falkTX
/
FFmpeg
mirror of https://github.com/falkTX/FFmpeg.git
1
0
Fork 0
Code Issues 0 Releases 338 Wiki Activity
You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
36500 Commits
32 Branches
388MB
Tree: ca1fe6c0e6
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from 'ca1fe6c0e6'
${ noResults }
FFmpeg/libavcodec/x86/dsputilenc_mmx.c

1059 lines
30KB
Raw Blame History 

  1. /*

  2.  * MMX optimized DSP utils

  3.  * Copyright (c) 2000, 2001 Fabrice Bellard

  4.  * Copyright (c) 2002-2004 Michael Niedermayer <michaelni@gmx.at>

  5.  *

  6.  * This file is part of Libav.

  7.  *

  8.  * Libav is free software; you can redistribute it and/or

  9.  * modify it under the terms of the GNU Lesser General Public

  10.  * License as published by the Free Software Foundation; either

  11.  * version 2.1 of the License, or (at your option) any later version.

  12.  *

  13.  * Libav is distributed in the hope that it will be useful,

  14.  * but WITHOUT ANY WARRANTY; without even the implied warranty of

  15.  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU

  16.  * Lesser General Public License for more details.

  17.  *

  18.  * You should have received a copy of the GNU Lesser General Public

  19.  * License along with Libav; if not, write to the Free Software

  20.  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA

  21.  *

  22.  * MMX optimization by Nick Kurshev <nickols_k@mail.ru>

  23.  */

  24. 

  25. #include "libavutil/attributes.h"

  26. #include "libavutil/cpu.h"

  27. #include "libavutil/x86/asm.h"

  28. #include "libavutil/x86/cpu.h"

  29. #include "libavcodec/dsputil.h"

  30. #include "libavcodec/mpegvideo.h"

  31. #include "libavcodec/mathops.h"

  32. #include "dsputil_mmx.h"

  33. 

  34. void ff_get_pixels_mmx(int16_t *block, const uint8_t *pixels, int line_size);

  35. void ff_get_pixels_sse2(int16_t *block, const uint8_t *pixels, int line_size);

  36. void ff_diff_pixels_mmx(int16_t *block, const uint8_t *s1, const uint8_t *s2, int stride);

  37. int ff_pix_sum16_mmx(uint8_t * pix, int line_size);

  38. int ff_pix_norm1_mmx(uint8_t *pix, int line_size);

  39. 

  40. #if HAVE_INLINE_ASM

  41. 

  42. static int sse8_mmx(void *v, uint8_t * pix1, uint8_t * pix2, int line_size, int h) {

  43.     int tmp;

  44.   __asm__ volatile (

  45.       "movl %4,%%ecx\n"

  46.       "shr $1,%%ecx\n"

  47.       "pxor %%mm0,%%mm0\n"      /* mm0 = 0 */

  48.       "pxor %%mm7,%%mm7\n"      /* mm7 holds the sum */

  49.       "1:\n"

  50.       "movq (%0),%%mm1\n"       /* mm1 = pix1[0][0-7] */

  51.       "movq (%1),%%mm2\n"       /* mm2 = pix2[0][0-7] */

  52.       "movq (%0,%3),%%mm3\n"    /* mm3 = pix1[1][0-7] */

  53.       "movq (%1,%3),%%mm4\n"    /* mm4 = pix2[1][0-7] */

  54. 

  55.       /* todo: mm1-mm2, mm3-mm4 */

  56.       /* algo: subtract mm1 from mm2 with saturation and vice versa */

  57.       /*       OR the results to get absolute difference */

  58.       "movq %%mm1,%%mm5\n"

  59.       "movq %%mm3,%%mm6\n"

  60.       "psubusb %%mm2,%%mm1\n"

  61.       "psubusb %%mm4,%%mm3\n"

  62.       "psubusb %%mm5,%%mm2\n"

  63.       "psubusb %%mm6,%%mm4\n"

  64. 

  65.       "por %%mm1,%%mm2\n"

  66.       "por %%mm3,%%mm4\n"

  67. 

  68.       /* now convert to 16-bit vectors so we can square them */

  69.       "movq %%mm2,%%mm1\n"

  70.       "movq %%mm4,%%mm3\n"

  71. 

  72.       "punpckhbw %%mm0,%%mm2\n"

  73.       "punpckhbw %%mm0,%%mm4\n"

  74.       "punpcklbw %%mm0,%%mm1\n" /* mm1 now spread over (mm1,mm2) */

  75.       "punpcklbw %%mm0,%%mm3\n" /* mm4 now spread over (mm3,mm4) */

  76. 

  77.       "pmaddwd %%mm2,%%mm2\n"

  78.       "pmaddwd %%mm4,%%mm4\n"

  79.       "pmaddwd %%mm1,%%mm1\n"

  80.       "pmaddwd %%mm3,%%mm3\n"

  81. 

  82.       "lea (%0,%3,2), %0\n"     /* pix1 += 2*line_size */

  83.       "lea (%1,%3,2), %1\n"     /* pix2 += 2*line_size */

  84. 

  85.       "paddd %%mm2,%%mm1\n"

  86.       "paddd %%mm4,%%mm3\n"

  87.       "paddd %%mm1,%%mm7\n"

  88.       "paddd %%mm3,%%mm7\n"

  89. 

  90.       "decl %%ecx\n"

  91.       "jnz 1b\n"

  92. 

  93.       "movq %%mm7,%%mm1\n"

  94.       "psrlq $32, %%mm7\n"      /* shift hi dword to lo */

  95.       "paddd %%mm7,%%mm1\n"

  96.       "movd %%mm1,%2\n"

  97.       : "+r" (pix1), "+r" (pix2), "=r"(tmp)

  98.       : "r" ((x86_reg)line_size) , "m" (h)

  99.       : "%ecx");

  100.     return tmp;

  101. }

  102. 

  103. static int sse16_mmx(void *v, uint8_t * pix1, uint8_t * pix2, int line_size, int h) {

  104.     int tmp;

  105.   __asm__ volatile (

  106.       "movl %4,%%ecx\n"

  107.       "pxor %%mm0,%%mm0\n"      /* mm0 = 0 */

  108.       "pxor %%mm7,%%mm7\n"      /* mm7 holds the sum */

  109.       "1:\n"

  110.       "movq (%0),%%mm1\n"       /* mm1 = pix1[0-7] */

  111.       "movq (%1),%%mm2\n"       /* mm2 = pix2[0-7] */

  112.       "movq 8(%0),%%mm3\n"      /* mm3 = pix1[8-15] */

  113.       "movq 8(%1),%%mm4\n"      /* mm4 = pix2[8-15] */

  114. 

