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.
28457 Commits
32 Branches
759MB
Tree: 44ca80df34
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from '44ca80df34'
${ noResults }
FFmpeg/libavcodec/x86/dsputil_mmx.c

2864 lines
119KB
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(8,  const uint64_t, ff_pw_1  ) = 0x0001000100010001ULL;

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  66. 

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

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

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

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

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

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

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

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

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

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

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

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

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

  80. 

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

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

  83. 

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

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

  86. 

  87. #define MOVQ_BFE(regd) \

  88.     __asm__ volatile ( \

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

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

  91. 

  92. #ifndef PIC

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

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

  95. #else

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

  97. // pcmpeqd -> -1

  98. #define MOVQ_BONE(regd) \

  99.     __asm__ volatile ( \

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

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

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

  103. 

  104. #define MOVQ_WTWO(regd) \

  105.     __asm__ volatile ( \

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

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

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

  109. 

  110. #endif

  111. 

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

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

  114. // regfe is supposed to contain 0xfefefefefefefefe

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

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

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

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

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

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

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

  122. 

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

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

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

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

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

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

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

  130. 

  131. // mm6 is supposed to contain 0xfefefefefefefefe

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

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

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

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

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

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

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

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

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

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

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

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

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

  145. 

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

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

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

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

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

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

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

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

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

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

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

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

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

  159. 

  160. /***********************************/

  161. /* MMX no rounding */

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

  163. #define SET_RND  MOVQ_WONE

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

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

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

  167. 

  168. #include "dsputil_mmx_rnd_template.c"

  169. 

  170. #undef DEF

  171. #undef SET_RND

  172. #undef PAVGBP

  173. #undef PAVGB

  174. /***********************************/

  175. /* MMX rounding */

  176. 

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

  178. #define SET_RND  MOVQ_WTWO

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

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

  181. 

  182. #include "dsputil_mmx_rnd_template.c"

  183. 

  184. #undef DEF

  185. #undef SET_RND

  186. #undef PAVGBP

  187. #undef PAVGB

  188. #undef OP_AVG

  189. 

  190. /***********************************/

  191. /* 3Dnow specific */

  192. 

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

  194. #define PAVGB "pavgusb"

  195. #define OP_AVG PAVGB

  196. 

  197. #include "dsputil_mmx_avg_template.c"

  198. 

  199. #undef DEF

  200. #undef PAVGB

  201. #undef OP_AVG

  202. 

  203. /***********************************/

  204. /* MMX2 specific */

  205. 

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

  207. 

  208. /* Introduced only in MMX2 set */

  209. #define PAVGB "pavgb"

  210. #define OP_AVG PAVGB

  211. 

  212. #include "dsputil_mmx_avg_template.c"

  213. 

  214. #undef DEF

  215. #undef PAVGB

  216. #undef OP_AVG

  217. 

  218. #define put_no_rnd_pixels16_mmx put_pixels16_mmx

  219. #define put_no_rnd_pixels8_mmx put_pixels8_mmx

  220. #define put_pixels16_mmx2 put_pixels16_mmx

  221. #define put_pixels8_mmx2 put_pixels8_mmx

  222. #define put_pixels4_mmx2 put_pixels4_mmx

  223. #define put_no_rnd_pixels16_mmx2 put_no_rnd_pixels16_mmx

  224. #define put_no_rnd_pixels8_mmx2 put_no_rnd_pixels8_mmx

  225. #define put_pixels16_3dnow put_pixels16_mmx

  226. #define put_pixels8_3dnow put_pixels8_mmx

  227. #define put_pixels4_3dnow put_pixels4_mmx

  228. #define put_no_rnd_pixels16_3dnow put_no_rnd_pixels16_mmx

  229. #define put_no_rnd_pixels8_3dnow put_no_rnd_pixels8_mmx

  230. 

  231. /***********************************/

  232. /* standard MMX */

  233. 

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

  235. {

  236.     const DCTELEM *p;

  237.     uint8_t *pix;

  238. 

  239.     /* read the pixels */

  240.     p = block;

  241.     pix = pixels;

  242.     /* unrolled loop */

  243.         __asm__ volatile(

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  261.                 :"memory");

  262.         pix += line_size*4;

  263.         p += 32;

  264. 

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

  266.     // compiler would generate some very strange code

  267.     // thus using "r"

  268.     __asm__ volatile(

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  286.             :"memory");

  287. }

  288. 

  289. #define put_signed_pixels_clamped_mmx_half(off) \

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  306. 

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

  308. {

  309.     x86_reg line_skip = line_size;

  310.     x86_reg line_skip3;

  311. 

  312.     __asm__ volatile (

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

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

  315.             put_signed_pixels_clamped_mmx_half(0)

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

  317.             put_signed_pixels_clamped_mmx_half(64)

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

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

  320.             :"memory");

  321. }

  322. 

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

  324. {

  325.     const DCTELEM *p;

  326.     uint8_t *pix;

  327.     int i;

  328. 

  329.     /* read the pixels */

  330.     p = block;

  331.     pix = pixels;

  332.     MOVQ_ZERO(mm7);

  333.     i = 4;

  334.     do {

  335.         __asm__ volatile(

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  357.                 :"r"(p)

  358.                 :"memory");

  359.         pix += line_size*2;

  360.         p += 16;

  361.     } while (--i);

  362. }

  363. 

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

  365. {

  366.     __asm__ volatile(

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

  368.          ".p2align 3                    \n\t"

  369.          "1:                            \n\t"

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

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

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

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

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

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

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

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

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

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

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

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

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

  383.          "jnz 1b                        \n\t"

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

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

  386.          : "%"REG_a, "memory"

  387.         );

  388. }

  389. 

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

  391. {

  392.     __asm__ volatile(

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

  394.          ".p2align 3                    \n\t"

  395.          "1:                            \n\t"

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

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

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

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

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

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

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

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

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

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

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

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

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

  409.          "jnz 1b                        \n\t"

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

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

  412.          : "%"REG_a, "memory"

  413.         );

  414. }

  415. 

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

  417. {

  418.     __asm__ volatile(

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

  420.          ".p2align 3                    \n\t"

  421.          "1:                            \n\t"

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  443.          "jnz 1b                        \n\t"

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

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

  446.          : "%"REG_a, "memory"

  447.         );

  448. }

  449. 

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

  451. {

  452.     __asm__ volatile(

  453.          "1:                            \n\t"

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

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

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

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

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

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

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

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

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

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

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

  465.          "jnz 1b                        \n\t"

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

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

  468.          : "memory"

  469.         );

  470. }

  471. 

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

  473. {

  474.     __asm__ volatile(

  475.          "1:                            \n\t"

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  491.          "jnz 1b                        \n\t"

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

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

  494.          : "memory"

  495.         );

  496. }

  497. 

  498. #define CLEAR_BLOCKS(name,n) \

  499. static void name(DCTELEM *blocks)\

  500. {\

  501.     __asm__ volatile(\

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

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

  504.                 "1:                             \n\t"\

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

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

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

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

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

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

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

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

  513.                 : "%"REG_a\

  514.         );\

  515. }

  516. CLEAR_BLOCKS(clear_blocks_mmx, 6)

  517. CLEAR_BLOCKS(clear_block_mmx, 1)

  518. 

  519. static void clear_block_sse(DCTELEM *block)

  520. {

  521.     __asm__ volatile(

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

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

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

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

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

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

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

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

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

  531.         :: "r"(block)

  532.         : "memory"

  533.     );

  534. }

  535. 

  536. static void clear_blocks_sse(DCTELEM *blocks)

  537. {\

  538.     __asm__ volatile(

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

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

  541.         "1:                     \n"

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

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

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

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

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

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

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

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

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

  551.         " js 1b                 \n"

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

  553.             "i" (-128*6)

  554.         : "%"REG_a

  555.     );

  556. }

  557. 

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

  559.     x86_reg i=0;

  560.     __asm__ volatile(

  561.         "jmp 2f                         \n\t"

  562.         "1:                             \n\t"

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

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

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

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

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

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

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

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

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

  572.         "2:                             \n\t"

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

  574.         " js 1b                         \n\t"

  575.         : "+r" (i)

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

  577.     );

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

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

  580. }

  581. 

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

  583.     x86_reg i=0;

  584.     __asm__ volatile(

  585.         "jmp 2f                         \n\t"

  586.         "1:                             \n\t"

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

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

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

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

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

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

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

  594.         "2:                             \n\t"

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

  596.         " js 1b                         \n\t"

  597.         : "+r" (i)

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

  599.     );

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

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

  602. }

  603. 

  604. #if HAVE_7REGS && HAVE_TEN_OPERANDS

  605. 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) {

  606.     x86_reg w2 = -w;

  607.     x86_reg x;

  608.     int l = *left & 0xff;

  609.     int tl = *left_top & 0xff;

  610.     int t;

  611.     __asm__ volatile(

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

  613.         "1: \n"

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  629.         "inc    %4 \n"

  630.         "jl 1b \n"

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

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

  633.     );

  634.     *left = l;

  635.     *left_top = tl;

  636. }

  637. #endif

  638. 

  639. #define H263_LOOP_FILTER \

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  709. 

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

  711.     if(CONFIG_H263_DECODER || CONFIG_H263_ENCODER) {

  712.     const int strength= ff_h263_loop_filter_strength[qscale];

  713. 

  714.     __asm__ volatile(

  715. 

  716.         H263_LOOP_FILTER

  717. 

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

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

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

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

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

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

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

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

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

  727.     );

  728.     }

  729. }

  730. 

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

  732.     if(CONFIG_H263_DECODER || CONFIG_H263_ENCODER) {

  733.     const int strength= ff_h263_loop_filter_strength[qscale];

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

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

  736. 

  737.     src -= 2;

  738. 

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

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

  741.     __asm__ volatile(

  742.         H263_LOOP_FILTER // 5 3 4 6

  743. 

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

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

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

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

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

  749.     );

  750. 

  751.     __asm__ volatile(

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  776.         :: "r" (src),

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

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

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

  780.     );

  781.     }

  782. }

  783. 

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

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

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

  787. {

  788.     uint8_t *ptr, *last_line;

  789.     int i;

  790. 

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

  792.     /* left and right */

  793.     ptr = buf;

  794.     if(w==8)

  795.     {

  796.         __asm__ volatile(

  797.                 "1:                             \n\t"

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

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

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

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

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

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

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

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

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

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

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

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

  810.                 " jb 1b                         \n\t"

  811.                 : "+r" (ptr)

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

  813.         );

  814.     }

  815.     else

  816.     {

  817.         __asm__ volatile(

  818.                 "1:                             \n\t"

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  833.                 " jb 1b                         \n\t"

  834.                 : "+r" (ptr)

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

  836.         );

  837.     }

  838. 

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

  840.     if (sides&EDGE_TOP) {

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

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

  843.             __asm__ volatile(

  844.                     "1:                             \n\t"

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

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

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

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

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

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

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

  852.                     " jb 1b                         \n\t"

  853.                     : "+r" (ptr)

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

  855.             );

  856.         }

  857.     }

  858. 

  859.     if (sides&EDGE_BOTTOM) {

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

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

  862.             __asm__ volatile(

  863.                     "1:                             \n\t"

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

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

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

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

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

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

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

  871.                     " jb 1b                         \n\t"

  872.                     : "+r" (ptr)

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

  874.             );

  875.         }

  876.     }

  877. }

  878. 

  879. #define PAETH(cpu, abs3)\

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

  881. {\

  882.     x86_reg i = -bpp;\

  883.     x86_reg end = w-3;\

  884.     __asm__ volatile(\

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

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

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

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

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

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

  891.         "1: \n"\

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

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

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

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

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

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

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

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

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

  901.         abs3\

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  924.         "jle 1b \n"\

  925.         :"+r"(i)\

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

  927.          "m"(ff_pw_255)\

  928.         :"memory"\

  929.     );\

  930. }

  931. 

  932. #define ABS3_MMX2\

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

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

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

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

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

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

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

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

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

  942. 

  943. #define ABS3_SSSE3\

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

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

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

  947. 

  948. PAETH(mmx2, ABS3_MMX2)

  949. #if HAVE_SSSE3

  950. PAETH(ssse3, ABS3_SSSE3)

  951. #endif

  952. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  973. 

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

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

  976.     uint64_t temp;\

  977. \

  978.     __asm__ volatile(\

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

  980.         "1:                               \n\t"\

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  1000.         "pshufw $0x06, %%mm0, %%mm5       \n\t" /* 0C0B0A0A */\

