Adding upstream version 1:115.7.0.upstream/1%115.7.0upstream

Signed-off-by: Daniel Baumann <daniel.baumann@progress-linux.org>

36 files changed, 14362 insertions, 0 deletions

diff --git a/media/libjpeg/simd/x86_64/jccolext-avx2.asm b/media/libjpeg/simd/x86_64/jccolext-avx2.asm
new file mode 100644
index 0000000000..ffb527db00
--- /dev/null
+++ b/media/libjpeg/simd/x86_64/jccolext-avx2.asm
@@ -0,0 +1,559 @@
+;
+; jccolext.asm - colorspace conversion (64-bit AVX2)
+;
+; Copyright (C) 2009, 2016, D. R. Commander.
+; Copyright (C) 2015, Intel Corporation.
+; Copyright (C) 2018, Matthias Räncker.
+;
+; Based on the x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+
+%include "jcolsamp.inc"
+
+; --------------------------------------------------------------------------
+;
+; Convert some rows of samples to the output colorspace.
+;
+; GLOBAL(void)
+; jsimd_rgb_ycc_convert_avx2(JDIMENSION img_width, JSAMPARRAY input_buf,
+;                            JSAMPIMAGE output_buf, JDIMENSION output_row,
+;                            int num_rows);
+;
+
+; r10d = JDIMENSION img_width
+; r11 = JSAMPARRAY input_buf
+; r12 = JSAMPIMAGE output_buf
+; r13d = JDIMENSION output_row
+; r14d = int num_rows
+
+%define wk(i)   rbp - (WK_NUM - (i)) * SIZEOF_YMMWORD  ; ymmword wk[WK_NUM]
+%define WK_NUM  8
+
+    align       32
+    GLOBAL_FUNCTION(jsimd_rgb_ycc_convert_avx2)
+
+EXTN(jsimd_rgb_ycc_convert_avx2):
+    push        rbp
+    mov         rax, rsp                     ; rax = original rbp
+    sub         rsp, byte 4
+    and         rsp, byte (-SIZEOF_YMMWORD)  ; align to 256 bits
+    mov         [rsp], rax
+    mov         rbp, rsp                     ; rbp = aligned rbp
+    lea         rsp, [wk(0)]
+    collect_args 5
+    push        rbx
+
+    mov         ecx, r10d
+    test        rcx, rcx
+    jz          near .return
+
+    push        rcx
+
+    mov         rsi, r12
+    mov         ecx, r13d
+    mov         rdip, JSAMPARRAY [rsi+0*SIZEOF_JSAMPARRAY]
+    mov         rbxp, JSAMPARRAY [rsi+1*SIZEOF_JSAMPARRAY]
+    mov         rdxp, JSAMPARRAY [rsi+2*SIZEOF_JSAMPARRAY]
+    lea         rdi, [rdi+rcx*SIZEOF_JSAMPROW]
+    lea         rbx, [rbx+rcx*SIZEOF_JSAMPROW]
+    lea         rdx, [rdx+rcx*SIZEOF_JSAMPROW]
+
+    pop         rcx
+
+    mov         rsi, r11
+    mov         eax, r14d
+    test        rax, rax
+    jle         near .return
+.rowloop:
+    push        rdx
+    push        rbx
+    push        rdi
+    push        rsi
+    push        rcx                     ; col
+
+    mov         rsip, JSAMPROW [rsi]    ; inptr
+    mov         rdip, JSAMPROW [rdi]    ; outptr0
+    mov         rbxp, JSAMPROW [rbx]    ; outptr1
+    mov         rdxp, JSAMPROW [rdx]    ; outptr2
+
+    cmp         rcx, byte SIZEOF_YMMWORD
+    jae         near .columnloop
+
+%if RGB_PIXELSIZE == 3  ; ---------------
+
+.column_ld1:
+    push        rax
+    push        rdx
+    lea         rcx, [rcx+rcx*2]        ; imul ecx,RGB_PIXELSIZE
+    test        cl, SIZEOF_BYTE
+    jz          short .column_ld2
+    sub         rcx, byte SIZEOF_BYTE
+    movzx       rax, byte [rsi+rcx]
+.column_ld2:
+    test        cl, SIZEOF_WORD
+    jz          short .column_ld4
+    sub         rcx, byte SIZEOF_WORD
+    movzx       rdx, word [rsi+rcx]
+    shl         rax, WORD_BIT
+    or          rax, rdx
+.column_ld4:
+    vmovd       xmmA, eax
+    pop         rdx
+    pop         rax
+    test        cl, SIZEOF_DWORD
+    jz          short .column_ld8
+    sub         rcx, byte SIZEOF_DWORD
+    vmovd       xmmF, XMM_DWORD [rsi+rcx]
+    vpslldq     xmmA, xmmA, SIZEOF_DWORD
+    vpor        xmmA, xmmA, xmmF
+.column_ld8:
+    test        cl, SIZEOF_MMWORD
+    jz          short .column_ld16
+    sub         rcx, byte SIZEOF_MMWORD
+    vmovq       xmmB, XMM_MMWORD [rsi+rcx]
+    vpslldq     xmmA, xmmA, SIZEOF_MMWORD
+    vpor        xmmA, xmmA, xmmB
+.column_ld16:
+    test        cl, SIZEOF_XMMWORD
+    jz          short .column_ld32
+    sub         rcx, byte SIZEOF_XMMWORD
+    vmovdqu     xmmB, XMM_MMWORD [rsi+rcx]
+    vperm2i128  ymmA, ymmA, ymmA, 1
+    vpor        ymmA, ymmB
+.column_ld32:
+    test        cl, SIZEOF_YMMWORD
+    jz          short .column_ld64
+    sub         rcx, byte SIZEOF_YMMWORD
+    vmovdqa     ymmF, ymmA
+    vmovdqu     ymmA, YMMWORD [rsi+0*SIZEOF_YMMWORD]
+.column_ld64:
+    test        cl, 2*SIZEOF_YMMWORD
+    mov         rcx, SIZEOF_YMMWORD
+    jz          short .rgb_ycc_cnv
+    vmovdqa     ymmB, ymmA
+    vmovdqu     ymmA, YMMWORD [rsi+0*SIZEOF_YMMWORD]
+    vmovdqu     ymmF, YMMWORD [rsi+1*SIZEOF_YMMWORD]
+    jmp         short .rgb_ycc_cnv
+
+.columnloop:
+    vmovdqu     ymmA, YMMWORD [rsi+0*SIZEOF_YMMWORD]
+    vmovdqu     ymmF, YMMWORD [rsi+1*SIZEOF_YMMWORD]
+    vmovdqu     ymmB, YMMWORD [rsi+2*SIZEOF_YMMWORD]
+
+.rgb_ycc_cnv:
+    ; ymmA=(00 10 20 01 11 21 02 12 22 03 13 23 04 14 24 05
+    ;       15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A)
+    ; ymmF=(2A 0B 1B 2B 0C 1C 2C 0D 1D 2D 0E 1E 2E 0F 1F 2F
+    ;       0G 1G 2G 0H 1H 2H 0I 1I 2I 0J 1J 2J 0K 1K 2K 0L)
+    ; ymmB=(1L 2L 0M 1M 2M 0N 1N 2N 0O 1O 2O 0P 1P 2P 0Q 1Q
+    ;       2Q 0R 1R 2R 0S 1S 2S 0T 1T 2T 0U 1U 2U 0V 1V 2V)
+
+    vmovdqu     ymmC, ymmA
+    vinserti128 ymmA, ymmF, xmmA, 0  ; ymmA=(00 10 20 01 11 21 02 12 22 03 13 23 04 14 24 05
+                                     ;       0G 1G 2G 0H 1H 2H 0I 1I 2I 0J 1J 2J 0K 1K 2K 0L)
+    vinserti128 ymmC, ymmC, xmmB, 0  ; ymmC=(1L 2L 0M 1M 2M 0N 1N 2N 0O 1O 2O 0P 1P 2P 0Q 1Q
+                                     ;       15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A)
+    vinserti128 ymmB, ymmB, xmmF, 0  ; ymmB=(2A 0B 1B 2B 0C 1C 2C 0D 1D 2D 0E 1E 2E 0F 1F 2F
+                                     ;       2Q 0R 1R 2R 0S 1S 2S 0T 1T 2T 0U 1U 2U 0V 1V 2V)
+    vperm2i128  ymmF, ymmC, ymmC, 1  ; ymmF=(15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A
+                                     ;       1L 2L 0M 1M 2M 0N 1N 2N 0O 1O 2O 0P 1P 2P 0Q 1Q)
+
+    vmovdqa     ymmG, ymmA
+    vpslldq     ymmA, ymmA, 8     ; ymmA=(-- -- -- -- -- -- -- -- 00 10 20 01 11 21 02 12
+                                  ;       22 03 13 23 04 14 24 05 0G 1G 2G 0H 1H 2H 0I 1I)
+    vpsrldq     ymmG, ymmG, 8     ; ymmG=(22 03 13 23 04 14 24 05 0G 1G 2G 0H 1H 2H 0I 1I
+                                  ;       2I 0J 1J 2J 0K 1K 2K 0L -- -- -- -- -- -- -- --)
+
+    vpunpckhbw  ymmA, ymmA, ymmF  ; ymmA=(00 08 10 18 20 28 01 09 11 19 21 29 02 0A 12 1A
+                                  ;       0G 0O 1G 1O 2G 2O 0H 0P 1H 1P 2H 2P 0I 0Q 1I 1Q)
+    vpslldq     ymmF, ymmF, 8     ; ymmF=(-- -- -- -- -- -- -- -- 15 25 06 16 26 07 17 27
+                                  ;       08 18 28 09 19 29 0A 1A 1L 2L 0M 1M 2M 0N 1N 2N)
+
+    vpunpcklbw  ymmG, ymmG, ymmB  ; ymmG=(22 2A 03 0B 13 1B 23 2B 04 0C 14 1C 24 2C 05 0D
+                                  ;       2I 2Q 0J 0R 1J 1R 2J 2R 0K 0S 1K 1S 2K 2S 0L 0T)
+    vpunpckhbw  ymmF, ymmF, ymmB  ; ymmF=(15 1D 25 2D 06 0E 16 1E 26 2E 07 0F 17 1F 27 2F
+                                  ;       1L 1T 2L 2T 0M 0U 1M 1U 2M 2U 0N 0V 1N 1V 2N 2V)
+
+    vmovdqa     ymmD, ymmA
+    vpslldq     ymmA, ymmA, 8     ; ymmA=(-- -- -- -- -- -- -- -- 00 08 10 18 20 28 01 09
+                                  ;       11 19 21 29 02 0A 12 1A 0G 0O 1G 1O 2G 2O 0H 0P)
+    vpsrldq     ymmD, ymmD, 8     ; ymmD=(11 19 21 29 02 0A 12 1A 0G 0O 1G 1O 2G 2O 0H 0P
+                                  ;       1H 1P 2H 2P 0I 0Q 1I 1Q -- -- -- -- -- -- -- --)
+
+    vpunpckhbw  ymmA, ymmA, ymmG  ; ymmA=(00 04 08 0C 10 14 18 1C 20 24 28 2C 01 05 09 0D
+                                  ;       0G 0K 0O 0S 1G 1K 1O 1S 2G 2K 2O 2S 0H 0L 0P 0T)
+    vpslldq     ymmG, ymmG, 8     ; ymmG=(-- -- -- -- -- -- -- -- 22 2A 03 0B 13 1B 23 2B
+                                  ;       04 0C 14 1C 24 2C 05 0D 2I 2Q 0J 0R 1J 1R 2J 2R)
+
+    vpunpcklbw  ymmD, ymmD, ymmF  ; ymmD=(11 15 19 1D 21 25 29 2D 02 06 0A 0E 12 16 1A 1E
+                                  ;       1H 1L 1P 1T 2H 2L 2P 2T 0I 0M 0Q 0U 1I 1M 1Q 1U)
+    vpunpckhbw  ymmG, ymmG, ymmF  ; ymmG=(22 26 2A 2E 03 07 0B 0F 13 17 1B 1F 23 27 2B 2F
+                                  ;       2I 2M 2Q 2U 0J 0N 0R 0V 1J 1N 1R 1V 2J 2N 2R 2V)
+
+    vmovdqa     ymmE, ymmA
+    vpslldq     ymmA, ymmA, 8     ; ymmA=(-- -- -- -- -- -- -- -- 00 04 08 0C 10 14 18 1C
+                                  ;       20 24 28 2C 01 05 09 0D 0G 0K 0O 0S 1G 1K 1O 1S)
+    vpsrldq     ymmE, ymmE, 8     ; ymmE=(20 24 28 2C 01 05 09 0D 0G 0K 0O 0S 1G 1K 1O 1S
+                                  ;       2G 2K 2O 2S 0H 0L 0P 0T -- -- -- -- -- -- -- --)
+
+    vpunpckhbw  ymmA, ymmA, ymmD  ; ymmA=(00 02 04 06 08 0A 0C 0E 10 12 14 16 18 1A 1C 1E
+                                  ;       0G 0I 0K 0M 0O 0Q 0S 0U 1G 1I 1K 1M 1O 1Q 1S 1U)
+    vpslldq     ymmD, ymmD, 8     ; ymmD=(-- -- -- -- -- -- -- -- 11 15 19 1D 21 25 29 2D
+                                  ;       02 06 0A 0E 12 16 1A 1E 1H 1L 1P 1T 2H 2L 2P 2T)
+
+    vpunpcklbw  ymmE, ymmE, ymmG  ; ymmE=(20 22 24 26 28 2A 2C 2E 01 03 05 07 09 0B 0D 0F
+                                  ;       2G 2I 2K 2M 2O 2Q 2S 2U 0H 0J 0L 0N 0P 0R 0T 0V)
+    vpunpckhbw  ymmD, ymmD, ymmG  ; ymmD=(11 13 15 17 19 1B 1D 1F 21 23 25 27 29 2B 2D 2F
+                                  ;       1H 1J 1L 1N 1P 1R 1T 1V 2H 2J 2L 2N 2P 2R 2T 2V)
+
+    vpxor       ymmH, ymmH, ymmH
+
+    vmovdqa     ymmC, ymmA
+    vpunpcklbw  ymmA, ymmA, ymmH  ; ymmA=(00 02 04 06 08 0A 0C 0E 0G 0I 0K 0M 0O 0Q 0S 0U)
+    vpunpckhbw  ymmC, ymmC, ymmH  ; ymmC=(10 12 14 16 18 1A 1C 1E 1G 1I 1K 1M 1O 1Q 1S 1U)
+
+    vmovdqa     ymmB, ymmE
+    vpunpcklbw  ymmE, ymmE, ymmH  ; ymmE=(20 22 24 26 28 2A 2C 2E 2G 2I 2K 2M 2O 2Q 2S 2U)
+    vpunpckhbw  ymmB, ymmB, ymmH  ; ymmB=(01 03 05 07 09 0B 0D 0F 0H 0J 0L 0N 0P 0R 0T 0V)
+
+    vmovdqa     ymmF, ymmD
+    vpunpcklbw  ymmD, ymmD, ymmH  ; ymmD=(11 13 15 17 19 1B 1D 1F 1H 1J 1L 1N 1P 1R 1T 1V)
+    vpunpckhbw  ymmF, ymmF, ymmH  ; ymmF=(21 23 25 27 29 2B 2D 2F 2H 2J 2L 2N 2P 2R 2T 2V)
+
+%else  ; RGB_PIXELSIZE == 4 ; -----------
+
+.column_ld1:
+    test        cl, SIZEOF_XMMWORD/16
+    jz          short .column_ld2
+    sub         rcx, byte SIZEOF_XMMWORD/16
+    vmovd       xmmA, XMM_DWORD [rsi+rcx*RGB_PIXELSIZE]
+.column_ld2:
+    test        cl, SIZEOF_XMMWORD/8
+    jz          short .column_ld4
+    sub         rcx, byte SIZEOF_XMMWORD/8
+    vmovq       xmmF, XMM_MMWORD [rsi+rcx*RGB_PIXELSIZE]
+    vpslldq     xmmA, xmmA, SIZEOF_MMWORD
+    vpor        xmmA, xmmA, xmmF
+.column_ld4:
+    test        cl, SIZEOF_XMMWORD/4
+    jz          short .column_ld8
+    sub         rcx, byte SIZEOF_XMMWORD/4
+    vmovdqa     xmmF, xmmA
+    vperm2i128  ymmF, ymmF, ymmF, 1
+    vmovdqu     xmmA, XMMWORD [rsi+rcx*RGB_PIXELSIZE]
+    vpor        ymmA, ymmA, ymmF
+.column_ld8:
+    test        cl, SIZEOF_XMMWORD/2
+    jz          short .column_ld16
+    sub         rcx, byte SIZEOF_XMMWORD/2
+    vmovdqa     ymmF, ymmA
+    vmovdqu     ymmA, YMMWORD [rsi+rcx*RGB_PIXELSIZE]
+.column_ld16:
+    test        cl, SIZEOF_XMMWORD
+    mov         rcx, SIZEOF_YMMWORD
+    jz          short .rgb_ycc_cnv
+    vmovdqa     ymmE, ymmA
+    vmovdqa     ymmH, ymmF
+    vmovdqu     ymmA, YMMWORD [rsi+0*SIZEOF_YMMWORD]
+    vmovdqu     ymmF, YMMWORD [rsi+1*SIZEOF_YMMWORD]
+    jmp         short .rgb_ycc_cnv
+
+.columnloop:
+    vmovdqu     ymmA, YMMWORD [rsi+0*SIZEOF_YMMWORD]
+    vmovdqu     ymmF, YMMWORD [rsi+1*SIZEOF_YMMWORD]
+    vmovdqu     ymmE, YMMWORD [rsi+2*SIZEOF_YMMWORD]
+    vmovdqu     ymmH, YMMWORD [rsi+3*SIZEOF_YMMWORD]
+
+.rgb_ycc_cnv:
+    ; ymmA=(00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33
+    ;       04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37)
+    ; ymmF=(08 18 28 38 09 19 29 39 0A 1A 2A 3A 0B 1B 2B 3B
+    ;       0C 1C 2C 3C 0D 1D 2D 3D 0E 1E 2E 3E 0F 1F 2F 3F)
+    ; ymmE=(0G 1G 2G 3G 0H 1H 2H 3H 0I 1I 2I 3I 0J 1J 2J 3J
+    ;       0K 1K 2K 3K 0L 1L 2L 3L 0M 1M 2M 3M 0N 1N 2N 3N)
+    ; ymmH=(0O 1O 2O 3O 0P 1P 2P 3P 0Q 1Q 2Q 3Q 0R 1R 2R 3R
+    ;       0S 1S 2S 3S 0T 1T 2T 3T 0U 1U 2U 3U 0V 1V 2V 3V)
+
+    vmovdqa     ymmB, ymmA
+    vinserti128 ymmA, ymmA, xmmE, 1     ; ymmA=(00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33
+                                        ;       0G 1G 2G 3G 0H 1H 2H 3H 0I 1I 2I 3I 0J 1J 2J 3J)
+    vperm2i128  ymmE, ymmB, ymmE, 0x31  ; ymmE=(04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37
+                                        ;       0K 1K 2K 3K 0L 1L 2L 3L 0M 1M 2M 3M 0N 1N 2N 3N)
+
+    vmovdqa     ymmB, ymmF
+    vinserti128 ymmF, ymmF, xmmH, 1     ; ymmF=(08 18 28 38 09 19 29 39 0A 1A 2A 3A 0B 1B 2B 3B
+                                        ;       0O 1O 2O 3O 0P 1P 2P 3P 0Q 1Q 2Q 3Q 0R 1R 2R 3R)
+    vperm2i128  ymmH, ymmB, ymmH, 0x31  ; ymmH=(0C 1C 2C 3C 0D 1D 2D 3D 0E 1E 2E 3E 0F 1F 2F 3F
+                                        ;       0S 1S 2S 3S 0T 1T 2T 3T 0U 1U 2U 3U 0V 1V 2V 3V)
+
+    vmovdqa     ymmD, ymmA
+    vpunpcklbw  ymmA, ymmA, ymmE      ; ymmA=(00 04 10 14 20 24 30 34 01 05 11 15 21 25 31 35
+                                      ;       0G 0K 1G 1K 2G 2K 3G 3K 0H 0L 1H 1L 2H 2L 3H 3L)
+    vpunpckhbw  ymmD, ymmD, ymmE      ; ymmD=(02 06 12 16 22 26 32 36 03 07 13 17 23 27 33 37
+                                      ;       0I 0M 1I 1M 2I 2M 3I 3M 0J 0N 1J 1N 2J 2N 3J 3N)
+
+    vmovdqa     ymmC, ymmF
+    vpunpcklbw  ymmF, ymmF, ymmH      ; ymmF=(08 0C 18 1C 28 2C 38 3C 09 0D 19 1D 29 2D 39 3D
+                                      ;       0O 0S 1O 1S 2O 2S 3O 3S 0P 0T 1P 1T 2P 2T 3P 3T)
+    vpunpckhbw  ymmC, ymmC, ymmH      ; ymmC=(0A 0E 1A 1E 2A 2E 3A 3E 0B 0F 1B 1F 2B 2F 3B 3F
+                                      ;       0Q 0U 1Q 1U 2Q 2U 3Q 3U 0R 0V 1R 1V 2R 2V 3R 3V)
+
+    vmovdqa     ymmB, ymmA
+    vpunpcklwd  ymmA, ymmA, ymmF      ; ymmA=(00 04 08 0C 10 14 18 1C 20 24 28 2C 30 34 38 3C
+                                      ;       0G 0K 0O 0S 1G 1K 1O 1S 2G 2K 2O 2S 3G 3K 3O 3S)
+    vpunpckhwd  ymmB, ymmB, ymmF      ; ymmB=(01 05 09 0D 11 15 19 1D 21 25 29 2D 31 35 39 3D
+                                      ;       0H 0L 0P 0T 1H 1L 1P 1T 2H 2L 2P 2T 3H 3L 3P 3T)
+
+    vmovdqa     ymmG, ymmD
+    vpunpcklwd  ymmD, ymmD, ymmC      ; ymmD=(02 06 0A 0E 12 16 1A 1E 22 26 2A 2E 32 36 3A 3E
+                                      ;       0I 0M 0Q 0U 1I 1M 1Q 1U 2I 2M 2Q 2U 3I 3M 3Q 3U)
+    vpunpckhwd  ymmG, ymmG, ymmC      ; ymmG=(03 07 0B 0F 13 17 1B 1F 23 27 2B 2F 33 37 3B 3F
+                                      ;       0J 0N 0R 0V 1J 1N 1R 1V 2J 2N 2R 2V 3J 3N 3R 3V)
+
+    vmovdqa     ymmE, ymmA
+    vpunpcklbw  ymmA, ymmA, ymmD      ; ymmA=(00 02 04 06 08 0A 0C 0E 10 12 14 16 18 1A 1C 1E
+                                      ;       0G 0I 0K 0M 0O 0Q 0S 0U 1G 1I 1K 1M 1O 1Q 1S 1U)
+    vpunpckhbw  ymmE, ymmE, ymmD      ; ymmE=(20 22 24 26 28 2A 2C 2E 30 32 34 36 38 3A 3C 3E
+                                      ;       2G 2I 2K 2M 2O 2Q 2S 2U 3G 3I 3K 3M 3O 3Q 3S 3U)
+
+    vmovdqa     ymmH, ymmB
+    vpunpcklbw  ymmB, ymmB, ymmG      ; ymmB=(01 03 05 07 09 0B 0D 0F 11 13 15 17 19 1B 1D 1F
+                                      ;       0H 0J 0L 0N 0P 0R 0T 0V 1H 1J 1L 1N 1P 1R 1T 1V)
+    vpunpckhbw  ymmH, ymmH, ymmG      ; ymmH=(21 23 25 27 29 2B 2D 2F 31 33 35 37 39 3B 3D 3F
+                                      ;       2H 2J 2L 2N 2P 2R 2T 2V 3H 3J 3L 3N 3P 3R 3T 3V)
+
+    vpxor       ymmF, ymmF, ymmF
+
+    vmovdqa     ymmC, ymmA
+    vpunpcklbw  ymmA, ymmA, ymmF      ; ymmA=(00 02 04 06 08 0A 0C 0E 0G 0I 0K 0M 0O 0Q 0S 0U)
+    vpunpckhbw  ymmC, ymmC, ymmF      ; ymmC=(10 12 14 16 18 1A 1C 1E 1G 1I 1K 1M 1O 1Q 1S 1U)
+
+    vmovdqa     ymmD, ymmB
+    vpunpcklbw  ymmB, ymmB, ymmF      ; ymmB=(01 03 05 07 09 0B 0D 0F 0H 0J 0L 0N 0P 0R 0T 0V)
+    vpunpckhbw  ymmD, ymmD, ymmF      ; ymmD=(11 13 15 17 19 1B 1D 1F 1H 1J 1L 1N 1P 1R 1T 1V)
+
+    vmovdqa     ymmG, ymmE
+    vpunpcklbw  ymmE, ymmE, ymmF      ; ymmE=(20 22 24 26 28 2A 2C 2E 2G 2I 2K 2M 2O 2Q 2S 2U)
+    vpunpckhbw  ymmG, ymmG, ymmF      ; ymmG=(30 32 34 36 38 3A 3C 3E 3G 3I 3K 3M 3O 3Q 3S 3U)
+
+    vpunpcklbw  ymmF, ymmF, ymmH
+    vpunpckhbw  ymmH, ymmH, ymmH
+    vpsrlw      ymmF, ymmF, BYTE_BIT  ; ymmF=(21 23 25 27 29 2B 2D 2F 2H 2J 2L 2N 2P 2R 2T 2V)
+    vpsrlw      ymmH, ymmH, BYTE_BIT  ; ymmH=(31 33 35 37 39 3B 3D 3F 3H 3J 3L 3N 3P 3R 3T 3V)
+
+%endif  ; RGB_PIXELSIZE ; ---------------
+
+    ; ymm0=R(02468ACEGIKMOQSU)=RE, ymm2=G(02468ACEGIKMOQSU)=GE, ymm4=B(02468ACEGIKMOQSU)=BE
+    ; ymm1=R(13579BDFHJLNPRTV)=RO, ymm3=G(13579BDFHJLNPRTV)=GO, ymm5=B(13579BDFHJLNPRTV)=BO
+
+    ; (Original)
+    ; Y  =  0.29900 * R + 0.58700 * G + 0.11400 * B
+    ; Cb = -0.16874 * R - 0.33126 * G + 0.50000 * B + CENTERJSAMPLE
+    ; Cr =  0.50000 * R - 0.41869 * G - 0.08131 * B + CENTERJSAMPLE
+    ;
+    ; (This implementation)
+    ; Y  =  0.29900 * R + 0.33700 * G + 0.11400 * B + 0.25000 * G
+    ; Cb = -0.16874 * R - 0.33126 * G + 0.50000 * B + CENTERJSAMPLE
+    ; Cr =  0.50000 * R - 0.41869 * G - 0.08131 * B + CENTERJSAMPLE
+
+    vmovdqa     YMMWORD [wk(0)], ymm0   ; wk(0)=RE
+    vmovdqa     YMMWORD [wk(1)], ymm1   ; wk(1)=RO
+    vmovdqa     YMMWORD [wk(2)], ymm4   ; wk(2)=BE
+    vmovdqa     YMMWORD [wk(3)], ymm5   ; wk(3)=BO
+
+    vmovdqa     ymm6, ymm1
+    vpunpcklwd  ymm1, ymm1, ymm3
+    vpunpckhwd  ymm6, ymm6, ymm3
+    vmovdqa     ymm7, ymm1
+    vmovdqa     ymm4, ymm6
+    vpmaddwd    ymm1, ymm1, [rel PW_F0299_F0337]  ; ymm1=ROL*FIX(0.299)+GOL*FIX(0.337)
+    vpmaddwd    ymm6, ymm6, [rel PW_F0299_F0337]  ; ymm6=ROH*FIX(0.299)+GOH*FIX(0.337)
+    vpmaddwd    ymm7, ymm7, [rel PW_MF016_MF033]  ; ymm7=ROL*-FIX(0.168)+GOL*-FIX(0.331)
+    vpmaddwd    ymm4, ymm4, [rel PW_MF016_MF033]  ; ymm4=ROH*-FIX(0.168)+GOH*-FIX(0.331)
+
+    vmovdqa     YMMWORD [wk(4)], ymm1   ; wk(4)=ROL*FIX(0.299)+GOL*FIX(0.337)
+    vmovdqa     YMMWORD [wk(5)], ymm6   ; wk(5)=ROH*FIX(0.299)+GOH*FIX(0.337)
+
+    vpxor       ymm1, ymm1, ymm1
+    vpxor       ymm6, ymm6, ymm6
+    vpunpcklwd  ymm1, ymm1, ymm5        ; ymm1=BOL
+    vpunpckhwd  ymm6, ymm6, ymm5        ; ymm6=BOH
+    vpsrld      ymm1, ymm1, 1           ; ymm1=BOL*FIX(0.500)
+    vpsrld      ymm6, ymm6, 1           ; ymm6=BOH*FIX(0.500)
+
+    vmovdqa     ymm5, [rel PD_ONEHALFM1_CJ]  ; ymm5=[PD_ONEHALFM1_CJ]
+
+    vpaddd      ymm7, ymm7, ymm1
+    vpaddd      ymm4, ymm4, ymm6
+    vpaddd      ymm7, ymm7, ymm5
+    vpaddd      ymm4, ymm4, ymm5
+    vpsrld      ymm7, ymm7, SCALEBITS   ; ymm7=CbOL
+    vpsrld      ymm4, ymm4, SCALEBITS   ; ymm4=CbOH
+    vpackssdw   ymm7, ymm7, ymm4        ; ymm7=CbO
+
+    vmovdqa     ymm1, YMMWORD [wk(2)]   ; ymm1=BE
+
+    vmovdqa     ymm6, ymm0
+    vpunpcklwd  ymm0, ymm0, ymm2
+    vpunpckhwd  ymm6, ymm6, ymm2
+    vmovdqa     ymm5, ymm0
+    vmovdqa     ymm4, ymm6
+    vpmaddwd    ymm0, ymm0, [rel PW_F0299_F0337]  ; ymm0=REL*FIX(0.299)+GEL*FIX(0.337)
+    vpmaddwd    ymm6, ymm6, [rel PW_F0299_F0337]  ; ymm6=REH*FIX(0.299)+GEH*FIX(0.337)
+    vpmaddwd    ymm5, ymm5, [rel PW_MF016_MF033]  ; ymm5=REL*-FIX(0.168)+GEL*-FIX(0.331)
+    vpmaddwd    ymm4, ymm4, [rel PW_MF016_MF033]  ; ymm4=REH*-FIX(0.168)+GEH*-FIX(0.331)
+
+    vmovdqa     YMMWORD [wk(6)], ymm0   ; wk(6)=REL*FIX(0.299)+GEL*FIX(0.337)
+    vmovdqa     YMMWORD [wk(7)], ymm6   ; wk(7)=REH*FIX(0.299)+GEH*FIX(0.337)
+
+    vpxor       ymm0, ymm0, ymm0
+    vpxor       ymm6, ymm6, ymm6
+    vpunpcklwd  ymm0, ymm0, ymm1        ; ymm0=BEL
+    vpunpckhwd  ymm6, ymm6, ymm1        ; ymm6=BEH
+    vpsrld      ymm0, ymm0, 1           ; ymm0=BEL*FIX(0.500)
+    vpsrld      ymm6, ymm6, 1           ; ymm6=BEH*FIX(0.500)
+
+    vmovdqa     ymm1, [rel PD_ONEHALFM1_CJ]  ; ymm1=[PD_ONEHALFM1_CJ]
+
+    vpaddd      ymm5, ymm5, ymm0
+    vpaddd      ymm4, ymm4, ymm6
+    vpaddd      ymm5, ymm5, ymm1
+    vpaddd      ymm4, ymm4, ymm1
+    vpsrld      ymm5, ymm5, SCALEBITS   ; ymm5=CbEL
+    vpsrld      ymm4, ymm4, SCALEBITS   ; ymm4=CbEH
+    vpackssdw   ymm5, ymm5, ymm4        ; ymm5=CbE
+
+    vpsllw      ymm7, ymm7, BYTE_BIT
+    vpor        ymm5, ymm5, ymm7        ; ymm5=Cb
+    vmovdqu     YMMWORD [rbx], ymm5     ; Save Cb
+
+    vmovdqa     ymm0, YMMWORD [wk(3)]   ; ymm0=BO
+    vmovdqa     ymm6, YMMWORD [wk(2)]   ; ymm6=BE
+    vmovdqa     ymm1, YMMWORD [wk(1)]   ; ymm1=RO
+
+    vmovdqa     ymm4, ymm0
+    vpunpcklwd  ymm0, ymm0, ymm3
+    vpunpckhwd  ymm4, ymm4, ymm3
+    vmovdqa     ymm7, ymm0
+    vmovdqa     ymm5, ymm4
+    vpmaddwd    ymm0, ymm0, [rel PW_F0114_F0250]  ; ymm0=BOL*FIX(0.114)+GOL*FIX(0.250)
+    vpmaddwd    ymm4, ymm4, [rel PW_F0114_F0250]  ; ymm4=BOH*FIX(0.114)+GOH*FIX(0.250)
+    vpmaddwd    ymm7, ymm7, [rel PW_MF008_MF041]  ; ymm7=BOL*-FIX(0.081)+GOL*-FIX(0.418)
+    vpmaddwd    ymm5, ymm5, [rel PW_MF008_MF041]  ; ymm5=BOH*-FIX(0.081)+GOH*-FIX(0.418)
+
+    vmovdqa     ymm3, [rel PD_ONEHALF]            ; ymm3=[PD_ONEHALF]
+
+    vpaddd      ymm0, ymm0, YMMWORD [wk(4)]
+    vpaddd      ymm4, ymm4, YMMWORD [wk(5)]
+    vpaddd      ymm0, ymm0, ymm3
+    vpaddd      ymm4, ymm4, ymm3
+    vpsrld      ymm0, ymm0, SCALEBITS   ; ymm0=YOL
+    vpsrld      ymm4, ymm4, SCALEBITS   ; ymm4=YOH
+    vpackssdw   ymm0, ymm0, ymm4        ; ymm0=YO
+
+    vpxor       ymm3, ymm3, ymm3
+    vpxor       ymm4, ymm4, ymm4
+    vpunpcklwd  ymm3, ymm3, ymm1        ; ymm3=ROL
+    vpunpckhwd  ymm4, ymm4, ymm1        ; ymm4=ROH
+    vpsrld      ymm3, ymm3, 1           ; ymm3=ROL*FIX(0.500)
+    vpsrld      ymm4, ymm4, 1           ; ymm4=ROH*FIX(0.500)
+
+    vmovdqa     ymm1, [rel PD_ONEHALFM1_CJ]  ; ymm1=[PD_ONEHALFM1_CJ]
+
+    vpaddd      ymm7, ymm7, ymm3
+    vpaddd      ymm5, ymm5, ymm4
+    vpaddd      ymm7, ymm7, ymm1
+    vpaddd      ymm5, ymm5, ymm1
+    vpsrld      ymm7, ymm7, SCALEBITS   ; ymm7=CrOL
+    vpsrld      ymm5, ymm5, SCALEBITS   ; ymm5=CrOH
+    vpackssdw   ymm7, ymm7, ymm5        ; ymm7=CrO
+
+    vmovdqa     ymm3, YMMWORD [wk(0)]   ; ymm3=RE
+
+    vmovdqa     ymm4, ymm6
+    vpunpcklwd  ymm6, ymm6, ymm2
+    vpunpckhwd  ymm4, ymm4, ymm2
+    vmovdqa     ymm1, ymm6
+    vmovdqa     ymm5, ymm4
+    vpmaddwd    ymm6, ymm6, [rel PW_F0114_F0250]  ; ymm6=BEL*FIX(0.114)+GEL*FIX(0.250)
+    vpmaddwd    ymm4, ymm4, [rel PW_F0114_F0250]  ; ymm4=BEH*FIX(0.114)+GEH*FIX(0.250)
+    vpmaddwd    ymm1, ymm1, [rel PW_MF008_MF041]  ; ymm1=BEL*-FIX(0.081)+GEL*-FIX(0.418)
+    vpmaddwd    ymm5, ymm5, [rel PW_MF008_MF041]  ; ymm5=BEH*-FIX(0.081)+GEH*-FIX(0.418)
+
+    vmovdqa     ymm2, [rel PD_ONEHALF]            ; ymm2=[PD_ONEHALF]
+
+    vpaddd      ymm6, ymm6, YMMWORD [wk(6)]
+    vpaddd      ymm4, ymm4, YMMWORD [wk(7)]
+    vpaddd      ymm6, ymm6, ymm2
+    vpaddd      ymm4, ymm4, ymm2
+    vpsrld      ymm6, ymm6, SCALEBITS   ; ymm6=YEL
+    vpsrld      ymm4, ymm4, SCALEBITS   ; ymm4=YEH
+    vpackssdw   ymm6, ymm6, ymm4        ; ymm6=YE
+
+    vpsllw      ymm0, ymm0, BYTE_BIT
+    vpor        ymm6, ymm6, ymm0        ; ymm6=Y
+    vmovdqu     YMMWORD [rdi], ymm6     ; Save Y
+
+    vpxor       ymm2, ymm2, ymm2
+    vpxor       ymm4, ymm4, ymm4
+    vpunpcklwd  ymm2, ymm2, ymm3        ; ymm2=REL
+    vpunpckhwd  ymm4, ymm4, ymm3        ; ymm4=REH
+    vpsrld      ymm2, ymm2, 1           ; ymm2=REL*FIX(0.500)
+    vpsrld      ymm4, ymm4, 1           ; ymm4=REH*FIX(0.500)
+
+    vmovdqa     ymm0, [rel PD_ONEHALFM1_CJ]  ; ymm0=[PD_ONEHALFM1_CJ]
+
+    vpaddd      ymm1, ymm1, ymm2
+    vpaddd      ymm5, ymm5, ymm4
+    vpaddd      ymm1, ymm1, ymm0
+    vpaddd      ymm5, ymm5, ymm0
+    vpsrld      ymm1, ymm1, SCALEBITS   ; ymm1=CrEL
+    vpsrld      ymm5, ymm5, SCALEBITS   ; ymm5=CrEH
+    vpackssdw   ymm1, ymm1, ymm5        ; ymm1=CrE
+
+    vpsllw      ymm7, ymm7, BYTE_BIT
+    vpor        ymm1, ymm1, ymm7        ; ymm1=Cr
+    vmovdqu     YMMWORD [rdx], ymm1     ; Save Cr
+
+    sub         rcx, byte SIZEOF_YMMWORD
+    add         rsi, RGB_PIXELSIZE*SIZEOF_YMMWORD  ; inptr
+    add         rdi, byte SIZEOF_YMMWORD           ; outptr0
+    add         rbx, byte SIZEOF_YMMWORD           ; outptr1
+    add         rdx, byte SIZEOF_YMMWORD           ; outptr2
+    cmp         rcx, byte SIZEOF_YMMWORD
+    jae         near .columnloop
+    test        rcx, rcx
+    jnz         near .column_ld1
+
+    pop         rcx                     ; col
+    pop         rsi
+    pop         rdi
+    pop         rbx
+    pop         rdx
+
+    add         rsi, byte SIZEOF_JSAMPROW  ; input_buf
+    add         rdi, byte SIZEOF_JSAMPROW
+    add         rbx, byte SIZEOF_JSAMPROW
+    add         rdx, byte SIZEOF_JSAMPROW
+    dec         rax                        ; num_rows
+    jg          near .rowloop
+
+.return:
+    pop         rbx
+    vzeroupper
+    uncollect_args 5
+    mov         rsp, rbp                ; rsp <- aligned rbp
+    pop         rsp                     ; rsp <- original rbp
+    pop         rbp
+    ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+    align       32
diff --git a/media/libjpeg/simd/x86_64/jccolext-sse2.asm b/media/libjpeg/simd/x86_64/jccolext-sse2.asm
new file mode 100644
index 0000000000..af70ed6010
--- /dev/null
+++ b/media/libjpeg/simd/x86_64/jccolext-sse2.asm
@@ -0,0 +1,484 @@
+;
+; jccolext.asm - colorspace conversion (64-bit SSE2)
+;
+; Copyright (C) 2009, 2016, D. R. Commander.
+; Copyright (C) 2018, Matthias Räncker.
+;
+; Based on the x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+
+%include "jcolsamp.inc"
+
+; --------------------------------------------------------------------------
+;
+; Convert some rows of samples to the output colorspace.
+;
+; GLOBAL(void)
+; jsimd_rgb_ycc_convert_sse2(JDIMENSION img_width, JSAMPARRAY input_buf,
+;                            JSAMPIMAGE output_buf, JDIMENSION output_row,
+;                            int num_rows);
+;
+
+; r10d = JDIMENSION img_width
+; r11 = JSAMPARRAY input_buf
+; r12 = JSAMPIMAGE output_buf
+; r13d = JDIMENSION output_row
+; r14d = int num_rows
+
+%define wk(i)   rbp - (WK_NUM - (i)) * SIZEOF_XMMWORD  ; xmmword wk[WK_NUM]
+%define WK_NUM  8
+
+    align       32
+    GLOBAL_FUNCTION(jsimd_rgb_ycc_convert_sse2)
+
+EXTN(jsimd_rgb_ycc_convert_sse2):
+    push        rbp
+    mov         rax, rsp                     ; rax = original rbp
+    sub         rsp, byte 4
+    and         rsp, byte (-SIZEOF_XMMWORD)  ; align to 128 bits
+    mov         [rsp], rax
+    mov         rbp, rsp                     ; rbp = aligned rbp
+    lea         rsp, [wk(0)]
+    collect_args 5
+    push        rbx
+
+    mov         ecx, r10d
+    test        rcx, rcx
+    jz          near .return
+
+    push        rcx
+
+    mov         rsi, r12
+    mov         ecx, r13d
+    mov         rdip, JSAMPARRAY [rsi+0*SIZEOF_JSAMPARRAY]
+    mov         rbxp, JSAMPARRAY [rsi+1*SIZEOF_JSAMPARRAY]
+    mov         rdxp, JSAMPARRAY [rsi+2*SIZEOF_JSAMPARRAY]
+    lea         rdi, [rdi+rcx*SIZEOF_JSAMPROW]
+    lea         rbx, [rbx+rcx*SIZEOF_JSAMPROW]
+    lea         rdx, [rdx+rcx*SIZEOF_JSAMPROW]
+
+    pop         rcx
+
+    mov         rsi, r11
+    mov         eax, r14d
+    test        rax, rax
+    jle         near .return
+.rowloop:
+    push        rdx
+    push        rbx
+    push        rdi
+    push        rsi
+    push        rcx                     ; col
+
+    mov         rsip, JSAMPROW [rsi]    ; inptr
+    mov         rdip, JSAMPROW [rdi]    ; outptr0
+    mov         rbxp, JSAMPROW [rbx]    ; outptr1
+    mov         rdxp, JSAMPROW [rdx]    ; outptr2
+
+    cmp         rcx, byte SIZEOF_XMMWORD
+    jae         near .columnloop
+
+%if RGB_PIXELSIZE == 3  ; ---------------
+
+.column_ld1:
+    push        rax
+    push        rdx
+    lea         rcx, [rcx+rcx*2]        ; imul ecx,RGB_PIXELSIZE
+    test        cl, SIZEOF_BYTE
+    jz          short .column_ld2
+    sub         rcx, byte SIZEOF_BYTE
+    movzx       rax, byte [rsi+rcx]
+.column_ld2:
+    test        cl, SIZEOF_WORD
+    jz          short .column_ld4
+    sub         rcx, byte SIZEOF_WORD
+    movzx       rdx, word [rsi+rcx]
+    shl         rax, WORD_BIT
+    or          rax, rdx
+.column_ld4:
+    movd        xmmA, eax
+    pop         rdx
+    pop         rax
+    test        cl, SIZEOF_DWORD
+    jz          short .column_ld8
+    sub         rcx, byte SIZEOF_DWORD
+    movd        xmmF, XMM_DWORD [rsi+rcx]
+    pslldq      xmmA, SIZEOF_DWORD
+    por         xmmA, xmmF
+.column_ld8:
+    test        cl, SIZEOF_MMWORD
+    jz          short .column_ld16
+    sub         rcx, byte SIZEOF_MMWORD
+    movq        xmmB, XMM_MMWORD [rsi+rcx]
+    pslldq      xmmA, SIZEOF_MMWORD
+    por         xmmA, xmmB
+.column_ld16:
+    test        cl, SIZEOF_XMMWORD
+    jz          short .column_ld32
+    movdqa      xmmF, xmmA
+    movdqu      xmmA, XMMWORD [rsi+0*SIZEOF_XMMWORD]
+    mov         rcx, SIZEOF_XMMWORD
+    jmp         short .rgb_ycc_cnv
+.column_ld32:
+    test        cl, 2*SIZEOF_XMMWORD
+    mov         rcx, SIZEOF_XMMWORD
+    jz          short .rgb_ycc_cnv
+    movdqa      xmmB, xmmA
+    movdqu      xmmA, XMMWORD [rsi+0*SIZEOF_XMMWORD]
+    movdqu      xmmF, XMMWORD [rsi+1*SIZEOF_XMMWORD]
+    jmp         short .rgb_ycc_cnv
+
+.columnloop:
+    movdqu      xmmA, XMMWORD [rsi+0*SIZEOF_XMMWORD]
+    movdqu      xmmF, XMMWORD [rsi+1*SIZEOF_XMMWORD]
+    movdqu      xmmB, XMMWORD [rsi+2*SIZEOF_XMMWORD]
+
+.rgb_ycc_cnv:
+    ; xmmA=(00 10 20 01 11 21 02 12 22 03 13 23 04 14 24 05)
+    ; xmmF=(15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A)
+    ; xmmB=(2A 0B 1B 2B 0C 1C 2C 0D 1D 2D 0E 1E 2E 0F 1F 2F)
+
+    movdqa      xmmG, xmmA
+    pslldq      xmmA, 8     ; xmmA=(-- -- -- -- -- -- -- -- 00 10 20 01 11 21 02 12)
+    psrldq      xmmG, 8     ; xmmG=(22 03 13 23 04 14 24 05 -- -- -- -- -- -- -- --)
+
+    punpckhbw   xmmA, xmmF  ; xmmA=(00 08 10 18 20 28 01 09 11 19 21 29 02 0A 12 1A)
+    pslldq      xmmF, 8     ; xmmF=(-- -- -- -- -- -- -- -- 15 25 06 16 26 07 17 27)
+
+    punpcklbw   xmmG, xmmB  ; xmmG=(22 2A 03 0B 13 1B 23 2B 04 0C 14 1C 24 2C 05 0D)
+    punpckhbw   xmmF, xmmB  ; xmmF=(15 1D 25 2D 06 0E 16 1E 26 2E 07 0F 17 1F 27 2F)
+
+    movdqa      xmmD, xmmA
+    pslldq      xmmA, 8     ; xmmA=(-- -- -- -- -- -- -- -- 00 08 10 18 20 28 01 09)
+    psrldq      xmmD, 8     ; xmmD=(11 19 21 29 02 0A 12 1A -- -- -- -- -- -- -- --)
+
+    punpckhbw   xmmA, xmmG  ; xmmA=(00 04 08 0C 10 14 18 1C 20 24 28 2C 01 05 09 0D)
+    pslldq      xmmG, 8     ; xmmG=(-- -- -- -- -- -- -- -- 22 2A 03 0B 13 1B 23 2B)
+
+    punpcklbw   xmmD, xmmF  ; xmmD=(11 15 19 1D 21 25 29 2D 02 06 0A 0E 12 16 1A 1E)
+    punpckhbw   xmmG, xmmF  ; xmmG=(22 26 2A 2E 03 07 0B 0F 13 17 1B 1F 23 27 2B 2F)
+
+    movdqa      xmmE, xmmA
+    pslldq      xmmA, 8     ; xmmA=(-- -- -- -- -- -- -- -- 00 04 08 0C 10 14 18 1C)
+    psrldq      xmmE, 8     ; xmmE=(20 24 28 2C 01 05 09 0D -- -- -- -- -- -- -- --)
+
+    punpckhbw   xmmA, xmmD  ; xmmA=(00 02 04 06 08 0A 0C 0E 10 12 14 16 18 1A 1C 1E)
+    pslldq      xmmD, 8     ; xmmD=(-- -- -- -- -- -- -- -- 11 15 19 1D 21 25 29 2D)
+
+    punpcklbw   xmmE, xmmG  ; xmmE=(20 22 24 26 28 2A 2C 2E 01 03 05 07 09 0B 0D 0F)
+    punpckhbw   xmmD, xmmG  ; xmmD=(11 13 15 17 19 1B 1D 1F 21 23 25 27 29 2B 2D 2F)
+
+    pxor        xmmH, xmmH
+
+    movdqa      xmmC, xmmA
+    punpcklbw   xmmA, xmmH  ; xmmA=(00 02 04 06 08 0A 0C 0E)
+    punpckhbw   xmmC, xmmH  ; xmmC=(10 12 14 16 18 1A 1C 1E)
+
+    movdqa      xmmB, xmmE
+    punpcklbw   xmmE, xmmH  ; xmmE=(20 22 24 26 28 2A 2C 2E)
+    punpckhbw   xmmB, xmmH  ; xmmB=(01 03 05 07 09 0B 0D 0F)
+
+    movdqa      xmmF, xmmD
+    punpcklbw   xmmD, xmmH  ; xmmD=(11 13 15 17 19 1B 1D 1F)
+    punpckhbw   xmmF, xmmH  ; xmmF=(21 23 25 27 29 2B 2D 2F)
+
+%else  ; RGB_PIXELSIZE == 4 ; -----------
+
+.column_ld1:
+    test        cl, SIZEOF_XMMWORD/16
+    jz          short .column_ld2
+    sub         rcx, byte SIZEOF_XMMWORD/16
+    movd        xmmA, XMM_DWORD [rsi+rcx*RGB_PIXELSIZE]
+.column_ld2:
+    test        cl, SIZEOF_XMMWORD/8
+    jz          short .column_ld4
+    sub         rcx, byte SIZEOF_XMMWORD/8
+    movq        xmmE, XMM_MMWORD [rsi+rcx*RGB_PIXELSIZE]
+    pslldq      xmmA, SIZEOF_MMWORD
+    por         xmmA, xmmE
+.column_ld4:
+    test        cl, SIZEOF_XMMWORD/4
+    jz          short .column_ld8
+    sub         rcx, byte SIZEOF_XMMWORD/4
+    movdqa      xmmE, xmmA
+    movdqu      xmmA, XMMWORD [rsi+rcx*RGB_PIXELSIZE]
+.column_ld8:
+    test        cl, SIZEOF_XMMWORD/2
+    mov         rcx, SIZEOF_XMMWORD
+    jz          short .rgb_ycc_cnv
+    movdqa      xmmF, xmmA
+    movdqa      xmmH, xmmE
+    movdqu      xmmA, XMMWORD [rsi+0*SIZEOF_XMMWORD]
+    movdqu      xmmE, XMMWORD [rsi+1*SIZEOF_XMMWORD]
+    jmp         short .rgb_ycc_cnv
+
+.columnloop:
+    movdqu      xmmA, XMMWORD [rsi+0*SIZEOF_XMMWORD]
+    movdqu      xmmE, XMMWORD [rsi+1*SIZEOF_XMMWORD]
+    movdqu      xmmF, XMMWORD [rsi+2*SIZEOF_XMMWORD]
+    movdqu      xmmH, XMMWORD [rsi+3*SIZEOF_XMMWORD]
+
+.rgb_ycc_cnv:
+    ; xmmA=(00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33)
+    ; xmmE=(04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37)
+    ; xmmF=(08 18 28 38 09 19 29 39 0A 1A 2A 3A 0B 1B 2B 3B)
+    ; xmmH=(0C 1C 2C 3C 0D 1D 2D 3D 0E 1E 2E 3E 0F 1F 2F 3F)
+
+    movdqa      xmmD, xmmA
+    punpcklbw   xmmA, xmmE      ; xmmA=(00 04 10 14 20 24 30 34 01 05 11 15 21 25 31 35)
+    punpckhbw   xmmD, xmmE      ; xmmD=(02 06 12 16 22 26 32 36 03 07 13 17 23 27 33 37)
+
+    movdqa      xmmC, xmmF
+    punpcklbw   xmmF, xmmH      ; xmmF=(08 0C 18 1C 28 2C 38 3C 09 0D 19 1D 29 2D 39 3D)
+    punpckhbw   xmmC, xmmH      ; xmmC=(0A 0E 1A 1E 2A 2E 3A 3E 0B 0F 1B 1F 2B 2F 3B 3F)
+
+    movdqa      xmmB, xmmA
+    punpcklwd   xmmA, xmmF      ; xmmA=(00 04 08 0C 10 14 18 1C 20 24 28 2C 30 34 38 3C)
+    punpckhwd   xmmB, xmmF      ; xmmB=(01 05 09 0D 11 15 19 1D 21 25 29 2D 31 35 39 3D)
+
+    movdqa      xmmG, xmmD
+    punpcklwd   xmmD, xmmC      ; xmmD=(02 06 0A 0E 12 16 1A 1E 22 26 2A 2E 32 36 3A 3E)
+    punpckhwd   xmmG, xmmC      ; xmmG=(03 07 0B 0F 13 17 1B 1F 23 27 2B 2F 33 37 3B 3F)
+
+    movdqa      xmmE, xmmA
+    punpcklbw   xmmA, xmmD      ; xmmA=(00 02 04 06 08 0A 0C 0E 10 12 14 16 18 1A 1C 1E)
+    punpckhbw   xmmE, xmmD      ; xmmE=(20 22 24 26 28 2A 2C 2E 30 32 34 36 38 3A 3C 3E)
+
+    movdqa      xmmH, xmmB
+    punpcklbw   xmmB, xmmG      ; xmmB=(01 03 05 07 09 0B 0D 0F 11 13 15 17 19 1B 1D 1F)
+    punpckhbw   xmmH, xmmG      ; xmmH=(21 23 25 27 29 2B 2D 2F 31 33 35 37 39 3B 3D 3F)
+
+    pxor        xmmF, xmmF
+
+    movdqa      xmmC, xmmA
+    punpcklbw   xmmA, xmmF      ; xmmA=(00 02 04 06 08 0A 0C 0E)
+    punpckhbw   xmmC, xmmF      ; xmmC=(10 12 14 16 18 1A 1C 1E)
+
+    movdqa      xmmD, xmmB
+    punpcklbw   xmmB, xmmF      ; xmmB=(01 03 05 07 09 0B 0D 0F)
+    punpckhbw   xmmD, xmmF      ; xmmD=(11 13 15 17 19 1B 1D 1F)
+
+    movdqa      xmmG, xmmE
+    punpcklbw   xmmE, xmmF      ; xmmE=(20 22 24 26 28 2A 2C 2E)
+    punpckhbw   xmmG, xmmF      ; xmmG=(30 32 34 36 38 3A 3C 3E)
+
+    punpcklbw   xmmF, xmmH
+    punpckhbw   xmmH, xmmH
+    psrlw       xmmF, BYTE_BIT  ; xmmF=(21 23 25 27 29 2B 2D 2F)
+    psrlw       xmmH, BYTE_BIT  ; xmmH=(31 33 35 37 39 3B 3D 3F)
+
+%endif  ; RGB_PIXELSIZE ; ---------------
+
+    ; xmm0=R(02468ACE)=RE, xmm2=G(02468ACE)=GE, xmm4=B(02468ACE)=BE
+    ; xmm1=R(13579BDF)=RO, xmm3=G(13579BDF)=GO, xmm5=B(13579BDF)=BO
+
+    ; (Original)
+    ; Y  =  0.29900 * R + 0.58700 * G + 0.11400 * B
+    ; Cb = -0.16874 * R - 0.33126 * G + 0.50000 * B + CENTERJSAMPLE
+    ; Cr =  0.50000 * R - 0.41869 * G - 0.08131 * B + CENTERJSAMPLE
+    ;
+    ; (This implementation)
+    ; Y  =  0.29900 * R + 0.33700 * G + 0.11400 * B + 0.25000 * G
+    ; Cb = -0.16874 * R - 0.33126 * G + 0.50000 * B + CENTERJSAMPLE
+    ; Cr =  0.50000 * R - 0.41869 * G - 0.08131 * B + CENTERJSAMPLE
+
+    movdqa      XMMWORD [wk(0)], xmm0   ; wk(0)=RE
+    movdqa      XMMWORD [wk(1)], xmm1   ; wk(1)=RO
+    movdqa      XMMWORD [wk(2)], xmm4   ; wk(2)=BE
+    movdqa      XMMWORD [wk(3)], xmm5   ; wk(3)=BO
+
+    movdqa      xmm6, xmm1
+    punpcklwd   xmm1, xmm3
+    punpckhwd   xmm6, xmm3
+    movdqa      xmm7, xmm1
+    movdqa      xmm4, xmm6
+    pmaddwd     xmm1, [rel PW_F0299_F0337]  ; xmm1=ROL*FIX(0.299)+GOL*FIX(0.337)
+    pmaddwd     xmm6, [rel PW_F0299_F0337]  ; xmm6=ROH*FIX(0.299)+GOH*FIX(0.337)
+    pmaddwd     xmm7, [rel PW_MF016_MF033]  ; xmm7=ROL*-FIX(0.168)+GOL*-FIX(0.331)
+    pmaddwd     xmm4, [rel PW_MF016_MF033]  ; xmm4=ROH*-FIX(0.168)+GOH*-FIX(0.331)
+
+    movdqa      XMMWORD [wk(4)], xmm1   ; wk(4)=ROL*FIX(0.299)+GOL*FIX(0.337)
+    movdqa      XMMWORD [wk(5)], xmm6   ; wk(5)=ROH*FIX(0.299)+GOH*FIX(0.337)
+
+    pxor        xmm1, xmm1
+    pxor        xmm6, xmm6
+    punpcklwd   xmm1, xmm5              ; xmm1=BOL
+    punpckhwd   xmm6, xmm5              ; xmm6=BOH
+    psrld       xmm1, 1                 ; xmm1=BOL*FIX(0.500)
+    psrld       xmm6, 1                 ; xmm6=BOH*FIX(0.500)
+
+    movdqa      xmm5, [rel PD_ONEHALFM1_CJ]  ; xmm5=[PD_ONEHALFM1_CJ]
+
+    paddd       xmm7, xmm1
+    paddd       xmm4, xmm6
+    paddd       xmm7, xmm5
+    paddd       xmm4, xmm5
+    psrld       xmm7, SCALEBITS         ; xmm7=CbOL
+    psrld       xmm4, SCALEBITS         ; xmm4=CbOH
+    packssdw    xmm7, xmm4              ; xmm7=CbO
+
+    movdqa      xmm1, XMMWORD [wk(2)]   ; xmm1=BE
+
+    movdqa      xmm6, xmm0
+    punpcklwd   xmm0, xmm2
+    punpckhwd   xmm6, xmm2
+    movdqa      xmm5, xmm0
+    movdqa      xmm4, xmm6
+    pmaddwd     xmm0, [rel PW_F0299_F0337]  ; xmm0=REL*FIX(0.299)+GEL*FIX(0.337)
+    pmaddwd     xmm6, [rel PW_F0299_F0337]  ; xmm6=REH*FIX(0.299)+GEH*FIX(0.337)
+    pmaddwd     xmm5, [rel PW_MF016_MF033]  ; xmm5=REL*-FIX(0.168)+GEL*-FIX(0.331)
+    pmaddwd     xmm4, [rel PW_MF016_MF033]  ; xmm4=REH*-FIX(0.168)+GEH*-FIX(0.331)
+
+    movdqa      XMMWORD [wk(6)], xmm0   ; wk(6)=REL*FIX(0.299)+GEL*FIX(0.337)
+    movdqa      XMMWORD [wk(7)], xmm6   ; wk(7)=REH*FIX(0.299)+GEH*FIX(0.337)
+
+    pxor        xmm0, xmm0
+    pxor        xmm6, xmm6
+    punpcklwd   xmm0, xmm1              ; xmm0=BEL
+    punpckhwd   xmm6, xmm1              ; xmm6=BEH
+    psrld       xmm0, 1                 ; xmm0=BEL*FIX(0.500)
+    psrld       xmm6, 1                 ; xmm6=BEH*FIX(0.500)
+
+    movdqa      xmm1, [rel PD_ONEHALFM1_CJ]  ; xmm1=[PD_ONEHALFM1_CJ]
+
+    paddd       xmm5, xmm0
+    paddd       xmm4, xmm6
+    paddd       xmm5, xmm1
+    paddd       xmm4, xmm1
+    psrld       xmm5, SCALEBITS         ; xmm5=CbEL
+    psrld       xmm4, SCALEBITS         ; xmm4=CbEH
+    packssdw    xmm5, xmm4              ; xmm5=CbE
+
+    psllw       xmm7, BYTE_BIT
+    por         xmm5, xmm7              ; xmm5=Cb
+    movdqa      XMMWORD [rbx], xmm5     ; Save Cb
+
+    movdqa      xmm0, XMMWORD [wk(3)]   ; xmm0=BO
+    movdqa      xmm6, XMMWORD [wk(2)]   ; xmm6=BE
+    movdqa      xmm1, XMMWORD [wk(1)]   ; xmm1=RO
+
+    movdqa      xmm4, xmm0
+    punpcklwd   xmm0, xmm3
+    punpckhwd   xmm4, xmm3
+    movdqa      xmm7, xmm0
+    movdqa      xmm5, xmm4
+    pmaddwd     xmm0, [rel PW_F0114_F0250]  ; xmm0=BOL*FIX(0.114)+GOL*FIX(0.250)
+    pmaddwd     xmm4, [rel PW_F0114_F0250]  ; xmm4=BOH*FIX(0.114)+GOH*FIX(0.250)
+    pmaddwd     xmm7, [rel PW_MF008_MF041]  ; xmm7=BOL*-FIX(0.081)+GOL*-FIX(0.418)
+    pmaddwd     xmm5, [rel PW_MF008_MF041]  ; xmm5=BOH*-FIX(0.081)+GOH*-FIX(0.418)
+
+    movdqa      xmm3, [rel PD_ONEHALF]      ; xmm3=[PD_ONEHALF]
+
+    paddd       xmm0, XMMWORD [wk(4)]
+    paddd       xmm4, XMMWORD [wk(5)]
+    paddd       xmm0, xmm3
+    paddd       xmm4, xmm3
+    psrld       xmm0, SCALEBITS         ; xmm0=YOL
+    psrld       xmm4, SCALEBITS         ; xmm4=YOH
+    packssdw    xmm0, xmm4              ; xmm0=YO
+
+    pxor        xmm3, xmm3
+    pxor        xmm4, xmm4
+    punpcklwd   xmm3, xmm1              ; xmm3=ROL
+    punpckhwd   xmm4, xmm1              ; xmm4=ROH
+    psrld       xmm3, 1                 ; xmm3=ROL*FIX(0.500)
+    psrld       xmm4, 1                 ; xmm4=ROH*FIX(0.500)
+
+    movdqa      xmm1, [rel PD_ONEHALFM1_CJ]  ; xmm1=[PD_ONEHALFM1_CJ]
+
+    paddd       xmm7, xmm3
+    paddd       xmm5, xmm4
+    paddd       xmm7, xmm1
+    paddd       xmm5, xmm1
+    psrld       xmm7, SCALEBITS         ; xmm7=CrOL
+    psrld       xmm5, SCALEBITS         ; xmm5=CrOH
+    packssdw    xmm7, xmm5              ; xmm7=CrO
+
+    movdqa      xmm3, XMMWORD [wk(0)]   ; xmm3=RE
+
+    movdqa      xmm4, xmm6
+    punpcklwd   xmm6, xmm2
+    punpckhwd   xmm4, xmm2
+    movdqa      xmm1, xmm6
+    movdqa      xmm5, xmm4
+    pmaddwd     xmm6, [rel PW_F0114_F0250]  ; xmm6=BEL*FIX(0.114)+GEL*FIX(0.250)
+    pmaddwd     xmm4, [rel PW_F0114_F0250]  ; xmm4=BEH*FIX(0.114)+GEH*FIX(0.250)
+    pmaddwd     xmm1, [rel PW_MF008_MF041]  ; xmm1=BEL*-FIX(0.081)+GEL*-FIX(0.418)
+    pmaddwd     xmm5, [rel PW_MF008_MF041]  ; xmm5=BEH*-FIX(0.081)+GEH*-FIX(0.418)
+
+    movdqa      xmm2, [rel PD_ONEHALF]      ; xmm2=[PD_ONEHALF]
+
+    paddd       xmm6, XMMWORD [wk(6)]
+    paddd       xmm4, XMMWORD [wk(7)]
+    paddd       xmm6, xmm2
+    paddd       xmm4, xmm2
+    psrld       xmm6, SCALEBITS         ; xmm6=YEL
+    psrld       xmm4, SCALEBITS         ; xmm4=YEH
+    packssdw    xmm6, xmm4              ; xmm6=YE
+
+    psllw       xmm0, BYTE_BIT
+    por         xmm6, xmm0              ; xmm6=Y
+    movdqa      XMMWORD [rdi], xmm6     ; Save Y
+
+    pxor        xmm2, xmm2
+    pxor        xmm4, xmm4
+    punpcklwd   xmm2, xmm3              ; xmm2=REL
+    punpckhwd   xmm4, xmm3              ; xmm4=REH
+    psrld       xmm2, 1                 ; xmm2=REL*FIX(0.500)
+    psrld       xmm4, 1                 ; xmm4=REH*FIX(0.500)
+
+    movdqa      xmm0, [rel PD_ONEHALFM1_CJ]  ; xmm0=[PD_ONEHALFM1_CJ]
+
+    paddd       xmm1, xmm2
+    paddd       xmm5, xmm4
+    paddd       xmm1, xmm0
+    paddd       xmm5, xmm0
+    psrld       xmm1, SCALEBITS         ; xmm1=CrEL
+    psrld       xmm5, SCALEBITS         ; xmm5=CrEH
+    packssdw    xmm1, xmm5              ; xmm1=CrE
+
+    psllw       xmm7, BYTE_BIT
+    por         xmm1, xmm7              ; xmm1=Cr
+    movdqa      XMMWORD [rdx], xmm1     ; Save Cr
+
+    sub         rcx, byte SIZEOF_XMMWORD
+    add         rsi, byte RGB_PIXELSIZE*SIZEOF_XMMWORD  ; inptr
+    add         rdi, byte SIZEOF_XMMWORD                ; outptr0
+    add         rbx, byte SIZEOF_XMMWORD                ; outptr1
+    add         rdx, byte SIZEOF_XMMWORD                ; outptr2
+    cmp         rcx, byte SIZEOF_XMMWORD
+    jae         near .columnloop
+    test        rcx, rcx
+    jnz         near .column_ld1
+
+    pop         rcx                     ; col
+    pop         rsi
+    pop         rdi
+    pop         rbx
+    pop         rdx
+
+    add         rsi, byte SIZEOF_JSAMPROW  ; input_buf
+    add         rdi, byte SIZEOF_JSAMPROW
+    add         rbx, byte SIZEOF_JSAMPROW
+    add         rdx, byte SIZEOF_JSAMPROW
+    dec         rax                        ; num_rows
+    jg          near .rowloop
+
+.return:
+    pop         rbx
+    uncollect_args 5
+    mov         rsp, rbp                ; rsp <- aligned rbp
+    pop         rsp                     ; rsp <- original rbp
+    pop         rbp
+    ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+    align       32
diff --git a/media/libjpeg/simd/x86_64/jccolor-avx2.asm b/media/libjpeg/simd/x86_64/jccolor-avx2.asm
new file mode 100644
index 0000000000..16b78298dc
--- /dev/null
+++ b/media/libjpeg/simd/x86_64/jccolor-avx2.asm
@@ -0,0 +1,121 @@
+;
+; jccolor.asm - colorspace conversion (64-bit AVX2)
+;
+; Copyright (C) 2009, 2016, D. R. Commander.
+; Copyright (C) 2015, Intel Corporation.
+;
+; Based on the x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+
+%include "jsimdext.inc"
+
+; --------------------------------------------------------------------------
+
+%define SCALEBITS  16
+
+F_0_081 equ  5329                ; FIX(0.08131)
+F_0_114 equ  7471                ; FIX(0.11400)
+F_0_168 equ 11059                ; FIX(0.16874)
+F_0_250 equ 16384                ; FIX(0.25000)
+F_0_299 equ 19595                ; FIX(0.29900)
+F_0_331 equ 21709                ; FIX(0.33126)
+F_0_418 equ 27439                ; FIX(0.41869)
+F_0_587 equ 38470                ; FIX(0.58700)
+F_0_337 equ (F_0_587 - F_0_250)  ; FIX(0.58700) - FIX(0.25000)
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_CONST
+
+    alignz      32
+    GLOBAL_DATA(jconst_rgb_ycc_convert_avx2)
+
+EXTN(jconst_rgb_ycc_convert_avx2):
+
+PW_F0299_F0337  times 8 dw  F_0_299,  F_0_337
+PW_F0114_F0250  times 8 dw  F_0_114,  F_0_250
+PW_MF016_MF033  times 8 dw -F_0_168, -F_0_331
+PW_MF008_MF041  times 8 dw -F_0_081, -F_0_418
+PD_ONEHALFM1_CJ times 8 dd  (1 << (SCALEBITS - 1)) - 1 + \
+                            (CENTERJSAMPLE << SCALEBITS)
+PD_ONEHALF      times 8 dd  (1 << (SCALEBITS - 1))
+
+    alignz      32
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_TEXT
+    BITS        64
+
+%include "jccolext-avx2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_RGB_RED
+%define RGB_GREEN  EXT_RGB_GREEN
+%define RGB_BLUE  EXT_RGB_BLUE
+%define RGB_PIXELSIZE  EXT_RGB_PIXELSIZE
+%define jsimd_rgb_ycc_convert_avx2  jsimd_extrgb_ycc_convert_avx2
+%include "jccolext-avx2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_RGBX_RED
+%define RGB_GREEN  EXT_RGBX_GREEN
+%define RGB_BLUE  EXT_RGBX_BLUE
+%define RGB_PIXELSIZE  EXT_RGBX_PIXELSIZE
+%define jsimd_rgb_ycc_convert_avx2  jsimd_extrgbx_ycc_convert_avx2
+%include "jccolext-avx2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_BGR_RED
+%define RGB_GREEN  EXT_BGR_GREEN
+%define RGB_BLUE  EXT_BGR_BLUE
+%define RGB_PIXELSIZE  EXT_BGR_PIXELSIZE
+%define jsimd_rgb_ycc_convert_avx2  jsimd_extbgr_ycc_convert_avx2
+%include "jccolext-avx2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_BGRX_RED
+%define RGB_GREEN  EXT_BGRX_GREEN
+%define RGB_BLUE  EXT_BGRX_BLUE
+%define RGB_PIXELSIZE  EXT_BGRX_PIXELSIZE
+%define jsimd_rgb_ycc_convert_avx2  jsimd_extbgrx_ycc_convert_avx2
+%include "jccolext-avx2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_XBGR_RED
+%define RGB_GREEN  EXT_XBGR_GREEN
+%define RGB_BLUE  EXT_XBGR_BLUE
+%define RGB_PIXELSIZE  EXT_XBGR_PIXELSIZE
+%define jsimd_rgb_ycc_convert_avx2  jsimd_extxbgr_ycc_convert_avx2
+%include "jccolext-avx2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_XRGB_RED
+%define RGB_GREEN  EXT_XRGB_GREEN
+%define RGB_BLUE  EXT_XRGB_BLUE
+%define RGB_PIXELSIZE  EXT_XRGB_PIXELSIZE
+%define jsimd_rgb_ycc_convert_avx2  jsimd_extxrgb_ycc_convert_avx2
+%include "jccolext-avx2.asm"
diff --git a/media/libjpeg/simd/x86_64/jccolor-sse2.asm b/media/libjpeg/simd/x86_64/jccolor-sse2.asm
new file mode 100644
index 0000000000..e2955c2134
--- /dev/null
+++ b/media/libjpeg/simd/x86_64/jccolor-sse2.asm
@@ -0,0 +1,120 @@
+;
+; jccolor.asm - colorspace conversion (64-bit SSE2)
+;
+; Copyright (C) 2009, 2016, D. R. Commander.
+;
+; Based on the x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+
+%include "jsimdext.inc"
+
+; --------------------------------------------------------------------------
+
+%define SCALEBITS  16
+
+F_0_081 equ  5329                ; FIX(0.08131)
+F_0_114 equ  7471                ; FIX(0.11400)
+F_0_168 equ 11059                ; FIX(0.16874)
+F_0_250 equ 16384                ; FIX(0.25000)
+F_0_299 equ 19595                ; FIX(0.29900)
+F_0_331 equ 21709                ; FIX(0.33126)
+F_0_418 equ 27439                ; FIX(0.41869)
+F_0_587 equ 38470                ; FIX(0.58700)
+F_0_337 equ (F_0_587 - F_0_250)  ; FIX(0.58700) - FIX(0.25000)
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_CONST
+
+    alignz      32
+    GLOBAL_DATA(jconst_rgb_ycc_convert_sse2)
+
+EXTN(jconst_rgb_ycc_convert_sse2):
+
+PW_F0299_F0337  times 4 dw  F_0_299,  F_0_337
+PW_F0114_F0250  times 4 dw  F_0_114,  F_0_250
+PW_MF016_MF033  times 4 dw -F_0_168, -F_0_331
+PW_MF008_MF041  times 4 dw -F_0_081, -F_0_418
+PD_ONEHALFM1_CJ times 4 dd  (1 << (SCALEBITS - 1)) - 1 + \
+                            (CENTERJSAMPLE << SCALEBITS)
+PD_ONEHALF      times 4 dd  (1 << (SCALEBITS - 1))
+
+    alignz      32
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_TEXT
+    BITS        64
+
+%include "jccolext-sse2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_RGB_RED
+%define RGB_GREEN  EXT_RGB_GREEN
+%define RGB_BLUE  EXT_RGB_BLUE
+%define RGB_PIXELSIZE  EXT_RGB_PIXELSIZE
+%define jsimd_rgb_ycc_convert_sse2  jsimd_extrgb_ycc_convert_sse2
+%include "jccolext-sse2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_RGBX_RED
+%define RGB_GREEN  EXT_RGBX_GREEN
+%define RGB_BLUE  EXT_RGBX_BLUE
+%define RGB_PIXELSIZE  EXT_RGBX_PIXELSIZE
+%define jsimd_rgb_ycc_convert_sse2  jsimd_extrgbx_ycc_convert_sse2
+%include "jccolext-sse2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_BGR_RED
+%define RGB_GREEN  EXT_BGR_GREEN
+%define RGB_BLUE  EXT_BGR_BLUE
+%define RGB_PIXELSIZE  EXT_BGR_PIXELSIZE
+%define jsimd_rgb_ycc_convert_sse2  jsimd_extbgr_ycc_convert_sse2
+%include "jccolext-sse2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_BGRX_RED
+%define RGB_GREEN  EXT_BGRX_GREEN
+%define RGB_BLUE  EXT_BGRX_BLUE
+%define RGB_PIXELSIZE  EXT_BGRX_PIXELSIZE
+%define jsimd_rgb_ycc_convert_sse2  jsimd_extbgrx_ycc_convert_sse2
+%include "jccolext-sse2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_XBGR_RED
+%define RGB_GREEN  EXT_XBGR_GREEN
+%define RGB_BLUE  EXT_XBGR_BLUE
+%define RGB_PIXELSIZE  EXT_XBGR_PIXELSIZE
+%define jsimd_rgb_ycc_convert_sse2  jsimd_extxbgr_ycc_convert_sse2
+%include "jccolext-sse2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_XRGB_RED
+%define RGB_GREEN  EXT_XRGB_GREEN
+%define RGB_BLUE  EXT_XRGB_BLUE
+%define RGB_PIXELSIZE  EXT_XRGB_PIXELSIZE
+%define jsimd_rgb_ycc_convert_sse2  jsimd_extxrgb_ycc_convert_sse2
+%include "jccolext-sse2.asm"
diff --git a/media/libjpeg/simd/x86_64/jcgray-avx2.asm b/media/libjpeg/simd/x86_64/jcgray-avx2.asm
new file mode 100644
index 0000000000..591255bb11
--- /dev/null
+++ b/media/libjpeg/simd/x86_64/jcgray-avx2.asm
@@ -0,0 +1,113 @@
+;
+; jcgray.asm - grayscale colorspace conversion (64-bit AVX2)
+;
+; Copyright (C) 2011, 2016, D. R. Commander.
+; Copyright (C) 2015, Intel Corporation.
+;
+; Based on the x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+
+%include "jsimdext.inc"
+
+; --------------------------------------------------------------------------
+
+%define SCALEBITS  16
+
+F_0_114 equ  7471                ; FIX(0.11400)
+F_0_250 equ 16384                ; FIX(0.25000)
+F_0_299 equ 19595                ; FIX(0.29900)
+F_0_587 equ 38470                ; FIX(0.58700)
+F_0_337 equ (F_0_587 - F_0_250)  ; FIX(0.58700) - FIX(0.25000)
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_CONST
+
+    alignz      32
+    GLOBAL_DATA(jconst_rgb_gray_convert_avx2)
+
+EXTN(jconst_rgb_gray_convert_avx2):
+
+PW_F0299_F0337 times 8 dw F_0_299, F_0_337
+PW_F0114_F0250 times 8 dw F_0_114, F_0_250
+PD_ONEHALF     times 8 dd (1 << (SCALEBITS - 1))
+
+    alignz      32
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_TEXT
+    BITS        64
+
+%include "jcgryext-avx2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_RGB_RED
+%define RGB_GREEN  EXT_RGB_GREEN
+%define RGB_BLUE  EXT_RGB_BLUE
+%define RGB_PIXELSIZE  EXT_RGB_PIXELSIZE
+%define jsimd_rgb_gray_convert_avx2  jsimd_extrgb_gray_convert_avx2
+%include "jcgryext-avx2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_RGBX_RED
+%define RGB_GREEN  EXT_RGBX_GREEN
+%define RGB_BLUE  EXT_RGBX_BLUE
+%define RGB_PIXELSIZE  EXT_RGBX_PIXELSIZE
+%define jsimd_rgb_gray_convert_avx2  jsimd_extrgbx_gray_convert_avx2
+%include "jcgryext-avx2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_BGR_RED
+%define RGB_GREEN  EXT_BGR_GREEN
+%define RGB_BLUE  EXT_BGR_BLUE
+%define RGB_PIXELSIZE  EXT_BGR_PIXELSIZE
+%define jsimd_rgb_gray_convert_avx2  jsimd_extbgr_gray_convert_avx2
+%include "jcgryext-avx2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_BGRX_RED
+%define RGB_GREEN  EXT_BGRX_GREEN
+%define RGB_BLUE  EXT_BGRX_BLUE
+%define RGB_PIXELSIZE  EXT_BGRX_PIXELSIZE
+%define jsimd_rgb_gray_convert_avx2  jsimd_extbgrx_gray_convert_avx2
+%include "jcgryext-avx2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_XBGR_RED
+%define RGB_GREEN  EXT_XBGR_GREEN
+%define RGB_BLUE  EXT_XBGR_BLUE
+%define RGB_PIXELSIZE  EXT_XBGR_PIXELSIZE
+%define jsimd_rgb_gray_convert_avx2  jsimd_extxbgr_gray_convert_avx2
+%include "jcgryext-avx2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_XRGB_RED
+%define RGB_GREEN  EXT_XRGB_GREEN
+%define RGB_BLUE  EXT_XRGB_BLUE
+%define RGB_PIXELSIZE  EXT_XRGB_PIXELSIZE
+%define jsimd_rgb_gray_convert_avx2  jsimd_extxrgb_gray_convert_avx2
+%include "jcgryext-avx2.asm"
diff --git a/media/libjpeg/simd/x86_64/jcgray-sse2.asm b/media/libjpeg/simd/x86_64/jcgray-sse2.asm
new file mode 100644
index 0000000000..e389904f2f
--- /dev/null
+++ b/media/libjpeg/simd/x86_64/jcgray-sse2.asm
@@ -0,0 +1,112 @@
+;
+; jcgray.asm - grayscale colorspace conversion (64-bit SSE2)
+;
+; Copyright (C) 2011, 2016, D. R. Commander.
+;
+; Based on the x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+
+%include "jsimdext.inc"
+
+; --------------------------------------------------------------------------
+
+%define SCALEBITS  16
+
+F_0_114 equ  7471                ; FIX(0.11400)
+F_0_250 equ 16384                ; FIX(0.25000)
+F_0_299 equ 19595                ; FIX(0.29900)
+F_0_587 equ 38470                ; FIX(0.58700)
+F_0_337 equ (F_0_587 - F_0_250)  ; FIX(0.58700) - FIX(0.25000)
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_CONST
+
+    alignz      32
+    GLOBAL_DATA(jconst_rgb_gray_convert_sse2)
+
+EXTN(jconst_rgb_gray_convert_sse2):
+
+PW_F0299_F0337 times 4 dw F_0_299, F_0_337
+PW_F0114_F0250 times 4 dw F_0_114, F_0_250
+PD_ONEHALF     times 4 dd (1 << (SCALEBITS - 1))
+
+    alignz      32
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_TEXT
+    BITS        64
+
+%include "jcgryext-sse2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_RGB_RED
+%define RGB_GREEN  EXT_RGB_GREEN
+%define RGB_BLUE  EXT_RGB_BLUE
+%define RGB_PIXELSIZE  EXT_RGB_PIXELSIZE
+%define jsimd_rgb_gray_convert_sse2  jsimd_extrgb_gray_convert_sse2
+%include "jcgryext-sse2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_RGBX_RED
+%define RGB_GREEN  EXT_RGBX_GREEN
+%define RGB_BLUE  EXT_RGBX_BLUE
+%define RGB_PIXELSIZE  EXT_RGBX_PIXELSIZE
+%define jsimd_rgb_gray_convert_sse2  jsimd_extrgbx_gray_convert_sse2
+%include "jcgryext-sse2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_BGR_RED
+%define RGB_GREEN  EXT_BGR_GREEN
+%define RGB_BLUE  EXT_BGR_BLUE
+%define RGB_PIXELSIZE  EXT_BGR_PIXELSIZE
+%define jsimd_rgb_gray_convert_sse2  jsimd_extbgr_gray_convert_sse2
+%include "jcgryext-sse2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_BGRX_RED
+%define RGB_GREEN  EXT_BGRX_GREEN
+%define RGB_BLUE  EXT_BGRX_BLUE
+%define RGB_PIXELSIZE  EXT_BGRX_PIXELSIZE
+%define jsimd_rgb_gray_convert_sse2  jsimd_extbgrx_gray_convert_sse2
+%include "jcgryext-sse2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_XBGR_RED
+%define RGB_GREEN  EXT_XBGR_GREEN
+%define RGB_BLUE  EXT_XBGR_BLUE
+%define RGB_PIXELSIZE  EXT_XBGR_PIXELSIZE
+%define jsimd_rgb_gray_convert_sse2  jsimd_extxbgr_gray_convert_sse2
+%include "jcgryext-sse2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_XRGB_RED
+%define RGB_GREEN  EXT_XRGB_GREEN
+%define RGB_BLUE  EXT_XRGB_BLUE
+%define RGB_PIXELSIZE  EXT_XRGB_PIXELSIZE
+%define jsimd_rgb_gray_convert_sse2  jsimd_extxrgb_gray_convert_sse2
+%include "jcgryext-sse2.asm"
diff --git a/media/libjpeg/simd/x86_64/jcgryext-avx2.asm b/media/libjpeg/simd/x86_64/jcgryext-avx2.asm
new file mode 100644
index 0000000000..ddcc2c0a2f
--- /dev/null
+++ b/media/libjpeg/simd/x86_64/jcgryext-avx2.asm
@@ -0,0 +1,438 @@
+;
+; jcgryext.asm - grayscale colorspace conversion (64-bit AVX2)
+;
+; Copyright (C) 2011, 2016, D. R. Commander.
+; Copyright (C) 2015, Intel Corporation.
+; Copyright (C) 2018, Matthias Räncker.
+;
+; Based on the x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+
+%include "jcolsamp.inc"
+
+; --------------------------------------------------------------------------
+;
+; Convert some rows of samples to the output colorspace.
