Home | History | Annotate | Download | only in simd
      1 ;
      2 ; jidctint.asm - accurate integer IDCT (64-bit SSE2)
      3 ;
      4 ; Copyright 2009 Pierre Ossman <ossman (a] cendio.se> for Cendio AB
      5 ; Copyright (C) 2009, D. R. Commander.
      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 ; This file contains a slow-but-accurate integer implementation of the
     18 ; inverse DCT (Discrete Cosine Transform). The following code is based
     19 ; directly on the IJG's original jidctint.c; see the jidctint.c for
     20 ; more details.
     21 ;
     22 ; [TAB8]
     23 
     24 %include "jsimdext.inc"
     25 %include "jdct.inc"
     26 
     27 ; --------------------------------------------------------------------------
     28 
     29 %define CONST_BITS      13
     30 %define PASS1_BITS      2
     31 
     32 %define DESCALE_P1      (CONST_BITS-PASS1_BITS)
     33 %define DESCALE_P2      (CONST_BITS+PASS1_BITS+3)
     34 
     35 %if CONST_BITS == 13
     36 F_0_298 equ      2446           ; FIX(0.298631336)
     37 F_0_390 equ      3196           ; FIX(0.390180644)
     38 F_0_541 equ      4433           ; FIX(0.541196100)
     39 F_0_765 equ      6270           ; FIX(0.765366865)
     40 F_0_899 equ      7373           ; FIX(0.899976223)
     41 F_1_175 equ      9633           ; FIX(1.175875602)
     42 F_1_501 equ     12299           ; FIX(1.501321110)
     43 F_1_847 equ     15137           ; FIX(1.847759065)
     44 F_1_961 equ     16069           ; FIX(1.961570560)
     45 F_2_053 equ     16819           ; FIX(2.053119869)
     46 F_2_562 equ     20995           ; FIX(2.562915447)
     47 F_3_072 equ     25172           ; FIX(3.072711026)
     48 %else
     49 ; NASM cannot do compile-time arithmetic on floating-point constants.
     50 %define DESCALE(x,n)  (((x)+(1<<((n)-1)))>>(n))
     51 F_0_298 equ     DESCALE( 320652955,30-CONST_BITS)       ; FIX(0.298631336)
     52 F_0_390 equ     DESCALE( 418953276,30-CONST_BITS)       ; FIX(0.390180644)
     53 F_0_541 equ     DESCALE( 581104887,30-CONST_BITS)       ; FIX(0.541196100)
     54 F_0_765 equ     DESCALE( 821806413,30-CONST_BITS)       ; FIX(0.765366865)
     55 F_0_899 equ     DESCALE( 966342111,30-CONST_BITS)       ; FIX(0.899976223)
     56 F_1_175 equ     DESCALE(1262586813,30-CONST_BITS)       ; FIX(1.175875602)
     57 F_1_501 equ     DESCALE(1612031267,30-CONST_BITS)       ; FIX(1.501321110)
     58 F_1_847 equ     DESCALE(1984016188,30-CONST_BITS)       ; FIX(1.847759065)
     59 F_1_961 equ     DESCALE(2106220350,30-CONST_BITS)       ; FIX(1.961570560)
     60 F_2_053 equ     DESCALE(2204520673,30-CONST_BITS)       ; FIX(2.053119869)
     61 F_2_562 equ     DESCALE(2751909506,30-CONST_BITS)       ; FIX(2.562915447)
     62 F_3_072 equ     DESCALE(3299298341,30-CONST_BITS)       ; FIX(3.072711026)
     63 %endif
     64 
     65 ; --------------------------------------------------------------------------
     66         SECTION SEG_CONST
     67 
     68         alignz  16
     69         global  EXTN(jconst_idct_islow_sse2)
     70 
     71 EXTN(jconst_idct_islow_sse2):
     72 
     73 PW_F130_F054    times 4 dw  (F_0_541+F_0_765), F_0_541
     74 PW_F054_MF130   times 4 dw  F_0_541, (F_0_541-F_1_847)
     75 PW_MF078_F117   times 4 dw  (F_1_175-F_1_961), F_1_175
     76 PW_F117_F078    times 4 dw  F_1_175, (F_1_175-F_0_390)
     77 PW_MF060_MF089  times 4 dw  (F_0_298-F_0_899),-F_0_899
     78 PW_MF089_F060   times 4 dw -F_0_899, (F_1_501-F_0_899)
     79 PW_MF050_MF256  times 4 dw  (F_2_053-F_2_562),-F_2_562
     80 PW_MF256_F050   times 4 dw -F_2_562, (F_3_072-F_2_562)
     81 PD_DESCALE_P1   times 4 dd  1 << (DESCALE_P1-1)
     82 PD_DESCALE_P2   times 4 dd  1 << (DESCALE_P2-1)
     83 PB_CENTERJSAMP  times 16 db CENTERJSAMPLE
     84 
     85         alignz  16
     86 
     87 ; --------------------------------------------------------------------------
     88         SECTION SEG_TEXT
     89         BITS    64
     90 ;
     91 ; Perform dequantization and inverse DCT on one block of coefficients.
     92 ;
     93 ; GLOBAL(void)
     94 ; jsimd_idct_islow_sse2 (void *dct_table, JCOEFPTR coef_block,
     95 ;                        JSAMPARRAY output_buf, JDIMENSION output_col)
     96 ;
     97 
     98 ; r10 = jpeg_component_info *compptr
     99 ; r11 = JCOEFPTR coef_block
    100 ; r12 = JSAMPARRAY output_buf
    101 ; r13 = JDIMENSION output_col
    102 
    103 %define original_rbp    rbp+0
    104 %define wk(i)           rbp-(WK_NUM-(i))*SIZEOF_XMMWORD ; xmmword wk[WK_NUM]
    105 %define WK_NUM          12
    106 
    107         align   16
    108         global  EXTN(jsimd_idct_islow_sse2)
    109 
    110 EXTN(jsimd_idct_islow_sse2):
    111         push    rbp
    112         mov     rax,rsp                         ; rax = original rbp
    113         sub     rsp, byte 4
    114         and     rsp, byte (-SIZEOF_XMMWORD)     ; align to 128 bits
    115         mov     [rsp],rax
    116         mov     rbp,rsp                         ; rbp = aligned rbp
    117         lea     rsp, [wk(0)]
    118         collect_args
    119 
    120         ; ---- Pass 1: process columns from input.
