503764817a119decdf9b38a427af15f4f1a57788
[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 uint64_t, ff_pdw_80000000)[2] =
42     { 0x8000000080000000ULL, 0x8000000080000000ULL };
43
44 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_1)    = { 0x0001000100010001ULL, 0x0001000100010001ULL };
45 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_2)    = { 0x0002000200020002ULL, 0x0002000200020002ULL };
46 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_3)    = { 0x0003000300030003ULL, 0x0003000300030003ULL };
47 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_4)    = { 0x0004000400040004ULL, 0x0004000400040004ULL };
48 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_5)    = { 0x0005000500050005ULL, 0x0005000500050005ULL };
49 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_8)    = { 0x0008000800080008ULL, 0x0008000800080008ULL };
50 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_9)    = { 0x0009000900090009ULL, 0x0009000900090009ULL };
51 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_15)   =   0x000F000F000F000FULL;
52 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_16)   = { 0x0010001000100010ULL, 0x0010001000100010ULL };
53 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_17)   = { 0x0011001100110011ULL, 0x0011001100110011ULL };
54 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_18)   = { 0x0012001200120012ULL, 0x0012001200120012ULL };
55 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_20)   =   0x0014001400140014ULL;
56 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_27)   = { 0x001B001B001B001BULL, 0x001B001B001B001BULL };
57 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_28)   = { 0x001C001C001C001CULL, 0x001C001C001C001CULL };
58 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_32)   = { 0x0020002000200020ULL, 0x0020002000200020ULL };
59 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_42)   =   0x002A002A002A002AULL;
60 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_53)   =   0x0035003500350035ULL;
61 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_63)   = { 0x003F003F003F003FULL, 0x003F003F003F003FULL };
62 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_64)   = { 0x0040004000400040ULL, 0x0040004000400040ULL };
63 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_96)   =   0x0060006000600060ULL;
64 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_128)  =   0x0080008000800080ULL;
65 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_255)  =   0x00ff00ff00ff00ffULL;
66 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_512)  = { 0x0200020002000200ULL, 0x0200020002000200ULL };
67 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_1019) = { 0x03FB03FB03FB03FBULL, 0x03FB03FB03FB03FBULL };
68
69 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_0)    = { 0x0000000000000000ULL, 0x0000000000000000ULL };
70 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_1)    = { 0x0101010101010101ULL, 0x0101010101010101ULL };
71 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_3)    = { 0x0303030303030303ULL, 0x0303030303030303ULL };
72 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_4)    = { 0x0404040404040404ULL, 0x0404040404040404ULL };
73 DECLARE_ALIGNED(8,  const uint64_t, ff_pb_7)    =   0x0707070707070707ULL;
74 DECLARE_ALIGNED(8,  const uint64_t, ff_pb_1F)   =   0x1F1F1F1F1F1F1F1FULL;
75 DECLARE_ALIGNED(8,  const uint64_t, ff_pb_3F)   =   0x3F3F3F3F3F3F3F3FULL;
76 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_80)   = { 0x8080808080808080ULL, 0x8080808080808080ULL };
77 DECLARE_ALIGNED(8,  const uint64_t, ff_pb_81)   =   0x8181818181818181ULL;
78 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_A1)   = { 0xA1A1A1A1A1A1A1A1ULL, 0xA1A1A1A1A1A1A1A1ULL };
79 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_F8)   = { 0xF8F8F8F8F8F8F8F8ULL, 0xF8F8F8F8F8F8F8F8ULL };
80 DECLARE_ALIGNED(8,  const uint64_t, ff_pb_FC)   =   0xFCFCFCFCFCFCFCFCULL;
81 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_FE)   = { 0xFEFEFEFEFEFEFEFEULL, 0xFEFEFEFEFEFEFEFEULL };
82
83 DECLARE_ALIGNED(16, const double, ff_pd_1)[2] = { 1.0, 1.0 };
84 DECLARE_ALIGNED(16, const double, ff_pd_2)[2] = { 2.0, 2.0 };
85
86 #if HAVE_INLINE_ASM
87
88 #define JUMPALIGN()     __asm__ volatile (".p2align 3"::)
89 #define MOVQ_ZERO(regd) __asm__ volatile ("pxor %%"#regd", %%"#regd ::)
90
91 #define MOVQ_BFE(regd)                                  \
92     __asm__ volatile (                                  \
93         "pcmpeqd %%"#regd", %%"#regd"   \n\t"           \
94         "paddb   %%"#regd", %%"#regd"   \n\t" ::)
95
96 #ifndef PIC
97 #define MOVQ_BONE(regd) __asm__ volatile ("movq %0, %%"#regd" \n\t" :: "m"(ff_bone))
98 #define MOVQ_WTWO(regd) __asm__ volatile ("movq %0, %%"#regd" \n\t" :: "m"(ff_wtwo))
99 #else
100 // for shared library it's better to use this way for accessing constants
101 // pcmpeqd -> -1
102 #define MOVQ_BONE(regd)                                 \
103     __asm__ volatile (                                  \
104         "pcmpeqd  %%"#regd", %%"#regd"  \n\t"           \
105         "psrlw          $15, %%"#regd"  \n\t"           \
106         "packuswb %%"#regd", %%"#regd"  \n\t" ::)
107
108 #define MOVQ_WTWO(regd)                                 \
109     __asm__ volatile (                                  \
110         "pcmpeqd %%"#regd", %%"#regd"   \n\t"           \
111         "psrlw         $15, %%"#regd"   \n\t"           \
112         "psllw          $1, %%"#regd"   \n\t"::)
113
114 #endif
115
116 // using regr as temporary and for the output result
117 // first argument is unmodifed and second is trashed
118 // regfe is supposed to contain 0xfefefefefefefefe
119 #define PAVGB_MMX_NO_RND(rega, regb, regr, regfe)                \
120     "movq   "#rega", "#regr"            \n\t"                    \
121     "pand   "#regb", "#regr"            \n\t"                    \
122     "pxor   "#rega", "#regb"            \n\t"                    \
123     "pand  "#regfe", "#regb"            \n\t"                    \
124     "psrlq       $1, "#regb"            \n\t"                    \
125     "paddb  "#regb", "#regr"            \n\t"
126
127 #define PAVGB_MMX(rega, regb, regr, regfe)                       \
128     "movq   "#rega", "#regr"            \n\t"                    \
129     "por    "#regb", "#regr"            \n\t"                    \
130     "pxor   "#rega", "#regb"            \n\t"                    \
131     "pand  "#regfe", "#regb"            \n\t"                    \
132     "psrlq       $1, "#regb"            \n\t"                    \
133     "psubb  "#regb", "#regr"            \n\t"
134
135 // mm6 is supposed to contain 0xfefefefefefefefe
136 #define PAVGBP_MMX_NO_RND(rega, regb, regr,  regc, regd, regp)   \
137     "movq  "#rega", "#regr"             \n\t"                    \
138     "movq  "#regc", "#regp"             \n\t"                    \
139     "pand  "#regb", "#regr"             \n\t"                    \
140     "pand  "#regd", "#regp"             \n\t"                    \
141     "pxor  "#rega", "#regb"             \n\t"                    \
142     "pxor  "#regc", "#regd"             \n\t"                    \
143     "pand    %%mm6, "#regb"             \n\t"                    \
144     "pand    %%mm6, "#regd"             \n\t"                    \
145     "psrlq      $1, "#regb"             \n\t"                    \
146     "psrlq      $1, "#regd"             \n\t"                    \
147     "paddb "#regb", "#regr"             \n\t"                    \
148     "paddb "#regd", "#regp"             \n\t"
149
150 #define PAVGBP_MMX(rega, regb, regr, regc, regd, regp)           \
151     "movq  "#rega", "#regr"             \n\t"                    \
152     "movq  "#regc", "#regp"             \n\t"                    \
153     "por   "#regb", "#regr"             \n\t"                    \
154     "por   "#regd", "#regp"             \n\t"                    \
155     "pxor  "#rega", "#regb"             \n\t"                    \
156     "pxor  "#regc", "#regd"             \n\t"                    \
157     "pand    %%mm6, "#regb"             \n\t"                    \
158     "pand    %%mm6, "#regd"             \n\t"                    \
159     "psrlq      $1, "#regd"             \n\t"                    \
160     "psrlq      $1, "#regb"             \n\t"                    \
161     "psubb "#regb", "#regr"             \n\t"                    \
162     "psubb "#regd", "#regp"             \n\t"
163
164 /***********************************/
165 /* MMX no rounding */
166 #define DEF(x, y) x ## _no_rnd_ ## y ## _mmx
167 #define SET_RND  MOVQ_WONE
168 #define PAVGBP(a, b, c, d, e, f)        PAVGBP_MMX_NO_RND(a, b, c, d, e, f)
169 #define PAVGB(a, b, c, e)               PAVGB_MMX_NO_RND(a, b, c, e)
170 #define OP_AVG(a, b, c, e)              PAVGB_MMX(a, b, c, e)
171
172 #include "dsputil_rnd_template.c"
173
174 #undef DEF
175 #undef SET_RND
176 #undef PAVGBP
177 #undef PAVGB
178 /***********************************/
179 /* MMX rounding */
180
181 #define DEF(x, y) x ## _ ## y ## _mmx
182 #define SET_RND  MOVQ_WTWO
183 #define PAVGBP(a, b, c, d, e, f)        PAVGBP_MMX(a, b, c, d, e, f)
184 #define PAVGB(a, b, c, e)               PAVGB_MMX(a, b, c, e)
185
186 #include "dsputil_rnd_template.c"
187
188 #undef DEF
189 #undef SET_RND
190 #undef PAVGBP
191 #undef PAVGB
192 #undef OP_AVG
193
194 /***********************************/
195 /* 3Dnow specific */
196
197 #define DEF(x) x ## _3dnow
198 #define PAVGB "pavgusb"
199 #define SKIP_FOR_3DNOW
200
201 #include "dsputil_avg_template.c"
202
203 #undef DEF
204 #undef PAVGB
205 #undef SKIP_FOR_3DNOW
206
207 /***********************************/
208 /* MMXEXT specific */
209
210 #define DEF(x) x ## _mmxext
211
212 /* Introduced only in MMXEXT set */
213 #define PAVGB "pavgb"
214
215 #include "dsputil_avg_template.c"
216
217 #undef DEF
218 #undef PAVGB
219
220 #define put_no_rnd_pixels16_mmx put_pixels16_mmx
221 #define put_no_rnd_pixels8_mmx put_pixels8_mmx
222 #define put_pixels16_mmxext put_pixels16_mmx
223 #define put_pixels8_mmxext put_pixels8_mmx
224 #define put_pixels4_mmxext put_pixels4_mmx
225 #define put_no_rnd_pixels16_mmxext put_no_rnd_pixels16_mmx
226 #define put_no_rnd_pixels8_mmxext put_no_rnd_pixels8_mmx
227
228 /***********************************/
229 /* standard MMX */
230
231 void ff_put_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels,
232                                int line_size)
233 {
234     const DCTELEM *p;
235     uint8_t *pix;
236
237     /* read the pixels */
238     p   = block;
239     pix = pixels;
240     /* unrolled loop */
241     __asm__ volatile (
242         "movq      (%3), %%mm0          \n\t"
243         "movq     8(%3), %%mm1          \n\t"
244         "movq    16(%3), %%mm2          \n\t"
245         "movq    24(%3), %%mm3          \n\t"
246         "movq    32(%3), %%mm4          \n\t"
247         "movq    40(%3), %%mm5          \n\t"
248         "movq    48(%3), %%mm6          \n\t"
249         "movq    56(%3), %%mm7          \n\t"
250         "packuswb %%mm1, %%mm0          \n\t"
251         "packuswb %%mm3, %%mm2          \n\t"
