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.
26966 Commits
32 Branches
283MB
Tree: b59dd1ea4f
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from 'b59dd1ea4f'
${ noResults }
FFmpeg/libavcodec/x86/dsputil_mmx.c

2962 lines
123KB
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 FFmpeg.

  7.  *

  8.  * FFmpeg 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.  * FFmpeg 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 FFmpeg; 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 "dsputil_mmx.h"

  32. #include "idct_xvid.h"

  33. 

  34. //#undef NDEBUG

  35. //#include <assert.h>

  36. 

  37. /* pixel operations */

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

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

  40. 

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

  42. {0x8000000080000000ULL, 0x8000000080000000ULL};

  43. 

  44. DECLARE_ALIGNED(8,  const xmm_reg,  ff_pw_3  ) = {0x0003000300030003ULL, 0x0003000300030003ULL};

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  64. 

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

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

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

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

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

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

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

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

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

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

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

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

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

  78. 

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

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

  81. 

  82. #define JUMPALIGN() __asm__ volatile (ASMALIGN(3)::)

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

  84. 

  85. #define MOVQ_BFE(regd) \

  86.     __asm__ volatile ( \

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

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

  89. 

  90. #ifndef PIC

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

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

  93. #else

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

  95. // pcmpeqd -> -1

  96. #define MOVQ_BONE(regd) \

  97.     __asm__ volatile ( \

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

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

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

  101. 

  102. #define MOVQ_WTWO(regd) \

  103.     __asm__ volatile ( \

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

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

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

  107. 

  108. #endif

  109. 

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

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

  112. // regfe is supposed to contain 0xfefefefefefefefe

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

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

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

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

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

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

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

  120. 

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

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

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

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

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

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

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

  128. 

  129. // mm6 is supposed to contain 0xfefefefefefefefe

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

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

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

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

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

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

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

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

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

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

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

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

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

  143. 

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

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

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

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

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

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

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

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

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

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

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

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

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

  157. 

  158. /***********************************/

  159. /* MMX no rounding */

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

  161. #define SET_RND  MOVQ_WONE

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

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

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

  165. 

  166. #include "dsputil_mmx_rnd_template.c"

  167. 

  168. #undef DEF

  169. #undef SET_RND

  170. #undef PAVGBP

  171. #undef PAVGB

  172. /***********************************/

  173. /* MMX rounding */

  174. 

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

  176. #define SET_RND  MOVQ_WTWO

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

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

  179. 

  180. #include "dsputil_mmx_rnd_template.c"

  181. 

  182. #undef DEF

  183. #undef SET_RND

  184. #undef PAVGBP

  185. #undef PAVGB

  186. #undef OP_AVG

  187. 

  188. /***********************************/

  189. /* 3Dnow specific */

  190. 

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

  192. #define PAVGB "pavgusb"

  193. #define OP_AVG PAVGB

  194. 

  195. #include "dsputil_mmx_avg_template.c"

  196. 

  197. #undef DEF

  198. #undef PAVGB

  199. #undef OP_AVG

  200. 

  201. /***********************************/

  202. /* MMX2 specific */

  203. 

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

  205. 

  206. /* Introduced only in MMX2 set */

  207. #define PAVGB "pavgb"

  208. #define OP_AVG PAVGB

  209. 

  210. #include "dsputil_mmx_avg_template.c"

  211. 

  212. #undef DEF

  213. #undef PAVGB

  214. #undef OP_AVG

  215. 

  216. #define put_no_rnd_pixels16_mmx put_pixels16_mmx

  217. #define put_no_rnd_pixels8_mmx put_pixels8_mmx

  218. #define put_pixels16_mmx2 put_pixels16_mmx

  219. #define put_pixels8_mmx2 put_pixels8_mmx

  220. #define put_pixels4_mmx2 put_pixels4_mmx

  221. #define put_no_rnd_pixels16_mmx2 put_no_rnd_pixels16_mmx

  222. #define put_no_rnd_pixels8_mmx2 put_no_rnd_pixels8_mmx

  223. #define put_pixels16_3dnow put_pixels16_mmx

  224. #define put_pixels8_3dnow put_pixels8_mmx

  225. #define put_pixels4_3dnow put_pixels4_mmx

  226. #define put_no_rnd_pixels16_3dnow put_no_rnd_pixels16_mmx

  227. #define put_no_rnd_pixels8_3dnow put_no_rnd_pixels8_mmx

  228. 

  229. /***********************************/

  230. /* standard MMX */

  231. 

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

  233. {

  234.     const DCTELEM *p;

  235.     uint8_t *pix;

  236. 

  237.     /* read the pixels */

  238.     p = block;

  239.     pix = pixels;

  240.     /* unrolled loop */

  241.         __asm__ volatile(

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  259.                 :"memory");

  260.         pix += line_size*4;

  261.         p += 32;

  262. 

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

  264.     // compiler would generate some very strange code

  265.     // thus using "r"

  266.     __asm__ volatile(

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  284.             :"memory");

  285. }

  286. 

  287. DECLARE_ASM_CONST(8, uint8_t, ff_vector128)[8] =

  288.   { 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80 };

  289. 

  290. #define put_signed_pixels_clamped_mmx_half(off) \

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  307. 

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

  309. {

  310.     x86_reg line_skip = line_size;

  311.     x86_reg line_skip3;

  312. 

  313.     __asm__ volatile (

  314.             "movq "MANGLE(ff_vector128)", %%mm0 \n\t"

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

  316.             put_signed_pixels_clamped_mmx_half(0)

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

  318.             put_signed_pixels_clamped_mmx_half(64)

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

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

  321.             :"memory");

  322. }

  323. 

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

  325. {

  326.     const DCTELEM *p;

  327.     uint8_t *pix;

  328.     int i;

  329. 

  330.     /* read the pixels */

  331.     p = block;

  332.     pix = pixels;

  333.     MOVQ_ZERO(mm7);

  334.     i = 4;

  335.     do {

  336.         __asm__ volatile(

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  358.                 :"r"(p)

  359.                 :"memory");

  360.         pix += line_size*2;

  361.         p += 16;

  362.     } while (--i);

  363. }

  364. 

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

  366. {

  367.     __asm__ volatile(

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

  369.          ASMALIGN(3)

  370.          "1:                            \n\t"

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

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

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

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

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

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

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

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

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

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

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

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

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

  384.          "jnz 1b                        \n\t"

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

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

  387.          : "%"REG_a, "memory"

  388.         );

  389. }

  390. 

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

  392. {

  393.     __asm__ volatile(

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

  395.          ASMALIGN(3)

  396.          "1:                            \n\t"

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

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

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

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

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

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

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

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

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

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

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

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

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

  410.          "jnz 1b                        \n\t"

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

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

  413.          : "%"REG_a, "memory"

  414.         );

  415. }

  416. 

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

  418. {

  419.     __asm__ volatile(

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

  421.          ASMALIGN(3)

  422.          "1:                            \n\t"

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  444.          "jnz 1b                        \n\t"

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

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

  447.          : "%"REG_a, "memory"

  448.         );

  449. }

  450. 

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

  452. {

  453.     __asm__ volatile(

  454.          "1:                            \n\t"

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

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

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

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

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

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

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

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

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

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

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

  466.          "jnz 1b                        \n\t"

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

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

  469.          : "memory"

  470.         );

  471. }

  472. 

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

  474. {

  475.     __asm__ volatile(

  476.          "1:                            \n\t"

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  492.          "jnz 1b                        \n\t"

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

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

  495.          : "memory"

  496.         );

  497. }

  498. 

  499. #define CLEAR_BLOCKS(name,n) \

  500. static void name(DCTELEM *blocks)\

  501. {\

  502.     __asm__ volatile(\

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

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

  505.                 "1:                             \n\t"\

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

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

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

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

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

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

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

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

  514.                 : "%"REG_a\

  515.         );\

  516. }

  517. CLEAR_BLOCKS(clear_blocks_mmx, 6)

  518. CLEAR_BLOCKS(clear_block_mmx, 1)

  519. 

  520. static void clear_block_sse(DCTELEM *block)

  521. {

  522.     __asm__ volatile(

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

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

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

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

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

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

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

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

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

  532.         :: "r"(block)

  533.         : "memory"

  534.     );

  535. }

  536. 

  537. static void clear_blocks_sse(DCTELEM *blocks)

  538. {\

  539.     __asm__ volatile(

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

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

  542.         "1:                     \n"

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

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

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

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

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

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

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

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

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

  552.         " js 1b                 \n"

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

  554.             "i" (-128*6)

  555.         : "%"REG_a

  556.     );

  557. }

  558. 

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

  560.     x86_reg i=0;

  561.     __asm__ volatile(

  562.         "jmp 2f                         \n\t"

  563.         "1:                             \n\t"

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

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

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

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

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

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

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

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

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

  573.         "2:                             \n\t"

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

  575.         " js 1b                         \n\t"

  576.         : "+r" (i)

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

  578.     );

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

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

  581. }

  582. 

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

  584.     x86_reg i=0;

  585.     __asm__ volatile(

  586.         "jmp 2f                         \n\t"

  587.         "1:                             \n\t"

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

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

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

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

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

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

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

  595.         "2:                             \n\t"

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

  597.         " js 1b                         \n\t"

  598.         : "+r" (i)

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

  600.     );

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

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

  603. }

  604. 

  605. #if HAVE_7REGS && HAVE_TEN_OPERANDS

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

  607.     x86_reg w2 = -w;

  608.     x86_reg x;

  609.     int l = *left & 0xff;

  610.     int tl = *left_top & 0xff;

  611.     int t;

  612.     __asm__ volatile(

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

  614.         "1: \n"

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  630.         "inc    %4 \n"

  631.         "jl 1b \n"

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

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

  634.     );

  635.     *left = l;

  636.     *left_top = tl;

  637. }

  638. #endif

  639. 

  640. #define H263_LOOP_FILTER \

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  710. 

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

  712.     if(CONFIG_H263_DECODER || CONFIG_H263_ENCODER) {

  713.     const int strength= ff_h263_loop_filter_strength[qscale];

  714. 

  715.     __asm__ volatile(

  716. 

  717.         H263_LOOP_FILTER

  718. 

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

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

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

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

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

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

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

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

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

  728.     );

  729.     }

  730. }

  731. 

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

  733.     if(CONFIG_H263_DECODER || CONFIG_H263_ENCODER) {

  734.     const int strength= ff_h263_loop_filter_strength[qscale];

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

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

  737. 

  738.     src -= 2;

  739. 

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

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

  742.     __asm__ volatile(

  743.         H263_LOOP_FILTER // 5 3 4 6

  744. 

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

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

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

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

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

  750.     );

  751. 

  752.     __asm__ volatile(

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  777.         :: "r" (src),

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

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

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

  781.     );

  782.     }

  783. }

  784. 

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

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

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

  788. {

  789.     uint8_t *ptr, *last_line;

  790.     int i;

  791. 

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

  793.     /* left and right */

  794.     ptr = buf;

  795.     if(w==8)

  796.     {

  797.         __asm__ volatile(

  798.                 "1:                             \n\t"

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

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

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

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

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

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

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

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

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

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

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

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

  811.                 " jb 1b                         \n\t"

  812.                 : "+r" (ptr)

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

  814.         );

  815.     }

  816.     else

  817.     {

  818.         __asm__ volatile(

  819.                 "1:                             \n\t"

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  834.                 " jb 1b                         \n\t"

  835.                 : "+r" (ptr)

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

  837.         );

  838.     }

  839. 

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

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

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

  857.         __asm__ volatile(

  858.                 "1:                             \n\t"

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

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

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

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

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

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

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

  866.                 " jb 1b                         \n\t"

  867.                 : "+r" (ptr)

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

  869.         );

  870.     }

  871. }

  872. 

  873. #define PAETH(cpu, abs3)\

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

  875. {\

  876.     x86_reg i = -bpp;\

  877.     x86_reg end = w-3;\

  878.     __asm__ volatile(\

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

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

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

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

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

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

  885.         "1: \n"\

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

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

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

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

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

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

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

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

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

  895.         abs3\

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  918.         "jle 1b \n"\

  919.         :"+r"(i)\

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

  921.          "m"(ff_pw_255)\

  922.         :"memory"\

  923.     );\

  924. }

  925. 

  926. #define ABS3_MMX2\

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

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

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

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

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

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

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

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

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

  936. 

  937. #define ABS3_SSSE3\

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

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

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

  941. 

  942. PAETH(mmx2, ABS3_MMX2)

  943. #if HAVE_SSSE3

  944. PAETH(ssse3, ABS3_SSSE3)

  945. #endif

  946. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  967. 

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

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

  970.     uint64_t temp;\

  971. \

  972.     __asm__ volatile(\

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

  974.         "1:                               \n\t"\

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  1000.         "paddw %6, %%mm6                  \n\t"\

