dsputil: x86: Convert mpeg4 qpel and dsputil avg to yasm
[ffmpeg.git] / libavcodec / x86 / dsputil_mmx.c
1 /*
2  * MMX optimized DSP utils
3  * Copyright (c) 2000, 2001 Fabrice Bellard
4  * Copyright (c) 2002-2004 Michael Niedermayer <michaelni@gmx.at>
5  *
6  * MMX optimization by Nick Kurshev <nickols_k@mail.ru>
7  *
8  * This file is part of Libav.
9  *
10  * Libav is free software; you can redistribute it and/or
11  * modify it under the terms of the GNU Lesser General Public
12  * License as published by the Free Software Foundation; either
13  * version 2.1 of the License, or (at your option) any later version.
14  *
15  * Libav is distributed in the hope that it will be useful,
16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
18  * Lesser General Public License for more details.
19  *
20  * You should have received a copy of the GNU Lesser General Public
21  * License along with Libav; if not, write to the Free Software
22  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
23  */
24
25 #include "libavutil/cpu.h"
26 #include "libavutil/x86/asm.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 xmm_reg,  ff_pw_1)    = { 0x0001000100010001ULL, 0x0001000100010001ULL };
42 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_2)    = { 0x0002000200020002ULL, 0x0002000200020002ULL };
43 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_3)    = { 0x0003000300030003ULL, 0x0003000300030003ULL };
44 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_4)    = { 0x0004000400040004ULL, 0x0004000400040004ULL };
45 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_5)    = { 0x0005000500050005ULL, 0x0005000500050005ULL };
46 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_8)    = { 0x0008000800080008ULL, 0x0008000800080008ULL };
47 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_9)    = { 0x0009000900090009ULL, 0x0009000900090009ULL };
48 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_15)   =   0x000F000F000F000FULL;
49 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_16)   = { 0x0010001000100010ULL, 0x0010001000100010ULL };
50 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_17)   = { 0x0011001100110011ULL, 0x0011001100110011ULL };
51 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_18)   = { 0x0012001200120012ULL, 0x0012001200120012ULL };
52 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_20)   =   0x0014001400140014ULL;
53 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_27)   = { 0x001B001B001B001BULL, 0x001B001B001B001BULL };
54 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_28)   = { 0x001C001C001C001CULL, 0x001C001C001C001CULL };
55 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_32)   = { 0x0020002000200020ULL, 0x0020002000200020ULL };
56 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_42)   =   0x002A002A002A002AULL;
57 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_53)   =   0x0035003500350035ULL;
58 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_63)   = { 0x003F003F003F003FULL, 0x003F003F003F003FULL };
59 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_64)   = { 0x0040004000400040ULL, 0x0040004000400040ULL };
60 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_96)   =   0x0060006000600060ULL;
61 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_128)  =   0x0080008000800080ULL;
62 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_255)  =   0x00ff00ff00ff00ffULL;
63 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_512)  = { 0x0200020002000200ULL, 0x0200020002000200ULL };
64 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_1019) = { 0x03FB03FB03FB03FBULL, 0x03FB03FB03FB03FBULL };
65
66 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_0)    = { 0x0000000000000000ULL, 0x0000000000000000ULL };
67 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_1)    = { 0x0101010101010101ULL, 0x0101010101010101ULL };
68 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_3)    = { 0x0303030303030303ULL, 0x0303030303030303ULL };
69 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_4)    = { 0x0404040404040404ULL, 0x0404040404040404ULL };
70 DECLARE_ALIGNED(8,  const uint64_t, ff_pb_7)    =   0x0707070707070707ULL;
71 DECLARE_ALIGNED(8,  const uint64_t, ff_pb_1F)   =   0x1F1F1F1F1F1F1F1FULL;
72 DECLARE_ALIGNED(8,  const uint64_t, ff_pb_3F)   =   0x3F3F3F3F3F3F3F3FULL;
73 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_80)   = { 0x8080808080808080ULL, 0x8080808080808080ULL };
74 DECLARE_ALIGNED(8,  const uint64_t, ff_pb_81)   =   0x8181818181818181ULL;
75 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_A1)   = { 0xA1A1A1A1A1A1A1A1ULL, 0xA1A1A1A1A1A1A1A1ULL };
76 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_F8)   = { 0xF8F8F8F8F8F8F8F8ULL, 0xF8F8F8F8F8F8F8F8ULL };
77 DECLARE_ALIGNED(8,  const uint64_t, ff_pb_FC)   =   0xFCFCFCFCFCFCFCFCULL;
78 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_FE)   = { 0xFEFEFEFEFEFEFEFEULL, 0xFEFEFEFEFEFEFEFEULL };
79
80 DECLARE_ALIGNED(16, const double, ff_pd_1)[2] = { 1.0, 1.0 };
81 DECLARE_ALIGNED(16, const double, ff_pd_2)[2] = { 2.0, 2.0 };
82
83
84 void ff_put_pixels8_x2_mmxext(uint8_t *block, const uint8_t *pixels,
85                               int line_size, int h);
86 void ff_put_pixels8_x2_3dnow(uint8_t *block, const uint8_t *pixels,
87                              int line_size, int h);
88 void ff_put_pixels8_l2_mmxext(uint8_t *dst, uint8_t *src1, uint8_t *src2,
89                               int dstStride, int src1Stride, int h);
90 void ff_put_no_rnd_pixels8_l2_mmxext(uint8_t *dst, uint8_t *src1,
91                                      uint8_t *src2, int dstStride,
92                                      int src1Stride, int h);
93 void ff_avg_pixels8_l2_mmxext(uint8_t *dst, uint8_t *src1, uint8_t *src2,
94                               int dstStride, int src1Stride, int h);
95 void ff_put_pixels16_x2_mmxext(uint8_t *block, const uint8_t *pixels,
96                                int line_size, int h);
97 void ff_put_pixels16_x2_3dnow(uint8_t *block, const uint8_t *pixels,
98                               int line_size, int h);
99 void ff_put_pixels16_l2_mmxext(uint8_t *dst, uint8_t *src1, uint8_t *src2,
100                                int dstStride, int src1Stride, int h);
101 void ff_avg_pixels16_l2_mmxext(uint8_t *dst, uint8_t *src1, uint8_t *src2,
102                                int dstStride, int src1Stride, int h);
103 void ff_put_no_rnd_pixels16_l2_mmxext(uint8_t *dst, uint8_t *src1, uint8_t *src2,
104                                       int dstStride, int src1Stride, int h);
105 void ff_put_no_rnd_pixels8_x2_mmxext(uint8_t *block, const uint8_t *pixels,
106                                      int line_size, int h);
107 void ff_put_no_rnd_pixels8_x2_3dnow(uint8_t *block, const uint8_t *pixels,
108                                     int line_size, int h);
109 void ff_put_no_rnd_pixels8_x2_exact_mmxext(uint8_t *block,
110                                            const uint8_t *pixels,
111                                            int line_size, int h);
112 void ff_put_no_rnd_pixels8_x2_exact_3dnow(uint8_t *block,
113                                           const uint8_t *pixels,
114                                           int line_size, int h);
115 void ff_put_pixels8_y2_mmxext(uint8_t *block, const uint8_t *pixels,
116                               int line_size, int h);
117 void ff_put_pixels8_y2_3dnow(uint8_t *block, const uint8_t *pixels,
118                              int line_size, int h);
119 void ff_put_no_rnd_pixels8_y2_mmxext(uint8_t *block, const uint8_t *pixels,
120                                      int line_size, int h);
121 void ff_put_no_rnd_pixels8_y2_3dnow(uint8_t *block, const uint8_t *pixels,
122                                     int line_size, int h);
123 void ff_put_no_rnd_pixels8_y2_exact_mmxext(uint8_t *block,
124                                            const uint8_t *pixels,
125                                            int line_size, int h);
126 void ff_put_no_rnd_pixels8_y2_exact_3dnow(uint8_t *block,
127                                           const uint8_t *pixels,
128                                           int line_size, int h);
129 void ff_avg_pixels8_mmxext(uint8_t *block, const uint8_t *pixels,
130                            int line_size, int h);
131 void ff_avg_pixels8_3dnow(uint8_t *block, const uint8_t *pixels,
132                           int line_size, int h);
133 void ff_avg_pixels8_x2_mmxext(uint8_t *block, const uint8_t *pixels,
134                               int line_size, int h);
135 void ff_avg_pixels8_x2_3dnow(uint8_t *block, const uint8_t *pixels,
136                              int line_size, int h);
137 void ff_avg_pixels8_y2_mmxext(uint8_t *block, const uint8_t *pixels,
138                               int line_size, int h);
139 void ff_avg_pixels8_y2_3dnow(uint8_t *block, const uint8_t *pixels,
140                              int line_size, int h);
141 void ff_avg_pixels8_xy2_mmxext(uint8_t *block, const uint8_t *pixels,
142                                int line_size, int h);
143 void ff_avg_pixels8_xy2_3dnow(uint8_t *block, const uint8_t *pixels,
144                               int line_size, int h);
145
146 void ff_put_pixels8_mmxext(uint8_t *block, const uint8_t *pixels, int line_size, int h);
147 static void ff_put_pixels16_mmxext(uint8_t *block, const uint8_t *pixels,
148                                    int line_size, int h)
149 {
150     ff_put_pixels8_mmxext(block,     pixels,     line_size, h);
151     ff_put_pixels8_mmxext(block + 8, pixels + 8, line_size, h);
152 }
153
154 void ff_put_mpeg4_qpel16_h_lowpass_mmxext(uint8_t *dst, uint8_t *src,
155                                          int dstStride, int srcStride, int h);
156 void ff_avg_mpeg4_qpel16_h_lowpass_mmxext(uint8_t *dst, uint8_t *src,
157                                          int dstStride, int srcStride, int h);
158 void ff_put_no_rnd_mpeg4_qpel16_h_lowpass_mmxext(uint8_t *dst, uint8_t *src,
159                                                  int dstStride, int srcStride,
160                                                  int h);
161 void ff_put_mpeg4_qpel8_h_lowpass_mmxext(uint8_t *dst, uint8_t *src,
162                                         int dstStride, int srcStride, int h);
163 void ff_avg_mpeg4_qpel8_h_lowpass_mmxext(uint8_t *dst, uint8_t *src,
164                                         int dstStride, int srcStride, int h);
165 void ff_put_no_rnd_mpeg4_qpel8_h_lowpass_mmxext(uint8_t *dst, uint8_t *src,
166                                                 int dstStride, int srcStride,
167                                                 int h);
168 void ff_put_mpeg4_qpel16_v_lowpass_mmxext(uint8_t *dst, uint8_t *src,
169                                          int dstStride, int srcStride);
170 void ff_avg_mpeg4_qpel16_v_lowpass_mmxext(uint8_t *dst, uint8_t *src,
171                                          int dstStride, int srcStride);
172 void ff_put_no_rnd_mpeg4_qpel16_v_lowpass_mmxext(uint8_t *dst, uint8_t *src,
173                                                  int dstStride, int srcStride);
174 void ff_put_mpeg4_qpel8_v_lowpass_mmxext(uint8_t *dst, uint8_t *src,
175                                         int dstStride, int srcStride);
176 void ff_avg_mpeg4_qpel8_v_lowpass_mmxext(uint8_t *dst, uint8_t *src,
177                                         int dstStride, int srcStride);
178 void ff_put_no_rnd_mpeg4_qpel8_v_lowpass_mmxext(uint8_t *dst, uint8_t *src,
179                                                 int dstStride, int srcStride);
180 #define ff_put_no_rnd_pixels16_mmxext ff_put_pixels16_mmxext
181 #define ff_put_no_rnd_pixels8_mmxext ff_put_pixels8_mmxext
182
183
184 #if HAVE_INLINE_ASM
185
186 #define JUMPALIGN()     __asm__ volatile (".p2align 3"::)
187 #define MOVQ_ZERO(regd) __asm__ volatile ("pxor %%"#regd", %%"#regd ::)
188
189 #define MOVQ_BFE(regd)                                  \
190     __asm__ volatile (                                  \
191         "pcmpeqd %%"#regd", %%"#regd"   \n\t"           \
192         "paddb   %%"#regd", %%"#regd"   \n\t" ::)
193
194 #ifndef PIC
195 #define MOVQ_BONE(regd) __asm__ volatile ("movq %0, %%"#regd" \n\t" :: "m"(ff_bone))
196 #define MOVQ_WTWO(regd) __asm__ volatile ("movq %0, %%"#regd" \n\t" :: "m"(ff_wtwo))
197 #else
198 // for shared library it's better to use this way for accessing constants
199 // pcmpeqd -> -1
200 #define MOVQ_BONE(regd)                                 \
201     __asm__ volatile (                                  \
202         "pcmpeqd  %%"#regd", %%"#regd"  \n\t"           \
203         "psrlw          $15, %%"#regd"  \n\t"           \
204         "packuswb %%"#regd", %%"#regd"  \n\t" ::)
205
206 #define MOVQ_WTWO(regd)                                 \
207     __asm__ volatile (                                  \
208         "pcmpeqd %%"#regd", %%"#regd"   \n\t"           \
209         "psrlw         $15, %%"#regd"   \n\t"           \
210         "psllw          $1, %%"#regd"   \n\t"::)
211
212 #endif
213
