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.
35406 Commits
32 Branches
821MB
Tree: c3e15f7b39
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from 'c3e15f7b39'
${ noResults }
FFmpeg/libavcodec/x86/dsputilenc_mmx.c

1212 lines
35KB
Raw Blame History 

  1. /*

  2.  * MMX optimized DSP utils

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

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

  5.  *

  6.  * This file is part of Libav.

  7.  *

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

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

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

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

  12.  *

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

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

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

  16.  * Lesser General Public License for more details.

  17.  *

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

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

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

  21.  *

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

  23.  */

  24. 

  25. #include "libavutil/cpu.h"

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

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

  28. #include "libavcodec/dsputil.h"

  29. #include "libavcodec/mpegvideo.h"

  30. #include "libavcodec/mathops.h"

  31. #include "dsputil_mmx.h"

  32. 

  33. 

  34. #if HAVE_INLINE_ASM

  35. 

  36. static void get_pixels_mmx(DCTELEM *block, const uint8_t *pixels, int line_size)

  37. {

  38.     __asm__ volatile(

  39.         "mov $-128, %%"REG_a"           \n\t"

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

  41.         ".p2align 4                     \n\t"

  42.         "1:                             \n\t"

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

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

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

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

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

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

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

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

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

  52.         "movq %%mm1, 8(%1, %%"REG_a")   \n\t"

  53.         "movq %%mm2, 16(%1, %%"REG_a")  \n\t"

  54.         "movq %%mm3, 24(%1, %%"REG_a")  \n\t"

  55.         "add %3, %0                     \n\t"

  56.         "add $32, %%"REG_a"             \n\t"

  57.         "js 1b                          \n\t"

  58.         : "+r" (pixels)

  59.         : "r" (block+64), "r" ((x86_reg)line_size), "r" ((x86_reg)line_size*2)

  60.         : "%"REG_a

  61.     );

  62. }

  63. 

  64. static void get_pixels_sse2(DCTELEM *block, const uint8_t *pixels, int line_size)

  65. {

  66.     __asm__ volatile(

  67.         "pxor %%xmm4,      %%xmm4         \n\t"

  68.         "movq (%0),        %%xmm0         \n\t"

  69.         "movq (%0, %2),    %%xmm1         \n\t"

  70.         "movq (%0, %2,2),  %%xmm2         \n\t"

  71.         "movq (%0, %3),    %%xmm3         \n\t"

  72.         "lea (%0,%2,4), %0                \n\t"

  73.         "punpcklbw %%xmm4, %%xmm0         \n\t"

  74.         "punpcklbw %%xmm4, %%xmm1         \n\t"

  75.         "punpcklbw %%xmm4, %%xmm2         \n\t"

  76.         "punpcklbw %%xmm4, %%xmm3         \n\t"

  77.         "movdqa %%xmm0,      (%1)         \n\t"

  78.         "movdqa %%xmm1,    16(%1)         \n\t"

  79.         "movdqa %%xmm2,    32(%1)         \n\t"

  80.         "movdqa %%xmm3,    48(%1)         \n\t"

  81.         "movq (%0),        %%xmm0         \n\t"

  82.         "movq (%0, %2),    %%xmm1         \n\t"

  83.         "movq (%0, %2,2),  %%xmm2         \n\t"

  84.         "movq (%0, %3),    %%xmm3         \n\t"

  85.         "punpcklbw %%xmm4, %%xmm0         \n\t"

  86.         "punpcklbw %%xmm4, %%xmm1         \n\t"

  87.         "punpcklbw %%xmm4, %%xmm2         \n\t"

  88.         "punpcklbw %%xmm4, %%xmm3         \n\t"

  89.         "movdqa %%xmm0,    64(%1)         \n\t"

  90.         "movdqa %%xmm1,    80(%1)         \n\t"

  91.         "movdqa %%xmm2,    96(%1)         \n\t"

  92.         "movdqa %%xmm3,   112(%1)         \n\t"

  93.         : "+r" (pixels)

  94.         : "r" (block), "r" ((x86_reg)line_size), "r" ((x86_reg)line_size*3)

  95.     );

  96. }

  97. 