  1001.         "pmullw "MANGLE(ff_pw_3)", %%mm6  \n\t" /* 3c - 6b */\

  1002.         "paddw %%mm4, %%mm0               \n\t" /* a */\

  1003.         "paddw %%mm1, %%mm5               \n\t" /* d */\

  1004.         "pmullw "MANGLE(ff_pw_20)", %%mm0 \n\t" /* 20a */\

  1005.         "psubw %%mm5, %%mm0               \n\t" /* 20a - d */\

  1006.         "paddw %6, %%mm6                  \n\t"\

  1007.         "paddw %%mm6, %%mm0               \n\t" /* 20a - 6b + 3c - d */\

  1008.         "psraw $5, %%mm0                  \n\t"\

  1009.         "movq %%mm0, %5                   \n\t"\

  1010.         /* mm1=EFGH, mm2=DEFG, mm3=CDEF, mm4=BCDE, mm7=0 */\

  1011.         \

  1012.         "movq 5(%0), %%mm0                \n\t" /* FGHIJKLM */\

  1013.         "movq %%mm0, %%mm5                \n\t" /* FGHIJKLM */\

  1014.         "movq %%mm0, %%mm6                \n\t" /* FGHIJKLM */\

  1015.         "psrlq $8, %%mm0                  \n\t" /* GHIJKLM0 */\

  1016.         "psrlq $16, %%mm5                 \n\t" /* HIJKLM00 */\

  1017.         "punpcklbw %%mm7, %%mm0           \n\t" /* 0G0H0I0J */\

  1018.         "punpcklbw %%mm7, %%mm5           \n\t" /* 0H0I0J0K */\

  1019.         "paddw %%mm0, %%mm2               \n\t" /* b */\

  1020.         "paddw %%mm5, %%mm3               \n\t" /* c */\

  1021.         "paddw %%mm2, %%mm2               \n\t" /* 2b */\

  1022.         "psubw %%mm2, %%mm3               \n\t" /* c - 2b */\

  1023.         "movq %%mm6, %%mm2                \n\t" /* FGHIJKLM */\

  1024.         "psrlq $24, %%mm6                 \n\t" /* IJKLM000 */\

  1025.         "punpcklbw %%mm7, %%mm2           \n\t" /* 0F0G0H0I */\

  1026.         "punpcklbw %%mm7, %%mm6           \n\t" /* 0I0J0K0L */\

  1027.         "pmullw "MANGLE(ff_pw_3)", %%mm3  \n\t" /* 3c - 6b */\

  1028.         "paddw %%mm2, %%mm1               \n\t" /* a */\

  1029.         "paddw %%mm6, %%mm4               \n\t" /* d */\

  1030.         "pmullw "MANGLE(ff_pw_20)", %%mm1 \n\t" /* 20a */\

  1031.         "psubw %%mm4, %%mm3               \n\t" /* - 6b +3c - d */\

  1032.         "paddw %6, %%mm1                  \n\t"\

  1033.         "paddw %%mm1, %%mm3               \n\t" /* 20a - 6b +3c - d */\

  1034.         "psraw $5, %%mm3                  \n\t"\

  1035.         "movq %5, %%mm1                   \n\t"\

  1036.         "packuswb %%mm3, %%mm1            \n\t"\

  1037.         OP_MMX2(%%mm1, (%1),%%mm4, q)\

  1038.         /* mm0= GHIJ, mm2=FGHI, mm5=HIJK, mm6=IJKL, mm7=0 */\

  1039.         \

  1040.         "movq 9(%0), %%mm1                \n\t" /* JKLMNOPQ */\

  1041.         "movq %%mm1, %%mm4                \n\t" /* JKLMNOPQ */\

  1042.         "movq %%mm1, %%mm3                \n\t" /* JKLMNOPQ */\

  1043.         "psrlq $8, %%mm1                  \n\t" /* KLMNOPQ0 */\

  1044.         "psrlq $16, %%mm4                 \n\t" /* LMNOPQ00 */\

  1045.         "punpcklbw %%mm7, %%mm1           \n\t" /* 0K0L0M0N */\

  1046.         "punpcklbw %%mm7, %%mm4           \n\t" /* 0L0M0N0O */\

  1047.         "paddw %%mm1, %%mm5               \n\t" /* b */\

  1048.         "paddw %%mm4, %%mm0               \n\t" /* c */\

  1049.         "paddw %%mm5, %%mm5               \n\t" /* 2b */\

  1050.         "psubw %%mm5, %%mm0               \n\t" /* c - 2b */\

  1051.         "movq %%mm3, %%mm5                \n\t" /* JKLMNOPQ */\

  1052.         "psrlq $24, %%mm3                 \n\t" /* MNOPQ000 */\

  1053.         "pmullw "MANGLE(ff_pw_3)", %%mm0  \n\t" /* 3c - 6b */\

  1054.         "punpcklbw %%mm7, %%mm3           \n\t" /* 0M0N0O0P */\

  1055.         "paddw %%mm3, %%mm2               \n\t" /* d */\

  1056.         "psubw %%mm2, %%mm0               \n\t" /* -6b + 3c - d */\

  1057.         "movq %%mm5, %%mm2                \n\t" /* JKLMNOPQ */\

  1058.         "punpcklbw %%mm7, %%mm2           \n\t" /* 0J0K0L0M */\

  1059.         "punpckhbw %%mm7, %%mm5           \n\t" /* 0N0O0P0Q */\

  1060.         "paddw %%mm2, %%mm6               \n\t" /* a */\

  1061.         "pmullw "MANGLE(ff_pw_20)", %%mm6 \n\t" /* 20a */\

  1062.         "paddw %6, %%mm0                  \n\t"\

  1063.         "paddw %%mm6, %%mm0               \n\t" /* 20a - 6b + 3c - d */\

  1064.         "psraw $5, %%mm0                  \n\t"\

  1065.         /* mm1=KLMN, mm2=JKLM, mm3=MNOP, mm4=LMNO, mm5=NOPQ mm7=0 */\

  1066.         \

  1067.         "paddw %%mm5, %%mm3               \n\t" /* a */\

  1068.         "pshufw $0xF9, %%mm5, %%mm6       \n\t" /* 0O0P0Q0Q */\

  1069.         "paddw %%mm4, %%mm6               \n\t" /* b */\

  1070.         "pshufw $0xBE, %%mm5, %%mm4       \n\t" /* 0P0Q0Q0P */\

  1071.         "pshufw $0x6F, %%mm5, %%mm5       \n\t" /* 0Q0Q0P0O */\

  1072.         "paddw %%mm1, %%mm4               \n\t" /* c */\

  1073.         "paddw %%mm2, %%mm5               \n\t" /* d */\

  1074.         "paddw %%mm6, %%mm6               \n\t" /* 2b */\

  1075.         "psubw %%mm6, %%mm4               \n\t" /* c - 2b */\

  1076.         "pmullw "MANGLE(ff_pw_20)", %%mm3 \n\t" /* 20a */\

  1077.         "pmullw "MANGLE(ff_pw_3)", %%mm4  \n\t" /* 3c - 6b */\

  1078.         "psubw %%mm5, %%mm3               \n\t" /* -6b + 3c - d */\

  1079.         "paddw %6, %%mm4                  \n\t"\

  1080.         "paddw %%mm3, %%mm4               \n\t" /* 20a - 6b + 3c - d */\

  1081.         "psraw $5, %%mm4                  \n\t"\

  1082.         "packuswb %%mm4, %%mm0            \n\t"\

  1083.         OP_MMX2(%%mm0, 8(%1), %%mm4, q)\

  1084.         \

  1085.         "add %3, %0                       \n\t"\

  1086.         "add %4, %1                       \n\t"\

  1087.         "decl %2                          \n\t"\

  1088.         " jnz 1b                          \n\t"\

  1089.         : "+a"(src), "+c"(dst), "+D"(h)\

  1090.         : "d"((x86_reg)srcStride), "S"((x86_reg)dstStride), /*"m"(ff_pw_20), "m"(ff_pw_3),*/ "m"(temp), "m"(ROUNDER)\

  1091.         : "memory"\

  1092.     );\

  1093. }\

  1094. \

  1095. static void OPNAME ## mpeg4_qpel16_h_lowpass_3dnow(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\

  1096.     int i;\

  1097.     int16_t temp[16];\

  1098.     /* quick HACK, XXX FIXME MUST be optimized */\

  1099.     for(i=0; i<h; i++)\

  1100.     {\

  1101.         temp[ 0]= (src[ 0]+src[ 1])*20 - (src[ 0]+src[ 2])*6 + (src[ 1]+src[ 3])*3 - (src[ 2]+src[ 4]);\

  1102.         temp[ 1]= (src[ 1]+src[ 2])*20 - (src[ 0]+src[ 3])*6 + (src[ 0]+src[ 4])*3 - (src[ 1]+src[ 5]);\

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

  1104.         temp[ 3]= (src[ 3]+src[ 4])*20 - (src[ 2]+src[ 5])*6 + (src[ 1]+src[ 6])*3 - (src[ 0]+src[ 7]);\

  1105.         temp[ 4]= (src[ 4]+src[ 5])*20 - (src[ 3]+src[ 6])*6 + (src[ 2]+src[ 7])*3 - (src[ 1]+src[ 8]);\

  1106.         temp[ 5]= (src[ 5]+src[ 6])*20 - (src[ 4]+src[ 7])*6 + (src[ 3]+src[ 8])*3 - (src[ 2]+src[ 9]);\

  1107.         temp[ 6]= (src[ 6]+src[ 7])*20 - (src[ 5]+src[ 8])*6 + (src[ 4]+src[ 9])*3 - (src[ 3]+src[10]);\

  1108.         temp[ 7]= (src[ 7]+src[ 8])*20 - (src[ 6]+src[ 9])*6 + (src[ 5]+src[10])*3 - (src[ 4]+src[11]);\

  1109.         temp[ 8]= (src[ 8]+src[ 9])*20 - (src[ 7]+src[10])*6 + (src[ 6]+src[11])*3 - (src[ 5]+src[12]);\

  1110.         temp[ 9]= (src[ 9]+src[10])*20 - (src[ 8]+src[11])*6 + (src[ 7]+src[12])*3 - (src[ 6]+src[13]);\

  1111.         temp[10]= (src[10]+src[11])*20 - (src[ 9]+src[12])*6 + (src[ 8]+src[13])*3 - (src[ 7]+src[14]);\

  1112.         temp[11]= (src[11]+src[12])*20 - (src[10]+src[13])*6 + (src[ 9]+src[14])*3 - (src[ 8]+src[15]);\

  1113.         temp[12]= (src[12]+src[13])*20 - (src[11]+src[14])*6 + (src[10]+src[15])*3 - (src[ 9]+src[16]);\

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

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

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

  1117.         __asm__ volatile(\

  1118.             "movq (%0), %%mm0               \n\t"\

  1119.             "movq 8(%0), %%mm1              \n\t"\

  1120.             "paddw %2, %%mm0                \n\t"\

  1121.             "paddw %2, %%mm1                \n\t"\

  1122.             "psraw $5, %%mm0                \n\t"\

  1123.             "psraw $5, %%mm1                \n\t"\

  1124.             "packuswb %%mm1, %%mm0          \n\t"\

  1125.             OP_3DNOW(%%mm0, (%1), %%mm1, q)\

  1126.             "movq 16(%0), %%mm0             \n\t"\

  1127.             "movq 24(%0), %%mm1             \n\t"\

  1128.             "paddw %2, %%mm0                \n\t"\

  1129.             "paddw %2, %%mm1                \n\t"\

  1130.             "psraw $5, %%mm0                \n\t"\

  1131.             "psraw $5, %%mm1                \n\t"\

  1132.             "packuswb %%mm1, %%mm0          \n\t"\

  1133.             OP_3DNOW(%%mm0, 8(%1), %%mm1, q)\

  1134.             :: "r"(temp), "r"(dst), "m"(ROUNDER)\

  1135.             : "memory"\

  1136.         );\

  1137.         dst+=dstStride;\

  1138.         src+=srcStride;\

  1139.     }\

  1140. }\

  1141. \

  1142. static void OPNAME ## mpeg4_qpel8_h_lowpass_mmx2(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\

  1143.     __asm__ volatile(\

  1144.         "pxor %%mm7, %%mm7                \n\t"\

  1145.         "1:                               \n\t"\

  1146.         "movq  (%0), %%mm0                \n\t" /* ABCDEFGH */\

  1147.         "movq %%mm0, %%mm1                \n\t" /* ABCDEFGH */\

  1148.         "movq %%mm0, %%mm2                \n\t" /* ABCDEFGH */\

  1149.         "punpcklbw %%mm7, %%mm0           \n\t" /* 0A0B0C0D */\

  1150.         "punpckhbw %%mm7, %%mm1           \n\t" /* 0E0F0G0H */\

  1151.         "pshufw $0x90, %%mm0, %%mm5       \n\t" /* 0A0A0B0C */\

  1152.         "pshufw $0x41, %%mm0, %%mm6       \n\t" /* 0B0A0A0B */\

  1153.         "movq %%mm2, %%mm3                \n\t" /* ABCDEFGH */\

  1154.         "movq %%mm2, %%mm4                \n\t" /* ABCDEFGH */\

  1155.         "psllq $8, %%mm2                  \n\t" /* 0ABCDEFG */\

  1156.         "psllq $16, %%mm3                 \n\t" /* 00ABCDEF */\

  1157.         "psllq $24, %%mm4                 \n\t" /* 000ABCDE */\

  1158.         "punpckhbw %%mm7, %%mm2           \n\t" /* 0D0E0F0G */\

  1159.         "punpckhbw %%mm7, %%mm3           \n\t" /* 0C0D0E0F */\

  1160.         "punpckhbw %%mm7, %%mm4           \n\t" /* 0B0C0D0E */\

  1161.         "paddw %%mm3, %%mm5               \n\t" /* b */\

  1162.         "paddw %%mm2, %%mm6               \n\t" /* c */\

  1163.         "paddw %%mm5, %%mm5               \n\t" /* 2b */\

  1164.         "psubw %%mm5, %%mm6               \n\t" /* c - 2b */\

  1165.         "pshufw $0x06, %%mm0, %%mm5       \n\t" /* 0C0B0A0A */\

  1166.         "pmullw "MANGLE(ff_pw_3)", %%mm6  \n\t" /* 3c - 6b */\

  1167.         "paddw %%mm4, %%mm0               \n\t" /* a */\

  1168.         "paddw %%mm1, %%mm5               \n\t" /* d */\

  1169.         "pmullw "MANGLE(ff_pw_20)", %%mm0 \n\t" /* 20a */\

  1170.         "psubw %%mm5, %%mm0               \n\t" /* 20a - d */\

  1171.         "paddw %5, %%mm6                  \n\t"\

  1172.         "paddw %%mm6, %%mm0               \n\t" /* 20a - 6b + 3c - d */\

  1173.         "psraw $5, %%mm0                  \n\t"\

  1174.         /* mm1=EFGH, mm2=DEFG, mm3=CDEF, mm4=BCDE, mm7=0 */\

  1175.         \

  1176.         "movd 5(%0), %%mm5                \n\t" /* FGHI */\

  1177.         "punpcklbw %%mm7, %%mm5           \n\t" /* 0F0G0H0I */\

  1178.         "pshufw $0xF9, %%mm5, %%mm6       \n\t" /* 0G0H0I0I */\

  1179.         "paddw %%mm5, %%mm1               \n\t" /* a */\

  1180.         "paddw %%mm6, %%mm2               \n\t" /* b */\

  1181.         "pshufw $0xBE, %%mm5, %%mm6       \n\t" /* 0H0I0I0H */\

  1182.         "pshufw $0x6F, %%mm5, %%mm5       \n\t" /* 0I0I0H0G */\

  1183.         "paddw %%mm6, %%mm3               \n\t" /* c */\

  1184.         "paddw %%mm5, %%mm4               \n\t" /* d */\

  1185.         "paddw %%mm2, %%mm2               \n\t" /* 2b */\

  1186.         "psubw %%mm2, %%mm3               \n\t" /* c - 2b */\

  1187.         "pmullw "MANGLE(ff_pw_20)", %%mm1 \n\t" /* 20a */\

  1188.         "pmullw "MANGLE(ff_pw_3)", %%mm3  \n\t" /* 3c - 6b */\

  1189.         "psubw %%mm4, %%mm3               \n\t" /* -6b + 3c - d */\

  1190.         "paddw %5, %%mm1                  \n\t"\

  1191.         "paddw %%mm1, %%mm3               \n\t" /* 20a - 6b + 3c - d */\

  1192.         "psraw $5, %%mm3                  \n\t"\

  1193.         "packuswb %%mm3, %%mm0            \n\t"\

  1194.         OP_MMX2(%%mm0, (%1), %%mm4, q)\

  1195.         \

  1196.         "add %3, %0                       \n\t"\

  1197.         "add %4, %1                       \n\t"\

  1198.         "decl %2                          \n\t"\

  1199.         " jnz 1b                          \n\t"\

  1200.         : "+a"(src), "+c"(dst), "+d"(h)\

  1201.         : "S"((x86_reg)srcStride), "D"((x86_reg)dstStride), /*"m"(ff_pw_20), "m"(ff_pw_3),*/ "m"(ROUNDER)\

  1202.         : "memory"\

  1203.     );\

  1204. }\

  1205. \

  1206. static void OPNAME ## mpeg4_qpel8_h_lowpass_3dnow(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\

  1207.     int i;\

  1208.     int16_t temp[8];\

  1209.     /* quick HACK, XXX FIXME MUST be optimized */\

  1210.     for(i=0; i<h; i++)\

  1211.     {\

  1212.         temp[ 0]= (src[ 0]+src[ 1])*20 - (src[ 0]+src[ 2])*6 + (src[ 1]+src[ 3])*3 - (src[ 2]+src[ 4]);\

  1213.         temp[ 1]= (src[ 1]+src[ 2])*20 - (src[ 0]+src[ 3])*6 + (src[ 0]+src[ 4])*3 - (src[ 1]+src[ 5]);\

  1214.         temp[ 2]= (src[ 2]+src[ 3])*20 - (src[ 1]+src[ 4])*6 + (src[ 0]+src[ 5])*3 - (src[ 0]+src[ 6]);\

  1215.         temp[ 3]= (src[ 3]+src[ 4])*20 - (src[ 2]+src[ 5])*6 + (src[ 1]+src[ 6])*3 - (src[ 0]+src[ 7]);\

  1216.         temp[ 4]= (src[ 4]+src[ 5])*20 - (src[ 3]+src[ 6])*6 + (src[ 2]+src[ 7])*3 - (src[ 1]+src[ 8]);\

  1217.         temp[ 5]= (src[ 5]+src[ 6])*20 - (src[ 4]+src[ 7])*6 + (src[ 3]+src[ 8])*3 - (src[ 2]+src[ 8]);\

  1218.         temp[ 6]= (src[ 6]+src[ 7])*20 - (src[ 5]+src[ 8])*6 + (src[ 4]+src[ 8])*3 - (src[ 3]+src[ 7]);\

  1219.         temp[ 7]= (src[ 7]+src[ 8])*20 - (src[ 6]+src[ 8])*6 + (src[ 5]+src[ 7])*3 - (src[ 4]+src[ 6]);\

  1220.         __asm__ volatile(\

  1221.             "movq (%0), %%mm0           \n\t"\

  1222.             "movq 8(%0), %%mm1          \n\t"\

  1223.             "paddw %2, %%mm0            \n\t"\

  1224.             "paddw %2, %%mm1            \n\t"\

  1225.             "psraw $5, %%mm0            \n\t"\

  1226.             "psraw $5, %%mm1            \n\t"\

  1227.             "packuswb %%mm1, %%mm0      \n\t"\

  1228.             OP_3DNOW(%%mm0, (%1), %%mm1, q)\

  1229.             :: "r"(temp), "r"(dst), "m"(ROUNDER)\

  1230.             :"memory"\

  1231.         );\

  1232.         dst+=dstStride;\

  1233.         src+=srcStride;\

  1234.     }\

  1235. }

  1236. 