  1001.         "paddw %%mm6, %%mm0               \n\t" /* 20a - 6b + 3c - d */\

  1002.         "psraw $5, %%mm0                  \n\t"\

  1003.         "movq %%mm0, %5                   \n\t"\

  1004.         /* mm1=EFGH, mm2=DEFG, mm3=CDEF, mm4=BCDE, mm7=0 */\

  1005.         \

  1006.         "movq 5(%0), %%mm0                \n\t" /* FGHIJKLM */\

  1007.         "movq %%mm0, %%mm5                \n\t" /* FGHIJKLM */\

  1008.         "movq %%mm0, %%mm6                \n\t" /* FGHIJKLM */\

  1009.         "psrlq $8, %%mm0                  \n\t" /* GHIJKLM0 */\

  1010.         "psrlq $16, %%mm5                 \n\t" /* HIJKLM00 */\

  1011.         "punpcklbw %%mm7, %%mm0           \n\t" /* 0G0H0I0J */\

  1012.         "punpcklbw %%mm7, %%mm5           \n\t" /* 0H0I0J0K */\

  1013.         "paddw %%mm0, %%mm2               \n\t" /* b */\

  1014.         "paddw %%mm5, %%mm3               \n\t" /* c */\

  1015.         "paddw %%mm2, %%mm2               \n\t" /* 2b */\

  1016.         "psubw %%mm2, %%mm3               \n\t" /* c - 2b */\

  1017.         "movq %%mm6, %%mm2                \n\t" /* FGHIJKLM */\

  1018.         "psrlq $24, %%mm6                 \n\t" /* IJKLM000 */\

  1019.         "punpcklbw %%mm7, %%mm2           \n\t" /* 0F0G0H0I */\

  1020.         "punpcklbw %%mm7, %%mm6           \n\t" /* 0I0J0K0L */\

  1021.         "pmullw "MANGLE(ff_pw_3)", %%mm3  \n\t" /* 3c - 6b */\

  1022.         "paddw %%mm2, %%mm1               \n\t" /* a */\

  1023.         "paddw %%mm6, %%mm4               \n\t" /* d */\

  1024.         "pmullw "MANGLE(ff_pw_20)", %%mm1 \n\t" /* 20a */\

  1025.         "psubw %%mm4, %%mm3               \n\t" /* - 6b +3c - d */\

  1026.         "paddw %6, %%mm1                  \n\t"\

  1027.         "paddw %%mm1, %%mm3               \n\t" /* 20a - 6b +3c - d */\

  1028.         "psraw $5, %%mm3                  \n\t"\

  1029.         "movq %5, %%mm1                   \n\t"\

  1030.         "packuswb %%mm3, %%mm1            \n\t"\

  1031.         OP_MMX2(%%mm1, (%1),%%mm4, q)\

  1032.         /* mm0= GHIJ, mm2=FGHI, mm5=HIJK, mm6=IJKL, mm7=0 */\

  1033.         \

  1034.         "movq 9(%0), %%mm1                \n\t" /* JKLMNOPQ */\

  1035.         "movq %%mm1, %%mm4                \n\t" /* JKLMNOPQ */\

  1036.         "movq %%mm1, %%mm3                \n\t" /* JKLMNOPQ */\

  1037.         "psrlq $8, %%mm1                  \n\t" /* KLMNOPQ0 */\

  1038.         "psrlq $16, %%mm4                 \n\t" /* LMNOPQ00 */\

  1039.         "punpcklbw %%mm7, %%mm1           \n\t" /* 0K0L0M0N */\

  1040.         "punpcklbw %%mm7, %%mm4           \n\t" /* 0L0M0N0O */\

  1041.         "paddw %%mm1, %%mm5               \n\t" /* b */\

  1042.         "paddw %%mm4, %%mm0               \n\t" /* c */\

  1043.         "paddw %%mm5, %%mm5               \n\t" /* 2b */\

  1044.         "psubw %%mm5, %%mm0               \n\t" /* c - 2b */\

  1045.         "movq %%mm3, %%mm5                \n\t" /* JKLMNOPQ */\

  1046.         "psrlq $24, %%mm3                 \n\t" /* MNOPQ000 */\

  1047.         "pmullw "MANGLE(ff_pw_3)", %%mm0  \n\t" /* 3c - 6b */\

  1048.         "punpcklbw %%mm7, %%mm3           \n\t" /* 0M0N0O0P */\

  1049.         "paddw %%mm3, %%mm2               \n\t" /* d */\

  1050.         "psubw %%mm2, %%mm0               \n\t" /* -6b + 3c - d */\

  1051.         "movq %%mm5, %%mm2                \n\t" /* JKLMNOPQ */\

  1052.         "punpcklbw %%mm7, %%mm2           \n\t" /* 0J0K0L0M */\

  1053.         "punpckhbw %%mm7, %%mm5           \n\t" /* 0N0O0P0Q */\

  1054.         "paddw %%mm2, %%mm6               \n\t" /* a */\

  1055.         "pmullw "MANGLE(ff_pw_20)", %%mm6 \n\t" /* 20a */\

  1056.         "paddw %6, %%mm0                  \n\t"\

  1057.         "paddw %%mm6, %%mm0               \n\t" /* 20a - 6b + 3c - d */\

  1058.         "psraw $5, %%mm0                  \n\t"\

  1059.         /* mm1=KLMN, mm2=JKLM, mm3=MNOP, mm4=LMNO, mm5=NOPQ mm7=0 */\

  1060.         \

  1061.         "paddw %%mm5, %%mm3               \n\t" /* a */\

  1062.         "pshufw $0xF9, %%mm5, %%mm6       \n\t" /* 0O0P0Q0Q */\

  1063.         "paddw %%mm4, %%mm6               \n\t" /* b */\

  1064.         "pshufw $0xBE, %%mm5, %%mm4       \n\t" /* 0P0Q0Q0P */\

  1065.         "pshufw $0x6F, %%mm5, %%mm5       \n\t" /* 0Q0Q0P0O */\

  1066.         "paddw %%mm1, %%mm4               \n\t" /* c */\

  1067.         "paddw %%mm2, %%mm5               \n\t" /* d */\

  1068.         "paddw %%mm6, %%mm6               \n\t" /* 2b */\

  1069.         "psubw %%mm6, %%mm4               \n\t" /* c - 2b */\

  1070.         "pmullw "MANGLE(ff_pw_20)", %%mm3 \n\t" /* 20a */\

  1071.         "pmullw "MANGLE(ff_pw_3)", %%mm4  \n\t" /* 3c - 6b */\

  1072.         "psubw %%mm5, %%mm3               \n\t" /* -6b + 3c - d */\

  1073.         "paddw %6, %%mm4                  \n\t"\

  1074.         "paddw %%mm3, %%mm4               \n\t" /* 20a - 6b + 3c - d */\

  1075.         "psraw $5, %%mm4                  \n\t"\

  1076.         "packuswb %%mm4, %%mm0            \n\t"\

  1077.         OP_MMX2(%%mm0, 8(%1), %%mm4, q)\

  1078.         \

  1079.         "add %3, %0                       \n\t"\

  1080.         "add %4, %1                       \n\t"\

  1081.         "decl %2                          \n\t"\

  1082.         " jnz 1b                          \n\t"\

  1083.         : "+a"(src), "+c"(dst), "+D"(h)\

  1084.         : "d"((x86_reg)srcStride), "S"((x86_reg)dstStride), /*"m"(ff_pw_20), "m"(ff_pw_3),*/ "m"(temp), "m"(ROUNDER)\

  1085.         : "memory"\

  1086.     );\

  1087. }\

  1088. \

  1089. static void OPNAME ## mpeg4_qpel16_h_lowpass_3dnow(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\

  1090.     int i;\

  1091.     int16_t temp[16];\

  1092.     /* quick HACK, XXX FIXME MUST be optimized */\

  1093.     for(i=0; i<h; i++)\

  1094.     {\

  1095.         temp[ 0]= (src[ 0]+src[ 1])*20 - (src[ 0]+src[ 2])*6 + (src[ 1]+src[ 3])*3 - (src[ 2]+src[ 4]);\

  1096.         temp[ 1]= (src[ 1]+src[ 2])*20 - (src[ 0]+src[ 3])*6 + (src[ 0]+src[ 4])*3 - (src[ 1]+src[ 5]);\

  1097.         temp[ 2]= (src[ 2]+src[ 3])*20 - (src[ 1]+src[ 4])*6 + (src[ 0]+src[ 5])*3 - (src[ 0]+src[ 6]);\

  1098.         temp[ 3]= (src[ 3]+src[ 4])*20 - (src[ 2]+src[ 5])*6 + (src[ 1]+src[ 6])*3 - (src[ 0]+src[ 7]);\

  1099.         temp[ 4]= (src[ 4]+src[ 5])*20 - (src[ 3]+src[ 6])*6 + (src[ 2]+src[ 7])*3 - (src[ 1]+src[ 8]);\

  1100.         temp[ 5]= (src[ 5]+src[ 6])*20 - (src[ 4]+src[ 7])*6 + (src[ 3]+src[ 8])*3 - (src[ 2]+src[ 9]);\

  1101.         temp[ 6]= (src[ 6]+src[ 7])*20 - (src[ 5]+src[ 8])*6 + (src[ 4]+src[ 9])*3 - (src[ 3]+src[10]);\

  1102.         temp[ 7]= (src[ 7]+src[ 8])*20 - (src[ 6]+src[ 9])*6 + (src[ 5]+src[10])*3 - (src[ 4]+src[11]);\

  1103.         temp[ 8]= (src[ 8]+src[ 9])*20 - (src[ 7]+src[10])*6 + (src[ 6]+src[11])*3 - (src[ 5]+src[12]);\

  1104.         temp[ 9]= (src[ 9]+src[10])*20 - (src[ 8]+src[11])*6 + (src[ 7]+src[12])*3 - (src[ 6]+src[13]);\

  1105.         temp[10]= (src[10]+src[11])*20 - (src[ 9]+src[12])*6 + (src[ 8]+src[13])*3 - (src[ 7]+src[14]);\

  1106.         temp[11]= (src[11]+src[12])*20 - (src[10]+src[13])*6 + (src[ 9]+src[14])*3 - (src[ 8]+src[15]);\

  1107.         temp[12]= (src[12]+src[13])*20 - (src[11]+src[14])*6 + (src[10]+src[15])*3 - (src[ 9]+src[16]);\

  1108.         temp[13]= (src[13]+src[14])*20 - (src[12]+src[15])*6 + (src[11]+src[16])*3 - (src[10]+src[16]);\

  1109.         temp[14]= (src[14]+src[15])*20 - (src[13]+src[16])*6 + (src[12]+src[16])*3 - (src[11]+src[15]);\

  1110.         temp[15]= (src[15]+src[16])*20 - (src[14]+src[16])*6 + (src[13]+src[15])*3 - (src[12]+src[14]);\

  1111.         __asm__ volatile(\

  1112.             "movq (%0), %%mm0               \n\t"\

  1113.             "movq 8(%0), %%mm1              \n\t"\

  1114.             "paddw %2, %%mm0                \n\t"\

  1115.             "paddw %2, %%mm1                \n\t"\

  1116.             "psraw $5, %%mm0                \n\t"\

  1117.             "psraw $5, %%mm1                \n\t"\

  1118.             "packuswb %%mm1, %%mm0          \n\t"\

  1119.             OP_3DNOW(%%mm0, (%1), %%mm1, q)\

  1120.             "movq 16(%0), %%mm0             \n\t"\

  1121.             "movq 24(%0), %%mm1             \n\t"\

  1122.             "paddw %2, %%mm0                \n\t"\

  1123.             "paddw %2, %%mm1                \n\t"\

  1124.             "psraw $5, %%mm0                \n\t"\

  1125.             "psraw $5, %%mm1                \n\t"\

  1126.             "packuswb %%mm1, %%mm0          \n\t"\

  1127.             OP_3DNOW(%%mm0, 8(%1), %%mm1, q)\

  1128.             :: "r"(temp), "r"(dst), "m"(ROUNDER)\

  1129.             : "memory"\

  1130.         );\

  1131.         dst+=dstStride;\

  1132.         src+=srcStride;\

  1133.     }\

  1134. }\

  1135. \

  1136. static void OPNAME ## mpeg4_qpel8_h_lowpass_mmx2(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\

  1137.     __asm__ volatile(\

  1138.         "pxor %%mm7, %%mm7                \n\t"\

  1139.         "1:                               \n\t"\

  1140.         "movq  (%0), %%mm0                \n\t" /* ABCDEFGH */\

  1141.         "movq %%mm0, %%mm1                \n\t" /* ABCDEFGH */\

  1142.         "movq %%mm0, %%mm2                \n\t" /* ABCDEFGH */\

  1143.         "punpcklbw %%mm7, %%mm0           \n\t" /* 0A0B0C0D */\

  1144.         "punpckhbw %%mm7, %%mm1           \n\t" /* 0E0F0G0H */\

  1145.         "pshufw $0x90, %%mm0, %%mm5       \n\t" /* 0A0A0B0C */\

  1146.         "pshufw $0x41, %%mm0, %%mm6       \n\t" /* 0B0A0A0B */\

  1147.         "movq %%mm2, %%mm3                \n\t" /* ABCDEFGH */\

  1148.         "movq %%mm2, %%mm4                \n\t" /* ABCDEFGH */\

  1149.         "psllq $8, %%mm2                  \n\t" /* 0ABCDEFG */\

  1150.         "psllq $16, %%mm3                 \n\t" /* 00ABCDEF */\

  1151.         "psllq $24, %%mm4                 \n\t" /* 000ABCDE */\

  1152.         "punpckhbw %%mm7, %%mm2           \n\t" /* 0D0E0F0G */\

  1153.         "punpckhbw %%mm7, %%mm3           \n\t" /* 0C0D0E0F */\

  1154.         "punpckhbw %%mm7, %%mm4           \n\t" /* 0B0C0D0E */\

  1155.         "paddw %%mm3, %%mm5               \n\t" /* b */\

  1156.         "paddw %%mm2, %%mm6               \n\t" /* c */\

  1157.         "paddw %%mm5, %%mm5               \n\t" /* 2b */\

  1158.         "psubw %%mm5, %%mm6               \n\t" /* c - 2b */\

  1159.         "pshufw $0x06, %%mm0, %%mm5       \n\t" /* 0C0B0A0A */\

  1160.         "pmullw "MANGLE(ff_pw_3)", %%mm6  \n\t" /* 3c - 6b */\

  1161.         "paddw %%mm4, %%mm0               \n\t" /* a */\

  1162.         "paddw %%mm1, %%mm5               \n\t" /* d */\

  1163.         "pmullw "MANGLE(ff_pw_20)", %%mm0 \n\t" /* 20a */\

  1164.         "psubw %%mm5, %%mm0               \n\t" /* 20a - d */\

  1165.         "paddw %5, %%mm6                  \n\t"\

  1166.         "paddw %%mm6, %%mm0               \n\t" /* 20a - 6b + 3c - d */\

  1167.         "psraw $5, %%mm0                  \n\t"\

  1168.         /* mm1=EFGH, mm2=DEFG, mm3=CDEF, mm4=BCDE, mm7=0 */\

  1169.         \

  1170.         "movd 5(%0), %%mm5                \n\t" /* FGHI */\

  1171.         "punpcklbw %%mm7, %%mm5           \n\t" /* 0F0G0H0I */\

  1172.         "pshufw $0xF9, %%mm5, %%mm6       \n\t" /* 0G0H0I0I */\

  1173.         "paddw %%mm5, %%mm1               \n\t" /* a */\

  1174.         "paddw %%mm6, %%mm2               \n\t" /* b */\

  1175.         "pshufw $0xBE, %%mm5, %%mm6       \n\t" /* 0H0I0I0H */\

  1176.         "pshufw $0x6F, %%mm5, %%mm5       \n\t" /* 0I0I0H0G */\

  1177.         "paddw %%mm6, %%mm3               \n\t" /* c */\

  1178.         "paddw %%mm5, %%mm4               \n\t" /* d */\

  1179.         "paddw %%mm2, %%mm2               \n\t" /* 2b */\

  1180.         "psubw %%mm2, %%mm3               \n\t" /* c - 2b */\

  1181.         "pmullw "MANGLE(ff_pw_20)", %%mm1 \n\t" /* 20a */\

  1182.         "pmullw "MANGLE(ff_pw_3)", %%mm3  \n\t" /* 3c - 6b */\

  1183.         "psubw %%mm4, %%mm3               \n\t" /* -6b + 3c - d */\

  1184.         "paddw %5, %%mm1                  \n\t"\

  1185.         "paddw %%mm1, %%mm3               \n\t" /* 20a - 6b + 3c - d */\

  1186.         "psraw $5, %%mm3                  \n\t"\

  1187.         "packuswb %%mm3, %%mm0            \n\t"\

  1188.         OP_MMX2(%%mm0, (%1), %%mm4, q)\

  1189.         \

  1190.         "add %3, %0                       \n\t"\

  1191.         "add %4, %1                       \n\t"\

  1192.         "decl %2                          \n\t"\

  1193.         " jnz 1b                          \n\t"\

  1194.         : "+a"(src), "+c"(dst), "+d"(h)\

  1195.         : "S"((x86_reg)srcStride), "D"((x86_reg)dstStride), /*"m"(ff_pw_20), "m"(ff_pw_3),*/ "m"(ROUNDER)\

  1196.         : "memory"\

  1197.     );\

  1198. }\

  1199. \

  1200. static void OPNAME ## mpeg4_qpel8_h_lowpass_3dnow(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\

  1201.     int i;\

  1202.     int16_t temp[8];\

  1203.     /* quick HACK, XXX FIXME MUST be optimized */\

  1204.     for(i=0; i<h; i++)\

  1205.     {\

  1206.         temp[ 0]= (src[ 0]+src[ 1])*20 - (src[ 0]+src[ 2])*6 + (src[ 1]+src[ 3])*3 - (src[ 2]+src[ 4]);\

  1207.         temp[ 1]= (src[ 1]+src[ 2])*20 - (src[ 0]+src[ 3])*6 + (src[ 0]+src[ 4])*3 - (src[ 1]+src[ 5]);\

  1208.         temp[ 2]= (src[ 2]+src[ 3])*20 - (src[ 1]+src[ 4])*6 + (src[ 0]+src[ 5])*3 - (src[ 0]+src[ 6]);\

  1209.         temp[ 3]= (src[ 3]+src[ 4])*20 - (src[ 2]+src[ 5])*6 + (src[ 1]+src[ 6])*3 - (src[ 0]+src[ 7]);\

  1210.         temp[ 4]= (src[ 4]+src[ 5])*20 - (src[ 3]+src[ 6])*6 + (src[ 2]+src[ 7])*3 - (src[ 1]+src[ 8]);\

  1211.         temp[ 5]= (src[ 5]+src[ 6])*20 - (src[ 4]+src[ 7])*6 + (src[ 3]+src[ 8])*3 - (src[ 2]+src[ 8]);\

  1212.         temp[ 6]= (src[ 6]+src[ 7])*20 - (src[ 5]+src[ 8])*6 + (src[ 4]+src[ 8])*3 - (src[ 3]+src[ 7]);\

  1213.         temp[ 7]= (src[ 7]+src[ 8])*20 - (src[ 6]+src[ 8])*6 + (src[ 5]+src[ 7])*3 - (src[ 4]+src[ 6]);\

  1214.         __asm__ volatile(\

  1215.             "movq (%0), %%mm0           \n\t"\

  1216.             "movq 8(%0), %%mm1          \n\t"\

  1217.             "paddw %2, %%mm0            \n\t"\

  1218.             "paddw %2, %%mm1            \n\t"\

  1219.             "psraw $5, %%mm0            \n\t"\

  1220.             "psraw $5, %%mm1            \n\t"\

  1221.             "packuswb %%mm1, %%mm0      \n\t"\

  1222.             OP_3DNOW(%%mm0, (%1), %%mm1, q)\

  1223.             :: "r"(temp), "r"(dst), "m"(ROUNDER)\

  1224.             :"memory"\

  1225.         );\

  1226.         dst+=dstStride;\

  1227.         src+=srcStride;\

  1228.     }\

  1229. }

  1230. 

