Home | History | Annotate | Download | only in i386
      1 ;
      2 ; jcgryext.asm - grayscale colorspace conversion (AVX2)
      3 ;
      4 ; Copyright (C) 2011, 2016, D. R. Commander.
      5 ; Copyright (C) 2015, Intel Corporation.
      6 ;
      7 ; Based on the x86 SIMD extension for IJG JPEG library
      8 ; Copyright (C) 1999-2006, MIYASAKA Masaru.
      9 ; For conditions of distribution and use, see copyright notice in jsimdext.inc
     10 ;
     11 ; This file should be assembled with NASM (Netwide Assembler),
     12 ; can *not* be assembled with Microsoft's MASM or any compatible
     13 ; assembler (including Borland's Turbo Assembler).
     14 ; NASM is available from http://nasm.sourceforge.net/ or
     15 ; http://sourceforge.net/project/showfiles.php?group_id=6208
     16 ;
     17 ; [TAB8]
     18 
     19 %include "jcolsamp.inc"
     20 
     21 ; --------------------------------------------------------------------------
     22 ;
     23 ; Convert some rows of samples to the output colorspace.
     24 ;
     25 ; GLOBAL(void)
     26 ; jsimd_rgb_gray_convert_avx2(JDIMENSION img_width, JSAMPARRAY input_buf,
     27 ;                             JSAMPIMAGE output_buf, JDIMENSION output_row,
     28 ;                             int num_rows);
     29 ;
     30 
     31 %define img_width(b)   (b) + 8          ; JDIMENSION img_width
     32 %define input_buf(b)   (b) + 12         ; JSAMPARRAY input_buf
     33 %define output_buf(b)  (b) + 16         ; JSAMPIMAGE output_buf
     34 %define output_row(b)  (b) + 20         ; JDIMENSION output_row
     35 %define num_rows(b)    (b) + 24         ; int num_rows
     36 
     37 %define original_ebp   ebp + 0
     38 %define wk(i)          ebp - (WK_NUM - (i)) * SIZEOF_YMMWORD
     39                                         ; ymmword wk[WK_NUM]
     40 %define WK_NUM         2
     41 %define gotptr         wk(0) - SIZEOF_POINTER  ; void * gotptr
     42 
     43     align       32
     44     GLOBAL_FUNCTION(jsimd_rgb_gray_convert_avx2)
     45 
     46 EXTN(jsimd_rgb_gray_convert_avx2):
     47     push        ebp
     48     mov         eax, esp                     ; eax = original ebp
     49     sub         esp, byte 4
     50     and         esp, byte (-SIZEOF_YMMWORD)  ; align to 256 bits
     51     mov         [esp], eax
     52     mov         ebp, esp                     ; ebp = aligned ebp
     53     lea         esp, [wk(0)]
     54     pushpic     eax                     ; make a room for GOT address
     55     push        ebx
     56 ;   push        ecx                     ; need not be preserved
     57 ;   push        edx                     ; need not be preserved
     58     push        esi
     59     push        edi
     60 
     61     get_GOT     ebx                     ; get GOT address
     62     movpic      POINTER [gotptr], ebx   ; save GOT address
     63 
     64     mov         ecx, JDIMENSION [img_width(eax)]
     65     test        ecx, ecx
     66     jz          near .return