    121 
    122         mov     rdx, r10                ; quantptr
    123         mov     rsi, r11                ; inptr
    124 
    125 %ifndef NO_ZERO_COLUMN_TEST_ISLOW_SSE2
    126         mov     eax, DWORD [DWBLOCK(1,0,rsi,SIZEOF_JCOEF)]
    127         or      eax, DWORD [DWBLOCK(2,0,rsi,SIZEOF_JCOEF)]
    128         jnz     near .columnDCT
    129 
    130         movdqa  xmm0, XMMWORD [XMMBLOCK(1,0,rsi,SIZEOF_JCOEF)]
    131         movdqa  xmm1, XMMWORD [XMMBLOCK(2,0,rsi,SIZEOF_JCOEF)]
    132         por     xmm0, XMMWORD [XMMBLOCK(3,0,rsi,SIZEOF_JCOEF)]
    133         por     xmm1, XMMWORD [XMMBLOCK(4,0,rsi,SIZEOF_JCOEF)]
    134         por     xmm0, XMMWORD [XMMBLOCK(5,0,rsi,SIZEOF_JCOEF)]
    135         por     xmm1, XMMWORD [XMMBLOCK(6,0,rsi,SIZEOF_JCOEF)]
    136         por     xmm0, XMMWORD [XMMBLOCK(7,0,rsi,SIZEOF_JCOEF)]
    137         por     xmm1,xmm0
    138         packsswb xmm1,xmm1
    139         packsswb xmm1,xmm1
    140         movd    eax,xmm1
    141         test    rax,rax
    142         jnz     short .columnDCT
    143 
    144         ; -- AC terms all zero
    145 
    146         movdqa  xmm5, XMMWORD [XMMBLOCK(0,0,rsi,SIZEOF_JCOEF)]
    147         pmullw  xmm5, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
    148 
    149         psllw   xmm5,PASS1_BITS
    150 
    151         movdqa    xmm4,xmm5             ; xmm5=in0=(00 01 02 03 04 05 06 07)
    152         punpcklwd xmm5,xmm5             ; xmm5=(00 00 01 01 02 02 03 03)
    153         punpckhwd xmm4,xmm4             ; xmm4=(04 04 05 05 06 06 07 07)
    154 
    155         pshufd  xmm7,xmm5,0x00          ; xmm7=col0=(00 00 00 00 00 00 00 00)
    156         pshufd  xmm6,xmm5,0x55          ; xmm6=col1=(01 01 01 01 01 01 01 01)
    157         pshufd  xmm1,xmm5,0xAA          ; xmm1=col2=(02 02 02 02 02 02 02 02)
    158         pshufd  xmm5,xmm5,0xFF          ; xmm5=col3=(03 03 03 03 03 03 03 03)
    159         pshufd  xmm0,xmm4,0x00          ; xmm0=col4=(04 04 04 04 04 04 04 04)
    160         pshufd  xmm3,xmm4,0x55          ; xmm3=col5=(05 05 05 05 05 05 05 05)
    161         pshufd  xmm2,xmm4,0xAA          ; xmm2=col6=(06 06 06 06 06 06 06 06)
    162         pshufd  xmm4,xmm4,0xFF          ; xmm4=col7=(07 07 07 07 07 07 07 07)
    163 
    164         movdqa  XMMWORD [wk(8)], xmm6   ; wk(8)=col1
    165         movdqa  XMMWORD [wk(9)], xmm5   ; wk(9)=col3
    166         movdqa  XMMWORD [wk(10)], xmm3  ; wk(10)=col5
    167         movdqa  XMMWORD [wk(11)], xmm4  ; wk(11)=col7
    168         jmp     near .column_end
    169 %endif
    170 .columnDCT:
    171 
    172         ; -- Even part
    173 
    174         movdqa  xmm0, XMMWORD [XMMBLOCK(0,0,rsi,SIZEOF_JCOEF)]
    175         movdqa  xmm1, XMMWORD [XMMBLOCK(2,0,rsi,SIZEOF_JCOEF)]
    176         pmullw  xmm0, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
    177         pmullw  xmm1, XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
    178         movdqa  xmm2, XMMWORD [XMMBLOCK(4,0,rsi,SIZEOF_JCOEF)]
    179         movdqa  xmm3, XMMWORD [XMMBLOCK(6,0,rsi,SIZEOF_JCOEF)]
    180         pmullw  xmm2, XMMWORD [XMMBLOCK(4,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
    181         pmullw  xmm3, XMMWORD [XMMBLOCK(6,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
    182 
    183         ; (Original)
    184         ; z1 = (z2 + z3) * 0.541196100;
    185         ; tmp2 = z1 + z3 * -1.847759065;
    186         ; tmp3 = z1 + z2 * 0.765366865;
    187         ;
    188         ; (This implementation)
    189         ; tmp2 = z2 * 0.541196100 + z3 * (0.541196100 - 1.847759065);
    190         ; tmp3 = z2 * (0.541196100 + 0.765366865) + z3 * 0.541196100;
    191 
    192         movdqa    xmm4,xmm1             ; xmm1=in2=z2
    193         movdqa    xmm5,xmm1
    194         punpcklwd xmm4,xmm3             ; xmm3=in6=z3
    195         punpckhwd xmm5,xmm3
    196         movdqa    xmm1,xmm4
    197         movdqa    xmm3,xmm5
    198         pmaddwd   xmm4,[rel PW_F130_F054]       ; xmm4=tmp3L
    199         pmaddwd   xmm5,[rel PW_F130_F054]       ; xmm5=tmp3H
    200         pmaddwd   xmm1,[rel PW_F054_MF130]      ; xmm1=tmp2L
    201         pmaddwd   xmm3,[rel PW_F054_MF130]      ; xmm3=tmp2H
    202 
    203         movdqa    xmm6,xmm0
    204         paddw     xmm0,xmm2             ; xmm0=in0+in4
    205         psubw     xmm6,xmm2             ; xmm6=in0-in4
    206 
    207         pxor      xmm7,xmm7
    208         pxor      xmm2,xmm2
    209         punpcklwd xmm7,xmm0             ; xmm7=tmp0L
    210         punpckhwd xmm2,xmm0             ; xmm2=tmp0H
    211         psrad     xmm7,(16-CONST_BITS)  ; psrad xmm7,16 & pslld xmm7,CONST_BITS
    212         psrad     xmm2,(16-CONST_BITS)  ; psrad xmm2,16 & pslld xmm2,CONST_BITS
    213 
    214         movdqa  xmm0,xmm7
    215         paddd   xmm7,xmm4               ; xmm7=tmp10L
    216         psubd   xmm0,xmm4               ; xmm0=tmp13L
    217         movdqa  xmm4,xmm2