252         "packuswb %%mm5, %%mm4          \n\t"
253         "packuswb %%mm7, %%mm6          \n\t"
254         "movq     %%mm0, (%0)           \n\t"
255         "movq     %%mm2, (%0, %1)       \n\t"
256         "movq     %%mm4, (%0, %1, 2)    \n\t"
257         "movq     %%mm6, (%0, %2)       \n\t"
258         :: "r"(pix), "r"((x86_reg)line_size), "r"((x86_reg)line_size * 3),
259            "r"(p)
260         : "memory");
261     pix += line_size * 4;
262     p   += 32;
263
264     // if here would be an exact copy of the code above
265     // compiler would generate some very strange code
266     // thus using "r"
267     __asm__ volatile (
268         "movq       (%3), %%mm0         \n\t"
269         "movq      8(%3), %%mm1         \n\t"
270         "movq     16(%3), %%mm2         \n\t"
271         "movq     24(%3), %%mm3         \n\t"
272         "movq     32(%3), %%mm4         \n\t"
273         "movq     40(%3), %%mm5         \n\t"
274         "movq     48(%3), %%mm6         \n\t"
275         "movq     56(%3), %%mm7         \n\t"
276         "packuswb  %%mm1, %%mm0         \n\t"
277         "packuswb  %%mm3, %%mm2         \n\t"
278         "packuswb  %%mm5, %%mm4         \n\t"
279         "packuswb  %%mm7, %%mm6         \n\t"
280         "movq      %%mm0, (%0)          \n\t"
281         "movq      %%mm2, (%0, %1)      \n\t"
282         "movq      %%mm4, (%0, %1, 2)   \n\t"
283         "movq      %%mm6, (%0, %2)      \n\t"
284         :: "r"(pix), "r"((x86_reg)line_size), "r"((x86_reg)line_size * 3), "r"(p)
285         : "memory");
286 }
287
288 #define put_signed_pixels_clamped_mmx_half(off)             \
289     "movq          "#off"(%2), %%mm1        \n\t"           \
290     "movq     16 + "#off"(%2), %%mm2        \n\t"           \
291     "movq     32 + "#off"(%2), %%mm3        \n\t"           \
292     "movq     48 + "#off"(%2), %%mm4        \n\t"           \
293     "packsswb  8 + "#off"(%2), %%mm1        \n\t"           \
294     "packsswb 24 + "#off"(%2), %%mm2        \n\t"           \
295     "packsswb 40 + "#off"(%2), %%mm3        \n\t"           \
296     "packsswb 56 + "#off"(%2), %%mm4        \n\t"           \
297     "paddb              %%mm0, %%mm1        \n\t"           \
298     "paddb              %%mm0, %%mm2        \n\t"           \
299     "paddb              %%mm0, %%mm3        \n\t"           \
300     "paddb              %%mm0, %%mm4        \n\t"           \
301     "movq               %%mm1, (%0)         \n\t"           \
302     "movq               %%mm2, (%0, %3)     \n\t"           \
303     "movq               %%mm3, (%0, %3, 2)  \n\t"           \
304     "movq               %%mm4, (%0, %1)     \n\t"
305
306 void ff_put_signed_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels,
307                                       int line_size)
308 {
309     x86_reg line_skip = line_size;
310     x86_reg line_skip3;
311
312     __asm__ volatile (
313         "movq "MANGLE(ff_pb_80)", %%mm0     \n\t"
314         "lea         (%3, %3, 2), %1        \n\t"
315         put_signed_pixels_clamped_mmx_half(0)
316         "lea         (%0, %3, 4), %0        \n\t"
317         put_signed_pixels_clamped_mmx_half(64)
318         : "+&r"(pixels), "=&r"(line_skip3)
319         : "r"(block), "r"(line_skip)
320         : "memory");
321 }
322
323 void ff_add_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels,
324                                int line_size)
325 {
326     const DCTELEM *p;
327     uint8_t *pix;
328     int i;
329
330     /* read the pixels */
331     p   = block;
332     pix = pixels;
333     MOVQ_ZERO(mm7);
334     i = 4;
335     do {
336         __asm__ volatile (
337             "movq        (%2), %%mm0    \n\t"
338             "movq       8(%2), %%mm1    \n\t"
339             "movq      16(%2), %%mm2    \n\t"
340             "movq      24(%2), %%mm3    \n\t"
341             "movq          %0, %%mm4    \n\t"
342             "movq          %1, %%mm6    \n\t"
343             "movq       %%mm4, %%mm5    \n\t"
344             "punpcklbw  %%mm7, %%mm4    \n\t"
345             "punpckhbw  %%mm7, %%mm5    \n\t"
346             "paddsw     %%mm4, %%mm0    \n\t"
347             "paddsw     %%mm5, %%mm1    \n\t"
348             "movq       %%mm6, %%mm5    \n\t"
349             "punpcklbw  %%mm7, %%mm6    \n\t"
350             "punpckhbw  %%mm7, %%mm5    \n\t"
351             "paddsw     %%mm6, %%mm2    \n\t"
352             "paddsw     %%mm5, %%mm3    \n\t"
353             "packuswb   %%mm1, %%mm0    \n\t"
354             "packuswb   %%mm3, %%mm2    \n\t"
355             "movq       %%mm0, %0       \n\t"
356             "movq       %%mm2, %1       \n\t"
357             : "+m"(*pix), "+m"(*(pix + line_size))
358             : "r"(p)
359             : "memory");
360         pix += line_size * 2;
361         p   += 16;
362     } while (--i);
363 }
364
365 static void put_pixels8_mmx(uint8_t *block, const uint8_t *pixels,
366                             int line_size, int h)
367 {
368     __asm__ volatile (
369         "lea   (%3, %3), %%"REG_a"      \n\t"
370         ".p2align     3                 \n\t"
371         "1:                             \n\t"
372         "movq  (%1    ), %%mm0          \n\t"
373         "movq  (%1, %3), %%mm1          \n\t"
374         "movq     %%mm0, (%2)           \n\t"
375         "movq     %%mm1, (%2, %3)       \n\t"
376         "add  %%"REG_a", %1             \n\t"
377         "add  %%"REG_a", %2             \n\t"
378         "movq  (%1    ), %%mm0          \n\t"
379         "movq  (%1, %3), %%mm1          \n\t"
380         "movq     %%mm0, (%2)           \n\t"
381         "movq     %%mm1, (%2, %3)       \n\t"
382         "add  %%"REG_a", %1             \n\t"
383         "add  %%"REG_a", %2             \n\t"
384         "subl        $4, %0             \n\t"
385         "jnz         1b                 \n\t"
386         : "+g"(h), "+r"(pixels),  "+r"(block)
387         : "r"((x86_reg)line_size)
388         : "%"REG_a, "memory"
389         );
390 }
391
392 static void put_pixels16_mmx(uint8_t *block, const uint8_t *pixels,
393                              int line_size, int h)
394 {
395     __asm__ volatile (
396         "lea   (%3, %3), %%"REG_a"      \n\t"
397         ".p2align     3                 \n\t"
398         "1:                             \n\t"
399         "movq  (%1    ), %%mm0          \n\t"
400         "movq 8(%1    ), %%mm4          \n\t"
401         "movq  (%1, %3), %%mm1          \n\t"
402         "movq 8(%1, %3), %%mm5          \n\t"
403         "movq     %%mm0,  (%2)          \n\t"
404         "movq     %%mm4, 8(%2)          \n\t"
405         "movq     %%mm1,  (%2, %3)      \n\t"
406         "movq     %%mm5, 8(%2, %3)      \n\t"
407         "add  %%"REG_a", %1             \n\t"
408         "add  %%"REG_a", %2             \n\t"
409         "movq  (%1    ), %%mm0          \n\t"
410         "movq 8(%1    ), %%mm4          \n\t"
411         "movq  (%1, %3), %%mm1          \n\t"
412         "movq 8(%1, %3), %%mm5          \n\t"
413         "movq     %%mm0,  (%2)          \n\t"
414         "movq     %%mm4, 8(%2)          \n\t"
415         "movq     %%mm1,  (%2, %3)      \n\t"
416         "movq     %%mm5, 8(%2, %3)      \n\t"
417         "add  %%"REG_a", %1             \n\t"
418         "add  %%"REG_a", %2             \n\t"
419         "subl        $4, %0             \n\t"
420         "jnz         1b                 \n\t"
421         : "+g"(h), "+r"(pixels),  "+r"(block)
422         : "r"((x86_reg)line_size)
423         : "%"REG_a, "memory"
424         );
425 }
426
427 #define CLEAR_BLOCKS(name, n)                           \
428 static void name(DCTELEM *blocks)                       \
429 {                                                       \
430     __asm__ volatile (                                  \
431         "pxor %%mm7, %%mm7              \n\t"           \
432         "mov     %1,        %%"REG_a"   \n\t"           \
433         "1:                             \n\t"           \
434         "movq %%mm7,   (%0, %%"REG_a")  \n\t"           \
435         "movq %%mm7,  8(%0, %%"REG_a")  \n\t"           \
436         "movq %%mm7, 16(%0, %%"REG_a")  \n\t"           \
437         "movq %%mm7, 24(%0, %%"REG_a")  \n\t"           \
438         "add    $32, %%"REG_a"          \n\t"           \
439         "js      1b                     \n\t"           \
440         :: "r"(((uint8_t *)blocks) + 128 * n),          \
441            "i"(-128 * n)                                \
442         : "%"REG_a                                      \
443         );                                              \
444 }
445 CLEAR_BLOCKS(clear_blocks_mmx, 6)
446 CLEAR_BLOCKS(clear_block_mmx, 1)
447
448 static void clear_block_sse(DCTELEM *block)
449 {
450     __asm__ volatile (
451         "xorps  %%xmm0, %%xmm0          \n"
452         "movaps %%xmm0,    (%0)         \n"
453         "movaps %%xmm0,  16(%0)         \n"
454         "movaps %%xmm0,  32(%0)         \n"
455         "movaps %%xmm0,  48(%0)         \n"
456         "movaps %%xmm0,  64(%0)         \n"
457         "movaps %%xmm0,  80(%0)         \n"
458         "movaps %%xmm0,  96(%0)         \n"
459         "movaps %%xmm0, 112(%0)         \n"
460         :: "r"(block)
461         : "memory"
462     );
463 }
464
465 static void clear_blocks_sse(DCTELEM *blocks)
466 {
467     __asm__ volatile (
468         "xorps  %%xmm0, %%xmm0              \n"
469         "mov        %1,         %%"REG_a"   \n"
470         "1:                                 \n"
471         "movaps %%xmm0,    (%0, %%"REG_a")  \n"
472         "movaps %%xmm0,  16(%0, %%"REG_a")  \n"
473         "movaps %%xmm0,  32(%0, %%"REG_a")  \n"
474         "movaps %%xmm0,  48(%0, %%"REG_a")  \n"
475         "movaps %%xmm0,  64(%0, %%"REG_a")  \n"
476         "movaps %%xmm0,  80(%0, %%"REG_a")  \n"
477         "movaps %%xmm0,  96(%0, %%"REG_a")  \n"
478         "movaps %%xmm0, 112(%0, %%"REG_a")  \n"
479         "add      $128,         %%"REG_a"   \n"
480         "js         1b                      \n"
481         :: "r"(((uint8_t *)blocks) + 128 * 6),
482            "i"(-128 * 6)
483         : "%"REG_a
484     );
485 }
486
487 static void add_bytes_mmx(uint8_t *dst, uint8_t *src, int w)
488 {
489     x86_reg i = 0;
490     __asm__ volatile (
491         "jmp          2f                \n\t"
492         "1:                             \n\t"
493         "movq   (%1, %0), %%mm0         \n\t"
494         "movq   (%2, %0), %%mm1         \n\t"
495         "paddb     %%mm0, %%mm1         \n\t"
496         "movq      %%mm1, (%2, %0)      \n\t"
497         "movq  8(%1, %0), %%mm0         \n\t"
498         "movq  8(%2, %0), %%mm1         \n\t"
499         "paddb     %%mm0, %%mm1         \n\t"
500         "movq      %%mm1, 8(%2, %0)     \n\t"
501         "add         $16, %0            \n\t"
502         "2:                             \n\t"
503         "cmp          %3, %0            \n\t"
504         "js           1b                \n\t"
505         : "+r"(i)
506         : "r"(src), "r"(dst), "r"((x86_reg)w - 15)
507     );
508     for ( ; i < w; i++)
509         dst[i + 0] += src[i + 0];
510 }
511
512 #if HAVE_7REGS
