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.
30020 Commits
32 Branches
748MB
Tree: a3e0135dbb
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from 'a3e0135dbb'
${ noResults }
FFmpeg/libavcodec/x86/dsputil_mmx.c

2933 lines
122KB
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_cpu.h"

  27. #include "libavcodec/dsputil.h"

  28. #include "libavcodec/h264dsp.h"

  29. #include "libavcodec/mpegvideo.h"

  30. #include "libavcodec/simple_idct.h"

  31. #include "libavcodec/ac3dec.h"

  32. #include "dsputil_mmx.h"

  33. #include "idct_xvid.h"

  34. 

  35. //#undef NDEBUG

  36. //#include <assert.h>

  37. 

  38. /* pixel operations */

  39. DECLARE_ALIGNED(8,  const uint64_t, ff_bone) = 0x0101010101010101ULL;

  40. DECLARE_ALIGNED(8,  const uint64_t, ff_wtwo) = 0x0002000200020002ULL;

  41. 

  42. DECLARE_ALIGNED(16, const uint64_t, ff_pdw_80000000)[2] =

  43. {0x8000000080000000ULL, 0x8000000080000000ULL};

  44. 

  45. DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_1  ) = {0x0001000100010001ULL, 0x0001000100010001ULL};

  46. DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_2  ) = {0x0002000200020002ULL, 0x0002000200020002ULL};

  47. DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_3  ) = {0x0003000300030003ULL, 0x0003000300030003ULL};

  48. DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_4  ) = {0x0004000400040004ULL, 0x0004000400040004ULL};

  49. DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_5  ) = {0x0005000500050005ULL, 0x0005000500050005ULL};

  50. DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_8  ) = {0x0008000800080008ULL, 0x0008000800080008ULL};

  51. DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_9  ) = {0x0009000900090009ULL, 0x0009000900090009ULL};

  52. DECLARE_ALIGNED(8,  const uint64_t, ff_pw_15 ) = 0x000F000F000F000FULL;

  53. DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_16 ) = {0x0010001000100010ULL, 0x0010001000100010ULL};

  54. DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_17 ) = {0x0011001100110011ULL, 0x0011001100110011ULL};

  55. DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_18 ) = {0x0012001200120012ULL, 0x0012001200120012ULL};

  56. DECLARE_ALIGNED(8,  const uint64_t, ff_pw_20 ) = 0x0014001400140014ULL;

  57. DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_27 ) = {0x001B001B001B001BULL, 0x001B001B001B001BULL};

  58. DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_28 ) = {0x001C001C001C001CULL, 0x001C001C001C001CULL};

  59. DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_32 ) = {0x0020002000200020ULL, 0x0020002000200020ULL};

  60. DECLARE_ALIGNED(8,  const uint64_t, ff_pw_42 ) = 0x002A002A002A002AULL;

  61. DECLARE_ALIGNED(8,  const uint64_t, ff_pw_53 ) = 0x0035003500350035ULL;

  62. DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_63 ) = {0x003F003F003F003FULL, 0x003F003F003F003FULL};

  63. DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_64 ) = {0x0040004000400040ULL, 0x0040004000400040ULL};

  64. DECLARE_ALIGNED(8,  const uint64_t, ff_pw_96 ) = 0x0060006000600060ULL;

  65. DECLARE_ALIGNED(8,  const uint64_t, ff_pw_128) = 0x0080008000800080ULL;

  66. DECLARE_ALIGNED(8,  const uint64_t, ff_pw_255) = 0x00ff00ff00ff00ffULL;

  67. 

  68. DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_0  ) = {0x0000000000000000ULL, 0x0000000000000000ULL};

  69. DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_1  ) = {0x0101010101010101ULL, 0x0101010101010101ULL};

  70. DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_3  ) = {0x0303030303030303ULL, 0x0303030303030303ULL};

  71. DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_4  ) = {0x0404040404040404ULL, 0x0404040404040404ULL};

  72. DECLARE_ALIGNED(8,  const uint64_t, ff_pb_7  ) = 0x0707070707070707ULL;

  73. DECLARE_ALIGNED(8,  const uint64_t, ff_pb_1F ) = 0x1F1F1F1F1F1F1F1FULL;

  74. DECLARE_ALIGNED(8,  const uint64_t, ff_pb_3F ) = 0x3F3F3F3F3F3F3F3FULL;

  75. DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_80 ) = {0x8080808080808080ULL, 0x8080808080808080ULL};

  76. DECLARE_ALIGNED(8,  const uint64_t, ff_pb_81 ) = 0x8181818181818181ULL;

  77. DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_A1 ) = {0xA1A1A1A1A1A1A1A1ULL, 0xA1A1A1A1A1A1A1A1ULL};

  78. DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_F8 ) = {0xF8F8F8F8F8F8F8F8ULL, 0xF8F8F8F8F8F8F8F8ULL};

  79. DECLARE_ALIGNED(8,  const uint64_t, ff_pb_FC ) = 0xFCFCFCFCFCFCFCFCULL;

  80. DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_FE ) = {0xFEFEFEFEFEFEFEFEULL, 0xFEFEFEFEFEFEFEFEULL};

  81. 

  82. DECLARE_ALIGNED(16, const double, ff_pd_1)[2] = { 1.0, 1.0 };

  83. DECLARE_ALIGNED(16, const double, ff_pd_2)[2] = { 2.0, 2.0 };

  84. 

  85. #define JUMPALIGN() __asm__ volatile (".p2align 3"::)

  86. #define MOVQ_ZERO(regd)  __asm__ volatile ("pxor %%" #regd ", %%" #regd ::)

  87. 

  88. #define MOVQ_BFE(regd) \

  89.     __asm__ volatile ( \

  90.     "pcmpeqd %%" #regd ", %%" #regd " \n\t"\

  91.     "paddb %%" #regd ", %%" #regd " \n\t" ::)

  92. 

  93. #ifndef PIC

  94. #define MOVQ_BONE(regd)  __asm__ volatile ("movq %0, %%" #regd " \n\t" ::"m"(ff_bone))

  95. #define MOVQ_WTWO(regd)  __asm__ volatile ("movq %0, %%" #regd " \n\t" ::"m"(ff_wtwo))

  96. #else

  97. // for shared library it's better to use this way for accessing constants

  98. // pcmpeqd -> -1

  99. #define MOVQ_BONE(regd) \

  100.     __asm__ volatile ( \

  101.     "pcmpeqd %%" #regd ", %%" #regd " \n\t" \

  102.     "psrlw $15, %%" #regd " \n\t" \

  103.     "packuswb %%" #regd ", %%" #regd " \n\t" ::)

  104. 

  105. #define MOVQ_WTWO(regd) \

  106.     __asm__ volatile ( \

  107.     "pcmpeqd %%" #regd ", %%" #regd " \n\t" \

  108.     "psrlw $15, %%" #regd " \n\t" \

  109.     "psllw $1, %%" #regd " \n\t"::)

  110. 

  111. #endif

  112. 

  113. // using regr as temporary and for the output result

  114. // first argument is unmodifed and second is trashed

  115. // regfe is supposed to contain 0xfefefefefefefefe

  116. #define PAVGB_MMX_NO_RND(rega, regb, regr, regfe) \

  117.     "movq " #rega ", " #regr "  \n\t"\

  118.     "pand " #regb ", " #regr "  \n\t"\

  119.     "pxor " #rega ", " #regb "  \n\t"\

  120.     "pand " #regfe "," #regb "  \n\t"\

  121.     "psrlq $1, " #regb "        \n\t"\

  122.     "paddb " #regb ", " #regr " \n\t"

  123. 

  124. #define PAVGB_MMX(rega, regb, regr, regfe) \

  125.     "movq " #rega ", " #regr "  \n\t"\

  126.     "por  " #regb ", " #regr "  \n\t"\

  127.     "pxor " #rega ", " #regb "  \n\t"\

  128.     "pand " #regfe "," #regb "  \n\t"\

  129.     "psrlq $1, " #regb "        \n\t"\

  130.     "psubb " #regb ", " #regr " \n\t"

  131. 

  132. // mm6 is supposed to contain 0xfefefefefefefefe

  133. #define PAVGBP_MMX_NO_RND(rega, regb, regr,  regc, regd, regp) \

  134.     "movq " #rega ", " #regr "  \n\t"\

  135.     "movq " #regc ", " #regp "  \n\t"\

  136.     "pand " #regb ", " #regr "  \n\t"\

  137.     "pand " #regd ", " #regp "  \n\t"\

  138.     "pxor " #rega ", " #regb "  \n\t"\

  139.     "pxor " #regc ", " #regd "  \n\t"\

  140.     "pand %%mm6, " #regb "      \n\t"\

  141.     "pand %%mm6, " #regd "      \n\t"\

  142.     "psrlq $1, " #regb "        \n\t"\

  143.     "psrlq $1, " #regd "        \n\t"\

  144.     "paddb " #regb ", " #regr " \n\t"\

  145.     "paddb " #regd ", " #regp " \n\t"

  146. 

  147. #define PAVGBP_MMX(rega, regb, regr, regc, regd, regp) \

  148.     "movq " #rega ", " #regr "  \n\t"\

  149.     "movq " #regc ", " #regp "  \n\t"\

  150.     "por  " #regb ", " #regr "  \n\t"\

  151.     "por  " #regd ", " #regp "  \n\t"\

  152.     "pxor " #rega ", " #regb "  \n\t"\

  153.     "pxor " #regc ", " #regd "  \n\t"\

  154.     "pand %%mm6, " #regb "      \n\t"\

  155.     "pand %%mm6, " #regd "      \n\t"\

  156.     "psrlq $1, " #regd "        \n\t"\

  157.     "psrlq $1, " #regb "        \n\t"\

  158.     "psubb " #regb ", " #regr " \n\t"\

  159.     "psubb " #regd ", " #regp " \n\t"

  160. 

  161. /***********************************/

  162. /* MMX no rounding */

  163. #define DEF(x, y) x ## _no_rnd_ ## y ##_mmx

  164. #define SET_RND  MOVQ_WONE

  165. #define PAVGBP(a, b, c, d, e, f)        PAVGBP_MMX_NO_RND(a, b, c, d, e, f)

  166. #define PAVGB(a, b, c, e)               PAVGB_MMX_NO_RND(a, b, c, e)

  167. #define OP_AVG(a, b, c, e)              PAVGB_MMX(a, b, c, e)

  168. 

  169. #include "dsputil_mmx_rnd_template.c"

  170. 

  171. #undef DEF

  172. #undef SET_RND

  173. #undef PAVGBP

  174. #undef PAVGB

  175. /***********************************/

  176. /* MMX rounding */

  177. 

  178. #define DEF(x, y) x ## _ ## y ##_mmx

  179. #define SET_RND  MOVQ_WTWO

  180. #define PAVGBP(a, b, c, d, e, f)        PAVGBP_MMX(a, b, c, d, e, f)

  181. #define PAVGB(a, b, c, e)               PAVGB_MMX(a, b, c, e)

  182. 

  183. #include "dsputil_mmx_rnd_template.c"

  184. 

  185. #undef DEF

  186. #undef SET_RND

  187. #undef PAVGBP

  188. #undef PAVGB

  189. #undef OP_AVG

  190. 

  191. /***********************************/

  192. /* 3Dnow specific */

  193. 

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

  195. #define PAVGB "pavgusb"

  196. #define OP_AVG PAVGB

  197. 

  198. #include "dsputil_mmx_avg_template.c"

  199. 

  200. #undef DEF

  201. #undef PAVGB

  202. #undef OP_AVG

  203. 

  204. /***********************************/

  205. /* MMX2 specific */

  206. 

  207. #define DEF(x) x ## _mmx2

  208. 

  209. /* Introduced only in MMX2 set */

  210. #define PAVGB "pavgb"

  211. #define OP_AVG PAVGB

  212. 

  213. #include "dsputil_mmx_avg_template.c"

  214. 

  215. #undef DEF

  216. #undef PAVGB

  217. #undef OP_AVG

  218. 

  219. #define put_no_rnd_pixels16_mmx put_pixels16_mmx

  220. #define put_no_rnd_pixels8_mmx put_pixels8_mmx

  221. #define put_pixels16_mmx2 put_pixels16_mmx

  222. #define put_pixels8_mmx2 put_pixels8_mmx

  223. #define put_pixels4_mmx2 put_pixels4_mmx

  224. #define put_no_rnd_pixels16_mmx2 put_no_rnd_pixels16_mmx

  225. #define put_no_rnd_pixels8_mmx2 put_no_rnd_pixels8_mmx

  226. #define put_pixels16_3dnow put_pixels16_mmx

  227. #define put_pixels8_3dnow put_pixels8_mmx

  228. #define put_pixels4_3dnow put_pixels4_mmx

  229. #define put_no_rnd_pixels16_3dnow put_no_rnd_pixels16_mmx

  230. #define put_no_rnd_pixels8_3dnow put_no_rnd_pixels8_mmx

  231. 

  232. /***********************************/

  233. /* standard MMX */

  234. 

  235. void ff_put_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size)

  236. {

  237.     const DCTELEM *p;

  238.     uint8_t *pix;

  239. 

  240.     /* read the pixels */

  241.     p = block;

  242.     pix = pixels;

  243.     /* unrolled loop */

  244.         __asm__ volatile(

  245.                 "movq   %3, %%mm0               \n\t"

  246.                 "movq   8%3, %%mm1              \n\t"

  247.                 "movq   16%3, %%mm2             \n\t"

  248.                 "movq   24%3, %%mm3             \n\t"

  249.                 "movq   32%3, %%mm4             \n\t"

  250.                 "movq   40%3, %%mm5             \n\t"

  251.                 "movq   48%3, %%mm6             \n\t"

  252.                 "movq   56%3, %%mm7             \n\t"

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

  254.                 "packuswb %%mm3, %%mm2          \n\t"

  255.                 "packuswb %%mm5, %%mm4          \n\t"

  256.                 "packuswb %%mm7, %%mm6          \n\t"

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

  258.                 "movq   %%mm2, (%0, %1)         \n\t"

  259.                 "movq   %%mm4, (%0, %1, 2)      \n\t"

  260.                 "movq   %%mm6, (%0, %2)         \n\t"

  261.                 ::"r" (pix), "r" ((x86_reg)line_size), "r" ((x86_reg)line_size*3), "m"(*p)

  262.                 :"memory");

  263.         pix += line_size*4;

  264.         p += 32;

  265. 

  266.     // if here would be an exact copy of the code above

  267.     // compiler would generate some very strange code

  268.     // thus using "r"

  269.     __asm__ volatile(

  270.             "movq       (%3), %%mm0             \n\t"

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

  272.             "movq       16(%3), %%mm2           \n\t"

  273.             "movq       24(%3), %%mm3           \n\t"

  274.             "movq       32(%3), %%mm4           \n\t"

  275.             "movq       40(%3), %%mm5           \n\t"

  276.             "movq       48(%3), %%mm6           \n\t"

  277.             "movq       56(%3), %%mm7           \n\t"

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

  279.             "packuswb %%mm3, %%mm2              \n\t"

  280.             "packuswb %%mm5, %%mm4              \n\t"

  281.             "packuswb %%mm7, %%mm6              \n\t"

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

  283.             "movq       %%mm2, (%0, %1)         \n\t"

  284.             "movq       %%mm4, (%0, %1, 2)      \n\t"

  285.             "movq       %%mm6, (%0, %2)         \n\t"

  286.             ::"r" (pix), "r" ((x86_reg)line_size), "r" ((x86_reg)line_size*3), "r"(p)

  287.             :"memory");

  288. }

  289. 

  290. #define put_signed_pixels_clamped_mmx_half(off) \

  291.             "movq    "#off"(%2), %%mm1          \n\t"\

  292.             "movq 16+"#off"(%2), %%mm2          \n\t"\

  293.             "movq 32+"#off"(%2), %%mm3          \n\t"\

  294.             "movq 48+"#off"(%2), %%mm4          \n\t"\

  295.             "packsswb  8+"#off"(%2), %%mm1      \n\t"\

  296.             "packsswb 24+"#off"(%2), %%mm2      \n\t"\

  297.             "packsswb 40+"#off"(%2), %%mm3      \n\t"\

  298.             "packsswb 56+"#off"(%2), %%mm4      \n\t"\

  299.             "paddb %%mm0, %%mm1                 \n\t"\

  300.             "paddb %%mm0, %%mm2                 \n\t"\

  301.             "paddb %%mm0, %%mm3                 \n\t"\

  302.             "paddb %%mm0, %%mm4                 \n\t"\

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

  304.             "movq %%mm2, (%0, %3)               \n\t"\

  305.             "movq %%mm3, (%0, %3, 2)            \n\t"\

  306.             "movq %%mm4, (%0, %1)               \n\t"

  307. 

  308. void ff_put_signed_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size)

  309. {

  310.     x86_reg line_skip = line_size;

  311.     x86_reg line_skip3;

  312. 

  313.     __asm__ volatile (

  314.             "movq "MANGLE(ff_pb_80)", %%mm0 \n\t"

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

  316.             put_signed_pixels_clamped_mmx_half(0)

  317.             "lea (%0, %3, 4), %0                \n\t"

  318.             put_signed_pixels_clamped_mmx_half(64)

  319.             :"+&r" (pixels), "=&r" (line_skip3)

  320.             :"r" (block), "r"(line_skip)

  321.             :"memory");

  322. }

  323. 

  324. void ff_add_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size)

  325. {

  326.     const DCTELEM *p;

  327.     uint8_t *pix;

  328.     int i;

  329. 

  330.     /* read the pixels */

  331.     p = block;

  332.     pix = pixels;

  333.     MOVQ_ZERO(mm7);

  334.     i = 4;

  335.     do {

  336.         __asm__ volatile(

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

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

  339.                 "movq   16(%2), %%mm2   \n\t"

  340.                 "movq   24(%2), %%mm3   \n\t"

  341.                 "movq   %0, %%mm4       \n\t"

  342.                 "movq   %1, %%mm6       \n\t"

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

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

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

  346.                 "paddsw %%mm4, %%mm0    \n\t"

  347.                 "paddsw %%mm5, %%mm1    \n\t"

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

  349.                 "punpcklbw %%mm7, %%mm6 \n\t"

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

  351.                 "paddsw %%mm6, %%mm2    \n\t"

  352.                 "paddsw %%mm5, %%mm3    \n\t"

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

  354.                 "packuswb %%mm3, %%mm2  \n\t"

  355.                 "movq   %%mm0, %0       \n\t"

  356.                 "movq   %%mm2, %1       \n\t"

  357.                 :"+m"(*pix), "+m"(*(pix+line_size))

  358.                 :"r"(p)

  359.                 :"memory");

  360.         pix += line_size*2;

  361.         p += 16;

  362.     } while (--i);

  363. }

  364. 

  365. static void put_pixels4_mmx(uint8_t *block, const uint8_t *pixels, int line_size, int h)

  366. {

  367.     __asm__ volatile(

  368.          "lea (%3, %3), %%"REG_a"       \n\t"

  369.          ".p2align 3                    \n\t"

  370.          "1:                            \n\t"

  371.          "movd (%1), %%mm0              \n\t"

  372.          "movd (%1, %3), %%mm1          \n\t"

  373.          "movd %%mm0, (%2)              \n\t"

  374.          "movd %%mm1, (%2, %3)          \n\t"

  375.          "add %%"REG_a", %1             \n\t"

  376.          "add %%"REG_a", %2             \n\t"

  377.          "movd (%1), %%mm0              \n\t"

  378.          "movd (%1, %3), %%mm1          \n\t"

  379.          "movd %%mm0, (%2)              \n\t"

  380.          "movd %%mm1, (%2, %3)          \n\t"

  381.          "add %%"REG_a", %1             \n\t"

  382.          "add %%"REG_a", %2             \n\t"

  383.          "subl $4, %0                   \n\t"

  384.          "jnz 1b                        \n\t"

  385.          : "+g"(h), "+r" (pixels),  "+r" (block)

  386.          : "r"((x86_reg)line_size)

  387.          : "%"REG_a, "memory"

  388.         );

  389. }

  390. 

  391. static void put_pixels8_mmx(uint8_t *block, const uint8_t *pixels, int line_size, int h)

  392. {

  393.     __asm__ volatile(

  394.          "lea (%3, %3), %%"REG_a"       \n\t"

  395.          ".p2align 3                    \n\t"

  396.          "1:                            \n\t"

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

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

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

  400.          "movq %%mm1, (%2, %3)          \n\t"

  401.          "add %%"REG_a", %1             \n\t"

  402.          "add %%"REG_a", %2             \n\t"

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

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

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

  406.          "movq %%mm1, (%2, %3)          \n\t"

  407.          "add %%"REG_a", %1             \n\t"

  408.          "add %%"REG_a", %2             \n\t"

  409.          "subl $4, %0                   \n\t"

  410.          "jnz 1b                        \n\t"

  411.          : "+g"(h), "+r" (pixels),  "+r" (block)

  412.          : "r"((x86_reg)line_size)

  413.          : "%"REG_a, "memory"

  414.         );

  415. }

  416. 

  417. static void put_pixels16_mmx(uint8_t *block, const uint8_t *pixels, int line_size, int h)

  418. {

  419.     __asm__ volatile(

  420.          "lea (%3, %3), %%"REG_a"       \n\t"

  421.          ".p2align 3                    \n\t"

  422.          "1:                            \n\t"

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

  424.          "movq 8(%1), %%mm4             \n\t"

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

  426.          "movq 8(%1, %3), %%mm5         \n\t"

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

  428.          "movq %%mm4, 8(%2)             \n\t"

  429.          "movq %%mm1, (%2, %3)          \n\t"

  430.          "movq %%mm5, 8(%2, %3)         \n\t"

  431.          "add %%"REG_a", %1             \n\t"

  432.          "add %%"REG_a", %2             \n\t"

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

  434.          "movq 8(%1), %%mm4             \n\t"

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

  436.          "movq 8(%1, %3), %%mm5         \n\t"

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

  438.          "movq %%mm4, 8(%2)             \n\t"

  439.          "movq %%mm1, (%2, %3)          \n\t"

  440.          "movq %%mm5, 8(%2, %3)         \n\t"

  441.          "add %%"REG_a", %1             \n\t"

  442.          "add %%"REG_a", %2             \n\t"

  443.          "subl $4, %0                   \n\t"

  444.          "jnz 1b                        \n\t"

  445.          : "+g"(h), "+r" (pixels),  "+r" (block)

  446.          : "r"((x86_reg)line_size)

  447.          : "%"REG_a, "memory"

  448.         );

  449. }

  450. 

