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.
40170 Commits
32 Branches
325MB
Tree: c166148409
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from 'c166148409'
${ noResults }
FFmpeg/libavcodec/x86/dsputilenc_mmx.c

911 lines
28KB
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/dct.h"

  30. #include "libavcodec/dsputil.h"

  31. #include "libavcodec/mpegvideo.h"

  32. #include "dsputil_x86.h"

  33. 

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

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

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

  37.                         int stride);

  38. 

  39. #if HAVE_INLINE_ASM

  40. 

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

  42.                     int line_size, int h)

  43. {

  44.     int tmp;

  45. 

  46.     __asm__ volatile (

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

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

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

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

  51.         "1:                              \n"

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

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

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

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

  56. 

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

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

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

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

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

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

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

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

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

  66. 

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

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

  69. 

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

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

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

  73. 

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

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

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

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

  78. 

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

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

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

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

  83. 

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

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

  86. 

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

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

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

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

  91. 

  92.         "decl      %%ecx                 \n"

  93.         "jnz       1b                    \n"

  94. 

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

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

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

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

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

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

  101.         : "%ecx");

  102. 

  103.     return tmp;

  104. }

  105. 

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

  107.                      int line_size, int h)

  108. {

  109.     int tmp;

  110. 

  111.     __asm__ volatile (

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

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

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

  115.         "1:\n"

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

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

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

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

  120. 

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

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

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

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

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

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

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

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

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

  130. 

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

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

  133. 

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

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

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

  137. 

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

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

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

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

  142. 

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

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

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

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

  147. 

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

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

  150. 

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

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

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

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

  155. 

  156.         "decl %%ecx\n"

  157.         "jnz 1b\n"

  158. 

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

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

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

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

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

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

  165.         : "%ecx");

  166. 

  167.     return tmp;

  168. }

  169. 

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

  171. {

  172.     int tmp;

  173. 

  174.     __asm__ volatile (

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

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

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

  178. 

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

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

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

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

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

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

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

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

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

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

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

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

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

  192. 

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

  194. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  220. 

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

  222.         "1:\n"

  223. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  249. 

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

  251. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  277. 

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

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

  280.         " jnz 1b\n"

  281. 

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

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

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

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

  286. 

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

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

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

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

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

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

  293.         : "%ecx");

  294. 

  295.     return tmp;

  296. }

  297. 

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

  299. {

  300.     int tmp;

  301.     uint8_t *pix = pix1;

  302. 

  303.     __asm__ volatile (

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

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

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

  307. 

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

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

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

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

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

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

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

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

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

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

  318. 

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

  320. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  343. 

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

  345.         "1:\n"

  346. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  369. 

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

  371. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  394. 

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

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

  397.         " jnz 1b\n"

  398. 

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

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

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

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

  403. 

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

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

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

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

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

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

  410.         : "%ecx");

  411. 

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

  413. }

  414. 

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

  416.                       int line_size, int h)

  417. {

  418.     int score1, score2;

  419. 

  420.     if (c)

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

  422.     else

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

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

  425.              hf_noise16_mmx(pix2, line_size, h);

  426. 

  427.     if (c)

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

  429.     else

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

  431. }

  432. 

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

  434.                      int line_size, int h)

  435. {

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

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

  438.                  hf_noise8_mmx(pix2, line_size, h);

  439. 

  440.     if (c)

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

  442.     else

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

  444. }

  445. 

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

  447.                             int line_size, int h)

  448. {

  449.     int tmp;

  450. 

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

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

  453. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  476. 

  477. 

  478.     __asm__ volatile (

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

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

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

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

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

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

  485.         "jmp 2f\n"

  486.         "1:\n"

  487. 

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

  489.         "2:\n"

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

  491. 

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

  493.         "jnz 1b\n"

  494. 

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

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

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

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

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

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

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

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

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

  504.         : "%ecx");

  505. 

  506.     return tmp & 0xFFFF;

  507. }

  508. #undef SUM

  509. 

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

  511.                                int line_size, int h)

  512. {

  513.     int tmp;

  514. 

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

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

  517. 

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

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

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

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

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

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

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

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

  526. 

  527.     __asm__ volatile (

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

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

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

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

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

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

  534.         "jmp 2f\n"

  535.         "1:\n"

  536. 

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

  538.         "2:\n"

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

  540. 

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

  542.         "jnz 1b\n"

  543. 

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

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

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

  547.         : "%ecx");

  548. 

  549.     return tmp;

  550. }

  551. #undef SUM

  552. 

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

  554.                       int line_size, int h)

  555. {

  556.     int tmp;

  557. 

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

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

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

  561. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  591. 

  592. 

  593.     __asm__ volatile (

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  609.         "jmp 2f\n"

  610.         "1:\n"

  611. 

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

  613.         "2:\n"

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

  615. 

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

  617.         "jnz 1b\n"

  618. 

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

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

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

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

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

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

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

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

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

  628.         : "%ecx");

  629. 

  630.     return tmp & 0x7FFF;

  631. }

  632. #undef SUM

  633. 

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

  635.                          int line_size, int h)

  636. {

  637.     int tmp;

  638. 

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

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

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

  642. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  658. 

  659.     __asm__ volatile (

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  675.         "jmp 2f\n"

  676.         "1:\n"

  677. 