  115.       /* todo: mm1-mm2, mm3-mm4 */

  116.       /* algo: subtract mm1 from mm2 with saturation and vice versa */

  117.       /*       OR the results to get absolute difference */

  118.       "movq %%mm1,%%mm5\n"

  119.       "movq %%mm3,%%mm6\n"

  120.       "psubusb %%mm2,%%mm1\n"

  121.       "psubusb %%mm4,%%mm3\n"

  122.       "psubusb %%mm5,%%mm2\n"

  123.       "psubusb %%mm6,%%mm4\n"

  124. 

  125.       "por %%mm1,%%mm2\n"

  126.       "por %%mm3,%%mm4\n"

  127. 

  128.       /* now convert to 16-bit vectors so we can square them */

  129.       "movq %%mm2,%%mm1\n"

  130.       "movq %%mm4,%%mm3\n"

  131. 

  132.       "punpckhbw %%mm0,%%mm2\n"

  133.       "punpckhbw %%mm0,%%mm4\n"

  134.       "punpcklbw %%mm0,%%mm1\n" /* mm1 now spread over (mm1,mm2) */

  135.       "punpcklbw %%mm0,%%mm3\n" /* mm4 now spread over (mm3,mm4) */

  136. 

  137.       "pmaddwd %%mm2,%%mm2\n"

  138.       "pmaddwd %%mm4,%%mm4\n"

  139.       "pmaddwd %%mm1,%%mm1\n"

  140.       "pmaddwd %%mm3,%%mm3\n"

  141. 

  142.       "add %3,%0\n"

  143.       "add %3,%1\n"

  144. 

  145.       "paddd %%mm2,%%mm1\n"

  146.       "paddd %%mm4,%%mm3\n"

  147.       "paddd %%mm1,%%mm7\n"

  148.       "paddd %%mm3,%%mm7\n"

  149. 

  150.       "decl %%ecx\n"

  151.       "jnz 1b\n"

  152. 

  153.       "movq %%mm7,%%mm1\n"

  154.       "psrlq $32, %%mm7\n"      /* shift hi dword to lo */

  155.       "paddd %%mm7,%%mm1\n"

  156.       "movd %%mm1,%2\n"

  157.       : "+r" (pix1), "+r" (pix2), "=r"(tmp)

  158.       : "r" ((x86_reg)line_size) , "m" (h)

  159.       : "%ecx");

  160.     return tmp;

  161. }

  162. 

  163. static int hf_noise8_mmx(uint8_t * pix1, int line_size, int h) {

  164.     int tmp;

  165.   __asm__ volatile (

  166.       "movl %3,%%ecx\n"

  167.       "pxor %%mm7,%%mm7\n"

  168.       "pxor %%mm6,%%mm6\n"

  169. 

  170.       "movq (%0),%%mm0\n"

  171.       "movq %%mm0, %%mm1\n"

  172.       "psllq $8, %%mm0\n"

  173.       "psrlq $8, %%mm1\n"

  174.       "psrlq $8, %%mm0\n"

  175.       "movq %%mm0, %%mm2\n"

  176.       "movq %%mm1, %%mm3\n"

  177.       "punpcklbw %%mm7,%%mm0\n"

  178.       "punpcklbw %%mm7,%%mm1\n"

  179.       "punpckhbw %%mm7,%%mm2\n"

  180.       "punpckhbw %%mm7,%%mm3\n"

  181.       "psubw %%mm1, %%mm0\n"

  182.       "psubw %%mm3, %%mm2\n"

  183. 

  184.       "add %2,%0\n"

  185. 

  186.       "movq (%0),%%mm4\n"

  187.       "movq %%mm4, %%mm1\n"

  188.       "psllq $8, %%mm4\n"

  189.       "psrlq $8, %%mm1\n"

  190.       "psrlq $8, %%mm4\n"

  191.       "movq %%mm4, %%mm5\n"

  192.       "movq %%mm1, %%mm3\n"

  193.       "punpcklbw %%mm7,%%mm4\n"

  194.       "punpcklbw %%mm7,%%mm1\n"

  195.       "punpckhbw %%mm7,%%mm5\n"

  196.       "punpckhbw %%mm7,%%mm3\n"

  197.       "psubw %%mm1, %%mm4\n"

  198.       "psubw %%mm3, %%mm5\n"

  199.       "psubw %%mm4, %%mm0\n"

  200.       "psubw %%mm5, %%mm2\n"

  201.       "pxor %%mm3, %%mm3\n"

  202.       "pxor %%mm1, %%mm1\n"

  203.       "pcmpgtw %%mm0, %%mm3\n\t"

  204.       "pcmpgtw %%mm2, %%mm1\n\t"

  205.       "pxor %%mm3, %%mm0\n"

  206.       "pxor %%mm1, %%mm2\n"

  207.       "psubw %%mm3, %%mm0\n"

  208.       "psubw %%mm1, %%mm2\n"

  209.       "paddw %%mm0, %%mm2\n"

  210.       "paddw %%mm2, %%mm6\n"

  211. 

  212.       "add %2,%0\n"

  213.       "1:\n"

  214. 

  215.       "movq (%0),%%mm0\n"

  216.       "movq %%mm0, %%mm1\n"

  217.       "psllq $8, %%mm0\n"

  218.       "psrlq $8, %%mm1\n"

  219.       "psrlq $8, %%mm0\n"

  220.       "movq %%mm0, %%mm2\n"

  221.       "movq %%mm1, %%mm3\n"

  222.       "punpcklbw %%mm7,%%mm0\n"

  223.       "punpcklbw %%mm7,%%mm1\n"

  224.       "punpckhbw %%mm7,%%mm2\n"

  225.       "punpckhbw %%mm7,%%mm3\n"

  226.       "psubw %%mm1, %%mm0\n"

  227.       "psubw %%mm3, %%mm2\n"

  228.       "psubw %%mm0, %%mm4\n"

  229.       "psubw %%mm2, %%mm5\n"

  230.       "pxor %%mm3, %%mm3\n"

  231.       "pxor %%mm1, %%mm1\n"

  232.       "pcmpgtw %%mm4, %%mm3\n\t"

  233.       "pcmpgtw %%mm5, %%mm1\n\t"

  234.       "pxor %%mm3, %%mm4\n"

  235.       "pxor %%mm1, %%mm5\n"

  236.       "psubw %%mm3, %%mm4\n"

  237.       "psubw %%mm1, %%mm5\n"

  238.       "paddw %%mm4, %%mm5\n"

  239.       "paddw %%mm5, %%mm6\n"

  240. 

  241.       "add %2,%0\n"

  242. 