214 // using regr as temporary and for the output result
215 // first argument is unmodifed and second is trashed
216 // regfe is supposed to contain 0xfefefefefefefefe
217 #define PAVGB_MMX_NO_RND(rega, regb, regr, regfe)                \
218     "movq   "#rega", "#regr"            \n\t"                    \
219     "pand   "#regb", "#regr"            \n\t"                    \
220     "pxor   "#rega", "#regb"            \n\t"                    \
221     "pand  "#regfe", "#regb"            \n\t"                    \
222     "psrlq       $1, "#regb"            \n\t"                    \
223     "paddb  "#regb", "#regr"            \n\t"
224
225 #define PAVGB_MMX(rega, regb, regr, regfe)                       \
226     "movq   "#rega", "#regr"            \n\t"                    \
227     "por    "#regb", "#regr"            \n\t"                    \
228     "pxor   "#rega", "#regb"            \n\t"                    \
229     "pand  "#regfe", "#regb"            \n\t"                    \
230     "psrlq       $1, "#regb"            \n\t"                    \
231     "psubb  "#regb", "#regr"            \n\t"
232
233 // mm6 is supposed to contain 0xfefefefefefefefe
234 #define PAVGBP_MMX_NO_RND(rega, regb, regr,  regc, regd, regp)   \
235     "movq  "#rega", "#regr"             \n\t"                    \
236     "movq  "#regc", "#regp"             \n\t"                    \
237     "pand  "#regb", "#regr"             \n\t"                    \
238     "pand  "#regd", "#regp"             \n\t"                    \
239     "pxor  "#rega", "#regb"             \n\t"                    \
240     "pxor  "#regc", "#regd"             \n\t"                    \
241     "pand    %%mm6, "#regb"             \n\t"                    \
242     "pand    %%mm6, "#regd"             \n\t"                    \
243     "psrlq      $1, "#regb"             \n\t"                    \
244     "psrlq      $1, "#regd"             \n\t"                    \
245     "paddb "#regb", "#regr"             \n\t"                    \
246     "paddb "#regd", "#regp"             \n\t"
247
248 #define PAVGBP_MMX(rega, regb, regr, regc, regd, regp)           \
249     "movq  "#rega", "#regr"             \n\t"                    \
250     "movq  "#regc", "#regp"             \n\t"                    \
251     "por   "#regb", "#regr"             \n\t"                    \
252     "por   "#regd", "#regp"             \n\t"                    \
253     "pxor  "#rega", "#regb"             \n\t"                    \
254     "pxor  "#regc", "#regd"             \n\t"                    \
255     "pand    %%mm6, "#regb"             \n\t"                    \
256     "pand    %%mm6, "#regd"             \n\t"                    \
257     "psrlq      $1, "#regd"             \n\t"                    \
258     "psrlq      $1, "#regb"             \n\t"                    \
259     "psubb "#regb", "#regr"             \n\t"                    \
260     "psubb "#regd", "#regp"             \n\t"
261
262 /***********************************/
263 /* MMX no rounding */
264 #define NO_RND 1
265 #define DEF(x, y) x ## _no_rnd_ ## y ## _mmx
266 #define SET_RND  MOVQ_WONE
267 #define PAVGBP(a, b, c, d, e, f)        PAVGBP_MMX_NO_RND(a, b, c, d, e, f)
268 #define PAVGB(a, b, c, e)               PAVGB_MMX_NO_RND(a, b, c, e)
269 #define OP_AVG(a, b, c, e)              PAVGB_MMX(a, b, c, e)
270
271 #include "dsputil_rnd_template.c"
272
273 #undef DEF
274 #undef SET_RND
275 #undef PAVGBP
276 #undef PAVGB
277 #undef NO_RND
278 /***********************************/
279 /* MMX rounding */
280
281 #define DEF(x, y) x ## _ ## y ## _mmx
282 #define SET_RND  MOVQ_WTWO
283 #define PAVGBP(a, b, c, d, e, f)        PAVGBP_MMX(a, b, c, d, e, f)
284 #define PAVGB(a, b, c, e)               PAVGB_MMX(a, b, c, e)
285
286 #include "dsputil_rnd_template.c"
287
288 #undef DEF
289 #undef SET_RND
290 #undef PAVGBP
291 #undef PAVGB
292 #undef OP_AVG
293
294 #endif /* HAVE_INLINE_ASM */
295
296
297 #if HAVE_YASM
298 #define ff_put_pixels8_mmx ff_put_pixels8_mmxext
299
300 /***********************************/
301 /* 3Dnow specific */
302
303 #define DEF(x) x ## _3dnow
304
305 #include "dsputil_avg_template.c"
306
307 #undef DEF
308
309 /***********************************/
310 /* MMXEXT specific */
311
312 #define DEF(x) x ## _mmxext
313
314 #include "dsputil_avg_template.c"
315
316 #undef DEF
317
318 #endif /* HAVE_YASM */
319
320
321 #if HAVE_INLINE_ASM
322 #define put_no_rnd_pixels16_mmx put_pixels16_mmx
323 #define put_no_rnd_pixels8_mmx put_pixels8_mmx
324 #define put_pixels16_mmxext put_pixels16_mmx
325 #define put_pixels8_mmxext put_pixels8_mmx
326 #define put_pixels4_mmxext put_pixels4_mmx
327 #define put_no_rnd_pixels16_mmxext put_no_rnd_pixels16_mmx
328 #define put_no_rnd_pixels8_mmxext put_no_rnd_pixels8_mmx
329
330 /***********************************/
331 /* standard MMX */
332
333 void ff_put_pixels_clamped_mmx(const int16_t *block, uint8_t *pixels,
334                                int line_size)
335 {
336     const int16_t *p;
337     uint8_t *pix;
338
339     /* read the pixels */
340     p   = block;
341     pix = pixels;
342     /* unrolled loop */
343     __asm__ volatile (
344         "movq      (%3), %%mm0          \n\t"
345         "movq     8(%3), %%mm1          \n\t"
346         "movq    16(%3), %%mm2          \n\t"
347         "movq    24(%3), %%mm3          \n\t"
348         "movq    32(%3), %%mm4          \n\t"
349         "movq    40(%3), %%mm5          \n\t"
350         "movq    48(%3), %%mm6          \n\t"
351         "movq    56(%3), %%mm7          \n\t"
352         "packuswb %%mm1, %%mm0          \n\t"
353         "packuswb %%mm3, %%mm2          \n\t"
354         "packuswb %%mm5, %%mm4          \n\t"
355         "packuswb %%mm7, %%mm6          \n\t"
356         "movq     %%mm0, (%0)           \n\t"
357         "movq     %%mm2, (%0, %1)       \n\t"
358         "movq     %%mm4, (%0, %1, 2)    \n\t"
359         "movq     %%mm6, (%0, %2)       \n\t"
360         :: "r"(pix), "r"((x86_reg)line_size), "r"((x86_reg)line_size * 3),
361            "r"(p)
362         : "memory");
363     pix += line_size * 4;
364     p   += 32;
365
366     // if here would be an exact copy of the code above
367     // compiler would generate some very strange code
368     // thus using "r"
369     __asm__ volatile (
370         "movq       (%3), %%mm0         \n\t"
371         "movq      8(%3), %%mm1         \n\t"
372         "movq     16(%3), %%mm2         \n\t"
373         "movq     24(%3), %%mm3         \n\t"
374         "movq     32(%3), %%mm4         \n\t"
375         "movq     40(%3), %%mm5         \n\t"
376         "movq     48(%3), %%mm6         \n\t"
377         "movq     56(%3), %%mm7         \n\t"
378         "packuswb  %%mm1, %%mm0         \n\t"
379         "packuswb  %%mm3, %%mm2         \n\t"
380         "packuswb  %%mm5, %%mm4         \n\t"
381         "packuswb  %%mm7, %%mm6         \n\t"
382         "movq      %%mm0, (%0)          \n\t"
383         "movq      %%mm2, (%0, %1)      \n\t"
384         "movq      %%mm4, (%0, %1, 2)   \n\t"
385         "movq      %%mm6, (%0, %2)      \n\t"
386         :: "r"(pix), "r"((x86_reg)line_size), "r"((x86_reg)line_size * 3), "r"(p)
387         : "memory");
388 }
389
390 #define put_signed_pixels_clamped_mmx_half(off)             \
391     "movq          "#off"(%2), %%mm1        \n\t"           \
392     "movq     16 + "#off"(%2), %%mm2        \n\t"           \
393     "movq     32 + "#off"(%2), %%mm3        \n\t"           \
394     "movq     48 + "#off"(%2), %%mm4        \n\t"           \
395     "packsswb  8 + "#off"(%2), %%mm1        \n\t"           \
396     "packsswb 24 + "#off"(%2), %%mm2        \n\t"           \
397     "packsswb 40 + "#off"(%2), %%mm3        \n\t"           \
398     "packsswb 56 + "#off"(%2), %%mm4        \n\t"           \
399     "paddb              %%mm0, %%mm1        \n\t"           \
400     "paddb              %%mm0, %%mm2        \n\t"           \
401     "paddb              %%mm0, %%mm3        \n\t"           \
402     "paddb              %%mm0, %%mm4        \n\t"           \
403     "movq               %%mm1, (%0)         \n\t"           \
404     "movq               %%mm2, (%0, %3)     \n\t"           \
405     "movq               %%mm3, (%0, %3, 2)  \n\t"           \
406     "movq               %%mm4, (%0, %1)     \n\t"
407
408 void ff_put_signed_pixels_clamped_mmx(const int16_t *block, uint8_t *pixels,
409                                       int line_size)
410 {
411     x86_reg line_skip = line_size;
412     x86_reg line_skip3;
413
414     __asm__ volatile (
415         "movq "MANGLE(ff_pb_80)", %%mm0     \n\t"
416         "lea         (%3, %3, 2), %1        \n\t"
417         put_signed_pixels_clamped_mmx_half(0)
418         "lea         (%0, %3, 4), %0        \n\t"
419         put_signed_pixels_clamped_mmx_half(64)
420         : "+&r"(pixels), "=&r"(line_skip3)
421         : "r"(block), "r"(line_skip)
422         : "memory");
423 }
424
425 void ff_add_pixels_clamped_mmx(const int16_t *block, uint8_t *pixels,
426                                int line_size)
427 {
428     const int16_t *p;
429     uint8_t *pix;
430     int i;
431
432     /* read the pixels */
433     p   = block;
434     pix = pixels;
435     MOVQ_ZERO(mm7);
436     i = 4;
437     do {
438         __asm__ volatile (
439             "movq        (%2), %%mm0    \n\t"
440             "movq       8(%2), %%mm1    \n\t"
441             "movq      16(%2), %%mm2    \n\t"
442             "movq      24(%2), %%mm3    \n\t"
443             "movq          %0, %%mm4    \n\t"
444             "movq          %1, %%mm6    \n\t"
445             "movq       %%mm4, %%mm5    \n\t"
446             "punpcklbw  %%mm7, %%mm4    \n\t"
447             "punpckhbw  %%mm7, %%mm5    \n\t"
448             "paddsw     %%mm4, %%mm0    \n\t"
449             "paddsw     %%mm5, %%mm1    \n\t"
450             "movq       %%mm6, %%mm5    \n\t"
451             "punpcklbw  %%mm7, %%mm6    \n\t"
452             "punpckhbw  %%mm7, %%mm5    \n\t"
453             "paddsw     %%mm6, %%mm2    \n\t"
454             "paddsw     %%mm5, %%mm3    \n\t"
455             "packuswb   %%mm1, %%mm0    \n\t"
456             "packuswb   %%mm3, %%mm2    \n\t"
457             "movq       %%mm0, %0       \n\t"
458             "movq       %%mm2, %1       \n\t"
459             : "+m"(*pix), "+m"(*(pix + line_size))
460             : "r"(p)
461             : "memory");
462         pix += line_size * 2;
463         p   += 16;
464     } while (--i);
465 }
466
467 static void put_pixels8_mmx(uint8_t *block, const uint8_t *pixels,
468                             int line_size, int h)
469 {
470     __asm__ volatile (
471         "lea   (%3, %3), %%"REG_a"      \n\t"
472         ".p2align     3                 \n\t"
473         "1:                             \n\t"
474         "movq  (%1    ), %%mm0          \n\t"
475         "movq  (%1, %3), %%mm1          \n\t"
476         "movq     %%mm0, (%2)           \n\t"
477         "movq     %%mm1, (%2, %3)       \n\t"
478         "add  %%"REG_a", %1             \n\t"
479         "add  %%"REG_a", %2             \n\t"
480         "movq  (%1    ), %%mm0          \n\t"
481         "movq  (%1, %3), %%mm1          \n\t"
482         "movq     %%mm0, (%2)           \n\t"
483         "movq     %%mm1, (%2, %3)       \n\t"
484         "add  %%"REG_a", %1             \n\t"
485         "add  %%"REG_a", %2             \n\t"
486         "subl        $4, %0             \n\t"
487         "jnz         1b                 \n\t"
488         : "+g"(h), "+r"(pixels),  "+r"(block)
489         : "r"((x86_reg)line_size)
490         : "%"REG_a, "memory"
491         );
492 }
493
494 static void put_pixels16_mmx(uint8_t *block, const uint8_t *pixels,
495                              int line_size, int h)
496 {
497     __asm__ volatile (
498         "lea   (%3, %3), %%"REG_a"      \n\t"
499         ".p2align     3                 \n\t"
500         "1:                             \n\t"
501         "movq  (%1    ), %%mm0          \n\t"
502         "movq 8(%1    ), %%mm4          \n\t"
503         "movq  (%1, %3), %%mm1          \n\t"
504         "movq 8(%1, %3), %%mm5          \n\t"
505         "movq     %%mm0,  (%2)          \n\t"
506         "movq     %%mm4, 8(%2)          \n\t"
507         "movq     %%mm1,  (%2, %3)      \n\t"
508         "movq     %%mm5, 8(%2, %3)      \n\t"
509         "add  %%"REG_a", %1             \n\t"
510         "add  %%"REG_a", %2             \n\t"
511         "movq  (%1    ), %%mm0          \n\t"
512         "movq 8(%1    ), %%mm4          \n\t"
513         "movq  (%1, %3), %%mm1          \n\t"
514         "movq 8(%1, %3), %%mm5          \n\t"
515         "movq     %%mm0,  (%2)          \n\t"
516         "movq     %%mm4, 8(%2)          \n\t"
517         "movq     %%mm1,  (%2, %3)      \n\t"
518         "movq     %%mm5, 8(%2, %3)      \n\t"
519         "add  %%"REG_a", %1             \n\t"
520         "add  %%"REG_a", %2             \n\t"
521         "subl        $4, %0             \n\t"
522         "jnz         1b                 \n\t"
523         : "+g"(h), "+r"(pixels),  "+r"(block)
