(const uint8_t *output_ptr, + const ptrdiff_t stride, const __m256i *a) { + _mm_storel_epi64((__m128i *)output_ptr, _mm256_castsi256_si128(*a)); + _mm_storel_epi64((__m128i *)(output_ptr + stride), + _mm256_extractf128_si256(*a, 1)); +} + +static INLINE __m256i xx_loadu2_mi128(const void *hi, const void *lo) { + __m256i a = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(lo))); + a = _mm256_inserti128_si256(a, _mm_loadu_si128((const __m128i *)(hi)), 1); + return a; +} + +static INLINE void xx_store2_mi128(const uint8_t *output_ptr, + const ptrdiff_t stride, const __m256i *a) { + _mm_store_si128((__m128i *)output_ptr, _mm256_castsi256_si128(*a)); + _mm_store_si128((__m128i *)(output_ptr + stride), + _mm256_extractf128_si256(*a, 1)); +} + +static void aom_filter_block1d4_h4_avx2( + const uint8_t *src_ptr, ptrdiff_t src_pixels_per_line, uint8_t *output_ptr, + ptrdiff_t output_pitch, uint32_t output_height, const int16_t *filter) { + __m128i filtersReg; + __m256i addFilterReg32, filt1Reg, firstFilters, srcReg32b1, srcRegFilt32b1_1; + unsigned int i; + ptrdiff_t src_stride, dst_stride; + src_ptr -= 3; + addFilterReg32 = _mm256_set1_epi16(32); + filtersReg = _mm_loadu_si128((const __m128i *)filter); + filtersReg = _mm_srai_epi16(filtersReg, 1); + // converting the 16 bit (short) to 8 bit (byte) and have the same data + // in both lanes of 128 bit register. + filtersReg = _mm_packs_epi16(filtersReg, filtersReg); + // have the same data in both lanes of a 256 bit register + const __m256i filtersReg32 = MM256_BROADCASTSI128_SI256(filtersReg); + + firstFilters = + _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi32(0x5040302u)); + filt1Reg = _mm256_load_si256((__m256i const *)(filt4_d4_global_avx2)); + + // multiple the size of the source and destination stride by two + src_stride = src_pixels_per_line << 1; + dst_stride = output_pitch << 1; + for (i = output_height; i > 1; i -= 2) { + // load the 2 strides of source + srcReg32b1 = xx_loadu2_mi128(src_ptr + src_pixels_per_line, src_ptr); + + // filter the source buffer + srcRegFilt32b1_1 = _mm256_shuffle_epi8(srcReg32b1, filt1Reg); + + // multiply 4 adjacent elements with the filter and add the result + srcRegFilt32b1_1 = _mm256_maddubs_epi16(srcRegFilt32b1_1, firstFilters); + + srcRegFilt32b1_1 = + _mm256_hadds_epi16(srcRegFilt32b1_1, _mm256_setzero_si256()); + + // shift by 6 bit each 16 bit + srcRegFilt32b1_1 = _mm256_adds_epi16(srcRegFilt32b1_1, addFilterReg32); + srcRegFilt32b1_1 = _mm256_srai_epi16(srcRegFilt32b1_1, 6); + + // shrink to 8 bit each 16 bits, the first lane contain the first + // convolve result and the second lane contain the second convolve result + srcRegFilt32b1_1 = + _mm256_packus_epi16(srcRegFilt32b1_1, _mm256_setzero_si256()); + + src_ptr += src_stride; + + xx_storeu2_epi32(output_ptr, output_pitch, &srcRegFilt32b1_1); + output_ptr += dst_stride; + } + + // if the number of strides is odd. + // process only 4 bytes + if (i > 0) { + __m128i srcReg1, srcRegFilt1_1; + + srcReg1 = _mm_loadu_si128((const __m128i *)(src_ptr)); + + // filter the source buffer + srcRegFilt1_1 = _mm_shuffle_epi8(srcReg1, _mm256_castsi256_si128(filt1Reg)); + + // multiply 4 adjacent elements with the filter and add the result + srcRegFilt1_1 = + _mm_maddubs_epi16(srcRegFilt1_1, _mm256_castsi256_si128(firstFilters)); + + srcRegFilt1_1 = _mm_hadds_epi16(srcRegFilt1_1, _mm_setzero_si128()); + // shift by 6 bit each 16 bit + srcRegFilt1_1 = + _mm_adds_epi16(srcRegFilt1_1, _mm256_castsi256_si128(addFilterReg32)); + srcRegFilt1_1 = _mm_srai_epi16(srcRegFilt1_1, 6); + + // shrink to 8 bit each 16 bits, the first lane contain the first + // convolve result and the second lane contain the second convolve result + srcRegFilt1_1 = _mm_packus_epi16(srcRegFilt1_1, _mm_setzero_si128()); + + // save 4 bytes + *((uint32_t *)(output_ptr)) = _mm_cvtsi128_si32(srcRegFilt1_1); + } +} + +static void aom_filter_block1d4_h8_avx2( + const uint8_t *src_ptr, ptrdiff_t src_pixels_per_line, uint8_t *output_ptr, + ptrdiff_t output_pitch, uint32_t output_height, const int16_t *filter) { + __m128i filtersReg; + __m256i addFilterReg32, filt1Reg, filt2Reg; + __m256i firstFilters, secondFilters; + __m256i srcRegFilt32b1_1, srcRegFilt32b2; + __m256i srcReg32b1; + unsigned int i; + ptrdiff_t src_stride, dst_stride; + src_ptr -= 3; + addFilterReg32 = _mm256_set1_epi16(32); + filtersReg = _mm_loadu_si128((const __m128i *)filter); + filtersReg = _mm_srai_epi16(filtersReg, 1); + // converting the 16 bit (short) to 8 bit (byte) and have the same data + // in both lanes of 128 bit register. + filtersReg = _mm_packs_epi16(filtersReg, filtersReg); + // have the same data in both lanes of a 256 bit register + const __m256i filtersReg32 = MM256_BROADCASTSI128_SI256(filtersReg); + + // duplicate only the first 32 bits + firstFilters = _mm256_shuffle_epi32(filtersReg32, 0); + // duplicate only the second 32 bits + secondFilters = _mm256_shuffle_epi32(filtersReg32, 0x55); + + filt1Reg = _mm256_load_si256((__m256i const *)filt_d4_global_avx2); + filt2Reg = _mm256_load_si256((__m256i const *)(filt_d4_global_avx2 + 32)); + + // multiple the size of the source and destination stride by two + src_stride = src_pixels_per_line << 1; + dst_stride = output_pitch << 1; + for (i = output_height; i > 1; i -= 2) { + // load the 2 strides of source + srcReg32b1 = xx_loadu2_mi128(src_ptr + src_pixels_per_line, src_ptr); + + // filter the source buffer + srcRegFilt32b1_1 = _mm256_shuffle_epi8(srcReg32b1, filt1Reg); + + // multiply 4 adjacent elements with the filter and add the result + srcRegFilt32b1_1 = _mm256_maddubs_epi16(srcRegFilt32b1_1, firstFilters); + + // filter the source buffer + srcRegFilt32b2 = _mm256_shuffle_epi8(srcReg32b1, filt2Reg); + + // multiply 4 adjacent elements with the filter and add the result + srcRegFilt32b2 = _mm256_maddubs_epi16(srcRegFilt32b2, secondFilters); + + srcRegFilt32b1_1 = _mm256_adds_epi16(srcRegFilt32b1_1, srcRegFilt32b2); + + srcRegFilt32b1_1 = + _mm256_hadds_epi16(srcRegFilt32b1_1, _mm256_setzero_si256()); + + // shift by 6 bit each 16 bit + srcRegFilt32b1_1 = _mm256_adds_epi16(srcRegFilt32b1_1, addFilterReg32); + srcRegFilt32b1_1 = _mm256_srai_epi16(srcRegFilt32b1_1, 6); + + // shrink to 8 bit each 16 bits, the first lane contain the first + // convolve result and the second lane contain the second convolve result + srcRegFilt32b1_1 = + _mm256_packus_epi16(srcRegFilt32b1_1, _mm256_setzero_si256()); + + src_ptr += src_stride; + + xx_storeu2_epi32(output_ptr, output_pitch, &srcRegFilt32b1_1); + output_ptr += dst_stride; + } + + // if the number of strides is odd. + // process only 4 bytes + if (i > 0) { + __m128i srcReg1, srcRegFilt1_1; + __m128i srcRegFilt2; + + srcReg1 = _mm_loadu_si128((const __m128i *)(src_ptr)); + + // filter the source buffer + srcRegFilt1_1 = _mm_shuffle_epi8(srcReg1, _mm256_castsi256_si128(filt1Reg)); + + // multiply 4 adjacent elements with the filter and add the result + srcRegFilt1_1 = + _mm_maddubs_epi16(srcRegFilt1_1, _mm256_castsi256_si128(firstFilters)); + + // filter the source buffer + srcRegFilt2 = _mm_shuffle_epi8(srcReg1, _mm256_castsi256_si128(filt2Reg)); + + // multiply 4 adjacent elements with the filter and add the result + srcRegFilt2 = + _mm_maddubs_epi16(srcRegFilt2, _mm256_castsi256_si128(secondFilters)); + + srcRegFilt1_1 = _mm_adds_epi16(srcRegFilt1_1, srcRegFilt2); + srcRegFilt1_1 = _mm_hadds_epi16(srcRegFilt1_1, _mm_setzero_si128()); + // shift by 6 bit each 16 bit + srcRegFilt1_1 = + _mm_adds_epi16(srcRegFilt1_1, _mm256_castsi256_si128(addFilterReg32)); + srcRegFilt1_1 = _mm_srai_epi16(srcRegFilt1_1, 6); + + // shrink to 8 bit each 16 bits, the first lane contain the first + // convolve result and the second lane contain the second convolve result + srcRegFilt1_1 = _mm_packus_epi16(srcRegFilt1_1, _mm_setzero_si128()); + + // save 4 bytes + *((uint32_t *)(output_ptr)) = _mm_cvtsi128_si32(srcRegFilt1_1); + } +} + +static void aom_filter_block1d8_h4_avx2( + const uint8_t *src_ptr, ptrdiff_t src_pixels_per_line, uint8_t *output_ptr, + ptrdiff_t output_pitch, uint32_t output_height, const int16_t *filter) { + __m128i filtersReg; + __m256i addFilterReg32, filt2Reg, filt3Reg; + __m256i secondFilters, thirdFilters; + __m256i srcRegFilt32b1_1, srcRegFilt32b2, srcRegFilt32b3; + __m256i srcReg32b1, filtersReg32; + unsigned int i; + ptrdiff_t src_stride, dst_stride; + src_ptr -= 3; + addFilterReg32 = _mm256_set1_epi16(32); + filtersReg = _mm_loadu_si128((const __m128i *)filter); + filtersReg = _mm_srai_epi16(filtersReg, 1); + // converting the 16 bit (short) to 8 bit (byte) and have the same data + // in both lanes of 128 bit register. + filtersReg = _mm_packs_epi16(filtersReg, filtersReg); + // have the same data in both lanes of a 256 bit register + filtersReg32 = MM256_BROADCASTSI128_SI256(filtersReg); + + // duplicate only the second 16 bits (third and forth byte) + // across 256 bit register + secondFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x302u)); + // duplicate only the third 16 bits (fifth and sixth byte) + // across 256 bit register + thirdFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x504u)); + + filt2Reg = _mm256_load_si256((__m256i const *)(filt_global_avx2 + 32)); + filt3Reg = _mm256_load_si256((__m256i const *)(filt_global_avx2 + 32 * 2)); + + // multiply the size of the source and destination stride by two + src_stride = src_pixels_per_line << 1; + dst_stride = output_pitch << 1; + for (i = output_height; i > 1; i -= 2) { + // load the 2 strides of source + srcReg32b1 = xx_loadu2_mi128(src_ptr + src_pixels_per_line, src_ptr); + + // filter the source buffer + srcRegFilt32b3 = _mm256_shuffle_epi8(srcReg32b1, filt2Reg); + srcRegFilt32b2 = _mm256_shuffle_epi8(srcReg32b1, filt3Reg); + + // multiply 2 adjacent elements with the filter and add the result + srcRegFilt32b3 = _mm256_maddubs_epi16(srcRegFilt32b3, secondFilters); + srcRegFilt32b2 = _mm256_maddubs_epi16(srcRegFilt32b2, thirdFilters); + + srcRegFilt32b1_1 = _mm256_adds_epi16(srcRegFilt32b3, srcRegFilt32b2); + + // shift by 6 bit each 16 bit + srcRegFilt32b1_1 = _mm256_adds_epi16(srcRegFilt32b1_1, addFilterReg32); + srcRegFilt32b1_1 = _mm256_srai_epi16(srcRegFilt32b1_1, 6); + + // shrink to 8 bit each 16 bits + srcRegFilt32b1_1 = _mm256_packus_epi16(srcRegFilt32b1_1, srcRegFilt32b1_1); + + src_ptr += src_stride; + + xx_storeu2_epi64(output_ptr, output_pitch, &srcRegFilt32b1_1); + output_ptr += dst_stride; + } + + // if the number of strides is odd. + // process only 8 bytes + if (i > 0) { + __m128i srcReg1, srcRegFilt1_1; + __m128i srcRegFilt2, srcRegFilt3; + + srcReg1 = _mm_loadu_si128((const __m128i *)(src_ptr)); + + // filter the source buffer + srcRegFilt2 = _mm_shuffle_epi8(srcReg1, _mm256_castsi256_si128(filt2Reg)); + srcRegFilt3 = _mm_shuffle_epi8(srcReg1, _mm256_castsi256_si128(filt3Reg)); + + // multiply 2 adjacent elements with the filter and add the result + srcRegFilt2 = + _mm_maddubs_epi16(srcRegFilt2, _mm256_castsi256_si128(secondFilters)); + srcRegFilt3 = + _mm_maddubs_epi16(srcRegFilt3, _mm256_castsi256_si128(thirdFilters)); + + // add and saturate the results together + srcRegFilt1_1 = _mm_adds_epi16(srcRegFilt2, srcRegFilt3); + + // shift by 6 bit each 16 bit + srcRegFilt1_1 = + _mm_adds_epi16(srcRegFilt1_1, _mm256_castsi256_si128(addFilterReg32)); + srcRegFilt1_1 = _mm_srai_epi16(srcRegFilt1_1, 6); + + // shrink to 8 bit each 16 bits + srcRegFilt1_1 = _mm_packus_epi16(srcRegFilt1_1, _mm_setzero_si128()); + + // save 8 bytes + _mm_storel_epi64((__m128i *)output_ptr, srcRegFilt1_1); + } +} + +static void aom_filter_block1d8_h8_avx2( + const uint8_t *src_ptr, ptrdiff_t src_pixels_per_line, uint8_t *output_ptr, + ptrdiff_t output_pitch, uint32_t output_height, const int16_t *filter) { + __m128i filtersReg; + __m256i addFilterReg32, filt1Reg, filt2Reg, filt3Reg, filt4Reg; + __m256i firstFilters, secondFilters, thirdFilters, forthFilters; + __m256i srcRegFilt32b1_1, srcRegFilt32b2, srcRegFilt32b3; + __m256i srcReg32b1; + unsigned int i; + ptrdiff_t src_stride, dst_stride; + src_ptr -= 3; + addFilterReg32 = _mm256_set1_epi16(32); + filtersReg = _mm_loadu_si128((const __m128i *)filter); + filtersReg = _mm_srai_epi16(filtersReg, 1); + // converting the 16 bit (short) to 8 bit (byte) and have the same data + // in both lanes of 128 bit register. + filtersReg = _mm_packs_epi16(filtersReg, filtersReg); + // have the same data in both lanes of a 256 bit register + const __m256i filtersReg32 = MM256_BROADCASTSI128_SI256(filtersReg); + + // duplicate only the first 16 bits (first and second byte) + // across 256 bit register + firstFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x100u)); + // duplicate only the second 16 bits (third and forth byte) + // across 256 bit register + secondFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x302u)); + // duplicate only the third 16 bits (fifth and sixth byte) + // across 256 bit register + thirdFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x504u)); + // duplicate only the forth 16 bits (seventh and eighth byte) + // across 256 bit register + forthFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x706u)); + + filt1Reg = _mm256_load_si256((__m256i const *)filt_global_avx2); + filt2Reg = _mm256_load_si256((__m256i const *)(filt_global_avx2 + 32)); + filt3Reg = _mm256_load_si256((__m256i const *)(filt_global_avx2 + 32 * 2)); + filt4Reg = _mm256_load_si256((__m256i const *)(filt_global_avx2 + 32 * 3)); + + // multiple the size of the source and destination stride by two + src_stride = src_pixels_per_line << 1; + dst_stride = output_pitch << 1; + for (i = output_height; i > 1; i -= 2) { + // load the 2 strides of source + srcReg32b1 = xx_loadu2_mi128(src_ptr + src_pixels_per_line, src_ptr); + + // filter the source buffer + srcRegFilt32b1_1 = _mm256_shuffle_epi8(srcReg32b1, filt1Reg); + srcRegFilt32b2 = _mm256_shuffle_epi8(srcReg32b1, filt4Reg); + + // multiply 2 adjacent elements with the filter and add the result + srcRegFilt32b1_1 = _mm256_maddubs_epi16(srcRegFilt32b1_1, firstFilters); + srcRegFilt32b2 = _mm256_maddubs_epi16(srcRegFilt32b2, forthFilters); + + // add and saturate the results together + srcRegFilt32b1_1 = _mm256_adds_epi16(srcRegFilt32b1_1, srcRegFilt32b2); + + // filter the source buffer + srcRegFilt32b3 = _mm256_shuffle_epi8(srcReg32b1, filt2Reg); + srcRegFilt32b2 = _mm256_shuffle_epi8(srcReg32b1, filt3Reg); + + // multiply 2 adjacent elements with the filter and add the result + srcRegFilt32b3 = _mm256_maddubs_epi16(srcRegFilt32b3, secondFilters); + srcRegFilt32b2 = _mm256_maddubs_epi16(srcRegFilt32b2, thirdFilters); + + __m256i sum23 = _mm256_adds_epi16(srcRegFilt32b3, srcRegFilt32b2); + srcRegFilt32b1_1 = _mm256_adds_epi16(srcRegFilt32b1_1, sum23); + + // shift by 6 bit each 16 bit + srcRegFilt32b1_1 = _mm256_adds_epi16(srcRegFilt32b1_1, addFilterReg32); + srcRegFilt32b1_1 = _mm256_srai_epi16(srcRegFilt32b1_1, 6); + + // shrink to 8 bit each 16 bits, the first lane contain the first + // convolve result and the second lane contain the second convolve result + srcRegFilt32b1_1 = + _mm256_packus_epi16(srcRegFilt32b1_1, _mm256_setzero_si256()); + + src_ptr += src_stride; + + xx_storeu2_epi64(output_ptr, output_pitch, &srcRegFilt32b1_1); + output_ptr += dst_stride; + } + + // if the number of strides is odd. + // process only 8 bytes + if (i > 0) { + __m128i srcReg1, srcRegFilt1_1; + __m128i srcRegFilt2, srcRegFilt3; + + srcReg1 = _mm_loadu_si128((const __m128i *)(src_ptr)); + + // filter the source buffer + srcRegFilt1_1 = _mm_shuffle_epi8(srcReg1, _mm256_castsi256_si128(filt1Reg)); + srcRegFilt2 = _mm_shuffle_epi8(srcReg1, _mm256_castsi256_si128(filt4Reg)); + + // multiply 2 adjacent elements with the filter and add the result + srcRegFilt1_1 = + _mm_maddubs_epi16(srcRegFilt1_1, _mm256_castsi256_si128(firstFilters)); + srcRegFilt2 = + _mm_maddubs_epi16(srcRegFilt2, _mm256_castsi256_si128(forthFilters)); + + // add and saturate the results together + srcRegFilt1_1 = _mm_adds_epi16(srcRegFilt1_1, srcRegFilt2); + + // filter the source buffer + srcRegFilt3 = _mm_shuffle_epi8(srcReg1, _mm256_castsi256_si128(filt2Reg)); + srcRegFilt2 = _mm_shuffle_epi8(srcReg1, _mm256_castsi256_si128(filt3Reg)); + + // multiply 2 adjacent elements with the filter and add the result + srcRegFilt3 = + _mm_maddubs_epi16(srcRegFilt3, _mm256_castsi256_si128(secondFilters)); + srcRegFilt2 = + _mm_maddubs_epi16(srcRegFilt2, _mm256_castsi256_si128(thirdFilters)); + + // add and saturate the results together + srcRegFilt1_1 = + _mm_adds_epi16(srcRegFilt1_1, _mm_adds_epi16(srcRegFilt3, srcRegFilt2)); + + // shift by 6 bit each 16 bit + srcRegFilt1_1 = + _mm_adds_epi16(srcRegFilt1_1, _mm256_castsi256_si128(addFilterReg32)); + srcRegFilt1_1 = _mm_srai_epi16(srcRegFilt1_1, 6); + + // shrink to 8 bit each 16 bits, the first lane contain the first + // convolve result and the second lane contain the second convolve + // result + srcRegFilt1_1 = _mm_packus_epi16(srcRegFilt1_1, _mm_setzero_si128()); + + // save 8 bytes + _mm_storel_epi64((__m128i *)output_ptr, srcRegFilt1_1); + } +} + +static void aom_filter_block1d16_h4_avx2( + const uint8_t *src_ptr, ptrdiff_t src_pixels_per_line, uint8_t *output_ptr, + ptrdiff_t output_pitch, uint32_t output_height, const int16_t *filter) { + __m128i filtersReg; + __m256i addFilterReg32, filt2Reg, filt3Reg; + __m256i secondFilters, thirdFilters; + __m256i srcRegFilt32b1_1, srcRegFilt32b2_1, srcRegFilt32b2, srcRegFilt32b3; + __m256i srcReg32b1, srcReg32b2, filtersReg32; + unsigned int i; + ptrdiff_t src_stride, dst_stride; + src_ptr -= 3; + addFilterReg32 = _mm256_set1_epi16(32); + filtersReg = _mm_loadu_si128((const __m128i *)filter); + filtersReg = _mm_srai_epi16(filtersReg, 1); + // converting the 16 bit (short) to 8 bit (byte) and have the same data + // in both lanes of 128 bit register. + filtersReg = _mm_packs_epi16(filtersReg, filtersReg); + // have the same data in both lanes of a 256 bit register + filtersReg32 = MM256_BROADCASTSI128_SI256(filtersReg); + + // duplicate only the second 16 bits (third and forth byte) + // across 256 bit register + secondFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x302u)); + // duplicate only the third 16 bits (fifth and sixth byte) + // across 256 bit register + thirdFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x504u)); + + filt2Reg = _mm256_load_si256((__m256i const *)(filt_global_avx2 + 32)); + filt3Reg = _mm256_load_si256((__m256i const *)(filt_global_avx2 + 32 * 2)); + + // multiply the size of the source and destination stride by two + src_stride = src_pixels_per_line << 1; + dst_stride = output_pitch << 1; + for (i = output_height; i > 1; i -= 2) { + // load the 2 strides of source + srcReg32b1 = xx_loadu2_mi128(src_ptr + src_pixels_per_line, src_ptr); + + // filter the source buffer + srcRegFilt32b3 = _mm256_shuffle_epi8(srcReg32b1, filt2Reg); + srcRegFilt32b2 = _mm256_shuffle_epi8(srcReg32b1, filt3Reg); + + // multiply 2 adjacent elements with the filter and add the result + srcRegFilt32b3 = _mm256_maddubs_epi16(srcRegFilt32b3, secondFilters); + srcRegFilt32b2 = _mm256_maddubs_epi16(srcRegFilt32b2, thirdFilters); + + srcRegFilt32b1_1 = _mm256_adds_epi16(srcRegFilt32b3, srcRegFilt32b2); + + // reading 2 strides of the next 16 bytes + // (part of it was being read by earlier read) + srcReg32b2 = + xx_loadu2_mi128(src_ptr + src_pixels_per_line + 8, src_ptr + 8); + + // filter the source buffer + srcRegFilt32b3 = _mm256_shuffle_epi8(srcReg32b2, filt2Reg); + srcRegFilt32b2 = _mm256_shuffle_epi8(srcReg32b2, filt3Reg); + + // multiply 2 adjacent elements with the filter and add the result + srcRegFilt32b3 = _mm256_maddubs_epi16(srcRegFilt32b3, secondFilters); + srcRegFilt32b2 = _mm256_maddubs_epi16(srcRegFilt32b2, thirdFilters); + + // add and saturate the results together + srcRegFilt32b2_1 = _mm256_adds_epi16(srcRegFilt32b3, srcRegFilt32b2); + + // shift by 6 bit each 16 bit + srcRegFilt32b1_1 = _mm256_adds_epi16(srcRegFilt32b1_1, addFilterReg32); + srcRegFilt32b2_1 = _mm256_adds_epi16(srcRegFilt32b2_1, addFilterReg32); + srcRegFilt32b1_1 = _mm256_srai_epi16(srcRegFilt32b1_1, 6); + srcRegFilt32b2_1 = _mm256_srai_epi16(srcRegFilt32b2_1, 6); + + // shrink to 8 bit each 16 bits, the first lane contain the first + // convolve result and the second lane contain the second convolve result + srcRegFilt32b1_1 = _mm256_packus_epi16(srcRegFilt32b1_1, srcRegFilt32b2_1); + + src_ptr += src_stride; + + xx_store2_mi128(output_ptr, output_pitch, &srcRegFilt32b1_1); + output_ptr += dst_stride; + } + + // if the number of strides is odd. + // process only 16 bytes + if (i > 0) { + __m256i srcReg1, srcReg12; + __m256i srcRegFilt2, srcRegFilt3, srcRegFilt1_1; + + srcReg1 = _mm256_loadu_si256((const __m256i *)(src_ptr)); + srcReg12 = _mm256_permute4x64_epi64(srcReg1, 0x94); + + // filter the source buffer + srcRegFilt2 = _mm256_shuffle_epi8(srcReg12, filt2Reg); + srcRegFilt3 = _mm256_shuffle_epi8(srcReg12, filt3Reg); + + // multiply 2 adjacent elements with the filter and add the result + srcRegFilt2 = _mm256_maddubs_epi16(srcRegFilt2, secondFilters); + srcRegFilt3 = _mm256_maddubs_epi16(srcRegFilt3, thirdFilters); + + // add and saturate the results together + srcRegFilt1_1 = _mm256_adds_epi16(srcRegFilt2, srcRegFilt3); + + // shift by 6 bit each 16 bit + srcRegFilt1_1 = _mm256_adds_epi16(srcRegFilt1_1, addFilterReg32); + srcRegFilt1_1 = _mm256_srai_epi16(srcRegFilt1_1, 6); + + // shrink to 8 bit each 16 bits, the first lane contain the first + // convolve result and the second lane contain the second convolve + // result + srcRegFilt1_1 = _mm256_packus_epi16(srcRegFilt1_1, srcRegFilt1_1); + srcRegFilt1_1 = _mm256_permute4x64_epi64(srcRegFilt1_1, 0x8); + + // save 16 bytes + _mm_store_si128((__m128i *)output_ptr, + _mm256_castsi256_si128(srcRegFilt1_1)); + } +} + +static void aom_filter_block1d16_h8_avx2( + const uint8_t *src_ptr, ptrdiff_t src_pixels_per_line, uint8_t *output_ptr, + ptrdiff_t output_pitch, uint32_t output_height, const int16_t *filter) { + __m128i filtersReg; + __m256i addFilterReg32, filt1Reg, filt2Reg, filt3Reg, filt4Reg; + __m256i firstFilters, secondFilters, thirdFilters, forthFilters; + __m256i srcRegFilt32b1_1, srcRegFilt32b2_1, srcRegFilt32b2, srcRegFilt32b3; + __m256i srcReg32b1, srcReg32b2, filtersReg32; + unsigned int i; + ptrdiff_t src_stride, dst_stride; + src_ptr -= 3; + addFilterReg32 = _mm256_set1_epi16(32); + filtersReg = _mm_loadu_si128((const __m128i *)filter); + filtersReg = _mm_srai_epi16(filtersReg, 1); + // converting the 16 bit (short) to 8 bit (byte) and have the same data + // in both lanes of 128 bit register. + filtersReg = _mm_packs_epi16(filtersReg, filtersReg); + // have the same data in both lanes of a 256 bit register + filtersReg32 = MM256_BROADCASTSI128_SI256(filtersReg); + + // duplicate only the first 16 bits (first and second byte) + // across 256 bit register + firstFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x100u)); + // duplicate only the second 16 bits (third and forth byte) + // across 256 bit register + secondFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x302u)); + // duplicate only the third 16 bits (fifth and sixth byte) + // across 256 bit register + thirdFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x504u)); + // duplicate only the forth 16 bits (seventh and eighth byte) + // across 256 bit register + forthFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x706u)); + + filt1Reg = _mm256_load_si256((__m256i const *)filt_global_avx2); + filt2Reg = _mm256_load_si256((__m256i const *)(filt_global_avx2 + 32)); + filt3Reg = _mm256_load_si256((__m256i const *)(filt_global_avx2 + 32 * 2)); + filt4Reg = _mm256_load_si256((__m256i const *)(filt_global_avx2 + 32 * 3)); + + // multiple the size of the source and destination stride by two + src_stride = src_pixels_per_line << 1; + dst_stride = output_pitch << 1; + for (i = output_height; i > 1; i -= 2) { + // load the 2 strides of source + srcReg32b1 = xx_loadu2_mi128(src_ptr + src_pixels_per_line, src_ptr); + + // filter the source buffer + srcRegFilt32b1_1 = _mm256_shuffle_epi8(srcReg32b1, filt1Reg); + srcRegFilt32b2 = _mm256_shuffle_epi8(srcReg32b1, filt4Reg); + + // multiply 2 adjacent elements with the filter and add the result + srcRegFilt32b1_1 = _mm256_maddubs_epi16(srcRegFilt32b1_1, firstFilters); + srcRegFilt32b2 = _mm256_maddubs_epi16(srcRegFilt32b2, forthFilters); + + // add and saturate the results together + srcRegFilt32b1_1 = _mm256_adds_epi16(srcRegFilt32b1_1, srcRegFilt32b2); + + // filter the source buffer + srcRegFilt32b3 = _mm256_shuffle_epi8(srcReg32b1, filt2Reg); + srcRegFilt32b2 = _mm256_shuffle_epi8(srcReg32b1, filt3Reg); + + // multiply 2 adjacent elements with the filter and add the result + srcRegFilt32b3 = _mm256_maddubs_epi16(srcRegFilt32b3, secondFilters); + srcRegFilt32b2 = _mm256_maddubs_epi16(srcRegFilt32b2, thirdFilters); + + __m256i sum23 = _mm256_adds_epi16(srcRegFilt32b3, srcRegFilt32b2); + srcRegFilt32b1_1 = _mm256_adds_epi16(srcRegFilt32b1_1, sum23); + + // reading 2 strides of the next 16 bytes + // (part of it was being read by earlier read) + srcReg32b2 = + xx_loadu2_mi128(src_ptr + src_pixels_per_line + 8, src_ptr + 8); + + // filter the source buffer + srcRegFilt32b2_1 = _mm256_shuffle_epi8(srcReg32b2, filt1Reg); + srcRegFilt32b2 = _mm256_shuffle_epi8(srcReg32b2, filt4Reg); + + // multiply 2 adjacent elements with the filter and add the result + srcRegFilt32b2_1 = _mm256_maddubs_epi16(srcRegFilt32b2_1, firstFilters); + srcRegFilt32b2 = _mm256_maddubs_epi16(srcRegFilt32b2, forthFilters); + + // add and saturate the results together + srcRegFilt32b2_1 = _mm256_adds_epi16(srcRegFilt32b2_1, srcRegFilt32b2); + + // filter the source buffer + srcRegFilt32b3 = _mm256_shuffle_epi8(srcReg32b2, filt2Reg); + srcRegFilt32b2 = _mm256_shuffle_epi8(srcReg32b2, filt3Reg); + + // multiply 2 adjacent elements with the filter and add the result + srcRegFilt32b3 = _mm256_maddubs_epi16(srcRegFilt32b3, secondFilters); + srcRegFilt32b2 = _mm256_maddubs_epi16(srcRegFilt32b2, thirdFilters); + + // add and saturate the results together + srcRegFilt32b2_1 = _mm256_adds_epi16( + srcRegFilt32b2_1, _mm256_adds_epi16(srcRegFilt32b3, srcRegFilt32b2)); + + // shift by 6 bit each 16 bit + srcRegFilt32b1_1 = _mm256_adds_epi16(srcRegFilt32b1_1, addFilterReg32); + srcRegFilt32b2_1 = _mm256_adds_epi16(srcRegFilt32b2_1, addFilterReg32); + srcRegFilt32b1_1 = _mm256_srai_epi16(srcRegFilt32b1_1, 6); + srcRegFilt32b2_1 = _mm256_srai_epi16(srcRegFilt32b2_1, 6); + + // shrink to 8 bit each 16 bits, the first lane contain the first + // convolve result and the second lane contain the second convolve result + srcRegFilt32b1_1 = _mm256_packus_epi16(srcRegFilt32b1_1, srcRegFilt32b2_1); + + src_ptr += src_stride; + + xx_store2_mi128(output_ptr, output_pitch, &srcRegFilt32b1_1); + output_ptr += dst_stride; + } + + // if the number of strides is odd. + // process only 16 bytes + if (i > 0) { + __m128i srcReg1, srcReg2, srcRegFilt1_1, srcRegFilt2_1; + __m128i srcRegFilt2, srcRegFilt3; + + srcReg1 = _mm_loadu_si128((const __m128i *)(src_ptr)); + + // filter the source buffer + srcRegFilt1_1 = _mm_shuffle_epi8(srcReg1, _mm256_castsi256_si128(filt1Reg)); + srcRegFilt2 = _mm_shuffle_epi8(srcReg1, _mm256_castsi256_si128(filt4Reg)); + + // multiply 2 adjacent elements with the filter and add the result + srcRegFilt1_1 = + _mm_maddubs_epi16(srcRegFilt1_1, _mm256_castsi256_si128(firstFilters)); + srcRegFilt2 = + _mm_maddubs_epi16(srcRegFilt2, _mm256_castsi256_si128(forthFilters)); + + // add and saturate the results together + srcRegFilt1_1 = _mm_adds_epi16(srcRegFilt1_1, srcRegFilt2); + + // filter the source buffer + srcRegFilt3 = _mm_shuffle_epi8(srcReg1, _mm256_castsi256_si128(filt2Reg)); + srcRegFilt2 = _mm_shuffle_epi8(srcReg1, _mm256_castsi256_si128(filt3Reg)); + + // multiply 2 adjacent elements with the filter and add the result + srcRegFilt3 = + _mm_maddubs_epi16(srcRegFilt3, _mm256_castsi256_si128(secondFilters)); + srcRegFilt2 = + _mm_maddubs_epi16(srcRegFilt2, _mm256_castsi256_si128(thirdFilters)); + + // add and saturate the results together + srcRegFilt1_1 = + _mm_adds_epi16(srcRegFilt1_1, _mm_adds_epi16(srcRegFilt3, srcRegFilt2)); + + // reading the next 16 bytes + // (part of it was being read by earlier read) + srcReg2 = _mm_loadu_si128((const __m128i *)(src_ptr + 8)); + + // filter the source buffer + srcRegFilt2_1 = _mm_shuffle_epi8(srcReg2, _mm256_castsi256_si128(filt1Reg)); + srcRegFilt2 = _mm_shuffle_epi8(srcReg2, _mm256_castsi256_si128(filt4Reg)); + + // multiply 2 adjacent elements with the filter and add the result + srcRegFilt2_1 = + _mm_maddubs_epi16(srcRegFilt2_1, _mm256_castsi256_si128(firstFilters)); + srcRegFilt2 = + _mm_maddubs_epi16(srcRegFilt2, _mm256_castsi256_si128(forthFilters)); + + // add and saturate the results together + srcRegFilt2_1 = _mm_adds_epi16(srcRegFilt2_1, srcRegFilt2); + + // filter the source buffer + srcRegFilt3 = _mm_shuffle_epi8(srcReg2, _mm256_castsi256_si128(filt2Reg)); + srcRegFilt2 = _mm_shuffle_epi8(srcReg2, _mm256_castsi256_si128(filt3Reg)); + + // multiply 2 adjacent elements with the filter and add the result + srcRegFilt3 = + _mm_maddubs_epi16(srcRegFilt3, _mm256_castsi256_si128(secondFilters)); + srcRegFilt2 = + _mm_maddubs_epi16(srcRegFilt2, _mm256_castsi256_si128(thirdFilters)); + + // add and saturate the results together + srcRegFilt2_1 = + _mm_adds_epi16(srcRegFilt2_1, _mm_adds_epi16(srcRegFilt3, srcRegFilt2)); + + // shift by 6 bit each 16 bit + srcRegFilt1_1 = + _mm_adds_epi16(srcRegFilt1_1, _mm256_castsi256_si128(addFilterReg32)); + srcRegFilt1_1 = _mm_srai_epi16(srcRegFilt1_1, 6); + + srcRegFilt2_1 = + _mm_adds_epi16(srcRegFilt2_1, _mm256_castsi256_si128(addFilterReg32)); + srcRegFilt2_1 = _mm_srai_epi16(srcRegFilt2_1, 6); + + // shrink to 8 bit each 16 bits, the first lane contain the first + // convolve result and the second lane contain the second convolve + // result + srcRegFilt1_1 = _mm_packus_epi16(srcRegFilt1_1, srcRegFilt2_1); + + // save 16 bytes + _mm_store_si128((__m128i *)output_ptr, srcRegFilt1_1); + } +} + +static void aom_filter_block1d8_v4_avx2( + const uint8_t *src_ptr, ptrdiff_t src_pitch, uint8_t *output_ptr, + ptrdiff_t out_pitch, uint32_t output_height, const int16_t *filter) { + __m128i filtersReg; + __m256i filtersReg32, addFilterReg32; + __m256i srcReg23, srcReg4x, srcReg34, srcReg5x, srcReg45, srcReg6x, srcReg56; + __m256i srcReg23_34_lo, srcReg45_56_lo; + __m256i resReg23_34_lo, resReg45_56_lo; + __m256i resReglo, resReg; + __m256i secondFilters, thirdFilters; + unsigned int i; + ptrdiff_t src_stride, dst_stride; + + addFilterReg32 = _mm256_set1_epi16(32); + filtersReg = _mm_loadu_si128((const __m128i *)filter); + // converting the 16 bit (short) to 8 bit (byte) and have the + // same data in both lanes of 128 bit register. + filtersReg = _mm_srai_epi16(filtersReg, 1); + filtersReg = _mm_packs_epi16(filtersReg, filtersReg); + // have the same data in both lanes of a 256 bit register + filtersReg32 = MM256_BROADCASTSI128_SI256(filtersReg); + + // duplicate only the second 16 bits (third and forth byte) + // across 256 bit register + secondFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x302u)); + // duplicate only the third 16 bits (fifth and sixth byte) + // across 256 bit register + thirdFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x504u)); + + // multiple the size of the source and destination stride by two + src_stride = src_pitch << 1; + dst_stride = out_pitch << 1; + + srcReg23 = xx_loadu2_epi64(src_ptr + src_pitch * 3, src_ptr + src_pitch * 2); + srcReg4x = _mm256_castsi128_si256( + _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 4))); + + // have consecutive loads on the same 256 register + srcReg34 = _mm256_permute2x128_si256(srcReg23, srcReg4x, 0x21); + + srcReg23_34_lo = _mm256_unpacklo_epi8(srcReg23, srcReg34); + + for (i = output_height; i > 1; i -= 2) { + // load the last 2 loads of 16 bytes and have every two + // consecutive loads in the same 256 bit register + srcReg5x = _mm256_castsi128_si256( + _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 5))); + srcReg45 = + _mm256_inserti128_si256(srcReg4x, _mm256_castsi256_si128(srcReg5x), 1); + + srcReg6x = _mm256_castsi128_si256( + _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 6))); + srcReg56 = + _mm256_inserti128_si256(srcReg5x, _mm256_castsi256_si128(srcReg6x), 1); + + // merge every two consecutive registers + srcReg45_56_lo = _mm256_unpacklo_epi8(srcReg45, srcReg56); + + // multiply 2 adjacent elements with the filter and add the result + resReg23_34_lo = _mm256_maddubs_epi16(srcReg23_34_lo, secondFilters); + resReg45_56_lo = _mm256_maddubs_epi16(srcReg45_56_lo, thirdFilters); + + // add and saturate the results together + resReglo = _mm256_adds_epi16(resReg23_34_lo, resReg45_56_lo); + + // shift by 6 bit each 16 bit + resReglo = _mm256_adds_epi16(resReglo, addFilterReg32); + resReglo = _mm256_srai_epi16(resReglo, 6); + + // shrink to 8 bit each 16 bits, the first lane contain the first + // convolve result and the second lane contain the second convolve + // result + resReg = _mm256_packus_epi16(resReglo, resReglo); + + src_ptr += src_stride; + + xx_storeu2_epi64(output_ptr, out_pitch, &resReg); + + output_ptr += dst_stride; + + // save part of the registers for next strides + srcReg23_34_lo = srcReg45_56_lo; + srcReg4x = srcReg6x; + } +} + +static void aom_filter_block1d8_v8_avx2( + const uint8_t *src_ptr, ptrdiff_t src_pitch, uint8_t *output_ptr, + ptrdiff_t out_pitch, uint32_t output_height, const int16_t *filter) { + __m128i filtersReg; + __m256i addFilterReg32; + __m256i srcReg32b1, srcReg32b2, srcReg32b3, srcReg32b4, srcReg32b5; + __m256i srcReg32b6, srcReg32b7, srcReg32b8, srcReg32b9, srcReg32b10; + __m256i srcReg32b11, srcReg32b12, filtersReg32; + __m256i firstFilters, secondFilters, thirdFilters, forthFilters; + unsigned int i; + ptrdiff_t src_stride, dst_stride; + + addFilterReg32 = _mm256_set1_epi16(32); + filtersReg = _mm_loadu_si128((const __m128i *)filter); + // converting the 16 bit (short) to 8 bit (byte) and have the + // same data in both lanes of 128 bit register. + filtersReg = _mm_srai_epi16(filtersReg, 1); + filtersReg = _mm_packs_epi16(filtersReg, filtersReg); + // have the same data in both lanes of a 256 bit register + filtersReg32 = MM256_BROADCASTSI128_SI256(filtersReg); + + // duplicate only the first 16 bits (first and second byte) + // across 256 bit register + firstFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x100u)); + // duplicate only the second 16 bits (third and forth byte) + // across 256 bit register + secondFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x302u)); + // duplicate only the third 16 bits (fifth and sixth byte) + // across 256 bit register + thirdFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x504u)); + // duplicate only the forth 16 bits (seventh and eighth byte) + // across 256 bit register + forthFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x706u)); + + // multiple the size of the source and destination stride by two + src_stride = src_pitch << 1; + dst_stride = out_pitch << 1; + + // load 16 bytes 7 times in stride of src_pitch + srcReg32b1 = xx_loadu2_epi64(src_ptr + src_pitch, src_ptr); + srcReg32b3 = + xx_loadu2_epi64(src_ptr + src_pitch * 3, src_ptr + src_pitch * 2); + srcReg32b5 = + xx_loadu2_epi64(src_ptr + src_pitch * 5, src_ptr + src_pitch * 4); + srcReg32b7 = _mm256_castsi128_si256( + _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 6))); + + // have each consecutive loads on the same 256 register + srcReg32b2 = _mm256_permute2x128_si256(srcReg32b1, srcReg32b3, 0x21); + srcReg32b4 = _mm256_permute2x128_si256(srcReg32b3, srcReg32b5, 0x21); + srcReg32b6 = _mm256_permute2x128_si256(srcReg32b5, srcReg32b7, 0x21); + // merge every two consecutive registers except the last one + srcReg32b10 = _mm256_unpacklo_epi8(srcReg32b1, srcReg32b2); + srcReg32b11 = _mm256_unpacklo_epi8(srcReg32b3, srcReg32b4); + srcReg32b2 = _mm256_unpacklo_epi8(srcReg32b5, srcReg32b6); + + for (i = output_height; i > 1; i -= 2) { + // load the last 2 loads of 16 bytes and have every two + // consecutive loads in the same 256 bit register + srcReg32b8 = _mm256_castsi128_si256( + _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 7))); + srcReg32b7 = _mm256_inserti128_si256(srcReg32b7, + _mm256_castsi256_si128(srcReg32b8), 1); + srcReg32b9 = _mm256_castsi128_si256( + _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 8))); + srcReg32b8 = _mm256_inserti128_si256(srcReg32b8, + _mm256_castsi256_si128(srcReg32b9), 1); + + // merge every two consecutive registers + // save + srcReg32b4 = _mm256_unpacklo_epi8(srcReg32b7, srcReg32b8); + + // multiply 2 adjacent elements with the filter and add the result + srcReg32b10 = _mm256_maddubs_epi16(srcReg32b10, firstFilters); + srcReg32b6 = _mm256_maddubs_epi16(srcReg32b4, forthFilters); + + // add and saturate the results together + srcReg32b10 = _mm256_adds_epi16(srcReg32b10, srcReg32b6); + + // multiply 2 adjacent elements with the filter and add the result + srcReg32b8 = _mm256_maddubs_epi16(srcReg32b11, secondFilters); + srcReg32b12 = _mm256_maddubs_epi16(srcReg32b2, thirdFilters); + + // add and saturate the results together + srcReg32b10 = _mm256_adds_epi16(srcReg32b10, + _mm256_adds_epi16(srcReg32b8, srcReg32b12)); + + // shift by 6 bit each 16 bit + srcReg32b10 = _mm256_adds_epi16(srcReg32b10, addFilterReg32); + srcReg32b10 = _mm256_srai_epi16(srcReg32b10, 6); + + // shrink to 8 bit each 16 bits, the first lane contain the first + // convolve result and the second lane contain the second convolve + // result + srcReg32b1 = _mm256_packus_epi16(srcReg32b10, _mm256_setzero_si256()); + + src_ptr += src_stride; + + xx_storeu2_epi64(output_ptr, out_pitch, &srcReg32b1); + + output_ptr += dst_stride; + + // save part of the registers for next strides + srcReg32b10 = srcReg32b11; + srcReg32b11 = srcReg32b2; + srcReg32b2 = srcReg32b4; + srcReg32b7 = srcReg32b9; + } + if (i > 0) { + __m128i srcRegFilt1, srcRegFilt4, srcRegFilt6, srcRegFilt8; + // load the last 16 bytes + srcRegFilt8 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 7)); + + // merge the last 2 results together + srcRegFilt4 = + _mm_unpacklo_epi8(_mm256_castsi256_si128(srcReg32b7), srcRegFilt8); + + // multiply 2 adjacent elements with the filter and add the result + srcRegFilt1 = _mm_maddubs_epi16(_mm256_castsi256_si128(srcReg32b10), + _mm256_castsi256_si128(firstFilters)); + srcRegFilt4 = + _mm_maddubs_epi16(srcRegFilt4, _mm256_castsi256_si128(forthFilters)); + + // add and saturate the results together + srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, srcRegFilt4); + + // multiply 2 adjacent elements with the filter and add the result + srcRegFilt4 = _mm_maddubs_epi16(_mm256_castsi256_si128(srcReg32b11), + _mm256_castsi256_si128(secondFilters)); + + // multiply 2 adjacent elements with the filter and add the result + srcRegFilt6 = _mm_maddubs_epi16(_mm256_castsi256_si128(srcReg32b2), + _mm256_castsi256_si128(thirdFilters)); + + // add and saturate the results together + srcRegFilt1 = + _mm_adds_epi16(srcRegFilt1, _mm_adds_epi16(srcRegFilt4, srcRegFilt6)); + + // shift by 6 bit each 16 bit + srcRegFilt1 = + _mm_adds_epi16(srcRegFilt1, _mm256_castsi256_si128(addFilterReg32)); + srcRegFilt1 = _mm_srai_epi16(srcRegFilt1, 6); + + // shrink to 8 bit each 16 bits, the first lane contain the first + // convolve result and the second lane contain the second convolve result + srcRegFilt1 = _mm_packus_epi16(srcRegFilt1, _mm_setzero_si128()); + + // save 8 bytes + _mm_storel_epi64((__m128i *)output_ptr, srcRegFilt1); + } +} + +static void aom_filter_block1d16_v4_avx2( + const uint8_t *src_ptr, ptrdiff_t src_pitch, uint8_t *output_ptr, + ptrdiff_t out_pitch, uint32_t output_height, const int16_t *filter) { + __m128i filtersReg; + __m256i filtersReg32, addFilterReg32; + __m256i srcReg23, srcReg4x, srcReg34, srcReg5x, srcReg45, srcReg6x, srcReg56; + __m256i srcReg23_34_lo, srcReg23_34_hi, srcReg45_56_lo, srcReg45_56_hi; + __m256i resReg23_34_lo, resReg23_34_hi, resReg45_56_lo, resReg45_56_hi; + __m256i resReglo, resReghi, resReg; + __m256i secondFilters, thirdFilters; + unsigned int i; + ptrdiff_t src_stride, dst_stride; + + addFilterReg32 = _mm256_set1_epi16(32); + filtersReg = _mm_loadu_si128((const __m128i *)filter); + // converting the 16 bit (short) to 8 bit (byte) and have the + // same data in both lanes of 128 bit register. + filtersReg = _mm_srai_epi16(filtersReg, 1); + filtersReg = _mm_packs_epi16(filtersReg, filtersReg); + // have the same data in both lanes of a 256 bit register + filtersReg32 = MM256_BROADCASTSI128_SI256(filtersReg); + + // duplicate only the second 16 bits (third and forth byte) + // across 256 bit register + secondFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x302u)); + // duplicate only the third 16 bits (fifth and sixth byte) + // across 256 bit register + thirdFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x504u)); + + // multiple the size of the source and destination stride by two + src_stride = src_pitch << 1; + dst_stride = out_pitch << 1; + + srcReg23 = xx_loadu2_mi128(src_ptr + src_pitch * 3, src_ptr + src_pitch * 2); + srcReg4x = _mm256_castsi128_si256( + _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 4))); + + // have consecutive loads on the same 256 register + srcReg34 = _mm256_permute2x128_si256(srcReg23, srcReg4x, 0x21); + + srcReg23_34_lo = _mm256_unpacklo_epi8(srcReg23, srcReg34); + srcReg23_34_hi = _mm256_unpackhi_epi8(srcReg23, srcReg34); + + for (i = output_height; i > 1; i -= 2) { + // load the last 2 loads of 16 bytes and have every two + // consecutive loads in the same 256 bit register + srcReg5x = _mm256_castsi128_si256( + _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 5))); + srcReg45 = + _mm256_inserti128_si256(srcReg4x, _mm256_castsi256_si128(srcReg5x), 1); + + srcReg6x = _mm256_castsi128_si256( + _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 6))); + srcReg56 = + _mm256_inserti128_si256(srcReg5x, _mm256_castsi256_si128(srcReg6x), 1); + + // merge every two consecutive registers + srcReg45_56_lo = _mm256_unpacklo_epi8(srcReg45, srcReg56); + srcReg45_56_hi = _mm256_unpackhi_epi8(srcReg45, srcReg56); + + // multiply 2 adjacent elements with the filter and add the result + resReg23_34_lo = _mm256_maddubs_epi16(srcReg23_34_lo, secondFilters); + resReg45_56_lo = _mm256_maddubs_epi16(srcReg45_56_lo, thirdFilters); + + // add and saturate the results together + resReglo = _mm256_adds_epi16(resReg23_34_lo, resReg45_56_lo); + + // multiply 2 adjacent elements with the filter and add the result + resReg23_34_hi = _mm256_maddubs_epi16(srcReg23_34_hi, secondFilters); + resReg45_56_hi = _mm256_maddubs_epi16(srcReg45_56_hi, thirdFilters); + + // add and saturate the results together + resReghi = _mm256_adds_epi16(resReg23_34_hi, resReg45_56_hi); + + // shift by 6 bit each 16 bit + resReglo = _mm256_adds_epi16(resReglo, addFilterReg32); + resReghi = _mm256_adds_epi16(resReghi, addFilterReg32); + resReglo = _mm256_srai_epi16(resReglo, 6); + resReghi = _mm256_srai_epi16(resReghi, 6); + + // shrink to 8 bit each 16 bits, the first lane contain the first + // convolve result and the second lane contain the second convolve + // result + resReg = _mm256_packus_epi16(resReglo, resReghi); + + src_ptr += src_stride; + + xx_store2_mi128(output_ptr, out_pitch, &resReg); + + output_ptr += dst_stride; + + // save part of the registers for next strides + srcReg23_34_lo = srcReg45_56_lo; + srcReg23_34_hi = srcReg45_56_hi; + srcReg4x = srcReg6x; + } +} + +static void aom_filter_block1d16_v8_avx2( + const uint8_t *src_ptr, ptrdiff_t src_pitch, uint8_t *output_ptr, + ptrdiff_t out_pitch, uint32_t output_height, const int16_t *filter) { + __m128i filtersReg; + __m256i addFilterReg32; + __m256i srcReg32b1, srcReg32b2, srcReg32b3, srcReg32b4, srcReg32b5; + __m256i srcReg32b6, srcReg32b7, srcReg32b8, srcReg32b9, srcReg32b10; + __m256i srcReg32b11, srcReg32b12, filtersReg32; + __m256i firstFilters, secondFilters, thirdFilters, forthFilters; + unsigned int i; + ptrdiff_t src_stride, dst_stride; + + addFilterReg32 = _mm256_set1_epi16(32); + filtersReg = _mm_loadu_si128((const __m128i *)filter); + // converting the 16 bit (short) to 8 bit (byte) and have the + // same data in both lanes of 128 bit register. + filtersReg = _mm_srai_epi16(filtersReg, 1); + filtersReg = _mm_packs_epi16(filtersReg, filtersReg); + // have the same data in both lanes of a 256 bit register + filtersReg32 = MM256_BROADCASTSI128_SI256(filtersReg); + + // duplicate only the first 16 bits (first and second byte) + // across 256 bit register + firstFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x100u)); + // duplicate only the second 16 bits (third and forth byte) + // across 256 bit register + secondFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x302u)); + // duplicate only the third 16 bits (fifth and sixth byte) + // across 256 bit register + thirdFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x504u)); + // duplicate only the forth 16 bits (seventh and eighth byte) + // across 256 bit register + forthFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x706u)); + + // multiple the size of the source and destination stride by two + src_stride = src_pitch << 1; + dst_stride = out_pitch << 1; + + // load 16 bytes 7 times in stride of src_pitch + srcReg32b1 = xx_loadu2_mi128(src_ptr + src_pitch, src_ptr); + srcReg32b3 = + xx_loadu2_mi128(src_ptr + src_pitch * 3, src_ptr + src_pitch * 2); + srcReg32b5 = + xx_loadu2_mi128(src_ptr + src_pitch * 5, src_ptr + src_pitch * 4); + srcReg32b7 = _mm256_castsi128_si256( + _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 6))); + + // have each consecutive loads on the same 256 register + srcReg32b2 = _mm256_permute2x128_si256(srcReg32b1, srcReg32b3, 0x21); + srcReg32b4 = _mm256_permute2x128_si256(srcReg32b3, srcReg32b5, 0x21); + srcReg32b6 = _mm256_permute2x128_si256(srcReg32b5, srcReg32b7, 0x21); + // merge every two consecutive registers except the last one + srcReg32b10 = _mm256_unpacklo_epi8(srcReg32b1, srcReg32b2); + srcReg32b1 = _mm256_unpackhi_epi8(srcReg32b1, srcReg32b2); + + // save + srcReg32b11 = _mm256_unpacklo_epi8(srcReg32b3, srcReg32b4); + srcReg32b3 = _mm256_unpackhi_epi8(srcReg32b3, srcReg32b4); + srcReg32b2 = _mm256_unpacklo_epi8(srcReg32b5, srcReg32b6); + srcReg32b5 = _mm256_unpackhi_epi8(srcReg32b5, srcReg32b6); + + for (i = output_height; i > 1; i -= 2) { + // load the last 2 loads of 16 bytes and have every two + // consecutive loads in the same 256 bit register + srcReg32b8 = _mm256_castsi128_si256( + _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 7))); + srcReg32b7 = _mm256_inserti128_si256(srcReg32b7, + _mm256_castsi256_si128(srcReg32b8), 1); + srcReg32b9 = _mm256_castsi128_si256( + _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 8))); + srcReg32b8 = _mm256_inserti128_si256(srcReg32b8, + _mm256_castsi256_si128(srcReg32b9), 1); + + // merge every two consecutive registers + // save + srcReg32b4 = _mm256_unpacklo_epi8(srcReg32b7, srcReg32b8); + srcReg32b7 = _mm256_unpackhi_epi8(srcReg32b7, srcReg32b8); + + // multiply 2 adjacent elements with the filter and add the result + srcReg32b10 = _mm256_maddubs_epi16(srcReg32b10, firstFilters); + srcReg32b6 = _mm256_maddubs_epi16(srcReg32b4, forthFilters); + + // add and saturate the results together + srcReg32b10 = _mm256_adds_epi16(srcReg32b10, srcReg32b6); + + // multiply 2 adjacent elements with the filter and add the result + srcReg32b8 = _mm256_maddubs_epi16(srcReg32b11, secondFilters); + srcReg32b12 = _mm256_maddubs_epi16(srcReg32b2, thirdFilters); + + // add and saturate the results together + srcReg32b10 = _mm256_adds_epi16(srcReg32b10, + _mm256_adds_epi16(srcReg32b8, srcReg32b12)); + + // multiply 2 adjacent elements with the filter and add the result + srcReg32b1 = _mm256_maddubs_epi16(srcReg32b1, firstFilters); + srcReg32b6 = _mm256_maddubs_epi16(srcReg32b7, forthFilters); + + srcReg32b1 = _mm256_adds_epi16(srcReg32b1, srcReg32b6); + + // multiply 2 adjacent elements with the filter and add the result + srcReg32b8 = _mm256_maddubs_epi16(srcReg32b3, secondFilters); + srcReg32b12 = _mm256_maddubs_epi16(srcReg32b5, thirdFilters); + + // add and saturate the results together + srcReg32b1 = _mm256_adds_epi16(srcReg32b1, + _mm256_adds_epi16(srcReg32b8, srcReg32b12)); + + // shift by 6 bit each 16 bit + srcReg32b10 = _mm256_adds_epi16(srcReg32b10, addFilterReg32); + srcReg32b1 = _mm256_adds_epi16(srcReg32b1, addFilterReg32); + srcReg32b10 = _mm256_srai_epi16(srcReg32b10, 6); + srcReg32b1 = _mm256_srai_epi16(srcReg32b1, 6); + + // shrink to 8 bit each 16 bits, the first lane contain the first + // convolve result and the second lane contain the second convolve + // result + srcReg32b1 = _mm256_packus_epi16(srcReg32b10, srcReg32b1); + + src_ptr += src_stride; + + xx_store2_mi128(output_ptr, out_pitch, &srcReg32b1); + + output_ptr += dst_stride; + + // save part of the registers for next strides + srcReg32b10 = srcReg32b11; + srcReg32b1 = srcReg32b3; + srcReg32b11 = srcReg32b2; + srcReg32b3 = srcReg32b5; + srcReg32b2 = srcReg32b4; + srcReg32b5 = srcReg32b7; + srcReg32b7 = srcReg32b9; + } + if (i > 0) { + __m128i srcRegFilt1, srcRegFilt3, srcRegFilt4, srcRegFilt5; + __m128i srcRegFilt6, srcRegFilt7, srcRegFilt8; + // load the last 16 bytes + srcRegFilt8 = _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 7)); + + // merge the last 2 results together + srcRegFilt4 = + _mm_unpacklo_epi8(_mm256_castsi256_si128(srcReg32b7), srcRegFilt8); + srcRegFilt7 = + _mm_unpackhi_epi8(_mm256_castsi256_si128(srcReg32b7), srcRegFilt8); + + // multiply 2 adjacent elements with the filter and add the result + srcRegFilt1 = _mm_maddubs_epi16(_mm256_castsi256_si128(srcReg32b10), + _mm256_castsi256_si128(firstFilters)); + srcRegFilt4 = + _mm_maddubs_epi16(srcRegFilt4, _mm256_castsi256_si128(forthFilters)); + srcRegFilt3 = _mm_maddubs_epi16(_mm256_castsi256_si128(srcReg32b1), + _mm256_castsi256_si128(firstFilters)); + srcRegFilt7 = + _mm_maddubs_epi16(srcRegFilt7, _mm256_castsi256_si128(forthFilters)); + + // add and saturate the results together + srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, srcRegFilt4); + srcRegFilt3 = _mm_adds_epi16(srcRegFilt3, srcRegFilt7); + + // multiply 2 adjacent elements with the filter and add the result + srcRegFilt4 = _mm_maddubs_epi16(_mm256_castsi256_si128(srcReg32b11), + _mm256_castsi256_si128(secondFilters)); + srcRegFilt5 = _mm_maddubs_epi16(_mm256_castsi256_si128(srcReg32b3), + _mm256_castsi256_si128(secondFilters)); + + // multiply 2 adjacent elements with the filter and add the result + srcRegFilt6 = _mm_maddubs_epi16(_mm256_castsi256_si128(srcReg32b2), + _mm256_castsi256_si128(thirdFilters)); + srcRegFilt7 = _mm_maddubs_epi16(_mm256_castsi256_si128(srcReg32b5), + _mm256_castsi256_si128(thirdFilters)); + + // add and saturate the results together + srcRegFilt1 = + _mm_adds_epi16(srcRegFilt1, _mm_adds_epi16(srcRegFilt4, srcRegFilt6)); + srcRegFilt3 = + _mm_adds_epi16(srcRegFilt3, _mm_adds_epi16(srcRegFilt5, srcRegFilt7)); + + // shift by 6 bit each 16 bit + srcRegFilt1 = + _mm_adds_epi16(srcRegFilt1, _mm256_castsi256_si128(addFilterReg32)); + srcRegFilt3 = + _mm_adds_epi16(srcRegFilt3, _mm256_castsi256_si128(addFilterReg32)); + srcRegFilt1 = _mm_srai_epi16(srcRegFilt1, 6); + srcRegFilt3 = _mm_srai_epi16(srcRegFilt3, 6); + + // shrink to 8 bit each 16 bits, the first lane contain the first + // convolve result and the second lane contain the second convolve + // result + srcRegFilt1 = _mm_packus_epi16(srcRegFilt1, srcRegFilt3); + + // save 16 bytes + _mm_store_si128((__m128i *)output_ptr, srcRegFilt1); + } +} + +static void aom_filter_block1d4_v4_avx2( + const uint8_t *src_ptr, ptrdiff_t src_pitch, uint8_t *output_ptr, + ptrdiff_t out_pitch, uint32_t output_height, const int16_t *filter) { + __m128i filtersReg; + __m256i filtersReg32, addFilterReg32; + __m256i srcReg23, srcReg4x, srcReg34, srcReg5x, srcReg45, srcReg6x, srcReg56; + __m256i srcReg23_34_lo, srcReg45_56_lo; + __m256i srcReg2345_3456_lo; + __m256i resReglo, resReg; + __m256i firstFilters; + unsigned int i; + ptrdiff_t src_stride, dst_stride; + + addFilterReg32 = _mm256_set1_epi16(32); + filtersReg = _mm_loadu_si128((const __m128i *)filter); + // converting the 16 bit (short) to 8 bit (byte) and have the + // same data in both lanes of 128 bit register. + filtersReg = _mm_srai_epi16(filtersReg, 1); + filtersReg = _mm_packs_epi16(filtersReg, filtersReg); + // have the same data in both lanes of a 256 bit register + filtersReg32 = MM256_BROADCASTSI128_SI256(filtersReg); + + firstFilters = + _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi32(0x5040302u)); + + // multiple the size of the source and destination stride by two + src_stride = src_pitch << 1; + dst_stride = out_pitch << 1; + + srcReg23 = xx_loadu2_epi64(src_ptr + src_pitch * 3, src_ptr + src_pitch * 2); + srcReg4x = _mm256_castsi128_si256( + _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 4))); + + // have consecutive loads on the same 256 register + srcReg34 = _mm256_permute2x128_si256(srcReg23, srcReg4x, 0x21); + + srcReg23_34_lo = _mm256_unpacklo_epi8(srcReg23, srcReg34); + + for (i = output_height; i > 1; i -= 2) { + // load the last 2 loads of 16 bytes and have every two + // consecutive loads in the same 256 bit register + srcReg5x = _mm256_castsi128_si256( + _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 5))); + srcReg45 = + _mm256_inserti128_si256(srcReg4x, _mm256_castsi256_si128(srcReg5x), 1); + + srcReg6x = _mm256_castsi128_si256( + _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 6))); + srcReg56 = + _mm256_inserti128_si256(srcReg5x, _mm256_castsi256_si128(srcReg6x), 1); + + // merge every two consecutive registers + srcReg45_56_lo = _mm256_unpacklo_epi8(srcReg45, srcReg56); + + srcReg2345_3456_lo = _mm256_unpacklo_epi16(srcReg23_34_lo, srcReg45_56_lo); + + // multiply 2 adjacent elements with the filter and add the result + resReglo = _mm256_maddubs_epi16(srcReg2345_3456_lo, firstFilters); + + resReglo = _mm256_hadds_epi16(resReglo, _mm256_setzero_si256()); + + // shift by 6 bit each 16 bit + resReglo = _mm256_adds_epi16(resReglo, addFilterReg32); + resReglo = _mm256_srai_epi16(resReglo, 6); + + // shrink to 8 bit each 16 bits, the first lane contain the first + // convolve result and the second lane contain the second convolve + // result + resReg = _mm256_packus_epi16(resReglo, resReglo); + + src_ptr += src_stride; + + xx_storeu2_epi32(output_ptr, out_pitch, &resReg); + + output_ptr += dst_stride; + + // save part of the registers for next strides + srcReg23_34_lo = srcReg45_56_lo; + srcReg4x = srcReg6x; + } +} + +#if HAVE_AVX2 && HAVE_SSSE3 +filter8_1dfunction aom_filter_block1d4_v8_ssse3; +filter8_1dfunction aom_filter_block1d16_v2_ssse3; +filter8_1dfunction aom_filter_block1d16_h2_ssse3; +filter8_1dfunction aom_filter_block1d8_v2_ssse3; +filter8_1dfunction aom_filter_block1d8_h2_ssse3; +filter8_1dfunction aom_filter_block1d4_v2_ssse3; +filter8_1dfunction aom_filter_block1d4_h2_ssse3; +#define aom_filter_block1d4_v8_avx2 aom_filter_block1d4_v8_ssse3 +#define aom_filter_block1d16_v2_avx2 aom_filter_block1d16_v2_ssse3 +#define aom_filter_block1d16_h2_avx2 aom_filter_block1d16_h2_ssse3 +#define aom_filter_block1d8_v2_avx2 aom_filter_block1d8_v2_ssse3 +#define aom_filter_block1d8_h2_avx2 aom_filter_block1d8_h2_ssse3 +#define aom_filter_block1d4_v2_avx2 aom_filter_block1d4_v2_ssse3 +#define aom_filter_block1d4_h2_avx2 aom_filter_block1d4_h2_ssse3 +// void aom_convolve8_horiz_avx2(const uint8_t *src, ptrdiff_t src_stride, +// uint8_t *dst, ptrdiff_t dst_stride, +// const int16_t *filter_x, int x_step_q4, +// const int16_t *filter_y, int y_step_q4, +// int w, int h); +// void aom_convolve8_vert_avx2(const uint8_t *src, ptrdiff_t src_stride, +// uint8_t *dst, ptrdiff_t dst_stride, +// const int16_t *filter_x, int x_step_q4, +// const int16_t *filter_y, int y_step_q4, +// int w, int h); +FUN_CONV_1D(horiz, x_step_q4, filter_x, h, src, , avx2); +FUN_CONV_1D(vert, y_step_q4, filter_y, v, src - src_stride * 3, , avx2); + +#endif // HAVE_AX2 && HAVE_SSSE3 diff --git a/third_party/aom/aom_dsp/x86/aom_subpixel_8t_intrin_ssse3.c b/third_party/aom/aom_dsp/x86/aom_subpixel_8t_intrin_ssse3.c new file mode 100644 index 0000000000..325a21b761 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/aom_subpixel_8t_intrin_ssse3.c @@ -0,0 +1,315 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include + +#include "config/aom_dsp_rtcd.h" + +#include "aom_dsp/aom_filter.h" +#include "aom_dsp/x86/convolve.h" +#include "aom_mem/aom_mem.h" +#include "aom_ports/mem.h" +#include "aom_ports/emmintrin_compat.h" + +// filters only for the 4_h8 convolution +DECLARE_ALIGNED(16, static const uint8_t, filt1_4_h8[16]) = { + 0, 1, 1, 2, 2, 3, 3, 4, 2, 3, 3, 4, 4, 5, 5, 6 +}; + +DECLARE_ALIGNED(16, static const uint8_t, filt2_4_h8[16]) = { + 4, 5, 5, 6, 6, 7, 7, 8, 6, 7, 7, 8, 8, 9, 9, 10 +}; + +// filters for 8_h8 and 16_h8 +DECLARE_ALIGNED(16, static const uint8_t, filt1_global[16]) = { + 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8 +}; + +DECLARE_ALIGNED(16, static const uint8_t, filt2_global[16]) = { + 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8, 8, 9, 9, 10 +}; + +DECLARE_ALIGNED(16, static const uint8_t, filt3_global[16]) = { + 4, 5, 5, 6, 6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12 +}; + +DECLARE_ALIGNED(16, static const uint8_t, filt4_global[16]) = { + 6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14 +}; + +// These are reused by the avx2 intrinsics. +filter8_1dfunction aom_filter_block1d8_v8_intrin_ssse3; +filter8_1dfunction aom_filter_block1d8_h8_intrin_ssse3; +filter8_1dfunction aom_filter_block1d4_h8_intrin_ssse3; + +void aom_filter_block1d4_h8_intrin_ssse3( + const uint8_t *src_ptr, ptrdiff_t src_pixels_per_line, uint8_t *output_ptr, + ptrdiff_t output_pitch, uint32_t output_height, const int16_t *filter) { + __m128i firstFilters, secondFilters, shuffle1, shuffle2; + __m128i srcRegFilt1, srcRegFilt2, srcRegFilt3, srcRegFilt4; + __m128i addFilterReg64, filtersReg, srcReg, minReg; + unsigned int i; + + // create a register with 0,64,0,64,0,64,0,64,0,64,0,64,0,64,0,64 + addFilterReg64 = _mm_set1_epi32((int)0x0400040u); + filtersReg = _mm_loadu_si128((const __m128i *)filter); + // converting the 16 bit (short) to 8 bit (byte) and have the same data + // in both lanes of 128 bit register. + filtersReg = _mm_packs_epi16(filtersReg, filtersReg); + + // duplicate only the first 16 bits in the filter into the first lane + firstFilters = _mm_shufflelo_epi16(filtersReg, 0); + // duplicate only the third 16 bit in the filter into the first lane + secondFilters = _mm_shufflelo_epi16(filtersReg, 0xAAu); + // duplicate only the seconds 16 bits in the filter into the second lane + // firstFilters: k0 k1 k0 k1 k0 k1 k0 k1 k2 k3 k2 k3 k2 k3 k2 k3 + firstFilters = _mm_shufflehi_epi16(firstFilters, 0x55u); + // duplicate only the forth 16 bits in the filter into the second lane + // secondFilters: k4 k5 k4 k5 k4 k5 k4 k5 k6 k7 k6 k7 k6 k7 k6 k7 + secondFilters = _mm_shufflehi_epi16(secondFilters, 0xFFu); + + // loading the local filters + shuffle1 = _mm_load_si128((__m128i const *)filt1_4_h8); + shuffle2 = _mm_load_si128((__m128i const *)filt2_4_h8); + + for (i = 0; i < output_height; i++) { + srcReg = _mm_loadu_si128((const __m128i *)(src_ptr - 3)); + + // filter the source buffer + srcRegFilt1 = _mm_shuffle_epi8(srcReg, shuffle1); + srcRegFilt2 = _mm_shuffle_epi8(srcReg, shuffle2); + + // multiply 2 adjacent elements with the filter and add the result + srcRegFilt1 = _mm_maddubs_epi16(srcRegFilt1, firstFilters); + srcRegFilt2 = _mm_maddubs_epi16(srcRegFilt2, secondFilters); + + // extract the higher half of the lane + srcRegFilt3 = _mm_srli_si128(srcRegFilt1, 8); + srcRegFilt4 = _mm_srli_si128(srcRegFilt2, 8); + + minReg = _mm_min_epi16(srcRegFilt3, srcRegFilt2); + + // add and saturate all the results together + srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, srcRegFilt4); + srcRegFilt3 = _mm_max_epi16(srcRegFilt3, srcRegFilt2); + srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, minReg); + srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, srcRegFilt3); + srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, addFilterReg64); + + // shift by 7 bit each 16 bits + srcRegFilt1 = _mm_srai_epi16(srcRegFilt1, 7); + + // shrink to 8 bit each 16 bits + srcRegFilt1 = _mm_packus_epi16(srcRegFilt1, srcRegFilt1); + src_ptr += src_pixels_per_line; + + // save only 4 bytes + *((int *)&output_ptr[0]) = _mm_cvtsi128_si32(srcRegFilt1); + + output_ptr += output_pitch; + } +} + +void aom_filter_block1d8_h8_intrin_ssse3( + const uint8_t *src_ptr, ptrdiff_t src_pixels_per_line, uint8_t *output_ptr, + ptrdiff_t output_pitch, uint32_t output_height, const int16_t *filter) { + __m128i firstFilters, secondFilters, thirdFilters, forthFilters, srcReg; + __m128i filt1Reg, filt2Reg, filt3Reg, filt4Reg; + __m128i srcRegFilt1, srcRegFilt2, srcRegFilt3, srcRegFilt4; + __m128i addFilterReg64, filtersReg, minReg; + unsigned int i; + + // create a register with 0,64,0,64,0,64,0,64,0,64,0,64,0,64,0,64 + addFilterReg64 = _mm_set1_epi32((int)0x0400040u); + filtersReg = _mm_loadu_si128((const __m128i *)filter); + // converting the 16 bit (short) to 8 bit (byte) and have the same data + // in both lanes of 128 bit register. + filtersReg = _mm_packs_epi16(filtersReg, filtersReg); + + // duplicate only the first 16 bits (first and second byte) + // across 128 bit register + firstFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi16(0x100u)); + // duplicate only the second 16 bits (third and forth byte) + // across 128 bit register + secondFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi16(0x302u)); + // duplicate only the third 16 bits (fifth and sixth byte) + // across 128 bit register + thirdFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi16(0x504u)); + // duplicate only the forth 16 bits (seventh and eighth byte) + // across 128 bit register + forthFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi16(0x706u)); + + filt1Reg = _mm_load_si128((__m128i const *)filt1_global); + filt2Reg = _mm_load_si128((__m128i const *)filt2_global); + filt3Reg = _mm_load_si128((__m128i const *)filt3_global); + filt4Reg = _mm_load_si128((__m128i const *)filt4_global); + + for (i = 0; i < output_height; i++) { + srcReg = _mm_loadu_si128((const __m128i *)(src_ptr - 3)); + + // filter the source buffer + srcRegFilt1 = _mm_shuffle_epi8(srcReg, filt1Reg); + srcRegFilt2 = _mm_shuffle_epi8(srcReg, filt2Reg); + + // multiply 2 adjacent elements with the filter and add the result + srcRegFilt1 = _mm_maddubs_epi16(srcRegFilt1, firstFilters); + srcRegFilt2 = _mm_maddubs_epi16(srcRegFilt2, secondFilters); + + // filter the source buffer + srcRegFilt3 = _mm_shuffle_epi8(srcReg, filt3Reg); + srcRegFilt4 = _mm_shuffle_epi8(srcReg, filt4Reg); + + // multiply 2 adjacent elements with the filter and add the result + srcRegFilt3 = _mm_maddubs_epi16(srcRegFilt3, thirdFilters); + srcRegFilt4 = _mm_maddubs_epi16(srcRegFilt4, forthFilters); + + // add and saturate all the results together + minReg = _mm_min_epi16(srcRegFilt2, srcRegFilt3); + srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, srcRegFilt4); + + srcRegFilt2 = _mm_max_epi16(srcRegFilt2, srcRegFilt3); + srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, minReg); + srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, srcRegFilt2); + srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, addFilterReg64); + + // shift by 7 bit each 16 bits + srcRegFilt1 = _mm_srai_epi16(srcRegFilt1, 7); + + // shrink to 8 bit each 16 bits + srcRegFilt1 = _mm_packus_epi16(srcRegFilt1, srcRegFilt1); + + src_ptr += src_pixels_per_line; + + // save only 8 bytes + _mm_storel_epi64((__m128i *)&output_ptr[0], srcRegFilt1); + + output_ptr += output_pitch; + } +} + +void aom_filter_block1d8_v8_intrin_ssse3( + const uint8_t *src_ptr, ptrdiff_t src_pitch, uint8_t *output_ptr, + ptrdiff_t out_pitch, uint32_t output_height, const int16_t *filter) { + __m128i addFilterReg64, filtersReg, minReg; + __m128i firstFilters, secondFilters, thirdFilters, forthFilters; + __m128i srcRegFilt1, srcRegFilt2, srcRegFilt3, srcRegFilt5; + __m128i srcReg1, srcReg2, srcReg3, srcReg4, srcReg5, srcReg6, srcReg7; + __m128i srcReg8; + unsigned int i; + + // create a register with 0,64,0,64,0,64,0,64,0,64,0,64,0,64,0,64 + addFilterReg64 = _mm_set1_epi32((int)0x0400040u); + filtersReg = _mm_loadu_si128((const __m128i *)filter); + // converting the 16 bit (short) to 8 bit (byte) and have the same data + // in both lanes of 128 bit register. + filtersReg = _mm_packs_epi16(filtersReg, filtersReg); + + // duplicate only the first 16 bits in the filter + firstFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi16(0x100u)); + // duplicate only the second 16 bits in the filter + secondFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi16(0x302u)); + // duplicate only the third 16 bits in the filter + thirdFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi16(0x504u)); + // duplicate only the forth 16 bits in the filter + forthFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi16(0x706u)); + + // load the first 7 rows of 8 bytes + srcReg1 = _mm_loadl_epi64((const __m128i *)src_ptr); + srcReg2 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch)); + srcReg3 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 2)); + srcReg4 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 3)); + srcReg5 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 4)); + srcReg6 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 5)); + srcReg7 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 6)); + + for (i = 0; i < output_height; i++) { + // load the last 8 bytes + srcReg8 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 7)); + + // merge the result together + srcRegFilt1 = _mm_unpacklo_epi8(srcReg1, srcReg2); + srcRegFilt3 = _mm_unpacklo_epi8(srcReg3, srcReg4); + + // merge the result together + srcRegFilt2 = _mm_unpacklo_epi8(srcReg5, srcReg6); + srcRegFilt5 = _mm_unpacklo_epi8(srcReg7, srcReg8); + + // multiply 2 adjacent elements with the filter and add the result + srcRegFilt1 = _mm_maddubs_epi16(srcRegFilt1, firstFilters); + srcRegFilt3 = _mm_maddubs_epi16(srcRegFilt3, secondFilters); + srcRegFilt2 = _mm_maddubs_epi16(srcRegFilt2, thirdFilters); + srcRegFilt5 = _mm_maddubs_epi16(srcRegFilt5, forthFilters); + + // add and saturate the results together + minReg = _mm_min_epi16(srcRegFilt2, srcRegFilt3); + srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, srcRegFilt5); + srcRegFilt2 = _mm_max_epi16(srcRegFilt2, srcRegFilt3); + srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, minReg); + srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, srcRegFilt2); + srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, addFilterReg64); + + // shift by 7 bit each 16 bit + srcRegFilt1 = _mm_srai_epi16(srcRegFilt1, 7); + + // shrink to 8 bit each 16 bits + srcRegFilt1 = _mm_packus_epi16(srcRegFilt1, srcRegFilt1); + + src_ptr += src_pitch; + + // shift down a row + srcReg1 = srcReg2; + srcReg2 = srcReg3; + srcReg3 = srcReg4; + srcReg4 = srcReg5; + srcReg5 = srcReg6; + srcReg6 = srcReg7; + srcReg7 = srcReg8; + + // save only 8 bytes convolve result + _mm_storel_epi64((__m128i *)&output_ptr[0], srcRegFilt1); + + output_ptr += out_pitch; + } +} + +filter8_1dfunction aom_filter_block1d16_v8_ssse3; +filter8_1dfunction aom_filter_block1d16_h8_ssse3; +filter8_1dfunction aom_filter_block1d8_v8_ssse3; +filter8_1dfunction aom_filter_block1d8_h8_ssse3; +filter8_1dfunction aom_filter_block1d4_v8_ssse3; +filter8_1dfunction aom_filter_block1d4_h8_ssse3; + +#define aom_filter_block1d16_h4_ssse3 aom_filter_block1d16_h8_ssse3 +#define aom_filter_block1d16_v4_ssse3 aom_filter_block1d16_v8_ssse3 +#define aom_filter_block1d8_h4_ssse3 aom_filter_block1d8_h8_ssse3 +#define aom_filter_block1d8_v4_ssse3 aom_filter_block1d8_v8_ssse3 +#define aom_filter_block1d4_h4_ssse3 aom_filter_block1d4_h8_ssse3 +#define aom_filter_block1d4_v4_ssse3 aom_filter_block1d4_v8_ssse3 + +filter8_1dfunction aom_filter_block1d16_v2_ssse3; +filter8_1dfunction aom_filter_block1d16_h2_ssse3; +filter8_1dfunction aom_filter_block1d8_v2_ssse3; +filter8_1dfunction aom_filter_block1d8_h2_ssse3; +filter8_1dfunction aom_filter_block1d4_v2_ssse3; +filter8_1dfunction aom_filter_block1d4_h2_ssse3; + +// void aom_convolve8_horiz_ssse3(const uint8_t *src, ptrdiff_t src_stride, +// uint8_t *dst, ptrdiff_t dst_stride, +// const int16_t *filter_x, int x_step_q4, +// const int16_t *filter_y, int y_step_q4, +// int w, int h); +// void aom_convolve8_vert_ssse3(const uint8_t *src, ptrdiff_t src_stride, +// uint8_t *dst, ptrdiff_t dst_stride, +// const int16_t *filter_x, int x_step_q4, +// const int16_t *filter_y, int y_step_q4, +// int w, int h); +FUN_CONV_1D(horiz, x_step_q4, filter_x, h, src, , ssse3); +FUN_CONV_1D(vert, y_step_q4, filter_y, v, src - src_stride * 3, , ssse3); diff --git a/third_party/aom/aom_dsp/x86/aom_subpixel_8t_sse2.asm b/third_party/aom/aom_dsp/x86/aom_subpixel_8t_sse2.asm new file mode 100644 index 0000000000..c88fc9ffb2 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/aom_subpixel_8t_sse2.asm @@ -0,0 +1,615 @@ +; +; Copyright (c) 2016, Alliance for Open Media. All rights reserved +; +; This source code is subject to the terms of the BSD 2 Clause License and +; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License +; was not distributed with this source code in the LICENSE file, you can +; obtain it at www.aomedia.org/license/software. If the Alliance for Open +; Media Patent License 1.0 was not distributed with this source code in the +; PATENTS file, you can obtain it at www.aomedia.org/license/patent. +; + +; + + +%include "aom_ports/x86_abi_support.asm" + +;Note: tap3 and tap4 have to be applied and added after other taps to avoid +;overflow. + +%macro GET_FILTERS_4 0 + mov rdx, arg(5) ;filter ptr + mov rcx, 0x0400040 + + movdqa xmm7, [rdx] ;load filters + pshuflw xmm0, xmm7, 0b ;k0 + pshuflw xmm1, xmm7, 01010101b ;k1 + pshuflw xmm2, xmm7, 10101010b ;k2 + pshuflw xmm3, xmm7, 11111111b ;k3 + psrldq xmm7, 8 + pshuflw xmm4, xmm7, 0b ;k4 + pshuflw xmm5, xmm7, 01010101b ;k5 + pshuflw xmm6, xmm7, 10101010b ;k6 + pshuflw xmm7, xmm7, 11111111b ;k7 + + punpcklqdq xmm0, xmm1 + punpcklqdq xmm2, xmm3 + punpcklqdq xmm5, xmm4 + punpcklqdq xmm6, xmm7 + + movdqa k0k1, xmm0 + movdqa k2k3, xmm2 + movdqa k5k4, xmm5 + movdqa k6k7, xmm6 + + movq xmm6, rcx + pshufd xmm6, xmm6, 0 + movdqa krd, xmm6 + + pxor xmm7, xmm7 + movdqa zero, xmm7 +%endm + +%macro APPLY_FILTER_4 1 + punpckldq xmm0, xmm1 ;two row in one register + punpckldq xmm6, xmm7 + punpckldq xmm2, xmm3 + punpckldq xmm5, xmm4 + + punpcklbw xmm0, zero ;unpack to word + punpcklbw xmm6, zero + punpcklbw xmm2, zero + punpcklbw xmm5, zero + + pmullw xmm0, k0k1 ;multiply the filter factors + pmullw xmm6, k6k7 + pmullw xmm2, k2k3 + pmullw xmm5, k5k4 + + paddsw xmm0, xmm6 ;sum + movdqa xmm1, xmm0 + psrldq xmm1, 8 + paddsw xmm0, xmm1 + paddsw xmm0, xmm2 + psrldq xmm2, 8 + paddsw xmm0, xmm5 + psrldq xmm5, 8 + paddsw xmm0, xmm2 + paddsw xmm0, xmm5 + + paddsw xmm0, krd ;rounding + psraw xmm0, 7 ;shift + packuswb xmm0, xmm0 ;pack to byte + +%if %1 + movd xmm1, [rdi] + pavgb xmm0, xmm1 +%endif + movd [rdi], xmm0 +%endm + +%macro GET_FILTERS 0 + mov rdx, arg(5) ;filter ptr + mov rsi, arg(0) ;src_ptr + mov rdi, arg(2) ;output_ptr + mov rcx, 0x0400040 + + movdqa xmm7, [rdx] ;load filters + pshuflw xmm0, xmm7, 0b ;k0 + pshuflw xmm1, xmm7, 01010101b ;k1 + pshuflw xmm2, xmm7, 10101010b ;k2 + pshuflw xmm3, xmm7, 11111111b ;k3 + pshufhw xmm4, xmm7, 0b ;k4 + pshufhw xmm5, xmm7, 01010101b ;k5 + pshufhw xmm6, xmm7, 10101010b ;k6 + pshufhw xmm7, xmm7, 11111111b ;k7 + + punpcklwd xmm0, xmm0 + punpcklwd xmm1, xmm1 + punpcklwd xmm2, xmm2 + punpcklwd xmm3, xmm3 + punpckhwd xmm4, xmm4 + punpckhwd xmm5, xmm5 + punpckhwd xmm6, xmm6 + punpckhwd xmm7, xmm7 + + movdqa k0, xmm0 ;store filter factors on stack + movdqa k1, xmm1 + movdqa k2, xmm2 + movdqa k3, xmm3 + movdqa k4, xmm4 + movdqa k5, xmm5 + movdqa k6, xmm6 + movdqa k7, xmm7 + + movq xmm6, rcx + pshufd xmm6, xmm6, 0 + movdqa krd, xmm6 ;rounding + + pxor xmm7, xmm7 + movdqa zero, xmm7 +%endm + +%macro LOAD_VERT_8 1 + movq xmm0, [rsi + %1] ;0 + movq xmm1, [rsi + rax + %1] ;1 + movq xmm6, [rsi + rdx * 2 + %1] ;6 + lea rsi, [rsi + rax] + movq xmm7, [rsi + rdx * 2 + %1] ;7 + movq xmm2, [rsi + rax + %1] ;2 + movq xmm3, [rsi + rax * 2 + %1] ;3 + movq xmm4, [rsi + rdx + %1] ;4 + movq xmm5, [rsi + rax * 4 + %1] ;5 +%endm + +%macro APPLY_FILTER_8 2 + punpcklbw xmm0, zero + punpcklbw xmm1, zero + punpcklbw xmm6, zero + punpcklbw xmm7, zero + punpcklbw xmm2, zero + punpcklbw xmm5, zero + punpcklbw xmm3, zero + punpcklbw xmm4, zero + + pmullw xmm0, k0 + pmullw xmm1, k1 + pmullw xmm6, k6 + pmullw xmm7, k7 + pmullw xmm2, k2 + pmullw xmm5, k5 + pmullw xmm3, k3 + pmullw xmm4, k4 + + paddsw xmm0, xmm1 + paddsw xmm0, xmm6 + paddsw xmm0, xmm7 + paddsw xmm0, xmm2 + paddsw xmm0, xmm5 + paddsw xmm0, xmm3 + paddsw xmm0, xmm4 + + paddsw xmm0, krd ;rounding + psraw xmm0, 7 ;shift + packuswb xmm0, xmm0 ;pack back to byte +%if %1 + movq xmm1, [rdi + %2] + pavgb xmm0, xmm1 +%endif + movq [rdi + %2], xmm0 +%endm + +SECTION .text + +;void aom_filter_block1d4_v8_sse2 +;( +; unsigned char *src_ptr, +; unsigned int src_pitch, +; unsigned char *output_ptr, +; unsigned int out_pitch, +; unsigned int output_height, +; short *filter +;) +global sym(aom_filter_block1d4_v8_sse2) PRIVATE +sym(aom_filter_block1d4_v8_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 6 + SAVE_XMM 7 + push rsi + push rdi + push rbx + ; end prolog + + ALIGN_STACK 16, rax + sub rsp, 16 * 6 + %define k0k1 [rsp + 16 * 0] + %define k2k3 [rsp + 16 * 1] + %define k5k4 [rsp + 16 * 2] + %define k6k7 [rsp + 16 * 3] + %define krd [rsp + 16 * 4] + %define zero [rsp + 16 * 5] + + GET_FILTERS_4 + + mov rsi, arg(0) ;src_ptr + mov rdi, arg(2) ;output_ptr + + movsxd rax, DWORD PTR arg(1) ;pixels_per_line + movsxd rbx, DWORD PTR arg(3) ;out_pitch + lea rdx, [rax + rax * 2] + movsxd rcx, DWORD PTR arg(4) ;output_height + +.loop: + movd xmm0, [rsi] ;load src: row 0 + movd xmm1, [rsi + rax] ;1 + movd xmm6, [rsi + rdx * 2] ;6 + lea rsi, [rsi + rax] + movd xmm7, [rsi + rdx * 2] ;7 + movd xmm2, [rsi + rax] ;2 + movd xmm3, [rsi + rax * 2] ;3 + movd xmm4, [rsi + rdx] ;4 + movd xmm5, [rsi + rax * 4] ;5 + + APPLY_FILTER_4 0 + + lea rdi, [rdi + rbx] + dec rcx + jnz .loop + + add rsp, 16 * 6 + pop rsp + pop rbx + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +;void aom_filter_block1d8_v8_sse2 +;( +; unsigned char *src_ptr, +; unsigned int src_pitch, +; unsigned char *output_ptr, +; unsigned int out_pitch, +; unsigned int output_height, +; short *filter +;) +global sym(aom_filter_block1d8_v8_sse2) PRIVATE +sym(aom_filter_block1d8_v8_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 6 + SAVE_XMM 7 + push rsi + push rdi + push rbx + ; end prolog + + ALIGN_STACK 16, rax + sub rsp, 16 * 10 + %define k0 [rsp + 16 * 0] + %define k1 [rsp + 16 * 1] + %define k2 [rsp + 16 * 2] + %define k3 [rsp + 16 * 3] + %define k4 [rsp + 16 * 4] + %define k5 [rsp + 16 * 5] + %define k6 [rsp + 16 * 6] + %define k7 [rsp + 16 * 7] + %define krd [rsp + 16 * 8] + %define zero [rsp + 16 * 9] + + GET_FILTERS + + movsxd rax, DWORD PTR arg(1) ;pixels_per_line + movsxd rbx, DWORD PTR arg(3) ;out_pitch + lea rdx, [rax + rax * 2] + movsxd rcx, DWORD PTR arg(4) ;output_height + +.loop: + LOAD_VERT_8 0 + APPLY_FILTER_8 0, 0 + + lea rdi, [rdi + rbx] + dec rcx + jnz .loop + + add rsp, 16 * 10 + pop rsp + pop rbx + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +;void aom_filter_block1d16_v8_sse2 +;( +; unsigned char *src_ptr, +; unsigned int src_pitch, +; unsigned char *output_ptr, +; unsigned int out_pitch, +; unsigned int output_height, +; short *filter +;) +global sym(aom_filter_block1d16_v8_sse2) PRIVATE +sym(aom_filter_block1d16_v8_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 6 + SAVE_XMM 7 + push rsi + push rdi + push rbx + ; end prolog + + ALIGN_STACK 16, rax + sub rsp, 16 * 10 + %define k0 [rsp + 16 * 0] + %define k1 [rsp + 16 * 1] + %define k2 [rsp + 16 * 2] + %define k3 [rsp + 16 * 3] + %define k4 [rsp + 16 * 4] + %define k5 [rsp + 16 * 5] + %define k6 [rsp + 16 * 6] + %define k7 [rsp + 16 * 7] + %define krd [rsp + 16 * 8] + %define zero [rsp + 16 * 9] + + GET_FILTERS + + movsxd rax, DWORD PTR arg(1) ;pixels_per_line + movsxd rbx, DWORD PTR arg(3) ;out_pitch + lea rdx, [rax + rax * 2] + movsxd rcx, DWORD PTR arg(4) ;output_height + +.loop: + LOAD_VERT_8 0 + APPLY_FILTER_8 0, 0 + sub rsi, rax + + LOAD_VERT_8 8 + APPLY_FILTER_8 0, 8 + add rdi, rbx + + dec rcx + jnz .loop + + add rsp, 16 * 10 + pop rsp + pop rbx + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +;void aom_filter_block1d4_h8_sse2 +;( +; unsigned char *src_ptr, +; unsigned int src_pixels_per_line, +; unsigned char *output_ptr, +; unsigned int output_pitch, +; unsigned int output_height, +; short *filter +;) +global sym(aom_filter_block1d4_h8_sse2) PRIVATE +sym(aom_filter_block1d4_h8_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 6 + SAVE_XMM 7 + push rsi + push rdi + ; end prolog + + ALIGN_STACK 16, rax + sub rsp, 16 * 6 + %define k0k1 [rsp + 16 * 0] + %define k2k3 [rsp + 16 * 1] + %define k5k4 [rsp + 16 * 2] + %define k6k7 [rsp + 16 * 3] + %define krd [rsp + 16 * 4] + %define zero [rsp + 16 * 5] + + GET_FILTERS_4 + + mov rsi, arg(0) ;src_ptr + mov rdi, arg(2) ;output_ptr + + movsxd rax, DWORD PTR arg(1) ;pixels_per_line + movsxd rdx, DWORD PTR arg(3) ;out_pitch + movsxd rcx, DWORD PTR arg(4) ;output_height + +.loop: + movdqu xmm0, [rsi - 3] ;load src + + movdqa xmm1, xmm0 + movdqa xmm6, xmm0 + movdqa xmm7, xmm0 + movdqa xmm2, xmm0 + movdqa xmm3, xmm0 + movdqa xmm5, xmm0 + movdqa xmm4, xmm0 + + psrldq xmm1, 1 + psrldq xmm6, 6 + psrldq xmm7, 7 + psrldq xmm2, 2 + psrldq xmm3, 3 + psrldq xmm5, 5 + psrldq xmm4, 4 + + APPLY_FILTER_4 0 + + lea rsi, [rsi + rax] + lea rdi, [rdi + rdx] + dec rcx + jnz .loop + + add rsp, 16 * 6 + pop rsp + + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +;void aom_filter_block1d8_h8_sse2 +;( +; unsigned char *src_ptr, +; unsigned int src_pixels_per_line, +; unsigned char *output_ptr, +; unsigned int output_pitch, +; unsigned int output_height, +; short *filter +;) +global sym(aom_filter_block1d8_h8_sse2) PRIVATE +sym(aom_filter_block1d8_h8_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 6 + SAVE_XMM 7 + push rsi + push rdi + ; end prolog + + ALIGN_STACK 16, rax + sub rsp, 16 * 10 + %define k0 [rsp + 16 * 0] + %define k1 [rsp + 16 * 1] + %define k2 [rsp + 16 * 2] + %define k3 [rsp + 16 * 3] + %define k4 [rsp + 16 * 4] + %define k5 [rsp + 16 * 5] + %define k6 [rsp + 16 * 6] + %define k7 [rsp + 16 * 7] + %define krd [rsp + 16 * 8] + %define zero [rsp + 16 * 9] + + GET_FILTERS + + movsxd rax, DWORD PTR arg(1) ;pixels_per_line + movsxd rdx, DWORD PTR arg(3) ;out_pitch + movsxd rcx, DWORD PTR arg(4) ;output_height + +.loop: + movdqu xmm0, [rsi - 3] ;load src + + movdqa xmm1, xmm0 + movdqa xmm6, xmm0 + movdqa xmm7, xmm0 + movdqa xmm2, xmm0 + movdqa xmm5, xmm0 + movdqa xmm3, xmm0 + movdqa xmm4, xmm0 + + psrldq xmm1, 1 + psrldq xmm6, 6 + psrldq xmm7, 7 + psrldq xmm2, 2 + psrldq xmm5, 5 + psrldq xmm3, 3 + psrldq xmm4, 4 + + APPLY_FILTER_8 0, 0 + + lea rsi, [rsi + rax] + lea rdi, [rdi + rdx] + dec rcx + jnz .loop + + add rsp, 16 * 10 + pop rsp + + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +;void aom_filter_block1d16_h8_sse2 +;( +; unsigned char *src_ptr, +; unsigned int src_pixels_per_line, +; unsigned char *output_ptr, +; unsigned int output_pitch, +; unsigned int output_height, +; short *filter +;) +global sym(aom_filter_block1d16_h8_sse2) PRIVATE +sym(aom_filter_block1d16_h8_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 6 + SAVE_XMM 7 + push rsi + push rdi + ; end prolog + + ALIGN_STACK 16, rax + sub rsp, 16 * 10 + %define k0 [rsp + 16 * 0] + %define k1 [rsp + 16 * 1] + %define k2 [rsp + 16 * 2] + %define k3 [rsp + 16 * 3] + %define k4 [rsp + 16 * 4] + %define k5 [rsp + 16 * 5] + %define k6 [rsp + 16 * 6] + %define k7 [rsp + 16 * 7] + %define krd [rsp + 16 * 8] + %define zero [rsp + 16 * 9] + + GET_FILTERS + + movsxd rax, DWORD PTR arg(1) ;pixels_per_line + movsxd rdx, DWORD PTR arg(3) ;out_pitch + movsxd rcx, DWORD PTR arg(4) ;output_height + +.loop: + movdqu xmm0, [rsi - 3] ;load src + + movdqa xmm1, xmm0 + movdqa xmm6, xmm0 + movdqa xmm7, xmm0 + movdqa xmm2, xmm0 + movdqa xmm5, xmm0 + movdqa xmm3, xmm0 + movdqa xmm4, xmm0 + + psrldq xmm1, 1 + psrldq xmm6, 6 + psrldq xmm7, 7 + psrldq xmm2, 2 + psrldq xmm5, 5 + psrldq xmm3, 3 + psrldq xmm4, 4 + + APPLY_FILTER_8 0, 0 + + movdqu xmm0, [rsi + 5] ;load src + + movdqa xmm1, xmm0 + movdqa xmm6, xmm0 + movdqa xmm7, xmm0 + movdqa xmm2, xmm0 + movdqa xmm5, xmm0 + movdqa xmm3, xmm0 + movdqa xmm4, xmm0 + + psrldq xmm1, 1 + psrldq xmm6, 6 + psrldq xmm7, 7 + psrldq xmm2, 2 + psrldq xmm5, 5 + psrldq xmm3, 3 + psrldq xmm4, 4 + + APPLY_FILTER_8 0, 8 + + lea rsi, [rsi + rax] + lea rdi, [rdi + rdx] + dec rcx + jnz .loop + + add rsp, 16 * 10 + pop rsp + + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret diff --git a/third_party/aom/aom_dsp/x86/aom_subpixel_8t_ssse3.asm b/third_party/aom/aom_dsp/x86/aom_subpixel_8t_ssse3.asm new file mode 100644 index 0000000000..3ca7921b66 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/aom_subpixel_8t_ssse3.asm @@ -0,0 +1,870 @@ +; +; Copyright (c) 2016, Alliance for Open Media. All rights reserved +; +; This source code is subject to the terms of the BSD 2 Clause License and +; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License +; was not distributed with this source code in the LICENSE file, you can +; obtain it at www.aomedia.org/license/software. If the Alliance for Open +; Media Patent License 1.0 was not distributed with this source code in the +; PATENTS file, you can obtain it at www.aomedia.org/license/patent. +; + +; + +%include "third_party/x86inc/x86inc.asm" + +SECTION_RODATA +pw_64: times 8 dw 64 +even_byte_mask: times 8 dw 0x00ff + +; %define USE_PMULHRSW +; NOTE: pmulhrsw has a latency of 5 cycles. Tests showed a performance loss +; when using this instruction. +; +; The add order below (based on ffav1) must be followed to prevent outranges. +; x = k0k1 + k4k5 +; y = k2k3 + k6k7 +; z = signed SAT(x + y) + +SECTION .text +%define LOCAL_VARS_SIZE 16*6 + +%macro SETUP_LOCAL_VARS 0 + ; TODO(slavarnway): using xmm registers for these on ARCH_X86_64 + + ; pmaddubsw has a higher latency on some platforms, this might be eased by + ; interleaving the instructions. + %define k0k1 [rsp + 16*0] + %define k2k3 [rsp + 16*1] + %define k4k5 [rsp + 16*2] + %define k6k7 [rsp + 16*3] + packsswb m4, m4 + ; TODO(slavarnway): multiple pshufb instructions had a higher latency on + ; some platforms. + pshuflw m0, m4, 0b ;k0_k1 + pshuflw m1, m4, 01010101b ;k2_k3 + pshuflw m2, m4, 10101010b ;k4_k5 + pshuflw m3, m4, 11111111b ;k6_k7 + punpcklqdq m0, m0 + punpcklqdq m1, m1 + punpcklqdq m2, m2 + punpcklqdq m3, m3 + mova k0k1, m0 + mova k2k3, m1 + mova k4k5, m2 + mova k6k7, m3 +%if ARCH_X86_64 + %define krd m12 + %define tmp0 [rsp + 16*4] + %define tmp1 [rsp + 16*5] + mova krd, [GLOBAL(pw_64)] +%else + %define krd [rsp + 16*4] +%if CONFIG_PIC=0 + mova m6, [GLOBAL(pw_64)] +%else + ; build constants without accessing global memory + pcmpeqb m6, m6 ;all ones + psrlw m6, 15 + psllw m6, 6 ;aka pw_64 +%endif + mova krd, m6 +%endif +%endm + +;------------------------------------------------------------------------------- +%if ARCH_X86_64 + %define LOCAL_VARS_SIZE_H4 0 +%else + %define LOCAL_VARS_SIZE_H4 16*4 +%endif + +%macro SUBPIX_HFILTER4 1 +cglobal filter_block1d4_%1, 6, 6, 11, LOCAL_VARS_SIZE_H4, \ + src, sstride, dst, dstride, height, filter + mova m4, [filterq] + packsswb m4, m4 +%if ARCH_X86_64 + %define k0k1k4k5 m8 + %define k2k3k6k7 m9 + %define krd m10 + mova krd, [GLOBAL(pw_64)] + pshuflw k0k1k4k5, m4, 0b ;k0_k1 + pshufhw k0k1k4k5, k0k1k4k5, 10101010b ;k0_k1_k4_k5 + pshuflw k2k3k6k7, m4, 01010101b ;k2_k3 + pshufhw k2k3k6k7, k2k3k6k7, 11111111b ;k2_k3_k6_k7 +%else + %define k0k1k4k5 [rsp + 16*0] + %define k2k3k6k7 [rsp + 16*1] + %define krd [rsp + 16*2] + pshuflw m6, m4, 0b ;k0_k1 + pshufhw m6, m6, 10101010b ;k0_k1_k4_k5 + pshuflw m7, m4, 01010101b ;k2_k3 + pshufhw m7, m7, 11111111b ;k2_k3_k6_k7 +%if CONFIG_PIC=0 + mova m1, [GLOBAL(pw_64)] +%else + ; build constants without accessing global memory + pcmpeqb m1, m1 ;all ones + psrlw m1, 15 + psllw m1, 6 ;aka pw_64 +%endif + mova k0k1k4k5, m6 + mova k2k3k6k7, m7 + mova krd, m1 +%endif + dec heightd + +.loop: + ;Do two rows at once + movu m4, [srcq - 3] + movu m5, [srcq + sstrideq - 3] + punpckhbw m1, m4, m4 + punpcklbw m4, m4 + punpckhbw m3, m5, m5 + punpcklbw m5, m5 + palignr m0, m1, m4, 1 + pmaddubsw m0, k0k1k4k5 + palignr m1, m4, 5 + pmaddubsw m1, k2k3k6k7 + palignr m2, m3, m5, 1 + pmaddubsw m2, k0k1k4k5 + palignr m3, m5, 5 + pmaddubsw m3, k2k3k6k7 + punpckhqdq m4, m0, m2 + punpcklqdq m0, m2 + punpckhqdq m5, m1, m3 + punpcklqdq m1, m3 + paddsw m0, m4 + paddsw m1, m5 +%ifidn %1, h8_avg + movd m4, [dstq] + movd m5, [dstq + dstrideq] +%endif + paddsw m0, m1 + paddsw m0, krd + psraw m0, 7 +%ifidn %1, h8_add_src + pxor m3, m3 + movu m4, [srcq] + movu m5, [srcq + sstrideq] + punpckldq m4, m5 ; Bytes 0,1,2,3 from row 0, then 0,1,2,3 from row 2 + punpcklbw m4, m3 + paddsw m0, m4 +%endif + packuswb m0, m0 + psrldq m1, m0, 4 + +%ifidn %1, h8_avg + pavgb m0, m4 + pavgb m1, m5 +%endif + movd [dstq], m0 + movd [dstq + dstrideq], m1 + + lea srcq, [srcq + sstrideq ] + prefetcht0 [srcq + 4 * sstrideq - 3] + lea srcq, [srcq + sstrideq ] + lea dstq, [dstq + 2 * dstrideq ] + prefetcht0 [srcq + 2 * sstrideq - 3] + + sub heightd, 2 + jg .loop + + ; Do last row if output_height is odd + jne .done + + movu m4, [srcq - 3] + punpckhbw m1, m4, m4 + punpcklbw m4, m4 + palignr m0, m1, m4, 1 + palignr m1, m4, 5 + pmaddubsw m0, k0k1k4k5 + pmaddubsw m1, k2k3k6k7 + psrldq m2, m0, 8 + psrldq m3, m1, 8 + paddsw m0, m2 + paddsw m1, m3 + paddsw m0, m1 + paddsw m0, krd + psraw m0, 7 +%ifidn %1, h8_add_src + pxor m3, m3 + movu m4, [srcq] + punpcklbw m4, m3 + paddsw m0, m4 +%endif + packuswb m0, m0 +%ifidn %1, h8_avg + movd m4, [dstq] + pavgb m0, m4 +%endif + movd [dstq], m0 +.done: + REP_RET +%endm + +;------------------------------------------------------------------------------- +%macro SUBPIX_HFILTER8 1 +cglobal filter_block1d8_%1, 6, 6, 14, LOCAL_VARS_SIZE, \ + src, sstride, dst, dstride, height, filter + mova m4, [filterq] + SETUP_LOCAL_VARS + dec heightd + +.loop: + ;Do two rows at once + movu m0, [srcq - 3] + movu m4, [srcq + sstrideq - 3] + punpckhbw m1, m0, m0 + punpcklbw m0, m0 + palignr m5, m1, m0, 13 + pmaddubsw m5, k6k7 + palignr m2, m1, m0, 5 + palignr m3, m1, m0, 9 + palignr m1, m0, 1 + pmaddubsw m1, k0k1 + punpckhbw m6, m4, m4 + punpcklbw m4, m4 + pmaddubsw m2, k2k3 + pmaddubsw m3, k4k5 + + palignr m7, m6, m4, 13 + palignr m0, m6, m4, 5 + pmaddubsw m7, k6k7 + paddsw m1, m3 + paddsw m2, m5 + paddsw m1, m2 +%ifidn %1, h8_avg + movh m2, [dstq] + movhps m2, [dstq + dstrideq] +%endif + palignr m5, m6, m4, 9 + palignr m6, m4, 1 + pmaddubsw m0, k2k3 + pmaddubsw m6, k0k1 + paddsw m1, krd + pmaddubsw m5, k4k5 + psraw m1, 7 + paddsw m0, m7 + paddsw m6, m5 + paddsw m6, m0 + paddsw m6, krd + psraw m6, 7 +%ifidn %1, h8_add_src + pxor m3, m3 + movu m4, [srcq] + movu m5, [srcq + sstrideq] + punpcklbw m4, m3 + punpcklbw m5, m3 + paddsw m1, m4 + paddsw m6, m5 +%endif + packuswb m1, m6 +%ifidn %1, h8_avg + pavgb m1, m2 +%endif + movh [dstq], m1 + movhps [dstq + dstrideq], m1 + + lea srcq, [srcq + sstrideq ] + prefetcht0 [srcq + 4 * sstrideq - 3] + lea srcq, [srcq + sstrideq ] + lea dstq, [dstq + 2 * dstrideq ] + prefetcht0 [srcq + 2 * sstrideq - 3] + sub heightd, 2 + jg .loop + + ; Do last row if output_height is odd + jne .done + + movu m0, [srcq - 3] + punpckhbw m3, m0, m0 + punpcklbw m0, m0 + palignr m1, m3, m0, 1 + palignr m2, m3, m0, 5 + palignr m4, m3, m0, 13 + palignr m3, m0, 9 + pmaddubsw m1, k0k1 + pmaddubsw m2, k2k3 + pmaddubsw m3, k4k5 + pmaddubsw m4, k6k7 + paddsw m1, m3 + paddsw m4, m2 + paddsw m1, m4 + paddsw m1, krd + psraw m1, 7 +%ifidn %1, h8_add_src + pxor m6, m6 + movu m5, [srcq] + punpcklbw m5, m6 + paddsw m1, m5 +%endif + packuswb m1, m1 +%ifidn %1, h8_avg + movh m0, [dstq] + pavgb m1, m0 +%endif + movh [dstq], m1 +.done: + REP_RET +%endm + +;------------------------------------------------------------------------------- +%macro SUBPIX_HFILTER16 1 +cglobal filter_block1d16_%1, 6, 6, 14, LOCAL_VARS_SIZE, \ + src, sstride, dst, dstride, height, filter + mova m4, [filterq] + SETUP_LOCAL_VARS + +.loop: + prefetcht0 [srcq + 2 * sstrideq -3] + + movu m0, [srcq - 3] + movu m4, [srcq - 2] + pmaddubsw m0, k0k1 + pmaddubsw m4, k0k1 + movu m1, [srcq - 1] + movu m5, [srcq + 0] + pmaddubsw m1, k2k3 + pmaddubsw m5, k2k3 + movu m2, [srcq + 1] + movu m6, [srcq + 2] + pmaddubsw m2, k4k5 + pmaddubsw m6, k4k5 + movu m3, [srcq + 3] + movu m7, [srcq + 4] + pmaddubsw m3, k6k7 + pmaddubsw m7, k6k7 + paddsw m0, m2 + paddsw m1, m3 + paddsw m0, m1 + paddsw m4, m6 + paddsw m5, m7 + paddsw m4, m5 + paddsw m0, krd + paddsw m4, krd + psraw m0, 7 + psraw m4, 7 +%ifidn %1, h8_add_src +%if ARCH_X86=1 && CONFIG_PIC=1 + pcmpeqb m2, m2 ;all ones + psrlw m2, 8 ;even_byte_mask +%else + mova m2, [GLOBAL(even_byte_mask)] +%endif + movu m5, [srcq] + mova m7, m5 + pand m5, m2 + psrlw m7, 8 + paddsw m0, m5 + paddsw m4, m7 +%endif + packuswb m0, m0 + packuswb m4, m4 + punpcklbw m0, m4 +%ifidn %1, h8_avg + pavgb m0, [dstq] +%endif + lea srcq, [srcq + sstrideq] + mova [dstq], m0 + lea dstq, [dstq + dstrideq] + dec heightd + jnz .loop + REP_RET +%endm + +INIT_XMM ssse3 +SUBPIX_HFILTER16 h8 +SUBPIX_HFILTER8 h8 +SUBPIX_HFILTER4 h8 + +;------------------------------------------------------------------------------- + +; TODO(Linfeng): Detect cpu type and choose the code with better performance. +%define X86_SUBPIX_VFILTER_PREFER_SLOW_CELERON 1 + +%if ARCH_X86_64 && X86_SUBPIX_VFILTER_PREFER_SLOW_CELERON + %define NUM_GENERAL_REG_USED 9 +%else + %define NUM_GENERAL_REG_USED 6 +%endif + +%macro SUBPIX_VFILTER 2 +cglobal filter_block1d%2_%1, 6, NUM_GENERAL_REG_USED, 15, LOCAL_VARS_SIZE, \ + src, sstride, dst, dstride, height, filter + mova m4, [filterq] + SETUP_LOCAL_VARS + +%ifidn %2, 8 + %define movx movh +%else + %define movx movd +%endif + + dec heightd + +%if ARCH_X86 || X86_SUBPIX_VFILTER_PREFER_SLOW_CELERON + +%if ARCH_X86_64 + %define src1q r7 + %define sstride6q r8 + %define dst_stride dstrideq +%else + %define src1q filterq + %define sstride6q dstrideq + %define dst_stride dstridemp +%endif + mov src1q, srcq + add src1q, sstrideq + lea sstride6q, [sstrideq + sstrideq * 4] + add sstride6q, sstrideq ;pitch * 6 + +.loop: + ;Do two rows at once + movx m0, [srcq ] ;A + movx m1, [src1q ] ;B + punpcklbw m0, m1 ;A B + movx m2, [srcq + sstrideq * 2 ] ;C + pmaddubsw m0, k0k1 + mova m6, m2 + movx m3, [src1q + sstrideq * 2] ;D + punpcklbw m2, m3 ;C D + pmaddubsw m2, k2k3 + movx m4, [srcq + sstrideq * 4 ] ;E + mova m7, m4 + movx m5, [src1q + sstrideq * 4] ;F + punpcklbw m4, m5 ;E F + pmaddubsw m4, k4k5 + punpcklbw m1, m6 ;A B next iter + movx m6, [srcq + sstride6q ] ;G + punpcklbw m5, m6 ;E F next iter + punpcklbw m3, m7 ;C D next iter + pmaddubsw m5, k4k5 + movx m7, [src1q + sstride6q ] ;H + punpcklbw m6, m7 ;G H + pmaddubsw m6, k6k7 + pmaddubsw m3, k2k3 + pmaddubsw m1, k0k1 + paddsw m0, m4 + paddsw m2, m6 + movx m6, [srcq + sstrideq * 8 ] ;H next iter + punpcklbw m7, m6 + pmaddubsw m7, k6k7 + paddsw m0, m2 + paddsw m0, krd + psraw m0, 7 + paddsw m1, m5 +%ifidn %1, v8_add_src + pxor m6, m6 + movu m4, [srcq] + punpcklbw m4, m6 + paddsw m0, m4 +%endif + packuswb m0, m0 + + paddsw m3, m7 + paddsw m1, m3 + paddsw m1, krd + psraw m1, 7 +%ifidn %1, v8_add_src + movu m4, [src1q] + punpcklbw m4, m6 + paddsw m1, m4 +%endif + lea srcq, [srcq + sstrideq * 2 ] + lea src1q, [src1q + sstrideq * 2] + packuswb m1, m1 + +%ifidn %1, v8_avg + movx m2, [dstq] + pavgb m0, m2 +%endif + movx [dstq], m0 + add dstq, dst_stride +%ifidn %1, v8_avg + movx m3, [dstq] + pavgb m1, m3 +%endif + movx [dstq], m1 + add dstq, dst_stride + sub heightd, 2 + jg .loop + + ; Do last row if output_height is odd + jne .done + + movx m0, [srcq ] ;A + movx m1, [srcq + sstrideq ] ;B + movx m6, [srcq + sstride6q ] ;G + punpcklbw m0, m1 ;A B + movx m7, [src1q + sstride6q ] ;H + pmaddubsw m0, k0k1 + movx m2, [srcq + sstrideq * 2 ] ;C + punpcklbw m6, m7 ;G H + movx m3, [src1q + sstrideq * 2] ;D + pmaddubsw m6, k6k7 + movx m4, [srcq + sstrideq * 4 ] ;E + punpcklbw m2, m3 ;C D + movx m5, [src1q + sstrideq * 4] ;F + punpcklbw m4, m5 ;E F + pmaddubsw m2, k2k3 + pmaddubsw m4, k4k5 + paddsw m2, m6 + paddsw m0, m4 + paddsw m0, m2 + paddsw m0, krd + psraw m0, 7 +%ifidn %1, v8_add_src + pxor m6, m6 + movu m4, [srcq] + punpcklbw m4, m6 + paddsw m0, m4 +%endif + packuswb m0, m0 +%ifidn %1, v8_avg + movx m1, [dstq] + pavgb m0, m1 +%endif + movx [dstq], m0 + +%else + ; ARCH_X86_64 + + movx m0, [srcq ] ;A + movx m1, [srcq + sstrideq ] ;B + lea srcq, [srcq + sstrideq * 2 ] + movx m2, [srcq] ;C + movx m3, [srcq + sstrideq] ;D + lea srcq, [srcq + sstrideq * 2 ] + movx m4, [srcq] ;E + movx m5, [srcq + sstrideq] ;F + lea srcq, [srcq + sstrideq * 2 ] + movx m6, [srcq] ;G + punpcklbw m0, m1 ;A B + punpcklbw m1, m2 ;A B next iter + punpcklbw m2, m3 ;C D + punpcklbw m3, m4 ;C D next iter + punpcklbw m4, m5 ;E F + punpcklbw m5, m6 ;E F next iter + +.loop: + ;Do two rows at once + movx m7, [srcq + sstrideq] ;H + lea srcq, [srcq + sstrideq * 2 ] + movx m14, [srcq] ;H next iter + punpcklbw m6, m7 ;G H + punpcklbw m7, m14 ;G H next iter + pmaddubsw m8, m0, k0k1 + pmaddubsw m9, m1, k0k1 + mova m0, m2 + mova m1, m3 + pmaddubsw m10, m2, k2k3 + pmaddubsw m11, m3, k2k3 + mova m2, m4 + mova m3, m5 + pmaddubsw m4, k4k5 + pmaddubsw m5, k4k5 + paddsw m8, m4 + paddsw m9, m5 + mova m4, m6 + mova m5, m7 + pmaddubsw m6, k6k7 + pmaddubsw m7, k6k7 + paddsw m10, m6 + paddsw m11, m7 + paddsw m8, m10 + paddsw m9, m11 + mova m6, m14 + paddsw m8, krd + paddsw m9, krd + psraw m8, 7 + psraw m9, 7 +%ifidn %2, 4 + packuswb m8, m8 + packuswb m9, m9 +%else + packuswb m8, m9 +%endif + +%ifidn %1, v8_avg + movx m7, [dstq] +%ifidn %2, 4 + movx m10, [dstq + dstrideq] + pavgb m9, m10 +%else + movhpd m7, [dstq + dstrideq] +%endif + pavgb m8, m7 +%endif + movx [dstq], m8 +%ifidn %2, 4 + movx [dstq + dstrideq], m9 +%else + movhpd [dstq + dstrideq], m8 +%endif + + lea dstq, [dstq + dstrideq * 2 ] + sub heightd, 2 + jg .loop + + ; Do last row if output_height is odd + jne .done + + movx m7, [srcq + sstrideq] ;H + punpcklbw m6, m7 ;G H + pmaddubsw m0, k0k1 + pmaddubsw m2, k2k3 + pmaddubsw m4, k4k5 + pmaddubsw m6, k6k7 + paddsw m0, m4 + paddsw m2, m6 + paddsw m0, m2 + paddsw m0, krd + psraw m0, 7 + packuswb m0, m0 +%ifidn %1, v8_avg + movx m1, [dstq] + pavgb m0, m1 +%endif + movx [dstq], m0 + +%endif ; ARCH_X86_64 + +.done: + REP_RET + +%endm + +;------------------------------------------------------------------------------- +%macro SUBPIX_VFILTER16 1 +cglobal filter_block1d16_%1, 6, NUM_GENERAL_REG_USED, 16, LOCAL_VARS_SIZE, \ + src, sstride, dst, dstride, height, filter + mova m4, [filterq] + SETUP_LOCAL_VARS + +%if ARCH_X86 || X86_SUBPIX_VFILTER_PREFER_SLOW_CELERON + +%if ARCH_X86_64 + %define src1q r7 + %define sstride6q r8 + %define dst_stride dstrideq +%else + %define src1q filterq + %define sstride6q dstrideq + %define dst_stride dstridemp +%endif + lea src1q, [srcq + sstrideq] + lea sstride6q, [sstrideq + sstrideq * 4] + add sstride6q, sstrideq ;pitch * 6 + +.loop: + movh m0, [srcq ] ;A + movh m1, [src1q ] ;B + movh m2, [srcq + sstrideq * 2 ] ;C + movh m3, [src1q + sstrideq * 2] ;D + movh m4, [srcq + sstrideq * 4 ] ;E + movh m5, [src1q + sstrideq * 4] ;F + + punpcklbw m0, m1 ;A B + movh m6, [srcq + sstride6q] ;G + punpcklbw m2, m3 ;C D + movh m7, [src1q + sstride6q] ;H + punpcklbw m4, m5 ;E F + pmaddubsw m0, k0k1 + movh m3, [srcq + 8] ;A + pmaddubsw m2, k2k3 + punpcklbw m6, m7 ;G H + movh m5, [srcq + sstrideq + 8] ;B + pmaddubsw m4, k4k5 + punpcklbw m3, m5 ;A B + movh m7, [srcq + sstrideq * 2 + 8] ;C + pmaddubsw m6, k6k7 + movh m5, [src1q + sstrideq * 2 + 8] ;D + punpcklbw m7, m5 ;C D + paddsw m2, m6 + pmaddubsw m3, k0k1 + movh m1, [srcq + sstrideq * 4 + 8] ;E + paddsw m0, m4 + pmaddubsw m7, k2k3 + movh m6, [src1q + sstrideq * 4 + 8] ;F + punpcklbw m1, m6 ;E F + paddsw m0, m2 + paddsw m0, krd + movh m2, [srcq + sstride6q + 8] ;G + pmaddubsw m1, k4k5 + movh m5, [src1q + sstride6q + 8] ;H + psraw m0, 7 + punpcklbw m2, m5 ;G H + pmaddubsw m2, k6k7 + paddsw m7, m2 + paddsw m3, m1 + paddsw m3, m7 + paddsw m3, krd + psraw m3, 7 +%ifidn %1, v8_add_src + pxor m6, m6 + movu m4, [src1q + 2 * sstrideq] ; Fetch from 3 rows down + mova m5, m4 + punpcklbw m4, m6 + punpckhbw m5, m6 + paddsw m0, m4 + paddsw m3, m5 +%endif + packuswb m0, m3 + + add srcq, sstrideq + add src1q, sstrideq +%ifidn %1, v8_avg + pavgb m0, [dstq] +%endif + mova [dstq], m0 + add dstq, dst_stride + dec heightd + jnz .loop + REP_RET + +%else + ; ARCH_X86_64 + dec heightd + + movu m1, [srcq ] ;A + movu m3, [srcq + sstrideq ] ;B + lea srcq, [srcq + sstrideq * 2] + punpcklbw m0, m1, m3 ;A B + punpckhbw m1, m3 ;A B + movu m5, [srcq] ;C + punpcklbw m2, m3, m5 ;A B next iter + punpckhbw m3, m5 ;A B next iter + mova tmp0, m2 ;store to stack + mova tmp1, m3 ;store to stack + movu m7, [srcq + sstrideq] ;D + lea srcq, [srcq + sstrideq * 2] + punpcklbw m4, m5, m7 ;C D + punpckhbw m5, m7 ;C D + movu m9, [srcq] ;E + punpcklbw m6, m7, m9 ;C D next iter + punpckhbw m7, m9 ;C D next iter + movu m11, [srcq + sstrideq] ;F + lea srcq, [srcq + sstrideq * 2] + punpcklbw m8, m9, m11 ;E F + punpckhbw m9, m11 ;E F + movu m2, [srcq] ;G + punpcklbw m10, m11, m2 ;E F next iter + punpckhbw m11, m2 ;E F next iter + +.loop: + ;Do two rows at once + pmaddubsw m13, m0, k0k1 + mova m0, m4 + pmaddubsw m14, m8, k4k5 + pmaddubsw m15, m4, k2k3 + mova m4, m8 + paddsw m13, m14 + movu m3, [srcq + sstrideq] ;H + lea srcq, [srcq + sstrideq * 2] + punpcklbw m14, m2, m3 ;G H + mova m8, m14 + pmaddubsw m14, k6k7 + paddsw m15, m14 + paddsw m13, m15 + paddsw m13, krd + psraw m13, 7 + + pmaddubsw m14, m1, k0k1 + pmaddubsw m1, m9, k4k5 + pmaddubsw m15, m5, k2k3 + paddsw m14, m1 + mova m1, m5 + mova m5, m9 + punpckhbw m2, m3 ;G H + mova m9, m2 + pmaddubsw m2, k6k7 + paddsw m15, m2 + paddsw m14, m15 + paddsw m14, krd + psraw m14, 7 + packuswb m13, m14 +%ifidn %1, v8_avg + pavgb m13, [dstq] +%endif + mova [dstq], m13 + + ; next iter + pmaddubsw m15, tmp0, k0k1 + pmaddubsw m14, m10, k4k5 + pmaddubsw m13, m6, k2k3 + paddsw m15, m14 + mova tmp0, m6 + mova m6, m10 + movu m2, [srcq] ;G next iter + punpcklbw m14, m3, m2 ;G H next iter + mova m10, m14 + pmaddubsw m14, k6k7 + paddsw m13, m14 + paddsw m15, m13 + paddsw m15, krd + psraw m15, 7 + + pmaddubsw m14, tmp1, k0k1 + mova tmp1, m7 + pmaddubsw m13, m7, k2k3 + mova m7, m11 + pmaddubsw m11, k4k5 + paddsw m14, m11 + punpckhbw m3, m2 ;G H next iter + mova m11, m3 + pmaddubsw m3, k6k7 + paddsw m13, m3 + paddsw m14, m13 + paddsw m14, krd + psraw m14, 7 + packuswb m15, m14 +%ifidn %1, v8_avg + pavgb m15, [dstq + dstrideq] +%endif + mova [dstq + dstrideq], m15 + lea dstq, [dstq + dstrideq * 2] + sub heightd, 2 + jg .loop + + ; Do last row if output_height is odd + jne .done + + movu m3, [srcq + sstrideq] ;H + punpcklbw m6, m2, m3 ;G H + punpckhbw m2, m3 ;G H + pmaddubsw m0, k0k1 + pmaddubsw m1, k0k1 + pmaddubsw m4, k2k3 + pmaddubsw m5, k2k3 + pmaddubsw m8, k4k5 + pmaddubsw m9, k4k5 + pmaddubsw m6, k6k7 + pmaddubsw m2, k6k7 + paddsw m0, m8 + paddsw m1, m9 + paddsw m4, m6 + paddsw m5, m2 + paddsw m0, m4 + paddsw m1, m5 + paddsw m0, krd + paddsw m1, krd + psraw m0, 7 + psraw m1, 7 + packuswb m0, m1 +%ifidn %1, v8_avg + pavgb m0, [dstq] +%endif + mova [dstq], m0 + +.done: + REP_RET + +%endif ; ARCH_X86_64 + +%endm + +INIT_XMM ssse3 +SUBPIX_VFILTER16 v8 +SUBPIX_VFILTER v8, 8 +SUBPIX_VFILTER v8, 4 diff --git a/third_party/aom/aom_dsp/x86/aom_subpixel_bilinear_sse2.asm b/third_party/aom/aom_dsp/x86/aom_subpixel_bilinear_sse2.asm new file mode 100644 index 0000000000..d0b4b2839e --- /dev/null +++ b/third_party/aom/aom_dsp/x86/aom_subpixel_bilinear_sse2.asm @@ -0,0 +1,295 @@ +; +; Copyright (c) 2016, Alliance for Open Media. All rights reserved +; +; This source code is subject to the terms of the BSD 2 Clause License and +; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License +; was not distributed with this source code in the LICENSE file, you can +; obtain it at www.aomedia.org/license/software. If the Alliance for Open +; Media Patent License 1.0 was not distributed with this source code in the +; PATENTS file, you can obtain it at www.aomedia.org/license/patent. +; + +; + +%include "aom_ports/x86_abi_support.asm" + +%macro GET_PARAM_4 0 + mov rdx, arg(5) ;filter ptr + mov rsi, arg(0) ;src_ptr + mov rdi, arg(2) ;output_ptr + mov rcx, 0x0400040 + + movdqa xmm3, [rdx] ;load filters + pshuflw xmm4, xmm3, 11111111b ;k3 + psrldq xmm3, 8 + pshuflw xmm3, xmm3, 0b ;k4 + punpcklqdq xmm4, xmm3 ;k3k4 + + movq xmm3, rcx ;rounding + pshufd xmm3, xmm3, 0 + + pxor xmm2, xmm2 + + movsxd rax, DWORD PTR arg(1) ;pixels_per_line + movsxd rdx, DWORD PTR arg(3) ;out_pitch + movsxd rcx, DWORD PTR arg(4) ;output_height +%endm + +%macro APPLY_FILTER_4 1 + + punpckldq xmm0, xmm1 ;two row in one register + punpcklbw xmm0, xmm2 ;unpack to word + pmullw xmm0, xmm4 ;multiply the filter factors + + movdqa xmm1, xmm0 + psrldq xmm1, 8 + paddsw xmm0, xmm1 + + paddsw xmm0, xmm3 ;rounding + psraw xmm0, 7 ;shift + packuswb xmm0, xmm0 ;pack to byte + +%if %1 + movd xmm1, [rdi] + pavgb xmm0, xmm1 +%endif + + movd [rdi], xmm0 + lea rsi, [rsi + rax] + lea rdi, [rdi + rdx] + dec rcx +%endm + +%macro GET_PARAM 0 + mov rdx, arg(5) ;filter ptr + mov rsi, arg(0) ;src_ptr + mov rdi, arg(2) ;output_ptr + mov rcx, 0x0400040 + + movdqa xmm7, [rdx] ;load filters + + pshuflw xmm6, xmm7, 11111111b ;k3 + pshufhw xmm7, xmm7, 0b ;k4 + punpcklwd xmm6, xmm6 + punpckhwd xmm7, xmm7 + + movq xmm4, rcx ;rounding + pshufd xmm4, xmm4, 0 + + pxor xmm5, xmm5 + + movsxd rax, DWORD PTR arg(1) ;pixels_per_line + movsxd rdx, DWORD PTR arg(3) ;out_pitch + movsxd rcx, DWORD PTR arg(4) ;output_height +%endm + +%macro APPLY_FILTER_8 1 + punpcklbw xmm0, xmm5 + punpcklbw xmm1, xmm5 + + pmullw xmm0, xmm6 + pmullw xmm1, xmm7 + paddsw xmm0, xmm1 + paddsw xmm0, xmm4 ;rounding + psraw xmm0, 7 ;shift + packuswb xmm0, xmm0 ;pack back to byte +%if %1 + movq xmm1, [rdi] + pavgb xmm0, xmm1 +%endif + movq [rdi], xmm0 ;store the result + + lea rsi, [rsi + rax] + lea rdi, [rdi + rdx] + dec rcx +%endm + +%macro APPLY_FILTER_16 1 + punpcklbw xmm0, xmm5 + punpcklbw xmm1, xmm5 + punpckhbw xmm2, xmm5 + punpckhbw xmm3, xmm5 + + pmullw xmm0, xmm6 + pmullw xmm1, xmm7 + pmullw xmm2, xmm6 + pmullw xmm3, xmm7 + + paddsw xmm0, xmm1 + paddsw xmm2, xmm3 + + paddsw xmm0, xmm4 ;rounding + paddsw xmm2, xmm4 + psraw xmm0, 7 ;shift + psraw xmm2, 7 + packuswb xmm0, xmm2 ;pack back to byte +%if %1 + movdqu xmm1, [rdi] + pavgb xmm0, xmm1 +%endif + movdqu [rdi], xmm0 ;store the result + + lea rsi, [rsi + rax] + lea rdi, [rdi + rdx] + dec rcx +%endm + +SECTION .text + +global sym(aom_filter_block1d4_v2_sse2) PRIVATE +sym(aom_filter_block1d4_v2_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 6 + push rsi + push rdi + ; end prolog + + GET_PARAM_4 +.loop: + movd xmm0, [rsi] ;load src + movd xmm1, [rsi + rax] + + APPLY_FILTER_4 0 + jnz .loop + + ; begin epilog + pop rdi + pop rsi + UNSHADOW_ARGS + pop rbp + ret + +global sym(aom_filter_block1d8_v2_sse2) PRIVATE +sym(aom_filter_block1d8_v2_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 6 + SAVE_XMM 7 + push rsi + push rdi + ; end prolog + + GET_PARAM +.loop: + movq xmm0, [rsi] ;0 + movq xmm1, [rsi + rax] ;1 + + APPLY_FILTER_8 0 + jnz .loop + + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +global sym(aom_filter_block1d16_v2_sse2) PRIVATE +sym(aom_filter_block1d16_v2_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 6 + SAVE_XMM 7 + push rsi + push rdi + ; end prolog + + GET_PARAM +.loop: + movdqu xmm0, [rsi] ;0 + movdqu xmm1, [rsi + rax] ;1 + movdqa xmm2, xmm0 + movdqa xmm3, xmm1 + + APPLY_FILTER_16 0 + jnz .loop + + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +global sym(aom_filter_block1d4_h2_sse2) PRIVATE +sym(aom_filter_block1d4_h2_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 6 + push rsi + push rdi + ; end prolog + + GET_PARAM_4 +.loop: + movdqu xmm0, [rsi] ;load src + movdqa xmm1, xmm0 + psrldq xmm1, 1 + + APPLY_FILTER_4 0 + jnz .loop + + ; begin epilog + pop rdi + pop rsi + UNSHADOW_ARGS + pop rbp + ret + +global sym(aom_filter_block1d8_h2_sse2) PRIVATE +sym(aom_filter_block1d8_h2_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 6 + SAVE_XMM 7 + push rsi + push rdi + ; end prolog + + GET_PARAM +.loop: + movdqu xmm0, [rsi] ;load src + movdqa xmm1, xmm0 + psrldq xmm1, 1 + + APPLY_FILTER_8 0 + jnz .loop + + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +global sym(aom_filter_block1d16_h2_sse2) PRIVATE +sym(aom_filter_block1d16_h2_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 6 + SAVE_XMM 7 + push rsi + push rdi + ; end prolog + + GET_PARAM +.loop: + movdqu xmm0, [rsi] ;load src + movdqu xmm1, [rsi + 1] + movdqa xmm2, xmm0 + movdqa xmm3, xmm1 + + APPLY_FILTER_16 0 + jnz .loop + + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret diff --git a/third_party/aom/aom_dsp/x86/aom_subpixel_bilinear_ssse3.asm b/third_party/aom/aom_dsp/x86/aom_subpixel_bilinear_ssse3.asm new file mode 100644 index 0000000000..59edc49a9b --- /dev/null +++ b/third_party/aom/aom_dsp/x86/aom_subpixel_bilinear_ssse3.asm @@ -0,0 +1,267 @@ +; +; Copyright (c) 2016, Alliance for Open Media. All rights reserved +; +; This source code is subject to the terms of the BSD 2 Clause License and +; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License +; was not distributed with this source code in the LICENSE file, you can +; obtain it at www.aomedia.org/license/software. If the Alliance for Open +; Media Patent License 1.0 was not distributed with this source code in the +; PATENTS file, you can obtain it at www.aomedia.org/license/patent. +; + +; + +%include "aom_ports/x86_abi_support.asm" + +%macro GET_PARAM_4 0 + mov rdx, arg(5) ;filter ptr + mov rsi, arg(0) ;src_ptr + mov rdi, arg(2) ;output_ptr + mov ecx, 0x01000100 + + movdqa xmm3, [rdx] ;load filters + psrldq xmm3, 6 + packsswb xmm3, xmm3 + pshuflw xmm3, xmm3, 0b ;k3_k4 + + movd xmm2, ecx ;rounding_shift + pshufd xmm2, xmm2, 0 + + movsxd rax, DWORD PTR arg(1) ;pixels_per_line + movsxd rdx, DWORD PTR arg(3) ;out_pitch + movsxd rcx, DWORD PTR arg(4) ;output_height +%endm + +%macro APPLY_FILTER_4 1 + punpcklbw xmm0, xmm1 + pmaddubsw xmm0, xmm3 + + pmulhrsw xmm0, xmm2 ;rounding(+64)+shift(>>7) + packuswb xmm0, xmm0 ;pack to byte + +%if %1 + movd xmm1, [rdi] + pavgb xmm0, xmm1 +%endif + movd [rdi], xmm0 + lea rsi, [rsi + rax] + lea rdi, [rdi + rdx] + dec rcx +%endm + +%macro GET_PARAM 0 + mov rdx, arg(5) ;filter ptr + mov rsi, arg(0) ;src_ptr + mov rdi, arg(2) ;output_ptr + mov ecx, 0x01000100 + + movdqa xmm7, [rdx] ;load filters + psrldq xmm7, 6 + packsswb xmm7, xmm7 + pshuflw xmm7, xmm7, 0b ;k3_k4 + punpcklwd xmm7, xmm7 + + movd xmm6, ecx ;rounding_shift + pshufd xmm6, xmm6, 0 + + movsxd rax, DWORD PTR arg(1) ;pixels_per_line + movsxd rdx, DWORD PTR arg(3) ;out_pitch + movsxd rcx, DWORD PTR arg(4) ;output_height +%endm + +%macro APPLY_FILTER_8 1 + punpcklbw xmm0, xmm1 + pmaddubsw xmm0, xmm7 + + pmulhrsw xmm0, xmm6 ;rounding(+64)+shift(>>7) + packuswb xmm0, xmm0 ;pack back to byte + +%if %1 + movq xmm1, [rdi] + pavgb xmm0, xmm1 +%endif + movq [rdi], xmm0 ;store the result + + lea rsi, [rsi + rax] + lea rdi, [rdi + rdx] + dec rcx +%endm + +%macro APPLY_FILTER_16 1 + punpcklbw xmm0, xmm1 + punpckhbw xmm2, xmm1 + pmaddubsw xmm0, xmm7 + pmaddubsw xmm2, xmm7 + + pmulhrsw xmm0, xmm6 ;rounding(+64)+shift(>>7) + pmulhrsw xmm2, xmm6 + packuswb xmm0, xmm2 ;pack back to byte + +%if %1 + movdqu xmm1, [rdi] + pavgb xmm0, xmm1 +%endif + movdqu [rdi], xmm0 ;store the result + + lea rsi, [rsi + rax] + lea rdi, [rdi + rdx] + dec rcx +%endm + +SECTION .text + +global sym(aom_filter_block1d4_v2_ssse3) PRIVATE +sym(aom_filter_block1d4_v2_ssse3): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 6 + push rsi + push rdi + ; end prolog + + GET_PARAM_4 +.loop: + movd xmm0, [rsi] ;load src + movd xmm1, [rsi + rax] + + APPLY_FILTER_4 0 + jnz .loop + + ; begin epilog + pop rdi + pop rsi + UNSHADOW_ARGS + pop rbp + ret + +global sym(aom_filter_block1d8_v2_ssse3) PRIVATE +sym(aom_filter_block1d8_v2_ssse3): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 6 + SAVE_XMM 7 + push rsi + push rdi + ; end prolog + + GET_PARAM +.loop: + movq xmm0, [rsi] ;0 + movq xmm1, [rsi + rax] ;1 + + APPLY_FILTER_8 0 + jnz .loop + + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +global sym(aom_filter_block1d16_v2_ssse3) PRIVATE +sym(aom_filter_block1d16_v2_ssse3): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 6 + SAVE_XMM 7 + push rsi + push rdi + ; end prolog + + GET_PARAM +.loop: + movdqu xmm0, [rsi] ;0 + movdqu xmm1, [rsi + rax] ;1 + movdqa xmm2, xmm0 + + APPLY_FILTER_16 0 + jnz .loop + + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +global sym(aom_filter_block1d4_h2_ssse3) PRIVATE +sym(aom_filter_block1d4_h2_ssse3): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 6 + push rsi + push rdi + ; end prolog + + GET_PARAM_4 +.loop: + movdqu xmm0, [rsi] ;load src + movdqa xmm1, xmm0 + psrldq xmm1, 1 + + APPLY_FILTER_4 0 + jnz .loop + + ; begin epilog + pop rdi + pop rsi + UNSHADOW_ARGS + pop rbp + ret + +global sym(aom_filter_block1d8_h2_ssse3) PRIVATE +sym(aom_filter_block1d8_h2_ssse3): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 6 + SAVE_XMM 7 + push rsi + push rdi + ; end prolog + + GET_PARAM +.loop: + movdqu xmm0, [rsi] ;load src + movdqa xmm1, xmm0 + psrldq xmm1, 1 + + APPLY_FILTER_8 0 + jnz .loop + + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +global sym(aom_filter_block1d16_h2_ssse3) PRIVATE +sym(aom_filter_block1d16_h2_ssse3): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 6 + SAVE_XMM 7 + push rsi + push rdi + ; end prolog + + GET_PARAM +.loop: + movdqu xmm0, [rsi] ;load src + movdqu xmm1, [rsi + 1] + movdqa xmm2, xmm0 + + APPLY_FILTER_16 0 + jnz .loop + + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret diff --git a/third_party/aom/aom_dsp/x86/blend_a64_hmask_sse4.c b/third_party/aom/aom_dsp/x86/blend_a64_hmask_sse4.c new file mode 100644 index 0000000000..4f5e3f8c1b --- /dev/null +++ b/third_party/aom/aom_dsp/x86/blend_a64_hmask_sse4.c @@ -0,0 +1,34 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include "aom/aom_integer.h" + +#include "config/aom_dsp_rtcd.h" + +// To start out, just dispatch to the function using the 2D mask and +// pass mask stride as 0. This can be improved upon if necessary. + +void aom_blend_a64_hmask_sse4_1(uint8_t *dst, uint32_t dst_stride, + const uint8_t *src0, uint32_t src0_stride, + const uint8_t *src1, uint32_t src1_stride, + const uint8_t *mask, int w, int h) { + aom_blend_a64_mask_sse4_1(dst, dst_stride, src0, src0_stride, src1, + src1_stride, mask, 0, w, h, 0, 0); +} + +void aom_highbd_blend_a64_hmask_sse4_1( + uint8_t *dst_8, uint32_t dst_stride, const uint8_t *src0_8, + uint32_t src0_stride, const uint8_t *src1_8, uint32_t src1_stride, + const uint8_t *mask, int w, int h, int bd) { + aom_highbd_blend_a64_mask_sse4_1(dst_8, dst_stride, src0_8, src0_stride, + src1_8, src1_stride, mask, 0, w, h, 0, 0, + bd); +} diff --git a/third_party/aom/aom_dsp/x86/blend_a64_mask_avx2.c b/third_party/aom/aom_dsp/x86/blend_a64_mask_avx2.c new file mode 100644 index 0000000000..67fb4d32bd --- /dev/null +++ b/third_party/aom/aom_dsp/x86/blend_a64_mask_avx2.c @@ -0,0 +1,900 @@ +/* + * Copyright (c) 2018, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include // SSE4.1 +#include // AVX2 + +#include + +#include "aom/aom_integer.h" +#include "aom_ports/mem.h" +#include "aom_dsp/aom_dsp_common.h" + +#include "aom_dsp/x86/synonyms.h" +#include "aom_dsp/x86/synonyms_avx2.h" +#include "aom_dsp/x86/blend_sse4.h" +#include "aom_dsp/x86/blend_mask_sse4.h" + +#include "config/aom_dsp_rtcd.h" + +static INLINE void blend_a64_d16_mask_w16_avx2( + uint8_t *dst, const CONV_BUF_TYPE *src0, const CONV_BUF_TYPE *src1, + const __m256i *m0, const __m256i *v_round_offset, const __m256i *v_maxval, + int shift) { + const __m256i max_minus_m0 = _mm256_sub_epi16(*v_maxval, *m0); + const __m256i s0_0 = yy_loadu_256(src0); + const __m256i s1_0 = yy_loadu_256(src1); + __m256i res0_lo = _mm256_madd_epi16(_mm256_unpacklo_epi16(s0_0, s1_0), + _mm256_unpacklo_epi16(*m0, max_minus_m0)); + __m256i res0_hi = _mm256_madd_epi16(_mm256_unpackhi_epi16(s0_0, s1_0), + _mm256_unpackhi_epi16(*m0, max_minus_m0)); + res0_lo = + _mm256_srai_epi32(_mm256_sub_epi32(res0_lo, *v_round_offset), shift); + res0_hi = + _mm256_srai_epi32(_mm256_sub_epi32(res0_hi, *v_round_offset), shift); + const __m256i res0 = _mm256_packs_epi32(res0_lo, res0_hi); + __m256i res = _mm256_packus_epi16(res0, res0); + res = _mm256_permute4x64_epi64(res, 0xd8); + _mm_storeu_si128((__m128i *)(dst), _mm256_castsi256_si128(res)); +} + +static INLINE void blend_a64_d16_mask_w32_avx2( + uint8_t *dst, const CONV_BUF_TYPE *src0, const CONV_BUF_TYPE *src1, + const __m256i *m0, const __m256i *m1, const __m256i *v_round_offset, + const __m256i *v_maxval, int shift) { + const __m256i max_minus_m0 = _mm256_sub_epi16(*v_maxval, *m0); + const __m256i max_minus_m1 = _mm256_sub_epi16(*v_maxval, *m1); + const __m256i s0_0 = yy_loadu_256(src0); + const __m256i s0_1 = yy_loadu_256(src0 + 16); + const __m256i s1_0 = yy_loadu_256(src1); + const __m256i s1_1 = yy_loadu_256(src1 + 16); + __m256i res0_lo = _mm256_madd_epi16(_mm256_unpacklo_epi16(s0_0, s1_0), + _mm256_unpacklo_epi16(*m0, max_minus_m0)); + __m256i res0_hi = _mm256_madd_epi16(_mm256_unpackhi_epi16(s0_0, s1_0), + _mm256_unpackhi_epi16(*m0, max_minus_m0)); + __m256i res1_lo = _mm256_madd_epi16(_mm256_unpacklo_epi16(s0_1, s1_1), + _mm256_unpacklo_epi16(*m1, max_minus_m1)); + __m256i res1_hi = _mm256_madd_epi16(_mm256_unpackhi_epi16(s0_1, s1_1), + _mm256_unpackhi_epi16(*m1, max_minus_m1)); + res0_lo = + _mm256_srai_epi32(_mm256_sub_epi32(res0_lo, *v_round_offset), shift); + res0_hi = + _mm256_srai_epi32(_mm256_sub_epi32(res0_hi, *v_round_offset), shift); + res1_lo = + _mm256_srai_epi32(_mm256_sub_epi32(res1_lo, *v_round_offset), shift); + res1_hi = + _mm256_srai_epi32(_mm256_sub_epi32(res1_hi, *v_round_offset), shift); + const __m256i res0 = _mm256_packs_epi32(res0_lo, res0_hi); + const __m256i res1 = _mm256_packs_epi32(res1_lo, res1_hi); + __m256i res = _mm256_packus_epi16(res0, res1); + res = _mm256_permute4x64_epi64(res, 0xd8); + _mm256_storeu_si256((__m256i *)(dst), res); +} + +static INLINE void lowbd_blend_a64_d16_mask_subw0_subh0_w16_avx2( + uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, + uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, + const __m256i *round_offset, int shift) { + const __m256i v_maxval = _mm256_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + for (int i = 0; i < h; ++i) { + const __m128i m = xx_loadu_128(mask); + const __m256i m0 = _mm256_cvtepu8_epi16(m); + + blend_a64_d16_mask_w16_avx2(dst, src0, src1, &m0, round_offset, &v_maxval, + shift); + mask += mask_stride; + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + } +} + +static INLINE void lowbd_blend_a64_d16_mask_subw0_subh0_w32_avx2( + uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, + uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, int w, + const __m256i *round_offset, int shift) { + const __m256i v_maxval = _mm256_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + for (int i = 0; i < h; ++i) { + for (int j = 0; j < w; j += 32) { + const __m256i m = yy_loadu_256(mask + j); + const __m256i m0 = _mm256_cvtepu8_epi16(_mm256_castsi256_si128(m)); + const __m256i m1 = _mm256_cvtepu8_epi16(_mm256_extracti128_si256(m, 1)); + + blend_a64_d16_mask_w32_avx2(dst + j, src0 + j, src1 + j, &m0, &m1, + round_offset, &v_maxval, shift); + } + mask += mask_stride; + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + } +} + +static INLINE void lowbd_blend_a64_d16_mask_subw1_subh1_w16_avx2( + uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, + uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, + const __m256i *round_offset, int shift) { + const __m256i v_maxval = _mm256_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + const __m256i one_b = _mm256_set1_epi8(1); + const __m256i two_w = _mm256_set1_epi16(2); + for (int i = 0; i < h; ++i) { + const __m256i m_i00 = yy_loadu_256(mask); + const __m256i m_i10 = yy_loadu_256(mask + mask_stride); + + const __m256i m0_ac = _mm256_adds_epu8(m_i00, m_i10); + const __m256i m0_acbd = _mm256_maddubs_epi16(m0_ac, one_b); + const __m256i m0 = _mm256_srli_epi16(_mm256_add_epi16(m0_acbd, two_w), 2); + + blend_a64_d16_mask_w16_avx2(dst, src0, src1, &m0, round_offset, &v_maxval, + shift); + mask += mask_stride << 1; + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + } +} + +static INLINE void lowbd_blend_a64_d16_mask_subw1_subh1_w32_avx2( + uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, + uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, int w, + const __m256i *round_offset, int shift) { + const __m256i v_maxval = _mm256_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + const __m256i one_b = _mm256_set1_epi8(1); + const __m256i two_w = _mm256_set1_epi16(2); + for (int i = 0; i < h; ++i) { + for (int j = 0; j < w; j += 32) { + const __m256i m_i00 = yy_loadu_256(mask + 2 * j); + const __m256i m_i01 = yy_loadu_256(mask + 2 * j + 32); + const __m256i m_i10 = yy_loadu_256(mask + mask_stride + 2 * j); + const __m256i m_i11 = yy_loadu_256(mask + mask_stride + 2 * j + 32); + + const __m256i m0_ac = _mm256_adds_epu8(m_i00, m_i10); + const __m256i m1_ac = _mm256_adds_epu8(m_i01, m_i11); + const __m256i m0_acbd = _mm256_maddubs_epi16(m0_ac, one_b); + const __m256i m1_acbd = _mm256_maddubs_epi16(m1_ac, one_b); + const __m256i m0 = _mm256_srli_epi16(_mm256_add_epi16(m0_acbd, two_w), 2); + const __m256i m1 = _mm256_srli_epi16(_mm256_add_epi16(m1_acbd, two_w), 2); + + blend_a64_d16_mask_w32_avx2(dst + j, src0 + j, src1 + j, &m0, &m1, + round_offset, &v_maxval, shift); + } + mask += mask_stride << 1; + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + } +} + +static INLINE void lowbd_blend_a64_d16_mask_subw1_subh0_w16_avx2( + uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, + uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, int w, + const __m256i *round_offset, int shift) { + const __m256i v_maxval = _mm256_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + const __m256i one_b = _mm256_set1_epi8(1); + const __m256i zeros = _mm256_setzero_si256(); + for (int i = 0; i < h; ++i) { + for (int j = 0; j < w; j += 16) { + const __m256i m_i00 = yy_loadu_256(mask + 2 * j); + const __m256i m0_ac = _mm256_maddubs_epi16(m_i00, one_b); + const __m256i m0 = _mm256_avg_epu16(m0_ac, zeros); + + blend_a64_d16_mask_w16_avx2(dst + j, src0 + j, src1 + j, &m0, + round_offset, &v_maxval, shift); + } + mask += mask_stride; + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + } +} + +static INLINE void lowbd_blend_a64_d16_mask_subw1_subh0_w32_avx2( + uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, + uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, int w, + const __m256i *round_offset, int shift) { + const __m256i v_maxval = _mm256_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + const __m256i one_b = _mm256_set1_epi8(1); + const __m256i zeros = _mm256_setzero_si256(); + for (int i = 0; i < h; ++i) { + for (int j = 0; j < w; j += 32) { + const __m256i m_i00 = yy_loadu_256(mask + 2 * j); + const __m256i m_i01 = yy_loadu_256(mask + 2 * j + 32); + const __m256i m0_ac = _mm256_maddubs_epi16(m_i00, one_b); + const __m256i m1_ac = _mm256_maddubs_epi16(m_i01, one_b); + const __m256i m0 = _mm256_avg_epu16(m0_ac, zeros); + const __m256i m1 = _mm256_avg_epu16(m1_ac, zeros); + + blend_a64_d16_mask_w32_avx2(dst + j, src0 + j, src1 + j, &m0, &m1, + round_offset, &v_maxval, shift); + } + mask += mask_stride; + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + } +} + +static INLINE void lowbd_blend_a64_d16_mask_subw0_subh1_w16_avx2( + uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, + uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, int w, + const __m256i *round_offset, int shift) { + const __m256i v_maxval = _mm256_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + const __m128i zeros = _mm_setzero_si128(); + for (int i = 0; i < h; ++i) { + for (int j = 0; j < w; j += 16) { + const __m128i m_i00 = xx_loadu_128(mask + j); + const __m128i m_i10 = xx_loadu_128(mask + mask_stride + j); + + const __m128i m_ac = _mm_avg_epu8(_mm_adds_epu8(m_i00, m_i10), zeros); + const __m256i m0 = _mm256_cvtepu8_epi16(m_ac); + + blend_a64_d16_mask_w16_avx2(dst + j, src0 + j, src1 + j, &m0, + round_offset, &v_maxval, shift); + } + mask += mask_stride << 1; + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + } +} + +static INLINE void lowbd_blend_a64_d16_mask_subw0_subh1_w32_avx2( + uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, + uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, int w, + const __m256i *round_offset, int shift) { + const __m256i v_maxval = _mm256_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + const __m256i zeros = _mm256_setzero_si256(); + for (int i = 0; i < h; ++i) { + for (int j = 0; j < w; j += 32) { + const __m256i m_i00 = yy_loadu_256(mask + j); + const __m256i m_i10 = yy_loadu_256(mask + mask_stride + j); + + const __m256i m_ac = + _mm256_avg_epu8(_mm256_adds_epu8(m_i00, m_i10), zeros); + const __m256i m0 = _mm256_cvtepu8_epi16(_mm256_castsi256_si128(m_ac)); + const __m256i m1 = + _mm256_cvtepu8_epi16(_mm256_extracti128_si256(m_ac, 1)); + + blend_a64_d16_mask_w32_avx2(dst + j, src0 + j, src1 + j, &m0, &m1, + round_offset, &v_maxval, shift); + } + mask += mask_stride << 1; + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + } +} + +void aom_lowbd_blend_a64_d16_mask_avx2( + uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, + uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, + ConvolveParams *conv_params) { + const int bd = 8; + const int round_bits = + 2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1; + + const int round_offset = + ((1 << (round_bits + bd)) + (1 << (round_bits + bd - 1)) - + (1 << (round_bits - 1))) + << AOM_BLEND_A64_ROUND_BITS; + + const int shift = round_bits + AOM_BLEND_A64_ROUND_BITS; + assert(IMPLIES((void *)src0 == dst, src0_stride == dst_stride)); + assert(IMPLIES((void *)src1 == dst, src1_stride == dst_stride)); + + assert(h >= 4); + assert(w >= 4); + assert(IS_POWER_OF_TWO(h)); + assert(IS_POWER_OF_TWO(w)); + const __m128i v_round_offset = _mm_set1_epi32(round_offset); + const __m256i y_round_offset = _mm256_set1_epi32(round_offset); + + if (subw == 0 && subh == 0) { + switch (w) { + case 4: + aom_lowbd_blend_a64_d16_mask_subw0_subh0_w4_sse4_1( + dst, dst_stride, src0, src0_stride, src1, src1_stride, mask, + mask_stride, h, &v_round_offset, shift); + break; + case 8: + aom_lowbd_blend_a64_d16_mask_subw0_subh0_w8_sse4_1( + dst, dst_stride, src0, src0_stride, src1, src1_stride, mask, + mask_stride, h, &v_round_offset, shift); + break; + case 16: + lowbd_blend_a64_d16_mask_subw0_subh0_w16_avx2( + dst, dst_stride, src0, src0_stride, src1, src1_stride, mask, + mask_stride, h, &y_round_offset, shift); + break; + default: + lowbd_blend_a64_d16_mask_subw0_subh0_w32_avx2( + dst, dst_stride, src0, src0_stride, src1, src1_stride, mask, + mask_stride, h, w, &y_round_offset, shift); + break; + } + } else if (subw == 1 && subh == 1) { + switch (w) { + case 4: + aom_lowbd_blend_a64_d16_mask_subw1_subh1_w4_sse4_1( + dst, dst_stride, src0, src0_stride, src1, src1_stride, mask, + mask_stride, h, &v_round_offset, shift); + break; + case 8: + aom_lowbd_blend_a64_d16_mask_subw1_subh1_w8_sse4_1( + dst, dst_stride, src0, src0_stride, src1, src1_stride, mask, + mask_stride, h, &v_round_offset, shift); + break; + case 16: + lowbd_blend_a64_d16_mask_subw1_subh1_w16_avx2( + dst, dst_stride, src0, src0_stride, src1, src1_stride, mask, + mask_stride, h, &y_round_offset, shift); + break; + default: + lowbd_blend_a64_d16_mask_subw1_subh1_w32_avx2( + dst, dst_stride, src0, src0_stride, src1, src1_stride, mask, + mask_stride, h, w, &y_round_offset, shift); + break; + } + } else if (subw == 1 && subh == 0) { + switch (w) { + case 4: + aom_lowbd_blend_a64_d16_mask_subw1_subh0_w4_sse4_1( + dst, dst_stride, src0, src0_stride, src1, src1_stride, mask, + mask_stride, h, &v_round_offset, shift); + break; + case 8: + aom_lowbd_blend_a64_d16_mask_subw1_subh0_w8_sse4_1( + dst, dst_stride, src0, src0_stride, src1, src1_stride, mask, + mask_stride, h, &v_round_offset, shift); + break; + case 16: + lowbd_blend_a64_d16_mask_subw1_subh0_w16_avx2( + dst, dst_stride, src0, src0_stride, src1, src1_stride, mask, + mask_stride, h, w, &y_round_offset, shift); + break; + default: + lowbd_blend_a64_d16_mask_subw1_subh0_w32_avx2( + dst, dst_stride, src0, src0_stride, src1, src1_stride, mask, + mask_stride, h, w, &y_round_offset, shift); + break; + } + } else { + switch (w) { + case 4: + aom_lowbd_blend_a64_d16_mask_subw0_subh1_w4_sse4_1( + dst, dst_stride, src0, src0_stride, src1, src1_stride, mask, + mask_stride, h, &v_round_offset, shift); + break; + case 8: + aom_lowbd_blend_a64_d16_mask_subw0_subh1_w8_sse4_1( + dst, dst_stride, src0, src0_stride, src1, src1_stride, mask, + mask_stride, h, &v_round_offset, shift); + break; + case 16: + lowbd_blend_a64_d16_mask_subw0_subh1_w16_avx2( + dst, dst_stride, src0, src0_stride, src1, src1_stride, mask, + mask_stride, h, w, &y_round_offset, shift); + break; + default: + lowbd_blend_a64_d16_mask_subw0_subh1_w32_avx2( + dst, dst_stride, src0, src0_stride, src1, src1_stride, mask, + mask_stride, h, w, &y_round_offset, shift); + break; + } + } +} + +static INLINE __m256i blend_16_u8_avx2(const uint8_t *src0, const uint8_t *src1, + const __m256i *v_m0_b, + const __m256i *v_m1_b, + const int32_t bits) { + const __m256i v_s0_b = _mm256_castsi128_si256(xx_loadu_128(src0)); + const __m256i v_s1_b = _mm256_castsi128_si256(xx_loadu_128(src1)); + const __m256i v_s0_s_b = _mm256_permute4x64_epi64(v_s0_b, 0xd8); + const __m256i v_s1_s_b = _mm256_permute4x64_epi64(v_s1_b, 0xd8); + + const __m256i v_p0_w = + _mm256_maddubs_epi16(_mm256_unpacklo_epi8(v_s0_s_b, v_s1_s_b), + _mm256_unpacklo_epi8(*v_m0_b, *v_m1_b)); + + const __m256i v_res0_w = yy_roundn_epu16(v_p0_w, bits); + const __m256i v_res_b = _mm256_packus_epi16(v_res0_w, v_res0_w); + const __m256i v_res = _mm256_permute4x64_epi64(v_res_b, 0xd8); + return v_res; +} + +static INLINE __m256i blend_32_u8_avx2(const uint8_t *src0, const uint8_t *src1, + const __m256i *v_m0_b, + const __m256i *v_m1_b, + const int32_t bits) { + const __m256i v_s0_b = yy_loadu_256(src0); + const __m256i v_s1_b = yy_loadu_256(src1); + + const __m256i v_p0_w = + _mm256_maddubs_epi16(_mm256_unpacklo_epi8(v_s0_b, v_s1_b), + _mm256_unpacklo_epi8(*v_m0_b, *v_m1_b)); + const __m256i v_p1_w = + _mm256_maddubs_epi16(_mm256_unpackhi_epi8(v_s0_b, v_s1_b), + _mm256_unpackhi_epi8(*v_m0_b, *v_m1_b)); + + const __m256i v_res0_w = yy_roundn_epu16(v_p0_w, bits); + const __m256i v_res1_w = yy_roundn_epu16(v_p1_w, bits); + const __m256i v_res = _mm256_packus_epi16(v_res0_w, v_res1_w); + return v_res; +} + +static INLINE void blend_a64_mask_sx_sy_w16_avx2( + uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, + uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h) { + const __m256i v_zmask_b = _mm256_set1_epi16(0xFF); + const __m256i v_maxval_b = _mm256_set1_epi8(AOM_BLEND_A64_MAX_ALPHA); + do { + const __m256i v_ral_b = yy_loadu_256(mask); + const __m256i v_rbl_b = yy_loadu_256(mask + mask_stride); + const __m256i v_rvsl_b = _mm256_add_epi8(v_ral_b, v_rbl_b); + const __m256i v_rvsal_w = _mm256_and_si256(v_rvsl_b, v_zmask_b); + const __m256i v_rvsbl_w = + _mm256_and_si256(_mm256_srli_si256(v_rvsl_b, 1), v_zmask_b); + const __m256i v_rsl_w = _mm256_add_epi16(v_rvsal_w, v_rvsbl_w); + + const __m256i v_m0_w = yy_roundn_epu16(v_rsl_w, 2); + const __m256i v_m0_b = _mm256_packus_epi16(v_m0_w, v_m0_w); + const __m256i v_m1_b = _mm256_sub_epi8(v_maxval_b, v_m0_b); + + const __m256i y_res_b = blend_16_u8_avx2(src0, src1, &v_m0_b, &v_m1_b, + AOM_BLEND_A64_ROUND_BITS); + + xx_storeu_128(dst, _mm256_castsi256_si128(y_res_b)); + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += 2 * mask_stride; + } while (--h); +} + +static INLINE void blend_a64_mask_sx_sy_w32n_avx2( + uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, + uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int w, int h) { + const __m256i v_maxval_b = _mm256_set1_epi8(AOM_BLEND_A64_MAX_ALPHA); + const __m256i v_zmask_b = _mm256_set1_epi16(0xFF); + do { + int c; + for (c = 0; c < w; c += 32) { + const __m256i v_ral_b = yy_loadu_256(mask + 2 * c); + const __m256i v_rah_b = yy_loadu_256(mask + 2 * c + 32); + const __m256i v_rbl_b = yy_loadu_256(mask + mask_stride + 2 * c); + const __m256i v_rbh_b = yy_loadu_256(mask + mask_stride + 2 * c + 32); + const __m256i v_rvsl_b = _mm256_add_epi8(v_ral_b, v_rbl_b); + const __m256i v_rvsh_b = _mm256_add_epi8(v_rah_b, v_rbh_b); + const __m256i v_rvsal_w = _mm256_and_si256(v_rvsl_b, v_zmask_b); + const __m256i v_rvsah_w = _mm256_and_si256(v_rvsh_b, v_zmask_b); + const __m256i v_rvsbl_w = + _mm256_and_si256(_mm256_srli_si256(v_rvsl_b, 1), v_zmask_b); + const __m256i v_rvsbh_w = + _mm256_and_si256(_mm256_srli_si256(v_rvsh_b, 1), v_zmask_b); + const __m256i v_rsl_w = _mm256_add_epi16(v_rvsal_w, v_rvsbl_w); + const __m256i v_rsh_w = _mm256_add_epi16(v_rvsah_w, v_rvsbh_w); + + const __m256i v_m0l_w = yy_roundn_epu16(v_rsl_w, 2); + const __m256i v_m0h_w = yy_roundn_epu16(v_rsh_w, 2); + const __m256i v_m0_b = + _mm256_permute4x64_epi64(_mm256_packus_epi16(v_m0l_w, v_m0h_w), 0xd8); + const __m256i v_m1_b = _mm256_sub_epi8(v_maxval_b, v_m0_b); + + const __m256i v_res_b = blend_32_u8_avx2( + src0 + c, src1 + c, &v_m0_b, &v_m1_b, AOM_BLEND_A64_ROUND_BITS); + + yy_storeu_256(dst + c, v_res_b); + } + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += 2 * mask_stride; + } while (--h); +} + +static INLINE void blend_a64_mask_sx_sy_avx2( + uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, + uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int w, int h) { + const __m128i v_shuffle_b = xx_loadu_128(g_blend_a64_mask_shuffle); + const __m128i v_maxval_b = _mm_set1_epi8(AOM_BLEND_A64_MAX_ALPHA); + const __m128i _r = _mm_set1_epi16(1 << (15 - AOM_BLEND_A64_ROUND_BITS)); + switch (w) { + case 4: + do { + const __m128i v_ra_b = xx_loadl_64(mask); + const __m128i v_rb_b = xx_loadl_64(mask + mask_stride); + const __m128i v_rvs_b = _mm_add_epi8(v_ra_b, v_rb_b); + const __m128i v_r_s_b = _mm_shuffle_epi8(v_rvs_b, v_shuffle_b); + const __m128i v_r0_s_w = _mm_cvtepu8_epi16(v_r_s_b); + const __m128i v_r1_s_w = _mm_cvtepu8_epi16(_mm_srli_si128(v_r_s_b, 8)); + const __m128i v_rs_w = _mm_add_epi16(v_r0_s_w, v_r1_s_w); + const __m128i v_m0_w = xx_roundn_epu16(v_rs_w, 2); + const __m128i v_m0_b = _mm_packus_epi16(v_m0_w, v_m0_w); + const __m128i v_m1_b = _mm_sub_epi8(v_maxval_b, v_m0_b); + + const __m128i v_res_b = blend_4_u8(src0, src1, &v_m0_b, &v_m1_b, &_r); + + xx_storel_32(dst, v_res_b); + + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += 2 * mask_stride; + } while (--h); + break; + case 8: + do { + const __m128i v_ra_b = xx_loadu_128(mask); + const __m128i v_rb_b = xx_loadu_128(mask + mask_stride); + const __m128i v_rvs_b = _mm_add_epi8(v_ra_b, v_rb_b); + const __m128i v_r_s_b = _mm_shuffle_epi8(v_rvs_b, v_shuffle_b); + const __m128i v_r0_s_w = _mm_cvtepu8_epi16(v_r_s_b); + const __m128i v_r1_s_w = _mm_cvtepu8_epi16(_mm_srli_si128(v_r_s_b, 8)); + const __m128i v_rs_w = _mm_add_epi16(v_r0_s_w, v_r1_s_w); + const __m128i v_m0_w = xx_roundn_epu16(v_rs_w, 2); + const __m128i v_m0_b = _mm_packus_epi16(v_m0_w, v_m0_w); + const __m128i v_m1_b = _mm_sub_epi8(v_maxval_b, v_m0_b); + + const __m128i v_res_b = blend_8_u8(src0, src1, &v_m0_b, &v_m1_b, &_r); + + xx_storel_64(dst, v_res_b); + + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += 2 * mask_stride; + } while (--h); + break; + case 16: + blend_a64_mask_sx_sy_w16_avx2(dst, dst_stride, src0, src0_stride, src1, + src1_stride, mask, mask_stride, h); + break; + default: + blend_a64_mask_sx_sy_w32n_avx2(dst, dst_stride, src0, src0_stride, src1, + src1_stride, mask, mask_stride, w, h); + break; + } +} + +static INLINE void blend_a64_mask_sx_w16_avx2( + uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, + uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h) { + const __m256i v_maxval_b = _mm256_set1_epi8(AOM_BLEND_A64_MAX_ALPHA); + const __m256i v_zmask_b = _mm256_set1_epi16(0xff); + do { + const __m256i v_rl_b = yy_loadu_256(mask); + const __m256i v_al_b = + _mm256_avg_epu8(v_rl_b, _mm256_srli_si256(v_rl_b, 1)); + + const __m256i v_m0_w = _mm256_and_si256(v_al_b, v_zmask_b); + const __m256i v_m0_b = _mm256_packus_epi16(v_m0_w, _mm256_setzero_si256()); + const __m256i v_m1_b = _mm256_sub_epi8(v_maxval_b, v_m0_b); + + const __m256i v_res_b = blend_16_u8_avx2(src0, src1, &v_m0_b, &v_m1_b, + AOM_BLEND_A64_ROUND_BITS); + + xx_storeu_128(dst, _mm256_castsi256_si128(v_res_b)); + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += mask_stride; + } while (--h); +} + +static INLINE void blend_a64_mask_sx_w32n_avx2( + uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, + uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int w, int h) { + const __m256i v_shuffle_b = yy_loadu_256(g_blend_a64_mask_shuffle); + const __m256i v_maxval_b = _mm256_set1_epi8(AOM_BLEND_A64_MAX_ALPHA); + do { + int c; + for (c = 0; c < w; c += 32) { + const __m256i v_r0_b = yy_loadu_256(mask + 2 * c); + const __m256i v_r1_b = yy_loadu_256(mask + 2 * c + 32); + const __m256i v_r0_s_b = _mm256_shuffle_epi8(v_r0_b, v_shuffle_b); + const __m256i v_r1_s_b = _mm256_shuffle_epi8(v_r1_b, v_shuffle_b); + const __m256i v_al_b = + _mm256_avg_epu8(v_r0_s_b, _mm256_srli_si256(v_r0_s_b, 8)); + const __m256i v_ah_b = + _mm256_avg_epu8(v_r1_s_b, _mm256_srli_si256(v_r1_s_b, 8)); + + const __m256i v_m0_b = + _mm256_permute4x64_epi64(_mm256_unpacklo_epi64(v_al_b, v_ah_b), 0xd8); + const __m256i v_m1_b = _mm256_sub_epi8(v_maxval_b, v_m0_b); + + const __m256i v_res_b = blend_32_u8_avx2( + src0 + c, src1 + c, &v_m0_b, &v_m1_b, AOM_BLEND_A64_ROUND_BITS); + + yy_storeu_256(dst + c, v_res_b); + } + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += mask_stride; + } while (--h); +} + +static INLINE void blend_a64_mask_sx_avx2( + uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, + uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int w, int h) { + const __m128i v_shuffle_b = xx_loadu_128(g_blend_a64_mask_shuffle); + const __m128i v_maxval_b = _mm_set1_epi8(AOM_BLEND_A64_MAX_ALPHA); + const __m128i _r = _mm_set1_epi16(1 << (15 - AOM_BLEND_A64_ROUND_BITS)); + switch (w) { + case 4: + do { + const __m128i v_r_b = xx_loadl_64(mask); + const __m128i v_r0_s_b = _mm_shuffle_epi8(v_r_b, v_shuffle_b); + const __m128i v_r_lo_b = _mm_unpacklo_epi64(v_r0_s_b, v_r0_s_b); + const __m128i v_r_hi_b = _mm_unpackhi_epi64(v_r0_s_b, v_r0_s_b); + const __m128i v_m0_b = _mm_avg_epu8(v_r_lo_b, v_r_hi_b); + const __m128i v_m1_b = _mm_sub_epi8(v_maxval_b, v_m0_b); + + const __m128i v_res_b = blend_4_u8(src0, src1, &v_m0_b, &v_m1_b, &_r); + + xx_storel_32(dst, v_res_b); + + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += mask_stride; + } while (--h); + break; + case 8: + do { + const __m128i v_r_b = xx_loadu_128(mask); + const __m128i v_r0_s_b = _mm_shuffle_epi8(v_r_b, v_shuffle_b); + const __m128i v_r_lo_b = _mm_unpacklo_epi64(v_r0_s_b, v_r0_s_b); + const __m128i v_r_hi_b = _mm_unpackhi_epi64(v_r0_s_b, v_r0_s_b); + const __m128i v_m0_b = _mm_avg_epu8(v_r_lo_b, v_r_hi_b); + const __m128i v_m1_b = _mm_sub_epi8(v_maxval_b, v_m0_b); + + const __m128i v_res_b = blend_8_u8(src0, src1, &v_m0_b, &v_m1_b, &_r); + + xx_storel_64(dst, v_res_b); + + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += mask_stride; + } while (--h); + break; + case 16: + blend_a64_mask_sx_w16_avx2(dst, dst_stride, src0, src0_stride, src1, + src1_stride, mask, mask_stride, h); + break; + default: + blend_a64_mask_sx_w32n_avx2(dst, dst_stride, src0, src0_stride, src1, + src1_stride, mask, mask_stride, w, h); + break; + } +} + +static INLINE void blend_a64_mask_sy_w16_avx2( + uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, + uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h) { + const __m128i _r = _mm_set1_epi16(1 << (15 - AOM_BLEND_A64_ROUND_BITS)); + const __m128i v_maxval_b = _mm_set1_epi8(AOM_BLEND_A64_MAX_ALPHA); + do { + const __m128i v_ra_b = xx_loadu_128(mask); + const __m128i v_rb_b = xx_loadu_128(mask + mask_stride); + const __m128i v_m0_b = _mm_avg_epu8(v_ra_b, v_rb_b); + + const __m128i v_m1_b = _mm_sub_epi16(v_maxval_b, v_m0_b); + const __m128i v_res_b = blend_16_u8(src0, src1, &v_m0_b, &v_m1_b, &_r); + + xx_storeu_128(dst, v_res_b); + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += 2 * mask_stride; + } while (--h); +} + +static INLINE void blend_a64_mask_sy_w32n_avx2( + uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, + uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int w, int h) { + const __m256i v_maxval_b = _mm256_set1_epi8(AOM_BLEND_A64_MAX_ALPHA); + do { + int c; + for (c = 0; c < w; c += 32) { + const __m256i v_ra_b = yy_loadu_256(mask + c); + const __m256i v_rb_b = yy_loadu_256(mask + c + mask_stride); + const __m256i v_m0_b = _mm256_avg_epu8(v_ra_b, v_rb_b); + const __m256i v_m1_b = _mm256_sub_epi8(v_maxval_b, v_m0_b); + const __m256i v_res_b = blend_32_u8_avx2( + src0 + c, src1 + c, &v_m0_b, &v_m1_b, AOM_BLEND_A64_ROUND_BITS); + + yy_storeu_256(dst + c, v_res_b); + } + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += 2 * mask_stride; + } while (--h); +} + +static INLINE void blend_a64_mask_sy_avx2( + uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, + uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int w, int h) { + const __m128i _r = _mm_set1_epi16(1 << (15 - AOM_BLEND_A64_ROUND_BITS)); + const __m128i v_maxval_b = _mm_set1_epi8(AOM_BLEND_A64_MAX_ALPHA); + switch (w) { + case 4: + do { + const __m128i v_ra_b = xx_loadl_32(mask); + const __m128i v_rb_b = xx_loadl_32(mask + mask_stride); + const __m128i v_m0_b = _mm_avg_epu8(v_ra_b, v_rb_b); + const __m128i v_m1_b = _mm_sub_epi8(v_maxval_b, v_m0_b); + const __m128i v_res_b = blend_4_u8(src0, src1, &v_m0_b, &v_m1_b, &_r); + + xx_storel_32(dst, v_res_b); + + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += 2 * mask_stride; + } while (--h); + break; + case 8: + do { + const __m128i v_ra_b = xx_loadl_64(mask); + const __m128i v_rb_b = xx_loadl_64(mask + mask_stride); + const __m128i v_m0_b = _mm_avg_epu8(v_ra_b, v_rb_b); + const __m128i v_m1_b = _mm_sub_epi8(v_maxval_b, v_m0_b); + const __m128i v_res_b = blend_8_u8(src0, src1, &v_m0_b, &v_m1_b, &_r); + + xx_storel_64(dst, v_res_b); + + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += 2 * mask_stride; + } while (--h); + break; + case 16: + blend_a64_mask_sy_w16_avx2(dst, dst_stride, src0, src0_stride, src1, + src1_stride, mask, mask_stride, h); + break; + default: + blend_a64_mask_sy_w32n_avx2(dst, dst_stride, src0, src0_stride, src1, + src1_stride, mask, mask_stride, w, h); + } +} + +static INLINE void blend_a64_mask_w32n_avx2( + uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, + uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int w, int h) { + const __m256i v_maxval_b = _mm256_set1_epi8(AOM_BLEND_A64_MAX_ALPHA); + do { + int c; + for (c = 0; c < w; c += 32) { + const __m256i v_m0_b = yy_loadu_256(mask + c); + const __m256i v_m1_b = _mm256_sub_epi8(v_maxval_b, v_m0_b); + + const __m256i v_res_b = blend_32_u8_avx2( + src0 + c, src1 + c, &v_m0_b, &v_m1_b, AOM_BLEND_A64_ROUND_BITS); + + yy_storeu_256(dst + c, v_res_b); + } + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += mask_stride; + } while (--h); +} + +static INLINE void blend_a64_mask_avx2( + uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, + uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int w, int h) { + const __m128i v_maxval_b = _mm_set1_epi8(AOM_BLEND_A64_MAX_ALPHA); + const __m128i _r = _mm_set1_epi16(1 << (15 - AOM_BLEND_A64_ROUND_BITS)); + switch (w) { + case 4: + do { + const __m128i v_m0_b = xx_loadl_32(mask); + const __m128i v_m1_b = _mm_sub_epi8(v_maxval_b, v_m0_b); + const __m128i v_res_b = blend_4_u8(src0, src1, &v_m0_b, &v_m1_b, &_r); + + xx_storel_32(dst, v_res_b); + + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += mask_stride; + } while (--h); + break; + case 8: + do { + const __m128i v_m0_b = xx_loadl_64(mask); + const __m128i v_m1_b = _mm_sub_epi8(v_maxval_b, v_m0_b); + const __m128i v_res_b = blend_8_u8(src0, src1, &v_m0_b, &v_m1_b, &_r); + + xx_storel_64(dst, v_res_b); + + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += mask_stride; + } while (--h); + break; + case 16: + do { + const __m128i v_m0_b = xx_loadu_128(mask); + const __m128i v_m1_b = _mm_sub_epi8(v_maxval_b, v_m0_b); + const __m128i v_res_b = blend_16_u8(src0, src1, &v_m0_b, &v_m1_b, &_r); + + xx_storeu_128(dst, v_res_b); + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += mask_stride; + } while (--h); + break; + default: + blend_a64_mask_w32n_avx2(dst, dst_stride, src0, src0_stride, src1, + src1_stride, mask, mask_stride, w, h); + } +} + +void aom_blend_a64_mask_avx2(uint8_t *dst, uint32_t dst_stride, + const uint8_t *src0, uint32_t src0_stride, + const uint8_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int w, + int h, int subx, int suby) { + assert(IMPLIES(src0 == dst, src0_stride == dst_stride)); + assert(IMPLIES(src1 == dst, src1_stride == dst_stride)); + + assert(h >= 1); + assert(w >= 1); + assert(IS_POWER_OF_TWO(h)); + assert(IS_POWER_OF_TWO(w)); + + if (UNLIKELY((h | w) & 3)) { // if (w <= 2 || h <= 2) + aom_blend_a64_mask_c(dst, dst_stride, src0, src0_stride, src1, src1_stride, + mask, mask_stride, w, h, subx, suby); + } else { + if (subx & suby) { + blend_a64_mask_sx_sy_avx2(dst, dst_stride, src0, src0_stride, src1, + src1_stride, mask, mask_stride, w, h); + } else if (subx) { + blend_a64_mask_sx_avx2(dst, dst_stride, src0, src0_stride, src1, + src1_stride, mask, mask_stride, w, h); + } else if (suby) { + blend_a64_mask_sy_avx2(dst, dst_stride, src0, src0_stride, src1, + src1_stride, mask, mask_stride, w, h); + } else { + blend_a64_mask_avx2(dst, dst_stride, src0, src0_stride, src1, src1_stride, + mask, mask_stride, w, h); + } + } +} diff --git a/third_party/aom/aom_dsp/x86/blend_a64_mask_sse4.c b/third_party/aom/aom_dsp/x86/blend_a64_mask_sse4.c new file mode 100644 index 0000000000..9d6b4c2f74 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/blend_a64_mask_sse4.c @@ -0,0 +1,1109 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include // SSE4.1 + +#include + +#include "aom/aom_integer.h" +#include "aom_ports/mem.h" +#include "aom_dsp/aom_dsp_common.h" +#include "aom_dsp/blend.h" + +#include "aom_dsp/x86/synonyms.h" +#include "aom_dsp/x86/blend_sse4.h" +#include "aom_dsp/x86/blend_mask_sse4.h" + +#include "config/aom_dsp_rtcd.h" + +////////////////////////////////////////////////////////////////////////////// +// No sub-sampling +////////////////////////////////////////////////////////////////////////////// + +static void blend_a64_mask_w4_sse4_1(uint8_t *dst, uint32_t dst_stride, + const uint8_t *src0, uint32_t src0_stride, + const uint8_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, + int w, int h) { + (void)w; + const __m128i v_maxval_b = _mm_set1_epi8(AOM_BLEND_A64_MAX_ALPHA); + const __m128i _r = _mm_set1_epi16(1 << (15 - AOM_BLEND_A64_ROUND_BITS)); + do { + const __m128i v_m0_b = xx_loadl_32(mask); + const __m128i v_m1_b = _mm_sub_epi8(v_maxval_b, v_m0_b); + const __m128i v_res_b = blend_4_u8(src0, src1, &v_m0_b, &v_m1_b, &_r); + xx_storel_32(dst, v_res_b); + + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += mask_stride; + } while (--h); +} + +static void blend_a64_mask_w8_sse4_1(uint8_t *dst, uint32_t dst_stride, + const uint8_t *src0, uint32_t src0_stride, + const uint8_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, + int w, int h) { + (void)w; + const __m128i v_maxval_b = _mm_set1_epi8(AOM_BLEND_A64_MAX_ALPHA); + const __m128i _r = _mm_set1_epi16(1 << (15 - AOM_BLEND_A64_ROUND_BITS)); + do { + const __m128i v_m0_b = xx_loadl_64(mask); + const __m128i v_m1_b = _mm_sub_epi8(v_maxval_b, v_m0_b); + const __m128i v_res_b = blend_8_u8(src0, src1, &v_m0_b, &v_m1_b, &_r); + xx_storel_64(dst, v_res_b); + + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += mask_stride; + } while (--h); +} + +static void blend_a64_mask_w16n_sse4_1( + uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, + uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int w, int h) { + const __m128i v_maxval_b = _mm_set1_epi8(AOM_BLEND_A64_MAX_ALPHA); + const __m128i _r = _mm_set1_epi16(1 << (15 - AOM_BLEND_A64_ROUND_BITS)); + + do { + int c; + for (c = 0; c < w; c += 16) { + const __m128i v_m0_b = xx_loadu_128(mask + c); + const __m128i v_m1_b = _mm_sub_epi8(v_maxval_b, v_m0_b); + + const __m128i v_res_b = + blend_16_u8(src0 + c, src1 + c, &v_m0_b, &v_m1_b, &_r); + + xx_storeu_128(dst + c, v_res_b); + } + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += mask_stride; + } while (--h); +} + +////////////////////////////////////////////////////////////////////////////// +// Horizontal sub-sampling +////////////////////////////////////////////////////////////////////////////// + +static void blend_a64_mask_sx_w4_sse4_1( + uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, + uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int w, int h) { + (void)w; + + const __m128i v_shuffle_b = xx_loadu_128(g_blend_a64_mask_shuffle); + const __m128i v_maxval_b = _mm_set1_epi8(AOM_BLEND_A64_MAX_ALPHA); + const __m128i _r = _mm_set1_epi16(1 << (15 - AOM_BLEND_A64_ROUND_BITS)); + do { + const __m128i v_r_b = xx_loadl_64(mask); + const __m128i v_r0_s_b = _mm_shuffle_epi8(v_r_b, v_shuffle_b); + const __m128i v_r_lo_b = _mm_unpacklo_epi64(v_r0_s_b, v_r0_s_b); + const __m128i v_r_hi_b = _mm_unpackhi_epi64(v_r0_s_b, v_r0_s_b); + const __m128i v_m0_b = _mm_avg_epu8(v_r_lo_b, v_r_hi_b); + const __m128i v_m1_b = _mm_sub_epi8(v_maxval_b, v_m0_b); + + const __m128i v_res_b = blend_4_u8(src0, src1, &v_m0_b, &v_m1_b, &_r); + xx_storel_32(dst, v_res_b); + + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += mask_stride; + } while (--h); +} + +static void blend_a64_mask_sx_w8_sse4_1( + uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, + uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int w, int h) { + (void)w; + + const __m128i v_shuffle_b = xx_loadu_128(g_blend_a64_mask_shuffle); + const __m128i v_maxval_b = _mm_set1_epi8(AOM_BLEND_A64_MAX_ALPHA); + const __m128i _r = _mm_set1_epi16(1 << (15 - AOM_BLEND_A64_ROUND_BITS)); + do { + const __m128i v_r_b = xx_loadu_128(mask); + const __m128i v_r0_s_b = _mm_shuffle_epi8(v_r_b, v_shuffle_b); + const __m128i v_r_lo_b = _mm_unpacklo_epi64(v_r0_s_b, v_r0_s_b); + const __m128i v_r_hi_b = _mm_unpackhi_epi64(v_r0_s_b, v_r0_s_b); + const __m128i v_m0_b = _mm_avg_epu8(v_r_lo_b, v_r_hi_b); + const __m128i v_m1_b = _mm_sub_epi8(v_maxval_b, v_m0_b); + + const __m128i v_res_b = blend_8_u8(src0, src1, &v_m0_b, &v_m1_b, &_r); + + xx_storel_64(dst, v_res_b); + + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += mask_stride; + } while (--h); +} + +static void blend_a64_mask_sx_w16n_sse4_1( + uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, + uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int w, int h) { + const __m128i v_shuffle_b = xx_loadu_128(g_blend_a64_mask_shuffle); + const __m128i v_maxval_b = _mm_set1_epi8(AOM_BLEND_A64_MAX_ALPHA); + const __m128i _r = _mm_set1_epi16(1 << (15 - AOM_BLEND_A64_ROUND_BITS)); + + do { + int c; + for (c = 0; c < w; c += 16) { + const __m128i v_r0_b = xx_loadu_128(mask + 2 * c); + const __m128i v_r1_b = xx_loadu_128(mask + 2 * c + 16); + const __m128i v_r0_s_b = _mm_shuffle_epi8(v_r0_b, v_shuffle_b); + const __m128i v_r1_s_b = _mm_shuffle_epi8(v_r1_b, v_shuffle_b); + const __m128i v_r_lo_b = _mm_unpacklo_epi64(v_r0_s_b, v_r1_s_b); + const __m128i v_r_hi_b = _mm_unpackhi_epi64(v_r0_s_b, v_r1_s_b); + const __m128i v_m0_b = _mm_avg_epu8(v_r_lo_b, v_r_hi_b); + const __m128i v_m1_b = _mm_sub_epi8(v_maxval_b, v_m0_b); + + const __m128i v_res_b = + blend_16_u8(src0 + c, src1 + c, &v_m0_b, &v_m1_b, &_r); + + xx_storeu_128(dst + c, v_res_b); + } + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += mask_stride; + } while (--h); +} + +////////////////////////////////////////////////////////////////////////////// +// Vertical sub-sampling +////////////////////////////////////////////////////////////////////////////// + +static void blend_a64_mask_sy_w4_sse4_1( + uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, + uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int w, int h) { + (void)w; + + const __m128i v_maxval_b = _mm_set1_epi8(AOM_BLEND_A64_MAX_ALPHA); + const __m128i _r = _mm_set1_epi16(1 << (15 - AOM_BLEND_A64_ROUND_BITS)); + + do { + const __m128i v_ra_b = xx_loadl_32(mask); + const __m128i v_rb_b = xx_loadl_32(mask + mask_stride); + const __m128i v_m0_b = _mm_avg_epu8(v_ra_b, v_rb_b); + const __m128i v_m1_b = _mm_sub_epi8(v_maxval_b, v_m0_b); + + const __m128i v_res_b = blend_4_u8(src0, src1, &v_m0_b, &v_m1_b, &_r); + + xx_storel_32(dst, v_res_b); + + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += 2 * mask_stride; + } while (--h); +} + +static void blend_a64_mask_sy_w8_sse4_1( + uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, + uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int w, int h) { + (void)w; + + const __m128i v_maxval_b = _mm_set1_epi8(AOM_BLEND_A64_MAX_ALPHA); + const __m128i _r = _mm_set1_epi16(1 << (15 - AOM_BLEND_A64_ROUND_BITS)); + do { + const __m128i v_ra_b = xx_loadl_64(mask); + const __m128i v_rb_b = xx_loadl_64(mask + mask_stride); + const __m128i v_m0_b = _mm_avg_epu8(v_ra_b, v_rb_b); + const __m128i v_m1_b = _mm_sub_epi8(v_maxval_b, v_m0_b); + const __m128i v_res_b = blend_8_u8(src0, src1, &v_m0_b, &v_m1_b, &_r); + + xx_storel_64(dst, v_res_b); + + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += 2 * mask_stride; + } while (--h); +} + +static void blend_a64_mask_sy_w16n_sse4_1( + uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, + uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int w, int h) { + const __m128i v_maxval_b = _mm_set1_epi8(AOM_BLEND_A64_MAX_ALPHA); + const __m128i _r = _mm_set1_epi16(1 << (15 - AOM_BLEND_A64_ROUND_BITS)); + do { + int c; + for (c = 0; c < w; c += 16) { + const __m128i v_ra_b = xx_loadu_128(mask + c); + const __m128i v_rb_b = xx_loadu_128(mask + c + mask_stride); + const __m128i v_m0_b = _mm_avg_epu8(v_ra_b, v_rb_b); + const __m128i v_m1_b = _mm_sub_epi8(v_maxval_b, v_m0_b); + + const __m128i v_res_b = + blend_16_u8(src0 + c, src1 + c, &v_m0_b, &v_m1_b, &_r); + + xx_storeu_128(dst + c, v_res_b); + } + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += 2 * mask_stride; + } while (--h); +} + +////////////////////////////////////////////////////////////////////////////// +// Horizontal and Vertical sub-sampling +////////////////////////////////////////////////////////////////////////////// + +static void blend_a64_mask_sx_sy_w4_sse4_1( + uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, + uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int w, int h) { + const __m128i v_shuffle_b = xx_loadu_128(g_blend_a64_mask_shuffle); + const __m128i v_maxval_b = _mm_set1_epi8(AOM_BLEND_A64_MAX_ALPHA); + const __m128i _r = _mm_set1_epi16(1 << (15 - AOM_BLEND_A64_ROUND_BITS)); + (void)w; + + do { + const __m128i v_ra_b = xx_loadl_64(mask); + const __m128i v_rb_b = xx_loadl_64(mask + mask_stride); + const __m128i v_rvs_b = _mm_add_epi8(v_ra_b, v_rb_b); + const __m128i v_r_s_b = _mm_shuffle_epi8(v_rvs_b, v_shuffle_b); + const __m128i v_r0_s_w = _mm_cvtepu8_epi16(v_r_s_b); + const __m128i v_r1_s_w = _mm_cvtepu8_epi16(_mm_srli_si128(v_r_s_b, 8)); + const __m128i v_rs_w = _mm_add_epi16(v_r0_s_w, v_r1_s_w); + const __m128i v_m0_w = xx_roundn_epu16(v_rs_w, 2); + const __m128i v_m0_b = _mm_packus_epi16(v_m0_w, v_m0_w); + const __m128i v_m1_b = _mm_sub_epi8(v_maxval_b, v_m0_b); + + const __m128i v_res_b = blend_4_u8(src0, src1, &v_m0_b, &v_m1_b, &_r); + + xx_storel_32(dst, v_res_b); + + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += 2 * mask_stride; + } while (--h); +} + +static void blend_a64_mask_sx_sy_w8_sse4_1( + uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, + uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int w, int h) { + const __m128i v_shuffle_b = xx_loadu_128(g_blend_a64_mask_shuffle); + const __m128i v_maxval_b = _mm_set1_epi8(AOM_BLEND_A64_MAX_ALPHA); + const __m128i _r = _mm_set1_epi16(1 << (15 - AOM_BLEND_A64_ROUND_BITS)); + (void)w; + + do { + const __m128i v_ra_b = xx_loadu_128(mask); + const __m128i v_rb_b = xx_loadu_128(mask + mask_stride); + + const __m128i v_rvs_b = _mm_add_epi8(v_ra_b, v_rb_b); + const __m128i v_r_s_b = _mm_shuffle_epi8(v_rvs_b, v_shuffle_b); + const __m128i v_r0_s_w = _mm_cvtepu8_epi16(v_r_s_b); + const __m128i v_r1_s_w = _mm_cvtepu8_epi16(_mm_srli_si128(v_r_s_b, 8)); + const __m128i v_rs_w = _mm_add_epi16(v_r0_s_w, v_r1_s_w); + const __m128i v_m0_w = xx_roundn_epu16(v_rs_w, 2); + const __m128i v_m0_b = _mm_packus_epi16(v_m0_w, v_m0_w); + const __m128i v_m1_b = _mm_sub_epi8(v_maxval_b, v_m0_b); + + const __m128i v_res_b = blend_8_u8(src0, src1, &v_m0_b, &v_m1_b, &_r); + + xx_storel_64(dst, v_res_b); + + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += 2 * mask_stride; + } while (--h); +} + +static void blend_a64_mask_sx_sy_w16n_sse4_1( + uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, + uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int w, int h) { + const __m128i v_zmask_b = _mm_set_epi8(0, 0xff, 0, 0xff, 0, 0xff, 0, 0xff, 0, + 0xff, 0, 0xff, 0, 0xff, 0, 0xff); + const __m128i v_maxval_b = _mm_set1_epi8(AOM_BLEND_A64_MAX_ALPHA); + const __m128i _r = _mm_set1_epi16(1 << (15 - AOM_BLEND_A64_ROUND_BITS)); + do { + int c; + for (c = 0; c < w; c += 16) { + const __m128i v_ral_b = xx_loadu_128(mask + 2 * c); + const __m128i v_rah_b = xx_loadu_128(mask + 2 * c + 16); + const __m128i v_rbl_b = xx_loadu_128(mask + mask_stride + 2 * c); + const __m128i v_rbh_b = xx_loadu_128(mask + mask_stride + 2 * c + 16); + const __m128i v_rvsl_b = _mm_add_epi8(v_ral_b, v_rbl_b); + const __m128i v_rvsh_b = _mm_add_epi8(v_rah_b, v_rbh_b); + const __m128i v_rvsal_w = _mm_and_si128(v_rvsl_b, v_zmask_b); + const __m128i v_rvsah_w = _mm_and_si128(v_rvsh_b, v_zmask_b); + const __m128i v_rvsbl_w = + _mm_and_si128(_mm_srli_si128(v_rvsl_b, 1), v_zmask_b); + const __m128i v_rvsbh_w = + _mm_and_si128(_mm_srli_si128(v_rvsh_b, 1), v_zmask_b); + const __m128i v_rsl_w = _mm_add_epi16(v_rvsal_w, v_rvsbl_w); + const __m128i v_rsh_w = _mm_add_epi16(v_rvsah_w, v_rvsbh_w); + + const __m128i v_m0l_w = xx_roundn_epu16(v_rsl_w, 2); + const __m128i v_m0h_w = xx_roundn_epu16(v_rsh_w, 2); + const __m128i v_m0_b = _mm_packus_epi16(v_m0l_w, v_m0h_w); + const __m128i v_m1_b = _mm_sub_epi8(v_maxval_b, v_m0_b); + + const __m128i v_res_b = + blend_16_u8(src0 + c, src1 + c, &v_m0_b, &v_m1_b, &_r); + + xx_storeu_128(dst + c, v_res_b); + } + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += 2 * mask_stride; + } while (--h); +} + +////////////////////////////////////////////////////////////////////////////// +// Dispatch +////////////////////////////////////////////////////////////////////////////// + +void aom_blend_a64_mask_sse4_1(uint8_t *dst, uint32_t dst_stride, + const uint8_t *src0, uint32_t src0_stride, + const uint8_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int w, + int h, int subx, int suby) { + typedef void (*blend_fn)( + uint8_t * dst, uint32_t dst_stride, const uint8_t *src0, + uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int w, int h); + + // Dimensions are: width_index X subx X suby + static const blend_fn blend[3][2][2] = { + { // w % 16 == 0 + { blend_a64_mask_w16n_sse4_1, blend_a64_mask_sy_w16n_sse4_1 }, + { blend_a64_mask_sx_w16n_sse4_1, blend_a64_mask_sx_sy_w16n_sse4_1 } }, + { // w == 4 + { blend_a64_mask_w4_sse4_1, blend_a64_mask_sy_w4_sse4_1 }, + { blend_a64_mask_sx_w4_sse4_1, blend_a64_mask_sx_sy_w4_sse4_1 } }, + { // w == 8 + { blend_a64_mask_w8_sse4_1, blend_a64_mask_sy_w8_sse4_1 }, + { blend_a64_mask_sx_w8_sse4_1, blend_a64_mask_sx_sy_w8_sse4_1 } } + }; + + assert(IMPLIES(src0 == dst, src0_stride == dst_stride)); + assert(IMPLIES(src1 == dst, src1_stride == dst_stride)); + + assert(h >= 1); + assert(w >= 1); + assert(IS_POWER_OF_TWO(h)); + assert(IS_POWER_OF_TWO(w)); + + if (UNLIKELY((h | w) & 3)) { // if (w <= 2 || h <= 2) + aom_blend_a64_mask_c(dst, dst_stride, src0, src0_stride, src1, src1_stride, + mask, mask_stride, w, h, subx, suby); + } else { + blend[(w >> 2) & 3][subx != 0][suby != 0](dst, dst_stride, src0, + src0_stride, src1, src1_stride, + mask, mask_stride, w, h); + } +} + +////////////////////////////////////////////////////////////////////////////// +// No sub-sampling +////////////////////////////////////////////////////////////////////////////// + +static INLINE void blend_a64_mask_bn_w4_sse4_1( + uint16_t *dst, uint32_t dst_stride, const uint16_t *src0, + uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, blend_unit_fn blend) { + const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + + do { + const __m128i v_m0_b = xx_loadl_32(mask); + const __m128i v_m0_w = _mm_cvtepu8_epi16(v_m0_b); + const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w); + + const __m128i v_res_w = blend(src0, src1, v_m0_w, v_m1_w); + + xx_storel_64(dst, v_res_w); + + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += mask_stride; + } while (--h); +} + +static void blend_a64_mask_b10_w4_sse4_1( + uint16_t *dst, uint32_t dst_stride, const uint16_t *src0, + uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int w, int h) { + (void)w; + blend_a64_mask_bn_w4_sse4_1(dst, dst_stride, src0, src0_stride, src1, + src1_stride, mask, mask_stride, h, blend_4_b10); +} + +static void blend_a64_mask_b12_w4_sse4_1( + uint16_t *dst, uint32_t dst_stride, const uint16_t *src0, + uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int w, int h) { + (void)w; + blend_a64_mask_bn_w4_sse4_1(dst, dst_stride, src0, src0_stride, src1, + src1_stride, mask, mask_stride, h, blend_4_b12); +} + +static INLINE void blend_a64_mask_bn_w8n_sse4_1( + uint16_t *dst, uint32_t dst_stride, const uint16_t *src0, + uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int w, int h, + blend_unit_fn blend) { + const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + + do { + int c; + for (c = 0; c < w; c += 8) { + const __m128i v_m0_b = xx_loadl_64(mask + c); + const __m128i v_m0_w = _mm_cvtepu8_epi16(v_m0_b); + const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w); + + const __m128i v_res_w = blend(src0 + c, src1 + c, v_m0_w, v_m1_w); + + xx_storeu_128(dst + c, v_res_w); + } + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += mask_stride; + } while (--h); +} + +static void blend_a64_mask_b10_w8n_sse4_1( + uint16_t *dst, uint32_t dst_stride, const uint16_t *src0, + uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int w, int h) { + blend_a64_mask_bn_w8n_sse4_1(dst, dst_stride, src0, src0_stride, src1, + src1_stride, mask, mask_stride, w, h, + blend_8_b10); +} + +static void blend_a64_mask_b12_w8n_sse4_1( + uint16_t *dst, uint32_t dst_stride, const uint16_t *src0, + uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int w, int h) { + blend_a64_mask_bn_w8n_sse4_1(dst, dst_stride, src0, src0_stride, src1, + src1_stride, mask, mask_stride, w, h, + blend_8_b12); +} + +////////////////////////////////////////////////////////////////////////////// +// Horizontal sub-sampling +////////////////////////////////////////////////////////////////////////////// + +static INLINE void blend_a64_mask_bn_sx_w4_sse4_1( + uint16_t *dst, uint32_t dst_stride, const uint16_t *src0, + uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, blend_unit_fn blend) { + const __m128i v_zmask_b = _mm_set_epi8(0, 0xff, 0, 0xff, 0, 0xff, 0, 0xff, 0, + 0xff, 0, 0xff, 0, 0xff, 0, 0xff); + const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + + do { + const __m128i v_r_b = xx_loadl_64(mask); + const __m128i v_a_b = _mm_avg_epu8(v_r_b, _mm_srli_si128(v_r_b, 1)); + + const __m128i v_m0_w = _mm_and_si128(v_a_b, v_zmask_b); + const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w); + + const __m128i v_res_w = blend(src0, src1, v_m0_w, v_m1_w); + + xx_storel_64(dst, v_res_w); + + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += mask_stride; + } while (--h); +} + +static void blend_a64_mask_b10_sx_w4_sse4_1( + uint16_t *dst, uint32_t dst_stride, const uint16_t *src0, + uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int w, int h) { + (void)w; + blend_a64_mask_bn_sx_w4_sse4_1(dst, dst_stride, src0, src0_stride, src1, + src1_stride, mask, mask_stride, h, + blend_4_b10); +} + +static void blend_a64_mask_b12_sx_w4_sse4_1( + uint16_t *dst, uint32_t dst_stride, const uint16_t *src0, + uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int w, int h) { + (void)w; + blend_a64_mask_bn_sx_w4_sse4_1(dst, dst_stride, src0, src0_stride, src1, + src1_stride, mask, mask_stride, h, + blend_4_b12); +} + +static INLINE void blend_a64_mask_bn_sx_w8n_sse4_1( + uint16_t *dst, uint32_t dst_stride, const uint16_t *src0, + uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int w, int h, + blend_unit_fn blend) { + const __m128i v_zmask_b = _mm_set_epi8(0, 0xff, 0, 0xff, 0, 0xff, 0, 0xff, 0, + 0xff, 0, 0xff, 0, 0xff, 0, 0xff); + const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + + do { + int c; + for (c = 0; c < w; c += 8) { + const __m128i v_r_b = xx_loadu_128(mask + 2 * c); + const __m128i v_a_b = _mm_avg_epu8(v_r_b, _mm_srli_si128(v_r_b, 1)); + + const __m128i v_m0_w = _mm_and_si128(v_a_b, v_zmask_b); + const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w); + + const __m128i v_res_w = blend(src0 + c, src1 + c, v_m0_w, v_m1_w); + + xx_storeu_128(dst + c, v_res_w); + } + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += mask_stride; + } while (--h); +} + +static void blend_a64_mask_b10_sx_w8n_sse4_1( + uint16_t *dst, uint32_t dst_stride, const uint16_t *src0, + uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int w, int h) { + blend_a64_mask_bn_sx_w8n_sse4_1(dst, dst_stride, src0, src0_stride, src1, + src1_stride, mask, mask_stride, w, h, + blend_8_b10); +} + +static void blend_a64_mask_b12_sx_w8n_sse4_1( + uint16_t *dst, uint32_t dst_stride, const uint16_t *src0, + uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int w, int h) { + blend_a64_mask_bn_sx_w8n_sse4_1(dst, dst_stride, src0, src0_stride, src1, + src1_stride, mask, mask_stride, w, h, + blend_8_b12); +} + +////////////////////////////////////////////////////////////////////////////// +// Vertical sub-sampling +////////////////////////////////////////////////////////////////////////////// + +static INLINE void blend_a64_mask_bn_sy_w4_sse4_1( + uint16_t *dst, uint32_t dst_stride, const uint16_t *src0, + uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, blend_unit_fn blend) { + const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + + do { + const __m128i v_ra_b = xx_loadl_32(mask); + const __m128i v_rb_b = xx_loadl_32(mask + mask_stride); + const __m128i v_a_b = _mm_avg_epu8(v_ra_b, v_rb_b); + + const __m128i v_m0_w = _mm_cvtepu8_epi16(v_a_b); + const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w); + + const __m128i v_res_w = blend(src0, src1, v_m0_w, v_m1_w); + + xx_storel_64(dst, v_res_w); + + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += 2 * mask_stride; + } while (--h); +} + +static void blend_a64_mask_b10_sy_w4_sse4_1( + uint16_t *dst, uint32_t dst_stride, const uint16_t *src0, + uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int w, int h) { + (void)w; + blend_a64_mask_bn_sy_w4_sse4_1(dst, dst_stride, src0, src0_stride, src1, + src1_stride, mask, mask_stride, h, + blend_4_b10); +} + +static void blend_a64_mask_b12_sy_w4_sse4_1( + uint16_t *dst, uint32_t dst_stride, const uint16_t *src0, + uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int w, int h) { + (void)w; + blend_a64_mask_bn_sy_w4_sse4_1(dst, dst_stride, src0, src0_stride, src1, + src1_stride, mask, mask_stride, h, + blend_4_b12); +} + +static INLINE void blend_a64_mask_bn_sy_w8n_sse4_1( + uint16_t *dst, uint32_t dst_stride, const uint16_t *src0, + uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int w, int h, + blend_unit_fn blend) { + const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + + do { + int c; + for (c = 0; c < w; c += 8) { + const __m128i v_ra_b = xx_loadl_64(mask + c); + const __m128i v_rb_b = xx_loadl_64(mask + c + mask_stride); + const __m128i v_a_b = _mm_avg_epu8(v_ra_b, v_rb_b); + + const __m128i v_m0_w = _mm_cvtepu8_epi16(v_a_b); + const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w); + + const __m128i v_res_w = blend(src0 + c, src1 + c, v_m0_w, v_m1_w); + + xx_storeu_128(dst + c, v_res_w); + } + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += 2 * mask_stride; + } while (--h); +} + +static void blend_a64_mask_b10_sy_w8n_sse4_1( + uint16_t *dst, uint32_t dst_stride, const uint16_t *src0, + uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int w, int h) { + blend_a64_mask_bn_sy_w8n_sse4_1(dst, dst_stride, src0, src0_stride, src1, + src1_stride, mask, mask_stride, w, h, + blend_8_b10); +} + +static void blend_a64_mask_b12_sy_w8n_sse4_1( + uint16_t *dst, uint32_t dst_stride, const uint16_t *src0, + uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int w, int h) { + blend_a64_mask_bn_sy_w8n_sse4_1(dst, dst_stride, src0, src0_stride, src1, + src1_stride, mask, mask_stride, w, h, + blend_8_b12); +} + +////////////////////////////////////////////////////////////////////////////// +// Horizontal and Vertical sub-sampling +////////////////////////////////////////////////////////////////////////////// + +static INLINE void blend_a64_mask_bn_sx_sy_w4_sse4_1( + uint16_t *dst, uint32_t dst_stride, const uint16_t *src0, + uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, blend_unit_fn blend) { + const __m128i v_zmask_b = _mm_set_epi8(0, 0xff, 0, 0xff, 0, 0xff, 0, 0xff, 0, + 0xff, 0, 0xff, 0, 0xff, 0, 0xff); + const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + + do { + const __m128i v_ra_b = xx_loadl_64(mask); + const __m128i v_rb_b = xx_loadl_64(mask + mask_stride); + const __m128i v_rvs_b = _mm_add_epi8(v_ra_b, v_rb_b); + const __m128i v_rvsa_w = _mm_and_si128(v_rvs_b, v_zmask_b); + const __m128i v_rvsb_w = + _mm_and_si128(_mm_srli_si128(v_rvs_b, 1), v_zmask_b); + const __m128i v_rs_w = _mm_add_epi16(v_rvsa_w, v_rvsb_w); + + const __m128i v_m0_w = xx_roundn_epu16(v_rs_w, 2); + const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w); + + const __m128i v_res_w = blend(src0, src1, v_m0_w, v_m1_w); + + xx_storel_64(dst, v_res_w); + + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += 2 * mask_stride; + } while (--h); +} + +static void blend_a64_mask_b10_sx_sy_w4_sse4_1( + uint16_t *dst, uint32_t dst_stride, const uint16_t *src0, + uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int w, int h) { + (void)w; + blend_a64_mask_bn_sx_sy_w4_sse4_1(dst, dst_stride, src0, src0_stride, src1, + src1_stride, mask, mask_stride, h, + blend_4_b10); +} + +static void blend_a64_mask_b12_sx_sy_w4_sse4_1( + uint16_t *dst, uint32_t dst_stride, const uint16_t *src0, + uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int w, int h) { + (void)w; + blend_a64_mask_bn_sx_sy_w4_sse4_1(dst, dst_stride, src0, src0_stride, src1, + src1_stride, mask, mask_stride, h, + blend_4_b12); +} + +static INLINE void blend_a64_mask_bn_sx_sy_w8n_sse4_1( + uint16_t *dst, uint32_t dst_stride, const uint16_t *src0, + uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int w, int h, + blend_unit_fn blend) { + const __m128i v_zmask_b = _mm_set_epi8(0, 0xff, 0, 0xff, 0, 0xff, 0, 0xff, 0, + 0xff, 0, 0xff, 0, 0xff, 0, 0xff); + const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + + do { + int c; + for (c = 0; c < w; c += 8) { + const __m128i v_ra_b = xx_loadu_128(mask + 2 * c); + const __m128i v_rb_b = xx_loadu_128(mask + 2 * c + mask_stride); + const __m128i v_rvs_b = _mm_add_epi8(v_ra_b, v_rb_b); + const __m128i v_rvsa_w = _mm_and_si128(v_rvs_b, v_zmask_b); + const __m128i v_rvsb_w = + _mm_and_si128(_mm_srli_si128(v_rvs_b, 1), v_zmask_b); + const __m128i v_rs_w = _mm_add_epi16(v_rvsa_w, v_rvsb_w); + + const __m128i v_m0_w = xx_roundn_epu16(v_rs_w, 2); + const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w); + + const __m128i v_res_w = blend(src0 + c, src1 + c, v_m0_w, v_m1_w); + + xx_storeu_128(dst + c, v_res_w); + } + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += 2 * mask_stride; + } while (--h); +} + +static void blend_a64_mask_b10_sx_sy_w8n_sse4_1( + uint16_t *dst, uint32_t dst_stride, const uint16_t *src0, + uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int w, int h) { + blend_a64_mask_bn_sx_sy_w8n_sse4_1(dst, dst_stride, src0, src0_stride, src1, + src1_stride, mask, mask_stride, w, h, + blend_8_b10); +} + +static void blend_a64_mask_b12_sx_sy_w8n_sse4_1( + uint16_t *dst, uint32_t dst_stride, const uint16_t *src0, + uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int w, int h) { + blend_a64_mask_bn_sx_sy_w8n_sse4_1(dst, dst_stride, src0, src0_stride, src1, + src1_stride, mask, mask_stride, w, h, + blend_8_b12); +} + +////////////////////////////////////////////////////////////////////////////// +// Dispatch +////////////////////////////////////////////////////////////////////////////// + +void aom_highbd_blend_a64_mask_sse4_1(uint8_t *dst_8, uint32_t dst_stride, + const uint8_t *src0_8, + uint32_t src0_stride, + const uint8_t *src1_8, + uint32_t src1_stride, const uint8_t *mask, + uint32_t mask_stride, int w, int h, + int subx, int suby, int bd) { + typedef void (*blend_fn)( + uint16_t * dst, uint32_t dst_stride, const uint16_t *src0, + uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int w, int h); + + // Dimensions are: bd_index X width_index X subx X suby + static const blend_fn blend[2][2][2][2] = { + { // bd == 8 or 10 + { // w % 8 == 0 + { blend_a64_mask_b10_w8n_sse4_1, blend_a64_mask_b10_sy_w8n_sse4_1 }, + { blend_a64_mask_b10_sx_w8n_sse4_1, + blend_a64_mask_b10_sx_sy_w8n_sse4_1 } }, + { // w == 4 + { blend_a64_mask_b10_w4_sse4_1, blend_a64_mask_b10_sy_w4_sse4_1 }, + { blend_a64_mask_b10_sx_w4_sse4_1, + blend_a64_mask_b10_sx_sy_w4_sse4_1 } } }, + { // bd == 12 + { // w % 8 == 0 + { blend_a64_mask_b12_w8n_sse4_1, blend_a64_mask_b12_sy_w8n_sse4_1 }, + { blend_a64_mask_b12_sx_w8n_sse4_1, + blend_a64_mask_b12_sx_sy_w8n_sse4_1 } }, + { // w == 4 + { blend_a64_mask_b12_w4_sse4_1, blend_a64_mask_b12_sy_w4_sse4_1 }, + { blend_a64_mask_b12_sx_w4_sse4_1, + blend_a64_mask_b12_sx_sy_w4_sse4_1 } } } + }; + + assert(IMPLIES(src0_8 == dst_8, src0_stride == dst_stride)); + assert(IMPLIES(src1_8 == dst_8, src1_stride == dst_stride)); + + assert(h >= 1); + assert(w >= 1); + assert(IS_POWER_OF_TWO(h)); + assert(IS_POWER_OF_TWO(w)); + + assert(bd == 8 || bd == 10 || bd == 12); + if (UNLIKELY((h | w) & 3)) { // if (w <= 2 || h <= 2) + aom_highbd_blend_a64_mask_c(dst_8, dst_stride, src0_8, src0_stride, src1_8, + src1_stride, mask, mask_stride, w, h, subx, + suby, bd); + } else { + uint16_t *const dst = CONVERT_TO_SHORTPTR(dst_8); + const uint16_t *const src0 = CONVERT_TO_SHORTPTR(src0_8); + const uint16_t *const src1 = CONVERT_TO_SHORTPTR(src1_8); + + blend[bd == 12][(w >> 2) & 1][subx != 0][suby != 0]( + dst, dst_stride, src0, src0_stride, src1, src1_stride, mask, + mask_stride, w, h); + } +} + +static INLINE void blend_a64_d16_mask_w16_sse41( + uint8_t *dst, const CONV_BUF_TYPE *src0, const CONV_BUF_TYPE *src1, + const __m128i *m0, const __m128i *m1, const __m128i *v_round_offset, + const __m128i *v_maxval, int shift) { + const __m128i max_minus_m0 = _mm_sub_epi16(*v_maxval, *m0); + const __m128i max_minus_m1 = _mm_sub_epi16(*v_maxval, *m1); + const __m128i s0_0 = xx_loadu_128(src0); + const __m128i s0_1 = xx_loadu_128(src0 + 8); + const __m128i s1_0 = xx_loadu_128(src1); + const __m128i s1_1 = xx_loadu_128(src1 + 8); + __m128i res0_lo = _mm_madd_epi16(_mm_unpacklo_epi16(s0_0, s1_0), + _mm_unpacklo_epi16(*m0, max_minus_m0)); + __m128i res0_hi = _mm_madd_epi16(_mm_unpackhi_epi16(s0_0, s1_0), + _mm_unpackhi_epi16(*m0, max_minus_m0)); + __m128i res1_lo = _mm_madd_epi16(_mm_unpacklo_epi16(s0_1, s1_1), + _mm_unpacklo_epi16(*m1, max_minus_m1)); + __m128i res1_hi = _mm_madd_epi16(_mm_unpackhi_epi16(s0_1, s1_1), + _mm_unpackhi_epi16(*m1, max_minus_m1)); + res0_lo = _mm_srai_epi32(_mm_sub_epi32(res0_lo, *v_round_offset), shift); + res0_hi = _mm_srai_epi32(_mm_sub_epi32(res0_hi, *v_round_offset), shift); + res1_lo = _mm_srai_epi32(_mm_sub_epi32(res1_lo, *v_round_offset), shift); + res1_hi = _mm_srai_epi32(_mm_sub_epi32(res1_hi, *v_round_offset), shift); + const __m128i res0 = _mm_packs_epi32(res0_lo, res0_hi); + const __m128i res1 = _mm_packs_epi32(res1_lo, res1_hi); + const __m128i res = _mm_packus_epi16(res0, res1); + + _mm_storeu_si128((__m128i *)(dst), res); +} + +static INLINE void lowbd_blend_a64_d16_mask_subw0_subh0_w16_sse4_1( + uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, + uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, int w, + const __m128i *round_offset, int shift) { + const __m128i v_maxval = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + for (int i = 0; i < h; ++i) { + for (int j = 0; j < w; j += 16) { + const __m128i m = xx_loadu_128(mask + j); + const __m128i m0 = _mm_cvtepu8_epi16(m); + const __m128i m1 = _mm_cvtepu8_epi16(_mm_srli_si128(m, 8)); + + blend_a64_d16_mask_w16_sse41(dst + j, src0 + j, src1 + j, &m0, &m1, + round_offset, &v_maxval, shift); + } + mask += mask_stride; + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + } +} + +static INLINE void lowbd_blend_a64_d16_mask_subw1_subh1_w16_sse4_1( + uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, + uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, int w, + const __m128i *round_offset, int shift) { + const __m128i v_maxval = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + const __m128i one_b = _mm_set1_epi8(1); + const __m128i two_w = _mm_set1_epi16(2); + for (int i = 0; i < h; ++i) { + for (int j = 0; j < w; j += 16) { + const __m128i m_i00 = xx_loadu_128(mask + 2 * j); + const __m128i m_i01 = xx_loadu_128(mask + 2 * j + 16); + const __m128i m_i10 = xx_loadu_128(mask + mask_stride + 2 * j); + const __m128i m_i11 = xx_loadu_128(mask + mask_stride + 2 * j + 16); + + const __m128i m0_ac = _mm_adds_epu8(m_i00, m_i10); + const __m128i m1_ac = _mm_adds_epu8(m_i01, m_i11); + const __m128i m0_acbd = _mm_maddubs_epi16(m0_ac, one_b); + const __m128i m1_acbd = _mm_maddubs_epi16(m1_ac, one_b); + const __m128i m0 = _mm_srli_epi16(_mm_add_epi16(m0_acbd, two_w), 2); + const __m128i m1 = _mm_srli_epi16(_mm_add_epi16(m1_acbd, two_w), 2); + + blend_a64_d16_mask_w16_sse41(dst + j, src0 + j, src1 + j, &m0, &m1, + round_offset, &v_maxval, shift); + } + mask += mask_stride << 1; + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + } +} + +static INLINE void lowbd_blend_a64_d16_mask_subw1_subh0_w16_sse4_1( + uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, + uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, int w, + const __m128i *round_offset, int shift) { + const __m128i v_maxval = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + const __m128i one_b = _mm_set1_epi8(1); + const __m128i zeros = _mm_setzero_si128(); + for (int i = 0; i < h; ++i) { + for (int j = 0; j < w; j += 16) { + const __m128i m_i00 = xx_loadu_128(mask + 2 * j); + const __m128i m_i01 = xx_loadu_128(mask + 2 * j + 16); + const __m128i m0_ac = _mm_maddubs_epi16(m_i00, one_b); + const __m128i m1_ac = _mm_maddubs_epi16(m_i01, one_b); + const __m128i m0 = _mm_avg_epu16(m0_ac, zeros); + const __m128i m1 = _mm_avg_epu16(m1_ac, zeros); + + blend_a64_d16_mask_w16_sse41(dst + j, src0 + j, src1 + j, &m0, &m1, + round_offset, &v_maxval, shift); + } + mask += mask_stride; + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + } +} + +static INLINE void lowbd_blend_a64_d16_mask_subw0_subh1_w16_sse4_1( + uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, + uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, int w, + const __m128i *round_offset, int shift) { + const __m128i v_maxval = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + const __m128i zeros = _mm_setzero_si128(); + for (int i = 0; i < h; ++i) { + for (int j = 0; j < w; j += 16) { + const __m128i m_i00 = xx_loadu_128(mask + j); + const __m128i m_i10 = xx_loadu_128(mask + mask_stride + j); + + const __m128i m_ac = _mm_avg_epu8(_mm_adds_epu8(m_i00, m_i10), zeros); + const __m128i m0 = _mm_cvtepu8_epi16(m_ac); + const __m128i m1 = _mm_cvtepu8_epi16(_mm_srli_si128(m_ac, 8)); + + blend_a64_d16_mask_w16_sse41(dst + j, src0 + j, src1 + j, &m0, &m1, + round_offset, &v_maxval, shift); + } + mask += mask_stride << 1; + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + } +} + +void aom_lowbd_blend_a64_d16_mask_sse4_1( + uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, + uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, + ConvolveParams *conv_params) { + const int bd = 8; + const int round_bits = + 2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1; + + const int round_offset = + ((1 << (round_bits + bd)) + (1 << (round_bits + bd - 1)) - + (1 << (round_bits - 1))) + << AOM_BLEND_A64_ROUND_BITS; + + const int shift = round_bits + AOM_BLEND_A64_ROUND_BITS; + assert(IMPLIES((void *)src0 == dst, src0_stride == dst_stride)); + assert(IMPLIES((void *)src1 == dst, src1_stride == dst_stride)); + + assert(h >= 4); + assert(w >= 4); + assert(IS_POWER_OF_TWO(h)); + assert(IS_POWER_OF_TWO(w)); + + const __m128i v_round_offset = _mm_set1_epi32(round_offset); + + if (subw == 0 && subh == 0) { + switch (w) { + case 4: + aom_lowbd_blend_a64_d16_mask_subw0_subh0_w4_sse4_1( + dst, dst_stride, src0, src0_stride, src1, src1_stride, mask, + mask_stride, h, &v_round_offset, shift); + break; + case 8: + aom_lowbd_blend_a64_d16_mask_subw0_subh0_w8_sse4_1( + dst, dst_stride, src0, src0_stride, src1, src1_stride, mask, + mask_stride, h, &v_round_offset, shift); + break; + default: + lowbd_blend_a64_d16_mask_subw0_subh0_w16_sse4_1( + dst, dst_stride, src0, src0_stride, src1, src1_stride, mask, + mask_stride, h, w, &v_round_offset, shift); + break; + } + + } else if (subw == 1 && subh == 1) { + switch (w) { + case 4: + aom_lowbd_blend_a64_d16_mask_subw1_subh1_w4_sse4_1( + dst, dst_stride, src0, src0_stride, src1, src1_stride, mask, + mask_stride, h, &v_round_offset, shift); + break; + case 8: + aom_lowbd_blend_a64_d16_mask_subw1_subh1_w8_sse4_1( + dst, dst_stride, src0, src0_stride, src1, src1_stride, mask, + mask_stride, h, &v_round_offset, shift); + break; + default: + lowbd_blend_a64_d16_mask_subw1_subh1_w16_sse4_1( + dst, dst_stride, src0, src0_stride, src1, src1_stride, mask, + mask_stride, h, w, &v_round_offset, shift); + break; + } + } else if (subw == 1 && subh == 0) { + switch (w) { + case 4: + aom_lowbd_blend_a64_d16_mask_subw1_subh0_w4_sse4_1( + dst, dst_stride, src0, src0_stride, src1, src1_stride, mask, + mask_stride, h, &v_round_offset, shift); + break; + case 8: + aom_lowbd_blend_a64_d16_mask_subw1_subh0_w8_sse4_1( + dst, dst_stride, src0, src0_stride, src1, src1_stride, mask, + mask_stride, h, &v_round_offset, shift); + break; + default: + lowbd_blend_a64_d16_mask_subw1_subh0_w16_sse4_1( + dst, dst_stride, src0, src0_stride, src1, src1_stride, mask, + mask_stride, h, w, &v_round_offset, shift); + break; + } + } else { + switch (w) { + case 4: + aom_lowbd_blend_a64_d16_mask_subw0_subh1_w4_sse4_1( + dst, dst_stride, src0, src0_stride, src1, src1_stride, mask, + mask_stride, h, &v_round_offset, shift); + break; + case 8: + aom_lowbd_blend_a64_d16_mask_subw0_subh1_w8_sse4_1( + dst, dst_stride, src0, src0_stride, src1, src1_stride, mask, + mask_stride, h, &v_round_offset, shift); + break; + default: + lowbd_blend_a64_d16_mask_subw0_subh1_w16_sse4_1( + dst, dst_stride, src0, src0_stride, src1, src1_stride, mask, + mask_stride, h, w, &v_round_offset, shift); + break; + } + } +} diff --git a/third_party/aom/aom_dsp/x86/blend_a64_vmask_sse4.c b/third_party/aom/aom_dsp/x86/blend_a64_vmask_sse4.c new file mode 100644 index 0000000000..0649102325 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/blend_a64_vmask_sse4.c @@ -0,0 +1,283 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include // SSE4.1 + +#include + +#include "aom/aom_integer.h" +#include "aom_ports/mem.h" +#include "aom_dsp/aom_dsp_common.h" +#include "aom_dsp/blend.h" + +#include "aom_dsp/x86/synonyms.h" +#include "aom_dsp/x86/blend_sse4.h" + +#include "config/aom_dsp_rtcd.h" + +////////////////////////////////////////////////////////////////////////////// +// Implementation - No sub-sampling +////////////////////////////////////////////////////////////////////////////// + +static void blend_a64_vmask_w4_sse4_1(uint8_t *dst, uint32_t dst_stride, + const uint8_t *src0, uint32_t src0_stride, + const uint8_t *src1, uint32_t src1_stride, + const uint8_t *mask, int w, int h) { + const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + + (void)w; + + do { + const __m128i v_m0_w = _mm_set1_epi16(*mask); + const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w); + + const __m128i v_res_w = blend_4(src0, src1, &v_m0_w, &v_m1_w); + + const __m128i v_res_b = _mm_packus_epi16(v_res_w, v_res_w); + + xx_storel_32(dst, v_res_b); + + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += 1; + } while (--h); +} + +static void blend_a64_vmask_w8_sse4_1(uint8_t *dst, uint32_t dst_stride, + const uint8_t *src0, uint32_t src0_stride, + const uint8_t *src1, uint32_t src1_stride, + const uint8_t *mask, int w, int h) { + const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + + (void)w; + + do { + const __m128i v_m0_w = _mm_set1_epi16(*mask); + const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w); + + const __m128i v_res_w = blend_8(src0, src1, &v_m0_w, &v_m1_w); + + const __m128i v_res_b = _mm_packus_epi16(v_res_w, v_res_w); + + xx_storel_64(dst, v_res_b); + + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += 1; + } while (--h); +} + +static void blend_a64_vmask_w16n_sse4_1(uint8_t *dst, uint32_t dst_stride, + const uint8_t *src0, + uint32_t src0_stride, + const uint8_t *src1, + uint32_t src1_stride, + const uint8_t *mask, int w, int h) { + const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + + do { + int c; + const __m128i v_m0_w = _mm_set1_epi16(*mask); + const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w); + for (c = 0; c < w; c += 16) { + const __m128i v_resl_w = blend_8(src0 + c, src1 + c, &v_m0_w, &v_m1_w); + const __m128i v_resh_w = + blend_8(src0 + c + 8, src1 + c + 8, &v_m0_w, &v_m1_w); + + const __m128i v_res_b = _mm_packus_epi16(v_resl_w, v_resh_w); + + xx_storeu_128(dst + c, v_res_b); + } + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += 1; + } while (--h); +} + +////////////////////////////////////////////////////////////////////////////// +// Dispatch +////////////////////////////////////////////////////////////////////////////// + +void aom_blend_a64_vmask_sse4_1(uint8_t *dst, uint32_t dst_stride, + const uint8_t *src0, uint32_t src0_stride, + const uint8_t *src1, uint32_t src1_stride, + const uint8_t *mask, int w, int h) { + typedef void (*blend_fn)(uint8_t * dst, uint32_t dst_stride, + const uint8_t *src0, uint32_t src0_stride, + const uint8_t *src1, uint32_t src1_stride, + const uint8_t *mask, int w, int h); + + // Dimension: width_index + static const blend_fn blend[9] = { + blend_a64_vmask_w16n_sse4_1, // w % 16 == 0 + aom_blend_a64_vmask_c, // w == 1 + aom_blend_a64_vmask_c, // w == 2 + NULL, // INVALID + blend_a64_vmask_w4_sse4_1, // w == 4 + NULL, // INVALID + NULL, // INVALID + NULL, // INVALID + blend_a64_vmask_w8_sse4_1, // w == 8 + }; + + assert(IMPLIES(src0 == dst, src0_stride == dst_stride)); + assert(IMPLIES(src1 == dst, src1_stride == dst_stride)); + + assert(h >= 1); + assert(w >= 1); + assert(IS_POWER_OF_TWO(h)); + assert(IS_POWER_OF_TWO(w)); + + blend[w & 0xf](dst, dst_stride, src0, src0_stride, src1, src1_stride, mask, w, + h); +} + +////////////////////////////////////////////////////////////////////////////// +// Implementation - No sub-sampling +////////////////////////////////////////////////////////////////////////////// + +static INLINE void blend_a64_vmask_bn_w4_sse4_1( + uint16_t *dst, uint32_t dst_stride, const uint16_t *src0, + uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride, + const uint8_t *mask, int h, blend_unit_fn blend) { + const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + + do { + const __m128i v_m0_w = _mm_set1_epi16(*mask); + const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w); + + const __m128i v_res_w = blend(src0, src1, v_m0_w, v_m1_w); + + xx_storel_64(dst, v_res_w); + + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += 1; + } while (--h); +} + +static void blend_a64_vmask_b10_w4_sse4_1(uint16_t *dst, uint32_t dst_stride, + const uint16_t *src0, + uint32_t src0_stride, + const uint16_t *src1, + uint32_t src1_stride, + const uint8_t *mask, int w, int h) { + (void)w; + blend_a64_vmask_bn_w4_sse4_1(dst, dst_stride, src0, src0_stride, src1, + src1_stride, mask, h, blend_4_b10); +} + +static void blend_a64_vmask_b12_w4_sse4_1(uint16_t *dst, uint32_t dst_stride, + const uint16_t *src0, + uint32_t src0_stride, + const uint16_t *src1, + uint32_t src1_stride, + const uint8_t *mask, int w, int h) { + (void)w; + blend_a64_vmask_bn_w4_sse4_1(dst, dst_stride, src0, src0_stride, src1, + src1_stride, mask, h, blend_4_b12); +} + +static INLINE void blend_a64_vmask_bn_w8n_sse4_1( + uint16_t *dst, uint32_t dst_stride, const uint16_t *src0, + uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride, + const uint8_t *mask, int w, int h, blend_unit_fn blend) { + const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + + do { + int c; + const __m128i v_m0_w = _mm_set1_epi16(*mask); + const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w); + for (c = 0; c < w; c += 8) { + const __m128i v_res_w = blend(src0 + c, src1 + c, v_m0_w, v_m1_w); + + xx_storeu_128(dst + c, v_res_w); + } + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += 1; + } while (--h); +} + +static void blend_a64_vmask_b10_w8n_sse4_1(uint16_t *dst, uint32_t dst_stride, + const uint16_t *src0, + uint32_t src0_stride, + const uint16_t *src1, + uint32_t src1_stride, + const uint8_t *mask, int w, int h) { + blend_a64_vmask_bn_w8n_sse4_1(dst, dst_stride, src0, src0_stride, src1, + src1_stride, mask, w, h, blend_8_b10); +} + +static void blend_a64_vmask_b12_w8n_sse4_1(uint16_t *dst, uint32_t dst_stride, + const uint16_t *src0, + uint32_t src0_stride, + const uint16_t *src1, + uint32_t src1_stride, + const uint8_t *mask, int w, int h) { + blend_a64_vmask_bn_w8n_sse4_1(dst, dst_stride, src0, src0_stride, src1, + src1_stride, mask, w, h, blend_8_b12); +} + +////////////////////////////////////////////////////////////////////////////// +// Dispatch +////////////////////////////////////////////////////////////////////////////// + +void aom_highbd_blend_a64_vmask_sse4_1( + uint8_t *dst_8, uint32_t dst_stride, const uint8_t *src0_8, + uint32_t src0_stride, const uint8_t *src1_8, uint32_t src1_stride, + const uint8_t *mask, int w, int h, int bd) { + typedef void (*blend_fn)(uint16_t * dst, uint32_t dst_stride, + const uint16_t *src0, uint32_t src0_stride, + const uint16_t *src1, uint32_t src1_stride, + const uint8_t *mask, int w, int h); + + // Dimensions are: bd_index X width_index + static const blend_fn blend[2][2] = { + { + // bd == 8 or 10 + blend_a64_vmask_b10_w8n_sse4_1, // w % 8 == 0 + blend_a64_vmask_b10_w4_sse4_1, // w == 4 + }, + { + // bd == 12 + blend_a64_vmask_b12_w8n_sse4_1, // w % 8 == 0 + blend_a64_vmask_b12_w4_sse4_1, // w == 4 + } + }; + + assert(IMPLIES(src0_8 == dst_8, src0_stride == dst_stride)); + assert(IMPLIES(src1_8 == dst_8, src1_stride == dst_stride)); + + assert(h >= 1); + assert(w >= 1); + assert(IS_POWER_OF_TWO(h)); + assert(IS_POWER_OF_TWO(w)); + + assert(bd == 8 || bd == 10 || bd == 12); + + if (UNLIKELY((h | w) & 3)) { // if (w <= 2 || h <= 2) + aom_highbd_blend_a64_vmask_c(dst_8, dst_stride, src0_8, src0_stride, src1_8, + src1_stride, mask, w, h, bd); + } else { + uint16_t *const dst = CONVERT_TO_SHORTPTR(dst_8); + const uint16_t *const src0 = CONVERT_TO_SHORTPTR(src0_8); + const uint16_t *const src1 = CONVERT_TO_SHORTPTR(src1_8); + + blend[bd == 12][(w >> 2) & 1](dst, dst_stride, src0, src0_stride, src1, + src1_stride, mask, w, h); + } +} diff --git a/third_party/aom/aom_dsp/x86/blend_mask_sse4.h b/third_party/aom/aom_dsp/x86/blend_mask_sse4.h new file mode 100644 index 0000000000..c071fdcfc4 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/blend_mask_sse4.h @@ -0,0 +1,237 @@ +/* + * Copyright (c) 2018, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_X86_BLEND_MASK_SSE4_H_ +#define AOM_AOM_DSP_X86_BLEND_MASK_SSE4_H_ +#include // SSE4.1 + +#include + +#include "aom/aom_integer.h" +#include "aom_ports/mem.h" +#include "aom_dsp/aom_dsp_common.h" +#include "aom_dsp/blend.h" + +#include "aom_dsp/x86/synonyms.h" + +#include "config/aom_dsp_rtcd.h" + +static INLINE void blend_a64_d16_mask_w4_sse41( + uint8_t *dst, const CONV_BUF_TYPE *src0, const CONV_BUF_TYPE *src1, + const __m128i *m, const __m128i *v_round_offset, const __m128i *v_maxval, + int shift) { + const __m128i max_minus_m = _mm_sub_epi16(*v_maxval, *m); + const __m128i s0 = xx_loadl_64(src0); + const __m128i s1 = xx_loadl_64(src1); + const __m128i s0_s1 = _mm_unpacklo_epi16(s0, s1); + const __m128i m_max_minus_m = _mm_unpacklo_epi16(*m, max_minus_m); + const __m128i res_a = _mm_madd_epi16(s0_s1, m_max_minus_m); + const __m128i res_c = _mm_sub_epi32(res_a, *v_round_offset); + const __m128i res_d = _mm_srai_epi32(res_c, shift); + const __m128i res_e = _mm_packs_epi32(res_d, res_d); + const __m128i res = _mm_packus_epi16(res_e, res_e); + + xx_storel_32(dst, res); +} + +static INLINE void blend_a64_d16_mask_w8_sse41( + uint8_t *dst, const CONV_BUF_TYPE *src0, const CONV_BUF_TYPE *src1, + const __m128i *m, const __m128i *v_round_offset, const __m128i *v_maxval, + int shift) { + const __m128i max_minus_m = _mm_sub_epi16(*v_maxval, *m); + const __m128i s0 = xx_loadu_128(src0); + const __m128i s1 = xx_loadu_128(src1); + __m128i res_lo = _mm_madd_epi16(_mm_unpacklo_epi16(s0, s1), + _mm_unpacklo_epi16(*m, max_minus_m)); + __m128i res_hi = _mm_madd_epi16(_mm_unpackhi_epi16(s0, s1), + _mm_unpackhi_epi16(*m, max_minus_m)); + res_lo = _mm_srai_epi32(_mm_sub_epi32(res_lo, *v_round_offset), shift); + res_hi = _mm_srai_epi32(_mm_sub_epi32(res_hi, *v_round_offset), shift); + const __m128i res_e = _mm_packs_epi32(res_lo, res_hi); + const __m128i res = _mm_packus_epi16(res_e, res_e); + + _mm_storel_epi64((__m128i *)(dst), res); +} + +static INLINE void aom_lowbd_blend_a64_d16_mask_subw0_subh0_w4_sse4_1( + uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, + uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, + const __m128i *round_offset, int shift) { + const __m128i v_maxval = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + for (int i = 0; i < h; ++i) { + const __m128i m0 = xx_loadl_32(mask); + const __m128i m = _mm_cvtepu8_epi16(m0); + + blend_a64_d16_mask_w4_sse41(dst, src0, src1, &m, round_offset, &v_maxval, + shift); + mask += mask_stride; + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + } +} + +static INLINE void aom_lowbd_blend_a64_d16_mask_subw0_subh0_w8_sse4_1( + uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, + uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, + const __m128i *round_offset, int shift) { + const __m128i v_maxval = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + for (int i = 0; i < h; ++i) { + const __m128i m0 = xx_loadl_64(mask); + const __m128i m = _mm_cvtepu8_epi16(m0); + blend_a64_d16_mask_w8_sse41(dst, src0, src1, &m, round_offset, &v_maxval, + shift); + mask += mask_stride; + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + } +} + +static INLINE void aom_lowbd_blend_a64_d16_mask_subw1_subh1_w4_sse4_1( + uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, + uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, + const __m128i *round_offset, int shift) { + const __m128i v_maxval = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + const __m128i one_b = _mm_set1_epi8(1); + const __m128i two_w = _mm_set1_epi16(2); + for (int i = 0; i < h; ++i) { + const __m128i m_i0 = xx_loadl_64(mask); + const __m128i m_i1 = xx_loadl_64(mask + mask_stride); + const __m128i m_ac = _mm_adds_epu8(m_i0, m_i1); + const __m128i m_acbd = _mm_maddubs_epi16(m_ac, one_b); + const __m128i m_acbd_2 = _mm_add_epi16(m_acbd, two_w); + const __m128i m = _mm_srli_epi16(m_acbd_2, 2); + + blend_a64_d16_mask_w4_sse41(dst, src0, src1, &m, round_offset, &v_maxval, + shift); + mask += mask_stride << 1; + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + } +} + +static INLINE void aom_lowbd_blend_a64_d16_mask_subw1_subh1_w8_sse4_1( + uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, + uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, + const __m128i *round_offset, int shift) { + const __m128i v_maxval = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + const __m128i one_b = _mm_set1_epi8(1); + const __m128i two_w = _mm_set1_epi16(2); + for (int i = 0; i < h; ++i) { + const __m128i m_i0 = xx_loadu_128(mask); + const __m128i m_i1 = xx_loadu_128(mask + mask_stride); + const __m128i m_ac = _mm_adds_epu8(m_i0, m_i1); + const __m128i m_acbd = _mm_maddubs_epi16(m_ac, one_b); + const __m128i m_acbd_2 = _mm_add_epi16(m_acbd, two_w); + const __m128i m = _mm_srli_epi16(m_acbd_2, 2); + + blend_a64_d16_mask_w8_sse41(dst, src0, src1, &m, round_offset, &v_maxval, + shift); + mask += mask_stride << 1; + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + } +} + +static INLINE void aom_lowbd_blend_a64_d16_mask_subw1_subh0_w4_sse4_1( + uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, + uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, + const __m128i *round_offset, int shift) { + const __m128i v_maxval = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + const __m128i one_b = _mm_set1_epi8(1); + const __m128i zeros = _mm_setzero_si128(); + for (int i = 0; i < h; ++i) { + const __m128i m_i0 = xx_loadl_64(mask); + const __m128i m_ac = _mm_maddubs_epi16(m_i0, one_b); + const __m128i m = _mm_avg_epu16(m_ac, zeros); + + blend_a64_d16_mask_w4_sse41(dst, src0, src1, &m, round_offset, &v_maxval, + shift); + mask += mask_stride; + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + } +} + +static INLINE void aom_lowbd_blend_a64_d16_mask_subw1_subh0_w8_sse4_1( + uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, + uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, + const __m128i *round_offset, int shift) { + const __m128i v_maxval = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + const __m128i one_b = _mm_set1_epi8(1); + const __m128i zeros = _mm_setzero_si128(); + for (int i = 0; i < h; ++i) { + const __m128i m_i0 = xx_loadu_128(mask); + const __m128i m_ac = _mm_maddubs_epi16(m_i0, one_b); + const __m128i m = _mm_avg_epu16(m_ac, zeros); + + blend_a64_d16_mask_w8_sse41(dst, src0, src1, &m, round_offset, &v_maxval, + shift); + mask += mask_stride; + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + } +} +static INLINE void aom_lowbd_blend_a64_d16_mask_subw0_subh1_w4_sse4_1( + uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, + uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, + const __m128i *round_offset, int shift) { + const __m128i v_maxval = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + const __m128i zeros = _mm_setzero_si128(); + for (int i = 0; i < h; ++i) { + const __m128i m_i0 = xx_loadl_64(mask); + const __m128i m_i1 = xx_loadl_64(mask + mask_stride); + const __m128i m_ac = _mm_adds_epu8(m_i0, m_i1); + const __m128i m = _mm_cvtepu8_epi16(_mm_avg_epu8(m_ac, zeros)); + + blend_a64_d16_mask_w4_sse41(dst, src0, src1, &m, round_offset, &v_maxval, + shift); + mask += mask_stride << 1; + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + } +} + +static INLINE void aom_lowbd_blend_a64_d16_mask_subw0_subh1_w8_sse4_1( + uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, + uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, + const __m128i *round_offset, int shift) { + const __m128i v_maxval = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + const __m128i zeros = _mm_setzero_si128(); + for (int i = 0; i < h; ++i) { + const __m128i m_i0 = xx_loadl_64(mask); + const __m128i m_i1 = xx_loadl_64(mask + mask_stride); + const __m128i m_ac = _mm_adds_epu8(m_i0, m_i1); + const __m128i m = _mm_cvtepu8_epi16(_mm_avg_epu8(m_ac, zeros)); + + blend_a64_d16_mask_w8_sse41(dst, src0, src1, &m, round_offset, &v_maxval, + shift); + mask += mask_stride << 1; + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + } +} +#endif // AOM_AOM_DSP_X86_BLEND_MASK_SSE4_H_ diff --git a/third_party/aom/aom_dsp/x86/blend_sse4.h b/third_party/aom/aom_dsp/x86/blend_sse4.h new file mode 100644 index 0000000000..8d9b325101 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/blend_sse4.h @@ -0,0 +1,191 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_X86_BLEND_SSE4_H_ +#define AOM_AOM_DSP_X86_BLEND_SSE4_H_ + +#include "aom_dsp/blend.h" +#include "aom_dsp/x86/synonyms.h" +static const uint8_t g_blend_a64_mask_shuffle[32] = { + 0, 2, 4, 6, 8, 10, 12, 14, 1, 3, 5, 7, 9, 11, 13, 15, + 0, 2, 4, 6, 8, 10, 12, 14, 1, 3, 5, 7, 9, 11, 13, 15, +}; + +////////////////////////////////////////////////////////////////////////////// +// Common kernels +////////////////////////////////////////////////////////////////////////////// + +static INLINE __m128i blend_4(const uint8_t *src0, const uint8_t *src1, + const __m128i *v_m0_w, const __m128i *v_m1_w) { + const __m128i v_s0_b = xx_loadl_32(src0); + const __m128i v_s1_b = xx_loadl_32(src1); + const __m128i v_s0_w = _mm_cvtepu8_epi16(v_s0_b); + const __m128i v_s1_w = _mm_cvtepu8_epi16(v_s1_b); + + const __m128i v_p0_w = _mm_mullo_epi16(v_s0_w, *v_m0_w); + const __m128i v_p1_w = _mm_mullo_epi16(v_s1_w, *v_m1_w); + const __m128i v_sum_w = _mm_add_epi16(v_p0_w, v_p1_w); + const __m128i v_res_w = xx_roundn_epu16(v_sum_w, AOM_BLEND_A64_ROUND_BITS); + + return v_res_w; +} + +static INLINE __m128i blend_8(const uint8_t *src0, const uint8_t *src1, + const __m128i *v_m0_w, const __m128i *v_m1_w) { + const __m128i v_s0_b = xx_loadl_64(src0); + const __m128i v_s1_b = xx_loadl_64(src1); + const __m128i v_s0_w = _mm_cvtepu8_epi16(v_s0_b); + const __m128i v_s1_w = _mm_cvtepu8_epi16(v_s1_b); + + const __m128i v_p0_w = _mm_mullo_epi16(v_s0_w, *v_m0_w); + const __m128i v_p1_w = _mm_mullo_epi16(v_s1_w, *v_m1_w); + + const __m128i v_sum_w = _mm_add_epi16(v_p0_w, v_p1_w); + + const __m128i v_res_w = xx_roundn_epu16(v_sum_w, AOM_BLEND_A64_ROUND_BITS); + + return v_res_w; +} + +static INLINE __m128i blend_4_u8(const uint8_t *src0, const uint8_t *src1, + const __m128i *v_m0_b, const __m128i *v_m1_b, + const __m128i *rounding) { + const __m128i v_s0_b = xx_loadl_32(src0); + const __m128i v_s1_b = xx_loadl_32(src1); + + const __m128i v_p0_w = _mm_maddubs_epi16(_mm_unpacklo_epi8(v_s0_b, v_s1_b), + _mm_unpacklo_epi8(*v_m0_b, *v_m1_b)); + + const __m128i v_res_w = _mm_mulhrs_epi16(v_p0_w, *rounding); + const __m128i v_res = _mm_packus_epi16(v_res_w, v_res_w); + return v_res; +} + +static INLINE __m128i blend_8_u8(const uint8_t *src0, const uint8_t *src1, + const __m128i *v_m0_b, const __m128i *v_m1_b, + const __m128i *rounding) { + const __m128i v_s0_b = xx_loadl_64(src0); + const __m128i v_s1_b = xx_loadl_64(src1); + + const __m128i v_p0_w = _mm_maddubs_epi16(_mm_unpacklo_epi8(v_s0_b, v_s1_b), + _mm_unpacklo_epi8(*v_m0_b, *v_m1_b)); + + const __m128i v_res_w = _mm_mulhrs_epi16(v_p0_w, *rounding); + const __m128i v_res = _mm_packus_epi16(v_res_w, v_res_w); + return v_res; +} + +static INLINE __m128i blend_16_u8(const uint8_t *src0, const uint8_t *src1, + const __m128i *v_m0_b, const __m128i *v_m1_b, + const __m128i *rounding) { + const __m128i v_s0_b = xx_loadu_128(src0); + const __m128i v_s1_b = xx_loadu_128(src1); + + const __m128i v_p0_w = _mm_maddubs_epi16(_mm_unpacklo_epi8(v_s0_b, v_s1_b), + _mm_unpacklo_epi8(*v_m0_b, *v_m1_b)); + const __m128i v_p1_w = _mm_maddubs_epi16(_mm_unpackhi_epi8(v_s0_b, v_s1_b), + _mm_unpackhi_epi8(*v_m0_b, *v_m1_b)); + + const __m128i v_res0_w = _mm_mulhrs_epi16(v_p0_w, *rounding); + const __m128i v_res1_w = _mm_mulhrs_epi16(v_p1_w, *rounding); + const __m128i v_res = _mm_packus_epi16(v_res0_w, v_res1_w); + return v_res; +} + +typedef __m128i (*blend_unit_fn)(const uint16_t *src0, const uint16_t *src1, + const __m128i v_m0_w, const __m128i v_m1_w); + +static INLINE __m128i blend_4_b10(const uint16_t *src0, const uint16_t *src1, + const __m128i v_m0_w, const __m128i v_m1_w) { + const __m128i v_s0_w = xx_loadl_64(src0); + const __m128i v_s1_w = xx_loadl_64(src1); + + const __m128i v_p0_w = _mm_mullo_epi16(v_s0_w, v_m0_w); + const __m128i v_p1_w = _mm_mullo_epi16(v_s1_w, v_m1_w); + + const __m128i v_sum_w = _mm_add_epi16(v_p0_w, v_p1_w); + + const __m128i v_res_w = xx_roundn_epu16(v_sum_w, AOM_BLEND_A64_ROUND_BITS); + + return v_res_w; +} + +static INLINE __m128i blend_8_b10(const uint16_t *src0, const uint16_t *src1, + const __m128i v_m0_w, const __m128i v_m1_w) { + const __m128i v_s0_w = xx_loadu_128(src0); + const __m128i v_s1_w = xx_loadu_128(src1); + + const __m128i v_p0_w = _mm_mullo_epi16(v_s0_w, v_m0_w); + const __m128i v_p1_w = _mm_mullo_epi16(v_s1_w, v_m1_w); + + const __m128i v_sum_w = _mm_add_epi16(v_p0_w, v_p1_w); + + const __m128i v_res_w = xx_roundn_epu16(v_sum_w, AOM_BLEND_A64_ROUND_BITS); + + return v_res_w; +} + +static INLINE __m128i blend_4_b12(const uint16_t *src0, const uint16_t *src1, + const __m128i v_m0_w, const __m128i v_m1_w) { + const __m128i v_s0_w = xx_loadl_64(src0); + const __m128i v_s1_w = xx_loadl_64(src1); + + // Interleave + const __m128i v_m01_w = _mm_unpacklo_epi16(v_m0_w, v_m1_w); + const __m128i v_s01_w = _mm_unpacklo_epi16(v_s0_w, v_s1_w); + + // Multiply-Add + const __m128i v_sum_d = _mm_madd_epi16(v_s01_w, v_m01_w); + + // Scale + const __m128i v_ssum_d = + _mm_srli_epi32(v_sum_d, AOM_BLEND_A64_ROUND_BITS - 1); + + // Pack + const __m128i v_pssum_d = _mm_packs_epi32(v_ssum_d, v_ssum_d); + + // Round + const __m128i v_res_w = xx_round_epu16(v_pssum_d); + + return v_res_w; +} + +static INLINE __m128i blend_8_b12(const uint16_t *src0, const uint16_t *src1, + const __m128i v_m0_w, const __m128i v_m1_w) { + const __m128i v_s0_w = xx_loadu_128(src0); + const __m128i v_s1_w = xx_loadu_128(src1); + + // Interleave + const __m128i v_m01l_w = _mm_unpacklo_epi16(v_m0_w, v_m1_w); + const __m128i v_m01h_w = _mm_unpackhi_epi16(v_m0_w, v_m1_w); + const __m128i v_s01l_w = _mm_unpacklo_epi16(v_s0_w, v_s1_w); + const __m128i v_s01h_w = _mm_unpackhi_epi16(v_s0_w, v_s1_w); + + // Multiply-Add + const __m128i v_suml_d = _mm_madd_epi16(v_s01l_w, v_m01l_w); + const __m128i v_sumh_d = _mm_madd_epi16(v_s01h_w, v_m01h_w); + + // Scale + const __m128i v_ssuml_d = + _mm_srli_epi32(v_suml_d, AOM_BLEND_A64_ROUND_BITS - 1); + const __m128i v_ssumh_d = + _mm_srli_epi32(v_sumh_d, AOM_BLEND_A64_ROUND_BITS - 1); + + // Pack + const __m128i v_pssum_d = _mm_packs_epi32(v_ssuml_d, v_ssumh_d); + + // Round + const __m128i v_res_w = xx_round_epu16(v_pssum_d); + + return v_res_w; +} + +#endif // AOM_AOM_DSP_X86_BLEND_SSE4_H_ diff --git a/third_party/aom/aom_dsp/x86/common_avx2.h b/third_party/aom/aom_dsp/x86/common_avx2.h new file mode 100644 index 0000000000..96fe4ebb67 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/common_avx2.h @@ -0,0 +1,147 @@ +/* + * Copyright (c) 2017, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_X86_COMMON_AVX2_H_ +#define AOM_AOM_DSP_X86_COMMON_AVX2_H_ + +#include + +#include "config/aom_config.h" + +// Note: in and out could have the same value +static INLINE void mm256_transpose_16x16(const __m256i *in, __m256i *out) { + __m256i tr0_0 = _mm256_unpacklo_epi16(in[0], in[1]); + __m256i tr0_1 = _mm256_unpackhi_epi16(in[0], in[1]); + __m256i tr0_2 = _mm256_unpacklo_epi16(in[2], in[3]); + __m256i tr0_3 = _mm256_unpackhi_epi16(in[2], in[3]); + __m256i tr0_4 = _mm256_unpacklo_epi16(in[4], in[5]); + __m256i tr0_5 = _mm256_unpackhi_epi16(in[4], in[5]); + __m256i tr0_6 = _mm256_unpacklo_epi16(in[6], in[7]); + __m256i tr0_7 = _mm256_unpackhi_epi16(in[6], in[7]); + + __m256i tr0_8 = _mm256_unpacklo_epi16(in[8], in[9]); + __m256i tr0_9 = _mm256_unpackhi_epi16(in[8], in[9]); + __m256i tr0_a = _mm256_unpacklo_epi16(in[10], in[11]); + __m256i tr0_b = _mm256_unpackhi_epi16(in[10], in[11]); + __m256i tr0_c = _mm256_unpacklo_epi16(in[12], in[13]); + __m256i tr0_d = _mm256_unpackhi_epi16(in[12], in[13]); + __m256i tr0_e = _mm256_unpacklo_epi16(in[14], in[15]); + __m256i tr0_f = _mm256_unpackhi_epi16(in[14], in[15]); + + // 00 10 01 11 02 12 03 13 08 18 09 19 0a 1a 0b 1b + // 04 14 05 15 06 16 07 17 0c 1c 0d 1d 0e 1e 0f 1f + // 20 30 21 31 22 32 23 33 28 38 29 39 2a 3a 2b 3b + // 24 34 25 35 26 36 27 37 2c 3c 2d 3d 2e 3e 2f 3f + // 40 50 41 51 42 52 43 53 48 58 49 59 4a 5a 4b 5b + // 44 54 45 55 46 56 47 57 4c 5c 4d 5d 4e 5e 4f 5f + // 60 70 61 71 62 72 63 73 68 78 69 79 6a 7a 6b 7b + // 64 74 65 75 66 76 67 77 6c 7c 6d 7d 6e 7e 6f 7f + + // 80 90 81 91 82 92 83 93 88 98 89 99 8a 9a 8b 9b + // 84 94 85 95 86 96 87 97 8c 9c 8d 9d 8e 9e 8f 9f + // a0 b0 a1 b1 a2 b2 a3 b3 a8 b8 a9 b9 aa ba ab bb + // a4 b4 a5 b5 a6 b6 a7 b7 ac bc ad bd ae be af bf + // c0 d0 c1 d1 c2 d2 c3 d3 c8 d8 c9 d9 ca da cb db + // c4 d4 c5 d5 c6 d6 c7 d7 cc dc cd dd ce de cf df + // e0 f0 e1 f1 e2 f2 e3 f3 e8 f8 e9 f9 ea fa eb fb + // e4 f4 e5 f5 e6 f6 e7 f7 ec fc ed fd ee fe ef ff + + __m256i tr1_0 = _mm256_unpacklo_epi32(tr0_0, tr0_2); + __m256i tr1_1 = _mm256_unpackhi_epi32(tr0_0, tr0_2); + __m256i tr1_2 = _mm256_unpacklo_epi32(tr0_1, tr0_3); + __m256i tr1_3 = _mm256_unpackhi_epi32(tr0_1, tr0_3); + __m256i tr1_4 = _mm256_unpacklo_epi32(tr0_4, tr0_6); + __m256i tr1_5 = _mm256_unpackhi_epi32(tr0_4, tr0_6); + __m256i tr1_6 = _mm256_unpacklo_epi32(tr0_5, tr0_7); + __m256i tr1_7 = _mm256_unpackhi_epi32(tr0_5, tr0_7); + + __m256i tr1_8 = _mm256_unpacklo_epi32(tr0_8, tr0_a); + __m256i tr1_9 = _mm256_unpackhi_epi32(tr0_8, tr0_a); + __m256i tr1_a = _mm256_unpacklo_epi32(tr0_9, tr0_b); + __m256i tr1_b = _mm256_unpackhi_epi32(tr0_9, tr0_b); + __m256i tr1_c = _mm256_unpacklo_epi32(tr0_c, tr0_e); + __m256i tr1_d = _mm256_unpackhi_epi32(tr0_c, tr0_e); + __m256i tr1_e = _mm256_unpacklo_epi32(tr0_d, tr0_f); + __m256i tr1_f = _mm256_unpackhi_epi32(tr0_d, tr0_f); + + // 00 10 20 30 01 11 21 31 08 18 28 38 09 19 29 39 + // 02 12 22 32 03 13 23 33 0a 1a 2a 3a 0b 1b 2b 3b + // 04 14 24 34 05 15 25 35 0c 1c 2c 3c 0d 1d 2d 3d + // 06 16 26 36 07 17 27 37 0e 1e 2e 3e 0f 1f 2f 3f + // 40 50 60 70 41 51 61 71 48 58 68 78 49 59 69 79 + // 42 52 62 72 43 53 63 73 4a 5a 6a 7a 4b 5b 6b 7b + // 44 54 64 74 45 55 65 75 4c 5c 6c 7c 4d 5d 6d 7d + // 46 56 66 76 47 57 67 77 4e 5e 6e 7e 4f 5f 6f 7f + + // 80 90 a0 b0 81 91 a1 b1 88 98 a8 b8 89 99 a9 b9 + // 82 92 a2 b2 83 93 a3 b3 8a 9a aa ba 8b 9b ab bb + // 84 94 a4 b4 85 95 a5 b5 8c 9c ac bc 8d 9d ad bd + // 86 96 a6 b6 87 97 a7 b7 8e ae 9e be 8f 9f af bf + // c0 d0 e0 f0 c1 d1 e1 f1 c8 d8 e8 f8 c9 d9 e9 f9 + // c2 d2 e2 f2 c3 d3 e3 f3 ca da ea fa cb db eb fb + // c4 d4 e4 f4 c5 d5 e5 f5 cc dc ef fc cd dd ed fd + // c6 d6 e6 f6 c7 d7 e7 f7 ce de ee fe cf df ef ff + + tr0_0 = _mm256_unpacklo_epi64(tr1_0, tr1_4); + tr0_1 = _mm256_unpackhi_epi64(tr1_0, tr1_4); + tr0_2 = _mm256_unpacklo_epi64(tr1_1, tr1_5); + tr0_3 = _mm256_unpackhi_epi64(tr1_1, tr1_5); + tr0_4 = _mm256_unpacklo_epi64(tr1_2, tr1_6); + tr0_5 = _mm256_unpackhi_epi64(tr1_2, tr1_6); + tr0_6 = _mm256_unpacklo_epi64(tr1_3, tr1_7); + tr0_7 = _mm256_unpackhi_epi64(tr1_3, tr1_7); + + tr0_8 = _mm256_unpacklo_epi64(tr1_8, tr1_c); + tr0_9 = _mm256_unpackhi_epi64(tr1_8, tr1_c); + tr0_a = _mm256_unpacklo_epi64(tr1_9, tr1_d); + tr0_b = _mm256_unpackhi_epi64(tr1_9, tr1_d); + tr0_c = _mm256_unpacklo_epi64(tr1_a, tr1_e); + tr0_d = _mm256_unpackhi_epi64(tr1_a, tr1_e); + tr0_e = _mm256_unpacklo_epi64(tr1_b, tr1_f); + tr0_f = _mm256_unpackhi_epi64(tr1_b, tr1_f); + + // 00 10 20 30 40 50 60 70 08 18 28 38 48 58 68 78 + // 01 11 21 31 41 51 61 71 09 19 29 39 49 59 69 79 + // 02 12 22 32 42 52 62 72 0a 1a 2a 3a 4a 5a 6a 7a + // 03 13 23 33 43 53 63 73 0b 1b 2b 3b 4b 5b 6b 7b + // 04 14 24 34 44 54 64 74 0c 1c 2c 3c 4c 5c 6c 7c + // 05 15 25 35 45 55 65 75 0d 1d 2d 3d 4d 5d 6d 7d + // 06 16 26 36 46 56 66 76 0e 1e 2e 3e 4e 5e 6e 7e + // 07 17 27 37 47 57 67 77 0f 1f 2f 3f 4f 5f 6f 7f + + // 80 90 a0 b0 c0 d0 e0 f0 88 98 a8 b8 c8 d8 e8 f8 + // 81 91 a1 b1 c1 d1 e1 f1 89 99 a9 b9 c9 d9 e9 f9 + // 82 92 a2 b2 c2 d2 e2 f2 8a 9a aa ba ca da ea fa + // 83 93 a3 b3 c3 d3 e3 f3 8b 9b ab bb cb db eb fb + // 84 94 a4 b4 c4 d4 e4 f4 8c 9c ac bc cc dc ef fc + // 85 95 a5 b5 c5 d5 e5 f5 8d 9d ad bd cd dd ed fd + // 86 96 a6 b6 c6 d6 e6 f6 8e ae 9e be ce de ee fe + // 87 97 a7 b7 c7 d7 e7 f7 8f 9f af bf cf df ef ff + + out[0] = _mm256_permute2x128_si256(tr0_0, tr0_8, 0x20); // 0010 0000 + out[8] = _mm256_permute2x128_si256(tr0_0, tr0_8, 0x31); // 0011 0001 + out[1] = _mm256_permute2x128_si256(tr0_1, tr0_9, 0x20); + out[9] = _mm256_permute2x128_si256(tr0_1, tr0_9, 0x31); + out[2] = _mm256_permute2x128_si256(tr0_2, tr0_a, 0x20); + out[10] = _mm256_permute2x128_si256(tr0_2, tr0_a, 0x31); + out[3] = _mm256_permute2x128_si256(tr0_3, tr0_b, 0x20); + out[11] = _mm256_permute2x128_si256(tr0_3, tr0_b, 0x31); + + out[4] = _mm256_permute2x128_si256(tr0_4, tr0_c, 0x20); + out[12] = _mm256_permute2x128_si256(tr0_4, tr0_c, 0x31); + out[5] = _mm256_permute2x128_si256(tr0_5, tr0_d, 0x20); + out[13] = _mm256_permute2x128_si256(tr0_5, tr0_d, 0x31); + out[6] = _mm256_permute2x128_si256(tr0_6, tr0_e, 0x20); + out[14] = _mm256_permute2x128_si256(tr0_6, tr0_e, 0x31); + out[7] = _mm256_permute2x128_si256(tr0_7, tr0_f, 0x20); + out[15] = _mm256_permute2x128_si256(tr0_7, tr0_f, 0x31); +} +#endif // AOM_AOM_DSP_X86_COMMON_AVX2_H_ diff --git a/third_party/aom/aom_dsp/x86/convolve.h b/third_party/aom/aom_dsp/x86/convolve.h new file mode 100644 index 0000000000..3e19682cd9 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/convolve.h @@ -0,0 +1,178 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ +#ifndef AOM_AOM_DSP_X86_CONVOLVE_H_ +#define AOM_AOM_DSP_X86_CONVOLVE_H_ + +#include + +#include "config/aom_config.h" + +#include "aom/aom_integer.h" +#include "aom_ports/mem.h" + +typedef void filter8_1dfunction(const uint8_t *src_ptr, ptrdiff_t src_pitch, + uint8_t *output_ptr, ptrdiff_t out_pitch, + uint32_t output_height, const int16_t *filter); + +#define FUN_CONV_1D(name, step_q4, filter, dir, src_start, avg, opt) \ + void aom_convolve8_##name##_##opt( \ + const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, \ + ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, \ + const int16_t *filter_y, int y_step_q4, int w, int h) { \ + (void)filter_x; \ + (void)x_step_q4; \ + (void)filter_y; \ + (void)y_step_q4; \ + assert((-128 <= filter[3]) && (filter[3] <= 127)); \ + assert(step_q4 == 16); \ + if (((filter[0] | filter[1] | filter[6] | filter[7]) == 0) && \ + (filter[2] | filter[5])) { \ + while (w >= 16) { \ + aom_filter_block1d16_##dir##4_##avg##opt(src_start, src_stride, dst, \ + dst_stride, h, filter); \ + src += 16; \ + dst += 16; \ + w -= 16; \ + } \ + while (w >= 8) { \ + aom_filter_block1d8_##dir##4_##avg##opt(src_start, src_stride, dst, \ + dst_stride, h, filter); \ + src += 8; \ + dst += 8; \ + w -= 8; \ + } \ + while (w >= 4) { \ + aom_filter_block1d4_##dir##4_##avg##opt(src_start, src_stride, dst, \ + dst_stride, h, filter); \ + src += 4; \ + dst += 4; \ + w -= 4; \ + } \ + } else if (filter[0] | filter[1] | filter[2]) { \ + while (w >= 16) { \ + aom_filter_block1d16_##dir##8_##avg##opt(src_start, src_stride, dst, \ + dst_stride, h, filter); \ + src += 16; \ + dst += 16; \ + w -= 16; \ + } \ + while (w >= 8) { \ + aom_filter_block1d8_##dir##8_##avg##opt(src_start, src_stride, dst, \ + dst_stride, h, filter); \ + src += 8; \ + dst += 8; \ + w -= 8; \ + } \ + while (w >= 4) { \ + aom_filter_block1d4_##dir##8_##avg##opt(src_start, src_stride, dst, \ + dst_stride, h, filter); \ + src += 4; \ + dst += 4; \ + w -= 4; \ + } \ + } else { \ + while (w >= 16) { \ + aom_filter_block1d16_##dir##2_##avg##opt(src, src_stride, dst, \ + dst_stride, h, filter); \ + src += 16; \ + dst += 16; \ + w -= 16; \ + } \ + while (w >= 8) { \ + aom_filter_block1d8_##dir##2_##avg##opt(src, src_stride, dst, \ + dst_stride, h, filter); \ + src += 8; \ + dst += 8; \ + w -= 8; \ + } \ + while (w >= 4) { \ + aom_filter_block1d4_##dir##2_##avg##opt(src, src_stride, dst, \ + dst_stride, h, filter); \ + src += 4; \ + dst += 4; \ + w -= 4; \ + } \ + } \ + if (w) { \ + aom_convolve8_##name##_c(src, src_stride, dst, dst_stride, filter_x, \ + x_step_q4, filter_y, y_step_q4, w, h); \ + } \ + } + +typedef void highbd_filter8_1dfunction(const uint16_t *src_ptr, + const ptrdiff_t src_pitch, + uint16_t *output_ptr, + ptrdiff_t out_pitch, + unsigned int output_height, + const int16_t *filter, int bd); + +#define HIGH_FUN_CONV_1D(name, step_q4, filter, dir, src_start, avg, opt) \ + void aom_highbd_convolve8_##name##_##opt( \ + const uint8_t *src8, ptrdiff_t src_stride, uint8_t *dst8, \ + ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, \ + const int16_t *filter_y, int y_step_q4, int w, int h, int bd) { \ + uint16_t *src = CONVERT_TO_SHORTPTR(src8); \ + uint16_t *dst = CONVERT_TO_SHORTPTR(dst8); \ + if (step_q4 == 16 && filter[3] != 128) { \ + if (filter[0] | filter[1] | filter[2]) { \ + while (w >= 16) { \ + aom_highbd_filter_block1d16_##dir##8_##avg##opt( \ + src_start, src_stride, dst, dst_stride, h, filter, bd); \ + src += 16; \ + dst += 16; \ + w -= 16; \ + } \ + while (w >= 8) { \ + aom_highbd_filter_block1d8_##dir##8_##avg##opt( \ + src_start, src_stride, dst, dst_stride, h, filter, bd); \ + src += 8; \ + dst += 8; \ + w -= 8; \ + } \ + while (w >= 4) { \ + aom_highbd_filter_block1d4_##dir##8_##avg##opt( \ + src_start, src_stride, dst, dst_stride, h, filter, bd); \ + src += 4; \ + dst += 4; \ + w -= 4; \ + } \ + } else { \ + while (w >= 16) { \ + aom_highbd_filter_block1d16_##dir##2_##avg##opt( \ + src, src_stride, dst, dst_stride, h, filter, bd); \ + src += 16; \ + dst += 16; \ + w -= 16; \ + } \ + while (w >= 8) { \ + aom_highbd_filter_block1d8_##dir##2_##avg##opt( \ + src, src_stride, dst, dst_stride, h, filter, bd); \ + src += 8; \ + dst += 8; \ + w -= 8; \ + } \ + while (w >= 4) { \ + aom_highbd_filter_block1d4_##dir##2_##avg##opt( \ + src, src_stride, dst, dst_stride, h, filter, bd); \ + src += 4; \ + dst += 4; \ + w -= 4; \ + } \ + } \ + } \ + if (w) { \ + aom_highbd_convolve8_##name##_c( \ + CONVERT_TO_BYTEPTR(src), src_stride, CONVERT_TO_BYTEPTR(dst), \ + dst_stride, filter_x, x_step_q4, filter_y, y_step_q4, w, h, bd); \ + } \ + } + +#endif // AOM_AOM_DSP_X86_CONVOLVE_H_ diff --git a/third_party/aom/aom_dsp/x86/convolve_avx2.h b/third_party/aom/aom_dsp/x86/convolve_avx2.h new file mode 100644 index 0000000000..30253f65c2 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/convolve_avx2.h @@ -0,0 +1,199 @@ +/* + * Copyright (c) 2018, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_X86_CONVOLVE_AVX2_H_ +#define AOM_AOM_DSP_X86_CONVOLVE_AVX2_H_ + +// filters for 16 +DECLARE_ALIGNED(32, static const uint8_t, filt_global_avx2[]) = { + 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8, 0, 1, 1, + 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8, 2, 3, 3, 4, 4, 5, + 5, 6, 6, 7, 7, 8, 8, 9, 9, 10, 2, 3, 3, 4, 4, 5, 5, 6, 6, + 7, 7, 8, 8, 9, 9, 10, 4, 5, 5, 6, 6, 7, 7, 8, 8, 9, 9, 10, + 10, 11, 11, 12, 4, 5, 5, 6, 6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, + 12, 6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14, 6, 7, + 7, 8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14 +}; + +DECLARE_ALIGNED(32, static const uint8_t, filt_d4_global_avx2[]) = { + 0, 1, 2, 3, 1, 2, 3, 4, 2, 3, 4, 5, 3, 4, 5, 6, 0, 1, 2, 3, 1, 2, + 3, 4, 2, 3, 4, 5, 3, 4, 5, 6, 4, 5, 6, 7, 5, 6, 7, 8, 6, 7, 8, 9, + 7, 8, 9, 10, 4, 5, 6, 7, 5, 6, 7, 8, 6, 7, 8, 9, 7, 8, 9, 10, +}; + +DECLARE_ALIGNED(32, static const uint8_t, filt4_d4_global_avx2[]) = { + 2, 3, 4, 5, 3, 4, 5, 6, 4, 5, 6, 7, 5, 6, 7, 8, + 2, 3, 4, 5, 3, 4, 5, 6, 4, 5, 6, 7, 5, 6, 7, 8, +}; + +static INLINE void prepare_coeffs_lowbd( + const InterpFilterParams *const filter_params, const int subpel_q4, + __m256i *const coeffs /* [4] */) { + const int16_t *const filter = av1_get_interp_filter_subpel_kernel( + filter_params, subpel_q4 & SUBPEL_MASK); + const __m128i coeffs_8 = _mm_loadu_si128((__m128i *)filter); + const __m256i filter_coeffs = _mm256_broadcastsi128_si256(coeffs_8); + + // right shift all filter co-efficients by 1 to reduce the bits required. + // This extra right shift will be taken care of at the end while rounding + // the result. + // Since all filter co-efficients are even, this change will not affect the + // end result + assert(_mm_test_all_zeros(_mm_and_si128(coeffs_8, _mm_set1_epi16(1)), + _mm_set1_epi16(0xffff))); + + const __m256i coeffs_1 = _mm256_srai_epi16(filter_coeffs, 1); + + // coeffs 0 1 0 1 0 1 0 1 + coeffs[0] = _mm256_shuffle_epi8(coeffs_1, _mm256_set1_epi16(0x0200u)); + // coeffs 2 3 2 3 2 3 2 3 + coeffs[1] = _mm256_shuffle_epi8(coeffs_1, _mm256_set1_epi16(0x0604u)); + // coeffs 4 5 4 5 4 5 4 5 + coeffs[2] = _mm256_shuffle_epi8(coeffs_1, _mm256_set1_epi16(0x0a08u)); + // coeffs 6 7 6 7 6 7 6 7 + coeffs[3] = _mm256_shuffle_epi8(coeffs_1, _mm256_set1_epi16(0x0e0cu)); +} + +static INLINE void prepare_coeffs(const InterpFilterParams *const filter_params, + const int subpel_q4, + __m256i *const coeffs /* [4] */) { + const int16_t *filter = av1_get_interp_filter_subpel_kernel( + filter_params, subpel_q4 & SUBPEL_MASK); + + const __m128i coeff_8 = _mm_loadu_si128((__m128i *)filter); + const __m256i coeff = _mm256_broadcastsi128_si256(coeff_8); + + // coeffs 0 1 0 1 0 1 0 1 + coeffs[0] = _mm256_shuffle_epi32(coeff, 0x00); + // coeffs 2 3 2 3 2 3 2 3 + coeffs[1] = _mm256_shuffle_epi32(coeff, 0x55); + // coeffs 4 5 4 5 4 5 4 5 + coeffs[2] = _mm256_shuffle_epi32(coeff, 0xaa); + // coeffs 6 7 6 7 6 7 6 7 + coeffs[3] = _mm256_shuffle_epi32(coeff, 0xff); +} + +static INLINE __m256i convolve_lowbd(const __m256i *const s, + const __m256i *const coeffs) { + const __m256i res_01 = _mm256_maddubs_epi16(s[0], coeffs[0]); + const __m256i res_23 = _mm256_maddubs_epi16(s[1], coeffs[1]); + const __m256i res_45 = _mm256_maddubs_epi16(s[2], coeffs[2]); + const __m256i res_67 = _mm256_maddubs_epi16(s[3], coeffs[3]); + + // order: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 + const __m256i res = _mm256_add_epi16(_mm256_add_epi16(res_01, res_45), + _mm256_add_epi16(res_23, res_67)); + + return res; +} + +static INLINE __m256i convolve(const __m256i *const s, + const __m256i *const coeffs) { + const __m256i res_0 = _mm256_madd_epi16(s[0], coeffs[0]); + const __m256i res_1 = _mm256_madd_epi16(s[1], coeffs[1]); + const __m256i res_2 = _mm256_madd_epi16(s[2], coeffs[2]); + const __m256i res_3 = _mm256_madd_epi16(s[3], coeffs[3]); + + const __m256i res = _mm256_add_epi32(_mm256_add_epi32(res_0, res_1), + _mm256_add_epi32(res_2, res_3)); + + return res; +} + +static INLINE __m256i convolve_lowbd_x(const __m256i data, + const __m256i *const coeffs, + const __m256i *const filt) { + __m256i s[4]; + + s[0] = _mm256_shuffle_epi8(data, filt[0]); + s[1] = _mm256_shuffle_epi8(data, filt[1]); + s[2] = _mm256_shuffle_epi8(data, filt[2]); + s[3] = _mm256_shuffle_epi8(data, filt[3]); + + return convolve_lowbd(s, coeffs); +} + +static INLINE void add_store_aligned_256(CONV_BUF_TYPE *const dst, + const __m256i *const res, + const int do_average) { + __m256i d; + if (do_average) { + d = _mm256_load_si256((__m256i *)dst); + d = _mm256_add_epi32(d, *res); + d = _mm256_srai_epi32(d, 1); + } else { + d = *res; + } + _mm256_store_si256((__m256i *)dst, d); +} + +static INLINE __m256i comp_avg(const __m256i *const data_ref_0, + const __m256i *const res_unsigned, + const __m256i *const wt, + const int use_jnt_comp_avg) { + __m256i res; + if (use_jnt_comp_avg) { + const __m256i data_lo = _mm256_unpacklo_epi16(*data_ref_0, *res_unsigned); + const __m256i data_hi = _mm256_unpackhi_epi16(*data_ref_0, *res_unsigned); + + const __m256i wt_res_lo = _mm256_madd_epi16(data_lo, *wt); + const __m256i wt_res_hi = _mm256_madd_epi16(data_hi, *wt); + + const __m256i res_lo = _mm256_srai_epi32(wt_res_lo, DIST_PRECISION_BITS); + const __m256i res_hi = _mm256_srai_epi32(wt_res_hi, DIST_PRECISION_BITS); + + res = _mm256_packs_epi32(res_lo, res_hi); + } else { + const __m256i wt_res = _mm256_add_epi16(*data_ref_0, *res_unsigned); + res = _mm256_srai_epi16(wt_res, 1); + } + return res; +} + +static INLINE __m256i convolve_rounding(const __m256i *const res_unsigned, + const __m256i *const offset_const, + const __m256i *const round_const, + const int round_shift) { + const __m256i res_signed = _mm256_sub_epi16(*res_unsigned, *offset_const); + const __m256i res_round = _mm256_srai_epi16( + _mm256_add_epi16(res_signed, *round_const), round_shift); + return res_round; +} + +static INLINE __m256i highbd_comp_avg(const __m256i *const data_ref_0, + const __m256i *const res_unsigned, + const __m256i *const wt0, + const __m256i *const wt1, + const int use_jnt_comp_avg) { + __m256i res; + if (use_jnt_comp_avg) { + const __m256i wt0_res = _mm256_mullo_epi32(*data_ref_0, *wt0); + const __m256i wt1_res = _mm256_mullo_epi32(*res_unsigned, *wt1); + const __m256i wt_res = _mm256_add_epi32(wt0_res, wt1_res); + res = _mm256_srai_epi32(wt_res, DIST_PRECISION_BITS); + } else { + const __m256i wt_res = _mm256_add_epi32(*data_ref_0, *res_unsigned); + res = _mm256_srai_epi32(wt_res, 1); + } + return res; +} + +static INLINE __m256i highbd_convolve_rounding( + const __m256i *const res_unsigned, const __m256i *const offset_const, + const __m256i *const round_const, const int round_shift) { + const __m256i res_signed = _mm256_sub_epi32(*res_unsigned, *offset_const); + const __m256i res_round = _mm256_srai_epi32( + _mm256_add_epi32(res_signed, *round_const), round_shift); + + return res_round; +} + +#endif // AOM_AOM_DSP_X86_CONVOLVE_AVX2_H_ diff --git a/third_party/aom/aom_dsp/x86/convolve_common_intrin.h b/third_party/aom/aom_dsp/x86/convolve_common_intrin.h new file mode 100644 index 0000000000..707bd2d782 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/convolve_common_intrin.h @@ -0,0 +1,31 @@ +/* + * Copyright (c) 2018, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_X86_CONVOLVE_COMMON_INTRIN_H_ +#define AOM_AOM_DSP_X86_CONVOLVE_COMMON_INTRIN_H_ + +// Note: +// This header file should be put below any x86 intrinsics head file + +static INLINE void add_store(CONV_BUF_TYPE *const dst, const __m128i *const res, + const int do_average) { + __m128i d; + if (do_average) { + d = _mm_load_si128((__m128i *)dst); + d = _mm_add_epi32(d, *res); + d = _mm_srai_epi32(d, 1); + } else { + d = *res; + } + _mm_store_si128((__m128i *)dst, d); +} + +#endif // AOM_AOM_DSP_X86_CONVOLVE_COMMON_INTRIN_H_ diff --git a/third_party/aom/aom_dsp/x86/convolve_sse2.h b/third_party/aom/aom_dsp/x86/convolve_sse2.h new file mode 100644 index 0000000000..445d04b103 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/convolve_sse2.h @@ -0,0 +1,121 @@ +/* + * Copyright (c) 2018, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_X86_CONVOLVE_SSE2_H_ +#define AOM_AOM_DSP_X86_CONVOLVE_SSE2_H_ + +// Note: +// This header file should be put below any x86 intrinsics head file + +static INLINE void prepare_coeffs(const InterpFilterParams *const filter_params, + const int subpel_q4, + __m128i *const coeffs /* [4] */) { + const int16_t *filter = av1_get_interp_filter_subpel_kernel( + filter_params, subpel_q4 & SUBPEL_MASK); + const __m128i coeff = _mm_loadu_si128((__m128i *)filter); + + // coeffs 0 1 0 1 0 1 0 1 + coeffs[0] = _mm_shuffle_epi32(coeff, 0x00); + // coeffs 2 3 2 3 2 3 2 3 + coeffs[1] = _mm_shuffle_epi32(coeff, 0x55); + // coeffs 4 5 4 5 4 5 4 5 + coeffs[2] = _mm_shuffle_epi32(coeff, 0xaa); + // coeffs 6 7 6 7 6 7 6 7 + coeffs[3] = _mm_shuffle_epi32(coeff, 0xff); +} + +static INLINE __m128i convolve(const __m128i *const s, + const __m128i *const coeffs) { + const __m128i res_0 = _mm_madd_epi16(s[0], coeffs[0]); + const __m128i res_1 = _mm_madd_epi16(s[1], coeffs[1]); + const __m128i res_2 = _mm_madd_epi16(s[2], coeffs[2]); + const __m128i res_3 = _mm_madd_epi16(s[3], coeffs[3]); + + const __m128i res = + _mm_add_epi32(_mm_add_epi32(res_0, res_1), _mm_add_epi32(res_2, res_3)); + + return res; +} + +static INLINE __m128i convolve_lo_x(const __m128i *const s, + const __m128i *const coeffs) { + __m128i ss[4]; + ss[0] = _mm_unpacklo_epi8(s[0], _mm_setzero_si128()); + ss[1] = _mm_unpacklo_epi8(s[1], _mm_setzero_si128()); + ss[2] = _mm_unpacklo_epi8(s[2], _mm_setzero_si128()); + ss[3] = _mm_unpacklo_epi8(s[3], _mm_setzero_si128()); + return convolve(ss, coeffs); +} + +static INLINE __m128i convolve_lo_y(const __m128i *const s, + const __m128i *const coeffs) { + __m128i ss[4]; + ss[0] = _mm_unpacklo_epi8(s[0], _mm_setzero_si128()); + ss[1] = _mm_unpacklo_epi8(s[2], _mm_setzero_si128()); + ss[2] = _mm_unpacklo_epi8(s[4], _mm_setzero_si128()); + ss[3] = _mm_unpacklo_epi8(s[6], _mm_setzero_si128()); + return convolve(ss, coeffs); +} + +static INLINE __m128i convolve_hi_y(const __m128i *const s, + const __m128i *const coeffs) { + __m128i ss[4]; + ss[0] = _mm_unpackhi_epi8(s[0], _mm_setzero_si128()); + ss[1] = _mm_unpackhi_epi8(s[2], _mm_setzero_si128()); + ss[2] = _mm_unpackhi_epi8(s[4], _mm_setzero_si128()); + ss[3] = _mm_unpackhi_epi8(s[6], _mm_setzero_si128()); + return convolve(ss, coeffs); +} + +static INLINE __m128i comp_avg(const __m128i *const data_ref_0, + const __m128i *const res_unsigned, + const __m128i *const wt, + const int use_jnt_comp_avg) { + __m128i res; + if (use_jnt_comp_avg) { + const __m128i data_lo = _mm_unpacklo_epi16(*data_ref_0, *res_unsigned); + const __m128i data_hi = _mm_unpackhi_epi16(*data_ref_0, *res_unsigned); + + const __m128i wt_res_lo = _mm_madd_epi16(data_lo, *wt); + const __m128i wt_res_hi = _mm_madd_epi16(data_hi, *wt); + + const __m128i res_lo = _mm_srai_epi32(wt_res_lo, DIST_PRECISION_BITS); + const __m128i res_hi = _mm_srai_epi32(wt_res_hi, DIST_PRECISION_BITS); + + res = _mm_packs_epi32(res_lo, res_hi); + } else { + const __m128i wt_res = _mm_add_epi16(*data_ref_0, *res_unsigned); + res = _mm_srai_epi16(wt_res, 1); + } + return res; +} + +static INLINE __m128i convolve_rounding(const __m128i *const res_unsigned, + const __m128i *const offset_const, + const __m128i *const round_const, + const int round_shift) { + const __m128i res_signed = _mm_sub_epi16(*res_unsigned, *offset_const); + const __m128i res_round = + _mm_srai_epi16(_mm_add_epi16(res_signed, *round_const), round_shift); + return res_round; +} + +static INLINE __m128i highbd_convolve_rounding_sse2( + const __m128i *const res_unsigned, const __m128i *const offset_const, + const __m128i *const round_const, const int round_shift) { + const __m128i res_signed = _mm_sub_epi32(*res_unsigned, *offset_const); + const __m128i res_round = + _mm_srai_epi32(_mm_add_epi32(res_signed, *round_const), round_shift); + + return res_round; +} + +#endif // AOM_AOM_DSP_X86_CONVOLVE_SSE2_H_ diff --git a/third_party/aom/aom_dsp/x86/convolve_sse4_1.h b/third_party/aom/aom_dsp/x86/convolve_sse4_1.h new file mode 100644 index 0000000000..6b8388d84a --- /dev/null +++ b/third_party/aom/aom_dsp/x86/convolve_sse4_1.h @@ -0,0 +1,53 @@ +/* + * Copyright (c) 2018, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_X86_CONVOLVE_SSE4_1_H_ +#define AOM_AOM_DSP_X86_CONVOLVE_SSE4_1_H_ + +// Note: +// This header file should be put below any x86 intrinsics head file + +static INLINE void mult_add_store(CONV_BUF_TYPE *const dst, + const __m128i *const res, + const __m128i *const wt0, + const __m128i *const wt1, + const int do_average) { + __m128i d; + if (do_average) { + d = _mm_load_si128((__m128i *)dst); + d = _mm_add_epi32(_mm_mullo_epi32(d, *wt0), _mm_mullo_epi32(*res, *wt1)); + d = _mm_srai_epi32(d, DIST_PRECISION_BITS); + } else { + d = *res; + } + _mm_store_si128((__m128i *)dst, d); +} + +static INLINE __m128i highbd_comp_avg_sse4_1(const __m128i *const data_ref_0, + const __m128i *const res_unsigned, + const __m128i *const wt0, + const __m128i *const wt1, + const int use_jnt_comp_avg) { + __m128i res; + if (use_jnt_comp_avg) { + const __m128i wt0_res = _mm_mullo_epi32(*data_ref_0, *wt0); + const __m128i wt1_res = _mm_mullo_epi32(*res_unsigned, *wt1); + + const __m128i wt_res = _mm_add_epi32(wt0_res, wt1_res); + res = _mm_srai_epi32(wt_res, DIST_PRECISION_BITS); + } else { + const __m128i wt_res = _mm_add_epi32(*data_ref_0, *res_unsigned); + res = _mm_srai_epi32(wt_res, 1); + } + return res; +} + +#endif // AOM_AOM_DSP_X86_CONVOLVE_SSE4_1_H_ diff --git a/third_party/aom/aom_dsp/x86/fft_avx2.c b/third_party/aom/aom_dsp/x86/fft_avx2.c new file mode 100644 index 0000000000..54da022538 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/fft_avx2.c @@ -0,0 +1,73 @@ +/* + * Copyright (c) 2018, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include + +#include "aom_dsp/aom_dsp_common.h" +#include "aom_dsp/fft_common.h" + +extern void aom_transpose_float_sse2(const float *A, float *B, int n); +extern void aom_fft_unpack_2d_output_sse2(const float *col_fft, float *output, + int n); + +// Generate the 1d forward transforms for float using _mm256 +GEN_FFT_8(static INLINE void, avx2, float, __m256, _mm256_load_ps, + _mm256_store_ps, _mm256_set1_ps, _mm256_add_ps, _mm256_sub_ps, + _mm256_mul_ps); +GEN_FFT_16(static INLINE void, avx2, float, __m256, _mm256_load_ps, + _mm256_store_ps, _mm256_set1_ps, _mm256_add_ps, _mm256_sub_ps, + _mm256_mul_ps); +GEN_FFT_32(static INLINE void, avx2, float, __m256, _mm256_load_ps, + _mm256_store_ps, _mm256_set1_ps, _mm256_add_ps, _mm256_sub_ps, + _mm256_mul_ps); + +void aom_fft8x8_float_avx2(const float *input, float *temp, float *output) { + aom_fft_2d_gen(input, temp, output, 8, aom_fft1d_8_avx2, + aom_transpose_float_sse2, aom_fft_unpack_2d_output_sse2, 8); +} + +void aom_fft16x16_float_avx2(const float *input, float *temp, float *output) { + aom_fft_2d_gen(input, temp, output, 16, aom_fft1d_16_avx2, + aom_transpose_float_sse2, aom_fft_unpack_2d_output_sse2, 8); +} + +void aom_fft32x32_float_avx2(const float *input, float *temp, float *output) { + aom_fft_2d_gen(input, temp, output, 32, aom_fft1d_32_avx2, + aom_transpose_float_sse2, aom_fft_unpack_2d_output_sse2, 8); +} + +// Generate the 1d inverse transforms for float using _mm256 +GEN_IFFT_8(static INLINE void, avx2, float, __m256, _mm256_load_ps, + _mm256_store_ps, _mm256_set1_ps, _mm256_add_ps, _mm256_sub_ps, + _mm256_mul_ps); +GEN_IFFT_16(static INLINE void, avx2, float, __m256, _mm256_load_ps, + _mm256_store_ps, _mm256_set1_ps, _mm256_add_ps, _mm256_sub_ps, + _mm256_mul_ps); +GEN_IFFT_32(static INLINE void, avx2, float, __m256, _mm256_load_ps, + _mm256_store_ps, _mm256_set1_ps, _mm256_add_ps, _mm256_sub_ps, + _mm256_mul_ps); + +void aom_ifft8x8_float_avx2(const float *input, float *temp, float *output) { + aom_ifft_2d_gen(input, temp, output, 8, aom_fft1d_8_float, aom_fft1d_8_avx2, + aom_ifft1d_8_avx2, aom_transpose_float_sse2, 8); +} + +void aom_ifft16x16_float_avx2(const float *input, float *temp, float *output) { + aom_ifft_2d_gen(input, temp, output, 16, aom_fft1d_16_float, + aom_fft1d_16_avx2, aom_ifft1d_16_avx2, + aom_transpose_float_sse2, 8); +} + +void aom_ifft32x32_float_avx2(const float *input, float *temp, float *output) { + aom_ifft_2d_gen(input, temp, output, 32, aom_fft1d_32_float, + aom_fft1d_32_avx2, aom_ifft1d_32_avx2, + aom_transpose_float_sse2, 8); +} diff --git a/third_party/aom/aom_dsp/x86/fft_sse2.c b/third_party/aom/aom_dsp/x86/fft_sse2.c new file mode 100644 index 0000000000..12bdc3e185 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/fft_sse2.c @@ -0,0 +1,166 @@ +/* + * Copyright (c) 2018, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the +s * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include + +#include "aom_dsp/aom_dsp_common.h" +#include "aom_dsp/fft_common.h" + +static INLINE void transpose4x4(const float *A, float *B, const int lda, + const int ldb) { + __m128 row1 = _mm_load_ps(&A[0 * lda]); + __m128 row2 = _mm_load_ps(&A[1 * lda]); + __m128 row3 = _mm_load_ps(&A[2 * lda]); + __m128 row4 = _mm_load_ps(&A[3 * lda]); + _MM_TRANSPOSE4_PS(row1, row2, row3, row4); + _mm_store_ps(&B[0 * ldb], row1); + _mm_store_ps(&B[1 * ldb], row2); + _mm_store_ps(&B[2 * ldb], row3); + _mm_store_ps(&B[3 * ldb], row4); +} + +void aom_transpose_float_sse2(const float *A, float *B, int n) { + for (int y = 0; y < n; y += 4) { + for (int x = 0; x < n; x += 4) { + transpose4x4(A + y * n + x, B + x * n + y, n, n); + } + } +} + +void aom_fft_unpack_2d_output_sse2(const float *packed, float *output, int n) { + const int n2 = n / 2; + output[0] = packed[0]; + output[1] = 0; + output[2 * (n2 * n)] = packed[n2 * n]; + output[2 * (n2 * n) + 1] = 0; + + output[2 * n2] = packed[n2]; + output[2 * n2 + 1] = 0; + output[2 * (n2 * n + n2)] = packed[n2 * n + n2]; + output[2 * (n2 * n + n2) + 1] = 0; + + for (int c = 1; c < n2; ++c) { + output[2 * (0 * n + c)] = packed[c]; + output[2 * (0 * n + c) + 1] = packed[c + n2]; + output[2 * (n2 * n + c) + 0] = packed[n2 * n + c]; + output[2 * (n2 * n + c) + 1] = packed[n2 * n + c + n2]; + } + for (int r = 1; r < n2; ++r) { + output[2 * (r * n + 0)] = packed[r * n]; + output[2 * (r * n + 0) + 1] = packed[(r + n2) * n]; + output[2 * (r * n + n2) + 0] = packed[r * n + n2]; + output[2 * (r * n + n2) + 1] = packed[(r + n2) * n + n2]; + + for (int c = 1; c < AOMMIN(n2, 4); ++c) { + output[2 * (r * n + c)] = + packed[r * n + c] - packed[(r + n2) * n + c + n2]; + output[2 * (r * n + c) + 1] = + packed[(r + n2) * n + c] + packed[r * n + c + n2]; + } + + for (int c = 4; c < n2; c += 4) { + __m128 real1 = _mm_load_ps(packed + r * n + c); + __m128 real2 = _mm_load_ps(packed + (r + n2) * n + c + n2); + __m128 imag1 = _mm_load_ps(packed + (r + n2) * n + c); + __m128 imag2 = _mm_load_ps(packed + r * n + c + n2); + real1 = _mm_sub_ps(real1, real2); + imag1 = _mm_add_ps(imag1, imag2); + _mm_store_ps(output + 2 * (r * n + c), _mm_unpacklo_ps(real1, imag1)); + _mm_store_ps(output + 2 * (r * n + c + 2), _mm_unpackhi_ps(real1, imag1)); + } + + int r2 = r + n2; + int r3 = n - r2; + output[2 * (r2 * n + 0)] = packed[r3 * n]; + output[2 * (r2 * n + 0) + 1] = -packed[(r3 + n2) * n]; + output[2 * (r2 * n + n2)] = packed[r3 * n + n2]; + output[2 * (r2 * n + n2) + 1] = -packed[(r3 + n2) * n + n2]; + for (int c = 1; c < AOMMIN(4, n2); ++c) { + output[2 * (r2 * n + c)] = + packed[r3 * n + c] + packed[(r3 + n2) * n + c + n2]; + output[2 * (r2 * n + c) + 1] = + -packed[(r3 + n2) * n + c] + packed[r3 * n + c + n2]; + } + for (int c = 4; c < n2; c += 4) { + __m128 real1 = _mm_load_ps(packed + r3 * n + c); + __m128 real2 = _mm_load_ps(packed + (r3 + n2) * n + c + n2); + __m128 imag1 = _mm_load_ps(packed + (r3 + n2) * n + c); + __m128 imag2 = _mm_load_ps(packed + r3 * n + c + n2); + real1 = _mm_add_ps(real1, real2); + imag1 = _mm_sub_ps(imag2, imag1); + _mm_store_ps(output + 2 * (r2 * n + c), _mm_unpacklo_ps(real1, imag1)); + _mm_store_ps(output + 2 * (r2 * n + c + 2), + _mm_unpackhi_ps(real1, imag1)); + } + } +} + +// Generate definitions for 1d transforms using float and __mm128 +GEN_FFT_4(static INLINE void, sse2, float, __m128, _mm_load_ps, _mm_store_ps, + _mm_set1_ps, _mm_add_ps, _mm_sub_ps); +GEN_FFT_8(static INLINE void, sse2, float, __m128, _mm_load_ps, _mm_store_ps, + _mm_set1_ps, _mm_add_ps, _mm_sub_ps, _mm_mul_ps); +GEN_FFT_16(static INLINE void, sse2, float, __m128, _mm_load_ps, _mm_store_ps, + _mm_set1_ps, _mm_add_ps, _mm_sub_ps, _mm_mul_ps); +GEN_FFT_32(static INLINE void, sse2, float, __m128, _mm_load_ps, _mm_store_ps, + _mm_set1_ps, _mm_add_ps, _mm_sub_ps, _mm_mul_ps); + +void aom_fft4x4_float_sse2(const float *input, float *temp, float *output) { + aom_fft_2d_gen(input, temp, output, 4, aom_fft1d_4_sse2, + aom_transpose_float_sse2, aom_fft_unpack_2d_output_sse2, 4); +} + +void aom_fft8x8_float_sse2(const float *input, float *temp, float *output) { + aom_fft_2d_gen(input, temp, output, 8, aom_fft1d_8_sse2, + aom_transpose_float_sse2, aom_fft_unpack_2d_output_sse2, 4); +} + +void aom_fft16x16_float_sse2(const float *input, float *temp, float *output) { + aom_fft_2d_gen(input, temp, output, 16, aom_fft1d_16_sse2, + aom_transpose_float_sse2, aom_fft_unpack_2d_output_sse2, 4); +} + +void aom_fft32x32_float_sse2(const float *input, float *temp, float *output) { + aom_fft_2d_gen(input, temp, output, 32, aom_fft1d_32_sse2, + aom_transpose_float_sse2, aom_fft_unpack_2d_output_sse2, 4); +} + +// Generate definitions for 1d inverse transforms using float and mm128 +GEN_IFFT_4(static INLINE void, sse2, float, __m128, _mm_load_ps, _mm_store_ps, + _mm_set1_ps, _mm_add_ps, _mm_sub_ps); +GEN_IFFT_8(static INLINE void, sse2, float, __m128, _mm_load_ps, _mm_store_ps, + _mm_set1_ps, _mm_add_ps, _mm_sub_ps, _mm_mul_ps); +GEN_IFFT_16(static INLINE void, sse2, float, __m128, _mm_load_ps, _mm_store_ps, + _mm_set1_ps, _mm_add_ps, _mm_sub_ps, _mm_mul_ps); +GEN_IFFT_32(static INLINE void, sse2, float, __m128, _mm_load_ps, _mm_store_ps, + _mm_set1_ps, _mm_add_ps, _mm_sub_ps, _mm_mul_ps); + +void aom_ifft4x4_float_sse2(const float *input, float *temp, float *output) { + aom_ifft_2d_gen(input, temp, output, 4, aom_fft1d_4_float, aom_fft1d_4_sse2, + aom_ifft1d_4_sse2, aom_transpose_float_sse2, 4); +} + +void aom_ifft8x8_float_sse2(const float *input, float *temp, float *output) { + aom_ifft_2d_gen(input, temp, output, 8, aom_fft1d_8_float, aom_fft1d_8_sse2, + aom_ifft1d_8_sse2, aom_transpose_float_sse2, 4); +} + +void aom_ifft16x16_float_sse2(const float *input, float *temp, float *output) { + aom_ifft_2d_gen(input, temp, output, 16, aom_fft1d_16_float, + aom_fft1d_16_sse2, aom_ifft1d_16_sse2, + aom_transpose_float_sse2, 4); +} + +void aom_ifft32x32_float_sse2(const float *input, float *temp, float *output) { + aom_ifft_2d_gen(input, temp, output, 32, aom_fft1d_32_float, + aom_fft1d_32_sse2, aom_ifft1d_32_sse2, + aom_transpose_float_sse2, 4); +} diff --git a/third_party/aom/aom_dsp/x86/fwd_txfm_impl_sse2.h b/third_party/aom/aom_dsp/x86/fwd_txfm_impl_sse2.h new file mode 100644 index 0000000000..1e3d13ec85 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/fwd_txfm_impl_sse2.h @@ -0,0 +1,344 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include // SSE2 + +#include "config/aom_dsp_rtcd.h" + +#include "aom_dsp/txfm_common.h" +#include "aom_dsp/x86/fwd_txfm_sse2.h" +#include "aom_dsp/x86/txfm_common_sse2.h" +#include "aom_ports/mem.h" + +// TODO(jingning) The high bit-depth functions need rework for performance. +// After we properly fix the high bit-depth function implementations, this +// file's dependency should be substantially simplified. +#if DCT_HIGH_BIT_DEPTH +#define ADD_EPI16 _mm_adds_epi16 +#define SUB_EPI16 _mm_subs_epi16 + +#else +#define ADD_EPI16 _mm_add_epi16 +#define SUB_EPI16 _mm_sub_epi16 +#endif + +void FDCT8x8_2D(const int16_t *input, tran_low_t *output, int stride) { + int pass; + // Constants + // When we use them, in one case, they are all the same. In all others + // it's a pair of them that we need to repeat four times. This is done + // by constructing the 32 bit constant corresponding to that pair. + const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64); + const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64); + const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64); + const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64); + const __m128i k__cospi_p28_p04 = pair_set_epi16(cospi_28_64, cospi_4_64); + const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64, cospi_28_64); + const __m128i k__cospi_p12_p20 = pair_set_epi16(cospi_12_64, cospi_20_64); + const __m128i k__cospi_m20_p12 = pair_set_epi16(-cospi_20_64, cospi_12_64); + const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING); +#if DCT_HIGH_BIT_DEPTH + int overflow; +#endif + // Load input + __m128i in0 = _mm_load_si128((const __m128i *)(input + 0 * stride)); + __m128i in1 = _mm_load_si128((const __m128i *)(input + 1 * stride)); + __m128i in2 = _mm_load_si128((const __m128i *)(input + 2 * stride)); + __m128i in3 = _mm_load_si128((const __m128i *)(input + 3 * stride)); + __m128i in4 = _mm_load_si128((const __m128i *)(input + 4 * stride)); + __m128i in5 = _mm_load_si128((const __m128i *)(input + 5 * stride)); + __m128i in6 = _mm_load_si128((const __m128i *)(input + 6 * stride)); + __m128i in7 = _mm_load_si128((const __m128i *)(input + 7 * stride)); + // Pre-condition input (shift by two) + in0 = _mm_slli_epi16(in0, 2); + in1 = _mm_slli_epi16(in1, 2); + in2 = _mm_slli_epi16(in2, 2); + in3 = _mm_slli_epi16(in3, 2); + in4 = _mm_slli_epi16(in4, 2); + in5 = _mm_slli_epi16(in5, 2); + in6 = _mm_slli_epi16(in6, 2); + in7 = _mm_slli_epi16(in7, 2); + + // We do two passes, first the columns, then the rows. The results of the + // first pass are transposed so that the same column code can be reused. The + // results of the second pass are also transposed so that the rows (processed + // as columns) are put back in row positions. + for (pass = 0; pass < 2; pass++) { + // To store results of each pass before the transpose. + __m128i res0, res1, res2, res3, res4, res5, res6, res7; + // Add/subtract + const __m128i q0 = ADD_EPI16(in0, in7); + const __m128i q1 = ADD_EPI16(in1, in6); + const __m128i q2 = ADD_EPI16(in2, in5); + const __m128i q3 = ADD_EPI16(in3, in4); + const __m128i q4 = SUB_EPI16(in3, in4); + const __m128i q5 = SUB_EPI16(in2, in5); + const __m128i q6 = SUB_EPI16(in1, in6); + const __m128i q7 = SUB_EPI16(in0, in7); +#if DCT_HIGH_BIT_DEPTH + if (pass == 1) { + overflow = + check_epi16_overflow_x8(&q0, &q1, &q2, &q3, &q4, &q5, &q6, &q7); + if (overflow) { + aom_highbd_fdct8x8_c(input, output, stride); + return; + } + } +#endif // DCT_HIGH_BIT_DEPTH + // Work on first four results + { + // Add/subtract + const __m128i r0 = ADD_EPI16(q0, q3); + const __m128i r1 = ADD_EPI16(q1, q2); + const __m128i r2 = SUB_EPI16(q1, q2); + const __m128i r3 = SUB_EPI16(q0, q3); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x4(&r0, &r1, &r2, &r3); + if (overflow) { + aom_highbd_fdct8x8_c(input, output, stride); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + // Interleave to do the multiply by constants which gets us into 32bits + { + const __m128i t0 = _mm_unpacklo_epi16(r0, r1); + const __m128i t1 = _mm_unpackhi_epi16(r0, r1); + const __m128i t2 = _mm_unpacklo_epi16(r2, r3); + const __m128i t3 = _mm_unpackhi_epi16(r2, r3); + const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p16_p16); + const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p16_p16); + const __m128i u2 = _mm_madd_epi16(t0, k__cospi_p16_m16); + const __m128i u3 = _mm_madd_epi16(t1, k__cospi_p16_m16); + const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p24_p08); + const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p24_p08); + const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m08_p24); + const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m08_p24); + // dct_const_round_shift + const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING); + const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING); + const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING); + const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING); + const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING); + const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING); + const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING); + const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING); + const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS); + const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS); + const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS); + const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS); + const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS); + const __m128i w5 = _mm_srai_epi32(v5, DCT_CONST_BITS); + const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS); + const __m128i w7 = _mm_srai_epi32(v7, DCT_CONST_BITS); + // Combine + res0 = _mm_packs_epi32(w0, w1); + res4 = _mm_packs_epi32(w2, w3); + res2 = _mm_packs_epi32(w4, w5); + res6 = _mm_packs_epi32(w6, w7); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x4(&res0, &res4, &res2, &res6); + if (overflow) { + aom_highbd_fdct8x8_c(input, output, stride); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + } + } + // Work on next four results + { + // Interleave to do the multiply by constants which gets us into 32bits + const __m128i d0 = _mm_unpacklo_epi16(q6, q5); + const __m128i d1 = _mm_unpackhi_epi16(q6, q5); + const __m128i e0 = _mm_madd_epi16(d0, k__cospi_p16_m16); + const __m128i e1 = _mm_madd_epi16(d1, k__cospi_p16_m16); + const __m128i e2 = _mm_madd_epi16(d0, k__cospi_p16_p16); + const __m128i e3 = _mm_madd_epi16(d1, k__cospi_p16_p16); + // dct_const_round_shift + const __m128i f0 = _mm_add_epi32(e0, k__DCT_CONST_ROUNDING); + const __m128i f1 = _mm_add_epi32(e1, k__DCT_CONST_ROUNDING); + const __m128i f2 = _mm_add_epi32(e2, k__DCT_CONST_ROUNDING); + const __m128i f3 = _mm_add_epi32(e3, k__DCT_CONST_ROUNDING); + const __m128i s0 = _mm_srai_epi32(f0, DCT_CONST_BITS); + const __m128i s1 = _mm_srai_epi32(f1, DCT_CONST_BITS); + const __m128i s2 = _mm_srai_epi32(f2, DCT_CONST_BITS); + const __m128i s3 = _mm_srai_epi32(f3, DCT_CONST_BITS); + // Combine + const __m128i r0 = _mm_packs_epi32(s0, s1); + const __m128i r1 = _mm_packs_epi32(s2, s3); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x2(&r0, &r1); + if (overflow) { + aom_highbd_fdct8x8_c(input, output, stride); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + { + // Add/subtract + const __m128i x0 = ADD_EPI16(q4, r0); + const __m128i x1 = SUB_EPI16(q4, r0); + const __m128i x2 = SUB_EPI16(q7, r1); + const __m128i x3 = ADD_EPI16(q7, r1); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x4(&x0, &x1, &x2, &x3); + if (overflow) { + aom_highbd_fdct8x8_c(input, output, stride); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + // Interleave to do the multiply by constants which gets us into 32bits + { + const __m128i t0 = _mm_unpacklo_epi16(x0, x3); + const __m128i t1 = _mm_unpackhi_epi16(x0, x3); + const __m128i t2 = _mm_unpacklo_epi16(x1, x2); + const __m128i t3 = _mm_unpackhi_epi16(x1, x2); + const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p28_p04); + const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p28_p04); + const __m128i u2 = _mm_madd_epi16(t0, k__cospi_m04_p28); + const __m128i u3 = _mm_madd_epi16(t1, k__cospi_m04_p28); + const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p12_p20); + const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p12_p20); + const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m20_p12); + const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m20_p12); + // dct_const_round_shift + const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING); + const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING); + const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING); + const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING); + const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING); + const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING); + const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING); + const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING); + const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS); + const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS); + const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS); + const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS); + const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS); + const __m128i w5 = _mm_srai_epi32(v5, DCT_CONST_BITS); + const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS); + const __m128i w7 = _mm_srai_epi32(v7, DCT_CONST_BITS); + // Combine + res1 = _mm_packs_epi32(w0, w1); + res7 = _mm_packs_epi32(w2, w3); + res5 = _mm_packs_epi32(w4, w5); + res3 = _mm_packs_epi32(w6, w7); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x4(&res1, &res7, &res5, &res3); + if (overflow) { + aom_highbd_fdct8x8_c(input, output, stride); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + } + } + } + // Transpose the 8x8. + { + // 00 01 02 03 04 05 06 07 + // 10 11 12 13 14 15 16 17 + // 20 21 22 23 24 25 26 27 + // 30 31 32 33 34 35 36 37 + // 40 41 42 43 44 45 46 47 + // 50 51 52 53 54 55 56 57 + // 60 61 62 63 64 65 66 67 + // 70 71 72 73 74 75 76 77 + const __m128i tr0_0 = _mm_unpacklo_epi16(res0, res1); + const __m128i tr0_1 = _mm_unpacklo_epi16(res2, res3); + const __m128i tr0_2 = _mm_unpackhi_epi16(res0, res1); + const __m128i tr0_3 = _mm_unpackhi_epi16(res2, res3); + const __m128i tr0_4 = _mm_unpacklo_epi16(res4, res5); + const __m128i tr0_5 = _mm_unpacklo_epi16(res6, res7); + const __m128i tr0_6 = _mm_unpackhi_epi16(res4, res5); + const __m128i tr0_7 = _mm_unpackhi_epi16(res6, res7); + // 00 10 01 11 02 12 03 13 + // 20 30 21 31 22 32 23 33 + // 04 14 05 15 06 16 07 17 + // 24 34 25 35 26 36 27 37 + // 40 50 41 51 42 52 43 53 + // 60 70 61 71 62 72 63 73 + // 54 54 55 55 56 56 57 57 + // 64 74 65 75 66 76 67 77 + const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1); + const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_2, tr0_3); + const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1); + const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_2, tr0_3); + const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5); + const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7); + const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5); + const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7); + // 00 10 20 30 01 11 21 31 + // 40 50 60 70 41 51 61 71 + // 02 12 22 32 03 13 23 33 + // 42 52 62 72 43 53 63 73 + // 04 14 24 34 05 15 21 36 + // 44 54 64 74 45 55 61 76 + // 06 16 26 36 07 17 27 37 + // 46 56 66 76 47 57 67 77 + in0 = _mm_unpacklo_epi64(tr1_0, tr1_4); + in1 = _mm_unpackhi_epi64(tr1_0, tr1_4); + in2 = _mm_unpacklo_epi64(tr1_2, tr1_6); + in3 = _mm_unpackhi_epi64(tr1_2, tr1_6); + in4 = _mm_unpacklo_epi64(tr1_1, tr1_5); + in5 = _mm_unpackhi_epi64(tr1_1, tr1_5); + in6 = _mm_unpacklo_epi64(tr1_3, tr1_7); + in7 = _mm_unpackhi_epi64(tr1_3, tr1_7); + // 00 10 20 30 40 50 60 70 + // 01 11 21 31 41 51 61 71 + // 02 12 22 32 42 52 62 72 + // 03 13 23 33 43 53 63 73 + // 04 14 24 34 44 54 64 74 + // 05 15 25 35 45 55 65 75 + // 06 16 26 36 46 56 66 76 + // 07 17 27 37 47 57 67 77 + } + } + // Post-condition output and store it + { + // Post-condition (division by two) + // division of two 16 bits signed numbers using shifts + // n / 2 = (n - (n >> 15)) >> 1 + const __m128i sign_in0 = _mm_srai_epi16(in0, 15); + const __m128i sign_in1 = _mm_srai_epi16(in1, 15); + const __m128i sign_in2 = _mm_srai_epi16(in2, 15); + const __m128i sign_in3 = _mm_srai_epi16(in3, 15); + const __m128i sign_in4 = _mm_srai_epi16(in4, 15); + const __m128i sign_in5 = _mm_srai_epi16(in5, 15); + const __m128i sign_in6 = _mm_srai_epi16(in6, 15); + const __m128i sign_in7 = _mm_srai_epi16(in7, 15); + in0 = _mm_sub_epi16(in0, sign_in0); + in1 = _mm_sub_epi16(in1, sign_in1); + in2 = _mm_sub_epi16(in2, sign_in2); + in3 = _mm_sub_epi16(in3, sign_in3); + in4 = _mm_sub_epi16(in4, sign_in4); + in5 = _mm_sub_epi16(in5, sign_in5); + in6 = _mm_sub_epi16(in6, sign_in6); + in7 = _mm_sub_epi16(in7, sign_in7); + in0 = _mm_srai_epi16(in0, 1); + in1 = _mm_srai_epi16(in1, 1); + in2 = _mm_srai_epi16(in2, 1); + in3 = _mm_srai_epi16(in3, 1); + in4 = _mm_srai_epi16(in4, 1); + in5 = _mm_srai_epi16(in5, 1); + in6 = _mm_srai_epi16(in6, 1); + in7 = _mm_srai_epi16(in7, 1); + // store results + store_output(&in0, (output + 0 * 8)); + store_output(&in1, (output + 1 * 8)); + store_output(&in2, (output + 2 * 8)); + store_output(&in3, (output + 3 * 8)); + store_output(&in4, (output + 4 * 8)); + store_output(&in5, (output + 5 * 8)); + store_output(&in6, (output + 6 * 8)); + store_output(&in7, (output + 7 * 8)); + } +} + +#undef ADD_EPI16 +#undef SUB_EPI16 diff --git a/third_party/aom/aom_dsp/x86/fwd_txfm_sse2.c b/third_party/aom/aom_dsp/x86/fwd_txfm_sse2.c new file mode 100644 index 0000000000..2d8f8f71e4 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/fwd_txfm_sse2.c @@ -0,0 +1,69 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include // SSE2 + +#include "config/aom_config.h" +#include "config/aom_dsp_rtcd.h" + +#include "aom_dsp/aom_dsp_common.h" +#include "aom_dsp/x86/fwd_txfm_sse2.h" + +void aom_fdct8x8_1_sse2(const int16_t *input, tran_low_t *output, int stride) { + __m128i in0 = _mm_load_si128((const __m128i *)(input + 0 * stride)); + __m128i in1 = _mm_load_si128((const __m128i *)(input + 1 * stride)); + __m128i in2 = _mm_load_si128((const __m128i *)(input + 2 * stride)); + __m128i in3 = _mm_load_si128((const __m128i *)(input + 3 * stride)); + __m128i u0, u1, sum; + + u0 = _mm_add_epi16(in0, in1); + u1 = _mm_add_epi16(in2, in3); + + in0 = _mm_load_si128((const __m128i *)(input + 4 * stride)); + in1 = _mm_load_si128((const __m128i *)(input + 5 * stride)); + in2 = _mm_load_si128((const __m128i *)(input + 6 * stride)); + in3 = _mm_load_si128((const __m128i *)(input + 7 * stride)); + + sum = _mm_add_epi16(u0, u1); + + in0 = _mm_add_epi16(in0, in1); + in2 = _mm_add_epi16(in2, in3); + sum = _mm_add_epi16(sum, in0); + + u0 = _mm_setzero_si128(); + sum = _mm_add_epi16(sum, in2); + + in0 = _mm_unpacklo_epi16(u0, sum); + in1 = _mm_unpackhi_epi16(u0, sum); + in0 = _mm_srai_epi32(in0, 16); + in1 = _mm_srai_epi32(in1, 16); + + sum = _mm_add_epi32(in0, in1); + in0 = _mm_unpacklo_epi32(sum, u0); + in1 = _mm_unpackhi_epi32(sum, u0); + + sum = _mm_add_epi32(in0, in1); + in0 = _mm_srli_si128(sum, 8); + + in1 = _mm_add_epi32(sum, in0); + output[0] = (tran_low_t)_mm_cvtsi128_si32(in1); +} + +#define DCT_HIGH_BIT_DEPTH 0 +#define FDCT8x8_2D aom_fdct8x8_sse2 +#include "aom_dsp/x86/fwd_txfm_impl_sse2.h" +#undef FDCT8x8_2D + +#undef DCT_HIGH_BIT_DEPTH +#define DCT_HIGH_BIT_DEPTH 1 +#define FDCT8x8_2D aom_highbd_fdct8x8_sse2 +#include "aom_dsp/x86/fwd_txfm_impl_sse2.h" // NOLINT +#undef FDCT8x8_2D diff --git a/third_party/aom/aom_dsp/x86/fwd_txfm_sse2.h b/third_party/aom/aom_dsp/x86/fwd_txfm_sse2.h new file mode 100644 index 0000000000..260d8dd58e --- /dev/null +++ b/third_party/aom/aom_dsp/x86/fwd_txfm_sse2.h @@ -0,0 +1,155 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_X86_FWD_TXFM_SSE2_H_ +#define AOM_AOM_DSP_X86_FWD_TXFM_SSE2_H_ + +#ifdef __cplusplus +extern "C" { +#endif + +static INLINE __m128i k_madd_epi32(__m128i a, __m128i b) { + __m128i buf0, buf1; + buf0 = _mm_mul_epu32(a, b); + a = _mm_srli_epi64(a, 32); + b = _mm_srli_epi64(b, 32); + buf1 = _mm_mul_epu32(a, b); + return _mm_add_epi64(buf0, buf1); +} + +static INLINE __m128i k_packs_epi64(__m128i a, __m128i b) { + __m128i buf0 = _mm_shuffle_epi32(a, _MM_SHUFFLE(0, 0, 2, 0)); + __m128i buf1 = _mm_shuffle_epi32(b, _MM_SHUFFLE(0, 0, 2, 0)); + return _mm_unpacklo_epi64(buf0, buf1); +} + +static INLINE int check_epi16_overflow_x2(const __m128i *preg0, + const __m128i *preg1) { + const __m128i max_overflow = _mm_set1_epi16(0x7fff); + const __m128i min_overflow = _mm_set1_epi16(0x8000); + __m128i cmp0 = _mm_or_si128(_mm_cmpeq_epi16(*preg0, max_overflow), + _mm_cmpeq_epi16(*preg0, min_overflow)); + __m128i cmp1 = _mm_or_si128(_mm_cmpeq_epi16(*preg1, max_overflow), + _mm_cmpeq_epi16(*preg1, min_overflow)); + cmp0 = _mm_or_si128(cmp0, cmp1); + return _mm_movemask_epi8(cmp0); +} + +static INLINE int check_epi16_overflow_x4(const __m128i *preg0, + const __m128i *preg1, + const __m128i *preg2, + const __m128i *preg3) { + const __m128i max_overflow = _mm_set1_epi16(0x7fff); + const __m128i min_overflow = _mm_set1_epi16(0x8000); + __m128i cmp0 = _mm_or_si128(_mm_cmpeq_epi16(*preg0, max_overflow), + _mm_cmpeq_epi16(*preg0, min_overflow)); + __m128i cmp1 = _mm_or_si128(_mm_cmpeq_epi16(*preg1, max_overflow), + _mm_cmpeq_epi16(*preg1, min_overflow)); + __m128i cmp2 = _mm_or_si128(_mm_cmpeq_epi16(*preg2, max_overflow), + _mm_cmpeq_epi16(*preg2, min_overflow)); + __m128i cmp3 = _mm_or_si128(_mm_cmpeq_epi16(*preg3, max_overflow), + _mm_cmpeq_epi16(*preg3, min_overflow)); + cmp0 = _mm_or_si128(_mm_or_si128(cmp0, cmp1), _mm_or_si128(cmp2, cmp3)); + return _mm_movemask_epi8(cmp0); +} + +static INLINE int check_epi16_overflow_x8( + const __m128i *preg0, const __m128i *preg1, const __m128i *preg2, + const __m128i *preg3, const __m128i *preg4, const __m128i *preg5, + const __m128i *preg6, const __m128i *preg7) { + int res0, res1; + res0 = check_epi16_overflow_x4(preg0, preg1, preg2, preg3); + res1 = check_epi16_overflow_x4(preg4, preg5, preg6, preg7); + return res0 + res1; +} + +static INLINE int check_epi16_overflow_x12( + const __m128i *preg0, const __m128i *preg1, const __m128i *preg2, + const __m128i *preg3, const __m128i *preg4, const __m128i *preg5, + const __m128i *preg6, const __m128i *preg7, const __m128i *preg8, + const __m128i *preg9, const __m128i *preg10, const __m128i *preg11) { + int res0, res1; + res0 = check_epi16_overflow_x4(preg0, preg1, preg2, preg3); + res1 = check_epi16_overflow_x4(preg4, preg5, preg6, preg7); + if (!res0) res0 = check_epi16_overflow_x4(preg8, preg9, preg10, preg11); + return res0 + res1; +} + +static INLINE int check_epi16_overflow_x16( + const __m128i *preg0, const __m128i *preg1, const __m128i *preg2, + const __m128i *preg3, const __m128i *preg4, const __m128i *preg5, + const __m128i *preg6, const __m128i *preg7, const __m128i *preg8, + const __m128i *preg9, const __m128i *preg10, const __m128i *preg11, + const __m128i *preg12, const __m128i *preg13, const __m128i *preg14, + const __m128i *preg15) { + int res0, res1; + res0 = check_epi16_overflow_x4(preg0, preg1, preg2, preg3); + res1 = check_epi16_overflow_x4(preg4, preg5, preg6, preg7); + if (!res0) { + res0 = check_epi16_overflow_x4(preg8, preg9, preg10, preg11); + if (!res1) res1 = check_epi16_overflow_x4(preg12, preg13, preg14, preg15); + } + return res0 + res1; +} + +static INLINE int check_epi16_overflow_x32( + const __m128i *preg0, const __m128i *preg1, const __m128i *preg2, + const __m128i *preg3, const __m128i *preg4, const __m128i *preg5, + const __m128i *preg6, const __m128i *preg7, const __m128i *preg8, + const __m128i *preg9, const __m128i *preg10, const __m128i *preg11, + const __m128i *preg12, const __m128i *preg13, const __m128i *preg14, + const __m128i *preg15, const __m128i *preg16, const __m128i *preg17, + const __m128i *preg18, const __m128i *preg19, const __m128i *preg20, + const __m128i *preg21, const __m128i *preg22, const __m128i *preg23, + const __m128i *preg24, const __m128i *preg25, const __m128i *preg26, + const __m128i *preg27, const __m128i *preg28, const __m128i *preg29, + const __m128i *preg30, const __m128i *preg31) { + int res0, res1; + res0 = check_epi16_overflow_x4(preg0, preg1, preg2, preg3); + res1 = check_epi16_overflow_x4(preg4, preg5, preg6, preg7); + if (!res0) { + res0 = check_epi16_overflow_x4(preg8, preg9, preg10, preg11); + if (!res1) { + res1 = check_epi16_overflow_x4(preg12, preg13, preg14, preg15); + if (!res0) { + res0 = check_epi16_overflow_x4(preg16, preg17, preg18, preg19); + if (!res1) { + res1 = check_epi16_overflow_x4(preg20, preg21, preg22, preg23); + if (!res0) { + res0 = check_epi16_overflow_x4(preg24, preg25, preg26, preg27); + if (!res1) + res1 = check_epi16_overflow_x4(preg28, preg29, preg30, preg31); + } + } + } + } + } + return res0 + res1; +} + +static INLINE void store_output(const __m128i *poutput, tran_low_t *dst_ptr) { + if (sizeof(tran_low_t) == 4) { + const __m128i zero = _mm_setzero_si128(); + const __m128i sign_bits = _mm_cmplt_epi16(*poutput, zero); + __m128i out0 = _mm_unpacklo_epi16(*poutput, sign_bits); + __m128i out1 = _mm_unpackhi_epi16(*poutput, sign_bits); + _mm_store_si128((__m128i *)(dst_ptr), out0); + _mm_store_si128((__m128i *)(dst_ptr + 4), out1); + } else { + _mm_store_si128((__m128i *)(dst_ptr), *poutput); + } +} + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif // AOM_AOM_DSP_X86_FWD_TXFM_SSE2_H_ diff --git a/third_party/aom/aom_dsp/x86/fwd_txfm_ssse3_x86_64.asm b/third_party/aom/aom_dsp/x86/fwd_txfm_ssse3_x86_64.asm new file mode 100644 index 0000000000..c1fb259a11 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/fwd_txfm_ssse3_x86_64.asm @@ -0,0 +1,379 @@ +; +; Copyright (c) 2016, Alliance for Open Media. All rights reserved +; +; This source code is subject to the terms of the BSD 2 Clause License and +; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License +; was not distributed with this source code in the LICENSE file, you can +; obtain it at www.aomedia.org/license/software. If the Alliance for Open +; Media Patent License 1.0 was not distributed with this source code in the +; PATENTS file, you can obtain it at www.aomedia.org/license/patent. +; + +; + +%include "third_party/x86inc/x86inc.asm" + +SECTION_RODATA + +pw_11585x2: times 8 dw 23170 +pd_8192: times 4 dd 8192 + +%macro TRANSFORM_COEFFS 2 +pw_%1_%2: dw %1, %2, %1, %2, %1, %2, %1, %2 +pw_%2_m%1: dw %2, -%1, %2, -%1, %2, -%1, %2, -%1 +%endmacro + +TRANSFORM_COEFFS 11585, 11585 +TRANSFORM_COEFFS 15137, 6270 +TRANSFORM_COEFFS 16069, 3196 +TRANSFORM_COEFFS 9102, 13623 + +%macro STORE_OUTPUT 2 ; index, result + ; const __m128i sign_bits = _mm_cmplt_epi16(*poutput, zero); + ; __m128i out0 = _mm_unpacklo_epi16(*poutput, sign_bits); + ; __m128i out1 = _mm_unpackhi_epi16(*poutput, sign_bits); + ; _mm_store_si128((__m128i *)(dst_ptr), out0); + ; _mm_store_si128((__m128i *)(dst_ptr + 4), out1); + pxor m11, m11 + pcmpgtw m11, m%2 + movdqa m12, m%2 + punpcklwd m%2, m11 + punpckhwd m12, m11 + mova [outputq + 4*%1 + 0], m%2 + mova [outputq + 4*%1 + 16], m12 +%endmacro + +SECTION .text + +%if ARCH_X86_64 +INIT_XMM ssse3 +cglobal fdct8x8, 3, 5, 13, input, output, stride + + mova m8, [GLOBAL(pd_8192)] + mova m12, [GLOBAL(pw_11585x2)] + + lea r3, [2 * strideq] + lea r4, [4 * strideq] + mova m0, [inputq] + mova m1, [inputq + r3] + lea inputq, [inputq + r4] + mova m2, [inputq] + mova m3, [inputq + r3] + lea inputq, [inputq + r4] + mova m4, [inputq] + mova m5, [inputq + r3] + lea inputq, [inputq + r4] + mova m6, [inputq] + mova m7, [inputq + r3] + + ; left shift by 2 to increase forward transformation precision + psllw m0, 2 + psllw m1, 2 + psllw m2, 2 + psllw m3, 2 + psllw m4, 2 + psllw m5, 2 + psllw m6, 2 + psllw m7, 2 + + ; column transform + ; stage 1 + paddw m10, m0, m7 + psubw m0, m7 + + paddw m9, m1, m6 + psubw m1, m6 + + paddw m7, m2, m5 + psubw m2, m5 + + paddw m6, m3, m4 + psubw m3, m4 + + ; stage 2 + paddw m5, m9, m7 + psubw m9, m7 + + paddw m4, m10, m6 + psubw m10, m6 + + paddw m7, m1, m2 + psubw m1, m2 + + ; stage 3 + paddw m6, m4, m5 + psubw m4, m5 + + pmulhrsw m1, m12 + pmulhrsw m7, m12 + + ; sin(pi / 8), cos(pi / 8) + punpcklwd m2, m10, m9 + punpckhwd m10, m9 + pmaddwd m5, m2, [GLOBAL(pw_15137_6270)] + pmaddwd m2, [GLOBAL(pw_6270_m15137)] + pmaddwd m9, m10, [GLOBAL(pw_15137_6270)] + pmaddwd m10, [GLOBAL(pw_6270_m15137)] + paddd m5, m8 + paddd m2, m8 + paddd m9, m8 + paddd m10, m8 + psrad m5, 14 + psrad m2, 14 + psrad m9, 14 + psrad m10, 14 + packssdw m5, m9 + packssdw m2, m10 + + pmulhrsw m6, m12 + pmulhrsw m4, m12 + + paddw m9, m3, m1 + psubw m3, m1 + + paddw m10, m0, m7 + psubw m0, m7 + + ; stage 4 + ; sin(pi / 16), cos(pi / 16) + punpcklwd m1, m10, m9 + punpckhwd m10, m9 + pmaddwd m7, m1, [GLOBAL(pw_16069_3196)] + pmaddwd m1, [GLOBAL(pw_3196_m16069)] + pmaddwd m9, m10, [GLOBAL(pw_16069_3196)] + pmaddwd m10, [GLOBAL(pw_3196_m16069)] + paddd m7, m8 + paddd m1, m8 + paddd m9, m8 + paddd m10, m8 + psrad m7, 14 + psrad m1, 14 + psrad m9, 14 + psrad m10, 14 + packssdw m7, m9 + packssdw m1, m10 + + ; sin(3 * pi / 16), cos(3 * pi / 16) + punpcklwd m11, m0, m3 + punpckhwd m0, m3 + pmaddwd m9, m11, [GLOBAL(pw_9102_13623)] + pmaddwd m11, [GLOBAL(pw_13623_m9102)] + pmaddwd m3, m0, [GLOBAL(pw_9102_13623)] + pmaddwd m0, [GLOBAL(pw_13623_m9102)] + paddd m9, m8 + paddd m11, m8 + paddd m3, m8 + paddd m0, m8 + psrad m9, 14 + psrad m11, 14 + psrad m3, 14 + psrad m0, 14 + packssdw m9, m3 + packssdw m11, m0 + + ; transpose + ; stage 1 + punpcklwd m0, m6, m7 + punpcklwd m3, m5, m11 + punpckhwd m6, m7 + punpckhwd m5, m11 + punpcklwd m7, m4, m9 + punpcklwd m10, m2, m1 + punpckhwd m4, m9 + punpckhwd m2, m1 + + ; stage 2 + punpckldq m9, m0, m3 + punpckldq m1, m6, m5 + punpckhdq m0, m3 + punpckhdq m6, m5 + punpckldq m3, m7, m10 + punpckldq m5, m4, m2 + punpckhdq m7, m10 + punpckhdq m4, m2 + + ; stage 3 + punpcklqdq m10, m9, m3 + punpckhqdq m9, m3 + punpcklqdq m2, m0, m7 + punpckhqdq m0, m7 + punpcklqdq m3, m1, m5 + punpckhqdq m1, m5 + punpcklqdq m7, m6, m4 + punpckhqdq m6, m4 + + ; row transform + ; stage 1 + paddw m5, m10, m6 + psubw m10, m6 + + paddw m4, m9, m7 + psubw m9, m7 + + paddw m6, m2, m1 + psubw m2, m1 + + paddw m7, m0, m3 + psubw m0, m3 + + ;stage 2 + paddw m1, m5, m7 + psubw m5, m7 + + paddw m3, m4, m6 + psubw m4, m6 + + paddw m7, m9, m2 + psubw m9, m2 + + ; stage 3 + punpcklwd m6, m1, m3 + punpckhwd m1, m3 + pmaddwd m2, m6, [GLOBAL(pw_11585_11585)] + pmaddwd m6, [GLOBAL(pw_11585_m11585)] + pmaddwd m3, m1, [GLOBAL(pw_11585_11585)] + pmaddwd m1, [GLOBAL(pw_11585_m11585)] + paddd m2, m8 + paddd m6, m8 + paddd m3, m8 + paddd m1, m8 + psrad m2, 14 + psrad m6, 14 + psrad m3, 14 + psrad m1, 14 + packssdw m2, m3 + packssdw m6, m1 + + pmulhrsw m7, m12 + pmulhrsw m9, m12 + + punpcklwd m3, m5, m4 + punpckhwd m5, m4 + pmaddwd m1, m3, [GLOBAL(pw_15137_6270)] + pmaddwd m3, [GLOBAL(pw_6270_m15137)] + pmaddwd m4, m5, [GLOBAL(pw_15137_6270)] + pmaddwd m5, [GLOBAL(pw_6270_m15137)] + paddd m1, m8 + paddd m3, m8 + paddd m4, m8 + paddd m5, m8 + psrad m1, 14 + psrad m3, 14 + psrad m4, 14 + psrad m5, 14 + packssdw m1, m4 + packssdw m3, m5 + + paddw m4, m0, m9 + psubw m0, m9 + + paddw m5, m10, m7 + psubw m10, m7 + + ; stage 4 + punpcklwd m9, m5, m4 + punpckhwd m5, m4 + pmaddwd m7, m9, [GLOBAL(pw_16069_3196)] + pmaddwd m9, [GLOBAL(pw_3196_m16069)] + pmaddwd m4, m5, [GLOBAL(pw_16069_3196)] + pmaddwd m5, [GLOBAL(pw_3196_m16069)] + paddd m7, m8 + paddd m9, m8 + paddd m4, m8 + paddd m5, m8 + psrad m7, 14 + psrad m9, 14 + psrad m4, 14 + psrad m5, 14 + packssdw m7, m4 + packssdw m9, m5 + + punpcklwd m4, m10, m0 + punpckhwd m10, m0 + pmaddwd m5, m4, [GLOBAL(pw_9102_13623)] + pmaddwd m4, [GLOBAL(pw_13623_m9102)] + pmaddwd m0, m10, [GLOBAL(pw_9102_13623)] + pmaddwd m10, [GLOBAL(pw_13623_m9102)] + paddd m5, m8 + paddd m4, m8 + paddd m0, m8 + paddd m10, m8 + psrad m5, 14 + psrad m4, 14 + psrad m0, 14 + psrad m10, 14 + packssdw m5, m0 + packssdw m4, m10 + + ; transpose + ; stage 1 + punpcklwd m0, m2, m7 + punpcklwd m10, m1, m4 + punpckhwd m2, m7 + punpckhwd m1, m4 + punpcklwd m7, m6, m5 + punpcklwd m4, m3, m9 + punpckhwd m6, m5 + punpckhwd m3, m9 + + ; stage 2 + punpckldq m5, m0, m10 + punpckldq m9, m2, m1 + punpckhdq m0, m10 + punpckhdq m2, m1 + punpckldq m10, m7, m4 + punpckldq m1, m6, m3 + punpckhdq m7, m4 + punpckhdq m6, m3 + + ; stage 3 + punpcklqdq m4, m5, m10 + punpckhqdq m5, m10 + punpcklqdq m3, m0, m7 + punpckhqdq m0, m7 + punpcklqdq m10, m9, m1 + punpckhqdq m9, m1 + punpcklqdq m7, m2, m6 + punpckhqdq m2, m6 + + psraw m1, m4, 15 + psraw m6, m5, 15 + psraw m8, m3, 15 + psraw m11, m0, 15 + + psubw m4, m1 + psubw m5, m6 + psubw m3, m8 + psubw m0, m11 + + psraw m4, 1 + psraw m5, 1 + psraw m3, 1 + psraw m0, 1 + + psraw m1, m10, 15 + psraw m6, m9, 15 + psraw m8, m7, 15 + psraw m11, m2, 15 + + psubw m10, m1 + psubw m9, m6 + psubw m7, m8 + psubw m2, m11 + + psraw m10, 1 + psraw m9, 1 + psraw m7, 1 + psraw m2, 1 + + STORE_OUTPUT 0, 4 + STORE_OUTPUT 8, 5 + STORE_OUTPUT 16, 3 + STORE_OUTPUT 24, 0 + STORE_OUTPUT 32, 10 + STORE_OUTPUT 40, 9 + STORE_OUTPUT 48, 7 + STORE_OUTPUT 56, 2 + + RET +%endif diff --git a/third_party/aom/aom_dsp/x86/highbd_convolve_avx2.c b/third_party/aom/aom_dsp/x86/highbd_convolve_avx2.c new file mode 100644 index 0000000000..099fcf7fc6 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/highbd_convolve_avx2.c @@ -0,0 +1,998 @@ +/* + * Copyright (c) 2017, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ +#include +#include + +#include "config/aom_dsp_rtcd.h" + +#include "aom_dsp/x86/convolve.h" +#include "aom_dsp/x86/convolve_avx2.h" +#include "aom_dsp/x86/synonyms.h" + +// ----------------------------------------------------------------------------- +// Copy and average + +void aom_highbd_convolve_copy_avx2(const uint8_t *src8, ptrdiff_t src_stride, + uint8_t *dst8, ptrdiff_t dst_stride, + const int16_t *filter_x, int filter_x_stride, + const int16_t *filter_y, int filter_y_stride, + int width, int h, int bd) { + const uint16_t *src = CONVERT_TO_SHORTPTR(src8); + uint16_t *dst = CONVERT_TO_SHORTPTR(dst8); + (void)filter_x; + (void)filter_y; + (void)filter_x_stride; + (void)filter_y_stride; + (void)bd; + + assert(width % 4 == 0); + if (width > 32) { // width = 64 + do { + const __m256i p0 = _mm256_loadu_si256((const __m256i *)src); + const __m256i p1 = _mm256_loadu_si256((const __m256i *)(src + 16)); + const __m256i p2 = _mm256_loadu_si256((const __m256i *)(src + 32)); + const __m256i p3 = _mm256_loadu_si256((const __m256i *)(src + 48)); + src += src_stride; + _mm256_storeu_si256((__m256i *)dst, p0); + _mm256_storeu_si256((__m256i *)(dst + 16), p1); + _mm256_storeu_si256((__m256i *)(dst + 32), p2); + _mm256_storeu_si256((__m256i *)(dst + 48), p3); + dst += dst_stride; + h--; + } while (h > 0); + } else if (width > 16) { // width = 32 + do { + const __m256i p0 = _mm256_loadu_si256((const __m256i *)src); + const __m256i p1 = _mm256_loadu_si256((const __m256i *)(src + 16)); + src += src_stride; + _mm256_storeu_si256((__m256i *)dst, p0); + _mm256_storeu_si256((__m256i *)(dst + 16), p1); + dst += dst_stride; + h--; + } while (h > 0); + } else if (width > 8) { // width = 16 + __m256i p0, p1; + do { + p0 = _mm256_loadu_si256((const __m256i *)src); + src += src_stride; + p1 = _mm256_loadu_si256((const __m256i *)src); + src += src_stride; + + _mm256_storeu_si256((__m256i *)dst, p0); + dst += dst_stride; + _mm256_storeu_si256((__m256i *)dst, p1); + dst += dst_stride; + h -= 2; + } while (h > 0); + } else if (width > 4) { // width = 8 + __m128i p0, p1; + do { + p0 = _mm_loadu_si128((const __m128i *)src); + src += src_stride; + p1 = _mm_loadu_si128((const __m128i *)src); + src += src_stride; + + _mm_storeu_si128((__m128i *)dst, p0); + dst += dst_stride; + _mm_storeu_si128((__m128i *)dst, p1); + dst += dst_stride; + h -= 2; + } while (h > 0); + } else { // width = 4 + __m128i p0, p1; + do { + p0 = _mm_loadl_epi64((const __m128i *)src); + src += src_stride; + p1 = _mm_loadl_epi64((const __m128i *)src); + src += src_stride; + + _mm_storel_epi64((__m128i *)dst, p0); + dst += dst_stride; + _mm_storel_epi64((__m128i *)dst, p1); + dst += dst_stride; + h -= 2; + } while (h > 0); + } +} + +void av1_highbd_convolve_y_sr_avx2(const uint16_t *src, int src_stride, + uint16_t *dst, int dst_stride, int w, int h, + const InterpFilterParams *filter_params_x, + const InterpFilterParams *filter_params_y, + const int subpel_x_q4, const int subpel_y_q4, + ConvolveParams *conv_params, int bd) { + int i, j; + const int fo_vert = filter_params_y->taps / 2 - 1; + const uint16_t *const src_ptr = src - fo_vert * src_stride; + (void)filter_params_x; + (void)subpel_x_q4; + (void)conv_params; + + assert(conv_params->round_0 <= FILTER_BITS); + assert(((conv_params->round_0 + conv_params->round_1) <= (FILTER_BITS + 1)) || + ((conv_params->round_0 + conv_params->round_1) == (2 * FILTER_BITS))); + + __m256i s[8], coeffs_y[4]; + + const int bits = FILTER_BITS; + + const __m128i round_shift_bits = _mm_cvtsi32_si128(bits); + const __m256i round_const_bits = _mm256_set1_epi32((1 << bits) >> 1); + const __m256i clip_pixel = + _mm256_set1_epi16(bd == 10 ? 1023 : (bd == 12 ? 4095 : 255)); + const __m256i zero = _mm256_setzero_si256(); + + prepare_coeffs(filter_params_y, subpel_y_q4, coeffs_y); + + for (j = 0; j < w; j += 8) { + const uint16_t *data = &src_ptr[j]; + /* Vertical filter */ + { + __m256i src6; + __m256i s01 = _mm256_permute2x128_si256( + _mm256_castsi128_si256( + _mm_loadu_si128((__m128i *)(data + 0 * src_stride))), + _mm256_castsi128_si256( + _mm_loadu_si128((__m128i *)(data + 1 * src_stride))), + 0x20); + __m256i s12 = _mm256_permute2x128_si256( + _mm256_castsi128_si256( + _mm_loadu_si128((__m128i *)(data + 1 * src_stride))), + _mm256_castsi128_si256( + _mm_loadu_si128((__m128i *)(data + 2 * src_stride))), + 0x20); + __m256i s23 = _mm256_permute2x128_si256( + _mm256_castsi128_si256( + _mm_loadu_si128((__m128i *)(data + 2 * src_stride))), + _mm256_castsi128_si256( + _mm_loadu_si128((__m128i *)(data + 3 * src_stride))), + 0x20); + __m256i s34 = _mm256_permute2x128_si256( + _mm256_castsi128_si256( + _mm_loadu_si128((__m128i *)(data + 3 * src_stride))), + _mm256_castsi128_si256( + _mm_loadu_si128((__m128i *)(data + 4 * src_stride))), + 0x20); + __m256i s45 = _mm256_permute2x128_si256( + _mm256_castsi128_si256( + _mm_loadu_si128((__m128i *)(data + 4 * src_stride))), + _mm256_castsi128_si256( + _mm_loadu_si128((__m128i *)(data + 5 * src_stride))), + 0x20); + src6 = _mm256_castsi128_si256( + _mm_loadu_si128((__m128i *)(data + 6 * src_stride))); + __m256i s56 = _mm256_permute2x128_si256( + _mm256_castsi128_si256( + _mm_loadu_si128((__m128i *)(data + 5 * src_stride))), + src6, 0x20); + + s[0] = _mm256_unpacklo_epi16(s01, s12); + s[1] = _mm256_unpacklo_epi16(s23, s34); + s[2] = _mm256_unpacklo_epi16(s45, s56); + + s[4] = _mm256_unpackhi_epi16(s01, s12); + s[5] = _mm256_unpackhi_epi16(s23, s34); + s[6] = _mm256_unpackhi_epi16(s45, s56); + + for (i = 0; i < h; i += 2) { + data = &src_ptr[i * src_stride + j]; + + const __m256i s67 = _mm256_permute2x128_si256( + src6, + _mm256_castsi128_si256( + _mm_loadu_si128((__m128i *)(data + 7 * src_stride))), + 0x20); + + src6 = _mm256_castsi128_si256( + _mm_loadu_si128((__m128i *)(data + 8 * src_stride))); + + const __m256i s78 = _mm256_permute2x128_si256( + _mm256_castsi128_si256( + _mm_loadu_si128((__m128i *)(data + 7 * src_stride))), + src6, 0x20); + + s[3] = _mm256_unpacklo_epi16(s67, s78); + s[7] = _mm256_unpackhi_epi16(s67, s78); + + const __m256i res_a = convolve(s, coeffs_y); + + __m256i res_a_round = _mm256_sra_epi32( + _mm256_add_epi32(res_a, round_const_bits), round_shift_bits); + + if (w - j > 4) { + const __m256i res_b = convolve(s + 4, coeffs_y); + __m256i res_b_round = _mm256_sra_epi32( + _mm256_add_epi32(res_b, round_const_bits), round_shift_bits); + + __m256i res_16bit = _mm256_packs_epi32(res_a_round, res_b_round); + res_16bit = _mm256_min_epi16(res_16bit, clip_pixel); + res_16bit = _mm256_max_epi16(res_16bit, zero); + + _mm_storeu_si128((__m128i *)&dst[i * dst_stride + j], + _mm256_castsi256_si128(res_16bit)); + _mm_storeu_si128((__m128i *)&dst[i * dst_stride + j + dst_stride], + _mm256_extracti128_si256(res_16bit, 1)); + } else if (w == 4) { + res_a_round = _mm256_packs_epi32(res_a_round, res_a_round); + res_a_round = _mm256_min_epi16(res_a_round, clip_pixel); + res_a_round = _mm256_max_epi16(res_a_round, zero); + + _mm_storel_epi64((__m128i *)&dst[i * dst_stride + j], + _mm256_castsi256_si128(res_a_round)); + _mm_storel_epi64((__m128i *)&dst[i * dst_stride + j + dst_stride], + _mm256_extracti128_si256(res_a_round, 1)); + } else { + res_a_round = _mm256_packs_epi32(res_a_round, res_a_round); + res_a_round = _mm256_min_epi16(res_a_round, clip_pixel); + res_a_round = _mm256_max_epi16(res_a_round, zero); + + xx_storel_32((__m128i *)&dst[i * dst_stride + j], + _mm256_castsi256_si128(res_a_round)); + xx_storel_32((__m128i *)&dst[i * dst_stride + j + dst_stride], + _mm256_extracti128_si256(res_a_round, 1)); + } + + s[0] = s[1]; + s[1] = s[2]; + s[2] = s[3]; + + s[4] = s[5]; + s[5] = s[6]; + s[6] = s[7]; + } + } + } +} + +void av1_highbd_convolve_x_sr_avx2(const uint16_t *src, int src_stride, + uint16_t *dst, int dst_stride, int w, int h, + const InterpFilterParams *filter_params_x, + const InterpFilterParams *filter_params_y, + const int subpel_x_q4, const int subpel_y_q4, + ConvolveParams *conv_params, int bd) { + int i, j; + const int fo_horiz = filter_params_x->taps / 2 - 1; + const uint16_t *const src_ptr = src - fo_horiz; + (void)subpel_y_q4; + (void)filter_params_y; + + // Check that, even with 12-bit input, the intermediate values will fit + // into an unsigned 16-bit intermediate array. + assert(bd + FILTER_BITS + 2 - conv_params->round_0 <= 16); + + __m256i s[4], coeffs_x[4]; + + const __m256i round_const_x = + _mm256_set1_epi32(((1 << conv_params->round_0) >> 1)); + const __m128i round_shift_x = _mm_cvtsi32_si128(conv_params->round_0); + + const int bits = FILTER_BITS - conv_params->round_0; + const __m128i round_shift_bits = _mm_cvtsi32_si128(bits); + const __m256i round_const_bits = _mm256_set1_epi32((1 << bits) >> 1); + const __m256i clip_pixel = + _mm256_set1_epi16(bd == 10 ? 1023 : (bd == 12 ? 