  243.       "movq (%0),%%mm4\n"

  244.       "movq %%mm4, %%mm1\n"

  245.       "psllq $8, %%mm4\n"

  246.       "psrlq $8, %%mm1\n"

  247.       "psrlq $8, %%mm4\n"

  248.       "movq %%mm4, %%mm5\n"

  249.       "movq %%mm1, %%mm3\n"

  250.       "punpcklbw %%mm7,%%mm4\n"

  251.       "punpcklbw %%mm7,%%mm1\n"

  252.       "punpckhbw %%mm7,%%mm5\n"

  253.       "punpckhbw %%mm7,%%mm3\n"

  254.       "psubw %%mm1, %%mm4\n"

  255.       "psubw %%mm3, %%mm5\n"

  256.       "psubw %%mm4, %%mm0\n"

  257.       "psubw %%mm5, %%mm2\n"

  258.       "pxor %%mm3, %%mm3\n"

  259.       "pxor %%mm1, %%mm1\n"

  260.       "pcmpgtw %%mm0, %%mm3\n\t"

  261.       "pcmpgtw %%mm2, %%mm1\n\t"

  262.       "pxor %%mm3, %%mm0\n"

  263.       "pxor %%mm1, %%mm2\n"

  264.       "psubw %%mm3, %%mm0\n"

  265.       "psubw %%mm1, %%mm2\n"

  266.       "paddw %%mm0, %%mm2\n"

  267.       "paddw %%mm2, %%mm6\n"

  268. 

  269.       "add %2,%0\n"

  270.       "subl $2, %%ecx\n"

  271.       " jnz 1b\n"

  272. 

  273.       "movq %%mm6, %%mm0\n"

  274.       "punpcklwd %%mm7,%%mm0\n"

  275.       "punpckhwd %%mm7,%%mm6\n"

  276.       "paddd %%mm0, %%mm6\n"

  277. 

  278.       "movq %%mm6,%%mm0\n"

  279.       "psrlq $32, %%mm6\n"

  280.       "paddd %%mm6,%%mm0\n"

  281.       "movd %%mm0,%1\n"

  282.       : "+r" (pix1), "=r"(tmp)

  283.       : "r" ((x86_reg)line_size) , "g" (h-2)

  284.       : "%ecx");

  285.       return tmp;

  286. }

  287. 

  288. static int hf_noise16_mmx(uint8_t * pix1, int line_size, int h) {

  289.     int tmp;

  290.     uint8_t * pix= pix1;

  291.   __asm__ volatile (

  292.       "movl %3,%%ecx\n"

  293.       "pxor %%mm7,%%mm7\n"

  294.       "pxor %%mm6,%%mm6\n"

  295. 

  296.       "movq (%0),%%mm0\n"

  297.       "movq 1(%0),%%mm1\n"

  298.       "movq %%mm0, %%mm2\n"

  299.       "movq %%mm1, %%mm3\n"

  300.       "punpcklbw %%mm7,%%mm0\n"

  301.       "punpcklbw %%mm7,%%mm1\n"

  302.       "punpckhbw %%mm7,%%mm2\n"

  303.       "punpckhbw %%mm7,%%mm3\n"

  304.       "psubw %%mm1, %%mm0\n"

  305.       "psubw %%mm3, %%mm2\n"

  306. 

  307.       "add %2,%0\n"

  308. 

  309.       "movq (%0),%%mm4\n"

  310.       "movq 1(%0),%%mm1\n"

  311.       "movq %%mm4, %%mm5\n"

  312.       "movq %%mm1, %%mm3\n"

  313.       "punpcklbw %%mm7,%%mm4\n"

  314.       "punpcklbw %%mm7,%%mm1\n"

  315.       "punpckhbw %%mm7,%%mm5\n"

  316.       "punpckhbw %%mm7,%%mm3\n"

  317.       "psubw %%mm1, %%mm4\n"

  318.       "psubw %%mm3, %%mm5\n"

  319.       "psubw %%mm4, %%mm0\n"

  320.       "psubw %%mm5, %%mm2\n"

  321.       "pxor %%mm3, %%mm3\n"

  322.       "pxor %%mm1, %%mm1\n"

  323.       "pcmpgtw %%mm0, %%mm3\n\t"

  324.       "pcmpgtw %%mm2, %%mm1\n\t"

  325.       "pxor %%mm3, %%mm0\n"

  326.       "pxor %%mm1, %%mm2\n"

  327.       "psubw %%mm3, %%mm0\n"

  328.       "psubw %%mm1, %%mm2\n"

  329.       "paddw %%mm0, %%mm2\n"

  330.       "paddw %%mm2, %%mm6\n"

  331. 

  332.       "add %2,%0\n"

  333.       "1:\n"

  334. 

  335.       "movq (%0),%%mm0\n"

  336.       "movq 1(%0),%%mm1\n"

  337.       "movq %%mm0, %%mm2\n"

  338.       "movq %%mm1, %%mm3\n"

  339.       "punpcklbw %%mm7,%%mm0\n"

  340.       "punpcklbw %%mm7,%%mm1\n"

  341.       "punpckhbw %%mm7,%%mm2\n"

  342.       "punpckhbw %%mm7,%%mm3\n"

  343.       "psubw %%mm1, %%mm0\n"

  344.       "psubw %%mm3, %%mm2\n"

  345.       "psubw %%mm0, %%mm4\n"

  346.       "psubw %%mm2, %%mm5\n"

  347.       "pxor %%mm3, %%mm3\n"

  348.       "pxor %%mm1, %%mm1\n"

  349.       "pcmpgtw %%mm4, %%mm3\n\t"

  350.       "pcmpgtw %%mm5, %%mm1\n\t"

  351.       "pxor %%mm3, %%mm4\n"

  352.       "pxor %%mm1, %%mm5\n"

  353.       "psubw %%mm3, %%mm4\n"

  354.       "psubw %%mm1, %%mm5\n"

  355.       "paddw %%mm4, %%mm5\n"

  356.       "paddw %%mm5, %%mm6\n"

  357. 

  358.       "add %2,%0\n"

  359. 

  360.       "movq (%0),%%mm4\n"

  361.       "movq 1(%0),%%mm1\n"

  362.       "movq %%mm4, %%mm5\n"

  363.       "movq %%mm1, %%mm3\n"

  364.       "punpcklbw %%mm7,%%mm4\n"

  365.       "punpcklbw %%mm7,%%mm1\n"

  366.       "punpckhbw %%mm7,%%mm5\n"

  367.       "punpckhbw %%mm7,%%mm3\n"

  368.       "psubw %%mm1, %%mm4\n"

  369.       "psubw %%mm3, %%mm5\n"

  370.       "psubw %%mm4, %%mm0\n"

  371.       "psubw %%mm5, %%mm2\n"

  372.       "pxor %%mm3, %%mm3\n"

  373.       "pxor %%mm1, %%mm1\n"

  374.       "pcmpgtw %%mm0, %%mm3\n\t"

  375.       "pcmpgtw %%mm2, %%mm1\n\t"

  376.       "pxor %%mm3, %%mm0\n"

  377.       "pxor %%mm1, %%mm2\n"

  378.       "psubw %%mm3, %%mm0\n"

  379.       "psubw %%mm1, %%mm2\n"

  380.       "paddw %%mm0, %%mm2\n"

  381.       "paddw %%mm2, %%mm6\n"

  382. 