513 static void add_hfyu_median_prediction_cmov(uint8_t *dst, const uint8_t *top,
514                                             const uint8_t *diff, int w,
515                                             int *left, int *left_top)
516 {
517     x86_reg w2 = -w;
518     x86_reg x;
519     int l  = *left     & 0xff;
520     int tl = *left_top & 0xff;
521     int t;
522     __asm__ volatile (
523         "mov          %7, %3            \n"
524         "1:                             \n"
525         "movzbl (%3, %4), %2            \n"
526         "mov          %2, %k3           \n"
527         "sub         %b1, %b3           \n"
528         "add         %b0, %b3           \n"
529         "mov          %2, %1            \n"
530         "cmp          %0, %2            \n"
531         "cmovg        %0, %2            \n"
532         "cmovg        %1, %0            \n"
533         "cmp         %k3, %0            \n"
534         "cmovg       %k3, %0            \n"
535         "mov          %7, %3            \n"
536         "cmp          %2, %0            \n"
537         "cmovl        %2, %0            \n"
538         "add    (%6, %4), %b0           \n"
539         "mov         %b0, (%5, %4)      \n"
540         "inc          %4                \n"
541         "jl           1b                \n"
542         : "+&q"(l), "+&q"(tl), "=&r"(t), "=&q"(x), "+&r"(w2)
543         : "r"(dst + w), "r"(diff + w), "rm"(top + w)
544     );
545     *left     = l;
546     *left_top = tl;
547 }
548 #endif
549
550 static inline void transpose4x4(uint8_t *dst, uint8_t *src, x86_reg dst_stride, x86_reg src_stride){
551     __asm__ volatile( //FIXME could save 1 instruction if done as 8x4 ...
552         "movd  (%1), %%mm0              \n\t"
553         "add   %3, %1                   \n\t"
554         "movd  (%1), %%mm1              \n\t"
555         "movd  (%1,%3,1), %%mm2         \n\t"
556         "movd  (%1,%3,2), %%mm3         \n\t"
557         "punpcklbw %%mm1, %%mm0         \n\t"
558         "punpcklbw %%mm3, %%mm2         \n\t"
559         "movq %%mm0, %%mm1              \n\t"
560         "punpcklwd %%mm2, %%mm0         \n\t"
561         "punpckhwd %%mm2, %%mm1         \n\t"
562         "movd  %%mm0, (%0)              \n\t"
563         "add   %2, %0                   \n\t"
564         "punpckhdq %%mm0, %%mm0         \n\t"
565         "movd  %%mm0, (%0)              \n\t"
566         "movd  %%mm1, (%0,%2,1)         \n\t"
567         "punpckhdq %%mm1, %%mm1         \n\t"
568         "movd  %%mm1, (%0,%2,2)         \n\t"
569
570         :  "+&r" (dst),
571            "+&r" (src)
572         :  "r" (dst_stride),
573            "r" (src_stride)
574         :  "memory"
575     );
576 }
577
578 #define H263_LOOP_FILTER                        \
579     "pxor      %%mm7, %%mm7             \n\t"   \
580     "movq         %0, %%mm0             \n\t"   \
581     "movq         %0, %%mm1             \n\t"   \
582     "movq         %3, %%mm2             \n\t"   \
583     "movq         %3, %%mm3             \n\t"   \
584     "punpcklbw %%mm7, %%mm0             \n\t"   \
585     "punpckhbw %%mm7, %%mm1             \n\t"   \
586     "punpcklbw %%mm7, %%mm2             \n\t"   \
587     "punpckhbw %%mm7, %%mm3             \n\t"   \
588     "psubw     %%mm2, %%mm0             \n\t"   \
589     "psubw     %%mm3, %%mm1             \n\t"   \
590     "movq         %1, %%mm2             \n\t"   \
591     "movq         %1, %%mm3             \n\t"   \
592     "movq         %2, %%mm4             \n\t"   \
593     "movq         %2, %%mm5             \n\t"   \
594     "punpcklbw %%mm7, %%mm2             \n\t"   \
595     "punpckhbw %%mm7, %%mm3             \n\t"   \
596     "punpcklbw %%mm7, %%mm4             \n\t"   \
597     "punpckhbw %%mm7, %%mm5             \n\t"   \
598     "psubw     %%mm2, %%mm4             \n\t"   \
599     "psubw     %%mm3, %%mm5             \n\t"   \
600     "psllw        $2, %%mm4             \n\t"   \
601     "psllw        $2, %%mm5             \n\t"   \
602     "paddw     %%mm0, %%mm4             \n\t"   \
603     "paddw     %%mm1, %%mm5             \n\t"   \
604     "pxor      %%mm6, %%mm6             \n\t"   \
605     "pcmpgtw   %%mm4, %%mm6             \n\t"   \
606     "pcmpgtw   %%mm5, %%mm7             \n\t"   \
607     "pxor      %%mm6, %%mm4             \n\t"   \
608     "pxor      %%mm7, %%mm5             \n\t"   \
609     "psubw     %%mm6, %%mm4             \n\t"   \
610     "psubw     %%mm7, %%mm5             \n\t"   \
611     "psrlw        $3, %%mm4             \n\t"   \
612     "psrlw        $3, %%mm5             \n\t"   \
613     "packuswb  %%mm5, %%mm4             \n\t"   \
614     "packsswb  %%mm7, %%mm6             \n\t"   \
615     "pxor      %%mm7, %%mm7             \n\t"   \
616     "movd         %4, %%mm2             \n\t"   \
617     "punpcklbw %%mm2, %%mm2             \n\t"   \
618     "punpcklbw %%mm2, %%mm2             \n\t"   \
619     "punpcklbw %%mm2, %%mm2             \n\t"   \
620     "psubusb   %%mm4, %%mm2             \n\t"   \
621     "movq      %%mm2, %%mm3             \n\t"   \
622     "psubusb   %%mm4, %%mm3             \n\t"   \
623     "psubb     %%mm3, %%mm2             \n\t"   \
624     "movq         %1, %%mm3             \n\t"   \
625     "movq         %2, %%mm4             \n\t"   \
626     "pxor      %%mm6, %%mm3             \n\t"   \
627     "pxor      %%mm6, %%mm4             \n\t"   \
628     "paddusb   %%mm2, %%mm3             \n\t"   \
629     "psubusb   %%mm2, %%mm4             \n\t"   \
630     "pxor      %%mm6, %%mm3             \n\t"   \
631     "pxor      %%mm6, %%mm4             \n\t"   \
632     "paddusb   %%mm2, %%mm2             \n\t"   \
633     "packsswb  %%mm1, %%mm0             \n\t"   \
634     "pcmpgtb   %%mm0, %%mm7             \n\t"   \
635     "pxor      %%mm7, %%mm0             \n\t"   \
636     "psubb     %%mm7, %%mm0             \n\t"   \
637     "movq      %%mm0, %%mm1             \n\t"   \
638     "psubusb   %%mm2, %%mm0             \n\t"   \
639     "psubb     %%mm0, %%mm1             \n\t"   \
640     "pand         %5, %%mm1             \n\t"   \
641     "psrlw        $2, %%mm1             \n\t"   \
642     "pxor      %%mm7, %%mm1             \n\t"   \
643     "psubb     %%mm7, %%mm1             \n\t"   \
644     "movq         %0, %%mm5             \n\t"   \
645     "movq         %3, %%mm6             \n\t"   \
646     "psubb     %%mm1, %%mm5             \n\t"   \
647     "paddb     %%mm1, %%mm6             \n\t"
648
649 static void h263_v_loop_filter_mmx(uint8_t *src, int stride, int qscale)
650 {
651     if (CONFIG_H263_DECODER || CONFIG_H263_ENCODER) {
652         const int strength = ff_h263_loop_filter_strength[qscale];
653
654         __asm__ volatile (
655             H263_LOOP_FILTER
656
657             "movq %%mm3, %1             \n\t"
658             "movq %%mm4, %2             \n\t"
659             "movq %%mm5, %0             \n\t"
660             "movq %%mm6, %3             \n\t"
661             : "+m"(*(uint64_t*)(src - 2 * stride)),
662               "+m"(*(uint64_t*)(src - 1 * stride)),
663               "+m"(*(uint64_t*)(src + 0 * stride)),
664               "+m"(*(uint64_t*)(src + 1 * stride))
665             : "g"(2 * strength), "m"(ff_pb_FC)
666             );
667     }
668 }
669
670 static void h263_h_loop_filter_mmx(uint8_t *src, int stride, int qscale)
671 {
672     if (CONFIG_H263_DECODER || CONFIG_H263_ENCODER) {
673         const int strength = ff_h263_loop_filter_strength[qscale];
674         DECLARE_ALIGNED(8, uint64_t, temp)[4];
675         uint8_t *btemp = (uint8_t*)temp;
676
677         src -= 2;
678
679         transpose4x4(btemp,     src,              8, stride);
680         transpose4x4(btemp + 4, src + 4 * stride, 8, stride);
681         __asm__ volatile (
682             H263_LOOP_FILTER // 5 3 4 6
683
684             : "+m"(temp[0]),
685               "+m"(temp[1]),
686               "+m"(temp[2]),
687               "+m"(temp[3])
688             : "g"(2 * strength), "m"(ff_pb_FC)
689             );
690
691         __asm__ volatile (
692             "movq      %%mm5, %%mm1         \n\t"
693             "movq      %%mm4, %%mm0         \n\t"
694             "punpcklbw %%mm3, %%mm5         \n\t"
695             "punpcklbw %%mm6, %%mm4         \n\t"
696             "punpckhbw %%mm3, %%mm1         \n\t"
697             "punpckhbw %%mm6, %%mm0         \n\t"
698             "movq      %%mm5, %%mm3         \n\t"
699             "movq      %%mm1, %%mm6         \n\t"
700             "punpcklwd %%mm4, %%mm5         \n\t"
701             "punpcklwd %%mm0, %%mm1         \n\t"
702             "punpckhwd %%mm4, %%mm3         \n\t"
703             "punpckhwd %%mm0, %%mm6         \n\t"
704             "movd      %%mm5, (%0)          \n\t"
705             "punpckhdq %%mm5, %%mm5         \n\t"
706             "movd      %%mm5, (%0, %2)      \n\t"
707             "movd      %%mm3, (%0, %2, 2)   \n\t"
708             "punpckhdq %%mm3, %%mm3         \n\t"
709             "movd      %%mm3, (%0, %3)      \n\t"
710             "movd      %%mm1, (%1)          \n\t"
711             "punpckhdq %%mm1, %%mm1         \n\t"
712             "movd      %%mm1, (%1, %2)      \n\t"
713             "movd      %%mm6, (%1, %2, 2)   \n\t"
714             "punpckhdq %%mm6, %%mm6         \n\t"
715             "movd      %%mm6, (%1, %3)      \n\t"
716             :: "r"(src),
717                "r"(src + 4 * stride),
718                "r"((x86_reg)stride),
719                "r"((x86_reg)(3 * stride))
720             );
721     }
722 }
723
724 /* Draw the edges of width 'w' of an image of size width, height
725  * this MMX version can only handle w == 8 || w == 16. */
726 static void draw_edges_mmx(uint8_t *buf, int wrap, int width, int height,
727                            int w, int h, int sides)
728 {
729     uint8_t *ptr, *last_line;
730     int i;
731
732     last_line = buf + (height - 1) * wrap;
733     /* left and right */
734     ptr = buf;
735     if (w == 8) {
736         __asm__ volatile (
737             "1:                             \n\t"
738             "movd            (%0), %%mm0    \n\t"
739             "punpcklbw      %%mm0, %%mm0    \n\t"
740             "punpcklwd      %%mm0, %%mm0    \n\t"
741             "punpckldq      %%mm0, %%mm0    \n\t"
742             "movq           %%mm0, -8(%0)   \n\t"
743             "movq      -8(%0, %2), %%mm1    \n\t"
744             "punpckhbw      %%mm1, %%mm1    \n\t"
745             "punpckhwd      %%mm1, %%mm1    \n\t"
746             "punpckhdq      %%mm1, %%mm1    \n\t"
747             "movq           %%mm1, (%0, %2) \n\t"
748             "add               %1, %0       \n\t"
749             "cmp               %3, %0       \n\t"
750             "jb                1b           \n\t"
751             : "+r"(ptr)
752             : "r"((x86_reg)wrap), "r"((x86_reg)width), "r"(ptr + wrap * height)
753             );
754     } else {
755         __asm__ volatile (
756             "1:                                 \n\t"
757             "movd            (%0), %%mm0        \n\t"
758             "punpcklbw      %%mm0, %%mm0        \n\t"
759             "punpcklwd      %%mm0, %%mm0        \n\t"