  1237. #define QPEL_OP(OPNAME, ROUNDER, RND, OP, MMX)\

  1238. \

  1239. static void OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\

  1240.     uint64_t temp[17*4];\

  1241.     uint64_t *temp_ptr= temp;\

  1242.     int count= 17;\

  1243. \

  1244.     /*FIXME unroll */\

  1245.     __asm__ volatile(\

  1246.         "pxor %%mm7, %%mm7              \n\t"\

  1247.         "1:                             \n\t"\

  1248.         "movq (%0), %%mm0               \n\t"\

  1249.         "movq (%0), %%mm1               \n\t"\

  1250.         "movq 8(%0), %%mm2              \n\t"\

  1251.         "movq 8(%0), %%mm3              \n\t"\

  1252.         "punpcklbw %%mm7, %%mm0         \n\t"\

  1253.         "punpckhbw %%mm7, %%mm1         \n\t"\

  1254.         "punpcklbw %%mm7, %%mm2         \n\t"\

  1255.         "punpckhbw %%mm7, %%mm3         \n\t"\

  1256.         "movq %%mm0, (%1)               \n\t"\

  1257.         "movq %%mm1, 17*8(%1)           \n\t"\

  1258.         "movq %%mm2, 2*17*8(%1)         \n\t"\

  1259.         "movq %%mm3, 3*17*8(%1)         \n\t"\

  1260.         "add $8, %1                     \n\t"\

  1261.         "add %3, %0                     \n\t"\

  1262.         "decl %2                        \n\t"\

  1263.         " jnz 1b                        \n\t"\

  1264.         : "+r" (src), "+r" (temp_ptr), "+r"(count)\

  1265.         : "r" ((x86_reg)srcStride)\

  1266.         : "memory"\

  1267.     );\

  1268.     \

  1269.     temp_ptr= temp;\

  1270.     count=4;\

  1271.     \

  1272. /*FIXME reorder for speed */\

  1273.     __asm__ volatile(\

  1274.         /*"pxor %%mm7, %%mm7              \n\t"*/\

  1275.         "1:                             \n\t"\

  1276.         "movq (%0), %%mm0               \n\t"\

  1277.         "movq 8(%0), %%mm1              \n\t"\

  1278.         "movq 16(%0), %%mm2             \n\t"\

  1279.         "movq 24(%0), %%mm3             \n\t"\

  1280.         QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 16(%0),  8(%0),   (%0), 32(%0), (%1), OP)\

  1281.         QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5,  8(%0),   (%0),   (%0), 40(%0), (%1, %3), OP)\

  1282.         "add %4, %1                     \n\t"\

  1283.         QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5,   (%0),   (%0),  8(%0), 48(%0), (%1), OP)\

  1284.         \

  1285.         QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5,   (%0),  8(%0), 16(%0), 56(%0), (%1, %3), OP)\

  1286.         "add %4, %1                     \n\t"\

  1287.         QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5,  8(%0), 16(%0), 24(%0), 64(%0), (%1), OP)\

  1288.         QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 16(%0), 24(%0), 32(%0), 72(%0), (%1, %3), OP)\

  1289.         "add %4, %1                     \n\t"\

  1290.         QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 24(%0), 32(%0), 40(%0), 80(%0), (%1), OP)\

  1291.         QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 32(%0), 40(%0), 48(%0), 88(%0), (%1, %3), OP)\

  1292.         "add %4, %1                     \n\t"\

  1293.         QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 40(%0), 48(%0), 56(%0), 96(%0), (%1), OP)\

  1294.         QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 48(%0), 56(%0), 64(%0),104(%0), (%1, %3), OP)\

  1295.         "add %4, %1                     \n\t"\

  1296.         QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 56(%0), 64(%0), 72(%0),112(%0), (%1), OP)\

  1297.         QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 64(%0), 72(%0), 80(%0),120(%0), (%1, %3), OP)\

  1298.         "add %4, %1                     \n\t"\

  1299.         QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 72(%0), 80(%0), 88(%0),128(%0), (%1), OP)\

  1300.         \

  1301.         QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 80(%0), 88(%0), 96(%0),128(%0), (%1, %3), OP)\

  1302.         "add %4, %1                     \n\t"  \

  1303.         QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 88(%0), 96(%0),104(%0),120(%0), (%1), OP)\

  1304.         QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 96(%0),104(%0),112(%0),112(%0), (%1, %3), OP)\

  1305.         \

  1306.         "add $136, %0                   \n\t"\

  1307.         "add %6, %1                     \n\t"\

  1308.         "decl %2                        \n\t"\

  1309.         " jnz 1b                        \n\t"\

  1310.         \

  1311.         : "+r"(temp_ptr), "+r"(dst), "+g"(count)\

  1312.         : "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)\

  1313.         :"memory"\

  1314.     );\

  1315. }\

  1316. \

  1317. static void OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\

  1318.     uint64_t temp[9*2];\

  1319.     uint64_t *temp_ptr= temp;\

  1320.     int count= 9;\

  1321. \

  1322.     /*FIXME unroll */\

  1323.     __asm__ volatile(\

  1324.         "pxor %%mm7, %%mm7              \n\t"\

  1325.         "1:                             \n\t"\

  1326.         "movq (%0), %%mm0               \n\t"\

  1327.         "movq (%0), %%mm1               \n\t"\

  1328.         "punpcklbw %%mm7, %%mm0         \n\t"\

  1329.         "punpckhbw %%mm7, %%mm1         \n\t"\

  1330.         "movq %%mm0, (%1)               \n\t"\

  1331.         "movq %%mm1, 9*8(%1)            \n\t"\

  1332.         "add $8, %1                     \n\t"\

  1333.         "add %3, %0                     \n\t"\

  1334.         "decl %2                        \n\t"\

  1335.         " jnz 1b                        \n\t"\

  1336.         : "+r" (src), "+r" (temp_ptr), "+r"(count)\

  1337.         : "r" ((x86_reg)srcStride)\

  1338.         : "memory"\

  1339.     );\

  1340.     \

  1341.     temp_ptr= temp;\

  1342.     count=2;\

  1343.     \

  1344. /*FIXME reorder for speed */\

  1345.     __asm__ volatile(\

  1346.         /*"pxor %%mm7, %%mm7              \n\t"*/\

  1347.         "1:                             \n\t"\

  1348.         "movq (%0), %%mm0               \n\t"\

  1349.         "movq 8(%0), %%mm1              \n\t"\

  1350.         "movq 16(%0), %%mm2             \n\t"\

  1351.         "movq 24(%0), %%mm3             \n\t"\

  1352.         QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 16(%0),  8(%0),   (%0), 32(%0), (%1), OP)\

  1353.         QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5,  8(%0),   (%0),   (%0), 40(%0), (%1, %3), OP)\

  1354.         "add %4, %1                     \n\t"\

  1355.         QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5,   (%0),   (%0),  8(%0), 48(%0), (%1), OP)\

  1356.         \

  1357.         QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5,   (%0),  8(%0), 16(%0), 56(%0), (%1, %3), OP)\

  1358.         "add %4, %1                     \n\t"\

  1359.         QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5,  8(%0), 16(%0), 24(%0), 64(%0), (%1), OP)\

  1360.         \

  1361.         QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 16(%0), 24(%0), 32(%0), 64(%0), (%1, %3), OP)\

  1362.         "add %4, %1                     \n\t"\

  1363.         QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 24(%0), 32(%0), 40(%0), 56(%0), (%1), OP)\

  1364.         QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 32(%0), 40(%0), 48(%0), 48(%0), (%1, %3), OP)\

  1365.                 \

  1366.         "add $72, %0                    \n\t"\

  1367.         "add %6, %1                     \n\t"\

  1368.         "decl %2                        \n\t"\

  1369.         " jnz 1b                        \n\t"\

  1370.          \

  1371.         : "+r"(temp_ptr), "+r"(dst), "+g"(count)\

  1372.         : "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)\

  1373.         : "memory"\

  1374.    );\

  1375. }\

  1376. \

  1377. static void OPNAME ## qpel8_mc00_ ## MMX (uint8_t *dst, uint8_t *src, int stride){\

  1378.     OPNAME ## pixels8_ ## MMX(dst, src, stride, 8);\

  1379. }\

  1380. \

  1381. static void OPNAME ## qpel8_mc10_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1382.     uint64_t temp[8];\

  1383.     uint8_t * const half= (uint8_t*)temp;\

  1384.     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(half, src, 8, stride, 8);\

  1385.     OPNAME ## pixels8_l2_ ## MMX(dst, src, half, stride, stride, 8);\

  1386. }\

  1387. \

  1388. static void OPNAME ## qpel8_mc20_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1389.     OPNAME ## mpeg4_qpel8_h_lowpass_ ## MMX(dst, src, stride, stride, 8);\

  1390. }\

  1391. \

  1392. static void OPNAME ## qpel8_mc30_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1393.     uint64_t temp[8];\

  1394.     uint8_t * const half= (uint8_t*)temp;\

  1395.     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(half, src, 8, stride, 8);\

  1396.     OPNAME ## pixels8_l2_ ## MMX(dst, src+1, half, stride, stride, 8);\

  1397. }\

  1398. \

  1399. static void OPNAME ## qpel8_mc01_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1400.     uint64_t temp[8];\

  1401.     uint8_t * const half= (uint8_t*)temp;\

  1402.     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(half, src, 8, stride);\

  1403.     OPNAME ## pixels8_l2_ ## MMX(dst, src, half, stride, stride, 8);\

  1404. }\

  1405. \

  1406. static void OPNAME ## qpel8_mc02_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1407.     OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, src, stride, stride);\

  1408. }\

  1409. \

  1410. static void OPNAME ## qpel8_mc03_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1411.     uint64_t temp[8];\

  1412.     uint8_t * const half= (uint8_t*)temp;\

  1413.     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(half, src, 8, stride);\

  1414.     OPNAME ## pixels8_l2_ ## MMX(dst, src+stride, half, stride, stride, 8);\

  1415. }\

  1416. static void OPNAME ## qpel8_mc11_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1417.     uint64_t half[8 + 9];\

  1418.     uint8_t * const halfH= ((uint8_t*)half) + 64;\

  1419.     uint8_t * const halfHV= ((uint8_t*)half);\

  1420.     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\

  1421.     put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8, stride, 9);\

  1422.     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\

  1423.     OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV, stride, 8, 8);\

  1424. }\

  1425. static void OPNAME ## qpel8_mc31_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1426.     uint64_t half[8 + 9];\

  1427.     uint8_t * const halfH= ((uint8_t*)half) + 64;\

  1428.     uint8_t * const halfHV= ((uint8_t*)half);\

  1429.     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\

  1430.     put ## RND ## pixels8_l2_ ## MMX(halfH, src+1, halfH, 8, stride, 9);\

  1431.     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\

  1432.     OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV, stride, 8, 8);\

  1433. }\

  1434. static void OPNAME ## qpel8_mc13_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1435.     uint64_t half[8 + 9];\

  1436.     uint8_t * const halfH= ((uint8_t*)half) + 64;\

  1437.     uint8_t * const halfHV= ((uint8_t*)half);\

  1438.     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\

  1439.     put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8, stride, 9);\

  1440.     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\

  1441.     OPNAME ## pixels8_l2_ ## MMX(dst, halfH+8, halfHV, stride, 8, 8);\

  1442. }\

  1443. static void OPNAME ## qpel8_mc33_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1444.     uint64_t half[8 + 9];\

  1445.     uint8_t * const halfH= ((uint8_t*)half) + 64;\

  1446.     uint8_t * const halfHV= ((uint8_t*)half);\

  1447.     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\

  1448.     put ## RND ## pixels8_l2_ ## MMX(halfH, src+1, halfH, 8, stride, 9);\

  1449.     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\

  1450.     OPNAME ## pixels8_l2_ ## MMX(dst, halfH+8, halfHV, stride, 8, 8);\

  1451. }\

  1452. static void OPNAME ## qpel8_mc21_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1453.     uint64_t half[8 + 9];\

  1454.     uint8_t * const halfH= ((uint8_t*)half) + 64;\

  1455.     uint8_t * const halfHV= ((uint8_t*)half);\

  1456.     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\

  1457.     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\

  1458.     OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV, stride, 8, 8);\

  1459. }\

  1460. static void OPNAME ## qpel8_mc23_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1461.     uint64_t half[8 + 9];\

  1462.     uint8_t * const halfH= ((uint8_t*)half) + 64;\

  1463.     uint8_t * const halfHV= ((uint8_t*)half);\

  1464.     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\

  1465.     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\

  1466.     OPNAME ## pixels8_l2_ ## MMX(dst, halfH+8, halfHV, stride, 8, 8);\

  1467. }\

  1468. static void OPNAME ## qpel8_mc12_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1469.     uint64_t half[8 + 9];\

  1470.     uint8_t * const halfH= ((uint8_t*)half);\

  1471.     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\

  1472.     put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8, stride, 9);\

  1473.     OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH, stride, 8);\

  1474. }\

  1475. static void OPNAME ## qpel8_mc32_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1476.     uint64_t half[8 + 9];\

  1477.     uint8_t * const halfH= ((uint8_t*)half);\

  1478.     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\

  1479.     put ## RND ## pixels8_l2_ ## MMX(halfH, src+1, halfH, 8, stride, 9);\

  1480.     OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH, stride, 8);\

  1481. }\

  1482. static void OPNAME ## qpel8_mc22_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1483.     uint64_t half[9];\

  1484.     uint8_t * const halfH= ((uint8_t*)half);\

  1485.     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\

  1486.     OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH, stride, 8);\

  1487. }\

  1488. static void OPNAME ## qpel16_mc00_ ## MMX (uint8_t *dst, uint8_t *src, int stride){\

  1489.     OPNAME ## pixels16_ ## MMX(dst, src, stride, 16);\

  1490. }\

  1491. \

  1492. static void OPNAME ## qpel16_mc10_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1493.     uint64_t temp[32];\

  1494.     uint8_t * const half= (uint8_t*)temp;\

  1495.     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(half, src, 16, stride, 16);\

  1496.     OPNAME ## pixels16_l2_ ## MMX(dst, src, half, stride, stride, 16);\

  1497. }\

  1498. \

  1499. static void OPNAME ## qpel16_mc20_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1500.     OPNAME ## mpeg4_qpel16_h_lowpass_ ## MMX(dst, src, stride, stride, 16);\

  1501. }\

  1502. \

  1503. static void OPNAME ## qpel16_mc30_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1504.     uint64_t temp[32];\

  1505.     uint8_t * const half= (uint8_t*)temp;\

  1506.     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(half, src, 16, stride, 16);\

  1507.     OPNAME ## pixels16_l2_ ## MMX(dst, src+1, half, stride, stride, 16);\

  1508. }\

  1509. \

  1510. static void OPNAME ## qpel16_mc01_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1511.     uint64_t temp[32];\

  1512.     uint8_t * const half= (uint8_t*)temp;\

  1513.     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(half, src, 16, stride);\

  1514.     OPNAME ## pixels16_l2_ ## MMX(dst, src, half, stride, stride, 16);\

  1515. }\

  1516. \

  1517. static void OPNAME ## qpel16_mc02_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1518.     OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, src, stride, stride);\

  1519. }\

  1520. \

  1521. static void OPNAME ## qpel16_mc03_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1522.     uint64_t temp[32];\

  1523.     uint8_t * const half= (uint8_t*)temp;\

  1524.     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(half, src, 16, stride);\

  1525.     OPNAME ## pixels16_l2_ ## MMX(dst, src+stride, half, stride, stride, 16);\

  1526. }\

  1527. static void OPNAME ## qpel16_mc11_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1528.     uint64_t half[16*2 + 17*2];\

  1529.     uint8_t * const halfH= ((uint8_t*)half) + 256;\

  1530.     uint8_t * const halfHV= ((uint8_t*)half);\

  1531.     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\

  1532.     put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16, stride, 17);\

  1533.     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\

  1534.     OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV, stride, 16, 16);\

  1535. }\

  1536. static void OPNAME ## qpel16_mc31_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1537.     uint64_t half[16*2 + 17*2];\

  1538.     uint8_t * const halfH= ((uint8_t*)half) + 256;\

  1539.     uint8_t * const halfHV= ((uint8_t*)half);\

  1540.     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\

  1541.     put ## RND ## pixels16_l2_ ## MMX(halfH, src+1, halfH, 16, stride, 17);\

  1542.     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\

  1543.     OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV, stride, 16, 16);\

  1544. }\

  1545. static void OPNAME ## qpel16_mc13_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1546.     uint64_t half[16*2 + 17*2];\

  1547.     uint8_t * const halfH= ((uint8_t*)half) + 256;\

  1548.     uint8_t * const halfHV= ((uint8_t*)half);\

  1549.     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\

  1550.     put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16, stride, 17);\