  383.       "add %2,%0\n"

  384.       "subl $2, %%ecx\n"

  385.       " jnz 1b\n"

  386. 

  387.       "movq %%mm6, %%mm0\n"

  388.       "punpcklwd %%mm7,%%mm0\n"

  389.       "punpckhwd %%mm7,%%mm6\n"

  390.       "paddd %%mm0, %%mm6\n"

  391. 

  392.       "movq %%mm6,%%mm0\n"

  393.       "psrlq $32, %%mm6\n"

  394.       "paddd %%mm6,%%mm0\n"

  395.       "movd %%mm0,%1\n"

  396.       : "+r" (pix1), "=r"(tmp)

  397.       : "r" ((x86_reg)line_size) , "g" (h-2)

  398.       : "%ecx");

  399.       return tmp + hf_noise8_mmx(pix+8, line_size, h);

  400. }

  401. 

  402. static int nsse16_mmx(void *p, uint8_t * pix1, uint8_t * pix2, int line_size, int h) {

  403.     MpegEncContext *c = p;

  404.     int score1, score2;

  405. 

  406.     if(c) score1 = c->dsp.sse[0](c, pix1, pix2, line_size, h);

  407.     else  score1 = sse16_mmx(c, pix1, pix2, line_size, h);

  408.     score2= hf_noise16_mmx(pix1, line_size, h) - hf_noise16_mmx(pix2, line_size, h);

  409. 

  410.     if(c) return score1 + FFABS(score2)*c->avctx->nsse_weight;

  411.     else  return score1 + FFABS(score2)*8;

  412. }

  413. 

  414. static int nsse8_mmx(void *p, uint8_t * pix1, uint8_t * pix2, int line_size, int h) {

  415.     MpegEncContext *c = p;

  416.     int score1= sse8_mmx(c, pix1, pix2, line_size, h);

  417.     int score2= hf_noise8_mmx(pix1, line_size, h) - hf_noise8_mmx(pix2, line_size, h);

  418. 

  419.     if(c) return score1 + FFABS(score2)*c->avctx->nsse_weight;

  420.     else  return score1 + FFABS(score2)*8;

  421. }

  422. 

  423. static int vsad_intra16_mmx(void *v, uint8_t * pix, uint8_t * dummy, int line_size, int h) {

  424.     int tmp;

  425. 

  426.     assert( (((int)pix) & 7) == 0);

  427.     assert((line_size &7) ==0);

  428. 

  429. #define SUM(in0, in1, out0, out1) \

  430.       "movq (%0), %%mm2\n"\

  431.       "movq 8(%0), %%mm3\n"\

  432.       "add %2,%0\n"\

  433.       "movq %%mm2, " #out0 "\n"\

  434.       "movq %%mm3, " #out1 "\n"\

  435.       "psubusb " #in0 ", %%mm2\n"\

  436.       "psubusb " #in1 ", %%mm3\n"\

  437.       "psubusb " #out0 ", " #in0 "\n"\

  438.       "psubusb " #out1 ", " #in1 "\n"\

  439.       "por %%mm2, " #in0 "\n"\

  440.       "por %%mm3, " #in1 "\n"\

  441.       "movq " #in0 ", %%mm2\n"\

  442.       "movq " #in1 ", %%mm3\n"\

  443.       "punpcklbw %%mm7, " #in0 "\n"\

  444.       "punpcklbw %%mm7, " #in1 "\n"\

  445.       "punpckhbw %%mm7, %%mm2\n"\

  446.       "punpckhbw %%mm7, %%mm3\n"\

  447.       "paddw " #in1 ", " #in0 "\n"\

  448.       "paddw %%mm3, %%mm2\n"\

  449.       "paddw %%mm2, " #in0 "\n"\

  450.       "paddw " #in0 ", %%mm6\n"

  451. 

  452. 

  453.   __asm__ volatile (

  454.       "movl %3,%%ecx\n"

  455.       "pxor %%mm6,%%mm6\n"

  456.       "pxor %%mm7,%%mm7\n"

  457.       "movq (%0),%%mm0\n"

  458.       "movq 8(%0),%%mm1\n"

  459.       "add %2,%0\n"

  460.       "jmp 2f\n"

  461.       "1:\n"

  462. 

  463.       SUM(%%mm4, %%mm5, %%mm0, %%mm1)

  464.       "2:\n"

  465.       SUM(%%mm0, %%mm1, %%mm4, %%mm5)

  466. 

  467.       "subl $2, %%ecx\n"

  468.       "jnz 1b\n"

  469. 

  470.       "movq %%mm6,%%mm0\n"

  471.       "psrlq $32, %%mm6\n"

  472.       "paddw %%mm6,%%mm0\n"

  473.       "movq %%mm0,%%mm6\n"

  474.       "psrlq $16, %%mm0\n"

  475.       "paddw %%mm6,%%mm0\n"

  476.       "movd %%mm0,%1\n"

  477.       : "+r" (pix), "=r"(tmp)

  478.       : "r" ((x86_reg)line_size) , "m" (h)

  479.       : "%ecx");

  480.     return tmp & 0xFFFF;

  481. }

  482. #undef SUM

  483. 

  484. static int vsad_intra16_mmxext(void *v, uint8_t *pix, uint8_t *dummy,

  485.                                int line_size, int h)

  486. {

  487.     int tmp;

  488. 

  489.     assert( (((int)pix) & 7) == 0);

  490.     assert((line_size &7) ==0);

  491. 

  492. #define SUM(in0, in1, out0, out1) \

  493.       "movq (%0), " #out0 "\n"\

  494.       "movq 8(%0), " #out1 "\n"\

  495.       "add %2,%0\n"\

  496.       "psadbw " #out0 ", " #in0 "\n"\

  497.       "psadbw " #out1 ", " #in1 "\n"\

  498.       "paddw " #in1 ", " #in0 "\n"\

  499.       "paddw " #in0 ", %%mm6\n"

  500. 

  501.   __asm__ volatile (

  502.       "movl %3,%%ecx\n"

  503.       "pxor %%mm6,%%mm6\n"

  504.       "pxor %%mm7,%%mm7\n"

  505.       "movq (%0),%%mm0\n"

  506.       "movq 8(%0),%%mm1\n"

  507.       "add %2,%0\n"

  508.       "jmp 2f\n"

  509.       "1:\n"

  510. 

  511.       SUM(%%mm4, %%mm5, %%mm0, %%mm1)

  512.       "2:\n"

  513.       SUM(%%mm0, %%mm1, %%mm4, %%mm5)

  514. 

  515.       "subl $2, %%ecx\n"

  516.       "jnz 1b\n"

  517. 

