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.
43415 Commits
32 Branches
283MB
Tree: 82b7525173
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from '82b7525173'
${ noResults }
FFmpeg/libavcodec/x86/me_cmp_init.c

1326 lines
47KB
Raw Blame History 

  1. /*

  2.  * SIMD-optimized motion estimation

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

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

  5.  *

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

  7.  *

  8.  * This file is part of Libav.

  9.  *

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

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

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

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

  14.  *

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

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

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

  18.  * Lesser General Public License for more details.

  19.  *

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

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

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

  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/me_cmp.h"

  30. #include "libavcodec/mpegvideo.h"

  31. 

  32. #if HAVE_INLINE_ASM

  33. 

  34. static int sse8_mmx(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,

  35.                     ptrdiff_t stride, int h)

  36. {

  37.     int tmp;

  38. 

  39.     __asm__ volatile (

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

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

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

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

  44.         "1:                              \n"

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

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

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

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

  49. 

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

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

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

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

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

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

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

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

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

  59. 

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

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

  62. 

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

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

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

  66. 

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

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

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

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

  71. 

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

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

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

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

  76. 

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

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

  79. 

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

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

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

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

  84. 

  85.         "decl      %%ecx                 \n"

  86.         "jnz       1b                    \n"

  87. 

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

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

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

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

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

  93.         : "r" (stride), "m" (h)

  94.         : "%ecx");

  95. 

  96.     return tmp;

  97. }

  98. 

  99. static int sse16_mmx(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,

  100.                      ptrdiff_t stride, int h)

  101. {

  102.     int tmp;

  103. 

  104.     __asm__ volatile (

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

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

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

  108.         "1:\n"

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

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

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

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

  113. 

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

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

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

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

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

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

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

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

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

  123. 

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

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

  126. 

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

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

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

  130. 

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

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

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

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

  135. 

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

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

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

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

  140. 

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

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

  143. 

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

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

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

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

  148. 

  149.         "decl %%ecx\n"

  150.         "jnz 1b\n"

  151. 

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

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

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

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

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

  157.         : "r" (stride), "m" (h)

  158.         : "%ecx");

  159. 

  160.     return tmp;

  161. }

  162. 

  163. static int hf_noise8_mmx(uint8_t *pix1, ptrdiff_t stride, int h)

  164. {

  165.     int tmp;

  166. 

  167.     __asm__ volatile (

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

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

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

  171. 

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

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

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

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

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

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

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

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

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

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

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

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

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

  185. 

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

  187. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  213. 

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

  215.         "1:\n"

  216. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  242. 

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

  244. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  270. 

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

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

  273.         " jnz 1b\n"

  274. 

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

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

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

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

  279. 

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

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

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

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

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

  285.         : "r" (stride), "g" (h - 2)

  286.         : "%ecx");

  287. 

  288.     return tmp;

  289. }

  290. 

  291. static int hf_noise16_mmx(uint8_t *pix1, ptrdiff_t stride, int h)

  292. {

  293.     int tmp;

  294.     uint8_t *pix = pix1;

  295. 

  296.     __asm__ volatile (

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

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

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

  300. 

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

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

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

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

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

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

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

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

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

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

  311. 

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

  313. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  336. 

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

  338.         "1:\n"

  339. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  362. 

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

  364. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  387. 

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

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

  390.         " jnz 1b\n"

  391. 

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

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

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

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

  396. 

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

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

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

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

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

  402.         : "r" (stride), "g" (h - 2)

  403.         : "%ecx");

  404. 

  405.     return tmp + hf_noise8_mmx(pix + 8, stride, h);

  406. }

  407. 

  408. static int nsse16_mmx(MpegEncContext *c, uint8_t *pix1, uint8_t *pix2,

  409.                       ptrdiff_t stride, int h)

  410. {

  411.     int score1, score2;

  412. 

  413.     if (c)

  414.         score1 = c->mecc.sse[0](c, pix1, pix2, stride, h);

  415.     else

  416.         score1 = sse16_mmx(c, pix1, pix2, stride, h);

  417.     score2 = hf_noise16_mmx(pix1, stride, h) -

  418.              hf_noise16_mmx(pix2, stride, h);

  419. 

  420.     if (c)

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

  422.     else

  423.         return score1 + FFABS(score2) * 8;

  424. }

  425. 

  426. static int nsse8_mmx(MpegEncContext *c, uint8_t *pix1, uint8_t *pix2,

  427.                      ptrdiff_t stride, int h)

  428. {

  429.     int score1 = sse8_mmx(c, pix1, pix2, stride, h);

  430.     int score2 = hf_noise8_mmx(pix1, stride, h) -

  431.                  hf_noise8_mmx(pix2, stride, h);

  432. 

  433.     if (c)

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

  435.     else

  436.         return score1 + FFABS(score2) * 8;

  437. }

  438. 