  1231. #define QPEL_OP(OPNAME, ROUNDER, RND, OP, MMX)\

  1232. \

  1233. static void OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\

  1234.     uint64_t temp[17*4];\

  1235.     uint64_t *temp_ptr= temp;\

  1236.     int count= 17;\

  1237. \

  1238.     /*FIXME unroll */\

  1239.     __asm__ volatile(\

  1240.         "pxor %%mm7, %%mm7              \n\t"\

  1241.         "1:                             \n\t"\

  1242.         "movq (%0), %%mm0               \n\t"\

  1243.         "movq (%0), %%mm1               \n\t"\

  1244.         "movq 8(%0), %%mm2              \n\t"\

  1245.         "movq 8(%0), %%mm3              \n\t"\

  1246.         "punpcklbw %%mm7, %%mm0         \n\t"\

  1247.         "punpckhbw %%mm7, %%mm1         \n\t"\

  1248.         "punpcklbw %%mm7, %%mm2         \n\t"\

  1249.         "punpckhbw %%mm7, %%mm3         \n\t"\

  1250.         "movq %%mm0, (%1)               \n\t"\

  1251.         "movq %%mm1, 17*8(%1)           \n\t"\

  1252.         "movq %%mm2, 2*17*8(%1)         \n\t"\

  1253.         "movq %%mm3, 3*17*8(%1)         \n\t"\

  1254.         "add $8, %1                     \n\t"\

  1255.         "add %3, %0                     \n\t"\

  1256.         "decl %2                        \n\t"\

  1257.         " jnz 1b                        \n\t"\

  1258.         : "+r" (src), "+r" (temp_ptr), "+r"(count)\

  1259.         : "r" ((x86_reg)srcStride)\

  1260.         : "memory"\

  1261.     );\

  1262.     \

  1263.     temp_ptr= temp;\

  1264.     count=4;\

  1265.     \

  1266. /*FIXME reorder for speed */\

  1267.     __asm__ volatile(\

  1268.         /*"pxor %%mm7, %%mm7              \n\t"*/\

  1269.         "1:                             \n\t"\

  1270.         "movq (%0), %%mm0               \n\t"\

  1271.         "movq 8(%0), %%mm1              \n\t"\

  1272.         "movq 16(%0), %%mm2             \n\t"\

  1273.         "movq 24(%0), %%mm3             \n\t"\

  1274.         QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 16(%0),  8(%0),   (%0), 32(%0), (%1), OP)\

  1275.         QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5,  8(%0),   (%0),   (%0), 40(%0), (%1, %3), OP)\

  1276.         "add %4, %1                     \n\t"\

  1277.         QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5,   (%0),   (%0),  8(%0), 48(%0), (%1), OP)\

  1278.         \

  1279.         QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5,   (%0),  8(%0), 16(%0), 56(%0), (%1, %3), OP)\

  1280.         "add %4, %1                     \n\t"\

  1281.         QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5,  8(%0), 16(%0), 24(%0), 64(%0), (%1), OP)\

  1282.         QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 16(%0), 24(%0), 32(%0), 72(%0), (%1, %3), OP)\

  1283.         "add %4, %1                     \n\t"\

  1284.         QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 24(%0), 32(%0), 40(%0), 80(%0), (%1), OP)\

  1285.         QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 32(%0), 40(%0), 48(%0), 88(%0), (%1, %3), OP)\

  1286.         "add %4, %1                     \n\t"\

  1287.         QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 40(%0), 48(%0), 56(%0), 96(%0), (%1), OP)\

  1288.         QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 48(%0), 56(%0), 64(%0),104(%0), (%1, %3), OP)\

  1289.         "add %4, %1                     \n\t"\

  1290.         QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 56(%0), 64(%0), 72(%0),112(%0), (%1), OP)\

  1291.         QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 64(%0), 72(%0), 80(%0),120(%0), (%1, %3), OP)\

  1292.         "add %4, %1                     \n\t"\

  1293.         QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 72(%0), 80(%0), 88(%0),128(%0), (%1), OP)\

  1294.         \

  1295.         QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 80(%0), 88(%0), 96(%0),128(%0), (%1, %3), OP)\

  1296.         "add %4, %1                     \n\t"  \

  1297.         QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 88(%0), 96(%0),104(%0),120(%0), (%1), OP)\

  1298.         QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 96(%0),104(%0),112(%0),112(%0), (%1, %3), OP)\

  1299.         \

  1300.         "add $136, %0                   \n\t"\

  1301.         "add %6, %1                     \n\t"\

  1302.         "decl %2                        \n\t"\

  1303.         " jnz 1b                        \n\t"\

  1304.         \

  1305.         : "+r"(temp_ptr), "+r"(dst), "+g"(count)\

  1306.         : "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)\

  1307.         :"memory"\

  1308.     );\

  1309. }\

  1310. \

  1311. static void OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\

  1312.     uint64_t temp[9*2];\

  1313.     uint64_t *temp_ptr= temp;\

  1314.     int count= 9;\

  1315. \

  1316.     /*FIXME unroll */\

  1317.     __asm__ volatile(\

  1318.         "pxor %%mm7, %%mm7              \n\t"\

  1319.         "1:                             \n\t"\

  1320.         "movq (%0), %%mm0               \n\t"\

  1321.         "movq (%0), %%mm1               \n\t"\

  1322.         "punpcklbw %%mm7, %%mm0         \n\t"\

  1323.         "punpckhbw %%mm7, %%mm1         \n\t"\

  1324.         "movq %%mm0, (%1)               \n\t"\

  1325.         "movq %%mm1, 9*8(%1)            \n\t"\

  1326.         "add $8, %1                     \n\t"\

  1327.         "add %3, %0                     \n\t"\

  1328.         "decl %2                        \n\t"\

  1329.         " jnz 1b                        \n\t"\

  1330.         : "+r" (src), "+r" (temp_ptr), "+r"(count)\

  1331.         : "r" ((x86_reg)srcStride)\

  1332.         : "memory"\

  1333.     );\

  1334.     \

  1335.     temp_ptr= temp;\

  1336.     count=2;\

  1337.     \

  1338. /*FIXME reorder for speed */\

  1339.     __asm__ volatile(\

  1340.         /*"pxor %%mm7, %%mm7              \n\t"*/\

  1341.         "1:                             \n\t"\

  1342.         "movq (%0), %%mm0               \n\t"\

  1343.         "movq 8(%0), %%mm1              \n\t"\

  1344.         "movq 16(%0), %%mm2             \n\t"\

  1345.         "movq 24(%0), %%mm3             \n\t"\

  1346.         QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 16(%0),  8(%0),   (%0), 32(%0), (%1), OP)\

  1347.         QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5,  8(%0),   (%0),   (%0), 40(%0), (%1, %3), OP)\

  1348.         "add %4, %1                     \n\t"\

  1349.         QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5,   (%0),   (%0),  8(%0), 48(%0), (%1), OP)\

  1350.         \

  1351.         QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5,   (%0),  8(%0), 16(%0), 56(%0), (%1, %3), OP)\

  1352.         "add %4, %1                     \n\t"\

  1353.         QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5,  8(%0), 16(%0), 24(%0), 64(%0), (%1), OP)\

  1354.         \

  1355.         QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 16(%0), 24(%0), 32(%0), 64(%0), (%1, %3), OP)\

  1356.         "add %4, %1                     \n\t"\

  1357.         QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 24(%0), 32(%0), 40(%0), 56(%0), (%1), OP)\

  1358.         QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 32(%0), 40(%0), 48(%0), 48(%0), (%1, %3), OP)\

  1359.                 \

  1360.         "add $72, %0                    \n\t"\

  1361.         "add %6, %1                     \n\t"\

  1362.         "decl %2                        \n\t"\

  1363.         " jnz 1b                        \n\t"\

  1364.          \

  1365.         : "+r"(temp_ptr), "+r"(dst), "+g"(count)\

  1366.         : "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)\

  1367.         : "memory"\

  1368.    );\

  1369. }\

  1370. \

  1371. static void OPNAME ## qpel8_mc00_ ## MMX (uint8_t *dst, uint8_t *src, int stride){\

  1372.     OPNAME ## pixels8_ ## MMX(dst, src, stride, 8);\

  1373. }\

  1374. \

  1375. static void OPNAME ## qpel8_mc10_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1376.     uint64_t temp[8];\

  1377.     uint8_t * const half= (uint8_t*)temp;\

  1378.     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(half, src, 8, stride, 8);\

  1379.     OPNAME ## pixels8_l2_ ## MMX(dst, src, half, stride, stride, 8);\

  1380. }\

  1381. \

  1382. static void OPNAME ## qpel8_mc20_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1383.     OPNAME ## mpeg4_qpel8_h_lowpass_ ## MMX(dst, src, stride, stride, 8);\

  1384. }\

  1385. \

  1386. static void OPNAME ## qpel8_mc30_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1387.     uint64_t temp[8];\

  1388.     uint8_t * const half= (uint8_t*)temp;\

  1389.     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(half, src, 8, stride, 8);\

  1390.     OPNAME ## pixels8_l2_ ## MMX(dst, src+1, half, stride, stride, 8);\

  1391. }\

  1392. \

  1393. static void OPNAME ## qpel8_mc01_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1394.     uint64_t temp[8];\

  1395.     uint8_t * const half= (uint8_t*)temp;\

  1396.     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(half, src, 8, stride);\

  1397.     OPNAME ## pixels8_l2_ ## MMX(dst, src, half, stride, stride, 8);\

  1398. }\

  1399. \

  1400. static void OPNAME ## qpel8_mc02_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1401.     OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, src, stride, stride);\

  1402. }\

  1403. \

  1404. static void OPNAME ## qpel8_mc03_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1405.     uint64_t temp[8];\

  1406.     uint8_t * const half= (uint8_t*)temp;\

  1407.     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(half, src, 8, stride);\

  1408.     OPNAME ## pixels8_l2_ ## MMX(dst, src+stride, half, stride, stride, 8);\

  1409. }\

  1410. static void OPNAME ## qpel8_mc11_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1411.     uint64_t half[8 + 9];\

  1412.     uint8_t * const halfH= ((uint8_t*)half) + 64;\

  1413.     uint8_t * const halfHV= ((uint8_t*)half);\

  1414.     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\

  1415.     put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8, stride, 9);\

  1416.     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\

  1417.     OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV, stride, 8, 8);\

  1418. }\

  1419. static void OPNAME ## qpel8_mc31_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1420.     uint64_t half[8 + 9];\

  1421.     uint8_t * const halfH= ((uint8_t*)half) + 64;\

  1422.     uint8_t * const halfHV= ((uint8_t*)half);\

  1423.     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\

  1424.     put ## RND ## pixels8_l2_ ## MMX(halfH, src+1, halfH, 8, stride, 9);\

  1425.     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\

  1426.     OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV, stride, 8, 8);\

  1427. }\

  1428. static void OPNAME ## qpel8_mc13_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1429.     uint64_t half[8 + 9];\

  1430.     uint8_t * const halfH= ((uint8_t*)half) + 64;\

  1431.     uint8_t * const halfHV= ((uint8_t*)half);\

  1432.     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\

  1433.     put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8, stride, 9);\

  1434.     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\

  1435.     OPNAME ## pixels8_l2_ ## MMX(dst, halfH+8, halfHV, stride, 8, 8);\

  1436. }\

  1437. static void OPNAME ## qpel8_mc33_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1438.     uint64_t half[8 + 9];\

  1439.     uint8_t * const halfH= ((uint8_t*)half) + 64;\

  1440.     uint8_t * const halfHV= ((uint8_t*)half);\

  1441.     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\

  1442.     put ## RND ## pixels8_l2_ ## MMX(halfH, src+1, halfH, 8, stride, 9);\

  1443.     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\

  1444.     OPNAME ## pixels8_l2_ ## MMX(dst, halfH+8, halfHV, stride, 8, 8);\

  1445. }\

  1446. static void OPNAME ## qpel8_mc21_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1447.     uint64_t half[8 + 9];\

  1448.     uint8_t * const halfH= ((uint8_t*)half) + 64;\

  1449.     uint8_t * const halfHV= ((uint8_t*)half);\

  1450.     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\

  1451.     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\

  1452.     OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV, stride, 8, 8);\

  1453. }\

  1454. static void OPNAME ## qpel8_mc23_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1455.     uint64_t half[8 + 9];\

  1456.     uint8_t * const halfH= ((uint8_t*)half) + 64;\

  1457.     uint8_t * const halfHV= ((uint8_t*)half);\

  1458.     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\

  1459.     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\

  1460.     OPNAME ## pixels8_l2_ ## MMX(dst, halfH+8, halfHV, stride, 8, 8);\

  1461. }\

  1462. static void OPNAME ## qpel8_mc12_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1463.     uint64_t half[8 + 9];\

  1464.     uint8_t * const halfH= ((uint8_t*)half);\

  1465.     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\

  1466.     put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8, stride, 9);\

  1467.     OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH, stride, 8);\

  1468. }\

  1469. static void OPNAME ## qpel8_mc32_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1470.     uint64_t half[8 + 9];\

  1471.     uint8_t * const halfH= ((uint8_t*)half);\

  1472.     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\

  1473.     put ## RND ## pixels8_l2_ ## MMX(halfH, src+1, halfH, 8, stride, 9);\

  1474.     OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH, stride, 8);\

  1475. }\

  1476. static void OPNAME ## qpel8_mc22_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1477.     uint64_t half[9];\

  1478.     uint8_t * const halfH= ((uint8_t*)half);\

  1479.     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\

  1480.     OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH, stride, 8);\

  1481. }\

  1482. static void OPNAME ## qpel16_mc00_ ## MMX (uint8_t *dst, uint8_t *src, int stride){\

  1483.     OPNAME ## pixels16_ ## MMX(dst, src, stride, 16);\

  1484. }\

  1485. \

  1486. static void OPNAME ## qpel16_mc10_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1487.     uint64_t temp[32];\

  1488.     uint8_t * const half= (uint8_t*)temp;\

  1489.     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(half, src, 16, stride, 16);\

  1490.     OPNAME ## pixels16_l2_ ## MMX(dst, src, half, stride, stride, 16);\

  1491. }\

  1492. \

  1493. static void OPNAME ## qpel16_mc20_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1494.     OPNAME ## mpeg4_qpel16_h_lowpass_ ## MMX(dst, src, stride, stride, 16);\

  1495. }\

  1496. \

  1497. static void OPNAME ## qpel16_mc30_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1498.     uint64_t temp[32];\

  1499.     uint8_t * const half= (uint8_t*)temp;\

  1500.     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(half, src, 16, stride, 16);\

  1501.     OPNAME ## pixels16_l2_ ## MMX(dst, src+1, half, stride, stride, 16);\

  1502. }\

  1503. \

  1504. static void OPNAME ## qpel16_mc01_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1505.     uint64_t temp[32];\

  1506.     uint8_t * const half= (uint8_t*)temp;\

  1507.     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(half, src, 16, stride);\

  1508.     OPNAME ## pixels16_l2_ ## MMX(dst, src, half, stride, stride, 16);\

  1509. }\

  1510. \

  1511. static void OPNAME ## qpel16_mc02_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1512.     OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, src, stride, stride);\

  1513. }\

  1514. \

  1515. static void OPNAME ## qpel16_mc03_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1516.     uint64_t temp[32];\

  1517.     uint8_t * const half= (uint8_t*)temp;\

  1518.     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(half, src, 16, stride);\

  1519.     OPNAME ## pixels16_l2_ ## MMX(dst, src+stride, half, stride, stride, 16);\

  1520. }\

  1521. static void OPNAME ## qpel16_mc11_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1522.     uint64_t half[16*2 + 17*2];\

  1523.     uint8_t * const halfH= ((uint8_t*)half) + 256;\

  1524.     uint8_t * const halfHV= ((uint8_t*)half);\

  1525.     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\

  1526.     put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16, stride, 17);\

  1527.     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\

  1528.     OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV, stride, 16, 16);\

  1529. }\

  1530. static void OPNAME ## qpel16_mc31_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1531.     uint64_t half[16*2 + 17*2];\

  1532.     uint8_t * const halfH= ((uint8_t*)half) + 256;\

  1533.     uint8_t * const halfHV= ((uint8_t*)half);\

  1534.     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\

  1535.     put ## RND ## pixels16_l2_ ## MMX(halfH, src+1, halfH, 16, stride, 17);\

  1536.     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\

  1537.     OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV, stride, 16, 16);\

  1538. }\

  1539. static void OPNAME ## qpel16_mc13_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1540.     uint64_t half[16*2 + 17*2];\

  1541.     uint8_t * const halfH= ((uint8_t*)half) + 256;\

  1542.     uint8_t * const halfHV= ((uint8_t*)half);\

  1543.     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\

  1544.     put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16, stride, 17);\

  1545.     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\

  1546.     OPNAME ## pixels16_l2_ ## MMX(dst, halfH+16, halfHV, stride, 16, 16);\

  1547. }\

  1548. static void OPNAME ## qpel16_mc33_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1549.     uint64_t half[16*2 + 17*2];\

  1550.     uint8_t * const halfH= ((uint8_t*)half) + 256;\

  1551.     uint8_t * const halfHV= ((uint8_t*)half);\

  1552.     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\

  1553.     put ## RND ## pixels16_l2_ ## MMX(halfH, src+1, halfH, 16, stride, 17);\

  1554.     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\

  1555.     OPNAME ## pixels16_l2_ ## MMX(dst, halfH+16, halfHV, stride, 16, 16);\

  1556. }\

  1557. static void OPNAME ## qpel16_mc21_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1558.     uint64_t half[16*2 + 17*2];\

  1559.     uint8_t * const halfH= ((uint8_t*)half) + 256;\

  1560.     uint8_t * const halfHV= ((uint8_t*)half);\

  1561.     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\

  1562.     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\

  1563.     OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV, stride, 16, 16);\

  1564. }\

  1565. static void OPNAME ## qpel16_mc23_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1566.     uint64_t half[16*2 + 17*2];\

  1567.     uint8_t * const halfH= ((uint8_t*)half) + 256;\

  1568.     uint8_t * const halfHV= ((uint8_t*)half);\

  1569.     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\

  1570.     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\

  1571.     OPNAME ## pixels16_l2_ ## MMX(dst, halfH+16, halfHV, stride, 16, 16);\

  1572. }\

  1573. static void OPNAME ## qpel16_mc12_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1574.     uint64_t half[17*2];\

  1575.     uint8_t * const halfH= ((uint8_t*)half);\

  1576.     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\

  1577.     put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16, stride, 17);\

  1578.     OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH, stride, 16);\

  1579. }\

  1580. static void OPNAME ## qpel16_mc32_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1581.     uint64_t half[17*2];\

  1582.     uint8_t * const halfH= ((uint8_t*)half);\

  1583.     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\

  1584.     put ## RND ## pixels16_l2_ ## MMX(halfH, src+1, halfH, 16, stride, 17);\

  1585.     OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH, stride, 16);\

  1586. }\

  1587. static void OPNAME ## qpel16_mc22_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1588.     uint64_t half[17*2];\

  1589.     uint8_t * const halfH= ((uint8_t*)half);\

  1590.     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\

  1591.     OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH, stride, 16);\

  1592. }

  1593. 