  451. static void put_pixels16_sse2(uint8_t *block, const uint8_t *pixels, int line_size, int h)

  452. {

  453.     __asm__ volatile(

  454.          "1:                            \n\t"

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

  456.          "movdqu (%1,%3), %%xmm1        \n\t"

  457.          "movdqu (%1,%3,2), %%xmm2      \n\t"

  458.          "movdqu (%1,%4), %%xmm3        \n\t"

  459.          "lea (%1,%3,4), %1             \n\t"

  460.          "movdqa %%xmm0, (%2)           \n\t"

  461.          "movdqa %%xmm1, (%2,%3)        \n\t"

  462.          "movdqa %%xmm2, (%2,%3,2)      \n\t"

  463.          "movdqa %%xmm3, (%2,%4)        \n\t"

  464.          "subl $4, %0                   \n\t"

  465.          "lea (%2,%3,4), %2             \n\t"

  466.          "jnz 1b                        \n\t"

  467.          : "+g"(h), "+r" (pixels),  "+r" (block)

  468.          : "r"((x86_reg)line_size), "r"((x86_reg)3L*line_size)

  469.          : "memory"

  470.         );

  471. }

  472. 

  473. static void avg_pixels16_sse2(uint8_t *block, const uint8_t *pixels, int line_size, int h)

  474. {

  475.     __asm__ volatile(

  476.          "1:                            \n\t"

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

  478.          "movdqu (%1,%3), %%xmm1        \n\t"

  479.          "movdqu (%1,%3,2), %%xmm2      \n\t"

  480.          "movdqu (%1,%4), %%xmm3        \n\t"

  481.          "lea (%1,%3,4), %1             \n\t"

  482.          "pavgb  (%2), %%xmm0           \n\t"

  483.          "pavgb  (%2,%3), %%xmm1        \n\t"

  484.          "pavgb  (%2,%3,2), %%xmm2      \n\t"

  485.          "pavgb  (%2,%4), %%xmm3        \n\t"

  486.          "movdqa %%xmm0, (%2)           \n\t"

  487.          "movdqa %%xmm1, (%2,%3)        \n\t"

  488.          "movdqa %%xmm2, (%2,%3,2)      \n\t"

  489.          "movdqa %%xmm3, (%2,%4)        \n\t"

  490.          "subl $4, %0                   \n\t"

  491.          "lea (%2,%3,4), %2             \n\t"

  492.          "jnz 1b                        \n\t"

  493.          : "+g"(h), "+r" (pixels),  "+r" (block)

  494.          : "r"((x86_reg)line_size), "r"((x86_reg)3L*line_size)

  495.          : "memory"

  496.         );

  497. }

  498. 

  499. #define CLEAR_BLOCKS(name,n) \

  500. static void name(DCTELEM *blocks)\

  501. {\

  502.     __asm__ volatile(\

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

  504.                 "mov     %1, %%"REG_a"          \n\t"\

  505.                 "1:                             \n\t"\

  506.                 "movq %%mm7, (%0, %%"REG_a")    \n\t"\

  507.                 "movq %%mm7, 8(%0, %%"REG_a")   \n\t"\

  508.                 "movq %%mm7, 16(%0, %%"REG_a")  \n\t"\

  509.                 "movq %%mm7, 24(%0, %%"REG_a")  \n\t"\

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

  511.                 " js 1b                         \n\t"\

  512.                 : : "r" (((uint8_t *)blocks)+128*n),\

  513.                     "i" (-128*n)\

  514.                 : "%"REG_a\

  515.         );\

  516. }

  517. CLEAR_BLOCKS(clear_blocks_mmx, 6)

  518. CLEAR_BLOCKS(clear_block_mmx, 1)

  519. 

  520. static void clear_block_sse(DCTELEM *block)

  521. {

  522.     __asm__ volatile(

  523.         "xorps  %%xmm0, %%xmm0  \n"

  524.         "movaps %%xmm0,    (%0) \n"

  525.         "movaps %%xmm0,  16(%0) \n"

  526.         "movaps %%xmm0,  32(%0) \n"

  527.         "movaps %%xmm0,  48(%0) \n"

  528.         "movaps %%xmm0,  64(%0) \n"

  529.         "movaps %%xmm0,  80(%0) \n"

  530.         "movaps %%xmm0,  96(%0) \n"

  531.         "movaps %%xmm0, 112(%0) \n"

  532.         :: "r"(block)

  533.         : "memory"

  534.     );

  535. }

  536. 

  537. static void clear_blocks_sse(DCTELEM *blocks)

  538. {\

  539.     __asm__ volatile(

  540.         "xorps  %%xmm0, %%xmm0  \n"

  541.         "mov     %1, %%"REG_a"  \n"

  542.         "1:                     \n"

  543.         "movaps %%xmm0,    (%0, %%"REG_a") \n"

  544.         "movaps %%xmm0,  16(%0, %%"REG_a") \n"

  545.         "movaps %%xmm0,  32(%0, %%"REG_a") \n"

  546.         "movaps %%xmm0,  48(%0, %%"REG_a") \n"

  547.         "movaps %%xmm0,  64(%0, %%"REG_a") \n"

  548.         "movaps %%xmm0,  80(%0, %%"REG_a") \n"

  549.         "movaps %%xmm0,  96(%0, %%"REG_a") \n"

  550.         "movaps %%xmm0, 112(%0, %%"REG_a") \n"

  551.         "add $128, %%"REG_a"    \n"

  552.         " js 1b                 \n"

  553.         : : "r" (((uint8_t *)blocks)+128*6),

  554.             "i" (-128*6)

  555.         : "%"REG_a

  556.     );

  557. }

  558. 

  559. static void add_bytes_mmx(uint8_t *dst, uint8_t *src, int w){

  560.     x86_reg i=0;

  561.     __asm__ volatile(

  562.         "jmp 2f                         \n\t"

  563.         "1:                             \n\t"

  564.         "movq  (%1, %0), %%mm0          \n\t"

  565.         "movq  (%2, %0), %%mm1          \n\t"

  566.         "paddb %%mm0, %%mm1             \n\t"

  567.         "movq %%mm1, (%2, %0)           \n\t"

  568.         "movq 8(%1, %0), %%mm0          \n\t"

  569.         "movq 8(%2, %0), %%mm1          \n\t"

  570.         "paddb %%mm0, %%mm1             \n\t"

  571.         "movq %%mm1, 8(%2, %0)          \n\t"

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

  573.         "2:                             \n\t"

  574.         "cmp %3, %0                     \n\t"

  575.         " js 1b                         \n\t"

  576.         : "+r" (i)

  577.         : "r"(src), "r"(dst), "r"((x86_reg)w-15)

  578.     );

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

  580.         dst[i+0] += src[i+0];

  581. }

  582. 

  583. static void add_bytes_l2_mmx(uint8_t *dst, uint8_t *src1, uint8_t *src2, int w){

  584.     x86_reg i=0;

  585.     __asm__ volatile(

  586.         "jmp 2f                         \n\t"

  587.         "1:                             \n\t"

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

  589.         "movq  8(%2, %0), %%mm1         \n\t"

  590.         "paddb  (%3, %0), %%mm0         \n\t"

  591.         "paddb 8(%3, %0), %%mm1         \n\t"

  592.         "movq %%mm0,  (%1, %0)          \n\t"

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

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

  595.         "2:                             \n\t"

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

  597.         " js 1b                         \n\t"

  598.         : "+r" (i)

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

  600.     );

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

  602.         dst[i] = src1[i] + src2[i];

  603. }

  604. 

  605. #if HAVE_7REGS

  606. static void add_hfyu_median_prediction_cmov(uint8_t *dst, const uint8_t *top, const uint8_t *diff, int w, int *left, int *left_top) {

  607.     x86_reg w2 = -w;

  608.     x86_reg x;

  609.     int l = *left & 0xff;

  610.     int tl = *left_top & 0xff;

  611.     int t;

  612.     __asm__ volatile(

  613.         "mov    %7, %3 \n"

  614.         "1: \n"

  615.         "movzbl (%3,%4), %2 \n"

  616.         "mov    %2, %k3 \n"

  617.         "sub   %b1, %b3 \n"

  618.         "add   %b0, %b3 \n"

  619.         "mov    %2, %1 \n"

  620.         "cmp    %0, %2 \n"

  621.         "cmovg  %0, %2 \n"

  622.         "cmovg  %1, %0 \n"

  623.         "cmp   %k3, %0 \n"

  624.         "cmovg %k3, %0 \n"

  625.         "mov    %7, %3 \n"

  626.         "cmp    %2, %0 \n"

  627.         "cmovl  %2, %0 \n"

  628.         "add (%6,%4), %b0 \n"

  629.         "mov   %b0, (%5,%4) \n"

  630.         "inc    %4 \n"

  631.         "jl 1b \n"

  632.         :"+&q"(l), "+&q"(tl), "=&r"(t), "=&q"(x), "+&r"(w2)

  633.         :"r"(dst+w), "r"(diff+w), "rm"(top+w)

  634.     );

  635.     *left = l;

  636.     *left_top = tl;

  637. }

  638. #endif

  639. 

  640. #define H263_LOOP_FILTER \

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

  642.         "movq  %0, %%mm0                \n\t"\

  643.         "movq  %0, %%mm1                \n\t"\

  644.         "movq  %3, %%mm2                \n\t"\

  645.         "movq  %3, %%mm3                \n\t"\

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

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

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

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

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

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

  652.         "movq  %1, %%mm2                \n\t"\

  653.         "movq  %1, %%mm3                \n\t"\

  654.         "movq  %2, %%mm4                \n\t"\

  655.         "movq  %2, %%mm5                \n\t"\

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

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

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

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

  660.         "psubw %%mm2, %%mm4             \n\t"\

  661.         "psubw %%mm3, %%mm5             \n\t"\

  662.         "psllw $2, %%mm4                \n\t"\

  663.         "psllw $2, %%mm5                \n\t"\

  664.         "paddw %%mm0, %%mm4             \n\t"\

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

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

  667.         "pcmpgtw %%mm4, %%mm6           \n\t"\

  668.         "pcmpgtw %%mm5, %%mm7           \n\t"\

  669.         "pxor %%mm6, %%mm4              \n\t"\

  670.         "pxor %%mm7, %%mm5              \n\t"\

  671.         "psubw %%mm6, %%mm4             \n\t"\

  672.         "psubw %%mm7, %%mm5             \n\t"\

  673.         "psrlw $3, %%mm4                \n\t"\

  674.         "psrlw $3, %%mm5                \n\t"\

  675.         "packuswb %%mm5, %%mm4          \n\t"\

  676.         "packsswb %%mm7, %%mm6          \n\t"\

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

  678.         "movd %4, %%mm2                 \n\t"\

  679.         "punpcklbw %%mm2, %%mm2         \n\t"\

  680.         "punpcklbw %%mm2, %%mm2         \n\t"\

  681.         "punpcklbw %%mm2, %%mm2         \n\t"\

  682.         "psubusb %%mm4, %%mm2           \n\t"\

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

  684.         "psubusb %%mm4, %%mm3           \n\t"\

  685.         "psubb %%mm3, %%mm2             \n\t"\

  686.         "movq %1, %%mm3                 \n\t"\

  687.         "movq %2, %%mm4                 \n\t"\

  688.         "pxor %%mm6, %%mm3              \n\t"\

  689.         "pxor %%mm6, %%mm4              \n\t"\

  690.         "paddusb %%mm2, %%mm3           \n\t"\

  691.         "psubusb %%mm2, %%mm4           \n\t"\

  692.         "pxor %%mm6, %%mm3              \n\t"\

  693.         "pxor %%mm6, %%mm4              \n\t"\

  694.         "paddusb %%mm2, %%mm2           \n\t"\

  695.         "packsswb %%mm1, %%mm0          \n\t"\

  696.         "pcmpgtb %%mm0, %%mm7           \n\t"\

  697.         "pxor %%mm7, %%mm0              \n\t"\

  698.         "psubb %%mm7, %%mm0             \n\t"\

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

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

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

  702.         "pand %5, %%mm1                 \n\t"\

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

  704.         "pxor %%mm7, %%mm1              \n\t"\

  705.         "psubb %%mm7, %%mm1             \n\t"\

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

  707.         "movq %3, %%mm6                 \n\t"\

  708.         "psubb %%mm1, %%mm5             \n\t"\

  709.         "paddb %%mm1, %%mm6             \n\t"

  710. 

  711. static void h263_v_loop_filter_mmx(uint8_t *src, int stride, int qscale){

  712.     if(CONFIG_H263_DECODER || CONFIG_H263_ENCODER) {

  713.     const int strength= ff_h263_loop_filter_strength[qscale];

  714. 

  715.     __asm__ volatile(

  716. 

  717.         H263_LOOP_FILTER

  718. 

  719.         "movq %%mm3, %1                 \n\t"

  720.         "movq %%mm4, %2                 \n\t"

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

  722.         "movq %%mm6, %3                 \n\t"

  723.         : "+m" (*(uint64_t*)(src - 2*stride)),

  724.           "+m" (*(uint64_t*)(src - 1*stride)),

  725.           "+m" (*(uint64_t*)(src + 0*stride)),

  726.           "+m" (*(uint64_t*)(src + 1*stride))

  727.         : "g" (2*strength), "m"(ff_pb_FC)

  728.     );

  729.     }

  730. }

  731. 

  732. static void h263_h_loop_filter_mmx(uint8_t *src, int stride, int qscale){

  733.     if(CONFIG_H263_DECODER || CONFIG_H263_ENCODER) {

  734.     const int strength= ff_h263_loop_filter_strength[qscale];

  735.     DECLARE_ALIGNED(8, uint64_t, temp)[4];

  736.     uint8_t *btemp= (uint8_t*)temp;

  737. 

  738.     src -= 2;

  739. 

  740.     transpose4x4(btemp  , src           , 8, stride);

  741.     transpose4x4(btemp+4, src + 4*stride, 8, stride);

  742.     __asm__ volatile(

  743.         H263_LOOP_FILTER // 5 3 4 6

  744. 

  745.         : "+m" (temp[0]),

  746.           "+m" (temp[1]),

  747.           "+m" (temp[2]),

  748.           "+m" (temp[3])

  749.         : "g" (2*strength), "m"(ff_pb_FC)

  750.     );

  751. 

  752.     __asm__ volatile(

  753.         "movq %%mm5, %%mm1              \n\t"

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

  755.         "punpcklbw %%mm3, %%mm5         \n\t"

  756.         "punpcklbw %%mm6, %%mm4         \n\t"

  757.         "punpckhbw %%mm3, %%mm1         \n\t"

  758.         "punpckhbw %%mm6, %%mm0         \n\t"

  759.         "movq %%mm5, %%mm3              \n\t"

  760.         "movq %%mm1, %%mm6              \n\t"

  761.         "punpcklwd %%mm4, %%mm5         \n\t"

  762.         "punpcklwd %%mm0, %%mm1         \n\t"

  763.         "punpckhwd %%mm4, %%mm3         \n\t"

  764.         "punpckhwd %%mm0, %%mm6         \n\t"

  765.         "movd %%mm5, (%0)               \n\t"

  766.         "punpckhdq %%mm5, %%mm5         \n\t"

  767.         "movd %%mm5, (%0,%2)            \n\t"

  768.         "movd %%mm3, (%0,%2,2)          \n\t"

  769.         "punpckhdq %%mm3, %%mm3         \n\t"

  770.         "movd %%mm3, (%0,%3)            \n\t"

  771.         "movd %%mm1, (%1)               \n\t"

  772.         "punpckhdq %%mm1, %%mm1         \n\t"

  773.         "movd %%mm1, (%1,%2)            \n\t"

  774.         "movd %%mm6, (%1,%2,2)          \n\t"

  775.         "punpckhdq %%mm6, %%mm6         \n\t"

  776.         "movd %%mm6, (%1,%3)            \n\t"

  777.         :: "r" (src),

  778.            "r" (src + 4*stride),

  779.            "r" ((x86_reg)   stride ),

  780.            "r" ((x86_reg)(3*stride))

  781.     );

  782.     }

  783. }

  784. 

  785. /* draw the edges of width 'w' of an image of size width, height

  786.    this mmx version can only handle w==8 || w==16 */

  787. static void draw_edges_mmx(uint8_t *buf, int wrap, int width, int height, int w, int h, int sides)

  788. {

  789.     uint8_t *ptr, *last_line;

  790.     int i;

  791. 

  792.     last_line = buf + (height - 1) * wrap;

  793.     /* left and right */

  794.     ptr = buf;

  795.     if(w==8)

  796.     {

  797.         __asm__ volatile(

  798.                 "1:                             \n\t"

  799.                 "movd (%0), %%mm0               \n\t"

  800.                 "punpcklbw %%mm0, %%mm0         \n\t"

  801.                 "punpcklwd %%mm0, %%mm0         \n\t"

  802.                 "punpckldq %%mm0, %%mm0         \n\t"

  803.                 "movq %%mm0, -8(%0)             \n\t"

  804.                 "movq -8(%0, %2), %%mm1         \n\t"

  805.                 "punpckhbw %%mm1, %%mm1         \n\t"

  806.                 "punpckhwd %%mm1, %%mm1         \n\t"

  807.                 "punpckhdq %%mm1, %%mm1         \n\t"

  808.                 "movq %%mm1, (%0, %2)           \n\t"

  809.                 "add %1, %0                     \n\t"

  810.                 "cmp %3, %0                     \n\t"

  811.                 " jb 1b                         \n\t"

  812.                 : "+r" (ptr)

  813.                 : "r" ((x86_reg)wrap), "r" ((x86_reg)width), "r" (ptr + wrap*height)

  814.         );

  815.     }

  816.     else

  817.     {

  818.         __asm__ volatile(

  819.                 "1:                             \n\t"

  820.                 "movd (%0), %%mm0               \n\t"

  821.                 "punpcklbw %%mm0, %%mm0         \n\t"

  822.                 "punpcklwd %%mm0, %%mm0         \n\t"

  823.                 "punpckldq %%mm0, %%mm0         \n\t"

  824.                 "movq %%mm0, -8(%0)             \n\t"

  825.                 "movq %%mm0, -16(%0)            \n\t"

  826.                 "movq -8(%0, %2), %%mm1         \n\t"

  827.                 "punpckhbw %%mm1, %%mm1         \n\t"

  828.                 "punpckhwd %%mm1, %%mm1         \n\t"

  829.                 "punpckhdq %%mm1, %%mm1         \n\t"

  830.                 "movq %%mm1, (%0, %2)           \n\t"

  831.                 "movq %%mm1, 8(%0, %2)          \n\t"

  832.                 "add %1, %0                     \n\t"

  833.                 "cmp %3, %0                     \n\t"

  834.                 " jb 1b                         \n\t"

  835.                 : "+r" (ptr)

  836.                 : "r" ((x86_reg)wrap), "r" ((x86_reg)width), "r" (ptr + wrap*height)

  837.         );

  838.     }

  839. 

  840.     /* top and bottom (and hopefully also the corners) */

  841.     if (sides&EDGE_TOP) {

  842.         for(i = 0; i < h; i += 4) {

  843.             ptr= buf - (i + 1) * wrap - w;

  844.             __asm__ volatile(

  845.                     "1:                             \n\t"

  846.                     "movq (%1, %0), %%mm0           \n\t"

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

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

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

  850.                     "movq %%mm0, (%0, %3)           \n\t"

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

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

  853.                     " jb 1b                         \n\t"

  854.                     : "+r" (ptr)

  855.                     : "r" ((x86_reg)buf - (x86_reg)ptr - w), "r" ((x86_reg)-wrap), "r" ((x86_reg)-wrap*3), "r" (ptr+width+2*w)

  856.             );

  857.         }

  858.     }

  859. 

  860.     if (sides&EDGE_BOTTOM) {

  861.         for(i = 0; i < w; i += 4) {

  862.             ptr= last_line + (i + 1) * wrap - w;

  863.             __asm__ volatile(

  864.                     "1:                             \n\t"

  865.                     "movq (%1, %0), %%mm0           \n\t"

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

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

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

  869.                     "movq %%mm0, (%0, %3)           \n\t"

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

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

  872.                     " jb 1b                         \n\t"

  873.                     : "+r" (ptr)

  874.                     : "r" ((x86_reg)last_line - (x86_reg)ptr - w), "r" ((x86_reg)wrap), "r" ((x86_reg)wrap*3), "r" (ptr+width+2*w)

  875.             );

  876.         }

  877.     }

  878. }

  879. 