  439. static int vsad_intra16_mmx(MpegEncContext *v, uint8_t *pix, uint8_t *dummy,

  440.                             ptrdiff_t stride, int h)

  441. {

  442.     int tmp;

  443. 

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

  445.     assert((stride & 7) == 0);

  446. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  469. 

  470. 

  471.     __asm__ volatile (

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

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

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

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

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

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

  478.         "jmp 2f\n"

  479.         "1:\n"

  480. 

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

  482.         "2:\n"

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

  484. 

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

  486.         "jnz 1b\n"

  487. 

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

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

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

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

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

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

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

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

  496.         : "r" (stride), "m" (h)

  497.         : "%ecx");

  498. 

  499.     return tmp & 0xFFFF;

  500. }

  501. #undef SUM

  502. 

  503. static int vsad_intra16_mmxext(MpegEncContext *v, uint8_t *pix, uint8_t *dummy,

  504.                                ptrdiff_t stride, int h)

  505. {

  506.     int tmp;

  507. 

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

  509.     assert((stride & 7) == 0);

  510. 

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

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

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

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

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

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

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

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

  519. 

  520.     __asm__ volatile (

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

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

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

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

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

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

  527.         "jmp 2f\n"

  528.         "1:\n"

  529. 

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

  531.         "2:\n"

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

  533. 

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

  535.         "jnz 1b\n"

  536. 

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

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

  539.         : "r" (stride), "m" (h)

  540.         : "%ecx");

  541. 

  542.     return tmp;

  543. }

  544. #undef SUM

  545. 

  546. static int vsad16_mmx(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,

  547.                       ptrdiff_t stride, int h)

  548. {

  549.     int tmp;

  550. 

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

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

  553.     assert((stride & 7) == 0);

  554. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  584. 

  585. 

  586.     __asm__ volatile (

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  602.         "jmp 2f\n"

  603.         "1:\n"

  604. 

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

  606.         "2:\n"

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

  608. 

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

  610.         "jnz 1b\n"

  611. 

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

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

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

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

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

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

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

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

  620.         : "r" (stride), "m" (h)

  621.         : "%ecx");

  622. 

  623.     return tmp & 0x7FFF;

  624. }

  625. #undef SUM

  626. 

  627. static int vsad16_mmxext(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,

  628.                          ptrdiff_t stride, int h)

  629. {

  630.     int tmp;

  631. 

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

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

  634.     assert((stride & 7) == 0);

  635. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  651. 

  652.     __asm__ volatile (

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  668.         "jmp 2f\n"

  669.         "1:\n"

  670. 

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

  672.         "2:\n"

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

  674. 

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

  676.         "jnz 1b\n"

  677. 

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

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

  680.         : "r" (stride), "m" (h)

  681.         : "%ecx");

  682. 

  683.     return tmp;

  684. }

  685. #undef SUM

  686. 

  687. #define MMABS_MMX(a,z)                          \

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

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

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

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

  692. 

  693. #define MMABS_MMXEXT(a, z)                      \

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

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

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

  697. 

  698. #define MMABS_SSSE3(a,z)                        \

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

  700. 

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

  702.     MMABS(a,z)                                  \

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

  704. 

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

  706.  * up to about 100k on extreme inputs. But that's very unlikely to occur in

  707.  * natural video, and it's even more unlikely to not have any alternative

  708.  * mvs/modes with lower cost. */

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

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

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

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

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

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

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

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

  717. 

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

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

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

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

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

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

  724. 

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

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

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

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

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

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

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

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

  733. 

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

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

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

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

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

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

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

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

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

  743. 

  744. #define DCT_SAD_MMX                             \

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

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

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

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

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

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

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

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

  753. 

  754. #define DCT_SAD_SSE2                            \

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

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

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

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

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

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

  761. 

  762. #define DCT_SAD_FUNC(cpu)                           \

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

  764. {                                                   \

  765.     int sum;                                        \

  766.     __asm__ volatile (                              \

  767.         DCT_SAD                                     \

  768.         :"=r"(sum)                                  \

  769.         :"r"(block));                               \

  770.     return sum & 0xFFFF;                            \

  771. }

  772. 

  773. #define DCT_SAD         DCT_SAD_MMX

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

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

  776. DCT_SAD_FUNC(mmx)

  777. #undef MMABS

  778. #undef HSUM

  779. 

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

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

  782. DCT_SAD_FUNC(mmxext)

  783. #undef HSUM

  784. #undef DCT_SAD

  785. 

  786. #define DCT_SAD         DCT_SAD_SSE2

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

  788. DCT_SAD_FUNC(sse2)

  789. #undef MMABS

  790. 

  791. #if HAVE_SSSE3_INLINE

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

  793. DCT_SAD_FUNC(ssse3)

  794. #undef MMABS

  795. #endif

  796. #undef HSUM

  797. #undef DCT_SAD

  798. 

  799. 