  1594. #define PUT_OP(a,b,temp, size) "mov" #size " " #a ", " #b "        \n\t"

  1595. #define AVG_3DNOW_OP(a,b,temp, size) \

  1596. "mov" #size " " #b ", " #temp "   \n\t"\

  1597. "pavgusb " #temp ", " #a "        \n\t"\

  1598. "mov" #size " " #a ", " #b "      \n\t"

  1599. #define AVG_MMX2_OP(a,b,temp, size) \

  1600. "mov" #size " " #b ", " #temp "   \n\t"\

  1601. "pavgb " #temp ", " #a "          \n\t"\

  1602. "mov" #size " " #a ", " #b "      \n\t"

  1603. 

  1604. QPEL_BASE(put_       , ff_pw_16, _       , PUT_OP, PUT_OP)

  1605. QPEL_BASE(avg_       , ff_pw_16, _       , AVG_MMX2_OP, AVG_3DNOW_OP)

  1606. QPEL_BASE(put_no_rnd_, ff_pw_15, _no_rnd_, PUT_OP, PUT_OP)

  1607. QPEL_OP(put_       , ff_pw_16, _       , PUT_OP, 3dnow)

  1608. QPEL_OP(avg_       , ff_pw_16, _       , AVG_3DNOW_OP, 3dnow)

  1609. QPEL_OP(put_no_rnd_, ff_pw_15, _no_rnd_, PUT_OP, 3dnow)

  1610. QPEL_OP(put_       , ff_pw_16, _       , PUT_OP, mmx2)

  1611. QPEL_OP(avg_       , ff_pw_16, _       , AVG_MMX2_OP, mmx2)

  1612. QPEL_OP(put_no_rnd_, ff_pw_15, _no_rnd_, PUT_OP, mmx2)

  1613. 

  1614. /***********************************/

  1615. /* bilinear qpel: not compliant to any spec, only for -lavdopts fast */

  1616. 

  1617. #define QPEL_2TAP_XY(OPNAME, SIZE, MMX, XY, HPEL)\

  1618. static void OPNAME ## 2tap_qpel ## SIZE ## _mc ## XY ## _ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1619.     OPNAME ## pixels ## SIZE ## HPEL(dst, src, stride, SIZE);\

  1620. }

  1621. #define QPEL_2TAP_L3(OPNAME, SIZE, MMX, XY, S0, S1, S2)\

  1622. static void OPNAME ## 2tap_qpel ## SIZE ## _mc ## XY ## _ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1623.     OPNAME ## 2tap_qpel ## SIZE ## _l3_ ## MMX(dst, src+S0, stride, SIZE, S1, S2);\

  1624. }

  1625. 

  1626. #define QPEL_2TAP(OPNAME, SIZE, MMX)\

  1627. QPEL_2TAP_XY(OPNAME, SIZE, MMX, 20, _x2_ ## MMX)\

  1628. QPEL_2TAP_XY(OPNAME, SIZE, MMX, 02, _y2_ ## MMX)\

  1629. QPEL_2TAP_XY(OPNAME, SIZE, MMX, 22, _xy2_mmx)\

  1630. static const qpel_mc_func OPNAME ## 2tap_qpel ## SIZE ## _mc00_ ## MMX =\

  1631.                           OPNAME ## qpel ## SIZE ## _mc00_ ## MMX;\

  1632. static const qpel_mc_func OPNAME ## 2tap_qpel ## SIZE ## _mc21_ ## MMX =\

  1633.                           OPNAME ## 2tap_qpel ## SIZE ## _mc20_ ## MMX;\

  1634. static const qpel_mc_func OPNAME ## 2tap_qpel ## SIZE ## _mc12_ ## MMX =\

  1635.                           OPNAME ## 2tap_qpel ## SIZE ## _mc02_ ## MMX;\

  1636. static void OPNAME ## 2tap_qpel ## SIZE ## _mc32_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1637.     OPNAME ## pixels ## SIZE ## _y2_ ## MMX(dst, src+1, stride, SIZE);\

  1638. }\

  1639. static void OPNAME ## 2tap_qpel ## SIZE ## _mc23_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\

  1640.     OPNAME ## pixels ## SIZE ## _x2_ ## MMX(dst, src+stride, stride, SIZE);\

  1641. }\

  1642. QPEL_2TAP_L3(OPNAME, SIZE, MMX, 10, 0,         1,       0)\

  1643. QPEL_2TAP_L3(OPNAME, SIZE, MMX, 30, 1,        -1,       0)\

  1644. QPEL_2TAP_L3(OPNAME, SIZE, MMX, 01, 0,         stride,  0)\

  1645. QPEL_2TAP_L3(OPNAME, SIZE, MMX, 03, stride,   -stride,  0)\

  1646. QPEL_2TAP_L3(OPNAME, SIZE, MMX, 11, 0,         stride,  1)\

  1647. QPEL_2TAP_L3(OPNAME, SIZE, MMX, 31, 1,         stride, -1)\

  1648. QPEL_2TAP_L3(OPNAME, SIZE, MMX, 13, stride,   -stride,  1)\

  1649. QPEL_2TAP_L3(OPNAME, SIZE, MMX, 33, stride+1, -stride, -1)\

  1650. 

  1651. QPEL_2TAP(put_, 16, mmx2)

  1652. QPEL_2TAP(avg_, 16, mmx2)

  1653. QPEL_2TAP(put_,  8, mmx2)

  1654. QPEL_2TAP(avg_,  8, mmx2)

  1655. QPEL_2TAP(put_, 16, 3dnow)

  1656. QPEL_2TAP(avg_, 16, 3dnow)

  1657. QPEL_2TAP(put_,  8, 3dnow)

  1658. QPEL_2TAP(avg_,  8, 3dnow)

  1659. 

  1660. 

  1661. #if 0

  1662. static void just_return(void) { return; }

  1663. #endif

  1664. 

  1665. static void gmc_mmx(uint8_t *dst, uint8_t *src, int stride, int h, int ox, int oy,

  1666.                     int dxx, int dxy, int dyx, int dyy, int shift, int r, int width, int height){

  1667.     const int w = 8;

  1668.     const int ix = ox>>(16+shift);

  1669.     const int iy = oy>>(16+shift);

  1670.     const int oxs = ox>>4;

  1671.     const int oys = oy>>4;

  1672.     const int dxxs = dxx>>4;

  1673.     const int dxys = dxy>>4;

  1674.     const int dyxs = dyx>>4;

  1675.     const int dyys = dyy>>4;

  1676.     const uint16_t r4[4] = {r,r,r,r};

  1677.     const uint16_t dxy4[4] = {dxys,dxys,dxys,dxys};

  1678.     const uint16_t dyy4[4] = {dyys,dyys,dyys,dyys};

  1679.     const uint64_t shift2 = 2*shift;

  1680.     uint8_t edge_buf[(h+1)*stride];

  1681.     int x, y;

  1682. 

  1683.     const int dxw = (dxx-(1<<(16+shift)))*(w-1);

  1684.     const int dyh = (dyy-(1<<(16+shift)))*(h-1);

  1685.     const int dxh = dxy*(h-1);

  1686.     const int dyw = dyx*(w-1);

  1687.     if( // non-constant fullpel offset (3% of blocks)

  1688.         ((ox^(ox+dxw)) | (ox^(ox+dxh)) | (ox^(ox+dxw+dxh)) |

  1689.          (oy^(oy+dyw)) | (oy^(oy+dyh)) | (oy^(oy+dyw+dyh))) >> (16+shift)

  1690.         // uses more than 16 bits of subpel mv (only at huge resolution)

  1691.         || (dxx|dxy|dyx|dyy)&15 )

  1692.     {

  1693.         //FIXME could still use mmx for some of the rows

  1694.         ff_gmc_c(dst, src, stride, h, ox, oy, dxx, dxy, dyx, dyy, shift, r, width, height);

  1695.         return;

  1696.     }

  1697. 

  1698.     src += ix + iy*stride;

  1699.     if( (unsigned)ix >= width-w ||

  1700.         (unsigned)iy >= height-h )

  1701.     {

  1702.         ff_emulated_edge_mc(edge_buf, src, stride, w+1, h+1, ix, iy, width, height);

  1703.         src = edge_buf;

  1704.     }

  1705. 

  1706.     __asm__ volatile(

  1707.         "movd         %0, %%mm6 \n\t"

  1708.         "pxor      %%mm7, %%mm7 \n\t"

  1709.         "punpcklwd %%mm6, %%mm6 \n\t"

  1710.         "punpcklwd %%mm6, %%mm6 \n\t"

  1711.         :: "r"(1<<shift)

  1712.     );

  1713. 

  1714.     for(x=0; x<w; x+=4){

  1715.         uint16_t dx4[4] = { oxs - dxys + dxxs*(x+0),

  1716.                             oxs - dxys + dxxs*(x+1),

  1717.                             oxs - dxys + dxxs*(x+2),

  1718.                             oxs - dxys + dxxs*(x+3) };

  1719.         uint16_t dy4[4] = { oys - dyys + dyxs*(x+0),

  1720.                             oys - dyys + dyxs*(x+1),

  1721.                             oys - dyys + dyxs*(x+2),

  1722.                             oys - dyys + dyxs*(x+3) };

  1723. 

  1724.         for(y=0; y<h; y++){

  1725.             __asm__ volatile(

  1726.                 "movq   %0,  %%mm4 \n\t"

  1727.                 "movq   %1,  %%mm5 \n\t"

  1728.                 "paddw  %2,  %%mm4 \n\t"

  1729.                 "paddw  %3,  %%mm5 \n\t"

  1730.                 "movq   %%mm4, %0  \n\t"

  1731.                 "movq   %%mm5, %1  \n\t"

  1732.                 "psrlw  $12, %%mm4 \n\t"

  1733.                 "psrlw  $12, %%mm5 \n\t"

  1734.                 : "+m"(*dx4), "+m"(*dy4)

  1735.                 : "m"(*dxy4), "m"(*dyy4)

  1736.             );

  1737. 

  1738.             __asm__ volatile(

  1739.                 "movq   %%mm6, %%mm2 \n\t"

  1740.                 "movq   %%mm6, %%mm1 \n\t"

  1741.                 "psubw  %%mm4, %%mm2 \n\t"

  1742.                 "psubw  %%mm5, %%mm1 \n\t"

  1743.                 "movq   %%mm2, %%mm0 \n\t"

  1744.                 "movq   %%mm4, %%mm3 \n\t"

  1745.                 "pmullw %%mm1, %%mm0 \n\t" // (s-dx)*(s-dy)

  1746.                 "pmullw %%mm5, %%mm3 \n\t" // dx*dy

  1747.                 "pmullw %%mm5, %%mm2 \n\t" // (s-dx)*dy

  1748.                 "pmullw %%mm4, %%mm1 \n\t" // dx*(s-dy)

  1749. 

  1750.                 "movd   %4,    %%mm5 \n\t"

  1751.                 "movd   %3,    %%mm4 \n\t"

  1752.                 "punpcklbw %%mm7, %%mm5 \n\t"

  1753.                 "punpcklbw %%mm7, %%mm4 \n\t"

  1754.                 "pmullw %%mm5, %%mm3 \n\t" // src[1,1] * dx*dy

  1755.                 "pmullw %%mm4, %%mm2 \n\t" // src[0,1] * (s-dx)*dy

  1756. 

  1757.                 "movd   %2,    %%mm5 \n\t"

  1758.                 "movd   %1,    %%mm4 \n\t"

  1759.                 "punpcklbw %%mm7, %%mm5 \n\t"

  1760.                 "punpcklbw %%mm7, %%mm4 \n\t"

  1761.                 "pmullw %%mm5, %%mm1 \n\t" // src[1,0] * dx*(s-dy)

  1762.                 "pmullw %%mm4, %%mm0 \n\t" // src[0,0] * (s-dx)*(s-dy)

  1763.                 "paddw  %5,    %%mm1 \n\t"

  1764.                 "paddw  %%mm3, %%mm2 \n\t"

  1765.                 "paddw  %%mm1, %%mm0 \n\t"

  1766.                 "paddw  %%mm2, %%mm0 \n\t"

  1767. 

  1768.                 "psrlw    %6,    %%mm0 \n\t"

  1769.                 "packuswb %%mm0, %%mm0 \n\t"

  1770.                 "movd     %%mm0, %0    \n\t"

  1771. 

  1772.                 : "=m"(dst[x+y*stride])

  1773.                 : "m"(src[0]), "m"(src[1]),

  1774.                   "m"(src[stride]), "m"(src[stride+1]),

  1775.                   "m"(*r4), "m"(shift2)

  1776.             );

  1777.             src += stride;

  1778.         }

  1779.         src += 4-h*stride;

  1780.     }

  1781. }

  1782. 

  1783. #define PREFETCH(name, op) \

  1784. static void name(void *mem, int stride, int h){\

  1785.     const uint8_t *p= mem;\

  1786.     do{\

  1787.         __asm__ volatile(#op" %0" :: "m"(*p));\

  1788.         p+= stride;\

  1789.     }while(--h);\

  1790. }

  1791. PREFETCH(prefetch_mmx2,  prefetcht0)

  1792. PREFETCH(prefetch_3dnow, prefetch)

  1793. #undef PREFETCH

  1794. 

  1795. #include "h264_qpel_mmx.c"

  1796. 

  1797. void ff_put_h264_chroma_mc8_mmx_rnd   (uint8_t *dst, uint8_t *src,

  1798.                                        int stride, int h, int x, int y);

  1799. void ff_put_vc1_chroma_mc8_mmx_nornd  (uint8_t *dst, uint8_t *src,

  1800.                                        int stride, int h, int x, int y);

  1801. void ff_put_rv40_chroma_mc8_mmx       (uint8_t *dst, uint8_t *src,

  1802.                                        int stride, int h, int x, int y);

  1803. void ff_avg_h264_chroma_mc8_mmx2_rnd  (uint8_t *dst, uint8_t *src,

  1804.                                        int stride, int h, int x, int y);

  1805. void ff_avg_vc1_chroma_mc8_mmx2_nornd (uint8_t *dst, uint8_t *src,

  1806.                                        int stride, int h, int x, int y);

  1807. void ff_avg_rv40_chroma_mc8_mmx2      (uint8_t *dst, uint8_t *src,

  1808.                                        int stride, int h, int x, int y);

  1809. void ff_avg_h264_chroma_mc8_3dnow_rnd (uint8_t *dst, uint8_t *src,

  1810.                                        int stride, int h, int x, int y);

  1811. void ff_avg_vc1_chroma_mc8_3dnow_nornd(uint8_t *dst, uint8_t *src,

  1812.                                        int stride, int h, int x, int y);

  1813. void ff_avg_rv40_chroma_mc8_3dnow     (uint8_t *dst, uint8_t *src,

  1814.                                        int stride, int h, int x, int y);

  1815. 

  1816. void ff_put_h264_chroma_mc4_mmx       (uint8_t *dst, uint8_t *src,

  1817.                                        int stride, int h, int x, int y);

  1818. void ff_put_rv40_chroma_mc4_mmx       (uint8_t *dst, uint8_t *src,

  1819.                                        int stride, int h, int x, int y);

  1820. void ff_avg_h264_chroma_mc4_mmx2      (uint8_t *dst, uint8_t *src,

  1821.                                        int stride, int h, int x, int y);

  1822. void ff_avg_rv40_chroma_mc4_mmx2      (uint8_t *dst, uint8_t *src,

  1823.                                        int stride, int h, int x, int y);

  1824. void ff_avg_h264_chroma_mc4_3dnow     (uint8_t *dst, uint8_t *src,

  1825.                                        int stride, int h, int x, int y);

  1826. void ff_avg_rv40_chroma_mc4_3dnow     (uint8_t *dst, uint8_t *src,

  1827.                                        int stride, int h, int x, int y);

  1828. 

  1829. void ff_put_h264_chroma_mc2_mmx2      (uint8_t *dst, uint8_t *src,

  1830.                                        int stride, int h, int x, int y);

  1831. void ff_avg_h264_chroma_mc2_mmx2      (uint8_t *dst, uint8_t *src,

  1832.                                        int stride, int h, int x, int y);

  1833. 

  1834. void ff_put_h264_chroma_mc8_ssse3_rnd (uint8_t *dst, uint8_t *src,

  1835.                                        int stride, int h, int x, int y);

  1836. void ff_put_vc1_chroma_mc8_ssse3_nornd(uint8_t *dst, uint8_t *src,

  1837.                                        int stride, int h, int x, int y);

  1838. void ff_put_h264_chroma_mc4_ssse3     (uint8_t *dst, uint8_t *src,

  1839.                                        int stride, int h, int x, int y);

  1840. 

  1841. void ff_avg_h264_chroma_mc8_ssse3_rnd (uint8_t *dst, uint8_t *src,

  1842.                                        int stride, int h, int x, int y);

  1843. void ff_avg_vc1_chroma_mc8_ssse3_nornd(uint8_t *dst, uint8_t *src,

  1844.                                        int stride, int h, int x, int y);

  1845. void ff_avg_h264_chroma_mc4_ssse3     (uint8_t *dst, uint8_t *src,

  1846.                                        int stride, int h, int x, int y);

  1847. 

  1848. 