  880. #define PAETH(cpu, abs3)\

  881. static void add_png_paeth_prediction_##cpu(uint8_t *dst, uint8_t *src, uint8_t *top, int w, int bpp)\

  882. {\

  883.     x86_reg i = -bpp;\

  884.     x86_reg end = w-3;\

  885.     __asm__ volatile(\

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

  887.         "movd    (%1,%0), %%mm0 \n"\

  888.         "movd    (%2,%0), %%mm1 \n"\

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

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

  891.         "add       %4, %0 \n"\

  892.         "1: \n"\

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

  894.         "movd    (%2,%0), %%mm1 \n"\

  895.         "movq      %%mm2, %%mm3 \n"\

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

  897.         "movq      %%mm2, %%mm4 \n"\

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

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

  900.         "movq      %%mm3, %%mm5 \n"\

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

  902.         abs3\

  903.         "movq      %%mm4, %%mm6 \n"\

  904.         "pminsw    %%mm5, %%mm6 \n"\

  905.         "pcmpgtw   %%mm6, %%mm3 \n"\

  906.         "pcmpgtw   %%mm5, %%mm4 \n"\

  907.         "movq      %%mm4, %%mm6 \n"\

  908.         "pand      %%mm3, %%mm4 \n"\

  909.         "pandn     %%mm3, %%mm6 \n"\

  910.         "pandn     %%mm0, %%mm3 \n"\

  911.         "movd    (%3,%0), %%mm0 \n"\

  912.         "pand      %%mm1, %%mm6 \n"\

  913.         "pand      %%mm4, %%mm2 \n"\

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

  915.         "movq      %6,    %%mm5 \n"\

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

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

  918.         "paddw     %%mm3, %%mm0 \n"\

  919.         "pand      %%mm5, %%mm0 \n"\

  920.         "movq      %%mm0, %%mm3 \n"\

  921.         "packuswb  %%mm3, %%mm3 \n"\

  922.         "movd      %%mm3, (%1,%0) \n"\

  923.         "add       %4, %0 \n"\

  924.         "cmp       %5, %0 \n"\

  925.         "jle 1b \n"\

  926.         :"+r"(i)\

  927.         :"r"(dst), "r"(top), "r"(src), "r"((x86_reg)bpp), "g"(end),\

  928.          "m"(ff_pw_255)\

  929.         :"memory"\

  930.     );\

  931. }

  932. 

  933. #define ABS3_MMX2\

  934.         "psubw     %%mm5, %%mm7 \n"\

  935.         "pmaxsw    %%mm7, %%mm5 \n"\

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

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

  938.         "psubw     %%mm3, %%mm6 \n"\

  939.         "psubw     %%mm4, %%mm7 \n"\

  940.         "pmaxsw    %%mm6, %%mm3 \n"\

  941.         "pmaxsw    %%mm7, %%mm4 \n"\

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

  943. 

  944. #define ABS3_SSSE3\

  945.         "pabsw     %%mm3, %%mm3 \n"\

  946.         "pabsw     %%mm4, %%mm4 \n"\

  947.         "pabsw     %%mm5, %%mm5 \n"

  948. 

  949. PAETH(mmx2, ABS3_MMX2)

  950. #if HAVE_SSSE3

  951. PAETH(ssse3, ABS3_SSSE3)

  952. #endif

  953. 

  954. #define QPEL_V_LOW(m3,m4,m5,m6, pw_20, pw_3, rnd, in0, in1, in2, in7, out, OP)\

  955.         "paddw " #m4 ", " #m3 "           \n\t" /* x1 */\

  956.         "movq "MANGLE(ff_pw_20)", %%mm4   \n\t" /* 20 */\

  957.         "pmullw " #m3 ", %%mm4            \n\t" /* 20x1 */\

  958.         "movq "#in7", " #m3 "             \n\t" /* d */\

  959.         "movq "#in0", %%mm5               \n\t" /* D */\

  960.         "paddw " #m3 ", %%mm5             \n\t" /* x4 */\

  961.         "psubw %%mm5, %%mm4               \n\t" /* 20x1 - x4 */\

  962.         "movq "#in1", %%mm5               \n\t" /* C */\

  963.         "movq "#in2", %%mm6               \n\t" /* B */\

  964.         "paddw " #m6 ", %%mm5             \n\t" /* x3 */\

  965.         "paddw " #m5 ", %%mm6             \n\t" /* x2 */\

  966.         "paddw %%mm6, %%mm6               \n\t" /* 2x2 */\

  967.         "psubw %%mm6, %%mm5               \n\t" /* -2x2 + x3 */\

  968.         "pmullw "MANGLE(ff_pw_3)", %%mm5  \n\t" /* -6x2 + 3x3 */\

  969.         "paddw " #rnd ", %%mm4            \n\t" /* x2 */\

  970.         "paddw %%mm4, %%mm5               \n\t" /* 20x1 - 6x2 + 3x3 - x4 */\

  971.         "psraw $5, %%mm5                  \n\t"\

  972.         "packuswb %%mm5, %%mm5            \n\t"\

  973.         OP(%%mm5, out, %%mm7, d)

  974. 

  975. #define QPEL_BASE(OPNAME, ROUNDER, RND, OP_MMX2, OP_3DNOW)\

  976. static void OPNAME ## mpeg4_qpel16_h_lowpass_mmx2(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\

  977.     uint64_t temp;\

  978. \

  979.     __asm__ volatile(\

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

  981.         "1:                               \n\t"\

  982.         "movq  (%0), %%mm0                \n\t" /* ABCDEFGH */\

  983.         "movq %%mm0, %%mm1                \n\t" /* ABCDEFGH */\

  984.         "movq %%mm0, %%mm2                \n\t" /* ABCDEFGH */\

  985.         "punpcklbw %%mm7, %%mm0           \n\t" /* 0A0B0C0D */\

  986.         "punpckhbw %%mm7, %%mm1           \n\t" /* 0E0F0G0H */\

  987.         "pshufw $0x90, %%mm0, %%mm5       \n\t" /* 0A0A0B0C */\

  988.         "pshufw $0x41, %%mm0, %%mm6       \n\t" /* 0B0A0A0B */\

  989.         "movq %%mm2, %%mm3                \n\t" /* ABCDEFGH */\

  990.         "movq %%mm2, %%mm4                \n\t" /* ABCDEFGH */\

  991.         "psllq $8, %%mm2                  \n\t" /* 0ABCDEFG */\

  992.         "psllq $16, %%mm3                 \n\t" /* 00ABCDEF */\

  993.         "psllq $24, %%mm4                 \n\t" /* 000ABCDE */\

  994.         "punpckhbw %%mm7, %%mm2           \n\t" /* 0D0E0F0G */\

  995.         "punpckhbw %%mm7, %%mm3           \n\t" /* 0C0D0E0F */\

  996.         "punpckhbw %%mm7, %%mm4           \n\t" /* 0B0C0D0E */\

  997.         "paddw %%mm3, %%mm5               \n\t" /* b */\

  998.         "paddw %%mm2, %%mm6               \n\t" /* c */\

  999.         "paddw %%mm5, %%mm5               \n\t" /* 2b */\

  1000.         "psubw %%mm5, %%mm6               \n\t" /* c - 2b */\

  1001.         "pshufw $0x06, %%mm0, %%mm5       \n\t" /* 0C0B0A0A */\

  1002.         "pmullw "MANGLE(ff_pw_3)", %%mm6  \n\t" /* 3c - 6b */\

  1003.         "paddw %%mm4, %%mm0               \n\t" /* a */\

  1004.         "paddw %%mm1, %%mm5               \n\t" /* d */\

  1005.         "pmullw "MANGLE(ff_pw_20)", %%mm0 \n\t" /* 20a */\

  1006.         "psubw %%mm5, %%mm0               \n\t" /* 20a - d */\

  1007.         "paddw %6, %%mm6                  \n\t"\

  1008.         "paddw %%mm6, %%mm0               \n\t" /* 20a - 6b + 3c - d */\

  1009.         "psraw $5, %%mm0                  \n\t"\

  1010.         "movq %%mm0, %5                   \n\t"\

  1011.         /* mm1=EFGH, mm2=DEFG, mm3=CDEF, mm4=BCDE, mm7=0 */\

  1012.         \

  1013.         "movq 5(%0), %%mm0                \n\t" /* FGHIJKLM */\

  1014.         "movq %%mm0, %%mm5                \n\t" /* FGHIJKLM */\

  1015.         "movq %%mm0, %%mm6                \n\t" /* FGHIJKLM */\

  1016.         "psrlq $8, %%mm0                  \n\t" /* GHIJKLM0 */\

  1017.         "psrlq $16, %%mm5                 \n\t" /* HIJKLM00 */\

  1018.         "punpcklbw %%mm7, %%mm0           \n\t" /* 0G0H0I0J */\

  1019.         "punpcklbw %%mm7, %%mm5           \n\t" /* 0H0I0J0K */\

  1020.         "paddw %%mm0, %%mm2               \n\t" /* b */\

  1021.         "paddw %%mm5, %%mm3               \n\t" /* c */\

  1022.         "paddw %%mm2, %%mm2               \n\t" /* 2b */\

  1023.         "psubw %%mm2, %%mm3               \n\t" /* c - 2b */\

  1024.         "movq %%mm6, %%mm2                \n\t" /* FGHIJKLM */\

  1025.         "psrlq $24, %%mm6                 \n\t" /* IJKLM000 */\

  1026.         "punpcklbw %%mm7, %%mm2           \n\t" /* 0F0G0H0I */\

  1027.         "punpcklbw %%mm7, %%mm6           \n\t" /* 0I0J0K0L */\

  1028.         "pmullw "MANGLE(ff_pw_3)", %%mm3  \n\t" /* 3c - 6b */\

  1029.         "paddw %%mm2, %%mm1               \n\t" /* a */\

  1030.         "paddw %%mm6, %%mm4               \n\t" /* d */\

  1031.         "pmullw "MANGLE(ff_pw_20)", %%mm1 \n\t" /* 20a */\

  1032.         "psubw %%mm4, %%mm3               \n\t" /* - 6b +3c - d */\

  1033.         "paddw %6, %%mm1                  \n\t"\

  1034.         "paddw %%mm1, %%mm3               \n\t" /* 20a - 6b +3c - d */\

  1035.         "psraw $5, %%mm3                  \n\t"\

  1036.         "movq %5, %%mm1                   \n\t"\

  1037.         "packuswb %%mm3, %%mm1            \n\t"\

  1038.         OP_MMX2(%%mm1, (%1),%%mm4, q)\

  1039.         /* mm0= GHIJ, mm2=FGHI, mm5=HIJK, mm6=IJKL, mm7=0 */\

  1040.         \

  1041.         "movq 9(%0), %%mm1                \n\t" /* JKLMNOPQ */\

  1042.         "movq %%mm1, %%mm4                \n\t" /* JKLMNOPQ */\

  1043.         "movq %%mm1, %%mm3                \n\t" /* JKLMNOPQ */\

  1044.         "psrlq $8, %%mm1                  \n\t" /* KLMNOPQ0 */\

  1045.         "psrlq $16, %%mm4                 \n\t" /* LMNOPQ00 */\

  1046.         "punpcklbw %%mm7, %%mm1           \n\t" /* 0K0L0M0N */\

  1047.         "punpcklbw %%mm7, %%mm4           \n\t" /* 0L0M0N0O */\

  1048.         "paddw %%mm1, %%mm5               \n\t" /* b */\

  1049.         "paddw %%mm4, %%mm0               \n\t" /* c */\

  1050.         "paddw %%mm5, %%mm5               \n\t" /* 2b */\

  1051.         "psubw %%mm5, %%mm0               \n\t" /* c - 2b */\

  1052.         "movq %%mm3, %%mm5                \n\t" /* JKLMNOPQ */\

  1053.         "psrlq $24, %%mm3                 \n\t" /* MNOPQ000 */\

  1054.         "pmullw "MANGLE(ff_pw_3)", %%mm0  \n\t" /* 3c - 6b */\

  1055.         "punpcklbw %%mm7, %%mm3           \n\t" /* 0M0N0O0P */\

  1056.         "paddw %%mm3, %%mm2               \n\t" /* d */\

  1057.         "psubw %%mm2, %%mm0               \n\t" /* -6b + 3c - d */\

  1058.         "movq %%mm5, %%mm2                \n\t" /* JKLMNOPQ */\

  1059.         "punpcklbw %%mm7, %%mm2           \n\t" /* 0J0K0L0M */\

  1060.         "punpckhbw %%mm7, %%mm5           \n\t" /* 0N0O0P0Q */\

  1061.         "paddw %%mm2, %%mm6               \n\t" /* a */\

  1062.         "pmullw "MANGLE(ff_pw_20)", %%mm6 \n\t" /* 20a */\

  1063.         "paddw %6, %%mm0                  \n\t"\

  1064.         "paddw %%mm6, %%mm0               \n\t" /* 20a - 6b + 3c - d */\

  1065.         "psraw $5, %%mm0                  \n\t"\

  1066.         /* mm1=KLMN, mm2=JKLM, mm3=MNOP, mm4=LMNO, mm5=NOPQ mm7=0 */\

  1067.         \

  1068.         "paddw %%mm5, %%mm3               \n\t" /* a */\

  1069.         "pshufw $0xF9, %%mm5, %%mm6       \n\t" /* 0O0P0Q0Q */\

  1070.         "paddw %%mm4, %%mm6               \n\t" /* b */\

  1071.         "pshufw $0xBE, %%mm5, %%mm4       \n\t" /* 0P0Q0Q0P */\

  1072.         "pshufw $0x6F, %%mm5, %%mm5       \n\t" /* 0Q0Q0P0O */\

  1073.         "paddw %%mm1, %%mm4               \n\t" /* c */\

  1074.         "paddw %%mm2, %%mm5               \n\t" /* d */\

  1075.         "paddw %%mm6, %%mm6               \n\t" /* 2b */\

  1076.         "psubw %%mm6, %%mm4               \n\t" /* c - 2b */\

  1077.         "pmullw "MANGLE(ff_pw_20)", %%mm3 \n\t" /* 20a */\

  1078.         "pmullw "MANGLE(ff_pw_3)", %%mm4  \n\t" /* 3c - 6b */\

  1079.         "psubw %%mm5, %%mm3               \n\t" /* -6b + 3c - d */\

  1080.         "paddw %6, %%mm4                  \n\t"\

  1081.         "paddw %%mm3, %%mm4               \n\t" /* 20a - 6b + 3c - d */\

  1082.         "psraw $5, %%mm4                  \n\t"\

  1083.         "packuswb %%mm4, %%mm0            \n\t"\

  1084.         OP_MMX2(%%mm0, 8(%1), %%mm4, q)\

  1085.         \

  1086.         "add %3, %0                       \n\t"\

  1087.         "add %4, %1                       \n\t"\

  1088.         "decl %2                          \n\t"\

  1089.         " jnz 1b                          \n\t"\

  1090.         : "+a"(src), "+c"(dst), "+D"(h)\

  1091.         : "d"((x86_reg)srcStride), "S"((x86_reg)dstStride), /*"m"(ff_pw_20), "m"(ff_pw_3),*/ "m"(temp), "m"(ROUNDER)\

  1092.         : "memory"\

  1093.     );\

  1094. }\

  1095. \

  1096. static void OPNAME ## mpeg4_qpel16_h_lowpass_3dnow(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\

  1097.     int i;\

  1098.     int16_t temp[16];\

  1099.     /* quick HACK, XXX FIXME MUST be optimized */\

  1100.     for(i=0; i<h; i++)\

  1101.     {\

  1102.         temp[ 0]= (src[ 0]+src[ 1])*20 - (src[ 0]+src[ 2])*6 + (src[ 1]+src[ 3])*3 - (src[ 2]+src[ 4]);\

  1103.         temp[ 1]= (src[ 1]+src[ 2])*20 - (src[ 0]+src[ 3])*6 + (src[ 0]+src[ 4])*3 - (src[ 1]+src[ 5]);\

  1104.         temp[ 2]= (src[ 2]+src[ 3])*20 - (src[ 1]+src[ 4])*6 + (src[ 0]+src[ 5])*3 - (src[ 0]+src[ 6]);\

  1105.         temp[ 3]= (src[ 3]+src[ 4])*20 - (src[ 2]+src[ 5])*6 + (src[ 1]+src[ 6])*3 - (src[ 0]+src[ 7]);\

  1106.         temp[ 4]= (src[ 4]+src[ 5])*20 - (src[ 3]+src[ 6])*6 + (src[ 2]+src[ 7])*3 - (src[ 1]+src[ 8]);\

  1107.         temp[ 5]= (src[ 5]+src[ 6])*20 - (src[ 4]+src[ 7])*6 + (src[ 3]+src[ 8])*3 - (src[ 2]+src[ 9]);\

  1108.         temp[ 6]= (src[ 6]+src[ 7])*20 - (src[ 5]+src[ 8])*6 + (src[ 4]+src[ 9])*3 - (src[ 3]+src[10]);\

  1109.         temp[ 7]= (src[ 7]+src[ 8])*20 - (src[ 6]+src[ 9])*6 + (src[ 5]+src[10])*3 - (src[ 4]+src[11]);\

  1110.         temp[ 8]= (src[ 8]+src[ 9])*20 - (src[ 7]+src[10])*6 + (src[ 6]+src[11])*3 - (src[ 5]+src[12]);\

  1111.         temp[ 9]= (src[ 9]+src[10])*20 - (src[ 8]+src[11])*6 + (src[ 7]+src[12])*3 - (src[ 6]+src[13]);\

  1112.         temp[10]= (src[10]+src[11])*20 - (src[ 9]+src[12])*6 + (src[ 8]+src[13])*3 - (src[ 7]+src[14]);\

  1113.         temp[11]= (src[11]+src[12])*20 - (src[10]+src[13])*6 + (src[ 9]+src[14])*3 - (src[ 8]+src[15]);\

  1114.         temp[12]= (src[12]+src[13])*20 - (src[11]+src[14])*6 + (src[10]+src[15])*3 - (src[ 9]+src[16]);\

  1115.         temp[13]= (src[13]+src[14])*20 - (src[12]+src[15])*6 + (src[11]+src[16])*3 - (src[10]+src[16]);\

  1116.         temp[14]= (src[14]+src[15])*20 - (src[13]+src[16])*6 + (src[12]+src[16])*3 - (src[11]+src[15]);\

  1117.         temp[15]= (src[15]+src[16])*20 - (src[14]+src[16])*6 + (src[13]+src[15])*3 - (src[12]+src[14]);\

  1118.         __asm__ volatile(\

  1119.             "movq (%0), %%mm0               \n\t"\

  1120.             "movq 8(%0), %%mm1              \n\t"\

  1121.             "paddw %2, %%mm0                \n\t"\

  1122.             "paddw %2, %%mm1                \n\t"\

  1123.             "psraw $5, %%mm0                \n\t"\

  1124.             "psraw $5, %%mm1                \n\t"\

  1125.             "packuswb %%mm1, %%mm0          \n\t"\

  1126.             OP_3DNOW(%%mm0, (%1), %%mm1, q)\

  1127.             "movq 16(%0), %%mm0             \n\t"\

  1128.             "movq 24(%0), %%mm1             \n\t"\

  1129.             "paddw %2, %%mm0                \n\t"\

  1130.             "paddw %2, %%mm1                \n\t"\

  1131.             "psraw $5, %%mm0                \n\t"\

  1132.             "psraw $5, %%mm1                \n\t"\

  1133.             "packuswb %%mm1, %%mm0          \n\t"\

  1134.             OP_3DNOW(%%mm0, 8(%1), %%mm1, q)\

  1135.             :: "r"(temp), "r"(dst), "m"(ROUNDER)\

  1136.             : "memory"\

  1137.         );\

  1138.         dst+=dstStride;\

  1139.         src+=srcStride;\

  1140.     }\

  1141. }\

  1142. \

  1143. static void OPNAME ## mpeg4_qpel8_h_lowpass_mmx2(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\

  1144.     __asm__ volatile(\

  1145.         "pxor %%mm7, %%mm7                \n\t"\

  1146.         "1:                               \n\t"\

  1147.         "movq  (%0), %%mm0                \n\t" /* ABCDEFGH */\

  1148.         "movq %%mm0, %%mm1                \n\t" /* ABCDEFGH */\

  1149.         "movq %%mm0, %%mm2                \n\t" /* ABCDEFGH */\

  1150.         "punpcklbw %%mm7, %%mm0           \n\t" /* 0A0B0C0D */\

  1151.         "punpckhbw %%mm7, %%mm1           \n\t" /* 0E0F0G0H */\

  1152.         "pshufw $0x90, %%mm0, %%mm5       \n\t" /* 0A0A0B0C */\

  1153.         "pshufw $0x41, %%mm0, %%mm6       \n\t" /* 0B0A0A0B */\

  1154.         "movq %%mm2, %%mm3                \n\t" /* ABCDEFGH */\

  1155.         "movq %%mm2, %%mm4                \n\t" /* ABCDEFGH */\

  1156.         "psllq $8, %%mm2                  \n\t" /* 0ABCDEFG */\

  1157.         "psllq $16, %%mm3                 \n\t" /* 00ABCDEF */\

  1158.         "psllq $24, %%mm4                 \n\t" /* 000ABCDE */\

  1159.         "punpckhbw %%mm7, %%mm2           \n\t" /* 0D0E0F0G */\

  1160.         "punpckhbw %%mm7, %%mm3           \n\t" /* 0C0D0E0F */\

  1161.         "punpckhbw %%mm7, %%mm4           \n\t" /* 0B0C0D0E */\

  1162.         "paddw %%mm3, %%mm5               \n\t" /* b */\

  1163.         "paddw %%mm2, %%mm6               \n\t" /* c */\

  1164.         "paddw %%mm5, %%mm5               \n\t" /* 2b */\

  1165.         "psubw %%mm5, %%mm6               \n\t" /* c - 2b */\

  1166.         "pshufw $0x06, %%mm0, %%mm5       \n\t" /* 0C0B0A0A */\

  1167.         "pmullw "MANGLE(ff_pw_3)", %%mm6  \n\t" /* 3c - 6b */\

  1168.         "paddw %%mm4, %%mm0               \n\t" /* a */\

  1169.         "paddw %%mm1, %%mm5               \n\t" /* d */\

  1170.         "pmullw "MANGLE(ff_pw_20)", %%mm0 \n\t" /* 20a */\

  1171.         "psubw %%mm5, %%mm0               \n\t" /* 20a - d */\

  1172.         "paddw %5, %%mm6                  \n\t"\

  1173.         "paddw %%mm6, %%mm0               \n\t" /* 20a - 6b + 3c - d */\

  1174.         "psraw $5, %%mm0                  \n\t"\

  1175.         /* mm1=EFGH, mm2=DEFG, mm3=CDEF, mm4=BCDE, mm7=0 */\

  1176.         \

  1177.         "movd 5(%0), %%mm5                \n\t" /* FGHI */\

  1178.         "punpcklbw %%mm7, %%mm5           \n\t" /* 0F0G0H0I */\

  1179.         "pshufw $0xF9, %%mm5, %%mm6       \n\t" /* 0G0H0I0I */\

  1180.         "paddw %%mm5, %%mm1               \n\t" /* a */\

  1181.         "paddw %%mm6, %%mm2               \n\t" /* b */\

  1182.         "pshufw $0xBE, %%mm5, %%mm6       \n\t" /* 0H0I0I0H */\

  1183.         "pshufw $0x6F, %%mm5, %%mm5       \n\t" /* 0I0I0H0G */\

  1184.         "paddw %%mm6, %%mm3               \n\t" /* c */\

  1185.         "paddw %%mm5, %%mm4               \n\t" /* d */\

  1186.         "paddw %%mm2, %%mm2               \n\t" /* 2b */\

  1187.         "psubw %%mm2, %%mm3               \n\t" /* c - 2b */\

  1188.         "pmullw "MANGLE(ff_pw_20)", %%mm1 \n\t" /* 20a */\

  1189.         "pmullw "MANGLE(ff_pw_3)", %%mm3  \n\t" /* 3c - 6b */\

  1190.         "psubw %%mm4, %%mm3               \n\t" /* -6b + 3c - d */\

  1191.         "paddw %5, %%mm1                  \n\t"\

  1192.         "paddw %%mm1, %%mm3               \n\t" /* 20a - 6b + 3c - d */\

  1193.         "psraw $5, %%mm3                  \n\t"\

  1194.         "packuswb %%mm3, %%mm0            \n\t"\

  1195.         OP_MMX2(%%mm0, (%1), %%mm4, q)\

  1196.         \

  1197.         "add %3, %0                       \n\t"\

  1198.         "add %4, %1                       \n\t"\

  1199.         "decl %2                          \n\t"\

  1200.         " jnz 1b                          \n\t"\

  1201.         : "+a"(src), "+c"(dst), "+d"(h)\

  1202.         : "S"((x86_reg)srcStride), "D"((x86_reg)dstStride), /*"m"(ff_pw_20), "m"(ff_pw_3),*/ "m"(ROUNDER)\