  1551.     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\

  1552.     OPNAME ## pixels16_l2_ ## MMX(dst, halfH+16, halfHV, stride, 16, 16);\

  1553. }\

  1554. static void OPNAME ## qpel16_mc33_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1555.     uint64_t half[16*2 + 17*2];\

  1556.     uint8_t * const halfH= ((uint8_t*)half) + 256;\

  1557.     uint8_t * const halfHV= ((uint8_t*)half);\

  1558.     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\

  1559.     put ## RND ## pixels16_l2_ ## MMX(halfH, src+1, halfH, 16, stride, 17);\

  1560.     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\

  1561.     OPNAME ## pixels16_l2_ ## MMX(dst, halfH+16, halfHV, stride, 16, 16);\

  1562. }\

  1563. static void OPNAME ## qpel16_mc21_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1564.     uint64_t half[16*2 + 17*2];\

  1565.     uint8_t * const halfH= ((uint8_t*)half) + 256;\

  1566.     uint8_t * const halfHV= ((uint8_t*)half);\

  1567.     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\

  1568.     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\

  1569.     OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV, stride, 16, 16);\

  1570. }\

  1571. static void OPNAME ## qpel16_mc23_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1572.     uint64_t half[16*2 + 17*2];\

  1573.     uint8_t * const halfH= ((uint8_t*)half) + 256;\

  1574.     uint8_t * const halfHV= ((uint8_t*)half);\

  1575.     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\

  1576.     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\

  1577.     OPNAME ## pixels16_l2_ ## MMX(dst, halfH+16, halfHV, stride, 16, 16);\

  1578. }\

  1579. static void OPNAME ## qpel16_mc12_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1580.     uint64_t half[17*2];\

  1581.     uint8_t * const halfH= ((uint8_t*)half);\

  1582.     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\

  1583.     put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16, stride, 17);\

  1584.     OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH, stride, 16);\

  1585. }\

  1586. static void OPNAME ## qpel16_mc32_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1587.     uint64_t half[17*2];\

  1588.     uint8_t * const halfH= ((uint8_t*)half);\

  1589.     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\

  1590.     put ## RND ## pixels16_l2_ ## MMX(halfH, src+1, halfH, 16, stride, 17);\

  1591.     OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH, stride, 16);\

  1592. }\

  1593. static void OPNAME ## qpel16_mc22_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1594.     uint64_t half[17*2];\

  1595.     uint8_t * const halfH= ((uint8_t*)half);\

  1596.     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\

  1597.     OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH, stride, 16);\

  1598. }

  1599. 

  1600. #define PUT_OP(a,b,temp, size) "mov" #size " " #a ", " #b "        \n\t"

  1601. #define AVG_3DNOW_OP(a,b,temp, size) \

  1602. "mov" #size " " #b ", " #temp "   \n\t"\

  1603. "pavgusb " #temp ", " #a "        \n\t"\

  1604. "mov" #size " " #a ", " #b "      \n\t"

  1605. #define AVG_MMX2_OP(a,b,temp, size) \

  1606. "mov" #size " " #b ", " #temp "   \n\t"\

  1607. "pavgb " #temp ", " #a "          \n\t"\

  1608. "mov" #size " " #a ", " #b "      \n\t"

  1609. 

  1610. QPEL_BASE(put_       , ff_pw_16, _       , PUT_OP, PUT_OP)

  1611. QPEL_BASE(avg_       , ff_pw_16, _       , AVG_MMX2_OP, AVG_3DNOW_OP)

  1612. QPEL_BASE(put_no_rnd_, ff_pw_15, _no_rnd_, PUT_OP, PUT_OP)

  1613. QPEL_OP(put_       , ff_pw_16, _       , PUT_OP, 3dnow)

  1614. QPEL_OP(avg_       , ff_pw_16, _       , AVG_3DNOW_OP, 3dnow)

  1615. QPEL_OP(put_no_rnd_, ff_pw_15, _no_rnd_, PUT_OP, 3dnow)

  1616. QPEL_OP(put_       , ff_pw_16, _       , PUT_OP, mmx2)

  1617. QPEL_OP(avg_       , ff_pw_16, _       , AVG_MMX2_OP, mmx2)

  1618. QPEL_OP(put_no_rnd_, ff_pw_15, _no_rnd_, PUT_OP, mmx2)

  1619. 

  1620. /***********************************/

  1621. /* bilinear qpel: not compliant to any spec, only for -lavdopts fast */

  1622. 

  1623. #define QPEL_2TAP_XY(OPNAME, SIZE, MMX, XY, HPEL)\

  1624. static void OPNAME ## 2tap_qpel ## SIZE ## _mc ## XY ## _ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1625.     OPNAME ## pixels ## SIZE ## HPEL(dst, src, stride, SIZE);\

  1626. }

  1627. #define QPEL_2TAP_L3(OPNAME, SIZE, MMX, XY, S0, S1, S2)\

  1628. static void OPNAME ## 2tap_qpel ## SIZE ## _mc ## XY ## _ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1629.     OPNAME ## 2tap_qpel ## SIZE ## _l3_ ## MMX(dst, src+S0, stride, SIZE, S1, S2);\

  1630. }

  1631. 

  1632. #define QPEL_2TAP(OPNAME, SIZE, MMX)\

  1633. QPEL_2TAP_XY(OPNAME, SIZE, MMX, 20, _x2_ ## MMX)\

  1634. QPEL_2TAP_XY(OPNAME, SIZE, MMX, 02, _y2_ ## MMX)\

  1635. QPEL_2TAP_XY(OPNAME, SIZE, MMX, 22, _xy2_mmx)\

  1636. static const qpel_mc_func OPNAME ## 2tap_qpel ## SIZE ## _mc00_ ## MMX =\

  1637.                           OPNAME ## qpel ## SIZE ## _mc00_ ## MMX;\

  1638. static const qpel_mc_func OPNAME ## 2tap_qpel ## SIZE ## _mc21_ ## MMX =\

  1639.                           OPNAME ## 2tap_qpel ## SIZE ## _mc20_ ## MMX;\

  1640. static const qpel_mc_func OPNAME ## 2tap_qpel ## SIZE ## _mc12_ ## MMX =\

  1641.                           OPNAME ## 2tap_qpel ## SIZE ## _mc02_ ## MMX;\

  1642. static void OPNAME ## 2tap_qpel ## SIZE ## _mc32_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1643.     OPNAME ## pixels ## SIZE ## _y2_ ## MMX(dst, src+1, stride, SIZE);\

  1644. }\

  1645. static void OPNAME ## 2tap_qpel ## SIZE ## _mc23_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1646.     OPNAME ## pixels ## SIZE ## _x2_ ## MMX(dst, src+stride, stride, SIZE);\

  1647. }\

  1648. QPEL_2TAP_L3(OPNAME, SIZE, MMX, 10, 0,         1,       0)\

  1649. QPEL_2TAP_L3(OPNAME, SIZE, MMX, 30, 1,        -1,       0)\

  1650. QPEL_2TAP_L3(OPNAME, SIZE, MMX, 01, 0,         stride,  0)\

  1651. QPEL_2TAP_L3(OPNAME, SIZE, MMX, 03, stride,   -stride,  0)\

  1652. QPEL_2TAP_L3(OPNAME, SIZE, MMX, 11, 0,         stride,  1)\

  1653. QPEL_2TAP_L3(OPNAME, SIZE, MMX, 31, 1,         stride, -1)\

  1654. QPEL_2TAP_L3(OPNAME, SIZE, MMX, 13, stride,   -stride,  1)\

  1655. QPEL_2TAP_L3(OPNAME, SIZE, MMX, 33, stride+1, -stride, -1)\

  1656. 

  1657. QPEL_2TAP(put_, 16, mmx2)

  1658. QPEL_2TAP(avg_, 16, mmx2)

  1659. QPEL_2TAP(put_,  8, mmx2)

  1660. QPEL_2TAP(avg_,  8, mmx2)

  1661. QPEL_2TAP(put_, 16, 3dnow)

  1662. QPEL_2TAP(avg_, 16, 3dnow)

  1663. QPEL_2TAP(put_,  8, 3dnow)

  1664. QPEL_2TAP(avg_,  8, 3dnow)

  1665. 

  1666. 

  1667. #if 0

  1668. static void just_return(void) { return; }

  1669. #endif

  1670. 

  1671. #if HAVE_YASM

  1672. typedef void emu_edge_core_func (uint8_t *buf, const uint8_t *src,

  1673.                                  x86_reg linesize, x86_reg start_y,

  1674.                                  x86_reg end_y, x86_reg block_h,

  1675.                                  x86_reg start_x, x86_reg end_x,

  1676.                                  x86_reg block_w);

  1677. extern emu_edge_core_func ff_emu_edge_core_mmx;

  1678. extern emu_edge_core_func ff_emu_edge_core_sse;

  1679. 

  1680. static av_always_inline

  1681. void emulated_edge_mc(uint8_t *buf, const uint8_t *src, int linesize,

  1682.                       int block_w, int block_h,

  1683.                       int src_x, int src_y, int w, int h,

  1684.                       emu_edge_core_func *core_fn)

  1685. {

  1686.     int start_y, start_x, end_y, end_x, src_y_add=0;

  1687. 

  1688.     if(src_y>= h){

  1689.         src_y_add = h-1-src_y;

  1690.         src_y=h-1;

  1691.     }else if(src_y<=-block_h){

  1692.         src_y_add = 1-block_h-src_y;

  1693.         src_y=1-block_h;

  1694.     }

  1695.     if(src_x>= w){

  1696.         src+= (w-1-src_x);

  1697.         src_x=w-1;

  1698.     }else if(src_x<=-block_w){

  1699.         src+= (1-block_w-src_x);

  1700.         src_x=1-block_w;

  1701.     }

  1702. 

  1703.     start_y= FFMAX(0, -src_y);

  1704.     start_x= FFMAX(0, -src_x);

  1705.     end_y= FFMIN(block_h, h-src_y);

  1706.     end_x= FFMIN(block_w, w-src_x);

  1707.     assert(start_x < end_x && block_w > 0);

  1708.     assert(start_y < end_y && block_h > 0);

  1709. 

  1710.     // fill in the to-be-copied part plus all above/below

  1711.     src += (src_y_add+start_y)*linesize + start_x;

  1712.     buf += start_x;

  1713.     core_fn(buf, src, linesize, start_y, end_y, block_h, start_x, end_x, block_w);

  1714. }

  1715. 

  1716. #if ARCH_X86_32

  1717. static av_noinline

  1718. void emulated_edge_mc_mmx(uint8_t *buf, const uint8_t *src, int linesize,

  1719.                           int block_w, int block_h,

  1720.                           int src_x, int src_y, int w, int h)

  1721. {

  1722.     emulated_edge_mc(buf, src, linesize, block_w, block_h, src_x, src_y,

  1723.                      w, h, &ff_emu_edge_core_mmx);

  1724. }

  1725. #endif

  1726. static av_noinline

  1727. void emulated_edge_mc_sse(uint8_t *buf, const uint8_t *src, int linesize,

  1728.                           int block_w, int block_h,

  1729.                           int src_x, int src_y, int w, int h)

  1730. {

  1731.     emulated_edge_mc(buf, src, linesize, block_w, block_h, src_x, src_y,

  1732.                      w, h, &ff_emu_edge_core_sse);

  1733. }

  1734. #endif /* HAVE_YASM */

  1735. 

  1736. typedef void emulated_edge_mc_func (uint8_t *dst, const uint8_t *src,

  1737.                                     int linesize, int block_w, int block_h,

  1738.                                     int src_x, int src_y, int w, int h);

  1739. 

  1740. static av_always_inline

  1741. void gmc(uint8_t *dst, uint8_t *src, int stride, int h, int ox, int oy,

  1742.          int dxx, int dxy, int dyx, int dyy, int shift, int r, int width, int height,

  1743.          emulated_edge_mc_func *emu_edge_fn)

  1744. {

  1745.     const int w = 8;

  1746.     const int ix = ox>>(16+shift);

  1747.     const int iy = oy>>(16+shift);

  1748.     const int oxs = ox>>4;

  1749.     const int oys = oy>>4;

  1750.     const int dxxs = dxx>>4;

  1751.     const int dxys = dxy>>4;

  1752.     const int dyxs = dyx>>4;

  1753.     const int dyys = dyy>>4;

  1754.     const uint16_t r4[4] = {r,r,r,r};

  1755.     const uint16_t dxy4[4] = {dxys,dxys,dxys,dxys};

  1756.     const uint16_t dyy4[4] = {dyys,dyys,dyys,dyys};

  1757.     const uint64_t shift2 = 2*shift;

  1758.     uint8_t edge_buf[(h+1)*stride];

  1759.     int x, y;

  1760. 

  1761.     const int dxw = (dxx-(1<<(16+shift)))*(w-1);

  1762.     const int dyh = (dyy-(1<<(16+shift)))*(h-1);

  1763.     const int dxh = dxy*(h-1);

  1764.     const int dyw = dyx*(w-1);

  1765.     if( // non-constant fullpel offset (3% of blocks)

  1766.         ((ox^(ox+dxw)) | (ox^(ox+dxh)) | (ox^(ox+dxw+dxh)) |

  1767.          (oy^(oy+dyw)) | (oy^(oy+dyh)) | (oy^(oy+dyw+dyh))) >> (16+shift)

  1768.         // uses more than 16 bits of subpel mv (only at huge resolution)

  1769.         || (dxx|dxy|dyx|dyy)&15 )

  1770.     {

  1771.         //FIXME could still use mmx for some of the rows

  1772.         ff_gmc_c(dst, src, stride, h, ox, oy, dxx, dxy, dyx, dyy, shift, r, width, height);

  1773.         return;

  1774.     }

  1775. 

  1776.     src += ix + iy*stride;

  1777.     if( (unsigned)ix >= width-w ||

  1778.         (unsigned)iy >= height-h )

  1779.     {

  1780.         emu_edge_fn(edge_buf, src, stride, w+1, h+1, ix, iy, width, height);

  1781.         src = edge_buf;

  1782.     }

  1783. 

  1784.     __asm__ volatile(

  1785.         "movd         %0, %%mm6 \n\t"

  1786.         "pxor      %%mm7, %%mm7 \n\t"

  1787.         "punpcklwd %%mm6, %%mm6 \n\t"

  1788.         "punpcklwd %%mm6, %%mm6 \n\t"

  1789.         :: "r"(1<<shift)

  1790.     );

  1791. 

  1792.     for(x=0; x<w; x+=4){

  1793.         uint16_t dx4[4] = { oxs - dxys + dxxs*(x+0),

  1794.                             oxs - dxys + dxxs*(x+1),

  1795.                             oxs - dxys + dxxs*(x+2),

  1796.                             oxs - dxys + dxxs*(x+3) };

  1797.         uint16_t dy4[4] = { oys - dyys + dyxs*(x+0),

  1798.                             oys - dyys + dyxs*(x+1),

  1799.                             oys - dyys + dyxs*(x+2),

  1800.                             oys - dyys + dyxs*(x+3) };

  1801. 

  1802.         for(y=0; y<h; y++){

  1803.             __asm__ volatile(

  1804.                 "movq   %0,  %%mm4 \n\t"

  1805.                 "movq   %1,  %%mm5 \n\t"

  1806.                 "paddw  %2,  %%mm4 \n\t"

  1807.                 "paddw  %3,  %%mm5 \n\t"

  1808.                 "movq   %%mm4, %0  \n\t"

  1809.                 "movq   %%mm5, %1  \n\t"

  1810.                 "psrlw  $12, %%mm4 \n\t"

  1811.                 "psrlw  $12, %%mm5 \n\t"

  1812.                 : "+m"(*dx4), "+m"(*dy4)

  1813.                 : "m"(*dxy4), "m"(*dyy4)

  1814.             );

  1815. 

  1816.             __asm__ volatile(

  1817.                 "movq   %%mm6, %%mm2 \n\t"

  1818.                 "movq   %%mm6, %%mm1 \n\t"

  1819.                 "psubw  %%mm4, %%mm2 \n\t"

  1820.                 "psubw  %%mm5, %%mm1 \n\t"

  1821.                 "movq   %%mm2, %%mm0 \n\t"

  1822.                 "movq   %%mm4, %%mm3 \n\t"

  1823.                 "pmullw %%mm1, %%mm0 \n\t" // (s-dx)*(s-dy)

  1824.                 "pmullw %%mm5, %%mm3 \n\t" // dx*dy

  1825.                 "pmullw %%mm5, %%mm2 \n\t" // (s-dx)*dy

  1826.                 "pmullw %%mm4, %%mm1 \n\t" // dx*(s-dy)

  1827. 