    218         paddd   xmm2,xmm5               ; xmm2=tmp10H
    219         psubd   xmm4,xmm5               ; xmm4=tmp13H
    220 
    221         movdqa  XMMWORD [wk(0)], xmm7   ; wk(0)=tmp10L
    222         movdqa  XMMWORD [wk(1)], xmm2   ; wk(1)=tmp10H
    223         movdqa  XMMWORD [wk(2)], xmm0   ; wk(2)=tmp13L
    224         movdqa  XMMWORD [wk(3)], xmm4   ; wk(3)=tmp13H
    225 
    226         pxor      xmm5,xmm5
    227         pxor      xmm7,xmm7
    228         punpcklwd xmm5,xmm6             ; xmm5=tmp1L
    229         punpckhwd xmm7,xmm6             ; xmm7=tmp1H
    230         psrad     xmm5,(16-CONST_BITS)  ; psrad xmm5,16 & pslld xmm5,CONST_BITS
    231         psrad     xmm7,(16-CONST_BITS)  ; psrad xmm7,16 & pslld xmm7,CONST_BITS
    232 
    233         movdqa  xmm2,xmm5
    234         paddd   xmm5,xmm1               ; xmm5=tmp11L
    235         psubd   xmm2,xmm1               ; xmm2=tmp12L
    236         movdqa  xmm0,xmm7
    237         paddd   xmm7,xmm3               ; xmm7=tmp11H
    238         psubd   xmm0,xmm3               ; xmm0=tmp12H
    239 
    240         movdqa  XMMWORD [wk(4)], xmm5   ; wk(4)=tmp11L
    241         movdqa  XMMWORD [wk(5)], xmm7   ; wk(5)=tmp11H
    242         movdqa  XMMWORD [wk(6)], xmm2   ; wk(6)=tmp12L
    243         movdqa  XMMWORD [wk(7)], xmm0   ; wk(7)=tmp12H
    244 
    245         ; -- Odd part
    246 
    247         movdqa  xmm4, XMMWORD [XMMBLOCK(1,0,rsi,SIZEOF_JCOEF)]
    248         movdqa  xmm6, XMMWORD [XMMBLOCK(3,0,rsi,SIZEOF_JCOEF)]
    249         pmullw  xmm4, XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
    250         pmullw  xmm6, XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
    251         movdqa  xmm1, XMMWORD [XMMBLOCK(5,0,rsi,SIZEOF_JCOEF)]
    252         movdqa  xmm3, XMMWORD [XMMBLOCK(7,0,rsi,SIZEOF_JCOEF)]
    253         pmullw  xmm1, XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
    254         pmullw  xmm3, XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
    255 
    256         movdqa  xmm5,xmm6
    257         movdqa  xmm7,xmm4
    258         paddw   xmm5,xmm3               ; xmm5=z3
    259         paddw   xmm7,xmm1               ; xmm7=z4
    260 
    261         ; (Original)
    262         ; z5 = (z3 + z4) * 1.175875602;
    263         ; z3 = z3 * -1.961570560;  z4 = z4 * -0.390180644;
    264         ; z3 += z5;  z4 += z5;
    265         ;
    266         ; (This implementation)
    267         ; z3 = z3 * (1.175875602 - 1.961570560) + z4 * 1.175875602;
    268         ; z4 = z3 * 1.175875602 + z4 * (1.175875602 - 0.390180644);
    269 
    270         movdqa    xmm2,xmm5
    271         movdqa    xmm0,xmm5
    272         punpcklwd xmm2,xmm7
    273         punpckhwd xmm0,xmm7
    274         movdqa    xmm5,xmm2
    275         movdqa    xmm7,xmm0
    276         pmaddwd   xmm2,[rel PW_MF078_F117]      ; xmm2=z3L
    277         pmaddwd   xmm0,[rel PW_MF078_F117]      ; xmm0=z3H
    278         pmaddwd   xmm5,[rel PW_F117_F078]       ; xmm5=z4L
    279         pmaddwd   xmm7,[rel PW_F117_F078]       ; xmm7=z4H
    280 
    281         movdqa  XMMWORD [wk(10)], xmm2  ; wk(10)=z3L
    282         movdqa  XMMWORD [wk(11)], xmm0  ; wk(11)=z3H
    283 
    284         ; (Original)
    285         ; z1 = tmp0 + tmp3;  z2 = tmp1 + tmp2;
    286         ; tmp0 = tmp0 * 0.298631336;  tmp1 = tmp1 * 2.053119869;
    287         ; tmp2 = tmp2 * 3.072711026;  tmp3 = tmp3 * 1.501321110;
    288         ; z1 = z1 * -0.899976223;  z2 = z2 * -2.562915447;
    289         ; tmp0 += z1 + z3;  tmp1 += z2 + z4;
    290         ; tmp2 += z2 + z3;  tmp3 += z1 + z4;
    291         ;
    292         ; (This implementation)
    293         ; tmp0 = tmp0 * (0.298631336 - 0.899976223) + tmp3 * -0.899976223;
    294         ; tmp1 = tmp1 * (2.053119869 - 2.562915447) + tmp2 * -2.562915447;
    295         ; tmp2 = tmp1 * -2.562915447 + tmp2 * (3.072711026 - 2.562915447);
    296         ; tmp3 = tmp0 * -0.899976223 + tmp3 * (1.501321110 - 0.899976223);
    297         ; tmp0 += z3;  tmp1 += z4;
    298         ; tmp2 += z3;  tmp3 += z4;
    299 
    300         movdqa    xmm2,xmm3
    301         movdqa    xmm0,xmm3
    302         punpcklwd xmm2,xmm4
    303         punpckhwd xmm0,xmm4
    304         movdqa    xmm3,xmm2
    305         movdqa    xmm4,xmm0
    306         pmaddwd   xmm2,[rel PW_MF060_MF089]     ; xmm2=tmp0L
    307         pmaddwd   xmm0,[rel PW_MF060_MF089]     ; xmm0=tmp0H
    308         pmaddwd   xmm3,[rel PW_MF089_F060]      ; xmm3=tmp3L
    309         pmaddwd   xmm4,[rel PW_MF089_F060]      ; xmm4=tmp3H
    310 
    311         paddd   xmm2, XMMWORD [wk(10)]  ; xmm2=tmp0L
    312         paddd   xmm0, XMMWORD [wk(11)]  ; xmm0=tmp0H
    313         paddd   xmm3,xmm5               ; xmm3=tmp3L
    314         paddd   xmm4,xmm7               ; xmm4=tmp3H
    315 
    316         movdqa  XMMWORD [wk(8)], xmm2   ; wk(8)=tmp0L
    317         movdqa  XMMWORD [wk(9)], xmm0   ; wk(9)=tmp0H
    318 
    319         movdqa    xmm2,xmm1