524         : "r"((x86_reg)line_size)
525         : "%"REG_a, "memory"
526         );
527 }
528
529 #define CLEAR_BLOCKS(name, n)                           \
530 static void name(int16_t *blocks)                       \
531 {                                                       \
532     __asm__ volatile (                                  \
533         "pxor %%mm7, %%mm7              \n\t"           \
534         "mov     %1,        %%"REG_a"   \n\t"           \
535         "1:                             \n\t"           \
536         "movq %%mm7,   (%0, %%"REG_a")  \n\t"           \
537         "movq %%mm7,  8(%0, %%"REG_a")  \n\t"           \
538         "movq %%mm7, 16(%0, %%"REG_a")  \n\t"           \
539         "movq %%mm7, 24(%0, %%"REG_a")  \n\t"           \
540         "add    $32, %%"REG_a"          \n\t"           \
541         "js      1b                     \n\t"           \
542         :: "r"(((uint8_t *)blocks) + 128 * n),          \
543            "i"(-128 * n)                                \
544         : "%"REG_a                                      \
545         );                                              \
546 }
547 CLEAR_BLOCKS(clear_blocks_mmx, 6)
548 CLEAR_BLOCKS(clear_block_mmx, 1)
549
550 static void clear_block_sse(int16_t *block)
551 {
552     __asm__ volatile (
553         "xorps  %%xmm0, %%xmm0          \n"
554         "movaps %%xmm0,    (%0)         \n"
555         "movaps %%xmm0,  16(%0)         \n"
556         "movaps %%xmm0,  32(%0)         \n"
557         "movaps %%xmm0,  48(%0)         \n"
558         "movaps %%xmm0,  64(%0)         \n"
559         "movaps %%xmm0,  80(%0)         \n"
560         "movaps %%xmm0,  96(%0)         \n"
561         "movaps %%xmm0, 112(%0)         \n"
562         :: "r"(block)
563         : "memory"
564     );
565 }
566
567 static void clear_blocks_sse(int16_t *blocks)
568 {
569     __asm__ volatile (
570         "xorps  %%xmm0, %%xmm0              \n"
571         "mov        %1,         %%"REG_a"   \n"
572         "1:                                 \n"
573         "movaps %%xmm0,    (%0, %%"REG_a")  \n"
574         "movaps %%xmm0,  16(%0, %%"REG_a")  \n"
575         "movaps %%xmm0,  32(%0, %%"REG_a")  \n"
576         "movaps %%xmm0,  48(%0, %%"REG_a")  \n"
577         "movaps %%xmm0,  64(%0, %%"REG_a")  \n"
578         "movaps %%xmm0,  80(%0, %%"REG_a")  \n"
579         "movaps %%xmm0,  96(%0, %%"REG_a")  \n"
580         "movaps %%xmm0, 112(%0, %%"REG_a")  \n"
581         "add      $128,         %%"REG_a"   \n"
582         "js         1b                      \n"
583         :: "r"(((uint8_t *)blocks) + 128 * 6),
584            "i"(-128 * 6)
585         : "%"REG_a
586     );
587 }
588
589 static void add_bytes_mmx(uint8_t *dst, uint8_t *src, int w)
590 {
591     x86_reg i = 0;
592     __asm__ volatile (
593         "jmp          2f                \n\t"
594         "1:                             \n\t"
595         "movq   (%1, %0), %%mm0         \n\t"
596         "movq   (%2, %0), %%mm1         \n\t"
597         "paddb     %%mm0, %%mm1         \n\t"
598         "movq      %%mm1, (%2, %0)      \n\t"
599         "movq  8(%1, %0), %%mm0         \n\t"
600         "movq  8(%2, %0), %%mm1         \n\t"
601         "paddb     %%mm0, %%mm1         \n\t"
602         "movq      %%mm1, 8(%2, %0)     \n\t"
603         "add         $16, %0            \n\t"
604         "2:                             \n\t"
605         "cmp          %3, %0            \n\t"
606         "js           1b                \n\t"
607         : "+r"(i)
608         : "r"(src), "r"(dst), "r"((x86_reg)w - 15)
609     );
610     for ( ; i < w; i++)
611         dst[i + 0] += src[i + 0];
612 }
613
614 #if HAVE_7REGS
615 static void add_hfyu_median_prediction_cmov(uint8_t *dst, const uint8_t *top,
616                                             const uint8_t *diff, int w,
617                                             int *left, int *left_top)
618 {
619     x86_reg w2 = -w;
620     x86_reg x;
621     int l  = *left     & 0xff;
622     int tl = *left_top & 0xff;
623     int t;
624     __asm__ volatile (
625         "mov          %7, %3            \n"
626         "1:                             \n"
627         "movzbl (%3, %4), %2            \n"
628         "mov          %2, %k3           \n"
629         "sub         %b1, %b3           \n"
630         "add         %b0, %b3           \n"
631         "mov          %2, %1            \n"
632         "cmp          %0, %2            \n"
633         "cmovg        %0, %2            \n"
634         "cmovg        %1, %0            \n"
635         "cmp         %k3, %0            \n"
636         "cmovg       %k3, %0            \n"
637         "mov          %7, %3            \n"
638         "cmp          %2, %0            \n"
639         "cmovl        %2, %0            \n"
640         "add    (%6, %4), %b0           \n"
641         "mov         %b0, (%5, %4)      \n"
642         "inc          %4                \n"
643         "jl           1b                \n"
644         : "+&q"(l), "+&q"(tl), "=&r"(t), "=&q"(x), "+&r"(w2)
645         : "r"(dst + w), "r"(diff + w), "rm"(top + w)
646     );
647     *left     = l;
648     *left_top = tl;
649 }
650 #endif
651
652 static inline void transpose4x4(uint8_t *dst, uint8_t *src, x86_reg dst_stride, x86_reg src_stride){
653     __asm__ volatile( //FIXME could save 1 instruction if done as 8x4 ...
654         "movd  (%1), %%mm0              \n\t"
655         "add   %3, %1                   \n\t"
656         "movd  (%1), %%mm1              \n\t"
657         "movd  (%1,%3,1), %%mm2         \n\t"
658         "movd  (%1,%3,2), %%mm3         \n\t"
659         "punpcklbw %%mm1, %%mm0         \n\t"
660         "punpcklbw %%mm3, %%mm2         \n\t"
661         "movq %%mm0, %%mm1              \n\t"
662         "punpcklwd %%mm2, %%mm0         \n\t"
663         "punpckhwd %%mm2, %%mm1         \n\t"
664         "movd  %%mm0, (%0)              \n\t"
665         "add   %2, %0                   \n\t"
666         "punpckhdq %%mm0, %%mm0         \n\t"
667         "movd  %%mm0, (%0)              \n\t"
668         "movd  %%mm1, (%0,%2,1)         \n\t"
669         "punpckhdq %%mm1, %%mm1         \n\t"
670         "movd  %%mm1, (%0,%2,2)         \n\t"
671
672         :  "+&r" (dst),
673            "+&r" (src)
674         :  "r" (dst_stride),
675            "r" (src_stride)
676         :  "memory"
677     );
678 }
679
680 #define H263_LOOP_FILTER                        \
681     "pxor      %%mm7, %%mm7             \n\t"   \
682     "movq         %0, %%mm0             \n\t"   \
683     "movq         %0, %%mm1             \n\t"   \
684     "movq         %3, %%mm2             \n\t"   \
685     "movq         %3, %%mm3             \n\t"   \
686     "punpcklbw %%mm7, %%mm0             \n\t"   \
687     "punpckhbw %%mm7, %%mm1             \n\t"   \
688     "punpcklbw %%mm7, %%mm2             \n\t"   \
689     "punpckhbw %%mm7, %%mm3             \n\t"   \
690     "psubw     %%mm2, %%mm0             \n\t"   \
691     "psubw     %%mm3, %%mm1             \n\t"   \
692     "movq         %1, %%mm2             \n\t"   \
693     "movq         %1, %%mm3             \n\t"   \
694     "movq         %2, %%mm4             \n\t"   \
695     "movq         %2, %%mm5             \n\t"   \
696     "punpcklbw %%mm7, %%mm2             \n\t"   \
697     "punpckhbw %%mm7, %%mm3             \n\t"   \
698     "punpcklbw %%mm7, %%mm4             \n\t"   \
699     "punpckhbw %%mm7, %%mm5             \n\t"   \
700     "psubw     %%mm2, %%mm4             \n\t"   \
701     "psubw     %%mm3, %%mm5             \n\t"   \
702     "psllw        $2, %%mm4             \n\t"   \
703     "psllw        $2, %%mm5             \n\t"   \
704     "paddw     %%mm0, %%mm4             \n\t"   \
705     "paddw     %%mm1, %%mm5             \n\t"   \
706     "pxor      %%mm6, %%mm6             \n\t"   \
707     "pcmpgtw   %%mm4, %%mm6             \n\t"   \
708     "pcmpgtw   %%mm5, %%mm7             \n\t"   \
709     "pxor      %%mm6, %%mm4             \n\t"   \
710     "pxor      %%mm7, %%mm5             \n\t"   \
711     "psubw     %%mm6, %%mm4             \n\t"   \
712     "psubw     %%mm7, %%mm5             \n\t"   \
713     "psrlw        $3, %%mm4             \n\t"   \
714     "psrlw        $3, %%mm5             \n\t"   \
715     "packuswb  %%mm5, %%mm4             \n\t"   \
716     "packsswb  %%mm7, %%mm6             \n\t"   \
717     "pxor      %%mm7, %%mm7             \n\t"   \
718     "movd         %4, %%mm2             \n\t"   \
719     "punpcklbw %%mm2, %%mm2             \n\t"   \
720     "punpcklbw %%mm2, %%mm2             \n\t"   \
721     "punpcklbw %%mm2, %%mm2             \n\t"   \
722     "psubusb   %%mm4, %%mm2             \n\t"   \
723     "movq      %%mm2, %%mm3             \n\t"   \
724     "psubusb   %%mm4, %%mm3             \n\t"   \
725     "psubb     %%mm3, %%mm2             \n\t"   \
726     "movq         %1, %%mm3             \n\t"   \
727     "movq         %2, %%mm4             \n\t"   \
728     "pxor      %%mm6, %%mm3             \n\t"   \
729     "pxor      %%mm6, %%mm4             \n\t"   \
730     "paddusb   %%mm2, %%mm3             \n\t"   \
731     "psubusb   %%mm2, %%mm4             \n\t"   \
732     "pxor      %%mm6, %%mm3             \n\t"   \
733     "pxor      %%mm6, %%mm4             \n\t"   \
734     "paddusb   %%mm2, %%mm2             \n\t"   \
735     "packsswb  %%mm1, %%mm0             \n\t"   \
736     "pcmpgtb   %%mm0, %%mm7             \n\t"   \
737     "pxor      %%mm7, %%mm0             \n\t"   \
738     "psubb     %%mm7, %%mm0             \n\t"   \
739     "movq      %%mm0, %%mm1             \n\t"   \
740     "psubusb   %%mm2, %%mm0             \n\t"   \
741     "psubb     %%mm0, %%mm1             \n\t"   \
742     "pand         %5, %%mm1             \n\t"   \
743     "psrlw        $2, %%mm1             \n\t"   \
744     "pxor      %%mm7, %%mm1             \n\t"   \
745     "psubb     %%mm7, %%mm1             \n\t"   \
746     "movq         %0, %%mm5             \n\t"   \
747     "movq         %3, %%mm6             \n\t"   \
748     "psubb     %%mm1, %%mm5             \n\t"   \
749     "paddb     %%mm1, %%mm6             \n\t"
750
751 static void h263_v_loop_filter_mmx(uint8_t *src, int stride, int qscale)
752 {
753     if (CONFIG_H263_DECODER || CONFIG_H263_ENCODER) {
754         const int strength = ff_h263_loop_filter_strength[qscale];
755
756         __asm__ volatile (
757             H263_LOOP_FILTER
758
759             "movq %%mm3, %1             \n\t"
760             "movq %%mm4, %2             \n\t"
761             "movq %%mm5, %0             \n\t"
762             "movq %%mm6, %3             \n\t"
763             : "+m"(*(uint64_t*)(src - 2 * stride)),
764               "+m"(*(uint64_t*)(src - 1 * stride)),
765               "+m"(*(uint64_t*)(src + 0 * stride)),
766               "+m"(*(uint64_t*)(src + 1 * stride))
767             : "g"(2 * strength), "m"(ff_pb_FC)
768             );
769     }
770 }
771
772 static void h263_h_loop_filter_mmx(uint8_t *src, int stride, int qscale)
773 {
774     if (CONFIG_H263_DECODER || CONFIG_H263_ENCODER) {
775         const int strength = ff_h263_loop_filter_strength[qscale];
776         DECLARE_ALIGNED(8, uint64_t, temp)[4];
777         uint8_t *btemp = (uint8_t*)temp;
778
779         src -= 2;
780
781         transpose4x4(btemp,     src,              8, stride);
782         transpose4x4(btemp + 4, src + 4 * stride, 8, stride);
783         __asm__ volatile (
784             H263_LOOP_FILTER // 5 3 4 6
785
786             : "+m"(temp[0]),
787               "+m"(temp[1]),
788               "+m"(temp[2]),
789               "+m"(temp[3])
790             : "g"(2 * strength), "m"(ff_pb_FC)
791             );
792
793         __asm__ volatile (
794             "movq      %%mm5, %%mm1         \n\t"
795             "movq      %%mm4, %%mm0         \n\t"
796             "punpcklbw %%mm3, %%mm5         \n\t"
797             "punpcklbw %%mm6, %%mm4         \n\t"
798             "punpckhbw %%mm3, %%mm1         \n\t"
799             "punpckhbw %%mm6, %%mm0         \n\t"
800             "movq      %%mm5, %%mm3         \n\t"
801             "movq      %%mm1, %%mm6         \n\t"
802             "punpcklwd %%mm4, %%mm5         \n\t"
803             "punpcklwd %%mm0, %%mm1         \n\t"
804             "punpckhwd %%mm4, %%mm3         \n\t"
805             "punpckhwd %%mm0, %%mm6         \n\t"
806             "movd      %%mm5, (%0)          \n\t"
807             "punpckhdq %%mm5, %%mm5         \n\t"
808             "movd      %%mm5, (%0, %2)      \n\t"
809             "movd      %%mm3, (%0, %2, 2)   \n\t"
810             "punpckhdq %%mm3, %%mm3         \n\t"
811             "movd      %%mm3, (%0, %3)      \n\t"
812             "movd      %%mm1, (%1)          \n\t"
813             "punpckhdq %%mm1, %%mm1         \n\t"
814             "movd      %%mm1, (%1, %2)      \n\t"
815             "movd      %%mm6, (%1, %2, 2)   \n\t"
816             "punpckhdq %%mm6, %%mm6         \n\t"
817             "movd      %%mm6, (%1, %3)      \n\t"
818             :: "r"(src),