  518.       "movd %%mm6,%1\n"

  519.       : "+r" (pix), "=r"(tmp)

  520.       : "r" ((x86_reg)line_size) , "m" (h)

  521.       : "%ecx");

  522.     return tmp;

  523. }

  524. #undef SUM

  525. 

  526. static int vsad16_mmx(void *v, uint8_t * pix1, uint8_t * pix2, int line_size, int h) {

  527.     int tmp;

  528. 

  529.     assert( (((int)pix1) & 7) == 0);

  530.     assert( (((int)pix2) & 7) == 0);

  531.     assert((line_size &7) ==0);

  532. 

  533. #define SUM(in0, in1, out0, out1) \

  534.       "movq (%0),%%mm2\n"\

  535.       "movq (%1)," #out0 "\n"\

  536.       "movq 8(%0),%%mm3\n"\

  537.       "movq 8(%1)," #out1 "\n"\

  538.       "add %3,%0\n"\

  539.       "add %3,%1\n"\

  540.       "psubb " #out0 ", %%mm2\n"\

  541.       "psubb " #out1 ", %%mm3\n"\

  542.       "pxor %%mm7, %%mm2\n"\

  543.       "pxor %%mm7, %%mm3\n"\

  544.       "movq %%mm2, " #out0 "\n"\

  545.       "movq %%mm3, " #out1 "\n"\

  546.       "psubusb " #in0 ", %%mm2\n"\

  547.       "psubusb " #in1 ", %%mm3\n"\

  548.       "psubusb " #out0 ", " #in0 "\n"\

  549.       "psubusb " #out1 ", " #in1 "\n"\

  550.       "por %%mm2, " #in0 "\n"\

  551.       "por %%mm3, " #in1 "\n"\

  552.       "movq " #in0 ", %%mm2\n"\

  553.       "movq " #in1 ", %%mm3\n"\

  554.       "punpcklbw %%mm7, " #in0 "\n"\

  555.       "punpcklbw %%mm7, " #in1 "\n"\

  556.       "punpckhbw %%mm7, %%mm2\n"\

  557.       "punpckhbw %%mm7, %%mm3\n"\

  558.       "paddw " #in1 ", " #in0 "\n"\

  559.       "paddw %%mm3, %%mm2\n"\

  560.       "paddw %%mm2, " #in0 "\n"\

  561.       "paddw " #in0 ", %%mm6\n"

  562. 

  563. 

  564.   __asm__ volatile (

  565.       "movl %4,%%ecx\n"

  566.       "pxor %%mm6,%%mm6\n"

  567.       "pcmpeqw %%mm7,%%mm7\n"

  568.       "psllw $15, %%mm7\n"

  569.       "packsswb %%mm7, %%mm7\n"

  570.       "movq (%0),%%mm0\n"

  571.       "movq (%1),%%mm2\n"

  572.       "movq 8(%0),%%mm1\n"

  573.       "movq 8(%1),%%mm3\n"

  574.       "add %3,%0\n"

  575.       "add %3,%1\n"

  576.       "psubb %%mm2, %%mm0\n"

  577.       "psubb %%mm3, %%mm1\n"

  578.       "pxor %%mm7, %%mm0\n"

  579.       "pxor %%mm7, %%mm1\n"

  580.       "jmp 2f\n"

  581.       "1:\n"

  582. 

  583.       SUM(%%mm4, %%mm5, %%mm0, %%mm1)

  584.       "2:\n"

  585.       SUM(%%mm0, %%mm1, %%mm4, %%mm5)

  586. 

  587.       "subl $2, %%ecx\n"

  588.       "jnz 1b\n"

  589. 

  590.       "movq %%mm6,%%mm0\n"

  591.       "psrlq $32, %%mm6\n"

  592.       "paddw %%mm6,%%mm0\n"

  593.       "movq %%mm0,%%mm6\n"

  594.       "psrlq $16, %%mm0\n"

  595.       "paddw %%mm6,%%mm0\n"

  596.       "movd %%mm0,%2\n"

  597.       : "+r" (pix1), "+r" (pix2), "=r"(tmp)

  598.       : "r" ((x86_reg)line_size) , "m" (h)

  599.       : "%ecx");

  600.     return tmp & 0x7FFF;

  601. }

  602. #undef SUM

  603. 

  604. static int vsad16_mmxext(void *v, uint8_t *pix1, uint8_t *pix2,

  605.                          int line_size, int h)

  606. {

  607.     int tmp;

  608. 

  609.     assert( (((int)pix1) & 7) == 0);

  610.     assert( (((int)pix2) & 7) == 0);

  611.     assert((line_size &7) ==0);

  612. 

  613. #define SUM(in0, in1, out0, out1) \

  614.       "movq (%0)," #out0 "\n"\

  615.       "movq (%1),%%mm2\n"\

  616.       "movq 8(%0)," #out1 "\n"\

  617.       "movq 8(%1),%%mm3\n"\

  618.       "add %3,%0\n"\

  619.       "add %3,%1\n"\

  620.       "psubb %%mm2, " #out0 "\n"\

  621.       "psubb %%mm3, " #out1 "\n"\

  622.       "pxor %%mm7, " #out0 "\n"\

  623.       "pxor %%mm7, " #out1 "\n"\

  624.       "psadbw " #out0 ", " #in0 "\n"\

  625.       "psadbw " #out1 ", " #in1 "\n"\

  626.       "paddw " #in1 ", " #in0 "\n"\

  627.       "paddw " #in0 ", %%mm6\n"

  628. 

  629.   __asm__ volatile (

  630.       "movl %4,%%ecx\n"

  631.       "pxor %%mm6,%%mm6\n"

  632.       "pcmpeqw %%mm7,%%mm7\n"

  633.       "psllw $15, %%mm7\n"

  634.       "packsswb %%mm7, %%mm7\n"

  635.       "movq (%0),%%mm0\n"

  636.       "movq (%1),%%mm2\n"

  637.       "movq 8(%0),%%mm1\n"

  638.       "movq 8(%1),%%mm3\n"

  639.       "add %3,%0\n"

  640.       "add %3,%1\n"

  641.       "psubb %%mm2, %%mm0\n"

  642.       "psubb %%mm3, %%mm1\n"

  643.       "pxor %%mm7, %%mm0\n"

  644.       "pxor %%mm7, %%mm1\n"

  645.       "jmp 2f\n"

  646.       "1:\n"

  647. 

  648.       SUM(%%mm4, %%mm5, %%mm0, %%mm1)

  649.       "2:\n"

  650.       SUM(%%mm0, %%mm1, %%mm4, %%mm5)

  651. 

  652.       "subl $2, %%ecx\n"

  653.       "jnz 1b\n"

  654. 