    320         movdqa    xmm0,xmm1
    321         punpcklwd xmm2,xmm6
    322         punpckhwd xmm0,xmm6
    323         movdqa    xmm1,xmm2
    324         movdqa    xmm6,xmm0
    325         pmaddwd   xmm2,[rel PW_MF050_MF256]     ; xmm2=tmp1L
    326         pmaddwd   xmm0,[rel PW_MF050_MF256]     ; xmm0=tmp1H
    327         pmaddwd   xmm1,[rel PW_MF256_F050]      ; xmm1=tmp2L
    328         pmaddwd   xmm6,[rel PW_MF256_F050]      ; xmm6=tmp2H
    329 
    330         paddd   xmm2,xmm5               ; xmm2=tmp1L
    331         paddd   xmm0,xmm7               ; xmm0=tmp1H
    332         paddd   xmm1, XMMWORD [wk(10)]  ; xmm1=tmp2L
    333         paddd   xmm6, XMMWORD [wk(11)]  ; xmm6=tmp2H
    334 
    335         movdqa  XMMWORD [wk(10)], xmm2  ; wk(10)=tmp1L
    336         movdqa  XMMWORD [wk(11)], xmm0  ; wk(11)=tmp1H
    337 
    338         ; -- Final output stage
    339 
    340         movdqa  xmm5, XMMWORD [wk(0)]   ; xmm5=tmp10L
    341         movdqa  xmm7, XMMWORD [wk(1)]   ; xmm7=tmp10H
    342 
    343         movdqa  xmm2,xmm5
    344         movdqa  xmm0,xmm7
    345         paddd   xmm5,xmm3               ; xmm5=data0L
    346         paddd   xmm7,xmm4               ; xmm7=data0H
    347         psubd   xmm2,xmm3               ; xmm2=data7L
    348         psubd   xmm0,xmm4               ; xmm0=data7H
    349 
    350         movdqa  xmm3,[rel PD_DESCALE_P1]        ; xmm3=[rel PD_DESCALE_P1]
    351 
    352         paddd   xmm5,xmm3
    353         paddd   xmm7,xmm3
    354         psrad   xmm5,DESCALE_P1
    355         psrad   xmm7,DESCALE_P1
    356         paddd   xmm2,xmm3
    357         paddd   xmm0,xmm3
    358         psrad   xmm2,DESCALE_P1
    359         psrad   xmm0,DESCALE_P1
    360 
    361         packssdw  xmm5,xmm7             ; xmm5=data0=(00 01 02 03 04 05 06 07)
    362         packssdw  xmm2,xmm0             ; xmm2=data7=(70 71 72 73 74 75 76 77)
    363 
    364         movdqa  xmm4, XMMWORD [wk(4)]   ; xmm4=tmp11L
    365         movdqa  xmm3, XMMWORD [wk(5)]   ; xmm3=tmp11H
    366 
    367         movdqa  xmm7,xmm4
    368         movdqa  xmm0,xmm3
    369         paddd   xmm4,xmm1               ; xmm4=data1L
    370         paddd   xmm3,xmm6               ; xmm3=data1H
    371         psubd   xmm7,xmm1               ; xmm7=data6L
    372         psubd   xmm0,xmm6               ; xmm0=data6H
    373 
    374         movdqa  xmm1,[rel PD_DESCALE_P1]        ; xmm1=[rel PD_DESCALE_P1]
    375 
    376         paddd   xmm4,xmm1
    377         paddd   xmm3,xmm1
    378         psrad   xmm4,DESCALE_P1
    379         psrad   xmm3,DESCALE_P1
    380         paddd   xmm7,xmm1
    381         paddd   xmm0,xmm1
    382         psrad   xmm7,DESCALE_P1
    383         psrad   xmm0,DESCALE_P1
    384 
    385         packssdw  xmm4,xmm3             ; xmm4=data1=(10 11 12 13 14 15 16 17)
    386         packssdw  xmm7,xmm0             ; xmm7=data6=(60 61 62 63 64 65 66 67)
    387 
    388         movdqa    xmm6,xmm5             ; transpose coefficients(phase 1)
    389         punpcklwd xmm5,xmm4             ; xmm5=(00 10 01 11 02 12 03 13)
    390         punpckhwd xmm6,xmm4             ; xmm6=(04 14 05 15 06 16 07 17)
    391         movdqa    xmm1,xmm7             ; transpose coefficients(phase 1)
    392         punpcklwd xmm7,xmm2             ; xmm7=(60 70 61 71 62 72 63 73)
    393         punpckhwd xmm1,xmm2             ; xmm1=(64 74 65 75 66 76 67 77)
    394 
    395         movdqa  xmm3, XMMWORD [wk(6)]   ; xmm3=tmp12L
    396         movdqa  xmm0, XMMWORD [wk(7)]   ; xmm0=tmp12H
    397         movdqa  xmm4, XMMWORD [wk(10)]  ; xmm4=tmp1L
    398         movdqa  xmm2, XMMWORD [wk(11)]  ; xmm2=tmp1H
    399 
    400         movdqa  XMMWORD [wk(0)], xmm5   ; wk(0)=(00 10 01 11 02 12 03 13)
    401         movdqa  XMMWORD [wk(1)], xmm6   ; wk(1)=(04 14 05 15 06 16 07 17)
    402         movdqa  XMMWORD [wk(4)], xmm7   ; wk(4)=(60 70 61 71 62 72 63 73)
    403         movdqa  XMMWORD [wk(5)], xmm1   ; wk(5)=(64 74 65 75 66 76 67 77)
    404 
    405         movdqa  xmm5,xmm3
    406         movdqa  xmm6,xmm0
    407         paddd   xmm3,xmm4               ; xmm3=data2L
    408         paddd   xmm0,xmm2               ; xmm0=data2H
    409         psubd   xmm5,xmm4               ; xmm5=data5L
    410         psubd   xmm6,xmm2               ; xmm6=data5H
    411 
    412         movdqa  xmm7,[rel PD_DESCALE_P1]        ; xmm7=[rel PD_DESCALE_P1]
    413 
    414         paddd   xmm3,xmm7
    415         paddd   xmm0,xmm7
    416         psrad   xmm3,DESCALE_P1
    417         psrad   xmm0,DESCALE_P1
    418         paddd   xmm5,xmm7
    419         paddd   xmm6,xmm7
    420         psrad   xmm5,DESCALE_P1
    421         psrad   xmm6,DESCALE_P1
    422 
    423         packssdw  xmm3,xmm0             ; xmm3=data2=(20 21 22 23 24 25 26 27)
    424         packssdw  xmm5,xmm6             ; xmm5=data5=(50 51 52 53 54 55 56 57)
    425 
    426         movdqa  xmm1, XMMWORD [wk(2)]   ; xmm1=tmp13L