     67 
     68     push        ecx
     69 
     70     mov         esi, JSAMPIMAGE [output_buf(eax)]
     71     mov         ecx, JDIMENSION [output_row(eax)]
     72     mov         edi, JSAMPARRAY [esi+0*SIZEOF_JSAMPARRAY]
     73     lea         edi, [edi+ecx*SIZEOF_JSAMPROW]
     74 
     75     pop         ecx
     76 
     77     mov         esi, JSAMPARRAY [input_buf(eax)]
     78     mov         eax, INT [num_rows(eax)]
     79     test        eax, eax
     80     jle         near .return
     81     alignx      16, 7
     82 .rowloop:
     83     pushpic     eax
     84     push        edi
     85     push        esi
     86     push        ecx                     ; col
     87 
     88     mov         esi, JSAMPROW [esi]     ; inptr
     89     mov         edi, JSAMPROW [edi]     ; outptr0
     90     movpic      eax, POINTER [gotptr]   ; load GOT address (eax)
     91 
     92     cmp         ecx, byte SIZEOF_YMMWORD
     93     jae         near .columnloop
     94     alignx      16, 7
     95 
     96 %if RGB_PIXELSIZE == 3  ; ---------------
     97 
     98 .column_ld1:
     99     push        eax
    100     push        edx
    101     lea         ecx, [ecx+ecx*2]        ; imul ecx,RGB_PIXELSIZE
    102     test        cl, SIZEOF_BYTE
    103     jz          short .column_ld2
    104     sub         ecx, byte SIZEOF_BYTE
    105     movzx       eax, BYTE [esi+ecx]
    106 .column_ld2:
    107     test        cl, SIZEOF_WORD
    108     jz          short .column_ld4
    109     sub         ecx, byte SIZEOF_WORD
    110     movzx       edx, WORD [esi+ecx]
    111     shl         eax, WORD_BIT
    112     or          eax, edx
    113 .column_ld4:
    114     vmovd       xmmA, eax
    115     pop         edx
    116     pop         eax
    117     test        cl, SIZEOF_DWORD
    118     jz          short .column_ld8
    119     sub         ecx, byte SIZEOF_DWORD
    120     vmovd       xmmF, XMM_DWORD [esi+ecx]
    121     vpslldq     xmmA, xmmA, SIZEOF_DWORD
    122     vpor        xmmA, xmmA, xmmF
    123 .column_ld8:
    124     test        cl, SIZEOF_MMWORD
    125     jz          short .column_ld16
    126     sub         ecx, byte SIZEOF_MMWORD
    127     vmovq       xmmB, XMM_MMWORD [esi+ecx]
    128     vpslldq     xmmA, xmmA, SIZEOF_MMWORD
    129     vpor        xmmA, xmmA, xmmB
    130 .column_ld16:
    131     test        cl, SIZEOF_XMMWORD
    132     jz          short .column_ld32
    133     sub         ecx, byte SIZEOF_XMMWORD
    134     vmovdqu     xmmB, XMM_MMWORD [esi+ecx]
    135     vperm2i128  ymmA, ymmA, ymmA, 1
    136     vpor        ymmA, ymmB
    137 .column_ld32:
    138     test        cl, SIZEOF_YMMWORD
    139     jz          short .column_ld64
    140     sub         ecx, byte SIZEOF_YMMWORD
    141     vmovdqa     ymmF, ymmA
    142     vmovdqu     ymmA, YMMWORD [esi+0*SIZEOF_YMMWORD]
    143 .column_ld64:
    144     test        cl, 2*SIZEOF_YMMWORD
    145     mov         ecx, SIZEOF_YMMWORD
    146     jz          short .rgb_gray_cnv