  1828.                 "movd   %4,    %%mm5 \n\t"

  1829.                 "movd   %3,    %%mm4 \n\t"

  1830.                 "punpcklbw %%mm7, %%mm5 \n\t"

  1831.                 "punpcklbw %%mm7, %%mm4 \n\t"

  1832.                 "pmullw %%mm5, %%mm3 \n\t" // src[1,1] * dx*dy

  1833.                 "pmullw %%mm4, %%mm2 \n\t" // src[0,1] * (s-dx)*dy

  1834. 

  1835.                 "movd   %2,    %%mm5 \n\t"

  1836.                 "movd   %1,    %%mm4 \n\t"

  1837.                 "punpcklbw %%mm7, %%mm5 \n\t"

  1838.                 "punpcklbw %%mm7, %%mm4 \n\t"

  1839.                 "pmullw %%mm5, %%mm1 \n\t" // src[1,0] * dx*(s-dy)

  1840.                 "pmullw %%mm4, %%mm0 \n\t" // src[0,0] * (s-dx)*(s-dy)

  1841.                 "paddw  %5,    %%mm1 \n\t"

  1842.                 "paddw  %%mm3, %%mm2 \n\t"

  1843.                 "paddw  %%mm1, %%mm0 \n\t"

  1844.                 "paddw  %%mm2, %%mm0 \n\t"

  1845. 

  1846.                 "psrlw    %6,    %%mm0 \n\t"

  1847.                 "packuswb %%mm0, %%mm0 \n\t"

  1848.                 "movd     %%mm0, %0    \n\t"

  1849. 

  1850.                 : "=m"(dst[x+y*stride])

  1851.                 : "m"(src[0]), "m"(src[1]),

  1852.                   "m"(src[stride]), "m"(src[stride+1]),

  1853.                   "m"(*r4), "m"(shift2)

  1854.             );

  1855.             src += stride;

  1856.         }

  1857.         src += 4-h*stride;

  1858.     }

  1859. }

  1860. 

  1861. #if HAVE_YASM

  1862. #if ARCH_X86_32

  1863. static void gmc_mmx(uint8_t *dst, uint8_t *src, int stride, int h, int ox, int oy,

  1864.                     int dxx, int dxy, int dyx, int dyy, int shift, int r, int width, int height)

  1865. {

  1866.     gmc(dst, src, stride, h, ox, oy, dxx, dxy, dyx, dyy, shift, r,

  1867.         width, height, &emulated_edge_mc_mmx);

  1868. }

  1869. #endif

  1870. static void gmc_sse(uint8_t *dst, uint8_t *src, int stride, int h, int ox, int oy,

  1871.                     int dxx, int dxy, int dyx, int dyy, int shift, int r, int width, int height)

  1872. {

  1873.     gmc(dst, src, stride, h, ox, oy, dxx, dxy, dyx, dyy, shift, r,

  1874.         width, height, &emulated_edge_mc_sse);

  1875. }

  1876. #else

  1877. static void gmc_mmx(uint8_t *dst, uint8_t *src, int stride, int h, int ox, int oy,

  1878.                     int dxx, int dxy, int dyx, int dyy, int shift, int r, int width, int height)

  1879. {

  1880.     gmc(dst, src, stride, h, ox, oy, dxx, dxy, dyx, dyy, shift, r,

  1881.         width, height, &ff_emulated_edge_mc);

  1882. }

  1883. #endif

  1884. 

  1885. #define PREFETCH(name, op) \

  1886. static void name(void *mem, int stride, int h){\

  1887.     const uint8_t *p= mem;\

  1888.     do{\

  1889.         __asm__ volatile(#op" %0" :: "m"(*p));\

  1890.         p+= stride;\

  1891.     }while(--h);\

  1892. }

  1893. PREFETCH(prefetch_mmx2,  prefetcht0)

  1894. PREFETCH(prefetch_3dnow, prefetch)

  1895. #undef PREFETCH

  1896. 

  1897. #include "h264_qpel_mmx.c"

  1898. 

  1899. void ff_put_h264_chroma_mc8_mmx_rnd   (uint8_t *dst, uint8_t *src,

  1900.                                        int stride, int h, int x, int y);

  1901. void ff_put_rv40_chroma_mc8_mmx       (uint8_t *dst, uint8_t *src,

  1902.                                        int stride, int h, int x, int y);

  1903. void ff_avg_h264_chroma_mc8_mmx2_rnd  (uint8_t *dst, uint8_t *src,

  1904.                                        int stride, int h, int x, int y);

  1905. void ff_avg_rv40_chroma_mc8_mmx2      (uint8_t *dst, uint8_t *src,

  1906.                                        int stride, int h, int x, int y);

  1907. void ff_avg_h264_chroma_mc8_3dnow_rnd (uint8_t *dst, uint8_t *src,

  1908.                                        int stride, int h, int x, int y);

  1909. void ff_avg_rv40_chroma_mc8_3dnow     (uint8_t *dst, uint8_t *src,

  1910.                                        int stride, int h, int x, int y);

  1911. 

  1912. void ff_put_h264_chroma_mc4_mmx       (uint8_t *dst, uint8_t *src,

  1913.                                        int stride, int h, int x, int y);

  1914. void ff_put_rv40_chroma_mc4_mmx       (uint8_t *dst, uint8_t *src,

  1915.                                        int stride, int h, int x, int y);

  1916. void ff_avg_h264_chroma_mc4_mmx2      (uint8_t *dst, uint8_t *src,

  1917.                                        int stride, int h, int x, int y);

  1918. void ff_avg_rv40_chroma_mc4_mmx2      (uint8_t *dst, uint8_t *src,

  1919.                                        int stride, int h, int x, int y);

  1920. void ff_avg_h264_chroma_mc4_3dnow     (uint8_t *dst, uint8_t *src,

  1921.                                        int stride, int h, int x, int y);

  1922. void ff_avg_rv40_chroma_mc4_3dnow     (uint8_t *dst, uint8_t *src,

  1923.                                        int stride, int h, int x, int y);

  1924. 

  1925. void ff_put_h264_chroma_mc2_mmx2      (uint8_t *dst, uint8_t *src,

  1926.                                        int stride, int h, int x, int y);

  1927. void ff_avg_h264_chroma_mc2_mmx2      (uint8_t *dst, uint8_t *src,

  1928.                                        int stride, int h, int x, int y);

  1929. 

  1930. void ff_put_h264_chroma_mc8_ssse3_rnd (uint8_t *dst, uint8_t *src,

  1931.                                        int stride, int h, int x, int y);

  1932. void ff_put_h264_chroma_mc4_ssse3     (uint8_t *dst, uint8_t *src,

  1933.                                        int stride, int h, int x, int y);

  1934. 

  1935. void ff_avg_h264_chroma_mc8_ssse3_rnd (uint8_t *dst, uint8_t *src,

  1936.                                        int stride, int h, int x, int y);

  1937. void ff_avg_h264_chroma_mc4_ssse3     (uint8_t *dst, uint8_t *src,

  1938.                                        int stride, int h, int x, int y);

  1939. 

  1940. 

  1941. /* CAVS specific */

  1942. void ff_put_cavs_qpel8_mc00_mmx2(uint8_t *dst, uint8_t *src, int stride) {

  1943.     put_pixels8_mmx(dst, src, stride, 8);

  1944. }

  1945. void ff_avg_cavs_qpel8_mc00_mmx2(uint8_t *dst, uint8_t *src, int stride) {

  1946.     avg_pixels8_mmx(dst, src, stride, 8);

  1947. }

  1948. void ff_put_cavs_qpel16_mc00_mmx2(uint8_t *dst, uint8_t *src, int stride) {

  1949.     put_pixels16_mmx(dst, src, stride, 16);

  1950. }

  1951. void ff_avg_cavs_qpel16_mc00_mmx2(uint8_t *dst, uint8_t *src, int stride) {

  1952.     avg_pixels16_mmx(dst, src, stride, 16);

  1953. }

  1954. 

  1955. /* VC1 specific */

  1956. void ff_put_vc1_mspel_mc00_mmx(uint8_t *dst, const uint8_t *src, int stride, int rnd) {

  1957.     put_pixels8_mmx(dst, src, stride, 8);

  1958. }

  1959. void ff_avg_vc1_mspel_mc00_mmx2(uint8_t *dst, const uint8_t *src, int stride, int rnd) {

  1960.     avg_pixels8_mmx2(dst, src, stride, 8);

  1961. }

  1962. 

  1963. /* XXX: those functions should be suppressed ASAP when all IDCTs are

  1964.    converted */

  1965. #if CONFIG_GPL

  1966. static void ff_libmpeg2mmx_idct_put(uint8_t *dest, int line_size, DCTELEM *block)

  1967. {

  1968.     ff_mmx_idct (block);

  1969.     ff_put_pixels_clamped_mmx(block, dest, line_size);

  1970. }

  1971. static void ff_libmpeg2mmx_idct_add(uint8_t *dest, int line_size, DCTELEM *block)

  1972. {

  1973.     ff_mmx_idct (block);

  1974.     ff_add_pixels_clamped_mmx(block, dest, line_size);

  1975. }

  1976. static void ff_libmpeg2mmx2_idct_put(uint8_t *dest, int line_size, DCTELEM *block)

  1977. {

  1978.     ff_mmxext_idct (block);

  1979.     ff_put_pixels_clamped_mmx(block, dest, line_size);

  1980. }

  1981. static void ff_libmpeg2mmx2_idct_add(uint8_t *dest, int line_size, DCTELEM *block)

  1982. {

  1983.     ff_mmxext_idct (block);

  1984.     ff_add_pixels_clamped_mmx(block, dest, line_size);

  1985. }

  1986. #endif

  1987. static void ff_idct_xvid_mmx_put(uint8_t *dest, int line_size, DCTELEM *block)

  1988. {

  1989.     ff_idct_xvid_mmx (block);

  1990.     ff_put_pixels_clamped_mmx(block, dest, line_size);

  1991. }

  1992. static void ff_idct_xvid_mmx_add(uint8_t *dest, int line_size, DCTELEM *block)

  1993. {

  1994.     ff_idct_xvid_mmx (block);

  1995.     ff_add_pixels_clamped_mmx(block, dest, line_size);

  1996. }

  1997. static void ff_idct_xvid_mmx2_put(uint8_t *dest, int line_size, DCTELEM *block)

  1998. {

  1999.     ff_idct_xvid_mmx2 (block);

  2000.     ff_put_pixels_clamped_mmx(block, dest, line_size);

  2001. }

  2002. static void ff_idct_xvid_mmx2_add(uint8_t *dest, int line_size, DCTELEM *block)

  2003. {

  2004.     ff_idct_xvid_mmx2 (block);

  2005.     ff_add_pixels_clamped_mmx(block, dest, line_size);

  2006. }

  2007. 

  2008. static void vorbis_inverse_coupling_3dnow(float *mag, float *ang, int blocksize)

  2009. {

  2010.     int i;

  2011.     __asm__ volatile("pxor %%mm7, %%mm7":);

  2012.     for(i=0; i<blocksize; i+=2) {

  2013.         __asm__ volatile(

  2014.             "movq    %0,    %%mm0 \n\t"

  2015.             "movq    %1,    %%mm1 \n\t"

  2016.             "movq    %%mm0, %%mm2 \n\t"

  2017.             "movq    %%mm1, %%mm3 \n\t"

  2018.             "pfcmpge %%mm7, %%mm2 \n\t" // m <= 0.0

  2019.             "pfcmpge %%mm7, %%mm3 \n\t" // a <= 0.0

  2020.             "pslld   $31,   %%mm2 \n\t" // keep only the sign bit

  2021.             "pxor    %%mm2, %%mm1 \n\t"

  2022.             "movq    %%mm3, %%mm4 \n\t"

  2023.             "pand    %%mm1, %%mm3 \n\t"

  2024.             "pandn   %%mm1, %%mm4 \n\t"

  2025.             "pfadd   %%mm0, %%mm3 \n\t" // a = m + ((a<0) & (a ^ sign(m)))

  2026.             "pfsub   %%mm4, %%mm0 \n\t" // m = m + ((a>0) & (a ^ sign(m)))

  2027.             "movq    %%mm3, %1    \n\t"

  2028.             "movq    %%mm0, %0    \n\t"

  2029.             :"+m"(mag[i]), "+m"(ang[i])

  2030.             ::"memory"

  2031.         );

  2032.     }

  2033.     __asm__ volatile("femms");

  2034. }

  2035. static void vorbis_inverse_coupling_sse(float *mag, float *ang, int blocksize)

  2036. {

  2037.     int i;

  2038. 

  2039.     __asm__ volatile(

  2040.             "movaps  %0,     %%xmm5 \n\t"

  2041.         ::"m"(ff_pdw_80000000[0])

  2042.     );

  2043.     for(i=0; i<blocksize; i+=4) {

  2044.         __asm__ volatile(

  2045.             "movaps  %0,     %%xmm0 \n\t"

  2046.             "movaps  %1,     %%xmm1 \n\t"

  2047.             "xorps   %%xmm2, %%xmm2 \n\t"

  2048.             "xorps   %%xmm3, %%xmm3 \n\t"

  2049.             "cmpleps %%xmm0, %%xmm2 \n\t" // m <= 0.0

  2050.             "cmpleps %%xmm1, %%xmm3 \n\t" // a <= 0.0

  2051.             "andps   %%xmm5, %%xmm2 \n\t" // keep only the sign bit

  2052.             "xorps   %%xmm2, %%xmm1 \n\t"

  2053.             "movaps  %%xmm3, %%xmm4 \n\t"

  2054.             "andps   %%xmm1, %%xmm3 \n\t"

  2055.             "andnps  %%xmm1, %%xmm4 \n\t"

  2056.             "addps   %%xmm0, %%xmm3 \n\t" // a = m + ((a<0) & (a ^ sign(m)))

  2057.             "subps   %%xmm4, %%xmm0 \n\t" // m = m + ((a>0) & (a ^ sign(m)))

  2058.             "movaps  %%xmm3, %1     \n\t"

  2059.             "movaps  %%xmm0, %0     \n\t"

  2060.             :"+m"(mag[i]), "+m"(ang[i])

  2061.             ::"memory"

  2062.         );

  2063.     }

  2064. }

  2065. 

  2066. #define IF1(x) x

  2067. #define IF0(x)

  2068. 

  2069. #define MIX5(mono,stereo)\

  2070.     __asm__ volatile(\

  2071.         "movss          0(%2), %%xmm5 \n"\

  2072.         "movss          8(%2), %%xmm6 \n"\

  2073.         "movss         24(%2), %%xmm7 \n"\

  2074.         "shufps    $0, %%xmm5, %%xmm5 \n"\

  2075.         "shufps    $0, %%xmm6, %%xmm6 \n"\

  2076.         "shufps    $0, %%xmm7, %%xmm7 \n"\

  2077.         "1: \n"\

  2078.         "movaps       (%0,%1), %%xmm0 \n"\

  2079.         "movaps  0x400(%0,%1), %%xmm1 \n"\

  2080.         "movaps  0x800(%0,%1), %%xmm2 \n"\

  2081.         "movaps  0xc00(%0,%1), %%xmm3 \n"\

  2082.         "movaps 0x1000(%0,%1), %%xmm4 \n"\

  2083.         "mulps         %%xmm5, %%xmm0 \n"\

  2084.         "mulps         %%xmm6, %%xmm1 \n"\

  2085.         "mulps         %%xmm5, %%xmm2 \n"\

  2086.         "mulps         %%xmm7, %%xmm3 \n"\

  2087.         "mulps         %%xmm7, %%xmm4 \n"\

  2088.  stereo("addps         %%xmm1, %%xmm0 \n")\

  2089.         "addps         %%xmm1, %%xmm2 \n"\

  2090.         "addps         %%xmm3, %%xmm0 \n"\

  2091.         "addps         %%xmm4, %%xmm2 \n"\

  2092.    mono("addps         %%xmm2, %%xmm0 \n")\

  2093.         "movaps  %%xmm0,      (%0,%1) \n"\

  2094.  stereo("movaps  %%xmm2, 0x400(%0,%1) \n")\

  2095.         "add $16, %0 \n"\

  2096.         "jl 1b \n"\

  2097.         :"+&r"(i)\

  2098.         :"r"(samples[0]+len), "r"(matrix)\

  2099.         :XMM_CLOBBERS("%xmm0", "%xmm1", "%xmm2", "%xmm3", \

  2100.                       "%xmm4", "%xmm5", "%xmm6", "%xmm7",)\

  2101.          "memory"\

  2102.     );

  2103. 