  1203.         : "memory"\

  1204.     );\

  1205. }\

  1206. \

  1207. static void OPNAME ## mpeg4_qpel8_h_lowpass_3dnow(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\

  1208.     int i;\

  1209.     int16_t temp[8];\

  1210.     /* quick HACK, XXX FIXME MUST be optimized */\

  1211.     for(i=0; i<h; i++)\

  1212.     {\

  1213.         temp[ 0]= (src[ 0]+src[ 1])*20 - (src[ 0]+src[ 2])*6 + (src[ 1]+src[ 3])*3 - (src[ 2]+src[ 4]);\

  1214.         temp[ 1]= (src[ 1]+src[ 2])*20 - (src[ 0]+src[ 3])*6 + (src[ 0]+src[ 4])*3 - (src[ 1]+src[ 5]);\

  1215.         temp[ 2]= (src[ 2]+src[ 3])*20 - (src[ 1]+src[ 4])*6 + (src[ 0]+src[ 5])*3 - (src[ 0]+src[ 6]);\

  1216.         temp[ 3]= (src[ 3]+src[ 4])*20 - (src[ 2]+src[ 5])*6 + (src[ 1]+src[ 6])*3 - (src[ 0]+src[ 7]);\

  1217.         temp[ 4]= (src[ 4]+src[ 5])*20 - (src[ 3]+src[ 6])*6 + (src[ 2]+src[ 7])*3 - (src[ 1]+src[ 8]);\

  1218.         temp[ 5]= (src[ 5]+src[ 6])*20 - (src[ 4]+src[ 7])*6 + (src[ 3]+src[ 8])*3 - (src[ 2]+src[ 8]);\

  1219.         temp[ 6]= (src[ 6]+src[ 7])*20 - (src[ 5]+src[ 8])*6 + (src[ 4]+src[ 8])*3 - (src[ 3]+src[ 7]);\

  1220.         temp[ 7]= (src[ 7]+src[ 8])*20 - (src[ 6]+src[ 8])*6 + (src[ 5]+src[ 7])*3 - (src[ 4]+src[ 6]);\

  1221.         __asm__ volatile(\

  1222.             "movq (%0), %%mm0           \n\t"\

  1223.             "movq 8(%0), %%mm1          \n\t"\

  1224.             "paddw %2, %%mm0            \n\t"\

  1225.             "paddw %2, %%mm1            \n\t"\

  1226.             "psraw $5, %%mm0            \n\t"\

  1227.             "psraw $5, %%mm1            \n\t"\

  1228.             "packuswb %%mm1, %%mm0      \n\t"\

  1229.             OP_3DNOW(%%mm0, (%1), %%mm1, q)\

  1230.             :: "r"(temp), "r"(dst), "m"(ROUNDER)\

  1231.             :"memory"\

  1232.         );\

  1233.         dst+=dstStride;\

  1234.         src+=srcStride;\

  1235.     }\

  1236. }

  1237. 

  1238. #define QPEL_OP(OPNAME, ROUNDER, RND, OP, MMX)\

  1239. \

  1240. static void OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\

  1241.     uint64_t temp[17*4];\

  1242.     uint64_t *temp_ptr= temp;\

  1243.     int count= 17;\

  1244. \

  1245.     /*FIXME unroll */\

  1246.     __asm__ volatile(\

  1247.         "pxor %%mm7, %%mm7              \n\t"\

  1248.         "1:                             \n\t"\

  1249.         "movq (%0), %%mm0               \n\t"\

  1250.         "movq (%0), %%mm1               \n\t"\

  1251.         "movq 8(%0), %%mm2              \n\t"\

  1252.         "movq 8(%0), %%mm3              \n\t"\

  1253.         "punpcklbw %%mm7, %%mm0         \n\t"\

  1254.         "punpckhbw %%mm7, %%mm1         \n\t"\

  1255.         "punpcklbw %%mm7, %%mm2         \n\t"\

  1256.         "punpckhbw %%mm7, %%mm3         \n\t"\

  1257.         "movq %%mm0, (%1)               \n\t"\

  1258.         "movq %%mm1, 17*8(%1)           \n\t"\

  1259.         "movq %%mm2, 2*17*8(%1)         \n\t"\

  1260.         "movq %%mm3, 3*17*8(%1)         \n\t"\

  1261.         "add $8, %1                     \n\t"\

  1262.         "add %3, %0                     \n\t"\

  1263.         "decl %2                        \n\t"\

  1264.         " jnz 1b                        \n\t"\

  1265.         : "+r" (src), "+r" (temp_ptr), "+r"(count)\

  1266.         : "r" ((x86_reg)srcStride)\

  1267.         : "memory"\

  1268.     );\

  1269.     \

  1270.     temp_ptr= temp;\

  1271.     count=4;\

  1272.     \

  1273. /*FIXME reorder for speed */\

  1274.     __asm__ volatile(\

  1275.         /*"pxor %%mm7, %%mm7              \n\t"*/\

  1276.         "1:                             \n\t"\

  1277.         "movq (%0), %%mm0               \n\t"\

  1278.         "movq 8(%0), %%mm1              \n\t"\

  1279.         "movq 16(%0), %%mm2             \n\t"\

  1280.         "movq 24(%0), %%mm3             \n\t"\

  1281.         QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 16(%0),  8(%0),   (%0), 32(%0), (%1), OP)\

  1282.         QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5,  8(%0),   (%0),   (%0), 40(%0), (%1, %3), OP)\

  1283.         "add %4, %1                     \n\t"\

  1284.         QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5,   (%0),   (%0),  8(%0), 48(%0), (%1), OP)\

  1285.         \

  1286.         QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5,   (%0),  8(%0), 16(%0), 56(%0), (%1, %3), OP)\

  1287.         "add %4, %1                     \n\t"\

  1288.         QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5,  8(%0), 16(%0), 24(%0), 64(%0), (%1), OP)\

  1289.         QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 16(%0), 24(%0), 32(%0), 72(%0), (%1, %3), OP)\

  1290.         "add %4, %1                     \n\t"\

  1291.         QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 24(%0), 32(%0), 40(%0), 80(%0), (%1), OP)\

  1292.         QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 32(%0), 40(%0), 48(%0), 88(%0), (%1, %3), OP)\

  1293.         "add %4, %1                     \n\t"\

  1294.         QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 40(%0), 48(%0), 56(%0), 96(%0), (%1), OP)\

  1295.         QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 48(%0), 56(%0), 64(%0),104(%0), (%1, %3), OP)\

  1296.         "add %4, %1                     \n\t"\

  1297.         QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 56(%0), 64(%0), 72(%0),112(%0), (%1), OP)\

  1298.         QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 64(%0), 72(%0), 80(%0),120(%0), (%1, %3), OP)\

  1299.         "add %4, %1                     \n\t"\

  1300.         QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 72(%0), 80(%0), 88(%0),128(%0), (%1), OP)\

  1301.         \

  1302.         QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 80(%0), 88(%0), 96(%0),128(%0), (%1, %3), OP)\

  1303.         "add %4, %1                     \n\t"  \

  1304.         QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 88(%0), 96(%0),104(%0),120(%0), (%1), OP)\

  1305.         QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 96(%0),104(%0),112(%0),112(%0), (%1, %3), OP)\

  1306.         \

  1307.         "add $136, %0                   \n\t"\

  1308.         "add %6, %1                     \n\t"\

  1309.         "decl %2                        \n\t"\

  1310.         " jnz 1b                        \n\t"\

  1311.         \

  1312.         : "+r"(temp_ptr), "+r"(dst), "+g"(count)\

  1313.         : "r"((x86_reg)dstStride), "r"(2*(x86_reg)dstStride), /*"m"(ff_pw_20), "m"(ff_pw_3),*/ "m"(ROUNDER), "g"(4-14*(x86_reg)dstStride)\

  1314.         :"memory"\

  1315.     );\

  1316. }\

  1317. \

  1318. static void OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\

  1319.     uint64_t temp[9*2];\

  1320.     uint64_t *temp_ptr= temp;\

  1321.     int count= 9;\

  1322. \

  1323.     /*FIXME unroll */\

  1324.     __asm__ volatile(\

  1325.         "pxor %%mm7, %%mm7              \n\t"\

  1326.         "1:                             \n\t"\

  1327.         "movq (%0), %%mm0               \n\t"\

  1328.         "movq (%0), %%mm1               \n\t"\

  1329.         "punpcklbw %%mm7, %%mm0         \n\t"\

  1330.         "punpckhbw %%mm7, %%mm1         \n\t"\

  1331.         "movq %%mm0, (%1)               \n\t"\

  1332.         "movq %%mm1, 9*8(%1)            \n\t"\

  1333.         "add $8, %1                     \n\t"\

  1334.         "add %3, %0                     \n\t"\

  1335.         "decl %2                        \n\t"\

  1336.         " jnz 1b                        \n\t"\

  1337.         : "+r" (src), "+r" (temp_ptr), "+r"(count)\

  1338.         : "r" ((x86_reg)srcStride)\

  1339.         : "memory"\

  1340.     );\

  1341.     \

  1342.     temp_ptr= temp;\

  1343.     count=2;\

  1344.     \

  1345. /*FIXME reorder for speed */\

  1346.     __asm__ volatile(\

  1347.         /*"pxor %%mm7, %%mm7              \n\t"*/\

  1348.         "1:                             \n\t"\

  1349.         "movq (%0), %%mm0               \n\t"\

  1350.         "movq 8(%0), %%mm1              \n\t"\

  1351.         "movq 16(%0), %%mm2             \n\t"\

  1352.         "movq 24(%0), %%mm3             \n\t"\

  1353.         QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 16(%0),  8(%0),   (%0), 32(%0), (%1), OP)\

  1354.         QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5,  8(%0),   (%0),   (%0), 40(%0), (%1, %3), OP)\

  1355.         "add %4, %1                     \n\t"\

  1356.         QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5,   (%0),   (%0),  8(%0), 48(%0), (%1), OP)\

  1357.         \

  1358.         QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5,   (%0),  8(%0), 16(%0), 56(%0), (%1, %3), OP)\

  1359.         "add %4, %1                     \n\t"\

  1360.         QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5,  8(%0), 16(%0), 24(%0), 64(%0), (%1), OP)\

  1361.         \

  1362.         QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 16(%0), 24(%0), 32(%0), 64(%0), (%1, %3), OP)\

  1363.         "add %4, %1                     \n\t"\

  1364.         QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 24(%0), 32(%0), 40(%0), 56(%0), (%1), OP)\

  1365.         QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 32(%0), 40(%0), 48(%0), 48(%0), (%1, %3), OP)\

  1366.                 \

  1367.         "add $72, %0                    \n\t"\

  1368.         "add %6, %1                     \n\t"\

  1369.         "decl %2                        \n\t"\

  1370.         " jnz 1b                        \n\t"\

  1371.          \

  1372.         : "+r"(temp_ptr), "+r"(dst), "+g"(count)\

  1373.         : "r"((x86_reg)dstStride), "r"(2*(x86_reg)dstStride), /*"m"(ff_pw_20), "m"(ff_pw_3),*/ "m"(ROUNDER), "g"(4-6*(x86_reg)dstStride)\

  1374.         : "memory"\

  1375.    );\

  1376. }\

  1377. \

  1378. static void OPNAME ## qpel8_mc00_ ## MMX (uint8_t *dst, uint8_t *src, int stride){\

  1379.     OPNAME ## pixels8_ ## MMX(dst, src, stride, 8);\

  1380. }\

  1381. \

  1382. static void OPNAME ## qpel8_mc10_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1383.     uint64_t temp[8];\

  1384.     uint8_t * const half= (uint8_t*)temp;\

  1385.     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(half, src, 8, stride, 8);\

  1386.     OPNAME ## pixels8_l2_ ## MMX(dst, src, half, stride, stride, 8);\

  1387. }\

  1388. \

  1389. static void OPNAME ## qpel8_mc20_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1390.     OPNAME ## mpeg4_qpel8_h_lowpass_ ## MMX(dst, src, stride, stride, 8);\

  1391. }\

  1392. \

  1393. static void OPNAME ## qpel8_mc30_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1394.     uint64_t temp[8];\

  1395.     uint8_t * const half= (uint8_t*)temp;\

  1396.     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(half, src, 8, stride, 8);\

  1397.     OPNAME ## pixels8_l2_ ## MMX(dst, src+1, half, stride, stride, 8);\

  1398. }\

  1399. \

  1400. static void OPNAME ## qpel8_mc01_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1401.     uint64_t temp[8];\

  1402.     uint8_t * const half= (uint8_t*)temp;\

  1403.     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(half, src, 8, stride);\

  1404.     OPNAME ## pixels8_l2_ ## MMX(dst, src, half, stride, stride, 8);\

  1405. }\

  1406. \

  1407. static void OPNAME ## qpel8_mc02_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1408.     OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, src, stride, stride);\

  1409. }\

  1410. \

  1411. static void OPNAME ## qpel8_mc03_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1412.     uint64_t temp[8];\

  1413.     uint8_t * const half= (uint8_t*)temp;\

  1414.     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(half, src, 8, stride);\

  1415.     OPNAME ## pixels8_l2_ ## MMX(dst, src+stride, half, stride, stride, 8);\

  1416. }\

  1417. static void OPNAME ## qpel8_mc11_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1418.     uint64_t half[8 + 9];\

  1419.     uint8_t * const halfH= ((uint8_t*)half) + 64;\

  1420.     uint8_t * const halfHV= ((uint8_t*)half);\

  1421.     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\

  1422.     put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8, stride, 9);\

  1423.     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\

  1424.     OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV, stride, 8, 8);\

  1425. }\

  1426. static void OPNAME ## qpel8_mc31_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1427.     uint64_t half[8 + 9];\

  1428.     uint8_t * const halfH= ((uint8_t*)half) + 64;\

  1429.     uint8_t * const halfHV= ((uint8_t*)half);\

  1430.     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\

  1431.     put ## RND ## pixels8_l2_ ## MMX(halfH, src+1, halfH, 8, stride, 9);\

  1432.     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\

  1433.     OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV, stride, 8, 8);\

  1434. }\

  1435. static void OPNAME ## qpel8_mc13_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1436.     uint64_t half[8 + 9];\

  1437.     uint8_t * const halfH= ((uint8_t*)half) + 64;\

  1438.     uint8_t * const halfHV= ((uint8_t*)half);\

  1439.     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\

  1440.     put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8, stride, 9);\

  1441.     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\

  1442.     OPNAME ## pixels8_l2_ ## MMX(dst, halfH+8, halfHV, stride, 8, 8);\

  1443. }\

  1444. static void OPNAME ## qpel8_mc33_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1445.     uint64_t half[8 + 9];\

  1446.     uint8_t * const halfH= ((uint8_t*)half) + 64;\

  1447.     uint8_t * const halfHV= ((uint8_t*)half);\

  1448.     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\

  1449.     put ## RND ## pixels8_l2_ ## MMX(halfH, src+1, halfH, 8, stride, 9);\

  1450.     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\

  1451.     OPNAME ## pixels8_l2_ ## MMX(dst, halfH+8, halfHV, stride, 8, 8);\

  1452. }\

  1453. static void OPNAME ## qpel8_mc21_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1454.     uint64_t half[8 + 9];\

  1455.     uint8_t * const halfH= ((uint8_t*)half) + 64;\

  1456.     uint8_t * const halfHV= ((uint8_t*)half);\

  1457.     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\

  1458.     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\

  1459.     OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV, stride, 8, 8);\

  1460. }\

  1461. static void OPNAME ## qpel8_mc23_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1462.     uint64_t half[8 + 9];\

  1463.     uint8_t * const halfH= ((uint8_t*)half) + 64;\

  1464.     uint8_t * const halfHV= ((uint8_t*)half);\

  1465.     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\

  1466.     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\

  1467.     OPNAME ## pixels8_l2_ ## MMX(dst, halfH+8, halfHV, stride, 8, 8);\

  1468. }\

  1469. static void OPNAME ## qpel8_mc12_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1470.     uint64_t half[8 + 9];\

  1471.     uint8_t * const halfH= ((uint8_t*)half);\

  1472.     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\

  1473.     put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8, stride, 9);\

  1474.     OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH, stride, 8);\

  1475. }\

  1476. static void OPNAME ## qpel8_mc32_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1477.     uint64_t half[8 + 9];\

  1478.     uint8_t * const halfH= ((uint8_t*)half);\

  1479.     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\

  1480.     put ## RND ## pixels8_l2_ ## MMX(halfH, src+1, halfH, 8, stride, 9);\

  1481.     OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH, stride, 8);\

  1482. }\

  1483. static void OPNAME ## qpel8_mc22_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1484.     uint64_t half[9];\

  1485.     uint8_t * const halfH= ((uint8_t*)half);\

  1486.     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\

  1487.     OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH, stride, 8);\

  1488. }\

  1489. static void OPNAME ## qpel16_mc00_ ## MMX (uint8_t *dst, uint8_t *src, int stride){\

  1490.     OPNAME ## pixels16_ ## MMX(dst, src, stride, 16);\

  1491. }\

  1492. \

  1493. static void OPNAME ## qpel16_mc10_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1494.     uint64_t temp[32];\

  1495.     uint8_t * const half= (uint8_t*)temp;\

  1496.     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(half, src, 16, stride, 16);\

  1497.     OPNAME ## pixels16_l2_ ## MMX(dst, src, half, stride, stride, 16);\

  1498. }\

  1499. \

  1500. static void OPNAME ## qpel16_mc20_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1501.     OPNAME ## mpeg4_qpel16_h_lowpass_ ## MMX(dst, src, stride, stride, 16);\

  1502. }\

  1503. \

  1504. static void OPNAME ## qpel16_mc30_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1505.     uint64_t temp[32];\

  1506.     uint8_t * const half= (uint8_t*)temp;\

  1507.     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(half, src, 16, stride, 16);\

  1508.     OPNAME ## pixels16_l2_ ## MMX(dst, src+1, half, stride, stride, 16);\

  1509. }\

  1510. \

  1511. static void OPNAME ## qpel16_mc01_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1512.     uint64_t temp[32];\

  1513.     uint8_t * const half= (uint8_t*)temp;\

  1514.     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(half, src, 16, stride);\

  1515.     OPNAME ## pixels16_l2_ ## MMX(dst, src, half, stride, stride, 16);\

  1516. }\

  1517. \

  1518. static void OPNAME ## qpel16_mc02_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1519.     OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, src, stride, stride);\

  1520. }\

  1521. \

  1522. static void OPNAME ## qpel16_mc03_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1523.     uint64_t temp[32];\

  1524.     uint8_t * const half= (uint8_t*)temp;\

  1525.     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(half, src, 16, stride);\

  1526.     OPNAME ## pixels16_l2_ ## MMX(dst, src+stride, half, stride, stride, 16);\

  1527. }\

  1528. static void OPNAME ## qpel16_mc11_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1529.     uint64_t half[16*2 + 17*2];\

  1530.     uint8_t * const halfH= ((uint8_t*)half) + 256;\

  1531.     uint8_t * const halfHV= ((uint8_t*)half);\

  1532.     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\

  1533.     put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16, stride, 17);\

  1534.     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\

  1535.     OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV, stride, 16, 16);\

  1536. }\

  1537. static void OPNAME ## qpel16_mc31_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1538.     uint64_t half[16*2 + 17*2];\

  1539.     uint8_t * const halfH= ((uint8_t*)half) + 256;\

  1540.     uint8_t * const halfHV= ((uint8_t*)half);\

  1541.     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\

  1542.     put ## RND ## pixels16_l2_ ## MMX(halfH, src+1, halfH, 16, stride, 17);\

  1543.     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\

  1544.     OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV, stride, 16, 16);\

  1545. }\

  1546. static void OPNAME ## qpel16_mc13_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1547.     uint64_t half[16*2 + 17*2];\

  1548.     uint8_t * const halfH= ((uint8_t*)half) + 256;\

  1549.     uint8_t * const halfHV= ((uint8_t*)half);\

  1550.     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\

  1551.     put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16, stride, 17);\

  1552.     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\

  1553.     OPNAME ## pixels16_l2_ ## MMX(dst, halfH+16, halfHV, stride, 16, 16);\

  1554. }\

  1555. static void OPNAME ## qpel16_mc33_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1556.     uint64_t half[16*2 + 17*2];\

  1557.     uint8_t * const halfH= ((uint8_t*)half) + 256;\

  1558.     uint8_t * const halfHV= ((uint8_t*)half);\

  1559.     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\

  1560.     put ## RND ## pixels16_l2_ ## MMX(halfH, src+1, halfH, 16, stride, 17);\

  1561.     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\

  1562.     OPNAME ## pixels16_l2_ ## MMX(dst, halfH+16, halfHV, stride, 16, 16);\

  1563. }\

  1564. static void OPNAME ## qpel16_mc21_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1565.     uint64_t half[16*2 + 17*2];\

  1566.     uint8_t * const halfH= ((uint8_t*)half) + 256;\

  1567.     uint8_t * const halfHV= ((uint8_t*)half);\

  1568.     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\

  1569.     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\

  1570.     OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV, stride, 16, 16);\

  1571. }\

  1572. static void OPNAME ## qpel16_mc23_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1573.     uint64_t half[16*2 + 17*2];\

  1574.     uint8_t * const halfH= ((uint8_t*)half) + 256;\

  1575.     uint8_t * const halfHV= ((uint8_t*)half);\

  1576.     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\

  1577.     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\

  1578.     OPNAME ## pixels16_l2_ ## MMX(dst, halfH+16, halfHV, stride, 16, 16);\

  1579. }\

  1580. static void OPNAME ## qpel16_mc12_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1581.     uint64_t half[17*2];\

  1582.     uint8_t * const halfH= ((uint8_t*)half);\

  1583.     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\

  1584.     put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16, stride, 17);\

  1585.     OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH, stride, 16);\

  1586. }\

  1587. static void OPNAME ## qpel16_mc32_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1588.     uint64_t half[17*2];\

  1589.     uint8_t * const halfH= ((uint8_t*)half);\

  1590.     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\

  1591.     put ## RND ## pixels16_l2_ ## MMX(halfH, src+1, halfH, 16, stride, 17);\

  1592.     OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH, stride, 16);\

  1593. }\

  1594. static void OPNAME ## qpel16_mc22_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1595.     uint64_t half[17*2];\

  1596.     uint8_t * const halfH= ((uint8_t*)half);\

  1597.     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\

  1598.     OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH, stride, 16);\

  1599. }

  1600. 