    147     vmovdqa     ymmB, ymmA
    148     vmovdqu     ymmA, YMMWORD [esi+0*SIZEOF_YMMWORD]
    149     vmovdqu     ymmF, YMMWORD [esi+1*SIZEOF_YMMWORD]
    150     jmp         short .rgb_gray_cnv
    151     alignx      16, 7
    152 
    153 .columnloop:
    154     vmovdqu     ymmA, YMMWORD [esi+0*SIZEOF_YMMWORD]
    155     vmovdqu     ymmF, YMMWORD [esi+1*SIZEOF_YMMWORD]
    156     vmovdqu     ymmB, YMMWORD [esi+2*SIZEOF_YMMWORD]
    157 
    158 .rgb_gray_cnv:
    159     ; ymmA=(00 10 20 01 11 21 02 12 22 03 13 23 04 14 24 05
    160     ;       15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A)
    161     ; ymmF=(2A 0B 1B 2B 0C 1C 2C 0D 1D 2D 0E 1E 2E 0F 1F 2F
    162     ;       0G 1G 2G 0H 1H 2H 0I 1I 2I 0J 1J 2J 0K 1K 2K 0L)
    163     ; ymmB=(1L 2L 0M 1M 2M 0N 1N 2N 0O 1O 2O 0P 1P 2P 0Q 1Q
    164     ;       2Q 0R 1R 2R 0S 1S 2S 0T 1T 2T 0U 1U 2U 0V 1V 2V)
    165 
    166     vmovdqu     ymmC, ymmA
    167     vinserti128 ymmA, ymmF, xmmA, 0  ; ymmA=(00 10 20 01 11 21 02 12 22 03 13 23 04 14 24 05
    168                                      ;       0G 1G 2G 0H 1H 2H 0I 1I 2I 0J 1J 2J 0K 1K 2K 0L)
    169     vinserti128 ymmC, ymmC, xmmB, 0  ; ymmC=(1L 2L 0M 1M 2M 0N 1N 2N 0O 1O 2O 0P 1P 2P 0Q 1Q
    170                                      ;       15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A)
    171     vinserti128 ymmB, ymmB, xmmF, 0  ; ymmB=(2A 0B 1B 2B 0C 1C 2C 0D 1D 2D 0E 1E 2E 0F 1F 2F
    172                                      ;       2Q 0R 1R 2R 0S 1S 2S 0T 1T 2T 0U 1U 2U 0V 1V 2V)
    173     vperm2i128  ymmF, ymmC, ymmC, 1  ; ymmF=(15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A
    174                                      ;       1L 2L 0M 1M 2M 0N 1N 2N 0O 1O 2O 0P 1P 2P 0Q 1Q)
    175 
    176     vmovdqa     ymmG, ymmA
    177     vpslldq     ymmA, ymmA, 8     ; ymmA=(-- -- -- -- -- -- -- -- 00 10 20 01 11 21 02 12
    178                                   ;       22 03 13 23 04 14 24 05 0G 1G 2G 0H 1H 2H 0I 1I)
    179     vpsrldq     ymmG, ymmG, 8     ; ymmG=(22 03 13 23 04 14 24 05 0G 1G 2G 0H 1H 2H 0I 1I
    180                                   ;       2I 0J 1J 2J 0K 1K 2K 0L -- -- -- -- -- -- -- --)
    181 
    182     vpunpckhbw  ymmA, ymmA, ymmF  ; ymmA=(00 08 10 18 20 28 01 09 11 19 21 29 02 0A 12 1A
    183                                   ;       0G 0O 1G 1O 2G 2O 0H 0P 1H 1P 2H 2P 0I 0Q 1I 1Q)
    184     vpslldq     ymmF, ymmF, 8     ; ymmF=(-- -- -- -- -- -- -- -- 15 25 06 16 26 07 17 27
    185                                   ;       08 18 28 09 19 29 0A 1A 1L 2L 0M 1M 2M 0N 1N 2N)
    186 
    187     vpunpcklbw  ymmG, ymmG, ymmB  ; ymmG=(22 2A 03 0B 13 1B 23 2B 04 0C 14 1C 24 2C 05 0D
    188                                   ;       2I 2Q 0J 0R 1J 1R 2J 2R 0K 0S 1K 1S 2K 2S 0L 0T)
    189     vpunpckhbw  ymmF, ymmF, ymmB  ; ymmF=(15 1D 25 2D 06 0E 16 1E 26 2E 07 0F 17 1F 27 2F
    190                                   ;       1L 1T 2L 2T 0M 0U 1M 1U 2M 2U 0N 0V 1N 1V 2N 2V)
    191 