+;
+; GLOBAL(void)
+; jsimd_rgb_gray_convert_avx2(JDIMENSION img_width, JSAMPARRAY input_buf,
+;                             JSAMPIMAGE output_buf, JDIMENSION output_row,
+;                             int num_rows);
+;
+
+; r10d = JDIMENSION img_width
+; r11 = JSAMPARRAY input_buf
+; r12 = JSAMPIMAGE output_buf
+; r13d = JDIMENSION output_row
+; r14d = int num_rows
+
+%define wk(i)   rbp - (WK_NUM - (i)) * SIZEOF_YMMWORD  ; ymmword wk[WK_NUM]
+%define WK_NUM  2
+
+    align       32
+    GLOBAL_FUNCTION(jsimd_rgb_gray_convert_avx2)
+
+EXTN(jsimd_rgb_gray_convert_avx2):
+    push        rbp
+    mov         rax, rsp                     ; rax = original rbp
+    sub         rsp, byte 4
+    and         rsp, byte (-SIZEOF_YMMWORD)  ; align to 256 bits
+    mov         [rsp], rax
+    mov         rbp, rsp                     ; rbp = aligned rbp
+    lea         rsp, [wk(0)]
+    collect_args 5
+    push        rbx
+
+    mov         ecx, r10d
+    test        rcx, rcx
+    jz          near .return
+
+    push        rcx
+
+    mov         rsi, r12
+    mov         ecx, r13d
+    mov         rdip, JSAMPARRAY [rsi+0*SIZEOF_JSAMPARRAY]
+    lea         rdi, [rdi+rcx*SIZEOF_JSAMPROW]
+
+    pop         rcx
+
+    mov         rsi, r11
+    mov         eax, r14d
+    test        rax, rax
+    jle         near .return
+.rowloop:
+    push        rdi
+    push        rsi
+    push        rcx                     ; col
+
+    mov         rsip, JSAMPROW [rsi]    ; inptr
+    mov         rdip, JSAMPROW [rdi]    ; outptr0
+
+    cmp         rcx, byte SIZEOF_YMMWORD
+    jae         near .columnloop
+
+%if RGB_PIXELSIZE == 3  ; ---------------
+
+.column_ld1:
+    push        rax
+    push        rdx
+    lea         rcx, [rcx+rcx*2]        ; imul ecx,RGB_PIXELSIZE
+    test        cl, SIZEOF_BYTE
+    jz          short .column_ld2
+    sub         rcx, byte SIZEOF_BYTE
+    movzx       rax, byte [rsi+rcx]
+.column_ld2:
+    test        cl, SIZEOF_WORD
+    jz          short .column_ld4
+    sub         rcx, byte SIZEOF_WORD
+    movzx       rdx, word [rsi+rcx]
+    shl         rax, WORD_BIT
+    or          rax, rdx
+.column_ld4:
+    vmovd       xmmA, eax
+    pop         rdx
+    pop         rax
+    test        cl, SIZEOF_DWORD
+    jz          short .column_ld8
+    sub         rcx, byte SIZEOF_DWORD
+    vmovd       xmmF, XMM_DWORD [rsi+rcx]
+    vpslldq     xmmA, xmmA, SIZEOF_DWORD
+    vpor        xmmA, xmmA, xmmF
+.column_ld8:
+    test        cl, SIZEOF_MMWORD
+    jz          short .column_ld16
+    sub         rcx, byte SIZEOF_MMWORD
+    vmovq       xmmB, XMM_MMWORD [rsi+rcx]
+    vpslldq     xmmA, xmmA, SIZEOF_MMWORD
+    vpor        xmmA, xmmA, xmmB
+.column_ld16:
+    test        cl, SIZEOF_XMMWORD
+    jz          short .column_ld32
+    sub         rcx, byte SIZEOF_XMMWORD
+    vmovdqu     xmmB, XMM_MMWORD [rsi+rcx]
+    vperm2i128  ymmA, ymmA, ymmA, 1
+    vpor        ymmA, ymmB
+.column_ld32:
+    test        cl, SIZEOF_YMMWORD
+    jz          short .column_ld64
+    sub         rcx, byte SIZEOF_YMMWORD
+    vmovdqa     ymmF, ymmA
+    vmovdqu     ymmA, YMMWORD [rsi+0*SIZEOF_YMMWORD]
+.column_ld64:
+    test        cl, 2*SIZEOF_YMMWORD
+    mov         rcx, SIZEOF_YMMWORD
+    jz          short .rgb_gray_cnv
+    vmovdqa     ymmB, ymmA
+    vmovdqu     ymmA, YMMWORD [rsi+0*SIZEOF_YMMWORD]
+    vmovdqu     ymmF, YMMWORD [rsi+1*SIZEOF_YMMWORD]
+    jmp         short .rgb_gray_cnv
+
+.columnloop:
+    vmovdqu     ymmA, YMMWORD [rsi+0*SIZEOF_YMMWORD]
+    vmovdqu     ymmF, YMMWORD [rsi+1*SIZEOF_YMMWORD]
+    vmovdqu     ymmB, YMMWORD [rsi+2*SIZEOF_YMMWORD]
+
+.rgb_gray_cnv:
+    ; ymmA=(00 10 20 01 11 21 02 12 22 03 13 23 04 14 24 05
+    ;       15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A)
+    ; ymmF=(2A 0B 1B 2B 0C 1C 2C 0D 1D 2D 0E 1E 2E 0F 1F 2F
+    ;       0G 1G 2G 0H 1H 2H 0I 1I 2I 0J 1J 2J 0K 1K 2K 0L)
+    ; ymmB=(1L 2L 0M 1M 2M 0N 1N 2N 0O 1O 2O 0P 1P 2P 0Q 1Q
+    ;       2Q 0R 1R 2R 0S 1S 2S 0T 1T 2T 0U 1U 2U 0V 1V 2V)
+
+    vmovdqu     ymmC, ymmA
+    vinserti128 ymmA, ymmF, xmmA, 0  ; ymmA=(00 10 20 01 11 21 02 12 22 03 13 23 04 14 24 05
+                                     ;       0G 1G 2G 0H 1H 2H 0I 1I 2I 0J 1J 2J 0K 1K 2K 0L)
+    vinserti128 ymmC, ymmC, xmmB, 0  ; ymmC=(1L 2L 0M 1M 2M 0N 1N 2N 0O 1O 2O 0P 1P 2P 0Q 1Q
+                                     ;       15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A)
+    vinserti128 ymmB, ymmB, xmmF, 0  ; ymmB=(2A 0B 1B 2B 0C 1C 2C 0D 1D 2D 0E 1E 2E 0F 1F 2F
+                                     ;       2Q 0R 1R 2R 0S 1S 2S 0T 1T 2T 0U 1U 2U 0V 1V 2V)
+    vperm2i128  ymmF, ymmC, ymmC, 1  ; ymmF=(15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A
+                                     ;       1L 2L 0M 1M 2M 0N 1N 2N 0O 1O 2O 0P 1P 2P 0Q 1Q)
+
+    vmovdqa     ymmG, ymmA
+    vpslldq     ymmA, ymmA, 8     ; ymmA=(-- -- -- -- -- -- -- -- 00 10 20 01 11 21 02 12
+                                  ;       22 03 13 23 04 14 24 05 0G 1G 2G 0H 1H 2H 0I 1I)
+    vpsrldq     ymmG, ymmG, 8     ; ymmG=(22 03 13 23 04 14 24 05 0G 1G 2G 0H 1H 2H 0I 1I
+                                  ;       2I 0J 1J 2J 0K 1K 2K 0L -- -- -- -- -- -- -- --)
+
+    vpunpckhbw  ymmA, ymmA, ymmF  ; ymmA=(00 08 10 18 20 28 01 09 11 19 21 29 02 0A 12 1A
+                                  ;       0G 0O 1G 1O 2G 2O 0H 0P 1H 1P 2H 2P 0I 0Q 1I 1Q)
+    vpslldq     ymmF, ymmF, 8     ; ymmF=(-- -- -- -- -- -- -- -- 15 25 06 16 26 07 17 27
+                                  ;       08 18 28 09 19 29 0A 1A 1L 2L 0M 1M 2M 0N 1N 2N)
+
+    vpunpcklbw  ymmG, ymmG, ymmB  ; ymmG=(22 2A 03 0B 13 1B 23 2B 04 0C 14 1C 24 2C 05 0D
+                                  ;       2I 2Q 0J 0R 1J 1R 2J 2R 0K 0S 1K 1S 2K 2S 0L 0T)
+    vpunpckhbw  ymmF, ymmF, ymmB  ; ymmF=(15 1D 25 2D 06 0E 16 1E 26 2E 07 0F 17 1F 27 2F
+                                  ;       1L 1T 2L 2T 0M 0U 1M 1U 2M 2U 0N 0V 1N 1V 2N 2V)
+
+    vmovdqa     ymmD, ymmA
+    vpslldq     ymmA, ymmA, 8     ; ymmA=(-- -- -- -- -- -- -- -- 00 08 10 18 20 28 01 09
+                                  ;       11 19 21 29 02 0A 12 1A 0G 0O 1G 1O 2G 2O 0H 0P)
+    vpsrldq     ymmD, ymmD, 8     ; ymmD=(11 19 21 29 02 0A 12 1A 0G 0O 1G 1O 2G 2O 0H 0P
+                                  ;       1H 1P 2H 2P 0I 0Q 1I 1Q -- -- -- -- -- -- -- --)
+
+    vpunpckhbw  ymmA, ymmA, ymmG  ; ymmA=(00 04 08 0C 10 14 18 1C 20 24 28 2C 01 05 09 0D
+                                  ;       0G 0K 0O 0S 1G 1K 1O 1S 2G 2K 2O 2S 0H 0L 0P 0T)
+    vpslldq     ymmG, ymmG, 8     ; ymmG=(-- -- -- -- -- -- -- -- 22 2A 03 0B 13 1B 23 2B
+                                  ;       04 0C 14 1C 24 2C 05 0D 2I 2Q 0J 0R 1J 1R 2J 2R)
+
+    vpunpcklbw  ymmD, ymmD, ymmF  ; ymmD=(11 15 19 1D 21 25 29 2D 02 06 0A 0E 12 16 1A 1E
+                                  ;       1H 1L 1P 1T 2H 2L 2P 2T 0I 0M 0Q 0U 1I 1M 1Q 1U)
+    vpunpckhbw  ymmG, ymmG, ymmF  ; ymmG=(22 26 2A 2E 03 07 0B 0F 13 17 1B 1F 23 27 2B 2F
+                                  ;       2I 2M 2Q 2U 0J 0N 0R 0V 1J 1N 1R 1V 2J 2N 2R 2V)
+
+    vmovdqa     ymmE, ymmA
+    vpslldq     ymmA, ymmA, 8     ; ymmA=(-- -- -- -- -- -- -- -- 00 04 08 0C 10 14 18 1C
+                                  ;       20 24 28 2C 01 05 09 0D 0G 0K 0O 0S 1G 1K 1O 1S)
+    vpsrldq     ymmE, ymmE, 8     ; ymmE=(20 24 28 2C 01 05 09 0D 0G 0K 0O 0S 1G 1K 1O 1S
+                                  ;       2G 2K 2O 2S 0H 0L 0P 0T -- -- -- -- -- -- -- --)
+
+    vpunpckhbw  ymmA, ymmA, ymmD  ; ymmA=(00 02 04 06 08 0A 0C 0E 10 12 14 16 18 1A 1C 1E
+                                  ;       0G 0I 0K 0M 0O 0Q 0S 0U 1G 1I 1K 1M 1O 1Q 1S 1U)
+    vpslldq     ymmD, ymmD, 8     ; ymmD=(-- -- -- -- -- -- -- -- 11 15 19 1D 21 25 29 2D
+                                  ;       02 06 0A 0E 12 16 1A 1E 1H 1L 1P 1T 2H 2L 2P 2T)
+
+    vpunpcklbw  ymmE, ymmE, ymmG  ; ymmE=(20 22 24 26 28 2A 2C 2E 01 03 05 07 09 0B 0D 0F
+                                  ;       2G 2I 2K 2M 2O 2Q 2S 2U 0H 0J 0L 0N 0P 0R 0T 0V)
+    vpunpckhbw  ymmD, ymmD, ymmG  ; ymmD=(11 13 15 17 19 1B 1D 1F 21 23 25 27 29 2B 2D 2F
+                                  ;       1H 1J 1L 1N 1P 1R 1T 1V 2H 2J 2L 2N 2P 2R 2T 2V)
+
+    vpxor       ymmH, ymmH, ymmH
+
+    vmovdqa     ymmC, ymmA
+    vpunpcklbw  ymmA, ymmA, ymmH  ; ymmA=(00 02 04 06 08 0A 0C 0E 0G 0I 0K 0M 0O 0Q 0S 0U)
+    vpunpckhbw  ymmC, ymmC, ymmH  ; ymmC=(10 12 14 16 18 1A 1C 1E 1G 1I 1K 1M 1O 1Q 1S 1U)
+
+    vmovdqa     ymmB, ymmE
+    vpunpcklbw  ymmE, ymmE, ymmH  ; ymmE=(20 22 24 26 28 2A 2C 2E 2G 2I 2K 2M 2O 2Q 2S 2U)
+    vpunpckhbw  ymmB, ymmB, ymmH  ; ymmB=(01 03 05 07 09 0B 0D 0F 0H 0J 0L 0N 0P 0R 0T 0V)
+
+    vmovdqa     ymmF, ymmD
+    vpunpcklbw  ymmD, ymmD, ymmH  ; ymmD=(11 13 15 17 19 1B 1D 1F 1H 1J 1L 1N 1P 1R 1T 1V)
+    vpunpckhbw  ymmF, ymmF, ymmH  ; ymmF=(21 23 25 27 29 2B 2D 2F 2H 2J 2L 2N 2P 2R 2T 2V)
+
+%else  ; RGB_PIXELSIZE == 4 ; -----------
+
+.column_ld1:
+    test        cl, SIZEOF_XMMWORD/16
+    jz          short .column_ld2
+    sub         rcx, byte SIZEOF_XMMWORD/16
+    vmovd       xmmA, XMM_DWORD [rsi+rcx*RGB_PIXELSIZE]
+.column_ld2:
+    test        cl, SIZEOF_XMMWORD/8
+    jz          short .column_ld4
+    sub         rcx, byte SIZEOF_XMMWORD/8
+    vmovq       xmmF, XMM_MMWORD [rsi+rcx*RGB_PIXELSIZE]
+    vpslldq     xmmA, xmmA, SIZEOF_MMWORD
+    vpor        xmmA, xmmA, xmmF
+.column_ld4:
+    test        cl, SIZEOF_XMMWORD/4
+    jz          short .column_ld8
+    sub         rcx, byte SIZEOF_XMMWORD/4
+    vmovdqa     xmmF, xmmA
+    vperm2i128  ymmF, ymmF, ymmF, 1
+    vmovdqu     xmmA, XMMWORD [rsi+rcx*RGB_PIXELSIZE]
+    vpor        ymmA, ymmA, ymmF
+.column_ld8:
+    test        cl, SIZEOF_XMMWORD/2
+    jz          short .column_ld16
+    sub         rcx, byte SIZEOF_XMMWORD/2
+    vmovdqa     ymmF, ymmA
+    vmovdqu     ymmA, YMMWORD [rsi+rcx*RGB_PIXELSIZE]
+.column_ld16:
+    test        cl, SIZEOF_XMMWORD
+    mov         rcx, SIZEOF_YMMWORD
+    jz          short .rgb_gray_cnv
+    vmovdqa     ymmE, ymmA
+    vmovdqa     ymmH, ymmF
+    vmovdqu     ymmA, YMMWORD [rsi+0*SIZEOF_YMMWORD]
+    vmovdqu     ymmF, YMMWORD [rsi+1*SIZEOF_YMMWORD]
+    jmp         short .rgb_gray_cnv
+
+.columnloop:
+    vmovdqu     ymmA, YMMWORD [rsi+0*SIZEOF_YMMWORD]
+    vmovdqu     ymmF, YMMWORD [rsi+1*SIZEOF_YMMWORD]
+    vmovdqu     ymmE, YMMWORD [rsi+2*SIZEOF_YMMWORD]
+    vmovdqu     ymmH, YMMWORD [rsi+3*SIZEOF_YMMWORD]
+
+.rgb_gray_cnv:
+    ; ymmA=(00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33
+    ;       04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37)
+    ; ymmF=(08 18 28 38 09 19 29 39 0A 1A 2A 3A 0B 1B 2B 3B
+    ;       0C 1C 2C 3C 0D 1D 2D 3D 0E 1E 2E 3E 0F 1F 2F 3F)
+    ; ymmE=(0G 1G 2G 3G 0H 1H 2H 3H 0I 1I 2I 3I 0J 1J 2J 3J
+    ;       0K 1K 2K 3K 0L 1L 2L 3L 0M 1M 2M 3M 0N 1N 2N 3N)
+    ; ymmH=(0O 1O 2O 3O 0P 1P 2P 3P 0Q 1Q 2Q 3Q 0R 1R 2R 3R
+    ;       0S 1S 2S 3S 0T 1T 2T 3T 0U 1U 2U 3U 0V 1V 2V 3V)
+
+    vmovdqa     ymmB, ymmA
+    vinserti128 ymmA, ymmA, xmmE, 1     ; ymmA=(00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33
+                                        ;       0G 1G 2G 3G 0H 1H 2H 3H 0I 1I 2I 3I 0J 1J 2J 3J)
+    vperm2i128  ymmE, ymmB, ymmE, 0x31  ; ymmE=(04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37
+                                        ;       0K 1K 2K 3K 0L 1L 2L 3L 0M 1M 2M 3M 0N 1N 2N 3N)
+
+    vmovdqa     ymmB, ymmF
+    vinserti128 ymmF, ymmF, xmmH, 1     ; ymmF=(08 18 28 38 09 19 29 39 0A 1A 2A 3A 0B 1B 2B 3B
+                                        ;       0O 1O 2O 3O 0P 1P 2P 3P 0Q 1Q 2Q 3Q 0R 1R 2R 3R)
+    vperm2i128  ymmH, ymmB, ymmH, 0x31  ; ymmH=(0C 1C 2C 3C 0D 1D 2D 3D 0E 1E 2E 3E 0F 1F 2F 3F
+                                        ;       0S 1S 2S 3S 0T 1T 2T 3T 0U 1U 2U 3U 0V 1V 2V 3V)
+
+    vmovdqa     ymmD, ymmA
+    vpunpcklbw  ymmA, ymmA, ymmE      ; ymmA=(00 04 10 14 20 24 30 34 01 05 11 15 21 25 31 35
+                                      ;       0G 0K 1G 1K 2G 2K 3G 3K 0H 0L 1H 1L 2H 2L 3H 3L)
+    vpunpckhbw  ymmD, ymmD, ymmE      ; ymmD=(02 06 12 16 22 26 32 36 03 07 13 17 23 27 33 37
+                                      ;       0I 0M 1I 1M 2I 2M 3I 3M 0J 0N 1J 1N 2J 2N 3J 3N)
+
+    vmovdqa     ymmC, ymmF
+    vpunpcklbw  ymmF, ymmF, ymmH      ; ymmF=(08 0C 18 1C 28 2C 38 3C 09 0D 19 1D 29 2D 39 3D
+                                      ;       0O 0S 1O 1S 2O 2S 3O 3S 0P 0T 1P 1T 2P 2T 3P 3T)
+    vpunpckhbw  ymmC, ymmC, ymmH      ; ymmC=(0A 0E 1A 1E 2A 2E 3A 3E 0B 0F 1B 1F 2B 2F 3B 3F
+                                      ;       0Q 0U 1Q 1U 2Q 2U 3Q 3U 0R 0V 1R 1V 2R 2V 3R 3V)
+
+    vmovdqa     ymmB, ymmA
+    vpunpcklwd  ymmA, ymmA, ymmF      ; ymmA=(00 04 08 0C 10 14 18 1C 20 24 28 2C 30 34 38 3C
+                                      ;       0G 0K 0O 0S 1G 1K 1O 1S 2G 2K 2O 2S 3G 3K 3O 3S)
+    vpunpckhwd  ymmB, ymmB, ymmF      ; ymmB=(01 05 09 0D 11 15 19 1D 21 25 29 2D 31 35 39 3D
+                                      ;       0H 0L 0P 0T 1H 1L 1P 1T 2H 2L 2P 2T 3H 3L 3P 3T)
+
+    vmovdqa     ymmG, ymmD
+    vpunpcklwd  ymmD, ymmD, ymmC      ; ymmD=(02 06 0A 0E 12 16 1A 1E 22 26 2A 2E 32 36 3A 3E
+                                      ;       0I 0M 0Q 0U 1I 1M 1Q 1U 2I 2M 2Q 2U 3I 3M 3Q 3U)
+    vpunpckhwd  ymmG, ymmG, ymmC      ; ymmG=(03 07 0B 0F 13 17 1B 1F 23 27 2B 2F 33 37 3B 3F
+                                      ;       0J 0N 0R 0V 1J 1N 1R 1V 2J 2N 2R 2V 3J 3N 3R 3V)
+
+    vmovdqa     ymmE, ymmA
+    vpunpcklbw  ymmA, ymmA, ymmD      ; ymmA=(00 02 04 06 08 0A 0C 0E 10 12 14 16 18 1A 1C 1E
+                                      ;       0G 0I 0K 0M 0O 0Q 0S 0U 1G 1I 1K 1M 1O 1Q 1S 1U)
+    vpunpckhbw  ymmE, ymmE, ymmD      ; ymmE=(20 22 24 26 28 2A 2C 2E 30 32 34 36 38 3A 3C 3E
+                                      ;       2G 2I 2K 2M 2O 2Q 2S 2U 3G 3I 3K 3M 3O 3Q 3S 3U)
+
+    vmovdqa     ymmH, ymmB
+    vpunpcklbw  ymmB, ymmB, ymmG      ; ymmB=(01 03 05 07 09 0B 0D 0F 11 13 15 17 19 1B 1D 1F
+                                      ;       0H 0J 0L 0N 0P 0R 0T 0V 1H 1J 1L 1N 1P 1R 1T 1V)
+    vpunpckhbw  ymmH, ymmH, ymmG      ; ymmH=(21 23 25 27 29 2B 2D 2F 31 33 35 37 39 3B 3D 3F
+                                      ;       2H 2J 2L 2N 2P 2R 2T 2V 3H 3J 3L 3N 3P 3R 3T 3V)
+
+    vpxor       ymmF, ymmF, ymmF
+
+    vmovdqa     ymmC, ymmA
+    vpunpcklbw  ymmA, ymmA, ymmF      ; ymmA=(00 02 04 06 08 0A 0C 0E 0G 0I 0K 0M 0O 0Q 0S 0U)
+    vpunpckhbw  ymmC, ymmC, ymmF      ; ymmC=(10 12 14 16 18 1A 1C 1E 1G 1I 1K 1M 1O 1Q 1S 1U)
+
+    vmovdqa     ymmD, ymmB
+    vpunpcklbw  ymmB, ymmB, ymmF      ; ymmB=(01 03 05 07 09 0B 0D 0F 0H 0J 0L 0N 0P 0R 0T 0V)
+    vpunpckhbw  ymmD, ymmD, ymmF      ; ymmD=(11 13 15 17 19 1B 1D 1F 1H 1J 1L 1N 1P 1R 1T 1V)
+
+    vmovdqa     ymmG, ymmE
+    vpunpcklbw  ymmE, ymmE, ymmF      ; ymmE=(20 22 24 26 28 2A 2C 2E 2G 2I 2K 2M 2O 2Q 2S 2U)
+    vpunpckhbw  ymmG, ymmG, ymmF      ; ymmG=(30 32 34 36 38 3A 3C 3E 3G 3I 3K 3M 3O 3Q 3S 3U)
+
+    vpunpcklbw  ymmF, ymmF, ymmH
+    vpunpckhbw  ymmH, ymmH, ymmH
+    vpsrlw      ymmF, ymmF, BYTE_BIT  ; ymmF=(21 23 25 27 29 2B 2D 2F 2H 2J 2L 2N 2P 2R 2T 2V)
+    vpsrlw      ymmH, ymmH, BYTE_BIT  ; ymmH=(31 33 35 37 39 3B 3D 3F 3H 3J 3L 3N 3P 3R 3T 3V)
+
+%endif  ; RGB_PIXELSIZE ; ---------------
+
+    ; ymm0=R(02468ACEGIKMOQSU)=RE, ymm2=G(02468ACEGIKMOQSU)=GE, ymm4=B(02468ACEGIKMOQSU)=BE
+    ; ymm1=R(13579BDFHJLNPRTV)=RO, ymm3=G(13579BDFHJLNPRTV)=GO, ymm5=B(13579BDFHJLNPRTV)=BO
+
+    ; (Original)
+    ; Y  =  0.29900 * R + 0.58700 * G + 0.11400 * B
+    ;
+    ; (This implementation)
+    ; Y  =  0.29900 * R + 0.33700 * G + 0.11400 * B + 0.25000 * G
+
+    vmovdqa     ymm6, ymm1
+    vpunpcklwd  ymm1, ymm1, ymm3
+    vpunpckhwd  ymm6, ymm6, ymm3
+    vpmaddwd    ymm1, ymm1, [rel PW_F0299_F0337]  ; ymm1=ROL*FIX(0.299)+GOL*FIX(0.337)
+    vpmaddwd    ymm6, ymm6, [rel PW_F0299_F0337]  ; ymm6=ROH*FIX(0.299)+GOH*FIX(0.337)
+
+    vmovdqa     ymm7, ymm6              ; ymm7=ROH*FIX(0.299)+GOH*FIX(0.337)
+
+    vmovdqa     ymm6, ymm0
+    vpunpcklwd  ymm0, ymm0, ymm2
+    vpunpckhwd  ymm6, ymm6, ymm2
+    vpmaddwd    ymm0, ymm0, [rel PW_F0299_F0337]  ; ymm0=REL*FIX(0.299)+GEL*FIX(0.337)
+    vpmaddwd    ymm6, ymm6, [rel PW_F0299_F0337]  ; ymm6=REH*FIX(0.299)+GEH*FIX(0.337)
+
+    vmovdqa     YMMWORD [wk(0)], ymm0   ; wk(0)=REL*FIX(0.299)+GEL*FIX(0.337)
+    vmovdqa     YMMWORD [wk(1)], ymm6   ; wk(1)=REH*FIX(0.299)+GEH*FIX(0.337)
+
+    vmovdqa     ymm0, ymm5              ; ymm0=BO
+    vmovdqa     ymm6, ymm4              ; ymm6=BE
+
+    vmovdqa     ymm4, ymm0
+    vpunpcklwd  ymm0, ymm0, ymm3
+    vpunpckhwd  ymm4, ymm4, ymm3
+    vpmaddwd    ymm0, ymm0, [rel PW_F0114_F0250]  ; ymm0=BOL*FIX(0.114)+GOL*FIX(0.250)
+    vpmaddwd    ymm4, ymm4, [rel PW_F0114_F0250]  ; ymm4=BOH*FIX(0.114)+GOH*FIX(0.250)
+
+    vmovdqa     ymm3, [rel PD_ONEHALF]            ; ymm3=[PD_ONEHALF]
+
+    vpaddd      ymm0, ymm0, ymm1
+    vpaddd      ymm4, ymm4, ymm7
+    vpaddd      ymm0, ymm0, ymm3
+    vpaddd      ymm4, ymm4, ymm3
+    vpsrld      ymm0, ymm0, SCALEBITS   ; ymm0=YOL
+    vpsrld      ymm4, ymm4, SCALEBITS   ; ymm4=YOH
+    vpackssdw   ymm0, ymm0, ymm4        ; ymm0=YO
+
+    vmovdqa     ymm4, ymm6
+    vpunpcklwd  ymm6, ymm6, ymm2
+    vpunpckhwd  ymm4, ymm4, ymm2
+    vpmaddwd    ymm6, ymm6, [rel PW_F0114_F0250]  ; ymm6=BEL*FIX(0.114)+GEL*FIX(0.250)
+    vpmaddwd    ymm4, ymm4, [rel PW_F0114_F0250]  ; ymm4=BEH*FIX(0.114)+GEH*FIX(0.250)
+
+    vmovdqa     ymm2, [rel PD_ONEHALF]            ; ymm2=[PD_ONEHALF]
+
+    vpaddd      ymm6, ymm6, YMMWORD [wk(0)]
+    vpaddd      ymm4, ymm4, YMMWORD [wk(1)]
+    vpaddd      ymm6, ymm6, ymm2
+    vpaddd      ymm4, ymm4, ymm2
+    vpsrld      ymm6, ymm6, SCALEBITS   ; ymm6=YEL
+    vpsrld      ymm4, ymm4, SCALEBITS   ; ymm4=YEH
+    vpackssdw   ymm6, ymm6, ymm4        ; ymm6=YE
+
+    vpsllw      ymm0, ymm0, BYTE_BIT
+    vpor        ymm6, ymm6, ymm0        ; ymm6=Y
+    vmovdqu     YMMWORD [rdi], ymm6     ; Save Y
+
+    sub         rcx, byte SIZEOF_YMMWORD
+    add         rsi, RGB_PIXELSIZE*SIZEOF_YMMWORD  ; inptr
+    add         rdi, byte SIZEOF_YMMWORD           ; outptr0
+    cmp         rcx, byte SIZEOF_YMMWORD
+    jae         near .columnloop
+    test        rcx, rcx
+    jnz         near .column_ld1
+
+    pop         rcx                     ; col
+    pop         rsi
+    pop         rdi
+
+    add         rsi, byte SIZEOF_JSAMPROW  ; input_buf
+    add         rdi, byte SIZEOF_JSAMPROW
+    dec         rax                        ; num_rows
+    jg          near .rowloop
+
+.return:
+    pop         rbx
+    vzeroupper
+    uncollect_args 5
+    mov         rsp, rbp                ; rsp <- aligned rbp
+    pop         rsp                     ; rsp <- original rbp
+    pop         rbp
+    ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+    align       32
diff --git a/media/libjpeg/simd/x86_64/jcgryext-sse2.asm b/media/libjpeg/simd/x86_64/jcgryext-sse2.asm
new file mode 100644
index 0000000000..f1d399a63b
--- /dev/null
+++ b/media/libjpeg/simd/x86_64/jcgryext-sse2.asm
@@ -0,0 +1,363 @@
+;
+; jcgryext.asm - grayscale colorspace conversion (64-bit SSE2)
+;
+; Copyright (C) 2011, 2016, D. R. Commander.
+; Copyright (C) 2018, Matthias Räncker.
+;
+; Based on the x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+
+%include "jcolsamp.inc"
+
+; --------------------------------------------------------------------------
+;
+; Convert some rows of samples to the output colorspace.
+;
+; GLOBAL(void)
+; jsimd_rgb_gray_convert_sse2(JDIMENSION img_width, JSAMPARRAY input_buf,
+;                             JSAMPIMAGE output_buf, JDIMENSION output_row,
+;                             int num_rows);
+;
+
+; r10d = JDIMENSION img_width
+; r11 = JSAMPARRAY input_buf
+; r12 = JSAMPIMAGE output_buf
+; r13d = JDIMENSION output_row
+; r14d = int num_rows
+
+%define wk(i)   rbp - (WK_NUM - (i)) * SIZEOF_XMMWORD  ; xmmword wk[WK_NUM]
+%define WK_NUM  2
+
+    align       32
+    GLOBAL_FUNCTION(jsimd_rgb_gray_convert_sse2)
+
+EXTN(jsimd_rgb_gray_convert_sse2):
+    push        rbp
+    mov         rax, rsp                     ; rax = original rbp
+    sub         rsp, byte 4
+    and         rsp, byte (-SIZEOF_XMMWORD)  ; align to 128 bits
+    mov         [rsp], rax
+    mov         rbp, rsp                     ; rbp = aligned rbp
+    lea         rsp, [wk(0)]
+    collect_args 5
+    push        rbx
+
+    mov         ecx, r10d
+    test        rcx, rcx
+    jz          near .return
+
+    push        rcx
+
+    mov         rsi, r12
+    mov         ecx, r13d
+    mov         rdip, JSAMPARRAY [rsi+0*SIZEOF_JSAMPARRAY]
+    lea         rdi, [rdi+rcx*SIZEOF_JSAMPROW]
+
+    pop         rcx
+
+    mov         rsi, r11
+    mov         eax, r14d
+    test        rax, rax
+    jle         near .return
+.rowloop:
+    push        rdi
+    push        rsi
+    push        rcx                     ; col
+
+    mov         rsip, JSAMPROW [rsi]    ; inptr
+    mov         rdip, JSAMPROW [rdi]    ; outptr0
+
+    cmp         rcx, byte SIZEOF_XMMWORD
+    jae         near .columnloop
+
+%if RGB_PIXELSIZE == 3  ; ---------------
+
+.column_ld1:
+    push        rax
+    push        rdx
+    lea         rcx, [rcx+rcx*2]        ; imul ecx,RGB_PIXELSIZE
+    test        cl, SIZEOF_BYTE
+    jz          short .column_ld2
+    sub         rcx, byte SIZEOF_BYTE
+    movzx       rax, byte [rsi+rcx]
+.column_ld2:
+    test        cl, SIZEOF_WORD
+    jz          short .column_ld4
+    sub         rcx, byte SIZEOF_WORD
+    movzx       rdx, word [rsi+rcx]
+    shl         rax, WORD_BIT
+    or          rax, rdx
+.column_ld4:
+    movd        xmmA, eax
+    pop         rdx
+    pop         rax
+    test        cl, SIZEOF_DWORD
+    jz          short .column_ld8
+    sub         rcx, byte SIZEOF_DWORD
+    movd        xmmF, XMM_DWORD [rsi+rcx]
+    pslldq      xmmA, SIZEOF_DWORD
+    por         xmmA, xmmF
+.column_ld8:
+    test        cl, SIZEOF_MMWORD
+    jz          short .column_ld16
+    sub         rcx, byte SIZEOF_MMWORD
+    movq        xmmB, XMM_MMWORD [rsi+rcx]
+    pslldq      xmmA, SIZEOF_MMWORD
+    por         xmmA, xmmB
+.column_ld16:
+    test        cl, SIZEOF_XMMWORD
+    jz          short .column_ld32
+    movdqa      xmmF, xmmA
+    movdqu      xmmA, XMMWORD [rsi+0*SIZEOF_XMMWORD]
+    mov         rcx, SIZEOF_XMMWORD
+    jmp         short .rgb_gray_cnv
+.column_ld32:
+    test        cl, 2*SIZEOF_XMMWORD
+    mov         rcx, SIZEOF_XMMWORD
+    jz          short .rgb_gray_cnv
+    movdqa      xmmB, xmmA
+    movdqu      xmmA, XMMWORD [rsi+0*SIZEOF_XMMWORD]
+    movdqu      xmmF, XMMWORD [rsi+1*SIZEOF_XMMWORD]
+    jmp         short .rgb_gray_cnv
+
+.columnloop:
+    movdqu      xmmA, XMMWORD [rsi+0*SIZEOF_XMMWORD]
+    movdqu      xmmF, XMMWORD [rsi+1*SIZEOF_XMMWORD]
+    movdqu      xmmB, XMMWORD [rsi+2*SIZEOF_XMMWORD]
+
+.rgb_gray_cnv:
+    ; xmmA=(00 10 20 01 11 21 02 12 22 03 13 23 04 14 24 05)
+    ; xmmF=(15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A)
+    ; xmmB=(2A 0B 1B 2B 0C 1C 2C 0D 1D 2D 0E 1E 2E 0F 1F 2F)
+
+    movdqa      xmmG, xmmA
+    pslldq      xmmA, 8     ; xmmA=(-- -- -- -- -- -- -- -- 00 10 20 01 11 21 02 12)
+    psrldq      xmmG, 8     ; xmmG=(22 03 13 23 04 14 24 05 -- -- -- -- -- -- -- --)
+
+    punpckhbw   xmmA, xmmF  ; xmmA=(00 08 10 18 20 28 01 09 11 19 21 29 02 0A 12 1A)
+    pslldq      xmmF, 8     ; xmmF=(-- -- -- -- -- -- -- -- 15 25 06 16 26 07 17 27)
+
+    punpcklbw   xmmG, xmmB  ; xmmG=(22 2A 03 0B 13 1B 23 2B 04 0C 14 1C 24 2C 05 0D)
+    punpckhbw   xmmF, xmmB  ; xmmF=(15 1D 25 2D 06 0E 16 1E 26 2E 07 0F 17 1F 27 2F)
+
+    movdqa      xmmD, xmmA
+    pslldq      xmmA, 8     ; xmmA=(-- -- -- -- -- -- -- -- 00 08 10 18 20 28 01 09)
+    psrldq      xmmD, 8     ; xmmD=(11 19 21 29 02 0A 12 1A -- -- -- -- -- -- -- --)
+
+    punpckhbw   xmmA, xmmG  ; xmmA=(00 04 08 0C 10 14 18 1C 20 24 28 2C 01 05 09 0D)
+    pslldq      xmmG, 8     ; xmmG=(-- -- -- -- -- -- -- -- 22 2A 03 0B 13 1B 23 2B)
+
+    punpcklbw   xmmD, xmmF  ; xmmD=(11 15 19 1D 21 25 29 2D 02 06 0A 0E 12 16 1A 1E)
+    punpckhbw   xmmG, xmmF  ; xmmG=(22 26 2A 2E 03 07 0B 0F 13 17 1B 1F 23 27 2B 2F)
+
+    movdqa      xmmE, xmmA
+    pslldq      xmmA, 8     ; xmmA=(-- -- -- -- -- -- -- -- 00 04 08 0C 10 14 18 1C)
+    psrldq      xmmE, 8     ; xmmE=(20 24 28 2C 01 05 09 0D -- -- -- -- -- -- -- --)
+
+    punpckhbw   xmmA, xmmD  ; xmmA=(00 02 04 06 08 0A 0C 0E 10 12 14 16 18 1A 1C 1E)
+    pslldq      xmmD, 8     ; xmmD=(-- -- -- -- -- -- -- -- 11 15 19 1D 21 25 29 2D)
+
+    punpcklbw   xmmE, xmmG  ; xmmE=(20 22 24 26 28 2A 2C 2E 01 03 05 07 09 0B 0D 0F)
+    punpckhbw   xmmD, xmmG  ; xmmD=(11 13 15 17 19 1B 1D 1F 21 23 25 27 29 2B 2D 2F)
+
+    pxor        xmmH, xmmH
+
+    movdqa      xmmC, xmmA
+    punpcklbw   xmmA, xmmH  ; xmmA=(00 02 04 06 08 0A 0C 0E)
+    punpckhbw   xmmC, xmmH  ; xmmC=(10 12 14 16 18 1A 1C 1E)
+
+    movdqa      xmmB, xmmE
+    punpcklbw   xmmE, xmmH  ; xmmE=(20 22 24 26 28 2A 2C 2E)
+    punpckhbw   xmmB, xmmH  ; xmmB=(01 03 05 07 09 0B 0D 0F)
+
+    movdqa      xmmF, xmmD
+    punpcklbw   xmmD, xmmH  ; xmmD=(11 13 15 17 19 1B 1D 1F)
+    punpckhbw   xmmF, xmmH  ; xmmF=(21 23 25 27 29 2B 2D 2F)
+
+%else  ; RGB_PIXELSIZE == 4 ; -----------
+
+.column_ld1:
+    test        cl, SIZEOF_XMMWORD/16
+    jz          short .column_ld2
+    sub         rcx, byte SIZEOF_XMMWORD/16
+    movd        xmmA, XMM_DWORD [rsi+rcx*RGB_PIXELSIZE]
+.column_ld2:
+    test        cl, SIZEOF_XMMWORD/8
+    jz          short .column_ld4
+    sub         rcx, byte SIZEOF_XMMWORD/8
+    movq        xmmE, XMM_MMWORD [rsi+rcx*RGB_PIXELSIZE]
+    pslldq      xmmA, SIZEOF_MMWORD
+    por         xmmA, xmmE
+.column_ld4:
+    test        cl, SIZEOF_XMMWORD/4
+    jz          short .column_ld8
+    sub         rcx, byte SIZEOF_XMMWORD/4
+    movdqa      xmmE, xmmA
+    movdqu      xmmA, XMMWORD [rsi+rcx*RGB_PIXELSIZE]
+.column_ld8:
+    test        cl, SIZEOF_XMMWORD/2
+    mov         rcx, SIZEOF_XMMWORD
+    jz          short .rgb_gray_cnv
+    movdqa      xmmF, xmmA
+    movdqa      xmmH, xmmE
+    movdqu      xmmA, XMMWORD [rsi+0*SIZEOF_XMMWORD]
+    movdqu      xmmE, XMMWORD [rsi+1*SIZEOF_XMMWORD]
+    jmp         short .rgb_gray_cnv
+
+.columnloop:
+    movdqu      xmmA, XMMWORD [rsi+0*SIZEOF_XMMWORD]
+    movdqu      xmmE, XMMWORD [rsi+1*SIZEOF_XMMWORD]
+    movdqu      xmmF, XMMWORD [rsi+2*SIZEOF_XMMWORD]
+    movdqu      xmmH, XMMWORD [rsi+3*SIZEOF_XMMWORD]
+
+.rgb_gray_cnv:
+    ; xmmA=(00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33)
+    ; xmmE=(04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37)
+    ; xmmF=(08 18 28 38 09 19 29 39 0A 1A 2A 3A 0B 1B 2B 3B)
+    ; xmmH=(0C 1C 2C 3C 0D 1D 2D 3D 0E 1E 2E 3E 0F 1F 2F 3F)
+
+    movdqa      xmmD, xmmA
+    punpcklbw   xmmA, xmmE      ; xmmA=(00 04 10 14 20 24 30 34 01 05 11 15 21 25 31 35)
+    punpckhbw   xmmD, xmmE      ; xmmD=(02 06 12 16 22 26 32 36 03 07 13 17 23 27 33 37)
+
+    movdqa      xmmC, xmmF
+    punpcklbw   xmmF, xmmH      ; xmmF=(08 0C 18 1C 28 2C 38 3C 09 0D 19 1D 29 2D 39 3D)
+    punpckhbw   xmmC, xmmH      ; xmmC=(0A 0E 1A 1E 2A 2E 3A 3E 0B 0F 1B 1F 2B 2F 3B 3F)
+
+    movdqa      xmmB, xmmA
+    punpcklwd   xmmA, xmmF      ; xmmA=(00 04 08 0C 10 14 18 1C 20 24 28 2C 30 34 38 3C)
+    punpckhwd   xmmB, xmmF      ; xmmB=(01 05 09 0D 11 15 19 1D 21 25 29 2D 31 35 39 3D)
+
+    movdqa      xmmG, xmmD
+    punpcklwd   xmmD, xmmC      ; xmmD=(02 06 0A 0E 12 16 1A 1E 22 26 2A 2E 32 36 3A 3E)
+    punpckhwd   xmmG, xmmC      ; xmmG=(03 07 0B 0F 13 17 1B 1F 23 27 2B 2F 33 37 3B 3F)
+
+    movdqa      xmmE, xmmA
+    punpcklbw   xmmA, xmmD      ; xmmA=(00 02 04 06 08 0A 0C 0E 10 12 14 16 18 1A 1C 1E)
+    punpckhbw   xmmE, xmmD      ; xmmE=(20 22 24 26 28 2A 2C 2E 30 32 34 36 38 3A 3C 3E)
+
+    movdqa      xmmH, xmmB
+    punpcklbw   xmmB, xmmG      ; xmmB=(01 03 05 07 09 0B 0D 0F 11 13 15 17 19 1B 1D 1F)
+    punpckhbw   xmmH, xmmG      ; xmmH=(21 23 25 27 29 2B 2D 2F 31 33 35 37 39 3B 3D 3F)
+
+    pxor        xmmF, xmmF
+
+    movdqa      xmmC, xmmA
+    punpcklbw   xmmA, xmmF      ; xmmA=(00 02 04 06 08 0A 0C 0E)
+    punpckhbw   xmmC, xmmF      ; xmmC=(10 12 14 16 18 1A 1C 1E)
+
+    movdqa      xmmD, xmmB
+    punpcklbw   xmmB, xmmF      ; xmmB=(01 03 05 07 09 0B 0D 0F)
+    punpckhbw   xmmD, xmmF      ; xmmD=(11 13 15 17 19 1B 1D 1F)
+
+    movdqa      xmmG, xmmE
+    punpcklbw   xmmE, xmmF      ; xmmE=(20 22 24 26 28 2A 2C 2E)
+    punpckhbw   xmmG, xmmF      ; xmmG=(30 32 34 36 38 3A 3C 3E)
+
+    punpcklbw   xmmF, xmmH
+    punpckhbw   xmmH, xmmH
+    psrlw       xmmF, BYTE_BIT  ; xmmF=(21 23 25 27 29 2B 2D 2F)
+    psrlw       xmmH, BYTE_BIT  ; xmmH=(31 33 35 37 39 3B 3D 3F)
+
+%endif  ; RGB_PIXELSIZE ; ---------------
+
+    ; xmm0=R(02468ACE)=RE, xmm2=G(02468ACE)=GE, xmm4=B(02468ACE)=BE
+    ; xmm1=R(13579BDF)=RO, xmm3=G(13579BDF)=GO, xmm5=B(13579BDF)=BO
+
+    ; (Original)
+    ; Y  =  0.29900 * R + 0.58700 * G + 0.11400 * B
+    ;
+    ; (This implementation)
+    ; Y  =  0.29900 * R + 0.33700 * G + 0.11400 * B + 0.25000 * G
+
+    movdqa      xmm6, xmm1
+    punpcklwd   xmm1, xmm3
+    punpckhwd   xmm6, xmm3
+    pmaddwd     xmm1, [rel PW_F0299_F0337]  ; xmm1=ROL*FIX(0.299)+GOL*FIX(0.337)
+    pmaddwd     xmm6, [rel PW_F0299_F0337]  ; xmm6=ROH*FIX(0.299)+GOH*FIX(0.337)
+
+    movdqa      xmm7, xmm6              ; xmm7=ROH*FIX(0.299)+GOH*FIX(0.337)
+
+    movdqa      xmm6, xmm0
+    punpcklwd   xmm0, xmm2
+    punpckhwd   xmm6, xmm2
+    pmaddwd     xmm0, [rel PW_F0299_F0337]  ; xmm0=REL*FIX(0.299)+GEL*FIX(0.337)
+    pmaddwd     xmm6, [rel PW_F0299_F0337]  ; xmm6=REH*FIX(0.299)+GEH*FIX(0.337)
+
+    movdqa      XMMWORD [wk(0)], xmm0   ; wk(0)=REL*FIX(0.299)+GEL*FIX(0.337)
+    movdqa      XMMWORD [wk(1)], xmm6   ; wk(1)=REH*FIX(0.299)+GEH*FIX(0.337)
+
+    movdqa      xmm0, xmm5              ; xmm0=BO
+    movdqa      xmm6, xmm4              ; xmm6=BE
+
+    movdqa      xmm4, xmm0
+    punpcklwd   xmm0, xmm3
+    punpckhwd   xmm4, xmm3
+    pmaddwd     xmm0, [rel PW_F0114_F0250]  ; xmm0=BOL*FIX(0.114)+GOL*FIX(0.250)
+    pmaddwd     xmm4, [rel PW_F0114_F0250]  ; xmm4=BOH*FIX(0.114)+GOH*FIX(0.250)
+
+    movdqa      xmm3, [rel PD_ONEHALF]      ; xmm3=[PD_ONEHALF]
+
+    paddd       xmm0, xmm1
+    paddd       xmm4, xmm7
+    paddd       xmm0, xmm3
+    paddd       xmm4, xmm3
+    psrld       xmm0, SCALEBITS         ; xmm0=YOL
+    psrld       xmm4, SCALEBITS         ; xmm4=YOH
+    packssdw    xmm0, xmm4              ; xmm0=YO
+
+    movdqa      xmm4, xmm6
+    punpcklwd   xmm6, xmm2
+    punpckhwd   xmm4, xmm2
+    pmaddwd     xmm6, [rel PW_F0114_F0250]  ; xmm6=BEL*FIX(0.114)+GEL*FIX(0.250)
+    pmaddwd     xmm4, [rel PW_F0114_F0250]  ; xmm4=BEH*FIX(0.114)+GEH*FIX(0.250)
+
+    movdqa      xmm2, [rel PD_ONEHALF]      ; xmm2=[PD_ONEHALF]
+
+    paddd       xmm6, XMMWORD [wk(0)]
+    paddd       xmm4, XMMWORD [wk(1)]
+    paddd       xmm6, xmm2
+    paddd       xmm4, xmm2
+    psrld       xmm6, SCALEBITS         ; xmm6=YEL
+    psrld       xmm4, SCALEBITS         ; xmm4=YEH
+    packssdw    xmm6, xmm4              ; xmm6=YE
+
+    psllw       xmm0, BYTE_BIT
+    por         xmm6, xmm0              ; xmm6=Y
+    movdqa      XMMWORD [rdi], xmm6     ; Save Y
+
+    sub         rcx, byte SIZEOF_XMMWORD
+    add         rsi, byte RGB_PIXELSIZE*SIZEOF_XMMWORD  ; inptr
+    add         rdi, byte SIZEOF_XMMWORD                ; outptr0
+    cmp         rcx, byte SIZEOF_XMMWORD
+    jae         near .columnloop
+    test        rcx, rcx
+    jnz         near .column_ld1
+
+    pop         rcx                     ; col
+    pop         rsi
+    pop         rdi
+
+    add         rsi, byte SIZEOF_JSAMPROW  ; input_buf
+    add         rdi, byte SIZEOF_JSAMPROW
+    dec         rax                        ; num_rows
+    jg          near .rowloop
+
+.return:
+    pop         rbx
+    uncollect_args 5
+    mov         rsp, rbp                ; rsp <- aligned rbp
+    pop         rsp                     ; rsp <- original rbp
+    pop         rbp
+    ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+    align       32
diff --git a/media/libjpeg/simd/x86_64/jchuff-sse2.asm b/media/libjpeg/simd/x86_64/jchuff-sse2.asm
new file mode 100644
index 0000000000..9ea6df946e
--- /dev/null
+++ b/media/libjpeg/simd/x86_64/jchuff-sse2.asm
@@ -0,0 +1,583 @@
+;
+; jchuff-sse2.asm - Huffman entropy encoding (64-bit SSE2)
+;
+; Copyright (C) 2009-2011, 2014-2016, 2019, 2021, D. R. Commander.
+; Copyright (C) 2015, Matthieu Darbois.
+; Copyright (C) 2018, Matthias Räncker.
+;
+; Based on the x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; This file contains an SSE2 implementation for Huffman coding of one block.
+; The following code is based on jchuff.c; see jchuff.c for more details.
+
+%include "jsimdext.inc"
+
+struc working_state
+.next_output_byte:   resp 1     ; => next byte to write in buffer
+.free_in_buffer:     resp 1     ; # of byte spaces remaining in buffer
+.cur.put_buffer.simd resq 1     ; current bit accumulation buffer
+.cur.free_bits       resd 1     ; # of bits available in it
+.cur.last_dc_val     resd 4     ; last DC coef for each component
+.cinfo:              resp 1     ; dump_buffer needs access to this
+endstruc
+
+struc c_derived_tbl
+.ehufco:             resd 256   ; code for each symbol
+.ehufsi:             resb 256   ; length of code for each symbol
+; If no code has been allocated for a symbol S, ehufsi[S] contains 0
+endstruc
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_CONST
+
+    alignz      32
+    GLOBAL_DATA(jconst_huff_encode_one_block)
+
+EXTN(jconst_huff_encode_one_block):
+
+jpeg_mask_bits dd 0x0000, 0x0001, 0x0003, 0x0007
+               dd 0x000f, 0x001f, 0x003f, 0x007f
+               dd 0x00ff, 0x01ff, 0x03ff, 0x07ff
+               dd 0x0fff, 0x1fff, 0x3fff, 0x7fff
+
+    alignz      32
+
+times 1 << 14 db 15
+times 1 << 13 db 14
+times 1 << 12 db 13
+times 1 << 11 db 12
+times 1 << 10 db 11
+times 1 <<  9 db 10
+times 1 <<  8 db  9
+times 1 <<  7 db  8
+times 1 <<  6 db  7
+times 1 <<  5 db  6
+times 1 <<  4 db  5
+times 1 <<  3 db  4
+times 1 <<  2 db  3
+times 1 <<  1 db  2
+times 1 <<  0 db  1
+times 1       db  0
+jpeg_nbits_table:
+times 1       db  0
+times 1 <<  0 db  1
+times 1 <<  1 db  2
+times 1 <<  2 db  3
+times 1 <<  3 db  4
+times 1 <<  4 db  5
+times 1 <<  5 db  6
+times 1 <<  6 db  7
+times 1 <<  7 db  8
+times 1 <<  8 db  9
+times 1 <<  9 db 10
+times 1 << 10 db 11
+times 1 << 11 db 12
+times 1 << 12 db 13
+times 1 << 13 db 14
+times 1 << 14 db 15
+times 1 << 15 db 16
+
+    alignz      32
+
+%define NBITS(x)      nbits_base + x
+%define MASK_BITS(x)  NBITS((x) * 4) + (jpeg_mask_bits - jpeg_nbits_table)
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_TEXT
+    BITS        64
+
+; Shorthand used to describe SIMD operations:
+; wN:  xmmN treated as eight signed 16-bit values
+; wN[i]:  perform the same operation on all eight signed 16-bit values, i=0..7
+; bN:  xmmN treated as 16 unsigned 8-bit values
+; bN[i]:  perform the same operation on all 16 unsigned 8-bit values, i=0..15
+; Contents of SIMD registers are shown in memory order.
+
+; Fill the bit buffer to capacity with the leading bits from code, then output
+; the bit buffer and put the remaining bits from code into the bit buffer.
+;
+; Usage:
+; code - contains the bits to shift into the bit buffer (LSB-aligned)
+; %1 - the label to which to jump when the macro completes
+; %2 (optional) - extra instructions to execute after nbits has been set
+;
+; Upon completion, free_bits will be set to the number of remaining bits from
+; code, and put_buffer will contain those remaining bits.  temp and code will
+; be clobbered.
+;
+; This macro encodes any 0xFF bytes as 0xFF 0x00, as does the EMIT_BYTE()
+; macro in jchuff.c.
+
+%macro EMIT_QWORD 1-2
+    add         nbitsb, free_bitsb      ; nbits += free_bits;
+    neg         free_bitsb              ; free_bits = -free_bits;
+    mov         tempd, code             ; temp = code;
+    shl         put_buffer, nbitsb      ; put_buffer <<= nbits;
+    mov         nbitsb, free_bitsb      ; nbits = free_bits;
+    neg         free_bitsb              ; free_bits = -free_bits;
+    shr         tempd, nbitsb           ; temp >>= nbits;
+    or          tempq, put_buffer       ; temp |= put_buffer;
+    movq        xmm0, tempq             ; xmm0.u64 = { temp, 0 };
+    bswap       tempq                   ; temp = htonl(temp);
+    mov         put_buffer, codeq       ; put_buffer = code;
+    pcmpeqb     xmm0, xmm1              ; b0[i] = (b0[i] == 0xFF ? 0xFF : 0);
+    %2
+    pmovmskb    code, xmm0              ; code = 0;  code |= ((b0[i] >> 7) << i);
+    mov         qword [buffer], tempq   ; memcpy(buffer, &temp, 8);
+                                        ; (speculative; will be overwritten if
+                                        ; code contains any 0xFF bytes)
+    add         free_bitsb, 64          ; free_bits += 64;
+    add         bufferp, 8              ; buffer += 8;
+    test        code, code              ; if (code == 0)  /* No 0xFF bytes */
+    jz          %1                      ;   return;
+    ; Execute the equivalent of the EMIT_BYTE() macro in jchuff.c for all 8
+    ; bytes in the qword.
+    cmp         tempb, 0xFF             ; Set CF if temp[0] < 0xFF
+    mov         byte [buffer-7], 0      ; buffer[-7] = 0;
+    sbb         bufferp, 6              ; buffer -= (6 + (temp[0] < 0xFF ? 1 : 0));
+    mov         byte [buffer], temph    ; buffer[0] = temp[1];
+    cmp         temph, 0xFF             ; Set CF if temp[1] < 0xFF
+    mov         byte [buffer+1], 0      ; buffer[1] = 0;
+    sbb         bufferp, -2             ; buffer -= (-2 + (temp[1] < 0xFF ? 1 : 0));
+    shr         tempq, 16               ; temp >>= 16;
+    mov         byte [buffer], tempb    ; buffer[0] = temp[0];
+    cmp         tempb, 0xFF             ; Set CF if temp[0] < 0xFF
+    mov         byte [buffer+1], 0      ; buffer[1] = 0;
+    sbb         bufferp, -2             ; buffer -= (-2 + (temp[0] < 0xFF ? 1 : 0));
+    mov         byte [buffer], temph    ; buffer[0] = temp[1];
+    cmp         temph, 0xFF             ; Set CF if temp[1] < 0xFF
+    mov         byte [buffer+1], 0      ; buffer[1] = 0;
+    sbb         bufferp, -2             ; buffer -= (-2 + (temp[1] < 0xFF ? 1 : 0));
+    shr         tempq, 16               ; temp >>= 16;
+    mov         byte [buffer], tempb    ; buffer[0] = temp[0];
+    cmp         tempb, 0xFF             ; Set CF if temp[0] < 0xFF
+    mov         byte [buffer+1], 0      ; buffer[1] = 0;
+    sbb         bufferp, -2             ; buffer -= (-2 + (temp[0] < 0xFF ? 1 : 0));
+    mov         byte [buffer], temph    ; buffer[0] = temp[1];
+    cmp         temph, 0xFF             ; Set CF if temp[1] < 0xFF
+    mov         byte [buffer+1], 0      ; buffer[1] = 0;
+    sbb         bufferp, -2             ; buffer -= (-2 + (temp[1] < 0xFF ? 1 : 0));
+    shr         tempd, 16               ; temp >>= 16;
+    mov         byte [buffer], tempb    ; buffer[0] = temp[0];
+    cmp         tempb, 0xFF             ; Set CF if temp[0] < 0xFF
+    mov         byte [buffer+1], 0      ; buffer[1] = 0;
+    sbb         bufferp, -2             ; buffer -= (-2 + (temp[0] < 0xFF ? 1 : 0));
+    mov         byte [buffer], temph    ; buffer[0] = temp[1];
+    cmp         temph, 0xFF             ; Set CF if temp[1] < 0xFF
+    mov         byte [buffer+1], 0      ; buffer[1] = 0;
+    sbb         bufferp, -2             ; buffer -= (-2 + (temp[1] < 0xFF ? 1 : 0));
+    jmp         %1                      ; return;
+%endmacro
+
+;
+; Encode a single block's worth of coefficients.
+;
+; GLOBAL(JOCTET *)
+; jsimd_huff_encode_one_block_sse2(working_state *state, JOCTET *buffer,
+;                                  JCOEFPTR block, int last_dc_val,
+;                                  c_derived_tbl *dctbl, c_derived_tbl *actbl)
+;
+; NOTES:
+; When shuffling data, we try to avoid pinsrw as much as possible, since it is
+; slow on many CPUs.  Its reciprocal throughput (issue latency) is 1 even on
+; modern CPUs, so chains of pinsrw instructions (even with different outputs)
+; can limit performance.  pinsrw is a VectorPath instruction on AMD K8 and
+; requires 2 µops (with memory operand) on Intel.  In either case, only one
+; pinsrw instruction can be decoded per cycle (and nothing else if they are
+; back-to-back), so out-of-order execution cannot be used to work around long
+; pinsrw chains (though for Sandy Bridge and later, this may be less of a
+; problem if the code runs from the µop cache.)
+;
+; We use tzcnt instead of bsf without checking for support.  The instruction is
+; executed as bsf on CPUs that don't support tzcnt (encoding is equivalent to
+; rep bsf.)  The destination (first) operand of bsf (and tzcnt on some CPUs) is
+; an input dependency (although the behavior is not formally defined, Intel
+; CPUs usually leave the destination unmodified if the source is zero.)  This
+; can prevent out-of-order execution, so we clear the destination before
+; invoking tzcnt.
+;
+; Initial register allocation
+; rax - buffer
+; rbx - temp
+; rcx - nbits
+; rdx - block --> free_bits
+; rsi - nbits_base
+; rdi - t
+; rbp - code
+; r8  - dctbl --> code_temp
+; r9  - actbl
+; r10 - state
+; r11 - index
+; r12 - put_buffer
+
+%define buffer       rax
+%ifdef WIN64
+%define bufferp      rax
+%else
+%define bufferp      raxp
+%endif
+%define tempq        rbx
+%define tempd        ebx
+%define tempb        bl
+%define temph        bh
+%define nbitsq       rcx
+%define nbits        ecx
+%define nbitsb       cl
+%define block        rdx
+%define nbits_base   rsi
+%define t            rdi
+%define td           edi
+%define codeq        rbp
+%define code         ebp
+%define dctbl        r8
+%define actbl        r9
+%define state        r10
+%define index        r11
+%define indexd       r11d
+%define put_buffer   r12
+%define put_bufferd  r12d
+
+; Step 1: Re-arrange input data according to jpeg_natural_order
+; xx 01 02 03 04 05 06 07      xx 01 08 16 09 02 03 10
+; 08 09 10 11 12 13 14 15      17 24 32 25 18 11 04 05
+; 16 17 18 19 20 21 22 23      12 19 26 33 40 48 41 34
+; 24 25 26 27 28 29 30 31 ==>  27 20 13 06 07 14 21 28
+; 32 33 34 35 36 37 38 39      35 42 49 56 57 50 43 36
+; 40 41 42 43 44 45 46 47      29 22 15 23 30 37 44 51
+; 48 49 50 51 52 53 54 55      58 59 52 45 38 31 39 46
+; 56 57 58 59 60 61 62 63      53 60 61 54 47 55 62 63
+
+    align       32
+    GLOBAL_FUNCTION(jsimd_huff_encode_one_block_sse2)
+
+EXTN(jsimd_huff_encode_one_block_sse2):
+
+%ifdef WIN64
+
+; rcx = working_state *state
+; rdx = JOCTET *buffer
+; r8 = JCOEFPTR block
+; r9 = int last_dc_val
+; [rax+48] = c_derived_tbl *dctbl
+; [rax+56] = c_derived_tbl *actbl
+
+                                                          ;X: X = code stream
+    mov         buffer, rdx
+    mov         block, r8
+    movups      xmm3, XMMWORD [block + 0 * SIZEOF_WORD]   ;D: w3 = xx 01 02 03 04 05 06 07
+    push        rbx
+    push        rbp
+    movdqa      xmm0, xmm3                                ;A: w0 = xx 01 02 03 04 05 06 07
+    push        rsi
+    push        rdi
+    push        r12
+    movups      xmm1, XMMWORD [block + 8 * SIZEOF_WORD]   ;B: w1 = 08 09 10 11 12 13 14 15
+    mov         state, rcx
+    movsx       code, word [block]                        ;Z:     code = block[0];
+    pxor        xmm4, xmm4                                ;A: w4[i] = 0;
+    sub         code, r9d                                 ;Z:     code -= last_dc_val;
+    mov         dctbl, POINTER [rsp+6*8+4*8]
+    mov         actbl, POINTER [rsp+6*8+5*8]
+    punpckldq   xmm0, xmm1                                ;A: w0 = xx 01 08 09 02 03 10 11
+    lea         nbits_base, [rel jpeg_nbits_table]
+    add         rsp, -DCTSIZE2 * SIZEOF_WORD
+    mov         t, rsp
+
+%else
+
+; rdi = working_state *state
+; rsi = JOCTET *buffer
+; rdx = JCOEFPTR block
+; rcx = int last_dc_val
+; r8 = c_derived_tbl *dctbl
+; r9 = c_derived_tbl *actbl
+
+                                                          ;X: X = code stream
+    movups      xmm3, XMMWORD [block + 0 * SIZEOF_WORD]   ;D: w3 = xx 01 02 03 04 05 06 07
+    push        rbx
+    push        rbp
+    movdqa      xmm0, xmm3                                ;A: w0 = xx 01 02 03 04 05 06 07
+    push        r12
+    mov         state, rdi
+    mov         buffer, rsi
+    movups      xmm1, XMMWORD [block + 8 * SIZEOF_WORD]   ;B: w1 = 08 09 10 11 12 13 14 15
+    movsx       codeq, word [block]                       ;Z:     code = block[0];
+    lea         nbits_base, [rel jpeg_nbits_table]
+    pxor        xmm4, xmm4                                ;A: w4[i] = 0;
+    sub         codeq, rcx                                ;Z:     code -= last_dc_val;
+    punpckldq   xmm0, xmm1                                ;A: w0 = xx 01 08 09 02 03 10 11
+    lea         t, [rsp - DCTSIZE2 * SIZEOF_WORD]         ;   use red zone for t_
+
+%endif
+
+    pshuflw     xmm0, xmm0, 11001001b                     ;A: w0 = 01 08 xx 09 02 03 10 11
+    pinsrw      xmm0, word [block + 16 * SIZEOF_WORD], 2  ;A: w0 = 01 08 16 09 02 03 10 11
+    punpckhdq   xmm3, xmm1                                ;D: w3 = 04 05 12 13 06 07 14 15
+    punpcklqdq  xmm1, xmm3                                ;B: w1 = 08 09 10 11 04 05 12 13
+    pinsrw      xmm0, word [block + 17 * SIZEOF_WORD], 7  ;A: w0 = 01 08 16 09 02 03 10 17
+                                                          ;A:      (Row 0, offset 1)
+    pcmpgtw     xmm4, xmm0                                ;A: w4[i] = (w0[i] < 0 ? -1 : 0);
+    paddw       xmm0, xmm4                                ;A: w0[i] += w4[i];
+    movaps      XMMWORD [t + 0 * SIZEOF_WORD], xmm0       ;A: t[i] = w0[i];
+
+    movq        xmm2, qword [block + 24 * SIZEOF_WORD]    ;B: w2 = 24 25 26 27 -- -- -- --
+    pshuflw     xmm2, xmm2, 11011000b                     ;B: w2 = 24 26 25 27 -- -- -- --
+    pslldq      xmm1, 1 * SIZEOF_WORD                     ;B: w1 = -- 08 09 10 11 04 05 12
+    movups      xmm5, XMMWORD [block + 48 * SIZEOF_WORD]  ;H: w5 = 48 49 50 51 52 53 54 55
+    movsd       xmm1, xmm2                                ;B: w1 = 24 26 25 27 11 04 05 12
+    punpcklqdq  xmm2, xmm5                                ;C: w2 = 24 26 25 27 48 49 50 51
+    pinsrw      xmm1, word [block + 32 * SIZEOF_WORD], 1  ;B: w1 = 24 32 25 27 11 04 05 12
+    pxor        xmm4, xmm4                                ;A: w4[i] = 0;
+    psrldq      xmm3, 2 * SIZEOF_WORD                     ;D: w3 = 12 13 06 07 14 15 -- --
+    pcmpeqw     xmm0, xmm4                                ;A: w0[i] = (w0[i] == 0 ? -1 : 0);
+    pinsrw      xmm1, word [block + 18 * SIZEOF_WORD], 3  ;B: w1 = 24 32 25 18 11 04 05 12
+                                                          ;        (Row 1, offset 1)
+    pcmpgtw     xmm4, xmm1                                ;B: w4[i] = (w1[i] < 0 ? -1 : 0);
+    paddw       xmm1, xmm4                                ;B: w1[i] += w4[i];
+    movaps      XMMWORD [t + 8 * SIZEOF_WORD], xmm1       ;B: t[i+8] = w1[i];
+    pxor        xmm4, xmm4                                ;B: w4[i] = 0;
+    pcmpeqw     xmm1, xmm4                                ;B: w1[i] = (w1[i] == 0 ? -1 : 0);
+
+    packsswb    xmm0, xmm1                                ;AB: b0[i] = w0[i], b0[i+8] = w1[i]
+                                                          ;    w/ signed saturation
+
+    pinsrw      xmm3, word [block + 20 * SIZEOF_WORD], 0  ;D: w3 = 20 13 06 07 14 15 -- --
+    pinsrw      xmm3, word [block + 21 * SIZEOF_WORD], 5  ;D: w3 = 20 13 06 07 14 21 -- --
+    pinsrw      xmm3, word [block + 28 * SIZEOF_WORD], 6  ;D: w3 = 20 13 06 07 14 21 28 --
+    pinsrw      xmm3, word [block + 35 * SIZEOF_WORD], 7  ;D: w3 = 20 13 06 07 14 21 28 35
+                                                          ;        (Row 3, offset 1)
+    pcmpgtw     xmm4, xmm3                                ;D: w4[i] = (w3[i] < 0 ? -1 : 0);
+    paddw       xmm3, xmm4                                ;D: w3[i] += w4[i];
+    movaps      XMMWORD [t + 24 * SIZEOF_WORD], xmm3      ;D: t[i+24] = w3[i];
+    pxor        xmm4, xmm4                                ;D: w4[i] = 0;
+    pcmpeqw     xmm3, xmm4                                ;D: w3[i] = (w3[i] == 0 ? -1 : 0);
+
+    pinsrw      xmm2, word [block + 19 * SIZEOF_WORD], 0  ;C: w2 = 19 26 25 27 48 49 50 51
+    cmp         code, 1 << 31                             ;Z:     Set CF if code < 0x80000000,
+                                                          ;Z:     i.e. if code is positive
+    pinsrw      xmm2, word [block + 33 * SIZEOF_WORD], 2  ;C: w2 = 19 26 33 27 48 49 50 51
+    pinsrw      xmm2, word [block + 40 * SIZEOF_WORD], 3  ;C: w2 = 19 26 33 40 48 49 50 51
+    adc         code, -1                                  ;Z:     code += -1 + (code >= 0 ? 1 : 0);
+    pinsrw      xmm2, word [block + 41 * SIZEOF_WORD], 5  ;C: w2 = 19 26 33 40 48 41 50 51
+    pinsrw      xmm2, word [block + 34 * SIZEOF_WORD], 6  ;C: w2 = 19 26 33 40 48 41 34 51
+    movsxd      codeq, code                               ;Z:     sign extend code
+    pinsrw      xmm2, word [block + 27 * SIZEOF_WORD], 7  ;C: w2 = 19 26 33 40 48 41 34 27
+                                                          ;        (Row 2, offset 1)
+    pcmpgtw     xmm4, xmm2                                ;C: w4[i] = (w2[i] < 0 ? -1 : 0);
+    paddw       xmm2, xmm4                                ;C: w2[i] += w4[i];
+    movaps      XMMWORD [t + 16 * SIZEOF_WORD], xmm2      ;C: t[i+16] = w2[i];
+    pxor        xmm4, xmm4                                ;C: w4[i] = 0;
+    pcmpeqw     xmm2, xmm4                                ;C: w2[i] = (w2[i] == 0 ? -1 : 0);
+
+    packsswb    xmm2, xmm3                                ;CD: b2[i] = w2[i], b2[i+8] = w3[i]
+                                                          ;    w/ signed saturation
+
+    movzx       nbitsq, byte [NBITS(codeq)]               ;Z:     nbits = JPEG_NBITS(code);
+    movdqa      xmm3, xmm5                                ;H: w3 = 48 49 50 51 52 53 54 55
+    pmovmskb    tempd, xmm2                               ;Z:     temp = 0;  temp |= ((b2[i] >> 7) << i);
+    pmovmskb    put_bufferd, xmm0                         ;Z:     put_buffer = 0;  put_buffer |= ((b0[i] >> 7) << i);
+    movups      xmm0, XMMWORD [block + 56 * SIZEOF_WORD]  ;H: w0 = 56 57 58 59 60 61 62 63
+    punpckhdq   xmm3, xmm0                                ;H: w3 = 52 53 60 61 54 55 62 63
+    shl         tempd, 16                                 ;Z:     temp <<= 16;
+    psrldq      xmm3, 1 * SIZEOF_WORD                     ;H: w3 = 53 60 61 54 55 62 63 --
+    pxor        xmm2, xmm2                                ;H: w2[i] = 0;
+    or          put_bufferd, tempd                        ;Z:     put_buffer |= temp;
+    pshuflw     xmm3, xmm3, 00111001b                     ;H: w3 = 60 61 54 53 55 62 63 --
+    movq        xmm1, qword [block + 44 * SIZEOF_WORD]    ;G: w1 = 44 45 46 47 -- -- -- --
+    unpcklps    xmm5, xmm0                                ;E: w5 = 48 49 56 57 50 51 58 59
+    pxor        xmm0, xmm0                                ;H: w0[i] = 0;
+    pinsrw      xmm3, word [block + 47 * SIZEOF_WORD], 3  ;H: w3 = 60 61 54 47 55 62 63 --
+                                                          ;        (Row 7, offset 1)
+    pcmpgtw     xmm2, xmm3                                ;H: w2[i] = (w3[i] < 0 ? -1 : 0);
+    paddw       xmm3, xmm2                                ;H: w3[i] += w2[i];
+    movaps      XMMWORD [t + 56 * SIZEOF_WORD], xmm3      ;H: t[i+56] = w3[i];
+    movq        xmm4, qword [block + 36 * SIZEOF_WORD]    ;G: w4 = 36 37 38 39 -- -- -- --
+    pcmpeqw     xmm3, xmm0                                ;H: w3[i] = (w3[i] == 0 ? -1 : 0);
+    punpckldq   xmm4, xmm1                                ;G: w4 = 36 37 44 45 38 39 46 47
+    mov         tempd, [dctbl + c_derived_tbl.ehufco + nbitsq * 4]
+                                                          ;Z:     temp = dctbl->ehufco[nbits];
+    movdqa      xmm1, xmm4                                ;F: w1 = 36 37 44 45 38 39 46 47
+    psrldq      xmm4, 1 * SIZEOF_WORD                     ;G: w4 = 37 44 45 38 39 46 47 --
+    shufpd      xmm1, xmm5, 10b                           ;F: w1 = 36 37 44 45 50 51 58 59
+    and         code, dword [MASK_BITS(nbitsq)]           ;Z:     code &= (1 << nbits) - 1;
+    pshufhw     xmm4, xmm4, 11010011b                     ;G: w4 = 37 44 45 38 -- 39 46 --
+    pslldq      xmm1, 1 * SIZEOF_WORD                     ;F: w1 = -- 36 37 44 45 50 51 58
+    shl         tempq, nbitsb                             ;Z:     temp <<= nbits;
+    pinsrw      xmm4, word [block + 59 * SIZEOF_WORD], 0  ;G: w4 = 59 44 45 38 -- 39 46 --
+    pshufd      xmm1, xmm1, 11011000b                     ;F: w1 = -- 36 45 50 37 44 51 58
+    pinsrw      xmm4, word [block + 52 * SIZEOF_WORD], 1  ;G: w4 = 59 52 45 38 -- 39 46 --
+    or          code, tempd                               ;Z:     code |= temp;
+    movlps      xmm1, qword [block + 20 * SIZEOF_WORD]    ;F: w1 = 20 21 22 23 37 44 51 58
+    pinsrw      xmm4, word [block + 31 * SIZEOF_WORD], 4  ;G: w4 = 59 52 45 38 31 39 46 --
+    pshuflw     xmm1, xmm1, 01110010b                     ;F: w1 = 22 20 23 21 37 44 51 58
+    pinsrw      xmm4, word [block + 53 * SIZEOF_WORD], 7  ;G: w4 = 59 52 45 38 31 39 46 53
+                                                          ;        (Row 6, offset 1)
+    pxor        xmm2, xmm2                                ;G: w2[i] = 0;
+    pcmpgtw     xmm0, xmm4                                ;G: w0[i] = (w4[i] < 0 ? -1 : 0);
+    pinsrw      xmm1, word [block + 15 * SIZEOF_WORD], 1  ;F: w1 = 22 15 23 21 37 44 51 58
+    paddw       xmm4, xmm0                                ;G: w4[i] += w0[i];
+    movaps      XMMWORD [t + 48 * SIZEOF_WORD], xmm4      ;G: t[48+i] = w4[i];
+    pinsrw      xmm1, word [block + 30 * SIZEOF_WORD], 3  ;F: w1 = 22 15 23 30 37 44 51 58
+                                                          ;        (Row 5, offset 1)
+    pcmpeqw     xmm4, xmm2                                ;G: w4[i] = (w4[i] == 0 ? -1 : 0);
+    pinsrw      xmm5, word [block + 42 * SIZEOF_WORD], 0  ;E: w5 = 42 49 56 57 50 51 58 59
+
+    packsswb    xmm4, xmm3                                ;GH: b4[i] = w4[i], b4[i+8] = w3[i]
+                                                          ;    w/ signed saturation
+
+    pxor        xmm0, xmm0                                ;F: w0[i] = 0;
+    pinsrw      xmm5, word [block + 43 * SIZEOF_WORD], 5  ;E: w5 = 42 49 56 57 50 43 58 59
+    pcmpgtw     xmm2, xmm1                                ;F: w2[i] = (w1[i] < 0 ? -1 : 0);
+    pmovmskb    tempd, xmm4                               ;Z:     temp = 0;  temp |= ((b4[i] >> 7) << i);
+    pinsrw      xmm5, word [block + 36 * SIZEOF_WORD], 6  ;E: w5 = 42 49 56 57 50 43 36 59
+    paddw       xmm1, xmm2                                ;F: w1[i] += w2[i];
+    movaps      XMMWORD [t + 40 * SIZEOF_WORD], xmm1      ;F: t[40+i] = w1[i];
+    pinsrw      xmm5, word [block + 29 * SIZEOF_WORD], 7  ;E: w5 = 42 49 56 57 50 43 36 29
+                                                          ;        (Row 4, offset 1)
+%undef block
+%define free_bitsq  rdx
+%define free_bitsd  edx
+%define free_bitsb  dl
+    pcmpeqw     xmm1, xmm0                                ;F: w1[i] = (w1[i] == 0 ? -1 : 0);
+    shl         tempq, 48                                 ;Z:     temp <<= 48;
+    pxor        xmm2, xmm2                                ;E: w2[i] = 0;
+    pcmpgtw     xmm0, xmm5                                ;E: w0[i] = (w5[i] < 0 ? -1 : 0);
+    paddw       xmm5, xmm0                                ;E: w5[i] += w0[i];
+    or          tempq, put_buffer                         ;Z:     temp |= put_buffer;
+    movaps      XMMWORD [t + 32 * SIZEOF_WORD], xmm5      ;E: t[32+i] = w5[i];
+    lea         t, [dword t - 2]                          ;Z:     t = &t[-1];
+    pcmpeqw     xmm5, xmm2                                ;E: w5[i] = (w5[i] == 0 ? -1 : 0);
+
+    packsswb    xmm5, xmm1                                ;EF: b5[i] = w5[i], b5[i+8] = w1[i]
+                                                          ;    w/ signed saturation
+
+    add         nbitsb, byte [dctbl + c_derived_tbl.ehufsi + nbitsq]
+                                                          ;Z:     nbits += dctbl->ehufsi[nbits];
+%undef dctbl
+%define code_temp  r8d
+    pmovmskb    indexd, xmm5                              ;Z:     index = 0;  index |= ((b5[i] >> 7) << i);
+    mov         free_bitsd, [state+working_state.cur.free_bits]
+                                                          ;Z:     free_bits = state->cur.free_bits;
+    pcmpeqw     xmm1, xmm1                                ;Z:     b1[i] = 0xFF;
+    shl         index, 32                                 ;Z:     index <<= 32;
+    mov         put_buffer, [state+working_state.cur.put_buffer.simd]
+                                                          ;Z:     put_buffer = state->cur.put_buffer.simd;
+    or          index, tempq                              ;Z:     index |= temp;
+    not         index                                     ;Z:     index = ~index;
+    sub         free_bitsb, nbitsb                        ;Z:     if ((free_bits -= nbits) >= 0)
+    jnl         .ENTRY_SKIP_EMIT_CODE                     ;Z:       goto .ENTRY_SKIP_EMIT_CODE;
+    align       16
+.EMIT_CODE:                                               ;Z:     .EMIT_CODE:
+    EMIT_QWORD  .BLOOP_COND                               ;Z:     insert code, flush buffer, goto .BLOOP_COND
+
+; ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+    align       16
+.BRLOOP:                                                  ; do {
+    lea         code_temp, [nbitsq - 16]                  ;   code_temp = nbits - 16;
+    movzx       nbits, byte [actbl + c_derived_tbl.ehufsi + 0xf0]
+                                                          ;   nbits = actbl->ehufsi[0xf0];
+    mov         code, [actbl + c_derived_tbl.ehufco + 0xf0 * 4]
+                                                          ;   code = actbl->ehufco[0xf0];
+    sub         free_bitsb, nbitsb                        ;   if ((free_bits -= nbits) <= 0)
+    jle         .EMIT_BRLOOP_CODE                         ;     goto .EMIT_BRLOOP_CODE;
+    shl         put_buffer, nbitsb                        ;   put_buffer <<= nbits;
+    mov         nbits, code_temp                          ;   nbits = code_temp;
+    or          put_buffer, codeq                         ;   put_buffer |= code;
+    cmp         nbits, 16                                 ;   if (nbits <= 16)
+    jle         .ERLOOP                                   ;     break;
+    jmp         .BRLOOP                                   ; } while (1);
+
+; ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+    align       16
+    times 5     nop
+.ENTRY_SKIP_EMIT_CODE:                                    ; .ENTRY_SKIP_EMIT_CODE:
+    shl         put_buffer, nbitsb                        ; put_buffer <<= nbits;
+    or          put_buffer, codeq                         ; put_buffer |= code;
+.BLOOP_COND:                                              ; .BLOOP_COND:
+    test        index, index                              ; if (index != 0)
+    jz          .ELOOP                                    ; {
+.BLOOP:                                                   ;   do {
+    xor         nbits, nbits                              ;     nbits = 0;  /* kill tzcnt input dependency */
+    tzcnt       nbitsq, index                             ;     nbits = # of trailing 0 bits in index
+    inc         nbits                                     ;     ++nbits;
+    lea         t, [t + nbitsq * 2]                       ;     t = &t[nbits];
+    shr         index, nbitsb                             ;     index >>= nbits;
+.EMIT_BRLOOP_CODE_END:                                    ; .EMIT_BRLOOP_CODE_END:
+    cmp         nbits, 16                                 ;     if (nbits > 16)
+    jg          .BRLOOP                                   ;       goto .BRLOOP;
+.ERLOOP:                                                  ; .ERLOOP:
+    movsx       codeq, word [t]                           ;     code = *t;
+    lea         tempd, [nbitsq * 2]                       ;     temp = nbits * 2;
+    movzx       nbits, byte [NBITS(codeq)]                ;     nbits = JPEG_NBITS(code);
+    lea         tempd, [nbitsq + tempq * 8]               ;     temp = temp * 8 + nbits;
+    mov         code_temp, [actbl + c_derived_tbl.ehufco + (tempq - 16) * 4]
+                                                          ;     code_temp = actbl->ehufco[temp-16];
+    shl         code_temp, nbitsb                         ;     code_temp <<= nbits;
+    and         code, dword [MASK_BITS(nbitsq)]           ;     code &= (1 << nbits) - 1;
+    add         nbitsb, [actbl + c_derived_tbl.ehufsi + (tempq - 16)]
+                                                          ;     free_bits -= actbl->ehufsi[temp-16];
+    or          code, code_temp                           ;     code |= code_temp;
+    sub         free_bitsb, nbitsb                        ;     if ((free_bits -= nbits) <= 0)
+    jle         .EMIT_CODE                                ;       goto .EMIT_CODE;
+    shl         put_buffer, nbitsb                        ;     put_buffer <<= nbits;
+    or          put_buffer, codeq                         ;     put_buffer |= code;
+    test        index, index
+    jnz         .BLOOP                                    ;   } while (index != 0);
+.ELOOP:                                                   ; }  /* index != 0 */
+    sub         td, esp                                   ; t -= (WIN64: &t_[0], UNIX: &t_[64]);
+%ifdef WIN64
+    cmp         td, (DCTSIZE2 - 2) * SIZEOF_WORD          ; if (t != 62)
+%else
+    cmp         td, -2 * SIZEOF_WORD                      ; if (t != -2)
+%endif
+    je          .EFN                                      ; {
+    movzx       nbits, byte [actbl + c_derived_tbl.ehufsi + 0]
+                                                          ;   nbits = actbl->ehufsi[0];
+    mov         code, [actbl + c_derived_tbl.ehufco + 0]  ;   code = actbl->ehufco[0];
+    sub         free_bitsb, nbitsb                        ;   if ((free_bits -= nbits) <= 0)
+    jg          .EFN_SKIP_EMIT_CODE                       ;   {
+    EMIT_QWORD  .EFN                                      ;     insert code, flush buffer
+    align       16
+.EFN_SKIP_EMIT_CODE:                                      ;   } else {
+    shl         put_buffer, nbitsb                        ;     put_buffer <<= nbits;
+    or          put_buffer, codeq                         ;     put_buffer |= code;
+.EFN:                                                     ; } }
+    mov         [state + working_state.cur.put_buffer.simd], put_buffer
+                                                          ; state->cur.put_buffer.simd = put_buffer;
+    mov         byte [state + working_state.cur.free_bits], free_bitsb
+                                                          ; state->cur.free_bits = free_bits;
+%ifdef WIN64
+    sub         rsp, -DCTSIZE2 * SIZEOF_WORD
+    pop         r12
+    pop         rdi
+    pop         rsi
+    pop         rbp
+    pop         rbx
+%else
+    pop         r12
+    pop         rbp
+    pop         rbx
+%endif
+    ret
+
+; ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+    align       16
+.EMIT_BRLOOP_CODE:
+    EMIT_QWORD  .EMIT_BRLOOP_CODE_END, { mov nbits, code_temp }
+                                                          ; insert code, flush buffer,
+                                                          ; nbits = code_temp, goto .EMIT_BRLOOP_CODE_END
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+    align       32
diff --git a/media/libjpeg/simd/x86_64/jcphuff-sse2.asm b/media/libjpeg/simd/x86_64/jcphuff-sse2.asm
new file mode 100644
index 0000000000..01b5c0235f
--- /dev/null
+++ b/media/libjpeg/simd/x86_64/jcphuff-sse2.asm
@@ -0,0 +1,639 @@
+;
+; jcphuff-sse2.asm - prepare data for progressive Huffman encoding
+; (64-bit SSE2)
+;
+; Copyright (C) 2016, 2018, Matthieu Darbois
+;
+; Based on the x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; This file contains an SSE2 implementation of data preparation for progressive
+; Huffman encoding.  See jcphuff.c for more details.
+
+%include "jsimdext.inc"
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_TEXT
+    BITS        64
+
+; --------------------------------------------------------------------------
+; Macros to load data for jsimd_encode_mcu_AC_first_prepare_sse2() and
+; jsimd_encode_mcu_AC_refine_prepare_sse2()
+
+%macro LOAD16 0
+    pxor        N0, N0
+    pxor        N1, N1
+
+    mov         T0d, INT [LUT +  0*SIZEOF_INT]
+    mov         T1d, INT [LUT +  8*SIZEOF_INT]
+    pinsrw      X0, word [BLOCK + T0 * 2], 0
+    pinsrw      X1, word [BLOCK + T1 * 2], 0
+
+    mov         T0d, INT [LUT +  1*SIZEOF_INT]
+    mov         T1d, INT [LUT +  9*SIZEOF_INT]
+    pinsrw      X0, word [BLOCK + T0 * 2], 1
+    pinsrw      X1, word [BLOCK + T1 * 2], 1
+
+    mov         T0d, INT [LUT +  2*SIZEOF_INT]
+    mov         T1d, INT [LUT + 10*SIZEOF_INT]
+    pinsrw      X0, word [BLOCK + T0 * 2], 2
+    pinsrw      X1, word [BLOCK + T1 * 2], 2
+
+    mov         T0d, INT [LUT +  3*SIZEOF_INT]
+    mov         T1d, INT [LUT + 11*SIZEOF_INT]
+    pinsrw      X0, word [BLOCK + T0 * 2], 3
+    pinsrw      X1, word [BLOCK + T1 * 2], 3
+
+    mov         T0d, INT [LUT +  4*SIZEOF_INT]
+    mov         T1d, INT [LUT + 12*SIZEOF_INT]
+    pinsrw      X0, word [BLOCK + T0 * 2], 4
+    pinsrw      X1, word [BLOCK + T1 * 2], 4
+
+    mov         T0d, INT [LUT +  5*SIZEOF_INT]
+    mov         T1d, INT [LUT + 13*SIZEOF_INT]
+    pinsrw      X0, word [BLOCK + T0 * 2], 5
+    pinsrw      X1, word [BLOCK + T1 * 2], 5
+
+    mov         T0d, INT [LUT +  6*SIZEOF_INT]
+    mov         T1d, INT [LUT + 14*SIZEOF_INT]
+    pinsrw      X0, word [BLOCK + T0 * 2], 6
+    pinsrw      X1, word [BLOCK + T1 * 2], 6
+
+    mov         T0d, INT [LUT +  7*SIZEOF_INT]
+    mov         T1d, INT [LUT + 15*SIZEOF_INT]
+    pinsrw      X0, word [BLOCK + T0 * 2], 7
+    pinsrw      X1, word [BLOCK + T1 * 2], 7
+%endmacro
+
+%macro LOAD15 0
+    pxor        N0, N0
+    pxor        N1, N1
+    pxor        X1, X1
+
+    mov         T0d, INT [LUT +  0*SIZEOF_INT]
+    mov         T1d, INT [LUT +  8*SIZEOF_INT]
+    pinsrw      X0, word [BLOCK + T0 * 2], 0
+    pinsrw      X1, word [BLOCK + T1 * 2], 0
+
+    mov         T0d, INT [LUT +  1*SIZEOF_INT]
+    pinsrw      X0, word [BLOCK + T0 * 2], 1
+
+    mov         T0d, INT [LUT +  2*SIZEOF_INT]
+    pinsrw      X0, word [BLOCK + T0 * 2], 2
+
+    mov         T0d, INT [LUT +  3*SIZEOF_INT]
+    pinsrw      X0, word [BLOCK + T0 * 2], 3
+
+    mov         T0d, INT [LUT +  4*SIZEOF_INT]
+    pinsrw      X0, word [BLOCK + T0 * 2], 4
+
+    mov         T0d, INT [LUT +  5*SIZEOF_INT]
+    pinsrw      X0, word [BLOCK + T0 * 2], 5
+
+    mov         T0d, INT [LUT +  6*SIZEOF_INT]
+    pinsrw      X0, word [BLOCK + T0 * 2], 6
+
+    mov         T0d, INT [LUT +  7*SIZEOF_INT]
+    pinsrw      X0, word [BLOCK + T0 * 2], 7
+
+    cmp         LENEND, 2
+    jl          %%.ELOAD15
+    mov         T1d, INT [LUT +  9*SIZEOF_INT]
+    pinsrw      X1, word [BLOCK + T1 * 2], 1
+
+    cmp         LENEND, 3
+    jl          %%.ELOAD15
+    mov         T1d, INT [LUT + 10*SIZEOF_INT]
+    pinsrw      X1, word [BLOCK + T1 * 2], 2
+
+    cmp         LENEND, 4
+    jl          %%.ELOAD15
+    mov         T1d, INT [LUT + 11*SIZEOF_INT]
+    pinsrw      X1, word [BLOCK + T1 * 2], 3
+
+    cmp         LENEND, 5
+    jl          %%.ELOAD15
+    mov         T1d, INT [LUT + 12*SIZEOF_INT]
+    pinsrw      X1, word [BLOCK + T1 * 2], 4
+
+    cmp         LENEND, 6
+    jl          %%.ELOAD15
+    mov         T1d, INT [LUT + 13*SIZEOF_INT]
+    pinsrw      X1, word [BLOCK + T1 * 2], 5
+
+    cmp         LENEND, 7
+    jl          %%.ELOAD15
+    mov         T1d, INT [LUT + 14*SIZEOF_INT]
+    pinsrw      X1, word [BLOCK + T1 * 2], 6
+%%.ELOAD15:
+%endmacro
+
+%macro LOAD8 0
+    pxor        N0, N0
+
+    mov         T0d, INT [LUT +  0*SIZEOF_INT]
+    pinsrw      X0, word [BLOCK + T0 * 2], 0
+
+    mov         T0d, INT [LUT +  1*SIZEOF_INT]
+    pinsrw      X0, word [BLOCK + T0 * 2], 1
+
+    mov         T0d, INT [LUT +  2*SIZEOF_INT]
+    pinsrw      X0, word [BLOCK + T0 * 2], 2
+
+    mov         T0d, INT [LUT +  3*SIZEOF_INT]
+    pinsrw      X0, word [BLOCK + T0 * 2], 3
+
+    mov         T0d, INT [LUT +  4*SIZEOF_INT]
+    pinsrw      X0, word [BLOCK + T0 * 2], 4
+
+    mov         T0d, INT [LUT +  5*SIZEOF_INT]
+    pinsrw      X0, word [BLOCK + T0 * 2], 5
+
+    mov         T0d, INT [LUT +  6*SIZEOF_INT]
+    pinsrw      X0, word [BLOCK + T0 * 2], 6
+
+    mov         T0d, INT [LUT +  7*SIZEOF_INT]
+    pinsrw      X0, word [BLOCK + T0 * 2], 7
+%endmacro
+
+%macro LOAD7 0
+    pxor        N0, N0
+    pxor        X0, X0
+
+    mov         T1d, INT [LUT +  0*SIZEOF_INT]
+    pinsrw      X0, word [BLOCK + T1 * 2], 0
+
+    cmp         LENEND, 2
+    jl          %%.ELOAD7
+    mov         T1d, INT [LUT +  1*SIZEOF_INT]
+    pinsrw      X0, word [BLOCK + T1 * 2], 1
+
+    cmp         LENEND, 3
+    jl          %%.ELOAD7
+    mov         T1d, INT [LUT +  2*SIZEOF_INT]
+    pinsrw      X0, word [BLOCK + T1 * 2], 2
+
+    cmp         LENEND, 4
+    jl          %%.ELOAD7
+    mov         T1d, INT [LUT +  3*SIZEOF_INT]
+    pinsrw      X0, word [BLOCK + T1 * 2], 3
+
+    cmp         LENEND, 5
+    jl          %%.ELOAD7
+    mov         T1d, INT [LUT +  4*SIZEOF_INT]
+    pinsrw      X0, word [BLOCK + T1 * 2], 4
+
+    cmp         LENEND, 6
+    jl          %%.ELOAD7
+    mov         T1d, INT [LUT +  5*SIZEOF_INT]
+    pinsrw      X0, word [BLOCK + T1 * 2], 5
+
+    cmp         LENEND, 7
+    jl          %%.ELOAD7
+    mov         T1d, INT [LUT +  6*SIZEOF_INT]
+    pinsrw      X0, word [BLOCK + T1 * 2], 6
+%%.ELOAD7:
+%endmacro
+
+%macro REDUCE0 0
+    movdqa      xmm0, XMMWORD [VALUES + ( 0*2)]
+    movdqa      xmm1, XMMWORD [VALUES + ( 8*2)]
+    movdqa      xmm2, XMMWORD [VALUES + (16*2)]
+    movdqa      xmm3, XMMWORD [VALUES + (24*2)]
+    movdqa      xmm4, XMMWORD [VALUES + (32*2)]
+    movdqa      xmm5, XMMWORD [VALUES + (40*2)]
+    movdqa      xmm6, XMMWORD [VALUES + (48*2)]
+    movdqa      xmm7, XMMWORD [VALUES + (56*2)]
+
+    pcmpeqw     xmm0, ZERO
+    pcmpeqw     xmm1, ZERO
+    pcmpeqw     xmm2, ZERO
+    pcmpeqw     xmm3, ZERO
+    pcmpeqw     xmm4, ZERO
+    pcmpeqw     xmm5, ZERO
+    pcmpeqw     xmm6, ZERO
+    pcmpeqw     xmm7, ZERO
+
+    packsswb    xmm0, xmm1
+    packsswb    xmm2, xmm3
+    packsswb    xmm4, xmm5
+    packsswb    xmm6, xmm7
+
+    pmovmskb    eax, xmm0
+    pmovmskb    ecx, xmm2
+    pmovmskb    edx, xmm4
+    pmovmskb    esi, xmm6
+
+    shl         rcx, 16
+    shl         rdx, 32
+    shl         rsi, 48
+
+    or          rax, rcx
+    or          rdx, rsi
+    or          rax, rdx
+
+    not         rax
+
+    mov         MMWORD [r15], rax
+%endmacro
+
+;
+; Prepare data for jsimd_encode_mcu_AC_first().
+;
+; GLOBAL(void)
+; jsimd_encode_mcu_AC_first_prepare_sse2(const JCOEF *block,
+;                                        const int *jpeg_natural_order_start,
+;                                        int Sl, int Al, JCOEF *values,
+;                                        size_t *zerobits)
+;
+; r10 = const JCOEF *block
+; r11 = const int *jpeg_natural_order_start
+; r12 = int Sl
+; r13 = int Al
+; r14 = JCOEF *values
+; r15 = size_t *zerobits
+
+%define ZERO    xmm9
+%define X0      xmm0
+%define X1      xmm1
+%define N0      xmm2
+%define N1      xmm3
+%define AL      xmm4
+%define K       eax
+%define LUT     r11
+%define T0      rcx
+%define T0d     ecx
+%define T1      rdx
+%define T1d     edx
+%define BLOCK   r10
+%define VALUES  r14
+%define LEN     r12d
+%define LENEND  r13d
+
+    align       32
+    GLOBAL_FUNCTION(jsimd_encode_mcu_AC_first_prepare_sse2)
+
+EXTN(jsimd_encode_mcu_AC_first_prepare_sse2):
+    push        rbp
+    mov         rax, rsp                     ; rax = original rbp
+    sub         rsp, byte 4
+    and         rsp, byte (-SIZEOF_XMMWORD)  ; align to 128 bits
+    mov         [rsp], rax
+    mov         rbp, rsp                     ; rbp = aligned rbp
+    lea         rsp, [rbp - 16]
+    collect_args 6
+
+    movdqa      XMMWORD [rbp - 16], ZERO
+
+    movd        AL, r13d
+    pxor        ZERO, ZERO
+    mov         K, LEN
+    mov         LENEND, LEN
+    and         K, -16
+    and         LENEND, 7
+    shr         K, 4
+    jz          .ELOOP16
+.BLOOP16:
+    LOAD16
+    pcmpgtw     N0, X0
+    pcmpgtw     N1, X1
+    paddw       X0, N0
+    paddw       X1, N1
+    pxor        X0, N0
+    pxor        X1, N1
+    psrlw       X0, AL
+    psrlw       X1, AL
+    pxor        N0, X0
+    pxor        N1, X1
+    movdqa      XMMWORD [VALUES + (0) * 2], X0
+    movdqa      XMMWORD [VALUES + (8) * 2], X1
+    movdqa      XMMWORD [VALUES + (0 + DCTSIZE2) * 2], N0
+    movdqa      XMMWORD [VALUES + (8 + DCTSIZE2) * 2], N1
+    add         VALUES, 16*2
+    add         LUT, 16*SIZEOF_INT
+    dec         K
+    jnz         .BLOOP16
+    test        LEN, 15
+    je          .PADDING
+.ELOOP16:
+    test        LEN, 8
+    jz          .TRY7
+    test        LEN, 7
+    jz          .TRY8
+
+    LOAD15
+    pcmpgtw     N0, X0
+    pcmpgtw     N1, X1
+    paddw       X0, N0
+    paddw       X1, N1
+    pxor        X0, N0
+    pxor        X1, N1
+    psrlw       X0, AL
+    psrlw       X1, AL
+    pxor        N0, X0
+    pxor        N1, X1
+    movdqa      XMMWORD [VALUES + (0) * 2], X0
+    movdqa      XMMWORD [VALUES + (8) * 2], X1
+    movdqa      XMMWORD [VALUES + (0 + DCTSIZE2) * 2], N0
+    movdqa      XMMWORD [VALUES + (8 + DCTSIZE2) * 2], N1
+    add         VALUES, 16*2
+    jmp         .PADDING
+.TRY8:
+    LOAD8
+    pcmpgtw     N0, X0
+    paddw       X0, N0
+    pxor        X0, N0
+    psrlw       X0, AL
+    pxor        N0, X0
+    movdqa      XMMWORD [VALUES + (0) * 2], X0
+    movdqa      XMMWORD [VALUES + (0 + DCTSIZE2) * 2], N0
+    add         VALUES, 8*2
+    jmp         .PADDING
+.TRY7:
+    LOAD7
+    pcmpgtw     N0, X0
+    paddw       X0, N0
+    pxor        X0, N0
+    psrlw       X0, AL
+    pxor        N0, X0
+    movdqa      XMMWORD [VALUES + (0) * 2], X0
+    movdqa      XMMWORD [VALUES + (0 + DCTSIZE2) * 2], N0
+    add         VALUES, 8*2
+.PADDING:
+    mov         K, LEN
+    add         K, 7
+    and         K, -8
+    shr         K, 3
+    sub         K, DCTSIZE2/8
+    jz          .EPADDING
+    align       16
+.ZEROLOOP:
+    movdqa      XMMWORD [VALUES + 0], ZERO
+    add         VALUES, 8*2
+    inc         K
+    jnz         .ZEROLOOP
+.EPADDING:
+    sub         VALUES, DCTSIZE2*2
+
+    REDUCE0
+
+    movdqa      ZERO, XMMWORD [rbp - 16]
+    uncollect_args 6
+    mov         rsp, rbp                ; rsp <- aligned rbp
+    pop         rsp                     ; rsp <- original rbp
+    pop         rbp
+    ret
+
+%undef ZERO
+%undef X0
+%undef X1
+%undef N0
+%undef N1
+%undef AL
+%undef K
+%undef LUT
+%undef T0
+%undef T0d
+%undef T1
+%undef T1d
+%undef BLOCK
+%undef VALUES
+%undef LEN
+%undef LENEND
+
+;
+; Prepare data for jsimd_encode_mcu_AC_refine().
+;
+; GLOBAL(int)
+; jsimd_encode_mcu_AC_refine_prepare_sse2(const JCOEF *block,
+;                                         const int *jpeg_natural_order_start,
+;                                         int Sl, int Al, JCOEF *absvalues,
+;                                         size_t *bits)
+;
+; r10 = const JCOEF *block
+; r11 = const int *jpeg_natural_order_start
+; r12 = int Sl
+; r13 = int Al
+; r14 = JCOEF *values
+; r15 = size_t *bits
+
+%define ZERO    xmm9
+%define ONE     xmm5
+%define X0      xmm0
+%define X1      xmm1
+%define N0      xmm2
+%define N1      xmm3
+%define AL      xmm4
+%define K       eax
+%define KK      r9d
+%define EOB     r8d
+%define SIGN    rdi
+%define LUT     r11
+%define T0      rcx
+%define T0d     ecx
+%define T1      rdx
+%define T1d     edx
+%define BLOCK   r10
+%define VALUES  r14
+%define LEN     r12d
+%define LENEND  r13d
+
+    align       32
+    GLOBAL_FUNCTION(jsimd_encode_mcu_AC_refine_prepare_sse2)
+
+EXTN(jsimd_encode_mcu_AC_refine_prepare_sse2):
+    push        rbp
+    mov         rax, rsp                     ; rax = original rbp
+    sub         rsp, byte 4
+    and         rsp, byte (-SIZEOF_XMMWORD)  ; align to 128 bits
+    mov         [rsp], rax
+    mov         rbp, rsp                     ; rbp = aligned rbp
+    lea         rsp, [rbp - 16]
+    collect_args 6
+
+    movdqa      XMMWORD [rbp - 16], ZERO
+
+    xor         SIGN, SIGN
+    xor         EOB, EOB
+    xor         KK, KK
+    movd        AL, r13d
+    pxor        ZERO, ZERO
+    pcmpeqw     ONE, ONE
+    psrlw       ONE, 15
+    mov         K, LEN
+    mov         LENEND, LEN
+    and         K, -16
+    and         LENEND, 7
+    shr         K, 4
+    jz          .ELOOPR16
+.BLOOPR16:
+    LOAD16
+    pcmpgtw     N0, X0
+    pcmpgtw     N1, X1
+    paddw       X0, N0
+    paddw       X1, N1
+    pxor        X0, N0
+    pxor        X1, N1
+    psrlw       X0, AL
+    psrlw       X1, AL
+    movdqa      XMMWORD [VALUES + (0) * 2], X0
+    movdqa      XMMWORD [VALUES + (8) * 2], X1
+    pcmpeqw     X0, ONE
+    pcmpeqw     X1, ONE
+    packsswb    N0, N1
+    packsswb    X0, X1
+    pmovmskb    T0d, N0                 ; lsignbits.val16u[k>>4] = _mm_movemask_epi8(neg);
+    pmovmskb    T1d, X0                 ; idx = _mm_movemask_epi8(x1);
+    shr         SIGN, 16                ; make room for sizebits
+    shl         T0, 48
+    or          SIGN, T0
+    bsr         T1d, T1d                ;  idx = 16 - (__builtin_clz(idx)>>1);
+    jz          .CONTINUER16            ; if (idx) {
+    mov         EOB, KK
+    add         EOB, T1d                ; EOB = k + idx;
+.CONTINUER16:
+    add         VALUES, 16*2
+    add         LUT, 16*SIZEOF_INT
+    add         KK, 16
+    dec         K
+    jnz         .BLOOPR16
+    test        LEN, 15
+    je          .PADDINGR
+.ELOOPR16:
+    test        LEN, 8
+    jz          .TRYR7
+    test        LEN, 7
+    jz          .TRYR8
+
+    LOAD15
+    pcmpgtw     N0, X0
+    pcmpgtw     N1, X1
+    paddw       X0, N0
+    paddw       X1, N1
+    pxor        X0, N0
+    pxor        X1, N1
+    psrlw       X0, AL
+    psrlw       X1, AL
+    movdqa      XMMWORD [VALUES + (0) * 2], X0
+    movdqa      XMMWORD [VALUES + (8) * 2], X1
+    pcmpeqw     X0, ONE
+    pcmpeqw     X1, ONE
+    packsswb    N0, N1
+    packsswb    X0, X1
+    pmovmskb    T0d, N0                 ; lsignbits.val16u[k>>4] = _mm_movemask_epi8(neg);
+    pmovmskb    T1d, X0                 ; idx = _mm_movemask_epi8(x1);
+    shr         SIGN, 16                ; make room for sizebits
+    shl         T0, 48
+    or          SIGN, T0
+    bsr         T1d, T1d                ;  idx = 16 - (__builtin_clz(idx)>>1);
+    jz          .CONTINUER15            ; if (idx) {
+    mov         EOB, KK
+    add         EOB, T1d                ; EOB = k + idx;
+.CONTINUER15:
+    add         VALUES, 16*2
+    jmp         .PADDINGR
+.TRYR8:
+    LOAD8
+
+    pcmpgtw     N0, X0
+    paddw       X0, N0
+    pxor        X0, N0
+    psrlw       X0, AL
+    movdqa      XMMWORD [VALUES + (0) * 2], X0
+    pcmpeqw     X0, ONE
+    packsswb    N0, ZERO
+    packsswb    X0, ZERO
+    pmovmskb    T0d, N0                 ; lsignbits.val16u[k>>4] = _mm_movemask_epi8(neg);
+    pmovmskb    T1d, X0                 ; idx = _mm_movemask_epi8(x1);
+    shr         SIGN, 8                 ; make room for sizebits
+    shl         T0, 56
+    or          SIGN, T0
+    bsr         T1d, T1d                ;  idx = 16 - (__builtin_clz(idx)>>1);
+    jz          .CONTINUER8             ; if (idx) {
+    mov         EOB, KK
+    add         EOB, T1d                ; EOB = k + idx;
+.CONTINUER8:
+    add         VALUES, 8*2
+    jmp         .PADDINGR
+.TRYR7:
+    LOAD7
+
+    pcmpgtw     N0, X0
+    paddw       X0, N0
+    pxor        X0, N0
+    psrlw       X0, AL
+    movdqa      XMMWORD [VALUES + (0) * 2], X0
+    pcmpeqw     X0, ONE
+    packsswb    N0, ZERO
+    packsswb    X0, ZERO
+    pmovmskb    T0d, N0                 ; lsignbits.val16u[k>>4] = _mm_movemask_epi8(neg);
+    pmovmskb    T1d, X0                 ; idx = _mm_movemask_epi8(x1);
+    shr         SIGN, 8                 ; make room for sizebits
+    shl         T0, 56
+    or          SIGN, T0
+    bsr         T1d, T1d                ; idx = 16 - (__builtin_clz(idx)>>1);
+    jz          .CONTINUER7             ; if (idx) {
+    mov         EOB, KK
+    add         EOB, T1d                ; EOB = k + idx;
+.CONTINUER7:
+    add         VALUES, 8*2
+.PADDINGR:
+    mov         K, LEN
+    add         K, 7
+    and         K, -8
+    shr         K, 3
+    sub         K, DCTSIZE2/8
+    jz          .EPADDINGR
+    align       16
+.ZEROLOOPR:
+    movdqa      XMMWORD [VALUES + 0], ZERO
+    shr         SIGN, 8
+    add         VALUES, 8*2
+    inc         K
+    jnz         .ZEROLOOPR
+.EPADDINGR:
+    not         SIGN
+    sub         VALUES, DCTSIZE2*2
+    mov         MMWORD [r15+SIZEOF_MMWORD], SIGN
+
+    REDUCE0
+
+    mov         eax, EOB
+    movdqa      ZERO, XMMWORD [rbp - 16]
+    uncollect_args 6
+    mov         rsp, rbp                ; rsp <- aligned rbp
+    pop         rsp                     ; rsp <- original rbp
+    pop         rbp
+    ret
+
+%undef ZERO
+%undef ONE
+%undef X0
+%undef X1
+%undef N0
+%undef N1
+%undef AL
+%undef K
+%undef KK
+%undef EOB
+%undef SIGN
+%undef LUT
+%undef T0
+%undef T0d
+%undef T1
+%undef T1d
+%undef BLOCK
+%undef VALUES
+%undef LEN
+%undef LENEND
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+    align       32
diff --git a/media/libjpeg/simd/x86_64/jcsample-avx2.asm b/media/libjpeg/simd/x86_64/jcsample-avx2.asm
new file mode 100644
index 0000000000..b32527aebe
--- /dev/null
+++ b/media/libjpeg/simd/x86_64/jcsample-avx2.asm
@@ -0,0 +1,367 @@
+;
+; jcsample.asm - downsampling (64-bit AVX2)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+; Copyright (C) 2009, 2016, D. R. Commander.