4095 : 255)); + const __m256i zero = _mm256_setzero_si256(); + + assert(bits >= 0); + assert((FILTER_BITS - conv_params->round_1) >= 0 || + ((conv_params->round_0 + conv_params->round_1) == 2 * FILTER_BITS)); + + prepare_coeffs(filter_params_x, subpel_x_q4, coeffs_x); + + for (j = 0; j < w; j += 8) { + /* Horizontal filter */ + for (i = 0; i < h; i += 2) { + const __m256i row0 = + _mm256_loadu_si256((__m256i *)&src_ptr[i * src_stride + j]); + __m256i row1 = + _mm256_loadu_si256((__m256i *)&src_ptr[(i + 1) * src_stride + j]); + + const __m256i r0 = _mm256_permute2x128_si256(row0, row1, 0x20); + const __m256i r1 = _mm256_permute2x128_si256(row0, row1, 0x31); + + // even pixels + s[0] = _mm256_alignr_epi8(r1, r0, 0); + s[1] = _mm256_alignr_epi8(r1, r0, 4); + s[2] = _mm256_alignr_epi8(r1, r0, 8); + s[3] = _mm256_alignr_epi8(r1, r0, 12); + + __m256i res_even = convolve(s, coeffs_x); + res_even = _mm256_sra_epi32(_mm256_add_epi32(res_even, round_const_x), + round_shift_x); + + // odd pixels + s[0] = _mm256_alignr_epi8(r1, r0, 2); + s[1] = _mm256_alignr_epi8(r1, r0, 6); + s[2] = _mm256_alignr_epi8(r1, r0, 10); + s[3] = _mm256_alignr_epi8(r1, r0, 14); + + __m256i res_odd = convolve(s, coeffs_x); + res_odd = _mm256_sra_epi32(_mm256_add_epi32(res_odd, round_const_x), + round_shift_x); + + res_even = _mm256_sra_epi32(_mm256_add_epi32(res_even, round_const_bits), + round_shift_bits); + res_odd = _mm256_sra_epi32(_mm256_add_epi32(res_odd, round_const_bits), + round_shift_bits); + + __m256i res_even1 = _mm256_packs_epi32(res_even, res_even); + __m256i res_odd1 = _mm256_packs_epi32(res_odd, res_odd); + + __m256i res = _mm256_unpacklo_epi16(res_even1, res_odd1); + res = _mm256_min_epi16(res, clip_pixel); + res = _mm256_max_epi16(res, zero); + + if (w - j > 4) { + _mm_storeu_si128((__m128i *)&dst[i * dst_stride + j], + _mm256_castsi256_si128(res)); + _mm_storeu_si128((__m128i *)&dst[i * dst_stride + j + dst_stride], + _mm256_extracti128_si256(res, 1)); + } else if (w == 4) { + _mm_storel_epi64((__m128i *)&dst[i * dst_stride + j], + _mm256_castsi256_si128(res)); + _mm_storel_epi64((__m128i *)&dst[i * dst_stride + j + dst_stride], + _mm256_extracti128_si256(res, 1)); + } else { + xx_storel_32((__m128i *)&dst[i * dst_stride + j], + _mm256_castsi256_si128(res)); + xx_storel_32((__m128i *)&dst[i * dst_stride + j + dst_stride], + _mm256_extracti128_si256(res, 1)); + } + } + } +} + +#define CONV8_ROUNDING_BITS (7) + +// ----------------------------------------------------------------------------- +// Horizontal and vertical filtering + +static const uint8_t signal_pattern_0[32] = { 0, 1, 2, 3, 2, 3, 4, 5, 4, 5, 6, + 7, 6, 7, 8, 9, 0, 1, 2, 3, 2, 3, + 4, 5, 4, 5, 6, 7, 6, 7, 8, 9 }; + +static const uint8_t signal_pattern_1[32] = { 4, 5, 6, 7, 6, 7, 8, 9, + 8, 9, 10, 11, 10, 11, 12, 13, + 4, 5, 6, 7, 6, 7, 8, 9, + 8, 9, 10, 11, 10, 11, 12, 13 }; + +static const uint8_t signal_pattern_2[32] = { 6, 7, 8, 9, 8, 9, 10, 11, + 10, 11, 12, 13, 12, 13, 14, 15, + 6, 7, 8, 9, 8, 9, 10, 11, + 10, 11, 12, 13, 12, 13, 14, 15 }; + +static const uint32_t signal_index[8] = { 2, 3, 4, 5, 2, 3, 4, 5 }; + +// ----------------------------------------------------------------------------- +// Horizontal Filtering + +static INLINE void pack_pixels(const __m256i *s, __m256i *p /*p[4]*/) { + const __m256i idx = _mm256_loadu_si256((const __m256i *)signal_index); + const __m256i sf0 = _mm256_loadu_si256((const __m256i *)signal_pattern_0); + const __m256i sf1 = _mm256_loadu_si256((const __m256i *)signal_pattern_1); + const __m256i c = _mm256_permutevar8x32_epi32(*s, idx); + + p[0] = _mm256_shuffle_epi8(*s, sf0); // x0x6 + p[1] = _mm256_shuffle_epi8(*s, sf1); // x1x7 + p[2] = _mm256_shuffle_epi8(c, sf0); // x2x4 + p[3] = _mm256_shuffle_epi8(c, sf1); // x3x5 +} + +// Note: +// Shared by 8x2 and 16x1 block +static INLINE void pack_16_pixels(const __m256i *s0, const __m256i *s1, + __m256i *x /*x[8]*/) { + __m256i pp[8]; + pack_pixels(s0, pp); + pack_pixels(s1, &pp[4]); + x[0] = _mm256_permute2x128_si256(pp[0], pp[4], 0x20); + x[1] = _mm256_permute2x128_si256(pp[1], pp[5], 0x20); + x[2] = _mm256_permute2x128_si256(pp[2], pp[6], 0x20); + x[3] = _mm256_permute2x128_si256(pp[3], pp[7], 0x20); + x[4] = x[2]; + x[5] = x[3]; + x[6] = _mm256_permute2x128_si256(pp[0], pp[4], 0x31); + x[7] = _mm256_permute2x128_si256(pp[1], pp[5], 0x31); +} + +static INLINE void pack_8x1_pixels(const uint16_t *src, __m256i *x) { + __m256i pp[8]; + __m256i s0; + s0 = _mm256_loadu_si256((const __m256i *)src); + pack_pixels(&s0, pp); + x[0] = _mm256_permute2x128_si256(pp[0], pp[2], 0x30); + x[1] = _mm256_permute2x128_si256(pp[1], pp[3], 0x30); + x[2] = _mm256_permute2x128_si256(pp[2], pp[0], 0x30); + x[3] = _mm256_permute2x128_si256(pp[3], pp[1], 0x30); +} + +static INLINE void pack_8x2_pixels(const uint16_t *src, ptrdiff_t stride, + __m256i *x) { + __m256i s0, s1; + s0 = _mm256_loadu_si256((const __m256i *)src); + s1 = _mm256_loadu_si256((const __m256i *)(src + stride)); + pack_16_pixels(&s0, &s1, x); +} + +static INLINE void pack_16x1_pixels(const uint16_t *src, __m256i *x) { + __m256i s0, s1; + s0 = _mm256_loadu_si256((const __m256i *)src); + s1 = _mm256_loadu_si256((const __m256i *)(src + 8)); + pack_16_pixels(&s0, &s1, x); +} + +// Note: +// Shared by horizontal and vertical filtering +static INLINE void pack_filters(const int16_t *filter, __m256i *f /*f[4]*/) { + const __m128i h = _mm_loadu_si128((const __m128i *)filter); + const __m256i hh = _mm256_insertf128_si256(_mm256_castsi128_si256(h), h, 1); + const __m256i p0 = _mm256_set1_epi32(0x03020100); + const __m256i p1 = _mm256_set1_epi32(0x07060504); + const __m256i p2 = _mm256_set1_epi32(0x0b0a0908); + const __m256i p3 = _mm256_set1_epi32(0x0f0e0d0c); + f[0] = _mm256_shuffle_epi8(hh, p0); + f[1] = _mm256_shuffle_epi8(hh, p1); + f[2] = _mm256_shuffle_epi8(hh, p2); + f[3] = _mm256_shuffle_epi8(hh, p3); +} + +static INLINE void filter_8x1_pixels(const __m256i *sig /*sig[4]*/, + const __m256i *fil /*fil[4]*/, + __m256i *y) { + __m256i a, a0, a1; + + a0 = _mm256_madd_epi16(fil[0], sig[0]); + a1 = _mm256_madd_epi16(fil[3], sig[3]); + a = _mm256_add_epi32(a0, a1); + + a0 = _mm256_madd_epi16(fil[1], sig[1]); + a1 = _mm256_madd_epi16(fil[2], sig[2]); + + { + const __m256i min = _mm256_min_epi32(a0, a1); + a = _mm256_add_epi32(a, min); + } + { + const __m256i max = _mm256_max_epi32(a0, a1); + a = _mm256_add_epi32(a, max); + } + { + const __m256i rounding = _mm256_set1_epi32(1 << (CONV8_ROUNDING_BITS - 1)); + a = _mm256_add_epi32(a, rounding); + *y = _mm256_srai_epi32(a, CONV8_ROUNDING_BITS); + } +} + +static INLINE void store_8x1_pixels(const __m256i *y, const __m256i *mask, + uint16_t *dst) { + const __m128i a0 = _mm256_castsi256_si128(*y); + const __m128i a1 = _mm256_extractf128_si256(*y, 1); + __m128i res = _mm_packus_epi32(a0, a1); + res = _mm_min_epi16(res, _mm256_castsi256_si128(*mask)); + _mm_storeu_si128((__m128i *)dst, res); +} + +static INLINE void store_8x2_pixels(const __m256i *y0, const __m256i *y1, + const __m256i *mask, uint16_t *dst, + ptrdiff_t pitch) { + __m256i a = _mm256_packus_epi32(*y0, *y1); + a = _mm256_min_epi16(a, *mask); + _mm_storeu_si128((__m128i *)dst, _mm256_castsi256_si128(a)); + _mm_storeu_si128((__m128i *)(dst + pitch), _mm256_extractf128_si256(a, 1)); +} + +static INLINE void store_16x1_pixels(const __m256i *y0, const __m256i *y1, + const __m256i *mask, uint16_t *dst) { + __m256i a = _mm256_packus_epi32(*y0, *y1); + a = _mm256_min_epi16(a, *mask); + _mm256_storeu_si256((__m256i *)dst, a); +} + +static void aom_highbd_filter_block1d8_h8_avx2( + const uint16_t *src_ptr, ptrdiff_t src_pitch, uint16_t *dst_ptr, + ptrdiff_t dst_pitch, uint32_t height, const int16_t *filter, int bd) { + __m256i signal[8], res0, res1; + const __m256i max = _mm256_set1_epi16((1 << bd) - 1); + + __m256i ff[4]; + pack_filters(filter, ff); + + src_ptr -= 3; + do { + pack_8x2_pixels(src_ptr, src_pitch, signal); + filter_8x1_pixels(signal, ff, &res0); + filter_8x1_pixels(&signal[4], ff, &res1); + store_8x2_pixels(&res0, &res1, &max, dst_ptr, dst_pitch); + height -= 2; + src_ptr += src_pitch << 1; + dst_ptr += dst_pitch << 1; + } while (height > 1); + + if (height > 0) { + pack_8x1_pixels(src_ptr, signal); + filter_8x1_pixels(signal, ff, &res0); + store_8x1_pixels(&res0, &max, dst_ptr); + } +} + +static void aom_highbd_filter_block1d16_h8_avx2( + const uint16_t *src_ptr, ptrdiff_t src_pitch, uint16_t *dst_ptr, + ptrdiff_t dst_pitch, uint32_t height, const int16_t *filter, int bd) { + __m256i signal[8], res0, res1; + const __m256i max = _mm256_set1_epi16((1 << bd) - 1); + + __m256i ff[4]; + pack_filters(filter, ff); + + src_ptr -= 3; + do { + pack_16x1_pixels(src_ptr, signal); + filter_8x1_pixels(signal, ff, &res0); + filter_8x1_pixels(&signal[4], ff, &res1); + store_16x1_pixels(&res0, &res1, &max, dst_ptr); + height -= 1; + src_ptr += src_pitch; + dst_ptr += dst_pitch; + } while (height > 0); +} + +// ----------------------------------------------------------------------------- +// 2-tap horizontal filtering + +static INLINE void pack_2t_filter(const int16_t *filter, __m256i *f) { + const __m128i h = _mm_loadu_si128((const __m128i *)filter); + const __m256i hh = _mm256_insertf128_si256(_mm256_castsi128_si256(h), h, 1); + const __m256i p = _mm256_set1_epi32(0x09080706); + f[0] = _mm256_shuffle_epi8(hh, p); +} + +// can be used by pack_8x2_2t_pixels() and pack_16x1_2t_pixels() +// the difference is s0/s1 specifies first and second rows or, +// first 16 samples and 8-sample shifted 16 samples +static INLINE void pack_16_2t_pixels(const __m256i *s0, const __m256i *s1, + __m256i *sig) { + const __m256i idx = _mm256_loadu_si256((const __m256i *)signal_index); + const __m256i sf2 = _mm256_loadu_si256((const __m256i *)signal_pattern_2); + __m256i x0 = _mm256_shuffle_epi8(*s0, sf2); + __m256i x1 = _mm256_shuffle_epi8(*s1, sf2); + __m256i r0 = _mm256_permutevar8x32_epi32(*s0, idx); + __m256i r1 = _mm256_permutevar8x32_epi32(*s1, idx); + r0 = _mm256_shuffle_epi8(r0, sf2); + r1 = _mm256_shuffle_epi8(r1, sf2); + sig[0] = _mm256_permute2x128_si256(x0, x1, 0x20); + sig[1] = _mm256_permute2x128_si256(r0, r1, 0x20); +} + +static INLINE void pack_8x2_2t_pixels(const uint16_t *src, + const ptrdiff_t pitch, __m256i *sig) { + const __m256i r0 = _mm256_loadu_si256((const __m256i *)src); + const __m256i r1 = _mm256_loadu_si256((const __m256i *)(src + pitch)); + pack_16_2t_pixels(&r0, &r1, sig); +} + +static INLINE void pack_16x1_2t_pixels(const uint16_t *src, + __m256i *sig /*sig[2]*/) { + const __m256i r0 = _mm256_loadu_si256((const __m256i *)src); + const __m256i r1 = _mm256_loadu_si256((const __m256i *)(src + 8)); + pack_16_2t_pixels(&r0, &r1, sig); +} + +static INLINE void pack_8x1_2t_pixels(const uint16_t *src, + __m256i *sig /*sig[2]*/) { + const __m256i idx = _mm256_loadu_si256((const __m256i *)signal_index); + const __m256i sf2 = _mm256_loadu_si256((const __m256i *)signal_pattern_2); + __m256i r0 = _mm256_loadu_si256((const __m256i *)src); + __m256i x0 = _mm256_shuffle_epi8(r0, sf2); + r0 = _mm256_permutevar8x32_epi32(r0, idx); + r0 = _mm256_shuffle_epi8(r0, sf2); + sig[0] = _mm256_permute2x128_si256(x0, r0, 0x20); +} + +// can be used by filter_8x2_2t_pixels() and filter_16x1_2t_pixels() +static INLINE void filter_16_2t_pixels(const __m256i *sig, const __m256i *f, + __m256i *y0, __m256i *y1) { + const __m256i rounding = _mm256_set1_epi32(1 << (CONV8_ROUNDING_BITS - 1)); + __m256i x0 = _mm256_madd_epi16(sig[0], *f); + __m256i x1 = _mm256_madd_epi16(sig[1], *f); + x0 = _mm256_add_epi32(x0, rounding); + x1 = _mm256_add_epi32(x1, rounding); + *y0 = _mm256_srai_epi32(x0, CONV8_ROUNDING_BITS); + *y1 = _mm256_srai_epi32(x1, CONV8_ROUNDING_BITS); +} + +static INLINE void filter_8x1_2t_pixels(const __m256i *sig, const __m256i *f, + __m256i *y0) { + const __m256i rounding = _mm256_set1_epi32(1 << (CONV8_ROUNDING_BITS - 1)); + __m256i x0 = _mm256_madd_epi16(sig[0], *f); + x0 = _mm256_add_epi32(x0, rounding); + *y0 = _mm256_srai_epi32(x0, CONV8_ROUNDING_BITS); +} + +static void aom_highbd_filter_block1d8_h2_avx2( + const uint16_t *src_ptr, ptrdiff_t src_pitch, uint16_t *dst_ptr, + ptrdiff_t dst_pitch, uint32_t height, const int16_t *filter, int bd) { + __m256i signal[2], res0, res1; + const __m256i max = _mm256_set1_epi16((1 << bd) - 1); + + __m256i ff; + pack_2t_filter(filter, &ff); + + src_ptr -= 3; + do { + pack_8x2_2t_pixels(src_ptr, src_pitch, signal); + filter_16_2t_pixels(signal, &ff, &res0, &res1); + store_8x2_pixels(&res0, &res1, &max, dst_ptr, dst_pitch); + height -= 2; + src_ptr += src_pitch << 1; + dst_ptr += dst_pitch << 1; + } while (height > 1); + + if (height > 0) { + pack_8x1_2t_pixels(src_ptr, signal); + filter_8x1_2t_pixels(signal, &ff, &res0); + store_8x1_pixels(&res0, &max, dst_ptr); + } +} + +static void aom_highbd_filter_block1d16_h2_avx2( + const uint16_t *src_ptr, ptrdiff_t src_pitch, uint16_t *dst_ptr, + ptrdiff_t dst_pitch, uint32_t height, const int16_t *filter, int bd) { + __m256i signal[2], res0, res1; + const __m256i max = _mm256_set1_epi16((1 << bd) - 1); + + __m256i ff; + pack_2t_filter(filter, &ff); + + src_ptr -= 3; + do { + pack_16x1_2t_pixels(src_ptr, signal); + filter_16_2t_pixels(signal, &ff, &res0, &res1); + store_16x1_pixels(&res0, &res1, &max, dst_ptr); + height -= 1; + src_ptr += src_pitch; + dst_ptr += dst_pitch; + } while (height > 0); +} + +// ----------------------------------------------------------------------------- +// Vertical Filtering + +static void pack_8x9_init(const uint16_t *src, ptrdiff_t pitch, __m256i *sig) { + __m256i s0 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)src)); + __m256i s1 = + _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(src + pitch))); + __m256i s2 = _mm256_castsi128_si256( + _mm_loadu_si128((const __m128i *)(src + 2 * pitch))); + __m256i s3 = _mm256_castsi128_si256( + _mm_loadu_si128((const __m128i *)(src + 3 * pitch))); + __m256i s4 = _mm256_castsi128_si256( + _mm_loadu_si128((const __m128i *)(src + 4 * pitch))); + __m256i s5 = _mm256_castsi128_si256( + _mm_loadu_si128((const __m128i *)(src + 5 * pitch))); + __m256i s6 = _mm256_castsi128_si256( + _mm_loadu_si128((const __m128i *)(src + 6 * pitch))); + + s0 = _mm256_inserti128_si256(s0, _mm256_castsi256_si128(s1), 1); + s1 = _mm256_inserti128_si256(s1, _mm256_castsi256_si128(s2), 1); + s2 = _mm256_inserti128_si256(s2, _mm256_castsi256_si128(s3), 1); + s3 = _mm256_inserti128_si256(s3, _mm256_castsi256_si128(s4), 1); + s4 = _mm256_inserti128_si256(s4, _mm256_castsi256_si128(s5), 1); + s5 = _mm256_inserti128_si256(s5, _mm256_castsi256_si128(s6), 1); + + sig[0] = _mm256_unpacklo_epi16(s0, s1); + sig[4] = _mm256_unpackhi_epi16(s0, s1); + sig[1] = _mm256_unpacklo_epi16(s2, s3); + sig[5] = _mm256_unpackhi_epi16(s2, s3); + sig[2] = _mm256_unpacklo_epi16(s4, s5); + sig[6] = _mm256_unpackhi_epi16(s4, s5); + sig[8] = s6; +} + +static INLINE void pack_8x9_pixels(const uint16_t *src, ptrdiff_t pitch, + __m256i *sig) { + // base + 7th row + __m256i s0 = _mm256_castsi128_si256( + _mm_loadu_si128((const __m128i *)(src + 7 * pitch))); + // base + 8th row + __m256i s1 = _mm256_castsi128_si256( + _mm_loadu_si128((const __m128i *)(src + 8 * pitch))); + __m256i s2 = _mm256_inserti128_si256(sig[8], _mm256_castsi256_si128(s0), 1); + __m256i s3 = _mm256_inserti128_si256(s0, _mm256_castsi256_si128(s1), 1); + sig[3] = _mm256_unpacklo_epi16(s2, s3); + sig[7] = _mm256_unpackhi_epi16(s2, s3); + sig[8] = s1; +} + +static INLINE void filter_8x9_pixels(const __m256i *sig, const __m256i *f, + __m256i *y0, __m256i *y1) { + filter_8x1_pixels(sig, f, y0); + filter_8x1_pixels(&sig[4], f, y1); +} + +static INLINE void update_pixels(__m256i *sig) { + int i; + for (i = 0; i < 3; ++i) { + sig[i] = sig[i + 1]; + sig[i + 4] = sig[i + 5]; + } +} + +static void aom_highbd_filter_block1d8_v8_avx2( + const uint16_t *src_ptr, ptrdiff_t src_pitch, uint16_t *dst_ptr, + ptrdiff_t dst_pitch, uint32_t height, const int16_t *filter, int bd) { + __m256i signal[9], res0, res1; + const __m256i max = _mm256_set1_epi16((1 << bd) - 1); + + __m256i ff[4]; + pack_filters(filter, ff); + + pack_8x9_init(src_ptr, src_pitch, signal); + + do { + pack_8x9_pixels(src_ptr, src_pitch, signal); + + filter_8x9_pixels(signal, ff, &res0, &res1); + store_8x2_pixels(&res0, &res1, &max, dst_ptr, dst_pitch); + update_pixels(signal); + + src_ptr += src_pitch << 1; + dst_ptr += dst_pitch << 1; + height -= 2; + } while (height > 0); +} + +static void pack_16x9_init(const uint16_t *src, ptrdiff_t pitch, __m256i *sig) { + __m256i u0, u1, u2, u3; + // load 0-6 rows + const __m256i s0 = _mm256_loadu_si256((const __m256i *)src); + const __m256i s1 = _mm256_loadu_si256((const __m256i *)(src + pitch)); + const __m256i s2 = _mm256_loadu_si256((const __m256i *)(src + 2 * pitch)); + const __m256i s3 = _mm256_loadu_si256((const __m256i *)(src + 3 * pitch)); + const __m256i s4 = _mm256_loadu_si256((const __m256i *)(src + 4 * pitch)); + const __m256i s5 = _mm256_loadu_si256((const __m256i *)(src + 5 * pitch)); + const __m256i s6 = _mm256_loadu_si256((const __m256i *)(src + 6 * pitch)); + + u0 = _mm256_permute2x128_si256(s0, s1, 0x20); // 0, 1 low + u1 = _mm256_permute2x128_si256(s0, s1, 0x31); // 0, 1 high + + u2 = _mm256_permute2x128_si256(s1, s2, 0x20); // 1, 2 low + u3 = _mm256_permute2x128_si256(s1, s2, 0x31); // 1, 2 high + + sig[0] = _mm256_unpacklo_epi16(u0, u2); + sig[4] = _mm256_unpackhi_epi16(u0, u2); + + sig[8] = _mm256_unpacklo_epi16(u1, u3); + sig[12] = _mm256_unpackhi_epi16(u1, u3); + + u0 = _mm256_permute2x128_si256(s2, s3, 0x20); + u1 = _mm256_permute2x128_si256(s2, s3, 0x31); + + u2 = _mm256_permute2x128_si256(s3, s4, 0x20); + u3 = _mm256_permute2x128_si256(s3, s4, 0x31); + + sig[1] = _mm256_unpacklo_epi16(u0, u2); + sig[5] = _mm256_unpackhi_epi16(u0, u2); + + sig[9] = _mm256_unpacklo_epi16(u1, u3); + sig[13] = _mm256_unpackhi_epi16(u1, u3); + + u0 = _mm256_permute2x128_si256(s4, s5, 0x20); + u1 = _mm256_permute2x128_si256(s4, s5, 0x31); + + u2 = _mm256_permute2x128_si256(s5, s6, 0x20); + u3 = _mm256_permute2x128_si256(s5, s6, 0x31); + + sig[2] = _mm256_unpacklo_epi16(u0, u2); + sig[6] = _mm256_unpackhi_epi16(u0, u2); + + sig[10] = _mm256_unpacklo_epi16(u1, u3); + sig[14] = _mm256_unpackhi_epi16(u1, u3); + + sig[16] = s6; +} + +static void pack_16x9_pixels(const uint16_t *src, ptrdiff_t pitch, + __m256i *sig) { + // base + 7th row + const __m256i s7 = _mm256_loadu_si256((const __m256i *)(src + 7 * pitch)); + // base + 8th row + const __m256i s8 = _mm256_loadu_si256((const __m256i *)(src + 8 * pitch)); + + __m256i u0, u1, u2, u3; + u0 = _mm256_permute2x128_si256(sig[16], s7, 0x20); + u1 = _mm256_permute2x128_si256(sig[16], s7, 0x31); + + u2 = _mm256_permute2x128_si256(s7, s8, 0x20); + u3 = _mm256_permute2x128_si256(s7, s8, 0x31); + + sig[3] = _mm256_unpacklo_epi16(u0, u2); + sig[7] = _mm256_unpackhi_epi16(u0, u2); + + sig[11] = _mm256_unpacklo_epi16(u1, u3); + sig[15] = _mm256_unpackhi_epi16(u1, u3); + + sig[16] = s8; +} + +static INLINE void filter_16x9_pixels(const __m256i *sig, const __m256i *f, + __m256i *y0, __m256i *y1) { + __m256i res[4]; + int i; + for (i = 0; i < 4; ++i) { + filter_8x1_pixels(&sig[i << 2], f, &res[i]); + } + + { + const __m256i l0l1 = _mm256_packus_epi32(res[0], res[1]); + const __m256i h0h1 = _mm256_packus_epi32(res[2], res[3]); + *y0 = _mm256_permute2x128_si256(l0l1, h0h1, 0x20); + *y1 = _mm256_permute2x128_si256(l0l1, h0h1, 0x31); + } +} + +static INLINE void store_16x2_pixels(const __m256i *y0, const __m256i *y1, + const __m256i *mask, uint16_t *dst, + ptrdiff_t pitch) { + __m256i p = _mm256_min_epi16(*y0, *mask); + _mm256_storeu_si256((__m256i *)dst, p); + p = _mm256_min_epi16(*y1, *mask); + _mm256_storeu_si256((__m256i *)(dst + pitch), p); +} + +static void update_16x9_pixels(__m256i *sig) { + update_pixels(&sig[0]); + update_pixels(&sig[8]); +} + +static void aom_highbd_filter_block1d16_v8_avx2( + const uint16_t *src_ptr, ptrdiff_t src_pitch, uint16_t *dst_ptr, + ptrdiff_t dst_pitch, uint32_t height, const int16_t *filter, int bd) { + __m256i signal[17], res0, res1; + const __m256i max = _mm256_set1_epi16((1 << bd) - 1); + + __m256i ff[4]; + pack_filters(filter, ff); + + pack_16x9_init(src_ptr, src_pitch, signal); + + do { + pack_16x9_pixels(src_ptr, src_pitch, signal); + filter_16x9_pixels(signal, ff, &res0, &res1); + store_16x2_pixels(&res0, &res1, &max, dst_ptr, dst_pitch); + update_16x9_pixels(signal); + + src_ptr += src_pitch << 1; + dst_ptr += dst_pitch << 1; + height -= 2; + } while (height > 0); +} + +// ----------------------------------------------------------------------------- +// 2-tap vertical filtering + +static void pack_16x2_init(const uint16_t *src, __m256i *sig) { + sig[2] = _mm256_loadu_si256((const __m256i *)src); +} + +static INLINE void pack_16x2_2t_pixels(const uint16_t *src, ptrdiff_t pitch, + __m256i *sig) { + // load the next row + const __m256i u = _mm256_loadu_si256((const __m256i *)(src + pitch)); + sig[0] = _mm256_unpacklo_epi16(sig[2], u); + sig[1] = _mm256_unpackhi_epi16(sig[2], u); + sig[2] = u; +} + +static INLINE void filter_16x2_2t_pixels(const __m256i *sig, const __m256i *f, + __m256i *y0, __m256i *y1) { + filter_16_2t_pixels(sig, f, y0, y1); +} + +static void aom_highbd_filter_block1d16_v2_avx2( + const uint16_t *src_ptr, ptrdiff_t src_pitch, uint16_t *dst_ptr, + ptrdiff_t dst_pitch, uint32_t height, const int16_t *filter, int bd) { + __m256i signal[3], res0, res1; + const __m256i max = _mm256_set1_epi16((1 << bd) - 1); + __m256i ff; + + pack_2t_filter(filter, &ff); + pack_16x2_init(src_ptr, signal); + + do { + pack_16x2_2t_pixels(src_ptr, src_pitch, signal); + filter_16x2_2t_pixels(signal, &ff, &res0, &res1); + store_16x1_pixels(&res0, &res1, &max, dst_ptr); + + src_ptr += src_pitch; + dst_ptr += dst_pitch; + height -= 1; + } while (height > 0); +} + +static INLINE void pack_8x1_2t_filter(const int16_t *filter, __m128i *f) { + const __m128i h = _mm_loadu_si128((const __m128i *)filter); + const __m128i p = _mm_set1_epi32(0x09080706); + f[0] = _mm_shuffle_epi8(h, p); +} + +static void pack_8x2_init(const uint16_t *src, __m128i *sig) { + sig[2] = _mm_loadu_si128((const __m128i *)src); +} + +static INLINE void pack_8x2_2t_pixels_ver(const uint16_t *src, ptrdiff_t pitch, + __m128i *sig) { + // load the next row + const __m128i u = _mm_loadu_si128((const __m128i *)(src + pitch)); + sig[0] = _mm_unpacklo_epi16(sig[2], u); + sig[1] = _mm_unpackhi_epi16(sig[2], u); + sig[2] = u; +} + +static INLINE void filter_8_2t_pixels(const __m128i *sig, const __m128i *f, + __m128i *y0, __m128i *y1) { + const __m128i rounding = _mm_set1_epi32(1 << (CONV8_ROUNDING_BITS - 1)); + __m128i x0 = _mm_madd_epi16(sig[0], *f); + __m128i x1 = _mm_madd_epi16(sig[1], *f); + x0 = _mm_add_epi32(x0, rounding); + x1 = _mm_add_epi32(x1, rounding); + *y0 = _mm_srai_epi32(x0, CONV8_ROUNDING_BITS); + *y1 = _mm_srai_epi32(x1, CONV8_ROUNDING_BITS); +} + +static INLINE void store_8x1_2t_pixels_ver(const __m128i *y0, const __m128i *y1, + const __m128i *mask, uint16_t *dst) { + __m128i res = _mm_packus_epi32(*y0, *y1); + res = _mm_min_epi16(res, *mask); + _mm_storeu_si128((__m128i *)dst, res); +} + +static void aom_highbd_filter_block1d8_v2_avx2( + const uint16_t *src_ptr, ptrdiff_t src_pitch, uint16_t *dst_ptr, + ptrdiff_t dst_pitch, uint32_t height, const int16_t *filter, int bd) { + __m128i signal[3], res0, res1; + const __m128i max = _mm_set1_epi16((1 << bd) - 1); + __m128i ff; + + pack_8x1_2t_filter(filter, &ff); + pack_8x2_init(src_ptr, signal); + + do { + pack_8x2_2t_pixels_ver(src_ptr, src_pitch, signal); + filter_8_2t_pixels(signal, &ff, &res0, &res1); + store_8x1_2t_pixels_ver(&res0, &res1, &max, dst_ptr); + + src_ptr += src_pitch; + dst_ptr += dst_pitch; + height -= 1; + } while (height > 0); +} + +void aom_highbd_filter_block1d4_h8_sse2(const uint16_t *, ptrdiff_t, uint16_t *, + ptrdiff_t, uint32_t, const int16_t *, + int); +void aom_highbd_filter_block1d4_h2_sse2(const uint16_t *, ptrdiff_t, uint16_t *, + ptrdiff_t, uint32_t, const int16_t *, + int); +void aom_highbd_filter_block1d4_v8_sse2(const uint16_t *, ptrdiff_t, uint16_t *, + ptrdiff_t, uint32_t, const int16_t *, + int); +void aom_highbd_filter_block1d4_v2_sse2(const uint16_t *, ptrdiff_t, uint16_t *, + ptrdiff_t, uint32_t, const int16_t *, + int); +#define aom_highbd_filter_block1d4_h8_avx2 aom_highbd_filter_block1d4_h8_sse2 +#define aom_highbd_filter_block1d4_h2_avx2 aom_highbd_filter_block1d4_h2_sse2 +#define aom_highbd_filter_block1d4_v8_avx2 aom_highbd_filter_block1d4_v8_sse2 +#define aom_highbd_filter_block1d4_v2_avx2 aom_highbd_filter_block1d4_v2_sse2 + +HIGH_FUN_CONV_1D(horiz, x_step_q4, filter_x, h, src, , avx2); +HIGH_FUN_CONV_1D(vert, y_step_q4, filter_y, v, src - src_stride * 3, , avx2); + +#undef HIGHBD_FUNC diff --git a/third_party/aom/aom_dsp/x86/highbd_convolve_ssse3.c b/third_party/aom/aom_dsp/x86/highbd_convolve_ssse3.c new file mode 100644 index 0000000000..e7b33d1c46 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/highbd_convolve_ssse3.c @@ -0,0 +1,251 @@ +/* + * Copyright (c) 2018, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include +#include + +#include "config/aom_dsp_rtcd.h" + +#include "aom_dsp/x86/convolve_sse2.h" + +void av1_highbd_convolve_y_sr_ssse3(const uint16_t *src, int src_stride, + uint16_t *dst, int dst_stride, int w, int h, + const InterpFilterParams *filter_params_x, + const InterpFilterParams *filter_params_y, + const int subpel_x_q4, + const int subpel_y_q4, + ConvolveParams *conv_params, int bd) { + int i, j; + const int fo_vert = filter_params_y->taps / 2 - 1; + const uint16_t *const src_ptr = src - fo_vert * src_stride; + (void)filter_params_x; + (void)subpel_x_q4; + (void)conv_params; + + assert(conv_params->round_0 <= FILTER_BITS); + assert(((conv_params->round_0 + conv_params->round_1) <= (FILTER_BITS + 1)) || + ((conv_params->round_0 + conv_params->round_1) == (2 * FILTER_BITS))); + + __m128i s[16], coeffs_y[4]; + + const int bits = FILTER_BITS; + + const __m128i round_shift_bits = _mm_cvtsi32_si128(bits); + const __m128i round_const_bits = _mm_set1_epi32((1 << bits) >> 1); + const __m128i clip_pixel = + _mm_set1_epi16(bd == 10 ? 1023 : (bd == 12 ? 4095 : 255)); + const __m128i zero = _mm_setzero_si128(); + + prepare_coeffs(filter_params_y, subpel_y_q4, coeffs_y); + + for (j = 0; j < w; j += 8) { + const uint16_t *data = &src_ptr[j]; + /* Vertical filter */ + { + __m128i s0 = _mm_loadu_si128((__m128i *)(data + 0 * src_stride)); + __m128i s1 = _mm_loadu_si128((__m128i *)(data + 1 * src_stride)); + __m128i s2 = _mm_loadu_si128((__m128i *)(data + 2 * src_stride)); + __m128i s3 = _mm_loadu_si128((__m128i *)(data + 3 * src_stride)); + __m128i s4 = _mm_loadu_si128((__m128i *)(data + 4 * src_stride)); + __m128i s5 = _mm_loadu_si128((__m128i *)(data + 5 * src_stride)); + __m128i s6 = _mm_loadu_si128((__m128i *)(data + 6 * src_stride)); + + s[0] = _mm_unpacklo_epi16(s0, s1); + s[1] = _mm_unpacklo_epi16(s2, s3); + s[2] = _mm_unpacklo_epi16(s4, s5); + + s[4] = _mm_unpackhi_epi16(s0, s1); + s[5] = _mm_unpackhi_epi16(s2, s3); + s[6] = _mm_unpackhi_epi16(s4, s5); + + s[0 + 8] = _mm_unpacklo_epi16(s1, s2); + s[1 + 8] = _mm_unpacklo_epi16(s3, s4); + s[2 + 8] = _mm_unpacklo_epi16(s5, s6); + + s[4 + 8] = _mm_unpackhi_epi16(s1, s2); + s[5 + 8] = _mm_unpackhi_epi16(s3, s4); + s[6 + 8] = _mm_unpackhi_epi16(s5, s6); + + for (i = 0; i < h; i += 2) { + data = &src_ptr[i * src_stride + j]; + + __m128i s7 = _mm_loadu_si128((__m128i *)(data + 7 * src_stride)); + __m128i s8 = _mm_loadu_si128((__m128i *)(data + 8 * src_stride)); + + s[3] = _mm_unpacklo_epi16(s6, s7); + s[7] = _mm_unpackhi_epi16(s6, s7); + + s[3 + 8] = _mm_unpacklo_epi16(s7, s8); + s[7 + 8] = _mm_unpackhi_epi16(s7, s8); + + const __m128i res_a0 = convolve(s, coeffs_y); + __m128i res_a_round0 = _mm_sra_epi32( + _mm_add_epi32(res_a0, round_const_bits), round_shift_bits); + + const __m128i res_a1 = convolve(s + 8, coeffs_y); + __m128i res_a_round1 = _mm_sra_epi32( + _mm_add_epi32(res_a1, round_const_bits), round_shift_bits); + + if (w - j > 4) { + const __m128i res_b0 = convolve(s + 4, coeffs_y); + __m128i res_b_round0 = _mm_sra_epi32( + _mm_add_epi32(res_b0, round_const_bits), round_shift_bits); + + const __m128i res_b1 = convolve(s + 4 + 8, coeffs_y); + __m128i res_b_round1 = _mm_sra_epi32( + _mm_add_epi32(res_b1, round_const_bits), round_shift_bits); + + __m128i res_16bit0 = _mm_packs_epi32(res_a_round0, res_b_round0); + res_16bit0 = _mm_min_epi16(res_16bit0, clip_pixel); + res_16bit0 = _mm_max_epi16(res_16bit0, zero); + + __m128i res_16bit1 = _mm_packs_epi32(res_a_round1, res_b_round1); + res_16bit1 = _mm_min_epi16(res_16bit1, clip_pixel); + res_16bit1 = _mm_max_epi16(res_16bit1, zero); + + _mm_storeu_si128((__m128i *)&dst[i * dst_stride + j], res_16bit0); + _mm_storeu_si128((__m128i *)&dst[i * dst_stride + j + dst_stride], + res_16bit1); + } else if (w == 4) { + res_a_round0 = _mm_packs_epi32(res_a_round0, res_a_round0); + res_a_round0 = _mm_min_epi16(res_a_round0, clip_pixel); + res_a_round0 = _mm_max_epi16(res_a_round0, zero); + + res_a_round1 = _mm_packs_epi32(res_a_round1, res_a_round1); + res_a_round1 = _mm_min_epi16(res_a_round1, clip_pixel); + res_a_round1 = _mm_max_epi16(res_a_round1, zero); + + _mm_storel_epi64((__m128i *)&dst[i * dst_stride + j], res_a_round0); + _mm_storel_epi64((__m128i *)&dst[i * dst_stride + j + dst_stride], + res_a_round1); + } else { + res_a_round0 = _mm_packs_epi32(res_a_round0, res_a_round0); + res_a_round0 = _mm_min_epi16(res_a_round0, clip_pixel); + res_a_round0 = _mm_max_epi16(res_a_round0, zero); + + res_a_round1 = _mm_packs_epi32(res_a_round1, res_a_round1); + res_a_round1 = _mm_min_epi16(res_a_round1, clip_pixel); + res_a_round1 = _mm_max_epi16(res_a_round1, zero); + + *((uint32_t *)(&dst[i * dst_stride + j])) = + _mm_cvtsi128_si32(res_a_round0); + + *((uint32_t *)(&dst[i * dst_stride + j + dst_stride])) = + _mm_cvtsi128_si32(res_a_round1); + } + + s[0] = s[1]; + s[1] = s[2]; + s[2] = s[3]; + + s[4] = s[5]; + s[5] = s[6]; + s[6] = s[7]; + + s[0 + 8] = s[1 + 8]; + s[1 + 8] = s[2 + 8]; + s[2 + 8] = s[3 + 8]; + + s[4 + 8] = s[5 + 8]; + s[5 + 8] = s[6 + 8]; + s[6 + 8] = s[7 + 8]; + + s6 = s8; + } + } + } +} + +void av1_highbd_convolve_x_sr_ssse3(const uint16_t *src, int src_stride, + uint16_t *dst, int dst_stride, int w, int h, + const InterpFilterParams *filter_params_x, + const InterpFilterParams *filter_params_y, + const int subpel_x_q4, + const int subpel_y_q4, + ConvolveParams *conv_params, int bd) { + int i, j; + const int fo_horiz = filter_params_x->taps / 2 - 1; + const uint16_t *const src_ptr = src - fo_horiz; + (void)subpel_y_q4; + (void)filter_params_y; + + // Check that, even with 12-bit input, the intermediate values will fit + // into an unsigned 16-bit intermediate array. + assert(bd + FILTER_BITS + 2 - conv_params->round_0 <= 16); + + __m128i s[4], coeffs_x[4]; + + const __m128i round_const_x = + _mm_set1_epi32(((1 << conv_params->round_0) >> 1)); + const __m128i round_shift_x = _mm_cvtsi32_si128(conv_params->round_0); + + const int bits = FILTER_BITS - conv_params->round_0; + + const __m128i round_shift_bits = _mm_cvtsi32_si128(bits); + const __m128i round_const_bits = _mm_set1_epi32((1 << bits) >> 1); + const __m128i clip_pixel = + _mm_set1_epi16(bd == 10 ? 1023 : (bd == 12 ? 4095 : 255)); + const __m128i zero = _mm_setzero_si128(); + + prepare_coeffs(filter_params_x, subpel_x_q4, coeffs_x); + + for (j = 0; j < w; j += 8) { + /* Horizontal filter */ + { + for (i = 0; i < h; i += 1) { + const __m128i row00 = + _mm_loadu_si128((__m128i *)&src_ptr[i * src_stride + j]); + const __m128i row01 = + _mm_loadu_si128((__m128i *)&src_ptr[i * src_stride + (j + 8)]); + + // even pixels + s[0] = _mm_alignr_epi8(row01, row00, 0); + s[1] = _mm_alignr_epi8(row01, row00, 4); + s[2] = _mm_alignr_epi8(row01, row00, 8); + s[3] = _mm_alignr_epi8(row01, row00, 12); + + __m128i res_even = convolve(s, coeffs_x); + res_even = _mm_sra_epi32(_mm_add_epi32(res_even, round_const_x), + round_shift_x); + + // odd pixels + s[0] = _mm_alignr_epi8(row01, row00, 2); + s[1] = _mm_alignr_epi8(row01, row00, 6); + s[2] = _mm_alignr_epi8(row01, row00, 10); + s[3] = _mm_alignr_epi8(row01, row00, 14); + + __m128i res_odd = convolve(s, coeffs_x); + res_odd = + _mm_sra_epi32(_mm_add_epi32(res_odd, round_const_x), round_shift_x); + + res_even = _mm_sra_epi32(_mm_add_epi32(res_even, round_const_bits), + round_shift_bits); + res_odd = _mm_sra_epi32(_mm_add_epi32(res_odd, round_const_bits), + round_shift_bits); + + __m128i res_even1 = _mm_packs_epi32(res_even, res_even); + __m128i res_odd1 = _mm_packs_epi32(res_odd, res_odd); + __m128i res = _mm_unpacklo_epi16(res_even1, res_odd1); + + res = _mm_min_epi16(res, clip_pixel); + res = _mm_max_epi16(res, zero); + + if (w - j > 4) { + _mm_storeu_si128((__m128i *)&dst[i * dst_stride + j], res); + } else if (w == 4) { + _mm_storel_epi64((__m128i *)&dst[i * dst_stride + j], res); + } else { + *((uint32_t *)(&dst[i * dst_stride + j])) = _mm_cvtsi128_si32(res); + } + } + } + } +} diff --git a/third_party/aom/aom_dsp/x86/highbd_intrapred_sse2.c b/third_party/aom/aom_dsp/x86/highbd_intrapred_sse2.c new file mode 100644 index 0000000000..5a55736c48 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/highbd_intrapred_sse2.c @@ -0,0 +1,984 @@ +/* + * Copyright (c) 2017, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include + +#include "config/aom_dsp_rtcd.h" + +// ----------------------------------------------------------------------------- +// H_PRED + +void aom_highbd_h_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, + const uint16_t *left, int bd) { + const __m128i left_u16 = _mm_loadl_epi64((const __m128i *)left); + const __m128i row0 = _mm_shufflelo_epi16(left_u16, 0x0); + const __m128i row1 = _mm_shufflelo_epi16(left_u16, 0x55); + const __m128i row2 = _mm_shufflelo_epi16(left_u16, 0xaa); + const __m128i row3 = _mm_shufflelo_epi16(left_u16, 0xff); + (void)above; + (void)bd; + _mm_storel_epi64((__m128i *)dst, row0); + dst += stride; + _mm_storel_epi64((__m128i *)dst, row1); + dst += stride; + _mm_storel_epi64((__m128i *)dst, row2); + dst += stride; + _mm_storel_epi64((__m128i *)dst, row3); +} + +void aom_highbd_h_predictor_4x8_sse2(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, + const uint16_t *left, int bd) { + aom_highbd_h_predictor_4x4_sse2(dst, stride, above, left, bd); + dst += stride << 2; + left += 4; + aom_highbd_h_predictor_4x4_sse2(dst, stride, above, left, bd); +} + +void aom_highbd_h_predictor_8x4_sse2(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, + const uint16_t *left, int bd) { + const __m128i left_u16 = _mm_load_si128((const __m128i *)left); + const __m128i row0 = _mm_shufflelo_epi16(left_u16, 0x0); + const __m128i row1 = _mm_shufflelo_epi16(left_u16, 0x55); + const __m128i row2 = _mm_shufflelo_epi16(left_u16, 0xaa); + const __m128i row3 = _mm_shufflelo_epi16(left_u16, 0xff); + (void)above; + (void)bd; + _mm_store_si128((__m128i *)dst, _mm_unpacklo_epi64(row0, row0)); + dst += stride; + _mm_store_si128((__m128i *)dst, _mm_unpacklo_epi64(row1, row1)); + dst += stride; + _mm_store_si128((__m128i *)dst, _mm_unpacklo_epi64(row2, row2)); + dst += stride; + _mm_store_si128((__m128i *)dst, _mm_unpacklo_epi64(row3, row3)); +} + +void aom_highbd_h_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, + const uint16_t *left, int bd) { + const __m128i left_u16 = _mm_load_si128((const __m128i *)left); + const __m128i row0 = _mm_shufflelo_epi16(left_u16, 0x0); + const __m128i row1 = _mm_shufflelo_epi16(left_u16, 0x55); + const __m128i row2 = _mm_shufflelo_epi16(left_u16, 0xaa); + const __m128i row3 = _mm_shufflelo_epi16(left_u16, 0xff); + const __m128i row4 = _mm_shufflehi_epi16(left_u16, 0x0); + const __m128i row5 = _mm_shufflehi_epi16(left_u16, 0x55); + const __m128i row6 = _mm_shufflehi_epi16(left_u16, 0xaa); + const __m128i row7 = _mm_shufflehi_epi16(left_u16, 0xff); + (void)above; + (void)bd; + _mm_store_si128((__m128i *)dst, _mm_unpacklo_epi64(row0, row0)); + dst += stride; + _mm_store_si128((__m128i *)dst, _mm_unpacklo_epi64(row1, row1)); + dst += stride; + _mm_store_si128((__m128i *)dst, _mm_unpacklo_epi64(row2, row2)); + dst += stride; + _mm_store_si128((__m128i *)dst, _mm_unpacklo_epi64(row3, row3)); + dst += stride; + _mm_store_si128((__m128i *)dst, _mm_unpackhi_epi64(row4, row4)); + dst += stride; + _mm_store_si128((__m128i *)dst, _mm_unpackhi_epi64(row5, row5)); + dst += stride; + _mm_store_si128((__m128i *)dst, _mm_unpackhi_epi64(row6, row6)); + dst += stride; + _mm_store_si128((__m128i *)dst, _mm_unpackhi_epi64(row7, row7)); +} + +void aom_highbd_h_predictor_8x16_sse2(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, + const uint16_t *left, int bd) { + aom_highbd_h_predictor_8x8_sse2(dst, stride, above, left, bd); + dst += stride << 3; + left += 8; + aom_highbd_h_predictor_8x8_sse2(dst, stride, above, left, bd); +} + +static INLINE void h_store_16_unpacklo(uint16_t **dst, const ptrdiff_t stride, + const __m128i *row) { + const __m128i val = _mm_unpacklo_epi64(*row, *row); + _mm_store_si128((__m128i *)*dst, val); + _mm_store_si128((__m128i *)(*dst + 8), val); + *dst += stride; +} + +static INLINE void h_store_16_unpackhi(uint16_t **dst, const ptrdiff_t stride, + const __m128i *row) { + const __m128i val = _mm_unpackhi_epi64(*row, *row); + _mm_store_si128((__m128i *)(*dst), val); + _mm_store_si128((__m128i *)(*dst + 8), val); + *dst += stride; +} + +static INLINE void h_predictor_16x8(uint16_t *dst, ptrdiff_t stride, + const uint16_t *left) { + const __m128i left_u16 = _mm_load_si128((const __m128i *)left); + const __m128i row0 = _mm_shufflelo_epi16(left_u16, 0x0); + const __m128i row1 = _mm_shufflelo_epi16(left_u16, 0x55); + const __m128i row2 = _mm_shufflelo_epi16(left_u16, 0xaa); + const __m128i row3 = _mm_shufflelo_epi16(left_u16, 0xff); + const __m128i row4 = _mm_shufflehi_epi16(left_u16, 0x0); + const __m128i row5 = _mm_shufflehi_epi16(left_u16, 0x55); + const __m128i row6 = _mm_shufflehi_epi16(left_u16, 0xaa); + const __m128i row7 = _mm_shufflehi_epi16(left_u16, 0xff); + h_store_16_unpacklo(&dst, stride, &row0); + h_store_16_unpacklo(&dst, stride, &row1); + h_store_16_unpacklo(&dst, stride, &row2); + h_store_16_unpacklo(&dst, stride, &row3); + h_store_16_unpackhi(&dst, stride, &row4); + h_store_16_unpackhi(&dst, stride, &row5); + h_store_16_unpackhi(&dst, stride, &row6); + h_store_16_unpackhi(&dst, stride, &row7); +} + +void aom_highbd_h_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, + const uint16_t *left, int bd) { + (void)above; + (void)bd; + h_predictor_16x8(dst, stride, left); +} + +void aom_highbd_h_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, + const uint16_t *left, int bd) { + int i; + (void)above; + (void)bd; + + for (i = 0; i < 2; i++, left += 8) { + h_predictor_16x8(dst, stride, left); + dst += stride << 3; + } +} + +void aom_highbd_h_predictor_16x32_sse2(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, + const uint16_t *left, int bd) { + int i; + (void)above; + (void)bd; + + for (i = 0; i < 4; i++, left += 8) { + h_predictor_16x8(dst, stride, left); + dst += stride << 3; + } +} + +static INLINE void h_store_32_unpacklo(uint16_t **dst, const ptrdiff_t stride, + const __m128i *row) { + const __m128i val = _mm_unpacklo_epi64(*row, *row); + _mm_store_si128((__m128i *)(*dst), val); + _mm_store_si128((__m128i *)(*dst + 8), val); + _mm_store_si128((__m128i *)(*dst + 16), val); + _mm_store_si128((__m128i *)(*dst + 24), val); + *dst += stride; +} + +static INLINE void h_store_32_unpackhi(uint16_t **dst, const ptrdiff_t stride, + const __m128i *row) { + const __m128i val = _mm_unpackhi_epi64(*row, *row); + _mm_store_si128((__m128i *)(*dst), val); + _mm_store_si128((__m128i *)(*dst + 8), val); + _mm_store_si128((__m128i *)(*dst + 16), val); + _mm_store_si128((__m128i *)(*dst + 24), val); + *dst += stride; +} + +static INLINE void h_predictor_32x8(uint16_t *dst, ptrdiff_t stride, + const uint16_t *left) { + const __m128i left_u16 = _mm_load_si128((const __m128i *)left); + const __m128i row0 = _mm_shufflelo_epi16(left_u16, 0x0); + const __m128i row1 = _mm_shufflelo_epi16(left_u16, 0x55); + const __m128i row2 = _mm_shufflelo_epi16(left_u16, 0xaa); + const __m128i row3 = _mm_shufflelo_epi16(left_u16, 0xff); + const __m128i row4 = _mm_shufflehi_epi16(left_u16, 0x0); + const __m128i row5 = _mm_shufflehi_epi16(left_u16, 0x55); + const __m128i row6 = _mm_shufflehi_epi16(left_u16, 0xaa); + const __m128i row7 = _mm_shufflehi_epi16(left_u16, 0xff); + h_store_32_unpacklo(&dst, stride, &row0); + h_store_32_unpacklo(&dst, stride, &row1); + h_store_32_unpacklo(&dst, stride, &row2); + h_store_32_unpacklo(&dst, stride, &row3); + h_store_32_unpackhi(&dst, stride, &row4); + h_store_32_unpackhi(&dst, stride, &row5); + h_store_32_unpackhi(&dst, stride, &row6); + h_store_32_unpackhi(&dst, stride, &row7); +} + +void aom_highbd_h_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, + const uint16_t *left, int bd) { + int i; + (void)above; + (void)bd; + + for (i = 0; i < 2; i++, left += 8) { + h_predictor_32x8(dst, stride, left); + dst += stride << 3; + } +} + +void aom_highbd_h_predictor_32x32_sse2(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, + const uint16_t *left, int bd) { + int i; + (void)above; + (void)bd; + + for (i = 0; i < 4; i++, left += 8) { + h_predictor_32x8(dst, stride, left); + dst += stride << 3; + } +} + +// ----------------------------------------------------------------------------- +// DC_TOP, DC_LEFT, DC_128 + +// 4x4 + +static INLINE __m128i dc_sum_4(const uint16_t *ref) { + const __m128i _dcba = _mm_loadl_epi64((const __m128i *)ref); + const __m128i _xxdc = _mm_shufflelo_epi16(_dcba, 0xe); + const __m128i a = _mm_add_epi16(_dcba, _xxdc); + return _mm_add_epi16(a, _mm_shufflelo_epi16(a, 0x1)); +} + +static INLINE void dc_store_4x4(uint16_t *dst, ptrdiff_t stride, + const __m128i *dc) { + const __m128i dc_dup = _mm_shufflelo_epi16(*dc, 0x0); + int i; + for (i = 0; i < 4; ++i, dst += stride) { + _mm_storel_epi64((__m128i *)dst, dc_dup); + } +} + +void aom_highbd_dc_left_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, + const uint16_t *left, int bd) { + const __m128i two = _mm_cvtsi32_si128(2); + const __m128i sum = dc_sum_4(left); + const __m128i dc = _mm_srli_epi16(_mm_add_epi16(sum, two), 2); + (void)above; + (void)bd; + dc_store_4x4(dst, stride, &dc); +} + +void aom_highbd_dc_top_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, + const uint16_t *left, int bd) { + const __m128i two = _mm_cvtsi32_si128(2); + const __m128i sum = dc_sum_4(above); + const __m128i dc = _mm_srli_epi16(_mm_add_epi16(sum, two), 2); + (void)left; + (void)bd; + dc_store_4x4(dst, stride, &dc); +} + +void aom_highbd_dc_128_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, + const uint16_t *left, int bd) { + const __m128i dc = _mm_cvtsi32_si128(1 << (bd - 1)); + const __m128i dc_dup = _mm_shufflelo_epi16(dc, 0x0); + (void)above; + (void)left; + dc_store_4x4(dst, stride, &dc_dup); +} + +// ----------------------------------------------------------------------------- +// 4x8 + +static INLINE void dc_store_4x8(uint16_t *dst, ptrdiff_t stride, + const __m128i *dc) { + const __m128i dc_dup = _mm_shufflelo_epi16(*dc, 0x0); + int i; + for (i = 0; i < 8; ++i, dst += stride) { + _mm_storel_epi64((__m128i *)dst, dc_dup); + } +} + +// Shared with DC 8xh +static INLINE __m128i dc_sum_8(const uint16_t *ref) { + const __m128i ref_u16 = _mm_load_si128((const __m128i *)ref); + const __m128i _dcba = _mm_add_epi16(ref_u16, _mm_srli_si128(ref_u16, 8)); + const __m128i _xxdc = _mm_shufflelo_epi16(_dcba, 0xe); + const __m128i a = _mm_add_epi16(_dcba, _xxdc); + + return _mm_add_epi16(a, _mm_shufflelo_epi16(a, 0x1)); +} + +void aom_highbd_dc_left_predictor_4x8_sse2(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, + const uint16_t *left, int bd) { + const __m128i sum = dc_sum_8(left); + const __m128i four = _mm_cvtsi32_si128(4); + const __m128i dc = _mm_srli_epi16(_mm_add_epi16(sum, four), 3); + (void)above; + (void)bd; + dc_store_4x8(dst, stride, &dc); +} + +void aom_highbd_dc_top_predictor_4x8_sse2(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, + const uint16_t *left, int bd) { + const __m128i two = _mm_cvtsi32_si128(2); + const __m128i sum = dc_sum_4(above); + const __m128i dc = _mm_srli_epi16(_mm_add_epi16(sum, two), 2); + (void)left; + (void)bd; + dc_store_4x8(dst, stride, &dc); +} + +void aom_highbd_dc_128_predictor_4x8_sse2(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, + const uint16_t *left, int bd) { + const __m128i dc = _mm_cvtsi32_si128(1 << (bd - 1)); + const __m128i dc_dup = _mm_shufflelo_epi16(dc, 0x0); + (void)above; + (void)left; + dc_store_4x8(dst, stride, &dc_dup); +} + +// ----------------------------------------------------------------------------- +// 8xh + +static INLINE void dc_store_8xh(uint16_t *dst, ptrdiff_t stride, int height, + const __m128i *dc) { + const __m128i dc_dup_lo = _mm_shufflelo_epi16(*dc, 0); + const __m128i dc_dup = _mm_unpacklo_epi64(dc_dup_lo, dc_dup_lo); + int i; + for (i = 0; i < height; ++i, dst += stride) { + _mm_store_si128((__m128i *)dst, dc_dup); + } +} + +// ----------------------------------------------------------------------------- +// DC_TOP + +static INLINE void dc_top_predictor_8xh(uint16_t *dst, ptrdiff_t stride, + int height, const uint16_t *above) { + const __m128i four = _mm_cvtsi32_si128(4); + const __m128i sum = dc_sum_8(above); + const __m128i dc = _mm_srli_epi16(_mm_add_epi16(sum, four), 3); + dc_store_8xh(dst, stride, height, &dc); +} + +void aom_highbd_dc_top_predictor_8x4_sse2(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, + const uint16_t *left, int bd) { + (void)left; + (void)bd; + dc_top_predictor_8xh(dst, stride, 4, above); +} + +void aom_highbd_dc_top_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, + const uint16_t *left, int bd) { + (void)left; + (void)bd; + dc_top_predictor_8xh(dst, stride, 8, above); +} + +void aom_highbd_dc_top_predictor_8x16_sse2(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, + const uint16_t *left, int bd) { + (void)left; + (void)bd; + dc_top_predictor_8xh(dst, stride, 16, above); +} + +// ----------------------------------------------------------------------------- +// DC_LEFT + +void aom_highbd_dc_left_predictor_8x4_sse2(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, + const uint16_t *left, int bd) { + const __m128i two = _mm_cvtsi32_si128(2); + const __m128i sum = dc_sum_4(left); + const __m128i dc = _mm_srli_epi16(_mm_add_epi16(sum, two), 2); + (void)above; + (void)bd; + dc_store_8xh(dst, stride, 4, &dc); +} + +void aom_highbd_dc_left_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, + const uint16_t *left, int bd) { + const __m128i four = _mm_cvtsi32_si128(4); + const __m128i sum = dc_sum_8(left); + const __m128i dc = _mm_srli_epi16(_mm_add_epi16(sum, four), 3); + (void)above; + (void)bd; + dc_store_8xh(dst, stride, 8, &dc); +} + +// Shared with DC 16xh +static INLINE __m128i dc_sum_16(const uint16_t *ref) { + const __m128i sum_lo = dc_sum_8(ref); + const __m128i sum_hi = dc_sum_8(ref + 8); + return _mm_add_epi16(sum_lo, sum_hi); +} + +void aom_highbd_dc_left_predictor_8x16_sse2(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, + const uint16_t *left, int bd) { + const __m128i eight = _mm_cvtsi32_si128(8); + const __m128i sum = dc_sum_16(left); + const __m128i dc = _mm_srli_epi16(_mm_add_epi16(sum, eight), 4); + (void)above; + (void)bd; + dc_store_8xh(dst, stride, 16, &dc); +} + +// ----------------------------------------------------------------------------- +// DC_128 + +static INLINE void dc_128_predictor_8xh(uint16_t *dst, ptrdiff_t stride, + int height, int bd) { + const __m128i dc = _mm_cvtsi32_si128(1 << (bd - 1)); + const __m128i dc_dup = _mm_shufflelo_epi16(dc, 0x0); + dc_store_8xh(dst, stride, height, &dc_dup); +} + +void aom_highbd_dc_128_predictor_8x4_sse2(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, + const uint16_t *left, int bd) { + (void)above; + (void)left; + dc_128_predictor_8xh(dst, stride, 4, bd); +} + +void aom_highbd_dc_128_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, + const uint16_t *left, int bd) { + (void)above; + (void)left; + dc_128_predictor_8xh(dst, stride, 8, bd); +} + +void aom_highbd_dc_128_predictor_8x16_sse2(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, + const uint16_t *left, int bd) { + (void)above; + (void)left; + dc_128_predictor_8xh(dst, stride, 16, bd); +} + +// ----------------------------------------------------------------------------- +// 16xh + +static INLINE void dc_store_16xh(uint16_t *dst, ptrdiff_t stride, int height, + const __m128i *dc) { + const __m128i dc_dup_lo = _mm_shufflelo_epi16(*dc, 0); + const __m128i dc_dup = _mm_unpacklo_epi64(dc_dup_lo, dc_dup_lo); + int i; + for (i = 0; i < height; ++i, dst += stride) { + _mm_store_si128((__m128i *)dst, dc_dup); + _mm_store_si128((__m128i *)(dst + 8), dc_dup); + } +} + +// ----------------------------------------------------------------------------- +// DC_LEFT + +void aom_highbd_dc_left_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, + const uint16_t *left, int bd) { + const __m128i four = _mm_cvtsi32_si128(4); + const __m128i sum = dc_sum_8(left); + const __m128i dc = _mm_srli_epi16(_mm_add_epi16(sum, four), 3); + (void)above; + (void)bd; + dc_store_16xh(dst, stride, 8, &dc); +} + +void aom_highbd_dc_left_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, + const uint16_t *left, int bd) { + const __m128i eight = _mm_cvtsi32_si128(8); + const __m128i sum = dc_sum_16(left); + const __m128i dc = _mm_srli_epi16(_mm_add_epi16(sum, eight), 4); + (void)above; + (void)bd; + dc_store_16xh(dst, stride, 16, &dc); +} + +// Shared with 32xh +static INLINE __m128i dc_sum_32(const uint16_t *ref) { + const __m128i zero = _mm_setzero_si128(); + const __m128i sum_a = dc_sum_16(ref); + const __m128i sum_b = dc_sum_16(ref + 16); + // 12 bit bd will outrange, so expand to 32 bit before adding final total + return _mm_add_epi32(_mm_unpacklo_epi16(sum_a, zero), + _mm_unpacklo_epi16(sum_b, zero)); +} + +void aom_highbd_dc_left_predictor_16x32_sse2(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, + const uint16_t *left, int bd) { + const __m128i sixteen = _mm_cvtsi32_si128(16); + const __m128i sum = dc_sum_32(left); + const __m128i dc = _mm_srli_epi32(_mm_add_epi32(sum, sixteen), 5); + (void)above; + (void)bd; + dc_store_16xh(dst, stride, 32, &dc); +} + +// ----------------------------------------------------------------------------- +// DC_TOP + +void aom_highbd_dc_top_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, + const uint16_t *left, int bd) { + const __m128i eight = _mm_cvtsi32_si128(8); + const __m128i sum = dc_sum_16(above); + const __m128i dc = _mm_srli_epi16(_mm_add_epi16(sum, eight), 4); + (void)left; + (void)bd; + dc_store_16xh(dst, stride, 8, &dc); +} + +void aom_highbd_dc_top_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, + const uint16_t *left, int bd) { + const __m128i eight = _mm_cvtsi32_si128(8); + const __m128i sum = dc_sum_16(above); + const __m128i dc = _mm_srli_epi16(_mm_add_epi16(sum, eight), 4); + (void)left; + (void)bd; + dc_store_16xh(dst, stride, 16, &dc); +} + +void aom_highbd_dc_top_predictor_16x32_sse2(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, + const uint16_t *left, int bd) { + const __m128i eight = _mm_cvtsi32_si128(8); + const __m128i sum = dc_sum_16(above); + const __m128i dc = _mm_srli_epi16(_mm_add_epi16(sum, eight), 4); + (void)left; + (void)bd; + dc_store_16xh(dst, stride, 32, &dc); +} + +// ----------------------------------------------------------------------------- +// DC_128 + +void aom_highbd_dc_128_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, + const uint16_t *left, int bd) { + const __m128i dc = _mm_cvtsi32_si128(1 << (bd - 1)); + const __m128i dc_dup = _mm_shufflelo_epi16(dc, 0x0); + (void)above; + (void)left; + dc_store_16xh(dst, stride, 8, &dc_dup); +} + +void aom_highbd_dc_128_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, + const uint16_t *left, int bd) { + const __m128i dc = _mm_cvtsi32_si128(1 << (bd - 1)); + const __m128i dc_dup = _mm_shufflelo_epi16(dc, 0x0); + (void)above; + (void)left; + dc_store_16xh(dst, stride, 16, &dc_dup); +} + +void aom_highbd_dc_128_predictor_16x32_sse2(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, + const uint16_t *left, int bd) { + const __m128i dc = _mm_cvtsi32_si128(1 << (bd - 1)); + const __m128i dc_dup = _mm_shufflelo_epi16(dc, 0x0); + (void)above; + (void)left; + dc_store_16xh(dst, stride, 32, &dc_dup); +} + +// ----------------------------------------------------------------------------- +// 32xh + +static INLINE void dc_store_32xh(uint16_t *dst, ptrdiff_t stride, int height, + const __m128i *dc) { + const __m128i dc_dup_lo = _mm_shufflelo_epi16(*dc, 0); + const __m128i dc_dup = _mm_unpacklo_epi64(dc_dup_lo, dc_dup_lo); + int i; + for (i = 0; i < height; ++i, dst += stride) { + _mm_store_si128((__m128i *)dst, dc_dup); + _mm_store_si128((__m128i *)(dst + 8), dc_dup); + _mm_store_si128((__m128i *)(dst + 16), dc_dup); + _mm_store_si128((__m128i *)(dst + 24), dc_dup); + } +} + +void aom_highbd_dc_left_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, + const uint16_t *left, int bd) { + const __m128i eight = _mm_cvtsi32_si128(8); + const __m128i sum = dc_sum_16(left); + const __m128i dc = _mm_srli_epi16(_mm_add_epi16(sum, eight), 4); + (void)above; + (void)bd; + dc_store_32xh(dst, stride, 16, &dc); +} + +void aom_highbd_dc_left_predictor_32x32_sse2(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, + const uint16_t *left, int bd) { + const __m128i sixteen = _mm_cvtsi32_si128(16); + const __m128i sum = dc_sum_32(left); + const __m128i dc = _mm_srli_epi32(_mm_add_epi32(sum, sixteen), 5); + (void)above; + (void)bd; + dc_store_32xh(dst, stride, 32, &dc); +} + +void aom_highbd_dc_top_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, + const uint16_t *left, int bd) { + const __m128i sixteen = _mm_cvtsi32_si128(16); + const __m128i sum = dc_sum_32(above); + const __m128i dc = _mm_srli_epi32(_mm_add_epi32(sum, sixteen), 5); + (void)left; + (void)bd; + dc_store_32xh(dst, stride, 16, &dc); +} + +void aom_highbd_dc_128_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, + const uint16_t *left, int bd) { + const __m128i dc = _mm_cvtsi32_si128(1 << (bd - 1)); + const __m128i dc_dup = _mm_shufflelo_epi16(dc, 0x0); + (void)above; + (void)left; + dc_store_32xh(dst, stride, 16, &dc_dup); +} + +void aom_highbd_dc_top_predictor_32x32_sse2(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, + const uint16_t *left, int bd) { + const __m128i sixteen = _mm_cvtsi32_si128(16); + const __m128i sum = dc_sum_32(above); + const __m128i dc = _mm_srli_epi32(_mm_add_epi32(sum, sixteen), 5); + (void)left; + (void)bd; + dc_store_32xh(dst, stride, 32, &dc); +} + +void aom_highbd_dc_128_predictor_32x32_sse2(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, + const uint16_t *left, int bd) { + const __m128i dc = _mm_cvtsi32_si128(1 << (bd - 1)); + const __m128i dc_dup = _mm_shufflelo_epi16(dc, 0x0); + (void)above; + (void)left; + dc_store_32xh(dst, stride, 32, &dc_dup); +} + +// ----------------------------------------------------------------------------- +// V_PRED + +void aom_highbd_v_predictor_4x8_sse2(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, + const uint16_t *left, int bd) { + (void)left; + (void)bd; + const __m128i above_u16 = _mm_loadl_epi64((const __m128i *)above); + int i; + for (i = 0; i < 2; ++i) { + _mm_storel_epi64((__m128i *)dst, above_u16); + _mm_storel_epi64((__m128i *)(dst + stride), above_u16); + _mm_storel_epi64((__m128i *)(dst + 2 * stride), above_u16); + _mm_storel_epi64((__m128i *)(dst + 3 * stride), above_u16); + dst += stride << 2; + } +} + +void aom_highbd_v_predictor_8x4_sse2(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, + const uint16_t *left, int bd) { + (void)left; + (void)bd; + const __m128i above_u16 = _mm_load_si128((const __m128i *)above); + _mm_store_si128((__m128i *)dst, above_u16); + _mm_store_si128((__m128i *)(dst + stride), above_u16); + _mm_store_si128((__m128i *)(dst + 2 * stride), above_u16); + _mm_store_si128((__m128i *)(dst + 3 * stride), above_u16); +} + +void aom_highbd_v_predictor_8x16_sse2(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, + const uint16_t *left, int bd) { + (void)left; + (void)bd; + const __m128i above_u16 = _mm_load_si128((const __m128i *)above); + int i; + for (i = 0; i < 4; ++i) { + _mm_store_si128((__m128i *)dst, above_u16); + _mm_store_si128((__m128i *)(dst + stride), above_u16); + _mm_store_si128((__m128i *)(dst + 2 * stride), above_u16); + _mm_store_si128((__m128i *)(dst + 3 * stride), above_u16); + dst += stride << 2; + } +} + +void aom_highbd_v_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, + const uint16_t *left, int bd) { + (void)left; + (void)bd; + const __m128i above0_u16 = _mm_load_si128((const __m128i *)above); + const __m128i above1_u16 = _mm_load_si128((const __m128i *)(above + 8)); + int i; + for (i = 0; i < 2; ++i) { + _mm_store_si128((__m128i *)dst, above0_u16); + _mm_store_si128((__m128i *)(dst + 8), above1_u16); + dst += stride; + _mm_store_si128((__m128i *)dst, above0_u16); + _mm_store_si128((__m128i *)(dst + 8), above1_u16); + dst += stride; + _mm_store_si128((__m128i *)dst, above0_u16); + _mm_store_si128((__m128i *)(dst + 8), above1_u16); + dst += stride; + _mm_store_si128((__m128i *)dst, above0_u16); + _mm_store_si128((__m128i *)(dst + 8), above1_u16); + dst += stride; + } +} + +void aom_highbd_v_predictor_16x32_sse2(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, + const uint16_t *left, int bd) { + (void)left; + (void)bd; + const __m128i above0_u16 = _mm_load_si128((const __m128i *)above); + const __m128i above1_u16 = _mm_load_si128((const __m128i *)(above + 8)); + int i; + for (i = 0; i < 8; ++i) { + _mm_store_si128((__m128i *)dst, above0_u16); + _mm_store_si128((__m128i *)(dst + 8), above1_u16); + dst += stride; + _mm_store_si128((__m128i *)dst, above0_u16); + _mm_store_si128((__m128i *)(dst + 8), above1_u16); + dst += stride; + _mm_store_si128((__m128i *)dst, above0_u16); + _mm_store_si128((__m128i *)(dst + 8), above1_u16); + dst += stride; + _mm_store_si128((__m128i *)dst, above0_u16); + _mm_store_si128((__m128i *)(dst + 8), above1_u16); + dst += stride; + } +} + +void aom_highbd_v_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, + const uint16_t *left, int bd) { + (void)left; + (void)bd; + const __m128i above0_u16 = _mm_load_si128((const __m128i *)above); + const __m128i above1_u16 = _mm_load_si128((const __m128i *)(above + 8)); + const __m128i above2_u16 = _mm_load_si128((const __m128i *)(above + 16)); + const __m128i above3_u16 = _mm_load_si128((const __m128i *)(above + 24)); + int i; + for (i = 0; i < 4; ++i) { + _mm_store_si128((__m128i *)dst, above0_u16); + _mm_store_si128((__m128i *)(dst + 8), above1_u16); + _mm_store_si128((__m128i *)(dst + 16), above2_u16); + _mm_store_si128((__m128i *)(dst + 24), above3_u16); + dst += stride; + _mm_store_si128((__m128i *)dst, above0_u16); + _mm_store_si128((__m128i *)(dst + 8), above1_u16); + _mm_store_si128((__m128i *)(dst + 16), above2_u16); + _mm_store_si128((__m128i *)(dst + 24), above3_u16); + dst += stride; + _mm_store_si128((__m128i *)dst, above0_u16); + _mm_store_si128((__m128i *)(dst + 8), above1_u16); + _mm_store_si128((__m128i *)(dst + 16), above2_u16); + _mm_store_si128((__m128i *)(dst + 24), above3_u16); + dst += stride; + _mm_store_si128((__m128i *)dst, above0_u16); + _mm_store_si128((__m128i *)(dst + 8), above1_u16); + _mm_store_si128((__m128i *)(dst + 16), above2_u16); + _mm_store_si128((__m128i *)(dst + 24), above3_u16); + dst += stride; + } +} + +// ----------------------------------------------------------------------------- +// DC_PRED + +void aom_highbd_dc_predictor_4x8_sse2(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, + const uint16_t *left, int bd) { + (void)bd; + const __m128i sum_above = dc_sum_4(above); + const __m128i sum_left = dc_sum_8(left); + const __m128i sum = _mm_add_epi16(sum_above, sum_left); + uint32_t sum32 = _mm_cvtsi128_si32(sum); + sum32 >>= 16; + sum32 += 6; + sum32 /= 12; + const __m128i row = _mm_set1_epi16((uint16_t)sum32); + int i; + for (i = 0; i < 4; ++i) { + _mm_storel_epi64((__m128i *)dst, row); + dst += stride; + _mm_storel_epi64((__m128i *)dst, row); + dst += stride; + } +} + +void aom_highbd_dc_predictor_8x4_sse2(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, + const uint16_t *left, int bd) { + (void)bd; + const __m128i sum_left = dc_sum_4(left); + const __m128i sum_above = dc_sum_8(above); + const __m128i sum = _mm_add_epi16(sum_above, sum_left); + uint32_t sum32 = _mm_cvtsi128_si32(sum); + sum32 >>= 16; + sum32 += 6; + sum32 /= 12; + const __m128i row = _mm_set1_epi16((uint16_t)sum32); + + _mm_store_si128((__m128i *)dst, row); + dst += stride; + _mm_store_si128((__m128i *)dst, row); + dst += stride; + _mm_store_si128((__m128i *)dst, row); + dst += stride; + _mm_store_si128((__m128i *)dst, row); +} + +void aom_highbd_dc_predictor_8x16_sse2(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, + const uint16_t *left, int bd) { + (void)bd; + __m128i sum_left = dc_sum_16(left); + __m128i sum_above = dc_sum_8(above); + const __m128i zero = _mm_setzero_si128(); + sum_left = _mm_unpacklo_epi16(sum_left, zero); + sum_above = _mm_unpacklo_epi16(sum_above, zero); + const __m128i sum = _mm_add_epi32(sum_left, sum_above); + uint32_t sum32 = _mm_cvtsi128_si32(sum); + sum32 += 12; + sum32 /= 24; + const __m128i row = _mm_set1_epi16((uint16_t)sum32); + int i; + for (i = 0; i < 4; ++i) { + _mm_store_si128((__m128i *)dst, row); + dst += stride; + _mm_store_si128((__m128i *)dst, row); + dst += stride; + _mm_store_si128((__m128i *)dst, row); + dst += stride; + _mm_store_si128((__m128i *)dst, row); + dst += stride; + } +} + +void aom_highbd_dc_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, + const uint16_t *left, int bd) { + (void)bd; + __m128i sum_left = dc_sum_8(left); + __m128i sum_above = dc_sum_16(above); + const __m128i zero = _mm_setzero_si128(); + sum_left = _mm_unpacklo_epi16(sum_left, zero); + sum_above = _mm_unpacklo_epi16(sum_above, zero); + const __m128i sum = _mm_add_epi32(sum_left, sum_above); + uint32_t sum32 = _mm_cvtsi128_si32(sum); + sum32 += 12; + sum32 /= 24; + const __m128i row = _mm_set1_epi16((uint16_t)sum32); + int i; + for (i = 0; i < 2; ++i) { + _mm_store_si128((__m128i *)dst, row); + _mm_store_si128((__m128i *)(dst + 8), row); + dst += stride; + _mm_store_si128((__m128i *)dst, row); + _mm_store_si128((__m128i *)(dst + 8), row); + dst += stride; + _mm_store_si128((__m128i *)dst, row); + _mm_store_si128((__m128i *)(dst + 8), row); + dst += stride; + _mm_store_si128((__m128i *)dst, row); + _mm_store_si128((__m128i *)(dst + 8), row); + dst += stride; + } +} + +void aom_highbd_dc_predictor_16x32_sse2(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, + const uint16_t *left, int bd) { + (void)bd; + __m128i sum_left = dc_sum_32(left); + __m128i sum_above = dc_sum_16(above); + const __m128i zero = _mm_setzero_si128(); + sum_above = _mm_unpacklo_epi16(sum_above, zero); + const __m128i sum = _mm_add_epi32(sum_left, sum_above); + uint32_t sum32 = _mm_cvtsi128_si32(sum); + sum32 += 24; + sum32 /= 48; + const __m128i row = _mm_set1_epi16((uint16_t)sum32); + int i; + for (i = 0; i < 8; ++i) { + _mm_store_si128((__m128i *)dst, row); + _mm_store_si128((__m128i *)(dst + 8), row); + dst += stride; + _mm_store_si128((__m128i *)dst, row); + _mm_store_si128((__m128i *)(dst + 8), row); + dst += stride; + _mm_store_si128((__m128i *)dst, row); + _mm_store_si128((__m128i *)(dst + 8), row); + dst += stride; + _mm_store_si128((__m128i *)dst, row); + _mm_store_si128((__m128i *)(dst + 8), row); + dst += stride; + } +} + +void aom_highbd_dc_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t stride, + const uint16_t *above, + const uint16_t *left, int bd) { + (void)bd; + __m128i sum_left = dc_sum_16(left); + __m128i sum_above = dc_sum_32(above); + const __m128i zero = _mm_setzero_si128(); + sum_left = _mm_unpacklo_epi16(sum_left, zero); + const __m128i sum = _mm_add_epi32(sum_left, sum_above); + uint32_t sum32 = _mm_cvtsi128_si32(sum); + sum32 += 24; + sum32 /= 48; + const __m128i row = _mm_set1_epi16((uint16_t)sum32); + int i; + for (i = 0; i < 4; ++i) { + _mm_store_si128((__m128i *)dst, row); + _mm_store_si128((__m128i *)(dst + 8), row); + _mm_store_si128((__m128i *)(dst + 16), row); + _mm_store_si128((__m128i *)(dst + 24), row); + dst += stride; + _mm_store_si128((__m128i *)dst, row); + _mm_store_si128((__m128i *)(dst + 8), row); + _mm_store_si128((__m128i *)(dst + 16), row); + _mm_store_si128((__m128i *)(dst + 24), row); + dst += stride; + _mm_store_si128((__m128i *)dst, row); + _mm_store_si128((__m128i *)(dst + 8), row); + _mm_store_si128((__m128i *)(dst + 16), row); + _mm_store_si128((__m128i *)(dst + 24), row); + dst += stride; + _mm_store_si128((__m128i *)dst, row); + _mm_store_si128((__m128i *)(dst + 8), row); + _mm_store_si128((__m128i *)(dst + 16), row); + _mm_store_si128((__m128i *)(dst + 24), row); + dst += stride; + } +} diff --git a/third_party/aom/aom_dsp/x86/highbd_intrapred_sse2_asm.asm b/third_party/aom/aom_dsp/x86/highbd_intrapred_sse2_asm.asm new file mode 100644 index 0000000000..91b3d126ca --- /dev/null +++ b/third_party/aom/aom_dsp/x86/highbd_intrapred_sse2_asm.asm @@ -0,0 +1,259 @@ +; +; Copyright (c) 2016, Alliance for Open Media. All rights reserved +; +; This source code is subject to the terms of the BSD 2 Clause License and +; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License +; was not distributed with this source code in the LICENSE file, you can +; obtain it at www.aomedia.org/license/software. If the Alliance for Open +; Media Patent License 1.0 was not distributed with this source code in the +; PATENTS file, you can obtain it at www.aomedia.org/license/patent. +; + +; + +%include "third_party/x86inc/x86inc.asm" + +SECTION_RODATA +pw_4: times 8 dw 4 +pw_8: times 8 dw 8 +pw_16: times 4 dd 16 +pw_32: times 4 dd 32 + +SECTION .text +INIT_XMM sse2 +cglobal highbd_dc_predictor_4x4, 4, 5, 4, dst, stride, above, left, goffset + GET_GOT goffsetq + + movq m0, [aboveq] + movq m2, [leftq] + paddw m0, m2 + pshuflw m1, m0, 0xe + paddw m0, m1 + pshuflw m1, m0, 0x1 + paddw m0, m1 + paddw m0, [GLOBAL(pw_4)] + psraw m0, 3 + pshuflw m0, m0, 0x0 + movq [dstq ], m0 + movq [dstq+strideq*2], m0 + lea dstq, [dstq+strideq*4] + movq [dstq ], m0 + movq [dstq+strideq*2], m0 + + RESTORE_GOT + RET + +INIT_XMM sse2 +cglobal highbd_dc_predictor_8x8, 4, 5, 4, dst, stride, above, left, goffset + GET_GOT goffsetq + + pxor m1, m1 + mova m0, [aboveq] + mova m2, [leftq] + DEFINE_ARGS dst, stride, stride3, one + mov oned, 0x00010001 + lea stride3q, [strideq*3] + movd m3, oned + pshufd m3, m3, 0x0 + paddw m0, m2 + pmaddwd m0, m3 + packssdw m0, m1 + pmaddwd m0, m3 + packssdw m0, m1 + pmaddwd m0, m3 + paddw m0, [GLOBAL(pw_8)] + psrlw m0, 4 + pshuflw m0, m0, 0x0 + punpcklqdq m0, m0 + mova [dstq ], m0 + mova [dstq+strideq*2 ], m0 + mova [dstq+strideq*4 ], m0 + mova [dstq+stride3q*2], m0 + lea dstq, [dstq+strideq*8] + mova [dstq ], m0 + mova [dstq+strideq*2 ], m0 + mova [dstq+strideq*4 ], m0 + mova [dstq+stride3q*2], m0 + + RESTORE_GOT + RET + +INIT_XMM sse2 +cglobal highbd_dc_predictor_16x16, 4, 5, 5, dst, stride, above, left, goffset + GET_GOT goffsetq + + pxor m1, m1 + mova m0, [aboveq] + mova m3, [aboveq+16] + mova m2, [leftq] + mova m4, [leftq+16] + DEFINE_ARGS dst, stride, stride3, lines4 + lea stride3q, [strideq*3] + mov lines4d, 4 + paddw m0, m2 + paddw m0, m3 + paddw m0, m4 + movhlps m2, m0 + paddw m0, m2 + punpcklwd m0, m1 + movhlps m2, m0 + paddd m0, m2 + punpckldq m0, m1 + movhlps m2, m0 + paddd m0, m2 + paddd m0, [GLOBAL(pw_16)] + psrad m0, 5 + pshuflw m0, m0, 0x0 + punpcklqdq m0, m0 +.loop: + mova [dstq ], m0 + mova [dstq +16], m0 + mova [dstq+strideq*2 ], m0 + mova [dstq+strideq*2 +16], m0 + mova [dstq+strideq*4 ], m0 + mova [dstq+strideq*4 +16], m0 + mova [dstq+stride3q*2 ], m0 + mova [dstq+stride3q*2+16], m0 + lea dstq, [dstq+strideq*8] + dec lines4d + jnz .loop + + RESTORE_GOT + REP_RET + +INIT_XMM sse2 +cglobal highbd_dc_predictor_32x32, 4, 5, 7, dst, stride, above, left, goffset + GET_GOT goffsetq + + mova m0, [aboveq] + mova m2, [aboveq+16] + mova m3, [aboveq+32] + mova m4, [aboveq+48] + paddw m0, m2 + paddw m3, m4 + mova m2, [leftq] + mova m4, [leftq+16] + mova m5, [leftq+32] + mova m6, [leftq+48] + paddw m2, m4 + paddw m5, m6 + paddw m0, m3 + paddw m2, m5 + pxor m1, m1 + paddw m0, m2 + DEFINE_ARGS dst, stride, stride3, lines4 + lea stride3q, [strideq*3] + mov lines4d, 8 + movhlps m2, m0 + paddw m0, m2 + punpcklwd m0, m1 + movhlps m2, m0 + paddd m0, m2 + punpckldq m0, m1 + movhlps m2, m0 + paddd m0, m2 + paddd m0, [GLOBAL(pw_32)] + psrad m0, 6 + pshuflw m0, m0, 0x0 + punpcklqdq m0, m0 +.loop: + mova [dstq ], m0 + mova [dstq +16 ], m0 + mova [dstq +32 ], m0 + mova [dstq +48 ], m0 + mova [dstq+strideq*2 ], m0 + mova [dstq+strideq*2+16 ], m0 + mova [dstq+strideq*2+32 ], m0 + mova [dstq+strideq*2+48 ], m0 + mova [dstq+strideq*4 ], m0 + mova [dstq+strideq*4+16 ], m0 + mova [dstq+strideq*4+32 ], m0 + mova [dstq+strideq*4+48 ], m0 + mova [dstq+stride3q*2 ], m0 + mova [dstq+stride3q*2 +16], m0 + mova [dstq+stride3q*2 +32], m0 + mova [dstq+stride3q*2 +48], m0 + lea dstq, [dstq+strideq*8] + dec lines4d + jnz .loop + + RESTORE_GOT + REP_RET + +INIT_XMM sse2 +cglobal highbd_v_predictor_4x4, 3, 3, 1, dst, stride, above + movq m0, [aboveq] + movq [dstq ], m0 + movq [dstq+strideq*2], m0 + lea dstq, [dstq+strideq*4] + movq [dstq ], m0 + movq [dstq+strideq*2], m0 + RET + +INIT_XMM sse2 +cglobal highbd_v_predictor_8x8, 3, 3, 1, dst, stride, above + mova m0, [aboveq] + DEFINE_ARGS dst, stride, stride3 + lea stride3q, [strideq*3] + mova [dstq ], m0 + mova [dstq+strideq*2 ], m0 + mova [dstq+strideq*4 ], m0 + mova [dstq+stride3q*2], m0 + lea dstq, [dstq+strideq*8] + mova [dstq ], m0 + mova [dstq+strideq*2 ], m0 + mova [dstq+strideq*4 ], m0 + mova [dstq+stride3q*2], m0 + RET + +INIT_XMM sse2 +cglobal highbd_v_predictor_16x16, 3, 4, 2, dst, stride, above + mova m0, [aboveq] + mova m1, [aboveq+16] + DEFINE_ARGS dst, stride, stride3, nlines4 + lea stride3q, [strideq*3] + mov nlines4d, 4 +.loop: + mova [dstq ], m0 + mova [dstq +16], m1 + mova [dstq+strideq*2 ], m0 + mova [dstq+strideq*2 +16], m1 + mova [dstq+strideq*4 ], m0 + mova [dstq+strideq*4 +16], m1 + mova [dstq+stride3q*2 ], m0 + mova [dstq+stride3q*2+16], m1 + lea dstq, [dstq+strideq*8] + dec nlines4d + jnz .loop + REP_RET + +INIT_XMM sse2 +cglobal highbd_v_predictor_32x32, 3, 4, 4, dst, stride, above + mova m0, [aboveq] + mova m1, [aboveq+16] + mova m2, [aboveq+32] + mova m3, [aboveq+48] + DEFINE_ARGS dst, stride, stride3, nlines4 + lea stride3q, [strideq*3] + mov nlines4d, 8 +.loop: + mova [dstq ], m0 + mova [dstq +16], m1 + mova [dstq +32], m2 + mova [dstq +48], m3 + mova [dstq+strideq*2 ], m0 + mova [dstq+strideq*2 +16], m1 + mova [dstq+strideq*2 +32], m2 + mova [dstq+strideq*2 +48], m3 + mova [dstq+strideq*4 ], m0 + mova [dstq+strideq*4 +16], m1 + mova [dstq+strideq*4 +32], m2 + mova [dstq+strideq*4 +48], m3 + mova [dstq+stride3q*2 ], m0 + mova [dstq+stride3q*2 +16], m1 + mova [dstq+stride3q*2 +32], m2 + mova [dstq+stride3q*2 +48], m3 + lea dstq, [dstq+strideq*8] + dec nlines4d + jnz .loop + REP_RET diff --git a/third_party/aom/aom_dsp/x86/highbd_loopfilter_avx2.c b/third_party/aom/aom_dsp/x86/highbd_loopfilter_avx2.c new file mode 100644 index 0000000000..c954da94e5 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/highbd_loopfilter_avx2.c @@ -0,0 +1,66 @@ +/* + * Copyright (c) 2017, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include + +#include "config/aom_dsp_rtcd.h" + +#include "aom_dsp/x86/common_avx2.h" +#include "aom_dsp/x86/lpf_common_sse2.h" +#include "aom/aom_integer.h" + +void aom_highbd_lpf_horizontal_14_dual_avx2( + uint16_t *s, int p, const uint8_t *blimit0, const uint8_t *limit0, + const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, + const uint8_t *thresh1, int bd) { + aom_highbd_lpf_horizontal_14_dual_sse2(s, p, blimit0, limit0, thresh0, + blimit1, limit1, thresh1, bd); +} + +void aom_highbd_lpf_vertical_14_dual_avx2( + uint16_t *s, int p, const uint8_t *blimit0, const uint8_t *limit0, + const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, + const uint8_t *thresh1, int bd) { + aom_highbd_lpf_vertical_14_dual_sse2(s, p, blimit0, limit0, thresh0, blimit1, + limit1, thresh1, bd); +} + +void aom_highbd_lpf_horizontal_4_dual_avx2( + uint16_t *s, int p, const uint8_t *blimit0, const uint8_t *limit0, + const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, + const uint8_t *thresh1, int bd) { + aom_highbd_lpf_horizontal_4_dual_sse2(s, p, blimit0, limit0, thresh0, blimit1, + limit1, thresh1, bd); +} + +void aom_highbd_lpf_horizontal_8_dual_avx2( + uint16_t *s, int p, const uint8_t *blimit0, const uint8_t *limit0, + const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, + const uint8_t *thresh1, int bd) { + aom_highbd_lpf_horizontal_8_dual_sse2(s, p, blimit0, limit0, thresh0, blimit1, + limit1, thresh1, bd); +} + +void aom_highbd_lpf_vertical_4_dual_avx2( + uint16_t *s, int p, const uint8_t *blimit0, const uint8_t *limit0, + const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, + const uint8_t *thresh1, int bd) { + aom_highbd_lpf_vertical_4_dual_sse2(s, p, blimit0, limit0, thresh0, blimit1, + limit1, thresh1, bd); +} + +void aom_highbd_lpf_vertical_8_dual_avx2( + uint16_t *s, int p, const uint8_t *blimit0, const uint8_t *limit0, + const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, + const uint8_t *thresh1, int bd) { + aom_highbd_lpf_vertical_8_dual_sse2(s, p, blimit0, limit0, thresh0, blimit1, + limit1, thresh1, bd); +} diff --git a/third_party/aom/aom_dsp/x86/highbd_loopfilter_sse2.c b/third_party/aom/aom_dsp/x86/highbd_loopfilter_sse2.c new file mode 100644 index 0000000000..097e0778ff --- /dev/null +++ b/third_party/aom/aom_dsp/x86/highbd_loopfilter_sse2.c @@ -0,0 +1,1697 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include // SSE2 + +#include "config/aom_dsp_rtcd.h" + +#include "aom_dsp/x86/lpf_common_sse2.h" + +static AOM_FORCE_INLINE void pixel_clamp(const __m128i *min, const __m128i *max, + __m128i *pixel) { + *pixel = _mm_min_epi16(*pixel, *max); + *pixel = _mm_max_epi16(*pixel, *min); +} + +static AOM_FORCE_INLINE __m128i abs_diff16(__m128i a, __m128i b) { + return _mm_or_si128(_mm_subs_epu16(a, b), _mm_subs_epu16(b, a)); +} + +static INLINE void get_limit(const uint8_t *bl, const uint8_t *l, + const uint8_t *t, int bd, __m128i *blt, + __m128i *lt, __m128i *thr, __m128i *t80_out) { + const int shift = bd - 8; + const __m128i zero = _mm_setzero_si128(); + + __m128i x = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)bl), zero); + *blt = _mm_slli_epi16(x, shift); + + x = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)l), zero); + *lt = _mm_slli_epi16(x, shift); + + x = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)t), zero); + *thr = _mm_slli_epi16(x, shift); + + *t80_out = _mm_set1_epi16(1 << (bd - 1)); +} + +static INLINE void get_limit_dual( + const uint8_t *_blimit0, const uint8_t *_limit0, const uint8_t *_thresh0, + const uint8_t *_blimit1, const uint8_t *_limit1, const uint8_t *_thresh1, + int bd, __m128i *blt_out, __m128i *lt_out, __m128i *thr_out, + __m128i *t80_out) { + const int shift = bd - 8; + const __m128i zero = _mm_setzero_si128(); + + __m128i x0 = + _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit0), zero); + __m128i x1 = + _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit1), zero); + x0 = _mm_unpacklo_epi64(x0, x1); + *blt_out = _mm_slli_epi16(x0, shift); + + x0 = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit0), zero); + x1 = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit1), zero); + x0 = _mm_unpacklo_epi64(x0, x1); + *lt_out = _mm_slli_epi16(x0, shift); + + x0 = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh0), zero); + x1 = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh1), zero); + x0 = _mm_unpacklo_epi64(x0, x1); + *thr_out = _mm_slli_epi16(x0, shift); + + *t80_out = _mm_set1_epi16(1 << (bd - 1)); +} + +static INLINE void load_highbd_pixel(const uint16_t *s, int size, int pitch, + __m128i *p, __m128i *q) { + int i; + for (i = 0; i < size; i++) { + p[i] = _mm_loadu_si128((__m128i *)(s - (i + 1) * pitch)); + q[i] = _mm_loadu_si128((__m128i *)(s + i * pitch)); + } +} + +static INLINE void highbd_filter_mask_dual(const __m128i *p, const __m128i *q, + const __m128i *l, const __m128i *bl, + __m128i *mask) { + __m128i abs_p0q0 = abs_diff16(p[0], q[0]); + __m128i abs_p1q1 = abs_diff16(p[1], q[1]); + abs_p0q0 = _mm_adds_epu16(abs_p0q0, abs_p0q0); + abs_p1q1 = _mm_srli_epi16(abs_p1q1, 1); + + const __m128i zero = _mm_setzero_si128(); + const __m128i one = _mm_set1_epi16(1); + const __m128i ffff = _mm_set1_epi16(0xFFFF); + + __m128i max = _mm_subs_epu16(_mm_adds_epu16(abs_p0q0, abs_p1q1), *bl); + max = _mm_xor_si128(_mm_cmpeq_epi16(max, zero), ffff); + max = _mm_and_si128(max, _mm_adds_epu16(*l, one)); + + int i; + for (i = 1; i < 4; ++i) { + max = _mm_max_epi16(max, abs_diff16(p[i], p[i - 1])); + max = _mm_max_epi16(max, abs_diff16(q[i], q[i - 1])); + } + max = _mm_subs_epu16(max, *l); + *mask = _mm_cmpeq_epi16(max, zero); // return ~mask +} + +static INLINE void highbd_hev_filter_mask_x_sse2(__m128i *pq, int x, + __m128i *p1p0, __m128i *q1q0, + __m128i *abs_p1p0, __m128i *l, + __m128i *bl, __m128i *t, + __m128i *hev, __m128i *mask) { + const __m128i zero = _mm_setzero_si128(); + const __m128i one = _mm_set1_epi16(1); + const __m128i ffff = _mm_set1_epi16(0xFFFF); + __m128i abs_p0q0_p1q1, abs_p0q0, abs_p1q1, abs_q1q0; + __m128i max, max01, h; + + *p1p0 = _mm_unpacklo_epi64(pq[0], pq[1]); + *q1q0 = _mm_unpackhi_epi64(pq[0], pq[1]); + + abs_p0q0_p1q1 = abs_diff16(*p1p0, *q1q0); + abs_p0q0 = _mm_adds_epu16(abs_p0q0_p1q1, abs_p0q0_p1q1); + abs_p0q0 = _mm_unpacklo_epi64(abs_p0q0, zero); + + abs_p1q1 = _mm_srli_si128(abs_p0q0_p1q1, 8); + abs_p1q1 = _mm_srli_epi16(abs_p1q1, 1); // divide by 2 + + max = _mm_subs_epu16(_mm_adds_epu16(abs_p0q0, abs_p1q1), *bl); + max = _mm_xor_si128(_mm_cmpeq_epi16(max, zero), ffff); + // mask |= (abs(*p0 - *q0) * 2 + abs(*p1 - *q1) / 2 > blimit) * -1; + // So taking maximums continues to work: + max = _mm_and_si128(max, _mm_adds_epu16(*l, one)); + + *abs_p1p0 = abs_diff16(pq[0], pq[1]); + abs_q1q0 = _mm_srli_si128(*abs_p1p0, 8); + max01 = _mm_max_epi16(*abs_p1p0, abs_q1q0); + // mask |= (abs(*p1 - *p0) > limit) * -1; + // mask |= (abs(*q1 - *q0) > limit) * -1; + h = _mm_subs_epu16(max01, *t); + + *hev = _mm_xor_si128(_mm_cmpeq_epi16(h, zero), ffff); + // replicate for the further "merged variables" usage + *hev = _mm_unpacklo_epi64(*hev, *hev); + + max = _mm_max_epi16(max, max01); + int i; + for (i = 2; i < x; ++i) { + max = _mm_max_epi16(max, abs_diff16(pq[i], pq[i - 1])); + } + max = _mm_max_epi16(max, _mm_srli_si128(max, 8)); + + max = _mm_subs_epu16(max, *l); + *mask = _mm_cmpeq_epi16(max, zero); // ~mask +} + +static INLINE void flat_mask_internal(const __m128i *th, const __m128i *pq, + int start, int end, __m128i *flat) { + int i; + __m128i max = _mm_max_epi16(abs_diff16(pq[start], pq[0]), + abs_diff16(pq[start + 1], pq[0])); + + for (i = start + 2; i < end; ++i) { + max = _mm_max_epi16(max, abs_diff16(pq[i], pq[0])); + } + max = _mm_max_epi16(max, _mm_srli_si128(max, 8)); + + __m128i ft; + ft = _mm_subs_epu16(max, *th); + + const __m128i zero = _mm_setzero_si128(); + *flat = _mm_cmpeq_epi16(ft, zero); +} + +static INLINE void flat_mask_internal_dual(const __m128i *th, const __m128i *p, + const __m128i *q, int start, int end, + __m128i *flat) { + int i; + __m128i max = + _mm_max_epi16(abs_diff16(q[start], q[0]), abs_diff16(p[start], p[0])); + + for (i = start + 1; i < end; ++i) { + max = _mm_max_epi16(max, abs_diff16(p[i], p[0])); + max = _mm_max_epi16(max, abs_diff16(q[i], q[0])); + } + + __m128i ft; + ft = _mm_subs_epu16(max, *th); + + const __m128i zero = _mm_setzero_si128(); + *flat = _mm_cmpeq_epi16(ft, zero); +} + +static INLINE void highbd_flat_mask4_sse2(__m128i *pq, __m128i *flat, + __m128i *flat2, int bd) { + // check the distance 1,2,3 against 0 + __m128i th = _mm_set1_epi16(1); + th = _mm_slli_epi16(th, bd - 8); + flat_mask_internal(&th, pq, 1, 4, flat); + flat_mask_internal(&th, pq, 4, 7, flat2); +} + +static INLINE void highbd_flat_mask4_dual_sse2(const __m128i *p, + const __m128i *q, __m128i *flat, + __m128i *flat2, int bd) { + // check the distance 1,2,3 against 0 + __m128i th = _mm_set1_epi16(1); + th = _mm_slli_epi16(th, bd - 8); + flat_mask_internal_dual(&th, p, q, 1, 4, flat); + flat_mask_internal_dual(&th, p, q, 4, 7, flat2); +} + +static AOM_FORCE_INLINE void highbd_filter4_sse2(__m128i *p1p0, __m128i *q1q0, + __m128i *hev, __m128i *mask, + __m128i *qs1qs0, + __m128i *ps1ps0, __m128i *t80, + int bd) { + const __m128i zero = _mm_setzero_si128(); + const __m128i one = _mm_set1_epi16(1); + const __m128i pmax = + _mm_subs_epi16(_mm_subs_epi16(_mm_slli_epi16(one, bd), one), *t80); + const __m128i pmin = _mm_subs_epi16(zero, *t80); + + const __m128i t3t4 = _mm_set_epi16(3, 3, 3, 3, 4, 4, 4, 4); + __m128i ps1ps0_work, qs1qs0_work, work; + __m128i filt, filter2filter1, filter2filt, filter1filt; + + ps1ps0_work = _mm_subs_epi16(*p1p0, *t80); + qs1qs0_work = _mm_subs_epi16(*q1q0, *t80); + + work = _mm_subs_epi16(ps1ps0_work, qs1qs0_work); + pixel_clamp(&pmin, &pmax, &work); + filt = _mm_and_si128(_mm_srli_si128(work, 8), *hev); + + filt = _mm_subs_epi16(filt, work); + filt = _mm_subs_epi16(filt, work); + filt = _mm_subs_epi16(filt, work); + // (aom_filter + 3 * (qs0 - ps0)) & mask + pixel_clamp(&pmin, &pmax, &filt); + filt = _mm_and_si128(filt, *mask); + filt = _mm_unpacklo_epi64(filt, filt); + + filter2filter1 = _mm_adds_epi16(filt, t3t4); /* signed_short_clamp */ + pixel_clamp(&pmin, &pmax, &filter2filter1); + filter2filter1 = _mm_srai_epi16(filter2filter1, 3); /* >> 3 */ + + filt = _mm_unpacklo_epi64(filter2filter1, filter2filter1); + + // filt >> 1 + filt = _mm_adds_epi16(filt, one); + filt = _mm_srai_epi16(filt, 1); + filt = _mm_andnot_si128(*hev, filt); + + filter2filt = _mm_unpackhi_epi64(filter2filter1, filt); + filter1filt = _mm_unpacklo_epi64(filter2filter1, filt); + + qs1qs0_work = _mm_subs_epi16(qs1qs0_work, filter1filt); + ps1ps0_work = _mm_adds_epi16(ps1ps0_work, filter2filt); + + pixel_clamp(&pmin, &pmax, &qs1qs0_work); + pixel_clamp(&pmin, &pmax, &ps1ps0_work); + + *qs1qs0 = _mm_adds_epi16(qs1qs0_work, *t80); + *ps1ps0 = _mm_adds_epi16(ps1ps0_work, *t80); +} + +static INLINE void highbd_filter4_dual_sse2(__m128i *p, __m128i *q, __m128i *ps, + __m128i *qs, const __m128i *mask, + const __m128i *th, int bd, + __m128i *t80) { + __m128i ps0 = _mm_subs_epi16(p[0], *t80); + __m128i ps1 = _mm_subs_epi16(p[1], *t80); + __m128i qs0 = _mm_subs_epi16(q[0], *t80); + __m128i qs1 = _mm_subs_epi16(q[1], *t80); + const __m128i one = _mm_set1_epi16(1); + const __m128i pmax = + _mm_subs_epi16(_mm_subs_epi16(_mm_slli_epi16(one, bd), one), *t80); + + const __m128i zero = _mm_setzero_si128(); + const __m128i pmin = _mm_subs_epi16(zero, *t80); + __m128i filter = _mm_subs_epi16(ps1, qs1); + pixel_clamp(&pmin, &pmax, &filter); + + // hev_filter + __m128i hev; + const __m128i abs_p1p0 = abs_diff16(p[1], p[0]); + const __m128i abs_q1q0 = abs_diff16(q[1], q[0]); + __m128i h = _mm_max_epi16(abs_p1p0, abs_q1q0); + h = _mm_subs_epu16(h, *th); + const __m128i ffff = _mm_cmpeq_epi16(h, h); + hev = _mm_xor_si128(_mm_cmpeq_epi16(h, zero), ffff); + + filter = _mm_and_si128(filter, hev); + + const __m128i x = _mm_subs_epi16(qs0, ps0); + filter = _mm_adds_epi16(filter, x); + filter = _mm_adds_epi16(filter, x); + filter = _mm_adds_epi16(filter, x); + pixel_clamp(&pmin, &pmax, &filter); + filter = _mm_and_si128(filter, *mask); + const __m128i t3 = _mm_set1_epi16(3); + const __m128i t4 = _mm_set1_epi16(4); + __m128i filter1 = _mm_adds_epi16(filter, t4); + __m128i filter2 = _mm_adds_epi16(filter, t3); + pixel_clamp(&pmin, &pmax, &filter1); + pixel_clamp(&pmin, &pmax, &filter2); + filter1 = _mm_srai_epi16(filter1, 3); + filter2 = _mm_srai_epi16(filter2, 3); + qs0 = _mm_subs_epi16(qs0, filter1); + pixel_clamp(&pmin, &pmax, &qs0); + ps0 = _mm_adds_epi16(ps0, filter2); + pixel_clamp(&pmin, &pmax, &ps0); + qs[0] = _mm_adds_epi16(qs0, *t80); + ps[0] = _mm_adds_epi16(ps0, *t80); + filter = _mm_adds_epi16(filter1, one); + filter = _mm_srai_epi16(filter, 1); + filter = _mm_andnot_si128(hev, filter); + qs1 = _mm_subs_epi16(qs1, filter); + pixel_clamp(&pmin, &pmax, &qs1); + ps1 = _mm_adds_epi16(ps1, filter); + pixel_clamp(&pmin, &pmax, &ps1); + qs[1] = _mm_adds_epi16(qs1, *t80); + ps[1] = _mm_adds_epi16(ps1, *t80); +} + +static AOM_FORCE_INLINE void highbd_lpf_internal_14_sse2( + __m128i *p, __m128i *q, __m128i *pq, const unsigned char *blt, + const unsigned char *lt, const unsigned char *thr, int bd) { + int i; + const __m128i zero = _mm_setzero_si128(); + __m128i blimit, limit, thresh; + __m128i t80; + get_limit(blt, lt, thr, bd, &blimit, &limit, &thresh, &t80); + + for (i = 0; i < 7; i++) { + pq[i] = _mm_unpacklo_epi64(p[i], q[i]); + } + __m128i mask, hevhev; + __m128i p1p0, q1q0, abs_p1p0; + + highbd_hev_filter_mask_x_sse2(pq, 4, &p1p0, &q1q0, &abs_p1p0, &limit, &blimit, + &thresh, &hevhev, &mask); + + __m128i ps0ps1, qs0qs1; + // filter4 + highbd_filter4_sse2(&p1p0, &q1q0, &hevhev, &mask, &qs0qs1, &ps0ps1, &t80, bd); + + __m128i flat, flat2; + highbd_flat_mask4_sse2(pq, &flat, &flat2, bd); + + flat = _mm_and_si128(flat, mask); + flat2 = _mm_and_si128(flat2, flat); + + // replicate for the further "merged variables" usage + flat = _mm_unpacklo_epi64(flat, flat); + flat2 = _mm_unpacklo_epi64(flat2, flat2); + + // flat and wide flat calculations + + // if flat ==0 then flat2 is zero as well and we don't need any calc below + // sse4.1 if (0==_mm_test_all_zeros(flat,ff)) + if (0xffff != _mm_movemask_epi8(_mm_cmpeq_epi16(flat, zero))) { + __m128i flat_p[3], flat_q[3], flat_pq[3]; + __m128i flat2_p[6], flat2_q[6]; + __m128i flat2_pq[6]; + __m128i sum_p6, sum_p3; + const __m128i eight = _mm_set1_epi16(8); + const __m128i four = _mm_set1_epi16(4); + + __m128i work0, work0_0, work0_1, sum_p_0; + __m128i sum_p = _mm_add_epi16(pq[5], _mm_add_epi16(pq[4], pq[3])); + __m128i sum_lp = _mm_add_epi16(pq[0], _mm_add_epi16(pq[2], pq[1])); + sum_p = _mm_add_epi16(sum_p, sum_lp); + + __m128i sum_lq = _mm_srli_si128(sum_lp, 8); + __m128i sum_q = _mm_srli_si128(sum_p, 8); + + sum_p_0 = _mm_add_epi16(eight, _mm_add_epi16(sum_p, sum_q)); + sum_lp = _mm_add_epi16(four, _mm_add_epi16(sum_lp, sum_lq)); + + flat_p[0] = _mm_add_epi16(sum_lp, _mm_add_epi16(pq[3], pq[0])); + flat_q[0] = _mm_add_epi16(sum_lp, _mm_add_epi16(q[3], q[0])); + + sum_p6 = _mm_add_epi16(pq[6], pq[6]); + sum_p3 = _mm_add_epi16(pq[3], pq[3]); + + sum_q = _mm_sub_epi16(sum_p_0, pq[5]); + sum_p = _mm_sub_epi16(sum_p_0, q[5]); + + work0_0 = _mm_add_epi16(_mm_add_epi16(pq[6], pq[0]), pq[1]); + work0_1 = _mm_add_epi16(sum_p6, + _mm_add_epi16(pq[1], _mm_add_epi16(pq[2], pq[0]))); + + sum_lq = _mm_sub_epi16(sum_lp, pq[2]); + sum_lp = _mm_sub_epi16(sum_lp, q[2]); + + work0 = _mm_add_epi16(sum_p3, pq[1]); + flat_p[1] = _mm_add_epi16(sum_lp, work0); + flat_q[1] = _mm_add_epi16(sum_lq, _mm_srli_si128(work0, 8)); + + flat_pq[0] = _mm_srli_epi16(_mm_unpacklo_epi64(flat_p[0], flat_q[0]), 3); + flat_pq[1] = _mm_srli_epi16(_mm_unpacklo_epi64(flat_p[1], flat_q[1]), 3); + + sum_lp = _mm_sub_epi16(sum_lp, q[1]); + sum_lq = _mm_sub_epi16(sum_lq, pq[1]); + + sum_p3 = _mm_add_epi16(sum_p3, pq[3]); + work0 = _mm_add_epi16(sum_p3, pq[2]); + + flat_p[2] = _mm_add_epi16(sum_lp, work0); + flat_q[2] = _mm_add_epi16(sum_lq, _mm_srli_si128(work0, 8)); + flat_pq[2] = _mm_srli_epi16(_mm_unpacklo_epi64(flat_p[2], flat_q[2]), 3); + + int flat2_mask = + (0xffff != _mm_movemask_epi8(_mm_cmpeq_epi16(flat2, zero))); + if (flat2_mask) { + flat2_p[0] = _mm_add_epi16(sum_p_0, _mm_add_epi16(work0_0, q[0])); + flat2_q[0] = _mm_add_epi16( + sum_p_0, _mm_add_epi16(_mm_srli_si128(work0_0, 8), pq[0])); + + flat2_p[1] = _mm_add_epi16(sum_p, work0_1); + flat2_q[1] = _mm_add_epi16(sum_q, _mm_srli_si128(work0_1, 8)); + + flat2_pq[0] = + _mm_srli_epi16(_mm_unpacklo_epi64(flat2_p[0], flat2_q[0]), 4); + flat2_pq[1] = + _mm_srli_epi16(_mm_unpacklo_epi64(flat2_p[1], flat2_q[1]), 4); + + sum_p = _mm_sub_epi16(sum_p, q[4]); + sum_q = _mm_sub_epi16(sum_q, pq[4]); + + sum_p6 = _mm_add_epi16(sum_p6, pq[6]); + work0 = _mm_add_epi16(sum_p6, + _mm_add_epi16(pq[2], _mm_add_epi16(pq[3], pq[1]))); + flat2_p[2] = _mm_add_epi16(sum_p, work0); + flat2_q[2] = _mm_add_epi16(sum_q, _mm_srli_si128(work0, 8)); + flat2_pq[2] = + _mm_srli_epi16(_mm_unpacklo_epi64(flat2_p[2], flat2_q[2]), 4); + + sum_p6 = _mm_add_epi16(sum_p6, pq[6]); + sum_p = _mm_sub_epi16(sum_p, q[3]); + sum_q = _mm_sub_epi16(sum_q, pq[3]); + + work0 = _mm_add_epi16(sum_p6, + _mm_add_epi16(pq[3], _mm_add_epi16(pq[4], pq[2]))); + flat2_p[3] = _mm_add_epi16(sum_p, work0); + flat2_q[3] = _mm_add_epi16(sum_q, _mm_srli_si128(work0, 8)); + flat2_pq[3] = + _mm_srli_epi16(_mm_unpacklo_epi64(flat2_p[3], flat2_q[3]), 4); + + sum_p6 = _mm_add_epi16(sum_p6, pq[6]); + sum_p = _mm_sub_epi16(sum_p, q[2]); + sum_q = _mm_sub_epi16(sum_q, pq[2]); + + work0 = _mm_add_epi16(sum_p6, + _mm_add_epi16(pq[4], _mm_add_epi16(pq[5], pq[3]))); + flat2_p[4] = _mm_add_epi16(sum_p, work0); + flat2_q[4] = _mm_add_epi16(sum_q, _mm_srli_si128(work0, 8)); + flat2_pq[4] = + _mm_srli_epi16(_mm_unpacklo_epi64(flat2_p[4], flat2_q[4]), 4); + + sum_p6 = _mm_add_epi16(sum_p6, pq[6]); + sum_p = _mm_sub_epi16(sum_p, q[1]); + sum_q = _mm_sub_epi16(sum_q, pq[1]); + + work0 = _mm_add_epi16(sum_p6, + _mm_add_epi16(pq[5], _mm_add_epi16(pq[6], pq[4]))); + flat2_p[5] = _mm_add_epi16(sum_p, work0); + flat2_q[5] = _mm_add_epi16(sum_q, _mm_srli_si128(work0, 8)); + flat2_pq[5] = + _mm_srli_epi16(_mm_unpacklo_epi64(flat2_p[5], flat2_q[5]), 4); + } // flat2 + // ~~~~~~~~~~ apply flat ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ + // highbd_filter8 + pq[0] = _mm_unpacklo_epi64(ps0ps1, qs0qs1); + pq[1] = _mm_unpackhi_epi64(ps0ps1, qs0qs1); + + for (i = 0; i < 3; i++) { + pq[i] = _mm_andnot_si128(flat, pq[i]); + flat_pq[i] = _mm_and_si128(flat, flat_pq[i]); + pq[i] = _mm_or_si128(pq[i], flat_pq[i]); + } + + // wide flat + // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ + if (flat2_mask) { + for (i = 0; i < 6; i++) { + pq[i] = _mm_andnot_si128(flat2, pq[i]); + flat2_pq[i] = _mm_and_si128(flat2, flat2_pq[i]); + pq[i] = _mm_or_si128(pq[i], flat2_pq[i]); // full list of pq values + } + } + } else { + pq[0] = _mm_unpacklo_epi64(ps0ps1, qs0qs1); + pq[1] = _mm_unpackhi_epi64(ps0ps1, qs0qs1); + } +} + +void aom_highbd_lpf_horizontal_14_sse2(uint16_t *s, int pitch, + const uint8_t *blt, const uint8_t *lt, + const uint8_t *thr, int bd) { + __m128i p[7], q[7], pq[7]; + int i; + + for (i = 0; i < 7; i++) { + p[i] = _mm_loadl_epi64((__m128i *)(s - (i + 1) * pitch)); + q[i] = _mm_loadl_epi64((__m128i *)(s + i * pitch)); + } + + highbd_lpf_internal_14_sse2(p, q, pq, blt, lt, thr, bd); + + for (i = 0; i < 6; i++) { + _mm_storel_epi64((__m128i *)(s - (i + 1) * pitch), pq[i]); + _mm_storel_epi64((__m128i *)(s + i * pitch), _mm_srli_si128(pq[i], 8)); + } +} + +static AOM_FORCE_INLINE void highbd_lpf_internal_14_dual_sse2( + __m128i *p, __m128i *q, const uint8_t *blt0, const uint8_t *lt0, + const uint8_t *thr0, const uint8_t *blt1, const uint8_t *lt1, + const uint8_t *thr1, int bd) { + __m128i blimit, limit, thresh, t80; + const __m128i zero = _mm_setzero_si128(); + + get_limit_dual(blt0, lt0, thr0, blt1, lt1, thr1, bd, &blimit, &limit, &thresh, + &t80); + __m128i mask; + highbd_filter_mask_dual(p, q, &limit, &blimit, &mask); + __m128i flat, flat2; + highbd_flat_mask4_dual_sse2(p, q, &flat, &flat2, bd); + + flat = _mm_and_si128(flat, mask); + flat2 = _mm_and_si128(flat2, flat); + __m128i ps[2], qs[2]; + highbd_filter4_dual_sse2(p, q, ps, qs, &mask, &thresh, bd, &t80); + // flat and wide flat calculations + + // if flat ==0 then flat2 is zero as well and we don't need any calc below + // sse4.1 if (0==_mm_test_all_zeros(flat,ff)) + if (0xffff != _mm_movemask_epi8(_mm_cmpeq_epi16(flat, zero))) { + __m128i flat_p[3], flat_q[3]; + __m128i flat2_p[6], flat2_q[6]; + const __m128i eight = _mm_set1_epi16(8); + const __m128i four = _mm_set1_epi16(4); + __m128i sum_p_0 = _mm_add_epi16(p[5], _mm_add_epi16(p[4], p[3])); + __m128i sum_q = _mm_add_epi16(q[5], _mm_add_epi16(q[4], q[3])); + __m128i sum_lp = _mm_add_epi16(p[0], _mm_add_epi16(p[2], p[1])); + sum_p_0 = _mm_add_epi16(sum_p_0, sum_lp); + __m128i sum_lq = _mm_add_epi16(q[0], _mm_add_epi16(q[2], q[1])); + sum_q = _mm_add_epi16(sum_q, sum_lq); + sum_p_0 = _mm_add_epi16(eight, _mm_add_epi16(sum_p_0, sum_q)); + sum_lp = _mm_add_epi16(four, _mm_add_epi16(sum_lp, sum_lq)); + flat_p[0] = + _mm_srli_epi16(_mm_add_epi16(sum_lp, _mm_add_epi16(p[3], p[0])), 3); + flat_q[0] = + _mm_srli_epi16(_mm_add_epi16(sum_lp, _mm_add_epi16(q[3], q[0])), 3); + __m128i sum_p6 = _mm_add_epi16(p[6], p[6]); + __m128i sum_q6 = _mm_add_epi16(q[6], q[6]); + __m128i sum_p3 = _mm_add_epi16(p[3], p[3]); + __m128i sum_q3 = _mm_add_epi16(q[3], q[3]); + + sum_q = _mm_sub_epi16(sum_p_0, p[5]); + __m128i sum_p = _mm_sub_epi16(sum_p_0, q[5]); + + sum_lq = _mm_sub_epi16(sum_lp, p[2]); + sum_lp = _mm_sub_epi16(sum_lp, q[2]); + flat_p[1] = + _mm_srli_epi16(_mm_add_epi16(sum_lp, _mm_add_epi16(sum_p3, p[1])), 3); + flat_q[1] = + _mm_srli_epi16(_mm_add_epi16(sum_lq, _mm_add_epi16(sum_q3, q[1])), 3); + + sum_lp = _mm_sub_epi16(sum_lp, q[1]); + sum_lq = _mm_sub_epi16(sum_lq, p[1]); + sum_p3 = _mm_add_epi16(sum_p3, p[3]); + sum_q3 = _mm_add_epi16(sum_q3, q[3]); + flat_p[2] = + _mm_srli_epi16(_mm_add_epi16(sum_lp, _mm_add_epi16(sum_p3, p[2])), 3); + flat_q[2] = + _mm_srli_epi16(_mm_add_epi16(sum_lq, _mm_add_epi16(sum_q3, q[2])), 3); + + int flat2_mask = + (0xffff != _mm_movemask_epi8(_mm_cmpeq_epi16(flat2, zero))); + if (flat2_mask) { + flat2_p[0] = _mm_srli_epi16( + _mm_add_epi16(sum_p_0, _mm_add_epi16(_mm_add_epi16(p[6], p[0]), + _mm_add_epi16(p[1], q[0]))), + 4); + flat2_q[0] = _mm_srli_epi16( + _mm_add_epi16(sum_p_0, _mm_add_epi16(_mm_add_epi16(q[6], q[0]), + _mm_add_epi16(p[0], q[1]))), + 4); + + flat2_p[1] = _mm_srli_epi16( + _mm_add_epi16( + sum_p, + _mm_add_epi16(sum_p6, + _mm_add_epi16(p[1], _mm_add_epi16(p[2], p[0])))), + 4); + flat2_q[1] = _mm_srli_epi16( + _mm_add_epi16( + sum_q, + _mm_add_epi16(sum_q6, + _mm_add_epi16(q[1], _mm_add_epi16(q[0], q[2])))), + 4); + sum_p6 = _mm_add_epi16(sum_p6, p[6]); + sum_q6 = _mm_add_epi16(sum_q6, q[6]); + sum_p = _mm_sub_epi16(sum_p, q[4]); + sum_q = _mm_sub_epi16(sum_q, p[4]); + flat2_p[2] = _mm_srli_epi16( + _mm_add_epi16( + sum_p, + _mm_add_epi16(sum_p6, + _mm_add_epi16(p[2], _mm_add_epi16(p[3], p[1])))), + 4); + flat2_q[2] = _mm_srli_epi16( + _mm_add_epi16( + sum_q, + _mm_add_epi16(sum_q6, + _mm_add_epi16(q[2], _mm_add_epi16(q[1], q[3])))), + 4); + sum_p6 = _mm_add_epi16(sum_p6, p[6]); + sum_q6 = _mm_add_epi16(sum_q6, q[6]); + sum_p = _mm_sub_epi16(sum_p, q[3]); + sum_q = _mm_sub_epi16(sum_q, p[3]); + flat2_p[3] = _mm_srli_epi16( + _mm_add_epi16( + sum_p, + _mm_add_epi16(sum_p6, + _mm_add_epi16(p[3], _mm_add_epi16(p[4], p[2])))), + 4); + flat2_q[3] = _mm_srli_epi16( + _mm_add_epi16( + sum_q, + _mm_add_epi16(sum_q6, + _mm_add_epi16(q[3], _mm_add_epi16(q[2], q[4])))), + 4); + sum_p6 = _mm_add_epi16(sum_p6, p[6]); + sum_q6 = _mm_add_epi16(sum_q6, q[6]); + sum_p = _mm_sub_epi16(sum_p, q[2]); + sum_q = _mm_sub_epi16(sum_q, p[2]); + flat2_p[4] = _mm_srli_epi16( + _mm_add_epi16( + sum_p, + _mm_add_epi16(sum_p6, + _mm_add_epi16(p[4], _mm_add_epi16(p[5], p[3])))), + 4); + flat2_q[4] = _mm_srli_epi16( + _mm_add_epi16( + sum_q, + _mm_add_epi16(sum_q6, + _mm_add_epi16(q[4], _mm_add_epi16(q[3], q[5])))), + 4); + sum_p6 = _mm_add_epi16(sum_p6, p[6]); + sum_q6 = _mm_add_epi16(sum_q6, q[6]); + sum_p = _mm_sub_epi16(sum_p, q[1]); + sum_q = _mm_sub_epi16(sum_q, p[1]); + flat2_p[5] = _mm_srli_epi16( + _mm_add_epi16( + sum_p, + _mm_add_epi16(sum_p6, + _mm_add_epi16(p[5], _mm_add_epi16(p[6], p[4])))), + 4); + flat2_q[5] = _mm_srli_epi16( + _mm_add_epi16( + sum_q, + _mm_add_epi16(sum_q6, + _mm_add_epi16(q[5], _mm_add_epi16(q[4], q[6])))), + 4); + } + // highbd_filter8 + int i; + for (i = 0; i < 2; i++) { + ps[i] = _mm_andnot_si128(flat, ps[i]); + flat_p[i] = _mm_and_si128(flat, flat_p[i]); + p[i] = _mm_or_si128(ps[i], flat_p[i]); + qs[i] = _mm_andnot_si128(flat, qs[i]); + flat_q[i] = _mm_and_si128(flat, flat_q[i]); + q[i] = _mm_or_si128(qs[i], flat_q[i]); + } + p[2] = _mm_andnot_si128(flat, p[2]); + // p2 remains unchanged if !(flat && mask) + flat_p[2] = _mm_and_si128(flat, flat_p[2]); + // when (flat && mask) + p[2] = _mm_or_si128(p[2], flat_p[2]); // full list of p2 values + q[2] = _mm_andnot_si128(flat, q[2]); + flat_q[2] = _mm_and_si128(flat, flat_q[2]); + q[2] = _mm_or_si128(q[2], flat_q[2]); // full list of q2 values + + for (i = 0; i < 2; i++) { + ps[i] = _mm_andnot_si128(flat, ps[i]); + flat_p[i] = _mm_and_si128(flat, flat_p[i]); + p[i] = _mm_or_si128(ps[i], flat_p[i]); + qs[i] = _mm_andnot_si128(flat, qs[i]); + flat_q[i] = _mm_and_si128(flat, flat_q[i]); + q[i] = _mm_or_si128(qs[i], flat_q[i]); + } + // highbd_filter16 + if (flat2_mask) { + for (i = 0; i < 6; i++) { + // p[i] remains unchanged if !(flat2 && flat && mask) + p[i] = _mm_andnot_si128(flat2, p[i]); + flat2_p[i] = _mm_and_si128(flat2, flat2_p[i]); + // get values for when (flat2 && flat && mask) + p[i] = _mm_or_si128(p[i], flat2_p[i]); // full list of p values + q[i] = _mm_andnot_si128(flat2, q[i]); + flat2_q[i] = _mm_and_si128(flat2, flat2_q[i]); + q[i] = _mm_or_si128(q[i], flat2_q[i]); + } + } + } else { + p[0] = ps[0]; + q[0] = qs[0]; + p[1] = ps[1]; + q[1] = qs[1]; + } +} + +void aom_highbd_lpf_horizontal_14_dual_sse2( + uint16_t *s, int pitch, const uint8_t *_blimit0, const uint8_t *_limit0, + const uint8_t *_thresh0, const uint8_t *_blimit1, const uint8_t *_limit1, + const uint8_t *_thresh1, int bd) { + __m128i p[7], q[7]; + int i; + load_highbd_pixel(s, 7, pitch, p, q); + + highbd_lpf_internal_14_dual_sse2(p, q, _blimit0, _limit0, _thresh0, _blimit1, + _limit1, _thresh1, bd); + + for (i = 0; i < 6; i++) { + _mm_store_si128((__m128i *)(s - (i + 1) * pitch), p[i]); + _mm_store_si128((__m128i *)(s + i * pitch), q[i]); + } +} + +static AOM_FORCE_INLINE void highbd_lpf_internal_6_sse2( + __m128i *p2, __m128i *p1, __m128i *p0, __m128i *q0, __m128i *q1, + __m128i *q2, __m128i *p1p0_out, __m128i *q1q0_out, const uint8_t *_blimit, + const uint8_t *_limit, const uint8_t *_thresh, int bd) { + __m128i blimit, limit, thresh; + __m128i mask, hev, flat; + __m128i pq[3]; + __m128i p1p0, q1q0, abs_p1p0, ps1ps0, qs1qs0; + __m128i flat_p1p0, flat_q0q1; + + pq[0] = _mm_unpacklo_epi64(*p0, *q0); + pq[1] = _mm_unpacklo_epi64(*p1, *q1); + pq[2] = _mm_unpacklo_epi64(*p2, *q2); + + const __m128i zero = _mm_setzero_si128(); + const __m128i four = _mm_set1_epi16(4); + __m128i t80; + const __m128i one = _mm_set1_epi16(0x1); + + get_limit(_blimit, _limit, _thresh, bd, &blimit, &limit, &thresh, &t80); + + highbd_hev_filter_mask_x_sse2(pq, 3, &p1p0, &q1q0, &abs_p1p0, &limit, &blimit, + &thresh, &hev, &mask); + + // lp filter + highbd_filter4_sse2(&p1p0, &q1q0, &hev, &mask, q1q0_out, p1p0_out, &t80, bd); + + // flat_mask + flat = _mm_max_epi16(abs_diff16(pq[2], pq[0]), abs_p1p0); + flat = _mm_max_epi16(flat, _mm_srli_si128(flat, 8)); + + flat = _mm_subs_epu16(flat, _mm_slli_epi16(one, bd - 8)); + + flat = _mm_cmpeq_epi16(flat, zero); + flat = _mm_and_si128(flat, mask); + // replicate for the further "merged variables" usage + flat = _mm_unpacklo_epi64(flat, flat); + + // 5 tap filter + // need it only if flat !=0 + if (0xffff != _mm_movemask_epi8(_mm_cmpeq_epi16(flat, zero))) { + __m128i workp_a, workp_b, workp_c; + __m128i pq0x2_pq1, pq1_pq2; + + // op1 + pq0x2_pq1 = + _mm_add_epi16(_mm_add_epi16(pq[0], pq[0]), pq[1]); // p0 *2 + p1 + pq1_pq2 = _mm_add_epi16(pq[1], pq[2]); // p1 + p2 + workp_a = _mm_add_epi16(_mm_add_epi16(pq0x2_pq1, four), + pq1_pq2); // p2 + p0 * 2 + p1 * 2 + 4 + + workp_b = _mm_add_epi16(_mm_add_epi16(pq[2], pq[2]), *q0); + workp_b = + _mm_add_epi16(workp_a, workp_b); // p2 * 3 + p1 * 2 + p0 * 2 + q0 + 4 + + // op0 + workp_c = _mm_srli_si128(pq0x2_pq1, 8); // q0 * 2 + q1 + workp_a = _mm_add_epi16(workp_a, + workp_c); // p2 + p0 * 2 + p1 * 2 + q0 * 2 + q1 + 4 + workp_b = _mm_unpacklo_epi64(workp_a, workp_b); + flat_p1p0 = _mm_srli_epi16(workp_b, 3); + + // oq0 + workp_a = _mm_sub_epi16(_mm_sub_epi16(workp_a, pq[2]), + pq[1]); // p0 * 2 + p1 + q0 * 2 + q1 + 4 + workp_b = _mm_srli_si128(pq1_pq2, 8); + workp_a = _mm_add_epi16( + workp_a, workp_b); // p0 * 2 + p1 + q0 * 2 + q1 * 2 + q2 + 4 + // workp_shft0 = _mm_srli_epi16(workp_a, 3); + + // oq1 + workp_c = _mm_sub_epi16(_mm_sub_epi16(workp_a, pq[1]), + pq[0]); // p0 + q0 * 2 + q1 * 2 + q2 + 4 + workp_b = _mm_add_epi16(*q2, *q2); + workp_b = + _mm_add_epi16(workp_c, workp_b); // p0 + q0 * 2 + q1 * 2 + q2 * 3 + 4 + + workp_a = _mm_unpacklo_epi64(workp_a, workp_b); + flat_q0q1 = _mm_srli_epi16(workp_a, 3); + + qs1qs0 = _mm_andnot_si128(flat, *q1q0_out); + q1q0 = _mm_and_si128(flat, flat_q0q1); + *q1q0_out = _mm_or_si128(qs1qs0, q1q0); + + ps1ps0 = _mm_andnot_si128(flat, *p1p0_out); + p1p0 = _mm_and_si128(flat, flat_p1p0); + *p1p0_out = _mm_or_si128(ps1ps0, p1p0); + } +} + +static AOM_FORCE_INLINE void highbd_lpf_internal_6_dual_sse2( + __m128i *p2, __m128i *p1, __m128i *p0, __m128i *q0, __m128i *q1, + __m128i *q2, const unsigned char *_blimit0, const unsigned char *_limit0, + const unsigned char *_thresh0, const unsigned char *_blimit1, + const unsigned char *_limit1, const unsigned char *_thresh1, int bd) { + const __m128i zero = _mm_setzero_si128(); + __m128i blimit0, limit0, thresh0; + __m128i t80; + __m128i mask, flat, work; + __m128i abs_p1q1, abs_p0q0, abs_p1p0, abs_p2p1, abs_q1q0, abs_q2q1; + __m128i op1, op0, oq0, oq1; + const __m128i four = _mm_set1_epi16(4); + const __m128i one = _mm_set1_epi16(0x1); + const __m128i ffff = _mm_cmpeq_epi16(one, one); + + get_limit_dual(_blimit0, _limit0, _thresh0, _blimit1, _limit1, _thresh1, bd, + &blimit0, &limit0, &thresh0, &t80); + + abs_p2p1 = abs_diff16(*p2, *p1); + abs_p1p0 = abs_diff16(*p1, *p0); + abs_q1q0 = abs_diff16(*q1, *q0); + abs_q2q1 = abs_diff16(*q2, *q1); + + abs_p0q0 = abs_diff16(*p0, *q0); + abs_p1q1 = abs_diff16(*p1, *q1); + + abs_p0q0 = _mm_adds_epu16(abs_p0q0, abs_p0q0); + abs_p1q1 = _mm_srli_epi16(abs_p1q1, 1); + mask = _mm_subs_epu16(_mm_adds_epu16(abs_p0q0, abs_p1q1), blimit0); + mask = _mm_xor_si128(_mm_cmpeq_epi16(mask, zero), ffff); + // mask |= (abs(*p0 - *q0) * 2 + abs(*p1 - *q1) / 2 > blimit) * -1; + // So taking maximums continues to work: + mask = _mm_and_si128(mask, _mm_adds_epu16(limit0, one)); + + mask = _mm_max_epi16(abs_q2q1, mask); + work = _mm_max_epi16(abs_p1p0, abs_q1q0); + mask = _mm_max_epi16(work, mask); + mask = _mm_max_epi16(mask, abs_p2p1); + mask = _mm_subs_epu16(mask, limit0); + mask = _mm_cmpeq_epi16(mask, zero); + + // lp filter + __m128i ps[2], qs[2], p[2], q[2]; + { + p[0] = *p0; + p[1] = *p1; + q[0] = *q0; + q[1] = *q1; + // filter_mask and hev_mask + highbd_filter4_dual_sse2(p, q, ps, qs, &mask, &thresh0, bd, &t80); + } + + // flat_mask + flat = _mm_max_epi16(abs_diff16(*q2, *q0), abs_diff16(*p2, *p0)); + flat = _mm_max_epi16(flat, work); + + flat = _mm_subs_epu16(flat, _mm_slli_epi16(one, bd - 8)); + + flat = _mm_cmpeq_epi16(flat, zero); + flat = _mm_and_si128(flat, mask); // flat & mask + + // 5 tap filter + // need it only if flat !=0 + if (0xffff != _mm_movemask_epi8(_mm_cmpeq_epi16(flat, zero))) { + __m128i workp_a, workp_b, workp_shft0, workp_shft1; + + // op1 + workp_a = _mm_add_epi16(_mm_add_epi16(*p0, *p0), + _mm_add_epi16(*p1, *p1)); // *p0 *2 + *p1 * 2 + workp_a = _mm_add_epi16(_mm_add_epi16(workp_a, four), + *p2); // *p2 + *p0 * 2 + *p1 * 2 + 4 + + workp_b = _mm_add_epi16(_mm_add_epi16(*p2, *p2), *q0); + workp_shft0 = _mm_add_epi16( + workp_a, workp_b); // *p2 * 3 + *p1 * 2 + *p0 * 2 + *q0 + 4 + op1 = _mm_srli_epi16(workp_shft0, 3); + + // op0 + workp_b = _mm_add_epi16(_mm_add_epi16(*q0, *q0), *q1); // *q0 * 2 + *q1 + workp_a = + _mm_add_epi16(workp_a, + workp_b); // *p2 + *p0 * 2 + *p1 * 2 + *q0 * 2 + *q1 + 4 + op0 = _mm_srli_epi16(workp_a, 3); + + // oq0 + workp_a = _mm_sub_epi16(_mm_sub_epi16(workp_a, *p2), + *p1); // *p0 * 2 + *p1 + *q0 * 2 + *q1 + 4 + workp_b = _mm_add_epi16(*q1, *q2); + workp_shft0 = _mm_add_epi16( + workp_a, workp_b); // *p0 * 2 + *p1 + *q0 * 2 + *q1 * 2 + *q2 + 4 + oq0 = _mm_srli_epi16(workp_shft0, 3); + + // oq1 + workp_a = _mm_sub_epi16(_mm_sub_epi16(workp_shft0, *p1), + *p0); // *p0 + *q0 * 2 + *q1 * 2 + *q2 + 4 + workp_b = _mm_add_epi16(*q2, *q2); + workp_shft1 = _mm_add_epi16( + workp_a, workp_b); // *p0 + *q0 * 2 + *q1 * 2 + *q2 * 3 + 4 + oq1 = _mm_srli_epi16(workp_shft1, 3); + + qs[0] = _mm_andnot_si128(flat, qs[0]); + oq0 = _mm_and_si128(flat, oq0); + *q0 = _mm_or_si128(qs[0], oq0); + + qs[1] = _mm_andnot_si128(flat, qs[1]); + oq1 = _mm_and_si128(flat, oq1); + *q1 = _mm_or_si128(qs[1], oq1); + + ps[0] = _mm_andnot_si128(flat, ps[0]); + op0 = _mm_and_si128(flat, op0); + *p0 = _mm_or_si128(ps[0], op0); + + ps[1] = _mm_andnot_si128(flat, ps[1]); + op1 = _mm_and_si128(flat, op1); + *p1 = _mm_or_si128(ps[1], op1); + } else { + *q0 = qs[0]; + *q1 = qs[1]; + *p0 = ps[0]; + *p1 = ps[1]; + } +} + +void aom_highbd_lpf_horizontal_6_sse2(uint16_t *s, int p, + const uint8_t *_blimit, + const uint8_t *_limit, + const uint8_t *_thresh, int bd) { + __m128i p2, p1, p0, q0, q1, q2, p1p0_out, q1q0_out; + + p2 = _mm_loadl_epi64((__m128i *)(s - 3 * p)); + p1 = _mm_loadl_epi64((__m128i *)(s - 2 * p)); + p0 = _mm_loadl_epi64((__m128i *)(s - 1 * p)); + q0 = _mm_loadl_epi64((__m128i *)(s + 0 * p)); + q1 = _mm_loadl_epi64((__m128i *)(s + 1 * p)); + q2 = _mm_loadl_epi64((__m128i *)(s + 2 * p)); + + highbd_lpf_internal_6_sse2(&p2, &p1, &p0, &q0, &q1, &q2, &p1p0_out, &q1q0_out, + _blimit, _limit, _thresh, bd); + + _mm_storel_epi64((__m128i *)(s - 2 * p), _mm_srli_si128(p1p0_out, 8)); + _mm_storel_epi64((__m128i *)(s - 1 * p), p1p0_out); + _mm_storel_epi64((__m128i *)(s + 0 * p), q1q0_out); + _mm_storel_epi64((__m128i *)(s + 1 * p), _mm_srli_si128(q1q0_out, 8)); +} + +void aom_highbd_lpf_horizontal_6_dual_sse2( + uint16_t *s, int p, const uint8_t *_blimit0, const uint8_t *_limit0, + const uint8_t *_thresh0, const uint8_t *_blimit1, const uint8_t *_limit1, + const uint8_t *_thresh1, int bd) { + __m128i p2, p1, p0, q0, q1, q2; + + p2 = _mm_loadu_si128((__m128i *)(s - 3 * p)); + p1 = _mm_loadu_si128((__m128i *)(s - 2 * p)); + p0 = _mm_loadu_si128((__m128i *)(s - 1 * p)); + q0 = _mm_loadu_si128((__m128i *)(s + 0 * p)); + q1 = _mm_loadu_si128((__m128i *)(s + 1 * p)); + q2 = _mm_loadu_si128((__m128i *)(s + 2 * p)); + + highbd_lpf_internal_6_dual_sse2(&p2, &p1, &p0, &q0, &q1, &q2, _blimit0, + _limit0, _thresh0, _blimit1, _limit1, + _thresh1, bd); + + _mm_storeu_si128((__m128i *)(s - 2 * p), p1); + _mm_storeu_si128((__m128i *)(s - 1 * p), p0); + _mm_storeu_si128((__m128i *)(s + 0 * p), q0); + _mm_storeu_si128((__m128i *)(s + 1 * p), q1); +} + +static AOM_FORCE_INLINE void highbd_lpf_internal_8_sse2( + __m128i *p3, __m128i *q3, __m128i *p2, __m128i *q2, __m128i *p1, + __m128i *q1, __m128i *p0, __m128i *q0, __m128i *q1q0_out, __m128i *p1p0_out, + const unsigned char *_blimit, const unsigned char *_limit, + const unsigned char *_thresh, int bd) { + const __m128i zero = _mm_setzero_si128(); + __m128i blimit, limit, thresh; + __m128i mask, hev, flat; + __m128i pq[4]; + __m128i p1p0, q1q0, ps1ps0, qs1qs0; + __m128i work_a, opq2, flat_p1p0, flat_q0q1; + + pq[0] = _mm_unpacklo_epi64(*p0, *q0); + pq[1] = _mm_unpacklo_epi64(*p1, *q1); + pq[2] = _mm_unpacklo_epi64(*p2, *q2); + pq[3] = _mm_unpacklo_epi64(*p3, *q3); + + __m128i abs_p1p0; + + const __m128i four = _mm_set1_epi16(4); + __m128i t80; + const __m128i one = _mm_set1_epi16(0x1); + + get_limit(_blimit, _limit, _thresh, bd, &blimit, &limit, &thresh, &t80); + + highbd_hev_filter_mask_x_sse2(pq, 4, &p1p0, &q1q0, &abs_p1p0, &limit, &blimit, + &thresh, &hev, &mask); + + // lp filter + highbd_filter4_sse2(&p1p0, &q1q0, &hev, &mask, q1q0_out, p1p0_out, &t80, bd); + + // flat_mask4 + flat = _mm_max_epi16(abs_diff16(pq[2], pq[0]), abs_diff16(pq[3], pq[0])); + flat = _mm_max_epi16(abs_p1p0, flat); + flat = _mm_max_epi16(flat, _mm_srli_si128(flat, 8)); + + flat = _mm_subs_epu16(flat, _mm_slli_epi16(one, bd - 8)); + + flat = _mm_cmpeq_epi16(flat, zero); + flat = _mm_and_si128(flat, mask); + // replicate for the further "merged variables" usage + flat = _mm_unpacklo_epi64(flat, flat); + + if (0xffff != _mm_movemask_epi8(_mm_cmpeq_epi16(flat, zero))) { + __m128i workp_a, workp_b, workp_c, workp_shft0, workp_shft1; + // Added before shift for rounding part of ROUND_POWER_OF_TWO + + // o*p2 + workp_a = _mm_add_epi16(_mm_add_epi16(*p3, *p3), _mm_add_epi16(*p2, *p1)); + workp_a = _mm_add_epi16(_mm_add_epi16(workp_a, four), *p0); + workp_c = _mm_add_epi16(_mm_add_epi16(*q0, *p2), *p3); + workp_c = _mm_add_epi16(workp_a, workp_c); + + // o*p1 + workp_b = _mm_add_epi16(_mm_add_epi16(*q0, *q1), *p1); + workp_shft0 = _mm_add_epi16(workp_a, workp_b); + + // o*p0 + workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, *p3), *q2); + workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, *p1), *p0); + workp_shft1 = _mm_add_epi16(workp_a, workp_b); + + flat_p1p0 = _mm_srli_epi16(_mm_unpacklo_epi64(workp_shft1, workp_shft0), 3); + + // oq0 + workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, *p3), *q3); + workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, *p0), *q0); + workp_shft0 = _mm_add_epi16(workp_a, workp_b); + + // oq1 + workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, *p2), *q3); + workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, *q0), *q1); + workp_shft1 = _mm_add_epi16(workp_a, workp_b); + + flat_q0q1 = _mm_srli_epi16(_mm_unpacklo_epi64(workp_shft0, workp_shft1), 3); + + // oq2 + workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, *p1), *q3); + workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, *q1), *q2); + workp_a = _mm_add_epi16(workp_a, workp_b); + opq2 = _mm_srli_epi16(_mm_unpacklo_epi64(workp_c, workp_a), 3); + + qs1qs0 = _mm_andnot_si128(flat, *q1q0_out); + q1q0 = _mm_and_si128(flat, flat_q0q1); + *q1q0_out = _mm_or_si128(qs1qs0, q1q0); + + ps1ps0 = _mm_andnot_si128(flat, *p1p0_out); + p1p0 = _mm_and_si128(flat, flat_p1p0); + *p1p0_out = _mm_or_si128(ps1ps0, p1p0); + + work_a = _mm_andnot_si128(flat, pq[2]); + *p2 = _mm_and_si128(flat, opq2); + *p2 = _mm_or_si128(work_a, *p2); + *q2 = _mm_srli_si128(*p2, 8); + } +} + +static AOM_FORCE_INLINE void highbd_lpf_internal_8_dual_sse2( + __m128i *p3, __m128i *q3, __m128i *p2, __m128i *q2, __m128i *p1, + __m128i *q1, __m128i *p0, __m128i *q0, const unsigned char *_blimit0, + const unsigned char *_limit0, const unsigned char *_thresh0, + const unsigned char *_blimit1, const unsigned char *_limit1, + const unsigned char *_thresh1, int bd) { + __m128i blimit0, limit0, thresh0; + __m128i t80; + __m128i mask, flat; + __m128i work_a, op2, oq2, op1, op0, oq0, oq1; + __m128i abs_p1q1, abs_p0q0, work0, work1, work2; + + const __m128i zero = _mm_setzero_si128(); + const __m128i four = _mm_set1_epi16(4); + const __m128i one = _mm_set1_epi16(0x1); + const __m128i ffff = _mm_cmpeq_epi16(one, one); + + get_limit_dual(_blimit0, _limit0, _thresh0, _blimit1, _limit1, _thresh1, bd, + &blimit0, &limit0, &thresh0, &t80); + + abs_p0q0 = abs_diff16(*p0, *q0); + abs_p1q1 = abs_diff16(*p1, *q1); + + abs_p0q0 = _mm_adds_epu16(abs_p0q0, abs_p0q0); + abs_p1q1 = _mm_srli_epi16(abs_p1q1, 1); + mask = _mm_subs_epu16(_mm_adds_epu16(abs_p0q0, abs_p1q1), blimit0); + mask = _mm_xor_si128(_mm_cmpeq_epi16(mask, zero), ffff); + // mask |= (abs(*p0 - q0) * 2 + abs(*p1 - q1) / 2 > blimit) * -1; + + // So taking maximums continues to work: + mask = _mm_and_si128(mask, _mm_adds_epu16(limit0, one)); + + work0 = _mm_max_epi16(abs_diff16(*p3, *p2), abs_diff16(*p2, *p1)); + work1 = + _mm_max_epi16(abs_diff16(*p1, *p0), abs_diff16(*q1, *q0)); // tbu 4 flat + work0 = _mm_max_epi16(work0, work1); + work2 = _mm_max_epi16(abs_diff16(*q2, *q1), abs_diff16(*q2, *q3)); + work2 = _mm_max_epi16(work2, work0); + mask = _mm_max_epi16(work2, mask); + + mask = _mm_subs_epu16(mask, limit0); + mask = _mm_cmpeq_epi16(mask, zero); + + // lp filter + __m128i ps[2], qs[2], p[2], q[2]; + { + p[0] = *p0; + p[1] = *p1; + q[0] = *q0; + q[1] = *q1; + // filter_mask and hev_mask + highbd_filter4_dual_sse2(p, q, ps, qs, &mask, &thresh0, bd, &t80); + } + + flat = _mm_max_epi16(abs_diff16(*p2, *p0), abs_diff16(*q2, *q0)); + flat = _mm_max_epi16(work1, flat); + work0 = _mm_max_epi16(abs_diff16(*p3, *p0), abs_diff16(*q3, *q0)); + flat = _mm_max_epi16(work0, flat); + + flat = _mm_subs_epu16(flat, _mm_slli_epi16(one, bd - 8)); + flat = _mm_cmpeq_epi16(flat, zero); + flat = _mm_and_si128(flat, mask); // flat & mask + + // filter8 need it only if flat !=0 + if (0xffff != _mm_movemask_epi8(_mm_cmpeq_epi16(flat, zero))) { + __m128i workp_a, workp_b; + // Added before shift for rounding part of ROUND_POWER_OF_TWO + + // o*p2 + workp_a = _mm_add_epi16(_mm_add_epi16(*p3, *p3), _mm_add_epi16(*p2, *p1)); + workp_a = _mm_add_epi16(_mm_add_epi16(workp_a, four), *p0); + workp_b = _mm_add_epi16(_mm_add_epi16(*q0, *p2), *p3); + op2 = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3); + + // o*p1 + workp_b = _mm_add_epi16(_mm_add_epi16(*q0, *q1), *p1); + op1 = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3); + + // o*p0 + workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, *p3), *q2); + workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, *p1), *p0); + op0 = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3); + + // oq0 + workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, *p3), *q3); + workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, *p0), *q0); + oq0 = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3); + + // oq1 + workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, *p2), *q3); + workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, *q0), *q1); + oq1 = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3); + + // oq2 + workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, *p1), *q3); + workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, *q1), *q2); + oq2 = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3); + + qs[0] = _mm_andnot_si128(flat, qs[0]); + oq0 = _mm_and_si128(flat, oq0); + *q0 = _mm_or_si128(qs[0], oq0); + + qs[1] = _mm_andnot_si128(flat, qs[1]); + oq1 = _mm_and_si128(flat, oq1); + *q1 = _mm_or_si128(qs[1], oq1); + + ps[0] = _mm_andnot_si128(flat, ps[0]); + op0 = _mm_and_si128(flat, op0); + *p0 = _mm_or_si128(ps[0], op0); + + ps[1] = _mm_andnot_si128(flat, ps[1]); + op1 = _mm_and_si128(flat, op1); + *p1 = _mm_or_si128(ps[1], op1); + + work_a = _mm_andnot_si128(flat, *q2); + *q2 = _mm_and_si128(flat, oq2); + *q2 = _mm_or_si128(work_a, *q2); + + work_a = _mm_andnot_si128(flat, *p2); + *p2 = _mm_and_si128(flat, op2); + *p2 = _mm_or_si128(work_a, *p2); + } else { + *q0 = qs[0]; + *q1 = qs[1]; + *p0 = ps[0]; + *p1 = ps[1]; + } +} + +void aom_highbd_lpf_horizontal_8_sse2(uint16_t *s, int p, + const uint8_t *_blimit, + const uint8_t *_limit, + const uint8_t *_thresh, int bd) { + __m128i p2, p1, p0, q0, q1, q2, p3, q3; + __m128i q1q0, p1p0; + + p3 = _mm_loadl_epi64((__m128i *)(s - 4 * p)); + q3 = _mm_loadl_epi64((__m128i *)(s + 3 * p)); + p2 = _mm_loadl_epi64((__m128i *)(s - 3 * p)); + q2 = _mm_loadl_epi64((__m128i *)(s + 2 * p)); + p1 = _mm_loadl_epi64((__m128i *)(s - 2 * p)); + q1 = _mm_loadl_epi64((__m128i *)(s + 1 * p)); + p0 = _mm_loadl_epi64((__m128i *)(s - 1 * p)); + q0 = _mm_loadl_epi64((__m128i *)(s + 0 * p)); + + highbd_lpf_internal_8_sse2(&p3, &q3, &p2, &q2, &p1, &q1, &p0, &q0, &q1q0, + &p1p0, _blimit, _limit, _thresh, bd); + + _mm_storel_epi64((__m128i *)(s - 3 * p), p2); + _mm_storel_epi64((__m128i *)(s - 2 * p), _mm_srli_si128(p1p0, 8)); + _mm_storel_epi64((__m128i *)(s - 1 * p), p1p0); + _mm_storel_epi64((__m128i *)(s + 0 * p), q1q0); + _mm_storel_epi64((__m128i *)(s + 1 * p), _mm_srli_si128(q1q0, 8)); + _mm_storel_epi64((__m128i *)(s + 2 * p), q2); +} + +void aom_highbd_lpf_horizontal_8_dual_sse2( + uint16_t *s, int p, const uint8_t *_blimit0, const uint8_t *_limit0, + const uint8_t *_thresh0, const uint8_t *_blimit1, const uint8_t *_limit1, + const uint8_t *_thresh1, int bd) { + __m128i p2, p1, p0, q0, q1, q2, p3, q3; + + p3 = _mm_loadu_si128((__m128i *)(s - 4 * p)); + q3 = _mm_loadu_si128((__m128i *)(s + 3 * p)); + p2 = _mm_loadu_si128((__m128i *)(s - 3 * p)); + q2 = _mm_loadu_si128((__m128i *)(s + 2 * p)); + p1 = _mm_loadu_si128((__m128i *)(s - 2 * p)); + q1 = _mm_loadu_si128((__m128i *)(s + 1 * p)); + p0 = _mm_loadu_si128((__m128i *)(s - 1 * p)); + q0 = _mm_loadu_si128((__m128i *)(s + 0 * p)); + + highbd_lpf_internal_8_dual_sse2(&p3, &q3, &p2, &q2, &p1, &q1, &p0, &q0, + _blimit0, _limit0, _thresh0, _blimit1, + _limit1, _thresh1, bd); + + _mm_storeu_si128((__m128i *)(s - 3 * p), p2); + _mm_storeu_si128((__m128i *)(s - 2 * p), p1); + _mm_storeu_si128((__m128i *)(s - 1 * p), p0); + _mm_storeu_si128((__m128i *)(s + 0 * p), q0); + _mm_storeu_si128((__m128i *)(s + 1 * p), q1); + _mm_storeu_si128((__m128i *)(s + 2 * p), q2); +} + +static AOM_FORCE_INLINE void highbd_lpf_internal_4_sse2( + __m128i *p1, __m128i *p0, __m128i *q0, __m128i *q1, __m128i *q1q0_out, + __m128i *p1p0_out, const uint8_t *_blimit, const uint8_t *_limit, + const uint8_t *_thresh, int bd) { + __m128i blimit, limit, thresh; + __m128i mask, hev; + __m128i p1p0, q1q0; + __m128i pq[2]; + + __m128i abs_p1p0; + + __m128i t80; + get_limit(_blimit, _limit, _thresh, bd, &blimit, &limit, &thresh, &t80); + + pq[0] = _mm_unpacklo_epi64(*p0, *q0); + pq[1] = _mm_unpacklo_epi64(*p1, *q1); + + highbd_hev_filter_mask_x_sse2(pq, 2, &p1p0, &q1q0, &abs_p1p0, &limit, &blimit, + &thresh, &hev, &mask); + + highbd_filter4_sse2(&p1p0, &q1q0, &hev, &mask, q1q0_out, p1p0_out, &t80, bd); +} + +static AOM_FORCE_INLINE void highbd_lpf_internal_4_dual_sse2( + __m128i *p1, __m128i *p0, __m128i *q0, __m128i *q1, __m128i *ps, + __m128i *qs, const uint8_t *_blimit0, const uint8_t *_limit0, + const uint8_t *_thresh0, const uint8_t *_blimit1, const uint8_t *_limit1, + const uint8_t *_thresh1, int bd) { + __m128i blimit0, limit0, thresh0; + __m128i mask, flat; + __m128i p[2], q[2]; + + const __m128i zero = _mm_setzero_si128(); + __m128i abs_p0q0 = abs_diff16(*q0, *p0); + __m128i abs_p1q1 = abs_diff16(*q1, *p1); + + __m128i abs_p1p0 = abs_diff16(*p1, *p0); + __m128i abs_q1q0 = abs_diff16(*q1, *q0); + + const __m128i ffff = _mm_cmpeq_epi16(abs_p1p0, abs_p1p0); + const __m128i one = _mm_set1_epi16(1); + + __m128i t80; + + get_limit_dual(_blimit0, _limit0, _thresh0, _blimit1, _limit1, _thresh1, bd, + &blimit0, &limit0, &thresh0, &t80); + + // filter_mask and hev_mask + flat = _mm_max_epi16(abs_p1p0, abs_q1q0); + + abs_p0q0 = _mm_adds_epu16(abs_p0q0, abs_p0q0); + abs_p1q1 = _mm_srli_epi16(abs_p1q1, 1); + + mask = _mm_subs_epu16(_mm_adds_epu16(abs_p0q0, abs_p1q1), blimit0); + mask = _mm_xor_si128(_mm_cmpeq_epi16(mask, zero), ffff); + // mask |= (abs(*p0 - *q0) * 2 + abs(*p1 - *q1) / 2 > blimit) * -1; + // So taking maximums continues to work: + mask = _mm_and_si128(mask, _mm_adds_epu16(limit0, one)); + mask = _mm_max_epi16(flat, mask); + + mask = _mm_subs_epu16(mask, limit0); + mask = _mm_cmpeq_epi16(mask, zero); + + p[0] = *p0; + p[1] = *p1; + q[0] = *q0; + q[1] = *q1; + + highbd_filter4_dual_sse2(p, q, ps, qs, &mask, &thresh0, bd, &t80); +} + +void aom_highbd_lpf_horizontal_4_sse2(uint16_t *s, int p, + const uint8_t *_blimit, + const uint8_t *_limit, + const uint8_t *_thresh, int bd) { + __m128i p1p0, q1q0; + __m128i p1 = _mm_loadl_epi64((__m128i *)(s - 2 * p)); + __m128i p0 = _mm_loadl_epi64((__m128i *)(s - 1 * p)); + __m128i q0 = _mm_loadl_epi64((__m128i *)(s - 0 * p)); + __m128i q1 = _mm_loadl_epi64((__m128i *)(s + 1 * p)); + + highbd_lpf_internal_4_sse2(&p1, &p0, &q0, &q1, &q1q0, &p1p0, _blimit, _limit, + _thresh, bd); + + _mm_storel_epi64((__m128i *)(s - 2 * p), _mm_srli_si128(p1p0, 8)); + _mm_storel_epi64((__m128i *)(s - 1 * p), p1p0); + _mm_storel_epi64((__m128i *)(s + 0 * p), q1q0); + _mm_storel_epi64((__m128i *)(s + 1 * p), _mm_srli_si128(q1q0, 8)); +} + +void aom_highbd_lpf_horizontal_4_dual_sse2( + uint16_t *s, int p, const uint8_t *_blimit0, const uint8_t *_limit0, + const uint8_t *_thresh0, const uint8_t *_blimit1, const uint8_t *_limit1, + const uint8_t *_thresh1, int bd) { + __m128i p1 = _mm_loadu_si128((__m128i *)(s - 2 * p)); + __m128i p0 = _mm_loadu_si128((__m128i *)(s - 1 * p)); + __m128i q0 = _mm_loadu_si128((__m128i *)(s - 0 * p)); + __m128i q1 = _mm_loadu_si128((__m128i *)(s + 1 * p)); + __m128i ps[2], qs[2]; + + highbd_lpf_internal_4_dual_sse2(&p1, &p0, &q0, &q1, ps, qs, _blimit0, _limit0, + _thresh0, _blimit1, _limit1, _thresh1, bd); + + _mm_storeu_si128((__m128i *)(s - 2 * p), ps[1]); + _mm_storeu_si128((__m128i *)(s - 1 * p), ps[0]); + _mm_storeu_si128((__m128i *)(s + 0 * p), qs[0]); + _mm_storeu_si128((__m128i *)(s + 1 * p), qs[1]); +} + +void aom_highbd_lpf_vertical_4_sse2(uint16_t *s, int p, const uint8_t *blimit, + const uint8_t *limit, const uint8_t *thresh, + int bd) { + __m128i x0, x1, x2, x3, d0, d1, d2, d3; + __m128i p1p0, q1q0; + __m128i p1, q1; + + x0 = _mm_loadl_epi64((__m128i *)(s - 2 + 0 * p)); + x1 = _mm_loadl_epi64((__m128i *)(s - 2 + 1 * p)); + x2 = _mm_loadl_epi64((__m128i *)(s - 2 + 2 * p)); + x3 = _mm_loadl_epi64((__m128i *)(s - 2 + 3 * p)); + + highbd_transpose4x8_8x4_low_sse2(&x0, &x1, &x2, &x3, &d0, &d1, &d2, &d3); + + highbd_lpf_internal_4_sse2(&d0, &d1, &d2, &d3, &q1q0, &p1p0, blimit, limit, + thresh, bd); + + p1 = _mm_srli_si128(p1p0, 8); + q1 = _mm_srli_si128(q1q0, 8); + + // transpose from 8x4 to 4x8 + highbd_transpose4x8_8x4_low_sse2(&p1, &p1p0, &q1q0, &q1, &d0, &d1, &d2, &d3); + + _mm_storel_epi64((__m128i *)(s - 2 + 0 * p), d0); + _mm_storel_epi64((__m128i *)(s - 2 + 1 * p), d1); + _mm_storel_epi64((__m128i *)(s - 2 + 2 * p), d2); + _mm_storel_epi64((__m128i *)(s - 2 + 3 * p), d3); +} + +void aom_highbd_lpf_vertical_4_dual_sse2( + uint16_t *s, int p, const uint8_t *blimit0, const uint8_t *limit0, + const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, + const uint8_t *thresh1, int bd) { + __m128i x0, x1, x2, x3, x4, x5, x6, x7; + __m128i d0, d1, d2, d3, d4, d5, d6, d7; + __m128i ps[2], qs[2]; + + x0 = _mm_loadl_epi64((__m128i *)(s - 2 + 0 * p)); + x1 = _mm_loadl_epi64((__m128i *)(s - 2 + 1 * p)); + x2 = _mm_loadl_epi64((__m128i *)(s - 2 + 2 * p)); + x3 = _mm_loadl_epi64((__m128i *)(s - 2 + 3 * p)); + x4 = _mm_loadl_epi64((__m128i *)(s - 2 + 4 * p)); + x5 = _mm_loadl_epi64((__m128i *)(s - 2 + 5 * p)); + x6 = _mm_loadl_epi64((__m128i *)(s - 2 + 6 * p)); + x7 = _mm_loadl_epi64((__m128i *)(s - 2 + 7 * p)); + + highbd_transpose8x8_low_sse2(&x0, &x1, &x2, &x3, &x4, &x5, &x6, &x7, &d0, &d1, + &d2, &d3); + + highbd_lpf_internal_4_dual_sse2(&d0, &d1, &d2, &d3, ps, qs, blimit0, limit0, + thresh0, blimit1, limit1, thresh1, bd); + + highbd_transpose4x8_8x4_sse2(&ps[1], &ps[0], &qs[0], &qs[1], &d0, &d1, &d2, + &d3, &d4, &d5, &d6, &d7); + + _mm_storel_epi64((__m128i *)(s - 2 + 0 * p), d0); + _mm_storel_epi64((__m128i *)(s - 2 + 1 * p), d1); + _mm_storel_epi64((__m128i *)(s - 2 + 2 * p), d2); + _mm_storel_epi64((__m128i *)(s - 2 + 3 * p), d3); + _mm_storel_epi64((__m128i *)(s - 2 + 4 * p), d4); + _mm_storel_epi64((__m128i *)(s - 2 + 5 * p), d5); + _mm_storel_epi64((__m128i *)(s - 2 + 6 * p), d6); + _mm_storel_epi64((__m128i *)(s - 2 + 7 * p), d7); +} + +void aom_highbd_lpf_vertical_6_sse2(uint16_t *s, int p, const uint8_t *blimit, + const uint8_t *limit, const uint8_t *thresh, + int bd) { + __m128i d0, d1, d2, d3, d4, d5, d6, d7; + __m128i x3, x2, x1, x0, p0, q0; + __m128i p1p0, q1q0; + + x3 = _mm_loadu_si128((__m128i *)((s - 3) + 0 * p)); + x2 = _mm_loadu_si128((__m128i *)((s - 3) + 1 * p)); + x1 = _mm_loadu_si128((__m128i *)((s - 3) + 2 * p)); + x0 = _mm_loadu_si128((__m128i *)((s - 3) + 3 * p)); + + highbd_transpose4x8_8x4_sse2(&x3, &x2, &x1, &x0, &d0, &d1, &d2, &d3, &d4, &d5, + &d6, &d7); + + highbd_lpf_internal_6_sse2(&d0, &d1, &d2, &d3, &d4, &d5, &p1p0, &q1q0, blimit, + limit, thresh, bd); + + p0 = _mm_srli_si128(p1p0, 8); + q0 = _mm_srli_si128(q1q0, 8); + + highbd_transpose4x8_8x4_low_sse2(&p0, &p1p0, &q1q0, &q0, &d0, &d1, &d2, &d3); + + _mm_storel_epi64((__m128i *)(s - 2 + 0 * p), d0); + _mm_storel_epi64((__m128i *)(s - 2 + 1 * p), d1); + _mm_storel_epi64((__m128i *)(s - 2 + 2 * p), d2); + _mm_storel_epi64((__m128i *)(s - 2 + 3 * p), d3); +} + +void aom_highbd_lpf_vertical_6_dual_sse2( + uint16_t *s, int p, const uint8_t *_blimit0, const uint8_t *_limit0, + const uint8_t *_thresh0, const uint8_t *_blimit1, const uint8_t *_limit1, + const uint8_t *_thresh1, int bd) { + __m128i d0, d1, d2, d3, d4, d5, d6, d7; + __m128i x0, x1, x2, x3, x4, x5, x6, x7; + __m128i p0, q0, p1, q1, p2, q2; + + x0 = _mm_loadu_si128((__m128i *)((s - 3) + 0 * p)); + x1 = _mm_loadu_si128((__m128i *)((s - 3) + 1 * p)); + x2 = _mm_loadu_si128((__m128i *)((s - 3) + 2 * p)); + x3 = _mm_loadu_si128((__m128i *)((s - 3) + 3 * p)); + x4 = _mm_loadu_si128((__m128i *)((s - 3) + 4 * p)); + x5 = _mm_loadu_si128((__m128i *)((s - 3) + 5 * p)); + x6 = _mm_loadu_si128((__m128i *)((s - 3) + 6 * p)); + x7 = _mm_loadu_si128((__m128i *)((s - 3) + 7 * p)); + + highbd_transpose8x8_sse2(&x0, &x1, &x2, &x3, &x4, &x5, &x6, &x7, &p2, &p1, + &p0, &q0, &q1, &q2, &d6, &d7); + + highbd_lpf_internal_6_dual_sse2(&p2, &p1, &p0, &q0, &q1, &q2, _blimit0, + _limit0, _thresh0, _blimit1, _limit1, + _thresh1, bd); + + highbd_transpose4x8_8x4_sse2(&p1, &p0, &q0, &q1, &d0, &d1, &d2, &d3, &d4, &d5, + &d6, &d7); + + _mm_storel_epi64((__m128i *)(s - 2 + 0 * p), d0); + _mm_storel_epi64((__m128i *)(s - 2 + 1 * p), d1); + _mm_storel_epi64((__m128i *)(s - 2 + 2 * p), d2); + _mm_storel_epi64((__m128i *)(s - 2 + 3 * p), d3); + _mm_storel_epi64((__m128i *)(s - 2 + 4 * p), d4); + _mm_storel_epi64((__m128i *)(s - 2 + 5 * p), d5); + _mm_storel_epi64((__m128i *)(s - 2 + 6 * p), d6); + _mm_storel_epi64((__m128i *)(s - 2 + 7 * p), d7); +} + +void aom_highbd_lpf_vertical_8_sse2(uint16_t *s, int p, const uint8_t *blimit, + const uint8_t *limit, const uint8_t *thresh, + int bd) { + __m128i d0, d1, d2, d3, d4, d5, d6, d7; + __m128i p2, p1, p0, p3, q0; + __m128i q1q0, p1p0; + + p3 = _mm_loadu_si128((__m128i *)((s - 4) + 0 * p)); + p2 = _mm_loadu_si128((__m128i *)((s - 4) + 1 * p)); + p1 = _mm_loadu_si128((__m128i *)((s - 4) + 2 * p)); + p0 = _mm_loadu_si128((__m128i *)((s - 4) + 3 * p)); + + highbd_transpose4x8_8x4_sse2(&p3, &p2, &p1, &p0, &d0, &d1, &d2, &d3, &d4, &d5, + &d6, &d7); + + // Loop filtering + highbd_lpf_internal_8_sse2(&d0, &d7, &d1, &d6, &d2, &d5, &d3, &d4, &q1q0, + &p1p0, blimit, limit, thresh, bd); + + p0 = _mm_srli_si128(p1p0, 8); + q0 = _mm_srli_si128(q1q0, 8); + + highbd_transpose8x8_low_sse2(&d0, &d1, &p0, &p1p0, &q1q0, &q0, &d6, &d7, &d0, + &d1, &d2, &d3); + + _mm_storeu_si128((__m128i *)(s - 4 + 0 * p), d0); + _mm_storeu_si128((__m128i *)(s - 4 + 1 * p), d1); + _mm_storeu_si128((__m128i *)(s - 4 + 2 * p), d2); + _mm_storeu_si128((__m128i *)(s - 4 + 3 * p), d3); +} + +void aom_highbd_lpf_vertical_8_dual_sse2( + uint16_t *s, int p, const uint8_t *blimit0, const uint8_t *limit0, + const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, + const uint8_t *thresh1, int bd) { + __m128i x0, x1, x2, x3, x4, x5, x6, x7; + __m128i d0, d1, d2, d3, d4, d5, d6, d7; + + x0 = _mm_loadu_si128((__m128i *)(s - 4 + 0 * p)); + x1 = _mm_loadu_si128((__m128i *)(s - 4 + 1 * p)); + x2 = _mm_loadu_si128((__m128i *)(s - 4 + 2 * p)); + x3 = _mm_loadu_si128((__m128i *)(s - 4 + 3 * p)); + x4 = _mm_loadu_si128((__m128i *)(s - 4 + 4 * p)); + x5 = _mm_loadu_si128((__m128i *)(s - 4 + 5 * p)); + x6 = _mm_loadu_si128((__m128i *)(s - 4 + 6 * p)); + x7 = _mm_loadu_si128((__m128i *)(s - 4 + 7 * p)); + + highbd_transpose8x8_sse2(&x0, &x1, &x2, &x3, &x4, &x5, &x6, &x7, &d0, &d1, + &d2, &d3, &d4, &d5, &d6, &d7); + + highbd_lpf_internal_8_dual_sse2(&d0, &d7, &d1, &d6, &d2, &d5, &d3, &d4, + blimit0, limit0, thresh0, blimit1, limit1, + thresh1, bd); + + highbd_transpose8x8_sse2(&d0, &d1, &d2, &d3, &d4, &d5, &d6, &d7, &x0, &x1, + &x2, &x3, &x4, &x5, &x6, &x7); + + _mm_storeu_si128((__m128i *)(s - 4 + 0 * p), x0); + _mm_storeu_si128((__m128i *)(s - 4 + 1 * p), x1); + _mm_storeu_si128((__m128i *)(s - 4 + 2 * p), x2); + _mm_storeu_si128((__m128i *)(s - 4 + 3 * p), x3); + _mm_storeu_si128((__m128i *)(s - 4 + 4 * p), x4); + _mm_storeu_si128((__m128i *)(s - 4 + 5 * p), x5); + _mm_storeu_si128((__m128i *)(s - 4 + 6 * p), x6); + _mm_storeu_si128((__m128i *)(s - 4 + 7 * p), x7); +} + +void aom_highbd_lpf_vertical_14_sse2(uint16_t *s, int pitch, + const uint8_t *blimit, + const uint8_t *limit, + const uint8_t *thresh, int bd) { + __m128i q[7], p[7], pq[7]; + __m128i p6, p5, p4, p3; + __m128i p6_2, p5_2, p4_2, p3_2; + __m128i d0, d1, d2, d3; + __m128i d0_2, d1_2, d2_2, d3_2, d7_2; + + p6 = _mm_loadu_si128((__m128i *)((s - 8) + 0 * pitch)); + p5 = _mm_loadu_si128((__m128i *)((s - 8) + 1 * pitch)); + p4 = _mm_loadu_si128((__m128i *)((s - 8) + 2 * pitch)); + p3 = _mm_loadu_si128((__m128i *)((s - 8) + 3 * pitch)); + + highbd_transpose4x8_8x4_sse2(&p6, &p5, &p4, &p3, &d0, &p[6], &p[5], &p[4], + &p[3], &p[2], &p[1], &p[0]); + + p6_2 = _mm_loadu_si128((__m128i *)(s + 0 * pitch)); + p5_2 = _mm_loadu_si128((__m128i *)(s + 1 * pitch)); + p4_2 = _mm_loadu_si128((__m128i *)(s + 2 * pitch)); + p3_2 = _mm_loadu_si128((__m128i *)(s + 3 * pitch)); + + highbd_transpose4x8_8x4_sse2(&p6_2, &p5_2, &p4_2, &p3_2, &q[0], &q[1], &q[2], + &q[3], &q[4], &q[5], &q[6], &d7_2); + + highbd_lpf_internal_14_sse2(p, q, pq, blimit, limit, thresh, bd); + + highbd_transpose8x8_low_sse2(&d0, &p[6], &pq[5], &pq[4], &pq[3], &pq[2], + &pq[1], &pq[0], &d0, &d1, &d2, &d3); + + q[0] = _mm_srli_si128(pq[0], 8); + q[1] = _mm_srli_si128(pq[1], 8); + q[2] = _mm_srli_si128(pq[2], 8); + q[3] = _mm_srli_si128(pq[3], 8); + q[4] = _mm_srli_si128(pq[4], 8); + q[5] = _mm_srli_si128(pq[5], 8); + + highbd_transpose8x8_low_sse2(&q[0], &q[1], &q[2], &q[3], &q[4], &q[5], &q[6], + &d7_2, &d0_2, &d1_2, &d2_2, &d3_2); + + _mm_storeu_si128((__m128i *)(s - 8 + 0 * pitch), d0); + _mm_storeu_si128((__m128i *)(s + 0 * pitch), d0_2); + + _mm_storeu_si128((__m128i *)(s - 8 + 1 * pitch), d1); + _mm_storeu_si128((__m128i *)(s + 1 * pitch), d1_2); + + _mm_storeu_si128((__m128i *)(s - 8 + 2 * pitch), d2); + _mm_storeu_si128((__m128i *)(s + 2 * pitch), d2_2); + + _mm_storeu_si128((__m128i *)(s - 8 + 3 * pitch), d3); + _mm_storeu_si128((__m128i *)(s + 3 * pitch), d3_2); +} + +void aom_highbd_lpf_vertical_14_dual_sse2( + uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, + const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, + const uint8_t *thresh1, int bd) { + __m128i q[7], p[7]; + __m128i p6, p5, p4, p3, p2, p1, p0, q0; + __m128i p6_2, p5_2, p4_2, p3_2, p2_2, p1_2, q0_2, p0_2; + __m128i d0, d7; + __m128i d0_out, d1_out, d2_out, d3_out, d4_out, d5_out, d6_out, d7_out; + + p6 = _mm_loadu_si128((__m128i *)((s - 8) + 0 * pitch)); + p5 = _mm_loadu_si128((__m128i *)((s - 8) + 1 * pitch)); + p4 = _mm_loadu_si128((__m128i *)((s - 8) + 2 * pitch)); + p3 = _mm_loadu_si128((__m128i *)((s - 8) + 3 * pitch)); + p2 = _mm_loadu_si128((__m128i *)((s - 8) + 4 * pitch)); + p1 = _mm_loadu_si128((__m128i *)((s - 8) + 5 * pitch)); + p0 = _mm_loadu_si128((__m128i *)((s - 8) + 6 * pitch)); + q0 = _mm_loadu_si128((__m128i *)((s - 8) + 7 * pitch)); + + highbd_transpose8x8_sse2(&p6, &p5, &p4, &p3, &p2, &p1, &p0, &q0, &d0, &p[6], + &p[5], &p[4], &p[3], &p[2], &p[1], &p[0]); + + p6_2 = _mm_loadu_si128((__m128i *)(s + 0 * pitch)); + p5_2 = _mm_loadu_si128((__m128i *)(s + 1 * pitch)); + p4_2 = _mm_loadu_si128((__m128i *)(s + 2 * pitch)); + p3_2 = _mm_loadu_si128((__m128i *)(s + 3 * pitch)); + p2_2 = _mm_loadu_si128((__m128i *)(s + 4 * pitch)); + p1_2 = _mm_loadu_si128((__m128i *)(s + 5 * pitch)); + p0_2 = _mm_loadu_si128((__m128i *)(s + 6 * pitch)); + q0_2 = _mm_loadu_si128((__m128i *)(s + 7 * pitch)); + + highbd_transpose8x8_sse2(&p6_2, &p5_2, &p4_2, &p3_2, &p2_2, &p1_2, &p0_2, + &q0_2, &q[0], &q[1], &q[2], &q[3], &q[4], &q[5], + &q[6], &d7); + + highbd_lpf_internal_14_dual_sse2(p, q, blimit0, limit0, thresh0, blimit1, + limit1, thresh1, bd); + + highbd_transpose8x8_sse2(&d0, &p[6], &p[5], &p[4], &p[3], &p[2], &p[1], &p[0], + &d0_out, &d1_out, &d2_out, &d3_out, &d4_out, &d5_out, + &d6_out, &d7_out); + + _mm_storeu_si128((__m128i *)(s - 8 + 0 * pitch), d0_out); + _mm_storeu_si128((__m128i *)(s - 8 + 1 * pitch), d1_out); + _mm_storeu_si128((__m128i *)(s - 8 + 2 * pitch), d2_out); + _mm_storeu_si128((__m128i *)(s - 8 + 3 * pitch), d3_out); + _mm_storeu_si128((__m128i *)(s - 8 + 4 * pitch), d4_out); + _mm_storeu_si128((__m128i *)(s - 8 + 5 * pitch), d5_out); + _mm_storeu_si128((__m128i *)(s - 8 + 6 * pitch), d6_out); + _mm_storeu_si128((__m128i *)(s - 8 + 7 * pitch), d7_out); + + highbd_transpose8x8_sse2(&q[0], &q[1], &q[2], &q[3], &q[4], &q[5], &q[6], &d7, + &d0_out, &d1_out, &d2_out, &d3_out, &d4_out, &d5_out, + &d6_out, &d7_out); + + _mm_storeu_si128((__m128i *)(s + 0 * pitch), d0_out); + _mm_storeu_si128((__m128i *)(s + 1 * pitch), d1_out); + _mm_storeu_si128((__m128i *)(s + 2 * pitch), d2_out); + _mm_storeu_si128((__m128i *)(s + 3 * pitch), d3_out); + _mm_storeu_si128((__m128i *)(s + 4 * pitch), d4_out); + _mm_storeu_si128((__m128i *)(s + 5 * pitch), d5_out); + _mm_storeu_si128((__m128i *)(s + 6 * pitch), d6_out); + _mm_storeu_si128((__m128i *)(s + 7 * pitch), d7_out); +} diff --git a/third_party/aom/aom_dsp/x86/highbd_quantize_intrin_avx2.c b/third_party/aom/aom_dsp/x86/highbd_quantize_intrin_avx2.c new file mode 100644 index 0000000000..b9689202a0 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/highbd_quantize_intrin_avx2.c @@ -0,0 +1,160 @@ +/* + * Copyright (c) 2017, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include + +#include "config/aom_dsp_rtcd.h" + +#include "aom/aom_integer.h" + +static INLINE void init_one_qp(const __m128i *p, __m256i *qp) { + const __m128i sign = _mm_srai_epi16(*p, 15); + const __m128i dc = _mm_unpacklo_epi16(*p, sign); + const __m128i ac = _mm_unpackhi_epi16(*p, sign); + *qp = _mm256_insertf128_si256(_mm256_castsi128_si256(dc), ac, 1); +} + +static INLINE void update_qp(__m256i *qp) { + int i; + for (i = 0; i < 5; ++i) { + qp[i] = _mm256_permute2x128_si256(qp[i], qp[i], 0x11); + } +} + +static INLINE void init_qp(const int16_t *zbin_ptr, const int16_t *round_ptr, + const int16_t *quant_ptr, const int16_t *dequant_ptr, + const int16_t *quant_shift_ptr, __m256i *qp) { + const __m128i zbin = _mm_loadu_si128((const __m128i *)zbin_ptr); + const __m128i round = _mm_loadu_si128((const __m128i *)round_ptr); + const __m128i quant = _mm_loadu_si128((const __m128i *)quant_ptr); + const __m128i dequant = _mm_loadu_si128((const __m128i *)dequant_ptr); + const __m128i quant_shift = _mm_loadu_si128((const __m128i *)quant_shift_ptr); + init_one_qp(&zbin, &qp[0]); + init_one_qp(&round, &qp[1]); + init_one_qp(&quant, &qp[2]); + init_one_qp(&dequant, &qp[3]); + init_one_qp(&quant_shift, &qp[4]); +} + +// Note: +// *x is vector multiplied by *y which is 16 int32_t parallel multiplication +// and right shift 16. The output, 16 int32_t is save in *p. +static INLINE void mm256_mul_shift_epi32(const __m256i *x, const __m256i *y, + __m256i *p) { + __m256i prod_lo = _mm256_mul_epi32(*x, *y); + __m256i prod_hi = _mm256_srli_epi64(*x, 32); + const __m256i mult_hi = _mm256_srli_epi64(*y, 32); + prod_hi = _mm256_mul_epi32(prod_hi, mult_hi); + + prod_lo = _mm256_srli_epi64(prod_lo, 16); + const __m256i mask = _mm256_set_epi32(0, -1, 0, -1, 0, -1, 0, -1); + prod_lo = _mm256_and_si256(prod_lo, mask); + prod_hi = _mm256_srli_epi64(prod_hi, 16); + + prod_hi = _mm256_slli_epi64(prod_hi, 32); + *p = _mm256_or_si256(prod_lo, prod_hi); +} + +static INLINE void quantize(const __m256i *qp, __m256i *c, + const int16_t *iscan_ptr, tran_low_t *qcoeff, + tran_low_t *dqcoeff, __m256i *eob) { + const __m256i abs = _mm256_abs_epi32(*c); + const __m256i flag1 = _mm256_cmpgt_epi32(abs, qp[0]); + __m256i flag2 = _mm256_cmpeq_epi32(abs, qp[0]); + flag2 = _mm256_or_si256(flag1, flag2); + const int32_t nzflag = _mm256_movemask_epi8(flag2); + + if (LIKELY(nzflag)) { + __m256i q = _mm256_add_epi32(abs, qp[1]); + __m256i tmp; + mm256_mul_shift_epi32(&q, &qp[2], &tmp); + q = _mm256_add_epi32(tmp, q); + + mm256_mul_shift_epi32(&q, &qp[4], &q); + __m256i dq = _mm256_mullo_epi32(q, qp[3]); + + q = _mm256_sign_epi32(q, *c); + dq = _mm256_sign_epi32(dq, *c); + q = _mm256_and_si256(q, flag2); + dq = _mm256_and_si256(dq, flag2); + + _mm256_storeu_si256((__m256i *)qcoeff, q); + _mm256_storeu_si256((__m256i *)dqcoeff, dq); + + const __m128i isc = _mm_loadu_si128((const __m128i *)iscan_ptr); + const __m128i zr = _mm_setzero_si128(); + const __m128i lo = _mm_unpacklo_epi16(isc, zr); + const __m128i hi = _mm_unpackhi_epi16(isc, zr); + const __m256i iscan = + _mm256_insertf128_si256(_mm256_castsi128_si256(lo), hi, 1); + + const __m256i zero = _mm256_setzero_si256(); + const __m256i zc = _mm256_cmpeq_epi32(dq, zero); + const __m256i nz = _mm256_cmpeq_epi32(zc, zero); + __m256i cur_eob = _mm256_sub_epi32(iscan, nz); + cur_eob = _mm256_and_si256(cur_eob, nz); + *eob = _mm256_max_epi32(cur_eob, *eob); + } else { + const __m256i zero = _mm256_setzero_si256(); + _mm256_storeu_si256((__m256i *)qcoeff, zero); + _mm256_storeu_si256((__m256i *)dqcoeff, zero); + } +} + +void aom_highbd_quantize_b_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, + const int16_t *zbin_ptr, + const int16_t *round_ptr, + const int16_t *quant_ptr, + const int16_t *quant_shift_ptr, + tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, + const int16_t *dequant_ptr, uint16_t *eob_ptr, + const int16_t *scan, const int16_t *iscan) { + (void)scan; + const unsigned int step = 8; + + __m256i qp[5], coeff; + init_qp(zbin_ptr, round_ptr, quant_ptr, dequant_ptr, quant_shift_ptr, qp); + coeff = _mm256_loadu_si256((const __m256i *)coeff_ptr); + + __m256i eob = _mm256_setzero_si256(); + quantize(qp, &coeff, iscan, qcoeff_ptr, dqcoeff_ptr, &eob); + + coeff_ptr += step; + qcoeff_ptr += step; + dqcoeff_ptr += step; + iscan += step; + n_coeffs -= step; + + update_qp(qp); + + while (n_coeffs > 0) { + coeff = _mm256_loadu_si256((const __m256i *)coeff_ptr); + quantize(qp, &coeff, iscan, qcoeff_ptr, dqcoeff_ptr, &eob); + + coeff_ptr += step; + qcoeff_ptr += step; + dqcoeff_ptr += step; + iscan += step; + n_coeffs -= step; + } + { + __m256i eob_s; + eob_s = _mm256_shuffle_epi32(eob, 0xe); + eob = _mm256_max_epi16(eob, eob_s); + eob_s = _mm256_shufflelo_epi16(eob, 0xe); + eob = _mm256_max_epi16(eob, eob_s); + eob_s = _mm256_shufflelo_epi16(eob, 1); + eob = _mm256_max_epi16(eob, eob_s); + const __m128i final_eob = _mm_max_epi16(_mm256_castsi256_si128(eob), + _mm256_extractf128_si256(eob, 1)); + *eob_ptr = _mm_extract_epi16(final_eob, 0); + } +} diff --git a/third_party/aom/aom_dsp/x86/highbd_quantize_intrin_sse2.c b/third_party/aom/aom_dsp/x86/highbd_quantize_intrin_sse2.c new file mode 100644 index 0000000000..58e5f98e58 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/highbd_quantize_intrin_sse2.c @@ -0,0 +1,148 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include + +#include "aom_dsp/aom_dsp_common.h" +#include "aom_mem/aom_mem.h" +#include "aom_ports/mem.h" + +void aom_highbd_quantize_b_sse2(const tran_low_t *coeff_ptr, intptr_t count, + const int16_t *zbin_ptr, + const int16_t *round_ptr, + const int16_t *quant_ptr, + const int16_t *quant_shift_ptr, + tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, + const int16_t *dequant_ptr, uint16_t *eob_ptr, + const int16_t *scan, const int16_t *iscan) { + int i, j, non_zero_regs = (int)count / 4, eob_i = -1; + __m128i zbins[2]; + __m128i nzbins[2]; + + zbins[0] = _mm_set_epi32((int)zbin_ptr[1], (int)zbin_ptr[1], (int)zbin_ptr[1], + (int)zbin_ptr[0]); + zbins[1] = _mm_set1_epi32((int)zbin_ptr[1]); + + nzbins[0] = _mm_setzero_si128(); + nzbins[1] = _mm_setzero_si128(); + nzbins[0] = _mm_sub_epi32(nzbins[0], zbins[0]); + nzbins[1] = _mm_sub_epi32(nzbins[1], zbins[1]); + + (void)scan; + + memset(qcoeff_ptr, 0, count * sizeof(*qcoeff_ptr)); + memset(dqcoeff_ptr, 0, count * sizeof(*dqcoeff_ptr)); + + // Pre-scan pass + for (i = ((int)count / 4) - 1; i >= 0; i--) { + __m128i coeffs, cmp1, cmp2; + int test; + coeffs = _mm_load_si128((const __m128i *)(coeff_ptr + i * 4)); + cmp1 = _mm_cmplt_epi32(coeffs, zbins[i != 0]); + cmp2 = _mm_cmpgt_epi32(coeffs, nzbins[i != 0]); + cmp1 = _mm_and_si128(cmp1, cmp2); + test = _mm_movemask_epi8(cmp1); + if (test == 0xffff) + non_zero_regs--; + else + break; + } + + // Quantization pass: + for (i = 0; i < non_zero_regs; i++) { + __m128i coeffs, coeffs_sign, tmp1, tmp2; + int test; + int abs_coeff[4]; + int coeff_sign[4]; + + coeffs = _mm_load_si128((const __m128i *)(coeff_ptr + i * 4)); + coeffs_sign = _mm_srai_epi32(coeffs, 31); + coeffs = _mm_sub_epi32(_mm_xor_si128(coeffs, coeffs_sign), coeffs_sign); + tmp1 = _mm_cmpgt_epi32(coeffs, zbins[i != 0]); + tmp2 = _mm_cmpeq_epi32(coeffs, zbins[i != 0]); + tmp1 = _mm_or_si128(tmp1, tmp2); + test = _mm_movemask_epi8(tmp1); + _mm_storeu_si128((__m128i *)abs_coeff, coeffs); + _mm_storeu_si128((__m128i *)coeff_sign, coeffs_sign); + + for (j = 0; j < 4; j++) { + if (test & (1 << (4 * j))) { + int k = 4 * i + j; + const int64_t tmp3 = abs_coeff[j] + round_ptr[k != 0]; + const int64_t tmp4 = ((tmp3 * quant_ptr[k != 0]) >> 16) + tmp3; + const uint32_t abs_qcoeff = + (uint32_t)((tmp4 * quant_shift_ptr[k != 0]) >> 16); + qcoeff_ptr[k] = (int)(abs_qcoeff ^ coeff_sign[j]) - coeff_sign[j]; + dqcoeff_ptr[k] = qcoeff_ptr[k] * dequant_ptr[k != 0]; + if (abs_qcoeff) eob_i = iscan[k] > eob_i ? iscan[k] : eob_i; + } + } + } + *eob_ptr = eob_i + 1; +} + +void aom_highbd_quantize_b_32x32_sse2( + const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, + const int16_t *round_ptr, const int16_t *quant_ptr, + const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, + tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, + const int16_t *scan, const int16_t *iscan) { + __m128i zbins[2]; + __m128i nzbins[2]; + int idx = 0; + int idx_arr[1024]; + int i, eob = -1; + const int zbin0_tmp = ROUND_POWER_OF_TWO(zbin_ptr[0], 1); + const int zbin1_tmp = ROUND_POWER_OF_TWO(zbin_ptr[1], 1); + (void)scan; + zbins[0] = _mm_set_epi32(zbin1_tmp, zbin1_tmp, zbin1_tmp, zbin0_tmp); + zbins[1] = _mm_set1_epi32(zbin1_tmp); + + nzbins[0] = _mm_setzero_si128(); + nzbins[1] = _mm_setzero_si128(); + nzbins[0] = _mm_sub_epi32(nzbins[0], zbins[0]); + nzbins[1] = _mm_sub_epi32(nzbins[1], zbins[1]); + + memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr)); + memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr)); + + // Pre-scan pass + for (i = 0; i < n_coeffs / 4; i++) { + __m128i coeffs, cmp1, cmp2; + int test; + coeffs = _mm_load_si128((const __m128i *)(coeff_ptr + i * 4)); + cmp1 = _mm_cmplt_epi32(coeffs, zbins[i != 0]); + cmp2 = _mm_cmpgt_epi32(coeffs, nzbins[i != 0]); + cmp1 = _mm_and_si128(cmp1, cmp2); + test = _mm_movemask_epi8(cmp1); + if (!(test & 0xf)) idx_arr[idx++] = i * 4; + if (!(test & 0xf0)) idx_arr[idx++] = i * 4 + 1; + if (!(test & 0xf00)) idx_arr[idx++] = i * 4 + 2; + if (!