    192     vmovdqa     ymmD, ymmA
    193     vpslldq     ymmA, ymmA, 8     ; ymmA=(-- -- -- -- -- -- -- -- 00 08 10 18 20 28 01 09
    194                                   ;       11 19 21 29 02 0A 12 1A 0G 0O 1G 1O 2G 2O 0H 0P)
    195     vpsrldq     ymmD, ymmD, 8     ; ymmD=(11 19 21 29 02 0A 12 1A 0G 0O 1G 1O 2G 2O 0H 0P
    196                                   ;       1H 1P 2H 2P 0I 0Q 1I 1Q -- -- -- -- -- -- -- --)
    197 
    198     vpunpckhbw  ymmA, ymmA, ymmG  ; ymmA=(00 04 08 0C 10 14 18 1C 20 24 28 2C 01 05 09 0D
    199                                   ;       0G 0K 0O 0S 1G 1K 1O 1S 2G 2K 2O 2S 0H 0L 0P 0T)
    200     vpslldq     ymmG, ymmG, 8     ; ymmG=(-- -- -- -- -- -- -- -- 22 2A 03 0B 13 1B 23 2B
    201                                   ;       04 0C 14 1C 24 2C 05 0D 2I 2Q 0J 0R 1J 1R 2J 2R)
    202 
    203     vpunpcklbw  ymmD, ymmD, ymmF  ; ymmD=(11 15 19 1D 21 25 29 2D 02 06 0A 0E 12 16 1A 1E
    204                                   ;       1H 1L 1P 1T 2H 2L 2P 2T 0I 0M 0Q 0U 1I 1M 1Q 1U)
    205     vpunpckhbw  ymmG, ymmG, ymmF  ; ymmG=(22 26 2A 2E 03 07 0B 0F 13 17 1B 1F 23 27 2B 2F
    206                                   ;       2I 2M 2Q 2U 0J 0N 0R 0V 1J 1N 1R 1V 2J 2N 2R 2V)
    207 
    208     vmovdqa     ymmE, ymmA
    209     vpslldq     ymmA, ymmA, 8     ; ymmA=(-- -- -- -- -- -- -- -- 00 04 08 0C 10 14 18 1C
    210                                   ;       20 24 28 2C 01 05 09 0D 0G 0K 0O 0S 1G 1K 1O 1S)
    211     vpsrldq     ymmE, ymmE, 8     ; ymmE=(20 24 28 2C 01 05 09 0D 0G 0K 0O 0S 1G 1K 1O 1S
    212                                   ;       2G 2K 2O 2S 0H 0L 0P 0T -- -- -- -- -- -- -- --)
    213 
    214     vpunpckhbw  ymmA, ymmA, ymmD  ; ymmA=(00 02 04 06 08 0A 0C 0E 10 12 14 16 18 1A 1C 1E
    215                                   ;       0G 0I 0K 0M 0O 0Q 0S 0U 1G 1I 1K 1M 1O 1Q 1S 1U)
    216     vpslldq     ymmD, ymmD, 8     ; ymmD=(-- -- -- -- -- -- -- -- 11 15 19 1D 21 25 29 2D
    217                                   ;       02 06 0A 0E 12 16 1A 1E 1H 1L 1P 1T 2H 2L 2P 2T)
    218 
    219     vpunpcklbw  ymmE, ymmE, ymmG  ; ymmE=(20 22 24 26 28 2A 2C 2E 01 03 05 07 09 0B 0D 0F
    220                                   ;       2G 2I 2K 2M 2O 2Q 2S 2U 0H 0J 0L 0N 0P 0R 0T 0V)
    221     vpunpckhbw  ymmD, ymmD, ymmG  ; ymmD=(11 13 15 17 19 1B 1D 1F 21 23 25 27 29 2B 2D 2F
    222                                   ;       1H 1J 1L 1N 1P 1R 1T 1V 2H 2J 2L 2N 2P 2R 2T 2V)
    223 
    224     vpxor       ymmH, ymmH, ymmH
    225 
    226     vmovdqa     ymmC, ymmA
    227     vpunpcklbw  ymmA, ymmA, ymmH  ; ymmA=(00 02 04 06 08 0A 0C 0E 0G 0I 0K 0M 0O 0Q 0S 0U)
    228     vpunpckhbw  ymmC, ymmC, ymmH  ; ymmC=(10 12 14 16 18 1A 1C 1E 1G 1I 1K 1M 1O 1Q 1S 1U)
    229 
    230     vmovdqa     ymmB, ymmE
    231     vpunpcklbw  ymmE, ymmE, ymmH  ; ymmE=(20 22 24 26 28 2A 2C 2E 2G 2I 2K 2M 2O 2Q 2S 2U)
    232     vpunpckhbw  ymmB, ymmB, ymmH  ; ymmB=(01 03 05 07 09 0B 0D 0F 0H 0J 0L 0N 0P 0R 0T 0V)
    233 