+; Copyright (C) 2015, Intel Corporation.
+; Copyright (C) 2018, Matthias Räncker.
+;
+; Based on the x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+
+%include "jsimdext.inc"
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_TEXT
+    BITS        64
+;
+; Downsample pixel values of a single component.
+; This version handles the common case of 2:1 horizontal and 1:1 vertical,
+; without smoothing.
+;
+; GLOBAL(void)
+; jsimd_h2v1_downsample_avx2(JDIMENSION image_width, int max_v_samp_factor,
+;                            JDIMENSION v_samp_factor,
+;                            JDIMENSION width_in_blocks, JSAMPARRAY input_data,
+;                            JSAMPARRAY output_data);
+;
+
+; r10d = JDIMENSION image_width
+; r11 = int max_v_samp_factor
+; r12d = JDIMENSION v_samp_factor
+; r13d = JDIMENSION width_in_blocks
+; r14 = JSAMPARRAY input_data
+; r15 = JSAMPARRAY output_data
+
+    align       32
+    GLOBAL_FUNCTION(jsimd_h2v1_downsample_avx2)
+
+EXTN(jsimd_h2v1_downsample_avx2):
+    push        rbp
+    mov         rax, rsp
+    mov         rbp, rsp
+    collect_args 6
+
+    mov         ecx, r13d
+    shl         rcx, 3                  ; imul rcx,DCTSIZE (rcx = output_cols)
+    jz          near .return
+
+    mov         edx, r10d
+
+    ; -- expand_right_edge
+
+    push        rcx
+    shl         rcx, 1                  ; output_cols * 2
+    sub         rcx, rdx
+    jle         short .expand_end
+
+    mov         rax, r11
+    test        rax, rax
+    jle         short .expand_end
+
+    cld
+    mov         rsi, r14                ; input_data
+.expandloop:
+    push        rax
+    push        rcx
+
+    mov         rdip, JSAMPROW [rsi]
+    add         rdi, rdx
+    mov         al, JSAMPLE [rdi-1]
+
+    rep stosb
+
+    pop         rcx
+    pop         rax
+
+    add         rsi, byte SIZEOF_JSAMPROW
+    dec         rax
+    jg          short .expandloop
+
+.expand_end:
+    pop         rcx                     ; output_cols
+
+    ; -- h2v1_downsample
+
+    mov         eax, r12d               ; rowctr
+    test        eax, eax
+    jle         near .return
+
+    mov         rdx, 0x00010000         ; bias pattern
+    vmovd       xmm7, edx
+    vpshufd     xmm7, xmm7, 0x00        ; xmm7={0, 1, 0, 1, 0, 1, 0, 1}
+    vperm2i128  ymm7, ymm7, ymm7, 0     ; ymm7={xmm7, xmm7}
+    vpcmpeqw    ymm6, ymm6, ymm6
+    vpsrlw      ymm6, ymm6, BYTE_BIT    ; ymm6={0xFF 0x00 0xFF 0x00 ..}
+
+    mov         rsi, r14                ; input_data
+    mov         rdi, r15                ; output_data
+.rowloop:
+    push        rcx
+    push        rdi
+    push        rsi
+
+    mov         rsip, JSAMPROW [rsi]    ; inptr
+    mov         rdip, JSAMPROW [rdi]    ; outptr
+
+    cmp         rcx, byte SIZEOF_YMMWORD
+    jae         short .columnloop
+
+.columnloop_r24:
+    ; rcx can possibly be 8, 16, 24
+    cmp         rcx, 24
+    jne         .columnloop_r16
+    vmovdqu     ymm0, YMMWORD [rsi+0*SIZEOF_YMMWORD]
+    vmovdqu     xmm1, XMMWORD [rsi+1*SIZEOF_YMMWORD]
+    mov         rcx, SIZEOF_YMMWORD
+    jmp         short .downsample
+
+.columnloop_r16:
+    cmp         rcx, 16
+    jne         .columnloop_r8
+    vmovdqu     ymm0, YMMWORD [rsi+0*SIZEOF_YMMWORD]
+    vpxor       ymm1, ymm1, ymm1
+    mov         rcx, SIZEOF_YMMWORD
+    jmp         short .downsample
+
+.columnloop_r8:
+    vmovdqu     xmm0, XMMWORD[rsi+0*SIZEOF_YMMWORD]
+    vpxor       ymm1, ymm1, ymm1
+    mov         rcx, SIZEOF_YMMWORD
+    jmp         short .downsample
+
+.columnloop:
+    vmovdqu     ymm0, YMMWORD [rsi+0*SIZEOF_YMMWORD]
+    vmovdqu     ymm1, YMMWORD [rsi+1*SIZEOF_YMMWORD]
+
+.downsample:
+    vpsrlw      ymm2, ymm0, BYTE_BIT
+    vpand       ymm0, ymm0, ymm6
+    vpsrlw      ymm3, ymm1, BYTE_BIT
+    vpand       ymm1, ymm1, ymm6
+
+    vpaddw      ymm0, ymm0, ymm2
+    vpaddw      ymm1, ymm1, ymm3
+    vpaddw      ymm0, ymm0, ymm7
+    vpaddw      ymm1, ymm1, ymm7
+    vpsrlw      ymm0, ymm0, 1
+    vpsrlw      ymm1, ymm1, 1
+
+    vpackuswb   ymm0, ymm0, ymm1
+    vpermq      ymm0, ymm0, 0xd8
+
+    vmovdqu     YMMWORD [rdi+0*SIZEOF_YMMWORD], ymm0
+
+    sub         rcx, byte SIZEOF_YMMWORD    ; outcol
+    add         rsi, byte 2*SIZEOF_YMMWORD  ; inptr
+    add         rdi, byte 1*SIZEOF_YMMWORD  ; outptr
+    cmp         rcx, byte SIZEOF_YMMWORD
+    jae         short .columnloop
+    test        rcx, rcx
+    jnz         near .columnloop_r24
+
+    pop         rsi
+    pop         rdi
+    pop         rcx
+
+    add         rsi, byte SIZEOF_JSAMPROW  ; input_data
+    add         rdi, byte SIZEOF_JSAMPROW  ; output_data
+    dec         rax                        ; rowctr
+    jg          near .rowloop
+
+.return:
+    vzeroupper
+    uncollect_args 6
+    pop         rbp
+    ret
+
+; --------------------------------------------------------------------------
+;
+; Downsample pixel values of a single component.
+; This version handles the standard case of 2:1 horizontal and 2:1 vertical,
+; without smoothing.
+;
+; GLOBAL(void)
+; jsimd_h2v2_downsample_avx2(JDIMENSION image_width, int max_v_samp_factor,
+;                            JDIMENSION v_samp_factor,
+;                            JDIMENSION width_in_blocks, JSAMPARRAY input_data,
+;                            JSAMPARRAY output_data);
+;
+
+; r10d = JDIMENSION image_width
+; r11 = int max_v_samp_factor
+; r12d = JDIMENSION v_samp_factor
+; r13d = JDIMENSION width_in_blocks
+; r14 = JSAMPARRAY input_data
+; r15 = JSAMPARRAY output_data
+
+    align       32
+    GLOBAL_FUNCTION(jsimd_h2v2_downsample_avx2)
+
+EXTN(jsimd_h2v2_downsample_avx2):
+    push        rbp
+    mov         rax, rsp
+    mov         rbp, rsp
+    collect_args 6
+
+    mov         ecx, r13d
+    shl         rcx, 3                  ; imul rcx,DCTSIZE (rcx = output_cols)
+    jz          near .return
+
+    mov         edx, r10d
+
+    ; -- expand_right_edge
+
+    push        rcx
+    shl         rcx, 1                  ; output_cols * 2
+    sub         rcx, rdx
+    jle         short .expand_end
+
+    mov         rax, r11
+    test        rax, rax
+    jle         short .expand_end
+
+    cld
+    mov         rsi, r14                ; input_data
+.expandloop:
+    push        rax
+    push        rcx
+
+    mov         rdip, JSAMPROW [rsi]
+    add         rdi, rdx
+    mov         al, JSAMPLE [rdi-1]
+
+    rep stosb
+
+    pop         rcx
+    pop         rax
+
+    add         rsi, byte SIZEOF_JSAMPROW
+    dec         rax
+    jg          short .expandloop
+
+.expand_end:
+    pop         rcx                     ; output_cols
+
+    ; -- h2v2_downsample
+
+    mov         eax, r12d               ; rowctr
+    test        rax, rax
+    jle         near .return
+
+    mov         rdx, 0x00020001         ; bias pattern
+    vmovd       xmm7, edx
+    vpcmpeqw    ymm6, ymm6, ymm6
+    vpshufd     xmm7, xmm7, 0x00        ; ymm7={1, 2, 1, 2, 1, 2, 1, 2}
+    vperm2i128  ymm7, ymm7, ymm7, 0
+    vpsrlw      ymm6, ymm6, BYTE_BIT    ; ymm6={0xFF 0x00 0xFF 0x00 ..}
+
+    mov         rsi, r14                ; input_data
+    mov         rdi, r15                ; output_data
+.rowloop:
+    push        rcx
+    push        rdi
+    push        rsi
+
+    mov         rdxp, JSAMPROW [rsi+0*SIZEOF_JSAMPROW]  ; inptr0
+    mov         rsip, JSAMPROW [rsi+1*SIZEOF_JSAMPROW]  ; inptr1
+    mov         rdip, JSAMPROW [rdi]                    ; outptr
+
+    cmp         rcx, byte SIZEOF_YMMWORD
+    jae         short .columnloop
+
+.columnloop_r24:
+    cmp         rcx, 24
+    jne         .columnloop_r16
+    vmovdqu     ymm0, YMMWORD [rdx+0*SIZEOF_YMMWORD]
+    vmovdqu     ymm1, YMMWORD [rsi+0*SIZEOF_YMMWORD]
+    vmovdqu     xmm2, XMMWORD [rdx+1*SIZEOF_YMMWORD]
+    vmovdqu     xmm3, XMMWORD [rsi+1*SIZEOF_YMMWORD]
+    mov         rcx, SIZEOF_YMMWORD
+    jmp         short .downsample
+
+.columnloop_r16:
+    cmp         rcx, 16
+    jne         .columnloop_r8
+    vmovdqu     ymm0, YMMWORD [rdx+0*SIZEOF_YMMWORD]
+    vmovdqu     ymm1, YMMWORD [rsi+0*SIZEOF_YMMWORD]
+    vpxor       ymm2, ymm2, ymm2
+    vpxor       ymm3, ymm3, ymm3
+    mov         rcx, SIZEOF_YMMWORD
+    jmp         short .downsample
+
+.columnloop_r8:
+    vmovdqu     xmm0, XMMWORD [rdx+0*SIZEOF_XMMWORD]
+    vmovdqu     xmm1, XMMWORD [rsi+0*SIZEOF_XMMWORD]
+    vpxor       ymm2, ymm2, ymm2
+    vpxor       ymm3, ymm3, ymm3
+    mov         rcx, SIZEOF_YMMWORD
+    jmp         short .downsample
+
+.columnloop:
+    vmovdqu     ymm0, YMMWORD [rdx+0*SIZEOF_YMMWORD]
+    vmovdqu     ymm1, YMMWORD [rsi+0*SIZEOF_YMMWORD]
+    vmovdqu     ymm2, YMMWORD [rdx+1*SIZEOF_YMMWORD]
+    vmovdqu     ymm3, YMMWORD [rsi+1*SIZEOF_YMMWORD]
+
+.downsample:
+    vpand       ymm4, ymm0, ymm6
+    vpsrlw      ymm0, ymm0, BYTE_BIT
+    vpand       ymm5, ymm1, ymm6
+    vpsrlw      ymm1, ymm1, BYTE_BIT
+    vpaddw      ymm0, ymm0, ymm4
+    vpaddw      ymm1, ymm1, ymm5
+
+    vpand       ymm4, ymm2, ymm6
+    vpsrlw      ymm2, ymm2, BYTE_BIT
+    vpand       ymm5, ymm3, ymm6
+    vpsrlw      ymm3, ymm3, BYTE_BIT
+    vpaddw      ymm2, ymm2, ymm4
+    vpaddw      ymm3, ymm3, ymm5
+
+    vpaddw      ymm0, ymm0, ymm1
+    vpaddw      ymm2, ymm2, ymm3
+    vpaddw      ymm0, ymm0, ymm7
+    vpaddw      ymm2, ymm2, ymm7
+    vpsrlw      ymm0, ymm0, 2
+    vpsrlw      ymm2, ymm2, 2
+
+    vpackuswb   ymm0, ymm0, ymm2
+    vpermq      ymm0, ymm0, 0xd8
+
+    vmovdqu     YMMWORD [rdi+0*SIZEOF_YMMWORD], ymm0
+
+    sub         rcx, byte SIZEOF_YMMWORD    ; outcol
+    add         rdx, byte 2*SIZEOF_YMMWORD  ; inptr0
+    add         rsi, byte 2*SIZEOF_YMMWORD  ; inptr1
+    add         rdi, byte 1*SIZEOF_YMMWORD  ; outptr
+    cmp         rcx, byte SIZEOF_YMMWORD
+    jae         near .columnloop
+    test        rcx, rcx
+    jnz         near .columnloop_r24
+
+    pop         rsi
+    pop         rdi
+    pop         rcx
+
+    add         rsi, byte 2*SIZEOF_JSAMPROW  ; input_data
+    add         rdi, byte 1*SIZEOF_JSAMPROW  ; output_data
+    dec         rax                          ; rowctr
+    jg          near .rowloop
+
+.return:
+    vzeroupper
+    uncollect_args 6
+    pop         rbp
+    ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+    align       32
diff --git a/media/libjpeg/simd/x86_64/jcsample-sse2.asm b/media/libjpeg/simd/x86_64/jcsample-sse2.asm
new file mode 100644
index 0000000000..2fcfe4567a
--- /dev/null
+++ b/media/libjpeg/simd/x86_64/jcsample-sse2.asm
@@ -0,0 +1,330 @@
+;
+; jcsample.asm - downsampling (64-bit SSE2)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+; Copyright (C) 2009, 2016, D. R. Commander.
+; Copyright (C) 2018, Matthias Räncker.
+;
+; Based on the x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+
+%include "jsimdext.inc"
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_TEXT
+    BITS        64
+;
+; Downsample pixel values of a single component.
+; This version handles the common case of 2:1 horizontal and 1:1 vertical,
+; without smoothing.
+;
+; GLOBAL(void)
+; jsimd_h2v1_downsample_sse2(JDIMENSION image_width, int max_v_samp_factor,
+;                            JDIMENSION v_samp_factor,
+;                            JDIMENSION width_in_blocks, JSAMPARRAY input_data,
+;                            JSAMPARRAY output_data);
+;
+
+; r10d = JDIMENSION image_width
+; r11 = int max_v_samp_factor
+; r12d = JDIMENSION v_samp_factor
+; r13d = JDIMENSION width_in_blocks
+; r14 = JSAMPARRAY input_data
+; r15 = JSAMPARRAY output_data
+
+    align       32
+    GLOBAL_FUNCTION(jsimd_h2v1_downsample_sse2)
+
+EXTN(jsimd_h2v1_downsample_sse2):
+    push        rbp
+    mov         rax, rsp
+    mov         rbp, rsp
+    collect_args 6
+
+    mov         ecx, r13d
+    shl         rcx, 3                  ; imul rcx,DCTSIZE (rcx = output_cols)
+    jz          near .return
+
+    mov         edx, r10d
+
+    ; -- expand_right_edge
+
+    push        rcx
+    shl         rcx, 1                  ; output_cols * 2
+    sub         rcx, rdx
+    jle         short .expand_end
+
+    mov         rax, r11
+    test        rax, rax
+    jle         short .expand_end
+
+    cld
+    mov         rsi, r14                ; input_data
+.expandloop:
+    push        rax
+    push        rcx
+
+    mov         rdip, JSAMPROW [rsi]
+    add         rdi, rdx
+    mov         al, JSAMPLE [rdi-1]
+
+    rep stosb
+
+    pop         rcx
+    pop         rax
+
+    add         rsi, byte SIZEOF_JSAMPROW
+    dec         rax
+    jg          short .expandloop
+
+.expand_end:
+    pop         rcx                     ; output_cols
+
+    ; -- h2v1_downsample
+
+    mov         eax, r12d               ; rowctr
+    test        eax, eax
+    jle         near .return
+
+    mov         rdx, 0x00010000         ; bias pattern
+    movd        xmm7, edx
+    pcmpeqw     xmm6, xmm6
+    pshufd      xmm7, xmm7, 0x00        ; xmm7={0, 1, 0, 1, 0, 1, 0, 1}
+    psrlw       xmm6, BYTE_BIT          ; xmm6={0xFF 0x00 0xFF 0x00 ..}
+
+    mov         rsi, r14                ; input_data
+    mov         rdi, r15                ; output_data
+.rowloop:
+    push        rcx
+    push        rdi
+    push        rsi
+
+    mov         rsip, JSAMPROW [rsi]    ; inptr
+    mov         rdip, JSAMPROW [rdi]    ; outptr
+
+    cmp         rcx, byte SIZEOF_XMMWORD
+    jae         short .columnloop
+
+.columnloop_r8:
+    movdqa      xmm0, XMMWORD [rsi+0*SIZEOF_XMMWORD]
+    pxor        xmm1, xmm1
+    mov         rcx, SIZEOF_XMMWORD
+    jmp         short .downsample
+
+.columnloop:
+    movdqa      xmm0, XMMWORD [rsi+0*SIZEOF_XMMWORD]
+    movdqa      xmm1, XMMWORD [rsi+1*SIZEOF_XMMWORD]
+
+.downsample:
+    movdqa      xmm2, xmm0
+    movdqa      xmm3, xmm1
+
+    pand        xmm0, xmm6
+    psrlw       xmm2, BYTE_BIT
+    pand        xmm1, xmm6
+    psrlw       xmm3, BYTE_BIT
+
+    paddw       xmm0, xmm2
+    paddw       xmm1, xmm3
+    paddw       xmm0, xmm7
+    paddw       xmm1, xmm7
+    psrlw       xmm0, 1
+    psrlw       xmm1, 1
+
+    packuswb    xmm0, xmm1
+
+    movdqa      XMMWORD [rdi+0*SIZEOF_XMMWORD], xmm0
+
+    sub         rcx, byte SIZEOF_XMMWORD    ; outcol
+    add         rsi, byte 2*SIZEOF_XMMWORD  ; inptr
+    add         rdi, byte 1*SIZEOF_XMMWORD  ; outptr
+    cmp         rcx, byte SIZEOF_XMMWORD
+    jae         short .columnloop
+    test        rcx, rcx
+    jnz         short .columnloop_r8
+
+    pop         rsi
+    pop         rdi
+    pop         rcx
+
+    add         rsi, byte SIZEOF_JSAMPROW  ; input_data
+    add         rdi, byte SIZEOF_JSAMPROW  ; output_data
+    dec         rax                        ; rowctr
+    jg          near .rowloop
+
+.return:
+    uncollect_args 6
+    pop         rbp
+    ret
+
+; --------------------------------------------------------------------------
+;
+; Downsample pixel values of a single component.
+; This version handles the standard case of 2:1 horizontal and 2:1 vertical,
+; without smoothing.
+;
+; GLOBAL(void)
+; jsimd_h2v2_downsample_sse2(JDIMENSION image_width, int max_v_samp_factor,
+;                            JDIMENSION v_samp_factor,
+;                            JDIMENSION width_in_blocks, JSAMPARRAY input_data,
+;                            JSAMPARRAY output_data);
+;
+
+; r10d = JDIMENSION image_width
+; r11 = int max_v_samp_factor
+; r12d = JDIMENSION v_samp_factor
+; r13d = JDIMENSION width_in_blocks
+; r14 = JSAMPARRAY input_data
+; r15 = JSAMPARRAY output_data
+
+    align       32
+    GLOBAL_FUNCTION(jsimd_h2v2_downsample_sse2)
+
+EXTN(jsimd_h2v2_downsample_sse2):
+    push        rbp
+    mov         rax, rsp
+    mov         rbp, rsp
+    collect_args 6
+
+    mov         ecx, r13d
+    shl         rcx, 3                  ; imul rcx,DCTSIZE (rcx = output_cols)
+    jz          near .return
+
+    mov         edx, r10d
+
+    ; -- expand_right_edge
+
+    push        rcx
+    shl         rcx, 1                  ; output_cols * 2
+    sub         rcx, rdx
+    jle         short .expand_end
+
+    mov         rax, r11
+    test        rax, rax
+    jle         short .expand_end
+
+    cld
+    mov         rsi, r14                ; input_data
+.expandloop:
+    push        rax
+    push        rcx
+
+    mov         rdip, JSAMPROW [rsi]
+    add         rdi, rdx
+    mov         al, JSAMPLE [rdi-1]
+
+    rep stosb
+
+    pop         rcx
+    pop         rax
+
+    add         rsi, byte SIZEOF_JSAMPROW
+    dec         rax
+    jg          short .expandloop
+
+.expand_end:
+    pop         rcx                     ; output_cols
+
+    ; -- h2v2_downsample
+
+    mov         eax, r12d               ; rowctr
+    test        rax, rax
+    jle         near .return
+
+    mov         rdx, 0x00020001         ; bias pattern
+    movd        xmm7, edx
+    pcmpeqw     xmm6, xmm6
+    pshufd      xmm7, xmm7, 0x00        ; xmm7={1, 2, 1, 2, 1, 2, 1, 2}
+    psrlw       xmm6, BYTE_BIT          ; xmm6={0xFF 0x00 0xFF 0x00 ..}
+
+    mov         rsi, r14                ; input_data
+    mov         rdi, r15                ; output_data
+.rowloop:
+    push        rcx
+    push        rdi
+    push        rsi
+
+    mov         rdxp, JSAMPROW [rsi+0*SIZEOF_JSAMPROW]  ; inptr0
+    mov         rsip, JSAMPROW [rsi+1*SIZEOF_JSAMPROW]  ; inptr1
+    mov         rdip, JSAMPROW [rdi]                    ; outptr
+
+    cmp         rcx, byte SIZEOF_XMMWORD
+    jae         short .columnloop
+
+.columnloop_r8:
+    movdqa      xmm0, XMMWORD [rdx+0*SIZEOF_XMMWORD]
+    movdqa      xmm1, XMMWORD [rsi+0*SIZEOF_XMMWORD]
+    pxor        xmm2, xmm2
+    pxor        xmm3, xmm3
+    mov         rcx, SIZEOF_XMMWORD
+    jmp         short .downsample
+
+.columnloop:
+    movdqa      xmm0, XMMWORD [rdx+0*SIZEOF_XMMWORD]
+    movdqa      xmm1, XMMWORD [rsi+0*SIZEOF_XMMWORD]
+    movdqa      xmm2, XMMWORD [rdx+1*SIZEOF_XMMWORD]
+    movdqa      xmm3, XMMWORD [rsi+1*SIZEOF_XMMWORD]
+
+.downsample:
+    movdqa      xmm4, xmm0
+    movdqa      xmm5, xmm1
+    pand        xmm0, xmm6
+    psrlw       xmm4, BYTE_BIT
+    pand        xmm1, xmm6
+    psrlw       xmm5, BYTE_BIT
+    paddw       xmm0, xmm4
+    paddw       xmm1, xmm5
+
+    movdqa      xmm4, xmm2
+    movdqa      xmm5, xmm3
+    pand        xmm2, xmm6
+    psrlw       xmm4, BYTE_BIT
+    pand        xmm3, xmm6
+    psrlw       xmm5, BYTE_BIT
+    paddw       xmm2, xmm4
+    paddw       xmm3, xmm5
+
+    paddw       xmm0, xmm1
+    paddw       xmm2, xmm3
+    paddw       xmm0, xmm7
+    paddw       xmm2, xmm7
+    psrlw       xmm0, 2
+    psrlw       xmm2, 2
+
+    packuswb    xmm0, xmm2
+
+    movdqa      XMMWORD [rdi+0*SIZEOF_XMMWORD], xmm0
+
+    sub         rcx, byte SIZEOF_XMMWORD    ; outcol
+    add         rdx, byte 2*SIZEOF_XMMWORD  ; inptr0
+    add         rsi, byte 2*SIZEOF_XMMWORD  ; inptr1
+    add         rdi, byte 1*SIZEOF_XMMWORD  ; outptr
+    cmp         rcx, byte SIZEOF_XMMWORD
+    jae         near .columnloop
+    test        rcx, rcx
+    jnz         near .columnloop_r8
+
+    pop         rsi
+    pop         rdi
+    pop         rcx
+
+    add         rsi, byte 2*SIZEOF_JSAMPROW  ; input_data
+    add         rdi, byte 1*SIZEOF_JSAMPROW  ; output_data
+    dec         rax                          ; rowctr
+    jg          near .rowloop
+
+.return:
+    uncollect_args 6
+    pop         rbp
+    ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+    align       32
diff --git a/media/libjpeg/simd/x86_64/jdcolext-avx2.asm b/media/libjpeg/simd/x86_64/jdcolext-avx2.asm
new file mode 100644
index 0000000000..2370fda642
--- /dev/null
+++ b/media/libjpeg/simd/x86_64/jdcolext-avx2.asm
@@ -0,0 +1,496 @@
+;
+; jdcolext.asm - colorspace conversion (64-bit AVX2)
+;
+; Copyright 2009, 2012 Pierre Ossman <ossman@cendio.se> for Cendio AB
+; Copyright (C) 2009, 2012, 2016, D. R. Commander.
+; Copyright (C) 2015, Intel Corporation.
+; Copyright (C) 2018, Matthias Räncker.
+;
+; Based on the x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+
+%include "jcolsamp.inc"
+
+; --------------------------------------------------------------------------
+;
+; Convert some rows of samples to the output colorspace.
+;
+; GLOBAL(void)
+; jsimd_ycc_rgb_convert_avx2(JDIMENSION out_width, JSAMPIMAGE input_buf,
+;                            JDIMENSION input_row, JSAMPARRAY output_buf,
+;                            int num_rows)
+;
+
+; r10d = JDIMENSION out_width
+; r11 = JSAMPIMAGE input_buf
+; r12d = JDIMENSION input_row
+; r13 = JSAMPARRAY output_buf
+; r14d = int num_rows
+
+%define wk(i)   rbp - (WK_NUM - (i)) * SIZEOF_YMMWORD  ; ymmword wk[WK_NUM]
+%define WK_NUM  2
+
+    align       32
+    GLOBAL_FUNCTION(jsimd_ycc_rgb_convert_avx2)
+
+EXTN(jsimd_ycc_rgb_convert_avx2):
+    push        rbp
+    mov         rax, rsp                     ; rax = original rbp
+    sub         rsp, byte 4
+    and         rsp, byte (-SIZEOF_YMMWORD)  ; align to 256 bits
+    mov         [rsp], rax
+    mov         rbp, rsp                     ; rbp = aligned rbp
+    lea         rsp, [wk(0)]
+    collect_args 5
+    push        rbx
+
+    mov         ecx, r10d               ; num_cols
+    test        rcx, rcx
+    jz          near .return
+
+    push        rcx
+
+    mov         rdi, r11
+    mov         ecx, r12d
+    mov         rsip, JSAMPARRAY [rdi+0*SIZEOF_JSAMPARRAY]
+    mov         rbxp, JSAMPARRAY [rdi+1*SIZEOF_JSAMPARRAY]
+    mov         rdxp, JSAMPARRAY [rdi+2*SIZEOF_JSAMPARRAY]
+    lea         rsi, [rsi+rcx*SIZEOF_JSAMPROW]
+    lea         rbx, [rbx+rcx*SIZEOF_JSAMPROW]
+    lea         rdx, [rdx+rcx*SIZEOF_JSAMPROW]
+
+    pop         rcx
+
+    mov         rdi, r13
+    mov         eax, r14d
+    test        rax, rax
+    jle         near .return
+.rowloop:
+    push        rax
+    push        rdi
+    push        rdx
+    push        rbx
+    push        rsi
+    push        rcx                     ; col
+
+    mov         rsip, JSAMPROW [rsi]    ; inptr0
+    mov         rbxp, JSAMPROW [rbx]    ; inptr1
+    mov         rdxp, JSAMPROW [rdx]    ; inptr2
+    mov         rdip, JSAMPROW [rdi]    ; outptr
+.columnloop:
+
+    vmovdqu     ymm5, YMMWORD [rbx]     ; ymm5=Cb(0123456789ABCDEFGHIJKLMNOPQRSTUV)
+    vmovdqu     ymm1, YMMWORD [rdx]     ; ymm1=Cr(0123456789ABCDEFGHIJKLMNOPQRSTUV)
+
+    vpcmpeqw    ymm0, ymm0, ymm0
+    vpcmpeqw    ymm7, ymm7, ymm7
+    vpsrlw      ymm0, ymm0, BYTE_BIT    ; ymm0={0xFF 0x00 0xFF 0x00 ..}
+    vpsllw      ymm7, ymm7, 7           ; ymm7={0xFF80 0xFF80 0xFF80 0xFF80 ..}
+
+    vpand       ymm4, ymm0, ymm5        ; ymm4=Cb(02468ACEGIKMOQSU)=CbE
+    vpsrlw      ymm5, ymm5, BYTE_BIT    ; ymm5=Cb(13579BDFHJLNPRTV)=CbO
+    vpand       ymm0, ymm0, ymm1        ; ymm0=Cr(02468ACEGIKMOQSU)=CrE
+    vpsrlw      ymm1, ymm1, BYTE_BIT    ; ymm1=Cr(13579BDFHJLNPRTV)=CrO
+
+    vpaddw      ymm2, ymm4, ymm7
+    vpaddw      ymm3, ymm5, ymm7
+    vpaddw      ymm6, ymm0, ymm7
+    vpaddw      ymm7, ymm1, ymm7
+
+    ; (Original)
+    ; R = Y                + 1.40200 * Cr
+    ; G = Y - 0.34414 * Cb - 0.71414 * Cr
+    ; B = Y + 1.77200 * Cb
+    ;
+    ; (This implementation)
+    ; R = Y                + 0.40200 * Cr + Cr
+    ; G = Y - 0.34414 * Cb + 0.28586 * Cr - Cr
+    ; B = Y - 0.22800 * Cb + Cb + Cb
+
+    vpaddw      ymm4, ymm2, ymm2             ; ymm4=2*CbE
+    vpaddw      ymm5, ymm3, ymm3             ; ymm5=2*CbO
+    vpaddw      ymm0, ymm6, ymm6             ; ymm0=2*CrE
+    vpaddw      ymm1, ymm7, ymm7             ; ymm1=2*CrO
+
+    vpmulhw     ymm4, ymm4, [rel PW_MF0228]  ; ymm4=(2*CbE * -FIX(0.22800))
+    vpmulhw     ymm5, ymm5, [rel PW_MF0228]  ; ymm5=(2*CbO * -FIX(0.22800))
+    vpmulhw     ymm0, ymm0, [rel PW_F0402]   ; ymm0=(2*CrE * FIX(0.40200))
+    vpmulhw     ymm1, ymm1, [rel PW_F0402]   ; ymm1=(2*CrO * FIX(0.40200))
+
+    vpaddw      ymm4, ymm4, [rel PW_ONE]
+    vpaddw      ymm5, ymm5, [rel PW_ONE]
+    vpsraw      ymm4, ymm4, 1                ; ymm4=(CbE * -FIX(0.22800))
+    vpsraw      ymm5, ymm5, 1                ; ymm5=(CbO * -FIX(0.22800))
+    vpaddw      ymm0, ymm0, [rel PW_ONE]
+    vpaddw      ymm1, ymm1, [rel PW_ONE]
+    vpsraw      ymm0, ymm0, 1                ; ymm0=(CrE * FIX(0.40200))
+    vpsraw      ymm1, ymm1, 1                ; ymm1=(CrO * FIX(0.40200))
+
+    vpaddw      ymm4, ymm4, ymm2
+    vpaddw      ymm5, ymm5, ymm3
+    vpaddw      ymm4, ymm4, ymm2             ; ymm4=(CbE * FIX(1.77200))=(B-Y)E
+    vpaddw      ymm5, ymm5, ymm3             ; ymm5=(CbO * FIX(1.77200))=(B-Y)O
+    vpaddw      ymm0, ymm0, ymm6             ; ymm0=(CrE * FIX(1.40200))=(R-Y)E
+    vpaddw      ymm1, ymm1, ymm7             ; ymm1=(CrO * FIX(1.40200))=(R-Y)O
+
+    vmovdqa     YMMWORD [wk(0)], ymm4        ; wk(0)=(B-Y)E
+    vmovdqa     YMMWORD [wk(1)], ymm5        ; wk(1)=(B-Y)O
+
+    vpunpckhwd  ymm4, ymm2, ymm6
+    vpunpcklwd  ymm2, ymm2, ymm6
+    vpmaddwd    ymm2, ymm2, [rel PW_MF0344_F0285]
+    vpmaddwd    ymm4, ymm4, [rel PW_MF0344_F0285]
+    vpunpckhwd  ymm5, ymm3, ymm7
+    vpunpcklwd  ymm3, ymm3, ymm7
+    vpmaddwd    ymm3, ymm3, [rel PW_MF0344_F0285]
+    vpmaddwd    ymm5, ymm5, [rel PW_MF0344_F0285]
+
+    vpaddd      ymm2, ymm2, [rel PD_ONEHALF]
+    vpaddd      ymm4, ymm4, [rel PD_ONEHALF]
+    vpsrad      ymm2, ymm2, SCALEBITS
+    vpsrad      ymm4, ymm4, SCALEBITS
+    vpaddd      ymm3, ymm3, [rel PD_ONEHALF]
+    vpaddd      ymm5, ymm5, [rel PD_ONEHALF]
+    vpsrad      ymm3, ymm3, SCALEBITS
+    vpsrad      ymm5, ymm5, SCALEBITS
+
+    vpackssdw   ymm2, ymm2, ymm4             ; ymm2=CbE*-FIX(0.344)+CrE*FIX(0.285)
+    vpackssdw   ymm3, ymm3, ymm5             ; ymm3=CbO*-FIX(0.344)+CrO*FIX(0.285)
+    vpsubw      ymm2, ymm2, ymm6             ; ymm2=CbE*-FIX(0.344)+CrE*-FIX(0.714)=(G-Y)E
+    vpsubw      ymm3, ymm3, ymm7             ; ymm3=CbO*-FIX(0.344)+CrO*-FIX(0.714)=(G-Y)O
+
+    vmovdqu     ymm5, YMMWORD [rsi]          ; ymm5=Y(0123456789ABCDEFGHIJKLMNOPQRSTUV)
+
+    vpcmpeqw    ymm4, ymm4, ymm4
+    vpsrlw      ymm4, ymm4, BYTE_BIT         ; ymm4={0xFF 0x00 0xFF 0x00 ..}
+    vpand       ymm4, ymm4, ymm5             ; ymm4=Y(02468ACEGIKMOQSU)=YE
+    vpsrlw      ymm5, ymm5, BYTE_BIT         ; ymm5=Y(13579BDFHJLNPRTV)=YO
+
+    vpaddw      ymm0, ymm0, ymm4             ; ymm0=((R-Y)E+YE)=RE=R(02468ACEGIKMOQSU)
+    vpaddw      ymm1, ymm1, ymm5             ; ymm1=((R-Y)O+YO)=RO=R(13579BDFHJLNPRTV)
+    vpackuswb   ymm0, ymm0, ymm0             ; ymm0=R(02468ACE********GIKMOQSU********)
+    vpackuswb   ymm1, ymm1, ymm1             ; ymm1=R(13579BDF********HJLNPRTV********)
+
+    vpaddw      ymm2, ymm2, ymm4             ; ymm2=((G-Y)E+YE)=GE=G(02468ACEGIKMOQSU)
+    vpaddw      ymm3, ymm3, ymm5             ; ymm3=((G-Y)O+YO)=GO=G(13579BDFHJLNPRTV)
+    vpackuswb   ymm2, ymm2, ymm2             ; ymm2=G(02468ACE********GIKMOQSU********)
+    vpackuswb   ymm3, ymm3, ymm3             ; ymm3=G(13579BDF********HJLNPRTV********)
+
+    vpaddw      ymm4, ymm4, YMMWORD [wk(0)]  ; ymm4=(YE+(B-Y)E)=BE=B(02468ACEGIKMOQSU)
+    vpaddw      ymm5, ymm5, YMMWORD [wk(1)]  ; ymm5=(YO+(B-Y)O)=BO=B(13579BDFHJLNPRTV)
+    vpackuswb   ymm4, ymm4, ymm4             ; ymm4=B(02468ACE********GIKMOQSU********)
+    vpackuswb   ymm5, ymm5, ymm5             ; ymm5=B(13579BDF********HJLNPRTV********)
+
+%if RGB_PIXELSIZE == 3  ; ---------------
+
+    ; ymmA=(00 02 04 06 08 0A 0C 0E ** 0G 0I 0K 0M 0O 0Q 0S 0U **)
+    ; ymmB=(01 03 05 07 09 0B 0D 0F ** 0H 0J 0L 0N 0P 0R 0T 0V **)
+    ; ymmC=(10 12 14 16 18 1A 1C 1E ** 1G 1I 1K 1M 1O 1Q 1S 1U **)
+    ; ymmD=(11 13 15 17 19 1B 1D 1F ** 1H 1J 1L 1N 1P 1R 1T 1V **)
+    ; ymmE=(20 22 24 26 28 2A 2C 2E ** 2G 2I 2K 2M 2O 2Q 2S 2U **)
+    ; ymmF=(21 23 25 27 29 2B 2D 2F ** 2H 2J 2L 2N 2P 2R 2T 2V **)
+    ; ymmG=(** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** **)
+    ; ymmH=(** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** **)
+
+    vpunpcklbw  ymmA, ymmA, ymmC        ; ymmA=(00 10 02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E
+                                        ;       0G 1G 0I 1I 0K 1K 0M 1M 0O 1O 0Q 1Q 0S 1S 0U 1U)
+    vpunpcklbw  ymmE, ymmE, ymmB        ; ymmE=(20 01 22 03 24 05 26 07 28 09 2A 0B 2C 0D 2E 0F
+                                        ;       2G 0H 2I 0J 2K 0L 2M 0N 2O 0P 2Q 0R 2S 0T 2U 0V)
+    vpunpcklbw  ymmD, ymmD, ymmF        ; ymmD=(11 21 13 23 15 25 17 27 19 29 1B 2B 1D 2D 1F 2F
+                                        ;       1H 2H 1J 2J 1L 2L 1N 2N 1P 2P 1R 2R 1T 2T 1V 2V)
+
+    vpsrldq     ymmH, ymmA, 2           ; ymmH=(02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E 0G 1G
+                                        ;       0I 1I 0K 1K 0M 1M 0O 1O 0Q 1Q 0S 1S 0U 1U -- --)
+    vpunpckhwd  ymmG, ymmA, ymmE        ; ymmG=(08 18 28 09 0A 1A 2A 0B 0C 1C 2C 0D 0E 1E 2E 0F
+                                        ;       0O 1O 2O 0P 0Q 1Q 2Q 0R 0S 1S 2S 0T 0U 1U 2U 0V)
+    vpunpcklwd  ymmA, ymmA, ymmE        ; ymmA=(00 10 20 01 02 12 22 03 04 14 24 05 06 16 26 07
+                                        ;       0G 1G 2G 0H 0I 1I 2I 0J 0K 1K 2K 0L 0M 1M 2M 0N)
+
+    vpsrldq     ymmE, ymmE, 2           ; ymmE=(22 03 24 05 26 07 28 09 2A 0B 2C 0D 2E 0F 2G 0H
+                                        ;       2I 0J 2K 0L 2M 0N 2O 0P 2Q 0R 2S 0T 2U 0V -- --)
+
+    vpsrldq     ymmB, ymmD, 2           ; ymmB=(13 23 15 25 17 27 19 29 1B 2B 1D 2D 1F 2F 1H 2H
+                                        ;       1J 2J 1L 2L 1N 2N 1P 2P 1R 2R 1T 2T 1V 2V -- --)
+    vpunpckhwd  ymmC, ymmD, ymmH        ; ymmC=(19 29 0A 1A 1B 2B 0C 1C 1D 2D 0E 1E 1F 2F 0G 1G
+                                        ;       1P 2P 0Q 1Q 1R 2R 0S 1S 1T 2T 0U 1U 1V 2V -- --)
+    vpunpcklwd  ymmD, ymmD, ymmH        ; ymmD=(11 21 02 12 13 23 04 14 15 25 06 16 17 27 08 18
+                                        ;       1H 2H 0I 1I 1J 2J 0K 1K 1L 2L 0M 1M 1N 2N 0O 1O)
+
+    vpunpckhwd  ymmF, ymmE, ymmB        ; ymmF=(2A 0B 1B 2B 2C 0D 1D 2D 2E 0F 1F 2F 2G 0H 1H 2H
+                                        ;       2Q 0R 1R 2R 2S 0T 1T 2T 2U 0V 1V 2V -- -- -- --)
+    vpunpcklwd  ymmE, ymmE, ymmB        ; ymmE=(22 03 13 23 24 05 15 25 26 07 17 27 28 09 19 29
+                                        ;       2I 0J 1J 2J 2K 0L 1L 2L 2M 0N 1N 2N 2O 0P 1P 2P)
+
+    vpshufd     ymmH, ymmA, 0x4E        ; ymmH=(04 14 24 05 06 16 26 07 00 10 20 01 02 12 22 03
+                                        ;       0K 1K 2K 0L 0M 1M 2M 0N 0G 1G 2G 0H 0I 1I 2I 0J)
+    vpunpckldq  ymmA, ymmA, ymmD        ; ymmA=(00 10 20 01 11 21 02 12 02 12 22 03 13 23 04 14
+                                        ;       0G 1G 2G 0H 1H 2H 0I 1I 0I 1I 2I 0J 1J 2J 0K 1K)
+    vpunpckhdq  ymmD, ymmD, ymmE        ; ymmD=(15 25 06 16 26 07 17 27 17 27 08 18 28 09 19 29
+                                        ;       1L 2L 0M 1M 2M 0N 1N 2N 1N 2N 0O 1O 2O 0P 1P 2P)
+    vpunpckldq  ymmE, ymmE, ymmH        ; ymmE=(22 03 13 23 04 14 24 05 24 05 15 25 06 16 26 07
+                                        ;       2I 0J 1J 2J 0K 1K 2K 0L 2K 0L 1L 2L 0M 1M 2M 0N)
+
+    vpshufd     ymmH, ymmG, 0x4E        ; ymmH=(0C 1C 2C 0D 0E 1E 2E 0F 08 18 28 09 0A 1A 2A 0B
+                                        ;       0S 1S 2S 0T 0U 1U 2U 0V 0O 1O 2O 0P 0Q 1Q 2Q 0R)
+    vpunpckldq  ymmG, ymmG, ymmC        ; ymmG=(08 18 28 09 19 29 0A 1A 0A 1A 2A 0B 1B 2B 0C 1C
+                                        ;       0O 1O 2O 0P 1P 2P 0Q 1Q 0Q 1Q 2Q 0R 1R 2R 0S 1S)
+    vpunpckhdq  ymmC, ymmC, ymmF        ; ymmC=(1D 2D 0E 1E 2E 0F 1F 2F 1F 2F 0G 1G 2G 0H 1H 2H
+                                        ;       1T 2T 0U 1U 2U 0V 1V 2V 1V 2V -- -- -- -- -- --)
+    vpunpckldq  ymmF, ymmF, ymmH        ; ymmF=(2A 0B 1B 2B 0C 1C 2C 0D 2C 0D 1D 2D 0E 1E 2E 0F
+                                        ;       2Q 0R 1R 2R 0S 1S 2S 0T 2S 0T 1T 2T 0U 1U 2U 0V)
+
+    vpunpcklqdq ymmH, ymmA, ymmE        ; ymmH=(00 10 20 01 11 21 02 12 22 03 13 23 04 14 24 05
+                                        ;       0G 1G 2G 0H 1H 2H 0I 1I 2I 0J 1J 2J 0K 1K 2K 0L)
+    vpunpcklqdq ymmG, ymmD, ymmG        ; ymmG=(15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A
+                                        ;       1L 2L 0M 1M 2M 0N 1N 2N 0O 1O 2O 0P 1P 2P 0Q 1Q)
+    vpunpcklqdq ymmC, ymmF, ymmC        ; ymmC=(2A 0B 1B 2B 0C 1C 2C 0D 1D 2D 0E 1E 2E 0F 1F 2F
+                                        ;       2Q 0R 1R 2R 0S 1S 2S 0T 1T 2T 0U 1U 2U 0V 1V 2V)
+
+    vperm2i128  ymmA, ymmH, ymmG, 0x20  ; ymmA=(00 10 20 01 11 21 02 12 22 03 13 23 04 14 24 05
+                                        ;       15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A)
+    vperm2i128  ymmD, ymmC, ymmH, 0x30  ; ymmD=(2A 0B 1B 2B 0C 1C 2C 0D 1D 2D 0E 1E 2E 0F 1F 2F
+                                        ;       0G 1G 2G 0H 1H 2H 0I 1I 2I 0J 1J 2J 0K 1K 2K 0L)
+    vperm2i128  ymmF, ymmG, ymmC, 0x31  ; ymmF=(1L 2L 0M 1M 2M 0N 1N 2N 0O 1O 2O 0P 1P 2P 0Q 1Q
+                                        ;       2Q 0R 1R 2R 0S 1S 2S 0T 1T 2T 0U 1U 2U 0V 1V 2V)
+
+    cmp         rcx, byte SIZEOF_YMMWORD
+    jb          short .column_st64
+
+    test        rdi, SIZEOF_YMMWORD-1
+    jnz         short .out1
+    ; --(aligned)-------------------
+    vmovntdq    YMMWORD [rdi+0*SIZEOF_YMMWORD], ymmA
+    vmovntdq    YMMWORD [rdi+1*SIZEOF_YMMWORD], ymmD
+    vmovntdq    YMMWORD [rdi+2*SIZEOF_YMMWORD], ymmF
+    jmp         short .out0
+.out1:  ; --(unaligned)-----------------
+    vmovdqu     YMMWORD [rdi+0*SIZEOF_YMMWORD], ymmA
+    vmovdqu     YMMWORD [rdi+1*SIZEOF_YMMWORD], ymmD
+    vmovdqu     YMMWORD [rdi+2*SIZEOF_YMMWORD], ymmF
+.out0:
+    add         rdi, byte RGB_PIXELSIZE*SIZEOF_YMMWORD  ; outptr
+    sub         rcx, byte SIZEOF_YMMWORD
+    jz          near .nextrow
+
+    add         rsi, byte SIZEOF_YMMWORD  ; inptr0
+    add         rbx, byte SIZEOF_YMMWORD  ; inptr1
+    add         rdx, byte SIZEOF_YMMWORD  ; inptr2
+    jmp         near .columnloop
+
+.column_st64:
+    lea         rcx, [rcx+rcx*2]            ; imul ecx, RGB_PIXELSIZE
+    cmp         rcx, byte 2*SIZEOF_YMMWORD
+    jb          short .column_st32
+    vmovdqu     YMMWORD [rdi+0*SIZEOF_YMMWORD], ymmA
+    vmovdqu     YMMWORD [rdi+1*SIZEOF_YMMWORD], ymmD
+    add         rdi, byte 2*SIZEOF_YMMWORD  ; outptr
+    vmovdqa     ymmA, ymmF
+    sub         rcx, byte 2*SIZEOF_YMMWORD
+    jmp         short .column_st31
+.column_st32:
+    cmp         rcx, byte SIZEOF_YMMWORD
+    jb          short .column_st31
+    vmovdqu     YMMWORD [rdi+0*SIZEOF_YMMWORD], ymmA
+    add         rdi, byte SIZEOF_YMMWORD    ; outptr
+    vmovdqa     ymmA, ymmD
+    sub         rcx, byte SIZEOF_YMMWORD
+    jmp         short .column_st31
+.column_st31:
+    cmp         rcx, byte SIZEOF_XMMWORD
+    jb          short .column_st15
+    vmovdqu     XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
+    add         rdi, byte SIZEOF_XMMWORD    ; outptr
+    vperm2i128  ymmA, ymmA, ymmA, 1
+    sub         rcx, byte SIZEOF_XMMWORD
+.column_st15:
+    ; Store the lower 8 bytes of xmmA to the output when it has enough
+    ; space.
+    cmp         rcx, byte SIZEOF_MMWORD
+    jb          short .column_st7
+    vmovq       XMM_MMWORD [rdi], xmmA
+    add         rdi, byte SIZEOF_MMWORD
+    sub         rcx, byte SIZEOF_MMWORD
+    vpsrldq     xmmA, xmmA, SIZEOF_MMWORD
+.column_st7:
+    ; Store the lower 4 bytes of xmmA to the output when it has enough
+    ; space.
+    cmp         rcx, byte SIZEOF_DWORD
+    jb          short .column_st3
+    vmovd       XMM_DWORD [rdi], xmmA
+    add         rdi, byte SIZEOF_DWORD
+    sub         rcx, byte SIZEOF_DWORD
+    vpsrldq     xmmA, xmmA, SIZEOF_DWORD
+.column_st3:
+    ; Store the lower 2 bytes of rax to the output when it has enough
+    ; space.
+    vmovd       eax, xmmA
+    cmp         rcx, byte SIZEOF_WORD
+    jb          short .column_st1
+    mov         word [rdi], ax
+    add         rdi, byte SIZEOF_WORD
+    sub         rcx, byte SIZEOF_WORD
+    shr         rax, 16
+.column_st1:
+    ; Store the lower 1 byte of rax to the output when it has enough
+    ; space.
+    test        rcx, rcx
+    jz          short .nextrow
+    mov         byte [rdi], al
+
+%else  ; RGB_PIXELSIZE == 4 ; -----------
+
+%ifdef RGBX_FILLER_0XFF
+    vpcmpeqb    ymm6, ymm6, ymm6        ; ymm6=XE=X(02468ACE********GIKMOQSU********)
+    vpcmpeqb    ymm7, ymm7, ymm7        ; ymm7=XO=X(13579BDF********HJLNPRTV********)
+%else
+    vpxor       ymm6, ymm6, ymm6        ; ymm6=XE=X(02468ACE********GIKMOQSU********)
+    vpxor       ymm7, ymm7, ymm7        ; ymm7=XO=X(13579BDF********HJLNPRTV********)
+%endif
+    ; ymmA=(00 02 04 06 08 0A 0C 0E ** 0G 0I 0K 0M 0O 0Q 0S 0U **)
+    ; ymmB=(01 03 05 07 09 0B 0D 0F ** 0H 0J 0L 0N 0P 0R 0T 0V **)
+    ; ymmC=(10 12 14 16 18 1A 1C 1E ** 1G 1I 1K 1M 1O 1Q 1S 1U **)
+    ; ymmD=(11 13 15 17 19 1B 1D 1F ** 1H 1J 1L 1N 1P 1R 1T 1V **)
+    ; ymmE=(20 22 24 26 28 2A 2C 2E ** 2G 2I 2K 2M 2O 2Q 2S 2U **)
+    ; ymmF=(21 23 25 27 29 2B 2D 2F ** 2H 2J 2L 2N 2P 2R 2T 2V **)
+    ; ymmG=(30 32 34 36 38 3A 3C 3E ** 3G 3I 3K 3M 3O 3Q 3S 3U **)
+    ; ymmH=(31 33 35 37 39 3B 3D 3F ** 3H 3J 3L 3N 3P 3R 3T 3V **)
+
+    vpunpcklbw  ymmA, ymmA, ymmC        ; ymmA=(00 10 02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E
+                                        ;       0G 1G 0I 1I 0K 1K 0M 1M 0O 1O 0Q 1Q 0S 1S 0U 1U)
+    vpunpcklbw  ymmE, ymmE, ymmG        ; ymmE=(20 30 22 32 24 34 26 36 28 38 2A 3A 2C 3C 2E 3E
+                                        ;       2G 3G 2I 3I 2K 3K 2M 3M 2O 3O 2Q 3Q 2S 3S 2U 3U)
+    vpunpcklbw  ymmB, ymmB, ymmD        ; ymmB=(01 11 03 13 05 15 07 17 09 19 0B 1B 0D 1D 0F 1F
+                                        ;       0H 1H 0J 1J 0L 1L 0N 1N 0P 1P 0R 1R 0T 1T 0V 1V)
+    vpunpcklbw  ymmF, ymmF, ymmH        ; ymmF=(21 31 23 33 25 35 27 37 29 39 2B 3B 2D 3D 2F 3F
+                                        ;       2H 3H 2J 3J 2L 3L 2N 3N 2P 3P 2R 3R 2T 3T 2V 3V)
+
+    vpunpckhwd  ymmC, ymmA, ymmE        ; ymmC=(08 18 28 38 0A 1A 2A 3A 0C 1C 2C 3C 0E 1E 2E 3E
+                                        ;       0O 1O 2O 3O 0Q 1Q 2Q 3Q 0S 1S 2S 3S 0U 1U 2U 3U)
+    vpunpcklwd  ymmA, ymmA, ymmE        ; ymmA=(00 10 20 30 02 12 22 32 04 14 24 34 06 16 26 36
+                                        ;       0G 1G 2G 3G 0I 1I 2I 3I 0K 1K 2K 3K 0M 1M 2M 3M)
+    vpunpckhwd  ymmG, ymmB, ymmF        ; ymmG=(09 19 29 39 0B 1B 2B 3B 0D 1D 2D 3D 0F 1F 2F 3F
+                                        ;       0P 1P 2P 3P 0R 1R 2R 3R 0T 1T 2T 3T 0V 1V 2V 3V)
+    vpunpcklwd  ymmB, ymmB, ymmF        ; ymmB=(01 11 21 31 03 13 23 33 05 15 25 35 07 17 27 37
+                                        ;       0H 1H 2H 3H 0J 1J 2J 3J 0L 1L 2L 3L 0N 1N 2N 3N)
+
+    vpunpckhdq  ymmE, ymmA, ymmB        ; ymmE=(04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37
+                                        ;       0K 1K 2K 3K 0L 1L 2L 3L 0M 1M 2M 3M 0N 1N 2N 3N)
+    vpunpckldq  ymmB, ymmA, ymmB        ; ymmB=(00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33
+                                        ;       0G 1G 2G 3G 0H 1H 2H 3H 0I 1I 2I 3I 0J 1J 2J 3J)
+    vpunpckhdq  ymmF, ymmC, ymmG        ; ymmF=(0C 1C 2C 3C 0D 1D 2D 3D 0E 1E 2E 3E 0F 1F 2F 3F
+                                        ;       0S 1S 2S 3S 0T 1T 2T 3T 0U 1U 2U 3U 0V 1V 2V 3V)
+    vpunpckldq  ymmG, ymmC, ymmG        ; ymmG=(08 18 28 38 09 19 29 39 0A 1A 2A 3A 0B 1B 2B 3B
+                                        ;       0O 1O 2O 3O 0P 1P 2P 3P 0Q 1Q 2Q 3Q 0R 1R 2R 3R)
+
+    vperm2i128  ymmA, ymmB, ymmE, 0x20  ; ymmA=(00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33
+                                        ;       04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37)
+    vperm2i128  ymmD, ymmG, ymmF, 0x20  ; ymmD=(08 18 28 38 09 19 29 39 0A 1A 2A 3A 0B 1B 2B 3B
+                                        ;       0C 1C 2C 3C 0D 1D 2D 3D 0E 1E 2E 3E 0F 1F 2F 3F)
+    vperm2i128  ymmC, ymmB, ymmE, 0x31  ; ymmC=(0G 1G 2G 3G 0H 1H 2H 3H 0I 1I 2I 3I 0J 1J 2J 3J
+                                        ;       0K 1K 2K 3K 0L 1L 2L 3L 0M 1M 2M 3M 0N 1N 2N 3N)
+    vperm2i128  ymmH, ymmG, ymmF, 0x31  ; ymmH=(0O 1O 2O 3O 0P 1P 2P 3P 0Q 1Q 2Q 3Q 0R 1R 2R 3R
+                                        ;       0S 1S 2S 3S 0T 1T 2T 3T 0U 1U 2U 3U 0V 1V 2V 3V)
+
+    cmp         rcx, byte SIZEOF_YMMWORD
+    jb          short .column_st64
+
+    test        rdi, SIZEOF_YMMWORD-1
+    jnz         short .out1
+    ; --(aligned)-------------------
+    vmovntdq    YMMWORD [rdi+0*SIZEOF_YMMWORD], ymmA
+    vmovntdq    YMMWORD [rdi+1*SIZEOF_YMMWORD], ymmD
+    vmovntdq    YMMWORD [rdi+2*SIZEOF_YMMWORD], ymmC
+    vmovntdq    YMMWORD [rdi+3*SIZEOF_YMMWORD], ymmH
+    jmp         short .out0
+.out1:  ; --(unaligned)-----------------
+    vmovdqu     YMMWORD [rdi+0*SIZEOF_YMMWORD], ymmA
+    vmovdqu     YMMWORD [rdi+1*SIZEOF_YMMWORD], ymmD
+    vmovdqu     YMMWORD [rdi+2*SIZEOF_YMMWORD], ymmC
+    vmovdqu     YMMWORD [rdi+3*SIZEOF_YMMWORD], ymmH
+.out0:
+    add         rdi, RGB_PIXELSIZE*SIZEOF_YMMWORD  ; outptr
+    sub         rcx, byte SIZEOF_YMMWORD
+    jz          near .nextrow
+
+    add         rsi, byte SIZEOF_YMMWORD  ; inptr0
+    add         rbx, byte SIZEOF_YMMWORD  ; inptr1
+    add         rdx, byte SIZEOF_YMMWORD  ; inptr2
+    jmp         near .columnloop
+
+.column_st64:
+    cmp         rcx, byte SIZEOF_YMMWORD/2
+    jb          short .column_st32
+    vmovdqu     YMMWORD [rdi+0*SIZEOF_YMMWORD], ymmA
+    vmovdqu     YMMWORD [rdi+1*SIZEOF_YMMWORD], ymmD
+    add         rdi, byte 2*SIZEOF_YMMWORD  ; outptr
+    vmovdqa     ymmA, ymmC
+    vmovdqa     ymmD, ymmH
+    sub         rcx, byte SIZEOF_YMMWORD/2
+.column_st32:
+    cmp         rcx, byte SIZEOF_YMMWORD/4
+    jb          short .column_st16
+    vmovdqu     YMMWORD [rdi+0*SIZEOF_YMMWORD], ymmA
+    add         rdi, byte SIZEOF_YMMWORD    ; outptr
+    vmovdqa     ymmA, ymmD
+    sub         rcx, byte SIZEOF_YMMWORD/4
+.column_st16:
+    cmp         rcx, byte SIZEOF_YMMWORD/8
+    jb          short .column_st15
+    vmovdqu     XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
+    vperm2i128  ymmA, ymmA, ymmA, 1
+    add         rdi, byte SIZEOF_XMMWORD    ; outptr
+    sub         rcx, byte SIZEOF_YMMWORD/8
+.column_st15:
+    ; Store two pixels (8 bytes) of ymmA to the output when it has enough
+    ; space.
+    cmp         rcx, byte SIZEOF_YMMWORD/16
+    jb          short .column_st7
+    vmovq       MMWORD [rdi], xmmA
+    add         rdi, byte SIZEOF_YMMWORD/16*4
+    sub         rcx, byte SIZEOF_YMMWORD/16
+    vpsrldq     xmmA, SIZEOF_YMMWORD/16*4
+.column_st7:
+    ; Store one pixel (4 bytes) of ymmA to the output when it has enough
+    ; space.
+    test        rcx, rcx
+    jz          short .nextrow
+    vmovd       XMM_DWORD [rdi], xmmA
+
+%endif  ; RGB_PIXELSIZE ; ---------------
+
+.nextrow:
+    pop         rcx
+    pop         rsi
+    pop         rbx
+    pop         rdx
+    pop         rdi
+    pop         rax
+
+    add         rsi, byte SIZEOF_JSAMPROW
+    add         rbx, byte SIZEOF_JSAMPROW
+    add         rdx, byte SIZEOF_JSAMPROW
+    add         rdi, byte SIZEOF_JSAMPROW  ; output_buf
+    dec         rax                        ; num_rows
+    jg          near .rowloop
+
+    sfence                              ; flush the write buffer
+
+.return:
+    pop         rbx
+    vzeroupper
+    uncollect_args 5
+    mov         rsp, rbp                ; rsp <- aligned rbp
+    pop         rsp                     ; rsp <- original rbp
+    pop         rbp
+    ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+    align       32
diff --git a/media/libjpeg/simd/x86_64/jdcolext-sse2.asm b/media/libjpeg/simd/x86_64/jdcolext-sse2.asm
new file mode 100644
index 0000000000..e07c8d7518
--- /dev/null
+++ b/media/libjpeg/simd/x86_64/jdcolext-sse2.asm
@@ -0,0 +1,439 @@
+;
+; jdcolext.asm - colorspace conversion (64-bit SSE2)
+;
+; Copyright 2009, 2012 Pierre Ossman <ossman@cendio.se> for Cendio AB
+; Copyright (C) 2009, 2012, 2016, D. R. Commander.
+; Copyright (C) 2018, Matthias Räncker.
+;
+; Based on the x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+
+%include "jcolsamp.inc"
+
+; --------------------------------------------------------------------------
+;
+; Convert some rows of samples to the output colorspace.
+;
+; GLOBAL(void)
+; jsimd_ycc_rgb_convert_sse2(JDIMENSION out_width, JSAMPIMAGE input_buf,
+;                            JDIMENSION input_row, JSAMPARRAY output_buf,
+;                            int num_rows)
+;
+
+; r10d = JDIMENSION out_width
+; r11 = JSAMPIMAGE input_buf
+; r12d = JDIMENSION input_row
+; r13 = JSAMPARRAY output_buf
+; r14d = int num_rows
+
+%define wk(i)   rbp - (WK_NUM - (i)) * SIZEOF_XMMWORD  ; xmmword wk[WK_NUM]
+%define WK_NUM  2
+
+    align       32
+    GLOBAL_FUNCTION(jsimd_ycc_rgb_convert_sse2)
+
+EXTN(jsimd_ycc_rgb_convert_sse2):
+    push        rbp
+    mov         rax, rsp                     ; rax = original rbp
+    sub         rsp, byte 4
+    and         rsp, byte (-SIZEOF_XMMWORD)  ; align to 128 bits
+    mov         [rsp], rax
+    mov         rbp, rsp                     ; rbp = aligned rbp
+    lea         rsp, [wk(0)]
+    collect_args 5
+    push        rbx
+
+    mov         ecx, r10d               ; num_cols
+    test        rcx, rcx
+    jz          near .return
+
+    push        rcx
+
+    mov         rdi, r11
+    mov         ecx, r12d
+    mov         rsip, JSAMPARRAY [rdi+0*SIZEOF_JSAMPARRAY]
+    mov         rbxp, JSAMPARRAY [rdi+1*SIZEOF_JSAMPARRAY]
+    mov         rdxp, JSAMPARRAY [rdi+2*SIZEOF_JSAMPARRAY]
+    lea         rsi, [rsi+rcx*SIZEOF_JSAMPROW]
+    lea         rbx, [rbx+rcx*SIZEOF_JSAMPROW]
+    lea         rdx, [rdx+rcx*SIZEOF_JSAMPROW]
+
+    pop         rcx
+
+    mov         rdi, r13
+    mov         eax, r14d
+    test        rax, rax
+    jle         near .return
+.rowloop:
+    push        rax
+    push        rdi
+    push        rdx
+    push        rbx
+    push        rsi
+    push        rcx                     ; col
+
+    mov         rsip, JSAMPROW [rsi]    ; inptr0
+    mov         rbxp, JSAMPROW [rbx]    ; inptr1
+    mov         rdxp, JSAMPROW [rdx]    ; inptr2
+    mov         rdip, JSAMPROW [rdi]    ; outptr
+.columnloop:
+
+    movdqa      xmm5, XMMWORD [rbx]     ; xmm5=Cb(0123456789ABCDEF)
+    movdqa      xmm1, XMMWORD [rdx]     ; xmm1=Cr(0123456789ABCDEF)
+
+    pcmpeqw     xmm4, xmm4
+    pcmpeqw     xmm7, xmm7
+    psrlw       xmm4, BYTE_BIT
+    psllw       xmm7, 7                 ; xmm7={0xFF80 0xFF80 0xFF80 0xFF80 ..}
+    movdqa      xmm0, xmm4              ; xmm0=xmm4={0xFF 0x00 0xFF 0x00 ..}
+
+    pand        xmm4, xmm5              ; xmm4=Cb(02468ACE)=CbE
+    psrlw       xmm5, BYTE_BIT          ; xmm5=Cb(13579BDF)=CbO
+    pand        xmm0, xmm1              ; xmm0=Cr(02468ACE)=CrE
+    psrlw       xmm1, BYTE_BIT          ; xmm1=Cr(13579BDF)=CrO
+
+    paddw       xmm4, xmm7
+    paddw       xmm5, xmm7
+    paddw       xmm0, xmm7
+    paddw       xmm1, xmm7
+
+    ; (Original)
+    ; R = Y                + 1.40200 * Cr
+    ; G = Y - 0.34414 * Cb - 0.71414 * Cr
+    ; B = Y + 1.77200 * Cb
+    ;
+    ; (This implementation)
+    ; R = Y                + 0.40200 * Cr + Cr
+    ; G = Y - 0.34414 * Cb + 0.28586 * Cr - Cr
+    ; B = Y - 0.22800 * Cb + Cb + Cb
+
+    movdqa      xmm2, xmm4              ; xmm2=CbE
+    movdqa      xmm3, xmm5              ; xmm3=CbO
+    paddw       xmm4, xmm4              ; xmm4=2*CbE
+    paddw       xmm5, xmm5              ; xmm5=2*CbO
+    movdqa      xmm6, xmm0              ; xmm6=CrE
+    movdqa      xmm7, xmm1              ; xmm7=CrO
+    paddw       xmm0, xmm0              ; xmm0=2*CrE
+    paddw       xmm1, xmm1              ; xmm1=2*CrO
+
+    pmulhw      xmm4, [rel PW_MF0228]   ; xmm4=(2*CbE * -FIX(0.22800))
+    pmulhw      xmm5, [rel PW_MF0228]   ; xmm5=(2*CbO * -FIX(0.22800))
+    pmulhw      xmm0, [rel PW_F0402]    ; xmm0=(2*CrE * FIX(0.40200))
+    pmulhw      xmm1, [rel PW_F0402]    ; xmm1=(2*CrO * FIX(0.40200))
+
+    paddw       xmm4, [rel PW_ONE]
+    paddw       xmm5, [rel PW_ONE]
+    psraw       xmm4, 1                 ; xmm4=(CbE * -FIX(0.22800))
+    psraw       xmm5, 1                 ; xmm5=(CbO * -FIX(0.22800))
+    paddw       xmm0, [rel PW_ONE]
+    paddw       xmm1, [rel PW_ONE]
+    psraw       xmm0, 1                 ; xmm0=(CrE * FIX(0.40200))
+    psraw       xmm1, 1                 ; xmm1=(CrO * FIX(0.40200))
+
+    paddw       xmm4, xmm2
+    paddw       xmm5, xmm3
+    paddw       xmm4, xmm2              ; xmm4=(CbE * FIX(1.77200))=(B-Y)E
+    paddw       xmm5, xmm3              ; xmm5=(CbO * FIX(1.77200))=(B-Y)O
+    paddw       xmm0, xmm6              ; xmm0=(CrE * FIX(1.40200))=(R-Y)E
+    paddw       xmm1, xmm7              ; xmm1=(CrO * FIX(1.40200))=(R-Y)O
+
+    movdqa      XMMWORD [wk(0)], xmm4   ; wk(0)=(B-Y)E
+    movdqa      XMMWORD [wk(1)], xmm5   ; wk(1)=(B-Y)O
+
+    movdqa      xmm4, xmm2
+    movdqa      xmm5, xmm3
+    punpcklwd   xmm2, xmm6
+    punpckhwd   xmm4, xmm6
+    pmaddwd     xmm2, [rel PW_MF0344_F0285]
+    pmaddwd     xmm4, [rel PW_MF0344_F0285]
+    punpcklwd   xmm3, xmm7
+    punpckhwd   xmm5, xmm7
+    pmaddwd     xmm3, [rel PW_MF0344_F0285]
+    pmaddwd     xmm5, [rel PW_MF0344_F0285]
+
+    paddd       xmm2, [rel PD_ONEHALF]
+    paddd       xmm4, [rel PD_ONEHALF]
+    psrad       xmm2, SCALEBITS
+    psrad       xmm4, SCALEBITS
+    paddd       xmm3, [rel PD_ONEHALF]
+    paddd       xmm5, [rel PD_ONEHALF]
+    psrad       xmm3, SCALEBITS
+    psrad       xmm5, SCALEBITS
+
+    packssdw    xmm2, xmm4              ; xmm2=CbE*-FIX(0.344)+CrE*FIX(0.285)
+    packssdw    xmm3, xmm5              ; xmm3=CbO*-FIX(0.344)+CrO*FIX(0.285)
+    psubw       xmm2, xmm6              ; xmm2=CbE*-FIX(0.344)+CrE*-FIX(0.714)=(G-Y)E
+    psubw       xmm3, xmm7              ; xmm3=CbO*-FIX(0.344)+CrO*-FIX(0.714)=(G-Y)O
+
+    movdqa      xmm5, XMMWORD [rsi]     ; xmm5=Y(0123456789ABCDEF)
+
+    pcmpeqw     xmm4, xmm4
+    psrlw       xmm4, BYTE_BIT          ; xmm4={0xFF 0x00 0xFF 0x00 ..}
+    pand        xmm4, xmm5              ; xmm4=Y(02468ACE)=YE
+    psrlw       xmm5, BYTE_BIT          ; xmm5=Y(13579BDF)=YO
+
+    paddw       xmm0, xmm4              ; xmm0=((R-Y)E+YE)=RE=R(02468ACE)
+    paddw       xmm1, xmm5              ; xmm1=((R-Y)O+YO)=RO=R(13579BDF)
+    packuswb    xmm0, xmm0              ; xmm0=R(02468ACE********)
+    packuswb    xmm1, xmm1              ; xmm1=R(13579BDF********)
+
+    paddw       xmm2, xmm4              ; xmm2=((G-Y)E+YE)=GE=G(02468ACE)
+    paddw       xmm3, xmm5              ; xmm3=((G-Y)O+YO)=GO=G(13579BDF)
+    packuswb    xmm2, xmm2              ; xmm2=G(02468ACE********)
+    packuswb    xmm3, xmm3              ; xmm3=G(13579BDF********)
+
+    paddw       xmm4, XMMWORD [wk(0)]   ; xmm4=(YE+(B-Y)E)=BE=B(02468ACE)
+    paddw       xmm5, XMMWORD [wk(1)]   ; xmm5=(YO+(B-Y)O)=BO=B(13579BDF)
+    packuswb    xmm4, xmm4              ; xmm4=B(02468ACE********)
+    packuswb    xmm5, xmm5              ; xmm5=B(13579BDF********)
+
+%if RGB_PIXELSIZE == 3  ; ---------------
+
+    ; xmmA=(00 02 04 06 08 0A 0C 0E **), xmmB=(01 03 05 07 09 0B 0D 0F **)
+    ; xmmC=(10 12 14 16 18 1A 1C 1E **), xmmD=(11 13 15 17 19 1B 1D 1F **)
+    ; xmmE=(20 22 24 26 28 2A 2C 2E **), xmmF=(21 23 25 27 29 2B 2D 2F **)
+    ; xmmG=(** ** ** ** ** ** ** ** **), xmmH=(** ** ** ** ** ** ** ** **)
+
+    punpcklbw   xmmA, xmmC        ; xmmA=(00 10 02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E)
+    punpcklbw   xmmE, xmmB        ; xmmE=(20 01 22 03 24 05 26 07 28 09 2A 0B 2C 0D 2E 0F)
+    punpcklbw   xmmD, xmmF        ; xmmD=(11 21 13 23 15 25 17 27 19 29 1B 2B 1D 2D 1F 2F)
+
+    movdqa      xmmG, xmmA
+    movdqa      xmmH, xmmA
+    punpcklwd   xmmA, xmmE        ; xmmA=(00 10 20 01 02 12 22 03 04 14 24 05 06 16 26 07)
+    punpckhwd   xmmG, xmmE        ; xmmG=(08 18 28 09 0A 1A 2A 0B 0C 1C 2C 0D 0E 1E 2E 0F)
+
+    psrldq      xmmH, 2           ; xmmH=(02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E -- --)
+    psrldq      xmmE, 2           ; xmmE=(22 03 24 05 26 07 28 09 2A 0B 2C 0D 2E 0F -- --)
+
+    movdqa      xmmC, xmmD
+    movdqa      xmmB, xmmD
+    punpcklwd   xmmD, xmmH        ; xmmD=(11 21 02 12 13 23 04 14 15 25 06 16 17 27 08 18)
+    punpckhwd   xmmC, xmmH        ; xmmC=(19 29 0A 1A 1B 2B 0C 1C 1D 2D 0E 1E 1F 2F -- --)
+
+    psrldq      xmmB, 2           ; xmmB=(13 23 15 25 17 27 19 29 1B 2B 1D 2D 1F 2F -- --)
+
+    movdqa      xmmF, xmmE
+    punpcklwd   xmmE, xmmB        ; xmmE=(22 03 13 23 24 05 15 25 26 07 17 27 28 09 19 29)
+    punpckhwd   xmmF, xmmB        ; xmmF=(2A 0B 1B 2B 2C 0D 1D 2D 2E 0F 1F 2F -- -- -- --)
+
+    pshufd      xmmH, xmmA, 0x4E  ; xmmH=(04 14 24 05 06 16 26 07 00 10 20 01 02 12 22 03)
+    movdqa      xmmB, xmmE
+    punpckldq   xmmA, xmmD        ; xmmA=(00 10 20 01 11 21 02 12 02 12 22 03 13 23 04 14)
+    punpckldq   xmmE, xmmH        ; xmmE=(22 03 13 23 04 14 24 05 24 05 15 25 06 16 26 07)
+    punpckhdq   xmmD, xmmB        ; xmmD=(15 25 06 16 26 07 17 27 17 27 08 18 28 09 19 29)
+
+    pshufd      xmmH, xmmG, 0x4E  ; xmmH=(0C 1C 2C 0D 0E 1E 2E 0F 08 18 28 09 0A 1A 2A 0B)
+    movdqa      xmmB, xmmF
+    punpckldq   xmmG, xmmC        ; xmmG=(08 18 28 09 19 29 0A 1A 0A 1A 2A 0B 1B 2B 0C 1C)
+    punpckldq   xmmF, xmmH        ; xmmF=(2A 0B 1B 2B 0C 1C 2C 0D 2C 0D 1D 2D 0E 1E 2E 0F)
+    punpckhdq   xmmC, xmmB        ; xmmC=(1D 2D 0E 1E 2E 0F 1F 2F 1F 2F -- -- -- -- -- --)
+
+    punpcklqdq  xmmA, xmmE        ; xmmA=(00 10 20 01 11 21 02 12 22 03 13 23 04 14 24 05)
+    punpcklqdq  xmmD, xmmG        ; xmmD=(15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A)
+    punpcklqdq  xmmF, xmmC        ; xmmF=(2A 0B 1B 2B 0C 1C 2C 0D 1D 2D 0E 1E 2E 0F 1F 2F)
+
+    cmp         rcx, byte SIZEOF_XMMWORD
+    jb          short .column_st32
+
+    test        rdi, SIZEOF_XMMWORD-1
+    jnz         short .out1
+    ; --(aligned)-------------------
+    movntdq     XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
+    movntdq     XMMWORD [rdi+1*SIZEOF_XMMWORD], xmmD
+    movntdq     XMMWORD [rdi+2*SIZEOF_XMMWORD], xmmF
+    jmp         short .out0
+.out1:  ; --(unaligned)-----------------
+    movdqu      XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
+    movdqu      XMMWORD [rdi+1*SIZEOF_XMMWORD], xmmD
+    movdqu      XMMWORD [rdi+2*SIZEOF_XMMWORD], xmmF
+.out0:
+    add         rdi, byte RGB_PIXELSIZE*SIZEOF_XMMWORD  ; outptr
+    sub         rcx, byte SIZEOF_XMMWORD
+    jz          near .nextrow
+
+    add         rsi, byte SIZEOF_XMMWORD  ; inptr0
+    add         rbx, byte SIZEOF_XMMWORD  ; inptr1
+    add         rdx, byte SIZEOF_XMMWORD  ; inptr2
+    jmp         near .columnloop
+
+.column_st32:
+    lea         rcx, [rcx+rcx*2]            ; imul ecx, RGB_PIXELSIZE
+    cmp         rcx, byte 2*SIZEOF_XMMWORD
+    jb          short .column_st16
+    movdqu      XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
+    movdqu      XMMWORD [rdi+1*SIZEOF_XMMWORD], xmmD
+    add         rdi, byte 2*SIZEOF_XMMWORD  ; outptr
+    movdqa      xmmA, xmmF
+    sub         rcx, byte 2*SIZEOF_XMMWORD
+    jmp         short .column_st15
+.column_st16:
+    cmp         rcx, byte SIZEOF_XMMWORD
+    jb          short .column_st15
+    movdqu      XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
+    add         rdi, byte SIZEOF_XMMWORD    ; outptr
+    movdqa      xmmA, xmmD
+    sub         rcx, byte SIZEOF_XMMWORD
+.column_st15:
+    ; Store the lower 8 bytes of xmmA to the output when it has enough
+    ; space.
+    cmp         rcx, byte SIZEOF_MMWORD
+    jb          short .column_st7
+    movq        XMM_MMWORD [rdi], xmmA
+    add         rdi, byte SIZEOF_MMWORD
+    sub         rcx, byte SIZEOF_MMWORD
+    psrldq      xmmA, SIZEOF_MMWORD
+.column_st7:
+    ; Store the lower 4 bytes of xmmA to the output when it has enough
+    ; space.
+    cmp         rcx, byte SIZEOF_DWORD
+    jb          short .column_st3
+    movd        XMM_DWORD [rdi], xmmA
+    add         rdi, byte SIZEOF_DWORD
+    sub         rcx, byte SIZEOF_DWORD
+    psrldq      xmmA, SIZEOF_DWORD
+.column_st3:
+    ; Store the lower 2 bytes of rax to the output when it has enough
+    ; space.
+    movd        eax, xmmA
+    cmp         rcx, byte SIZEOF_WORD
+    jb          short .column_st1
+    mov         word [rdi], ax
+    add         rdi, byte SIZEOF_WORD
+    sub         rcx, byte SIZEOF_WORD
+    shr         rax, 16
+.column_st1:
+    ; Store the lower 1 byte of rax to the output when it has enough
+    ; space.
+    test        rcx, rcx
+    jz          short .nextrow
+    mov         byte [rdi], al
+
+%else  ; RGB_PIXELSIZE == 4 ; -----------
+
+%ifdef RGBX_FILLER_0XFF
+    pcmpeqb     xmm6, xmm6              ; xmm6=XE=X(02468ACE********)
+    pcmpeqb     xmm7, xmm7              ; xmm7=XO=X(13579BDF********)
+%else
+    pxor        xmm6, xmm6              ; xmm6=XE=X(02468ACE********)
+    pxor        xmm7, xmm7              ; xmm7=XO=X(13579BDF********)
+%endif
+    ; xmmA=(00 02 04 06 08 0A 0C 0E **), xmmB=(01 03 05 07 09 0B 0D 0F **)
+    ; xmmC=(10 12 14 16 18 1A 1C 1E **), xmmD=(11 13 15 17 19 1B 1D 1F **)
+    ; xmmE=(20 22 24 26 28 2A 2C 2E **), xmmF=(21 23 25 27 29 2B 2D 2F **)
+    ; xmmG=(30 32 34 36 38 3A 3C 3E **), xmmH=(31 33 35 37 39 3B 3D 3F **)
+
+    punpcklbw   xmmA, xmmC  ; xmmA=(00 10 02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E)
+    punpcklbw   xmmE, xmmG  ; xmmE=(20 30 22 32 24 34 26 36 28 38 2A 3A 2C 3C 2E 3E)
+    punpcklbw   xmmB, xmmD  ; xmmB=(01 11 03 13 05 15 07 17 09 19 0B 1B 0D 1D 0F 1F)
+    punpcklbw   xmmF, xmmH  ; xmmF=(21 31 23 33 25 35 27 37 29 39 2B 3B 2D 3D 2F 3F)
+
+    movdqa      xmmC, xmmA
+    punpcklwd   xmmA, xmmE  ; xmmA=(00 10 20 30 02 12 22 32 04 14 24 34 06 16 26 36)
+    punpckhwd   xmmC, xmmE  ; xmmC=(08 18 28 38 0A 1A 2A 3A 0C 1C 2C 3C 0E 1E 2E 3E)
+    movdqa      xmmG, xmmB
+    punpcklwd   xmmB, xmmF  ; xmmB=(01 11 21 31 03 13 23 33 05 15 25 35 07 17 27 37)
+    punpckhwd   xmmG, xmmF  ; xmmG=(09 19 29 39 0B 1B 2B 3B 0D 1D 2D 3D 0F 1F 2F 3F)
+
+    movdqa      xmmD, xmmA
+    punpckldq   xmmA, xmmB  ; xmmA=(00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33)
+    punpckhdq   xmmD, xmmB  ; xmmD=(04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37)
+    movdqa      xmmH, xmmC
+    punpckldq   xmmC, xmmG  ; xmmC=(08 18 28 38 09 19 29 39 0A 1A 2A 3A 0B 1B 2B 3B)
+    punpckhdq   xmmH, xmmG  ; xmmH=(0C 1C 2C 3C 0D 1D 2D 3D 0E 1E 2E 3E 0F 1F 2F 3F)
+
+    cmp         rcx, byte SIZEOF_XMMWORD
+    jb          short .column_st32
+
+    test        rdi, SIZEOF_XMMWORD-1
+    jnz         short .out1
+    ; --(aligned)-------------------
+    movntdq     XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
+    movntdq     XMMWORD [rdi+1*SIZEOF_XMMWORD], xmmD
+    movntdq     XMMWORD [rdi+2*SIZEOF_XMMWORD], xmmC
+    movntdq     XMMWORD [rdi+3*SIZEOF_XMMWORD], xmmH
+    jmp         short .out0
+.out1:  ; --(unaligned)-----------------
+    movdqu      XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
+    movdqu      XMMWORD [rdi+1*SIZEOF_XMMWORD], xmmD
+    movdqu      XMMWORD [rdi+2*SIZEOF_XMMWORD], xmmC
+    movdqu      XMMWORD [rdi+3*SIZEOF_XMMWORD], xmmH
+.out0:
+    add         rdi, byte RGB_PIXELSIZE*SIZEOF_XMMWORD  ; outptr
+    sub         rcx, byte SIZEOF_XMMWORD
+    jz          near .nextrow
+
+    add         rsi, byte SIZEOF_XMMWORD  ; inptr0
+    add         rbx, byte SIZEOF_XMMWORD  ; inptr1
+    add         rdx, byte SIZEOF_XMMWORD  ; inptr2
+    jmp         near .columnloop
+
+.column_st32:
+    cmp         rcx, byte SIZEOF_XMMWORD/2
+    jb          short .column_st16
+    movdqu      XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
+    movdqu      XMMWORD [rdi+1*SIZEOF_XMMWORD], xmmD
+    add         rdi, byte 2*SIZEOF_XMMWORD  ; outptr
+    movdqa      xmmA, xmmC
+    movdqa      xmmD, xmmH
+    sub         rcx, byte SIZEOF_XMMWORD/2
+.column_st16:
+    cmp         rcx, byte SIZEOF_XMMWORD/4
+    jb          short .column_st15
+    movdqu      XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
+    add         rdi, byte SIZEOF_XMMWORD    ; outptr
+    movdqa      xmmA, xmmD
+    sub         rcx, byte SIZEOF_XMMWORD/4
+.column_st15:
+    ; Store two pixels (8 bytes) of xmmA to the output when it has enough
+    ; space.
+    cmp         rcx, byte SIZEOF_XMMWORD/8
+    jb          short .column_st7
+    movq        MMWORD [rdi], xmmA
+    add         rdi, byte SIZEOF_XMMWORD/8*4
+    sub         rcx, byte SIZEOF_XMMWORD/8
+    psrldq      xmmA, SIZEOF_XMMWORD/8*4
+.column_st7:
+    ; Store one pixel (4 bytes) of xmmA to the output when it has enough
+    ; space.
+    test        rcx, rcx
+    jz          short .nextrow
+    movd        XMM_DWORD [rdi], xmmA
+
+%endif  ; RGB_PIXELSIZE ; ---------------
+
+.nextrow:
+    pop         rcx
+    pop         rsi
+    pop         rbx
+    pop         rdx
+    pop         rdi
+    pop         rax
+
+    add         rsi, byte SIZEOF_JSAMPROW
+    add         rbx, byte SIZEOF_JSAMPROW
+    add         rdx, byte SIZEOF_JSAMPROW
+    add         rdi, byte SIZEOF_JSAMPROW  ; output_buf
+    dec         rax                        ; num_rows
+    jg          near .rowloop
+
+    sfence                              ; flush the write buffer
+
+.return:
+    pop         rbx
+    uncollect_args 5
+    mov         rsp, rbp                ; rsp <- aligned rbp
+    pop         rsp                     ; rsp <- original rbp
+    pop         rbp
+    ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+    align       32
diff --git a/media/libjpeg/simd/x86_64/jdcolor-avx2.asm b/media/libjpeg/simd/x86_64/jdcolor-avx2.asm
new file mode 100644
index 0000000000..43de9db04d
--- /dev/null
+++ b/media/libjpeg/simd/x86_64/jdcolor-avx2.asm
@@ -0,0 +1,118 @@
+;
+; jdcolor.asm - colorspace conversion (64-bit AVX2)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+; Copyright (C) 2009, 2016, D. R. Commander.
+; Copyright (C) 2015, Intel Corporation.
+;
+; Based on the x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+
+%include "jsimdext.inc"
+
+; --------------------------------------------------------------------------
+
+%define SCALEBITS  16
+
+F_0_344 equ  22554              ; FIX(0.34414)
+F_0_714 equ  46802              ; FIX(0.71414)
+F_1_402 equ  91881              ; FIX(1.40200)
+F_1_772 equ 116130              ; FIX(1.77200)
+F_0_402 equ (F_1_402 - 65536)   ; FIX(1.40200) - FIX(1)
+F_0_285 equ ( 65536 - F_0_714)  ; FIX(1) - FIX(0.71414)
+F_0_228 equ (131072 - F_1_772)  ; FIX(2) - FIX(1.77200)
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_CONST
+
+    alignz      32
+    GLOBAL_DATA(jconst_ycc_rgb_convert_avx2)
+
+EXTN(jconst_ycc_rgb_convert_avx2):
+
+PW_F0402        times 16 dw  F_0_402
+PW_MF0228       times 16 dw -F_0_228
+PW_MF0344_F0285 times 8  dw -F_0_344, F_0_285
+PW_ONE          times 16 dw  1
+PD_ONEHALF      times 8  dd  1 << (SCALEBITS - 1)
+
+    alignz      32
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_TEXT
+    BITS        64
+
+%include "jdcolext-avx2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_RGB_RED
+%define RGB_GREEN  EXT_RGB_GREEN
+%define RGB_BLUE  EXT_RGB_BLUE
+%define RGB_PIXELSIZE  EXT_RGB_PIXELSIZE
+%define jsimd_ycc_rgb_convert_avx2  jsimd_ycc_extrgb_convert_avx2
+%include "jdcolext-avx2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_RGBX_RED
+%define RGB_GREEN  EXT_RGBX_GREEN
+%define RGB_BLUE  EXT_RGBX_BLUE
+%define RGB_PIXELSIZE  EXT_RGBX_PIXELSIZE
+%define jsimd_ycc_rgb_convert_avx2  jsimd_ycc_extrgbx_convert_avx2
+%include "jdcolext-avx2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_BGR_RED
+%define RGB_GREEN  EXT_BGR_GREEN
+%define RGB_BLUE  EXT_BGR_BLUE
+%define RGB_PIXELSIZE  EXT_BGR_PIXELSIZE
+%define jsimd_ycc_rgb_convert_avx2  jsimd_ycc_extbgr_convert_avx2
+%include "jdcolext-avx2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_BGRX_RED
+%define RGB_GREEN  EXT_BGRX_GREEN
+%define RGB_BLUE  EXT_BGRX_BLUE
+%define RGB_PIXELSIZE  EXT_BGRX_PIXELSIZE
+%define jsimd_ycc_rgb_convert_avx2  jsimd_ycc_extbgrx_convert_avx2
+%include "jdcolext-avx2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_XBGR_RED
+%define RGB_GREEN  EXT_XBGR_GREEN
+%define RGB_BLUE  EXT_XBGR_BLUE
+%define RGB_PIXELSIZE  EXT_XBGR_PIXELSIZE
+%define jsimd_ycc_rgb_convert_avx2  jsimd_ycc_extxbgr_convert_avx2
+%include "jdcolext-avx2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_XRGB_RED
+%define RGB_GREEN  EXT_XRGB_GREEN
+%define RGB_BLUE  EXT_XRGB_BLUE
+%define RGB_PIXELSIZE  EXT_XRGB_PIXELSIZE
+%define jsimd_ycc_rgb_convert_avx2  jsimd_ycc_extxrgb_convert_avx2
+%include "jdcolext-avx2.asm"
diff --git a/media/libjpeg/simd/x86_64/jdcolor-sse2.asm b/media/libjpeg/simd/x86_64/jdcolor-sse2.asm
new file mode 100644
index 0000000000..b3f1fec07e
--- /dev/null
+++ b/media/libjpeg/simd/x86_64/jdcolor-sse2.asm
@@ -0,0 +1,117 @@
+;
+; jdcolor.asm - colorspace conversion (64-bit SSE2)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+; Copyright (C) 2009, 2016, D. R. Commander.