  2104. #define MIX_MISC(stereo)\

  2105.     __asm__ volatile(\

  2106.         "1: \n"\

  2107.         "movaps  (%3,%0), %%xmm0 \n"\

  2108.  stereo("movaps   %%xmm0, %%xmm1 \n")\

  2109.         "mulps    %%xmm4, %%xmm0 \n"\

  2110.  stereo("mulps    %%xmm5, %%xmm1 \n")\

  2111.         "lea 1024(%3,%0), %1 \n"\

  2112.         "mov %5, %2 \n"\

  2113.         "2: \n"\

  2114.         "movaps   (%1),   %%xmm2 \n"\

  2115.  stereo("movaps   %%xmm2, %%xmm3 \n")\

  2116.         "mulps   (%4,%2), %%xmm2 \n"\

  2117.  stereo("mulps 16(%4,%2), %%xmm3 \n")\

  2118.         "addps    %%xmm2, %%xmm0 \n"\

  2119.  stereo("addps    %%xmm3, %%xmm1 \n")\

  2120.         "add $1024, %1 \n"\

  2121.         "add $32, %2 \n"\

  2122.         "jl 2b \n"\

  2123.         "movaps   %%xmm0,     (%3,%0) \n"\

  2124.  stereo("movaps   %%xmm1, 1024(%3,%0) \n")\

  2125.         "add $16, %0 \n"\

  2126.         "jl 1b \n"\

  2127.         :"+&r"(i), "=&r"(j), "=&r"(k)\

  2128.         :"r"(samples[0]+len), "r"(matrix_simd+in_ch), "g"((intptr_t)-32*(in_ch-1))\

  2129.         :"memory"\

  2130.     );

  2131. 

  2132. static void ac3_downmix_sse(float (*samples)[256], float (*matrix)[2], int out_ch, int in_ch, int len)

  2133. {

  2134.     int (*matrix_cmp)[2] = (int(*)[2])matrix;

  2135.     intptr_t i,j,k;

  2136. 

  2137.     i = -len*sizeof(float);

  2138.     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]))) {

  2139.         MIX5(IF0,IF1);

  2140.     } else if(in_ch == 5 && out_ch == 1 && matrix_cmp[0][0]==matrix_cmp[2][0] && matrix_cmp[3][0]==matrix_cmp[4][0]) {

  2141.         MIX5(IF1,IF0);

  2142.     } else {

  2143.         DECLARE_ALIGNED(16, float, matrix_simd)[AC3_MAX_CHANNELS][2][4];

  2144.         j = 2*in_ch*sizeof(float);

  2145.         __asm__ volatile(

  2146.             "1: \n"

  2147.             "sub $8, %0 \n"

  2148.             "movss     (%2,%0), %%xmm4 \n"

  2149.             "movss    4(%2,%0), %%xmm5 \n"

  2150.             "shufps $0, %%xmm4, %%xmm4 \n"

  2151.             "shufps $0, %%xmm5, %%xmm5 \n"

  2152.             "movaps %%xmm4,   (%1,%0,4) \n"

  2153.             "movaps %%xmm5, 16(%1,%0,4) \n"

  2154.             "jg 1b \n"

  2155.             :"+&r"(j)

  2156.             :"r"(matrix_simd), "r"(matrix)

  2157.             :"memory"

  2158.         );

  2159.         if(out_ch == 2) {

  2160.             MIX_MISC(IF1);

  2161.         } else {

  2162.             MIX_MISC(IF0);

  2163.         }

  2164.     }

  2165. }

  2166. 

  2167. static void vector_fmul_3dnow(float *dst, const float *src0, const float *src1, int len){

  2168.     x86_reg i = (len-4)*4;

  2169.     __asm__ volatile(

  2170.         "1: \n\t"

  2171.         "movq    (%2,%0), %%mm0 \n\t"

  2172.         "movq   8(%2,%0), %%mm1 \n\t"

  2173.         "pfmul   (%3,%0), %%mm0 \n\t"

  2174.         "pfmul  8(%3,%0), %%mm1 \n\t"

  2175.         "movq   %%mm0,  (%1,%0) \n\t"

  2176.         "movq   %%mm1, 8(%1,%0) \n\t"

  2177.         "sub  $16, %0 \n\t"

  2178.         "jge 1b \n\t"

  2179.         "femms  \n\t"

  2180.         :"+r"(i)

  2181.         :"r"(dst), "r"(src0), "r"(src1)

  2182.         :"memory"

  2183.     );

  2184. }

  2185. static void vector_fmul_sse(float *dst, const float *src0, const float *src1, int len){

  2186.     x86_reg i = (len-8)*4;

  2187.     __asm__ volatile(

  2188.         "1: \n\t"

  2189.         "movaps    (%2,%0), %%xmm0 \n\t"

  2190.         "movaps  16(%2,%0), %%xmm1 \n\t"

  2191.         "mulps     (%3,%0), %%xmm0 \n\t"

  2192.         "mulps   16(%3,%0), %%xmm1 \n\t"

  2193.         "movaps  %%xmm0,   (%1,%0) \n\t"

  2194.         "movaps  %%xmm1, 16(%1,%0) \n\t"

  2195.         "sub  $32, %0 \n\t"

  2196.         "jge 1b \n\t"

  2197.         :"+r"(i)

  2198.         :"r"(dst), "r"(src0), "r"(src1)

  2199.         :"memory"

  2200.     );

  2201. }

  2202. 

  2203. static void vector_fmul_reverse_3dnow2(float *dst, const float *src0, const float *src1, int len){

  2204.     x86_reg i = len*4-16;

  2205.     __asm__ volatile(

  2206.         "1: \n\t"

  2207.         "pswapd   8(%1), %%mm0 \n\t"

  2208.         "pswapd    (%1), %%mm1 \n\t"

  2209.         "pfmul  (%3,%0), %%mm0 \n\t"

  2210.         "pfmul 8(%3,%0), %%mm1 \n\t"

  2211.         "movq  %%mm0,  (%2,%0) \n\t"

  2212.         "movq  %%mm1, 8(%2,%0) \n\t"

  2213.         "add   $16, %1 \n\t"

  2214.         "sub   $16, %0 \n\t"

  2215.         "jge   1b \n\t"

  2216.         :"+r"(i), "+r"(src1)

  2217.         :"r"(dst), "r"(src0)

  2218.     );

  2219.     __asm__ volatile("femms");

  2220. }

  2221. static void vector_fmul_reverse_sse(float *dst, const float *src0, const float *src1, int len){

  2222.     x86_reg i = len*4-32;

  2223.     __asm__ volatile(

  2224.         "1: \n\t"

  2225.         "movaps        16(%1), %%xmm0 \n\t"

  2226.         "movaps          (%1), %%xmm1 \n\t"

  2227.         "shufps $0x1b, %%xmm0, %%xmm0 \n\t"

  2228.         "shufps $0x1b, %%xmm1, %%xmm1 \n\t"

  2229.         "mulps        (%3,%0), %%xmm0 \n\t"

  2230.         "mulps      16(%3,%0), %%xmm1 \n\t"

  2231.         "movaps     %%xmm0,   (%2,%0) \n\t"

  2232.         "movaps     %%xmm1, 16(%2,%0) \n\t"

  2233.         "add    $32, %1 \n\t"

  2234.         "sub    $32, %0 \n\t"

  2235.         "jge    1b \n\t"

  2236.         :"+r"(i), "+r"(src1)

  2237.         :"r"(dst), "r"(src0)

  2238.     );

  2239. }

  2240. 

  2241. static void vector_fmul_add_3dnow(float *dst, const float *src0, const float *src1,

  2242.                                   const float *src2, int len){

  2243.     x86_reg i = (len-4)*4;

  2244.     __asm__ volatile(

  2245.         "1: \n\t"

  2246.         "movq    (%2,%0), %%mm0 \n\t"

  2247.         "movq   8(%2,%0), %%mm1 \n\t"

  2248.         "pfmul   (%3,%0), %%mm0 \n\t"

  2249.         "pfmul  8(%3,%0), %%mm1 \n\t"

  2250.         "pfadd   (%4,%0), %%mm0 \n\t"

  2251.         "pfadd  8(%4,%0), %%mm1 \n\t"

  2252.         "movq  %%mm0,   (%1,%0) \n\t"

  2253.         "movq  %%mm1,  8(%1,%0) \n\t"

  2254.         "sub  $16, %0 \n\t"

  2255.         "jge  1b \n\t"

  2256.         :"+r"(i)

  2257.         :"r"(dst), "r"(src0), "r"(src1), "r"(src2)

  2258.         :"memory"

  2259.     );

  2260.     __asm__ volatile("femms");

  2261. }

  2262. static void vector_fmul_add_sse(float *dst, const float *src0, const float *src1,

  2263.                                 const float *src2, int len){

  2264.     x86_reg i = (len-8)*4;

  2265.     __asm__ volatile(

  2266.         "1: \n\t"

  2267.         "movaps   (%2,%0), %%xmm0 \n\t"

  2268.         "movaps 16(%2,%0), %%xmm1 \n\t"

  2269.         "mulps    (%3,%0), %%xmm0 \n\t"

  2270.         "mulps  16(%3,%0), %%xmm1 \n\t"

  2271.         "addps    (%4,%0), %%xmm0 \n\t"

  2272.         "addps  16(%4,%0), %%xmm1 \n\t"

  2273.         "movaps %%xmm0,   (%1,%0) \n\t"

  2274.         "movaps %%xmm1, 16(%1,%0) \n\t"

  2275.         "sub  $32, %0 \n\t"

  2276.         "jge  1b \n\t"

  2277.         :"+r"(i)

  2278.         :"r"(dst), "r"(src0), "r"(src1), "r"(src2)

  2279.         :"memory"

  2280.     );

  2281. }

  2282. 

  2283. #if HAVE_6REGS

  2284. static void vector_fmul_window_3dnow2(float *dst, const float *src0, const float *src1,

  2285.                                       const float *win, int len){

  2286.     x86_reg i = -len*4;

  2287.     x86_reg j = len*4-8;

  2288.     __asm__ volatile(

  2289.         "1: \n"

  2290.         "pswapd  (%5,%1), %%mm1 \n"

  2291.         "movq    (%5,%0), %%mm0 \n"

  2292.         "pswapd  (%4,%1), %%mm5 \n"

  2293.         "movq    (%3,%0), %%mm4 \n"

  2294.         "movq      %%mm0, %%mm2 \n"

  2295.         "movq      %%mm1, %%mm3 \n"

  2296.         "pfmul     %%mm4, %%mm2 \n" // src0[len+i]*win[len+i]

  2297.         "pfmul     %%mm5, %%mm3 \n" // src1[    j]*win[len+j]

  2298.         "pfmul     %%mm4, %%mm1 \n" // src0[len+i]*win[len+j]

  2299.         "pfmul     %%mm5, %%mm0 \n" // src1[    j]*win[len+i]

  2300.         "pfadd     %%mm3, %%mm2 \n"

  2301.         "pfsub     %%mm0, %%mm1 \n"

  2302.         "pswapd    %%mm2, %%mm2 \n"

  2303.         "movq      %%mm1, (%2,%0) \n"

  2304.         "movq      %%mm2, (%2,%1) \n"

  2305.         "sub $8, %1 \n"

  2306.         "add $8, %0 \n"

  2307.         "jl 1b \n"

  2308.         "femms \n"

  2309.         :"+r"(i), "+r"(j)

  2310.         :"r"(dst+len), "r"(src0+len), "r"(src1), "r"(win+len)

  2311.     );

  2312. }

  2313. 

  2314. static void vector_fmul_window_sse(float *dst, const float *src0, const float *src1,

  2315.                                    const float *win, int len){

  2316.     x86_reg i = -len*4;

  2317.     x86_reg j = len*4-16;

  2318.     __asm__ volatile(

  2319.         "1: \n"

  2320.         "movaps       (%5,%1), %%xmm1 \n"

  2321.         "movaps       (%5,%0), %%xmm0 \n"

  2322.         "movaps       (%4,%1), %%xmm5 \n"

  2323.         "movaps       (%3,%0), %%xmm4 \n"

  2324.         "shufps $0x1b, %%xmm1, %%xmm1 \n"

  2325.         "shufps $0x1b, %%xmm5, %%xmm5 \n"

  2326.         "movaps        %%xmm0, %%xmm2 \n"

  2327.         "movaps        %%xmm1, %%xmm3 \n"

  2328.         "mulps         %%xmm4, %%xmm2 \n" // src0[len+i]*win[len+i]

  2329.         "mulps         %%xmm5, %%xmm3 \n" // src1[    j]*win[len+j]

  2330.         "mulps         %%xmm4, %%xmm1 \n" // src0[len+i]*win[len+j]

  2331.         "mulps         %%xmm5, %%xmm0 \n" // src1[    j]*win[len+i]

  2332.         "addps         %%xmm3, %%xmm2 \n"

  2333.         "subps         %%xmm0, %%xmm1 \n"

  2334.         "shufps $0x1b, %%xmm2, %%xmm2 \n"

  2335.         "movaps        %%xmm1, (%2,%0) \n"

  2336.         "movaps        %%xmm2, (%2,%1) \n"

  2337.         "sub $16, %1 \n"

  2338.         "add $16, %0 \n"

  2339.         "jl 1b \n"

  2340.         :"+r"(i), "+r"(j)

  2341.         :"r"(dst+len), "r"(src0+len), "r"(src1), "r"(win+len)

  2342.     );

  2343. }

  2344. #endif /* HAVE_6REGS */

  2345. 

  2346. static void vector_clipf_sse(float *dst, const float *src, float min, float max,

  2347.                              int len)

  2348. {

  2349.     x86_reg i = (len-16)*4;

  2350.     __asm__ volatile(

  2351.         "movss  %3, %%xmm4 \n"

  2352.         "movss  %4, %%xmm5 \n"

  2353.         "shufps $0, %%xmm4, %%xmm4 \n"

  2354.         "shufps $0, %%xmm5, %%xmm5 \n"

  2355.         "1: \n\t"

  2356.         "movaps    (%2,%0), %%xmm0 \n\t" // 3/1 on intel

  2357.         "movaps  16(%2,%0), %%xmm1 \n\t"

  2358.         "movaps  32(%2,%0), %%xmm2 \n\t"

  2359.         "movaps  48(%2,%0), %%xmm3 \n\t"

  2360.         "maxps      %%xmm4, %%xmm0 \n\t"

  2361.         "maxps      %%xmm4, %%xmm1 \n\t"

  2362.         "maxps      %%xmm4, %%xmm2 \n\t"

  2363.         "maxps      %%xmm4, %%xmm3 \n\t"

  2364.         "minps      %%xmm5, %%xmm0 \n\t"

  2365.         "minps      %%xmm5, %%xmm1 \n\t"

  2366.         "minps      %%xmm5, %%xmm2 \n\t"

  2367.         "minps      %%xmm5, %%xmm3 \n\t"

  2368.         "movaps  %%xmm0,   (%1,%0) \n\t"

  2369.         "movaps  %%xmm1, 16(%1,%0) \n\t"

  2370.         "movaps  %%xmm2, 32(%1,%0) \n\t"

  2371.         "movaps  %%xmm3, 48(%1,%0) \n\t"

  2372.         "sub  $64, %0 \n\t"

  2373.         "jge 1b \n\t"

  2374.         :"+&r"(i)

  2375.         :"r"(dst), "r"(src), "m"(min), "m"(max)

  2376.         :"memory"

  2377.     );

  2378. }

  2379. 

  2380. void ff_vp3_idct_mmx(int16_t *input_data);

  2381. void ff_vp3_idct_put_mmx(uint8_t *dest, int line_size, DCTELEM *block);

  2382. void ff_vp3_idct_add_mmx(uint8_t *dest, int line_size, DCTELEM *block);

  2383. 

  2384. void ff_vp3_idct_dc_add_mmx2(uint8_t *dest, int line_size, const DCTELEM *block);

  2385. 

  2386. void ff_vp3_v_loop_filter_mmx2(uint8_t *src, int stride, int *bounding_values);

  2387. void ff_vp3_h_loop_filter_mmx2(uint8_t *src, int stride, int *bounding_values);

  2388. 