    427         movdqa  xmm4, XMMWORD [wk(3)]   ; xmm4=tmp13H
    428         movdqa  xmm2, XMMWORD [wk(8)]   ; xmm2=tmp0L
    429         movdqa  xmm7, XMMWORD [wk(9)]   ; xmm7=tmp0H
    430 
    431         movdqa  xmm0,xmm1
    432         movdqa  xmm6,xmm4
    433         paddd   xmm1,xmm2               ; xmm1=data3L
    434         paddd   xmm4,xmm7               ; xmm4=data3H
    435         psubd   xmm0,xmm2               ; xmm0=data4L
    436         psubd   xmm6,xmm7               ; xmm6=data4H
    437 
    438         movdqa  xmm2,[rel PD_DESCALE_P1]        ; xmm2=[rel PD_DESCALE_P1]
    439 
    440         paddd   xmm1,xmm2
    441         paddd   xmm4,xmm2
    442         psrad   xmm1,DESCALE_P1
    443         psrad   xmm4,DESCALE_P1
    444         paddd   xmm0,xmm2
    445         paddd   xmm6,xmm2
    446         psrad   xmm0,DESCALE_P1
    447         psrad   xmm6,DESCALE_P1
    448 
    449         packssdw  xmm1,xmm4             ; xmm1=data3=(30 31 32 33 34 35 36 37)
    450         packssdw  xmm0,xmm6             ; xmm0=data4=(40 41 42 43 44 45 46 47)
    451 
    452         movdqa  xmm7, XMMWORD [wk(0)]   ; xmm7=(00 10 01 11 02 12 03 13)
    453         movdqa  xmm2, XMMWORD [wk(1)]   ; xmm2=(04 14 05 15 06 16 07 17)
    454 
    455         movdqa    xmm4,xmm3             ; transpose coefficients(phase 1)
    456         punpcklwd xmm3,xmm1             ; xmm3=(20 30 21 31 22 32 23 33)
    457         punpckhwd xmm4,xmm1             ; xmm4=(24 34 25 35 26 36 27 37)
    458         movdqa    xmm6,xmm0             ; transpose coefficients(phase 1)
    459         punpcklwd xmm0,xmm5             ; xmm0=(40 50 41 51 42 52 43 53)
    460         punpckhwd xmm6,xmm5             ; xmm6=(44 54 45 55 46 56 47 57)
    461 
    462         movdqa    xmm1,xmm7             ; transpose coefficients(phase 2)
    463         punpckldq xmm7,xmm3             ; xmm7=(00 10 20 30 01 11 21 31)
    464         punpckhdq xmm1,xmm3             ; xmm1=(02 12 22 32 03 13 23 33)
    465         movdqa    xmm5,xmm2             ; transpose coefficients(phase 2)
    466         punpckldq xmm2,xmm4             ; xmm2=(04 14 24 34 05 15 25 35)
    467         punpckhdq xmm5,xmm4             ; xmm5=(06 16 26 36 07 17 27 37)
    468 
    469         movdqa  xmm3, XMMWORD [wk(4)]   ; xmm3=(60 70 61 71 62 72 63 73)
    470         movdqa  xmm4, XMMWORD [wk(5)]   ; xmm4=(64 74 65 75 66 76 67 77)
    471 
    472         movdqa  XMMWORD [wk(6)], xmm2   ; wk(6)=(04 14 24 34 05 15 25 35)
    473         movdqa  XMMWORD [wk(7)], xmm5   ; wk(7)=(06 16 26 36 07 17 27 37)
    474 
    475         movdqa    xmm2,xmm0             ; transpose coefficients(phase 2)
    476         punpckldq xmm0,xmm3             ; xmm0=(40 50 60 70 41 51 61 71)
    477         punpckhdq xmm2,xmm3             ; xmm2=(42 52 62 72 43 53 63 73)
    478         movdqa    xmm5,xmm6             ; transpose coefficients(phase 2)
    479         punpckldq xmm6,xmm4             ; xmm6=(44 54 64 74 45 55 65 75)
    480         punpckhdq xmm5,xmm4             ; xmm5=(46 56 66 76 47 57 67 77)
    481 
    482         movdqa     xmm3,xmm7            ; transpose coefficients(phase 3)
    483         punpcklqdq xmm7,xmm0            ; xmm7=col0=(00 10 20 30 40 50 60 70)
    484         punpckhqdq xmm3,xmm0            ; xmm3=col1=(01 11 21 31 41 51 61 71)
    485         movdqa     xmm4,xmm1            ; transpose coefficients(phase 3)
    486         punpcklqdq xmm1,xmm2            ; xmm1=col2=(02 12 22 32 42 52 62 72)
    487         punpckhqdq xmm4,xmm2            ; xmm4=col3=(03 13 23 33 43 53 63 73)
    488 
    489         movdqa  xmm0, XMMWORD [wk(6)]   ; xmm0=(04 14 24 34 05 15 25 35)
    490         movdqa  xmm2, XMMWORD [wk(7)]   ; xmm2=(06 16 26 36 07 17 27 37)
    491 
    492         movdqa  XMMWORD [wk(8)], xmm3   ; wk(8)=col1
    493         movdqa  XMMWORD [wk(9)], xmm4   ; wk(9)=col3
    494 
    495         movdqa     xmm3,xmm0            ; transpose coefficients(phase 3)
    496         punpcklqdq xmm0,xmm6            ; xmm0=col4=(04 14 24 34 44 54 64 74)
    497         punpckhqdq xmm3,xmm6            ; xmm3=col5=(05 15 25 35 45 55 65 75)
    498         movdqa     xmm4,xmm2            ; transpose coefficients(phase 3)
    499         punpcklqdq xmm2,xmm5            ; xmm2=col6=(06 16 26 36 46 56 66 76)
    500         punpckhqdq xmm4,xmm5            ; xmm4=col7=(07 17 27 37 47 57 67 77)
    501 
    502         movdqa  XMMWORD [wk(10)], xmm3  ; wk(10)=col5
    503         movdqa  XMMWORD [wk(11)], xmm4  ; wk(11)=col7
    504 .column_end:
    505 
    506         ; -- Prefetch the next coefficient block
    507 
    508         prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 0*32]
    509         prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 1*32]
    510         prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 2*32]
    511         prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 3*32]
    512 
    513         ; ---- Pass 2: process rows from work array, store into output array.
