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.
40197 Commits
32 Branches
872MB
Tree: f46bb608d9
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from 'f46bb608d9'
${ noResults }
FFmpeg/libavcodec/x86/dsputilenc_mmx.c

882 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. #if HAVE_INLINE_ASM

  34. 

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

  36.                     int line_size, int h)

  37. {

  38.     int tmp;

  39. 

  40.     __asm__ volatile (

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

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

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

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

  45.         "1:                              \n"

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

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

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

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

  50. 

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

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

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

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

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

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

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

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

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

  60. 

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

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

  63. 

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

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

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

  67. 

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

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

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

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

  72. 

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

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

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

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

  77. 

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

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

  80. 

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

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

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

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

  85. 

  86.         "decl      %%ecx                 \n"

  87.         "jnz       1b                    \n"

  88. 

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

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

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

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

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

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

  95.         : "%ecx");

  96. 

  97.     return tmp;

  98. }

  99. 

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

  101.                      int line_size, int h)

  102. {

  103.     int tmp;

  104. 

  105.     __asm__ volatile (

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

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

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

  109.         "1:\n"

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

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

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

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

  114. 

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

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

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

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

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

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

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

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

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

  124. 

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

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

  127. 

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

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

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

  131. 

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

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

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

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

  136. 

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

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

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

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

  141. 

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

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

  144. 

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

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

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

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

  149. 

  150.         "decl %%ecx\n"

  151.         "jnz 1b\n"

  152. 

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

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

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

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

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

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

  159.         : "%ecx");

  160. 

  161.     return tmp;

  162. }

  163. 

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

  165. {

  166.     int tmp;

  167. 

  168.     __asm__ volatile (

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

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

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

  172. 

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

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

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

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

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

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

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

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

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

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

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

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

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

  186. 

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

  188. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  214. 

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

  216.         "1:\n"

  217. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  243. 

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

  245. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  271. 

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

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

  274.         " jnz 1b\n"

  275. 

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

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

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

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

  280. 

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

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

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

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

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

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

  287.         : "%ecx");

  288. 

  289.     return tmp;

  290. }

  291. 

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

  293. {

  294.     int tmp;

  295.     uint8_t *pix = pix1;

  296. 

  297.     __asm__ volatile (

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

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

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

  301. 

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

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

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

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

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

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

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

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

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

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

  312. 

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

  314. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  337. 

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

  339.         "1:\n"

  340. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  363. 

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

  365. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  388. 

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

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

  391.         " jnz 1b\n"

  392. 

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

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

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

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

  397. 

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

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

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

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

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

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

  404.         : "%ecx");

  405. 

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

  407. }

  408. 

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

  410.                       int line_size, int h)

  411. {

  412.     int score1, score2;

  413. 

  414.     if (c)

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

  416.     else

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

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

  419.              hf_noise16_mmx(pix2, line_size, h);

  420. 

  421.     if (c)

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

  423.     else

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

  425. }

  426. 

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

  428.                      int line_size, int h)

  429. {

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

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

  432.                  hf_noise8_mmx(pix2, line_size, h);

  433. 

  434.     if (c)

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

  436.     else

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

  438. }

  439. 

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

  441.                             int line_size, int h)

  442. {

  443.     int tmp;

  444. 

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

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

  447. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  470. 

  471. 

  472.     __asm__ volatile (

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

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

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

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

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

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

  479.         "jmp 2f\n"

  480.         "1:\n"

  481. 

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

  483.         "2:\n"

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

  485. 

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

  487.         "jnz 1b\n"

  488. 

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

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

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

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

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

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

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

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

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

  498.         : "%ecx");

  499. 

  500.     return tmp & 0xFFFF;

  501. }

  502. #undef SUM

  503. 

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

  505.                                int line_size, int h)

  506. {

  507.     int tmp;

  508. 

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

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

  511. 

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

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

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

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

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

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

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

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

  520. 

  521.     __asm__ volatile (

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

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

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

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

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

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

  528.         "jmp 2f\n"

  529.         "1:\n"

  530. 

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

  532.         "2:\n"

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

  534. 

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

  536.         "jnz 1b\n"

  537. 

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

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

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

  541.         : "%ecx");

  542. 

  543.     return tmp;

  544. }

  545. #undef SUM

  546. 

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

  548.                       int line_size, int h)

  549. {

  550.     int tmp;

  551. 

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

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

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

  555. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  585. 

  586. 

  587.     __asm__ volatile (

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  603.         "jmp 2f\n"

  604.         "1:\n"

  605. 

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

  607.         "2:\n"

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

  609. 