819                "r"(src + 4 * stride),
820                "r"((x86_reg)stride),
821                "r"((x86_reg)(3 * stride))
822             );
823     }
824 }
825
826 /* Draw the edges of width 'w' of an image of size width, height
827  * this MMX version can only handle w == 8 || w == 16. */
828 static void draw_edges_mmx(uint8_t *buf, int wrap, int width, int height,
829                            int w, int h, int sides)
830 {
831     uint8_t *ptr, *last_line;
832     int i;
833
834     last_line = buf + (height - 1) * wrap;
835     /* left and right */
836     ptr = buf;
837     if (w == 8) {
838         __asm__ volatile (
839             "1:                             \n\t"
840             "movd            (%0), %%mm0    \n\t"
841             "punpcklbw      %%mm0, %%mm0    \n\t"
842             "punpcklwd      %%mm0, %%mm0    \n\t"
843             "punpckldq      %%mm0, %%mm0    \n\t"
844             "movq           %%mm0, -8(%0)   \n\t"
845             "movq      -8(%0, %2), %%mm1    \n\t"
846             "punpckhbw      %%mm1, %%mm1    \n\t"
847             "punpckhwd      %%mm1, %%mm1    \n\t"
848             "punpckhdq      %%mm1, %%mm1    \n\t"
849             "movq           %%mm1, (%0, %2) \n\t"
850             "add               %1, %0       \n\t"
851             "cmp               %3, %0       \n\t"
852             "jb                1b           \n\t"
853             : "+r"(ptr)
854             : "r"((x86_reg)wrap), "r"((x86_reg)width), "r"(ptr + wrap * height)
855             );
856     } else {
857         __asm__ volatile (
858             "1:                                 \n\t"
859             "movd            (%0), %%mm0        \n\t"
860             "punpcklbw      %%mm0, %%mm0        \n\t"
861             "punpcklwd      %%mm0, %%mm0        \n\t"
862             "punpckldq      %%mm0, %%mm0        \n\t"
863             "movq           %%mm0, -8(%0)       \n\t"
864             "movq           %%mm0, -16(%0)      \n\t"
865             "movq      -8(%0, %2), %%mm1        \n\t"
866             "punpckhbw      %%mm1, %%mm1        \n\t"
867             "punpckhwd      %%mm1, %%mm1        \n\t"
868             "punpckhdq      %%mm1, %%mm1        \n\t"
869             "movq           %%mm1,  (%0, %2)    \n\t"
870             "movq           %%mm1, 8(%0, %2)    \n\t"
871             "add               %1, %0           \n\t"
872             "cmp               %3, %0           \n\t"
873             "jb                1b               \n\t"
874             : "+r"(ptr)
875             : "r"((x86_reg)wrap), "r"((x86_reg)width), "r"(ptr + wrap * height)
876             );
877     }
878
879     /* top and bottom (and hopefully also the corners) */
880     if (sides & EDGE_TOP) {
881         for (i = 0; i < h; i += 4) {
882             ptr = buf - (i + 1) * wrap - w;
883             __asm__ volatile (
884                 "1:                             \n\t"
885                 "movq (%1, %0), %%mm0           \n\t"
886                 "movq    %%mm0, (%0)            \n\t"
887                 "movq    %%mm0, (%0, %2)        \n\t"
888                 "movq    %%mm0, (%0, %2, 2)     \n\t"
889                 "movq    %%mm0, (%0, %3)        \n\t"
890                 "add        $8, %0              \n\t"
891                 "cmp        %4, %0              \n\t"
892                 "jb         1b                  \n\t"
893                 : "+r"(ptr)
894                 : "r"((x86_reg)buf - (x86_reg)ptr - w), "r"((x86_reg) -wrap),
895                   "r"((x86_reg) -wrap * 3), "r"(ptr + width + 2 * w)
896                 );
897         }
898     }
899
900     if (sides & EDGE_BOTTOM) {
901         for (i = 0; i < h; i += 4) {
902             ptr = last_line + (i + 1) * wrap - w;
903             __asm__ volatile (
904                 "1:                             \n\t"
905                 "movq (%1, %0), %%mm0           \n\t"
906                 "movq    %%mm0, (%0)            \n\t"
907                 "movq    %%mm0, (%0, %2)        \n\t"
908                 "movq    %%mm0, (%0, %2, 2)     \n\t"
909                 "movq    %%mm0, (%0, %3)        \n\t"
910                 "add        $8, %0              \n\t"
911                 "cmp        %4, %0              \n\t"
912                 "jb         1b                  \n\t"
913                 : "+r"(ptr)
914                 : "r"((x86_reg)last_line - (x86_reg)ptr - w),
915                   "r"((x86_reg)wrap), "r"((x86_reg)wrap * 3),
916                   "r"(ptr + width + 2 * w)
917                 );
918         }
919     }
920 }
921 #endif /* HAVE_INLINE_ASM */
922
923
924 #if HAVE_YASM
925 #define QPEL_OP(OPNAME, ROUNDER, RND, OP, MMX)                          \
926 static void OPNAME ## qpel8_mc00_ ## MMX (uint8_t *dst, uint8_t *src,   \
927                                           int stride)                   \
928 {                                                                       \
929     ff_ ## OPNAME ## pixels8_ ## MMX(dst, src, stride, 8);              \
930 }                                                                       \
931                                                                         \
932 static void OPNAME ## qpel8_mc10_ ## MMX(uint8_t *dst, uint8_t *src,    \
933                                          int stride)                    \
934 {                                                                       \
935     uint64_t temp[8];                                                   \
936     uint8_t * const half = (uint8_t*)temp;                              \
937     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(half, src, 8,        \
938                                                    stride, 8);          \
939     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, src, half,                 \
940                                         stride, stride, 8);             \
941 }                                                                       \
942                                                                         \
943 static void OPNAME ## qpel8_mc20_ ## MMX(uint8_t *dst, uint8_t *src,    \
944                                          int stride)                    \
945 {                                                                       \
946     ff_ ## OPNAME ## mpeg4_qpel8_h_lowpass_ ## MMX(dst, src, stride,    \
947                                                    stride, 8);          \
948 }                                                                       \
949                                                                         \
950 static void OPNAME ## qpel8_mc30_ ## MMX(uint8_t *dst, uint8_t *src,    \
951                                          int stride)                    \
952 {                                                                       \
953     uint64_t temp[8];                                                   \
954     uint8_t * const half = (uint8_t*)temp;                              \
955     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(half, src, 8,        \
956                                                    stride, 8);          \
957     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, src + 1, half, stride,     \
958                                         stride, 8);                     \
959 }                                                                       \
960                                                                         \
961 static void OPNAME ## qpel8_mc01_ ## MMX(uint8_t *dst, uint8_t *src,    \
962                                          int stride)                    \
963 {                                                                       \
964     uint64_t temp[8];                                                   \
965     uint8_t * const half = (uint8_t*)temp;                              \
966     ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(half, src,           \
967                                                    8, stride);          \
968     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, src, half,                 \
969                                         stride, stride, 8);             \
970 }                                                                       \
971                                                                         \
972 static void OPNAME ## qpel8_mc02_ ## MMX(uint8_t *dst, uint8_t *src,    \
973                                          int stride)                    \
974 {                                                                       \
975     ff_ ## OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, src,            \
976                                                    stride, stride);     \
977 }                                                                       \
978                                                                         \
979 static void OPNAME ## qpel8_mc03_ ## MMX(uint8_t *dst, uint8_t *src,    \
980                                          int stride)                    \
981 {                                                                       \
982     uint64_t temp[8];                                                   \
983     uint8_t * const half = (uint8_t*)temp;                              \
984     ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(half, src,           \
985                                                    8, stride);          \
986     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, src + stride, half, stride,\
987                                         stride, 8);                     \
988 }                                                                       \
989                                                                         \
990 static void OPNAME ## qpel8_mc11_ ## MMX(uint8_t *dst, uint8_t *src,    \
991                                          int stride)                    \
992 {                                                                       \
993     uint64_t half[8 + 9];                                               \
994     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
995     uint8_t * const halfHV = ((uint8_t*)half);                          \
996     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,       \
997                                                    stride, 9);          \
998     ff_put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8,           \
999                                         stride, 9);                     \
1000     ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
1001     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV,             \
1002                                         stride, 8, 8);                  \
1003 }                                                                       \
1004                                                                         \
1005 static void OPNAME ## qpel8_mc31_ ## MMX(uint8_t *dst, uint8_t *src,    \
1006                                          int stride)                    \
1007 {                                                                       \
1008     uint64_t half[8 + 9];                                               \
1009     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
1010     uint8_t * const halfHV = ((uint8_t*)half);                          \
1011     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,       \
1012                                                    stride, 9);          \
1013     ff_put ## RND ## pixels8_l2_ ## MMX(halfH, src + 1, halfH, 8,       \
1014                                         stride, 9);                     \
1015     ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
1016     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV,             \
1017                                         stride, 8, 8);                  \
1018 }                                                                       \
1019                                                                         \
1020 static void OPNAME ## qpel8_mc13_ ## MMX(uint8_t *dst, uint8_t *src,    \
1021                                          int stride)                    \
1022 {                                                                       \
1023     uint64_t half[8 + 9];                                               \
1024     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
1025     uint8_t * const halfHV = ((uint8_t*)half);                          \
1026     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,       \
1027                                                    stride, 9);          \
1028     ff_put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8,           \
1029                                         stride, 9);                     \
1030     ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
1031     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, halfH + 8, halfHV,         \
1032                                         stride, 8, 8);                  \
1033 }                                                                       \