  98. static inline void diff_pixels_mmx(DCTELEM *block, const uint8_t *s1, const uint8_t *s2, int stride)

  99. {

  100.     __asm__ volatile(

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

  102.         "mov $-128, %%"REG_a"           \n\t"

  103.         ".p2align 4                     \n\t"

  104.         "1:                             \n\t"

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

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

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

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

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

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

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

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

  113.         "psubw %%mm2, %%mm0             \n\t"

  114.         "psubw %%mm3, %%mm1             \n\t"

  115.         "movq %%mm0, (%2, %%"REG_a")    \n\t"

  116.         "movq %%mm1, 8(%2, %%"REG_a")   \n\t"

  117.         "add %3, %0                     \n\t"

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

  119.         "add $16, %%"REG_a"             \n\t"

  120.         "jnz 1b                         \n\t"

  121.         : "+r" (s1), "+r" (s2)

  122.         : "r" (block+64), "r" ((x86_reg)stride)

  123.         : "%"REG_a

  124.     );

  125. }

  126. 

  127. static int pix_sum16_mmx(uint8_t * pix, int line_size){

  128.     const int h=16;

  129.     int sum;

  130.     x86_reg index= -line_size*h;

  131. 

  132.     __asm__ volatile(

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

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

  135.                 "1:                             \n\t"

  136.                 "movq (%2, %1), %%mm0           \n\t"

  137.                 "movq (%2, %1), %%mm1           \n\t"

  138.                 "movq 8(%2, %1), %%mm2          \n\t"

  139.                 "movq 8(%2, %1), %%mm3          \n\t"

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

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

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

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

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

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

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

  147.                 "paddw %%mm3, %%mm6             \n\t"

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

  149.                 " js 1b                         \n\t"

  150.                 "movq %%mm6, %%mm5              \n\t"

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

  152.                 "paddw %%mm5, %%mm6             \n\t"

  153.                 "movq %%mm6, %%mm5              \n\t"

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

  155.                 "paddw %%mm5, %%mm6             \n\t"

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

  157.                 "andl $0xFFFF, %0               \n\t"

  158.                 : "=&r" (sum), "+r" (index)

  159.                 : "r" (pix - index), "r" ((x86_reg)line_size)

  160.         );

  161. 

  162.         return sum;

  163. }

  164. 

  165. static int pix_norm1_mmx(uint8_t *pix, int line_size) {

  166.     int tmp;

  167.   __asm__ volatile (

  168.       "movl $16,%%ecx\n"

  169.       "pxor %%mm0,%%mm0\n"

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

  171.       "1:\n"

  172.       "movq (%0),%%mm2\n"       /* mm2 = pix[0-7] */

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

  174. 

  175.       "movq %%mm2,%%mm1\n"      /* mm1 = mm2 = pix[0-7] */

  176. 

  177.       "punpckhbw %%mm0,%%mm1\n" /* mm1 = [pix4-7] */

  178.       "punpcklbw %%mm0,%%mm2\n" /* mm2 = [pix0-3] */

  179. 

  180.       "movq %%mm3,%%mm4\n"      /* mm4 = mm3 = pix[8-15] */

  181.       "punpckhbw %%mm0,%%mm3\n" /* mm3 = [pix12-15] */

  182.       "punpcklbw %%mm0,%%mm4\n" /* mm4 = [pix8-11] */

  183. 

  184.       "pmaddwd %%mm1,%%mm1\n"   /* mm1 = (pix0^2+pix1^2,pix2^2+pix3^2) */

  185.       "pmaddwd %%mm2,%%mm2\n"   /* mm2 = (pix4^2+pix5^2,pix6^2+pix7^2) */

  186. 

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

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

  189. 

  190.       "paddd %%mm1,%%mm2\n"     /* mm2 = (pix0^2+pix1^2+pix4^2+pix5^2,

  191.                                           pix2^2+pix3^2+pix6^2+pix7^2) */

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

  193.       "paddd %%mm2,%%mm7\n"

  194. 

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

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

  197.       "dec %%ecx\n"

  198.       "jnz 1b\n"

  199. 

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

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

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

  203.       "movd %%mm1,%1\n"

  204.       : "+r" (pix), "=r"(tmp) : "r" ((x86_reg)line_size) : "%ecx" );

  205.     return tmp;

  206. }

  207. 

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

  209.     int tmp;

  210.   __asm__ volatile (

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

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

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

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

  215.       "1:\n"

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

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

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

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

  220. 

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

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

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

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

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

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

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

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

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

  230. 

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

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

  233. 

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

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

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

  237. 

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

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

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

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

  242. 

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

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

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

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

  247. 

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

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

  250. 

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

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

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

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

  255. 

  256.       "decl %%ecx\n"

  257.       "jnz 1b\n"

  258. 

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

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

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

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

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

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

  265.       : "%ecx");

  266.     return tmp;

  267. }

  268. 