+;
+; Based on the x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+
+%include "jsimdext.inc"
+
+; --------------------------------------------------------------------------
+
+%define SCALEBITS  16
+
+F_0_344 equ  22554              ; FIX(0.34414)
+F_0_714 equ  46802              ; FIX(0.71414)
+F_1_402 equ  91881              ; FIX(1.40200)
+F_1_772 equ 116130              ; FIX(1.77200)
+F_0_402 equ (F_1_402 - 65536)   ; FIX(1.40200) - FIX(1)
+F_0_285 equ ( 65536 - F_0_714)  ; FIX(1) - FIX(0.71414)
+F_0_228 equ (131072 - F_1_772)  ; FIX(2) - FIX(1.77200)
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_CONST
+
+    alignz      32
+    GLOBAL_DATA(jconst_ycc_rgb_convert_sse2)
+
+EXTN(jconst_ycc_rgb_convert_sse2):
+
+PW_F0402        times 8 dw  F_0_402
+PW_MF0228       times 8 dw -F_0_228
+PW_MF0344_F0285 times 4 dw -F_0_344, F_0_285
+PW_ONE          times 8 dw  1
+PD_ONEHALF      times 4 dd  1 << (SCALEBITS - 1)
+
+    alignz      32
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_TEXT
+    BITS        64
+
+%include "jdcolext-sse2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_RGB_RED
+%define RGB_GREEN  EXT_RGB_GREEN
+%define RGB_BLUE  EXT_RGB_BLUE
+%define RGB_PIXELSIZE  EXT_RGB_PIXELSIZE
+%define jsimd_ycc_rgb_convert_sse2  jsimd_ycc_extrgb_convert_sse2
+%include "jdcolext-sse2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_RGBX_RED
+%define RGB_GREEN  EXT_RGBX_GREEN
+%define RGB_BLUE  EXT_RGBX_BLUE
+%define RGB_PIXELSIZE  EXT_RGBX_PIXELSIZE
+%define jsimd_ycc_rgb_convert_sse2  jsimd_ycc_extrgbx_convert_sse2
+%include "jdcolext-sse2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_BGR_RED
+%define RGB_GREEN  EXT_BGR_GREEN
+%define RGB_BLUE  EXT_BGR_BLUE
+%define RGB_PIXELSIZE  EXT_BGR_PIXELSIZE
+%define jsimd_ycc_rgb_convert_sse2  jsimd_ycc_extbgr_convert_sse2
+%include "jdcolext-sse2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_BGRX_RED
+%define RGB_GREEN  EXT_BGRX_GREEN
+%define RGB_BLUE  EXT_BGRX_BLUE
+%define RGB_PIXELSIZE  EXT_BGRX_PIXELSIZE
+%define jsimd_ycc_rgb_convert_sse2  jsimd_ycc_extbgrx_convert_sse2
+%include "jdcolext-sse2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_XBGR_RED
+%define RGB_GREEN  EXT_XBGR_GREEN
+%define RGB_BLUE  EXT_XBGR_BLUE
+%define RGB_PIXELSIZE  EXT_XBGR_PIXELSIZE
+%define jsimd_ycc_rgb_convert_sse2  jsimd_ycc_extxbgr_convert_sse2
+%include "jdcolext-sse2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_XRGB_RED
+%define RGB_GREEN  EXT_XRGB_GREEN
+%define RGB_BLUE  EXT_XRGB_BLUE
+%define RGB_PIXELSIZE  EXT_XRGB_PIXELSIZE
+%define jsimd_ycc_rgb_convert_sse2  jsimd_ycc_extxrgb_convert_sse2
+%include "jdcolext-sse2.asm"
diff --git a/media/libjpeg/simd/x86_64/jdmerge-avx2.asm b/media/libjpeg/simd/x86_64/jdmerge-avx2.asm
new file mode 100644
index 0000000000..9515a17013
--- /dev/null
+++ b/media/libjpeg/simd/x86_64/jdmerge-avx2.asm
@@ -0,0 +1,136 @@
+;
+; jdmerge.asm - merged upsampling/color conversion (64-bit AVX2)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+; Copyright (C) 2009, 2016, D. R. Commander.
+; Copyright (C) 2015, Intel Corporation.
+;
+; Based on the x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+
+%include "jsimdext.inc"
+
+; --------------------------------------------------------------------------
+
+%define SCALEBITS  16
+
+F_0_344 equ  22554              ; FIX(0.34414)
+F_0_714 equ  46802              ; FIX(0.71414)
+F_1_402 equ  91881              ; FIX(1.40200)
+F_1_772 equ 116130              ; FIX(1.77200)
+F_0_402 equ (F_1_402 - 65536)   ; FIX(1.40200) - FIX(1)
+F_0_285 equ ( 65536 - F_0_714)  ; FIX(1) - FIX(0.71414)
+F_0_228 equ (131072 - F_1_772)  ; FIX(2) - FIX(1.77200)
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_CONST
+
+    alignz      32
+    GLOBAL_DATA(jconst_merged_upsample_avx2)
+
+EXTN(jconst_merged_upsample_avx2):
+
+PW_F0402        times 16 dw  F_0_402
+PW_MF0228       times 16 dw -F_0_228
+PW_MF0344_F0285 times 8  dw -F_0_344, F_0_285
+PW_ONE          times 16 dw  1
+PD_ONEHALF      times 8  dd  1 << (SCALEBITS - 1)
+
+    alignz      32
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_TEXT
+    BITS        64
+
+%include "jdmrgext-avx2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_RGB_RED
+%define RGB_GREEN  EXT_RGB_GREEN
+%define RGB_BLUE  EXT_RGB_BLUE
+%define RGB_PIXELSIZE  EXT_RGB_PIXELSIZE
+%define jsimd_h2v1_merged_upsample_avx2 \
+  jsimd_h2v1_extrgb_merged_upsample_avx2
+%define jsimd_h2v2_merged_upsample_avx2 \
+  jsimd_h2v2_extrgb_merged_upsample_avx2
+%include "jdmrgext-avx2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_RGBX_RED
+%define RGB_GREEN  EXT_RGBX_GREEN
+%define RGB_BLUE  EXT_RGBX_BLUE
+%define RGB_PIXELSIZE  EXT_RGBX_PIXELSIZE
+%define jsimd_h2v1_merged_upsample_avx2 \
+  jsimd_h2v1_extrgbx_merged_upsample_avx2
+%define jsimd_h2v2_merged_upsample_avx2 \
+  jsimd_h2v2_extrgbx_merged_upsample_avx2
+%include "jdmrgext-avx2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_BGR_RED
+%define RGB_GREEN  EXT_BGR_GREEN
+%define RGB_BLUE  EXT_BGR_BLUE
+%define RGB_PIXELSIZE  EXT_BGR_PIXELSIZE
+%define jsimd_h2v1_merged_upsample_avx2 \
+  jsimd_h2v1_extbgr_merged_upsample_avx2
+%define jsimd_h2v2_merged_upsample_avx2 \
+  jsimd_h2v2_extbgr_merged_upsample_avx2
+%include "jdmrgext-avx2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_BGRX_RED
+%define RGB_GREEN  EXT_BGRX_GREEN
+%define RGB_BLUE  EXT_BGRX_BLUE
+%define RGB_PIXELSIZE  EXT_BGRX_PIXELSIZE
+%define jsimd_h2v1_merged_upsample_avx2 \
+  jsimd_h2v1_extbgrx_merged_upsample_avx2
+%define jsimd_h2v2_merged_upsample_avx2 \
+  jsimd_h2v2_extbgrx_merged_upsample_avx2
+%include "jdmrgext-avx2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_XBGR_RED
+%define RGB_GREEN  EXT_XBGR_GREEN
+%define RGB_BLUE  EXT_XBGR_BLUE
+%define RGB_PIXELSIZE  EXT_XBGR_PIXELSIZE
+%define jsimd_h2v1_merged_upsample_avx2 \
+  jsimd_h2v1_extxbgr_merged_upsample_avx2
+%define jsimd_h2v2_merged_upsample_avx2 \
+  jsimd_h2v2_extxbgr_merged_upsample_avx2
+%include "jdmrgext-avx2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_XRGB_RED
+%define RGB_GREEN  EXT_XRGB_GREEN
+%define RGB_BLUE  EXT_XRGB_BLUE
+%define RGB_PIXELSIZE  EXT_XRGB_PIXELSIZE
+%define jsimd_h2v1_merged_upsample_avx2 \
+  jsimd_h2v1_extxrgb_merged_upsample_avx2
+%define jsimd_h2v2_merged_upsample_avx2 \
+  jsimd_h2v2_extxrgb_merged_upsample_avx2
+%include "jdmrgext-avx2.asm"
diff --git a/media/libjpeg/simd/x86_64/jdmerge-sse2.asm b/media/libjpeg/simd/x86_64/jdmerge-sse2.asm
new file mode 100644
index 0000000000..aedccc20f6
--- /dev/null
+++ b/media/libjpeg/simd/x86_64/jdmerge-sse2.asm
@@ -0,0 +1,135 @@
+;
+; jdmerge.asm - merged upsampling/color conversion (64-bit SSE2)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+; Copyright (C) 2009, 2016, D. R. Commander.
+;
+; Based on the x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+
+%include "jsimdext.inc"
+
+; --------------------------------------------------------------------------
+
+%define SCALEBITS  16
+
+F_0_344 equ  22554              ; FIX(0.34414)
+F_0_714 equ  46802              ; FIX(0.71414)
+F_1_402 equ  91881              ; FIX(1.40200)
+F_1_772 equ 116130              ; FIX(1.77200)
+F_0_402 equ (F_1_402 - 65536)   ; FIX(1.40200) - FIX(1)
+F_0_285 equ ( 65536 - F_0_714)  ; FIX(1) - FIX(0.71414)
+F_0_228 equ (131072 - F_1_772)  ; FIX(2) - FIX(1.77200)
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_CONST
+
+    alignz      32
+    GLOBAL_DATA(jconst_merged_upsample_sse2)
+
+EXTN(jconst_merged_upsample_sse2):
+
+PW_F0402        times 8 dw  F_0_402
+PW_MF0228       times 8 dw -F_0_228
+PW_MF0344_F0285 times 4 dw -F_0_344, F_0_285
+PW_ONE          times 8 dw  1
+PD_ONEHALF      times 4 dd  1 << (SCALEBITS - 1)
+
+    alignz      32
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_TEXT
+    BITS        64
+
+%include "jdmrgext-sse2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_RGB_RED
+%define RGB_GREEN  EXT_RGB_GREEN
+%define RGB_BLUE  EXT_RGB_BLUE
+%define RGB_PIXELSIZE  EXT_RGB_PIXELSIZE
+%define jsimd_h2v1_merged_upsample_sse2 \
+  jsimd_h2v1_extrgb_merged_upsample_sse2
+%define jsimd_h2v2_merged_upsample_sse2 \
+  jsimd_h2v2_extrgb_merged_upsample_sse2
+%include "jdmrgext-sse2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_RGBX_RED
+%define RGB_GREEN  EXT_RGBX_GREEN
+%define RGB_BLUE  EXT_RGBX_BLUE
+%define RGB_PIXELSIZE  EXT_RGBX_PIXELSIZE
+%define jsimd_h2v1_merged_upsample_sse2 \
+  jsimd_h2v1_extrgbx_merged_upsample_sse2
+%define jsimd_h2v2_merged_upsample_sse2 \
+  jsimd_h2v2_extrgbx_merged_upsample_sse2
+%include "jdmrgext-sse2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_BGR_RED
+%define RGB_GREEN  EXT_BGR_GREEN
+%define RGB_BLUE  EXT_BGR_BLUE
+%define RGB_PIXELSIZE  EXT_BGR_PIXELSIZE
+%define jsimd_h2v1_merged_upsample_sse2 \
+  jsimd_h2v1_extbgr_merged_upsample_sse2
+%define jsimd_h2v2_merged_upsample_sse2 \
+  jsimd_h2v2_extbgr_merged_upsample_sse2
+%include "jdmrgext-sse2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_BGRX_RED
+%define RGB_GREEN  EXT_BGRX_GREEN
+%define RGB_BLUE  EXT_BGRX_BLUE
+%define RGB_PIXELSIZE  EXT_BGRX_PIXELSIZE
+%define jsimd_h2v1_merged_upsample_sse2 \
+  jsimd_h2v1_extbgrx_merged_upsample_sse2
+%define jsimd_h2v2_merged_upsample_sse2 \
+  jsimd_h2v2_extbgrx_merged_upsample_sse2
+%include "jdmrgext-sse2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_XBGR_RED
+%define RGB_GREEN  EXT_XBGR_GREEN
+%define RGB_BLUE  EXT_XBGR_BLUE
+%define RGB_PIXELSIZE  EXT_XBGR_PIXELSIZE
+%define jsimd_h2v1_merged_upsample_sse2 \
+  jsimd_h2v1_extxbgr_merged_upsample_sse2
+%define jsimd_h2v2_merged_upsample_sse2 \
+  jsimd_h2v2_extxbgr_merged_upsample_sse2
+%include "jdmrgext-sse2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_XRGB_RED
+%define RGB_GREEN  EXT_XRGB_GREEN
+%define RGB_BLUE  EXT_XRGB_BLUE
+%define RGB_PIXELSIZE  EXT_XRGB_PIXELSIZE
+%define jsimd_h2v1_merged_upsample_sse2 \
+  jsimd_h2v1_extxrgb_merged_upsample_sse2
+%define jsimd_h2v2_merged_upsample_sse2 \
+  jsimd_h2v2_extxrgb_merged_upsample_sse2
+%include "jdmrgext-sse2.asm"
diff --git a/media/libjpeg/simd/x86_64/jdmrgext-avx2.asm b/media/libjpeg/simd/x86_64/jdmrgext-avx2.asm
new file mode 100644
index 0000000000..8b264b4f03
--- /dev/null
+++ b/media/libjpeg/simd/x86_64/jdmrgext-avx2.asm
@@ -0,0 +1,596 @@
+;
+; jdmrgext.asm - merged upsampling/color conversion (64-bit AVX2)
+;
+; Copyright 2009, 2012 Pierre Ossman <ossman@cendio.se> for Cendio AB
+; Copyright (C) 2009, 2012, 2016, D. R. Commander.
+; Copyright (C) 2015, Intel Corporation.
+; Copyright (C) 2018, Matthias Räncker.
+;
+; Based on the x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+
+%include "jcolsamp.inc"
+
+; --------------------------------------------------------------------------
+;
+; Upsample and color convert for the case of 2:1 horizontal and 1:1 vertical.
+;
+; GLOBAL(void)
+; jsimd_h2v1_merged_upsample_avx2(JDIMENSION output_width,
+;                                 JSAMPIMAGE input_buf,
+;                                 JDIMENSION in_row_group_ctr,
+;                                 JSAMPARRAY output_buf);
+;
+
+; r10d = JDIMENSION output_width
+; r11 = JSAMPIMAGE input_buf
+; r12d = JDIMENSION in_row_group_ctr
+; r13 = JSAMPARRAY output_buf
+
+%define wk(i)   rbp - (WK_NUM - (i)) * SIZEOF_YMMWORD  ; ymmword wk[WK_NUM]
+%define WK_NUM  3
+
+    align       32
+    GLOBAL_FUNCTION(jsimd_h2v1_merged_upsample_avx2)
+
+EXTN(jsimd_h2v1_merged_upsample_avx2):
+    push        rbp
+    mov         rax, rsp                     ; rax = original rbp
+    sub         rsp, byte 4
+    and         rsp, byte (-SIZEOF_YMMWORD)  ; align to 256 bits
+    mov         [rsp], rax
+    mov         rbp, rsp                     ; rbp = aligned rbp
+    lea         rsp, [wk(0)]
+    collect_args 4
+    push        rbx
+
+    mov         ecx, r10d               ; col
+    test        rcx, rcx
+    jz          near .return
+
+    push        rcx
+
+    mov         rdi, r11
+    mov         ecx, r12d
+    mov         rsip, JSAMPARRAY [rdi+0*SIZEOF_JSAMPARRAY]
+    mov         rbxp, JSAMPARRAY [rdi+1*SIZEOF_JSAMPARRAY]
+    mov         rdxp, JSAMPARRAY [rdi+2*SIZEOF_JSAMPARRAY]
+    mov         rdi, r13
+    mov         rsip, JSAMPROW [rsi+rcx*SIZEOF_JSAMPROW]  ; inptr0
+    mov         rbxp, JSAMPROW [rbx+rcx*SIZEOF_JSAMPROW]  ; inptr1
+    mov         rdxp, JSAMPROW [rdx+rcx*SIZEOF_JSAMPROW]  ; inptr2
+    mov         rdip, JSAMPROW [rdi]                      ; outptr
+
+    pop         rcx                     ; col
+
+.columnloop:
+
+    vmovdqu     ymm6, YMMWORD [rbx]     ; ymm6=Cb(0123456789ABCDEFGHIJKLMNOPQRSTUV)
+    vmovdqu     ymm7, YMMWORD [rdx]     ; ymm7=Cr(0123456789ABCDEFGHIJKLMNOPQRSTUV)
+
+    vpxor       ymm1, ymm1, ymm1        ; ymm1=(all 0's)
+    vpcmpeqw    ymm3, ymm3, ymm3
+    vpsllw      ymm3, ymm3, 7           ; ymm3={0xFF80 0xFF80 0xFF80 0xFF80 ..}
+
+    vpermq      ymm6, ymm6, 0xd8        ; ymm6=Cb(01234567GHIJKLMN89ABCDEFOPQRSTUV)
+    vpermq      ymm7, ymm7, 0xd8        ; ymm7=Cr(01234567GHIJKLMN89ABCDEFOPQRSTUV)
+    vpunpcklbw  ymm4, ymm6, ymm1        ; ymm4=Cb(0123456789ABCDEF)=CbL
+    vpunpckhbw  ymm6, ymm6, ymm1        ; ymm6=Cb(GHIJKLMNOPQRSTUV)=CbH
+    vpunpcklbw  ymm0, ymm7, ymm1        ; ymm0=Cr(0123456789ABCDEF)=CrL
+    vpunpckhbw  ymm7, ymm7, ymm1        ; ymm7=Cr(GHIJKLMNOPQRSTUV)=CrH
+
+    vpaddw      ymm5, ymm6, ymm3
+    vpaddw      ymm2, ymm4, ymm3
+    vpaddw      ymm1, ymm7, ymm3
+    vpaddw      ymm3, ymm0, ymm3
+
+    ; (Original)
+    ; R = Y                + 1.40200 * Cr
+    ; G = Y - 0.34414 * Cb - 0.71414 * Cr
+    ; B = Y + 1.77200 * Cb
+    ;
+    ; (This implementation)
+    ; R = Y                + 0.40200 * Cr + Cr
+    ; G = Y - 0.34414 * Cb + 0.28586 * Cr - Cr
+    ; B = Y - 0.22800 * Cb + Cb + Cb
+
+    vpaddw      ymm6, ymm5, ymm5             ; ymm6=2*CbH
+    vpaddw      ymm4, ymm2, ymm2             ; ymm4=2*CbL
+    vpaddw      ymm7, ymm1, ymm1             ; ymm7=2*CrH
+    vpaddw      ymm0, ymm3, ymm3             ; ymm0=2*CrL
+
+    vpmulhw     ymm6, ymm6, [rel PW_MF0228]  ; ymm6=(2*CbH * -FIX(0.22800))
+    vpmulhw     ymm4, ymm4, [rel PW_MF0228]  ; ymm4=(2*CbL * -FIX(0.22800))
+    vpmulhw     ymm7, ymm7, [rel PW_F0402]   ; ymm7=(2*CrH * FIX(0.40200))
+    vpmulhw     ymm0, ymm0, [rel PW_F0402]   ; ymm0=(2*CrL * FIX(0.40200))
+
+    vpaddw      ymm6, ymm6, [rel PW_ONE]
+    vpaddw      ymm4, ymm4, [rel PW_ONE]
+    vpsraw      ymm6, ymm6, 1                ; ymm6=(CbH * -FIX(0.22800))
+    vpsraw      ymm4, ymm4, 1                ; ymm4=(CbL * -FIX(0.22800))
+    vpaddw      ymm7, ymm7, [rel PW_ONE]
+    vpaddw      ymm0, ymm0, [rel PW_ONE]
+    vpsraw      ymm7, ymm7, 1                ; ymm7=(CrH * FIX(0.40200))
+    vpsraw      ymm0, ymm0, 1                ; ymm0=(CrL * FIX(0.40200))
+
+    vpaddw      ymm6, ymm6, ymm5
+    vpaddw      ymm4, ymm4, ymm2
+    vpaddw      ymm6, ymm6, ymm5             ; ymm6=(CbH * FIX(1.77200))=(B-Y)H
+    vpaddw      ymm4, ymm4, ymm2             ; ymm4=(CbL * FIX(1.77200))=(B-Y)L
+    vpaddw      ymm7, ymm7, ymm1             ; ymm7=(CrH * FIX(1.40200))=(R-Y)H
+    vpaddw      ymm0, ymm0, ymm3             ; ymm0=(CrL * FIX(1.40200))=(R-Y)L
+
+    vmovdqa     YMMWORD [wk(0)], ymm6        ; wk(0)=(B-Y)H
+    vmovdqa     YMMWORD [wk(1)], ymm7        ; wk(1)=(R-Y)H
+
+    vpunpckhwd  ymm6, ymm5, ymm1
+    vpunpcklwd  ymm5, ymm5, ymm1
+    vpmaddwd    ymm5, ymm5, [rel PW_MF0344_F0285]
+    vpmaddwd    ymm6, ymm6, [rel PW_MF0344_F0285]
+    vpunpckhwd  ymm7, ymm2, ymm3
+    vpunpcklwd  ymm2, ymm2, ymm3
+    vpmaddwd    ymm2, ymm2, [rel PW_MF0344_F0285]
+    vpmaddwd    ymm7, ymm7, [rel PW_MF0344_F0285]
+
+    vpaddd      ymm5, ymm5, [rel PD_ONEHALF]
+    vpaddd      ymm6, ymm6, [rel PD_ONEHALF]
+    vpsrad      ymm5, ymm5, SCALEBITS
+    vpsrad      ymm6, ymm6, SCALEBITS
+    vpaddd      ymm2, ymm2, [rel PD_ONEHALF]
+    vpaddd      ymm7, ymm7, [rel PD_ONEHALF]
+    vpsrad      ymm2, ymm2, SCALEBITS
+    vpsrad      ymm7, ymm7, SCALEBITS
+
+    vpackssdw   ymm5, ymm5, ymm6        ; ymm5=CbH*-FIX(0.344)+CrH*FIX(0.285)
+    vpackssdw   ymm2, ymm2, ymm7        ; ymm2=CbL*-FIX(0.344)+CrL*FIX(0.285)
+    vpsubw      ymm5, ymm5, ymm1        ; ymm5=CbH*-FIX(0.344)+CrH*-FIX(0.714)=(G-Y)H
+    vpsubw      ymm2, ymm2, ymm3        ; ymm2=CbL*-FIX(0.344)+CrL*-FIX(0.714)=(G-Y)L
+
+    vmovdqa     YMMWORD [wk(2)], ymm5   ; wk(2)=(G-Y)H
+
+    mov         al, 2                   ; Yctr
+    jmp         short .Yloop_1st
+
+.Yloop_2nd:
+    vmovdqa     ymm0, YMMWORD [wk(1)]   ; ymm0=(R-Y)H
+    vmovdqa     ymm2, YMMWORD [wk(2)]   ; ymm2=(G-Y)H
+    vmovdqa     ymm4, YMMWORD [wk(0)]   ; ymm4=(B-Y)H
+
+.Yloop_1st:
+    vmovdqu     ymm7, YMMWORD [rsi]     ; ymm7=Y(0123456789ABCDEFGHIJKLMNOPQRSTUV)
+
+    vpcmpeqw    ymm6, ymm6, ymm6
+    vpsrlw      ymm6, ymm6, BYTE_BIT    ; ymm6={0xFF 0x00 0xFF 0x00 ..}
+    vpand       ymm6, ymm6, ymm7        ; ymm6=Y(02468ACEGIKMOQSU)=YE
+    vpsrlw      ymm7, ymm7, BYTE_BIT    ; ymm7=Y(13579BDFHJLNPRTV)=YO
+
+    vmovdqa     ymm1, ymm0              ; ymm1=ymm0=(R-Y)(L/H)
+    vmovdqa     ymm3, ymm2              ; ymm3=ymm2=(G-Y)(L/H)
+    vmovdqa     ymm5, ymm4              ; ymm5=ymm4=(B-Y)(L/H)
+
+    vpaddw      ymm0, ymm0, ymm6        ; ymm0=((R-Y)+YE)=RE=R(02468ACEGIKMOQSU)
+    vpaddw      ymm1, ymm1, ymm7        ; ymm1=((R-Y)+YO)=RO=R(13579BDFHJLNPRTV)
+    vpackuswb   ymm0, ymm0, ymm0        ; ymm0=R(02468ACE********GIKMOQSU********)
+    vpackuswb   ymm1, ymm1, ymm1        ; ymm1=R(13579BDF********HJLNPRTV********)
+
+    vpaddw      ymm2, ymm2, ymm6        ; ymm2=((G-Y)+YE)=GE=G(02468ACEGIKMOQSU)
+    vpaddw      ymm3, ymm3, ymm7        ; ymm3=((G-Y)+YO)=GO=G(13579BDFHJLNPRTV)
+    vpackuswb   ymm2, ymm2, ymm2        ; ymm2=G(02468ACE********GIKMOQSU********)
+    vpackuswb   ymm3, ymm3, ymm3        ; ymm3=G(13579BDF********HJLNPRTV********)
+
+    vpaddw      ymm4, ymm4, ymm6        ; ymm4=((B-Y)+YE)=BE=B(02468ACEGIKMOQSU)
+    vpaddw      ymm5, ymm5, ymm7        ; ymm5=((B-Y)+YO)=BO=B(13579BDFHJLNPRTV)
+    vpackuswb   ymm4, ymm4, ymm4        ; ymm4=B(02468ACE********GIKMOQSU********)
+    vpackuswb   ymm5, ymm5, ymm5        ; ymm5=B(13579BDF********HJLNPRTV********)
+
+%if RGB_PIXELSIZE == 3  ; ---------------
+
+    ; ymmA=(00 02 04 06 08 0A 0C 0E ** 0G 0I 0K 0M 0O 0Q 0S 0U **)
+    ; ymmB=(01 03 05 07 09 0B 0D 0F ** 0H 0J 0L 0N 0P 0R 0T 0V **)
+    ; ymmC=(10 12 14 16 18 1A 1C 1E ** 1G 1I 1K 1M 1O 1Q 1S 1U **)
+    ; ymmD=(11 13 15 17 19 1B 1D 1F ** 1H 1J 1L 1N 1P 1R 1T 1V **)
+    ; ymmE=(20 22 24 26 28 2A 2C 2E ** 2G 2I 2K 2M 2O 2Q 2S 2U **)
+    ; ymmF=(21 23 25 27 29 2B 2D 2F ** 2H 2J 2L 2N 2P 2R 2T 2V **)
+    ; ymmG=(** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** **)
+    ; ymmH=(** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** **)
+
+    vpunpcklbw  ymmA, ymmA, ymmC        ; ymmA=(00 10 02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E
+                                        ;       0G 1G 0I 1I 0K 1K 0M 1M 0O 1O 0Q 1Q 0S 1S 0U 1U)
+    vpunpcklbw  ymmE, ymmE, ymmB        ; ymmE=(20 01 22 03 24 05 26 07 28 09 2A 0B 2C 0D 2E 0F
+                                        ;       2G 0H 2I 0J 2K 0L 2M 0N 2O 0P 2Q 0R 2S 0T 2U 0V)
+    vpunpcklbw  ymmD, ymmD, ymmF        ; ymmD=(11 21 13 23 15 25 17 27 19 29 1B 2B 1D 2D 1F 2F
+                                        ;       1H 2H 1J 2J 1L 2L 1N 2N 1P 2P 1R 2R 1T 2T 1V 2V)
+
+    vpsrldq     ymmH, ymmA, 2           ; ymmH=(02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E 0G 1G
+                                        ;       0I 1I 0K 1K 0M 1M 0O 1O 0Q 1Q 0S 1S 0U 1U -- --)
+    vpunpckhwd  ymmG, ymmA, ymmE        ; ymmG=(08 18 28 09 0A 1A 2A 0B 0C 1C 2C 0D 0E 1E 2E 0F
+                                        ;       0O 1O 2O 0P 0Q 1Q 2Q 0R 0S 1S 2S 0T 0U 1U 2U 0V)
+    vpunpcklwd  ymmA, ymmA, ymmE        ; ymmA=(00 10 20 01 02 12 22 03 04 14 24 05 06 16 26 07
+                                        ;       0G 1G 2G 0H 0I 1I 2I 0J 0K 1K 2K 0L 0M 1M 2M 0N)
+
+    vpsrldq     ymmE, ymmE, 2           ; ymmE=(22 03 24 05 26 07 28 09 2A 0B 2C 0D 2E 0F 2G 0H
+                                        ;       2I 0J 2K 0L 2M 0N 2O 0P 2Q 0R 2S 0T 2U 0V -- --)
+
+    vpsrldq     ymmB, ymmD, 2           ; ymmB=(13 23 15 25 17 27 19 29 1B 2B 1D 2D 1F 2F 1H 2H
+                                        ;       1J 2J 1L 2L 1N 2N 1P 2P 1R 2R 1T 2T 1V 2V -- --)
+    vpunpckhwd  ymmC, ymmD, ymmH        ; ymmC=(19 29 0A 1A 1B 2B 0C 1C 1D 2D 0E 1E 1F 2F 0G 1G
+                                        ;       1P 2P 0Q 1Q 1R 2R 0S 1S 1T 2T 0U 1U 1V 2V -- --)
+    vpunpcklwd  ymmD, ymmD, ymmH        ; ymmD=(11 21 02 12 13 23 04 14 15 25 06 16 17 27 08 18
+                                        ;       1H 2H 0I 1I 1J 2J 0K 1K 1L 2L 0M 1M 1N 2N 0O 1O)
+
+    vpunpckhwd  ymmF, ymmE, ymmB        ; ymmF=(2A 0B 1B 2B 2C 0D 1D 2D 2E 0F 1F 2F 2G 0H 1H 2H
+                                        ;       2Q 0R 1R 2R 2S 0T 1T 2T 2U 0V 1V 2V -- -- -- --)
+    vpunpcklwd  ymmE, ymmE, ymmB        ; ymmE=(22 03 13 23 24 05 15 25 26 07 17 27 28 09 19 29
+                                        ;       2I 0J 1J 2J 2K 0L 1L 2L 2M 0N 1N 2N 2O 0P 1P 2P)
+
+    vpshufd     ymmH, ymmA, 0x4E        ; ymmH=(04 14 24 05 06 16 26 07 00 10 20 01 02 12 22 03
+                                        ;       0K 1K 2K 0L 0M 1M 2M 0N 0G 1G 2G 0H 0I 1I 2I 0J)
+    vpunpckldq  ymmA, ymmA, ymmD        ; ymmA=(00 10 20 01 11 21 02 12 02 12 22 03 13 23 04 14
+                                        ;       0G 1G 2G 0H 1H 2H 0I 1I 0I 1I 2I 0J 1J 2J 0K 1K)
+    vpunpckhdq  ymmD, ymmD, ymmE        ; ymmD=(15 25 06 16 26 07 17 27 17 27 08 18 28 09 19 29
+                                        ;       1L 2L 0M 1M 2M 0N 1N 2N 1N 2N 0O 1O 2O 0P 1P 2P)
+    vpunpckldq  ymmE, ymmE, ymmH        ; ymmE=(22 03 13 23 04 14 24 05 24 05 15 25 06 16 26 07
+                                        ;       2I 0J 1J 2J 0K 1K 2K 0L 2K 0L 1L 2L 0M 1M 2M 0N)
+
+    vpshufd     ymmH, ymmG, 0x4E        ; ymmH=(0C 1C 2C 0D 0E 1E 2E 0F 08 18 28 09 0A 1A 2A 0B
+                                        ;       0S 1S 2S 0T 0U 1U 2U 0V 0O 1O 2O 0P 0Q 1Q 2Q 0R)
+    vpunpckldq  ymmG, ymmG, ymmC        ; ymmG=(08 18 28 09 19 29 0A 1A 0A 1A 2A 0B 1B 2B 0C 1C
+                                        ;       0O 1O 2O 0P 1P 2P 0Q 1Q 0Q 1Q 2Q 0R 1R 2R 0S 1S)
+    vpunpckhdq  ymmC, ymmC, ymmF        ; ymmC=(1D 2D 0E 1E 2E 0F 1F 2F 1F 2F 0G 1G 2G 0H 1H 2H
+                                        ;       1T 2T 0U 1U 2U 0V 1V 2V 1V 2V -- -- -- -- -- --)
+    vpunpckldq  ymmF, ymmF, ymmH        ; ymmF=(2A 0B 1B 2B 0C 1C 2C 0D 2C 0D 1D 2D 0E 1E 2E 0F
+                                        ;       2Q 0R 1R 2R 0S 1S 2S 0T 2S 0T 1T 2T 0U 1U 2U 0V)
+
+    vpunpcklqdq ymmH, ymmA, ymmE        ; ymmH=(00 10 20 01 11 21 02 12 22 03 13 23 04 14 24 05
+                                        ;       0G 1G 2G 0H 1H 2H 0I 1I 2I 0J 1J 2J 0K 1K 2K 0L)
+    vpunpcklqdq ymmG, ymmD, ymmG        ; ymmG=(15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A
+                                        ;       1L 2L 0M 1M 2M 0N 1N 2N 0O 1O 2O 0P 1P 2P 0Q 1Q)
+    vpunpcklqdq ymmC, ymmF, ymmC        ; ymmC=(2A 0B 1B 2B 0C 1C 2C 0D 1D 2D 0E 1E 2E 0F 1F 2F
+                                        ;       2Q 0R 1R 2R 0S 1S 2S 0T 1T 2T 0U 1U 2U 0V 1V 2V)
+
+    vperm2i128  ymmA, ymmH, ymmG, 0x20  ; ymmA=(00 10 20 01 11 21 02 12 22 03 13 23 04 14 24 05
+                                        ;       15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A)
+    vperm2i128  ymmD, ymmC, ymmH, 0x30  ; ymmD=(2A 0B 1B 2B 0C 1C 2C 0D 1D 2D 0E 1E 2E 0F 1F 2F
+                                        ;       0G 1G 2G 0H 1H 2H 0I 1I 2I 0J 1J 2J 0K 1K 2K 0L)
+    vperm2i128  ymmF, ymmG, ymmC, 0x31  ; ymmF=(1L 2L 0M 1M 2M 0N 1N 2N 0O 1O 2O 0P 1P 2P 0Q 1Q
+                                        ;       2Q 0R 1R 2R 0S 1S 2S 0T 1T 2T 0U 1U 2U 0V 1V 2V)
+
+    cmp         rcx, byte SIZEOF_YMMWORD
+    jb          short .column_st64
+
+    test        rdi, SIZEOF_YMMWORD-1
+    jnz         short .out1
+    ; --(aligned)-------------------
+    vmovntdq    YMMWORD [rdi+0*SIZEOF_YMMWORD], ymmA
+    vmovntdq    YMMWORD [rdi+1*SIZEOF_YMMWORD], ymmD
+    vmovntdq    YMMWORD [rdi+2*SIZEOF_YMMWORD], ymmF
+    jmp         short .out0
+.out1:  ; --(unaligned)-----------------
+    vmovdqu     YMMWORD [rdi+0*SIZEOF_YMMWORD], ymmA
+    vmovdqu     YMMWORD [rdi+1*SIZEOF_YMMWORD], ymmD
+    vmovdqu     YMMWORD [rdi+2*SIZEOF_YMMWORD], ymmF
+.out0:
+    add         rdi, byte RGB_PIXELSIZE*SIZEOF_YMMWORD  ; outptr
+    sub         rcx, byte SIZEOF_YMMWORD
+    jz          near .endcolumn
+
+    add         rsi, byte SIZEOF_YMMWORD  ; inptr0
+    dec         al                        ; Yctr
+    jnz         near .Yloop_2nd
+
+    add         rbx, byte SIZEOF_YMMWORD  ; inptr1
+    add         rdx, byte SIZEOF_YMMWORD  ; inptr2
+    jmp         near .columnloop
+
+.column_st64:
+    lea         rcx, [rcx+rcx*2]            ; imul ecx, RGB_PIXELSIZE
+    cmp         rcx, byte 2*SIZEOF_YMMWORD
+    jb          short .column_st32
+    vmovdqu     YMMWORD [rdi+0*SIZEOF_YMMWORD], ymmA
+    vmovdqu     YMMWORD [rdi+1*SIZEOF_YMMWORD], ymmD
+    add         rdi, byte 2*SIZEOF_YMMWORD  ; outptr
+    vmovdqa     ymmA, ymmF
+    sub         rcx, byte 2*SIZEOF_YMMWORD
+    jmp         short .column_st31
+.column_st32:
+    cmp         rcx, byte SIZEOF_YMMWORD
+    jb          short .column_st31
+    vmovdqu     YMMWORD [rdi+0*SIZEOF_YMMWORD], ymmA
+    add         rdi, byte SIZEOF_YMMWORD    ; outptr
+    vmovdqa     ymmA, ymmD
+    sub         rcx, byte SIZEOF_YMMWORD
+    jmp         short .column_st31
+.column_st31:
+    cmp         rcx, byte SIZEOF_XMMWORD
+    jb          short .column_st15
+    vmovdqu     XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
+    add         rdi, byte SIZEOF_XMMWORD    ; outptr
+    vperm2i128  ymmA, ymmA, ymmA, 1
+    sub         rcx, byte SIZEOF_XMMWORD
+.column_st15:
+    ; Store the lower 8 bytes of xmmA to the output when it has enough
+    ; space.
+    cmp         rcx, byte SIZEOF_MMWORD
+    jb          short .column_st7
+    vmovq       XMM_MMWORD [rdi], xmmA
+    add         rdi, byte SIZEOF_MMWORD
+    sub         rcx, byte SIZEOF_MMWORD
+    vpsrldq     xmmA, xmmA, SIZEOF_MMWORD
+.column_st7:
+    ; Store the lower 4 bytes of xmmA to the output when it has enough
+    ; space.
+    cmp         rcx, byte SIZEOF_DWORD
+    jb          short .column_st3
+    vmovd       XMM_DWORD [rdi], xmmA
+    add         rdi, byte SIZEOF_DWORD
+    sub         rcx, byte SIZEOF_DWORD
+    vpsrldq     xmmA, xmmA, SIZEOF_DWORD
+.column_st3:
+    ; Store the lower 2 bytes of rax to the output when it has enough
+    ; space.
+    vmovd       eax, xmmA
+    cmp         rcx, byte SIZEOF_WORD
+    jb          short .column_st1
+    mov         word [rdi], ax
+    add         rdi, byte SIZEOF_WORD
+    sub         rcx, byte SIZEOF_WORD
+    shr         rax, 16
+.column_st1:
+    ; Store the lower 1 byte of rax to the output when it has enough
+    ; space.
+    test        rcx, rcx
+    jz          short .endcolumn
+    mov         byte [rdi], al
+
+%else  ; RGB_PIXELSIZE == 4 ; -----------
+
+%ifdef RGBX_FILLER_0XFF
+    vpcmpeqb    ymm6, ymm6, ymm6        ; ymm6=XE=X(02468ACE********GIKMOQSU********)
+    vpcmpeqb    ymm7, ymm7, ymm7        ; ymm7=XO=X(13579BDF********HJLNPRTV********)
+%else
+    vpxor       ymm6, ymm6, ymm6        ; ymm6=XE=X(02468ACE********GIKMOQSU********)
+    vpxor       ymm7, ymm7, ymm7        ; ymm7=XO=X(13579BDF********HJLNPRTV********)
+%endif
+    ; ymmA=(00 02 04 06 08 0A 0C 0E ** 0G 0I 0K 0M 0O 0Q 0S 0U **)
+    ; ymmB=(01 03 05 07 09 0B 0D 0F ** 0H 0J 0L 0N 0P 0R 0T 0V **)
+    ; ymmC=(10 12 14 16 18 1A 1C 1E ** 1G 1I 1K 1M 1O 1Q 1S 1U **)
+    ; ymmD=(11 13 15 17 19 1B 1D 1F ** 1H 1J 1L 1N 1P 1R 1T 1V **)
+    ; ymmE=(20 22 24 26 28 2A 2C 2E ** 2G 2I 2K 2M 2O 2Q 2S 2U **)
+    ; ymmF=(21 23 25 27 29 2B 2D 2F ** 2H 2J 2L 2N 2P 2R 2T 2V **)
+    ; ymmG=(30 32 34 36 38 3A 3C 3E ** 3G 3I 3K 3M 3O 3Q 3S 3U **)
+    ; ymmH=(31 33 35 37 39 3B 3D 3F ** 3H 3J 3L 3N 3P 3R 3T 3V **)
+
+    vpunpcklbw  ymmA, ymmA, ymmC        ; ymmA=(00 10 02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E
+                                        ;       0G 1G 0I 1I 0K 1K 0M 1M 0O 1O 0Q 1Q 0S 1S 0U 1U)
+    vpunpcklbw  ymmE, ymmE, ymmG        ; ymmE=(20 30 22 32 24 34 26 36 28 38 2A 3A 2C 3C 2E 3E
+                                        ;       2G 3G 2I 3I 2K 3K 2M 3M 2O 3O 2Q 3Q 2S 3S 2U 3U)
+    vpunpcklbw  ymmB, ymmB, ymmD        ; ymmB=(01 11 03 13 05 15 07 17 09 19 0B 1B 0D 1D 0F 1F
+                                        ;       0H 1H 0J 1J 0L 1L 0N 1N 0P 1P 0R 1R 0T 1T 0V 1V)
+    vpunpcklbw  ymmF, ymmF, ymmH        ; ymmF=(21 31 23 33 25 35 27 37 29 39 2B 3B 2D 3D 2F 3F
+                                        ;       2H 3H 2J 3J 2L 3L 2N 3N 2P 3P 2R 3R 2T 3T 2V 3V)
+
+    vpunpckhwd  ymmC, ymmA, ymmE        ; ymmC=(08 18 28 38 0A 1A 2A 3A 0C 1C 2C 3C 0E 1E 2E 3E
+                                        ;       0O 1O 2O 3O 0Q 1Q 2Q 3Q 0S 1S 2S 3S 0U 1U 2U 3U)
+    vpunpcklwd  ymmA, ymmA, ymmE        ; ymmA=(00 10 20 30 02 12 22 32 04 14 24 34 06 16 26 36
+                                        ;       0G 1G 2G 3G 0I 1I 2I 3I 0K 1K 2K 3K 0M 1M 2M 3M)
+    vpunpckhwd  ymmG, ymmB, ymmF        ; ymmG=(09 19 29 39 0B 1B 2B 3B 0D 1D 2D 3D 0F 1F 2F 3F
+                                        ;       0P 1P 2P 3P 0R 1R 2R 3R 0T 1T 2T 3T 0V 1V 2V 3V)
+    vpunpcklwd  ymmB, ymmB, ymmF        ; ymmB=(01 11 21 31 03 13 23 33 05 15 25 35 07 17 27 37
+                                        ;       0H 1H 2H 3H 0J 1J 2J 3J 0L 1L 2L 3L 0N 1N 2N 3N)
+
+    vpunpckhdq  ymmE, ymmA, ymmB        ; ymmE=(04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37
+                                        ;       0K 1K 2K 3K 0L 1L 2L 3L 0M 1M 2M 3M 0N 1N 2N 3N)
+    vpunpckldq  ymmB, ymmA, ymmB        ; ymmB=(00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33
+                                        ;       0G 1G 2G 3G 0H 1H 2H 3H 0I 1I 2I 3I 0J 1J 2J 3J)
+    vpunpckhdq  ymmF, ymmC, ymmG        ; ymmF=(0C 1C 2C 3C 0D 1D 2D 3D 0E 1E 2E 3E 0F 1F 2F 3F
+                                        ;       0S 1S 2S 3S 0T 1T 2T 3T 0U 1U 2U 3U 0V 1V 2V 3V)
+    vpunpckldq  ymmG, ymmC, ymmG        ; ymmG=(08 18 28 38 09 19 29 39 0A 1A 2A 3A 0B 1B 2B 3B
+                                        ;       0O 1O 2O 3O 0P 1P 2P 3P 0Q 1Q 2Q 3Q 0R 1R 2R 3R)
+
+    vperm2i128  ymmA, ymmB, ymmE, 0x20  ; ymmA=(00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33
+                                        ;       04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37)
+    vperm2i128  ymmD, ymmG, ymmF, 0x20  ; ymmD=(08 18 28 38 09 19 29 39 0A 1A 2A 3A 0B 1B 2B 3B
+                                        ;       0C 1C 2C 3C 0D 1D 2D 3D 0E 1E 2E 3E 0F 1F 2F 3F)
+    vperm2i128  ymmC, ymmB, ymmE, 0x31  ; ymmC=(0G 1G 2G 3G 0H 1H 2H 3H 0I 1I 2I 3I 0J 1J 2J 3J
+                                        ;       0K 1K 2K 3K 0L 1L 2L 3L 0M 1M 2M 3M 0N 1N 2N 3N)
+    vperm2i128  ymmH, ymmG, ymmF, 0x31  ; ymmH=(0O 1O 2O 3O 0P 1P 2P 3P 0Q 1Q 2Q 3Q 0R 1R 2R 3R
+                                        ;       0S 1S 2S 3S 0T 1T 2T 3T 0U 1U 2U 3U 0V 1V 2V 3V)
+
+    cmp         rcx, byte SIZEOF_YMMWORD
+    jb          short .column_st64
+
+    test        rdi, SIZEOF_YMMWORD-1
+    jnz         short .out1
+    ; --(aligned)-------------------
+    vmovntdq    YMMWORD [rdi+0*SIZEOF_YMMWORD], ymmA
+    vmovntdq    YMMWORD [rdi+1*SIZEOF_YMMWORD], ymmD
+    vmovntdq    YMMWORD [rdi+2*SIZEOF_YMMWORD], ymmC
+    vmovntdq    YMMWORD [rdi+3*SIZEOF_YMMWORD], ymmH
+    jmp         short .out0
+.out1:  ; --(unaligned)-----------------
+    vmovdqu     YMMWORD [rdi+0*SIZEOF_YMMWORD], ymmA
+    vmovdqu     YMMWORD [rdi+1*SIZEOF_YMMWORD], ymmD
+    vmovdqu     YMMWORD [rdi+2*SIZEOF_YMMWORD], ymmC
+    vmovdqu     YMMWORD [rdi+3*SIZEOF_YMMWORD], ymmH
+.out0:
+    add         rdi, RGB_PIXELSIZE*SIZEOF_YMMWORD  ; outptr
+    sub         rcx, byte SIZEOF_YMMWORD
+    jz          near .endcolumn
+
+    add         rsi, byte SIZEOF_YMMWORD  ; inptr0
+    dec         al
+    jnz         near .Yloop_2nd
+
+    add         rbx, byte SIZEOF_YMMWORD  ; inptr1
+    add         rdx, byte SIZEOF_YMMWORD  ; inptr2
+    jmp         near .columnloop
+
+.column_st64:
+    cmp         rcx, byte SIZEOF_YMMWORD/2
+    jb          short .column_st32
+    vmovdqu     YMMWORD [rdi+0*SIZEOF_YMMWORD], ymmA
+    vmovdqu     YMMWORD [rdi+1*SIZEOF_YMMWORD], ymmD
+    add         rdi, byte 2*SIZEOF_YMMWORD  ; outptr
+    vmovdqa     ymmA, ymmC
+    vmovdqa     ymmD, ymmH
+    sub         rcx, byte SIZEOF_YMMWORD/2
+.column_st32:
+    cmp         rcx, byte SIZEOF_YMMWORD/4
+    jb          short .column_st16
+    vmovdqu     YMMWORD [rdi+0*SIZEOF_YMMWORD], ymmA
+    add         rdi, byte SIZEOF_YMMWORD    ; outptr
+    vmovdqa     ymmA, ymmD
+    sub         rcx, byte SIZEOF_YMMWORD/4
+.column_st16:
+    cmp         rcx, byte SIZEOF_YMMWORD/8
+    jb          short .column_st15
+    vmovdqu     XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
+    add         rdi, byte SIZEOF_XMMWORD    ; outptr
+    vperm2i128  ymmA, ymmA, ymmA, 1
+    sub         rcx, byte SIZEOF_YMMWORD/8
+.column_st15:
+    ; Store two pixels (8 bytes) of ymmA to the output when it has enough
+    ; space.
+    cmp         rcx, byte SIZEOF_YMMWORD/16
+    jb          short .column_st7
+    vmovq       MMWORD [rdi], xmmA
+    add         rdi, byte SIZEOF_YMMWORD/16*4
+    sub         rcx, byte SIZEOF_YMMWORD/16
+    vpsrldq     xmmA, SIZEOF_YMMWORD/16*4
+.column_st7:
+    ; Store one pixel (4 bytes) of ymmA to the output when it has enough
+    ; space.
+    test        rcx, rcx
+    jz          short .endcolumn
+    vmovd       XMM_DWORD [rdi], xmmA
+
+%endif  ; RGB_PIXELSIZE ; ---------------
+
+.endcolumn:
+    sfence                              ; flush the write buffer
+
+.return:
+    pop         rbx
+    vzeroupper
+    uncollect_args 4
+    mov         rsp, rbp                ; rsp <- aligned rbp
+    pop         rsp                     ; rsp <- original rbp
+    pop         rbp
+    ret
+
+; --------------------------------------------------------------------------
+;
+; Upsample and color convert for the case of 2:1 horizontal and 2:1 vertical.
+;
+; GLOBAL(void)
+; jsimd_h2v2_merged_upsample_avx2(JDIMENSION output_width,
+;                                 JSAMPIMAGE input_buf,
+;                                 JDIMENSION in_row_group_ctr,
+;                                 JSAMPARRAY output_buf);
+;
+
+; r10d = JDIMENSION output_width
+; r11 = JSAMPIMAGE input_buf
+; r12d = JDIMENSION in_row_group_ctr
+; r13 = JSAMPARRAY output_buf
+
+    align       32
+    GLOBAL_FUNCTION(jsimd_h2v2_merged_upsample_avx2)
+
+EXTN(jsimd_h2v2_merged_upsample_avx2):
+    push        rbp
+    mov         rax, rsp
+    mov         rbp, rsp
+    collect_args 4
+    push        rbx
+
+    mov         eax, r10d
+
+    mov         rdi, r11
+    mov         ecx, r12d
+    mov         rsip, JSAMPARRAY [rdi+0*SIZEOF_JSAMPARRAY]
+    mov         rbxp, JSAMPARRAY [rdi+1*SIZEOF_JSAMPARRAY]
+    mov         rdxp, JSAMPARRAY [rdi+2*SIZEOF_JSAMPARRAY]
+    mov         rdi, r13
+    lea         rsi, [rsi+rcx*SIZEOF_JSAMPROW]
+
+    sub         rsp, SIZEOF_JSAMPARRAY*4
+    mov         JSAMPARRAY [rsp+0*SIZEOF_JSAMPARRAY], rsip  ; intpr00
+    mov         JSAMPARRAY [rsp+1*SIZEOF_JSAMPARRAY], rbxp  ; intpr1
+    mov         JSAMPARRAY [rsp+2*SIZEOF_JSAMPARRAY], rdxp  ; intpr2
+    mov         rbx, rsp
+
+    push        rdi
+    push        rcx
+    push        rax
+
+    %ifdef WIN64
+    mov         r8, rcx
+    mov         r9, rdi
+    mov         rcx, rax
+    mov         rdx, rbx
+    %else
+    mov         rdx, rcx
+    mov         rcx, rdi
+    mov         rdi, rax
+    mov         rsi, rbx
+    %endif
+
+    call        EXTN(jsimd_h2v1_merged_upsample_avx2)
+
+    pop         rax
+    pop         rcx
+    pop         rdi
+    mov         rsip, JSAMPARRAY [rsp+0*SIZEOF_JSAMPARRAY]
+    mov         rbxp, JSAMPARRAY [rsp+1*SIZEOF_JSAMPARRAY]
+    mov         rdxp, JSAMPARRAY [rsp+2*SIZEOF_JSAMPARRAY]
+
+    add         rdi, byte SIZEOF_JSAMPROW  ; outptr1
+    add         rsi, byte SIZEOF_JSAMPROW  ; inptr01
+
+    mov         JSAMPARRAY [rsp+0*SIZEOF_JSAMPARRAY], rsip  ; intpr00
+    mov         JSAMPARRAY [rsp+1*SIZEOF_JSAMPARRAY], rbxp  ; intpr1
+    mov         JSAMPARRAY [rsp+2*SIZEOF_JSAMPARRAY], rdxp  ; intpr2
+    mov         rbx, rsp
+
+    push        rdi
+    push        rcx
+    push        rax
+
+    %ifdef WIN64
+    mov         r8, rcx
+    mov         r9, rdi
+    mov         rcx, rax
+    mov         rdx, rbx
+    %else
+    mov         rdx, rcx
+    mov         rcx, rdi
+    mov         rdi, rax
+    mov         rsi, rbx
+    %endif
+
+    call        EXTN(jsimd_h2v1_merged_upsample_avx2)
+
+    pop         rax
+    pop         rcx
+    pop         rdi
+    mov         rsip, JSAMPARRAY [rsp+0*SIZEOF_JSAMPARRAY]
+    mov         rbxp, JSAMPARRAY [rsp+1*SIZEOF_JSAMPARRAY]
+    mov         rdxp, JSAMPARRAY [rsp+2*SIZEOF_JSAMPARRAY]
+    add         rsp, SIZEOF_JSAMPARRAY*4
+
+    pop         rbx
+    uncollect_args 4
+    pop         rbp
+    ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+    align       32
diff --git a/media/libjpeg/simd/x86_64/jdmrgext-sse2.asm b/media/libjpeg/simd/x86_64/jdmrgext-sse2.asm
new file mode 100644
index 0000000000..eb3ab9dbd9
--- /dev/null
+++ b/media/libjpeg/simd/x86_64/jdmrgext-sse2.asm
@@ -0,0 +1,538 @@
+;
+; jdmrgext.asm - merged upsampling/color conversion (64-bit SSE2)
+;
+; Copyright 2009, 2012 Pierre Ossman <ossman@cendio.se> for Cendio AB
+; Copyright (C) 2009, 2012, 2016, D. R. Commander.
+; Copyright (C) 2018, Matthias Räncker.
+;
+; Based on the x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+
+%include "jcolsamp.inc"
+
+; --------------------------------------------------------------------------
+;
+; Upsample and color convert for the case of 2:1 horizontal and 1:1 vertical.
+;
+; GLOBAL(void)
+; jsimd_h2v1_merged_upsample_sse2(JDIMENSION output_width,
+;                                 JSAMPIMAGE input_buf,
+;                                 JDIMENSION in_row_group_ctr,
+;                                 JSAMPARRAY output_buf);
+;
+
+; r10d = JDIMENSION output_width
+; r11 = JSAMPIMAGE input_buf
+; r12d = JDIMENSION in_row_group_ctr
+; r13 = JSAMPARRAY output_buf
+
+%define wk(i)   rbp - (WK_NUM - (i)) * SIZEOF_XMMWORD  ; xmmword wk[WK_NUM]
+%define WK_NUM  3
+
+    align       32
+    GLOBAL_FUNCTION(jsimd_h2v1_merged_upsample_sse2)
+
+EXTN(jsimd_h2v1_merged_upsample_sse2):
+    push        rbp
+    mov         rax, rsp                     ; rax = original rbp
+    sub         rsp, byte 4
+    and         rsp, byte (-SIZEOF_XMMWORD)  ; align to 128 bits
+    mov         [rsp], rax
+    mov         rbp, rsp                     ; rbp = aligned rbp
+    lea         rsp, [wk(0)]
+    collect_args 4
+    push        rbx
+
+    mov         ecx, r10d               ; col
+    test        rcx, rcx
+    jz          near .return
+
+    push        rcx
+
+    mov         rdi, r11
+    mov         ecx, r12d
+    mov         rsip, JSAMPARRAY [rdi+0*SIZEOF_JSAMPARRAY]
+    mov         rbxp, JSAMPARRAY [rdi+1*SIZEOF_JSAMPARRAY]
+    mov         rdxp, JSAMPARRAY [rdi+2*SIZEOF_JSAMPARRAY]
+    mov         rdi, r13
+    mov         rsip, JSAMPROW [rsi+rcx*SIZEOF_JSAMPROW]  ; inptr0
+    mov         rbxp, JSAMPROW [rbx+rcx*SIZEOF_JSAMPROW]  ; inptr1
+    mov         rdxp, JSAMPROW [rdx+rcx*SIZEOF_JSAMPROW]  ; inptr2
+    mov         rdip, JSAMPROW [rdi]                      ; outptr
+
+    pop         rcx                     ; col
+
+.columnloop:
+
+    movdqa      xmm6, XMMWORD [rbx]     ; xmm6=Cb(0123456789ABCDEF)
+    movdqa      xmm7, XMMWORD [rdx]     ; xmm7=Cr(0123456789ABCDEF)
+
+    pxor        xmm1, xmm1              ; xmm1=(all 0's)
+    pcmpeqw     xmm3, xmm3
+    psllw       xmm3, 7                 ; xmm3={0xFF80 0xFF80 0xFF80 0xFF80 ..}
+
+    movdqa      xmm4, xmm6
+    punpckhbw   xmm6, xmm1              ; xmm6=Cb(89ABCDEF)=CbH
+    punpcklbw   xmm4, xmm1              ; xmm4=Cb(01234567)=CbL
+    movdqa      xmm0, xmm7
+    punpckhbw   xmm7, xmm1              ; xmm7=Cr(89ABCDEF)=CrH
+    punpcklbw   xmm0, xmm1              ; xmm0=Cr(01234567)=CrL
+
+    paddw       xmm6, xmm3
+    paddw       xmm4, xmm3
+    paddw       xmm7, xmm3
+    paddw       xmm0, xmm3
+
+    ; (Original)
+    ; R = Y                + 1.40200 * Cr
+    ; G = Y - 0.34414 * Cb - 0.71414 * Cr
+    ; B = Y + 1.77200 * Cb
+    ;
+    ; (This implementation)
+    ; R = Y                + 0.40200 * Cr + Cr
+    ; G = Y - 0.34414 * Cb + 0.28586 * Cr - Cr
+    ; B = Y - 0.22800 * Cb + Cb + Cb
+
+    movdqa      xmm5, xmm6              ; xmm5=CbH
+    movdqa      xmm2, xmm4              ; xmm2=CbL
+    paddw       xmm6, xmm6              ; xmm6=2*CbH
+    paddw       xmm4, xmm4              ; xmm4=2*CbL
+    movdqa      xmm1, xmm7              ; xmm1=CrH
+    movdqa      xmm3, xmm0              ; xmm3=CrL
+    paddw       xmm7, xmm7              ; xmm7=2*CrH
+    paddw       xmm0, xmm0              ; xmm0=2*CrL
+
+    pmulhw      xmm6, [rel PW_MF0228]   ; xmm6=(2*CbH * -FIX(0.22800))
+    pmulhw      xmm4, [rel PW_MF0228]   ; xmm4=(2*CbL * -FIX(0.22800))
+    pmulhw      xmm7, [rel PW_F0402]    ; xmm7=(2*CrH * FIX(0.40200))
+    pmulhw      xmm0, [rel PW_F0402]    ; xmm0=(2*CrL * FIX(0.40200))
+
+    paddw       xmm6, [rel PW_ONE]
+    paddw       xmm4, [rel PW_ONE]
+    psraw       xmm6, 1                 ; xmm6=(CbH * -FIX(0.22800))
+    psraw       xmm4, 1                 ; xmm4=(CbL * -FIX(0.22800))
+    paddw       xmm7, [rel PW_ONE]
+    paddw       xmm0, [rel PW_ONE]
+    psraw       xmm7, 1                 ; xmm7=(CrH * FIX(0.40200))
+    psraw       xmm0, 1                 ; xmm0=(CrL * FIX(0.40200))
+
+    paddw       xmm6, xmm5
+    paddw       xmm4, xmm2
+    paddw       xmm6, xmm5              ; xmm6=(CbH * FIX(1.77200))=(B-Y)H
+    paddw       xmm4, xmm2              ; xmm4=(CbL * FIX(1.77200))=(B-Y)L
+    paddw       xmm7, xmm1              ; xmm7=(CrH * FIX(1.40200))=(R-Y)H
+    paddw       xmm0, xmm3              ; xmm0=(CrL * FIX(1.40200))=(R-Y)L
+
+    movdqa      XMMWORD [wk(0)], xmm6   ; wk(0)=(B-Y)H
+    movdqa      XMMWORD [wk(1)], xmm7   ; wk(1)=(R-Y)H
+
+    movdqa      xmm6, xmm5
+    movdqa      xmm7, xmm2
+    punpcklwd   xmm5, xmm1
+    punpckhwd   xmm6, xmm1
+    pmaddwd     xmm5, [rel PW_MF0344_F0285]
+    pmaddwd     xmm6, [rel PW_MF0344_F0285]
+    punpcklwd   xmm2, xmm3
+    punpckhwd   xmm7, xmm3
+    pmaddwd     xmm2, [rel PW_MF0344_F0285]
+    pmaddwd     xmm7, [rel PW_MF0344_F0285]
+
+    paddd       xmm5, [rel PD_ONEHALF]
+    paddd       xmm6, [rel PD_ONEHALF]
+    psrad       xmm5, SCALEBITS
+    psrad       xmm6, SCALEBITS
+    paddd       xmm2, [rel PD_ONEHALF]
+    paddd       xmm7, [rel PD_ONEHALF]
+    psrad       xmm2, SCALEBITS
+    psrad       xmm7, SCALEBITS
+
+    packssdw    xmm5, xmm6              ; xmm5=CbH*-FIX(0.344)+CrH*FIX(0.285)
+    packssdw    xmm2, xmm7              ; xmm2=CbL*-FIX(0.344)+CrL*FIX(0.285)
+    psubw       xmm5, xmm1              ; xmm5=CbH*-FIX(0.344)+CrH*-FIX(0.714)=(G-Y)H
+    psubw       xmm2, xmm3              ; xmm2=CbL*-FIX(0.344)+CrL*-FIX(0.714)=(G-Y)L
+
+    movdqa      XMMWORD [wk(2)], xmm5   ; wk(2)=(G-Y)H
+
+    mov         al, 2                   ; Yctr
+    jmp         short .Yloop_1st
+
+.Yloop_2nd:
+    movdqa      xmm0, XMMWORD [wk(1)]   ; xmm0=(R-Y)H
+    movdqa      xmm2, XMMWORD [wk(2)]   ; xmm2=(G-Y)H
+    movdqa      xmm4, XMMWORD [wk(0)]   ; xmm4=(B-Y)H
+
+.Yloop_1st:
+    movdqa      xmm7, XMMWORD [rsi]     ; xmm7=Y(0123456789ABCDEF)
+
+    pcmpeqw     xmm6, xmm6
+    psrlw       xmm6, BYTE_BIT          ; xmm6={0xFF 0x00 0xFF 0x00 ..}
+    pand        xmm6, xmm7              ; xmm6=Y(02468ACE)=YE
+    psrlw       xmm7, BYTE_BIT          ; xmm7=Y(13579BDF)=YO
+
+    movdqa      xmm1, xmm0              ; xmm1=xmm0=(R-Y)(L/H)
+    movdqa      xmm3, xmm2              ; xmm3=xmm2=(G-Y)(L/H)
+    movdqa      xmm5, xmm4              ; xmm5=xmm4=(B-Y)(L/H)
+
+    paddw       xmm0, xmm6              ; xmm0=((R-Y)+YE)=RE=R(02468ACE)
+    paddw       xmm1, xmm7              ; xmm1=((R-Y)+YO)=RO=R(13579BDF)
+    packuswb    xmm0, xmm0              ; xmm0=R(02468ACE********)
+    packuswb    xmm1, xmm1              ; xmm1=R(13579BDF********)
+
+    paddw       xmm2, xmm6              ; xmm2=((G-Y)+YE)=GE=G(02468ACE)
+    paddw       xmm3, xmm7              ; xmm3=((G-Y)+YO)=GO=G(13579BDF)
+    packuswb    xmm2, xmm2              ; xmm2=G(02468ACE********)
+    packuswb    xmm3, xmm3              ; xmm3=G(13579BDF********)
+
+    paddw       xmm4, xmm6              ; xmm4=((B-Y)+YE)=BE=B(02468ACE)
+    paddw       xmm5, xmm7              ; xmm5=((B-Y)+YO)=BO=B(13579BDF)
+    packuswb    xmm4, xmm4              ; xmm4=B(02468ACE********)
+    packuswb    xmm5, xmm5              ; xmm5=B(13579BDF********)
+
+%if RGB_PIXELSIZE == 3  ; ---------------
+
+    ; xmmA=(00 02 04 06 08 0A 0C 0E **), xmmB=(01 03 05 07 09 0B 0D 0F **)
+    ; xmmC=(10 12 14 16 18 1A 1C 1E **), xmmD=(11 13 15 17 19 1B 1D 1F **)
+    ; xmmE=(20 22 24 26 28 2A 2C 2E **), xmmF=(21 23 25 27 29 2B 2D 2F **)
+    ; xmmG=(** ** ** ** ** ** ** ** **), xmmH=(** ** ** ** ** ** ** ** **)
+
+    punpcklbw   xmmA, xmmC        ; xmmA=(00 10 02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E)
+    punpcklbw   xmmE, xmmB        ; xmmE=(20 01 22 03 24 05 26 07 28 09 2A 0B 2C 0D 2E 0F)
+    punpcklbw   xmmD, xmmF        ; xmmD=(11 21 13 23 15 25 17 27 19 29 1B 2B 1D 2D 1F 2F)
+
+    movdqa      xmmG, xmmA
+    movdqa      xmmH, xmmA
+    punpcklwd   xmmA, xmmE        ; xmmA=(00 10 20 01 02 12 22 03 04 14 24 05 06 16 26 07)
+    punpckhwd   xmmG, xmmE        ; xmmG=(08 18 28 09 0A 1A 2A 0B 0C 1C 2C 0D 0E 1E 2E 0F)
+
+    psrldq      xmmH, 2           ; xmmH=(02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E -- --)
+    psrldq      xmmE, 2           ; xmmE=(22 03 24 05 26 07 28 09 2A 0B 2C 0D 2E 0F -- --)
+
+    movdqa      xmmC, xmmD
+    movdqa      xmmB, xmmD
+    punpcklwd   xmmD, xmmH        ; xmmD=(11 21 02 12 13 23 04 14 15 25 06 16 17 27 08 18)
+    punpckhwd   xmmC, xmmH        ; xmmC=(19 29 0A 1A 1B 2B 0C 1C 1D 2D 0E 1E 1F 2F -- --)
+
+    psrldq      xmmB, 2           ; xmmB=(13 23 15 25 17 27 19 29 1B 2B 1D 2D 1F 2F -- --)
+
+    movdqa      xmmF, xmmE
+    punpcklwd   xmmE, xmmB        ; xmmE=(22 03 13 23 24 05 15 25 26 07 17 27 28 09 19 29)
+    punpckhwd   xmmF, xmmB        ; xmmF=(2A 0B 1B 2B 2C 0D 1D 2D 2E 0F 1F 2F -- -- -- --)
+
+    pshufd      xmmH, xmmA, 0x4E  ; xmmH=(04 14 24 05 06 16 26 07 00 10 20 01 02 12 22 03)
+    movdqa      xmmB, xmmE
+    punpckldq   xmmA, xmmD        ; xmmA=(00 10 20 01 11 21 02 12 02 12 22 03 13 23 04 14)
+    punpckldq   xmmE, xmmH        ; xmmE=(22 03 13 23 04 14 24 05 24 05 15 25 06 16 26 07)
+    punpckhdq   xmmD, xmmB        ; xmmD=(15 25 06 16 26 07 17 27 17 27 08 18 28 09 19 29)
+
+    pshufd      xmmH, xmmG, 0x4E  ; xmmH=(0C 1C 2C 0D 0E 1E 2E 0F 08 18 28 09 0A 1A 2A 0B)
+    movdqa      xmmB, xmmF
+    punpckldq   xmmG, xmmC        ; xmmG=(08 18 28 09 19 29 0A 1A 0A 1A 2A 0B 1B 2B 0C 1C)
+    punpckldq   xmmF, xmmH        ; xmmF=(2A 0B 1B 2B 0C 1C 2C 0D 2C 0D 1D 2D 0E 1E 2E 0F)
+    punpckhdq   xmmC, xmmB        ; xmmC=(1D 2D 0E 1E 2E 0F 1F 2F 1F 2F -- -- -- -- -- --)
+
+    punpcklqdq  xmmA, xmmE        ; xmmA=(00 10 20 01 11 21 02 12 22 03 13 23 04 14 24 05)
+    punpcklqdq  xmmD, xmmG        ; xmmD=(15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A)
+    punpcklqdq  xmmF, xmmC        ; xmmF=(2A 0B 1B 2B 0C 1C 2C 0D 1D 2D 0E 1E 2E 0F 1F 2F)
+
+    cmp         rcx, byte SIZEOF_XMMWORD
+    jb          short .column_st32
+
+    test        rdi, SIZEOF_XMMWORD-1
+    jnz         short .out1
+    ; --(aligned)-------------------
+    movntdq     XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
+    movntdq     XMMWORD [rdi+1*SIZEOF_XMMWORD], xmmD
+    movntdq     XMMWORD [rdi+2*SIZEOF_XMMWORD], xmmF
+    jmp         short .out0
+.out1:  ; --(unaligned)-----------------
+    movdqu      XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
+    movdqu      XMMWORD [rdi+1*SIZEOF_XMMWORD], xmmD
+    movdqu      XMMWORD [rdi+2*SIZEOF_XMMWORD], xmmF
+.out0:
+    add         rdi, byte RGB_PIXELSIZE*SIZEOF_XMMWORD  ; outptr
+    sub         rcx, byte SIZEOF_XMMWORD
+    jz          near .endcolumn
+
+    add         rsi, byte SIZEOF_XMMWORD  ; inptr0
+    dec         al                        ; Yctr
+    jnz         near .Yloop_2nd
+
+    add         rbx, byte SIZEOF_XMMWORD  ; inptr1
+    add         rdx, byte SIZEOF_XMMWORD  ; inptr2
+    jmp         near .columnloop
+
+.column_st32:
+    lea         rcx, [rcx+rcx*2]            ; imul ecx, RGB_PIXELSIZE
+    cmp         rcx, byte 2*SIZEOF_XMMWORD
+    jb          short .column_st16
+    movdqu      XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
+    movdqu      XMMWORD [rdi+1*SIZEOF_XMMWORD], xmmD
+    add         rdi, byte 2*SIZEOF_XMMWORD  ; outptr
+    movdqa      xmmA, xmmF
+    sub         rcx, byte 2*SIZEOF_XMMWORD
+    jmp         short .column_st15
+.column_st16:
+    cmp         rcx, byte SIZEOF_XMMWORD
+    jb          short .column_st15
+    movdqu      XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
+    add         rdi, byte SIZEOF_XMMWORD    ; outptr
+    movdqa      xmmA, xmmD
+    sub         rcx, byte SIZEOF_XMMWORD
+.column_st15:
+    ; Store the lower 8 bytes of xmmA to the output when it has enough
+    ; space.
+    cmp         rcx, byte SIZEOF_MMWORD
+    jb          short .column_st7
+    movq        XMM_MMWORD [rdi], xmmA
+    add         rdi, byte SIZEOF_MMWORD
+    sub         rcx, byte SIZEOF_MMWORD
+    psrldq      xmmA, SIZEOF_MMWORD
+.column_st7:
+    ; Store the lower 4 bytes of xmmA to the output when it has enough
+    ; space.
+    cmp         rcx, byte SIZEOF_DWORD
+    jb          short .column_st3
+    movd        XMM_DWORD [rdi], xmmA
+    add         rdi, byte SIZEOF_DWORD
+    sub         rcx, byte SIZEOF_DWORD
+    psrldq      xmmA, SIZEOF_DWORD
+.column_st3:
+    ; Store the lower 2 bytes of rax to the output when it has enough
+    ; space.
+    movd        eax, xmmA
+    cmp         rcx, byte SIZEOF_WORD
+    jb          short .column_st1
+    mov         word [rdi], ax
+    add         rdi, byte SIZEOF_WORD
+    sub         rcx, byte SIZEOF_WORD
+    shr         rax, 16
+.column_st1:
+    ; Store the lower 1 byte of rax to the output when it has enough
+    ; space.
+    test        rcx, rcx
+    jz          short .endcolumn
+    mov         byte [rdi], al
+
+%else  ; RGB_PIXELSIZE == 4 ; -----------
+
+%ifdef RGBX_FILLER_0XFF
+    pcmpeqb     xmm6, xmm6              ; xmm6=XE=X(02468ACE********)
+    pcmpeqb     xmm7, xmm7              ; xmm7=XO=X(13579BDF********)
+%else
+    pxor        xmm6, xmm6              ; xmm6=XE=X(02468ACE********)
+    pxor        xmm7, xmm7              ; xmm7=XO=X(13579BDF********)
+%endif
+    ; xmmA=(00 02 04 06 08 0A 0C 0E **), xmmB=(01 03 05 07 09 0B 0D 0F **)
+    ; xmmC=(10 12 14 16 18 1A 1C 1E **), xmmD=(11 13 15 17 19 1B 1D 1F **)
+    ; xmmE=(20 22 24 26 28 2A 2C 2E **), xmmF=(21 23 25 27 29 2B 2D 2F **)
+    ; xmmG=(30 32 34 36 38 3A 3C 3E **), xmmH=(31 33 35 37 39 3B 3D 3F **)
+
+    punpcklbw   xmmA, xmmC  ; xmmA=(00 10 02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E)
+    punpcklbw   xmmE, xmmG  ; xmmE=(20 30 22 32 24 34 26 36 28 38 2A 3A 2C 3C 2E 3E)
+    punpcklbw   xmmB, xmmD  ; xmmB=(01 11 03 13 05 15 07 17 09 19 0B 1B 0D 1D 0F 1F)
+    punpcklbw   xmmF, xmmH  ; xmmF=(21 31 23 33 25 35 27 37 29 39 2B 3B 2D 3D 2F 3F)
+
+    movdqa      xmmC, xmmA
+    punpcklwd   xmmA, xmmE  ; xmmA=(00 10 20 30 02 12 22 32 04 14 24 34 06 16 26 36)
+    punpckhwd   xmmC, xmmE  ; xmmC=(08 18 28 38 0A 1A 2A 3A 0C 1C 2C 3C 0E 1E 2E 3E)
+    movdqa      xmmG, xmmB
+    punpcklwd   xmmB, xmmF  ; xmmB=(01 11 21 31 03 13 23 33 05 15 25 35 07 17 27 37)
+    punpckhwd   xmmG, xmmF  ; xmmG=(09 19 29 39 0B 1B 2B 3B 0D 1D 2D 3D 0F 1F 2F 3F)
+
+    movdqa      xmmD, xmmA
+    punpckldq   xmmA, xmmB  ; xmmA=(00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33)
+    punpckhdq   xmmD, xmmB  ; xmmD=(04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37)
+    movdqa      xmmH, xmmC
+    punpckldq   xmmC, xmmG  ; xmmC=(08 18 28 38 09 19 29 39 0A 1A 2A 3A 0B 1B 2B 3B)
+    punpckhdq   xmmH, xmmG  ; xmmH=(0C 1C 2C 3C 0D 1D 2D 3D 0E 1E 2E 3E 0F 1F 2F 3F)
+
+    cmp         rcx, byte SIZEOF_XMMWORD
+    jb          short .column_st32
+
+    test        rdi, SIZEOF_XMMWORD-1
+    jnz         short .out1
+    ; --(aligned)-------------------
+    movntdq     XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
+    movntdq     XMMWORD [rdi+1*SIZEOF_XMMWORD], xmmD
+    movntdq     XMMWORD [rdi+2*SIZEOF_XMMWORD], xmmC
+    movntdq     XMMWORD [rdi+3*SIZEOF_XMMWORD], xmmH
+    jmp         short .out0
+.out1:  ; --(unaligned)-----------------
+    movdqu      XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
+    movdqu      XMMWORD [rdi+1*SIZEOF_XMMWORD], xmmD
+    movdqu      XMMWORD [rdi+2*SIZEOF_XMMWORD], xmmC
+    movdqu      XMMWORD [rdi+3*SIZEOF_XMMWORD], xmmH
+.out0:
+    add         rdi, byte RGB_PIXELSIZE*SIZEOF_XMMWORD  ; outptr
+    sub         rcx, byte SIZEOF_XMMWORD
+    jz          near .endcolumn
+
+    add         rsi, byte SIZEOF_XMMWORD  ; inptr0
+    dec         al                        ; Yctr
+    jnz         near .Yloop_2nd
+
+    add         rbx, byte SIZEOF_XMMWORD  ; inptr1
+    add         rdx, byte SIZEOF_XMMWORD  ; inptr2
+    jmp         near .columnloop
+
+.column_st32:
+    cmp         rcx, byte SIZEOF_XMMWORD/2
+    jb          short .column_st16
+    movdqu      XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
+    movdqu      XMMWORD [rdi+1*SIZEOF_XMMWORD], xmmD
+    add         rdi, byte 2*SIZEOF_XMMWORD  ; outptr
+    movdqa      xmmA, xmmC
+    movdqa      xmmD, xmmH
+    sub         rcx, byte SIZEOF_XMMWORD/2
+.column_st16:
+    cmp         rcx, byte SIZEOF_XMMWORD/4
+    jb          short .column_st15
+    movdqu      XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
+    add         rdi, byte SIZEOF_XMMWORD    ; outptr
+    movdqa      xmmA, xmmD
+    sub         rcx, byte SIZEOF_XMMWORD/4
+.column_st15:
+    ; Store two pixels (8 bytes) of xmmA to the output when it has enough
+    ; space.
+    cmp         rcx, byte SIZEOF_XMMWORD/8
+    jb          short .column_st7
+    movq        XMM_MMWORD [rdi], xmmA
+    add         rdi, byte SIZEOF_XMMWORD/8*4
+    sub         rcx, byte SIZEOF_XMMWORD/8
+    psrldq      xmmA, SIZEOF_XMMWORD/8*4
+.column_st7:
+    ; Store one pixel (4 bytes) of xmmA to the output when it has enough
+    ; space.
+    test        rcx, rcx
+    jz          short .endcolumn
+    movd        XMM_DWORD [rdi], xmmA
+
+%endif  ; RGB_PIXELSIZE ; ---------------
+
+.endcolumn:
+    sfence                              ; flush the write buffer
+
+.return:
+    pop         rbx
+    uncollect_args 4
+    mov         rsp, rbp                ; rsp <- aligned rbp
+    pop         rsp                     ; rsp <- original rbp
+    pop         rbp
+    ret
+
+; --------------------------------------------------------------------------
+;
+; Upsample and color convert for the case of 2:1 horizontal and 2:1 vertical.
+;
+; GLOBAL(void)
+; jsimd_h2v2_merged_upsample_sse2(JDIMENSION output_width,
+;                                 JSAMPIMAGE input_buf,
+;                                 JDIMENSION in_row_group_ctr,
+;                                 JSAMPARRAY output_buf);
+;
+
+; r10d = JDIMENSION output_width
+; r11 = JSAMPIMAGE input_buf
+; r12d = JDIMENSION in_row_group_ctr
+; r13 = JSAMPARRAY output_buf
+
+    align       32
+    GLOBAL_FUNCTION(jsimd_h2v2_merged_upsample_sse2)
+
+EXTN(jsimd_h2v2_merged_upsample_sse2):
+    push        rbp
+    mov         rax, rsp
+    mov         rbp, rsp
+    collect_args 4
+    push        rbx
+
+    mov         eax, r10d
+
+    mov         rdi, r11
+    mov         ecx, r12d
+    mov         rsip, JSAMPARRAY [rdi+0*SIZEOF_JSAMPARRAY]
+    mov         rbxp, JSAMPARRAY [rdi+1*SIZEOF_JSAMPARRAY]
+    mov         rdxp, JSAMPARRAY [rdi+2*SIZEOF_JSAMPARRAY]
+    mov         rdi, r13
+    lea         rsi, [rsi+rcx*SIZEOF_JSAMPROW]
+
+    sub         rsp, SIZEOF_JSAMPARRAY*4
+    mov         JSAMPARRAY [rsp+0*SIZEOF_JSAMPARRAY], rsip  ; intpr00
+    mov         JSAMPARRAY [rsp+1*SIZEOF_JSAMPARRAY], rbxp  ; intpr1
+    mov         JSAMPARRAY [rsp+2*SIZEOF_JSAMPARRAY], rdxp  ; intpr2
+    mov         rbx, rsp
+
+    push        rdi
+    push        rcx
+    push        rax
+
+    %ifdef WIN64
+    mov         r8, rcx
+    mov         r9, rdi
+    mov         rcx, rax
+    mov         rdx, rbx
+    %else
+    mov         rdx, rcx
+    mov         rcx, rdi
+    mov         rdi, rax
+    mov         rsi, rbx
+    %endif
+
+    call        EXTN(jsimd_h2v1_merged_upsample_sse2)
+
+    pop         rax
+    pop         rcx
+    pop         rdi
+    mov         rsip, JSAMPARRAY [rsp+0*SIZEOF_JSAMPARRAY]
+    mov         rbxp, JSAMPARRAY [rsp+1*SIZEOF_JSAMPARRAY]
+    mov         rdxp, JSAMPARRAY [rsp+2*SIZEOF_JSAMPARRAY]
+
+    add         rdi, byte SIZEOF_JSAMPROW  ; outptr1
+    add         rsi, byte SIZEOF_JSAMPROW  ; inptr01
+
+    mov         JSAMPARRAY [rsp+0*SIZEOF_JSAMPARRAY], rsip  ; intpr00
+    mov         JSAMPARRAY [rsp+1*SIZEOF_JSAMPARRAY], rbxp  ; intpr1
+    mov         JSAMPARRAY [rsp+2*SIZEOF_JSAMPARRAY], rdxp  ; intpr2
+    mov         rbx, rsp
+
+    push        rdi
+    push        rcx
+    push        rax
+
+    %ifdef WIN64
+    mov         r8, rcx
+    mov         r9, rdi
+    mov         rcx, rax
+    mov         rdx, rbx
+    %else
+    mov         rdx, rcx
+    mov         rcx, rdi
+    mov         rdi, rax
+    mov         rsi, rbx
+    %endif
+
+    call        EXTN(jsimd_h2v1_merged_upsample_sse2)
+
+    pop         rax
+    pop         rcx
+    pop         rdi
+    mov         rsip, JSAMPARRAY [rsp+0*SIZEOF_JSAMPARRAY]
+    mov         rbxp, JSAMPARRAY [rsp+1*SIZEOF_JSAMPARRAY]
+    mov         rdxp, JSAMPARRAY [rsp+2*SIZEOF_JSAMPARRAY]
+    add         rsp, SIZEOF_JSAMPARRAY*4
+
+    pop         rbx
+    uncollect_args 4
+    pop         rbp
+    ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+    align       32
diff --git a/media/libjpeg/simd/x86_64/jdsample-avx2.asm b/media/libjpeg/simd/x86_64/jdsample-avx2.asm
new file mode 100644
index 0000000000..1e4979f933
--- /dev/null
+++ b/media/libjpeg/simd/x86_64/jdsample-avx2.asm
@@ -0,0 +1,696 @@
+;
+; jdsample.asm - upsampling (64-bit AVX2)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+; Copyright (C) 2009, 2016, D. R. Commander.
+; Copyright (C) 2015, Intel Corporation.
+; Copyright (C) 2018, Matthias Räncker.
+;
+; Based on the x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+
+%include "jsimdext.inc"
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_CONST
+
+    alignz      32
+    GLOBAL_DATA(jconst_fancy_upsample_avx2)
+
+EXTN(jconst_fancy_upsample_avx2):
+
+PW_ONE   times 16 dw 1
+PW_TWO   times 16 dw 2
+PW_THREE times 16 dw 3
+PW_SEVEN times 16 dw 7
+PW_EIGHT times 16 dw 8
+
+    alignz      32
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_TEXT
+    BITS        64
+;
+; Fancy processing for the common case of 2:1 horizontal and 1:1 vertical.
+;
+; The upsampling algorithm is linear interpolation between pixel centers,
+; also known as a "triangle filter".  This is a good compromise between
+; speed and visual quality.  The centers of the output pixels are 1/4 and 3/4
+; of the way between input pixel centers.
+;
+; GLOBAL(void)
+; jsimd_h2v1_fancy_upsample_avx2(int max_v_samp_factor,
+;                                JDIMENSION downsampled_width,
+;                                JSAMPARRAY input_data,
+;                                JSAMPARRAY *output_data_ptr);
+;
+
+; r10 = int max_v_samp_factor
+; r11d = JDIMENSION downsampled_width
+; r12 = JSAMPARRAY input_data
+; r13 = JSAMPARRAY *output_data_ptr
+
+    align       32
+    GLOBAL_FUNCTION(jsimd_h2v1_fancy_upsample_avx2)
+
+EXTN(jsimd_h2v1_fancy_upsample_avx2):
+    push        rbp
+    mov         rax, rsp
+    mov         rbp, rsp
+    push_xmm    3
+    collect_args 4
+
+    mov         eax, r11d               ; colctr
+    test        rax, rax
+    jz          near .return
+
+    mov         rcx, r10                ; rowctr
+    test        rcx, rcx
+    jz          near .return
+
+    mov         rsi, r12                ; input_data
+    mov         rdi, r13
+    mov         rdip, JSAMPARRAY [rdi]  ; output_data
+
+    vpxor       ymm0, ymm0, ymm0                 ; ymm0=(all 0's)
+    vpcmpeqb    xmm9, xmm9, xmm9
+    vpsrldq     xmm10, xmm9, (SIZEOF_XMMWORD-1)  ; (ff -- -- -- ... -- --) LSB is ff
+
+    vpslldq     xmm9, xmm9, (SIZEOF_XMMWORD-1)
+    vperm2i128  ymm9, ymm9, ymm9, 1              ; (---- ---- ... ---- ---- ff) MSB is ff
+
+.rowloop:
+    push        rax                     ; colctr
+    push        rdi
+    push        rsi
+
+    mov         rsip, JSAMPROW [rsi]    ; inptr
+    mov         rdip, JSAMPROW [rdi]    ; outptr
+
+    test        rax, SIZEOF_YMMWORD-1
+    jz          short .skip
+    mov         dl, JSAMPLE [rsi+(rax-1)*SIZEOF_JSAMPLE]
+    mov         JSAMPLE [rsi+rax*SIZEOF_JSAMPLE], dl    ; insert a dummy sample
+.skip:
+    vpand       ymm7, ymm10, YMMWORD [rsi+0*SIZEOF_YMMWORD]
+
+    add         rax, byte SIZEOF_YMMWORD-1
+    and         rax, byte -SIZEOF_YMMWORD
+    cmp         rax, byte SIZEOF_YMMWORD
+    ja          short .columnloop
+
+.columnloop_last:
+    vpand       ymm6, ymm9, YMMWORD [rsi+0*SIZEOF_YMMWORD]
+    jmp         short .upsample
+
+.columnloop:
+    vmovdqu     ymm6, YMMWORD [rsi+1*SIZEOF_YMMWORD]
+    vperm2i128  ymm6, ymm0, ymm6, 0x20
+    vpslldq     ymm6, ymm6, 15
+
+.upsample:
+    vmovdqu     ymm1, YMMWORD [rsi+0*SIZEOF_YMMWORD]  ; ymm1=( 0  1  2 ... 29 30 31)
+
+    vperm2i128  ymm2, ymm0, ymm1, 0x20
+    vpalignr    ymm2, ymm1, ymm2, 15            ; ymm2=(--  0  1 ... 28 29 30)
+    vperm2i128  ymm4, ymm0, ymm1, 0x03
+    vpalignr    ymm3, ymm4, ymm1, 1             ; ymm3=( 1  2  3 ... 30 31 --)
+
+    vpor        ymm2, ymm2, ymm7                ; ymm2=(-1  0  1 ... 28 29 30)
+    vpor        ymm3, ymm3, ymm6                ; ymm3=( 1  2  3 ... 30 31 32)
+
+    vpsrldq     ymm7, ymm4, (SIZEOF_XMMWORD-1)  ; ymm7=(31 -- -- ... -- -- --)
+
+    vpunpckhbw  ymm4, ymm1, ymm0                ; ymm4=( 8  9 10 11 12 13 14 15 24 25 26 27 28 29 30 31)
+    vpunpcklbw  ymm5, ymm1, ymm0                ; ymm5=( 0  1  2  3  4  5  6  7 16 17 18 19 20 21 22 23)
+    vperm2i128  ymm1, ymm5, ymm4, 0x20          ; ymm1=( 0  1  2  3  4  5  6  7  8  9 10 11 12 13 14 15)
+    vperm2i128  ymm4, ymm5, ymm4, 0x31          ; ymm4=(16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31)
+
+    vpunpckhbw  ymm5, ymm2, ymm0                ; ymm5=( 7  8  9 10 11 12 13 14 23 24 25 26 27 28 29 30)
+    vpunpcklbw  ymm6, ymm2, ymm0                ; ymm6=(-1  0  1  2  3  4  5  6 15 16 17 18 19 20 21 22)
+    vperm2i128  ymm2, ymm6, ymm5, 0x20          ; ymm2=(-1  0  1  2  3  4  5  6  7  8  9 10 11 12 13 14)
+    vperm2i128  ymm5, ymm6, ymm5, 0x31          ; ymm5=(15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30)
+
+    vpunpckhbw  ymm6, ymm3, ymm0                ; ymm6=( 1  2  3  4  5  6  7  8 17 18 19 20 21 22 23 24)
+    vpunpcklbw  ymm8, ymm3, ymm0                ; ymm8=( 9 10 11 12 13 14 15 16 25 26 27 28 29 30 31 32)
+    vperm2i128  ymm3, ymm8, ymm6, 0x20          ; ymm3=( 1  2  3  4  5  6  7  8  9 10 11 12 13 14 15 16)
+    vperm2i128  ymm6, ymm8, ymm6, 0x31          ; ymm6=(17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32)
+
+    vpmullw     ymm1, ymm1, [rel PW_THREE]
+    vpmullw     ymm4, ymm4, [rel PW_THREE]
+    vpaddw      ymm2, ymm2, [rel PW_ONE]
+    vpaddw      ymm5, ymm5, [rel PW_ONE]
+    vpaddw      ymm3, ymm3, [rel PW_TWO]
+    vpaddw      ymm6, ymm6, [rel PW_TWO]
+
+    vpaddw      ymm2, ymm2, ymm1
+    vpaddw      ymm5, ymm5, ymm4
+    vpsrlw      ymm2, ymm2, 2                   ; ymm2=OutLE=( 0  2  4  6  8 10 12 14 16 18 20 22 24 26 28 30)
+    vpsrlw      ymm5, ymm5, 2                   ; ymm5=OutHE=(32 34 36 38 40 42 44 46 48 50 52 54 56 58 60 62)
+    vpaddw      ymm3, ymm3, ymm1
+    vpaddw      ymm6, ymm6, ymm4
+    vpsrlw      ymm3, ymm3, 2                   ; ymm3=OutLO=( 1  3  5  7  9 11 13 15 17 19 21 23 25 27 29 31)
+    vpsrlw      ymm6, ymm6, 2                   ; ymm6=OutHO=(33 35 37 39 41 43 45 47 49 51 53 55 57 59 61 63)
+
+    vpsllw      ymm3, ymm3, BYTE_BIT
+    vpsllw      ymm6, ymm6, BYTE_BIT
+    vpor        ymm2, ymm2, ymm3                ; ymm2=OutL=( 0  1  2 ... 29 30 31)
+    vpor        ymm5, ymm5, ymm6                ; ymm5=OutH=(32 33 34 ... 61 62 63)
+
+    vmovdqu     YMMWORD [rdi+0*SIZEOF_YMMWORD], ymm2
+    vmovdqu     YMMWORD [rdi+1*SIZEOF_YMMWORD], ymm5
+
+    sub         rax, byte SIZEOF_YMMWORD
+    add         rsi, byte 1*SIZEOF_YMMWORD  ; inptr
+    add         rdi, byte 2*SIZEOF_YMMWORD  ; outptr
+    cmp         rax, byte SIZEOF_YMMWORD
+    ja          near .columnloop
+    test        eax, eax
+    jnz         near .columnloop_last
+
+    pop         rsi
+    pop         rdi
+    pop         rax
+
+    add         rsi, byte SIZEOF_JSAMPROW  ; input_data
+    add         rdi, byte SIZEOF_JSAMPROW  ; output_data
+    dec         rcx                        ; rowctr
+    jg          near .rowloop
+
+.return:
+    vzeroupper
+    uncollect_args 4
+    pop_xmm     3
+    pop         rbp
+    ret
+
+; --------------------------------------------------------------------------
+;
+; Fancy processing for the common case of 2:1 horizontal and 2:1 vertical.
+; Again a triangle filter; see comments for h2v1 case, above.