1034                                                                         \
1035 static void OPNAME ## qpel8_mc33_ ## MMX(uint8_t *dst, uint8_t *src,    \
1036                                          int stride)                    \
1037 {                                                                       \
1038     uint64_t half[8 + 9];                                               \
1039     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
1040     uint8_t * const halfHV = ((uint8_t*)half);                          \
1041     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,       \
1042                                                    stride, 9);          \
1043     ff_put ## RND ## pixels8_l2_ ## MMX(halfH, src + 1, halfH, 8,       \
1044                                         stride, 9);                     \
1045     ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
1046     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, halfH + 8, halfHV,         \
1047                                         stride, 8, 8);                  \
1048 }                                                                       \
1049                                                                         \
1050 static void OPNAME ## qpel8_mc21_ ## MMX(uint8_t *dst, uint8_t *src,    \
1051                                          int stride)                    \
1052 {                                                                       \
1053     uint64_t half[8 + 9];                                               \
1054     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
1055     uint8_t * const halfHV = ((uint8_t*)half);                          \
1056     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,       \
1057                                                    stride, 9);          \
1058     ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
1059     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV,             \
1060                                         stride, 8, 8);                  \
1061 }                                                                       \
1062                                                                         \
1063 static void OPNAME ## qpel8_mc23_ ## MMX(uint8_t *dst, uint8_t *src,    \
1064                                          int stride)                    \
1065 {                                                                       \
1066     uint64_t half[8 + 9];                                               \
1067     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
1068     uint8_t * const halfHV = ((uint8_t*)half);                          \
1069     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,       \
1070                                                    stride, 9);          \
1071     ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
1072     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, halfH + 8, halfHV,         \
1073                                         stride, 8, 8);                  \
1074 }                                                                       \
1075                                                                         \
1076 static void OPNAME ## qpel8_mc12_ ## MMX(uint8_t *dst, uint8_t *src,    \
1077                                          int stride)                    \
1078 {                                                                       \
1079     uint64_t half[8 + 9];                                               \
1080     uint8_t * const halfH = ((uint8_t*)half);                           \
1081     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,       \
1082                                                    stride, 9);          \
1083     ff_put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH,              \
1084                                         8, stride, 9);                  \
1085     ff_ ## OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH,          \
1086                                                    stride, 8);          \
1087 }                                                                       \
1088                                                                         \
1089 static void OPNAME ## qpel8_mc32_ ## MMX(uint8_t *dst, uint8_t *src,    \
1090                                          int stride)                    \
1091 {                                                                       \
1092     uint64_t half[8 + 9];                                               \
1093     uint8_t * const halfH = ((uint8_t*)half);                           \
1094     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,       \
1095                                                    stride, 9);          \
1096     ff_put ## RND ## pixels8_l2_ ## MMX(halfH, src + 1, halfH, 8,       \
1097                                         stride, 9);                     \
1098     ff_ ## OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH,          \
1099                                                    stride, 8);          \
1100 }                                                                       \
1101                                                                         \
1102 static void OPNAME ## qpel8_mc22_ ## MMX(uint8_t *dst, uint8_t *src,    \
1103                                          int stride)                    \
1104 {                                                                       \
1105     uint64_t half[9];                                                   \
1106     uint8_t * const halfH = ((uint8_t*)half);                           \
1107     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,       \
1108                                                    stride, 9);          \
1109     ff_ ## OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH,          \
1110                                                    stride, 8);          \
1111 }                                                                       \
1112                                                                         \
1113 static void OPNAME ## qpel16_mc00_ ## MMX (uint8_t *dst, uint8_t *src,  \
1114                                            int stride)                  \
1115 {                                                                       \
1116     ff_ ## OPNAME ## pixels16_ ## MMX(dst, src, stride, 16);            \
1117 }                                                                       \
1118                                                                         \
1119 static void OPNAME ## qpel16_mc10_ ## MMX(uint8_t *dst, uint8_t *src,   \
1120                                           int stride)                   \
1121 {                                                                       \
1122     uint64_t temp[32];                                                  \
1123     uint8_t * const half = (uint8_t*)temp;                              \
1124     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(half, src, 16,      \
1125                                                     stride, 16);        \
1126     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, src, half, stride,        \
1127                                          stride, 16);                   \
1128 }                                                                       \
1129                                                                         \
1130 static void OPNAME ## qpel16_mc20_ ## MMX(uint8_t *dst, uint8_t *src,   \
1131                                           int stride)                   \
1132 {                                                                       \
1133     ff_ ## OPNAME ## mpeg4_qpel16_h_lowpass_ ## MMX(dst, src,           \
1134                                                     stride, stride, 16);\
1135 }                                                                       \
1136                                                                         \
1137 static void OPNAME ## qpel16_mc30_ ## MMX(uint8_t *dst, uint8_t *src,   \
1138                                           int stride)                   \
1139 {                                                                       \
1140     uint64_t temp[32];                                                  \
1141     uint8_t * const half = (uint8_t*)temp;                              \
1142     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(half, src, 16,      \
1143                                                     stride, 16);        \
1144     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, src + 1, half,            \
1145                                          stride, stride, 16);           \
1146 }                                                                       \
1147                                                                         \
1148 static void OPNAME ## qpel16_mc01_ ## MMX(uint8_t *dst, uint8_t *src,   \
1149                                           int stride)                   \
1150 {                                                                       \
1151     uint64_t temp[32];                                                  \
1152     uint8_t * const half = (uint8_t*)temp;                              \
1153     ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(half, src, 16,      \
1154                                                     stride);            \
1155     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, src, half, stride,        \
1156                                          stride, 16);                   \
1157 }                                                                       \
1158                                                                         \
1159 static void OPNAME ## qpel16_mc02_ ## MMX(uint8_t *dst, uint8_t *src,   \
1160                                           int stride)                   \
1161 {                                                                       \
1162     ff_ ## OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, src,           \
1163                                                     stride, stride);    \
1164 }                                                                       \
1165                                                                         \
1166 static void OPNAME ## qpel16_mc03_ ## MMX(uint8_t *dst, uint8_t *src,   \
1167                                           int stride)                   \
1168 {                                                                       \
1169     uint64_t temp[32];                                                  \
1170     uint8_t * const half = (uint8_t*)temp;                              \
1171     ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(half, src, 16,      \
1172                                                     stride);            \
1173     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, src+stride, half,         \
1174                                          stride, stride, 16);           \
1175 }                                                                       \
1176                                                                         \
1177 static void OPNAME ## qpel16_mc11_ ## MMX(uint8_t *dst, uint8_t *src,   \
1178                                           int stride)                   \
1179 {                                                                       \
1180     uint64_t half[16 * 2 + 17 * 2];                                     \
1181     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1182     uint8_t * const halfHV = ((uint8_t*)half);                          \
1183     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
1184                                                     stride, 17);        \
1185     ff_put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16,         \
1186                                          stride, 17);                   \
1187     ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,      \
1188                                                     16, 16);            \
1189     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV,            \
1190                                          stride, 16, 16);               \
1191 }                                                                       \
1192                                                                         \
1193 static void OPNAME ## qpel16_mc31_ ## MMX(uint8_t *dst, uint8_t *src,   \
1194                                           int stride)                   \
1195 {                                                                       \
1196     uint64_t half[16 * 2 + 17 * 2];                                     \
1197     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1198     uint8_t * const halfHV = ((uint8_t*)half);                          \
1199     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
1200                                                     stride, 17);        \