  800. DECLARE_ASM_CONST(8, uint64_t, round_tab)[3] = {

  801.     0x0000000000000000ULL,

  802.     0x0001000100010001ULL,

  803.     0x0002000200020002ULL,

  804. };

  805. 

  806. DECLARE_ASM_CONST(8, uint64_t, bone) = 0x0101010101010101LL;

  807. 

  808. static inline void sad8_1_mmx(uint8_t *blk1, uint8_t *blk2,

  809.                               ptrdiff_t stride, int h)

  810. {

  811.     x86_reg len = -(stride * h);

  812.     __asm__ volatile (

  813.         ".p2align 4                     \n\t"

  814.         "1:                             \n\t"

  815.         "movq (%1, %%"FF_REG_a"), %%mm0 \n\t"

  816.         "movq (%2, %%"FF_REG_a"), %%mm2 \n\t"

  817.         "movq (%2, %%"FF_REG_a"), %%mm4 \n\t"

  818.         "add %3, %%"FF_REG_a"           \n\t"

  819.         "psubusb %%mm0, %%mm2           \n\t"

  820.         "psubusb %%mm4, %%mm0           \n\t"

  821.         "movq (%1, %%"FF_REG_a"), %%mm1 \n\t"

  822.         "movq (%2, %%"FF_REG_a"), %%mm3 \n\t"

  823.         "movq (%2, %%"FF_REG_a"), %%mm5 \n\t"

  824.         "psubusb %%mm1, %%mm3           \n\t"

  825.         "psubusb %%mm5, %%mm1           \n\t"

  826.         "por %%mm2, %%mm0               \n\t"

  827.         "por %%mm1, %%mm3               \n\t"

  828.         "movq %%mm0, %%mm1              \n\t"

  829.         "movq %%mm3, %%mm2              \n\t"

  830.         "punpcklbw %%mm7, %%mm0         \n\t"

  831.         "punpckhbw %%mm7, %%mm1         \n\t"

  832.         "punpcklbw %%mm7, %%mm3         \n\t"

  833.         "punpckhbw %%mm7, %%mm2         \n\t"

  834.         "paddw %%mm1, %%mm0             \n\t"

  835.         "paddw %%mm3, %%mm2             \n\t"

  836.         "paddw %%mm2, %%mm0             \n\t"

  837.         "paddw %%mm0, %%mm6             \n\t"

  838.         "add %3, %%"FF_REG_a"           \n\t"

  839.         " js 1b                         \n\t"

  840.         : "+a" (len)

  841.         : "r" (blk1 - len), "r" (blk2 - len), "r" (stride));

  842. }

  843. 

  844. static inline void sad8_1_mmxext(uint8_t *blk1, uint8_t *blk2,

  845.                                  ptrdiff_t stride, int h)

  846. {

  847.     __asm__ volatile (

  848.         ".p2align 4                     \n\t"

  849.         "1:                             \n\t"

  850.         "movq (%1), %%mm0               \n\t"

  851.         "movq (%1, %3), %%mm1           \n\t"

  852.         "psadbw (%2), %%mm0             \n\t"

  853.         "psadbw (%2, %3), %%mm1         \n\t"

  854.         "paddw %%mm0, %%mm6             \n\t"

  855.         "paddw %%mm1, %%mm6             \n\t"

  856.         "lea (%1,%3,2), %1              \n\t"

  857.         "lea (%2,%3,2), %2              \n\t"

  858.         "sub $2, %0                     \n\t"

  859.         " jg 1b                         \n\t"

  860.         : "+r" (h), "+r" (blk1), "+r" (blk2)

  861.         : "r" (stride));

  862. }

  863. 

  864. static int sad16_sse2(MpegEncContext *v, uint8_t *blk2, uint8_t *blk1,

  865.                       ptrdiff_t stride, int h)

  866. {

  867.     int ret;

  868.     __asm__ volatile (

  869.         "pxor %%xmm2, %%xmm2            \n\t"

  870.         ".p2align 4                     \n\t"

  871.         "1:                             \n\t"

  872.         "movdqu (%1), %%xmm0            \n\t"

  873.         "movdqu (%1, %4), %%xmm1        \n\t"

  874.         "psadbw (%2), %%xmm0            \n\t"

  875.         "psadbw (%2, %4), %%xmm1        \n\t"

  876.         "paddw %%xmm0, %%xmm2           \n\t"

  877.         "paddw %%xmm1, %%xmm2           \n\t"

  878.         "lea (%1,%4,2), %1              \n\t"

  879.         "lea (%2,%4,2), %2              \n\t"

  880.         "sub $2, %0                     \n\t"

  881.         " jg 1b                         \n\t"

  882.         "movhlps %%xmm2, %%xmm0         \n\t"

  883.         "paddw   %%xmm0, %%xmm2         \n\t"

  884.         "movd    %%xmm2, %3             \n\t"

  885.         : "+r" (h), "+r" (blk1), "+r" (blk2), "=r" (ret)

  886.         : "r" (stride));

  887.     return ret;

  888. }

  889. 