+;
+; GLOBAL(void)
+; jsimd_h2v2_fancy_upsample_avx2(int max_v_samp_factor,
+;                                JDIMENSION downsampled_width,
+;                                JSAMPARRAY input_data,
+;                                JSAMPARRAY *output_data_ptr);
+;
+
+; r10 = int max_v_samp_factor
+; r11d = JDIMENSION downsampled_width
+; r12 = JSAMPARRAY input_data
+; r13 = JSAMPARRAY *output_data_ptr
+
+%define wk(i)   rbp - (WK_NUM - (i)) * SIZEOF_YMMWORD  ; ymmword wk[WK_NUM]
+%define WK_NUM  4
+
+    align       32
+    GLOBAL_FUNCTION(jsimd_h2v2_fancy_upsample_avx2)
+
+EXTN(jsimd_h2v2_fancy_upsample_avx2):
+    push        rbp
+    mov         rax, rsp                     ; rax = original rbp
+    sub         rsp, byte 4
+    and         rsp, byte (-SIZEOF_YMMWORD)  ; align to 256 bits
+    mov         [rsp], rax
+    mov         rbp, rsp                     ; rbp = aligned rbp
+    lea         rsp, [wk(0)]
+    push_xmm    3
+    collect_args 4
+    push        rbx
+
+    mov         eax, r11d               ; colctr
+    test        rax, rax
+    jz          near .return
+
+    mov         rcx, r10                ; rowctr
+    test        rcx, rcx
+    jz          near .return
+
+    mov         rsi, r12                ; input_data
+    mov         rdi, r13
+    mov         rdip, JSAMPARRAY [rdi]  ; output_data
+.rowloop:
+    push        rax                     ; colctr
+    push        rcx
+    push        rdi
+    push        rsi
+
+    mov         rcxp, JSAMPROW [rsi-1*SIZEOF_JSAMPROW]  ; inptr1(above)
+    mov         rbxp, JSAMPROW [rsi+0*SIZEOF_JSAMPROW]  ; inptr0
+    mov         rsip, JSAMPROW [rsi+1*SIZEOF_JSAMPROW]  ; inptr1(below)
+    mov         rdxp, JSAMPROW [rdi+0*SIZEOF_JSAMPROW]  ; outptr0
+    mov         rdip, JSAMPROW [rdi+1*SIZEOF_JSAMPROW]  ; outptr1
+
+    vpxor       ymm8, ymm8, ymm8                 ; ymm8=(all 0's)
+    vpcmpeqb    xmm9, xmm9, xmm9
+    vpsrldq     xmm10, xmm9, (SIZEOF_XMMWORD-2)  ; (ffff ---- ---- ... ---- ----) LSB is ffff
+    vpslldq     xmm9, xmm9, (SIZEOF_XMMWORD-2)
+    vperm2i128  ymm9, ymm9, ymm9, 1              ; (---- ---- ... ---- ---- ffff) MSB is ffff
+
+    test        rax, SIZEOF_YMMWORD-1
+    jz          short .skip
+    push        rdx
+    mov         dl, JSAMPLE [rcx+(rax-1)*SIZEOF_JSAMPLE]
+    mov         JSAMPLE [rcx+rax*SIZEOF_JSAMPLE], dl
+    mov         dl, JSAMPLE [rbx+(rax-1)*SIZEOF_JSAMPLE]
+    mov         JSAMPLE [rbx+rax*SIZEOF_JSAMPLE], dl
+    mov         dl, JSAMPLE [rsi+(rax-1)*SIZEOF_JSAMPLE]
+    mov         JSAMPLE [rsi+rax*SIZEOF_JSAMPLE], dl    ; insert a dummy sample
+    pop         rdx
+.skip:
+    ; -- process the first column block
+
+    vmovdqu     ymm0, YMMWORD [rbx+0*SIZEOF_YMMWORD]  ; ymm0=row[ 0][0]
+    vmovdqu     ymm1, YMMWORD [rcx+0*SIZEOF_YMMWORD]  ; ymm1=row[-1][0]
+    vmovdqu     ymm2, YMMWORD [rsi+0*SIZEOF_YMMWORD]  ; ymm2=row[+1][0]
+
+    vpunpckhbw  ymm4, ymm0, ymm8        ; ymm4=row[ 0]( 8  9 10 11 12 13 14 15 24 25 26 27 28 29 30 31)
+    vpunpcklbw  ymm5, ymm0, ymm8        ; ymm5=row[ 0]( 0  1  2  3  4  5  6  7 16 17 18 19 20 21 22 23)
+    vperm2i128  ymm0, ymm5, ymm4, 0x20  ; ymm0=row[ 0]( 0  1  2  3  4  5  6  7  8  9 10 11 12 13 14 15)
+    vperm2i128  ymm4, ymm5, ymm4, 0x31  ; ymm4=row[ 0](16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31)
+
+    vpunpckhbw  ymm5, ymm1, ymm8        ; ymm5=row[-1]( 8  9 10 11 12 13 14 15 24 25 26 27 28 29 30 31)
+    vpunpcklbw  ymm6, ymm1, ymm8        ; ymm6=row[-1]( 0  1  2  3  4  5  6  7 16 17 18 19 20 21 22 23)
+    vperm2i128  ymm1, ymm6, ymm5, 0x20  ; ymm1=row[-1]( 0  1  2  3  4  5  6  7  8  9 10 11 12 13 14 15)
+    vperm2i128  ymm5, ymm6, ymm5, 0x31  ; ymm5=row[-1](16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31)
+
+    vpunpckhbw  ymm6, ymm2, ymm8        ; ymm6=row[+1]( 8  9 10 11 12 13 14 15 24 25 26 27 28 29 30 31)
+    vpunpcklbw  ymm3, ymm2, ymm8        ; ymm3=row[+1]( 0  1  2  3  4  5  6  7 16 17 18 19 20 21 22 23)
+    vperm2i128  ymm2, ymm3, ymm6, 0x20  ; ymm2=row[+1]( 0  1  2  3  4  5  6  7  8  9 10 11 12 13 14 15)
+    vperm2i128  ymm6, ymm3, ymm6, 0x31  ; ymm6=row[+1](16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31)
+
+    vpmullw     ymm0, ymm0, [rel PW_THREE]
+    vpmullw     ymm4, ymm4, [rel PW_THREE]
+
+    vpaddw      ymm1, ymm1, ymm0        ; ymm1=Int0L=( 0  1  2  3  4  5  6  7  8  9 10 11 12 13 14 15)
+    vpaddw      ymm5, ymm5, ymm4        ; ymm5=Int0H=(16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31)
+    vpaddw      ymm2, ymm2, ymm0        ; ymm2=Int1L=( 0  1  2  3  4  5  6  7  8  9 10 11 12 13 14 15)
+    vpaddw      ymm6, ymm6, ymm4        ; ymm6=Int1H=(16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31)
+
+    vmovdqu     YMMWORD [rdx+0*SIZEOF_YMMWORD], ymm1  ; temporarily save
+    vmovdqu     YMMWORD [rdx+1*SIZEOF_YMMWORD], ymm5  ; the intermediate data
+    vmovdqu     YMMWORD [rdi+0*SIZEOF_YMMWORD], ymm2
+    vmovdqu     YMMWORD [rdi+1*SIZEOF_YMMWORD], ymm6
+
+    vpand       ymm1, ymm1, ymm10       ; ymm1=( 0 -- -- -- -- -- -- -- -- -- -- -- -- -- -- --)
+    vpand       ymm2, ymm2, ymm10       ; ymm2=( 0 -- -- -- -- -- -- -- -- -- -- -- -- -- -- --)
+
+    vmovdqa     YMMWORD [wk(0)], ymm1
+    vmovdqa     YMMWORD [wk(1)], ymm2
+
+    add         rax, byte SIZEOF_YMMWORD-1
+    and         rax, byte -SIZEOF_YMMWORD
+    cmp         rax, byte SIZEOF_YMMWORD
+    ja          short .columnloop
+
+.columnloop_last:
+    ; -- process the last column block
+
+    vpand       ymm1, ymm9, YMMWORD [rdx+1*SIZEOF_YMMWORD]
+    vpand       ymm2, ymm9, YMMWORD [rdi+1*SIZEOF_YMMWORD]
+
+    vmovdqa     YMMWORD [wk(2)], ymm1   ; ymm1=(-- -- -- -- -- -- -- -- -- -- -- -- -- -- -- 31)
+    vmovdqa     YMMWORD [wk(3)], ymm2   ; ymm2=(-- -- -- -- -- -- -- -- -- -- -- -- -- -- -- 31)
+
+    jmp         near .upsample
+
+.columnloop:
+    ; -- process the next column block
+
+    vmovdqu     ymm0, YMMWORD [rbx+1*SIZEOF_YMMWORD]  ; ymm0=row[ 0][1]
+    vmovdqu     ymm1, YMMWORD [rcx+1*SIZEOF_YMMWORD]  ; ymm1=row[-1][1]
+    vmovdqu     ymm2, YMMWORD [rsi+1*SIZEOF_YMMWORD]  ; ymm2=row[+1][1]
+
+    vpunpckhbw  ymm4, ymm0, ymm8        ; ymm4=row[ 0]( 8  9 10 11 12 13 14 15 24 25 26 27 28 29 30 31)
+    vpunpcklbw  ymm5, ymm0, ymm8        ; ymm5=row[ 0]( 0  1  2  3  4  5  6  7 16 17 18 19 20 21 22 23)
+    vperm2i128  ymm0, ymm5, ymm4, 0x20  ; ymm0=row[ 0]( 0  1  2  3  4  5  6  7  8  9 10 11 12 13 14 15)
+    vperm2i128  ymm4, ymm5, ymm4, 0x31  ; ymm4=row[ 0](16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31)
+
+    vpunpckhbw  ymm5, ymm1, ymm8        ; ymm5=row[-1]( 8  9 10 11 12 13 14 15 24 25 26 27 28 29 30 31)
+    vpunpcklbw  ymm6, ymm1, ymm8        ; ymm6=row[-1]( 0  1  2  3  4  5  6  7 16 17 18 19 20 21 22 23)
+    vperm2i128  ymm1, ymm6, ymm5, 0x20  ; ymm1=row[-1]( 0  1  2  3  4  5  6  7  8  9 10 11 12 13 14 15)
+    vperm2i128  ymm5, ymm6, ymm5, 0x31  ; ymm5=row[-1](16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31)
+
+    vpunpckhbw  ymm6, ymm2, ymm8        ; ymm6=row[+1]( 8  9 10 11 12 13 14 15 24 25 26 27 28 29 30 31)
+    vpunpcklbw  ymm7, ymm2, ymm8        ; ymm7=row[+1]( 0  1  2  3  4  5  6  7 16 17 18 19 20 21 22 23)
+    vperm2i128  ymm2, ymm7, ymm6, 0x20  ; ymm2=row[+1]( 0  1  2  3  4  5  6  7  8  9 10 11 12 13 14 15)
+    vperm2i128  ymm6, ymm7, ymm6, 0x31  ; ymm6=row[+1](16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31)
+
+    vpmullw     ymm0, ymm0, [rel PW_THREE]
+    vpmullw     ymm4, ymm4, [rel PW_THREE]
+
+    vpaddw      ymm1, ymm1, ymm0        ; ymm1=Int0L=( 0  1  2  3  4  5  6  7  8  9 10 11 12 13 14 15)
+    vpaddw      ymm5, ymm5, ymm4        ; ymm5=Int0H=(16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31)
+    vpaddw      ymm2, ymm2, ymm0        ; ymm2=Int1L=( 0  1  2  3  4  5  6  7  8  9 10 11 12 13 14 15)
+    vpaddw      ymm6, ymm6, ymm4        ; ymm6=Int1H=(16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31)
+
+    vmovdqu     YMMWORD [rdx+2*SIZEOF_YMMWORD], ymm1  ; temporarily save
+    vmovdqu     YMMWORD [rdx+3*SIZEOF_YMMWORD], ymm5  ; the intermediate data
+    vmovdqu     YMMWORD [rdi+2*SIZEOF_YMMWORD], ymm2
+    vmovdqu     YMMWORD [rdi+3*SIZEOF_YMMWORD], ymm6
+
+    vperm2i128  ymm1, ymm8, ymm1, 0x20
+    vpslldq     ymm1, ymm1, 14          ; ymm1=(-- -- -- -- -- -- -- -- -- -- -- -- -- -- --  0)
+    vperm2i128  ymm2, ymm8, ymm2, 0x20
+    vpslldq     ymm2, ymm2, 14          ; ymm2=(-- -- -- -- -- -- -- -- -- -- -- -- -- -- --  0)
+
+    vmovdqa     YMMWORD [wk(2)], ymm1
+    vmovdqa     YMMWORD [wk(3)], ymm2
+
+.upsample:
+    ; -- process the upper row
+
+    vmovdqu     ymm7, YMMWORD [rdx+0*SIZEOF_YMMWORD]  ; ymm7=Int0L=( 0  1  2  3  4  5  6  7  8  9 10 11 12 13 14 15)
+    vmovdqu     ymm3, YMMWORD [rdx+1*SIZEOF_YMMWORD]  ; ymm3=Int0H=(16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31)
+
+    vperm2i128  ymm0, ymm8, ymm7, 0x03
+    vpalignr    ymm0, ymm0, ymm7, 2     ; ymm0=( 1  2  3  4  5  6  7  8  9 10 11 12 13 14 15 --)
+    vperm2i128  ymm4, ymm8, ymm3, 0x20
+    vpslldq     ymm4, ymm4, 14          ; ymm4=(-- -- -- -- -- -- -- -- -- -- -- -- -- -- -- 16)
+
+    vperm2i128  ymm5, ymm8, ymm7, 0x03
+    vpsrldq     ymm5, ymm5, 14          ; ymm5=(15 -- -- -- -- -- -- -- -- -- -- -- -- -- -- --)
+    vperm2i128  ymm6, ymm8, ymm3, 0x20
+    vpalignr    ymm6, ymm3, ymm6, 14    ; ymm6=(-- 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30)
+
+    vpor        ymm0, ymm0, ymm4        ; ymm0=( 1  2  3  4  5  6  7  8  9 10 11 12 13 14 15 16)
+    vpor        ymm5, ymm5, ymm6        ; ymm5=(15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30)
+
+    vperm2i128  ymm2, ymm8, ymm3, 0x03
+    vpalignr    ymm2, ymm2, ymm3, 2     ; ymm2=(17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 --)
+    vperm2i128  ymm4, ymm8, ymm3, 0x03
+    vpsrldq     ymm4, ymm4, 14          ; ymm4=(31 -- -- -- -- -- -- -- -- -- -- -- -- -- -- --)
+    vperm2i128  ymm1, ymm8, ymm7, 0x20
+    vpalignr    ymm1, ymm7, ymm1, 14    ; ymm1=(--  0  1  2  3  4  5  6  7  8  9 10 11 12 13 14)
+
+    vpor        ymm1, ymm1, YMMWORD [wk(0)]  ; ymm1=(-1  0  1  2  3  4  5  6  7  8  9 10 11 12 13 14)
+    vpor        ymm2, ymm2, YMMWORD [wk(2)]  ; ymm2=(17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32)
+
+    vmovdqa     YMMWORD [wk(0)], ymm4
+
+    vpmullw     ymm7, ymm7, [rel PW_THREE]
+    vpmullw     ymm3, ymm3, [rel PW_THREE]
+    vpaddw      ymm1, ymm1, [rel PW_EIGHT]
+    vpaddw      ymm5, ymm5, [rel PW_EIGHT]
+    vpaddw      ymm0, ymm0, [rel PW_SEVEN]
+    vpaddw      ymm2, [rel PW_SEVEN]
+
+    vpaddw      ymm1, ymm1, ymm7
+    vpaddw      ymm5, ymm5, ymm3
+    vpsrlw      ymm1, ymm1, 4           ; ymm1=Out0LE=( 0  2  4  6  8 10 12 14 16 18 20 22 24 26 28 30)
+    vpsrlw      ymm5, ymm5, 4           ; ymm5=Out0HE=(32 34 36 38 40 42 44 46 48 50 52 54 56 58 60 62)
+    vpaddw      ymm0, ymm0, ymm7
+    vpaddw      ymm2, ymm2, ymm3
+    vpsrlw      ymm0, ymm0, 4           ; ymm0=Out0LO=( 1  3  5  7  9 11 13 15 17 19 21 23 25 27 29 31)
+    vpsrlw      ymm2, ymm2, 4           ; ymm2=Out0HO=(33 35 37 39 41 43 45 47 49 51 53 55 57 59 61 63)
+
+    vpsllw      ymm0, ymm0, BYTE_BIT
+    vpsllw      ymm2, ymm2, BYTE_BIT
+    vpor        ymm1, ymm1, ymm0        ; ymm1=Out0L=( 0  1  2 ... 29 30 31)
+    vpor        ymm5, ymm5, ymm2        ; ymm5=Out0H=(32 33 34 ... 61 62 63)
+
+    vmovdqu     YMMWORD [rdx+0*SIZEOF_YMMWORD], ymm1
+    vmovdqu     YMMWORD [rdx+1*SIZEOF_YMMWORD], ymm5
+
+    ; -- process the lower row
+
+    vmovdqu     ymm6, YMMWORD [rdi+0*SIZEOF_YMMWORD]  ; ymm6=Int1L=( 0  1  2  3  4  5  6  7  8  9 10 11 12 13 14 15)
+    vmovdqu     ymm4, YMMWORD [rdi+1*SIZEOF_YMMWORD]  ; ymm4=Int1H=(16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31)
+
+    vperm2i128  ymm7, ymm8, ymm6, 0x03
+    vpalignr    ymm7, ymm7, ymm6, 2     ; ymm7=( 1  2  3  4  5  6  7  8  9 10 11 12 13 14 15 --)
+    vperm2i128  ymm3, ymm8, ymm4, 0x20
+    vpslldq     ymm3, ymm3, 14          ; ymm3=(-- -- -- -- -- -- -- -- -- -- -- -- -- -- -- 16)
+
+    vperm2i128  ymm0, ymm8, ymm6, 0x03
+    vpsrldq     ymm0, ymm0, 14          ; ymm0=(15 -- -- -- -- -- -- -- -- -- -- -- -- -- -- --)
+    vperm2i128  ymm2, ymm8, ymm4, 0x20
+    vpalignr    ymm2, ymm4, ymm2, 14    ; ymm2=(-- 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30)
+
+    vpor        ymm7, ymm7, ymm3        ; ymm7=( 1  2  3  4  5  6  7  8  9 10 11 12 13 14 15 16)
+    vpor        ymm0, ymm0, ymm2        ; ymm0=(15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30)
+
+    vperm2i128  ymm5, ymm8, ymm4, 0x03
+    vpalignr    ymm5, ymm5, ymm4, 2     ; ymm5=(17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 --)
+    vperm2i128  ymm3, ymm8, ymm4, 0x03
+    vpsrldq     ymm3, ymm3, 14          ; ymm3=(31 -- -- -- -- -- -- -- -- -- -- -- -- -- -- --)
+    vperm2i128  ymm1, ymm8, ymm6, 0x20
+    vpalignr    ymm1, ymm6, ymm1, 14    ; ymm1=(--  0  1  2  3  4  5  6  7  8  9 10 11 12 13 14)
+
+    vpor        ymm1, ymm1, YMMWORD [wk(1)]  ; ymm1=(-1  0  1  2  3  4  5  6  7  8  9 10 11 12 13 14)
+    vpor        ymm5, ymm5, YMMWORD [wk(3)]  ; ymm5=(17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32)
+
+    vmovdqa     YMMWORD [wk(1)], ymm3
+
+    vpmullw     ymm6, ymm6, [rel PW_THREE]
+    vpmullw     ymm4, ymm4, [rel PW_THREE]
+    vpaddw      ymm1, ymm1, [rel PW_EIGHT]
+    vpaddw      ymm0, ymm0, [rel PW_EIGHT]
+    vpaddw      ymm7, ymm7, [rel PW_SEVEN]
+    vpaddw      ymm5, ymm5, [rel PW_SEVEN]
+
+    vpaddw      ymm1, ymm1, ymm6
+    vpaddw      ymm0, ymm0, ymm4
+    vpsrlw      ymm1, ymm1, 4           ; ymm1=Out1LE=( 0  2  4  6  8 10 12 14 16 18 20 22 24 26 28 30)
+    vpsrlw      ymm0, ymm0, 4           ; ymm0=Out1HE=(32 34 36 38 40 42 44 46 48 50 52 54 56 58 60 62)
+    vpaddw      ymm7, ymm7, ymm6
+    vpaddw      ymm5, ymm5, ymm4
+    vpsrlw      ymm7, ymm7, 4           ; ymm7=Out1LO=( 1  3  5  7  9 11 13 15 17 19 21 23 25 27 29 31)
+    vpsrlw      ymm5, ymm5, 4           ; ymm5=Out1HO=(33 35 37 39 41 43 45 47 49 51 53 55 57 59 61 63)
+
+    vpsllw      ymm7, ymm7, BYTE_BIT
+    vpsllw      ymm5, ymm5, BYTE_BIT
+    vpor        ymm1, ymm1, ymm7        ; ymm1=Out1L=( 0  1  2 ... 29 30 31)
+    vpor        ymm0, ymm0, ymm5        ; ymm0=Out1H=(32 33 34 ... 61 62 63)
+
+    vmovdqu     YMMWORD [rdi+0*SIZEOF_YMMWORD], ymm1
+    vmovdqu     YMMWORD [rdi+1*SIZEOF_YMMWORD], ymm0
+
+    sub         rax, byte SIZEOF_YMMWORD
+    add         rcx, byte 1*SIZEOF_YMMWORD  ; inptr1(above)
+    add         rbx, byte 1*SIZEOF_YMMWORD  ; inptr0
+    add         rsi, byte 1*SIZEOF_YMMWORD  ; inptr1(below)
+    add         rdx, byte 2*SIZEOF_YMMWORD  ; outptr0
+    add         rdi, byte 2*SIZEOF_YMMWORD  ; outptr1
+    cmp         rax, byte SIZEOF_YMMWORD
+    ja          near .columnloop
+    test        rax, rax
+    jnz         near .columnloop_last
+
+    pop         rsi
+    pop         rdi
+    pop         rcx
+    pop         rax
+
+    add         rsi, byte 1*SIZEOF_JSAMPROW  ; input_data
+    add         rdi, byte 2*SIZEOF_JSAMPROW  ; output_data
+    sub         rcx, byte 2                  ; rowctr
+    jg          near .rowloop
+
+.return:
+    pop         rbx
+    vzeroupper
+    uncollect_args 4
+    pop_xmm     3
+    mov         rsp, rbp                ; rsp <- aligned rbp
+    pop         rsp                     ; rsp <- original rbp
+    pop         rbp
+    ret
+
+; --------------------------------------------------------------------------
+;
+; Fast processing for the common case of 2:1 horizontal and 1:1 vertical.
+; It's still a box filter.
+;
+; GLOBAL(void)
+; jsimd_h2v1_upsample_avx2(int max_v_samp_factor, JDIMENSION output_width,
+;                          JSAMPARRAY input_data, JSAMPARRAY *output_data_ptr);
+;
+
+; r10 = int max_v_samp_factor
+; r11d = JDIMENSION output_width
+; r12 = JSAMPARRAY input_data
+; r13 = JSAMPARRAY *output_data_ptr
+
+    align       32
+    GLOBAL_FUNCTION(jsimd_h2v1_upsample_avx2)
+
+EXTN(jsimd_h2v1_upsample_avx2):
+    push        rbp
+    mov         rax, rsp
+    mov         rbp, rsp
+    collect_args 4
+
+    mov         edx, r11d
+    add         rdx, byte (SIZEOF_YMMWORD-1)
+    and         rdx, -SIZEOF_YMMWORD
+    jz          near .return
+
+    mov         rcx, r10                ; rowctr
+    test        rcx, rcx
+    jz          short .return
+
+    mov         rsi, r12                ; input_data
+    mov         rdi, r13
+    mov         rdip, JSAMPARRAY [rdi]  ; output_data
+.rowloop:
+    push        rdi
+    push        rsi
+
+    mov         rsip, JSAMPROW [rsi]    ; inptr
+    mov         rdip, JSAMPROW [rdi]    ; outptr
+    mov         rax, rdx                ; colctr
+.columnloop:
+
+    cmp         rax, byte SIZEOF_YMMWORD
+    ja          near .above_16
+
+    vmovdqu     xmm0, XMMWORD [rsi+0*SIZEOF_YMMWORD]
+    vpunpckhbw  xmm1, xmm0, xmm0
+    vpunpcklbw  xmm0, xmm0, xmm0
+
+    vmovdqu     XMMWORD [rdi+0*SIZEOF_XMMWORD], xmm0
+    vmovdqu     XMMWORD [rdi+1*SIZEOF_XMMWORD], xmm1
+
+    jmp         short .nextrow
+
+.above_16:
+    vmovdqu     ymm0, YMMWORD [rsi+0*SIZEOF_YMMWORD]
+
+    vpermq      ymm0, ymm0, 0xd8
+    vpunpckhbw  ymm1, ymm0, ymm0
+    vpunpcklbw  ymm0, ymm0, ymm0
+
+    vmovdqu     YMMWORD [rdi+0*SIZEOF_YMMWORD], ymm0
+    vmovdqu     YMMWORD [rdi+1*SIZEOF_YMMWORD], ymm1
+
+    sub         rax, byte 2*SIZEOF_YMMWORD
+    jz          short .nextrow
+
+    add         rsi, byte SIZEOF_YMMWORD    ; inptr
+    add         rdi, byte 2*SIZEOF_YMMWORD  ; outptr
+    jmp         short .columnloop
+
+.nextrow:
+    pop         rsi
+    pop         rdi
+
+    add         rsi, byte SIZEOF_JSAMPROW  ; input_data
+    add         rdi, byte SIZEOF_JSAMPROW  ; output_data
+    dec         rcx                        ; rowctr
+    jg          short .rowloop
+
+.return:
+    vzeroupper
+    uncollect_args 4
+    pop         rbp
+    ret
+
+; --------------------------------------------------------------------------
+;
+; Fast processing for the common case of 2:1 horizontal and 2:1 vertical.
+; It's still a box filter.
+;
+; GLOBAL(void)
+; jsimd_h2v2_upsample_avx2(int max_v_samp_factor, JDIMENSION output_width,
+;                          JSAMPARRAY input_data, JSAMPARRAY *output_data_ptr);
+;
+
+; r10 = int max_v_samp_factor
+; r11d = JDIMENSION output_width
+; r12 = JSAMPARRAY input_data
+; r13 = JSAMPARRAY *output_data_ptr
+
+    align       32
+    GLOBAL_FUNCTION(jsimd_h2v2_upsample_avx2)
+
+EXTN(jsimd_h2v2_upsample_avx2):
+    push        rbp
+    mov         rax, rsp
+    mov         rbp, rsp
+    collect_args 4
+    push        rbx
+
+    mov         edx, r11d
+    add         rdx, byte (SIZEOF_YMMWORD-1)
+    and         rdx, -SIZEOF_YMMWORD
+    jz          near .return
+
+    mov         rcx, r10                ; rowctr
+    test        rcx, rcx
+    jz          near .return
+
+    mov         rsi, r12                ; input_data
+    mov         rdi, r13
+    mov         rdip, JSAMPARRAY [rdi]  ; output_data
+.rowloop:
+    push        rdi
+    push        rsi
+
+    mov         rsip, JSAMPROW [rsi]                   ; inptr
+    mov         rbxp, JSAMPROW [rdi+0*SIZEOF_JSAMPROW] ; outptr0
+    mov         rdip, JSAMPROW [rdi+1*SIZEOF_JSAMPROW] ; outptr1
+    mov         rax, rdx                               ; colctr
+.columnloop:
+
+    cmp         rax, byte SIZEOF_YMMWORD
+    ja          short .above_16
+
+    vmovdqu     xmm0, XMMWORD [rsi+0*SIZEOF_XMMWORD]
+    vpunpckhbw  xmm1, xmm0, xmm0
+    vpunpcklbw  xmm0, xmm0, xmm0
+
+    vmovdqu     XMMWORD [rbx+0*SIZEOF_XMMWORD], xmm0
+    vmovdqu     XMMWORD [rbx+1*SIZEOF_XMMWORD], xmm1
+    vmovdqu     XMMWORD [rdi+0*SIZEOF_XMMWORD], xmm0
+    vmovdqu     XMMWORD [rdi+1*SIZEOF_XMMWORD], xmm1
+
+    jmp         near .nextrow
+
+.above_16:
+    vmovdqu     ymm0, YMMWORD [rsi+0*SIZEOF_YMMWORD]
+
+    vpermq      ymm0, ymm0, 0xd8
+    vpunpckhbw  ymm1, ymm0, ymm0
+    vpunpcklbw  ymm0, ymm0, ymm0
+
+    vmovdqu     YMMWORD [rbx+0*SIZEOF_YMMWORD], ymm0
+    vmovdqu     YMMWORD [rbx+1*SIZEOF_YMMWORD], ymm1
+    vmovdqu     YMMWORD [rdi+0*SIZEOF_YMMWORD], ymm0
+    vmovdqu     YMMWORD [rdi+1*SIZEOF_YMMWORD], ymm1
+
+    sub         rax, byte 2*SIZEOF_YMMWORD
+    jz          short .nextrow
+
+    add         rsi, byte SIZEOF_YMMWORD  ; inptr
+    add         rbx, 2*SIZEOF_YMMWORD     ; outptr0
+    add         rdi, 2*SIZEOF_YMMWORD     ; outptr1
+    jmp         short .columnloop
+
+.nextrow:
+    pop         rsi
+    pop         rdi
+
+    add         rsi, byte 1*SIZEOF_JSAMPROW  ; input_data
+    add         rdi, byte 2*SIZEOF_JSAMPROW  ; output_data
+    sub         rcx, byte 2                  ; rowctr
+    jg          near .rowloop
+
+.return:
+    pop         rbx
+    vzeroupper
+    uncollect_args 4
+    pop         rbp
+    ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+    align       32
diff --git a/media/libjpeg/simd/x86_64/jdsample-sse2.asm b/media/libjpeg/simd/x86_64/jdsample-sse2.asm
new file mode 100644
index 0000000000..38dbceec26
--- /dev/null
+++ b/media/libjpeg/simd/x86_64/jdsample-sse2.asm
@@ -0,0 +1,665 @@
+;
+; jdsample.asm - upsampling (64-bit SSE2)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+; Copyright (C) 2009, 2016, D. R. Commander.
+; Copyright (C) 2018, Matthias Räncker.
+;
+; Based on the x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+
+%include "jsimdext.inc"
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_CONST
+
+    alignz      32
+    GLOBAL_DATA(jconst_fancy_upsample_sse2)
+
+EXTN(jconst_fancy_upsample_sse2):
+
+PW_ONE   times 8 dw 1
+PW_TWO   times 8 dw 2
+PW_THREE times 8 dw 3
+PW_SEVEN times 8 dw 7
+PW_EIGHT times 8 dw 8
+
+    alignz      32
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_TEXT
+    BITS        64
+;
+; Fancy processing for the common case of 2:1 horizontal and 1:1 vertical.
+;
+; The upsampling algorithm is linear interpolation between pixel centers,
+; also known as a "triangle filter".  This is a good compromise between
+; speed and visual quality.  The centers of the output pixels are 1/4 and 3/4
+; of the way between input pixel centers.
+;
+; GLOBAL(void)
+; jsimd_h2v1_fancy_upsample_sse2(int max_v_samp_factor,
+;                                JDIMENSION downsampled_width,
+;                                JSAMPARRAY input_data,
+;                                JSAMPARRAY *output_data_ptr);
+;
+
+; r10 = int max_v_samp_factor
+; r11d = JDIMENSION downsampled_width
+; r12 = JSAMPARRAY input_data
+; r13 = JSAMPARRAY *output_data_ptr
+
+    align       32
+    GLOBAL_FUNCTION(jsimd_h2v1_fancy_upsample_sse2)
+
+EXTN(jsimd_h2v1_fancy_upsample_sse2):
+    push        rbp
+    mov         rax, rsp
+    mov         rbp, rsp
+    collect_args 4
+
+    mov         eax, r11d               ; colctr
+    test        rax, rax
+    jz          near .return
+
+    mov         rcx, r10                ; rowctr
+    test        rcx, rcx
+    jz          near .return
+
+    mov         rsi, r12                ; input_data
+    mov         rdi, r13
+    mov         rdip, JSAMPARRAY [rdi]  ; output_data
+.rowloop:
+    push        rax                     ; colctr
+    push        rdi
+    push        rsi
+
+    mov         rsip, JSAMPROW [rsi]    ; inptr
+    mov         rdip, JSAMPROW [rdi]    ; outptr
+
+    test        rax, SIZEOF_XMMWORD-1
+    jz          short .skip
+    mov         dl, JSAMPLE [rsi+(rax-1)*SIZEOF_JSAMPLE]
+    mov         JSAMPLE [rsi+rax*SIZEOF_JSAMPLE], dl    ; insert a dummy sample
+.skip:
+    pxor        xmm0, xmm0              ; xmm0=(all 0's)
+    pcmpeqb     xmm7, xmm7
+    psrldq      xmm7, (SIZEOF_XMMWORD-1)
+    pand        xmm7, XMMWORD [rsi+0*SIZEOF_XMMWORD]
+
+    add         rax, byte SIZEOF_XMMWORD-1
+    and         rax, byte -SIZEOF_XMMWORD
+    cmp         rax, byte SIZEOF_XMMWORD
+    ja          short .columnloop
+
+.columnloop_last:
+    pcmpeqb     xmm6, xmm6
+    pslldq      xmm6, (SIZEOF_XMMWORD-1)
+    pand        xmm6, XMMWORD [rsi+0*SIZEOF_XMMWORD]
+    jmp         short .upsample
+
+.columnloop:
+    movdqa      xmm6, XMMWORD [rsi+1*SIZEOF_XMMWORD]
+    pslldq      xmm6, (SIZEOF_XMMWORD-1)
+
+.upsample:
+    movdqa      xmm1, XMMWORD [rsi+0*SIZEOF_XMMWORD]
+    movdqa      xmm2, xmm1
+    movdqa      xmm3, xmm1                ; xmm1=( 0  1  2 ... 13 14 15)
+    pslldq      xmm2, 1                   ; xmm2=(--  0  1 ... 12 13 14)
+    psrldq      xmm3, 1                   ; xmm3=( 1  2  3 ... 14 15 --)
+
+    por         xmm2, xmm7                ; xmm2=(-1  0  1 ... 12 13 14)
+    por         xmm3, xmm6                ; xmm3=( 1  2  3 ... 14 15 16)
+
+    movdqa      xmm7, xmm1
+    psrldq      xmm7, (SIZEOF_XMMWORD-1)  ; xmm7=(15 -- -- ... -- -- --)
+
+    movdqa      xmm4, xmm1
+    punpcklbw   xmm1, xmm0                ; xmm1=( 0  1  2  3  4  5  6  7)
+    punpckhbw   xmm4, xmm0                ; xmm4=( 8  9 10 11 12 13 14 15)
+    movdqa      xmm5, xmm2
+    punpcklbw   xmm2, xmm0                ; xmm2=(-1  0  1  2  3  4  5  6)
+    punpckhbw   xmm5, xmm0                ; xmm5=( 7  8  9 10 11 12 13 14)
+    movdqa      xmm6, xmm3
+    punpcklbw   xmm3, xmm0                ; xmm3=( 1  2  3  4  5  6  7  8)
+    punpckhbw   xmm6, xmm0                ; xmm6=( 9 10 11 12 13 14 15 16)
+
+    pmullw      xmm1, [rel PW_THREE]
+    pmullw      xmm4, [rel PW_THREE]
+    paddw       xmm2, [rel PW_ONE]
+    paddw       xmm5, [rel PW_ONE]
+    paddw       xmm3, [rel PW_TWO]
+    paddw       xmm6, [rel PW_TWO]
+
+    paddw       xmm2, xmm1
+    paddw       xmm5, xmm4
+    psrlw       xmm2, 2                 ; xmm2=OutLE=( 0  2  4  6  8 10 12 14)
+    psrlw       xmm5, 2                 ; xmm5=OutHE=(16 18 20 22 24 26 28 30)
+    paddw       xmm3, xmm1
+    paddw       xmm6, xmm4
+    psrlw       xmm3, 2                 ; xmm3=OutLO=( 1  3  5  7  9 11 13 15)
+    psrlw       xmm6, 2                 ; xmm6=OutHO=(17 19 21 23 25 27 29 31)
+
+    psllw       xmm3, BYTE_BIT
+    psllw       xmm6, BYTE_BIT
+    por         xmm2, xmm3              ; xmm2=OutL=( 0  1  2 ... 13 14 15)
+    por         xmm5, xmm6              ; xmm5=OutH=(16 17 18 ... 29 30 31)
+
+    movdqa      XMMWORD [rdi+0*SIZEOF_XMMWORD], xmm2
+    movdqa      XMMWORD [rdi+1*SIZEOF_XMMWORD], xmm5
+
+    sub         rax, byte SIZEOF_XMMWORD
+    add         rsi, byte 1*SIZEOF_XMMWORD  ; inptr
+    add         rdi, byte 2*SIZEOF_XMMWORD  ; outptr
+    cmp         rax, byte SIZEOF_XMMWORD
+    ja          near .columnloop
+    test        eax, eax
+    jnz         near .columnloop_last
+
+    pop         rsi
+    pop         rdi
+    pop         rax
+
+    add         rsi, byte SIZEOF_JSAMPROW  ; input_data
+    add         rdi, byte SIZEOF_JSAMPROW  ; output_data
+    dec         rcx                        ; rowctr
+    jg          near .rowloop
+
+.return:
+    uncollect_args 4
+    pop         rbp
+    ret
+
+; --------------------------------------------------------------------------
+;
+; Fancy processing for the common case of 2:1 horizontal and 2:1 vertical.
+; Again a triangle filter; see comments for h2v1 case, above.
+;
+; GLOBAL(void)
+; jsimd_h2v2_fancy_upsample_sse2(int max_v_samp_factor,
+;                                JDIMENSION downsampled_width,
+;                                JSAMPARRAY input_data,
+;                                JSAMPARRAY *output_data_ptr);
+;
+
+; r10 = int max_v_samp_factor
+; r11d = JDIMENSION downsampled_width
+; r12 = JSAMPARRAY input_data
+; r13 = JSAMPARRAY *output_data_ptr
+
+%define wk(i)   rbp - (WK_NUM - (i)) * SIZEOF_XMMWORD  ; xmmword wk[WK_NUM]
+%define WK_NUM  4
+
+    align       32
+    GLOBAL_FUNCTION(jsimd_h2v2_fancy_upsample_sse2)
+
+EXTN(jsimd_h2v2_fancy_upsample_sse2):
+    push        rbp
+    mov         rax, rsp                     ; rax = original rbp
+    sub         rsp, byte 4
+    and         rsp, byte (-SIZEOF_XMMWORD)  ; align to 128 bits
+    mov         [rsp], rax
+    mov         rbp, rsp                     ; rbp = aligned rbp
+    lea         rsp, [wk(0)]
+    collect_args 4
+    push        rbx
+
+    mov         eax, r11d               ; colctr
+    test        rax, rax
+    jz          near .return
+
+    mov         rcx, r10                ; rowctr
+    test        rcx, rcx
+    jz          near .return
+
+    mov         rsi, r12                ; input_data
+    mov         rdi, r13
+    mov         rdip, JSAMPARRAY [rdi]  ; output_data
+.rowloop:
+    push        rax                     ; colctr
+    push        rcx
+    push        rdi
+    push        rsi
+
+    mov         rcxp, JSAMPROW [rsi-1*SIZEOF_JSAMPROW]  ; inptr1(above)
+    mov         rbxp, JSAMPROW [rsi+0*SIZEOF_JSAMPROW]  ; inptr0
+    mov         rsip, JSAMPROW [rsi+1*SIZEOF_JSAMPROW]  ; inptr1(below)
+    mov         rdxp, JSAMPROW [rdi+0*SIZEOF_JSAMPROW]  ; outptr0
+    mov         rdip, JSAMPROW [rdi+1*SIZEOF_JSAMPROW]  ; outptr1
+
+    test        rax, SIZEOF_XMMWORD-1
+    jz          short .skip
+    push        rdx
+    mov         dl, JSAMPLE [rcx+(rax-1)*SIZEOF_JSAMPLE]
+    mov         JSAMPLE [rcx+rax*SIZEOF_JSAMPLE], dl
+    mov         dl, JSAMPLE [rbx+(rax-1)*SIZEOF_JSAMPLE]
+    mov         JSAMPLE [rbx+rax*SIZEOF_JSAMPLE], dl
+    mov         dl, JSAMPLE [rsi+(rax-1)*SIZEOF_JSAMPLE]
+    mov         JSAMPLE [rsi+rax*SIZEOF_JSAMPLE], dl    ; insert a dummy sample
+    pop         rdx
+.skip:
+    ; -- process the first column block
+
+    movdqa      xmm0, XMMWORD [rbx+0*SIZEOF_XMMWORD]  ; xmm0=row[ 0][0]
+    movdqa      xmm1, XMMWORD [rcx+0*SIZEOF_XMMWORD]  ; xmm1=row[-1][0]
+    movdqa      xmm2, XMMWORD [rsi+0*SIZEOF_XMMWORD]  ; xmm2=row[+1][0]
+
+    pxor        xmm3, xmm3              ; xmm3=(all 0's)
+    movdqa      xmm4, xmm0
+    punpcklbw   xmm0, xmm3              ; xmm0=row[ 0]( 0  1  2  3  4  5  6  7)
+    punpckhbw   xmm4, xmm3              ; xmm4=row[ 0]( 8  9 10 11 12 13 14 15)
+    movdqa      xmm5, xmm1
+    punpcklbw   xmm1, xmm3              ; xmm1=row[-1]( 0  1  2  3  4  5  6  7)
+    punpckhbw   xmm5, xmm3              ; xmm5=row[-1]( 8  9 10 11 12 13 14 15)
+    movdqa      xmm6, xmm2
+    punpcklbw   xmm2, xmm3              ; xmm2=row[+1]( 0  1  2  3  4  5  6  7)
+    punpckhbw   xmm6, xmm3              ; xmm6=row[+1]( 8  9 10 11 12 13 14 15)
+
+    pmullw      xmm0, [rel PW_THREE]
+    pmullw      xmm4, [rel PW_THREE]
+
+    pcmpeqb     xmm7, xmm7
+    psrldq      xmm7, (SIZEOF_XMMWORD-2)
+
+    paddw       xmm1, xmm0              ; xmm1=Int0L=( 0  1  2  3  4  5  6  7)
+    paddw       xmm5, xmm4              ; xmm5=Int0H=( 8  9 10 11 12 13 14 15)
+    paddw       xmm2, xmm0              ; xmm2=Int1L=( 0  1  2  3  4  5  6  7)
+    paddw       xmm6, xmm4              ; xmm6=Int1H=( 8  9 10 11 12 13 14 15)
+
+    movdqa      XMMWORD [rdx+0*SIZEOF_XMMWORD], xmm1  ; temporarily save
+    movdqa      XMMWORD [rdx+1*SIZEOF_XMMWORD], xmm5  ; the intermediate data
+    movdqa      XMMWORD [rdi+0*SIZEOF_XMMWORD], xmm2
+    movdqa      XMMWORD [rdi+1*SIZEOF_XMMWORD], xmm6
+
+    pand        xmm1, xmm7              ; xmm1=( 0 -- -- -- -- -- -- --)
+    pand        xmm2, xmm7              ; xmm2=( 0 -- -- -- -- -- -- --)
+
+    movdqa      XMMWORD [wk(0)], xmm1
+    movdqa      XMMWORD [wk(1)], xmm2
+
+    add         rax, byte SIZEOF_XMMWORD-1
+    and         rax, byte -SIZEOF_XMMWORD
+    cmp         rax, byte SIZEOF_XMMWORD
+    ja          short .columnloop
+
+.columnloop_last:
+    ; -- process the last column block
+
+    pcmpeqb     xmm1, xmm1
+    pslldq      xmm1, (SIZEOF_XMMWORD-2)
+    movdqa      xmm2, xmm1
+
+    pand        xmm1, XMMWORD [rdx+1*SIZEOF_XMMWORD]
+    pand        xmm2, XMMWORD [rdi+1*SIZEOF_XMMWORD]
+
+    movdqa      XMMWORD [wk(2)], xmm1   ; xmm1=(-- -- -- -- -- -- -- 15)
+    movdqa      XMMWORD [wk(3)], xmm2   ; xmm2=(-- -- -- -- -- -- -- 15)
+
+    jmp         near .upsample
+
+.columnloop:
+    ; -- process the next column block
+
+    movdqa      xmm0, XMMWORD [rbx+1*SIZEOF_XMMWORD]  ; xmm0=row[ 0][1]
+    movdqa      xmm1, XMMWORD [rcx+1*SIZEOF_XMMWORD]  ; xmm1=row[-1][1]
+    movdqa      xmm2, XMMWORD [rsi+1*SIZEOF_XMMWORD]  ; xmm2=row[+1][1]
+
+    pxor        xmm3, xmm3              ; xmm3=(all 0's)
+    movdqa      xmm4, xmm0
+    punpcklbw   xmm0, xmm3              ; xmm0=row[ 0]( 0  1  2  3  4  5  6  7)
+    punpckhbw   xmm4, xmm3              ; xmm4=row[ 0]( 8  9 10 11 12 13 14 15)
+    movdqa      xmm5, xmm1
+    punpcklbw   xmm1, xmm3              ; xmm1=row[-1]( 0  1  2  3  4  5  6  7)
+    punpckhbw   xmm5, xmm3              ; xmm5=row[-1]( 8  9 10 11 12 13 14 15)
+    movdqa      xmm6, xmm2
+    punpcklbw   xmm2, xmm3              ; xmm2=row[+1]( 0  1  2  3  4  5  6  7)
+    punpckhbw   xmm6, xmm3              ; xmm6=row[+1]( 8  9 10 11 12 13 14 15)
+
+    pmullw      xmm0, [rel PW_THREE]
+    pmullw      xmm4, [rel PW_THREE]
+
+    paddw       xmm1, xmm0              ; xmm1=Int0L=( 0  1  2  3  4  5  6  7)
+    paddw       xmm5, xmm4              ; xmm5=Int0H=( 8  9 10 11 12 13 14 15)
+    paddw       xmm2, xmm0              ; xmm2=Int1L=( 0  1  2  3  4  5  6  7)
+    paddw       xmm6, xmm4              ; xmm6=Int1H=( 8  9 10 11 12 13 14 15)
+
+    movdqa      XMMWORD [rdx+2*SIZEOF_XMMWORD], xmm1  ; temporarily save
+    movdqa      XMMWORD [rdx+3*SIZEOF_XMMWORD], xmm5  ; the intermediate data
+    movdqa      XMMWORD [rdi+2*SIZEOF_XMMWORD], xmm2
+    movdqa      XMMWORD [rdi+3*SIZEOF_XMMWORD], xmm6
+
+    pslldq      xmm1, (SIZEOF_XMMWORD-2)  ; xmm1=(-- -- -- -- -- -- --  0)
+    pslldq      xmm2, (SIZEOF_XMMWORD-2)  ; xmm2=(-- -- -- -- -- -- --  0)
+
+    movdqa      XMMWORD [wk(2)], xmm1
+    movdqa      XMMWORD [wk(3)], xmm2
+
+.upsample:
+    ; -- process the upper row
+
+    movdqa      xmm7, XMMWORD [rdx+0*SIZEOF_XMMWORD]
+    movdqa      xmm3, XMMWORD [rdx+1*SIZEOF_XMMWORD]
+
+    movdqa      xmm0, xmm7                ; xmm7=Int0L=( 0  1  2  3  4  5  6  7)
+    movdqa      xmm4, xmm3                ; xmm3=Int0H=( 8  9 10 11 12 13 14 15)
+    psrldq      xmm0, 2                   ; xmm0=( 1  2  3  4  5  6  7 --)
+    pslldq      xmm4, (SIZEOF_XMMWORD-2)  ; xmm4=(-- -- -- -- -- -- --  8)
+    movdqa      xmm5, xmm7
+    movdqa      xmm6, xmm3
+    psrldq      xmm5, (SIZEOF_XMMWORD-2)  ; xmm5=( 7 -- -- -- -- -- -- --)
+    pslldq      xmm6, 2                   ; xmm6=(--  8  9 10 11 12 13 14)
+
+    por         xmm0, xmm4                ; xmm0=( 1  2  3  4  5  6  7  8)
+    por         xmm5, xmm6                ; xmm5=( 7  8  9 10 11 12 13 14)
+
+    movdqa      xmm1, xmm7
+    movdqa      xmm2, xmm3
+    pslldq      xmm1, 2                   ; xmm1=(--  0  1  2  3  4  5  6)
+    psrldq      xmm2, 2                   ; xmm2=( 9 10 11 12 13 14 15 --)
+    movdqa      xmm4, xmm3
+    psrldq      xmm4, (SIZEOF_XMMWORD-2)  ; xmm4=(15 -- -- -- -- -- -- --)
+
+    por         xmm1, XMMWORD [wk(0)]     ; xmm1=(-1  0  1  2  3  4  5  6)
+    por         xmm2, XMMWORD [wk(2)]     ; xmm2=( 9 10 11 12 13 14 15 16)
+
+    movdqa      XMMWORD [wk(0)], xmm4
+
+    pmullw      xmm7, [rel PW_THREE]
+    pmullw      xmm3, [rel PW_THREE]
+    paddw       xmm1, [rel PW_EIGHT]
+    paddw       xmm5, [rel PW_EIGHT]
+    paddw       xmm0, [rel PW_SEVEN]
+    paddw       xmm2, [rel PW_SEVEN]
+
+    paddw       xmm1, xmm7
+    paddw       xmm5, xmm3
+    psrlw       xmm1, 4                 ; xmm1=Out0LE=( 0  2  4  6  8 10 12 14)
+    psrlw       xmm5, 4                 ; xmm5=Out0HE=(16 18 20 22 24 26 28 30)
+    paddw       xmm0, xmm7
+    paddw       xmm2, xmm3
+    psrlw       xmm0, 4                 ; xmm0=Out0LO=( 1  3  5  7  9 11 13 15)
+    psrlw       xmm2, 4                 ; xmm2=Out0HO=(17 19 21 23 25 27 29 31)
+
+    psllw       xmm0, BYTE_BIT
+    psllw       xmm2, BYTE_BIT
+    por         xmm1, xmm0              ; xmm1=Out0L=( 0  1  2 ... 13 14 15)
+    por         xmm5, xmm2              ; xmm5=Out0H=(16 17 18 ... 29 30 31)
+
+    movdqa      XMMWORD [rdx+0*SIZEOF_XMMWORD], xmm1
+    movdqa      XMMWORD [rdx+1*SIZEOF_XMMWORD], xmm5
+
+    ; -- process the lower row
+
+    movdqa      xmm6, XMMWORD [rdi+0*SIZEOF_XMMWORD]
+    movdqa      xmm4, XMMWORD [rdi+1*SIZEOF_XMMWORD]
+
+    movdqa      xmm7, xmm6                ; xmm6=Int1L=( 0  1  2  3  4  5  6  7)
+    movdqa      xmm3, xmm4                ; xmm4=Int1H=( 8  9 10 11 12 13 14 15)
+    psrldq      xmm7, 2                   ; xmm7=( 1  2  3  4  5  6  7 --)
+    pslldq      xmm3, (SIZEOF_XMMWORD-2)  ; xmm3=(-- -- -- -- -- -- --  8)
+    movdqa      xmm0, xmm6
+    movdqa      xmm2, xmm4
+    psrldq      xmm0, (SIZEOF_XMMWORD-2)  ; xmm0=( 7 -- -- -- -- -- -- --)
+    pslldq      xmm2, 2                   ; xmm2=(--  8  9 10 11 12 13 14)
+
+    por         xmm7, xmm3                ; xmm7=( 1  2  3  4  5  6  7  8)
+    por         xmm0, xmm2                ; xmm0=( 7  8  9 10 11 12 13 14)
+
+    movdqa      xmm1, xmm6
+    movdqa      xmm5, xmm4
+    pslldq      xmm1, 2                   ; xmm1=(--  0  1  2  3  4  5  6)
+    psrldq      xmm5, 2                   ; xmm5=( 9 10 11 12 13 14 15 --)
+    movdqa      xmm3, xmm4
+    psrldq      xmm3, (SIZEOF_XMMWORD-2)  ; xmm3=(15 -- -- -- -- -- -- --)
+
+    por         xmm1, XMMWORD [wk(1)]     ; xmm1=(-1  0  1  2  3  4  5  6)
+    por         xmm5, XMMWORD [wk(3)]     ; xmm5=( 9 10 11 12 13 14 15 16)
+
+    movdqa      XMMWORD [wk(1)], xmm3
+
+    pmullw      xmm6, [rel PW_THREE]
+    pmullw      xmm4, [rel PW_THREE]
+    paddw       xmm1, [rel PW_EIGHT]
+    paddw       xmm0, [rel PW_EIGHT]
+    paddw       xmm7, [rel PW_SEVEN]
+    paddw       xmm5, [rel PW_SEVEN]
+
+    paddw       xmm1, xmm6
+    paddw       xmm0, xmm4
+    psrlw       xmm1, 4                 ; xmm1=Out1LE=( 0  2  4  6  8 10 12 14)
+    psrlw       xmm0, 4                 ; xmm0=Out1HE=(16 18 20 22 24 26 28 30)
+    paddw       xmm7, xmm6
+    paddw       xmm5, xmm4
+    psrlw       xmm7, 4                 ; xmm7=Out1LO=( 1  3  5  7  9 11 13 15)
+    psrlw       xmm5, 4                 ; xmm5=Out1HO=(17 19 21 23 25 27 29 31)
+
+    psllw       xmm7, BYTE_BIT
+    psllw       xmm5, BYTE_BIT
+    por         xmm1, xmm7              ; xmm1=Out1L=( 0  1  2 ... 13 14 15)
+    por         xmm0, xmm5              ; xmm0=Out1H=(16 17 18 ... 29 30 31)
+
+    movdqa      XMMWORD [rdi+0*SIZEOF_XMMWORD], xmm1
+    movdqa      XMMWORD [rdi+1*SIZEOF_XMMWORD], xmm0
+
+    sub         rax, byte SIZEOF_XMMWORD
+    add         rcx, byte 1*SIZEOF_XMMWORD  ; inptr1(above)
+    add         rbx, byte 1*SIZEOF_XMMWORD  ; inptr0
+    add         rsi, byte 1*SIZEOF_XMMWORD  ; inptr1(below)
+    add         rdx, byte 2*SIZEOF_XMMWORD  ; outptr0
+    add         rdi, byte 2*SIZEOF_XMMWORD  ; outptr1
+    cmp         rax, byte SIZEOF_XMMWORD
+    ja          near .columnloop
+    test        rax, rax
+    jnz         near .columnloop_last
+
+    pop         rsi
+    pop         rdi
+    pop         rcx
+    pop         rax
+
+    add         rsi, byte 1*SIZEOF_JSAMPROW  ; input_data
+    add         rdi, byte 2*SIZEOF_JSAMPROW  ; output_data
+    sub         rcx, byte 2                  ; rowctr
+    jg          near .rowloop
+
+.return:
+    pop         rbx
+    uncollect_args 4
+    mov         rsp, rbp                ; rsp <- aligned rbp
+    pop         rsp                     ; rsp <- original rbp
+    pop         rbp
+    ret
+
+; --------------------------------------------------------------------------
+;
+; Fast processing for the common case of 2:1 horizontal and 1:1 vertical.
+; It's still a box filter.
+;
+; GLOBAL(void)
+; jsimd_h2v1_upsample_sse2(int max_v_samp_factor, JDIMENSION output_width,
+;                          JSAMPARRAY input_data, JSAMPARRAY *output_data_ptr);
+;
+
+; r10 = int max_v_samp_factor
+; r11d = JDIMENSION output_width
+; r12 = JSAMPARRAY input_data
+; r13 = JSAMPARRAY *output_data_ptr
+
+    align       32
+    GLOBAL_FUNCTION(jsimd_h2v1_upsample_sse2)
+
+EXTN(jsimd_h2v1_upsample_sse2):
+    push        rbp
+    mov         rax, rsp
+    mov         rbp, rsp
+    collect_args 4
+
+    mov         edx, r11d
+    add         rdx, byte (2*SIZEOF_XMMWORD)-1
+    and         rdx, byte -(2*SIZEOF_XMMWORD)
+    jz          near .return
+
+    mov         rcx, r10                ; rowctr
+    test        rcx, rcx
+    jz          short .return
+
+    mov         rsi, r12                ; input_data
+    mov         rdi, r13
+    mov         rdip, JSAMPARRAY [rdi]  ; output_data
+.rowloop:
+    push        rdi
+    push        rsi
+
+    mov         rsip, JSAMPROW [rsi]    ; inptr
+    mov         rdip, JSAMPROW [rdi]    ; outptr
+    mov         rax, rdx                ; colctr
+.columnloop:
+
+    movdqa      xmm0, XMMWORD [rsi+0*SIZEOF_XMMWORD]
+
+    movdqa      xmm1, xmm0
+    punpcklbw   xmm0, xmm0
+    punpckhbw   xmm1, xmm1
+
+    movdqa      XMMWORD [rdi+0*SIZEOF_XMMWORD], xmm0
+    movdqa      XMMWORD [rdi+1*SIZEOF_XMMWORD], xmm1
+
+    sub         rax, byte 2*SIZEOF_XMMWORD
+    jz          short .nextrow
+
+    movdqa      xmm2, XMMWORD [rsi+1*SIZEOF_XMMWORD]
+
+    movdqa      xmm3, xmm2
+    punpcklbw   xmm2, xmm2
+    punpckhbw   xmm3, xmm3
+
+    movdqa      XMMWORD [rdi+2*SIZEOF_XMMWORD], xmm2
+    movdqa      XMMWORD [rdi+3*SIZEOF_XMMWORD], xmm3
+
+    sub         rax, byte 2*SIZEOF_XMMWORD
+    jz          short .nextrow
+
+    add         rsi, byte 2*SIZEOF_XMMWORD  ; inptr
+    add         rdi, byte 4*SIZEOF_XMMWORD  ; outptr
+    jmp         short .columnloop
+
+.nextrow:
+    pop         rsi
+    pop         rdi
+
+    add         rsi, byte SIZEOF_JSAMPROW  ; input_data
+    add         rdi, byte SIZEOF_JSAMPROW  ; output_data
+    dec         rcx                        ; rowctr
+    jg          short .rowloop
+
+.return:
+    uncollect_args 4
+    pop         rbp
+    ret
+
+; --------------------------------------------------------------------------
+;
+; Fast processing for the common case of 2:1 horizontal and 2:1 vertical.
+; It's still a box filter.
+;
+; GLOBAL(void)
+; jsimd_h2v2_upsample_sse2(int max_v_samp_factor, JDIMENSION output_width,
+;                          JSAMPARRAY input_data, JSAMPARRAY *output_data_ptr);
+;
+
+; r10 = int max_v_samp_factor
+; r11d = JDIMENSION output_width
+; r12 = JSAMPARRAY input_data
+; r13 = JSAMPARRAY *output_data_ptr
+
+    align       32
+    GLOBAL_FUNCTION(jsimd_h2v2_upsample_sse2)
+
+EXTN(jsimd_h2v2_upsample_sse2):
+    push        rbp
+    mov         rax, rsp
+    mov         rbp, rsp
+    collect_args 4
+    push        rbx
+
+    mov         edx, r11d
+    add         rdx, byte (2*SIZEOF_XMMWORD)-1
+    and         rdx, byte -(2*SIZEOF_XMMWORD)
+    jz          near .return
+
+    mov         rcx, r10                ; rowctr
+    test        rcx, rcx
+    jz          near .return
+
+    mov         rsi, r12                ; input_data
+    mov         rdi, r13
+    mov         rdip, JSAMPARRAY [rdi]  ; output_data
+.rowloop:
+    push        rdi
+    push        rsi
+
+    mov         rsip, JSAMPROW [rsi]                   ; inptr
+    mov         rbxp, JSAMPROW [rdi+0*SIZEOF_JSAMPROW] ; outptr0
+    mov         rdip, JSAMPROW [rdi+1*SIZEOF_JSAMPROW] ; outptr1
+    mov         rax, rdx                               ; colctr
+.columnloop:
+
+    movdqa      xmm0, XMMWORD [rsi+0*SIZEOF_XMMWORD]
+
+    movdqa      xmm1, xmm0
+    punpcklbw   xmm0, xmm0
+    punpckhbw   xmm1, xmm1
+
+    movdqa      XMMWORD [rbx+0*SIZEOF_XMMWORD], xmm0
+    movdqa      XMMWORD [rbx+1*SIZEOF_XMMWORD], xmm1
+    movdqa      XMMWORD [rdi+0*SIZEOF_XMMWORD], xmm0
+    movdqa      XMMWORD [rdi+1*SIZEOF_XMMWORD], xmm1
+
+    sub         rax, byte 2*SIZEOF_XMMWORD
+    jz          short .nextrow
+
+    movdqa      xmm2, XMMWORD [rsi+1*SIZEOF_XMMWORD]
+
+    movdqa      xmm3, xmm2
+    punpcklbw   xmm2, xmm2
+    punpckhbw   xmm3, xmm3
+
+    movdqa      XMMWORD [rbx+2*SIZEOF_XMMWORD], xmm2
+    movdqa      XMMWORD [rbx+3*SIZEOF_XMMWORD], xmm3
+    movdqa      XMMWORD [rdi+2*SIZEOF_XMMWORD], xmm2
+    movdqa      XMMWORD [rdi+3*SIZEOF_XMMWORD], xmm3
+
+    sub         rax, byte 2*SIZEOF_XMMWORD
+    jz          short .nextrow
+
+    add         rsi, byte 2*SIZEOF_XMMWORD  ; inptr
+    add         rbx, byte 4*SIZEOF_XMMWORD  ; outptr0
+    add         rdi, byte 4*SIZEOF_XMMWORD  ; outptr1
+    jmp         short .columnloop
+
+.nextrow:
+    pop         rsi
+    pop         rdi
+
+    add         rsi, byte 1*SIZEOF_JSAMPROW  ; input_data
+    add         rdi, byte 2*SIZEOF_JSAMPROW  ; output_data
+    sub         rcx, byte 2                  ; rowctr
+    jg          near .rowloop
+
+.return:
+    pop         rbx
+    uncollect_args 4
+    pop         rbp
+    ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+    align       32
diff --git a/media/libjpeg/simd/x86_64/jfdctflt-sse.asm b/media/libjpeg/simd/x86_64/jfdctflt-sse.asm
new file mode 100644
index 0000000000..ef2796649b
--- /dev/null
+++ b/media/libjpeg/simd/x86_64/jfdctflt-sse.asm
@@ -0,0 +1,355 @@
+;
+; jfdctflt.asm - floating-point FDCT (64-bit SSE)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+; Copyright (C) 2009, 2016, D. R. Commander.
+;
+; Based on the x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; This file contains a floating-point implementation of the forward DCT
+; (Discrete Cosine Transform). The following code is based directly on
+; the IJG's original jfdctflt.c; see the jfdctflt.c for more details.
+
+%include "jsimdext.inc"
+%include "jdct.inc"
+
+; --------------------------------------------------------------------------
+
+%macro  unpcklps2 2  ; %1=(0 1 2 3) / %2=(4 5 6 7) => %1=(0 1 4 5)
+    shufps      %1, %2, 0x44
+%endmacro
+
+%macro  unpckhps2 2  ; %1=(0 1 2 3) / %2=(4 5 6 7) => %1=(2 3 6 7)
+    shufps      %1, %2, 0xEE
+%endmacro
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_CONST
+
+    alignz      32
+    GLOBAL_DATA(jconst_fdct_float_sse)
+
+EXTN(jconst_fdct_float_sse):
+
+PD_0_382 times 4 dd 0.382683432365089771728460
+PD_0_707 times 4 dd 0.707106781186547524400844
+PD_0_541 times 4 dd 0.541196100146196984399723
+PD_1_306 times 4 dd 1.306562964876376527856643
+
+    alignz      32
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_TEXT
+    BITS        64
+;
+; Perform the forward DCT on one block of samples.
+;
+; GLOBAL(void)
+; jsimd_fdct_float_sse(FAST_FLOAT *data)
+;
+
+; r10 = FAST_FLOAT *data
+
+%define wk(i)   rbp - (WK_NUM - (i)) * SIZEOF_XMMWORD  ; xmmword wk[WK_NUM]
+%define WK_NUM  2
+
+    align       32
+    GLOBAL_FUNCTION(jsimd_fdct_float_sse)
+
+EXTN(jsimd_fdct_float_sse):
+    push        rbp
+    mov         rax, rsp                     ; rax = original rbp
+    sub         rsp, byte 4
+    and         rsp, byte (-SIZEOF_XMMWORD)  ; align to 128 bits
+    mov         [rsp], rax
+    mov         rbp, rsp                     ; rbp = aligned rbp
+    lea         rsp, [wk(0)]
+    collect_args 1
+
+    ; ---- Pass 1: process rows.
+
+    mov         rdx, r10                ; (FAST_FLOAT *)
+    mov         rcx, DCTSIZE/4
+.rowloop:
+
+    movaps      xmm0, XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_FAST_FLOAT)]
+    movaps      xmm1, XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_FAST_FLOAT)]
+    movaps      xmm2, XMMWORD [XMMBLOCK(2,1,rdx,SIZEOF_FAST_FLOAT)]
+    movaps      xmm3, XMMWORD [XMMBLOCK(3,1,rdx,SIZEOF_FAST_FLOAT)]
+
+    ; xmm0=(20 21 22 23), xmm2=(24 25 26 27)
+    ; xmm1=(30 31 32 33), xmm3=(34 35 36 37)
+
+    movaps      xmm4, xmm0              ; transpose coefficients(phase 1)
+    unpcklps    xmm0, xmm1              ; xmm0=(20 30 21 31)
+    unpckhps    xmm4, xmm1              ; xmm4=(22 32 23 33)
+    movaps      xmm5, xmm2              ; transpose coefficients(phase 1)
+    unpcklps    xmm2, xmm3              ; xmm2=(24 34 25 35)
+    unpckhps    xmm5, xmm3              ; xmm5=(26 36 27 37)
+
+    movaps      xmm6, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_FAST_FLOAT)]
+    movaps      xmm7, XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_FAST_FLOAT)]
+    movaps      xmm1, XMMWORD [XMMBLOCK(0,1,rdx,SIZEOF_FAST_FLOAT)]
+    movaps      xmm3, XMMWORD [XMMBLOCK(1,1,rdx,SIZEOF_FAST_FLOAT)]
+
+    ; xmm6=(00 01 02 03), xmm1=(04 05 06 07)
+    ; xmm7=(10 11 12 13), xmm3=(14 15 16 17)
+
+    movaps      XMMWORD [wk(0)], xmm4   ; wk(0)=(22 32 23 33)
+    movaps      XMMWORD [wk(1)], xmm2   ; wk(1)=(24 34 25 35)
+
+    movaps      xmm4, xmm6              ; transpose coefficients(phase 1)
+    unpcklps    xmm6, xmm7              ; xmm6=(00 10 01 11)
+    unpckhps    xmm4, xmm7              ; xmm4=(02 12 03 13)
+    movaps      xmm2, xmm1              ; transpose coefficients(phase 1)
+    unpcklps    xmm1, xmm3              ; xmm1=(04 14 05 15)
+    unpckhps    xmm2, xmm3              ; xmm2=(06 16 07 17)
+
+    movaps      xmm7, xmm6              ; transpose coefficients(phase 2)
+    unpcklps2   xmm6, xmm0              ; xmm6=(00 10 20 30)=data0
+    unpckhps2   xmm7, xmm0              ; xmm7=(01 11 21 31)=data1
+    movaps      xmm3, xmm2              ; transpose coefficients(phase 2)
+    unpcklps2   xmm2, xmm5              ; xmm2=(06 16 26 36)=data6
+    unpckhps2   xmm3, xmm5              ; xmm3=(07 17 27 37)=data7
+
+    movaps      xmm0, xmm7
+    movaps      xmm5, xmm6
+    subps       xmm7, xmm2              ; xmm7=data1-data6=tmp6
+    subps       xmm6, xmm3              ; xmm6=data0-data7=tmp7
+    addps       xmm0, xmm2              ; xmm0=data1+data6=tmp1
+    addps       xmm5, xmm3              ; xmm5=data0+data7=tmp0
+
+    movaps      xmm2, XMMWORD [wk(0)]   ; xmm2=(22 32 23 33)
+    movaps      xmm3, XMMWORD [wk(1)]   ; xmm3=(24 34 25 35)
+    movaps      XMMWORD [wk(0)], xmm7   ; wk(0)=tmp6
+    movaps      XMMWORD [wk(1)], xmm6   ; wk(1)=tmp7
+
+    movaps      xmm7, xmm4              ; transpose coefficients(phase 2)
+    unpcklps2   xmm4, xmm2              ; xmm4=(02 12 22 32)=data2
+    unpckhps2   xmm7, xmm2              ; xmm7=(03 13 23 33)=data3
+    movaps      xmm6, xmm1              ; transpose coefficients(phase 2)
+    unpcklps2   xmm1, xmm3              ; xmm1=(04 14 24 34)=data4
+    unpckhps2   xmm6, xmm3              ; xmm6=(05 15 25 35)=data5
+
+    movaps      xmm2, xmm7
+    movaps      xmm3, xmm4
+    addps       xmm7, xmm1              ; xmm7=data3+data4=tmp3
+    addps       xmm4, xmm6              ; xmm4=data2+data5=tmp2
+    subps       xmm2, xmm1              ; xmm2=data3-data4=tmp4
+    subps       xmm3, xmm6              ; xmm3=data2-data5=tmp5
+
+    ; -- Even part
+
+    movaps      xmm1, xmm5
+    movaps      xmm6, xmm0
+    subps       xmm5, xmm7              ; xmm5=tmp13
+    subps       xmm0, xmm4              ; xmm0=tmp12
+    addps       xmm1, xmm7              ; xmm1=tmp10
+    addps       xmm6, xmm4              ; xmm6=tmp11
+
+    addps       xmm0, xmm5
+    mulps       xmm0, [rel PD_0_707]    ; xmm0=z1
+
+    movaps      xmm7, xmm1
+    movaps      xmm4, xmm5
+    subps       xmm1, xmm6              ; xmm1=data4
+    subps       xmm5, xmm0              ; xmm5=data6
+    addps       xmm7, xmm6              ; xmm7=data0
+    addps       xmm4, xmm0              ; xmm4=data2
+
+    movaps      XMMWORD [XMMBLOCK(0,1,rdx,SIZEOF_FAST_FLOAT)], xmm1
+    movaps      XMMWORD [XMMBLOCK(2,1,rdx,SIZEOF_FAST_FLOAT)], xmm5
+    movaps      XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_FAST_FLOAT)], xmm7
+    movaps      XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_FAST_FLOAT)], xmm4
+
+    ; -- Odd part
+
+    movaps      xmm6, XMMWORD [wk(0)]   ; xmm6=tmp6
+    movaps      xmm0, XMMWORD [wk(1)]   ; xmm0=tmp7
+
+    addps       xmm2, xmm3              ; xmm2=tmp10
+    addps       xmm3, xmm6              ; xmm3=tmp11
+    addps       xmm6, xmm0              ; xmm6=tmp12, xmm0=tmp7
+
+    mulps       xmm3, [rel PD_0_707]    ; xmm3=z3
+
+    movaps      xmm1, xmm2              ; xmm1=tmp10
+    subps       xmm2, xmm6
+    mulps       xmm2, [rel PD_0_382]    ; xmm2=z5
+    mulps       xmm1, [rel PD_0_541]    ; xmm1=MULTIPLY(tmp10,FIX_0_541196)
+    mulps       xmm6, [rel PD_1_306]    ; xmm6=MULTIPLY(tmp12,FIX_1_306562)
+    addps       xmm1, xmm2              ; xmm1=z2
+    addps       xmm6, xmm2              ; xmm6=z4
+
+    movaps      xmm5, xmm0
+    subps       xmm0, xmm3              ; xmm0=z13
+    addps       xmm5, xmm3              ; xmm5=z11
+
+    movaps      xmm7, xmm0
+    movaps      xmm4, xmm5
+    subps       xmm0, xmm1              ; xmm0=data3
+    subps       xmm5, xmm6              ; xmm5=data7
+    addps       xmm7, xmm1              ; xmm7=data5
+    addps       xmm4, xmm6              ; xmm4=data1
+
+    movaps      XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_FAST_FLOAT)], xmm0
+    movaps      XMMWORD [XMMBLOCK(3,1,rdx,SIZEOF_FAST_FLOAT)], xmm5
+    movaps      XMMWORD [XMMBLOCK(1,1,rdx,SIZEOF_FAST_FLOAT)], xmm7
+    movaps      XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_FAST_FLOAT)], xmm4
+
+    add         rdx, 4*DCTSIZE*SIZEOF_FAST_FLOAT
+    dec         rcx
+    jnz         near .rowloop
+
+    ; ---- Pass 2: process columns.
+
+    mov         rdx, r10                ; (FAST_FLOAT *)
+    mov         rcx, DCTSIZE/4
+.columnloop:
+
+    movaps      xmm0, XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_FAST_FLOAT)]
+    movaps      xmm1, XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_FAST_FLOAT)]
+    movaps      xmm2, XMMWORD [XMMBLOCK(6,0,rdx,SIZEOF_FAST_FLOAT)]
+    movaps      xmm3, XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_FAST_FLOAT)]
+
+    ; xmm0=(02 12 22 32), xmm2=(42 52 62 72)
+    ; xmm1=(03 13 23 33), xmm3=(43 53 63 73)
+
+    movaps      xmm4, xmm0              ; transpose coefficients(phase 1)
+    unpcklps    xmm0, xmm1              ; xmm0=(02 03 12 13)
+    unpckhps    xmm4, xmm1              ; xmm4=(22 23 32 33)
+    movaps      xmm5, xmm2              ; transpose coefficients(phase 1)
+    unpcklps    xmm2, xmm3              ; xmm2=(42 43 52 53)
+    unpckhps    xmm5, xmm3              ; xmm5=(62 63 72 73)
+
+    movaps      xmm6, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_FAST_FLOAT)]
+    movaps      xmm7, XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_FAST_FLOAT)]
+    movaps      xmm1, XMMWORD [XMMBLOCK(4,0,rdx,SIZEOF_FAST_FLOAT)]
+    movaps      xmm3, XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_FAST_FLOAT)]
+
+    ; xmm6=(00 10 20 30), xmm1=(40 50 60 70)
+    ; xmm7=(01 11 21 31), xmm3=(41 51 61 71)
+
+    movaps      XMMWORD [wk(0)], xmm4   ; wk(0)=(22 23 32 33)
+    movaps      XMMWORD [wk(1)], xmm2   ; wk(1)=(42 43 52 53)
+
+    movaps      xmm4, xmm6              ; transpose coefficients(phase 1)
+    unpcklps    xmm6, xmm7              ; xmm6=(00 01 10 11)
+    unpckhps    xmm4, xmm7              ; xmm4=(20 21 30 31)
+    movaps      xmm2, xmm1              ; transpose coefficients(phase 1)
+    unpcklps    xmm1, xmm3              ; xmm1=(40 41 50 51)
+    unpckhps    xmm2, xmm3              ; xmm2=(60 61 70 71)
+
+    movaps      xmm7, xmm6              ; transpose coefficients(phase 2)
+    unpcklps2   xmm6, xmm0              ; xmm6=(00 01 02 03)=data0
+    unpckhps2   xmm7, xmm0              ; xmm7=(10 11 12 13)=data1
+    movaps      xmm3, xmm2              ; transpose coefficients(phase 2)
+    unpcklps2   xmm2, xmm5              ; xmm2=(60 61 62 63)=data6
+    unpckhps2   xmm3, xmm5              ; xmm3=(70 71 72 73)=data7
+
+    movaps      xmm0, xmm7
+    movaps      xmm5, xmm6
+    subps       xmm7, xmm2              ; xmm7=data1-data6=tmp6
+    subps       xmm6, xmm3              ; xmm6=data0-data7=tmp7
+    addps       xmm0, xmm2              ; xmm0=data1+data6=tmp1
+    addps       xmm5, xmm3              ; xmm5=data0+data7=tmp0
+
+    movaps      xmm2, XMMWORD [wk(0)]   ; xmm2=(22 23 32 33)
+    movaps      xmm3, XMMWORD [wk(1)]   ; xmm3=(42 43 52 53)
+    movaps      XMMWORD [wk(0)], xmm7   ; wk(0)=tmp6
+    movaps      XMMWORD [wk(1)], xmm6   ; wk(1)=tmp7
+
+    movaps      xmm7, xmm4              ; transpose coefficients(phase 2)
+    unpcklps2   xmm4, xmm2              ; xmm4=(20 21 22 23)=data2
+    unpckhps2   xmm7, xmm2              ; xmm7=(30 31 32 33)=data3
+    movaps      xmm6, xmm1              ; transpose coefficients(phase 2)
+    unpcklps2   xmm1, xmm3              ; xmm1=(40 41 42 43)=data4
+    unpckhps2   xmm6, xmm3              ; xmm6=(50 51 52 53)=data5
+
+    movaps      xmm2, xmm7
+    movaps      xmm3, xmm4
+    addps       xmm7, xmm1              ; xmm7=data3+data4=tmp3
+    addps       xmm4, xmm6              ; xmm4=data2+data5=tmp2
+    subps       xmm2, xmm1              ; xmm2=data3-data4=tmp4
+    subps       xmm3, xmm6              ; xmm3=data2-data5=tmp5
+
+    ; -- Even part
+
+    movaps      xmm1, xmm5
+    movaps      xmm6, xmm0
+    subps       xmm5, xmm7              ; xmm5=tmp13
+    subps       xmm0, xmm4              ; xmm0=tmp12
+    addps       xmm1, xmm7              ; xmm1=tmp10
+    addps       xmm6, xmm4              ; xmm6=tmp11
+
+    addps       xmm0, xmm5
+    mulps       xmm0, [rel PD_0_707]    ; xmm0=z1
+
+    movaps      xmm7, xmm1
+    movaps      xmm4, xmm5
+    subps       xmm1, xmm6              ; xmm1=data4
+    subps       xmm5, xmm0              ; xmm5=data6
+    addps       xmm7, xmm6              ; xmm7=data0
+    addps       xmm4, xmm0              ; xmm4=data2
+
+    movaps      XMMWORD [XMMBLOCK(4,0,rdx,SIZEOF_FAST_FLOAT)], xmm1
+    movaps      XMMWORD [XMMBLOCK(6,0,rdx,SIZEOF_FAST_FLOAT)], xmm5
+    movaps      XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_FAST_FLOAT)], xmm7
+    movaps      XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_FAST_FLOAT)], xmm4
+
+    ; -- Odd part
+
+    movaps      xmm6, XMMWORD [wk(0)]   ; xmm6=tmp6
+    movaps      xmm0, XMMWORD [wk(1)]   ; xmm0=tmp7
+
+    addps       xmm2, xmm3              ; xmm2=tmp10
+    addps       xmm3, xmm6              ; xmm3=tmp11
+    addps       xmm6, xmm0              ; xmm6=tmp12, xmm0=tmp7
+
+    mulps       xmm3, [rel PD_0_707]    ; xmm3=z3
+
+    movaps      xmm1, xmm2              ; xmm1=tmp10
+    subps       xmm2, xmm6
+    mulps       xmm2, [rel PD_0_382]    ; xmm2=z5
+    mulps       xmm1, [rel PD_0_541]    ; xmm1=MULTIPLY(tmp10,FIX_0_541196)
+    mulps       xmm6, [rel PD_1_306]    ; xmm6=MULTIPLY(tmp12,FIX_1_306562)
+    addps       xmm1, xmm2              ; xmm1=z2
+    addps       xmm6, xmm2              ; xmm6=z4
+
+    movaps      xmm5, xmm0
+    subps       xmm0, xmm3              ; xmm0=z13
+    addps       xmm5, xmm3              ; xmm5=z11
+
+    movaps      xmm7, xmm0
+    movaps      xmm4, xmm5
+    subps       xmm0, xmm1              ; xmm0=data3
+    subps       xmm5, xmm6              ; xmm5=data7
+    addps       xmm7, xmm1              ; xmm7=data5
+    addps       xmm4, xmm6              ; xmm4=data1
+
+    movaps      XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_FAST_FLOAT)], xmm0
+    movaps      XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_FAST_FLOAT)], xmm5
+    movaps      XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_FAST_FLOAT)], xmm7
+    movaps      XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_FAST_FLOAT)], xmm4
+
+    add         rdx, byte 4*SIZEOF_FAST_FLOAT
+    dec         rcx
+    jnz         near .columnloop
+
+    uncollect_args 1
+    mov         rsp, rbp                ; rsp <- aligned rbp
+    pop         rsp                     ; rsp <- original rbp
+    pop         rbp
+    ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+    align       32
diff --git a/media/libjpeg/simd/x86_64/jfdctfst-sse2.asm b/media/libjpeg/simd/x86_64/jfdctfst-sse2.asm
new file mode 100644
index 0000000000..2e1bfe6e8c
--- /dev/null
+++ b/media/libjpeg/simd/x86_64/jfdctfst-sse2.asm
@@ -0,0 +1,389 @@
+;
+; jfdctfst.asm - fast integer FDCT (64-bit SSE2)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+; Copyright (C) 2009, 2016, D. R. Commander.
+;
+; Based on the x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; This file contains a fast, not so accurate integer implementation of
+; the forward DCT (Discrete Cosine Transform). The following code is
+; based directly on the IJG's original jfdctfst.c; see the jfdctfst.c
+; for more details.
+
+%include "jsimdext.inc"
+%include "jdct.inc"
+
+; --------------------------------------------------------------------------
+
+%define CONST_BITS  8  ; 14 is also OK.
+
+%if CONST_BITS == 8
+F_0_382 equ  98  ; FIX(0.382683433)
+F_0_541 equ 139  ; FIX(0.541196100)
+F_0_707 equ 181  ; FIX(0.707106781)
+F_1_306 equ 334  ; FIX(1.306562965)
+%else
+; NASM cannot do compile-time arithmetic on floating-point constants.
+%define DESCALE(x, n)  (((x) + (1 << ((n) - 1))) >> (n))
+F_0_382 equ DESCALE( 410903207, 30 - CONST_BITS)  ; FIX(0.382683433)
+F_0_541 equ DESCALE( 581104887, 30 - CONST_BITS)  ; FIX(0.541196100)
+F_0_707 equ DESCALE( 759250124, 30 - CONST_BITS)  ; FIX(0.707106781)
+F_1_306 equ DESCALE(1402911301, 30 - CONST_BITS)  ; FIX(1.306562965)
+%endif
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_CONST
+
+; PRE_MULTIPLY_SCALE_BITS <= 2 (to avoid overflow)
+; CONST_BITS + CONST_SHIFT + PRE_MULTIPLY_SCALE_BITS == 16 (for pmulhw)
+
+%define PRE_MULTIPLY_SCALE_BITS  2
+%define CONST_SHIFT              (16 - PRE_MULTIPLY_SCALE_BITS - CONST_BITS)
+
+    alignz      32
+    GLOBAL_DATA(jconst_fdct_ifast_sse2)
+
+EXTN(jconst_fdct_ifast_sse2):
+
+PW_F0707 times 8 dw F_0_707 << CONST_SHIFT
+PW_F0382 times 8 dw F_0_382 << CONST_SHIFT
+PW_F0541 times 8 dw F_0_541 << CONST_SHIFT
+PW_F1306 times 8 dw F_1_306 << CONST_SHIFT
+
+    alignz      32
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_TEXT
+    BITS        64
+;
+; Perform the forward DCT on one block of samples.
+;
+; GLOBAL(void)
+; jsimd_fdct_ifast_sse2(DCTELEM *data)
+;
+
+; r10 = DCTELEM *data
+
+%define wk(i)   rbp - (WK_NUM - (i)) * SIZEOF_XMMWORD  ; xmmword wk[WK_NUM]
+%define WK_NUM  2
+
+    align       32
+    GLOBAL_FUNCTION(jsimd_fdct_ifast_sse2)
+
+EXTN(jsimd_fdct_ifast_sse2):
+    push        rbp
+    mov         rax, rsp                     ; rax = original rbp
+    sub         rsp, byte 4
+    and         rsp, byte (-SIZEOF_XMMWORD)  ; align to 128 bits
+    mov         [rsp], rax
+    mov         rbp, rsp                     ; rbp = aligned rbp
+    lea         rsp, [wk(0)]
+    collect_args 1
+
+    ; ---- Pass 1: process rows.
+
+    mov         rdx, r10                ; (DCTELEM *)
+
+    movdqa      xmm0, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_DCTELEM)]
+    movdqa      xmm1, XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_DCTELEM)]
+    movdqa      xmm2, XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_DCTELEM)]
+    movdqa      xmm3, XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_DCTELEM)]
+
+    ; xmm0=(00 01 02 03 04 05 06 07), xmm2=(20 21 22 23 24 25 26 27)
+    ; xmm1=(10 11 12 13 14 15 16 17), xmm3=(30 31 32 33 34 35 36 37)
+
+    movdqa      xmm4, xmm0              ; transpose coefficients(phase 1)
+    punpcklwd   xmm0, xmm1              ; xmm0=(00 10 01 11 02 12 03 13)
+    punpckhwd   xmm4, xmm1              ; xmm4=(04 14 05 15 06 16 07 17)
+    movdqa      xmm5, xmm2              ; transpose coefficients(phase 1)
+    punpcklwd   xmm2, xmm3              ; xmm2=(20 30 21 31 22 32 23 33)
+    punpckhwd   xmm5, xmm3              ; xmm5=(24 34 25 35 26 36 27 37)
+
+    movdqa      xmm6, XMMWORD [XMMBLOCK(4,0,rdx,SIZEOF_DCTELEM)]
+    movdqa      xmm7, XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_DCTELEM)]
+    movdqa      xmm1, XMMWORD [XMMBLOCK(6,0,rdx,SIZEOF_DCTELEM)]
+    movdqa      xmm3, XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_DCTELEM)]
+
+    ; xmm6=( 4 12 20 28 36 44 52 60), xmm1=( 6 14 22 30 38 46 54 62)
+    ; xmm7=( 5 13 21 29 37 45 53 61), xmm3=( 7 15 23 31 39 47 55 63)
+
+    movdqa      XMMWORD [wk(0)], xmm2   ; wk(0)=(20 30 21 31 22 32 23 33)
+    movdqa      XMMWORD [wk(1)], xmm5   ; wk(1)=(24 34 25 35 26 36 27 37)
+
+    movdqa      xmm2, xmm6              ; transpose coefficients(phase 1)
+    punpcklwd   xmm6, xmm7              ; xmm6=(40 50 41 51 42 52 43 53)
+    punpckhwd   xmm2, xmm7              ; xmm2=(44 54 45 55 46 56 47 57)
+    movdqa      xmm5, xmm1              ; transpose coefficients(phase 1)
+    punpcklwd   xmm1, xmm3              ; xmm1=(60 70 61 71 62 72 63 73)
+    punpckhwd   xmm5, xmm3              ; xmm5=(64 74 65 75 66 76 67 77)
+
+    movdqa      xmm7, xmm6              ; transpose coefficients(phase 2)
+    punpckldq   xmm6, xmm1              ; xmm6=(40 50 60 70 41 51 61 71)
+    punpckhdq   xmm7, xmm1              ; xmm7=(42 52 62 72 43 53 63 73)
+    movdqa      xmm3, xmm2              ; transpose coefficients(phase 2)
+    punpckldq   xmm2, xmm5              ; xmm2=(44 54 64 74 45 55 65 75)
+    punpckhdq   xmm3, xmm5              ; xmm3=(46 56 66 76 47 57 67 77)
+
+    movdqa      xmm1, XMMWORD [wk(0)]   ; xmm1=(20 30 21 31 22 32 23 33)
+    movdqa      xmm5, XMMWORD [wk(1)]   ; xmm5=(24 34 25 35 26 36 27 37)
+    movdqa      XMMWORD [wk(0)], xmm7   ; wk(0)=(42 52 62 72 43 53 63 73)
+    movdqa      XMMWORD [wk(1)], xmm2   ; wk(1)=(44 54 64 74 45 55 65 75)
+
+    movdqa      xmm7, xmm0              ; transpose coefficients(phase 2)
+    punpckldq   xmm0, xmm1              ; xmm0=(00 10 20 30 01 11 21 31)
+    punpckhdq   xmm7, xmm1              ; xmm7=(02 12 22 32 03 13 23 33)
+    movdqa      xmm2, xmm4              ; transpose coefficients(phase 2)
+    punpckldq   xmm4, xmm5              ; xmm4=(04 14 24 34 05 15 25 35)
+    punpckhdq   xmm2, xmm5              ; xmm2=(06 16 26 36 07 17 27 37)
+
+    movdqa      xmm1, xmm0              ; transpose coefficients(phase 3)
+    punpcklqdq  xmm0, xmm6              ; xmm0=(00 10 20 30 40 50 60 70)=data0
+    punpckhqdq  xmm1, xmm6              ; xmm1=(01 11 21 31 41 51 61 71)=data1
+    movdqa      xmm5, xmm2              ; transpose coefficients(phase 3)
+    punpcklqdq  xmm2, xmm3              ; xmm2=(06 16 26 36 46 56 66 76)=data6
+    punpckhqdq  xmm5, xmm3              ; xmm5=(07 17 27 37 47 57 67 77)=data7
+
+    movdqa      xmm6, xmm1
+    movdqa      xmm3, xmm0
+    psubw       xmm1, xmm2              ; xmm1=data1-data6=tmp6
+    psubw       xmm0, xmm5              ; xmm0=data0-data7=tmp7
+    paddw       xmm6, xmm2              ; xmm6=data1+data6=tmp1
+    paddw       xmm3, xmm5              ; xmm3=data0+data7=tmp0
+
+    movdqa      xmm2, XMMWORD [wk(0)]   ; xmm2=(42 52 62 72 43 53 63 73)
+    movdqa      xmm5, XMMWORD [wk(1)]   ; xmm5=(44 54 64 74 45 55 65 75)
+    movdqa      XMMWORD [wk(0)], xmm1   ; wk(0)=tmp6
+    movdqa      XMMWORD [wk(1)], xmm0   ; wk(1)=tmp7
+
+    movdqa      xmm1, xmm7              ; transpose coefficients(phase 3)
+    punpcklqdq  xmm7, xmm2              ; xmm7=(02 12 22 32 42 52 62 72)=data2
+    punpckhqdq  xmm1, xmm2              ; xmm1=(03 13 23 33 43 53 63 73)=data3
+    movdqa      xmm0, xmm4              ; transpose coefficients(phase 3)
+    punpcklqdq  xmm4, xmm5              ; xmm4=(04 14 24 34 44 54 64 74)=data4
+    punpckhqdq  xmm0, xmm5              ; xmm0=(05 15 25 35 45 55 65 75)=data5
+
+    movdqa      xmm2, xmm1
+    movdqa      xmm5, xmm7
+    paddw       xmm1, xmm4              ; xmm1=data3+data4=tmp3
+    paddw       xmm7, xmm0              ; xmm7=data2+data5=tmp2
+    psubw       xmm2, xmm4              ; xmm2=data3-data4=tmp4
+    psubw       xmm5, xmm0              ; xmm5=data2-data5=tmp5
+
+    ; -- Even part
+
+    movdqa      xmm4, xmm3
+    movdqa      xmm0, xmm6
+    psubw       xmm3, xmm1              ; xmm3=tmp13
+    psubw       xmm6, xmm7              ; xmm6=tmp12
+    paddw       xmm4, xmm1              ; xmm4=tmp10
+    paddw       xmm0, xmm7              ; xmm0=tmp11
+
+    paddw       xmm6, xmm3
+    psllw       xmm6, PRE_MULTIPLY_SCALE_BITS
+    pmulhw      xmm6, [rel PW_F0707]    ; xmm6=z1
+
+    movdqa      xmm1, xmm4
+    movdqa      xmm7, xmm3
+    psubw       xmm4, xmm0              ; xmm4=data4
+    psubw       xmm3, xmm6              ; xmm3=data6
+    paddw       xmm1, xmm0              ; xmm1=data0
+    paddw       xmm7, xmm6              ; xmm7=data2
+
+    movdqa      xmm0, XMMWORD [wk(0)]   ; xmm0=tmp6
+    movdqa      xmm6, XMMWORD [wk(1)]   ; xmm6=tmp7
+    movdqa      XMMWORD [wk(0)], xmm4   ; wk(0)=data4
+    movdqa      XMMWORD [wk(1)], xmm3   ; wk(1)=data6
+
+    ; -- Odd part
+
+    paddw       xmm2, xmm5              ; xmm2=tmp10
+    paddw       xmm5, xmm0              ; xmm5=tmp11
+    paddw       xmm0, xmm6              ; xmm0=tmp12, xmm6=tmp7
+
+    psllw       xmm2, PRE_MULTIPLY_SCALE_BITS
+    psllw       xmm0, PRE_MULTIPLY_SCALE_BITS
+
+    psllw       xmm5, PRE_MULTIPLY_SCALE_BITS
+    pmulhw      xmm5, [rel PW_F0707]    ; xmm5=z3
+
+    movdqa      xmm4, xmm2              ; xmm4=tmp10
+    psubw       xmm2, xmm0
+    pmulhw      xmm2, [rel PW_F0382]    ; xmm2=z5
+    pmulhw      xmm4, [rel PW_F0541]    ; xmm4=MULTIPLY(tmp10,FIX_0_541196)
+    pmulhw      xmm0, [rel PW_F1306]    ; xmm0=MULTIPLY(tmp12,FIX_1_306562)
+    paddw       xmm4, xmm2              ; xmm4=z2
+    paddw       xmm0, xmm2              ; xmm0=z4
+
+    movdqa      xmm3, xmm6
+    psubw       xmm6, xmm5              ; xmm6=z13
+    paddw       xmm3, xmm5              ; xmm3=z11
+
+    movdqa      xmm2, xmm6
+    movdqa      xmm5, xmm3
+    psubw       xmm6, xmm4              ; xmm6=data3
+    psubw       xmm3, xmm0              ; xmm3=data7
+    paddw       xmm2, xmm4              ; xmm2=data5
+    paddw       xmm5, xmm0              ; xmm5=data1
+
+    ; ---- Pass 2: process columns.
+
+    ; xmm1=(00 10 20 30 40 50 60 70), xmm7=(02 12 22 32 42 52 62 72)
+    ; xmm5=(01 11 21 31 41 51 61 71), xmm6=(03 13 23 33 43 53 63 73)
+
+    movdqa      xmm4, xmm1              ; transpose coefficients(phase 1)
+    punpcklwd   xmm1, xmm5              ; xmm1=(00 01 10 11 20 21 30 31)
+    punpckhwd   xmm4, xmm5              ; xmm4=(40 41 50 51 60 61 70 71)
+    movdqa      xmm0, xmm7              ; transpose coefficients(phase 1)
+    punpcklwd   xmm7, xmm6              ; xmm7=(02 03 12 13 22 23 32 33)
+    punpckhwd   xmm0, xmm6              ; xmm0=(42 43 52 53 62 63 72 73)
+
+    movdqa      xmm5, XMMWORD [wk(0)]   ; xmm5=col4
+    movdqa      xmm6, XMMWORD [wk(1)]   ; xmm6=col6
+
+    ; xmm5=(04 14 24 34 44 54 64 74), xmm6=(06 16 26 36 46 56 66 76)
+    ; xmm2=(05 15 25 35 45 55 65 75), xmm3=(07 17 27 37 47 57 67 77)
+
+    movdqa      XMMWORD [wk(0)], xmm7   ; wk(0)=(02 03 12 13 22 23 32 33)
+    movdqa      XMMWORD [wk(1)], xmm0   ; wk(1)=(42 43 52 53 62 63 72 73)
+
+    movdqa      xmm7, xmm5              ; transpose coefficients(phase 1)
+    punpcklwd   xmm5, xmm2              ; xmm5=(04 05 14 15 24 25 34 35)
+    punpckhwd   xmm7, xmm2              ; xmm7=(44 45 54 55 64 65 74 75)
+    movdqa      xmm0, xmm6              ; transpose coefficients(phase 1)
+    punpcklwd   xmm6, xmm3              ; xmm6=(06 07 16 17 26 27 36 37)
+    punpckhwd   xmm0, xmm3              ; xmm0=(46 47 56 57 66 67 76 77)
+
+    movdqa      xmm2, xmm5              ; transpose coefficients(phase 2)
+    punpckldq   xmm5, xmm6              ; xmm5=(04 05 06 07 14 15 16 17)
+    punpckhdq   xmm2, xmm6              ; xmm2=(24 25 26 27 34 35 36 37)
+    movdqa      xmm3, xmm7              ; transpose coefficients(phase 2)
+    punpckldq   xmm7, xmm0              ; xmm7=(44 45 46 47 54 55 56 57)
+    punpckhdq   xmm3, xmm0              ; xmm3=(64 65 66 67 74 75 76 77)
+
+    movdqa      xmm6, XMMWORD [wk(0)]   ; xmm6=(02 03 12 13 22 23 32 33)
+    movdqa      xmm0, XMMWORD [wk(1)]   ; xmm0=(42 43 52 53 62 63 72 73)
+    movdqa      XMMWORD [wk(0)], xmm2   ; wk(0)=(24 25 26 27 34 35 36 37)
+    movdqa      XMMWORD [wk(1)], xmm7   ; wk(1)=(44 45 46 47 54 55 56 57)
+
+    movdqa      xmm2, xmm1              ; transpose coefficients(phase 2)
+    punpckldq   xmm1, xmm6              ; xmm1=(00 01 02 03 10 11 12 13)
+    punpckhdq   xmm2, xmm6              ; xmm2=(20 21 22 23 30 31 32 33)
+    movdqa      xmm7, xmm4              ; transpose coefficients(phase 2)
+    punpckldq   xmm4, xmm0              ; xmm4=(40 41 42 43 50 51 52 53)
+    punpckhdq   xmm7, xmm0              ; xmm7=(60 61 62 63 70 71 72 73)
+
+    movdqa      xmm6, xmm1              ; transpose coefficients(phase 3)
+    punpcklqdq  xmm1, xmm5              ; xmm1=(00 01 02 03 04 05 06 07)=data0
+    punpckhqdq  xmm6, xmm5              ; xmm6=(10 11 12 13 14 15 16 17)=data1
+    movdqa      xmm0, xmm7              ; transpose coefficients(phase 3)
+    punpcklqdq  xmm7, xmm3              ; xmm7=(60 61 62 63 64 65 66 67)=data6
+    punpckhqdq  xmm0, xmm3              ; xmm0=(70 71 72 73 74 75 76 77)=data7
+
+    movdqa      xmm5, xmm6
+    movdqa      xmm3, xmm1
+    psubw       xmm6, xmm7              ; xmm6=data1-data6=tmp6
+    psubw       xmm1, xmm0              ; xmm1=data0-data7=tmp7
+    paddw       xmm5, xmm7              ; xmm5=data1+data6=tmp1
+    paddw       xmm3, xmm0              ; xmm3=data0+data7=tmp0
+
+    movdqa      xmm7, XMMWORD [wk(0)]   ; xmm7=(24 25 26 27 34 35 36 37)
+    movdqa      xmm0, XMMWORD [wk(1)]   ; xmm0=(44 45 46 47 54 55 56 57)
+    movdqa      XMMWORD [wk(0)], xmm6   ; wk(0)=tmp6
+    movdqa      XMMWORD [wk(1)], xmm1   ; wk(1)=tmp7
+
+    movdqa      xmm6, xmm2              ; transpose coefficients(phase 3)
+    punpcklqdq  xmm2, xmm7              ; xmm2=(20 21 22 23 24 25 26 27)=data2
+    punpckhqdq  xmm6, xmm7              ; xmm6=(30 31 32 33 34 35 36 37)=data3
+    movdqa      xmm1, xmm4              ; transpose coefficients(phase 3)
+    punpcklqdq  xmm4, xmm0              ; xmm4=(40 41 42 43 44 45 46 47)=data4
+    punpckhqdq  xmm1, xmm0              ; xmm1=(50 51 52 53 54 55 56 57)=data5
+
+    movdqa      xmm7, xmm6
+    movdqa      xmm0, xmm2
+    paddw       xmm6, xmm4              ; xmm6=data3+data4=tmp3
+    paddw       xmm2, xmm1              ; xmm2=data2+data5=tmp2
+    psubw       xmm7, xmm4              ; xmm7=data3-data4=tmp4
+    psubw       xmm0, xmm1              ; xmm0=data2-data5=tmp5
+
+    ; -- Even part
+
+    movdqa      xmm4, xmm3
+    movdqa      xmm1, xmm5
+    psubw       xmm3, xmm6              ; xmm3=tmp13
+    psubw       xmm5, xmm2              ; xmm5=tmp12
+    paddw       xmm4, xmm6              ; xmm4=tmp10
+    paddw       xmm1, xmm2              ; xmm1=tmp11
+
+    paddw       xmm5, xmm3
+    psllw       xmm5, PRE_MULTIPLY_SCALE_BITS
+    pmulhw      xmm5, [rel PW_F0707]    ; xmm5=z1
+
+    movdqa      xmm6, xmm4
+    movdqa      xmm2, xmm3
+    psubw       xmm4, xmm1              ; xmm4=data4
+    psubw       xmm3, xmm5              ; xmm3=data6
+    paddw       xmm6, xmm1              ; xmm6=data0
+    paddw       xmm2, xmm5              ; xmm2=data2
+
+    movdqa      XMMWORD [XMMBLOCK(4,0,rdx,SIZEOF_DCTELEM)], xmm4
+    movdqa      XMMWORD [XMMBLOCK(6,0,rdx,SIZEOF_DCTELEM)], xmm3
+    movdqa      XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_DCTELEM)], xmm6
+    movdqa      XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_DCTELEM)], xmm2
+
+    ; -- Odd part
+
+    movdqa      xmm1, XMMWORD [wk(0)]   ; xmm1=tmp6
+    movdqa      xmm5, XMMWORD [wk(1)]   ; xmm5=tmp7
+
+    paddw       xmm7, xmm0              ; xmm7=tmp10
+    paddw       xmm0, xmm1              ; xmm0=tmp11
+    paddw       xmm1, xmm5              ; xmm1=tmp12, xmm5=tmp7
+
+    psllw       xmm7, PRE_MULTIPLY_SCALE_BITS
+    psllw       xmm1, PRE_MULTIPLY_SCALE_BITS
+
+    psllw       xmm0, PRE_MULTIPLY_SCALE_BITS
+    pmulhw      xmm0, [rel PW_F0707]    ; xmm0=z3
+
+    movdqa      xmm4, xmm7              ; xmm4=tmp10
+    psubw       xmm7, xmm1
+    pmulhw      xmm7, [rel PW_F0382]    ; xmm7=z5
+    pmulhw      xmm4, [rel PW_F0541]    ; xmm4=MULTIPLY(tmp10,FIX_0_541196)
+    pmulhw      xmm1, [rel PW_F1306]    ; xmm1=MULTIPLY(tmp12,FIX_1_306562)
+    paddw       xmm4, xmm7              ; xmm4=z2
+    paddw       xmm1, xmm7              ; xmm1=z4
+
+    movdqa      xmm3, xmm5
+    psubw       xmm5, xmm0              ; xmm5=z13
+    paddw       xmm3, xmm0              ; xmm3=z11
+
+    movdqa      xmm6, xmm5
+    movdqa      xmm2, xmm3
+    psubw       xmm5, xmm4              ; xmm5=data3
+    psubw       xmm3, xmm1              ; xmm3=data7
+    paddw       xmm6, xmm4              ; xmm6=data5
+    paddw       xmm2, xmm1              ; xmm2=data1
+
+    movdqa      XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_DCTELEM)], xmm5
+    movdqa      XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_DCTELEM)], xmm3
+    movdqa      XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_DCTELEM)], xmm6
+    movdqa      XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_DCTELEM)], xmm2
+
+    uncollect_args 1
+    mov         rsp, rbp                ; rsp <- aligned rbp
+    pop         rsp                     ; rsp <- original rbp
+    pop         rbp
+    ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+    align       32
diff --git a/media/libjpeg/simd/x86_64/jfdctint-avx2.asm b/media/libjpeg/simd/x86_64/jfdctint-avx2.asm
new file mode 100644
index 0000000000..e56258b48a
--- /dev/null
+++ b/media/libjpeg/simd/x86_64/jfdctint-avx2.asm
@@ -0,0 +1,320 @@
+;
+; jfdctint.asm - accurate integer FDCT (64-bit AVX2)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+; Copyright (C) 2009, 2016, 2018, 2020, D. R. Commander.