  1601. #define PUT_OP(a,b,temp, size) "mov" #size " " #a ", " #b "        \n\t"

  1602. #define AVG_3DNOW_OP(a,b,temp, size) \

  1603. "mov" #size " " #b ", " #temp "   \n\t"\

  1604. "pavgusb " #temp ", " #a "        \n\t"\

  1605. "mov" #size " " #a ", " #b "      \n\t"

  1606. #define AVG_MMX2_OP(a,b,temp, size) \

  1607. "mov" #size " " #b ", " #temp "   \n\t"\

  1608. "pavgb " #temp ", " #a "          \n\t"\

  1609. "mov" #size " " #a ", " #b "      \n\t"

  1610. 

  1611. QPEL_BASE(put_       , ff_pw_16, _       , PUT_OP, PUT_OP)

  1612. QPEL_BASE(avg_       , ff_pw_16, _       , AVG_MMX2_OP, AVG_3DNOW_OP)

  1613. QPEL_BASE(put_no_rnd_, ff_pw_15, _no_rnd_, PUT_OP, PUT_OP)

  1614. QPEL_OP(put_       , ff_pw_16, _       , PUT_OP, 3dnow)

  1615. QPEL_OP(avg_       , ff_pw_16, _       , AVG_3DNOW_OP, 3dnow)

  1616. QPEL_OP(put_no_rnd_, ff_pw_15, _no_rnd_, PUT_OP, 3dnow)

  1617. QPEL_OP(put_       , ff_pw_16, _       , PUT_OP, mmx2)

  1618. QPEL_OP(avg_       , ff_pw_16, _       , AVG_MMX2_OP, mmx2)

  1619. QPEL_OP(put_no_rnd_, ff_pw_15, _no_rnd_, PUT_OP, mmx2)

  1620. 

  1621. /***********************************/

  1622. /* bilinear qpel: not compliant to any spec, only for -lavdopts fast */

  1623. 

  1624. #define QPEL_2TAP_XY(OPNAME, SIZE, MMX, XY, HPEL)\

  1625. static void OPNAME ## 2tap_qpel ## SIZE ## _mc ## XY ## _ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1626.     OPNAME ## pixels ## SIZE ## HPEL(dst, src, stride, SIZE);\

  1627. }

  1628. #define QPEL_2TAP_L3(OPNAME, SIZE, MMX, XY, S0, S1, S2)\

  1629. static void OPNAME ## 2tap_qpel ## SIZE ## _mc ## XY ## _ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1630.     OPNAME ## 2tap_qpel ## SIZE ## _l3_ ## MMX(dst, src+S0, stride, SIZE, S1, S2);\

  1631. }

  1632. 

  1633. #define QPEL_2TAP(OPNAME, SIZE, MMX)\

  1634. QPEL_2TAP_XY(OPNAME, SIZE, MMX, 20, _x2_ ## MMX)\

  1635. QPEL_2TAP_XY(OPNAME, SIZE, MMX, 02, _y2_ ## MMX)\

  1636. QPEL_2TAP_XY(OPNAME, SIZE, MMX, 22, _xy2_mmx)\

  1637. static const qpel_mc_func OPNAME ## 2tap_qpel ## SIZE ## _mc00_ ## MMX =\

  1638.                           OPNAME ## qpel ## SIZE ## _mc00_ ## MMX;\

  1639. static const qpel_mc_func OPNAME ## 2tap_qpel ## SIZE ## _mc21_ ## MMX =\

  1640.                           OPNAME ## 2tap_qpel ## SIZE ## _mc20_ ## MMX;\

  1641. static const qpel_mc_func OPNAME ## 2tap_qpel ## SIZE ## _mc12_ ## MMX =\

  1642.                           OPNAME ## 2tap_qpel ## SIZE ## _mc02_ ## MMX;\

  1643. static void OPNAME ## 2tap_qpel ## SIZE ## _mc32_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1644.     OPNAME ## pixels ## SIZE ## _y2_ ## MMX(dst, src+1, stride, SIZE);\

  1645. }\

  1646. static void OPNAME ## 2tap_qpel ## SIZE ## _mc23_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1647.     OPNAME ## pixels ## SIZE ## _x2_ ## MMX(dst, src+stride, stride, SIZE);\

  1648. }\

  1649. QPEL_2TAP_L3(OPNAME, SIZE, MMX, 10, 0,         1,       0)\

  1650. QPEL_2TAP_L3(OPNAME, SIZE, MMX, 30, 1,        -1,       0)\

  1651. QPEL_2TAP_L3(OPNAME, SIZE, MMX, 01, 0,         stride,  0)\

  1652. QPEL_2TAP_L3(OPNAME, SIZE, MMX, 03, stride,   -stride,  0)\

  1653. QPEL_2TAP_L3(OPNAME, SIZE, MMX, 11, 0,         stride,  1)\

  1654. QPEL_2TAP_L3(OPNAME, SIZE, MMX, 31, 1,         stride, -1)\

  1655. QPEL_2TAP_L3(OPNAME, SIZE, MMX, 13, stride,   -stride,  1)\

  1656. QPEL_2TAP_L3(OPNAME, SIZE, MMX, 33, stride+1, -stride, -1)\

  1657. 

  1658. QPEL_2TAP(put_, 16, mmx2)

  1659. QPEL_2TAP(avg_, 16, mmx2)

  1660. QPEL_2TAP(put_,  8, mmx2)

  1661. QPEL_2TAP(avg_,  8, mmx2)

  1662. QPEL_2TAP(put_, 16, 3dnow)

  1663. QPEL_2TAP(avg_, 16, 3dnow)

  1664. QPEL_2TAP(put_,  8, 3dnow)

  1665. QPEL_2TAP(avg_,  8, 3dnow)

  1666. 

  1667. 

  1668. #if HAVE_YASM

  1669. typedef void emu_edge_core_func (uint8_t *buf, const uint8_t *src,

  1670.                                  x86_reg linesize, x86_reg start_y,

  1671.                                  x86_reg end_y, x86_reg block_h,

  1672.                                  x86_reg start_x, x86_reg end_x,

  1673.                                  x86_reg block_w);

  1674. extern emu_edge_core_func ff_emu_edge_core_mmx;

  1675. extern emu_edge_core_func ff_emu_edge_core_sse;

  1676. 

  1677. static av_always_inline

  1678. void emulated_edge_mc(uint8_t *buf, const uint8_t *src, int linesize,

  1679.                       int block_w, int block_h,

  1680.                       int src_x, int src_y, int w, int h,

  1681.                       emu_edge_core_func *core_fn)

  1682. {

  1683.     int start_y, start_x, end_y, end_x, src_y_add=0;

  1684. 

  1685.     if(src_y>= h){

  1686.         src_y_add = h-1-src_y;

  1687.         src_y=h-1;

  1688.     }else if(src_y<=-block_h){

  1689.         src_y_add = 1-block_h-src_y;

  1690.         src_y=1-block_h;

  1691.     }

  1692.     if(src_x>= w){

  1693.         src+= (w-1-src_x);

  1694.         src_x=w-1;

  1695.     }else if(src_x<=-block_w){

  1696.         src+= (1-block_w-src_x);

  1697.         src_x=1-block_w;

  1698.     }

  1699. 

  1700.     start_y= FFMAX(0, -src_y);

  1701.     start_x= FFMAX(0, -src_x);

  1702.     end_y= FFMIN(block_h, h-src_y);

  1703.     end_x= FFMIN(block_w, w-src_x);

  1704.     assert(start_x < end_x && block_w > 0);

  1705.     assert(start_y < end_y && block_h > 0);

  1706. 

  1707.     // fill in the to-be-copied part plus all above/below

  1708.     src += (src_y_add+start_y)*linesize + start_x;

  1709.     buf += start_x;

  1710.     core_fn(buf, src, linesize, start_y, end_y, block_h, start_x, end_x, block_w);

  1711. }

  1712. 

  1713. #if ARCH_X86_32

  1714. static av_noinline

  1715. void emulated_edge_mc_mmx(uint8_t *buf, const uint8_t *src, int linesize,

  1716.                           int block_w, int block_h,

  1717.                           int src_x, int src_y, int w, int h)

  1718. {

  1719.     emulated_edge_mc(buf, src, linesize, block_w, block_h, src_x, src_y,

  1720.                      w, h, &ff_emu_edge_core_mmx);

  1721. }

  1722. #endif

  1723. static av_noinline

  1724. void emulated_edge_mc_sse(uint8_t *buf, const uint8_t *src, int linesize,

  1725.                           int block_w, int block_h,

  1726.                           int src_x, int src_y, int w, int h)

  1727. {

  1728.     emulated_edge_mc(buf, src, linesize, block_w, block_h, src_x, src_y,

  1729.                      w, h, &ff_emu_edge_core_sse);

  1730. }

  1731. #endif /* HAVE_YASM */

  1732. 

  1733. typedef void emulated_edge_mc_func (uint8_t *dst, const uint8_t *src,

  1734.                                     int linesize, int block_w, int block_h,

  1735.                                     int src_x, int src_y, int w, int h);

  1736. 

  1737. static av_always_inline

  1738. void gmc(uint8_t *dst, uint8_t *src, int stride, int h, int ox, int oy,

  1739.          int dxx, int dxy, int dyx, int dyy, int shift, int r, int width, int height,

  1740.          emulated_edge_mc_func *emu_edge_fn)

  1741. {

  1742.     const int w = 8;

  1743.     const int ix = ox>>(16+shift);

  1744.     const int iy = oy>>(16+shift);

  1745.     const int oxs = ox>>4;

  1746.     const int oys = oy>>4;

  1747.     const int dxxs = dxx>>4;

  1748.     const int dxys = dxy>>4;

  1749.     const int dyxs = dyx>>4;

  1750.     const int dyys = dyy>>4;

  1751.     const uint16_t r4[4] = {r,r,r,r};

  1752.     const uint16_t dxy4[4] = {dxys,dxys,dxys,dxys};

  1753.     const uint16_t dyy4[4] = {dyys,dyys,dyys,dyys};

  1754.     const uint64_t shift2 = 2*shift;

  1755.     uint8_t edge_buf[(h+1)*stride];

  1756.     int x, y;

  1757. 

  1758.     const int dxw = (dxx-(1<<(16+shift)))*(w-1);

  1759.     const int dyh = (dyy-(1<<(16+shift)))*(h-1);

  1760.     const int dxh = dxy*(h-1);

  1761.     const int dyw = dyx*(w-1);

  1762.     if( // non-constant fullpel offset (3% of blocks)

  1763.         ((ox^(ox+dxw)) | (ox^(ox+dxh)) | (ox^(ox+dxw+dxh)) |

  1764.          (oy^(oy+dyw)) | (oy^(oy+dyh)) | (oy^(oy+dyw+dyh))) >> (16+shift)

  1765.         // uses more than 16 bits of subpel mv (only at huge resolution)

  1766.         || (dxx|dxy|dyx|dyy)&15 )

  1767.     {

  1768.         //FIXME could still use mmx for some of the rows

  1769.         ff_gmc_c(dst, src, stride, h, ox, oy, dxx, dxy, dyx, dyy, shift, r, width, height);

  1770.         return;

  1771.     }

  1772. 

  1773.     src += ix + iy*stride;

  1774.     if( (unsigned)ix >= width-w ||

  1775.         (unsigned)iy >= height-h )

  1776.     {

  1777.         emu_edge_fn(edge_buf, src, stride, w+1, h+1, ix, iy, width, height);

  1778.         src = edge_buf;

  1779.     }

  1780. 

  1781.     __asm__ volatile(

  1782.         "movd         %0, %%mm6 \n\t"

  1783.         "pxor      %%mm7, %%mm7 \n\t"

  1784.         "punpcklwd %%mm6, %%mm6 \n\t"

  1785.         "punpcklwd %%mm6, %%mm6 \n\t"

  1786.         :: "r"(1<<shift)

  1787.     );

  1788. 

  1789.     for(x=0; x<w; x+=4){

  1790.         uint16_t dx4[4] = { oxs - dxys + dxxs*(x+0),

  1791.                             oxs - dxys + dxxs*(x+1),

  1792.                             oxs - dxys + dxxs*(x+2),

  1793.                             oxs - dxys + dxxs*(x+3) };

  1794.         uint16_t dy4[4] = { oys - dyys + dyxs*(x+0),

  1795.                             oys - dyys + dyxs*(x+1),

  1796.                             oys - dyys + dyxs*(x+2),

  1797.                             oys - dyys + dyxs*(x+3) };

  1798. 

  1799.         for(y=0; y<h; y++){

  1800.             __asm__ volatile(

  1801.                 "movq   %0,  %%mm4 \n\t"

  1802.                 "movq   %1,  %%mm5 \n\t"

  1803.                 "paddw  %2,  %%mm4 \n\t"

  1804.                 "paddw  %3,  %%mm5 \n\t"

  1805.                 "movq   %%mm4, %0  \n\t"

  1806.                 "movq   %%mm5, %1  \n\t"

  1807.                 "psrlw  $12, %%mm4 \n\t"

  1808.                 "psrlw  $12, %%mm5 \n\t"

  1809.                 : "+m"(*dx4), "+m"(*dy4)

  1810.                 : "m"(*dxy4), "m"(*dyy4)

  1811.             );

  1812. 

  1813.             __asm__ volatile(

  1814.                 "movq   %%mm6, %%mm2 \n\t"

  1815.                 "movq   %%mm6, %%mm1 \n\t"

  1816.                 "psubw  %%mm4, %%mm2 \n\t"

  1817.                 "psubw  %%mm5, %%mm1 \n\t"

  1818.                 "movq   %%mm2, %%mm0 \n\t"

  1819.                 "movq   %%mm4, %%mm3 \n\t"

  1820.                 "pmullw %%mm1, %%mm0 \n\t" // (s-dx)*(s-dy)

  1821.                 "pmullw %%mm5, %%mm3 \n\t" // dx*dy

  1822.                 "pmullw %%mm5, %%mm2 \n\t" // (s-dx)*dy

  1823.                 "pmullw %%mm4, %%mm1 \n\t" // dx*(s-dy)

  1824. 

  1825.                 "movd   %4,    %%mm5 \n\t"

  1826.                 "movd   %3,    %%mm4 \n\t"

  1827.                 "punpcklbw %%mm7, %%mm5 \n\t"

  1828.                 "punpcklbw %%mm7, %%mm4 \n\t"

  1829.                 "pmullw %%mm5, %%mm3 \n\t" // src[1,1] * dx*dy

  1830.                 "pmullw %%mm4, %%mm2 \n\t" // src[0,1] * (s-dx)*dy

  1831. 

  1832.                 "movd   %2,    %%mm5 \n\t"

  1833.                 "movd   %1,    %%mm4 \n\t"

  1834.                 "punpcklbw %%mm7, %%mm5 \n\t"

  1835.                 "punpcklbw %%mm7, %%mm4 \n\t"

  1836.                 "pmullw %%mm5, %%mm1 \n\t" // src[1,0] * dx*(s-dy)

  1837.                 "pmullw %%mm4, %%mm0 \n\t" // src[0,0] * (s-dx)*(s-dy)

  1838.                 "paddw  %5,    %%mm1 \n\t"

  1839.                 "paddw  %%mm3, %%mm2 \n\t"

  1840.                 "paddw  %%mm1, %%mm0 \n\t"

  1841.                 "paddw  %%mm2, %%mm0 \n\t"

  1842. 

  1843.                 "psrlw    %6,    %%mm0 \n\t"

  1844.                 "packuswb %%mm0, %%mm0 \n\t"

  1845.                 "movd     %%mm0, %0    \n\t"

  1846. 

  1847.                 : "=m"(dst[x+y*stride])

  1848.                 : "m"(src[0]), "m"(src[1]),

  1849.                   "m"(src[stride]), "m"(src[stride+1]),

  1850.                   "m"(*r4), "m"(shift2)

  1851.             );

  1852.             src += stride;

  1853.         }

  1854.         src += 4-h*stride;

  1855.     }

  1856. }

  1857. 

  1858. #if HAVE_YASM

  1859. #if ARCH_X86_32

  1860. static void gmc_mmx(uint8_t *dst, uint8_t *src, int stride, int h, int ox, int oy,

  1861.                     int dxx, int dxy, int dyx, int dyy, int shift, int r, int width, int height)

  1862. {

  1863.     gmc(dst, src, stride, h, ox, oy, dxx, dxy, dyx, dyy, shift, r,

  1864.         width, height, &emulated_edge_mc_mmx);

  1865. }

  1866. #endif

  1867. static void gmc_sse(uint8_t *dst, uint8_t *src, int stride, int h, int ox, int oy,

  1868.                     int dxx, int dxy, int dyx, int dyy, int shift, int r, int width, int height)

  1869. {

  1870.     gmc(dst, src, stride, h, ox, oy, dxx, dxy, dyx, dyy, shift, r,

  1871.         width, height, &emulated_edge_mc_sse);

  1872. }

  1873. #else

  1874. static void gmc_mmx(uint8_t *dst, uint8_t *src, int stride, int h, int ox, int oy,

  1875.                     int dxx, int dxy, int dyx, int dyy, int shift, int r, int width, int height)

  1876. {

  1877.     gmc(dst, src, stride, h, ox, oy, dxx, dxy, dyx, dyy, shift, r,

  1878.         width, height, &ff_emulated_edge_mc_8);

  1879. }

  1880. #endif

  1881. 

  1882. #define PREFETCH(name, op) \

  1883. static void name(void *mem, int stride, int h){\

  1884.     const uint8_t *p= mem;\

  1885.     do{\

  1886.         __asm__ volatile(#op" %0" :: "m"(*p));\

  1887.         p+= stride;\

  1888.     }while(--h);\

  1889. }

  1890. PREFETCH(prefetch_mmx2,  prefetcht0)

  1891. PREFETCH(prefetch_3dnow, prefetch)

  1892. #undef PREFETCH

  1893. 

  1894. #include "h264_qpel_mmx.c"

  1895. 

  1896. void ff_put_h264_chroma_mc8_mmx_rnd   (uint8_t *dst, uint8_t *src,

  1897.                                        int stride, int h, int x, int y);

  1898. void ff_avg_h264_chroma_mc8_mmx2_rnd  (uint8_t *dst, uint8_t *src,

  1899.                                        int stride, int h, int x, int y);

  1900. void ff_avg_h264_chroma_mc8_3dnow_rnd (uint8_t *dst, uint8_t *src,

  1901.                                        int stride, int h, int x, int y);

  1902. 

  1903. void ff_put_h264_chroma_mc4_mmx       (uint8_t *dst, uint8_t *src,

  1904.                                        int stride, int h, int x, int y);

  1905. void ff_avg_h264_chroma_mc4_mmx2      (uint8_t *dst, uint8_t *src,

  1906.                                        int stride, int h, int x, int y);

  1907. void ff_avg_h264_chroma_mc4_3dnow     (uint8_t *dst, uint8_t *src,

  1908.                                        int stride, int h, int x, int y);

  1909. 