    514 
    515         mov     rax, [original_rbp]
    516         mov     rdi, r12        ; (JSAMPROW *)
    517         mov     eax, r13d
    518 
    519         ; -- Even part
    520 
    521         ; xmm7=col0, xmm1=col2, xmm0=col4, xmm2=col6
    522 
    523         ; (Original)
    524         ; z1 = (z2 + z3) * 0.541196100;
    525         ; tmp2 = z1 + z3 * -1.847759065;
    526         ; tmp3 = z1 + z2 * 0.765366865;
    527         ;
    528         ; (This implementation)
    529         ; tmp2 = z2 * 0.541196100 + z3 * (0.541196100 - 1.847759065);
    530         ; tmp3 = z2 * (0.541196100 + 0.765366865) + z3 * 0.541196100;
    531 
    532         movdqa    xmm6,xmm1             ; xmm1=in2=z2
    533         movdqa    xmm5,xmm1
    534         punpcklwd xmm6,xmm2             ; xmm2=in6=z3
    535         punpckhwd xmm5,xmm2
    536         movdqa    xmm1,xmm6
    537         movdqa    xmm2,xmm5
    538         pmaddwd   xmm6,[rel PW_F130_F054]       ; xmm6=tmp3L
    539         pmaddwd   xmm5,[rel PW_F130_F054]       ; xmm5=tmp3H
    540         pmaddwd   xmm1,[rel PW_F054_MF130]      ; xmm1=tmp2L
    541         pmaddwd   xmm2,[rel PW_F054_MF130]      ; xmm2=tmp2H
    542 
    543         movdqa    xmm3,xmm7
    544         paddw     xmm7,xmm0             ; xmm7=in0+in4
    545         psubw     xmm3,xmm0             ; xmm3=in0-in4
    546 
    547         pxor      xmm4,xmm4
    548         pxor      xmm0,xmm0
    549         punpcklwd xmm4,xmm7             ; xmm4=tmp0L
    550         punpckhwd xmm0,xmm7             ; xmm0=tmp0H
    551         psrad     xmm4,(16-CONST_BITS)  ; psrad xmm4,16 & pslld xmm4,CONST_BITS
    552         psrad     xmm0,(16-CONST_BITS)  ; psrad xmm0,16 & pslld xmm0,CONST_BITS
    553 
    554         movdqa  xmm7,xmm4
    555         paddd   xmm4,xmm6               ; xmm4=tmp10L
    556         psubd   xmm7,xmm6               ; xmm7=tmp13L
    557         movdqa  xmm6,xmm0
    558         paddd   xmm0,xmm5               ; xmm0=tmp10H
    559         psubd   xmm6,xmm5               ; xmm6=tmp13H
    560 
    561         movdqa  XMMWORD [wk(0)], xmm4   ; wk(0)=tmp10L
    562         movdqa  XMMWORD [wk(1)], xmm0   ; wk(1)=tmp10H
    563         movdqa  XMMWORD [wk(2)], xmm7   ; wk(2)=tmp13L
    564         movdqa  XMMWORD [wk(3)], xmm6   ; wk(3)=tmp13H
    565 
    566         pxor      xmm5,xmm5
    567         pxor      xmm4,xmm4
    568         punpcklwd xmm5,xmm3             ; xmm5=tmp1L
    569         punpckhwd xmm4,xmm3             ; xmm4=tmp1H
    570         psrad     xmm5,(16-CONST_BITS)  ; psrad xmm5,16 & pslld xmm5,CONST_BITS
    571         psrad     xmm4,(16-CONST_BITS)  ; psrad xmm4,16 & pslld xmm4,CONST_BITS
    572 
    573         movdqa  xmm0,xmm5
    574         paddd   xmm5,xmm1               ; xmm5=tmp11L
    575         psubd   xmm0,xmm1               ; xmm0=tmp12L
    576         movdqa  xmm7,xmm4
    577         paddd   xmm4,xmm2               ; xmm4=tmp11H
    578         psubd   xmm7,xmm2               ; xmm7=tmp12H
    579 
    580         movdqa  XMMWORD [wk(4)], xmm5   ; wk(4)=tmp11L
    581         movdqa  XMMWORD [wk(5)], xmm4   ; wk(5)=tmp11H
    582         movdqa  XMMWORD [wk(6)], xmm0   ; wk(6)=tmp12L
    583         movdqa  XMMWORD [wk(7)], xmm7   ; wk(7)=tmp12H
    584 
    585         ; -- Odd part
    586 
    587         movdqa  xmm6, XMMWORD [wk(9)]   ; xmm6=col3
    588         movdqa  xmm3, XMMWORD [wk(8)]   ; xmm3=col1
    589         movdqa  xmm1, XMMWORD [wk(11)]  ; xmm1=col7
    590         movdqa  xmm2, XMMWORD [wk(10)]  ; xmm2=col5
    591 
    592         movdqa  xmm5,xmm6
    593         movdqa  xmm4,xmm3
    594         paddw   xmm5,xmm1               ; xmm5=z3
    595         paddw   xmm4,xmm2               ; xmm4=z4
    596 
    597         ; (Original)
    598         ; z5 = (z3 + z4) * 1.175875602;
    599         ; z3 = z3 * -1.961570560;  z4 = z4 * -0.390180644;
    600         ; z3 += z5;  z4 += z5;
    601         ;
    602         ; (This implementation)
    603         ; z3 = z3 * (1.175875602 - 1.961570560) + z4 * 1.175875602;
    604         ; z4 = z3 * 1.175875602 + z4 * (1.175875602 - 0.390180644);
    605 
    606         movdqa    xmm0,xmm5
    607         movdqa    xmm7,xmm5
    608         punpcklwd xmm0,xmm4
    609         punpckhwd xmm7,xmm4
    610         movdqa    xmm5,xmm0
    611         movdqa    xmm4,xmm7
    612         pmaddwd   xmm0,[rel PW_MF078_F117]      ; xmm0=z3L
    613         pmaddwd   xmm7,[rel PW_MF078_F117]      ; xmm7=z3H
    614         pmaddwd   xmm5,[rel PW_F117_F078]       ; xmm5=z4L
    615         pmaddwd   xmm4,[rel PW_F117_F078]       ; xmm4=z4H
    616 
    617         movdqa  XMMWORD [wk(10)], xmm0  ; wk(10)=z3L
    618         movdqa  XMMWORD [wk(11)], xmm7  ; wk(11)=z3H
    619 
    620         ; (Original)
    621         ; z1 = tmp0 + tmp3;  z2 = tmp1 + tmp2;
    622         ; tmp0 = tmp0 * 0.298631336;  tmp1 = tmp1 * 2.053119869;
    623         ; tmp2 = tmp2 * 3.072711026;  tmp3 = tmp3 * 1.501321110;
    624         ; z1 = z1 * -0.899976223;  z2 = z2 * -2.562915447;
    625         ; tmp0 += z1 + z3;  tmp1 += z2 + z4;
    626         ; tmp2 += z2 + z3;  tmp3 += z1 + z4;
    627         ;
    628         ; (This implementation)
    629         ; tmp0 = tmp0 * (0.298631336 - 0.899976223) + tmp3 * -0.899976223;
    630         ; tmp1 = tmp1 * (2.053119869 - 2.562915447) + tmp2 * -2.562915447;
    631         ; tmp2 = tmp1 * -2.562915447 + tmp2 * (3.072711026 - 2.562915447);