+;
+; Based on the x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; This file contains a slower but more accurate integer implementation of the
+; forward DCT (Discrete Cosine Transform). The following code is based
+; directly on the IJG's original jfdctint.c; see the jfdctint.c for
+; more details.
+
+%include "jsimdext.inc"
+%include "jdct.inc"
+
+; --------------------------------------------------------------------------
+
+%define CONST_BITS  13
+%define PASS1_BITS  2
+
+%define DESCALE_P1  (CONST_BITS - PASS1_BITS)
+%define DESCALE_P2  (CONST_BITS + PASS1_BITS)
+
+%if CONST_BITS == 13
+F_0_298 equ  2446  ; FIX(0.298631336)
+F_0_390 equ  3196  ; FIX(0.390180644)
+F_0_541 equ  4433  ; FIX(0.541196100)
+F_0_765 equ  6270  ; FIX(0.765366865)
+F_0_899 equ  7373  ; FIX(0.899976223)
+F_1_175 equ  9633  ; FIX(1.175875602)
+F_1_501 equ 12299  ; FIX(1.501321110)
+F_1_847 equ 15137  ; FIX(1.847759065)
+F_1_961 equ 16069  ; FIX(1.961570560)
+F_2_053 equ 16819  ; FIX(2.053119869)
+F_2_562 equ 20995  ; FIX(2.562915447)
+F_3_072 equ 25172  ; FIX(3.072711026)
+%else
+; NASM cannot do compile-time arithmetic on floating-point constants.
+%define DESCALE(x, n)  (((x) + (1 << ((n) - 1))) >> (n))
+F_0_298 equ DESCALE( 320652955, 30 - CONST_BITS)  ; FIX(0.298631336)
+F_0_390 equ DESCALE( 418953276, 30 - CONST_BITS)  ; FIX(0.390180644)
+F_0_541 equ DESCALE( 581104887, 30 - CONST_BITS)  ; FIX(0.541196100)
+F_0_765 equ DESCALE( 821806413, 30 - CONST_BITS)  ; FIX(0.765366865)
+F_0_899 equ DESCALE( 966342111, 30 - CONST_BITS)  ; FIX(0.899976223)
+F_1_175 equ DESCALE(1262586813, 30 - CONST_BITS)  ; FIX(1.175875602)
+F_1_501 equ DESCALE(1612031267, 30 - CONST_BITS)  ; FIX(1.501321110)
+F_1_847 equ DESCALE(1984016188, 30 - CONST_BITS)  ; FIX(1.847759065)
+F_1_961 equ DESCALE(2106220350, 30 - CONST_BITS)  ; FIX(1.961570560)
+F_2_053 equ DESCALE(2204520673, 30 - CONST_BITS)  ; FIX(2.053119869)
+F_2_562 equ DESCALE(2751909506, 30 - CONST_BITS)  ; FIX(2.562915447)
+F_3_072 equ DESCALE(3299298341, 30 - CONST_BITS)  ; FIX(3.072711026)
+%endif
+
+; --------------------------------------------------------------------------
+; In-place 8x8x16-bit matrix transpose using AVX2 instructions
+; %1-%4: Input/output registers
+; %5-%8: Temp registers
+
+%macro dotranspose 8
+    ; %1=(00 01 02 03 04 05 06 07  40 41 42 43 44 45 46 47)
+    ; %2=(10 11 12 13 14 15 16 17  50 51 52 53 54 55 56 57)
+    ; %3=(20 21 22 23 24 25 26 27  60 61 62 63 64 65 66 67)
+    ; %4=(30 31 32 33 34 35 36 37  70 71 72 73 74 75 76 77)
+
+    vpunpcklwd  %5, %1, %2
+    vpunpckhwd  %6, %1, %2
+    vpunpcklwd  %7, %3, %4
+    vpunpckhwd  %8, %3, %4
+    ; transpose coefficients(phase 1)
+    ; %5=(00 10 01 11 02 12 03 13  40 50 41 51 42 52 43 53)
+    ; %6=(04 14 05 15 06 16 07 17  44 54 45 55 46 56 47 57)
+    ; %7=(20 30 21 31 22 32 23 33  60 70 61 71 62 72 63 73)
+    ; %8=(24 34 25 35 26 36 27 37  64 74 65 75 66 76 67 77)
+
+    vpunpckldq  %1, %5, %7
+    vpunpckhdq  %2, %5, %7
+    vpunpckldq  %3, %6, %8
+    vpunpckhdq  %4, %6, %8
+    ; transpose coefficients(phase 2)
+    ; %1=(00 10 20 30 01 11 21 31  40 50 60 70 41 51 61 71)
+    ; %2=(02 12 22 32 03 13 23 33  42 52 62 72 43 53 63 73)
+    ; %3=(04 14 24 34 05 15 25 35  44 54 64 74 45 55 65 75)
+    ; %4=(06 16 26 36 07 17 27 37  46 56 66 76 47 57 67 77)
+
+    vpermq      %1, %1, 0x8D
+    vpermq      %2, %2, 0x8D
+    vpermq      %3, %3, 0xD8
+    vpermq      %4, %4, 0xD8
+    ; transpose coefficients(phase 3)
+    ; %1=(01 11 21 31 41 51 61 71  00 10 20 30 40 50 60 70)
+    ; %2=(03 13 23 33 43 53 63 73  02 12 22 32 42 52 62 72)
+    ; %3=(04 14 24 34 44 54 64 74  05 15 25 35 45 55 65 75)
+    ; %4=(06 16 26 36 46 56 66 76  07 17 27 37 47 57 67 77)
+%endmacro
+
+; --------------------------------------------------------------------------
+; In-place 8x8x16-bit accurate integer forward DCT using AVX2 instructions
+; %1-%4: Input/output registers
+; %5-%8: Temp registers
+; %9:    Pass (1 or 2)
+
+%macro dodct 9
+    vpsubw      %5, %1, %4              ; %5=data1_0-data6_7=tmp6_7
+    vpaddw      %6, %1, %4              ; %6=data1_0+data6_7=tmp1_0
+    vpaddw      %7, %2, %3              ; %7=data3_2+data4_5=tmp3_2
+    vpsubw      %8, %2, %3              ; %8=data3_2-data4_5=tmp4_5
+
+    ; -- Even part
+
+    vperm2i128  %6, %6, %6, 0x01        ; %6=tmp0_1
+    vpaddw      %1, %6, %7              ; %1=tmp0_1+tmp3_2=tmp10_11
+    vpsubw      %6, %6, %7              ; %6=tmp0_1-tmp3_2=tmp13_12
+
+    vperm2i128  %7, %1, %1, 0x01        ; %7=tmp11_10
+    vpsignw     %1, %1, [rel PW_1_NEG1]  ; %1=tmp10_neg11
+    vpaddw      %7, %7, %1              ; %7=(tmp10+tmp11)_(tmp10-tmp11)
+%if %9 == 1
+    vpsllw      %1, %7, PASS1_BITS      ; %1=data0_4
+%else
+    vpaddw      %7, %7, [rel PW_DESCALE_P2X]
+    vpsraw      %1, %7, PASS1_BITS      ; %1=data0_4
+%endif
+
+    ; (Original)
+    ; z1 = (tmp12 + tmp13) * 0.541196100;
+    ; data2 = z1 + tmp13 * 0.765366865;
+    ; data6 = z1 + tmp12 * -1.847759065;
+    ;
+    ; (This implementation)
+    ; data2 = tmp13 * (0.541196100 + 0.765366865) + tmp12 * 0.541196100;
+    ; data6 = tmp13 * 0.541196100 + tmp12 * (0.541196100 - 1.847759065);
+
+    vperm2i128  %7, %6, %6, 0x01        ; %7=tmp12_13
+    vpunpcklwd  %2, %6, %7
+    vpunpckhwd  %6, %6, %7
+    vpmaddwd    %2, %2, [rel PW_F130_F054_MF130_F054]  ; %2=data2_6L
+    vpmaddwd    %6, %6, [rel PW_F130_F054_MF130_F054]  ; %6=data2_6H
+
+    vpaddd      %2, %2, [rel PD_DESCALE_P %+ %9]
+    vpaddd      %6, %6, [rel PD_DESCALE_P %+ %9]
+    vpsrad      %2, %2, DESCALE_P %+ %9
+    vpsrad      %6, %6, DESCALE_P %+ %9
+
+    vpackssdw   %3, %2, %6              ; %6=data2_6
+
+    ; -- Odd part
+
+    vpaddw      %7, %8, %5              ; %7=tmp4_5+tmp6_7=z3_4
+
+    ; (Original)
+    ; z5 = (z3 + z4) * 1.175875602;
+    ; z3 = z3 * -1.961570560;  z4 = z4 * -0.390180644;
+    ; z3 += z5;  z4 += z5;
+    ;
+    ; (This implementation)
+    ; z3 = z3 * (1.175875602 - 1.961570560) + z4 * 1.175875602;
+    ; z4 = z3 * 1.175875602 + z4 * (1.175875602 - 0.390180644);
+
+    vperm2i128  %2, %7, %7, 0x01        ; %2=z4_3
+    vpunpcklwd  %6, %7, %2
+    vpunpckhwd  %7, %7, %2
+    vpmaddwd    %6, %6, [rel PW_MF078_F117_F078_F117]  ; %6=z3_4L
+    vpmaddwd    %7, %7, [rel PW_MF078_F117_F078_F117]  ; %7=z3_4H
+
+    ; (Original)
+    ; z1 = tmp4 + tmp7;  z2 = tmp5 + tmp6;
+    ; tmp4 = tmp4 * 0.298631336;  tmp5 = tmp5 * 2.053119869;
+    ; tmp6 = tmp6 * 3.072711026;  tmp7 = tmp7 * 1.501321110;
+    ; z1 = z1 * -0.899976223;  z2 = z2 * -2.562915447;
+    ; data7 = tmp4 + z1 + z3;  data5 = tmp5 + z2 + z4;
+    ; data3 = tmp6 + z2 + z3;  data1 = tmp7 + z1 + z4;
+    ;
+    ; (This implementation)
+    ; tmp4 = tmp4 * (0.298631336 - 0.899976223) + tmp7 * -0.899976223;
+    ; tmp5 = tmp5 * (2.053119869 - 2.562915447) + tmp6 * -2.562915447;
+    ; tmp6 = tmp5 * -2.562915447 + tmp6 * (3.072711026 - 2.562915447);
+    ; tmp7 = tmp4 * -0.899976223 + tmp7 * (1.501321110 - 0.899976223);
+    ; data7 = tmp4 + z3;  data5 = tmp5 + z4;
+    ; data3 = tmp6 + z3;  data1 = tmp7 + z4;
+
+    vperm2i128  %4, %5, %5, 0x01        ; %4=tmp7_6
+    vpunpcklwd  %2, %8, %4
+    vpunpckhwd  %4, %8, %4
+    vpmaddwd    %2, %2, [rel PW_MF060_MF089_MF050_MF256]  ; %2=tmp4_5L
+    vpmaddwd    %4, %4, [rel PW_MF060_MF089_MF050_MF256]  ; %4=tmp4_5H
+
+    vpaddd      %2, %2, %6              ; %2=data7_5L
+    vpaddd      %4, %4, %7              ; %4=data7_5H
+
+    vpaddd      %2, %2, [rel PD_DESCALE_P %+ %9]
+    vpaddd      %4, %4, [rel PD_DESCALE_P %+ %9]
+    vpsrad      %2, %2, DESCALE_P %+ %9
+    vpsrad      %4, %4, DESCALE_P %+ %9
+
+    vpackssdw   %4, %2, %4              ; %4=data7_5
+
+    vperm2i128  %2, %8, %8, 0x01        ; %2=tmp5_4
+    vpunpcklwd  %8, %5, %2
+    vpunpckhwd  %5, %5, %2
+    vpmaddwd    %8, %8, [rel PW_F050_MF256_F060_MF089]  ; %8=tmp6_7L
+    vpmaddwd    %5, %5, [rel PW_F050_MF256_F060_MF089]  ; %5=tmp6_7H
+
+    vpaddd      %8, %8, %6              ; %8=data3_1L
+    vpaddd      %5, %5, %7              ; %5=data3_1H
+
+    vpaddd      %8, %8, [rel PD_DESCALE_P %+ %9]
+    vpaddd      %5, %5, [rel PD_DESCALE_P %+ %9]
+    vpsrad      %8, %8, DESCALE_P %+ %9
+    vpsrad      %5, %5, DESCALE_P %+ %9
+
+    vpackssdw   %2, %8, %5              ; %2=data3_1
+%endmacro
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_CONST
+
+    alignz      32
+    GLOBAL_DATA(jconst_fdct_islow_avx2)
+
+EXTN(jconst_fdct_islow_avx2):
+
+PW_F130_F054_MF130_F054    times 4  dw  (F_0_541 + F_0_765),  F_0_541
+                           times 4  dw  (F_0_541 - F_1_847),  F_0_541
+PW_MF078_F117_F078_F117    times 4  dw  (F_1_175 - F_1_961),  F_1_175
+                           times 4  dw  (F_1_175 - F_0_390),  F_1_175
+PW_MF060_MF089_MF050_MF256 times 4  dw  (F_0_298 - F_0_899), -F_0_899
+                           times 4  dw  (F_2_053 - F_2_562), -F_2_562
+PW_F050_MF256_F060_MF089   times 4  dw  (F_3_072 - F_2_562), -F_2_562
+                           times 4  dw  (F_1_501 - F_0_899), -F_0_899
+PD_DESCALE_P1              times 8  dd  1 << (DESCALE_P1 - 1)
+PD_DESCALE_P2              times 8  dd  1 << (DESCALE_P2 - 1)
+PW_DESCALE_P2X             times 16 dw  1 << (PASS1_BITS - 1)
+PW_1_NEG1                  times 8  dw  1
+                           times 8  dw -1
+
+    alignz      32
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_TEXT
+    BITS        64
+;
+; Perform the forward DCT on one block of samples.
+;
+; GLOBAL(void)
+; jsimd_fdct_islow_avx2(DCTELEM *data)
+;
+
+; r10 = DCTELEM *data
+
+    align       32
+    GLOBAL_FUNCTION(jsimd_fdct_islow_avx2)
+
+EXTN(jsimd_fdct_islow_avx2):
+    push        rbp
+    mov         rax, rsp
+    mov         rbp, rsp
+    collect_args 1
+
+    ; ---- Pass 1: process rows.
+
+    vmovdqu     ymm4, YMMWORD [YMMBLOCK(0,0,r10,SIZEOF_DCTELEM)]
+    vmovdqu     ymm5, YMMWORD [YMMBLOCK(2,0,r10,SIZEOF_DCTELEM)]
+    vmovdqu     ymm6, YMMWORD [YMMBLOCK(4,0,r10,SIZEOF_DCTELEM)]
+    vmovdqu     ymm7, YMMWORD [YMMBLOCK(6,0,r10,SIZEOF_DCTELEM)]
+    ; ymm4=(00 01 02 03 04 05 06 07  10 11 12 13 14 15 16 17)
+    ; ymm5=(20 21 22 23 24 25 26 27  30 31 32 33 34 35 36 37)
+    ; ymm6=(40 41 42 43 44 45 46 47  50 51 52 53 54 55 56 57)
+    ; ymm7=(60 61 62 63 64 65 66 67  70 71 72 73 74 75 76 77)
+
+    vperm2i128  ymm0, ymm4, ymm6, 0x20
+    vperm2i128  ymm1, ymm4, ymm6, 0x31
+    vperm2i128  ymm2, ymm5, ymm7, 0x20
+    vperm2i128  ymm3, ymm5, ymm7, 0x31
+    ; ymm0=(00 01 02 03 04 05 06 07  40 41 42 43 44 45 46 47)
+    ; ymm1=(10 11 12 13 14 15 16 17  50 51 52 53 54 55 56 57)
+    ; ymm2=(20 21 22 23 24 25 26 27  60 61 62 63 64 65 66 67)
+    ; ymm3=(30 31 32 33 34 35 36 37  70 71 72 73 74 75 76 77)
+
+    dotranspose ymm0, ymm1, ymm2, ymm3, ymm4, ymm5, ymm6, ymm7
+
+    dodct       ymm0, ymm1, ymm2, ymm3, ymm4, ymm5, ymm6, ymm7, 1
+    ; ymm0=data0_4, ymm1=data3_1, ymm2=data2_6, ymm3=data7_5
+
+    ; ---- Pass 2: process columns.
+
+    vperm2i128  ymm4, ymm1, ymm3, 0x20  ; ymm4=data3_7
+    vperm2i128  ymm1, ymm1, ymm3, 0x31  ; ymm1=data1_5
+
+    dotranspose ymm0, ymm1, ymm2, ymm4, ymm3, ymm5, ymm6, ymm7
+
+    dodct       ymm0, ymm1, ymm2, ymm4, ymm3, ymm5, ymm6, ymm7, 2
+    ; ymm0=data0_4, ymm1=data3_1, ymm2=data2_6, ymm4=data7_5
+
+    vperm2i128 ymm3, ymm0, ymm1, 0x30   ; ymm3=data0_1
+    vperm2i128 ymm5, ymm2, ymm1, 0x20   ; ymm5=data2_3
+    vperm2i128 ymm6, ymm0, ymm4, 0x31   ; ymm6=data4_5
+    vperm2i128 ymm7, ymm2, ymm4, 0x21   ; ymm7=data6_7
+
+    vmovdqu     YMMWORD [YMMBLOCK(0,0,r10,SIZEOF_DCTELEM)], ymm3
+    vmovdqu     YMMWORD [YMMBLOCK(2,0,r10,SIZEOF_DCTELEM)], ymm5
+    vmovdqu     YMMWORD [YMMBLOCK(4,0,r10,SIZEOF_DCTELEM)], ymm6
+    vmovdqu     YMMWORD [YMMBLOCK(6,0,r10,SIZEOF_DCTELEM)], ymm7
+
+    vzeroupper
+    uncollect_args 1
+    pop         rbp
+    ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+    align       32
diff --git a/media/libjpeg/simd/x86_64/jfdctint-sse2.asm b/media/libjpeg/simd/x86_64/jfdctint-sse2.asm
new file mode 100644
index 0000000000..ec1f383ccb
--- /dev/null
+++ b/media/libjpeg/simd/x86_64/jfdctint-sse2.asm
@@ -0,0 +1,619 @@
+;
+; jfdctint.asm - accurate integer FDCT (64-bit SSE2)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+; Copyright (C) 2009, 2016, 2020, D. R. Commander.
+;
+; Based on the x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; This file contains a slower but more accurate integer implementation of the
+; forward DCT (Discrete Cosine Transform). The following code is based
+; directly on the IJG's original jfdctint.c; see the jfdctint.c for
+; more details.
+
+%include "jsimdext.inc"
+%include "jdct.inc"
+
+; --------------------------------------------------------------------------
+
+%define CONST_BITS  13
+%define PASS1_BITS  2
+
+%define DESCALE_P1  (CONST_BITS - PASS1_BITS)
+%define DESCALE_P2  (CONST_BITS + PASS1_BITS)
+
+%if CONST_BITS == 13
+F_0_298 equ  2446  ; FIX(0.298631336)
+F_0_390 equ  3196  ; FIX(0.390180644)
+F_0_541 equ  4433  ; FIX(0.541196100)
+F_0_765 equ  6270  ; FIX(0.765366865)
+F_0_899 equ  7373  ; FIX(0.899976223)
+F_1_175 equ  9633  ; FIX(1.175875602)
+F_1_501 equ 12299  ; FIX(1.501321110)
+F_1_847 equ 15137  ; FIX(1.847759065)
+F_1_961 equ 16069  ; FIX(1.961570560)
+F_2_053 equ 16819  ; FIX(2.053119869)
+F_2_562 equ 20995  ; FIX(2.562915447)
+F_3_072 equ 25172  ; FIX(3.072711026)
+%else
+; NASM cannot do compile-time arithmetic on floating-point constants.
+%define DESCALE(x, n)  (((x) + (1 << ((n) - 1))) >> (n))
+F_0_298 equ DESCALE( 320652955, 30 - CONST_BITS)  ; FIX(0.298631336)
+F_0_390 equ DESCALE( 418953276, 30 - CONST_BITS)  ; FIX(0.390180644)
+F_0_541 equ DESCALE( 581104887, 30 - CONST_BITS)  ; FIX(0.541196100)
+F_0_765 equ DESCALE( 821806413, 30 - CONST_BITS)  ; FIX(0.765366865)
+F_0_899 equ DESCALE( 966342111, 30 - CONST_BITS)  ; FIX(0.899976223)
+F_1_175 equ DESCALE(1262586813, 30 - CONST_BITS)  ; FIX(1.175875602)
+F_1_501 equ DESCALE(1612031267, 30 - CONST_BITS)  ; FIX(1.501321110)
+F_1_847 equ DESCALE(1984016188, 30 - CONST_BITS)  ; FIX(1.847759065)
+F_1_961 equ DESCALE(2106220350, 30 - CONST_BITS)  ; FIX(1.961570560)
+F_2_053 equ DESCALE(2204520673, 30 - CONST_BITS)  ; FIX(2.053119869)
+F_2_562 equ DESCALE(2751909506, 30 - CONST_BITS)  ; FIX(2.562915447)
+F_3_072 equ DESCALE(3299298341, 30 - CONST_BITS)  ; FIX(3.072711026)
+%endif
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_CONST
+
+    alignz      32
+    GLOBAL_DATA(jconst_fdct_islow_sse2)
+
+EXTN(jconst_fdct_islow_sse2):
+
+PW_F130_F054   times 4 dw  (F_0_541 + F_0_765),  F_0_541
+PW_F054_MF130  times 4 dw  F_0_541, (F_0_541 - F_1_847)
+PW_MF078_F117  times 4 dw  (F_1_175 - F_1_961),  F_1_175
+PW_F117_F078   times 4 dw  F_1_175, (F_1_175 - F_0_390)
+PW_MF060_MF089 times 4 dw  (F_0_298 - F_0_899), -F_0_899
+PW_MF089_F060  times 4 dw -F_0_899, (F_1_501 - F_0_899)
+PW_MF050_MF256 times 4 dw  (F_2_053 - F_2_562), -F_2_562
+PW_MF256_F050  times 4 dw -F_2_562, (F_3_072 - F_2_562)
+PD_DESCALE_P1  times 4 dd  1 << (DESCALE_P1 - 1)
+PD_DESCALE_P2  times 4 dd  1 << (DESCALE_P2 - 1)
+PW_DESCALE_P2X times 8 dw  1 << (PASS1_BITS - 1)
+
+    alignz      32
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_TEXT
+    BITS        64
+;
+; Perform the forward DCT on one block of samples.
+;
+; GLOBAL(void)
+; jsimd_fdct_islow_sse2(DCTELEM *data)
+;
+
+; r10 = DCTELEM *data
+
+%define wk(i)   rbp - (WK_NUM - (i)) * SIZEOF_XMMWORD  ; xmmword wk[WK_NUM]
+%define WK_NUM  6
+
+    align       32
+    GLOBAL_FUNCTION(jsimd_fdct_islow_sse2)
+
+EXTN(jsimd_fdct_islow_sse2):
+    push        rbp
+    mov         rax, rsp                     ; rax = original rbp
+    sub         rsp, byte 4
+    and         rsp, byte (-SIZEOF_XMMWORD)  ; align to 128 bits
+    mov         [rsp], rax
+    mov         rbp, rsp                     ; rbp = aligned rbp
+    lea         rsp, [wk(0)]
+    collect_args 1
+
+    ; ---- Pass 1: process rows.
+
+    mov         rdx, r10                ; (DCTELEM *)
+
+    movdqa      xmm0, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_DCTELEM)]
+    movdqa      xmm1, XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_DCTELEM)]
+    movdqa      xmm2, XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_DCTELEM)]
+    movdqa      xmm3, XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_DCTELEM)]
+
+    ; xmm0=(00 01 02 03 04 05 06 07), xmm2=(20 21 22 23 24 25 26 27)
+    ; xmm1=(10 11 12 13 14 15 16 17), xmm3=(30 31 32 33 34 35 36 37)
+
+    movdqa      xmm4, xmm0              ; transpose coefficients(phase 1)
+    punpcklwd   xmm0, xmm1              ; xmm0=(00 10 01 11 02 12 03 13)
+    punpckhwd   xmm4, xmm1              ; xmm4=(04 14 05 15 06 16 07 17)
+    movdqa      xmm5, xmm2              ; transpose coefficients(phase 1)
+    punpcklwd   xmm2, xmm3              ; xmm2=(20 30 21 31 22 32 23 33)
+    punpckhwd   xmm5, xmm3              ; xmm5=(24 34 25 35 26 36 27 37)
+
+    movdqa      xmm6, XMMWORD [XMMBLOCK(4,0,rdx,SIZEOF_DCTELEM)]
+    movdqa      xmm7, XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_DCTELEM)]
+    movdqa      xmm1, XMMWORD [XMMBLOCK(6,0,rdx,SIZEOF_DCTELEM)]
+    movdqa      xmm3, XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_DCTELEM)]
+
+    ; xmm6=( 4 12 20 28 36 44 52 60), xmm1=( 6 14 22 30 38 46 54 62)
+    ; xmm7=( 5 13 21 29 37 45 53 61), xmm3=( 7 15 23 31 39 47 55 63)
+
+    movdqa      XMMWORD [wk(0)], xmm2   ; wk(0)=(20 30 21 31 22 32 23 33)
+    movdqa      XMMWORD [wk(1)], xmm5   ; wk(1)=(24 34 25 35 26 36 27 37)
+
+    movdqa      xmm2, xmm6              ; transpose coefficients(phase 1)
+    punpcklwd   xmm6, xmm7              ; xmm6=(40 50 41 51 42 52 43 53)
+    punpckhwd   xmm2, xmm7              ; xmm2=(44 54 45 55 46 56 47 57)
+    movdqa      xmm5, xmm1              ; transpose coefficients(phase 1)
+    punpcklwd   xmm1, xmm3              ; xmm1=(60 70 61 71 62 72 63 73)
+    punpckhwd   xmm5, xmm3              ; xmm5=(64 74 65 75 66 76 67 77)
+
+    movdqa      xmm7, xmm6              ; transpose coefficients(phase 2)
+    punpckldq   xmm6, xmm1              ; xmm6=(40 50 60 70 41 51 61 71)
+    punpckhdq   xmm7, xmm1              ; xmm7=(42 52 62 72 43 53 63 73)
+    movdqa      xmm3, xmm2              ; transpose coefficients(phase 2)
+    punpckldq   xmm2, xmm5              ; xmm2=(44 54 64 74 45 55 65 75)
+    punpckhdq   xmm3, xmm5              ; xmm3=(46 56 66 76 47 57 67 77)
+
+    movdqa      xmm1, XMMWORD [wk(0)]   ; xmm1=(20 30 21 31 22 32 23 33)
+    movdqa      xmm5, XMMWORD [wk(1)]   ; xmm5=(24 34 25 35 26 36 27 37)
+    movdqa      XMMWORD [wk(2)], xmm7   ; wk(2)=(42 52 62 72 43 53 63 73)
+    movdqa      XMMWORD [wk(3)], xmm2   ; wk(3)=(44 54 64 74 45 55 65 75)
+
+    movdqa      xmm7, xmm0              ; transpose coefficients(phase 2)
+    punpckldq   xmm0, xmm1              ; xmm0=(00 10 20 30 01 11 21 31)
+    punpckhdq   xmm7, xmm1              ; xmm7=(02 12 22 32 03 13 23 33)
+    movdqa      xmm2, xmm4              ; transpose coefficients(phase 2)
+    punpckldq   xmm4, xmm5              ; xmm4=(04 14 24 34 05 15 25 35)
+    punpckhdq   xmm2, xmm5              ; xmm2=(06 16 26 36 07 17 27 37)
+
+    movdqa      xmm1, xmm0              ; transpose coefficients(phase 3)
+    punpcklqdq  xmm0, xmm6              ; xmm0=(00 10 20 30 40 50 60 70)=data0
+    punpckhqdq  xmm1, xmm6              ; xmm1=(01 11 21 31 41 51 61 71)=data1
+    movdqa      xmm5, xmm2              ; transpose coefficients(phase 3)
+    punpcklqdq  xmm2, xmm3              ; xmm2=(06 16 26 36 46 56 66 76)=data6
+    punpckhqdq  xmm5, xmm3              ; xmm5=(07 17 27 37 47 57 67 77)=data7
+
+    movdqa      xmm6, xmm1
+    movdqa      xmm3, xmm0
+    psubw       xmm1, xmm2              ; xmm1=data1-data6=tmp6
+    psubw       xmm0, xmm5              ; xmm0=data0-data7=tmp7
+    paddw       xmm6, xmm2              ; xmm6=data1+data6=tmp1
+    paddw       xmm3, xmm5              ; xmm3=data0+data7=tmp0
+
+    movdqa      xmm2, XMMWORD [wk(2)]   ; xmm2=(42 52 62 72 43 53 63 73)
+    movdqa      xmm5, XMMWORD [wk(3)]   ; xmm5=(44 54 64 74 45 55 65 75)
+    movdqa      XMMWORD [wk(0)], xmm1   ; wk(0)=tmp6
+    movdqa      XMMWORD [wk(1)], xmm0   ; wk(1)=tmp7
+
+    movdqa      xmm1, xmm7              ; transpose coefficients(phase 3)
+    punpcklqdq  xmm7, xmm2              ; xmm7=(02 12 22 32 42 52 62 72)=data2
+    punpckhqdq  xmm1, xmm2              ; xmm1=(03 13 23 33 43 53 63 73)=data3
+    movdqa      xmm0, xmm4              ; transpose coefficients(phase 3)
+    punpcklqdq  xmm4, xmm5              ; xmm4=(04 14 24 34 44 54 64 74)=data4
+    punpckhqdq  xmm0, xmm5              ; xmm0=(05 15 25 35 45 55 65 75)=data5
+
+    movdqa      xmm2, xmm1
+    movdqa      xmm5, xmm7
+    paddw       xmm1, xmm4              ; xmm1=data3+data4=tmp3
+    paddw       xmm7, xmm0              ; xmm7=data2+data5=tmp2
+    psubw       xmm2, xmm4              ; xmm2=data3-data4=tmp4
+    psubw       xmm5, xmm0              ; xmm5=data2-data5=tmp5
+
+    ; -- Even part
+
+    movdqa      xmm4, xmm3
+    movdqa      xmm0, xmm6
+    paddw       xmm3, xmm1              ; xmm3=tmp10
+    paddw       xmm6, xmm7              ; xmm6=tmp11
+    psubw       xmm4, xmm1              ; xmm4=tmp13
+    psubw       xmm0, xmm7              ; xmm0=tmp12
+
+    movdqa      xmm1, xmm3
+    paddw       xmm3, xmm6              ; xmm3=tmp10+tmp11
+    psubw       xmm1, xmm6              ; xmm1=tmp10-tmp11
+
+    psllw       xmm3, PASS1_BITS        ; xmm3=data0
+    psllw       xmm1, PASS1_BITS        ; xmm1=data4
+
+    movdqa      XMMWORD [wk(2)], xmm3   ; wk(2)=data0
+    movdqa      XMMWORD [wk(3)], xmm1   ; wk(3)=data4
+
+    ; (Original)
+    ; z1 = (tmp12 + tmp13) * 0.541196100;
+    ; data2 = z1 + tmp13 * 0.765366865;
+    ; data6 = z1 + tmp12 * -1.847759065;
+    ;
+    ; (This implementation)
+    ; data2 = tmp13 * (0.541196100 + 0.765366865) + tmp12 * 0.541196100;
+    ; data6 = tmp13 * 0.541196100 + tmp12 * (0.541196100 - 1.847759065);
+
+    movdqa      xmm7, xmm4              ; xmm4=tmp13
+    movdqa      xmm6, xmm4
+    punpcklwd   xmm7, xmm0              ; xmm0=tmp12
+    punpckhwd   xmm6, xmm0
+    movdqa      xmm4, xmm7
+    movdqa      xmm0, xmm6
+    pmaddwd     xmm7, [rel PW_F130_F054]   ; xmm7=data2L
+    pmaddwd     xmm6, [rel PW_F130_F054]   ; xmm6=data2H
+    pmaddwd     xmm4, [rel PW_F054_MF130]  ; xmm4=data6L
+    pmaddwd     xmm0, [rel PW_F054_MF130]  ; xmm0=data6H
+
+    paddd       xmm7, [rel PD_DESCALE_P1]
+    paddd       xmm6, [rel PD_DESCALE_P1]
+    psrad       xmm7, DESCALE_P1
+    psrad       xmm6, DESCALE_P1
+    paddd       xmm4, [rel PD_DESCALE_P1]
+    paddd       xmm0, [rel PD_DESCALE_P1]
+    psrad       xmm4, DESCALE_P1
+    psrad       xmm0, DESCALE_P1
+
+    packssdw    xmm7, xmm6              ; xmm7=data2
+    packssdw    xmm4, xmm0              ; xmm4=data6
+
+    movdqa      XMMWORD [wk(4)], xmm7   ; wk(4)=data2
+    movdqa      XMMWORD [wk(5)], xmm4   ; wk(5)=data6
+
+    ; -- Odd part
+
+    movdqa      xmm3, XMMWORD [wk(0)]   ; xmm3=tmp6
+    movdqa      xmm1, XMMWORD [wk(1)]   ; xmm1=tmp7
+
+    movdqa      xmm6, xmm2              ; xmm2=tmp4
+    movdqa      xmm0, xmm5              ; xmm5=tmp5
+    paddw       xmm6, xmm3              ; xmm6=z3
+    paddw       xmm0, xmm1              ; xmm0=z4
+
+    ; (Original)
+    ; z5 = (z3 + z4) * 1.175875602;
+    ; z3 = z3 * -1.961570560;  z4 = z4 * -0.390180644;
+    ; z3 += z5;  z4 += z5;
+    ;
+    ; (This implementation)
+    ; z3 = z3 * (1.175875602 - 1.961570560) + z4 * 1.175875602;
+    ; z4 = z3 * 1.175875602 + z4 * (1.175875602 - 0.390180644);
+
+    movdqa      xmm7, xmm6
+    movdqa      xmm4, xmm6
+    punpcklwd   xmm7, xmm0
+    punpckhwd   xmm4, xmm0
+    movdqa      xmm6, xmm7
+    movdqa      xmm0, xmm4
+    pmaddwd     xmm7, [rel PW_MF078_F117]  ; xmm7=z3L
+    pmaddwd     xmm4, [rel PW_MF078_F117]  ; xmm4=z3H
+    pmaddwd     xmm6, [rel PW_F117_F078]   ; xmm6=z4L
+    pmaddwd     xmm0, [rel PW_F117_F078]   ; xmm0=z4H
+
+    movdqa      XMMWORD [wk(0)], xmm7   ; wk(0)=z3L
+    movdqa      XMMWORD [wk(1)], xmm4   ; wk(1)=z3H
+
+    ; (Original)
+    ; z1 = tmp4 + tmp7;  z2 = tmp5 + tmp6;
+    ; tmp4 = tmp4 * 0.298631336;  tmp5 = tmp5 * 2.053119869;
+    ; tmp6 = tmp6 * 3.072711026;  tmp7 = tmp7 * 1.501321110;
+    ; z1 = z1 * -0.899976223;  z2 = z2 * -2.562915447;
+    ; data7 = tmp4 + z1 + z3;  data5 = tmp5 + z2 + z4;
+    ; data3 = tmp6 + z2 + z3;  data1 = tmp7 + z1 + z4;
+    ;
+    ; (This implementation)
+    ; tmp4 = tmp4 * (0.298631336 - 0.899976223) + tmp7 * -0.899976223;
+    ; tmp5 = tmp5 * (2.053119869 - 2.562915447) + tmp6 * -2.562915447;
+    ; tmp6 = tmp5 * -2.562915447 + tmp6 * (3.072711026 - 2.562915447);
+    ; tmp7 = tmp4 * -0.899976223 + tmp7 * (1.501321110 - 0.899976223);
+    ; data7 = tmp4 + z3;  data5 = tmp5 + z4;
+    ; data3 = tmp6 + z3;  data1 = tmp7 + z4;
+
+    movdqa      xmm7, xmm2
+    movdqa      xmm4, xmm2
+    punpcklwd   xmm7, xmm1
+    punpckhwd   xmm4, xmm1
+    movdqa      xmm2, xmm7
+    movdqa      xmm1, xmm4
+    pmaddwd     xmm7, [rel PW_MF060_MF089]  ; xmm7=tmp4L
+    pmaddwd     xmm4, [rel PW_MF060_MF089]  ; xmm4=tmp4H
+    pmaddwd     xmm2, [rel PW_MF089_F060]   ; xmm2=tmp7L
+    pmaddwd     xmm1, [rel PW_MF089_F060]   ; xmm1=tmp7H
+
+    paddd       xmm7, XMMWORD [wk(0)]   ; xmm7=data7L
+    paddd       xmm4, XMMWORD [wk(1)]   ; xmm4=data7H
+    paddd       xmm2, xmm6              ; xmm2=data1L
+    paddd       xmm1, xmm0              ; xmm1=data1H
+
+    paddd       xmm7, [rel PD_DESCALE_P1]
+    paddd       xmm4, [rel PD_DESCALE_P1]
+    psrad       xmm7, DESCALE_P1
+    psrad       xmm4, DESCALE_P1
+    paddd       xmm2, [rel PD_DESCALE_P1]
+    paddd       xmm1, [rel PD_DESCALE_P1]
+    psrad       xmm2, DESCALE_P1
+    psrad       xmm1, DESCALE_P1
+
+    packssdw    xmm7, xmm4              ; xmm7=data7
+    packssdw    xmm2, xmm1              ; xmm2=data1
+
+    movdqa      xmm4, xmm5
+    movdqa      xmm1, xmm5
+    punpcklwd   xmm4, xmm3
+    punpckhwd   xmm1, xmm3
+    movdqa      xmm5, xmm4
+    movdqa      xmm3, xmm1
+    pmaddwd     xmm4, [rel PW_MF050_MF256]  ; xmm4=tmp5L
+    pmaddwd     xmm1, [rel PW_MF050_MF256]  ; xmm1=tmp5H
+    pmaddwd     xmm5, [rel PW_MF256_F050]   ; xmm5=tmp6L
+    pmaddwd     xmm3, [rel PW_MF256_F050]   ; xmm3=tmp6H
+
+    paddd       xmm4, xmm6              ; xmm4=data5L
+    paddd       xmm1, xmm0              ; xmm1=data5H
+    paddd       xmm5, XMMWORD [wk(0)]   ; xmm5=data3L
+    paddd       xmm3, XMMWORD [wk(1)]   ; xmm3=data3H
+
+    paddd       xmm4, [rel PD_DESCALE_P1]
+    paddd       xmm1, [rel PD_DESCALE_P1]
+    psrad       xmm4, DESCALE_P1
+    psrad       xmm1, DESCALE_P1
+    paddd       xmm5, [rel PD_DESCALE_P1]
+    paddd       xmm3, [rel PD_DESCALE_P1]
+    psrad       xmm5, DESCALE_P1
+    psrad       xmm3, DESCALE_P1
+
+    packssdw    xmm4, xmm1              ; xmm4=data5
+    packssdw    xmm5, xmm3              ; xmm5=data3
+
+    ; ---- Pass 2: process columns.
+
+    movdqa      xmm6, XMMWORD [wk(2)]   ; xmm6=col0
+    movdqa      xmm0, XMMWORD [wk(4)]   ; xmm0=col2
+
+    ; xmm6=(00 10 20 30 40 50 60 70), xmm0=(02 12 22 32 42 52 62 72)
+    ; xmm2=(01 11 21 31 41 51 61 71), xmm5=(03 13 23 33 43 53 63 73)
+
+    movdqa      xmm1, xmm6              ; transpose coefficients(phase 1)
+    punpcklwd   xmm6, xmm2              ; xmm6=(00 01 10 11 20 21 30 31)
+    punpckhwd   xmm1, xmm2              ; xmm1=(40 41 50 51 60 61 70 71)
+    movdqa      xmm3, xmm0              ; transpose coefficients(phase 1)
+    punpcklwd   xmm0, xmm5              ; xmm0=(02 03 12 13 22 23 32 33)
+    punpckhwd   xmm3, xmm5              ; xmm3=(42 43 52 53 62 63 72 73)
+
+    movdqa      xmm2, XMMWORD [wk(3)]   ; xmm2=col4
+    movdqa      xmm5, XMMWORD [wk(5)]   ; xmm5=col6
+
+    ; xmm2=(04 14 24 34 44 54 64 74), xmm5=(06 16 26 36 46 56 66 76)
+    ; xmm4=(05 15 25 35 45 55 65 75), xmm7=(07 17 27 37 47 57 67 77)
+
+    movdqa      XMMWORD [wk(0)], xmm0   ; wk(0)=(02 03 12 13 22 23 32 33)
+    movdqa      XMMWORD [wk(1)], xmm3   ; wk(1)=(42 43 52 53 62 63 72 73)
+
+    movdqa      xmm0, xmm2              ; transpose coefficients(phase 1)
+    punpcklwd   xmm2, xmm4              ; xmm2=(04 05 14 15 24 25 34 35)
+    punpckhwd   xmm0, xmm4              ; xmm0=(44 45 54 55 64 65 74 75)
+    movdqa      xmm3, xmm5              ; transpose coefficients(phase 1)
+    punpcklwd   xmm5, xmm7              ; xmm5=(06 07 16 17 26 27 36 37)
+    punpckhwd   xmm3, xmm7              ; xmm3=(46 47 56 57 66 67 76 77)
+
+    movdqa      xmm4, xmm2              ; transpose coefficients(phase 2)
+    punpckldq   xmm2, xmm5              ; xmm2=(04 05 06 07 14 15 16 17)
+    punpckhdq   xmm4, xmm5              ; xmm4=(24 25 26 27 34 35 36 37)
+    movdqa      xmm7, xmm0              ; transpose coefficients(phase 2)
+    punpckldq   xmm0, xmm3              ; xmm0=(44 45 46 47 54 55 56 57)
+    punpckhdq   xmm7, xmm3              ; xmm7=(64 65 66 67 74 75 76 77)
+
+    movdqa      xmm5, XMMWORD [wk(0)]   ; xmm5=(02 03 12 13 22 23 32 33)
+    movdqa      xmm3, XMMWORD [wk(1)]   ; xmm3=(42 43 52 53 62 63 72 73)
+    movdqa      XMMWORD [wk(2)], xmm4   ; wk(2)=(24 25 26 27 34 35 36 37)
+    movdqa      XMMWORD [wk(3)], xmm0   ; wk(3)=(44 45 46 47 54 55 56 57)
+
+    movdqa      xmm4, xmm6              ; transpose coefficients(phase 2)
+    punpckldq   xmm6, xmm5              ; xmm6=(00 01 02 03 10 11 12 13)
+    punpckhdq   xmm4, xmm5              ; xmm4=(20 21 22 23 30 31 32 33)
+    movdqa      xmm0, xmm1              ; transpose coefficients(phase 2)
+    punpckldq   xmm1, xmm3              ; xmm1=(40 41 42 43 50 51 52 53)
+    punpckhdq   xmm0, xmm3              ; xmm0=(60 61 62 63 70 71 72 73)
+
+    movdqa      xmm5, xmm6              ; transpose coefficients(phase 3)
+    punpcklqdq  xmm6, xmm2              ; xmm6=(00 01 02 03 04 05 06 07)=data0
+    punpckhqdq  xmm5, xmm2              ; xmm5=(10 11 12 13 14 15 16 17)=data1
+    movdqa      xmm3, xmm0              ; transpose coefficients(phase 3)
+    punpcklqdq  xmm0, xmm7              ; xmm0=(60 61 62 63 64 65 66 67)=data6
+    punpckhqdq  xmm3, xmm7              ; xmm3=(70 71 72 73 74 75 76 77)=data7
+
+    movdqa      xmm2, xmm5
+    movdqa      xmm7, xmm6
+    psubw       xmm5, xmm0              ; xmm5=data1-data6=tmp6
+    psubw       xmm6, xmm3              ; xmm6=data0-data7=tmp7
+    paddw       xmm2, xmm0              ; xmm2=data1+data6=tmp1
+    paddw       xmm7, xmm3              ; xmm7=data0+data7=tmp0
+
+    movdqa      xmm0, XMMWORD [wk(2)]   ; xmm0=(24 25 26 27 34 35 36 37)
+    movdqa      xmm3, XMMWORD [wk(3)]   ; xmm3=(44 45 46 47 54 55 56 57)
+    movdqa      XMMWORD [wk(0)], xmm5   ; wk(0)=tmp6
+    movdqa      XMMWORD [wk(1)], xmm6   ; wk(1)=tmp7
+
+    movdqa      xmm5, xmm4              ; transpose coefficients(phase 3)
+    punpcklqdq  xmm4, xmm0              ; xmm4=(20 21 22 23 24 25 26 27)=data2
+    punpckhqdq  xmm5, xmm0              ; xmm5=(30 31 32 33 34 35 36 37)=data3
+    movdqa      xmm6, xmm1              ; transpose coefficients(phase 3)
+    punpcklqdq  xmm1, xmm3              ; xmm1=(40 41 42 43 44 45 46 47)=data4
+    punpckhqdq  xmm6, xmm3              ; xmm6=(50 51 52 53 54 55 56 57)=data5
+
+    movdqa      xmm0, xmm5
+    movdqa      xmm3, xmm4
+    paddw       xmm5, xmm1              ; xmm5=data3+data4=tmp3
+    paddw       xmm4, xmm6              ; xmm4=data2+data5=tmp2
+    psubw       xmm0, xmm1              ; xmm0=data3-data4=tmp4
+    psubw       xmm3, xmm6              ; xmm3=data2-data5=tmp5
+
+    ; -- Even part
+
+    movdqa      xmm1, xmm7
+    movdqa      xmm6, xmm2
+    paddw       xmm7, xmm5              ; xmm7=tmp10
+    paddw       xmm2, xmm4              ; xmm2=tmp11
+    psubw       xmm1, xmm5              ; xmm1=tmp13
+    psubw       xmm6, xmm4              ; xmm6=tmp12
+
+    movdqa      xmm5, xmm7
+    paddw       xmm7, xmm2              ; xmm7=tmp10+tmp11
+    psubw       xmm5, xmm2              ; xmm5=tmp10-tmp11
+
+    paddw       xmm7, [rel PW_DESCALE_P2X]
+    paddw       xmm5, [rel PW_DESCALE_P2X]
+    psraw       xmm7, PASS1_BITS        ; xmm7=data0
+    psraw       xmm5, PASS1_BITS        ; xmm5=data4
+
+    movdqa      XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_DCTELEM)], xmm7
+    movdqa      XMMWORD [XMMBLOCK(4,0,rdx,SIZEOF_DCTELEM)], xmm5
+
+    ; (Original)
+    ; z1 = (tmp12 + tmp13) * 0.541196100;
+    ; data2 = z1 + tmp13 * 0.765366865;
+    ; data6 = z1 + tmp12 * -1.847759065;
+    ;
+    ; (This implementation)
+    ; data2 = tmp13 * (0.541196100 + 0.765366865) + tmp12 * 0.541196100;
+    ; data6 = tmp13 * 0.541196100 + tmp12 * (0.541196100 - 1.847759065);
+
+    movdqa      xmm4, xmm1              ; xmm1=tmp13
+    movdqa      xmm2, xmm1
+    punpcklwd   xmm4, xmm6              ; xmm6=tmp12
+    punpckhwd   xmm2, xmm6
+    movdqa      xmm1, xmm4
+    movdqa      xmm6, xmm2
+    pmaddwd     xmm4, [rel PW_F130_F054]   ; xmm4=data2L
+    pmaddwd     xmm2, [rel PW_F130_F054]   ; xmm2=data2H
+    pmaddwd     xmm1, [rel PW_F054_MF130]  ; xmm1=data6L
+    pmaddwd     xmm6, [rel PW_F054_MF130]  ; xmm6=data6H
+
+    paddd       xmm4, [rel PD_DESCALE_P2]
+    paddd       xmm2, [rel PD_DESCALE_P2]
+    psrad       xmm4, DESCALE_P2
+    psrad       xmm2, DESCALE_P2
+    paddd       xmm1, [rel PD_DESCALE_P2]
+    paddd       xmm6, [rel PD_DESCALE_P2]
+    psrad       xmm1, DESCALE_P2
+    psrad       xmm6, DESCALE_P2
+
+    packssdw    xmm4, xmm2              ; xmm4=data2
+    packssdw    xmm1, xmm6              ; xmm1=data6
+
+    movdqa      XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_DCTELEM)], xmm4
+    movdqa      XMMWORD [XMMBLOCK(6,0,rdx,SIZEOF_DCTELEM)], xmm1
+
+    ; -- Odd part
+
+    movdqa      xmm7, XMMWORD [wk(0)]   ; xmm7=tmp6
+    movdqa      xmm5, XMMWORD [wk(1)]   ; xmm5=tmp7
+
+    movdqa      xmm2, xmm0              ; xmm0=tmp4
+    movdqa      xmm6, xmm3              ; xmm3=tmp5
+    paddw       xmm2, xmm7              ; xmm2=z3
+    paddw       xmm6, xmm5              ; xmm6=z4
+
+    ; (Original)
+    ; z5 = (z3 + z4) * 1.175875602;
+    ; z3 = z3 * -1.961570560;  z4 = z4 * -0.390180644;
+    ; z3 += z5;  z4 += z5;
+    ;
+    ; (This implementation)
+    ; z3 = z3 * (1.175875602 - 1.961570560) + z4 * 1.175875602;
+    ; z4 = z3 * 1.175875602 + z4 * (1.175875602 - 0.390180644);
+
+    movdqa      xmm4, xmm2
+    movdqa      xmm1, xmm2
+    punpcklwd   xmm4, xmm6
+    punpckhwd   xmm1, xmm6
+    movdqa      xmm2, xmm4
+    movdqa      xmm6, xmm1
+    pmaddwd     xmm4, [rel PW_MF078_F117]  ; xmm4=z3L
+    pmaddwd     xmm1, [rel PW_MF078_F117]  ; xmm1=z3H
+    pmaddwd     xmm2, [rel PW_F117_F078]   ; xmm2=z4L
+    pmaddwd     xmm6, [rel PW_F117_F078]   ; xmm6=z4H
+
+    movdqa      XMMWORD [wk(0)], xmm4   ; wk(0)=z3L
+    movdqa      XMMWORD [wk(1)], xmm1   ; wk(1)=z3H
+
+    ; (Original)
+    ; z1 = tmp4 + tmp7;  z2 = tmp5 + tmp6;
+    ; tmp4 = tmp4 * 0.298631336;  tmp5 = tmp5 * 2.053119869;
+    ; tmp6 = tmp6 * 3.072711026;  tmp7 = tmp7 * 1.501321110;
+    ; z1 = z1 * -0.899976223;  z2 = z2 * -2.562915447;
+    ; data7 = tmp4 + z1 + z3;  data5 = tmp5 + z2 + z4;
+    ; data3 = tmp6 + z2 + z3;  data1 = tmp7 + z1 + z4;
+    ;
+    ; (This implementation)
+    ; tmp4 = tmp4 * (0.298631336 - 0.899976223) + tmp7 * -0.899976223;
+    ; tmp5 = tmp5 * (2.053119869 - 2.562915447) + tmp6 * -2.562915447;
+    ; tmp6 = tmp5 * -2.562915447 + tmp6 * (3.072711026 - 2.562915447);
+    ; tmp7 = tmp4 * -0.899976223 + tmp7 * (1.501321110 - 0.899976223);
+    ; data7 = tmp4 + z3;  data5 = tmp5 + z4;
+    ; data3 = tmp6 + z3;  data1 = tmp7 + z4;
+
+    movdqa      xmm4, xmm0
+    movdqa      xmm1, xmm0
+    punpcklwd   xmm4, xmm5
+    punpckhwd   xmm1, xmm5
+    movdqa      xmm0, xmm4
+    movdqa      xmm5, xmm1
+    pmaddwd     xmm4, [rel PW_MF060_MF089]  ; xmm4=tmp4L
+    pmaddwd     xmm1, [rel PW_MF060_MF089]  ; xmm1=tmp4H
+    pmaddwd     xmm0, [rel PW_MF089_F060]   ; xmm0=tmp7L
+    pmaddwd     xmm5, [rel PW_MF089_F060]   ; xmm5=tmp7H
+
+    paddd       xmm4,  XMMWORD [wk(0)]  ; xmm4=data7L
+    paddd       xmm1,  XMMWORD [wk(1)]  ; xmm1=data7H
+    paddd       xmm0, xmm2              ; xmm0=data1L
+    paddd       xmm5, xmm6              ; xmm5=data1H
+
+    paddd       xmm4, [rel PD_DESCALE_P2]
+    paddd       xmm1, [rel PD_DESCALE_P2]
+    psrad       xmm4, DESCALE_P2
+    psrad       xmm1, DESCALE_P2
+    paddd       xmm0, [rel PD_DESCALE_P2]
+    paddd       xmm5, [rel PD_DESCALE_P2]
+    psrad       xmm0, DESCALE_P2
+    psrad       xmm5, DESCALE_P2
+
+    packssdw    xmm4, xmm1              ; xmm4=data7
+    packssdw    xmm0, xmm5              ; xmm0=data1
+
+    movdqa      XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_DCTELEM)], xmm4
+    movdqa      XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_DCTELEM)], xmm0
+
+    movdqa      xmm1, xmm3
+    movdqa      xmm5, xmm3
+    punpcklwd   xmm1, xmm7
+    punpckhwd   xmm5, xmm7
+    movdqa      xmm3, xmm1
+    movdqa      xmm7, xmm5
+    pmaddwd     xmm1, [rel PW_MF050_MF256]  ; xmm1=tmp5L
+    pmaddwd     xmm5, [rel PW_MF050_MF256]  ; xmm5=tmp5H
+    pmaddwd     xmm3, [rel PW_MF256_F050]   ; xmm3=tmp6L
+    pmaddwd     xmm7, [rel PW_MF256_F050]   ; xmm7=tmp6H
+
+    paddd       xmm1, xmm2              ; xmm1=data5L
+    paddd       xmm5, xmm6              ; xmm5=data5H
+    paddd       xmm3, XMMWORD [wk(0)]   ; xmm3=data3L
+    paddd       xmm7, XMMWORD [wk(1)]   ; xmm7=data3H
+
+    paddd       xmm1, [rel PD_DESCALE_P2]
+    paddd       xmm5, [rel PD_DESCALE_P2]
+    psrad       xmm1, DESCALE_P2
+    psrad       xmm5, DESCALE_P2
+    paddd       xmm3, [rel PD_DESCALE_P2]
+    paddd       xmm7, [rel PD_DESCALE_P2]
+    psrad       xmm3, DESCALE_P2
+    psrad       xmm7, DESCALE_P2
+
+    packssdw    xmm1, xmm5              ; xmm1=data5
+    packssdw    xmm3, xmm7              ; xmm3=data3
+
+    movdqa      XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_DCTELEM)], xmm1
+    movdqa      XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_DCTELEM)], xmm3
+
+    uncollect_args 1
+    mov         rsp, rbp                ; rsp <- aligned rbp
+    pop         rsp                     ; rsp <- original rbp
+    pop         rbp
+    ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+    align       32
diff --git a/media/libjpeg/simd/x86_64/jidctflt-sse2.asm b/media/libjpeg/simd/x86_64/jidctflt-sse2.asm
new file mode 100644
index 0000000000..60bf961896
--- /dev/null
+++ b/media/libjpeg/simd/x86_64/jidctflt-sse2.asm
@@ -0,0 +1,482 @@
+;
+; jidctflt.asm - floating-point IDCT (64-bit SSE & SSE2)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+; Copyright (C) 2009, 2016, D. R. Commander.
+; Copyright (C) 2018, Matthias Räncker.
+;
+; Based on the x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; This file contains a floating-point implementation of the inverse DCT
+; (Discrete Cosine Transform). The following code is based directly on
+; the IJG's original jidctflt.c; see the jidctflt.c for more details.
+
+%include "jsimdext.inc"
+%include "jdct.inc"
+
+; --------------------------------------------------------------------------
+
+%macro unpcklps2 2  ; %1=(0 1 2 3) / %2=(4 5 6 7) => %1=(0 1 4 5)
+    shufps      %1, %2, 0x44
+%endmacro
+
+%macro unpckhps2 2  ; %1=(0 1 2 3) / %2=(4 5 6 7) => %1=(2 3 6 7)
+    shufps      %1, %2, 0xEE
+%endmacro
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_CONST
+
+    alignz      32
+    GLOBAL_DATA(jconst_idct_float_sse2)
+
+EXTN(jconst_idct_float_sse2):
+
+PD_1_414        times 4  dd  1.414213562373095048801689
+PD_1_847        times 4  dd  1.847759065022573512256366
+PD_1_082        times 4  dd  1.082392200292393968799446
+PD_M2_613       times 4  dd -2.613125929752753055713286
+PD_RNDINT_MAGIC times 4  dd  100663296.0  ; (float)(0x00C00000 << 3)
+PB_CENTERJSAMP  times 16 db  CENTERJSAMPLE
+
+    alignz      32
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_TEXT
+    BITS        64
+;
+; Perform dequantization and inverse DCT on one block of coefficients.
+;
+; GLOBAL(void)
+; jsimd_idct_float_sse2(void *dct_table, JCOEFPTR coef_block,
+;                       JSAMPARRAY output_buf, JDIMENSION output_col)
+;
+
+; r10 = void *dct_table
+; r11 = JCOEFPTR coef_block
+; r12 = JSAMPARRAY output_buf
+; r13d = JDIMENSION output_col
+
+%define original_rbp  rbp + 0
+%define wk(i)         rbp - (WK_NUM - (i)) * SIZEOF_XMMWORD
+                                        ; xmmword wk[WK_NUM]
+%define WK_NUM        2
+%define workspace     wk(0) - DCTSIZE2 * SIZEOF_FAST_FLOAT
+                                        ; FAST_FLOAT workspace[DCTSIZE2]
+
+    align       32
+    GLOBAL_FUNCTION(jsimd_idct_float_sse2)
+
+EXTN(jsimd_idct_float_sse2):
+    push        rbp
+    mov         rax, rsp                     ; rax = original rbp
+    sub         rsp, byte 4
+    and         rsp, byte (-SIZEOF_XMMWORD)  ; align to 128 bits
+    mov         [rsp], rax
+    mov         rbp, rsp                     ; rbp = aligned rbp
+    lea         rsp, [workspace]
+    collect_args 4
+    push        rbx
+
+    ; ---- Pass 1: process columns from input, store into work array.
+
+    mov         rdx, r10                ; quantptr
+    mov         rsi, r11                ; inptr
+    lea         rdi, [workspace]        ; FAST_FLOAT *wsptr
+    mov         rcx, DCTSIZE/4          ; ctr
+.columnloop:
+%ifndef NO_ZERO_COLUMN_TEST_FLOAT_SSE
+    mov         eax, dword [DWBLOCK(1,0,rsi,SIZEOF_JCOEF)]
+    or          eax, dword [DWBLOCK(2,0,rsi,SIZEOF_JCOEF)]
+    jnz         near .columnDCT
+
+    movq        xmm1, XMM_MMWORD [MMBLOCK(1,0,rsi,SIZEOF_JCOEF)]
+    movq        xmm2, XMM_MMWORD [MMBLOCK(2,0,rsi,SIZEOF_JCOEF)]
+    movq        xmm3, XMM_MMWORD [MMBLOCK(3,0,rsi,SIZEOF_JCOEF)]
+    movq        xmm4, XMM_MMWORD [MMBLOCK(4,0,rsi,SIZEOF_JCOEF)]
+    movq        xmm5, XMM_MMWORD [MMBLOCK(5,0,rsi,SIZEOF_JCOEF)]
+    movq        xmm6, XMM_MMWORD [MMBLOCK(6,0,rsi,SIZEOF_JCOEF)]
+    movq        xmm7, XMM_MMWORD [MMBLOCK(7,0,rsi,SIZEOF_JCOEF)]
+    por         xmm1, xmm2
+    por         xmm3, xmm4
+    por         xmm5, xmm6
+    por         xmm1, xmm3
+    por         xmm5, xmm7
+    por         xmm1, xmm5
+    packsswb    xmm1, xmm1
+    movd        eax, xmm1
+    test        rax, rax
+    jnz         short .columnDCT
+
+    ; -- AC terms all zero
+
+    movq        xmm0, XMM_MMWORD [MMBLOCK(0,0,rsi,SIZEOF_JCOEF)]
+
+    punpcklwd   xmm0, xmm0                  ; xmm0=(00 00 01 01 02 02 03 03)
+    psrad       xmm0, (DWORD_BIT-WORD_BIT)  ; xmm0=in0=(00 01 02 03)
+    cvtdq2ps    xmm0, xmm0                  ; xmm0=in0=(00 01 02 03)
+
+    mulps       xmm0, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_FLOAT_MULT_TYPE)]
+
+    movaps      xmm1, xmm0
+    movaps      xmm2, xmm0
+    movaps      xmm3, xmm0
+
+    shufps      xmm0, xmm0, 0x00        ; xmm0=(00 00 00 00)
+    shufps      xmm1, xmm1, 0x55        ; xmm1=(01 01 01 01)
+    shufps      xmm2, xmm2, 0xAA        ; xmm2=(02 02 02 02)
+    shufps      xmm3, xmm3, 0xFF        ; xmm3=(03 03 03 03)
+
+    movaps      XMMWORD [XMMBLOCK(0,0,rdi,SIZEOF_FAST_FLOAT)], xmm0
+    movaps      XMMWORD [XMMBLOCK(0,1,rdi,SIZEOF_FAST_FLOAT)], xmm0
+    movaps      XMMWORD [XMMBLOCK(1,0,rdi,SIZEOF_FAST_FLOAT)], xmm1
+    movaps      XMMWORD [XMMBLOCK(1,1,rdi,SIZEOF_FAST_FLOAT)], xmm1
+    movaps      XMMWORD [XMMBLOCK(2,0,rdi,SIZEOF_FAST_FLOAT)], xmm2
+    movaps      XMMWORD [XMMBLOCK(2,1,rdi,SIZEOF_FAST_FLOAT)], xmm2
+    movaps      XMMWORD [XMMBLOCK(3,0,rdi,SIZEOF_FAST_FLOAT)], xmm3
+    movaps      XMMWORD [XMMBLOCK(3,1,rdi,SIZEOF_FAST_FLOAT)], xmm3
+    jmp         near .nextcolumn
+%endif
+.columnDCT:
+
+    ; -- Even part
+
+    movq        xmm0, XMM_MMWORD [MMBLOCK(0,0,rsi,SIZEOF_JCOEF)]
+    movq        xmm1, XMM_MMWORD [MMBLOCK(2,0,rsi,SIZEOF_JCOEF)]
+    movq        xmm2, XMM_MMWORD [MMBLOCK(4,0,rsi,SIZEOF_JCOEF)]
+    movq        xmm3, XMM_MMWORD [MMBLOCK(6,0,rsi,SIZEOF_JCOEF)]
+
+    punpcklwd   xmm0, xmm0                  ; xmm0=(00 00 01 01 02 02 03 03)
+    punpcklwd   xmm1, xmm1                  ; xmm1=(20 20 21 21 22 22 23 23)
+    psrad       xmm0, (DWORD_BIT-WORD_BIT)  ; xmm0=in0=(00 01 02 03)
+    psrad       xmm1, (DWORD_BIT-WORD_BIT)  ; xmm1=in2=(20 21 22 23)
+    cvtdq2ps    xmm0, xmm0                  ; xmm0=in0=(00 01 02 03)
+    cvtdq2ps    xmm1, xmm1                  ; xmm1=in2=(20 21 22 23)
+
+    punpcklwd   xmm2, xmm2                  ; xmm2=(40 40 41 41 42 42 43 43)
+    punpcklwd   xmm3, xmm3                  ; xmm3=(60 60 61 61 62 62 63 63)
+    psrad       xmm2, (DWORD_BIT-WORD_BIT)  ; xmm2=in4=(40 41 42 43)
+    psrad       xmm3, (DWORD_BIT-WORD_BIT)  ; xmm3=in6=(60 61 62 63)
+    cvtdq2ps    xmm2, xmm2                  ; xmm2=in4=(40 41 42 43)
+    cvtdq2ps    xmm3, xmm3                  ; xmm3=in6=(60 61 62 63)
+
+    mulps       xmm0, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_FLOAT_MULT_TYPE)]
+    mulps       xmm1, XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_FLOAT_MULT_TYPE)]
+    mulps       xmm2, XMMWORD [XMMBLOCK(4,0,rdx,SIZEOF_FLOAT_MULT_TYPE)]
+    mulps       xmm3, XMMWORD [XMMBLOCK(6,0,rdx,SIZEOF_FLOAT_MULT_TYPE)]
+
+    movaps      xmm4, xmm0
+    movaps      xmm5, xmm1
+    subps       xmm0, xmm2              ; xmm0=tmp11
+    subps       xmm1, xmm3
+    addps       xmm4, xmm2              ; xmm4=tmp10
+    addps       xmm5, xmm3              ; xmm5=tmp13
+
+    mulps       xmm1, [rel PD_1_414]
+    subps       xmm1, xmm5              ; xmm1=tmp12
+
+    movaps      xmm6, xmm4
+    movaps      xmm7, xmm0
+    subps       xmm4, xmm5              ; xmm4=tmp3
+    subps       xmm0, xmm1              ; xmm0=tmp2
+    addps       xmm6, xmm5              ; xmm6=tmp0
+    addps       xmm7, xmm1              ; xmm7=tmp1
+
+    movaps      XMMWORD [wk(1)], xmm4   ; tmp3
+    movaps      XMMWORD [wk(0)], xmm0   ; tmp2
+
+    ; -- Odd part
+
+    movq        xmm2, XMM_MMWORD [MMBLOCK(1,0,rsi,SIZEOF_JCOEF)]
+    movq        xmm3, XMM_MMWORD [MMBLOCK(3,0,rsi,SIZEOF_JCOEF)]
+    movq        xmm5, XMM_MMWORD [MMBLOCK(5,0,rsi,SIZEOF_JCOEF)]
+    movq        xmm1, XMM_MMWORD [MMBLOCK(7,0,rsi,SIZEOF_JCOEF)]
+
+    punpcklwd   xmm2, xmm2                  ; xmm2=(10 10 11 11 12 12 13 13)
+    punpcklwd   xmm3, xmm3                  ; xmm3=(30 30 31 31 32 32 33 33)
+    psrad       xmm2, (DWORD_BIT-WORD_BIT)  ; xmm2=in1=(10 11 12 13)
+    psrad       xmm3, (DWORD_BIT-WORD_BIT)  ; xmm3=in3=(30 31 32 33)
+    cvtdq2ps    xmm2, xmm2                  ; xmm2=in1=(10 11 12 13)
+    cvtdq2ps    xmm3, xmm3                  ; xmm3=in3=(30 31 32 33)
+
+    punpcklwd   xmm5, xmm5                  ; xmm5=(50 50 51 51 52 52 53 53)
+    punpcklwd   xmm1, xmm1                  ; xmm1=(70 70 71 71 72 72 73 73)
+    psrad       xmm5, (DWORD_BIT-WORD_BIT)  ; xmm5=in5=(50 51 52 53)
+    psrad       xmm1, (DWORD_BIT-WORD_BIT)  ; xmm1=in7=(70 71 72 73)
+    cvtdq2ps    xmm5, xmm5                  ; xmm5=in5=(50 51 52 53)
+    cvtdq2ps    xmm1, xmm1                  ; xmm1=in7=(70 71 72 73)
+
+    mulps       xmm2, XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_FLOAT_MULT_TYPE)]
+    mulps       xmm3, XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_FLOAT_MULT_TYPE)]
+    mulps       xmm5, XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_FLOAT_MULT_TYPE)]
+    mulps       xmm1, XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_FLOAT_MULT_TYPE)]
+
+    movaps      xmm4, xmm2
+    movaps      xmm0, xmm5
+    addps       xmm2, xmm1              ; xmm2=z11
+    addps       xmm5, xmm3              ; xmm5=z13
+    subps       xmm4, xmm1              ; xmm4=z12
+    subps       xmm0, xmm3              ; xmm0=z10
+
+    movaps      xmm1, xmm2
+    subps       xmm2, xmm5
+    addps       xmm1, xmm5              ; xmm1=tmp7
+
+    mulps       xmm2, [rel PD_1_414]    ; xmm2=tmp11
+
+    movaps      xmm3, xmm0
+    addps       xmm0, xmm4
+    mulps       xmm0, [rel PD_1_847]    ; xmm0=z5
+    mulps       xmm3, [rel PD_M2_613]   ; xmm3=(z10 * -2.613125930)
+    mulps       xmm4, [rel PD_1_082]    ; xmm4=(z12 * 1.082392200)
+    addps       xmm3, xmm0              ; xmm3=tmp12
+    subps       xmm4, xmm0              ; xmm4=tmp10
+
+    ; -- Final output stage
+
+    subps       xmm3, xmm1              ; xmm3=tmp6
+    movaps      xmm5, xmm6
+    movaps      xmm0, xmm7
+    addps       xmm6, xmm1              ; xmm6=data0=(00 01 02 03)
+    addps       xmm7, xmm3              ; xmm7=data1=(10 11 12 13)
+    subps       xmm5, xmm1              ; xmm5=data7=(70 71 72 73)
+    subps       xmm0, xmm3              ; xmm0=data6=(60 61 62 63)
+    subps       xmm2, xmm3              ; xmm2=tmp5
+
+    movaps      xmm1, xmm6              ; transpose coefficients(phase 1)
+    unpcklps    xmm6, xmm7              ; xmm6=(00 10 01 11)
+    unpckhps    xmm1, xmm7              ; xmm1=(02 12 03 13)
+    movaps      xmm3, xmm0              ; transpose coefficients(phase 1)
+    unpcklps    xmm0, xmm5              ; xmm0=(60 70 61 71)
+    unpckhps    xmm3, xmm5              ; xmm3=(62 72 63 73)
+
+    movaps      xmm7, XMMWORD [wk(0)]   ; xmm7=tmp2
+    movaps      xmm5, XMMWORD [wk(1)]   ; xmm5=tmp3
+
+    movaps      XMMWORD [wk(0)], xmm0   ; wk(0)=(60 70 61 71)
+    movaps      XMMWORD [wk(1)], xmm3   ; wk(1)=(62 72 63 73)
+
+    addps       xmm4, xmm2              ; xmm4=tmp4
+    movaps      xmm0, xmm7
+    movaps      xmm3, xmm5
+    addps       xmm7, xmm2              ; xmm7=data2=(20 21 22 23)
+    addps       xmm5, xmm4              ; xmm5=data4=(40 41 42 43)
+    subps       xmm0, xmm2              ; xmm0=data5=(50 51 52 53)
+    subps       xmm3, xmm4              ; xmm3=data3=(30 31 32 33)
+
+    movaps      xmm2, xmm7              ; transpose coefficients(phase 1)
+    unpcklps    xmm7, xmm3              ; xmm7=(20 30 21 31)
+    unpckhps    xmm2, xmm3              ; xmm2=(22 32 23 33)
+    movaps      xmm4, xmm5              ; transpose coefficients(phase 1)
+    unpcklps    xmm5, xmm0              ; xmm5=(40 50 41 51)
+    unpckhps    xmm4, xmm0              ; xmm4=(42 52 43 53)
+
+    movaps      xmm3, xmm6              ; transpose coefficients(phase 2)
+    unpcklps2   xmm6, xmm7              ; xmm6=(00 10 20 30)
+    unpckhps2   xmm3, xmm7              ; xmm3=(01 11 21 31)
+    movaps      xmm0, xmm1              ; transpose coefficients(phase 2)
+    unpcklps2   xmm1, xmm2              ; xmm1=(02 12 22 32)
+    unpckhps2   xmm0, xmm2              ; xmm0=(03 13 23 33)
+
+    movaps      xmm7, XMMWORD [wk(0)]   ; xmm7=(60 70 61 71)
+    movaps      xmm2, XMMWORD [wk(1)]   ; xmm2=(62 72 63 73)
+
+    movaps      XMMWORD [XMMBLOCK(0,0,rdi,SIZEOF_FAST_FLOAT)], xmm6
+    movaps      XMMWORD [XMMBLOCK(1,0,rdi,SIZEOF_FAST_FLOAT)], xmm3
+    movaps      XMMWORD [XMMBLOCK(2,0,rdi,SIZEOF_FAST_FLOAT)], xmm1
+    movaps      XMMWORD [XMMBLOCK(3,0,rdi,SIZEOF_FAST_FLOAT)], xmm0
+
+    movaps      xmm6, xmm5              ; transpose coefficients(phase 2)
+    unpcklps2   xmm5, xmm7              ; xmm5=(40 50 60 70)
+    unpckhps2   xmm6, xmm7              ; xmm6=(41 51 61 71)
+    movaps      xmm3, xmm4              ; transpose coefficients(phase 2)
+    unpcklps2   xmm4, xmm2              ; xmm4=(42 52 62 72)
+    unpckhps2   xmm3, xmm2              ; xmm3=(43 53 63 73)
+
+    movaps      XMMWORD [XMMBLOCK(0,1,rdi,SIZEOF_FAST_FLOAT)], xmm5
+    movaps      XMMWORD [XMMBLOCK(1,1,rdi,SIZEOF_FAST_FLOAT)], xmm6
+    movaps      XMMWORD [XMMBLOCK(2,1,rdi,SIZEOF_FAST_FLOAT)], xmm4
+    movaps      XMMWORD [XMMBLOCK(3,1,rdi,SIZEOF_FAST_FLOAT)], xmm3
+
+.nextcolumn:
+    add         rsi, byte 4*SIZEOF_JCOEF               ; coef_block
+    add         rdx, byte 4*SIZEOF_FLOAT_MULT_TYPE     ; quantptr
+    add         rdi,      4*DCTSIZE*SIZEOF_FAST_FLOAT  ; wsptr
+    dec         rcx                                    ; ctr
+    jnz         near .columnloop
+
+    ; -- Prefetch the next coefficient block
+
+    prefetchnta [rsi + (DCTSIZE2-8)*SIZEOF_JCOEF + 0*32]
+    prefetchnta [rsi + (DCTSIZE2-8)*SIZEOF_JCOEF + 1*32]
+    prefetchnta [rsi + (DCTSIZE2-8)*SIZEOF_JCOEF + 2*32]
+    prefetchnta [rsi + (DCTSIZE2-8)*SIZEOF_JCOEF + 3*32]
+
+    ; ---- Pass 2: process rows from work array, store into output array.
+
+    mov         rax, [original_rbp]
+    lea         rsi, [workspace]        ; FAST_FLOAT *wsptr
+    mov         rdi, r12                ; (JSAMPROW *)
+    mov         eax, r13d
+    mov         rcx, DCTSIZE/4          ; ctr
+.rowloop:
+
+    ; -- Even part
+
+    movaps      xmm0, XMMWORD [XMMBLOCK(0,0,rsi,SIZEOF_FAST_FLOAT)]
+    movaps      xmm1, XMMWORD [XMMBLOCK(2,0,rsi,SIZEOF_FAST_FLOAT)]
+    movaps      xmm2, XMMWORD [XMMBLOCK(4,0,rsi,SIZEOF_FAST_FLOAT)]
+    movaps      xmm3, XMMWORD [XMMBLOCK(6,0,rsi,SIZEOF_FAST_FLOAT)]
+
+    movaps      xmm4, xmm0
+    movaps      xmm5, xmm1
+    subps       xmm0, xmm2              ; xmm0=tmp11
+    subps       xmm1, xmm3
+    addps       xmm4, xmm2              ; xmm4=tmp10
+    addps       xmm5, xmm3              ; xmm5=tmp13
+
+    mulps       xmm1, [rel PD_1_414]
+    subps       xmm1, xmm5              ; xmm1=tmp12
+
+    movaps      xmm6, xmm4
+    movaps      xmm7, xmm0
+    subps       xmm4, xmm5              ; xmm4=tmp3
+    subps       xmm0, xmm1              ; xmm0=tmp2
+    addps       xmm6, xmm5              ; xmm6=tmp0
+    addps       xmm7, xmm1              ; xmm7=tmp1
+
+    movaps      XMMWORD [wk(1)], xmm4   ; tmp3
+    movaps      XMMWORD [wk(0)], xmm0   ; tmp2
+
+    ; -- Odd part
+
+    movaps      xmm2, XMMWORD [XMMBLOCK(1,0,rsi,SIZEOF_FAST_FLOAT)]
+    movaps      xmm3, XMMWORD [XMMBLOCK(3,0,rsi,SIZEOF_FAST_FLOAT)]
+    movaps      xmm5, XMMWORD [XMMBLOCK(5,0,rsi,SIZEOF_FAST_FLOAT)]
+    movaps      xmm1, XMMWORD [XMMBLOCK(7,0,rsi,SIZEOF_FAST_FLOAT)]
+
+    movaps      xmm4, xmm2
+    movaps      xmm0, xmm5
+    addps       xmm2, xmm1              ; xmm2=z11
+    addps       xmm5, xmm3              ; xmm5=z13
+    subps       xmm4, xmm1              ; xmm4=z12
+    subps       xmm0, xmm3              ; xmm0=z10
+
+    movaps      xmm1, xmm2
+    subps       xmm2, xmm5
+    addps       xmm1, xmm5              ; xmm1=tmp7
+
+    mulps       xmm2, [rel PD_1_414]    ; xmm2=tmp11
+
+    movaps      xmm3, xmm0
+    addps       xmm0, xmm4
+    mulps       xmm0, [rel PD_1_847]    ; xmm0=z5
+    mulps       xmm3, [rel PD_M2_613]   ; xmm3=(z10 * -2.613125930)
+    mulps       xmm4, [rel PD_1_082]    ; xmm4=(z12 * 1.082392200)
+    addps       xmm3, xmm0              ; xmm3=tmp12
+    subps       xmm4, xmm0              ; xmm4=tmp10
+
+    ; -- Final output stage
+
+    subps       xmm3, xmm1              ; xmm3=tmp6
+    movaps      xmm5, xmm6
+    movaps      xmm0, xmm7
+    addps       xmm6, xmm1              ; xmm6=data0=(00 10 20 30)
+    addps       xmm7, xmm3              ; xmm7=data1=(01 11 21 31)
+    subps       xmm5, xmm1              ; xmm5=data7=(07 17 27 37)
+    subps       xmm0, xmm3              ; xmm0=data6=(06 16 26 36)
+    subps       xmm2, xmm3              ; xmm2=tmp5
+
+    movaps      xmm1, [rel PD_RNDINT_MAGIC]  ; xmm1=[rel PD_RNDINT_MAGIC]
+    pcmpeqd     xmm3, xmm3
+    psrld       xmm3, WORD_BIT          ; xmm3={0xFFFF 0x0000 0xFFFF 0x0000 ..}
+
+    addps       xmm6, xmm1              ; xmm6=roundint(data0/8)=(00 ** 10 ** 20 ** 30 **)
+    addps       xmm7, xmm1              ; xmm7=roundint(data1/8)=(01 ** 11 ** 21 ** 31 **)
+    addps       xmm0, xmm1              ; xmm0=roundint(data6/8)=(06 ** 16 ** 26 ** 36 **)
+    addps       xmm5, xmm1              ; xmm5=roundint(data7/8)=(07 ** 17 ** 27 ** 37 **)
+
+    pand        xmm6, xmm3              ; xmm6=(00 -- 10 -- 20 -- 30 --)
+    pslld       xmm7, WORD_BIT          ; xmm7=(-- 01 -- 11 -- 21 -- 31)
+    pand        xmm0, xmm3              ; xmm0=(06 -- 16 -- 26 -- 36 --)
+    pslld       xmm5, WORD_BIT          ; xmm5=(-- 07 -- 17 -- 27 -- 37)
+    por         xmm6, xmm7              ; xmm6=(00 01 10 11 20 21 30 31)
+    por         xmm0, xmm5              ; xmm0=(06 07 16 17 26 27 36 37)
+
+    movaps      xmm1,  XMMWORD [wk(0)]  ; xmm1=tmp2
+    movaps      xmm3,  XMMWORD [wk(1)]  ; xmm3=tmp3
+
+    addps       xmm4, xmm2              ; xmm4=tmp4
+    movaps      xmm7, xmm1
+    movaps      xmm5, xmm3
+    addps       xmm1, xmm2              ; xmm1=data2=(02 12 22 32)
+    addps       xmm3, xmm4              ; xmm3=data4=(04 14 24 34)
+    subps       xmm7, xmm2              ; xmm7=data5=(05 15 25 35)
+    subps       xmm5, xmm4              ; xmm5=data3=(03 13 23 33)
+
+    movaps      xmm2, [rel PD_RNDINT_MAGIC]  ; xmm2=[rel PD_RNDINT_MAGIC]
+    pcmpeqd     xmm4, xmm4
+    psrld       xmm4, WORD_BIT          ; xmm4={0xFFFF 0x0000 0xFFFF 0x0000 ..}
+
+    addps       xmm3, xmm2              ; xmm3=roundint(data4/8)=(04 ** 14 ** 24 ** 34 **)
+    addps       xmm7, xmm2              ; xmm7=roundint(data5/8)=(05 ** 15 ** 25 ** 35 **)
+    addps       xmm1, xmm2              ; xmm1=roundint(data2/8)=(02 ** 12 ** 22 ** 32 **)
+    addps       xmm5, xmm2              ; xmm5=roundint(data3/8)=(03 ** 13 ** 23 ** 33 **)
+
+    pand        xmm3, xmm4              ; xmm3=(04 -- 14 -- 24 -- 34 --)
+    pslld       xmm7, WORD_BIT          ; xmm7=(-- 05 -- 15 -- 25 -- 35)
+    pand        xmm1, xmm4              ; xmm1=(02 -- 12 -- 22 -- 32 --)
+    pslld       xmm5, WORD_BIT          ; xmm5=(-- 03 -- 13 -- 23 -- 33)
+    por         xmm3, xmm7              ; xmm3=(04 05 14 15 24 25 34 35)
+    por         xmm1, xmm5              ; xmm1=(02 03 12 13 22 23 32 33)
+
+    movdqa      xmm2, [rel PB_CENTERJSAMP]  ; xmm2=[rel PB_CENTERJSAMP]
+
+    packsswb    xmm6, xmm3        ; xmm6=(00 01 10 11 20 21 30 31 04 05 14 15 24 25 34 35)
+    packsswb    xmm1, xmm0        ; xmm1=(02 03 12 13 22 23 32 33 06 07 16 17 26 27 36 37)
+    paddb       xmm6, xmm2
+    paddb       xmm1, xmm2
+
+    movdqa      xmm4, xmm6        ; transpose coefficients(phase 2)
+    punpcklwd   xmm6, xmm1        ; xmm6=(00 01 02 03 10 11 12 13 20 21 22 23 30 31 32 33)
+    punpckhwd   xmm4, xmm1        ; xmm4=(04 05 06 07 14 15 16 17 24 25 26 27 34 35 36 37)
+
+    movdqa      xmm7, xmm6        ; transpose coefficients(phase 3)
+    punpckldq   xmm6, xmm4        ; xmm6=(00 01 02 03 04 05 06 07 10 11 12 13 14 15 16 17)
+    punpckhdq   xmm7, xmm4        ; xmm7=(20 21 22 23 24 25 26 27 30 31 32 33 34 35 36 37)
+
+    pshufd      xmm5, xmm6, 0x4E  ; xmm5=(10 11 12 13 14 15 16 17 00 01 02 03 04 05 06 07)
+    pshufd      xmm3, xmm7, 0x4E  ; xmm3=(30 31 32 33 34 35 36 37 20 21 22 23 24 25 26 27)
+
+    mov         rdxp, JSAMPROW [rdi+0*SIZEOF_JSAMPROW]
+    mov         rbxp, JSAMPROW [rdi+2*SIZEOF_JSAMPROW]
+    movq        XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm6
+    movq        XMM_MMWORD [rbx+rax*SIZEOF_JSAMPLE], xmm7
+    mov         rdxp, JSAMPROW [rdi+1*SIZEOF_JSAMPROW]
+    mov         rbxp, JSAMPROW [rdi+3*SIZEOF_JSAMPROW]
+    movq        XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm5
+    movq        XMM_MMWORD [rbx+rax*SIZEOF_JSAMPLE], xmm3
+
+    add         rsi, byte 4*SIZEOF_FAST_FLOAT  ; wsptr
+    add         rdi, byte 4*SIZEOF_JSAMPROW
+    dec         rcx                            ; ctr
+    jnz         near .rowloop
+
+    pop         rbx
+    uncollect_args 4
+    mov         rsp, rbp                ; rsp <- aligned rbp
+    pop         rsp                     ; rsp <- original rbp
+    pop         rbp
+    ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+    align       32
diff --git a/media/libjpeg/simd/x86_64/jidctfst-sse2.asm b/media/libjpeg/simd/x86_64/jidctfst-sse2.asm
new file mode 100644
index 0000000000..cb97fdfbb2
--- /dev/null
+++ b/media/libjpeg/simd/x86_64/jidctfst-sse2.asm
@@ -0,0 +1,491 @@
+;
+; jidctfst.asm - fast integer IDCT (64-bit SSE2)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+; Copyright (C) 2009, 2016, D. R. Commander.
+; Copyright (C) 2018, Matthias Räncker.
+;
+; Based on the x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; This file contains a fast, not so accurate integer implementation of
+; the inverse DCT (Discrete Cosine Transform). The following code is
+; based directly on the IJG's original jidctfst.c; see the jidctfst.c
+; for more details.
+
+%include "jsimdext.inc"
+%include "jdct.inc"
+
+; --------------------------------------------------------------------------
+
+%define CONST_BITS  8  ; 14 is also OK.
+%define PASS1_BITS  2
+
+%if IFAST_SCALE_BITS != PASS1_BITS
+%error "'IFAST_SCALE_BITS' must be equal to 'PASS1_BITS'."
+%endif
+
+%if CONST_BITS == 8
+F_1_082 equ 277              ; FIX(1.082392200)
+F_1_414 equ 362              ; FIX(1.414213562)
+F_1_847 equ 473              ; FIX(1.847759065)
+F_2_613 equ 669              ; FIX(2.613125930)
+F_1_613 equ (F_2_613 - 256)  ; FIX(2.613125930) - FIX(1)
+%else
+; NASM cannot do compile-time arithmetic on floating-point constants.
+%define DESCALE(x, n)  (((x) + (1 << ((n) - 1))) >> (n))
+F_1_082 equ DESCALE(1162209775, 30 - CONST_BITS)  ; FIX(1.082392200)
+F_1_414 equ DESCALE(1518500249, 30 - CONST_BITS)  ; FIX(1.414213562)
+F_1_847 equ DESCALE(1984016188, 30 - CONST_BITS)  ; FIX(1.847759065)
+F_2_613 equ DESCALE(2805822602, 30 - CONST_BITS)  ; FIX(2.613125930)
+F_1_613 equ (F_2_613 - (1 << CONST_BITS))         ; FIX(2.613125930) - FIX(1)
+%endif
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_CONST
+
+; PRE_MULTIPLY_SCALE_BITS <= 2 (to avoid overflow)
+; CONST_BITS + CONST_SHIFT + PRE_MULTIPLY_SCALE_BITS == 16 (for pmulhw)
+
+%define PRE_MULTIPLY_SCALE_BITS  2
+%define CONST_SHIFT              (16 - PRE_MULTIPLY_SCALE_BITS - CONST_BITS)
+
+    alignz      32
+    GLOBAL_DATA(jconst_idct_ifast_sse2)
+
+EXTN(jconst_idct_ifast_sse2):
+
+PW_F1414       times 8  dw  F_1_414 << CONST_SHIFT
+PW_F1847       times 8  dw  F_1_847 << CONST_SHIFT
+PW_MF1613      times 8  dw -F_1_613 << CONST_SHIFT
+PW_F1082       times 8  dw  F_1_082 << CONST_SHIFT
+PB_CENTERJSAMP times 16 db  CENTERJSAMPLE
+
+    alignz      32
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_TEXT
+    BITS        64
+;
+; Perform dequantization and inverse DCT on one block of coefficients.
+;
+; GLOBAL(void)
+; jsimd_idct_ifast_sse2(void *dct_table, JCOEFPTR coef_block,
+;                      JSAMPARRAY output_buf, JDIMENSION output_col)
+;
+
+; r10 = jpeg_component_info *compptr
+; r11 = JCOEFPTR coef_block
+; r12 = JSAMPARRAY output_buf
+; r13d = JDIMENSION output_col
+
+%define original_rbp  rbp + 0
+%define wk(i)         rbp - (WK_NUM - (i)) * SIZEOF_XMMWORD
+                                        ; xmmword wk[WK_NUM]
+%define WK_NUM        2
+
+    align       32
+    GLOBAL_FUNCTION(jsimd_idct_ifast_sse2)
+
+EXTN(jsimd_idct_ifast_sse2):
+    push        rbp
+    mov         rax, rsp                     ; rax = original rbp
+    sub         rsp, byte 4
+    and         rsp, byte (-SIZEOF_XMMWORD)  ; align to 128 bits
+    mov         [rsp], rax
+    mov         rbp, rsp                     ; rbp = aligned rbp
+    lea         rsp, [wk(0)]
+    collect_args 4
+
+    ; ---- Pass 1: process columns from input.
