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.
40175 Commits
32 Branches
844MB
Tree: a9aee08d90
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from 'a9aee08d90'
${ noResults }
FFmpeg/libavcodec/x86/dsputilenc_mmx.c

897 lines
27KB
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.  * 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/dsputil.h"

  30. #include "libavcodec/mpegvideo.h"

  31. #include "dsputil_x86.h"

  32. 

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

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

  35. void ff_diff_pixels_mmx(int16_t *block, const uint8_t *s1, const uint8_t *s2,

  36.                         int stride);

  37. 

  38. #if HAVE_INLINE_ASM

  39. 

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

  41.                     int line_size, int h)

  42. {

  43.     int tmp;

  44. 

  45.     __asm__ volatile (

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

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

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

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

  50.         "1:                              \n"

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

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

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

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

  55. 

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

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

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

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

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

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

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

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

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

  65. 

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

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

  68. 

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

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

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

  72. 

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

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

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

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

  77. 

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

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

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

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

  82. 

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

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

  85. 

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

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

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

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

  90. 

  91.         "decl      %%ecx                 \n"

  92.         "jnz       1b                    \n"

  93. 

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

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

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

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

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

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

  100.         : "%ecx");

  101. 

  102.     return tmp;

  103. }

  104. 

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

  106.                      int line_size, int h)

  107. {

  108.     int tmp;

  109. 

  110.     __asm__ volatile (

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

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

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

  114.         "1:\n"

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

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

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

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

  119. 

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

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

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

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

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

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

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

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

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

  129. 

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

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

  132. 

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

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

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

  136. 

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

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

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

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

  141. 

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

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

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

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

  146. 

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

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

  149. 

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

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

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

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

  154. 

  155.         "decl %%ecx\n"

  156.         "jnz 1b\n"

  157. 

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

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

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

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

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

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

  164.         : "%ecx");

  165. 

  166.     return tmp;

  167. }

  168. 

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

  170. {

  171.     int tmp;

  172. 

  173.     __asm__ volatile (

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

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

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

  177. 

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

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

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

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

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

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

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

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

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

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

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

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

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

  191. 

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

  193. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  219. 

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

  221.         "1:\n"

  222. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  248. 

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

  250. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  276. 

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

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

  279.         " jnz 1b\n"

  280. 

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

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

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

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

  285. 

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

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

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

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

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

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

  292.         : "%ecx");

  293. 

  294.     return tmp;

  295. }

  296. 

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

  298. {

  299.     int tmp;

  300.     uint8_t *pix = pix1;

  301. 

  302.     __asm__ volatile (

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

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

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

  306. 

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

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

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

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

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

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

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

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

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

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

  317. 

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

  319. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  342. 

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

  344.         "1:\n"

  345. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  368. 

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

  370. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  393. 

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

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

  396.         " jnz 1b\n"

  397. 

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

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

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

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

  402. 

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

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

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

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

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

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

  409.         : "%ecx");

  410. 

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

  412. }

  413. 

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

  415.                       int line_size, int h)

  416. {

  417.     int score1, score2;

  418. 

  419.     if (c)

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

  421.     else

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

  423.     score2 = hf_noise16_mmx(pix1, line_size, h) -

  424.              hf_noise16_mmx(pix2, line_size, h);

  425. 

  426.     if (c)

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

  428.     else

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

  430. }

  431. 

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

  433.                      int line_size, int h)

  434. {

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

  436.     int score2 = hf_noise8_mmx(pix1, line_size, h) -

  437.                  hf_noise8_mmx(pix2, line_size, h);

  438. 

  439.     if (c)

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

  441.     else

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

  443. }

  444. 

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

  446.                             int line_size, int h)

  447. {

  448.     int tmp;

  449. 

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

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

  452. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  475. 

  476. 

  477.     __asm__ volatile (

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

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

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

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

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

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

  484.         "jmp 2f\n"

  485.         "1:\n"

  486. 

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

  488.         "2:\n"

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

  490. 

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

  492.         "jnz 1b\n"

  493. 

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

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

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

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

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

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

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

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

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

  503.         : "%ecx");

  504. 

  505.     return tmp & 0xFFFF;

  506. }

  507. #undef SUM

  508. 

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

  510.                                int line_size, int h)

  511. {

  512.     int tmp;

  513. 

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

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

  516. 

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

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

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

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

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

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

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

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

  525. 

  526.     __asm__ volatile (

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

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

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

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

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

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

  533.         "jmp 2f\n"

  534.         "1:\n"

  535. 

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

  537.         "2:\n"

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

  539. 

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

  541.         "jnz 1b\n"

  542. 

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

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

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

  546.         : "%ecx");

  547. 

  548.     return tmp;

  549. }

  550. #undef SUM

  551. 

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

  553.                       int line_size, int h)

  554. {

  555.     int tmp;

  556. 

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

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

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

  560. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  590. 

  591. 

  592.     __asm__ volatile (

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  608.         "jmp 2f\n"

  609.         "1:\n"

  610. 

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

  612.         "2:\n"

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

  614. 

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

  616.         "jnz 1b\n"

  617. 

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

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

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

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

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

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

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

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

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

  627.         : "%ecx");

  628. 