  890. static inline void sad8_x2a_mmxext(uint8_t *blk1, uint8_t *blk2,

  891.                                    ptrdiff_t stride, int h)

  892. {

  893.     __asm__ volatile (

  894.         ".p2align 4                     \n\t"

  895.         "1:                             \n\t"

  896.         "movq (%1), %%mm0               \n\t"

  897.         "movq (%1, %3), %%mm1           \n\t"

  898.         "pavgb 1(%1), %%mm0             \n\t"

  899.         "pavgb 1(%1, %3), %%mm1         \n\t"

  900.         "psadbw (%2), %%mm0             \n\t"

  901.         "psadbw (%2, %3), %%mm1         \n\t"

  902.         "paddw %%mm0, %%mm6             \n\t"

  903.         "paddw %%mm1, %%mm6             \n\t"

  904.         "lea (%1,%3,2), %1              \n\t"

  905.         "lea (%2,%3,2), %2              \n\t"

  906.         "sub $2, %0                     \n\t"

  907.         " jg 1b                         \n\t"

  908.         : "+r" (h), "+r" (blk1), "+r" (blk2)

  909.         : "r" (stride));

  910. }

  911. 

  912. static inline void sad8_y2a_mmxext(uint8_t *blk1, uint8_t *blk2,

  913.                                    ptrdiff_t stride, int h)

  914. {

  915.     __asm__ volatile (

  916.         "movq (%1), %%mm0               \n\t"

  917.         "add %3, %1                     \n\t"

  918.         ".p2align 4                     \n\t"

  919.         "1:                             \n\t"

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

  921.         "movq (%1, %3), %%mm2           \n\t"

  922.         "pavgb %%mm1, %%mm0             \n\t"

  923.         "pavgb %%mm2, %%mm1             \n\t"

  924.         "psadbw (%2), %%mm0             \n\t"

  925.         "psadbw (%2, %3), %%mm1         \n\t"

  926.         "paddw %%mm0, %%mm6             \n\t"

  927.         "paddw %%mm1, %%mm6             \n\t"

  928.         "movq %%mm2, %%mm0              \n\t"

  929.         "lea (%1,%3,2), %1              \n\t"

  930.         "lea (%2,%3,2), %2              \n\t"

  931.         "sub $2, %0                     \n\t"

  932.         " jg 1b                         \n\t"

  933.         : "+r" (h), "+r" (blk1), "+r" (blk2)

  934.         : "r" (stride));

  935. }

  936. 

  937. static inline void sad8_4_mmxext(uint8_t *blk1, uint8_t *blk2,

  938.                                  ptrdiff_t stride, int h)

  939. {

  940.     __asm__ volatile (

  941.         "movq "MANGLE(bone)", %%mm5     \n\t"

  942.         "movq (%1), %%mm0               \n\t"

  943.         "pavgb 1(%1), %%mm0             \n\t"

  944.         "add %3, %1                     \n\t"

  945.         ".p2align 4                     \n\t"

  946.         "1:                             \n\t"

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

  948.         "movq (%1,%3), %%mm2            \n\t"

  949.         "pavgb 1(%1), %%mm1             \n\t"

  950.         "pavgb 1(%1,%3), %%mm2          \n\t"

  951.         "psubusb %%mm5, %%mm1           \n\t"

  952.         "pavgb %%mm1, %%mm0             \n\t"

  953.         "pavgb %%mm2, %%mm1             \n\t"

  954.         "psadbw (%2), %%mm0             \n\t"

  955.         "psadbw (%2,%3), %%mm1          \n\t"

  956.         "paddw %%mm0, %%mm6             \n\t"

  957.         "paddw %%mm1, %%mm6             \n\t"

  958.         "movq %%mm2, %%mm0              \n\t"

  959.         "lea (%1,%3,2), %1              \n\t"

  960.         "lea (%2,%3,2), %2              \n\t"

  961.         "sub $2, %0                     \n\t"

  962.         " jg 1b                         \n\t"

  963.         : "+r" (h), "+r" (blk1), "+r" (blk2)

  964.         : "r" (stride));

  965. }

  966. 

  967. static inline void sad8_2_mmx(uint8_t *blk1a, uint8_t *blk1b, uint8_t *blk2,

  968.                               ptrdiff_t stride, int h)

  969. {

  970.     x86_reg len = -(stride * h);

  971.     __asm__ volatile (

  972.         ".p2align 4                     \n\t"

  973.         "1:                             \n\t"

  974.         "movq (%1, %%"FF_REG_a"), %%mm0 \n\t"

  975.         "movq (%2, %%"FF_REG_a"), %%mm1 \n\t"

  976.         "movq (%1, %%"FF_REG_a"), %%mm2 \n\t"

  977.         "movq (%2, %%"FF_REG_a"), %%mm3 \n\t"

  978.         "punpcklbw %%mm7, %%mm0         \n\t"