  655.       "movd %%mm6,%2\n"

  656.       : "+r" (pix1), "+r" (pix2), "=r"(tmp)

  657.       : "r" ((x86_reg)line_size) , "m" (h)

  658.       : "%ecx");

  659.     return tmp;

  660. }

  661. #undef SUM

  662. 

  663. static void diff_bytes_mmx(uint8_t *dst, uint8_t *src1, uint8_t *src2, int w){

  664.     x86_reg i=0;

  665.     __asm__ volatile(

  666.         "1:                             \n\t"

  667.         "movq  (%2, %0), %%mm0          \n\t"

  668.         "movq  (%1, %0), %%mm1          \n\t"

  669.         "psubb %%mm0, %%mm1             \n\t"

  670.         "movq %%mm1, (%3, %0)           \n\t"

  671.         "movq 8(%2, %0), %%mm0          \n\t"

  672.         "movq 8(%1, %0), %%mm1          \n\t"

  673.         "psubb %%mm0, %%mm1             \n\t"

  674.         "movq %%mm1, 8(%3, %0)          \n\t"

  675.         "add $16, %0                    \n\t"

  676.         "cmp %4, %0                     \n\t"

  677.         " jb 1b                         \n\t"

  678.         : "+r" (i)

  679.         : "r"(src1), "r"(src2), "r"(dst), "r"((x86_reg)w-15)

  680.     );

  681.     for(; i<w; i++)

  682.         dst[i+0] = src1[i+0]-src2[i+0];

  683. }

  684. 

  685. static void sub_hfyu_median_prediction_mmxext(uint8_t *dst, const uint8_t *src1,

  686.                                               const uint8_t *src2, int w,

  687.                                               int *left, int *left_top)

  688. {

  689.     x86_reg i=0;

  690.     uint8_t l, lt;

  691. 

  692.     __asm__ volatile(

  693.         "movq  (%1, %0), %%mm0          \n\t" // LT

  694.         "psllq $8, %%mm0                \n\t"

  695.         "1:                             \n\t"

  696.         "movq  (%1, %0), %%mm1          \n\t" // T

  697.         "movq  -1(%2, %0), %%mm2        \n\t" // L

  698.         "movq  (%2, %0), %%mm3          \n\t" // X

  699.         "movq %%mm2, %%mm4              \n\t" // L

  700.         "psubb %%mm0, %%mm2             \n\t"

  701.         "paddb %%mm1, %%mm2             \n\t" // L + T - LT

  702.         "movq %%mm4, %%mm5              \n\t" // L

  703.         "pmaxub %%mm1, %%mm4            \n\t" // max(T, L)

  704.         "pminub %%mm5, %%mm1            \n\t" // min(T, L)

  705.         "pminub %%mm2, %%mm4            \n\t"

  706.         "pmaxub %%mm1, %%mm4            \n\t"

  707.         "psubb %%mm4, %%mm3             \n\t" // dst - pred

  708.         "movq %%mm3, (%3, %0)           \n\t"

  709.         "add $8, %0                     \n\t"

  710.         "movq -1(%1, %0), %%mm0         \n\t" // LT

  711.         "cmp %4, %0                     \n\t"

  712.         " jb 1b                         \n\t"

  713.         : "+r" (i)

  714.         : "r"(src1), "r"(src2), "r"(dst), "r"((x86_reg)w)

  715.     );

  716. 

  717.     l= *left;

  718.     lt= *left_top;

  719. 

  720.     dst[0]= src2[0] - mid_pred(l, src1[0], (l + src1[0] - lt)&0xFF);

  721. 

  722.     *left_top= src1[w-1];

  723.     *left    = src2[w-1];

  724. }

  725. 

  726. #define MMABS_MMX(a,z)\

  727.     "pxor " #z ", " #z "              \n\t"\

  728.     "pcmpgtw " #a ", " #z "           \n\t"\

  729.     "pxor " #z ", " #a "              \n\t"\

  730.     "psubw " #z ", " #a "             \n\t"

  731. 

  732. #define MMABS_MMXEXT(a, z)                 \

  733.     "pxor " #z ", " #z "              \n\t"\

  734.     "psubw " #a ", " #z "             \n\t"\

  735.     "pmaxsw " #z ", " #a "            \n\t"

  736. 

  737. #define MMABS_SSSE3(a,z)\

  738.     "pabsw " #a ", " #a "             \n\t"

  739. 

  740. #define MMABS_SUM(a,z, sum)\

  741.     MMABS(a,z)\

  742.     "paddusw " #a ", " #sum "         \n\t"

  743. 

  744. /* FIXME: HSUM_* saturates at 64k, while an 8x8 hadamard or dct block can get up to

  745.  * about 100k on extreme inputs. But that's very unlikely to occur in natural video,

  746.  * and it's even more unlikely to not have any alternative mvs/modes with lower cost. */

  747. #define HSUM_MMX(a, t, dst)\

  748.     "movq "#a", "#t"                  \n\t"\

  749.     "psrlq $32, "#a"                  \n\t"\

  750.     "paddusw "#t", "#a"               \n\t"\

  751.     "movq "#a", "#t"                  \n\t"\

  752.     "psrlq $16, "#a"                  \n\t"\

  753.     "paddusw "#t", "#a"               \n\t"\

  754.     "movd "#a", "#dst"                \n\t"\

  755. 

  756. #define HSUM_MMXEXT(a, t, dst)             \

  757.     "pshufw $0x0E, "#a", "#t"         \n\t"\

  758.     "paddusw "#t", "#a"               \n\t"\

  759.     "pshufw $0x01, "#a", "#t"         \n\t"\

  760.     "paddusw "#t", "#a"               \n\t"\

  761.     "movd "#a", "#dst"                \n\t"\

  762. 

  763. #define HSUM_SSE2(a, t, dst)\

  764.     "movhlps "#a", "#t"               \n\t"\

  765.     "paddusw "#t", "#a"               \n\t"\

  766.     "pshuflw $0x0E, "#a", "#t"        \n\t"\

  767.     "paddusw "#t", "#a"               \n\t"\

  768.     "pshuflw $0x01, "#a", "#t"        \n\t"\

  769.     "paddusw "#t", "#a"               \n\t"\

  770.     "movd "#a", "#dst"                \n\t"\

  771. 

  772. #define DCT_SAD4(m,mm,o)\

  773.     "mov"#m" "#o"+ 0(%1), "#mm"2      \n\t"\

  774.     "mov"#m" "#o"+16(%1), "#mm"3      \n\t"\

  775.     "mov"#m" "#o"+32(%1), "#mm"4      \n\t"\

  776.     "mov"#m" "#o"+48(%1), "#mm"5      \n\t"\

  777.     MMABS_SUM(mm##2, mm##6, mm##0)\

  778.     MMABS_SUM(mm##3, mm##7, mm##1)\

  779.     MMABS_SUM(mm##4, mm##6, mm##0)\

  780.     MMABS_SUM(mm##5, mm##7, mm##1)\

  781. 