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

  611.         "jnz 1b\n"

  612. 

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

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

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

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

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

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

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

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

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

  622.         : "%ecx");

  623. 

  624.     return tmp & 0x7FFF;

  625. }

  626. #undef SUM

  627. 

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

  629.                          int line_size, int h)

  630. {

  631.     int tmp;

  632. 

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

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

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

  636. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  652. 

  653.     __asm__ volatile (

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  669.         "jmp 2f\n"

  670.         "1:\n"

  671. 

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

  673.         "2:\n"

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

  675. 

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

  677.         "jnz 1b\n"

  678. 

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

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

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

  682.         : "%ecx");

  683. 

  684.     return tmp;

  685. }

  686. #undef SUM

  687. 

  688. #define MMABS_MMX(a,z)                          \

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

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

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

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

  693. 

  694. #define MMABS_MMXEXT(a, z)                      \

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

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

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

  698. 

  699. #define MMABS_SSSE3(a,z)                        \

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

  701. 

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

  703.     MMABS(a,z)                                  \

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

  705. 

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

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

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

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

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

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

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

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

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

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

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

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

  718. 

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

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

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

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

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

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

  725. 

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

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

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

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

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

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

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

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

  734. 

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

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

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

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

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

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

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

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

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

  744. 

  745. #define DCT_SAD_MMX                             \

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

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

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

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

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

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

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

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

  754. 

  755. #define DCT_SAD_SSE2                            \

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

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

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

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

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

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

  762. 

  763. #define DCT_SAD_FUNC(cpu)                           \

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

  765. {                                                   \

  766.     int sum;                                        \

  767.     __asm__ volatile (                              \

  768.         DCT_SAD                                     \

  769.         :"=r"(sum)                                  \

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

  771.     return sum & 0xFFFF;                            \

  772. }

  773. 

  774. #define DCT_SAD         DCT_SAD_MMX

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

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

  777. DCT_SAD_FUNC(mmx)

  778. #undef MMABS

  779. #undef HSUM

  780. 

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

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

  783. DCT_SAD_FUNC(mmxext)

  784. #undef HSUM

  785. #undef DCT_SAD

  786. 

  787. #define DCT_SAD         DCT_SAD_SSE2

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

  789. DCT_SAD_FUNC(sse2)

  790. #undef MMABS

  791. 

  792. #if HAVE_SSSE3_INLINE

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

  794. DCT_SAD_FUNC(ssse3)

  795. #undef MMABS

  796. #endif

  797. #undef HSUM

  798. #undef DCT_SAD

  799. 

  800. #endif /* HAVE_INLINE_ASM */

  801. 

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

  803.                   int line_size, int h);

  804. 

  805. #define hadamard_func(cpu)                                              \

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

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

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

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

  810. 

  811. hadamard_func(mmx)

  812. hadamard_func(mmxext)

  813. hadamard_func(sse2)

  814. hadamard_func(ssse3)

  815. 

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

  817.                                     unsigned high_bit_depth)

  818. {

  819.     int cpu_flags = av_get_cpu_flags();

  820. 

  821. #if HAVE_INLINE_ASM

  822.     if (INLINE_MMX(cpu_flags)) {

  823.         c->sum_abs_dctelem = sum_abs_dctelem_mmx;

  824. 

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

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

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

  828. 

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

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

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

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

  833.         }

  834.     }

  835. 

  836.     if (INLINE_MMXEXT(cpu_flags)) {

  837.         c->sum_abs_dctelem = sum_abs_dctelem_mmxext;

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

  839. 

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

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

  842.         }

  843.     }

  844. 

  845.     if (INLINE_SSE2(cpu_flags)) {

  846.         c->sum_abs_dctelem = sum_abs_dctelem_sse2;

  847.     }

  848. 

  849. #if HAVE_SSSE3_INLINE

  850.     if (INLINE_SSSE3(cpu_flags)) {

  851.         c->sum_abs_dctelem = sum_abs_dctelem_ssse3;

  852.     }

  853. #endif

  854. #endif /* HAVE_INLINE_ASM */

  855. 

  856.     if (EXTERNAL_MMX(cpu_flags)) {

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

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

  859.     }

  860. 

  861.     if (EXTERNAL_MMXEXT(cpu_flags)) {

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

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

  864.     }

  865. 

  866.     if (EXTERNAL_SSE2(cpu_flags)) {

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

  868. 

  869. #if HAVE_ALIGNED_STACK

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

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

  872. #endif

  873.     }

  874. 

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

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

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

  878.     }

  879. 

  880.     ff_dsputil_init_pix_mmx(c, avctx);

  881. }