  1849. /* CAVS specific */

  1850. void ff_put_cavs_qpel8_mc00_mmx2(uint8_t *dst, uint8_t *src, int stride) {

  1851.     put_pixels8_mmx(dst, src, stride, 8);

  1852. }

  1853. void ff_avg_cavs_qpel8_mc00_mmx2(uint8_t *dst, uint8_t *src, int stride) {

  1854.     avg_pixels8_mmx(dst, src, stride, 8);

  1855. }

  1856. void ff_put_cavs_qpel16_mc00_mmx2(uint8_t *dst, uint8_t *src, int stride) {

  1857.     put_pixels16_mmx(dst, src, stride, 16);

  1858. }

  1859. void ff_avg_cavs_qpel16_mc00_mmx2(uint8_t *dst, uint8_t *src, int stride) {

  1860.     avg_pixels16_mmx(dst, src, stride, 16);

  1861. }

  1862. 

  1863. /* VC1 specific */

  1864. void ff_put_vc1_mspel_mc00_mmx(uint8_t *dst, const uint8_t *src, int stride, int rnd) {

  1865.     put_pixels8_mmx(dst, src, stride, 8);

  1866. }

  1867. void ff_avg_vc1_mspel_mc00_mmx2(uint8_t *dst, const uint8_t *src, int stride, int rnd) {

  1868.     avg_pixels8_mmx2(dst, src, stride, 8);

  1869. }

  1870. 

  1871. /* XXX: those functions should be suppressed ASAP when all IDCTs are

  1872.    converted */

  1873. #if CONFIG_GPL

  1874. static void ff_libmpeg2mmx_idct_put(uint8_t *dest, int line_size, DCTELEM *block)

  1875. {

  1876.     ff_mmx_idct (block);

  1877.     ff_put_pixels_clamped_mmx(block, dest, line_size);

  1878. }

  1879. static void ff_libmpeg2mmx_idct_add(uint8_t *dest, int line_size, DCTELEM *block)

  1880. {

  1881.     ff_mmx_idct (block);

  1882.     ff_add_pixels_clamped_mmx(block, dest, line_size);

  1883. }

  1884. static void ff_libmpeg2mmx2_idct_put(uint8_t *dest, int line_size, DCTELEM *block)

  1885. {

  1886.     ff_mmxext_idct (block);

  1887.     ff_put_pixels_clamped_mmx(block, dest, line_size);

  1888. }

  1889. static void ff_libmpeg2mmx2_idct_add(uint8_t *dest, int line_size, DCTELEM *block)

  1890. {

  1891.     ff_mmxext_idct (block);

  1892.     ff_add_pixels_clamped_mmx(block, dest, line_size);

  1893. }

  1894. #endif

  1895. static void ff_idct_xvid_mmx_put(uint8_t *dest, int line_size, DCTELEM *block)

  1896. {

  1897.     ff_idct_xvid_mmx (block);

  1898.     ff_put_pixels_clamped_mmx(block, dest, line_size);

  1899. }

  1900. static void ff_idct_xvid_mmx_add(uint8_t *dest, int line_size, DCTELEM *block)

  1901. {

  1902.     ff_idct_xvid_mmx (block);

  1903.     ff_add_pixels_clamped_mmx(block, dest, line_size);

  1904. }

  1905. static void ff_idct_xvid_mmx2_put(uint8_t *dest, int line_size, DCTELEM *block)

  1906. {

  1907.     ff_idct_xvid_mmx2 (block);

  1908.     ff_put_pixels_clamped_mmx(block, dest, line_size);

  1909. }

  1910. static void ff_idct_xvid_mmx2_add(uint8_t *dest, int line_size, DCTELEM *block)

  1911. {

  1912.     ff_idct_xvid_mmx2 (block);

  1913.     ff_add_pixels_clamped_mmx(block, dest, line_size);

  1914. }

  1915. 

  1916. static void vorbis_inverse_coupling_3dnow(float *mag, float *ang, int blocksize)

  1917. {

  1918.     int i;

  1919.     __asm__ volatile("pxor %%mm7, %%mm7":);

  1920.     for(i=0; i<blocksize; i+=2) {

  1921.         __asm__ volatile(

  1922.             "movq    %0,    %%mm0 \n\t"

  1923.             "movq    %1,    %%mm1 \n\t"

  1924.             "movq    %%mm0, %%mm2 \n\t"

  1925.             "movq    %%mm1, %%mm3 \n\t"

  1926.             "pfcmpge %%mm7, %%mm2 \n\t" // m <= 0.0

  1927.             "pfcmpge %%mm7, %%mm3 \n\t" // a <= 0.0

  1928.             "pslld   $31,   %%mm2 \n\t" // keep only the sign bit

  1929.             "pxor    %%mm2, %%mm1 \n\t"

  1930.             "movq    %%mm3, %%mm4 \n\t"

  1931.             "pand    %%mm1, %%mm3 \n\t"

  1932.             "pandn   %%mm1, %%mm4 \n\t"

  1933.             "pfadd   %%mm0, %%mm3 \n\t" // a = m + ((a<0) & (a ^ sign(m)))

  1934.             "pfsub   %%mm4, %%mm0 \n\t" // m = m + ((a>0) & (a ^ sign(m)))

  1935.             "movq    %%mm3, %1    \n\t"

  1936.             "movq    %%mm0, %0    \n\t"

  1937.             :"+m"(mag[i]), "+m"(ang[i])

  1938.             ::"memory"

  1939.         );

  1940.     }

  1941.     __asm__ volatile("femms");

  1942. }

  1943. static void vorbis_inverse_coupling_sse(float *mag, float *ang, int blocksize)

  1944. {

  1945.     int i;

  1946. 

  1947.     __asm__ volatile(

  1948.             "movaps  %0,     %%xmm5 \n\t"

  1949.         ::"m"(ff_pdw_80000000[0])

  1950.     );

  1951.     for(i=0; i<blocksize; i+=4) {

  1952.         __asm__ volatile(

  1953.             "movaps  %0,     %%xmm0 \n\t"

  1954.             "movaps  %1,     %%xmm1 \n\t"

  1955.             "xorps   %%xmm2, %%xmm2 \n\t"

  1956.             "xorps   %%xmm3, %%xmm3 \n\t"

  1957.             "cmpleps %%xmm0, %%xmm2 \n\t" // m <= 0.0

  1958.             "cmpleps %%xmm1, %%xmm3 \n\t" // a <= 0.0

  1959.             "andps   %%xmm5, %%xmm2 \n\t" // keep only the sign bit

  1960.             "xorps   %%xmm2, %%xmm1 \n\t"

  1961.             "movaps  %%xmm3, %%xmm4 \n\t"

  1962.             "andps   %%xmm1, %%xmm3 \n\t"

  1963.             "andnps  %%xmm1, %%xmm4 \n\t"

  1964.             "addps   %%xmm0, %%xmm3 \n\t" // a = m + ((a<0) & (a ^ sign(m)))

  1965.             "subps   %%xmm4, %%xmm0 \n\t" // m = m + ((a>0) & (a ^ sign(m)))

  1966.             "movaps  %%xmm3, %1     \n\t"

  1967.             "movaps  %%xmm0, %0     \n\t"

  1968.             :"+m"(mag[i]), "+m"(ang[i])

  1969.             ::"memory"

  1970.         );

  1971.     }

  1972. }

  1973. 

  1974. #define IF1(x) x

  1975. #define IF0(x)

  1976. 

  1977. #define MIX5(mono,stereo)\

  1978.     __asm__ volatile(\

  1979.         "movss          0(%2), %%xmm5 \n"\

  1980.         "movss          8(%2), %%xmm6 \n"\

  1981.         "movss         24(%2), %%xmm7 \n"\

  1982.         "shufps    $0, %%xmm5, %%xmm5 \n"\

  1983.         "shufps    $0, %%xmm6, %%xmm6 \n"\

  1984.         "shufps    $0, %%xmm7, %%xmm7 \n"\

  1985.         "1: \n"\

  1986.         "movaps       (%0,%1), %%xmm0 \n"\

  1987.         "movaps  0x400(%0,%1), %%xmm1 \n"\

  1988.         "movaps  0x800(%0,%1), %%xmm2 \n"\

  1989.         "movaps  0xc00(%0,%1), %%xmm3 \n"\

  1990.         "movaps 0x1000(%0,%1), %%xmm4 \n"\

  1991.         "mulps         %%xmm5, %%xmm0 \n"\

  1992.         "mulps         %%xmm6, %%xmm1 \n"\

  1993.         "mulps         %%xmm5, %%xmm2 \n"\

  1994.         "mulps         %%xmm7, %%xmm3 \n"\

  1995.         "mulps         %%xmm7, %%xmm4 \n"\

  1996.  stereo("addps         %%xmm1, %%xmm0 \n")\

  1997.         "addps         %%xmm1, %%xmm2 \n"\

  1998.         "addps         %%xmm3, %%xmm0 \n"\

  1999.         "addps         %%xmm4, %%xmm2 \n"\

  2000.    mono("addps         %%xmm2, %%xmm0 \n")\

  2001.         "movaps  %%xmm0,      (%0,%1) \n"\

  2002.  stereo("movaps  %%xmm2, 0x400(%0,%1) \n")\

  2003.         "add $16, %0 \n"\

  2004.         "jl 1b \n"\

  2005.         :"+&r"(i)\

  2006.         :"r"(samples[0]+len), "r"(matrix)\

  2007.         :XMM_CLOBBERS("%xmm0", "%xmm1", "%xmm2", "%xmm3", \

  2008.                       "%xmm4", "%xmm5", "%xmm6", "%xmm7",)\

  2009.          "memory"\

  2010.     );

  2011. 

  2012. #define MIX_MISC(stereo)\

  2013.     __asm__ volatile(\

  2014.         "1: \n"\

  2015.         "movaps  (%3,%0), %%xmm0 \n"\

  2016.  stereo("movaps   %%xmm0, %%xmm1 \n")\

  2017.         "mulps    %%xmm4, %%xmm0 \n"\

  2018.  stereo("mulps    %%xmm5, %%xmm1 \n")\

  2019.         "lea 1024(%3,%0), %1 \n"\

  2020.         "mov %5, %2 \n"\

  2021.         "2: \n"\

  2022.         "movaps   (%1),   %%xmm2 \n"\

  2023.  stereo("movaps   %%xmm2, %%xmm3 \n")\

  2024.         "mulps   (%4,%2), %%xmm2 \n"\

  2025.  stereo("mulps 16(%4,%2), %%xmm3 \n")\

  2026.         "addps    %%xmm2, %%xmm0 \n"\

  2027.  stereo("addps    %%xmm3, %%xmm1 \n")\

  2028.         "add $1024, %1 \n"\

  2029.         "add $32, %2 \n"\

  2030.         "jl 2b \n"\

  2031.         "movaps   %%xmm0,     (%3,%0) \n"\

  2032.  stereo("movaps   %%xmm1, 1024(%3,%0) \n")\

  2033.         "add $16, %0 \n"\

  2034.         "jl 1b \n"\

  2035.         :"+&r"(i), "=&r"(j), "=&r"(k)\

  2036.         :"r"(samples[0]+len), "r"(matrix_simd+in_ch), "g"((intptr_t)-32*(in_ch-1))\

  2037.         :"memory"\

  2038.     );

  2039. 

  2040. static void ac3_downmix_sse(float (*samples)[256], float (*matrix)[2], int out_ch, int in_ch, int len)

  2041. {

  2042.     int (*matrix_cmp)[2] = (int(*)[2])matrix;

  2043.     intptr_t i,j,k;

  2044. 

  2045.     i = -len*sizeof(float);

  2046.     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]))) {

  2047.         MIX5(IF0,IF1);

  2048.     } else if(in_ch == 5 && out_ch == 1 && matrix_cmp[0][0]==matrix_cmp[2][0] && matrix_cmp[3][0]==matrix_cmp[4][0]) {

  2049.         MIX5(IF1,IF0);

  2050.     } else {

  2051.         DECLARE_ALIGNED(16, float, matrix_simd)[in_ch][2][4];

  2052.         j = 2*in_ch*sizeof(float);

  2053.         __asm__ volatile(

  2054.             "1: \n"

  2055.             "sub $8, %0 \n"

  2056.             "movss     (%2,%0), %%xmm4 \n"

  2057.             "movss    4(%2,%0), %%xmm5 \n"

  2058.             "shufps $0, %%xmm4, %%xmm4 \n"

  2059.             "shufps $0, %%xmm5, %%xmm5 \n"

  2060.             "movaps %%xmm4,   (%1,%0,4) \n"

  2061.             "movaps %%xmm5, 16(%1,%0,4) \n"

  2062.             "jg 1b \n"

  2063.             :"+&r"(j)

  2064.             :"r"(matrix_simd), "r"(matrix)

  2065.             :"memory"

  2066.         );

  2067.         if(out_ch == 2) {

  2068.             MIX_MISC(IF1);

  2069.         } else {

  2070.             MIX_MISC(IF0);

  2071.         }

  2072.     }

  2073. }

  2074. 

  2075. static void vector_fmul_3dnow(float *dst, const float *src, int len){

  2076.     x86_reg i = (len-4)*4;

  2077.     __asm__ volatile(

  2078.         "1: \n\t"

  2079.         "movq    (%1,%0), %%mm0 \n\t"

  2080.         "movq   8(%1,%0), %%mm1 \n\t"

  2081.         "pfmul   (%2,%0), %%mm0 \n\t"

  2082.         "pfmul  8(%2,%0), %%mm1 \n\t"

  2083.         "movq   %%mm0,  (%1,%0) \n\t"

  2084.         "movq   %%mm1, 8(%1,%0) \n\t"

  2085.         "sub  $16, %0 \n\t"

  2086.         "jge 1b \n\t"

  2087.         "femms  \n\t"

  2088.         :"+r"(i)

  2089.         :"r"(dst), "r"(src)

  2090.         :"memory"

  2091.     );

  2092. }

  2093. static void vector_fmul_sse(float *dst, const float *src, int len){

  2094.     x86_reg i = (len-8)*4;

  2095.     __asm__ volatile(

  2096.         "1: \n\t"

  2097.         "movaps    (%1,%0), %%xmm0 \n\t"

  2098.         "movaps  16(%1,%0), %%xmm1 \n\t"

  2099.         "mulps     (%2,%0), %%xmm0 \n\t"

  2100.         "mulps   16(%2,%0), %%xmm1 \n\t"

  2101.         "movaps  %%xmm0,   (%1,%0) \n\t"

  2102.         "movaps  %%xmm1, 16(%1,%0) \n\t"

  2103.         "sub  $32, %0 \n\t"

  2104.         "jge 1b \n\t"

  2105.         :"+r"(i)

  2106.         :"r"(dst), "r"(src)

  2107.         :"memory"

  2108.     );

  2109. }

  2110. 

  2111. static void vector_fmul_reverse_3dnow2(float *dst, const float *src0, const float *src1, int len){

  2112.     x86_reg i = len*4-16;

  2113.     __asm__ volatile(

  2114.         "1: \n\t"

  2115.         "pswapd   8(%1), %%mm0 \n\t"

  2116.         "pswapd    (%1), %%mm1 \n\t"

  2117.         "pfmul  (%3,%0), %%mm0 \n\t"

  2118.         "pfmul 8(%3,%0), %%mm1 \n\t"

  2119.         "movq  %%mm0,  (%2,%0) \n\t"

  2120.         "movq  %%mm1, 8(%2,%0) \n\t"

  2121.         "add   $16, %1 \n\t"

  2122.         "sub   $16, %0 \n\t"

  2123.         "jge   1b \n\t"

  2124.         :"+r"(i), "+r"(src1)

  2125.         :"r"(dst), "r"(src0)

  2126.     );

  2127.     __asm__ volatile("femms");

  2128. }

  2129. static void vector_fmul_reverse_sse(float *dst, const float *src0, const float *src1, int len){

  2130.     x86_reg i = len*4-32;

  2131.     __asm__ volatile(

  2132.         "1: \n\t"

  2133.         "movaps        16(%1), %%xmm0 \n\t"

  2134.         "movaps          (%1), %%xmm1 \n\t"

  2135.         "shufps $0x1b, %%xmm0, %%xmm0 \n\t"

  2136.         "shufps $0x1b, %%xmm1, %%xmm1 \n\t"

  2137.         "mulps        (%3,%0), %%xmm0 \n\t"

  2138.         "mulps      16(%3,%0), %%xmm1 \n\t"

  2139.         "movaps     %%xmm0,   (%2,%0) \n\t"

  2140.         "movaps     %%xmm1, 16(%2,%0) \n\t"

  2141.         "add    $32, %1 \n\t"

  2142.         "sub    $32, %0 \n\t"

  2143.         "jge    1b \n\t"

  2144.         :"+r"(i), "+r"(src1)

  2145.         :"r"(dst), "r"(src0)

  2146.     );

  2147. }

  2148. 