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

  270.     int tmp;

  271.   __asm__ volatile (

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

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

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

  275.       "1:\n"

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

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

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

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

  280. 

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

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

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

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

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

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

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

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

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

  290. 

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

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

  293. 

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

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

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

  297. 

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

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

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

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

  302. 

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

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

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

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

  307. 

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

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

  310. 

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

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

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

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

  315. 

  316.       "decl %%ecx\n"

  317.       "jnz 1b\n"

  318. 

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

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

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

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

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

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

  325.       : "%ecx");

  326.     return tmp;

  327. }

  328. 

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

  330.     int tmp;

  331.   __asm__ volatile (

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

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

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

  335. 

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

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

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

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

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

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

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

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

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

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

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

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

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

  349. 

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

  351. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  377. 

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

  379.       "1:\n"

  380. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  406. 

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

  408. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  434. 

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

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

  437.       " jnz 1b\n"

  438. 

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

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

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

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

  443. 

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

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

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

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

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

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

  450.       : "%ecx");

  451.       return tmp;

  452. }

  453. 

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

  455.     int tmp;

  456.     uint8_t * pix= pix1;

  457.   __asm__ volatile (

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

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

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

  461. 

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

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

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

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

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

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

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

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

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

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

  472. 

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

  474. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  497. 

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

  499.       "1:\n"

  500. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  523. 

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

  525. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  548. 

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

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

  551.       " jnz 1b\n"

  552. 

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

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

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

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

  557. 

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

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

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

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

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

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

  564.       : "%ecx");

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

  566. }

  567. 

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

  569.     MpegEncContext *c = p;

  570.     int score1, score2;

  571. 

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

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

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

  575. 

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

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

  578. }

  579. 

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

  581.     MpegEncContext *c = p;

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

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

  584. 

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

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

  587. }

  588. 

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

  590.     int tmp;

  591. 

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

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

  594. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  617. 

  618. 

  619.   __asm__ volatile (

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

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

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

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

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

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

  626.       "jmp 2f\n"

  627.       "1:\n"

  628. 

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

  630.       "2:\n"

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

  632. 

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

  634.       "jnz 1b\n"

  635. 

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

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

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

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

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

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

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

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

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

  645.       : "%ecx");

  646.     return tmp & 0xFFFF;

  647. }

  648. #undef SUM

  649. 

  650. static int vsad_intra16_mmx2(void *v, uint8_t * pix, uint8_t * dummy, int line_size, int h) {

  651.     int tmp;

  652. 

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

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

  655. 

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

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

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

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

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

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

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

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

  664. 

  665.   __asm__ volatile (

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

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

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

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

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

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

  672.       "jmp 2f\n"

  673.       "1:\n"

  674. 

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

  676.       "2:\n"

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

  678. 

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

  680.       "jnz 1b\n"

  681. 

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

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

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

  685.       : "%ecx");

  686.     return tmp;

  687. }

  688. #undef SUM

  689. 

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

  691.     int tmp;

  692. 

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

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

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

  696. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  726. 

  727. 

  728.   __asm__ volatile (

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  744.       "jmp 2f\n"

  745.       "1:\n"

  746. 

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

  748.       "2:\n"

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

  750. 

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

  752.       "jnz 1b\n"

  753. 

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

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

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

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

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

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

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

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

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

  763.       : "%ecx");

  764.     return tmp & 0x7FFF;

  765. }

  766. #undef SUM

  767. 

  768. static int vsad16_mmx2(void *v, uint8_t * pix1, uint8_t * pix2, int line_size, int h) {

  769.     int tmp;

  770. 

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

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

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

  774. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  790. 

  791.   __asm__ volatile (

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  807.       "jmp 2f\n"

  808.       "1:\n"

  809. 

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

  811.       "2:\n"

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

  813. 

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

  815.       "jnz 1b\n"

  816. 

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

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

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

  820.       : "%ecx");

  821.     return tmp;

  822. }

  823. #undef SUM

  824. 

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

  826.     x86_reg i=0;

  827.     __asm__ volatile(

  828.         "1:                             \n\t"

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

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

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

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

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

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

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

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

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

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

  839.         " jb 1b                         \n\t"

  840.         : "+r" (i)

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

  842.     );

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

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

  845. }

  846. 

  847. static void sub_hfyu_median_prediction_mmx2(uint8_t *dst, const uint8_t *src1, const uint8_t *src2, int w, int *left, int *left_top){

  848.     x86_reg i=0;

  849.     uint8_t l, lt;

  850. 