+
+    mov         rdx, r10                ; quantptr
+    mov         rsi, r11                ; inptr
+
+%ifndef NO_ZERO_COLUMN_TEST_IFAST_SSE2
+    mov         eax, dword [DWBLOCK(1,0,rsi,SIZEOF_JCOEF)]
+    or          eax, dword [DWBLOCK(2,0,rsi,SIZEOF_JCOEF)]
+    jnz         near .columnDCT
+
+    movdqa      xmm0, XMMWORD [XMMBLOCK(1,0,rsi,SIZEOF_JCOEF)]
+    movdqa      xmm1, XMMWORD [XMMBLOCK(2,0,rsi,SIZEOF_JCOEF)]
+    por         xmm0, XMMWORD [XMMBLOCK(3,0,rsi,SIZEOF_JCOEF)]
+    por         xmm1, XMMWORD [XMMBLOCK(4,0,rsi,SIZEOF_JCOEF)]
+    por         xmm0, XMMWORD [XMMBLOCK(5,0,rsi,SIZEOF_JCOEF)]
+    por         xmm1, XMMWORD [XMMBLOCK(6,0,rsi,SIZEOF_JCOEF)]
+    por         xmm0, XMMWORD [XMMBLOCK(7,0,rsi,SIZEOF_JCOEF)]
+    por         xmm1, xmm0
+    packsswb    xmm1, xmm1
+    packsswb    xmm1, xmm1
+    movd        eax, xmm1
+    test        rax, rax
+    jnz         short .columnDCT
+
+    ; -- AC terms all zero
+
+    movdqa      xmm0, XMMWORD [XMMBLOCK(0,0,rsi,SIZEOF_JCOEF)]
+    pmullw      xmm0, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
+
+    movdqa      xmm7, xmm0              ; xmm0=in0=(00 01 02 03 04 05 06 07)
+    punpcklwd   xmm0, xmm0              ; xmm0=(00 00 01 01 02 02 03 03)
+    punpckhwd   xmm7, xmm7              ; xmm7=(04 04 05 05 06 06 07 07)
+
+    pshufd      xmm6, xmm0, 0x00        ; xmm6=col0=(00 00 00 00 00 00 00 00)
+    pshufd      xmm2, xmm0, 0x55        ; xmm2=col1=(01 01 01 01 01 01 01 01)
+    pshufd      xmm5, xmm0, 0xAA        ; xmm5=col2=(02 02 02 02 02 02 02 02)
+    pshufd      xmm0, xmm0, 0xFF        ; xmm0=col3=(03 03 03 03 03 03 03 03)
+    pshufd      xmm1, xmm7, 0x00        ; xmm1=col4=(04 04 04 04 04 04 04 04)
+    pshufd      xmm4, xmm7, 0x55        ; xmm4=col5=(05 05 05 05 05 05 05 05)
+    pshufd      xmm3, xmm7, 0xAA        ; xmm3=col6=(06 06 06 06 06 06 06 06)
+    pshufd      xmm7, xmm7, 0xFF        ; xmm7=col7=(07 07 07 07 07 07 07 07)
+
+    movdqa      XMMWORD [wk(0)], xmm2   ; wk(0)=col1
+    movdqa      XMMWORD [wk(1)], xmm0   ; wk(1)=col3
+    jmp         near .column_end
+%endif
+.columnDCT:
+
+    ; -- Even part
+
+    movdqa      xmm0, XMMWORD [XMMBLOCK(0,0,rsi,SIZEOF_JCOEF)]
+    movdqa      xmm1, XMMWORD [XMMBLOCK(2,0,rsi,SIZEOF_JCOEF)]
+    pmullw      xmm0, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_IFAST_MULT_TYPE)]
+    pmullw      xmm1, XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_IFAST_MULT_TYPE)]
+    movdqa      xmm2, XMMWORD [XMMBLOCK(4,0,rsi,SIZEOF_JCOEF)]
+    movdqa      xmm3, XMMWORD [XMMBLOCK(6,0,rsi,SIZEOF_JCOEF)]
+    pmullw      xmm2, XMMWORD [XMMBLOCK(4,0,rdx,SIZEOF_IFAST_MULT_TYPE)]
+    pmullw      xmm3, XMMWORD [XMMBLOCK(6,0,rdx,SIZEOF_IFAST_MULT_TYPE)]
+
+    movdqa      xmm4, xmm0
+    movdqa      xmm5, xmm1
+    psubw       xmm0, xmm2              ; xmm0=tmp11
+    psubw       xmm1, xmm3
+    paddw       xmm4, xmm2              ; xmm4=tmp10
+    paddw       xmm5, xmm3              ; xmm5=tmp13
+
+    psllw       xmm1, PRE_MULTIPLY_SCALE_BITS
+    pmulhw      xmm1, [rel PW_F1414]
+    psubw       xmm1, xmm5              ; xmm1=tmp12
+
+    movdqa      xmm6, xmm4
+    movdqa      xmm7, xmm0
+    psubw       xmm4, xmm5              ; xmm4=tmp3
+    psubw       xmm0, xmm1              ; xmm0=tmp2
+    paddw       xmm6, xmm5              ; xmm6=tmp0
+    paddw       xmm7, xmm1              ; xmm7=tmp1
+
+    movdqa      XMMWORD [wk(1)], xmm4   ; wk(1)=tmp3
+    movdqa      XMMWORD [wk(0)], xmm0   ; wk(0)=tmp2
+
+    ; -- Odd part
+
+    movdqa      xmm2, XMMWORD [XMMBLOCK(1,0,rsi,SIZEOF_JCOEF)]
+    movdqa      xmm3, XMMWORD [XMMBLOCK(3,0,rsi,SIZEOF_JCOEF)]
+    pmullw      xmm2, XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_IFAST_MULT_TYPE)]
+    pmullw      xmm3, XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_IFAST_MULT_TYPE)]
+    movdqa      xmm5, XMMWORD [XMMBLOCK(5,0,rsi,SIZEOF_JCOEF)]
+    movdqa      xmm1, XMMWORD [XMMBLOCK(7,0,rsi,SIZEOF_JCOEF)]
+    pmullw      xmm5, XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_IFAST_MULT_TYPE)]
+    pmullw      xmm1, XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_IFAST_MULT_TYPE)]
+
+    movdqa      xmm4, xmm2
+    movdqa      xmm0, xmm5
+    psubw       xmm2, xmm1              ; xmm2=z12
+    psubw       xmm5, xmm3              ; xmm5=z10
+    paddw       xmm4, xmm1              ; xmm4=z11
+    paddw       xmm0, xmm3              ; xmm0=z13
+
+    movdqa      xmm1, xmm5              ; xmm1=z10(unscaled)
+    psllw       xmm2, PRE_MULTIPLY_SCALE_BITS
+    psllw       xmm5, PRE_MULTIPLY_SCALE_BITS
+
+    movdqa      xmm3, xmm4
+    psubw       xmm4, xmm0
+    paddw       xmm3, xmm0              ; xmm3=tmp7
+
+    psllw       xmm4, PRE_MULTIPLY_SCALE_BITS
+    pmulhw      xmm4, [rel PW_F1414]    ; xmm4=tmp11
+
+    ; To avoid overflow...
+    ;
+    ; (Original)
+    ; tmp12 = -2.613125930 * z10 + z5;
+    ;
+    ; (This implementation)
+    ; tmp12 = (-1.613125930 - 1) * z10 + z5;
+    ;       = -1.613125930 * z10 - z10 + z5;
+
+    movdqa      xmm0, xmm5
+    paddw       xmm5, xmm2
+    pmulhw      xmm5, [rel PW_F1847]    ; xmm5=z5
+    pmulhw      xmm0, [rel PW_MF1613]
+    pmulhw      xmm2, [rel PW_F1082]
+    psubw       xmm0, xmm1
+    psubw       xmm2, xmm5              ; xmm2=tmp10
+    paddw       xmm0, xmm5              ; xmm0=tmp12
+
+    ; -- Final output stage
+
+    psubw       xmm0, xmm3              ; xmm0=tmp6
+    movdqa      xmm1, xmm6
+    movdqa      xmm5, xmm7
+    paddw       xmm6, xmm3              ; xmm6=data0=(00 01 02 03 04 05 06 07)
+    paddw       xmm7, xmm0              ; xmm7=data1=(10 11 12 13 14 15 16 17)
+    psubw       xmm1, xmm3              ; xmm1=data7=(70 71 72 73 74 75 76 77)
+    psubw       xmm5, xmm0              ; xmm5=data6=(60 61 62 63 64 65 66 67)
+    psubw       xmm4, xmm0              ; xmm4=tmp5
+
+    movdqa      xmm3, xmm6              ; transpose coefficients(phase 1)
+    punpcklwd   xmm6, xmm7              ; xmm6=(00 10 01 11 02 12 03 13)
+    punpckhwd   xmm3, xmm7              ; xmm3=(04 14 05 15 06 16 07 17)
+    movdqa      xmm0, xmm5              ; transpose coefficients(phase 1)
+    punpcklwd   xmm5, xmm1              ; xmm5=(60 70 61 71 62 72 63 73)
+    punpckhwd   xmm0, xmm1              ; xmm0=(64 74 65 75 66 76 67 77)
+
+    movdqa      xmm7, XMMWORD [wk(0)]   ; xmm7=tmp2
+    movdqa      xmm1, XMMWORD [wk(1)]   ; xmm1=tmp3
+
+    movdqa      XMMWORD [wk(0)], xmm5   ; wk(0)=(60 70 61 71 62 72 63 73)
+    movdqa      XMMWORD [wk(1)], xmm0   ; wk(1)=(64 74 65 75 66 76 67 77)
+
+    paddw       xmm2, xmm4              ; xmm2=tmp4
+    movdqa      xmm5, xmm7
+    movdqa      xmm0, xmm1
+    paddw       xmm7, xmm4              ; xmm7=data2=(20 21 22 23 24 25 26 27)
+    paddw       xmm1, xmm2              ; xmm1=data4=(40 41 42 43 44 45 46 47)
+    psubw       xmm5, xmm4              ; xmm5=data5=(50 51 52 53 54 55 56 57)
+    psubw       xmm0, xmm2              ; xmm0=data3=(30 31 32 33 34 35 36 37)
+
+    movdqa      xmm4, xmm7              ; transpose coefficients(phase 1)
+    punpcklwd   xmm7, xmm0              ; xmm7=(20 30 21 31 22 32 23 33)
+    punpckhwd   xmm4, xmm0              ; xmm4=(24 34 25 35 26 36 27 37)
+    movdqa      xmm2, xmm1              ; transpose coefficients(phase 1)
+    punpcklwd   xmm1, xmm5              ; xmm1=(40 50 41 51 42 52 43 53)
+    punpckhwd   xmm2, xmm5              ; xmm2=(44 54 45 55 46 56 47 57)
+
+    movdqa      xmm0, xmm3              ; transpose coefficients(phase 2)
+    punpckldq   xmm3, xmm4              ; xmm3=(04 14 24 34 05 15 25 35)
+    punpckhdq   xmm0, xmm4              ; xmm0=(06 16 26 36 07 17 27 37)
+    movdqa      xmm5, xmm6              ; transpose coefficients(phase 2)
+    punpckldq   xmm6, xmm7              ; xmm6=(00 10 20 30 01 11 21 31)
+    punpckhdq   xmm5, xmm7              ; xmm5=(02 12 22 32 03 13 23 33)
+
+    movdqa      xmm4, XMMWORD [wk(0)]   ; xmm4=(60 70 61 71 62 72 63 73)
+    movdqa      xmm7, XMMWORD [wk(1)]   ; xmm7=(64 74 65 75 66 76 67 77)
+
+    movdqa      XMMWORD [wk(0)], xmm3   ; wk(0)=(04 14 24 34 05 15 25 35)
+    movdqa      XMMWORD [wk(1)], xmm0   ; wk(1)=(06 16 26 36 07 17 27 37)
+
+    movdqa      xmm3, xmm1              ; transpose coefficients(phase 2)
+    punpckldq   xmm1, xmm4              ; xmm1=(40 50 60 70 41 51 61 71)
+    punpckhdq   xmm3, xmm4              ; xmm3=(42 52 62 72 43 53 63 73)
+    movdqa      xmm0, xmm2              ; transpose coefficients(phase 2)
+    punpckldq   xmm2, xmm7              ; xmm2=(44 54 64 74 45 55 65 75)
+    punpckhdq   xmm0, xmm7              ; xmm0=(46 56 66 76 47 57 67 77)
+
+    movdqa      xmm4, xmm6              ; transpose coefficients(phase 3)
+    punpcklqdq  xmm6, xmm1              ; xmm6=col0=(00 10 20 30 40 50 60 70)
+    punpckhqdq  xmm4, xmm1              ; xmm4=col1=(01 11 21 31 41 51 61 71)
+    movdqa      xmm7, xmm5              ; transpose coefficients(phase 3)
+    punpcklqdq  xmm5, xmm3              ; xmm5=col2=(02 12 22 32 42 52 62 72)
+    punpckhqdq  xmm7, xmm3              ; xmm7=col3=(03 13 23 33 43 53 63 73)
+
+    movdqa      xmm1, XMMWORD [wk(0)]   ; xmm1=(04 14 24 34 05 15 25 35)
+    movdqa      xmm3, XMMWORD [wk(1)]   ; xmm3=(06 16 26 36 07 17 27 37)
+
+    movdqa      XMMWORD [wk(0)], xmm4   ; wk(0)=col1
+    movdqa      XMMWORD [wk(1)], xmm7   ; wk(1)=col3
+
+    movdqa      xmm4, xmm1              ; transpose coefficients(phase 3)
+    punpcklqdq  xmm1, xmm2              ; xmm1=col4=(04 14 24 34 44 54 64 74)
+    punpckhqdq  xmm4, xmm2              ; xmm4=col5=(05 15 25 35 45 55 65 75)
+    movdqa      xmm7, xmm3              ; transpose coefficients(phase 3)
+    punpcklqdq  xmm3, xmm0              ; xmm3=col6=(06 16 26 36 46 56 66 76)
+    punpckhqdq  xmm7, xmm0              ; xmm7=col7=(07 17 27 37 47 57 67 77)
+.column_end:
+
+    ; -- Prefetch the next coefficient block
+
+    prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 0*32]
+    prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 1*32]
+    prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 2*32]
+    prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 3*32]
+
+    ; ---- Pass 2: process rows from work array, store into output array.
+
+    mov         rax, [original_rbp]
+    mov         rdi, r12                ; (JSAMPROW *)
+    mov         eax, r13d
+
+    ; -- Even part
+
+    ; xmm6=col0, xmm5=col2, xmm1=col4, xmm3=col6
+
+    movdqa      xmm2, xmm6
+    movdqa      xmm0, xmm5
+    psubw       xmm6, xmm1              ; xmm6=tmp11
+    psubw       xmm5, xmm3
+    paddw       xmm2, xmm1              ; xmm2=tmp10
+    paddw       xmm0, xmm3              ; xmm0=tmp13
+
+    psllw       xmm5, PRE_MULTIPLY_SCALE_BITS
+    pmulhw      xmm5, [rel PW_F1414]
+    psubw       xmm5, xmm0              ; xmm5=tmp12
+
+    movdqa      xmm1, xmm2
+    movdqa      xmm3, xmm6
+    psubw       xmm2, xmm0              ; xmm2=tmp3
+    psubw       xmm6, xmm5              ; xmm6=tmp2
+    paddw       xmm1, xmm0              ; xmm1=tmp0
+    paddw       xmm3, xmm5              ; xmm3=tmp1
+
+    movdqa      xmm0, XMMWORD [wk(0)]   ; xmm0=col1
+    movdqa      xmm5, XMMWORD [wk(1)]   ; xmm5=col3
+
+    movdqa      XMMWORD [wk(0)], xmm2   ; wk(0)=tmp3
+    movdqa      XMMWORD [wk(1)], xmm6   ; wk(1)=tmp2
+
+    ; -- Odd part
+
+    ; xmm0=col1, xmm5=col3, xmm4=col5, xmm7=col7
+
+    movdqa      xmm2, xmm0
+    movdqa      xmm6, xmm4
+    psubw       xmm0, xmm7              ; xmm0=z12
+    psubw       xmm4, xmm5              ; xmm4=z10
+    paddw       xmm2, xmm7              ; xmm2=z11
+    paddw       xmm6, xmm5              ; xmm6=z13
+
+    movdqa      xmm7, xmm4              ; xmm7=z10(unscaled)
+    psllw       xmm0, PRE_MULTIPLY_SCALE_BITS
+    psllw       xmm4, PRE_MULTIPLY_SCALE_BITS
+
+    movdqa      xmm5, xmm2
+    psubw       xmm2, xmm6
+    paddw       xmm5, xmm6              ; xmm5=tmp7
+
+    psllw       xmm2, PRE_MULTIPLY_SCALE_BITS
+    pmulhw      xmm2, [rel PW_F1414]    ; xmm2=tmp11
+
+    ; To avoid overflow...
+    ;
+    ; (Original)
+    ; tmp12 = -2.613125930 * z10 + z5;
+    ;
+    ; (This implementation)
+    ; tmp12 = (-1.613125930 - 1) * z10 + z5;
+    ;       = -1.613125930 * z10 - z10 + z5;
+
+    movdqa      xmm6, xmm4
+    paddw       xmm4, xmm0
+    pmulhw      xmm4, [rel PW_F1847]    ; xmm4=z5
+    pmulhw      xmm6, [rel PW_MF1613]
+    pmulhw      xmm0, [rel PW_F1082]
+    psubw       xmm6, xmm7
+    psubw       xmm0, xmm4              ; xmm0=tmp10
+    paddw       xmm6, xmm4              ; xmm6=tmp12
+
+    ; -- Final output stage
+
+    psubw       xmm6, xmm5              ; xmm6=tmp6
+    movdqa      xmm7, xmm1
+    movdqa      xmm4, xmm3
+    paddw       xmm1, xmm5              ; xmm1=data0=(00 10 20 30 40 50 60 70)
+    paddw       xmm3, xmm6              ; xmm3=data1=(01 11 21 31 41 51 61 71)
+    psraw       xmm1, (PASS1_BITS+3)    ; descale
+    psraw       xmm3, (PASS1_BITS+3)    ; descale
+    psubw       xmm7, xmm5              ; xmm7=data7=(07 17 27 37 47 57 67 77)
+    psubw       xmm4, xmm6              ; xmm4=data6=(06 16 26 36 46 56 66 76)
+    psraw       xmm7, (PASS1_BITS+3)    ; descale
+    psraw       xmm4, (PASS1_BITS+3)    ; descale
+    psubw       xmm2, xmm6              ; xmm2=tmp5
+
+    packsswb    xmm1, xmm4        ; xmm1=(00 10 20 30 40 50 60 70 06 16 26 36 46 56 66 76)
+    packsswb    xmm3, xmm7        ; xmm3=(01 11 21 31 41 51 61 71 07 17 27 37 47 57 67 77)
+
+    movdqa      xmm5, XMMWORD [wk(1)]   ; xmm5=tmp2
+    movdqa      xmm6, XMMWORD [wk(0)]   ; xmm6=tmp3
+
+    paddw       xmm0, xmm2              ; xmm0=tmp4
+    movdqa      xmm4, xmm5
+    movdqa      xmm7, xmm6
+    paddw       xmm5, xmm2              ; xmm5=data2=(02 12 22 32 42 52 62 72)
+    paddw       xmm6, xmm0              ; xmm6=data4=(04 14 24 34 44 54 64 74)
+    psraw       xmm5, (PASS1_BITS+3)    ; descale
+    psraw       xmm6, (PASS1_BITS+3)    ; descale
+    psubw       xmm4, xmm2              ; xmm4=data5=(05 15 25 35 45 55 65 75)
+    psubw       xmm7, xmm0              ; xmm7=data3=(03 13 23 33 43 53 63 73)
+    psraw       xmm4, (PASS1_BITS+3)    ; descale
+    psraw       xmm7, (PASS1_BITS+3)    ; descale
+
+    movdqa      xmm2, [rel PB_CENTERJSAMP]  ; xmm2=[rel PB_CENTERJSAMP]
+
+    packsswb    xmm5, xmm6        ; xmm5=(02 12 22 32 42 52 62 72 04 14 24 34 44 54 64 74)
+    packsswb    xmm7, xmm4        ; xmm7=(03 13 23 33 43 53 63 73 05 15 25 35 45 55 65 75)
+
+    paddb       xmm1, xmm2
+    paddb       xmm3, xmm2
+    paddb       xmm5, xmm2
+    paddb       xmm7, xmm2
+
+    movdqa      xmm0, xmm1        ; transpose coefficients(phase 1)
+    punpcklbw   xmm1, xmm3        ; xmm1=(00 01 10 11 20 21 30 31 40 41 50 51 60 61 70 71)
+    punpckhbw   xmm0, xmm3        ; xmm0=(06 07 16 17 26 27 36 37 46 47 56 57 66 67 76 77)
+    movdqa      xmm6, xmm5        ; transpose coefficients(phase 1)
+    punpcklbw   xmm5, xmm7        ; xmm5=(02 03 12 13 22 23 32 33 42 43 52 53 62 63 72 73)
+    punpckhbw   xmm6, xmm7        ; xmm6=(04 05 14 15 24 25 34 35 44 45 54 55 64 65 74 75)
+
+    movdqa      xmm4, xmm1        ; transpose coefficients(phase 2)
+    punpcklwd   xmm1, xmm5        ; xmm1=(00 01 02 03 10 11 12 13 20 21 22 23 30 31 32 33)
+    punpckhwd   xmm4, xmm5        ; xmm4=(40 41 42 43 50 51 52 53 60 61 62 63 70 71 72 73)
+    movdqa      xmm2, xmm6        ; transpose coefficients(phase 2)
+    punpcklwd   xmm6, xmm0        ; xmm6=(04 05 06 07 14 15 16 17 24 25 26 27 34 35 36 37)
+    punpckhwd   xmm2, xmm0        ; xmm2=(44 45 46 47 54 55 56 57 64 65 66 67 74 75 76 77)
+
+    movdqa      xmm3, xmm1        ; transpose coefficients(phase 3)
+    punpckldq   xmm1, xmm6        ; xmm1=(00 01 02 03 04 05 06 07 10 11 12 13 14 15 16 17)
+    punpckhdq   xmm3, xmm6        ; xmm3=(20 21 22 23 24 25 26 27 30 31 32 33 34 35 36 37)
+    movdqa      xmm7, xmm4        ; transpose coefficients(phase 3)
+    punpckldq   xmm4, xmm2        ; xmm4=(40 41 42 43 44 45 46 47 50 51 52 53 54 55 56 57)
+    punpckhdq   xmm7, xmm2        ; xmm7=(60 61 62 63 64 65 66 67 70 71 72 73 74 75 76 77)
+
+    pshufd      xmm5, xmm1, 0x4E  ; xmm5=(10 11 12 13 14 15 16 17 00 01 02 03 04 05 06 07)
+    pshufd      xmm0, xmm3, 0x4E  ; xmm0=(30 31 32 33 34 35 36 37 20 21 22 23 24 25 26 27)
+    pshufd      xmm6, xmm4, 0x4E  ; xmm6=(50 51 52 53 54 55 56 57 40 41 42 43 44 45 46 47)
+    pshufd      xmm2, xmm7, 0x4E  ; xmm2=(70 71 72 73 74 75 76 77 60 61 62 63 64 65 66 67)
+
+    mov         rdxp, JSAMPROW [rdi+0*SIZEOF_JSAMPROW]
+    mov         rsip, JSAMPROW [rdi+2*SIZEOF_JSAMPROW]
+    movq        XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm1
+    movq        XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm3
+    mov         rdxp, JSAMPROW [rdi+4*SIZEOF_JSAMPROW]
+    mov         rsip, JSAMPROW [rdi+6*SIZEOF_JSAMPROW]
+    movq        XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm4
+    movq        XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm7
+
+    mov         rdxp, JSAMPROW [rdi+1*SIZEOF_JSAMPROW]
+    mov         rsip, JSAMPROW [rdi+3*SIZEOF_JSAMPROW]
+    movq        XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm5
+    movq        XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm0
+    mov         rdxp, JSAMPROW [rdi+5*SIZEOF_JSAMPROW]
+    mov         rsip, JSAMPROW [rdi+7*SIZEOF_JSAMPROW]
+    movq        XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm6
+    movq        XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm2
+
+    uncollect_args 4
+    mov         rsp, rbp                ; rsp <- aligned rbp
+    pop         rsp                     ; rsp <- original rbp
+    pop         rbp
+    ret
+    ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+    align       32
diff --git a/media/libjpeg/simd/x86_64/jidctint-avx2.asm b/media/libjpeg/simd/x86_64/jidctint-avx2.asm
new file mode 100644
index 0000000000..ca7e317f6e
--- /dev/null
+++ b/media/libjpeg/simd/x86_64/jidctint-avx2.asm
@@ -0,0 +1,418 @@
+;
+; jidctint.asm - accurate integer IDCT (64-bit AVX2)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+; Copyright (C) 2009, 2016, 2018, 2020, D. R. Commander.
+; Copyright (C) 2018, Matthias Räncker.
+;
+; Based on the x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; This file contains a slower but more accurate integer implementation of the
+; inverse DCT (Discrete Cosine Transform). The following code is based
+; directly on the IJG's original jidctint.c; see the jidctint.c for
+; more details.
+
+%include "jsimdext.inc"
+%include "jdct.inc"
+
+; --------------------------------------------------------------------------
+
+%define CONST_BITS  13
+%define PASS1_BITS  2
+
+%define DESCALE_P1  (CONST_BITS - PASS1_BITS)
+%define DESCALE_P2  (CONST_BITS + PASS1_BITS + 3)
+
+%if CONST_BITS == 13
+F_0_298 equ  2446  ; FIX(0.298631336)
+F_0_390 equ  3196  ; FIX(0.390180644)
+F_0_541 equ  4433  ; FIX(0.541196100)
+F_0_765 equ  6270  ; FIX(0.765366865)
+F_0_899 equ  7373  ; FIX(0.899976223)
+F_1_175 equ  9633  ; FIX(1.175875602)
+F_1_501 equ 12299  ; FIX(1.501321110)
+F_1_847 equ 15137  ; FIX(1.847759065)
+F_1_961 equ 16069  ; FIX(1.961570560)
+F_2_053 equ 16819  ; FIX(2.053119869)
+F_2_562 equ 20995  ; FIX(2.562915447)
+F_3_072 equ 25172  ; FIX(3.072711026)
+%else
+; NASM cannot do compile-time arithmetic on floating-point constants.
+%define DESCALE(x, n)  (((x) + (1 << ((n) - 1))) >> (n))
+F_0_298 equ DESCALE( 320652955, 30 - CONST_BITS)  ; FIX(0.298631336)
+F_0_390 equ DESCALE( 418953276, 30 - CONST_BITS)  ; FIX(0.390180644)
+F_0_541 equ DESCALE( 581104887, 30 - CONST_BITS)  ; FIX(0.541196100)
+F_0_765 equ DESCALE( 821806413, 30 - CONST_BITS)  ; FIX(0.765366865)
+F_0_899 equ DESCALE( 966342111, 30 - CONST_BITS)  ; FIX(0.899976223)
+F_1_175 equ DESCALE(1262586813, 30 - CONST_BITS)  ; FIX(1.175875602)
+F_1_501 equ DESCALE(1612031267, 30 - CONST_BITS)  ; FIX(1.501321110)
+F_1_847 equ DESCALE(1984016188, 30 - CONST_BITS)  ; FIX(1.847759065)
+F_1_961 equ DESCALE(2106220350, 30 - CONST_BITS)  ; FIX(1.961570560)
+F_2_053 equ DESCALE(2204520673, 30 - CONST_BITS)  ; FIX(2.053119869)
+F_2_562 equ DESCALE(2751909506, 30 - CONST_BITS)  ; FIX(2.562915447)
+F_3_072 equ DESCALE(3299298341, 30 - CONST_BITS)  ; FIX(3.072711026)
+%endif
+
+; --------------------------------------------------------------------------
+; In-place 8x8x16-bit inverse matrix transpose using AVX2 instructions
+; %1-%4: Input/output registers
+; %5-%8: Temp registers
+
+%macro dotranspose 8
+    ; %5=(00 10 20 30 40 50 60 70  01 11 21 31 41 51 61 71)
+    ; %6=(03 13 23 33 43 53 63 73  02 12 22 32 42 52 62 72)
+    ; %7=(04 14 24 34 44 54 64 74  05 15 25 35 45 55 65 75)
+    ; %8=(07 17 27 37 47 57 67 77  06 16 26 36 46 56 66 76)
+
+    vpermq      %5, %1, 0xD8
+    vpermq      %6, %2, 0x72
+    vpermq      %7, %3, 0xD8
+    vpermq      %8, %4, 0x72
+    ; transpose coefficients(phase 1)
+    ; %5=(00 10 20 30 01 11 21 31  40 50 60 70 41 51 61 71)
+    ; %6=(02 12 22 32 03 13 23 33  42 52 62 72 43 53 63 73)
+    ; %7=(04 14 24 34 05 15 25 35  44 54 64 74 45 55 65 75)
+    ; %8=(06 16 26 36 07 17 27 37  46 56 66 76 47 57 67 77)
+
+    vpunpcklwd  %1, %5, %6
+    vpunpckhwd  %2, %5, %6
+    vpunpcklwd  %3, %7, %8
+    vpunpckhwd  %4, %7, %8
+    ; transpose coefficients(phase 2)
+    ; %1=(00 02 10 12 20 22 30 32  40 42 50 52 60 62 70 72)
+    ; %2=(01 03 11 13 21 23 31 33  41 43 51 53 61 63 71 73)
+    ; %3=(04 06 14 16 24 26 34 36  44 46 54 56 64 66 74 76)
+    ; %4=(05 07 15 17 25 27 35 37  45 47 55 57 65 67 75 77)
+
+    vpunpcklwd  %5, %1, %2
+    vpunpcklwd  %6, %3, %4
+    vpunpckhwd  %7, %1, %2
+    vpunpckhwd  %8, %3, %4
+    ; transpose coefficients(phase 3)
+    ; %5=(00 01 02 03 10 11 12 13  40 41 42 43 50 51 52 53)
+    ; %6=(04 05 06 07 14 15 16 17  44 45 46 47 54 55 56 57)
+    ; %7=(20 21 22 23 30 31 32 33  60 61 62 63 70 71 72 73)
+    ; %8=(24 25 26 27 34 35 36 37  64 65 66 67 74 75 76 77)
+
+    vpunpcklqdq %1, %5, %6
+    vpunpckhqdq %2, %5, %6
+    vpunpcklqdq %3, %7, %8
+    vpunpckhqdq %4, %7, %8
+    ; transpose coefficients(phase 4)
+    ; %1=(00 01 02 03 04 05 06 07  40 41 42 43 44 45 46 47)
+    ; %2=(10 11 12 13 14 15 16 17  50 51 52 53 54 55 56 57)
+    ; %3=(20 21 22 23 24 25 26 27  60 61 62 63 64 65 66 67)
+    ; %4=(30 31 32 33 34 35 36 37  70 71 72 73 74 75 76 77)
+%endmacro
+
+; --------------------------------------------------------------------------
+; In-place 8x8x16-bit accurate integer inverse DCT using AVX2 instructions
+; %1-%4:  Input/output registers
+; %5-%12: Temp registers
+; %9:     Pass (1 or 2)
+
+%macro dodct 13
+    ; -- Even part
+
+    ; (Original)
+    ; z1 = (z2 + z3) * 0.541196100;
+    ; tmp2 = z1 + z3 * -1.847759065;
+    ; tmp3 = z1 + z2 * 0.765366865;
+    ;
+    ; (This implementation)
+    ; tmp2 = z2 * 0.541196100 + z3 * (0.541196100 - 1.847759065);
+    ; tmp3 = z2 * (0.541196100 + 0.765366865) + z3 * 0.541196100;
+
+    vperm2i128  %6, %3, %3, 0x01        ; %6=in6_2
+    vpunpcklwd  %5, %3, %6              ; %5=in26_62L
+    vpunpckhwd  %6, %3, %6              ; %6=in26_62H
+    vpmaddwd    %5, %5, [rel PW_F130_F054_MF130_F054]  ; %5=tmp3_2L
+    vpmaddwd    %6, %6, [rel PW_F130_F054_MF130_F054]  ; %6=tmp3_2H
+
+    vperm2i128  %7, %1, %1, 0x01        ; %7=in4_0
+    vpsignw     %1, %1, [rel PW_1_NEG1]
+    vpaddw      %7, %7, %1              ; %7=(in0+in4)_(in0-in4)
+
+    vpxor       %1, %1, %1
+    vpunpcklwd  %8, %1, %7              ; %8=tmp0_1L
+    vpunpckhwd  %1, %1, %7              ; %1=tmp0_1H
+    vpsrad      %8, %8, (16-CONST_BITS)  ; vpsrad %8,16 & vpslld %8,CONST_BITS
+    vpsrad      %1, %1, (16-CONST_BITS)  ; vpsrad %1,16 & vpslld %1,CONST_BITS
+
+    vpsubd      %11, %8, %5             ; %11=tmp0_1L-tmp3_2L=tmp13_12L
+    vpaddd      %9, %8, %5              ; %9=tmp0_1L+tmp3_2L=tmp10_11L
+    vpsubd      %12, %1, %6             ; %12=tmp0_1H-tmp3_2H=tmp13_12H
+    vpaddd      %10, %1, %6             ; %10=tmp0_1H+tmp3_2H=tmp10_11H
+
+    ; -- Odd part
+
+    vpaddw      %1, %4, %2              ; %1=in7_5+in3_1=z3_4
+
+    ; (Original)
+    ; z5 = (z3 + z4) * 1.175875602;
+    ; z3 = z3 * -1.961570560;  z4 = z4 * -0.390180644;
+    ; z3 += z5;  z4 += z5;
+    ;
+    ; (This implementation)
+    ; z3 = z3 * (1.175875602 - 1.961570560) + z4 * 1.175875602;
+    ; z4 = z3 * 1.175875602 + z4 * (1.175875602 - 0.390180644);
+
+    vperm2i128  %8, %1, %1, 0x01        ; %8=z4_3
+    vpunpcklwd  %7, %1, %8              ; %7=z34_43L
+    vpunpckhwd  %8, %1, %8              ; %8=z34_43H
+    vpmaddwd    %7, %7, [rel PW_MF078_F117_F078_F117]  ; %7=z3_4L
+    vpmaddwd    %8, %8, [rel PW_MF078_F117_F078_F117]  ; %8=z3_4H
+
+    ; (Original)
+    ; z1 = tmp0 + tmp3;  z2 = tmp1 + tmp2;
+    ; tmp0 = tmp0 * 0.298631336;  tmp1 = tmp1 * 2.053119869;
+    ; tmp2 = tmp2 * 3.072711026;  tmp3 = tmp3 * 1.501321110;
+    ; z1 = z1 * -0.899976223;  z2 = z2 * -2.562915447;
+    ; tmp0 += z1 + z3;  tmp1 += z2 + z4;
+    ; tmp2 += z2 + z3;  tmp3 += z1 + z4;
+    ;
+    ; (This implementation)
+    ; tmp0 = tmp0 * (0.298631336 - 0.899976223) + tmp3 * -0.899976223;
+    ; tmp1 = tmp1 * (2.053119869 - 2.562915447) + tmp2 * -2.562915447;
+    ; tmp2 = tmp1 * -2.562915447 + tmp2 * (3.072711026 - 2.562915447);
+    ; tmp3 = tmp0 * -0.899976223 + tmp3 * (1.501321110 - 0.899976223);
+    ; tmp0 += z3;  tmp1 += z4;
+    ; tmp2 += z3;  tmp3 += z4;
+
+    vperm2i128  %2, %2, %2, 0x01        ; %2=in1_3
+    vpunpcklwd  %3, %4, %2              ; %3=in71_53L
+    vpunpckhwd  %4, %4, %2              ; %4=in71_53H
+
+    vpmaddwd    %5, %3, [rel PW_MF060_MF089_MF050_MF256]  ; %5=tmp0_1L
+    vpmaddwd    %6, %4, [rel PW_MF060_MF089_MF050_MF256]  ; %6=tmp0_1H
+    vpaddd      %5, %5, %7              ; %5=tmp0_1L+z3_4L=tmp0_1L
+    vpaddd      %6, %6, %8              ; %6=tmp0_1H+z3_4H=tmp0_1H
+
+    vpmaddwd    %3, %3, [rel PW_MF089_F060_MF256_F050]  ; %3=tmp3_2L
+    vpmaddwd    %4, %4, [rel PW_MF089_F060_MF256_F050]  ; %4=tmp3_2H
+    vperm2i128  %7, %7, %7, 0x01        ; %7=z4_3L
+    vperm2i128  %8, %8, %8, 0x01        ; %8=z4_3H
+    vpaddd      %7, %3, %7              ; %7=tmp3_2L+z4_3L=tmp3_2L
+    vpaddd      %8, %4, %8              ; %8=tmp3_2H+z4_3H=tmp3_2H
+
+    ; -- Final output stage
+
+    vpaddd      %1, %9, %7              ; %1=tmp10_11L+tmp3_2L=data0_1L
+    vpaddd      %2, %10, %8             ; %2=tmp10_11H+tmp3_2H=data0_1H
+    vpaddd      %1, %1, [rel PD_DESCALE_P %+ %13]
+    vpaddd      %2, %2, [rel PD_DESCALE_P %+ %13]
+    vpsrad      %1, %1, DESCALE_P %+ %13
+    vpsrad      %2, %2, DESCALE_P %+ %13
+    vpackssdw   %1, %1, %2              ; %1=data0_1
+
+    vpsubd      %3, %9, %7              ; %3=tmp10_11L-tmp3_2L=data7_6L
+    vpsubd      %4, %10, %8             ; %4=tmp10_11H-tmp3_2H=data7_6H
+    vpaddd      %3, %3, [rel PD_DESCALE_P %+ %13]
+    vpaddd      %4, %4, [rel PD_DESCALE_P %+ %13]
+    vpsrad      %3, %3, DESCALE_P %+ %13
+    vpsrad      %4, %4, DESCALE_P %+ %13
+    vpackssdw   %4, %3, %4              ; %4=data7_6
+
+    vpaddd      %7, %11, %5             ; %7=tmp13_12L+tmp0_1L=data3_2L
+    vpaddd      %8, %12, %6             ; %8=tmp13_12H+tmp0_1H=data3_2H
+    vpaddd      %7, %7, [rel PD_DESCALE_P %+ %13]
+    vpaddd      %8, %8, [rel PD_DESCALE_P %+ %13]
+    vpsrad      %7, %7, DESCALE_P %+ %13
+    vpsrad      %8, %8, DESCALE_P %+ %13
+    vpackssdw   %2, %7, %8              ; %2=data3_2
+
+    vpsubd      %7, %11, %5             ; %7=tmp13_12L-tmp0_1L=data4_5L
+    vpsubd      %8, %12, %6             ; %8=tmp13_12H-tmp0_1H=data4_5H
+    vpaddd      %7, %7, [rel PD_DESCALE_P %+ %13]
+    vpaddd      %8, %8, [rel PD_DESCALE_P %+ %13]
+    vpsrad      %7, %7, DESCALE_P %+ %13
+    vpsrad      %8, %8, DESCALE_P %+ %13
+    vpackssdw   %3, %7, %8              ; %3=data4_5
+%endmacro
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_CONST
+
+    alignz      32
+    GLOBAL_DATA(jconst_idct_islow_avx2)
+
+EXTN(jconst_idct_islow_avx2):
+
+PW_F130_F054_MF130_F054    times 4  dw  (F_0_541 + F_0_765),  F_0_541
+                           times 4  dw  (F_0_541 - F_1_847),  F_0_541
+PW_MF078_F117_F078_F117    times 4  dw  (F_1_175 - F_1_961),  F_1_175
+                           times 4  dw  (F_1_175 - F_0_390),  F_1_175
+PW_MF060_MF089_MF050_MF256 times 4  dw  (F_0_298 - F_0_899), -F_0_899
+                           times 4  dw  (F_2_053 - F_2_562), -F_2_562
+PW_MF089_F060_MF256_F050   times 4  dw -F_0_899, (F_1_501 - F_0_899)
+                           times 4  dw -F_2_562, (F_3_072 - F_2_562)
+PD_DESCALE_P1              times 8  dd  1 << (DESCALE_P1 - 1)
+PD_DESCALE_P2              times 8  dd  1 << (DESCALE_P2 - 1)
+PB_CENTERJSAMP             times 32 db  CENTERJSAMPLE
+PW_1_NEG1                  times 8  dw  1
+                           times 8  dw -1
+
+    alignz      32
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_TEXT
+    BITS        64
+;
+; Perform dequantization and inverse DCT on one block of coefficients.
+;
+; GLOBAL(void)
+; jsimd_idct_islow_avx2(void *dct_table, JCOEFPTR coef_block,
+;                       JSAMPARRAY output_buf, JDIMENSION output_col)
+;
+
+; r10 = jpeg_component_info *compptr
+; r11 = JCOEFPTR coef_block
+; r12 = JSAMPARRAY output_buf
+; r13d = JDIMENSION output_col
+
+    align       32
+    GLOBAL_FUNCTION(jsimd_idct_islow_avx2)
+
+EXTN(jsimd_idct_islow_avx2):
+    push        rbp
+    mov         rax, rsp                     ; rax = original rbp
+    mov         rbp, rsp                     ; rbp = aligned rbp
+    push_xmm    4
+    collect_args 4
+
+    ; ---- Pass 1: process columns.
+
+%ifndef NO_ZERO_COLUMN_TEST_ISLOW_AVX2
+    mov         eax, dword [DWBLOCK(1,0,r11,SIZEOF_JCOEF)]
+    or          eax, dword [DWBLOCK(2,0,r11,SIZEOF_JCOEF)]
+    jnz         near .columnDCT
+
+    movdqa      xmm0, XMMWORD [XMMBLOCK(1,0,r11,SIZEOF_JCOEF)]
+    movdqa      xmm1, XMMWORD [XMMBLOCK(2,0,r11,SIZEOF_JCOEF)]
+    vpor        xmm0, xmm0, XMMWORD [XMMBLOCK(3,0,r11,SIZEOF_JCOEF)]
+    vpor        xmm1, xmm1, XMMWORD [XMMBLOCK(4,0,r11,SIZEOF_JCOEF)]
+    vpor        xmm0, xmm0, XMMWORD [XMMBLOCK(5,0,r11,SIZEOF_JCOEF)]
+    vpor        xmm1, xmm1, XMMWORD [XMMBLOCK(6,0,r11,SIZEOF_JCOEF)]
+    vpor        xmm0, xmm0, XMMWORD [XMMBLOCK(7,0,r11,SIZEOF_JCOEF)]
+    vpor        xmm1, xmm1, xmm0
+    vpacksswb   xmm1, xmm1, xmm1
+    vpacksswb   xmm1, xmm1, xmm1
+    movd        eax, xmm1
+    test        rax, rax
+    jnz         short .columnDCT
+
+    ; -- AC terms all zero
+
+    movdqa      xmm5, XMMWORD [XMMBLOCK(0,0,r11,SIZEOF_JCOEF)]
+    vpmullw     xmm5, xmm5, XMMWORD [XMMBLOCK(0,0,r10,SIZEOF_ISLOW_MULT_TYPE)]
+
+    vpsllw      xmm5, xmm5, PASS1_BITS
+
+    vpunpcklwd  xmm4, xmm5, xmm5        ; xmm4=(00 00 01 01 02 02 03 03)
+    vpunpckhwd  xmm5, xmm5, xmm5        ; xmm5=(04 04 05 05 06 06 07 07)
+    vinserti128 ymm4, ymm4, xmm5, 1
+
+    vpshufd     ymm0, ymm4, 0x00        ; ymm0=col0_4=(00 00 00 00 00 00 00 00  04 04 04 04 04 04 04 04)
+    vpshufd     ymm1, ymm4, 0x55        ; ymm1=col1_5=(01 01 01 01 01 01 01 01  05 05 05 05 05 05 05 05)
+    vpshufd     ymm2, ymm4, 0xAA        ; ymm2=col2_6=(02 02 02 02 02 02 02 02  06 06 06 06 06 06 06 06)
+    vpshufd     ymm3, ymm4, 0xFF        ; ymm3=col3_7=(03 03 03 03 03 03 03 03  07 07 07 07 07 07 07 07)
+
+    jmp         near .column_end
+%endif
+.columnDCT:
+
+    vmovdqu     ymm4, YMMWORD [YMMBLOCK(0,0,r11,SIZEOF_JCOEF)]  ; ymm4=in0_1
+    vmovdqu     ymm5, YMMWORD [YMMBLOCK(2,0,r11,SIZEOF_JCOEF)]  ; ymm5=in2_3
+    vmovdqu     ymm6, YMMWORD [YMMBLOCK(4,0,r11,SIZEOF_JCOEF)]  ; ymm6=in4_5
+    vmovdqu     ymm7, YMMWORD [YMMBLOCK(6,0,r11,SIZEOF_JCOEF)]  ; ymm7=in6_7
+    vpmullw     ymm4, ymm4, YMMWORD [YMMBLOCK(0,0,r10,SIZEOF_ISLOW_MULT_TYPE)]
+    vpmullw     ymm5, ymm5, YMMWORD [YMMBLOCK(2,0,r10,SIZEOF_ISLOW_MULT_TYPE)]
+    vpmullw     ymm6, ymm6, YMMWORD [YMMBLOCK(4,0,r10,SIZEOF_ISLOW_MULT_TYPE)]
+    vpmullw     ymm7, ymm7, YMMWORD [YMMBLOCK(6,0,r10,SIZEOF_ISLOW_MULT_TYPE)]
+
+    vperm2i128  ymm0, ymm4, ymm6, 0x20  ; ymm0=in0_4
+    vperm2i128  ymm1, ymm5, ymm4, 0x31  ; ymm1=in3_1
+    vperm2i128  ymm2, ymm5, ymm7, 0x20  ; ymm2=in2_6
+    vperm2i128  ymm3, ymm7, ymm6, 0x31  ; ymm3=in7_5
+
+    dodct ymm0, ymm1, ymm2, ymm3, ymm4, ymm5, ymm6, ymm7, ymm8, ymm9, ymm10, ymm11, 1
+    ; ymm0=data0_1, ymm1=data3_2, ymm2=data4_5, ymm3=data7_6
+
+    dotranspose ymm0, ymm1, ymm2, ymm3, ymm4, ymm5, ymm6, ymm7
+    ; ymm0=data0_4, ymm1=data1_5, ymm2=data2_6, ymm3=data3_7
+
+.column_end:
+
+    ; -- Prefetch the next coefficient block
+
+    prefetchnta [r11 + DCTSIZE2*SIZEOF_JCOEF + 0*32]
+    prefetchnta [r11 + DCTSIZE2*SIZEOF_JCOEF + 1*32]
+    prefetchnta [r11 + DCTSIZE2*SIZEOF_JCOEF + 2*32]
+    prefetchnta [r11 + DCTSIZE2*SIZEOF_JCOEF + 3*32]
+
+    ; ---- Pass 2: process rows.
+
+    vperm2i128  ymm4, ymm3, ymm1, 0x31  ; ymm3=in7_5
+    vperm2i128  ymm1, ymm3, ymm1, 0x20  ; ymm1=in3_1
+
+    dodct ymm0, ymm1, ymm2, ymm4, ymm3, ymm5, ymm6, ymm7, ymm8, ymm9, ymm10, ymm11, 2
+    ; ymm0=data0_1, ymm1=data3_2, ymm2=data4_5, ymm4=data7_6
+
+    dotranspose ymm0, ymm1, ymm2, ymm4, ymm3, ymm5, ymm6, ymm7
+    ; ymm0=data0_4, ymm1=data1_5, ymm2=data2_6, ymm4=data3_7
+
+    vpacksswb   ymm0, ymm0, ymm1        ; ymm0=data01_45
+    vpacksswb   ymm1, ymm2, ymm4        ; ymm1=data23_67
+    vpaddb      ymm0, ymm0, [rel PB_CENTERJSAMP]
+    vpaddb      ymm1, ymm1, [rel PB_CENTERJSAMP]
+
+    vextracti128 xmm6, ymm1, 1          ; xmm3=data67
+    vextracti128 xmm4, ymm0, 1          ; xmm2=data45
+    vextracti128 xmm2, ymm1, 0          ; xmm1=data23
+    vextracti128 xmm0, ymm0, 0          ; xmm0=data01
+
+    vpshufd     xmm1, xmm0, 0x4E  ; xmm1=(10 11 12 13 14 15 16 17 00 01 02 03 04 05 06 07)
+    vpshufd     xmm3, xmm2, 0x4E  ; xmm3=(30 31 32 33 34 35 36 37 20 21 22 23 24 25 26 27)
+    vpshufd     xmm5, xmm4, 0x4E  ; xmm5=(50 51 52 53 54 55 56 57 40 41 42 43 44 45 46 47)
+    vpshufd     xmm7, xmm6, 0x4E  ; xmm7=(70 71 72 73 74 75 76 77 60 61 62 63 64 65 66 67)
+
+    vzeroupper
+
+    mov         eax, r13d
+
+    mov         rdxp, JSAMPROW [r12+0*SIZEOF_JSAMPROW]  ; (JSAMPLE *)
+    mov         rsip, JSAMPROW [r12+1*SIZEOF_JSAMPROW]  ; (JSAMPLE *)
+    movq        XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm0
+    movq        XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm1
+
+    mov         rdxp, JSAMPROW [r12+2*SIZEOF_JSAMPROW]  ; (JSAMPLE *)
+    mov         rsip, JSAMPROW [r12+3*SIZEOF_JSAMPROW]  ; (JSAMPLE *)
+    movq        XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm2
+    movq        XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm3
+
+    mov         rdxp, JSAMPROW [r12+4*SIZEOF_JSAMPROW]  ; (JSAMPLE *)
+    mov         rsip, JSAMPROW [r12+5*SIZEOF_JSAMPROW]  ; (JSAMPLE *)
+    movq        XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm4
+    movq        XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm5
+
+    mov         rdxp, JSAMPROW [r12+6*SIZEOF_JSAMPROW]  ; (JSAMPLE *)
+    mov         rsip, JSAMPROW [r12+7*SIZEOF_JSAMPROW]  ; (JSAMPLE *)
+    movq        XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm6
+    movq        XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm7
+
+    uncollect_args 4
+    pop_xmm     4
+    pop         rbp
+    ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+    align       32
diff --git a/media/libjpeg/simd/x86_64/jidctint-sse2.asm b/media/libjpeg/simd/x86_64/jidctint-sse2.asm
new file mode 100644
index 0000000000..7aa869bc0b
--- /dev/null
+++ b/media/libjpeg/simd/x86_64/jidctint-sse2.asm
@@ -0,0 +1,847 @@
+;
+; jidctint.asm - accurate integer IDCT (64-bit SSE2)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+; Copyright (C) 2009, 2016, 2020, D. R. Commander.
+; Copyright (C) 2018, Matthias Räncker.
+;
+; Based on the x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; This file contains a slower but more accurate integer implementation of the
+; inverse DCT (Discrete Cosine Transform). The following code is based
+; directly on the IJG's original jidctint.c; see the jidctint.c for
+; more details.
+
+%include "jsimdext.inc"
+%include "jdct.inc"
+
+; --------------------------------------------------------------------------
+
+%define CONST_BITS  13
+%define PASS1_BITS  2
+
+%define DESCALE_P1  (CONST_BITS - PASS1_BITS)
+%define DESCALE_P2  (CONST_BITS + PASS1_BITS + 3)
+
+%if CONST_BITS == 13
+F_0_298 equ  2446  ; FIX(0.298631336)
+F_0_390 equ  3196  ; FIX(0.390180644)
+F_0_541 equ  4433  ; FIX(0.541196100)
+F_0_765 equ  6270  ; FIX(0.765366865)
+F_0_899 equ  7373  ; FIX(0.899976223)
+F_1_175 equ  9633  ; FIX(1.175875602)
+F_1_501 equ 12299  ; FIX(1.501321110)
+F_1_847 equ 15137  ; FIX(1.847759065)
+F_1_961 equ 16069  ; FIX(1.961570560)
+F_2_053 equ 16819  ; FIX(2.053119869)
+F_2_562 equ 20995  ; FIX(2.562915447)
+F_3_072 equ 25172  ; FIX(3.072711026)
+%else
+; NASM cannot do compile-time arithmetic on floating-point constants.
+%define DESCALE(x, n)  (((x) + (1 << ((n) - 1))) >> (n))
+F_0_298 equ DESCALE( 320652955, 30 - CONST_BITS)  ; FIX(0.298631336)
+F_0_390 equ DESCALE( 418953276, 30 - CONST_BITS)  ; FIX(0.390180644)
+F_0_541 equ DESCALE( 581104887, 30 - CONST_BITS)  ; FIX(0.541196100)
+F_0_765 equ DESCALE( 821806413, 30 - CONST_BITS)  ; FIX(0.765366865)
+F_0_899 equ DESCALE( 966342111, 30 - CONST_BITS)  ; FIX(0.899976223)
+F_1_175 equ DESCALE(1262586813, 30 - CONST_BITS)  ; FIX(1.175875602)
+F_1_501 equ DESCALE(1612031267, 30 - CONST_BITS)  ; FIX(1.501321110)
+F_1_847 equ DESCALE(1984016188, 30 - CONST_BITS)  ; FIX(1.847759065)
+F_1_961 equ DESCALE(2106220350, 30 - CONST_BITS)  ; FIX(1.961570560)
+F_2_053 equ DESCALE(2204520673, 30 - CONST_BITS)  ; FIX(2.053119869)
+F_2_562 equ DESCALE(2751909506, 30 - CONST_BITS)  ; FIX(2.562915447)
+F_3_072 equ DESCALE(3299298341, 30 - CONST_BITS)  ; FIX(3.072711026)
+%endif
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_CONST
+
+    alignz      32
+    GLOBAL_DATA(jconst_idct_islow_sse2)
+
+EXTN(jconst_idct_islow_sse2):
+
+PW_F130_F054   times 4  dw  (F_0_541 + F_0_765),  F_0_541
+PW_F054_MF130  times 4  dw  F_0_541, (F_0_541 - F_1_847)
+PW_MF078_F117  times 4  dw  (F_1_175 - F_1_961),  F_1_175
+PW_F117_F078   times 4  dw  F_1_175, (F_1_175 - F_0_390)
+PW_MF060_MF089 times 4  dw  (F_0_298 - F_0_899), -F_0_899
+PW_MF089_F060  times 4  dw -F_0_899, (F_1_501 - F_0_899)
+PW_MF050_MF256 times 4  dw  (F_2_053 - F_2_562), -F_2_562
+PW_MF256_F050  times 4  dw -F_2_562, (F_3_072 - F_2_562)
+PD_DESCALE_P1  times 4  dd  1 << (DESCALE_P1 - 1)
+PD_DESCALE_P2  times 4  dd  1 << (DESCALE_P2 - 1)
+PB_CENTERJSAMP times 16 db  CENTERJSAMPLE
+
+    alignz      32
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_TEXT
+    BITS        64
+;
+; Perform dequantization and inverse DCT on one block of coefficients.
+;
+; GLOBAL(void)
+; jsimd_idct_islow_sse2(void *dct_table, JCOEFPTR coef_block,
+;                       JSAMPARRAY output_buf, JDIMENSION output_col)
+;
+
+; r10 = jpeg_component_info *compptr
+; r11 = JCOEFPTR coef_block
+; r12 = JSAMPARRAY output_buf
+; r13d = JDIMENSION output_col
+
+%define original_rbp  rbp + 0
+%define wk(i)         rbp - (WK_NUM - (i)) * SIZEOF_XMMWORD
+                                        ; xmmword wk[WK_NUM]
+%define WK_NUM        12
+
+    align       32
+    GLOBAL_FUNCTION(jsimd_idct_islow_sse2)
+
+EXTN(jsimd_idct_islow_sse2):
+    push        rbp
+    mov         rax, rsp                     ; rax = original rbp
+    sub         rsp, byte 4
+    and         rsp, byte (-SIZEOF_XMMWORD)  ; align to 128 bits
+    mov         [rsp], rax
+    mov         rbp, rsp                     ; rbp = aligned rbp
+    lea         rsp, [wk(0)]
+    collect_args 4
+
+    ; ---- Pass 1: process columns from input.
+
+    mov         rdx, r10                ; quantptr
+    mov         rsi, r11                ; inptr
+
+%ifndef NO_ZERO_COLUMN_TEST_ISLOW_SSE2
+    mov         eax, dword [DWBLOCK(1,0,rsi,SIZEOF_JCOEF)]
+    or          eax, dword [DWBLOCK(2,0,rsi,SIZEOF_JCOEF)]
+    jnz         near .columnDCT
+
+    movdqa      xmm0, XMMWORD [XMMBLOCK(1,0,rsi,SIZEOF_JCOEF)]
+    movdqa      xmm1, XMMWORD [XMMBLOCK(2,0,rsi,SIZEOF_JCOEF)]
+    por         xmm0, XMMWORD [XMMBLOCK(3,0,rsi,SIZEOF_JCOEF)]
+    por         xmm1, XMMWORD [XMMBLOCK(4,0,rsi,SIZEOF_JCOEF)]
+    por         xmm0, XMMWORD [XMMBLOCK(5,0,rsi,SIZEOF_JCOEF)]
+    por         xmm1, XMMWORD [XMMBLOCK(6,0,rsi,SIZEOF_JCOEF)]
+    por         xmm0, XMMWORD [XMMBLOCK(7,0,rsi,SIZEOF_JCOEF)]
+    por         xmm1, xmm0
+    packsswb    xmm1, xmm1
+    packsswb    xmm1, xmm1
+    movd        eax, xmm1
+    test        rax, rax
+    jnz         short .columnDCT
+
+    ; -- AC terms all zero
+
+    movdqa      xmm5, XMMWORD [XMMBLOCK(0,0,rsi,SIZEOF_JCOEF)]
+    pmullw      xmm5, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
+
+    psllw       xmm5, PASS1_BITS
+
+    movdqa      xmm4, xmm5              ; xmm5=in0=(00 01 02 03 04 05 06 07)
+    punpcklwd   xmm5, xmm5              ; xmm5=(00 00 01 01 02 02 03 03)
+    punpckhwd   xmm4, xmm4              ; xmm4=(04 04 05 05 06 06 07 07)
+
+    pshufd      xmm7, xmm5, 0x00        ; xmm7=col0=(00 00 00 00 00 00 00 00)
+    pshufd      xmm6, xmm5, 0x55        ; xmm6=col1=(01 01 01 01 01 01 01 01)
+    pshufd      xmm1, xmm5, 0xAA        ; xmm1=col2=(02 02 02 02 02 02 02 02)
+    pshufd      xmm5, xmm5, 0xFF        ; xmm5=col3=(03 03 03 03 03 03 03 03)
+    pshufd      xmm0, xmm4, 0x00        ; xmm0=col4=(04 04 04 04 04 04 04 04)
+    pshufd      xmm3, xmm4, 0x55        ; xmm3=col5=(05 05 05 05 05 05 05 05)
+    pshufd      xmm2, xmm4, 0xAA        ; xmm2=col6=(06 06 06 06 06 06 06 06)
+    pshufd      xmm4, xmm4, 0xFF        ; xmm4=col7=(07 07 07 07 07 07 07 07)
+
+    movdqa      XMMWORD [wk(8)], xmm6   ; wk(8)=col1
+    movdqa      XMMWORD [wk(9)], xmm5   ; wk(9)=col3
+    movdqa      XMMWORD [wk(10)], xmm3  ; wk(10)=col5
+    movdqa      XMMWORD [wk(11)], xmm4  ; wk(11)=col7
+    jmp         near .column_end
+%endif
+.columnDCT:
+
+    ; -- Even part
+
+    movdqa      xmm0, XMMWORD [XMMBLOCK(0,0,rsi,SIZEOF_JCOEF)]
+    movdqa      xmm1, XMMWORD [XMMBLOCK(2,0,rsi,SIZEOF_JCOEF)]
+    pmullw      xmm0, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
+    pmullw      xmm1, XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
+    movdqa      xmm2, XMMWORD [XMMBLOCK(4,0,rsi,SIZEOF_JCOEF)]
+    movdqa      xmm3, XMMWORD [XMMBLOCK(6,0,rsi,SIZEOF_JCOEF)]
+    pmullw      xmm2, XMMWORD [XMMBLOCK(4,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
+    pmullw      xmm3, XMMWORD [XMMBLOCK(6,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
+
+    ; (Original)
+    ; z1 = (z2 + z3) * 0.541196100;
+    ; tmp2 = z1 + z3 * -1.847759065;
+    ; tmp3 = z1 + z2 * 0.765366865;
+    ;
+    ; (This implementation)
+    ; tmp2 = z2 * 0.541196100 + z3 * (0.541196100 - 1.847759065);
+    ; tmp3 = z2 * (0.541196100 + 0.765366865) + z3 * 0.541196100;
+
+    movdqa      xmm4, xmm1              ; xmm1=in2=z2
+    movdqa      xmm5, xmm1
+    punpcklwd   xmm4, xmm3              ; xmm3=in6=z3
+    punpckhwd   xmm5, xmm3
+    movdqa      xmm1, xmm4
+    movdqa      xmm3, xmm5
+    pmaddwd     xmm4, [rel PW_F130_F054]   ; xmm4=tmp3L
+    pmaddwd     xmm5, [rel PW_F130_F054]   ; xmm5=tmp3H
+    pmaddwd     xmm1, [rel PW_F054_MF130]  ; xmm1=tmp2L
+    pmaddwd     xmm3, [rel PW_F054_MF130]  ; xmm3=tmp2H
+
+    movdqa      xmm6, xmm0
+    paddw       xmm0, xmm2              ; xmm0=in0+in4
+    psubw       xmm6, xmm2              ; xmm6=in0-in4
+
+    pxor        xmm7, xmm7
+    pxor        xmm2, xmm2
+    punpcklwd   xmm7, xmm0              ; xmm7=tmp0L
+    punpckhwd   xmm2, xmm0              ; xmm2=tmp0H
+    psrad       xmm7, (16-CONST_BITS)   ; psrad xmm7,16 & pslld xmm7,CONST_BITS
+    psrad       xmm2, (16-CONST_BITS)   ; psrad xmm2,16 & pslld xmm2,CONST_BITS
+
+    movdqa      xmm0, xmm7
+    paddd       xmm7, xmm4              ; xmm7=tmp10L
+    psubd       xmm0, xmm4              ; xmm0=tmp13L
+    movdqa      xmm4, xmm2
+    paddd       xmm2, xmm5              ; xmm2=tmp10H
+    psubd       xmm4, xmm5              ; xmm4=tmp13H
+
+    movdqa      XMMWORD [wk(0)], xmm7   ; wk(0)=tmp10L
+    movdqa      XMMWORD [wk(1)], xmm2   ; wk(1)=tmp10H
+    movdqa      XMMWORD [wk(2)], xmm0   ; wk(2)=tmp13L
+    movdqa      XMMWORD [wk(3)], xmm4   ; wk(3)=tmp13H
+
+    pxor        xmm5, xmm5
+    pxor        xmm7, xmm7
+    punpcklwd   xmm5, xmm6              ; xmm5=tmp1L
+    punpckhwd   xmm7, xmm6              ; xmm7=tmp1H
+    psrad       xmm5, (16-CONST_BITS)   ; psrad xmm5,16 & pslld xmm5,CONST_BITS
+    psrad       xmm7, (16-CONST_BITS)   ; psrad xmm7,16 & pslld xmm7,CONST_BITS
+
+    movdqa      xmm2, xmm5
+    paddd       xmm5, xmm1              ; xmm5=tmp11L
+    psubd       xmm2, xmm1              ; xmm2=tmp12L
+    movdqa      xmm0, xmm7
+    paddd       xmm7, xmm3              ; xmm7=tmp11H
+    psubd       xmm0, xmm3              ; xmm0=tmp12H
+
+    movdqa      XMMWORD [wk(4)], xmm5   ; wk(4)=tmp11L
+    movdqa      XMMWORD [wk(5)], xmm7   ; wk(5)=tmp11H
+    movdqa      XMMWORD [wk(6)], xmm2   ; wk(6)=tmp12L
+    movdqa      XMMWORD [wk(7)], xmm0   ; wk(7)=tmp12H
+
+    ; -- Odd part
+
+    movdqa      xmm4, XMMWORD [XMMBLOCK(1,0,rsi,SIZEOF_JCOEF)]
+    movdqa      xmm6, XMMWORD [XMMBLOCK(3,0,rsi,SIZEOF_JCOEF)]
+    pmullw      xmm4, XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
+    pmullw      xmm6, XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
+    movdqa      xmm1, XMMWORD [XMMBLOCK(5,0,rsi,SIZEOF_JCOEF)]
+    movdqa      xmm3, XMMWORD [XMMBLOCK(7,0,rsi,SIZEOF_JCOEF)]
+    pmullw      xmm1, XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
+    pmullw      xmm3, XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
+
+    movdqa      xmm5, xmm6
+    movdqa      xmm7, xmm4
+    paddw       xmm5, xmm3              ; xmm5=z3
+    paddw       xmm7, xmm1              ; xmm7=z4
+
+    ; (Original)
+    ; z5 = (z3 + z4) * 1.175875602;
+    ; z3 = z3 * -1.961570560;  z4 = z4 * -0.390180644;
+    ; z3 += z5;  z4 += z5;
+    ;
+    ; (This implementation)
+    ; z3 = z3 * (1.175875602 - 1.961570560) + z4 * 1.175875602;
+    ; z4 = z3 * 1.175875602 + z4 * (1.175875602 - 0.390180644);
+
+    movdqa      xmm2, xmm5
+    movdqa      xmm0, xmm5
+    punpcklwd   xmm2, xmm7
+    punpckhwd   xmm0, xmm7
+    movdqa      xmm5, xmm2
+    movdqa      xmm7, xmm0
+    pmaddwd     xmm2, [rel PW_MF078_F117]  ; xmm2=z3L
+    pmaddwd     xmm0, [rel PW_MF078_F117]  ; xmm0=z3H
+    pmaddwd     xmm5, [rel PW_F117_F078]   ; xmm5=z4L
+    pmaddwd     xmm7, [rel PW_F117_F078]   ; xmm7=z4H
+
+    movdqa      XMMWORD [wk(10)], xmm2  ; wk(10)=z3L
+    movdqa      XMMWORD [wk(11)], xmm0  ; wk(11)=z3H
+
+    ; (Original)
+    ; z1 = tmp0 + tmp3;  z2 = tmp1 + tmp2;
+    ; tmp0 = tmp0 * 0.298631336;  tmp1 = tmp1 * 2.053119869;
+    ; tmp2 = tmp2 * 3.072711026;  tmp3 = tmp3 * 1.501321110;
+    ; z1 = z1 * -0.899976223;  z2 = z2 * -2.562915447;
+    ; tmp0 += z1 + z3;  tmp1 += z2 + z4;
+    ; tmp2 += z2 + z3;  tmp3 += z1 + z4;
+    ;
+    ; (This implementation)
+    ; tmp0 = tmp0 * (0.298631336 - 0.899976223) + tmp3 * -0.899976223;
+    ; tmp1 = tmp1 * (2.053119869 - 2.562915447) + tmp2 * -2.562915447;
+    ; tmp2 = tmp1 * -2.562915447 + tmp2 * (3.072711026 - 2.562915447);
+    ; tmp3 = tmp0 * -0.899976223 + tmp3 * (1.501321110 - 0.899976223);
+    ; tmp0 += z3;  tmp1 += z4;
+    ; tmp2 += z3;  tmp3 += z4;
+
+    movdqa      xmm2, xmm3
+    movdqa      xmm0, xmm3
+    punpcklwd   xmm2, xmm4
+    punpckhwd   xmm0, xmm4
+    movdqa      xmm3, xmm2
+    movdqa      xmm4, xmm0
+    pmaddwd     xmm2, [rel PW_MF060_MF089]  ; xmm2=tmp0L
+    pmaddwd     xmm0, [rel PW_MF060_MF089]  ; xmm0=tmp0H
+    pmaddwd     xmm3, [rel PW_MF089_F060]   ; xmm3=tmp3L
+    pmaddwd     xmm4, [rel PW_MF089_F060]   ; xmm4=tmp3H
+
+    paddd       xmm2, XMMWORD [wk(10)]  ; xmm2=tmp0L
+    paddd       xmm0, XMMWORD [wk(11)]  ; xmm0=tmp0H
+    paddd       xmm3, xmm5              ; xmm3=tmp3L
+    paddd       xmm4, xmm7              ; xmm4=tmp3H
+
+    movdqa      XMMWORD [wk(8)], xmm2   ; wk(8)=tmp0L
+    movdqa      XMMWORD [wk(9)], xmm0   ; wk(9)=tmp0H
+
+    movdqa      xmm2, xmm1
+    movdqa      xmm0, xmm1
+    punpcklwd   xmm2, xmm6
+    punpckhwd   xmm0, xmm6
+    movdqa      xmm1, xmm2
+    movdqa      xmm6, xmm0
+    pmaddwd     xmm2, [rel PW_MF050_MF256]  ; xmm2=tmp1L
+    pmaddwd     xmm0, [rel PW_MF050_MF256]  ; xmm0=tmp1H
+    pmaddwd     xmm1, [rel PW_MF256_F050]   ; xmm1=tmp2L
+    pmaddwd     xmm6, [rel PW_MF256_F050]   ; xmm6=tmp2H
+
+    paddd       xmm2, xmm5              ; xmm2=tmp1L
+    paddd       xmm0, xmm7              ; xmm0=tmp1H
+    paddd       xmm1, XMMWORD [wk(10)]  ; xmm1=tmp2L
+    paddd       xmm6, XMMWORD [wk(11)]  ; xmm6=tmp2H
+
+    movdqa      XMMWORD [wk(10)], xmm2  ; wk(10)=tmp1L
+    movdqa      XMMWORD [wk(11)], xmm0  ; wk(11)=tmp1H
+
+    ; -- Final output stage
+
+    movdqa      xmm5, XMMWORD [wk(0)]   ; xmm5=tmp10L
+    movdqa      xmm7, XMMWORD [wk(1)]   ; xmm7=tmp10H
+
+    movdqa      xmm2, xmm5
+    movdqa      xmm0, xmm7
+    paddd       xmm5, xmm3              ; xmm5=data0L
+    paddd       xmm7, xmm4              ; xmm7=data0H
+    psubd       xmm2, xmm3              ; xmm2=data7L
+    psubd       xmm0, xmm4              ; xmm0=data7H
+
+    movdqa      xmm3, [rel PD_DESCALE_P1]  ; xmm3=[rel PD_DESCALE_P1]
+
+    paddd       xmm5, xmm3
+    paddd       xmm7, xmm3
+    psrad       xmm5, DESCALE_P1
+    psrad       xmm7, DESCALE_P1
+    paddd       xmm2, xmm3
+    paddd       xmm0, xmm3
+    psrad       xmm2, DESCALE_P1
+    psrad       xmm0, DESCALE_P1
+
+    packssdw    xmm5, xmm7              ; xmm5=data0=(00 01 02 03 04 05 06 07)
+    packssdw    xmm2, xmm0              ; xmm2=data7=(70 71 72 73 74 75 76 77)
+
+    movdqa      xmm4, XMMWORD [wk(4)]   ; xmm4=tmp11L
+    movdqa      xmm3, XMMWORD [wk(5)]   ; xmm3=tmp11H
+
+    movdqa      xmm7, xmm4
+    movdqa      xmm0, xmm3
+    paddd       xmm4, xmm1              ; xmm4=data1L
+    paddd       xmm3, xmm6              ; xmm3=data1H
+    psubd       xmm7, xmm1              ; xmm7=data6L
+    psubd       xmm0, xmm6              ; xmm0=data6H
+
+    movdqa      xmm1, [rel PD_DESCALE_P1]  ; xmm1=[rel PD_DESCALE_P1]
+
+    paddd       xmm4, xmm1
+    paddd       xmm3, xmm1
+    psrad       xmm4, DESCALE_P1
+    psrad       xmm3, DESCALE_P1
+    paddd       xmm7, xmm1
+    paddd       xmm0, xmm1
+    psrad       xmm7, DESCALE_P1
+    psrad       xmm0, DESCALE_P1
+
+    packssdw    xmm4, xmm3              ; xmm4=data1=(10 11 12 13 14 15 16 17)
+    packssdw    xmm7, xmm0              ; xmm7=data6=(60 61 62 63 64 65 66 67)
+
+    movdqa      xmm6, xmm5              ; transpose coefficients(phase 1)
+    punpcklwd   xmm5, xmm4              ; xmm5=(00 10 01 11 02 12 03 13)
+    punpckhwd   xmm6, xmm4              ; xmm6=(04 14 05 15 06 16 07 17)
+    movdqa      xmm1, xmm7              ; transpose coefficients(phase 1)
+    punpcklwd   xmm7, xmm2              ; xmm7=(60 70 61 71 62 72 63 73)
+    punpckhwd   xmm1, xmm2              ; xmm1=(64 74 65 75 66 76 67 77)
+
+    movdqa      xmm3, XMMWORD [wk(6)]   ; xmm3=tmp12L
+    movdqa      xmm0, XMMWORD [wk(7)]   ; xmm0=tmp12H
+    movdqa      xmm4, XMMWORD [wk(10)]  ; xmm4=tmp1L
+    movdqa      xmm2, XMMWORD [wk(11)]  ; xmm2=tmp1H
+
+    movdqa      XMMWORD [wk(0)], xmm5   ; wk(0)=(00 10 01 11 02 12 03 13)
+    movdqa      XMMWORD [wk(1)], xmm6   ; wk(1)=(04 14 05 15 06 16 07 17)
+    movdqa      XMMWORD [wk(4)], xmm7   ; wk(4)=(60 70 61 71 62 72 63 73)
+    movdqa      XMMWORD [wk(5)], xmm1   ; wk(5)=(64 74 65 75 66 76 67 77)
+
+    movdqa      xmm5, xmm3
+    movdqa      xmm6, xmm0
+    paddd       xmm3, xmm4              ; xmm3=data2L
+    paddd       xmm0, xmm2              ; xmm0=data2H
+    psubd       xmm5, xmm4              ; xmm5=data5L
+    psubd       xmm6, xmm2              ; xmm6=data5H
+
+    movdqa      xmm7, [rel PD_DESCALE_P1]  ; xmm7=[rel PD_DESCALE_P1]
+
+    paddd       xmm3, xmm7
+    paddd       xmm0, xmm7
+    psrad       xmm3, DESCALE_P1
+    psrad       xmm0, DESCALE_P1
+    paddd       xmm5, xmm7
+    paddd       xmm6, xmm7
+    psrad       xmm5, DESCALE_P1
+    psrad       xmm6, DESCALE_P1
+
+    packssdw    xmm3, xmm0              ; xmm3=data2=(20 21 22 23 24 25 26 27)
+    packssdw    xmm5, xmm6              ; xmm5=data5=(50 51 52 53 54 55 56 57)
+
+    movdqa      xmm1, XMMWORD [wk(2)]   ; xmm1=tmp13L
+    movdqa      xmm4, XMMWORD [wk(3)]   ; xmm4=tmp13H
+    movdqa      xmm2, XMMWORD [wk(8)]   ; xmm2=tmp0L
+    movdqa      xmm7, XMMWORD [wk(9)]   ; xmm7=tmp0H
+
+    movdqa      xmm0, xmm1
+    movdqa      xmm6, xmm4
+    paddd       xmm1, xmm2              ; xmm1=data3L
+    paddd       xmm4, xmm7              ; xmm4=data3H
+    psubd       xmm0, xmm2              ; xmm0=data4L
+    psubd       xmm6, xmm7              ; xmm6=data4H
+
+    movdqa      xmm2, [rel PD_DESCALE_P1]  ; xmm2=[rel PD_DESCALE_P1]
+
+    paddd       xmm1, xmm2
+    paddd       xmm4, xmm2
+    psrad       xmm1, DESCALE_P1
+    psrad       xmm4, DESCALE_P1
+    paddd       xmm0, xmm2
+    paddd       xmm6, xmm2
+    psrad       xmm0, DESCALE_P1
+    psrad       xmm6, DESCALE_P1
+
+    packssdw    xmm1, xmm4              ; xmm1=data3=(30 31 32 33 34 35 36 37)
+    packssdw    xmm0, xmm6              ; xmm0=data4=(40 41 42 43 44 45 46 47)
+
+    movdqa      xmm7, XMMWORD [wk(0)]   ; xmm7=(00 10 01 11 02 12 03 13)
+    movdqa      xmm2, XMMWORD [wk(1)]   ; xmm2=(04 14 05 15 06 16 07 17)
+
+    movdqa      xmm4, xmm3              ; transpose coefficients(phase 1)
+    punpcklwd   xmm3, xmm1              ; xmm3=(20 30 21 31 22 32 23 33)
+    punpckhwd   xmm4, xmm1              ; xmm4=(24 34 25 35 26 36 27 37)
+    movdqa      xmm6, xmm0              ; transpose coefficients(phase 1)
+    punpcklwd   xmm0, xmm5              ; xmm0=(40 50 41 51 42 52 43 53)
+    punpckhwd   xmm6, xmm5              ; xmm6=(44 54 45 55 46 56 47 57)
+
+    movdqa      xmm1, xmm7              ; transpose coefficients(phase 2)
+    punpckldq   xmm7, xmm3              ; xmm7=(00 10 20 30 01 11 21 31)
+    punpckhdq   xmm1, xmm3              ; xmm1=(02 12 22 32 03 13 23 33)
+    movdqa      xmm5, xmm2              ; transpose coefficients(phase 2)
+    punpckldq   xmm2, xmm4              ; xmm2=(04 14 24 34 05 15 25 35)
+    punpckhdq   xmm5, xmm4              ; xmm5=(06 16 26 36 07 17 27 37)
+
+    movdqa      xmm3, XMMWORD [wk(4)]   ; xmm3=(60 70 61 71 62 72 63 73)
+    movdqa      xmm4, XMMWORD [wk(5)]   ; xmm4=(64 74 65 75 66 76 67 77)
+
+    movdqa      XMMWORD [wk(6)], xmm2   ; wk(6)=(04 14 24 34 05 15 25 35)
+    movdqa      XMMWORD [wk(7)], xmm5   ; wk(7)=(06 16 26 36 07 17 27 37)
+
+    movdqa      xmm2, xmm0              ; transpose coefficients(phase 2)
+    punpckldq   xmm0, xmm3              ; xmm0=(40 50 60 70 41 51 61 71)
+    punpckhdq   xmm2, xmm3              ; xmm2=(42 52 62 72 43 53 63 73)
+    movdqa      xmm5, xmm6              ; transpose coefficients(phase 2)
+    punpckldq   xmm6, xmm4              ; xmm6=(44 54 64 74 45 55 65 75)
+    punpckhdq   xmm5, xmm4              ; xmm5=(46 56 66 76 47 57 67 77)
+
+    movdqa      xmm3, xmm7              ; transpose coefficients(phase 3)
+    punpcklqdq  xmm7, xmm0              ; xmm7=col0=(00 10 20 30 40 50 60 70)
+    punpckhqdq  xmm3, xmm0              ; xmm3=col1=(01 11 21 31 41 51 61 71)
+    movdqa      xmm4, xmm1              ; transpose coefficients(phase 3)
+    punpcklqdq  xmm1, xmm2              ; xmm1=col2=(02 12 22 32 42 52 62 72)
+    punpckhqdq  xmm4, xmm2              ; xmm4=col3=(03 13 23 33 43 53 63 73)
+
+    movdqa      xmm0, XMMWORD [wk(6)]   ; xmm0=(04 14 24 34 05 15 25 35)
+    movdqa      xmm2, XMMWORD [wk(7)]   ; xmm2=(06 16 26 36 07 17 27 37)
+
+    movdqa      XMMWORD [wk(8)], xmm3   ; wk(8)=col1
+    movdqa      XMMWORD [wk(9)], xmm4   ; wk(9)=col3
+
+    movdqa      xmm3, xmm0              ; transpose coefficients(phase 3)
+    punpcklqdq  xmm0, xmm6              ; xmm0=col4=(04 14 24 34 44 54 64 74)
+    punpckhqdq  xmm3, xmm6              ; xmm3=col5=(05 15 25 35 45 55 65 75)
+    movdqa      xmm4, xmm2              ; transpose coefficients(phase 3)
+    punpcklqdq  xmm2, xmm5              ; xmm2=col6=(06 16 26 36 46 56 66 76)
+    punpckhqdq  xmm4, xmm5              ; xmm4=col7=(07 17 27 37 47 57 67 77)
+
+    movdqa      XMMWORD [wk(10)], xmm3  ; wk(10)=col5
+    movdqa      XMMWORD [wk(11)], xmm4  ; wk(11)=col7
+.column_end:
+
+    ; -- Prefetch the next coefficient block
+
+    prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 0*32]
+    prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 1*32]
+    prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 2*32]
+    prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 3*32]
+
+    ; ---- Pass 2: process rows from work array, store into output array.
+
+    mov         rax, [original_rbp]
+    mov         rdi, r12                ; (JSAMPROW *)
+    mov         eax, r13d
+
+    ; -- Even part
+
+    ; xmm7=col0, xmm1=col2, xmm0=col4, xmm2=col6
+
+    ; (Original)
+    ; z1 = (z2 + z3) * 0.541196100;
+    ; tmp2 = z1 + z3 * -1.847759065;
+    ; tmp3 = z1 + z2 * 0.765366865;
+    ;
+    ; (This implementation)
+    ; tmp2 = z2 * 0.541196100 + z3 * (0.541196100 - 1.847759065);
+    ; tmp3 = z2 * (0.541196100 + 0.765366865) + z3 * 0.541196100;
+
+    movdqa      xmm6, xmm1              ; xmm1=in2=z2
+    movdqa      xmm5, xmm1
+    punpcklwd   xmm6, xmm2              ; xmm2=in6=z3
+    punpckhwd   xmm5, xmm2
+    movdqa      xmm1, xmm6
+    movdqa      xmm2, xmm5
+    pmaddwd     xmm6, [rel PW_F130_F054]   ; xmm6=tmp3L
+    pmaddwd     xmm5, [rel PW_F130_F054]   ; xmm5=tmp3H
+    pmaddwd     xmm1, [rel PW_F054_MF130]  ; xmm1=tmp2L
+    pmaddwd     xmm2, [rel PW_F054_MF130]  ; xmm2=tmp2H
+
+    movdqa      xmm3, xmm7
+    paddw       xmm7, xmm0              ; xmm7=in0+in4
+    psubw       xmm3, xmm0              ; xmm3=in0-in4
+
+    pxor        xmm4, xmm4
+    pxor        xmm0, xmm0
+    punpcklwd   xmm4, xmm7              ; xmm4=tmp0L
+    punpckhwd   xmm0, xmm7              ; xmm0=tmp0H
+    psrad       xmm4, (16-CONST_BITS)   ; psrad xmm4,16 & pslld xmm4,CONST_BITS
+    psrad       xmm0, (16-CONST_BITS)   ; psrad xmm0,16 & pslld xmm0,CONST_BITS
+
+    movdqa      xmm7, xmm4
+    paddd       xmm4, xmm6              ; xmm4=tmp10L
+    psubd       xmm7, xmm6              ; xmm7=tmp13L
+    movdqa      xmm6, xmm0
+    paddd       xmm0, xmm5              ; xmm0=tmp10H
+    psubd       xmm6, xmm5              ; xmm6=tmp13H
+
+    movdqa      XMMWORD [wk(0)], xmm4   ; wk(0)=tmp10L
+    movdqa      XMMWORD [wk(1)], xmm0   ; wk(1)=tmp10H
+    movdqa      XMMWORD [wk(2)], xmm7   ; wk(2)=tmp13L
+    movdqa      XMMWORD [wk(3)], xmm6   ; wk(3)=tmp13H
+
+    pxor        xmm5, xmm5
+    pxor        xmm4, xmm4
+    punpcklwd   xmm5, xmm3              ; xmm5=tmp1L
+    punpckhwd   xmm4, xmm3              ; xmm4=tmp1H
+    psrad       xmm5, (16-CONST_BITS)   ; psrad xmm5,16 & pslld xmm5,CONST_BITS
+    psrad       xmm4, (16-CONST_BITS)   ; psrad xmm4,16 & pslld xmm4,CONST_BITS
+
+    movdqa      xmm0, xmm5
+    paddd       xmm5, xmm1              ; xmm5=tmp11L
+    psubd       xmm0, xmm1              ; xmm0=tmp12L
+    movdqa      xmm7, xmm4
+    paddd       xmm4, xmm2              ; xmm4=tmp11H
+    psubd       xmm7, xmm2              ; xmm7=tmp12H
+
+    movdqa      XMMWORD [wk(4)], xmm5   ; wk(4)=tmp11L
+    movdqa      XMMWORD [wk(5)], xmm4   ; wk(5)=tmp11H
+    movdqa      XMMWORD [wk(6)], xmm0   ; wk(6)=tmp12L
+    movdqa      XMMWORD [wk(7)], xmm7   ; wk(7)=tmp12H
+
+    ; -- Odd part
+
+    movdqa      xmm6, XMMWORD [wk(9)]   ; xmm6=col3
+    movdqa      xmm3, XMMWORD [wk(8)]   ; xmm3=col1
+    movdqa      xmm1, XMMWORD [wk(11)]  ; xmm1=col7
+    movdqa      xmm2, XMMWORD [wk(10)]  ; xmm2=col5
+
+    movdqa      xmm5, xmm6
+    movdqa      xmm4, xmm3
+    paddw       xmm5, xmm1              ; xmm5=z3
+    paddw       xmm4, xmm2              ; xmm4=z4
+
+    ; (Original)
+    ; z5 = (z3 + z4) * 1.175875602;
+    ; z3 = z3 * -1.961570560;  z4 = z4 * -0.390180644;
+    ; z3 += z5;  z4 += z5;
+    ;
+    ; (This implementation)
+    ; z3 = z3 * (1.175875602 - 1.961570560) + z4 * 1.175875602;
+    ; z4 = z3 * 1.175875602 + z4 * (1.175875602 - 0.390180644);
+
+    movdqa      xmm0, xmm5
+    movdqa      xmm7, xmm5
+    punpcklwd   xmm0, xmm4
+    punpckhwd   xmm7, xmm4
+    movdqa      xmm5, xmm0
+    movdqa      xmm4, xmm7
+    pmaddwd     xmm0, [rel PW_MF078_F117]  ; xmm0=z3L
+    pmaddwd     xmm7, [rel PW_MF078_F117]  ; xmm7=z3H
+    pmaddwd     xmm5, [rel PW_F117_F078]   ; xmm5=z4L
+    pmaddwd     xmm4, [rel PW_F117_F078]   ; xmm4=z4H
+
+    movdqa      XMMWORD [wk(10)], xmm0  ; wk(10)=z3L
+    movdqa      XMMWORD [wk(11)], xmm7  ; wk(11)=z3H
+
+    ; (Original)
+    ; z1 = tmp0 + tmp3;  z2 = tmp1 + tmp2;
+    ; tmp0 = tmp0 * 0.298631336;  tmp1 = tmp1 * 2.053119869;
+    ; tmp2 = tmp2 * 3.072711026;  tmp3 = tmp3 * 1.501321110;
+    ; z1 = z1 * -0.899976223;  z2 = z2 * -2.562915447;
+    ; tmp0 += z1 + z3;  tmp1 += z2 + z4;
+    ; tmp2 += z2 + z3;  tmp3 += z1 + z4;
+    ;
+    ; (This implementation)
+    ; tmp0 = tmp0 * (0.298631336 - 0.899976223) + tmp3 * -0.899976223;
+    ; tmp1 = tmp1 * (2.053119869 - 2.562915447) + tmp2 * -2.562915447;
+    ; tmp2 = tmp1 * -2.562915447 + tmp2 * (3.072711026 - 2.562915447);
+    ; tmp3 = tmp0 * -0.899976223 + tmp3 * (1.501321110 - 0.899976223);
+    ; tmp0 += z3;  tmp1 += z4;
+    ; tmp2 += z3;  tmp3 += z4;
+
+    movdqa      xmm0, xmm1
+    movdqa      xmm7, xmm1
+    punpcklwd   xmm0, xmm3
+    punpckhwd   xmm7, xmm3
+    movdqa      xmm1, xmm0
+    movdqa      xmm3, xmm7
+    pmaddwd     xmm0, [rel PW_MF060_MF089]  ; xmm0=tmp0L
+    pmaddwd     xmm7, [rel PW_MF060_MF089]  ; xmm7=tmp0H
+    pmaddwd     xmm1, [rel PW_MF089_F060]   ; xmm1=tmp3L
+    pmaddwd     xmm3, [rel PW_MF089_F060]   ; xmm3=tmp3H
+
+    paddd       xmm0, XMMWORD [wk(10)]  ; xmm0=tmp0L
+    paddd       xmm7, XMMWORD [wk(11)]  ; xmm7=tmp0H
+    paddd       xmm1, xmm5              ; xmm1=tmp3L
+    paddd       xmm3, xmm4              ; xmm3=tmp3H
+
+    movdqa      XMMWORD [wk(8)], xmm0   ; wk(8)=tmp0L
+    movdqa      XMMWORD [wk(9)], xmm7   ; wk(9)=tmp0H
+
+    movdqa      xmm0, xmm2
+    movdqa      xmm7, xmm2
+    punpcklwd   xmm0, xmm6
+    punpckhwd   xmm7, xmm6
+    movdqa      xmm2, xmm0
+    movdqa      xmm6, xmm7
+    pmaddwd     xmm0, [rel PW_MF050_MF256]  ; xmm0=tmp1L
+    pmaddwd     xmm7, [rel PW_MF050_MF256]  ; xmm7=tmp1H
+    pmaddwd     xmm2, [rel PW_MF256_F050]   ; xmm2=tmp2L
+    pmaddwd     xmm6, [rel PW_MF256_F050]   ; xmm6=tmp2H
+
+    paddd       xmm0, xmm5              ; xmm0=tmp1L
+    paddd       xmm7, xmm4              ; xmm7=tmp1H
+    paddd       xmm2, XMMWORD [wk(10)]  ; xmm2=tmp2L
+    paddd       xmm6, XMMWORD [wk(11)]  ; xmm6=tmp2H
+
+    movdqa      XMMWORD [wk(10)], xmm0  ; wk(10)=tmp1L
+    movdqa      XMMWORD [wk(11)], xmm7  ; wk(11)=tmp1H
+
+    ; -- Final output stage
+
+    movdqa      xmm5, XMMWORD [wk(0)]   ; xmm5=tmp10L
+    movdqa      xmm4, XMMWORD [wk(1)]   ; xmm4=tmp10H
+
+    movdqa      xmm0, xmm5
+    movdqa      xmm7, xmm4
+    paddd       xmm5, xmm1              ; xmm5=data0L
+    paddd       xmm4, xmm3              ; xmm4=data0H
+    psubd       xmm0, xmm1              ; xmm0=data7L
+    psubd       xmm7, xmm3              ; xmm7=data7H
+
+    movdqa      xmm1, [rel PD_DESCALE_P2]  ; xmm1=[rel PD_DESCALE_P2]
+
+    paddd       xmm5, xmm1
+    paddd       xmm4, xmm1
+    psrad       xmm5, DESCALE_P2
+    psrad       xmm4, DESCALE_P2
+    paddd       xmm0, xmm1
+    paddd       xmm7, xmm1
+    psrad       xmm0, DESCALE_P2
+    psrad       xmm7, DESCALE_P2
+
+    packssdw    xmm5, xmm4              ; xmm5=data0=(00 10 20 30 40 50 60 70)
+    packssdw    xmm0, xmm7              ; xmm0=data7=(07 17 27 37 47 57 67 77)
+
+    movdqa      xmm3, XMMWORD [wk(4)]   ; xmm3=tmp11L
+    movdqa      xmm1, XMMWORD [wk(5)]   ; xmm1=tmp11H
+
+    movdqa      xmm4, xmm3
+    movdqa      xmm7, xmm1
+    paddd       xmm3, xmm2              ; xmm3=data1L
+    paddd       xmm1, xmm6              ; xmm1=data1H
+    psubd       xmm4, xmm2              ; xmm4=data6L
+    psubd       xmm7, xmm6              ; xmm7=data6H
+
+    movdqa      xmm2, [rel PD_DESCALE_P2]  ; xmm2=[rel PD_DESCALE_P2]
+
+    paddd       xmm3, xmm2
+    paddd       xmm1, xmm2
+    psrad       xmm3, DESCALE_P2
+    psrad       xmm1, DESCALE_P2
+    paddd       xmm4, xmm2
+    paddd       xmm7, xmm2
+    psrad       xmm4, DESCALE_P2
+    psrad       xmm7, DESCALE_P2
+
+    packssdw    xmm3, xmm1              ; xmm3=data1=(01 11 21 31 41 51 61 71)
+    packssdw    xmm4, xmm7              ; xmm4=data6=(06 16 26 36 46 56 66 76)
+
+    packsswb    xmm5, xmm4              ; xmm5=(00 10 20 30 40 50 60 70 06 16 26 36 46 56 66 76)
+    packsswb    xmm3, xmm0              ; xmm3=(01 11 21 31 41 51 61 71 07 17 27 37 47 57 67 77)
+
+    movdqa      xmm6, XMMWORD [wk(6)]   ; xmm6=tmp12L
+    movdqa      xmm2, XMMWORD [wk(7)]   ; xmm2=tmp12H
+    movdqa      xmm1, XMMWORD [wk(10)]  ; xmm1=tmp1L
+    movdqa      xmm7, XMMWORD [wk(11)]  ; xmm7=tmp1H
+
+    movdqa      XMMWORD [wk(0)], xmm5   ; wk(0)=(00 10 20 30 40 50 60 70 06 16 26 36 46 56 66 76)
+    movdqa      XMMWORD [wk(1)], xmm3   ; wk(1)=(01 11 21 31 41 51 61 71 07 17 27 37 47 57 67 77)
+
+    movdqa      xmm4, xmm6
+    movdqa      xmm0, xmm2
+    paddd       xmm6, xmm1              ; xmm6=data2L
+    paddd       xmm2, xmm7              ; xmm2=data2H
+    psubd       xmm4, xmm1              ; xmm4=data5L
+    psubd       xmm0, xmm7              ; xmm0=data5H
+
+    movdqa      xmm5, [rel PD_DESCALE_P2]  ; xmm5=[rel PD_DESCALE_P2]
+
+    paddd       xmm6, xmm5
+    paddd       xmm2, xmm5
+    psrad       xmm6, DESCALE_P2
+    psrad       xmm2, DESCALE_P2
+    paddd       xmm4, xmm5
+    paddd       xmm0, xmm5
+    psrad       xmm4, DESCALE_P2
+    psrad       xmm0, DESCALE_P2
+
+    packssdw    xmm6, xmm2              ; xmm6=data2=(02 12 22 32 42 52 62 72)
+    packssdw    xmm4, xmm0              ; xmm4=data5=(05 15 25 35 45 55 65 75)
+
+    movdqa      xmm3, XMMWORD [wk(2)]   ; xmm3=tmp13L
+    movdqa      xmm1, XMMWORD [wk(3)]   ; xmm1=tmp13H
+    movdqa      xmm7, XMMWORD [wk(8)]   ; xmm7=tmp0L
+    movdqa      xmm5, XMMWORD [wk(9)]   ; xmm5=tmp0H
+
+    movdqa      xmm2, xmm3
+    movdqa      xmm0, xmm1
+    paddd       xmm3, xmm7              ; xmm3=data3L
+    paddd       xmm1, xmm5              ; xmm1=data3H
+    psubd       xmm2, xmm7              ; xmm2=data4L
+    psubd       xmm0, xmm5              ; xmm0=data4H
+
+    movdqa      xmm7, [rel PD_DESCALE_P2]  ; xmm7=[rel PD_DESCALE_P2]
+
+    paddd       xmm3, xmm7
+    paddd       xmm1, xmm7
+    psrad       xmm3, DESCALE_P2
+    psrad       xmm1, DESCALE_P2
+    paddd       xmm2, xmm7
+    paddd       xmm0, xmm7
+    psrad       xmm2, DESCALE_P2
+    psrad       xmm0, DESCALE_P2
+
+    movdqa      xmm5, [rel PB_CENTERJSAMP]  ; xmm5=[rel PB_CENTERJSAMP]
+
+    packssdw    xmm3, xmm1             ; xmm3=data3=(03 13 23 33 43 53 63 73)
+    packssdw    xmm2, xmm0             ; xmm2=data4=(04 14 24 34 44 54 64 74)
+
+    movdqa      xmm7, XMMWORD [wk(0)]  ; xmm7=(00 10 20 30 40 50 60 70 06 16 26 36 46 56 66 76)
+    movdqa      xmm1, XMMWORD [wk(1)]  ; xmm1=(01 11 21 31 41 51 61 71 07 17 27 37 47 57 67 77)
+
+    packsswb    xmm6, xmm2             ; xmm6=(02 12 22 32 42 52 62 72 04 14 24 34 44 54 64 74)
+    packsswb    xmm3, xmm4             ; xmm3=(03 13 23 33 43 53 63 73 05 15 25 35 45 55 65 75)
+
+    paddb       xmm7, xmm5
+    paddb       xmm1, xmm5
+    paddb       xmm6, xmm5
+    paddb       xmm3, xmm5
+
+    movdqa      xmm0, xmm7        ; transpose coefficients(phase 1)
+    punpcklbw   xmm7, xmm1        ; xmm7=(00 01 10 11 20 21 30 31 40 41 50 51 60 61 70 71)
+    punpckhbw   xmm0, xmm1        ; xmm0=(06 07 16 17 26 27 36 37 46 47 56 57 66 67 76 77)
+    movdqa      xmm2, xmm6        ; transpose coefficients(phase 1)
+    punpcklbw   xmm6, xmm3        ; xmm6=(02 03 12 13 22 23 32 33 42 43 52 53 62 63 72 73)
+    punpckhbw   xmm2, xmm3        ; xmm2=(04 05 14 15 24 25 34 35 44 45 54 55 64 65 74 75)
+
+    movdqa      xmm4, xmm7        ; transpose coefficients(phase 2)
+    punpcklwd   xmm7, xmm6        ; xmm7=(00 01 02 03 10 11 12 13 20 21 22 23 30 31 32 33)
+    punpckhwd   xmm4, xmm6        ; xmm4=(40 41 42 43 50 51 52 53 60 61 62 63 70 71 72 73)
+    movdqa      xmm5, xmm2        ; transpose coefficients(phase 2)
+    punpcklwd   xmm2, xmm0        ; xmm2=(04 05 06 07 14 15 16 17 24 25 26 27 34 35 36 37)
+    punpckhwd   xmm5, xmm0        ; xmm5=(44 45 46 47 54 55 56 57 64 65 66 67 74 75 76 77)
+
+    movdqa      xmm1, xmm7        ; transpose coefficients(phase 3)
+    punpckldq   xmm7, xmm2        ; xmm7=(00 01 02 03 04 05 06 07 10 11 12 13 14 15 16 17)
+    punpckhdq   xmm1, xmm2        ; xmm1=(20 21 22 23 24 25 26 27 30 31 32 33 34 35 36 37)
+    movdqa      xmm3, xmm4        ; transpose coefficients(phase 3)
+    punpckldq   xmm4, xmm5        ; xmm4=(40 41 42 43 44 45 46 47 50 51 52 53 54 55 56 57)
+    punpckhdq   xmm3, xmm5        ; xmm3=(60 61 62 63 64 65 66 67 70 71 72 73 74 75 76 77)
+
+    pshufd      xmm6, xmm7, 0x4E  ; xmm6=(10 11 12 13 14 15 16 17 00 01 02 03 04 05 06 07)
+    pshufd      xmm0, xmm1, 0x4E  ; xmm0=(30 31 32 33 34 35 36 37 20 21 22 23 24 25 26 27)
+    pshufd      xmm2, xmm4, 0x4E  ; xmm2=(50 51 52 53 54 55 56 57 40 41 42 43 44 45 46 47)
+    pshufd      xmm5, xmm3, 0x4E  ; xmm5=(70 71 72 73 74 75 76 77 60 61 62 63 64 65 66 67)
+
+    mov         rdxp, JSAMPROW [rdi+0*SIZEOF_JSAMPROW]
+    mov         rsip, JSAMPROW [rdi+2*SIZEOF_JSAMPROW]
+    movq        XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm7
+    movq        XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm1
+    mov         rdxp, JSAMPROW [rdi+4*SIZEOF_JSAMPROW]
+    mov         rsip, JSAMPROW [rdi+6*SIZEOF_JSAMPROW]
+    movq        XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm4
+    movq        XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm3
+
+    mov         rdxp, JSAMPROW [rdi+1*SIZEOF_JSAMPROW]
+    mov         rsip, JSAMPROW [rdi+3*SIZEOF_JSAMPROW]
+    movq        XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm6
+    movq        XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm0
+    mov         rdxp, JSAMPROW [rdi+5*SIZEOF_JSAMPROW]
+    mov         rsip, JSAMPROW [rdi+7*SIZEOF_JSAMPROW]
+    movq        XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm2
+    movq        XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm5
+
+    uncollect_args 4
+    mov         rsp, rbp                ; rsp <- aligned rbp
+    pop         rsp                     ; rsp <- original rbp
+    pop         rbp
+    ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+    align       32
diff --git a/media/libjpeg/simd/x86_64/jidctred-sse2.asm b/media/libjpeg/simd/x86_64/jidctred-sse2.asm
new file mode 100644
index 0000000000..4ece9d891c
--- /dev/null
+++ b/media/libjpeg/simd/x86_64/jidctred-sse2.asm
@@ -0,0 +1,574 @@
+;
+; jidctred.asm - reduced-size IDCT (64-bit SSE2)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+; Copyright (C) 2009, 2016, D. R. Commander.