760             "punpckldq      %%mm0, %%mm0        \n\t"
761             "movq           %%mm0, -8(%0)       \n\t"
762             "movq           %%mm0, -16(%0)      \n\t"
763             "movq      -8(%0, %2), %%mm1        \n\t"
764             "punpckhbw      %%mm1, %%mm1        \n\t"
765             "punpckhwd      %%mm1, %%mm1        \n\t"
766             "punpckhdq      %%mm1, %%mm1        \n\t"
767             "movq           %%mm1,  (%0, %2)    \n\t"
768             "movq           %%mm1, 8(%0, %2)    \n\t"
769             "add               %1, %0           \n\t"
770             "cmp               %3, %0           \n\t"
771             "jb                1b               \n\t"
772             : "+r"(ptr)
773             : "r"((x86_reg)wrap), "r"((x86_reg)width), "r"(ptr + wrap * height)
774             );
775     }
776
777     /* top and bottom (and hopefully also the corners) */
778     if (sides & EDGE_TOP) {
779         for (i = 0; i < h; i += 4) {
780             ptr = buf - (i + 1) * wrap - w;
781             __asm__ volatile (
782                 "1:                             \n\t"
783                 "movq (%1, %0), %%mm0           \n\t"
784                 "movq    %%mm0, (%0)            \n\t"
785                 "movq    %%mm0, (%0, %2)        \n\t"
786                 "movq    %%mm0, (%0, %2, 2)     \n\t"
787                 "movq    %%mm0, (%0, %3)        \n\t"
788                 "add        $8, %0              \n\t"
789                 "cmp        %4, %0              \n\t"
790                 "jb         1b                  \n\t"
791                 : "+r"(ptr)
792                 : "r"((x86_reg)buf - (x86_reg)ptr - w), "r"((x86_reg) -wrap),
793                   "r"((x86_reg) -wrap * 3), "r"(ptr + width + 2 * w)
794                 );
795         }
796     }
797
798     if (sides & EDGE_BOTTOM) {
799         for (i = 0; i < h; i += 4) {
800             ptr = last_line + (i + 1) * wrap - w;
801             __asm__ volatile (
802                 "1:                             \n\t"
803                 "movq (%1, %0), %%mm0           \n\t"
804                 "movq    %%mm0, (%0)            \n\t"
805                 "movq    %%mm0, (%0, %2)        \n\t"
806                 "movq    %%mm0, (%0, %2, 2)     \n\t"
807                 "movq    %%mm0, (%0, %3)        \n\t"
808                 "add        $8, %0              \n\t"
809                 "cmp        %4, %0              \n\t"
810                 "jb         1b                  \n\t"
811                 : "+r"(ptr)
812                 : "r"((x86_reg)last_line - (x86_reg)ptr - w),
813                   "r"((x86_reg)wrap), "r"((x86_reg)wrap * 3),
814                   "r"(ptr + width + 2 * w)
815                 );
816         }
817     }
818 }
819
820 #define QPEL_V_LOW(m3, m4, m5, m6, pw_20, pw_3, rnd,                      \
821                    in0, in1, in2, in7, out, OP)                           \
822     "paddw               "#m4", "#m3"   \n\t" /* x1 */                    \
823     "movq   "MANGLE(ff_pw_20)", %%mm4   \n\t" /* 20 */                    \
824     "pmullw              "#m3", %%mm4   \n\t" /* 20x1 */                  \
825     "movq               "#in7", "#m3"   \n\t" /* d */                     \
826     "movq               "#in0", %%mm5   \n\t" /* D */                     \
827     "paddw               "#m3", %%mm5   \n\t" /* x4 */                    \
828     "psubw               %%mm5, %%mm4   \n\t" /* 20x1 - x4 */             \
829     "movq               "#in1", %%mm5   \n\t" /* C */                     \
830     "movq               "#in2", %%mm6   \n\t" /* B */                     \
831     "paddw               "#m6", %%mm5   \n\t" /* x3 */                    \
832     "paddw               "#m5", %%mm6   \n\t" /* x2 */                    \
833     "paddw               %%mm6, %%mm6   \n\t" /* 2x2 */                   \
834     "psubw               %%mm6, %%mm5   \n\t" /* -2x2 + x3 */             \
835     "pmullw  "MANGLE(ff_pw_3)", %%mm5   \n\t" /* -6x2 + 3x3 */            \
836     "paddw              "#rnd", %%mm4   \n\t" /* x2 */                    \
837     "paddw               %%mm4, %%mm5   \n\t" /* 20x1 - 6x2 + 3x3 - x4 */ \
838     "psraw                  $5, %%mm5   \n\t"                             \
839     "packuswb            %%mm5, %%mm5   \n\t"                             \
840     OP(%%mm5, out, %%mm7, d)
841
842 #define QPEL_BASE(OPNAME, ROUNDER, RND, OP_MMXEXT)                        \
843 static void OPNAME ## mpeg4_qpel16_h_lowpass_mmxext(uint8_t *dst,         \
844                                                     uint8_t *src,         \
845                                                     int dstStride,        \
846                                                     int srcStride,        \
847                                                     int h)                \
848 {                                                                         \
849     uint64_t temp;                                                        \
850                                                                           \
851     __asm__ volatile (                                                    \
852         "pxor      %%mm7, %%mm7             \n\t"                         \
853         "1:                                 \n\t"                         \
854         "movq       (%0), %%mm0             \n\t" /* ABCDEFGH */          \
855         "movq      %%mm0, %%mm1             \n\t" /* ABCDEFGH */          \
856         "movq      %%mm0, %%mm2             \n\t" /* ABCDEFGH */          \
857         "punpcklbw %%mm7, %%mm0             \n\t" /* 0A0B0C0D */          \
858         "punpckhbw %%mm7, %%mm1             \n\t" /* 0E0F0G0H */          \
859         "pshufw    $0x90, %%mm0, %%mm5      \n\t" /* 0A0A0B0C */          \
860         "pshufw    $0x41, %%mm0, %%mm6      \n\t" /* 0B0A0A0B */          \
861         "movq      %%mm2, %%mm3             \n\t" /* ABCDEFGH */          \
862         "movq      %%mm2, %%mm4             \n\t" /* ABCDEFGH */          \
863         "psllq        $8, %%mm2             \n\t" /* 0ABCDEFG */          \
864         "psllq       $16, %%mm3             \n\t" /* 00ABCDEF */          \
865         "psllq       $24, %%mm4             \n\t" /* 000ABCDE */          \
866         "punpckhbw %%mm7, %%mm2             \n\t" /* 0D0E0F0G */          \
867         "punpckhbw %%mm7, %%mm3             \n\t" /* 0C0D0E0F */          \
868         "punpckhbw %%mm7, %%mm4             \n\t" /* 0B0C0D0E */          \
869         "paddw     %%mm3, %%mm5             \n\t" /* b */                 \
870         "paddw     %%mm2, %%mm6             \n\t" /* c */                 \
871         "paddw     %%mm5, %%mm5             \n\t" /* 2b */                \
872         "psubw     %%mm5, %%mm6             \n\t" /* c - 2b */            \
873         "pshufw    $0x06, %%mm0, %%mm5      \n\t" /* 0C0B0A0A */          \
874         "pmullw "MANGLE(ff_pw_3)", %%mm6    \n\t" /* 3c - 6b */           \
875         "paddw     %%mm4, %%mm0             \n\t" /* a */                 \
876         "paddw     %%mm1, %%mm5             \n\t" /* d */                 \
877         "pmullw "MANGLE(ff_pw_20)", %%mm0   \n\t" /* 20a */               \
878         "psubw     %%mm5, %%mm0             \n\t" /* 20a - d */           \
879         "paddw        %6, %%mm6             \n\t"                         \
880         "paddw     %%mm6, %%mm0             \n\t" /* 20a - 6b + 3c - d */ \
881         "psraw        $5, %%mm0             \n\t"                         \
882         "movq      %%mm0, %5                \n\t"                         \
883         /* mm1 = EFGH, mm2 = DEFG, mm3 = CDEF, mm4 = BCDE, mm7 = 0 */     \
884                                                                           \
885         "movq      5(%0), %%mm0             \n\t" /* FGHIJKLM */          \
886         "movq      %%mm0, %%mm5             \n\t" /* FGHIJKLM */          \
887         "movq      %%mm0, %%mm6             \n\t" /* FGHIJKLM */          \
888         "psrlq        $8, %%mm0             \n\t" /* GHIJKLM0 */          \
889         "psrlq       $16, %%mm5             \n\t" /* HIJKLM00 */          \
890         "punpcklbw %%mm7, %%mm0             \n\t" /* 0G0H0I0J */          \
891         "punpcklbw %%mm7, %%mm5             \n\t" /* 0H0I0J0K */          \
892         "paddw     %%mm0, %%mm2             \n\t" /* b */                 \
893         "paddw     %%mm5, %%mm3             \n\t" /* c */                 \
894         "paddw     %%mm2, %%mm2             \n\t" /* 2b */                \
895         "psubw     %%mm2, %%mm3             \n\t" /* c - 2b */            \
896         "movq      %%mm6, %%mm2             \n\t" /* FGHIJKLM */          \
897         "psrlq       $24, %%mm6             \n\t" /* IJKLM000 */          \
898         "punpcklbw %%mm7, %%mm2             \n\t" /* 0F0G0H0I */          \
899         "punpcklbw %%mm7, %%mm6             \n\t" /* 0I0J0K0L */          \
900         "pmullw "MANGLE(ff_pw_3)", %%mm3    \n\t" /* 3c - 6b */           \
901         "paddw     %%mm2, %%mm1             \n\t" /* a */                 \
902         "paddw     %%mm6, %%mm4             \n\t" /* d */                 \
903         "pmullw "MANGLE(ff_pw_20)", %%mm1   \n\t" /* 20a */               \
904         "psubw     %%mm4, %%mm3             \n\t" /* - 6b +3c - d */      \
905         "paddw        %6, %%mm1             \n\t"                         \
906         "paddw     %%mm1, %%mm3             \n\t" /* 20a - 6b +3c - d */  \
907         "psraw        $5, %%mm3             \n\t"                         \
908         "movq         %5, %%mm1             \n\t"                         \
909         "packuswb  %%mm3, %%mm1             \n\t"                         \
910         OP_MMXEXT(%%mm1, (%1), %%mm4, q)                                  \
911         /* mm0 = GHIJ, mm2 = FGHI, mm5 = HIJK, mm6 = IJKL, mm7 = 0 */     \
912                                                                           \
913         "movq      9(%0), %%mm1             \n\t" /* JKLMNOPQ */          \
914         "movq      %%mm1, %%mm4             \n\t" /* JKLMNOPQ */          \
915         "movq      %%mm1, %%mm3             \n\t" /* JKLMNOPQ */          \
916         "psrlq        $8, %%mm1             \n\t" /* KLMNOPQ0 */          \
917         "psrlq       $16, %%mm4             \n\t" /* LMNOPQ00 */          \
918         "punpcklbw %%mm7, %%mm1             \n\t" /* 0K0L0M0N */          \
919         "punpcklbw %%mm7, %%mm4             \n\t" /* 0L0M0N0O */          \
920         "paddw     %%mm1, %%mm5             \n\t" /* b */                 \
921         "paddw     %%mm4, %%mm0             \n\t" /* c */                 \
922         "paddw     %%mm5, %%mm5             \n\t" /* 2b */                \
923         "psubw     %%mm5, %%mm0             \n\t" /* c - 2b */            \
924         "movq      %%mm3, %%mm5             \n\t" /* JKLMNOPQ */          \
925         "psrlq       $24, %%mm3             \n\t" /* MNOPQ000 */          \
926         "pmullw "MANGLE(ff_pw_3)", %%mm0    \n\t" /* 3c - 6b */           \