  782. #define DCT_SAD_MMX\

  783.     "pxor %%mm0, %%mm0                \n\t"\

  784.     "pxor %%mm1, %%mm1                \n\t"\

  785.     DCT_SAD4(q, %%mm, 0)\

  786.     DCT_SAD4(q, %%mm, 8)\

  787.     DCT_SAD4(q, %%mm, 64)\

  788.     DCT_SAD4(q, %%mm, 72)\

  789.     "paddusw %%mm1, %%mm0             \n\t"\

  790.     HSUM(%%mm0, %%mm1, %0)

  791. 

  792. #define DCT_SAD_SSE2\

  793.     "pxor %%xmm0, %%xmm0              \n\t"\

  794.     "pxor %%xmm1, %%xmm1              \n\t"\

  795.     DCT_SAD4(dqa, %%xmm, 0)\

  796.     DCT_SAD4(dqa, %%xmm, 64)\

  797.     "paddusw %%xmm1, %%xmm0           \n\t"\

  798.     HSUM(%%xmm0, %%xmm1, %0)

  799. 

  800. #define DCT_SAD_FUNC(cpu) \

  801. static int sum_abs_dctelem_##cpu(int16_t *block){\

  802.     int sum;\

  803.     __asm__ volatile(\

  804.         DCT_SAD\

  805.         :"=r"(sum)\

  806.         :"r"(block)\

  807.     );\

  808.     return sum&0xFFFF;\

  809. }

  810. 

  811. #define DCT_SAD       DCT_SAD_MMX

  812. #define HSUM(a,t,dst) HSUM_MMX(a,t,dst)

  813. #define MMABS(a,z)    MMABS_MMX(a,z)

  814. DCT_SAD_FUNC(mmx)

  815. #undef MMABS

  816. #undef HSUM

  817. 

  818. #define HSUM(a,t,dst) HSUM_MMXEXT(a,t,dst)

  819. #define MMABS(a,z)    MMABS_MMXEXT(a,z)

  820. DCT_SAD_FUNC(mmxext)

  821. #undef HSUM

  822. #undef DCT_SAD

  823. 

  824. #define DCT_SAD       DCT_SAD_SSE2

  825. #define HSUM(a,t,dst) HSUM_SSE2(a,t,dst)

  826. DCT_SAD_FUNC(sse2)

  827. #undef MMABS

  828. 

  829. #if HAVE_SSSE3_INLINE

  830. #define MMABS(a,z)    MMABS_SSSE3(a,z)

  831. DCT_SAD_FUNC(ssse3)

  832. #undef MMABS

  833. #endif

  834. #undef HSUM

  835. #undef DCT_SAD

  836. 

  837. static int ssd_int8_vs_int16_mmx(const int8_t *pix1, const int16_t *pix2, int size){

  838.     int sum;

  839.     x86_reg i=size;

  840.     __asm__ volatile(

  841.         "pxor %%mm4, %%mm4 \n"

  842.         "1: \n"

  843.         "sub $8, %0 \n"

  844.         "movq (%2,%0), %%mm2 \n"

  845.         "movq (%3,%0,2), %%mm0 \n"

  846.         "movq 8(%3,%0,2), %%mm1 \n"

  847.         "punpckhbw %%mm2, %%mm3 \n"

  848.         "punpcklbw %%mm2, %%mm2 \n"

  849.         "psraw $8, %%mm3 \n"

  850.         "psraw $8, %%mm2 \n"

  851.         "psubw %%mm3, %%mm1 \n"

  852.         "psubw %%mm2, %%mm0 \n"

  853.         "pmaddwd %%mm1, %%mm1 \n"

  854.         "pmaddwd %%mm0, %%mm0 \n"

  855.         "paddd %%mm1, %%mm4 \n"

  856.         "paddd %%mm0, %%mm4 \n"

  857.         "jg 1b \n"

  858.         "movq %%mm4, %%mm3 \n"

  859.         "psrlq $32, %%mm3 \n"

  860.         "paddd %%mm3, %%mm4 \n"

  861.         "movd %%mm4, %1 \n"

  862.         :"+r"(i), "=r"(sum)

  863.         :"r"(pix1), "r"(pix2)

  864.     );

  865.     return sum;

  866. }

  867. 

  868. #define PHADDD(a, t)\

  869.     "movq "#a", "#t"                  \n\t"\

  870.     "psrlq $32, "#a"                  \n\t"\

  871.     "paddd "#t", "#a"                 \n\t"

  872. /*

  873.    pmulhw: dst[0-15]=(src[0-15]*dst[0-15])[16-31]

  874.    pmulhrw: dst[0-15]=(src[0-15]*dst[0-15] + 0x8000)[16-31]

  875.    pmulhrsw: dst[0-15]=(src[0-15]*dst[0-15] + 0x4000)[15-30]

  876.  */

  877. #define PMULHRW(x, y, s, o)\

  878.     "pmulhw " #s ", "#x "            \n\t"\

  879.     "pmulhw " #s ", "#y "            \n\t"\

  880.     "paddw " #o ", "#x "             \n\t"\

  881.     "paddw " #o ", "#y "             \n\t"\

  882.     "psraw $1, "#x "                 \n\t"\

  883.     "psraw $1, "#y "                 \n\t"

  884. #define DEF(x) x ## _mmx

  885. #define SET_RND MOVQ_WONE

  886. #define SCALE_OFFSET 1

  887. 

  888. #include "dsputil_qns_template.c"

  889. 

  890. #undef DEF

  891. #undef SET_RND

  892. #undef SCALE_OFFSET

  893. #undef PMULHRW

  894. 

  895. #define DEF(x) x ## _3dnow

  896. #define SET_RND(x)

  897. #define SCALE_OFFSET 0

  898. #define PMULHRW(x, y, s, o)\

  899.     "pmulhrw " #s ", "#x "           \n\t"\

  900.     "pmulhrw " #s ", "#y "           \n\t"

  901. 

  902. #include "dsputil_qns_template.c"

  903. 

  904. #undef DEF

  905. #undef SET_RND

  906. #undef SCALE_OFFSET

  907. #undef PMULHRW

  908. 

  909. #if HAVE_SSSE3_INLINE

  910. #undef PHADDD

  911. #define DEF(x) x ## _ssse3

  912. #define SET_RND(x)

  913. #define SCALE_OFFSET -1

  914. #define PHADDD(a, t)\

  915.     "pshufw $0x0E, "#a", "#t"         \n\t"\

  916.     "paddd "#t", "#a"                 \n\t" /* faster than phaddd on core2 */

  917. #define PMULHRW(x, y, s, o)\

  918.     "pmulhrsw " #s ", "#x "          \n\t"\

  919.     "pmulhrsw " #s ", "#y "          \n\t"

  920. 

  921. #include "dsputil_qns_template.c"

  922. 