    234     vmovdqa     ymmF, ymmD
    235     vpunpcklbw  ymmD, ymmD, ymmH  ; ymmD=(11 13 15 17 19 1B 1D 1F 1H 1J 1L 1N 1P 1R 1T 1V)
    236     vpunpckhbw  ymmF, ymmF, ymmH  ; ymmF=(21 23 25 27 29 2B 2D 2F 2H 2J 2L 2N 2P 2R 2T 2V)
    237 
    238 %else  ; RGB_PIXELSIZE == 4 ; -----------
    239 
    240 .column_ld1:
    241     test        cl, SIZEOF_XMMWORD/16
    242     jz          short .column_ld2
    243     sub         ecx, byte SIZEOF_XMMWORD/16
    244     vmovd       xmmA, XMM_DWORD [esi+ecx*RGB_PIXELSIZE]
    245 .column_ld2:
    246     test        cl, SIZEOF_XMMWORD/8
    247     jz          short .column_ld4
    248     sub         ecx, byte SIZEOF_XMMWORD/8
    249     vmovq       xmmF, XMM_MMWORD [esi+ecx*RGB_PIXELSIZE]
    250     vpslldq     xmmA, xmmA, SIZEOF_MMWORD
    251     vpor        xmmA, xmmA, xmmF
    252 .column_ld4:
    253     test        cl, SIZEOF_XMMWORD/4
    254     jz          short .column_ld8
    255     sub         ecx, byte SIZEOF_XMMWORD/4
    256     vmovdqa     xmmF, xmmA
    257     vperm2i128  ymmF, ymmF, ymmF, 1
    258     vmovdqu     xmmA, XMMWORD [esi+ecx*RGB_PIXELSIZE]
    259     vpor        ymmA, ymmA, ymmF
    260 .column_ld8:
    261     test        cl, SIZEOF_XMMWORD/2
    262     jz          short .column_ld16
    263     sub         ecx, byte SIZEOF_XMMWORD/2
    264     vmovdqa     ymmF, ymmA
    265     vmovdqu     ymmA, YMMWORD [esi+ecx*RGB_PIXELSIZE]
    266 .column_ld16:
    267     test        cl, SIZEOF_XMMWORD
    268     mov         ecx, SIZEOF_YMMWORD
    269     jz          short .rgb_gray_cnv
    270     vmovdqa     ymmE, ymmA
    271     vmovdqa     ymmH, ymmF
    272     vmovdqu     ymmA, YMMWORD [esi+0*SIZEOF_YMMWORD]
    273     vmovdqu     ymmF, YMMWORD [esi+1*SIZEOF_YMMWORD]
    274     jmp         short .rgb_gray_cnv
    275     alignx      16, 7
    276 
    277 .columnloop:
    278     vmovdqu     ymmA, YMMWORD [esi+0*SIZEOF_YMMWORD]
    279     vmovdqu     ymmF, YMMWORD [esi+1*SIZEOF_YMMWORD]
    280     vmovdqu     ymmE, YMMWORD [esi+2*SIZEOF_YMMWORD]
    281     vmovdqu     ymmH, YMMWORD [esi+3*SIZEOF_YMMWORD]
    282 
    283 .rgb_gray_cnv:
    284     ; ymmA=(00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33
    285     ;       04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37)
    286     ; ymmF=(08 18 28 38 09 19 29 39 0A 1A 2A 3A 0B 1B 2B 3B
    287     ;       0C 1C 2C 3C 0D 1D 2D 3D 0E 1E 2E 3E 0F 1F 2F 3F)
    288     ; ymmE=(0G 1G 2G 3G 0H 1H 2H 3H 0I 1I 2I 3I 0J 1J 2J 3J
    289     ;       0K 1K 2K 3K 0L 1L 2L 3L 0M 1M 2M 3M 0N 1N 2N 3N)
    290     ; ymmH=(0O 1O 2O 3O 0P 1P 2P 3P 0Q 1Q 2Q 3Q 0R 1R 2R 3R
    291     ;       0S 1S 2S 3S 0T 1T 2T 3T 0U 1U 2U 3U 0V 1V 2V 3V)
    292 
    293     vmovdqa     ymmB, ymmA
    294     vinserti128 ymmA, ymmA, xmmE, 1     ; ymmA=(00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33
    295                                         ;       0G 1G 2G 3G 0H 1H 2H 3H 0I 1I 2I 3I 0J 1J 2J 3J)