    632         ; tmp3 = tmp0 * -0.899976223 + tmp3 * (1.501321110 - 0.899976223);
    633         ; tmp0 += z3;  tmp1 += z4;
    634         ; tmp2 += z3;  tmp3 += z4;
    635 
    636         movdqa    xmm0,xmm1
    637         movdqa    xmm7,xmm1
    638         punpcklwd xmm0,xmm3
    639         punpckhwd xmm7,xmm3
    640         movdqa    xmm1,xmm0
    641         movdqa    xmm3,xmm7
    642         pmaddwd   xmm0,[rel PW_MF060_MF089]     ; xmm0=tmp0L
    643         pmaddwd   xmm7,[rel PW_MF060_MF089]     ; xmm7=tmp0H
    644         pmaddwd   xmm1,[rel PW_MF089_F060]      ; xmm1=tmp3L
    645         pmaddwd   xmm3,[rel PW_MF089_F060]      ; xmm3=tmp3H
    646 
    647         paddd   xmm0, XMMWORD [wk(10)]  ; xmm0=tmp0L
    648         paddd   xmm7, XMMWORD [wk(11)]  ; xmm7=tmp0H
    649         paddd   xmm1,xmm5               ; xmm1=tmp3L
    650         paddd   xmm3,xmm4               ; xmm3=tmp3H
    651 
    652         movdqa  XMMWORD [wk(8)], xmm0   ; wk(8)=tmp0L
    653         movdqa  XMMWORD [wk(9)], xmm7   ; wk(9)=tmp0H
    654 
    655         movdqa    xmm0,xmm2
    656         movdqa    xmm7,xmm2
    657         punpcklwd xmm0,xmm6
    658         punpckhwd xmm7,xmm6
    659         movdqa    xmm2,xmm0
    660         movdqa    xmm6,xmm7
    661         pmaddwd   xmm0,[rel PW_MF050_MF256]     ; xmm0=tmp1L
    662         pmaddwd   xmm7,[rel PW_MF050_MF256]     ; xmm7=tmp1H
    663         pmaddwd   xmm2,[rel PW_MF256_F050]      ; xmm2=tmp2L
    664         pmaddwd   xmm6,[rel PW_MF256_F050]      ; xmm6=tmp2H
    665 
    666         paddd   xmm0,xmm5               ; xmm0=tmp1L
    667         paddd   xmm7,xmm4               ; xmm7=tmp1H
    668         paddd   xmm2, XMMWORD [wk(10)]  ; xmm2=tmp2L
    669         paddd   xmm6, XMMWORD [wk(11)]  ; xmm6=tmp2H
    670 
    671         movdqa  XMMWORD [wk(10)], xmm0  ; wk(10)=tmp1L
    672         movdqa  XMMWORD [wk(11)], xmm7  ; wk(11)=tmp1H
    673 
    674         ; -- Final output stage
    675 
    676         movdqa  xmm5, XMMWORD [wk(0)]   ; xmm5=tmp10L
    677         movdqa  xmm4, XMMWORD [wk(1)]   ; xmm4=tmp10H
    678 
    679         movdqa  xmm0,xmm5
    680         movdqa  xmm7,xmm4
    681         paddd   xmm5,xmm1               ; xmm5=data0L
    682         paddd   xmm4,xmm3               ; xmm4=data0H
    683         psubd   xmm0,xmm1               ; xmm0=data7L
    684         psubd   xmm7,xmm3               ; xmm7=data7H
    685 
    686         movdqa  xmm1,[rel PD_DESCALE_P2]        ; xmm1=[rel PD_DESCALE_P2]
    687 
    688         paddd   xmm5,xmm1
    689         paddd   xmm4,xmm1
    690         psrad   xmm5,DESCALE_P2
    691         psrad   xmm4,DESCALE_P2
    692         paddd   xmm0,xmm1
    693         paddd   xmm7,xmm1
    694         psrad   xmm0,DESCALE_P2
    695         psrad   xmm7,DESCALE_P2
    696 
    697         packssdw  xmm5,xmm4             ; xmm5=data0=(00 10 20 30 40 50 60 70)
    698         packssdw  xmm0,xmm7             ; xmm0=data7=(07 17 27 37 47 57 67 77)
    699 
    700         movdqa  xmm3, XMMWORD [wk(4)]   ; xmm3=tmp11L
    701         movdqa  xmm1, XMMWORD [wk(5)]   ; xmm1=tmp11H
    702 
    703         movdqa  xmm4,xmm3
    704         movdqa  xmm7,xmm1
    705         paddd   xmm3,xmm2               ; xmm3=data1L
    706         paddd   xmm1,xmm6               ; xmm1=data1H
    707         psubd   xmm4,xmm2               ; xmm4=data6L
    708         psubd   xmm7,xmm6               ; xmm7=data6H
    709 
    710         movdqa  xmm2,[rel PD_DESCALE_P2]        ; xmm2=[rel PD_DESCALE_P2]
    711 
    712         paddd   xmm3,xmm2
    713         paddd   xmm1,xmm2
    714         psrad   xmm3,DESCALE_P2
    715         psrad   xmm1,DESCALE_P2
    716         paddd   xmm4,xmm2
    717         paddd   xmm7,xmm2
    718         psrad   xmm4,DESCALE_P2
    719         psrad   xmm7,DESCALE_P2
    720 
    721         packssdw  xmm3,xmm1             ; xmm3=data1=(01 11 21 31 41 51 61 71)
    722         packssdw  xmm4,xmm7             ; xmm4=data6=(06 16 26 36 46 56 66 76)
    723 
    724         packsswb  xmm5,xmm4             ; xmm5=(00 10 20 30 40 50 60 70 06 16 26 36 46 56 66 76)
    725         packsswb  xmm3,xmm0             ; xmm3=(01 11 21 31 41 51 61 71 07 17 27 37 47 57 67 77)
    726 
    727         movdqa  xmm6, XMMWORD [wk(6)]   ; xmm6=tmp12L
    728         movdqa  xmm2, XMMWORD [wk(7)]   ; xmm2=tmp12H
    729         movdqa  xmm1, XMMWORD [wk(10)]  ; xmm1=tmp1L
    730         movdqa  xmm7, XMMWORD [wk(11)]  ; xmm7=tmp1H
    731 
    732         movdqa  XMMWORD [wk(0)], xmm5   ; wk(0)=(00 10 20 30 40 50 60 70 06 16 26 36 46 56 66 76)
    733         movdqa  XMMWORD [wk(1)], xmm3   ; wk(1)=(01 11 21 31 41 51 61 71 07 17 27 37 47 57 67 77)
    734 
    735         movdqa  xmm4,xmm6
    736         movdqa  xmm0,xmm2
    737         paddd   xmm6,xmm1               ; xmm6=data2L
    738         paddd   xmm2,xmm7               ; xmm2=data2H
    739         psubd   xmm4,xmm1               ; xmm4=data5L
    740         psubd   xmm0,xmm7               ; xmm0=data5H
    741 
    742         movdqa  xmm5,[rel PD_DESCALE_P2]        ; xmm5=[rel PD_DESCALE_P2]
    743 
    744         paddd   xmm6,xmm5
    745         paddd   xmm2,xmm5
    746         psrad   xmm6,DESCALE_P2
    747         psrad   xmm2,DESCALE_P2
    748         paddd   xmm4,xmm5
    749         paddd   xmm0,xmm5
    750         psrad   xmm4,DESCALE_P2
    751         psrad   xmm0,DESCALE_P2
    752 
    753         packssdw  xmm6,xmm2             ; xmm6=data2=(02 12 22 32 42 52 62 72)