  2389. void ff_vp3_idct_sse2(int16_t *input_data);

  2390. void ff_vp3_idct_put_sse2(uint8_t *dest, int line_size, DCTELEM *block);

  2391. void ff_vp3_idct_add_sse2(uint8_t *dest, int line_size, DCTELEM *block);

  2392. 

  2393. int32_t ff_scalarproduct_int16_mmx2(const int16_t *v1, const int16_t *v2, int order, int shift);

  2394. int32_t ff_scalarproduct_int16_sse2(const int16_t *v1, const int16_t *v2, int order, int shift);

  2395. int32_t ff_scalarproduct_and_madd_int16_mmx2(int16_t *v1, const int16_t *v2, const int16_t *v3, int order, int mul);

  2396. int32_t ff_scalarproduct_and_madd_int16_sse2(int16_t *v1, const int16_t *v2, const int16_t *v3, int order, int mul);

  2397. int32_t ff_scalarproduct_and_madd_int16_ssse3(int16_t *v1, const int16_t *v2, const int16_t *v3, int order, int mul);

  2398. 

  2399. void ff_apply_window_int16_mmxext    (int16_t *output, const int16_t *input,

  2400.                                       const int16_t *window, unsigned int len);

  2401. void ff_apply_window_int16_mmxext_ba (int16_t *output, const int16_t *input,

  2402.                                       const int16_t *window, unsigned int len);

  2403. void ff_apply_window_int16_sse2      (int16_t *output, const int16_t *input,

  2404.                                       const int16_t *window, unsigned int len);

  2405. void ff_apply_window_int16_sse2_ba   (int16_t *output, const int16_t *input,

  2406.                                       const int16_t *window, unsigned int len);

  2407. void ff_apply_window_int16_ssse3     (int16_t *output, const int16_t *input,

  2408.                                       const int16_t *window, unsigned int len);

  2409. void ff_apply_window_int16_ssse3_atom(int16_t *output, const int16_t *input,

  2410.                                       const int16_t *window, unsigned int len);

  2411. 

  2412. 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);

  2413. int  ff_add_hfyu_left_prediction_ssse3(uint8_t *dst, const uint8_t *src, int w, int left);

  2414. int  ff_add_hfyu_left_prediction_sse4(uint8_t *dst, const uint8_t *src, int w, int left);

  2415. 

  2416. float ff_scalarproduct_float_sse(const float *v1, const float *v2, int order);

  2417. 

  2418. void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx)

  2419. {

  2420.     int mm_flags = av_get_cpu_flags();

  2421. 

  2422.     if (avctx->dsp_mask) {

  2423.         if (avctx->dsp_mask & AV_CPU_FLAG_FORCE)

  2424.             mm_flags |= (avctx->dsp_mask & 0xffff);

  2425.         else

  2426.             mm_flags &= ~(avctx->dsp_mask & 0xffff);

  2427.     }

  2428. 

  2429. #if 0

  2430.     av_log(avctx, AV_LOG_INFO, "libavcodec: CPU flags:");

  2431.     if (mm_flags & AV_CPU_FLAG_MMX)

  2432.         av_log(avctx, AV_LOG_INFO, " mmx");

  2433.     if (mm_flags & AV_CPU_FLAG_MMX2)

  2434.         av_log(avctx, AV_LOG_INFO, " mmx2");

  2435.     if (mm_flags & AV_CPU_FLAG_3DNOW)

  2436.         av_log(avctx, AV_LOG_INFO, " 3dnow");

  2437.     if (mm_flags & AV_CPU_FLAG_SSE)

  2438.         av_log(avctx, AV_LOG_INFO, " sse");

  2439.     if (mm_flags & AV_CPU_FLAG_SSE2)

  2440.         av_log(avctx, AV_LOG_INFO, " sse2");

  2441.     av_log(avctx, AV_LOG_INFO, "\n");

  2442. #endif

  2443. 

  2444.     if (mm_flags & AV_CPU_FLAG_MMX) {

  2445.         const int idct_algo= avctx->idct_algo;

  2446. 

  2447.         if(avctx->lowres==0){

  2448.             if(idct_algo==FF_IDCT_AUTO || idct_algo==FF_IDCT_SIMPLEMMX){

  2449.                 c->idct_put= ff_simple_idct_put_mmx;

  2450.                 c->idct_add= ff_simple_idct_add_mmx;

  2451.                 c->idct    = ff_simple_idct_mmx;

  2452.                 c->idct_permutation_type= FF_SIMPLE_IDCT_PERM;

  2453. #if CONFIG_GPL

  2454.             }else if(idct_algo==FF_IDCT_LIBMPEG2MMX){

  2455.                 if(mm_flags & AV_CPU_FLAG_MMX2){

  2456.                     c->idct_put= ff_libmpeg2mmx2_idct_put;

  2457.                     c->idct_add= ff_libmpeg2mmx2_idct_add;

  2458.                     c->idct    = ff_mmxext_idct;

  2459.                 }else{

  2460.                     c->idct_put= ff_libmpeg2mmx_idct_put;

  2461.                     c->idct_add= ff_libmpeg2mmx_idct_add;

  2462.                     c->idct    = ff_mmx_idct;

  2463.                 }

  2464.                 c->idct_permutation_type= FF_LIBMPEG2_IDCT_PERM;

  2465. #endif

  2466.             }else if((CONFIG_VP3_DECODER || CONFIG_VP5_DECODER || CONFIG_VP6_DECODER) &&

  2467.                      idct_algo==FF_IDCT_VP3 && HAVE_YASM){

  2468.                 if(mm_flags & AV_CPU_FLAG_SSE2){

  2469.                     c->idct_put= ff_vp3_idct_put_sse2;

  2470.                     c->idct_add= ff_vp3_idct_add_sse2;

  2471.                     c->idct    = ff_vp3_idct_sse2;

  2472.                     c->idct_permutation_type= FF_TRANSPOSE_IDCT_PERM;

  2473.                 }else{

  2474.                     c->idct_put= ff_vp3_idct_put_mmx;

  2475.                     c->idct_add= ff_vp3_idct_add_mmx;

  2476.                     c->idct    = ff_vp3_idct_mmx;

  2477.                     c->idct_permutation_type= FF_PARTTRANS_IDCT_PERM;

  2478.                 }

  2479.             }else if(idct_algo==FF_IDCT_CAVS){

  2480.                     c->idct_permutation_type= FF_TRANSPOSE_IDCT_PERM;

  2481.             }else if(idct_algo==FF_IDCT_XVIDMMX){

  2482.                 if(mm_flags & AV_CPU_FLAG_SSE2){

  2483.                     c->idct_put= ff_idct_xvid_sse2_put;

  2484.                     c->idct_add= ff_idct_xvid_sse2_add;

  2485.                     c->idct    = ff_idct_xvid_sse2;

  2486.                     c->idct_permutation_type= FF_SSE2_IDCT_PERM;

  2487.                 }else if(mm_flags & AV_CPU_FLAG_MMX2){

  2488.                     c->idct_put= ff_idct_xvid_mmx2_put;

  2489.                     c->idct_add= ff_idct_xvid_mmx2_add;

  2490.                     c->idct    = ff_idct_xvid_mmx2;

  2491.                 }else{

  2492.                     c->idct_put= ff_idct_xvid_mmx_put;

  2493.                     c->idct_add= ff_idct_xvid_mmx_add;

  2494.                     c->idct    = ff_idct_xvid_mmx;

  2495.                 }

  2496.             }

  2497.         }

  2498. 

  2499.         c->put_pixels_clamped = ff_put_pixels_clamped_mmx;

  2500.         c->put_signed_pixels_clamped = ff_put_signed_pixels_clamped_mmx;

  2501.         c->add_pixels_clamped = ff_add_pixels_clamped_mmx;

  2502.         c->clear_block  = clear_block_mmx;

  2503.         c->clear_blocks = clear_blocks_mmx;

  2504.         if ((mm_flags & AV_CPU_FLAG_SSE) &&

  2505.             !(CONFIG_MPEG_XVMC_DECODER && avctx->xvmc_acceleration > 1)){

  2506.             /* XvMCCreateBlocks() may not allocate 16-byte aligned blocks */

  2507.             c->clear_block  = clear_block_sse;

  2508.             c->clear_blocks = clear_blocks_sse;

  2509.         }

  2510. 

  2511. #define SET_HPEL_FUNCS(PFX, IDX, SIZE, CPU) \

  2512.         c->PFX ## _pixels_tab[IDX][0] = PFX ## _pixels ## SIZE ## _ ## CPU; \

  2513.         c->PFX ## _pixels_tab[IDX][1] = PFX ## _pixels ## SIZE ## _x2_ ## CPU; \

  2514.         c->PFX ## _pixels_tab[IDX][2] = PFX ## _pixels ## SIZE ## _y2_ ## CPU; \

  2515.         c->PFX ## _pixels_tab[IDX][3] = PFX ## _pixels ## SIZE ## _xy2_ ## CPU

  2516. 

  2517.         SET_HPEL_FUNCS(put, 0, 16, mmx);

  2518.         SET_HPEL_FUNCS(put_no_rnd, 0, 16, mmx);

  2519.         SET_HPEL_FUNCS(avg, 0, 16, mmx);

  2520.         SET_HPEL_FUNCS(avg_no_rnd, 0, 16, mmx);

  2521.         SET_HPEL_FUNCS(put, 1, 8, mmx);

  2522.         SET_HPEL_FUNCS(put_no_rnd, 1, 8, mmx);

  2523.         SET_HPEL_FUNCS(avg, 1, 8, mmx);

  2524.         SET_HPEL_FUNCS(avg_no_rnd, 1, 8, mmx);

  2525. 

  2526. #if ARCH_X86_32 || !HAVE_YASM

  2527.         c->gmc= gmc_mmx;

  2528. #endif

  2529. #if ARCH_X86_32 && HAVE_YASM

  2530.         c->emulated_edge_mc = emulated_edge_mc_mmx;

  2531. #endif

  2532. 

  2533.         c->add_bytes= add_bytes_mmx;

  2534.         c->add_bytes_l2= add_bytes_l2_mmx;

  2535. 

  2536.         c->draw_edges = draw_edges_mmx;

  2537. 

  2538.         if (CONFIG_H263_DECODER || CONFIG_H263_ENCODER) {

  2539.             c->h263_v_loop_filter= h263_v_loop_filter_mmx;

  2540.             c->h263_h_loop_filter= h263_h_loop_filter_mmx;

  2541.         }

  2542. 

  2543. #if HAVE_YASM

  2544.         c->put_h264_chroma_pixels_tab[0]= ff_put_h264_chroma_mc8_mmx_rnd;

  2545.         c->put_h264_chroma_pixels_tab[1]= ff_put_h264_chroma_mc4_mmx;

  2546. 

  2547.         c->put_rv40_chroma_pixels_tab[0]= ff_put_rv40_chroma_mc8_mmx;

  2548.         c->put_rv40_chroma_pixels_tab[1]= ff_put_rv40_chroma_mc4_mmx;

  2549. #endif

  2550. 

  2551.         if (mm_flags & AV_CPU_FLAG_MMX2) {

  2552.             c->prefetch = prefetch_mmx2;

  2553. 

  2554.             c->put_pixels_tab[0][1] = put_pixels16_x2_mmx2;

  2555.             c->put_pixels_tab[0][2] = put_pixels16_y2_mmx2;

  2556. 

  2557.             c->avg_pixels_tab[0][0] = avg_pixels16_mmx2;

  2558.             c->avg_pixels_tab[0][1] = avg_pixels16_x2_mmx2;

  2559.             c->avg_pixels_tab[0][2] = avg_pixels16_y2_mmx2;

  2560. 

  2561.             c->put_pixels_tab[1][1] = put_pixels8_x2_mmx2;

  2562.             c->put_pixels_tab[1][2] = put_pixels8_y2_mmx2;

  2563. 

  2564.             c->avg_pixels_tab[1][0] = avg_pixels8_mmx2;

  2565.             c->avg_pixels_tab[1][1] = avg_pixels8_x2_mmx2;

  2566.             c->avg_pixels_tab[1][2] = avg_pixels8_y2_mmx2;

  2567. 

  2568.             if(!(avctx->flags & CODEC_FLAG_BITEXACT)){

  2569.                 c->put_no_rnd_pixels_tab[0][1] = put_no_rnd_pixels16_x2_mmx2;

  2570.                 c->put_no_rnd_pixels_tab[0][2] = put_no_rnd_pixels16_y2_mmx2;

  2571.                 c->put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_mmx2;

  2572.                 c->put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_mmx2;

  2573.                 c->avg_pixels_tab[0][3] = avg_pixels16_xy2_mmx2;

  2574.                 c->avg_pixels_tab[1][3] = avg_pixels8_xy2_mmx2;

  2575. 

  2576.                 if (CONFIG_VP3_DECODER && HAVE_YASM) {

  2577.                     c->vp3_v_loop_filter= ff_vp3_v_loop_filter_mmx2;

  2578.                     c->vp3_h_loop_filter= ff_vp3_h_loop_filter_mmx2;

  2579.                 }

  2580.             }

  2581.             if (CONFIG_VP3_DECODER && HAVE_YASM) {

  2582.                 c->vp3_idct_dc_add = ff_vp3_idct_dc_add_mmx2;

  2583.             }

  2584. 

  2585.             if (CONFIG_VP3_DECODER

  2586.                 && (avctx->codec_id == CODEC_ID_VP3 || avctx->codec_id == CODEC_ID_THEORA)) {

  2587.                 c->put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_exact_mmx2;

  2588.                 c->put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_exact_mmx2;

  2589.             }

  2590. 

  2591. #define SET_QPEL_FUNCS(PFX, IDX, SIZE, CPU) \

  2592.             c->PFX ## _pixels_tab[IDX][ 0] = PFX ## SIZE ## _mc00_ ## CPU; \

  2593.             c->PFX ## _pixels_tab[IDX][ 1] = PFX ## SIZE ## _mc10_ ## CPU; \

  2594.             c->PFX ## _pixels_tab[IDX][ 2] = PFX ## SIZE ## _mc20_ ## CPU; \

  2595.             c->PFX ## _pixels_tab[IDX][ 3] = PFX ## SIZE ## _mc30_ ## CPU; \

  2596.             c->PFX ## _pixels_tab[IDX][ 4] = PFX ## SIZE ## _mc01_ ## CPU; \

  2597.             c->PFX ## _pixels_tab[IDX][ 5] = PFX ## SIZE ## _mc11_ ## CPU; \

  2598.             c->PFX ## _pixels_tab[IDX][ 6] = PFX ## SIZE ## _mc21_ ## CPU; \

  2599.             c->PFX ## _pixels_tab[IDX][ 7] = PFX ## SIZE ## _mc31_ ## CPU; \

  2600.             c->PFX ## _pixels_tab[IDX][ 8] = PFX ## SIZE ## _mc02_ ## CPU; \

  2601.             c->PFX ## _pixels_tab[IDX][ 9] = PFX ## SIZE ## _mc12_ ## CPU; \

  2602.             c->PFX ## _pixels_tab[IDX][10] = PFX ## SIZE ## _mc22_ ## CPU; \

  2603.             c->PFX ## _pixels_tab[IDX][11] = PFX ## SIZE ## _mc32_ ## CPU; \

  2604.             c->PFX ## _pixels_tab[IDX][12] = PFX ## SIZE ## _mc03_ ## CPU; \

  2605.             c->PFX ## _pixels_tab[IDX][13] = PFX ## SIZE ## _mc13_ ## CPU; \

  2606.             c->PFX ## _pixels_tab[IDX][14] = PFX ## SIZE ## _mc23_ ## CPU; \

  2607.             c->PFX ## _pixels_tab[IDX][15] = PFX ## SIZE ## _mc33_ ## CPU

  2608. 

  2609.             SET_QPEL_FUNCS(put_qpel, 0, 16, mmx2);

  2610.             SET_QPEL_FUNCS(put_qpel, 1, 8, mmx2);

  2611.             SET_QPEL_FUNCS(put_no_rnd_qpel, 0, 16, mmx2);

  2612.             SET_QPEL_FUNCS(put_no_rnd_qpel, 1, 8, mmx2);

  2613.             SET_QPEL_FUNCS(avg_qpel, 0, 16, mmx2);

  2614.             SET_QPEL_FUNCS(avg_qpel, 1, 8, mmx2);

  2615. 