1201     ff_put ## RND ## pixels16_l2_ ## MMX(halfH, src + 1, halfH, 16,     \
1202                                          stride, 17);                   \
1203     ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,      \
1204                                                     16, 16);            \
1205     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV,            \
1206                                          stride, 16, 16);               \
1207 }                                                                       \
1208                                                                         \
1209 static void OPNAME ## qpel16_mc13_ ## MMX(uint8_t *dst, uint8_t *src,   \
1210                                           int stride)                   \
1211 {                                                                       \
1212     uint64_t half[16 * 2 + 17 * 2];                                     \
1213     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1214     uint8_t * const halfHV = ((uint8_t*)half);                          \
1215     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
1216                                                     stride, 17);        \
1217     ff_put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16,         \
1218                                          stride, 17);                   \
1219     ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,      \
1220                                                     16, 16);            \
1221     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, halfH + 16, halfHV,       \
1222                                          stride, 16, 16);               \
1223 }                                                                       \
1224                                                                         \
1225 static void OPNAME ## qpel16_mc33_ ## MMX(uint8_t *dst, uint8_t *src,   \
1226                                           int stride)                   \
1227 {                                                                       \
1228     uint64_t half[16 * 2 + 17 * 2];                                     \
1229     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1230     uint8_t * const halfHV = ((uint8_t*)half);                          \
1231     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
1232                                                     stride, 17);        \
1233     ff_put ## RND ## pixels16_l2_ ## MMX(halfH, src + 1, halfH, 16,     \
1234                                          stride, 17);                   \
1235     ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,      \
1236                                                     16, 16);            \
1237     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, halfH + 16, halfHV,       \
1238                                          stride, 16, 16);               \
1239 }                                                                       \
1240                                                                         \
1241 static void OPNAME ## qpel16_mc21_ ## MMX(uint8_t *dst, uint8_t *src,   \
1242                                           int stride)                   \
1243 {                                                                       \
1244     uint64_t half[16 * 2 + 17 * 2];                                     \
1245     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1246     uint8_t * const halfHV = ((uint8_t*)half);                          \
1247     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
1248                                                     stride, 17);        \
1249     ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,      \
1250                                                     16, 16);            \
1251     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV,            \
1252                                          stride, 16, 16);               \
1253 }                                                                       \
1254                                                                         \
1255 static void OPNAME ## qpel16_mc23_ ## MMX(uint8_t *dst, uint8_t *src,   \
1256                                           int stride)                   \
1257 {                                                                       \
1258     uint64_t half[16 * 2 + 17 * 2];                                     \
1259     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1260     uint8_t * const halfHV = ((uint8_t*)half);                          \
1261     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
1262                                                     stride, 17);        \
1263     ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,      \
1264                                                     16, 16);            \
1265     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, halfH + 16, halfHV,       \
1266                                          stride, 16, 16);               \
1267 }                                                                       \
1268                                                                         \
1269 static void OPNAME ## qpel16_mc12_ ## MMX(uint8_t *dst, uint8_t *src,   \
1270                                           int stride)                   \
1271 {                                                                       \
1272     uint64_t half[17 * 2];                                              \
1273     uint8_t * const halfH = ((uint8_t*)half);                           \
1274     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
1275                                                     stride, 17);        \
1276     ff_put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16,         \
1277                                          stride, 17);                   \
1278     ff_ ## OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH,         \
1279                                                     stride, 16);        \
1280 }                                                                       \
1281                                                                         \
1282 static void OPNAME ## qpel16_mc32_ ## MMX(uint8_t *dst, uint8_t *src,   \
1283                                           int stride)                   \
1284 {                                                                       \
1285     uint64_t half[17 * 2];                                              \
1286     uint8_t * const halfH = ((uint8_t*)half);                           \
1287     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
1288                                                     stride, 17);        \
1289     ff_put ## RND ## pixels16_l2_ ## MMX(halfH, src + 1, halfH, 16,     \
1290                                          stride, 17);                   \
1291     ff_ ## OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH,         \
1292                                                     stride, 16);        \
1293 }                                                                       \
1294                                                                         \
1295 static void OPNAME ## qpel16_mc22_ ## MMX(uint8_t *dst, uint8_t *src,   \
1296                                           int stride)                   \
1297 {                                                                       \
1298     uint64_t half[17 * 2];                                              \
1299     uint8_t * const halfH = ((uint8_t*)half);                           \
1300     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
1301                                                     stride, 17);        \
1302     ff_ ## OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH,         \
1303                                                     stride, 16);        \
1304 }
1305
1306 #define PUT_OP(a, b, temp, size)                \
1307     "mov"#size"        "#a", "#b"       \n\t"
1308
1309 #define AVG_MMXEXT_OP(a, b, temp, size)         \
1310     "mov"#size"        "#b", "#temp"    \n\t"   \
1311     "pavgb          "#temp", "#a"       \n\t"   \
1312     "mov"#size"        "#a", "#b"       \n\t"
1313
1314 QPEL_OP(put_,          ff_pw_16, _,        PUT_OP,        mmxext)
1315 QPEL_OP(avg_,          ff_pw_16, _,        AVG_MMXEXT_OP, mmxext)
1316 QPEL_OP(put_no_rnd_,   ff_pw_15, _no_rnd_, PUT_OP,        mmxext)
1317 #endif /* HAVE_YASM */
1318
1319
1320 #if HAVE_INLINE_ASM
1321 void ff_put_rv40_qpel8_mc33_mmx(uint8_t *dst, uint8_t *src, int stride)
1322 {
1323   put_pixels8_xy2_mmx(dst, src, stride, 8);
1324 }
1325 void ff_put_rv40_qpel16_mc33_mmx(uint8_t *dst, uint8_t *src, int stride)
1326 {
1327   put_pixels16_xy2_mmx(dst, src, stride, 16);
1328 }
1329 void ff_avg_rv40_qpel8_mc33_mmx(uint8_t *dst, uint8_t *src, int stride)
1330 {
1331   avg_pixels8_xy2_mmx(dst, src, stride, 8);
1332 }
1333 void ff_avg_rv40_qpel16_mc33_mmx(uint8_t *dst, uint8_t *src, int stride)
1334 {
1335   avg_pixels16_xy2_mmx(dst, src, stride, 16);
1336 }
1337
1338 static void gmc_mmx(uint8_t *dst, uint8_t *src,
1339                     int stride, int h, int ox, int oy,
1340                     int dxx, int dxy, int dyx, int dyy,
1341                     int shift, int r, int width, int height)
1342 {
1343     const int w    = 8;
1344     const int ix   = ox  >> (16 + shift);
1345     const int iy   = oy  >> (16 + shift);
1346     const int oxs  = ox  >> 4;
1347     const int oys  = oy  >> 4;
1348     const int dxxs = dxx >> 4;
1349     const int dxys = dxy >> 4;
1350     const int dyxs = dyx >> 4;
1351     const int dyys = dyy >> 4;
1352     const uint16_t r4[4]   = { r, r, r, r };
1353     const uint16_t dxy4[4] = { dxys, dxys, dxys, dxys };
1354     const uint16_t dyy4[4] = { dyys, dyys, dyys, dyys };
1355     const uint64_t shift2 = 2 * shift;
1356     int x, y;
1357
1358     const int dxw = (dxx - (1 << (16 + shift))) * (w - 1);
1359     const int dyh = (dyy - (1 << (16 + shift))) * (h - 1);
1360     const int dxh = dxy * (h - 1);
1361     const int dyw = dyx * (w - 1);
1362     if ( // non-constant fullpel offset (3% of blocks)
1363         ((ox ^ (ox + dxw)) | (ox ^ (ox + dxh)) | (ox ^ (ox + dxw + dxh)) |
1364          (oy ^ (oy + dyw)) | (oy ^ (oy + dyh)) | (oy ^ (oy + dyw + dyh))) >> (16 + shift)
1365         // uses more than 16 bits of subpel mv (only at huge resolution)
1366         || (dxx | dxy | dyx | dyy) & 15 ||
1367         (unsigned)ix >= width  - w ||
1368         (unsigned)iy >= height - h) {
1369         // FIXME could still use mmx for some of the rows
1370         ff_gmc_c(dst, src, stride, h, ox, oy, dxx, dxy, dyx, dyy,
1371                  shift, r, width, height);
1372         return;
1373     }
1374
1375     src += ix + iy * stride;
1376
1377     __asm__ volatile (
1378         "movd         %0, %%mm6         \n\t"
1379         "pxor      %%mm7, %%mm7         \n\t"
1380         "punpcklwd %%mm6, %%mm6         \n\t"
1381         "punpcklwd %%mm6, %%mm6         \n\t"
1382         :: "r"(1<<shift)
1383     );
1384
1385     for (x = 0; x < w; x += 4) {
1386         uint16_t dx4[4] = { oxs - dxys + dxxs * (x + 0),
1387                             oxs - dxys + dxxs * (x + 1),
1388                             oxs - dxys + dxxs * (x + 2),
1389                             oxs - dxys + dxxs * (x + 3) };
1390         uint16_t dy4[4] = { oys - dyys + dyxs * (x + 0),
1391                             oys - dyys + dyxs * (x + 1),
1392                             oys - dyys + dyxs * (x + 2),
1393                             oys - dyys + dyxs * (x + 3) };
1394
1395         for (y = 0; y < h; y++) {
1396             __asm__ volatile (
1397                 "movq      %0, %%mm4    \n\t"
1398                 "movq      %1, %%mm5    \n\t"
1399                 "paddw     %2, %%mm4    \n\t"
1400                 "paddw     %3, %%mm5    \n\t"
1401                 "movq   %%mm4, %0       \n\t"
1402                 "movq   %%mm5, %1       \n\t"
1403                 "psrlw    $12, %%mm4    \n\t"
1404                 "psrlw    $12, %%mm5    \n\t"
1405                 : "+m"(*dx4), "+m"(*dy4)
1406                 : "m"(*dxy4), "m"(*dyy4)
1407             );
1408
1409             __asm__ volatile (
1410                 "movq      %%mm6, %%mm2 \n\t"
1411                 "movq      %%mm6, %%mm1 \n\t"
1412                 "psubw     %%mm4, %%mm2 \n\t"
1413                 "psubw     %%mm5, %%mm1 \n\t"
1414                 "movq      %%mm2, %%mm0 \n\t"
1415                 "movq      %%mm4, %%mm3 \n\t"
1416                 "pmullw    %%mm1, %%mm0 \n\t" // (s - dx) * (s - dy)
1417                 "pmullw    %%mm5, %%mm3 \n\t" // dx * dy
1418                 "pmullw    %%mm5, %%mm2 \n\t" // (s - dx) * dy
1419                 "pmullw    %%mm4, %%mm1 \n\t" // dx * (s - dy)
1420
1421                 "movd         %4, %%mm5 \n\t"
1422                 "movd         %3, %%mm4 \n\t"
1423                 "punpcklbw %%mm7, %%mm5 \n\t"
1424                 "punpcklbw %%mm7, %%mm4 \n\t"