    754         packssdw  xmm4,xmm0             ; xmm4=data5=(05 15 25 35 45 55 65 75)
    755 
    756         movdqa  xmm3, XMMWORD [wk(2)]   ; xmm3=tmp13L
    757         movdqa  xmm1, XMMWORD [wk(3)]   ; xmm1=tmp13H
    758         movdqa  xmm7, XMMWORD [wk(8)]   ; xmm7=tmp0L
    759         movdqa  xmm5, XMMWORD [wk(9)]   ; xmm5=tmp0H
    760 
    761         movdqa  xmm2,xmm3
    762         movdqa  xmm0,xmm1
    763         paddd   xmm3,xmm7               ; xmm3=data3L
    764         paddd   xmm1,xmm5               ; xmm1=data3H
    765         psubd   xmm2,xmm7               ; xmm2=data4L
    766         psubd   xmm0,xmm5               ; xmm0=data4H
    767 
    768         movdqa  xmm7,[rel PD_DESCALE_P2]        ; xmm7=[rel PD_DESCALE_P2]
    769 
    770         paddd   xmm3,xmm7
    771         paddd   xmm1,xmm7
    772         psrad   xmm3,DESCALE_P2
    773         psrad   xmm1,DESCALE_P2
    774         paddd   xmm2,xmm7
    775         paddd   xmm0,xmm7
    776         psrad   xmm2,DESCALE_P2
    777         psrad   xmm0,DESCALE_P2
    778 
    779         movdqa    xmm5,[rel PB_CENTERJSAMP]     ; xmm5=[rel PB_CENTERJSAMP]
    780 
    781         packssdw  xmm3,xmm1             ; xmm3=data3=(03 13 23 33 43 53 63 73)
    782         packssdw  xmm2,xmm0             ; xmm2=data4=(04 14 24 34 44 54 64 74)
    783 
    784         movdqa    xmm7, XMMWORD [wk(0)] ; xmm7=(00 10 20 30 40 50 60 70 06 16 26 36 46 56 66 76)
    785         movdqa    xmm1, XMMWORD [wk(1)] ; xmm1=(01 11 21 31 41 51 61 71 07 17 27 37 47 57 67 77)
    786 
    787         packsswb  xmm6,xmm2             ; xmm6=(02 12 22 32 42 52 62 72 04 14 24 34 44 54 64 74)
    788         packsswb  xmm3,xmm4             ; xmm3=(03 13 23 33 43 53 63 73 05 15 25 35 45 55 65 75)
    789 
    790         paddb     xmm7,xmm5
    791         paddb     xmm1,xmm5
    792         paddb     xmm6,xmm5
    793         paddb     xmm3,xmm5
    794 
    795         movdqa    xmm0,xmm7     ; transpose coefficients(phase 1)
    796         punpcklbw xmm7,xmm1     ; xmm7=(00 01 10 11 20 21 30 31 40 41 50 51 60 61 70 71)
    797         punpckhbw xmm0,xmm1     ; xmm0=(06 07 16 17 26 27 36 37 46 47 56 57 66 67 76 77)
    798         movdqa    xmm2,xmm6     ; transpose coefficients(phase 1)
    799         punpcklbw xmm6,xmm3     ; xmm6=(02 03 12 13 22 23 32 33 42 43 52 53 62 63 72 73)
    800         punpckhbw xmm2,xmm3     ; xmm2=(04 05 14 15 24 25 34 35 44 45 54 55 64 65 74 75)
    801 
    802         movdqa    xmm4,xmm7     ; transpose coefficients(phase 2)
    803         punpcklwd xmm7,xmm6     ; xmm7=(00 01 02 03 10 11 12 13 20 21 22 23 30 31 32 33)
    804         punpckhwd xmm4,xmm6     ; xmm4=(40 41 42 43 50 51 52 53 60 61 62 63 70 71 72 73)
    805         movdqa    xmm5,xmm2     ; transpose coefficients(phase 2)
    806         punpcklwd xmm2,xmm0     ; xmm2=(04 05 06 07 14 15 16 17 24 25 26 27 34 35 36 37)
    807         punpckhwd xmm5,xmm0     ; xmm5=(44 45 46 47 54 55 56 57 64 65 66 67 74 75 76 77)
    808 
    809         movdqa    xmm1,xmm7     ; transpose coefficients(phase 3)
    810         punpckldq xmm7,xmm2     ; xmm7=(00 01 02 03 04 05 06 07 10 11 12 13 14 15 16 17)
    811         punpckhdq xmm1,xmm2     ; xmm1=(20 21 22 23 24 25 26 27 30 31 32 33 34 35 36 37)
    812         movdqa    xmm3,xmm4     ; transpose coefficients(phase 3)
    813         punpckldq xmm4,xmm5     ; xmm4=(40 41 42 43 44 45 46 47 50 51 52 53 54 55 56 57)
    814         punpckhdq xmm3,xmm5     ; xmm3=(60 61 62 63 64 65 66 67 70 71 72 73 74 75 76 77)
    815 
    816         pshufd  xmm6,xmm7,0x4E  ; xmm6=(10 11 12 13 14 15 16 17 00 01 02 03 04 05 06 07)
    817         pshufd  xmm0,xmm1,0x4E  ; xmm0=(30 31 32 33 34 35 36 37 20 21 22 23 24 25 26 27)
    818         pshufd  xmm2,xmm4,0x4E  ; xmm2=(50 51 52 53 54 55 56 57 40 41 42 43 44 45 46 47)
    819         pshufd  xmm5,xmm3,0x4E  ; xmm5=(70 71 72 73 74 75 76 77 60 61 62 63 64 65 66 67)
    820 
    821         mov     rdx, JSAMPROW [rdi+0*SIZEOF_JSAMPROW]
    822         mov     rsi, JSAMPROW [rdi+2*SIZEOF_JSAMPROW]
    823         movq    XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm7
    824         movq    XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm1
    825         mov     rdx, JSAMPROW [rdi+4*SIZEOF_JSAMPROW]
    826         mov     rsi, JSAMPROW [rdi+6*SIZEOF_JSAMPROW]
    827         movq    XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm4
    828         movq    XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm3
    829 
    830         mov     rdx, JSAMPROW [rdi+1*SIZEOF_JSAMPROW]
    831         mov     rsi, JSAMPROW [rdi+3*SIZEOF_JSAMPROW]
    832         movq    XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm6
    833         movq    XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm0
    834         mov     rdx, JSAMPROW [rdi+5*SIZEOF_JSAMPROW]
    835         mov     rsi, JSAMPROW [rdi+7*SIZEOF_JSAMPROW]
    836         movq    XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm2
    837         movq    XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm5
    838 
    839         uncollect_args
    840         mov     rsp,rbp         ; rsp <- aligned rbp
    841         pop     rsp             ; rsp <- original rbp
    842         pop     rbp
    843         ret
    844 
    845 ; For some reason, the OS X linker does not honor the request to align the
    846 ; segment unless we do this.
    847         align   16
    848