    296     vperm2i128  ymmE, ymmB, ymmE, 0x31  ; ymmE=(04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37
    297                                         ;       0K 1K 2K 3K 0L 1L 2L 3L 0M 1M 2M 3M 0N 1N 2N 3N)
    298 
    299     vmovdqa     ymmB, ymmF
    300     vinserti128 ymmF, ymmF, xmmH, 1     ; ymmF=(08 18 28 38 09 19 29 39 0A 1A 2A 3A 0B 1B 2B 3B
    301                                         ;       0O 1O 2O 3O 0P 1P 2P 3P 0Q 1Q 2Q 3Q 0R 1R 2R 3R)
    302     vperm2i128  ymmH, ymmB, ymmH, 0x31  ; ymmH=(0C 1C 2C 3C 0D 1D 2D 3D 0E 1E 2E 3E 0F 1F 2F 3F
    303                                         ;       0S 1S 2S 3S 0T 1T 2T 3T 0U 1U 2U 3U 0V 1V 2V 3V)
    304 
    305     vmovdqa     ymmD, ymmA
    306     vpunpcklbw  ymmA, ymmA, ymmE      ; ymmA=(00 04 10 14 20 24 30 34 01 05 11 15 21 25 31 35
    307                                       ;       0G 0K 1G 1K 2G 2K 3G 3K 0H 0L 1H 1L 2H 2L 3H 3L)
    308     vpunpckhbw  ymmD, ymmD, ymmE      ; ymmD=(02 06 12 16 22 26 32 36 03 07 13 17 23 27 33 37
    309                                       ;       0I 0M 1I 1M 2I 2M 3I 3M 0J 0N 1J 1N 2J 2N 3J 3N)
    310 
    311     vmovdqa     ymmC, ymmF
    312     vpunpcklbw  ymmF, ymmF, ymmH      ; ymmF=(08 0C 18 1C 28 2C 38 3C 09 0D 19 1D 29 2D 39 3D
    313                                       ;       0O 0S 1O 1S 2O 2S 3O 3S 0P 0T 1P 1T 2P 2T 3P 3T)
    314     vpunpckhbw  ymmC, ymmC, ymmH      ; ymmC=(0A 0E 1A 1E 2A 2E 3A 3E 0B 0F 1B 1F 2B 2F 3B 3F
    315                                       ;       0Q 0U 1Q 1U 2Q 2U 3Q 3U 0R 0V 1R 1V 2R 2V 3R 3V)
    316 
    317     vmovdqa     ymmB, ymmA
    318     vpunpcklwd  ymmA, ymmA, ymmF      ; ymmA=(00 04 08 0C 10 14 18 1C 20 24 28 2C 30 34 38 3C
    319                                       ;       0G 0K 0O 0S 1G 1K 1O 1S 2G 2K 2O 2S 3G 3K 3O 3S)
    320     vpunpckhwd  ymmB, ymmB, ymmF      ; ymmB=(01 05 09 0D 11 15 19 1D 21 25 29 2D 31 35 39 3D
    321                                       ;       0H 0L 0P 0T 1H 1L 1P 1T 2H 2L 2P 2T 3H 3L 3P 3T)
    322 
    323     vmovdqa     ymmG, ymmD
    324     vpunpcklwd  ymmD, ymmD, ymmC      ; ymmD=(02 06 0A 0E 12 16 1A 1E 22 26 2A 2E 32 36 3A 3E
    325                                       ;       0I 0M 0Q 0U 1I 1M 1Q 1U 2I 2M 2Q 2U 3I 3M 3Q 3U)
    326     vpunpckhwd  ymmG, ymmG, ymmC      ; ymmG=(03 07 0B 0F 13 17 1B 1F 23 27 2B 2F 33 37 3B 3F
    327                                       ;       0J 0N 0R 0V 1J 1N 1R 1V 2J 2N 2R 2V 3J 3N 3R 3V)
    328 
    329     vmovdqa     ymmE, ymmA
    330     vpunpcklbw  ymmA, ymmA, ymmD      ; ymmA=(00 02 04 06 08 0A 0C 0E 10 12 14 16 18 1A 1C 1E
    331                                       ;       0G 0I 0K 0M 0O 0Q 0S 0U 1G 1I 1K 1M 1O 1Q 1S 1U)
    332     vpunpckhbw  ymmE, ymmE, ymmD      ; ymmE=(20 22 24 26 28 2A 2C 2E 30 32 34 36 38 3A 3C 3E
    333                                       ;       2G 2I 2K 2M 2O 2Q 2S 2U 3G 3I 3K 3M 3O 3Q 3S 3U)
    334 
    335     vmovdqa     ymmH, ymmB
    336     vpunpcklbw  ymmB, ymmB, ymmG      ; ymmB=(01 03 05 07 09 0B 0D 0F 11 13 15 17 19 1B 1D 1F