  851.     __asm__ volatile(

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

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

  854.         "1:                             \n\t"

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  871.         " jb 1b                         \n\t"

  872.         : "+r" (i)

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

  874.     );

  875. 

  876.     l= *left;

  877.     lt= *left_top;

  878. 

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

  880. 

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

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

  883. }

  884. 

  885. #define MMABS_MMX(a,z)\

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

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

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

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

  890. 

  891. #define MMABS_MMX2(a,z)\

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

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

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

  895. 

  896. #define MMABS_SSSE3(a,z)\

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

  898. 

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

  900.     MMABS(a,z)\

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

  902. 

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

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

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

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

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

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

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

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

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

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

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

  914. 

  915. #define HSUM_MMX2(a, t, dst)\

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

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

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

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

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

  921. 

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

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

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

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

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

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

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

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

  930. 

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

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

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

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

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

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

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

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

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

  940. 

  941. #define DCT_SAD_MMX\

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

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

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

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

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

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

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

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

  950. 

  951. #define DCT_SAD_SSE2\

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

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

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

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

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

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

  958. 

  959. #define DCT_SAD_FUNC(cpu) \

  960. static int sum_abs_dctelem_##cpu(DCTELEM *block){\

  961.     int sum;\

  962.     __asm__ volatile(\

  963.         DCT_SAD\

  964.         :"=r"(sum)\

  965.         :"r"(block)\

  966.     );\

  967.     return sum&0xFFFF;\

  968. }

  969. 

  970. #define DCT_SAD       DCT_SAD_MMX

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

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

  973. DCT_SAD_FUNC(mmx)

  974. #undef MMABS

  975. #undef HSUM

  976. 

  977. #define HSUM(a,t,dst) HSUM_MMX2(a,t,dst)

  978. #define MMABS(a,z)    MMABS_MMX2(a,z)

  979. DCT_SAD_FUNC(mmx2)

  980. #undef HSUM

  981. #undef DCT_SAD

  982. 

  983. #define DCT_SAD       DCT_SAD_SSE2

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

  985. DCT_SAD_FUNC(sse2)

  986. #undef MMABS

  987. 

  988. #if HAVE_SSSE3_INLINE

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

  990. DCT_SAD_FUNC(ssse3)

  991. #undef MMABS

  992. #endif

  993. #undef HSUM

  994. #undef DCT_SAD

  995. 

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

  997.     int sum;

  998.     x86_reg i=size;

  999.     __asm__ volatile(

  1000.         "pxor %%mm4, %%mm4 \n"

  1001.         "1: \n"

  1002.         "sub $8, %0 \n"

  1003.         "movq (%2,%0), %%mm2 \n"

  1004.         "movq (%3,%0,2), %%mm0 \n"

  1005.         "movq 8(%3,%0,2), %%mm1 \n"

  1006.         "punpckhbw %%mm2, %%mm3 \n"

  1007.         "punpcklbw %%mm2, %%mm2 \n"

  1008.         "psraw $8, %%mm3 \n"

  1009.         "psraw $8, %%mm2 \n"

  1010.         "psubw %%mm3, %%mm1 \n"

  1011.         "psubw %%mm2, %%mm0 \n"

  1012.         "pmaddwd %%mm1, %%mm1 \n"

  1013.         "pmaddwd %%mm0, %%mm0 \n"

  1014.         "paddd %%mm1, %%mm4 \n"

  1015.         "paddd %%mm0, %%mm4 \n"

  1016.         "jg 1b \n"

  1017.         "movq %%mm4, %%mm3 \n"

  1018.         "psrlq $32, %%mm3 \n"

  1019.         "paddd %%mm3, %%mm4 \n"

  1020.         "movd %%mm4, %1 \n"

  1021.         :"+r"(i), "=r"(sum)

  1022.         :"r"(pix1), "r"(pix2)

  1023.     );

  1024.     return sum;

  1025. }

  1026. 