1425                 "pmullw    %%mm5, %%mm3 \n\t" // src[1, 1] * dx * dy
1426                 "pmullw    %%mm4, %%mm2 \n\t" // src[0, 1] * (s - dx) * dy
1427
1428                 "movd         %2, %%mm5 \n\t"
1429                 "movd         %1, %%mm4 \n\t"
1430                 "punpcklbw %%mm7, %%mm5 \n\t"
1431                 "punpcklbw %%mm7, %%mm4 \n\t"
1432                 "pmullw    %%mm5, %%mm1 \n\t" // src[1, 0] * dx * (s - dy)
1433                 "pmullw    %%mm4, %%mm0 \n\t" // src[0, 0] * (s - dx) * (s - dy)
1434                 "paddw        %5, %%mm1 \n\t"
1435                 "paddw     %%mm3, %%mm2 \n\t"
1436                 "paddw     %%mm1, %%mm0 \n\t"
1437                 "paddw     %%mm2, %%mm0 \n\t"
1438
1439                 "psrlw        %6, %%mm0 \n\t"
1440                 "packuswb  %%mm0, %%mm0 \n\t"
1441                 "movd      %%mm0, %0    \n\t"
1442
1443                 : "=m"(dst[x + y * stride])
1444                 : "m"(src[0]), "m"(src[1]),
1445                   "m"(src[stride]), "m"(src[stride + 1]),
1446                   "m"(*r4), "m"(shift2)
1447             );
1448             src += stride;
1449         }
1450         src += 4 - h * stride;
1451     }
1452 }
1453 #endif /* HAVE_INLINE_ASM */
1454
1455 void ff_put_pixels16_sse2(uint8_t *block, const uint8_t *pixels,
1456                           int line_size, int h);
1457 void ff_avg_pixels16_sse2(uint8_t *block, const uint8_t *pixels,
1458                           int line_size, int h);
1459
1460 void ff_put_h264_chroma_mc8_rnd_mmx  (uint8_t *dst, uint8_t *src,
1461                                       int stride, int h, int x, int y);
1462 void ff_avg_h264_chroma_mc8_rnd_mmxext(uint8_t *dst, uint8_t *src,
1463                                        int stride, int h, int x, int y);
1464 void ff_avg_h264_chroma_mc8_rnd_3dnow(uint8_t *dst, uint8_t *src,
1465                                       int stride, int h, int x, int y);
1466
1467 void ff_put_h264_chroma_mc4_mmx      (uint8_t *dst, uint8_t *src,
1468                                       int stride, int h, int x, int y);
1469 void ff_avg_h264_chroma_mc4_mmxext   (uint8_t *dst, uint8_t *src,
1470                                       int stride, int h, int x, int y);
1471 void ff_avg_h264_chroma_mc4_3dnow    (uint8_t *dst, uint8_t *src,
1472                                       int stride, int h, int x, int y);
1473
1474 void ff_put_h264_chroma_mc2_mmxext   (uint8_t *dst, uint8_t *src,
1475                                       int stride, int h, int x, int y);
1476 void ff_avg_h264_chroma_mc2_mmxext   (uint8_t *dst, uint8_t *src,
1477                                       int stride, int h, int x, int y);
1478
1479 void ff_put_h264_chroma_mc8_rnd_ssse3(uint8_t *dst, uint8_t *src,
1480                                       int stride, int h, int x, int y);
1481 void ff_put_h264_chroma_mc4_ssse3    (uint8_t *dst, uint8_t *src,
1482                                       int stride, int h, int x, int y);
1483
1484 void ff_avg_h264_chroma_mc8_rnd_ssse3(uint8_t *dst, uint8_t *src,
1485                                       int stride, int h, int x, int y);
1486 void ff_avg_h264_chroma_mc4_ssse3    (uint8_t *dst, uint8_t *src,
1487                                       int stride, int h, int x, int y);
1488
1489 #define CHROMA_MC(OP, NUM, DEPTH, OPT)                                  \
1490 void ff_ ## OP ## _h264_chroma_mc ## NUM ## _ ## DEPTH ## _ ## OPT      \
1491                                       (uint8_t *dst, uint8_t *src,      \
1492                                        int stride, int h, int x, int y);
1493
1494 CHROMA_MC(put, 2, 10, mmxext)
1495 CHROMA_MC(avg, 2, 10, mmxext)
1496 CHROMA_MC(put, 4, 10, mmxext)
1497 CHROMA_MC(avg, 4, 10, mmxext)
1498 CHROMA_MC(put, 8, 10, sse2)
1499 CHROMA_MC(avg, 8, 10, sse2)
1500 CHROMA_MC(put, 8, 10, avx)
1501 CHROMA_MC(avg, 8, 10, avx)
1502
1503 #if HAVE_INLINE_ASM
1504
1505 /* CAVS-specific */
1506 void ff_put_cavs_qpel8_mc00_mmxext(uint8_t *dst, uint8_t *src, int stride)
1507 {
1508     put_pixels8_mmx(dst, src, stride, 8);
1509 }
1510
1511 void ff_avg_cavs_qpel8_mc00_mmxext(uint8_t *dst, uint8_t *src, int stride)
1512 {
1513     avg_pixels8_mmx(dst, src, stride, 8);
1514 }
1515
1516 void ff_put_cavs_qpel16_mc00_mmxext(uint8_t *dst, uint8_t *src, int stride)
1517 {
1518     put_pixels16_mmx(dst, src, stride, 16);
1519 }
1520
1521 void ff_avg_cavs_qpel16_mc00_mmxext(uint8_t *dst, uint8_t *src, int stride)
1522 {
1523     avg_pixels16_mmx(dst, src, stride, 16);
1524 }
1525 #endif /* HAVE_INLINE_ASM */
1526
1527 #if HAVE_YASM
1528 /* VC-1-specific */
1529 void ff_put_vc1_mspel_mc00_mmx(uint8_t *dst, const uint8_t *src,
1530                                int stride, int rnd)
1531 {
1532     ff_put_pixels8_mmx(dst, src, stride, 8);
1533 }
1534
1535 void ff_avg_vc1_mspel_mc00_mmxext(uint8_t *dst, const uint8_t *src,
1536                                   int stride, int rnd)
1537 {
1538     ff_avg_pixels8_mmxext(dst, src, stride, 8);
1539 }
1540 #endif /* HAVE_YASM */
1541
1542 #if HAVE_INLINE_ASM
1543 static void vector_clipf_sse(float *dst, const float *src,
1544                              float min, float max, int len)
1545 {
1546     x86_reg i = (len - 16) * 4;
1547     __asm__ volatile (
1548         "movss          %3, %%xmm4      \n\t"
1549         "movss          %4, %%xmm5      \n\t"
1550         "shufps $0, %%xmm4, %%xmm4      \n\t"
1551         "shufps $0, %%xmm5, %%xmm5      \n\t"
1552         "1:                             \n\t"
1553         "movaps   (%2, %0), %%xmm0      \n\t" // 3/1 on intel
1554         "movaps 16(%2, %0), %%xmm1      \n\t"
1555         "movaps 32(%2, %0), %%xmm2      \n\t"
1556         "movaps 48(%2, %0), %%xmm3      \n\t"
1557         "maxps      %%xmm4, %%xmm0      \n\t"
1558         "maxps      %%xmm4, %%xmm1      \n\t"
1559         "maxps      %%xmm4, %%xmm2      \n\t"
1560         "maxps      %%xmm4, %%xmm3      \n\t"
1561         "minps      %%xmm5, %%xmm0      \n\t"
1562         "minps      %%xmm5, %%xmm1      \n\t"
1563         "minps      %%xmm5, %%xmm2      \n\t"
1564         "minps      %%xmm5, %%xmm3      \n\t"
1565         "movaps     %%xmm0,   (%1, %0)  \n\t"
1566         "movaps     %%xmm1, 16(%1, %0)  \n\t"
1567         "movaps     %%xmm2, 32(%1, %0)  \n\t"
1568         "movaps     %%xmm3, 48(%1, %0)  \n\t"
1569         "sub           $64, %0          \n\t"
1570         "jge            1b              \n\t"
1571         : "+&r"(i)
1572         : "r"(dst), "r"(src), "m"(min), "m"(max)
1573         : "memory"
1574     );
1575 }
1576
1577 #endif /* HAVE_INLINE_ASM */
1578
1579 int32_t ff_scalarproduct_int16_mmxext(const int16_t *v1, const int16_t *v2,
1580                                       int order);
1581 int32_t ff_scalarproduct_int16_sse2(const int16_t *v1, const int16_t *v2,
1582                                     int order);
1583 int32_t ff_scalarproduct_and_madd_int16_mmxext(int16_t *v1, const int16_t *v2,
1584                                                const int16_t *v3,
1585                                                int order, int mul);
1586 int32_t ff_scalarproduct_and_madd_int16_sse2(int16_t *v1, const int16_t *v2,
1587                                              const int16_t *v3,
1588                                              int order, int mul);
1589 int32_t ff_scalarproduct_and_madd_int16_ssse3(int16_t *v1, const int16_t *v2,
1590                                               const int16_t *v3,
1591                                               int order, int mul);
1592
1593 void ff_apply_window_int16_round_mmxext(int16_t *output, const int16_t *input,
1594                                         const int16_t *window, unsigned int len);
1595 void ff_apply_window_int16_round_sse2(int16_t *output, const int16_t *input,
1596                                       const int16_t *window, unsigned int len);
1597 void ff_apply_window_int16_mmxext(int16_t *output, const int16_t *input,
1598                                   const int16_t *window, unsigned int len);
1599 void ff_apply_window_int16_sse2(int16_t *output, const int16_t *input,
1600                                 const int16_t *window, unsigned int len);
1601 void ff_apply_window_int16_ssse3(int16_t *output, const int16_t *input,
1602                                  const int16_t *window, unsigned int len);
1603 void ff_apply_window_int16_ssse3_atom(int16_t *output, const int16_t *input,
1604                                       const int16_t *window, unsigned int len);
1605
1606 void ff_bswap32_buf_ssse3(uint32_t *dst, const uint32_t *src, int w);
1607 void ff_bswap32_buf_sse2(uint32_t *dst, const uint32_t *src, int w);
1608
1609 void ff_add_hfyu_median_prediction_mmxext(uint8_t *dst, const uint8_t *top,
1610                                           const uint8_t *diff, int w,
1611                                           int *left, int *left_top);
1612 int  ff_add_hfyu_left_prediction_ssse3(uint8_t *dst, const uint8_t *src,
1613                                        int w, int left);
1614 int  ff_add_hfyu_left_prediction_sse4(uint8_t *dst, const uint8_t *src,
1615                                       int w, int left);
1616
1617 void ff_vector_clip_int32_mmx     (int32_t *dst, const int32_t *src,
1618                                    int32_t min, int32_t max, unsigned int len);
1619 void ff_vector_clip_int32_sse2    (int32_t *dst, const int32_t *src,
1620                                    int32_t min, int32_t max, unsigned int len);
1621 void ff_vector_clip_int32_int_sse2(int32_t *dst, const int32_t *src,
1622                                    int32_t min, int32_t max, unsigned int len);
1623 void ff_vector_clip_int32_sse4    (int32_t *dst, const int32_t *src,
1624                                    int32_t min, int32_t max, unsigned int len);
1625
1626 #define SET_QPEL_FUNCS(PFX, IDX, SIZE, CPU, PREFIX)                          \
1627     do {                                                                     \
1628     c->PFX ## _pixels_tab[IDX][ 0] = PREFIX ## PFX ## SIZE ## _mc00_ ## CPU; \
1629     c->PFX ## _pixels_tab[IDX][ 1] = PREFIX ## PFX ## SIZE ## _mc10_ ## CPU; \
1630     c->PFX ## _pixels_tab[IDX][ 2] = PREFIX ## PFX ## SIZE ## _mc20_ ## CPU; \
1631     c->PFX ## _pixels_tab[IDX][ 3] = PREFIX ## PFX ## SIZE ## _mc30_ ## CPU; \
1632     c->PFX ## _pixels_tab[IDX][ 4] = PREFIX ## PFX ## SIZE ## _mc01_ ## CPU; \
1633     c->PFX ## _pixels_tab[IDX][ 5] = PREFIX ## PFX ## SIZE ## _mc11_ ## CPU; \
1634     c->PFX ## _pixels_tab[IDX][ 6] = PREFIX ## PFX ## SIZE ## _mc21_ ## CPU; \
1635     c->PFX ## _pixels_tab[IDX][ 7] = PREFIX ## PFX ## SIZE ## _mc31_ ## CPU; \
1636     c->PFX ## _pixels_tab[IDX][ 8] = PREFIX ## PFX ## SIZE ## _mc02_ ## CPU; \
1637     c->PFX ## _pixels_tab[IDX][ 9] = PREFIX ## PFX ## SIZE ## _mc12_ ## CPU; \
1638     c->PFX ## _pixels_tab[IDX][10] = PREFIX ## PFX ## SIZE ## _mc22_ ## CPU; \
1639     c->PFX ## _pixels_tab[IDX][11] = PREFIX ## PFX ## SIZE ## _mc32_ ## CPU; \
1640     c->PFX ## _pixels_tab[IDX][12] = PREFIX ## PFX ## SIZE ## _mc03_ ## CPU; \
1641     c->PFX ## _pixels_tab[IDX][13] = PREFIX ## PFX ## SIZE ## _mc13_ ## CPU; \
1642     c->PFX ## _pixels_tab[IDX][14] = PREFIX ## PFX ## SIZE ## _mc23_ ## CPU; \
1643     c->PFX ## _pixels_tab[IDX][15] = PREFIX ## PFX ## SIZE ## _mc33_ ## CPU; \
1644     } while (0)
1645
1646 #define SET_HPEL_FUNCS(PFX, IDX, SIZE, CPU)                                     \
1647     do {                                                                        \
1648         c->PFX ## _pixels_tab IDX [0] = PFX ## _pixels ## SIZE ## _     ## CPU; \
1649         c->PFX ## _pixels_tab IDX [1] = PFX ## _pixels ## SIZE ## _x2_  ## CPU; \
1650         c->PFX ## _pixels_tab IDX [2] = PFX ## _pixels ## SIZE ## _y2_  ## CPU; \
1651         c->PFX ## _pixels_tab IDX [3] = PFX ## _pixels ## SIZE ## _xy2_ ## CPU; \
1652     } while (0)
1653
1654 static void dsputil_init_mmx(DSPContext *c, AVCodecContext *avctx, int mm_flags)
1655 {
1656     const int high_bit_depth = avctx->bits_per_raw_sample > 8;
1657
1658 #if HAVE_INLINE_ASM
1659     c->put_pixels_clamped        = ff_put_pixels_clamped_mmx;
1660     c->put_signed_pixels_clamped = ff_put_signed_pixels_clamped_mmx;
1661     c->add_pixels_clamped        = ff_add_pixels_clamped_mmx;
1662
1663     if (!high_bit_depth) {
1664         c->clear_block  = clear_block_mmx;
1665         c->clear_blocks = clear_blocks_mmx;
1666         c->draw_edges   = draw_edges_mmx;
1667
1668         SET_HPEL_FUNCS(put,        [0], 16, mmx);
1669         SET_HPEL_FUNCS(put_no_rnd, [0], 16, mmx);
1670         SET_HPEL_FUNCS(avg,        [0], 16, mmx);
1671         SET_HPEL_FUNCS(avg_no_rnd,    , 16, mmx);
1672         SET_HPEL_FUNCS(put,        [1],  8, mmx);
1673         SET_HPEL_FUNCS(put_no_rnd, [1],  8, mmx);
1674         SET_HPEL_FUNCS(avg,        [1],  8, mmx);
1675
1676         switch (avctx->idct_algo) {
1677         case FF_IDCT_AUTO:
1678         case FF_IDCT_SIMPLEMMX:
1679             c->idct_put              = ff_simple_idct_put_mmx;