    337                                       ;       0H 0J 0L 0N 0P 0R 0T 0V 1H 1J 1L 1N 1P 1R 1T 1V)
    338     vpunpckhbw  ymmH, ymmH, ymmG      ; ymmH=(21 23 25 27 29 2B 2D 2F 31 33 35 37 39 3B 3D 3F
    339                                       ;       2H 2J 2L 2N 2P 2R 2T 2V 3H 3J 3L 3N 3P 3R 3T 3V)
    340 
    341     vpxor       ymmF, ymmF, ymmF
    342 
    343     vmovdqa     ymmC, ymmA
    344     vpunpcklbw  ymmA, ymmA, ymmF      ; ymmA=(00 02 04 06 08 0A 0C 0E 0G 0I 0K 0M 0O 0Q 0S 0U)
    345     vpunpckhbw  ymmC, ymmC, ymmF      ; ymmC=(10 12 14 16 18 1A 1C 1E 1G 1I 1K 1M 1O 1Q 1S 1U)
    346 
    347     vmovdqa     ymmD, ymmB
    348     vpunpcklbw  ymmB, ymmB, ymmF      ; ymmB=(01 03 05 07 09 0B 0D 0F 0H 0J 0L 0N 0P 0R 0T 0V)
    349     vpunpckhbw  ymmD, ymmD, ymmF      ; ymmD=(11 13 15 17 19 1B 1D 1F 1H 1J 1L 1N 1P 1R 1T 1V)
    350 
    351     vmovdqa     ymmG, ymmE
    352     vpunpcklbw  ymmE, ymmE, ymmF      ; ymmE=(20 22 24 26 28 2A 2C 2E 2G 2I 2K 2M 2O 2Q 2S 2U)
    353     vpunpckhbw  ymmG, ymmG, ymmF      ; ymmG=(30 32 34 36 38 3A 3C 3E 3G 3I 3K 3M 3O 3Q 3S 3U)
    354 
    355     vpunpcklbw  ymmF, ymmF, ymmH
    356     vpunpckhbw  ymmH, ymmH, ymmH
    357     vpsrlw      ymmF, ymmF, BYTE_BIT  ; ymmF=(21 23 25 27 29 2B 2D 2F 2H 2J 2L 2N 2P 2R 2T 2V)
    358     vpsrlw      ymmH, ymmH, BYTE_BIT  ; ymmH=(31 33 35 37 39 3B 3D 3F 3H 3J 3L 3N 3P 3R 3T 3V)
    359 
    360 %endif  ; RGB_PIXELSIZE ; ---------------
    361 
    362     ; ymm0=R(02468ACEGIKMOQSU)=RE, ymm2=G(02468ACEGIKMOQSU)=GE, ymm4=B(02468ACEGIKMOQSU)=BE
    363     ; ymm1=R(13579BDFHJLNPRTV)=RO, ymm3=G(13579BDFHJLNPRTV)=GO, ymm5=B(13579BDFHJLNPRTV)=BO
    364 
    365     ; (Original)
    366     ; Y  =  0.29900 * R + 0.58700 * G + 0.11400 * B
    367     ;
    368     ; (This implementation)
    369     ; Y  =  0.29900 * R + 0.33700 * G + 0.11400 * B + 0.25000 * G
    370 
    371     vmovdqa     ymm6, ymm1
    372     vpunpcklwd  ymm1, ymm1, ymm3
    373     vpunpckhwd  ymm6, ymm6, ymm3
    374     vpmaddwd    ymm1, ymm1, [GOTOFF(eax,PW_F0299_F0337)]  ; ymm1=ROL*FIX(0.299)+GOL*FIX(0.337)
    375     vpmaddwd    ymm6, ymm6, [GOTOFF(eax,PW_F0299_F0337)]  ; ymm6=ROH*FIX(0.299)+GOH*FIX(0.337)
    376 
    377     vmovdqa     ymm7, ymm6              ; ymm7=ROH*FIX(0.299)+GOH*FIX(0.337)
    378 
    379     vmovdqa     ymm6, ymm0
    380     vpunpcklwd  ymm0, ymm0, ymm2
    381     vpunpckhwd  ymm6, ymm6, ymm2
    382     vpmaddwd    ymm0, ymm0, [GOTOFF(eax,PW_F0299_F0337)]  ; ymm0=REL*FIX(0.299)+GEL*FIX(0.337)
    383     vpmaddwd    ymm6, ymm6, [GOTOFF(eax,PW_F0299_F0337)]  ; ymm6=REH*FIX(0.299)+GEH*FIX(0.337)
    384 
    385     vmovdqa     YMMWORD [wk(0)], ymm0   ; wk(0)=REL*FIX(0.299)+GEL*FIX(0.337)
    386     vmovdqa     YMMWORD [wk(1)], ymm6   ; wk(1)=REH*FIX(0.299)+GEH*FIX(0.337)
    387 
    388     vmovdqa     ymm0, ymm5              ; ymm0=BO
    389     vmovdqa     ymm6, ymm4              ; ymm6=BE
    390 
    391     vmovdqa     ymm4, ymm0