+; Copyright (C) 2018, Matthias Räncker.
+;
+; Based on the x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; This file contains inverse-DCT routines that produce reduced-size
+; output: either 4x4 or 2x2 pixels from an 8x8 DCT block.
+; The following code is based directly on the IJG's original jidctred.c;
+; see the jidctred.c for more details.
+
+%include "jsimdext.inc"
+%include "jdct.inc"
+
+; --------------------------------------------------------------------------
+
+%define CONST_BITS    13
+%define PASS1_BITS    2
+
+%define DESCALE_P1_4  (CONST_BITS - PASS1_BITS + 1)
+%define DESCALE_P2_4  (CONST_BITS + PASS1_BITS + 3 + 1)
+%define DESCALE_P1_2  (CONST_BITS - PASS1_BITS + 2)
+%define DESCALE_P2_2  (CONST_BITS + PASS1_BITS + 3 + 2)
+
+%if CONST_BITS == 13
+F_0_211 equ  1730  ; FIX(0.211164243)
+F_0_509 equ  4176  ; FIX(0.509795579)
+F_0_601 equ  4926  ; FIX(0.601344887)
+F_0_720 equ  5906  ; FIX(0.720959822)
+F_0_765 equ  6270  ; FIX(0.765366865)
+F_0_850 equ  6967  ; FIX(0.850430095)
+F_0_899 equ  7373  ; FIX(0.899976223)
+F_1_061 equ  8697  ; FIX(1.061594337)
+F_1_272 equ 10426  ; FIX(1.272758580)
+F_1_451 equ 11893  ; FIX(1.451774981)
+F_1_847 equ 15137  ; FIX(1.847759065)
+F_2_172 equ 17799  ; FIX(2.172734803)
+F_2_562 equ 20995  ; FIX(2.562915447)
+F_3_624 equ 29692  ; FIX(3.624509785)
+%else
+; NASM cannot do compile-time arithmetic on floating-point constants.
+%define DESCALE(x, n)  (((x) + (1 << ((n) - 1))) >> (n))
+F_0_211 equ DESCALE( 226735879, 30 - CONST_BITS)  ; FIX(0.211164243)
+F_0_509 equ DESCALE( 547388834, 30 - CONST_BITS)  ; FIX(0.509795579)
+F_0_601 equ DESCALE( 645689155, 30 - CONST_BITS)  ; FIX(0.601344887)
+F_0_720 equ DESCALE( 774124714, 30 - CONST_BITS)  ; FIX(0.720959822)
+F_0_765 equ DESCALE( 821806413, 30 - CONST_BITS)  ; FIX(0.765366865)
+F_0_850 equ DESCALE( 913142361, 30 - CONST_BITS)  ; FIX(0.850430095)
+F_0_899 equ DESCALE( 966342111, 30 - CONST_BITS)  ; FIX(0.899976223)
+F_1_061 equ DESCALE(1139878239, 30 - CONST_BITS)  ; FIX(1.061594337)
+F_1_272 equ DESCALE(1366614119, 30 - CONST_BITS)  ; FIX(1.272758580)
+F_1_451 equ DESCALE(1558831516, 30 - CONST_BITS)  ; FIX(1.451774981)
+F_1_847 equ DESCALE(1984016188, 30 - CONST_BITS)  ; FIX(1.847759065)
+F_2_172 equ DESCALE(2332956230, 30 - CONST_BITS)  ; FIX(2.172734803)
+F_2_562 equ DESCALE(2751909506, 30 - CONST_BITS)  ; FIX(2.562915447)
+F_3_624 equ DESCALE(3891787747, 30 - CONST_BITS)  ; FIX(3.624509785)
+%endif
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_CONST
+
+    alignz      32
+    GLOBAL_DATA(jconst_idct_red_sse2)
+
+EXTN(jconst_idct_red_sse2):
+
+PW_F184_MF076   times 4  dw  F_1_847, -F_0_765
+PW_F256_F089    times 4  dw  F_2_562,  F_0_899
+PW_F106_MF217   times 4  dw  F_1_061, -F_2_172
+PW_MF060_MF050  times 4  dw -F_0_601, -F_0_509
+PW_F145_MF021   times 4  dw  F_1_451, -F_0_211
+PW_F362_MF127   times 4  dw  F_3_624, -F_1_272
+PW_F085_MF072   times 4  dw  F_0_850, -F_0_720
+PD_DESCALE_P1_4 times 4  dd  1 << (DESCALE_P1_4 - 1)
+PD_DESCALE_P2_4 times 4  dd  1 << (DESCALE_P2_4 - 1)
+PD_DESCALE_P1_2 times 4  dd  1 << (DESCALE_P1_2 - 1)
+PD_DESCALE_P2_2 times 4  dd  1 << (DESCALE_P2_2 - 1)
+PB_CENTERJSAMP  times 16 db  CENTERJSAMPLE
+
+    alignz      32
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_TEXT
+    BITS        64
+;
+; Perform dequantization and inverse DCT on one block of coefficients,
+; producing a reduced-size 4x4 output block.
+;
+; GLOBAL(void)
+; jsimd_idct_4x4_sse2(void *dct_table, JCOEFPTR coef_block,
+;                     JSAMPARRAY output_buf, JDIMENSION output_col)
+;
+
+; r10 = void *dct_table
+; r11 = JCOEFPTR coef_block
+; r12 = JSAMPARRAY output_buf
+; r13d = JDIMENSION output_col
+
+%define original_rbp  rbp + 0
+%define wk(i)         rbp - (WK_NUM - (i)) * SIZEOF_XMMWORD
+                                        ; xmmword wk[WK_NUM]
+%define WK_NUM        2
+
+    align       32
+    GLOBAL_FUNCTION(jsimd_idct_4x4_sse2)
+
+EXTN(jsimd_idct_4x4_sse2):
+    push        rbp
+    mov         rax, rsp                     ; rax = original rbp
+    sub         rsp, byte 4
+    and         rsp, byte (-SIZEOF_XMMWORD)  ; align to 128 bits
+    mov         [rsp], rax
+    mov         rbp, rsp                     ; rbp = aligned rbp
+    lea         rsp, [wk(0)]
+    collect_args 4
+
+    ; ---- Pass 1: process columns from input.
+
+    mov         rdx, r10                ; quantptr
+    mov         rsi, r11                ; inptr
+
+%ifndef NO_ZERO_COLUMN_TEST_4X4_SSE2
+    mov         eax, dword [DWBLOCK(1,0,rsi,SIZEOF_JCOEF)]
+    or          eax, dword [DWBLOCK(2,0,rsi,SIZEOF_JCOEF)]
+    jnz         short .columnDCT
+
+    movdqa      xmm0, XMMWORD [XMMBLOCK(1,0,rsi,SIZEOF_JCOEF)]
+    movdqa      xmm1, XMMWORD [XMMBLOCK(2,0,rsi,SIZEOF_JCOEF)]
+    por         xmm0, XMMWORD [XMMBLOCK(3,0,rsi,SIZEOF_JCOEF)]
+    por         xmm1, XMMWORD [XMMBLOCK(5,0,rsi,SIZEOF_JCOEF)]
+    por         xmm0, XMMWORD [XMMBLOCK(6,0,rsi,SIZEOF_JCOEF)]
+    por         xmm1, XMMWORD [XMMBLOCK(7,0,rsi,SIZEOF_JCOEF)]
+    por         xmm0, xmm1
+    packsswb    xmm0, xmm0
+    packsswb    xmm0, xmm0
+    movd        eax, xmm0
+    test        rax, rax
+    jnz         short .columnDCT
+
+    ; -- AC terms all zero
+
+    movdqa      xmm0, XMMWORD [XMMBLOCK(0,0,rsi,SIZEOF_JCOEF)]
+    pmullw      xmm0, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
+
+    psllw       xmm0, PASS1_BITS
+
+    movdqa      xmm3, xmm0        ; xmm0=in0=(00 01 02 03 04 05 06 07)
+    punpcklwd   xmm0, xmm0        ; xmm0=(00 00 01 01 02 02 03 03)
+    punpckhwd   xmm3, xmm3        ; xmm3=(04 04 05 05 06 06 07 07)
+
+    pshufd      xmm1, xmm0, 0x50  ; xmm1=[col0 col1]=(00 00 00 00 01 01 01 01)
+    pshufd      xmm0, xmm0, 0xFA  ; xmm0=[col2 col3]=(02 02 02 02 03 03 03 03)
+    pshufd      xmm6, xmm3, 0x50  ; xmm6=[col4 col5]=(04 04 04 04 05 05 05 05)
+    pshufd      xmm3, xmm3, 0xFA  ; xmm3=[col6 col7]=(06 06 06 06 07 07 07 07)
+
+    jmp         near .column_end
+%endif
+.columnDCT:
+
+    ; -- Odd part
+
+    movdqa      xmm0, XMMWORD [XMMBLOCK(1,0,rsi,SIZEOF_JCOEF)]
+    movdqa      xmm1, XMMWORD [XMMBLOCK(3,0,rsi,SIZEOF_JCOEF)]
+    pmullw      xmm0, XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
+    pmullw      xmm1, XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
+    movdqa      xmm2, XMMWORD [XMMBLOCK(5,0,rsi,SIZEOF_JCOEF)]
+    movdqa      xmm3, XMMWORD [XMMBLOCK(7,0,rsi,SIZEOF_JCOEF)]
+    pmullw      xmm2, XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
+    pmullw      xmm3, XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
+
+    movdqa      xmm4, xmm0
+    movdqa      xmm5, xmm0
+    punpcklwd   xmm4, xmm1
+    punpckhwd   xmm5, xmm1
+    movdqa      xmm0, xmm4
+    movdqa      xmm1, xmm5
+    pmaddwd     xmm4, [rel PW_F256_F089]   ; xmm4=(tmp2L)
+    pmaddwd     xmm5, [rel PW_F256_F089]   ; xmm5=(tmp2H)
+    pmaddwd     xmm0, [rel PW_F106_MF217]  ; xmm0=(tmp0L)
+    pmaddwd     xmm1, [rel PW_F106_MF217]  ; xmm1=(tmp0H)
+
+    movdqa      xmm6, xmm2
+    movdqa      xmm7, xmm2
+    punpcklwd   xmm6, xmm3
+    punpckhwd   xmm7, xmm3
+    movdqa      xmm2, xmm6
+    movdqa      xmm3, xmm7
+    pmaddwd     xmm6, [rel PW_MF060_MF050]  ; xmm6=(tmp2L)
+    pmaddwd     xmm7, [rel PW_MF060_MF050]  ; xmm7=(tmp2H)
+    pmaddwd     xmm2, [rel PW_F145_MF021]   ; xmm2=(tmp0L)
+    pmaddwd     xmm3, [rel PW_F145_MF021]   ; xmm3=(tmp0H)
+
+    paddd       xmm6, xmm4              ; xmm6=tmp2L
+    paddd       xmm7, xmm5              ; xmm7=tmp2H
+    paddd       xmm2, xmm0              ; xmm2=tmp0L
+    paddd       xmm3, xmm1              ; xmm3=tmp0H
+
+    movdqa      XMMWORD [wk(0)], xmm2   ; wk(0)=tmp0L
+    movdqa      XMMWORD [wk(1)], xmm3   ; wk(1)=tmp0H
+
+    ; -- Even part
+
+    movdqa      xmm4, XMMWORD [XMMBLOCK(0,0,rsi,SIZEOF_JCOEF)]
+    movdqa      xmm5, XMMWORD [XMMBLOCK(2,0,rsi,SIZEOF_JCOEF)]
+    movdqa      xmm0, XMMWORD [XMMBLOCK(6,0,rsi,SIZEOF_JCOEF)]
+    pmullw      xmm4, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
+    pmullw      xmm5, XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
+    pmullw      xmm0, XMMWORD [XMMBLOCK(6,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
+
+    pxor        xmm1, xmm1
+    pxor        xmm2, xmm2
+    punpcklwd   xmm1, xmm4               ; xmm1=tmp0L
+    punpckhwd   xmm2, xmm4               ; xmm2=tmp0H
+    psrad       xmm1, (16-CONST_BITS-1)  ; psrad xmm1,16 & pslld xmm1,CONST_BITS+1
+    psrad       xmm2, (16-CONST_BITS-1)  ; psrad xmm2,16 & pslld xmm2,CONST_BITS+1
+
+    movdqa      xmm3, xmm5              ; xmm5=in2=z2
+    punpcklwd   xmm5, xmm0              ; xmm0=in6=z3
+    punpckhwd   xmm3, xmm0
+    pmaddwd     xmm5, [rel PW_F184_MF076]  ; xmm5=tmp2L
+    pmaddwd     xmm3, [rel PW_F184_MF076]  ; xmm3=tmp2H
+
+    movdqa      xmm4, xmm1
+    movdqa      xmm0, xmm2
+    paddd       xmm1, xmm5              ; xmm1=tmp10L
+    paddd       xmm2, xmm3              ; xmm2=tmp10H
+    psubd       xmm4, xmm5              ; xmm4=tmp12L
+    psubd       xmm0, xmm3              ; xmm0=tmp12H
+
+    ; -- Final output stage
+
+    movdqa      xmm5, xmm1
+    movdqa      xmm3, xmm2
+    paddd       xmm1, xmm6              ; xmm1=data0L
+    paddd       xmm2, xmm7              ; xmm2=data0H
+    psubd       xmm5, xmm6              ; xmm5=data3L
+    psubd       xmm3, xmm7              ; xmm3=data3H
+
+    movdqa      xmm6, [rel PD_DESCALE_P1_4]  ; xmm6=[rel PD_DESCALE_P1_4]
+
+    paddd       xmm1, xmm6
+    paddd       xmm2, xmm6
+    psrad       xmm1, DESCALE_P1_4
+    psrad       xmm2, DESCALE_P1_4
+    paddd       xmm5, xmm6
+    paddd       xmm3, xmm6
+    psrad       xmm5, DESCALE_P1_4
+    psrad       xmm3, DESCALE_P1_4
+
+    packssdw    xmm1, xmm2              ; xmm1=data0=(00 01 02 03 04 05 06 07)
+    packssdw    xmm5, xmm3              ; xmm5=data3=(30 31 32 33 34 35 36 37)
+
+    movdqa      xmm7, XMMWORD [wk(0)]   ; xmm7=tmp0L
+    movdqa      xmm6, XMMWORD [wk(1)]   ; xmm6=tmp0H
+
+    movdqa      xmm2, xmm4
+    movdqa      xmm3, xmm0
+    paddd       xmm4, xmm7              ; xmm4=data1L
+    paddd       xmm0, xmm6              ; xmm0=data1H
+    psubd       xmm2, xmm7              ; xmm2=data2L
+    psubd       xmm3, xmm6              ; xmm3=data2H
+
+    movdqa      xmm7, [rel PD_DESCALE_P1_4]  ; xmm7=[rel PD_DESCALE_P1_4]
+
+    paddd       xmm4, xmm7
+    paddd       xmm0, xmm7
+    psrad       xmm4, DESCALE_P1_4
+    psrad       xmm0, DESCALE_P1_4
+    paddd       xmm2, xmm7
+    paddd       xmm3, xmm7
+    psrad       xmm2, DESCALE_P1_4
+    psrad       xmm3, DESCALE_P1_4
+
+    packssdw    xmm4, xmm0        ; xmm4=data1=(10 11 12 13 14 15 16 17)
+    packssdw    xmm2, xmm3        ; xmm2=data2=(20 21 22 23 24 25 26 27)
+
+    movdqa      xmm6, xmm1        ; transpose coefficients(phase 1)
+    punpcklwd   xmm1, xmm4        ; xmm1=(00 10 01 11 02 12 03 13)
+    punpckhwd   xmm6, xmm4        ; xmm6=(04 14 05 15 06 16 07 17)
+    movdqa      xmm7, xmm2        ; transpose coefficients(phase 1)
+    punpcklwd   xmm2, xmm5        ; xmm2=(20 30 21 31 22 32 23 33)
+    punpckhwd   xmm7, xmm5        ; xmm7=(24 34 25 35 26 36 27 37)
+
+    movdqa      xmm0, xmm1        ; transpose coefficients(phase 2)
+    punpckldq   xmm1, xmm2        ; xmm1=[col0 col1]=(00 10 20 30 01 11 21 31)
+    punpckhdq   xmm0, xmm2        ; xmm0=[col2 col3]=(02 12 22 32 03 13 23 33)
+    movdqa      xmm3, xmm6        ; transpose coefficients(phase 2)
+    punpckldq   xmm6, xmm7        ; xmm6=[col4 col5]=(04 14 24 34 05 15 25 35)
+    punpckhdq   xmm3, xmm7        ; xmm3=[col6 col7]=(06 16 26 36 07 17 27 37)
+.column_end:
+
+    ; -- Prefetch the next coefficient block
+
+    prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 0*32]
+    prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 1*32]
+    prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 2*32]
+    prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 3*32]
+
+    ; ---- Pass 2: process rows, store into output array.
+
+    mov         rax, [original_rbp]
+    mov         rdi, r12                ; (JSAMPROW *)
+    mov         eax, r13d
+
+    ; -- Even part
+
+    pxor        xmm4, xmm4
+    punpcklwd   xmm4, xmm1               ; xmm4=tmp0
+    psrad       xmm4, (16-CONST_BITS-1)  ; psrad xmm4,16 & pslld xmm4,CONST_BITS+1
+
+    ; -- Odd part
+
+    punpckhwd   xmm1, xmm0
+    punpckhwd   xmm6, xmm3
+    movdqa      xmm5, xmm1
+    movdqa      xmm2, xmm6
+    pmaddwd     xmm1, [rel PW_F256_F089]    ; xmm1=(tmp2)
+    pmaddwd     xmm6, [rel PW_MF060_MF050]  ; xmm6=(tmp2)
+    pmaddwd     xmm5, [rel PW_F106_MF217]   ; xmm5=(tmp0)
+    pmaddwd     xmm2, [rel PW_F145_MF021]   ; xmm2=(tmp0)
+
+    paddd       xmm6, xmm1              ; xmm6=tmp2
+    paddd       xmm2, xmm5              ; xmm2=tmp0
+
+    ; -- Even part
+
+    punpcklwd   xmm0, xmm3
+    pmaddwd     xmm0, [rel PW_F184_MF076]  ; xmm0=tmp2
+
+    movdqa      xmm7, xmm4
+    paddd       xmm4, xmm0              ; xmm4=tmp10
+    psubd       xmm7, xmm0              ; xmm7=tmp12
+
+    ; -- Final output stage
+
+    movdqa      xmm1, [rel PD_DESCALE_P2_4]  ; xmm1=[rel PD_DESCALE_P2_4]
+
+    movdqa      xmm5, xmm4
+    movdqa      xmm3, xmm7
+    paddd       xmm4, xmm6              ; xmm4=data0=(00 10 20 30)
+    paddd       xmm7, xmm2              ; xmm7=data1=(01 11 21 31)
+    psubd       xmm5, xmm6              ; xmm5=data3=(03 13 23 33)
+    psubd       xmm3, xmm2              ; xmm3=data2=(02 12 22 32)
+
+    paddd       xmm4, xmm1
+    paddd       xmm7, xmm1
+    psrad       xmm4, DESCALE_P2_4
+    psrad       xmm7, DESCALE_P2_4
+    paddd       xmm5, xmm1
+    paddd       xmm3, xmm1
+    psrad       xmm5, DESCALE_P2_4
+    psrad       xmm3, DESCALE_P2_4
+
+    packssdw    xmm4, xmm3              ; xmm4=(00 10 20 30 02 12 22 32)
+    packssdw    xmm7, xmm5              ; xmm7=(01 11 21 31 03 13 23 33)
+
+    movdqa      xmm0, xmm4              ; transpose coefficients(phase 1)
+    punpcklwd   xmm4, xmm7              ; xmm4=(00 01 10 11 20 21 30 31)
+    punpckhwd   xmm0, xmm7              ; xmm0=(02 03 12 13 22 23 32 33)
+
+    movdqa      xmm6, xmm4              ; transpose coefficients(phase 2)
+    punpckldq   xmm4, xmm0              ; xmm4=(00 01 02 03 10 11 12 13)
+    punpckhdq   xmm6, xmm0              ; xmm6=(20 21 22 23 30 31 32 33)
+
+    packsswb    xmm4, xmm6              ; xmm4=(00 01 02 03 10 11 12 13 20 ..)
+    paddb       xmm4, [rel PB_CENTERJSAMP]
+
+    pshufd      xmm2, xmm4, 0x39        ; xmm2=(10 11 12 13 20 21 22 23 30 ..)
+    pshufd      xmm1, xmm4, 0x4E        ; xmm1=(20 21 22 23 30 31 32 33 00 ..)
+    pshufd      xmm3, xmm4, 0x93        ; xmm3=(30 31 32 33 00 01 02 03 10 ..)
+
+    mov         rdxp, JSAMPROW [rdi+0*SIZEOF_JSAMPROW]
+    mov         rsip, JSAMPROW [rdi+1*SIZEOF_JSAMPROW]
+    movd        XMM_DWORD [rdx+rax*SIZEOF_JSAMPLE], xmm4
+    movd        XMM_DWORD [rsi+rax*SIZEOF_JSAMPLE], xmm2
+    mov         rdxp, JSAMPROW [rdi+2*SIZEOF_JSAMPROW]
+    mov         rsip, JSAMPROW [rdi+3*SIZEOF_JSAMPROW]
+    movd        XMM_DWORD [rdx+rax*SIZEOF_JSAMPLE], xmm1
+    movd        XMM_DWORD [rsi+rax*SIZEOF_JSAMPLE], xmm3
+
+    uncollect_args 4
+    mov         rsp, rbp                ; rsp <- aligned rbp
+    pop         rsp                     ; rsp <- original rbp
+    pop         rbp
+    ret
+
+; --------------------------------------------------------------------------
+;
+; Perform dequantization and inverse DCT on one block of coefficients,
+; producing a reduced-size 2x2 output block.
+;
+; GLOBAL(void)
+; jsimd_idct_2x2_sse2(void *dct_table, JCOEFPTR coef_block,
+;                     JSAMPARRAY output_buf, JDIMENSION output_col)
+;
+
+; r10 = void *dct_table
+; r11 = JCOEFPTR coef_block
+; r12 = JSAMPARRAY output_buf
+; r13d = JDIMENSION output_col
+
+    align       32
+    GLOBAL_FUNCTION(jsimd_idct_2x2_sse2)
+
+EXTN(jsimd_idct_2x2_sse2):
+    push        rbp
+    mov         rax, rsp
+    mov         rbp, rsp
+    collect_args 4
+    push        rbx
+
+    ; ---- Pass 1: process columns from input.
+
+    mov         rdx, r10                ; quantptr
+    mov         rsi, r11                ; inptr
+
+    ; | input:                  | result:        |
+    ; | 00 01 ** 03 ** 05 ** 07 |                |
+    ; | 10 11 ** 13 ** 15 ** 17 |                |
+    ; | ** ** ** ** ** ** ** ** |                |
+    ; | 30 31 ** 33 ** 35 ** 37 | A0 A1 A3 A5 A7 |
+    ; | ** ** ** ** ** ** ** ** | B0 B1 B3 B5 B7 |
+    ; | 50 51 ** 53 ** 55 ** 57 |                |
+    ; | ** ** ** ** ** ** ** ** |                |
+    ; | 70 71 ** 73 ** 75 ** 77 |                |
+
+    ; -- Odd part
+
+    movdqa      xmm0, XMMWORD [XMMBLOCK(1,0,rsi,SIZEOF_JCOEF)]
+    movdqa      xmm1, XMMWORD [XMMBLOCK(3,0,rsi,SIZEOF_JCOEF)]
+    pmullw      xmm0, XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
+    pmullw      xmm1, XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
+    movdqa      xmm2, XMMWORD [XMMBLOCK(5,0,rsi,SIZEOF_JCOEF)]
+    movdqa      xmm3, XMMWORD [XMMBLOCK(7,0,rsi,SIZEOF_JCOEF)]
+    pmullw      xmm2, XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
+    pmullw      xmm3, XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
+
+    ; xmm0=(10 11 ** 13 ** 15 ** 17), xmm1=(30 31 ** 33 ** 35 ** 37)
+    ; xmm2=(50 51 ** 53 ** 55 ** 57), xmm3=(70 71 ** 73 ** 75 ** 77)
+
+    pcmpeqd     xmm7, xmm7
+    pslld       xmm7, WORD_BIT          ; xmm7={0x0000 0xFFFF 0x0000 0xFFFF ..}
+
+    movdqa      xmm4, xmm0              ; xmm4=(10 11 ** 13 ** 15 ** 17)
+    movdqa      xmm5, xmm2              ; xmm5=(50 51 ** 53 ** 55 ** 57)
+    punpcklwd   xmm4, xmm1              ; xmm4=(10 30 11 31 ** ** 13 33)
+    punpcklwd   xmm5, xmm3              ; xmm5=(50 70 51 71 ** ** 53 73)
+    pmaddwd     xmm4, [rel PW_F362_MF127]
+    pmaddwd     xmm5, [rel PW_F085_MF072]
+
+    psrld       xmm0, WORD_BIT          ; xmm0=(11 -- 13 -- 15 -- 17 --)
+    pand        xmm1, xmm7              ; xmm1=(-- 31 -- 33 -- 35 -- 37)
+    psrld       xmm2, WORD_BIT          ; xmm2=(51 -- 53 -- 55 -- 57 --)
+    pand        xmm3, xmm7              ; xmm3=(-- 71 -- 73 -- 75 -- 77)
+    por         xmm0, xmm1              ; xmm0=(11 31 13 33 15 35 17 37)
+    por         xmm2, xmm3              ; xmm2=(51 71 53 73 55 75 57 77)
+    pmaddwd     xmm0, [rel PW_F362_MF127]
+    pmaddwd     xmm2, [rel PW_F085_MF072]
+
+    paddd       xmm4, xmm5              ; xmm4=tmp0[col0 col1 **** col3]
+    paddd       xmm0, xmm2              ; xmm0=tmp0[col1 col3 col5 col7]
+
+    ; -- Even part
+
+    movdqa      xmm6, XMMWORD [XMMBLOCK(0,0,rsi,SIZEOF_JCOEF)]
+    pmullw      xmm6, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
+
+    ; xmm6=(00 01 ** 03 ** 05 ** 07)
+
+    movdqa      xmm1, xmm6              ; xmm1=(00 01 ** 03 ** 05 ** 07)
+    pslld       xmm6, WORD_BIT          ; xmm6=(-- 00 -- ** -- ** -- **)
+    pand        xmm1, xmm7              ; xmm1=(-- 01 -- 03 -- 05 -- 07)
+    psrad       xmm6, (WORD_BIT-CONST_BITS-2)  ; xmm6=tmp10[col0 **** **** ****]
+    psrad       xmm1, (WORD_BIT-CONST_BITS-2)  ; xmm1=tmp10[col1 col3 col5 col7]
+
+    ; -- Final output stage
+
+    movdqa      xmm3, xmm6
+    movdqa      xmm5, xmm1
+    paddd       xmm6, xmm4      ; xmm6=data0[col0 **** **** ****]=(A0 ** ** **)
+    paddd       xmm1, xmm0      ; xmm1=data0[col1 col3 col5 col7]=(A1 A3 A5 A7)
+    psubd       xmm3, xmm4      ; xmm3=data1[col0 **** **** ****]=(B0 ** ** **)
+    psubd       xmm5, xmm0      ; xmm5=data1[col1 col3 col5 col7]=(B1 B3 B5 B7)
+
+    movdqa      xmm2, [rel PD_DESCALE_P1_2]  ; xmm2=[rel PD_DESCALE_P1_2]
+
+    punpckldq   xmm6, xmm3              ; xmm6=(A0 B0 ** **)
+
+    movdqa      xmm7, xmm1
+    punpcklqdq  xmm1, xmm5              ; xmm1=(A1 A3 B1 B3)
+    punpckhqdq  xmm7, xmm5              ; xmm7=(A5 A7 B5 B7)
+
+    paddd       xmm6, xmm2
+    psrad       xmm6, DESCALE_P1_2
+
+    paddd       xmm1, xmm2
+    paddd       xmm7, xmm2
+    psrad       xmm1, DESCALE_P1_2
+    psrad       xmm7, DESCALE_P1_2
+
+    ; -- Prefetch the next coefficient block
+
+    prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 0*32]
+    prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 1*32]
+    prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 2*32]
+    prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 3*32]
+
+    ; ---- Pass 2: process rows, store into output array.
+
+    mov         rdi, r12                ; (JSAMPROW *)
+    mov         eax, r13d
+
+    ; | input:| result:|
+    ; | A0 B0 |        |
+    ; | A1 B1 | C0 C1  |
+    ; | A3 B3 | D0 D1  |
+    ; | A5 B5 |        |
+    ; | A7 B7 |        |
+
+    ; -- Odd part
+
+    packssdw    xmm1, xmm1              ; xmm1=(A1 A3 B1 B3 A1 A3 B1 B3)
+    packssdw    xmm7, xmm7              ; xmm7=(A5 A7 B5 B7 A5 A7 B5 B7)
+    pmaddwd     xmm1, [rel PW_F362_MF127]
+    pmaddwd     xmm7, [rel PW_F085_MF072]
+
+    paddd       xmm1, xmm7              ; xmm1=tmp0[row0 row1 row0 row1]
+
+    ; -- Even part
+
+    pslld       xmm6, (CONST_BITS+2)    ; xmm6=tmp10[row0 row1 **** ****]
+
+    ; -- Final output stage
+
+    movdqa      xmm4, xmm6
+    paddd       xmm6, xmm1     ; xmm6=data0[row0 row1 **** ****]=(C0 C1 ** **)
+    psubd       xmm4, xmm1     ; xmm4=data1[row0 row1 **** ****]=(D0 D1 ** **)
+
+    punpckldq   xmm6, xmm4     ; xmm6=(C0 D0 C1 D1)
+
+    paddd       xmm6, [rel PD_DESCALE_P2_2]
+    psrad       xmm6, DESCALE_P2_2
+
+    packssdw    xmm6, xmm6              ; xmm6=(C0 D0 C1 D1 C0 D0 C1 D1)
+    packsswb    xmm6, xmm6              ; xmm6=(C0 D0 C1 D1 C0 D0 C1 D1 ..)
+    paddb       xmm6, [rel PB_CENTERJSAMP]
+
+    pextrw      ebx, xmm6, 0x00         ; ebx=(C0 D0 -- --)
+    pextrw      ecx, xmm6, 0x01         ; ecx=(C1 D1 -- --)
+
+    mov         rdxp, JSAMPROW [rdi+0*SIZEOF_JSAMPROW]
+    mov         rsip, JSAMPROW [rdi+1*SIZEOF_JSAMPROW]
+    mov         word [rdx+rax*SIZEOF_JSAMPLE], bx
+    mov         word [rsi+rax*SIZEOF_JSAMPLE], cx
+
+    pop         rbx
+    uncollect_args 4
+    pop         rbp
+    ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+    align       32
diff --git a/media/libjpeg/simd/x86_64/jquantf-sse2.asm b/media/libjpeg/simd/x86_64/jquantf-sse2.asm
new file mode 100644
index 0000000000..ab2e3954f6
--- /dev/null
+++ b/media/libjpeg/simd/x86_64/jquantf-sse2.asm
@@ -0,0 +1,155 @@
+;
+; jquantf.asm - sample data conversion and quantization (64-bit SSE & SSE2)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+; Copyright (C) 2009, 2016, D. R. Commander.
+; Copyright (C) 2018, Matthias Räncker.
+;
+; Based on the x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+
+%include "jsimdext.inc"
+%include "jdct.inc"
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_TEXT
+    BITS        64
+;
+; Load data into workspace, applying unsigned->signed conversion
+;
+; GLOBAL(void)
+; jsimd_convsamp_float_sse2(JSAMPARRAY sample_data, JDIMENSION start_col,
+;                           FAST_FLOAT *workspace);
+;
+
+; r10 = JSAMPARRAY sample_data
+; r11d = JDIMENSION start_col
+; r12 = FAST_FLOAT *workspace
+
+    align       32
+    GLOBAL_FUNCTION(jsimd_convsamp_float_sse2)
+
+EXTN(jsimd_convsamp_float_sse2):
+    push        rbp
+    mov         rax, rsp
+    mov         rbp, rsp
+    collect_args 3
+    push        rbx
+
+    pcmpeqw     xmm7, xmm7
+    psllw       xmm7, 7
+    packsswb    xmm7, xmm7              ; xmm7 = PB_CENTERJSAMPLE (0x808080..)
+
+    mov         rsi, r10
+    mov         eax, r11d
+    mov         rdi, r12
+    mov         rcx, DCTSIZE/2
+.convloop:
+    mov         rbxp, JSAMPROW [rsi+0*SIZEOF_JSAMPROW]  ; (JSAMPLE *)
+    mov         rdxp, JSAMPROW [rsi+1*SIZEOF_JSAMPROW]  ; (JSAMPLE *)
+
+    movq        xmm0, XMM_MMWORD [rbx+rax*SIZEOF_JSAMPLE]
+    movq        xmm1, XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE]
+
+    psubb       xmm0, xmm7              ; xmm0=(01234567)
+    psubb       xmm1, xmm7              ; xmm1=(89ABCDEF)
+
+    punpcklbw   xmm0, xmm0              ; xmm0=(*0*1*2*3*4*5*6*7)
+    punpcklbw   xmm1, xmm1              ; xmm1=(*8*9*A*B*C*D*E*F)
+
+    punpcklwd   xmm2, xmm0              ; xmm2=(***0***1***2***3)
+    punpckhwd   xmm0, xmm0              ; xmm0=(***4***5***6***7)
+    punpcklwd   xmm3, xmm1              ; xmm3=(***8***9***A***B)
+    punpckhwd   xmm1, xmm1              ; xmm1=(***C***D***E***F)
+
+    psrad       xmm2, (DWORD_BIT-BYTE_BIT)  ; xmm2=(0123)
+    psrad       xmm0, (DWORD_BIT-BYTE_BIT)  ; xmm0=(4567)
+    cvtdq2ps    xmm2, xmm2                  ; xmm2=(0123)
+    cvtdq2ps    xmm0, xmm0                  ; xmm0=(4567)
+    psrad       xmm3, (DWORD_BIT-BYTE_BIT)  ; xmm3=(89AB)
+    psrad       xmm1, (DWORD_BIT-BYTE_BIT)  ; xmm1=(CDEF)
+    cvtdq2ps    xmm3, xmm3                  ; xmm3=(89AB)
+    cvtdq2ps    xmm1, xmm1                  ; xmm1=(CDEF)
+
+    movaps      XMMWORD [XMMBLOCK(0,0,rdi,SIZEOF_FAST_FLOAT)], xmm2
+    movaps      XMMWORD [XMMBLOCK(0,1,rdi,SIZEOF_FAST_FLOAT)], xmm0
+    movaps      XMMWORD [XMMBLOCK(1,0,rdi,SIZEOF_FAST_FLOAT)], xmm3
+    movaps      XMMWORD [XMMBLOCK(1,1,rdi,SIZEOF_FAST_FLOAT)], xmm1
+
+    add         rsi, byte 2*SIZEOF_JSAMPROW
+    add         rdi, byte 2*DCTSIZE*SIZEOF_FAST_FLOAT
+    dec         rcx
+    jnz         short .convloop
+
+    pop         rbx
+    uncollect_args 3
+    pop         rbp
+    ret
+
+; --------------------------------------------------------------------------
+;
+; Quantize/descale the coefficients, and store into coef_block
+;
+; GLOBAL(void)
+; jsimd_quantize_float_sse2(JCOEFPTR coef_block, FAST_FLOAT *divisors,
+;                           FAST_FLOAT *workspace);
+;
+
+; r10 = JCOEFPTR coef_block
+; r11 = FAST_FLOAT *divisors
+; r12 = FAST_FLOAT *workspace
+
+    align       32
+    GLOBAL_FUNCTION(jsimd_quantize_float_sse2)
+
+EXTN(jsimd_quantize_float_sse2):
+    push        rbp
+    mov         rax, rsp
+    mov         rbp, rsp
+    collect_args 3
+
+    mov         rsi, r12
+    mov         rdx, r11
+    mov         rdi, r10
+    mov         rax, DCTSIZE2/16
+.quantloop:
+    movaps      xmm0, XMMWORD [XMMBLOCK(0,0,rsi,SIZEOF_FAST_FLOAT)]
+    movaps      xmm1, XMMWORD [XMMBLOCK(0,1,rsi,SIZEOF_FAST_FLOAT)]
+    mulps       xmm0, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_FAST_FLOAT)]
+    mulps       xmm1, XMMWORD [XMMBLOCK(0,1,rdx,SIZEOF_FAST_FLOAT)]
+    movaps      xmm2, XMMWORD [XMMBLOCK(1,0,rsi,SIZEOF_FAST_FLOAT)]
+    movaps      xmm3, XMMWORD [XMMBLOCK(1,1,rsi,SIZEOF_FAST_FLOAT)]
+    mulps       xmm2, XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_FAST_FLOAT)]
+    mulps       xmm3, XMMWORD [XMMBLOCK(1,1,rdx,SIZEOF_FAST_FLOAT)]
+
+    cvtps2dq    xmm0, xmm0
+    cvtps2dq    xmm1, xmm1
+    cvtps2dq    xmm2, xmm2
+    cvtps2dq    xmm3, xmm3
+
+    packssdw    xmm0, xmm1
+    packssdw    xmm2, xmm3
+
+    movdqa      XMMWORD [XMMBLOCK(0,0,rdi,SIZEOF_JCOEF)], xmm0
+    movdqa      XMMWORD [XMMBLOCK(1,0,rdi,SIZEOF_JCOEF)], xmm2
+
+    add         rsi, byte 16*SIZEOF_FAST_FLOAT
+    add         rdx, byte 16*SIZEOF_FAST_FLOAT
+    add         rdi, byte 16*SIZEOF_JCOEF
+    dec         rax
+    jnz         short .quantloop
+
+    uncollect_args 3
+    pop         rbp
+    ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+    align       32
diff --git a/media/libjpeg/simd/x86_64/jquanti-avx2.asm b/media/libjpeg/simd/x86_64/jquanti-avx2.asm
new file mode 100644
index 0000000000..70fe81139c
--- /dev/null
+++ b/media/libjpeg/simd/x86_64/jquanti-avx2.asm
@@ -0,0 +1,163 @@
+;
+; jquanti.asm - sample data conversion and quantization (64-bit AVX2)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+; Copyright (C) 2009, 2016, 2018, D. R. Commander.
+; Copyright (C) 2016, Matthieu Darbois.
+; Copyright (C) 2018, Matthias Räncker.
+;
+; Based on the x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+
+%include "jsimdext.inc"
+%include "jdct.inc"
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_TEXT
+    BITS        64
+;
+; Load data into workspace, applying unsigned->signed conversion
+;
+; GLOBAL(void)
+; jsimd_convsamp_avx2(JSAMPARRAY sample_data, JDIMENSION start_col,
+;                     DCTELEM *workspace);
+;
+
+; r10 = JSAMPARRAY sample_data
+; r11d = JDIMENSION start_col
+; r12 = DCTELEM *workspace
+
+    align       32
+    GLOBAL_FUNCTION(jsimd_convsamp_avx2)
+
+EXTN(jsimd_convsamp_avx2):
+    push        rbp
+    mov         rax, rsp
+    mov         rbp, rsp
+    collect_args 3
+
+    mov         eax, r11d
+
+    mov         rsip, JSAMPROW [r10+0*SIZEOF_JSAMPROW]  ; (JSAMPLE *)
+    mov         rdip, JSAMPROW [r10+1*SIZEOF_JSAMPROW]  ; (JSAMPLE *)
+    movq        xmm0, XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE]
+    pinsrq      xmm0, XMM_MMWORD [rdi+rax*SIZEOF_JSAMPLE], 1
+
+    mov         rsip, JSAMPROW [r10+2*SIZEOF_JSAMPROW]  ; (JSAMPLE *)
+    mov         rdip, JSAMPROW [r10+3*SIZEOF_JSAMPROW]  ; (JSAMPLE *)
+    movq        xmm1, XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE]
+    pinsrq      xmm1, XMM_MMWORD [rdi+rax*SIZEOF_JSAMPLE], 1
+
+    mov         rsip, JSAMPROW [r10+4*SIZEOF_JSAMPROW]  ; (JSAMPLE *)
+    mov         rdip, JSAMPROW [r10+5*SIZEOF_JSAMPROW]  ; (JSAMPLE *)
+    movq        xmm2, XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE]
+    pinsrq      xmm2, XMM_MMWORD [rdi+rax*SIZEOF_JSAMPLE], 1
+
+    mov         rsip, JSAMPROW [r10+6*SIZEOF_JSAMPROW]  ; (JSAMPLE *)
+    mov         rdip, JSAMPROW [r10+7*SIZEOF_JSAMPROW]  ; (JSAMPLE *)
+    movq        xmm3, XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE]
+    pinsrq      xmm3, XMM_MMWORD [rdi+rax*SIZEOF_JSAMPLE], 1
+
+    vpmovzxbw   ymm0, xmm0              ; ymm0=(00 01 02 03 04 05 06 07 10 11 12 13 14 15 16 17)
+    vpmovzxbw   ymm1, xmm1              ; ymm1=(20 21 22 23 24 25 26 27 30 31 32 33 34 35 36 37)
+    vpmovzxbw   ymm2, xmm2              ; ymm2=(40 41 42 43 44 45 46 47 50 51 52 53 54 55 56 57)
+    vpmovzxbw   ymm3, xmm3              ; ymm3=(60 61 62 63 64 65 66 67 70 71 72 73 74 75 76 77)
+
+    vpcmpeqw    ymm7, ymm7, ymm7
+    vpsllw      ymm7, ymm7, 7           ; ymm7={0xFF80 0xFF80 0xFF80 0xFF80 ..}
+
+    vpaddw      ymm0, ymm0, ymm7
+    vpaddw      ymm1, ymm1, ymm7
+    vpaddw      ymm2, ymm2, ymm7
+    vpaddw      ymm3, ymm3, ymm7
+
+    vmovdqu     YMMWORD [YMMBLOCK(0,0,r12,SIZEOF_DCTELEM)], ymm0
+    vmovdqu     YMMWORD [YMMBLOCK(2,0,r12,SIZEOF_DCTELEM)], ymm1
+    vmovdqu     YMMWORD [YMMBLOCK(4,0,r12,SIZEOF_DCTELEM)], ymm2
+    vmovdqu     YMMWORD [YMMBLOCK(6,0,r12,SIZEOF_DCTELEM)], ymm3
+
+    vzeroupper
+    uncollect_args 3
+    pop         rbp
+    ret
+
+; --------------------------------------------------------------------------
+;
+; Quantize/descale the coefficients, and store into coef_block
+;
+; This implementation is based on an algorithm described in
+;   "How to optimize for the Pentium family of microprocessors"
+;   (http://www.agner.org/assem/).
+;
+; GLOBAL(void)
+; jsimd_quantize_avx2(JCOEFPTR coef_block, DCTELEM *divisors,
+;                     DCTELEM *workspace);
+;
+
+%define RECIPROCAL(m, n, b) \
+  YMMBLOCK(DCTSIZE * 0 + (m), (n), (b), SIZEOF_DCTELEM)
+%define CORRECTION(m, n, b) \
+  YMMBLOCK(DCTSIZE * 1 + (m), (n), (b), SIZEOF_DCTELEM)
+%define SCALE(m, n, b) \
+  YMMBLOCK(DCTSIZE * 2 + (m), (n), (b), SIZEOF_DCTELEM)
+
+; r10 = JCOEFPTR coef_block
+; r11 = DCTELEM *divisors
+; r12 = DCTELEM *workspace
+
+    align       32
+    GLOBAL_FUNCTION(jsimd_quantize_avx2)
+
+EXTN(jsimd_quantize_avx2):
+    push        rbp
+    mov         rax, rsp
+    mov         rbp, rsp
+    collect_args 3
+
+    vmovdqu     ymm4, [YMMBLOCK(0,0,r12,SIZEOF_DCTELEM)]
+    vmovdqu     ymm5, [YMMBLOCK(2,0,r12,SIZEOF_DCTELEM)]
+    vmovdqu     ymm6, [YMMBLOCK(4,0,r12,SIZEOF_DCTELEM)]
+    vmovdqu     ymm7, [YMMBLOCK(6,0,r12,SIZEOF_DCTELEM)]
+    vpabsw      ymm0, ymm4
+    vpabsw      ymm1, ymm5
+    vpabsw      ymm2, ymm6
+    vpabsw      ymm3, ymm7
+
+    vpaddw      ymm0, YMMWORD [CORRECTION(0,0,r11)]  ; correction + roundfactor
+    vpaddw      ymm1, YMMWORD [CORRECTION(2,0,r11)]
+    vpaddw      ymm2, YMMWORD [CORRECTION(4,0,r11)]
+    vpaddw      ymm3, YMMWORD [CORRECTION(6,0,r11)]
+    vpmulhuw    ymm0, YMMWORD [RECIPROCAL(0,0,r11)]  ; reciprocal
+    vpmulhuw    ymm1, YMMWORD [RECIPROCAL(2,0,r11)]
+    vpmulhuw    ymm2, YMMWORD [RECIPROCAL(4,0,r11)]
+    vpmulhuw    ymm3, YMMWORD [RECIPROCAL(6,0,r11)]
+    vpmulhuw    ymm0, YMMWORD [SCALE(0,0,r11)]       ; scale
+    vpmulhuw    ymm1, YMMWORD [SCALE(2,0,r11)]
+    vpmulhuw    ymm2, YMMWORD [SCALE(4,0,r11)]
+    vpmulhuw    ymm3, YMMWORD [SCALE(6,0,r11)]
+
+    vpsignw     ymm0, ymm0, ymm4
+    vpsignw     ymm1, ymm1, ymm5
+    vpsignw     ymm2, ymm2, ymm6
+    vpsignw     ymm3, ymm3, ymm7
+
+    vmovdqu     [YMMBLOCK(0,0,r10,SIZEOF_DCTELEM)], ymm0
+    vmovdqu     [YMMBLOCK(2,0,r10,SIZEOF_DCTELEM)], ymm1
+    vmovdqu     [YMMBLOCK(4,0,r10,SIZEOF_DCTELEM)], ymm2
+    vmovdqu     [YMMBLOCK(6,0,r10,SIZEOF_DCTELEM)], ymm3
+
+    vzeroupper
+    uncollect_args 3
+    pop         rbp
+    ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+    align       32
diff --git a/media/libjpeg/simd/x86_64/jquanti-sse2.asm b/media/libjpeg/simd/x86_64/jquanti-sse2.asm
new file mode 100644
index 0000000000..3ee442027a
--- /dev/null
+++ b/media/libjpeg/simd/x86_64/jquanti-sse2.asm
@@ -0,0 +1,188 @@
+;
+; jquanti.asm - sample data conversion and quantization (64-bit SSE2)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+; Copyright (C) 2009, 2016, D. R. Commander.
+; Copyright (C) 2018, Matthias Räncker.
+;
+; Based on the x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+
+%include "jsimdext.inc"
+%include "jdct.inc"
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_TEXT
+    BITS        64
+;
+; Load data into workspace, applying unsigned->signed conversion
+;
+; GLOBAL(void)
+; jsimd_convsamp_sse2(JSAMPARRAY sample_data, JDIMENSION start_col,
+;                     DCTELEM *workspace);
+;
+
+; r10 = JSAMPARRAY sample_data
+; r11d = JDIMENSION start_col
+; r12 = DCTELEM *workspace
+
+    align       32
+    GLOBAL_FUNCTION(jsimd_convsamp_sse2)
+
+EXTN(jsimd_convsamp_sse2):
+    push        rbp
+    mov         rax, rsp
+    mov         rbp, rsp
+    collect_args 3
+    push        rbx
+
+    pxor        xmm6, xmm6              ; xmm6=(all 0's)
+    pcmpeqw     xmm7, xmm7
+    psllw       xmm7, 7                 ; xmm7={0xFF80 0xFF80 0xFF80 0xFF80 ..}
+
+    mov         rsi, r10
+    mov         eax, r11d
+    mov         rdi, r12
+    mov         rcx, DCTSIZE/4
+.convloop:
+    mov         rbxp, JSAMPROW [rsi+0*SIZEOF_JSAMPROW]  ; (JSAMPLE *)
+    mov         rdxp, JSAMPROW [rsi+1*SIZEOF_JSAMPROW]  ; (JSAMPLE *)
+
+    movq        xmm0, XMM_MMWORD [rbx+rax*SIZEOF_JSAMPLE]  ; xmm0=(01234567)
+    movq        xmm1, XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE]  ; xmm1=(89ABCDEF)
+
+    mov         rbxp, JSAMPROW [rsi+2*SIZEOF_JSAMPROW]  ; (JSAMPLE *)
+    mov         rdxp, JSAMPROW [rsi+3*SIZEOF_JSAMPROW]  ; (JSAMPLE *)
+
+    movq        xmm2, XMM_MMWORD [rbx+rax*SIZEOF_JSAMPLE]  ; xmm2=(GHIJKLMN)
+    movq        xmm3, XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE]  ; xmm3=(OPQRSTUV)
+
+    punpcklbw   xmm0, xmm6              ; xmm0=(01234567)
+    punpcklbw   xmm1, xmm6              ; xmm1=(89ABCDEF)
+    paddw       xmm0, xmm7
+    paddw       xmm1, xmm7
+    punpcklbw   xmm2, xmm6              ; xmm2=(GHIJKLMN)
+    punpcklbw   xmm3, xmm6              ; xmm3=(OPQRSTUV)
+    paddw       xmm2, xmm7
+    paddw       xmm3, xmm7
+
+    movdqa      XMMWORD [XMMBLOCK(0,0,rdi,SIZEOF_DCTELEM)], xmm0
+    movdqa      XMMWORD [XMMBLOCK(1,0,rdi,SIZEOF_DCTELEM)], xmm1
+    movdqa      XMMWORD [XMMBLOCK(2,0,rdi,SIZEOF_DCTELEM)], xmm2
+    movdqa      XMMWORD [XMMBLOCK(3,0,rdi,SIZEOF_DCTELEM)], xmm3
+
+    add         rsi, byte 4*SIZEOF_JSAMPROW
+    add         rdi, byte 4*DCTSIZE*SIZEOF_DCTELEM
+    dec         rcx
+    jnz         short .convloop
+
+    pop         rbx
+    uncollect_args 3
+    pop         rbp
+    ret
+
+; --------------------------------------------------------------------------
+;
+; Quantize/descale the coefficients, and store into coef_block
+;
+; This implementation is based on an algorithm described in
+;   "How to optimize for the Pentium family of microprocessors"
+;   (http://www.agner.org/assem/).
+;
+; GLOBAL(void)
+; jsimd_quantize_sse2(JCOEFPTR coef_block, DCTELEM *divisors,
+;                     DCTELEM *workspace);
+;
+
+%define RECIPROCAL(m, n, b) \
+  XMMBLOCK(DCTSIZE * 0 + (m), (n), (b), SIZEOF_DCTELEM)
+%define CORRECTION(m, n, b) \
+  XMMBLOCK(DCTSIZE * 1 + (m), (n), (b), SIZEOF_DCTELEM)
+%define SCALE(m, n, b) \
+  XMMBLOCK(DCTSIZE * 2 + (m), (n), (b), SIZEOF_DCTELEM)
+
+; r10 = JCOEFPTR coef_block
+; r11 = DCTELEM *divisors
+; r12 = DCTELEM *workspace
+
+    align       32
+    GLOBAL_FUNCTION(jsimd_quantize_sse2)
+
+EXTN(jsimd_quantize_sse2):
+    push        rbp
+    mov         rax, rsp
+    mov         rbp, rsp
+    collect_args 3
+
+    mov         rsi, r12
+    mov         rdx, r11
+    mov         rdi, r10
+    mov         rax, DCTSIZE2/32
+.quantloop:
+    movdqa      xmm4, XMMWORD [XMMBLOCK(0,0,rsi,SIZEOF_DCTELEM)]
+    movdqa      xmm5, XMMWORD [XMMBLOCK(1,0,rsi,SIZEOF_DCTELEM)]
+    movdqa      xmm6, XMMWORD [XMMBLOCK(2,0,rsi,SIZEOF_DCTELEM)]
+    movdqa      xmm7, XMMWORD [XMMBLOCK(3,0,rsi,SIZEOF_DCTELEM)]
+    movdqa      xmm0, xmm4
+    movdqa      xmm1, xmm5
+    movdqa      xmm2, xmm6
+    movdqa      xmm3, xmm7
+    psraw       xmm4, (WORD_BIT-1)
+    psraw       xmm5, (WORD_BIT-1)
+    psraw       xmm6, (WORD_BIT-1)
+    psraw       xmm7, (WORD_BIT-1)
+    pxor        xmm0, xmm4
+    pxor        xmm1, xmm5
+    pxor        xmm2, xmm6
+    pxor        xmm3, xmm7
+    psubw       xmm0, xmm4              ; if (xmm0 < 0) xmm0 = -xmm0;
+    psubw       xmm1, xmm5              ; if (xmm1 < 0) xmm1 = -xmm1;
+    psubw       xmm2, xmm6              ; if (xmm2 < 0) xmm2 = -xmm2;
+    psubw       xmm3, xmm7              ; if (xmm3 < 0) xmm3 = -xmm3;
+
+    paddw       xmm0, XMMWORD [CORRECTION(0,0,rdx)]  ; correction + roundfactor
+    paddw       xmm1, XMMWORD [CORRECTION(1,0,rdx)]
+    paddw       xmm2, XMMWORD [CORRECTION(2,0,rdx)]
+    paddw       xmm3, XMMWORD [CORRECTION(3,0,rdx)]
+    pmulhuw     xmm0, XMMWORD [RECIPROCAL(0,0,rdx)]  ; reciprocal
+    pmulhuw     xmm1, XMMWORD [RECIPROCAL(1,0,rdx)]
+    pmulhuw     xmm2, XMMWORD [RECIPROCAL(2,0,rdx)]
+    pmulhuw     xmm3, XMMWORD [RECIPROCAL(3,0,rdx)]
+    pmulhuw     xmm0, XMMWORD [SCALE(0,0,rdx)]       ; scale
+    pmulhuw     xmm1, XMMWORD [SCALE(1,0,rdx)]
+    pmulhuw     xmm2, XMMWORD [SCALE(2,0,rdx)]
+    pmulhuw     xmm3, XMMWORD [SCALE(3,0,rdx)]
+
+    pxor        xmm0, xmm4
+    pxor        xmm1, xmm5
+    pxor        xmm2, xmm6
+    pxor        xmm3, xmm7
+    psubw       xmm0, xmm4
+    psubw       xmm1, xmm5
+    psubw       xmm2, xmm6
+    psubw       xmm3, xmm7
+    movdqa      XMMWORD [XMMBLOCK(0,0,rdi,SIZEOF_DCTELEM)], xmm0
+    movdqa      XMMWORD [XMMBLOCK(1,0,rdi,SIZEOF_DCTELEM)], xmm1
+    movdqa      XMMWORD [XMMBLOCK(2,0,rdi,SIZEOF_DCTELEM)], xmm2
+    movdqa      XMMWORD [XMMBLOCK(3,0,rdi,SIZEOF_DCTELEM)], xmm3
+
+    add         rsi, byte 32*SIZEOF_DCTELEM
+    add         rdx, byte 32*SIZEOF_DCTELEM
+    add         rdi, byte 32*SIZEOF_JCOEF
+    dec         rax
+    jnz         near .quantloop
+
+    uncollect_args 3
+    pop         rbp
+    ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+    align       32
diff --git a/media/libjpeg/simd/x86_64/jsimd.c b/media/libjpeg/simd/x86_64/jsimd.c
new file mode 100644
index 0000000000..3f5ee77eb9
--- /dev/null
+++ b/media/libjpeg/simd/x86_64/jsimd.c
@@ -0,0 +1,1110 @@
+/*
+ * jsimd_x86_64.c
+ *
+ * Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+ * Copyright (C) 2009-2011, 2014, 2016, 2018, 2022-2023, D. R. Commander.
+ * Copyright (C) 2015-2016, 2018, 2022, Matthieu Darbois.
+ *
+ * Based on the x86 SIMD extension for IJG JPEG library,
+ * Copyright (C) 1999-2006, MIYASAKA Masaru.
+ * For conditions of distribution and use, see copyright notice in jsimdext.inc
+ *
+ * This file contains the interface between the "normal" portions
+ * of the library and the SIMD implementations when running on a
+ * 64-bit x86 architecture.
+ */
+
+#define JPEG_INTERNALS
+#include "../../jinclude.h"
+#include "../../jpeglib.h"
+#include "../../jsimd.h"
+#include "../../jdct.h"
+#include "../../jsimddct.h"
+#include "../jsimd.h"
+
+/*
+ * In the PIC cases, we have no guarantee that constants will keep
+ * their alignment. This macro allows us to verify it at runtime.
+ */
+#define IS_ALIGNED(ptr, order)  (((size_t)ptr & ((1 << order) - 1)) == 0)
+
+#define IS_ALIGNED_SSE(ptr)  (IS_ALIGNED(ptr, 4)) /* 16 byte alignment */
+#define IS_ALIGNED_AVX(ptr)  (IS_ALIGNED(ptr, 5)) /* 32 byte alignment */
+
+static THREAD_LOCAL unsigned int simd_support = (unsigned int)(~0);
+static THREAD_LOCAL unsigned int simd_huffman = 1;
+
+/*
+ * Check what SIMD accelerations are supported.
+ */
+LOCAL(void)
+init_simd(void)
+{
+#ifndef NO_GETENV
+  char env[2] = { 0 };
+#endif
+
+  if (simd_support != ~0U)
+    return;
+
+  simd_support = jpeg_simd_cpu_support();
+
+#ifndef NO_GETENV
+  /* Force different settings through environment variables */
+  if (!GETENV_S(env, 2, "JSIMD_FORCESSE2") && !strcmp(env, "1"))
+    simd_support &= JSIMD_SSE2;
+  if (!GETENV_S(env, 2, "JSIMD_FORCEAVX2") && !strcmp(env, "1"))
+    simd_support &= JSIMD_AVX2;
+  if (!GETENV_S(env, 2, "JSIMD_FORCENONE") && !strcmp(env, "1"))
+    simd_support = 0;
+  if (!GETENV_S(env, 2, "JSIMD_NOHUFFENC") && !strcmp(env, "1"))
+    simd_huffman = 0;
+#endif
+}
+
+GLOBAL(int)
+jsimd_can_rgb_ycc(void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+  if ((RGB_PIXELSIZE != 3) && (RGB_PIXELSIZE != 4))
+    return 0;
+
+  if ((simd_support & JSIMD_AVX2) &&
+      IS_ALIGNED_AVX(jconst_rgb_ycc_convert_avx2))
+    return 1;
+  if ((simd_support & JSIMD_SSE2) &&
+      IS_ALIGNED_SSE(jconst_rgb_ycc_convert_sse2))
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(int)
+jsimd_can_rgb_gray(void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+  if ((RGB_PIXELSIZE != 3) && (RGB_PIXELSIZE != 4))
+    return 0;
+
+  if ((simd_support & JSIMD_AVX2) &&
+      IS_ALIGNED_AVX(jconst_rgb_gray_convert_avx2))
+    return 1;
+  if ((simd_support & JSIMD_SSE2) &&
+      IS_ALIGNED_SSE(jconst_rgb_gray_convert_sse2))
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(int)
+jsimd_can_ycc_rgb(void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+  if ((RGB_PIXELSIZE != 3) && (RGB_PIXELSIZE != 4))
+    return 0;
+
+  if ((simd_support & JSIMD_AVX2) &&
+      IS_ALIGNED_AVX(jconst_ycc_rgb_convert_avx2))
+    return 1;
+  if ((simd_support & JSIMD_SSE2) &&
+      IS_ALIGNED_SSE(jconst_ycc_rgb_convert_sse2))
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(int)
+jsimd_can_ycc_rgb565(void)
+{
+  return 0;
+}
+
+GLOBAL(void)
+jsimd_rgb_ycc_convert(j_compress_ptr cinfo, JSAMPARRAY input_buf,
+                      JSAMPIMAGE output_buf, JDIMENSION output_row,
+                      int num_rows)
+{
+  void (*avx2fct) (JDIMENSION, JSAMPARRAY, JSAMPIMAGE, JDIMENSION, int);
+  void (*sse2fct) (JDIMENSION, JSAMPARRAY, JSAMPIMAGE, JDIMENSION, int);
+
+  if (simd_support == ~0U)
+    init_simd();
+
+  switch (cinfo->in_color_space) {
+  case JCS_EXT_RGB:
+    avx2fct = jsimd_extrgb_ycc_convert_avx2;
+    sse2fct = jsimd_extrgb_ycc_convert_sse2;
+    break;
+  case JCS_EXT_RGBX:
+  case JCS_EXT_RGBA:
+    avx2fct = jsimd_extrgbx_ycc_convert_avx2;
+    sse2fct = jsimd_extrgbx_ycc_convert_sse2;
+    break;
+  case JCS_EXT_BGR:
+    avx2fct = jsimd_extbgr_ycc_convert_avx2;
+    sse2fct = jsimd_extbgr_ycc_convert_sse2;
+    break;
+  case JCS_EXT_BGRX:
+  case JCS_EXT_BGRA:
+    avx2fct = jsimd_extbgrx_ycc_convert_avx2;
+    sse2fct = jsimd_extbgrx_ycc_convert_sse2;
+    break;
+  case JCS_EXT_XBGR:
+  case JCS_EXT_ABGR:
+    avx2fct = jsimd_extxbgr_ycc_convert_avx2;
+    sse2fct = jsimd_extxbgr_ycc_convert_sse2;
+    break;
+  case JCS_EXT_XRGB:
+  case JCS_EXT_ARGB:
+    avx2fct = jsimd_extxrgb_ycc_convert_avx2;
+    sse2fct = jsimd_extxrgb_ycc_convert_sse2;
+    break;
+  default:
+    avx2fct = jsimd_rgb_ycc_convert_avx2;
+    sse2fct = jsimd_rgb_ycc_convert_sse2;
+    break;
+  }
+
+  if (simd_support & JSIMD_AVX2)
+    avx2fct(cinfo->image_width, input_buf, output_buf, output_row, num_rows);
+  else
+    sse2fct(cinfo->image_width, input_buf, output_buf, output_row, num_rows);
+}
+
+GLOBAL(void)
+jsimd_rgb_gray_convert(j_compress_ptr cinfo, JSAMPARRAY input_buf,
+                       JSAMPIMAGE output_buf, JDIMENSION output_row,
+                       int num_rows)
+{
+  void (*avx2fct) (JDIMENSION, JSAMPARRAY, JSAMPIMAGE, JDIMENSION, int);
+  void (*sse2fct) (JDIMENSION, JSAMPARRAY, JSAMPIMAGE, JDIMENSION, int);
+
+  if (simd_support == ~0U)
+    init_simd();
+
+  switch (cinfo->in_color_space) {
+  case JCS_EXT_RGB:
+    avx2fct = jsimd_extrgb_gray_convert_avx2;
+    sse2fct = jsimd_extrgb_gray_convert_sse2;
+    break;
+  case JCS_EXT_RGBX:
+  case JCS_EXT_RGBA:
+    avx2fct = jsimd_extrgbx_gray_convert_avx2;
+    sse2fct = jsimd_extrgbx_gray_convert_sse2;
+    break;
+  case JCS_EXT_BGR:
+    avx2fct = jsimd_extbgr_gray_convert_avx2;
+    sse2fct = jsimd_extbgr_gray_convert_sse2;
+    break;
+  case JCS_EXT_BGRX:
+  case JCS_EXT_BGRA:
+    avx2fct = jsimd_extbgrx_gray_convert_avx2;
+    sse2fct = jsimd_extbgrx_gray_convert_sse2;
+    break;
+  case JCS_EXT_XBGR:
+  case JCS_EXT_ABGR:
+    avx2fct = jsimd_extxbgr_gray_convert_avx2;
+    sse2fct = jsimd_extxbgr_gray_convert_sse2;
+    break;
+  case JCS_EXT_XRGB:
+  case JCS_EXT_ARGB:
+    avx2fct = jsimd_extxrgb_gray_convert_avx2;
+    sse2fct = jsimd_extxrgb_gray_convert_sse2;
+    break;
+  default:
+    avx2fct = jsimd_rgb_gray_convert_avx2;
+    sse2fct = jsimd_rgb_gray_convert_sse2;
+    break;
+  }
+
+  if (simd_support & JSIMD_AVX2)
+    avx2fct(cinfo->image_width, input_buf, output_buf, output_row, num_rows);
+  else
+    sse2fct(cinfo->image_width, input_buf, output_buf, output_row, num_rows);
+}
+
+GLOBAL(void)
+jsimd_ycc_rgb_convert(j_decompress_ptr cinfo, JSAMPIMAGE input_buf,
+                      JDIMENSION input_row, JSAMPARRAY output_buf,
+                      int num_rows)
+{
+  void (*avx2fct) (JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY, int);
+  void (*sse2fct) (JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY, int);
+
+  if (simd_support == ~0U)
+    init_simd();
+
+  switch (cinfo->out_color_space) {
+  case JCS_EXT_RGB:
+    avx2fct = jsimd_ycc_extrgb_convert_avx2;
+    sse2fct = jsimd_ycc_extrgb_convert_sse2;
+    break;
+  case JCS_EXT_RGBX:
+  case JCS_EXT_RGBA:
+    avx2fct = jsimd_ycc_extrgbx_convert_avx2;
+    sse2fct = jsimd_ycc_extrgbx_convert_sse2;
+    break;
+  case JCS_EXT_BGR:
+    avx2fct = jsimd_ycc_extbgr_convert_avx2;
+    sse2fct = jsimd_ycc_extbgr_convert_sse2;
+    break;
+  case JCS_EXT_BGRX:
+  case JCS_EXT_BGRA:
+    avx2fct = jsimd_ycc_extbgrx_convert_avx2;
+    sse2fct = jsimd_ycc_extbgrx_convert_sse2;
+    break;
+  case JCS_EXT_XBGR:
+  case JCS_EXT_ABGR:
+    avx2fct = jsimd_ycc_extxbgr_convert_avx2;
+    sse2fct = jsimd_ycc_extxbgr_convert_sse2;
+    break;
+  case JCS_EXT_XRGB:
+  case JCS_EXT_ARGB:
+    avx2fct = jsimd_ycc_extxrgb_convert_avx2;
+    sse2fct = jsimd_ycc_extxrgb_convert_sse2;
+    break;
+  default:
+    avx2fct = jsimd_ycc_rgb_convert_avx2;
+    sse2fct = jsimd_ycc_rgb_convert_sse2;
+    break;
+  }
+
+  if (simd_support & JSIMD_AVX2)
+    avx2fct(cinfo->output_width, input_buf, input_row, output_buf, num_rows);
+  else
+    sse2fct(cinfo->output_width, input_buf, input_row, output_buf, num_rows);
+}
+
+GLOBAL(void)
+jsimd_ycc_rgb565_convert(j_decompress_ptr cinfo, JSAMPIMAGE input_buf,
+                         JDIMENSION input_row, JSAMPARRAY output_buf,
+                         int num_rows)
+{
+}
+
+GLOBAL(int)
+jsimd_can_h2v2_downsample(void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+
+  if (simd_support & JSIMD_AVX2)
+    return 1;
+  if (simd_support & JSIMD_SSE2)
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(int)
+jsimd_can_h2v1_downsample(void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+
+  if (simd_support & JSIMD_AVX2)
+    return 1;
+  if (simd_support & JSIMD_SSE2)
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(void)
+jsimd_h2v2_downsample(j_compress_ptr cinfo, jpeg_component_info *compptr,
+                      JSAMPARRAY input_data, JSAMPARRAY output_data)
+{
+  if (simd_support == ~0U)
+    init_simd();
+
+  if (simd_support & JSIMD_AVX2)
+    jsimd_h2v2_downsample_avx2(cinfo->image_width, cinfo->max_v_samp_factor,
+                               compptr->v_samp_factor,
+                               compptr->width_in_blocks, input_data,
+                               output_data);
+  else
+    jsimd_h2v2_downsample_sse2(cinfo->image_width, cinfo->max_v_samp_factor,
+                               compptr->v_samp_factor,
+                               compptr->width_in_blocks, input_data,
+                               output_data);
+}
+
+GLOBAL(void)
+jsimd_h2v1_downsample(j_compress_ptr cinfo, jpeg_component_info *compptr,
+                      JSAMPARRAY input_data, JSAMPARRAY output_data)
+{
+  if (simd_support == ~0U)
+    init_simd();
+
+  if (simd_support & JSIMD_AVX2)
+    jsimd_h2v1_downsample_avx2(cinfo->image_width, cinfo->max_v_samp_factor,
+                               compptr->v_samp_factor,
+                               compptr->width_in_blocks, input_data,
+                               output_data);
+  else
+    jsimd_h2v1_downsample_sse2(cinfo->image_width, cinfo->max_v_samp_factor,
+                               compptr->v_samp_factor,
+                               compptr->width_in_blocks, input_data,
+                               output_data);
+}
+
+GLOBAL(int)
+jsimd_can_h2v2_upsample(void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+
+  if (simd_support & JSIMD_AVX2)
+    return 1;
+  if (simd_support & JSIMD_SSE2)
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(int)
+jsimd_can_h2v1_upsample(void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+
+  if (simd_support & JSIMD_AVX2)
+    return 1;
+  if (simd_support & JSIMD_SSE2)
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(void)
+jsimd_h2v2_upsample(j_decompress_ptr cinfo, jpeg_component_info *compptr,
+                    JSAMPARRAY input_data, JSAMPARRAY *output_data_ptr)
+{
+  if (simd_support == ~0U)
+    init_simd();
+
+  if (simd_support & JSIMD_AVX2)
+    jsimd_h2v2_upsample_avx2(cinfo->max_v_samp_factor, cinfo->output_width,
+                             input_data, output_data_ptr);
+  else
+    jsimd_h2v2_upsample_sse2(cinfo->max_v_samp_factor, cinfo->output_width,
+                             input_data, output_data_ptr);
+}
+
+GLOBAL(void)
+jsimd_h2v1_upsample(j_decompress_ptr cinfo, jpeg_component_info *compptr,
+                    JSAMPARRAY input_data, JSAMPARRAY *output_data_ptr)
+{
+  if (simd_support == ~0U)
+    init_simd();
+
+  if (simd_support & JSIMD_AVX2)
+    jsimd_h2v1_upsample_avx2(cinfo->max_v_samp_factor, cinfo->output_width,
+                             input_data, output_data_ptr);
+  else
+    jsimd_h2v1_upsample_sse2(cinfo->max_v_samp_factor, cinfo->output_width,
+                             input_data, output_data_ptr);
+}
+
+GLOBAL(int)
+jsimd_can_h2v2_fancy_upsample(void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+
+  if ((simd_support & JSIMD_AVX2) &&
+      IS_ALIGNED_AVX(jconst_fancy_upsample_avx2))
+    return 1;
+  if ((simd_support & JSIMD_SSE2) &&
+      IS_ALIGNED_SSE(jconst_fancy_upsample_sse2))
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(int)
+jsimd_can_h2v1_fancy_upsample(void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+
+  if ((simd_support & JSIMD_AVX2) &&
+      IS_ALIGNED_AVX(jconst_fancy_upsample_avx2))
+    return 1;
+  if ((simd_support & JSIMD_SSE2) &&
+      IS_ALIGNED_SSE(jconst_fancy_upsample_sse2))
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(void)
+jsimd_h2v2_fancy_upsample(j_decompress_ptr cinfo, jpeg_component_info *compptr,
+                          JSAMPARRAY input_data, JSAMPARRAY *output_data_ptr)
+{
+  if (simd_support == ~0U)
+    init_simd();
+
+  if (simd_support & JSIMD_AVX2)
+    jsimd_h2v2_fancy_upsample_avx2(cinfo->max_v_samp_factor,
+                                   compptr->downsampled_width, input_data,
+                                   output_data_ptr);
+  else
+    jsimd_h2v2_fancy_upsample_sse2(cinfo->max_v_samp_factor,
+                                   compptr->downsampled_width, input_data,
+                                   output_data_ptr);
+}
+
+GLOBAL(void)
+jsimd_h2v1_fancy_upsample(j_decompress_ptr cinfo, jpeg_component_info *compptr,
+                          JSAMPARRAY input_data, JSAMPARRAY *output_data_ptr)
+{
+  if (simd_support == ~0U)
+    init_simd();
+
+  if (simd_support & JSIMD_AVX2)
+    jsimd_h2v1_fancy_upsample_avx2(cinfo->max_v_samp_factor,
+                                   compptr->downsampled_width, input_data,
+                                   output_data_ptr);
+  else
+    jsimd_h2v1_fancy_upsample_sse2(cinfo->max_v_samp_factor,
+                                   compptr->downsampled_width, input_data,
+                                   output_data_ptr);
+}
+
+GLOBAL(int)
+jsimd_can_h2v2_merged_upsample(void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+
+  if ((simd_support & JSIMD_AVX2) &&
+      IS_ALIGNED_AVX(jconst_merged_upsample_avx2))
+    return 1;
+  if ((simd_support & JSIMD_SSE2) &&
+      IS_ALIGNED_SSE(jconst_merged_upsample_sse2))
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(int)
+jsimd_can_h2v1_merged_upsample(void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+
+  if ((simd_support & JSIMD_AVX2) &&
+      IS_ALIGNED_AVX(jconst_merged_upsample_avx2))
+    return 1;
+  if ((simd_support & JSIMD_SSE2) &&
+      IS_ALIGNED_SSE(jconst_merged_upsample_sse2))
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(void)
+jsimd_h2v2_merged_upsample(j_decompress_ptr cinfo, JSAMPIMAGE input_buf,
+                           JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf)
+{
+  void (*avx2fct) (JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY);
+  void (*sse2fct) (JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY);
+
+  if (simd_support == ~0U)
+    init_simd();
+
+  switch (cinfo->out_color_space) {
+  case JCS_EXT_RGB:
+    avx2fct = jsimd_h2v2_extrgb_merged_upsample_avx2;
+    sse2fct = jsimd_h2v2_extrgb_merged_upsample_sse2;
+    break;
+  case JCS_EXT_RGBX:
+  case JCS_EXT_RGBA:
+    avx2fct = jsimd_h2v2_extrgbx_merged_upsample_avx2;
+    sse2fct = jsimd_h2v2_extrgbx_merged_upsample_sse2;
+    break;
+  case JCS_EXT_BGR:
+    avx2fct = jsimd_h2v2_extbgr_merged_upsample_avx2;
+    sse2fct = jsimd_h2v2_extbgr_merged_upsample_sse2;
+    break;
+  case JCS_EXT_BGRX:
+  case JCS_EXT_BGRA:
+    avx2fct = jsimd_h2v2_extbgrx_merged_upsample_avx2;
+    sse2fct = jsimd_h2v2_extbgrx_merged_upsample_sse2;
+    break;
+  case JCS_EXT_XBGR:
+  case JCS_EXT_ABGR:
+    avx2fct = jsimd_h2v2_extxbgr_merged_upsample_avx2;
+    sse2fct = jsimd_h2v2_extxbgr_merged_upsample_sse2;
+    break;
+  case JCS_EXT_XRGB:
+  case JCS_EXT_ARGB:
+    avx2fct = jsimd_h2v2_extxrgb_merged_upsample_avx2;
+    sse2fct = jsimd_h2v2_extxrgb_merged_upsample_sse2;
+    break;
+  default:
+    avx2fct = jsimd_h2v2_merged_upsample_avx2;
+    sse2fct = jsimd_h2v2_merged_upsample_sse2;
+    break;
+  }
+
+  if (simd_support & JSIMD_AVX2)
+    avx2fct(cinfo->output_width, input_buf, in_row_group_ctr, output_buf);
+  else
+    sse2fct(cinfo->output_width, input_buf, in_row_group_ctr, output_buf);
+}
+
+GLOBAL(void)
+jsimd_h2v1_merged_upsample(j_decompress_ptr cinfo, JSAMPIMAGE input_buf,
+                           JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf)
+{
+  void (*avx2fct) (JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY);
+  void (*sse2fct) (JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY);
+
+  if (simd_support == ~0U)
+    init_simd();
+
+  switch (cinfo->out_color_space) {
+  case JCS_EXT_RGB:
+    avx2fct = jsimd_h2v1_extrgb_merged_upsample_avx2;
+    sse2fct = jsimd_h2v1_extrgb_merged_upsample_sse2;
+    break;
+  case JCS_EXT_RGBX:
+  case JCS_EXT_RGBA:
+    avx2fct = jsimd_h2v1_extrgbx_merged_upsample_avx2;
+    sse2fct = jsimd_h2v1_extrgbx_merged_upsample_sse2;
+    break;
+  case JCS_EXT_BGR:
+    avx2fct = jsimd_h2v1_extbgr_merged_upsample_avx2;
+    sse2fct = jsimd_h2v1_extbgr_merged_upsample_sse2;
+    break;
+  case JCS_EXT_BGRX:
+  case JCS_EXT_BGRA:
+    avx2fct = jsimd_h2v1_extbgrx_merged_upsample_avx2;
+    sse2fct = jsimd_h2v1_extbgrx_merged_upsample_sse2;
+    break;
+  case JCS_EXT_XBGR:
+  case JCS_EXT_ABGR:
+    avx2fct = jsimd_h2v1_extxbgr_merged_upsample_avx2;
+    sse2fct = jsimd_h2v1_extxbgr_merged_upsample_sse2;
+    break;
+  case JCS_EXT_XRGB:
+  case JCS_EXT_ARGB:
+    avx2fct = jsimd_h2v1_extxrgb_merged_upsample_avx2;
+    sse2fct = jsimd_h2v1_extxrgb_merged_upsample_sse2;
+    break;
+  default:
+    avx2fct = jsimd_h2v1_merged_upsample_avx2;
+    sse2fct = jsimd_h2v1_merged_upsample_sse2;
+    break;
+  }
+
+  if (simd_support & JSIMD_AVX2)
+    avx2fct(cinfo->output_width, input_buf, in_row_group_ctr, output_buf);
+  else
+    sse2fct(cinfo->output_width, input_buf, in_row_group_ctr, output_buf);
+}
+
+GLOBAL(int)
+jsimd_can_convsamp(void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (DCTSIZE != 8)
+    return 0;
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+  if (sizeof(DCTELEM) != 2)
+    return 0;
+
+  if (simd_support & JSIMD_AVX2)
+    return 1;
+  if (simd_support & JSIMD_SSE2)
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(int)
+jsimd_can_convsamp_float(void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (DCTSIZE != 8)
+    return 0;
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+  if (sizeof(FAST_FLOAT) != 4)
+    return 0;
+
+  if (simd_support & JSIMD_SSE2)
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(void)
+jsimd_convsamp(JSAMPARRAY sample_data, JDIMENSION start_col,
+               DCTELEM *workspace)
+{
+  if (simd_support == ~0U)
+    init_simd();
+
+  if (simd_support & JSIMD_AVX2)
+    jsimd_convsamp_avx2(sample_data, start_col, workspace);
+  else
+    jsimd_convsamp_sse2(sample_data, start_col, workspace);
+}
+
+GLOBAL(void)
+jsimd_convsamp_float(JSAMPARRAY sample_data, JDIMENSION start_col,
+                     FAST_FLOAT *workspace)
+{
+  jsimd_convsamp_float_sse2(sample_data, start_col, workspace);
+}
+
+GLOBAL(int)
+jsimd_can_fdct_islow(void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (DCTSIZE != 8)
+    return 0;
+  if (sizeof(DCTELEM) != 2)
+    return 0;
+
+  if ((simd_support & JSIMD_AVX2) && IS_ALIGNED_AVX(jconst_fdct_islow_avx2))
+    return 1;
+  if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_fdct_islow_sse2))
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(int)
+jsimd_can_fdct_ifast(void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (DCTSIZE != 8)
+    return 0;
+  if (sizeof(DCTELEM) != 2)
+    return 0;
+
+  if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_fdct_ifast_sse2))
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(int)
+jsimd_can_fdct_float(void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (DCTSIZE != 8)
+    return 0;
+  if (sizeof(FAST_FLOAT) != 4)
+    return 0;
+
+  if ((simd_support & JSIMD_SSE) && IS_ALIGNED_SSE(jconst_fdct_float_sse))
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(void)
+jsimd_fdct_islow(DCTELEM *data)
+{
+  if (simd_support == ~0U)
+    init_simd();
+
+  if (simd_support & JSIMD_AVX2)
+    jsimd_fdct_islow_avx2(data);
+  else
+    jsimd_fdct_islow_sse2(data);
+}
+
+GLOBAL(void)
+jsimd_fdct_ifast(DCTELEM *data)
+{
+  jsimd_fdct_ifast_sse2(data);
+}
+
+GLOBAL(void)
+jsimd_fdct_float(FAST_FLOAT *data)
+{
+  jsimd_fdct_float_sse(data);
+}
+
+GLOBAL(int)
+jsimd_can_quantize(void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (DCTSIZE != 8)
+    return 0;
+  if (sizeof(JCOEF) != 2)
+    return 0;
+  if (sizeof(DCTELEM) != 2)
+    return 0;
+
+  if (simd_support & JSIMD_AVX2)
+    return 1;
+  if (simd_support & JSIMD_SSE2)
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(int)
+jsimd_can_quantize_float(void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (DCTSIZE != 8)
+    return 0;
+  if (sizeof(JCOEF) != 2)
+    return 0;
+  if (sizeof(FAST_FLOAT) != 4)
+    return 0;
+
+  if (simd_support & JSIMD_SSE2)
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(void)
+jsimd_quantize(JCOEFPTR coef_block, DCTELEM *divisors, DCTELEM *workspace)
+{
+  if (simd_support == ~0U)
+    init_simd();
+
+  if (simd_support & JSIMD_AVX2)
+    jsimd_quantize_avx2(coef_block, divisors, workspace);
+  else
+    jsimd_quantize_sse2(coef_block, divisors, workspace);
+}
+
+GLOBAL(void)
+jsimd_quantize_float(JCOEFPTR coef_block, FAST_FLOAT *divisors,
+                     FAST_FLOAT *workspace)
+{
+  jsimd_quantize_float_sse2(coef_block, divisors, workspace);
+}
+
+GLOBAL(int)
+jsimd_can_idct_2x2(void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (DCTSIZE != 8)
+    return 0;
+  if (sizeof(JCOEF) != 2)
+    return 0;
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+  if (sizeof(ISLOW_MULT_TYPE) != 2)
+    return 0;
+
+  if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_red_sse2))
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(int)
+jsimd_can_idct_4x4(void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (DCTSIZE != 8)
+    return 0;
+  if (sizeof(JCOEF) != 2)
+    return 0;
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+  if (sizeof(ISLOW_MULT_TYPE) != 2)
+    return 0;
+
+  if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_red_sse2))
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(void)
+jsimd_idct_2x2(j_decompress_ptr cinfo, jpeg_component_info *compptr,
+               JCOEFPTR coef_block, JSAMPARRAY output_buf,
+               JDIMENSION output_col)
+{
+  jsimd_idct_2x2_sse2(compptr->dct_table, coef_block, output_buf, output_col);
+}
+
+GLOBAL(void)
+jsimd_idct_4x4(j_decompress_ptr cinfo, jpeg_component_info *compptr,
+               JCOEFPTR coef_block, JSAMPARRAY output_buf,
+               JDIMENSION output_col)
+{
+  jsimd_idct_4x4_sse2(compptr->dct_table, coef_block, output_buf, output_col);
+}
+
+GLOBAL(int)
+jsimd_can_idct_islow(void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (DCTSIZE != 8)
+    return 0;
+  if (sizeof(JCOEF) != 2)
+    return 0;
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+  if (sizeof(ISLOW_MULT_TYPE) != 2)
+    return 0;
+
+  if ((simd_support & JSIMD_AVX2) && IS_ALIGNED_AVX(jconst_idct_islow_avx2))
+    return 1;
+  if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_islow_sse2))
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(int)
+jsimd_can_idct_ifast(void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (DCTSIZE != 8)
+    return 0;
+  if (sizeof(JCOEF) != 2)
+    return 0;
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+  if (sizeof(IFAST_MULT_TYPE) != 2)
+    return 0;
+  if (IFAST_SCALE_BITS != 2)
+    return 0;
+
+  if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_ifast_sse2))
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(int)
+jsimd_can_idct_float(void)
+{
+  init_simd();
+
+  if (DCTSIZE != 8)
+    return 0;
+  if (sizeof(JCOEF) != 2)
+    return 0;
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+  if (sizeof(FAST_FLOAT) != 4)
+    return 0;
+  if (sizeof(FLOAT_MULT_TYPE) != 4)
+    return 0;
+
+  if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_float_sse2))
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(void)
+jsimd_idct_islow(j_decompress_ptr cinfo, jpeg_component_info *compptr,
+                 JCOEFPTR coef_block, JSAMPARRAY output_buf,
+                 JDIMENSION output_col)
+{
+  if (simd_support == ~0U)
+    init_simd();
+
+  if (simd_support & JSIMD_AVX2)
+    jsimd_idct_islow_avx2(compptr->dct_table, coef_block, output_buf,
+                          output_col);
+  else
+    jsimd_idct_islow_sse2(compptr->dct_table, coef_block, output_buf,
+                          output_col);
+}
+
+GLOBAL(void)
+jsimd_idct_ifast(j_decompress_ptr cinfo, jpeg_component_info *compptr,
+                 JCOEFPTR coef_block, JSAMPARRAY output_buf,
+                 JDIMENSION output_col)
+{
+  jsimd_idct_ifast_sse2(compptr->dct_table, coef_block, output_buf,
+                        output_col);
+}
+
+GLOBAL(void)
+jsimd_idct_float(j_decompress_ptr cinfo, jpeg_component_info *compptr,
+                 JCOEFPTR coef_block, JSAMPARRAY output_buf,
+                 JDIMENSION output_col)
+{
+  jsimd_idct_float_sse2(compptr->dct_table, coef_block, output_buf,
+                        output_col);
+}
+
+GLOBAL(int)
+jsimd_can_huff_encode_one_block(void)
+{
+  init_simd();
+
+  if (DCTSIZE != 8)
+    return 0;
+  if (sizeof(JCOEF) != 2)
+    return 0;
+
+  if ((simd_support & JSIMD_SSE2) && simd_huffman &&
+      IS_ALIGNED_SSE(jconst_huff_encode_one_block))
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(JOCTET *)
+jsimd_huff_encode_one_block(void *state, JOCTET *buffer, JCOEFPTR block,
+                            int last_dc_val, c_derived_tbl *dctbl,
+                            c_derived_tbl *actbl)
+{
+  return jsimd_huff_encode_one_block_sse2(state, buffer, block, last_dc_val,
+                                          dctbl, actbl);
+}
+
+GLOBAL(int)
+jsimd_can_encode_mcu_AC_first_prepare(void)
+{
+  init_simd();
+
+  if (DCTSIZE != 8)
+    return 0;
+  if (sizeof(JCOEF) != 2)
+    return 0;
+  if (simd_support & JSIMD_SSE2)
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(void)
+jsimd_encode_mcu_AC_first_prepare(const JCOEF *block,
+                                  const int *jpeg_natural_order_start, int Sl,
+                                  int Al, UJCOEF *values, size_t *zerobits)
+{
+  jsimd_encode_mcu_AC_first_prepare_sse2(block, jpeg_natural_order_start,
+                                         Sl, Al, values, zerobits);
+}
+
+GLOBAL(int)
+jsimd_can_encode_mcu_AC_refine_prepare(void)
+{
+  init_simd();
+
+  if (DCTSIZE != 8)
+    return 0;
+  if (sizeof(JCOEF) != 2)
+    return 0;
+  if (simd_support & JSIMD_SSE2)
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(int)
+jsimd_encode_mcu_AC_refine_prepare(const JCOEF *block,
+                                   const int *jpeg_natural_order_start, int Sl,
+                                   int Al, UJCOEF *absvalues, size_t *bits)
+{
+  return jsimd_encode_mcu_AC_refine_prepare_sse2(block,
+                                                 jpeg_natural_order_start,
+                                                 Sl, Al, absvalues, bits);
+}
diff --git a/media/libjpeg/simd/x86_64/jsimdcpu.asm b/media/libjpeg/simd/x86_64/jsimdcpu.asm
new file mode 100644
index 0000000000..705f813d7d
--- /dev/null
+++ b/media/libjpeg/simd/x86_64/jsimdcpu.asm
@@ -0,0 +1,86 @@
+;
+; jsimdcpu.asm - SIMD instruction support check
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+; Copyright (C) 2016, D. R. Commander.
+;
+; Based on
+; x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+
+%include "jsimdext.inc"
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_TEXT
+    BITS        64
+;
+; Check if the CPU supports SIMD instructions
+;
+; GLOBAL(unsigned int)
+; jpeg_simd_cpu_support(void)
+;
+
+    align       32
+    GLOBAL_FUNCTION(jpeg_simd_cpu_support)
+
+EXTN(jpeg_simd_cpu_support):
+    push        rbx
+    push        rdi
+
+    xor         rdi, rdi                ; simd support flag
+
+    ; Assume that all x86-64 processors support SSE & SSE2 instructions
+    or          rdi, JSIMD_SSE2
+    or          rdi, JSIMD_SSE
+
+    ; Check whether CPUID leaf 07H is supported
+    ; (leaf 07H is used to check for AVX2 instruction support)
+    mov         rax, 0
+    cpuid
+    cmp         rax, 7
+    jl          short .return           ; Maximum leaf < 07H
+
+    ; Check for AVX2 instruction support
+    mov         rax, 7
+    xor         rcx, rcx
+    cpuid
+    mov         rax, rbx                ; rax = Extended feature flags
+
+    test        rax, 1<<5               ; bit5:AVX2
+    jz          short .return
+
+    ; Check for AVX2 O/S support
+    mov         rax, 1
+    xor         rcx, rcx
+    cpuid
+    test        rcx, 1<<27
+    jz          short .return           ; O/S does not support XSAVE
+    test        rcx, 1<<28
+    jz          short .return           ; CPU does not support AVX2
+
+    xor         rcx, rcx
+    xgetbv
+    and         rax, 6
+    cmp         rax, 6                  ; O/S does not manage XMM/YMM state
+                                        ; using XSAVE
+    jnz         short .return
+
+    or          rdi, JSIMD_AVX2
+
+.return:
+    mov         rax, rdi
+
+    pop         rdi
+    pop         rbx
+    ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+    align       32