  2149. static void vector_fmul_add_3dnow(float *dst, const float *src0, const float *src1,

  2150.                                   const float *src2, int len){

  2151.     x86_reg i = (len-4)*4;

  2152.     __asm__ volatile(

  2153.         "1: \n\t"

  2154.         "movq    (%2,%0), %%mm0 \n\t"

  2155.         "movq   8(%2,%0), %%mm1 \n\t"

  2156.         "pfmul   (%3,%0), %%mm0 \n\t"

  2157.         "pfmul  8(%3,%0), %%mm1 \n\t"

  2158.         "pfadd   (%4,%0), %%mm0 \n\t"

  2159.         "pfadd  8(%4,%0), %%mm1 \n\t"

  2160.         "movq  %%mm0,   (%1,%0) \n\t"

  2161.         "movq  %%mm1,  8(%1,%0) \n\t"

  2162.         "sub  $16, %0 \n\t"

  2163.         "jge  1b \n\t"

  2164.         :"+r"(i)

  2165.         :"r"(dst), "r"(src0), "r"(src1), "r"(src2)

  2166.         :"memory"

  2167.     );

  2168.     __asm__ volatile("femms");

  2169. }

  2170. static void vector_fmul_add_sse(float *dst, const float *src0, const float *src1,

  2171.                                 const float *src2, int len){

  2172.     x86_reg i = (len-8)*4;

  2173.     __asm__ volatile(

  2174.         "1: \n\t"

  2175.         "movaps   (%2,%0), %%xmm0 \n\t"

  2176.         "movaps 16(%2,%0), %%xmm1 \n\t"

  2177.         "mulps    (%3,%0), %%xmm0 \n\t"

  2178.         "mulps  16(%3,%0), %%xmm1 \n\t"

  2179.         "addps    (%4,%0), %%xmm0 \n\t"

  2180.         "addps  16(%4,%0), %%xmm1 \n\t"

  2181.         "movaps %%xmm0,   (%1,%0) \n\t"

  2182.         "movaps %%xmm1, 16(%1,%0) \n\t"

  2183.         "sub  $32, %0 \n\t"

  2184.         "jge  1b \n\t"

  2185.         :"+r"(i)

  2186.         :"r"(dst), "r"(src0), "r"(src1), "r"(src2)

  2187.         :"memory"

  2188.     );

  2189. }

  2190. 

  2191. static void vector_fmul_window_3dnow2(float *dst, const float *src0, const float *src1,

  2192.                                       const float *win, float add_bias, int len){

  2193. #if HAVE_6REGS

  2194.     if(add_bias == 0){

  2195.         x86_reg i = -len*4;

  2196.         x86_reg j = len*4-8;

  2197.         __asm__ volatile(

  2198.             "1: \n"

  2199.             "pswapd  (%5,%1), %%mm1 \n"

  2200.             "movq    (%5,%0), %%mm0 \n"

  2201.             "pswapd  (%4,%1), %%mm5 \n"

  2202.             "movq    (%3,%0), %%mm4 \n"

  2203.             "movq      %%mm0, %%mm2 \n"

  2204.             "movq      %%mm1, %%mm3 \n"

  2205.             "pfmul     %%mm4, %%mm2 \n" // src0[len+i]*win[len+i]

  2206.             "pfmul     %%mm5, %%mm3 \n" // src1[    j]*win[len+j]

  2207.             "pfmul     %%mm4, %%mm1 \n" // src0[len+i]*win[len+j]

  2208.             "pfmul     %%mm5, %%mm0 \n" // src1[    j]*win[len+i]

  2209.             "pfadd     %%mm3, %%mm2 \n"

  2210.             "pfsub     %%mm0, %%mm1 \n"

  2211.             "pswapd    %%mm2, %%mm2 \n"

  2212.             "movq      %%mm1, (%2,%0) \n"

  2213.             "movq      %%mm2, (%2,%1) \n"

  2214.             "sub $8, %1 \n"

  2215.             "add $8, %0 \n"

  2216.             "jl 1b \n"

  2217.             "femms \n"

  2218.             :"+r"(i), "+r"(j)

  2219.             :"r"(dst+len), "r"(src0+len), "r"(src1), "r"(win+len)

  2220.         );

  2221.     }else

  2222. #endif

  2223.         ff_vector_fmul_window_c(dst, src0, src1, win, add_bias, len);

  2224. }

  2225. 

  2226. static void vector_fmul_window_sse(float *dst, const float *src0, const float *src1,

  2227.                                    const float *win, float add_bias, int len){

  2228. #if HAVE_6REGS

  2229.     if(add_bias == 0){

  2230.         x86_reg i = -len*4;

  2231.         x86_reg j = len*4-16;

  2232.         __asm__ volatile(

  2233.             "1: \n"

  2234.             "movaps       (%5,%1), %%xmm1 \n"

  2235.             "movaps       (%5,%0), %%xmm0 \n"

  2236.             "movaps       (%4,%1), %%xmm5 \n"

  2237.             "movaps       (%3,%0), %%xmm4 \n"

  2238.             "shufps $0x1b, %%xmm1, %%xmm1 \n"

  2239.             "shufps $0x1b, %%xmm5, %%xmm5 \n"

  2240.             "movaps        %%xmm0, %%xmm2 \n"

  2241.             "movaps        %%xmm1, %%xmm3 \n"

  2242.             "mulps         %%xmm4, %%xmm2 \n" // src0[len+i]*win[len+i]

  2243.             "mulps         %%xmm5, %%xmm3 \n" // src1[    j]*win[len+j]

  2244.             "mulps         %%xmm4, %%xmm1 \n" // src0[len+i]*win[len+j]

  2245.             "mulps         %%xmm5, %%xmm0 \n" // src1[    j]*win[len+i]

  2246.             "addps         %%xmm3, %%xmm2 \n"

  2247.             "subps         %%xmm0, %%xmm1 \n"

  2248.             "shufps $0x1b, %%xmm2, %%xmm2 \n"

  2249.             "movaps        %%xmm1, (%2,%0) \n"

  2250.             "movaps        %%xmm2, (%2,%1) \n"

  2251.             "sub $16, %1 \n"

  2252.             "add $16, %0 \n"

  2253.             "jl 1b \n"

  2254.             :"+r"(i), "+r"(j)

  2255.             :"r"(dst+len), "r"(src0+len), "r"(src1), "r"(win+len)

  2256.         );

  2257.     }else

  2258. #endif

  2259.         ff_vector_fmul_window_c(dst, src0, src1, win, add_bias, len);

  2260. }

  2261. 

  2262. static void int32_to_float_fmul_scalar_sse(float *dst, const int *src, float mul, int len)

  2263. {

  2264.     x86_reg i = -4*len;

  2265.     __asm__ volatile(

  2266.         "movss  %3, %%xmm4 \n"

  2267.         "shufps $0, %%xmm4, %%xmm4 \n"

  2268.         "1: \n"

  2269.         "cvtpi2ps   (%2,%0), %%xmm0 \n"

  2270.         "cvtpi2ps  8(%2,%0), %%xmm1 \n"

  2271.         "cvtpi2ps 16(%2,%0), %%xmm2 \n"

  2272.         "cvtpi2ps 24(%2,%0), %%xmm3 \n"

  2273.         "movlhps  %%xmm1,    %%xmm0 \n"

  2274.         "movlhps  %%xmm3,    %%xmm2 \n"

  2275.         "mulps    %%xmm4,    %%xmm0 \n"

  2276.         "mulps    %%xmm4,    %%xmm2 \n"

  2277.         "movaps   %%xmm0,   (%1,%0) \n"

  2278.         "movaps   %%xmm2, 16(%1,%0) \n"

  2279.         "add $32, %0 \n"

  2280.         "jl 1b \n"

  2281.         :"+r"(i)

  2282.         :"r"(dst+len), "r"(src+len), "m"(mul)

  2283.     );

  2284. }

  2285. 

  2286. static void int32_to_float_fmul_scalar_sse2(float *dst, const int *src, float mul, int len)

  2287. {

  2288.     x86_reg i = -4*len;

  2289.     __asm__ volatile(

  2290.         "movss  %3, %%xmm4 \n"

  2291.         "shufps $0, %%xmm4, %%xmm4 \n"

  2292.         "1: \n"

  2293.         "cvtdq2ps   (%2,%0), %%xmm0 \n"

  2294.         "cvtdq2ps 16(%2,%0), %%xmm1 \n"

  2295.         "mulps    %%xmm4,    %%xmm0 \n"

  2296.         "mulps    %%xmm4,    %%xmm1 \n"

  2297.         "movaps   %%xmm0,   (%1,%0) \n"

  2298.         "movaps   %%xmm1, 16(%1,%0) \n"

  2299.         "add $32, %0 \n"

  2300.         "jl 1b \n"

  2301.         :"+r"(i)

  2302.         :"r"(dst+len), "r"(src+len), "m"(mul)

  2303.     );

  2304. }

  2305. 

  2306. static void vector_clipf_sse(float *dst, const float *src, float min, float max,

  2307.                              int len)

  2308. {

  2309.     x86_reg i = (len-16)*4;

  2310.     __asm__ volatile(

  2311.         "movss  %3, %%xmm4 \n"

  2312.         "movss  %4, %%xmm5 \n"

  2313.         "shufps $0, %%xmm4, %%xmm4 \n"

  2314.         "shufps $0, %%xmm5, %%xmm5 \n"

  2315.         "1: \n\t"

  2316.         "movaps    (%2,%0), %%xmm0 \n\t" // 3/1 on intel

  2317.         "movaps  16(%2,%0), %%xmm1 \n\t"

  2318.         "movaps  32(%2,%0), %%xmm2 \n\t"

  2319.         "movaps  48(%2,%0), %%xmm3 \n\t"

  2320.         "maxps      %%xmm4, %%xmm0 \n\t"

  2321.         "maxps      %%xmm4, %%xmm1 \n\t"

  2322.         "maxps      %%xmm4, %%xmm2 \n\t"

  2323.         "maxps      %%xmm4, %%xmm3 \n\t"

  2324.         "minps      %%xmm5, %%xmm0 \n\t"

  2325.         "minps      %%xmm5, %%xmm1 \n\t"

  2326.         "minps      %%xmm5, %%xmm2 \n\t"

  2327.         "minps      %%xmm5, %%xmm3 \n\t"

  2328.         "movaps  %%xmm0,   (%1,%0) \n\t"

  2329.         "movaps  %%xmm1, 16(%1,%0) \n\t"

  2330.         "movaps  %%xmm2, 32(%1,%0) \n\t"

  2331.         "movaps  %%xmm3, 48(%1,%0) \n\t"

  2332.         "sub  $64, %0 \n\t"

  2333.         "jge 1b \n\t"

  2334.         :"+&r"(i)

  2335.         :"r"(dst), "r"(src), "m"(min), "m"(max)

  2336.         :"memory"

  2337.     );

  2338. }

  2339. 

  2340. static void float_to_int16_3dnow(int16_t *dst, const float *src, long len){

  2341.     x86_reg reglen = len;

  2342.     // not bit-exact: pf2id uses different rounding than C and SSE

  2343.     __asm__ volatile(

  2344.         "add        %0          , %0        \n\t"

  2345.         "lea         (%2,%0,2)  , %2        \n\t"

  2346.         "add        %0          , %1        \n\t"

  2347.         "neg        %0                      \n\t"

  2348.         "1:                                 \n\t"

  2349.         "pf2id       (%2,%0,2)  , %%mm0     \n\t"

  2350.         "pf2id      8(%2,%0,2)  , %%mm1     \n\t"

  2351.         "pf2id     16(%2,%0,2)  , %%mm2     \n\t"

  2352.         "pf2id     24(%2,%0,2)  , %%mm3     \n\t"

  2353.         "packssdw   %%mm1       , %%mm0     \n\t"

  2354.         "packssdw   %%mm3       , %%mm2     \n\t"

  2355.         "movq       %%mm0       ,  (%1,%0)  \n\t"

  2356.         "movq       %%mm2       , 8(%1,%0)  \n\t"

  2357.         "add        $16         , %0        \n\t"

  2358.         " js 1b                             \n\t"

  2359.         "femms                              \n\t"

  2360.         :"+r"(reglen), "+r"(dst), "+r"(src)

  2361.     );

  2362. }

  2363. static void float_to_int16_sse(int16_t *dst, const float *src, long len){

  2364.     x86_reg reglen = len;

  2365.     __asm__ volatile(

  2366.         "add        %0          , %0        \n\t"

  2367.         "lea         (%2,%0,2)  , %2        \n\t"

  2368.         "add        %0          , %1        \n\t"

  2369.         "neg        %0                      \n\t"

  2370.         "1:                                 \n\t"

  2371.         "cvtps2pi    (%2,%0,2)  , %%mm0     \n\t"

  2372.         "cvtps2pi   8(%2,%0,2)  , %%mm1     \n\t"

  2373.         "cvtps2pi  16(%2,%0,2)  , %%mm2     \n\t"

  2374.         "cvtps2pi  24(%2,%0,2)  , %%mm3     \n\t"

  2375.         "packssdw   %%mm1       , %%mm0     \n\t"

  2376.         "packssdw   %%mm3       , %%mm2     \n\t"

  2377.         "movq       %%mm0       ,  (%1,%0)  \n\t"

  2378.         "movq       %%mm2       , 8(%1,%0)  \n\t"

  2379.         "add        $16         , %0        \n\t"

  2380.         " js 1b                             \n\t"

  2381.         "emms                               \n\t"

  2382.         :"+r"(reglen), "+r"(dst), "+r"(src)

  2383.     );

  2384. }

  2385. 

  2386. static void float_to_int16_sse2(int16_t *dst, const float *src, long len){

  2387.     x86_reg reglen = len;

  2388.     __asm__ volatile(

  2389.         "add        %0          , %0        \n\t"

  2390.         "lea         (%2,%0,2)  , %2        \n\t"

  2391.         "add        %0          , %1        \n\t"

  2392.         "neg        %0                      \n\t"

  2393.         "1:                                 \n\t"

  2394.         "cvtps2dq    (%2,%0,2)  , %%xmm0    \n\t"

  2395.         "cvtps2dq  16(%2,%0,2)  , %%xmm1    \n\t"

  2396.         "packssdw   %%xmm1      , %%xmm0    \n\t"

  2397.         "movdqa     %%xmm0      ,  (%1,%0)  \n\t"

  2398.         "add        $16         , %0        \n\t"

  2399.         " js 1b                             \n\t"

  2400.         :"+r"(reglen), "+r"(dst), "+r"(src)

  2401.     );

  2402. }

  2403. 

  2404. void ff_vp3_idct_mmx(int16_t *input_data);

  2405. void ff_vp3_idct_put_mmx(uint8_t *dest, int line_size, DCTELEM *block);

  2406. void ff_vp3_idct_add_mmx(uint8_t *dest, int line_size, DCTELEM *block);

  2407. 

  2408. void ff_vp3_idct_dc_add_mmx2(uint8_t *dest, int line_size, const DCTELEM *block);

  2409. 

  2410. void ff_vp3_v_loop_filter_mmx2(uint8_t *src, int stride, int *bounding_values);

  2411. void ff_vp3_h_loop_filter_mmx2(uint8_t *src, int stride, int *bounding_values);

  2412. 

  2413. void ff_vp3_idct_sse2(int16_t *input_data);

  2414. void ff_vp3_idct_put_sse2(uint8_t *dest, int line_size, DCTELEM *block);

  2415. void ff_vp3_idct_add_sse2(uint8_t *dest, int line_size, DCTELEM *block);

  2416. 

  2417. void ff_float_to_int16_interleave6_sse(int16_t *dst, const float **src, int len);

  2418. void ff_float_to_int16_interleave6_3dnow(int16_t *dst, const float **src, int len);

  2419. void ff_float_to_int16_interleave6_3dn2(int16_t *dst, const float **src, int len);

  2420. int32_t ff_scalarproduct_int16_mmx2(const int16_t *v1, const int16_t *v2, int order, int shift);

  2421. int32_t ff_scalarproduct_int16_sse2(const int16_t *v1, const int16_t *v2, int order, int shift);

  2422. int32_t ff_scalarproduct_and_madd_int16_mmx2(int16_t *v1, const int16_t *v2, const int16_t *v3, int order, int mul);

  2423. int32_t ff_scalarproduct_and_madd_int16_sse2(int16_t *v1, const int16_t *v2, const int16_t *v3, int order, int mul);

  2424. int32_t ff_scalarproduct_and_madd_int16_ssse3(int16_t *v1, const int16_t *v2, const int16_t *v3, int order, int mul);

  2425. 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);

  2426. int  ff_add_hfyu_left_prediction_ssse3(uint8_t *dst, const uint8_t *src, int w, int left);

  2427. int  ff_add_hfyu_left_prediction_sse4(uint8_t *dst, const uint8_t *src, int w, int left);

  2428. 

  2429. #if !HAVE_YASM

  2430. #define ff_float_to_int16_interleave6_sse(a,b,c)   float_to_int16_interleave_misc_sse(a,b,c,6)

  2431. #define ff_float_to_int16_interleave6_3dnow(a,b,c) float_to_int16_interleave_misc_3dnow(a,b,c,6)

  2432. #define ff_float_to_int16_interleave6_3dn2(a,b,c)  float_to_int16_interleave_misc_3dnow(a,b,c,6)

  2433. #endif

  2434. #define ff_float_to_int16_interleave6_sse2 ff_float_to_int16_interleave6_sse

  2435. 