(test & 0xf000)) idx_arr[idx++] = i * 4 + 3; + } + + // Quantization pass: only process the coefficients selected in + // pre-scan pass. Note: idx can be zero. + for (i = 0; i < idx; i++) { + const int rc = idx_arr[i]; + const int coeff = coeff_ptr[rc]; + const int coeff_sign = (coeff >> 31); + const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign; + const int64_t tmp1 = abs_coeff + ROUND_POWER_OF_TWO(round_ptr[rc != 0], 1); + const int64_t tmp2 = ((tmp1 * quant_ptr[rc != 0]) >> 16) + tmp1; + const uint32_t abs_qcoeff = + (uint32_t)((tmp2 * quant_shift_ptr[rc != 0]) >> 15); + qcoeff_ptr[rc] = (int)(abs_qcoeff ^ coeff_sign) - coeff_sign; + dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0] / 2; + if (abs_qcoeff) eob = iscan[idx_arr[i]] > eob ? iscan[idx_arr[i]] : eob; + } + *eob_ptr = eob + 1; +} diff --git a/third_party/aom/aom_dsp/x86/highbd_sad4d_sse2.asm b/third_party/aom/aom_dsp/x86/highbd_sad4d_sse2.asm new file mode 100644 index 0000000000..e0d22522d2 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/highbd_sad4d_sse2.asm @@ -0,0 +1,296 @@ +; +; Copyright (c) 2016, Alliance for Open Media. All rights reserved +; +; This source code is subject to the terms of the BSD 2 Clause License and +; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License +; was not distributed with this source code in the LICENSE file, you can +; obtain it at www.aomedia.org/license/software. If the Alliance for Open +; Media Patent License 1.0 was not distributed with this source code in the +; PATENTS file, you can obtain it at www.aomedia.org/license/patent. +; + +; + +%include "third_party/x86inc/x86inc.asm" + +SECTION .text + +; HIGH_PROCESS_4x2x4 first, off_{first,second}_{src,ref}, advance_at_end +%macro HIGH_PROCESS_4x2x4 5-6 0 + movh m0, [srcq +%2*2] +%if %1 == 1 + movu m4, [ref1q+%3*2] + movu m5, [ref2q+%3*2] + movu m6, [ref3q+%3*2] + movu m7, [ref4q+%3*2] + movhps m0, [srcq +%4*2] + movhps m4, [ref1q+%5*2] + movhps m5, [ref2q+%5*2] + movhps m6, [ref3q+%5*2] + movhps m7, [ref4q+%5*2] + mova m3, m0 + mova m2, m0 + psubusw m3, m4 + psubusw m2, m5 + psubusw m4, m0 + psubusw m5, m0 + por m4, m3 + por m5, m2 + pmaddwd m4, m1 + pmaddwd m5, m1 + mova m3, m0 + mova m2, m0 + psubusw m3, m6 + psubusw m2, m7 + psubusw m6, m0 + psubusw m7, m0 + por m6, m3 + por m7, m2 + pmaddwd m6, m1 + pmaddwd m7, m1 +%else + movu m2, [ref1q+%3*2] + movhps m0, [srcq +%4*2] + movhps m2, [ref1q+%5*2] + mova m3, m0 + psubusw m3, m2 + psubusw m2, m0 + por m2, m3 + pmaddwd m2, m1 + paddd m4, m2 + + movu m2, [ref2q+%3*2] + mova m3, m0 + movhps m2, [ref2q+%5*2] + psubusw m3, m2 + psubusw m2, m0 + por m2, m3 + pmaddwd m2, m1 + paddd m5, m2 + + movu m2, [ref3q+%3*2] + mova m3, m0 + movhps m2, [ref3q+%5*2] + psubusw m3, m2 + psubusw m2, m0 + por m2, m3 + pmaddwd m2, m1 + paddd m6, m2 + + movu m2, [ref4q+%3*2] + mova m3, m0 + movhps m2, [ref4q+%5*2] + psubusw m3, m2 + psubusw m2, m0 + por m2, m3 + pmaddwd m2, m1 + paddd m7, m2 +%endif +%if %6 == 1 + lea srcq, [srcq +src_strideq*4] + lea ref1q, [ref1q+ref_strideq*4] + lea ref2q, [ref2q+ref_strideq*4] + lea ref3q, [ref3q+ref_strideq*4] + lea ref4q, [ref4q+ref_strideq*4] +%endif +%endmacro + +; PROCESS_8x2x4 first, off_{first,second}_{src,ref}, advance_at_end +%macro HIGH_PROCESS_8x2x4 5-6 0 + ; 1st 8 px + mova m0, [srcq +%2*2] +%if %1 == 1 + movu m4, [ref1q+%3*2] + movu m5, [ref2q+%3*2] + movu m6, [ref3q+%3*2] + movu m7, [ref4q+%3*2] + mova m3, m0 + mova m2, m0 + psubusw m3, m4 + psubusw m2, m5 + psubusw m4, m0 + psubusw m5, m0 + por m4, m3 + por m5, m2 + pmaddwd m4, m1 + pmaddwd m5, m1 + mova m3, m0 + mova m2, m0 + psubusw m3, m6 + psubusw m2, m7 + psubusw m6, m0 + psubusw m7, m0 + por m6, m3 + por m7, m2 + pmaddwd m6, m1 + pmaddwd m7, m1 +%else + mova m3, m0 + movu m2, [ref1q+%3*2] + psubusw m3, m2 + psubusw m2, m0 + por m2, m3 + mova m3, m0 + pmaddwd m2, m1 + paddd m4, m2 + movu m2, [ref2q+%3*2] + psubusw m3, m2 + psubusw m2, m0 + por m2, m3 + mova m3, m0 + pmaddwd m2, m1 + paddd m5, m2 + movu m2, [ref3q+%3*2] + psubusw m3, m2 + psubusw m2, m0 + por m2, m3 + mova m3, m0 + pmaddwd m2, m1 + paddd m6, m2 + movu m2, [ref4q+%3*2] + psubusw m3, m2 + psubusw m2, m0 + por m2, m3 + pmaddwd m2, m1 + paddd m7, m2 +%endif + + ; 2nd 8 px + mova m0, [srcq +(%4)*2] + mova m3, m0 + movu m2, [ref1q+(%5)*2] + psubusw m3, m2 + psubusw m2, m0 + por m2, m3 + mova m3, m0 + pmaddwd m2, m1 + paddd m4, m2 + movu m2, [ref2q+(%5)*2] + psubusw m3, m2 + psubusw m2, m0 + por m2, m3 + mova m3, m0 + pmaddwd m2, m1 + paddd m5, m2 + movu m2, [ref3q+(%5)*2] + psubusw m3, m2 + psubusw m2, m0 + por m2, m3 + mova m3, m0 + pmaddwd m2, m1 + paddd m6, m2 + movu m2, [ref4q+(%5)*2] + psubusw m3, m2 + psubusw m2, m0 +%if %6 == 1 + lea srcq, [srcq +src_strideq*4] + lea ref1q, [ref1q+ref_strideq*4] + lea ref2q, [ref2q+ref_strideq*4] + lea ref3q, [ref3q+ref_strideq*4] + lea ref4q, [ref4q+ref_strideq*4] +%endif + por m2, m3 + pmaddwd m2, m1 + paddd m7, m2 +%endmacro + +; HIGH_PROCESS_16x2x4 first, off_{first,second}_{src,ref}, advance_at_end +%macro HIGH_PROCESS_16x2x4 5-6 0 + HIGH_PROCESS_8x2x4 %1, %2, %3, (%2 + 8), (%3 + 8) + HIGH_PROCESS_8x2x4 0, %4, %5, (%4 + 8), (%5 + 8), %6 +%endmacro + +; HIGH_PROCESS_32x2x4 first, off_{first,second}_{src,ref}, advance_at_end +%macro HIGH_PROCESS_32x2x4 5-6 0 + HIGH_PROCESS_16x2x4 %1, %2, %3, (%2 + 16), (%3 + 16) + HIGH_PROCESS_16x2x4 0, %4, %5, (%4 + 16), (%5 + 16), %6 +%endmacro + +; HIGH_PROCESS_64x2x4 first, off_{first,second}_{src,ref}, advance_at_end +%macro HIGH_PROCESS_64x2x4 5-6 0 + HIGH_PROCESS_32x2x4 %1, %2, %3, (%2 + 32), (%3 + 32) + HIGH_PROCESS_32x2x4 0, %4, %5, (%4 + 32), (%5 + 32), %6 +%endmacro + +; void aom_highbd_sadNxNx4d_sse2(uint8_t *src, int src_stride, +; uint8_t *ref[4], int ref_stride, +; uint32_t res[4]); +; where NxN = 64x64, 32x32, 16x16, 16x8, 8x16 or 8x8 +%macro HIGH_SADNXN4D 2 +%if UNIX64 +cglobal highbd_sad%1x%2x4d, 5, 8, 8, src, src_stride, ref1, ref_stride, \ + res, ref2, ref3, ref4 +%else +cglobal highbd_sad%1x%2x4d, 4, 7, 8, src, src_stride, ref1, ref_stride, \ + ref2, ref3, ref4 +%endif + +; set m1 + push srcq + mov srcd, 0x00010001 + movd m1, srcd + pshufd m1, m1, 0x0 + pop srcq + + movsxdifnidn src_strideq, src_strided + movsxdifnidn ref_strideq, ref_strided + mov ref2q, [ref1q+gprsize*1] + mov ref3q, [ref1q+gprsize*2] + mov ref4q, [ref1q+gprsize*3] + mov ref1q, [ref1q+gprsize*0] + +; convert byte pointers to short pointers + shl srcq, 1 + shl ref2q, 1 + shl ref3q, 1 + shl ref4q, 1 + shl ref1q, 1 + + HIGH_PROCESS_%1x2x4 1, 0, 0, src_strideq, ref_strideq, 1 +%rep (%2-4)/2 + HIGH_PROCESS_%1x2x4 0, 0, 0, src_strideq, ref_strideq, 1 +%endrep + HIGH_PROCESS_%1x2x4 0, 0, 0, src_strideq, ref_strideq, 0 + ; N.B. HIGH_PROCESS outputs dwords (32 bits) + ; so in high bit depth even the smallest width (4) needs 128bits i.e. XMM + movhlps m0, m4 + movhlps m1, m5 + movhlps m2, m6 + movhlps m3, m7 + paddd m4, m0 + paddd m5, m1 + paddd m6, m2 + paddd m7, m3 + punpckldq m4, m5 + punpckldq m6, m7 + movhlps m0, m4 + movhlps m1, m6 + paddd m4, m0 + paddd m6, m1 + punpcklqdq m4, m6 + movifnidn r4, r4mp + movu [r4], m4 + RET +%endmacro + + +INIT_XMM sse2 +HIGH_SADNXN4D 64, 64 +HIGH_SADNXN4D 64, 32 +HIGH_SADNXN4D 32, 64 +HIGH_SADNXN4D 32, 32 +HIGH_SADNXN4D 32, 16 +HIGH_SADNXN4D 16, 32 +HIGH_SADNXN4D 16, 16 +HIGH_SADNXN4D 16, 8 +HIGH_SADNXN4D 8, 16 +HIGH_SADNXN4D 8, 8 +HIGH_SADNXN4D 8, 4 +HIGH_SADNXN4D 4, 8 +HIGH_SADNXN4D 4, 4 +HIGH_SADNXN4D 4, 16 +HIGH_SADNXN4D 16, 4 +HIGH_SADNXN4D 8, 32 +HIGH_SADNXN4D 32, 8 +HIGH_SADNXN4D 16, 64 +HIGH_SADNXN4D 64, 16 diff --git a/third_party/aom/aom_dsp/x86/highbd_sad_sse2.asm b/third_party/aom/aom_dsp/x86/highbd_sad_sse2.asm new file mode 100644 index 0000000000..3398d8a2ae --- /dev/null +++ b/third_party/aom/aom_dsp/x86/highbd_sad_sse2.asm @@ -0,0 +1,374 @@ +; +; Copyright (c) 2016, Alliance for Open Media. All rights reserved +; +; This source code is subject to the terms of the BSD 2 Clause License and +; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License +; was not distributed with this source code in the LICENSE file, you can +; obtain it at www.aomedia.org/license/software. If the Alliance for Open +; Media Patent License 1.0 was not distributed with this source code in the +; PATENTS file, you can obtain it at www.aomedia.org/license/patent. +; + +; + +%include "third_party/x86inc/x86inc.asm" + +SECTION .text + +%macro HIGH_SAD_FN 4 +%if %4 == 0 +%if %3 == 5 +cglobal highbd_sad%1x%2, 4, %3, 7, src, src_stride, ref, ref_stride, n_rows +%else ; %3 == 7 +cglobal highbd_sad%1x%2, 4, %3, 7, src, src_stride, ref, ref_stride, \ + src_stride3, ref_stride3, n_rows +%endif ; %3 == 5/7 +%else ; avg +%if %3 == 5 +cglobal highbd_sad%1x%2_avg, 5, 1 + %3, 7, src, src_stride, ref, ref_stride, \ + second_pred, n_rows +%else ; %3 == 7 +cglobal highbd_sad%1x%2_avg, 5, ARCH_X86_64 + %3, 7, src, src_stride, \ + ref, ref_stride, \ + second_pred, \ + src_stride3, ref_stride3 +%if ARCH_X86_64 +%define n_rowsd r7d +%else ; x86-32 +%define n_rowsd dword r0m +%endif ; x86-32/64 +%endif ; %3 == 5/7 +%endif ; avg/sad + movsxdifnidn src_strideq, src_strided + movsxdifnidn ref_strideq, ref_strided +%if %3 == 7 + lea src_stride3q, [src_strideq*3] + lea ref_stride3q, [ref_strideq*3] +%endif ; %3 == 7 +; convert src, ref & second_pred to short ptrs (from byte ptrs) + shl srcq, 1 + shl refq, 1 +%if %4 == 1 + shl second_predq, 1 +%endif +%endmacro + +; unsigned int aom_highbd_sad64x{16,32,64}_sse2(uint8_t *src, int src_stride, +; uint8_t *ref, int ref_stride); +%macro HIGH_SAD64XN 1-2 0 + HIGH_SAD_FN 64, %1, 5, %2 + mov n_rowsd, %1 + pxor m0, m0 + pxor m6, m6 + +.loop: + ; first half of each row + movu m1, [refq] + movu m2, [refq+16] + movu m3, [refq+32] + movu m4, [refq+48] +%if %2 == 1 + pavgw m1, [second_predq+mmsize*0] + pavgw m2, [second_predq+mmsize*1] + pavgw m3, [second_predq+mmsize*2] + pavgw m4, [second_predq+mmsize*3] + lea second_predq, [second_predq+mmsize*4] +%endif + mova m5, [srcq] + psubusw m5, m1 + psubusw m1, [srcq] + por m1, m5 + mova m5, [srcq+16] + psubusw m5, m2 + psubusw m2, [srcq+16] + por m2, m5 + mova m5, [srcq+32] + psubusw m5, m3 + psubusw m3, [srcq+32] + por m3, m5 + mova m5, [srcq+48] + psubusw m5, m4 + psubusw m4, [srcq+48] + por m4, m5 + paddw m1, m2 + paddw m3, m4 + movhlps m2, m1 + movhlps m4, m3 + paddw m1, m2 + paddw m3, m4 + punpcklwd m1, m6 + punpcklwd m3, m6 + paddd m0, m1 + paddd m0, m3 + ; second half of each row + movu m1, [refq+64] + movu m2, [refq+80] + movu m3, [refq+96] + movu m4, [refq+112] +%if %2 == 1 + pavgw m1, [second_predq+mmsize*0] + pavgw m2, [second_predq+mmsize*1] + pavgw m3, [second_predq+mmsize*2] + pavgw m4, [second_predq+mmsize*3] + lea second_predq, [second_predq+mmsize*4] +%endif + mova m5, [srcq+64] + psubusw m5, m1 + psubusw m1, [srcq+64] + por m1, m5 + mova m5, [srcq+80] + psubusw m5, m2 + psubusw m2, [srcq+80] + por m2, m5 + mova m5, [srcq+96] + psubusw m5, m3 + psubusw m3, [srcq+96] + por m3, m5 + mova m5, [srcq+112] + psubusw m5, m4 + psubusw m4, [srcq+112] + por m4, m5 + paddw m1, m2 + paddw m3, m4 + movhlps m2, m1 + movhlps m4, m3 + paddw m1, m2 + paddw m3, m4 + punpcklwd m1, m6 + punpcklwd m3, m6 + lea refq, [refq+ref_strideq*2] + paddd m0, m1 + lea srcq, [srcq+src_strideq*2] + paddd m0, m3 + + dec n_rowsd + jg .loop + + movhlps m1, m0 + paddd m0, m1 + punpckldq m0, m6 + movhlps m1, m0 + paddd m0, m1 + movd eax, m0 + RET +%endmacro + +INIT_XMM sse2 +HIGH_SAD64XN 64 ; highbd_sad64x64_sse2 +HIGH_SAD64XN 32 ; highbd_sad64x32_sse2 +HIGH_SAD64XN 64, 1 ; highbd_sad64x64_avg_sse2 +HIGH_SAD64XN 32, 1 ; highbd_sad64x32_avg_sse2 +HIGH_SAD64XN 16 ; highbd_sad_64x16_sse2 +HIGH_SAD64XN 16, 1 ; highbd_sad_64x16_avg_sse2 + +; unsigned int aom_highbd_sad32x{16,32,64}_sse2(uint8_t *src, int src_stride, +; uint8_t *ref, int ref_stride); +%macro HIGH_SAD32XN 1-2 0 + HIGH_SAD_FN 32, %1, 5, %2 + mov n_rowsd, %1 + pxor m0, m0 + pxor m6, m6 + +.loop: + movu m1, [refq] + movu m2, [refq+16] + movu m3, [refq+32] + movu m4, [refq+48] +%if %2 == 1 + pavgw m1, [second_predq+mmsize*0] + pavgw m2, [second_predq+mmsize*1] + pavgw m3, [second_predq+mmsize*2] + pavgw m4, [second_predq+mmsize*3] + lea second_predq, [second_predq+mmsize*4] +%endif + mova m5, [srcq] + psubusw m5, m1 + psubusw m1, [srcq] + por m1, m5 + mova m5, [srcq+16] + psubusw m5, m2 + psubusw m2, [srcq+16] + por m2, m5 + mova m5, [srcq+32] + psubusw m5, m3 + psubusw m3, [srcq+32] + por m3, m5 + mova m5, [srcq+48] + psubusw m5, m4 + psubusw m4, [srcq+48] + por m4, m5 + paddw m1, m2 + paddw m3, m4 + movhlps m2, m1 + movhlps m4, m3 + paddw m1, m2 + paddw m3, m4 + punpcklwd m1, m6 + punpcklwd m3, m6 + lea refq, [refq+ref_strideq*2] + paddd m0, m1 + lea srcq, [srcq+src_strideq*2] + paddd m0, m3 + dec n_rowsd + jg .loop + + movhlps m1, m0 + paddd m0, m1 + punpckldq m0, m6 + movhlps m1, m0 + paddd m0, m1 + movd eax, m0 + RET +%endmacro + +INIT_XMM sse2 +HIGH_SAD32XN 64 ; highbd_sad32x64_sse2 +HIGH_SAD32XN 32 ; highbd_sad32x32_sse2 +HIGH_SAD32XN 16 ; highbd_sad32x16_sse2 +HIGH_SAD32XN 64, 1 ; highbd_sad32x64_avg_sse2 +HIGH_SAD32XN 32, 1 ; highbd_sad32x32_avg_sse2 +HIGH_SAD32XN 16, 1 ; highbd_sad32x16_avg_sse2 +HIGH_SAD32XN 8 ; highbd_sad_32x8_sse2 +HIGH_SAD32XN 8, 1 ; highbd_sad_32x8_avg_sse2 + +; unsigned int aom_highbd_sad16x{8,16,32}_sse2(uint8_t *src, int src_stride, +; uint8_t *ref, int ref_stride); +%macro HIGH_SAD16XN 1-2 0 + HIGH_SAD_FN 16, %1, 5, %2 + mov n_rowsd, %1/2 + pxor m0, m0 + pxor m6, m6 + +.loop: + movu m1, [refq] + movu m2, [refq+16] + movu m3, [refq+ref_strideq*2] + movu m4, [refq+ref_strideq*2+16] +%if %2 == 1 + pavgw m1, [second_predq+mmsize*0] + pavgw m2, [second_predq+16] + pavgw m3, [second_predq+mmsize*2] + pavgw m4, [second_predq+mmsize*2+16] + lea second_predq, [second_predq+mmsize*4] +%endif + mova m5, [srcq] + psubusw m5, m1 + psubusw m1, [srcq] + por m1, m5 + mova m5, [srcq+16] + psubusw m5, m2 + psubusw m2, [srcq+16] + por m2, m5 + mova m5, [srcq+src_strideq*2] + psubusw m5, m3 + psubusw m3, [srcq+src_strideq*2] + por m3, m5 + mova m5, [srcq+src_strideq*2+16] + psubusw m5, m4 + psubusw m4, [srcq+src_strideq*2+16] + por m4, m5 + paddw m1, m2 + paddw m3, m4 + movhlps m2, m1 + movhlps m4, m3 + paddw m1, m2 + paddw m3, m4 + punpcklwd m1, m6 + punpcklwd m3, m6 + lea refq, [refq+ref_strideq*4] + paddd m0, m1 + lea srcq, [srcq+src_strideq*4] + paddd m0, m3 + dec n_rowsd + jg .loop + + movhlps m1, m0 + paddd m0, m1 + punpckldq m0, m6 + movhlps m1, m0 + paddd m0, m1 + movd eax, m0 + RET +%endmacro + +INIT_XMM sse2 +HIGH_SAD16XN 32 ; highbd_sad16x32_sse2 +HIGH_SAD16XN 16 ; highbd_sad16x16_sse2 +HIGH_SAD16XN 8 ; highbd_sad16x8_sse2 +HIGH_SAD16XN 32, 1 ; highbd_sad16x32_avg_sse2 +HIGH_SAD16XN 16, 1 ; highbd_sad16x16_avg_sse2 +HIGH_SAD16XN 8, 1 ; highbd_sad16x8_avg_sse2 +HIGH_SAD16XN 4 ; highbd_sad_16x4_sse2 +HIGH_SAD16XN 4, 1 ; highbd_sad_16x4_avg_sse2 +HIGH_SAD16XN 64 ; highbd_sad_16x64_sse2 +HIGH_SAD16XN 64, 1 ; highbd_sad_16x64_avg_sse2 + +; unsigned int aom_highbd_sad8x{4,8,16}_sse2(uint8_t *src, int src_stride, +; uint8_t *ref, int ref_stride); +%macro HIGH_SAD8XN 1-2 0 + HIGH_SAD_FN 8, %1, 7, %2 + mov n_rowsd, %1/4 + pxor m0, m0 + pxor m6, m6 + +.loop: + movu m1, [refq] + movu m2, [refq+ref_strideq*2] + movu m3, [refq+ref_strideq*4] + movu m4, [refq+ref_stride3q*2] +%if %2 == 1 + pavgw m1, [second_predq+mmsize*0] + pavgw m2, [second_predq+mmsize*1] + pavgw m3, [second_predq+mmsize*2] + pavgw m4, [second_predq+mmsize*3] + lea second_predq, [second_predq+mmsize*4] +%endif + mova m5, [srcq] + psubusw m5, m1 + psubusw m1, [srcq] + por m1, m5 + mova m5, [srcq+src_strideq*2] + psubusw m5, m2 + psubusw m2, [srcq+src_strideq*2] + por m2, m5 + mova m5, [srcq+src_strideq*4] + psubusw m5, m3 + psubusw m3, [srcq+src_strideq*4] + por m3, m5 + mova m5, [srcq+src_stride3q*2] + psubusw m5, m4 + psubusw m4, [srcq+src_stride3q*2] + por m4, m5 + paddw m1, m2 + paddw m3, m4 + movhlps m2, m1 + movhlps m4, m3 + paddw m1, m2 + paddw m3, m4 + punpcklwd m1, m6 + punpcklwd m3, m6 + lea refq, [refq+ref_strideq*8] + paddd m0, m1 + lea srcq, [srcq+src_strideq*8] + paddd m0, m3 + dec n_rowsd + jg .loop + + movhlps m1, m0 + paddd m0, m1 + punpckldq m0, m6 + movhlps m1, m0 + paddd m0, m1 + movd eax, m0 + RET +%endmacro + +INIT_XMM sse2 +HIGH_SAD8XN 16 ; highbd_sad8x16_sse2 +HIGH_SAD8XN 8 ; highbd_sad8x8_sse2 +HIGH_SAD8XN 4 ; highbd_sad8x4_sse2 +HIGH_SAD8XN 16, 1 ; highbd_sad8x16_avg_sse2 +HIGH_SAD8XN 8, 1 ; highbd_sad8x8_avg_sse2 +HIGH_SAD8XN 4, 1 ; highbd_sad8x4_avg_sse2 +HIGH_SAD8XN 32 ; highbd_sad_8x32_sse2 +HIGH_SAD8XN 32, 1 ; highbd_sad_8x32_avg_sse2 diff --git a/third_party/aom/aom_dsp/x86/highbd_subpel_variance_impl_sse2.asm b/third_party/aom/aom_dsp/x86/highbd_subpel_variance_impl_sse2.asm new file mode 100644 index 0000000000..61f5b8e865 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/highbd_subpel_variance_impl_sse2.asm @@ -0,0 +1,1036 @@ +; +; Copyright (c) 2016, Alliance for Open Media. All rights reserved +; +; This source code is subject to the terms of the BSD 2 Clause License and +; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License +; was not distributed with this source code in the LICENSE file, you can +; obtain it at www.aomedia.org/license/software. If the Alliance for Open +; Media Patent License 1.0 was not distributed with this source code in the +; PATENTS file, you can obtain it at www.aomedia.org/license/patent. +; + +; + +%include "third_party/x86inc/x86inc.asm" + +SECTION_RODATA +pw_8: times 8 dw 8 +bilin_filter_m_sse2: times 8 dw 16 + times 8 dw 0 + times 8 dw 14 + times 8 dw 2 + times 8 dw 12 + times 8 dw 4 + times 8 dw 10 + times 8 dw 6 + times 16 dw 8 + times 8 dw 6 + times 8 dw 10 + times 8 dw 4 + times 8 dw 12 + times 8 dw 2 + times 8 dw 14 + +SECTION .text + +; int aom_sub_pixel_varianceNxh(const uint8_t *src, ptrdiff_t src_stride, +; int x_offset, int y_offset, +; const uint8_t *dst, ptrdiff_t dst_stride, +; int height, unsigned int *sse); +; +; This function returns the SE and stores SSE in the given pointer. + +%macro SUM_SSE 6 ; src1, dst1, src2, dst2, sum, sse + psubw %3, %4 + psubw %1, %2 + mova %4, %3 ; make copies to manipulate to calc sum + mova %2, %1 ; use originals for calc sse + pmaddwd %3, %3 + paddw %4, %2 + pmaddwd %1, %1 + movhlps %2, %4 + paddd %6, %3 + paddw %4, %2 + pxor %2, %2 + pcmpgtw %2, %4 ; mask for 0 > %4 (sum) + punpcklwd %4, %2 ; sign-extend word to dword + paddd %6, %1 + paddd %5, %4 + +%endmacro + +%macro STORE_AND_RET 0 +%if mmsize == 16 + ; if H=64 and W=16, we have 8 words of each 2(1bit)x64(6bit)x9bit=16bit + ; in m6, i.e. it _exactly_ fits in a signed word per word in the xmm reg. + ; We have to sign-extend it before adding the words within the register + ; and outputing to a dword. + movhlps m3, m7 + movhlps m4, m6 + paddd m7, m3 + paddd m6, m4 + pshufd m3, m7, 0x1 + pshufd m4, m6, 0x1 + paddd m7, m3 + paddd m6, m4 + mov r1, ssem ; r1 = unsigned int *sse + movd [r1], m7 ; store sse + movd eax, m6 ; store sum as return value +%endif + RET +%endmacro + +%macro INC_SRC_BY_SRC_STRIDE 0 +%if ARCH_X86=1 && CONFIG_PIC=1 + add srcq, src_stridemp + add srcq, src_stridemp +%else + lea srcq, [srcq + src_strideq*2] +%endif +%endmacro + +%macro SUBPEL_VARIANCE 1-2 0 ; W +%define bilin_filter_m bilin_filter_m_sse2 +%define filter_idx_shift 5 + + +%if ARCH_X86_64 + %if %2 == 1 ; avg + cglobal highbd_sub_pixel_avg_variance%1xh, 9, 10, 13, src, src_stride, \ + x_offset, y_offset, \ + dst, dst_stride, \ + sec, sec_stride, height, sse + %define sec_str sec_strideq + %else + cglobal highbd_sub_pixel_variance%1xh, 7, 8, 13, src, src_stride, \ + x_offset, y_offset, \ + dst, dst_stride, height, sse + %endif + %define block_height heightd + %define bilin_filter sseq +%else + %if CONFIG_PIC=1 + %if %2 == 1 ; avg + cglobal highbd_sub_pixel_avg_variance%1xh, 7, 7, 13, src, src_stride, \ + x_offset, y_offset, \ + dst, dst_stride, \ + sec, sec_stride, height, sse, \ + g_bilin_filter, g_pw_8 + %define block_height dword heightm + %define sec_str sec_stridemp + + ; Store bilin_filter and pw_8 location in stack + %if GET_GOT_DEFINED == 1 + GET_GOT eax + add esp, 4 ; restore esp + %endif + + lea ecx, [GLOBAL(bilin_filter_m)] + mov g_bilin_filterm, ecx + + lea ecx, [GLOBAL(pw_8)] + mov g_pw_8m, ecx + + LOAD_IF_USED 0, 1 ; load eax, ecx back + %else + cglobal highbd_sub_pixel_variance%1xh, 7, 7, 13, src, src_stride, \ + x_offset, y_offset, \ + dst, dst_stride, height, sse, \ + g_bilin_filter, g_pw_8 + %define block_height heightd + + ; Store bilin_filter and pw_8 location in stack + %if GET_GOT_DEFINED == 1 + GET_GOT eax + add esp, 4 ; restore esp + %endif + + lea ecx, [GLOBAL(bilin_filter_m)] + mov g_bilin_filterm, ecx + + lea ecx, [GLOBAL(pw_8)] + mov g_pw_8m, ecx + + LOAD_IF_USED 0, 1 ; load eax, ecx back + %endif + %else + %if %2 == 1 ; avg + cglobal highbd_sub_pixel_avg_variance%1xh, 7, 7, 13, src, src_stride, \ + x_offset, y_offset, \ + dst, dst_stride, \ + sec, sec_stride, height, sse + %define block_height dword heightm + %define sec_str sec_stridemp + %else + cglobal highbd_sub_pixel_variance%1xh, 7, 7, 13, src, src_stride, \ + x_offset, y_offset, \ + dst, dst_stride, height, sse + %define block_height heightd + %endif + + %define bilin_filter bilin_filter_m + %endif +%endif + + ASSERT %1 <= 16 ; m6 overflows if w > 16 + pxor m6, m6 ; sum + pxor m7, m7 ; sse + +%if %1 < 16 + sar block_height, 1 +%endif +%if %2 == 1 ; avg + shl sec_str, 1 +%endif + + ; FIXME(rbultje) replace by jumptable? + test x_offsetd, x_offsetd + jnz .x_nonzero + ; x_offset == 0 + test y_offsetd, y_offsetd + jnz .x_zero_y_nonzero + + ; x_offset == 0 && y_offset == 0 +.x_zero_y_zero_loop: +%if %1 == 16 + movu m0, [srcq] + movu m2, [srcq + 16] + mova m1, [dstq] + mova m3, [dstq + 16] +%if %2 == 1 ; avg + pavgw m0, [secq] + pavgw m2, [secq+16] +%endif + SUM_SSE m0, m1, m2, m3, m6, m7 + + lea srcq, [srcq + src_strideq*2] + lea dstq, [dstq + dst_strideq*2] +%if %2 == 1 ; avg + add secq, sec_str +%endif +%else ; %1 < 16 + movu m0, [srcq] + movu m2, [srcq + src_strideq*2] + mova m1, [dstq] + mova m3, [dstq + dst_strideq*2] +%if %2 == 1 ; avg + pavgw m0, [secq] + add secq, sec_str + pavgw m2, [secq] +%endif + SUM_SSE m0, m1, m2, m3, m6, m7 + + lea srcq, [srcq + src_strideq*4] + lea dstq, [dstq + dst_strideq*4] +%if %2 == 1 ; avg + add secq, sec_str +%endif +%endif + dec block_height + jg .x_zero_y_zero_loop + STORE_AND_RET + +.x_zero_y_nonzero: + cmp y_offsetd, 8 + jne .x_zero_y_nonhalf + + ; x_offset == 0 && y_offset == 0.5 +.x_zero_y_half_loop: +%if %1 == 16 + movu m0, [srcq] + movu m1, [srcq+16] + movu m4, [srcq+src_strideq*2] + movu m5, [srcq+src_strideq*2+16] + mova m2, [dstq] + mova m3, [dstq+16] + pavgw m0, m4 + pavgw m1, m5 +%if %2 == 1 ; avg + pavgw m0, [secq] + pavgw m1, [secq+16] +%endif + SUM_SSE m0, m2, m1, m3, m6, m7 + + lea srcq, [srcq + src_strideq*2] + lea dstq, [dstq + dst_strideq*2] +%if %2 == 1 ; avg + add secq, sec_str +%endif +%else ; %1 < 16 + movu m0, [srcq] + movu m1, [srcq+src_strideq*2] + movu m5, [srcq+src_strideq*4] + mova m2, [dstq] + mova m3, [dstq+dst_strideq*2] + pavgw m0, m1 + pavgw m1, m5 +%if %2 == 1 ; avg + pavgw m0, [secq] + add secq, sec_str + pavgw m1, [secq] +%endif + SUM_SSE m0, m2, m1, m3, m6, m7 + + lea srcq, [srcq + src_strideq*4] + lea dstq, [dstq + dst_strideq*4] +%if %2 == 1 ; avg + add secq, sec_str +%endif +%endif + dec block_height + jg .x_zero_y_half_loop + STORE_AND_RET + +.x_zero_y_nonhalf: + ; x_offset == 0 && y_offset == bilin interpolation +%if ARCH_X86_64 + lea bilin_filter, [GLOBAL(bilin_filter_m)] +%endif + shl y_offsetd, filter_idx_shift +%if ARCH_X86_64 && mmsize == 16 + mova m8, [bilin_filter+y_offsetq] + mova m9, [bilin_filter+y_offsetq+16] + mova m10, [GLOBAL(pw_8)] +%define filter_y_a m8 +%define filter_y_b m9 +%define filter_rnd m10 +%else ; x86-32 or mmx +%if ARCH_X86=1 && CONFIG_PIC=1 +; x_offset == 0, reuse x_offset reg +%define tempq x_offsetq + add y_offsetq, g_bilin_filterm +%define filter_y_a [y_offsetq] +%define filter_y_b [y_offsetq+16] + mov tempq, g_pw_8m +%define filter_rnd [tempq] +%else + add y_offsetq, bilin_filter +%define filter_y_a [y_offsetq] +%define filter_y_b [y_offsetq+16] +%define filter_rnd [GLOBAL(pw_8)] +%endif +%endif + +.x_zero_y_other_loop: +%if %1 == 16 + movu m0, [srcq] + movu m1, [srcq + 16] + movu m4, [srcq+src_strideq*2] + movu m5, [srcq+src_strideq*2+16] + mova m2, [dstq] + mova m3, [dstq+16] + ; FIXME(rbultje) instead of out=((num-x)*in1+x*in2+rnd)>>log2(num), we can + ; also do out=in1+(((num-x)*(in2-in1)+rnd)>>log2(num)). Total number of + ; instructions is the same (5), but it is 1 mul instead of 2, so might be + ; slightly faster because of pmullw latency. It would also cut our rodata + ; tables in half for this function, and save 1-2 registers on x86-64. + pmullw m1, filter_y_a + pmullw m5, filter_y_b + paddw m1, filter_rnd + pmullw m0, filter_y_a + pmullw m4, filter_y_b + paddw m0, filter_rnd + paddw m1, m5 + paddw m0, m4 + psrlw m1, 4 + psrlw m0, 4 +%if %2 == 1 ; avg + pavgw m0, [secq] + pavgw m1, [secq+16] +%endif + SUM_SSE m0, m2, m1, m3, m6, m7 + + lea srcq, [srcq + src_strideq*2] + lea dstq, [dstq + dst_strideq*2] +%if %2 == 1 ; avg + add secq, sec_str +%endif +%else ; %1 < 16 + movu m0, [srcq] + movu m1, [srcq+src_strideq*2] + movu m5, [srcq+src_strideq*4] + mova m4, m1 + mova m2, [dstq] + mova m3, [dstq+dst_strideq*2] + pmullw m1, filter_y_a + pmullw m5, filter_y_b + paddw m1, filter_rnd + pmullw m0, filter_y_a + pmullw m4, filter_y_b + paddw m0, filter_rnd + paddw m1, m5 + paddw m0, m4 + psrlw m1, 4 + psrlw m0, 4 +%if %2 == 1 ; avg + pavgw m0, [secq] + add secq, sec_str + pavgw m1, [secq] +%endif + SUM_SSE m0, m2, m1, m3, m6, m7 + + lea srcq, [srcq + src_strideq*4] + lea dstq, [dstq + dst_strideq*4] +%if %2 == 1 ; avg + add secq, sec_str +%endif +%endif + dec block_height + jg .x_zero_y_other_loop +%undef filter_y_a +%undef filter_y_b +%undef filter_rnd + STORE_AND_RET + +.x_nonzero: + cmp x_offsetd, 8 + jne .x_nonhalf + ; x_offset == 0.5 + test y_offsetd, y_offsetd + jnz .x_half_y_nonzero + + ; x_offset == 0.5 && y_offset == 0 +.x_half_y_zero_loop: +%if %1 == 16 + movu m0, [srcq] + movu m1, [srcq + 16] + movu m4, [srcq + 2] + movu m5, [srcq + 18] + mova m2, [dstq] + mova m3, [dstq + 16] + pavgw m0, m4 + pavgw m1, m5 +%if %2 == 1 ; avg + pavgw m0, [secq] + pavgw m1, [secq+16] +%endif + SUM_SSE m0, m2, m1, m3, m6, m7 + + lea srcq, [srcq + src_strideq*2] + lea dstq, [dstq + dst_strideq*2] +%if %2 == 1 ; avg + add secq, sec_str +%endif +%else ; %1 < 16 + movu m0, [srcq] + movu m1, [srcq + src_strideq*2] + movu m4, [srcq + 2] + movu m5, [srcq + src_strideq*2 + 2] + mova m2, [dstq] + mova m3, [dstq + dst_strideq*2] + pavgw m0, m4 + pavgw m1, m5 +%if %2 == 1 ; avg + pavgw m0, [secq] + add secq, sec_str + pavgw m1, [secq] +%endif + SUM_SSE m0, m2, m1, m3, m6, m7 + + lea srcq, [srcq + src_strideq*4] + lea dstq, [dstq + dst_strideq*4] +%if %2 == 1 ; avg + add secq, sec_str +%endif +%endif + dec block_height + jg .x_half_y_zero_loop + STORE_AND_RET + +.x_half_y_nonzero: + cmp y_offsetd, 8 + jne .x_half_y_nonhalf + + ; x_offset == 0.5 && y_offset == 0.5 +%if %1 == 16 + movu m0, [srcq] + movu m1, [srcq+16] + movu m2, [srcq+2] + movu m3, [srcq+18] + lea srcq, [srcq + src_strideq*2] + pavgw m0, m2 + pavgw m1, m3 +.x_half_y_half_loop: + movu m2, [srcq] + movu m3, [srcq + 16] + movu m4, [srcq + 2] + movu m5, [srcq + 18] + pavgw m2, m4 + pavgw m3, m5 + pavgw m0, m2 + pavgw m1, m3 + mova m4, [dstq] + mova m5, [dstq + 16] +%if %2 == 1 ; avg + pavgw m0, [secq] + pavgw m1, [secq+16] +%endif + SUM_SSE m0, m4, m1, m5, m6, m7 + mova m0, m2 + mova m1, m3 + + lea srcq, [srcq + src_strideq*2] + lea dstq, [dstq + dst_strideq*2] +%if %2 == 1 ; avg + add secq, sec_str +%endif +%else ; %1 < 16 + movu m0, [srcq] + movu m2, [srcq+2] + lea srcq, [srcq + src_strideq*2] + pavgw m0, m2 +.x_half_y_half_loop: + movu m2, [srcq] + movu m3, [srcq + src_strideq*2] + movu m4, [srcq + 2] + movu m5, [srcq + src_strideq*2 + 2] + pavgw m2, m4 + pavgw m3, m5 + pavgw m0, m2 + pavgw m2, m3 + mova m4, [dstq] + mova m5, [dstq + dst_strideq*2] +%if %2 == 1 ; avg + pavgw m0, [secq] + add secq, sec_str + pavgw m2, [secq] +%endif + SUM_SSE m0, m4, m2, m5, m6, m7 + mova m0, m3 + + lea srcq, [srcq + src_strideq*4] + lea dstq, [dstq + dst_strideq*4] +%if %2 == 1 ; avg + add secq, sec_str +%endif +%endif + dec block_height + jg .x_half_y_half_loop + STORE_AND_RET + +.x_half_y_nonhalf: + ; x_offset == 0.5 && y_offset == bilin interpolation +%if ARCH_X86_64 + lea bilin_filter, [GLOBAL(bilin_filter_m)] +%endif + shl y_offsetd, filter_idx_shift +%if ARCH_X86_64 && mmsize == 16 + mova m8, [bilin_filter+y_offsetq] + mova m9, [bilin_filter+y_offsetq+16] + mova m10, [GLOBAL(pw_8)] +%define filter_y_a m8 +%define filter_y_b m9 +%define filter_rnd m10 +%else ; x86_32 +%if ARCH_X86=1 && CONFIG_PIC=1 +; x_offset == 0.5. We can reuse x_offset reg +%define tempq x_offsetq + add y_offsetq, g_bilin_filterm +%define filter_y_a [y_offsetq] +%define filter_y_b [y_offsetq+16] + mov tempq, g_pw_8m +%define filter_rnd [tempq] +%else + add y_offsetq, bilin_filter +%define filter_y_a [y_offsetq] +%define filter_y_b [y_offsetq+16] +%define filter_rnd [GLOBAL(pw_8)] +%endif +%endif + +%if %1 == 16 + movu m0, [srcq] + movu m1, [srcq+16] + movu m2, [srcq+2] + movu m3, [srcq+18] + lea srcq, [srcq + src_strideq*2] + pavgw m0, m2 + pavgw m1, m3 +.x_half_y_other_loop: + movu m2, [srcq] + movu m3, [srcq+16] + movu m4, [srcq+2] + movu m5, [srcq+18] + pavgw m2, m4 + pavgw m3, m5 + mova m4, m2 + mova m5, m3 + pmullw m1, filter_y_a + pmullw m3, filter_y_b + paddw m1, filter_rnd + paddw m1, m3 + pmullw m0, filter_y_a + pmullw m2, filter_y_b + paddw m0, filter_rnd + psrlw m1, 4 + paddw m0, m2 + mova m2, [dstq] + psrlw m0, 4 + mova m3, [dstq+16] +%if %2 == 1 ; avg + pavgw m0, [secq] + pavgw m1, [secq+16] +%endif + SUM_SSE m0, m2, m1, m3, m6, m7 + mova m0, m4 + mova m1, m5 + + lea srcq, [srcq + src_strideq*2] + lea dstq, [dstq + dst_strideq*2] +%if %2 == 1 ; avg + add secq, sec_str +%endif +%else ; %1 < 16 + movu m0, [srcq] + movu m2, [srcq+2] + lea srcq, [srcq + src_strideq*2] + pavgw m0, m2 +.x_half_y_other_loop: + movu m2, [srcq] + movu m3, [srcq+src_strideq*2] + movu m4, [srcq+2] + movu m5, [srcq+src_strideq*2+2] + pavgw m2, m4 + pavgw m3, m5 + mova m4, m2 + mova m5, m3 + pmullw m4, filter_y_a + pmullw m3, filter_y_b + paddw m4, filter_rnd + paddw m4, m3 + pmullw m0, filter_y_a + pmullw m2, filter_y_b + paddw m0, filter_rnd + psrlw m4, 4 + paddw m0, m2 + mova m2, [dstq] + psrlw m0, 4 + mova m3, [dstq+dst_strideq*2] +%if %2 == 1 ; avg + pavgw m0, [secq] + add secq, sec_str + pavgw m4, [secq] +%endif + SUM_SSE m0, m2, m4, m3, m6, m7 + mova m0, m5 + + lea srcq, [srcq + src_strideq*4] + lea dstq, [dstq + dst_strideq*4] +%if %2 == 1 ; avg + add secq, sec_str +%endif +%endif + dec block_height + jg .x_half_y_other_loop +%undef filter_y_a +%undef filter_y_b +%undef filter_rnd + STORE_AND_RET + +.x_nonhalf: + test y_offsetd, y_offsetd + jnz .x_nonhalf_y_nonzero + + ; x_offset == bilin interpolation && y_offset == 0 +%if ARCH_X86_64 + lea bilin_filter, [GLOBAL(bilin_filter_m)] +%endif + shl x_offsetd, filter_idx_shift +%if ARCH_X86_64 && mmsize == 16 + mova m8, [bilin_filter+x_offsetq] + mova m9, [bilin_filter+x_offsetq+16] + mova m10, [GLOBAL(pw_8)] +%define filter_x_a m8 +%define filter_x_b m9 +%define filter_rnd m10 +%else ; x86-32 +%if ARCH_X86=1 && CONFIG_PIC=1 +; y_offset == 0. We can reuse y_offset reg. +%define tempq y_offsetq + add x_offsetq, g_bilin_filterm +%define filter_x_a [x_offsetq] +%define filter_x_b [x_offsetq+16] + mov tempq, g_pw_8m +%define filter_rnd [tempq] +%else + add x_offsetq, bilin_filter +%define filter_x_a [x_offsetq] +%define filter_x_b [x_offsetq+16] +%define filter_rnd [GLOBAL(pw_8)] +%endif +%endif + +.x_other_y_zero_loop: +%if %1 == 16 + movu m0, [srcq] + movu m1, [srcq+16] + movu m2, [srcq+2] + movu m3, [srcq+18] + mova m4, [dstq] + mova m5, [dstq+16] + pmullw m1, filter_x_a + pmullw m3, filter_x_b + paddw m1, filter_rnd + pmullw m0, filter_x_a + pmullw m2, filter_x_b + paddw m0, filter_rnd + paddw m1, m3 + paddw m0, m2 + psrlw m1, 4 + psrlw m0, 4 +%if %2 == 1 ; avg + pavgw m0, [secq] + pavgw m1, [secq+16] +%endif + SUM_SSE m0, m4, m1, m5, m6, m7 + + lea srcq, [srcq+src_strideq*2] + lea dstq, [dstq+dst_strideq*2] +%if %2 == 1 ; avg + add secq, sec_str +%endif +%else ; %1 < 16 + movu m0, [srcq] + movu m1, [srcq+src_strideq*2] + movu m2, [srcq+2] + movu m3, [srcq+src_strideq*2+2] + mova m4, [dstq] + mova m5, [dstq+dst_strideq*2] + pmullw m1, filter_x_a + pmullw m3, filter_x_b + paddw m1, filter_rnd + pmullw m0, filter_x_a + pmullw m2, filter_x_b + paddw m0, filter_rnd + paddw m1, m3 + paddw m0, m2 + psrlw m1, 4 + psrlw m0, 4 +%if %2 == 1 ; avg + pavgw m0, [secq] + add secq, sec_str + pavgw m1, [secq] +%endif + SUM_SSE m0, m4, m1, m5, m6, m7 + + lea srcq, [srcq+src_strideq*4] + lea dstq, [dstq+dst_strideq*4] +%if %2 == 1 ; avg + add secq, sec_str +%endif +%endif + dec block_height + jg .x_other_y_zero_loop +%undef filter_x_a +%undef filter_x_b +%undef filter_rnd + STORE_AND_RET + +.x_nonhalf_y_nonzero: + cmp y_offsetd, 8 + jne .x_nonhalf_y_nonhalf + + ; x_offset == bilin interpolation && y_offset == 0.5 +%if ARCH_X86_64 + lea bilin_filter, [GLOBAL(bilin_filter_m)] +%endif + shl x_offsetd, filter_idx_shift +%if ARCH_X86_64 && mmsize == 16 + mova m8, [bilin_filter+x_offsetq] + mova m9, [bilin_filter+x_offsetq+16] + mova m10, [GLOBAL(pw_8)] +%define filter_x_a m8 +%define filter_x_b m9 +%define filter_rnd m10 +%else ; x86-32 +%if ARCH_X86=1 && CONFIG_PIC=1 +; y_offset == 0.5. We can reuse y_offset reg. +%define tempq y_offsetq + add x_offsetq, g_bilin_filterm +%define filter_x_a [x_offsetq] +%define filter_x_b [x_offsetq+16] + mov tempq, g_pw_8m +%define filter_rnd [tempq] +%else + add x_offsetq, bilin_filter +%define filter_x_a [x_offsetq] +%define filter_x_b [x_offsetq+16] +%define filter_rnd [GLOBAL(pw_8)] +%endif +%endif + +%if %1 == 16 + movu m0, [srcq] + movu m1, [srcq+16] + movu m2, [srcq+2] + movu m3, [srcq+18] + pmullw m0, filter_x_a + pmullw m2, filter_x_b + paddw m0, filter_rnd + pmullw m1, filter_x_a + pmullw m3, filter_x_b + paddw m1, filter_rnd + paddw m0, m2 + paddw m1, m3 + psrlw m0, 4 + psrlw m1, 4 + lea srcq, [srcq+src_strideq*2] +.x_other_y_half_loop: + movu m2, [srcq] + movu m3, [srcq+16] + movu m4, [srcq+2] + movu m5, [srcq+18] + pmullw m2, filter_x_a + pmullw m4, filter_x_b + paddw m2, filter_rnd + pmullw m3, filter_x_a + pmullw m5, filter_x_b + paddw m3, filter_rnd + paddw m2, m4 + paddw m3, m5 + mova m4, [dstq] + mova m5, [dstq+16] + psrlw m2, 4 + psrlw m3, 4 + pavgw m0, m2 + pavgw m1, m3 +%if %2 == 1 ; avg + pavgw m0, [secq] + pavgw m1, [secq+16] +%endif + SUM_SSE m0, m4, m1, m5, m6, m7 + mova m0, m2 + mova m1, m3 + + lea srcq, [srcq+src_strideq*2] + lea dstq, [dstq+dst_strideq*2] +%if %2 == 1 ; avg + add secq, sec_str +%endif +%else ; %1 < 16 + movu m0, [srcq] + movu m2, [srcq+2] + pmullw m0, filter_x_a + pmullw m2, filter_x_b + paddw m0, filter_rnd + paddw m0, m2 + psrlw m0, 4 + lea srcq, [srcq+src_strideq*2] +.x_other_y_half_loop: + movu m2, [srcq] + movu m3, [srcq+src_strideq*2] + movu m4, [srcq+2] + movu m5, [srcq+src_strideq*2+2] + pmullw m2, filter_x_a + pmullw m4, filter_x_b + paddw m2, filter_rnd + pmullw m3, filter_x_a + pmullw m5, filter_x_b + paddw m3, filter_rnd + paddw m2, m4 + paddw m3, m5 + mova m4, [dstq] + mova m5, [dstq+dst_strideq*2] + psrlw m2, 4 + psrlw m3, 4 + pavgw m0, m2 + pavgw m2, m3 +%if %2 == 1 ; avg + pavgw m0, [secq] + add secq, sec_str + pavgw m2, [secq] +%endif + SUM_SSE m0, m4, m2, m5, m6, m7 + mova m0, m3 + + lea srcq, [srcq+src_strideq*4] + lea dstq, [dstq+dst_strideq*4] +%if %2 == 1 ; avg + add secq, sec_str +%endif +%endif + dec block_height + jg .x_other_y_half_loop +%undef filter_x_a +%undef filter_x_b +%undef filter_rnd + STORE_AND_RET + +.x_nonhalf_y_nonhalf: +; loading filter - this is same as in 8-bit depth +%if ARCH_X86_64 + lea bilin_filter, [GLOBAL(bilin_filter_m)] +%endif + shl x_offsetd, filter_idx_shift ; filter_idx_shift = 5 + shl y_offsetd, filter_idx_shift +%if ARCH_X86_64 && mmsize == 16 + mova m8, [bilin_filter+x_offsetq] + mova m9, [bilin_filter+x_offsetq+16] + mova m10, [bilin_filter+y_offsetq] + mova m11, [bilin_filter+y_offsetq+16] + mova m12, [GLOBAL(pw_8)] +%define filter_x_a m8 +%define filter_x_b m9 +%define filter_y_a m10 +%define filter_y_b m11 +%define filter_rnd m12 +%else ; x86-32 +%if ARCH_X86=1 && CONFIG_PIC=1 +; In this case, there is NO unused register. Used src_stride register. Later, +; src_stride has to be loaded from stack when it is needed. +%define tempq src_strideq + mov tempq, g_bilin_filterm + add x_offsetq, tempq + add y_offsetq, tempq +%define filter_x_a [x_offsetq] +%define filter_x_b [x_offsetq+16] +%define filter_y_a [y_offsetq] +%define filter_y_b [y_offsetq+16] + + mov tempq, g_pw_8m +%define filter_rnd [tempq] +%else + add x_offsetq, bilin_filter + add y_offsetq, bilin_filter +%define filter_x_a [x_offsetq] +%define filter_x_b [x_offsetq+16] +%define filter_y_a [y_offsetq] +%define filter_y_b [y_offsetq+16] +%define filter_rnd [GLOBAL(pw_8)] +%endif +%endif +; end of load filter + + ; x_offset == bilin interpolation && y_offset == bilin interpolation +%if %1 == 16 + movu m0, [srcq] + movu m2, [srcq+2] + movu m1, [srcq+16] + movu m3, [srcq+18] + pmullw m0, filter_x_a + pmullw m2, filter_x_b + paddw m0, filter_rnd + pmullw m1, filter_x_a + pmullw m3, filter_x_b + paddw m1, filter_rnd + paddw m0, m2 + paddw m1, m3 + psrlw m0, 4 + psrlw m1, 4 + + INC_SRC_BY_SRC_STRIDE + +.x_other_y_other_loop: + movu m2, [srcq] + movu m4, [srcq+2] + movu m3, [srcq+16] + movu m5, [srcq+18] + pmullw m2, filter_x_a + pmullw m4, filter_x_b + paddw m2, filter_rnd + pmullw m3, filter_x_a + pmullw m5, filter_x_b + paddw m3, filter_rnd + paddw m2, m4 + paddw m3, m5 + psrlw m2, 4 + psrlw m3, 4 + mova m4, m2 + mova m5, m3 + pmullw m0, filter_y_a + pmullw m2, filter_y_b + paddw m0, filter_rnd + pmullw m1, filter_y_a + pmullw m3, filter_y_b + paddw m0, m2 + paddw m1, filter_rnd + mova m2, [dstq] + paddw m1, m3 + psrlw m0, 4 + psrlw m1, 4 + mova m3, [dstq+16] +%if %2 == 1 ; avg + pavgw m0, [secq] + pavgw m1, [secq+16] +%endif + SUM_SSE m0, m2, m1, m3, m6, m7 + mova m0, m4 + mova m1, m5 + + INC_SRC_BY_SRC_STRIDE + lea dstq, [dstq + dst_strideq * 2] +%if %2 == 1 ; avg + add secq, sec_str +%endif +%else ; %1 < 16 + movu m0, [srcq] + movu m2, [srcq+2] + pmullw m0, filter_x_a + pmullw m2, filter_x_b + paddw m0, filter_rnd + paddw m0, m2 + psrlw m0, 4 + + INC_SRC_BY_SRC_STRIDE + +.x_other_y_other_loop: + movu m2, [srcq] + movu m4, [srcq+2] + INC_SRC_BY_SRC_STRIDE + movu m3, [srcq] + movu m5, [srcq+2] + pmullw m2, filter_x_a + pmullw m4, filter_x_b + paddw m2, filter_rnd + pmullw m3, filter_x_a + pmullw m5, filter_x_b + paddw m3, filter_rnd + paddw m2, m4 + paddw m3, m5 + psrlw m2, 4 + psrlw m3, 4 + mova m4, m2 + mova m5, m3 + pmullw m0, filter_y_a + pmullw m2, filter_y_b + paddw m0, filter_rnd + pmullw m4, filter_y_a + pmullw m3, filter_y_b + paddw m0, m2 + paddw m4, filter_rnd + mova m2, [dstq] + paddw m4, m3 + psrlw m0, 4 + psrlw m4, 4 + mova m3, [dstq+dst_strideq*2] +%if %2 == 1 ; avg + pavgw m0, [secq] + add secq, sec_str + pavgw m4, [secq] +%endif + SUM_SSE m0, m2, m4, m3, m6, m7 + mova m0, m5 + + INC_SRC_BY_SRC_STRIDE + lea dstq, [dstq + dst_strideq * 4] +%if %2 == 1 ; avg + add secq, sec_str +%endif +%endif + dec block_height + jg .x_other_y_other_loop +%undef filter_x_a +%undef filter_x_b +%undef filter_y_a +%undef filter_y_b +%undef filter_rnd + STORE_AND_RET +%endmacro + +INIT_XMM sse2 +SUBPEL_VARIANCE 8 +SUBPEL_VARIANCE 16 + +INIT_XMM sse2 +SUBPEL_VARIANCE 8, 1 +SUBPEL_VARIANCE 16, 1 diff --git a/third_party/aom/aom_dsp/x86/highbd_subtract_sse2.c b/third_party/aom/aom_dsp/x86/highbd_subtract_sse2.c new file mode 100644 index 0000000000..18eb03d12c --- /dev/null +++ b/third_party/aom/aom_dsp/x86/highbd_subtract_sse2.c @@ -0,0 +1,267 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include +#include +#include + +#include "config/aom_config.h" +#include "config/aom_dsp_rtcd.h" + +typedef void (*SubtractWxHFuncType)(int16_t *diff, ptrdiff_t diff_stride, + const uint16_t *src, ptrdiff_t src_stride, + const uint16_t *pred, + ptrdiff_t pred_stride); + +static void subtract_4x4(int16_t *diff, ptrdiff_t diff_stride, + const uint16_t *src, ptrdiff_t src_stride, + const uint16_t *pred, ptrdiff_t pred_stride) { + __m128i u0, u1, u2, u3; + __m128i v0, v1, v2, v3; + __m128i x0, x1, x2, x3; + int64_t *store_diff = (int64_t *)(diff + 0 * diff_stride); + + u0 = _mm_loadu_si128((__m128i const *)(src + 0 * src_stride)); + u1 = _mm_loadu_si128((__m128i const *)(src + 1 * src_stride)); + u2 = _mm_loadu_si128((__m128i const *)(src + 2 * src_stride)); + u3 = _mm_loadu_si128((__m128i const *)(src + 3 * src_stride)); + + v0 = _mm_loadu_si128((__m128i const *)(pred + 0 * pred_stride)); + v1 = _mm_loadu_si128((__m128i const *)(pred + 1 * pred_stride)); + v2 = _mm_loadu_si128((__m128i const *)(pred + 2 * pred_stride)); + v3 = _mm_loadu_si128((__m128i const *)(pred + 3 * pred_stride)); + + x0 = _mm_sub_epi16(u0, v0); + x1 = _mm_sub_epi16(u1, v1); + x2 = _mm_sub_epi16(u2, v2); + x3 = _mm_sub_epi16(u3, v3); + + _mm_storel_epi64((__m128i *)store_diff, x0); + store_diff = (int64_t *)(diff + 1 * diff_stride); + _mm_storel_epi64((__m128i *)store_diff, x1); + store_diff = (int64_t *)(diff + 2 * diff_stride); + _mm_storel_epi64((__m128i *)store_diff, x2); + store_diff = (int64_t *)(diff + 3 * diff_stride); + _mm_storel_epi64((__m128i *)store_diff, x3); +} + +static void subtract_4x8(int16_t *diff, ptrdiff_t diff_stride, + const uint16_t *src, ptrdiff_t src_stride, + const uint16_t *pred, ptrdiff_t pred_stride) { + __m128i u0, u1, u2, u3, u4, u5, u6, u7; + __m128i v0, v1, v2, v3, v4, v5, v6, v7; + __m128i x0, x1, x2, x3, x4, x5, x6, x7; + int64_t *store_diff = (int64_t *)(diff + 0 * diff_stride); + + u0 = _mm_loadu_si128((__m128i const *)(src + 0 * src_stride)); + u1 = _mm_loadu_si128((__m128i const *)(src + 1 * src_stride)); + u2 = _mm_loadu_si128((__m128i const *)(src + 2 * src_stride)); + u3 = _mm_loadu_si128((__m128i const *)(src + 3 * src_stride)); + u4 = _mm_loadu_si128((__m128i const *)(src + 4 * src_stride)); + u5 = _mm_loadu_si128((__m128i const *)(src + 5 * src_stride)); + u6 = _mm_loadu_si128((__m128i const *)(src + 6 * src_stride)); + u7 = _mm_loadu_si128((__m128i const *)(src + 7 * src_stride)); + + v0 = _mm_loadu_si128((__m128i const *)(pred + 0 * pred_stride)); + v1 = _mm_loadu_si128((__m128i const *)(pred + 1 * pred_stride)); + v2 = _mm_loadu_si128((__m128i const *)(pred + 2 * pred_stride)); + v3 = _mm_loadu_si128((__m128i const *)(pred + 3 * pred_stride)); + v4 = _mm_loadu_si128((__m128i const *)(pred + 4 * pred_stride)); + v5 = _mm_loadu_si128((__m128i const *)(pred + 5 * pred_stride)); + v6 = _mm_loadu_si128((__m128i const *)(pred + 6 * pred_stride)); + v7 = _mm_loadu_si128((__m128i const *)(pred + 7 * pred_stride)); + + x0 = _mm_sub_epi16(u0, v0); + x1 = _mm_sub_epi16(u1, v1); + x2 = _mm_sub_epi16(u2, v2); + x3 = _mm_sub_epi16(u3, v3); + x4 = _mm_sub_epi16(u4, v4); + x5 = _mm_sub_epi16(u5, v5); + x6 = _mm_sub_epi16(u6, v6); + x7 = _mm_sub_epi16(u7, v7); + + _mm_storel_epi64((__m128i *)store_diff, x0); + store_diff = (int64_t *)(diff + 1 * diff_stride); + _mm_storel_epi64((__m128i *)store_diff, x1); + store_diff = (int64_t *)(diff + 2 * diff_stride); + _mm_storel_epi64((__m128i *)store_diff, x2); + store_diff = (int64_t *)(diff + 3 * diff_stride); + _mm_storel_epi64((__m128i *)store_diff, x3); + store_diff = (int64_t *)(diff + 4 * diff_stride); + _mm_storel_epi64((__m128i *)store_diff, x4); + store_diff = (int64_t *)(diff + 5 * diff_stride); + _mm_storel_epi64((__m128i *)store_diff, x5); + store_diff = (int64_t *)(diff + 6 * diff_stride); + _mm_storel_epi64((__m128i *)store_diff, x6); + store_diff = (int64_t *)(diff + 7 * diff_stride); + _mm_storel_epi64((__m128i *)store_diff, x7); +} + +static void subtract_8x4(int16_t *diff, ptrdiff_t diff_stride, + const uint16_t *src, ptrdiff_t src_stride, + const uint16_t *pred, ptrdiff_t pred_stride) { + __m128i u0, u1, u2, u3; + __m128i v0, v1, v2, v3; + __m128i x0, x1, x2, x3; + + u0 = _mm_loadu_si128((__m128i const *)(src + 0 * src_stride)); + u1 = _mm_loadu_si128((__m128i const *)(src + 1 * src_stride)); + u2 = _mm_loadu_si128((__m128i const *)(src + 2 * src_stride)); + u3 = _mm_loadu_si128((__m128i const *)(src + 3 * src_stride)); + + v0 = _mm_loadu_si128((__m128i const *)(pred + 0 * pred_stride)); + v1 = _mm_loadu_si128((__m128i const *)(pred + 1 * pred_stride)); + v2 = _mm_loadu_si128((__m128i const *)(pred + 2 * pred_stride)); + v3 = _mm_loadu_si128((__m128i const *)(pred + 3 * pred_stride)); + + x0 = _mm_sub_epi16(u0, v0); + x1 = _mm_sub_epi16(u1, v1); + x2 = _mm_sub_epi16(u2, v2); + x3 = _mm_sub_epi16(u3, v3); + + _mm_storeu_si128((__m128i *)(diff + 0 * diff_stride), x0); + _mm_storeu_si128((__m128i *)(diff + 1 * diff_stride), x1); + _mm_storeu_si128((__m128i *)(diff + 2 * diff_stride), x2); + _mm_storeu_si128((__m128i *)(diff + 3 * diff_stride), x3); +} + +static void subtract_8x8(int16_t *diff, ptrdiff_t diff_stride, + const uint16_t *src, ptrdiff_t src_stride, + const uint16_t *pred, ptrdiff_t pred_stride) { + __m128i u0, u1, u2, u3, u4, u5, u6, u7; + __m128i v0, v1, v2, v3, v4, v5, v6, v7; + __m128i x0, x1, x2, x3, x4, x5, x6, x7; + + u0 = _mm_loadu_si128((__m128i const *)(src + 0 * src_stride)); + u1 = _mm_loadu_si128((__m128i const *)(src + 1 * src_stride)); + u2 = _mm_loadu_si128((__m128i const *)(src + 2 * src_stride)); + u3 = _mm_loadu_si128((__m128i const *)(src + 3 * src_stride)); + u4 = _mm_loadu_si128((__m128i const *)(src + 4 * src_stride)); + u5 = _mm_loadu_si128((__m128i const *)(src + 5 * src_stride)); + u6 = _mm_loadu_si128((__m128i const *)(src + 6 * src_stride)); + u7 = _mm_loadu_si128((__m128i const *)(src + 7 * src_stride)); + + v0 = _mm_loadu_si128((__m128i const *)(pred + 0 * pred_stride)); + v1 = _mm_loadu_si128((__m128i const *)(pred + 1 * pred_stride)); + v2 = _mm_loadu_si128((__m128i const *)(pred + 2 * pred_stride)); + v3 = _mm_loadu_si128((__m128i const *)(pred + 3 * pred_stride)); + v4 = _mm_loadu_si128((__m128i const *)(pred + 4 * pred_stride)); + v5 = _mm_loadu_si128((__m128i const *)(pred + 5 * pred_stride)); + v6 = _mm_loadu_si128((__m128i const *)(pred + 6 * pred_stride)); + v7 = _mm_loadu_si128((__m128i const *)(pred + 7 * pred_stride)); + + x0 = _mm_sub_epi16(u0, v0); + x1 = _mm_sub_epi16(u1, v1); + x2 = _mm_sub_epi16(u2, v2); + x3 = _mm_sub_epi16(u3, v3); + x4 = _mm_sub_epi16(u4, v4); + x5 = _mm_sub_epi16(u5, v5); + x6 = _mm_sub_epi16(u6, v6); + x7 = _mm_sub_epi16(u7, v7); + + _mm_storeu_si128((__m128i *)(diff + 0 * diff_stride), x0); + _mm_storeu_si128((__m128i *)(diff + 1 * diff_stride), x1); + _mm_storeu_si128((__m128i *)(diff + 2 * diff_stride), x2); + _mm_storeu_si128((__m128i *)(diff + 3 * diff_stride), x3); + _mm_storeu_si128((__m128i *)(diff + 4 * diff_stride), x4); + _mm_storeu_si128((__m128i *)(diff + 5 * diff_stride), x5); + _mm_storeu_si128((__m128i *)(diff + 6 * diff_stride), x6); + _mm_storeu_si128((__m128i *)(diff + 7 * diff_stride), x7); +} + +#define STACK_V(h, fun) \ + do { \ + fun(diff, diff_stride, src, src_stride, pred, pred_stride); \ + fun(diff + diff_stride * h, diff_stride, src + src_stride * h, src_stride, \ + pred + pred_stride * h, pred_stride); \ + } while (0) + +#define STACK_H(w, fun) \ + do { \ + fun(diff, diff_stride, src, src_stride, pred, pred_stride); \ + fun(diff + w, diff_stride, src + w, src_stride, pred + w, pred_stride); \ + } while (0) + +#define SUBTRACT_FUN(size) \ + static void subtract_##size(int16_t *diff, ptrdiff_t diff_stride, \ + const uint16_t *src, ptrdiff_t src_stride, \ + const uint16_t *pred, ptrdiff_t pred_stride) + +SUBTRACT_FUN(8x16) { STACK_V(8, subtract_8x8); } +SUBTRACT_FUN(16x8) { STACK_H(8, subtract_8x8); } +SUBTRACT_FUN(16x16) { STACK_V(8, subtract_16x8); } +SUBTRACT_FUN(16x32) { STACK_V(16, subtract_16x16); } +SUBTRACT_FUN(32x16) { STACK_H(16, subtract_16x16); } +SUBTRACT_FUN(32x32) { STACK_V(16, subtract_32x16); } +SUBTRACT_FUN(32x64) { STACK_V(32, subtract_32x32); } +SUBTRACT_FUN(64x32) { STACK_H(32, subtract_32x32); } +SUBTRACT_FUN(64x64) { STACK_V(32, subtract_64x32); } +SUBTRACT_FUN(64x128) { STACK_V(64, subtract_64x64); } +SUBTRACT_FUN(128x64) { STACK_H(64, subtract_64x64); } +SUBTRACT_FUN(128x128) { STACK_V(64, subtract_128x64); } +SUBTRACT_FUN(4x16) { STACK_V(8, subtract_4x8); } +SUBTRACT_FUN(16x4) { STACK_H(8, subtract_8x4); } +SUBTRACT_FUN(8x32) { STACK_V(16, subtract_8x16); } +SUBTRACT_FUN(32x8) { STACK_H(16, subtract_16x8); } +SUBTRACT_FUN(16x64) { STACK_V(32, subtract_16x32); } +SUBTRACT_FUN(64x16) { STACK_H(32, subtract_32x16); } + +static SubtractWxHFuncType getSubtractFunc(int rows, int cols) { + if (rows == 4) { + if (cols == 4) return subtract_4x4; + if (cols == 8) return subtract_8x4; + if (cols == 16) return subtract_16x4; + } + if (rows == 8) { + if (cols == 4) return subtract_4x8; + if (cols == 8) return subtract_8x8; + if (cols == 16) return subtract_16x8; + if (cols == 32) return subtract_32x8; + } + if (rows == 16) { + if (cols == 4) return subtract_4x16; + if (cols == 8) return subtract_8x16; + if (cols == 16) return subtract_16x16; + if (cols == 32) return subtract_32x16; + if (cols == 64) return subtract_64x16; + } + if (rows == 32) { + if (cols == 8) return subtract_8x32; + if (cols == 16) return subtract_16x32; + if (cols == 32) return subtract_32x32; + if (cols == 64) return subtract_64x32; + } + if (rows == 64) { + if (cols == 16) return subtract_16x64; + if (cols == 32) return subtract_32x64; + if (cols == 64) return subtract_64x64; + if (cols == 128) return subtract_128x64; + } + if (rows == 128) { + if (cols == 64) return subtract_64x128; + if (cols == 128) return subtract_128x128; + } + assert(0); + return NULL; +} + +void aom_highbd_subtract_block_sse2(int rows, int cols, int16_t *diff, + ptrdiff_t diff_stride, const uint8_t *src8, + ptrdiff_t src_stride, const uint8_t *pred8, + ptrdiff_t pred_stride, int bd) { + uint16_t *src = CONVERT_TO_SHORTPTR(src8); + uint16_t *pred = CONVERT_TO_SHORTPTR(pred8); + SubtractWxHFuncType func; + (void)bd; + + func = getSubtractFunc(rows, cols); + func(diff, diff_stride, src, src_stride, pred, pred_stride); +} diff --git a/third_party/aom/aom_dsp/x86/highbd_variance_avx2.c b/third_party/aom/aom_dsp/x86/highbd_variance_avx2.c new file mode 100644 index 0000000000..9b1b4c9de7 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/highbd_variance_avx2.c @@ -0,0 +1,140 @@ +/* + * Copyright (c) 2018, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include +#include // AVX2 + +#include "config/aom_dsp_rtcd.h" + +typedef void (*high_variance_fn_t)(const uint16_t *src, int src_stride, + const uint16_t *ref, int ref_stride, + uint32_t *sse, int *sum); + +void aom_highbd_calc8x8var_avx2(const uint16_t *src, int src_stride, + const uint16_t *ref, int ref_stride, + uint32_t *sse, int *sum) { + __m256i v_sum_d = _mm256_setzero_si256(); + __m256i v_sse_d = _mm256_setzero_si256(); + for (int i = 0; i < 8; i += 2) { + const __m128i v_p_a0 = _mm_loadu_si128((const __m128i *)src); + const __m128i v_p_a1 = _mm_loadu_si128((const __m128i *)(src + src_stride)); + const __m128i v_p_b0 = _mm_loadu_si128((const __m128i *)ref); + const __m128i v_p_b1 = _mm_loadu_si128((const __m128i *)(ref + ref_stride)); + __m256i v_p_a = _mm256_castsi128_si256(v_p_a0); + __m256i v_p_b = _mm256_castsi128_si256(v_p_b0); + v_p_a = _mm256_inserti128_si256(v_p_a, v_p_a1, 1); + v_p_b = _mm256_inserti128_si256(v_p_b, v_p_b1, 1); + const __m256i v_diff = _mm256_sub_epi16(v_p_a, v_p_b); + const __m256i v_sqrdiff = _mm256_madd_epi16(v_diff, v_diff); + v_sum_d = _mm256_add_epi16(v_sum_d, v_diff); + v_sse_d = _mm256_add_epi32(v_sse_d, v_sqrdiff); + src += src_stride * 2; + ref += ref_stride * 2; + } + __m256i v_sum00 = _mm256_cvtepi16_epi32(_mm256_castsi256_si128(v_sum_d)); + __m256i v_sum01 = _mm256_cvtepi16_epi32(_mm256_extracti128_si256(v_sum_d, 1)); + __m256i v_sum0 = _mm256_add_epi32(v_sum00, v_sum01); + __m256i v_d_l = _mm256_unpacklo_epi32(v_sum0, v_sse_d); + __m256i v_d_h = _mm256_unpackhi_epi32(v_sum0, v_sse_d); + __m256i v_d_lh = _mm256_add_epi32(v_d_l, v_d_h); + const __m128i v_d0_d = _mm256_castsi256_si128(v_d_lh); + const __m128i v_d1_d = _mm256_extracti128_si256(v_d_lh, 1); + __m128i v_d = _mm_add_epi32(v_d0_d, v_d1_d); + v_d = _mm_add_epi32(v_d, _mm_srli_si128(v_d, 8)); + *sum = _mm_extract_epi32(v_d, 0); + *sse = _mm_extract_epi32(v_d, 1); +} + +void aom_highbd_calc16x16var_avx2(const uint16_t *src, int src_stride, + const uint16_t *ref, int ref_stride, + uint32_t *sse, int *sum) { + __m256i v_sum_d = _mm256_setzero_si256(); + __m256i v_sse_d = _mm256_setzero_si256(); + const __m256i one = _mm256_set1_epi16(1); + for (int i = 0; i < 16; ++i) { + const __m256i v_p_a = _mm256_loadu_si256((const __m256i *)src); + const __m256i v_p_b = _mm256_loadu_si256((const __m256i *)ref); + const __m256i v_diff = _mm256_sub_epi16(v_p_a, v_p_b); + const __m256i v_sqrdiff = _mm256_madd_epi16(v_diff, v_diff); + v_sum_d = _mm256_add_epi16(v_sum_d, v_diff); + v_sse_d = _mm256_add_epi32(v_sse_d, v_sqrdiff); + src += src_stride; + ref += ref_stride; + } + __m256i v_sum0 = _mm256_madd_epi16(v_sum_d, one); + __m256i v_d_l = _mm256_unpacklo_epi32(v_sum0, v_sse_d); + __m256i v_d_h = _mm256_unpackhi_epi32(v_sum0, v_sse_d); + __m256i v_d_lh = _mm256_add_epi32(v_d_l, v_d_h); + const __m128i v_d0_d = _mm256_castsi256_si128(v_d_lh); + const __m128i v_d1_d = _mm256_extracti128_si256(v_d_lh, 1); + __m128i v_d = _mm_add_epi32(v_d0_d, v_d1_d); + v_d = _mm_add_epi32(v_d, _mm_srli_si128(v_d, 8)); + *sum = _mm_extract_epi32(v_d, 0); + *sse = _mm_extract_epi32(v_d, 1); +} + +static void highbd_10_variance_avx2(const uint16_t *src, int src_stride, + const uint16_t *ref, int ref_stride, int w, + int h, uint32_t *sse, int *sum, + high_variance_fn_t var_fn, int block_size) { + int i, j; + uint64_t sse_long = 0; + int32_t sum_long = 0; + + for (i = 0; i < h; i += block_size) { + for (j = 0; j < w; j += block_size) { + unsigned int sse0; + int sum0; + var_fn(src + src_stride * i + j, src_stride, ref + ref_stride * i + j, + ref_stride, &sse0, &sum0); + sse_long += sse0; + sum_long += sum0; + } + } + *sum = ROUND_POWER_OF_TWO(sum_long, 2); + *sse = (uint32_t)ROUND_POWER_OF_TWO(sse_long, 4); +} + +#define VAR_FN(w, h, block_size, shift) \ + uint32_t aom_highbd_10_variance##w##x##h##_avx2( \ + const uint8_t *src8, int src_stride, const uint8_t *ref8, \ + int ref_stride, uint32_t *sse) { \ + int sum; \ + int64_t var; \ + uint16_t *src = CONVERT_TO_SHORTPTR(src8); \ + uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); \ + highbd_10_variance_avx2( \ + src, src_stride, ref, ref_stride, w, h, sse, &sum, \ + aom_highbd_calc##block_size##x##block_size##var_avx2, block_size); \ + var = (int64_t)(*sse) - (((int64_t)sum * sum) >> shift); \ + return (var >= 0) ? (uint32_t)var : 0; \ + } + +VAR_FN(128, 128, 16, 14); +VAR_FN(128, 64, 16, 13); +VAR_FN(64, 128, 16, 13); +VAR_FN(64, 64, 16, 12); +VAR_FN(64, 32, 16, 11); +VAR_FN(32, 64, 16, 11); +VAR_FN(32, 32, 16, 10); +VAR_FN(32, 16, 16, 9); +VAR_FN(16, 32, 16, 9); +VAR_FN(16, 16, 16, 8); +VAR_FN(16, 8, 8, 7); +VAR_FN(8, 16, 8, 7); +VAR_FN(8, 8, 8, 6); +VAR_FN(16, 4, 16, 6); +VAR_FN(8, 32, 8, 8); +VAR_FN(32, 8, 8, 8); +VAR_FN(16, 64, 16, 10); +VAR_FN(64, 16, 16, 10); + +#undef VAR_FN diff --git a/third_party/aom/aom_dsp/x86/highbd_variance_impl_sse2.asm b/third_party/aom/aom_dsp/x86/highbd_variance_impl_sse2.asm new file mode 100644 index 0000000000..0d954e1788 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/highbd_variance_impl_sse2.asm @@ -0,0 +1,318 @@ +; +; Copyright (c) 2016, Alliance for Open Media. All rights reserved +; +; This source code is subject to the terms of the BSD 2 Clause License and +; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License +; was not distributed with this source code in the LICENSE file, you can +; obtain it at www.aomedia.org/license/software. If the Alliance for Open +; Media Patent License 1.0 was not distributed with this source code in the +; PATENTS file, you can obtain it at www.aomedia.org/license/patent. +; + +; + + +%include "aom_ports/x86_abi_support.asm" + +SECTION .text + +;unsigned int aom_highbd_calc16x16var_sse2 +;( +; unsigned char * src_ptr, +; int source_stride, +; unsigned char * ref_ptr, +; int recon_stride, +; unsigned int * SSE, +; int * Sum +;) +global sym(aom_highbd_calc16x16var_sse2) PRIVATE +sym(aom_highbd_calc16x16var_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 6 + SAVE_XMM 7 + push rbx + push rsi + push rdi + ; end prolog + + mov rsi, arg(0) ;[src_ptr] + mov rdi, arg(2) ;[ref_ptr] + + movsxd rax, DWORD PTR arg(1) ;[source_stride] + movsxd rdx, DWORD PTR arg(3) ;[recon_stride] + add rax, rax ; source stride in bytes + add rdx, rdx ; recon stride in bytes + + ; Prefetch data + prefetcht0 [rsi] + prefetcht0 [rsi+16] + prefetcht0 [rsi+rax] + prefetcht0 [rsi+rax+16] + lea rbx, [rsi+rax*2] + prefetcht0 [rbx] + prefetcht0 [rbx+16] + prefetcht0 [rbx+rax] + prefetcht0 [rbx+rax+16] + + prefetcht0 [rdi] + prefetcht0 [rdi+16] + prefetcht0 [rdi+rdx] + prefetcht0 [rdi+rdx+16] + lea rbx, [rdi+rdx*2] + prefetcht0 [rbx] + prefetcht0 [rbx+16] + prefetcht0 [rbx+rdx] + prefetcht0 [rbx+rdx+16] + + pxor xmm0, xmm0 ; clear xmm0 for unpack + pxor xmm7, xmm7 ; clear xmm7 for accumulating diffs + + pxor xmm6, xmm6 ; clear xmm6 for accumulating sse + mov rcx, 16 + +.var16loop: + movdqu xmm1, XMMWORD PTR [rsi] + movdqu xmm2, XMMWORD PTR [rdi] + + lea rbx, [rsi+rax*2] + prefetcht0 [rbx] + prefetcht0 [rbx+16] + prefetcht0 [rbx+rax] + prefetcht0 [rbx+rax+16] + lea rbx, [rdi+rdx*2] + prefetcht0 [rbx] + prefetcht0 [rbx+16] + prefetcht0 [rbx+rdx] + prefetcht0 [rbx+rdx+16] + + pxor xmm5, xmm5 + + psubw xmm1, xmm2 + movdqu xmm3, XMMWORD PTR [rsi+16] + paddw xmm5, xmm1 + pmaddwd xmm1, xmm1 + movdqu xmm2, XMMWORD PTR [rdi+16] + paddd xmm6, xmm1 + + psubw xmm3, xmm2 + movdqu xmm1, XMMWORD PTR [rsi+rax] + paddw xmm5, xmm3 + pmaddwd xmm3, xmm3 + movdqu xmm2, XMMWORD PTR [rdi+rdx] + paddd xmm6, xmm3 + + psubw xmm1, xmm2 + movdqu xmm3, XMMWORD PTR [rsi+rax+16] + paddw xmm5, xmm1 + pmaddwd xmm1, xmm1 + movdqu xmm2, XMMWORD PTR [rdi+rdx+16] + paddd xmm6, xmm1 + + psubw xmm3, xmm2 + paddw xmm5, xmm3 + pmaddwd xmm3, xmm3 + paddd xmm6, xmm3 + + movdqa xmm1, xmm5 + movdqa xmm2, xmm5 + pcmpgtw xmm1, xmm0 + pcmpeqw xmm2, xmm0 + por xmm1, xmm2 + pcmpeqw xmm1, xmm0 + movdqa xmm2, xmm5 + punpcklwd xmm5, xmm1 + punpckhwd xmm2, xmm1 + paddd xmm7, xmm5 + paddd xmm7, xmm2 + + lea rsi, [rsi + 2*rax] + lea rdi, [rdi + 2*rdx] + sub rcx, 2 + jnz .var16loop + + movdqa xmm4, xmm6 + punpckldq xmm6, xmm0 + + punpckhdq xmm4, xmm0 + movdqa xmm5, xmm7 + + paddd xmm6, xmm4 + punpckldq xmm7, xmm0 + + punpckhdq xmm5, xmm0 + paddd xmm7, xmm5 + + movdqa xmm4, xmm6 + movdqa xmm5, xmm7 + + psrldq xmm4, 8 + psrldq xmm5, 8 + + paddd xmm6, xmm4 + paddd xmm7, xmm5 + + mov rdi, arg(4) ; [SSE] + mov rax, arg(5) ; [Sum] + + movd DWORD PTR [rdi], xmm6 + movd DWORD PTR [rax], xmm7 + + + ; begin epilog + pop rdi + pop rsi + pop rbx + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + + +;unsigned int aom_highbd_calc8x8var_sse2 +;( +; unsigned char * src_ptr, +; int source_stride, +; unsigned char * ref_ptr, +; int recon_stride, +; unsigned int * SSE, +; int * Sum +;) +global sym(aom_highbd_calc8x8var_sse2) PRIVATE +sym(aom_highbd_calc8x8var_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 6 + SAVE_XMM 7 + push rbx + push rsi + push rdi + ; end prolog + + mov rsi, arg(0) ;[src_ptr] + mov rdi, arg(2) ;[ref_ptr] + + movsxd rax, DWORD PTR arg(1) ;[source_stride] + movsxd rdx, DWORD PTR arg(3) ;[recon_stride] + add rax, rax ; source stride in bytes + add rdx, rdx ; recon stride in bytes + + ; Prefetch data + prefetcht0 [rsi] + prefetcht0 [rsi+rax] + lea rbx, [rsi+rax*2] + prefetcht0 [rbx] + prefetcht0 [rbx+rax] + + prefetcht0 [rdi] + prefetcht0 [rdi+rdx] + lea rbx, [rdi+rdx*2] + prefetcht0 [rbx] + prefetcht0 [rbx+rdx] + + pxor xmm0, xmm0 ; clear xmm0 for unpack + pxor xmm7, xmm7 ; clear xmm7 for accumulating diffs + + pxor xmm6, xmm6 ; clear xmm6 for accumulating sse + mov rcx, 8 + +.var8loop: + movdqu xmm1, XMMWORD PTR [rsi] + movdqu xmm2, XMMWORD PTR [rdi] + + lea rbx, [rsi+rax*4] + prefetcht0 [rbx] + prefetcht0 [rbx+rax] + lea rbx, [rbx+rax*2] + prefetcht0 [rbx] + prefetcht0 [rbx+rax] + lea rbx, [rdi+rdx*4] + prefetcht0 [rbx] + prefetcht0 [rbx+rdx] + lea rbx, [rbx+rdx*2] + prefetcht0 [rbx] + prefetcht0 [rbx+rdx] + + pxor xmm5, xmm5 + + psubw xmm1, xmm2 + movdqu xmm3, XMMWORD PTR [rsi+rax] + paddw xmm5, xmm1 + pmaddwd xmm1, xmm1 + movdqu xmm2, XMMWORD PTR [rdi+rdx] + paddd xmm6, xmm1 + + lea rsi, [rsi + 2*rax] + lea rdi, [rdi + 2*rdx] + + psubw xmm3, xmm2 + movdqu xmm1, XMMWORD PTR [rsi] + paddw xmm5, xmm3 + pmaddwd xmm3, xmm3 + movdqu xmm2, XMMWORD PTR [rdi] + paddd xmm6, xmm3 + + psubw xmm1, xmm2 + movdqu xmm3, XMMWORD PTR [rsi+rax] + paddw xmm5, xmm1 + pmaddwd xmm1, xmm1 + movdqu xmm2, XMMWORD PTR [rdi+rdx] + paddd xmm6, xmm1 + + psubw xmm3, xmm2 + paddw xmm5, xmm3 + pmaddwd xmm3, xmm3 + paddd xmm6, xmm3 + + movdqa xmm1, xmm5 + movdqa xmm2, xmm5 + pcmpgtw xmm1, xmm0 + pcmpeqw xmm2, xmm0 + por xmm1, xmm2 + pcmpeqw xmm1, xmm0 + movdqa xmm2, xmm5 + punpcklwd xmm5, xmm1 + punpckhwd xmm2, xmm1 + paddd xmm7, xmm5 + paddd xmm7, xmm2 + + lea rsi, [rsi + 2*rax] + lea rdi, [rdi + 2*rdx] + sub rcx, 4 + jnz .var8loop + + movdqa xmm4, xmm6 + punpckldq xmm6, xmm0 + + punpckhdq xmm4, xmm0 + movdqa xmm5, xmm7 + + paddd xmm6, xmm4 + punpckldq xmm7, xmm0 + + punpckhdq xmm5, xmm0 + paddd xmm7, xmm5 + + movdqa xmm4, xmm6 + movdqa xmm5, xmm7 + + psrldq xmm4, 8 + psrldq xmm5, 8 + + paddd xmm6, xmm4 + paddd xmm7, xmm5 + + mov rdi, arg(4) ; [SSE] + mov rax, arg(5) ; [Sum] + + movd DWORD PTR [rdi], xmm6 + movd DWORD PTR [rax], xmm7 + + ; begin epilog + pop rdi + pop rsi + pop rbx + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret diff --git a/third_party/aom/aom_dsp/x86/highbd_variance_sse2.c b/third_party/aom/aom_dsp/x86/highbd_variance_sse2.c new file mode 100644 index 0000000000..47b052abc9 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/highbd_variance_sse2.c @@ -0,0 +1,868 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include +#include // SSE2 + +#include "config/aom_config.h" +#include "config/aom_dsp_rtcd.h" +#include "config/av1_rtcd.h" + +#include "aom_dsp/x86/synonyms.h" + +#include "aom_ports/mem.h" + +#include "av1/common/filter.h" +#include "av1/common/onyxc_int.h" +#include "av1/common/reconinter.h" + +typedef uint32_t (*high_variance_fn_t)(const uint16_t *src, int src_stride, + const uint16_t *ref, int ref_stride, + uint32_t *sse, int *sum); + +uint32_t aom_highbd_calc8x8var_sse2(const uint16_t *src, int src_stride, + const uint16_t *ref, int ref_stride, + uint32_t *sse, int *sum); + +uint32_t aom_highbd_calc16x16var_sse2(const uint16_t *src, int src_stride, + const uint16_t *ref, int ref_stride, + uint32_t *sse, int *sum); + +static void highbd_8_variance_sse2(const uint16_t *src, int src_stride, + const uint16_t *ref, int ref_stride, int w, + int h, uint32_t *sse, int *sum, + high_variance_fn_t var_fn, int block_size) { + int i, j; + + *sse = 0; + *sum = 0; + + for (i = 0; i < h; i += block_size) { + for (j = 0; j < w; j += block_size) { + unsigned int sse0; + int sum0; + var_fn(src + src_stride * i + j, src_stride, ref + ref_stride * i + j, + ref_stride, &sse0, &sum0); + *sse += sse0; + *sum += sum0; + } + } +} + +static void highbd_10_variance_sse2(const uint16_t *src, int src_stride, + const uint16_t *ref, int ref_stride, int w, + int h, uint32_t *sse, int *sum, + high_variance_fn_t var_fn, int block_size) { + int i, j; + uint64_t sse_long = 0; + int32_t sum_long = 0; + + for (i = 0; i < h; i += block_size) { + for (j = 0; j < w; j += block_size) { + unsigned int sse0; + int sum0; + var_fn(src + src_stride * i + j, src_stride, ref + ref_stride * i + j, + ref_stride, &sse0, &sum0); + sse_long += sse0; + sum_long += sum0; + } + } + *sum = ROUND_POWER_OF_TWO(sum_long, 2); + *sse = (uint32_t)ROUND_POWER_OF_TWO(sse_long, 4); +} + +static void highbd_12_variance_sse2(const uint16_t *src, int src_stride, + const uint16_t *ref, int ref_stride, int w, + int h, uint32_t *sse, int *sum, + high_variance_fn_t var_fn, int block_size) { + int i, j; + uint64_t sse_long = 0; + int32_t sum_long = 0; + + for (i = 0; i < h; i += block_size) { + for (j = 0; j < w; j += block_size) { + unsigned int sse0; + int sum0; + var_fn(src + src_stride * i + j, src_stride, ref + ref_stride * i + j, + ref_stride, &sse0, &sum0); + sse_long += sse0; + sum_long += sum0; + } + } + *sum = ROUND_POWER_OF_TWO(sum_long, 4); + *sse = (uint32_t)ROUND_POWER_OF_TWO(sse_long, 8); +} + +#define HIGH_GET_VAR(S) \ + void aom_highbd_get##S##x##S##var_sse2(const uint8_t *src8, int src_stride, \ + const uint8_t *ref8, int ref_stride, \ + uint32_t *sse, int *sum) { \ + uint16_t *src = CONVERT_TO_SHORTPTR(src8); \ + uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); \ + aom_highbd_calc##S##x##S##var_sse2(src, src_stride, ref, ref_stride, sse, \ + sum); \ + } \ + \ + void aom_highbd_10_get##S##x##S##var_sse2( \ + const uint8_t *src8, int src_stride, const uint8_t *ref8, \ + int ref_stride, uint32_t *sse, int *sum) { \ + uint16_t *src = CONVERT_TO_SHORTPTR(src8); \ + uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); \ + aom_highbd_calc##S##x##S##var_sse2(src, src_stride, ref, ref_stride, sse, \ + sum); \ + *sum = ROUND_POWER_OF_TWO(*sum, 2); \ + *sse = ROUND_POWER_OF_TWO(*sse, 4); \ + } \ + \ + void aom_highbd_12_get##S##x##S##var_sse2( \ + const uint8_t *src8, int src_stride, const uint8_t *ref8, \ + int ref_stride, uint32_t *sse, int *sum) { \ + uint16_t *src = CONVERT_TO_SHORTPTR(src8); \ + uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); \ + aom_highbd_calc##S##x##S##var_sse2(src, src_stride, ref, ref_stride, sse, \ + sum); \ + *sum = ROUND_POWER_OF_TWO(*sum, 4); \ + *sse = ROUND_POWER_OF_TWO(*sse, 8); \ + } + +HIGH_GET_VAR(16); +HIGH_GET_VAR(8); + +#undef HIGH_GET_VAR + +#define VAR_FN(w, h, block_size, shift) \ + uint32_t aom_highbd_8_variance##w##x##h##_sse2( \ + const uint8_t *src8, int src_stride, const uint8_t *ref8, \ + int ref_stride, uint32_t *sse) { \ + int sum; \ + uint16_t *src = CONVERT_TO_SHORTPTR(src8); \ + uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); \ + highbd_8_variance_sse2( \ + src, src_stride, ref, ref_stride, w, h, sse, &sum, \ + aom_highbd_calc##block_size##x##block_size##var_sse2, block_size); \ + return *sse - (uint32_t)(((int64_t)sum * sum) >> shift); \ + } \ + \ + uint32_t aom_highbd_10_variance##w##x##h##_sse2( \ + const uint8_t *src8, int src_stride, const uint8_t *ref8, \ + int ref_stride, uint32_t *sse) { \ + int sum; \ + int64_t var; \ + uint16_t *src = CONVERT_TO_SHORTPTR(src8); \ + uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); \ + highbd_10_variance_sse2( \ + src, src_stride, ref, ref_stride, w, h, sse, &sum, \ + aom_highbd_calc##block_size##x##block_size##var_sse2, block_size); \ + var = (int64_t)(*sse) - (((int64_t)sum * sum) >> shift); \ + return (var >= 0) ? (uint32_t)var : 0; \ + } \ + \ + uint32_t aom_highbd_12_variance##w##x##h##_sse2( \ + const uint8_t *src8, int src_stride, const uint8_t *ref8, \ + int ref_stride, uint32_t *sse) { \ + int sum; \ + int64_t var; \ + uint16_t *src = CONVERT_TO_SHORTPTR(src8); \ + uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); \ + highbd_12_variance_sse2( \ + src, src_stride, ref, ref_stride, w, h, sse, &sum, \ + aom_highbd_calc##block_size##x##block_size##var_sse2, block_size); \ + var = (int64_t)(*sse) - (((int64_t)sum * sum) >> shift); \ + return (var >= 0) ? (uint32_t)var : 0; \ + } + +VAR_FN(128, 128, 16, 14); +VAR_FN(128, 64, 16, 13); +VAR_FN(64, 128, 16, 13); +VAR_FN(64, 64, 16, 12); +VAR_FN(64, 32, 16, 11); +VAR_FN(32, 64, 16, 11); +VAR_FN(32, 32, 16, 10); +VAR_FN(32, 16, 16, 9); +VAR_FN(16, 32, 16, 9); +VAR_FN(16, 16, 16, 8); +VAR_FN(16, 8, 8, 7); +VAR_FN(8, 16, 8, 7); +VAR_FN(8, 8, 8, 6); +VAR_FN(16, 4, 16, 6); +VAR_FN(8, 32, 8, 8); +VAR_FN(32, 8, 8, 8); +VAR_FN(16, 64, 16, 10); +VAR_FN(64, 16, 16, 10); + +#undef VAR_FN + +unsigned int aom_highbd_8_mse16x16_sse2(const uint8_t *src8, int src_stride, + const uint8_t *ref8, int ref_stride, + unsigned int *sse) { + int sum; + uint16_t *src = CONVERT_TO_SHORTPTR(src8); + uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); + highbd_8_variance_sse2(src, src_stride, ref, ref_stride, 16, 16, sse, &sum, + aom_highbd_calc16x16var_sse2, 16); + return *sse; +} + +unsigned int aom_highbd_10_mse16x16_sse2(const uint8_t *src8, int src_stride, + const uint8_t *ref8, int ref_stride, + unsigned int *sse) { + int sum; + uint16_t *src = CONVERT_TO_SHORTPTR(src8); + uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); + highbd_10_variance_sse2(src, src_stride, ref, ref_stride, 16, 16, sse, &sum, + aom_highbd_calc16x16var_sse2, 16); + return *sse; +} + +unsigned int aom_highbd_12_mse16x16_sse2(const uint8_t *src8, int src_stride, + const uint8_t *ref8, int ref_stride, + unsigned int *sse) { + int sum; + uint16_t *src = CONVERT_TO_SHORTPTR(src8); + uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); + highbd_12_variance_sse2(src, src_stride, ref, ref_stride, 16, 16, sse, &sum, + aom_highbd_calc16x16var_sse2, 16); + return *sse; +} + +unsigned int aom_highbd_8_mse8x8_sse2(const uint8_t *src8, int src_stride, + const uint8_t *ref8, int ref_stride, + unsigned int *sse) { + int sum; + uint16_t *src = CONVERT_TO_SHORTPTR(src8); + uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); + highbd_8_variance_sse2(src, src_stride, ref, ref_stride, 8, 8, sse, &sum, + aom_highbd_calc8x8var_sse2, 8); + return *sse; +} + +unsigned int aom_highbd_10_mse8x8_sse2(const uint8_t *src8, int src_stride, + const uint8_t *ref8, int ref_stride, + unsigned int *sse) { + int sum; + uint16_t *src = CONVERT_TO_SHORTPTR(src8); + uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); + highbd_10_variance_sse2(src, src_stride, ref, ref_stride, 8, 8, sse, &sum, + aom_highbd_calc8x8var_sse2, 8); + return *sse; +} + +unsigned int aom_highbd_12_mse8x8_sse2(const uint8_t *src8, int src_stride, + const uint8_t *ref8, int ref_stride, + unsigned int *sse) { + int sum; + uint16_t *src = CONVERT_TO_SHORTPTR(src8); + uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); + highbd_12_variance_sse2(src, src_stride, ref, ref_stride, 8, 8, sse, &sum, + aom_highbd_calc8x8var_sse2, 8); + return *sse; +} + +// The 2 unused parameters are place holders for PIC enabled build. +// These definitions are for functions defined in +// highbd_subpel_variance_impl_sse2.asm +#define DECL(w, opt) \ + int aom_highbd_sub_pixel_variance##w##xh_##opt( \ + const uint16_t *src, ptrdiff_t src_stride, int x_offset, int y_offset, \ + const uint16_t *dst, ptrdiff_t dst_stride, int height, \ + unsigned int *sse, void *unused0, void *unused); +#define DECLS(opt) \ + DECL(8, opt); \ + DECL(16, opt) + +DECLS(sse2); + +#undef DECLS +#undef DECL + +#define FN(w, h, wf, wlog2, hlog2, opt, cast) \ + uint32_t aom_highbd_8_sub_pixel_variance##w##x##h##_##opt( \ + const uint8_t *src8, int src_stride, int x_offset, int y_offset, \ + const uint8_t *dst8, int dst_stride, uint32_t *sse_ptr) { \ + uint32_t sse; \ + uint16_t *src = CONVERT_TO_SHORTPTR(src8); \ + uint16_t *dst = CONVERT_TO_SHORTPTR(dst8); \ + int se = aom_highbd_sub_pixel_variance##wf##xh_##opt( \ + src, src_stride, x_offset, y_offset, dst, dst_stride, h, &sse, NULL, \ + NULL); \ + if (w > wf) { \ + unsigned int sse2; \ + int se2 = aom_highbd_sub_pixel_variance##wf##xh_##opt( \ + src + 16, src_stride, x_offset, y_offset, dst + 16, dst_stride, h, \ + &sse2, NULL, NULL); \ + se += se2; \ + sse += sse2; \ + if (w > wf * 2) { \ + se2 = aom_highbd_sub_pixel_variance##wf##xh_##opt( \ + src + 32, src_stride, x_offset, y_offset, dst + 32, dst_stride, h, \ + &sse2, NULL, NULL); \ + se += se2; \ + sse += sse2; \ + se2 = aom_highbd_sub_pixel_variance##wf##xh_##opt( \ + src + 48, src_stride, x_offset, y_offset, dst + 48, dst_stride, h, \ + &sse2, NULL, NULL); \ + se += se2; \ + sse += sse2; \ + } \ + } \ + *sse_ptr = sse; \ + return sse - (uint32_t)((cast se * se) >> (wlog2 + hlog2)); \ + } \ + \ + uint32_t aom_highbd_10_sub_pixel_variance##w##x##h##_##opt( \ + const uint8_t *src8, int src_stride, int x_offset, int y_offset, \ + const uint8_t *dst8, int dst_stride, uint32_t *sse_ptr) { \ + int64_t var; \ + uint32_t sse; \ + uint16_t *src = CONVERT_TO_SHORTPTR(src8); \ + uint16_t *dst = CONVERT_TO_SHORTPTR(dst8); \ + int se = aom_highbd_sub_pixel_variance##wf##xh_##opt( \ + src, src_stride, x_offset, y_offset, dst, dst_stride, h, &sse, NULL, \ + NULL); \ + if (w > wf) { \ + uint32_t sse2; \ + int se2 = aom_highbd_sub_pixel_variance##wf##xh_##opt( \ + src + 16, src_stride, x_offset, y_offset, dst + 16, dst_stride, h, \ + &sse2, NULL, NULL); \ + se += se2; \ + sse += sse2; \ + if (w > wf * 2) { \ + se2 = aom_highbd_sub_pixel_variance##wf##xh_##opt( \ + src + 32, src_stride, x_offset, y_offset, dst + 32, dst_stride, h, \ + &sse2, NULL, NULL); \ + se += se2; \ + sse += sse2; \ + se2 = aom_highbd_sub_pixel_variance##wf##xh_##opt( \ + src + 48, src_stride, x_offset, y_offset, dst + 48, dst_stride, h, \ + &sse2, NULL, NULL); \ + se += se2; \ + sse += sse2; \ + } \ + } \ + se = ROUND_POWER_OF_TWO(se, 2); \ + sse = ROUND_POWER_OF_TWO(sse, 4); \ + *sse_ptr = sse; \ + var = (int64_t)(sse) - ((cast se * se) >> (wlog2 + hlog2)); \ + return (var >= 0) ? (uint32_t)var : 0; \ + } \ + \ + uint32_t aom_highbd_12_sub_pixel_variance##w##x##h##_##opt( \ + const uint8_t *src8, int src_stride, int x_offset, int y_offset, \ + const uint8_t *dst8, int dst_stride, uint32_t *sse_ptr) { \ + int start_row; \ + uint32_t sse; \ + int se = 0; \ + int64_t var; \ + uint64_t long_sse = 0; \ + uint16_t *src = CONVERT_TO_SHORTPTR(src8); \ + uint16_t *dst = CONVERT_TO_SHORTPTR(dst8); \ + for (start_row = 0; start_row < h; start_row += 16) { \ + uint32_t sse2; \ + int height = h - start_row < 16 ? h - start_row : 16; \ + int se2 = aom_highbd_sub_pixel_variance##wf##xh_##opt( \ + src + (start_row * src_stride), src_stride, x_offset, y_offset, \ + dst + (start_row * dst_stride), dst_stride, height, &sse2, NULL, \ + NULL); \ + se += se2; \ + long_sse += sse2; \ + if (w > wf) { \ + se2 = aom_highbd_sub_pixel_variance##wf##xh_##opt( \ + src + 16 + (start_row * src_stride), src_stride, x_offset, \ + y_offset, dst + 16 + (start_row * dst_stride), dst_stride, height, \ + &sse2, NULL, NULL); \ + se += se2; \ + long_sse += sse2; \ + if (w > wf * 2) { \ + se2 = aom_highbd_sub_pixel_variance##wf##xh_##opt( \ + src + 32 + (start_row * src_stride), src_stride, x_offset, \ + y_offset, dst + 32 + (start_row * dst_stride), dst_stride, \ + height, &sse2, NULL, NULL); \ + se += se2; \ + long_sse += sse2; \ + se2 = aom_highbd_sub_pixel_variance##wf##xh_##opt( \ + src + 48 + (start_row * src_stride), src_stride, x_offset, \ + y_offset, dst + 48 + (start_row * dst_stride), dst_stride, \ + height, &sse2, NULL, NULL); \ + se += se2; \ + long_sse += sse2; \ + } \ + } \ + } \ + se = ROUND_POWER_OF_TWO(se, 4); \ + sse = (uint32_t)ROUND_POWER_OF_TWO(long_sse, 8); \ + *sse_ptr = sse; \ + var = (int64_t)(sse) - ((cast se * se) >> (wlog2 + hlog2)); \ + return (var >= 0) ? (uint32_t)var : 0; \ + } + +#define FNS(opt) \ + FN(64, 64, 16, 6, 6, opt, (int64_t)); \ + FN(64, 32, 16, 6, 5, opt, (int64_t)); \ + FN(32, 64, 16, 5, 6, opt, (int64_t)); \ + FN(32, 32, 16, 5, 5, opt, (int64_t)); \ + FN(32, 16, 16, 5, 4, opt, (int64_t)); \ + FN(16, 32, 16, 4, 5, opt, (int64_t)); \ + FN(16, 16, 16, 4, 4, opt, (int64_t)); \ + FN(16, 8, 16, 4, 3, opt, (int64_t)); \ + FN(8, 16, 8, 3, 4, opt, (int64_t)); \ + FN(8, 8, 8, 3, 3, opt, (int64_t)); \ + FN(8, 4, 8, 3, 2, opt, (int64_t)); \ + FN(16, 4, 16, 4, 2, opt, (int64_t)); \ + FN(8, 32, 8, 3, 5, opt, (int64_t)); \ + FN(32, 8, 16, 5, 3, opt, (int64_t)); \ + FN(16, 64, 16, 4, 6, opt, (int64_t)); \ + FN(64, 16, 16, 6, 4, opt, (int64_t)) + +FNS(sse2); + +#undef FNS +#undef FN + +// The 2 unused parameters are place holders for PIC enabled build. +#define DECL(w, opt) \ + int aom_highbd_sub_pixel_avg_variance##w##xh_##opt( \ + const uint16_t *src, ptrdiff_t src_stride, int x_offset, int y_offset, \ + const uint16_t *dst, ptrdiff_t dst_stride, const uint16_t *sec, \ + ptrdiff_t sec_stride, int height, unsigned int *sse, void *unused0, \ + void *unused); +#define DECLS(opt) \ + DECL(16, opt) \ + DECL(8, opt) + +DECLS(sse2); +#undef DECL +#undef DECLS + +#define FN(w, h, wf, wlog2, hlog2, opt, cast) \ + uint32_t aom_highbd_8_sub_pixel_avg_variance##w##x##h##_##opt( \ + const uint8_t *src8, int src_stride, int x_offset, int y_offset, \ + const uint8_t *dst8, int dst_stride, uint32_t *sse_ptr, \ + const uint8_t *sec8) { \ + uint32_t sse; \ + uint16_t *src = CONVERT_TO_SHORTPTR(src8); \ + uint16_t *dst = CONVERT_TO_SHORTPTR(dst8); \ + uint16_t *sec = CONVERT_TO_SHORTPTR(sec8); \ + int se = aom_highbd_sub_pixel_avg_variance##wf##xh_##opt( \ + src, src_stride, x_offset, y_offset, dst, dst_stride, sec, w, h, &sse, \ + NULL, NULL); \ + if (w > wf) { \ + uint32_t sse2; \ + int se2 = aom_highbd_sub_pixel_avg_variance##wf##xh_##opt( \ + src + 16, src_stride, x_offset, y_offset, dst + 16, dst_stride, \ + sec + 16, w, h, &sse2, NULL, NULL); \ + se += se2; \ + sse += sse2; \ + if (w > wf * 2) { \ + se2 = aom_highbd_sub_pixel_avg_variance##wf##xh_##opt( \ + src + 32, src_stride, x_offset, y_offset, dst + 32, dst_stride, \ + sec + 32, w, h, &sse2, NULL, NULL); \ + se += se2; \ + sse += sse2; \ + se2 = aom_highbd_sub_pixel_avg_variance##wf##xh_##opt( \ + src + 48, src_stride, x_offset, y_offset, dst + 48, dst_stride, \ + sec + 48, w, h, &sse2, NULL, NULL); \ + se += se2; \ + sse += sse2; \ + } \ + } \ + *sse_ptr = sse; \ + return sse - (uint32_t)((cast se * se) >> (wlog2 + hlog2)); \ + } \ + \ + uint32_t aom_highbd_10_sub_pixel_avg_variance##w##x##h##_##opt( \ + const uint8_t *src8, int src_stride, int x_offset, int y_offset, \ + const uint8_t *dst8, int dst_stride, uint32_t *sse_ptr, \ + const uint8_t *sec8) { \ + int64_t var; \ + uint32_t sse; \ + uint16_t *src = CONVERT_TO_SHORTPTR(src8); \ + uint16_t *dst = CONVERT_TO_SHORTPTR(dst8); \ + uint16_t *sec = CONVERT_TO_SHORTPTR(sec8); \ + int se = aom_highbd_sub_pixel_avg_variance##wf##xh_##opt( \ + src, src_stride, x_offset, y_offset, dst, dst_stride, sec, w, h, &sse, \ + NULL, NULL); \ + if (w > wf) { \ + uint32_t sse2; \ + int se2 = aom_highbd_sub_pixel_avg_variance##wf##xh_##opt( \ + src + 16, src_stride, x_offset, y_offset, dst + 16, dst_stride, \ + sec + 16, w, h, &sse2, NULL, NULL); \ + se += se2; \ + sse += sse2; \ + if (w > wf * 2) { \ + se2 = aom_highbd_sub_pixel_avg_variance##wf##xh_##opt( \ + src + 32, src_stride, x_offset, y_offset, dst + 32, dst_stride, \ + sec + 32, w, h, &sse2, NULL, NULL); \ + se += se2; \ + sse += sse2; \ + se2 = aom_highbd_sub_pixel_avg_variance##wf##xh_##opt( \ + src + 48, src_stride, x_offset, y_offset, dst + 48, dst_stride, \ + sec + 48, w, h, &sse2, NULL, NULL); \ + se += se2; \ + sse += sse2; \ + } \ + } \ + se = ROUND_POWER_OF_TWO(se, 2); \ + sse = ROUND_POWER_OF_TWO(sse, 4); \ + *sse_ptr = sse; \ + var = (int64_t)(sse) - ((cast se * se) >> (wlog2 + hlog2)); \ + return (var >= 0) ? (uint32_t)var : 0; \ + } \ + \ + uint32_t aom_highbd_12_sub_pixel_avg_variance##w##x##h##_##opt( \ + const uint8_t *src8, int src_stride, int x_offset, int y_offset, \ + const uint8_t *dst8, int dst_stride, uint32_t *sse_ptr, \ + const uint8_t *sec8) { \ + int start_row; \ + int64_t var; \ + uint32_t sse; \ + int se = 0; \ + uint64_t long_sse = 0; \ + uint16_t *src = CONVERT_TO_SHORTPTR(src8); \ + uint16_t *dst = CONVERT_TO_SHORTPTR(dst8); \ + uint16_t *sec = CONVERT_TO_SHORTPTR(sec8); \ + for (start_row = 0; start_row < h; start_row += 16) { \ + uint32_t sse2; \ + int height = h - start_row < 16 ? h - start_row : 16; \ + int se2 = aom_highbd_sub_pixel_avg_variance##wf##xh_##opt( \ + src + (start_row * src_stride), src_stride, x_offset, y_offset, \ + dst + (start_row * dst_stride), dst_stride, sec + (start_row * w), \ + w, height, &sse2, NULL, NULL); \ + se += se2; \ + long_sse += sse2; \ + if (w > wf) { \ + se2 = aom_highbd_sub_pixel_avg_variance##wf##xh_##opt( \ + src + 16 + (start_row * src_stride), src_stride, x_offset, \ + y_offset, dst + 16 + (start_row * dst_stride), dst_stride, \ + sec + 16 + (start_row * w), w, height, &sse2, NULL, NULL); \ + se += se2; \ + long_sse += sse2; \ + if (w > wf * 2) { \ + se2 = aom_highbd_sub_pixel_avg_variance##wf##xh_##opt( \ + src + 32 + (start_row * src_stride), src_stride, x_offset, \ + y_offset, dst + 32 + (start_row * dst_stride), dst_stride, \ + sec + 32 + (start_row * w), w, height, &sse2, NULL, NULL); \ + se += se2; \ + long_sse += sse2; \ + se2 = aom_highbd_sub_pixel_avg_variance##wf##xh_##opt( \ + src + 48 + (start_row * src_stride), src_stride, x_offset, \ + y_offset, dst + 48 + (start_row * dst_stride), dst_stride, \ + sec + 48 + (start_row * w), w, height, &sse2, NULL, NULL); \ + se += se2; \ + long_sse += sse2; \ + } \ + } \ + } \ + se = ROUND_POWER_OF_TWO(se, 4); \ + sse = (uint32_t)ROUND_POWER_OF_TWO(long_sse, 8); \ + *sse_ptr = sse; \ + var = (int64_t)(sse) - ((cast se * se) >> (wlog2 + hlog2)); \ + return (var >= 0) ? (uint32_t)var : 0; \ + } + +#define FNS(opt) \ + FN(64, 64, 16, 6, 6, opt, (int64_t)); \ + FN(64, 32, 16, 6, 5, opt, (int64_t)); \ + FN(32, 64, 16, 5, 6, opt, (int64_t)); \ + FN(32, 32, 16, 5, 5, opt, (int64_t)); \ + FN(32, 16, 16, 5, 4, opt, (int64_t)); \ + FN(16, 32, 16, 4, 5, opt, (int64_t)); \ + FN(16, 16, 16, 4, 4, opt, (int64_t)); \ + FN(16, 8, 16, 4, 3, opt, (int64_t)); \ + FN(8, 16, 8, 3, 4, opt, (int64_t)); \ + FN(8, 8, 8, 3, 3, opt, (int64_t)); \ + FN(8, 4, 8, 3, 2, opt, (int64_t)); \ + FN(16, 4, 16, 4, 2, opt, (int64_t)); \ + FN(8, 32, 8, 3, 5, opt, (int64_t)); \ + FN(32, 8, 16, 5, 3, opt, (int64_t)); \ + FN(16, 64, 16, 4, 6, opt, (int64_t)); \ + FN(64, 16, 16, 6, 4, opt, (int64_t)); + +FNS(sse2); + +#undef FNS +#undef FN + +void aom_highbd_upsampled_pred_sse2(MACROBLOCKD *xd, + const struct AV1Common *const cm, + int mi_row, int mi_col, const MV *const mv, + uint8_t *comp_pred8, int width, int height, + int subpel_x_q3, int subpel_y_q3, + const uint8_t *ref8, int ref_stride, int bd, + int subpel_search) { + // expect xd == NULL only in tests + if (xd != NULL) { + const MB_MODE_INFO *mi = xd->mi[0]; + const int ref_num = 0; + const int is_intrabc = is_intrabc_block(mi); + const struct scale_factors *const sf = + is_intrabc ? &cm->sf_identity : &xd->block_refs[ref_num]->sf; + const int is_scaled = av1_is_scaled(sf); + + if (is_scaled) { + // Note: This is mostly a copy from the >=8X8 case in + // build_inter_predictors() function, with some small tweaks. + // Some assumptions. + const int plane = 0; + + // Get pre-requisites. + const struct macroblockd_plane *const pd = &xd->plane[plane]; + const int ssx = pd->subsampling_x; + const int ssy = pd->subsampling_y; + assert(ssx == 0 && ssy == 0); + const struct buf_2d *const dst_buf = &pd->dst; + const struct buf_2d *const pre_buf = + is_intrabc ? dst_buf : &pd->pre[ref_num]; + const int mi_x = mi_col * MI_SIZE; + const int mi_y = mi_row * MI_SIZE; + + // Calculate subpel_x/y and x/y_step. + const int row_start = 0; // Because ss_y is 0. + const int col_start = 0; // Because ss_x is 0. + const int pre_x = (mi_x + MI_SIZE * col_start) >> ssx; + const int pre_y = (mi_y + MI_SIZE * row_start) >> ssy; + int orig_pos_y = pre_y << SUBPEL_BITS; + orig_pos_y += mv->row * (1 << (1 - ssy)); + int orig_pos_x = pre_x << SUBPEL_BITS; + orig_pos_x += mv->col * (1 << (1 - ssx)); + int pos_y = sf->scale_value_y(orig_pos_y, sf); + int pos_x = sf->scale_value_x(orig_pos_x, sf); + pos_x += SCALE_EXTRA_OFF; + pos_y += SCALE_EXTRA_OFF; + + const int top = -AOM_LEFT_TOP_MARGIN_SCALED(ssy); + const int left = -AOM_LEFT_TOP_MARGIN_SCALED(ssx); + const int bottom = (pre_buf->height + AOM_INTERP_EXTEND) + << SCALE_SUBPEL_BITS; + const int right = (pre_buf->width + AOM_INTERP_EXTEND) + << SCALE_SUBPEL_BITS; + pos_y = clamp(pos_y, top, bottom); + pos_x = clamp(pos_x, left, right); + + const uint8_t *const pre = + pre_buf->buf0 + (pos_y >> SCALE_SUBPEL_BITS) * pre_buf->stride + + (pos_x >> SCALE_SUBPEL_BITS); + const SubpelParams subpel_params = { sf->x_step_q4, sf->y_step_q4, + pos_x & SCALE_SUBPEL_MASK, + pos_y & SCALE_SUBPEL_MASK }; + + // Get warp types. + const WarpedMotionParams *const wm = + &xd->global_motion[mi->ref_frame[ref_num]]; + const int is_global = is_global_mv_block(mi, wm->wmtype); + WarpTypesAllowed warp_types; + warp_types.global_warp_allowed = is_global; + warp_types.local_warp_allowed = mi->motion_mode == WARPED_CAUSAL; + + // Get convolve parameters. + ConvolveParams conv_params = get_conv_params(0, plane, xd->bd); + const InterpFilters filters = + av1_broadcast_interp_filter(EIGHTTAP_REGULAR); + + // Get the inter predictor. + const int build_for_obmc = 0; + av1_make_inter_predictor(pre, pre_buf->stride, comp_pred8, width, + &subpel_params, sf, width, height, &conv_params, + filters, &warp_types, mi_x >> pd->subsampling_x, + mi_y >> pd->subsampling_y, plane, ref_num, mi, + build_for_obmc, xd, cm->allow_warped_motion); + return; + } + } + + const InterpFilterParams *filter = + (subpel_search == 1) + ? av1_get_4tap_interp_filter_params(EIGHTTAP_REGULAR) + : av1_get_interp_filter_params_with_block_size(EIGHTTAP_REGULAR, 8); + + if (!subpel_x_q3 && !subpel_y_q3) { + uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); + uint16_t *comp_pred = CONVERT_TO_SHORTPTR(comp_pred8); + if (width >= 8) { + int i; + assert(!(width & 7)); + /*Read 8 pixels one row at a time.*/ + for (i = 0; i < height; i++) { + int j; + for (j = 0; j < width; j += 8) { + __m128i s0 = _mm_loadu_si128((const __m128i *)ref); + _mm_storeu_si128((__m128i *)comp_pred, s0); + comp_pred += 8; + ref += 8; + } + ref += ref_stride - width; + } + } else { + int i; + assert(!(width & 3)); + /*Read 4 pixels two rows at a time.*/ + for (i = 0; i < height; i += 2) { + __m128i s0 = _mm_loadl_epi64((const __m128i *)ref); + __m128i s1 = _mm_loadl_epi64((const __m128i *)(ref + ref_stride)); + __m128i t0 = _mm_unpacklo_epi64(s0, s1); + _mm_storeu_si128((__m128i *)comp_pred, t0); + comp_pred += 8; + ref += 2 * ref_stride; + } + } + } else if (!subpel_y_q3) { + const int16_t *const kernel = + av1_get_interp_filter_subpel_kernel(filter, subpel_x_q3 << 1); + aom_highbd_convolve8_horiz(ref8, ref_stride, comp_pred8, width, kernel, 16, + NULL, -1, width, height, bd); + } else if (!subpel_x_q3) { + const int16_t *const kernel = + av1_get_interp_filter_subpel_kernel(filter, subpel_y_q3 << 1); + aom_highbd_convolve8_vert(ref8, ref_stride, comp_pred8, width, NULL, -1, + kernel, 16, width, height, bd); + } else { + DECLARE_ALIGNED(16, uint16_t, + temp[((MAX_SB_SIZE + 16) + 16) * MAX_SB_SIZE]); + const int16_t *const kernel_x = + av1_get_interp_filter_subpel_kernel(filter, subpel_x_q3 << 1); + const int16_t *const kernel_y = + av1_get_interp_filter_subpel_kernel(filter, subpel_y_q3 << 1); + const int intermediate_height = + (((height - 1) * 8 + subpel_y_q3) >> 3) + filter->taps; + assert(intermediate_height <= (MAX_SB_SIZE * 2 + 16) + 16); + aom_highbd_convolve8_horiz(ref8 - ref_stride * ((filter->taps >> 1) - 1), + ref_stride, CONVERT_TO_BYTEPTR(temp), + MAX_SB_SIZE, kernel_x, 16, NULL, -1, width, + intermediate_height, bd); + aom_highbd_convolve8_vert( + CONVERT_TO_BYTEPTR(temp + MAX_SB_SIZE * ((filter->taps >> 1) - 1)), + MAX_SB_SIZE, comp_pred8, width, NULL, -1, kernel_y, 16, width, height, + bd); + } +} + +void aom_highbd_comp_avg_upsampled_pred_sse2( + MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col, + const MV *const mv, uint8_t *comp_pred8, const uint8_t *pred8, int width, + int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref8, + int ref_stride, int bd, int subpel_search) { + aom_highbd_upsampled_pred(xd, cm, mi_row, mi_col, mv, comp_pred8, width, + height, subpel_x_q3, subpel_y_q3, ref8, ref_stride, + bd, subpel_search); + uint16_t *pred = CONVERT_TO_SHORTPTR(pred8); + uint16_t *comp_pred16 = CONVERT_TO_SHORTPTR(comp_pred8); + /*The total number of pixels must be a multiple of 8 (e.g., 4x4).*/ + assert(!(width * height & 7)); + int n = width * height >> 3; + for (int i = 0; i < n; i++) { + __m128i s0 = _mm_loadu_si128((const __m128i *)comp_pred16); + __m128i p0 = _mm_loadu_si128((const __m128i *)pred); + _mm_storeu_si128((__m128i *)comp_pred16, _mm_avg_epu16(s0, p0)); + comp_pred16 += 8; + pred += 8; + } +} + +static INLINE void highbd_compute_jnt_comp_avg(__m128i *p0, __m128i *p1, + const __m128i *w0, + const __m128i *w1, + const __m128i *r, + void *const result) { + assert(DIST_PRECISION_BITS <= 4); + __m128i mult0 = _mm_mullo_epi16(*p0, *w0); + __m128i mult1 = _mm_mullo_epi16(*p1, *w1); + __m128i sum = _mm_adds_epu16(mult0, mult1); + __m128i round = _mm_adds_epu16(sum, *r); + __m128i shift = _mm_srli_epi16(round, DIST_PRECISION_BITS); + + xx_storeu_128(result, shift); +} + +void aom_highbd_jnt_comp_avg_pred_sse2(uint8_t *comp_pred8, + const uint8_t *pred8, int width, + int height, const uint8_t *ref8, + int ref_stride, + const JNT_COMP_PARAMS *jcp_param) { + int i; + const uint16_t wt0 = (uint16_t)jcp_param->fwd_offset; + const uint16_t wt1 = (uint16_t)jcp_param->bck_offset; + const __m128i w0 = _mm_set_epi16(wt0, wt0, wt0, wt0, wt0, wt0, wt0, wt0); + const __m128i w1 = _mm_set_epi16(wt1, wt1, wt1, wt1, wt1, wt1, wt1, wt1); + const uint16_t round = ((1 << DIST_PRECISION_BITS) >> 1); + const __m128i r = + _mm_set_epi16(round, round, round, round, round, round, round, round); + uint16_t *pred = CONVERT_TO_SHORTPTR(pred8); + uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); + uint16_t *comp_pred = CONVERT_TO_SHORTPTR(comp_pred8); + + if (width >= 8) { + // Read 8 pixels one row at a time + assert(!(width & 7)); + for (i = 0; i < height; ++i) { + int j; + for (j = 0; j < width; j += 8) { + __m128i p0 = xx_loadu_128(ref); + __m128i p1 = xx_loadu_128(pred); + + highbd_compute_jnt_comp_avg(&p0, &p1, &w0, &w1, &r, comp_pred); + + comp_pred += 8; + pred += 8; + ref += 8; + } + ref += ref_stride - width; + } + } else { + // Read 4 pixels two rows at a time + assert(!(width & 3)); + for (i = 0; i < height; i += 2) { + __m128i p0_0 = xx_loadl_64(ref + 0 * ref_stride); + __m128i p0_1 = xx_loadl_64(ref + 1 * ref_stride); + __m128i p0 = _mm_unpacklo_epi64(p0_0, p0_1); + __m128i p1 = xx_loadu_128(pred); + + highbd_compute_jnt_comp_avg(&p0, &p1, &w0, &w1, &r, comp_pred); + + comp_pred += 8; + pred += 8; + ref += 2 * ref_stride; + } + } +} + +void aom_highbd_jnt_comp_avg_upsampled_pred_sse2( + MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col, + const MV *const mv, uint8_t *comp_pred8, const uint8_t *pred8, int width, + int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref8, + int ref_stride, int bd, const JNT_COMP_PARAMS *jcp_param, + int subpel_search) { + uint16_t *pred = CONVERT_TO_SHORTPTR(pred8); + int n; + int i; + aom_highbd_upsampled_pred(xd, cm, mi_row, mi_col, mv, comp_pred8, width, + height, subpel_x_q3, subpel_y_q3, ref8, ref_stride, + bd, subpel_search); + assert(!(width * height & 7)); + n = width * height >> 3; + + const uint16_t wt0 = (uint16_t)jcp_param->fwd_offset; + const uint16_t wt1 = (uint16_t)jcp_param->bck_offset; + const __m128i w0 = _mm_set_epi16(wt0, wt0, wt0, wt0, wt0, wt0, wt0, wt0); + const __m128i w1 = _mm_set_epi16(wt1, wt1, wt1, wt1, wt1, wt1, wt1, wt1); + const uint16_t round = ((1 << DIST_PRECISION_BITS) >> 1); + const __m128i r = + _mm_set_epi16(round, round, round, round, round, round, round, round); + + uint16_t *comp_pred16 = CONVERT_TO_SHORTPTR(comp_pred8); + for (i = 0; i < n; i++) { + __m128i p0 = xx_loadu_128(comp_pred16); + __m128i p1 = xx_loadu_128(pred); + + highbd_compute_jnt_comp_avg(&p0, &p1, &w0, &w1, &r, comp_pred16); + + comp_pred16 += 8; + pred += 8; + } +} diff --git a/third_party/aom/aom_dsp/x86/highbd_variance_sse4.c b/third_party/aom/aom_dsp/x86/highbd_variance_sse4.c new file mode 100644 index 0000000000..df5449a9df --- /dev/null +++ b/third_party/aom/aom_dsp/x86/highbd_variance_sse4.c @@ -0,0 +1,216 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include /* SSE4.1 */ + +#include "config/aom_config.h" +#include "config/aom_dsp_rtcd.h" + +#include "aom_dsp/variance.h" +#include "aom_dsp/aom_filter.h" + +static INLINE void variance4x4_64_sse4_1(const uint8_t *a8, int a_stride, + const uint8_t *b8, int b_stride, + uint64_t *sse, int64_t *sum) { + __m128i u0, u1, u2, u3; + __m128i s0, s1, s2, s3; + __m128i t0, t1, x0, y0; + __m128i a0, a1, a2, a3; + __m128i b0, b1, b2, b3; + __m128i k_one_epi16 = _mm_set1_epi16((int16_t)1); + + uint16_t *a = CONVERT_TO_SHORTPTR(a8); + uint16_t *b = CONVERT_TO_SHORTPTR(b8); + + a0 = _mm_loadl_epi64((__m128i const *)(a + 0 * a_stride)); + a1 = _mm_loadl_epi64((__m128i const *)(a + 1 * a_stride)); + a2 = _mm_loadl_epi64((__m128i const *)(a + 2 * a_stride)); + a3 = _mm_loadl_epi64((__m128i const *)(a + 3 * a_stride)); + + b0 = _mm_loadl_epi64((__m128i const *)(b + 0 * b_stride)); + b1 = _mm_loadl_epi64((__m128i const *)(b + 1 * b_stride)); + b2 = _mm_loadl_epi64((__m128i const *)(b + 2 * b_stride)); + b3 = _mm_loadl_epi64((__m128i const *)(b + 3 * b_stride)); + + u0 = _mm_unpacklo_epi16(a0, a1); + u1 = _mm_unpacklo_epi16(a2, a3); + u2 = _mm_unpacklo_epi16(b0, b1); + u3 = _mm_unpacklo_epi16(b2, b3); + + s0 = _mm_sub_epi16(u0, u2); + s1 = _mm_sub_epi16(u1, u3); + + t0 = _mm_madd_epi16(s0, k_one_epi16); + t1 = _mm_madd_epi16(s1, k_one_epi16); + + s2 = _mm_hadd_epi32(t0, t1); + s3 = _mm_hadd_epi32(s2, s2); + y0 = _mm_hadd_epi32(s3, s3); + + t0 = _mm_madd_epi16(s0, s0); + t1 = _mm_madd_epi16(s1, s1); + + s2 = _mm_hadd_epi32(t0, t1); + s3 = _mm_hadd_epi32(s2, s2); + x0 = _mm_hadd_epi32(s3, s3); + + *sse = (uint64_t)_mm_extract_epi32(x0, 0); + *sum = (int64_t)_mm_extract_epi32(y0, 0); +} + +uint32_t aom_highbd_8_variance4x4_sse4_1(const uint8_t *a, int a_stride, + const uint8_t *b, int b_stride, + uint32_t *sse) { + int64_t sum, diff; + uint64_t local_sse; + + variance4x4_64_sse4_1(a, a_stride, b, b_stride, &local_sse, &sum); + *sse = (uint32_t)local_sse; + + diff = (int64_t)*sse - ((sum * sum) >> 4); + return (diff >= 0) ? (uint32_t)diff : 0; +} + +uint32_t aom_highbd_10_variance4x4_sse4_1(const uint8_t *a, int a_stride, + const uint8_t *b, int b_stride, + uint32_t *sse) { + int64_t sum, diff; + uint64_t local_sse; + + variance4x4_64_sse4_1(a, a_stride, b, b_stride, &local_sse, &sum); + *sse = (uint32_t)ROUND_POWER_OF_TWO(local_sse, 4); + sum = ROUND_POWER_OF_TWO(sum, 2); + + diff = (int64_t)*sse - ((sum * sum) >> 4); + return (diff >= 0) ? (uint32_t)diff : 0; +} + +uint32_t aom_highbd_12_variance4x4_sse4_1(const uint8_t *a, int a_stride, + const uint8_t *b, int b_stride, + uint32_t *sse) { + int64_t sum, diff; + uint64_t local_sse; + + variance4x4_64_sse4_1(a, a_stride, b, b_stride, &local_sse, &sum); + *sse = (uint32_t)ROUND_POWER_OF_TWO(local_sse, 8); + sum = ROUND_POWER_OF_TWO(sum, 4); + + diff = (int64_t)*sse - ((sum * sum) >> 4); + return diff >= 0 ? (uint32_t)diff : 0; +} + +// Sub-pixel +uint32_t aom_highbd_8_sub_pixel_variance4x4_sse4_1( + const uint8_t *src, int src_stride, int xoffset, int yoffset, + const uint8_t *dst, int dst_stride, uint32_t *sse) { + uint16_t fdata3[(4 + 1) * 4]; + uint16_t temp2[4 * 4]; + + aom_highbd_var_filter_block2d_bil_first_pass( + src, fdata3, src_stride, 1, 4 + 1, 4, bilinear_filters_2t[xoffset]); + aom_highbd_var_filter_block2d_bil_second_pass(fdata3, temp2, 4, 4, 4, 4, + bilinear_filters_2t[yoffset]); + + return aom_highbd_8_variance4x4(CONVERT_TO_BYTEPTR(temp2), 4, dst, dst_stride, + sse); +} + +uint32_t aom_highbd_10_sub_pixel_variance4x4_sse4_1( + const uint8_t *src, int src_stride, int xoffset, int yoffset, + const uint8_t *dst, int dst_stride, uint32_t *sse) { + uint16_t fdata3[(4 + 1) * 4]; + uint16_t temp2[4 * 4]; + + aom_highbd_var_filter_block2d_bil_first_pass( + src, fdata3, src_stride, 1, 4 + 1, 4, bilinear_filters_2t[xoffset]); + aom_highbd_var_filter_block2d_bil_second_pass(fdata3, temp2, 4, 4, 4, 4, + bilinear_filters_2t[yoffset]); + + return aom_highbd_10_variance4x4(CONVERT_TO_BYTEPTR(temp2), 4, dst, + dst_stride, sse); +} + +uint32_t aom_highbd_12_sub_pixel_variance4x4_sse4_1( + const uint8_t *src, int src_stride, int xoffset, int yoffset, + const uint8_t *dst, int dst_stride, uint32_t *sse) { + uint16_t fdata3[(4 + 1) * 4]; + uint16_t temp2[4 * 4]; + + aom_highbd_var_filter_block2d_bil_first_pass( + src, fdata3, src_stride, 1, 4 + 1, 4, bilinear_filters_2t[xoffset]); + aom_highbd_var_filter_block2d_bil_second_pass(fdata3, temp2, 4, 4, 4, 4, + bilinear_filters_2t[yoffset]); + + return aom_highbd_12_variance4x4(CONVERT_TO_BYTEPTR(temp2), 4, dst, + dst_stride, sse); +} + +// Sub-pixel average + +uint32_t aom_highbd_8_sub_pixel_avg_variance4x4_sse4_1( + const uint8_t *src, int src_stride, int xoffset, int yoffset, + const uint8_t *dst, int dst_stride, uint32_t *sse, + const uint8_t *second_pred) { + uint16_t fdata3[(4 + 1) * 4]; + uint16_t temp2[4 * 4]; + DECLARE_ALIGNED(16, uint16_t, temp3[4 * 4]); + + aom_highbd_var_filter_block2d_bil_first_pass( + src, fdata3, src_stride, 1, 4 + 1, 4, bilinear_filters_2t[xoffset]); + aom_highbd_var_filter_block2d_bil_second_pass(fdata3, temp2, 4, 4, 4, 4, + bilinear_filters_2t[yoffset]); + + aom_highbd_comp_avg_pred(CONVERT_TO_BYTEPTR(temp3), second_pred, 4, 4, + CONVERT_TO_BYTEPTR(temp2), 4); + + return aom_highbd_8_variance4x4(CONVERT_TO_BYTEPTR(temp3), 4, dst, dst_stride, + sse); +} + +uint32_t aom_highbd_10_sub_pixel_avg_variance4x4_sse4_1( + const uint8_t *src, int src_stride, int xoffset, int yoffset, + const uint8_t *dst, int dst_stride, uint32_t *sse, + const uint8_t *second_pred) { + uint16_t fdata3[(4 + 1) * 4]; + uint16_t temp2[4 * 4]; + DECLARE_ALIGNED(16, uint16_t, temp3[4 * 4]); + + aom_highbd_var_filter_block2d_bil_first_pass( + src, fdata3, src_stride, 1, 4 + 1, 4, bilinear_filters_2t[xoffset]); + aom_highbd_var_filter_block2d_bil_second_pass(fdata3, temp2, 4, 4, 4, 4, + bilinear_filters_2t[yoffset]); + + aom_highbd_comp_avg_pred(CONVERT_TO_BYTEPTR(temp3), second_pred, 4, 4, + CONVERT_TO_BYTEPTR(temp2), 4); + + return aom_highbd_10_variance4x4(CONVERT_TO_BYTEPTR(temp3), 4, dst, + dst_stride, sse); +} + +uint32_t aom_highbd_12_sub_pixel_avg_variance4x4_sse4_1( + const uint8_t *src, int src_stride, int xoffset, int yoffset, + const uint8_t *dst, int dst_stride, uint32_t *sse, + const uint8_t *second_pred) { + uint16_t fdata3[(4 + 1) * 4]; + uint16_t temp2[4 * 4]; + DECLARE_ALIGNED(16, uint16_t, temp3[4 * 4]); + + aom_highbd_var_filter_block2d_bil_first_pass( + src, fdata3, src_stride, 1, 4 + 1, 4, bilinear_filters_2t[xoffset]); + aom_highbd_var_filter_block2d_bil_second_pass(fdata3, temp2, 4, 4, 4, 4, + bilinear_filters_2t[yoffset]); + + aom_highbd_comp_avg_pred(CONVERT_TO_BYTEPTR(temp3), second_pred, 4, 4, + CONVERT_TO_BYTEPTR(temp2), 4); + + return aom_highbd_12_variance4x4(CONVERT_TO_BYTEPTR(temp3), 4, dst, + dst_stride, sse); +} diff --git a/third_party/aom/aom_dsp/x86/intrapred_avx2.c b/third_party/aom/aom_dsp/x86/intrapred_avx2.c new file mode 100644 index 0000000000..1e67d392e8 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/intrapred_avx2.c @@ -0,0 +1,811 @@ +/* + * Copyright (c) 2017, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include + +#include "config/aom_dsp_rtcd.h" + +static INLINE __m256i dc_sum_64(const uint8_t *ref) { + const __m256i x0 = _mm256_loadu_si256((const __m256i *)ref); + const __m256i x1 = _mm256_loadu_si256((const __m256i *)(ref + 32)); + const __m256i zero = _mm256_setzero_si256(); + __m256i y0 = _mm256_sad_epu8(x0, zero); + __m256i y1 = _mm256_sad_epu8(x1, zero); + y0 = _mm256_add_epi64(y0, y1); + __m256i u0 = _mm256_permute2x128_si256(y0, y0, 1); + y0 = _mm256_add_epi64(u0, y0); + u0 = _mm256_unpackhi_epi64(y0, y0); + return _mm256_add_epi16(y0, u0); +} + +static INLINE __m256i dc_sum_32(const uint8_t *ref) { + const __m256i x = _mm256_loadu_si256((const __m256i *)ref); + const __m256i zero = _mm256_setzero_si256(); + __m256i y = _mm256_sad_epu8(x, zero); + __m256i u = _mm256_permute2x128_si256(y, y, 1); + y = _mm256_add_epi64(u, y); + u = _mm256_unpackhi_epi64(y, y); + return _mm256_add_epi16(y, u); +} + +static INLINE void row_store_32xh(const __m256i *r, int height, uint8_t *dst, + ptrdiff_t stride) { + for (int i = 0; i < height; ++i) { + _mm256_storeu_si256((__m256i *)dst, *r); + dst += stride; + } +} + +static INLINE void row_store_32x2xh(const __m256i *r0, const __m256i *r1, + int height, uint8_t *dst, + ptrdiff_t stride) { + for (int i = 0; i < height; ++i) { + _mm256_storeu_si256((__m256i *)dst, *r0); + _mm256_storeu_si256((__m256i *)(dst + 32), *r1); + dst += stride; + } +} + +static INLINE void row_store_64xh(const __m256i *r, int height, uint8_t *dst, + ptrdiff_t stride) { + for (int i = 0; i < height; ++i) { + _mm256_storeu_si256((__m256i *)dst, *r); + _mm256_storeu_si256((__m256i *)(dst + 32), *r); + dst += stride; + } +} + +void aom_dc_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + const __m256i sum_above = dc_sum_32(above); + __m256i sum_left = dc_sum_32(left); + sum_left = _mm256_add_epi16(sum_left, sum_above); + const __m256i thirtytwo = _mm256_set1_epi16(32); + sum_left = _mm256_add_epi16(sum_left, thirtytwo); + sum_left = _mm256_srai_epi16(sum_left, 6); + const __m256i zero = _mm256_setzero_si256(); + __m256i row = _mm256_shuffle_epi8(sum_left, zero); + row_store_32xh(&row, 32, dst, stride); +} + +void aom_dc_top_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + __m256i sum = dc_sum_32(above); + (void)left; + + const __m256i sixteen = _mm256_set1_epi16(16); + sum = _mm256_add_epi16(sum, sixteen); + sum = _mm256_srai_epi16(sum, 5); + const __m256i zero = _mm256_setzero_si256(); + __m256i row = _mm256_shuffle_epi8(sum, zero); + row_store_32xh(&row, 32, dst, stride); +} + +void aom_dc_left_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + __m256i sum = dc_sum_32(left); + (void)above; + + const __m256i sixteen = _mm256_set1_epi16(16); + sum = _mm256_add_epi16(sum, sixteen); + sum = _mm256_srai_epi16(sum, 5); + const __m256i zero = _mm256_setzero_si256(); + __m256i row = _mm256_shuffle_epi8(sum, zero); + row_store_32xh(&row, 32, dst, stride); +} + +void aom_dc_128_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + (void)above; + (void)left; + const __m256i row = _mm256_set1_epi8((uint8_t)0x80); + row_store_32xh(&row, 32, dst, stride); +} + +void aom_v_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + const __m256i row = _mm256_loadu_si256((const __m256i *)above); + (void)left; + row_store_32xh(&row, 32, dst, stride); +} + +// There are 32 rows togeter. This function does line: +// 0,1,2,3, and 16,17,18,19. The next call would do +// 4,5,6,7, and 20,21,22,23. So 4 times of calling +// would finish 32 rows. +static INLINE void h_predictor_32x8line(const __m256i *row, uint8_t *dst, + ptrdiff_t stride) { + __m256i t[4]; + __m256i m = _mm256_setzero_si256(); + const __m256i inc = _mm256_set1_epi8(4); + int i; + + for (i = 0; i < 4; i++) { + t[i] = _mm256_shuffle_epi8(*row, m); + __m256i r0 = _mm256_permute2x128_si256(t[i], t[i], 0); + __m256i r1 = _mm256_permute2x128_si256(t[i], t[i], 0x11); + _mm256_storeu_si256((__m256i *)dst, r0); + _mm256_storeu_si256((__m256i *)(dst + (stride << 4)), r1); + dst += stride; + m = _mm256_add_epi8(m, inc); + } +} + +void aom_h_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + (void)above; + const __m256i left_col = _mm256_loadu_si256((__m256i const *)left); + + __m256i u = _mm256_unpacklo_epi8(left_col, left_col); + + __m256i v = _mm256_unpacklo_epi8(u, u); + h_predictor_32x8line(&v, dst, stride); + dst += stride << 2; + + v = _mm256_unpackhi_epi8(u, u); + h_predictor_32x8line(&v, dst, stride); + dst += stride << 2; + + u = _mm256_unpackhi_epi8(left_col, left_col); + + v = _mm256_unpacklo_epi8(u, u); + h_predictor_32x8line(&v, dst, stride); + dst += stride << 2; + + v = _mm256_unpackhi_epi8(u, u); + h_predictor_32x8line(&v, dst, stride); +} + +// ----------------------------------------------------------------------------- +// Rectangle + +// TODO(luoyi) The following two functions are shared with intrapred_sse2.c. +// Use a header file, intrapred_common_x86.h +static INLINE __m128i dc_sum_16_sse2(const uint8_t *ref) { + __m128i x = _mm_load_si128((__m128i const *)ref); + const __m128i zero = _mm_setzero_si128(); + x = _mm_sad_epu8(x, zero); + const __m128i high = _mm_unpackhi_epi64(x, x); + return _mm_add_epi16(x, high); +} + +static INLINE __m128i dc_sum_32_sse2(const uint8_t *ref) { + __m128i x0 = _mm_load_si128((__m128i const *)ref); + __m128i x1 = _mm_load_si128((__m128i const *)(ref + 16)); + const __m128i zero = _mm_setzero_si128(); + x0 = _mm_sad_epu8(x0, zero); + x1 = _mm_sad_epu8(x1, zero); + x0 = _mm_add_epi16(x0, x1); + const __m128i high = _mm_unpackhi_epi64(x0, x0); + return _mm_add_epi16(x0, high); +} + +void aom_dc_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + const __m128i top_sum = dc_sum_32_sse2(above); + __m128i left_sum = dc_sum_16_sse2(left); + left_sum = _mm_add_epi16(top_sum, left_sum); + uint32_t sum = _mm_cvtsi128_si32(left_sum); + sum += 24; + sum /= 48; + const __m256i row = _mm256_set1_epi8((uint8_t)sum); + row_store_32xh(&row, 16, dst, stride); +} + +void aom_dc_predictor_32x64_avx2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + const __m256i sum_above = dc_sum_32(above); + __m256i sum_left = dc_sum_64(left); + sum_left = _mm256_add_epi16(sum_left, sum_above); + uint32_t sum = _mm_cvtsi128_si32(_mm256_castsi256_si128(sum_left)); + sum += 48; + sum /= 96; + const __m256i row = _mm256_set1_epi8((uint8_t)sum); + row_store_32xh(&row, 64, dst, stride); +} + +void aom_dc_predictor_64x64_avx2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + const __m256i sum_above = dc_sum_64(above); + __m256i sum_left = dc_sum_64(left); + sum_left = _mm256_add_epi16(sum_left, sum_above); + uint32_t sum = _mm_cvtsi128_si32(_mm256_castsi256_si128(sum_left)); + sum += 64; + sum /= 128; + const __m256i row = _mm256_set1_epi8((uint8_t)sum); + row_store_64xh(&row, 64, dst, stride); +} + +void aom_dc_predictor_64x32_avx2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + const __m256i sum_above = dc_sum_64(above); + __m256i sum_left = dc_sum_32(left); + sum_left = _mm256_add_epi16(sum_left, sum_above); + uint32_t sum = _mm_cvtsi128_si32(_mm256_castsi256_si128(sum_left)); + sum += 48; + sum /= 96; + const __m256i row = _mm256_set1_epi8((uint8_t)sum); + row_store_64xh(&row, 32, dst, stride); +} + +void aom_dc_predictor_64x16_avx2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + const __m256i sum_above = dc_sum_64(above); + __m256i sum_left = _mm256_castsi128_si256(dc_sum_16_sse2(left)); + sum_left = _mm256_add_epi16(sum_left, sum_above); + uint32_t sum = _mm_cvtsi128_si32(_mm256_castsi256_si128(sum_left)); + sum += 40; + sum /= 80; + const __m256i row = _mm256_set1_epi8((uint8_t)sum); + row_store_64xh(&row, 16, dst, stride); +} + +void aom_dc_top_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + __m256i sum = dc_sum_32(above); + (void)left; + + const __m256i sixteen = _mm256_set1_epi16(16); + sum = _mm256_add_epi16(sum, sixteen); + sum = _mm256_srai_epi16(sum, 5); + const __m256i zero = _mm256_setzero_si256(); + __m256i row = _mm256_shuffle_epi8(sum, zero); + row_store_32xh(&row, 16, dst, stride); +} + +void aom_dc_top_predictor_32x64_avx2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + __m256i sum = dc_sum_32(above); + (void)left; + + const __m256i sixteen = _mm256_set1_epi16(16); + sum = _mm256_add_epi16(sum, sixteen); + sum = _mm256_srai_epi16(sum, 5); + const __m256i zero = _mm256_setzero_si256(); + __m256i row = _mm256_shuffle_epi8(sum, zero); + row_store_32xh(&row, 64, dst, stride); +} + +void aom_dc_top_predictor_64x64_avx2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + __m256i sum = dc_sum_64(above); + (void)left; + + const __m256i thirtytwo = _mm256_set1_epi16(32); + sum = _mm256_add_epi16(sum, thirtytwo); + sum = _mm256_srai_epi16(sum, 6); + const __m256i zero = _mm256_setzero_si256(); + __m256i row = _mm256_shuffle_epi8(sum, zero); + row_store_64xh(&row, 64, dst, stride); +} + +void aom_dc_top_predictor_64x32_avx2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + __m256i sum = dc_sum_64(above); + (void)left; + + const __m256i thirtytwo = _mm256_set1_epi16(32); + sum = _mm256_add_epi16(sum, thirtytwo); + sum = _mm256_srai_epi16(sum, 6); + const __m256i zero = _mm256_setzero_si256(); + __m256i row = _mm256_shuffle_epi8(sum, zero); + row_store_64xh(&row, 32, dst, stride); +} + +void aom_dc_top_predictor_64x16_avx2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + __m256i sum = dc_sum_64(above); + (void)left; + + const __m256i thirtytwo = _mm256_set1_epi16(32); + sum = _mm256_add_epi16(sum, thirtytwo); + sum = _mm256_srai_epi16(sum, 6); + const __m256i zero = _mm256_setzero_si256(); + __m256i row = _mm256_shuffle_epi8(sum, zero); + row_store_64xh(&row, 16, dst, stride); +} + +void aom_dc_left_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + __m128i sum = dc_sum_16_sse2(left); + (void)above; + + const __m128i eight = _mm_set1_epi16(8); + sum = _mm_add_epi16(sum, eight); + sum = _mm_srai_epi16(sum, 4); + const __m128i zero = _mm_setzero_si128(); + const __m128i r = _mm_shuffle_epi8(sum, zero); + const __m256i row = _mm256_inserti128_si256(_mm256_castsi128_si256(r), r, 1); + row_store_32xh(&row, 16, dst, stride); +} + +void aom_dc_left_predictor_32x64_avx2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + __m256i sum = dc_sum_64(left); + (void)above; + + const __m256i thirtytwo = _mm256_set1_epi16(32); + sum = _mm256_add_epi16(sum, thirtytwo); + sum = _mm256_srai_epi16(sum, 6); + const __m256i zero = _mm256_setzero_si256(); + __m256i row = _mm256_shuffle_epi8(sum, zero); + row_store_32xh(&row, 64, dst, stride); +} + +void aom_dc_left_predictor_64x64_avx2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + __m256i sum = dc_sum_64(left); + (void)above; + + const __m256i thirtytwo = _mm256_set1_epi16(32); + sum = _mm256_add_epi16(sum, thirtytwo); + sum = _mm256_srai_epi16(sum, 6); + const __m256i zero = _mm256_setzero_si256(); + __m256i row = _mm256_shuffle_epi8(sum, zero); + row_store_64xh(&row, 64, dst, stride); +} + +void aom_dc_left_predictor_64x32_avx2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + __m256i sum = dc_sum_32(left); + (void)above; + + const __m256i sixteen = _mm256_set1_epi16(16); + sum = _mm256_add_epi16(sum, sixteen); + sum = _mm256_srai_epi16(sum, 5); + const __m256i zero = _mm256_setzero_si256(); + __m256i row = _mm256_shuffle_epi8(sum, zero); + row_store_64xh(&row, 32, dst, stride); +} + +void aom_dc_left_predictor_64x16_avx2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + __m128i sum = dc_sum_16_sse2(left); + (void)above; + + const __m128i eight = _mm_set1_epi16(8); + sum = _mm_add_epi16(sum, eight); + sum = _mm_srai_epi16(sum, 4); + const __m128i zero = _mm_setzero_si128(); + const __m128i r = _mm_shuffle_epi8(sum, zero); + const __m256i row = _mm256_inserti128_si256(_mm256_castsi128_si256(r), r, 1); + row_store_64xh(&row, 16, dst, stride); +} + +void aom_dc_128_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + (void)above; + (void)left; + const __m256i row = _mm256_set1_epi8((uint8_t)0x80); + row_store_32xh(&row, 16, dst, stride); +} + +void aom_dc_128_predictor_32x64_avx2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + (void)above; + (void)left; + const __m256i row = _mm256_set1_epi8((uint8_t)0x80); + row_store_32xh(&row, 64, dst, stride); +} + +void aom_dc_128_predictor_64x64_avx2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + (void)above; + (void)left; + const __m256i row = _mm256_set1_epi8((uint8_t)0x80); + row_store_64xh(&row, 64, dst, stride); +} + +void aom_dc_128_predictor_64x32_avx2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + (void)above; + (void)left; + const __m256i row = _mm256_set1_epi8((uint8_t)0x80); + row_store_64xh(&row, 32, dst, stride); +} + +void aom_dc_128_predictor_64x16_avx2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + (void)above; + (void)left; + const __m256i row = _mm256_set1_epi8((uint8_t)0x80); + row_store_64xh(&row, 16, dst, stride); +} + +void aom_v_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + const __m256i row = _mm256_loadu_si256((const __m256i *)above); + (void)left; + row_store_32xh(&row, 16, dst, stride); +} + +void aom_v_predictor_32x64_avx2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + const __m256i row = _mm256_loadu_si256((const __m256i *)above); + (void)left; + row_store_32xh(&row, 64, dst, stride); +} + +void aom_v_predictor_64x64_avx2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + const __m256i row0 = _mm256_loadu_si256((const __m256i *)above); + const __m256i row1 = _mm256_loadu_si256((const __m256i *)(above + 32)); + (void)left; + row_store_32x2xh(&row0, &row1, 64, dst, stride); +} + +void aom_v_predictor_64x32_avx2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + const __m256i row0 = _mm256_loadu_si256((const __m256i *)above); + const __m256i row1 = _mm256_loadu_si256((const __m256i *)(above + 32)); + (void)left; + row_store_32x2xh(&row0, &row1, 32, dst, stride); +} + +void aom_v_predictor_64x16_avx2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + const __m256i row0 = _mm256_loadu_si256((const __m256i *)above); + const __m256i row1 = _mm256_loadu_si256((const __m256i *)(above + 32)); + (void)left; + row_store_32x2xh(&row0, &row1, 16, dst, stride); +} + +// ----------------------------------------------------------------------------- +// PAETH_PRED + +// Return 16 16-bit pixels in one row (__m256i) +static INLINE __m256i paeth_pred(const __m256i *left, const __m256i *top, + const __m256i *topleft) { + const __m256i base = + _mm256_sub_epi16(_mm256_add_epi16(*top, *left), *topleft); + + __m256i pl = _mm256_abs_epi16(_mm256_sub_epi16(base, *left)); + __m256i pt = _mm256_abs_epi16(_mm256_sub_epi16(base, *top)); + __m256i ptl = _mm256_abs_epi16(_mm256_sub_epi16(base, *topleft)); + + __m256i mask1 = _mm256_cmpgt_epi16(pl, pt); + mask1 = _mm256_or_si256(mask1, _mm256_cmpgt_epi16(pl, ptl)); + __m256i mask2 = _mm256_cmpgt_epi16(pt, ptl); + + pl = _mm256_andnot_si256(mask1, *left); + + ptl = _mm256_and_si256(mask2, *topleft); + pt = _mm256_andnot_si256(mask2, *top); + pt = _mm256_or_si256(pt, ptl); + pt = _mm256_and_si256(mask1, pt); + + return _mm256_or_si256(pt, pl); +} + +// Return 16 8-bit pixels in one row (__m128i) +static INLINE __m128i paeth_16x1_pred(const __m256i *left, const __m256i *top, + const __m256i *topleft) { + const __m256i p0 = paeth_pred(left, top, topleft); + const __m256i p1 = _mm256_permute4x64_epi64(p0, 0xe); + const __m256i p = _mm256_packus_epi16(p0, p1); + return _mm256_castsi256_si128(p); +} + +static INLINE __m256i get_top_vector(const uint8_t *above) { + const __m128i x = _mm_load_si128((const __m128i *)above); + const __m128i zero = _mm_setzero_si128(); + const __m128i t0 = _mm_unpacklo_epi8(x, zero); + const __m128i t1 = _mm_unpackhi_epi8(x, zero); + return _mm256_inserti128_si256(_mm256_castsi128_si256(t0), t1, 1); +} + +void aom_paeth_predictor_16x8_avx2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + __m128i x = _mm_loadl_epi64((const __m128i *)left); + const __m256i l = _mm256_inserti128_si256(_mm256_castsi128_si256(x), x, 1); + const __m256i tl16 = _mm256_set1_epi16((uint16_t)above[-1]); + __m256i rep = _mm256_set1_epi16(0x8000); + const __m256i one = _mm256_set1_epi16(1); + const __m256i top = get_top_vector(above); + + int i; + for (i = 0; i < 8; ++i) { + const __m256i l16 = _mm256_shuffle_epi8(l, rep); + const __m128i row = paeth_16x1_pred(&l16, &top, &tl16); + + _mm_store_si128((__m128i *)dst, row); + dst += stride; + rep = _mm256_add_epi16(rep, one); + } +} + +static INLINE __m256i get_left_vector(const uint8_t *left) { + const __m128i x = _mm_load_si128((const __m128i *)left); + return _mm256_inserti128_si256(_mm256_castsi128_si256(x), x, 1); +} + +void aom_paeth_predictor_16x16_avx2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + const __m256i l = get_left_vector(left); + const __m256i tl16 = _mm256_set1_epi16((uint16_t)above[-1]); + __m256i rep = _mm256_set1_epi16(0x8000); + const __m256i one = _mm256_set1_epi16(1); + const __m256i top = get_top_vector(above); + + int i; + for (i = 0; i < 16; ++i) { + const __m256i l16 = _mm256_shuffle_epi8(l, rep); + const __m128i row = paeth_16x1_pred(&l16, &top, &tl16); + + _mm_store_si128((__m128i *)dst, row); + dst += stride; + rep = _mm256_add_epi16(rep, one); + } +} + +void aom_paeth_predictor_16x32_avx2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + __m256i l = get_left_vector(left); + const __m256i tl16 = _mm256_set1_epi16((uint16_t)above[-1]); + __m256i rep = _mm256_set1_epi16(0x8000); + const __m256i one = _mm256_set1_epi16(1); + const __m256i top = get_top_vector(above); + + int i; + for (i = 0; i < 16; ++i) { + const __m256i l16 = _mm256_shuffle_epi8(l, rep); + const __m128i row = paeth_16x1_pred(&l16, &top, &tl16); + + _mm_store_si128((__m128i *)dst, row); + dst += stride; + rep = _mm256_add_epi16(rep, one); + } + + l = get_left_vector(left + 16); + rep = _mm256_set1_epi16(0x8000); + for (i = 0; i < 16; ++i) { + const __m256i l16 = _mm256_shuffle_epi8(l, rep); + const __m128i row = paeth_16x1_pred(&l16, &top, &tl16); + + _mm_store_si128((__m128i *)dst, row); + dst += stride; + rep = _mm256_add_epi16(rep, one); + } +} + +void aom_paeth_predictor_16x64_avx2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + const __m256i tl16 = _mm256_set1_epi16((uint16_t)above[-1]); + const __m256i one = _mm256_set1_epi16(1); + const __m256i top = get_top_vector(above); + + for (int j = 0; j < 4; ++j) { + const __m256i l = get_left_vector(left + j * 16); + __m256i rep = _mm256_set1_epi16(0x8000); + for (int i = 0; i < 16; ++i) { + const __m256i l16 = _mm256_shuffle_epi8(l, rep); + const __m128i row = paeth_16x1_pred(&l16, &top, &tl16); + + _mm_store_si128((__m128i *)dst, row); + dst += stride; + rep = _mm256_add_epi16(rep, one); + } + } +} + +// Return 32 8-bit pixels in one row (__m256i) +static INLINE __m256i paeth_32x1_pred(const __m256i *left, const __m256i *top0, + const __m256i *top1, + const __m256i *topleft) { + __m256i p0 = paeth_pred(left, top0, topleft); + __m256i p1 = _mm256_permute4x64_epi64(p0, 0xe); + const __m256i x0 = _mm256_packus_epi16(p0, p1); + + p0 = paeth_pred(left, top1, topleft); + p1 = _mm256_permute4x64_epi64(p0, 0xe); + const __m256i x1 = _mm256_packus_epi16(p0, p1); + + return _mm256_permute2x128_si256(x0, x1, 0x20); +} + +void aom_paeth_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + const __m256i l = get_left_vector(left); + const __m256i t0 = get_top_vector(above); + const __m256i t1 = get_top_vector(above + 16); + const __m256i tl = _mm256_set1_epi16((uint16_t)above[-1]); + __m256i rep = _mm256_set1_epi16(0x8000); + const __m256i one = _mm256_set1_epi16(1); + + int i; + for (i = 0; i < 16; ++i) { + const __m256i l16 = _mm256_shuffle_epi8(l, rep); + + const __m256i r = paeth_32x1_pred(&l16, &t0, &t1, &tl); + + _mm256_storeu_si256((__m256i *)dst, r); + + dst += stride; + rep = _mm256_add_epi16(rep, one); + } +} + +void aom_paeth_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + __m256i l = get_left_vector(left); + const __m256i t0 = get_top_vector(above); + const __m256i t1 = get_top_vector(above + 16); + const __m256i tl = _mm256_set1_epi16((uint16_t)above[-1]); + __m256i rep = _mm256_set1_epi16(0x8000); + const __m256i one = _mm256_set1_epi16(1); + + int i; + for (i = 0; i < 16; ++i) { + const __m256i l16 = _mm256_shuffle_epi8(l, rep); + + const __m128i r0 = paeth_16x1_pred(&l16, &t0, &tl); + const __m128i r1 = paeth_16x1_pred(&l16, &t1, &tl); + + _mm_store_si128((__m128i *)dst, r0); + _mm_store_si128((__m128i *)(dst + 16), r1); + + dst += stride; + rep = _mm256_add_epi16(rep, one); + } + + l = get_left_vector(left + 16); + rep = _mm256_set1_epi16(0x8000); + for (i = 0; i < 16; ++i) { + const __m256i l16 = _mm256_shuffle_epi8(l, rep); + + const __m128i r0 = paeth_16x1_pred(&l16, &t0, &tl); + const __m128i r1 = paeth_16x1_pred(&l16, &t1, &tl); + + _mm_store_si128((__m128i *)dst, r0); + _mm_store_si128((__m128i *)(dst + 16), r1); + + dst += stride; + rep = _mm256_add_epi16(rep, one); + } +} + +void aom_paeth_predictor_32x64_avx2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + const __m256i t0 = get_top_vector(above); + const __m256i t1 = get_top_vector(above + 16); + const __m256i tl = _mm256_set1_epi16((uint16_t)above[-1]); + const __m256i one = _mm256_set1_epi16(1); + + int i, j; + for (j = 0; j < 4; ++j) { + const __m256i l = get_left_vector(left + j * 16); + __m256i rep = _mm256_set1_epi16(0x8000); + for (i = 0; i < 16; ++i) { + const __m256i l16 = _mm256_shuffle_epi8(l, rep); + + const __m128i r0 = paeth_16x1_pred(&l16, &t0, &tl); + const __m128i r1 = paeth_16x1_pred(&l16, &t1, &tl); + + _mm_store_si128((__m128i *)dst, r0); + _mm_store_si128((__m128i *)(dst + 16), r1); + + dst += stride; + rep = _mm256_add_epi16(rep, one); + } + } +} + +void aom_paeth_predictor_64x32_avx2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + const __m256i t0 = get_top_vector(above); + const __m256i t1 = get_top_vector(above + 16); + const __m256i t2 = get_top_vector(above + 32); + const __m256i t3 = get_top_vector(above + 48); + const __m256i tl = _mm256_set1_epi16((uint16_t)above[-1]); + const __m256i one = _mm256_set1_epi16(1); + + int i, j; + for (j = 0; j < 2; ++j) { + const __m256i l = get_left_vector(left + j * 16); + __m256i rep = _mm256_set1_epi16(0x8000); + for (i = 0; i < 16; ++i) { + const __m256i l16 = _mm256_shuffle_epi8(l, rep); + + const __m128i r0 = paeth_16x1_pred(&l16, &t0, &tl); + const __m128i r1 = paeth_16x1_pred(&l16, &t1, &tl); + const __m128i r2 = paeth_16x1_pred(&l16, &t2, &tl); + const __m128i r3 = paeth_16x1_pred(&l16, &t3, &tl); + + _mm_store_si128((__m128i *)dst, r0); + _mm_store_si128((__m128i *)(dst + 16), r1); + _mm_store_si128((__m128i *)(dst + 32), r2); + _mm_store_si128((__m128i *)(dst + 48), r3); + + dst += stride; + rep = _mm256_add_epi16(rep, one); + } + } +} + +void aom_paeth_predictor_64x64_avx2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + const __m256i t0 = get_top_vector(above); + const __m256i t1 = get_top_vector(above + 16); + const __m256i t2 = get_top_vector(above + 32); + const __m256i t3 = get_top_vector(above + 48); + const __m256i tl = _mm256_set1_epi16((uint16_t)above[-1]); + const __m256i one = _mm256_set1_epi16(1); + + int i, j; + for (j = 0; j < 4; ++j) { + const __m256i l = get_left_vector(left + j * 16); + __m256i rep = _mm256_set1_epi16(0x8000); + for (i = 0; i < 16; ++i) { + const __m256i l16 = _mm256_shuffle_epi8(l, rep); + + const __m128i r0 = paeth_16x1_pred(&l16, &t0, &tl); + const __m128i r1 = paeth_16x1_pred(&l16, &t1, &tl); + const __m128i r2 = paeth_16x1_pred(&l16, &t2, &tl); + const __m128i r3 = paeth_16x1_pred(&l16, &t3, &tl); + + _mm_store_si128((__m128i *)dst, r0); + _mm_store_si128((__m128i *)(dst + 16), r1); + _mm_store_si128((__m128i *)(dst + 32), r2); + _mm_store_si128((__m128i *)(dst + 48), r3); + + dst += stride; + rep = _mm256_add_epi16(rep, one); + } + } +} + +void aom_paeth_predictor_64x16_avx2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + const __m256i t0 = get_top_vector(above); + const __m256i t1 = get_top_vector(above + 16); + const __m256i t2 = get_top_vector(above + 32); + const __m256i t3 = get_top_vector(above + 48); + const __m256i tl = _mm256_set1_epi16((uint16_t)above[-1]); + const __m256i one = _mm256_set1_epi16(1); + + int i; + const __m256i l = get_left_vector(left); + __m256i rep = _mm256_set1_epi16(0x8000); + for (i = 0; i < 16; ++i) { + const __m256i l16 = _mm256_shuffle_epi8(l, rep); + + const __m128i r0 = paeth_16x1_pred(&l16, &t0, &tl); + const __m128i r1 = paeth_16x1_pred(&l16, &t1, &tl); + const __m128i r2 = paeth_16x1_pred(&l16, &t2, &tl); + const __m128i r3 = paeth_16x1_pred(&l16, &t3, &tl); + + _mm_store_si128((__m128i *)dst, r0); + _mm_store_si128((__m128i *)(dst + 16), r1); + _mm_store_si128((__m128i *)(dst + 32), r2); + _mm_store_si128((__m128i *)(dst + 48), r3); + + dst += stride; + rep = _mm256_add_epi16(rep, one); + } +} diff --git a/third_party/aom/aom_dsp/x86/intrapred_sse2.c b/third_party/aom/aom_dsp/x86/intrapred_sse2.c new file mode 100644 index 0000000000..5b2452c8eb --- /dev/null +++ b/third_party/aom/aom_dsp/x86/intrapred_sse2.c @@ -0,0 +1,1430 @@ +/* + * Copyright (c) 2017, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include + +#include "config/aom_dsp_rtcd.h" + +static INLINE void dc_store_4xh(uint32_t dc, int height, uint8_t *dst, + ptrdiff_t stride) { + for (int i = 0; i < height; i += 2) { + *(uint32_t *)dst = dc; + dst += stride; + *(uint32_t *)dst = dc; + dst += stride; + } +} + +static INLINE void dc_store_8xh(const __m128i *row, int height, uint8_t *dst, + ptrdiff_t stride) { + int i; + for (i = 0; i < height; ++i) { + _mm_storel_epi64((__m128i *)dst, *row); + dst += stride; + } +} + +static INLINE void dc_store_16xh(const __m128i *row, int height, uint8_t *dst, + ptrdiff_t stride) { + int i; + for (i = 0; i < height; ++i) { + _mm_store_si128((__m128i *)dst, *row); + dst += stride; + } +} + +static INLINE void dc_store_32xh(const __m128i *row, int height, uint8_t *dst, + ptrdiff_t stride) { + int i; + for (i = 0; i < height; ++i) { + _mm_store_si128((__m128i *)dst, *row); + _mm_store_si128((__m128i *)(dst + 16), *row); + dst += stride; + } +} + +static INLINE void dc_store_64xh(const __m128i *row, int height, uint8_t *dst, + ptrdiff_t stride) { + for (int i = 0; i < height; ++i) { + _mm_store_si128((__m128i *)dst, *row); + _mm_store_si128((__m128i *)(dst + 16), *row); + _mm_store_si128((__m128i *)(dst + 32), *row); + _mm_store_si128((__m128i *)(dst + 48), *row); + dst += stride; + } +} + +static INLINE __m128i dc_sum_4(const uint8_t *ref) { + __m128i x = _mm_loadl_epi64((__m128i const *)ref); + const __m128i zero = _mm_setzero_si128(); + x = _mm_unpacklo_epi8(x, zero); + return _mm_sad_epu8(x, zero); +} + +static INLINE __m128i dc_sum_8(const uint8_t *ref) { + __m128i x = _mm_loadl_epi64((__m128i const *)ref); + const __m128i zero = _mm_setzero_si128(); + return _mm_sad_epu8(x, zero); +} + +static INLINE __m128i dc_sum_16(const uint8_t *ref) { + __m128i x = _mm_load_si128((__m128i const *)ref); + const __m128i zero = _mm_setzero_si128(); + x = _mm_sad_epu8(x, zero); + const __m128i high = _mm_unpackhi_epi64(x, x); + return _mm_add_epi16(x, high); +} + +static INLINE __m128i dc_sum_32(const uint8_t *ref) { + __m128i x0 = _mm_load_si128((__m128i const *)ref); + __m128i x1 = _mm_load_si128((__m128i const *)(ref + 16)); + const __m128i zero = _mm_setzero_si128(); + x0 = _mm_sad_epu8(x0, zero); + x1 = _mm_sad_epu8(x1, zero); + x0 = _mm_add_epi16(x0, x1); + const __m128i high = _mm_unpackhi_epi64(x0, x0); + return _mm_add_epi16(x0, high); +} + +static INLINE __m128i dc_sum_64(const uint8_t *ref) { + __m128i x0 = _mm_load_si128((__m128i const *)ref); + __m128i x1 = _mm_load_si128((__m128i const *)(ref + 16)); + __m128i x2 = _mm_load_si128((__m128i const *)(ref + 32)); + __m128i x3 = _mm_load_si128((__m128i const *)(ref + 48)); + const __m128i zero = _mm_setzero_si128(); + x0 = _mm_sad_epu8(x0, zero); + x1 = _mm_sad_epu8(x1, zero); + x2 = _mm_sad_epu8(x2, zero); + x3 = _mm_sad_epu8(x3, zero); + x0 = _mm_add_epi16(x0, x1); + x2 = _mm_add_epi16(x2, x3); + x0 = _mm_add_epi16(x0, x2); + const __m128i high = _mm_unpackhi_epi64(x0, x0); + return _mm_add_epi16(x0, high); +} + +#define DC_MULTIPLIER_1X2 0x5556 +#define DC_MULTIPLIER_1X4 0x3334 + +#define DC_SHIFT2 16 + +static INLINE int divide_using_multiply_shift(int num, int shift1, + int multiplier) { + const int interm = num >> shift1; + return interm * multiplier >> DC_SHIFT2; +} + +// ----------------------------------------------------------------------------- +// DC_PRED + +void aom_dc_predictor_4x8_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + const __m128i sum_left = dc_sum_8(left); + __m128i sum_above = dc_sum_4(above); + sum_above = _mm_add_epi16(sum_left, sum_above); + + uint32_t sum = _mm_cvtsi128_si32(sum_above); + sum += 6; + sum = divide_using_multiply_shift(sum, 2, DC_MULTIPLIER_1X2); + + const __m128i row = _mm_set1_epi8((uint8_t)sum); + const uint32_t pred = _mm_cvtsi128_si32(row); + dc_store_4xh(pred, 8, dst, stride); +} + +void aom_dc_predictor_4x16_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + const __m128i sum_left = dc_sum_16(left); + __m128i sum_above = dc_sum_4(above); + sum_above = _mm_add_epi16(sum_left, sum_above); + + uint32_t sum = _mm_cvtsi128_si32(sum_above); + sum += 10; + sum = divide_using_multiply_shift(sum, 2, DC_MULTIPLIER_1X4); + + const __m128i row = _mm_set1_epi8((uint8_t)sum); + const uint32_t pred = _mm_cvtsi128_si32(row); + dc_store_4xh(pred, 16, dst, stride); +} + +void aom_dc_predictor_8x4_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + const __m128i sum_left = dc_sum_4(left); + __m128i sum_above = dc_sum_8(above); + sum_above = _mm_add_epi16(sum_above, sum_left); + + uint32_t sum = _mm_cvtsi128_si32(sum_above); + sum += 6; + sum = divide_using_multiply_shift(sum, 2, DC_MULTIPLIER_1X2); + + const __m128i row = _mm_set1_epi8((uint8_t)sum); + dc_store_8xh(&row, 4, dst, stride); +} + +void aom_dc_predictor_8x16_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + const __m128i sum_left = dc_sum_16(left); + __m128i sum_above = dc_sum_8(above); + sum_above = _mm_add_epi16(sum_above, sum_left); + + uint32_t sum = _mm_cvtsi128_si32(sum_above); + sum += 12; + sum = divide_using_multiply_shift(sum, 3, DC_MULTIPLIER_1X2); + const __m128i row = _mm_set1_epi8((uint8_t)sum); + dc_store_8xh(&row, 16, dst, stride); +} + +void aom_dc_predictor_8x32_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + const __m128i sum_left = dc_sum_32(left); + __m128i sum_above = dc_sum_8(above); + sum_above = _mm_add_epi16(sum_above, sum_left); + + uint32_t sum = _mm_cvtsi128_si32(sum_above); + sum += 20; + sum = divide_using_multiply_shift(sum, 3, DC_MULTIPLIER_1X4); + const __m128i row = _mm_set1_epi8((uint8_t)sum); + dc_store_8xh(&row, 32, dst, stride); +} + +void aom_dc_predictor_16x4_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + const __m128i sum_left = dc_sum_4(left); + __m128i sum_above = dc_sum_16(above); + sum_above = _mm_add_epi16(sum_above, sum_left); + + uint32_t sum = _mm_cvtsi128_si32(sum_above); + sum += 10; + sum = divide_using_multiply_shift(sum, 2, DC_MULTIPLIER_1X4); + const __m128i row = _mm_set1_epi8((uint8_t)sum); + dc_store_16xh(&row, 4, dst, stride); +} + +void aom_dc_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + const __m128i sum_left = dc_sum_8(left); + __m128i sum_above = dc_sum_16(above); + sum_above = _mm_add_epi16(sum_above, sum_left); + + uint32_t sum = _mm_cvtsi128_si32(sum_above); + sum += 12; + sum = divide_using_multiply_shift(sum, 3, DC_MULTIPLIER_1X2); + const __m128i row = _mm_set1_epi8((uint8_t)sum); + dc_store_16xh(&row, 8, dst, stride); +} + +void aom_dc_predictor_16x32_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + const __m128i sum_left = dc_sum_32(left); + __m128i sum_above = dc_sum_16(above); + sum_above = _mm_add_epi16(sum_left, sum_above); + + uint32_t sum = _mm_cvtsi128_si32(sum_above); + sum += 24; + sum = divide_using_multiply_shift(sum, 4, DC_MULTIPLIER_1X2); + const __m128i row = _mm_set1_epi8((uint8_t)sum); + dc_store_16xh(&row, 32, dst, stride); +} + +void aom_dc_predictor_16x64_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + const __m128i sum_left = dc_sum_64(left); + __m128i sum_above = dc_sum_16(above); + sum_above = _mm_add_epi16(sum_left, sum_above); + + uint32_t sum = _mm_cvtsi128_si32(sum_above); + sum += 40; + sum = divide_using_multiply_shift(sum, 4, DC_MULTIPLIER_1X4); + const __m128i row = _mm_set1_epi8((uint8_t)sum); + dc_store_16xh(&row, 64, dst, stride); +} + +void aom_dc_predictor_32x8_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + __m128i sum_above = dc_sum_32(above); + const __m128i sum_left = dc_sum_8(left); + sum_above = _mm_add_epi16(sum_above, sum_left); + + uint32_t sum = _mm_cvtsi128_si32(sum_above); + sum += 20; + sum = divide_using_multiply_shift(sum, 3, DC_MULTIPLIER_1X4); + const __m128i row = _mm_set1_epi8((uint8_t)sum); + dc_store_32xh(&row, 8, dst, stride); +} + +void aom_dc_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + __m128i sum_above = dc_sum_32(above); + const __m128i sum_left = dc_sum_16(left); + sum_above = _mm_add_epi16(sum_above, sum_left); + + uint32_t sum = _mm_cvtsi128_si32(sum_above); + sum += 24; + sum = divide_using_multiply_shift(sum, 4, DC_MULTIPLIER_1X2); + const __m128i row = _mm_set1_epi8((uint8_t)sum); + dc_store_32xh(&row, 16, dst, stride); +} + +void aom_dc_predictor_32x64_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + __m128i sum_above = dc_sum_32(above); + const __m128i sum_left = dc_sum_64(left); + sum_above = _mm_add_epi16(sum_above, sum_left); + + uint32_t sum = _mm_cvtsi128_si32(sum_above); + sum += 48; + sum = divide_using_multiply_shift(sum, 5, DC_MULTIPLIER_1X2); + const __m128i row = _mm_set1_epi8((uint8_t)sum); + dc_store_32xh(&row, 64, dst, stride); +} + +void aom_dc_predictor_64x64_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + __m128i sum_above = dc_sum_64(above); + const __m128i sum_left = dc_sum_64(left); + sum_above = _mm_add_epi16(sum_above, sum_left); + + uint32_t sum = _mm_cvtsi128_si32(sum_above); + sum += 64; + sum /= 128; + const __m128i row = _mm_set1_epi8((uint8_t)sum); + dc_store_64xh(&row, 64, dst, stride); +} + +void aom_dc_predictor_64x32_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + __m128i sum_above = dc_sum_64(above); + const __m128i sum_left = dc_sum_32(left); + sum_above = _mm_add_epi16(sum_above, sum_left); + + uint32_t sum = _mm_cvtsi128_si32(sum_above); + sum += 48; + sum = divide_using_multiply_shift(sum, 5, DC_MULTIPLIER_1X2); + const __m128i row = _mm_set1_epi8((uint8_t)sum); + dc_store_64xh(&row, 32, dst, stride); +} + +void aom_dc_predictor_64x16_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + __m128i sum_above = dc_sum_64(above); + const __m128i sum_left = dc_sum_16(left); + sum_above = _mm_add_epi16(sum_above, sum_left); + + uint32_t sum = _mm_cvtsi128_si32(sum_above); + sum += 40; + sum = divide_using_multiply_shift(sum, 4, DC_MULTIPLIER_1X4); + const __m128i row = _mm_set1_epi8((uint8_t)sum); + dc_store_64xh(&row, 16, dst, stride); +} + +// ----------------------------------------------------------------------------- +// DC_TOP + +void aom_dc_top_predictor_4x8_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + (void)left; + __m128i sum_above = dc_sum_4(above); + const __m128i two = _mm_set1_epi16((int16_t)2); + sum_above = _mm_add_epi16(sum_above, two); + sum_above = _mm_srai_epi16(sum_above, 2); + sum_above = _mm_shufflelo_epi16(sum_above, 0); + sum_above = _mm_packus_epi16(sum_above, sum_above); + + const uint32_t pred = _mm_cvtsi128_si32(sum_above); + dc_store_4xh(pred, 8, dst, stride); +} + +void aom_dc_top_predictor_4x16_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + (void)left; + __m128i sum_above = dc_sum_4(above); + const __m128i two = _mm_set1_epi16((int16_t)2); + sum_above = _mm_add_epi16(sum_above, two); + sum_above = _mm_srai_epi16(sum_above, 2); + sum_above = _mm_shufflelo_epi16(sum_above, 0); + sum_above = _mm_packus_epi16(sum_above, sum_above); + + const uint32_t pred = _mm_cvtsi128_si32(sum_above); + dc_store_4xh(pred, 16, dst, stride); +} + +void aom_dc_top_predictor_8x4_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + (void)left; + __m128i sum_above = dc_sum_8(above); + const __m128i four = _mm_set1_epi16((uint16_t)4); + sum_above = _mm_add_epi16(sum_above, four); + sum_above = _mm_srai_epi16(sum_above, 3); + sum_above = _mm_unpacklo_epi8(sum_above, sum_above); + const __m128i row = _mm_shufflelo_epi16(sum_above, 0); + dc_store_8xh(&row, 4, dst, stride); +} + +void aom_dc_top_predictor_8x16_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + (void)left; + __m128i sum_above = dc_sum_8(above); + const __m128i four = _mm_set1_epi16((uint16_t)4); + sum_above = _mm_add_epi16(sum_above, four); + sum_above = _mm_srai_epi16(sum_above, 3); + sum_above = _mm_unpacklo_epi8(sum_above, sum_above); + const __m128i row = _mm_shufflelo_epi16(sum_above, 0); + dc_store_8xh(&row, 16, dst, stride); +} + +void aom_dc_top_predictor_8x32_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + (void)left; + __m128i sum_above = dc_sum_8(above); + const __m128i four = _mm_set1_epi16((uint16_t)4); + sum_above = _mm_add_epi16(sum_above, four); + sum_above = _mm_srai_epi16(sum_above, 3); + sum_above = _mm_unpacklo_epi8(sum_above, sum_above); + const __m128i row = _mm_shufflelo_epi16(sum_above, 0); + dc_store_8xh(&row, 32, dst, stride); +} + +void aom_dc_top_predictor_16x4_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + (void)left; + __m128i sum_above = dc_sum_16(above); + const __m128i eight = _mm_set1_epi16((uint16_t)8); + sum_above = _mm_add_epi16(sum_above, eight); + sum_above = _mm_srai_epi16(sum_above, 4); + sum_above = _mm_unpacklo_epi8(sum_above, sum_above); + sum_above = _mm_shufflelo_epi16(sum_above, 0); + const __m128i row = _mm_unpacklo_epi64(sum_above, sum_above); + dc_store_16xh(&row, 4, dst, stride); +} + +void aom_dc_top_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + (void)left; + __m128i sum_above = dc_sum_16(above); + const __m128i eight = _mm_set1_epi16((uint16_t)8); + sum_above = _mm_add_epi16(sum_above, eight); + sum_above = _mm_srai_epi16(sum_above, 4); + sum_above = _mm_unpacklo_epi8(sum_above, sum_above); + sum_above = _mm_shufflelo_epi16(sum_above, 0); + const __m128i row = _mm_unpacklo_epi64(sum_above, sum_above); + dc_store_16xh(&row, 8, dst, stride); +} + +void aom_dc_top_predictor_16x32_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + (void)left; + __m128i sum_above = dc_sum_16(above); + const __m128i eight = _mm_set1_epi16((uint16_t)8); + sum_above = _mm_add_epi16(sum_above, eight); + sum_above = _mm_srai_epi16(sum_above, 4); + sum_above = _mm_unpacklo_epi8(sum_above, sum_above); + sum_above = _mm_shufflelo_epi16(sum_above, 0); + const __m128i row = _mm_unpacklo_epi64(sum_above, sum_above); + dc_store_16xh(&row, 32, dst, stride); +} + +void aom_dc_top_predictor_16x64_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + (void)left; + __m128i sum_above = dc_sum_16(above); + const __m128i eight = _mm_set1_epi16((uint16_t)8); + sum_above = _mm_add_epi16(sum_above, eight); + sum_above = _mm_srai_epi16(sum_above, 4); + sum_above = _mm_unpacklo_epi8(sum_above, sum_above); + sum_above = _mm_shufflelo_epi16(sum_above, 0); + const __m128i row = _mm_unpacklo_epi64(sum_above, sum_above); + dc_store_16xh(&row, 64, dst, stride); +} + +void aom_dc_top_predictor_32x8_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + (void)left; + __m128i sum_above = dc_sum_32(above); + const __m128i sixteen = _mm_set1_epi16((uint16_t)16); + sum_above = _mm_add_epi16(sum_above, sixteen); + sum_above = _mm_srai_epi16(sum_above, 5); + sum_above = _mm_unpacklo_epi8(sum_above, sum_above); + sum_above = _mm_shufflelo_epi16(sum_above, 0); + const __m128i row = _mm_unpacklo_epi64(sum_above, sum_above); + dc_store_32xh(&row, 8, dst, stride); +} + +void aom_dc_top_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + (void)left; + __m128i sum_above = dc_sum_32(above); + const __m128i sixteen = _mm_set1_epi16((uint16_t)16); + sum_above = _mm_add_epi16(sum_above, sixteen); + sum_above = _mm_srai_epi16(sum_above, 5); + sum_above = _mm_unpacklo_epi8(sum_above, sum_above); + sum_above = _mm_shufflelo_epi16(sum_above, 0); + const __m128i row = _mm_unpacklo_epi64(sum_above, sum_above); + dc_store_32xh(&row, 16, dst, stride); +} + +void aom_dc_top_predictor_32x64_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + (void)left; + __m128i sum_above = dc_sum_32(above); + const __m128i sixteen = _mm_set1_epi16((uint16_t)16); + sum_above = _mm_add_epi16(sum_above, sixteen); + sum_above = _mm_srai_epi16(sum_above, 5); + sum_above = _mm_unpacklo_epi8(sum_above, sum_above); + sum_above = _mm_shufflelo_epi16(sum_above, 0); + const __m128i row = _mm_unpacklo_epi64(sum_above, sum_above); + dc_store_32xh(&row, 64, dst, stride); +} + +void aom_dc_top_predictor_64x64_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + (void)left; + __m128i sum_above = dc_sum_64(above); + const __m128i thirtytwo = _mm_set1_epi16((uint16_t)32); + sum_above = _mm_add_epi16(sum_above, thirtytwo); + sum_above = _mm_srai_epi16(sum_above, 6); + sum_above = _mm_unpacklo_epi8(sum_above, sum_above); + sum_above = _mm_shufflelo_epi16(sum_above, 0); + const __m128i row = _mm_unpacklo_epi64(sum_above, sum_above); + dc_store_64xh(&row, 64, dst, stride); +} + +void aom_dc_top_predictor_64x32_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + (void)left; + __m128i sum_above = dc_sum_64(above); + const __m128i thirtytwo = _mm_set1_epi16((uint16_t)32); + sum_above = _mm_add_epi16(sum_above, thirtytwo); + sum_above = _mm_srai_epi16(sum_above, 6); + sum_above = _mm_unpacklo_epi8(sum_above, sum_above); + sum_above = _mm_shufflelo_epi16(sum_above, 0); + const __m128i row = _mm_unpacklo_epi64(sum_above, sum_above); + dc_store_64xh(&row, 32, dst, stride); +} + +void aom_dc_top_predictor_64x16_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + (void)left; + __m128i sum_above = dc_sum_64(above); + const __m128i thirtytwo = _mm_set1_epi16((uint16_t)32); + sum_above = _mm_add_epi16(sum_above, thirtytwo); + sum_above = _mm_srai_epi16(sum_above, 6); + sum_above = _mm_unpacklo_epi8(sum_above, sum_above); + sum_above = _mm_shufflelo_epi16(sum_above, 0); + const __m128i row = _mm_unpacklo_epi64(sum_above, sum_above); + dc_store_64xh(&row, 16, dst, stride); +} + +// ----------------------------------------------------------------------------- +// DC_LEFT + +void aom_dc_left_predictor_4x8_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + (void)above; + __m128i sum_left = dc_sum_8(left); + const __m128i four = _mm_set1_epi16((uint16_t)4); + sum_left = _mm_add_epi16(sum_left, four); + sum_left = _mm_srai_epi16(sum_left, 3); + sum_left = _mm_shufflelo_epi16(sum_left, 0); + sum_left = _mm_packus_epi16(sum_left, sum_left); + + const uint32_t pred = _mm_cvtsi128_si32(sum_left); + dc_store_4xh(pred, 8, dst, stride); +} + +void aom_dc_left_predictor_4x16_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + (void)above; + __m128i sum_left = dc_sum_16(left); + const __m128i eight = _mm_set1_epi16((uint16_t)8); + sum_left = _mm_add_epi16(sum_left, eight); + sum_left = _mm_srai_epi16(sum_left, 4); + sum_left = _mm_shufflelo_epi16(sum_left, 0); + sum_left = _mm_packus_epi16(sum_left, sum_left); + + const uint32_t pred = _mm_cvtsi128_si32(sum_left); + dc_store_4xh(pred, 16, dst, stride); +} + +void aom_dc_left_predictor_8x4_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + (void)above; + __m128i sum_left = dc_sum_4(left); + const __m128i two = _mm_set1_epi16((uint16_t)2); + sum_left = _mm_add_epi16(sum_left, two); + sum_left = _mm_srai_epi16(sum_left, 2); + sum_left = _mm_unpacklo_epi8(sum_left, sum_left); + const __m128i row = _mm_shufflelo_epi16(sum_left, 0); + dc_store_8xh(&row, 4, dst, stride); +} + +void aom_dc_left_predictor_8x16_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + (void)above; + __m128i sum_left = dc_sum_16(left); + const __m128i eight = _mm_set1_epi16((uint16_t)8); + sum_left = _mm_add_epi16(sum_left, eight); + sum_left = _mm_srai_epi16(sum_left, 4); + sum_left = _mm_unpacklo_epi8(sum_left, sum_left); + const __m128i row = _mm_shufflelo_epi16(sum_left, 0); + dc_store_8xh(&row, 16, dst, stride); +} + +void aom_dc_left_predictor_8x32_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + (void)above; + __m128i sum_left = dc_sum_32(left); + const __m128i sixteen = _mm_set1_epi16((uint16_t)16); + sum_left = _mm_add_epi16(sum_left, sixteen); + sum_left = _mm_srai_epi16(sum_left, 5); + sum_left = _mm_unpacklo_epi8(sum_left, sum_left); + const __m128i row = _mm_shufflelo_epi16(sum_left, 0); + dc_store_8xh(&row, 32, dst, stride); +} + +void aom_dc_left_predictor_16x4_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + (void)above; + __m128i sum_left = dc_sum_4(left); + const __m128i two = _mm_set1_epi16((uint16_t)2); + sum_left = _mm_add_epi16(sum_left, two); + sum_left = _mm_srai_epi16(sum_left, 2); + sum_left = _mm_unpacklo_epi8(sum_left, sum_left); + sum_left = _mm_shufflelo_epi16(sum_left, 0); + const __m128i row = _mm_unpacklo_epi64(sum_left, sum_left); + dc_store_16xh(&row, 4, dst, stride); +} + +void aom_dc_left_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + (void)above; + __m128i sum_left = dc_sum_8(left); + const __m128i four = _mm_set1_epi16((uint16_t)4); + sum_left = _mm_add_epi16(sum_left, four); + sum_left = _mm_srai_epi16(sum_left, 3); + sum_left = _mm_unpacklo_epi8(sum_left, sum_left); + sum_left = _mm_shufflelo_epi16(sum_left, 0); + const __m128i row = _mm_unpacklo_epi64(sum_left, sum_left); + dc_store_16xh(&row, 8, dst, stride); +} + +void aom_dc_left_predictor_16x32_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + (void)above; + __m128i sum_left = dc_sum_32(left); + const __m128i sixteen = _mm_set1_epi16((uint16_t)16); + sum_left = _mm_add_epi16(sum_left, sixteen); + sum_left = _mm_srai_epi16(sum_left, 5); + sum_left = _mm_unpacklo_epi8(sum_left, sum_left); + sum_left = _mm_shufflelo_epi16(sum_left, 0); + const __m128i row = _mm_unpacklo_epi64(sum_left, sum_left); + dc_store_16xh(&row, 32, dst, stride); +} + +void aom_dc_left_predictor_16x64_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + (void)above; + __m128i sum_left = dc_sum_64(left); + const __m128i thirtytwo = _mm_set1_epi16((uint16_t)32); + sum_left = _mm_add_epi16(sum_left, thirtytwo); + sum_left = _mm_srai_epi16(sum_left, 6); + sum_left = _mm_unpacklo_epi8(sum_left, sum_left); + sum_left = _mm_shufflelo_epi16(sum_left, 0); + const __m128i row = _mm_unpacklo_epi64(sum_left, sum_left); + dc_store_16xh(&row, 64, dst, stride); +} + +void aom_dc_left_predictor_32x8_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + (void)above; + __m128i sum_left = dc_sum_8(left); + const __m128i four = _mm_set1_epi16((uint16_t)4); + sum_left = _mm_add_epi16(sum_left, four); + sum_left = _mm_srai_epi16(sum_left, 3); + sum_left = _mm_unpacklo_epi8(sum_left, sum_left); + sum_left = _mm_shufflelo_epi16(sum_left, 0); + const __m128i row = _mm_unpacklo_epi64(sum_left, sum_left); + dc_store_32xh(&row, 8, dst, stride); +} + +void aom_dc_left_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + (void)above; + __m128i sum_left = dc_sum_16(left); + const __m128i eight = _mm_set1_epi16((uint16_t)8); + sum_left = _mm_add_epi16(sum_left, eight); + sum_left = _mm_srai_epi16(sum_left, 4); + sum_left = _mm_unpacklo_epi8(sum_left, sum_left); + sum_left = _mm_shufflelo_epi16(sum_left, 0); + const __m128i row = _mm_unpacklo_epi64(sum_left, sum_left); + dc_store_32xh(&row, 16, dst, stride); +} + +void aom_dc_left_predictor_32x64_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + (void)above; + __m128i sum_left = dc_sum_64(left); + const __m128i thirtytwo = _mm_set1_epi16((uint16_t)32); + sum_left = _mm_add_epi16(sum_left, thirtytwo); + sum_left = _mm_srai_epi16(sum_left, 6); + sum_left = _mm_unpacklo_epi8(sum_left, sum_left); + sum_left = _mm_shufflelo_epi16(sum_left, 0); + const __m128i row = _mm_unpacklo_epi64(sum_left, sum_left); + dc_store_32xh(&row, 64, dst, stride); +} + +void aom_dc_left_predictor_64x64_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + (void)above; + __m128i sum_left = dc_sum_64(left); + const __m128i thirtytwo = _mm_set1_epi16((uint16_t)32); + sum_left = _mm_add_epi16(sum_left, thirtytwo); + sum_left = _mm_srai_epi16(sum_left, 6); + sum_left = _mm_unpacklo_epi8(sum_left, sum_left); + sum_left = _mm_shufflelo_epi16(sum_left, 0); + const __m128i row = _mm_unpacklo_epi64(sum_left, sum_left); + dc_store_64xh(&row, 64, dst, stride); +} + +void aom_dc_left_predictor_64x32_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + (void)above; + __m128i sum_left = dc_sum_32(left); + const __m128i sixteen = _mm_set1_epi16((uint16_t)16); + sum_left = _mm_add_epi16(sum_left, sixteen); + sum_left = _mm_srai_epi16(sum_left, 5); + sum_left = _mm_unpacklo_epi8(sum_left, sum_left); + sum_left = _mm_shufflelo_epi16(sum_left, 0); + const __m128i row = _mm_unpacklo_epi64(sum_left, sum_left); + dc_store_64xh(&row, 32, dst, stride); +} + +void aom_dc_left_predictor_64x16_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + (void)above; + __m128i sum_left = dc_sum_16(left); + const __m128i eight = _mm_set1_epi16((uint16_t)8); + sum_left = _mm_add_epi16(sum_left, eight); + sum_left = _mm_srai_epi16(sum_left, 4); + sum_left = _mm_unpacklo_epi8(sum_left, sum_left); + sum_left = _mm_shufflelo_epi16(sum_left, 0); + const __m128i row = _mm_unpacklo_epi64(sum_left, sum_left); + dc_store_64xh(&row, 16, dst, stride); +} + +// ----------------------------------------------------------------------------- +// DC_128 + +void aom_dc_128_predictor_4x8_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + (void)above; + (void)left; + const uint32_t pred = 0x80808080; + dc_store_4xh(pred, 8, dst, stride); +} + +void aom_dc_128_predictor_4x16_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + (void)above; + (void)left; + const uint32_t pred = 0x80808080; + dc_store_4xh(pred, 16, dst, stride); +} + +void aom_dc_128_predictor_8x4_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + (void)above; + (void)left; + const __m128i row = _mm_set1_epi8((uint8_t)128); + dc_store_8xh(&row, 4, dst, stride); +} + +void aom_dc_128_predictor_8x16_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + (void)above; + (void)left; + const __m128i row = _mm_set1_epi8((uint8_t)128); + dc_store_8xh(&row, 16, dst, stride); +} + +void aom_dc_128_predictor_8x32_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + (void)above; + (void)left; + const __m128i row = _mm_set1_epi8((uint8_t)128); + dc_store_8xh(&row, 32, dst, stride); +} + +void aom_dc_128_predictor_16x4_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + (void)above; + (void)left; + const __m128i row = _mm_set1_epi8((uint8_t)128); + dc_store_16xh(&row, 4, dst, stride); +} + +void aom_dc_128_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + (void)above; + (void)left; + const __m128i row = _mm_set1_epi8((uint8_t)128); + dc_store_16xh(&row, 8, dst, stride); +} + +void aom_dc_128_predictor_16x32_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + (void)above; + (void)left; + const __m128i row = _mm_set1_epi8((uint8_t)128); + dc_store_16xh(&row, 32, dst, stride); +} + +void aom_dc_128_predictor_16x64_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + (void)above; + (void)left; + const __m128i row = _mm_set1_epi8((uint8_t)128); + dc_store_16xh(&row, 64, dst, stride); +} + +void aom_dc_128_predictor_32x8_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + (void)above; + (void)left; + const __m128i row = _mm_set1_epi8((uint8_t)128); + dc_store_32xh(&row, 8, dst, stride); +} + +void aom_dc_128_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + (void)above; + (void)left; + const __m128i row = _mm_set1_epi8((uint8_t)128); + dc_store_32xh(&row, 16, dst, stride); +} + +void aom_dc_128_predictor_32x64_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + (void)above; + (void)left; + const __m128i row = _mm_set1_epi8((uint8_t)128); + dc_store_32xh(&row, 64, dst, stride); +} + +void aom_dc_128_predictor_64x64_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + (void)above; + (void)left; + const __m128i row = _mm_set1_epi8((uint8_t)128); + dc_store_64xh(&row, 64, dst, stride); +} + +void aom_dc_128_predictor_64x32_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + (void)above; + (void)left; + const __m128i row = _mm_set1_epi8((uint8_t)128); + dc_store_64xh(&row, 32, dst, stride); +} + +void aom_dc_128_predictor_64x16_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + (void)above; + (void)left; + const __m128i row = _mm_set1_epi8((uint8_t)128); + dc_store_64xh(&row, 16, dst, stride); +} + +// ----------------------------------------------------------------------------- +// V_PRED + +void aom_v_predictor_4x8_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + const uint32_t pred = *(uint32_t *)above; + (void)left; + dc_store_4xh(pred, 8, dst, stride); +} + +void aom_v_predictor_4x16_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + const uint32_t pred = *(uint32_t *)above; + (void)left; + dc_store_4xh(pred, 16, dst, stride); +} + +void aom_v_predictor_8x4_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + const __m128i row = _mm_loadl_epi64((__m128i const *)above); + (void)left; + dc_store_8xh(&row, 4, dst, stride); +} + +void aom_v_predictor_8x16_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + const __m128i row = _mm_loadl_epi64((__m128i const *)above); + (void)left; + dc_store_8xh(&row, 16, dst, stride); +} + +void aom_v_predictor_8x32_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + const __m128i row = _mm_loadl_epi64((__m128i const *)above); + (void)left; + dc_store_8xh(&row, 32, dst, stride); +} + +void aom_v_predictor_16x4_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + const __m128i row = _mm_load_si128((__m128i const *)above); + (void)left; + dc_store_16xh(&row, 4, dst, stride); +} + +void aom_v_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + const __m128i row = _mm_load_si128((__m128i const *)above); + (void)left; + dc_store_16xh(&row, 8, dst, stride); +} + +void aom_v_predictor_16x32_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + const __m128i row = _mm_load_si128((__m128i const *)above); + (void)left; + dc_store_16xh(&row, 32, dst, stride); +} + +void aom_v_predictor_16x64_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + const __m128i row = _mm_load_si128((__m128i const *)above); + (void)left; + dc_store_16xh(&row, 64, dst, stride); +} + +static INLINE void v_predictor_32xh(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, int height) { + const __m128i row0 = _mm_load_si128((__m128i const *)above); + const __m128i row1 = _mm_load_si128((__m128i const *)(above + 16)); + for (int i = 0; i < height; ++i) { + _mm_store_si128((__m128i *)dst, row0); + _mm_store_si128((__m128i *)(dst + 16), row1); + dst += stride; + } +} + +void aom_v_predictor_32x8_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + (void)left; + v_predictor_32xh(dst, stride, above, 8); +} + +void aom_v_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + (void)left; + v_predictor_32xh(dst, stride, above, 16); +} + +void aom_v_predictor_32x64_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + (void)left; + v_predictor_32xh(dst, stride, above, 64); +} + +static INLINE void v_predictor_64xh(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, int height) { + const __m128i row0 = _mm_load_si128((__m128i const *)above); + const __m128i row1 = _mm_load_si128((__m128i const *)(above + 16)); + const __m128i row2 = _mm_load_si128((__m128i const *)(above + 32)); + const __m128i row3 = _mm_load_si128((__m128i const *)(above + 48)); + for (int i = 0; i < height; ++i) { + _mm_store_si128((__m128i *)dst, row0); + _mm_store_si128((__m128i *)(dst + 16), row1); + _mm_store_si128((__m128i *)(dst + 32), row2); + _mm_store_si128((__m128i *)(dst + 48), row3); + dst += stride; + } +} + +void aom_v_predictor_64x64_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + (void)left; + v_predictor_64xh(dst, stride, above, 64); +} + +void aom_v_predictor_64x32_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + (void)left; + v_predictor_64xh(dst, stride, above, 32); +} + +void aom_v_predictor_64x16_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + (void)left; + v_predictor_64xh(dst, stride, above, 16); +} + +// ----------------------------------------------------------------------------- +// H_PRED + +void aom_h_predictor_4x8_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + (void)above; + __m128i left_col = _mm_loadl_epi64((__m128i const *)left); + left_col = _mm_unpacklo_epi8(left_col, left_col); + __m128i row0 = _mm_shufflelo_epi16(left_col, 0); + __m128i row1 = _mm_shufflelo_epi16(left_col, 0x55); + __m128i row2 = _mm_shufflelo_epi16(left_col, 0xaa); + __m128i row3 = _mm_shufflelo_epi16(left_col, 0xff); + *(uint32_t *)dst = _mm_cvtsi128_si32(row0); + dst += stride; + *(uint32_t *)dst = _mm_cvtsi128_si32(row1); + dst += stride; + *(uint32_t *)dst = _mm_cvtsi128_si32(row2); + dst += stride; + *(uint32_t *)dst = _mm_cvtsi128_si32(row3); + dst += stride; + left_col = _mm_unpackhi_epi64(left_col, left_col); + row0 = _mm_shufflelo_epi16(left_col, 0); + row1 = _mm_shufflelo_epi16(left_col, 0x55); + row2 = _mm_shufflelo_epi16(left_col, 0xaa); + row3 = _mm_shufflelo_epi16(left_col, 0xff); + *(uint32_t *)dst = _mm_cvtsi128_si32(row0); + dst += stride; + *(uint32_t *)dst = _mm_cvtsi128_si32(row1); + dst += stride; + *(uint32_t *)dst = _mm_cvtsi128_si32(row2); + dst += stride; + *(uint32_t *)dst = _mm_cvtsi128_si32(row3); +} + +void aom_h_predictor_4x16_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + (void)above; + const __m128i left_col = _mm_load_si128((__m128i const *)left); + __m128i left_col_low = _mm_unpacklo_epi8(left_col, left_col); + __m128i left_col_high = _mm_unpackhi_epi8(left_col, left_col); + + __m128i row0 = _mm_shufflelo_epi16(left_col_low, 0); + __m128i row1 = _mm_shufflelo_epi16(left_col_low, 0x55); + __m128i row2 = _mm_shufflelo_epi16(left_col_low, 0xaa); + __m128i row3 = _mm_shufflelo_epi16(left_col_low, 0xff); + *(uint32_t *)dst = _mm_cvtsi128_si32(row0); + dst += stride; + *(uint32_t *)dst = _mm_cvtsi128_si32(row1); + dst += stride; + *(uint32_t *)dst = _mm_cvtsi128_si32(row2); + dst += stride; + *(uint32_t *)dst = _mm_cvtsi128_si32(row3); + dst += stride; + + left_col_low = _mm_unpackhi_epi64(left_col_low, left_col_low); + row0 = _mm_shufflelo_epi16(left_col_low, 0); + row1 = _mm_shufflelo_epi16(left_col_low, 0x55); + row2 = _mm_shufflelo_epi16(left_col_low, 0xaa); + row3 = _mm_shufflelo_epi16(left_col_low, 0xff); + *(uint32_t *)dst = _mm_cvtsi128_si32(row0); + dst += stride; + *(uint32_t *)dst = _mm_cvtsi128_si32(row1); + dst += stride; + *(uint32_t *)dst = _mm_cvtsi128_si32(row2); + dst += stride; + *(uint32_t *)dst = _mm_cvtsi128_si32(row3); + dst += stride; + + row0 = _mm_shufflelo_epi16(left_col_high, 0); + row1 = _mm_shufflelo_epi16(left_col_high, 0x55); + row2 = _mm_shufflelo_epi16(left_col_high, 0xaa); + row3 = _mm_shufflelo_epi16(left_col_high, 0xff); + *(uint32_t *)dst = _mm_cvtsi128_si32(row0); + dst += stride; + *(uint32_t *)dst = _mm_cvtsi128_si32(row1); + dst += stride; + *(uint32_t *)dst = _mm_cvtsi128_si32(row2); + dst += stride; + *(uint32_t *)dst = _mm_cvtsi128_si32(row3); + dst += stride; + + left_col_high = _mm_unpackhi_epi64(left_col_high, left_col_high); + row0 = _mm_shufflelo_epi16(left_col_high, 0); + row1 = _mm_shufflelo_epi16(left_col_high, 0x55); + row2 = _mm_shufflelo_epi16(left_col_high, 0xaa); + row3 = _mm_shufflelo_epi16(left_col_high, 0xff); + *(uint32_t *)dst = _mm_cvtsi128_si32(row0); + dst += stride; + *(uint32_t *)dst = _mm_cvtsi128_si32(row1); + dst += stride; + *(uint32_t *)dst = _mm_cvtsi128_si32(row2); + dst += stride; + *(uint32_t *)dst = _mm_cvtsi128_si32(row3); +} + +void aom_h_predictor_8x4_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + (void)above; + __m128i left_col = _mm_loadl_epi64((__m128i const *)left); + left_col = _mm_unpacklo_epi8(left_col, left_col); + __m128i row0 = _mm_shufflelo_epi16(left_col, 0); + __m128i row1 = _mm_shufflelo_epi16(left_col, 0x55); + __m128i row2 = _mm_shufflelo_epi16(left_col, 0xaa); + __m128i row3 = _mm_shufflelo_epi16(left_col, 0xff); + _mm_storel_epi64((__m128i *)dst, row0); + dst += stride; + _mm_storel_epi64((__m128i *)dst, row1); + dst += stride; + _mm_storel_epi64((__m128i *)dst, row2); + dst += stride; + _mm_storel_epi64((__m128i *)dst, row3); +} + +static INLINE void h_predictor_8x16xc(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left, + int count) { + (void)above; + for (int i = 0; i < count; ++i) { + const __m128i left_col = _mm_load_si128((__m128i const *)left); + __m128i left_col_low = _mm_unpacklo_epi8(left_col, left_col); + __m128i left_col_high = _mm_unpackhi_epi8(left_col, left_col); + + __m128i row0 = _mm_shufflelo_epi16(left_col_low, 0); + __m128i row1 = _mm_shufflelo_epi16(left_col_low, 0x55); + __m128i row2 = _mm_shufflelo_epi16(left_col_low, 0xaa); + __m128i row3 = _mm_shufflelo_epi16(left_col_low, 0xff); + _mm_storel_epi64((__m128i *)dst, row0); + dst += stride; + _mm_storel_epi64((__m128i *)dst, row1); + dst += stride; + _mm_storel_epi64((__m128i *)dst, row2); + dst += stride; + _mm_storel_epi64((__m128i *)dst, row3); + dst += stride; + + left_col_low = _mm_unpackhi_epi64(left_col_low, left_col_low); + row0 = _mm_shufflelo_epi16(left_col_low, 0); + row1 = _mm_shufflelo_epi16(left_col_low, 0x55); + row2 = _mm_shufflelo_epi16(left_col_low, 0xaa); + row3 = _mm_shufflelo_epi16(left_col_low, 0xff); + _mm_storel_epi64((__m128i *)dst, row0); + dst += stride; + _mm_storel_epi64((__m128i *)dst, row1); + dst += stride; + _mm_storel_epi64((__m128i *)dst, row2); + dst += stride; + _mm_storel_epi64((__m128i *)dst, row3); + dst += stride; + + row0 = _mm_shufflelo_epi16(left_col_high, 0); + row1 = _mm_shufflelo_epi16(left_col_high, 0x55); + row2 = _mm_shufflelo_epi16(left_col_high, 0xaa); + row3 = _mm_shufflelo_epi16(left_col_high, 0xff); + _mm_storel_epi64((__m128i *)dst, row0); + dst += stride; + _mm_storel_epi64((__m128i *)dst, row1); + dst += stride; + _mm_storel_epi64((__m128i *)dst, row2); + dst += stride; + _mm_storel_epi64((__m128i *)dst, row3); + dst += stride; + + left_col_high = _mm_unpackhi_epi64(left_col_high, left_col_high); + row0 = _mm_shufflelo_epi16(left_col_high, 0); + row1 = _mm_shufflelo_epi16(left_col_high, 0x55); + row2 = _mm_shufflelo_epi16(left_col_high, 0xaa); + row3 = _mm_shufflelo_epi16(left_col_high, 0xff); + _mm_storel_epi64((__m128i *)dst, row0); + dst += stride; + _mm_storel_epi64((__m128i *)dst, row1); + dst += stride; + _mm_storel_epi64((__m128i *)dst, row2); + dst += stride; + _mm_storel_epi64((__m128i *)dst, row3); + dst += stride; + left += 16; + } +} + +void aom_h_predictor_8x16_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + h_predictor_8x16xc(dst, stride, above, left, 1); +} + +void aom_h_predictor_8x32_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + h_predictor_8x16xc(dst, stride, above, left, 2); +} + +static INLINE void h_pred_store_16xh(const __m128i *row, int h, uint8_t *dst, + ptrdiff_t stride) { + int i; + for (i = 0; i < h; ++i) { + _mm_store_si128((__m128i *)dst, row[i]); + dst += stride; + } +} + +static INLINE void repeat_low_4pixels(const __m128i *x, __m128i *row) { + const __m128i u0 = _mm_shufflelo_epi16(*x, 0); + const __m128i u1 = _mm_shufflelo_epi16(*x, 0x55); + const __m128i u2 = _mm_shufflelo_epi16(*x, 0xaa); + const __m128i u3 = _mm_shufflelo_epi16(*x, 0xff); + + row[0] = _mm_unpacklo_epi64(u0, u0); + row[1] = _mm_unpacklo_epi64(u1, u1); + row[2] = _mm_unpacklo_epi64(u2, u2); + row[3] = _mm_unpacklo_epi64(u3, u3); +} + +static INLINE void repeat_high_4pixels(const __m128i *x, __m128i *row) { + const __m128i u0 = _mm_shufflehi_epi16(*x, 0); + const __m128i u1 = _mm_shufflehi_epi16(*x, 0x55); + const __m128i u2 = _mm_shufflehi_epi16(*x, 0xaa); + const __m128i u3 = _mm_shufflehi_epi16(*x, 0xff); + + row[0] = _mm_unpackhi_epi64(u0, u0); + row[1] = _mm_unpackhi_epi64(u1, u1); + row[2] = _mm_unpackhi_epi64(u2, u2); + row[3] = _mm_unpackhi_epi64(u3, u3); +} + +// Process 16x8, first 4 rows +// Use first 8 bytes of left register: xxxxxxxx33221100 +static INLINE void h_prediction_16x8_1(const __m128i *left, uint8_t *dst, + ptrdiff_t stride) { + __m128i row[4]; + repeat_low_4pixels(left, row); + h_pred_store_16xh(row, 4, dst, stride); +} + +// Process 16x8, second 4 rows +// Use second 8 bytes of left register: 77665544xxxxxxxx +static INLINE void h_prediction_16x8_2(const __m128i *left, uint8_t *dst, + ptrdiff_t stride) { + __m128i row[4]; + repeat_high_4pixels(left, row); + h_pred_store_16xh(row, 4, dst, stride); +} + +void aom_h_predictor_16x4_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + (void)above; + const __m128i left_col = _mm_loadl_epi64((const __m128i *)left); + const __m128i left_col_8p = _mm_unpacklo_epi8(left_col, left_col); + h_prediction_16x8_1(&left_col_8p, dst, stride); +} + +void aom_h_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + (void)above; + const __m128i left_col = _mm_loadl_epi64((const __m128i *)left); + const __m128i left_col_8p = _mm_unpacklo_epi8(left_col, left_col); + h_prediction_16x8_1(&left_col_8p, dst, stride); + dst += stride << 2; + h_prediction_16x8_2(&left_col_8p, dst, stride); +} + +static INLINE void h_predictor_16xh(uint8_t *dst, ptrdiff_t stride, + const uint8_t *left, int count) { + int i = 0; + do { + const __m128i left_col = _mm_load_si128((const __m128i *)left); + const __m128i left_col_8p_lo = _mm_unpacklo_epi8(left_col, left_col); + h_prediction_16x8_1(&left_col_8p_lo, dst, stride); + dst += stride << 2; + h_prediction_16x8_2(&left_col_8p_lo, dst, stride); + dst += stride << 2; + + const __m128i left_col_8p_hi = _mm_unpackhi_epi8(left_col, left_col); + h_prediction_16x8_1(&left_col_8p_hi, dst, stride); + dst += stride << 2; + h_prediction_16x8_2(&left_col_8p_hi, dst, stride); + dst += stride << 2; + + left += 16; + i++; + } while (i < count); +} + +void aom_h_predictor_16x32_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + (void)above; + h_predictor_16xh(dst, stride, left, 2); +} + +void aom_h_predictor_16x64_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + (void)above; + h_predictor_16xh(dst, stride, left, 4); +} + +static INLINE void h_pred_store_32xh(const __m128i *row, int h, uint8_t *dst, + ptrdiff_t stride) { + int i; + for (i = 0; i < h; ++i) { + _mm_store_si128((__m128i *)dst, row[i]); + _mm_store_si128((__m128i *)(dst + 16), row[i]); + dst += stride; + } +} + +// Process 32x8, first 4 rows +// Use first 8 bytes of left register: xxxxxxxx33221100 +static INLINE void h_prediction_32x8_1(const __m128i *left, uint8_t *dst, + ptrdiff_t stride) { + __m128i row[4]; + repeat_low_4pixels(left, row); + h_pred_store_32xh(row, 4, dst, stride); +} + +// Process 32x8, second 4 rows +// Use second 8 bytes of left register: 77665544xxxxxxxx +static INLINE void h_prediction_32x8_2(const __m128i *left, uint8_t *dst, + ptrdiff_t stride) { + __m128i row[4]; + repeat_high_4pixels(left, row); + h_pred_store_32xh(row, 4, dst, stride); +} + +void aom_h_predictor_32x8_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + __m128i left_col, left_col_8p; + (void)above; + + left_col = _mm_load_si128((const __m128i *)left); + + left_col_8p = _mm_unpacklo_epi8(left_col, left_col); + h_prediction_32x8_1(&left_col_8p, dst, stride); + dst += stride << 2; + h_prediction_32x8_2(&left_col_8p, dst, stride); +} + +void aom_h_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + __m128i left_col, left_col_8p; + (void)above; + + left_col = _mm_load_si128((const __m128i *)left); + + left_col_8p = _mm_unpacklo_epi8(left_col, left_col); + h_prediction_32x8_1(&left_col_8p, dst, stride); + dst += stride << 2; + h_prediction_32x8_2(&left_col_8p, dst, stride); + dst += stride << 2; + + left_col_8p = _mm_unpackhi_epi8(left_col, left_col); + h_prediction_32x8_1(&left_col_8p, dst, stride); + dst += stride << 2; + h_prediction_32x8_2(&left_col_8p, dst, stride); +} + +static INLINE void h_predictor_32xh(uint8_t *dst, ptrdiff_t stride, + const uint8_t *left, int height) { + int i = height >> 2; + do { + __m128i left4 = _mm_cvtsi32_si128(((uint32_t *)left)[0]); + left4 = _mm_unpacklo_epi8(left4, left4); + left4 = _mm_unpacklo_epi8(left4, left4); + const __m128i r0 = _mm_shuffle_epi32(left4, 0x0); + const __m128i r1 = _mm_shuffle_epi32(left4, 0x55); + _mm_store_si128((__m128i *)dst, r0); + _mm_store_si128((__m128i *)(dst + 16), r0); + _mm_store_si128((__m128i *)(dst + stride), r1); + _mm_store_si128((__m128i *)(dst + stride + 16), r1); + const __m128i r2 = _mm_shuffle_epi32(left4, 0xaa); + const __m128i r3 = _mm_shuffle_epi32(left4, 0xff); + _mm_store_si128((__m128i *)(dst + stride * 2), r2); + _mm_store_si128((__m128i *)(dst + stride * 2 + 16), r2); + _mm_store_si128((__m128i *)(dst + stride * 3), r3); + _mm_store_si128((__m128i *)(dst + stride * 3 + 16), r3); + left += 4; + dst += stride * 4; + } while (--i); +} + +void aom_h_predictor_32x64_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + (void)above; + h_predictor_32xh(dst, stride, left, 64); +} + +static INLINE void h_predictor_64xh(uint8_t *dst, ptrdiff_t stride, + const uint8_t *left, int height) { + int i = height >> 2; + do { + __m128i left4 = _mm_cvtsi32_si128(((uint32_t *)left)[0]); + left4 = _mm_unpacklo_epi8(left4, left4); + left4 = _mm_unpacklo_epi8(left4, left4); + const __m128i r0 = _mm_shuffle_epi32(left4, 0x0); + const __m128i r1 = _mm_shuffle_epi32(left4, 0x55); + _mm_store_si128((__m128i *)dst, r0); + _mm_store_si128((__m128i *)(dst + 16), r0); + _mm_store_si128((__m128i *)(dst + 32), r0); + _mm_store_si128((__m128i *)(dst + 48), r0); + _mm_store_si128((__m128i *)(dst + stride), r1); + _mm_store_si128((__m128i *)(dst + stride + 16), r1); + _mm_store_si128((__m128i *)(dst + stride + 32), r1); + _mm_store_si128((__m128i *)(dst + stride + 48), r1); + const __m128i r2 = _mm_shuffle_epi32(left4, 0xaa); + const __m128i r3 = _mm_shuffle_epi32(left4, 0xff); + _mm_store_si128((__m128i *)(dst + stride * 2), r2); + _mm_store_si128((__m128i *)(dst + stride * 2 + 16), r2); + _mm_store_si128((__m128i *)(dst + stride * 2 + 32), r2); + _mm_store_si128((__m128i *)(dst + stride * 2 + 48), r2); + _mm_store_si128((__m128i *)(dst + stride * 3), r3); + _mm_store_si128((__m128i *)(dst + stride * 3 + 16), r3); + _mm_store_si128((__m128i *)(dst + stride * 3 + 32), r3); + _mm_store_si128((__m128i *)(dst + stride * 3 + 48), r3); + left += 4; + dst += stride * 4; + } while (--i); +} + +void aom_h_predictor_64x64_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + (void)above; + h_predictor_64xh(dst, stride, left, 64); +} + +void aom_h_predictor_64x32_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + (void)above; + h_predictor_64xh(dst, stride, left, 32); +} + +void aom_h_predictor_64x16_sse2(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + (void)above; + h_predictor_64xh(dst, stride, left, 16); +} diff --git a/third_party/aom/aom_dsp/x86/intrapred_sse2_asm.asm b/third_party/aom/aom_dsp/x86/intrapred_sse2_asm.asm new file mode 100644 index 0000000000..9aece27beb --- /dev/null +++ b/third_party/aom/aom_dsp/x86/intrapred_sse2_asm.asm @@ -0,0 +1,625 @@ +; +; Copyright (c) 2016, Alliance for Open Media. All rights reserved +; +; This source code is subject to the terms of the BSD 2 Clause License and +; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License +; was not distributed with this source code in the LICENSE file, you can +; obtain it at www.aomedia.org/license/software. If the Alliance for Open +; Media Patent License 1.0 was not distributed with this source code in the +; PATENTS file, you can obtain it at www.aomedia.org/license/patent. +; + +; + +%include "third_party/x86inc/x86inc.asm" + +SECTION_RODATA +pb_1: times 16 db 1 +pw_4: times 8 dw 4 +pw_8: times 8 dw 8 +pw_16: times 8 dw 16 +pw_32: times 8 dw 32 +dc_128: times 16 db 128 +pw2_4: times 8 dw 2 +pw2_8: times 8 dw 4 +pw2_16: times 8 dw 8 +pw2_32: times 8 dw 16 + +SECTION .text + +; ------------------------------------------ +; input: x, y, z, result +; +; trick from pascal +; (x+2y+z+2)>>2 can be calculated as: +; result = avg(x,z) +; result -= xor(x,z) & 1 +; result = avg(result,y) +; ------------------------------------------ +%macro X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 4 + pavgb %4, %1, %3 + pxor %3, %1 + pand %3, [GLOBAL(pb_1)] + psubb %4, %3 + pavgb %4, %2 +%endmacro + +INIT_XMM sse2 +cglobal dc_predictor_4x4, 4, 5, 3, dst, stride, above, left, goffset + GET_GOT goffsetq + + movd m2, [leftq] + movd m0, [aboveq] + pxor m1, m1 + punpckldq m0, m2 + psadbw m0, m1 + paddw m0, [GLOBAL(pw_4)] + psraw m0, 3 + pshuflw m0, m0, 0x0 + packuswb m0, m0 + movd [dstq ], m0 + movd [dstq+strideq], m0 + lea dstq, [dstq+strideq*2] + movd [dstq ], m0 + movd [dstq+strideq], m0 + + RESTORE_GOT + RET + +INIT_XMM sse2 +cglobal dc_left_predictor_4x4, 2, 5, 2, dst, stride, above, left, goffset + movifnidn leftq, leftmp + GET_GOT goffsetq + + pxor m1, m1 + movd m0, [leftq] + psadbw m0, m1 + paddw m0, [GLOBAL(pw2_4)] + psraw m0, 2 + pshuflw m0, m0, 0x0 + packuswb m0, m0 + movd [dstq ], m0 + movd [dstq+strideq], m0 + lea dstq, [dstq+strideq*2] + movd [dstq ], m0 + movd [dstq+strideq], m0 + + RESTORE_GOT + RET + +INIT_XMM sse2 +cglobal dc_top_predictor_4x4, 3, 5, 2, dst, stride, above, left, goffset + GET_GOT goffsetq + + pxor m1, m1 + movd m0, [aboveq] + psadbw m0, m1 + paddw m0, [GLOBAL(pw2_4)] + psraw m0, 2 + pshuflw m0, m0, 0x0 + packuswb m0, m0 + movd [dstq ], m0 + movd [dstq+strideq], m0 + lea dstq, [dstq+strideq*2] + movd [dstq ], m0 + movd [dstq+strideq], m0 + + RESTORE_GOT + RET + +INIT_XMM sse2 +cglobal dc_predictor_8x8, 4, 5, 3, dst, stride, above, left, goffset + GET_GOT goffsetq + + pxor m1, m1 + movq m0, [aboveq] + movq m2, [leftq] + DEFINE_ARGS dst, stride, stride3 + lea stride3q, [strideq*3] + psadbw m0, m1 + psadbw m2, m1 + paddw m0, m2 + paddw m0, [GLOBAL(pw_8)] + psraw m0, 4 + punpcklbw m0, m0 + pshuflw m0, m0, 0x0 + movq [dstq ], m0 + movq [dstq+strideq ], m0 + movq [dstq+strideq*2], m0 + movq [dstq+stride3q ], m0 + lea dstq, [dstq+strideq*4] + movq [dstq ], m0 + movq [dstq+strideq ], m0 + movq [dstq+strideq*2], m0 + movq [dstq+stride3q ], m0 + + RESTORE_GOT + RET + +INIT_XMM sse2 +cglobal dc_top_predictor_8x8, 3, 5, 2, dst, stride, above, left, goffset + GET_GOT goffsetq + + pxor m1, m1 + movq m0, [aboveq] + DEFINE_ARGS dst, stride, stride3 + lea stride3q, [strideq*3] + psadbw m0, m1 + paddw m0, [GLOBAL(pw2_8)] + psraw m0, 3 + punpcklbw m0, m0 + pshuflw m0, m0, 0x0 + movq [dstq ], m0 + movq [dstq+strideq ], m0 + movq [dstq+strideq*2], m0 + movq [dstq+stride3q ], m0 + lea dstq, [dstq+strideq*4] + movq [dstq ], m0 + movq [dstq+strideq ], m0 + movq [dstq+strideq*2], m0 + movq [dstq+stride3q ], m0 + + RESTORE_GOT + RET + +INIT_XMM sse2 +cglobal dc_left_predictor_8x8, 2, 5, 2, dst, stride, above, left, goffset + movifnidn leftq, leftmp + GET_GOT goffsetq + + pxor m1, m1 + movq m0, [leftq] + DEFINE_ARGS dst, stride, stride3 + lea stride3q, [strideq*3] + psadbw m0, m1 + paddw m0, [GLOBAL(pw2_8)] + psraw m0, 3 + punpcklbw m0, m0 + pshuflw m0, m0, 0x0 + movq [dstq ], m0 + movq [dstq+strideq ], m0 + movq [dstq+strideq*2], m0 + movq [dstq+stride3q ], m0 + lea dstq, [dstq+strideq*4] + movq [dstq ], m0 + movq [dstq+strideq ], m0 + movq [dstq+strideq*2], m0 + movq [dstq+stride3q ], m0 + + RESTORE_GOT + RET + +INIT_XMM sse2 +cglobal dc_128_predictor_4x4, 2, 5, 1, dst, stride, above, left, goffset + GET_GOT goffsetq + + DEFINE_ARGS dst, stride, stride3 + lea stride3q, [strideq*3] + movd m0, [GLOBAL(dc_128)] + movd [dstq ], m0 + movd [dstq+strideq ], m0 + movd [dstq+strideq*2], m0 + movd [dstq+stride3q ], m0 + RESTORE_GOT + RET + +INIT_XMM sse2 +cglobal dc_128_predictor_8x8, 2, 5, 1, dst, stride, above, left, goffset + GET_GOT goffsetq + + DEFINE_ARGS dst, stride, stride3 + lea stride3q, [strideq*3] + movq m0, [GLOBAL(dc_128)] + movq [dstq ], m0 + movq [dstq+strideq ], m0 + movq [dstq+strideq*2], m0 + movq [dstq+stride3q ], m0 + lea dstq, [dstq+strideq*4] + movq [dstq ], m0 + movq [dstq+strideq ], m0 + movq [dstq+strideq*2], m0 + movq [dstq+stride3q ], m0 + RESTORE_GOT + RET + +INIT_XMM sse2 +cglobal dc_predictor_16x16, 4, 5, 3, dst, stride, above, left, goffset + GET_GOT goffsetq + + pxor m1, m1 + mova m0, [aboveq] + mova m2, [leftq] + DEFINE_ARGS dst, stride, stride3, lines4 + lea stride3q, [strideq*3] + mov lines4d, 4 + psadbw m0, m1 + psadbw m2, m1 + paddw m0, m2 + movhlps m2, m0 + paddw m0, m2 + paddw m0, [GLOBAL(pw_16)] + psraw m0, 5 + pshuflw m0, m0, 0x0 + punpcklqdq m0, m0 + packuswb m0, m0 +.loop: + mova [dstq ], m0 + mova [dstq+strideq ], m0 + mova [dstq+strideq*2], m0 + mova [dstq+stride3q ], m0 + lea dstq, [dstq+strideq*4] + dec lines4d + jnz .loop + + RESTORE_GOT + REP_RET + + +INIT_XMM sse2 +cglobal dc_top_predictor_16x16, 4, 5, 3, dst, stride, above, left, goffset + GET_GOT goffsetq + + pxor m1, m1 + mova m0, [aboveq] + DEFINE_ARGS dst, stride, stride3, lines4 + lea stride3q, [strideq*3] + mov lines4d, 4 + psadbw m0, m1 + movhlps m2, m0 + paddw m0, m2 + paddw m0, [GLOBAL(pw2_16)] + psraw m0, 4 + pshuflw m0, m0, 0x0 + punpcklqdq m0, m0 + packuswb m0, m0 +.loop: + mova [dstq ], m0 + mova [dstq+strideq ], m0 + mova [dstq+strideq*2], m0 + mova [dstq+stride3q ], m0 + lea dstq, [dstq+strideq*4] + dec lines4d + jnz .loop + + RESTORE_GOT + REP_RET + +INIT_XMM sse2 +cglobal dc_left_predictor_16x16, 4, 5, 3, dst, stride, above, left, goffset + GET_GOT goffsetq + + pxor m1, m1 + mova m0, [leftq] + DEFINE_ARGS dst, stride, stride3, lines4 + lea stride3q, [strideq*3] + mov lines4d, 4 + psadbw m0, m1 + movhlps m2, m0 + paddw m0, m2 + paddw m0, [GLOBAL(pw2_16)] + psraw m0, 4 + pshuflw m0, m0, 0x0 + punpcklqdq m0, m0 + packuswb m0, m0 +.loop: + mova [dstq ], m0 + mova [dstq+strideq ], m0 + mova [dstq+strideq*2], m0 + mova [dstq+stride3q ], m0 + lea dstq, [dstq+strideq*4] + dec lines4d + jnz .loop + + RESTORE_GOT + REP_RET + +INIT_XMM sse2 +cglobal dc_128_predictor_16x16, 4, 5, 3, dst, stride, above, left, goffset + GET_GOT goffsetq + + DEFINE_ARGS dst, stride, stride3, lines4 + lea stride3q, [strideq*3] + mov lines4d, 4 + mova m0, [GLOBAL(dc_128)] +.loop: + mova [dstq ], m0 + mova [dstq+strideq ], m0 + mova [dstq+strideq*2], m0 + mova [dstq+stride3q ], m0 + lea dstq, [dstq+strideq*4] + dec lines4d + jnz .loop + RESTORE_GOT + RET + + +INIT_XMM sse2 +cglobal dc_predictor_32x32, 4, 5, 5, dst, stride, above, left, goffset + GET_GOT goffsetq + + pxor m1, m1 + mova m0, [aboveq] + mova m2, [aboveq+16] + mova m3, [leftq] + mova m4, [leftq+16] + DEFINE_ARGS dst, stride, stride3, lines4 + lea stride3q, [strideq*3] + mov lines4d, 8 + psadbw m0, m1 + psadbw m2, m1 + psadbw m3, m1 + psadbw m4, m1 + paddw m0, m2 + paddw m0, m3 + paddw m0, m4 + movhlps m2, m0 + paddw m0, m2 + paddw m0, [GLOBAL(pw_32)] + psraw m0, 6 + pshuflw m0, m0, 0x0 + punpcklqdq m0, m0 + packuswb m0, m0 +.loop: + mova [dstq ], m0 + mova [dstq +16], m0 + mova [dstq+strideq ], m0 + mova [dstq+strideq +16], m0 + mova [dstq+strideq*2 ], m0 + mova [dstq+strideq*2+16], m0 + mova [dstq+stride3q ], m0 + mova [dstq+stride3q +16], m0 + lea dstq, [dstq+strideq*4] + dec lines4d + jnz .loop + + RESTORE_GOT + REP_RET + +INIT_XMM sse2 +cglobal dc_top_predictor_32x32, 4, 5, 5, dst, stride, above, left, goffset + GET_GOT goffsetq + + pxor m1, m1 + mova m0, [aboveq] + mova m2, [aboveq+16] + DEFINE_ARGS dst, stride, stride3, lines4 + lea stride3q, [strideq*3] + mov lines4d, 8 + psadbw m0, m1 + psadbw m2, m1 + paddw m0, m2 + movhlps m2, m0 + paddw m0, m2 + paddw m0, [GLOBAL(pw2_32)] + psraw m0, 5 + pshuflw m0, m0, 0x0 + punpcklqdq m0, m0 + packuswb m0, m0 +.loop: + mova [dstq ], m0 + mova [dstq +16], m0 + mova [dstq+strideq ], m0 + mova [dstq+strideq +16], m0 + mova [dstq+strideq*2 ], m0 + mova [dstq+strideq*2+16], m0 + mova [dstq+stride3q ], m0 + mova [dstq+stride3q +16], m0 + lea dstq, [dstq+strideq*4] + dec lines4d + jnz .loop + + RESTORE_GOT + REP_RET + +INIT_XMM sse2 +cglobal dc_left_predictor_32x32, 4, 5, 5, dst, stride, above, left, goffset + GET_GOT goffsetq + + pxor m1, m1 + mova m0, [leftq] + mova m2, [leftq+16] + DEFINE_ARGS dst, stride, stride3, lines4 + lea stride3q, [strideq*3] + mov lines4d, 8 + psadbw m0, m1 + psadbw m2, m1 + paddw m0, m2 + movhlps m2, m0 + paddw m0, m2 + paddw m0, [GLOBAL(pw2_32)] + psraw m0, 5 + pshuflw m0, m0, 0x0 + punpcklqdq m0, m0 + packuswb m0, m0 +.loop: + mova [dstq ], m0 + mova [dstq +16], m0 + mova [dstq+strideq ], m0 + mova [dstq+strideq +16], m0 + mova [dstq+strideq*2 ], m0 + mova [dstq+strideq*2+16], m0 + mova [dstq+stride3q ], m0 + mova [dstq+stride3q +16], m0 + lea dstq, [dstq+strideq*4] + dec lines4d + jnz .loop + + RESTORE_GOT + REP_RET + +INIT_XMM sse2 +cglobal dc_128_predictor_32x32, 4, 5, 3, dst, stride, above, left, goffset + GET_GOT goffsetq + + DEFINE_ARGS dst, stride, stride3, lines4 + lea stride3q, [strideq*3] + mov lines4d, 8 + mova m0, [GLOBAL(dc_128)] +.loop: + mova [dstq ], m0 + mova [dstq +16], m0 + mova [dstq+strideq ], m0 + mova [dstq+strideq +16], m0 + mova [dstq+strideq*2 ], m0 + mova [dstq+strideq*2+16], m0 + mova [dstq+stride3q ], m0 + mova [dstq+stride3q +16], m0 + lea dstq, [dstq+strideq*4] + dec lines4d + jnz .loop + RESTORE_GOT + RET + +INIT_XMM sse2 +cglobal v_predictor_4x4, 3, 3, 1, dst, stride, above + movd m0, [aboveq] + movd [dstq ], m0 + movd [dstq+strideq], m0 + lea dstq, [dstq+strideq*2] + movd [dstq ], m0 + movd [dstq+strideq], m0 + RET + +INIT_XMM sse2 +cglobal v_predictor_8x8, 3, 3, 1, dst, stride, above + movq m0, [aboveq] + DEFINE_ARGS dst, stride, stride3 + lea stride3q, [strideq*3] + movq [dstq ], m0 + movq [dstq+strideq ], m0 + movq [dstq+strideq*2], m0 + movq [dstq+stride3q ], m0 + lea dstq, [dstq+strideq*4] + movq [dstq ], m0 + movq [dstq+strideq ], m0 + movq [dstq+strideq*2], m0 + movq [dstq+stride3q ], m0 + RET + +INIT_XMM sse2 +cglobal v_predictor_16x16, 3, 4, 1, dst, stride, above + mova m0, [aboveq] + DEFINE_ARGS dst, stride, stride3, nlines4 + lea stride3q, [strideq*3] + mov nlines4d, 4 +.loop: + mova [dstq ], m0 + mova [dstq+strideq ], m0 + mova [dstq+strideq*2], m0 + mova [dstq+stride3q ], m0 + lea dstq, [dstq+strideq*4] + dec nlines4d + jnz .loop + REP_RET + +INIT_XMM sse2 +cglobal v_predictor_32x32, 3, 4, 2, dst, stride, above + mova m0, [aboveq] + mova m1, [aboveq+16] + DEFINE_ARGS dst, stride, stride3, nlines4 + lea stride3q, [strideq*3] + mov nlines4d, 8 +.loop: + mova [dstq ], m0 + mova [dstq +16], m1 + mova [dstq+strideq ], m0 + mova [dstq+strideq +16], m1 + mova [dstq+strideq*2 ], m0 + mova [dstq+strideq*2+16], m1 + mova [dstq+stride3q ], m0 + mova [dstq+stride3q +16], m1 + lea dstq, [dstq+strideq*4] + dec nlines4d + jnz .loop + REP_RET + +INIT_XMM sse2 +cglobal h_predictor_4x4, 2, 4, 4, dst, stride, line, left + movifnidn leftq, leftmp + movd m0, [leftq] + punpcklbw m0, m0 + punpcklbw m0, m0 + pshufd m1, m0, 0x1 + movd [dstq ], m0 + movd [dstq+strideq], m1 + pshufd m2, m0, 0x2 + lea dstq, [dstq+strideq*2] + pshufd m3, m0, 0x3 + movd [dstq ], m2 + movd [dstq+strideq], m3 + RET + +INIT_XMM sse2 +cglobal h_predictor_8x8, 2, 5, 3, dst, stride, line, left + movifnidn leftq, leftmp + mov lineq, -2 + DEFINE_ARGS dst, stride, line, left, stride3 + lea stride3q, [strideq*3] + movq m0, [leftq ] + punpcklbw m0, m0 ; l1 l1 l2 l2 ... l8 l8 +.loop: + pshuflw m1, m0, 0x0 ; l1 l1 l1 l1 l1 l1 l1 l1 + pshuflw m2, m0, 0x55 ; l2 l2 l2 l2 l2 l2 l2 l2 + movq [dstq ], m1 + movq [dstq+strideq], m2 + pshuflw m1, m0, 0xaa + pshuflw m2, m0, 0xff + movq [dstq+strideq*2], m1 + movq [dstq+stride3q ], m2 + pshufd m0, m0, 0xe ; [63:0] l5 l5 l6 l6 l7 l7 l8 l8 + inc lineq + lea dstq, [dstq+strideq*4] + jnz .loop + REP_RET + +INIT_XMM sse2 +cglobal h_predictor_16x16, 2, 5, 3, dst, stride, line, left + movifnidn leftq, leftmp + mov lineq, -4 + DEFINE_ARGS dst, stride, line, left, stride3 + lea stride3q, [strideq*3] +.loop: + movd m0, [leftq] + punpcklbw m0, m0 + punpcklbw m0, m0 ; l1 to l4 each repeated 4 times + pshufd m1, m0, 0x0 ; l1 repeated 16 times + pshufd m2, m0, 0x55 ; l2 repeated 16 times + mova [dstq ], m1 + mova [dstq+strideq ], m2 + pshufd m1, m0, 0xaa + pshufd m2, m0, 0xff + mova [dstq+strideq*2], m1 + mova [dstq+stride3q ], m2 + inc lineq + lea leftq, [leftq+4 ] + lea dstq, [dstq+strideq*4] + jnz .loop + REP_RET + +INIT_XMM sse2 +cglobal h_predictor_32x32, 2, 5, 3, dst, stride, line, left + movifnidn leftq, leftmp + mov lineq, -8 + DEFINE_ARGS dst, stride, line, left, stride3 + lea stride3q, [strideq*3] +.loop: + movd m0, [leftq] + punpcklbw m0, m0 + punpcklbw m0, m0 ; l1 to l4 each repeated 4 times + pshufd m1, m0, 0x0 ; l1 repeated 16 times + pshufd m2, m0, 0x55 ; l2 repeated 16 times + mova [dstq ], m1 + mova [dstq+16 ], m1 + mova [dstq+strideq ], m2 + mova [dstq+strideq+16 ], m2 + pshufd m1, m0, 0xaa + pshufd m2, m0, 0xff + mova [dstq+strideq*2 ], m1 + mova [dstq+strideq*2+16], m1 + mova [dstq+stride3q ], m2 + mova [dstq+stride3q+16 ], m2 + inc lineq + lea leftq, [leftq+4 ] + lea dstq, [dstq+strideq*4] + jnz .loop + REP_RET diff --git a/third_party/aom/aom_dsp/x86/intrapred_ssse3.c b/third_party/aom/aom_dsp/x86/intrapred_ssse3.c new file mode 100644 index 0000000000..807ed1770f --- /dev/null +++ b/third_party/aom/aom_dsp/x86/intrapred_ssse3.c @@ -0,0 +1,1692 @@ +/* + * Copyright (c) 2017, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include + +#include "config/aom_dsp_rtcd.h" + +#include "aom_dsp/intrapred_common.h" + +// ----------------------------------------------------------------------------- +// PAETH_PRED + +// Return 8 16-bit pixels in one row +static INLINE __m128i paeth_8x1_pred(const __m128i *left, const __m128i *top, + const __m128i *topleft) { + const __m128i base = _mm_sub_epi16(_mm_add_epi16(*top, *left), *topleft); + + __m128i pl = _mm_abs_epi16(_mm_sub_epi16(base, *left)); + __m128i pt = _mm_abs_epi16(_mm_sub_epi16(base, *top)); + __m128i ptl = _mm_abs_epi16(_mm_sub_epi16(base, *topleft)); + + __m128i mask1 = _mm_cmpgt_epi16(pl, pt); + mask1 = _mm_or_si128(mask1, _mm_cmpgt_epi16(pl, ptl)); + __m128i mask2 = _mm_cmpgt_epi16(pt, ptl); + + pl = _mm_andnot_si128(mask1, *left); + + ptl = _mm_and_si128(mask2, *topleft); + pt = _mm_andnot_si128(mask2, *top); + pt = _mm_or_si128(pt, ptl); + pt = _mm_and_si128(mask1, pt); + + return _mm_or_si128(pl, pt); +} + +void aom_paeth_predictor_4x4_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + __m128i l = _mm_loadl_epi64((const __m128i *)left); + const __m128i t = _mm_loadl_epi64((const __m128i *)above); + const __m128i zero = _mm_setzero_si128(); + const __m128i t16 = _mm_unpacklo_epi8(t, zero); + const __m128i tl16 = _mm_set1_epi16((uint16_t)above[-1]); + __m128i rep = _mm_set1_epi16(0x8000); + const __m128i one = _mm_set1_epi16(1); + + int i; + for (i = 0; i < 4; ++i) { + const __m128i l16 = _mm_shuffle_epi8(l, rep); + const __m128i row = paeth_8x1_pred(&l16, &t16, &tl16); + + *(uint32_t *)dst = _mm_cvtsi128_si32(_mm_packus_epi16(row, row)); + dst += stride; + rep = _mm_add_epi16(rep, one); + } +} + +void aom_paeth_predictor_4x8_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + __m128i l = _mm_loadl_epi64((const __m128i *)left); + const __m128i t = _mm_loadl_epi64((const __m128i *)above); + const __m128i zero = _mm_setzero_si128(); + const __m128i t16 = _mm_unpacklo_epi8(t, zero); + const __m128i tl16 = _mm_set1_epi16((uint16_t)above[-1]); + __m128i rep = _mm_set1_epi16(0x8000); + const __m128i one = _mm_set1_epi16(1); + + int i; + for (i = 0; i < 8; ++i) { + const __m128i l16 = _mm_shuffle_epi8(l, rep); + const __m128i row = paeth_8x1_pred(&l16, &t16, &tl16); + + *(uint32_t *)dst = _mm_cvtsi128_si32(_mm_packus_epi16(row, row)); + dst += stride; + rep = _mm_add_epi16(rep, one); + } +} + +void aom_paeth_predictor_4x16_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + __m128i l = _mm_load_si128((const __m128i *)left); + const __m128i t = _mm_cvtsi32_si128(((const uint32_t *)above)[0]); + const __m128i zero = _mm_setzero_si128(); + const __m128i t16 = _mm_unpacklo_epi8(t, zero); + const __m128i tl16 = _mm_set1_epi16((uint16_t)above[-1]); + __m128i rep = _mm_set1_epi16(0x8000); + const __m128i one = _mm_set1_epi16(1); + + for (int i = 0; i < 16; ++i) { + const __m128i l16 = _mm_shuffle_epi8(l, rep); + const __m128i row = paeth_8x1_pred(&l16, &t16, &tl16); + + *(uint32_t *)dst = _mm_cvtsi128_si32(_mm_packus_epi16(row, row)); + dst += stride; + rep = _mm_add_epi16(rep, one); + } +} + +void aom_paeth_predictor_8x4_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + __m128i l = _mm_loadl_epi64((const __m128i *)left); + const __m128i t = _mm_loadl_epi64((const __m128i *)above); + const __m128i zero = _mm_setzero_si128(); + const __m128i t16 = _mm_unpacklo_epi8(t, zero); + const __m128i tl16 = _mm_set1_epi16((uint16_t)above[-1]); + __m128i rep = _mm_set1_epi16(0x8000); + const __m128i one = _mm_set1_epi16(1); + + int i; + for (i = 0; i < 4; ++i) { + const __m128i l16 = _mm_shuffle_epi8(l, rep); + const __m128i row = paeth_8x1_pred(&l16, &t16, &tl16); + + _mm_storel_epi64((__m128i *)dst, _mm_packus_epi16(row, row)); + dst += stride; + rep = _mm_add_epi16(rep, one); + } +} + +void aom_paeth_predictor_8x8_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + __m128i l = _mm_loadl_epi64((const __m128i *)left); + const __m128i t = _mm_loadl_epi64((const __m128i *)above); + const __m128i zero = _mm_setzero_si128(); + const __m128i t16 = _mm_unpacklo_epi8(t, zero); + const __m128i tl16 = _mm_set1_epi16((uint16_t)above[-1]); + __m128i rep = _mm_set1_epi16(0x8000); + const __m128i one = _mm_set1_epi16(1); + + int i; + for (i = 0; i < 8; ++i) { + const __m128i l16 = _mm_shuffle_epi8(l, rep); + const __m128i row = paeth_8x1_pred(&l16, &t16, &tl16); + + _mm_storel_epi64((__m128i *)dst, _mm_packus_epi16(row, row)); + dst += stride; + rep = _mm_add_epi16(rep, one); + } +} + +void aom_paeth_predictor_8x16_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + __m128i l = _mm_load_si128((const __m128i *)left); + const __m128i t = _mm_loadl_epi64((const __m128i *)above); + const __m128i zero = _mm_setzero_si128(); + const __m128i t16 = _mm_unpacklo_epi8(t, zero); + const __m128i tl16 = _mm_set1_epi16((uint16_t)above[-1]); + __m128i rep = _mm_set1_epi16(0x8000); + const __m128i one = _mm_set1_epi16(1); + + int i; + for (i = 0; i < 16; ++i) { + const __m128i l16 = _mm_shuffle_epi8(l, rep); + const __m128i row = paeth_8x1_pred(&l16, &t16, &tl16); + + _mm_storel_epi64((__m128i *)dst, _mm_packus_epi16(row, row)); + dst += stride; + rep = _mm_add_epi16(rep, one); + } +} + +void aom_paeth_predictor_8x32_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + const __m128i t = _mm_loadl_epi64((const __m128i *)above); + const __m128i zero = _mm_setzero_si128(); + const __m128i t16 = _mm_unpacklo_epi8(t, zero); + const __m128i tl16 = _mm_set1_epi16((uint16_t)above[-1]); + const __m128i one = _mm_set1_epi16(1); + + for (int j = 0; j < 2; ++j) { + const __m128i l = _mm_load_si128((const __m128i *)(left + j * 16)); + __m128i rep = _mm_set1_epi16(0x8000); + for (int i = 0; i < 16; ++i) { + const __m128i l16 = _mm_shuffle_epi8(l, rep); + const __m128i row = paeth_8x1_pred(&l16, &t16, &tl16); + + _mm_storel_epi64((__m128i *)dst, _mm_packus_epi16(row, row)); + dst += stride; + rep = _mm_add_epi16(rep, one); + } + } +} + +// Return 16 8-bit pixels in one row +static INLINE __m128i paeth_16x1_pred(const __m128i *left, const __m128i *top0, + const __m128i *top1, + const __m128i *topleft) { + const __m128i p0 = paeth_8x1_pred(left, top0, topleft); + const __m128i p1 = paeth_8x1_pred(left, top1, topleft); + return _mm_packus_epi16(p0, p1); +} + +void aom_paeth_predictor_16x4_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + __m128i l = _mm_cvtsi32_si128(((const uint32_t *)left)[0]); + const __m128i t = _mm_load_si128((const __m128i *)above); + const __m128i zero = _mm_setzero_si128(); + const __m128i top0 = _mm_unpacklo_epi8(t, zero); + const __m128i top1 = _mm_unpackhi_epi8(t, zero); + const __m128i tl16 = _mm_set1_epi16((uint16_t)above[-1]); + __m128i rep = _mm_set1_epi16(0x8000); + const __m128i one = _mm_set1_epi16(1); + + for (int i = 0; i < 4; ++i) { + const __m128i l16 = _mm_shuffle_epi8(l, rep); + const __m128i row = paeth_16x1_pred(&l16, &top0, &top1, &tl16); + + _mm_store_si128((__m128i *)dst, row); + dst += stride; + rep = _mm_add_epi16(rep, one); + } +} + +void aom_paeth_predictor_16x8_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + __m128i l = _mm_loadl_epi64((const __m128i *)left); + const __m128i t = _mm_load_si128((const __m128i *)above); + const __m128i zero = _mm_setzero_si128(); + const __m128i top0 = _mm_unpacklo_epi8(t, zero); + const __m128i top1 = _mm_unpackhi_epi8(t, zero); + const __m128i tl16 = _mm_set1_epi16((uint16_t)above[-1]); + __m128i rep = _mm_set1_epi16(0x8000); + const __m128i one = _mm_set1_epi16(1); + + int i; + for (i = 0; i < 8; ++i) { + const __m128i l16 = _mm_shuffle_epi8(l, rep); + const __m128i row = paeth_16x1_pred(&l16, &top0, &top1, &tl16); + + _mm_store_si128((__m128i *)dst, row); + dst += stride; + rep = _mm_add_epi16(rep, one); + } +} + +void aom_paeth_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + __m128i l = _mm_load_si128((const __m128i *)left); + const __m128i t = _mm_load_si128((const __m128i *)above); + const __m128i zero = _mm_setzero_si128(); + const __m128i top0 = _mm_unpacklo_epi8(t, zero); + const __m128i top1 = _mm_unpackhi_epi8(t, zero); + const __m128i tl16 = _mm_set1_epi16((uint16_t)above[-1]); + __m128i rep = _mm_set1_epi16(0x8000); + const __m128i one = _mm_set1_epi16(1); + + int i; + for (i = 0; i < 16; ++i) { + const __m128i l16 = _mm_shuffle_epi8(l, rep); + const __m128i row = paeth_16x1_pred(&l16, &top0, &top1, &tl16); + + _mm_store_si128((__m128i *)dst, row); + dst += stride; + rep = _mm_add_epi16(rep, one); + } +} + +void aom_paeth_predictor_16x32_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + __m128i l = _mm_load_si128((const __m128i *)left); + const __m128i t = _mm_load_si128((const __m128i *)above); + const __m128i zero = _mm_setzero_si128(); + const __m128i top0 = _mm_unpacklo_epi8(t, zero); + const __m128i top1 = _mm_unpackhi_epi8(t, zero); + const __m128i tl16 = _mm_set1_epi16((uint16_t)above[-1]); + __m128i rep = _mm_set1_epi16(0x8000); + const __m128i one = _mm_set1_epi16(1); + __m128i l16; + + int i; + for (i = 0; i < 16; ++i) { + l16 = _mm_shuffle_epi8(l, rep); + const __m128i row = paeth_16x1_pred(&l16, &top0, &top1, &tl16); + + _mm_store_si128((__m128i *)dst, row); + dst += stride; + rep = _mm_add_epi16(rep, one); + } + + l = _mm_load_si128((const __m128i *)(left + 16)); + rep = _mm_set1_epi16(0x8000); + for (i = 0; i < 16; ++i) { + l16 = _mm_shuffle_epi8(l, rep); + const __m128i row = paeth_16x1_pred(&l16, &top0, &top1, &tl16); + + _mm_store_si128((__m128i *)dst, row); + dst += stride; + rep = _mm_add_epi16(rep, one); + } +} + +void aom_paeth_predictor_16x64_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + const __m128i t = _mm_load_si128((const __m128i *)above); + const __m128i zero = _mm_setzero_si128(); + const __m128i top0 = _mm_unpacklo_epi8(t, zero); + const __m128i top1 = _mm_unpackhi_epi8(t, zero); + const __m128i tl16 = _mm_set1_epi16((uint16_t)above[-1]); + const __m128i one = _mm_set1_epi16(1); + + for (int j = 0; j < 4; ++j) { + const __m128i l = _mm_load_si128((const __m128i *)(left + j * 16)); + __m128i rep = _mm_set1_epi16(0x8000); + for (int i = 0; i < 16; ++i) { + const __m128i l16 = _mm_shuffle_epi8(l, rep); + const __m128i row = paeth_16x1_pred(&l16, &top0, &top1, &tl16); + _mm_store_si128((__m128i *)dst, row); + dst += stride; + rep = _mm_add_epi16(rep, one); + } + } +} + +void aom_paeth_predictor_32x8_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + const __m128i a = _mm_load_si128((const __m128i *)above); + const __m128i b = _mm_load_si128((const __m128i *)(above + 16)); + const __m128i zero = _mm_setzero_si128(); + const __m128i al = _mm_unpacklo_epi8(a, zero); + const __m128i ah = _mm_unpackhi_epi8(a, zero); + const __m128i bl = _mm_unpacklo_epi8(b, zero); + const __m128i bh = _mm_unpackhi_epi8(b, zero); + + const __m128i tl16 = _mm_set1_epi16((uint16_t)above[-1]); + __m128i rep = _mm_set1_epi16(0x8000); + const __m128i one = _mm_set1_epi16(1); + const __m128i l = _mm_loadl_epi64((const __m128i *)left); + __m128i l16; + + for (int i = 0; i < 8; ++i) { + l16 = _mm_shuffle_epi8(l, rep); + const __m128i r32l = paeth_16x1_pred(&l16, &al, &ah, &tl16); + const __m128i r32h = paeth_16x1_pred(&l16, &bl, &bh, &tl16); + + _mm_store_si128((__m128i *)dst, r32l); + _mm_store_si128((__m128i *)(dst + 16), r32h); + dst += stride; + rep = _mm_add_epi16(rep, one); + } +} + +void aom_paeth_predictor_32x16_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + const __m128i a = _mm_load_si128((const __m128i *)above); + const __m128i b = _mm_load_si128((const __m128i *)(above + 16)); + const __m128i zero = _mm_setzero_si128(); + const __m128i al = _mm_unpacklo_epi8(a, zero); + const __m128i ah = _mm_unpackhi_epi8(a, zero); + const __m128i bl = _mm_unpacklo_epi8(b, zero); + const __m128i bh = _mm_unpackhi_epi8(b, zero); + + const __m128i tl16 = _mm_set1_epi16((uint16_t)above[-1]); + __m128i rep = _mm_set1_epi16(0x8000); + const __m128i one = _mm_set1_epi16(1); + __m128i l = _mm_load_si128((const __m128i *)left); + __m128i l16; + + int i; + for (i = 0; i < 16; ++i) { + l16 = _mm_shuffle_epi8(l, rep); + const __m128i r32l = paeth_16x1_pred(&l16, &al, &ah, &tl16); + const __m128i r32h = paeth_16x1_pred(&l16, &bl, &bh, &tl16); + + _mm_store_si128((__m128i *)dst, r32l); + _mm_store_si128((__m128i *)(dst + 16), r32h); + dst += stride; + rep = _mm_add_epi16(rep, one); + } +} + +void aom_paeth_predictor_32x32_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + const __m128i a = _mm_load_si128((const __m128i *)above); + const __m128i b = _mm_load_si128((const __m128i *)(above + 16)); + const __m128i zero = _mm_setzero_si128(); + const __m128i al = _mm_unpacklo_epi8(a, zero); + const __m128i ah = _mm_unpackhi_epi8(a, zero); + const __m128i bl = _mm_unpacklo_epi8(b, zero); + const __m128i bh = _mm_unpackhi_epi8(b, zero); + + const __m128i tl16 = _mm_set1_epi16((uint16_t)above[-1]); + __m128i rep = _mm_set1_epi16(0x8000); + const __m128i one = _mm_set1_epi16(1); + __m128i l = _mm_load_si128((const __m128i *)left); + __m128i l16; + + int i; + for (i = 0; i < 16; ++i) { + l16 = _mm_shuffle_epi8(l, rep); + const __m128i r32l = paeth_16x1_pred(&l16, &al, &ah, &tl16); + const __m128i r32h = paeth_16x1_pred(&l16, &bl, &bh, &tl16); + + _mm_store_si128((__m128i *)dst, r32l); + _mm_store_si128((__m128i *)(dst + 16), r32h); + dst += stride; + rep = _mm_add_epi16(rep, one); + } + + rep = _mm_set1_epi16(0x8000); + l = _mm_load_si128((const __m128i *)(left + 16)); + for (i = 0; i < 16; ++i) { + l16 = _mm_shuffle_epi8(l, rep); + const __m128i r32l = paeth_16x1_pred(&l16, &al, &ah, &tl16); + const __m128i r32h = paeth_16x1_pred(&l16, &bl, &bh, &tl16); + + _mm_store_si128((__m128i *)dst, r32l); + _mm_store_si128((__m128i *)(dst + 16), r32h); + dst += stride; + rep = _mm_add_epi16(rep, one); + } +} + +void aom_paeth_predictor_32x64_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + const __m128i a = _mm_load_si128((const __m128i *)above); + const __m128i b = _mm_load_si128((const __m128i *)(above + 16)); + const __m128i zero = _mm_setzero_si128(); + const __m128i al = _mm_unpacklo_epi8(a, zero); + const __m128i ah = _mm_unpackhi_epi8(a, zero); + const __m128i bl = _mm_unpacklo_epi8(b, zero); + const __m128i bh = _mm_unpackhi_epi8(b, zero); + + const __m128i tl16 = _mm_set1_epi16((uint16_t)above[-1]); + const __m128i one = _mm_set1_epi16(1); + __m128i l16; + + int i, j; + for (j = 0; j < 4; ++j) { + const __m128i l = _mm_load_si128((const __m128i *)(left + j * 16)); + __m128i rep = _mm_set1_epi16(0x8000); + for (i = 0; i < 16; ++i) { + l16 = _mm_shuffle_epi8(l, rep); + const __m128i r32l = paeth_16x1_pred(&l16, &al, &ah, &tl16); + const __m128i r32h = paeth_16x1_pred(&l16, &bl, &bh, &tl16); + + _mm_store_si128((__m128i *)dst, r32l); + _mm_store_si128((__m128i *)(dst + 16), r32h); + dst += stride; + rep = _mm_add_epi16(rep, one); + } + } +} + +void aom_paeth_predictor_64x32_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + const __m128i a = _mm_load_si128((const __m128i *)above); + const __m128i b = _mm_load_si128((const __m128i *)(above + 16)); + const __m128i c = _mm_load_si128((const __m128i *)(above + 32)); + const __m128i d = _mm_load_si128((const __m128i *)(above + 48)); + const __m128i zero = _mm_setzero_si128(); + const __m128i al = _mm_unpacklo_epi8(a, zero); + const __m128i ah = _mm_unpackhi_epi8(a, zero); + const __m128i bl = _mm_unpacklo_epi8(b, zero); + const __m128i bh = _mm_unpackhi_epi8(b, zero); + const __m128i cl = _mm_unpacklo_epi8(c, zero); + const __m128i ch = _mm_unpackhi_epi8(c, zero); + const __m128i dl = _mm_unpacklo_epi8(d, zero); + const __m128i dh = _mm_unpackhi_epi8(d, zero); + + const __m128i tl16 = _mm_set1_epi16((uint16_t)above[-1]); + const __m128i one = _mm_set1_epi16(1); + __m128i l16; + + int i, j; + for (j = 0; j < 2; ++j) { + const __m128i l = _mm_load_si128((const __m128i *)(left + j * 16)); + __m128i rep = _mm_set1_epi16(0x8000); + for (i = 0; i < 16; ++i) { + l16 = _mm_shuffle_epi8(l, rep); + const __m128i r0 = paeth_16x1_pred(&l16, &al, &ah, &tl16); + const __m128i r1 = paeth_16x1_pred(&l16, &bl, &bh, &tl16); + const __m128i r2 = paeth_16x1_pred(&l16, &cl, &ch, &tl16); + const __m128i r3 = paeth_16x1_pred(&l16, &dl, &dh, &tl16); + + _mm_store_si128((__m128i *)dst, r0); + _mm_store_si128((__m128i *)(dst + 16), r1); + _mm_store_si128((__m128i *)(dst + 32), r2); + _mm_store_si128((__m128i *)(dst + 48), r3); + dst += stride; + rep = _mm_add_epi16(rep, one); + } + } +} + +void aom_paeth_predictor_64x64_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + const __m128i a = _mm_load_si128((const __m128i *)above); + const __m128i b = _mm_load_si128((const __m128i *)(above + 16)); + const __m128i c = _mm_load_si128((const __m128i *)(above + 32)); + const __m128i d = _mm_load_si128((const __m128i *)(above + 48)); + const __m128i zero = _mm_setzero_si128(); + const __m128i al = _mm_unpacklo_epi8(a, zero); + const __m128i ah = _mm_unpackhi_epi8(a, zero); + const __m128i bl = _mm_unpacklo_epi8(b, zero); + const __m128i bh = _mm_unpackhi_epi8(b, zero); + const __m128i cl = _mm_unpacklo_epi8(c, zero); + const __m128i ch = _mm_unpackhi_epi8(c, zero); + const __m128i dl = _mm_unpacklo_epi8(d, zero); + const __m128i dh = _mm_unpackhi_epi8(d, zero); + + const __m128i tl16 = _mm_set1_epi16((uint16_t)above[-1]); + const __m128i one = _mm_set1_epi16(1); + __m128i l16; + + int i, j; + for (j = 0; j < 4; ++j) { + const __m128i l = _mm_load_si128((const __m128i *)(left + j * 16)); + __m128i rep = _mm_set1_epi16(0x8000); + for (i = 0; i < 16; ++i) { + l16 = _mm_shuffle_epi8(l, rep); + const __m128i r0 = paeth_16x1_pred(&l16, &al, &ah, &tl16); + const __m128i r1 = paeth_16x1_pred(&l16, &bl, &bh, &tl16); + const __m128i r2 = paeth_16x1_pred(&l16, &cl, &ch, &tl16); + const __m128i r3 = paeth_16x1_pred(&l16, &dl, &dh, &tl16); + + _mm_store_si128((__m128i *)dst, r0); + _mm_store_si128((__m128i *)(dst + 16), r1); + _mm_store_si128((__m128i *)(dst + 32), r2); + _mm_store_si128((__m128i *)(dst + 48), r3); + dst += stride; + rep = _mm_add_epi16(rep, one); + } + } +} + +void aom_paeth_predictor_64x16_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + const __m128i a = _mm_load_si128((const __m128i *)above); + const __m128i b = _mm_load_si128((const __m128i *)(above + 16)); + const __m128i c = _mm_load_si128((const __m128i *)(above + 32)); + const __m128i d = _mm_load_si128((const __m128i *)(above + 48)); + const __m128i zero = _mm_setzero_si128(); + const __m128i al = _mm_unpacklo_epi8(a, zero); + const __m128i ah = _mm_unpackhi_epi8(a, zero); + const __m128i bl = _mm_unpacklo_epi8(b, zero); + const __m128i bh = _mm_unpackhi_epi8(b, zero); + const __m128i cl = _mm_unpacklo_epi8(c, zero); + const __m128i ch = _mm_unpackhi_epi8(c, zero); + const __m128i dl = _mm_unpacklo_epi8(d, zero); + const __m128i dh = _mm_unpackhi_epi8(d, zero); + + const __m128i tl16 = _mm_set1_epi16((uint16_t)above[-1]); + const __m128i one = _mm_set1_epi16(1); + __m128i l16; + + int i; + const __m128i l = _mm_load_si128((const __m128i *)left); + __m128i rep = _mm_set1_epi16(0x8000); + for (i = 0; i < 16; ++i) { + l16 = _mm_shuffle_epi8(l, rep); + const __m128i r0 = paeth_16x1_pred(&l16, &al, &ah, &tl16); + const __m128i r1 = paeth_16x1_pred(&l16, &bl, &bh, &tl16); + const __m128i r2 = paeth_16x1_pred(&l16, &cl, &ch, &tl16); + const __m128i r3 = paeth_16x1_pred(&l16, &dl, &dh, &tl16); + + _mm_store_si128((__m128i *)dst, r0); + _mm_store_si128((__m128i *)(dst + 16), r1); + _mm_store_si128((__m128i *)(dst + 32), r2); + _mm_store_si128((__m128i *)(dst + 48), r3); + dst += stride; + rep = _mm_add_epi16(rep, one); + } +} + +// ----------------------------------------------------------------------------- +// SMOOTH_PRED + +// pixels[0]: above and below_pred interleave vector +// pixels[1]: left vector +// pixels[2]: right_pred vector +static INLINE void load_pixel_w4(const uint8_t *above, const uint8_t *left, + int height, __m128i *pixels) { + __m128i d = _mm_cvtsi32_si128(((const uint32_t *)above)[0]); + if (height == 4) + pixels[1] = _mm_cvtsi32_si128(((const uint32_t *)left)[0]); + else if (height == 8) + pixels[1] = _mm_loadl_epi64(((const __m128i *)left)); + else + pixels[1] = _mm_loadu_si128(((const __m128i *)left)); + + pixels[2] = _mm_set1_epi16((uint16_t)above[3]); + + const __m128i bp = _mm_set1_epi16((uint16_t)left[height - 1]); + const __m128i zero = _mm_setzero_si128(); + d = _mm_unpacklo_epi8(d, zero); + pixels[0] = _mm_unpacklo_epi16(d, bp); +} + +// weight_h[0]: weight_h vector +// weight_h[1]: scale - weight_h vector +// weight_h[2]: same as [0], second half for height = 16 only +// weight_h[3]: same as [1], second half for height = 16 only +// weight_w[0]: weights_w and scale - weights_w interleave vector +static INLINE void load_weight_w4(const uint8_t *weight_array, int height, + __m128i *weight_h, __m128i *weight_w) { + const __m128i zero = _mm_setzero_si128(); + const __m128i d = _mm_set1_epi16((uint16_t)(1 << sm_weight_log2_scale)); + const __m128i t = _mm_cvtsi32_si128(((const uint32_t *)weight_array)[1]); + weight_h[0] = _mm_unpacklo_epi8(t, zero); + weight_h[1] = _mm_sub_epi16(d, weight_h[0]); + weight_w[0] = _mm_unpacklo_epi16(weight_h[0], weight_h[1]); + + if (height == 8) { + const __m128i weight = _mm_loadl_epi64((const __m128i *)&weight_array[8]); + weight_h[0] = _mm_unpacklo_epi8(weight, zero); + weight_h[1] = _mm_sub_epi16(d, weight_h[0]); + } else if (height == 16) { + const __m128i weight = _mm_loadu_si128((const __m128i *)&weight_array[16]); + weight_h[0] = _mm_unpacklo_epi8(weight, zero); + weight_h[1] = _mm_sub_epi16(d, weight_h[0]); + weight_h[2] = _mm_unpackhi_epi8(weight, zero); + weight_h[3] = _mm_sub_epi16(d, weight_h[2]); + } +} + +static INLINE void smooth_pred_4xh(const __m128i *pixel, const __m128i *wh, + const __m128i *ww, int h, uint8_t *dst, + ptrdiff_t stride, int second_half) { + const __m128i round = _mm_set1_epi32((1 << sm_weight_log2_scale)); + const __m128i one = _mm_set1_epi16(1); + const __m128i inc = _mm_set1_epi16(0x202); + const __m128i gat = _mm_set1_epi32(0xc080400); + __m128i rep = second_half ? _mm_set1_epi16(0x8008) : _mm_set1_epi16(0x8000); + __m128i d = _mm_set1_epi16(0x100); + + for (int i = 0; i < h; ++i) { + const __m128i wg_wg = _mm_shuffle_epi8(wh[0], d); + const __m128i sc_sc = _mm_shuffle_epi8(wh[1], d); + const __m128i wh_sc = _mm_unpacklo_epi16(wg_wg, sc_sc); + __m128i s = _mm_madd_epi16(pixel[0], wh_sc); + + __m128i b = _mm_shuffle_epi8(pixel[1], rep); + b = _mm_unpacklo_epi16(b, pixel[2]); + __m128i sum = _mm_madd_epi16(b, ww[0]); + + sum = _mm_add_epi32(s, sum); + sum = _mm_add_epi32(sum, round); + sum = _mm_srai_epi32(sum, 1 + sm_weight_log2_scale); + + sum = _mm_shuffle_epi8(sum, gat); + *(uint32_t *)dst = _mm_cvtsi128_si32(sum); + dst += stride; + + rep = _mm_add_epi16(rep, one); + d = _mm_add_epi16(d, inc); + } +} + +void aom_smooth_predictor_4x4_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + __m128i pixels[3]; + load_pixel_w4(above, left, 4, pixels); + + __m128i wh[4], ww[2]; + load_weight_w4(sm_weight_arrays, 4, wh, ww); + + smooth_pred_4xh(pixels, wh, ww, 4, dst, stride, 0); +} + +void aom_smooth_predictor_4x8_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + __m128i pixels[3]; + load_pixel_w4(above, left, 8, pixels); + + __m128i wh[4], ww[2]; + load_weight_w4(sm_weight_arrays, 8, wh, ww); + + smooth_pred_4xh(pixels, wh, ww, 8, dst, stride, 0); +} + +void aom_smooth_predictor_4x16_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + __m128i pixels[3]; + load_pixel_w4(above, left, 16, pixels); + + __m128i wh[4], ww[2]; + load_weight_w4(sm_weight_arrays, 16, wh, ww); + + smooth_pred_4xh(pixels, wh, ww, 8, dst, stride, 0); + dst += stride << 3; + smooth_pred_4xh(pixels, &wh[2], ww, 8, dst, stride, 1); +} + +// pixels[0]: above and below_pred interleave vector, first half +// pixels[1]: above and below_pred interleave vector, second half +// pixels[2]: left vector +// pixels[3]: right_pred vector +// pixels[4]: above and below_pred interleave vector, first half +// pixels[5]: above and below_pred interleave vector, second half +// pixels[6]: left vector + 16 +// pixels[7]: right_pred vector +static INLINE void load_pixel_w8(const uint8_t *above, const uint8_t *left, + int height, __m128i *pixels) { + const __m128i zero = _mm_setzero_si128(); + const __m128i bp = _mm_set1_epi16((uint16_t)left[height - 1]); + __m128i d = _mm_loadl_epi64((const __m128i *)above); + d = _mm_unpacklo_epi8(d, zero); + pixels[0] = _mm_unpacklo_epi16(d, bp); + pixels[1] = _mm_unpackhi_epi16(d, bp); + + pixels[3] = _mm_set1_epi16((uint16_t)above[7]); + + if (height == 4) { + pixels[2] = _mm_cvtsi32_si128(((const uint32_t *)left)[0]); + } else if (height == 8) { + pixels[2] = _mm_loadl_epi64((const __m128i *)left); + } else if (height == 16) { + pixels[2] = _mm_load_si128((const __m128i *)left); + } else { + pixels[2] = _mm_load_si128((const __m128i *)left); + pixels[4] = pixels[0]; + pixels[5] = pixels[1]; + pixels[6] = _mm_load_si128((const __m128i *)(left + 16)); + pixels[7] = pixels[3]; + } +} + +// weight_h[0]: weight_h vector +// weight_h[1]: scale - weight_h vector +// weight_h[2]: same as [0], offset 8 +// weight_h[3]: same as [1], offset 8 +// weight_h[4]: same as [0], offset 16 +// weight_h[5]: same as [1], offset 16 +// weight_h[6]: same as [0], offset 24 +// weight_h[7]: same as [1], offset 24 +// weight_w[0]: weights_w and scale - weights_w interleave vector, first half +// weight_w[1]: weights_w and scale - weights_w interleave vector, second half +static INLINE void load_weight_w8(const uint8_t *weight_array, int height, + __m128i *weight_h, __m128i *weight_w) { + const __m128i zero = _mm_setzero_si128(); + const int we_offset = height < 8 ? 4 : 8; + __m128i we = _mm_loadu_si128((const __m128i *)&weight_array[we_offset]); + weight_h[0] = _mm_unpacklo_epi8(we, zero); + const __m128i d = _mm_set1_epi16((uint16_t)(1 << sm_weight_log2_scale)); + weight_h[1] = _mm_sub_epi16(d, weight_h[0]); + + if (height == 4) { + we = _mm_srli_si128(we, 4); + __m128i tmp1 = _mm_unpacklo_epi8(we, zero); + __m128i tmp2 = _mm_sub_epi16(d, tmp1); + weight_w[0] = _mm_unpacklo_epi16(tmp1, tmp2); + weight_w[1] = _mm_unpackhi_epi16(tmp1, tmp2); + } else { + weight_w[0] = _mm_unpacklo_epi16(weight_h[0], weight_h[1]); + weight_w[1] = _mm_unpackhi_epi16(weight_h[0], weight_h[1]); + } + + if (height == 16) { + we = _mm_loadu_si128((const __m128i *)&weight_array[16]); + weight_h[0] = _mm_unpacklo_epi8(we, zero); + weight_h[1] = _mm_sub_epi16(d, weight_h[0]); + weight_h[2] = _mm_unpackhi_epi8(we, zero); + weight_h[3] = _mm_sub_epi16(d, weight_h[2]); + } else if (height == 32) { + const __m128i weight_lo = + _mm_loadu_si128((const __m128i *)&weight_array[32]); + weight_h[0] = _mm_unpacklo_epi8(weight_lo, zero); + weight_h[1] = _mm_sub_epi16(d, weight_h[0]); + weight_h[2] = _mm_unpackhi_epi8(weight_lo, zero); + weight_h[3] = _mm_sub_epi16(d, weight_h[2]); + const __m128i weight_hi = + _mm_loadu_si128((const __m128i *)&weight_array[32 + 16]); + weight_h[4] = _mm_unpacklo_epi8(weight_hi, zero); + weight_h[5] = _mm_sub_epi16(d, weight_h[4]); + weight_h[6] = _mm_unpackhi_epi8(weight_hi, zero); + weight_h[7] = _mm_sub_epi16(d, weight_h[6]); + } +} + +static INLINE void smooth_pred_8xh(const __m128i *pixels, const __m128i *wh, + const __m128i *ww, int h, uint8_t *dst, + ptrdiff_t stride, int second_half) { + const __m128i round = _mm_set1_epi32((1 << sm_weight_log2_scale)); + const __m128i one = _mm_set1_epi16(1); + const __m128i inc = _mm_set1_epi16(0x202); + const __m128i gat = _mm_set_epi32(0, 0, 0xe0c0a08, 0x6040200); + + __m128i rep = second_half ? _mm_set1_epi16(0x8008) : _mm_set1_epi16(0x8000); + __m128i d = _mm_set1_epi16(0x100); + + int i; + for (i = 0; i < h; ++i) { + const __m128i wg_wg = _mm_shuffle_epi8(wh[0], d); + const __m128i sc_sc = _mm_shuffle_epi8(wh[1], d); + const __m128i wh_sc = _mm_unpacklo_epi16(wg_wg, sc_sc); + __m128i s0 = _mm_madd_epi16(pixels[0], wh_sc); + __m128i s1 = _mm_madd_epi16(pixels[1], wh_sc); + + __m128i b = _mm_shuffle_epi8(pixels[2], rep); + b = _mm_unpacklo_epi16(b, pixels[3]); + __m128i sum0 = _mm_madd_epi16(b, ww[0]); + __m128i sum1 = _mm_madd_epi16(b, ww[1]); + + s0 = _mm_add_epi32(s0, sum0); + s0 = _mm_add_epi32(s0, round); + s0 = _mm_srai_epi32(s0, 1 + sm_weight_log2_scale); + + s1 = _mm_add_epi32(s1, sum1); + s1 = _mm_add_epi32(s1, round); + s1 = _mm_srai_epi32(s1, 1 + sm_weight_log2_scale); + + sum0 = _mm_packus_epi16(s0, s1); + sum0 = _mm_shuffle_epi8(sum0, gat); + _mm_storel_epi64((__m128i *)dst, sum0); + dst += stride; + + rep = _mm_add_epi16(rep, one); + d = _mm_add_epi16(d, inc); + } +} + +void aom_smooth_predictor_8x4_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + __m128i pixels[4]; + load_pixel_w8(above, left, 4, pixels); + + __m128i wh[4], ww[2]; + load_weight_w8(sm_weight_arrays, 4, wh, ww); + + smooth_pred_8xh(pixels, wh, ww, 4, dst, stride, 0); +} + +void aom_smooth_predictor_8x8_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + __m128i pixels[4]; + load_pixel_w8(above, left, 8, pixels); + + __m128i wh[4], ww[2]; + load_weight_w8(sm_weight_arrays, 8, wh, ww); + + smooth_pred_8xh(pixels, wh, ww, 8, dst, stride, 0); +} + +void aom_smooth_predictor_8x16_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + __m128i pixels[4]; + load_pixel_w8(above, left, 16, pixels); + + __m128i wh[4], ww[2]; + load_weight_w8(sm_weight_arrays, 16, wh, ww); + + smooth_pred_8xh(pixels, wh, ww, 8, dst, stride, 0); + dst += stride << 3; + smooth_pred_8xh(pixels, &wh[2], ww, 8, dst, stride, 1); +} + +void aom_smooth_predictor_8x32_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + __m128i pixels[8]; + load_pixel_w8(above, left, 32, pixels); + + __m128i wh[8], ww[2]; + load_weight_w8(sm_weight_arrays, 32, wh, ww); + + smooth_pred_8xh(&pixels[0], wh, ww, 8, dst, stride, 0); + dst += stride << 3; + smooth_pred_8xh(&pixels[0], &wh[2], ww, 8, dst, stride, 1); + dst += stride << 3; + smooth_pred_8xh(&pixels[4], &wh[4], ww, 8, dst, stride, 0); + dst += stride << 3; + smooth_pred_8xh(&pixels[4], &wh[6], ww, 8, dst, stride, 1); +} + +static INLINE void smooth_predictor_wxh(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left, uint32_t bw, + uint32_t bh) { + const uint8_t *const sm_weights_w = sm_weight_arrays + bw; + const uint8_t *const sm_weights_h = sm_weight_arrays + bh; + const __m128i zero = _mm_setzero_si128(); + const __m128i scale_value = + _mm_set1_epi16((uint16_t)(1 << sm_weight_log2_scale)); + const __m128i bottom_left = _mm_cvtsi32_si128((uint32_t)left[bh - 1]); + const __m128i dup16 = _mm_set1_epi32(0x01000100); + const __m128i top_right = + _mm_shuffle_epi8(_mm_cvtsi32_si128((uint32_t)above[bw - 1]), dup16); + const __m128i gat = _mm_set_epi32(0, 0, 0xe0c0a08, 0x6040200); + const __m128i round = _mm_set1_epi32((uint16_t)(1 << sm_weight_log2_scale)); + + for (uint32_t y = 0; y < bh; ++y) { + const __m128i weights_y = _mm_cvtsi32_si128((uint32_t)sm_weights_h[y]); + const __m128i left_y = _mm_cvtsi32_si128((uint32_t)left[y]); + const __m128i scale_m_weights_y = _mm_sub_epi16(scale_value, weights_y); + __m128i pred_scaled_bl = _mm_mullo_epi16(scale_m_weights_y, bottom_left); + const __m128i wl_y = + _mm_shuffle_epi32(_mm_unpacklo_epi16(weights_y, left_y), 0); + pred_scaled_bl = _mm_add_epi32(pred_scaled_bl, round); + pred_scaled_bl = _mm_shuffle_epi32(pred_scaled_bl, 0); + + for (uint32_t x = 0; x < bw; x += 8) { + const __m128i top_x = _mm_loadl_epi64((const __m128i *)(above + x)); + const __m128i weights_x = + _mm_loadl_epi64((const __m128i *)(sm_weights_w + x)); + const __m128i tw_x = _mm_unpacklo_epi8(top_x, weights_x); + const __m128i tw_x_lo = _mm_unpacklo_epi8(tw_x, zero); + const __m128i tw_x_hi = _mm_unpackhi_epi8(tw_x, zero); + + __m128i pred_lo = _mm_madd_epi16(tw_x_lo, wl_y); + __m128i pred_hi = _mm_madd_epi16(tw_x_hi, wl_y); + + const __m128i scale_m_weights_x = + _mm_sub_epi16(scale_value, _mm_unpacklo_epi8(weights_x, zero)); + const __m128i swxtr = _mm_mullo_epi16(scale_m_weights_x, top_right); + const __m128i swxtr_lo = _mm_unpacklo_epi16(swxtr, zero); + const __m128i swxtr_hi = _mm_unpackhi_epi16(swxtr, zero); + + pred_lo = _mm_add_epi32(pred_lo, pred_scaled_bl); + pred_hi = _mm_add_epi32(pred_hi, pred_scaled_bl); + + pred_lo = _mm_add_epi32(pred_lo, swxtr_lo); + pred_hi = _mm_add_epi32(pred_hi, swxtr_hi); + + pred_lo = _mm_srai_epi32(pred_lo, (1 + sm_weight_log2_scale)); + pred_hi = _mm_srai_epi32(pred_hi, (1 + sm_weight_log2_scale)); + + __m128i pred = _mm_packus_epi16(pred_lo, pred_hi); + pred = _mm_shuffle_epi8(pred, gat); + _mm_storel_epi64((__m128i *)(dst + x), pred); + } + dst += stride; + } +} + +void aom_smooth_predictor_16x4_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + smooth_predictor_wxh(dst, stride, above, left, 16, 4); +} + +void aom_smooth_predictor_16x8_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + smooth_predictor_wxh(dst, stride, above, left, 16, 8); +} + +void aom_smooth_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + smooth_predictor_wxh(dst, stride, above, left, 16, 16); +} + +void aom_smooth_predictor_16x32_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + smooth_predictor_wxh(dst, stride, above, left, 16, 32); +} + +void aom_smooth_predictor_32x8_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + smooth_predictor_wxh(dst, stride, above, left, 32, 8); +} + +void aom_smooth_predictor_32x16_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + smooth_predictor_wxh(dst, stride, above, left, 32, 16); +} + +void aom_smooth_predictor_32x32_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + smooth_predictor_wxh(dst, stride, above, left, 32, 32); +} + +void aom_smooth_predictor_32x64_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + smooth_predictor_wxh(dst, stride, above, left, 32, 64); +} + +void aom_smooth_predictor_64x64_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + smooth_predictor_wxh(dst, stride, above, left, 64, 64); +} + +void aom_smooth_predictor_64x32_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + smooth_predictor_wxh(dst, stride, above, left, 64, 32); +} + +void aom_smooth_predictor_64x16_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + smooth_predictor_wxh(dst, stride, above, left, 64, 16); +} + +void aom_smooth_predictor_16x64_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + smooth_predictor_wxh(dst, stride, above, left, 16, 64); +} + +// ----------------------------------------------------------------------------- +// SMOOTH_V_PRED + +// pixels[0]: above and below_pred interleave vector +static INLINE void load_pixel_v_w4(const uint8_t *above, const uint8_t *left, + int height, __m128i *pixels) { + const __m128i zero = _mm_setzero_si128(); + __m128i d = _mm_cvtsi32_si128(((const uint32_t *)above)[0]); + const __m128i bp = _mm_set1_epi16((uint16_t)left[height - 1]); + d = _mm_unpacklo_epi8(d, zero); + pixels[0] = _mm_unpacklo_epi16(d, bp); +} + +// weights[0]: weights_h vector +// weights[1]: scale - weights_h vector +static INLINE void load_weight_v_w4(const uint8_t *weight_array, int height, + __m128i *weights) { + const __m128i zero = _mm_setzero_si128(); + const __m128i d = _mm_set1_epi16((uint16_t)(1 << sm_weight_log2_scale)); + + if (height == 4) { + const __m128i weight = + _mm_cvtsi32_si128(((const uint32_t *)weight_array)[1]); + weights[0] = _mm_unpacklo_epi8(weight, zero); + weights[1] = _mm_sub_epi16(d, weights[0]); + } else if (height == 8) { + const __m128i weight = _mm_loadl_epi64((const __m128i *)&weight_array[8]); + weights[0] = _mm_unpacklo_epi8(weight, zero); + weights[1] = _mm_sub_epi16(d, weights[0]); + } else { + const __m128i weight = _mm_loadu_si128((const __m128i *)&weight_array[16]); + weights[0] = _mm_unpacklo_epi8(weight, zero); + weights[1] = _mm_sub_epi16(d, weights[0]); + weights[2] = _mm_unpackhi_epi8(weight, zero); + weights[3] = _mm_sub_epi16(d, weights[2]); + } +} + +static INLINE void smooth_v_pred_4xh(const __m128i *pixel, + const __m128i *weight, int h, uint8_t *dst, + ptrdiff_t stride) { + const __m128i pred_round = _mm_set1_epi32((1 << (sm_weight_log2_scale - 1))); + const __m128i inc = _mm_set1_epi16(0x202); + const __m128i gat = _mm_set1_epi32(0xc080400); + __m128i d = _mm_set1_epi16(0x100); + + for (int i = 0; i < h; ++i) { + const __m128i wg_wg = _mm_shuffle_epi8(weight[0], d); + const __m128i sc_sc = _mm_shuffle_epi8(weight[1], d); + const __m128i wh_sc = _mm_unpacklo_epi16(wg_wg, sc_sc); + __m128i sum = _mm_madd_epi16(pixel[0], wh_sc); + sum = _mm_add_epi32(sum, pred_round); + sum = _mm_srai_epi32(sum, sm_weight_log2_scale); + sum = _mm_shuffle_epi8(sum, gat); + *(uint32_t *)dst = _mm_cvtsi128_si32(sum); + dst += stride; + d = _mm_add_epi16(d, inc); + } +} + +void aom_smooth_v_predictor_4x4_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + __m128i pixels; + load_pixel_v_w4(above, left, 4, &pixels); + + __m128i weights[2]; + load_weight_v_w4(sm_weight_arrays, 4, weights); + + smooth_v_pred_4xh(&pixels, weights, 4, dst, stride); +} + +void aom_smooth_v_predictor_4x8_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + __m128i pixels; + load_pixel_v_w4(above, left, 8, &pixels); + + __m128i weights[2]; + load_weight_v_w4(sm_weight_arrays, 8, weights); + + smooth_v_pred_4xh(&pixels, weights, 8, dst, stride); +} + +void aom_smooth_v_predictor_4x16_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + __m128i pixels; + load_pixel_v_w4(above, left, 16, &pixels); + + __m128i weights[4]; + load_weight_v_w4(sm_weight_arrays, 16, weights); + + smooth_v_pred_4xh(&pixels, weights, 8, dst, stride); + dst += stride << 3; + smooth_v_pred_4xh(&pixels, &weights[2], 8, dst, stride); +} + +// pixels[0]: above and below_pred interleave vector, first half +// pixels[1]: above and below_pred interleave vector, second half +static INLINE void load_pixel_v_w8(const uint8_t *above, const uint8_t *left, + int height, __m128i *pixels) { + const __m128i zero = _mm_setzero_si128(); + __m128i d = _mm_loadl_epi64((const __m128i *)above); + const __m128i bp = _mm_set1_epi16((uint16_t)left[height - 1]); + d = _mm_unpacklo_epi8(d, zero); + pixels[0] = _mm_unpacklo_epi16(d, bp); + pixels[1] = _mm_unpackhi_epi16(d, bp); +} + +// weight_h[0]: weight_h vector +// weight_h[1]: scale - weight_h vector +// weight_h[2]: same as [0], offset 8 +// weight_h[3]: same as [1], offset 8 +// weight_h[4]: same as [0], offset 16 +// weight_h[5]: same as [1], offset 16 +// weight_h[6]: same as [0], offset 24 +// weight_h[7]: same as [1], offset 24 +static INLINE void load_weight_v_w8(const uint8_t *weight_array, int height, + __m128i *weight_h) { + const __m128i zero = _mm_setzero_si128(); + const __m128i d = _mm_set1_epi16((uint16_t)(1 << sm_weight_log2_scale)); + + if (height < 16) { + const int offset = height < 8 ? 4 : 8; + const __m128i weight = + _mm_loadu_si128((const __m128i *)&weight_array[offset]); + weight_h[0] = _mm_unpacklo_epi8(weight, zero); + weight_h[1] = _mm_sub_epi16(d, weight_h[0]); + } else if (height == 16) { + const __m128i weight = _mm_loadu_si128((const __m128i *)&weight_array[16]); + weight_h[0] = _mm_unpacklo_epi8(weight, zero); + weight_h[1] = _mm_sub_epi16(d, weight_h[0]); + weight_h[2] = _mm_unpackhi_epi8(weight, zero); + weight_h[3] = _mm_sub_epi16(d, weight_h[2]); + } else { + const __m128i weight_lo = + _mm_loadu_si128((const __m128i *)&weight_array[32]); + weight_h[0] = _mm_unpacklo_epi8(weight_lo, zero); + weight_h[1] = _mm_sub_epi16(d, weight_h[0]); + weight_h[2] = _mm_unpackhi_epi8(weight_lo, zero); + weight_h[3] = _mm_sub_epi16(d, weight_h[2]); + const __m128i weight_hi = + _mm_loadu_si128((const __m128i *)&weight_array[32 + 16]); + weight_h[4] = _mm_unpacklo_epi8(weight_hi, zero); + weight_h[5] = _mm_sub_epi16(d, weight_h[4]); + weight_h[6] = _mm_unpackhi_epi8(weight_hi, zero); + weight_h[7] = _mm_sub_epi16(d, weight_h[6]); + } +} + +static INLINE void smooth_v_pred_8xh(const __m128i *pixels, const __m128i *wh, + int h, uint8_t *dst, ptrdiff_t stride) { + const __m128i pred_round = _mm_set1_epi32((1 << (sm_weight_log2_scale - 1))); + const __m128i inc = _mm_set1_epi16(0x202); + const __m128i gat = _mm_set_epi32(0, 0, 0xe0c0a08, 0x6040200); + __m128i d = _mm_set1_epi16(0x100); + + for (int i = 0; i < h; ++i) { + const __m128i wg_wg = _mm_shuffle_epi8(wh[0], d); + const __m128i sc_sc = _mm_shuffle_epi8(wh[1], d); + const __m128i wh_sc = _mm_unpacklo_epi16(wg_wg, sc_sc); + __m128i s0 = _mm_madd_epi16(pixels[0], wh_sc); + __m128i s1 = _mm_madd_epi16(pixels[1], wh_sc); + + s0 = _mm_add_epi32(s0, pred_round); + s0 = _mm_srai_epi32(s0, sm_weight_log2_scale); + + s1 = _mm_add_epi32(s1, pred_round); + s1 = _mm_srai_epi32(s1, sm_weight_log2_scale); + + __m128i sum01 = _mm_packus_epi16(s0, s1); + sum01 = _mm_shuffle_epi8(sum01, gat); + _mm_storel_epi64((__m128i *)dst, sum01); + dst += stride; + + d = _mm_add_epi16(d, inc); + } +} + +void aom_smooth_v_predictor_8x4_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + __m128i pixels[2]; + load_pixel_v_w8(above, left, 4, pixels); + + __m128i wh[2]; + load_weight_v_w8(sm_weight_arrays, 4, wh); + + smooth_v_pred_8xh(pixels, wh, 4, dst, stride); +} + +void aom_smooth_v_predictor_8x8_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + __m128i pixels[2]; + load_pixel_v_w8(above, left, 8, pixels); + + __m128i wh[2]; + load_weight_v_w8(sm_weight_arrays, 8, wh); + + smooth_v_pred_8xh(pixels, wh, 8, dst, stride); +} + +void aom_smooth_v_predictor_8x16_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + __m128i pixels[2]; + load_pixel_v_w8(above, left, 16, pixels); + + __m128i wh[4]; + load_weight_v_w8(sm_weight_arrays, 16, wh); + + smooth_v_pred_8xh(pixels, wh, 8, dst, stride); + dst += stride << 3; + smooth_v_pred_8xh(pixels, &wh[2], 8, dst, stride); +} + +void aom_smooth_v_predictor_8x32_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + __m128i pixels[2]; + load_pixel_v_w8(above, left, 32, pixels); + + __m128i wh[8]; + load_weight_v_w8(sm_weight_arrays, 32, wh); + + smooth_v_pred_8xh(pixels, &wh[0], 8, dst, stride); + dst += stride << 3; + smooth_v_pred_8xh(pixels, &wh[2], 8, dst, stride); + dst += stride << 3; + smooth_v_pred_8xh(pixels, &wh[4], 8, dst, stride); + dst += stride << 3; + smooth_v_pred_8xh(pixels, &wh[6], 8, dst, stride); +} + +static INLINE void smooth_v_predictor_wxh(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left, uint32_t bw, + uint32_t bh) { + const uint8_t *const sm_weights_h = sm_weight_arrays + bh; + const __m128i zero = _mm_setzero_si128(); + const __m128i scale_value = + _mm_set1_epi16((uint16_t)(1 << sm_weight_log2_scale)); + const __m128i dup16 = _mm_set1_epi32(0x01000100); + const __m128i bottom_left = + _mm_shuffle_epi8(_mm_cvtsi32_si128((uint32_t)left[bh - 1]), dup16); + const __m128i gat = _mm_set_epi32(0, 0, 0xe0c0a08, 0x6040200); + const __m128i round = + _mm_set1_epi32((uint16_t)(1 << (sm_weight_log2_scale - 1))); + + for (uint32_t y = 0; y < bh; ++y) { + const __m128i weights_y = _mm_cvtsi32_si128((uint32_t)sm_weights_h[y]); + const __m128i scale_m_weights_y = + _mm_shuffle_epi8(_mm_sub_epi16(scale_value, weights_y), dup16); + const __m128i wl_y = + _mm_shuffle_epi32(_mm_unpacklo_epi16(weights_y, bottom_left), 0); + + for (uint32_t x = 0; x < bw; x += 8) { + const __m128i top_x = _mm_loadl_epi64((const __m128i *)(above + x)); + // 8 -> 16 + const __m128i tw_x = _mm_unpacklo_epi8(top_x, zero); + const __m128i tw_x_lo = _mm_unpacklo_epi16(tw_x, scale_m_weights_y); + const __m128i tw_x_hi = _mm_unpackhi_epi16(tw_x, scale_m_weights_y); + // top_x * weights_y + scale_m_weights_y * bottom_left + __m128i pred_lo = _mm_madd_epi16(tw_x_lo, wl_y); + __m128i pred_hi = _mm_madd_epi16(tw_x_hi, wl_y); + + pred_lo = _mm_add_epi32(pred_lo, round); + pred_hi = _mm_add_epi32(pred_hi, round); + pred_lo = _mm_srai_epi32(pred_lo, sm_weight_log2_scale); + pred_hi = _mm_srai_epi32(pred_hi, sm_weight_log2_scale); + + __m128i pred = _mm_packus_epi16(pred_lo, pred_hi); + pred = _mm_shuffle_epi8(pred, gat); + _mm_storel_epi64((__m128i *)(dst + x), pred); + } + dst += stride; + } +} + +void aom_smooth_v_predictor_16x4_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + smooth_v_predictor_wxh(dst, stride, above, left, 16, 4); +} + +void aom_smooth_v_predictor_16x8_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + smooth_v_predictor_wxh(dst, stride, above, left, 16, 8); +} + +void aom_smooth_v_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + smooth_v_predictor_wxh(dst, stride, above, left, 16, 16); +} + +void aom_smooth_v_predictor_16x32_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + smooth_v_predictor_wxh(dst, stride, above, left, 16, 32); +} + +void aom_smooth_v_predictor_32x8_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + smooth_v_predictor_wxh(dst, stride, above, left, 32, 8); +} + +void aom_smooth_v_predictor_32x16_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + smooth_v_predictor_wxh(dst, stride, above, left, 32, 16); +} + +void aom_smooth_v_predictor_32x32_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + smooth_v_predictor_wxh(dst, stride, above, left, 32, 32); +} + +void aom_smooth_v_predictor_32x64_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + smooth_v_predictor_wxh(dst, stride, above, left, 32, 64); +} + +void aom_smooth_v_predictor_64x64_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + smooth_v_predictor_wxh(dst, stride, above, left, 64, 64); +} + +void aom_smooth_v_predictor_64x32_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + smooth_v_predictor_wxh(dst, stride, above, left, 64, 32); +} + +void aom_smooth_v_predictor_64x16_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + smooth_v_predictor_wxh(dst, stride, above, left, 64, 16); +} + +void aom_smooth_v_predictor_16x64_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + smooth_v_predictor_wxh(dst, stride, above, left, 16, 64); +} + +// ----------------------------------------------------------------------------- +// SMOOTH_H_PRED + +// pixels[0]: left vector +// pixels[1]: right_pred vector +static INLINE void load_pixel_h_w4(const uint8_t *above, const uint8_t *left, + int height, __m128i *pixels) { + if (height == 4) + pixels[0] = _mm_cvtsi32_si128(((const uint32_t *)left)[0]); + else if (height == 8) + pixels[0] = _mm_loadl_epi64(((const __m128i *)left)); + else + pixels[0] = _mm_loadu_si128(((const __m128i *)left)); + pixels[1] = _mm_set1_epi16((uint16_t)above[3]); +} + +// weights[0]: weights_w and scale - weights_w interleave vector +static INLINE void load_weight_h_w4(const uint8_t *weight_array, int height, + __m128i *weights) { + (void)height; + const __m128i t = _mm_loadu_si128((const __m128i *)&weight_array[4]); + const __m128i zero = _mm_setzero_si128(); + + const __m128i weights_0 = _mm_unpacklo_epi8(t, zero); + const __m128i d = _mm_set1_epi16((uint16_t)(1 << sm_weight_log2_scale)); + const __m128i weights_1 = _mm_sub_epi16(d, weights_0); + weights[0] = _mm_unpacklo_epi16(weights_0, weights_1); +} + +static INLINE void smooth_h_pred_4xh(const __m128i *pixel, + const __m128i *weight, int h, uint8_t *dst, + ptrdiff_t stride) { + const __m128i pred_round = _mm_set1_epi32((1 << (sm_weight_log2_scale - 1))); + const __m128i one = _mm_set1_epi16(1); + const __m128i gat = _mm_set1_epi32(0xc080400); + __m128i rep = _mm_set1_epi16(0x8000); + + for (int i = 0; i < h; ++i) { + __m128i b = _mm_shuffle_epi8(pixel[0], rep); + b = _mm_unpacklo_epi16(b, pixel[1]); + __m128i sum = _mm_madd_epi16(b, weight[0]); + + sum = _mm_add_epi32(sum, pred_round); + sum = _mm_srai_epi32(sum, sm_weight_log2_scale); + + sum = _mm_shuffle_epi8(sum, gat); + *(uint32_t *)dst = _mm_cvtsi128_si32(sum); + dst += stride; + + rep = _mm_add_epi16(rep, one); + } +} + +void aom_smooth_h_predictor_4x4_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + __m128i pixels[2]; + load_pixel_h_w4(above, left, 4, pixels); + + __m128i weights; + load_weight_h_w4(sm_weight_arrays, 4, &weights); + + smooth_h_pred_4xh(pixels, &weights, 4, dst, stride); +} + +void aom_smooth_h_predictor_4x8_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + __m128i pixels[2]; + load_pixel_h_w4(above, left, 8, pixels); + + __m128i weights; + load_weight_h_w4(sm_weight_arrays, 8, &weights); + + smooth_h_pred_4xh(pixels, &weights, 8, dst, stride); +} + +void aom_smooth_h_predictor_4x16_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + __m128i pixels[2]; + load_pixel_h_w4(above, left, 16, pixels); + + __m128i weights; + load_weight_h_w4(sm_weight_arrays, 8, &weights); + + smooth_h_pred_4xh(pixels, &weights, 8, dst, stride); + dst += stride << 3; + + pixels[0] = _mm_srli_si128(pixels[0], 8); + smooth_h_pred_4xh(pixels, &weights, 8, dst, stride); +} + +// pixels[0]: left vector +// pixels[1]: right_pred vector +// pixels[2]: left vector + 16 +// pixels[3]: right_pred vector +static INLINE void load_pixel_h_w8(const uint8_t *above, const uint8_t *left, + int height, __m128i *pixels) { + pixels[1] = _mm_set1_epi16((uint16_t)above[7]); + + if (height == 4) { + pixels[0] = _mm_cvtsi32_si128(((const uint32_t *)left)[0]); + } else if (height == 8) { + pixels[0] = _mm_loadl_epi64((const __m128i *)left); + } else if (height == 16) { + pixels[0] = _mm_load_si128((const __m128i *)left); + } else { + pixels[0] = _mm_load_si128((const __m128i *)left); + pixels[2] = _mm_load_si128((const __m128i *)(left + 16)); + pixels[3] = pixels[1]; + } +} + +// weight_w[0]: weights_w and scale - weights_w interleave vector, first half +// weight_w[1]: weights_w and scale - weights_w interleave vector, second half +static INLINE void load_weight_h_w8(const uint8_t *weight_array, int height, + __m128i *weight_w) { + (void)height; + const __m128i zero = _mm_setzero_si128(); + const __m128i d = _mm_set1_epi16((uint16_t)(1 << sm_weight_log2_scale)); + const __m128i we = _mm_loadu_si128((const __m128i *)&weight_array[8]); + const __m128i tmp1 = _mm_unpacklo_epi8(we, zero); + const __m128i tmp2 = _mm_sub_epi16(d, tmp1); + weight_w[0] = _mm_unpacklo_epi16(tmp1, tmp2); + weight_w[1] = _mm_unpackhi_epi16(tmp1, tmp2); +} + +static INLINE void smooth_h_pred_8xh(const __m128i *pixels, const __m128i *ww, + int h, uint8_t *dst, ptrdiff_t stride, + int second_half) { + const __m128i pred_round = _mm_set1_epi32((1 << (sm_weight_log2_scale - 1))); + const __m128i one = _mm_set1_epi16(1); + const __m128i gat = _mm_set_epi32(0, 0, 0xe0c0a08, 0x6040200); + __m128i rep = second_half ? _mm_set1_epi16(0x8008) : _mm_set1_epi16(0x8000); + + for (int i = 0; i < h; ++i) { + __m128i b = _mm_shuffle_epi8(pixels[0], rep); + b = _mm_unpacklo_epi16(b, pixels[1]); + __m128i sum0 = _mm_madd_epi16(b, ww[0]); + __m128i sum1 = _mm_madd_epi16(b, ww[1]); + + sum0 = _mm_add_epi32(sum0, pred_round); + sum0 = _mm_srai_epi32(sum0, sm_weight_log2_scale); + + sum1 = _mm_add_epi32(sum1, pred_round); + sum1 = _mm_srai_epi32(sum1, sm_weight_log2_scale); + + sum0 = _mm_packus_epi16(sum0, sum1); + sum0 = _mm_shuffle_epi8(sum0, gat); + _mm_storel_epi64((__m128i *)dst, sum0); + dst += stride; + + rep = _mm_add_epi16(rep, one); + } +} + +void aom_smooth_h_predictor_8x4_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + __m128i pixels[2]; + load_pixel_h_w8(above, left, 4, pixels); + + __m128i ww[2]; + load_weight_h_w8(sm_weight_arrays, 4, ww); + + smooth_h_pred_8xh(pixels, ww, 4, dst, stride, 0); +} + +void aom_smooth_h_predictor_8x8_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + __m128i pixels[2]; + load_pixel_h_w8(above, left, 8, pixels); + + __m128i ww[2]; + load_weight_h_w8(sm_weight_arrays, 8, ww); + + smooth_h_pred_8xh(pixels, ww, 8, dst, stride, 0); +} + +void aom_smooth_h_predictor_8x16_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + __m128i pixels[2]; + load_pixel_h_w8(above, left, 16, pixels); + + __m128i ww[2]; + load_weight_h_w8(sm_weight_arrays, 16, ww); + + smooth_h_pred_8xh(pixels, ww, 8, dst, stride, 0); + dst += stride << 3; + smooth_h_pred_8xh(pixels, ww, 8, dst, stride, 1); +} + +void aom_smooth_h_predictor_8x32_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + __m128i pixels[4]; + load_pixel_h_w8(above, left, 32, pixels); + + __m128i ww[2]; + load_weight_h_w8(sm_weight_arrays, 32, ww); + + smooth_h_pred_8xh(&pixels[0], ww, 8, dst, stride, 0); + dst += stride << 3; + smooth_h_pred_8xh(&pixels[0], ww, 8, dst, stride, 1); + dst += stride << 3; + smooth_h_pred_8xh(&pixels[2], ww, 8, dst, stride, 0); + dst += stride << 3; + smooth_h_pred_8xh(&pixels[2], ww, 8, dst, stride, 1); +} + +static INLINE void smooth_h_predictor_wxh(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left, uint32_t bw, + uint32_t bh) { + const uint8_t *const sm_weights_w = sm_weight_arrays + bw; + const __m128i zero = _mm_setzero_si128(); + const __m128i scale_value = + _mm_set1_epi16((uint16_t)(1 << sm_weight_log2_scale)); + const __m128i top_right = _mm_cvtsi32_si128((uint32_t)above[bw - 1]); + const __m128i gat = _mm_set_epi32(0, 0, 0xe0c0a08, 0x6040200); + const __m128i pred_round = _mm_set1_epi32((1 << (sm_weight_log2_scale - 1))); + + for (uint32_t y = 0; y < bh; ++y) { + const __m128i left_y = _mm_cvtsi32_si128((uint32_t)left[y]); + const __m128i tr_ly = + _mm_shuffle_epi32(_mm_unpacklo_epi16(top_right, left_y), 0); + + for (uint32_t x = 0; x < bw; x += 8) { + const __m128i weights_x = + _mm_loadl_epi64((const __m128i *)(sm_weights_w + x)); + const __m128i weights_xw = _mm_unpacklo_epi8(weights_x, zero); + const __m128i scale_m_weights_x = _mm_sub_epi16(scale_value, weights_xw); + const __m128i wx_lo = _mm_unpacklo_epi16(scale_m_weights_x, weights_xw); + const __m128i wx_hi = _mm_unpackhi_epi16(scale_m_weights_x, weights_xw); + __m128i pred_lo = _mm_madd_epi16(wx_lo, tr_ly); + __m128i pred_hi = _mm_madd_epi16(wx_hi, tr_ly); + + pred_lo = _mm_add_epi32(pred_lo, pred_round); + pred_hi = _mm_add_epi32(pred_hi, pred_round); + + pred_lo = _mm_srai_epi32(pred_lo, sm_weight_log2_scale); + pred_hi = _mm_srai_epi32(pred_hi, sm_weight_log2_scale); + + __m128i pred = _mm_packus_epi16(pred_lo, pred_hi); + pred = _mm_shuffle_epi8(pred, gat); + _mm_storel_epi64((__m128i *)(dst + x), pred); + } + dst += stride; + } +} + +void aom_smooth_h_predictor_16x4_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + smooth_h_predictor_wxh(dst, stride, above, left, 16, 4); +} + +void aom_smooth_h_predictor_16x8_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + smooth_h_predictor_wxh(dst, stride, above, left, 16, 8); +} + +void aom_smooth_h_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + smooth_h_predictor_wxh(dst, stride, above, left, 16, 16); +} + +void aom_smooth_h_predictor_16x32_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + smooth_h_predictor_wxh(dst, stride, above, left, 16, 32); +} + +void aom_smooth_h_predictor_16x64_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + smooth_h_predictor_wxh(dst, stride, above, left, 16, 64); +} + +void aom_smooth_h_predictor_32x8_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + smooth_h_predictor_wxh(dst, stride, above, left, 32, 8); +} + +void aom_smooth_h_predictor_32x16_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + smooth_h_predictor_wxh(dst, stride, above, left, 32, 16); +} + +void aom_smooth_h_predictor_32x32_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + smooth_h_predictor_wxh(dst, stride, above, left, 32, 32); +} + +void aom_smooth_h_predictor_32x64_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + smooth_h_predictor_wxh(dst, stride, above, left, 32, 64); +} + +void aom_smooth_h_predictor_64x64_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + smooth_h_predictor_wxh(dst, stride, above, left, 64, 64); +} + +void aom_smooth_h_predictor_64x32_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + smooth_h_predictor_wxh(dst, stride, above, left, 64, 32); +} + +void aom_smooth_h_predictor_64x16_ssse3(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, + const uint8_t *left) { + smooth_h_predictor_wxh(dst, stride, above, left, 64, 16); +} diff --git a/third_party/aom/aom_dsp/x86/inv_wht_sse2.asm b/third_party/aom/aom_dsp/x86/inv_wht_sse2.asm new file mode 100644 index 0000000000..0bc841a7a4 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/inv_wht_sse2.asm @@ -0,0 +1,107 @@ +; +; Copyright (c) 2016, Alliance for Open Media. All rights reserved +; +; This source code is subject to the terms of the BSD 2 Clause License and +; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License +; was not distributed with this source code in the LICENSE file, you can +; obtain it at www.aomedia.org/license/software. If the Alliance for Open +; Media Patent License 1.0 was not distributed with this source code in the +; PATENTS file, you can obtain it at www.aomedia.org/license/patent. +; + +; + +%include "third_party/x86inc/x86inc.asm" + +SECTION .text + +%macro REORDER_INPUTS 0 + ; a c d b to a b c d + SWAP 1, 3, 2 +%endmacro + +%macro TRANSFORM_COLS 0 + ; input: + ; m0 a + ; m1 b + ; m2 c + ; m3 d + paddw m0, m2 + psubw m3, m1 + + ; wide subtract + punpcklwd m4, m0 + punpcklwd m5, m3 + psrad m4, 16 + psrad m5, 16 + psubd m4, m5 + psrad m4, 1 + packssdw m4, m4 ; e + + psubw m5, m4, m1 ; b + psubw m4, m2 ; c + psubw m0, m5 + paddw m3, m4 + ; m0 a + SWAP 1, 5 ; m1 b + SWAP 2, 4 ; m2 c + ; m3 d +%endmacro + +%macro TRANSPOSE_4X4 0 + punpcklwd m0, m2 + punpcklwd m1, m3 + mova m2, m0 + punpcklwd m0, m1 + punpckhwd m2, m1 + pshufd m1, m0, 0x0e + pshufd m3, m2, 0x0e +%endmacro + +; transpose a 4x4 int16 matrix in xmm0 and xmm1 to the bottom half of xmm0-xmm3 +%macro TRANSPOSE_4X4_WIDE 0 + mova m3, m0 + punpcklwd m0, m1 + punpckhwd m3, m1 + mova m2, m0 + punpcklwd m0, m3 + punpckhwd m2, m3 + pshufd m1, m0, 0x0e + pshufd m3, m2, 0x0e +%endmacro + +%macro ADD_STORE_4P_2X 5 ; src1, src2, tmp1, tmp2, zero + movd m%3, [outputq] + movd m%4, [outputq + strideq] + punpcklbw m%3, m%5 + punpcklbw m%4, m%5 + paddw m%1, m%3 + paddw m%2, m%4 + packuswb m%1, m%5 + packuswb m%2, m%5 + movd [outputq], m%1 + movd [outputq + strideq], m%2 +%endmacro + +INIT_XMM sse2 +cglobal iwht4x4_16_add, 3, 3, 7, input, output, stride + mova m0, [inputq + 0] + packssdw m0, [inputq + 16] + mova m1, [inputq + 32] + packssdw m1, [inputq + 48] + psraw m0, 2 + psraw m1, 2 + + TRANSPOSE_4X4_WIDE + REORDER_INPUTS + TRANSFORM_COLS + TRANSPOSE_4X4 + REORDER_INPUTS + TRANSFORM_COLS + + pxor m4, m4 + ADD_STORE_4P_2X 0, 1, 5, 6, 4 + lea outputq, [outputq + 2 * strideq] + ADD_STORE_4P_2X 2, 3, 5, 6, 4 + + RET diff --git a/third_party/aom/aom_dsp/x86/jnt_sad_ssse3.c b/third_party/aom/aom_dsp/x86/jnt_sad_ssse3.c new file mode 100644 index 0000000000..c3c88245a4 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/jnt_sad_ssse3.c @@ -0,0 +1,238 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include +#include // SSE2 +#include + +#include "config/aom_config.h" +#include "config/aom_dsp_rtcd.h" +#include "config/av1_rtcd.h" + +#include "aom_dsp/x86/synonyms.h" + +unsigned int aom_sad4xh_sse2(const uint8_t *a, int a_stride, const uint8_t *b, + int b_stride, int width, int height) { + int i; + assert(width == 4); + (void)width; + + __m128i sad = _mm_setzero_si128(); + for (i = 0; i < height; i += 4) { + __m128i x0 = xx_loadl_32(a + 0 * a_stride); + __m128i x1 = xx_loadl_32(a + 1 * a_stride); + __m128i x2 = xx_loadl_32(a + 2 * a_stride); + __m128i x3 = xx_loadl_32(a + 3 * a_stride); + __m128i x_lo = _mm_unpacklo_epi32(x0, x1); + __m128i x_hi = _mm_unpacklo_epi32(x2, x3); + + __m128i x = _mm_unpacklo_epi64(x_lo, x_hi); + + x0 = xx_loadl_32(b + 0 * b_stride); + x1 = xx_loadl_32(b + 1 * b_stride); + x2 = xx_loadl_32(b + 2 * b_stride); + x3 = xx_loadl_32(b + 3 * b_stride); + x_lo = _mm_unpacklo_epi32(x0, x1); + x_hi = _mm_unpacklo_epi32(x2, x3); + + __m128i y = _mm_unpacklo_epi64(x_lo, x_hi); + + __m128i sad4x4 = _mm_sad_epu8(x, y); + sad = _mm_add_epi32(sad, sad4x4); + + a += 4 * a_stride; + b += 4 * b_stride; + } + + // At this point, we have two 32-bit partial SADs at bit[0:31] and [64:95]. + const unsigned int res = + _mm_cvtsi128_si32(sad) + _mm_cvtsi128_si32(_mm_srli_si128(sad, 8)); + + return res; +} + +unsigned int aom_sad8xh_sse2(const uint8_t *a, int a_stride, const uint8_t *b, + int b_stride, int width, int height) { + int i; + assert(width == 8); + (void)width; + + __m128i sad = _mm_setzero_si128(); + for (i = 0; i < height; i += 2) { + __m128i x0 = xx_loadl_64(a + 0 * a_stride); + __m128i x1 = xx_loadl_64(a + 1 * a_stride); + + __m128i x = _mm_unpacklo_epi64(x0, x1); + + x0 = xx_loadl_64(b + 0 * b_stride); + x1 = xx_loadl_64(b + 1 * b_stride); + + __m128i y = _mm_unpacklo_epi64(x0, x1); + + __m128i sad8x2 = _mm_sad_epu8(x, y); + sad = _mm_add_epi32(sad, sad8x2); + + a += 2 * a_stride; + b += 2 * b_stride; + } + + const unsigned int res = + _mm_cvtsi128_si32(sad) + _mm_cvtsi128_si32(_mm_srli_si128(sad, 8)); + + return res; +} + +unsigned int aom_sad16xh_sse2(const uint8_t *a, int a_stride, const uint8_t *b, + int b_stride, int width, int height) { + int i; + assert(width == 16); + (void)width; + + __m128i sad = _mm_setzero_si128(); + for (i = 0; i < height; ++i) { + __m128i x = xx_loadu_128(a); + __m128i y = xx_loadu_128(b); + + __m128i sad16x1 = _mm_sad_epu8(x, y); + sad = _mm_add_epi32(sad, sad16x1); + + a += a_stride; + b += b_stride; + } + + const unsigned int res = + _mm_cvtsi128_si32(sad) + _mm_cvtsi128_si32(_mm_srli_si128(sad, 8)); + + return res; +} + +unsigned int aom_sad32xh_sse2(const uint8_t *a, int a_stride, const uint8_t *b, + int b_stride, int width, int height) { + int i, j; + assert(width == 32); + (void)width; + + __m128i sad = _mm_setzero_si128(); + for (i = 0; i < height; ++i) { + for (j = 0; j < 2; ++j) { + __m128i x = xx_loadu_128(a + j * 16); + __m128i y = xx_loadu_128(b + j * 16); + + __m128i sad32_half = _mm_sad_epu8(x, y); + sad = _mm_add_epi32(sad, sad32_half); + } + + a += a_stride; + b += b_stride; + } + + const unsigned int res = + _mm_cvtsi128_si32(sad) + _mm_cvtsi128_si32(_mm_srli_si128(sad, 8)); + + return res; +} + +unsigned int aom_sad64xh_sse2(const uint8_t *a, int a_stride, const uint8_t *b, + int b_stride, int width, int height) { + int i, j; + assert(width == 64); + (void)width; + + __m128i sad = _mm_setzero_si128(); + for (i = 0; i < height; ++i) { + for (j = 0; j < 4; ++j) { + __m128i x = xx_loadu_128(a + j * 16); + __m128i y = xx_loadu_128(b + j * 16); + + __m128i sad64_quarter = _mm_sad_epu8(x, y); + sad = _mm_add_epi32(sad, sad64_quarter); + } + + a += a_stride; + b += b_stride; + } + + const unsigned int res = + _mm_cvtsi128_si32(sad) + _mm_cvtsi128_si32(_mm_srli_si128(sad, 8)); + + return res; +} + +unsigned int aom_sad128xh_sse2(const uint8_t *a, int a_stride, const uint8_t *b, + int b_stride, int width, int height) { + int i, j; + assert(width == 128); + (void)width; + + __m128i sad = _mm_setzero_si128(); + for (i = 0; i < height; ++i) { + for (j = 0; j < 8; ++j) { + __m128i x = xx_loadu_128(a + j * 16); + __m128i y = xx_loadu_128(b + j * 16); + + __m128i sad64_quarter = _mm_sad_epu8(x, y); + sad = _mm_add_epi32(sad, sad64_quarter); + } + + a += a_stride; + b += b_stride; + } + + const unsigned int res = + _mm_cvtsi128_si32(sad) + _mm_cvtsi128_si32(_mm_srli_si128(sad, 8)); + + return res; +} + +#define jnt_sadMxN_sse2(m, n) \ + unsigned int aom_jnt_sad##m##x##n##_avg_ssse3( \ + const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, \ + const uint8_t *second_pred, const JNT_COMP_PARAMS *jcp_param) { \ + uint8_t comp_pred[m * n]; \ + aom_jnt_comp_avg_pred(comp_pred, second_pred, m, n, ref, ref_stride, \ + jcp_param); \ + return aom_sad##m##xh_sse2(src, src_stride, comp_pred, m, m, n); \ + } + +#define jnt_sadMxN_avx2(m, n) \ + unsigned int aom_jnt_sad##m##x##n##_avg_avx2( \ + const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, \ + const uint8_t *second_pred, const JNT_COMP_PARAMS *jcp_param) { \ + uint8_t comp_pred[m * n]; \ + aom_jnt_comp_avg_pred(comp_pred, second_pred, m, n, ref, ref_stride, \ + jcp_param); \ + return aom_sad##m##xh_avx2(src, src_stride, comp_pred, m, m, n); \ + } + +/* clang-format off */ +jnt_sadMxN_sse2(128, 128) +jnt_sadMxN_sse2(128, 64) +jnt_sadMxN_sse2(64, 128) +jnt_sadMxN_sse2(64, 64) +jnt_sadMxN_sse2(64, 32) +jnt_sadMxN_sse2(32, 64) +jnt_sadMxN_sse2(32, 32) +jnt_sadMxN_sse2(32, 16) +jnt_sadMxN_sse2(16, 32) +jnt_sadMxN_sse2(16, 16) +jnt_sadMxN_sse2(16, 8) +jnt_sadMxN_sse2(8, 16) +jnt_sadMxN_sse2(8, 8) +jnt_sadMxN_sse2(8, 4) +jnt_sadMxN_sse2(4, 8) +jnt_sadMxN_sse2(4, 4) +jnt_sadMxN_sse2(4, 16) +jnt_sadMxN_sse2(16, 4) +jnt_sadMxN_sse2(8, 32) +jnt_sadMxN_sse2(32, 8) +jnt_sadMxN_sse2(16, 64) +jnt_sadMxN_sse2(64, 16) + /* clang-format on */ diff --git a/third_party/aom/aom_dsp/x86/jnt_variance_ssse3.c b/third_party/aom/aom_dsp/x86/jnt_variance_ssse3.c new file mode 100644 index 0000000000..f9a41a210b --- /dev/null +++ b/third_party/aom/aom_dsp/x86/jnt_variance_ssse3.c @@ -0,0 +1,192 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include +#include // SSE2 +#include + +#include "config/aom_config.h" +#include "config/aom_dsp_rtcd.h" +#include "config/av1_rtcd.h" + +#include "aom_dsp/x86/synonyms.h" + +void aom_var_filter_block2d_bil_first_pass_ssse3( + const uint8_t *a, uint16_t *b, unsigned int src_pixels_per_line, + unsigned int pixel_step, unsigned int output_height, + unsigned int output_width, const uint8_t *filter); + +void aom_var_filter_block2d_bil_second_pass_ssse3( + const uint16_t *a, uint8_t *b, unsigned int src_pixels_per_line, + unsigned int pixel_step, unsigned int output_height, + unsigned int output_width, const uint8_t *filter); + +static INLINE void compute_jnt_comp_avg(__m128i *p0, __m128i *p1, + const __m128i *w, const __m128i *r, + void *const result) { + __m128i p_lo = _mm_unpacklo_epi8(*p0, *p1); + __m128i mult_lo = _mm_maddubs_epi16(p_lo, *w); + __m128i round_lo = _mm_add_epi16(mult_lo, *r); + __m128i shift_lo = _mm_srai_epi16(round_lo, DIST_PRECISION_BITS); + + __m128i p_hi = _mm_unpackhi_epi8(*p0, *p1); + __m128i mult_hi = _mm_maddubs_epi16(p_hi, *w); + __m128i round_hi = _mm_add_epi16(mult_hi, *r); + __m128i shift_hi = _mm_srai_epi16(round_hi, DIST_PRECISION_BITS); + + xx_storeu_128(result, _mm_packus_epi16(shift_lo, shift_hi)); +} + +void aom_jnt_comp_avg_pred_ssse3(uint8_t *comp_pred, const uint8_t *pred, + int width, int height, const uint8_t *ref, + int ref_stride, + const JNT_COMP_PARAMS *jcp_param) { + int i; + const uint8_t w0 = (uint8_t)jcp_param->fwd_offset; + const uint8_t w1 = (uint8_t)jcp_param->bck_offset; + const __m128i w = _mm_set_epi8(w1, w0, w1, w0, w1, w0, w1, w0, w1, w0, w1, w0, + w1, w0, w1, w0); + const uint16_t round = ((1 << DIST_PRECISION_BITS) >> 1); + const __m128i r = + _mm_set_epi16(round, round, round, round, round, round, round, round); + + if (width >= 16) { + // Read 16 pixels one row at a time + assert(!(width & 15)); + for (i = 0; i < height; ++i) { + int j; + for (j = 0; j < width; j += 16) { + __m128i p0 = xx_loadu_128(ref); + __m128i p1 = xx_loadu_128(pred); + + compute_jnt_comp_avg(&p0, &p1, &w, &r, comp_pred); + + comp_pred += 16; + pred += 16; + ref += 16; + } + ref += ref_stride - width; + } + } else if (width >= 8) { + // Read 8 pixels two row at a time + assert(!(width & 7)); + assert(!(width & 1)); + for (i = 0; i < height; i += 2) { + __m128i p0_0 = xx_loadl_64(ref + 0 * ref_stride); + __m128i p0_1 = xx_loadl_64(ref + 1 * ref_stride); + __m128i p0 = _mm_unpacklo_epi64(p0_0, p0_1); + __m128i p1 = xx_loadu_128(pred); + + compute_jnt_comp_avg(&p0, &p1, &w, &r, comp_pred); + + comp_pred += 16; + pred += 16; + ref += 2 * ref_stride; + } + } else { + // Read 4 pixels four row at a time + assert(!(width & 3)); + assert(!(height & 3)); + for (i = 0; i < height; i += 4) { + const uint8_t *row0 = ref + 0 * ref_stride; + const uint8_t *row1 = ref + 1 * ref_stride; + const uint8_t *row2 = ref + 2 * ref_stride; + const uint8_t *row3 = ref + 3 * ref_stride; + + __m128i p0 = + _mm_setr_epi8(row0[0], row0[1], row0[2], row0[3], row1[0], row1[1], + row1[2], row1[3], row2[0], row2[1], row2[2], row2[3], + row3[0], row3[1], row3[2], row3[3]); + __m128i p1 = xx_loadu_128(pred); + + compute_jnt_comp_avg(&p0, &p1, &w, &r, comp_pred); + + comp_pred += 16; + pred += 16; + ref += 4 * ref_stride; + } + } +} + +void aom_jnt_comp_avg_upsampled_pred_ssse3( + MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col, + const MV *const mv, uint8_t *comp_pred, const uint8_t *pred, int width, + int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref, + int ref_stride, const JNT_COMP_PARAMS *jcp_param, int subpel_search) { + int n; + int i; + aom_upsampled_pred(xd, cm, mi_row, mi_col, mv, comp_pred, width, height, + subpel_x_q3, subpel_y_q3, ref, ref_stride, subpel_search); + /*The total number of pixels must be a multiple of 16 (e.g., 4x4).*/ + assert(!(width * height & 15)); + n = width * height >> 4; + + const uint8_t w0 = (uint8_t)jcp_param->fwd_offset; + const uint8_t w1 = (uint8_t)jcp_param->bck_offset; + const __m128i w = _mm_set_epi8(w1, w0, w1, w0, w1, w0, w1, w0, w1, w0, w1, w0, + w1, w0, w1, w0); + const uint16_t round = ((1 << DIST_PRECISION_BITS) >> 1); + const __m128i r = + _mm_set_epi16(round, round, round, round, round, round, round, round); + + for (i = 0; i < n; i++) { + __m128i p0 = xx_loadu_128(comp_pred); + __m128i p1 = xx_loadu_128(pred); + + compute_jnt_comp_avg(&p0, &p1, &w, &r, comp_pred); + + comp_pred += 16; + pred += 16; + } +} + +#define JNT_SUBPIX_AVG_VAR(W, H) \ + uint32_t aom_jnt_sub_pixel_avg_variance##W##x##H##_ssse3( \ + const uint8_t *a, int a_stride, int xoffset, int yoffset, \ + const uint8_t *b, int b_stride, uint32_t *sse, \ + const uint8_t *second_pred, const JNT_COMP_PARAMS *jcp_param) { \ + uint16_t fdata3[(H + 1) * W]; \ + uint8_t temp2[H * W]; \ + DECLARE_ALIGNED(16, uint8_t, temp3[H * W]); \ + \ + aom_var_filter_block2d_bil_first_pass_ssse3( \ + a, fdata3, a_stride, 1, H + 1, W, bilinear_filters_2t[xoffset]); \ + aom_var_filter_block2d_bil_second_pass_ssse3( \ + fdata3, temp2, W, W, H, W, bilinear_filters_2t[yoffset]); \ + \ + aom_jnt_comp_avg_pred_ssse3(temp3, second_pred, W, H, temp2, W, \ + jcp_param); \ + \ + return aom_variance##W##x##H(temp3, W, b, b_stride, sse); \ + } + +JNT_SUBPIX_AVG_VAR(128, 128) +JNT_SUBPIX_AVG_VAR(128, 64) +JNT_SUBPIX_AVG_VAR(64, 128) +JNT_SUBPIX_AVG_VAR(64, 64) +JNT_SUBPIX_AVG_VAR(64, 32) +JNT_SUBPIX_AVG_VAR(32, 64) +JNT_SUBPIX_AVG_VAR(32, 32) +JNT_SUBPIX_AVG_VAR(32, 16) +JNT_SUBPIX_AVG_VAR(16, 32) +JNT_SUBPIX_AVG_VAR(16, 16) +JNT_SUBPIX_AVG_VAR(16, 8) +JNT_SUBPIX_AVG_VAR(8, 16) +JNT_SUBPIX_AVG_VAR(8, 8) +JNT_SUBPIX_AVG_VAR(8, 4) +JNT_SUBPIX_AVG_VAR(4, 8) +JNT_SUBPIX_AVG_VAR(4, 4) +JNT_SUBPIX_AVG_VAR(4, 16) +JNT_SUBPIX_AVG_VAR(16, 4) +JNT_SUBPIX_AVG_VAR(8, 32) +JNT_SUBPIX_AVG_VAR(32, 8) +JNT_SUBPIX_AVG_VAR(16, 64) +JNT_SUBPIX_AVG_VAR(64, 16) diff --git a/third_party/aom/aom_dsp/x86/loopfilter_sse2.c b/third_party/aom/aom_dsp/x86/loopfilter_sse2.c new file mode 100644 index 0000000000..9d88b5e493 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/loopfilter_sse2.c @@ -0,0 +1,2385 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include // SSE2 + +#include "config/aom_dsp_rtcd.h" + +#include "aom_dsp/x86/synonyms.h" +#include "aom_ports/mem.h" +#include "aom_ports/emmintrin_compat.h" + +static INLINE __m128i abs_diff(__m128i a, __m128i b) { + return _mm_or_si128(_mm_subs_epu8(a, b), _mm_subs_epu8(b, a)); +} + +static INLINE void transpose4x8_8x4_low_sse2(__m128i *x0, __m128i *x1, + __m128i *x2, __m128i *x3, + __m128i *d0, __m128i *d1, + __m128i *d2, __m128i *d3) { + // input + // x0 00 01 02 03 04 05 06 07 xx xx xx xx xx xx xx xx + // x1 10 11 12 13 14 15 16 17 xx xx xx xx xx xx xx xx + // x2 20 21 22 23 24 25 26 27 xx xx xx xx xx xx xx xx + // x3 30 31 32 33 34 35 36 37 xx xx xx xx xx xx xx xx + // output + // 00 10 20 30 xx xx xx xx xx xx xx xx xx xx xx xx + // 01 11 21 31 xx xx xx xx xx xx xx xx xx xx xx xx + // 02 12 22 32 xx xx xx xx xx xx xx xx xx xx xx xx + // 03 13 23 33 xx xx xx xx xx xx xx xx xx xx xx xx + + __m128i w0, w1; + + w0 = _mm_unpacklo_epi8( + *x0, *x1); // 00 10 01 11 02 12 03 13 04 14 05 15 06 16 07 17 + w1 = _mm_unpacklo_epi8( + *x2, *x3); // 20 30 21 31 22 32 23 33 24 34 25 35 26 36 27 37 + + *d0 = _mm_unpacklo_epi16( + w0, w1); // 00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33 + + *d1 = _mm_srli_si128(*d0, + 4); // 01 11 21 31 xx xx xx xx xx xx xx xx xx xx xx xx + *d2 = _mm_srli_si128(*d0, + 8); // 02 12 22 32 xx xx xx xx xx xx xx xx xx xx xx xx + *d3 = _mm_srli_si128(*d0, + 12); // 03 13 23 33 xx xx xx xx xx xx xx xx xx xx xx xx +} + +static INLINE void transpose4x8_8x4_sse2(__m128i *x0, __m128i *x1, __m128i *x2, + __m128i *x3, __m128i *d0, __m128i *d1, + __m128i *d2, __m128i *d3, __m128i *d4, + __m128i *d5, __m128i *d6, + __m128i *d7) { + // input + // x0 00 01 02 03 04 05 06 07 xx xx xx xx xx xx xx xx + // x1 10 11 12 13 14 15 16 17 xx xx xx xx xx xx xx xx + // x2 20 21 22 23 24 25 26 27 xx xx xx xx xx xx xx xx + // x3 30 31 32 33 34 35 36 37 xx xx xx xx xx xx xx xx + // output + // 00 10 20 30 xx xx xx xx xx xx xx xx xx xx xx xx + // 01 11 21 31 xx xx xx xx xx xx xx xx xx xx xx xx + // 02 12 22 32 xx xx xx xx xx xx xx xx xx xx xx xx + // 03 13 23 33 xx xx xx xx xx xx xx xx xx xx xx xx + // 04 14 24 34 xx xx xx xx xx xx xx xx xx xx xx xx + // 05 15 25 35 xx xx xx xx xx xx xx xx xx xx xx xx + // 06 16 26 36 xx xx xx xx xx xx xx xx xx xx xx xx + // 07 17 27 37 xx xx xx xx xx xx xx xx xx xx xx xx + + __m128i w0, w1, ww0, ww1; + + w0 = _mm_unpacklo_epi8( + *x0, *x1); // 00 10 01 11 02 12 03 13 04 14 05 15 06 16 07 17 + w1 = _mm_unpacklo_epi8( + *x2, *x3); // 20 30 21 31 22 32 23 33 24 34 25 35 26 36 27 37 + + ww0 = _mm_unpacklo_epi16( + w0, w1); // 00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33 + ww1 = _mm_unpackhi_epi16( + w0, w1); // 04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37 + + *d0 = ww0; // 00 10 20 30 xx xx xx xx xx xx xx xx xx xx xx xx + *d1 = _mm_srli_si128(ww0, + 4); // 01 11 21 31 xx xx xx xx xx xx xx xx xx xx xx xx + *d2 = _mm_srli_si128(ww0, + 8); // 02 12 22 32 xx xx xx xx xx xx xx xx xx xx xx xx + *d3 = _mm_srli_si128(ww0, + 12); // 03 13 23 33 xx xx xx xx xx xx xx xx xx xx xx xx + + *d4 = ww1; // 04 14 24 34 xx xx xx xx xx xx xx xx xx xx xx xx + *d5 = _mm_srli_si128(ww1, + 4); // 05 15 25 35 xx xx xx xx xx xx xx xx xx xx xx xx + *d6 = _mm_srli_si128(ww1, + 8); // 06 16 26 36 xx xx xx xx xx xx xx xx xx xx xx xx + *d7 = _mm_srli_si128(ww1, + 12); // 07 17 27 37 xx xx xx xx xx xx xx xx xx xx xx xx +} + +static INLINE void transpose8x8_low_sse2(__m128i *x0, __m128i *x1, __m128i *x2, + __m128i *x3, __m128i *x4, __m128i *x5, + __m128i *x6, __m128i *x7, __m128i *d0, + __m128i *d1, __m128i *d2, + __m128i *d3) { + // input + // x0 00 01 02 03 04 05 06 07 + // x1 10 11 12 13 14 15 16 17 + // x2 20 21 22 23 24 25 26 27 + // x3 30 31 32 33 34 35 36 37 + // x4 40 41 42 43 44 45 46 47 + // x5 50 51 52 53 54 55 56 57 + // x6 60 61 62 63 64 65 66 67 + // x7 70 71 72 73 74 75 76 77 + // output + // d0 00 10 20 30 40 50 60 70 xx xx xx xx xx xx xx + // d1 01 11 21 31 41 51 61 71 xx xx xx xx xx xx xx xx + // d2 02 12 22 32 42 52 62 72 xx xx xx xx xx xx xx xx + // d3 03 13 23 33 43 53 63 73 xx xx xx xx xx xx xx xx + + __m128i w0, w1, w2, w3, w4, w5; + + w0 = _mm_unpacklo_epi8( + *x0, *x1); // 00 10 01 11 02 12 03 13 04 14 05 15 06 16 07 17 + + w1 = _mm_unpacklo_epi8( + *x2, *x3); // 20 30 21 31 22 32 23 33 24 34 25 35 26 36 27 37 + + w2 = _mm_unpacklo_epi8( + *x4, *x5); // 40 50 41 51 42 52 43 53 44 54 45 55 46 56 47 57 + + w3 = _mm_unpacklo_epi8( + *x6, *x7); // 60 70 61 71 62 72 63 73 64 74 65 75 66 76 67 77 + + w4 = _mm_unpacklo_epi16( + w0, w1); // 00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33 + w5 = _mm_unpacklo_epi16( + w2, w3); // 40 50 60 70 41 51 61 71 42 52 62 72 43 53 63 73 + + *d0 = _mm_unpacklo_epi32( + w4, w5); // 00 10 20 30 40 50 60 70 01 11 21 31 41 51 61 71 + *d1 = _mm_srli_si128(*d0, 8); + *d2 = _mm_unpackhi_epi32( + w4, w5); // 02 12 22 32 42 52 62 72 03 13 23 33 43 53 63 73 + *d3 = _mm_srli_si128(*d2, 8); +} + +static INLINE void transpose8x8_sse2(__m128i *x0, __m128i *x1, __m128i *x2, + __m128i *x3, __m128i *x4, __m128i *x5, + __m128i *x6, __m128i *x7, __m128i *d0d1, + __m128i *d2d3, __m128i *d4d5, + __m128i *d6d7) { + __m128i w0, w1, w2, w3, w4, w5, w6, w7; + // x0 00 01 02 03 04 05 06 07 + // x1 10 11 12 13 14 15 16 17 + w0 = _mm_unpacklo_epi8( + *x0, *x1); // 00 10 01 11 02 12 03 13 04 14 05 15 06 16 07 17 + + // x2 20 21 22 23 24 25 26 27 + // x3 30 31 32 33 34 35 36 37 + w1 = _mm_unpacklo_epi8( + *x2, *x3); // 20 30 21 31 22 32 23 33 24 34 25 35 26 36 27 37 + + // x4 40 41 42 43 44 45 46 47 + // x5 50 51 52 53 54 55 56 57 + w2 = _mm_unpacklo_epi8( + *x4, *x5); // 40 50 41 51 42 52 43 53 44 54 45 55 46 56 47 57 + + // x6 60 61 62 63 64 65 66 67 + // x7 70 71 72 73 74 75 76 77 + w3 = _mm_unpacklo_epi8( + *x6, *x7); // 60 70 61 71 62 72 63 73 64 74 65 75 66 76 67 77 + + w4 = _mm_unpacklo_epi16( + w0, w1); // 00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33 + w5 = _mm_unpacklo_epi16( + w2, w3); // 40 50 60 70 41 51 61 71 42 52 62 72 43 53 63 73 + + *d0d1 = _mm_unpacklo_epi32( + w4, w5); // 00 10 20 30 40 50 60 70 01 11 21 31 41 51 61 71 + *d2d3 = _mm_unpackhi_epi32( + w4, w5); // 02 12 22 32 42 52 62 72 03 13 23 33 43 53 63 73 + + w6 = _mm_unpackhi_epi16( + w0, w1); // 04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37 + w7 = _mm_unpackhi_epi16( + w2, w3); // 44 54 64 74 45 55 65 75 46 56 66 76 47 57 67 77 + + *d4d5 = _mm_unpacklo_epi32( + w6, w7); // 04 14 24 34 44 54 64 74 05 15 25 35 45 55 65 75 + *d6d7 = _mm_unpackhi_epi32( + w6, w7); // 06 16 26 36 46 56 66 76 07 17 27 37 47 57 67 77 +} + +static INLINE void transpose16x8_8x16_sse2( + __m128i *x0, __m128i *x1, __m128i *x2, __m128i *x3, __m128i *x4, + __m128i *x5, __m128i *x6, __m128i *x7, __m128i *x8, __m128i *x9, + __m128i *x10, __m128i *x11, __m128i *x12, __m128i *x13, __m128i *x14, + __m128i *x15, __m128i *d0, __m128i *d1, __m128i *d2, __m128i *d3, + __m128i *d4, __m128i *d5, __m128i *d6, __m128i *d7) { + __m128i w0, w1, w2, w3, w4, w5, w6, w7, w8, w9; + __m128i w10, w11, w12, w13, w14, w15; + + w0 = _mm_unpacklo_epi8(*x0, *x1); + w1 = _mm_unpacklo_epi8(*x2, *x3); + w2 = _mm_unpacklo_epi8(*x4, *x5); + w3 = _mm_unpacklo_epi8(*x6, *x7); + + w8 = _mm_unpacklo_epi8(*x8, *x9); + w9 = _mm_unpacklo_epi8(*x10, *x11); + w10 = _mm_unpacklo_epi8(*x12, *x13); + w11 = _mm_unpacklo_epi8(*x14, *x15); + + w4 = _mm_unpacklo_epi16(w0, w1); + w5 = _mm_unpacklo_epi16(w2, w3); + w12 = _mm_unpacklo_epi16(w8, w9); + w13 = _mm_unpacklo_epi16(w10, w11); + + w6 = _mm_unpacklo_epi32(w4, w5); + w7 = _mm_unpackhi_epi32(w4, w5); + w14 = _mm_unpacklo_epi32(w12, w13); + w15 = _mm_unpackhi_epi32(w12, w13); + + // Store first 4-line result + *d0 = _mm_unpacklo_epi64(w6, w14); + *d1 = _mm_unpackhi_epi64(w6, w14); + *d2 = _mm_unpacklo_epi64(w7, w15); + *d3 = _mm_unpackhi_epi64(w7, w15); + + w4 = _mm_unpackhi_epi16(w0, w1); + w5 = _mm_unpackhi_epi16(w2, w3); + w12 = _mm_unpackhi_epi16(w8, w9); + w13 = _mm_unpackhi_epi16(w10, w11); + + w6 = _mm_unpacklo_epi32(w4, w5); + w7 = _mm_unpackhi_epi32(w4, w5); + w14 = _mm_unpacklo_epi32(w12, w13); + w15 = _mm_unpackhi_epi32(w12, w13); + + // Store second 4-line result + *d4 = _mm_unpacklo_epi64(w6, w14); + *d5 = _mm_unpackhi_epi64(w6, w14); + *d6 = _mm_unpacklo_epi64(w7, w15); + *d7 = _mm_unpackhi_epi64(w7, w15); +} + +// this function treats its input as 2 parallel 8x4 matrices, transposes each of +// them independently while flipping the second matrix horizontaly Used for 14 +// taps filter pq pairs inverse +static INLINE void transpose_pq_14_inv_sse2(__m128i *x0, __m128i *x1, + __m128i *x2, __m128i *x3, + __m128i *x4, __m128i *x5, + __m128i *x6, __m128i *x7, + __m128i *pq0, __m128i *pq1, + __m128i *pq2, __m128i *pq3) { + __m128i w10, w11, w12, w13; + __m128i w0, w1, w2, w3, w4, w5; + __m128i d0, d1, d2, d3; + + w0 = _mm_unpacklo_epi8( + *x0, *x1); // p 00 10 01 11 02 12 03 13 04 14 05 15 06 16 07 17 + w1 = _mm_unpacklo_epi8( + *x2, *x3); // p 20 30 21 31 22 32 23 33 24 34 25 35 26 36 27 37 + w2 = _mm_unpacklo_epi8( + *x4, *x5); // p 40 50 41 51 42 52 43 53 44 54 45 55 46 56 47 57 + w3 = _mm_unpacklo_epi8( + *x6, *x7); // p 60 70 61 71 62 72 63 73 64 74 65 75 66 76 67 77 + + w4 = _mm_unpacklo_epi16( + w0, w1); // 00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33 + w5 = _mm_unpacklo_epi16( + w2, w3); // 40 50 60 70 41 51 61 71 42 52 62 72 43 53 63 73 + + d0 = _mm_unpacklo_epi32( + w4, w5); // 00 10 20 30 40 50 60 70 01 11 21 31 41 51 61 71 + d2 = _mm_unpackhi_epi32( + w4, w5); // 02 12 22 32 42 52 62 72 03 13 23 33 43 53 63 73 + + w10 = _mm_unpacklo_epi8( + *x7, *x6); // q xx xx xx xx xx xx xx xx 00 10 01 11 02 12 03 13 + w11 = _mm_unpacklo_epi8( + *x5, *x4); // q xx xx xx xx xx xx xx xx 20 30 21 31 22 32 23 33 + w12 = _mm_unpacklo_epi8( + *x3, *x2); // q xx xx xx xx xx xx xx xx 40 50 41 51 42 52 43 53 + w13 = _mm_unpacklo_epi8( + *x1, *x0); // q xx xx xx xx xx xx xx xx 60 70 61 71 62 72 63 73 + + w4 = _mm_unpackhi_epi16( + w10, w11); // 00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33 + w5 = _mm_unpackhi_epi16( + w12, w13); // 40 50 60 70 41 51 61 71 42 52 62 72 43 53 63 73 + + d1 = _mm_unpacklo_epi32( + w4, w5); // 00 10 20 30 40 50 60 70 01 11 21 31 41 51 61 71 + d3 = _mm_unpackhi_epi32( + w4, w5); // 02 12 22 32 42 52 62 72 03 13 23 33 43 53 63 73 + + *pq0 = _mm_unpacklo_epi64(d0, d1); // pq + *pq1 = _mm_unpackhi_epi64(d0, d1); // pq + *pq2 = _mm_unpacklo_epi64(d2, d3); // pq + *pq3 = _mm_unpackhi_epi64(d2, d3); // pq +} + +static INLINE void transpose8x16_16x8_sse2( + __m128i *x0, __m128i *x1, __m128i *x2, __m128i *x3, __m128i *x4, + __m128i *x5, __m128i *x6, __m128i *x7, __m128i *d0d1, __m128i *d2d3, + __m128i *d4d5, __m128i *d6d7, __m128i *d8d9, __m128i *d10d11, + __m128i *d12d13, __m128i *d14d15) { + __m128i w0, w1, w2, w3, w4, w5, w6, w7, w8, w9; + __m128i w10, w11, w12, w13, w14, w15; + + w0 = _mm_unpacklo_epi8(*x0, *x1); + w1 = _mm_unpacklo_epi8(*x2, *x3); + w2 = _mm_unpacklo_epi8(*x4, *x5); + w3 = _mm_unpacklo_epi8(*x6, *x7); + + w8 = _mm_unpackhi_epi8(*x0, *x1); + w9 = _mm_unpackhi_epi8(*x2, *x3); + w10 = _mm_unpackhi_epi8(*x4, *x5); + w11 = _mm_unpackhi_epi8(*x6, *x7); + + w4 = _mm_unpacklo_epi16(w0, w1); + w5 = _mm_unpacklo_epi16(w2, w3); + w12 = _mm_unpacklo_epi16(w8, w9); + w13 = _mm_unpacklo_epi16(w10, w11); + + w6 = _mm_unpacklo_epi32(w4, w5); + w7 = _mm_unpackhi_epi32(w4, w5); + w14 = _mm_unpacklo_epi32(w12, w13); + w15 = _mm_unpackhi_epi32(w12, w13); + + // Store first 4-line result + *d0d1 = _mm_unpacklo_epi64(w6, w14); + *d2d3 = _mm_unpackhi_epi64(w6, w14); + *d4d5 = _mm_unpacklo_epi64(w7, w15); + *d6d7 = _mm_unpackhi_epi64(w7, w15); + + w4 = _mm_unpackhi_epi16(w0, w1); + w5 = _mm_unpackhi_epi16(w2, w3); + w12 = _mm_unpackhi_epi16(w8, w9); + w13 = _mm_unpackhi_epi16(w10, w11); + + w6 = _mm_unpacklo_epi32(w4, w5); + w7 = _mm_unpackhi_epi32(w4, w5); + w14 = _mm_unpacklo_epi32(w12, w13); + w15 = _mm_unpackhi_epi32(w12, w13); + + // Store second 4-line result + *d8d9 = _mm_unpacklo_epi64(w6, w14); + *d10d11 = _mm_unpackhi_epi64(w6, w14); + *d12d13 = _mm_unpacklo_epi64(w7, w15); + *d14d15 = _mm_unpackhi_epi64(w7, w15); +} + +// this function treats its input as 2 parallel 8x4 matrices, transposes each of +// them to 4x8 independently while flipping the second matrix horizontaly. Used +// for 14 taps pq pairs creation +static INLINE void transpose_pq_14_sse2(__m128i *x0, __m128i *x1, __m128i *x2, + __m128i *x3, __m128i *q0p0, + __m128i *q1p1, __m128i *q2p2, + __m128i *q3p3, __m128i *q4p4, + __m128i *q5p5, __m128i *q6p6, + __m128i *q7p7) { + __m128i w0, w1, ww0, ww1, w2, w3, ww2, ww3; + w0 = _mm_unpacklo_epi8( + *x0, *x1); // 00 10 01 11 02 12 03 13 04 14 05 15 06 16 07 17 + w1 = _mm_unpacklo_epi8( + *x2, *x3); // 20 30 21 31 22 32 23 33 24 34 25 35 26 36 27 37 + w2 = _mm_unpackhi_epi8( + *x0, *x1); // 08 18 09 19 010 110 011 111 012 112 013 113 014 114 015 115 + w3 = _mm_unpackhi_epi8( + *x2, *x3); // 28 38 29 39 210 310 211 311 212 312 213 313 214 314 215 315 + + ww0 = _mm_unpacklo_epi16( + w0, w1); // 00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33 + ww1 = _mm_unpackhi_epi16( + w0, w1); // 04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37 + ww2 = _mm_unpacklo_epi16( + w2, w3); // 08 18 28 38 09 19 29 39 010 110 210 310 011 111 211 311 + ww3 = _mm_unpackhi_epi16( + w2, + w3); // 012 112 212 312 013 113 213 313 014 114 214 314 015 115 215 315 + + *q7p7 = _mm_unpacklo_epi32( + ww0, + _mm_srli_si128( + ww3, 12)); // 00 10 20 30 015 115 215 315 xx xx xx xx xx xx xx xx + *q6p6 = _mm_unpackhi_epi32( + _mm_slli_si128(ww0, 4), + ww3); // 01 11 21 31 014 114 214 314 xx xx xx xxxx xx xx xx + *q5p5 = _mm_unpackhi_epi32( + ww0, + _mm_slli_si128( + ww3, 4)); // 02 12 22 32 013 113 213 313 xx xx xx x xx xx xx xxx + *q4p4 = _mm_unpacklo_epi32( + _mm_srli_si128(ww0, 12), + ww3); // 03 13 23 33 012 112 212 312 xx xx xx xx xx xx xx xx + *q3p3 = _mm_unpacklo_epi32( + ww1, + _mm_srli_si128( + ww2, 12)); // 04 14 24 34 011 111 211 311 xx xx xx xx xx xx xx xx + *q2p2 = _mm_unpackhi_epi32( + _mm_slli_si128(ww1, 4), + ww2); // 05 15 25 35 010 110 210 310 xx xx xx xx xx xx xx xx + *q1p1 = _mm_unpackhi_epi32( + ww1, + _mm_slli_si128( + ww2, 4)); // 06 16 26 36 09 19 29 39 xx xx xx xx xx xx xx xx + *q0p0 = _mm_unpacklo_epi32( + _mm_srli_si128(ww1, 12), + ww2); // 07 17 27 37 08 18 28 38 xx xx xx xx xx xx xx xx +} + +static AOM_FORCE_INLINE void filter4_sse2(__m128i *p1p0, __m128i *q1q0, + __m128i *hev, __m128i *mask, + __m128i *qs1qs0, __m128i *ps1ps0) { + __m128i filter, filter2filter1, work; + __m128i ps1ps0_work, qs1qs0_work; + __m128i hev1; + const __m128i t3t4 = + _mm_set_epi8(0, 0, 0, 0, 0, 0, 0, 0, 3, 3, 3, 3, 4, 4, 4, 4); + const __m128i t80 = _mm_set1_epi8(0x80); + const __m128i ff = _mm_cmpeq_epi8(t80, t80); + + ps1ps0_work = _mm_xor_si128(*p1p0, t80); /* ^ 0x80 */ + qs1qs0_work = _mm_xor_si128(*q1q0, t80); + + /* int8_t filter = signed_char_clamp(ps1 - qs1) & hev; */ + work = _mm_subs_epi8(ps1ps0_work, qs1qs0_work); + filter = _mm_and_si128(_mm_srli_si128(work, 4), *hev); + /* filter = signed_char_clamp(filter + 3 * (qs0 - ps0)) & mask; */ + filter = _mm_subs_epi8(filter, work); + filter = _mm_subs_epi8(filter, work); + filter = _mm_subs_epi8(filter, work); /* + 3 * (qs0 - ps0) */ + filter = _mm_and_si128(filter, *mask); /* & mask */ + filter = _mm_unpacklo_epi32(filter, filter); + + /* filter1 = signed_char_clamp(filter + 4) >> 3; */ + /* filter2 = signed_char_clamp(filter + 3) >> 3; */ + filter2filter1 = _mm_adds_epi8(filter, t3t4); /* signed_char_clamp */ + filter2filter1 = + _mm_unpacklo_epi8(filter2filter1, filter2filter1); // goto 16 bit + filter2filter1 = _mm_srai_epi16(filter2filter1, 11); /* >> 3 */ + filter2filter1 = _mm_packs_epi16(filter2filter1, filter2filter1); + + /* filter = ROUND_POWER_OF_TWO(filter1, 1) & ~hev; */ + filter = _mm_subs_epi8(filter2filter1, ff); /* + 1 */ + filter = _mm_unpacklo_epi8(filter, filter); // goto 16 bit + filter = _mm_srai_epi16(filter, 9); /* round */ + filter = _mm_packs_epi16(filter, filter); + filter = _mm_andnot_si128(*hev, filter); + filter = _mm_unpacklo_epi32(filter, filter); + + filter2filter1 = _mm_unpacklo_epi32(filter2filter1, filter); + hev1 = _mm_srli_si128(filter2filter1, 8); + /* signed_char_clamp(qs1 - filter), signed_char_clamp(qs0 - filter1) */ + qs1qs0_work = _mm_subs_epi8(qs1qs0_work, filter2filter1); + /* signed_char_clamp(ps1 + filter), signed_char_clamp(ps0 + filter2) */ + ps1ps0_work = _mm_adds_epi8(ps1ps0_work, hev1); + + *qs1qs0 = _mm_xor_si128(qs1qs0_work, t80); /* ^ 0x80 */ + *ps1ps0 = _mm_xor_si128(ps1ps0_work, t80); /* ^ 0x80 */ +} + +static AOM_FORCE_INLINE void filter4_dual_sse2(__m128i *p1p0, __m128i *q1q0, + __m128i *hev, __m128i *mask, + __m128i *qs1qs0, + __m128i *ps1ps0) { + const __m128i t3t4 = + _mm_set_epi8(3, 3, 3, 3, 3, 3, 3, 3, 4, 4, 4, 4, 4, 4, 4, 4); + const __m128i t80 = _mm_set1_epi8(0x80); + __m128i filter, filter2filter1, work; + __m128i ps1ps0_work, qs1qs0_work; + __m128i hev1; + const __m128i ff = _mm_cmpeq_epi8(t80, t80); + + ps1ps0_work = _mm_xor_si128(*p1p0, t80); /* ^ 0x80 */ + qs1qs0_work = _mm_xor_si128(*q1q0, t80); + + /* int8_t filter = signed_char_clamp(ps1 - qs1) & hev; */ + work = _mm_subs_epi8(ps1ps0_work, qs1qs0_work); + filter = _mm_and_si128(_mm_srli_si128(work, 8), *hev); + /* filter = signed_char_clamp(filter + 3 * (qs0 - ps0)) & mask; */ + filter = _mm_subs_epi8(filter, work); + filter = _mm_subs_epi8(filter, work); + filter = _mm_subs_epi8(filter, work); /* + 3 * (qs0 - ps0) */ + filter = _mm_and_si128(filter, *mask); /* & mask */ + filter = _mm_unpacklo_epi64(filter, filter); + + /* filter1 = signed_char_clamp(filter + 4) >> 3; */ + /* filter2 = signed_char_clamp(filter + 3) >> 3; */ + filter2filter1 = _mm_adds_epi8(filter, t3t4); /* signed_char_clamp */ + filter = _mm_unpackhi_epi8(filter2filter1, filter2filter1); + filter2filter1 = _mm_unpacklo_epi8(filter2filter1, filter2filter1); + filter2filter1 = _mm_srai_epi16(filter2filter1, 11); /* >> 3 */ + filter = _mm_srai_epi16(filter, 11); /* >> 3 */ + filter2filter1 = _mm_packs_epi16(filter2filter1, filter); + + /* filter = ROUND_POWER_OF_TWO(filter1, 1) & ~hev; */ + filter = _mm_subs_epi8(filter2filter1, ff); /* + 1 */ + filter = _mm_unpacklo_epi8(filter, filter); + filter = _mm_srai_epi16(filter, 9); /* round */ + filter = _mm_packs_epi16(filter, filter); + filter = _mm_andnot_si128(*hev, filter); + + hev1 = _mm_unpackhi_epi64(filter2filter1, filter); + filter2filter1 = _mm_unpacklo_epi64(filter2filter1, filter); + + /* signed_char_clamp(qs1 - filter), signed_char_clamp(qs0 - filter1) */ + qs1qs0_work = _mm_subs_epi8(qs1qs0_work, filter2filter1); + /* signed_char_clamp(ps1 + filter), signed_char_clamp(ps0 + filter2) */ + ps1ps0_work = _mm_adds_epi8(ps1ps0_work, hev1); + *qs1qs0 = _mm_xor_si128(qs1qs0_work, t80); /* ^ 0x80 */ + *ps1ps0 = _mm_xor_si128(ps1ps0_work, t80); /* ^ 0x80 */ +} + +static AOM_FORCE_INLINE void lpf_internal_4_sse2( + __m128i *p1, __m128i *p0, __m128i *q0, __m128i *q1, __m128i *limit, + __m128i *thresh, __m128i *q1q0_out, __m128i *p1p0_out) { + __m128i q1p1, q0p0, p1p0, q1q0; + __m128i abs_p0q0, abs_p1q1; + __m128i mask, flat, hev; + const __m128i zero = _mm_setzero_si128(); + + q1p1 = _mm_unpacklo_epi32(*p1, *q1); + q0p0 = _mm_unpacklo_epi32(*p0, *q0); + + p1p0 = _mm_unpacklo_epi32(q0p0, q1p1); + q1q0 = _mm_srli_si128(p1p0, 8); + + /* (abs(q1 - q0), abs(p1 - p0) */ + flat = abs_diff(q1p1, q0p0); + /* abs(p1 - q1), abs(p0 - q0) */ + __m128i abs_p1q1p0q0 = abs_diff(p1p0, q1q0); + + /* const uint8_t hev = hev_mask(thresh, *op1, *op0, *oq0, *oq1); */ + flat = _mm_max_epu8(flat, _mm_srli_si128(flat, 4)); + hev = _mm_unpacklo_epi8(flat, zero); + + hev = _mm_cmpgt_epi16(hev, *thresh); + hev = _mm_packs_epi16(hev, hev); + hev = _mm_unpacklo_epi32(hev, hev); + + abs_p0q0 = _mm_adds_epu8(abs_p1q1p0q0, abs_p1q1p0q0); /* abs(p0 - q0) * 2 */ + abs_p1q1 = _mm_srli_si128(abs_p1q1p0q0, 4); /* abs(p1 - q1) */ + abs_p1q1 = _mm_unpacklo_epi8(abs_p1q1, abs_p1q1); + abs_p1q1 = _mm_srli_epi16(abs_p1q1, 9); + abs_p1q1 = _mm_packs_epi16(abs_p1q1, abs_p1q1); /* abs(p1 - q1) / 2 */ + /* abs(p0 - q0) * 2 + abs(p1 - q1) / 2 */ + + mask = _mm_adds_epu8(abs_p0q0, abs_p1q1); + mask = _mm_unpacklo_epi32(mask, flat); + mask = _mm_subs_epu8(mask, *limit); + mask = _mm_cmpeq_epi8(mask, zero); + mask = _mm_and_si128(mask, _mm_srli_si128(mask, 4)); + + filter4_sse2(&p1p0, &q1q0, &hev, &mask, q1q0_out, p1p0_out); +} + +static AOM_FORCE_INLINE void lpf_internal_4_dual_sse2( + __m128i *p1, __m128i *p0, __m128i *q0, __m128i *q1, __m128i *limit, + __m128i *thresh, __m128i *q1q0_out, __m128i *p1p0_out) { + __m128i q1p1, q0p0, p1p0, q1q0; + __m128i abs_p0q0, abs_p1q1; + __m128i mask, hev; + const __m128i zero = _mm_setzero_si128(); + + q1p1 = _mm_unpacklo_epi64(*p1, *q1); + q0p0 = _mm_unpacklo_epi64(*p0, *q0); + + p1p0 = _mm_unpacklo_epi64(q0p0, q1p1); + q1q0 = _mm_unpackhi_epi64(q0p0, q1p1); + + /* (abs(q1 - q0), abs(p1 - p0) */ + __m128i flat = abs_diff(q1p1, q0p0); + /* abs(p1 - q1), abs(p0 - q0) */ + const __m128i abs_p1q1p0q0 = abs_diff(p1p0, q1q0); + + /* const uint8_t hev = hev_mask(thresh, *op1, *op0, *oq0, *oq1); */ + flat = _mm_max_epu8(flat, _mm_srli_si128(flat, 8)); + hev = _mm_unpacklo_epi8(flat, zero); + + hev = _mm_cmpgt_epi16(hev, *thresh); + hev = _mm_packs_epi16(hev, hev); + + /* const int8_t mask = filter_mask2(*limit, *blimit, */ + /* p1, p0, q0, q1); */ + abs_p0q0 = _mm_adds_epu8(abs_p1q1p0q0, abs_p1q1p0q0); /* abs(p0 - q0) * 2 */ + abs_p1q1 = _mm_unpackhi_epi8(abs_p1q1p0q0, abs_p1q1p0q0); /* abs(p1 - q1) */ + abs_p1q1 = _mm_srli_epi16(abs_p1q1, 9); + abs_p1q1 = _mm_packs_epi16(abs_p1q1, abs_p1q1); /* abs(p1 - q1) / 2 */ + /* abs(p0 - q0) * 2 + abs(p1 - q1) / 2 */ + mask = _mm_adds_epu8(abs_p0q0, abs_p1q1); + mask = _mm_unpacklo_epi64(mask, flat); + mask = _mm_subs_epu8(mask, *limit); + mask = _mm_cmpeq_epi8(mask, zero); + mask = _mm_and_si128(mask, _mm_srli_si128(mask, 8)); + + filter4_dual_sse2(&p1p0, &q1q0, &hev, &mask, q1q0_out, p1p0_out); +} + +void aom_lpf_horizontal_4_sse2(uint8_t *s, int p /* pitch */, + const uint8_t *_blimit, const uint8_t *_limit, + const uint8_t *_thresh) { + const __m128i zero = _mm_setzero_si128(); + __m128i limit = _mm_unpacklo_epi32(_mm_loadl_epi64((const __m128i *)_blimit), + _mm_loadl_epi64((const __m128i *)_limit)); + __m128i thresh = + _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)_thresh), zero); + + __m128i qs1qs0, ps1ps0; + __m128i p1, p0, q0, q1; + + p1 = _mm_cvtsi32_si128(*(int *)(s - 2 * p)); + p0 = _mm_cvtsi32_si128(*(int *)(s - 1 * p)); + q0 = _mm_cvtsi32_si128(*(int *)(s + 0 * p)); + q1 = _mm_cvtsi32_si128(*(int *)(s + 1 * p)); + + lpf_internal_4_sse2(&p1, &p0, &q0, &q1, &limit, &thresh, &qs1qs0, &ps1ps0); + + xx_storel_32(s - 1 * p, ps1ps0); + xx_storel_32(s - 2 * p, _mm_srli_si128(ps1ps0, 4)); + xx_storel_32(s + 0 * p, qs1qs0); + xx_storel_32(s + 1 * p, _mm_srli_si128(qs1qs0, 4)); +} + +void aom_lpf_vertical_4_sse2(uint8_t *s, int p /* pitch */, + const uint8_t *_blimit, const uint8_t *_limit, + const uint8_t *_thresh) { + __m128i p1p0, q1q0; + __m128i p1, p0, q0, q1; + + const __m128i zero = _mm_setzero_si128(); + __m128i limit = _mm_unpacklo_epi32(_mm_loadl_epi64((const __m128i *)_blimit), + _mm_loadl_epi64((const __m128i *)_limit)); + __m128i thresh = + _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)_thresh), zero); + + __m128i x0, x1, x2, x3; + __m128i d0, d1, d2, d3; + x0 = _mm_loadl_epi64((__m128i *)(s - 2 + 0 * p)); + x1 = _mm_loadl_epi64((__m128i *)(s - 2 + 1 * p)); + x2 = _mm_loadl_epi64((__m128i *)(s - 2 + 2 * p)); + x3 = _mm_loadl_epi64((__m128i *)(s - 2 + 3 * p)); + + transpose4x8_8x4_low_sse2(&x0, &x1, &x2, &x3, &p1, &p0, &q0, &q1); + + lpf_internal_4_sse2(&p1, &p0, &q0, &q1, &limit, &thresh, &q1q0, &p1p0); + + // Transpose 8x4 to 4x8 + p1 = _mm_srli_si128(p1p0, 4); + q1 = _mm_srli_si128(q1q0, 4); + + transpose4x8_8x4_low_sse2(&p1, &p1p0, &q1q0, &q1, &d0, &d1, &d2, &d3); + + xx_storel_32(s + 0 * p - 2, d0); + xx_storel_32(s + 1 * p - 2, d1); + xx_storel_32(s + 2 * p - 2, d2); + xx_storel_32(s + 3 * p - 2, d3); +} + +static INLINE void store_buffer_horz_8(__m128i x, int p, int num, uint8_t *s) { + xx_storel_32(s - (num + 1) * p, x); + xx_storel_32(s + num * p, _mm_srli_si128(x, 4)); +} + +static AOM_FORCE_INLINE void lpf_internal_14_dual_sse2( + __m128i *q6p6, __m128i *q5p5, __m128i *q4p4, __m128i *q3p3, __m128i *q2p2, + __m128i *q1p1, __m128i *q0p0, __m128i *blimit, __m128i *limit, + __m128i *thresh) { + const __m128i zero = _mm_setzero_si128(); + const __m128i one = _mm_set1_epi8(1); + __m128i mask, hev, flat, flat2; + __m128i qs0ps0, qs1ps1; + __m128i p1p0, q1q0, qs1qs0, ps1ps0; + __m128i abs_p1p0; + + p1p0 = _mm_unpacklo_epi64(*q0p0, *q1p1); + q1q0 = _mm_unpackhi_epi64(*q0p0, *q1p1); + + { + __m128i abs_p1q1, abs_p0q0, abs_q1q0; + __m128i fe, ff, work; + abs_p1p0 = abs_diff(*q1p1, *q0p0); + abs_q1q0 = _mm_srli_si128(abs_p1p0, 8); + fe = _mm_set1_epi8(0xfe); + ff = _mm_cmpeq_epi8(abs_p1p0, abs_p1p0); + abs_p0q0 = abs_diff(p1p0, q1q0); + abs_p1q1 = _mm_srli_si128(abs_p0q0, 8); + abs_p0q0 = _mm_unpacklo_epi64(abs_p0q0, zero); + + flat = _mm_max_epu8(abs_p1p0, abs_q1q0); + hev = _mm_subs_epu8(flat, *thresh); + hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff); + // replicate for the further "merged variables" usage + hev = _mm_unpacklo_epi64(hev, hev); + + abs_p0q0 = _mm_adds_epu8(abs_p0q0, abs_p0q0); + abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1); + mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), *blimit); + mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff); + // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit) * -1; + mask = _mm_max_epu8(abs_p1p0, mask); + // mask |= (abs(p1 - p0) > limit) * -1; + // mask |= (abs(q1 - q0) > limit) * -1; + + work = _mm_max_epu8(abs_diff(*q2p2, *q1p1), abs_diff(*q3p3, *q2p2)); + mask = _mm_max_epu8(work, mask); + mask = _mm_max_epu8(mask, _mm_srli_si128(mask, 8)); + mask = _mm_subs_epu8(mask, *limit); + mask = _mm_cmpeq_epi8(mask, zero); + } + + // lp filter - the same for 6, 8 and 14 versions + filter4_dual_sse2(&p1p0, &q1q0, &hev, &mask, &qs1qs0, &ps1ps0); + qs0ps0 = _mm_unpacklo_epi64(ps1ps0, qs1qs0); + qs1ps1 = _mm_unpackhi_epi64(ps1ps0, qs1qs0); + // loopfilter done + + __m128i flat2_q5p5, flat2_q4p4, flat2_q3p3, flat2_q2p2; + __m128i flat2_q1p1, flat2_q0p0, flat_q2p2, flat_q1p1, flat_q0p0; + + __m128i work; + flat = _mm_max_epu8(abs_diff(*q2p2, *q0p0), abs_diff(*q3p3, *q0p0)); + flat = _mm_max_epu8(abs_p1p0, flat); + flat = _mm_max_epu8(flat, _mm_srli_si128(flat, 8)); + flat = _mm_subs_epu8(flat, one); + flat = _mm_cmpeq_epi8(flat, zero); + flat = _mm_and_si128(flat, mask); + + // if flat ==0 then flat2 is zero as well and we don't need any calc below + // sse4.1 if (0==_mm_test_all_zeros(flat,ff)) + if (0xffff != _mm_movemask_epi8(_mm_cmpeq_epi8(flat, zero))) { + // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ + // flat and wide flat calculations + + const __m128i eight = _mm_set1_epi16(8); + const __m128i four = _mm_set1_epi16(4); + __m128i p6_16, p5_16, p4_16, p3_16, p2_16, p1_16, p0_16; + __m128i q6_16, q5_16, q4_16, q3_16, q2_16, q1_16, q0_16; + __m128i pixelFilter_p, pixelFilter_q; + __m128i pixetFilter_p2p1p0, pixetFilter_q2q1q0; + __m128i sum_p6, sum_q6; + __m128i sum_p3, sum_q3, res_p, res_q; + + p6_16 = _mm_unpacklo_epi8(*q6p6, zero); + p5_16 = _mm_unpacklo_epi8(*q5p5, zero); + p4_16 = _mm_unpacklo_epi8(*q4p4, zero); + p3_16 = _mm_unpacklo_epi8(*q3p3, zero); + p2_16 = _mm_unpacklo_epi8(*q2p2, zero); + p1_16 = _mm_unpacklo_epi8(*q1p1, zero); + p0_16 = _mm_unpacklo_epi8(*q0p0, zero); + q0_16 = _mm_unpackhi_epi8(*q0p0, zero); + q1_16 = _mm_unpackhi_epi8(*q1p1, zero); + q2_16 = _mm_unpackhi_epi8(*q2p2, zero); + q3_16 = _mm_unpackhi_epi8(*q3p3, zero); + q4_16 = _mm_unpackhi_epi8(*q4p4, zero); + q5_16 = _mm_unpackhi_epi8(*q5p5, zero); + q6_16 = _mm_unpackhi_epi8(*q6p6, zero); + pixelFilter_p = _mm_add_epi16(p5_16, _mm_add_epi16(p4_16, p3_16)); + pixelFilter_q = _mm_add_epi16(q5_16, _mm_add_epi16(q4_16, q3_16)); + + pixetFilter_p2p1p0 = _mm_add_epi16(p0_16, _mm_add_epi16(p2_16, p1_16)); + pixelFilter_p = _mm_add_epi16(pixelFilter_p, pixetFilter_p2p1p0); + + pixetFilter_q2q1q0 = _mm_add_epi16(q0_16, _mm_add_epi16(q2_16, q1_16)); + pixelFilter_q = _mm_add_epi16(pixelFilter_q, pixetFilter_q2q1q0); + pixelFilter_p = + _mm_add_epi16(eight, _mm_add_epi16(pixelFilter_p, pixelFilter_q)); + pixetFilter_p2p1p0 = _mm_add_epi16( + four, _mm_add_epi16(pixetFilter_p2p1p0, pixetFilter_q2q1q0)); + res_p = _mm_srli_epi16( + _mm_add_epi16(pixelFilter_p, + _mm_add_epi16(_mm_add_epi16(p6_16, p0_16), + _mm_add_epi16(p1_16, q0_16))), + 4); + res_q = _mm_srli_epi16( + _mm_add_epi16(pixelFilter_p, + _mm_add_epi16(_mm_add_epi16(q6_16, q0_16), + _mm_add_epi16(p0_16, q1_16))), + 4); + flat2_q0p0 = _mm_packus_epi16(res_p, res_q); + + res_p = _mm_srli_epi16( + _mm_add_epi16(pixetFilter_p2p1p0, _mm_add_epi16(p3_16, p0_16)), 3); + res_q = _mm_srli_epi16( + _mm_add_epi16(pixetFilter_p2p1p0, _mm_add_epi16(q3_16, q0_16)), 3); + + flat_q0p0 = _mm_packus_epi16(res_p, res_q); + + sum_p6 = _mm_add_epi16(p6_16, p6_16); + sum_q6 = _mm_add_epi16(q6_16, q6_16); + sum_p3 = _mm_add_epi16(p3_16, p3_16); + sum_q3 = _mm_add_epi16(q3_16, q3_16); + + pixelFilter_q = _mm_sub_epi16(pixelFilter_p, p5_16); + pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q5_16); + + res_p = _mm_srli_epi16( + _mm_add_epi16( + pixelFilter_p, + _mm_add_epi16(sum_p6, + _mm_add_epi16(p1_16, _mm_add_epi16(p2_16, p0_16)))), + 4); + res_q = _mm_srli_epi16( + _mm_add_epi16( + pixelFilter_q, + _mm_add_epi16(sum_q6, + _mm_add_epi16(q1_16, _mm_add_epi16(q0_16, q2_16)))), + 4); + flat2_q1p1 = _mm_packus_epi16(res_p, res_q); + + pixetFilter_q2q1q0 = _mm_sub_epi16(pixetFilter_p2p1p0, p2_16); + pixetFilter_p2p1p0 = _mm_sub_epi16(pixetFilter_p2p1p0, q2_16); + res_p = _mm_srli_epi16( + _mm_add_epi16(pixetFilter_p2p1p0, _mm_add_epi16(sum_p3, p1_16)), 3); + res_q = _mm_srli_epi16( + _mm_add_epi16(pixetFilter_q2q1q0, _mm_add_epi16(sum_q3, q1_16)), 3); + flat_q1p1 = _mm_packus_epi16(res_p, res_q); + + pixetFilter_p2p1p0 = _mm_sub_epi16(pixetFilter_p2p1p0, q1_16); + pixetFilter_q2q1q0 = _mm_sub_epi16(pixetFilter_q2q1q0, p1_16); + + sum_p3 = _mm_add_epi16(sum_p3, p3_16); + sum_q3 = _mm_add_epi16(sum_q3, q3_16); + + res_p = _mm_srli_epi16( + _mm_add_epi16(pixetFilter_p2p1p0, _mm_add_epi16(sum_p3, p2_16)), 3); + res_q = _mm_srli_epi16( + _mm_add_epi16(pixetFilter_q2q1q0, _mm_add_epi16(sum_q3, q2_16)), 3); + flat_q2p2 = _mm_packus_epi16(res_p, res_q); + + // work with flat2 + flat2 = _mm_max_epu8(abs_diff(*q4p4, *q0p0), abs_diff(*q5p5, *q0p0)); + work = abs_diff(*q6p6, *q0p0); + flat2 = _mm_max_epu8(work, flat2); + flat2 = _mm_max_epu8(flat2, _mm_srli_si128(flat2, 8)); + flat2 = _mm_subs_epu8(flat2, one); + flat2 = _mm_cmpeq_epi8(flat2, zero); + flat2 = _mm_and_si128(flat2, flat); // flat2 & flat & mask + + // ~~~~~~~~~~ apply flat ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ + flat = _mm_unpacklo_epi64(flat, flat); + *q2p2 = _mm_andnot_si128(flat, *q2p2); + flat_q2p2 = _mm_and_si128(flat, flat_q2p2); + *q2p2 = _mm_or_si128(*q2p2, flat_q2p2); + + qs1ps1 = _mm_andnot_si128(flat, qs1ps1); + flat_q1p1 = _mm_and_si128(flat, flat_q1p1); + *q1p1 = _mm_or_si128(qs1ps1, flat_q1p1); + + qs0ps0 = _mm_andnot_si128(flat, qs0ps0); + flat_q0p0 = _mm_and_si128(flat, flat_q0p0); + *q0p0 = _mm_or_si128(qs0ps0, flat_q0p0); + + if (0xffff != _mm_movemask_epi8(_mm_cmpeq_epi8(flat2, zero))) { + pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q4_16); + pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p4_16); + + sum_p6 = _mm_add_epi16(sum_p6, p6_16); + sum_q6 = _mm_add_epi16(sum_q6, q6_16); + + res_p = _mm_srli_epi16( + _mm_add_epi16( + pixelFilter_p, + _mm_add_epi16(sum_p6, + _mm_add_epi16(p2_16, _mm_add_epi16(p3_16, p1_16)))), + 4); + res_q = _mm_srli_epi16( + _mm_add_epi16( + pixelFilter_q, + _mm_add_epi16(sum_q6, + _mm_add_epi16(q2_16, _mm_add_epi16(q1_16, q3_16)))), + 4); + flat2_q2p2 = _mm_packus_epi16(res_p, res_q); + + sum_p6 = _mm_add_epi16(sum_p6, p6_16); + sum_q6 = _mm_add_epi16(sum_q6, q6_16); + + pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q3_16); + pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p3_16); + + res_p = _mm_srli_epi16( + _mm_add_epi16( + pixelFilter_p, + _mm_add_epi16(sum_p6, + _mm_add_epi16(p3_16, _mm_add_epi16(p4_16, p2_16)))), + 4); + res_q = _mm_srli_epi16( + _mm_add_epi16( + pixelFilter_q, + _mm_add_epi16(sum_q6, + _mm_add_epi16(q3_16, _mm_add_epi16(q2_16, q4_16)))), + 4); + flat2_q3p3 = _mm_packus_epi16(res_p, res_q); + + sum_p6 = _mm_add_epi16(sum_p6, p6_16); + sum_q6 = _mm_add_epi16(sum_q6, q6_16); + + pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q2_16); + pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p2_16); + + res_p = _mm_srli_epi16( + _mm_add_epi16( + pixelFilter_p, + _mm_add_epi16(sum_p6, + _mm_add_epi16(p4_16, _mm_add_epi16(p5_16, p3_16)))), + 4); + res_q = _mm_srli_epi16( + _mm_add_epi16( + pixelFilter_q, + _mm_add_epi16(sum_q6, + _mm_add_epi16(q4_16, _mm_add_epi16(q3_16, q5_16)))), + 4); + flat2_q4p4 = _mm_packus_epi16(res_p, res_q); + + sum_p6 = _mm_add_epi16(sum_p6, p6_16); + sum_q6 = _mm_add_epi16(sum_q6, q6_16); + pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q1_16); + pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p1_16); + + res_p = _mm_srli_epi16( + _mm_add_epi16( + pixelFilter_p, + _mm_add_epi16(sum_p6, + _mm_add_epi16(p5_16, _mm_add_epi16(p6_16, p4_16)))), + 4); + res_q = _mm_srli_epi16( + _mm_add_epi16( + pixelFilter_q, + _mm_add_epi16(sum_q6, + _mm_add_epi16(q5_16, _mm_add_epi16(q6_16, q4_16)))), + 4); + flat2_q5p5 = _mm_packus_epi16(res_p, res_q); + + // wide flat + // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ + flat2 = _mm_unpacklo_epi64(flat2, flat2); + + *q5p5 = _mm_andnot_si128(flat2, *q5p5); + flat2_q5p5 = _mm_and_si128(flat2, flat2_q5p5); + *q5p5 = _mm_or_si128(*q5p5, flat2_q5p5); + + *q4p4 = _mm_andnot_si128(flat2, *q4p4); + flat2_q4p4 = _mm_and_si128(flat2, flat2_q4p4); + *q4p4 = _mm_or_si128(*q4p4, flat2_q4p4); + + *q3p3 = _mm_andnot_si128(flat2, *q3p3); + flat2_q3p3 = _mm_and_si128(flat2, flat2_q3p3); + *q3p3 = _mm_or_si128(*q3p3, flat2_q3p3); + + *q2p2 = _mm_andnot_si128(flat2, *q2p2); + flat2_q2p2 = _mm_and_si128(flat2, flat2_q2p2); + *q2p2 = _mm_or_si128(*q2p2, flat2_q2p2); + + *q1p1 = _mm_andnot_si128(flat2, *q1p1); + flat2_q1p1 = _mm_and_si128(flat2, flat2_q1p1); + *q1p1 = _mm_or_si128(*q1p1, flat2_q1p1); + + *q0p0 = _mm_andnot_si128(flat2, *q0p0); + flat2_q0p0 = _mm_and_si128(flat2, flat2_q0p0); + *q0p0 = _mm_or_si128(*q0p0, flat2_q0p0); + } + } else { + *q0p0 = qs0ps0; + *q1p1 = qs1ps1; + } +} + +static AOM_FORCE_INLINE void lpf_internal_14_sse2( + __m128i *q6p6, __m128i *q5p5, __m128i *q4p4, __m128i *q3p3, __m128i *q2p2, + __m128i *q1p1, __m128i *q0p0, __m128i *blimit, __m128i *limit, + __m128i *thresh) { + const __m128i zero = _mm_setzero_si128(); + const __m128i one = _mm_set1_epi8(1); + __m128i mask, hev, flat, flat2; + __m128i flat2_pq[6], flat_pq[3]; + __m128i qs0ps0, qs1ps1; + __m128i p1p0, q1q0, qs1qs0, ps1ps0; + __m128i abs_p1p0; + + p1p0 = _mm_unpacklo_epi32(*q0p0, *q1p1); + q1q0 = _mm_srli_si128(p1p0, 8); + + __m128i fe, ff, work; + { + __m128i abs_p1q1, abs_p0q0, abs_q1q0; + abs_p1p0 = abs_diff(*q1p1, *q0p0); + abs_q1q0 = _mm_srli_si128(abs_p1p0, 4); + fe = _mm_set1_epi8(0xfe); + ff = _mm_cmpeq_epi8(fe, fe); + abs_p0q0 = abs_diff(p1p0, q1q0); + abs_p1q1 = _mm_srli_si128(abs_p0q0, 4); + + flat = _mm_max_epu8(abs_p1p0, abs_q1q0); + + hev = _mm_subs_epu8(flat, *thresh); + hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff); + // replicate for the further "merged variables" usage + hev = _mm_unpacklo_epi32(hev, hev); + + abs_p0q0 = _mm_adds_epu8(abs_p0q0, abs_p0q0); + abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1); + mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), *blimit); + mask = _mm_unpacklo_epi32(mask, zero); + mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff); + // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit) * -1; + mask = _mm_max_epu8(abs_p1p0, mask); + // mask |= (abs(p1 - p0) > limit) * -1; + // mask |= (abs(q1 - q0) > limit) * -1; + + work = _mm_max_epu8(abs_diff(*q2p2, *q1p1), abs_diff(*q3p3, *q2p2)); + mask = _mm_max_epu8(work, mask); + mask = _mm_max_epu8(mask, _mm_srli_si128(mask, 4)); + mask = _mm_subs_epu8(mask, *limit); + mask = _mm_cmpeq_epi8(mask, zero); + } + + // lp filter - the same for 6, 8 and 14 versions + filter4_sse2(&p1p0, &q1q0, &hev, &mask, &qs1qs0, &ps1ps0); + qs0ps0 = _mm_unpacklo_epi32(ps1ps0, qs1qs0); + qs1ps1 = _mm_srli_si128(qs0ps0, 8); + // loopfilter done + + flat = _mm_max_epu8(abs_diff(*q2p2, *q0p0), abs_diff(*q3p3, *q0p0)); + flat = _mm_max_epu8(abs_p1p0, flat); + flat = _mm_max_epu8(flat, _mm_srli_si128(flat, 4)); + flat = _mm_subs_epu8(flat, one); + flat = _mm_cmpeq_epi8(flat, zero); + flat = _mm_and_si128(flat, mask); + flat = _mm_unpacklo_epi32(flat, flat); + flat = _mm_unpacklo_epi64(flat, flat); + + // if flat ==0 then flat2 is zero as well and we don't need any calc below + // sse4.1 if (0==_mm_test_all_zeros(flat,ff)) + if (0xffff != _mm_movemask_epi8(_mm_cmpeq_epi8(flat, zero))) { + // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ + // flat and wide flat calculations + __m128i q5_16, q4_16, q3_16, q2_16, q1_16, q0_16; + __m128i pq_16[7]; + const __m128i eight = _mm_set1_epi16(8); + const __m128i four = _mm_set1_epi16(4); + __m128i sum_p6; + __m128i sum_p3; + + pq_16[0] = _mm_unpacklo_epi8(*q0p0, zero); + pq_16[1] = _mm_unpacklo_epi8(*q1p1, zero); + pq_16[2] = _mm_unpacklo_epi8(*q2p2, zero); + pq_16[3] = _mm_unpacklo_epi8(*q3p3, zero); + pq_16[4] = _mm_unpacklo_epi8(*q4p4, zero); + pq_16[5] = _mm_unpacklo_epi8(*q5p5, zero); + pq_16[6] = _mm_unpacklo_epi8(*q6p6, zero); + q0_16 = _mm_srli_si128(pq_16[0], 8); + q1_16 = _mm_srli_si128(pq_16[1], 8); + q2_16 = _mm_srli_si128(pq_16[2], 8); + q3_16 = _mm_srli_si128(pq_16[3], 8); + q4_16 = _mm_srli_si128(pq_16[4], 8); + q5_16 = _mm_srli_si128(pq_16[5], 8); + + __m128i flat_p[3], flat_q[3]; + __m128i flat2_p[6], flat2_q[6]; + + __m128i work0, work0_0, work0_1, sum_p_0; + __m128i sum_p = _mm_add_epi16(pq_16[5], _mm_add_epi16(pq_16[4], pq_16[3])); + __m128i sum_lp = _mm_add_epi16(pq_16[0], _mm_add_epi16(pq_16[2], pq_16[1])); + sum_p = _mm_add_epi16(sum_p, sum_lp); + + __m128i sum_lq = _mm_srli_si128(sum_lp, 8); + __m128i sum_q = _mm_srli_si128(sum_p, 8); + + sum_p_0 = _mm_add_epi16(eight, _mm_add_epi16(sum_p, sum_q)); + sum_lp = _mm_add_epi16(four, _mm_add_epi16(sum_lp, sum_lq)); + + flat_p[0] = _mm_add_epi16(sum_lp, _mm_add_epi16(pq_16[3], pq_16[0])); + flat_q[0] = _mm_add_epi16(sum_lp, _mm_add_epi16(q3_16, q0_16)); + + sum_p6 = _mm_add_epi16(pq_16[6], pq_16[6]); + sum_p3 = _mm_add_epi16(pq_16[3], pq_16[3]); + + sum_q = _mm_sub_epi16(sum_p_0, pq_16[5]); + sum_p = _mm_sub_epi16(sum_p_0, q5_16); + + work0_0 = _mm_add_epi16(_mm_add_epi16(pq_16[6], pq_16[0]), pq_16[1]); + work0_1 = _mm_add_epi16( + sum_p6, _mm_add_epi16(pq_16[1], _mm_add_epi16(pq_16[2], pq_16[0]))); + + sum_lq = _mm_sub_epi16(sum_lp, pq_16[2]); + sum_lp = _mm_sub_epi16(sum_lp, q2_16); + + work0 = _mm_add_epi16(sum_p3, pq_16[1]); + flat_p[1] = _mm_add_epi16(sum_lp, work0); + flat_q[1] = _mm_add_epi16(sum_lq, _mm_srli_si128(work0, 8)); + + flat_pq[0] = _mm_srli_epi16(_mm_unpacklo_epi64(flat_p[0], flat_q[0]), 3); + flat_pq[1] = _mm_srli_epi16(_mm_unpacklo_epi64(flat_p[1], flat_q[1]), 3); + flat_pq[0] = _mm_packus_epi16(flat_pq[0], flat_pq[0]); + flat_pq[1] = _mm_packus_epi16(flat_pq[1], flat_pq[1]); + + sum_lp = _mm_sub_epi16(sum_lp, q1_16); + sum_lq = _mm_sub_epi16(sum_lq, pq_16[1]); + + sum_p3 = _mm_add_epi16(sum_p3, pq_16[3]); + work0 = _mm_add_epi16(sum_p3, pq_16[2]); + + flat_p[2] = _mm_add_epi16(sum_lp, work0); + flat_q[2] = _mm_add_epi16(sum_lq, _mm_srli_si128(work0, 8)); + flat_pq[2] = _mm_srli_epi16(_mm_unpacklo_epi64(flat_p[2], flat_q[2]), 3); + flat_pq[2] = _mm_packus_epi16(flat_pq[2], flat_pq[2]); + + // ~~~~~~~~~~~~~~~~~~~~~~~~~~~ flat 2 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ + flat2 = _mm_max_epu8(abs_diff(*q4p4, *q0p0), abs_diff(*q5p5, *q0p0)); + + work = abs_diff(*q6p6, *q0p0); + flat2 = _mm_max_epu8(work, flat2); + flat2 = _mm_max_epu8(flat2, _mm_srli_si128(flat2, 4)); + flat2 = _mm_subs_epu8(flat2, one); + flat2 = _mm_cmpeq_epi8(flat2, zero); + flat2 = _mm_and_si128(flat2, flat); // flat2 & flat & mask + flat2 = _mm_unpacklo_epi32(flat2, flat2); + + // ~~~~~~~~~~ apply flat ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ + qs0ps0 = _mm_andnot_si128(flat, qs0ps0); + flat_pq[0] = _mm_and_si128(flat, flat_pq[0]); + *q0p0 = _mm_or_si128(qs0ps0, flat_pq[0]); + + qs1ps1 = _mm_andnot_si128(flat, qs1ps1); + flat_pq[1] = _mm_and_si128(flat, flat_pq[1]); + *q1p1 = _mm_or_si128(qs1ps1, flat_pq[1]); + + *q2p2 = _mm_andnot_si128(flat, *q2p2); + flat_pq[2] = _mm_and_si128(flat, flat_pq[2]); + *q2p2 = _mm_or_si128(*q2p2, flat_pq[2]); + + if (0xffff != _mm_movemask_epi8(_mm_cmpeq_epi8(flat2, zero))) { + flat2_p[0] = _mm_add_epi16(sum_p_0, _mm_add_epi16(work0_0, q0_16)); + flat2_q[0] = _mm_add_epi16( + sum_p_0, _mm_add_epi16(_mm_srli_si128(work0_0, 8), pq_16[0])); + + flat2_p[1] = _mm_add_epi16(sum_p, work0_1); + flat2_q[1] = _mm_add_epi16(sum_q, _mm_srli_si128(work0_1, 8)); + + flat2_pq[0] = + _mm_srli_epi16(_mm_unpacklo_epi64(flat2_p[0], flat2_q[0]), 4); + flat2_pq[1] = + _mm_srli_epi16(_mm_unpacklo_epi64(flat2_p[1], flat2_q[1]), 4); + flat2_pq[0] = _mm_packus_epi16(flat2_pq[0], flat2_pq[0]); + flat2_pq[1] = _mm_packus_epi16(flat2_pq[1], flat2_pq[1]); + + sum_p = _mm_sub_epi16(sum_p, q4_16); + sum_q = _mm_sub_epi16(sum_q, pq_16[4]); + + sum_p6 = _mm_add_epi16(sum_p6, pq_16[6]); + work0 = _mm_add_epi16( + sum_p6, _mm_add_epi16(pq_16[2], _mm_add_epi16(pq_16[3], pq_16[1]))); + flat2_p[2] = _mm_add_epi16(sum_p, work0); + flat2_q[2] = _mm_add_epi16(sum_q, _mm_srli_si128(work0, 8)); + flat2_pq[2] = + _mm_srli_epi16(_mm_unpacklo_epi64(flat2_p[2], flat2_q[2]), 4); + flat2_pq[2] = _mm_packus_epi16(flat2_pq[2], flat2_pq[2]); + + sum_p6 = _mm_add_epi16(sum_p6, pq_16[6]); + sum_p = _mm_sub_epi16(sum_p, q3_16); + sum_q = _mm_sub_epi16(sum_q, pq_16[3]); + + work0 = _mm_add_epi16( + sum_p6, _mm_add_epi16(pq_16[3], _mm_add_epi16(pq_16[4], pq_16[2]))); + flat2_p[3] = _mm_add_epi16(sum_p, work0); + flat2_q[3] = _mm_add_epi16(sum_q, _mm_srli_si128(work0, 8)); + flat2_pq[3] = + _mm_srli_epi16(_mm_unpacklo_epi64(flat2_p[3], flat2_q[3]), 4); + flat2_pq[3] = _mm_packus_epi16(flat2_pq[3], flat2_pq[3]); + + sum_p6 = _mm_add_epi16(sum_p6, pq_16[6]); + sum_p = _mm_sub_epi16(sum_p, q2_16); + sum_q = _mm_sub_epi16(sum_q, pq_16[2]); + + work0 = _mm_add_epi16( + sum_p6, _mm_add_epi16(pq_16[4], _mm_add_epi16(pq_16[5], pq_16[3]))); + flat2_p[4] = _mm_add_epi16(sum_p, work0); + flat2_q[4] = _mm_add_epi16(sum_q, _mm_srli_si128(work0, 8)); + flat2_pq[4] = + _mm_srli_epi16(_mm_unpacklo_epi64(flat2_p[4], flat2_q[4]), 4); + flat2_pq[4] = _mm_packus_epi16(flat2_pq[4], flat2_pq[4]); + + sum_p6 = _mm_add_epi16(sum_p6, pq_16[6]); + sum_p = _mm_sub_epi16(sum_p, q1_16); + sum_q = _mm_sub_epi16(sum_q, pq_16[1]); + + work0 = _mm_add_epi16( + sum_p6, _mm_add_epi16(pq_16[5], _mm_add_epi16(pq_16[6], pq_16[4]))); + flat2_p[5] = _mm_add_epi16(sum_p, work0); + flat2_q[5] = _mm_add_epi16(sum_q, _mm_srli_si128(work0, 8)); + flat2_pq[5] = + _mm_srli_epi16(_mm_unpacklo_epi64(flat2_p[5], flat2_q[5]), 4); + flat2_pq[5] = _mm_packus_epi16(flat2_pq[5], flat2_pq[5]); + + // wide flat + // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ + + *q0p0 = _mm_andnot_si128(flat2, *q0p0); + flat2_pq[0] = _mm_and_si128(flat2, flat2_pq[0]); + *q0p0 = _mm_or_si128(*q0p0, flat2_pq[0]); + + *q1p1 = _mm_andnot_si128(flat2, *q1p1); + flat2_pq[1] = _mm_and_si128(flat2, flat2_pq[1]); + *q1p1 = _mm_or_si128(*q1p1, flat2_pq[1]); + + *q2p2 = _mm_andnot_si128(flat2, *q2p2); + flat2_pq[2] = _mm_and_si128(flat2, flat2_pq[2]); + *q2p2 = _mm_or_si128(*q2p2, flat2_pq[2]); + + *q3p3 = _mm_andnot_si128(flat2, *q3p3); + flat2_pq[3] = _mm_and_si128(flat2, flat2_pq[3]); + *q3p3 = _mm_or_si128(*q3p3, flat2_pq[3]); + + *q4p4 = _mm_andnot_si128(flat2, *q4p4); + flat2_pq[4] = _mm_and_si128(flat2, flat2_pq[4]); + *q4p4 = _mm_or_si128(*q4p4, flat2_pq[4]); + + *q5p5 = _mm_andnot_si128(flat2, *q5p5); + flat2_pq[5] = _mm_and_si128(flat2, flat2_pq[5]); + *q5p5 = _mm_or_si128(*q5p5, flat2_pq[5]); + } + } else { + *q0p0 = qs0ps0; + *q1p1 = qs1ps1; + } +} + +void aom_lpf_horizontal_14_sse2(unsigned char *s, int p, + const unsigned char *_blimit, + const unsigned char *_limit, + const unsigned char *_thresh) { + __m128i q6p6, q5p5, q4p4, q3p3, q2p2, q1p1, q0p0; + __m128i blimit = _mm_load_si128((const __m128i *)_blimit); + __m128i limit = _mm_load_si128((const __m128i *)_limit); + __m128i thresh = _mm_load_si128((const __m128i *)_thresh); + + q4p4 = _mm_unpacklo_epi32(_mm_cvtsi32_si128(*(int *)(s - 5 * p)), + _mm_cvtsi32_si128(*(int *)(s + 4 * p))); + q3p3 = _mm_unpacklo_epi32(_mm_cvtsi32_si128(*(int *)(s - 4 * p)), + _mm_cvtsi32_si128(*(int *)(s + 3 * p))); + q2p2 = _mm_unpacklo_epi32(_mm_cvtsi32_si128(*(int *)(s - 3 * p)), + _mm_cvtsi32_si128(*(int *)(s + 2 * p))); + q1p1 = _mm_unpacklo_epi32(_mm_cvtsi32_si128(*(int *)(s - 2 * p)), + _mm_cvtsi32_si128(*(int *)(s + 1 * p))); + + q0p0 = _mm_unpacklo_epi32(_mm_cvtsi32_si128(*(int *)(s - 1 * p)), + _mm_cvtsi32_si128(*(int *)(s - 0 * p))); + + q5p5 = _mm_unpacklo_epi32(_mm_cvtsi32_si128(*(int *)(s - 6 * p)), + _mm_cvtsi32_si128(*(int *)(s + 5 * p))); + + q6p6 = _mm_unpacklo_epi32(_mm_cvtsi32_si128(*(int *)(s - 7 * p)), + _mm_cvtsi32_si128(*(int *)(s + 6 * p))); + + lpf_internal_14_sse2(&q6p6, &q5p5, &q4p4, &q3p3, &q2p2, &q1p1, &q0p0, &blimit, + &limit, &thresh); + + store_buffer_horz_8(q0p0, p, 0, s); + store_buffer_horz_8(q1p1, p, 1, s); + store_buffer_horz_8(q2p2, p, 2, s); + store_buffer_horz_8(q3p3, p, 3, s); + store_buffer_horz_8(q4p4, p, 4, s); + store_buffer_horz_8(q5p5, p, 5, s); +} + +static AOM_FORCE_INLINE void lpf_internal_6_dual_sse2( + __m128i *p2, __m128i *q2, __m128i *p1, __m128i *q1, __m128i *p0, + __m128i *q0, __m128i *q1q0, __m128i *p1p0, __m128i *blimit, __m128i *limit, + __m128i *thresh) { + const __m128i zero = _mm_setzero_si128(); + __m128i mask, hev, flat; + __m128i q2p2, q1p1, q0p0, flat_p1p0, flat_q0q1; + __m128i p2_16, q2_16, p1_16, q1_16, p0_16, q0_16; + __m128i ps1ps0, qs1qs0; + + q2p2 = _mm_unpacklo_epi64(*p2, *q2); + q1p1 = _mm_unpacklo_epi64(*p1, *q1); + q0p0 = _mm_unpacklo_epi64(*p0, *q0); + + *p1p0 = _mm_unpacklo_epi64(q0p0, q1p1); + *q1q0 = _mm_unpackhi_epi64(q0p0, q1p1); + + const __m128i one = _mm_set1_epi8(1); + const __m128i fe = _mm_set1_epi8(0xfe); + const __m128i ff = _mm_cmpeq_epi8(fe, fe); + + { + // filter_mask and hev_mask + __m128i abs_p1q1, abs_p0q0, abs_q1q0, abs_p1p0, work; + abs_p1p0 = abs_diff(q1p1, q0p0); + abs_q1q0 = _mm_srli_si128(abs_p1p0, 8); + + abs_p0q0 = abs_diff(*p1p0, *q1q0); + abs_p1q1 = _mm_srli_si128(abs_p0q0, 8); + abs_p0q0 = _mm_unpacklo_epi64(abs_p0q0, zero); + + // considering sse doesn't have unsigned elements comparison the idea is + // to find at least one case when X > limit, it means the corresponding + // mask bit is set. + // to achieve that we find global max value of all inputs of abs(x-y) or + // (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 If it is > limit the mask is set + // otherwise - not + + flat = _mm_max_epu8(abs_p1p0, abs_q1q0); + hev = _mm_subs_epu8(flat, *thresh); + hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff); + // replicate for the further "merged variables" usage + hev = _mm_unpacklo_epi64(hev, hev); + + abs_p0q0 = _mm_adds_epu8(abs_p0q0, abs_p0q0); + abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1); + mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), *blimit); + mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff); + // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit) * -1; + mask = _mm_max_epu8(abs_p1p0, mask); + // mask |= (abs(p1 - p0) > limit) * -1; + // mask |= (abs(q1 - q0) > limit) * -1; + + work = abs_diff(q2p2, q1p1); + mask = _mm_max_epu8(work, mask); + mask = _mm_max_epu8(mask, _mm_srli_si128(mask, 8)); + mask = _mm_subs_epu8(mask, *limit); + mask = _mm_cmpeq_epi8(mask, zero); + + // lp filter - the same for 6, 8 and 14 versions + filter4_dual_sse2(p1p0, q1q0, &hev, &mask, q1q0, p1p0); + + // flat_mask + flat = _mm_max_epu8(abs_diff(q2p2, q0p0), abs_p1p0); + flat = _mm_max_epu8(flat, _mm_srli_si128(flat, 8)); + flat = _mm_subs_epu8(flat, one); + flat = _mm_cmpeq_epi8(flat, zero); + flat = _mm_and_si128(flat, mask); + // replicate for the further "merged variables" usage + flat = _mm_unpacklo_epi64(flat, flat); + } + + // 5 tap filter + // need it only if flat !