  2616.             SET_QPEL_FUNCS(put_h264_qpel, 0, 16, mmx2);

  2617.             SET_QPEL_FUNCS(put_h264_qpel, 1, 8, mmx2);

  2618.             SET_QPEL_FUNCS(put_h264_qpel, 2, 4, mmx2);

  2619.             SET_QPEL_FUNCS(avg_h264_qpel, 0, 16, mmx2);

  2620.             SET_QPEL_FUNCS(avg_h264_qpel, 1, 8, mmx2);

  2621.             SET_QPEL_FUNCS(avg_h264_qpel, 2, 4, mmx2);

  2622. 

  2623.             SET_QPEL_FUNCS(put_2tap_qpel, 0, 16, mmx2);

  2624.             SET_QPEL_FUNCS(put_2tap_qpel, 1, 8, mmx2);

  2625.             SET_QPEL_FUNCS(avg_2tap_qpel, 0, 16, mmx2);

  2626.             SET_QPEL_FUNCS(avg_2tap_qpel, 1, 8, mmx2);

  2627. 

  2628. #if HAVE_YASM

  2629.             c->avg_rv40_chroma_pixels_tab[0]= ff_avg_rv40_chroma_mc8_mmx2;

  2630.             c->avg_rv40_chroma_pixels_tab[1]= ff_avg_rv40_chroma_mc4_mmx2;

  2631. 

  2632.             c->avg_h264_chroma_pixels_tab[0]= ff_avg_h264_chroma_mc8_mmx2_rnd;

  2633.             c->avg_h264_chroma_pixels_tab[1]= ff_avg_h264_chroma_mc4_mmx2;

  2634.             c->avg_h264_chroma_pixels_tab[2]= ff_avg_h264_chroma_mc2_mmx2;

  2635.             c->put_h264_chroma_pixels_tab[2]= ff_put_h264_chroma_mc2_mmx2;

  2636. 

  2637.             c->add_hfyu_median_prediction = ff_add_hfyu_median_prediction_mmx2;

  2638. #endif

  2639. #if HAVE_7REGS && HAVE_TEN_OPERANDS

  2640.             if( mm_flags&AV_CPU_FLAG_3DNOW )

  2641.                 c->add_hfyu_median_prediction = add_hfyu_median_prediction_cmov;

  2642. #endif

  2643. 

  2644.             c->add_png_paeth_prediction= add_png_paeth_prediction_mmx2;

  2645.         } else if (mm_flags & AV_CPU_FLAG_3DNOW) {

  2646.             c->prefetch = prefetch_3dnow;

  2647. 

  2648.             c->put_pixels_tab[0][1] = put_pixels16_x2_3dnow;

  2649.             c->put_pixels_tab[0][2] = put_pixels16_y2_3dnow;

  2650. 

  2651.             c->avg_pixels_tab[0][0] = avg_pixels16_3dnow;

  2652.             c->avg_pixels_tab[0][1] = avg_pixels16_x2_3dnow;

  2653.             c->avg_pixels_tab[0][2] = avg_pixels16_y2_3dnow;

  2654. 

  2655.             c->put_pixels_tab[1][1] = put_pixels8_x2_3dnow;

  2656.             c->put_pixels_tab[1][2] = put_pixels8_y2_3dnow;

  2657. 

  2658.             c->avg_pixels_tab[1][0] = avg_pixels8_3dnow;

  2659.             c->avg_pixels_tab[1][1] = avg_pixels8_x2_3dnow;

  2660.             c->avg_pixels_tab[1][2] = avg_pixels8_y2_3dnow;

  2661. 

  2662.             if(!(avctx->flags & CODEC_FLAG_BITEXACT)){

  2663.                 c->put_no_rnd_pixels_tab[0][1] = put_no_rnd_pixels16_x2_3dnow;

  2664.                 c->put_no_rnd_pixels_tab[0][2] = put_no_rnd_pixels16_y2_3dnow;

  2665.                 c->put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_3dnow;

  2666.                 c->put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_3dnow;

  2667.                 c->avg_pixels_tab[0][3] = avg_pixels16_xy2_3dnow;

  2668.                 c->avg_pixels_tab[1][3] = avg_pixels8_xy2_3dnow;

  2669.             }

  2670. 

  2671.             if (CONFIG_VP3_DECODER

  2672.                 && (avctx->codec_id == CODEC_ID_VP3 || avctx->codec_id == CODEC_ID_THEORA)) {

  2673.                 c->put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_exact_3dnow;

  2674.                 c->put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_exact_3dnow;

  2675.             }

  2676. 

  2677.             SET_QPEL_FUNCS(put_qpel, 0, 16, 3dnow);

  2678.             SET_QPEL_FUNCS(put_qpel, 1, 8, 3dnow);

  2679.             SET_QPEL_FUNCS(put_no_rnd_qpel, 0, 16, 3dnow);

  2680.             SET_QPEL_FUNCS(put_no_rnd_qpel, 1, 8, 3dnow);

  2681.             SET_QPEL_FUNCS(avg_qpel, 0, 16, 3dnow);

  2682.             SET_QPEL_FUNCS(avg_qpel, 1, 8, 3dnow);

  2683. 

  2684.             SET_QPEL_FUNCS(put_h264_qpel, 0, 16, 3dnow);

  2685.             SET_QPEL_FUNCS(put_h264_qpel, 1, 8, 3dnow);

  2686.             SET_QPEL_FUNCS(put_h264_qpel, 2, 4, 3dnow);

  2687.             SET_QPEL_FUNCS(avg_h264_qpel, 0, 16, 3dnow);

  2688.             SET_QPEL_FUNCS(avg_h264_qpel, 1, 8, 3dnow);

  2689.             SET_QPEL_FUNCS(avg_h264_qpel, 2, 4, 3dnow);

  2690. 

  2691.             SET_QPEL_FUNCS(put_2tap_qpel, 0, 16, 3dnow);

  2692.             SET_QPEL_FUNCS(put_2tap_qpel, 1, 8, 3dnow);

  2693.             SET_QPEL_FUNCS(avg_2tap_qpel, 0, 16, 3dnow);

  2694.             SET_QPEL_FUNCS(avg_2tap_qpel, 1, 8, 3dnow);

  2695. 

  2696. #if HAVE_YASM

  2697.             c->avg_h264_chroma_pixels_tab[0]= ff_avg_h264_chroma_mc8_3dnow_rnd;

  2698.             c->avg_h264_chroma_pixels_tab[1]= ff_avg_h264_chroma_mc4_3dnow;

  2699. 

  2700.             c->avg_rv40_chroma_pixels_tab[0]= ff_avg_rv40_chroma_mc8_3dnow;

  2701.             c->avg_rv40_chroma_pixels_tab[1]= ff_avg_rv40_chroma_mc4_3dnow;

  2702. #endif

  2703.         }

  2704. 

  2705. 

  2706. #define H264_QPEL_FUNCS(x, y, CPU)\

  2707.             c->put_h264_qpel_pixels_tab[0][x+y*4] = put_h264_qpel16_mc##x##y##_##CPU;\

  2708.             c->put_h264_qpel_pixels_tab[1][x+y*4] = put_h264_qpel8_mc##x##y##_##CPU;\

  2709.             c->avg_h264_qpel_pixels_tab[0][x+y*4] = avg_h264_qpel16_mc##x##y##_##CPU;\

  2710.             c->avg_h264_qpel_pixels_tab[1][x+y*4] = avg_h264_qpel8_mc##x##y##_##CPU;

  2711.         if((mm_flags & AV_CPU_FLAG_SSE2) && !(mm_flags & AV_CPU_FLAG_3DNOW)){

  2712.             // these functions are slower than mmx on AMD, but faster on Intel

  2713.             c->put_pixels_tab[0][0] = put_pixels16_sse2;

  2714.             c->put_no_rnd_pixels_tab[0][0] = put_pixels16_sse2;

  2715.             c->avg_pixels_tab[0][0] = avg_pixels16_sse2;

  2716.             H264_QPEL_FUNCS(0, 0, sse2);

  2717.         }

  2718.         if(mm_flags & AV_CPU_FLAG_SSE2){

  2719.             H264_QPEL_FUNCS(0, 1, sse2);

  2720.             H264_QPEL_FUNCS(0, 2, sse2);

  2721.             H264_QPEL_FUNCS(0, 3, sse2);

  2722.             H264_QPEL_FUNCS(1, 1, sse2);

  2723.             H264_QPEL_FUNCS(1, 2, sse2);

  2724.             H264_QPEL_FUNCS(1, 3, sse2);

  2725.             H264_QPEL_FUNCS(2, 1, sse2);

  2726.             H264_QPEL_FUNCS(2, 2, sse2);

  2727.             H264_QPEL_FUNCS(2, 3, sse2);

  2728.             H264_QPEL_FUNCS(3, 1, sse2);

  2729.             H264_QPEL_FUNCS(3, 2, sse2);

  2730.             H264_QPEL_FUNCS(3, 3, sse2);

  2731.         }

  2732. #if HAVE_SSSE3

  2733.         if(mm_flags & AV_CPU_FLAG_SSSE3){

  2734.             H264_QPEL_FUNCS(1, 0, ssse3);

  2735.             H264_QPEL_FUNCS(1, 1, ssse3);

  2736.             H264_QPEL_FUNCS(1, 2, ssse3);

  2737.             H264_QPEL_FUNCS(1, 3, ssse3);

  2738.             H264_QPEL_FUNCS(2, 0, ssse3);

  2739.             H264_QPEL_FUNCS(2, 1, ssse3);

  2740.             H264_QPEL_FUNCS(2, 2, ssse3);

  2741.             H264_QPEL_FUNCS(2, 3, ssse3);

  2742.             H264_QPEL_FUNCS(3, 0, ssse3);

  2743.             H264_QPEL_FUNCS(3, 1, ssse3);

  2744.             H264_QPEL_FUNCS(3, 2, ssse3);

  2745.             H264_QPEL_FUNCS(3, 3, ssse3);

  2746.             c->add_png_paeth_prediction= add_png_paeth_prediction_ssse3;

  2747. #if HAVE_YASM

  2748.             c->put_h264_chroma_pixels_tab[0]= ff_put_h264_chroma_mc8_ssse3_rnd;

  2749.             c->avg_h264_chroma_pixels_tab[0]= ff_avg_h264_chroma_mc8_ssse3_rnd;

  2750.             c->put_h264_chroma_pixels_tab[1]= ff_put_h264_chroma_mc4_ssse3;

  2751.             c->avg_h264_chroma_pixels_tab[1]= ff_avg_h264_chroma_mc4_ssse3;

  2752.             c->add_hfyu_left_prediction = ff_add_hfyu_left_prediction_ssse3;

  2753.             if (mm_flags & AV_CPU_FLAG_SSE4) // not really sse4, just slow on Conroe

  2754.                 c->add_hfyu_left_prediction = ff_add_hfyu_left_prediction_sse4;

  2755. #endif

  2756.         }

  2757. #endif

  2758. 

  2759.         if(mm_flags & AV_CPU_FLAG_3DNOW){

  2760.             c->vorbis_inverse_coupling = vorbis_inverse_coupling_3dnow;

  2761.             c->vector_fmul = vector_fmul_3dnow;

  2762.         }

  2763.         if(mm_flags & AV_CPU_FLAG_3DNOWEXT){

  2764.             c->vector_fmul_reverse = vector_fmul_reverse_3dnow2;

  2765. #if HAVE_6REGS

  2766.             c->vector_fmul_window = vector_fmul_window_3dnow2;

  2767. #endif

  2768.         }

  2769.         if(mm_flags & AV_CPU_FLAG_MMX2){

  2770. #if HAVE_YASM

  2771.             c->scalarproduct_int16 = ff_scalarproduct_int16_mmx2;

  2772.             c->scalarproduct_and_madd_int16 = ff_scalarproduct_and_madd_int16_mmx2;

  2773.             if (avctx->flags & CODEC_FLAG_BITEXACT) {

  2774.                 c->apply_window_int16 = ff_apply_window_int16_mmxext_ba;

  2775.             } else {

  2776.                 c->apply_window_int16 = ff_apply_window_int16_mmxext;

  2777.             }

  2778. #endif

  2779.         }

  2780.         if(mm_flags & AV_CPU_FLAG_SSE){

  2781.             c->vorbis_inverse_coupling = vorbis_inverse_coupling_sse;

  2782.             c->ac3_downmix = ac3_downmix_sse;

  2783.             c->vector_fmul = vector_fmul_sse;

  2784.             c->vector_fmul_reverse = vector_fmul_reverse_sse;

  2785.             c->vector_fmul_add = vector_fmul_add_sse;

  2786. #if HAVE_6REGS

  2787.             c->vector_fmul_window = vector_fmul_window_sse;

  2788. #endif

  2789.             c->vector_clipf = vector_clipf_sse;

  2790. #if HAVE_YASM

  2791.             c->scalarproduct_float = ff_scalarproduct_float_sse;

  2792. #endif

  2793.         }

  2794.         if(mm_flags & AV_CPU_FLAG_3DNOW)

  2795.             c->vector_fmul_add = vector_fmul_add_3dnow; // faster than sse

  2796.         if(mm_flags & AV_CPU_FLAG_SSE2){

  2797. #if HAVE_YASM

  2798.             c->scalarproduct_int16 = ff_scalarproduct_int16_sse2;

  2799.             c->scalarproduct_and_madd_int16 = ff_scalarproduct_and_madd_int16_sse2;

  2800.             if (avctx->flags & CODEC_FLAG_BITEXACT) {

  2801.                 c->apply_window_int16 = ff_apply_window_int16_sse2_ba;

  2802.             } else {

  2803.                 if (!(mm_flags & AV_CPU_FLAG_SSE2SLOW)) {

  2804.                     c->apply_window_int16 = ff_apply_window_int16_sse2;

  2805.                 }

  2806.             }

  2807. 

  2808.             c->emulated_edge_mc = emulated_edge_mc_sse;

  2809.             c->gmc= gmc_sse;

  2810. #endif

  2811.         }

  2812.         if (mm_flags & AV_CPU_FLAG_SSSE3) {

  2813. #if HAVE_YASM

  2814.             if (mm_flags & AV_CPU_FLAG_ATOM) {

  2815.                 c->apply_window_int16 = ff_apply_window_int16_ssse3_atom;

  2816.             } else {

  2817.                 c->apply_window_int16 = ff_apply_window_int16_ssse3;

  2818.             }

  2819.             if (!(mm_flags & (AV_CPU_FLAG_SSE42|AV_CPU_FLAG_3DNOW))) { // cachesplit

  2820.                 c->scalarproduct_and_madd_int16 = ff_scalarproduct_and_madd_int16_ssse3;

  2821.             }

  2822. #endif

  2823.         }

  2824.     }

  2825. 

  2826.     if (CONFIG_ENCODERS)

  2827.         dsputilenc_init_mmx(c, avctx);

  2828. 

  2829. #if 0

  2830.     // for speed testing

  2831.     get_pixels = just_return;

  2832.     put_pixels_clamped = just_return;

  2833.     add_pixels_clamped = just_return;

  2834. 

  2835.     pix_abs16x16 = just_return;

  2836.     pix_abs16x16_x2 = just_return;

  2837.     pix_abs16x16_y2 = just_return;

  2838.     pix_abs16x16_xy2 = just_return;

  2839. 

  2840.     put_pixels_tab[0] = just_return;

  2841.     put_pixels_tab[1] = just_return;

  2842.     put_pixels_tab[2] = just_return;

  2843.     put_pixels_tab[3] = just_return;

  2844. 

  2845.     put_no_rnd_pixels_tab[0] = just_return;

  2846.     put_no_rnd_pixels_tab[1] = just_return;

  2847.     put_no_rnd_pixels_tab[2] = just_return;

  2848.     put_no_rnd_pixels_tab[3] = just_return;

  2849. 

  2850.     avg_pixels_tab[0] = just_return;

  2851.     avg_pixels_tab[1] = just_return;

  2852.     avg_pixels_tab[2] = just_return;

  2853.     avg_pixels_tab[3] = just_return;

  2854. 

  2855.     avg_no_rnd_pixels_tab[0] = just_return;

  2856.     avg_no_rnd_pixels_tab[1] = just_return;

  2857.     avg_no_rnd_pixels_tab[2] = just_return;

  2858.     avg_no_rnd_pixels_tab[3] = just_return;

  2859. 

  2860.     //av_fdct = just_return;

  2861.     //ff_idct = just_return;

  2862. #endif

  2863. }