  1027. #define PHADDD(a, t)\

  1028.     "movq "#a", "#t"                  \n\t"\

  1029.     "psrlq $32, "#a"                  \n\t"\

  1030.     "paddd "#t", "#a"                 \n\t"

  1031. /*

  1032.    pmulhw: dst[0-15]=(src[0-15]*dst[0-15])[16-31]

  1033.    pmulhrw: dst[0-15]=(src[0-15]*dst[0-15] + 0x8000)[16-31]

  1034.    pmulhrsw: dst[0-15]=(src[0-15]*dst[0-15] + 0x4000)[15-30]

  1035.  */

  1036. #define PMULHRW(x, y, s, o)\

  1037.     "pmulhw " #s ", "#x "            \n\t"\

  1038.     "pmulhw " #s ", "#y "            \n\t"\

  1039.     "paddw " #o ", "#x "             \n\t"\

  1040.     "paddw " #o ", "#y "             \n\t"\

  1041.     "psraw $1, "#x "                 \n\t"\

  1042.     "psraw $1, "#y "                 \n\t"

  1043. #define DEF(x) x ## _mmx

  1044. #define SET_RND MOVQ_WONE

  1045. #define SCALE_OFFSET 1

  1046. 

  1047. #include "dsputil_qns_template.c"

  1048. 

  1049. #undef DEF

  1050. #undef SET_RND

  1051. #undef SCALE_OFFSET

  1052. #undef PMULHRW

  1053. 

  1054. #define DEF(x) x ## _3dnow

  1055. #define SET_RND(x)

  1056. #define SCALE_OFFSET 0

  1057. #define PMULHRW(x, y, s, o)\

  1058.     "pmulhrw " #s ", "#x "           \n\t"\

  1059.     "pmulhrw " #s ", "#y "           \n\t"

  1060. 

  1061. #include "dsputil_qns_template.c"

  1062. 

  1063. #undef DEF

  1064. #undef SET_RND

  1065. #undef SCALE_OFFSET

  1066. #undef PMULHRW

  1067. 

  1068. #if HAVE_SSSE3_INLINE

  1069. #undef PHADDD

  1070. #define DEF(x) x ## _ssse3

  1071. #define SET_RND(x)

  1072. #define SCALE_OFFSET -1

  1073. #define PHADDD(a, t)\

  1074.     "pshufw $0x0E, "#a", "#t"         \n\t"\

  1075.     "paddd "#t", "#a"                 \n\t" /* faster than phaddd on core2 */

  1076. #define PMULHRW(x, y, s, o)\

  1077.     "pmulhrsw " #s ", "#x "          \n\t"\

  1078.     "pmulhrsw " #s ", "#y "          \n\t"

  1079. 

  1080. #include "dsputil_qns_template.c"

  1081. 

  1082. #undef DEF

  1083. #undef SET_RND

  1084. #undef SCALE_OFFSET

  1085. #undef PMULHRW

  1086. #undef PHADDD

  1087. #endif /* HAVE_SSSE3_INLINE */

  1088. 

  1089. #endif /* HAVE_INLINE_ASM */

  1090. 

  1091. int ff_sse16_sse2(void *v, uint8_t * pix1, uint8_t * pix2, int line_size, int h);

  1092. 

  1093. #define hadamard_func(cpu) \

  1094. int ff_hadamard8_diff_##cpu  (void *s, uint8_t *src1, uint8_t *src2, \

  1095.                               int stride, int h); \

  1096. int ff_hadamard8_diff16_##cpu(void *s, uint8_t *src1, uint8_t *src2, \

  1097.                               int stride, int h);

  1098. 

  1099. hadamard_func(mmx)

  1100. hadamard_func(mmx2)

  1101. hadamard_func(sse2)

  1102. hadamard_func(ssse3)

  1103. 

  1104. void ff_dsputilenc_init_mmx(DSPContext* c, AVCodecContext *avctx)

  1105. {

  1106.     int mm_flags = av_get_cpu_flags();

  1107. 

  1108. #if HAVE_INLINE_ASM

  1109.     int bit_depth = avctx->bits_per_raw_sample;

  1110. 

  1111.     if (mm_flags & AV_CPU_FLAG_MMX) {

  1112.         const int dct_algo = avctx->dct_algo;

  1113.         if (avctx->bits_per_raw_sample <= 8 &&

  1114.             (dct_algo==FF_DCT_AUTO || dct_algo==FF_DCT_MMX)) {

  1115.             if(mm_flags & AV_CPU_FLAG_SSE2){

  1116.                 c->fdct = ff_fdct_sse2;

  1117.             } else if (mm_flags & AV_CPU_FLAG_MMXEXT) {

  1118.                 c->fdct = ff_fdct_mmx2;

  1119.             }else{

  1120.                 c->fdct = ff_fdct_mmx;

  1121.             }

  1122.         }

  1123. 

  1124.         if (bit_depth <= 8)

  1125.             c->get_pixels = get_pixels_mmx;

  1126.         c->diff_pixels = diff_pixels_mmx;

  1127.         c->pix_sum = pix_sum16_mmx;

  1128. 