=0 + if (0xffff != _mm_movemask_epi8(_mm_cmpeq_epi8(flat, zero))) { + const __m128i four = _mm_set1_epi16(4); + __m128i workp_a, workp_b, workp_shft0, workp_shft1; + p2_16 = _mm_unpacklo_epi8(*p2, zero); + p1_16 = _mm_unpacklo_epi8(*p1, zero); + p0_16 = _mm_unpacklo_epi8(*p0, zero); + q0_16 = _mm_unpacklo_epi8(*q0, zero); + q1_16 = _mm_unpacklo_epi8(*q1, zero); + q2_16 = _mm_unpacklo_epi8(*q2, zero); + + // op1 + workp_a = _mm_add_epi16(_mm_add_epi16(p0_16, p0_16), + _mm_add_epi16(p1_16, p1_16)); // p0 *2 + p1 * 2 + workp_a = _mm_add_epi16(_mm_add_epi16(workp_a, four), + p2_16); // p2 + p0 * 2 + p1 * 2 + 4 + + workp_b = _mm_add_epi16(_mm_add_epi16(p2_16, p2_16), q0_16); + workp_shft0 = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), + 3); // p2 * 3 + p1 * 2 + p0 * 2 + q0 + 4 + + // op0 + workp_b = _mm_add_epi16(_mm_add_epi16(q0_16, q0_16), q1_16); // q0 * 2 + q1 + workp_a = _mm_add_epi16(workp_a, + workp_b); // p2 + p0 * 2 + p1 * 2 + q0 * 2 + q1 + 4 + workp_shft1 = _mm_srli_epi16(workp_a, 3); + + flat_p1p0 = _mm_packus_epi16(workp_shft1, workp_shft0); + + // oq0 + workp_a = _mm_sub_epi16(_mm_sub_epi16(workp_a, p2_16), + p1_16); // p0 * 2 + p1 + q0 * 2 + q1 + 4 + workp_b = _mm_add_epi16(q1_16, q2_16); + workp_a = _mm_add_epi16( + workp_a, workp_b); // p0 * 2 + p1 + q0 * 2 + q1 * 2 + q2 + 4 + workp_shft0 = _mm_srli_epi16(workp_a, 3); + + // oq1 + workp_a = _mm_sub_epi16(_mm_sub_epi16(workp_a, p1_16), + p0_16); // p0 + q0 * 2 + q1 * 2 + q2 + 4 + workp_b = _mm_add_epi16(q2_16, q2_16); + workp_shft1 = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), + 3); // p0 + q0 * 2 + q1 * 2 + q2 * 3 + 4 + + flat_q0q1 = _mm_packus_epi16(workp_shft0, workp_shft1); + + qs1qs0 = _mm_andnot_si128(flat, *q1q0); + *q1q0 = _mm_and_si128(flat, flat_q0q1); + *q1q0 = _mm_or_si128(qs1qs0, *q1q0); + + ps1ps0 = _mm_andnot_si128(flat, *p1p0); + *p1p0 = _mm_and_si128(flat, flat_p1p0); + *p1p0 = _mm_or_si128(ps1ps0, *p1p0); + } +} + +static AOM_FORCE_INLINE void lpf_internal_6_sse2( + __m128i *p2, __m128i *q2, __m128i *p1, __m128i *q1, __m128i *p0, + __m128i *q0, __m128i *q1q0, __m128i *p1p0, __m128i *blimit, __m128i *limit, + __m128i *thresh) { + const __m128i zero = _mm_setzero_si128(); + __m128i mask, hev, flat; + __m128i q2p2, q1p1, q0p0, flat_p1p0, flat_q0q1; + __m128i pq2_16, q2_16, pq1_16, pq0_16, q0_16; + __m128i ps1ps0, qs1qs0; + + q2p2 = _mm_unpacklo_epi32(*p2, *q2); + q1p1 = _mm_unpacklo_epi32(*p1, *q1); + q0p0 = _mm_unpacklo_epi32(*p0, *q0); + + *p1p0 = _mm_unpacklo_epi32(*p0, *p1); + *q1q0 = _mm_unpacklo_epi32(*q0, *q1); + + const __m128i one = _mm_set1_epi8(1); + const __m128i fe = _mm_set1_epi8(0xfe); + const __m128i ff = _mm_cmpeq_epi8(fe, fe); + { + // filter_mask and hev_mask + __m128i abs_p1q1, abs_p0q0, abs_q1q0, abs_p1p0, work; + abs_p1p0 = abs_diff(q1p1, q0p0); + abs_q1q0 = _mm_srli_si128(abs_p1p0, 4); + + abs_p0q0 = abs_diff(*p1p0, *q1q0); + abs_p1q1 = _mm_srli_si128(abs_p0q0, 4); + + // considering sse doesn't have unsigned elements comparison the idea is + // to find at least one case when X > limit, it means the corresponding + // mask bit is set. + // to achieve that we find global max value of all inputs of abs(x-y) or + // (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 If it is > limit the mask is set + // otherwise - not + + flat = _mm_max_epu8(abs_p1p0, abs_q1q0); + hev = _mm_subs_epu8(flat, *thresh); + hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff); + // replicate for the further "merged variables" usage + hev = _mm_unpacklo_epi32(hev, hev); + + abs_p0q0 = _mm_adds_epu8(abs_p0q0, abs_p0q0); + abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1); + mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), *blimit); + mask = _mm_unpacklo_epi32(mask, zero); + mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff); + // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit) * -1; + mask = _mm_max_epu8(abs_p1p0, mask); + // mask |= (abs(p1 - p0) > limit) * -1; + // mask |= (abs(q1 - q0) > limit) * -1; + + work = abs_diff(q2p2, q1p1); + mask = _mm_max_epu8(work, mask); + mask = _mm_max_epu8(mask, _mm_srli_si128(mask, 4)); + mask = _mm_subs_epu8(mask, *limit); + mask = _mm_cmpeq_epi8(mask, zero); + + // lp filter - the same for 6, 8 and 14 versions + filter4_sse2(p1p0, q1q0, &hev, &mask, q1q0, p1p0); + + // flat_mask + flat = _mm_max_epu8(abs_diff(q2p2, q0p0), abs_p1p0); + flat = _mm_max_epu8(flat, _mm_srli_si128(flat, 4)); + flat = _mm_subs_epu8(flat, one); + flat = _mm_cmpeq_epi8(flat, zero); + flat = _mm_and_si128(flat, mask); + // replicate for the further "merged variables" usage + flat = _mm_unpacklo_epi32(flat, flat); + flat = _mm_unpacklo_epi64(flat, flat); + } + + // 5 tap filter + // need it only if flat !=0 + if (0xffff != _mm_movemask_epi8(_mm_cmpeq_epi8(flat, zero))) { + const __m128i four = _mm_set1_epi16(4); + __m128i workp_a, workp_b, workp_c; + __m128i pq0x2_pq1, pq1_pq2; + pq2_16 = _mm_unpacklo_epi8(q2p2, zero); + pq1_16 = _mm_unpacklo_epi8(q1p1, zero); + pq0_16 = _mm_unpacklo_epi8(q0p0, zero); + q0_16 = _mm_srli_si128(pq0_16, 8); + q2_16 = _mm_srli_si128(pq2_16, 8); + + // op1 + pq0x2_pq1 = + _mm_add_epi16(_mm_add_epi16(pq0_16, pq0_16), pq1_16); // p0 *2 + p1 + pq1_pq2 = _mm_add_epi16(pq1_16, pq2_16); // p1 + p2 + workp_a = _mm_add_epi16(_mm_add_epi16(pq0x2_pq1, four), + pq1_pq2); // p2 + p0 * 2 + p1 * 2 + 4 + + workp_b = _mm_add_epi16(_mm_add_epi16(pq2_16, pq2_16), q0_16); + workp_b = + _mm_add_epi16(workp_a, workp_b); // p2 * 3 + p1 * 2 + p0 * 2 + q0 + 4 + + // op0 + workp_c = _mm_srli_si128(pq0x2_pq1, 8); // q0 * 2 + q1 + workp_a = _mm_add_epi16(workp_a, + workp_c); // p2 + p0 * 2 + p1 * 2 + q0 * 2 + q1 + 4 + workp_b = _mm_unpacklo_epi64(workp_a, workp_b); + workp_b = _mm_srli_epi16(workp_b, 3); + + flat_p1p0 = _mm_packus_epi16(workp_b, workp_b); + + // oq0 + workp_a = _mm_sub_epi16(_mm_sub_epi16(workp_a, pq2_16), + pq1_16); // p0 * 2 + p1 + q0 * 2 + q1 + 4 + workp_b = _mm_srli_si128(pq1_pq2, 8); + workp_a = _mm_add_epi16( + workp_a, workp_b); // p0 * 2 + p1 + q0 * 2 + q1 * 2 + q2 + 4 + // workp_shft0 = _mm_srli_epi16(workp_a, 3); + + // oq1 + workp_c = _mm_sub_epi16(_mm_sub_epi16(workp_a, pq1_16), + pq0_16); // p0 + q0 * 2 + q1 * 2 + q2 + 4 + workp_b = _mm_add_epi16(q2_16, q2_16); + workp_b = + _mm_add_epi16(workp_c, workp_b); // p0 + q0 * 2 + q1 * 2 + q2 * 3 + 4 + + workp_a = _mm_unpacklo_epi64(workp_a, workp_b); + workp_a = _mm_srli_epi16(workp_a, 3); + + flat_q0q1 = _mm_packus_epi16(workp_a, workp_a); + + qs1qs0 = _mm_andnot_si128(flat, *q1q0); + *q1q0 = _mm_and_si128(flat, flat_q0q1); + *q1q0 = _mm_or_si128(qs1qs0, *q1q0); + + ps1ps0 = _mm_andnot_si128(flat, *p1p0); + *p1p0 = _mm_and_si128(flat, flat_p1p0); + *p1p0 = _mm_or_si128(ps1ps0, *p1p0); + } +} + +void aom_lpf_horizontal_6_sse2(unsigned char *s, int p, + const unsigned char *_blimit, + const unsigned char *_limit, + const unsigned char *_thresh) { + __m128i p2, p1, p0, q0, q1, q2; + __m128i p1p0, q1q0; + __m128i blimit = _mm_load_si128((__m128i *)_blimit); + __m128i limit = _mm_load_si128((__m128i *)_limit); + __m128i thresh = _mm_load_si128((__m128i *)_thresh); + + p2 = _mm_cvtsi32_si128(*(int *)(s - 3 * p)); + p1 = _mm_cvtsi32_si128(*(int *)(s - 2 * p)); + p0 = _mm_cvtsi32_si128(*(int *)(s - 1 * p)); + q0 = _mm_cvtsi32_si128(*(int *)(s - 0 * p)); + q1 = _mm_cvtsi32_si128(*(int *)(s + 1 * p)); + q2 = _mm_cvtsi32_si128(*(int *)(s + 2 * p)); + + lpf_internal_6_sse2(&p2, &q2, &p1, &q1, &p0, &q0, &q1q0, &p1p0, &blimit, + &limit, &thresh); + + xx_storel_32(s - 1 * p, p1p0); + xx_storel_32(s - 2 * p, _mm_srli_si128(p1p0, 4)); + xx_storel_32(s + 0 * p, q1q0); + xx_storel_32(s + 1 * p, _mm_srli_si128(q1q0, 4)); +} + +void aom_lpf_horizontal_6_dual_sse2(unsigned char *s, int p, + const unsigned char *_blimit0, + const unsigned char *_limit0, + const unsigned char *_thresh0, + const unsigned char *_blimit1, + const unsigned char *_limit1, + const unsigned char *_thresh1) { + __m128i blimit = _mm_unpacklo_epi32(_mm_load_si128((__m128i *)_blimit0), + _mm_load_si128((__m128i *)_blimit1)); + __m128i limit = _mm_unpacklo_epi32(_mm_load_si128((__m128i *)_limit0), + _mm_load_si128((__m128i *)_limit1)); + __m128i thresh = _mm_unpacklo_epi32(_mm_load_si128((__m128i *)_thresh0), + _mm_load_si128((__m128i *)_thresh1)); + + __m128i p2, p1, p0, q0, q1, q2; + __m128i p1p0, q1q0; + + p2 = _mm_loadl_epi64((__m128i *)(s - 3 * p)); + p1 = _mm_loadl_epi64((__m128i *)(s - 2 * p)); + p0 = _mm_loadl_epi64((__m128i *)(s - 1 * p)); + q0 = _mm_loadl_epi64((__m128i *)(s - 0 * p)); + q1 = _mm_loadl_epi64((__m128i *)(s + 1 * p)); + q2 = _mm_loadl_epi64((__m128i *)(s + 2 * p)); + + lpf_internal_6_dual_sse2(&p2, &q2, &p1, &q1, &p0, &q0, &q1q0, &p1p0, &blimit, + &limit, &thresh); + + _mm_storel_epi64((__m128i *)(s - 1 * p), p1p0); + _mm_storel_epi64((__m128i *)(s - 2 * p), _mm_srli_si128(p1p0, 8)); + _mm_storel_epi64((__m128i *)(s + 0 * p), q1q0); + _mm_storel_epi64((__m128i *)(s + 1 * p), _mm_srli_si128(q1q0, 8)); +} + +static AOM_FORCE_INLINE void lpf_internal_8_sse2( + __m128i *p3, __m128i *q3, __m128i *p2, __m128i *q2, __m128i *p1, + __m128i *q1, __m128i *p0, __m128i *q0, __m128i *q1q0_out, __m128i *p1p0_out, + __m128i *blimit, __m128i *limit, __m128i *thresh) { + const __m128i zero = _mm_setzero_si128(); + __m128i mask, hev, flat; + __m128i p2_16, q2_16, p1_16, p0_16, q0_16, q1_16, p3_16, q3_16, q3p3, + flat_p1p0, flat_q0q1; + __m128i q2p2, q1p1, q0p0; + __m128i q1q0, p1p0, ps1ps0, qs1qs0; + __m128i work_pq, opq2, pq2; + + q3p3 = _mm_unpacklo_epi32(*p3, *q3); + q2p2 = _mm_unpacklo_epi32(*p2, *q2); + q1p1 = _mm_unpacklo_epi32(*p1, *q1); + q0p0 = _mm_unpacklo_epi32(*p0, *q0); + + p1p0 = _mm_unpacklo_epi32(q0p0, q1p1); // p1p0 q1q0 + q1q0 = _mm_srli_si128(p1p0, 8); + + // filter_mask and hev_mask + + // considering sse doesn't have unsigned elements comparison the idea is to + // find at least one case when X > limit, it means the corresponding mask + // bit is set. + // to achieve that we find global max value of all inputs of abs(x-y) or + // (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 If it is > limit the mask is set + // otherwise - not + + const __m128i one = _mm_set1_epi8(1); + const __m128i fe = _mm_set1_epi8(0xfe); + const __m128i ff = _mm_cmpeq_epi8(fe, fe); + __m128i abs_p1q1, abs_p0q0, abs_q1q0, abs_p1p0, work; + + abs_p1p0 = abs_diff(q1p1, q0p0); + abs_q1q0 = _mm_srli_si128(abs_p1p0, 4); + + abs_p0q0 = abs_diff(p1p0, q1q0); + abs_p1q1 = _mm_srli_si128(abs_p0q0, 4); + + flat = _mm_max_epu8(abs_p1p0, abs_q1q0); + hev = _mm_subs_epu8(flat, *thresh); + hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff); + // replicate for the further "merged variables" usage + hev = _mm_unpacklo_epi32(hev, hev); + + abs_p0q0 = _mm_adds_epu8(abs_p0q0, abs_p0q0); + abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1); + mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), *blimit); + mask = _mm_unpacklo_epi32(mask, zero); + mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff); + // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit) * -1; + mask = _mm_max_epu8(abs_p1p0, mask); + // mask |= (abs(p1 - p0) > limit) * -1; + // mask |= (abs(q1 - q0) > limit) * -1; + + work = _mm_max_epu8(abs_diff(q2p2, q1p1), abs_diff(q3p3, q2p2)); + + mask = _mm_max_epu8(work, mask); + mask = _mm_max_epu8(mask, _mm_srli_si128(mask, 4)); + mask = _mm_subs_epu8(mask, *limit); + mask = _mm_cmpeq_epi8(mask, zero); + + // lp filter - the same for 6, 8 and 14 versions + filter4_sse2(&p1p0, &q1q0, &hev, &mask, q1q0_out, p1p0_out); + + // flat_mask4 + flat = _mm_max_epu8(abs_diff(q2p2, q0p0), abs_diff(q3p3, q0p0)); + flat = _mm_max_epu8(abs_p1p0, flat); + + flat = _mm_max_epu8(flat, _mm_srli_si128(flat, 4)); + flat = _mm_subs_epu8(flat, one); + flat = _mm_cmpeq_epi8(flat, zero); + flat = _mm_and_si128(flat, mask); + // replicate for the further "merged variables" usage + flat = _mm_unpacklo_epi32(flat, flat); + flat = _mm_unpacklo_epi64(flat, flat); + + // filter8 need it only if flat !=0 + if (0xffff != _mm_movemask_epi8(_mm_cmpeq_epi8(flat, zero))) { + const __m128i four = _mm_set1_epi16(4); + __m128i workp_a, workp_b, workp_c, workp_d, workp_shft1, workp_shft2; + p2_16 = _mm_unpacklo_epi8(*p2, zero); + p1_16 = _mm_unpacklo_epi8(*p1, zero); + p0_16 = _mm_unpacklo_epi8(*p0, zero); + q0_16 = _mm_unpacklo_epi8(*q0, zero); + q1_16 = _mm_unpacklo_epi8(*q1, zero); + q2_16 = _mm_unpacklo_epi8(*q2, zero); + p3_16 = _mm_unpacklo_epi8(*p3, zero); + q3_16 = _mm_unpacklo_epi8(*q3, zero); + + // op2 + workp_a = + _mm_add_epi16(_mm_add_epi16(p3_16, p3_16), _mm_add_epi16(p2_16, p1_16)); + workp_a = _mm_add_epi16(_mm_add_epi16(workp_a, four), p0_16); + workp_b = _mm_add_epi16(_mm_add_epi16(q0_16, p2_16), p3_16); + workp_shft2 = _mm_add_epi16(workp_a, workp_b); + + // op1 + workp_b = _mm_add_epi16(_mm_add_epi16(q0_16, q1_16), p1_16); + workp_c = _mm_add_epi16(workp_a, workp_b); + // workp_shft0 = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3); + + // op0 + workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p3_16), q2_16); + workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p1_16), p0_16); + workp_d = _mm_add_epi16(workp_a, workp_b); + // workp_shft1 = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3); + + workp_c = _mm_unpacklo_epi64(workp_d, workp_c); + workp_c = _mm_srli_epi16(workp_c, 3); + flat_p1p0 = _mm_packus_epi16(workp_c, workp_c); + + // oq0 + workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p3_16), q3_16); + workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p0_16), q0_16); + // workp_shft0 = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3); + workp_c = _mm_add_epi16(workp_a, workp_b); + + // oq1 + workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p2_16), q3_16); + workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q0_16), q1_16); + workp_d = _mm_add_epi16(workp_a, workp_b); + // workp_shft1 = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3); + + workp_c = _mm_unpacklo_epi64(workp_c, workp_d); + workp_c = _mm_srli_epi16(workp_c, 3); + flat_q0q1 = _mm_packus_epi16(workp_c, workp_c); + + // oq2 + workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p1_16), q3_16); + workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q1_16), q2_16); + workp_shft1 = _mm_add_epi16(workp_a, workp_b); + + workp_c = _mm_unpacklo_epi64(workp_shft2, workp_shft1); + workp_c = _mm_srli_epi16(workp_c, 3); + + opq2 = _mm_packus_epi16(workp_c, workp_c); + + work_pq = _mm_andnot_si128(flat, q2p2); + pq2 = _mm_and_si128(flat, opq2); + *p2 = _mm_or_si128(work_pq, pq2); + *q2 = _mm_srli_si128(*p2, 4); + + qs1qs0 = _mm_andnot_si128(flat, *q1q0_out); + q1q0 = _mm_and_si128(flat, flat_q0q1); + *q1q0_out = _mm_or_si128(qs1qs0, q1q0); + + ps1ps0 = _mm_andnot_si128(flat, *p1p0_out); + p1p0 = _mm_and_si128(flat, flat_p1p0); + *p1p0_out = _mm_or_si128(ps1ps0, p1p0); + } +} + +static AOM_FORCE_INLINE void lpf_internal_8_dual_sse2( + __m128i *p3, __m128i *q3, __m128i *p2, __m128i *q2, __m128i *p1, + __m128i *q1, __m128i *p0, __m128i *q0, __m128i *q1q0_out, __m128i *p1p0_out, + __m128i *blimit, __m128i *limit, __m128i *thresh) { + const __m128i zero = _mm_setzero_si128(); + __m128i mask, hev, flat; + __m128i p2_16, q2_16, p1_16, p0_16, q0_16, q1_16, p3_16, q3_16, q3p3, + flat_p1p0, flat_q0q1; + __m128i q2p2, q1p1, q0p0; + __m128i q1q0, p1p0, ps1ps0, qs1qs0; + __m128i work_pq, opq2, pq2; + + q3p3 = _mm_unpacklo_epi64(*p3, *q3); + q2p2 = _mm_unpacklo_epi64(*p2, *q2); + q1p1 = _mm_unpacklo_epi64(*p1, *q1); + q0p0 = _mm_unpacklo_epi64(*p0, *q0); + + p1p0 = _mm_unpacklo_epi64(q0p0, q1p1); + q1q0 = _mm_unpackhi_epi64(q0p0, q1p1); + + { + // filter_mask and hev_mask + + // considering sse doesn't have unsigned elements comparison the idea is to + // find at least one case when X > limit, it means the corresponding mask + // bit is set. + // to achieve that we find global max value of all inputs of abs(x-y) or + // (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 If it is > limit the mask is set + // otherwise - not + + const __m128i one = _mm_set1_epi8(1); + const __m128i fe = _mm_set1_epi8(0xfe); + const __m128i ff = _mm_cmpeq_epi8(fe, fe); + __m128i abs_p1q1, abs_p0q0, abs_q1q0, abs_p1p0, work; + + abs_p1p0 = abs_diff(q1p1, q0p0); + abs_q1q0 = _mm_srli_si128(abs_p1p0, 8); + + abs_p0q0 = abs_diff(p1p0, q1q0); + abs_p1q1 = _mm_srli_si128(abs_p0q0, 8); + abs_p0q0 = _mm_unpacklo_epi64(abs_p0q0, abs_p0q0); + + flat = _mm_max_epu8(abs_p1p0, abs_q1q0); + hev = _mm_subs_epu8(flat, *thresh); + hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff); + // replicate for the further "merged variables" usage + hev = _mm_unpacklo_epi64(hev, hev); + + abs_p0q0 = _mm_adds_epu8(abs_p0q0, abs_p0q0); + abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1); + mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), *blimit); + mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff); + // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit) * -1; + mask = _mm_max_epu8(abs_p1p0, mask); + // mask |= (abs(p1 - p0) > limit) * -1; + // mask |= (abs(q1 - q0) > limit) * -1; + + work = _mm_max_epu8(abs_diff(q2p2, q1p1), abs_diff(q3p3, q2p2)); + + mask = _mm_max_epu8(work, mask); + mask = _mm_max_epu8(mask, _mm_srli_si128(mask, 8)); + mask = _mm_subs_epu8(mask, *limit); + mask = _mm_cmpeq_epi8(mask, zero); + + // lp filter - the same for 6, 8 and 14 versions + filter4_dual_sse2(&p1p0, &q1q0, &hev, &mask, q1q0_out, p1p0_out); + + // flat_mask4 + flat = _mm_max_epu8(abs_diff(q2p2, q0p0), abs_diff(q3p3, q0p0)); + flat = _mm_max_epu8(abs_p1p0, flat); + + flat = _mm_max_epu8(flat, _mm_srli_si128(flat, 8)); + flat = _mm_subs_epu8(flat, one); + flat = _mm_cmpeq_epi8(flat, zero); + flat = _mm_and_si128(flat, mask); + // replicate for the further "merged variables" usage + flat = _mm_unpacklo_epi64(flat, flat); + } + + // filter8 need it only if flat !=0 + if (0xffff != _mm_movemask_epi8(_mm_cmpeq_epi8(flat, zero))) { + const __m128i four = _mm_set1_epi16(4); + + __m128i workp_a, workp_b, workp_shft0, workp_shft1, workp_shft2; + p2_16 = _mm_unpacklo_epi8(*p2, zero); + p1_16 = _mm_unpacklo_epi8(*p1, zero); + p0_16 = _mm_unpacklo_epi8(*p0, zero); + q0_16 = _mm_unpacklo_epi8(*q0, zero); + q1_16 = _mm_unpacklo_epi8(*q1, zero); + q2_16 = _mm_unpacklo_epi8(*q2, zero); + p3_16 = _mm_unpacklo_epi8(*p3, zero); + q3_16 = _mm_unpacklo_epi8(*q3, zero); + + // op2 + workp_a = + _mm_add_epi16(_mm_add_epi16(p3_16, p3_16), _mm_add_epi16(p2_16, p1_16)); + workp_a = _mm_add_epi16(_mm_add_epi16(workp_a, four), p0_16); + workp_b = _mm_add_epi16(_mm_add_epi16(q0_16, p2_16), p3_16); + workp_shft2 = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3); + + // op1 + workp_b = _mm_add_epi16(_mm_add_epi16(q0_16, q1_16), p1_16); + workp_shft0 = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3); + + // op0 + workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p3_16), q2_16); + workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p1_16), p0_16); + workp_shft1 = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3); + + flat_p1p0 = _mm_packus_epi16(workp_shft1, workp_shft0); + + // oq0 + workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p3_16), q3_16); + workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p0_16), q0_16); + workp_shft0 = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3); + + // oq1 + workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p2_16), q3_16); + workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q0_16), q1_16); + workp_shft1 = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3); + + flat_q0q1 = _mm_packus_epi16(workp_shft0, workp_shft1); + + // oq2 + workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p1_16), q3_16); + workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q1_16), q2_16); + workp_shft1 = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3); + + opq2 = _mm_packus_epi16(workp_shft2, workp_shft1); + + work_pq = _mm_andnot_si128(flat, q2p2); + pq2 = _mm_and_si128(flat, opq2); + *p2 = _mm_or_si128(work_pq, pq2); + *q2 = _mm_srli_si128(*p2, 8); + + qs1qs0 = _mm_andnot_si128(flat, *q1q0_out); + q1q0 = _mm_and_si128(flat, flat_q0q1); + *q1q0_out = _mm_or_si128(qs1qs0, q1q0); + + ps1ps0 = _mm_andnot_si128(flat, *p1p0_out); + p1p0 = _mm_and_si128(flat, flat_p1p0); + *p1p0_out = _mm_or_si128(ps1ps0, p1p0); + } +} + +void aom_lpf_horizontal_8_sse2(unsigned char *s, int p, + const unsigned char *_blimit, + const unsigned char *_limit, + const unsigned char *_thresh) { + __m128i p2, p1, p0, q0, q1, q2, p3, q3; + __m128i q1q0, p1p0; + __m128i blimit = _mm_load_si128((const __m128i *)_blimit); + __m128i limit = _mm_load_si128((const __m128i *)_limit); + __m128i thresh = _mm_load_si128((const __m128i *)_thresh); + + p3 = _mm_cvtsi32_si128(*(int *)(s - 4 * p)); + p2 = _mm_cvtsi32_si128(*(int *)(s - 3 * p)); + p1 = _mm_cvtsi32_si128(*(int *)(s - 2 * p)); + p0 = _mm_cvtsi32_si128(*(int *)(s - 1 * p)); + q0 = _mm_cvtsi32_si128(*(int *)(s - 0 * p)); + q1 = _mm_cvtsi32_si128(*(int *)(s + 1 * p)); + q2 = _mm_cvtsi32_si128(*(int *)(s + 2 * p)); + q3 = _mm_cvtsi32_si128(*(int *)(s + 3 * p)); + + lpf_internal_8_sse2(&p3, &q3, &p2, &q2, &p1, &q1, &p0, &q0, &q1q0, &p1p0, + &blimit, &limit, &thresh); + + xx_storel_32(s - 1 * p, p1p0); + xx_storel_32(s - 2 * p, _mm_srli_si128(p1p0, 4)); + xx_storel_32(s + 0 * p, q1q0); + xx_storel_32(s + 1 * p, _mm_srli_si128(q1q0, 4)); + xx_storel_32(s - 3 * p, p2); + xx_storel_32(s + 2 * p, q2); +} + +void aom_lpf_horizontal_14_dual_sse2(unsigned char *s, int p, + const unsigned char *_blimit0, + const unsigned char *_limit0, + const unsigned char *_thresh0, + const unsigned char *_blimit1, + const unsigned char *_limit1, + const unsigned char *_thresh1) { + __m128i q6p6, q5p5, q4p4, q3p3, q2p2, q1p1, q0p0; + __m128i blimit = + _mm_unpacklo_epi32(_mm_load_si128((const __m128i *)_blimit0), + _mm_load_si128((const __m128i *)_blimit1)); + __m128i limit = _mm_unpacklo_epi32(_mm_load_si128((const __m128i *)_limit0), + _mm_load_si128((const __m128i *)_limit1)); + __m128i thresh = + _mm_unpacklo_epi32(_mm_load_si128((const __m128i *)_thresh0), + _mm_load_si128((const __m128i *)_thresh1)); + + q4p4 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s - 5 * p)), + _mm_loadl_epi64((__m128i *)(s + 4 * p))); + q3p3 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s - 4 * p)), + _mm_loadl_epi64((__m128i *)(s + 3 * p))); + q2p2 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s - 3 * p)), + _mm_loadl_epi64((__m128i *)(s + 2 * p))); + q1p1 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s - 2 * p)), + _mm_loadl_epi64((__m128i *)(s + 1 * p))); + + q0p0 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s - 1 * p)), + _mm_loadl_epi64((__m128i *)(s - 0 * p))); + + q5p5 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s - 6 * p)), + _mm_loadl_epi64((__m128i *)(s + 5 * p))); + + q6p6 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s - 7 * p)), + _mm_loadl_epi64((__m128i *)(s + 6 * p))); + + lpf_internal_14_dual_sse2(&q6p6, &q5p5, &q4p4, &q3p3, &q2p2, &q1p1, &q0p0, + &blimit, &limit, &thresh); + + _mm_storel_epi64((__m128i *)(s - 1 * p), q0p0); + _mm_storel_epi64((__m128i *)(s + 0 * p), _mm_srli_si128(q0p0, 8)); + _mm_storel_epi64((__m128i *)(s - 2 * p), q1p1); + _mm_storel_epi64((__m128i *)(s + 1 * p), _mm_srli_si128(q1p1, 8)); + _mm_storel_epi64((__m128i *)(s - 3 * p), q2p2); + _mm_storel_epi64((__m128i *)(s + 2 * p), _mm_srli_si128(q2p2, 8)); + _mm_storel_epi64((__m128i *)(s - 4 * p), q3p3); + _mm_storel_epi64((__m128i *)(s + 3 * p), _mm_srli_si128(q3p3, 8)); + _mm_storel_epi64((__m128i *)(s - 5 * p), q4p4); + _mm_storel_epi64((__m128i *)(s + 4 * p), _mm_srli_si128(q4p4, 8)); + _mm_storel_epi64((__m128i *)(s - 6 * p), q5p5); + _mm_storel_epi64((__m128i *)(s + 5 * p), _mm_srli_si128(q5p5, 8)); +} + +void aom_lpf_horizontal_8_dual_sse2(uint8_t *s, int p, const uint8_t *_blimit0, + const uint8_t *_limit0, + const uint8_t *_thresh0, + const uint8_t *_blimit1, + const uint8_t *_limit1, + const uint8_t *_thresh1) { + __m128i blimit = _mm_unpacklo_epi32(_mm_load_si128((__m128i *)_blimit0), + _mm_load_si128((__m128i *)_blimit1)); + __m128i limit = _mm_unpacklo_epi32(_mm_load_si128((__m128i *)_limit0), + _mm_load_si128((__m128i *)_limit1)); + __m128i thresh = _mm_unpacklo_epi32(_mm_load_si128((__m128i *)_thresh0), + _mm_load_si128((__m128i *)_thresh1)); + + __m128i p2, p1, p0, q0, q1, q2, p3, q3; + __m128i q1q0, p1p0; + + p3 = _mm_loadl_epi64((__m128i *)(s - 4 * p)); + p2 = _mm_loadl_epi64((__m128i *)(s - 3 * p)); + p1 = _mm_loadl_epi64((__m128i *)(s - 2 * p)); + p0 = _mm_loadl_epi64((__m128i *)(s - 1 * p)); + q0 = _mm_loadl_epi64((__m128i *)(s - 0 * p)); + q1 = _mm_loadl_epi64((__m128i *)(s + 1 * p)); + q2 = _mm_loadl_epi64((__m128i *)(s + 2 * p)); + q3 = _mm_loadl_epi64((__m128i *)(s + 3 * p)); + + lpf_internal_8_dual_sse2(&p3, &q3, &p2, &q2, &p1, &q1, &p0, &q0, &q1q0, &p1p0, + &blimit, &limit, &thresh); + + _mm_storel_epi64((__m128i *)(s - 1 * p), p1p0); + _mm_storel_epi64((__m128i *)(s - 2 * p), _mm_srli_si128(p1p0, 8)); + _mm_storel_epi64((__m128i *)(s + 0 * p), q1q0); + _mm_storel_epi64((__m128i *)(s + 1 * p), _mm_srli_si128(q1q0, 8)); + _mm_storel_epi64((__m128i *)(s - 3 * p), p2); + _mm_storel_epi64((__m128i *)(s + 2 * p), q2); +} + +void aom_lpf_horizontal_4_dual_sse2(unsigned char *s, int p, + const unsigned char *_blimit0, + const unsigned char *_limit0, + const unsigned char *_thresh0, + const unsigned char *_blimit1, + const unsigned char *_limit1, + const unsigned char *_thresh1) { + __m128i p1, p0, q0, q1; + __m128i qs1qs0, ps1ps0; + + p1 = _mm_loadl_epi64((__m128i *)(s - 2 * p)); + p0 = _mm_loadl_epi64((__m128i *)(s - 1 * p)); + q0 = _mm_loadl_epi64((__m128i *)(s - 0 * p)); + q1 = _mm_loadl_epi64((__m128i *)(s + 1 * p)); + + const __m128i zero = _mm_setzero_si128(); + const __m128i blimit = + _mm_unpacklo_epi32(_mm_load_si128((const __m128i *)_blimit0), + _mm_load_si128((const __m128i *)_blimit1)); + const __m128i limit = + _mm_unpacklo_epi32(_mm_load_si128((const __m128i *)_limit0), + _mm_load_si128((const __m128i *)_limit1)); + + __m128i l = _mm_unpacklo_epi64(blimit, limit); + + __m128i thresh0 = + _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)_thresh0), zero); + + __m128i thresh1 = + _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)_thresh1), zero); + + __m128i t = _mm_unpacklo_epi64(thresh0, thresh1); + + lpf_internal_4_dual_sse2(&p1, &p0, &q0, &q1, &l, &t, &qs1qs0, &ps1ps0); + + _mm_storel_epi64((__m128i *)(s - 1 * p), ps1ps0); + _mm_storel_epi64((__m128i *)(s - 2 * p), _mm_srli_si128(ps1ps0, 8)); + _mm_storel_epi64((__m128i *)(s + 0 * p), qs1qs0); + _mm_storel_epi64((__m128i *)(s + 1 * p), _mm_srli_si128(qs1qs0, 8)); +} + +void aom_lpf_vertical_4_dual_sse2(uint8_t *s, int p, const uint8_t *_blimit0, + const uint8_t *_limit0, + const uint8_t *_thresh0, + const uint8_t *_blimit1, + const uint8_t *_limit1, + const uint8_t *_thresh1) { + __m128i p0, q0, q1, p1; + __m128i x0, x1, x2, x3, x4, x5, x6, x7; + __m128i d0, d1, d2, d3, d4, d5, d6, d7; + __m128i qs1qs0, ps1ps0; + + const __m128i zero = _mm_setzero_si128(); + const __m128i blimit = + _mm_unpacklo_epi32(_mm_load_si128((const __m128i *)_blimit0), + _mm_load_si128((const __m128i *)_blimit1)); + const __m128i limit = + _mm_unpacklo_epi32(_mm_load_si128((const __m128i *)_limit0), + _mm_load_si128((const __m128i *)_limit1)); + + __m128i l = _mm_unpacklo_epi64(blimit, limit); + + __m128i thresh0 = + _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)_thresh0), zero); + + __m128i thresh1 = + _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)_thresh1), zero); + + __m128i t = _mm_unpacklo_epi64(thresh0, thresh1); + + x0 = _mm_loadl_epi64((__m128i *)((s - 2))); + x1 = _mm_loadl_epi64((__m128i *)((s - 2) + p)); + x2 = _mm_loadl_epi64((__m128i *)((s - 2) + 2 * p)); + x3 = _mm_loadl_epi64((__m128i *)((s - 2) + 3 * p)); + x4 = _mm_loadl_epi64((__m128i *)((s - 2) + 4 * p)); + x5 = _mm_loadl_epi64((__m128i *)((s - 2) + 5 * p)); + x6 = _mm_loadl_epi64((__m128i *)((s - 2) + 6 * p)); + x7 = _mm_loadl_epi64((__m128i *)((s - 2) + 7 * p)); + + transpose8x8_low_sse2(&x0, &x1, &x2, &x3, &x4, &x5, &x6, &x7, &p1, &p0, &q0, + &q1); + + lpf_internal_4_dual_sse2(&p1, &p0, &q0, &q1, &l, &t, &qs1qs0, &ps1ps0); + + p1 = _mm_srli_si128(ps1ps0, 8); + q1 = _mm_srli_si128(qs1qs0, 8); + + transpose4x8_8x4_sse2(&p1, &ps1ps0, &qs1qs0, &q1, &d0, &d1, &d2, &d3, &d4, + &d5, &d6, &d7); + + xx_storel_32((s - 2 + 0 * p), d0); + xx_storel_32((s - 2 + 1 * p), d1); + xx_storel_32((s - 2 + 2 * p), d2); + xx_storel_32((s - 2 + 3 * p), d3); + xx_storel_32((s - 2 + 4 * p), d4); + xx_storel_32((s - 2 + 5 * p), d5); + xx_storel_32((s - 2 + 6 * p), d6); + xx_storel_32((s - 2 + 7 * p), d7); +} + +void aom_lpf_vertical_6_sse2(unsigned char *s, int p, + const unsigned char *_blimit, + const unsigned char *_limit, + const unsigned char *_thresh) { + __m128i d0, d1, d2, d3, d4, d5, d6, d7; + __m128i x2, x1, x0, x3; + __m128i p0, q0; + __m128i p1p0, q1q0; + __m128i blimit = _mm_load_si128((__m128i *)_blimit); + __m128i limit = _mm_load_si128((__m128i *)_limit); + __m128i thresh = _mm_load_si128((__m128i *)_thresh); + + x3 = _mm_loadl_epi64((__m128i *)((s - 3) + 0 * p)); + x2 = _mm_loadl_epi64((__m128i *)((s - 3) + 1 * p)); + x1 = _mm_loadl_epi64((__m128i *)((s - 3) + 2 * p)); + x0 = _mm_loadl_epi64((__m128i *)((s - 3) + 3 * p)); + + transpose4x8_8x4_sse2(&x3, &x2, &x1, &x0, &d0, &d1, &d2, &d3, &d4, &d5, &d6, + &d7); + + lpf_internal_6_sse2(&d0, &d5, &d1, &d4, &d2, &d3, &q1q0, &p1p0, &blimit, + &limit, &thresh); + + p0 = _mm_srli_si128(p1p0, 4); + q0 = _mm_srli_si128(q1q0, 4); + + transpose4x8_8x4_low_sse2(&p0, &p1p0, &q1q0, &q0, &d0, &d1, &d2, &d3); + + xx_storel_32(s + 0 * p - 2, d0); + xx_storel_32(s + 1 * p - 2, d1); + xx_storel_32(s + 2 * p - 2, d2); + xx_storel_32(s + 3 * p - 2, d3); +} + +void aom_lpf_vertical_6_dual_sse2(uint8_t *s, int p, const uint8_t *_blimit0, + const uint8_t *_limit0, + const uint8_t *_thresh0, + const uint8_t *_blimit1, + const uint8_t *_limit1, + const uint8_t *_thresh1) { + __m128i blimit = _mm_unpacklo_epi32(_mm_load_si128((__m128i *)_blimit0), + _mm_load_si128((__m128i *)_blimit1)); + __m128i limit = _mm_unpacklo_epi32(_mm_load_si128((__m128i *)_limit0), + _mm_load_si128((__m128i *)_limit1)); + __m128i thresh = _mm_unpacklo_epi32(_mm_load_si128((__m128i *)_thresh0), + _mm_load_si128((__m128i *)_thresh1)); + + __m128i d0, d1, d2, d3, d4, d5, d6, d7; + __m128i x0, x1, x2, x3, x4, x5, x6, x7; + __m128i p0, q0; + __m128i p1p0, q1q0; + __m128i d0d1, d2d3, d4d5, d6d7; + + x0 = _mm_loadl_epi64((__m128i *)((s - 3) + 0 * p)); + x1 = _mm_loadl_epi64((__m128i *)((s - 3) + 1 * p)); + x2 = _mm_loadl_epi64((__m128i *)((s - 3) + 2 * p)); + x3 = _mm_loadl_epi64((__m128i *)((s - 3) + 3 * p)); + x4 = _mm_loadl_epi64((__m128i *)((s - 3) + 4 * p)); + x5 = _mm_loadl_epi64((__m128i *)((s - 3) + 5 * p)); + x6 = _mm_loadl_epi64((__m128i *)((s - 3) + 6 * p)); + x7 = _mm_loadl_epi64((__m128i *)((s - 3) + 7 * p)); + + transpose8x8_sse2(&x0, &x1, &x2, &x3, &x4, &x5, &x6, &x7, &d0d1, &d2d3, &d4d5, + &d6d7); + + d1 = _mm_srli_si128(d0d1, 8); + d3 = _mm_srli_si128(d2d3, 8); + d5 = _mm_srli_si128(d4d5, 8); + d7 = _mm_srli_si128(d6d7, 8); + + lpf_internal_6_dual_sse2(&d0d1, &d5, &d1, &d4d5, &d2d3, &d3, &q1q0, &p1p0, + &blimit, &limit, &thresh); + + p0 = _mm_srli_si128(p1p0, 8); + q0 = _mm_srli_si128(q1q0, 8); + + transpose4x8_8x4_sse2(&p0, &p1p0, &q1q0, &q0, &d0, &d1, &d2, &d3, &d4, &d5, + &d6, &d7); + + xx_storel_32((s - 2 + 0 * p), d0); + xx_storel_32((s - 2 + 1 * p), d1); + xx_storel_32((s - 2 + 2 * p), d2); + xx_storel_32((s - 2 + 3 * p), d3); + xx_storel_32((s - 2 + 4 * p), d4); + xx_storel_32((s - 2 + 5 * p), d5); + xx_storel_32((s - 2 + 6 * p), d6); + xx_storel_32((s - 2 + 7 * p), d7); +} + +void aom_lpf_vertical_8_sse2(unsigned char *s, int p, + const unsigned char *_blimit, + const unsigned char *_limit, + const unsigned char *_thresh) { + __m128i d0, d1, d2, d3, d4, d5, d6, d7; + + __m128i p0, q0; + __m128i x2, x1, x0, x3; + __m128i q1q0, p1p0; + __m128i blimit = _mm_load_si128((const __m128i *)_blimit); + __m128i limit = _mm_load_si128((const __m128i *)_limit); + __m128i thresh = _mm_load_si128((const __m128i *)_thresh); + + x3 = _mm_loadl_epi64((__m128i *)((s - 4) + 0 * p)); + x2 = _mm_loadl_epi64((__m128i *)((s - 4) + 1 * p)); + x1 = _mm_loadl_epi64((__m128i *)((s - 4) + 2 * p)); + x0 = _mm_loadl_epi64((__m128i *)((s - 4) + 3 * p)); + + transpose4x8_8x4_sse2(&x3, &x2, &x1, &x0, &d0, &d1, &d2, &d3, &d4, &d5, &d6, + &d7); + // Loop filtering + lpf_internal_8_sse2(&d0, &d7, &d1, &d6, &d2, &d5, &d3, &d4, &q1q0, &p1p0, + &blimit, &limit, &thresh); + + p0 = _mm_srli_si128(p1p0, 4); + q0 = _mm_srli_si128(q1q0, 4); + + transpose8x8_low_sse2(&d0, &d1, &p0, &p1p0, &q1q0, &q0, &d6, &d7, &d0, &d1, + &d2, &d3); + + _mm_storel_epi64((__m128i *)(s - 4 + 0 * p), d0); + _mm_storel_epi64((__m128i *)(s - 4 + 1 * p), d1); + _mm_storel_epi64((__m128i *)(s - 4 + 2 * p), d2); + _mm_storel_epi64((__m128i *)(s - 4 + 3 * p), d3); +} + +void aom_lpf_vertical_8_dual_sse2(uint8_t *s, int p, const uint8_t *_blimit0, + const uint8_t *_limit0, + const uint8_t *_thresh0, + const uint8_t *_blimit1, + const uint8_t *_limit1, + const uint8_t *_thresh1) { + __m128i blimit = _mm_unpacklo_epi32(_mm_load_si128((__m128i *)_blimit0), + _mm_load_si128((__m128i *)_blimit1)); + __m128i limit = _mm_unpacklo_epi32(_mm_load_si128((__m128i *)_limit0), + _mm_load_si128((__m128i *)_limit1)); + __m128i thresh = _mm_unpacklo_epi32(_mm_load_si128((__m128i *)_thresh0), + _mm_load_si128((__m128i *)_thresh1)); + + __m128i x0, x1, x2, x3, x4, x5, x6, x7; + __m128i d1, d3, d5, d7; + __m128i q1q0, p1p0; + __m128i p1, q1; + __m128i d0d1, d2d3, d4d5, d6d7; + + x0 = _mm_loadl_epi64((__m128i *)(s - 4 + 0 * p)); + x1 = _mm_loadl_epi64((__m128i *)(s - 4 + 1 * p)); + x2 = _mm_loadl_epi64((__m128i *)(s - 4 + 2 * p)); + x3 = _mm_loadl_epi64((__m128i *)(s - 4 + 3 * p)); + x4 = _mm_loadl_epi64((__m128i *)(s - 4 + 4 * p)); + x5 = _mm_loadl_epi64((__m128i *)(s - 4 + 5 * p)); + x6 = _mm_loadl_epi64((__m128i *)(s - 4 + 6 * p)); + x7 = _mm_loadl_epi64((__m128i *)(s - 4 + 7 * p)); + + transpose8x8_sse2(&x0, &x1, &x2, &x3, &x4, &x5, &x6, &x7, &d0d1, &d2d3, &d4d5, + &d6d7); + + d1 = _mm_srli_si128(d0d1, 8); + d3 = _mm_srli_si128(d2d3, 8); + d5 = _mm_srli_si128(d4d5, 8); + d7 = _mm_srli_si128(d6d7, 8); + + lpf_internal_8_dual_sse2(&d0d1, &d7, &d1, &d6d7, &d2d3, &d5, &d3, &d4d5, + &q1q0, &p1p0, &blimit, &limit, &thresh); + + p1 = _mm_srli_si128(p1p0, 8); + q1 = _mm_srli_si128(q1q0, 8); + + transpose8x8_sse2(&d0d1, &d1, &p1, &p1p0, &q1q0, &q1, &d6d7, &d7, &d0d1, + &d2d3, &d4d5, &d6d7); + + _mm_storel_epi64((__m128i *)(s - 4 + 0 * p), d0d1); + _mm_storel_epi64((__m128i *)(s - 4 + 1 * p), _mm_srli_si128(d0d1, 8)); + _mm_storel_epi64((__m128i *)(s - 4 + 2 * p), d2d3); + _mm_storel_epi64((__m128i *)(s - 4 + 3 * p), _mm_srli_si128(d2d3, 8)); + _mm_storel_epi64((__m128i *)(s - 4 + 4 * p), d4d5); + _mm_storel_epi64((__m128i *)(s - 4 + 5 * p), _mm_srli_si128(d4d5, 8)); + _mm_storel_epi64((__m128i *)(s - 4 + 6 * p), d6d7); + _mm_storel_epi64((__m128i *)(s - 4 + 7 * p), _mm_srli_si128(d6d7, 8)); +} + +void aom_lpf_vertical_14_sse2(unsigned char *s, int p, + const unsigned char *_blimit, + const unsigned char *_limit, + const unsigned char *_thresh) { + __m128i q7p7, q6p6, q5p5, q4p4, q3p3, q2p2, q1p1, q0p0; + __m128i x6, x5, x4, x3; + __m128i pq0, pq1, pq2, pq3; + __m128i blimit = _mm_load_si128((__m128i *)_blimit); + __m128i limit = _mm_load_si128((__m128i *)_limit); + __m128i thresh = _mm_load_si128((__m128i *)_thresh); + + x6 = _mm_loadu_si128((__m128i *)((s - 8) + 0 * p)); + x5 = _mm_loadu_si128((__m128i *)((s - 8) + 1 * p)); + x4 = _mm_loadu_si128((__m128i *)((s - 8) + 2 * p)); + x3 = _mm_loadu_si128((__m128i *)((s - 8) + 3 * p)); + + transpose_pq_14_sse2(&x6, &x5, &x4, &x3, &q0p0, &q1p1, &q2p2, &q3p3, &q4p4, + &q5p5, &q6p6, &q7p7); + + lpf_internal_14_sse2(&q6p6, &q5p5, &q4p4, &q3p3, &q2p2, &q1p1, &q0p0, &blimit, + &limit, &thresh); + + transpose_pq_14_inv_sse2(&q7p7, &q6p6, &q5p5, &q4p4, &q3p3, &q2p2, &q1p1, + &q0p0, &pq0, &pq1, &pq2, &pq3); + _mm_storeu_si128((__m128i *)(s - 8 + 0 * p), pq0); + _mm_storeu_si128((__m128i *)(s - 8 + 1 * p), pq1); + _mm_storeu_si128((__m128i *)(s - 8 + 2 * p), pq2); + _mm_storeu_si128((__m128i *)(s - 8 + 3 * p), pq3); +} + +void aom_lpf_vertical_14_dual_sse2( + unsigned char *s, int p, const uint8_t *_blimit0, const uint8_t *_limit0, + const uint8_t *_thresh0, const uint8_t *_blimit1, const uint8_t *_limit1, + const uint8_t *_thresh1) { + __m128i q6p6, q5p5, q4p4, q3p3, q2p2, q1p1, q0p0; + __m128i x7, x6, x5, x4, x3, x2, x1, x0; + __m128i d0d1, d2d3, d4d5, d6d7, d8d9, d10d11, d12d13, d14d15; + __m128i q0, q1, q2, q3, q7; + __m128i p0p1, p2p3, p4p5, p6p7; + + __m128i blimit = + _mm_unpacklo_epi32(_mm_load_si128((const __m128i *)_blimit0), + _mm_load_si128((const __m128i *)_blimit1)); + __m128i limit = _mm_unpacklo_epi32(_mm_load_si128((const __m128i *)_limit0), + _mm_load_si128((const __m128i *)_limit1)); + __m128i thresh = + _mm_unpacklo_epi32(_mm_load_si128((const __m128i *)_thresh0), + _mm_load_si128((const __m128i *)_thresh1)); + + x7 = _mm_loadu_si128((__m128i *)((s - 8) + 0 * p)); + x6 = _mm_loadu_si128((__m128i *)((s - 8) + 1 * p)); + x5 = _mm_loadu_si128((__m128i *)((s - 8) + 2 * p)); + x4 = _mm_loadu_si128((__m128i *)((s - 8) + 3 * p)); + x3 = _mm_loadu_si128((__m128i *)((s - 8) + 4 * p)); + x2 = _mm_loadu_si128((__m128i *)((s - 8) + 5 * p)); + x1 = _mm_loadu_si128((__m128i *)((s - 8) + 6 * p)); + x0 = _mm_loadu_si128((__m128i *)((s - 8) + 7 * p)); + + transpose8x16_16x8_sse2(&x7, &x6, &x5, &x4, &x3, &x2, &x1, &x0, &d0d1, &d2d3, + &d4d5, &d6d7, &d8d9, &d10d11, &d12d13, &d14d15); + + q6p6 = _mm_unpacklo_epi64(d2d3, _mm_srli_si128(d12d13, 8)); + q5p5 = _mm_unpacklo_epi64(d4d5, _mm_srli_si128(d10d11, 8)); + q4p4 = _mm_unpacklo_epi64(d6d7, _mm_srli_si128(d8d9, 8)); + q3p3 = _mm_unpacklo_epi64(d8d9, _mm_srli_si128(d6d7, 8)); + q2p2 = _mm_unpacklo_epi64(d10d11, _mm_srli_si128(d4d5, 8)); + q1p1 = _mm_unpacklo_epi64(d12d13, _mm_srli_si128(d2d3, 8)); + q0p0 = _mm_unpacklo_epi64(d14d15, _mm_srli_si128(d0d1, 8)); + q7 = _mm_srli_si128(d14d15, 8); + + lpf_internal_14_dual_sse2(&q6p6, &q5p5, &q4p4, &q3p3, &q2p2, &q1p1, &q0p0, + &blimit, &limit, &thresh); + + x0 = _mm_srli_si128(q0p0, 8); + x1 = _mm_srli_si128(q1p1, 8); + x2 = _mm_srli_si128(q2p2, 8); + x3 = _mm_srli_si128(q3p3, 8); + x4 = _mm_srli_si128(q4p4, 8); + x5 = _mm_srli_si128(q5p5, 8); + x6 = _mm_srli_si128(q6p6, 8); + + transpose16x8_8x16_sse2(&d0d1, &q6p6, &q5p5, &q4p4, &q3p3, &q2p2, &q1p1, + &q0p0, &x0, &x1, &x2, &x3, &x4, &x5, &x6, &q7, &p0p1, + &p2p3, &p4p5, &p6p7, &q0, &q1, &q2, &q3); + + _mm_storeu_si128((__m128i *)(s - 8 + 0 * p), p0p1); + _mm_storeu_si128((__m128i *)(s - 8 + 1 * p), p2p3); + _mm_storeu_si128((__m128i *)(s - 8 + 2 * p), p4p5); + _mm_storeu_si128((__m128i *)(s - 8 + 3 * p), p6p7); + _mm_storeu_si128((__m128i *)(s - 8 + 4 * p), q0); + _mm_storeu_si128((__m128i *)(s - 8 + 5 * p), q1); + _mm_storeu_si128((__m128i *)(s - 8 + 6 * p), q2); + _mm_storeu_si128((__m128i *)(s - 8 + 7 * p), q3); +} diff --git a/third_party/aom/aom_dsp/x86/lpf_common_sse2.h b/third_party/aom/aom_dsp/x86/lpf_common_sse2.h new file mode 100644 index 0000000000..8970fe7dd6 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/lpf_common_sse2.h @@ -0,0 +1,215 @@ +/* + * Copyright (c) 2017, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_X86_LPF_COMMON_SSE2_H_ +#define AOM_AOM_DSP_X86_LPF_COMMON_SSE2_H_ + +#include // SSE2 + +#include "config/aom_config.h" + +static INLINE void highbd_transpose6x6_sse2(__m128i *x0, __m128i *x1, + __m128i *x2, __m128i *x3, + __m128i *x4, __m128i *x5, + __m128i *d0, __m128i *d1, + __m128i *d2, __m128i *d3, + __m128i *d4, __m128i *d5) { + __m128i w0, w1, w2, w3, w4, w5, ww0; + + // 00 01 02 03 04 05 xx xx + // 10 11 12 13 14 15 xx xx + // 20 21 22 23 24 25 xx xx + // 30 31 32 33 34 35 xx xx + // 40 41 42 43 44 45 xx xx + // 50 51 52 53 54 55 xx xx + + w0 = _mm_unpacklo_epi16(*x0, *x1); // 00 10 01 11 02 12 03 13 + w1 = _mm_unpacklo_epi16(*x2, *x3); // 20 30 21 31 22 32 23 33 + w2 = _mm_unpacklo_epi16(*x4, *x5); // 40 50 41 51 42 52 43 53 + + ww0 = _mm_unpacklo_epi32(w0, w1); // 00 10 20 30 01 11 21 31 + *d0 = _mm_unpacklo_epi64(ww0, w2); // 00 10 20 30 40 50 41 51 + *d1 = _mm_unpackhi_epi64(ww0, + _mm_srli_si128(w2, 4)); // 01 11 21 31 41 51 xx xx + + ww0 = _mm_unpackhi_epi32(w0, w1); // 02 12 22 32 03 13 23 33 + *d2 = _mm_unpacklo_epi64(ww0, + _mm_srli_si128(w2, 8)); // 02 12 22 32 42 52 xx xx + + w3 = _mm_unpackhi_epi16(*x0, *x1); // 04 14 05 15 xx xx xx xx + w4 = _mm_unpackhi_epi16(*x2, *x3); // 24 34 25 35 xx xx xx xx + w5 = _mm_unpackhi_epi16(*x4, *x5); // 44 54 45 55 xx xx xx xx + + *d3 = _mm_unpackhi_epi64(ww0, _mm_srli_si128(w2, 4)); // 03 13 23 33 43 53 + + ww0 = _mm_unpacklo_epi32(w3, w4); // 04 14 24 34 05 15 25 35 + *d4 = _mm_unpacklo_epi64(ww0, w5); // 04 14 24 34 44 54 45 55 + *d5 = _mm_unpackhi_epi64(ww0, + _mm_slli_si128(w5, 4)); // 05 15 25 35 45 55 xx xx +} + +static INLINE void highbd_transpose4x8_8x4_low_sse2(__m128i *x0, __m128i *x1, + __m128i *x2, __m128i *x3, + __m128i *d0, __m128i *d1, + __m128i *d2, __m128i *d3) { + __m128i zero = _mm_setzero_si128(); + __m128i w0, w1, ww0, ww1; + + w0 = _mm_unpacklo_epi16(*x0, *x1); // 00 10 01 11 02 12 03 13 + w1 = _mm_unpacklo_epi16(*x2, *x3); // 20 30 21 31 22 32 23 33 + + ww0 = _mm_unpacklo_epi32(w0, w1); // 00 10 20 30 01 11 21 31 + ww1 = _mm_unpackhi_epi32(w0, w1); // 02 12 22 32 03 13 23 33 + + *d0 = _mm_unpacklo_epi64(ww0, zero); // 00 10 20 30 xx xx xx xx + *d1 = _mm_unpackhi_epi64(ww0, zero); // 01 11 21 31 xx xx xx xx + *d2 = _mm_unpacklo_epi64(ww1, zero); // 02 12 22 32 xx xx xx xx + *d3 = _mm_unpackhi_epi64(ww1, zero); // 03 13 23 33 xx xx xx xx +} + +static INLINE void highbd_transpose4x8_8x4_high_sse2(__m128i *x0, __m128i *x1, + __m128i *x2, __m128i *x3, + __m128i *d4, __m128i *d5, + __m128i *d6, __m128i *d7) { + __m128i w0, w1, ww2, ww3; + __m128i zero = _mm_setzero_si128(); + + w0 = _mm_unpackhi_epi16(*x0, *x1); // 04 14 05 15 06 16 07 17 + w1 = _mm_unpackhi_epi16(*x2, *x3); // 24 34 25 35 26 36 27 37 + + ww2 = _mm_unpacklo_epi32(w0, w1); // 04 14 24 34 05 15 25 35 + ww3 = _mm_unpackhi_epi32(w0, w1); // 06 16 26 36 07 17 27 37 + + *d4 = _mm_unpacklo_epi64(ww2, zero); // 04 14 24 34 xx xx xx xx + *d5 = _mm_unpackhi_epi64(ww2, zero); // 05 15 25 35 xx xx xx xx + *d6 = _mm_unpacklo_epi64(ww3, zero); // 06 16 26 36 xx xx xx xx + *d7 = _mm_unpackhi_epi64(ww3, zero); // 07 17 27 37 xx xx xx xx +} + +// here in and out pointers (x and d) should be different! we don't store their +// values inside +static INLINE void highbd_transpose4x8_8x4_sse2(__m128i *x0, __m128i *x1, + __m128i *x2, __m128i *x3, + __m128i *d0, __m128i *d1, + __m128i *d2, __m128i *d3, + __m128i *d4, __m128i *d5, + __m128i *d6, __m128i *d7) { + // input + // x0 00 01 02 03 04 05 06 07 + // x1 10 11 12 13 14 15 16 17 + // x2 20 21 22 23 24 25 26 27 + // x3 30 31 32 33 34 35 36 37 + // output + // 00 10 20 30 xx xx xx xx + // 01 11 21 31 xx xx xx xx + // 02 12 22 32 xx xx xx xx + // 03 13 23 33 xx xx xx xx + // 04 14 24 34 xx xx xx xx + // 05 15 25 35 xx xx xx xx + // 06 16 26 36 xx xx xx xx + // 07 17 27 37 xx xx xx xx + highbd_transpose4x8_8x4_low_sse2(x0, x1, x2, x3, d0, d1, d2, d3); + highbd_transpose4x8_8x4_high_sse2(x0, x1, x2, x3, d4, d5, d6, d7); +} + +static INLINE void highbd_transpose8x8_low_sse2(__m128i *x0, __m128i *x1, + __m128i *x2, __m128i *x3, + __m128i *x4, __m128i *x5, + __m128i *x6, __m128i *x7, + __m128i *d0, __m128i *d1, + __m128i *d2, __m128i *d3) { + __m128i w0, w1, w2, w3, ww0, ww1; + // x0 00 01 02 03 04 05 06 07 + // x1 10 11 12 13 14 15 16 17 + // x2 20 21 22 23 24 25 26 27 + // x3 30 31 32 33 34 35 36 37 + // x4 40 41 42 43 44 45 46 47 + // x5 50 51 52 53 54 55 56 57 + // x6 60 61 62 63 64 65 66 67 + // x7 70 71 72 73 74 75 76 77 + + w0 = _mm_unpacklo_epi16(*x0, *x1); // 00 10 01 11 02 12 03 13 + w1 = _mm_unpacklo_epi16(*x2, *x3); // 20 30 21 31 22 32 23 33 + w2 = _mm_unpacklo_epi16(*x4, *x5); // 40 50 41 51 42 52 43 53 + w3 = _mm_unpacklo_epi16(*x6, *x7); // 60 70 61 71 62 72 63 73 + + ww0 = _mm_unpacklo_epi32(w0, w1); // 00 10 20 30 01 11 21 31 + ww1 = _mm_unpacklo_epi32(w2, w3); // 40 50 60 70 41 51 61 71 + + *d0 = _mm_unpacklo_epi64(ww0, ww1); // 00 10 20 30 40 50 60 70 + *d1 = _mm_unpackhi_epi64(ww0, ww1); // 01 11 21 31 41 51 61 71 + + ww0 = _mm_unpackhi_epi32(w0, w1); // 02 12 22 32 03 13 23 33 + ww1 = _mm_unpackhi_epi32(w2, w3); // 42 52 62 72 43 53 63 73 + + *d2 = _mm_unpacklo_epi64(ww0, ww1); // 02 12 22 32 42 52 62 72 + *d3 = _mm_unpackhi_epi64(ww0, ww1); // 03 13 23 33 43 53 63 73 +} + +static INLINE void highbd_transpose8x8_high_sse2(__m128i *x0, __m128i *x1, + __m128i *x2, __m128i *x3, + __m128i *x4, __m128i *x5, + __m128i *x6, __m128i *x7, + __m128i *d4, __m128i *d5, + __m128i *d6, __m128i *d7) { + __m128i w0, w1, w2, w3, ww0, ww1; + // x0 00 01 02 03 04 05 06 07 + // x1 10 11 12 13 14 15 16 17 + // x2 20 21 22 23 24 25 26 27 + // x3 30 31 32 33 34 35 36 37 + // x4 40 41 42 43 44 45 46 47 + // x5 50 51 52 53 54 55 56 57 + // x6 60 61 62 63 64 65 66 67 + // x7 70 71 72 73 74 75 76 77 + w0 = _mm_unpackhi_epi16(*x0, *x1); // 04 14 05 15 06 16 07 17 + w1 = _mm_unpackhi_epi16(*x2, *x3); // 24 34 25 35 26 36 27 37 + w2 = _mm_unpackhi_epi16(*x4, *x5); // 44 54 45 55 46 56 47 57 + w3 = _mm_unpackhi_epi16(*x6, *x7); // 64 74 65 75 66 76 67 77 + + ww0 = _mm_unpacklo_epi32(w0, w1); // 04 14 24 34 05 15 25 35 + ww1 = _mm_unpacklo_epi32(w2, w3); // 44 54 64 74 45 55 65 75 + + *d4 = _mm_unpacklo_epi64(ww0, ww1); // 04 14 24 34 44 54 64 74 + *d5 = _mm_unpackhi_epi64(ww0, ww1); // 05 15 25 35 45 55 65 75 + + ww0 = _mm_unpackhi_epi32(w0, w1); // 06 16 26 36 07 17 27 37 + ww1 = _mm_unpackhi_epi32(w2, w3); // 46 56 66 76 47 57 67 77 + + *d6 = _mm_unpacklo_epi64(ww0, ww1); // 06 16 26 36 46 56 66 76 + *d7 = _mm_unpackhi_epi64(ww0, ww1); // 07 17 27 37 47 57 67 77 +} + +// here in and out pointers (x and d) should be different! we don't store their +// values inside +static INLINE void highbd_transpose8x8_sse2( + __m128i *x0, __m128i *x1, __m128i *x2, __m128i *x3, __m128i *x4, + __m128i *x5, __m128i *x6, __m128i *x7, __m128i *d0, __m128i *d1, + __m128i *d2, __m128i *d3, __m128i *d4, __m128i *d5, __m128i *d6, + __m128i *d7) { + highbd_transpose8x8_low_sse2(x0, x1, x2, x3, x4, x5, x6, x7, d0, d1, d2, d3); + highbd_transpose8x8_high_sse2(x0, x1, x2, x3, x4, x5, x6, x7, d4, d5, d6, d7); +} + +// here in and out pointers (x and d arrays) should be different! we don't store +// their values inside +static INLINE void highbd_transpose8x16_sse2( + __m128i *x0, __m128i *x1, __m128i *x2, __m128i *x3, __m128i *x4, + __m128i *x5, __m128i *x6, __m128i *x7, __m128i *d0, __m128i *d1, + __m128i *d2, __m128i *d3, __m128i *d4, __m128i *d5, __m128i *d6, + __m128i *d7) { + highbd_transpose8x8_sse2(x0, x1, x2, x3, x4, x5, x6, x7, d0, d1, d2, d3, d4, + d5, d6, d7); + highbd_transpose8x8_sse2(x0 + 1, x1 + 1, x2 + 1, x3 + 1, x4 + 1, x5 + 1, + x6 + 1, x7 + 1, d0 + 1, d1 + 1, d2 + 1, d3 + 1, + d4 + 1, d5 + 1, d6 + 1, d7 + 1); +} + +#endif // AOM_AOM_DSP_X86_LPF_COMMON_SSE2_H_ diff --git a/third_party/aom/aom_dsp/x86/masked_sad_intrin_avx2.c b/third_party/aom/aom_dsp/x86/masked_sad_intrin_avx2.c new file mode 100644 index 0000000000..584b5e7e37 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/masked_sad_intrin_avx2.c @@ -0,0 +1,389 @@ +/* + * Copyright (c) 2018, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include + +#include "config/aom_config.h" +#include "config/aom_dsp_rtcd.h" + +#include "aom_dsp/blend.h" +#include "aom/aom_integer.h" +#include "aom_dsp/x86/synonyms.h" +#include "aom_dsp/x86//masked_sad_intrin_ssse3.h" + +static INLINE unsigned int masked_sad32xh_avx2( + const uint8_t *src_ptr, int src_stride, const uint8_t *a_ptr, int a_stride, + const uint8_t *b_ptr, int b_stride, const uint8_t *m_ptr, int m_stride, + int width, int height) { + int x, y; + __m256i res = _mm256_setzero_si256(); + const __m256i mask_max = _mm256_set1_epi8((1 << AOM_BLEND_A64_ROUND_BITS)); + const __m256i round_scale = + _mm256_set1_epi16(1 << (15 - AOM_BLEND_A64_ROUND_BITS)); + for (y = 0; y < height; y++) { + for (x = 0; x < width; x += 32) { + const __m256i src = _mm256_lddqu_si256((const __m256i *)&src_ptr[x]); + const __m256i a = _mm256_lddqu_si256((const __m256i *)&a_ptr[x]); + const __m256i b = _mm256_lddqu_si256((const __m256i *)&b_ptr[x]); + const __m256i m = _mm256_lddqu_si256((const __m256i *)&m_ptr[x]); + const __m256i m_inv = _mm256_sub_epi8(mask_max, m); + + // Calculate 16 predicted pixels. + // Note that the maximum value of any entry of 'pred_l' or 'pred_r' + // is 64 * 255, so we have plenty of space to add rounding constants. + const __m256i data_l = _mm256_unpacklo_epi8(a, b); + const __m256i mask_l = _mm256_unpacklo_epi8(m, m_inv); + __m256i pred_l = _mm256_maddubs_epi16(data_l, mask_l); + pred_l = _mm256_mulhrs_epi16(pred_l, round_scale); + + const __m256i data_r = _mm256_unpackhi_epi8(a, b); + const __m256i mask_r = _mm256_unpackhi_epi8(m, m_inv); + __m256i pred_r = _mm256_maddubs_epi16(data_r, mask_r); + pred_r = _mm256_mulhrs_epi16(pred_r, round_scale); + + const __m256i pred = _mm256_packus_epi16(pred_l, pred_r); + res = _mm256_add_epi32(res, _mm256_sad_epu8(pred, src)); + } + + src_ptr += src_stride; + a_ptr += a_stride; + b_ptr += b_stride; + m_ptr += m_stride; + } + // At this point, we have two 32-bit partial SADs in lanes 0 and 2 of 'res'. + res = _mm256_shuffle_epi32(res, 0xd8); + res = _mm256_permute4x64_epi64(res, 0xd8); + res = _mm256_hadd_epi32(res, res); + res = _mm256_hadd_epi32(res, res); + int32_t sad = _mm256_extract_epi32(res, 0); + return (sad + 31) >> 6; +} + +static INLINE __m256i xx_loadu2_m128i(const void *hi, const void *lo) { + __m128i a0 = _mm_lddqu_si128((const __m128i *)(lo)); + __m128i a1 = _mm_lddqu_si128((const __m128i *)(hi)); + __m256i a = _mm256_castsi128_si256(a0); + return _mm256_inserti128_si256(a, a1, 1); +} + +static INLINE unsigned int masked_sad16xh_avx2( + const uint8_t *src_ptr, int src_stride, const uint8_t *a_ptr, int a_stride, + const uint8_t *b_ptr, int b_stride, const uint8_t *m_ptr, int m_stride, + int height) { + int y; + __m256i res = _mm256_setzero_si256(); + const __m256i mask_max = _mm256_set1_epi8((1 << AOM_BLEND_A64_ROUND_BITS)); + const __m256i round_scale = + _mm256_set1_epi16(1 << (15 - AOM_BLEND_A64_ROUND_BITS)); + for (y = 0; y < height; y += 2) { + const __m256i src = xx_loadu2_m128i(src_ptr + src_stride, src_ptr); + const __m256i a = xx_loadu2_m128i(a_ptr + a_stride, a_ptr); + const __m256i b = xx_loadu2_m128i(b_ptr + b_stride, b_ptr); + const __m256i m = xx_loadu2_m128i(m_ptr + m_stride, m_ptr); + const __m256i m_inv = _mm256_sub_epi8(mask_max, m); + + // Calculate 16 predicted pixels. + // Note that the maximum value of any entry of 'pred_l' or 'pred_r' + // is 64 * 255, so we have plenty of space to add rounding constants. + const __m256i data_l = _mm256_unpacklo_epi8(a, b); + const __m256i mask_l = _mm256_unpacklo_epi8(m, m_inv); + __m256i pred_l = _mm256_maddubs_epi16(data_l, mask_l); + pred_l = _mm256_mulhrs_epi16(pred_l, round_scale); + + const __m256i data_r = _mm256_unpackhi_epi8(a, b); + const __m256i mask_r = _mm256_unpackhi_epi8(m, m_inv); + __m256i pred_r = _mm256_maddubs_epi16(data_r, mask_r); + pred_r = _mm256_mulhrs_epi16(pred_r, round_scale); + + const __m256i pred = _mm256_packus_epi16(pred_l, pred_r); + res = _mm256_add_epi32(res, _mm256_sad_epu8(pred, src)); + + src_ptr += src_stride << 1; + a_ptr += a_stride << 1; + b_ptr += b_stride << 1; + m_ptr += m_stride << 1; + } + // At this point, we have two 32-bit partial SADs in lanes 0 and 2 of 'res'. + res = _mm256_shuffle_epi32(res, 0xd8); + res = _mm256_permute4x64_epi64(res, 0xd8); + res = _mm256_hadd_epi32(res, res); + res = _mm256_hadd_epi32(res, res); + int32_t sad = _mm256_extract_epi32(res, 0); + return (sad + 31) >> 6; +} + +static INLINE unsigned int aom_masked_sad_avx2( + const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, + const uint8_t *second_pred, const uint8_t *msk, int msk_stride, + int invert_mask, int m, int n) { + unsigned int sad; + if (!invert_mask) { + switch (m) { + case 4: + sad = aom_masked_sad4xh_ssse3(src, src_stride, ref, ref_stride, + second_pred, m, msk, msk_stride, n); + break; + case 8: + sad = aom_masked_sad8xh_ssse3(src, src_stride, ref, ref_stride, + second_pred, m, msk, msk_stride, n); + break; + case 16: + sad = masked_sad16xh_avx2(src, src_stride, ref, ref_stride, second_pred, + m, msk, msk_stride, n); + break; + default: + sad = masked_sad32xh_avx2(src, src_stride, ref, ref_stride, second_pred, + m, msk, msk_stride, m, n); + break; + } + } else { + switch (m) { + case 4: + sad = aom_masked_sad4xh_ssse3(src, src_stride, second_pred, m, ref, + ref_stride, msk, msk_stride, n); + break; + case 8: + sad = aom_masked_sad8xh_ssse3(src, src_stride, second_pred, m, ref, + ref_stride, msk, msk_stride, n); + break; + case 16: + sad = masked_sad16xh_avx2(src, src_stride, second_pred, m, ref, + ref_stride, msk, msk_stride, n); + break; + default: + sad = masked_sad32xh_avx2(src, src_stride, second_pred, m, ref, + ref_stride, msk, msk_stride, m, n); + break; + } + } + return sad; +} + +#define MASKSADMXN_AVX2(m, n) \ + unsigned int aom_masked_sad##m##x##n##_avx2( \ + const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, \ + const uint8_t *second_pred, const uint8_t *msk, int msk_stride, \ + int invert_mask) { \ + return aom_masked_sad_avx2(src, src_stride, ref, ref_stride, second_pred, \ + msk, msk_stride, invert_mask, m, n); \ + } + +MASKSADMXN_AVX2(4, 4) +MASKSADMXN_AVX2(4, 8) +MASKSADMXN_AVX2(8, 4) +MASKSADMXN_AVX2(8, 8) +MASKSADMXN_AVX2(8, 16) +MASKSADMXN_AVX2(16, 8) +MASKSADMXN_AVX2(16, 16) +MASKSADMXN_AVX2(16, 32) +MASKSADMXN_AVX2(32, 16) +MASKSADMXN_AVX2(32, 32) +MASKSADMXN_AVX2(32, 64) +MASKSADMXN_AVX2(64, 32) +MASKSADMXN_AVX2(64, 64) +MASKSADMXN_AVX2(64, 128) +MASKSADMXN_AVX2(128, 64) +MASKSADMXN_AVX2(128, 128) +MASKSADMXN_AVX2(4, 16) +MASKSADMXN_AVX2(16, 4) +MASKSADMXN_AVX2(8, 32) +MASKSADMXN_AVX2(32, 8) +MASKSADMXN_AVX2(16, 64) +MASKSADMXN_AVX2(64, 16) + +static INLINE unsigned int highbd_masked_sad8xh_avx2( + const uint8_t *src8, int src_stride, const uint8_t *a8, int a_stride, + const uint8_t *b8, int b_stride, const uint8_t *m_ptr, int m_stride, + int height) { + const uint16_t *src_ptr = CONVERT_TO_SHORTPTR(src8); + const uint16_t *a_ptr = CONVERT_TO_SHORTPTR(a8); + const uint16_t *b_ptr = CONVERT_TO_SHORTPTR(b8); + int y; + __m256i res = _mm256_setzero_si256(); + const __m256i mask_max = _mm256_set1_epi16((1 << AOM_BLEND_A64_ROUND_BITS)); + const __m256i round_const = + _mm256_set1_epi32((1 << AOM_BLEND_A64_ROUND_BITS) >> 1); + const __m256i one = _mm256_set1_epi16(1); + + for (y = 0; y < height; y += 2) { + const __m256i src = xx_loadu2_m128i(src_ptr + src_stride, src_ptr); + const __m256i a = xx_loadu2_m128i(a_ptr + a_stride, a_ptr); + const __m256i b = xx_loadu2_m128i(b_ptr + b_stride, b_ptr); + // Zero-extend mask to 16 bits + const __m256i m = _mm256_cvtepu8_epi16(_mm_unpacklo_epi64( + _mm_loadl_epi64((const __m128i *)(m_ptr)), + _mm_loadl_epi64((const __m128i *)(m_ptr + m_stride)))); + const __m256i m_inv = _mm256_sub_epi16(mask_max, m); + + const __m256i data_l = _mm256_unpacklo_epi16(a, b); + const __m256i mask_l = _mm256_unpacklo_epi16(m, m_inv); + __m256i pred_l = _mm256_madd_epi16(data_l, mask_l); + pred_l = _mm256_srai_epi32(_mm256_add_epi32(pred_l, round_const), + AOM_BLEND_A64_ROUND_BITS); + + const __m256i data_r = _mm256_unpackhi_epi16(a, b); + const __m256i mask_r = _mm256_unpackhi_epi16(m, m_inv); + __m256i pred_r = _mm256_madd_epi16(data_r, mask_r); + pred_r = _mm256_srai_epi32(_mm256_add_epi32(pred_r, round_const), + AOM_BLEND_A64_ROUND_BITS); + + // Note: the maximum value in pred_l/r is (2^bd)-1 < 2^15, + // so it is safe to do signed saturation here. + const __m256i pred = _mm256_packs_epi32(pred_l, pred_r); + // There is no 16-bit SAD instruction, so we have to synthesize + // an 8-element SAD. We do this by storing 4 32-bit partial SADs, + // and accumulating them at the end + const __m256i diff = _mm256_abs_epi16(_mm256_sub_epi16(pred, src)); + res = _mm256_add_epi32(res, _mm256_madd_epi16(diff, one)); + + src_ptr += src_stride << 1; + a_ptr += a_stride << 1; + b_ptr += b_stride << 1; + m_ptr += m_stride << 1; + } + // At this point, we have four 32-bit partial SADs stored in 'res'. + res = _mm256_hadd_epi32(res, res); + res = _mm256_hadd_epi32(res, res); + int sad = _mm256_extract_epi32(res, 0) + _mm256_extract_epi32(res, 4); + return (sad + 31) >> 6; +} + +static INLINE unsigned int highbd_masked_sad16xh_avx2( + const uint8_t *src8, int src_stride, const uint8_t *a8, int a_stride, + const uint8_t *b8, int b_stride, const uint8_t *m_ptr, int m_stride, + int width, int height) { + const uint16_t *src_ptr = CONVERT_TO_SHORTPTR(src8); + const uint16_t *a_ptr = CONVERT_TO_SHORTPTR(a8); + const uint16_t *b_ptr = CONVERT_TO_SHORTPTR(b8); + int x, y; + __m256i res = _mm256_setzero_si256(); + const __m256i mask_max = _mm256_set1_epi16((1 << AOM_BLEND_A64_ROUND_BITS)); + const __m256i round_const = + _mm256_set1_epi32((1 << AOM_BLEND_A64_ROUND_BITS) >> 1); + const __m256i one = _mm256_set1_epi16(1); + + for (y = 0; y < height; y++) { + for (x = 0; x < width; x += 16) { + const __m256i src = _mm256_lddqu_si256((const __m256i *)&src_ptr[x]); + const __m256i a = _mm256_lddqu_si256((const __m256i *)&a_ptr[x]); + const __m256i b = _mm256_lddqu_si256((const __m256i *)&b_ptr[x]); + // Zero-extend mask to 16 bits + const __m256i m = + _mm256_cvtepu8_epi16(_mm_lddqu_si128((const __m128i *)&m_ptr[x])); + const __m256i m_inv = _mm256_sub_epi16(mask_max, m); + + const __m256i data_l = _mm256_unpacklo_epi16(a, b); + const __m256i mask_l = _mm256_unpacklo_epi16(m, m_inv); + __m256i pred_l = _mm256_madd_epi16(data_l, mask_l); + pred_l = _mm256_srai_epi32(_mm256_add_epi32(pred_l, round_const), + AOM_BLEND_A64_ROUND_BITS); + + const __m256i data_r = _mm256_unpackhi_epi16(a, b); + const __m256i mask_r = _mm256_unpackhi_epi16(m, m_inv); + __m256i pred_r = _mm256_madd_epi16(data_r, mask_r); + pred_r = _mm256_srai_epi32(_mm256_add_epi32(pred_r, round_const), + AOM_BLEND_A64_ROUND_BITS); + + // Note: the maximum value in pred_l/r is (2^bd)-1 < 2^15, + // so it is safe to do signed saturation here. + const __m256i pred = _mm256_packs_epi32(pred_l, pred_r); + // There is no 16-bit SAD instruction, so we have to synthesize + // an 8-element SAD. We do this by storing 4 32-bit partial SADs, + // and accumulating them at the end + const __m256i diff = _mm256_abs_epi16(_mm256_sub_epi16(pred, src)); + res = _mm256_add_epi32(res, _mm256_madd_epi16(diff, one)); + } + + src_ptr += src_stride; + a_ptr += a_stride; + b_ptr += b_stride; + m_ptr += m_stride; + } + // At this point, we have four 32-bit partial SADs stored in 'res'. + res = _mm256_hadd_epi32(res, res); + res = _mm256_hadd_epi32(res, res); + int sad = _mm256_extract_epi32(res, 0) + _mm256_extract_epi32(res, 4); + return (sad + 31) >> 6; +} + +static INLINE unsigned int aom_highbd_masked_sad_avx2( + const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, + const uint8_t *second_pred, const uint8_t *msk, int msk_stride, + int invert_mask, int m, int n) { + unsigned int sad; + if (!invert_mask) { + switch (m) { + case 4: + sad = + aom_highbd_masked_sad4xh_ssse3(src, src_stride, ref, ref_stride, + second_pred, m, msk, msk_stride, n); + break; + case 8: + sad = highbd_masked_sad8xh_avx2(src, src_stride, ref, ref_stride, + second_pred, m, msk, msk_stride, n); + break; + default: + sad = highbd_masked_sad16xh_avx2(src, src_stride, ref, ref_stride, + second_pred, m, msk, msk_stride, m, n); + break; + } + } else { + switch (m) { + case 4: + sad = + aom_highbd_masked_sad4xh_ssse3(src, src_stride, second_pred, m, ref, + ref_stride, msk, msk_stride, n); + break; + case 8: + sad = highbd_masked_sad8xh_avx2(src, src_stride, second_pred, m, ref, + ref_stride, msk, msk_stride, n); + break; + default: + sad = highbd_masked_sad16xh_avx2(src, src_stride, second_pred, m, ref, + ref_stride, msk, msk_stride, m, n); + break; + } + } + return sad; +} + +#define HIGHBD_MASKSADMXN_AVX2(m, n) \ + unsigned int aom_highbd_masked_sad##m##x##n##_avx2( \ + const uint8_t *src8, int src_stride, const uint8_t *ref8, \ + int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, \ + int msk_stride, int invert_mask) { \ + return aom_highbd_masked_sad_avx2(src8, src_stride, ref8, ref_stride, \ + second_pred8, msk, msk_stride, \ + invert_mask, m, n); \ + } + +HIGHBD_MASKSADMXN_AVX2(4, 4); +HIGHBD_MASKSADMXN_AVX2(4, 8); +HIGHBD_MASKSADMXN_AVX2(8, 4); +HIGHBD_MASKSADMXN_AVX2(8, 8); +HIGHBD_MASKSADMXN_AVX2(8, 16); +HIGHBD_MASKSADMXN_AVX2(16, 8); +HIGHBD_MASKSADMXN_AVX2(16, 16); +HIGHBD_MASKSADMXN_AVX2(16, 32); +HIGHBD_MASKSADMXN_AVX2(32, 16); +HIGHBD_MASKSADMXN_AVX2(32, 32); +HIGHBD_MASKSADMXN_AVX2(32, 64); +HIGHBD_MASKSADMXN_AVX2(64, 32); +HIGHBD_MASKSADMXN_AVX2(64, 64); +HIGHBD_MASKSADMXN_AVX2(64, 128); +HIGHBD_MASKSADMXN_AVX2(128, 64); +HIGHBD_MASKSADMXN_AVX2(128, 128); +HIGHBD_MASKSADMXN_AVX2(4, 16); +HIGHBD_MASKSADMXN_AVX2(16, 4); +HIGHBD_MASKSADMXN_AVX2(8, 32); +HIGHBD_MASKSADMXN_AVX2(32, 8); +HIGHBD_MASKSADMXN_AVX2(16, 64); +HIGHBD_MASKSADMXN_AVX2(64, 16); diff --git a/third_party/aom/aom_dsp/x86/masked_sad_intrin_ssse3.c b/third_party/aom/aom_dsp/x86/masked_sad_intrin_ssse3.c new file mode 100644 index 0000000000..493f9bd8f2 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/masked_sad_intrin_ssse3.c @@ -0,0 +1,402 @@ +/* + * Copyright (c) 2017, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include +#include + +#include "config/aom_config.h" +#include "config/aom_dsp_rtcd.h" + +#include "aom_dsp/blend.h" +#include "aom/aom_integer.h" +#include "aom_dsp/x86/synonyms.h" + +#include "aom_dsp/x86//masked_sad_intrin_ssse3.h" + +// For width a multiple of 16 +static INLINE unsigned int masked_sad_ssse3(const uint8_t *src_ptr, + int src_stride, + const uint8_t *a_ptr, int a_stride, + const uint8_t *b_ptr, int b_stride, + const uint8_t *m_ptr, int m_stride, + int width, int height); + +#define MASKSADMXN_SSSE3(m, n) \ + unsigned int aom_masked_sad##m##x##n##_ssse3( \ + const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, \ + const uint8_t *second_pred, const uint8_t *msk, int msk_stride, \ + int invert_mask) { \ + if (!invert_mask) \ + return masked_sad_ssse3(src, src_stride, ref, ref_stride, second_pred, \ + m, msk, msk_stride, m, n); \ + else \ + return masked_sad_ssse3(src, src_stride, second_pred, m, ref, \ + ref_stride, msk, msk_stride, m, n); \ + } + +#define MASKSAD8XN_SSSE3(n) \ + unsigned int aom_masked_sad8x##n##_ssse3( \ + const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, \ + const uint8_t *second_pred, const uint8_t *msk, int msk_stride, \ + int invert_mask) { \ + if (!invert_mask) \ + return aom_masked_sad8xh_ssse3(src, src_stride, ref, ref_stride, \ + second_pred, 8, msk, msk_stride, n); \ + else \ + return aom_masked_sad8xh_ssse3(src, src_stride, second_pred, 8, ref, \ + ref_stride, msk, msk_stride, n); \ + } + +#define MASKSAD4XN_SSSE3(n) \ + unsigned int aom_masked_sad4x##n##_ssse3( \ + const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, \ + const uint8_t *second_pred, const uint8_t *msk, int msk_stride, \ + int invert_mask) { \ + if (!invert_mask) \ + return aom_masked_sad4xh_ssse3(src, src_stride, ref, ref_stride, \ + second_pred, 4, msk, msk_stride, n); \ + else \ + return aom_masked_sad4xh_ssse3(src, src_stride, second_pred, 4, ref, \ + ref_stride, msk, msk_stride, n); \ + } + +MASKSADMXN_SSSE3(128, 128) +MASKSADMXN_SSSE3(128, 64) +MASKSADMXN_SSSE3(64, 128) +MASKSADMXN_SSSE3(64, 64) +MASKSADMXN_SSSE3(64, 32) +MASKSADMXN_SSSE3(32, 64) +MASKSADMXN_SSSE3(32, 32) +MASKSADMXN_SSSE3(32, 16) +MASKSADMXN_SSSE3(16, 32) +MASKSADMXN_SSSE3(16, 16) +MASKSADMXN_SSSE3(16, 8) +MASKSAD8XN_SSSE3(16) +MASKSAD8XN_SSSE3(8) +MASKSAD8XN_SSSE3(4) +MASKSAD4XN_SSSE3(8) +MASKSAD4XN_SSSE3(4) +MASKSAD4XN_SSSE3(16) +MASKSADMXN_SSSE3(16, 4) +MASKSAD8XN_SSSE3(32) +MASKSADMXN_SSSE3(32, 8) +MASKSADMXN_SSSE3(16, 64) +MASKSADMXN_SSSE3(64, 16) + +static INLINE unsigned int masked_sad_ssse3(const uint8_t *src_ptr, + int src_stride, + const uint8_t *a_ptr, int a_stride, + const uint8_t *b_ptr, int b_stride, + const uint8_t *m_ptr, int m_stride, + int width, int height) { + int x, y; + __m128i res = _mm_setzero_si128(); + const __m128i mask_max = _mm_set1_epi8((1 << AOM_BLEND_A64_ROUND_BITS)); + + for (y = 0; y < height; y++) { + for (x = 0; x < width; x += 16) { + const __m128i src = _mm_loadu_si128((const __m128i *)&src_ptr[x]); + const __m128i a = _mm_loadu_si128((const __m128i *)&a_ptr[x]); + const __m128i b = _mm_loadu_si128((const __m128i *)&b_ptr[x]); + const __m128i m = _mm_loadu_si128((const __m128i *)&m_ptr[x]); + const __m128i m_inv = _mm_sub_epi8(mask_max, m); + + // Calculate 16 predicted pixels. + // Note that the maximum value of any entry of 'pred_l' or 'pred_r' + // is 64 * 255, so we have plenty of space to add rounding constants. + const __m128i data_l = _mm_unpacklo_epi8(a, b); + const __m128i mask_l = _mm_unpacklo_epi8(m, m_inv); + __m128i pred_l = _mm_maddubs_epi16(data_l, mask_l); + pred_l = xx_roundn_epu16(pred_l, AOM_BLEND_A64_ROUND_BITS); + + const __m128i data_r = _mm_unpackhi_epi8(a, b); + const __m128i mask_r = _mm_unpackhi_epi8(m, m_inv); + __m128i pred_r = _mm_maddubs_epi16(data_r, mask_r); + pred_r = xx_roundn_epu16(pred_r, AOM_BLEND_A64_ROUND_BITS); + + const __m128i pred = _mm_packus_epi16(pred_l, pred_r); + res = _mm_add_epi32(res, _mm_sad_epu8(pred, src)); + } + + src_ptr += src_stride; + a_ptr += a_stride; + b_ptr += b_stride; + m_ptr += m_stride; + } + // At this point, we have two 32-bit partial SADs in lanes 0 and 2 of 'res'. + int32_t sad = + _mm_cvtsi128_si32(res) + _mm_cvtsi128_si32(_mm_srli_si128(res, 8)); + return (sad + 31) >> 6; +} + +unsigned int aom_masked_sad8xh_ssse3(const uint8_t *src_ptr, int src_stride, + const uint8_t *a_ptr, int a_stride, + const uint8_t *b_ptr, int b_stride, + const uint8_t *m_ptr, int m_stride, + int height) { + int y; + __m128i res = _mm_setzero_si128(); + const __m128i mask_max = _mm_set1_epi8((1 << AOM_BLEND_A64_ROUND_BITS)); + + for (y = 0; y < height; y += 2) { + const __m128i src = _mm_unpacklo_epi64( + _mm_loadl_epi64((const __m128i *)src_ptr), + _mm_loadl_epi64((const __m128i *)&src_ptr[src_stride])); + const __m128i a0 = _mm_loadl_epi64((const __m128i *)a_ptr); + const __m128i a1 = _mm_loadl_epi64((const __m128i *)&a_ptr[a_stride]); + const __m128i b0 = _mm_loadl_epi64((const __m128i *)b_ptr); + const __m128i b1 = _mm_loadl_epi64((const __m128i *)&b_ptr[b_stride]); + const __m128i m = + _mm_unpacklo_epi64(_mm_loadl_epi64((const __m128i *)m_ptr), + _mm_loadl_epi64((const __m128i *)&m_ptr[m_stride])); + const __m128i m_inv = _mm_sub_epi8(mask_max, m); + + const __m128i data_l = _mm_unpacklo_epi8(a0, b0); + const __m128i mask_l = _mm_unpacklo_epi8(m, m_inv); + __m128i pred_l = _mm_maddubs_epi16(data_l, mask_l); + pred_l = xx_roundn_epu16(pred_l, AOM_BLEND_A64_ROUND_BITS); + + const __m128i data_r = _mm_unpacklo_epi8(a1, b1); + const __m128i mask_r = _mm_unpackhi_epi8(m, m_inv); + __m128i pred_r = _mm_maddubs_epi16(data_r, mask_r); + pred_r = xx_roundn_epu16(pred_r, AOM_BLEND_A64_ROUND_BITS); + + const __m128i pred = _mm_packus_epi16(pred_l, pred_r); + res = _mm_add_epi32(res, _mm_sad_epu8(pred, src)); + + src_ptr += src_stride * 2; + a_ptr += a_stride * 2; + b_ptr += b_stride * 2; + m_ptr += m_stride * 2; + } + int32_t sad = + _mm_cvtsi128_si32(res) + _mm_cvtsi128_si32(_mm_srli_si128(res, 8)); + return (sad + 31) >> 6; +} + +unsigned int aom_masked_sad4xh_ssse3(const uint8_t *src_ptr, int src_stride, + const uint8_t *a_ptr, int a_stride, + const uint8_t *b_ptr, int b_stride, + const uint8_t *m_ptr, int m_stride, + int height) { + int y; + __m128i res = _mm_setzero_si128(); + const __m128i mask_max = _mm_set1_epi8((1 << AOM_BLEND_A64_ROUND_BITS)); + + for (y = 0; y < height; y += 2) { + // Load two rows at a time, this seems to be a bit faster + // than four rows at a time in this case. + const __m128i src = _mm_unpacklo_epi32( + _mm_cvtsi32_si128(*(uint32_t *)src_ptr), + _mm_cvtsi32_si128(*(uint32_t *)&src_ptr[src_stride])); + const __m128i a = + _mm_unpacklo_epi32(_mm_cvtsi32_si128(*(uint32_t *)a_ptr), + _mm_cvtsi32_si128(*(uint32_t *)&a_ptr[a_stride])); + const __m128i b = + _mm_unpacklo_epi32(_mm_cvtsi32_si128(*(uint32_t *)b_ptr), + _mm_cvtsi32_si128(*(uint32_t *)&b_ptr[b_stride])); + const __m128i m = + _mm_unpacklo_epi32(_mm_cvtsi32_si128(*(uint32_t *)m_ptr), + _mm_cvtsi32_si128(*(uint32_t *)&m_ptr[m_stride])); + const __m128i m_inv = _mm_sub_epi8(mask_max, m); + + const __m128i data = _mm_unpacklo_epi8(a, b); + const __m128i mask = _mm_unpacklo_epi8(m, m_inv); + __m128i pred_16bit = _mm_maddubs_epi16(data, mask); + pred_16bit = xx_roundn_epu16(pred_16bit, AOM_BLEND_A64_ROUND_BITS); + + const __m128i pred = _mm_packus_epi16(pred_16bit, _mm_setzero_si128()); + res = _mm_add_epi32(res, _mm_sad_epu8(pred, src)); + + src_ptr += src_stride * 2; + a_ptr += a_stride * 2; + b_ptr += b_stride * 2; + m_ptr += m_stride * 2; + } + // At this point, the SAD is stored in lane 0 of 'res' + int32_t sad = _mm_cvtsi128_si32(res); + return (sad + 31) >> 6; +} + +// For width a multiple of 8 +static INLINE unsigned int highbd_masked_sad_ssse3( + const uint8_t *src8, int src_stride, const uint8_t *a8, int a_stride, + const uint8_t *b8, int b_stride, const uint8_t *m_ptr, int m_stride, + int width, int height); + +#define HIGHBD_MASKSADMXN_SSSE3(m, n) \ + unsigned int aom_highbd_masked_sad##m##x##n##_ssse3( \ + const uint8_t *src8, int src_stride, const uint8_t *ref8, \ + int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, \ + int msk_stride, int invert_mask) { \ + if (!invert_mask) \ + return highbd_masked_sad_ssse3(src8, src_stride, ref8, ref_stride, \ + second_pred8, m, msk, msk_stride, m, n); \ + else \ + return highbd_masked_sad_ssse3(src8, src_stride, second_pred8, m, ref8, \ + ref_stride, msk, msk_stride, m, n); \ + } + +#define HIGHBD_MASKSAD4XN_SSSE3(n) \ + unsigned int aom_highbd_masked_sad4x##n##_ssse3( \ + const uint8_t *src8, int src_stride, const uint8_t *ref8, \ + int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, \ + int msk_stride, int invert_mask) { \ + if (!invert_mask) \ + return aom_highbd_masked_sad4xh_ssse3(src8, src_stride, ref8, \ + ref_stride, second_pred8, 4, msk, \ + msk_stride, n); \ + else \ + return aom_highbd_masked_sad4xh_ssse3(src8, src_stride, second_pred8, 4, \ + ref8, ref_stride, msk, msk_stride, \ + n); \ + } + +HIGHBD_MASKSADMXN_SSSE3(128, 128) +HIGHBD_MASKSADMXN_SSSE3(128, 64) +HIGHBD_MASKSADMXN_SSSE3(64, 128) +HIGHBD_MASKSADMXN_SSSE3(64, 64) +HIGHBD_MASKSADMXN_SSSE3(64, 32) +HIGHBD_MASKSADMXN_SSSE3(32, 64) +HIGHBD_MASKSADMXN_SSSE3(32, 32) +HIGHBD_MASKSADMXN_SSSE3(32, 16) +HIGHBD_MASKSADMXN_SSSE3(16, 32) +HIGHBD_MASKSADMXN_SSSE3(16, 16) +HIGHBD_MASKSADMXN_SSSE3(16, 8) +HIGHBD_MASKSADMXN_SSSE3(8, 16) +HIGHBD_MASKSADMXN_SSSE3(8, 8) +HIGHBD_MASKSADMXN_SSSE3(8, 4) +HIGHBD_MASKSAD4XN_SSSE3(8) +HIGHBD_MASKSAD4XN_SSSE3(4) +HIGHBD_MASKSAD4XN_SSSE3(16) +HIGHBD_MASKSADMXN_SSSE3(16, 4) +HIGHBD_MASKSADMXN_SSSE3(8, 32) +HIGHBD_MASKSADMXN_SSSE3(32, 8) +HIGHBD_MASKSADMXN_SSSE3(16, 64) +HIGHBD_MASKSADMXN_SSSE3(64, 16) + +static INLINE unsigned int highbd_masked_sad_ssse3( + const uint8_t *src8, int src_stride, const uint8_t *a8, int a_stride, + const uint8_t *b8, int b_stride, const uint8_t *m_ptr, int m_stride, + int width, int height) { + const uint16_t *src_ptr = CONVERT_TO_SHORTPTR(src8); + const uint16_t *a_ptr = CONVERT_TO_SHORTPTR(a8); + const uint16_t *b_ptr = CONVERT_TO_SHORTPTR(b8); + int x, y; + __m128i res = _mm_setzero_si128(); + const __m128i mask_max = _mm_set1_epi16((1 << AOM_BLEND_A64_ROUND_BITS)); + const __m128i round_const = + _mm_set1_epi32((1 << AOM_BLEND_A64_ROUND_BITS) >> 1); + const __m128i one = _mm_set1_epi16(1); + + for (y = 0; y < height; y++) { + for (x = 0; x < width; x += 8) { + const __m128i src = _mm_loadu_si128((const __m128i *)&src_ptr[x]); + const __m128i a = _mm_loadu_si128((const __m128i *)&a_ptr[x]); + const __m128i b = _mm_loadu_si128((const __m128i *)&b_ptr[x]); + // Zero-extend mask to 16 bits + const __m128i m = _mm_unpacklo_epi8( + _mm_loadl_epi64((const __m128i *)&m_ptr[x]), _mm_setzero_si128()); + const __m128i m_inv = _mm_sub_epi16(mask_max, m); + + const __m128i data_l = _mm_unpacklo_epi16(a, b); + const __m128i mask_l = _mm_unpacklo_epi16(m, m_inv); + __m128i pred_l = _mm_madd_epi16(data_l, mask_l); + pred_l = _mm_srai_epi32(_mm_add_epi32(pred_l, round_const), + AOM_BLEND_A64_ROUND_BITS); + + const __m128i data_r = _mm_unpackhi_epi16(a, b); + const __m128i mask_r = _mm_unpackhi_epi16(m, m_inv); + __m128i pred_r = _mm_madd_epi16(data_r, mask_r); + pred_r = _mm_srai_epi32(_mm_add_epi32(pred_r, round_const), + AOM_BLEND_A64_ROUND_BITS); + + // Note: the maximum value in pred_l/r is (2^bd)-1 < 2^15, + // so it is safe to do signed saturation here. + const __m128i pred = _mm_packs_epi32(pred_l, pred_r); + // There is no 16-bit SAD instruction, so we have to synthesize + // an 8-element SAD. We do this by storing 4 32-bit partial SADs, + // and accumulating them at the end + const __m128i diff = _mm_abs_epi16(_mm_sub_epi16(pred, src)); + res = _mm_add_epi32(res, _mm_madd_epi16(diff, one)); + } + + src_ptr += src_stride; + a_ptr += a_stride; + b_ptr += b_stride; + m_ptr += m_stride; + } + // At this point, we have four 32-bit partial SADs stored in 'res'. + res = _mm_hadd_epi32(res, res); + res = _mm_hadd_epi32(res, res); + int sad = _mm_cvtsi128_si32(res); + return (sad + 31) >> 6; +} + +unsigned int aom_highbd_masked_sad4xh_ssse3(const uint8_t *src8, int src_stride, + const uint8_t *a8, int a_stride, + const uint8_t *b8, int b_stride, + const uint8_t *m_ptr, int m_stride, + int height) { + const uint16_t *src_ptr = CONVERT_TO_SHORTPTR(src8); + const uint16_t *a_ptr = CONVERT_TO_SHORTPTR(a8); + const uint16_t *b_ptr = CONVERT_TO_SHORTPTR(b8); + int y; + __m128i res = _mm_setzero_si128(); + const __m128i mask_max = _mm_set1_epi16((1 << AOM_BLEND_A64_ROUND_BITS)); + const __m128i round_const = + _mm_set1_epi32((1 << AOM_BLEND_A64_ROUND_BITS) >> 1); + const __m128i one = _mm_set1_epi16(1); + + for (y = 0; y < height; y += 2) { + const __m128i src = _mm_unpacklo_epi64( + _mm_loadl_epi64((const __m128i *)src_ptr), + _mm_loadl_epi64((const __m128i *)&src_ptr[src_stride])); + const __m128i a = + _mm_unpacklo_epi64(_mm_loadl_epi64((const __m128i *)a_ptr), + _mm_loadl_epi64((const __m128i *)&a_ptr[a_stride])); + const __m128i b = + _mm_unpacklo_epi64(_mm_loadl_epi64((const __m128i *)b_ptr), + _mm_loadl_epi64((const __m128i *)&b_ptr[b_stride])); + // Zero-extend mask to 16 bits + const __m128i m = _mm_unpacklo_epi8( + _mm_unpacklo_epi32( + _mm_cvtsi32_si128(*(const uint32_t *)m_ptr), + _mm_cvtsi32_si128(*(const uint32_t *)&m_ptr[m_stride])), + _mm_setzero_si128()); + const __m128i m_inv = _mm_sub_epi16(mask_max, m); + + const __m128i data_l = _mm_unpacklo_epi16(a, b); + const __m128i mask_l = _mm_unpacklo_epi16(m, m_inv); + __m128i pred_l = _mm_madd_epi16(data_l, mask_l); + pred_l = _mm_srai_epi32(_mm_add_epi32(pred_l, round_const), + AOM_BLEND_A64_ROUND_BITS); + + const __m128i data_r = _mm_unpackhi_epi16(a, b); + const __m128i mask_r = _mm_unpackhi_epi16(m, m_inv); + __m128i pred_r = _mm_madd_epi16(data_r, mask_r); + pred_r = _mm_srai_epi32(_mm_add_epi32(pred_r, round_const), + AOM_BLEND_A64_ROUND_BITS); + + const __m128i pred = _mm_packs_epi32(pred_l, pred_r); + const __m128i diff = _mm_abs_epi16(_mm_sub_epi16(pred, src)); + res = _mm_add_epi32(res, _mm_madd_epi16(diff, one)); + + src_ptr += src_stride * 2; + a_ptr += a_stride * 2; + b_ptr += b_stride * 2; + m_ptr += m_stride * 2; + } + res = _mm_hadd_epi32(res, res); + res = _mm_hadd_epi32(res, res); + int sad = _mm_cvtsi128_si32(res); + return (sad + 31) >> 6; +} diff --git a/third_party/aom/aom_dsp/x86/masked_sad_intrin_ssse3.h b/third_party/aom/aom_dsp/x86/masked_sad_intrin_ssse3.h new file mode 100644 index 0000000000..cffbd9672c --- /dev/null +++ b/third_party/aom/aom_dsp/x86/masked_sad_intrin_ssse3.h @@ -0,0 +1,33 @@ +/* + * Copyright (c) 2018, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_X86_MASKED_SAD_INTRIN_SSSE3_H_ +#define AOM_AOM_DSP_X86_MASKED_SAD_INTRIN_SSSE3_H_ + +unsigned int aom_masked_sad8xh_ssse3(const uint8_t *src_ptr, int src_stride, + const uint8_t *a_ptr, int a_stride, + const uint8_t *b_ptr, int b_stride, + const uint8_t *m_ptr, int m_stride, + int height); + +unsigned int aom_masked_sad4xh_ssse3(const uint8_t *src_ptr, int src_stride, + const uint8_t *a_ptr, int a_stride, + const uint8_t *b_ptr, int b_stride, + const uint8_t *m_ptr, int m_stride, + int height); + +unsigned int aom_highbd_masked_sad4xh_ssse3(const uint8_t *src8, int src_stride, + const uint8_t *a8, int a_stride, + const uint8_t *b8, int b_stride, + const uint8_t *m_ptr, int m_stride, + int height); + +#endif // AOM_AOM_DSP_X86_MASKED_SAD_INTRIN_SSSE3_H_ diff --git a/third_party/aom/aom_dsp/x86/masked_variance_intrin_ssse3.c b/third_party/aom/aom_dsp/x86/masked_variance_intrin_ssse3.c new file mode 100644 index 0000000000..ebf4631ba7 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/masked_variance_intrin_ssse3.c @@ -0,0 +1,1065 @@ +/* + * Copyright (c) 2017, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include +#include +#include + +#include "config/aom_config.h" +#include "config/aom_dsp_rtcd.h" + +#include "aom/aom_integer.h" +#include "aom_dsp/aom_filter.h" +#include "aom_dsp/blend.h" +#include "aom_dsp/x86/masked_variance_intrin_ssse3.h" +#include "aom_dsp/x86/synonyms.h" +#include "aom_ports/mem.h" + +// For width a multiple of 16 +static void bilinear_filter(const uint8_t *src, int src_stride, int xoffset, + int yoffset, uint8_t *dst, int w, int h); + +static void bilinear_filter8xh(const uint8_t *src, int src_stride, int xoffset, + int yoffset, uint8_t *dst, int h); + +static void bilinear_filter4xh(const uint8_t *src, int src_stride, int xoffset, + int yoffset, uint8_t *dst, int h); + +// For width a multiple of 16 +static void masked_variance(const uint8_t *src_ptr, int src_stride, + const uint8_t *a_ptr, int a_stride, + const uint8_t *b_ptr, int b_stride, + const uint8_t *m_ptr, int m_stride, int width, + int height, unsigned int *sse, int *sum_); + +static void masked_variance8xh(const uint8_t *src_ptr, int src_stride, + const uint8_t *a_ptr, const uint8_t *b_ptr, + const uint8_t *m_ptr, int m_stride, int height, + unsigned int *sse, int *sum_); + +static void masked_variance4xh(const uint8_t *src_ptr, int src_stride, + const uint8_t *a_ptr, const uint8_t *b_ptr, + const uint8_t *m_ptr, int m_stride, int height, + unsigned int *sse, int *sum_); + +#define MASK_SUBPIX_VAR_SSSE3(W, H) \ + unsigned int aom_masked_sub_pixel_variance##W##x##H##_ssse3( \ + const uint8_t *src, int src_stride, int xoffset, int yoffset, \ + const uint8_t *ref, int ref_stride, const uint8_t *second_pred, \ + const uint8_t *msk, int msk_stride, int invert_mask, \ + unsigned int *sse) { \ + int sum; \ + uint8_t temp[(H + 1) * W]; \ + \ + bilinear_filter(src, src_stride, xoffset, yoffset, temp, W, H); \ + \ + if (!invert_mask) \ + masked_variance(ref, ref_stride, temp, W, second_pred, W, msk, \ + msk_stride, W, H, sse, &sum); \ + else \ + masked_variance(ref, ref_stride, second_pred, W, temp, W, msk, \ + msk_stride, W, H, sse, &sum); \ + return *sse - (uint32_t)(((int64_t)sum * sum) / (W * H)); \ + } + +#define MASK_SUBPIX_VAR8XH_SSSE3(H) \ + unsigned int aom_masked_sub_pixel_variance8x##H##_ssse3( \ + const uint8_t *src, int src_stride, int xoffset, int yoffset, \ + const uint8_t *ref, int ref_stride, const uint8_t *second_pred, \ + const uint8_t *msk, int msk_stride, int invert_mask, \ + unsigned int *sse) { \ + int sum; \ + uint8_t temp[(H + 1) * 8]; \ + \ + bilinear_filter8xh(src, src_stride, xoffset, yoffset, temp, H); \ + \ + if (!invert_mask) \ + masked_variance8xh(ref, ref_stride, temp, second_pred, msk, msk_stride, \ + H, sse, &sum); \ + else \ + masked_variance8xh(ref, ref_stride, second_pred, temp, msk, msk_stride, \ + H, sse, &sum); \ + return *sse - (uint32_t)(((int64_t)sum * sum) / (8 * H)); \ + } + +#define MASK_SUBPIX_VAR4XH_SSSE3(H) \ + unsigned int aom_masked_sub_pixel_variance4x##H##_ssse3( \ + const uint8_t *src, int src_stride, int xoffset, int yoffset, \ + const uint8_t *ref, int ref_stride, const uint8_t *second_pred, \ + const uint8_t *msk, int msk_stride, int invert_mask, \ + unsigned int *sse) { \ + int sum; \ + uint8_t temp[(H + 1) * 4]; \ + \ + bilinear_filter4xh(src, src_stride, xoffset, yoffset, temp, H); \ + \ + if (!invert_mask) \ + masked_variance4xh(ref, ref_stride, temp, second_pred, msk, msk_stride, \ + H, sse, &sum); \ + else \ + masked_variance4xh(ref, ref_stride, second_pred, temp, msk, msk_stride, \ + H, sse, &sum); \ + return *sse - (uint32_t)(((int64_t)sum * sum) / (4 * H)); \ + } + +MASK_SUBPIX_VAR_SSSE3(128, 128) +MASK_SUBPIX_VAR_SSSE3(128, 64) +MASK_SUBPIX_VAR_SSSE3(64, 128) +MASK_SUBPIX_VAR_SSSE3(64, 64) +MASK_SUBPIX_VAR_SSSE3(64, 32) +MASK_SUBPIX_VAR_SSSE3(32, 64) +MASK_SUBPIX_VAR_SSSE3(32, 32) +MASK_SUBPIX_VAR_SSSE3(32, 16) +MASK_SUBPIX_VAR_SSSE3(16, 32) +MASK_SUBPIX_VAR_SSSE3(16, 16) +MASK_SUBPIX_VAR_SSSE3(16, 8) +MASK_SUBPIX_VAR8XH_SSSE3(16) +MASK_SUBPIX_VAR8XH_SSSE3(8) +MASK_SUBPIX_VAR8XH_SSSE3(4) +MASK_SUBPIX_VAR4XH_SSSE3(8) +MASK_SUBPIX_VAR4XH_SSSE3(4) +MASK_SUBPIX_VAR4XH_SSSE3(16) +MASK_SUBPIX_VAR_SSSE3(16, 4) +MASK_SUBPIX_VAR8XH_SSSE3(32) +MASK_SUBPIX_VAR_SSSE3(32, 8) +MASK_SUBPIX_VAR_SSSE3(64, 16) +MASK_SUBPIX_VAR_SSSE3(16, 64) + +static INLINE __m128i filter_block(const __m128i a, const __m128i b, + const __m128i filter) { + __m128i v0 = _mm_unpacklo_epi8(a, b); + v0 = _mm_maddubs_epi16(v0, filter); + v0 = xx_roundn_epu16(v0, FILTER_BITS); + + __m128i v1 = _mm_unpackhi_epi8(a, b); + v1 = _mm_maddubs_epi16(v1, filter); + v1 = xx_roundn_epu16(v1, FILTER_BITS); + + return _mm_packus_epi16(v0, v1); +} + +static void bilinear_filter(const uint8_t *src, int src_stride, int xoffset, + int yoffset, uint8_t *dst, int w, int h) { + int i, j; + // Horizontal filter + if (xoffset == 0) { + uint8_t *b = dst; + for (i = 0; i < h + 1; ++i) { + for (j = 0; j < w; j += 16) { + __m128i x = _mm_loadu_si128((__m128i *)&src[j]); + _mm_storeu_si128((__m128i *)&b[j], x); + } + src += src_stride; + b += w; + } + } else if (xoffset == 4) { + uint8_t *b = dst; + for (i = 0; i < h + 1; ++i) { + for (j = 0; j < w; j += 16) { + __m128i x = _mm_loadu_si128((__m128i *)&src[j]); + __m128i y = _mm_loadu_si128((__m128i *)&src[j + 16]); + __m128i z = _mm_alignr_epi8(y, x, 1); + _mm_storeu_si128((__m128i *)&b[j], _mm_avg_epu8(x, z)); + } + src += src_stride; + b += w; + } + } else { + uint8_t *b = dst; + const uint8_t *hfilter = bilinear_filters_2t[xoffset]; + const __m128i hfilter_vec = _mm_set1_epi16(hfilter[0] | (hfilter[1] << 8)); + for (i = 0; i < h + 1; ++i) { + for (j = 0; j < w; j += 16) { + const __m128i x = _mm_loadu_si128((__m128i *)&src[j]); + const __m128i y = _mm_loadu_si128((__m128i *)&src[j + 16]); + const __m128i z = _mm_alignr_epi8(y, x, 1); + const __m128i res = filter_block(x, z, hfilter_vec); + _mm_storeu_si128((__m128i *)&b[j], res); + } + + src += src_stride; + b += w; + } + } + + // Vertical filter + if (yoffset == 0) { + // The data is already in 'dst', so no need to filter + } else if (yoffset == 4) { + for (i = 0; i < h; ++i) { + for (j = 0; j < w; j += 16) { + __m128i x = _mm_loadu_si128((__m128i *)&dst[j]); + __m128i y = _mm_loadu_si128((__m128i *)&dst[j + w]); + _mm_storeu_si128((__m128i *)&dst[j], _mm_avg_epu8(x, y)); + } + dst += w; + } + } else { + const uint8_t *vfilter = bilinear_filters_2t[yoffset]; + const __m128i vfilter_vec = _mm_set1_epi16(vfilter[0] | (vfilter[1] << 8)); + for (i = 0; i < h; ++i) { + for (j = 0; j < w; j += 16) { + const __m128i x = _mm_loadu_si128((__m128i *)&dst[j]); + const __m128i y = _mm_loadu_si128((__m128i *)&dst[j + w]); + const __m128i res = filter_block(x, y, vfilter_vec); + _mm_storeu_si128((__m128i *)&dst[j], res); + } + + dst += w; + } + } +} + +static INLINE __m128i filter_block_2rows(const __m128i *a0, const __m128i *b0, + const __m128i *a1, const __m128i *b1, + const __m128i *filter) { + __m128i v0 = _mm_unpacklo_epi8(*a0, *b0); + v0 = _mm_maddubs_epi16(v0, *filter); + v0 = xx_roundn_epu16(v0, FILTER_BITS); + + __m128i v1 = _mm_unpacklo_epi8(*a1, *b1); + v1 = _mm_maddubs_epi16(v1, *filter); + v1 = xx_roundn_epu16(v1, FILTER_BITS); + + return _mm_packus_epi16(v0, v1); +} + +static void bilinear_filter8xh(const uint8_t *src, int src_stride, int xoffset, + int yoffset, uint8_t *dst, int h) { + int i; + // Horizontal filter + if (xoffset == 0) { + uint8_t *b = dst; + for (i = 0; i < h + 1; ++i) { + __m128i x = _mm_loadl_epi64((__m128i *)src); + _mm_storel_epi64((__m128i *)b, x); + src += src_stride; + b += 8; + } + } else if (xoffset == 4) { + uint8_t *b = dst; + for (i = 0; i < h + 1; ++i) { + __m128i x = _mm_loadu_si128((__m128i *)src); + __m128i z = _mm_srli_si128(x, 1); + _mm_storel_epi64((__m128i *)b, _mm_avg_epu8(x, z)); + src += src_stride; + b += 8; + } + } else { + uint8_t *b = dst; + const uint8_t *hfilter = bilinear_filters_2t[xoffset]; + const __m128i hfilter_vec = _mm_set1_epi16(hfilter[0] | (hfilter[1] << 8)); + for (i = 0; i < h; i += 2) { + const __m128i x0 = _mm_loadu_si128((__m128i *)src); + const __m128i z0 = _mm_srli_si128(x0, 1); + const __m128i x1 = _mm_loadu_si128((__m128i *)&src[src_stride]); + const __m128i z1 = _mm_srli_si128(x1, 1); + const __m128i res = filter_block_2rows(&x0, &z0, &x1, &z1, &hfilter_vec); + _mm_storeu_si128((__m128i *)b, res); + + src += src_stride * 2; + b += 16; + } + // Handle i = h separately + const __m128i x0 = _mm_loadu_si128((__m128i *)src); + const __m128i z0 = _mm_srli_si128(x0, 1); + + __m128i v0 = _mm_unpacklo_epi8(x0, z0); + v0 = _mm_maddubs_epi16(v0, hfilter_vec); + v0 = xx_roundn_epu16(v0, FILTER_BITS); + + _mm_storel_epi64((__m128i *)b, _mm_packus_epi16(v0, v0)); + } + + // Vertical filter + if (yoffset == 0) { + // The data is already in 'dst', so no need to filter + } else if (yoffset == 4) { + for (i = 0; i < h; ++i) { + __m128i x = _mm_loadl_epi64((__m128i *)dst); + __m128i y = _mm_loadl_epi64((__m128i *)&dst[8]); + _mm_storel_epi64((__m128i *)dst, _mm_avg_epu8(x, y)); + dst += 8; + } + } else { + const uint8_t *vfilter = bilinear_filters_2t[yoffset]; + const __m128i vfilter_vec = _mm_set1_epi16(vfilter[0] | (vfilter[1] << 8)); + for (i = 0; i < h; i += 2) { + const __m128i x = _mm_loadl_epi64((__m128i *)dst); + const __m128i y = _mm_loadl_epi64((__m128i *)&dst[8]); + const __m128i z = _mm_loadl_epi64((__m128i *)&dst[16]); + const __m128i res = filter_block_2rows(&x, &y, &y, &z, &vfilter_vec); + _mm_storeu_si128((__m128i *)dst, res); + + dst += 16; + } + } +} + +static void bilinear_filter4xh(const uint8_t *src, int src_stride, int xoffset, + int yoffset, uint8_t *dst, int h) { + int i; + // Horizontal filter + if (xoffset == 0) { + uint8_t *b = dst; + for (i = 0; i < h + 1; ++i) { + __m128i x = xx_loadl_32((__m128i *)src); + xx_storel_32((__m128i *)b, x); + src += src_stride; + b += 4; + } + } else if (xoffset == 4) { + uint8_t *b = dst; + for (i = 0; i < h + 1; ++i) { + __m128i x = _mm_loadl_epi64((__m128i *)src); + __m128i z = _mm_srli_si128(x, 1); + xx_storel_32((__m128i *)b, _mm_avg_epu8(x, z)); + src += src_stride; + b += 4; + } + } else { + uint8_t *b = dst; + const uint8_t *hfilter = bilinear_filters_2t[xoffset]; + const __m128i hfilter_vec = _mm_set1_epi16(hfilter[0] | (hfilter[1] << 8)); + for (i = 0; i < h; i += 4) { + const __m128i x0 = _mm_loadl_epi64((__m128i *)src); + const __m128i z0 = _mm_srli_si128(x0, 1); + const __m128i x1 = _mm_loadl_epi64((__m128i *)&src[src_stride]); + const __m128i z1 = _mm_srli_si128(x1, 1); + const __m128i x2 = _mm_loadl_epi64((__m128i *)&src[src_stride * 2]); + const __m128i z2 = _mm_srli_si128(x2, 1); + const __m128i x3 = _mm_loadl_epi64((__m128i *)&src[src_stride * 3]); + const __m128i z3 = _mm_srli_si128(x3, 1); + + const __m128i a0 = _mm_unpacklo_epi32(x0, x1); + const __m128i b0 = _mm_unpacklo_epi32(z0, z1); + const __m128i a1 = _mm_unpacklo_epi32(x2, x3); + const __m128i b1 = _mm_unpacklo_epi32(z2, z3); + const __m128i res = filter_block_2rows(&a0, &b0, &a1, &b1, &hfilter_vec); + _mm_storeu_si128((__m128i *)b, res); + + src += src_stride * 4; + b += 16; + } + // Handle i = h separately + const __m128i x = _mm_loadl_epi64((__m128i *)src); + const __m128i z = _mm_srli_si128(x, 1); + + __m128i v0 = _mm_unpacklo_epi8(x, z); + v0 = _mm_maddubs_epi16(v0, hfilter_vec); + v0 = xx_roundn_epu16(v0, FILTER_BITS); + + xx_storel_32((__m128i *)b, _mm_packus_epi16(v0, v0)); + } + + // Vertical filter + if (yoffset == 0) { + // The data is already in 'dst', so no need to filter + } else if (yoffset == 4) { + for (i = 0; i < h; ++i) { + __m128i x = xx_loadl_32((__m128i *)dst); + __m128i y = xx_loadl_32((__m128i *)&dst[4]); + xx_storel_32((__m128i *)dst, _mm_avg_epu8(x, y)); + dst += 4; + } + } else { + const uint8_t *vfilter = bilinear_filters_2t[yoffset]; + const __m128i vfilter_vec = _mm_set1_epi16(vfilter[0] | (vfilter[1] << 8)); + for (i = 0; i < h; i += 4) { + const __m128i a = xx_loadl_32((__m128i *)dst); + const __m128i b = xx_loadl_32((__m128i *)&dst[4]); + const __m128i c = xx_loadl_32((__m128i *)&dst[8]); + const __m128i d = xx_loadl_32((__m128i *)&dst[12]); + const __m128i e = xx_loadl_32((__m128i *)&dst[16]); + + const __m128i a0 = _mm_unpacklo_epi32(a, b); + const __m128i b0 = _mm_unpacklo_epi32(b, c); + const __m128i a1 = _mm_unpacklo_epi32(c, d); + const __m128i b1 = _mm_unpacklo_epi32(d, e); + const __m128i res = filter_block_2rows(&a0, &b0, &a1, &b1, &vfilter_vec); + _mm_storeu_si128((__m128i *)dst, res); + + dst += 16; + } + } +} + +static INLINE void accumulate_block(const __m128i *src, const __m128i *a, + const __m128i *b, const __m128i *m, + __m128i *sum, __m128i *sum_sq) { + const __m128i zero = _mm_setzero_si128(); + const __m128i one = _mm_set1_epi16(1); + const __m128i mask_max = _mm_set1_epi8((1 << AOM_BLEND_A64_ROUND_BITS)); + const __m128i m_inv = _mm_sub_epi8(mask_max, *m); + + // Calculate 16 predicted pixels. + // Note that the maximum value of any entry of 'pred_l' or 'pred_r' + // is 64 * 255, so we have plenty of space to add rounding constants. + const __m128i data_l = _mm_unpacklo_epi8(*a, *b); + const __m128i mask_l = _mm_unpacklo_epi8(*m, m_inv); + __m128i pred_l = _mm_maddubs_epi16(data_l, mask_l); + pred_l = xx_roundn_epu16(pred_l, AOM_BLEND_A64_ROUND_BITS); + + const __m128i data_r = _mm_unpackhi_epi8(*a, *b); + const __m128i mask_r = _mm_unpackhi_epi8(*m, m_inv); + __m128i pred_r = _mm_maddubs_epi16(data_r, mask_r); + pred_r = xx_roundn_epu16(pred_r, AOM_BLEND_A64_ROUND_BITS); + + const __m128i src_l = _mm_unpacklo_epi8(*src, zero); + const __m128i src_r = _mm_unpackhi_epi8(*src, zero); + const __m128i diff_l = _mm_sub_epi16(pred_l, src_l); + const __m128i diff_r = _mm_sub_epi16(pred_r, src_r); + + // Update partial sums and partial sums of squares + *sum = + _mm_add_epi32(*sum, _mm_madd_epi16(_mm_add_epi16(diff_l, diff_r), one)); + *sum_sq = + _mm_add_epi32(*sum_sq, _mm_add_epi32(_mm_madd_epi16(diff_l, diff_l), + _mm_madd_epi16(diff_r, diff_r))); +} + +static void masked_variance(const uint8_t *src_ptr, int src_stride, + const uint8_t *a_ptr, int a_stride, + const uint8_t *b_ptr, int b_stride, + const uint8_t *m_ptr, int m_stride, int width, + int height, unsigned int *sse, int *sum_) { + int x, y; + __m128i sum = _mm_setzero_si128(), sum_sq = _mm_setzero_si128(); + + for (y = 0; y < height; y++) { + for (x = 0; x < width; x += 16) { + const __m128i src = _mm_loadu_si128((const __m128i *)&src_ptr[x]); + const __m128i a = _mm_loadu_si128((const __m128i *)&a_ptr[x]); + const __m128i b = _mm_loadu_si128((const __m128i *)&b_ptr[x]); + const __m128i m = _mm_loadu_si128((const __m128i *)&m_ptr[x]); + accumulate_block(&src, &a, &b, &m, &sum, &sum_sq); + } + + src_ptr += src_stride; + a_ptr += a_stride; + b_ptr += b_stride; + m_ptr += m_stride; + } + // Reduce down to a single sum and sum of squares + sum = _mm_hadd_epi32(sum, sum_sq); + sum = _mm_hadd_epi32(sum, sum); + *sum_ = _mm_cvtsi128_si32(sum); + *sse = _mm_cvtsi128_si32(_mm_srli_si128(sum, 4)); +} + +static void masked_variance8xh(const uint8_t *src_ptr, int src_stride, + const uint8_t *a_ptr, const uint8_t *b_ptr, + const uint8_t *m_ptr, int m_stride, int height, + unsigned int *sse, int *sum_) { + int y; + __m128i sum = _mm_setzero_si128(), sum_sq = _mm_setzero_si128(); + + for (y = 0; y < height; y += 2) { + __m128i src = _mm_unpacklo_epi64( + _mm_loadl_epi64((const __m128i *)src_ptr), + _mm_loadl_epi64((const __m128i *)&src_ptr[src_stride])); + const __m128i a = _mm_loadu_si128((const __m128i *)a_ptr); + const __m128i b = _mm_loadu_si128((const __m128i *)b_ptr); + const __m128i m = + _mm_unpacklo_epi64(_mm_loadl_epi64((const __m128i *)m_ptr), + _mm_loadl_epi64((const __m128i *)&m_ptr[m_stride])); + accumulate_block(&src, &a, &b, &m, &sum, &sum_sq); + + src_ptr += src_stride * 2; + a_ptr += 16; + b_ptr += 16; + m_ptr += m_stride * 2; + } + // Reduce down to a single sum and sum of squares + sum = _mm_hadd_epi32(sum, sum_sq); + sum = _mm_hadd_epi32(sum, sum); + *sum_ = _mm_cvtsi128_si32(sum); + *sse = _mm_cvtsi128_si32(_mm_srli_si128(sum, 4)); +} + +static void masked_variance4xh(const uint8_t *src_ptr, int src_stride, + const uint8_t *a_ptr, const uint8_t *b_ptr, + const uint8_t *m_ptr, int m_stride, int height, + unsigned int *sse, int *sum_) { + int y; + __m128i sum = _mm_setzero_si128(), sum_sq = _mm_setzero_si128(); + + for (y = 0; y < height; y += 4) { + // Load four rows at a time + __m128i src = + _mm_setr_epi32(*(uint32_t *)src_ptr, *(uint32_t *)&src_ptr[src_stride], + *(uint32_t *)&src_ptr[src_stride * 2], + *(uint32_t *)&src_ptr[src_stride * 3]); + const __m128i a = _mm_loadu_si128((const __m128i *)a_ptr); + const __m128i b = _mm_loadu_si128((const __m128i *)b_ptr); + const __m128i m = _mm_setr_epi32( + *(uint32_t *)m_ptr, *(uint32_t *)&m_ptr[m_stride], + *(uint32_t *)&m_ptr[m_stride * 2], *(uint32_t *)&m_ptr[m_stride * 3]); + accumulate_block(&src, &a, &b, &m, &sum, &sum_sq); + + src_ptr += src_stride * 4; + a_ptr += 16; + b_ptr += 16; + m_ptr += m_stride * 4; + } + // Reduce down to a single sum and sum of squares + sum = _mm_hadd_epi32(sum, sum_sq); + sum = _mm_hadd_epi32(sum, sum); + *sum_ = _mm_cvtsi128_si32(sum); + *sse = _mm_cvtsi128_si32(_mm_srli_si128(sum, 4)); +} + +// For width a multiple of 8 +static void highbd_bilinear_filter(const uint16_t *src, int src_stride, + int xoffset, int yoffset, uint16_t *dst, + int w, int h); + +static void highbd_bilinear_filter4xh(const uint16_t *src, int src_stride, + int xoffset, int yoffset, uint16_t *dst, + int h); + +// For width a multiple of 8 +static void highbd_masked_variance(const uint16_t *src_ptr, int src_stride, + const uint16_t *a_ptr, int a_stride, + const uint16_t *b_ptr, int b_stride, + const uint8_t *m_ptr, int m_stride, + int width, int height, uint64_t *sse, + int *sum_); + +static void highbd_masked_variance4xh(const uint16_t *src_ptr, int src_stride, + const uint16_t *a_ptr, + const uint16_t *b_ptr, + const uint8_t *m_ptr, int m_stride, + int height, int *sse, int *sum_); + +#define HIGHBD_MASK_SUBPIX_VAR_SSSE3(W, H) \ + unsigned int aom_highbd_8_masked_sub_pixel_variance##W##x##H##_ssse3( \ + const uint8_t *src8, int src_stride, int xoffset, int yoffset, \ + const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, \ + const uint8_t *msk, int msk_stride, int invert_mask, uint32_t *sse) { \ + uint64_t sse64; \ + int sum; \ + uint16_t temp[(H + 1) * W]; \ + const uint16_t *src = CONVERT_TO_SHORTPTR(src8); \ + const uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); \ + const uint16_t *second_pred = CONVERT_TO_SHORTPTR(second_pred8); \ + \ + highbd_bilinear_filter(src, src_stride, xoffset, yoffset, temp, W, H); \ + \ + if (!invert_mask) \ + highbd_masked_variance(ref, ref_stride, temp, W, second_pred, W, msk, \ + msk_stride, W, H, &sse64, &sum); \ + else \ + highbd_masked_variance(ref, ref_stride, second_pred, W, temp, W, msk, \ + msk_stride, W, H, &sse64, &sum); \ + *sse = (uint32_t)sse64; \ + return *sse - (uint32_t)(((int64_t)sum * sum) / (W * H)); \ + } \ + unsigned int aom_highbd_10_masked_sub_pixel_variance##W##x##H##_ssse3( \ + const uint8_t *src8, int src_stride, int xoffset, int yoffset, \ + const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, \ + const uint8_t *msk, int msk_stride, int invert_mask, uint32_t *sse) { \ + uint64_t sse64; \ + int sum; \ + int64_t var; \ + uint16_t temp[(H + 1) * W]; \ + const uint16_t *src = CONVERT_TO_SHORTPTR(src8); \ + const uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); \ + const uint16_t *second_pred = CONVERT_TO_SHORTPTR(second_pred8); \ + \ + highbd_bilinear_filter(src, src_stride, xoffset, yoffset, temp, W, H); \ + \ + if (!invert_mask) \ + highbd_masked_variance(ref, ref_stride, temp, W, second_pred, W, msk, \ + msk_stride, W, H, &sse64, &sum); \ + else \ + highbd_masked_variance(ref, ref_stride, second_pred, W, temp, W, msk, \ + msk_stride, W, H, &sse64, &sum); \ + *sse = (uint32_t)ROUND_POWER_OF_TWO(sse64, 4); \ + sum = ROUND_POWER_OF_TWO(sum, 2); \ + var = (int64_t)(*sse) - (((int64_t)sum * sum) / (W * H)); \ + return (var >= 0) ? (uint32_t)var : 0; \ + } \ + unsigned int aom_highbd_12_masked_sub_pixel_variance##W##x##H##_ssse3( \ + const uint8_t *src8, int src_stride, int xoffset, int yoffset, \ + const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, \ + const uint8_t *msk, int msk_stride, int invert_mask, uint32_t *sse) { \ + uint64_t sse64; \ + int sum; \ + int64_t var; \ + uint16_t temp[(H + 1) * W]; \ + const uint16_t *src = CONVERT_TO_SHORTPTR(src8); \ + const uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); \ + const uint16_t *second_pred = CONVERT_TO_SHORTPTR(second_pred8); \ + \ + highbd_bilinear_filter(src, src_stride, xoffset, yoffset, temp, W, H); \ + \ + if (!invert_mask) \ + highbd_masked_variance(ref, ref_stride, temp, W, second_pred, W, msk, \ + msk_stride, W, H, &sse64, &sum); \ + else \ + highbd_masked_variance(ref, ref_stride, second_pred, W, temp, W, msk, \ + msk_stride, W, H, &sse64, &sum); \ + *sse = (uint32_t)ROUND_POWER_OF_TWO(sse64, 8); \ + sum = ROUND_POWER_OF_TWO(sum, 4); \ + var = (int64_t)(*sse) - (((int64_t)sum * sum) / (W * H)); \ + return (var >= 0) ? (uint32_t)var : 0; \ + } + +#define HIGHBD_MASK_SUBPIX_VAR4XH_SSSE3(H) \ + unsigned int aom_highbd_8_masked_sub_pixel_variance4x##H##_ssse3( \ + const uint8_t *src8, int src_stride, int xoffset, int yoffset, \ + const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, \ + const uint8_t *msk, int msk_stride, int invert_mask, uint32_t *sse) { \ + int sse_; \ + int sum; \ + uint16_t temp[(H + 1) * 4]; \ + const uint16_t *src = CONVERT_TO_SHORTPTR(src8); \ + const uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); \ + const uint16_t *second_pred = CONVERT_TO_SHORTPTR(second_pred8); \ + \ + highbd_bilinear_filter4xh(src, src_stride, xoffset, yoffset, temp, H); \ + \ + if (!invert_mask) \ + highbd_masked_variance4xh(ref, ref_stride, temp, second_pred, msk, \ + msk_stride, H, &sse_, &sum); \ + else \ + highbd_masked_variance4xh(ref, ref_stride, second_pred, temp, msk, \ + msk_stride, H, &sse_, &sum); \ + *sse = (uint32_t)sse_; \ + return *sse - (uint32_t)(((int64_t)sum * sum) / (4 * H)); \ + } \ + unsigned int aom_highbd_10_masked_sub_pixel_variance4x##H##_ssse3( \ + const uint8_t *src8, int src_stride, int xoffset, int yoffset, \ + const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, \ + const uint8_t *msk, int msk_stride, int invert_mask, uint32_t *sse) { \ + int sse_; \ + int sum; \ + int64_t var; \ + uint16_t temp[(H + 1) * 4]; \ + const uint16_t *src = CONVERT_TO_SHORTPTR(src8); \ + const uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); \ + const uint16_t *second_pred = CONVERT_TO_SHORTPTR(second_pred8); \ + \ + highbd_bilinear_filter4xh(src, src_stride, xoffset, yoffset, temp, H); \ + \ + if (!invert_mask) \ + highbd_masked_variance4xh(ref, ref_stride, temp, second_pred, msk, \ + msk_stride, H, &sse_, &sum); \ + else \ + highbd_masked_variance4xh(ref, ref_stride, second_pred, temp, msk, \ + msk_stride, H, &sse_, &sum); \ + *sse = (uint32_t)ROUND_POWER_OF_TWO(sse_, 4); \ + sum = ROUND_POWER_OF_TWO(sum, 2); \ + var = (int64_t)(*sse) - (((int64_t)sum * sum) / (4 * H)); \ + return (var >= 0) ? (uint32_t)var : 0; \ + } \ + unsigned int aom_highbd_12_masked_sub_pixel_variance4x##H##_ssse3( \ + const uint8_t *src8, int src_stride, int xoffset, int yoffset, \ + const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, \ + const uint8_t *msk, int msk_stride, int invert_mask, uint32_t *sse) { \ + int sse_; \ + int sum; \ + int64_t var; \ + uint16_t temp[(H + 1) * 4]; \ + const uint16_t *src = CONVERT_TO_SHORTPTR(src8); \ + const uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); \ + const uint16_t *second_pred = CONVERT_TO_SHORTPTR(second_pred8); \ + \ + highbd_bilinear_filter4xh(src, src_stride, xoffset, yoffset, temp, H); \ + \ + if (!invert_mask) \ + highbd_masked_variance4xh(ref, ref_stride, temp, second_pred, msk, \ + msk_stride, H, &sse_, &sum); \ + else \ + highbd_masked_variance4xh(ref, ref_stride, second_pred, temp, msk, \ + msk_stride, H, &sse_, &sum); \ + *sse = (uint32_t)ROUND_POWER_OF_TWO(sse_, 8); \ + sum = ROUND_POWER_OF_TWO(sum, 4); \ + var = (int64_t)(*sse) - (((int64_t)sum * sum) / (4 * H)); \ + return (var >= 0) ? (uint32_t)var : 0; \ + } + +HIGHBD_MASK_SUBPIX_VAR_SSSE3(128, 128) +HIGHBD_MASK_SUBPIX_VAR_SSSE3(128, 64) +HIGHBD_MASK_SUBPIX_VAR_SSSE3(64, 128) +HIGHBD_MASK_SUBPIX_VAR_SSSE3(64, 64) +HIGHBD_MASK_SUBPIX_VAR_SSSE3(64, 32) +HIGHBD_MASK_SUBPIX_VAR_SSSE3(32, 64) +HIGHBD_MASK_SUBPIX_VAR_SSSE3(32, 32) +HIGHBD_MASK_SUBPIX_VAR_SSSE3(32, 16) +HIGHBD_MASK_SUBPIX_VAR_SSSE3(16, 32) +HIGHBD_MASK_SUBPIX_VAR_SSSE3(16, 16) +HIGHBD_MASK_SUBPIX_VAR_SSSE3(16, 8) +HIGHBD_MASK_SUBPIX_VAR_SSSE3(8, 16) +HIGHBD_MASK_SUBPIX_VAR_SSSE3(8, 8) +HIGHBD_MASK_SUBPIX_VAR_SSSE3(8, 4) +HIGHBD_MASK_SUBPIX_VAR4XH_SSSE3(8) +HIGHBD_MASK_SUBPIX_VAR4XH_SSSE3(4) +HIGHBD_MASK_SUBPIX_VAR4XH_SSSE3(16) +HIGHBD_MASK_SUBPIX_VAR_SSSE3(16, 4) +HIGHBD_MASK_SUBPIX_VAR_SSSE3(8, 32) +HIGHBD_MASK_SUBPIX_VAR_SSSE3(32, 8) +HIGHBD_MASK_SUBPIX_VAR_SSSE3(16, 64) +HIGHBD_MASK_SUBPIX_VAR_SSSE3(64, 16) + +static INLINE __m128i highbd_filter_block(const __m128i a, const __m128i b, + const __m128i filter) { + __m128i v0 = _mm_unpacklo_epi16(a, b); + v0 = _mm_madd_epi16(v0, filter); + v0 = xx_roundn_epu32(v0, FILTER_BITS); + + __m128i v1 = _mm_unpackhi_epi16(a, b); + v1 = _mm_madd_epi16(v1, filter); + v1 = xx_roundn_epu32(v1, FILTER_BITS); + + return _mm_packs_epi32(v0, v1); +} + +static void highbd_bilinear_filter(const uint16_t *src, int src_stride, + int xoffset, int yoffset, uint16_t *dst, + int w, int h) { + int i, j; + // Horizontal filter + if (xoffset == 0) { + uint16_t *b = dst; + for (i = 0; i < h + 1; ++i) { + for (j = 0; j < w; j += 8) { + __m128i x = _mm_loadu_si128((__m128i *)&src[j]); + _mm_storeu_si128((__m128i *)&b[j], x); + } + src += src_stride; + b += w; + } + } else if (xoffset == 4) { + uint16_t *b = dst; + for (i = 0; i < h + 1; ++i) { + for (j = 0; j < w; j += 8) { + __m128i x = _mm_loadu_si128((__m128i *)&src[j]); + __m128i y = _mm_loadu_si128((__m128i *)&src[j + 8]); + __m128i z = _mm_alignr_epi8(y, x, 2); + _mm_storeu_si128((__m128i *)&b[j], _mm_avg_epu16(x, z)); + } + src += src_stride; + b += w; + } + } else { + uint16_t *b = dst; + const uint8_t *hfilter = bilinear_filters_2t[xoffset]; + const __m128i hfilter_vec = _mm_set1_epi32(hfilter[0] | (hfilter[1] << 16)); + for (i = 0; i < h + 1; ++i) { + for (j = 0; j < w; j += 8) { + const __m128i x = _mm_loadu_si128((__m128i *)&src[j]); + const __m128i y = _mm_loadu_si128((__m128i *)&src[j + 8]); + const __m128i z = _mm_alignr_epi8(y, x, 2); + const __m128i res = highbd_filter_block(x, z, hfilter_vec); + _mm_storeu_si128((__m128i *)&b[j], res); + } + + src += src_stride; + b += w; + } + } + + // Vertical filter + if (yoffset == 0) { + // The data is already in 'dst', so no need to filter + } else if (yoffset == 4) { + for (i = 0; i < h; ++i) { + for (j = 0; j < w; j += 8) { + __m128i x = _mm_loadu_si128((__m128i *)&dst[j]); + __m128i y = _mm_loadu_si128((__m128i *)&dst[j + w]); + _mm_storeu_si128((__m128i *)&dst[j], _mm_avg_epu16(x, y)); + } + dst += w; + } + } else { + const uint8_t *vfilter = bilinear_filters_2t[yoffset]; + const __m128i vfilter_vec = _mm_set1_epi32(vfilter[0] | (vfilter[1] << 16)); + for (i = 0; i < h; ++i) { + for (j = 0; j < w; j += 8) { + const __m128i x = _mm_loadu_si128((__m128i *)&dst[j]); + const __m128i y = _mm_loadu_si128((__m128i *)&dst[j + w]); + const __m128i res = highbd_filter_block(x, y, vfilter_vec); + _mm_storeu_si128((__m128i *)&dst[j], res); + } + + dst += w; + } + } +} + +static INLINE __m128i highbd_filter_block_2rows(const __m128i *a0, + const __m128i *b0, + const __m128i *a1, + const __m128i *b1, + const __m128i *filter) { + __m128i v0 = _mm_unpacklo_epi16(*a0, *b0); + v0 = _mm_madd_epi16(v0, *filter); + v0 = xx_roundn_epu32(v0, FILTER_BITS); + + __m128i v1 = _mm_unpacklo_epi16(*a1, *b1); + v1 = _mm_madd_epi16(v1, *filter); + v1 = xx_roundn_epu32(v1, FILTER_BITS); + + return _mm_packs_epi32(v0, v1); +} + +static void highbd_bilinear_filter4xh(const uint16_t *src, int src_stride, + int xoffset, int yoffset, uint16_t *dst, + int h) { + int i; + // Horizontal filter + if (xoffset == 0) { + uint16_t *b = dst; + for (i = 0; i < h + 1; ++i) { + __m128i x = _mm_loadl_epi64((__m128i *)src); + _mm_storel_epi64((__m128i *)b, x); + src += src_stride; + b += 4; + } + } else if (xoffset == 4) { + uint16_t *b = dst; + for (i = 0; i < h + 1; ++i) { + __m128i x = _mm_loadu_si128((__m128i *)src); + __m128i z = _mm_srli_si128(x, 2); + _mm_storel_epi64((__m128i *)b, _mm_avg_epu16(x, z)); + src += src_stride; + b += 4; + } + } else { + uint16_t *b = dst; + const uint8_t *hfilter = bilinear_filters_2t[xoffset]; + const __m128i hfilter_vec = _mm_set1_epi32(hfilter[0] | (hfilter[1] << 16)); + for (i = 0; i < h; i += 2) { + const __m128i x0 = _mm_loadu_si128((__m128i *)src); + const __m128i z0 = _mm_srli_si128(x0, 2); + const __m128i x1 = _mm_loadu_si128((__m128i *)&src[src_stride]); + const __m128i z1 = _mm_srli_si128(x1, 2); + const __m128i res = + highbd_filter_block_2rows(&x0, &z0, &x1, &z1, &hfilter_vec); + _mm_storeu_si128((__m128i *)b, res); + + src += src_stride * 2; + b += 8; + } + // Process i = h separately + __m128i x = _mm_loadu_si128((__m128i *)src); + __m128i z = _mm_srli_si128(x, 2); + + __m128i v0 = _mm_unpacklo_epi16(x, z); + v0 = _mm_madd_epi16(v0, hfilter_vec); + v0 = xx_roundn_epu32(v0, FILTER_BITS); + + _mm_storel_epi64((__m128i *)b, _mm_packs_epi32(v0, v0)); + } + + // Vertical filter + if (yoffset == 0) { + // The data is already in 'dst', so no need to filter + } else if (yoffset == 4) { + for (i = 0; i < h; ++i) { + __m128i x = _mm_loadl_epi64((__m128i *)dst); + __m128i y = _mm_loadl_epi64((__m128i *)&dst[4]); + _mm_storel_epi64((__m128i *)dst, _mm_avg_epu16(x, y)); + dst += 4; + } + } else { + const uint8_t *vfilter = bilinear_filters_2t[yoffset]; + const __m128i vfilter_vec = _mm_set1_epi32(vfilter[0] | (vfilter[1] << 16)); + for (i = 0; i < h; i += 2) { + const __m128i x = _mm_loadl_epi64((__m128i *)dst); + const __m128i y = _mm_loadl_epi64((__m128i *)&dst[4]); + const __m128i z = _mm_loadl_epi64((__m128i *)&dst[8]); + const __m128i res = + highbd_filter_block_2rows(&x, &y, &y, &z, &vfilter_vec); + _mm_storeu_si128((__m128i *)dst, res); + + dst += 8; + } + } +} + +static void highbd_masked_variance(const uint16_t *src_ptr, int src_stride, + const uint16_t *a_ptr, int a_stride, + const uint16_t *b_ptr, int b_stride, + const uint8_t *m_ptr, int m_stride, + int width, int height, uint64_t *sse, + int *sum_) { + int x, y; + // Note on bit widths: + // The maximum value of 'sum' is (2^12 - 1) * 128 * 128 =~ 2^26, + // so this can be kept as four 32-bit values. + // But the maximum value of 'sum_sq' is (2^12 - 1)^2 * 128 * 128 =~ 2^38, + // so this must be stored as two 64-bit values. + __m128i sum = _mm_setzero_si128(), sum_sq = _mm_setzero_si128(); + const __m128i mask_max = _mm_set1_epi16((1 << AOM_BLEND_A64_ROUND_BITS)); + const __m128i round_const = + _mm_set1_epi32((1 << AOM_BLEND_A64_ROUND_BITS) >> 1); + const __m128i zero = _mm_setzero_si128(); + + for (y = 0; y < height; y++) { + for (x = 0; x < width; x += 8) { + const __m128i src = _mm_loadu_si128((const __m128i *)&src_ptr[x]); + const __m128i a = _mm_loadu_si128((const __m128i *)&a_ptr[x]); + const __m128i b = _mm_loadu_si128((const __m128i *)&b_ptr[x]); + const __m128i m = + _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)&m_ptr[x]), zero); + const __m128i m_inv = _mm_sub_epi16(mask_max, m); + + // Calculate 8 predicted pixels. + const __m128i data_l = _mm_unpacklo_epi16(a, b); + const __m128i mask_l = _mm_unpacklo_epi16(m, m_inv); + __m128i pred_l = _mm_madd_epi16(data_l, mask_l); + pred_l = _mm_srai_epi32(_mm_add_epi32(pred_l, round_const), + AOM_BLEND_A64_ROUND_BITS); + + const __m128i data_r = _mm_unpackhi_epi16(a, b); + const __m128i mask_r = _mm_unpackhi_epi16(m, m_inv); + __m128i pred_r = _mm_madd_epi16(data_r, mask_r); + pred_r = _mm_srai_epi32(_mm_add_epi32(pred_r, round_const), + AOM_BLEND_A64_ROUND_BITS); + + const __m128i src_l = _mm_unpacklo_epi16(src, zero); + const __m128i src_r = _mm_unpackhi_epi16(src, zero); + __m128i diff_l = _mm_sub_epi32(pred_l, src_l); + __m128i diff_r = _mm_sub_epi32(pred_r, src_r); + + // Update partial sums and partial sums of squares + sum = _mm_add_epi32(sum, _mm_add_epi32(diff_l, diff_r)); + // A trick: Now each entry of diff_l and diff_r is stored in a 32-bit + // field, but the range of values is only [-(2^12 - 1), 2^12 - 1]. + // So we can re-pack into 16-bit fields and use _mm_madd_epi16 + // to calculate the squares and partially sum them. + const __m128i tmp = _mm_packs_epi32(diff_l, diff_r); + const __m128i prod = _mm_madd_epi16(tmp, tmp); + // Then we want to sign-extend to 64 bits and accumulate + const __m128i sign = _mm_srai_epi32(prod, 31); + const __m128i tmp_0 = _mm_unpacklo_epi32(prod, sign); + const __m128i tmp_1 = _mm_unpackhi_epi32(prod, sign); + sum_sq = _mm_add_epi64(sum_sq, _mm_add_epi64(tmp_0, tmp_1)); + } + + src_ptr += src_stride; + a_ptr += a_stride; + b_ptr += b_stride; + m_ptr += m_stride; + } + // Reduce down to a single sum and sum of squares + sum = _mm_hadd_epi32(sum, zero); + sum = _mm_hadd_epi32(sum, zero); + *sum_ = _mm_cvtsi128_si32(sum); + sum_sq = _mm_add_epi64(sum_sq, _mm_srli_si128(sum_sq, 8)); + _mm_storel_epi64((__m128i *)sse, sum_sq); +} + +static void highbd_masked_variance4xh(const uint16_t *src_ptr, int src_stride, + const uint16_t *a_ptr, + const uint16_t *b_ptr, + const uint8_t *m_ptr, int m_stride, + int height, int *sse, int *sum_) { + int y; + // Note: For this function, h <= 8 (or maybe 16 if we add 4:1 partitions). + // So the maximum value of sum is (2^12 - 1) * 4 * 16 =~ 2^18 + // and the maximum value of sum_sq is (2^12 - 1)^2 * 4 * 16 =~ 2^30. + // So we can safely pack sum_sq into 32-bit fields, which is slightly more + // convenient. + __m128i sum = _mm_setzero_si128(), sum_sq = _mm_setzero_si128(); + const __m128i mask_max = _mm_set1_epi16((1 << AOM_BLEND_A64_ROUND_BITS)); + const __m128i round_const = + _mm_set1_epi32((1 << AOM_BLEND_A64_ROUND_BITS) >> 1); + const __m128i zero = _mm_setzero_si128(); + + for (y = 0; y < height; y += 2) { + __m128i src = _mm_unpacklo_epi64( + _mm_loadl_epi64((const __m128i *)src_ptr), + _mm_loadl_epi64((const __m128i *)&src_ptr[src_stride])); + const __m128i a = _mm_loadu_si128((const __m128i *)a_ptr); + const __m128i b = _mm_loadu_si128((const __m128i *)b_ptr); + const __m128i m = _mm_unpacklo_epi8( + _mm_unpacklo_epi32( + _mm_cvtsi32_si128(*(const uint32_t *)m_ptr), + _mm_cvtsi32_si128(*(const uint32_t *)&m_ptr[m_stride])), + zero); + const __m128i m_inv = _mm_sub_epi16(mask_max, m); + + const __m128i data_l = _mm_unpacklo_epi16(a, b); + const __m128i mask_l = _mm_unpacklo_epi16(m, m_inv); + __m128i pred_l = _mm_madd_epi16(data_l, mask_l); + pred_l = _mm_srai_epi32(_mm_add_epi32(pred_l, round_const), + AOM_BLEND_A64_ROUND_BITS); + + const __m128i data_r = _mm_unpackhi_epi16(a, b); + const __m128i mask_r = _mm_unpackhi_epi16(m, m_inv); + __m128i pred_r = _mm_madd_epi16(data_r, mask_r); + pred_r = _mm_srai_epi32(_mm_add_epi32(pred_r, round_const), + AOM_BLEND_A64_ROUND_BITS); + + const __m128i src_l = _mm_unpacklo_epi16(src, zero); + const __m128i src_r = _mm_unpackhi_epi16(src, zero); + __m128i diff_l = _mm_sub_epi32(pred_l, src_l); + __m128i diff_r = _mm_sub_epi32(pred_r, src_r); + + // Update partial sums and partial sums of squares + sum = _mm_add_epi32(sum, _mm_add_epi32(diff_l, diff_r)); + const __m128i tmp = _mm_packs_epi32(diff_l, diff_r); + const __m128i prod = _mm_madd_epi16(tmp, tmp); + sum_sq = _mm_add_epi32(sum_sq, prod); + + src_ptr += src_stride * 2; + a_ptr += 8; + b_ptr += 8; + m_ptr += m_stride * 2; + } + // Reduce down to a single sum and sum of squares + sum = _mm_hadd_epi32(sum, sum_sq); + sum = _mm_hadd_epi32(sum, zero); + *sum_ = _mm_cvtsi128_si32(sum); + *sse = _mm_cvtsi128_si32(_mm_srli_si128(sum, 4)); +} + +void aom_comp_mask_pred_ssse3(uint8_t *comp_pred, const uint8_t *pred, + int width, int height, const uint8_t *ref, + int ref_stride, const uint8_t *mask, + int mask_stride, int invert_mask) { + const uint8_t *src0 = invert_mask ? pred : ref; + const uint8_t *src1 = invert_mask ? ref : pred; + const int stride0 = invert_mask ? width : ref_stride; + const int stride1 = invert_mask ? ref_stride : width; + assert(height % 2 == 0); + int i = 0; + if (width == 8) { + comp_mask_pred_8_ssse3(comp_pred, height, src0, stride0, src1, stride1, + mask, mask_stride); + } else if (width == 16) { + do { + comp_mask_pred_16_ssse3(src0, src1, mask, comp_pred); + comp_mask_pred_16_ssse3(src0 + stride0, src1 + stride1, + mask + mask_stride, comp_pred + width); + comp_pred += (width << 1); + src0 += (stride0 << 1); + src1 += (stride1 << 1); + mask += (mask_stride << 1); + i += 2; + } while (i < height); + } else { // width == 32 + assert(width == 32); + do { + comp_mask_pred_16_ssse3(src0, src1, mask, comp_pred); + comp_mask_pred_16_ssse3(src0 + 16, src1 + 16, mask + 16, comp_pred + 16); + comp_pred += (width); + src0 += (stride0); + src1 += (stride1); + mask += (mask_stride); + i += 1; + } while (i < height); + } +} diff --git a/third_party/aom/aom_dsp/x86/masked_variance_intrin_ssse3.h b/third_party/aom/aom_dsp/x86/masked_variance_intrin_ssse3.h new file mode 100644 index 0000000000..4faa098ace --- /dev/null +++ b/third_party/aom/aom_dsp/x86/masked_variance_intrin_ssse3.h @@ -0,0 +1,92 @@ +/* + * Copyright (c) 2018, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_X86_MASKED_VARIANCE_INTRIN_SSSE3_H_ +#define AOM_AOM_DSP_X86_MASKED_VARIANCE_INTRIN_SSSE3_H_ + +#include +#include +#include + +#include "config/aom_config.h" +#include "config/aom_dsp_rtcd.h" + +#include "aom_dsp/blend.h" + +static INLINE void comp_mask_pred_16_ssse3(const uint8_t *src0, + const uint8_t *src1, + const uint8_t *mask, uint8_t *dst) { + const __m128i alpha_max = _mm_set1_epi8(AOM_BLEND_A64_MAX_ALPHA); + const __m128i round_offset = + _mm_set1_epi16(1 << (15 - AOM_BLEND_A64_ROUND_BITS)); + + const __m128i sA0 = _mm_lddqu_si128((const __m128i *)(src0)); + const __m128i sA1 = _mm_lddqu_si128((const __m128i *)(src1)); + const __m128i aA = _mm_load_si128((const __m128i *)(mask)); + + const __m128i maA = _mm_sub_epi8(alpha_max, aA); + + const __m128i ssAL = _mm_unpacklo_epi8(sA0, sA1); + const __m128i aaAL = _mm_unpacklo_epi8(aA, maA); + const __m128i ssAH = _mm_unpackhi_epi8(sA0, sA1); + const __m128i aaAH = _mm_unpackhi_epi8(aA, maA); + + const __m128i blendAL = _mm_maddubs_epi16(ssAL, aaAL); + const __m128i blendAH = _mm_maddubs_epi16(ssAH, aaAH); + + const __m128i roundAL = _mm_mulhrs_epi16(blendAL, round_offset); + const __m128i roundAH = _mm_mulhrs_epi16(blendAH, round_offset); + _mm_store_si128((__m128i *)dst, _mm_packus_epi16(roundAL, roundAH)); +} + +static INLINE void comp_mask_pred_8_ssse3(uint8_t *comp_pred, int height, + const uint8_t *src0, int stride0, + const uint8_t *src1, int stride1, + const uint8_t *mask, + int mask_stride) { + int i = 0; + const __m128i alpha_max = _mm_set1_epi8(AOM_BLEND_A64_MAX_ALPHA); + const __m128i round_offset = + _mm_set1_epi16(1 << (15 - AOM_BLEND_A64_ROUND_BITS)); + do { + // odd line A + const __m128i sA0 = _mm_loadl_epi64((const __m128i *)(src0)); + const __m128i sA1 = _mm_loadl_epi64((const __m128i *)(src1)); + const __m128i aA = _mm_loadl_epi64((const __m128i *)(mask)); + // even line B + const __m128i sB0 = _mm_loadl_epi64((const __m128i *)(src0 + stride0)); + const __m128i sB1 = _mm_loadl_epi64((const __m128i *)(src1 + stride1)); + const __m128i a = _mm_castps_si128(_mm_loadh_pi( + _mm_castsi128_ps(aA), (const __m64 *)(mask + mask_stride))); + + const __m128i ssA = _mm_unpacklo_epi8(sA0, sA1); + const __m128i ssB = _mm_unpacklo_epi8(sB0, sB1); + + const __m128i ma = _mm_sub_epi8(alpha_max, a); + const __m128i aaA = _mm_unpacklo_epi8(a, ma); + const __m128i aaB = _mm_unpackhi_epi8(a, ma); + + const __m128i blendA = _mm_maddubs_epi16(ssA, aaA); + const __m128i blendB = _mm_maddubs_epi16(ssB, aaB); + const __m128i roundA = _mm_mulhrs_epi16(blendA, round_offset); + const __m128i roundB = _mm_mulhrs_epi16(blendB, round_offset); + const __m128i round = _mm_packus_epi16(roundA, roundB); + // comp_pred's stride == width == 8 + _mm_store_si128((__m128i *)(comp_pred), round); + comp_pred += (8 << 1); + src0 += (stride0 << 1); + src1 += (stride1 << 1); + mask += (mask_stride << 1); + i += 2; + } while (i < height); +} + +#endif // AOM_AOM_DSP_X86_MASKED_VARIANCE_INTRIN_SSSE3_H_ diff --git a/third_party/aom/aom_dsp/x86/mem_sse2.h b/third_party/aom/aom_dsp/x86/mem_sse2.h new file mode 100644 index 0000000000..6c821673e8 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/mem_sse2.h @@ -0,0 +1,42 @@ +/* + * Copyright (c) 2017, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_X86_MEM_SSE2_H_ +#define AOM_AOM_DSP_X86_MEM_SSE2_H_ + +#include // SSE2 + +#include "config/aom_config.h" + +#include "aom/aom_integer.h" + +static INLINE __m128i loadh_epi64(const void *const src, const __m128i s) { + return _mm_castps_si128( + _mm_loadh_pi(_mm_castsi128_ps(s), (const __m64 *)src)); +} + +static INLINE __m128i load_8bit_4x4_to_1_reg_sse2(const void *const src, + const int byte_stride) { + return _mm_setr_epi32(*(const int32_t *)((int8_t *)src + 0 * byte_stride), + *(const int32_t *)((int8_t *)src + 1 * byte_stride), + *(const int32_t *)((int8_t *)src + 2 * byte_stride), + *(const int32_t *)((int8_t *)src + 3 * byte_stride)); +} + +static INLINE __m128i load_8bit_8x2_to_1_reg_sse2(const void *const src, + const int byte_stride) { + __m128i dst; + dst = _mm_loadl_epi64((__m128i *)((int8_t *)src + 0 * byte_stride)); + dst = loadh_epi64((int8_t *)src + 1 * byte_stride, dst); + return dst; +} + +#endif // AOM_AOM_DSP_X86_MEM_SSE2_H_ diff --git a/third_party/aom/aom_dsp/x86/obmc_intrinsic_sse4.h b/third_party/aom/aom_dsp/x86/obmc_intrinsic_sse4.h new file mode 100644 index 0000000000..5181e444c0 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/obmc_intrinsic_sse4.h @@ -0,0 +1,58 @@ +/* + * Copyright (c) 2018, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_X86_OBMC_INTRINSIC_SSE4_H_ +#define AOM_AOM_DSP_X86_OBMC_INTRINSIC_SSE4_H_ + +#include + +#include "aom_dsp/x86/obmc_intrinsic_ssse3.h" + +static INLINE void obmc_variance_w4(const uint8_t *pre, const int pre_stride, + const int32_t *wsrc, const int32_t *mask, + unsigned int *const sse, int *const sum, + const int h) { + const int pre_step = pre_stride - 4; + int n = 0; + __m128i v_sum_d = _mm_setzero_si128(); + __m128i v_sse_d = _mm_setzero_si128(); + + assert(IS_POWER_OF_TWO(h)); + + do { + const __m128i v_p_b = _mm_cvtsi32_si128(*(const uint32_t *)(pre + n)); + const __m128i v_m_d = _mm_load_si128((const __m128i *)(mask + n)); + const __m128i v_w_d = _mm_load_si128((const __m128i *)(wsrc + n)); + + const __m128i v_p_d = _mm_cvtepu8_epi32(v_p_b); + + // Values in both pre and mask fit in 15 bits, and are packed at 32 bit + // boundaries. We use pmaddwd, as it has lower latency on Haswell + // than pmulld but produces the same result with these inputs. + const __m128i v_pm_d = _mm_madd_epi16(v_p_d, v_m_d); + + const __m128i v_diff_d = _mm_sub_epi32(v_w_d, v_pm_d); + const __m128i v_rdiff_d = xx_roundn_epi32(v_diff_d, 12); + const __m128i v_sqrdiff_d = _mm_mullo_epi32(v_rdiff_d, v_rdiff_d); + + v_sum_d = _mm_add_epi32(v_sum_d, v_rdiff_d); + v_sse_d = _mm_add_epi32(v_sse_d, v_sqrdiff_d); + + n += 4; + + if (n % 4 == 0) pre += pre_step; + } while (n < 4 * h); + + *sum = xx_hsum_epi32_si32(v_sum_d); + *sse = xx_hsum_epi32_si32(v_sse_d); +} + +#endif // AOM_AOM_DSP_X86_OBMC_INTRINSIC_SSE4_H_ diff --git a/third_party/aom/aom_dsp/x86/obmc_intrinsic_ssse3.h b/third_party/aom/aom_dsp/x86/obmc_intrinsic_ssse3.h new file mode 100644 index 0000000000..48486c6c4c --- /dev/null +++ b/third_party/aom/aom_dsp/x86/obmc_intrinsic_ssse3.h @@ -0,0 +1,54 @@ +/* + * Copyright (c) 2017, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_X86_OBMC_INTRINSIC_SSSE3_H_ +#define AOM_AOM_DSP_X86_OBMC_INTRINSIC_SSSE3_H_ + +#include + +#include "config/aom_config.h" + +static INLINE int32_t xx_hsum_epi32_si32(__m128i v_d) { + v_d = _mm_hadd_epi32(v_d, v_d); + v_d = _mm_hadd_epi32(v_d, v_d); + return _mm_cvtsi128_si32(v_d); +} + +static INLINE int64_t xx_hsum_epi64_si64(__m128i v_q) { + v_q = _mm_add_epi64(v_q, _mm_srli_si128(v_q, 8)); +#if ARCH_X86_64 + return _mm_cvtsi128_si64(v_q); +#else + { + int64_t tmp; + _mm_storel_epi64((__m128i *)&tmp, v_q); + return tmp; + } +#endif +} + +static INLINE int64_t xx_hsum_epi32_si64(__m128i v_d) { + const __m128i v_sign_d = _mm_cmplt_epi32(v_d, _mm_setzero_si128()); + const __m128i v_0_q = _mm_unpacklo_epi32(v_d, v_sign_d); + const __m128i v_1_q = _mm_unpackhi_epi32(v_d, v_sign_d); + return xx_hsum_epi64_si64(_mm_add_epi64(v_0_q, v_1_q)); +} + +// This is equivalent to ROUND_POWER_OF_TWO_SIGNED(v_val_d, bits) +static INLINE __m128i xx_roundn_epi32(__m128i v_val_d, int bits) { + const __m128i v_bias_d = _mm_set1_epi32((1 << bits) >> 1); + const __m128i v_sign_d = _mm_srai_epi32(v_val_d, 31); + const __m128i v_tmp_d = + _mm_add_epi32(_mm_add_epi32(v_val_d, v_bias_d), v_sign_d); + return _mm_srai_epi32(v_tmp_d, bits); +} + +#endif // AOM_AOM_DSP_X86_OBMC_INTRINSIC_SSSE3_H_ diff --git a/third_party/aom/aom_dsp/x86/obmc_sad_avx2.c b/third_party/aom/aom_dsp/x86/obmc_sad_avx2.c new file mode 100644 index 0000000000..2aa2a05551 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/obmc_sad_avx2.c @@ -0,0 +1,270 @@ +/* + * Copyright (c) 2018, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include +#include + +#include "config/aom_config.h" + +#include "aom_ports/mem.h" +#include "aom/aom_integer.h" + +#include "aom_dsp/aom_dsp_common.h" +#include "aom_dsp/x86/obmc_intrinsic_ssse3.h" +#include "aom_dsp/x86/synonyms.h" + +//////////////////////////////////////////////////////////////////////////////// +// 8 bit +//////////////////////////////////////////////////////////////////////////////// + +static INLINE unsigned int obmc_sad_w4_avx2(const uint8_t *pre, + const int pre_stride, + const int32_t *wsrc, + const int32_t *mask, + const int height) { + int n = 0; + __m256i v_sad_d = _mm256_setzero_si256(); + const __m256i v_bias_d = _mm256_set1_epi32((1 << 12) >> 1); + + do { + const __m128i v_p_b_0 = xx_loadl_32(pre); + const __m128i v_p_b_1 = xx_loadl_32(pre + pre_stride); + const __m128i v_p_b = _mm_unpacklo_epi32(v_p_b_0, v_p_b_1); + const __m256i v_m_d = _mm256_lddqu_si256((__m256i *)(mask + n)); + const __m256i v_w_d = _mm256_lddqu_si256((__m256i *)(wsrc + n)); + + const __m256i v_p_d = _mm256_cvtepu8_epi32(v_p_b); + + // Values in both pre and mask fit in 15 bits, and are packed at 32 bit + // boundaries. We use pmaddwd, as it has lower latency on Haswell + // than pmulld but produces the same result with these inputs. + const __m256i v_pm_d = _mm256_madd_epi16(v_p_d, v_m_d); + + const __m256i v_diff_d = _mm256_sub_epi32(v_w_d, v_pm_d); + const __m256i v_absdiff_d = _mm256_abs_epi32(v_diff_d); + + // Rounded absolute difference + const __m256i v_tmp_d = _mm256_add_epi32(v_absdiff_d, v_bias_d); + const __m256i v_rad_d = _mm256_srli_epi32(v_tmp_d, 12); + + v_sad_d = _mm256_add_epi32(v_sad_d, v_rad_d); + + n += 8; + pre += pre_stride << 1; + } while (n < 8 * (height >> 1)); + + __m128i v_sad_d_0 = _mm256_castsi256_si128(v_sad_d); + __m128i v_sad_d_1 = _mm256_extracti128_si256(v_sad_d, 1); + v_sad_d_0 = _mm_add_epi32(v_sad_d_0, v_sad_d_1); + return xx_hsum_epi32_si32(v_sad_d_0); +} + +static INLINE unsigned int obmc_sad_w8n_avx2( + const uint8_t *pre, const int pre_stride, const int32_t *wsrc, + const int32_t *mask, const int width, const int height) { + const int pre_step = pre_stride - width; + int n = 0; + __m256i v_sad_d = _mm256_setzero_si256(); + const __m256i v_bias_d = _mm256_set1_epi32((1 << 12) >> 1); + assert(width >= 8); + assert(IS_POWER_OF_TWO(width)); + + do { + const __m128i v_p0_b = xx_loadl_64(pre + n); + const __m256i v_m0_d = _mm256_lddqu_si256((__m256i *)(mask + n)); + const __m256i v_w0_d = _mm256_lddqu_si256((__m256i *)(wsrc + n)); + + const __m256i v_p0_d = _mm256_cvtepu8_epi32(v_p0_b); + + // Values in both pre and mask fit in 15 bits, and are packed at 32 bit + // boundaries. We use pmaddwd, as it has lower latency on Haswell + // than pmulld but produces the same result with these inputs. + const __m256i v_pm0_d = _mm256_madd_epi16(v_p0_d, v_m0_d); + + const __m256i v_diff0_d = _mm256_sub_epi32(v_w0_d, v_pm0_d); + const __m256i v_absdiff0_d = _mm256_abs_epi32(v_diff0_d); + + // Rounded absolute difference + const __m256i v_tmp_d = _mm256_add_epi32(v_absdiff0_d, v_bias_d); + const __m256i v_rad0_d = _mm256_srli_epi32(v_tmp_d, 12); + + v_sad_d = _mm256_add_epi32(v_sad_d, v_rad0_d); + + n += 8; + + if ((n & (width - 1)) == 0) pre += pre_step; + } while (n < width * height); + + __m128i v_sad_d_0 = _mm256_castsi256_si128(v_sad_d); + __m128i v_sad_d_1 = _mm256_extracti128_si256(v_sad_d, 1); + v_sad_d_0 = _mm_add_epi32(v_sad_d_0, v_sad_d_1); + return xx_hsum_epi32_si32(v_sad_d_0); +} + +#define OBMCSADWXH(w, h) \ + unsigned int aom_obmc_sad##w##x##h##_avx2( \ + const uint8_t *pre, int pre_stride, const int32_t *wsrc, \ + const int32_t *msk) { \ + if (w == 4) { \ + return obmc_sad_w4_avx2(pre, pre_stride, wsrc, msk, h); \ + } else { \ + return obmc_sad_w8n_avx2(pre, pre_stride, wsrc, msk, w, h); \ + } \ + } + +OBMCSADWXH(128, 128) +OBMCSADWXH(128, 64) +OBMCSADWXH(64, 128) +OBMCSADWXH(64, 64) +OBMCSADWXH(64, 32) +OBMCSADWXH(32, 64) +OBMCSADWXH(32, 32) +OBMCSADWXH(32, 16) +OBMCSADWXH(16, 32) +OBMCSADWXH(16, 16) +OBMCSADWXH(16, 8) +OBMCSADWXH(8, 16) +OBMCSADWXH(8, 8) +OBMCSADWXH(8, 4) +OBMCSADWXH(4, 8) +OBMCSADWXH(4, 4) +OBMCSADWXH(4, 16) +OBMCSADWXH(16, 4) +OBMCSADWXH(8, 32) +OBMCSADWXH(32, 8) +OBMCSADWXH(16, 64) +OBMCSADWXH(64, 16) + +//////////////////////////////////////////////////////////////////////////////// +// High bit-depth +//////////////////////////////////////////////////////////////////////////////// + +static INLINE unsigned int hbd_obmc_sad_w4_avx2(const uint8_t *pre8, + const int pre_stride, + const int32_t *wsrc, + const int32_t *mask, + const int height) { + const uint16_t *pre = CONVERT_TO_SHORTPTR(pre8); + int n = 0; + __m256i v_sad_d = _mm256_setzero_si256(); + const __m256i v_bias_d = _mm256_set1_epi32((1 << 12) >> 1); + do { + const __m128i v_p_w_0 = xx_loadl_64(pre); + const __m128i v_p_w_1 = xx_loadl_64(pre + pre_stride); + const __m128i v_p_w = _mm_unpacklo_epi64(v_p_w_0, v_p_w_1); + const __m256i v_m_d = _mm256_lddqu_si256((__m256i *)(mask + n)); + const __m256i v_w_d = _mm256_lddqu_si256((__m256i *)(wsrc + n)); + + const __m256i v_p_d = _mm256_cvtepu16_epi32(v_p_w); + + // Values in both pre and mask fit in 15 bits, and are packed at 32 bit + // boundaries. We use pmaddwd, as it has lower latency on Haswell + // than pmulld but produces the same result with these inputs. + const __m256i v_pm_d = _mm256_madd_epi16(v_p_d, v_m_d); + + const __m256i v_diff_d = _mm256_sub_epi32(v_w_d, v_pm_d); + const __m256i v_absdiff_d = _mm256_abs_epi32(v_diff_d); + + // Rounded absolute difference + + const __m256i v_tmp_d = _mm256_add_epi32(v_absdiff_d, v_bias_d); + const __m256i v_rad_d = _mm256_srli_epi32(v_tmp_d, 12); + + v_sad_d = _mm256_add_epi32(v_sad_d, v_rad_d); + + n += 8; + + pre += pre_stride << 1; + } while (n < 8 * (height >> 1)); + + __m128i v_sad_d_0 = _mm256_castsi256_si128(v_sad_d); + __m128i v_sad_d_1 = _mm256_extracti128_si256(v_sad_d, 1); + v_sad_d_0 = _mm_add_epi32(v_sad_d_0, v_sad_d_1); + return xx_hsum_epi32_si32(v_sad_d_0); +} + +static INLINE unsigned int hbd_obmc_sad_w8n_avx2( + const uint8_t *pre8, const int pre_stride, const int32_t *wsrc, + const int32_t *mask, const int width, const int height) { + const uint16_t *pre = CONVERT_TO_SHORTPTR(pre8); + const int pre_step = pre_stride - width; + int n = 0; + __m256i v_sad_d = _mm256_setzero_si256(); + const __m256i v_bias_d = _mm256_set1_epi32((1 << 12) >> 1); + + assert(width >= 8); + assert(IS_POWER_OF_TWO(width)); + + do { + const __m128i v_p0_w = _mm_lddqu_si128((__m128i *)(pre + n)); + const __m256i v_m0_d = _mm256_lddqu_si256((__m256i *)(mask + n)); + const __m256i v_w0_d = _mm256_lddqu_si256((__m256i *)(wsrc + n)); + + const __m256i v_p0_d = _mm256_cvtepu16_epi32(v_p0_w); + + // Values in both pre and mask fit in 15 bits, and are packed at 32 bit + // boundaries. We use pmaddwd, as it has lower latency on Haswell + // than pmulld but produces the same result with these inputs. + const __m256i v_pm0_d = _mm256_madd_epi16(v_p0_d, v_m0_d); + + const __m256i v_diff0_d = _mm256_sub_epi32(v_w0_d, v_pm0_d); + const __m256i v_absdiff0_d = _mm256_abs_epi32(v_diff0_d); + + // Rounded absolute difference + const __m256i v_tmp_d = _mm256_add_epi32(v_absdiff0_d, v_bias_d); + const __m256i v_rad0_d = _mm256_srli_epi32(v_tmp_d, 12); + + v_sad_d = _mm256_add_epi32(v_sad_d, v_rad0_d); + + n += 8; + + if (n % width == 0) pre += pre_step; + } while (n < width * height); + + __m128i v_sad_d_0 = _mm256_castsi256_si128(v_sad_d); + __m128i v_sad_d_1 = _mm256_extracti128_si256(v_sad_d, 1); + v_sad_d_0 = _mm_add_epi32(v_sad_d_0, v_sad_d_1); + return xx_hsum_epi32_si32(v_sad_d_0); +} + +#define HBD_OBMCSADWXH(w, h) \ + unsigned int aom_highbd_obmc_sad##w##x##h##_avx2( \ + const uint8_t *pre, int pre_stride, const int32_t *wsrc, \ + const int32_t *mask) { \ + if (w == 4) { \ + return hbd_obmc_sad_w4_avx2(pre, pre_stride, wsrc, mask, h); \ + } else { \ + return hbd_obmc_sad_w8n_avx2(pre, pre_stride, wsrc, mask, w, h); \ + } \ + } + +HBD_OBMCSADWXH(128, 128) +HBD_OBMCSADWXH(128, 64) +HBD_OBMCSADWXH(64, 128) +HBD_OBMCSADWXH(64, 64) +HBD_OBMCSADWXH(64, 32) +HBD_OBMCSADWXH(32, 64) +HBD_OBMCSADWXH(32, 32) +HBD_OBMCSADWXH(32, 16) +HBD_OBMCSADWXH(16, 32) +HBD_OBMCSADWXH(16, 16) +HBD_OBMCSADWXH(16, 8) +HBD_OBMCSADWXH(8, 16) +HBD_OBMCSADWXH(8, 8) +HBD_OBMCSADWXH(8, 4) +HBD_OBMCSADWXH(4, 8) +HBD_OBMCSADWXH(4, 4) +HBD_OBMCSADWXH(4, 16) +HBD_OBMCSADWXH(16, 4) +HBD_OBMCSADWXH(8, 32) +HBD_OBMCSADWXH(32, 8) +HBD_OBMCSADWXH(16, 64) +HBD_OBMCSADWXH(64, 16) diff --git a/third_party/aom/aom_dsp/x86/obmc_sad_sse4.c b/third_party/aom/aom_dsp/x86/obmc_sad_sse4.c new file mode 100644 index 0000000000..0338a8c777 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/obmc_sad_sse4.c @@ -0,0 +1,268 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include +#include + +#include "config/aom_config.h" + +#include "aom_ports/mem.h" +#include "aom/aom_integer.h" + +#include "aom_dsp/aom_dsp_common.h" +#include "aom_dsp/x86/obmc_intrinsic_ssse3.h" +#include "aom_dsp/x86/synonyms.h" + +//////////////////////////////////////////////////////////////////////////////// +// 8 bit +//////////////////////////////////////////////////////////////////////////////// + +static AOM_FORCE_INLINE unsigned int obmc_sad_w4(const uint8_t *pre, + const int pre_stride, + const int32_t *wsrc, + const int32_t *mask, + const int height) { + const int pre_step = pre_stride - 4; + int n = 0; + __m128i v_sad_d = _mm_setzero_si128(); + + do { + const __m128i v_p_b = xx_loadl_32(pre + n); + const __m128i v_m_d = xx_load_128(mask + n); + const __m128i v_w_d = xx_load_128(wsrc + n); + + const __m128i v_p_d = _mm_cvtepu8_epi32(v_p_b); + + // Values in both pre and mask fit in 15 bits, and are packed at 32 bit + // boundaries. We use pmaddwd, as it has lower latency on Haswell + // than pmulld but produces the same result with these inputs. + const __m128i v_pm_d = _mm_madd_epi16(v_p_d, v_m_d); + + const __m128i v_diff_d = _mm_sub_epi32(v_w_d, v_pm_d); + const __m128i v_absdiff_d = _mm_abs_epi32(v_diff_d); + + // Rounded absolute difference + const __m128i v_rad_d = xx_roundn_epu32(v_absdiff_d, 12); + + v_sad_d = _mm_add_epi32(v_sad_d, v_rad_d); + + n += 4; + + if (n % 4 == 0) pre += pre_step; + } while (n < 4 * height); + + return xx_hsum_epi32_si32(v_sad_d); +} + +static AOM_FORCE_INLINE unsigned int obmc_sad_w8n( + const uint8_t *pre, const int pre_stride, const int32_t *wsrc, + const int32_t *mask, const int width, const int height) { + const int pre_step = pre_stride - width; + int n = 0; + __m128i v_sad_d = _mm_setzero_si128(); + + assert(width >= 8); + assert(IS_POWER_OF_TWO(width)); + + do { + const __m128i v_p1_b = xx_loadl_32(pre + n + 4); + const __m128i v_m1_d = xx_load_128(mask + n + 4); + const __m128i v_w1_d = xx_load_128(wsrc + n + 4); + const __m128i v_p0_b = xx_loadl_32(pre + n); + const __m128i v_m0_d = xx_load_128(mask + n); + const __m128i v_w0_d = xx_load_128(wsrc + n); + + const __m128i