  923. #undef DEF

  924. #undef SET_RND

  925. #undef SCALE_OFFSET

  926. #undef PMULHRW

  927. #undef PHADDD

  928. #endif /* HAVE_SSSE3_INLINE */

  929. 

  930. #endif /* HAVE_INLINE_ASM */

  931. 

  932. int ff_sse16_sse2(void *v, uint8_t * pix1, uint8_t * pix2, int line_size, int h);

  933. 

  934. #define hadamard_func(cpu) \

  935. int ff_hadamard8_diff_##cpu  (void *s, uint8_t *src1, uint8_t *src2, \

  936.                               int stride, int h); \

  937. int ff_hadamard8_diff16_##cpu(void *s, uint8_t *src1, uint8_t *src2, \

  938.                               int stride, int h);

  939. 

  940. hadamard_func(mmx)

  941. hadamard_func(mmxext)

  942. hadamard_func(sse2)

  943. hadamard_func(ssse3)

  944. 

  945. av_cold void ff_dsputilenc_init_mmx(DSPContext *c, AVCodecContext *avctx)

  946. {

  947.     int mm_flags = av_get_cpu_flags();

  948.     int bit_depth = avctx->bits_per_raw_sample;

  949. 

  950. #if HAVE_YASM

  951.     if (EXTERNAL_MMX(mm_flags)) {

  952.         if (bit_depth <= 8)

  953.             c->get_pixels = ff_get_pixels_mmx;

  954.         c->diff_pixels = ff_diff_pixels_mmx;

  955.         c->pix_sum = ff_pix_sum16_mmx;

  956. 

  957.         c->pix_norm1 = ff_pix_norm1_mmx;

  958.     }

  959.     if (EXTERNAL_SSE2(mm_flags))

  960.         if (bit_depth <= 8)

  961.             c->get_pixels = ff_get_pixels_sse2;

  962. #endif /* HAVE_YASM */

  963. 

  964. #if HAVE_INLINE_ASM

  965.     if (mm_flags & AV_CPU_FLAG_MMX) {

  966.         const int dct_algo = avctx->dct_algo;

  967.         if (avctx->bits_per_raw_sample <= 8 &&

  968.             (dct_algo==FF_DCT_AUTO || dct_algo==FF_DCT_MMX)) {

  969.             if(mm_flags & AV_CPU_FLAG_SSE2){

  970.                 c->fdct = ff_fdct_sse2;

  971.             } else if (mm_flags & AV_CPU_FLAG_MMXEXT) {

  972.                 c->fdct = ff_fdct_mmxext;

  973.             }else{

  974.                 c->fdct = ff_fdct_mmx;

  975.             }

  976.         }

  977. 

  978. 

  979.         c->diff_bytes= diff_bytes_mmx;

  980.         c->sum_abs_dctelem= sum_abs_dctelem_mmx;

  981. 

  982.         c->sse[0] = sse16_mmx;

  983.         c->sse[1] = sse8_mmx;

  984.         c->vsad[4]= vsad_intra16_mmx;

  985. 

  986.         c->nsse[0] = nsse16_mmx;

  987.         c->nsse[1] = nsse8_mmx;

  988.         if(!(avctx->flags & CODEC_FLAG_BITEXACT)){

  989.             c->vsad[0] = vsad16_mmx;

  990.         }

  991. 

  992.         if(!(avctx->flags & CODEC_FLAG_BITEXACT)){

  993.             c->try_8x8basis= try_8x8basis_mmx;

  994.         }

  995.         c->add_8x8basis= add_8x8basis_mmx;

  996. 

  997.         c->ssd_int8_vs_int16 = ssd_int8_vs_int16_mmx;

  998. 

  999.         if (mm_flags & AV_CPU_FLAG_MMXEXT) {

  1000.             c->sum_abs_dctelem = sum_abs_dctelem_mmxext;

  1001.             c->vsad[4]         = vsad_intra16_mmxext;

  1002. 

  1003.             if(!(avctx->flags & CODEC_FLAG_BITEXACT)){

  1004.                 c->vsad[0] = vsad16_mmxext;

  1005.             }

  1006. 

  1007.             c->sub_hfyu_median_prediction = sub_hfyu_median_prediction_mmxext;

  1008.         }

  1009. 

  1010.         if(mm_flags & AV_CPU_FLAG_SSE2){

  1011.             c->sum_abs_dctelem= sum_abs_dctelem_sse2;

  1012.         }

  1013. 

  1014. #if HAVE_SSSE3_INLINE

  1015.         if(mm_flags & AV_CPU_FLAG_SSSE3){

  1016.             if(!(avctx->flags & CODEC_FLAG_BITEXACT)){

  1017.                 c->try_8x8basis= try_8x8basis_ssse3;

  1018.             }

  1019.             c->add_8x8basis= add_8x8basis_ssse3;

  1020.             c->sum_abs_dctelem= sum_abs_dctelem_ssse3;

  1021.         }

  1022. #endif

  1023. 

  1024.         if(mm_flags & AV_CPU_FLAG_3DNOW){

  1025.             if(!(avctx->flags & CODEC_FLAG_BITEXACT)){

  1026.                 c->try_8x8basis= try_8x8basis_3dnow;

  1027.             }

  1028.             c->add_8x8basis= add_8x8basis_3dnow;

  1029.         }

  1030.     }

  1031. #endif /* HAVE_INLINE_ASM */

  1032. 

  1033.     if (EXTERNAL_MMX(mm_flags)) {

  1034.         c->hadamard8_diff[0] = ff_hadamard8_diff16_mmx;

  1035.         c->hadamard8_diff[1] = ff_hadamard8_diff_mmx;

  1036. 

  1037.         if (EXTERNAL_MMXEXT(mm_flags)) {

  1038.             c->hadamard8_diff[0] = ff_hadamard8_diff16_mmxext;

  1039.             c->hadamard8_diff[1] = ff_hadamard8_diff_mmxext;

  1040.         }

  1041. 

  1042.         if (EXTERNAL_SSE2(mm_flags)) {

  1043.             c->sse[0] = ff_sse16_sse2;

  1044. 

  1045. #if HAVE_ALIGNED_STACK

  1046.             c->hadamard8_diff[0] = ff_hadamard8_diff16_sse2;

  1047.             c->hadamard8_diff[1] = ff_hadamard8_diff_sse2;

  1048. #endif

  1049.         }

  1050. 

  1051.         if (EXTERNAL_SSSE3(mm_flags) && HAVE_ALIGNED_STACK) {

  1052.             c->hadamard8_diff[0] = ff_hadamard8_diff16_ssse3;

  1053.             c->hadamard8_diff[1] = ff_hadamard8_diff_ssse3;

  1054.         }

  1055.     }

  1056. 

  1057.     ff_dsputil_init_pix_mmx(c, avctx);

  1058. }