927         "punpcklbw %%mm7, %%mm3             \n\t" /* 0M0N0O0P */          \
928         "paddw     %%mm3, %%mm2             \n\t" /* d */                 \
929         "psubw     %%mm2, %%mm0             \n\t" /* -6b + 3c - d */      \
930         "movq      %%mm5, %%mm2             \n\t" /* JKLMNOPQ */          \
931         "punpcklbw %%mm7, %%mm2             \n\t" /* 0J0K0L0M */          \
932         "punpckhbw %%mm7, %%mm5             \n\t" /* 0N0O0P0Q */          \
933         "paddw     %%mm2, %%mm6             \n\t" /* a */                 \
934         "pmullw "MANGLE(ff_pw_20)", %%mm6   \n\t" /* 20a */               \
935         "paddw        %6, %%mm0             \n\t"                         \
936         "paddw     %%mm6, %%mm0             \n\t" /* 20a - 6b + 3c - d */ \
937         "psraw        $5, %%mm0             \n\t"                         \
938         /* mm1 = KLMN, mm2 = JKLM, mm3 = MNOP, */                         \
939         /* mm4 = LMNO, mm5 = NOPQ mm7 = 0 */                              \
940                                                                           \
941         "paddw    %%mm5, %%mm3              \n\t" /* a */                 \
942         "pshufw   $0xF9, %%mm5, %%mm6       \n\t" /* 0O0P0Q0Q */          \
943         "paddw    %%mm4, %%mm6              \n\t" /* b */                 \
944         "pshufw   $0xBE, %%mm5, %%mm4       \n\t" /* 0P0Q0Q0P */          \
945         "pshufw   $0x6F, %%mm5, %%mm5       \n\t" /* 0Q0Q0P0O */          \
946         "paddw    %%mm1, %%mm4              \n\t" /* c */                 \
947         "paddw    %%mm2, %%mm5              \n\t" /* d */                 \
948         "paddw    %%mm6, %%mm6              \n\t" /* 2b */                \
949         "psubw    %%mm6, %%mm4              \n\t" /* c - 2b */            \
950         "pmullw "MANGLE(ff_pw_20)", %%mm3   \n\t" /* 20a */               \
951         "pmullw  "MANGLE(ff_pw_3)", %%mm4   \n\t" /* 3c - 6b */           \
952         "psubw    %%mm5, %%mm3              \n\t" /* -6b + 3c - d */      \
953         "paddw       %6, %%mm4              \n\t"                         \
954         "paddw    %%mm3, %%mm4              \n\t" /* 20a - 6b + 3c - d */ \
955         "psraw       $5, %%mm4              \n\t"                         \
956         "packuswb %%mm4, %%mm0              \n\t"                         \
957         OP_MMXEXT(%%mm0, 8(%1), %%mm4, q)                                 \
958                                                                           \
959         "add         %3, %0                 \n\t"                         \
960         "add         %4, %1                 \n\t"                         \
961         "decl        %2                     \n\t"                         \
962         "jnz         1b                     \n\t"                         \
963         : "+a"(src), "+c"(dst), "+D"(h)                                   \
964         : "d"((x86_reg)srcStride), "S"((x86_reg)dstStride),               \
965           /* "m"(ff_pw_20), "m"(ff_pw_3), */ "m"(temp), "m"(ROUNDER)      \
966         : "memory"                                                        \
967         );                                                                \
968 }                                                                         \
969                                                                           \
970 static void OPNAME ## mpeg4_qpel8_h_lowpass_mmxext(uint8_t *dst,          \
971                                                    uint8_t *src,          \
972                                                    int dstStride,         \
973                                                    int srcStride,         \
974                                                    int h)                 \
975 {                                                                         \
976     __asm__ volatile (                                                    \
977         "pxor      %%mm7, %%mm7             \n\t"                         \
978         "1:                                 \n\t"                         \
979         "movq       (%0), %%mm0             \n\t" /* ABCDEFGH */          \
980         "movq      %%mm0, %%mm1             \n\t" /* ABCDEFGH */          \
981         "movq      %%mm0, %%mm2             \n\t" /* ABCDEFGH */          \
982         "punpcklbw %%mm7, %%mm0             \n\t" /* 0A0B0C0D */          \
983         "punpckhbw %%mm7, %%mm1             \n\t" /* 0E0F0G0H */          \
984         "pshufw    $0x90, %%mm0, %%mm5      \n\t" /* 0A0A0B0C */          \
985         "pshufw    $0x41, %%mm0, %%mm6      \n\t" /* 0B0A0A0B */          \
986         "movq      %%mm2, %%mm3             \n\t" /* ABCDEFGH */          \
987         "movq      %%mm2, %%mm4             \n\t" /* ABCDEFGH */          \
988         "psllq        $8, %%mm2             \n\t" /* 0ABCDEFG */          \
989         "psllq       $16, %%mm3             \n\t" /* 00ABCDEF */          \
990         "psllq       $24, %%mm4             \n\t" /* 000ABCDE */          \
991         "punpckhbw %%mm7, %%mm2             \n\t" /* 0D0E0F0G */          \
992         "punpckhbw %%mm7, %%mm3             \n\t" /* 0C0D0E0F */          \
993         "punpckhbw %%mm7, %%mm4             \n\t" /* 0B0C0D0E */          \
994         "paddw     %%mm3, %%mm5             \n\t" /* b */                 \
995         "paddw     %%mm2, %%mm6             \n\t" /* c */                 \
996         "paddw     %%mm5, %%mm5             \n\t" /* 2b */                \
997         "psubw     %%mm5, %%mm6             \n\t" /* c - 2b */            \
998         "pshufw    $0x06, %%mm0, %%mm5      \n\t" /* 0C0B0A0A */          \
999         "pmullw "MANGLE(ff_pw_3)", %%mm6    \n\t" /* 3c - 6b */           \
1000         "paddw     %%mm4, %%mm0             \n\t" /* a */                 \
1001         "paddw     %%mm1, %%mm5             \n\t" /* d */                 \
1002         "pmullw "MANGLE(ff_pw_20)", %%mm0   \n\t" /* 20a */               \
1003         "psubw     %%mm5, %%mm0             \n\t" /* 20a - d */           \
1004         "paddw        %5, %%mm6             \n\t"                         \
1005         "paddw     %%mm6, %%mm0             \n\t" /* 20a - 6b + 3c - d */ \
1006         "psraw        $5, %%mm0             \n\t"                         \
1007         /* mm1 = EFGH, mm2 = DEFG, mm3 = CDEF, mm4 = BCDE, mm7 = 0 */     \
1008                                                                           \
1009         "movd      5(%0), %%mm5             \n\t" /* FGHI */              \
1010         "punpcklbw %%mm7, %%mm5             \n\t" /* 0F0G0H0I */          \
1011         "pshufw    $0xF9, %%mm5, %%mm6      \n\t" /* 0G0H0I0I */          \
1012         "paddw     %%mm5, %%mm1             \n\t" /* a */                 \
1013         "paddw     %%mm6, %%mm2             \n\t" /* b */                 \
1014         "pshufw    $0xBE, %%mm5, %%mm6      \n\t" /* 0H0I0I0H */          \
1015         "pshufw    $0x6F, %%mm5, %%mm5      \n\t" /* 0I0I0H0G */          \
1016         "paddw     %%mm6, %%mm3             \n\t" /* c */                 \
1017         "paddw     %%mm5, %%mm4             \n\t" /* d */                 \
1018         "paddw     %%mm2, %%mm2             \n\t" /* 2b */                \
1019         "psubw     %%mm2, %%mm3             \n\t" /* c - 2b */            \
1020         "pmullw "MANGLE(ff_pw_20)", %%mm1   \n\t" /* 20a */               \
1021         "pmullw  "MANGLE(ff_pw_3)", %%mm3   \n\t" /* 3c - 6b */           \
1022         "psubw     %%mm4, %%mm3             \n\t" /* -6b + 3c - d */      \
1023         "paddw        %5, %%mm1             \n\t"                         \
1024         "paddw     %%mm1, %%mm3             \n\t" /* 20a - 6b + 3c - d */ \
1025         "psraw        $5, %%mm3             \n\t"                         \
1026         "packuswb  %%mm3, %%mm0             \n\t"                         \
1027         OP_MMXEXT(%%mm0, (%1), %%mm4, q)                                  \
1028                                                                           \
1029         "add          %3, %0                \n\t"                         \
1030         "add          %4, %1                \n\t"                         \
1031         "decl         %2                    \n\t"                         \
1032         "jnz          1b                    \n\t"                         \
1033         : "+a"(src), "+c"(dst), "+d"(h)                                   \
1034         : "S"((x86_reg)srcStride), "D"((x86_reg)dstStride),               \
1035           /* "m"(ff_pw_20), "m"(ff_pw_3), */ "m"(ROUNDER)                 \
1036         : "memory"                                                        \
1037         );                                                                \
1038 }
1039
1040 #define QPEL_OP(OPNAME, ROUNDER, RND, OP, MMX)                          \
1041 static void OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(uint8_t *dst,      \
1042                                                      uint8_t *src,      \
1043                                                      int dstStride,     \
1044                                                      int srcStride)     \
1045 {                                                                       \
1046     uint64_t temp[17 * 4];                                              \
1047     uint64_t *temp_ptr = temp;                                          \
1048     int count = 17;                                                     \
1049                                                                         \
1050     /* FIXME unroll */                                                  \
1051     __asm__ volatile (                                                  \
1052         "pxor      %%mm7, %%mm7             \n\t"                       \
1053         "1:                                 \n\t"                       \
1054         "movq       (%0), %%mm0             \n\t"                       \
1055         "movq       (%0), %%mm1             \n\t"                       \
1056         "movq      8(%0), %%mm2             \n\t"                       \
1057         "movq      8(%0), %%mm3             \n\t"                       \
1058         "punpcklbw %%mm7, %%mm0             \n\t"                       \
1059         "punpckhbw %%mm7, %%mm1             \n\t"                       \
1060         "punpcklbw %%mm7, %%mm2             \n\t"                       \
1061         "punpckhbw %%mm7, %%mm3             \n\t"                       \
1062         "movq      %%mm0, (%1)              \n\t"                       \
1063         "movq      %%mm1, 17 * 8(%1)        \n\t"                       \
1064         "movq      %%mm2, 2 * 17 * 8(%1)    \n\t"                       \
1065         "movq      %%mm3, 3 * 17 * 8(%1)    \n\t"                       \
1066         "add          $8, %1                \n\t"                       \
1067         "add          %3, %0                \n\t"                       \