  1910. void ff_put_h264_chroma_mc2_mmx2      (uint8_t *dst, uint8_t *src,

  1911.                                        int stride, int h, int x, int y);

  1912. void ff_avg_h264_chroma_mc2_mmx2      (uint8_t *dst, uint8_t *src,

  1913.                                        int stride, int h, int x, int y);

  1914. 

  1915. void ff_put_h264_chroma_mc8_ssse3_rnd (uint8_t *dst, uint8_t *src,

  1916.                                        int stride, int h, int x, int y);

  1917. void ff_put_h264_chroma_mc4_ssse3     (uint8_t *dst, uint8_t *src,

  1918.                                        int stride, int h, int x, int y);

  1919. 

  1920. void ff_avg_h264_chroma_mc8_ssse3_rnd (uint8_t *dst, uint8_t *src,

  1921.                                        int stride, int h, int x, int y);

  1922. void ff_avg_h264_chroma_mc4_ssse3     (uint8_t *dst, uint8_t *src,

  1923.                                        int stride, int h, int x, int y);

  1924. 

  1925. #define CHROMA_MC(OP, NUM, DEPTH, OPT) \

  1926. void ff_ ## OP ## _h264_chroma_mc ## NUM ## _ ## DEPTH ## _ ## OPT \

  1927.                                       (uint8_t *dst, uint8_t *src,\

  1928.                                        int stride, int h, int x, int y);

  1929. 

  1930. CHROMA_MC(put, 2, 10, mmxext)

  1931. CHROMA_MC(avg, 2, 10, mmxext)

  1932. CHROMA_MC(put, 4, 10, mmxext)

  1933. CHROMA_MC(avg, 4, 10, mmxext)

  1934. CHROMA_MC(put, 8, 10, sse2)

  1935. CHROMA_MC(avg, 8, 10, sse2)

  1936. CHROMA_MC(put, 8, 10, avx)

  1937. CHROMA_MC(avg, 8, 10, avx)

  1938. 

  1939. /* CAVS specific */

  1940. void ff_put_cavs_qpel8_mc00_mmx2(uint8_t *dst, uint8_t *src, int stride) {

  1941.     put_pixels8_mmx(dst, src, stride, 8);

  1942. }

  1943. void ff_avg_cavs_qpel8_mc00_mmx2(uint8_t *dst, uint8_t *src, int stride) {

  1944.     avg_pixels8_mmx(dst, src, stride, 8);

  1945. }

  1946. void ff_put_cavs_qpel16_mc00_mmx2(uint8_t *dst, uint8_t *src, int stride) {

  1947.     put_pixels16_mmx(dst, src, stride, 16);

  1948. }

  1949. void ff_avg_cavs_qpel16_mc00_mmx2(uint8_t *dst, uint8_t *src, int stride) {

  1950.     avg_pixels16_mmx(dst, src, stride, 16);

  1951. }

  1952. 

  1953. /* VC1 specific */

  1954. void ff_put_vc1_mspel_mc00_mmx(uint8_t *dst, const uint8_t *src, int stride, int rnd) {

  1955.     put_pixels8_mmx(dst, src, stride, 8);

  1956. }

  1957. void ff_avg_vc1_mspel_mc00_mmx2(uint8_t *dst, const uint8_t *src, int stride, int rnd) {

  1958.     avg_pixels8_mmx2(dst, src, stride, 8);

  1959. }

  1960. 

  1961. /* XXX: those functions should be suppressed ASAP when all IDCTs are

  1962.    converted */

  1963. #if CONFIG_GPL

  1964. static void ff_libmpeg2mmx_idct_put(uint8_t *dest, int line_size, DCTELEM *block)

  1965. {

  1966.     ff_mmx_idct (block);

  1967.     ff_put_pixels_clamped_mmx(block, dest, line_size);

  1968. }

  1969. static void ff_libmpeg2mmx_idct_add(uint8_t *dest, int line_size, DCTELEM *block)

  1970. {

  1971.     ff_mmx_idct (block);

  1972.     ff_add_pixels_clamped_mmx(block, dest, line_size);

  1973. }

  1974. static void ff_libmpeg2mmx2_idct_put(uint8_t *dest, int line_size, DCTELEM *block)

  1975. {

  1976.     ff_mmxext_idct (block);

  1977.     ff_put_pixels_clamped_mmx(block, dest, line_size);

  1978. }

  1979. static void ff_libmpeg2mmx2_idct_add(uint8_t *dest, int line_size, DCTELEM *block)

  1980. {

  1981.     ff_mmxext_idct (block);

  1982.     ff_add_pixels_clamped_mmx(block, dest, line_size);

  1983. }

  1984. #endif

  1985. static void ff_idct_xvid_mmx_put(uint8_t *dest, int line_size, DCTELEM *block)

  1986. {

  1987.     ff_idct_xvid_mmx (block);

  1988.     ff_put_pixels_clamped_mmx(block, dest, line_size);

  1989. }

  1990. static void ff_idct_xvid_mmx_add(uint8_t *dest, int line_size, DCTELEM *block)

  1991. {

  1992.     ff_idct_xvid_mmx (block);

  1993.     ff_add_pixels_clamped_mmx(block, dest, line_size);

  1994. }

  1995. static void ff_idct_xvid_mmx2_put(uint8_t *dest, int line_size, DCTELEM *block)

  1996. {

  1997.     ff_idct_xvid_mmx2 (block);

  1998.     ff_put_pixels_clamped_mmx(block, dest, line_size);

  1999. }

  2000. static void ff_idct_xvid_mmx2_add(uint8_t *dest, int line_size, DCTELEM *block)

  2001. {

  2002.     ff_idct_xvid_mmx2 (block);

  2003.     ff_add_pixels_clamped_mmx(block, dest, line_size);

  2004. }

  2005. 

  2006. static void vorbis_inverse_coupling_3dnow(float *mag, float *ang, int blocksize)

  2007. {

  2008.     int i;

  2009.     __asm__ volatile("pxor %%mm7, %%mm7":);

  2010.     for(i=0; i<blocksize; i+=2) {

  2011.         __asm__ volatile(

  2012.             "movq    %0,    %%mm0 \n\t"

  2013.             "movq    %1,    %%mm1 \n\t"

  2014.             "movq    %%mm0, %%mm2 \n\t"

  2015.             "movq    %%mm1, %%mm3 \n\t"

  2016.             "pfcmpge %%mm7, %%mm2 \n\t" // m <= 0.0

  2017.             "pfcmpge %%mm7, %%mm3 \n\t" // a <= 0.0

  2018.             "pslld   $31,   %%mm2 \n\t" // keep only the sign bit

  2019.             "pxor    %%mm2, %%mm1 \n\t"

  2020.             "movq    %%mm3, %%mm4 \n\t"

  2021.             "pand    %%mm1, %%mm3 \n\t"

  2022.             "pandn   %%mm1, %%mm4 \n\t"

  2023.             "pfadd   %%mm0, %%mm3 \n\t" // a = m + ((a<0) & (a ^ sign(m)))

  2024.             "pfsub   %%mm4, %%mm0 \n\t" // m = m + ((a>0) & (a ^ sign(m)))

  2025.             "movq    %%mm3, %1    \n\t"

  2026.             "movq    %%mm0, %0    \n\t"

  2027.             :"+m"(mag[i]), "+m"(ang[i])

  2028.             ::"memory"

  2029.         );

  2030.     }

  2031.     __asm__ volatile("femms");

  2032. }

  2033. static void vorbis_inverse_coupling_sse(float *mag, float *ang, int blocksize)

  2034. {

  2035.     int i;

  2036. 

  2037.     __asm__ volatile(

  2038.             "movaps  %0,     %%xmm5 \n\t"

  2039.         ::"m"(ff_pdw_80000000[0])

  2040.     );

  2041.     for(i=0; i<blocksize; i+=4) {

  2042.         __asm__ volatile(

  2043.             "movaps  %0,     %%xmm0 \n\t"

  2044.             "movaps  %1,     %%xmm1 \n\t"

  2045.             "xorps   %%xmm2, %%xmm2 \n\t"

  2046.             "xorps   %%xmm3, %%xmm3 \n\t"

  2047.             "cmpleps %%xmm0, %%xmm2 \n\t" // m <= 0.0

  2048.             "cmpleps %%xmm1, %%xmm3 \n\t" // a <= 0.0

  2049.             "andps   %%xmm5, %%xmm2 \n\t" // keep only the sign bit

  2050.             "xorps   %%xmm2, %%xmm1 \n\t"

  2051.             "movaps  %%xmm3, %%xmm4 \n\t"

  2052.             "andps   %%xmm1, %%xmm3 \n\t"

  2053.             "andnps  %%xmm1, %%xmm4 \n\t"

  2054.             "addps   %%xmm0, %%xmm3 \n\t" // a = m + ((a<0) & (a ^ sign(m)))

  2055.             "subps   %%xmm4, %%xmm0 \n\t" // m = m + ((a>0) & (a ^ sign(m)))

  2056.             "movaps  %%xmm3, %1     \n\t"

  2057.             "movaps  %%xmm0, %0     \n\t"

  2058.             :"+m"(mag[i]), "+m"(ang[i])

  2059.             ::"memory"

  2060.         );

  2061.     }

  2062. }

  2063. 

  2064. #define IF1(x) x

  2065. #define IF0(x)

  2066. 

  2067. #define MIX5(mono,stereo)\

  2068.     __asm__ volatile(\

  2069.         "movss          0(%2), %%xmm5 \n"\

  2070.         "movss          8(%2), %%xmm6 \n"\

  2071.         "movss         24(%2), %%xmm7 \n"\

  2072.         "shufps    $0, %%xmm5, %%xmm5 \n"\

  2073.         "shufps    $0, %%xmm6, %%xmm6 \n"\

  2074.         "shufps    $0, %%xmm7, %%xmm7 \n"\

  2075.         "1: \n"\

  2076.         "movaps       (%0,%1), %%xmm0 \n"\

  2077.         "movaps  0x400(%0,%1), %%xmm1 \n"\

  2078.         "movaps  0x800(%0,%1), %%xmm2 \n"\

  2079.         "movaps  0xc00(%0,%1), %%xmm3 \n"\

  2080.         "movaps 0x1000(%0,%1), %%xmm4 \n"\

  2081.         "mulps         %%xmm5, %%xmm0 \n"\

  2082.         "mulps         %%xmm6, %%xmm1 \n"\

  2083.         "mulps         %%xmm5, %%xmm2 \n"\

  2084.         "mulps         %%xmm7, %%xmm3 \n"\

  2085.         "mulps         %%xmm7, %%xmm4 \n"\

  2086.  stereo("addps         %%xmm1, %%xmm0 \n")\

  2087.         "addps         %%xmm1, %%xmm2 \n"\

  2088.         "addps         %%xmm3, %%xmm0 \n"\

  2089.         "addps         %%xmm4, %%xmm2 \n"\

  2090.    mono("addps         %%xmm2, %%xmm0 \n")\

  2091.         "movaps  %%xmm0,      (%0,%1) \n"\

  2092.  stereo("movaps  %%xmm2, 0x400(%0,%1) \n")\

  2093.         "add $16, %0 \n"\

  2094.         "jl 1b \n"\

  2095.         :"+&r"(i)\

  2096.         :"r"(samples[0]+len), "r"(matrix)\

  2097.         :XMM_CLOBBERS("%xmm0", "%xmm1", "%xmm2", "%xmm3", \

  2098.                       "%xmm4", "%xmm5", "%xmm6", "%xmm7",)\

  2099.          "memory"\

  2100.     );

  2101. 

  2102. #define MIX_MISC(stereo)\

  2103.     __asm__ volatile(\

  2104.         "1: \n"\

  2105.         "movaps  (%3,%0), %%xmm0 \n"\

  2106.  stereo("movaps   %%xmm0, %%xmm1 \n")\

  2107.         "mulps    %%xmm4, %%xmm0 \n"\

  2108.  stereo("mulps    %%xmm5, %%xmm1 \n")\

  2109.         "lea 1024(%3,%0), %1 \n"\

  2110.         "mov %5, %2 \n"\

  2111.         "2: \n"\

  2112.         "movaps   (%1),   %%xmm2 \n"\

  2113.  stereo("movaps   %%xmm2, %%xmm3 \n")\

  2114.         "mulps   (%4,%2), %%xmm2 \n"\

  2115.  stereo("mulps 16(%4,%2), %%xmm3 \n")\

  2116.         "addps    %%xmm2, %%xmm0 \n"\

  2117.  stereo("addps    %%xmm3, %%xmm1 \n")\

  2118.         "add $1024, %1 \n"\

  2119.         "add $32, %2 \n"\

  2120.         "jl 2b \n"\

  2121.         "movaps   %%xmm0,     (%3,%0) \n"\

  2122.  stereo("movaps   %%xmm1, 1024(%3,%0) \n")\

  2123.         "add $16, %0 \n"\

  2124.         "jl 1b \n"\

  2125.         :"+&r"(i), "=&r"(j), "=&r"(k)\

  2126.         :"r"(samples[0]+len), "r"(matrix_simd+in_ch), "g"((intptr_t)-32*(in_ch-1))\

  2127.         :"memory"\

  2128.     );

  2129. 

  2130. static void ac3_downmix_sse(float (*samples)[256], float (*matrix)[2], int out_ch, int in_ch, int len)

  2131. {

  2132.     int (*matrix_cmp)[2] = (int(*)[2])matrix;

  2133.     intptr_t i,j,k;

  2134. 

  2135.     i = -len*sizeof(float);

  2136.     if(in_ch == 5 && out_ch == 2 && !(matrix_cmp[0][1]|matrix_cmp[2][0]|matrix_cmp[3][1]|matrix_cmp[4][0]|(matrix_cmp[1][0]^matrix_cmp[1][1])|(matrix_cmp[0][0]^matrix_cmp[2][1]))) {

  2137.         MIX5(IF0,IF1);

  2138.     } else if(in_ch == 5 && out_ch == 1 && matrix_cmp[0][0]==matrix_cmp[2][0] && matrix_cmp[3][0]==matrix_cmp[4][0]) {

  2139.         MIX5(IF1,IF0);

  2140.     } else {

  2141.         DECLARE_ALIGNED(16, float, matrix_simd)[AC3_MAX_CHANNELS][2][4];

  2142.         j = 2*in_ch*sizeof(float);

  2143.         __asm__ volatile(

  2144.             "1: \n"

  2145.             "sub $8, %0 \n"

  2146.             "movss     (%2,%0), %%xmm4 \n"

  2147.             "movss    4(%2,%0), %%xmm5 \n"

  2148.             "shufps $0, %%xmm4, %%xmm4 \n"

  2149.             "shufps $0, %%xmm5, %%xmm5 \n"

  2150.             "movaps %%xmm4,   (%1,%0,4) \n"

  2151.             "movaps %%xmm5, 16(%1,%0,4) \n"

  2152.             "jg 1b \n"

  2153.             :"+&r"(j)

  2154.             :"r"(matrix_simd), "r"(matrix)

  2155.             :"memory"

  2156.         );

  2157.         if(out_ch == 2) {

  2158.             MIX_MISC(IF1);

  2159.         } else {

  2160.             MIX_MISC(IF0);

  2161.         }

  2162.     }

  2163. }

  2164. 

  2165. static void vector_fmul_3dnow(float *dst, const float *src0, const float *src1, int len){

  2166.     x86_reg i = (len-4)*4;

  2167.     __asm__ volatile(

  2168.         "1: \n\t"

  2169.         "movq    (%2,%0), %%mm0 \n\t"

  2170.         "movq   8(%2,%0), %%mm1 \n\t"

  2171.         "pfmul   (%3,%0), %%mm0 \n\t"

  2172.         "pfmul  8(%3,%0), %%mm1 \n\t"

  2173.         "movq   %%mm0,  (%1,%0) \n\t"

  2174.         "movq   %%mm1, 8(%1,%0) \n\t"

  2175.         "sub  $16, %0 \n\t"

  2176.         "jge 1b \n\t"

  2177.         "femms  \n\t"

  2178.         :"+r"(i)

  2179.         :"r"(dst), "r"(src0), "r"(src1)

  2180.         :"memory"

  2181.     );

  2182. }

  2183. static void vector_fmul_sse(float *dst, const float *src0, const float *src1, int len){

  2184.     x86_reg i = (len-8)*4;

  2185.     __asm__ volatile(

  2186.         "1: \n\t"

  2187.         "movaps    (%2,%0), %%xmm0 \n\t"

  2188.         "movaps  16(%2,%0), %%xmm1 \n\t"

  2189.         "mulps     (%3,%0), %%xmm0 \n\t"

  2190.         "mulps   16(%3,%0), %%xmm1 \n\t"

  2191.         "movaps  %%xmm0,   (%1,%0) \n\t"

  2192.         "movaps  %%xmm1, 16(%1,%0) \n\t"

  2193.         "sub  $32, %0 \n\t"

  2194.         "jge 1b \n\t"

  2195.         :"+r"(i)

  2196.         :"r"(dst), "r"(src0), "r"(src1)

  2197.         :"memory"

  2198.     );

  2199. }

  2200. 

  2201. static void vector_fmul_reverse_3dnow2(float *dst, const float *src0, const float *src1, int len){

  2202.     x86_reg i = len*4-16;

  2203.     __asm__ volatile(

  2204.         "1: \n\t"

  2205.         "pswapd   8(%1), %%mm0 \n\t"

  2206.         "pswapd    (%1), %%mm1 \n\t"

  2207.         "pfmul  (%3,%0), %%mm0 \n\t"

  2208.         "pfmul 8(%3,%0), %%mm1 \n\t"

  2209.         "movq  %%mm0,  (%2,%0) \n\t"

  2210.         "movq  %%mm1, 8(%2,%0) \n\t"

  2211.         "add   $16, %1 \n\t"

  2212.         "sub   $16, %0 \n\t"

  2213.         "jge   1b \n\t"

  2214.         :"+r"(i), "+r"(src1)

  2215.         :"r"(dst), "r"(src0)

  2216.     );

  2217.     __asm__ volatile("femms");

  2218. }

  2219. static void vector_fmul_reverse_sse(float *dst, const float *src0, const float *src1, int len){

  2220.     x86_reg i = len*4-32;

  2221.     __asm__ volatile(

  2222.         "1: \n\t"

  2223.         "movaps        16(%1), %%xmm0 \n\t"

  2224.         "movaps          (%1), %%xmm1 \n\t"

  2225.         "shufps $0x1b, %%xmm0, %%xmm0 \n\t"

  2226.         "shufps $0x1b, %%xmm1, %%xmm1 \n\t"

  2227.         "mulps        (%3,%0), %%xmm0 \n\t"

  2228.         "mulps      16(%3,%0), %%xmm1 \n\t"

  2229.         "movaps     %%xmm0,   (%2,%0) \n\t"

  2230.         "movaps     %%xmm1, 16(%2,%0) \n\t"

  2231.         "add    $32, %1 \n\t"

  2232.         "sub    $32, %0 \n\t"

  2233.         "jge    1b \n\t"

  2234.         :"+r"(i), "+r"(src1)

  2235.         :"r"(dst), "r"(src0)

  2236.     );

  2237. }

  2238. 

  2239. static void vector_fmul_add_3dnow(float *dst, const float *src0, const float *src1,

  2240.                                   const float *src2, int len){

  2241.     x86_reg i = (len-4)*4;

  2242.     __asm__ volatile(

  2243.         "1: \n\t"

  2244.         "movq    (%2,%0), %%mm0 \n\t"

  2245.         "movq   8(%2,%0), %%mm1 \n\t"

  2246.         "pfmul   (%3,%0), %%mm0 \n\t"

  2247.         "pfmul  8(%3,%0), %%mm1 \n\t"

  2248.         "pfadd   (%4,%0), %%mm0 \n\t"

  2249.         "pfadd  8(%4,%0), %%mm1 \n\t"

  2250.         "movq  %%mm0,   (%1,%0) \n\t"

  2251.         "movq  %%mm1,  8(%1,%0) \n\t"

  2252.         "sub  $16, %0 \n\t"

  2253.         "jge  1b \n\t"

  2254.         :"+r"(i)

  2255.         :"r"(dst), "r"(src0), "r"(src1), "r"(src2)

  2256.         :"memory"

  2257.     );

  2258.     __asm__ volatile("femms");

  2259. }

  2260. static void vector_fmul_add_sse(float *dst, const float *src0, const float *src1,

  2261.                                 const float *src2, int len){

  2262.     x86_reg i = (len-8)*4;

  2263.     __asm__ volatile(

  2264.         "1: \n\t"

  2265.         "movaps   (%2,%0), %%xmm0 \n\t"

  2266.         "movaps 16(%2,%0), %%xmm1 \n\t"

  2267.         "mulps    (%3,%0), %%xmm0 \n\t"

  2268.         "mulps  16(%3,%0), %%xmm1 \n\t"

  2269.         "addps    (%4,%0), %%xmm0 \n\t"

  2270.         "addps  16(%4,%0), %%xmm1 \n\t"

  2271.         "movaps %%xmm0,   (%1,%0) \n\t"

  2272.         "movaps %%xmm1, 16(%1,%0) \n\t"

  2273.         "sub  $32, %0 \n\t"

  2274.         "jge  1b \n\t"

  2275.         :"+r"(i)

  2276.         :"r"(dst), "r"(src0), "r"(src1), "r"(src2)

  2277.         :"memory"

  2278.     );

  2279. }

  2280. 