  629.     return tmp & 0x7FFF;

  630. }

  631. #undef SUM

  632. 

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

  634.                          int line_size, int h)

  635. {

  636.     int tmp;

  637. 

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

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

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

  641. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  657. 

  658.     __asm__ volatile (

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  674.         "jmp 2f\n"

  675.         "1:\n"

  676. 

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

  678.         "2:\n"

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

  680. 

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

  682.         "jnz 1b\n"

  683. 

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

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

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

  687.         : "%ecx");

  688. 

  689.     return tmp;

  690. }

  691. #undef SUM

  692. 

  693. #define MMABS_MMX(a,z)                          \

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

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

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

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

  698. 

  699. #define MMABS_MMXEXT(a, z)                      \

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

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

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

  703. 

  704. #define MMABS_SSSE3(a,z)                        \

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

  706. 

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

  708.     MMABS(a,z)                                  \

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

  710. 

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

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

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

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

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

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

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

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

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

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

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

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

  723. 

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

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

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

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

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

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

  730. 

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

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

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

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

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

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

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

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

  739. 

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

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

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

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

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

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

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

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

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

  749. 

  750. #define DCT_SAD_MMX                             \

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

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

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

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

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

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

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

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

  759. 

  760. #define DCT_SAD_SSE2                            \

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

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

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

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

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

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

  767. 

  768. #define DCT_SAD_FUNC(cpu)                           \

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

  770. {                                                   \

  771.     int sum;                                        \

  772.     __asm__ volatile (                              \

  773.         DCT_SAD                                     \

  774.         :"=r"(sum)                                  \

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

  776.     return sum & 0xFFFF;                            \

  777. }

  778. 

  779. #define DCT_SAD         DCT_SAD_MMX

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

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

  782. DCT_SAD_FUNC(mmx)

  783. #undef MMABS

  784. #undef HSUM

  785. 

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

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

  788. DCT_SAD_FUNC(mmxext)

  789. #undef HSUM

  790. #undef DCT_SAD

  791. 

  792. #define DCT_SAD         DCT_SAD_SSE2

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

  794. DCT_SAD_FUNC(sse2)

  795. #undef MMABS

  796. 

  797. #if HAVE_SSSE3_INLINE

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

  799. DCT_SAD_FUNC(ssse3)

  800. #undef MMABS

  801. #endif

  802. #undef HSUM

  803. #undef DCT_SAD

  804. 

  805. #endif /* HAVE_INLINE_ASM */

  806. 

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

  808.                   int line_size, int h);

  809. 

  810. #define hadamard_func(cpu)                                              \

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

  812.                                   uint8_t *src2, int stride, int h);    \

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

  814.                                     uint8_t *src2, int stride, int h);

  815. 

  816. hadamard_func(mmx)

  817. hadamard_func(mmxext)

  818. hadamard_func(sse2)

  819. hadamard_func(ssse3)

  820. 

  821. av_cold void ff_dsputilenc_init_mmx(DSPContext *c, AVCodecContext *avctx,

  822.                                     unsigned high_bit_depth)

  823. {

  824.     int cpu_flags = av_get_cpu_flags();

  825. 

  826.     if (EXTERNAL_MMX(cpu_flags)) {

  827.         if (!high_bit_depth)

  828.             c->get_pixels = ff_get_pixels_mmx;

  829.         c->diff_pixels = ff_diff_pixels_mmx;

  830.     }

  831. 

  832.     if (EXTERNAL_SSE2(cpu_flags))

  833.         if (!high_bit_depth)

  834.             c->get_pixels = ff_get_pixels_sse2;

  835. 

  836. #if HAVE_INLINE_ASM

  837.     if (INLINE_MMX(cpu_flags)) {

  838.         c->sum_abs_dctelem = sum_abs_dctelem_mmx;

  839. 

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

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

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

  843. 

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

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

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

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

  848.         }

  849.     }

  850. 

  851.     if (INLINE_MMXEXT(cpu_flags)) {

  852.         c->sum_abs_dctelem = sum_abs_dctelem_mmxext;

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

  854. 

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

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

  857.         }

  858.     }

  859. 

  860.     if (INLINE_SSE2(cpu_flags)) {

  861.         c->sum_abs_dctelem = sum_abs_dctelem_sse2;

  862.     }

  863. 

  864. #if HAVE_SSSE3_INLINE

  865.     if (INLINE_SSSE3(cpu_flags)) {

  866.         c->sum_abs_dctelem = sum_abs_dctelem_ssse3;

  867.     }

  868. #endif

  869. #endif /* HAVE_INLINE_ASM */

  870. 

  871.     if (EXTERNAL_MMX(cpu_flags)) {

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

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

  874.     }

  875. 

  876.     if (EXTERNAL_MMXEXT(cpu_flags)) {

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

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

  879.     }

  880. 

  881.     if (EXTERNAL_SSE2(cpu_flags)) {

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

  883. 

  884. #if HAVE_ALIGNED_STACK

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

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

  887. #endif

  888.     }

  889. 

  890.     if (EXTERNAL_SSSE3(cpu_flags) && HAVE_ALIGNED_STACK) {

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

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

  893.     }

  894. 

  895.     ff_dsputil_init_pix_mmx(c, avctx);

  896. }