1680             c->idct_add              = ff_simple_idct_add_mmx;
1681             c->idct                  = ff_simple_idct_mmx;
1682             c->idct_permutation_type = FF_SIMPLE_IDCT_PERM;
1683             break;
1684         case FF_IDCT_XVIDMMX:
1685             c->idct_put              = ff_idct_xvid_mmx_put;
1686             c->idct_add              = ff_idct_xvid_mmx_add;
1687             c->idct                  = ff_idct_xvid_mmx;
1688             break;
1689         }
1690     }
1691
1692     c->gmc = gmc_mmx;
1693
1694     c->add_bytes = add_bytes_mmx;
1695
1696     if (CONFIG_H263_DECODER || CONFIG_H263_ENCODER) {
1697         c->h263_v_loop_filter = h263_v_loop_filter_mmx;
1698         c->h263_h_loop_filter = h263_h_loop_filter_mmx;
1699     }
1700 #endif /* HAVE_INLINE_ASM */
1701
1702 #if HAVE_YASM
1703     if (!high_bit_depth && CONFIG_H264CHROMA) {
1704         c->put_h264_chroma_pixels_tab[0] = ff_put_h264_chroma_mc8_rnd_mmx;
1705         c->put_h264_chroma_pixels_tab[1] = ff_put_h264_chroma_mc4_mmx;
1706     }
1707
1708     c->vector_clip_int32 = ff_vector_clip_int32_mmx;
1709 #endif
1710
1711 }
1712
1713 static void dsputil_init_mmxext(DSPContext *c, AVCodecContext *avctx,
1714                                 int mm_flags)
1715 {
1716     const int bit_depth      = avctx->bits_per_raw_sample;
1717     const int high_bit_depth = bit_depth > 8;
1718
1719 #if HAVE_YASM
1720     SET_QPEL_FUNCS(avg_qpel,        0, 16, mmxext, );
1721     SET_QPEL_FUNCS(avg_qpel,        1,  8, mmxext, );
1722
1723     SET_QPEL_FUNCS(put_qpel,        0, 16, mmxext, );
1724     SET_QPEL_FUNCS(put_qpel,        1,  8, mmxext, );
1725     SET_QPEL_FUNCS(put_no_rnd_qpel, 0, 16, mmxext, );
1726     SET_QPEL_FUNCS(put_no_rnd_qpel, 1,  8, mmxext, );
1727
1728     if (!high_bit_depth) {
1729         c->put_pixels_tab[0][1] = ff_put_pixels16_x2_mmxext;
1730         c->put_pixels_tab[0][2] = ff_put_pixels16_y2_mmxext;
1731
1732         c->avg_pixels_tab[0][0] = ff_avg_pixels16_mmxext;
1733         c->avg_pixels_tab[0][1] = ff_avg_pixels16_x2_mmxext;
1734         c->avg_pixels_tab[0][2] = ff_avg_pixels16_y2_mmxext;
1735
1736         c->put_pixels_tab[1][1] = ff_put_pixels8_x2_mmxext;
1737         c->put_pixels_tab[1][2] = ff_put_pixels8_y2_mmxext;
1738
1739         c->avg_pixels_tab[1][0] = ff_avg_pixels8_mmxext;
1740         c->avg_pixels_tab[1][1] = ff_avg_pixels8_x2_mmxext;
1741         c->avg_pixels_tab[1][2] = ff_avg_pixels8_y2_mmxext;
1742     }
1743
1744     if (!(avctx->flags & CODEC_FLAG_BITEXACT)) {
1745         if (!high_bit_depth) {
1746             c->put_no_rnd_pixels_tab[0][1] = ff_put_no_rnd_pixels16_x2_mmxext;
1747             c->put_no_rnd_pixels_tab[0][2] = ff_put_no_rnd_pixels16_y2_mmxext;
1748             c->put_no_rnd_pixels_tab[1][1] = ff_put_no_rnd_pixels8_x2_mmxext;
1749             c->put_no_rnd_pixels_tab[1][2] = ff_put_no_rnd_pixels8_y2_mmxext;
1750
1751             c->avg_pixels_tab[0][3] = ff_avg_pixels16_xy2_mmxext;
1752             c->avg_pixels_tab[1][3] = ff_avg_pixels8_xy2_mmxext;
1753         }
1754     }
1755 #endif /* HAVE_YASM */
1756
1757 #if HAVE_INLINE_ASM
1758     if (!high_bit_depth && avctx->idct_algo == FF_IDCT_XVIDMMX) {
1759         c->idct_put = ff_idct_xvid_mmxext_put;
1760         c->idct_add = ff_idct_xvid_mmxext_add;
1761         c->idct     = ff_idct_xvid_mmxext;
1762     }
1763 #endif /* HAVE_INLINE_ASM */
1764
1765 #if HAVE_MMXEXT_EXTERNAL
1766     if (CONFIG_VP3_DECODER && (avctx->codec_id == AV_CODEC_ID_VP3 ||
1767                                avctx->codec_id == AV_CODEC_ID_THEORA)) {
1768         c->put_no_rnd_pixels_tab[1][1] = ff_put_no_rnd_pixels8_x2_exact_mmxext;
1769         c->put_no_rnd_pixels_tab[1][2] = ff_put_no_rnd_pixels8_y2_exact_mmxext;
1770     }
1771
1772     if (!high_bit_depth && CONFIG_H264CHROMA) {
1773         c->avg_h264_chroma_pixels_tab[0] = ff_avg_h264_chroma_mc8_rnd_mmxext;
1774         c->avg_h264_chroma_pixels_tab[1] = ff_avg_h264_chroma_mc4_mmxext;
1775         c->avg_h264_chroma_pixels_tab[2] = ff_avg_h264_chroma_mc2_mmxext;
1776         c->put_h264_chroma_pixels_tab[2] = ff_put_h264_chroma_mc2_mmxext;
1777     }
1778     if (bit_depth == 10 && CONFIG_H264CHROMA) {
1779         c->put_h264_chroma_pixels_tab[2] = ff_put_h264_chroma_mc2_10_mmxext;
1780         c->avg_h264_chroma_pixels_tab[2] = ff_avg_h264_chroma_mc2_10_mmxext;
1781         c->put_h264_chroma_pixels_tab[1] = ff_put_h264_chroma_mc4_10_mmxext;
1782         c->avg_h264_chroma_pixels_tab[1] = ff_avg_h264_chroma_mc4_10_mmxext;
1783     }
1784
1785     /* slower than cmov version on AMD */
1786     if (!(mm_flags & AV_CPU_FLAG_3DNOW))
1787         c->add_hfyu_median_prediction = ff_add_hfyu_median_prediction_mmxext;
1788
1789     c->scalarproduct_int16          = ff_scalarproduct_int16_mmxext;
1790     c->scalarproduct_and_madd_int16 = ff_scalarproduct_and_madd_int16_mmxext;
1791
1792     if (avctx->flags & CODEC_FLAG_BITEXACT) {
1793         c->apply_window_int16 = ff_apply_window_int16_mmxext;
1794     } else {
1795         c->apply_window_int16 = ff_apply_window_int16_round_mmxext;
1796     }
1797 #endif /* HAVE_MMXEXT_EXTERNAL */
1798 }
1799
1800 static void dsputil_init_3dnow(DSPContext *c, AVCodecContext *avctx,
1801                                int mm_flags)
1802 {
1803     const int high_bit_depth = avctx->bits_per_raw_sample > 8;
1804
1805 #if HAVE_YASM
1806     if (!high_bit_depth) {
1807         c->put_pixels_tab[0][1] = ff_put_pixels16_x2_3dnow;
1808         c->put_pixels_tab[0][2] = ff_put_pixels16_y2_3dnow;
1809
1810         c->avg_pixels_tab[0][0] = ff_avg_pixels16_3dnow;
1811         c->avg_pixels_tab[0][1] = ff_avg_pixels16_x2_3dnow;
1812         c->avg_pixels_tab[0][2] = ff_avg_pixels16_y2_3dnow;
1813
1814         c->put_pixels_tab[1][1] = ff_put_pixels8_x2_3dnow;
1815         c->put_pixels_tab[1][2] = ff_put_pixels8_y2_3dnow;
1816
1817         c->avg_pixels_tab[1][0] = ff_avg_pixels8_3dnow;
1818         c->avg_pixels_tab[1][1] = ff_avg_pixels8_x2_3dnow;
1819         c->avg_pixels_tab[1][2] = ff_avg_pixels8_y2_3dnow;
1820
1821         if (!(avctx->flags & CODEC_FLAG_BITEXACT)){
1822             c->put_no_rnd_pixels_tab[0][1] = ff_put_no_rnd_pixels16_x2_3dnow;
1823             c->put_no_rnd_pixels_tab[0][2] = ff_put_no_rnd_pixels16_y2_3dnow;
1824             c->put_no_rnd_pixels_tab[1][1] = ff_put_no_rnd_pixels8_x2_3dnow;
1825             c->put_no_rnd_pixels_tab[1][2] = ff_put_no_rnd_pixels8_y2_3dnow;
1826
1827             c->avg_pixels_tab[0][3] = ff_avg_pixels16_xy2_3dnow;
1828             c->avg_pixels_tab[1][3] = ff_avg_pixels8_xy2_3dnow;
1829         }
1830     }
1831
1832     if (CONFIG_VP3_DECODER && (avctx->codec_id == AV_CODEC_ID_VP3 ||
1833                                avctx->codec_id == AV_CODEC_ID_THEORA)) {
1834         c->put_no_rnd_pixels_tab[1][1] = ff_put_no_rnd_pixels8_x2_exact_3dnow;
1835         c->put_no_rnd_pixels_tab[1][2] = ff_put_no_rnd_pixels8_y2_exact_3dnow;
1836     }
1837
1838     if (!high_bit_depth && CONFIG_H264CHROMA) {
1839         c->avg_h264_chroma_pixels_tab[0] = ff_avg_h264_chroma_mc8_rnd_3dnow;
1840         c->avg_h264_chroma_pixels_tab[1] = ff_avg_h264_chroma_mc4_3dnow;
1841     }
1842 #endif /* HAVE_YASM */
1843 }
1844
1845 static void dsputil_init_sse(DSPContext *c, AVCodecContext *avctx, int mm_flags)
1846 {
1847     const int high_bit_depth = avctx->bits_per_raw_sample > 8;
1848
1849 #if HAVE_INLINE_ASM
1850     if (!high_bit_depth) {
1851         if (!(CONFIG_MPEG_XVMC_DECODER && avctx->xvmc_acceleration > 1)) {
1852             /* XvMCCreateBlocks() may not allocate 16-byte aligned blocks */
1853             c->clear_block  = clear_block_sse;
1854             c->clear_blocks = clear_blocks_sse;
1855         }
1856     }
1857
1858     c->vector_clipf = vector_clipf_sse;
1859 #endif /* HAVE_INLINE_ASM */
1860 }
1861
1862 static void dsputil_init_sse2(DSPContext *c, AVCodecContext *avctx,
1863                               int mm_flags)
1864 {
1865     const int bit_depth      = avctx->bits_per_raw_sample;
1866     const int high_bit_depth = bit_depth > 8;
1867
1868 #if HAVE_SSE2_INLINE
1869     if (!high_bit_depth && avctx->idct_algo == FF_IDCT_XVIDMMX) {
1870         c->idct_put              = ff_idct_xvid_sse2_put;
1871         c->idct_add              = ff_idct_xvid_sse2_add;
1872         c->idct                  = ff_idct_xvid_sse2;
1873         c->idct_permutation_type = FF_SSE2_IDCT_PERM;
1874     }
1875 #endif /* HAVE_SSE2_INLINE */
1876
1877 #if HAVE_SSE2_EXTERNAL
1878     if (!(mm_flags & AV_CPU_FLAG_SSE2SLOW)) {
1879         // these functions are slower than mmx on AMD, but faster on Intel
1880         if (!high_bit_depth) {
1881             c->put_pixels_tab[0][0]        = ff_put_pixels16_sse2;
1882             c->put_no_rnd_pixels_tab[0][0] = ff_put_pixels16_sse2;
1883             c->avg_pixels_tab[0][0]        = ff_avg_pixels16_sse2;
1884         }
1885     }
1886
1887     if (bit_depth == 10) {
1888         if (CONFIG_H264CHROMA) {
1889             c->put_h264_chroma_pixels_tab[0] = ff_put_h264_chroma_mc8_10_sse2;
1890             c->avg_h264_chroma_pixels_tab[0] = ff_avg_h264_chroma_mc8_10_sse2;
1891         }
1892     }
1893
1894     c->scalarproduct_int16          = ff_scalarproduct_int16_sse2;
1895     c->scalarproduct_and_madd_int16 = ff_scalarproduct_and_madd_int16_sse2;
1896     if (mm_flags & AV_CPU_FLAG_ATOM) {
1897         c->vector_clip_int32 = ff_vector_clip_int32_int_sse2;
1898     } else {
1899         c->vector_clip_int32 = ff_vector_clip_int32_sse2;
1900     }
1901     if (avctx->flags & CODEC_FLAG_BITEXACT) {
1902         c->apply_window_int16 = ff_apply_window_int16_sse2;
1903     } else if (!(mm_flags & AV_CPU_FLAG_SSE2SLOW)) {
1904         c->apply_window_int16 = ff_apply_window_int16_round_sse2;
1905     }
1906     c->bswap_buf = ff_bswap32_buf_sse2;
1907 #endif /* HAVE_SSE2_EXTERNAL */
1908 }
1909
1910 static void dsputil_init_ssse3(DSPContext *c, AVCodecContext *avctx,
1911                                int mm_flags)
1912 {
1913 #if HAVE_SSSE3_EXTERNAL
1914     const int high_bit_depth = avctx->bits_per_raw_sample > 8;
1915
1916     if (!high_bit_depth && CONFIG_H264CHROMA) {
1917         c->put_h264_chroma_pixels_tab[0] = ff_put_h264_chroma_mc8_rnd_ssse3;
1918         c->avg_h264_chroma_pixels_tab[0] = ff_avg_h264_chroma_mc8_rnd_ssse3;
1919         c->put_h264_chroma_pixels_tab[1] = ff_put_h264_chroma_mc4_ssse3;
1920         c->avg_h264_chroma_pixels_tab[1] = ff_avg_h264_chroma_mc4_ssse3;
1921     }
1922     c->add_hfyu_left_prediction = ff_add_hfyu_left_prediction_ssse3;
1923     if (mm_flags & AV_CPU_FLAG_SSE4) // not really sse4, just slow on Conroe
1924         c->add_hfyu_left_prediction = ff_add_hfyu_left_prediction_sse4;
1925
1926     if (mm_flags & AV_CPU_FLAG_ATOM)
1927         c->apply_window_int16 = ff_apply_window_int16_ssse3_atom;
1928     else
1929         c->apply_window_int16 = ff_apply_window_int16_ssse3;
1930     if (!(mm_flags & (AV_CPU_FLAG_SSE42|AV_CPU_FLAG_3DNOW))) // cachesplit
1931         c->scalarproduct_and_madd_int16 = ff_scalarproduct_and_madd_int16_ssse3;
1932     c->bswap_buf = ff_bswap32_buf_ssse3;
1933 #endif /* HAVE_SSSE3_EXTERNAL */
1934 }
1935
1936 static void dsputil_init_sse4(DSPContext *c, AVCodecContext *avctx,
1937                               int mm_flags)
1938 {
1939 #if HAVE_SSE4_EXTERNAL
1940     c->vector_clip_int32 = ff_vector_clip_int32_sse4;
1941 #endif /* HAVE_SSE4_EXTERNAL */
1942 }
1943
1944 static void dsputil_init_avx(DSPContext *c, AVCodecContext *avctx, int mm_flags)
1945 {
1946 #if HAVE_AVX_EXTERNAL
1947     const int bit_depth = avctx->bits_per_raw_sample;
1948
1949     if (bit_depth == 10) {
1950         if (CONFIG_H264CHROMA) {
1951             c->put_h264_chroma_pixels_tab[0] = ff_put_h264_chroma_mc8_10_avx;
1952             c->avg_h264_chroma_pixels_tab[0] = ff_avg_h264_chroma_mc8_10_avx;
1953         }
1954     }
1955 #endif /* HAVE_AVX_EXTERNAL */
1956 }
1957
1958 void ff_dsputil_init_mmx(DSPContext *c, AVCodecContext *avctx)
1959 {
1960     int mm_flags = av_get_cpu_flags();
1961
1962 #if HAVE_7REGS && HAVE_INLINE_ASM
1963     if (mm_flags & AV_CPU_FLAG_CMOV)
1964         c->add_hfyu_median_prediction = add_hfyu_median_prediction_cmov;
1965 #endif
1966
1967     if (mm_flags & AV_CPU_FLAG_MMX)
1968         dsputil_init_mmx(c, avctx, mm_flags);
1969
1970     if (mm_flags & AV_CPU_FLAG_MMXEXT)
1971         dsputil_init_mmxext(c, avctx, mm_flags);
1972
1973     if (mm_flags & AV_CPU_FLAG_3DNOW)
1974         dsputil_init_3dnow(c, avctx, mm_flags);
1975
1976     if (mm_flags & AV_CPU_FLAG_SSE)
1977         dsputil_init_sse(c, avctx, mm_flags);
1978
1979     if (mm_flags & AV_CPU_FLAG_SSE2)
1980         dsputil_init_sse2(c, avctx, mm_flags);
1981
1982     if (mm_flags & AV_CPU_FLAG_SSSE3)
1983         dsputil_init_ssse3(c, avctx, mm_flags);
1984
1985     if (mm_flags & AV_CPU_FLAG_SSE4)
1986         dsputil_init_sse4(c, avctx, mm_flags);
1987
1988     if (mm_flags & AV_CPU_FLAG_AVX)
1989         dsputil_init_avx(c, avctx, mm_flags);
1990
1991     if (CONFIG_ENCODERS)
1992         ff_dsputilenc_init_mmx(c, avctx);
1993 }