  2281. #if HAVE_6REGS

  2282. static void vector_fmul_window_3dnow2(float *dst, const float *src0, const float *src1,

  2283.                                       const float *win, int len){

  2284.     x86_reg i = -len*4;

  2285.     x86_reg j = len*4-8;

  2286.     __asm__ volatile(

  2287.         "1: \n"

  2288.         "pswapd  (%5,%1), %%mm1 \n"

  2289.         "movq    (%5,%0), %%mm0 \n"

  2290.         "pswapd  (%4,%1), %%mm5 \n"

  2291.         "movq    (%3,%0), %%mm4 \n"

  2292.         "movq      %%mm0, %%mm2 \n"

  2293.         "movq      %%mm1, %%mm3 \n"

  2294.         "pfmul     %%mm4, %%mm2 \n" // src0[len+i]*win[len+i]

  2295.         "pfmul     %%mm5, %%mm3 \n" // src1[    j]*win[len+j]

  2296.         "pfmul     %%mm4, %%mm1 \n" // src0[len+i]*win[len+j]

  2297.         "pfmul     %%mm5, %%mm0 \n" // src1[    j]*win[len+i]

  2298.         "pfadd     %%mm3, %%mm2 \n"

  2299.         "pfsub     %%mm0, %%mm1 \n"

  2300.         "pswapd    %%mm2, %%mm2 \n"

  2301.         "movq      %%mm1, (%2,%0) \n"

  2302.         "movq      %%mm2, (%2,%1) \n"

  2303.         "sub $8, %1 \n"

  2304.         "add $8, %0 \n"

  2305.         "jl 1b \n"

  2306.         "femms \n"

  2307.         :"+r"(i), "+r"(j)

  2308.         :"r"(dst+len), "r"(src0+len), "r"(src1), "r"(win+len)

  2309.     );

  2310. }

  2311. 

  2312. static void vector_fmul_window_sse(float *dst, const float *src0, const float *src1,

  2313.                                    const float *win, int len){

  2314.     x86_reg i = -len*4;

  2315.     x86_reg j = len*4-16;

  2316.     __asm__ volatile(

  2317.         "1: \n"

  2318.         "movaps       (%5,%1), %%xmm1 \n"

  2319.         "movaps       (%5,%0), %%xmm0 \n"

  2320.         "movaps       (%4,%1), %%xmm5 \n"

  2321.         "movaps       (%3,%0), %%xmm4 \n"

  2322.         "shufps $0x1b, %%xmm1, %%xmm1 \n"

  2323.         "shufps $0x1b, %%xmm5, %%xmm5 \n"

  2324.         "movaps        %%xmm0, %%xmm2 \n"

  2325.         "movaps        %%xmm1, %%xmm3 \n"

  2326.         "mulps         %%xmm4, %%xmm2 \n" // src0[len+i]*win[len+i]

  2327.         "mulps         %%xmm5, %%xmm3 \n" // src1[    j]*win[len+j]

  2328.         "mulps         %%xmm4, %%xmm1 \n" // src0[len+i]*win[len+j]

  2329.         "mulps         %%xmm5, %%xmm0 \n" // src1[    j]*win[len+i]

  2330.         "addps         %%xmm3, %%xmm2 \n"

  2331.         "subps         %%xmm0, %%xmm1 \n"

  2332.         "shufps $0x1b, %%xmm2, %%xmm2 \n"

  2333.         "movaps        %%xmm1, (%2,%0) \n"

  2334.         "movaps        %%xmm2, (%2,%1) \n"

  2335.         "sub $16, %1 \n"

  2336.         "add $16, %0 \n"

  2337.         "jl 1b \n"

  2338.         :"+r"(i), "+r"(j)

  2339.         :"r"(dst+len), "r"(src0+len), "r"(src1), "r"(win+len)

  2340.     );

  2341. }

  2342. #endif /* HAVE_6REGS */

  2343. 

  2344. static void vector_clipf_sse(float *dst, const float *src, float min, float max,

  2345.                              int len)

  2346. {

  2347.     x86_reg i = (len-16)*4;

  2348.     __asm__ volatile(

  2349.         "movss  %3, %%xmm4 \n"

  2350.         "movss  %4, %%xmm5 \n"

  2351.         "shufps $0, %%xmm4, %%xmm4 \n"

  2352.         "shufps $0, %%xmm5, %%xmm5 \n"

  2353.         "1: \n\t"

  2354.         "movaps    (%2,%0), %%xmm0 \n\t" // 3/1 on intel

  2355.         "movaps  16(%2,%0), %%xmm1 \n\t"

  2356.         "movaps  32(%2,%0), %%xmm2 \n\t"

  2357.         "movaps  48(%2,%0), %%xmm3 \n\t"

  2358.         "maxps      %%xmm4, %%xmm0 \n\t"

  2359.         "maxps      %%xmm4, %%xmm1 \n\t"

  2360.         "maxps      %%xmm4, %%xmm2 \n\t"

  2361.         "maxps      %%xmm4, %%xmm3 \n\t"

  2362.         "minps      %%xmm5, %%xmm0 \n\t"

  2363.         "minps      %%xmm5, %%xmm1 \n\t"

  2364.         "minps      %%xmm5, %%xmm2 \n\t"

  2365.         "minps      %%xmm5, %%xmm3 \n\t"

  2366.         "movaps  %%xmm0,   (%1,%0) \n\t"

  2367.         "movaps  %%xmm1, 16(%1,%0) \n\t"

  2368.         "movaps  %%xmm2, 32(%1,%0) \n\t"

  2369.         "movaps  %%xmm3, 48(%1,%0) \n\t"

  2370.         "sub  $64, %0 \n\t"

  2371.         "jge 1b \n\t"

  2372.         :"+&r"(i)

  2373.         :"r"(dst), "r"(src), "m"(min), "m"(max)

  2374.         :"memory"

  2375.     );

  2376. }

  2377. 

  2378. void ff_vp3_idct_mmx(int16_t *input_data);

  2379. void ff_vp3_idct_put_mmx(uint8_t *dest, int line_size, DCTELEM *block);

  2380. void ff_vp3_idct_add_mmx(uint8_t *dest, int line_size, DCTELEM *block);

  2381. 

  2382. void ff_vp3_idct_dc_add_mmx2(uint8_t *dest, int line_size, const DCTELEM *block);

  2383. 

  2384. void ff_vp3_v_loop_filter_mmx2(uint8_t *src, int stride, int *bounding_values);

  2385. void ff_vp3_h_loop_filter_mmx2(uint8_t *src, int stride, int *bounding_values);

  2386. 

  2387. void ff_vp3_idct_sse2(int16_t *input_data);

  2388. void ff_vp3_idct_put_sse2(uint8_t *dest, int line_size, DCTELEM *block);

  2389. void ff_vp3_idct_add_sse2(uint8_t *dest, int line_size, DCTELEM *block);

  2390. 

  2391. int32_t ff_scalarproduct_int16_mmx2(const int16_t *v1, const int16_t *v2, int order, int shift);

  2392. int32_t ff_scalarproduct_int16_sse2(const int16_t *v1, const int16_t *v2, int order, int shift);

  2393. int32_t ff_scalarproduct_and_madd_int16_mmx2(int16_t *v1, const int16_t *v2, const int16_t *v3, int order, int mul);

  2394. int32_t ff_scalarproduct_and_madd_int16_sse2(int16_t *v1, const int16_t *v2, const int16_t *v3, int order, int mul);

  2395. int32_t ff_scalarproduct_and_madd_int16_ssse3(int16_t *v1, const int16_t *v2, const int16_t *v3, int order, int mul);

  2396. 

  2397. void ff_apply_window_int16_mmxext    (int16_t *output, const int16_t *input,

  2398.                                       const int16_t *window, unsigned int len);

  2399. void ff_apply_window_int16_mmxext_ba (int16_t *output, const int16_t *input,

  2400.                                       const int16_t *window, unsigned int len);

  2401. void ff_apply_window_int16_sse2      (int16_t *output, const int16_t *input,

  2402.                                       const int16_t *window, unsigned int len);

  2403. void ff_apply_window_int16_sse2_ba   (int16_t *output, const int16_t *input,

  2404.                                       const int16_t *window, unsigned int len);

  2405. void ff_apply_window_int16_ssse3     (int16_t *output, const int16_t *input,

  2406.                                       const int16_t *window, unsigned int len);

  2407. void ff_apply_window_int16_ssse3_atom(int16_t *output, const int16_t *input,

  2408.                                       const int16_t *window, unsigned int len);

  2409. 

  2410. void ff_add_hfyu_median_prediction_mmx2(uint8_t *dst, const uint8_t *top, const uint8_t *diff, int w, int *left, int *left_top);

  2411. int  ff_add_hfyu_left_prediction_ssse3(uint8_t *dst, const uint8_t *src, int w, int left);

  2412. int  ff_add_hfyu_left_prediction_sse4(uint8_t *dst, const uint8_t *src, int w, int left);

  2413. 

  2414. float ff_scalarproduct_float_sse(const float *v1, const float *v2, int order);

  2415. 

  2416. void ff_vector_clip_int32_mmx     (int32_t *dst, const int32_t *src, int32_t min,

  2417.                                    int32_t max, unsigned int len);

  2418. void ff_vector_clip_int32_sse2    (int32_t *dst, const int32_t *src, int32_t min,

  2419.                                    int32_t max, unsigned int len);

  2420. void ff_vector_clip_int32_sse2_int(int32_t *dst, const int32_t *src, int32_t min,

  2421.                                    int32_t max, unsigned int len);

  2422. void ff_vector_clip_int32_sse41   (int32_t *dst, const int32_t *src, int32_t min,

  2423.                                    int32_t max, unsigned int len);

  2424. 

  2425. void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx)

  2426. {

  2427.     int mm_flags = av_get_cpu_flags();

  2428.     const int high_bit_depth = avctx->bits_per_raw_sample > 8;

  2429.     const int bit_depth = avctx->bits_per_raw_sample;

  2430. 

  2431.     if (avctx->dsp_mask) {

  2432.         if (avctx->dsp_mask & AV_CPU_FLAG_FORCE)

  2433.             mm_flags |= (avctx->dsp_mask & 0xffff);

  2434.         else

  2435.             mm_flags &= ~(avctx->dsp_mask & 0xffff);

  2436.     }

  2437. 

  2438. #if 0

  2439.     av_log(avctx, AV_LOG_INFO, "libavcodec: CPU flags:");

  2440.     if (mm_flags & AV_CPU_FLAG_MMX)

  2441.         av_log(avctx, AV_LOG_INFO, " mmx");

  2442.     if (mm_flags & AV_CPU_FLAG_MMX2)

  2443.         av_log(avctx, AV_LOG_INFO, " mmx2");

  2444.     if (mm_flags & AV_CPU_FLAG_3DNOW)

  2445.         av_log(avctx, AV_LOG_INFO, " 3dnow");

  2446.     if (mm_flags & AV_CPU_FLAG_SSE)

  2447.         av_log(avctx, AV_LOG_INFO, " sse");

  2448.     if (mm_flags & AV_CPU_FLAG_SSE2)

  2449.         av_log(avctx, AV_LOG_INFO, " sse2");

  2450.     av_log(avctx, AV_LOG_INFO, "\n");

  2451. #endif

  2452. 

  2453.     if (mm_flags & AV_CPU_FLAG_MMX) {

  2454.         const int idct_algo= avctx->idct_algo;

  2455. 

  2456.         if (avctx->lowres == 0 && avctx->bits_per_raw_sample <= 8) {

  2457.             if(idct_algo==FF_IDCT_AUTO || idct_algo==FF_IDCT_SIMPLEMMX){

  2458.                 c->idct_put= ff_simple_idct_put_mmx;

  2459.                 c->idct_add= ff_simple_idct_add_mmx;

  2460.                 c->idct    = ff_simple_idct_mmx;

  2461.                 c->idct_permutation_type= FF_SIMPLE_IDCT_PERM;

  2462. #if CONFIG_GPL

  2463.             }else if(idct_algo==FF_IDCT_LIBMPEG2MMX){

  2464.                 if(mm_flags & AV_CPU_FLAG_MMX2){

  2465.                     c->idct_put= ff_libmpeg2mmx2_idct_put;

  2466.                     c->idct_add= ff_libmpeg2mmx2_idct_add;

  2467.                     c->idct    = ff_mmxext_idct;

  2468.                 }else{

  2469.                     c->idct_put= ff_libmpeg2mmx_idct_put;

  2470.                     c->idct_add= ff_libmpeg2mmx_idct_add;

  2471.                     c->idct    = ff_mmx_idct;

  2472.                 }

  2473.                 c->idct_permutation_type= FF_LIBMPEG2_IDCT_PERM;

  2474. #endif

  2475.             }else if((CONFIG_VP3_DECODER || CONFIG_VP5_DECODER || CONFIG_VP6_DECODER) &&

  2476.                      idct_algo==FF_IDCT_VP3 && HAVE_YASM){

  2477.                 if(mm_flags & AV_CPU_FLAG_SSE2){

  2478.                     c->idct_put= ff_vp3_idct_put_sse2;

  2479.                     c->idct_add= ff_vp3_idct_add_sse2;

  2480.                     c->idct    = ff_vp3_idct_sse2;

  2481.                     c->idct_permutation_type= FF_TRANSPOSE_IDCT_PERM;

  2482.                 }else{

  2483.                     c->idct_put= ff_vp3_idct_put_mmx;

  2484.                     c->idct_add= ff_vp3_idct_add_mmx;

  2485.                     c->idct    = ff_vp3_idct_mmx;

  2486.                     c->idct_permutation_type= FF_PARTTRANS_IDCT_PERM;

  2487.                 }

  2488.             }else if(idct_algo==FF_IDCT_CAVS){

  2489.                     c->idct_permutation_type= FF_TRANSPOSE_IDCT_PERM;

  2490.             }else if(idct_algo==FF_IDCT_XVIDMMX){

  2491.                 if(mm_flags & AV_CPU_FLAG_SSE2){

  2492.                     c->idct_put= ff_idct_xvid_sse2_put;

  2493.                     c->idct_add= ff_idct_xvid_sse2_add;

  2494.                     c->idct    = ff_idct_xvid_sse2;

  2495.                     c->idct_permutation_type= FF_SSE2_IDCT_PERM;

  2496.                 }else if(mm_flags & AV_CPU_FLAG_MMX2){

  2497.                     c->idct_put= ff_idct_xvid_mmx2_put;

  2498.                     c->idct_add= ff_idct_xvid_mmx2_add;

  2499.                     c->idct    = ff_idct_xvid_mmx2;

  2500.                 }else{

  2501.                     c->idct_put= ff_idct_xvid_mmx_put;

  2502.                     c->idct_add= ff_idct_xvid_mmx_add;

  2503.                     c->idct    = ff_idct_xvid_mmx;

  2504.                 }

  2505.             }

  2506.         }

  2507. 

  2508.         c->put_pixels_clamped = ff_put_pixels_clamped_mmx;

  2509.         c->put_signed_pixels_clamped = ff_put_signed_pixels_clamped_mmx;

  2510.         c->add_pixels_clamped = ff_add_pixels_clamped_mmx;

  2511.         if (!high_bit_depth) {

  2512.         c->clear_block  = clear_block_mmx;

  2513.         c->clear_blocks = clear_blocks_mmx;

  2514.         if ((mm_flags & AV_CPU_FLAG_SSE) &&

  2515.             !(CONFIG_MPEG_XVMC_DECODER && avctx->xvmc_acceleration > 1)){

  2516.             /* XvMCCreateBlocks() may not allocate 16-byte aligned blocks */

  2517.             c->clear_block  = clear_block_sse;

  2518.             c->clear_blocks = clear_blocks_sse;

  2519.         }

  2520.         }

  2521. 

  2522. #define SET_HPEL_FUNCS(PFX, IDX, SIZE, CPU) \

  2523.         c->PFX ## _pixels_tab[IDX][0] = PFX ## _pixels ## SIZE ## _ ## CPU; \

  2524.         c->PFX ## _pixels_tab[IDX][1] = PFX ## _pixels ## SIZE ## _x2_ ## CPU; \

  2525.         c->PFX ## _pixels_tab[IDX][2] = PFX ## _pixels ## SIZE ## _y2_ ## CPU; \

  2526.         c->PFX ## _pixels_tab[IDX][3] = PFX ## _pixels ## SIZE ## _xy2_ ## CPU

  2527. 

  2528.         if (!high_bit_depth) {

  2529.         SET_HPEL_FUNCS(put, 0, 16, mmx);

  2530.         SET_HPEL_FUNCS(put_no_rnd, 0, 16, mmx);

  2531.         SET_HPEL_FUNCS(avg, 0, 16, mmx);

  2532.         SET_HPEL_FUNCS(avg_no_rnd, 0, 16, mmx);

  2533.         SET_HPEL_FUNCS(put, 1, 8, mmx);

  2534.         SET_HPEL_FUNCS(put_no_rnd, 1, 8, mmx);

  2535.         SET_HPEL_FUNCS(avg, 1, 8, mmx);

  2536.         SET_HPEL_FUNCS(avg_no_rnd, 1, 8, mmx);

  2537.         }

  2538. 

  2539. #if ARCH_X86_32 || !HAVE_YASM

  2540.         c->gmc= gmc_mmx;

  2541. #endif

  2542. #if ARCH_X86_32 && HAVE_YASM

  2543.         if (!high_bit_depth)

  2544.         c->emulated_edge_mc = emulated_edge_mc_mmx;

  2545. #endif

  2546. 

  2547.         c->add_bytes= add_bytes_mmx;

  2548.         c->add_bytes_l2= add_bytes_l2_mmx;

  2549. 

  2550.         if (!high_bit_depth)

  2551.         c->draw_edges = draw_edges_mmx;

  2552. 

  2553.         if (CONFIG_H263_DECODER || CONFIG_H263_ENCODER) {

  2554.             c->h263_v_loop_filter= h263_v_loop_filter_mmx;

  2555.             c->h263_h_loop_filter= h263_h_loop_filter_mmx;

  2556.         }

  2557. 

  2558. #if HAVE_YASM

  2559.         if (!high_bit_depth) {

  2560.         c->put_h264_chroma_pixels_tab[0]= ff_put_h264_chroma_mc8_mmx_rnd;

  2561.         c->put_h264_chroma_pixels_tab[1]= ff_put_h264_chroma_mc4_mmx;

  2562.         }

  2563. 

  2564.         c->vector_clip_int32 = ff_vector_clip_int32_mmx;

  2565. #endif

  2566. 

  2567.         if (mm_flags & AV_CPU_FLAG_MMX2) {

  2568.             c->prefetch = prefetch_mmx2;

  2569. 

  2570.             if (!high_bit_depth) {

  2571.             c->put_pixels_tab[0][1] = put_pixels16_x2_mmx2;

  2572.             c->put_pixels_tab[0][2] = put_pixels16_y2_mmx2;

  2573. 

  2574.             c->avg_pixels_tab[0][0] = avg_pixels16_mmx2;

  2575.             c->avg_pixels_tab[0][1] = avg_pixels16_x2_mmx2;

  2576.             c->avg_pixels_tab[0][2] = avg_pixels16_y2_mmx2;

  2577. 

  2578.             c->put_pixels_tab[1][1] = put_pixels8_x2_mmx2;

  2579.             c->put_pixels_tab[1][2] = put_pixels8_y2_mmx2;

  2580. 

  2581.             c->avg_pixels_tab[1][0] = avg_pixels8_mmx2;

  2582.             c->avg_pixels_tab[1][1] = avg_pixels8_x2_mmx2;

  2583.             c->avg_pixels_tab[1][2] = avg_pixels8_y2_mmx2;

  2584.             }

  2585. 

  2586.             if(!(avctx->flags & CODEC_FLAG_BITEXACT)){

  2587.                 if (!high_bit_depth) {

  2588.                 c->put_no_rnd_pixels_tab[0][1] = put_no_rnd_pixels16_x2_mmx2;

  2589.                 c->put_no_rnd_pixels_tab[0][2] = put_no_rnd_pixels16_y2_mmx2;

  2590.                 c->put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_mmx2;

  2591.                 c->put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_mmx2;

  2592.                 c->avg_pixels_tab[0][3] = avg_pixels16_xy2_mmx2;

  2593.                 c->avg_pixels_tab[1][3] = avg_pixels8_xy2_mmx2;

  2594.                 }

  2595. 

  2596.                 if (CONFIG_VP3_DECODER && HAVE_YASM) {

  2597.                     c->vp3_v_loop_filter= ff_vp3_v_loop_filter_mmx2;

  2598.                     c->vp3_h_loop_filter= ff_vp3_h_loop_filter_mmx2;

  2599.                 }

  2600.             }

  2601.             if (CONFIG_VP3_DECODER && HAVE_YASM) {

  2602.                 c->vp3_idct_dc_add = ff_vp3_idct_dc_add_mmx2;

  2603.             }

  2604. 

  2605.             if (CONFIG_VP3_DECODER

  2606.                 && (avctx->codec_id == CODEC_ID_VP3 || avctx->codec_id == CODEC_ID_THEORA)) {

  2607.                 c->put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_exact_mmx2;

  2608.                 c->put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_exact_mmx2;

  2609.             }

  2610. 

  2611. #define SET_QPEL_FUNCS(PFX, IDX, SIZE, CPU, PREFIX) \

  2612.             c->PFX ## _pixels_tab[IDX][ 0] = PREFIX ## PFX ## SIZE ## _mc00_ ## CPU; \

  2613.             c->PFX ## _pixels_tab[IDX][ 1] = PREFIX ## PFX ## SIZE ## _mc10_ ## CPU; \

  2614.             c->PFX ## _pixels_tab[IDX][ 2] = PREFIX ## PFX ## SIZE ## _mc20_ ## CPU; \

  2615.             c->PFX ## _pixels_tab[IDX][ 3] = PREFIX ## PFX ## SIZE ## _mc30_ ## CPU; \

  2616.             c->PFX ## _pixels_tab[IDX][ 4] = PREFIX ## PFX ## SIZE ## _mc01_ ## CPU; \

  2617.             c->PFX ## _pixels_tab[IDX][ 5] = PREFIX ## PFX ## SIZE ## _mc11_ ## CPU; \

  2618.             c->PFX ## _pixels_tab[IDX][ 6] = PREFIX ## PFX ## SIZE ## _mc21_ ## CPU; \

  2619.             c->PFX ## _pixels_tab[IDX][ 7] = PREFIX ## PFX ## SIZE ## _mc31_ ## CPU; \

  2620.             c->PFX ## _pixels_tab[IDX][ 8] = PREFIX ## PFX ## SIZE ## _mc02_ ## CPU; \

  2621.             c->PFX ## _pixels_tab[IDX][ 9] = PREFIX ## PFX ## SIZE ## _mc12_ ## CPU; \

  2622.             c->PFX ## _pixels_tab[IDX][10] = PREFIX ## PFX ## SIZE ## _mc22_ ## CPU; \

  2623.             c->PFX ## _pixels_tab[IDX][11] = PREFIX ## PFX ## SIZE ## _mc32_ ## CPU; \

  2624.             c->PFX ## _pixels_tab[IDX][12] = PREFIX ## PFX ## SIZE ## _mc03_ ## CPU; \

  2625.             c->PFX ## _pixels_tab[IDX][13] = PREFIX ## PFX ## SIZE ## _mc13_ ## CPU; \

  2626.             c->PFX ## _pixels_tab[IDX][14] = PREFIX ## PFX ## SIZE ## _mc23_ ## CPU; \

  2627.             c->PFX ## _pixels_tab[IDX][15] = PREFIX ## PFX ## SIZE ## _mc33_ ## CPU

  2628. 