    392     vpunpcklwd  ymm0, ymm0, ymm3
    393     vpunpckhwd  ymm4, ymm4, ymm3
    394     vpmaddwd    ymm0, ymm0, [GOTOFF(eax,PW_F0114_F0250)]  ; ymm0=BOL*FIX(0.114)+GOL*FIX(0.250)
    395     vpmaddwd    ymm4, ymm4, [GOTOFF(eax,PW_F0114_F0250)]  ; ymm4=BOH*FIX(0.114)+GOH*FIX(0.250)
    396 
    397     vmovdqa     ymm3, [GOTOFF(eax,PD_ONEHALF)]            ; ymm3=[PD_ONEHALF]
    398 
    399     vpaddd      ymm0, ymm0, ymm1
    400     vpaddd      ymm4, ymm4, ymm7
    401     vpaddd      ymm0, ymm0, ymm3
    402     vpaddd      ymm4, ymm4, ymm3
    403     vpsrld      ymm0, ymm0, SCALEBITS   ; ymm0=YOL
    404     vpsrld      ymm4, ymm4, SCALEBITS   ; ymm4=YOH
    405     vpackssdw   ymm0, ymm0, ymm4        ; ymm0=YO
    406 
    407     vmovdqa     ymm4, ymm6
    408     vpunpcklwd  ymm6, ymm6, ymm2
    409     vpunpckhwd  ymm4, ymm4, ymm2
    410     vpmaddwd    ymm6, ymm6, [GOTOFF(eax,PW_F0114_F0250)]  ; ymm6=BEL*FIX(0.114)+GEL*FIX(0.250)
    411     vpmaddwd    ymm4, ymm4, [GOTOFF(eax,PW_F0114_F0250)]  ; ymm4=BEH*FIX(0.114)+GEH*FIX(0.250)
    412 
    413     vmovdqa     ymm2, [GOTOFF(eax,PD_ONEHALF)]            ; ymm2=[PD_ONEHALF]
    414 
    415     vpaddd      ymm6, ymm6, YMMWORD [wk(0)]
    416     vpaddd      ymm4, ymm4, YMMWORD [wk(1)]
    417     vpaddd      ymm6, ymm6, ymm2
    418     vpaddd      ymm4, ymm4, ymm2
    419     vpsrld      ymm6, ymm6, SCALEBITS   ; ymm6=YEL
    420     vpsrld      ymm4, ymm4, SCALEBITS   ; ymm4=YEH
    421     vpackssdw   ymm6, ymm6, ymm4        ; ymm6=YE
    422 
    423     vpsllw      ymm0, ymm0, BYTE_BIT
    424     vpor        ymm6, ymm6, ymm0        ; ymm6=Y
    425     vmovdqu     YMMWORD [edi], ymm6     ; Save Y
    426 
    427     sub         ecx, byte SIZEOF_YMMWORD
    428     add         esi, RGB_PIXELSIZE*SIZEOF_YMMWORD  ; inptr
    429     add         edi, byte SIZEOF_YMMWORD           ; outptr0
    430     cmp         ecx, byte SIZEOF_YMMWORD
    431     jae         near .columnloop
    432     test        ecx, ecx
    433     jnz         near .column_ld1
    434 
    435     pop         ecx                     ; col
    436     pop         esi
    437     pop         edi
    438     poppic      eax
    439 
    440     add         esi, byte SIZEOF_JSAMPROW  ; input_buf
    441     add         edi, byte SIZEOF_JSAMPROW
    442     dec         eax                        ; num_rows
    443     jg          near .rowloop
    444 
    445 .return:
    446     vzeroupper
    447     pop         edi
    448     pop         esi
    449 ;   pop         edx                     ; need not be preserved
    450 ;   pop         ecx                     ; need not be preserved
    451     pop         ebx
    452     mov         esp, ebp                ; esp <- aligned ebp
    453     pop         esp                     ; esp <- original ebp
    454     pop         ebp
    455     ret
    456 
    457 ; For some reason, the OS X linker does not honor the request to align the
    458 ; segment unless we do this.
    459     align       32
    460