  979.         "punpcklbw %%mm7, %%mm1         \n\t"

  980.         "punpckhbw %%mm7, %%mm2         \n\t"

  981.         "punpckhbw %%mm7, %%mm3         \n\t"

  982.         "paddw %%mm0, %%mm1             \n\t"

  983.         "paddw %%mm2, %%mm3             \n\t"

  984.         "movq (%3, %%"FF_REG_a"), %%mm4 \n\t"

  985.         "movq (%3, %%"FF_REG_a"), %%mm2 \n\t"

  986.         "paddw %%mm5, %%mm1             \n\t"

  987.         "paddw %%mm5, %%mm3             \n\t"

  988.         "psrlw $1, %%mm1                \n\t"

  989.         "psrlw $1, %%mm3                \n\t"

  990.         "packuswb %%mm3, %%mm1          \n\t"

  991.         "psubusb %%mm1, %%mm4           \n\t"

  992.         "psubusb %%mm2, %%mm1           \n\t"

  993.         "por %%mm4, %%mm1               \n\t"

  994.         "movq %%mm1, %%mm0              \n\t"

  995.         "punpcklbw %%mm7, %%mm0         \n\t"

  996.         "punpckhbw %%mm7, %%mm1         \n\t"

  997.         "paddw %%mm1, %%mm0             \n\t"

  998.         "paddw %%mm0, %%mm6             \n\t"

  999.         "add %4, %%"FF_REG_a"           \n\t"

  1000.         " js 1b                         \n\t"

  1001.         : "+a" (len)

  1002.         : "r" (blk1a - len), "r" (blk1b - len), "r" (blk2 - len),

  1003.           "r" (stride));

  1004. }

  1005. 

  1006. static inline void sad8_4_mmx(uint8_t *blk1, uint8_t *blk2,

  1007.                               ptrdiff_t stride, int h)

  1008. {

  1009.     x86_reg len = -(stride * h);

  1010.     __asm__ volatile (

  1011.         "movq  (%1, %%"FF_REG_a"), %%mm0\n\t"

  1012.         "movq 1(%1, %%"FF_REG_a"), %%mm2\n\t"

  1013.         "movq %%mm0, %%mm1              \n\t"

  1014.         "movq %%mm2, %%mm3              \n\t"

  1015.         "punpcklbw %%mm7, %%mm0         \n\t"

  1016.         "punpckhbw %%mm7, %%mm1         \n\t"

  1017.         "punpcklbw %%mm7, %%mm2         \n\t"

  1018.         "punpckhbw %%mm7, %%mm3         \n\t"

  1019.         "paddw %%mm2, %%mm0             \n\t"

  1020.         "paddw %%mm3, %%mm1             \n\t"

  1021.         ".p2align 4                     \n\t"

  1022.         "1:                             \n\t"

  1023.         "movq  (%2, %%"FF_REG_a"), %%mm2\n\t"

  1024.         "movq 1(%2, %%"FF_REG_a"), %%mm4\n\t"

  1025.         "movq %%mm2, %%mm3              \n\t"

  1026.         "movq %%mm4, %%mm5              \n\t"

  1027.         "punpcklbw %%mm7, %%mm2         \n\t"

  1028.         "punpckhbw %%mm7, %%mm3         \n\t"

  1029.         "punpcklbw %%mm7, %%mm4         \n\t"

  1030.         "punpckhbw %%mm7, %%mm5         \n\t"

  1031.         "paddw %%mm4, %%mm2             \n\t"

  1032.         "paddw %%mm5, %%mm3             \n\t"

  1033.         "movq 16+"MANGLE(round_tab)", %%mm5 \n\t"

  1034.         "paddw %%mm2, %%mm0             \n\t"

  1035.         "paddw %%mm3, %%mm1             \n\t"

  1036.         "paddw %%mm5, %%mm0             \n\t"

  1037.         "paddw %%mm5, %%mm1             \n\t"

  1038.         "movq (%3, %%"FF_REG_a"), %%mm4 \n\t"

  1039.         "movq (%3, %%"FF_REG_a"), %%mm5 \n\t"

  1040.         "psrlw $2, %%mm0                \n\t"

  1041.         "psrlw $2, %%mm1                \n\t"

  1042.         "packuswb %%mm1, %%mm0          \n\t"

  1043.         "psubusb %%mm0, %%mm4           \n\t"

  1044.         "psubusb %%mm5, %%mm0           \n\t"

  1045.         "por %%mm4, %%mm0               \n\t"

  1046.         "movq %%mm0, %%mm4              \n\t"

  1047.         "punpcklbw %%mm7, %%mm0         \n\t"

  1048.         "punpckhbw %%mm7, %%mm4         \n\t"

  1049.         "paddw %%mm0, %%mm6             \n\t"

  1050.         "paddw %%mm4, %%mm6             \n\t"

  1051.         "movq  %%mm2, %%mm0             \n\t"

  1052.         "movq  %%mm3, %%mm1             \n\t"

  1053.         "add %4, %%"FF_REG_a"           \n\t"

  1054.         " js 1b                         \n\t"

  1055.         : "+a" (len)

  1056.         : "r" (blk1 - len), "r" (blk1 - len + stride), "r" (blk2 - len),

  1057.           "r" (stride));

  1058. }

  1059. 