  2629.             SET_QPEL_FUNCS(put_qpel, 0, 16, mmx2, );

  2630.             SET_QPEL_FUNCS(put_qpel, 1, 8, mmx2, );

  2631.             SET_QPEL_FUNCS(put_no_rnd_qpel, 0, 16, mmx2, );

  2632.             SET_QPEL_FUNCS(put_no_rnd_qpel, 1, 8, mmx2, );

  2633.             SET_QPEL_FUNCS(avg_qpel, 0, 16, mmx2, );

  2634.             SET_QPEL_FUNCS(avg_qpel, 1, 8, mmx2, );

  2635. 

  2636.             if (!high_bit_depth) {

  2637.             SET_QPEL_FUNCS(put_h264_qpel, 0, 16, mmx2, );

  2638.             SET_QPEL_FUNCS(put_h264_qpel, 1, 8, mmx2, );

  2639.             SET_QPEL_FUNCS(put_h264_qpel, 2, 4, mmx2, );

  2640.             SET_QPEL_FUNCS(avg_h264_qpel, 0, 16, mmx2, );

  2641.             SET_QPEL_FUNCS(avg_h264_qpel, 1, 8, mmx2, );

  2642.             SET_QPEL_FUNCS(avg_h264_qpel, 2, 4, mmx2, );

  2643.             }

  2644.             else if (bit_depth == 10) {

  2645. #if HAVE_YASM

  2646. #if !ARCH_X86_64

  2647.                 SET_QPEL_FUNCS(avg_h264_qpel, 0, 16, 10_mmxext, ff_);

  2648.                 SET_QPEL_FUNCS(put_h264_qpel, 0, 16, 10_mmxext, ff_);

  2649.                 SET_QPEL_FUNCS(put_h264_qpel, 1, 8,  10_mmxext, ff_);

  2650.                 SET_QPEL_FUNCS(avg_h264_qpel, 1, 8,  10_mmxext, ff_);

  2651. #endif

  2652.                 SET_QPEL_FUNCS(put_h264_qpel, 2, 4,  10_mmxext, ff_);

  2653.                 SET_QPEL_FUNCS(avg_h264_qpel, 2, 4,  10_mmxext, ff_);

  2654. #endif

  2655.             }

  2656. 

  2657.             SET_QPEL_FUNCS(put_2tap_qpel, 0, 16, mmx2, );

  2658.             SET_QPEL_FUNCS(put_2tap_qpel, 1, 8, mmx2, );

  2659.             SET_QPEL_FUNCS(avg_2tap_qpel, 0, 16, mmx2, );

  2660.             SET_QPEL_FUNCS(avg_2tap_qpel, 1, 8, mmx2, );

  2661. 

  2662. #if HAVE_YASM

  2663.             if (!high_bit_depth) {

  2664.             c->avg_h264_chroma_pixels_tab[0]= ff_avg_h264_chroma_mc8_mmx2_rnd;

  2665.             c->avg_h264_chroma_pixels_tab[1]= ff_avg_h264_chroma_mc4_mmx2;

  2666.             c->avg_h264_chroma_pixels_tab[2]= ff_avg_h264_chroma_mc2_mmx2;

  2667.             c->put_h264_chroma_pixels_tab[2]= ff_put_h264_chroma_mc2_mmx2;

  2668.             }

  2669.             if (bit_depth == 10) {

  2670.                 c->put_h264_chroma_pixels_tab[2]= ff_put_h264_chroma_mc2_10_mmxext;

  2671.                 c->avg_h264_chroma_pixels_tab[2]= ff_avg_h264_chroma_mc2_10_mmxext;

  2672.                 c->put_h264_chroma_pixels_tab[1]= ff_put_h264_chroma_mc4_10_mmxext;

  2673.                 c->avg_h264_chroma_pixels_tab[1]= ff_avg_h264_chroma_mc4_10_mmxext;

  2674.             }

  2675. 

  2676.             c->add_hfyu_median_prediction = ff_add_hfyu_median_prediction_mmx2;

  2677. #endif

  2678. #if HAVE_7REGS

  2679.             if (HAVE_AMD3DNOW && (mm_flags & AV_CPU_FLAG_3DNOW))

  2680.                 c->add_hfyu_median_prediction = add_hfyu_median_prediction_cmov;

  2681. #endif

  2682. 

  2683.             c->add_png_paeth_prediction= add_png_paeth_prediction_mmx2;

  2684.         } else if (HAVE_AMD3DNOW && (mm_flags & AV_CPU_FLAG_3DNOW)) {

  2685.             c->prefetch = prefetch_3dnow;

  2686. 

  2687.             if (!high_bit_depth) {

  2688.             c->put_pixels_tab[0][1] = put_pixels16_x2_3dnow;

  2689.             c->put_pixels_tab[0][2] = put_pixels16_y2_3dnow;

  2690. 

  2691.             c->avg_pixels_tab[0][0] = avg_pixels16_3dnow;

  2692.             c->avg_pixels_tab[0][1] = avg_pixels16_x2_3dnow;

  2693.             c->avg_pixels_tab[0][2] = avg_pixels16_y2_3dnow;

  2694. 

  2695.             c->put_pixels_tab[1][1] = put_pixels8_x2_3dnow;

  2696.             c->put_pixels_tab[1][2] = put_pixels8_y2_3dnow;

  2697. 

  2698.             c->avg_pixels_tab[1][0] = avg_pixels8_3dnow;

  2699.             c->avg_pixels_tab[1][1] = avg_pixels8_x2_3dnow;

  2700.             c->avg_pixels_tab[1][2] = avg_pixels8_y2_3dnow;

  2701. 

  2702.             if(!(avctx->flags & CODEC_FLAG_BITEXACT)){

  2703.                 c->put_no_rnd_pixels_tab[0][1] = put_no_rnd_pixels16_x2_3dnow;

  2704.                 c->put_no_rnd_pixels_tab[0][2] = put_no_rnd_pixels16_y2_3dnow;

  2705.                 c->put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_3dnow;

  2706.                 c->put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_3dnow;

  2707.                 c->avg_pixels_tab[0][3] = avg_pixels16_xy2_3dnow;

  2708.                 c->avg_pixels_tab[1][3] = avg_pixels8_xy2_3dnow;

  2709.             }

  2710.             }

  2711. 

  2712.             if (CONFIG_VP3_DECODER

  2713.                 && (avctx->codec_id == CODEC_ID_VP3 || avctx->codec_id == CODEC_ID_THEORA)) {

  2714.                 c->put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_exact_3dnow;

  2715.                 c->put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_exact_3dnow;

  2716.             }

  2717. 

  2718.             SET_QPEL_FUNCS(put_qpel, 0, 16, 3dnow, );

  2719.             SET_QPEL_FUNCS(put_qpel, 1, 8, 3dnow, );

  2720.             SET_QPEL_FUNCS(put_no_rnd_qpel, 0, 16, 3dnow, );

  2721.             SET_QPEL_FUNCS(put_no_rnd_qpel, 1, 8, 3dnow, );

  2722.             SET_QPEL_FUNCS(avg_qpel, 0, 16, 3dnow, );

  2723.             SET_QPEL_FUNCS(avg_qpel, 1, 8, 3dnow, );

  2724. 

  2725.             if (!high_bit_depth) {

  2726.             SET_QPEL_FUNCS(put_h264_qpel, 0, 16, 3dnow, );

  2727.             SET_QPEL_FUNCS(put_h264_qpel, 1, 8, 3dnow, );

  2728.             SET_QPEL_FUNCS(put_h264_qpel, 2, 4, 3dnow, );

  2729.             SET_QPEL_FUNCS(avg_h264_qpel, 0, 16, 3dnow, );

  2730.             SET_QPEL_FUNCS(avg_h264_qpel, 1, 8, 3dnow, );

  2731.             SET_QPEL_FUNCS(avg_h264_qpel, 2, 4, 3dnow, );

  2732.             }

  2733. 

  2734.             SET_QPEL_FUNCS(put_2tap_qpel, 0, 16, 3dnow, );

  2735.             SET_QPEL_FUNCS(put_2tap_qpel, 1, 8, 3dnow, );

  2736.             SET_QPEL_FUNCS(avg_2tap_qpel, 0, 16, 3dnow, );

  2737.             SET_QPEL_FUNCS(avg_2tap_qpel, 1, 8, 3dnow, );

  2738. 

  2739. #if HAVE_YASM

  2740.             if (!high_bit_depth) {

  2741.             c->avg_h264_chroma_pixels_tab[0]= ff_avg_h264_chroma_mc8_3dnow_rnd;

  2742.             c->avg_h264_chroma_pixels_tab[1]= ff_avg_h264_chroma_mc4_3dnow;

  2743.             }

  2744. 

  2745. #endif

  2746.         }

  2747. 

  2748. 

  2749. #define H264_QPEL_FUNCS(x, y, CPU)\

  2750.             c->put_h264_qpel_pixels_tab[0][x+y*4] = put_h264_qpel16_mc##x##y##_##CPU;\

  2751.             c->put_h264_qpel_pixels_tab[1][x+y*4] = put_h264_qpel8_mc##x##y##_##CPU;\

  2752.             c->avg_h264_qpel_pixels_tab[0][x+y*4] = avg_h264_qpel16_mc##x##y##_##CPU;\

  2753.             c->avg_h264_qpel_pixels_tab[1][x+y*4] = avg_h264_qpel8_mc##x##y##_##CPU;

  2754.         if((mm_flags & AV_CPU_FLAG_SSE2) && !(mm_flags & AV_CPU_FLAG_3DNOW)){

  2755.             // these functions are slower than mmx on AMD, but faster on Intel

  2756.             if (!high_bit_depth) {

  2757.             c->put_pixels_tab[0][0] = put_pixels16_sse2;

  2758.             c->put_no_rnd_pixels_tab[0][0] = put_pixels16_sse2;

  2759.             c->avg_pixels_tab[0][0] = avg_pixels16_sse2;

  2760.             H264_QPEL_FUNCS(0, 0, sse2);

  2761.             }

  2762.         }

  2763.         if(mm_flags & AV_CPU_FLAG_SSE2){

  2764.             if (!high_bit_depth) {

  2765.             H264_QPEL_FUNCS(0, 1, sse2);

  2766.             H264_QPEL_FUNCS(0, 2, sse2);

  2767.             H264_QPEL_FUNCS(0, 3, sse2);

  2768.             H264_QPEL_FUNCS(1, 1, sse2);

  2769.             H264_QPEL_FUNCS(1, 2, sse2);

  2770.             H264_QPEL_FUNCS(1, 3, sse2);

  2771.             H264_QPEL_FUNCS(2, 1, sse2);

  2772.             H264_QPEL_FUNCS(2, 2, sse2);

  2773.             H264_QPEL_FUNCS(2, 3, sse2);

  2774.             H264_QPEL_FUNCS(3, 1, sse2);

  2775.             H264_QPEL_FUNCS(3, 2, sse2);

  2776.             H264_QPEL_FUNCS(3, 3, sse2);

  2777.             }

  2778. #if HAVE_YASM

  2779. #define H264_QPEL_FUNCS_10(x, y, CPU)\

  2780.             c->put_h264_qpel_pixels_tab[0][x+y*4] = ff_put_h264_qpel16_mc##x##y##_10_##CPU;\

  2781.             c->put_h264_qpel_pixels_tab[1][x+y*4] = ff_put_h264_qpel8_mc##x##y##_10_##CPU;\

  2782.             c->avg_h264_qpel_pixels_tab[0][x+y*4] = ff_avg_h264_qpel16_mc##x##y##_10_##CPU;\

  2783.             c->avg_h264_qpel_pixels_tab[1][x+y*4] = ff_avg_h264_qpel8_mc##x##y##_10_##CPU;

  2784.             if (bit_depth == 10) {

  2785.                 SET_QPEL_FUNCS(put_h264_qpel, 0, 16, 10_sse2, ff_);

  2786.                 SET_QPEL_FUNCS(put_h264_qpel, 1, 8,  10_sse2, ff_);

  2787.                 SET_QPEL_FUNCS(avg_h264_qpel, 0, 16, 10_sse2, ff_);

  2788.                 SET_QPEL_FUNCS(avg_h264_qpel, 1, 8,  10_sse2, ff_);

  2789.                 H264_QPEL_FUNCS_10(1, 0, sse2_cache64)

  2790.                 H264_QPEL_FUNCS_10(2, 0, sse2_cache64)

  2791.                 H264_QPEL_FUNCS_10(3, 0, sse2_cache64)

  2792. 

  2793.                 c->put_h264_chroma_pixels_tab[0]= ff_put_h264_chroma_mc8_10_sse2;

  2794.                 c->avg_h264_chroma_pixels_tab[0]= ff_avg_h264_chroma_mc8_10_sse2;

  2795.             }

  2796. #endif

  2797.         }

  2798. #if HAVE_SSSE3

  2799.         if(mm_flags & AV_CPU_FLAG_SSSE3){

  2800.             if (!high_bit_depth) {

  2801.             H264_QPEL_FUNCS(1, 0, ssse3);

  2802.             H264_QPEL_FUNCS(1, 1, ssse3);

  2803.             H264_QPEL_FUNCS(1, 2, ssse3);

  2804.             H264_QPEL_FUNCS(1, 3, ssse3);

  2805.             H264_QPEL_FUNCS(2, 0, ssse3);

  2806.             H264_QPEL_FUNCS(2, 1, ssse3);

  2807.             H264_QPEL_FUNCS(2, 2, ssse3);

  2808.             H264_QPEL_FUNCS(2, 3, ssse3);

  2809.             H264_QPEL_FUNCS(3, 0, ssse3);

  2810.             H264_QPEL_FUNCS(3, 1, ssse3);

  2811.             H264_QPEL_FUNCS(3, 2, ssse3);

  2812.             H264_QPEL_FUNCS(3, 3, ssse3);

  2813.             }

  2814. #if HAVE_YASM

  2815.             else if (bit_depth == 10) {

  2816.                 H264_QPEL_FUNCS_10(1, 0, ssse3_cache64)

  2817.                 H264_QPEL_FUNCS_10(2, 0, ssse3_cache64)

  2818.                 H264_QPEL_FUNCS_10(3, 0, ssse3_cache64)

  2819.             }

  2820. #endif

  2821.             c->add_png_paeth_prediction= add_png_paeth_prediction_ssse3;

  2822. #if HAVE_YASM

  2823.             if (!high_bit_depth) {

  2824.             c->put_h264_chroma_pixels_tab[0]= ff_put_h264_chroma_mc8_ssse3_rnd;

  2825.             c->avg_h264_chroma_pixels_tab[0]= ff_avg_h264_chroma_mc8_ssse3_rnd;

  2826.             c->put_h264_chroma_pixels_tab[1]= ff_put_h264_chroma_mc4_ssse3;

  2827.             c->avg_h264_chroma_pixels_tab[1]= ff_avg_h264_chroma_mc4_ssse3;

  2828.             }

  2829.             c->add_hfyu_left_prediction = ff_add_hfyu_left_prediction_ssse3;

  2830.             if (mm_flags & AV_CPU_FLAG_SSE4) // not really sse4, just slow on Conroe

  2831.                 c->add_hfyu_left_prediction = ff_add_hfyu_left_prediction_sse4;

  2832. #endif

  2833.         }

  2834. #endif

  2835. 

  2836.         if (HAVE_AMD3DNOW && (mm_flags & AV_CPU_FLAG_3DNOW)) {

  2837.             c->vorbis_inverse_coupling = vorbis_inverse_coupling_3dnow;

  2838.             c->vector_fmul = vector_fmul_3dnow;

  2839.         }

  2840.         if (HAVE_AMD3DNOWEXT && (mm_flags & AV_CPU_FLAG_3DNOWEXT)) {

  2841.             c->vector_fmul_reverse = vector_fmul_reverse_3dnow2;

  2842. #if HAVE_6REGS

  2843.             c->vector_fmul_window = vector_fmul_window_3dnow2;

  2844. #endif

  2845.         }

  2846.         if(mm_flags & AV_CPU_FLAG_MMX2){

  2847. #if HAVE_YASM

  2848.             c->scalarproduct_int16 = ff_scalarproduct_int16_mmx2;

  2849.             c->scalarproduct_and_madd_int16 = ff_scalarproduct_and_madd_int16_mmx2;

  2850.             if (avctx->flags & CODEC_FLAG_BITEXACT) {

  2851.                 c->apply_window_int16 = ff_apply_window_int16_mmxext_ba;

  2852.             } else {

  2853.                 c->apply_window_int16 = ff_apply_window_int16_mmxext;

  2854.             }

  2855. #endif

  2856.         }

  2857.         if(mm_flags & AV_CPU_FLAG_SSE){

  2858.             c->vorbis_inverse_coupling = vorbis_inverse_coupling_sse;

  2859.             c->ac3_downmix = ac3_downmix_sse;

  2860.             c->vector_fmul = vector_fmul_sse;

  2861.             c->vector_fmul_reverse = vector_fmul_reverse_sse;

  2862.             c->vector_fmul_add = vector_fmul_add_sse;

  2863. #if HAVE_6REGS

  2864.             c->vector_fmul_window = vector_fmul_window_sse;

  2865. #endif

  2866.             c->vector_clipf = vector_clipf_sse;

  2867. #if HAVE_YASM

  2868.             c->scalarproduct_float = ff_scalarproduct_float_sse;

  2869. #endif

  2870.         }

  2871.         if (HAVE_AMD3DNOW && (mm_flags & AV_CPU_FLAG_3DNOW))

  2872.             c->vector_fmul_add = vector_fmul_add_3dnow; // faster than sse

  2873.         if(mm_flags & AV_CPU_FLAG_SSE2){

  2874. #if HAVE_YASM

  2875.             c->scalarproduct_int16 = ff_scalarproduct_int16_sse2;

  2876.             c->scalarproduct_and_madd_int16 = ff_scalarproduct_and_madd_int16_sse2;

  2877.             if (mm_flags & AV_CPU_FLAG_ATOM) {

  2878.                 c->vector_clip_int32 = ff_vector_clip_int32_sse2_int;

  2879.             } else {

  2880.                 c->vector_clip_int32 = ff_vector_clip_int32_sse2;

  2881.             }

  2882.             if (avctx->flags & CODEC_FLAG_BITEXACT) {

  2883.                 c->apply_window_int16 = ff_apply_window_int16_sse2_ba;

  2884.             } else {

  2885.                 if (!(mm_flags & AV_CPU_FLAG_SSE2SLOW)) {

  2886.                     c->apply_window_int16 = ff_apply_window_int16_sse2;

  2887.                 }

  2888.             }

  2889. 

  2890.             if (!high_bit_depth)

  2891.             c->emulated_edge_mc = emulated_edge_mc_sse;

  2892.             c->gmc= gmc_sse;

  2893. #endif

  2894.         }

  2895.         if (mm_flags & AV_CPU_FLAG_SSSE3) {

  2896. #if HAVE_YASM

  2897.             if (mm_flags & AV_CPU_FLAG_ATOM) {

  2898.                 c->apply_window_int16 = ff_apply_window_int16_ssse3_atom;

  2899.             } else {

  2900.                 c->apply_window_int16 = ff_apply_window_int16_ssse3;

  2901.             }

  2902.             if (!(mm_flags & (AV_CPU_FLAG_SSE42|AV_CPU_FLAG_3DNOW))) { // cachesplit

  2903.                 c->scalarproduct_and_madd_int16 = ff_scalarproduct_and_madd_int16_ssse3;

  2904.             }

  2905. #endif

  2906.         }

  2907. 

  2908.         if (mm_flags & AV_CPU_FLAG_SSE4 && HAVE_SSE) {

  2909. #if HAVE_YASM

  2910.             c->vector_clip_int32 = ff_vector_clip_int32_sse41;

  2911. #endif

  2912.         }

  2913. 

  2914. #if HAVE_AVX && HAVE_YASM

  2915.         if (mm_flags & AV_CPU_FLAG_AVX) {

  2916.             if (bit_depth == 10) {

  2917.                 //AVX implies !cache64.

  2918.                 //TODO: Port cache(32|64) detection from x264.

  2919.                 H264_QPEL_FUNCS_10(1, 0, sse2)

  2920.                 H264_QPEL_FUNCS_10(2, 0, sse2)

  2921.                 H264_QPEL_FUNCS_10(3, 0, sse2)

  2922. 

  2923.                 c->put_h264_chroma_pixels_tab[0]= ff_put_h264_chroma_mc8_10_avx;

  2924.                 c->avg_h264_chroma_pixels_tab[0]= ff_avg_h264_chroma_mc8_10_avx;

  2925.             }

  2926.         }

  2927. #endif

  2928.     }

  2929. 

  2930.     if (CONFIG_ENCODERS)

  2931.         dsputilenc_init_mmx(c, avctx);

  2932. }