1068         "decl         %2                    \n\t"                       \
1069         "jnz          1b                    \n\t"                       \
1070         : "+r"(src), "+r"(temp_ptr), "+r"(count)                        \
1071         : "r"((x86_reg)srcStride)                                       \
1072         : "memory"                                                      \
1073         );                                                              \
1074                                                                         \
1075     temp_ptr = temp;                                                    \
1076     count    = 4;                                                       \
1077                                                                         \
1078     /* FIXME reorder for speed */                                       \
1079     __asm__ volatile (                                                  \
1080         /* "pxor  %%mm7, %%mm7            \n\t" */                      \
1081         "1:                             \n\t"                           \
1082         "movq    (%0), %%mm0            \n\t"                           \
1083         "movq   8(%0), %%mm1            \n\t"                           \
1084         "movq  16(%0), %%mm2            \n\t"                           \
1085         "movq  24(%0), %%mm3            \n\t"                           \
1086         QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 16(%0),   8(%0),    (%0),  32(%0), (%1),     OP) \
1087         QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5,  8(%0),    (%0),    (%0),  40(%0), (%1, %3), OP) \
1088         "add       %4, %1               \n\t"                           \
1089         QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5,   (%0),    (%0),   8(%0),  48(%0), (%1),     OP) \
1090                                                                         \
1091         QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5,   (%0),   8(%0),  16(%0),  56(%0), (%1, %3), OP) \
1092         "add       %4, %1               \n\t"                           \
1093         QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5,  8(%0),  16(%0),  24(%0),  64(%0), (%1),     OP) \
1094         QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 16(%0),  24(%0),  32(%0),  72(%0), (%1, %3), OP) \
1095         "add       %4, %1               \n\t"                           \
1096         QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 24(%0),  32(%0),  40(%0),  80(%0), (%1),     OP) \
1097         QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 32(%0),  40(%0),  48(%0),  88(%0), (%1, %3), OP) \
1098         "add       %4, %1               \n\t"                           \
1099         QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 40(%0),  48(%0),  56(%0),  96(%0), (%1),     OP) \
1100         QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 48(%0),  56(%0),  64(%0), 104(%0), (%1, %3), OP) \
1101         "add       %4, %1               \n\t"                           \
1102         QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 56(%0),  64(%0),  72(%0), 112(%0), (%1),     OP) \
1103         QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 64(%0),  72(%0),  80(%0), 120(%0), (%1, %3), OP) \
1104         "add       %4, %1               \n\t"                           \
1105         QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 72(%0),  80(%0),  88(%0), 128(%0), (%1),     OP) \
1106                                                                         \
1107         QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 80(%0),  88(%0),  96(%0), 128(%0), (%1, %3), OP) \
1108         "add       %4, %1               \n\t"                           \
1109         QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 88(%0),  96(%0), 104(%0), 120(%0), (%1),     OP) \
1110         QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 96(%0), 104(%0), 112(%0), 112(%0), (%1, %3), OP) \
1111                                                                         \
1112         "add     $136, %0               \n\t"                           \
1113         "add       %6, %1               \n\t"                           \
1114         "decl      %2                   \n\t"                           \
1115         "jnz       1b                   \n\t"                           \
1116                                                                         \
1117         : "+r"(temp_ptr), "+r"(dst), "+g"(count)                        \
1118         : "r"((x86_reg)dstStride), "r"(2 * (x86_reg)dstStride),         \
1119           /* "m"(ff_pw_20), "m"(ff_pw_3), */ "m"(ROUNDER),              \
1120           "g"(4 - 14 * (x86_reg)dstStride)                              \
1121         : "memory"                                                      \
1122         );                                                              \
1123 }                                                                       \
1124                                                                         \
1125 static void OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(uint8_t *dst,       \
1126                                                     uint8_t *src,       \
1127                                                     int dstStride,      \
1128                                                     int srcStride)      \
1129 {                                                                       \
1130     uint64_t temp[9 * 2];                                               \
1131     uint64_t *temp_ptr = temp;                                          \
1132     int count = 9;                                                      \
1133                                                                         \
1134     /* FIXME unroll */                                                  \
1135     __asm__ volatile (                                                  \
1136         "pxor      %%mm7, %%mm7         \n\t"                           \
1137         "1:                             \n\t"                           \
1138         "movq       (%0), %%mm0         \n\t"                           \
1139         "movq       (%0), %%mm1         \n\t"                           \
1140         "punpcklbw %%mm7, %%mm0         \n\t"                           \
1141         "punpckhbw %%mm7, %%mm1         \n\t"                           \
1142         "movq      %%mm0, (%1)          \n\t"                           \
1143         "movq      %%mm1, 9*8(%1)       \n\t"                           \
1144         "add          $8, %1            \n\t"                           \
1145         "add          %3, %0            \n\t"                           \
1146         "decl         %2                \n\t"                           \
1147         "jnz          1b                \n\t"                           \
1148         : "+r"(src), "+r"(temp_ptr), "+r"(count)                        \
1149         : "r"((x86_reg)srcStride)                                       \
1150         : "memory"                                                      \
1151         );                                                              \
1152                                                                         \
1153     temp_ptr = temp;                                                    \
1154     count    = 2;                                                       \
1155                                                                         \
1156     /* FIXME reorder for speed */                                       \
1157     __asm__ volatile (                                                  \
1158         /* "pxor  %%mm7, %%mm7            \n\t" */                      \
1159         "1:                             \n\t"                           \
1160         "movq    (%0), %%mm0            \n\t"                           \
1161         "movq   8(%0), %%mm1            \n\t"                           \
1162         "movq  16(%0), %%mm2            \n\t"                           \
1163         "movq  24(%0), %%mm3            \n\t"                           \
1164         QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 16(%0),  8(%0),   (%0), 32(%0), (%1), OP)     \
1165         QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5,  8(%0),   (%0),   (%0), 40(%0), (%1, %3), OP) \
1166         "add       %4, %1               \n\t"                           \
1167         QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5,   (%0),   (%0),  8(%0), 48(%0), (%1), OP)     \
1168                                                                         \
1169         QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5,   (%0),  8(%0), 16(%0), 56(%0), (%1, %3), OP) \
1170         "add       %4, %1               \n\t"                           \
1171         QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5,  8(%0), 16(%0), 24(%0), 64(%0), (%1), OP)     \
1172                                                                         \
1173         QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 16(%0), 24(%0), 32(%0), 64(%0), (%1, %3), OP) \
1174         "add       %4, %1               \n\t"                           \
1175         QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 24(%0), 32(%0), 40(%0), 56(%0), (%1), OP)     \
1176         QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 32(%0), 40(%0), 48(%0), 48(%0), (%1, %3), OP) \
1177                                                                         \
1178         "add      $72, %0               \n\t"                           \
1179         "add       %6, %1               \n\t"                           \
1180         "decl      %2                   \n\t"                           \
1181         "jnz       1b                   \n\t"                           \
1182                                                                         \
1183         : "+r"(temp_ptr), "+r"(dst), "+g"(count)                        \
1184         : "r"((x86_reg)dstStride), "r"(2 * (x86_reg)dstStride),         \
1185           /* "m"(ff_pw_20), "m"(ff_pw_3), */ "m"(ROUNDER),              \
1186           "g"(4 - 6 * (x86_reg)dstStride)                               \
1187         : "memory"                                                      \
1188         );                                                              \
1189 }                                                                       \
1190                                                                         \
1191 static void OPNAME ## qpel8_mc00_ ## MMX (uint8_t *dst, uint8_t *src,   \
1192                                           int stride)                   \
1193 {                                                                       \
1194     OPNAME ## pixels8_ ## MMX(dst, src, stride, 8);                     \
1195 }                                                                       \
1196                                                                         \
1197 static void OPNAME ## qpel8_mc10_ ## MMX(uint8_t *dst, uint8_t *src,    \