  1060. static inline int sum_mmx(void)

  1061. {

  1062.     int ret;

  1063.     __asm__ volatile (

  1064.         "movq %%mm6, %%mm0              \n\t"

  1065.         "psrlq $32, %%mm6               \n\t"

  1066.         "paddw %%mm0, %%mm6             \n\t"

  1067.         "movq %%mm6, %%mm0              \n\t"

  1068.         "psrlq $16, %%mm6               \n\t"

  1069.         "paddw %%mm0, %%mm6             \n\t"

  1070.         "movd %%mm6, %0                 \n\t"

  1071.         : "=r" (ret));

  1072.     return ret & 0xFFFF;

  1073. }

  1074. 

  1075. static inline int sum_mmxext(void)

  1076. {

  1077.     int ret;

  1078.     __asm__ volatile (

  1079.         "movd %%mm6, %0                 \n\t"

  1080.         : "=r" (ret));

  1081.     return ret;

  1082. }

  1083. 

  1084. static inline void sad8_x2a_mmx(uint8_t *blk1, uint8_t *blk2,

  1085.                                 ptrdiff_t stride, int h)

  1086. {

  1087.     sad8_2_mmx(blk1, blk1 + 1, blk2, stride, h);

  1088. }

  1089. 

  1090. static inline void sad8_y2a_mmx(uint8_t *blk1, uint8_t *blk2,

  1091.                                 ptrdiff_t stride, int h)

  1092. {

  1093.     sad8_2_mmx(blk1, blk1 + stride, blk2, stride, h);

  1094. }

  1095. 