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

  679.         "2:\n"

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

  681. 

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

  683.         "jnz 1b\n"

  684. 

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

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

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

  688.         : "%ecx");

  689. 

  690.     return tmp;

  691. }

  692. #undef SUM

  693. 

  694. #define MMABS_MMX(a,z)                          \

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

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

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

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

  699. 

  700. #define MMABS_MMXEXT(a, z)                      \

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

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

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

  704. 

  705. #define MMABS_SSSE3(a,z)                        \

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

  707. 

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

  709.     MMABS(a,z)                                  \

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

  711. 

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

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

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

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

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

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

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

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

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

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

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

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

  724. 

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

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

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

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

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

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

  731. 

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

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

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

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

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

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

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

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

  740. 

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

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

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

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

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

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

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

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

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

  750. 

  751. #define DCT_SAD_MMX                             \

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

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

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

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

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

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

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

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

  760. 

  761. #define DCT_SAD_SSE2                            \

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

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

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

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

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

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

  768. 

  769. #define DCT_SAD_FUNC(cpu)                           \

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

  771. {                                                   \

  772.     int sum;                                        \

  773.     __asm__ volatile (                              \

  774.         DCT_SAD                                     \

  775.         :"=r"(sum)                                  \

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

  777.     return sum & 0xFFFF;                            \

  778. }

  779. 

  780. #define DCT_SAD         DCT_SAD_MMX

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

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

  783. DCT_SAD_FUNC(mmx)

  784. #undef MMABS

  785. #undef HSUM

  786. 

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

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

  789. DCT_SAD_FUNC(mmxext)

  790. #undef HSUM

  791. #undef DCT_SAD

  792. 

  793. #define DCT_SAD         DCT_SAD_SSE2

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

  795. DCT_SAD_FUNC(sse2)

  796. #undef MMABS

  797. 

  798. #if HAVE_SSSE3_INLINE

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

  800. DCT_SAD_FUNC(ssse3)

  801. #undef MMABS

  802. #endif

  803. #undef HSUM

  804. #undef DCT_SAD

  805. 

  806. #endif /* HAVE_INLINE_ASM */

  807. 

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

  809.                   int line_size, int h);

  810. 

  811. #define hadamard_func(cpu)                                              \

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

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

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

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

  816. 

  817. hadamard_func(mmx)

  818. hadamard_func(mmxext)

  819. hadamard_func(sse2)

  820. hadamard_func(ssse3)

  821. 

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

  823.                                     unsigned high_bit_depth)

  824. {

  825.     int cpu_flags = av_get_cpu_flags();

  826.     const int dct_algo = avctx->dct_algo;

  827. 

  828.     if (EXTERNAL_MMX(cpu_flags)) {

  829.         if (!high_bit_depth)

  830.             c->get_pixels = ff_get_pixels_mmx;

  831.         c->diff_pixels = ff_diff_pixels_mmx;

  832.     }

  833. 

  834.     if (EXTERNAL_SSE2(cpu_flags))

  835.         if (!high_bit_depth)

  836.             c->get_pixels = ff_get_pixels_sse2;

  837. 

  838. #if HAVE_INLINE_ASM

  839.     if (INLINE_MMX(cpu_flags)) {

  840.         if (!high_bit_depth &&

  841.             (dct_algo == FF_DCT_AUTO || dct_algo == FF_DCT_MMX))

  842.             c->fdct = ff_fdct_mmx;

  843. 

  844.         c->sum_abs_dctelem = sum_abs_dctelem_mmx;

  845. 

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

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

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

  849. 

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

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

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

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

  854.         }

  855.     }

  856. 

  857.     if (INLINE_MMXEXT(cpu_flags)) {

  858.         if (!high_bit_depth &&

  859.             (dct_algo == FF_DCT_AUTO || dct_algo == FF_DCT_MMX))

  860.             c->fdct = ff_fdct_mmxext;

  861. 

  862.         c->sum_abs_dctelem = sum_abs_dctelem_mmxext;

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

  864. 

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

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

  867.         }

  868.     }

  869. 

  870.     if (INLINE_SSE2(cpu_flags)) {

  871.         if (!high_bit_depth &&

  872.             (dct_algo == FF_DCT_AUTO || dct_algo == FF_DCT_MMX))

  873.             c->fdct = ff_fdct_sse2;

  874. 

  875.         c->sum_abs_dctelem = sum_abs_dctelem_sse2;

  876.     }

  877. 

  878. #if HAVE_SSSE3_INLINE

  879.     if (INLINE_SSSE3(cpu_flags)) {

  880.         c->sum_abs_dctelem = sum_abs_dctelem_ssse3;

  881.     }

  882. #endif

  883. #endif /* HAVE_INLINE_ASM */

  884. 

  885.     if (EXTERNAL_MMX(cpu_flags)) {

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

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

  888.     }

  889. 

  890.     if (EXTERNAL_MMXEXT(cpu_flags)) {

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

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

  893.     }

  894. 

  895.     if (EXTERNAL_SSE2(cpu_flags)) {

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

  897. 

  898. #if HAVE_ALIGNED_STACK

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

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

  901. #endif

  902.     }

  903. 

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

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

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

  907.     }

  908. 

  909.     ff_dsputil_init_pix_mmx(c, avctx);

  910. }