  2436. #define FLOAT_TO_INT16_INTERLEAVE(cpu, body) \

  2437. /* gcc pessimizes register allocation if this is in the same function as float_to_int16_interleave_sse2*/\

  2438. static av_noinline void float_to_int16_interleave_misc_##cpu(int16_t *dst, const float **src, long len, int channels){\

  2439.     DECLARE_ALIGNED(16, int16_t, tmp)[len];\

  2440.     int i,j,c;\

  2441.     for(c=0; c<channels; c++){\

  2442.         float_to_int16_##cpu(tmp, src[c], len);\

  2443.         for(i=0, j=c; i<len; i++, j+=channels)\

  2444.             dst[j] = tmp[i];\

  2445.     }\

  2446. }\

  2447. \

  2448. static void float_to_int16_interleave_##cpu(int16_t *dst, const float **src, long len, int channels){\

  2449.     if(channels==1)\

  2450.         float_to_int16_##cpu(dst, src[0], len);\

  2451.     else if(channels==2){\

  2452.         x86_reg reglen = len; \

  2453.         const float *src0 = src[0];\

  2454.         const float *src1 = src[1];\

  2455.         __asm__ volatile(\

  2456.             "shl $2, %0 \n"\

  2457.             "add %0, %1 \n"\

  2458.             "add %0, %2 \n"\

  2459.             "add %0, %3 \n"\

  2460.             "neg %0 \n"\

  2461.             body\

  2462.             :"+r"(reglen), "+r"(dst), "+r"(src0), "+r"(src1)\

  2463.         );\

  2464.     }else if(channels==6){\

  2465.         ff_float_to_int16_interleave6_##cpu(dst, src, len);\

  2466.     }else\

  2467.         float_to_int16_interleave_misc_##cpu(dst, src, len, channels);\

  2468. }

  2469. 

  2470. FLOAT_TO_INT16_INTERLEAVE(3dnow,

  2471.     "1:                         \n"

  2472.     "pf2id     (%2,%0), %%mm0   \n"

  2473.     "pf2id    8(%2,%0), %%mm1   \n"

  2474.     "pf2id     (%3,%0), %%mm2   \n"

  2475.     "pf2id    8(%3,%0), %%mm3   \n"

  2476.     "packssdw    %%mm1, %%mm0   \n"

  2477.     "packssdw    %%mm3, %%mm2   \n"

  2478.     "movq        %%mm0, %%mm1   \n"

  2479.     "punpcklwd   %%mm2, %%mm0   \n"

  2480.     "punpckhwd   %%mm2, %%mm1   \n"

  2481.     "movq        %%mm0,  (%1,%0)\n"

  2482.     "movq        %%mm1, 8(%1,%0)\n"

  2483.     "add $16, %0                \n"

  2484.     "js 1b                      \n"

  2485.     "femms                      \n"

  2486. )

  2487. 

  2488. FLOAT_TO_INT16_INTERLEAVE(sse,

  2489.     "1:                         \n"

  2490.     "cvtps2pi  (%2,%0), %%mm0   \n"

  2491.     "cvtps2pi 8(%2,%0), %%mm1   \n"

  2492.     "cvtps2pi  (%3,%0), %%mm2   \n"

  2493.     "cvtps2pi 8(%3,%0), %%mm3   \n"

  2494.     "packssdw    %%mm1, %%mm0   \n"

  2495.     "packssdw    %%mm3, %%mm2   \n"

  2496.     "movq        %%mm0, %%mm1   \n"

  2497.     "punpcklwd   %%mm2, %%mm0   \n"

  2498.     "punpckhwd   %%mm2, %%mm1   \n"

  2499.     "movq        %%mm0,  (%1,%0)\n"

  2500.     "movq        %%mm1, 8(%1,%0)\n"

  2501.     "add $16, %0                \n"

  2502.     "js 1b                      \n"

  2503.     "emms                       \n"

  2504. )

  2505. 

  2506. FLOAT_TO_INT16_INTERLEAVE(sse2,

  2507.     "1:                         \n"

  2508.     "cvtps2dq  (%2,%0), %%xmm0  \n"

  2509.     "cvtps2dq  (%3,%0), %%xmm1  \n"

  2510.     "packssdw   %%xmm1, %%xmm0  \n"

  2511.     "movhlps    %%xmm0, %%xmm1  \n"

  2512.     "punpcklwd  %%xmm1, %%xmm0  \n"

  2513.     "movdqa     %%xmm0, (%1,%0) \n"

  2514.     "add $16, %0                \n"

  2515.     "js 1b                      \n"

  2516. )

  2517. 

  2518. static void float_to_int16_interleave_3dn2(int16_t *dst, const float **src, long len, int channels){

  2519.     if(channels==6)

  2520.         ff_float_to_int16_interleave6_3dn2(dst, src, len);

  2521.     else

  2522.         float_to_int16_interleave_3dnow(dst, src, len, channels);

  2523. }

  2524. 

  2525. float ff_scalarproduct_float_sse(const float *v1, const float *v2, int order);

  2526. 

  2527. void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx)

  2528. {

  2529.     int mm_flags = av_get_cpu_flags();

  2530. 

  2531.     if (avctx->dsp_mask) {

  2532.         if (avctx->dsp_mask & AV_CPU_FLAG_FORCE)

  2533.             mm_flags |= (avctx->dsp_mask & 0xffff);

  2534.         else

  2535.             mm_flags &= ~(avctx->dsp_mask & 0xffff);

  2536.     }

  2537. 

  2538. #if 0

  2539.     av_log(avctx, AV_LOG_INFO, "libavcodec: CPU flags:");

  2540.     if (mm_flags & AV_CPU_FLAG_MMX)

  2541.         av_log(avctx, AV_LOG_INFO, " mmx");

  2542.     if (mm_flags & AV_CPU_FLAG_MMX2)

  2543.         av_log(avctx, AV_LOG_INFO, " mmx2");

  2544.     if (mm_flags & AV_CPU_FLAG_3DNOW)

  2545.         av_log(avctx, AV_LOG_INFO, " 3dnow");

  2546.     if (mm_flags & AV_CPU_FLAG_SSE)

  2547.         av_log(avctx, AV_LOG_INFO, " sse");

  2548.     if (mm_flags & AV_CPU_FLAG_SSE2)

  2549.         av_log(avctx, AV_LOG_INFO, " sse2");

  2550.     av_log(avctx, AV_LOG_INFO, "\n");

  2551. #endif

  2552. 

  2553.     if (mm_flags & AV_CPU_FLAG_MMX) {

  2554.         const int idct_algo= avctx->idct_algo;

  2555. 

  2556.         if(avctx->lowres==0){

  2557.             if(idct_algo==FF_IDCT_AUTO || idct_algo==FF_IDCT_SIMPLEMMX){

  2558.                 c->idct_put= ff_simple_idct_put_mmx;

  2559.                 c->idct_add= ff_simple_idct_add_mmx;

  2560.                 c->idct    = ff_simple_idct_mmx;

  2561.                 c->idct_permutation_type= FF_SIMPLE_IDCT_PERM;

  2562. #if CONFIG_GPL

  2563.             }else if(idct_algo==FF_IDCT_LIBMPEG2MMX){

  2564.                 if(mm_flags & AV_CPU_FLAG_MMX2){

  2565.                     c->idct_put= ff_libmpeg2mmx2_idct_put;

  2566.                     c->idct_add= ff_libmpeg2mmx2_idct_add;

  2567.                     c->idct    = ff_mmxext_idct;

  2568.                 }else{

  2569.                     c->idct_put= ff_libmpeg2mmx_idct_put;

  2570.                     c->idct_add= ff_libmpeg2mmx_idct_add;

  2571.                     c->idct    = ff_mmx_idct;

  2572.                 }

  2573.                 c->idct_permutation_type= FF_LIBMPEG2_IDCT_PERM;

  2574. #endif

  2575.             }else if((CONFIG_VP3_DECODER || CONFIG_VP5_DECODER || CONFIG_VP6_DECODER) &&

  2576.                      idct_algo==FF_IDCT_VP3 && HAVE_YASM){

  2577.                 if(mm_flags & AV_CPU_FLAG_SSE2){

  2578.                     c->idct_put= ff_vp3_idct_put_sse2;

  2579.                     c->idct_add= ff_vp3_idct_add_sse2;

  2580.                     c->idct    = ff_vp3_idct_sse2;

  2581.                     c->idct_permutation_type= FF_TRANSPOSE_IDCT_PERM;

  2582.                 }else{

  2583.                     c->idct_put= ff_vp3_idct_put_mmx;

  2584.                     c->idct_add= ff_vp3_idct_add_mmx;

  2585.                     c->idct    = ff_vp3_idct_mmx;

  2586.                     c->idct_permutation_type= FF_PARTTRANS_IDCT_PERM;

  2587.                 }

  2588.             }else if(idct_algo==FF_IDCT_CAVS){

  2589.                     c->idct_permutation_type= FF_TRANSPOSE_IDCT_PERM;

  2590.             }else if(idct_algo==FF_IDCT_XVIDMMX){

  2591.                 if(mm_flags & AV_CPU_FLAG_SSE2){

  2592.                     c->idct_put= ff_idct_xvid_sse2_put;

  2593.                     c->idct_add= ff_idct_xvid_sse2_add;

  2594.                     c->idct    = ff_idct_xvid_sse2;

  2595.                     c->idct_permutation_type= FF_SSE2_IDCT_PERM;

  2596.                 }else if(mm_flags & AV_CPU_FLAG_MMX2){

  2597.                     c->idct_put= ff_idct_xvid_mmx2_put;

  2598.                     c->idct_add= ff_idct_xvid_mmx2_add;

  2599.                     c->idct    = ff_idct_xvid_mmx2;

  2600.                 }else{

  2601.                     c->idct_put= ff_idct_xvid_mmx_put;

  2602.                     c->idct_add= ff_idct_xvid_mmx_add;

  2603.                     c->idct    = ff_idct_xvid_mmx;

  2604.                 }

  2605.             }

  2606.         }

  2607. 

  2608.         c->put_pixels_clamped = ff_put_pixels_clamped_mmx;

  2609.         c->put_signed_pixels_clamped = ff_put_signed_pixels_clamped_mmx;

  2610.         c->add_pixels_clamped = ff_add_pixels_clamped_mmx;

  2611.         c->clear_block  = clear_block_mmx;

  2612.         c->clear_blocks = clear_blocks_mmx;

  2613.         if ((mm_flags & AV_CPU_FLAG_SSE) &&

  2614.             !(CONFIG_MPEG_XVMC_DECODER && avctx->xvmc_acceleration > 1)){

  2615.             /* XvMCCreateBlocks() may not allocate 16-byte aligned blocks */

  2616.             c->clear_block  = clear_block_sse;

  2617.             c->clear_blocks = clear_blocks_sse;

  2618.         }

  2619. 

  2620. #define SET_HPEL_FUNCS(PFX, IDX, SIZE, CPU) \

  2621.         c->PFX ## _pixels_tab[IDX][0] = PFX ## _pixels ## SIZE ## _ ## CPU; \

  2622.         c->PFX ## _pixels_tab[IDX][1] = PFX ## _pixels ## SIZE ## _x2_ ## CPU; \

  2623.         c->PFX ## _pixels_tab[IDX][2] = PFX ## _pixels ## SIZE ## _y2_ ## CPU; \

  2624.         c->PFX ## _pixels_tab[IDX][3] = PFX ## _pixels ## SIZE ## _xy2_ ## CPU

  2625. 

  2626.         SET_HPEL_FUNCS(put, 0, 16, mmx);

  2627.         SET_HPEL_FUNCS(put_no_rnd, 0, 16, mmx);

  2628.         SET_HPEL_FUNCS(avg, 0, 16, mmx);

  2629.         SET_HPEL_FUNCS(avg_no_rnd, 0, 16, mmx);

  2630.         SET_HPEL_FUNCS(put, 1, 8, mmx);

  2631.         SET_HPEL_FUNCS(put_no_rnd, 1, 8, mmx);

  2632.         SET_HPEL_FUNCS(avg, 1, 8, mmx);

  2633.         SET_HPEL_FUNCS(avg_no_rnd, 1, 8, mmx);

  2634. 

  2635.         c->gmc= gmc_mmx;

  2636. 

  2637.         c->add_bytes= add_bytes_mmx;

  2638.         c->add_bytes_l2= add_bytes_l2_mmx;

  2639. 

  2640.         c->draw_edges = draw_edges_mmx;

  2641. 

  2642.         if (CONFIG_H263_DECODER || CONFIG_H263_ENCODER) {

  2643.             c->h263_v_loop_filter= h263_v_loop_filter_mmx;

  2644.             c->h263_h_loop_filter= h263_h_loop_filter_mmx;

  2645.         }

  2646. 

  2647. #if HAVE_YASM

  2648.         c->put_h264_chroma_pixels_tab[0]= ff_put_h264_chroma_mc8_mmx_rnd;

  2649.         c->put_h264_chroma_pixels_tab[1]= ff_put_h264_chroma_mc4_mmx;

  2650.         c->put_no_rnd_vc1_chroma_pixels_tab[0]= ff_put_vc1_chroma_mc8_mmx_nornd;

  2651. 

  2652.         c->put_rv40_chroma_pixels_tab[0]= ff_put_rv40_chroma_mc8_mmx;

  2653.         c->put_rv40_chroma_pixels_tab[1]= ff_put_rv40_chroma_mc4_mmx;

  2654. #endif

  2655. 

  2656.         if (mm_flags & AV_CPU_FLAG_MMX2) {

  2657.             c->prefetch = prefetch_mmx2;

  2658. 

  2659.             c->put_pixels_tab[0][1] = put_pixels16_x2_mmx2;

  2660.             c->put_pixels_tab[0][2] = put_pixels16_y2_mmx2;

  2661. 

  2662.             c->avg_pixels_tab[0][0] = avg_pixels16_mmx2;

  2663.             c->avg_pixels_tab[0][1] = avg_pixels16_x2_mmx2;

  2664.             c->avg_pixels_tab[0][2] = avg_pixels16_y2_mmx2;

  2665. 

  2666.             c->put_pixels_tab[1][1] = put_pixels8_x2_mmx2;

  2667.             c->put_pixels_tab[1][2] = put_pixels8_y2_mmx2;

  2668. 

  2669.             c->avg_pixels_tab[1][0] = avg_pixels8_mmx2;

  2670.             c->avg_pixels_tab[1][1] = avg_pixels8_x2_mmx2;

  2671.             c->avg_pixels_tab[1][2] = avg_pixels8_y2_mmx2;

  2672. 

  2673.             if(!(avctx->flags & CODEC_FLAG_BITEXACT)){

  2674.                 c->put_no_rnd_pixels_tab[0][1] = put_no_rnd_pixels16_x2_mmx2;

  2675.                 c->put_no_rnd_pixels_tab[0][2] = put_no_rnd_pixels16_y2_mmx2;

  2676.                 c->put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_mmx2;

  2677.                 c->put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_mmx2;

  2678.                 c->avg_pixels_tab[0][3] = avg_pixels16_xy2_mmx2;

  2679.                 c->avg_pixels_tab[1][3] = avg_pixels8_xy2_mmx2;

  2680. 

  2681.                 if (CONFIG_VP3_DECODER && HAVE_YASM) {

  2682.                     c->vp3_v_loop_filter= ff_vp3_v_loop_filter_mmx2;

  2683.                     c->vp3_h_loop_filter= ff_vp3_h_loop_filter_mmx2;

  2684.                 }

  2685.             }

  2686.             if (CONFIG_VP3_DECODER && HAVE_YASM) {

  2687.                 c->vp3_idct_dc_add = ff_vp3_idct_dc_add_mmx2;

  2688.             }

  2689. 

  2690.             if (CONFIG_VP3_DECODER

  2691.                 && (avctx->codec_id == CODEC_ID_VP3 || avctx->codec_id == CODEC_ID_THEORA)) {

  2692.                 c->put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_exact_mmx2;

  2693.                 c->put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_exact_mmx2;

  2694.             }

  2695. 

  2696. #define SET_QPEL_FUNCS(PFX, IDX, SIZE, CPU) \

  2697.             c->PFX ## _pixels_tab[IDX][ 0] = PFX ## SIZE ## _mc00_ ## CPU; \

  2698.             c->PFX ## _pixels_tab[IDX][ 1] = PFX ## SIZE ## _mc10_ ## CPU; \

  2699.             c->PFX ## _pixels_tab[IDX][ 2] = PFX ## SIZE ## _mc20_ ## CPU; \

  2700.             c->PFX ## _pixels_tab[IDX][ 3] = PFX ## SIZE ## _mc30_ ## CPU; \

  2701.             c->PFX ## _pixels_tab[IDX][ 4] = PFX ## SIZE ## _mc01_ ## CPU; \

  2702.             c->PFX ## _pixels_tab[IDX][ 5] = PFX ## SIZE ## _mc11_ ## CPU; \

  2703.             c->PFX ## _pixels_tab[IDX][ 6] = PFX ## SIZE ## _mc21_ ## CPU; \

  2704.             c->PFX ## _pixels_tab[IDX][ 7] = PFX ## SIZE ## _mc31_ ## CPU; \

  2705.             c->PFX ## _pixels_tab[IDX][ 8] = PFX ## SIZE ## _mc02_ ## CPU; \

  2706.             c->PFX ## _pixels_tab[IDX][ 9] = PFX ## SIZE ## _mc12_ ## CPU; \

  2707.             c->PFX ## _pixels_tab[IDX][10] = PFX ## SIZE ## _mc22_ ## CPU; \

  2708.             c->PFX ## _pixels_tab[IDX][11] = PFX ## SIZE ## _mc32_ ## CPU; \

  2709.             c->PFX ## _pixels_tab[IDX][12] = PFX ## SIZE ## _mc03_ ## CPU; \

  2710.             c->PFX ## _pixels_tab[IDX][13] = PFX ## SIZE ## _mc13_ ## CPU; \

  2711.             c->PFX ## _pixels_tab[IDX][14] = PFX ## SIZE ## _mc23_ ## CPU; \

  2712.             c->PFX ## _pixels_tab[IDX][15] = PFX ## SIZE ## _mc33_ ## CPU

  2713. 