  1096. #define PIX_SAD(suf)                                                    \

  1097. static int sad8_ ## suf(MpegEncContext *v, uint8_t *blk2,               \

  1098.                         uint8_t *blk1, ptrdiff_t stride, int h)         \

  1099. {                                                                       \

  1100.     assert(h == 8);                                                     \

  1101.     __asm__ volatile (                                                  \

  1102.         "pxor %%mm7, %%mm7     \n\t"                                    \

  1103.         "pxor %%mm6, %%mm6     \n\t"                                    \

  1104.         :);                                                             \

  1105.                                                                         \

  1106.     sad8_1_ ## suf(blk1, blk2, stride, 8);                              \

  1107.                                                                         \

  1108.     return sum_ ## suf();                                               \

  1109. }                                                                       \

  1110.                                                                         \

  1111. static int sad8_x2_ ## suf(MpegEncContext *v, uint8_t *blk2,            \

  1112.                            uint8_t *blk1, ptrdiff_t stride, int h)      \

  1113. {                                                                       \

  1114.     assert(h == 8);                                                     \

  1115.     __asm__ volatile (                                                  \

  1116.         "pxor %%mm7, %%mm7     \n\t"                                    \

  1117.         "pxor %%mm6, %%mm6     \n\t"                                    \

  1118.         "movq %0, %%mm5        \n\t"                                    \

  1119.         :: "m" (round_tab[1]));                                         \

  1120.                                                                         \

  1121.     sad8_x2a_ ## suf(blk1, blk2, stride, 8);                            \

  1122.                                                                         \

  1123.     return sum_ ## suf();                                               \

  1124. }                                                                       \

  1125.                                                                         \

  1126. static int sad8_y2_ ## suf(MpegEncContext *v, uint8_t *blk2,            \

  1127.                            uint8_t *blk1, ptrdiff_t stride, int h)      \

  1128. {                                                                       \

  1129.     assert(h == 8);                                                     \

  1130.     __asm__ volatile (                                                  \

  1131.         "pxor %%mm7, %%mm7     \n\t"                                    \

  1132.         "pxor %%mm6, %%mm6     \n\t"                                    \

  1133.         "movq %0, %%mm5        \n\t"                                    \

  1134.         :: "m" (round_tab[1]));                                         \

  1135.                                                                         \

  1136.     sad8_y2a_ ## suf(blk1, blk2, stride, 8);                            \

  1137.                                                                         \

  1138.     return sum_ ## suf();                                               \

  1139. }                                                                       \

  1140.                                                                         \

  1141. static int sad8_xy2_ ## suf(MpegEncContext *v, uint8_t *blk2,           \

  1142.                             uint8_t *blk1, ptrdiff_t stride, int h)     \

  1143. {                                                                       \

  1144.     assert(h == 8);                                                     \

  1145.     __asm__ volatile (                                                  \

  1146.         "pxor %%mm7, %%mm7     \n\t"                                    \

  1147.         "pxor %%mm6, %%mm6     \n\t"                                    \

  1148.         ::);                                                            \

  1149.                                                                         \

  1150.     sad8_4_ ## suf(blk1, blk2, stride, 8);                              \

  1151.                                                                         \

  1152.     return sum_ ## suf();                                               \

  1153. }                                                                       \

  1154.                                                                         \

  1155. static int sad16_ ## suf(MpegEncContext *v, uint8_t *blk2,              \

  1156.                          uint8_t *blk1, ptrdiff_t stride, int h)        \

  1157. {                                                                       \

  1158.     __asm__ volatile (                                                  \

  1159.         "pxor %%mm7, %%mm7     \n\t"                                    \

  1160.         "pxor %%mm6, %%mm6     \n\t"                                    \

  1161.         :);                                                             \

  1162.                                                                         \

  1163.     sad8_1_ ## suf(blk1,     blk2,     stride, h);                      \

  1164.     sad8_1_ ## suf(blk1 + 8, blk2 + 8, stride, h);                      \

  1165.                                                                         \

  1166.     return sum_ ## suf();                                               \

  1167. }                                                                       \

  1168.                                                                         \

  1169. static int sad16_x2_ ## suf(MpegEncContext *v, uint8_t *blk2,           \

  1170.                             uint8_t *blk1, ptrdiff_t stride, int h)     \

  1171. {                                                                       \

  1172.     __asm__ volatile (                                                  \

  1173.         "pxor %%mm7, %%mm7     \n\t"                                    \

  1174.         "pxor %%mm6, %%mm6     \n\t"                                    \

  1175.         "movq %0, %%mm5        \n\t"                                    \

  1176.         :: "m" (round_tab[1]));                                         \

  1177.                                                                         \

  1178.     sad8_x2a_ ## suf(blk1,     blk2,     stride, h);                    \

  1179.     sad8_x2a_ ## suf(blk1 + 8, blk2 + 8, stride, h);                    \

  1180.                                                                         \

  1181.     return sum_ ## suf();                                               \

  1182. }                                                                       \

  1183.                                                                         \

  1184. static int sad16_y2_ ## suf(MpegEncContext *v, uint8_t *blk2,           \

  1185.                             uint8_t *blk1, ptrdiff_t stride, int h)     \

  1186. {                                                                       \

  1187.     __asm__ volatile (                                                  \

  1188.         "pxor %%mm7, %%mm7     \n\t"                                    \

  1189.         "pxor %%mm6, %%mm6     \n\t"                                    \

  1190.         "movq %0, %%mm5        \n\t"                                    \

  1191.         :: "m" (round_tab[1]));                                         \

  1192.                                                                         \

  1193.     sad8_y2a_ ## suf(blk1,     blk2,     stride, h);                    \

  1194.     sad8_y2a_ ## suf(blk1 + 8, blk2 + 8, stride, h);                    \

  1195.                                                                         \

  1196.     return sum_ ## suf();                                               \

  1197. }                                                                       \

  1198.                                                                         \

  1199. static int sad16_xy2_ ## suf(MpegEncContext *v, uint8_t *blk2,          \

  1200.                              uint8_t *blk1, ptrdiff_t stride, int h)    \

  1201. {                                                                       \

  1202.     __asm__ volatile (                                                  \

  1203.         "pxor %%mm7, %%mm7     \n\t"                                    \

  1204.         "pxor %%mm6, %%mm6     \n\t"                                    \

  1205.         ::);                                                            \

  1206.                                                                         \

  1207.     sad8_4_ ## suf(blk1,     blk2,     stride, h);                      \

  1208.     sad8_4_ ## suf(blk1 + 8, blk2 + 8, stride, h);                      \

  1209.                                                                         \

  1210.     return sum_ ## suf();                                               \

  1211. }                                                                       \

  1212. 

  1213. PIX_SAD(mmx)

  1214. PIX_SAD(mmxext)

  1215. 

  1216. #endif /* HAVE_INLINE_ASM */

  1217. 

  1218. int ff_sse16_sse2(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,

  1219.                   ptrdiff_t stride, int h);

  1220. 

  1221. #define hadamard_func(cpu)                                                    \

  1222.     int ff_hadamard8_diff_ ## cpu(MpegEncContext *s, uint8_t *src1,           \

  1223.                                   uint8_t *src2, ptrdiff_t stride, int h);    \

  1224.     int ff_hadamard8_diff16_ ## cpu(MpegEncContext *s, uint8_t *src1,         \

  1225.                                     uint8_t *src2, ptrdiff_t stride, int h);

  1226. 

  1227. hadamard_func(mmx)

  1228. hadamard_func(mmxext)

  1229. hadamard_func(sse2)

  1230. hadamard_func(ssse3)

  1231. 

  1232. av_cold void ff_me_cmp_init_x86(MECmpContext *c, AVCodecContext *avctx)

  1233. {

  1234.     int cpu_flags = av_get_cpu_flags();

  1235. 