1198                                          int stride)                    \
1199 {                                                                       \
1200     uint64_t temp[8];                                                   \
1201     uint8_t * const half = (uint8_t*)temp;                              \
1202     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(half, src, 8,           \
1203                                                 stride, 8);             \
1204     OPNAME ## pixels8_l2_ ## MMX(dst, src, half, stride, stride, 8);    \
1205 }                                                                       \
1206                                                                         \
1207 static void OPNAME ## qpel8_mc20_ ## MMX(uint8_t *dst, uint8_t *src,    \
1208                                          int stride)                    \
1209 {                                                                       \
1210     OPNAME ## mpeg4_qpel8_h_lowpass_ ## MMX(dst, src, stride,           \
1211                                             stride, 8);                 \
1212 }                                                                       \
1213                                                                         \
1214 static void OPNAME ## qpel8_mc30_ ## MMX(uint8_t *dst, uint8_t *src,    \
1215                                          int stride)                    \
1216 {                                                                       \
1217     uint64_t temp[8];                                                   \
1218     uint8_t * const half = (uint8_t*)temp;                              \
1219     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(half, src, 8,           \
1220                                                 stride, 8);             \
1221     OPNAME ## pixels8_l2_ ## MMX(dst, src + 1, half, stride,            \
1222                                  stride, 8);                            \
1223 }                                                                       \
1224                                                                         \
1225 static void OPNAME ## qpel8_mc01_ ## MMX(uint8_t *dst, uint8_t *src,    \
1226                                          int stride)                    \
1227 {                                                                       \
1228     uint64_t temp[8];                                                   \
1229     uint8_t * const half = (uint8_t*)temp;                              \
1230     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(half, src, 8, stride);  \
1231     OPNAME ## pixels8_l2_ ## MMX(dst, src, half, stride, stride, 8);    \
1232 }                                                                       \
1233                                                                         \
1234 static void OPNAME ## qpel8_mc02_ ## MMX(uint8_t *dst, uint8_t *src,    \
1235                                          int stride)                    \
1236 {                                                                       \
1237     OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, src, stride, stride);  \
1238 }                                                                       \
1239                                                                         \
1240 static void OPNAME ## qpel8_mc03_ ## MMX(uint8_t *dst, uint8_t *src,    \
1241                                          int stride)                    \
1242 {                                                                       \
1243     uint64_t temp[8];                                                   \
1244     uint8_t * const half = (uint8_t*)temp;                              \
1245     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(half, src, 8, stride);  \
1246     OPNAME ## pixels8_l2_ ## MMX(dst, src + stride, half, stride,       \
1247                                  stride, 8);                            \
1248 }                                                                       \
1249                                                                         \
1250 static void OPNAME ## qpel8_mc11_ ## MMX(uint8_t *dst, uint8_t *src,    \
1251                                          int stride)                    \
1252 {                                                                       \
1253     uint64_t half[8 + 9];                                               \
1254     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
1255     uint8_t * const halfHV = ((uint8_t*)half);                          \
1256     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,          \
1257                                                 stride, 9);             \
1258     put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8, stride, 9);  \
1259     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);   \
1260     OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV, stride, 8, 8);     \
1261 }                                                                       \
1262                                                                         \
1263 static void OPNAME ## qpel8_mc31_ ## MMX(uint8_t *dst, uint8_t *src,    \
1264                                          int stride)                    \
1265 {                                                                       \
1266     uint64_t half[8 + 9];                                               \
1267     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
1268     uint8_t * const halfHV = ((uint8_t*)half);                          \
1269     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,          \
1270                                                 stride, 9);             \
1271     put ## RND ## pixels8_l2_ ## MMX(halfH, src + 1, halfH, 8,          \
1272                                      stride, 9);                        \
1273     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);   \
1274     OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV, stride, 8, 8);     \
1275 }                                                                       \
1276                                                                         \
1277 static void OPNAME ## qpel8_mc13_ ## MMX(uint8_t *dst, uint8_t *src,    \
1278                                          int stride)                    \
1279 {                                                                       \
1280     uint64_t half[8 + 9];                                               \
1281     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
1282     uint8_t * const halfHV = ((uint8_t*)half);                          \
1283     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,          \
1284                                                 stride, 9);             \
1285     put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8, stride, 9);  \
1286     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);   \
1287     OPNAME ## pixels8_l2_ ## MMX(dst, halfH + 8, halfHV, stride, 8, 8); \
1288 }                                                                       \
1289                                                                         \
1290 static void OPNAME ## qpel8_mc33_ ## MMX(uint8_t *dst, uint8_t *src,    \
1291                                          int stride)                    \
1292 {                                                                       \
1293     uint64_t half[8 + 9];                                               \
1294     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
1295     uint8_t * const halfHV = ((uint8_t*)half);                          \
1296     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,          \
1297                                                 stride, 9);             \
1298     put ## RND ## pixels8_l2_ ## MMX(halfH, src + 1, halfH, 8,          \
1299                                      stride, 9);                        \
1300     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);   \
1301     OPNAME ## pixels8_l2_ ## MMX(dst, halfH + 8, halfHV, stride, 8, 8); \
1302 }                                                                       \
1303                                                                         \
1304 static void OPNAME ## qpel8_mc21_ ## MMX(uint8_t *dst, uint8_t *src,    \
1305                                          int stride)                    \
1306 {                                                                       \
1307     uint64_t half[8 + 9];                                               \
1308     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
1309     uint8_t * const halfHV = ((uint8_t*)half);                          \
1310     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,          \
1311                                                 stride, 9);             \
1312     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);   \
1313     OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV, stride, 8, 8);     \
1314 }                                                                       \
1315                                                                         \
1316 static void OPNAME ## qpel8_mc23_ ## MMX(uint8_t *dst, uint8_t *src,    \
1317                                          int stride)                    \
1318 {                                                                       \
1319     uint64_t half[8 + 9];                                               \
1320     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
1321     uint8_t * const halfHV = ((uint8_t*)half);                          \
1322     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,          \
1323                                                 stride, 9);             \
1324     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);   \
1325     OPNAME ## pixels8_l2_ ## MMX(dst, halfH + 8, halfHV, stride, 8, 8); \
1326 }                                                                       \
1327                                                                         \
1328 static void OPNAME ## qpel8_mc12_ ## MMX(uint8_t *dst, uint8_t *src,    \
1329                                          int stride)                    \
1330 {                                                                       \
1331     uint64_t half[8 + 9];                                               \
1332     uint8_t * const halfH = ((uint8_t*)half);                           \
1333     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,          \
1334                                                 stride, 9);             \
1335     put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8, stride, 9);  \
1336     OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH, stride, 8);     \
1337 }                                                                       \
1338                                                                         \
1339 static void OPNAME ## qpel8_mc32_ ## MMX(uint8_t *dst, uint8_t *src,    \
1340                                          int stride)                    \