  1129.         c->diff_bytes= diff_bytes_mmx;

  1130.         c->sum_abs_dctelem= sum_abs_dctelem_mmx;

  1131. 

  1132.         c->pix_norm1 = pix_norm1_mmx;

  1133.         c->sse[0] = sse16_mmx;

  1134.         c->sse[1] = sse8_mmx;

  1135.         c->vsad[4]= vsad_intra16_mmx;

  1136. 

  1137.         c->nsse[0] = nsse16_mmx;

  1138.         c->nsse[1] = nsse8_mmx;

  1139.         if(!(avctx->flags & CODEC_FLAG_BITEXACT)){

  1140.             c->vsad[0] = vsad16_mmx;

  1141.         }

  1142. 

  1143.         if(!(avctx->flags & CODEC_FLAG_BITEXACT)){

  1144.             c->try_8x8basis= try_8x8basis_mmx;

  1145.         }

  1146.         c->add_8x8basis= add_8x8basis_mmx;

  1147. 

  1148.         c->ssd_int8_vs_int16 = ssd_int8_vs_int16_mmx;

  1149. 

  1150.         if (mm_flags & AV_CPU_FLAG_MMXEXT) {

  1151.             c->sum_abs_dctelem= sum_abs_dctelem_mmx2;

  1152.             c->vsad[4]= vsad_intra16_mmx2;

  1153. 

  1154.             if(!(avctx->flags & CODEC_FLAG_BITEXACT)){

  1155.                 c->vsad[0] = vsad16_mmx2;

  1156.             }

  1157. 

  1158.             c->sub_hfyu_median_prediction= sub_hfyu_median_prediction_mmx2;

  1159.         }

  1160. 

  1161.         if(mm_flags & AV_CPU_FLAG_SSE2){

  1162.             if (bit_depth <= 8)

  1163.                 c->get_pixels = get_pixels_sse2;

  1164.             c->sum_abs_dctelem= sum_abs_dctelem_sse2;

  1165.         }

  1166. 

  1167. #if HAVE_SSSE3_INLINE

  1168.         if(mm_flags & AV_CPU_FLAG_SSSE3){

  1169.             if(!(avctx->flags & CODEC_FLAG_BITEXACT)){

  1170.                 c->try_8x8basis= try_8x8basis_ssse3;

  1171.             }

  1172.             c->add_8x8basis= add_8x8basis_ssse3;

  1173.             c->sum_abs_dctelem= sum_abs_dctelem_ssse3;

  1174.         }

  1175. #endif

  1176. 

  1177.         if(mm_flags & AV_CPU_FLAG_3DNOW){

  1178.             if(!(avctx->flags & CODEC_FLAG_BITEXACT)){

  1179.                 c->try_8x8basis= try_8x8basis_3dnow;

  1180.             }

  1181.             c->add_8x8basis= add_8x8basis_3dnow;

  1182.         }

  1183.     }

  1184. #endif /* HAVE_INLINE_ASM */

  1185. 

  1186.     if (EXTERNAL_MMX(mm_flags)) {

  1187.         c->hadamard8_diff[0] = ff_hadamard8_diff16_mmx;

  1188.         c->hadamard8_diff[1] = ff_hadamard8_diff_mmx;

  1189. 

  1190.         if (EXTERNAL_MMXEXT(mm_flags)) {

  1191.             c->hadamard8_diff[0] = ff_hadamard8_diff16_mmx2;

  1192.             c->hadamard8_diff[1] = ff_hadamard8_diff_mmx2;

  1193.         }

  1194. 

  1195.         if (EXTERNAL_SSE2(mm_flags)) {

  1196.             c->sse[0] = ff_sse16_sse2;

  1197. 

  1198. #if HAVE_ALIGNED_STACK

  1199.             c->hadamard8_diff[0] = ff_hadamard8_diff16_sse2;

  1200.             c->hadamard8_diff[1] = ff_hadamard8_diff_sse2;

  1201. #endif

  1202.         }

  1203. 

  1204.         if (EXTERNAL_SSSE3(mm_flags) && HAVE_ALIGNED_STACK) {

  1205.             c->hadamard8_diff[0] = ff_hadamard8_diff16_ssse3;

  1206.             c->hadamard8_diff[1] = ff_hadamard8_diff_ssse3;

  1207.         }

  1208.     }

  1209. 

  1210.     ff_dsputil_init_pix_mmx(c, avctx);

  1211. }