  1236. #if HAVE_INLINE_ASM

  1237.     if (INLINE_MMX(cpu_flags)) {

  1238.         c->sum_abs_dctelem = sum_abs_dctelem_mmx;

  1239. 

  1240.         c->pix_abs[0][0] = sad16_mmx;

  1241.         c->pix_abs[0][1] = sad16_x2_mmx;

  1242.         c->pix_abs[0][2] = sad16_y2_mmx;

  1243.         c->pix_abs[0][3] = sad16_xy2_mmx;

  1244.         c->pix_abs[1][0] = sad8_mmx;

  1245.         c->pix_abs[1][1] = sad8_x2_mmx;

  1246.         c->pix_abs[1][2] = sad8_y2_mmx;

  1247.         c->pix_abs[1][3] = sad8_xy2_mmx;

  1248. 

  1249.         c->sad[0] = sad16_mmx;

  1250.         c->sad[1] = sad8_mmx;

  1251. 

  1252.         c->sse[0]  = sse16_mmx;

  1253.         c->sse[1]  = sse8_mmx;

  1254.         c->vsad[4] = vsad_intra16_mmx;

  1255. 

  1256.         c->nsse[0] = nsse16_mmx;

  1257.         c->nsse[1] = nsse8_mmx;

  1258. 

  1259.         if (!(avctx->flags & AV_CODEC_FLAG_BITEXACT)) {

  1260.             c->vsad[0] = vsad16_mmx;

  1261.         }

  1262.     }

  1263. 

  1264.     if (INLINE_MMXEXT(cpu_flags)) {

  1265.         c->sum_abs_dctelem = sum_abs_dctelem_mmxext;

  1266. 

  1267.         c->vsad[4] = vsad_intra16_mmxext;

  1268. 

  1269.         c->pix_abs[0][0] = sad16_mmxext;

  1270.         c->pix_abs[1][0] = sad8_mmxext;

  1271. 

  1272.         c->sad[0] = sad16_mmxext;

  1273.         c->sad[1] = sad8_mmxext;

  1274. 

  1275.         if (!(avctx->flags & AV_CODEC_FLAG_BITEXACT)) {

  1276.             c->pix_abs[0][1] = sad16_x2_mmxext;

  1277.             c->pix_abs[0][2] = sad16_y2_mmxext;

  1278.             c->pix_abs[0][3] = sad16_xy2_mmxext;

  1279.             c->pix_abs[1][1] = sad8_x2_mmxext;

  1280.             c->pix_abs[1][2] = sad8_y2_mmxext;

  1281.             c->pix_abs[1][3] = sad8_xy2_mmxext;

  1282. 

  1283.             c->vsad[0] = vsad16_mmxext;

  1284.         }

  1285.     }

  1286. 

  1287.     if (INLINE_SSE2(cpu_flags)) {

  1288.         c->sum_abs_dctelem = sum_abs_dctelem_sse2;

  1289.     }

  1290. 

  1291.     if (INLINE_SSE2(cpu_flags) && !(cpu_flags & AV_CPU_FLAG_3DNOW)) {

  1292.         c->sad[0] = sad16_sse2;

  1293.     }

  1294. 

  1295. #if HAVE_SSSE3_INLINE

  1296.     if (INLINE_SSSE3(cpu_flags)) {

  1297.         c->sum_abs_dctelem = sum_abs_dctelem_ssse3;

  1298.     }

  1299. #endif

  1300. #endif /* HAVE_INLINE_ASM */

  1301. 

  1302.     if (EXTERNAL_MMX(cpu_flags)) {

  1303.         c->hadamard8_diff[0] = ff_hadamard8_diff16_mmx;

  1304.         c->hadamard8_diff[1] = ff_hadamard8_diff_mmx;

  1305.     }

  1306. 

  1307.     if (EXTERNAL_MMXEXT(cpu_flags)) {

  1308.         c->hadamard8_diff[0] = ff_hadamard8_diff16_mmxext;

  1309.         c->hadamard8_diff[1] = ff_hadamard8_diff_mmxext;

  1310.     }

  1311. 

  1312.     if (EXTERNAL_SSE2(cpu_flags)) {

  1313.         c->sse[0] = ff_sse16_sse2;

  1314. 

  1315. #if HAVE_ALIGNED_STACK

  1316.         c->hadamard8_diff[0] = ff_hadamard8_diff16_sse2;

  1317.         c->hadamard8_diff[1] = ff_hadamard8_diff_sse2;

  1318. #endif

  1319.     }

  1320. 

  1321.     if (EXTERNAL_SSSE3(cpu_flags) && HAVE_ALIGNED_STACK) {

  1322.         c->hadamard8_diff[0] = ff_hadamard8_diff16_ssse3;

  1323.         c->hadamard8_diff[1] = ff_hadamard8_diff_ssse3;

  1324.     }

  1325. }