  2714.             SET_QPEL_FUNCS(put_qpel, 0, 16, mmx2);

  2715.             SET_QPEL_FUNCS(put_qpel, 1, 8, mmx2);

  2716.             SET_QPEL_FUNCS(put_no_rnd_qpel, 0, 16, mmx2);

  2717.             SET_QPEL_FUNCS(put_no_rnd_qpel, 1, 8, mmx2);

  2718.             SET_QPEL_FUNCS(avg_qpel, 0, 16, mmx2);

  2719.             SET_QPEL_FUNCS(avg_qpel, 1, 8, mmx2);

  2720. 

  2721.             SET_QPEL_FUNCS(put_h264_qpel, 0, 16, mmx2);

  2722.             SET_QPEL_FUNCS(put_h264_qpel, 1, 8, mmx2);

  2723.             SET_QPEL_FUNCS(put_h264_qpel, 2, 4, mmx2);

  2724.             SET_QPEL_FUNCS(avg_h264_qpel, 0, 16, mmx2);

  2725.             SET_QPEL_FUNCS(avg_h264_qpel, 1, 8, mmx2);

  2726.             SET_QPEL_FUNCS(avg_h264_qpel, 2, 4, mmx2);

  2727. 

  2728.             SET_QPEL_FUNCS(put_2tap_qpel, 0, 16, mmx2);

  2729.             SET_QPEL_FUNCS(put_2tap_qpel, 1, 8, mmx2);

  2730.             SET_QPEL_FUNCS(avg_2tap_qpel, 0, 16, mmx2);

  2731.             SET_QPEL_FUNCS(avg_2tap_qpel, 1, 8, mmx2);

  2732. 

  2733. #if HAVE_YASM

  2734.             c->avg_rv40_chroma_pixels_tab[0]= ff_avg_rv40_chroma_mc8_mmx2;

  2735.             c->avg_rv40_chroma_pixels_tab[1]= ff_avg_rv40_chroma_mc4_mmx2;

  2736. 

  2737.             c->avg_no_rnd_vc1_chroma_pixels_tab[0]= ff_avg_vc1_chroma_mc8_mmx2_nornd;

  2738. 

  2739.             c->avg_h264_chroma_pixels_tab[0]= ff_avg_h264_chroma_mc8_mmx2_rnd;

  2740.             c->avg_h264_chroma_pixels_tab[1]= ff_avg_h264_chroma_mc4_mmx2;

  2741.             c->avg_h264_chroma_pixels_tab[2]= ff_avg_h264_chroma_mc2_mmx2;

  2742.             c->put_h264_chroma_pixels_tab[2]= ff_put_h264_chroma_mc2_mmx2;

  2743. 

  2744.             c->add_hfyu_median_prediction = ff_add_hfyu_median_prediction_mmx2;

  2745. #endif

  2746. #if HAVE_7REGS && HAVE_TEN_OPERANDS

  2747.             if( mm_flags&AV_CPU_FLAG_3DNOW )

  2748.                 c->add_hfyu_median_prediction = add_hfyu_median_prediction_cmov;

  2749. #endif

  2750. 

  2751.             if (CONFIG_VC1_DECODER)

  2752.                 ff_vc1dsp_init_mmx(c, avctx);

  2753. 

  2754.             c->add_png_paeth_prediction= add_png_paeth_prediction_mmx2;

  2755.         } else if (mm_flags & AV_CPU_FLAG_3DNOW) {

  2756.             c->prefetch = prefetch_3dnow;

  2757. 

  2758.             c->put_pixels_tab[0][1] = put_pixels16_x2_3dnow;

  2759.             c->put_pixels_tab[0][2] = put_pixels16_y2_3dnow;

  2760. 

  2761.             c->avg_pixels_tab[0][0] = avg_pixels16_3dnow;

  2762.             c->avg_pixels_tab[0][1] = avg_pixels16_x2_3dnow;

  2763.             c->avg_pixels_tab[0][2] = avg_pixels16_y2_3dnow;

  2764. 

  2765.             c->put_pixels_tab[1][1] = put_pixels8_x2_3dnow;

  2766.             c->put_pixels_tab[1][2] = put_pixels8_y2_3dnow;

  2767. 

  2768.             c->avg_pixels_tab[1][0] = avg_pixels8_3dnow;

  2769.             c->avg_pixels_tab[1][1] = avg_pixels8_x2_3dnow;

  2770.             c->avg_pixels_tab[1][2] = avg_pixels8_y2_3dnow;

  2771. 

  2772.             if(!(avctx->flags & CODEC_FLAG_BITEXACT)){

  2773.                 c->put_no_rnd_pixels_tab[0][1] = put_no_rnd_pixels16_x2_3dnow;

  2774.                 c->put_no_rnd_pixels_tab[0][2] = put_no_rnd_pixels16_y2_3dnow;

  2775.                 c->put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_3dnow;

  2776.                 c->put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_3dnow;

  2777.                 c->avg_pixels_tab[0][3] = avg_pixels16_xy2_3dnow;

  2778.                 c->avg_pixels_tab[1][3] = avg_pixels8_xy2_3dnow;

  2779.             }

  2780. 

  2781.             if (CONFIG_VP3_DECODER

  2782.                 && (avctx->codec_id == CODEC_ID_VP3 || avctx->codec_id == CODEC_ID_THEORA)) {

  2783.                 c->put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_exact_3dnow;

  2784.                 c->put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_exact_3dnow;

  2785.             }

  2786. 

  2787.             SET_QPEL_FUNCS(put_qpel, 0, 16, 3dnow);

  2788.             SET_QPEL_FUNCS(put_qpel, 1, 8, 3dnow);

  2789.             SET_QPEL_FUNCS(put_no_rnd_qpel, 0, 16, 3dnow);

  2790.             SET_QPEL_FUNCS(put_no_rnd_qpel, 1, 8, 3dnow);

  2791.             SET_QPEL_FUNCS(avg_qpel, 0, 16, 3dnow);

  2792.             SET_QPEL_FUNCS(avg_qpel, 1, 8, 3dnow);

  2793. 

  2794.             SET_QPEL_FUNCS(put_h264_qpel, 0, 16, 3dnow);

  2795.             SET_QPEL_FUNCS(put_h264_qpel, 1, 8, 3dnow);

  2796.             SET_QPEL_FUNCS(put_h264_qpel, 2, 4, 3dnow);

  2797.             SET_QPEL_FUNCS(avg_h264_qpel, 0, 16, 3dnow);

  2798.             SET_QPEL_FUNCS(avg_h264_qpel, 1, 8, 3dnow);

  2799.             SET_QPEL_FUNCS(avg_h264_qpel, 2, 4, 3dnow);

  2800. 

  2801.             SET_QPEL_FUNCS(put_2tap_qpel, 0, 16, 3dnow);

  2802.             SET_QPEL_FUNCS(put_2tap_qpel, 1, 8, 3dnow);

  2803.             SET_QPEL_FUNCS(avg_2tap_qpel, 0, 16, 3dnow);

  2804.             SET_QPEL_FUNCS(avg_2tap_qpel, 1, 8, 3dnow);

  2805. 

  2806. #if HAVE_YASM

  2807.             c->avg_h264_chroma_pixels_tab[0]= ff_avg_h264_chroma_mc8_3dnow_rnd;

  2808.             c->avg_h264_chroma_pixels_tab[1]= ff_avg_h264_chroma_mc4_3dnow;

  2809. 

  2810.             c->avg_no_rnd_vc1_chroma_pixels_tab[0]= ff_avg_vc1_chroma_mc8_3dnow_nornd;

  2811. 

  2812.             c->avg_rv40_chroma_pixels_tab[0]= ff_avg_rv40_chroma_mc8_3dnow;

  2813.             c->avg_rv40_chroma_pixels_tab[1]= ff_avg_rv40_chroma_mc4_3dnow;

  2814. #endif

  2815.         }

  2816. 

  2817. 

  2818. #define H264_QPEL_FUNCS(x, y, CPU)\

  2819.             c->put_h264_qpel_pixels_tab[0][x+y*4] = put_h264_qpel16_mc##x##y##_##CPU;\

  2820.             c->put_h264_qpel_pixels_tab[1][x+y*4] = put_h264_qpel8_mc##x##y##_##CPU;\

  2821.             c->avg_h264_qpel_pixels_tab[0][x+y*4] = avg_h264_qpel16_mc##x##y##_##CPU;\

  2822.             c->avg_h264_qpel_pixels_tab[1][x+y*4] = avg_h264_qpel8_mc##x##y##_##CPU;

  2823.         if((mm_flags & AV_CPU_FLAG_SSE2) && !(mm_flags & AV_CPU_FLAG_3DNOW)){

  2824.             // these functions are slower than mmx on AMD, but faster on Intel

  2825.             c->put_pixels_tab[0][0] = put_pixels16_sse2;

  2826.             c->put_no_rnd_pixels_tab[0][0] = put_pixels16_sse2;

  2827.             c->avg_pixels_tab[0][0] = avg_pixels16_sse2;

  2828.             H264_QPEL_FUNCS(0, 0, sse2);

  2829.         }

  2830.         if(mm_flags & AV_CPU_FLAG_SSE2){

  2831.             H264_QPEL_FUNCS(0, 1, sse2);

  2832.             H264_QPEL_FUNCS(0, 2, sse2);

  2833.             H264_QPEL_FUNCS(0, 3, sse2);

  2834.             H264_QPEL_FUNCS(1, 1, sse2);

  2835.             H264_QPEL_FUNCS(1, 2, sse2);

  2836.             H264_QPEL_FUNCS(1, 3, sse2);

  2837.             H264_QPEL_FUNCS(2, 1, sse2);

  2838.             H264_QPEL_FUNCS(2, 2, sse2);

  2839.             H264_QPEL_FUNCS(2, 3, sse2);

  2840.             H264_QPEL_FUNCS(3, 1, sse2);

  2841.             H264_QPEL_FUNCS(3, 2, sse2);

  2842.             H264_QPEL_FUNCS(3, 3, sse2);

  2843.         }

  2844. #if HAVE_SSSE3

  2845.         if(mm_flags & AV_CPU_FLAG_SSSE3){

  2846.             H264_QPEL_FUNCS(1, 0, ssse3);

  2847.             H264_QPEL_FUNCS(1, 1, ssse3);

  2848.             H264_QPEL_FUNCS(1, 2, ssse3);

  2849.             H264_QPEL_FUNCS(1, 3, ssse3);

  2850.             H264_QPEL_FUNCS(2, 0, ssse3);

  2851.             H264_QPEL_FUNCS(2, 1, ssse3);

  2852.             H264_QPEL_FUNCS(2, 2, ssse3);

  2853.             H264_QPEL_FUNCS(2, 3, ssse3);

  2854.             H264_QPEL_FUNCS(3, 0, ssse3);

  2855.             H264_QPEL_FUNCS(3, 1, ssse3);

  2856.             H264_QPEL_FUNCS(3, 2, ssse3);

  2857.             H264_QPEL_FUNCS(3, 3, ssse3);

  2858.             c->add_png_paeth_prediction= add_png_paeth_prediction_ssse3;

  2859. #if HAVE_YASM

  2860.             c->put_no_rnd_vc1_chroma_pixels_tab[0]= ff_put_vc1_chroma_mc8_ssse3_nornd;

  2861.             c->avg_no_rnd_vc1_chroma_pixels_tab[0]= ff_avg_vc1_chroma_mc8_ssse3_nornd;

  2862.             c->put_h264_chroma_pixels_tab[0]= ff_put_h264_chroma_mc8_ssse3_rnd;

  2863.             c->avg_h264_chroma_pixels_tab[0]= ff_avg_h264_chroma_mc8_ssse3_rnd;

  2864.             c->put_h264_chroma_pixels_tab[1]= ff_put_h264_chroma_mc4_ssse3;

  2865.             c->avg_h264_chroma_pixels_tab[1]= ff_avg_h264_chroma_mc4_ssse3;

  2866.             c->add_hfyu_left_prediction = ff_add_hfyu_left_prediction_ssse3;

  2867.             if (mm_flags & AV_CPU_FLAG_SSE4) // not really sse4, just slow on Conroe

  2868.                 c->add_hfyu_left_prediction = ff_add_hfyu_left_prediction_sse4;

  2869. #endif

  2870.         }

  2871. #endif

  2872. 

  2873.         if(mm_flags & AV_CPU_FLAG_3DNOW){

  2874.             c->vorbis_inverse_coupling = vorbis_inverse_coupling_3dnow;

  2875.             c->vector_fmul = vector_fmul_3dnow;

  2876.             if(!(avctx->flags & CODEC_FLAG_BITEXACT)){

  2877.                 c->float_to_int16 = float_to_int16_3dnow;

  2878.                 c->float_to_int16_interleave = float_to_int16_interleave_3dnow;

  2879.             }

  2880.         }

  2881.         if(mm_flags & AV_CPU_FLAG_3DNOWEXT){

  2882.             c->vector_fmul_reverse = vector_fmul_reverse_3dnow2;

  2883.             c->vector_fmul_window = vector_fmul_window_3dnow2;

  2884.             if(!(avctx->flags & CODEC_FLAG_BITEXACT)){

  2885.                 c->float_to_int16_interleave = float_to_int16_interleave_3dn2;

  2886.             }

  2887.         }

  2888.         if(mm_flags & AV_CPU_FLAG_MMX2){

  2889. #if HAVE_YASM

  2890.             c->scalarproduct_int16 = ff_scalarproduct_int16_mmx2;

  2891.             c->scalarproduct_and_madd_int16 = ff_scalarproduct_and_madd_int16_mmx2;

  2892. #endif

  2893.         }

  2894.         if(mm_flags & AV_CPU_FLAG_SSE){

  2895.             c->vorbis_inverse_coupling = vorbis_inverse_coupling_sse;

  2896.             c->ac3_downmix = ac3_downmix_sse;

  2897.             c->vector_fmul = vector_fmul_sse;

  2898.             c->vector_fmul_reverse = vector_fmul_reverse_sse;

  2899.             c->vector_fmul_add = vector_fmul_add_sse;

  2900.             c->vector_fmul_window = vector_fmul_window_sse;

  2901.             c->int32_to_float_fmul_scalar = int32_to_float_fmul_scalar_sse;

  2902.             c->vector_clipf = vector_clipf_sse;

  2903.             c->float_to_int16 = float_to_int16_sse;

  2904.             c->float_to_int16_interleave = float_to_int16_interleave_sse;

  2905. #if HAVE_YASM

  2906.             c->scalarproduct_float = ff_scalarproduct_float_sse;

  2907. #endif

  2908.         }

  2909.         if(mm_flags & AV_CPU_FLAG_3DNOW)

  2910.             c->vector_fmul_add = vector_fmul_add_3dnow; // faster than sse

  2911.         if(mm_flags & AV_CPU_FLAG_SSE2){

  2912.             c->int32_to_float_fmul_scalar = int32_to_float_fmul_scalar_sse2;

  2913.             c->float_to_int16 = float_to_int16_sse2;

  2914.             c->float_to_int16_interleave = float_to_int16_interleave_sse2;

  2915. #if HAVE_YASM

  2916.             c->scalarproduct_int16 = ff_scalarproduct_int16_sse2;

  2917.             c->scalarproduct_and_madd_int16 = ff_scalarproduct_and_madd_int16_sse2;

  2918. #endif

  2919.         }

  2920.         if((mm_flags & AV_CPU_FLAG_SSSE3) && !(mm_flags & (AV_CPU_FLAG_SSE42|AV_CPU_FLAG_3DNOW)) && HAVE_YASM) // cachesplit

  2921.             c->scalarproduct_and_madd_int16 = ff_scalarproduct_and_madd_int16_ssse3;

  2922.     }

  2923. 

  2924.     if (CONFIG_ENCODERS)

  2925.         dsputilenc_init_mmx(c, avctx);

  2926. 

  2927. #if 0

  2928.     // for speed testing

  2929.     get_pixels = just_return;

  2930.     put_pixels_clamped = just_return;

  2931.     add_pixels_clamped = just_return;

  2932. 

  2933.     pix_abs16x16 = just_return;

  2934.     pix_abs16x16_x2 = just_return;

  2935.     pix_abs16x16_y2 = just_return;

  2936.     pix_abs16x16_xy2 = just_return;

  2937. 

  2938.     put_pixels_tab[0] = just_return;

  2939.     put_pixels_tab[1] = just_return;

  2940.     put_pixels_tab[2] = just_return;

  2941.     put_pixels_tab[3] = just_return;

  2942. 

  2943.     put_no_rnd_pixels_tab[0] = just_return;

  2944.     put_no_rnd_pixels_tab[1] = just_return;

  2945.     put_no_rnd_pixels_tab[2] = just_return;

  2946.     put_no_rnd_pixels_tab[3] = just_return;

  2947. 

  2948.     avg_pixels_tab[0] = just_return;

  2949.     avg_pixels_tab[1] = just_return;

  2950.     avg_pixels_tab[2] = just_return;

  2951.     avg_pixels_tab[3] = just_return;

  2952. 

  2953.     avg_no_rnd_pixels_tab[0] = just_return;

  2954.     avg_no_rnd_pixels_tab[1] = just_return;

  2955.     avg_no_rnd_pixels_tab[2] = just_return;

  2956.     avg_no_rnd_pixels_tab[3] = just_return;

  2957. 

  2958.     //av_fdct = just_return;

  2959.     //ff_idct = just_return;

  2960. #endif

  2961. }