v_p0_d = _mm_cvtepu8_epi32(v_p0_b); + const __m128i v_p1_d = _mm_cvtepu8_epi32(v_p1_b); + + // Values in both pre and mask fit in 15 bits, and are packed at 32 bit + // boundaries. We use pmaddwd, as it has lower latency on Haswell + // than pmulld but produces the same result with these inputs. + const __m128i v_pm0_d = _mm_madd_epi16(v_p0_d, v_m0_d); + const __m128i v_pm1_d = _mm_madd_epi16(v_p1_d, v_m1_d); + + const __m128i v_diff0_d = _mm_sub_epi32(v_w0_d, v_pm0_d); + const __m128i v_diff1_d = _mm_sub_epi32(v_w1_d, v_pm1_d); + const __m128i v_absdiff0_d = _mm_abs_epi32(v_diff0_d); + const __m128i v_absdiff1_d = _mm_abs_epi32(v_diff1_d); + + // Rounded absolute difference + const __m128i v_rad0_d = xx_roundn_epu32(v_absdiff0_d, 12); + const __m128i v_rad1_d = xx_roundn_epu32(v_absdiff1_d, 12); + + v_sad_d = _mm_add_epi32(v_sad_d, v_rad0_d); + v_sad_d = _mm_add_epi32(v_sad_d, v_rad1_d); + + n += 8; + + if (n % width == 0) pre += pre_step; + } while (n < width * height); + + return xx_hsum_epi32_si32(v_sad_d); +} + +#define OBMCSADWXH(w, h) \ + unsigned int aom_obmc_sad##w##x##h##_sse4_1( \ + const uint8_t *pre, int pre_stride, const int32_t *wsrc, \ + const int32_t *msk) { \ + if (w == 4) { \ + return obmc_sad_w4(pre, pre_stride, wsrc, msk, h); \ + } else { \ + return obmc_sad_w8n(pre, pre_stride, wsrc, msk, w, h); \ + } \ + } + +OBMCSADWXH(128, 128) +OBMCSADWXH(128, 64) +OBMCSADWXH(64, 128) +OBMCSADWXH(64, 64) +OBMCSADWXH(64, 32) +OBMCSADWXH(32, 64) +OBMCSADWXH(32, 32) +OBMCSADWXH(32, 16) +OBMCSADWXH(16, 32) +OBMCSADWXH(16, 16) +OBMCSADWXH(16, 8) +OBMCSADWXH(8, 16) +OBMCSADWXH(8, 8) +OBMCSADWXH(8, 4) +OBMCSADWXH(4, 8) +OBMCSADWXH(4, 4) +OBMCSADWXH(4, 16) +OBMCSADWXH(16, 4) +OBMCSADWXH(8, 32) +OBMCSADWXH(32, 8) +OBMCSADWXH(16, 64) +OBMCSADWXH(64, 16) + +//////////////////////////////////////////////////////////////////////////////// +// High bit-depth +//////////////////////////////////////////////////////////////////////////////// + +static AOM_FORCE_INLINE unsigned int hbd_obmc_sad_w4(const uint8_t *pre8, + const int pre_stride, + const int32_t *wsrc, + const int32_t *mask, + const int height) { + const uint16_t *pre = CONVERT_TO_SHORTPTR(pre8); + const int pre_step = pre_stride - 4; + int n = 0; + __m128i v_sad_d = _mm_setzero_si128(); + + do { + const __m128i v_p_w = xx_loadl_64(pre + n); + const __m128i v_m_d = xx_load_128(mask + n); + const __m128i v_w_d = xx_load_128(wsrc + n); + + const __m128i v_p_d = _mm_cvtepu16_epi32(v_p_w); + + // Values in both pre and mask fit in 15 bits, and are packed at 32 bit + // boundaries. We use pmaddwd, as it has lower latency on Haswell + // than pmulld but produces the same result with these inputs. + const __m128i v_pm_d = _mm_madd_epi16(v_p_d, v_m_d); + + const __m128i v_diff_d = _mm_sub_epi32(v_w_d, v_pm_d); + const __m128i v_absdiff_d = _mm_abs_epi32(v_diff_d); + + // Rounded absolute difference + const __m128i v_rad_d = xx_roundn_epu32(v_absdiff_d, 12); + + v_sad_d = _mm_add_epi32(v_sad_d, v_rad_d); + + n += 4; + + if (n % 4 == 0) pre += pre_step; + } while (n < 4 * height); + + return xx_hsum_epi32_si32(v_sad_d); +} + +static AOM_FORCE_INLINE unsigned int hbd_obmc_sad_w8n( + const uint8_t *pre8, const int pre_stride, const int32_t *wsrc, + const int32_t *mask, const int width, const int height) { + const uint16_t *pre = CONVERT_TO_SHORTPTR(pre8); + const int pre_step = pre_stride - width; + int n = 0; + __m128i v_sad_d = _mm_setzero_si128(); + + assert(width >= 8); + assert(IS_POWER_OF_TWO(width)); + + do { + const __m128i v_p1_w = xx_loadl_64(pre + n + 4); + const __m128i v_m1_d = xx_load_128(mask + n + 4); + const __m128i v_w1_d = xx_load_128(wsrc + n + 4); + const __m128i v_p0_w = xx_loadl_64(pre + n); + const __m128i v_m0_d = xx_load_128(mask + n); + const __m128i v_w0_d = xx_load_128(wsrc + n); + + const __m128i v_p0_d = _mm_cvtepu16_epi32(v_p0_w); + const __m128i v_p1_d = _mm_cvtepu16_epi32(v_p1_w); + + // Values in both pre and mask fit in 15 bits, and are packed at 32 bit + // boundaries. We use pmaddwd, as it has lower latency on Haswell + // than pmulld but produces the same result with these inputs. + const __m128i v_pm0_d = _mm_madd_epi16(v_p0_d, v_m0_d); + const __m128i v_pm1_d = _mm_madd_epi16(v_p1_d, v_m1_d); + + const __m128i v_diff0_d = _mm_sub_epi32(v_w0_d, v_pm0_d); + const __m128i v_diff1_d = _mm_sub_epi32(v_w1_d, v_pm1_d); + const __m128i v_absdiff0_d = _mm_abs_epi32(v_diff0_d); + const __m128i v_absdiff1_d = _mm_abs_epi32(v_diff1_d); + + // Rounded absolute difference + const __m128i v_rad0_d = xx_roundn_epu32(v_absdiff0_d, 12); + const __m128i v_rad1_d = xx_roundn_epu32(v_absdiff1_d, 12); + + v_sad_d = _mm_add_epi32(v_sad_d, v_rad0_d); + v_sad_d = _mm_add_epi32(v_sad_d, v_rad1_d); + + n += 8; + + if (n % width == 0) pre += pre_step; + } while (n < width * height); + + return xx_hsum_epi32_si32(v_sad_d); +} + +#define HBD_OBMCSADWXH(w, h) \ + unsigned int aom_highbd_obmc_sad##w##x##h##_sse4_1( \ + const uint8_t *pre, int pre_stride, const int32_t *wsrc, \ + const int32_t *mask) { \ + if (w == 4) { \ + return hbd_obmc_sad_w4(pre, pre_stride, wsrc, mask, h); \ + } else { \ + return hbd_obmc_sad_w8n(pre, pre_stride, wsrc, mask, w, h); \ + } \ + } + +HBD_OBMCSADWXH(128, 128) +HBD_OBMCSADWXH(128, 64) +HBD_OBMCSADWXH(64, 128) +HBD_OBMCSADWXH(64, 64) +HBD_OBMCSADWXH(64, 32) +HBD_OBMCSADWXH(32, 64) +HBD_OBMCSADWXH(32, 32) +HBD_OBMCSADWXH(32, 16) +HBD_OBMCSADWXH(16, 32) +HBD_OBMCSADWXH(16, 16) +HBD_OBMCSADWXH(16, 8) +HBD_OBMCSADWXH(8, 16) +HBD_OBMCSADWXH(8, 8) +HBD_OBMCSADWXH(8, 4) +HBD_OBMCSADWXH(4, 8) +HBD_OBMCSADWXH(4, 4) +HBD_OBMCSADWXH(4, 16) +HBD_OBMCSADWXH(16, 4) +HBD_OBMCSADWXH(8, 32) +HBD_OBMCSADWXH(32, 8) +HBD_OBMCSADWXH(16, 64) +HBD_OBMCSADWXH(64, 16) diff --git a/third_party/aom/aom_dsp/x86/obmc_variance_avx2.c b/third_party/aom/aom_dsp/x86/obmc_variance_avx2.c new file mode 100644 index 0000000000..bfec0e8a84 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/obmc_variance_avx2.c @@ -0,0 +1,190 @@ +/* + * Copyright (c) 2018, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include +#include + +#include "config/aom_config.h" + +#include "aom_ports/mem.h" +#include "aom/aom_integer.h" + +#include "aom_dsp/aom_dsp_common.h" +#include "aom_dsp/aom_filter.h" +#include "aom_dsp/x86/obmc_intrinsic_sse4.h" + +//////////////////////////////////////////////////////////////////////////////// +// 8 bit +//////////////////////////////////////////////////////////////////////////////// + +static INLINE void obmc_variance_w8n(const uint8_t *pre, const int pre_stride, + const int32_t *wsrc, const int32_t *mask, + unsigned int *const sse, int *const sum, + const int w, const int h) { + int n = 0, width, height = h; + __m128i v_sum_d = _mm_setzero_si128(); + __m128i v_sse_d = _mm_setzero_si128(); + const __m256i v_bias_d = _mm256_set1_epi32((1 << 12) >> 1); + __m128i v_d; + const uint8_t *pre_temp; + assert(w >= 8); + assert(IS_POWER_OF_TWO(w)); + assert(IS_POWER_OF_TWO(h)); + do { + width = w; + pre_temp = pre; + do { + const __m128i v_p_b = _mm_loadl_epi64((const __m128i *)pre_temp); + const __m256i v_m_d = _mm256_loadu_si256((__m256i const *)(mask + n)); + const __m256i v_w_d = _mm256_loadu_si256((__m256i const *)(wsrc + n)); + const __m256i v_p0_d = _mm256_cvtepu8_epi32(v_p_b); + + // Values in both pre and mask fit in 15 bits, and are packed at 32 bit + // boundaries. We use pmaddwd, as it has lower latency on Haswell + // than pmulld but produces the same result with these inputs. + const __m256i v_pm_d = _mm256_madd_epi16(v_p0_d, v_m_d); + const __m256i v_diff0_d = _mm256_sub_epi32(v_w_d, v_pm_d); + + const __m256i v_sign_d = _mm256_srai_epi32(v_diff0_d, 31); + const __m256i v_tmp_d = + _mm256_add_epi32(_mm256_add_epi32(v_diff0_d, v_bias_d), v_sign_d); + const __m256i v_rdiff0_d = _mm256_srai_epi32(v_tmp_d, 12); + const __m128i v_rdiff_d = _mm256_castsi256_si128(v_rdiff0_d); + const __m128i v_rdiff1_d = _mm256_extracti128_si256(v_rdiff0_d, 1); + + const __m128i v_rdiff01_w = _mm_packs_epi32(v_rdiff_d, v_rdiff1_d); + const __m128i v_sqrdiff_d = _mm_madd_epi16(v_rdiff01_w, v_rdiff01_w); + + v_sum_d = _mm_add_epi32(v_sum_d, v_rdiff_d); + v_sum_d = _mm_add_epi32(v_sum_d, v_rdiff1_d); + v_sse_d = _mm_add_epi32(v_sse_d, v_sqrdiff_d); + + pre_temp += 8; + n += 8; + width -= 8; + } while (width > 0); + pre += pre_stride; + height -= 1; + } while (height > 0); + v_d = _mm_hadd_epi32(v_sum_d, v_sse_d); + v_d = _mm_hadd_epi32(v_d, v_d); + *sum = _mm_cvtsi128_si32(v_d); + *sse = _mm_cvtsi128_si32(_mm_srli_si128(v_d, 4)); +} + +static INLINE void obmc_variance_w16n(const uint8_t *pre, const int pre_stride, + const int32_t *wsrc, const int32_t *mask, + unsigned int *const sse, int *const sum, + const int w, const int h) { + int n = 0, width, height = h; + __m256i v_d; + __m128i res0; + const uint8_t *pre_temp; + const __m256i v_bias_d = _mm256_set1_epi32((1 << 12) >> 1); + __m256i v_sum_d = _mm256_setzero_si256(); + __m256i v_sse_d = _mm256_setzero_si256(); + + assert(w >= 16); + assert(IS_POWER_OF_TWO(w)); + assert(IS_POWER_OF_TWO(h)); + do { + width = w; + pre_temp = pre; + do { + const __m128i v_p_b = _mm_loadu_si128((__m128i *)pre_temp); + const __m256i v_m0_d = _mm256_loadu_si256((__m256i const *)(mask + n)); + const __m256i v_w0_d = _mm256_loadu_si256((__m256i const *)(wsrc + n)); + const __m256i v_m1_d = + _mm256_loadu_si256((__m256i const *)(mask + n + 8)); + const __m256i v_w1_d = + _mm256_loadu_si256((__m256i const *)(wsrc + n + 8)); + + const __m256i v_p0_d = _mm256_cvtepu8_epi32(v_p_b); + const __m256i v_p1_d = _mm256_cvtepu8_epi32(_mm_srli_si128(v_p_b, 8)); + + const __m256i v_pm0_d = _mm256_madd_epi16(v_p0_d, v_m0_d); + const __m256i v_pm1_d = _mm256_madd_epi16(v_p1_d, v_m1_d); + + const __m256i v_diff0_d = _mm256_sub_epi32(v_w0_d, v_pm0_d); + const __m256i v_diff1_d = _mm256_sub_epi32(v_w1_d, v_pm1_d); + + const __m256i v_sign0_d = _mm256_srai_epi32(v_diff0_d, 31); + const __m256i v_sign1_d = _mm256_srai_epi32(v_diff1_d, 31); + + const __m256i v_tmp0_d = + _mm256_add_epi32(_mm256_add_epi32(v_diff0_d, v_bias_d), v_sign0_d); + const __m256i v_tmp1_d = + _mm256_add_epi32(_mm256_add_epi32(v_diff1_d, v_bias_d), v_sign1_d); + + const __m256i v_rdiff0_d = _mm256_srai_epi32(v_tmp0_d, 12); + const __m256i v_rdiff2_d = _mm256_srai_epi32(v_tmp1_d, 12); + + const __m256i v_rdiff1_d = _mm256_add_epi32(v_rdiff0_d, v_rdiff2_d); + const __m256i v_rdiff01_w = _mm256_packs_epi32(v_rdiff0_d, v_rdiff2_d); + const __m256i v_sqrdiff_d = _mm256_madd_epi16(v_rdiff01_w, v_rdiff01_w); + + v_sum_d = _mm256_add_epi32(v_sum_d, v_rdiff1_d); + v_sse_d = _mm256_add_epi32(v_sse_d, v_sqrdiff_d); + + pre_temp += 16; + n += 16; + width -= 16; + } while (width > 0); + pre += pre_stride; + height -= 1; + } while (height > 0); + + v_d = _mm256_hadd_epi32(v_sum_d, v_sse_d); + v_d = _mm256_hadd_epi32(v_d, v_d); + res0 = _mm256_castsi256_si128(v_d); + res0 = _mm_add_epi32(res0, _mm256_extractf128_si256(v_d, 1)); + *sum = _mm_cvtsi128_si32(res0); + *sse = _mm_cvtsi128_si32(_mm_srli_si128(res0, 4)); +} + +#define OBMCVARWXH(W, H) \ + unsigned int aom_obmc_variance##W##x##H##_avx2( \ + const uint8_t *pre, int pre_stride, const int32_t *wsrc, \ + const int32_t *mask, unsigned int *sse) { \ + int sum; \ + if (W == 4) { \ + obmc_variance_w4(pre, pre_stride, wsrc, mask, sse, &sum, H); \ + } else if (W == 8) { \ + obmc_variance_w8n(pre, pre_stride, wsrc, mask, sse, &sum, W, H); \ + } else { \ + obmc_variance_w16n(pre, pre_stride, wsrc, mask, sse, &sum, W, H); \ + } \ + \ + return *sse - (unsigned int)(((int64_t)sum * sum) / (W * H)); \ + } + +OBMCVARWXH(128, 128) +OBMCVARWXH(128, 64) +OBMCVARWXH(64, 128) +OBMCVARWXH(64, 64) +OBMCVARWXH(64, 32) +OBMCVARWXH(32, 64) +OBMCVARWXH(32, 32) +OBMCVARWXH(32, 16) +OBMCVARWXH(16, 32) +OBMCVARWXH(16, 16) +OBMCVARWXH(16, 8) +OBMCVARWXH(8, 16) +OBMCVARWXH(8, 8) +OBMCVARWXH(8, 4) +OBMCVARWXH(4, 8) +OBMCVARWXH(4, 4) +OBMCVARWXH(4, 16) +OBMCVARWXH(16, 4) +OBMCVARWXH(8, 32) +OBMCVARWXH(32, 8) +OBMCVARWXH(16, 64) +OBMCVARWXH(64, 16) diff --git a/third_party/aom/aom_dsp/x86/obmc_variance_sse4.c b/third_party/aom/aom_dsp/x86/obmc_variance_sse4.c new file mode 100644 index 0000000000..72eda0e578 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/obmc_variance_sse4.c @@ -0,0 +1,380 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include +#include + +#include "config/aom_config.h" + +#include "aom_ports/mem.h" +#include "aom/aom_integer.h" + +#include "aom_dsp/aom_dsp_common.h" +#include "aom_dsp/aom_filter.h" +#include "aom_dsp/x86/obmc_intrinsic_sse4.h" +#include "aom_dsp/x86/synonyms.h" + +//////////////////////////////////////////////////////////////////////////////// +// 8 bit +//////////////////////////////////////////////////////////////////////////////// + +void aom_var_filter_block2d_bil_first_pass_ssse3( + const uint8_t *a, uint16_t *b, unsigned int src_pixels_per_line, + unsigned int pixel_step, unsigned int output_height, + unsigned int output_width, const uint8_t *filter); + +void aom_var_filter_block2d_bil_second_pass_ssse3( + const uint16_t *a, uint8_t *b, unsigned int src_pixels_per_line, + unsigned int pixel_step, unsigned int output_height, + unsigned int output_width, const uint8_t *filter); + +static INLINE void obmc_variance_w8n(const uint8_t *pre, const int pre_stride, + const int32_t *wsrc, const int32_t *mask, + unsigned int *const sse, int *const sum, + const int w, const int h) { + const int pre_step = pre_stride - w; + int n = 0; + __m128i v_sum_d = _mm_setzero_si128(); + __m128i v_sse_d = _mm_setzero_si128(); + + assert(w >= 8); + assert(IS_POWER_OF_TWO(w)); + assert(IS_POWER_OF_TWO(h)); + + do { + const __m128i v_p1_b = xx_loadl_32(pre + n + 4); + const __m128i v_m1_d = xx_load_128(mask + n + 4); + const __m128i v_w1_d = xx_load_128(wsrc + n + 4); + const __m128i v_p0_b = xx_loadl_32(pre + n); + const __m128i v_m0_d = xx_load_128(mask + n); + const __m128i v_w0_d = xx_load_128(wsrc + n); + + const __m128i v_p0_d = _mm_cvtepu8_epi32(v_p0_b); + const __m128i v_p1_d = _mm_cvtepu8_epi32(v_p1_b); + + // Values in both pre and mask fit in 15 bits, and are packed at 32 bit + // boundaries. We use pmaddwd, as it has lower latency on Haswell + // than pmulld but produces the same result with these inputs. + const __m128i v_pm0_d = _mm_madd_epi16(v_p0_d, v_m0_d); + const __m128i v_pm1_d = _mm_madd_epi16(v_p1_d, v_m1_d); + + const __m128i v_diff0_d = _mm_sub_epi32(v_w0_d, v_pm0_d); + const __m128i v_diff1_d = _mm_sub_epi32(v_w1_d, v_pm1_d); + + const __m128i v_rdiff0_d = xx_roundn_epi32(v_diff0_d, 12); + const __m128i v_rdiff1_d = xx_roundn_epi32(v_diff1_d, 12); + const __m128i v_rdiff01_w = _mm_packs_epi32(v_rdiff0_d, v_rdiff1_d); + const __m128i v_sqrdiff_d = _mm_madd_epi16(v_rdiff01_w, v_rdiff01_w); + + v_sum_d = _mm_add_epi32(v_sum_d, v_rdiff0_d); + v_sum_d = _mm_add_epi32(v_sum_d, v_rdiff1_d); + v_sse_d = _mm_add_epi32(v_sse_d, v_sqrdiff_d); + + n += 8; + + if (n % w == 0) pre += pre_step; + } while (n < w * h); + + *sum = xx_hsum_epi32_si32(v_sum_d); + *sse = xx_hsum_epi32_si32(v_sse_d); +} + +#define OBMCVARWXH(W, H) \ + unsigned int aom_obmc_variance##W##x##H##_sse4_1( \ + const uint8_t *pre, int pre_stride, const int32_t *wsrc, \ + const int32_t *mask, unsigned int *sse) { \ + int sum; \ + if (W == 4) { \ + obmc_variance_w4(pre, pre_stride, wsrc, mask, sse, &sum, H); \ + } else { \ + obmc_variance_w8n(pre, pre_stride, wsrc, mask, sse, &sum, W, H); \ + } \ + return *sse - (unsigned int)(((int64_t)sum * sum) / (W * H)); \ + } + +OBMCVARWXH(128, 128) +OBMCVARWXH(128, 64) +OBMCVARWXH(64, 128) +OBMCVARWXH(64, 64) +OBMCVARWXH(64, 32) +OBMCVARWXH(32, 64) +OBMCVARWXH(32, 32) +OBMCVARWXH(32, 16) +OBMCVARWXH(16, 32) +OBMCVARWXH(16, 16) +OBMCVARWXH(16, 8) +OBMCVARWXH(8, 16) +OBMCVARWXH(8, 8) +OBMCVARWXH(8, 4) +OBMCVARWXH(4, 8) +OBMCVARWXH(4, 4) +OBMCVARWXH(4, 16) +OBMCVARWXH(16, 4) +OBMCVARWXH(8, 32) +OBMCVARWXH(32, 8) +OBMCVARWXH(16, 64) +OBMCVARWXH(64, 16) + +#include "config/aom_dsp_rtcd.h" + +#define OBMC_SUBPIX_VAR(W, H) \ + uint32_t aom_obmc_sub_pixel_variance##W##x##H##_sse4_1( \ + const uint8_t *pre, int pre_stride, int xoffset, int yoffset, \ + const int32_t *wsrc, const int32_t *mask, unsigned int *sse) { \ + uint16_t fdata3[(H + 1) * W]; \ + uint8_t temp2[H * W]; \ + \ + aom_var_filter_block2d_bil_first_pass_ssse3( \ + pre, fdata3, pre_stride, 1, H + 1, W, bilinear_filters_2t[xoffset]); \ + aom_var_filter_block2d_bil_second_pass_ssse3( \ + fdata3, temp2, W, W, H, W, bilinear_filters_2t[yoffset]); \ + \ + return aom_obmc_variance##W##x##H##_sse4_1(temp2, W, wsrc, mask, sse); \ + } + +OBMC_SUBPIX_VAR(128, 128) +OBMC_SUBPIX_VAR(128, 64) +OBMC_SUBPIX_VAR(64, 128) +OBMC_SUBPIX_VAR(64, 64) +OBMC_SUBPIX_VAR(64, 32) +OBMC_SUBPIX_VAR(32, 64) +OBMC_SUBPIX_VAR(32, 32) +OBMC_SUBPIX_VAR(32, 16) +OBMC_SUBPIX_VAR(16, 32) +OBMC_SUBPIX_VAR(16, 16) +OBMC_SUBPIX_VAR(16, 8) +OBMC_SUBPIX_VAR(8, 16) +OBMC_SUBPIX_VAR(8, 8) +OBMC_SUBPIX_VAR(8, 4) +OBMC_SUBPIX_VAR(4, 8) +OBMC_SUBPIX_VAR(4, 4) +OBMC_SUBPIX_VAR(4, 16) +OBMC_SUBPIX_VAR(16, 4) +OBMC_SUBPIX_VAR(8, 32) +OBMC_SUBPIX_VAR(32, 8) +OBMC_SUBPIX_VAR(16, 64) +OBMC_SUBPIX_VAR(64, 16) + +//////////////////////////////////////////////////////////////////////////////// +// High bit-depth +//////////////////////////////////////////////////////////////////////////////// + +static INLINE void hbd_obmc_variance_w4( + const uint8_t *pre8, const int pre_stride, const int32_t *wsrc, + const int32_t *mask, uint64_t *const sse, int64_t *const sum, const int h) { + const uint16_t *pre = CONVERT_TO_SHORTPTR(pre8); + const int pre_step = pre_stride - 4; + int n = 0; + __m128i v_sum_d = _mm_setzero_si128(); + __m128i v_sse_d = _mm_setzero_si128(); + + assert(IS_POWER_OF_TWO(h)); + + do { + const __m128i v_p_w = xx_loadl_64(pre + n); + const __m128i v_m_d = xx_load_128(mask + n); + const __m128i v_w_d = xx_load_128(wsrc + n); + + const __m128i v_p_d = _mm_cvtepu16_epi32(v_p_w); + + // Values in both pre and mask fit in 15 bits, and are packed at 32 bit + // boundaries. We use pmaddwd, as it has lower latency on Haswell + // than pmulld but produces the same result with these inputs. + const __m128i v_pm_d = _mm_madd_epi16(v_p_d, v_m_d); + + const __m128i v_diff_d = _mm_sub_epi32(v_w_d, v_pm_d); + const __m128i v_rdiff_d = xx_roundn_epi32(v_diff_d, 12); + const __m128i v_sqrdiff_d = _mm_mullo_epi32(v_rdiff_d, v_rdiff_d); + + v_sum_d = _mm_add_epi32(v_sum_d, v_rdiff_d); + v_sse_d = _mm_add_epi32(v_sse_d, v_sqrdiff_d); + + n += 4; + + if (n % 4 == 0) pre += pre_step; + } while (n < 4 * h); + + *sum = xx_hsum_epi32_si32(v_sum_d); + *sse = xx_hsum_epi32_si32(v_sse_d); +} + +static INLINE void hbd_obmc_variance_w8n( + const uint8_t *pre8, const int pre_stride, const int32_t *wsrc, + const int32_t *mask, uint64_t *const sse, int64_t *const sum, const int w, + const int h) { + const uint16_t *pre = CONVERT_TO_SHORTPTR(pre8); + const int pre_step = pre_stride - w; + int n = 0; + __m128i v_sum_d = _mm_setzero_si128(); + __m128i v_sse_d = _mm_setzero_si128(); + + assert(w >= 8); + assert(IS_POWER_OF_TWO(w)); + assert(IS_POWER_OF_TWO(h)); + + do { + const __m128i v_p1_w = xx_loadl_64(pre + n + 4); + const __m128i v_m1_d = xx_load_128(mask + n + 4); + const __m128i v_w1_d = xx_load_128(wsrc + n + 4); + const __m128i v_p0_w = xx_loadl_64(pre + n); + const __m128i v_m0_d = xx_load_128(mask + n); + const __m128i v_w0_d = xx_load_128(wsrc + n); + + const __m128i v_p0_d = _mm_cvtepu16_epi32(v_p0_w); + const __m128i v_p1_d = _mm_cvtepu16_epi32(v_p1_w); + + // Values in both pre and mask fit in 15 bits, and are packed at 32 bit + // boundaries. We use pmaddwd, as it has lower latency on Haswell + // than pmulld but produces the same result with these inputs. + const __m128i v_pm0_d = _mm_madd_epi16(v_p0_d, v_m0_d); + const __m128i v_pm1_d = _mm_madd_epi16(v_p1_d, v_m1_d); + + const __m128i v_diff0_d = _mm_sub_epi32(v_w0_d, v_pm0_d); + const __m128i v_diff1_d = _mm_sub_epi32(v_w1_d, v_pm1_d); + + const __m128i v_rdiff0_d = xx_roundn_epi32(v_diff0_d, 12); + const __m128i v_rdiff1_d = xx_roundn_epi32(v_diff1_d, 12); + const __m128i v_rdiff01_w = _mm_packs_epi32(v_rdiff0_d, v_rdiff1_d); + const __m128i v_sqrdiff_d = _mm_madd_epi16(v_rdiff01_w, v_rdiff01_w); + + v_sum_d = _mm_add_epi32(v_sum_d, v_rdiff0_d); + v_sum_d = _mm_add_epi32(v_sum_d, v_rdiff1_d); + v_sse_d = _mm_add_epi32(v_sse_d, v_sqrdiff_d); + + n += 8; + + if (n % w == 0) pre += pre_step; + } while (n < w * h); + + *sum += xx_hsum_epi32_si64(v_sum_d); + *sse += xx_hsum_epi32_si64(v_sse_d); +} + +static INLINE void highbd_obmc_variance(const uint8_t *pre8, int pre_stride, + const int32_t *wsrc, + const int32_t *mask, int w, int h, + unsigned int *sse, int *sum) { + int64_t sum64 = 0; + uint64_t sse64 = 0; + if (w == 4) { + hbd_obmc_variance_w4(pre8, pre_stride, wsrc, mask, &sse64, &sum64, h); + } else { + hbd_obmc_variance_w8n(pre8, pre_stride, wsrc, mask, &sse64, &sum64, w, h); + } + *sum = (int)sum64; + *sse = (unsigned int)sse64; +} + +static INLINE void highbd_10_obmc_variance(const uint8_t *pre8, int pre_stride, + const int32_t *wsrc, + const int32_t *mask, int w, int h, + unsigned int *sse, int *sum) { + int64_t sum64 = 0; + uint64_t sse64 = 0; + if (w == 4) { + hbd_obmc_variance_w4(pre8, pre_stride, wsrc, mask, &sse64, &sum64, h); + } else if (w < 128 || h < 128) { + hbd_obmc_variance_w8n(pre8, pre_stride, wsrc, mask, &sse64, &sum64, w, h); + } else { + assert(w == 128 && h == 128); + + do { + hbd_obmc_variance_w8n(pre8, pre_stride, wsrc, mask, &sse64, &sum64, w, + 64); + pre8 += 64 * pre_stride; + wsrc += 64 * w; + mask += 64 * w; + h -= 64; + } while (h > 0); + } + *sum = (int)ROUND_POWER_OF_TWO(sum64, 2); + *sse = (unsigned int)ROUND_POWER_OF_TWO(sse64, 4); +} + +static INLINE void highbd_12_obmc_variance(const uint8_t *pre8, int pre_stride, + const int32_t *wsrc, + const int32_t *mask, int w, int h, + unsigned int *sse, int *sum) { + int64_t sum64 = 0; + uint64_t sse64 = 0; + int max_pel_allowed_per_ovf = 512; + if (w == 4) { + hbd_obmc_variance_w4(pre8, pre_stride, wsrc, mask, &sse64, &sum64, h); + } else if (w * h <= max_pel_allowed_per_ovf) { + hbd_obmc_variance_w8n(pre8, pre_stride, wsrc, mask, &sse64, &sum64, w, h); + } else { + int h_per_ovf = max_pel_allowed_per_ovf / w; + + assert(max_pel_allowed_per_ovf % w == 0); + do { + hbd_obmc_variance_w8n(pre8, pre_stride, wsrc, mask, &sse64, &sum64, w, + h_per_ovf); + pre8 += h_per_ovf * pre_stride; + wsrc += h_per_ovf * w; + mask += h_per_ovf * w; + h -= h_per_ovf; + } while (h > 0); + } + *sum = (int)ROUND_POWER_OF_TWO(sum64, 4); + *sse = (unsigned int)ROUND_POWER_OF_TWO(sse64, 8); +} + +#define HBD_OBMCVARWXH(W, H) \ + unsigned int aom_highbd_obmc_variance##W##x##H##_sse4_1( \ + const uint8_t *pre, int pre_stride, const int32_t *wsrc, \ + const int32_t *mask, unsigned int *sse) { \ + int sum; \ + highbd_obmc_variance(pre, pre_stride, wsrc, mask, W, H, sse, &sum); \ + return *sse - (unsigned int)(((int64_t)sum * sum) / (W * H)); \ + } \ + \ + unsigned int aom_highbd_10_obmc_variance##W##x##H##_sse4_1( \ + const uint8_t *pre, int pre_stride, const int32_t *wsrc, \ + const int32_t *mask, unsigned int *sse) { \ + int sum; \ + int64_t var; \ + highbd_10_obmc_variance(pre, pre_stride, wsrc, mask, W, H, sse, &sum); \ + var = (int64_t)(*sse) - (((int64_t)sum * sum) / (W * H)); \ + return (var >= 0) ? (uint32_t)var : 0; \ + } \ + \ + unsigned int aom_highbd_12_obmc_variance##W##x##H##_sse4_1( \ + const uint8_t *pre, int pre_stride, const int32_t *wsrc, \ + const int32_t *mask, unsigned int *sse) { \ + int sum; \ + int64_t var; \ + highbd_12_obmc_variance(pre, pre_stride, wsrc, mask, W, H, sse, &sum); \ + var = (int64_t)(*sse) - (((int64_t)sum * sum) / (W * H)); \ + return (var >= 0) ? (uint32_t)var : 0; \ + } + +HBD_OBMCVARWXH(128, 128) +HBD_OBMCVARWXH(128, 64) +HBD_OBMCVARWXH(64, 128) +HBD_OBMCVARWXH(64, 64) +HBD_OBMCVARWXH(64, 32) +HBD_OBMCVARWXH(32, 64) +HBD_OBMCVARWXH(32, 32) +HBD_OBMCVARWXH(32, 16) +HBD_OBMCVARWXH(16, 32) +HBD_OBMCVARWXH(16, 16) +HBD_OBMCVARWXH(16, 8) +HBD_OBMCVARWXH(8, 16) +HBD_OBMCVARWXH(8, 8) +HBD_OBMCVARWXH(8, 4) +HBD_OBMCVARWXH(4, 8) +HBD_OBMCVARWXH(4, 4) +HBD_OBMCVARWXH(4, 16) +HBD_OBMCVARWXH(16, 4) +HBD_OBMCVARWXH(8, 32) +HBD_OBMCVARWXH(32, 8) +HBD_OBMCVARWXH(16, 64) +HBD_OBMCVARWXH(64, 16) diff --git a/third_party/aom/aom_dsp/x86/quantize_avx_x86_64.asm b/third_party/aom/aom_dsp/x86/quantize_avx_x86_64.asm new file mode 100644 index 0000000000..216a0bd8f9 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/quantize_avx_x86_64.asm @@ -0,0 +1,435 @@ +; +; Copyright (c) 2016, Alliance for Open Media. All rights reserved +; +; This source code is subject to the terms of the BSD 2 Clause License and +; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License +; was not distributed with this source code in the LICENSE file, you can +; obtain it at www.aomedia.org/license/software. If the Alliance for Open +; Media Patent License 1.0 was not distributed with this source code in the +; PATENTS file, you can obtain it at www.aomedia.org/license/patent. +; + +; + +%include "third_party/x86inc/x86inc.asm" + +SECTION .text + +%macro QUANTIZE_FN 2 +cglobal quantize_%1, 0, %2, 15, coeff, ncoeff, zbin, round, quant, \ + shift, qcoeff, dqcoeff, dequant, \ + eob, scan, iscan + + vzeroupper + +%ifnidn %1, b_32x32 + + ; Special case for ncoeff == 16, as it is frequent and we can save on + ; not setting up a loop. + cmp ncoeffmp, 16 + jne .generic + + ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; + ;; Special case of ncoeff == 16 + ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; + +.single: + + movifnidn coeffq, coeffmp + movifnidn zbinq, zbinmp + mova m0, [zbinq] ; m0 = zbin + + ; Get DC and first 15 AC coeffs - in this special case, that is all. + ; coeff stored as 32bit numbers but we process them as 16 bit numbers + mova m9, [coeffq] + packssdw m9, [coeffq+16] ; m9 = c[i] + mova m10, [coeffq+32] + packssdw m10, [coeffq+48] ; m10 = c[i] + + mov r0, eobmp ; Output pointer + mov r1, qcoeffmp ; Output pointer + mov r2, dqcoeffmp ; Output pointer + + pxor m5, m5 ; m5 = dedicated zero + + pcmpeqw m4, m4 ; All word lanes -1 + paddw m0, m4 ; m0 = zbin - 1 + + pabsw m6, m9 ; m6 = abs(m9) + pabsw m11, m10 ; m11 = abs(m10) + pcmpgtw m7, m6, m0 ; m7 = c[i] >= zbin + punpckhqdq m0, m0 + pcmpgtw m12, m11, m0 ; m12 = c[i] >= zbin + + ; Check if all coeffs are less than zbin. If yes, we just write zeros + ; to the outputs and we are done. + por m14, m7, m12 + ptest m14, m14 + jnz .single_nonzero + + mova [r1 ], ymm5 + mova [r1+32], ymm5 + mova [r2 ], ymm5 + mova [r2+32], ymm5 + mov [r0], word 0 + + vzeroupper + RET + +.single_nonzero: + + ; Actual quantization of size 16 block - setup pointers, rounders, etc. + movifnidn r3, roundmp + movifnidn r4, quantmp + mov r6, dequantmp + mov r5, shiftmp + mova m1, [r3] ; m1 = round + mova m2, [r4] ; m2 = quant + mova m3, [r6] ; m3 = dequant + mova m4, [r5] ; m4 = shift + + mov r3, iscanmp + + DEFINE_ARGS eob, qcoeff, dqcoeff, iscan + + ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; + + paddsw m6, m1 ; m6 += round + punpckhqdq m1, m1 + paddsw m11, m1 ; m11 += round + pmulhw m8, m6, m2 ; m8 = m6*q>>16 + punpckhqdq m2, m2 + pmulhw m13, m11, m2 ; m13 = m11*q>>16 + paddw m8, m6 ; m8 += m6 + paddw m13, m11 ; m13 += m11 + pmulhw m8, m4 ; m8 = m8*qsh>>16 + punpckhqdq m4, m4 + pmulhw m13, m4 ; m13 = m13*qsh>>16 + psignw m8, m9 ; m8 = reinsert sign + psignw m13, m10 ; m13 = reinsert sign + pand m8, m7 + pand m13, m12 + + ; Store 16bit numbers as 32bit numbers in array pointed to by qcoeff + pcmpgtw m6, m5, m8 + punpckhwd m6, m8, m6 + pmovsxwd m11, m8 + mova [qcoeffq ], m11 + mova [qcoeffq+16], m6 + pcmpgtw m6, m5, m13 + punpckhwd m6, m13, m6 + pmovsxwd m11, m13 + mova [qcoeffq+32], m11 + mova [qcoeffq+48], m6 + + pmullw m8, m3 ; dqc[i] = qc[i] * q + punpckhqdq m3, m3 + pmullw m13, m3 ; dqc[i] = qc[i] * q + + ; Store 16bit numbers as 32bit numbers in array pointed to by qcoeff + pcmpgtw m6, m5, m8 + punpckhwd m6, m8, m6 + pmovsxwd m11, m8 + mova [dqcoeffq ], m11 + mova [dqcoeffq+16], m6 + pcmpgtw m6, m5, m13 + punpckhwd m6, m13, m6 + pmovsxwd m11, m13 + mova [dqcoeffq+32], m11 + mova [dqcoeffq+48], m6 + + mova m6, [iscanq] ; m6 = scan[i] + mova m11, [iscanq+16] ; m11 = scan[i] + + pcmpeqw m8, m8, m5 ; m8 = c[i] == 0 + pcmpeqw m13, m13, m5 ; m13 = c[i] == 0 + psubw m6, m6, m7 ; m6 = scan[i] + 1 + psubw m11, m11, m12 ; m11 = scan[i] + 1 + pandn m8, m8, m6 ; m8 = max(eob) + pandn m13, m13, m11 ; m13 = max(eob) + pmaxsw m8, m8, m13 + + ; Horizontally accumulate/max eobs and write into [eob] memory pointer + pshufd m7, m8, 0xe + pmaxsw m8, m7 + pshuflw m7, m8, 0xe + pmaxsw m8, m7 + pshuflw m7, m8, 0x1 + pmaxsw m8, m7 + movq rax, m8 + mov [eobq], ax + + vzeroupper + RET + + ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; + ;; Generic case of ncoeff != 16 + ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; + +.generic: + +%endif ; %ifnidn %1, b_32x32 + +DEFINE_ARGS coeff, ncoeff, zbin, round, quant, shift, \ + qcoeff, dqcoeff, dequant, eob, scan, iscan + + ; Actual quantization loop - setup pointers, rounders, etc. + movifnidn coeffq, coeffmp + movifnidn ncoeffq, ncoeffmp + movifnidn zbinq, zbinmp + movifnidn roundq, roundmp + movifnidn quantq, quantmp + movifnidn dequantq, dequantmp + mova m0, [zbinq] ; m0 = zbin + mova m1, [roundq] ; m1 = round + mova m2, [quantq] ; m2 = quant + mova m3, [dequantq] ; m3 = dequant + pcmpeqw m4, m4 ; All lanes -1 +%ifidn %1, b_32x32 + psubw m0, m4 + psubw m1, m4 + psrlw m0, 1 ; m0 = (m0 + 1) / 2 + psrlw m1, 1 ; m1 = (m1 + 1) / 2 +%endif + paddw m0, m4 ; m0 = m0 + 1 + + mov r2, shiftmp + mov r3, qcoeffmp + mova m4, [r2] ; m4 = shift + mov r4, dqcoeffmp + mov r5, iscanmp +%ifidn %1, b_32x32 + psllw m4, 1 +%endif + pxor m5, m5 ; m5 = dedicated zero + + DEFINE_ARGS coeff, ncoeff, d1, qcoeff, dqcoeff, iscan, d2, d3, d4, eob + + + lea coeffq, [ coeffq+ncoeffq*4] + lea qcoeffq, [ qcoeffq+ncoeffq*4] + lea dqcoeffq, [dqcoeffq+ncoeffq*4] + + lea iscanq, [ iscanq+ncoeffq*2] + neg ncoeffq + + ; get DC and first 15 AC coeffs + ; coeff stored as 32bit numbers & require 16bit numbers + mova m9, [coeffq+ncoeffq*4+ 0] + packssdw m9, [coeffq+ncoeffq*4+16] + mova m10, [coeffq+ncoeffq*4+32] + packssdw m10, [coeffq+ncoeffq*4+48] + + pabsw m6, m9 ; m6 = abs(m9) + pabsw m11, m10 ; m11 = abs(m10) + pcmpgtw m7, m6, m0 ; m7 = c[i] >= zbin + punpckhqdq m0, m0 + pcmpgtw m12, m11, m0 ; m12 = c[i] >= zbin + + ; Check if all coeffs are less than zbin. If yes, skip forward quickly. + por m14, m7, m12 + ptest m14, m14 + jnz .first_nonzero + + mova [qcoeffq+ncoeffq*4 ], ymm5 + mova [qcoeffq+ncoeffq*4+32], ymm5 + mova [dqcoeffq+ncoeffq*4 ], ymm5 + mova [dqcoeffq+ncoeffq*4+32], ymm5 + add ncoeffq, mmsize + + punpckhqdq m1, m1 + punpckhqdq m2, m2 + punpckhqdq m3, m3 + punpckhqdq m4, m4 + pxor m8, m8 + + jmp .ac_only_loop + +.first_nonzero: + + paddsw m6, m1 ; m6 += round + punpckhqdq m1, m1 + paddsw m11, m1 ; m11 += round + pmulhw m8, m6, m2 ; m8 = m6*q>>16 + punpckhqdq m2, m2 + pmulhw m13, m11, m2 ; m13 = m11*q>>16 + paddw m8, m6 ; m8 += m6 + paddw m13, m11 ; m13 += m11 + pmulhw m8, m4 ; m8 = m8*qsh>>16 + punpckhqdq m4, m4 + pmulhw m13, m4 ; m13 = m13*qsh>>16 + psignw m8, m9 ; m8 = reinsert sign + psignw m13, m10 ; m13 = reinsert sign + pand m8, m7 + pand m13, m12 + + ; store 16bit numbers as 32bit numbers in array pointed to by qcoeff + pcmpgtw m6, m5, m8 + punpckhwd m6, m8, m6 + pmovsxwd m11, m8 + mova [qcoeffq+ncoeffq*4+ 0], m11 + mova [qcoeffq+ncoeffq*4+16], m6 + pcmpgtw m6, m5, m13 + punpckhwd m6, m13, m6 + pmovsxwd m11, m13 + mova [qcoeffq+ncoeffq*4+32], m11 + mova [qcoeffq+ncoeffq*4+48], m6 + +%ifidn %1, b_32x32 + pabsw m8, m8 + pabsw m13, m13 +%endif + pmullw m8, m3 ; dqc[i] = qc[i] * q + punpckhqdq m3, m3 + pmullw m13, m3 ; dqc[i] = qc[i] * q +%ifidn %1, b_32x32 + psrlw m8, 1 + psrlw m13, 1 + psignw m8, m9 + psignw m13, m10 +%endif + + ; store 16bit numbers as 32bit numbers in array pointed to by qcoeff + pcmpgtw m6, m5, m8 + punpckhwd m6, m8, m6 + pmovsxwd m11, m8 + mova [dqcoeffq+ncoeffq*4+ 0], m11 + mova [dqcoeffq+ncoeffq*4+16], m6 + pcmpgtw m6, m5, m13 + punpckhwd m6, m13, m6 + pmovsxwd m11, m13 + mova [dqcoeffq+ncoeffq*4+32], m11 + mova [dqcoeffq+ncoeffq*4+48], m6 + + pcmpeqw m8, m5 ; m8 = c[i] == 0 + pcmpeqw m13, m5 ; m13 = c[i] == 0 + mova m6, [iscanq+ncoeffq*2] ; m6 = scan[i] + mova m11, [iscanq+ncoeffq*2+16] ; m11 = scan[i] + psubw m6, m7 ; m6 = scan[i] + 1 + psubw m11, m12 ; m11 = scan[i] + 1 + pandn m8, m6 ; m8 = max(eob) + pandn m13, m11 ; m13 = max(eob) + pmaxsw m8, m13 + add ncoeffq, mmsize + +.ac_only_loop: + + ; pack coeff from 32bit to 16bit array + mova m9, [coeffq+ncoeffq*4+ 0] + packssdw m9, [coeffq+ncoeffq*4+16] + mova m10, [coeffq+ncoeffq*4+32] + packssdw m10, [coeffq+ncoeffq*4+48] + + pabsw m6, m9 ; m6 = abs(m9) + pabsw m11, m10 ; m11 = abs(m10) + pcmpgtw m7, m6, m0 ; m7 = c[i] >= zbin + pcmpgtw m12, m11, m0 ; m12 = c[i] >= zbin + + ; Check if all coeffs are less than zbin. If yes, skip this itertion. + ; And just write zeros as the result would be. + por m14, m7, m12 + ptest m14, m14 + jnz .rest_nonzero + + mova [qcoeffq+ncoeffq*4+ 0], ymm5 + mova [qcoeffq+ncoeffq*4+32], ymm5 + mova [dqcoeffq+ncoeffq*4+ 0], ymm5 + mova [dqcoeffq+ncoeffq*4+32], ymm5 + + add ncoeffq, mmsize + jnz .ac_only_loop + + ; Horizontally accumulate/max eobs and write into [eob] memory pointer + mov r2, eobmp + pshufd m7, m8, 0xe + pmaxsw m8, m7 + pshuflw m7, m8, 0xe + pmaxsw m8, m7 + pshuflw m7, m8, 0x1 + pmaxsw m8, m7 + movq rax, m8 + mov [r2], ax + vzeroupper + RET + +.rest_nonzero: + paddsw m6, m1 ; m6 += round + paddsw m11, m1 ; m11 += round + pmulhw m14, m6, m2 ; m14 = m6*q>>16 + pmulhw m13, m11, m2 ; m13 = m11*q>>16 + paddw m14, m6 ; m14 += m6 + paddw m13, m11 ; m13 += m11 + pmulhw m14, m4 ; m14 = m14*qsh>>16 + pmulhw m13, m4 ; m13 = m13*qsh>>16 + psignw m14, m9 ; m14 = reinsert sign + psignw m13, m10 ; m13 = reinsert sign + pand m14, m7 + pand m13, m12 + + ; store 16bit numbers as 32bit numbers in array pointed to by qcoeff + pcmpgtw m6, m5, m14 + punpckhwd m6, m14, m6 + pmovsxwd m11, m14 + mova [qcoeffq+ncoeffq*4+ 0], m11 + mova [qcoeffq+ncoeffq*4+16], m6 + pcmpgtw m6, m5, m13 + punpckhwd m6, m13, m6 + pmovsxwd m11, m13 + mova [qcoeffq+ncoeffq*4+32], m11 + mova [qcoeffq+ncoeffq*4+48], m6 + +%ifidn %1, b_32x32 + pabsw m14, m14 + pabsw m13, m13 +%endif + pmullw m14, m3 ; dqc[i] = qc[i] * q + pmullw m13, m3 ; dqc[i] = qc[i] * q +%ifidn %1, b_32x32 + psrlw m14, 1 + psrlw m13, 1 + psignw m14, m9 + psignw m13, m10 +%endif + + ; store 16bit numbers as 32bit numbers in array pointed to by qcoeff + pcmpgtw m6, m5, m14 + punpckhwd m6, m14, m6 + pmovsxwd m11, m14 + mova [dqcoeffq+ncoeffq*4+ 0], m11 + mova [dqcoeffq+ncoeffq*4+16], m6 + pcmpgtw m6, m5, m13 + punpckhwd m6, m13, m6 + pmovsxwd m11, m13 + mova [dqcoeffq+ncoeffq*4+32], m11 + mova [dqcoeffq+ncoeffq*4+48], m6 + + pcmpeqw m14, m5 ; m14 = c[i] == 0 + pcmpeqw m13, m5 ; m13 = c[i] == 0 + mova m6, [iscanq+ncoeffq*2+ 0] ; m6 = scan[i] + mova m11, [iscanq+ncoeffq*2+16] ; m11 = scan[i] + psubw m6, m7 ; m6 = scan[i] + 1 + psubw m11, m12 ; m11 = scan[i] + 1 + pandn m14, m6 ; m14 = max(eob) + pandn m13, m11 ; m13 = max(eob) + pmaxsw m8, m14 + pmaxsw m8, m13 + add ncoeffq, mmsize + jnz .ac_only_loop + + ; Horizontally accumulate/max eobs and write into [eob] memory pointer + mov r2, eobmp + pshufd m7, m8, 0xe + pmaxsw m8, m7 + pshuflw m7, m8, 0xe + pmaxsw m8, m7 + pshuflw m7, m8, 0x1 + pmaxsw m8, m7 + movq rax, m8 + mov [r2], ax + vzeroupper + RET +%endmacro + +INIT_XMM avx +QUANTIZE_FN b, 9 +QUANTIZE_FN b_32x32, 9 diff --git a/third_party/aom/aom_dsp/x86/quantize_sse2.c b/third_party/aom/aom_dsp/x86/quantize_sse2.c new file mode 100644 index 0000000000..d3de6e24db --- /dev/null +++ b/third_party/aom/aom_dsp/x86/quantize_sse2.c @@ -0,0 +1,147 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include +#include +#include + +#include "config/aom_dsp_rtcd.h" + +#include "aom/aom_integer.h" +#include "aom_dsp/x86/quantize_x86.h" + +static INLINE __m128i load_coefficients(const tran_low_t *coeff_ptr) { + assert(sizeof(tran_low_t) == 4); + + return _mm_setr_epi16((int16_t)coeff_ptr[0], (int16_t)coeff_ptr[1], + (int16_t)coeff_ptr[2], (int16_t)coeff_ptr[3], + (int16_t)coeff_ptr[4], (int16_t)coeff_ptr[5], + (int16_t)coeff_ptr[6], (int16_t)coeff_ptr[7]); +} + +static INLINE void store_coefficients(__m128i coeff_vals, + tran_low_t *coeff_ptr) { + assert(sizeof(tran_low_t) == 4); + + __m128i one = _mm_set1_epi16(1); + __m128i coeff_vals_hi = _mm_mulhi_epi16(coeff_vals, one); + __m128i coeff_vals_lo = _mm_mullo_epi16(coeff_vals, one); + __m128i coeff_vals_1 = _mm_unpacklo_epi16(coeff_vals_lo, coeff_vals_hi); + __m128i coeff_vals_2 = _mm_unpackhi_epi16(coeff_vals_lo, coeff_vals_hi); + _mm_store_si128((__m128i *)(coeff_ptr), coeff_vals_1); + _mm_store_si128((__m128i *)(coeff_ptr + 4), coeff_vals_2); +} + +void aom_quantize_b_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, + const int16_t *zbin_ptr, const int16_t *round_ptr, + const int16_t *quant_ptr, + const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, + tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, + uint16_t *eob_ptr, const int16_t *scan_ptr, + const int16_t *iscan_ptr) { + const __m128i zero = _mm_setzero_si128(); + int index = 16; + + __m128i zbin, round, quant, dequant, shift; + __m128i coeff0, coeff1, coeff0_sign, coeff1_sign; + __m128i qcoeff0, qcoeff1; + __m128i cmp_mask0, cmp_mask1; + __m128i eob, eob0; + + (void)scan_ptr; + + // Setup global values. + load_b_values(zbin_ptr, &zbin, round_ptr, &round, quant_ptr, &quant, + dequant_ptr, &dequant, quant_shift_ptr, &shift); + + // Do DC and first 15 AC. + coeff0 = load_coefficients(coeff_ptr); + coeff1 = load_coefficients(coeff_ptr + 8); + + // Poor man's abs(). + coeff0_sign = _mm_srai_epi16(coeff0, 15); + coeff1_sign = _mm_srai_epi16(coeff1, 15); + qcoeff0 = invert_sign_sse2(coeff0, coeff0_sign); + qcoeff1 = invert_sign_sse2(coeff1, coeff1_sign); + + cmp_mask0 = _mm_cmpgt_epi16(qcoeff0, zbin); + zbin = _mm_unpackhi_epi64(zbin, zbin); // Switch DC to AC + cmp_mask1 = _mm_cmpgt_epi16(qcoeff1, zbin); + + calculate_qcoeff(&qcoeff0, round, quant, shift); + + round = _mm_unpackhi_epi64(round, round); + quant = _mm_unpackhi_epi64(quant, quant); + shift = _mm_unpackhi_epi64(shift, shift); + + calculate_qcoeff(&qcoeff1, round, quant, shift); + + // Reinsert signs + qcoeff0 = invert_sign_sse2(qcoeff0, coeff0_sign); + qcoeff1 = invert_sign_sse2(qcoeff1, coeff1_sign); + + // Mask out zbin threshold coeffs + qcoeff0 = _mm_and_si128(qcoeff0, cmp_mask0); + qcoeff1 = _mm_and_si128(qcoeff1, cmp_mask1); + + store_coefficients(qcoeff0, qcoeff_ptr); + store_coefficients(qcoeff1, qcoeff_ptr + 8); + + coeff0 = calculate_dqcoeff(qcoeff0, dequant); + dequant = _mm_unpackhi_epi64(dequant, dequant); + coeff1 = calculate_dqcoeff(qcoeff1, dequant); + + store_coefficients(coeff0, dqcoeff_ptr); + store_coefficients(coeff1, dqcoeff_ptr + 8); + + eob = + scan_for_eob(&coeff0, &coeff1, cmp_mask0, cmp_mask1, iscan_ptr, 0, zero); + + // AC only loop. + while (index < n_coeffs) { + coeff0 = load_coefficients(coeff_ptr + index); + coeff1 = load_coefficients(coeff_ptr + index + 8); + + coeff0_sign = _mm_srai_epi16(coeff0, 15); + coeff1_sign = _mm_srai_epi16(coeff1, 15); + qcoeff0 = invert_sign_sse2(coeff0, coeff0_sign); + qcoeff1 = invert_sign_sse2(coeff1, coeff1_sign); + + cmp_mask0 = _mm_cmpgt_epi16(qcoeff0, zbin); + cmp_mask1 = _mm_cmpgt_epi16(qcoeff1, zbin); + + calculate_qcoeff(&qcoeff0, round, quant, shift); + calculate_qcoeff(&qcoeff1, round, quant, shift); + + qcoeff0 = invert_sign_sse2(qcoeff0, coeff0_sign); + qcoeff1 = invert_sign_sse2(qcoeff1, coeff1_sign); + + qcoeff0 = _mm_and_si128(qcoeff0, cmp_mask0); + qcoeff1 = _mm_and_si128(qcoeff1, cmp_mask1); + + store_coefficients(qcoeff0, qcoeff_ptr + index); + store_coefficients(qcoeff1, qcoeff_ptr + index + 8); + + coeff0 = calculate_dqcoeff(qcoeff0, dequant); + coeff1 = calculate_dqcoeff(qcoeff1, dequant); + + store_coefficients(coeff0, dqcoeff_ptr + index); + store_coefficients(coeff1, dqcoeff_ptr + index + 8); + + eob0 = scan_for_eob(&coeff0, &coeff1, cmp_mask0, cmp_mask1, iscan_ptr, + index, zero); + eob = _mm_max_epi16(eob, eob0); + + index += 16; + } + + *eob_ptr = accumulate_eob(eob); +} diff --git a/third_party/aom/aom_dsp/x86/quantize_ssse3_x86_64.asm b/third_party/aom/aom_dsp/x86/quantize_ssse3_x86_64.asm new file mode 100644 index 0000000000..39d4ca674c --- /dev/null +++ b/third_party/aom/aom_dsp/x86/quantize_ssse3_x86_64.asm @@ -0,0 +1,272 @@ +; +; Copyright (c) 2016, Alliance for Open Media. All rights reserved +; +; This source code is subject to the terms of the BSD 2 Clause License and +; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License +; was not distributed with this source code in the LICENSE file, you can +; obtain it at www.aomedia.org/license/software. If the Alliance for Open +; Media Patent License 1.0 was not distributed with this source code in the +; PATENTS file, you can obtain it at www.aomedia.org/license/patent. +; + +; + +%include "third_party/x86inc/x86inc.asm" + +SECTION_RODATA +pw_1: times 8 dw 1 + +SECTION .text + +%macro QUANTIZE_FN 2 +cglobal quantize_%1, 0, %2, 15, coeff, ncoeff, zbin, round, quant, \ + shift, qcoeff, dqcoeff, dequant, \ + eob, scan, iscan + + ; actual quantize loop - setup pointers, rounders, etc. + movifnidn coeffq, coeffmp + movifnidn ncoeffq, ncoeffmp + movifnidn zbinq, zbinmp + movifnidn roundq, roundmp + movifnidn quantq, quantmp + movifnidn dequantq, dequantmp + mova m0, [zbinq] ; m0 = zbin + mova m1, [roundq] ; m1 = round + mova m2, [quantq] ; m2 = quant +%ifidn %1, b_32x32 + pcmpeqw m5, m5 + psrlw m5, 15 + paddw m0, m5 + paddw m1, m5 + psrlw m0, 1 ; m0 = (m0 + 1) / 2 + psrlw m1, 1 ; m1 = (m1 + 1) / 2 +%endif + mova m3, [dequantq] ; m3 = dequant + mov r2, shiftmp + psubw m0, [GLOBAL(pw_1)] + mova m4, [r2] ; m4 = shift + mov r3, qcoeffmp + mov r4, dqcoeffmp + mov r5, iscanmp +%ifidn %1, b_32x32 + psllw m4, 1 +%endif + pxor m5, m5 ; m5 = dedicated zero + DEFINE_ARGS coeff, ncoeff, d1, qcoeff, dqcoeff, iscan, d2, d3, d4, eob + lea coeffq, [ coeffq+ncoeffq*4] + lea qcoeffq, [ qcoeffq+ncoeffq*4] + lea dqcoeffq, [dqcoeffq+ncoeffq*4] + lea iscanq, [ iscanq+ncoeffq*2] + neg ncoeffq + + ; get DC and first 15 AC coeffs + ; coeff stored as 32bit numbers & require 16bit numbers + mova m9, [ coeffq+ncoeffq*4+ 0] + packssdw m9, [ coeffq+ncoeffq*4+16] + mova m10, [ coeffq+ncoeffq*4+32] + packssdw m10, [ coeffq+ncoeffq*4+48] + pabsw m6, m9 ; m6 = abs(m9) + pabsw m11, m10 ; m11 = abs(m10) + pcmpgtw m7, m6, m0 ; m7 = c[i] >= zbin + punpckhqdq m0, m0 + pcmpgtw m12, m11, m0 ; m12 = c[i] >= zbin + paddsw m6, m1 ; m6 += round + punpckhqdq m1, m1 + paddsw m11, m1 ; m11 += round + pmulhw m8, m6, m2 ; m8 = m6*q>>16 + punpckhqdq m2, m2 + pmulhw m13, m11, m2 ; m13 = m11*q>>16 + paddw m8, m6 ; m8 += m6 + paddw m13, m11 ; m13 += m11 + pmulhw m8, m4 ; m8 = m8*qsh>>16 + punpckhqdq m4, m4 + pmulhw m13, m4 ; m13 = m13*qsh>>16 + psignw m8, m9 ; m8 = reinsert sign + psignw m13, m10 ; m13 = reinsert sign + pand m8, m7 + pand m13, m12 + + ; store 16bit numbers as 32bit numbers in array pointed to by qcoeff + mova m11, m8 + mova m6, m8 + pcmpgtw m5, m8 + punpcklwd m11, m5 + punpckhwd m6, m5 + mova [qcoeffq+ncoeffq*4+ 0], m11 + mova [qcoeffq+ncoeffq*4+16], m6 + pxor m5, m5 + mova m11, m13 + mova m6, m13 + pcmpgtw m5, m13 + punpcklwd m11, m5 + punpckhwd m6, m5 + mova [qcoeffq+ncoeffq*4+32], m11 + mova [qcoeffq+ncoeffq*4+48], m6 + pxor m5, m5 ; reset m5 to zero register + +%ifidn %1, b_32x32 + pabsw m8, m8 + pabsw m13, m13 +%endif + pmullw m8, m3 ; dqc[i] = qc[i] * q + punpckhqdq m3, m3 + pmullw m13, m3 ; dqc[i] = qc[i] * q +%ifidn %1, b_32x32 + psrlw m8, 1 + psrlw m13, 1 + psignw m8, m9 + psignw m13, m10 +%endif + ; store 16bit numbers as 32bit numbers in array pointed to by qcoeff + mova m11, m8 + mova m6, m8 + pcmpgtw m5, m8 + punpcklwd m11, m5 + punpckhwd m6, m5 + mova [dqcoeffq+ncoeffq*4+ 0], m11 + mova [dqcoeffq+ncoeffq*4+16], m6 + pxor m5, m5 + mova m11, m13 + mova m6, m13 + pcmpgtw m5, m13 + punpcklwd m11, m5 + punpckhwd m6, m5 + mova [dqcoeffq+ncoeffq*4+32], m11 + mova [dqcoeffq+ncoeffq*4+48], m6 + pxor m5, m5 ; reset m5 to zero register + pcmpeqw m8, m5 ; m8 = c[i] == 0 + pcmpeqw m13, m5 ; m13 = c[i] == 0 + mova m6, [ iscanq+ncoeffq*2+ 0] ; m6 = scan[i] + mova m11, [ iscanq+ncoeffq*2+16] ; m11 = scan[i] + psubw m6, m7 ; m6 = scan[i] + 1 + psubw m11, m12 ; m11 = scan[i] + 1 + pandn m8, m6 ; m8 = max(eob) + pandn m13, m11 ; m13 = max(eob) + pmaxsw m8, m13 + add ncoeffq, mmsize + jz .accumulate_eob + +.ac_only_loop: + ; pack coeff from 32bit to 16bit array + mova m9, [ coeffq+ncoeffq*4+ 0] + packssdw m9, [ coeffq+ncoeffq*4+16] + mova m10, [ coeffq+ncoeffq*4+32] + packssdw m10, [ coeffq+ncoeffq*4+48] + + pabsw m6, m9 ; m6 = abs(m9) + pabsw m11, m10 ; m11 = abs(m10) + pcmpgtw m7, m6, m0 ; m7 = c[i] >= zbin + pcmpgtw m12, m11, m0 ; m12 = c[i] >= zbin +%ifidn %1, b_32x32 + pmovmskb r6d, m7 + pmovmskb r2d, m12 + or r6, r2 + jz .skip_iter +%endif + paddsw m6, m1 ; m6 += round + paddsw m11, m1 ; m11 += round + pmulhw m14, m6, m2 ; m14 = m6*q>>16 + pmulhw m13, m11, m2 ; m13 = m11*q>>16 + paddw m14, m6 ; m14 += m6 + paddw m13, m11 ; m13 += m11 + pmulhw m14, m4 ; m14 = m14*qsh>>16 + pmulhw m13, m4 ; m13 = m13*qsh>>16 + psignw m14, m9 ; m14 = reinsert sign + psignw m13, m10 ; m13 = reinsert sign + pand m14, m7 + pand m13, m12 + ; store 16bit numbers as 32bit numbers in array pointed to by qcoeff + pxor m11, m11 + mova m11, m14 + mova m6, m14 + pcmpgtw m5, m14 + punpcklwd m11, m5 + punpckhwd m6, m5 + mova [qcoeffq+ncoeffq*4+ 0], m11 + mova [qcoeffq+ncoeffq*4+16], m6 + pxor m5, m5 + mova m11, m13 + mova m6, m13 + pcmpgtw m5, m13 + punpcklwd m11, m5 + punpckhwd m6, m5 + mova [qcoeffq+ncoeffq*4+32], m11 + mova [qcoeffq+ncoeffq*4+48], m6 + pxor m5, m5 ; reset m5 to zero register + +%ifidn %1, b_32x32 + pabsw m14, m14 + pabsw m13, m13 +%endif + pmullw m14, m3 ; dqc[i] = qc[i] * q + pmullw m13, m3 ; dqc[i] = qc[i] * q +%ifidn %1, b_32x32 + psrlw m14, 1 + psrlw m13, 1 + psignw m14, m9 + psignw m13, m10 +%endif + + ; store 16bit numbers as 32bit numbers in array pointed to by qcoeff + mova m11, m14 + mova m6, m14 + pcmpgtw m5, m14 + punpcklwd m11, m5 + punpckhwd m6, m5 + mova [dqcoeffq+ncoeffq*4+ 0], m11 + mova [dqcoeffq+ncoeffq*4+16], m6 + pxor m5, m5 + mova m11, m13 + mova m6, m13 + pcmpgtw m5, m13 + punpcklwd m11, m5 + punpckhwd m6, m5 + mova [dqcoeffq+ncoeffq*4+32], m11 + mova [dqcoeffq+ncoeffq*4+48], m6 + pxor m5, m5 + + pcmpeqw m14, m5 ; m14 = c[i] == 0 + pcmpeqw m13, m5 ; m13 = c[i] == 0 + mova m6, [ iscanq+ncoeffq*2+ 0] ; m6 = scan[i] + mova m11, [ iscanq+ncoeffq*2+16] ; m11 = scan[i] + psubw m6, m7 ; m6 = scan[i] + 1 + psubw m11, m12 ; m11 = scan[i] + 1 + pandn m14, m6 ; m14 = max(eob) + pandn m13, m11 ; m13 = max(eob) + pmaxsw m8, m14 + pmaxsw m8, m13 + add ncoeffq, mmsize + jl .ac_only_loop + +%ifidn %1, b_32x32 + jmp .accumulate_eob +.skip_iter: + mova [qcoeffq+ncoeffq*4+ 0], m5 + mova [qcoeffq+ncoeffq*4+16], m5 + mova [qcoeffq+ncoeffq*4+32], m5 + mova [qcoeffq+ncoeffq*4+48], m5 + mova [dqcoeffq+ncoeffq*4+ 0], m5 + mova [dqcoeffq+ncoeffq*4+16], m5 + mova [dqcoeffq+ncoeffq*4+32], m5 + mova [dqcoeffq+ncoeffq*4+48], m5 + add ncoeffq, mmsize + jl .ac_only_loop +%endif + +.accumulate_eob: + ; horizontally accumulate/max eobs and write into [eob] memory pointer + mov r2, eobmp + pshufd m7, m8, 0xe + pmaxsw m8, m7 + pshuflw m7, m8, 0xe + pmaxsw m8, m7 + pshuflw m7, m8, 0x1 + pmaxsw m8, m7 + pextrw r6, m8, 0 + mov [r2], r6 + RET +%endmacro + +INIT_XMM ssse3 +QUANTIZE_FN b, 9 +QUANTIZE_FN b_32x32, 9 diff --git a/third_party/aom/aom_dsp/x86/quantize_x86.h b/third_party/aom/aom_dsp/x86/quantize_x86.h new file mode 100644 index 0000000000..4eed7dd29a --- /dev/null +++ b/third_party/aom/aom_dsp/x86/quantize_x86.h @@ -0,0 +1,77 @@ +/* + * Copyright (c) 2018, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include + +#include "aom/aom_integer.h" + +static INLINE void load_b_values(const int16_t *zbin_ptr, __m128i *zbin, + const int16_t *round_ptr, __m128i *round, + const int16_t *quant_ptr, __m128i *quant, + const int16_t *dequant_ptr, __m128i *dequant, + const int16_t *shift_ptr, __m128i *shift) { + *zbin = _mm_load_si128((const __m128i *)zbin_ptr); + *round = _mm_load_si128((const __m128i *)round_ptr); + *quant = _mm_load_si128((const __m128i *)quant_ptr); + *zbin = _mm_sub_epi16(*zbin, _mm_set1_epi16(1)); + *dequant = _mm_load_si128((const __m128i *)dequant_ptr); + *shift = _mm_load_si128((const __m128i *)shift_ptr); +} + +// With ssse3 and later abs() and sign() are preferred. +static INLINE __m128i invert_sign_sse2(__m128i a, __m128i sign) { + a = _mm_xor_si128(a, sign); + return _mm_sub_epi16(a, sign); +} + +static INLINE void calculate_qcoeff(__m128i *coeff, const __m128i round, + const __m128i quant, const __m128i shift) { + __m128i tmp, qcoeff; + qcoeff = _mm_adds_epi16(*coeff, round); + tmp = _mm_mulhi_epi16(qcoeff, quant); + qcoeff = _mm_add_epi16(tmp, qcoeff); + *coeff = _mm_mulhi_epi16(qcoeff, shift); +} + +static INLINE __m128i calculate_dqcoeff(__m128i qcoeff, __m128i dequant) { + return _mm_mullo_epi16(qcoeff, dequant); +} + +// Scan 16 values for eob reference in scan_ptr. Use masks (-1) from comparing +// to zbin to add 1 to the index in 'scan'. +static INLINE __m128i scan_for_eob(__m128i *coeff0, __m128i *coeff1, + const __m128i zbin_mask0, + const __m128i zbin_mask1, + const int16_t *scan_ptr, const int index, + const __m128i zero) { + const __m128i zero_coeff0 = _mm_cmpeq_epi16(*coeff0, zero); + const __m128i zero_coeff1 = _mm_cmpeq_epi16(*coeff1, zero); + __m128i scan0 = _mm_load_si128((const __m128i *)(scan_ptr + index)); + __m128i scan1 = _mm_load_si128((const __m128i *)(scan_ptr + index + 8)); + __m128i eob0, eob1; + // Add one to convert from indices to counts + scan0 = _mm_sub_epi16(scan0, zbin_mask0); + scan1 = _mm_sub_epi16(scan1, zbin_mask1); + eob0 = _mm_andnot_si128(zero_coeff0, scan0); + eob1 = _mm_andnot_si128(zero_coeff1, scan1); + return _mm_max_epi16(eob0, eob1); +} + +static INLINE int16_t accumulate_eob(__m128i eob) { + __m128i eob_shuffled; + eob_shuffled = _mm_shuffle_epi32(eob, 0xe); + eob = _mm_max_epi16(eob, eob_shuffled); + eob_shuffled = _mm_shufflelo_epi16(eob, 0xe); + eob = _mm_max_epi16(eob, eob_shuffled); + eob_shuffled = _mm_shufflelo_epi16(eob, 0x1); + eob = _mm_max_epi16(eob, eob_shuffled); + return _mm_extract_epi16(eob, 1); +} diff --git a/third_party/aom/aom_dsp/x86/sad4d_avx2.c b/third_party/aom/aom_dsp/x86/sad4d_avx2.c new file mode 100644 index 0000000000..f662b62b16 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/sad4d_avx2.c @@ -0,0 +1,218 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ +#include // AVX2 + +#include "config/aom_dsp_rtcd.h" + +#include "aom/aom_integer.h" + +void aom_sad32x32x4d_avx2(const uint8_t *src, int src_stride, + const uint8_t *const ref[4], int ref_stride, + uint32_t res[4]) { + __m256i src_reg, ref0_reg, ref1_reg, ref2_reg, ref3_reg; + __m256i sum_ref0, sum_ref1, sum_ref2, sum_ref3; + __m256i sum_mlow, sum_mhigh; + int i; + const uint8_t *ref0, *ref1, *ref2, *ref3; + + ref0 = ref[0]; + ref1 = ref[1]; + ref2 = ref[2]; + ref3 = ref[3]; + sum_ref0 = _mm256_set1_epi16(0); + sum_ref1 = _mm256_set1_epi16(0); + sum_ref2 = _mm256_set1_epi16(0); + sum_ref3 = _mm256_set1_epi16(0); + for (i = 0; i < 32; i++) { + // load src and all refs + src_reg = _mm256_loadu_si256((const __m256i *)src); + ref0_reg = _mm256_loadu_si256((const __m256i *)ref0); + ref1_reg = _mm256_loadu_si256((const __m256i *)ref1); + ref2_reg = _mm256_loadu_si256((const __m256i *)ref2); + ref3_reg = _mm256_loadu_si256((const __m256i *)ref3); + // sum of the absolute differences between every ref-i to src + ref0_reg = _mm256_sad_epu8(ref0_reg, src_reg); + ref1_reg = _mm256_sad_epu8(ref1_reg, src_reg); + ref2_reg = _mm256_sad_epu8(ref2_reg, src_reg); + ref3_reg = _mm256_sad_epu8(ref3_reg, src_reg); + // sum every ref-i + sum_ref0 = _mm256_add_epi32(sum_ref0, ref0_reg); + sum_ref1 = _mm256_add_epi32(sum_ref1, ref1_reg); + sum_ref2 = _mm256_add_epi32(sum_ref2, ref2_reg); + sum_ref3 = _mm256_add_epi32(sum_ref3, ref3_reg); + + src += src_stride; + ref0 += ref_stride; + ref1 += ref_stride; + ref2 += ref_stride; + ref3 += ref_stride; + } + { + __m128i sum; + // in sum_ref-i the result is saved in the first 4 bytes + // the other 4 bytes are zeroed. + // sum_ref1 and sum_ref3 are shifted left by 4 bytes + sum_ref1 = _mm256_slli_si256(sum_ref1, 4); + sum_ref3 = _mm256_slli_si256(sum_ref3, 4); + + // merge sum_ref0 and sum_ref1 also sum_ref2 and sum_ref3 + sum_ref0 = _mm256_or_si256(sum_ref0, sum_ref1); + sum_ref2 = _mm256_or_si256(sum_ref2, sum_ref3); + + // merge every 64 bit from each sum_ref-i + sum_mlow = _mm256_unpacklo_epi64(sum_ref0, sum_ref2); + sum_mhigh = _mm256_unpackhi_epi64(sum_ref0, sum_ref2); + + // add the low 64 bit to the high 64 bit + sum_mlow = _mm256_add_epi32(sum_mlow, sum_mhigh); + + // add the low 128 bit to the high 128 bit + sum = _mm_add_epi32(_mm256_castsi256_si128(sum_mlow), + _mm256_extractf128_si256(sum_mlow, 1)); + + _mm_storeu_si128((__m128i *)(res), sum); + } + _mm256_zeroupper(); +} + +void aom_sad64x64x4d_avx2(const uint8_t *src, int src_stride, + const uint8_t *const ref[4], int ref_stride, + uint32_t res[4]) { + __m256i src_reg, srcnext_reg, ref0_reg, ref0next_reg; + __m256i ref1_reg, ref1next_reg, ref2_reg, ref2next_reg; + __m256i ref3_reg, ref3next_reg; + __m256i sum_ref0, sum_ref1, sum_ref2, sum_ref3; + __m256i sum_mlow, sum_mhigh; + int i; + const uint8_t *ref0, *ref1, *ref2, *ref3; + + ref0 = ref[0]; + ref1 = ref[1]; + ref2 = ref[2]; + ref3 = ref[3]; + sum_ref0 = _mm256_set1_epi16(0); + sum_ref1 = _mm256_set1_epi16(0); + sum_ref2 = _mm256_set1_epi16(0); + sum_ref3 = _mm256_set1_epi16(0); + for (i = 0; i < 64; i++) { + // load 64 bytes from src and all refs + src_reg = _mm256_loadu_si256((const __m256i *)src); + srcnext_reg = _mm256_loadu_si256((const __m256i *)(src + 32)); + ref0_reg = _mm256_loadu_si256((const __m256i *)ref0); + ref0next_reg = _mm256_loadu_si256((const __m256i *)(ref0 + 32)); + ref1_reg = _mm256_loadu_si256((const __m256i *)ref1); + ref1next_reg = _mm256_loadu_si256((const __m256i *)(ref1 + 32)); + ref2_reg = _mm256_loadu_si256((const __m256i *)ref2); + ref2next_reg = _mm256_loadu_si256((const __m256i *)(ref2 + 32)); + ref3_reg = _mm256_loadu_si256((const __m256i *)ref3); + ref3next_reg = _mm256_loadu_si256((const __m256i *)(ref3 + 32)); + // sum of the absolute differences between every ref-i to src + ref0_reg = _mm256_sad_epu8(ref0_reg, src_reg); + ref1_reg = _mm256_sad_epu8(ref1_reg, src_reg); + ref2_reg = _mm256_sad_epu8(ref2_reg, src_reg); + ref3_reg = _mm256_sad_epu8(ref3_reg, src_reg); + ref0next_reg = _mm256_sad_epu8(ref0next_reg, srcnext_reg); + ref1next_reg = _mm256_sad_epu8(ref1next_reg, srcnext_reg); + ref2next_reg = _mm256_sad_epu8(ref2next_reg, srcnext_reg); + ref3next_reg = _mm256_sad_epu8(ref3next_reg, srcnext_reg); + + // sum every ref-i + sum_ref0 = _mm256_add_epi32(sum_ref0, ref0_reg); + sum_ref1 = _mm256_add_epi32(sum_ref1, ref1_reg); + sum_ref2 = _mm256_add_epi32(sum_ref2, ref2_reg); + sum_ref3 = _mm256_add_epi32(sum_ref3, ref3_reg); + sum_ref0 = _mm256_add_epi32(sum_ref0, ref0next_reg); + sum_ref1 = _mm256_add_epi32(sum_ref1, ref1next_reg); + sum_ref2 = _mm256_add_epi32(sum_ref2, ref2next_reg); + sum_ref3 = _mm256_add_epi32(sum_ref3, ref3next_reg); + src += src_stride; + ref0 += ref_stride; + ref1 += ref_stride; + ref2 += ref_stride; + ref3 += ref_stride; + } + { + __m128i sum; + + // in sum_ref-i the result is saved in the first 4 bytes + // the other 4 bytes are zeroed. + // sum_ref1 and sum_ref3 are shifted left by 4 bytes + sum_ref1 = _mm256_slli_si256(sum_ref1, 4); + sum_ref3 = _mm256_slli_si256(sum_ref3, 4); + + // merge sum_ref0 and sum_ref1 also sum_ref2 and sum_ref3 + sum_ref0 = _mm256_or_si256(sum_ref0, sum_ref1); + sum_ref2 = _mm256_or_si256(sum_ref2, sum_ref3); + + // merge every 64 bit from each sum_ref-i + sum_mlow = _mm256_unpacklo_epi64(sum_ref0, sum_ref2); + sum_mhigh = _mm256_unpackhi_epi64(sum_ref0, sum_ref2); + + // add the low 64 bit to the high 64 bit + sum_mlow = _mm256_add_epi32(sum_mlow, sum_mhigh); + + // add the low 128 bit to the high 128 bit + sum = _mm_add_epi32(_mm256_castsi256_si128(sum_mlow), + _mm256_extractf128_si256(sum_mlow, 1)); + + _mm_storeu_si128((__m128i *)(res), sum); + } + _mm256_zeroupper(); +} + +void aom_sad32x64x4d_avx2(const uint8_t *src, int src_stride, + const uint8_t *const ref[4], int ref_stride, + uint32_t res[4]) { + const uint8_t *rf[4]; + uint32_t sum0[4]; + uint32_t sum1[4]; + + rf[0] = ref[0]; + rf[1] = ref[1]; + rf[2] = ref[2]; + rf[3] = ref[3]; + aom_sad32x32x4d_avx2(src, src_stride, rf, ref_stride, sum0); + src += src_stride << 5; + rf[0] += ref_stride << 5; + rf[1] += ref_stride << 5; + rf[2] += ref_stride << 5; + rf[3] += ref_stride << 5; + aom_sad32x32x4d_avx2(src, src_stride, rf, ref_stride, sum1); + res[0] = sum0[0] + sum1[0]; + res[1] = sum0[1] + sum1[1]; + res[2] = sum0[2] + sum1[2]; + res[3] = sum0[3] + sum1[3]; +} + +void aom_sad64x32x4d_avx2(const uint8_t *src, int src_stride, + const uint8_t *const ref[4], int ref_stride, + uint32_t res[4]) { + const uint8_t *rf[4]; + uint32_t sum0[4]; + uint32_t sum1[4]; + unsigned int half_width = 32; + + rf[0] = ref[0]; + rf[1] = ref[1]; + rf[2] = ref[2]; + rf[3] = ref[3]; + aom_sad32x32x4d_avx2(src, src_stride, rf, ref_stride, sum0); + src += half_width; + rf[0] += half_width; + rf[1] += half_width; + rf[2] += half_width; + rf[3] += half_width; + aom_sad32x32x4d_avx2(src, src_stride, rf, ref_stride, sum1); + res[0] = sum0[0] + sum1[0]; + res[1] = sum0[1] + sum1[1]; + res[2] = sum0[2] + sum1[2]; + res[3] = sum0[3] + sum1[3]; +} diff --git a/third_party/aom/aom_dsp/x86/sad4d_sse2.asm b/third_party/aom/aom_dsp/x86/sad4d_sse2.asm new file mode 100644 index 0000000000..55a856985a --- /dev/null +++ b/third_party/aom/aom_dsp/x86/sad4d_sse2.asm @@ -0,0 +1,257 @@ +; +; Copyright (c) 2016, Alliance for Open Media. All rights reserved +; +; This source code is subject to the terms of the BSD 2 Clause License and +; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License +; was not distributed with this source code in the LICENSE file, you can +; obtain it at www.aomedia.org/license/software. If the Alliance for Open +; Media Patent License 1.0 was not distributed with this source code in the +; PATENTS file, you can obtain it at www.aomedia.org/license/patent. +; + +; + +%include "third_party/x86inc/x86inc.asm" + +SECTION .text + +; PROCESS_4x2x4 first, off_{first,second}_{src,ref}, advance_at_end +%macro PROCESS_4x2x4 5-6 0 + movd m0, [srcq +%2] +%if %1 == 1 + movd m6, [ref1q+%3] + movd m4, [ref2q+%3] + movd m7, [ref3q+%3] + movd m5, [ref4q+%3] + movd m1, [srcq +%4] + movd m2, [ref1q+%5] + punpckldq m0, m1 + punpckldq m6, m2 + movd m1, [ref2q+%5] + movd m2, [ref3q+%5] + movd m3, [ref4q+%5] + punpckldq m4, m1 + punpckldq m7, m2 + punpckldq m5, m3 + movlhps m0, m0 + movlhps m6, m4 + movlhps m7, m5 + psadbw m6, m0 + psadbw m7, m0 +%else + movd m1, [ref1q+%3] + movd m5, [ref1q+%5] + movd m2, [ref2q+%3] + movd m4, [ref2q+%5] + punpckldq m1, m5 + punpckldq m2, m4 + movd m3, [ref3q+%3] + movd m5, [ref3q+%5] + punpckldq m3, m5 + movd m4, [ref4q+%3] + movd m5, [ref4q+%5] + punpckldq m4, m5 + movd m5, [srcq +%4] + punpckldq m0, m5 + movlhps m0, m0 + movlhps m1, m2 + movlhps m3, m4 + psadbw m1, m0 + psadbw m3, m0 + paddd m6, m1 + paddd m7, m3 +%endif +%if %6 == 1 + lea srcq, [srcq +src_strideq*2] + lea ref1q, [ref1q+ref_strideq*2] + lea ref2q, [ref2q+ref_strideq*2] + lea ref3q, [ref3q+ref_strideq*2] + lea ref4q, [ref4q+ref_strideq*2] +%endif +%endmacro + +; PROCESS_8x2x4 first, off_{first,second}_{src,ref}, advance_at_end +%macro PROCESS_8x2x4 5-6 0 + movh m0, [srcq +%2] +%if %1 == 1 + movh m4, [ref1q+%3] + movh m5, [ref2q+%3] + movh m6, [ref3q+%3] + movh m7, [ref4q+%3] + movhps m0, [srcq +%4] + movhps m4, [ref1q+%5] + movhps m5, [ref2q+%5] + movhps m6, [ref3q+%5] + movhps m7, [ref4q+%5] + psadbw m4, m0 + psadbw m5, m0 + psadbw m6, m0 + psadbw m7, m0 +%else + movh m1, [ref1q+%3] + movh m2, [ref2q+%3] + movh m3, [ref3q+%3] + movhps m0, [srcq +%4] + movhps m1, [ref1q+%5] + movhps m2, [ref2q+%5] + movhps m3, [ref3q+%5] + psadbw m1, m0 + psadbw m2, m0 + psadbw m3, m0 + paddd m4, m1 + movh m1, [ref4q+%3] + movhps m1, [ref4q+%5] + paddd m5, m2 + paddd m6, m3 + psadbw m1, m0 + paddd m7, m1 +%endif +%if %6 == 1 + lea srcq, [srcq +src_strideq*2] + lea ref1q, [ref1q+ref_strideq*2] + lea ref2q, [ref2q+ref_strideq*2] + lea ref3q, [ref3q+ref_strideq*2] + lea ref4q, [ref4q+ref_strideq*2] +%endif +%endmacro + +; PROCESS_16x2x4 first, off_{first,second}_{src,ref}, advance_at_end +%macro PROCESS_16x2x4 5-6 0 + ; 1st 16 px + mova m0, [srcq +%2] +%if %1 == 1 + movu m4, [ref1q+%3] + movu m5, [ref2q+%3] + movu m6, [ref3q+%3] + movu m7, [ref4q+%3] + psadbw m4, m0 + psadbw m5, m0 + psadbw m6, m0 + psadbw m7, m0 +%else + movu m1, [ref1q+%3] + movu m2, [ref2q+%3] + movu m3, [ref3q+%3] + psadbw m1, m0 + psadbw m2, m0 + psadbw m3, m0 + paddd m4, m1 + movu m1, [ref4q+%3] + paddd m5, m2 + paddd m6, m3 + psadbw m1, m0 + paddd m7, m1 +%endif + + ; 2nd 16 px + mova m0, [srcq +%4] + movu m1, [ref1q+%5] + movu m2, [ref2q+%5] + movu m3, [ref3q+%5] + psadbw m1, m0 + psadbw m2, m0 + psadbw m3, m0 + paddd m4, m1 + movu m1, [ref4q+%5] + paddd m5, m2 + paddd m6, m3 +%if %6 == 1 + lea srcq, [srcq +src_strideq*2] + lea ref1q, [ref1q+ref_strideq*2] + lea ref2q, [ref2q+ref_strideq*2] + lea ref3q, [ref3q+ref_strideq*2] + lea ref4q, [ref4q+ref_strideq*2] +%endif + psadbw m1, m0 + paddd m7, m1 +%endmacro + +; PROCESS_32x2x4 first, off_{first,second}_{src,ref}, advance_at_end +%macro PROCESS_32x2x4 5-6 0 + PROCESS_16x2x4 %1, %2, %3, %2 + 16, %3 + 16 + PROCESS_16x2x4 0, %4, %5, %4 + 16, %5 + 16, %6 +%endmacro + +; PROCESS_64x2x4 first, off_{first,second}_{src,ref}, advance_at_end +%macro PROCESS_64x2x4 5-6 0 + PROCESS_32x2x4 %1, %2, %3, %2 + 32, %3 + 32 + PROCESS_32x2x4 0, %4, %5, %4 + 32, %5 + 32, %6 +%endmacro + +; PROCESS_128x2x4 first, off_{first,second}_{src,ref}, advance_at_end +%macro PROCESS_128x2x4 5-6 0 + PROCESS_64x2x4 %1, %2, %3, %2 + 64, %3 + 64 + PROCESS_64x2x4 0, %4, %5, %4 + 64, %5 + 64, %6 +%endmacro + +; void aom_sadNxNx4d_sse2(uint8_t *src, int src_stride, +; uint8_t *ref[4], int ref_stride, +; uint32_t res[4]); +; where NxN = 64x64, 32x32, 16x16, 16x8, 8x16, 8x8, 8x4, 4x8 and 4x4 +%macro SADNXN4D 2 +%if UNIX64 +cglobal sad%1x%2x4d, 5, 8, 8, src, src_stride, ref1, ref_stride, \ + res, ref2, ref3, ref4 +%else +cglobal sad%1x%2x4d, 4, 7, 8, src, src_stride, ref1, ref_stride, \ + ref2, ref3, ref4 +%endif + movsxdifnidn src_strideq, src_strided + movsxdifnidn ref_strideq, ref_strided + mov ref2q, [ref1q+gprsize*1] + mov ref3q, [ref1q+gprsize*2] + mov ref4q, [ref1q+gprsize*3] + mov ref1q, [ref1q+gprsize*0] + + PROCESS_%1x2x4 1, 0, 0, src_strideq, ref_strideq, 1 +%rep (%2-4)/2 + PROCESS_%1x2x4 0, 0, 0, src_strideq, ref_strideq, 1 +%endrep + PROCESS_%1x2x4 0, 0, 0, src_strideq, ref_strideq, 0 + +%if %1 > 4 + pslldq m5, 4 + pslldq m7, 4 + por m4, m5 + por m6, m7 + mova m5, m4 + mova m7, m6 + punpcklqdq m4, m6 + punpckhqdq m5, m7 + movifnidn r4, r4mp + paddd m4, m5 + movu [r4], m4 + RET +%else + movifnidn r4, r4mp + pshufd m6, m6, 0x08 + pshufd m7, m7, 0x08 + movq [r4+0], m6 + movq [r4+8], m7 + RET +%endif +%endmacro + +INIT_XMM sse2 +SADNXN4D 128, 128 +SADNXN4D 128, 64 +SADNXN4D 64, 128 +SADNXN4D 64, 64 +SADNXN4D 64, 32 +SADNXN4D 32, 64 +SADNXN4D 32, 32 +SADNXN4D 32, 16 +SADNXN4D 16, 32 +SADNXN4D 16, 16 +SADNXN4D 16, 8 +SADNXN4D 8, 16 +SADNXN4D 8, 8 +SADNXN4D 8, 4 +SADNXN4D 4, 8 +SADNXN4D 4, 4 +SADNXN4D 4, 16 +SADNXN4D 16, 4 +SADNXN4D 8, 32 +SADNXN4D 32, 8 +SADNXN4D 16, 64 +SADNXN4D 64, 16 diff --git a/third_party/aom/aom_dsp/x86/sad_avx2.c b/third_party/aom/aom_dsp/x86/sad_avx2.c new file mode 100644 index 0000000000..a50dba64a1 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/sad_avx2.c @@ -0,0 +1,189 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ +#include + +#include "config/aom_dsp_rtcd.h" + +#include "aom_ports/mem.h" + +#define FSAD64_H(h) \ + unsigned int aom_sad64x##h##_avx2(const uint8_t *src_ptr, int src_stride, \ + const uint8_t *ref_ptr, int ref_stride) { \ + int i, res; \ + __m256i sad1_reg, sad2_reg, ref1_reg, ref2_reg; \ + __m256i sum_sad = _mm256_setzero_si256(); \ + __m256i sum_sad_h; \ + __m128i sum_sad128; \ + for (i = 0; i < h; i++) { \ + ref1_reg = _mm256_loadu_si256((__m256i const *)ref_ptr); \ + ref2_reg = _mm256_loadu_si256((__m256i const *)(ref_ptr + 32)); \ + sad1_reg = _mm256_sad_epu8( \ + ref1_reg, _mm256_loadu_si256((__m256i const *)src_ptr)); \ + sad2_reg = _mm256_sad_epu8( \ + ref2_reg, _mm256_loadu_si256((__m256i const *)(src_ptr + 32))); \ + sum_sad = \ + _mm256_add_epi32(sum_sad, _mm256_add_epi32(sad1_reg, sad2_reg)); \ + ref_ptr += ref_stride; \ + src_ptr += src_stride; \ + } \ + sum_sad_h = _mm256_srli_si256(sum_sad, 8); \ + sum_sad = _mm256_add_epi32(sum_sad, sum_sad_h); \ + sum_sad128 = _mm256_extracti128_si256(sum_sad, 1); \ + sum_sad128 = _mm_add_epi32(_mm256_castsi256_si128(sum_sad), sum_sad128); \ + res = _mm_cvtsi128_si32(sum_sad128); \ + _mm256_zeroupper(); \ + return res; \ + } + +#define FSAD32_H(h) \ + unsigned int aom_sad32x##h##_avx2(const uint8_t *src_ptr, int src_stride, \ + const uint8_t *ref_ptr, int ref_stride) { \ + int i, res; \ + __m256i sad1_reg, sad2_reg, ref1_reg, ref2_reg; \ + __m256i sum_sad = _mm256_setzero_si256(); \ + __m256i sum_sad_h; \ + __m128i sum_sad128; \ + int ref2_stride = ref_stride << 1; \ + int src2_stride = src_stride << 1; \ + int max = h >> 1; \ + for (i = 0; i < max; i++) { \ + ref1_reg = _mm256_loadu_si256((__m256i const *)ref_ptr); \ + ref2_reg = _mm256_loadu_si256((__m256i const *)(ref_ptr + ref_stride)); \ + sad1_reg = _mm256_sad_epu8( \ + ref1_reg, _mm256_loadu_si256((__m256i const *)src_ptr)); \ + sad2_reg = _mm256_sad_epu8( \ + ref2_reg, \ + _mm256_loadu_si256((__m256i const *)(src_ptr + src_stride))); \ + sum_sad = \ + _mm256_add_epi32(sum_sad, _mm256_add_epi32(sad1_reg, sad2_reg)); \ + ref_ptr += ref2_stride; \ + src_ptr += src2_stride; \ + } \ + sum_sad_h = _mm256_srli_si256(sum_sad, 8); \ + sum_sad = _mm256_add_epi32(sum_sad, sum_sad_h); \ + sum_sad128 = _mm256_extracti128_si256(sum_sad, 1); \ + sum_sad128 = _mm_add_epi32(_mm256_castsi256_si128(sum_sad), sum_sad128); \ + res = _mm_cvtsi128_si32(sum_sad128); \ + _mm256_zeroupper(); \ + return res; \ + } + +#define FSAD64 \ + FSAD64_H(64); \ + FSAD64_H(32); + +#define FSAD32 \ + FSAD32_H(64); \ + FSAD32_H(32); \ + FSAD32_H(16); + +/* clang-format off */ +FSAD64 +FSAD32 +/* clang-format on */ + +#undef FSAD64 +#undef FSAD32 +#undef FSAD64_H +#undef FSAD32_H + +#define FSADAVG64_H(h) \ + unsigned int aom_sad64x##h##_avg_avx2( \ + const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, \ + int ref_stride, const uint8_t *second_pred) { \ + int i, res; \ + __m256i sad1_reg, sad2_reg, ref1_reg, ref2_reg; \ + __m256i sum_sad = _mm256_setzero_si256(); \ + __m256i sum_sad_h; \ + __m128i sum_sad128; \ + for (i = 0; i < h; i++) { \ + ref1_reg = _mm256_loadu_si256((__m256i const *)ref_ptr); \ + ref2_reg = _mm256_loadu_si256((__m256i const *)(ref_ptr + 32)); \ + ref1_reg = _mm256_avg_epu8( \ + ref1_reg, _mm256_loadu_si256((__m256i const *)second_pred)); \ + ref2_reg = _mm256_avg_epu8( \ + ref2_reg, _mm256_loadu_si256((__m256i const *)(second_pred + 32))); \ + sad1_reg = _mm256_sad_epu8( \ + ref1_reg, _mm256_loadu_si256((__m256i const *)src_ptr)); \ + sad2_reg = _mm256_sad_epu8( \ + ref2_reg, _mm256_loadu_si256((__m256i const *)(src_ptr + 32))); \ + sum_sad = \ + _mm256_add_epi32(sum_sad, _mm256_add_epi32(sad1_reg, sad2_reg)); \ + ref_ptr += ref_stride; \ + src_ptr += src_stride; \ + second_pred += 64; \ + } \ + sum_sad_h = _mm256_srli_si256(sum_sad, 8); \ + sum_sad = _mm256_add_epi32(sum_sad, sum_sad_h); \ + sum_sad128 = _mm256_extracti128_si256(sum_sad, 1); \ + sum_sad128 = _mm_add_epi32(_mm256_castsi256_si128(sum_sad), sum_sad128); \ + res = _mm_cvtsi128_si32(sum_sad128); \ + _mm256_zeroupper(); \ + return res; \ + } + +#define FSADAVG32_H(h) \ + unsigned int aom_sad32x##h##_avg_avx2( \ + const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, \ + int ref_stride, const uint8_t *second_pred) { \ + int i, res; \ + __m256i sad1_reg, sad2_reg, ref1_reg, ref2_reg; \ + __m256i sum_sad = _mm256_setzero_si256(); \ + __m256i sum_sad_h; \ + __m128i sum_sad128; \ + int ref2_stride = ref_stride << 1; \ + int src2_stride = src_stride << 1; \ + int max = h >> 1; \ + for (i = 0; i < max; i++) { \ + ref1_reg = _mm256_loadu_si256((__m256i const *)ref_ptr); \ + ref2_reg = _mm256_loadu_si256((__m256i const *)(ref_ptr + ref_stride)); \ + ref1_reg = _mm256_avg_epu8( \ + ref1_reg, _mm256_loadu_si256((__m256i const *)second_pred)); \ + ref2_reg = _mm256_avg_epu8( \ + ref2_reg, _mm256_loadu_si256((__m256i const *)(second_pred + 32))); \ + sad1_reg = _mm256_sad_epu8( \ + ref1_reg, _mm256_loadu_si256((__m256i const *)src_ptr)); \ + sad2_reg = _mm256_sad_epu8( \ + ref2_reg, \ + _mm256_loadu_si256((__m256i const *)(src_ptr + src_stride))); \ + sum_sad = \ + _mm256_add_epi32(sum_sad, _mm256_add_epi32(sad1_reg, sad2_reg)); \ + ref_ptr += ref2_stride; \ + src_ptr += src2_stride; \ + second_pred += 64; \ + } \ + sum_sad_h = _mm256_srli_si256(sum_sad, 8); \ + sum_sad = _mm256_add_epi32(sum_sad, sum_sad_h); \ + sum_sad128 = _mm256_extracti128_si256(sum_sad, 1); \ + sum_sad128 = _mm_add_epi32(_mm256_castsi256_si128(sum_sad), sum_sad128); \ + res = _mm_cvtsi128_si32(sum_sad128); \ + _mm256_zeroupper(); \ + return res; \ + } + +#define FSADAVG64 \ + FSADAVG64_H(64); \ + FSADAVG64_H(32); + +#define FSADAVG32 \ + FSADAVG32_H(64); \ + FSADAVG32_H(32); \ + FSADAVG32_H(16); + +/* clang-format off */ +FSADAVG64 +FSADAVG32 +/* clang-format on */ + +#undef FSADAVG64 +#undef FSADAVG32 +#undef FSADAVG64_H +#undef FSADAVG32_H diff --git a/third_party/aom/aom_dsp/x86/sad_highbd_avx2.c b/third_party/aom/aom_dsp/x86/sad_highbd_avx2.c new file mode 100644 index 0000000000..b506d46639 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/sad_highbd_avx2.c @@ -0,0 +1,1038 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include + +#include "config/aom_config.h" +#include "config/aom_dsp_rtcd.h" + +#include "aom/aom_integer.h" +#include "aom_dsp/x86/synonyms_avx2.h" +#include "aom_ports/mem.h" + +// SAD +static INLINE unsigned int get_sad_from_mm256_epi32(const __m256i *v) { + // input 8 32-bit summation + __m128i lo128, hi128; + __m256i u = _mm256_srli_si256(*v, 8); + u = _mm256_add_epi32(u, *v); + + // 4 32-bit summation + hi128 = _mm256_extracti128_si256(u, 1); + lo128 = _mm256_castsi256_si128(u); + lo128 = _mm_add_epi32(hi128, lo128); + + // 2 32-bit summation + hi128 = _mm_srli_si128(lo128, 4); + lo128 = _mm_add_epi32(lo128, hi128); + + return (unsigned int)_mm_cvtsi128_si32(lo128); +} + +unsigned int aom_highbd_sad16x8_avx2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride) { + const uint16_t *src_ptr = CONVERT_TO_SHORTPTR(src); + const uint16_t *ref_ptr = CONVERT_TO_SHORTPTR(ref); + + // first 4 rows + __m256i s0 = _mm256_loadu_si256((const __m256i *)src_ptr); + __m256i s1 = _mm256_loadu_si256((const __m256i *)(src_ptr + src_stride)); + __m256i s2 = _mm256_loadu_si256((const __m256i *)(src_ptr + 2 * src_stride)); + __m256i s3 = _mm256_loadu_si256((const __m256i *)(src_ptr + 3 * src_stride)); + + __m256i r0 = _mm256_loadu_si256((const __m256i *)ref_ptr); + __m256i r1 = _mm256_loadu_si256((const __m256i *)(ref_ptr + ref_stride)); + __m256i r2 = _mm256_loadu_si256((const __m256i *)(ref_ptr + 2 * ref_stride)); + __m256i r3 = _mm256_loadu_si256((const __m256i *)(ref_ptr + 3 * ref_stride)); + + __m256i u0 = _mm256_sub_epi16(s0, r0); + __m256i u1 = _mm256_sub_epi16(s1, r1); + __m256i u2 = _mm256_sub_epi16(s2, r2); + __m256i u3 = _mm256_sub_epi16(s3, r3); + __m256i zero = _mm256_setzero_si256(); + __m256i sum0, sum1; + + u0 = _mm256_abs_epi16(u0); + u1 = _mm256_abs_epi16(u1); + u2 = _mm256_abs_epi16(u2); + u3 = _mm256_abs_epi16(u3); + + sum0 = _mm256_add_epi16(u0, u1); + sum0 = _mm256_add_epi16(sum0, u2); + sum0 = _mm256_add_epi16(sum0, u3); + + // second 4 rows + src_ptr += src_stride << 2; + ref_ptr += ref_stride << 2; + s0 = _mm256_loadu_si256((const __m256i *)src_ptr); + s1 = _mm256_loadu_si256((const __m256i *)(src_ptr + src_stride)); + s2 = _mm256_loadu_si256((const __m256i *)(src_ptr + 2 * src_stride)); + s3 = _mm256_loadu_si256((const __m256i *)(src_ptr + 3 * src_stride)); + + r0 = _mm256_loadu_si256((const __m256i *)ref_ptr); + r1 = _mm256_loadu_si256((const __m256i *)(ref_ptr + ref_stride)); + r2 = _mm256_loadu_si256((const __m256i *)(ref_ptr + 2 * ref_stride)); + r3 = _mm256_loadu_si256((const __m256i *)(ref_ptr + 3 * ref_stride)); + + u0 = _mm256_sub_epi16(s0, r0); + u1 = _mm256_sub_epi16(s1, r1); + u2 = _mm256_sub_epi16(s2, r2); + u3 = _mm256_sub_epi16(s3, r3); + + u0 = _mm256_abs_epi16(u0); + u1 = _mm256_abs_epi16(u1); + u2 = _mm256_abs_epi16(u2); + u3 = _mm256_abs_epi16(u3); + + sum1 = _mm256_add_epi16(u0, u1); + sum1 = _mm256_add_epi16(sum1, u2); + sum1 = _mm256_add_epi16(sum1, u3); + + // find out the SAD + s0 = _mm256_unpacklo_epi16(sum0, zero); + s1 = _mm256_unpackhi_epi16(sum0, zero); + r0 = _mm256_unpacklo_epi16(sum1, zero); + r1 = _mm256_unpackhi_epi16(sum1, zero); + s0 = _mm256_add_epi32(s0, s1); + r0 = _mm256_add_epi32(r0, r1); + sum0 = _mm256_add_epi32(s0, r0); + // 8 32-bit summation + + return (unsigned int)get_sad_from_mm256_epi32(&sum0); +} + +unsigned int aom_highbd_sad16x16_avx2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride) { + const uint16_t *src_ptr = CONVERT_TO_SHORTPTR(src); + const uint16_t *ref_ptr = CONVERT_TO_SHORTPTR(ref); + __m256i s0, s1, s2, s3, r0, r1, r2, r3, u0, u1, u2, u3; + __m256i sum0; + __m256i sum = _mm256_setzero_si256(); + const __m256i zero = _mm256_setzero_si256(); + int row = 0; + + // Loop for every 4 rows + while (row < 16) { + s0 = _mm256_loadu_si256((const __m256i *)src_ptr); + s1 = _mm256_loadu_si256((const __m256i *)(src_ptr + src_stride)); + s2 = _mm256_loadu_si256((const __m256i *)(src_ptr + 2 * src_stride)); + s3 = _mm256_loadu_si256((const __m256i *)(src_ptr + 3 * src_stride)); + + r0 = _mm256_loadu_si256((const __m256i *)ref_ptr); + r1 = _mm256_loadu_si256((const __m256i *)(ref_ptr + ref_stride)); + r2 = _mm256_loadu_si256((const __m256i *)(ref_ptr + 2 * ref_stride)); + r3 = _mm256_loadu_si256((const __m256i *)(ref_ptr + 3 * ref_stride)); + + u0 = _mm256_sub_epi16(s0, r0); + u1 = _mm256_sub_epi16(s1, r1); + u2 = _mm256_sub_epi16(s2, r2); + u3 = _mm256_sub_epi16(s3, r3); + + u0 = _mm256_abs_epi16(u0); + u1 = _mm256_abs_epi16(u1); + u2 = _mm256_abs_epi16(u2); + u3 = _mm256_abs_epi16(u3); + + sum0 = _mm256_add_epi16(u0, u1); + sum0 = _mm256_add_epi16(sum0, u2); + sum0 = _mm256_add_epi16(sum0, u3); + + s0 = _mm256_unpacklo_epi16(sum0, zero); + s1 = _mm256_unpackhi_epi16(sum0, zero); + sum = _mm256_add_epi32(sum, s0); + sum = _mm256_add_epi32(sum, s1); + // 8 32-bit summation + + row += 4; + src_ptr += src_stride << 2; + ref_ptr += ref_stride << 2; + } + return get_sad_from_mm256_epi32(&sum); +} + +static void sad32x4(const uint16_t *src_ptr, int src_stride, + const uint16_t *ref_ptr, int ref_stride, + const uint16_t *sec_ptr, __m256i *sad_acc) { + __m256i s0, s1, s2, s3, r0, r1, r2, r3; + const __m256i zero = _mm256_setzero_si256(); + int row_sections = 0; + + while (row_sections < 2) { + s0 = _mm256_loadu_si256((const __m256i *)src_ptr); + s1 = _mm256_loadu_si256((const __m256i *)(src_ptr + 16)); + s2 = _mm256_loadu_si256((const __m256i *)(src_ptr + src_stride)); + s3 = _mm256_loadu_si256((const __m256i *)(src_ptr + src_stride + 16)); + + r0 = _mm256_loadu_si256((const __m256i *)ref_ptr); + r1 = _mm256_loadu_si256((const __m256i *)(ref_ptr + 16)); + r2 = _mm256_loadu_si256((const __m256i *)(ref_ptr + ref_stride)); + r3 = _mm256_loadu_si256((const __m256i *)(ref_ptr + ref_stride + 16)); + + if (sec_ptr) { + r0 = _mm256_avg_epu16(r0, _mm256_loadu_si256((const __m256i *)sec_ptr)); + r1 = _mm256_avg_epu16( + r1, _mm256_loadu_si256((const __m256i *)(sec_ptr + 16))); + r2 = _mm256_avg_epu16( + r2, _mm256_loadu_si256((const __m256i *)(sec_ptr + 32))); + r3 = _mm256_avg_epu16( + r3, _mm256_loadu_si256((const __m256i *)(sec_ptr + 48))); + } + s0 = _mm256_sub_epi16(s0, r0); + s1 = _mm256_sub_epi16(s1, r1); + s2 = _mm256_sub_epi16(s2, r2); + s3 = _mm256_sub_epi16(s3, r3); + + s0 = _mm256_abs_epi16(s0); + s1 = _mm256_abs_epi16(s1); + s2 = _mm256_abs_epi16(s2); + s3 = _mm256_abs_epi16(s3); + + s0 = _mm256_add_epi16(s0, s1); + s0 = _mm256_add_epi16(s0, s2); + s0 = _mm256_add_epi16(s0, s3); + + r0 = _mm256_unpacklo_epi16(s0, zero); + r1 = _mm256_unpackhi_epi16(s0, zero); + + r0 = _mm256_add_epi32(r0, r1); + *sad_acc = _mm256_add_epi32(*sad_acc, r0); + + row_sections += 1; + src_ptr += src_stride << 1; + ref_ptr += ref_stride << 1; + if (sec_ptr) sec_ptr += 32 << 1; + } +} + +unsigned int aom_highbd_sad32x16_avx2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride) { + __m256i sad = _mm256_setzero_si256(); + uint16_t *srcp = CONVERT_TO_SHORTPTR(src); + uint16_t *refp = CONVERT_TO_SHORTPTR(ref); + const int left_shift = 2; + int row_section = 0; + + while (row_section < 4) { + sad32x4(srcp, src_stride, refp, ref_stride, NULL, &sad); + srcp += src_stride << left_shift; + refp += ref_stride << left_shift; + row_section += 1; + } + return get_sad_from_mm256_epi32(&sad); +} + +unsigned int aom_highbd_sad16x32_avx2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride) { + uint32_t sum = aom_highbd_sad16x16_avx2(src, src_stride, ref, ref_stride); + src += src_stride << 4; + ref += ref_stride << 4; + sum += aom_highbd_sad16x16_avx2(src, src_stride, ref, ref_stride); + return sum; +} + +unsigned int aom_highbd_sad32x32_avx2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride) { + uint32_t sum = aom_highbd_sad32x16_avx2(src, src_stride, ref, ref_stride); + src += src_stride << 4; + ref += ref_stride << 4; + sum += aom_highbd_sad32x16_avx2(src, src_stride, ref, ref_stride); + return sum; +} + +unsigned int aom_highbd_sad32x64_avx2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride) { + uint32_t sum = aom_highbd_sad32x32_avx2(src, src_stride, ref, ref_stride); + src += src_stride << 5; + ref += ref_stride << 5; + sum += aom_highbd_sad32x32_avx2(src, src_stride, ref, ref_stride); + return sum; +} + +static void sad64x2(const uint16_t *src_ptr, int src_stride, + const uint16_t *ref_ptr, int ref_stride, + const uint16_t *sec_ptr, __m256i *sad_acc) { + __m256i s[8], r[8]; + const __m256i zero = _mm256_setzero_si256(); + + s[0] = _mm256_loadu_si256((const __m256i *)src_ptr); + s[1] = _mm256_loadu_si256((const __m256i *)(src_ptr + 16)); + s[2] = _mm256_loadu_si256((const __m256i *)(src_ptr + 32)); + s[3] = _mm256_loadu_si256((const __m256i *)(src_ptr + 48)); + s[4] = _mm256_loadu_si256((const __m256i *)(src_ptr + src_stride)); + s[5] = _mm256_loadu_si256((const __m256i *)(src_ptr + src_stride + 16)); + s[6] = _mm256_loadu_si256((const __m256i *)(src_ptr + src_stride + 32)); + s[7] = _mm256_loadu_si256((const __m256i *)(src_ptr + src_stride + 48)); + + r[0] = _mm256_loadu_si256((const __m256i *)ref_ptr); + r[1] = _mm256_loadu_si256((const __m256i *)(ref_ptr + 16)); + r[2] = _mm256_loadu_si256((const __m256i *)(ref_ptr + 32)); + r[3] = _mm256_loadu_si256((const __m256i *)(ref_ptr + 48)); + r[4] = _mm256_loadu_si256((const __m256i *)(ref_ptr + ref_stride)); + r[5] = _mm256_loadu_si256((const __m256i *)(ref_ptr + ref_stride + 16)); + r[6] = _mm256_loadu_si256((const __m256i *)(ref_ptr + ref_stride + 32)); + r[7] = _mm256_loadu_si256((const __m256i *)(ref_ptr + ref_stride + 48)); + + if (sec_ptr) { + r[0] = _mm256_avg_epu16(r[0], _mm256_loadu_si256((const __m256i *)sec_ptr)); + r[1] = _mm256_avg_epu16( + r[1], _mm256_loadu_si256((const __m256i *)(sec_ptr + 16))); + r[2] = _mm256_avg_epu16( + r[2], _mm256_loadu_si256((const __m256i *)(sec_ptr + 32))); + r[3] = _mm256_avg_epu16( + r[3], _mm256_loadu_si256((const __m256i *)(sec_ptr + 48))); + r[4] = _mm256_avg_epu16( + r[4], _mm256_loadu_si256((const __m256i *)(sec_ptr + 64))); + r[5] = _mm256_avg_epu16( + r[5], _mm256_loadu_si256((const __m256i *)(sec_ptr + 80))); + r[6] = _mm256_avg_epu16( + r[6], _mm256_loadu_si256((const __m256i *)(sec_ptr + 96))); + r[7] = _mm256_avg_epu16( + r[7], _mm256_loadu_si256((const __m256i *)(sec_ptr + 112))); + } + + s[0] = _mm256_sub_epi16(s[0], r[0]); + s[1] = _mm256_sub_epi16(s[1], r[1]); + s[2] = _mm256_sub_epi16(s[2], r[2]); + s[3] = _mm256_sub_epi16(s[3], r[3]); + s[4] = _mm256_sub_epi16(s[4], r[4]); + s[5] = _mm256_sub_epi16(s[5], r[5]); + s[6] = _mm256_sub_epi16(s[6], r[6]); + s[7] = _mm256_sub_epi16(s[7], r[7]); + + s[0] = _mm256_abs_epi16(s[0]); + s[1] = _mm256_abs_epi16(s[1]); + s[2] = _mm256_abs_epi16(s[2]); + s[3] = _mm256_abs_epi16(s[3]); + s[4] = _mm256_abs_epi16(s[4]); + s[5] = _mm256_abs_epi16(s[5]); + s[6] = _mm256_abs_epi16(s[6]); + s[7] = _mm256_abs_epi16(s[7]); + + s[0] = _mm256_add_epi16(s[0], s[1]); + s[0] = _mm256_add_epi16(s[0], s[2]); + s[0] = _mm256_add_epi16(s[0], s[3]); + + s[4] = _mm256_add_epi16(s[4], s[5]); + s[4] = _mm256_add_epi16(s[4], s[6]); + s[4] = _mm256_add_epi16(s[4], s[7]); + + r[0] = _mm256_unpacklo_epi16(s[0], zero); + r[1] = _mm256_unpackhi_epi16(s[0], zero); + r[2] = _mm256_unpacklo_epi16(s[4], zero); + r[3] = _mm256_unpackhi_epi16(s[4], zero); + + r[0] = _mm256_add_epi32(r[0], r[1]); + r[0] = _mm256_add_epi32(r[0], r[2]); + r[0] = _mm256_add_epi32(r[0], r[3]); + *sad_acc = _mm256_add_epi32(*sad_acc, r[0]); +} + +unsigned int aom_highbd_sad64x32_avx2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride) { + __m256i sad = _mm256_setzero_si256(); + uint16_t *srcp = CONVERT_TO_SHORTPTR(src); + uint16_t *refp = CONVERT_TO_SHORTPTR(ref); + const int left_shift = 1; + int row_section = 0; + + while (row_section < 16) { + sad64x2(srcp, src_stride, refp, ref_stride, NULL, &sad); + srcp += src_stride << left_shift; + refp += ref_stride << left_shift; + row_section += 1; + } + return get_sad_from_mm256_epi32(&sad); +} + +unsigned int aom_highbd_sad64x64_avx2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride) { + uint32_t sum = aom_highbd_sad64x32_avx2(src, src_stride, ref, ref_stride); + src += src_stride << 5; + ref += ref_stride << 5; + sum += aom_highbd_sad64x32_avx2(src, src_stride, ref, ref_stride); + return sum; +} + +static void sad128x1(const uint16_t *src_ptr, const uint16_t *ref_ptr, + const uint16_t *sec_ptr, __m256i *sad_acc) { + __m256i s[8], r[8]; + const __m256i zero = _mm256_setzero_si256(); + + s[0] = _mm256_loadu_si256((const __m256i *)src_ptr); + s[1] = _mm256_loadu_si256((const __m256i *)(src_ptr + 16)); + s[2] = _mm256_loadu_si256((const __m256i *)(src_ptr + 32)); + s[3] = _mm256_loadu_si256((const __m256i *)(src_ptr + 48)); + s[4] = _mm256_loadu_si256((const __m256i *)(src_ptr + 64)); + s[5] = _mm256_loadu_si256((const __m256i *)(src_ptr + 80)); + s[6] = _mm256_loadu_si256((const __m256i *)(src_ptr + 96)); + s[7] = _mm256_loadu_si256((const __m256i *)(src_ptr + 112)); + + r[0] = _mm256_loadu_si256((const __m256i *)ref_ptr); + r[1] = _mm256_loadu_si256((const __m256i *)(ref_ptr + 16)); + r[2] = _mm256_loadu_si256((const __m256i *)(ref_ptr + 32)); + r[3] = _mm256_loadu_si256((const __m256i *)(ref_ptr + 48)); + r[4] = _mm256_loadu_si256((const __m256i *)(ref_ptr + 64)); + r[5] = _mm256_loadu_si256((const __m256i *)(ref_ptr + 80)); + r[6] = _mm256_loadu_si256((const __m256i *)(ref_ptr + 96)); + r[7] = _mm256_loadu_si256((const __m256i *)(ref_ptr + 112)); + + if (sec_ptr) { + r[0] = _mm256_avg_epu16(r[0], _mm256_loadu_si256((const __m256i *)sec_ptr)); + r[1] = _mm256_avg_epu16( + r[1], _mm256_loadu_si256((const __m256i *)(sec_ptr + 16))); + r[2] = _mm256_avg_epu16( + r[2], _mm256_loadu_si256((const __m256i *)(sec_ptr + 32))); + r[3] = _mm256_avg_epu16( + r[3], _mm256_loadu_si256((const __m256i *)(sec_ptr + 48))); + r[4] = _mm256_avg_epu16( + r[4], _mm256_loadu_si256((const __m256i *)(sec_ptr + 64))); + r[5] = _mm256_avg_epu16( + r[5], _mm256_loadu_si256((const __m256i *)(sec_ptr + 80))); + r[6] = _mm256_avg_epu16( + r[6], _mm256_loadu_si256((const __m256i *)(sec_ptr + 96))); + r[7] = _mm256_avg_epu16( + r[7], _mm256_loadu_si256((const __m256i *)(sec_ptr + 112))); + } + + s[0] = _mm256_sub_epi16(s[0], r[0]); + s[1] = _mm256_sub_epi16(s[1], r[1]); + s[2] = _mm256_sub_epi16(s[2], r[2]); + s[3] = _mm256_sub_epi16(s[3], r[3]); + s[4] = _mm256_sub_epi16(s[4], r[4]); + s[5] = _mm256_sub_epi16(s[5], r[5]); + s[6] = _mm256_sub_epi16(s[6], r[6]); + s[7] = _mm256_sub_epi16(s[7], r[7]); + + s[0] = _mm256_abs_epi16(s[0]); + s[1] = _mm256_abs_epi16(s[1]); + s[2] = _mm256_abs_epi16(s[2]); + s[3] = _mm256_abs_epi16(s[3]); + s[4] = _mm256_abs_epi16(s[4]); + s[5] = _mm256_abs_epi16(s[5]); + s[6] = _mm256_abs_epi16(s[6]); + s[7] = _mm256_abs_epi16(s[7]); + + s[0] = _mm256_add_epi16(s[0], s[1]); + s[0] = _mm256_add_epi16(s[0], s[2]); + s[0] = _mm256_add_epi16(s[0], s[3]); + + s[4] = _mm256_add_epi16(s[4], s[5]); + s[4] = _mm256_add_epi16(s[4], s[6]); + s[4] = _mm256_add_epi16(s[4], s[7]); + + r[0] = _mm256_unpacklo_epi16(s[0], zero); + r[1] = _mm256_unpackhi_epi16(s[0], zero); + r[2] = _mm256_unpacklo_epi16(s[4], zero); + r[3] = _mm256_unpackhi_epi16(s[4], zero); + + r[0] = _mm256_add_epi32(r[0], r[1]); + r[0] = _mm256_add_epi32(r[0], r[2]); + r[0] = _mm256_add_epi32(r[0], r[3]); + *sad_acc = _mm256_add_epi32(*sad_acc, r[0]); +} + +unsigned int aom_highbd_sad128x64_avx2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride) { + __m256i sad = _mm256_setzero_si256(); + uint16_t *srcp = CONVERT_TO_SHORTPTR(src); + uint16_t *refp = CONVERT_TO_SHORTPTR(ref); + int row = 0; + while (row < 64) { + sad128x1(srcp, refp, NULL, &sad); + srcp += src_stride; + refp += ref_stride; + row += 1; + } + return get_sad_from_mm256_epi32(&sad); +} + +unsigned int aom_highbd_sad64x128_avx2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride) { + uint32_t sum = aom_highbd_sad64x64_avx2(src, src_stride, ref, ref_stride); + src += src_stride << 6; + ref += ref_stride << 6; + sum += aom_highbd_sad64x64_avx2(src, src_stride, ref, ref_stride); + return sum; +} + +unsigned int aom_highbd_sad128x128_avx2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride) { + uint32_t sum = aom_highbd_sad128x64_avx2(src, src_stride, ref, ref_stride); + src += src_stride << 6; + ref += ref_stride << 6; + sum += aom_highbd_sad128x64_avx2(src, src_stride, ref, ref_stride); + return sum; +} + +// If sec_ptr = 0, calculate regular SAD. Otherwise, calculate average SAD. +static INLINE void sad16x4(const uint16_t *src_ptr, int src_stride, + const uint16_t *ref_ptr, int ref_stride, + const uint16_t *sec_ptr, __m256i *sad_acc) { + __m256i s0, s1, s2, s3, r0, r1, r2, r3; + const __m256i zero = _mm256_setzero_si256(); + + s0 = _mm256_loadu_si256((const __m256i *)src_ptr); + s1 = _mm256_loadu_si256((const __m256i *)(src_ptr + src_stride)); + s2 = _mm256_loadu_si256((const __m256i *)(src_ptr + 2 * src_stride)); + s3 = _mm256_loadu_si256((const __m256i *)(src_ptr + 3 * src_stride)); + + r0 = _mm256_loadu_si256((const __m256i *)ref_ptr); + r1 = _mm256_loadu_si256((const __m256i *)(ref_ptr + ref_stride)); + r2 = _mm256_loadu_si256((const __m256i *)(ref_ptr + 2 * ref_stride)); + r3 = _mm256_loadu_si256((const __m256i *)(ref_ptr + 3 * ref_stride)); + + if (sec_ptr) { + r0 = _mm256_avg_epu16(r0, _mm256_loadu_si256((const __m256i *)sec_ptr)); + r1 = _mm256_avg_epu16(r1, + _mm256_loadu_si256((const __m256i *)(sec_ptr + 16))); + r2 = _mm256_avg_epu16(r2, + _mm256_loadu_si256((const __m256i *)(sec_ptr + 32))); + r3 = _mm256_avg_epu16(r3, + _mm256_loadu_si256((const __m256i *)(sec_ptr + 48))); + } + + s0 = _mm256_sub_epi16(s0, r0); + s1 = _mm256_sub_epi16(s1, r1); + s2 = _mm256_sub_epi16(s2, r2); + s3 = _mm256_sub_epi16(s3, r3); + + s0 = _mm256_abs_epi16(s0); + s1 = _mm256_abs_epi16(s1); + s2 = _mm256_abs_epi16(s2); + s3 = _mm256_abs_epi16(s3); + + s0 = _mm256_add_epi16(s0, s1); + s0 = _mm256_add_epi16(s0, s2); + s0 = _mm256_add_epi16(s0, s3); + + r0 = _mm256_unpacklo_epi16(s0, zero); + r1 = _mm256_unpackhi_epi16(s0, zero); + + r0 = _mm256_add_epi32(r0, r1); + *sad_acc = _mm256_add_epi32(*sad_acc, r0); +} + +unsigned int aom_highbd_sad16x8_avg_avx2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride, + const uint8_t *second_pred) { + __m256i sad = _mm256_setzero_si256(); + uint16_t *srcp = CONVERT_TO_SHORTPTR(src); + uint16_t *refp = CONVERT_TO_SHORTPTR(ref); + uint16_t *secp = CONVERT_TO_SHORTPTR(second_pred); + + sad16x4(srcp, src_stride, refp, ref_stride, secp, &sad); + + // Next 4 rows + srcp += src_stride << 2; + refp += ref_stride << 2; + secp += 64; + sad16x4(srcp, src_stride, refp, ref_stride, secp, &sad); + return get_sad_from_mm256_epi32(&sad); +} + +unsigned int aom_highbd_sad16x16_avg_avx2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride, + const uint8_t *second_pred) { + const int left_shift = 3; + uint32_t sum = aom_highbd_sad16x8_avg_avx2(src, src_stride, ref, ref_stride, + second_pred); + src += src_stride << left_shift; + ref += ref_stride << left_shift; + second_pred += 16 << left_shift; + sum += aom_highbd_sad16x8_avg_avx2(src, src_stride, ref, ref_stride, + second_pred); + return sum; +} + +unsigned int aom_highbd_sad16x32_avg_avx2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride, + const uint8_t *second_pred) { + const int left_shift = 4; + uint32_t sum = aom_highbd_sad16x16_avg_avx2(src, src_stride, ref, ref_stride, + second_pred); + src += src_stride << left_shift; + ref += ref_stride << left_shift; + second_pred += 16 << left_shift; + sum += aom_highbd_sad16x16_avg_avx2(src, src_stride, ref, ref_stride, + second_pred); + return sum; +} + +unsigned int aom_highbd_sad32x16_avg_avx2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride, + const uint8_t *second_pred) { + __m256i sad = _mm256_setzero_si256(); + uint16_t *srcp = CONVERT_TO_SHORTPTR(src); + uint16_t *refp = CONVERT_TO_SHORTPTR(ref); + uint16_t *secp = CONVERT_TO_SHORTPTR(second_pred); + const int left_shift = 2; + int row_section = 0; + + while (row_section < 4) { + sad32x4(srcp, src_stride, refp, ref_stride, secp, &sad); + srcp += src_stride << left_shift; + refp += ref_stride << left_shift; + secp += 32 << left_shift; + row_section += 1; + } + return get_sad_from_mm256_epi32(&sad); +} + +unsigned int aom_highbd_sad32x32_avg_avx2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride, + const uint8_t *second_pred) { + const int left_shift = 4; + uint32_t sum = aom_highbd_sad32x16_avg_avx2(src, src_stride, ref, ref_stride, + second_pred); + src += src_stride << left_shift; + ref += ref_stride << left_shift; + second_pred += 32 << left_shift; + sum += aom_highbd_sad32x16_avg_avx2(src, src_stride, ref, ref_stride, + second_pred); + return sum; +} + +unsigned int aom_highbd_sad32x64_avg_avx2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride, + const uint8_t *second_pred) { + const int left_shift = 5; + uint32_t sum = aom_highbd_sad32x32_avg_avx2(src, src_stride, ref, ref_stride, + second_pred); + src += src_stride << left_shift; + ref += ref_stride << left_shift; + second_pred += 32 << left_shift; + sum += aom_highbd_sad32x32_avg_avx2(src, src_stride, ref, ref_stride, + second_pred); + return sum; +} + +unsigned int aom_highbd_sad64x32_avg_avx2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride, + const uint8_t *second_pred) { + __m256i sad = _mm256_setzero_si256(); + uint16_t *srcp = CONVERT_TO_SHORTPTR(src); + uint16_t *refp = CONVERT_TO_SHORTPTR(ref); + uint16_t *secp = CONVERT_TO_SHORTPTR(second_pred); + const int left_shift = 1; + int row_section = 0; + + while (row_section < 16) { + sad64x2(srcp, src_stride, refp, ref_stride, secp, &sad); + srcp += src_stride << left_shift; + refp += ref_stride << left_shift; + secp += 64 << left_shift; + row_section += 1; + } + return get_sad_from_mm256_epi32(&sad); +} + +unsigned int aom_highbd_sad64x64_avg_avx2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride, + const uint8_t *second_pred) { + const int left_shift = 5; + uint32_t sum = aom_highbd_sad64x32_avg_avx2(src, src_stride, ref, ref_stride, + second_pred); + src += src_stride << left_shift; + ref += ref_stride << left_shift; + second_pred += 64 << left_shift; + sum += aom_highbd_sad64x32_avg_avx2(src, src_stride, ref, ref_stride, + second_pred); + return sum; +} + +unsigned int aom_highbd_sad64x128_avg_avx2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride, + const uint8_t *second_pred) { + const int left_shift = 6; + uint32_t sum = aom_highbd_sad64x64_avg_avx2(src, src_stride, ref, ref_stride, + second_pred); + src += src_stride << left_shift; + ref += ref_stride << left_shift; + second_pred += 64 << left_shift; + sum += aom_highbd_sad64x64_avg_avx2(src, src_stride, ref, ref_stride, + second_pred); + return sum; +} + +unsigned int aom_highbd_sad128x64_avg_avx2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride, + const uint8_t *second_pred) { + __m256i sad = _mm256_setzero_si256(); + uint16_t *srcp = CONVERT_TO_SHORTPTR(src); + uint16_t *refp = CONVERT_TO_SHORTPTR(ref); + uint16_t *secp = CONVERT_TO_SHORTPTR(second_pred); + int row = 0; + while (row < 64) { + sad128x1(srcp, refp, secp, &sad); + srcp += src_stride; + refp += ref_stride; + secp += 16 << 3; + row += 1; + } + return get_sad_from_mm256_epi32(&sad); +} + +unsigned int aom_highbd_sad128x128_avg_avx2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride, + const uint8_t *second_pred) { + unsigned int sum; + const int left_shift = 6; + + sum = aom_highbd_sad128x64_avg_avx2(src, src_stride, ref, ref_stride, + second_pred); + src += src_stride << left_shift; + ref += ref_stride << left_shift; + second_pred += 128 << left_shift; + sum += aom_highbd_sad128x64_avg_avx2(src, src_stride, ref, ref_stride, + second_pred); + return sum; +} + +// SAD 4D +// Combine 4 __m256i vectors to uint32_t result[4] +static INLINE void get_4d_sad_from_mm256_epi32(const __m256i *v, + uint32_t *res) { + __m256i u0, u1, u2, u3; + const __m256i mask = yy_set1_64_from_32i(UINT32_MAX); + __m128i sad; + + // 8 32-bit summation + u0 = _mm256_srli_si256(v[0], 4); + u1 = _mm256_srli_si256(v[1], 4); + u2 = _mm256_srli_si256(v[2], 4); + u3 = _mm256_srli_si256(v[3], 4); + + u0 = _mm256_add_epi32(u0, v[0]); + u1 = _mm256_add_epi32(u1, v[1]); + u2 = _mm256_add_epi32(u2, v[2]); + u3 = _mm256_add_epi32(u3, v[3]); + + u0 = _mm256_and_si256(u0, mask); + u1 = _mm256_and_si256(u1, mask); + u2 = _mm256_and_si256(u2, mask); + u3 = _mm256_and_si256(u3, mask); + // 4 32-bit summation, evenly positioned + + u1 = _mm256_slli_si256(u1, 4); + u3 = _mm256_slli_si256(u3, 4); + + u0 = _mm256_or_si256(u0, u1); + u2 = _mm256_or_si256(u2, u3); + // 8 32-bit summation, interleaved + + u1 = _mm256_unpacklo_epi64(u0, u2); + u3 = _mm256_unpackhi_epi64(u0, u2); + + u0 = _mm256_add_epi32(u1, u3); + sad = _mm_add_epi32(_mm256_extractf128_si256(u0, 1), + _mm256_castsi256_si128(u0)); + _mm_storeu_si128((__m128i *)res, sad); +} + +static void convert_pointers(const uint8_t *const ref8[], + const uint16_t *ref[]) { + ref[0] = CONVERT_TO_SHORTPTR(ref8[0]); + ref[1] = CONVERT_TO_SHORTPTR(ref8[1]); + ref[2] = CONVERT_TO_SHORTPTR(ref8[2]); + ref[3] = CONVERT_TO_SHORTPTR(ref8[3]); +} + +static void init_sad(__m256i *s) { + s[0] = _mm256_setzero_si256(); + s[1] = _mm256_setzero_si256(); + s[2] = _mm256_setzero_si256(); + s[3] = _mm256_setzero_si256(); +} + +void aom_highbd_sad16x8x4d_avx2(const uint8_t *src, int src_stride, + const uint8_t *const ref_array[], + int ref_stride, uint32_t *sad_array) { + __m256i sad_vec[4]; + const uint16_t *refp[4]; + const uint16_t *keep = CONVERT_TO_SHORTPTR(src); + const uint16_t *srcp; + const int shift_for_4_rows = 2; + int i; + + init_sad(sad_vec); + convert_pointers(ref_array, refp); + + for (i = 0; i < 4; ++i) { + srcp = keep; + sad16x4(srcp, src_stride, refp[i], ref_stride, 0, &sad_vec[i]); + srcp += src_stride << shift_for_4_rows; + refp[i] += ref_stride << shift_for_4_rows; + sad16x4(srcp, src_stride, refp[i], ref_stride, 0, &sad_vec[i]); + } + get_4d_sad_from_mm256_epi32(sad_vec, sad_array); +} + +void aom_highbd_sad16x16x4d_avx2(const uint8_t *src, int src_stride, + const uint8_t *const ref_array[], + int ref_stride, uint32_t *sad_array) { + uint32_t first8rows[4]; + uint32_t second8rows[4]; + const uint8_t *ref[4]; + const int shift_for_8_rows = 3; + + ref[0] = ref_array[0]; + ref[1] = ref_array[1]; + ref[2] = ref_array[2]; + ref[3] = ref_array[3]; + + aom_highbd_sad16x8x4d_avx2(src, src_stride, ref, ref_stride, first8rows); + src += src_stride << shift_for_8_rows; + ref[0] += ref_stride << shift_for_8_rows; + ref[1] += ref_stride << shift_for_8_rows; + ref[2] += ref_stride << shift_for_8_rows; + ref[3] += ref_stride << shift_for_8_rows; + aom_highbd_sad16x8x4d_avx2(src, src_stride, ref, ref_stride, second8rows); + sad_array[0] = first8rows[0] + second8rows[0]; + sad_array[1] = first8rows[1] + second8rows[1]; + sad_array[2] = first8rows[2] + second8rows[2]; + sad_array[3] = first8rows[3] + second8rows[3]; +} + +void aom_highbd_sad16x32x4d_avx2(const uint8_t *src, int src_stride, + const uint8_t *const ref_array[], + int ref_stride, uint32_t *sad_array) { + uint32_t first_half[4]; + uint32_t second_half[4]; + const uint8_t *ref[4]; + const int shift_for_rows = 4; + + ref[0] = ref_array[0]; + ref[1] = ref_array[1]; + ref[2] = ref_array[2]; + ref[3] = ref_array[3]; + + aom_highbd_sad16x16x4d_avx2(src, src_stride, ref, ref_stride, first_half); + src += src_stride << shift_for_rows; + ref[0] += ref_stride << shift_for_rows; + ref[1] += ref_stride << shift_for_rows; + ref[2] += ref_stride << shift_for_rows; + ref[3] += ref_stride << shift_for_rows; + aom_highbd_sad16x16x4d_avx2(src, src_stride, ref, ref_stride, second_half); + sad_array[0] = first_half[0] + second_half[0]; + sad_array[1] = first_half[1] + second_half[1]; + sad_array[2] = first_half[2] + second_half[2]; + sad_array[3] = first_half[3] + second_half[3]; +} + +void aom_highbd_sad32x16x4d_avx2(const uint8_t *src, int src_stride, + const uint8_t *const ref_array[], + int ref_stride, uint32_t *sad_array) { + __m256i sad_vec[4]; + const uint16_t *refp[4]; + const uint16_t *keep = CONVERT_TO_SHORTPTR(src); + const uint16_t *srcp; + const int shift_for_4_rows = 2; + int i; + int rows_section; + + init_sad(sad_vec); + convert_pointers(ref_array, refp); + + for (i = 0; i < 4; ++i) { + srcp = keep; + rows_section = 0; + while (rows_section < 4) { + sad32x4(srcp, src_stride, refp[i], ref_stride, 0, &sad_vec[i]); + srcp += src_stride << shift_for_4_rows; + refp[i] += ref_stride << shift_for_4_rows; + rows_section++; + } + } + get_4d_sad_from_mm256_epi32(sad_vec, sad_array); +} + +void aom_highbd_sad32x32x4d_avx2(const uint8_t *src, int src_stride, + const uint8_t *const ref_array[], + int ref_stride, uint32_t *sad_array) { + uint32_t first_half[4]; + uint32_t second_half[4]; + const uint8_t *ref[4]; + const int shift_for_rows = 4; + + ref[0] = ref_array[0]; + ref[1] = ref_array[1]; + ref[2] = ref_array[2]; + ref[3] = ref_array[3]; + + aom_highbd_sad32x16x4d_avx2(src, src_stride, ref, ref_stride, first_half); + src += src_stride << shift_for_rows; + ref[0] += ref_stride << shift_for_rows; + ref[1] += ref_stride << shift_for_rows; + ref[2] += ref_stride << shift_for_rows; + ref[3] += ref_stride << shift_for_rows; + aom_highbd_sad32x16x4d_avx2(src, src_stride, ref, ref_stride, second_half); + sad_array[0] = first_half[0] + second_half[0]; + sad_array[1] = first_half[1] + second_half[1]; + sad_array[2] = first_half[2] + second_half[2]; + sad_array[3] = first_half[3] + second_half[3]; +} + +void aom_highbd_sad32x64x4d_avx2(const uint8_t *src, int src_stride, + const uint8_t *const ref_array[], + int ref_stride, uint32_t *sad_array) { + uint32_t first_half[4]; + uint32_t second_half[4]; + const uint8_t *ref[4]; + const int shift_for_rows = 5; + + ref[0] = ref_array[0]; + ref[1] = ref_array[1]; + ref[2] = ref_array[2]; + ref[3] = ref_array[3]; + + aom_highbd_sad32x32x4d_avx2(src, src_stride, ref, ref_stride, first_half); + src += src_stride << shift_for_rows; + ref[0] += ref_stride << shift_for_rows; + ref[1] += ref_stride << shift_for_rows; + ref[2] += ref_stride << shift_for_rows; + ref[3] += ref_stride << shift_for_rows; + aom_highbd_sad32x32x4d_avx2(src, src_stride, ref, ref_stride, second_half); + sad_array[0] = first_half[0] + second_half[0]; + sad_array[1] = first_half[1] + second_half[1]; + sad_array[2] = first_half[2] + second_half[2]; + sad_array[3] = first_half[3] + second_half[3]; +} + +void aom_highbd_sad64x32x4d_avx2(const uint8_t *src, int src_stride, + const uint8_t *const ref_array[], + int ref_stride, uint32_t *sad_array) { + __m256i sad_vec[4]; + const uint16_t *refp[4]; + const uint16_t *keep = CONVERT_TO_SHORTPTR(src); + const uint16_t *srcp; + const int shift_for_rows = 1; + int i; + int rows_section; + + init_sad(sad_vec); + convert_pointers(ref_array, refp); + + for (i = 0; i < 4; ++i) { + srcp = keep; + rows_section = 0; + while (rows_section < 16) { + sad64x2(srcp, src_stride, refp[i], ref_stride, NULL, &sad_vec[i]); + srcp += src_stride << shift_for_rows; + refp[i] += ref_stride << shift_for_rows; + rows_section++; + } + } + get_4d_sad_from_mm256_epi32(sad_vec, sad_array); +} + +void aom_highbd_sad64x64x4d_avx2(const uint8_t *src, int src_stride, + const uint8_t *const ref_array[], + int ref_stride, uint32_t *sad_array) { + uint32_t first_half[4]; + uint32_t second_half[4]; + const uint8_t *ref[4]; + const int shift_for_rows = 5; + + ref[0] = ref_array[0]; + ref[1] = ref_array[1]; + ref[2] = ref_array[2]; + ref[3] = ref_array[3]; + + aom_highbd_sad64x32x4d_avx2(src, src_stride, ref, ref_stride, first_half); + src += src_stride << shift_for_rows; + ref[0] += ref_stride << shift_for_rows; + ref[1] += ref_stride << shift_for_rows; + ref[2] += ref_stride << shift_for_rows; + ref[3] += ref_stride << shift_for_rows; + aom_highbd_sad64x32x4d_avx2(src, src_stride, ref, ref_stride, second_half); + sad_array[0] = first_half[0] + second_half[0]; + sad_array[1] = first_half[1] + second_half[1]; + sad_array[2] = first_half[2] + second_half[2]; + sad_array[3] = first_half[3] + second_half[3]; +} + +void aom_highbd_sad64x128x4d_avx2(const uint8_t *src, int src_stride, + const uint8_t *const ref_array[], + int ref_stride, uint32_t *sad_array) { + uint32_t first_half[4]; + uint32_t second_half[4]; + const uint8_t *ref[4]; + const int shift_for_rows = 6; + + ref[0] = ref_array[0]; + ref[1] = ref_array[1]; + ref[2] = ref_array[2]; + ref[3] = ref_array[3]; + + aom_highbd_sad64x64x4d_avx2(src, src_stride, ref, ref_stride, first_half); + src += src_stride << shift_for_rows; + ref[0] += ref_stride << shift_for_rows; + ref[1] += ref_stride << shift_for_rows; + ref[2] += ref_stride << shift_for_rows; + ref[3] += ref_stride << shift_for_rows; + aom_highbd_sad64x64x4d_avx2(src, src_stride, ref, ref_stride, second_half); + sad_array[0] = first_half[0] + second_half[0]; + sad_array[1] = first_half[1] + second_half[1]; + sad_array[2] = first_half[2] + second_half[2]; + sad_array[3] = first_half[3] + second_half[3]; +} + +void aom_highbd_sad128x64x4d_avx2(const uint8_t *src, int src_stride, + const uint8_t *const ref_array[], + int ref_stride, uint32_t *sad_array) { + __m256i sad_vec[4]; + const uint16_t *refp[4]; + const uint16_t *keep = CONVERT_TO_SHORTPTR(src); + const uint16_t *srcp; + int i; + int rows_section; + + init_sad(sad_vec); + convert_pointers(ref_array, refp); + + for (i = 0; i < 4; ++i) { + srcp = keep; + rows_section = 0; + while (rows_section < 64) { + sad128x1(srcp, refp[i], NULL, &sad_vec[i]); + srcp += src_stride; + refp[i] += ref_stride; + rows_section++; + } + } + get_4d_sad_from_mm256_epi32(sad_vec, sad_array); +} + +void aom_highbd_sad128x128x4d_avx2(const uint8_t *src, int src_stride, + const uint8_t *const ref_array[], + int ref_stride, uint32_t *sad_array) { + uint32_t first_half[4]; + uint32_t second_half[4]; + const uint8_t *ref[4]; + const int shift_for_rows = 6; + + ref[0] = ref_array[0]; + ref[1] = ref_array[1]; + ref[2] = ref_array[2]; + ref[3] = ref_array[3]; + + aom_highbd_sad128x64x4d_avx2(src, src_stride, ref, ref_stride, first_half); + src += src_stride << shift_for_rows; + ref[0] += ref_stride << shift_for_rows; + ref[1] += ref_stride << shift_for_rows; + ref[2] += ref_stride << shift_for_rows; + ref[3] += ref_stride << shift_for_rows; + aom_highbd_sad128x64x4d_avx2(src, src_stride, ref, ref_stride, second_half); + sad_array[0] = first_half[0] + second_half[0]; + sad_array[1] = first_half[1] + second_half[1]; + sad_array[2] = first_half[2] + second_half[2]; + sad_array[3] = first_half[3] + second_half[3]; +} diff --git a/third_party/aom/aom_dsp/x86/sad_impl_avx2.c b/third_party/aom/aom_dsp/x86/sad_impl_avx2.c new file mode 100644 index 0000000000..c6fd62c9e2 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/sad_impl_avx2.c @@ -0,0 +1,234 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include + +#include "config/aom_dsp_rtcd.h" + +static unsigned int sad32x32(const uint8_t *src_ptr, int src_stride, + const uint8_t *ref_ptr, int ref_stride) { + __m256i s1, s2, r1, r2; + __m256i sum = _mm256_setzero_si256(); + __m128i sum_i128; + int i; + + for (i = 0; i < 16; ++i) { + r1 = _mm256_loadu_si256((__m256i const *)ref_ptr); + r2 = _mm256_loadu_si256((__m256i const *)(ref_ptr + ref_stride)); + s1 = _mm256_sad_epu8(r1, _mm256_loadu_si256((__m256i const *)src_ptr)); + s2 = _mm256_sad_epu8( + r2, _mm256_loadu_si256((__m256i const *)(src_ptr + src_stride))); + sum = _mm256_add_epi32(sum, _mm256_add_epi32(s1, s2)); + ref_ptr += ref_stride << 1; + src_ptr += src_stride << 1; + } + + sum = _mm256_add_epi32(sum, _mm256_srli_si256(sum, 8)); + sum_i128 = _mm_add_epi32(_mm256_extracti128_si256(sum, 1), + _mm256_castsi256_si128(sum)); + return _mm_cvtsi128_si32(sum_i128); +} + +static unsigned int sad64x32(const uint8_t *src_ptr, int src_stride, + const uint8_t *ref_ptr, int ref_stride) { + unsigned int half_width = 32; + uint32_t sum = sad32x32(src_ptr, src_stride, ref_ptr, ref_stride); + src_ptr += half_width; + ref_ptr += half_width; + sum += sad32x32(src_ptr, src_stride, ref_ptr, ref_stride); + return sum; +} + +static unsigned int sad64x64(const uint8_t *src_ptr, int src_stride, + const uint8_t *ref_ptr, int ref_stride) { + uint32_t sum = sad64x32(src_ptr, src_stride, ref_ptr, ref_stride); + src_ptr += src_stride << 5; + ref_ptr += ref_stride << 5; + sum += sad64x32(src_ptr, src_stride, ref_ptr, ref_stride); + return sum; +} + +unsigned int aom_sad128x64_avx2(const uint8_t *src_ptr, int src_stride, + const uint8_t *ref_ptr, int ref_stride) { + unsigned int half_width = 64; + uint32_t sum = sad64x64(src_ptr, src_stride, ref_ptr, ref_stride); + src_ptr += half_width; + ref_ptr += half_width; + sum += sad64x64(src_ptr, src_stride, ref_ptr, ref_stride); + return sum; +} + +unsigned int aom_sad64x128_avx2(const uint8_t *src_ptr, int src_stride, + const uint8_t *ref_ptr, int ref_stride) { + uint32_t sum = sad64x64(src_ptr, src_stride, ref_ptr, ref_stride); + src_ptr += src_stride << 6; + ref_ptr += ref_stride << 6; + sum += sad64x64(src_ptr, src_stride, ref_ptr, ref_stride); + return sum; +} + +unsigned int aom_sad128x128_avx2(const uint8_t *src_ptr, int src_stride, + const uint8_t *ref_ptr, int ref_stride) { + uint32_t sum = aom_sad128x64_avx2(src_ptr, src_stride, ref_ptr, ref_stride); + src_ptr += src_stride << 6; + ref_ptr += ref_stride << 6; + sum += aom_sad128x64_avx2(src_ptr, src_stride, ref_ptr, ref_stride); + return sum; +} + +static void sad64x64x4d(const uint8_t *src, int src_stride, + const uint8_t *const ref[4], int ref_stride, + __m128i *res) { + uint32_t sum[4]; + aom_sad64x64x4d_avx2(src, src_stride, ref, ref_stride, sum); + *res = _mm_loadu_si128((const __m128i *)sum); +} + +void aom_sad64x128x4d_avx2(const uint8_t *src, int src_stride, + const uint8_t *const ref[4], int ref_stride, + uint32_t res[4]) { + __m128i sum0, sum1; + const uint8_t *rf[4]; + + rf[0] = ref[0]; + rf[1] = ref[1]; + rf[2] = ref[2]; + rf[3] = ref[3]; + sad64x64x4d(src, src_stride, rf, ref_stride, &sum0); + src += src_stride << 6; + rf[0] += ref_stride << 6; + rf[1] += ref_stride << 6; + rf[2] += ref_stride << 6; + rf[3] += ref_stride << 6; + sad64x64x4d(src, src_stride, rf, ref_stride, &sum1); + sum0 = _mm_add_epi32(sum0, sum1); + _mm_storeu_si128((__m128i *)res, sum0); +} + +void aom_sad128x64x4d_avx2(const uint8_t *src, int src_stride, + const uint8_t *const ref[4], int ref_stride, + uint32_t res[4]) { + __m128i sum0, sum1; + unsigned int half_width = 64; + const uint8_t *rf[4]; + + rf[0] = ref[0]; + rf[1] = ref[1]; + rf[2] = ref[2]; + rf[3] = ref[3]; + sad64x64x4d(src, src_stride, rf, ref_stride, &sum0); + src += half_width; + rf[0] += half_width; + rf[1] += half_width; + rf[2] += half_width; + rf[3] += half_width; + sad64x64x4d(src, src_stride, rf, ref_stride, &sum1); + sum0 = _mm_add_epi32(sum0, sum1); + _mm_storeu_si128((__m128i *)res, sum0); +} + +void aom_sad128x128x4d_avx2(const uint8_t *src, int src_stride, + const uint8_t *const ref[4], int ref_stride, + uint32_t res[4]) { + const uint8_t *rf[4]; + uint32_t sum0[4]; + uint32_t sum1[4]; + + rf[0] = ref[0]; + rf[1] = ref[1]; + rf[2] = ref[2]; + rf[3] = ref[3]; + aom_sad128x64x4d_avx2(src, src_stride, rf, ref_stride, sum0); + src += src_stride << 6; + rf[0] += ref_stride << 6; + rf[1] += ref_stride << 6; + rf[2] += ref_stride << 6; + rf[3] += ref_stride << 6; + aom_sad128x64x4d_avx2(src, src_stride, rf, ref_stride, sum1); + res[0] = sum0[0] + sum1[0]; + res[1] = sum0[1] + sum1[1]; + res[2] = sum0[2] + sum1[2]; + res[3] = sum0[3] + sum1[3]; +} + +static unsigned int sad_w64_avg_avx2(const uint8_t *src_ptr, int src_stride, + const uint8_t *ref_ptr, int ref_stride, + const int h, const uint8_t *second_pred, + const int second_pred_stride) { + int i, res; + __m256i sad1_reg, sad2_reg, ref1_reg, ref2_reg; + __m256i sum_sad = _mm256_setzero_si256(); + __m256i sum_sad_h; + __m128i sum_sad128; + for (i = 0; i < h; i++) { + ref1_reg = _mm256_loadu_si256((__m256i const *)ref_ptr); + ref2_reg = _mm256_loadu_si256((__m256i const *)(ref_ptr + 32)); + ref1_reg = _mm256_avg_epu8( + ref1_reg, _mm256_loadu_si256((__m256i const *)second_pred)); + ref2_reg = _mm256_avg_epu8( + ref2_reg, _mm256_loadu_si256((__m256i const *)(second_pred + 32))); + sad1_reg = + _mm256_sad_epu8(ref1_reg, _mm256_loadu_si256((__m256i const *)src_ptr)); + sad2_reg = _mm256_sad_epu8( + ref2_reg, _mm256_loadu_si256((__m256i const *)(src_ptr + 32))); + sum_sad = _mm256_add_epi32(sum_sad, _mm256_add_epi32(sad1_reg, sad2_reg)); + ref_ptr += ref_stride; + src_ptr += src_stride; + second_pred += second_pred_stride; + } + sum_sad_h = _mm256_srli_si256(sum_sad, 8); + sum_sad = _mm256_add_epi32(sum_sad, sum_sad_h); + sum_sad128 = _mm256_extracti128_si256(sum_sad, 1); + sum_sad128 = _mm_add_epi32(_mm256_castsi256_si128(sum_sad), sum_sad128); + res = _mm_cvtsi128_si32(sum_sad128); + + return res; +} + +unsigned int aom_sad64x128_avg_avx2(const uint8_t *src_ptr, int src_stride, + const uint8_t *ref_ptr, int ref_stride, + const uint8_t *second_pred) { + uint32_t sum = sad_w64_avg_avx2(src_ptr, src_stride, ref_ptr, ref_stride, 64, + second_pred, 64); + src_ptr += src_stride << 6; + ref_ptr += ref_stride << 6; + second_pred += 64 << 6; + sum += sad_w64_avg_avx2(src_ptr, src_stride, ref_ptr, ref_stride, 64, + second_pred, 64); + return sum; +} + +unsigned int aom_sad128x64_avg_avx2(const uint8_t *src_ptr, int src_stride, + const uint8_t *ref_ptr, int ref_stride, + const uint8_t *second_pred) { + unsigned int half_width = 64; + uint32_t sum = sad_w64_avg_avx2(src_ptr, src_stride, ref_ptr, ref_stride, 64, + second_pred, 128); + src_ptr += half_width; + ref_ptr += half_width; + second_pred += half_width; + sum += sad_w64_avg_avx2(src_ptr, src_stride, ref_ptr, ref_stride, 64, + second_pred, 128); + return sum; +} + +unsigned int aom_sad128x128_avg_avx2(const uint8_t *src_ptr, int src_stride, + const uint8_t *ref_ptr, int ref_stride, + const uint8_t *second_pred) { + uint32_t sum = aom_sad128x64_avg_avx2(src_ptr, src_stride, ref_ptr, + ref_stride, second_pred); + src_ptr += src_stride << 6; + ref_ptr += ref_stride << 6; + second_pred += 128 << 6; + sum += aom_sad128x64_avg_avx2(src_ptr, src_stride, ref_ptr, ref_stride, + second_pred); + return sum; +} diff --git a/third_party/aom/aom_dsp/x86/sad_sse2.asm b/third_party/aom/aom_dsp/x86/sad_sse2.asm new file mode 100644 index 0000000000..3251b76559 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/sad_sse2.asm @@ -0,0 +1,353 @@ +; +; Copyright (c) 2016, Alliance for Open Media. All rights reserved +; +; This source code is subject to the terms of the BSD 2 Clause License and +; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License +; was not distributed with this source code in the LICENSE file, you can +; obtain it at www.aomedia.org/license/software. If the Alliance for Open +; Media Patent License 1.0 was not distributed with this source code in the +; PATENTS file, you can obtain it at www.aomedia.org/license/patent. +; + +; + +%include "third_party/x86inc/x86inc.asm" + +SECTION .text + +%macro SAD_FN 4 +%if %4 == 0 +%if %3 == 5 +cglobal sad%1x%2, 4, %3, 5, src, src_stride, ref, ref_stride, n_rows +%else ; %3 == 7 +cglobal sad%1x%2, 4, %3, 6, src, src_stride, ref, ref_stride, \ + src_stride3, ref_stride3, n_rows +%endif ; %3 == 5/7 +%else ; avg +%if %3 == 5 +cglobal sad%1x%2_avg, 5, 1 + %3, 5, src, src_stride, ref, ref_stride, \ + second_pred, n_rows +%else ; %3 == 7 +cglobal sad%1x%2_avg, 5, ARCH_X86_64 + %3, 6, src, src_stride, \ + ref, ref_stride, \ + second_pred, \ + src_stride3, ref_stride3 +%if ARCH_X86_64 +%define n_rowsd r7d +%else ; x86-32 +%define n_rowsd dword r0m +%endif ; x86-32/64 +%endif ; %3 == 5/7 +%endif ; avg/sad + movsxdifnidn src_strideq, src_strided + movsxdifnidn ref_strideq, ref_strided +%if %3 == 7 + lea src_stride3q, [src_strideq*3] + lea ref_stride3q, [ref_strideq*3] +%endif ; %3 == 7 +%endmacro + +; unsigned int aom_sad128x128_sse2(uint8_t *src, int src_stride, +; uint8_t *ref, int ref_stride); +%macro SAD128XN 1-2 0 + SAD_FN 128, %1, 5, %2 + mov n_rowsd, %1 + pxor m0, m0 + +.loop: + movu m1, [refq] + movu m2, [refq+16] + movu m3, [refq+32] + movu m4, [refq+48] +%if %2 == 1 + pavgb m1, [second_predq+mmsize*0] + pavgb m2, [second_predq+mmsize*1] + pavgb m3, [second_predq+mmsize*2] + pavgb m4, [second_predq+mmsize*3] +%endif + psadbw m1, [srcq] + psadbw m2, [srcq+16] + psadbw m3, [srcq+32] + psadbw m4, [srcq+48] + + paddd m1, m2 + paddd m3, m4 + paddd m0, m1 + paddd m0, m3 + + movu m1, [refq+64] + movu m2, [refq+80] + movu m3, [refq+96] + movu m4, [refq+112] +%if %2 == 1 + pavgb m1, [second_predq+mmsize*4] + pavgb m2, [second_predq+mmsize*5] + pavgb m3, [second_predq+mmsize*6] + pavgb m4, [second_predq+mmsize*7] + lea second_predq, [second_predq+mmsize*8] +%endif + psadbw m1, [srcq+64] + psadbw m2, [srcq+80] + psadbw m3, [srcq+96] + psadbw m4, [srcq+112] + + add refq, ref_strideq + add srcq, src_strideq + + paddd m1, m2 + paddd m3, m4 + paddd m0, m1 + paddd m0, m3 + + sub n_rowsd, 1 + jg .loop + + movhlps m1, m0 + paddd m0, m1 + movd eax, m0 + RET +%endmacro + +INIT_XMM sse2 +SAD128XN 128 ; sad128x128_sse2 +SAD128XN 128, 1 ; sad128x128_avg_sse2 +SAD128XN 64 ; sad128x64_sse2 +SAD128XN 64, 1 ; sad128x64_avg_sse2 + + +; unsigned int aom_sad64x64_sse2(uint8_t *src, int src_stride, +; uint8_t *ref, int ref_stride); +%macro SAD64XN 1-2 0 + SAD_FN 64, %1, 5, %2 + mov n_rowsd, %1 + pxor m0, m0 +.loop: + movu m1, [refq] + movu m2, [refq+16] + movu m3, [refq+32] + movu m4, [refq+48] +%if %2 == 1 + pavgb m1, [second_predq+mmsize*0] + pavgb m2, [second_predq+mmsize*1] + pavgb m3, [second_predq+mmsize*2] + pavgb m4, [second_predq+mmsize*3] + lea second_predq, [second_predq+mmsize*4] +%endif + psadbw m1, [srcq] + psadbw m2, [srcq+16] + psadbw m3, [srcq+32] + psadbw m4, [srcq+48] + paddd m1, m2 + paddd m3, m4 + add refq, ref_strideq + paddd m0, m1 + add srcq, src_strideq + paddd m0, m3 + dec n_rowsd + jg .loop + + movhlps m1, m0 + paddd m0, m1 + movd eax, m0 + RET +%endmacro + +INIT_XMM sse2 +SAD64XN 128 ; sad64x128_sse2 +SAD64XN 128, 1 ; sad64x128_avg_sse2 +SAD64XN 64 ; sad64x64_sse2 +SAD64XN 32 ; sad64x32_sse2 +SAD64XN 64, 1 ; sad64x64_avg_sse2 +SAD64XN 32, 1 ; sad64x32_avg_sse2 +SAD64XN 16 ; sad64x16_sse2 +SAD64XN 16, 1 ; sad64x16_avg_sse2 + +; unsigned int aom_sad32x32_sse2(uint8_t *src, int src_stride, +; uint8_t *ref, int ref_stride); +%macro SAD32XN 1-2 0 + SAD_FN 32, %1, 5, %2 + mov n_rowsd, %1/2 + pxor m0, m0 +.loop: + movu m1, [refq] + movu m2, [refq+16] + movu m3, [refq+ref_strideq] + movu m4, [refq+ref_strideq+16] +%if %2 == 1 + pavgb m1, [second_predq+mmsize*0] + pavgb m2, [second_predq+mmsize*1] + pavgb m3, [second_predq+mmsize*2] + pavgb m4, [second_predq+mmsize*3] + lea second_predq, [second_predq+mmsize*4] +%endif + psadbw m1, [srcq] + psadbw m2, [srcq+16] + psadbw m3, [srcq+src_strideq] + psadbw m4, [srcq+src_strideq+16] + paddd m1, m2 + paddd m3, m4 + lea refq, [refq+ref_strideq*2] + paddd m0, m1 + lea srcq, [srcq+src_strideq*2] + paddd m0, m3 + dec n_rowsd + jg .loop + + movhlps m1, m0 + paddd m0, m1 + movd eax, m0 + RET +%endmacro + +INIT_XMM sse2 +SAD32XN 64 ; sad32x64_sse2 +SAD32XN 32 ; sad32x32_sse2 +SAD32XN 16 ; sad32x16_sse2 +SAD32XN 64, 1 ; sad32x64_avg_sse2 +SAD32XN 32, 1 ; sad32x32_avg_sse2 +SAD32XN 16, 1 ; sad32x16_avg_sse2 +SAD32XN 8 ; sad_32x8_sse2 +SAD32XN 8, 1 ; sad_32x8_avg_sse2 + +; unsigned int aom_sad16x{8,16}_sse2(uint8_t *src, int src_stride, +; uint8_t *ref, int ref_stride); +%macro SAD16XN 1-2 0 + SAD_FN 16, %1, 7, %2 + mov n_rowsd, %1/4 + pxor m0, m0 + +.loop: + movu m1, [refq] + movu m2, [refq+ref_strideq] + movu m3, [refq+ref_strideq*2] + movu m4, [refq+ref_stride3q] +%if %2 == 1 + pavgb m1, [second_predq+mmsize*0] + pavgb m2, [second_predq+mmsize*1] + pavgb m3, [second_predq+mmsize*2] + pavgb m4, [second_predq+mmsize*3] + lea second_predq, [second_predq+mmsize*4] +%endif + psadbw m1, [srcq] + psadbw m2, [srcq+src_strideq] + psadbw m3, [srcq+src_strideq*2] + psadbw m4, [srcq+src_stride3q] + paddd m1, m2 + paddd m3, m4 + lea refq, [refq+ref_strideq*4] + paddd m0, m1 + lea srcq, [srcq+src_strideq*4] + paddd m0, m3 + dec n_rowsd + jg .loop + + movhlps m1, m0 + paddd m0, m1 + movd eax, m0 + RET +%endmacro + +INIT_XMM sse2 +SAD16XN 32 ; sad16x32_sse2 +SAD16XN 16 ; sad16x16_sse2 +SAD16XN 8 ; sad16x8_sse2 +SAD16XN 32, 1 ; sad16x32_avg_sse2 +SAD16XN 16, 1 ; sad16x16_avg_sse2 +SAD16XN 8, 1 ; sad16x8_avg_sse2 +SAD16XN 4 ; sad_16x4_sse2 +SAD16XN 4, 1 ; sad_16x4_avg_sse2 +SAD16XN 64 ; sad_16x64_sse2 +SAD16XN 64, 1 ; sad_16x64_avg_sse2 + +; unsigned int aom_sad8x{8,16}_sse2(uint8_t *src, int src_stride, +; uint8_t *ref, int ref_stride); +%macro SAD8XN 1-2 0 + SAD_FN 8, %1, 7, %2 + mov n_rowsd, %1/4 + pxor m0, m0 + +.loop: + movh m1, [refq] + movhps m1, [refq+ref_strideq] + movh m2, [refq+ref_strideq*2] + movhps m2, [refq+ref_stride3q] +%if %2 == 1 + pavgb m1, [second_predq+mmsize*0] + pavgb m2, [second_predq+mmsize*1] + lea second_predq, [second_predq+mmsize*2] +%endif + movh m3, [srcq] + movhps m3, [srcq+src_strideq] + movh m4, [srcq+src_strideq*2] + movhps m4, [srcq+src_stride3q] + psadbw m1, m3 + psadbw m2, m4 + lea refq, [refq+ref_strideq*4] + paddd m0, m1 + lea srcq, [srcq+src_strideq*4] + paddd m0, m2 + dec n_rowsd + jg .loop + + movhlps m1, m0 + paddd m0, m1 + movd eax, m0 + RET +%endmacro + +INIT_XMM sse2 +SAD8XN 16 ; sad8x16_sse2 +SAD8XN 8 ; sad8x8_sse2 +SAD8XN 4 ; sad8x4_sse2 +SAD8XN 16, 1 ; sad8x16_avg_sse2 +SAD8XN 8, 1 ; sad8x8_avg_sse2 +SAD8XN 4, 1 ; sad8x4_avg_sse2 +SAD8XN 32 ; sad_8x32_sse2 +SAD8XN 32, 1 ; sad_8x32_avg_sse2 + +; unsigned int aom_sad4x{4, 8}_sse2(uint8_t *src, int src_stride, +; uint8_t *ref, int ref_stride); +%macro SAD4XN 1-2 0 + SAD_FN 4, %1, 7, %2 + mov n_rowsd, %1/4 + pxor m0, m0 + +.loop: + movd m1, [refq] + movd m2, [refq+ref_strideq] + movd m3, [refq+ref_strideq*2] + movd m4, [refq+ref_stride3q] + punpckldq m1, m2 + punpckldq m3, m4 + movlhps m1, m3 +%if %2 == 1 + pavgb m1, [second_predq+mmsize*0] + lea second_predq, [second_predq+mmsize*1] +%endif + movd m2, [srcq] + movd m5, [srcq+src_strideq] + movd m4, [srcq+src_strideq*2] + movd m3, [srcq+src_stride3q] + punpckldq m2, m5 + punpckldq m4, m3 + movlhps m2, m4 + psadbw m1, m2 + lea refq, [refq+ref_strideq*4] + paddd m0, m1 + lea srcq, [srcq+src_strideq*4] + dec n_rowsd + jg .loop + + movhlps m1, m0 + paddd m0, m1 + movd eax, m0 + RET +%endmacro + +INIT_XMM sse2 +SAD4XN 8 ; sad4x8_sse +SAD4XN 4 ; sad4x4_sse +SAD4XN 8, 1 ; sad4x8_avg_sse +SAD4XN 4, 1 ; sad4x4_avg_sse +SAD4XN 16 ; sad_4x16_sse2 +SAD4XN 16, 1 ; sad_4x16_avg_sse2 diff --git a/third_party/aom/aom_dsp/x86/sse_avx2.c b/third_party/aom/aom_dsp/x86/sse_avx2.c new file mode 100644 index 0000000000..305dde5c08 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/sse_avx2.c @@ -0,0 +1,250 @@ +/* + * Copyright (c) 2018, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ +#include +#include + +#include "config/aom_dsp_rtcd.h" + +#include "aom_ports/mem.h" +#include "aom_dsp/x86/synonyms.h" +#include "aom_dsp/x86/synonyms_avx2.h" + +static INLINE void sse_w32_avx2(__m256i *sum, const uint8_t *a, + const uint8_t *b) { + const __m256i v_a0 = yy_loadu_256(a); + const __m256i v_b0 = yy_loadu_256(b); + const __m256i v_a00_w = _mm256_cvtepu8_epi16(_mm256_castsi256_si128(v_a0)); + const __m256i v_a01_w = + _mm256_cvtepu8_epi16(_mm256_extracti128_si256(v_a0, 1)); + const __m256i v_b00_w = _mm256_cvtepu8_epi16(_mm256_castsi256_si128(v_b0)); + const __m256i v_b01_w = + _mm256_cvtepu8_epi16(_mm256_extracti128_si256(v_b0, 1)); + const __m256i v_d00_w = _mm256_sub_epi16(v_a00_w, v_b00_w); + const __m256i v_d01_w = _mm256_sub_epi16(v_a01_w, v_b01_w); + *sum = _mm256_add_epi32(*sum, _mm256_madd_epi16(v_d00_w, v_d00_w)); + *sum = _mm256_add_epi32(*sum, _mm256_madd_epi16(v_d01_w, v_d01_w)); +} + +static INLINE int64_t summary_all_avx2(const __m256i *sum_all) { + int64_t sum; + const __m256i sum0_4x64 = + _mm256_cvtepu32_epi64(_mm256_castsi256_si128(*sum_all)); + const __m256i sum1_4x64 = + _mm256_cvtepu32_epi64(_mm256_extracti128_si256(*sum_all, 1)); + const __m256i sum_4x64 = _mm256_add_epi64(sum0_4x64, sum1_4x64); + const __m128i sum_2x64 = _mm_add_epi64(_mm256_castsi256_si128(sum_4x64), + _mm256_extracti128_si256(sum_4x64, 1)); + const __m128i sum_1x64 = _mm_add_epi64(sum_2x64, _mm_srli_si128(sum_2x64, 8)); + + xx_storel_64(&sum, sum_1x64); + return sum; +} + +int64_t aom_sse_avx2(const uint8_t *a, int a_stride, const uint8_t *b, + int b_stride, int width, int height) { + int32_t y = 0; + int64_t sse = 0; + __m256i sum = _mm256_setzero_si256(); + switch (width) { + case 4: + do { + const __m128i v_a0 = xx_loadl_32(a); + const __m128i v_a1 = xx_loadl_32(a + a_stride); + const __m128i v_a2 = xx_loadl_32(a + a_stride * 2); + const __m128i v_a3 = xx_loadl_32(a + a_stride * 3); + const __m128i v_b0 = xx_loadl_32(b); + const __m128i v_b1 = xx_loadl_32(b + b_stride); + const __m128i v_b2 = xx_loadl_32(b + b_stride * 2); + const __m128i v_b3 = xx_loadl_32(b + b_stride * 3); + const __m128i v_a0123 = _mm_unpacklo_epi64( + _mm_unpacklo_epi32(v_a0, v_a1), _mm_unpacklo_epi32(v_a2, v_a3)); + const __m128i v_b0123 = _mm_unpacklo_epi64( + _mm_unpacklo_epi32(v_b0, v_b1), _mm_unpacklo_epi32(v_b2, v_b3)); + const __m256i v_a_w = _mm256_cvtepu8_epi16(v_a0123); + const __m256i v_b_w = _mm256_cvtepu8_epi16(v_b0123); + const __m256i v_d_w = _mm256_sub_epi16(v_a_w, v_b_w); + sum = _mm256_add_epi32(sum, _mm256_madd_epi16(v_d_w, v_d_w)); + a += a_stride << 2; + b += b_stride << 2; + y += 4; + } while (y < height); + sse = summary_all_avx2(&sum); + break; + case 8: + do { + const __m128i v_a0 = xx_loadl_64(a); + const __m128i v_a1 = xx_loadl_64(a + a_stride); + const __m128i v_b0 = xx_loadl_64(b); + const __m128i v_b1 = xx_loadl_64(b + b_stride); + const __m256i v_a_w = + _mm256_cvtepu8_epi16(_mm_unpacklo_epi64(v_a0, v_a1)); + const __m256i v_b_w = + _mm256_cvtepu8_epi16(_mm_unpacklo_epi64(v_b0, v_b1)); + const __m256i v_d_w = _mm256_sub_epi16(v_a_w, v_b_w); + sum = _mm256_add_epi32(sum, _mm256_madd_epi16(v_d_w, v_d_w)); + a += a_stride << 1; + b += b_stride << 1; + y += 2; + } while (y < height); + sse = summary_all_avx2(&sum); + break; + case 16: + do { + const __m128i v_a0 = xx_loadu_128(a); + const __m128i v_b0 = xx_loadu_128(b); + const __m256i v_a_w = _mm256_cvtepu8_epi16(v_a0); + const __m256i v_b_w = _mm256_cvtepu8_epi16(v_b0); + const __m256i v_d_w = _mm256_sub_epi16(v_a_w, v_b_w); + sum = _mm256_add_epi32(sum, _mm256_madd_epi16(v_d_w, v_d_w)); + a += a_stride; + b += b_stride; + y += 1; + } while (y < height); + sse = summary_all_avx2(&sum); + break; + case 32: + do { + sse_w32_avx2(&sum, a, b); + a += a_stride; + b += b_stride; + y += 1; + } while (y < height); + sse = summary_all_avx2(&sum); + break; + case 64: + do { + sse_w32_avx2(&sum, a, b); + sse_w32_avx2(&sum, a + 32, b + 32); + a += a_stride; + b += b_stride; + y += 1; + } while (y < height); + sse = summary_all_avx2(&sum); + break; + case 128: + do { + sse_w32_avx2(&sum, a, b); + sse_w32_avx2(&sum, a + 32, b + 32); + sse_w32_avx2(&sum, a + 64, b + 64); + sse_w32_avx2(&sum, a + 96, b + 96); + a += a_stride; + b += b_stride; + y += 1; + } while (y < height); + sse = summary_all_avx2(&sum); + break; + default: break; + } + + return sse; +} + +static INLINE void highbd_sse_w16_avx2(__m256i *sum, const uint16_t *a, + const uint16_t *b) { + const __m256i v_a_w = yy_loadu_256(a); + const __m256i v_b_w = yy_loadu_256(b); + const __m256i v_d_w = _mm256_sub_epi16(v_a_w, v_b_w); + *sum = _mm256_add_epi32(*sum, _mm256_madd_epi16(v_d_w, v_d_w)); +} + +int64_t aom_highbd_sse_avx2(const uint8_t *a8, int a_stride, const uint8_t *b8, + int b_stride, int width, int height) { + int32_t y = 0; + int64_t sse = 0; + uint16_t *a = CONVERT_TO_SHORTPTR(a8); + uint16_t *b = CONVERT_TO_SHORTPTR(b8); + __m256i sum = _mm256_setzero_si256(); + switch (width) { + case 4: + do { + const __m128i v_a0 = xx_loadl_64(a); + const __m128i v_a1 = xx_loadl_64(a + a_stride); + const __m128i v_a2 = xx_loadl_64(a + a_stride * 2); + const __m128i v_a3 = xx_loadl_64(a + a_stride * 3); + const __m128i v_b0 = xx_loadl_64(b); + const __m128i v_b1 = xx_loadl_64(b + b_stride); + const __m128i v_b2 = xx_loadl_64(b + b_stride * 2); + const __m128i v_b3 = xx_loadl_64(b + b_stride * 3); + const __m256i v_a_w = yy_set_m128i(_mm_unpacklo_epi64(v_a0, v_a1), + _mm_unpacklo_epi64(v_a2, v_a3)); + const __m256i v_b_w = yy_set_m128i(_mm_unpacklo_epi64(v_b0, v_b1), + _mm_unpacklo_epi64(v_b2, v_b3)); + const __m256i v_d_w = _mm256_sub_epi16(v_a_w, v_b_w); + sum = _mm256_add_epi32(sum, _mm256_madd_epi16(v_d_w, v_d_w)); + a += a_stride << 2; + b += b_stride << 2; + y += 4; + } while (y < height); + sse = summary_all_avx2(&sum); + break; + case 8: + do { + const __m256i v_a_w = yy_loadu2_128(a + a_stride, a); + const __m256i v_b_w = yy_loadu2_128(b + b_stride, b); + const __m256i v_d_w = _mm256_sub_epi16(v_a_w, v_b_w); + sum = _mm256_add_epi32(sum, _mm256_madd_epi16(v_d_w, v_d_w)); + a += a_stride << 1; + b += b_stride << 1; + y += 2; + } while (y < height); + sse = summary_all_avx2(&sum); + break; + case 16: + do { + highbd_sse_w16_avx2(&sum, a, b); + a += a_stride; + b += b_stride; + y += 1; + } while (y < height); + sse = summary_all_avx2(&sum); + break; + case 32: + do { + highbd_sse_w16_avx2(&sum, a, b); + highbd_sse_w16_avx2(&sum, a + 16, b + 16); + a += a_stride; + b += b_stride; + y += 1; + } while (y < height); + sse = summary_all_avx2(&sum); + break; + case 64: + do { + highbd_sse_w16_avx2(&sum, a, b); + highbd_sse_w16_avx2(&sum, a + 16 * 1, b + 16 * 1); + highbd_sse_w16_avx2(&sum, a + 16 * 2, b + 16 * 2); + highbd_sse_w16_avx2(&sum, a + 16 * 3, b + 16 * 3); + a += a_stride; + b += b_stride; + y += 1; + } while (y < height); + sse = summary_all_avx2(&sum); + break; + case 128: + do { + highbd_sse_w16_avx2(&sum, a, b); + highbd_sse_w16_avx2(&sum, a + 16 * 1, b + 16 * 1); + highbd_sse_w16_avx2(&sum, a + 16 * 2, b + 16 * 2); + highbd_sse_w16_avx2(&sum, a + 16 * 3, b + 16 * 3); + highbd_sse_w16_avx2(&sum, a + 16 * 4, b + 16 * 4); + highbd_sse_w16_avx2(&sum, a + 16 * 5, b + 16 * 5); + highbd_sse_w16_avx2(&sum, a + 16 * 6, b + 16 * 6); + highbd_sse_w16_avx2(&sum, a + 16 * 7, b + 16 * 7); + a += a_stride; + b += b_stride; + y += 1; + } while (y < height); + sse = summary_all_avx2(&sum); + break; + default: break; + } + return sse; +} diff --git a/third_party/aom/aom_dsp/x86/sse_sse4.c b/third_party/aom/aom_dsp/x86/sse_sse4.c new file mode 100644 index 0000000000..8b5af84691 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/sse_sse4.c @@ -0,0 +1,241 @@ +/* + * Copyright (c) 2018, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include +#include + +#include "config/aom_config.h" + +#include "aom_ports/mem.h" +#include "aom/aom_integer.h" +#include "aom_dsp/x86/synonyms.h" + +static INLINE int64_t summary_all_sse4(const __m128i *sum_all) { + int64_t sum; + const __m128i sum0 = _mm_cvtepu32_epi64(*sum_all); + const __m128i sum1 = _mm_cvtepu32_epi64(_mm_srli_si128(*sum_all, 8)); + const __m128i sum_2x64 = _mm_add_epi64(sum0, sum1); + const __m128i sum_1x64 = _mm_add_epi64(sum_2x64, _mm_srli_si128(sum_2x64, 8)); + xx_storel_64(&sum, sum_1x64); + return sum; +} + +static INLINE void sse_w16_sse4_1(__m128i *sum, const uint8_t *a, + const uint8_t *b) { + const __m128i v_a0 = xx_loadu_128(a); + const __m128i v_b0 = xx_loadu_128(b); + const __m128i v_a00_w = _mm_cvtepu8_epi16(v_a0); + const __m128i v_a01_w = _mm_cvtepu8_epi16(_mm_srli_si128(v_a0, 8)); + const __m128i v_b00_w = _mm_cvtepu8_epi16(v_b0); + const __m128i v_b01_w = _mm_cvtepu8_epi16(_mm_srli_si128(v_b0, 8)); + const __m128i v_d00_w = _mm_sub_epi16(v_a00_w, v_b00_w); + const __m128i v_d01_w = _mm_sub_epi16(v_a01_w, v_b01_w); + *sum = _mm_add_epi32(*sum, _mm_madd_epi16(v_d00_w, v_d00_w)); + *sum = _mm_add_epi32(*sum, _mm_madd_epi16(v_d01_w, v_d01_w)); +} + +int64_t aom_sse_sse4_1(const uint8_t *a, int a_stride, const uint8_t *b, + int b_stride, int width, int height) { + int y = 0; + int64_t sse = 0; + __m128i sum = _mm_setzero_si128(); + switch (width) { + case 4: + do { + const __m128i v_a0 = xx_loadl_32(a); + const __m128i v_a1 = xx_loadl_32(a + a_stride); + const __m128i v_b0 = xx_loadl_32(b); + const __m128i v_b1 = xx_loadl_32(b + b_stride); + const __m128i v_a_w = _mm_cvtepu8_epi16(_mm_unpacklo_epi32(v_a0, v_a1)); + const __m128i v_b_w = _mm_cvtepu8_epi16(_mm_unpacklo_epi32(v_b0, v_b1)); + const __m128i v_d_w = _mm_sub_epi16(v_a_w, v_b_w); + sum = _mm_add_epi32(sum, _mm_madd_epi16(v_d_w, v_d_w)); + a += a_stride << 1; + b += b_stride << 1; + y += 2; + } while (y < height); + sse = summary_all_sse4(&sum); + break; + case 8: + do { + const __m128i v_a0 = xx_loadl_64(a); + const __m128i v_b0 = xx_loadl_64(b); + const __m128i v_a_w = _mm_cvtepu8_epi16(v_a0); + const __m128i v_b_w = _mm_cvtepu8_epi16(v_b0); + const __m128i v_d_w = _mm_sub_epi16(v_a_w, v_b_w); + sum = _mm_add_epi32(sum, _mm_madd_epi16(v_d_w, v_d_w)); + a += a_stride; + b += b_stride; + y += 1; + } while (y < height); + sse = summary_all_sse4(&sum); + break; + case 16: + do { + sse_w16_sse4_1(&sum, a, b); + a += a_stride; + b += b_stride; + y += 1; + } while (y < height); + sse = summary_all_sse4(&sum); + break; + case 32: + do { + sse_w16_sse4_1(&sum, a, b); + sse_w16_sse4_1(&sum, a + 16, b + 16); + a += a_stride; + b += b_stride; + y += 1; + } while (y < height); + sse = summary_all_sse4(&sum); + break; + case 64: + do { + sse_w16_sse4_1(&sum, a, b); + sse_w16_sse4_1(&sum, a + 16 * 1, b + 16 * 1); + sse_w16_sse4_1(&sum, a + 16 * 2, b + 16 * 2); + sse_w16_sse4_1(&sum, a + 16 * 3, b + 16 * 3); + a += a_stride; + b += b_stride; + y += 1; + } while (y < height); + sse = summary_all_sse4(&sum); + break; + case 128: + do { + sse_w16_sse4_1(&sum, a, b); + sse_w16_sse4_1(&sum, a + 16 * 1, b + 16 * 1); + sse_w16_sse4_1(&sum, a + 16 * 2, b + 16 * 2); + sse_w16_sse4_1(&sum, a + 16 * 3, b + 16 * 3); + sse_w16_sse4_1(&sum, a + 16 * 4, b + 16 * 4); + sse_w16_sse4_1(&sum, a + 16 * 5, b + 16 * 5); + sse_w16_sse4_1(&sum, a + 16 * 6, b + 16 * 6); + sse_w16_sse4_1(&sum, a + 16 * 7, b + 16 * 7); + a += a_stride; + b += b_stride; + y += 1; + } while (y < height); + sse = summary_all_sse4(&sum); + break; + default: break; + } + + return sse; +} + +static INLINE void highbd_sse_w8_sse4_1(__m128i *sum, const uint16_t *a, + const uint16_t *b) { + const __m128i v_a_w = xx_loadu_128(a); + const __m128i v_b_w = xx_loadu_128(b); + const __m128i v_d_w = _mm_sub_epi16(v_a_w, v_b_w); + *sum = _mm_add_epi32(*sum, _mm_madd_epi16(v_d_w, v_d_w)); +} + +int64_t aom_highbd_sse_sse4_1(const uint8_t *a8, int a_stride, + const uint8_t *b8, int b_stride, int width, + int height) { + int32_t y = 0; + int64_t sse = 0; + uint16_t *a = CONVERT_TO_SHORTPTR(a8); + uint16_t *b = CONVERT_TO_SHORTPTR(b8); + __m128i sum = _mm_setzero_si128(); + switch (width) { + case 4: + do { + const __m128i v_a0 = xx_loadl_64(a); + const __m128i v_a1 = xx_loadl_64(a + a_stride); + const __m128i v_b0 = xx_loadl_64(b); + const __m128i v_b1 = xx_loadl_64(b + b_stride); + const __m128i v_a_w = _mm_unpacklo_epi64(v_a0, v_a1); + const __m128i v_b_w = _mm_unpacklo_epi64(v_b0, v_b1); + const __m128i v_d_w = _mm_sub_epi16(v_a_w, v_b_w); + sum = _mm_add_epi32(sum, _mm_madd_epi16(v_d_w, v_d_w)); + a += a_stride << 1; + b += b_stride << 1; + y += 2; + } while (y < height); + sse = summary_all_sse4(&sum); + break; + case 8: + do { + highbd_sse_w8_sse4_1(&sum, a, b); + a += a_stride; + b += b_stride; + y += 1; + } while (y < height); + sse = summary_all_sse4(&sum); + break; + case 16: + do { + highbd_sse_w8_sse4_1(&sum, a, b); + highbd_sse_w8_sse4_1(&sum, a + 8, b + 8); + a += a_stride; + b += b_stride; + y += 1; + } while (y < height); + sse = summary_all_sse4(&sum); + break; + case 32: + do { + highbd_sse_w8_sse4_1(&sum, a, b); + highbd_sse_w8_sse4_1(&sum, a + 8 * 1, b + 8 * 1); + highbd_sse_w8_sse4_1(&sum, a + 8 * 2, b + 8 * 2); + highbd_sse_w8_sse4_1(&sum, a + 8 * 3, b + 8 * 3); + a += a_stride; + b += b_stride; + y += 1; + } while (y < height); + sse = summary_all_sse4(&sum); + break; + case 64: + do { + highbd_sse_w8_sse4_1(&sum, a, b); + highbd_sse_w8_sse4_1(&sum, a + 8 * 1, b + 8 * 1); + highbd_sse_w8_sse4_1(&sum, a + 8 * 2, b + 8 * 2); + highbd_sse_w8_sse4_1(&sum, a + 8 * 3, b + 8 * 3); + highbd_sse_w8_sse4_1(&sum, a + 8 * 4, b + 8 * 4); + highbd_sse_w8_sse4_1(&sum, a + 8 * 5, b + 8 * 5); + highbd_sse_w8_sse4_1(&sum, a + 8 * 6, b + 8 * 6); + highbd_sse_w8_sse4_1(&sum, a + 8 * 7, b + 8 * 7); + a += a_stride; + b += b_stride; + y += 1; + } while (y < height); + sse = summary_all_sse4(&sum); + break; + case 128: + do { + highbd_sse_w8_sse4_1(&sum, a, b); + highbd_sse_w8_sse4_1(&sum, a + 8 * 1, b + 8 * 1); + highbd_sse_w8_sse4_1(&sum, a + 8 * 2, b + 8 * 2); + highbd_sse_w8_sse4_1(&sum, a + 8 * 3, b + 8 * 3); + highbd_sse_w8_sse4_1(&sum, a + 8 * 4, b + 8 * 4); + highbd_sse_w8_sse4_1(&sum, a + 8 * 5, b + 8 * 5); + highbd_sse_w8_sse4_1(&sum, a + 8 * 6, b + 8 * 6); + highbd_sse_w8_sse4_1(&sum, a + 8 * 7, b + 8 * 7); + highbd_sse_w8_sse4_1(&sum, a + 8 * 8, b + 8 * 8); + highbd_sse_w8_sse4_1(&sum, a + 8 * 9, b + 8 * 9); + highbd_sse_w8_sse4_1(&sum, a + 8 * 10, b + 8 * 10); + highbd_sse_w8_sse4_1(&sum, a + 8 * 11, b + 8 * 11); + highbd_sse_w8_sse4_1(&sum, a + 8 * 12, b + 8 * 12); + highbd_sse_w8_sse4_1(&sum, a + 8 * 13, b + 8 * 13); + highbd_sse_w8_sse4_1(&sum, a + 8 * 14, b + 8 * 14); + highbd_sse_w8_sse4_1(&sum, a + 8 * 15, b + 8 * 15); + a += a_stride; + b += b_stride; + y += 1; + } while (y < height); + sse = summary_all_sse4(&sum); + break; + default: break; + } + return sse; +} diff --git a/third_party/aom/aom_dsp/x86/ssim_opt_x86_64.asm b/third_party/aom/aom_dsp/x86/ssim_opt_x86_64.asm new file mode 100644 index 0000000000..6d9b5a12f1 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/ssim_opt_x86_64.asm @@ -0,0 +1,222 @@ +; +; Copyright (c) 2016, Alliance for Open Media. All rights reserved +; +; This source code is subject to the terms of the BSD 2 Clause License and +; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License +; was not distributed with this source code in the LICENSE file, you can +; obtain it at www.aomedia.org/license/software. If the Alliance for Open +; Media Patent License 1.0 was not distributed with this source code in the +; PATENTS file, you can obtain it at www.aomedia.org/license/patent. +; + +; + +%include "aom_ports/x86_abi_support.asm" + +; tabulate_ssim - sums sum_s,sum_r,sum_sq_s,sum_sq_r, sum_sxr +%macro TABULATE_SSIM 0 + paddusw xmm15, xmm3 ; sum_s + paddusw xmm14, xmm4 ; sum_r + movdqa xmm1, xmm3 + pmaddwd xmm1, xmm1 + paddd xmm13, xmm1 ; sum_sq_s + movdqa xmm2, xmm4 + pmaddwd xmm2, xmm2 + paddd xmm12, xmm2 ; sum_sq_r + pmaddwd xmm3, xmm4 + paddd xmm11, xmm3 ; sum_sxr +%endmacro + +; Sum across the register %1 starting with q words +%macro SUM_ACROSS_Q 1 + movdqa xmm2,%1 + punpckldq %1,xmm0 + punpckhdq xmm2,xmm0 + paddq %1,xmm2 + movdqa xmm2,%1 + punpcklqdq %1,xmm0 + punpckhqdq xmm2,xmm0 + paddq %1,xmm2 +%endmacro + +; Sum across the register %1 starting with q words +%macro SUM_ACROSS_W 1 + movdqa xmm1, %1 + punpcklwd %1,xmm0 + punpckhwd xmm1,xmm0 + paddd %1, xmm1 + SUM_ACROSS_Q %1 +%endmacro + +SECTION .text + +;void ssim_parms_sse2( +; unsigned char *s, +; int sp, +; unsigned char *r, +; int rp +; uint32_t *sum_s, +; uint32_t *sum_r, +; uint32_t *sum_sq_s, +; uint32_t *sum_sq_r, +; uint32_t *sum_sxr); +; +; TODO: Use parm passing through structure, probably don't need the pxors +; ( calling app will initialize to 0 ) could easily fit everything in sse2 +; without too much hastle, and can probably do better estimates with psadw +; or pavgb At this point this is just meant to be first pass for calculating +; all the parms needed for 16x16 ssim so we can play with dssim as distortion +; in mode selection code. +global sym(aom_ssim_parms_16x16_sse2) PRIVATE +sym(aom_ssim_parms_16x16_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 9 + SAVE_XMM 15 + push rsi + push rdi + ; end prolog + + mov rsi, arg(0) ;s + mov rcx, arg(1) ;sp + mov rdi, arg(2) ;r + mov rax, arg(3) ;rp + + pxor xmm0, xmm0 + pxor xmm15,xmm15 ;sum_s + pxor xmm14,xmm14 ;sum_r + pxor xmm13,xmm13 ;sum_sq_s + pxor xmm12,xmm12 ;sum_sq_r + pxor xmm11,xmm11 ;sum_sxr + + mov rdx, 16 ;row counter +.NextRow: + + ;grab source and reference pixels + movdqu xmm5, [rsi] + movdqu xmm6, [rdi] + movdqa xmm3, xmm5 + movdqa xmm4, xmm6 + punpckhbw xmm3, xmm0 ; high_s + punpckhbw xmm4, xmm0 ; high_r + + TABULATE_SSIM + + movdqa xmm3, xmm5 + movdqa xmm4, xmm6 + punpcklbw xmm3, xmm0 ; low_s + punpcklbw xmm4, xmm0 ; low_r + + TABULATE_SSIM + + add rsi, rcx ; next s row + add rdi, rax ; next r row + + dec rdx ; counter + jnz .NextRow + + SUM_ACROSS_W xmm15 + SUM_ACROSS_W xmm14 + SUM_ACROSS_Q xmm13 + SUM_ACROSS_Q xmm12 + SUM_ACROSS_Q xmm11 + + mov rdi,arg(4) + movd [rdi], xmm15; + mov rdi,arg(5) + movd [rdi], xmm14; + mov rdi,arg(6) + movd [rdi], xmm13; + mov rdi,arg(7) + movd [rdi], xmm12; + mov rdi,arg(8) + movd [rdi], xmm11; + + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +;void ssim_parms_sse2( +; unsigned char *s, +; int sp, +; unsigned char *r, +; int rp +; uint32_t *sum_s, +; uint32_t *sum_r, +; uint32_t *sum_sq_s, +; uint32_t *sum_sq_r, +; uint32_t *sum_sxr); +; +; TODO: Use parm passing through structure, probably don't need the pxors +; ( calling app will initialize to 0 ) could easily fit everything in sse2 +; without too much hastle, and can probably do better estimates with psadw +; or pavgb At this point this is just meant to be first pass for calculating +; all the parms needed for 16x16 ssim so we can play with dssim as distortion +; in mode selection code. +global sym(aom_ssim_parms_8x8_sse2) PRIVATE +sym(aom_ssim_parms_8x8_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 9 + SAVE_XMM 15 + push rsi + push rdi + ; end prolog + + mov rsi, arg(0) ;s + mov rcx, arg(1) ;sp + mov rdi, arg(2) ;r + mov rax, arg(3) ;rp + + pxor xmm0, xmm0 + pxor xmm15,xmm15 ;sum_s + pxor xmm14,xmm14 ;sum_r + pxor xmm13,xmm13 ;sum_sq_s + pxor xmm12,xmm12 ;sum_sq_r + pxor xmm11,xmm11 ;sum_sxr + + mov rdx, 8 ;row counter +.NextRow: + + ;grab source and reference pixels + movq xmm3, [rsi] + movq xmm4, [rdi] + punpcklbw xmm3, xmm0 ; low_s + punpcklbw xmm4, xmm0 ; low_r + + TABULATE_SSIM + + add rsi, rcx ; next s row + add rdi, rax ; next r row + + dec rdx ; counter + jnz .NextRow + + SUM_ACROSS_W xmm15 + SUM_ACROSS_W xmm14 + SUM_ACROSS_Q xmm13 + SUM_ACROSS_Q xmm12 + SUM_ACROSS_Q xmm11 + + mov rdi,arg(4) + movd [rdi], xmm15; + mov rdi,arg(5) + movd [rdi], xmm14; + mov rdi,arg(6) + movd [rdi], xmm13; + mov rdi,arg(7) + movd [rdi], xmm12; + mov rdi,arg(8) + movd [rdi], xmm11; + + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret diff --git a/third_party/aom/aom_dsp/x86/subpel_variance_sse2.asm b/third_party/aom/aom_dsp/x86/subpel_variance_sse2.asm new file mode 100644 index 0000000000..45bf6ec3c5 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/subpel_variance_sse2.asm @@ -0,0 +1,1481 @@ +; +; Copyright (c) 2016, Alliance for Open Media. All rights reserved +; +; This source code is subject to the terms of the BSD 2 Clause License and +; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License +; was not distributed with this source code in the LICENSE file, you can +; obtain it at www.aomedia.org/license/software. If the Alliance for Open +; Media Patent License 1.0 was not distributed with this source code in the +; PATENTS file, you can obtain it at www.aomedia.org/license/patent. +; + +; + +%include "third_party/x86inc/x86inc.asm" + +SECTION_RODATA +pw_8: times 8 dw 8 +bilin_filter_m_sse2: times 8 dw 16 + times 8 dw 0 + times 8 dw 14 + times 8 dw 2 + times 8 dw 12 + times 8 dw 4 + times 8 dw 10 + times 8 dw 6 + times 16 dw 8 + times 8 dw 6 + times 8 dw 10 + times 8 dw 4 + times 8 dw 12 + times 8 dw 2 + times 8 dw 14 + +bilin_filter_m_ssse3: times 8 db 16, 0 + times 8 db 14, 2 + times 8 db 12, 4 + times 8 db 10, 6 + times 16 db 8 + times 8 db 6, 10 + times 8 db 4, 12 + times 8 db 2, 14 + +SECTION .text + +; int aom_sub_pixel_varianceNxh(const uint8_t *src, ptrdiff_t src_stride, +; int x_offset, int y_offset, +; const uint8_t *dst, ptrdiff_t dst_stride, +; int height, unsigned int *sse); +; +; This function returns the SE and stores SSE in the given pointer. + +%macro SUM_SSE 6 ; src1, dst1, src2, dst2, sum, sse + psubw %3, %4 + psubw %1, %2 + paddw %5, %3 + pmaddwd %3, %3 + paddw %5, %1 + pmaddwd %1, %1 + paddd %6, %3 + paddd %6, %1 +%endmacro + +%macro STORE_AND_RET 1 +%if %1 > 4 + ; if H=64 and W=16, we have 8 words of each 2(1bit)x64(6bit)x9bit=16bit + ; in m6, i.e. it _exactly_ fits in a signed word per word in the xmm reg. + ; We have to sign-extend it before adding the words within the register + ; and outputing to a dword. + pcmpgtw m5, m6 ; mask for 0 > x + movhlps m3, m7 + punpcklwd m4, m6, m5 + punpckhwd m6, m5 ; sign-extend m6 word->dword + paddd m7, m3 + paddd m6, m4 + pshufd m3, m7, 0x1 + movhlps m4, m6 + paddd m7, m3 + paddd m6, m4 + mov r1, ssem ; r1 = unsigned int *sse + pshufd m4, m6, 0x1 + movd [r1], m7 ; store sse + paddd m6, m4 + movd raxd, m6 ; store sum as return value +%else ; 4xh + pshuflw m4, m6, 0xe + pshuflw m3, m7, 0xe + paddw m6, m4 + paddd m7, m3 + pcmpgtw m5, m6 ; mask for 0 > x + mov r1, ssem ; r1 = unsigned int *sse + punpcklwd m6, m5 ; sign-extend m6 word->dword + movd [r1], m7 ; store sse + pshuflw m4, m6, 0xe + paddd m6, m4 + movd raxd, m6 ; store sum as return value +%endif + RET +%endmacro + +%macro INC_SRC_BY_SRC_STRIDE 0 +%if ARCH_X86=1 && CONFIG_PIC=1 + add srcq, src_stridemp +%else + add srcq, src_strideq +%endif +%endmacro + +%macro SUBPEL_VARIANCE 1-2 0 ; W +%if cpuflag(ssse3) +%define bilin_filter_m bilin_filter_m_ssse3 +%define filter_idx_shift 4 +%else +%define bilin_filter_m bilin_filter_m_sse2 +%define filter_idx_shift 5 +%endif +; FIXME(rbultje) only bilinear filters use >8 registers, and ssse3 only uses +; 11, not 13, if the registers are ordered correctly. May make a minor speed +; difference on Win64 + +%if ARCH_X86_64 + %if %2 == 1 ; avg + cglobal sub_pixel_avg_variance%1xh, 9, 10, 13, src, src_stride, \ + x_offset, y_offset, dst, dst_stride, \ + sec, sec_stride, height, sse + %define sec_str sec_strideq + %else + cglobal sub_pixel_variance%1xh, 7, 8, 13, src, src_stride, \ + x_offset, y_offset, dst, dst_stride, \ + height, sse + %endif + %define block_height heightd + %define bilin_filter sseq +%else + %if CONFIG_PIC=1 + %if %2 == 1 ; avg + cglobal sub_pixel_avg_variance%1xh, 7, 7, 13, src, src_stride, \ + x_offset, y_offset, dst, dst_stride, \ + sec, sec_stride, height, sse, \ + g_bilin_filter, g_pw_8 + %define block_height dword heightm + %define sec_str sec_stridemp + + ;Store bilin_filter and pw_8 location in stack + %if GET_GOT_DEFINED == 1 + GET_GOT eax + add esp, 4 ; restore esp + %endif + + lea ecx, [GLOBAL(bilin_filter_m)] + mov g_bilin_filterm, ecx + + lea ecx, [GLOBAL(pw_8)] + mov g_pw_8m, ecx + + LOAD_IF_USED 0, 1 ; load eax, ecx back + %else + cglobal sub_pixel_variance%1xh, 7, 7, 13, src, src_stride, \ + x_offset, y_offset, dst, dst_stride, \ + height, sse, g_bilin_filter, g_pw_8 + %define block_height heightd + + ;Store bilin_filter and pw_8 location in stack + %if GET_GOT_DEFINED == 1 + GET_GOT eax + add esp, 4 ; restore esp + %endif + + lea ecx, [GLOBAL(bilin_filter_m)] + mov g_bilin_filterm, ecx + + lea ecx, [GLOBAL(pw_8)] + mov g_pw_8m, ecx + + LOAD_IF_USED 0, 1 ; load eax, ecx back + %endif + %else + %if %2 == 1 ; avg + cglobal sub_pixel_avg_variance%1xh, 7, 7, 13, src, src_stride, \ + x_offset, y_offset, \ + dst, dst_stride, sec, sec_stride, \ + height, sse + %define block_height dword heightm + %define sec_str sec_stridemp + %else + cglobal sub_pixel_variance%1xh, 7, 7, 13, src, src_stride, \ + x_offset, y_offset, dst, dst_stride, \ + height, sse + %define block_height heightd + %endif + %define bilin_filter bilin_filter_m + %endif +%endif + +%if %1 == 4 + %define movx movd +%else + %define movx movh +%endif + + ASSERT %1 <= 16 ; m6 overflows if w > 16 + pxor m6, m6 ; sum + pxor m7, m7 ; sse + ; FIXME(rbultje) if both filters are bilinear, we don't actually use m5; we + ; could perhaps use it for something more productive then + pxor m5, m5 ; dedicated zero register +%if %1 < 16 + sar block_height, 1 +%if %2 == 1 ; avg + shl sec_str, 1 +%endif +%endif + + ; FIXME(rbultje) replace by jumptable? + test x_offsetd, x_offsetd + jnz .x_nonzero + ; x_offset == 0 + test y_offsetd, y_offsetd + jnz .x_zero_y_nonzero + + ; x_offset == 0 && y_offset == 0 +.x_zero_y_zero_loop: +%if %1 == 16 + movu m0, [srcq] + mova m1, [dstq] +%if %2 == 1 ; avg + pavgb m0, [secq] + punpckhbw m3, m1, m5 + punpcklbw m1, m5 +%endif + punpckhbw m2, m0, m5 + punpcklbw m0, m5 + +%if %2 == 0 ; !avg + punpckhbw m3, m1, m5 + punpcklbw m1, m5 +%endif + SUM_SSE m0, m1, m2, m3, m6, m7 + + add srcq, src_strideq + add dstq, dst_strideq +%else ; %1 < 16 + movx m0, [srcq] +%if %2 == 1 ; avg +%if %1 > 4 + movhps m0, [srcq+src_strideq] +%else ; 4xh + movx m1, [srcq+src_strideq] + punpckldq m0, m1 +%endif +%else ; !avg + movx m2, [srcq+src_strideq] +%endif + + movx m1, [dstq] + movx m3, [dstq+dst_strideq] + +%if %2 == 1 ; avg +%if %1 > 4 + pavgb m0, [secq] +%else + movh m2, [secq] + pavgb m0, m2 +%endif + punpcklbw m3, m5 + punpcklbw m1, m5 +%if %1 > 4 + punpckhbw m2, m0, m5 + punpcklbw m0, m5 +%else ; 4xh + punpcklbw m0, m5 + movhlps m2, m0 +%endif +%else ; !avg + punpcklbw m0, m5 + punpcklbw m2, m5 + punpcklbw m3, m5 + punpcklbw m1, m5 +%endif + SUM_SSE m0, m1, m2, m3, m6, m7 + + lea srcq, [srcq+src_strideq*2] + lea dstq, [dstq+dst_strideq*2] +%endif +%if %2 == 1 ; avg + add secq, sec_str +%endif + dec block_height + jg .x_zero_y_zero_loop + STORE_AND_RET %1 + +.x_zero_y_nonzero: + cmp y_offsetd, 4 + jne .x_zero_y_nonhalf + + ; x_offset == 0 && y_offset == 0.5 +.x_zero_y_half_loop: +%if %1 == 16 + movu m0, [srcq] + movu m4, [srcq+src_strideq] + mova m1, [dstq] + pavgb m0, m4 + punpckhbw m3, m1, m5 +%if %2 == 1 ; avg + pavgb m0, [secq] +%endif + punpcklbw m1, m5 + punpckhbw m2, m0, m5 + punpcklbw m0, m5 + SUM_SSE m0, m1, m2, m3, m6, m7 + + add srcq, src_strideq + add dstq, dst_strideq +%else ; %1 < 16 + movx m0, [srcq] + movx m2, [srcq+src_strideq] +%if %2 == 1 ; avg +%if %1 > 4 + movhps m2, [srcq+src_strideq*2] +%else ; 4xh + movx m1, [srcq+src_strideq*2] + punpckldq m2, m1 +%endif + movx m1, [dstq] +%if %1 > 4 + movlhps m0, m2 +%else ; 4xh + punpckldq m0, m2 +%endif + movx m3, [dstq+dst_strideq] + pavgb m0, m2 + punpcklbw m1, m5 +%if %1 > 4 + pavgb m0, [secq] + punpcklbw m3, m5 + punpckhbw m2, m0, m5 + punpcklbw m0, m5 +%else ; 4xh + movh m4, [secq] + pavgb m0, m4 + punpcklbw m3, m5 + punpcklbw m0, m5 + movhlps m2, m0 +%endif +%else ; !avg + movx m4, [srcq+src_strideq*2] + movx m1, [dstq] + pavgb m0, m2 + movx m3, [dstq+dst_strideq] + pavgb m2, m4 + punpcklbw m0, m5 + punpcklbw m2, m5 + punpcklbw m3, m5 + punpcklbw m1, m5 +%endif + SUM_SSE m0, m1, m2, m3, m6, m7 + + lea srcq, [srcq+src_strideq*2] + lea dstq, [dstq+dst_strideq*2] +%endif +%if %2 == 1 ; avg + add secq, sec_str +%endif + dec block_height + jg .x_zero_y_half_loop + STORE_AND_RET %1 + +.x_zero_y_nonhalf: + ; x_offset == 0 && y_offset == bilin interpolation +%if ARCH_X86_64 + lea bilin_filter, [GLOBAL(bilin_filter_m)] +%endif + shl y_offsetd, filter_idx_shift +%if ARCH_X86_64 && %1 > 4 + mova m8, [bilin_filter+y_offsetq] +%if notcpuflag(ssse3) ; FIXME(rbultje) don't scatter registers on x86-64 + mova m9, [bilin_filter+y_offsetq+16] +%endif + mova m10, [GLOBAL(pw_8)] +%define filter_y_a m8 +%define filter_y_b m9 +%define filter_rnd m10 +%else ; x86-32 or mmx +%if ARCH_X86=1 && CONFIG_PIC=1 +; x_offset == 0, reuse x_offset reg +%define tempq x_offsetq + add y_offsetq, g_bilin_filterm +%define filter_y_a [y_offsetq] +%define filter_y_b [y_offsetq+16] + mov tempq, g_pw_8m +%define filter_rnd [tempq] +%else + add y_offsetq, bilin_filter +%define filter_y_a [y_offsetq] +%define filter_y_b [y_offsetq+16] +%define filter_rnd [GLOBAL(pw_8)] +%endif +%endif + +.x_zero_y_other_loop: +%if %1 == 16 + movu m0, [srcq] + movu m4, [srcq+src_strideq] + mova m1, [dstq] +%if cpuflag(ssse3) + punpckhbw m2, m0, m4 + punpcklbw m0, m4 + pmaddubsw m2, filter_y_a + pmaddubsw m0, filter_y_a + paddw m2, filter_rnd + paddw m0, filter_rnd +%else + punpckhbw m2, m0, m5 + punpckhbw m3, m4, m5 + punpcklbw m0, m5 + punpcklbw m4, m5 + ; FIXME(rbultje) instead of out=((num-x)*in1+x*in2+rnd)>>log2(num), we can + ; also do out=in1+(((num-x)*(in2-in1)+rnd)>>log2(num)). Total number of + ; instructions is the same (5), but it is 1 mul instead of 2, so might be + ; slightly faster because of pmullw latency. It would also cut our rodata + ; tables in half for this function, and save 1-2 registers on x86-64. + pmullw m2, filter_y_a + pmullw m3, filter_y_b + paddw m2, filter_rnd + pmullw m0, filter_y_a + pmullw m4, filter_y_b + paddw m0, filter_rnd + paddw m2, m3 + paddw m0, m4 +%endif + psraw m2, 4 + psraw m0, 4 +%if %2 == 1 ; avg + ; FIXME(rbultje) pipeline + packuswb m0, m2 + pavgb m0, [secq] + punpckhbw m2, m0, m5 + punpcklbw m0, m5 +%endif + punpckhbw m3, m1, m5 + punpcklbw m1, m5 + SUM_SSE m0, m1, m2, m3, m6, m7 + + add srcq, src_strideq + add dstq, dst_strideq +%else ; %1 < 16 + movx m0, [srcq] + movx m2, [srcq+src_strideq] + movx m4, [srcq+src_strideq*2] + movx m3, [dstq+dst_strideq] +%if cpuflag(ssse3) + movx m1, [dstq] + punpcklbw m0, m2 + punpcklbw m2, m4 + pmaddubsw m0, filter_y_a + pmaddubsw m2, filter_y_a + punpcklbw m3, m5 + paddw m2, filter_rnd + paddw m0, filter_rnd +%else + punpcklbw m0, m5 + punpcklbw m2, m5 + punpcklbw m4, m5 + pmullw m0, filter_y_a + pmullw m1, m2, filter_y_b + punpcklbw m3, m5 + paddw m0, filter_rnd + pmullw m2, filter_y_a + pmullw m4, filter_y_b + paddw m0, m1 + paddw m2, filter_rnd + movx m1, [dstq] + paddw m2, m4 +%endif + psraw m0, 4 + psraw m2, 4 +%if %2 == 1 ; avg + ; FIXME(rbultje) pipeline +%if %1 == 4 + movlhps m0, m2 +%endif + packuswb m0, m2 +%if %1 > 4 + pavgb m0, [secq] + punpckhbw m2, m0, m5 + punpcklbw m0, m5 +%else ; 4xh + movh m2, [secq] + pavgb m0, m2 + punpcklbw m0, m5 + movhlps m2, m0 +%endif +%endif + punpcklbw m1, m5 + SUM_SSE m0, m1, m2, m3, m6, m7 + + lea srcq, [srcq+src_strideq*2] + lea dstq, [dstq+dst_strideq*2] +%endif +%if %2 == 1 ; avg + add secq, sec_str +%endif + dec block_height + jg .x_zero_y_other_loop +%undef filter_y_a +%undef filter_y_b +%undef filter_rnd + STORE_AND_RET %1 + +.x_nonzero: + cmp x_offsetd, 4 + jne .x_nonhalf + ; x_offset == 0.5 + test y_offsetd, y_offsetd + jnz .x_half_y_nonzero + + ; x_offset == 0.5 && y_offset == 0 +.x_half_y_zero_loop: +%if %1 == 16 + movu m0, [srcq] + movu m4, [srcq+1] + mova m1, [dstq] + pavgb m0, m4 + punpckhbw m3, m1, m5 +%if %2 == 1 ; avg + pavgb m0, [secq] +%endif + punpcklbw m1, m5 + punpckhbw m2, m0, m5 + punpcklbw m0, m5 + SUM_SSE m0, m1, m2, m3, m6, m7 + + add srcq, src_strideq + add dstq, dst_strideq +%else ; %1 < 16 + movx m0, [srcq] + movx m4, [srcq+1] +%if %2 == 1 ; avg +%if %1 > 4 + movhps m0, [srcq+src_strideq] + movhps m4, [srcq+src_strideq+1] +%else ; 4xh + movx m1, [srcq+src_strideq] + punpckldq m0, m1 + movx m2, [srcq+src_strideq+1] + punpckldq m4, m2 +%endif + movx m1, [dstq] + movx m3, [dstq+dst_strideq] + pavgb m0, m4 + punpcklbw m3, m5 +%if %1 > 4 + pavgb m0, [secq] + punpcklbw m1, m5 + punpckhbw m2, m0, m5 + punpcklbw m0, m5 +%else ; 4xh + movh m2, [secq] + pavgb m0, m2 + punpcklbw m1, m5 + punpcklbw m0, m5 + movhlps m2, m0 +%endif +%else ; !avg + movx m2, [srcq+src_strideq] + movx m1, [dstq] + pavgb m0, m4 + movx m4, [srcq+src_strideq+1] + movx m3, [dstq+dst_strideq] + pavgb m2, m4 + punpcklbw m0, m5 + punpcklbw m2, m5 + punpcklbw m3, m5 + punpcklbw m1, m5 +%endif + SUM_SSE m0, m1, m2, m3, m6, m7 + + lea srcq, [srcq+src_strideq*2] + lea dstq, [dstq+dst_strideq*2] +%endif +%if %2 == 1 ; avg + add secq, sec_str +%endif + dec block_height + jg .x_half_y_zero_loop + STORE_AND_RET %1 + +.x_half_y_nonzero: + cmp y_offsetd, 4 + jne .x_half_y_nonhalf + + ; x_offset == 0.5 && y_offset == 0.5 +%if %1 == 16 + movu m0, [srcq] + movu m3, [srcq+1] + add srcq, src_strideq + pavgb m0, m3 +.x_half_y_half_loop: + movu m4, [srcq] + movu m3, [srcq+1] + mova m1, [dstq] + pavgb m4, m3 + punpckhbw m3, m1, m5 + pavgb m0, m4 +%if %2 == 1 ; avg + punpcklbw m1, m5 + pavgb m0, [secq] + punpckhbw m2, m0, m5 + punpcklbw m0, m5 +%else + punpckhbw m2, m0, m5 + punpcklbw m0, m5 + punpcklbw m1, m5 +%endif + SUM_SSE m0, m1, m2, m3, m6, m7 + mova m0, m4 + + add srcq, src_strideq + add dstq, dst_strideq +%else ; %1 < 16 + movx m0, [srcq] + movx m3, [srcq+1] + add srcq, src_strideq + pavgb m0, m3 +.x_half_y_half_loop: + movx m2, [srcq] + movx m3, [srcq+1] +%if %2 == 1 ; avg +%if %1 > 4 + movhps m2, [srcq+src_strideq] + movhps m3, [srcq+src_strideq+1] +%else + movx m1, [srcq+src_strideq] + punpckldq m2, m1 + movx m1, [srcq+src_strideq+1] + punpckldq m3, m1 +%endif + pavgb m2, m3 +%if %1 > 4 + movlhps m0, m2 + movhlps m4, m2 +%else ; 4xh + punpckldq m0, m2 + pshuflw m4, m2, 0xe +%endif + movx m1, [dstq] + pavgb m0, m2 + movx m3, [dstq+dst_strideq] +%if %1 > 4 + pavgb m0, [secq] +%else + movh m2, [secq] + pavgb m0, m2 +%endif + punpcklbw m3, m5 + punpcklbw m1, m5 +%if %1 > 4 + punpckhbw m2, m0, m5 + punpcklbw m0, m5 +%else + punpcklbw m0, m5 + movhlps m2, m0 +%endif +%else ; !avg + movx m4, [srcq+src_strideq] + movx m1, [srcq+src_strideq+1] + pavgb m2, m3 + pavgb m4, m1 + pavgb m0, m2 + pavgb m2, m4 + movx m1, [dstq] + movx m3, [dstq+dst_strideq] + punpcklbw m0, m5 + punpcklbw m2, m5 + punpcklbw m3, m5 + punpcklbw m1, m5 +%endif + SUM_SSE m0, m1, m2, m3, m6, m7 + mova m0, m4 + + lea srcq, [srcq+src_strideq*2] + lea dstq, [dstq+dst_strideq*2] +%endif +%if %2 == 1 ; avg + add secq, sec_str +%endif + dec block_height + jg .x_half_y_half_loop + STORE_AND_RET %1 + +.x_half_y_nonhalf: + ; x_offset == 0.5 && y_offset == bilin interpolation +%if ARCH_X86_64 + lea bilin_filter, [GLOBAL(bilin_filter_m)] +%endif + shl y_offsetd, filter_idx_shift +%if ARCH_X86_64 && %1 > 4 + mova m8, [bilin_filter+y_offsetq] +%if notcpuflag(ssse3) ; FIXME(rbultje) don't scatter registers on x86-64 + mova m9, [bilin_filter+y_offsetq+16] +%endif + mova m10, [GLOBAL(pw_8)] +%define filter_y_a m8 +%define filter_y_b m9 +%define filter_rnd m10 +%else ;x86_32 +%if ARCH_X86=1 && CONFIG_PIC=1 +; x_offset == 0.5. We can reuse x_offset reg +%define tempq x_offsetq + add y_offsetq, g_bilin_filterm +%define filter_y_a [y_offsetq] +%define filter_y_b [y_offsetq+16] + mov tempq, g_pw_8m +%define filter_rnd [tempq] +%else + add y_offsetq, bilin_filter +%define filter_y_a [y_offsetq] +%define filter_y_b [y_offsetq+16] +%define filter_rnd [GLOBAL(pw_8)] +%endif +%endif + +%if %1 == 16 + movu m0, [srcq] + movu m3, [srcq+1] + add srcq, src_strideq + pavgb m0, m3 +.x_half_y_other_loop: + movu m4, [srcq] + movu m2, [srcq+1] + mova m1, [dstq] + pavgb m4, m2 +%if cpuflag(ssse3) + punpckhbw m2, m0, m4 + punpcklbw m0, m4 + pmaddubsw m2, filter_y_a + pmaddubsw m0, filter_y_a + paddw m2, filter_rnd + paddw m0, filter_rnd + psraw m2, 4 +%else + punpckhbw m2, m0, m5 + punpckhbw m3, m4, m5 + pmullw m2, filter_y_a + pmullw m3, filter_y_b + paddw m2, filter_rnd + punpcklbw m0, m5 + paddw m2, m3 + punpcklbw m3, m4, m5 + pmullw m0, filter_y_a + pmullw m3, filter_y_b + paddw m0, filter_rnd + psraw m2, 4 + paddw m0, m3 +%endif + punpckhbw m3, m1, m5 + psraw m0, 4 +%if %2 == 1 ; avg + ; FIXME(rbultje) pipeline + packuswb m0, m2 + pavgb m0, [secq] + punpckhbw m2, m0, m5 + punpcklbw m0, m5 +%endif + punpcklbw m1, m5 + SUM_SSE m0, m1, m2, m3, m6, m7 + mova m0, m4 + + add srcq, src_strideq + add dstq, dst_strideq +%else ; %1 < 16 + movx m0, [srcq] + movx m3, [srcq+1] + add srcq, src_strideq + pavgb m0, m3 +%if notcpuflag(ssse3) + punpcklbw m0, m5 +%endif +.x_half_y_other_loop: + movx m2, [srcq] + movx m1, [srcq+1] + movx m4, [srcq+src_strideq] + movx m3, [srcq+src_strideq+1] + pavgb m2, m1 + pavgb m4, m3 + movx m3, [dstq+dst_strideq] +%if cpuflag(ssse3) + movx m1, [dstq] + punpcklbw m0, m2 + punpcklbw m2, m4 + pmaddubsw m0, filter_y_a + pmaddubsw m2, filter_y_a + punpcklbw m3, m5 + paddw m0, filter_rnd + paddw m2, filter_rnd +%else + punpcklbw m2, m5 + punpcklbw m4, m5 + pmullw m0, filter_y_a + pmullw m1, m2, filter_y_b + punpcklbw m3, m5 + paddw m0, filter_rnd + pmullw m2, filter_y_a + paddw m0, m1 + pmullw m1, m4, filter_y_b + paddw m2, filter_rnd + paddw m2, m1 + movx m1, [dstq] +%endif + psraw m0, 4 + psraw m2, 4 +%if %2 == 1 ; avg + ; FIXME(rbultje) pipeline +%if %1 == 4 + movlhps m0, m2 +%endif + packuswb m0, m2 +%if %1 > 4 + pavgb m0, [secq] + punpckhbw m2, m0, m5 + punpcklbw m0, m5 +%else + movh m2, [secq] + pavgb m0, m2 + punpcklbw m0, m5 + movhlps m2, m0 +%endif +%endif + punpcklbw m1, m5 + SUM_SSE m0, m1, m2, m3, m6, m7 + mova m0, m4 + + lea srcq, [srcq+src_strideq*2] + lea dstq, [dstq+dst_strideq*2] +%endif +%if %2 == 1 ; avg + add secq, sec_str +%endif + dec block_height + jg .x_half_y_other_loop +%undef filter_y_a +%undef filter_y_b +%undef filter_rnd + STORE_AND_RET %1 + +.x_nonhalf: + test y_offsetd, y_offsetd + jnz .x_nonhalf_y_nonzero + + ; x_offset == bilin interpolation && y_offset == 0 +%if ARCH_X86_64 + lea bilin_filter, [GLOBAL(bilin_filter_m)] +%endif + shl x_offsetd, filter_idx_shift +%if ARCH_X86_64 && %1 > 4 + mova m8, [bilin_filter+x_offsetq] +%if notcpuflag(ssse3) ; FIXME(rbultje) don't scatter registers on x86-64 + mova m9, [bilin_filter+x_offsetq+16] +%endif + mova m10, [GLOBAL(pw_8)] +%define filter_x_a m8 +%define filter_x_b m9 +%define filter_rnd m10 +%else ; x86-32 +%if ARCH_X86=1 && CONFIG_PIC=1 +;y_offset == 0. We can reuse y_offset reg. +%define tempq y_offsetq + add x_offsetq, g_bilin_filterm +%define filter_x_a [x_offsetq] +%define filter_x_b [x_offsetq+16] + mov tempq, g_pw_8m +%define filter_rnd [tempq] +%else + add x_offsetq, bilin_filter +%define filter_x_a [x_offsetq] +%define filter_x_b [x_offsetq+16] +%define filter_rnd [GLOBAL(pw_8)] +%endif +%endif + +.x_other_y_zero_loop: +%if %1 == 16 + movu m0, [srcq] + movu m4, [srcq+1] + mova m1, [dstq] +%if cpuflag(ssse3) + punpckhbw m2, m0, m4 + punpcklbw m0, m4 + pmaddubsw m2, filter_x_a + pmaddubsw m0, filter_x_a + paddw m2, filter_rnd + paddw m0, filter_rnd +%else + punpckhbw m2, m0, m5 + punpckhbw m3, m4, m5 + punpcklbw m0, m5 + punpcklbw m4, m5 + pmullw m2, filter_x_a + pmullw m3, filter_x_b + paddw m2, filter_rnd + pmullw m0, filter_x_a + pmullw m4, filter_x_b + paddw m0, filter_rnd + paddw m2, m3 + paddw m0, m4 +%endif + psraw m2, 4 + psraw m0, 4 +%if %2 == 1 ; avg + ; FIXME(rbultje) pipeline + packuswb m0, m2 + pavgb m0, [secq] + punpckhbw m2, m0, m5 + punpcklbw m0, m5 +%endif + punpckhbw m3, m1, m5 + punpcklbw m1, m5 + SUM_SSE m0, m1, m2, m3, m6, m7 + + add srcq, src_strideq + add dstq, dst_strideq +%else ; %1 < 16 + movx m0, [srcq] + movx m1, [srcq+1] + movx m2, [srcq+src_strideq] + movx m4, [srcq+src_strideq+1] + movx m3, [dstq+dst_strideq] +%if cpuflag(ssse3) + punpcklbw m0, m1 + movx m1, [dstq] + punpcklbw m2, m4 + pmaddubsw m0, filter_x_a + pmaddubsw m2, filter_x_a + punpcklbw m3, m5 + paddw m0, filter_rnd + paddw m2, filter_rnd +%else + punpcklbw m0, m5 + punpcklbw m1, m5 + punpcklbw m2, m5 + punpcklbw m4, m5 + pmullw m0, filter_x_a + pmullw m1, filter_x_b + punpcklbw m3, m5 + paddw m0, filter_rnd + pmullw m2, filter_x_a + pmullw m4, filter_x_b + paddw m0, m1 + paddw m2, filter_rnd + movx m1, [dstq] + paddw m2, m4 +%endif + psraw m0, 4 + psraw m2, 4 +%if %2 == 1 ; avg + ; FIXME(rbultje) pipeline +%if %1 == 4 + movlhps m0, m2 +%endif + packuswb m0, m2 +%if %1 > 4 + pavgb m0, [secq] + punpckhbw m2, m0, m5 + punpcklbw m0, m5 +%else + movh m2, [secq] + pavgb m0, m2 + punpcklbw m0, m5 + movhlps m2, m0 +%endif +%endif + punpcklbw m1, m5 + SUM_SSE m0, m1, m2, m3, m6, m7 + + lea srcq, [srcq+src_strideq*2] + lea dstq, [dstq+dst_strideq*2] +%endif +%if %2 == 1 ; avg + add secq, sec_str +%endif + dec block_height + jg .x_other_y_zero_loop +%undef filter_x_a +%undef filter_x_b +%undef filter_rnd + STORE_AND_RET %1 + +.x_nonhalf_y_nonzero: + cmp y_offsetd, 4 + jne .x_nonhalf_y_nonhalf + + ; x_offset == bilin interpolation && y_offset == 0.5 +%if ARCH_X86_64 + lea bilin_filter, [GLOBAL(bilin_filter_m)] +%endif + shl x_offsetd, filter_idx_shift +%if ARCH_X86_64 && %1 > 4 + mova m8, [bilin_filter+x_offsetq] +%if notcpuflag(ssse3) ; FIXME(rbultje) don't scatter registers on x86-64 + mova m9, [bilin_filter+x_offsetq+16] +%endif + mova m10, [GLOBAL(pw_8)] +%define filter_x_a m8 +%define filter_x_b m9 +%define filter_rnd m10 +%else ; x86-32 +%if ARCH_X86=1 && CONFIG_PIC=1 +; y_offset == 0.5. We can reuse y_offset reg. +%define tempq y_offsetq + add x_offsetq, g_bilin_filterm +%define filter_x_a [x_offsetq] +%define filter_x_b [x_offsetq+16] + mov tempq, g_pw_8m +%define filter_rnd [tempq] +%else + add x_offsetq, bilin_filter +%define filter_x_a [x_offsetq] +%define filter_x_b [x_offsetq+16] +%define filter_rnd [GLOBAL(pw_8)] +%endif +%endif + +%if %1 == 16 + movu m0, [srcq] + movu m1, [srcq+1] +%if cpuflag(ssse3) + punpckhbw m2, m0, m1 + punpcklbw m0, m1 + pmaddubsw m2, filter_x_a + pmaddubsw m0, filter_x_a + paddw m2, filter_rnd + paddw m0, filter_rnd +%else + punpckhbw m2, m0, m5 + punpckhbw m3, m1, m5 + punpcklbw m0, m5 + punpcklbw m1, m5 + pmullw m0, filter_x_a + pmullw m1, filter_x_b + paddw m0, filter_rnd + pmullw m2, filter_x_a + pmullw m3, filter_x_b + paddw m2, filter_rnd + paddw m0, m1 + paddw m2, m3 +%endif + psraw m0, 4 + psraw m2, 4 + add srcq, src_strideq + packuswb m0, m2 +.x_other_y_half_loop: + movu m4, [srcq] + movu m3, [srcq+1] +%if cpuflag(ssse3) + mova m1, [dstq] + punpckhbw m2, m4, m3 + punpcklbw m4, m3 + pmaddubsw m2, filter_x_a + pmaddubsw m4, filter_x_a + paddw m2, filter_rnd + paddw m4, filter_rnd + psraw m2, 4 + psraw m4, 4 + packuswb m4, m2 + pavgb m0, m4 + punpckhbw m3, m1, m5 + punpcklbw m1, m5 +%else + punpckhbw m2, m4, m5 + punpckhbw m1, m3, m5 + punpcklbw m4, m5 + punpcklbw m3, m5 + pmullw m4, filter_x_a + pmullw m3, filter_x_b + paddw m4, filter_rnd + pmullw m2, filter_x_a + pmullw m1, filter_x_b + paddw m2, filter_rnd + paddw m4, m3 + paddw m2, m1 + mova m1, [dstq] + psraw m4, 4 + psraw m2, 4 + punpckhbw m3, m1, m5 + ; FIXME(rbultje) the repeated pack/unpack here around m0/m2 is because we + ; have a 1-register shortage to be able to store the backup of the bilin + ; filtered second line as words as cache for the next line. Packing into + ; a byte costs 1 pack and 2 unpacks, but saves a register. + packuswb m4, m2 + punpcklbw m1, m5 + pavgb m0, m4 +%endif +%if %2 == 1 ; avg + ; FIXME(rbultje) pipeline + pavgb m0, [secq] +%endif + punpckhbw m2, m0, m5 + punpcklbw m0, m5 + SUM_SSE m0, m1, m2, m3, m6, m7 + mova m0, m4 + + add srcq, src_strideq + add dstq, dst_strideq +%else ; %1 < 16 + movx m0, [srcq] + movx m1, [srcq+1] +%if cpuflag(ssse3) + punpcklbw m0, m1 + pmaddubsw m0, filter_x_a + paddw m0, filter_rnd +%else + punpcklbw m0, m5 + punpcklbw m1, m5 + pmullw m0, filter_x_a + pmullw m1, filter_x_b + paddw m0, filter_rnd + paddw m0, m1 +%endif + add srcq, src_strideq + psraw m0, 4 +.x_other_y_half_loop: + movx m2, [srcq] + movx m1, [srcq+1] + movx m4, [srcq+src_strideq] + movx m3, [srcq+src_strideq+1] +%if cpuflag(ssse3) + punpcklbw m2, m1 + punpcklbw m4, m3 + pmaddubsw m2, filter_x_a + pmaddubsw m4, filter_x_a + movx m1, [dstq] + movx m3, [dstq+dst_strideq] + paddw m2, filter_rnd + paddw m4, filter_rnd +%else + punpcklbw m2, m5 + punpcklbw m1, m5 + punpcklbw m4, m5 + punpcklbw m3, m5 + pmullw m2, filter_x_a + pmullw m1, filter_x_b + paddw m2, filter_rnd + pmullw m4, filter_x_a + pmullw m3, filter_x_b + paddw m4, filter_rnd + paddw m2, m1 + movx m1, [dstq] + paddw m4, m3 + movx m3, [dstq+dst_strideq] +%endif + psraw m2, 4 + psraw m4, 4 + pavgw m0, m2 + pavgw m2, m4 +%if %2 == 1 ; avg + ; FIXME(rbultje) pipeline - also consider going to bytes here +%if %1 == 4 + movlhps m0, m2 +%endif + packuswb m0, m2 +%if %1 > 4 + pavgb m0, [secq] + punpckhbw m2, m0, m5 + punpcklbw m0, m5 +%else + movh m2, [secq] + pavgb m0, m2 + punpcklbw m0, m5 + movhlps m2, m0 +%endif +%endif + punpcklbw m3, m5 + punpcklbw m1, m5 + SUM_SSE m0, m1, m2, m3, m6, m7 + mova m0, m4 + + lea srcq, [srcq+src_strideq*2] + lea dstq, [dstq+dst_strideq*2] +%endif +%if %2 == 1 ; avg + add secq, sec_str +%endif + dec block_height + jg .x_other_y_half_loop +%undef filter_x_a +%undef filter_x_b +%undef filter_rnd + STORE_AND_RET %1 + +.x_nonhalf_y_nonhalf: +%if ARCH_X86_64 + lea bilin_filter, [GLOBAL(bilin_filter_m)] +%endif + shl x_offsetd, filter_idx_shift + shl y_offsetd, filter_idx_shift +%if ARCH_X86_64 && %1 > 4 + mova m8, [bilin_filter+x_offsetq] +%if notcpuflag(ssse3) ; FIXME(rbultje) don't scatter registers on x86-64 + mova m9, [bilin_filter+x_offsetq+16] +%endif + mova m10, [bilin_filter+y_offsetq] +%if notcpuflag(ssse3) ; FIXME(rbultje) don't scatter registers on x86-64 + mova m11, [bilin_filter+y_offsetq+16] +%endif + mova m12, [GLOBAL(pw_8)] +%define filter_x_a m8 +%define filter_x_b m9 +%define filter_y_a m10 +%define filter_y_b m11 +%define filter_rnd m12 +%else ; x86-32 +%if ARCH_X86=1 && CONFIG_PIC=1 +; In this case, there is NO unused register. Used src_stride register. Later, +; src_stride has to be loaded from stack when it is needed. +%define tempq src_strideq + mov tempq, g_bilin_filterm + add x_offsetq, tempq + add y_offsetq, tempq +%define filter_x_a [x_offsetq] +%define filter_x_b [x_offsetq+16] +%define filter_y_a [y_offsetq] +%define filter_y_b [y_offsetq+16] + + mov tempq, g_pw_8m +%define filter_rnd [tempq] +%else + add x_offsetq, bilin_filter + add y_offsetq, bilin_filter +%define filter_x_a [x_offsetq] +%define filter_x_b [x_offsetq+16] +%define filter_y_a [y_offsetq] +%define filter_y_b [y_offsetq+16] +%define filter_rnd [GLOBAL(pw_8)] +%endif +%endif + + ; x_offset == bilin interpolation && y_offset == bilin interpolation +%if %1 == 16 + movu m0, [srcq] + movu m1, [srcq+1] +%if cpuflag(ssse3) + punpckhbw m2, m0, m1 + punpcklbw m0, m1 + pmaddubsw m2, filter_x_a + pmaddubsw m0, filter_x_a + paddw m2, filter_rnd + paddw m0, filter_rnd +%else + punpckhbw m2, m0, m5 + punpckhbw m3, m1, m5 + punpcklbw m0, m5 + punpcklbw m1, m5 + pmullw m0, filter_x_a + pmullw m1, filter_x_b + paddw m0, filter_rnd + pmullw m2, filter_x_a + pmullw m3, filter_x_b + paddw m2, filter_rnd + paddw m0, m1 + paddw m2, m3 +%endif + psraw m0, 4 + psraw m2, 4 + + INC_SRC_BY_SRC_STRIDE + + packuswb m0, m2 +.x_other_y_other_loop: +%if cpuflag(ssse3) + movu m4, [srcq] + movu m3, [srcq+1] + mova m1, [dstq] + punpckhbw m2, m4, m3 + punpcklbw m4, m3 + pmaddubsw m2, filter_x_a + pmaddubsw m4, filter_x_a + punpckhbw m3, m1, m5 + paddw m2, filter_rnd + paddw m4, filter_rnd + psraw m2, 4 + psraw m4, 4 + packuswb m4, m2 + punpckhbw m2, m0, m4 + punpcklbw m0, m4 + pmaddubsw m2, filter_y_a + pmaddubsw m0, filter_y_a + punpcklbw m1, m5 + paddw m2, filter_rnd + paddw m0, filter_rnd + psraw m2, 4 + psraw m0, 4 +%else + movu m3, [srcq] + movu m4, [srcq+1] + punpckhbw m1, m3, m5 + punpckhbw m2, m4, m5 + punpcklbw m3, m5 + punpcklbw m4, m5 + pmullw m3, filter_x_a + pmullw m4, filter_x_b + paddw m3, filter_rnd + pmullw m1, filter_x_a + pmullw m2, filter_x_b + paddw m1, filter_rnd + paddw m3, m4 + paddw m1, m2 + psraw m3, 4 + psraw m1, 4 + packuswb m4, m3, m1 + punpckhbw m2, m0, m5 + punpcklbw m0, m5 + pmullw m2, filter_y_a + pmullw m1, filter_y_b + paddw m2, filter_rnd + pmullw m0, filter_y_a + pmullw m3, filter_y_b + paddw m2, m1 + mova m1, [dstq] + paddw m0, filter_rnd + psraw m2, 4 + paddw m0, m3 + punpckhbw m3, m1, m5 + psraw m0, 4 + punpcklbw m1, m5 +%endif +%if %2 == 1 ; avg + ; FIXME(rbultje) pipeline + packuswb m0, m2 + pavgb m0, [secq] + punpckhbw m2, m0, m5 + punpcklbw m0, m5 +%endif + SUM_SSE m0, m1, m2, m3, m6, m7 + mova m0, m4 + + INC_SRC_BY_SRC_STRIDE + add dstq, dst_strideq +%else ; %1 < 16 + movx m0, [srcq] + movx m1, [srcq+1] +%if cpuflag(ssse3) + punpcklbw m0, m1 + pmaddubsw m0, filter_x_a + paddw m0, filter_rnd +%else + punpcklbw m0, m5 + punpcklbw m1, m5 + pmullw m0, filter_x_a + pmullw m1, filter_x_b + paddw m0, filter_rnd + paddw m0, m1 +%endif + psraw m0, 4 +%if cpuflag(ssse3) + packuswb m0, m0 +%endif + + INC_SRC_BY_SRC_STRIDE + +.x_other_y_other_loop: + movx m2, [srcq] + movx m1, [srcq+1] + + INC_SRC_BY_SRC_STRIDE + movx m4, [srcq] + movx m3, [srcq+1] + +%if cpuflag(ssse3) + punpcklbw m2, m1 + punpcklbw m4, m3 + pmaddubsw m2, filter_x_a + pmaddubsw m4, filter_x_a + movx m3, [dstq+dst_strideq] + movx m1, [dstq] + paddw m2, filter_rnd + paddw m4, filter_rnd + psraw m2, 4 + psraw m4, 4 + packuswb m2, m2 + packuswb m4, m4 + punpcklbw m0, m2 + punpcklbw m2, m4 + pmaddubsw m0, filter_y_a + pmaddubsw m2, filter_y_a + punpcklbw m3, m5 + paddw m0, filter_rnd + paddw m2, filter_rnd + psraw m0, 4 + psraw m2, 4 + punpcklbw m1, m5 +%else + punpcklbw m2, m5 + punpcklbw m1, m5 + punpcklbw m4, m5 + punpcklbw m3, m5 + pmullw m2, filter_x_a + pmullw m1, filter_x_b + paddw m2, filter_rnd + pmullw m4, filter_x_a + pmullw m3, filter_x_b + paddw m4, filter_rnd + paddw m2, m1 + paddw m4, m3 + psraw m2, 4 + psraw m4, 4 + pmullw m0, filter_y_a + pmullw m3, m2, filter_y_b + paddw m0, filter_rnd + pmullw m2, filter_y_a + pmullw m1, m4, filter_y_b + paddw m2, filter_rnd + paddw m0, m3 + movx m3, [dstq+dst_strideq] + paddw m2, m1 + movx m1, [dstq] + psraw m0, 4 + psraw m2, 4 + punpcklbw m3, m5 + punpcklbw m1, m5 +%endif +%if %2 == 1 ; avg + ; FIXME(rbultje) pipeline +%if %1 == 4 + movlhps m0, m2 +%endif + packuswb m0, m2 +%if %1 > 4 + pavgb m0, [secq] + punpckhbw m2, m0, m5 + punpcklbw m0, m5 +%else + movh m2, [secq] + pavgb m0, m2 + punpcklbw m0, m5 + movhlps m2, m0 +%endif +%endif + SUM_SSE m0, m1, m2, m3, m6, m7 + mova m0, m4 + + INC_SRC_BY_SRC_STRIDE + lea dstq, [dstq+dst_strideq*2] +%endif +%if %2 == 1 ; avg + add secq, sec_str +%endif + dec block_height + jg .x_other_y_other_loop +%undef filter_x_a +%undef filter_x_b +%undef filter_y_a +%undef filter_y_b +%undef filter_rnd +%undef movx + STORE_AND_RET %1 +%endmacro + +; FIXME(rbultje) the non-bilinear versions (i.e. x=0,8&&y=0,8) are identical +; between the ssse3 and non-ssse3 version. It may make sense to merge their +; code in the sense that the ssse3 version would jump to the appropriate +; location in the sse/2 version, rather than duplicating that code in the +; binary. + +INIT_XMM sse2 +SUBPEL_VARIANCE 4 +SUBPEL_VARIANCE 8 +SUBPEL_VARIANCE 16 + +INIT_XMM ssse3 +SUBPEL_VARIANCE 4 +SUBPEL_VARIANCE 8 +SUBPEL_VARIANCE 16 + +INIT_XMM sse2 +SUBPEL_VARIANCE 4, 1 +SUBPEL_VARIANCE 8, 1 +SUBPEL_VARIANCE 16, 1 + +INIT_XMM ssse3 +SUBPEL_VARIANCE 4, 1 +SUBPEL_VARIANCE 8, 1 +SUBPEL_VARIANCE 16, 1 diff --git a/third_party/aom/aom_dsp/x86/subtract_avx2.c b/third_party/aom/aom_dsp/x86/subtract_avx2.c new file mode 100644 index 0000000000..4389d123db --- /dev/null +++ b/third_party/aom/aom_dsp/x86/subtract_avx2.c @@ -0,0 +1,108 @@ +/* + * Copyright (c) 2018, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ +#include + +#include "config/aom_dsp_rtcd.h" + +static INLINE void subtract32_avx2(int16_t *diff_ptr, const uint8_t *src_ptr, + const uint8_t *pred_ptr) { + __m256i s = _mm256_lddqu_si256((__m256i *)(src_ptr)); + __m256i p = _mm256_lddqu_si256((__m256i *)(pred_ptr)); + __m256i s_0 = _mm256_cvtepu8_epi16(_mm256_castsi256_si128(s)); + __m256i s_1 = _mm256_cvtepu8_epi16(_mm256_extracti128_si256(s, 1)); + __m256i p_0 = _mm256_cvtepu8_epi16(_mm256_castsi256_si128(p)); + __m256i p_1 = _mm256_cvtepu8_epi16(_mm256_extracti128_si256(p, 1)); + const __m256i d_0 = _mm256_sub_epi16(s_0, p_0); + const __m256i d_1 = _mm256_sub_epi16(s_1, p_1); + _mm256_store_si256((__m256i *)(diff_ptr), d_0); + _mm256_store_si256((__m256i *)(diff_ptr + 16), d_1); +} + +static INLINE void aom_subtract_block_16xn_avx2( + int rows, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, + ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride) { + for (int32_t j = 0; j < rows; ++j) { + __m128i s = _mm_lddqu_si128((__m128i *)(src_ptr)); + __m128i p = _mm_lddqu_si128((__m128i *)(pred_ptr)); + __m256i s_0 = _mm256_cvtepu8_epi16(s); + __m256i p_0 = _mm256_cvtepu8_epi16(p); + const __m256i d_0 = _mm256_sub_epi16(s_0, p_0); + _mm256_store_si256((__m256i *)(diff_ptr), d_0); + src_ptr += src_stride; + pred_ptr += pred_stride; + diff_ptr += diff_stride; + } +} + +static INLINE void aom_subtract_block_32xn_avx2( + int rows, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, + ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride) { + for (int32_t j = 0; j < rows; ++j) { + subtract32_avx2(diff_ptr, src_ptr, pred_ptr); + src_ptr += src_stride; + pred_ptr += pred_stride; + diff_ptr += diff_stride; + } +} + +static INLINE void aom_subtract_block_64xn_avx2( + int rows, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, + ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride) { + for (int32_t j = 0; j < rows; ++j) { + subtract32_avx2(diff_ptr, src_ptr, pred_ptr); + subtract32_avx2(diff_ptr + 32, src_ptr + 32, pred_ptr + 32); + src_ptr += src_stride; + pred_ptr += pred_stride; + diff_ptr += diff_stride; + } +} + +static INLINE void aom_subtract_block_128xn_avx2( + int rows, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, + ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride) { + for (int32_t j = 0; j < rows; ++j) { + subtract32_avx2(diff_ptr, src_ptr, pred_ptr); + subtract32_avx2(diff_ptr + 32, src_ptr + 32, pred_ptr + 32); + subtract32_avx2(diff_ptr + 64, src_ptr + 64, pred_ptr + 64); + subtract32_avx2(diff_ptr + 96, src_ptr + 96, pred_ptr + 96); + src_ptr += src_stride; + pred_ptr += pred_stride; + diff_ptr += diff_stride; + } +} + +void aom_subtract_block_avx2(int rows, int cols, int16_t *diff_ptr, + ptrdiff_t diff_stride, const uint8_t *src_ptr, + ptrdiff_t src_stride, const uint8_t *pred_ptr, + ptrdiff_t pred_stride) { + switch (cols) { + case 16: + aom_subtract_block_16xn_avx2(rows, diff_ptr, diff_stride, src_ptr, + src_stride, pred_ptr, pred_stride); + break; + case 32: + aom_subtract_block_32xn_avx2(rows, diff_ptr, diff_stride, src_ptr, + src_stride, pred_ptr, pred_stride); + break; + case 64: + aom_subtract_block_64xn_avx2(rows, diff_ptr, diff_stride, src_ptr, + src_stride, pred_ptr, pred_stride); + break; + case 128: + aom_subtract_block_128xn_avx2(rows, diff_ptr, diff_stride, src_ptr, + src_stride, pred_ptr, pred_stride); + break; + default: + aom_subtract_block_sse2(rows, cols, diff_ptr, diff_stride, src_ptr, + src_stride, pred_ptr, pred_stride); + break; + } +} diff --git a/third_party/aom/aom_dsp/x86/subtract_sse2.asm b/third_party/aom/aom_dsp/x86/subtract_sse2.asm new file mode 100644 index 0000000000..1a75a234f9 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/subtract_sse2.asm @@ -0,0 +1,146 @@ +; +; Copyright (c) 2016, Alliance for Open Media. All rights reserved +; +; This source code is subject to the terms of the BSD 2 Clause License and +; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License +; was not distributed with this source code in the LICENSE file, you can +; obtain it at www.aomedia.org/license/software. If the Alliance for Open +; Media Patent License 1.0 was not distributed with this source code in the +; PATENTS file, you can obtain it at www.aomedia.org/license/patent. +; + +; + +%include "third_party/x86inc/x86inc.asm" + +SECTION .text + +; void aom_subtract_block(int rows, int cols, +; int16_t *diff, ptrdiff_t diff_stride, +; const uint8_t *src, ptrdiff_t src_stride, +; const uint8_t *pred, ptrdiff_t pred_stride) + +INIT_XMM sse2 +cglobal subtract_block, 7, 7, 8, \ + rows, cols, diff, diff_stride, src, src_stride, \ + pred, pred_stride +%define pred_str colsq + pxor m7, m7 ; dedicated zero register + cmp colsd, 4 + je .case_4 + cmp colsd, 8 + je .case_8 + cmp colsd, 16 + je .case_16 + cmp colsd, 32 + je .case_32 + cmp colsd, 64 + je .case_64 + +%macro loop16 6 + mova m0, [srcq+%1] + mova m4, [srcq+%2] + mova m1, [predq+%3] + mova m5, [predq+%4] + punpckhbw m2, m0, m7 + punpckhbw m3, m1, m7 + punpcklbw m0, m7 + punpcklbw m1, m7 + psubw m2, m3 + psubw m0, m1 + punpckhbw m1, m4, m7 + punpckhbw m3, m5, m7 + punpcklbw m4, m7 + punpcklbw m5, m7 + psubw m1, m3 + psubw m4, m5 + mova [diffq+mmsize*0+%5], m0 + mova [diffq+mmsize*1+%5], m2 + mova [diffq+mmsize*0+%6], m4 + mova [diffq+mmsize*1+%6], m1 +%endmacro + + mov pred_str, pred_stridemp +.loop_128: + loop16 0*mmsize, 1*mmsize, 0*mmsize, 1*mmsize, 0*mmsize, 2*mmsize + loop16 2*mmsize, 3*mmsize, 2*mmsize, 3*mmsize, 4*mmsize, 6*mmsize + loop16 4*mmsize, 5*mmsize, 4*mmsize, 5*mmsize, 8*mmsize, 10*mmsize + loop16 6*mmsize, 7*mmsize, 6*mmsize, 7*mmsize, 12*mmsize, 14*mmsize + lea diffq, [diffq+diff_strideq*2] + add predq, pred_str + add srcq, src_strideq + sub rowsd, 1 + jnz .loop_128 + RET + +.case_64: + mov pred_str, pred_stridemp +.loop_64: + loop16 0*mmsize, 1*mmsize, 0*mmsize, 1*mmsize, 0*mmsize, 2*mmsize + loop16 2*mmsize, 3*mmsize, 2*mmsize, 3*mmsize, 4*mmsize, 6*mmsize + lea diffq, [diffq+diff_strideq*2] + add predq, pred_str + add srcq, src_strideq + dec rowsd + jg .loop_64 + RET + +.case_32: + mov pred_str, pred_stridemp +.loop_32: + loop16 0, mmsize, 0, mmsize, 0, 2*mmsize + lea diffq, [diffq+diff_strideq*2] + add predq, pred_str + add srcq, src_strideq + dec rowsd + jg .loop_32 + RET + +.case_16: + mov pred_str, pred_stridemp +.loop_16: + loop16 0, src_strideq, 0, pred_str, 0, diff_strideq*2 + lea diffq, [diffq+diff_strideq*4] + lea predq, [predq+pred_str*2] + lea srcq, [srcq+src_strideq*2] + sub rowsd, 2 + jg .loop_16 + RET + +%macro loop_h 0 + movh m0, [srcq] + movh m2, [srcq+src_strideq] + movh m1, [predq] + movh m3, [predq+pred_str] + punpcklbw m0, m7 + punpcklbw m1, m7 + punpcklbw m2, m7 + punpcklbw m3, m7 + psubw m0, m1 + psubw m2, m3 + mova [diffq], m0 + mova [diffq+diff_strideq*2], m2 +%endmacro + +.case_8: + mov pred_str, pred_stridemp +.loop_8: + loop_h + lea diffq, [diffq+diff_strideq*4] + lea srcq, [srcq+src_strideq*2] + lea predq, [predq+pred_str*2] + sub rowsd, 2 + jg .loop_8 + RET + +INIT_MMX +.case_4: + mov pred_str, pred_stridemp +.loop_4: + loop_h + lea diffq, [diffq+diff_strideq*4] + lea srcq, [srcq+src_strideq*2] + lea predq, [predq+pred_str*2] + sub rowsd, 2 + jg .loop_4 + RET diff --git a/third_party/aom/aom_dsp/x86/sum_squares_avx2.c b/third_party/aom/aom_dsp/x86/sum_squares_avx2.c new file mode 100644 index 0000000000..0af44e3a4c --- /dev/null +++ b/third_party/aom/aom_dsp/x86/sum_squares_avx2.c @@ -0,0 +1,79 @@ +/* + * Copyright (c) 2018, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include +#include + +#include "aom_dsp/x86/synonyms.h" +#include "aom_dsp/x86/synonyms_avx2.h" +#include "aom_dsp/x86/sum_squares_sse2.h" +#include "config/aom_dsp_rtcd.h" + +static uint64_t aom_sum_squares_2d_i16_nxn_avx2(const int16_t *src, int stride, + int width, int height) { + uint64_t result; + __m256i v_acc_q = _mm256_setzero_si256(); + const __m256i v_zext_mask_q = yy_set1_64_from_32i(0xffffffff); + for (int col = 0; col < height; col += 4) { + __m256i v_acc_d = _mm256_setzero_si256(); + for (int row = 0; row < width; row += 16) { + const int16_t *tempsrc = src + row; + const __m256i v_val_0_w = + _mm256_loadu_si256((const __m256i *)(tempsrc + 0 * stride)); + const __m256i v_val_1_w = + _mm256_loadu_si256((const __m256i *)(tempsrc + 1 * stride)); + const __m256i v_val_2_w = + _mm256_loadu_si256((const __m256i *)(tempsrc + 2 * stride)); + const __m256i v_val_3_w = + _mm256_loadu_si256((const __m256i *)(tempsrc + 3 * stride)); + + const __m256i v_sq_0_d = _mm256_madd_epi16(v_val_0_w, v_val_0_w); + const __m256i v_sq_1_d = _mm256_madd_epi16(v_val_1_w, v_val_1_w); + const __m256i v_sq_2_d = _mm256_madd_epi16(v_val_2_w, v_val_2_w); + const __m256i v_sq_3_d = _mm256_madd_epi16(v_val_3_w, v_val_3_w); + + const __m256i v_sum_01_d = _mm256_add_epi32(v_sq_0_d, v_sq_1_d); + const __m256i v_sum_23_d = _mm256_add_epi32(v_sq_2_d, v_sq_3_d); + const __m256i v_sum_0123_d = _mm256_add_epi32(v_sum_01_d, v_sum_23_d); + + v_acc_d = _mm256_add_epi32(v_acc_d, v_sum_0123_d); + } + v_acc_q = + _mm256_add_epi64(v_acc_q, _mm256_and_si256(v_acc_d, v_zext_mask_q)); + v_acc_q = _mm256_add_epi64(v_acc_q, _mm256_srli_epi64(v_acc_d, 32)); + src += 4 * stride; + } + __m128i lower_64_2_Value = _mm256_castsi256_si128(v_acc_q); + __m128i higher_64_2_Value = _mm256_extracti128_si256(v_acc_q, 1); + __m128i result_64_2_int = _mm_add_epi64(lower_64_2_Value, higher_64_2_Value); + + result_64_2_int = _mm_add_epi64( + result_64_2_int, _mm_unpackhi_epi64(result_64_2_int, result_64_2_int)); + + xx_storel_64(&result, result_64_2_int); + + return result; +} + +uint64_t aom_sum_squares_2d_i16_avx2(const int16_t *src, int stride, int width, + int height) { + if (LIKELY(width == 4 && height == 4)) { + return aom_sum_squares_2d_i16_4x4_sse2(src, stride); + } else if (LIKELY(width == 4 && (height & 3) == 0)) { + return aom_sum_squares_2d_i16_4xn_sse2(src, stride, height); + } else if (LIKELY(width == 8 && (height & 3) == 0)) { + return aom_sum_squares_2d_i16_nxn_sse2(src, stride, width, height); + } else if (LIKELY(((width & 15) == 0) && ((height & 3) == 0))) { + return aom_sum_squares_2d_i16_nxn_avx2(src, stride, width, height); + } else { + return aom_sum_squares_2d_i16_c(src, stride, width, height); + } +} diff --git a/third_party/aom/aom_dsp/x86/sum_squares_sse2.c b/third_party/aom/aom_dsp/x86/sum_squares_sse2.c new file mode 100644 index 0000000000..22d7739ec4 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/sum_squares_sse2.c @@ -0,0 +1,203 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include +#include +#include + +#include "aom_dsp/x86/synonyms.h" +#include "aom_dsp/x86/sum_squares_sse2.h" +#include "config/aom_dsp_rtcd.h" + +static INLINE __m128i xx_loadh_64(__m128i a, const void *b) { + const __m128d ad = _mm_castsi128_pd(a); + return _mm_castpd_si128(_mm_loadh_pd(ad, (double *)b)); +} + +static INLINE uint64_t xx_cvtsi128_si64(__m128i a) { +#if ARCH_X86_64 + return (uint64_t)_mm_cvtsi128_si64(a); +#else + { + uint64_t tmp; + _mm_storel_epi64((__m128i *)&tmp, a); + return tmp; + } +#endif +} + +static INLINE __m128i sum_squares_i16_4x4_sse2(const int16_t *src, int stride) { + const __m128i v_val_0_w = xx_loadl_64(src + 0 * stride); + const __m128i v_val_2_w = xx_loadl_64(src + 2 * stride); + const __m128i v_val_01_w = xx_loadh_64(v_val_0_w, src + 1 * stride); + const __m128i v_val_23_w = xx_loadh_64(v_val_2_w, src + 3 * stride); + const __m128i v_sq_01_d = _mm_madd_epi16(v_val_01_w, v_val_01_w); + const __m128i v_sq_23_d = _mm_madd_epi16(v_val_23_w, v_val_23_w); + + return _mm_add_epi32(v_sq_01_d, v_sq_23_d); +} + +uint64_t aom_sum_squares_2d_i16_4x4_sse2(const int16_t *src, int stride) { + const __m128i v_sum_0123_d = sum_squares_i16_4x4_sse2(src, stride); + __m128i v_sum_d = + _mm_add_epi32(v_sum_0123_d, _mm_srli_epi64(v_sum_0123_d, 32)); + v_sum_d = _mm_add_epi32(v_sum_d, _mm_srli_si128(v_sum_d, 8)); + return (uint64_t)_mm_cvtsi128_si32(v_sum_d); +} + +uint64_t aom_sum_squares_2d_i16_4xn_sse2(const int16_t *src, int stride, + int height) { + int r = 0; + __m128i v_acc_q = _mm_setzero_si128(); + do { + const __m128i v_acc_d = sum_squares_i16_4x4_sse2(src, stride); + v_acc_q = _mm_add_epi32(v_acc_q, v_acc_d); + src += stride << 2; + r += 4; + } while (r < height); + const __m128i v_zext_mask_q = xx_set1_64_from_32i(0xffffffff); + __m128i v_acc_64 = _mm_add_epi64(_mm_srli_epi64(v_acc_q, 32), + _mm_and_si128(v_acc_q, v_zext_mask_q)); + v_acc_64 = _mm_add_epi64(v_acc_64, _mm_srli_si128(v_acc_64, 8)); + return xx_cvtsi128_si64(v_acc_64); +} + +#ifdef __GNUC__ +// This prevents GCC/Clang from inlining this function into +// aom_sum_squares_2d_i16_sse2, which in turn saves some stack +// maintenance instructions in the common case of 4x4. +__attribute__((noinline)) +#endif +uint64_t +aom_sum_squares_2d_i16_nxn_sse2(const int16_t *src, int stride, int width, + int height) { + int r = 0; + + const __m128i v_zext_mask_q = xx_set1_64_from_32i(0xffffffff); + __m128i v_acc_q = _mm_setzero_si128(); + + do { + __m128i v_acc_d = _mm_setzero_si128(); + int c = 0; + do { + const int16_t *b = src + c; + + const __m128i v_val_0_w = xx_load_128(b + 0 * stride); + const __m128i v_val_1_w = xx_load_128(b + 1 * stride); + const __m128i v_val_2_w = xx_load_128(b + 2 * stride); + const __m128i v_val_3_w = xx_load_128(b + 3 * stride); + + const __m128i v_sq_0_d = _mm_madd_epi16(v_val_0_w, v_val_0_w); + const __m128i v_sq_1_d = _mm_madd_epi16(v_val_1_w, v_val_1_w); + const __m128i v_sq_2_d = _mm_madd_epi16(v_val_2_w, v_val_2_w); + const __m128i v_sq_3_d = _mm_madd_epi16(v_val_3_w, v_val_3_w); + + const __m128i v_sum_01_d = _mm_add_epi32(v_sq_0_d, v_sq_1_d); + const __m128i v_sum_23_d = _mm_add_epi32(v_sq_2_d, v_sq_3_d); + + const __m128i v_sum_0123_d = _mm_add_epi32(v_sum_01_d, v_sum_23_d); + + v_acc_d = _mm_add_epi32(v_acc_d, v_sum_0123_d); + c += 8; + } while (c < width); + + v_acc_q = _mm_add_epi64(v_acc_q, _mm_and_si128(v_acc_d, v_zext_mask_q)); + v_acc_q = _mm_add_epi64(v_acc_q, _mm_srli_epi64(v_acc_d, 32)); + + src += 4 * stride; + r += 4; + } while (r < height); + + v_acc_q = _mm_add_epi64(v_acc_q, _mm_srli_si128(v_acc_q, 8)); + return xx_cvtsi128_si64(v_acc_q); +} + +uint64_t aom_sum_squares_2d_i16_sse2(const int16_t *src, int stride, int width, + int height) { + // 4 elements per row only requires half an XMM register, so this + // must be a special case, but also note that over 75% of all calls + // are with size == 4, so it is also the common case. + if (LIKELY(width == 4 && height == 4)) { + return aom_sum_squares_2d_i16_4x4_sse2(src, stride); + } else if (LIKELY(width == 4 && (height & 3) == 0)) { + return aom_sum_squares_2d_i16_4xn_sse2(src, stride, height); + } else if (LIKELY((width & 7) == 0 && (height & 3) == 0)) { + // Generic case + return aom_sum_squares_2d_i16_nxn_sse2(src, stride, width, height); + } else { + return aom_sum_squares_2d_i16_c(src, stride, width, height); + } +} + +////////////////////////////////////////////////////////////////////////////// +// 1D version +////////////////////////////////////////////////////////////////////////////// + +static uint64_t aom_sum_squares_i16_64n_sse2(const int16_t *src, uint32_t n) { + const __m128i v_zext_mask_q = xx_set1_64_from_32i(0xffffffff); + __m128i v_acc0_q = _mm_setzero_si128(); + __m128i v_acc1_q = _mm_setzero_si128(); + + const int16_t *const end = src + n; + + assert(n % 64 == 0); + + while (src < end) { + const __m128i v_val_0_w = xx_load_128(src); + const __m128i v_val_1_w = xx_load_128(src + 8); + const __m128i v_val_2_w = xx_load_128(src + 16); + const __m128i v_val_3_w = xx_load_128(src + 24); + const __m128i v_val_4_w = xx_load_128(src + 32); + const __m128i v_val_5_w = xx_load_128(src + 40); + const __m128i v_val_6_w = xx_load_128(src + 48); + const __m128i v_val_7_w = xx_load_128(src + 56); + + const __m128i v_sq_0_d = _mm_madd_epi16(v_val_0_w, v_val_0_w); + const __m128i v_sq_1_d = _mm_madd_epi16(v_val_1_w, v_val_1_w); + const __m128i v_sq_2_d = _mm_madd_epi16(v_val_2_w, v_val_2_w); + const __m128i v_sq_3_d = _mm_madd_epi16(v_val_3_w, v_val_3_w); + const __m128i v_sq_4_d = _mm_madd_epi16(v_val_4_w, v_val_4_w); + const __m128i v_sq_5_d = _mm_madd_epi16(v_val_5_w, v_val_5_w); + const __m128i v_sq_6_d = _mm_madd_epi16(v_val_6_w, v_val_6_w); + const __m128i v_sq_7_d = _mm_madd_epi16(v_val_7_w, v_val_7_w); + + const __m128i v_sum_01_d = _mm_add_epi32(v_sq_0_d, v_sq_1_d); + const __m128i v_sum_23_d = _mm_add_epi32(v_sq_2_d, v_sq_3_d); + const __m128i v_sum_45_d = _mm_add_epi32(v_sq_4_d, v_sq_5_d); + const __m128i v_sum_67_d = _mm_add_epi32(v_sq_6_d, v_sq_7_d); + + const __m128i v_sum_0123_d = _mm_add_epi32(v_sum_01_d, v_sum_23_d); + const __m128i v_sum_4567_d = _mm_add_epi32(v_sum_45_d, v_sum_67_d); + + const __m128i v_sum_d = _mm_add_epi32(v_sum_0123_d, v_sum_4567_d); + + v_acc0_q = _mm_add_epi64(v_acc0_q, _mm_and_si128(v_sum_d, v_zext_mask_q)); + v_acc1_q = _mm_add_epi64(v_acc1_q, _mm_srli_epi64(v_sum_d, 32)); + + src += 64; + } + + v_acc0_q = _mm_add_epi64(v_acc0_q, v_acc1_q); + v_acc0_q = _mm_add_epi64(v_acc0_q, _mm_srli_si128(v_acc0_q, 8)); + return xx_cvtsi128_si64(v_acc0_q); +} + +uint64_t aom_sum_squares_i16_sse2(const int16_t *src, uint32_t n) { + if (n % 64 == 0) { + return aom_sum_squares_i16_64n_sse2(src, n); + } else if (n > 64) { + int k = n & ~(64 - 1); + return aom_sum_squares_i16_64n_sse2(src, k) + + aom_sum_squares_i16_c(src + k, n - k); + } else { + return aom_sum_squares_i16_c(src, n); + } +} diff --git a/third_party/aom/aom_dsp/x86/sum_squares_sse2.h b/third_party/aom/aom_dsp/x86/sum_squares_sse2.h new file mode 100644 index 0000000000..491e31cc5e --- /dev/null +++ b/third_party/aom/aom_dsp/x86/sum_squares_sse2.h @@ -0,0 +1,22 @@ +/* + * Copyright (c) 2018, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_DSP_X86_SUM_SQUARES_SSE2_H_ +#define AOM_DSP_X86_SUM_SQUARES_SSE2_H_ + +uint64_t aom_sum_squares_2d_i16_nxn_sse2(const int16_t *src, int stride, + int width, int height); + +uint64_t aom_sum_squares_2d_i16_4xn_sse2(const int16_t *src, int stride, + int height); +uint64_t aom_sum_squares_2d_i16_4x4_sse2(const int16_t *src, int stride); + +#endif // AOM_DSP_X86_SUM_SQUARES_SSE2_H_ diff --git a/third_party/aom/aom_dsp/x86/synonyms.h b/third_party/aom/aom_dsp/x86/synonyms.h new file mode 100644 index 0000000000..1e9f1e27b8 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/synonyms.h @@ -0,0 +1,114 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_X86_SYNONYMS_H_ +#define AOM_AOM_DSP_X86_SYNONYMS_H_ + +#include + +#include "config/aom_config.h" + +#include "aom/aom_integer.h" + +/** + * Various reusable shorthands for x86 SIMD intrinsics. + * + * Intrinsics prefixed with xx_ operate on or return 128bit XMM registers. + * Intrinsics prefixed with yy_ operate on or return 256bit YMM registers. + */ + +// Loads and stores to do away with the tedium of casting the address +// to the right type. +static INLINE __m128i xx_loadl_32(const void *a) { + return _mm_cvtsi32_si128(*(const uint32_t *)a); +} + +static INLINE __m128i xx_loadl_64(const void *a) { + return _mm_loadl_epi64((const __m128i *)a); +} + +static INLINE __m128i xx_load_128(const void *a) { + return _mm_load_si128((const __m128i *)a); +} + +static INLINE __m128i xx_loadu_128(const void *a) { + return _mm_loadu_si128((const __m128i *)a); +} + +static INLINE void xx_storel_32(void *const a, const __m128i v) { + *(uint32_t *)a = _mm_cvtsi128_si32(v); +} + +static INLINE void xx_storel_64(void *const a, const __m128i v) { + _mm_storel_epi64((__m128i *)a, v); +} + +static INLINE void xx_store_128(void *const a, const __m128i v) { + _mm_store_si128((__m128i *)a, v); +} + +static INLINE void xx_storeu_128(void *const a, const __m128i v) { + _mm_storeu_si128((__m128i *)a, v); +} + +// The _mm_set_epi64x() intrinsic is undefined for some Visual Studio +// compilers. The following function is equivalent to _mm_set_epi64x() +// acting on 32-bit integers. +static INLINE __m128i xx_set_64_from_32i(int32_t e1, int32_t e0) { +#if defined(_MSC_VER) && _MSC_VER < 1900 + return _mm_set_epi32(0, e1, 0, e0); +#else + return _mm_set_epi64x((uint32_t)e1, (uint32_t)e0); +#endif +} + +// The _mm_set1_epi64x() intrinsic is undefined for some Visual Studio +// compilers. The following function is equivalent to _mm_set1_epi64x() +// acting on a 32-bit integer. +static INLINE __m128i xx_set1_64_from_32i(int32_t a) { +#if defined(_MSC_VER) && _MSC_VER < 1900 + return _mm_set_epi32(0, a, 0, a); +#else + return _mm_set1_epi64x((uint32_t)a); +#endif +} + +static INLINE __m128i xx_round_epu16(__m128i v_val_w) { + return _mm_avg_epu16(v_val_w, _mm_setzero_si128()); +} + +static INLINE __m128i xx_roundn_epu16(__m128i v_val_w, int bits) { + const __m128i v_s_w = _mm_srli_epi16(v_val_w, bits - 1); + return _mm_avg_epu16(v_s_w, _mm_setzero_si128()); +} + +static INLINE __m128i xx_roundn_epu32(__m128i v_val_d, int bits) { + const __m128i v_bias_d = _mm_set1_epi32((1 << bits) >> 1); + const __m128i v_tmp_d = _mm_add_epi32(v_val_d, v_bias_d); + return _mm_srli_epi32(v_tmp_d, bits); +} + +// This is equivalent to ROUND_POWER_OF_TWO(v_val_d, bits) +static INLINE __m128i xx_roundn_epi32_unsigned(__m128i v_val_d, int bits) { + const __m128i v_bias_d = _mm_set1_epi32((1 << bits) >> 1); + const __m128i v_tmp_d = _mm_add_epi32(v_val_d, v_bias_d); + return _mm_srai_epi32(v_tmp_d, bits); +} + +static INLINE __m128i xx_roundn_epi16(__m128i v_val_d, int bits) { + const __m128i v_bias_d = _mm_set1_epi16((1 << bits) >> 1); + const __m128i v_sign_d = _mm_srai_epi16(v_val_d, 15); + const __m128i v_tmp_d = + _mm_add_epi16(_mm_add_epi16(v_val_d, v_bias_d), v_sign_d); + return _mm_srai_epi16(v_tmp_d, bits); +} + +#endif // AOM_AOM_DSP_X86_SYNONYMS_H_ diff --git a/third_party/aom/aom_dsp/x86/synonyms_avx2.h b/third_party/aom/aom_dsp/x86/synonyms_avx2.h new file mode 100644 index 0000000000..3f69b120ea --- /dev/null +++ b/third_party/aom/aom_dsp/x86/synonyms_avx2.h @@ -0,0 +1,74 @@ +/* + * Copyright (c) 2018, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_X86_SYNONYMS_AVX2_H_ +#define AOM_AOM_DSP_X86_SYNONYMS_AVX2_H_ + +#include + +#include "config/aom_config.h" + +#include "aom/aom_integer.h" + +/** + * Various reusable shorthands for x86 SIMD intrinsics. + * + * Intrinsics prefixed with xx_ operate on or return 128bit XMM registers. + * Intrinsics prefixed with yy_ operate on or return 256bit YMM registers. + */ + +// Loads and stores to do away with the tedium of casting the address +// to the right type. +static INLINE __m256i yy_load_256(const void *a) { + return _mm256_load_si256((const __m256i *)a); +} + +static INLINE __m256i yy_loadu_256(const void *a) { + return _mm256_loadu_si256((const __m256i *)a); +} + +static INLINE void yy_store_256(void *const a, const __m256i v) { + _mm256_store_si256((__m256i *)a, v); +} + +static INLINE void yy_storeu_256(void *const a, const __m256i v) { + _mm256_storeu_si256((__m256i *)a, v); +} + +// The _mm256_set1_epi64x() intrinsic is undefined for some Visual Studio +// compilers. The following function is equivalent to _mm256_set1_epi64x() +// acting on a 32-bit integer. +static INLINE __m256i yy_set1_64_from_32i(int32_t a) { +#if defined(_MSC_VER) && defined(_M_IX86) && _MSC_VER < 1900 + return _mm256_set_epi32(0, a, 0, a, 0, a, 0, a); +#else + return _mm256_set1_epi64x((uint32_t)a); +#endif +} + +// Some compilers don't have _mm256_set_m128i defined in immintrin.h. We +// therefore define an equivalent function using a different intrinsic. +// ([ hi ], [ lo ]) -> [ hi ][ lo ] +static INLINE __m256i yy_set_m128i(__m128i hi, __m128i lo) { + return _mm256_insertf128_si256(_mm256_castsi128_si256(lo), hi, 1); +} + +static INLINE __m256i yy_loadu2_128(const void *hi, const void *lo) { + __m128i mhi = _mm_loadu_si128((__m128i *)(hi)); + __m128i mlo = _mm_loadu_si128((__m128i *)(lo)); + return yy_set_m128i(mhi, mlo); +} + +static INLINE __m256i yy_roundn_epu16(__m256i v_val_w, int bits) { + const __m256i v_s_w = _mm256_srli_epi16(v_val_w, bits - 1); + return _mm256_avg_epu16(v_s_w, _mm256_setzero_si256()); +} +#endif // AOM_AOM_DSP_X86_SYNONYMS_AVX2_H_ diff --git a/third_party/aom/aom_dsp/x86/transpose_sse2.h b/third_party/aom/aom_dsp/x86/transpose_sse2.h new file mode 100644 index 0000000000..d0d1ee6845 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/transpose_sse2.h @@ -0,0 +1,420 @@ +/* + * Copyright (c) 2018, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_X86_TRANSPOSE_SSE2_H_ +#define AOM_AOM_DSP_X86_TRANSPOSE_SSE2_H_ + +#include // SSE2 + +#include "config/aom_config.h" + +static INLINE __m128i transpose_8bit_4x4(const __m128i *const in) { + // Unpack 16 bit elements. Goes from: + // in[0]: 00 01 02 03 + // in[1]: 10 11 12 13 + // in[2]: 20 21 22 23 + // in[3]: 30 31 32 33 + // to: + // a0: 00 10 01 11 02 12 03 13 + // a1: 20 30 21 31 22 32 23 33 + const __m128i a0 = _mm_unpacklo_epi8(in[0], in[1]); + const __m128i a1 = _mm_unpacklo_epi8(in[2], in[3]); + + // Unpack 32 bit elements resulting in: + // 00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33 + return _mm_unpacklo_epi16(a0, a1); +} + +static INLINE void transpose_8bit_8x8(const __m128i *const in, + __m128i *const out) { + // Unpack 8 bit elements. Goes from: + // in[0]: 00 01 02 03 04 05 06 07 + // in[1]: 10 11 12 13 14 15 16 17 + // in[2]: 20 21 22 23 24 25 26 27 + // in[3]: 30 31 32 33 34 35 36 37 + // in[4]: 40 41 42 43 44 45 46 47 + // in[5]: 50 51 52 53 54 55 56 57 + // in[6]: 60 61 62 63 64 65 66 67 + // in[7]: 70 71 72 73 74 75 76 77 + // to: + // a0: 00 10 01 11 02 12 03 13 04 14 05 15 06 16 07 17 + // a1: 20 30 21 31 22 32 23 33 24 34 25 35 26 36 27 37 + // a2: 40 50 41 51 42 52 43 53 44 54 45 55 46 56 47 57 + // a3: 60 70 61 71 62 72 63 73 64 74 65 75 66 76 67 77 + const __m128i a0 = _mm_unpacklo_epi8(in[0], in[1]); + const __m128i a1 = _mm_unpacklo_epi8(in[2], in[3]); + const __m128i a2 = _mm_unpacklo_epi8(in[4], in[5]); + const __m128i a3 = _mm_unpacklo_epi8(in[6], in[7]); + + // Unpack 16 bit elements resulting in: + // b0: 00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33 + // b1: 40 50 60 70 41 51 61 71 42 52 62 72 43 53 63 73 + // b2: 04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37 + // b3: 44 54 64 74 45 55 65 75 46 56 66 76 47 57 67 77 + const __m128i b0 = _mm_unpacklo_epi16(a0, a1); + const __m128i b1 = _mm_unpackhi_epi16(a0, a1); + const __m128i b2 = _mm_unpacklo_epi16(a2, a3); + const __m128i b3 = _mm_unpackhi_epi16(a2, a3); + + // Unpack 32 bit elements resulting in: + // c0: 00 10 20 30 40 50 60 70 01 11 21 31 41 51 61 71 + // c1: 02 12 22 32 42 52 62 72 03 13 23 33 43 53 63 73 + // c2: 04 14 24 34 44 54 64 74 05 15 25 35 45 55 65 75 + // c3: 06 16 26 36 46 56 66 76 07 17 27 37 47 57 67 77 + const __m128i c0 = _mm_unpacklo_epi32(b0, b2); + const __m128i c1 = _mm_unpackhi_epi32(b0, b2); + const __m128i c2 = _mm_unpacklo_epi32(b1, b3); + const __m128i c3 = _mm_unpackhi_epi32(b1, b3); + + // Unpack 64 bit elements resulting in: + // out[0]: 00 10 20 30 40 50 60 70 + // out[1]: 01 11 21 31 41 51 61 71 + // out[2]: 02 12 22 32 42 52 62 72 + // out[3]: 03 13 23 33 43 53 63 73 + // out[4]: 04 14 24 34 44 54 64 74 + // out[5]: 05 15 25 35 45 55 65 75 + // out[6]: 06 16 26 36 46 56 66 76 + // out[7]: 07 17 27 37 47 57 67 77 + out[0] = _mm_unpacklo_epi64(c0, c0); + out[1] = _mm_unpackhi_epi64(c0, c0); + out[2] = _mm_unpacklo_epi64(c1, c1); + out[3] = _mm_unpackhi_epi64(c1, c1); + out[4] = _mm_unpacklo_epi64(c2, c2); + out[5] = _mm_unpackhi_epi64(c2, c2); + out[6] = _mm_unpacklo_epi64(c3, c3); + out[7] = _mm_unpackhi_epi64(c3, c3); +} + +static INLINE void transpose_16bit_4x4(const __m128i *const in, + __m128i *const out) { + // Unpack 16 bit elements. Goes from: + // in[0]: 00 01 02 03 XX XX XX XX + // in[1]: 10 11 12 13 XX XX XX XX + // in[2]: 20 21 22 23 XX XX XX XX + // in[3]: 30 31 32 33 XX XX XX XX + // to: + // a0: 00 10 01 11 02 12 03 13 + // a1: 20 30 21 31 22 32 23 33 + const __m128i a0 = _mm_unpacklo_epi16(in[0], in[1]); + const __m128i a1 = _mm_unpacklo_epi16(in[2], in[3]); + + // Unpack 32 bit elements resulting in: + // out[0]: 00 10 20 30 + // out[1]: 01 11 21 31 + // out[2]: 02 12 22 32 + // out[3]: 03 13 23 33 + out[0] = _mm_unpacklo_epi32(a0, a1); + out[1] = _mm_srli_si128(out[0], 8); + out[2] = _mm_unpackhi_epi32(a0, a1); + out[3] = _mm_srli_si128(out[2], 8); +} + +static INLINE void transpose_16bit_4x8(const __m128i *const in, + __m128i *const out) { + // Unpack 16 bit elements. Goes from: + // in[0]: 00 01 02 03 XX XX XX XX + // in[1]: 10 11 12 13 XX XX XX XX + // in[2]: 20 21 22 23 XX XX XX XX + // in[3]: 30 31 32 33 XX XX XX XX + // in[4]: 40 41 42 43 XX XX XX XX + // in[5]: 50 51 52 53 XX XX XX XX + // in[6]: 60 61 62 63 XX XX XX XX + // in[7]: 70 71 72 73 XX XX XX XX + // to: + // a0: 00 10 01 11 02 12 03 13 + // a1: 20 30 21 31 22 32 23 33 + // a2: 40 50 41 51 42 52 43 53 + // a3: 60 70 61 71 62 72 63 73 + const __m128i a0 = _mm_unpacklo_epi16(in[0], in[1]); + const __m128i a1 = _mm_unpacklo_epi16(in[2], in[3]); + const __m128i a2 = _mm_unpacklo_epi16(in[4], in[5]); + const __m128i a3 = _mm_unpacklo_epi16(in[6], in[7]); + + // Unpack 32 bit elements resulting in: + // b0: 00 10 20 30 01 11 21 31 + // b1: 40 50 60 70 41 51 61 71 + // b2: 02 12 22 32 03 13 23 33 + // b3: 42 52 62 72 43 53 63 73 + const __m128i b0 = _mm_unpacklo_epi32(a0, a1); + const __m128i b1 = _mm_unpacklo_epi32(a2, a3); + const __m128i b2 = _mm_unpackhi_epi32(a0, a1); + const __m128i b3 = _mm_unpackhi_epi32(a2, a3); + + // Unpack 64 bit elements resulting in: + // out[0]: 00 10 20 30 40 50 60 70 + // out[1]: 01 11 21 31 41 51 61 71 + // out[2]: 02 12 22 32 42 52 62 72 + // out[3]: 03 13 23 33 43 53 63 73 + out[0] = _mm_unpacklo_epi64(b0, b1); + out[1] = _mm_unpackhi_epi64(b0, b1); + out[2] = _mm_unpacklo_epi64(b2, b3); + out[3] = _mm_unpackhi_epi64(b2, b3); +} + +static INLINE void transpose_16bit_8x4(const __m128i *const in, + __m128i *const out) { + // Unpack 16 bit elements. Goes from: + // in[0]: 00 01 02 03 04 05 06 07 + // in[1]: 10 11 12 13 14 15 16 17 + // in[2]: 20 21 22 23 24 25 26 27 + // in[3]: 30 31 32 33 34 35 36 37 + + // to: + // a0: 00 10 01 11 02 12 03 13 + // a1: 20 30 21 31 22 32 23 33 + // a4: 04 14 05 15 06 16 07 17 + // a5: 24 34 25 35 26 36 27 37 + const __m128i a0 = _mm_unpacklo_epi16(in[0], in[1]); + const __m128i a1 = _mm_unpacklo_epi16(in[2], in[3]); + const __m128i a4 = _mm_unpackhi_epi16(in[0], in[1]); + const __m128i a5 = _mm_unpackhi_epi16(in[2], in[3]); + + // Unpack 32 bit elements resulting in: + // b0: 00 10 20 30 01 11 21 31 + // b2: 04 14 24 34 05 15 25 35 + // b4: 02 12 22 32 03 13 23 33 + // b6: 06 16 26 36 07 17 27 37 + const __m128i b0 = _mm_unpacklo_epi32(a0, a1); + const __m128i b2 = _mm_unpacklo_epi32(a4, a5); + const __m128i b4 = _mm_unpackhi_epi32(a0, a1); + const __m128i b6 = _mm_unpackhi_epi32(a4, a5); + + // Unpack 64 bit elements resulting in: + // out[0]: 00 10 20 30 XX XX XX XX + // out[1]: 01 11 21 31 XX XX XX XX + // out[2]: 02 12 22 32 XX XX XX XX + // out[3]: 03 13 23 33 XX XX XX XX + // out[4]: 04 14 24 34 XX XX XX XX + // out[5]: 05 15 25 35 XX XX XX XX + // out[6]: 06 16 26 36 XX XX XX XX + // out[7]: 07 17 27 37 XX XX XX XX + const __m128i zeros = _mm_setzero_si128(); + out[0] = _mm_unpacklo_epi64(b0, zeros); + out[1] = _mm_unpackhi_epi64(b0, zeros); + out[2] = _mm_unpacklo_epi64(b4, zeros); + out[3] = _mm_unpackhi_epi64(b4, zeros); + out[4] = _mm_unpacklo_epi64(b2, zeros); + out[5] = _mm_unpackhi_epi64(b2, zeros); + out[6] = _mm_unpacklo_epi64(b6, zeros); + out[7] = _mm_unpackhi_epi64(b6, zeros); +} + +static INLINE void transpose_16bit_8x8(const __m128i *const in, + __m128i *const out) { + // Unpack 16 bit elements. Goes from: + // in[0]: 00 01 02 03 04 05 06 07 + // in[1]: 10 11 12 13 14 15 16 17 + // in[2]: 20 21 22 23 24 25 26 27 + // in[3]: 30 31 32 33 34 35 36 37 + // in[4]: 40 41 42 43 44 45 46 47 + // in[5]: 50 51 52 53 54 55 56 57 + // in[6]: 60 61 62 63 64 65 66 67 + // in[7]: 70 71 72 73 74 75 76 77 + // to: + // a0: 00 10 01 11 02 12 03 13 + // a1: 20 30 21 31 22 32 23 33 + // a2: 40 50 41 51 42 52 43 53 + // a3: 60 70 61 71 62 72 63 73 + // a4: 04 14 05 15 06 16 07 17 + // a5: 24 34 25 35 26 36 27 37 + // a6: 44 54 45 55 46 56 47 57 + // a7: 64 74 65 75 66 76 67 77 + const __m128i a0 = _mm_unpacklo_epi16(in[0], in[1]); + const __m128i a1 = _mm_unpacklo_epi16(in[2], in[3]); + const __m128i a2 = _mm_unpacklo_epi16(in[4], in[5]); + const __m128i a3 = _mm_unpacklo_epi16(in[6], in[7]); + const __m128i a4 = _mm_unpackhi_epi16(in[0], in[1]); + const __m128i a5 = _mm_unpackhi_epi16(in[2], in[3]); + const __m128i a6 = _mm_unpackhi_epi16(in[4], in[5]); + const __m128i a7 = _mm_unpackhi_epi16(in[6], in[7]); + + // Unpack 32 bit elements resulting in: + // b0: 00 10 20 30 01 11 21 31 + // b1: 40 50 60 70 41 51 61 71 + // b2: 04 14 24 34 05 15 25 35 + // b3: 44 54 64 74 45 55 65 75 + // b4: 02 12 22 32 03 13 23 33 + // b5: 42 52 62 72 43 53 63 73 + // b6: 06 16 26 36 07 17 27 37 + // b7: 46 56 66 76 47 57 67 77 + const __m128i b0 = _mm_unpacklo_epi32(a0, a1); + const __m128i b1 = _mm_unpacklo_epi32(a2, a3); + const __m128i b2 = _mm_unpacklo_epi32(a4, a5); + const __m128i b3 = _mm_unpacklo_epi32(a6, a7); + const __m128i b4 = _mm_unpackhi_epi32(a0, a1); + const __m128i b5 = _mm_unpackhi_epi32(a2, a3); + const __m128i b6 = _mm_unpackhi_epi32(a4, a5); + const __m128i b7 = _mm_unpackhi_epi32(a6, a7); + + // Unpack 64 bit elements resulting in: + // out[0]: 00 10 20 30 40 50 60 70 + // out[1]: 01 11 21 31 41 51 61 71 + // out[2]: 02 12 22 32 42 52 62 72 + // out[3]: 03 13 23 33 43 53 63 73 + // out[4]: 04 14 24 34 44 54 64 74 + // out[5]: 05 15 25 35 45 55 65 75 + // out[6]: 06 16 26 36 46 56 66 76 + // out[7]: 07 17 27 37 47 57 67 77 + out[0] = _mm_unpacklo_epi64(b0, b1); + out[1] = _mm_unpackhi_epi64(b0, b1); + out[2] = _mm_unpacklo_epi64(b4, b5); + out[3] = _mm_unpackhi_epi64(b4, b5); + out[4] = _mm_unpacklo_epi64(b2, b3); + out[5] = _mm_unpackhi_epi64(b2, b3); + out[6] = _mm_unpacklo_epi64(b6, b7); + out[7] = _mm_unpackhi_epi64(b6, b7); +} + +// Transpose in-place +static INLINE void transpose_16bit_16x16(__m128i *const left, + __m128i *const right) { + __m128i tbuf[8]; + transpose_16bit_8x8(left, left); + transpose_16bit_8x8(right, tbuf); + transpose_16bit_8x8(left + 8, right); + transpose_16bit_8x8(right + 8, right + 8); + + left[8] = tbuf[0]; + left[9] = tbuf[1]; + left[10] = tbuf[2]; + left[11] = tbuf[3]; + left[12] = tbuf[4]; + left[13] = tbuf[5]; + left[14] = tbuf[6]; + left[15] = tbuf[7]; +} + +static INLINE void transpose_32bit_4x4(const __m128i *const in, + __m128i *const out) { + // Unpack 32 bit elements. Goes from: + // in[0]: 00 01 02 03 + // in[1]: 10 11 12 13 + // in[2]: 20 21 22 23 + // in[3]: 30 31 32 33 + // to: + // a0: 00 10 01 11 + // a1: 20 30 21 31 + // a2: 02 12 03 13 + // a3: 22 32 23 33 + + const __m128i a0 = _mm_unpacklo_epi32(in[0], in[1]); + const __m128i a1 = _mm_unpacklo_epi32(in[2], in[3]); + const __m128i a2 = _mm_unpackhi_epi32(in[0], in[1]); + const __m128i a3 = _mm_unpackhi_epi32(in[2], in[3]); + + // Unpack 64 bit elements resulting in: + // out[0]: 00 10 20 30 + // out[1]: 01 11 21 31 + // out[2]: 02 12 22 32 + // out[3]: 03 13 23 33 + out[0] = _mm_unpacklo_epi64(a0, a1); + out[1] = _mm_unpackhi_epi64(a0, a1); + out[2] = _mm_unpacklo_epi64(a2, a3); + out[3] = _mm_unpackhi_epi64(a2, a3); +} + +static INLINE void transpose_32bit_4x4x2(const __m128i *const in, + __m128i *const out) { + // Unpack 32 bit elements. Goes from: + // in[0]: 00 01 02 03 + // in[1]: 10 11 12 13 + // in[2]: 20 21 22 23 + // in[3]: 30 31 32 33 + // in[4]: 04 05 06 07 + // in[5]: 14 15 16 17 + // in[6]: 24 25 26 27 + // in[7]: 34 35 36 37 + // to: + // a0: 00 10 01 11 + // a1: 20 30 21 31 + // a2: 02 12 03 13 + // a3: 22 32 23 33 + // a4: 04 14 05 15 + // a5: 24 34 25 35 + // a6: 06 16 07 17 + // a7: 26 36 27 37 + const __m128i a0 = _mm_unpacklo_epi32(in[0], in[1]); + const __m128i a1 = _mm_unpacklo_epi32(in[2], in[3]); + const __m128i a2 = _mm_unpackhi_epi32(in[0], in[1]); + const __m128i a3 = _mm_unpackhi_epi32(in[2], in[3]); + const __m128i a4 = _mm_unpacklo_epi32(in[4], in[5]); + const __m128i a5 = _mm_unpacklo_epi32(in[6], in[7]); + const __m128i a6 = _mm_unpackhi_epi32(in[4], in[5]); + const __m128i a7 = _mm_unpackhi_epi32(in[6], in[7]); + + // Unpack 64 bit elements resulting in: + // out[0]: 00 10 20 30 + // out[1]: 01 11 21 31 + // out[2]: 02 12 22 32 + // out[3]: 03 13 23 33 + // out[4]: 04 14 24 34 + // out[5]: 05 15 25 35 + // out[6]: 06 16 26 36 + // out[7]: 07 17 27 37 + out[0] = _mm_unpacklo_epi64(a0, a1); + out[1] = _mm_unpackhi_epi64(a0, a1); + out[2] = _mm_unpacklo_epi64(a2, a3); + out[3] = _mm_unpackhi_epi64(a2, a3); + out[4] = _mm_unpacklo_epi64(a4, a5); + out[5] = _mm_unpackhi_epi64(a4, a5); + out[6] = _mm_unpacklo_epi64(a6, a7); + out[7] = _mm_unpackhi_epi64(a6, a7); +} + +static INLINE void transpose_32bit_8x4(const __m128i *const in, + __m128i *const out) { + // Unpack 32 bit elements. Goes from: + // in[0]: 00 01 02 03 + // in[1]: 04 05 06 07 + // in[2]: 10 11 12 13 + // in[3]: 14 15 16 17 + // in[4]: 20 21 22 23 + // in[5]: 24 25 26 27 + // in[6]: 30 31 32 33 + // in[7]: 34 35 36 37 + // to: + // a0: 00 10 01 11 + // a1: 20 30 21 31 + // a2: 02 12 03 13 + // a3: 22 32 23 33 + // a4: 04 14 05 15 + // a5: 24 34 25 35 + // a6: 06 16 07 17 + // a7: 26 36 27 37 + const __m128i a0 = _mm_unpacklo_epi32(in[0], in[2]); + const __m128i a1 = _mm_unpacklo_epi32(in[4], in[6]); + const __m128i a2 = _mm_unpackhi_epi32(in[0], in[2]); + const __m128i a3 = _mm_unpackhi_epi32(in[4], in[6]); + const __m128i a4 = _mm_unpacklo_epi32(in[1], in[3]); + const __m128i a5 = _mm_unpacklo_epi32(in[5], in[7]); + const __m128i a6 = _mm_unpackhi_epi32(in[1], in[3]); + const __m128i a7 = _mm_unpackhi_epi32(in[5], in[7]); + + // Unpack 64 bit elements resulting in: + // out[0]: 00 10 20 30 + // out[1]: 01 11 21 31 + // out[2]: 02 12 22 32 + // out[3]: 03 13 23 33 + // out[4]: 04 14 24 34 + // out[5]: 05 15 25 35 + // out[6]: 06 16 26 36 + // out[7]: 07 17 27 37 + out[0] = _mm_unpacklo_epi64(a0, a1); + out[1] = _mm_unpackhi_epi64(a0, a1); + out[2] = _mm_unpacklo_epi64(a2, a3); + out[3] = _mm_unpackhi_epi64(a2, a3); + out[4] = _mm_unpacklo_epi64(a4, a5); + out[5] = _mm_unpackhi_epi64(a4, a5); + out[6] = _mm_unpacklo_epi64(a6, a7); + out[7] = _mm_unpackhi_epi64(a6, a7); +} + +#endif // AOM_AOM_DSP_X86_TRANSPOSE_SSE2_H_ diff --git a/third_party/aom/aom_dsp/x86/txfm_common_avx2.h b/third_party/aom/aom_dsp/x86/txfm_common_avx2.h new file mode 100644 index 0000000000..b1611ba870 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/txfm_common_avx2.h @@ -0,0 +1,199 @@ +/* + * Copyright (c) 2018, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_X86_TXFM_COMMON_AVX2_H_ +#define AOM_AOM_DSP_X86_TXFM_COMMON_AVX2_H_ + +#include +#include "aom/aom_integer.h" +#include "aom_dsp/x86/synonyms.h" + +#ifdef __cplusplus +extern "C" { +#endif + +typedef void (*transform_1d_avx2)(const __m256i *input, __m256i *output, + int8_t cos_bit); + +static INLINE __m256i pair_set_w16_epi16(int16_t a, int16_t b) { + return _mm256_set1_epi32( + (int32_t)(((uint16_t)(a)) | (((uint32_t)(b)) << 16))); +} + +static INLINE void btf_16_w16_avx2(const __m256i w0, const __m256i w1, + __m256i *in0, __m256i *in1, const __m256i _r, + const int32_t cos_bit) { + __m256i t0 = _mm256_unpacklo_epi16(*in0, *in1); + __m256i t1 = _mm256_unpackhi_epi16(*in0, *in1); + __m256i u0 = _mm256_madd_epi16(t0, w0); + __m256i u1 = _mm256_madd_epi16(t1, w0); + __m256i v0 = _mm256_madd_epi16(t0, w1); + __m256i v1 = _mm256_madd_epi16(t1, w1); + + __m256i a0 = _mm256_add_epi32(u0, _r); + __m256i a1 = _mm256_add_epi32(u1, _r); + __m256i b0 = _mm256_add_epi32(v0, _r); + __m256i b1 = _mm256_add_epi32(v1, _r); + + __m256i c0 = _mm256_srai_epi32(a0, cos_bit); + __m256i c1 = _mm256_srai_epi32(a1, cos_bit); + __m256i d0 = _mm256_srai_epi32(b0, cos_bit); + __m256i d1 = _mm256_srai_epi32(b1, cos_bit); + + *in0 = _mm256_packs_epi32(c0, c1); + *in1 = _mm256_packs_epi32(d0, d1); +} + +static INLINE void btf_16_adds_subs_avx2(__m256i *in0, __m256i *in1) { + const __m256i _in0 = *in0; + const __m256i _in1 = *in1; + *in0 = _mm256_adds_epi16(_in0, _in1); + *in1 = _mm256_subs_epi16(_in0, _in1); +} + +static INLINE void btf_32_add_sub_avx2(__m256i *in0, __m256i *in1) { + const __m256i _in0 = *in0; + const __m256i _in1 = *in1; + *in0 = _mm256_add_epi32(_in0, _in1); + *in1 = _mm256_sub_epi32(_in0, _in1); +} + +static INLINE void btf_16_adds_subs_out_avx2(__m256i *out0, __m256i *out1, + __m256i in0, __m256i in1) { + const __m256i _in0 = in0; + const __m256i _in1 = in1; + *out0 = _mm256_adds_epi16(_in0, _in1); + *out1 = _mm256_subs_epi16(_in0, _in1); +} + +static INLINE void btf_32_add_sub_out_avx2(__m256i *out0, __m256i *out1, + __m256i in0, __m256i in1) { + const __m256i _in0 = in0; + const __m256i _in1 = in1; + *out0 = _mm256_add_epi32(_in0, _in1); + *out1 = _mm256_sub_epi32(_in0, _in1); +} + +static INLINE __m256i load_16bit_to_16bit_avx2(const int16_t *a) { + return _mm256_load_si256((const __m256i *)a); +} + +static INLINE void load_buffer_16bit_to_16bit_avx2(const int16_t *in, + int stride, __m256i *out, + int out_size) { + for (int i = 0; i < out_size; ++i) { + out[i] = load_16bit_to_16bit_avx2(in + i * stride); + } +} + +static INLINE void load_buffer_16bit_to_16bit_flip_avx2(const int16_t *in, + int stride, + __m256i *out, + int out_size) { + for (int i = 0; i < out_size; ++i) { + out[out_size - i - 1] = load_16bit_to_16bit_avx2(in + i * stride); + } +} + +static INLINE __m256i load_32bit_to_16bit_w16_avx2(const int32_t *a) { + const __m256i a_low = _mm256_lddqu_si256((const __m256i *)a); + const __m256i b = _mm256_packs_epi32(a_low, *(const __m256i *)(a + 8)); + return _mm256_permute4x64_epi64(b, 0xD8); +} + +static INLINE void load_buffer_32bit_to_16bit_w16_avx2(const int32_t *in, + int stride, __m256i *out, + int out_size) { + for (int i = 0; i < out_size; ++i) { + out[i] = load_32bit_to_16bit_w16_avx2(in + i * stride); + } +} + +static INLINE void transpose_16bit_16x16_avx2(const __m256i *const in, + __m256i *const out) { + // Unpack 16 bit elements. Goes from: + // in[0]: 00 01 02 03 08 09 0a 0b 04 05 06 07 0c 0d 0e 0f + // in[1]: 10 11 12 13 18 19 1a 1b 14 15 16 17 1c 1d 1e 1f + // in[2]: 20 21 22 23 28 29 2a 2b 24 25 26 27 2c 2d 2e 2f + // in[3]: 30 31 32 33 38 39 3a 3b 34 35 36 37 3c 3d 3e 3f + // in[4]: 40 41 42 43 48 49 4a 4b 44 45 46 47 4c 4d 4e 4f + // in[5]: 50 51 52 53 58 59 5a 5b 54 55 56 57 5c 5d 5e 5f + // in[6]: 60 61 62 63 68 69 6a 6b 64 65 66 67 6c 6d 6e 6f + // in[7]: 70 71 72 73 78 79 7a 7b 74 75 76 77 7c 7d 7e 7f + // in[8]: 80 81 82 83 88 89 8a 8b 84 85 86 87 8c 8d 8e 8f + // to: + // a0: 00 10 01 11 02 12 03 13 04 14 05 15 06 16 07 17 + // a1: 20 30 21 31 22 32 23 33 24 34 25 35 26 36 27 37 + // a2: 40 50 41 51 42 52 43 53 44 54 45 55 46 56 47 57 + // a3: 60 70 61 71 62 72 63 73 64 74 65 75 66 76 67 77 + // ... + __m256i a[16]; + for (int i = 0; i < 16; i += 2) { + a[i / 2 + 0] = _mm256_unpacklo_epi16(in[i], in[i + 1]); + a[i / 2 + 8] = _mm256_unpackhi_epi16(in[i], in[i + 1]); + } + __m256i b[16]; + for (int i = 0; i < 16; i += 2) { + b[i / 2 + 0] = _mm256_unpacklo_epi32(a[i], a[i + 1]); + b[i / 2 + 8] = _mm256_unpackhi_epi32(a[i], a[i + 1]); + } + __m256i c[16]; + for (int i = 0; i < 16; i += 2) { + c[i / 2 + 0] = _mm256_unpacklo_epi64(b[i], b[i + 1]); + c[i / 2 + 8] = _mm256_unpackhi_epi64(b[i], b[i + 1]); + } + out[0 + 0] = _mm256_permute2x128_si256(c[0], c[1], 0x20); + out[1 + 0] = _mm256_permute2x128_si256(c[8], c[9], 0x20); + out[2 + 0] = _mm256_permute2x128_si256(c[4], c[5], 0x20); + out[3 + 0] = _mm256_permute2x128_si256(c[12], c[13], 0x20); + + out[0 + 8] = _mm256_permute2x128_si256(c[0], c[1], 0x31); + out[1 + 8] = _mm256_permute2x128_si256(c[8], c[9], 0x31); + out[2 + 8] = _mm256_permute2x128_si256(c[4], c[5], 0x31); + out[3 + 8] = _mm256_permute2x128_si256(c[12], c[13], 0x31); + + out[4 + 0] = _mm256_permute2x128_si256(c[0 + 2], c[1 + 2], 0x20); + out[5 + 0] = _mm256_permute2x128_si256(c[8 + 2], c[9 + 2], 0x20); + out[6 + 0] = _mm256_permute2x128_si256(c[4 + 2], c[5 + 2], 0x20); + out[7 + 0] = _mm256_permute2x128_si256(c[12 + 2], c[13 + 2], 0x20); + + out[4 + 8] = _mm256_permute2x128_si256(c[0 + 2], c[1 + 2], 0x31); + out[5 + 8] = _mm256_permute2x128_si256(c[8 + 2], c[9 + 2], 0x31); + out[6 + 8] = _mm256_permute2x128_si256(c[4 + 2], c[5 + 2], 0x31); + out[7 + 8] = _mm256_permute2x128_si256(c[12 + 2], c[13 + 2], 0x31); +} + +static INLINE void flip_buf_avx2(__m256i *in, __m256i *out, int size) { + for (int i = 0; i < size; ++i) { + out[size - i - 1] = in[i]; + } +} + +static INLINE void round_shift_16bit_w16_avx2(__m256i *in, int size, int bit) { + if (bit < 0) { + bit = -bit; + __m256i round = _mm256_set1_epi16(1 << (bit - 1)); + for (int i = 0; i < size; ++i) { + in[i] = _mm256_adds_epi16(in[i], round); + in[i] = _mm256_srai_epi16(in[i], bit); + } + } else if (bit > 0) { + for (int i = 0; i < size; ++i) { + in[i] = _mm256_slli_epi16(in[i], bit); + } + } +} + +#ifdef __cplusplus +} +#endif + +#endif // AOM_AOM_DSP_X86_TXFM_COMMON_AVX2_H_ diff --git a/third_party/aom/aom_dsp/x86/txfm_common_sse2.h b/third_party/aom/aom_dsp/x86/txfm_common_sse2.h new file mode 100644 index 0000000000..ed82eee962 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/txfm_common_sse2.h @@ -0,0 +1,29 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AOM_DSP_X86_TXFM_COMMON_SSE2_H_ +#define AOM_AOM_DSP_X86_TXFM_COMMON_SSE2_H_ + +#include +#include "aom/aom_integer.h" +#include "aom_dsp/x86/synonyms.h" + +#define pair_set_epi16(a, b) \ + _mm_set1_epi32((int32_t)(((uint16_t)(a)) | (((uint32_t)(b)) << 16))) + +// Reverse the 8 16 bit words in __m128i +static INLINE __m128i mm_reverse_epi16(const __m128i x) { + const __m128i a = _mm_shufflelo_epi16(x, 0x1b); + const __m128i b = _mm_shufflehi_epi16(a, 0x1b); + return _mm_shuffle_epi32(b, 0x4e); +} + +#endif // AOM_AOM_DSP_X86_TXFM_COMMON_SSE2_H_ diff --git a/third_party/aom/aom_dsp/x86/variance_avx2.c b/third_party/aom/aom_dsp/x86/variance_avx2.c new file mode 100644 index 0000000000..800aef1266 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/variance_avx2.c @@ -0,0 +1,517 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include + +#include "config/aom_dsp_rtcd.h" + +#include "aom_dsp/x86/masked_variance_intrin_ssse3.h" + +static INLINE __m128i mm256_add_hi_lo_epi16(const __m256i val) { + return _mm_add_epi16(_mm256_castsi256_si128(val), + _mm256_extractf128_si256(val, 1)); +} + +static INLINE __m128i mm256_add_hi_lo_epi32(const __m256i val) { + return _mm_add_epi32(_mm256_castsi256_si128(val), + _mm256_extractf128_si256(val, 1)); +} + +static INLINE void variance_kernel_avx2(const __m256i src, const __m256i ref, + __m256i *const sse, + __m256i *const sum) { + const __m256i adj_sub = _mm256_set1_epi16(0xff01); // (1,-1) + + // unpack into pairs of source and reference values + const __m256i src_ref0 = _mm256_unpacklo_epi8(src, ref); + const __m256i src_ref1 = _mm256_unpackhi_epi8(src, ref); + + // subtract adjacent elements using src*1 + ref*-1 + const __m256i diff0 = _mm256_maddubs_epi16(src_ref0, adj_sub); + const __m256i diff1 = _mm256_maddubs_epi16(src_ref1, adj_sub); + const __m256i madd0 = _mm256_madd_epi16(diff0, diff0); + const __m256i madd1 = _mm256_madd_epi16(diff1, diff1); + + // add to the running totals + *sum = _mm256_add_epi16(*sum, _mm256_add_epi16(diff0, diff1)); + *sse = _mm256_add_epi32(*sse, _mm256_add_epi32(madd0, madd1)); +} + +static INLINE int variance_final_from_32bit_sum_avx2(__m256i vsse, __m128i vsum, + unsigned int *const sse) { + // extract the low lane and add it to the high lane + const __m128i sse_reg_128 = mm256_add_hi_lo_epi32(vsse); + + // unpack sse and sum registers and add + const __m128i sse_sum_lo = _mm_unpacklo_epi32(sse_reg_128, vsum); + const __m128i sse_sum_hi = _mm_unpackhi_epi32(sse_reg_128, vsum); + const __m128i sse_sum = _mm_add_epi32(sse_sum_lo, sse_sum_hi); + + // perform the final summation and extract the results + const __m128i res = _mm_add_epi32(sse_sum, _mm_srli_si128(sse_sum, 8)); + *((int *)sse) = _mm_cvtsi128_si32(res); + return _mm_extract_epi32(res, 1); +} + +// handle pixels (<= 512) +static INLINE int variance_final_512_avx2(__m256i vsse, __m256i vsum, + unsigned int *const sse) { + // extract the low lane and add it to the high lane + const __m128i vsum_128 = mm256_add_hi_lo_epi16(vsum); + const __m128i vsum_64 = _mm_add_epi16(vsum_128, _mm_srli_si128(vsum_128, 8)); + const __m128i sum_int32 = _mm_cvtepi16_epi32(vsum_64); + return variance_final_from_32bit_sum_avx2(vsse, sum_int32, sse); +} + +// handle 1024 pixels (32x32, 16x64, 64x16) +static INLINE int variance_final_1024_avx2(__m256i vsse, __m256i vsum, + unsigned int *const sse) { + // extract the low lane and add it to the high lane + const __m128i vsum_128 = mm256_add_hi_lo_epi16(vsum); + const __m128i vsum_64 = + _mm_add_epi32(_mm_cvtepi16_epi32(vsum_128), + _mm_cvtepi16_epi32(_mm_srli_si128(vsum_128, 8))); + return variance_final_from_32bit_sum_avx2(vsse, vsum_64, sse); +} + +static INLINE __m256i sum_to_32bit_avx2(const __m256i sum) { + const __m256i sum_lo = _mm256_cvtepi16_epi32(_mm256_castsi256_si128(sum)); + const __m256i sum_hi = + _mm256_cvtepi16_epi32(_mm256_extractf128_si256(sum, 1)); + return _mm256_add_epi32(sum_lo, sum_hi); +} + +// handle 2048 pixels (32x64, 64x32) +static INLINE int variance_final_2048_avx2(__m256i vsse, __m256i vsum, + unsigned int *const sse) { + vsum = sum_to_32bit_avx2(vsum); + const __m128i vsum_128 = mm256_add_hi_lo_epi32(vsum); + return variance_final_from_32bit_sum_avx2(vsse, vsum_128, sse); +} + +static INLINE void variance16_kernel_avx2( + const uint8_t *const src, const int src_stride, const uint8_t *const ref, + const int ref_stride, __m256i *const sse, __m256i *const sum) { + const __m128i s0 = _mm_loadu_si128((__m128i const *)(src + 0 * src_stride)); + const __m128i s1 = _mm_loadu_si128((__m128i const *)(src + 1 * src_stride)); + const __m128i r0 = _mm_loadu_si128((__m128i const *)(ref + 0 * ref_stride)); + const __m128i r1 = _mm_loadu_si128((__m128i const *)(ref + 1 * ref_stride)); + const __m256i s = _mm256_inserti128_si256(_mm256_castsi128_si256(s0), s1, 1); + const __m256i r = _mm256_inserti128_si256(_mm256_castsi128_si256(r0), r1, 1); + variance_kernel_avx2(s, r, sse, sum); +} + +static INLINE void variance32_kernel_avx2(const uint8_t *const src, + const uint8_t *const ref, + __m256i *const sse, + __m256i *const sum) { + const __m256i s = _mm256_loadu_si256((__m256i const *)(src)); + const __m256i r = _mm256_loadu_si256((__m256i const *)(ref)); + variance_kernel_avx2(s, r, sse, sum); +} + +static INLINE void variance16_avx2(const uint8_t *src, const int src_stride, + const uint8_t *ref, const int ref_stride, + const int h, __m256i *const vsse, + __m256i *const vsum) { + *vsum = _mm256_setzero_si256(); + + for (int i = 0; i < h; i += 2) { + variance16_kernel_avx2(src, src_stride, ref, ref_stride, vsse, vsum); + src += 2 * src_stride; + ref += 2 * ref_stride; + } +} + +static INLINE void variance32_avx2(const uint8_t *src, const int src_stride, + const uint8_t *ref, const int ref_stride, + const int h, __m256i *const vsse, + __m256i *const vsum) { + *vsum = _mm256_setzero_si256(); + + for (int i = 0; i < h; i++) { + variance32_kernel_avx2(src, ref, vsse, vsum); + src += src_stride; + ref += ref_stride; + } +} + +static INLINE void variance64_avx2(const uint8_t *src, const int src_stride, + const uint8_t *ref, const int ref_stride, + const int h, __m256i *const vsse, + __m256i *const vsum) { + *vsum = _mm256_setzero_si256(); + + for (int i = 0; i < h; i++) { + variance32_kernel_avx2(src + 0, ref + 0, vsse, vsum); + variance32_kernel_avx2(src + 32, ref + 32, vsse, vsum); + src += src_stride; + ref += ref_stride; + } +} + +static INLINE void variance128_avx2(const uint8_t *src, const int src_stride, + const uint8_t *ref, const int ref_stride, + const int h, __m256i *const vsse, + __m256i *const vsum) { + *vsum = _mm256_setzero_si256(); + + for (int i = 0; i < h; i++) { + variance32_kernel_avx2(src + 0, ref + 0, vsse, vsum); + variance32_kernel_avx2(src + 32, ref + 32, vsse, vsum); + variance32_kernel_avx2(src + 64, ref + 64, vsse, vsum); + variance32_kernel_avx2(src + 96, ref + 96, vsse, vsum); + src += src_stride; + ref += ref_stride; + } +} + +#define AOM_VAR_NO_LOOP_AVX2(bw, bh, bits, max_pixel) \ + unsigned int aom_variance##bw##x##bh##_avx2( \ + const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, \ + unsigned int *sse) { \ + __m256i vsse = _mm256_setzero_si256(); \ + __m256i vsum; \ + variance##bw##_avx2(src, src_stride, ref, ref_stride, bh, &vsse, &vsum); \ + const int sum = variance_final_##max_pixel##_avx2(vsse, vsum, sse); \ + return *sse - (uint32_t)(((int64_t)sum * sum) >> bits); \ + } + +AOM_VAR_NO_LOOP_AVX2(16, 4, 6, 512); +AOM_VAR_NO_LOOP_AVX2(16, 8, 7, 512); +AOM_VAR_NO_LOOP_AVX2(16, 16, 8, 512); +AOM_VAR_NO_LOOP_AVX2(16, 32, 9, 512); +AOM_VAR_NO_LOOP_AVX2(16, 64, 10, 1024); + +AOM_VAR_NO_LOOP_AVX2(32, 8, 8, 512); +AOM_VAR_NO_LOOP_AVX2(32, 16, 9, 512); +AOM_VAR_NO_LOOP_AVX2(32, 32, 10, 1024); +AOM_VAR_NO_LOOP_AVX2(32, 64, 11, 2048); + +AOM_VAR_NO_LOOP_AVX2(64, 16, 10, 1024); +AOM_VAR_NO_LOOP_AVX2(64, 32, 11, 2048); + +#define AOM_VAR_LOOP_AVX2(bw, bh, bits, uh) \ + unsigned int aom_variance##bw##x##bh##_avx2( \ + const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, \ + unsigned int *sse) { \ + __m256i vsse = _mm256_setzero_si256(); \ + __m256i vsum = _mm256_setzero_si256(); \ + for (int i = 0; i < (bh / uh); i++) { \ + __m256i vsum16; \ + variance##bw##_avx2(src, src_stride, ref, ref_stride, uh, &vsse, \ + &vsum16); \ + vsum = _mm256_add_epi32(vsum, sum_to_32bit_avx2(vsum16)); \ + src += uh * src_stride; \ + ref += uh * ref_stride; \ + } \ + const __m128i vsum_128 = mm256_add_hi_lo_epi32(vsum); \ + const int sum = variance_final_from_32bit_sum_avx2(vsse, vsum_128, sse); \ + return *sse - (unsigned int)(((int64_t)sum * sum) >> bits); \ + } + +AOM_VAR_LOOP_AVX2(64, 64, 12, 32); // 64x32 * ( 64/32) +AOM_VAR_LOOP_AVX2(64, 128, 13, 32); // 64x32 * (128/32) +AOM_VAR_LOOP_AVX2(128, 64, 13, 16); // 128x16 * ( 64/16) +AOM_VAR_LOOP_AVX2(128, 128, 14, 16); // 128x16 * (128/16) + +unsigned int aom_mse16x16_avx2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride, + unsigned int *sse) { + aom_variance16x16_avx2(src, src_stride, ref, ref_stride, sse); + return *sse; +} + +unsigned int aom_sub_pixel_variance32xh_avx2(const uint8_t *src, int src_stride, + int x_offset, int y_offset, + const uint8_t *dst, int dst_stride, + int height, unsigned int *sse); + +unsigned int aom_sub_pixel_avg_variance32xh_avx2( + const uint8_t *src, int src_stride, int x_offset, int y_offset, + const uint8_t *dst, int dst_stride, const uint8_t *sec, int sec_stride, + int height, unsigned int *sseptr); + +#define AOM_SUB_PIXEL_VAR_AVX2(w, h, wf, wlog2, hlog2) \ + unsigned int aom_sub_pixel_variance##w##x##h##_avx2( \ + const uint8_t *src, int src_stride, int x_offset, int y_offset, \ + const uint8_t *dst, int dst_stride, unsigned int *sse_ptr) { \ + /*Avoid overflow in helper by capping height.*/ \ + const int hf = AOMMIN(h, 64); \ + unsigned int sse = 0; \ + int se = 0; \ + for (int i = 0; i < (w / wf); ++i) { \ + const uint8_t *src_ptr = src; \ + const uint8_t *dst_ptr = dst; \ + for (int j = 0; j < (h / hf); ++j) { \ + unsigned int sse2; \ + const int se2 = aom_sub_pixel_variance##wf##xh_avx2( \ + src_ptr, src_stride, x_offset, y_offset, dst_ptr, dst_stride, hf, \ + &sse2); \ + dst_ptr += hf * dst_stride; \ + src_ptr += hf * src_stride; \ + se += se2; \ + sse += sse2; \ + } \ + src += wf; \ + dst += wf; \ + } \ + *sse_ptr = sse; \ + return sse - (unsigned int)(((int64_t)se * se) >> (wlog2 + hlog2)); \ + } + +AOM_SUB_PIXEL_VAR_AVX2(128, 128, 32, 7, 7); +AOM_SUB_PIXEL_VAR_AVX2(128, 64, 32, 7, 6); +AOM_SUB_PIXEL_VAR_AVX2(64, 128, 32, 6, 7); +AOM_SUB_PIXEL_VAR_AVX2(64, 64, 32, 6, 6); +AOM_SUB_PIXEL_VAR_AVX2(64, 32, 32, 6, 5); +AOM_SUB_PIXEL_VAR_AVX2(32, 64, 32, 5, 6); +AOM_SUB_PIXEL_VAR_AVX2(32, 32, 32, 5, 5); +AOM_SUB_PIXEL_VAR_AVX2(32, 16, 32, 5, 4); + +#define AOM_SUB_PIXEL_AVG_VAR_AVX2(w, h, wf, wlog2, hlog2) \ + unsigned int aom_sub_pixel_avg_variance##w##x##h##_avx2( \ + const uint8_t *src, int src_stride, int x_offset, int y_offset, \ + const uint8_t *dst, int dst_stride, unsigned int *sse_ptr, \ + const uint8_t *sec) { \ + /*Avoid overflow in helper by capping height.*/ \ + const int hf = AOMMIN(h, 64); \ + unsigned int sse = 0; \ + int se = 0; \ + for (int i = 0; i < (w / wf); ++i) { \ + const uint8_t *src_ptr = src; \ + const uint8_t *dst_ptr = dst; \ + const uint8_t *sec_ptr = sec; \ + for (int j = 0; j < (h / hf); ++j) { \ + unsigned int sse2; \ + const int se2 = aom_sub_pixel_avg_variance##wf##xh_avx2( \ + src_ptr, src_stride, x_offset, y_offset, dst_ptr, dst_stride, \ + sec_ptr, w, hf, &sse2); \ + dst_ptr += hf * dst_stride; \ + src_ptr += hf * src_stride; \ + sec_ptr += hf * w; \ + se += se2; \ + sse += sse2; \ + } \ + src += wf; \ + dst += wf; \ + sec += wf; \ + } \ + *sse_ptr = sse; \ + return sse - (unsigned int)(((int64_t)se * se) >> (wlog2 + hlog2)); \ + } + +AOM_SUB_PIXEL_AVG_VAR_AVX2(128, 128, 32, 7, 7); +AOM_SUB_PIXEL_AVG_VAR_AVX2(128, 64, 32, 7, 6); +AOM_SUB_PIXEL_AVG_VAR_AVX2(64, 128, 32, 6, 7); +AOM_SUB_PIXEL_AVG_VAR_AVX2(64, 64, 32, 6, 6); +AOM_SUB_PIXEL_AVG_VAR_AVX2(64, 32, 32, 6, 5); +AOM_SUB_PIXEL_AVG_VAR_AVX2(32, 64, 32, 5, 6); +AOM_SUB_PIXEL_AVG_VAR_AVX2(32, 32, 32, 5, 5); +AOM_SUB_PIXEL_AVG_VAR_AVX2(32, 16, 32, 5, 4); + +static INLINE __m256i mm256_loadu2(const uint8_t *p0, const uint8_t *p1) { + const __m256i d = + _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)p1)); + return _mm256_insertf128_si256(d, _mm_loadu_si128((const __m128i *)p0), 1); +} + +static INLINE __m256i mm256_loadu2_16(const uint16_t *p0, const uint16_t *p1) { + const __m256i d = + _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)p1)); + return _mm256_insertf128_si256(d, _mm_loadu_si128((const __m128i *)p0), 1); +} + +static INLINE void comp_mask_pred_line_avx2(const __m256i s0, const __m256i s1, + const __m256i a, + uint8_t *comp_pred) { + const __m256i alpha_max = _mm256_set1_epi8(AOM_BLEND_A64_MAX_ALPHA); + const int16_t round_bits = 15 - AOM_BLEND_A64_ROUND_BITS; + const __m256i round_offset = _mm256_set1_epi16(1 << (round_bits)); + + const __m256i ma = _mm256_sub_epi8(alpha_max, a); + + const __m256i ssAL = _mm256_unpacklo_epi8(s0, s1); + const __m256i aaAL = _mm256_unpacklo_epi8(a, ma); + const __m256i ssAH = _mm256_unpackhi_epi8(s0, s1); + const __m256i aaAH = _mm256_unpackhi_epi8(a, ma); + + const __m256i blendAL = _mm256_maddubs_epi16(ssAL, aaAL); + const __m256i blendAH = _mm256_maddubs_epi16(ssAH, aaAH); + const __m256i roundAL = _mm256_mulhrs_epi16(blendAL, round_offset); + const __m256i roundAH = _mm256_mulhrs_epi16(blendAH, round_offset); + + const __m256i roundA = _mm256_packus_epi16(roundAL, roundAH); + _mm256_storeu_si256((__m256i *)(comp_pred), roundA); +} + +void aom_comp_mask_pred_avx2(uint8_t *comp_pred, const uint8_t *pred, int width, + int height, const uint8_t *ref, int ref_stride, + const uint8_t *mask, int mask_stride, + int invert_mask) { + int i = 0; + const uint8_t *src0 = invert_mask ? pred : ref; + const uint8_t *src1 = invert_mask ? ref : pred; + const int stride0 = invert_mask ? width : ref_stride; + const int stride1 = invert_mask ? ref_stride : width; + if (width == 8) { + comp_mask_pred_8_ssse3(comp_pred, height, src0, stride0, src1, stride1, + mask, mask_stride); + } else if (width == 16) { + do { + const __m256i sA0 = mm256_loadu2(src0 + stride0, src0); + const __m256i sA1 = mm256_loadu2(src1 + stride1, src1); + const __m256i aA = mm256_loadu2(mask + mask_stride, mask); + src0 += (stride0 << 1); + src1 += (stride1 << 1); + mask += (mask_stride << 1); + const __m256i sB0 = mm256_loadu2(src0 + stride0, src0); + const __m256i sB1 = mm256_loadu2(src1 + stride1, src1); + const __m256i aB = mm256_loadu2(mask + mask_stride, mask); + src0 += (stride0 << 1); + src1 += (stride1 << 1); + mask += (mask_stride << 1); + // comp_pred's stride == width == 16 + comp_mask_pred_line_avx2(sA0, sA1, aA, comp_pred); + comp_mask_pred_line_avx2(sB0, sB1, aB, comp_pred + 32); + comp_pred += (16 << 2); + i += 4; + } while (i < height); + } else { // for width == 32 + do { + const __m256i sA0 = _mm256_lddqu_si256((const __m256i *)(src0)); + const __m256i sA1 = _mm256_lddqu_si256((const __m256i *)(src1)); + const __m256i aA = _mm256_lddqu_si256((const __m256i *)(mask)); + + const __m256i sB0 = _mm256_lddqu_si256((const __m256i *)(src0 + stride0)); + const __m256i sB1 = _mm256_lddqu_si256((const __m256i *)(src1 + stride1)); + const __m256i aB = + _mm256_lddqu_si256((const __m256i *)(mask + mask_stride)); + + comp_mask_pred_line_avx2(sA0, sA1, aA, comp_pred); + comp_mask_pred_line_avx2(sB0, sB1, aB, comp_pred + 32); + comp_pred += (32 << 1); + + src0 += (stride0 << 1); + src1 += (stride1 << 1); + mask += (mask_stride << 1); + i += 2; + } while (i < height); + } +} + +static INLINE __m256i highbd_comp_mask_pred_line_avx2(const __m256i s0, + const __m256i s1, + const __m256i a) { + const __m256i alpha_max = _mm256_set1_epi16((1 << AOM_BLEND_A64_ROUND_BITS)); + const __m256i round_const = + _mm256_set1_epi32((1 << AOM_BLEND_A64_ROUND_BITS) >> 1); + const __m256i a_inv = _mm256_sub_epi16(alpha_max, a); + + const __m256i s_lo = _mm256_unpacklo_epi16(s0, s1); + const __m256i a_lo = _mm256_unpacklo_epi16(a, a_inv); + const __m256i pred_lo = _mm256_madd_epi16(s_lo, a_lo); + const __m256i pred_l = _mm256_srai_epi32( + _mm256_add_epi32(pred_lo, round_const), AOM_BLEND_A64_ROUND_BITS); + + const __m256i s_hi = _mm256_unpackhi_epi16(s0, s1); + const __m256i a_hi = _mm256_unpackhi_epi16(a, a_inv); + const __m256i pred_hi = _mm256_madd_epi16(s_hi, a_hi); + const __m256i pred_h = _mm256_srai_epi32( + _mm256_add_epi32(pred_hi, round_const), AOM_BLEND_A64_ROUND_BITS); + + const __m256i comp = _mm256_packs_epi32(pred_l, pred_h); + + return comp; +} + +void aom_highbd_comp_mask_pred_avx2(uint8_t *comp_pred8, const uint8_t *pred8, + int width, int height, const uint8_t *ref8, + int ref_stride, const uint8_t *mask, + int mask_stride, int invert_mask) { + int i = 0; + uint16_t *pred = CONVERT_TO_SHORTPTR(pred8); + uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); + uint16_t *comp_pred = CONVERT_TO_SHORTPTR(comp_pred8); + const uint16_t *src0 = invert_mask ? pred : ref; + const uint16_t *src1 = invert_mask ? ref : pred; + const int stride0 = invert_mask ? width : ref_stride; + const int stride1 = invert_mask ? ref_stride : width; + const __m256i zero = _mm256_setzero_si256(); + + if (width == 8) { + do { + const __m256i s0 = mm256_loadu2_16(src0 + stride0, src0); + const __m256i s1 = mm256_loadu2_16(src1 + stride1, src1); + + const __m128i m_l = _mm_loadl_epi64((const __m128i *)mask); + const __m128i m_h = _mm_loadl_epi64((const __m128i *)(mask + 8)); + + __m256i m = _mm256_castsi128_si256(m_l); + m = _mm256_insertf128_si256(m, m_h, 1); + const __m256i m_16 = _mm256_unpacklo_epi8(m, zero); + + const __m256i comp = highbd_comp_mask_pred_line_avx2(s0, s1, m_16); + + _mm_storeu_si128((__m128i *)(comp_pred), _mm256_castsi256_si128(comp)); + + _mm_storeu_si128((__m128i *)(comp_pred + width), + _mm256_extractf128_si256(comp, 1)); + + src0 += (stride0 << 1); + src1 += (stride1 << 1); + mask += (mask_stride << 1); + comp_pred += (width << 1); + i += 2; + } while (i < height); + } else if (width == 16) { + do { + const __m256i s0 = _mm256_loadu_si256((const __m256i *)(src0)); + const __m256i s1 = _mm256_loadu_si256((const __m256i *)(src1)); + const __m256i m_16 = + _mm256_cvtepu8_epi16(_mm_loadu_si128((const __m128i *)mask)); + + const __m256i comp = highbd_comp_mask_pred_line_avx2(s0, s1, m_16); + + _mm256_storeu_si256((__m256i *)comp_pred, comp); + + src0 += stride0; + src1 += stride1; + mask += mask_stride; + comp_pred += width; + i += 1; + } while (i < height); + } else if (width == 32) { + do { + const __m256i s0 = _mm256_loadu_si256((const __m256i *)src0); + const __m256i s2 = _mm256_loadu_si256((const __m256i *)(src0 + 16)); + const __m256i s1 = _mm256_loadu_si256((const __m256i *)src1); + const __m256i s3 = _mm256_loadu_si256((const __m256i *)(src1 + 16)); + + const __m256i m01_16 = + _mm256_cvtepu8_epi16(_mm_loadu_si128((const __m128i *)mask)); + const __m256i m23_16 = + _mm256_cvtepu8_epi16(_mm_loadu_si128((const __m128i *)(mask + 16))); + + const __m256i comp = highbd_comp_mask_pred_line_avx2(s0, s1, m01_16); + const __m256i comp1 = highbd_comp_mask_pred_line_avx2(s2, s3, m23_16); + + _mm256_storeu_si256((__m256i *)comp_pred, comp); + _mm256_storeu_si256((__m256i *)(comp_pred + 16), comp1); + + src0 += stride0; + src1 += stride1; + mask += mask_stride; + comp_pred += width; + i += 1; + } while (i < height); + } +} diff --git a/third_party/aom/aom_dsp/x86/variance_impl_avx2.c b/third_party/aom/aom_dsp/x86/variance_impl_avx2.c new file mode 100644 index 0000000000..88e27aef3a --- /dev/null +++ b/third_party/aom/aom_dsp/x86/variance_impl_avx2.c @@ -0,0 +1,517 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include // AVX2 + +#include "config/aom_dsp_rtcd.h" + +#include "aom_ports/mem.h" + +/* clang-format off */ +DECLARE_ALIGNED(32, static const uint8_t, bilinear_filters_avx2[512]) = { + 16, 0, 16, 0, 16, 0, 16, 0, 16, 0, 16, 0, 16, 0, 16, 0, + 16, 0, 16, 0, 16, 0, 16, 0, 16, 0, 16, 0, 16, 0, 16, 0, + 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, + 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, + 12, 4, 12, 4, 12, 4, 12, 4, 12, 4, 12, 4, 12, 4, 12, 4, + 12, 4, 12, 4, 12, 4, 12, 4, 12, 4, 12, 4, 12, 4, 12, 4, + 10, 6, 10, 6, 10, 6, 10, 6, 10, 6, 10, 6, 10, 6, 10, 6, + 10, 6, 10, 6, 10, 6, 10, 6, 10, 6, 10, 6, 10, 6, 10, 6, + 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, + 6, 10, 6, 10, 6, 10, 6, 10, 6, 10, 6, 10, 6, 10, 6, 10, + 6, 10, 6, 10, 6, 10, 6, 10, 6, 10, 6, 10, 6, 10, 6, 10, + 4, 12, 4, 12, 4, 12, 4, 12, 4, 12, 4, 12, 4, 12, 4, 12, + 4, 12, 4, 12, 4, 12, 4, 12, 4, 12, 4, 12, 4, 12, 4, 12, + 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, + 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, +}; +/* clang-format on */ + +#define FILTER_SRC(filter) \ + /* filter the source */ \ + exp_src_lo = _mm256_maddubs_epi16(exp_src_lo, filter); \ + exp_src_hi = _mm256_maddubs_epi16(exp_src_hi, filter); \ + \ + /* add 8 to source */ \ + exp_src_lo = _mm256_add_epi16(exp_src_lo, pw8); \ + exp_src_hi = _mm256_add_epi16(exp_src_hi, pw8); \ + \ + /* divide source by 16 */ \ + exp_src_lo = _mm256_srai_epi16(exp_src_lo, 4); \ + exp_src_hi = _mm256_srai_epi16(exp_src_hi, 4); + +#define MERGE_WITH_SRC(src_reg, reg) \ + exp_src_lo = _mm256_unpacklo_epi8(src_reg, reg); \ + exp_src_hi = _mm256_unpackhi_epi8(src_reg, reg); + +#define LOAD_SRC_DST \ + /* load source and destination */ \ + src_reg = _mm256_loadu_si256((__m256i const *)(src)); \ + dst_reg = _mm256_loadu_si256((__m256i const *)(dst)); + +#define AVG_NEXT_SRC(src_reg, size_stride) \ + src_next_reg = _mm256_loadu_si256((__m256i const *)(src + size_stride)); \ + /* average between current and next stride source */ \ + src_reg = _mm256_avg_epu8(src_reg, src_next_reg); + +#define MERGE_NEXT_SRC(src_reg, size_stride) \ + src_next_reg = _mm256_loadu_si256((__m256i const *)(src + size_stride)); \ + MERGE_WITH_SRC(src_reg, src_next_reg) + +#define CALC_SUM_SSE_INSIDE_LOOP \ + /* expand each byte to 2 bytes */ \ + exp_dst_lo = _mm256_unpacklo_epi8(dst_reg, zero_reg); \ + exp_dst_hi = _mm256_unpackhi_epi8(dst_reg, zero_reg); \ + /* source - dest */ \ + exp_src_lo = _mm256_sub_epi16(exp_src_lo, exp_dst_lo); \ + exp_src_hi = _mm256_sub_epi16(exp_src_hi, exp_dst_hi); \ + /* caculate sum */ \ + sum_reg = _mm256_add_epi16(sum_reg, exp_src_lo); \ + exp_src_lo = _mm256_madd_epi16(exp_src_lo, exp_src_lo); \ + sum_reg = _mm256_add_epi16(sum_reg, exp_src_hi); \ + exp_src_hi = _mm256_madd_epi16(exp_src_hi, exp_src_hi); \ + /* calculate sse */ \ + sse_reg = _mm256_add_epi32(sse_reg, exp_src_lo); \ + sse_reg = _mm256_add_epi32(sse_reg, exp_src_hi); + +// final calculation to sum and sse +#define CALC_SUM_AND_SSE \ + res_cmp = _mm256_cmpgt_epi16(zero_reg, sum_reg); \ + sse_reg_hi = _mm256_srli_si256(sse_reg, 8); \ + sum_reg_lo = _mm256_unpacklo_epi16(sum_reg, res_cmp); \ + sum_reg_hi = _mm256_unpackhi_epi16(sum_reg, res_cmp); \ + sse_reg = _mm256_add_epi32(sse_reg, sse_reg_hi); \ + sum_reg = _mm256_add_epi32(sum_reg_lo, sum_reg_hi); \ + \ + sse_reg_hi = _mm256_srli_si256(sse_reg, 4); \ + sum_reg_hi = _mm256_srli_si256(sum_reg, 8); \ + \ + sse_reg = _mm256_add_epi32(sse_reg, sse_reg_hi); \ + sum_reg = _mm256_add_epi32(sum_reg, sum_reg_hi); \ + *((int *)sse) = _mm_cvtsi128_si32(_mm256_castsi256_si128(sse_reg)) + \ + _mm_cvtsi128_si32(_mm256_extractf128_si256(sse_reg, 1)); \ + sum_reg_hi = _mm256_srli_si256(sum_reg, 4); \ + sum_reg = _mm256_add_epi32(sum_reg, sum_reg_hi); \ + sum = _mm_cvtsi128_si32(_mm256_castsi256_si128(sum_reg)) + \ + _mm_cvtsi128_si32(_mm256_extractf128_si256(sum_reg, 1)); + +unsigned int aom_sub_pixel_variance32xh_avx2(const uint8_t *src, int src_stride, + int x_offset, int y_offset, + const uint8_t *dst, int dst_stride, + int height, unsigned int *sse) { + __m256i src_reg, dst_reg, exp_src_lo, exp_src_hi, exp_dst_lo, exp_dst_hi; + __m256i sse_reg, sum_reg, sse_reg_hi, res_cmp, sum_reg_lo, sum_reg_hi; + __m256i zero_reg; + int i, sum; + sum_reg = _mm256_set1_epi16(0); + sse_reg = _mm256_set1_epi16(0); + zero_reg = _mm256_set1_epi16(0); + + // x_offset = 0 and y_offset = 0 + if (x_offset == 0) { + if (y_offset == 0) { + for (i = 0; i < height; i++) { + LOAD_SRC_DST + // expend each byte to 2 bytes + MERGE_WITH_SRC(src_reg, zero_reg) + CALC_SUM_SSE_INSIDE_LOOP + src += src_stride; + dst += dst_stride; + } + // x_offset = 0 and y_offset = 8 + } else if (y_offset == 8) { + __m256i src_next_reg; + for (i = 0; i < height; i++) { + LOAD_SRC_DST + AVG_NEXT_SRC(src_reg, src_stride) + // expend each byte to 2 bytes + MERGE_WITH_SRC(src_reg, zero_reg) + CALC_SUM_SSE_INSIDE_LOOP + src += src_stride; + dst += dst_stride; + } + // x_offset = 0 and y_offset = bilin interpolation + } else { + __m256i filter, pw8, src_next_reg; + + y_offset <<= 5; + filter = _mm256_load_si256( + (__m256i const *)(bilinear_filters_avx2 + y_offset)); + pw8 = _mm256_set1_epi16(8); + for (i = 0; i < height; i++) { + LOAD_SRC_DST + MERGE_NEXT_SRC(src_reg, src_stride) + FILTER_SRC(filter) + CALC_SUM_SSE_INSIDE_LOOP + src += src_stride; + dst += dst_stride; + } + } + // x_offset = 8 and y_offset = 0 + } else if (x_offset == 8) { + if (y_offset == 0) { + __m256i src_next_reg; + for (i = 0; i < height; i++) { + LOAD_SRC_DST + AVG_NEXT_SRC(src_reg, 1) + // expand each byte to 2 bytes + MERGE_WITH_SRC(src_reg, zero_reg) + CALC_SUM_SSE_INSIDE_LOOP + src += src_stride; + dst += dst_stride; + } + // x_offset = 8 and y_offset = 8 + } else if (y_offset == 8) { + __m256i src_next_reg, src_avg; + // load source and another source starting from the next + // following byte + src_reg = _mm256_loadu_si256((__m256i const *)(src)); + AVG_NEXT_SRC(src_reg, 1) + for (i = 0; i < height; i++) { + src_avg = src_reg; + src += src_stride; + LOAD_SRC_DST + AVG_NEXT_SRC(src_reg, 1) + // average between previous average to current average + src_avg = _mm256_avg_epu8(src_avg, src_reg); + // expand each byte to 2 bytes + MERGE_WITH_SRC(src_avg, zero_reg) + // save current source average + CALC_SUM_SSE_INSIDE_LOOP + dst += dst_stride; + } + // x_offset = 8 and y_offset = bilin interpolation + } else { + __m256i filter, pw8, src_next_reg, src_avg; + y_offset <<= 5; + filter = _mm256_load_si256( + (__m256i const *)(bilinear_filters_avx2 + y_offset)); + pw8 = _mm256_set1_epi16(8); + // load source and another source starting from the next + // following byte + src_reg = _mm256_loadu_si256((__m256i const *)(src)); + AVG_NEXT_SRC(src_reg, 1) + for (i = 0; i < height; i++) { + // save current source average + src_avg = src_reg; + src += src_stride; + LOAD_SRC_DST + AVG_NEXT_SRC(src_reg, 1) + MERGE_WITH_SRC(src_avg, src_reg) + FILTER_SRC(filter) + CALC_SUM_SSE_INSIDE_LOOP + dst += dst_stride; + } + } + // x_offset = bilin interpolation and y_offset = 0 + } else { + if (y_offset == 0) { + __m256i filter, pw8, src_next_reg; + x_offset <<= 5; + filter = _mm256_load_si256( + (__m256i const *)(bilinear_filters_avx2 + x_offset)); + pw8 = _mm256_set1_epi16(8); + for (i = 0; i < height; i++) { + LOAD_SRC_DST + MERGE_NEXT_SRC(src_reg, 1) + FILTER_SRC(filter) + CALC_SUM_SSE_INSIDE_LOOP + src += src_stride; + dst += dst_stride; + } + // x_offset = bilin interpolation and y_offset = 8 + } else if (y_offset == 8) { + __m256i filter, pw8, src_next_reg, src_pack; + x_offset <<= 5; + filter = _mm256_load_si256( + (__m256i const *)(bilinear_filters_avx2 + x_offset)); + pw8 = _mm256_set1_epi16(8); + src_reg = _mm256_loadu_si256((__m256i const *)(src)); + MERGE_NEXT_SRC(src_reg, 1) + FILTER_SRC(filter) + // convert each 16 bit to 8 bit to each low and high lane source + src_pack = _mm256_packus_epi16(exp_src_lo, exp_src_hi); + for (i = 0; i < height; i++) { + src += src_stride; + LOAD_SRC_DST + MERGE_NEXT_SRC(src_reg, 1) + FILTER_SRC(filter) + src_reg = _mm256_packus_epi16(exp_src_lo, exp_src_hi); + // average between previous pack to the current + src_pack = _mm256_avg_epu8(src_pack, src_reg); + MERGE_WITH_SRC(src_pack, zero_reg) + CALC_SUM_SSE_INSIDE_LOOP + src_pack = src_reg; + dst += dst_stride; + } + // x_offset = bilin interpolation and y_offset = bilin interpolation + } else { + __m256i xfilter, yfilter, pw8, src_next_reg, src_pack; + x_offset <<= 5; + xfilter = _mm256_load_si256( + (__m256i const *)(bilinear_filters_avx2 + x_offset)); + y_offset <<= 5; + yfilter = _mm256_load_si256( + (__m256i const *)(bilinear_filters_avx2 + y_offset)); + pw8 = _mm256_set1_epi16(8); + // load source and another source starting from the next + // following byte + src_reg = _mm256_loadu_si256((__m256i const *)(src)); + MERGE_NEXT_SRC(src_reg, 1) + + FILTER_SRC(xfilter) + // convert each 16 bit to 8 bit to each low and high lane source + src_pack = _mm256_packus_epi16(exp_src_lo, exp_src_hi); + for (i = 0; i < height; i++) { + src += src_stride; + LOAD_SRC_DST + MERGE_NEXT_SRC(src_reg, 1) + FILTER_SRC(xfilter) + src_reg = _mm256_packus_epi16(exp_src_lo, exp_src_hi); + // merge previous pack to current pack source + MERGE_WITH_SRC(src_pack, src_reg) + // filter the source + FILTER_SRC(yfilter) + src_pack = src_reg; + CALC_SUM_SSE_INSIDE_LOOP + dst += dst_stride; + } + } + } + CALC_SUM_AND_SSE + _mm256_zeroupper(); + return sum; +} + +unsigned int aom_sub_pixel_avg_variance32xh_avx2( + const uint8_t *src, int src_stride, int x_offset, int y_offset, + const uint8_t *dst, int dst_stride, const uint8_t *sec, int sec_stride, + int height, unsigned int *sse) { + __m256i sec_reg; + __m256i src_reg, dst_reg, exp_src_lo, exp_src_hi, exp_dst_lo, exp_dst_hi; + __m256i sse_reg, sum_reg, sse_reg_hi, res_cmp, sum_reg_lo, sum_reg_hi; + __m256i zero_reg; + int i, sum; + sum_reg = _mm256_set1_epi16(0); + sse_reg = _mm256_set1_epi16(0); + zero_reg = _mm256_set1_epi16(0); + + // x_offset = 0 and y_offset = 0 + if (x_offset == 0) { + if (y_offset == 0) { + for (i = 0; i < height; i++) { + LOAD_SRC_DST + sec_reg = _mm256_loadu_si256((__m256i const *)(sec)); + src_reg = _mm256_avg_epu8(src_reg, sec_reg); + sec += sec_stride; + // expend each byte to 2 bytes + MERGE_WITH_SRC(src_reg, zero_reg) + CALC_SUM_SSE_INSIDE_LOOP + src += src_stride; + dst += dst_stride; + } + } else if (y_offset == 8) { + __m256i src_next_reg; + for (i = 0; i < height; i++) { + LOAD_SRC_DST + AVG_NEXT_SRC(src_reg, src_stride) + sec_reg = _mm256_loadu_si256((__m256i const *)(sec)); + src_reg = _mm256_avg_epu8(src_reg, sec_reg); + sec += sec_stride; + // expend each byte to 2 bytes + MERGE_WITH_SRC(src_reg, zero_reg) + CALC_SUM_SSE_INSIDE_LOOP + src += src_stride; + dst += dst_stride; + } + // x_offset = 0 and y_offset = bilin interpolation + } else { + __m256i filter, pw8, src_next_reg; + + y_offset <<= 5; + filter = _mm256_load_si256( + (__m256i const *)(bilinear_filters_avx2 + y_offset)); + pw8 = _mm256_set1_epi16(8); + for (i = 0; i < height; i++) { + LOAD_SRC_DST + MERGE_NEXT_SRC(src_reg, src_stride) + FILTER_SRC(filter) + src_reg = _mm256_packus_epi16(exp_src_lo, exp_src_hi); + sec_reg = _mm256_loadu_si256((__m256i const *)(sec)); + src_reg = _mm256_avg_epu8(src_reg, sec_reg); + sec += sec_stride; + MERGE_WITH_SRC(src_reg, zero_reg) + CALC_SUM_SSE_INSIDE_LOOP + src += src_stride; + dst += dst_stride; + } + } + // x_offset = 8 and y_offset = 0 + } else if (x_offset == 8) { + if (y_offset == 0) { + __m256i src_next_reg; + for (i = 0; i < height; i++) { + LOAD_SRC_DST + AVG_NEXT_SRC(src_reg, 1) + sec_reg = _mm256_loadu_si256((__m256i const *)(sec)); + src_reg = _mm256_avg_epu8(src_reg, sec_reg); + sec += sec_stride; + // expand each byte to 2 bytes + MERGE_WITH_SRC(src_reg, zero_reg) + CALC_SUM_SSE_INSIDE_LOOP + src += src_stride; + dst += dst_stride; + } + // x_offset = 8 and y_offset = 8 + } else if (y_offset == 8) { + __m256i src_next_reg, src_avg; + // load source and another source starting from the next + // following byte + src_reg = _mm256_loadu_si256((__m256i const *)(src)); + AVG_NEXT_SRC(src_reg, 1) + for (i = 0; i < height; i++) { + // save current source average + src_avg = src_reg; + src += src_stride; + LOAD_SRC_DST + AVG_NEXT_SRC(src_reg, 1) + // average between previous average to current average + src_avg = _mm256_avg_epu8(src_avg, src_reg); + sec_reg = _mm256_loadu_si256((__m256i const *)(sec)); + src_avg = _mm256_avg_epu8(src_avg, sec_reg); + sec += sec_stride; + // expand each byte to 2 bytes + MERGE_WITH_SRC(src_avg, zero_reg) + CALC_SUM_SSE_INSIDE_LOOP + dst += dst_stride; + } + // x_offset = 8 and y_offset = bilin interpolation + } else { + __m256i filter, pw8, src_next_reg, src_avg; + y_offset <<= 5; + filter = _mm256_load_si256( + (__m256i const *)(bilinear_filters_avx2 + y_offset)); + pw8 = _mm256_set1_epi16(8); + // load source and another source starting from the next + // following byte + src_reg = _mm256_loadu_si256((__m256i const *)(src)); + AVG_NEXT_SRC(src_reg, 1) + for (i = 0; i < height; i++) { + // save current source average + src_avg = src_reg; + src += src_stride; + LOAD_SRC_DST + AVG_NEXT_SRC(src_reg, 1) + MERGE_WITH_SRC(src_avg, src_reg) + FILTER_SRC(filter) + src_avg = _mm256_packus_epi16(exp_src_lo, exp_src_hi); + sec_reg = _mm256_loadu_si256((__m256i const *)(sec)); + src_avg = _mm256_avg_epu8(src_avg, sec_reg); + // expand each byte to 2 bytes + MERGE_WITH_SRC(src_avg, zero_reg) + sec += sec_stride; + CALC_SUM_SSE_INSIDE_LOOP + dst += dst_stride; + } + } + // x_offset = bilin interpolation and y_offset = 0 + } else { + if (y_offset == 0) { + __m256i filter, pw8, src_next_reg; + x_offset <<= 5; + filter = _mm256_load_si256( + (__m256i const *)(bilinear_filters_avx2 + x_offset)); + pw8 = _mm256_set1_epi16(8); + for (i = 0; i < height; i++) { + LOAD_SRC_DST + MERGE_NEXT_SRC(src_reg, 1) + FILTER_SRC(filter) + src_reg = _mm256_packus_epi16(exp_src_lo, exp_src_hi); + sec_reg = _mm256_loadu_si256((__m256i const *)(sec)); + src_reg = _mm256_avg_epu8(src_reg, sec_reg); + MERGE_WITH_SRC(src_reg, zero_reg) + sec += sec_stride; + CALC_SUM_SSE_INSIDE_LOOP + src += src_stride; + dst += dst_stride; + } + // x_offset = bilin interpolation and y_offset = 8 + } else if (y_offset == 8) { + __m256i filter, pw8, src_next_reg, src_pack; + x_offset <<= 5; + filter = _mm256_load_si256( + (__m256i const *)(bilinear_filters_avx2 + x_offset)); + pw8 = _mm256_set1_epi16(8); + src_reg = _mm256_loadu_si256((__m256i const *)(src)); + MERGE_NEXT_SRC(src_reg, 1) + FILTER_SRC(filter) + // convert each 16 bit to 8 bit to each low and high lane source + src_pack = _mm256_packus_epi16(exp_src_lo, exp_src_hi); + for (i = 0; i < height; i++) { + src += src_stride; + LOAD_SRC_DST + MERGE_NEXT_SRC(src_reg, 1) + FILTER_SRC(filter) + src_reg = _mm256_packus_epi16(exp_src_lo, exp_src_hi); + // average between previous pack to the current + src_pack = _mm256_avg_epu8(src_pack, src_reg); + sec_reg = _mm256_loadu_si256((__m256i const *)(sec)); + src_pack = _mm256_avg_epu8(src_pack, sec_reg); + sec += sec_stride; + MERGE_WITH_SRC(src_pack, zero_reg) + src_pack = src_reg; + CALC_SUM_SSE_INSIDE_LOOP + dst += dst_stride; + } + // x_offset = bilin interpolation and y_offset = bilin interpolation + } else { + __m256i xfilter, yfilter, pw8, src_next_reg, src_pack; + x_offset <<= 5; + xfilter = _mm256_load_si256( + (__m256i const *)(bilinear_filters_avx2 + x_offset)); + y_offset <<= 5; + yfilter = _mm256_load_si256( + (__m256i const *)(bilinear_filters_avx2 + y_offset)); + pw8 = _mm256_set1_epi16(8); + // load source and another source starting from the next + // following byte + src_reg = _mm256_loadu_si256((__m256i const *)(src)); + MERGE_NEXT_SRC(src_reg, 1) + + FILTER_SRC(xfilter) + // convert each 16 bit to 8 bit to each low and high lane source + src_pack = _mm256_packus_epi16(exp_src_lo, exp_src_hi); + for (i = 0; i < height; i++) { + src += src_stride; + LOAD_SRC_DST + MERGE_NEXT_SRC(src_reg, 1) + FILTER_SRC(xfilter) + src_reg = _mm256_packus_epi16(exp_src_lo, exp_src_hi); + // merge previous pack to current pack source + MERGE_WITH_SRC(src_pack, src_reg) + // filter the source + FILTER_SRC(yfilter) + src_pack = _mm256_packus_epi16(exp_src_lo, exp_src_hi); + sec_reg = _mm256_loadu_si256((__m256i const *)(sec)); + src_pack = _mm256_avg_epu8(src_pack, sec_reg); + MERGE_WITH_SRC(src_pack, zero_reg) + src_pack = src_reg; + sec += sec_stride; + CALC_SUM_SSE_INSIDE_LOOP + dst += dst_stride; + } + } + } + CALC_SUM_AND_SSE + _mm256_zeroupper(); + return sum; +} diff --git a/third_party/aom/aom_dsp/x86/variance_impl_ssse3.c b/third_party/aom/aom_dsp/x86/variance_impl_ssse3.c new file mode 100644 index 0000000000..66b0d7d845 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/variance_impl_ssse3.c @@ -0,0 +1,129 @@ +/* + * Copyright (c) 2018, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include + +#include "config/aom_config.h" +#include "config/aom_dsp_rtcd.h" + +#include "aom_dsp/x86/synonyms.h" + +void aom_var_filter_block2d_bil_first_pass_ssse3( + const uint8_t *a, uint16_t *b, unsigned int src_pixels_per_line, + unsigned int pixel_step, unsigned int output_height, + unsigned int output_width, const uint8_t *filter) { + // Note: filter[0], filter[1] could be {128, 0}, where 128 will overflow + // in computation using _mm_maddubs_epi16. + // Change {128, 0} to {64, 0} and reduce FILTER_BITS by 1 to avoid overflow. + const int16_t round = (1 << (FILTER_BITS - 1)) >> 1; + const __m128i r = _mm_set1_epi16(round); + const uint8_t f0 = filter[0] >> 1; + const uint8_t f1 = filter[1] >> 1; + const __m128i filters = _mm_setr_epi8(f0, f1, f0, f1, f0, f1, f0, f1, f0, f1, + f0, f1, f0, f1, f0, f1); + unsigned int i, j; + (void)pixel_step; + + if (output_width >= 8) { + for (i = 0; i < output_height; ++i) { + for (j = 0; j < output_width; j += 8) { + // load source + __m128i source_low = xx_loadl_64(a); + __m128i source_hi = xx_loadl_64(a + 1); + + // unpack to: + // { a[0], a[1], a[1], a[2], a[2], a[3], a[3], a[4], + // a[4], a[5], a[5], a[6], a[6], a[7], a[7], a[8] } + __m128i source = _mm_unpacklo_epi8(source_low, source_hi); + + // b[i] = a[i] * filter[0] + a[i + 1] * filter[1] + __m128i res = _mm_maddubs_epi16(source, filters); + + // round + res = _mm_srai_epi16(_mm_add_epi16(res, r), FILTER_BITS - 1); + + xx_storeu_128(b, res); + + a += 8; + b += 8; + } + + a += src_pixels_per_line - output_width; + } + } else { + const __m128i shuffle_mask = + _mm_setr_epi8(0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8); + for (i = 0; i < output_height; ++i) { + // load source, only first 5 values are meaningful: + // { a[0], a[1], a[2], a[3], a[4], xxxx } + __m128i source = xx_loadl_64(a); + + // shuffle, up to the first 8 are useful + // { a[0], a[1], a[1], a[2], a[2], a[3], a[3], a[4], + // a[4], a[5], a[5], a[6], a[6], a[7], a[7], a[8] } + __m128i source_shuffle = _mm_shuffle_epi8(source, shuffle_mask); + + __m128i res = _mm_maddubs_epi16(source_shuffle, filters); + res = _mm_srai_epi16(_mm_add_epi16(res, r), FILTER_BITS - 1); + + xx_storel_64(b, res); + + a += src_pixels_per_line; + b += output_width; + } + } +} + +void aom_var_filter_block2d_bil_second_pass_ssse3( + const uint16_t *a, uint8_t *b, unsigned int src_pixels_per_line, + unsigned int pixel_step, unsigned int output_height, + unsigned int output_width, const uint8_t *filter) { + const int16_t round = (1 << FILTER_BITS) >> 1; + const __m128i r = _mm_set1_epi32(round); + const __m128i filters = + _mm_setr_epi16(filter[0], filter[1], filter[0], filter[1], filter[0], + filter[1], filter[0], filter[1]); + const __m128i shuffle_mask = + _mm_setr_epi8(0, 1, 8, 9, 2, 3, 10, 11, 4, 5, 12, 13, 6, 7, 14, 15); + const __m128i mask = + _mm_setr_epi8(0, 4, 8, 12, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1); + unsigned int i, j; + + for (i = 0; i < output_height; ++i) { + for (j = 0; j < output_width; j += 4) { + // load source as: + // { a[0], a[1], a[2], a[3], a[w], a[w+1], a[w+2], a[w+3] } + __m128i source1 = xx_loadl_64(a); + __m128i source2 = xx_loadl_64(a + pixel_step); + __m128i source = _mm_unpacklo_epi64(source1, source2); + + // shuffle source to: + // { a[0], a[w], a[1], a[w+1], a[2], a[w+2], a[3], a[w+3] } + __m128i source_shuffle = _mm_shuffle_epi8(source, shuffle_mask); + + // b[i] = a[i] * filter[0] + a[w + i] * filter[1] + __m128i res = _mm_madd_epi16(source_shuffle, filters); + + // round + res = _mm_srai_epi32(_mm_add_epi32(res, r), FILTER_BITS); + + // shuffle to get each lower 8 bit of every 32 bit + res = _mm_shuffle_epi8(res, mask); + + xx_storel_32(b, res); + + a += 4; + b += 4; + } + + a += src_pixels_per_line - output_width; + } +} diff --git a/third_party/aom/aom_dsp/x86/variance_sse2.c b/third_party/aom/aom_dsp/x86/variance_sse2.c new file mode 100644 index 0000000000..3c37e77c06 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/variance_sse2.c @@ -0,0 +1,806 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include +#include // SSE2 + +#include "config/aom_config.h" +#include "config/aom_dsp_rtcd.h" +#include "config/av1_rtcd.h" + +#include "aom_dsp/blend.h" +#include "aom_dsp/x86/synonyms.h" + +#include "aom_ports/mem.h" + +#include "av1/common/filter.h" +#include "av1/common/onyxc_int.h" +#include "av1/common/reconinter.h" + +unsigned int aom_get_mb_ss_sse2(const int16_t *src) { + __m128i vsum = _mm_setzero_si128(); + int i; + + for (i = 0; i < 32; ++i) { + const __m128i v = xx_loadu_128(src); + vsum = _mm_add_epi32(vsum, _mm_madd_epi16(v, v)); + src += 8; + } + + vsum = _mm_add_epi32(vsum, _mm_srli_si128(vsum, 8)); + vsum = _mm_add_epi32(vsum, _mm_srli_si128(vsum, 4)); + return _mm_cvtsi128_si32(vsum); +} + +static INLINE __m128i load4x2_sse2(const uint8_t *const p, const int stride) { + const __m128i p0 = _mm_cvtsi32_si128(*(const uint32_t *)(p + 0 * stride)); + const __m128i p1 = _mm_cvtsi32_si128(*(const uint32_t *)(p + 1 * stride)); + return _mm_unpacklo_epi8(_mm_unpacklo_epi32(p0, p1), _mm_setzero_si128()); +} + +static INLINE __m128i load8_8to16_sse2(const uint8_t *const p) { + const __m128i p0 = _mm_loadl_epi64((const __m128i *)p); + return _mm_unpacklo_epi8(p0, _mm_setzero_si128()); +} + +// Accumulate 4 32bit numbers in val to 1 32bit number +static INLINE unsigned int add32x4_sse2(__m128i val) { + val = _mm_add_epi32(val, _mm_srli_si128(val, 8)); + val = _mm_add_epi32(val, _mm_srli_si128(val, 4)); + return _mm_cvtsi128_si32(val); +} + +// Accumulate 8 16bit in sum to 4 32bit number +static INLINE __m128i sum_to_32bit_sse2(const __m128i sum) { + const __m128i sum_lo = _mm_srai_epi32(_mm_unpacklo_epi16(sum, sum), 16); + const __m128i sum_hi = _mm_srai_epi32(_mm_unpackhi_epi16(sum, sum), 16); + return _mm_add_epi32(sum_lo, sum_hi); +} + +static INLINE void variance_kernel_sse2(const __m128i src, const __m128i ref, + __m128i *const sse, + __m128i *const sum) { + const __m128i diff = _mm_sub_epi16(src, ref); + *sse = _mm_add_epi32(*sse, _mm_madd_epi16(diff, diff)); + *sum = _mm_add_epi16(*sum, diff); +} + +// Can handle 128 pixels' diff sum (such as 8x16 or 16x8) +// Slightly faster than variance_final_256_pel_sse2() +// diff sum of 128 pixels can still fit in 16bit integer +static INLINE void variance_final_128_pel_sse2(__m128i vsse, __m128i vsum, + unsigned int *const sse, + int *const sum) { + *sse = add32x4_sse2(vsse); + + vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 8)); + vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 4)); + vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 2)); + *sum = (int16_t)_mm_extract_epi16(vsum, 0); +} + +// Can handle 256 pixels' diff sum (such as 16x16) +static INLINE void variance_final_256_pel_sse2(__m128i vsse, __m128i vsum, + unsigned int *const sse, + int *const sum) { + *sse = add32x4_sse2(vsse); + + vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 8)); + vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 4)); + *sum = (int16_t)_mm_extract_epi16(vsum, 0); + *sum += (int16_t)_mm_extract_epi16(vsum, 1); +} + +// Can handle 512 pixels' diff sum (such as 16x32 or 32x16) +static INLINE void variance_final_512_pel_sse2(__m128i vsse, __m128i vsum, + unsigned int *const sse, + int *const sum) { + *sse = add32x4_sse2(vsse); + + vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 8)); + vsum = _mm_unpacklo_epi16(vsum, vsum); + vsum = _mm_srai_epi32(vsum, 16); + *sum = add32x4_sse2(vsum); +} + +// Can handle 1024 pixels' diff sum (such as 32x32) +static INLINE void variance_final_1024_pel_sse2(__m128i vsse, __m128i vsum, + unsigned int *const sse, + int *const sum) { + *sse = add32x4_sse2(vsse); + + vsum = sum_to_32bit_sse2(vsum); + *sum = add32x4_sse2(vsum); +} + +static INLINE void variance4_sse2(const uint8_t *src, const int src_stride, + const uint8_t *ref, const int ref_stride, + const int h, __m128i *const sse, + __m128i *const sum) { + assert(h <= 256); // May overflow for larger height. + *sum = _mm_setzero_si128(); + + for (int i = 0; i < h; i += 2) { + const __m128i s = load4x2_sse2(src, src_stride); + const __m128i r = load4x2_sse2(ref, ref_stride); + + variance_kernel_sse2(s, r, sse, sum); + src += 2 * src_stride; + ref += 2 * ref_stride; + } +} + +static INLINE void variance8_sse2(const uint8_t *src, const int src_stride, + const uint8_t *ref, const int ref_stride, + const int h, __m128i *const sse, + __m128i *const sum) { + assert(h <= 128); // May overflow for larger height. + *sum = _mm_setzero_si128(); + for (int i = 0; i < h; i++) { + const __m128i s = load8_8to16_sse2(src); + const __m128i r = load8_8to16_sse2(ref); + + variance_kernel_sse2(s, r, sse, sum); + src += src_stride; + ref += ref_stride; + } +} + +static INLINE void variance16_kernel_sse2(const uint8_t *const src, + const uint8_t *const ref, + __m128i *const sse, + __m128i *const sum) { + const __m128i zero = _mm_setzero_si128(); + const __m128i s = _mm_loadu_si128((const __m128i *)src); + const __m128i r = _mm_loadu_si128((const __m128i *)ref); + const __m128i src0 = _mm_unpacklo_epi8(s, zero); + const __m128i ref0 = _mm_unpacklo_epi8(r, zero); + const __m128i src1 = _mm_unpackhi_epi8(s, zero); + const __m128i ref1 = _mm_unpackhi_epi8(r, zero); + + variance_kernel_sse2(src0, ref0, sse, sum); + variance_kernel_sse2(src1, ref1, sse, sum); +} + +static INLINE void variance16_sse2(const uint8_t *src, const int src_stride, + const uint8_t *ref, const int ref_stride, + const int h, __m128i *const sse, + __m128i *const sum) { + assert(h <= 64); // May overflow for larger height. + *sum = _mm_setzero_si128(); + + for (int i = 0; i < h; ++i) { + variance16_kernel_sse2(src, ref, sse, sum); + src += src_stride; + ref += ref_stride; + } +} + +static INLINE void variance32_sse2(const uint8_t *src, const int src_stride, + const uint8_t *ref, const int ref_stride, + const int h, __m128i *const sse, + __m128i *const sum) { + assert(h <= 32); // May overflow for larger height. + // Don't initialize sse here since it's an accumulation. + *sum = _mm_setzero_si128(); + + for (int i = 0; i < h; ++i) { + variance16_kernel_sse2(src + 0, ref + 0, sse, sum); + variance16_kernel_sse2(src + 16, ref + 16, sse, sum); + src += src_stride; + ref += ref_stride; + } +} + +static INLINE void variance64_sse2(const uint8_t *src, const int src_stride, + const uint8_t *ref, const int ref_stride, + const int h, __m128i *const sse, + __m128i *const sum) { + assert(h <= 16); // May overflow for larger height. + *sum = _mm_setzero_si128(); + + for (int i = 0; i < h; ++i) { + variance16_kernel_sse2(src + 0, ref + 0, sse, sum); + variance16_kernel_sse2(src + 16, ref + 16, sse, sum); + variance16_kernel_sse2(src + 32, ref + 32, sse, sum); + variance16_kernel_sse2(src + 48, ref + 48, sse, sum); + src += src_stride; + ref += ref_stride; + } +} + +static INLINE void variance128_sse2(const uint8_t *src, const int src_stride, + const uint8_t *ref, const int ref_stride, + const int h, __m128i *const sse, + __m128i *const sum) { + assert(h <= 8); // May overflow for larger height. + *sum = _mm_setzero_si128(); + + for (int i = 0; i < h; ++i) { + for (int j = 0; j < 4; ++j) { + const int offset0 = j << 5; + const int offset1 = offset0 + 16; + variance16_kernel_sse2(src + offset0, ref + offset0, sse, sum); + variance16_kernel_sse2(src + offset1, ref + offset1, sse, sum); + } + src += src_stride; + ref += ref_stride; + } +} + +#define AOM_VAR_NO_LOOP_SSE2(bw, bh, bits, max_pixels) \ + unsigned int aom_variance##bw##x##bh##_sse2( \ + const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, \ + unsigned int *sse) { \ + __m128i vsse = _mm_setzero_si128(); \ + __m128i vsum; \ + int sum = 0; \ + variance##bw##_sse2(src, src_stride, ref, ref_stride, bh, &vsse, &vsum); \ + variance_final_##max_pixels##_pel_sse2(vsse, vsum, sse, &sum); \ + assert(sum <= 255 * bw * bh); \ + assert(sum >= -255 * bw * bh); \ + return *sse - (uint32_t)(((int64_t)sum * sum) >> bits); \ + } + +AOM_VAR_NO_LOOP_SSE2(4, 4, 4, 128); +AOM_VAR_NO_LOOP_SSE2(4, 8, 5, 128); +AOM_VAR_NO_LOOP_SSE2(4, 16, 6, 128); + +AOM_VAR_NO_LOOP_SSE2(8, 4, 5, 128); +AOM_VAR_NO_LOOP_SSE2(8, 8, 6, 128); +AOM_VAR_NO_LOOP_SSE2(8, 16, 7, 128); +AOM_VAR_NO_LOOP_SSE2(8, 32, 8, 256); + +AOM_VAR_NO_LOOP_SSE2(16, 4, 6, 128); +AOM_VAR_NO_LOOP_SSE2(16, 8, 7, 128); +AOM_VAR_NO_LOOP_SSE2(16, 16, 8, 256); +AOM_VAR_NO_LOOP_SSE2(16, 32, 9, 512); +AOM_VAR_NO_LOOP_SSE2(16, 64, 10, 1024); + +AOM_VAR_NO_LOOP_SSE2(32, 8, 8, 256); +AOM_VAR_NO_LOOP_SSE2(32, 16, 9, 512); +AOM_VAR_NO_LOOP_SSE2(32, 32, 10, 1024); + +#define AOM_VAR_LOOP_SSE2(bw, bh, bits, uh) \ + unsigned int aom_variance##bw##x##bh##_sse2( \ + const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, \ + unsigned int *sse) { \ + __m128i vsse = _mm_setzero_si128(); \ + __m128i vsum = _mm_setzero_si128(); \ + for (int i = 0; i < (bh / uh); ++i) { \ + __m128i vsum16; \ + variance##bw##_sse2(src, src_stride, ref, ref_stride, uh, &vsse, \ + &vsum16); \ + vsum = _mm_add_epi32(vsum, sum_to_32bit_sse2(vsum16)); \ + src += (src_stride * uh); \ + ref += (ref_stride * uh); \ + } \ + *sse = add32x4_sse2(vsse); \ + int sum = add32x4_sse2(vsum); \ + assert(sum <= 255 * bw * bh); \ + assert(sum >= -255 * bw * bh); \ + return *sse - (uint32_t)(((int64_t)sum * sum) >> bits); \ + } + +AOM_VAR_LOOP_SSE2(32, 64, 11, 32); // 32x32 * ( 64/32 ) + +AOM_VAR_NO_LOOP_SSE2(64, 16, 10, 1024); +AOM_VAR_LOOP_SSE2(64, 32, 11, 16); // 64x16 * ( 32/16 ) +AOM_VAR_LOOP_SSE2(64, 64, 12, 16); // 64x16 * ( 64/16 ) +AOM_VAR_LOOP_SSE2(64, 128, 13, 16); // 64x16 * ( 128/16 ) + +AOM_VAR_LOOP_SSE2(128, 64, 13, 8); // 128x8 * ( 64/8 ) +AOM_VAR_LOOP_SSE2(128, 128, 14, 8); // 128x8 * ( 128/8 ) + +unsigned int aom_mse8x8_sse2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride, + unsigned int *sse) { + aom_variance8x8_sse2(src, src_stride, ref, ref_stride, sse); + return *sse; +} + +unsigned int aom_mse8x16_sse2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride, + unsigned int *sse) { + aom_variance8x16_sse2(src, src_stride, ref, ref_stride, sse); + return *sse; +} + +unsigned int aom_mse16x8_sse2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride, + unsigned int *sse) { + aom_variance16x8_sse2(src, src_stride, ref, ref_stride, sse); + return *sse; +} + +unsigned int aom_mse16x16_sse2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride, + unsigned int *sse) { + aom_variance16x16_sse2(src, src_stride, ref, ref_stride, sse); + return *sse; +} + +// The 2 unused parameters are place holders for PIC enabled build. +// These definitions are for functions defined in subpel_variance.asm +#define DECL(w, opt) \ + int aom_sub_pixel_variance##w##xh_##opt( \ + const uint8_t *src, ptrdiff_t src_stride, int x_offset, int y_offset, \ + const uint8_t *dst, ptrdiff_t dst_stride, int height, unsigned int *sse, \ + void *unused0, void *unused) +#define DECLS(opt) \ + DECL(4, opt); \ + DECL(8, opt); \ + DECL(16, opt) + +DECLS(sse2); +DECLS(ssse3); +#undef DECLS +#undef DECL + +#define FN(w, h, wf, wlog2, hlog2, opt, cast_prod, cast) \ + unsigned int aom_sub_pixel_variance##w##x##h##_##opt( \ + const uint8_t *src, int src_stride, int x_offset, int y_offset, \ + const uint8_t *dst, int dst_stride, unsigned int *sse_ptr) { \ + /*Avoid overflow in helper by capping height.*/ \ + const int hf = AOMMIN(h, 64); \ + unsigned int sse = 0; \ + int se = 0; \ + for (int i = 0; i < (w / wf); ++i) { \ + const uint8_t *src_ptr = src; \ + const uint8_t *dst_ptr = dst; \ + for (int j = 0; j < (h / hf); ++j) { \ + unsigned int sse2; \ + const int se2 = aom_sub_pixel_variance##wf##xh_##opt( \ + src_ptr, src_stride, x_offset, y_offset, dst_ptr, dst_stride, hf, \ + &sse2, NULL, NULL); \ + dst_ptr += hf * dst_stride; \ + src_ptr += hf * src_stride; \ + se += se2; \ + sse += sse2; \ + } \ + src += wf; \ + dst += wf; \ + } \ + *sse_ptr = sse; \ + return sse - (unsigned int)(cast_prod(cast se * se) >> (wlog2 + hlog2)); \ + } + +#define FNS(opt) \ + FN(128, 128, 16, 7, 7, opt, (int64_t), (int64_t)); \ + FN(128, 64, 16, 7, 6, opt, (int64_t), (int64_t)); \ + FN(64, 128, 16, 6, 7, opt, (int64_t), (int64_t)); \ + FN(64, 64, 16, 6, 6, opt, (int64_t), (int64_t)); \ + FN(64, 32, 16, 6, 5, opt, (int64_t), (int64_t)); \ + FN(32, 64, 16, 5, 6, opt, (int64_t), (int64_t)); \ + FN(32, 32, 16, 5, 5, opt, (int64_t), (int64_t)); \ + FN(32, 16, 16, 5, 4, opt, (int64_t), (int64_t)); \ + FN(16, 32, 16, 4, 5, opt, (int64_t), (int64_t)); \ + FN(16, 16, 16, 4, 4, opt, (uint32_t), (int64_t)); \ + FN(16, 8, 16, 4, 3, opt, (int32_t), (int32_t)); \ + FN(8, 16, 8, 3, 4, opt, (int32_t), (int32_t)); \ + FN(8, 8, 8, 3, 3, opt, (int32_t), (int32_t)); \ + FN(8, 4, 8, 3, 2, opt, (int32_t), (int32_t)); \ + FN(4, 8, 4, 2, 3, opt, (int32_t), (int32_t)); \ + FN(4, 4, 4, 2, 2, opt, (int32_t), (int32_t)); \ + FN(4, 16, 4, 2, 4, opt, (int32_t), (int32_t)); \ + FN(16, 4, 16, 4, 2, opt, (int32_t), (int32_t)); \ + FN(8, 32, 8, 3, 5, opt, (uint32_t), (int64_t)); \ + FN(32, 8, 16, 5, 3, opt, (uint32_t), (int64_t)); \ + FN(16, 64, 16, 4, 6, opt, (int64_t), (int64_t)); \ + FN(64, 16, 16, 6, 4, opt, (int64_t), (int64_t)) + +FNS(sse2); +FNS(ssse3); + +#undef FNS +#undef FN + +// The 2 unused parameters are place holders for PIC enabled build. +#define DECL(w, opt) \ + int aom_sub_pixel_avg_variance##w##xh_##opt( \ + const uint8_t *src, ptrdiff_t src_stride, int x_offset, int y_offset, \ + const uint8_t *dst, ptrdiff_t dst_stride, const uint8_t *sec, \ + ptrdiff_t sec_stride, int height, unsigned int *sse, void *unused0, \ + void *unused) +#define DECLS(opt) \ + DECL(4, opt); \ + DECL(8, opt); \ + DECL(16, opt) + +DECLS(sse2); +DECLS(ssse3); +#undef DECL +#undef DECLS + +#define FN(w, h, wf, wlog2, hlog2, opt, cast_prod, cast) \ + unsigned int aom_sub_pixel_avg_variance##w##x##h##_##opt( \ + const uint8_t *src, int src_stride, int x_offset, int y_offset, \ + const uint8_t *dst, int dst_stride, unsigned int *sse_ptr, \ + const uint8_t *sec) { \ + /*Avoid overflow in helper by capping height.*/ \ + const int hf = AOMMIN(h, 64); \ + unsigned int sse = 0; \ + int se = 0; \ + for (int i = 0; i < (w / wf); ++i) { \ + const uint8_t *src_ptr = src; \ + const uint8_t *dst_ptr = dst; \ + const uint8_t *sec_ptr = sec; \ + for (int j = 0; j < (h / hf); ++j) { \ + unsigned int sse2; \ + const int se2 = aom_sub_pixel_avg_variance##wf##xh_##opt( \ + src_ptr, src_stride, x_offset, y_offset, dst_ptr, dst_stride, \ + sec_ptr, w, hf, &sse2, NULL, NULL); \ + dst_ptr += hf * dst_stride; \ + src_ptr += hf * src_stride; \ + sec_ptr += hf * w; \ + se += se2; \ + sse += sse2; \ + } \ + src += wf; \ + dst += wf; \ + sec += wf; \ + } \ + *sse_ptr = sse; \ + return sse - (unsigned int)(cast_prod(cast se * se) >> (wlog2 + hlog2)); \ + } + +#define FNS(opt) \ + FN(128, 128, 16, 7, 7, opt, (int64_t), (int64_t)); \ + FN(128, 64, 16, 7, 6, opt, (int64_t), (int64_t)); \ + FN(64, 128, 16, 6, 7, opt, (int64_t), (int64_t)); \ + FN(64, 64, 16, 6, 6, opt, (int64_t), (int64_t)); \ + FN(64, 32, 16, 6, 5, opt, (int64_t), (int64_t)); \ + FN(32, 64, 16, 5, 6, opt, (int64_t), (int64_t)); \ + FN(32, 32, 16, 5, 5, opt, (int64_t), (int64_t)); \ + FN(32, 16, 16, 5, 4, opt, (int64_t), (int64_t)); \ + FN(16, 32, 16, 4, 5, opt, (int64_t), (int64_t)); \ + FN(16, 16, 16, 4, 4, opt, (uint32_t), (int64_t)); \ + FN(16, 8, 16, 4, 3, opt, (uint32_t), (int32_t)); \ + FN(8, 16, 8, 3, 4, opt, (uint32_t), (int32_t)); \ + FN(8, 8, 8, 3, 3, opt, (uint32_t), (int32_t)); \ + FN(8, 4, 8, 3, 2, opt, (uint32_t), (int32_t)); \ + FN(4, 8, 4, 2, 3, opt, (uint32_t), (int32_t)); \ + FN(4, 4, 4, 2, 2, opt, (uint32_t), (int32_t)); \ + FN(4, 16, 4, 2, 4, opt, (int32_t), (int32_t)); \ + FN(16, 4, 16, 4, 2, opt, (int32_t), (int32_t)); \ + FN(8, 32, 8, 3, 5, opt, (uint32_t), (int64_t)); \ + FN(32, 8, 16, 5, 3, opt, (uint32_t), (int64_t)); \ + FN(16, 64, 16, 4, 6, opt, (int64_t), (int64_t)); \ + FN(64, 16, 16, 6, 4, opt, (int64_t), (int64_t)) + +FNS(sse2); +FNS(ssse3); + +#undef FNS +#undef FN + +void aom_upsampled_pred_sse2(MACROBLOCKD *xd, const struct AV1Common *const cm, + int mi_row, int mi_col, const MV *const mv, + uint8_t *comp_pred, int width, int height, + int subpel_x_q3, int subpel_y_q3, + const uint8_t *ref, int ref_stride, + int subpel_search) { + // expect xd == NULL only in tests + if (xd != NULL) { + const MB_MODE_INFO *mi = xd->mi[0]; + const int ref_num = 0; + const int is_intrabc = is_intrabc_block(mi); + const struct scale_factors *const sf = + is_intrabc ? &cm->sf_identity : &xd->block_refs[ref_num]->sf; + const int is_scaled = av1_is_scaled(sf); + + if (is_scaled) { + // Note: This is mostly a copy from the >=8X8 case in + // build_inter_predictors() function, with some small tweaks. + + // Some assumptions. + const int plane = 0; + + // Get pre-requisites. + const struct macroblockd_plane *const pd = &xd->plane[plane]; + const int ssx = pd->subsampling_x; + const int ssy = pd->subsampling_y; + assert(ssx == 0 && ssy == 0); + const struct buf_2d *const dst_buf = &pd->dst; + const struct buf_2d *const pre_buf = + is_intrabc ? dst_buf : &pd->pre[ref_num]; + const int mi_x = mi_col * MI_SIZE; + const int mi_y = mi_row * MI_SIZE; + + // Calculate subpel_x/y and x/y_step. + const int row_start = 0; // Because ss_y is 0. + const int col_start = 0; // Because ss_x is 0. + const int pre_x = (mi_x + MI_SIZE * col_start) >> ssx; + const int pre_y = (mi_y + MI_SIZE * row_start) >> ssy; + int orig_pos_y = pre_y << SUBPEL_BITS; + orig_pos_y += mv->row * (1 << (1 - ssy)); + int orig_pos_x = pre_x << SUBPEL_BITS; + orig_pos_x += mv->col * (1 << (1 - ssx)); + int pos_y = sf->scale_value_y(orig_pos_y, sf); + int pos_x = sf->scale_value_x(orig_pos_x, sf); + pos_x += SCALE_EXTRA_OFF; + pos_y += SCALE_EXTRA_OFF; + + const int top = -AOM_LEFT_TOP_MARGIN_SCALED(ssy); + const int left = -AOM_LEFT_TOP_MARGIN_SCALED(ssx); + const int bottom = (pre_buf->height + AOM_INTERP_EXTEND) + << SCALE_SUBPEL_BITS; + const int right = (pre_buf->width + AOM_INTERP_EXTEND) + << SCALE_SUBPEL_BITS; + pos_y = clamp(pos_y, top, bottom); + pos_x = clamp(pos_x, left, right); + + const uint8_t *const pre = + pre_buf->buf0 + (pos_y >> SCALE_SUBPEL_BITS) * pre_buf->stride + + (pos_x >> SCALE_SUBPEL_BITS); + + const SubpelParams subpel_params = { sf->x_step_q4, sf->y_step_q4, + pos_x & SCALE_SUBPEL_MASK, + pos_y & SCALE_SUBPEL_MASK }; + + // Get warp types. + const WarpedMotionParams *const wm = + &xd->global_motion[mi->ref_frame[ref_num]]; + const int is_global = is_global_mv_block(mi, wm->wmtype); + WarpTypesAllowed warp_types; + warp_types.global_warp_allowed = is_global; + warp_types.local_warp_allowed = mi->motion_mode == WARPED_CAUSAL; + + // Get convolve parameters. + ConvolveParams conv_params = get_conv_params(0, plane, xd->bd); + const InterpFilters filters = + av1_broadcast_interp_filter(EIGHTTAP_REGULAR); + + // Get the inter predictor. + const int build_for_obmc = 0; + av1_make_inter_predictor(pre, pre_buf->stride, comp_pred, width, + &subpel_params, sf, width, height, &conv_params, + filters, &warp_types, mi_x >> pd->subsampling_x, + mi_y >> pd->subsampling_y, plane, ref_num, mi, + build_for_obmc, xd, cm->allow_warped_motion); + + return; + } + } + + const InterpFilterParams *filter = + (subpel_search == 1) + ? av1_get_4tap_interp_filter_params(EIGHTTAP_REGULAR) + : av1_get_interp_filter_params_with_block_size(EIGHTTAP_REGULAR, 8); + int filter_taps = (subpel_search == 1) ? 4 : SUBPEL_TAPS; + + if (!subpel_x_q3 && !subpel_y_q3) { + if (width >= 16) { + int i; + assert(!(width & 15)); + /*Read 16 pixels one row at a time.*/ + for (i = 0; i < height; i++) { + int j; + for (j = 0; j < width; j += 16) { + xx_storeu_128(comp_pred, xx_loadu_128(ref)); + comp_pred += 16; + ref += 16; + } + ref += ref_stride - width; + } + } else if (width >= 8) { + int i; + assert(!(width & 7)); + assert(!(height & 1)); + /*Read 8 pixels two rows at a time.*/ + for (i = 0; i < height; i += 2) { + __m128i s0 = xx_loadl_64(ref + 0 * ref_stride); + __m128i s1 = xx_loadl_64(ref + 1 * ref_stride); + xx_storeu_128(comp_pred, _mm_unpacklo_epi64(s0, s1)); + comp_pred += 16; + ref += 2 * ref_stride; + } + } else { + int i; + assert(!(width & 3)); + assert(!(height & 3)); + /*Read 4 pixels four rows at a time.*/ + for (i = 0; i < height; i++) { + const __m128i row0 = xx_loadl_64(ref + 0 * ref_stride); + const __m128i row1 = xx_loadl_64(ref + 1 * ref_stride); + const __m128i row2 = xx_loadl_64(ref + 2 * ref_stride); + const __m128i row3 = xx_loadl_64(ref + 3 * ref_stride); + const __m128i reg = _mm_unpacklo_epi64(_mm_unpacklo_epi32(row0, row1), + _mm_unpacklo_epi32(row2, row3)); + xx_storeu_128(comp_pred, reg); + comp_pred += 16; + ref += 4 * ref_stride; + } + } + } else if (!subpel_y_q3) { + const int16_t *const kernel = + av1_get_interp_filter_subpel_kernel(filter, subpel_x_q3 << 1); + aom_convolve8_horiz(ref, ref_stride, comp_pred, width, kernel, 16, NULL, -1, + width, height); + } else if (!subpel_x_q3) { + const int16_t *const kernel = + av1_get_interp_filter_subpel_kernel(filter, subpel_y_q3 << 1); + aom_convolve8_vert(ref, ref_stride, comp_pred, width, NULL, -1, kernel, 16, + width, height); + } else { + DECLARE_ALIGNED(16, uint8_t, + temp[((MAX_SB_SIZE * 2 + 16) + 16) * MAX_SB_SIZE]); + const int16_t *const kernel_x = + av1_get_interp_filter_subpel_kernel(filter, subpel_x_q3 << 1); + const int16_t *const kernel_y = + av1_get_interp_filter_subpel_kernel(filter, subpel_y_q3 << 1); + const uint8_t *ref_start = ref - ref_stride * ((filter_taps >> 1) - 1); + uint8_t *temp_start_horiz = + (subpel_search == 1) ? temp + (filter_taps >> 1) * MAX_SB_SIZE : temp; + uint8_t *temp_start_vert = temp + MAX_SB_SIZE * ((filter->taps >> 1) - 1); + int intermediate_height = + (((height - 1) * 8 + subpel_y_q3) >> 3) + filter_taps; + assert(intermediate_height <= (MAX_SB_SIZE * 2 + 16) + 16); + // TODO(Deepa): Remove the memset below when we have + // 4 tap simd for sse2 and ssse3. + if (subpel_search == 1) { + memset(temp_start_vert - 3 * MAX_SB_SIZE, 0, width); + memset(temp_start_vert - 2 * MAX_SB_SIZE, 0, width); + memset(temp_start_vert + (height + 2) * MAX_SB_SIZE, 0, width); + memset(temp_start_vert + (height + 3) * MAX_SB_SIZE, 0, width); + } + aom_convolve8_horiz(ref_start, ref_stride, temp_start_horiz, MAX_SB_SIZE, + kernel_x, 16, NULL, -1, width, intermediate_height); + aom_convolve8_vert(temp_start_vert, MAX_SB_SIZE, comp_pred, width, NULL, -1, + kernel_y, 16, width, height); + } +} + +void aom_comp_avg_upsampled_pred_sse2( + MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col, + const MV *const mv, uint8_t *comp_pred, const uint8_t *pred, int width, + int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref, + int ref_stride, int subpel_search) { + int n; + int i; + aom_upsampled_pred(xd, cm, mi_row, mi_col, mv, comp_pred, width, height, + subpel_x_q3, subpel_y_q3, ref, ref_stride, subpel_search); + /*The total number of pixels must be a multiple of 16 (e.g., 4x4).*/ + assert(!(width * height & 15)); + n = width * height >> 4; + for (i = 0; i < n; i++) { + __m128i s0 = xx_loadu_128(comp_pred); + __m128i p0 = xx_loadu_128(pred); + xx_storeu_128(comp_pred, _mm_avg_epu8(s0, p0)); + comp_pred += 16; + pred += 16; + } +} + +void aom_comp_mask_upsampled_pred_sse2( + MACROBLOCKD *xd, const AV1_COMMON *const cm, int mi_row, int mi_col, + const MV *const mv, uint8_t *comp_pred, const uint8_t *pred, int width, + int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref, + int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask, + int subpel_search) { + if (subpel_x_q3 | subpel_y_q3) { + aom_upsampled_pred(xd, cm, mi_row, mi_col, mv, comp_pred, width, height, + subpel_x_q3, subpel_y_q3, ref, ref_stride, + subpel_search); + ref = comp_pred; + ref_stride = width; + } + aom_comp_mask_pred(comp_pred, pred, width, height, ref, ref_stride, mask, + mask_stride, invert_mask); +} + +static INLINE __m128i highbd_comp_mask_pred_line_sse2(const __m128i s0, + const __m128i s1, + const __m128i a) { + const __m128i alpha_max = _mm_set1_epi16((1 << AOM_BLEND_A64_ROUND_BITS)); + const __m128i round_const = + _mm_set1_epi32((1 << AOM_BLEND_A64_ROUND_BITS) >> 1); + const __m128i a_inv = _mm_sub_epi16(alpha_max, a); + + const __m128i s_lo = _mm_unpacklo_epi16(s0, s1); + const __m128i a_lo = _mm_unpacklo_epi16(a, a_inv); + const __m128i pred_lo = _mm_madd_epi16(s_lo, a_lo); + const __m128i pred_l = _mm_srai_epi32(_mm_add_epi32(pred_lo, round_const), + AOM_BLEND_A64_ROUND_BITS); + + const __m128i s_hi = _mm_unpackhi_epi16(s0, s1); + const __m128i a_hi = _mm_unpackhi_epi16(a, a_inv); + const __m128i pred_hi = _mm_madd_epi16(s_hi, a_hi); + const __m128i pred_h = _mm_srai_epi32(_mm_add_epi32(pred_hi, round_const), + AOM_BLEND_A64_ROUND_BITS); + + const __m128i comp = _mm_packs_epi32(pred_l, pred_h); + + return comp; +} + +void aom_highbd_comp_mask_pred_sse2(uint8_t *comp_pred8, const uint8_t *pred8, + int width, int height, const uint8_t *ref8, + int ref_stride, const uint8_t *mask, + int mask_stride, int invert_mask) { + int i = 0; + uint16_t *comp_pred = CONVERT_TO_SHORTPTR(comp_pred8); + uint16_t *pred = CONVERT_TO_SHORTPTR(pred8); + uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); + const uint16_t *src0 = invert_mask ? pred : ref; + const uint16_t *src1 = invert_mask ? ref : pred; + const int stride0 = invert_mask ? width : ref_stride; + const int stride1 = invert_mask ? ref_stride : width; + const __m128i zero = _mm_setzero_si128(); + + if (width == 8) { + do { + const __m128i s0 = _mm_loadu_si128((const __m128i *)(src0)); + const __m128i s1 = _mm_loadu_si128((const __m128i *)(src1)); + const __m128i m_8 = _mm_loadl_epi64((const __m128i *)mask); + const __m128i m_16 = _mm_unpacklo_epi8(m_8, zero); + + const __m128i comp = highbd_comp_mask_pred_line_sse2(s0, s1, m_16); + + _mm_storeu_si128((__m128i *)comp_pred, comp); + + src0 += stride0; + src1 += stride1; + mask += mask_stride; + comp_pred += width; + i += 1; + } while (i < height); + } else if (width == 16) { + do { + const __m128i s0 = _mm_loadu_si128((const __m128i *)(src0)); + const __m128i s2 = _mm_loadu_si128((const __m128i *)(src0 + 8)); + const __m128i s1 = _mm_loadu_si128((const __m128i *)(src1)); + const __m128i s3 = _mm_loadu_si128((const __m128i *)(src1 + 8)); + + const __m128i m_8 = _mm_loadu_si128((const __m128i *)mask); + const __m128i m01_16 = _mm_unpacklo_epi8(m_8, zero); + const __m128i m23_16 = _mm_unpackhi_epi8(m_8, zero); + + const __m128i comp = highbd_comp_mask_pred_line_sse2(s0, s1, m01_16); + const __m128i comp1 = highbd_comp_mask_pred_line_sse2(s2, s3, m23_16); + + _mm_storeu_si128((__m128i *)comp_pred, comp); + _mm_storeu_si128((__m128i *)(comp_pred + 8), comp1); + + src0 += stride0; + src1 += stride1; + mask += mask_stride; + comp_pred += width; + i += 1; + } while (i < height); + } else if (width == 32) { + do { + for (int j = 0; j < 2; j++) { + const __m128i s0 = _mm_loadu_si128((const __m128i *)(src0 + j * 16)); + const __m128i s2 = + _mm_loadu_si128((const __m128i *)(src0 + 8 + j * 16)); + const __m128i s1 = _mm_loadu_si128((const __m128i *)(src1 + j * 16)); + const __m128i s3 = + _mm_loadu_si128((const __m128i *)(src1 + 8 + j * 16)); + + const __m128i m_8 = _mm_loadu_si128((const __m128i *)(mask + j * 16)); + const __m128i m01_16 = _mm_unpacklo_epi8(m_8, zero); + const __m128i m23_16 = _mm_unpackhi_epi8(m_8, zero); + + const __m128i comp = highbd_comp_mask_pred_line_sse2(s0, s1, m01_16); + const __m128i comp1 = highbd_comp_mask_pred_line_sse2(s2, s3, m23_16); + + _mm_storeu_si128((__m128i *)(comp_pred + j * 16), comp); + _mm_storeu_si128((__m128i *)(comp_pred + 8 + j * 16), comp1); + } + src0 += stride0; + src1 += stride1; + mask += mask_stride; + comp_pred += width; + i += 1; + } while (i < height); + } +} -- cgit v1.2.3