1341 {                                                                       \
1342     uint64_t half[8 + 9];                                               \
1343     uint8_t * const halfH = ((uint8_t*)half);                           \
1344     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,          \
1345                                                 stride, 9);             \
1346     put ## RND ## pixels8_l2_ ## MMX(halfH, src + 1, halfH, 8,          \
1347                                      stride, 9);                        \
1348     OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH, stride, 8);     \
1349 }                                                                       \
1350                                                                         \
1351 static void OPNAME ## qpel8_mc22_ ## MMX(uint8_t *dst, uint8_t *src,    \
1352                                          int stride)                    \
1353 {                                                                       \
1354     uint64_t half[9];                                                   \
1355     uint8_t * const halfH = ((uint8_t*)half);                           \
1356     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,          \
1357                                                 stride, 9);             \
1358     OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH, stride, 8);     \
1359 }                                                                       \
1360                                                                         \
1361 static void OPNAME ## qpel16_mc00_ ## MMX (uint8_t *dst, uint8_t *src,  \
1362                                            int stride)                  \
1363 {                                                                       \
1364     OPNAME ## pixels16_ ## MMX(dst, src, stride, 16);                   \
1365 }                                                                       \
1366                                                                         \
1367 static void OPNAME ## qpel16_mc10_ ## MMX(uint8_t *dst, uint8_t *src,   \
1368                                           int stride)                   \
1369 {                                                                       \
1370     uint64_t temp[32];                                                  \
1371     uint8_t * const half = (uint8_t*)temp;                              \
1372     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(half, src, 16,         \
1373                                                  stride, 16);           \
1374     OPNAME ## pixels16_l2_ ## MMX(dst, src, half, stride, stride, 16);  \
1375 }                                                                       \
1376                                                                         \
1377 static void OPNAME ## qpel16_mc20_ ## MMX(uint8_t *dst, uint8_t *src,   \
1378                                           int stride)                   \
1379 {                                                                       \
1380     OPNAME ## mpeg4_qpel16_h_lowpass_ ## MMX(dst, src,                  \
1381                                              stride, stride, 16);       \
1382 }                                                                       \
1383                                                                         \
1384 static void OPNAME ## qpel16_mc30_ ## MMX(uint8_t *dst, uint8_t *src,   \
1385                                           int stride)                   \
1386 {                                                                       \
1387     uint64_t temp[32];                                                  \
1388     uint8_t * const half = (uint8_t*)temp;                              \
1389     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(half, src, 16,         \
1390                                                  stride, 16);           \
1391     OPNAME ## pixels16_l2_ ## MMX(dst, src + 1, half,                   \
1392                                   stride, stride, 16);                  \
1393 }                                                                       \
1394                                                                         \
1395 static void OPNAME ## qpel16_mc01_ ## MMX(uint8_t *dst, uint8_t *src,   \
1396                                           int stride)                   \
1397 {                                                                       \
1398     uint64_t temp[32];                                                  \
1399     uint8_t * const half = (uint8_t*)temp;                              \
1400     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(half, src, 16,         \
1401                                                  stride);               \
1402     OPNAME ## pixels16_l2_ ## MMX(dst, src, half, stride, stride, 16);  \
1403 }                                                                       \
1404                                                                         \
1405 static void OPNAME ## qpel16_mc02_ ## MMX(uint8_t *dst, uint8_t *src,   \
1406                                           int stride)                   \
1407 {                                                                       \
1408     OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, src, stride, stride); \
1409 }                                                                       \
1410                                                                         \
1411 static void OPNAME ## qpel16_mc03_ ## MMX(uint8_t *dst, uint8_t *src,   \
1412                                           int stride)                   \
1413 {                                                                       \
1414     uint64_t temp[32];                                                  \
1415     uint8_t * const half = (uint8_t*)temp;                              \
1416     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(half, src, 16,         \
1417                                                  stride);               \
1418     OPNAME ## pixels16_l2_ ## MMX(dst, src+stride, half,                \
1419                                   stride, stride, 16);                  \
1420 }                                                                       \
1421                                                                         \
1422 static void OPNAME ## qpel16_mc11_ ## MMX(uint8_t *dst, uint8_t *src,   \
1423                                           int stride)                   \
1424 {                                                                       \
1425     uint64_t half[16 * 2 + 17 * 2];                                     \
1426     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1427     uint8_t * const halfHV = ((uint8_t*)half);                          \
1428     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,        \
1429                                                  stride, 17);           \
1430     put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16,            \
1431                                       stride, 17);                      \
1432     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,         \
1433                                                  16, 16);               \
1434     OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV, stride, 16, 16);  \
1435 }                                                                       \
1436                                                                         \
1437 static void OPNAME ## qpel16_mc31_ ## MMX(uint8_t *dst, uint8_t *src,   \
1438                                           int stride)                   \
1439 {                                                                       \
1440     uint64_t half[16 * 2 + 17 * 2];                                     \
1441     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1442     uint8_t * const halfHV = ((uint8_t*)half);                          \
1443     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,        \
1444                                                  stride, 17);           \
1445     put ## RND ## pixels16_l2_ ## MMX(halfH, src + 1, halfH, 16,        \
1446                                       stride, 17);                      \
1447     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,         \
1448                                                  16, 16);               \
1449     OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV, stride, 16, 16);  \
1450 }                                                                       \
1451                                                                         \
1452 static void OPNAME ## qpel16_mc13_ ## MMX(uint8_t *dst, uint8_t *src,   \
1453                                           int stride)                   \
1454 {                                                                       \
1455     uint64_t half[16 * 2 + 17 * 2];                                     \
1456     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1457     uint8_t * const halfHV = ((uint8_t*)half);                          \
1458     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,        \
1459                                                  stride, 17);           \
1460     put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16,            \
1461                                       stride, 17);                      \
1462     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,         \
1463                                                  16, 16);               \
1464     OPNAME ## pixels16_l2_ ## MMX(dst, halfH + 16, halfHV, stride,      \
1465                                   16, 16);                              \
1466 }                                                                       \
1467                                                                         \
1468 static void OPNAME ## qpel16_mc33_ ## MMX(uint8_t *dst, uint8_t *src,   \
1469                                           int stride)                   \
1470 {                                                                       \
1471     uint64_t half[16 * 2 + 17 * 2];                                     \
1472     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1473     uint8_t * const halfHV = ((uint8_t*)half);                          \
1474     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,        \
1475                                                  stride, 17);           \
1476     put ## RND ## pixels16_l2_ ## MMX(halfH, src + 1, halfH, 16,        \
1477                                       stride, 17);                      \
1478     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,         \
1479                                                  16, 16);               \
1480     OPNAME ## pixels16_l2_ ## MMX(dst, halfH + 16, halfHV, stride,      \
1481                                   16, 16);                              \
1482 }                                                                       \
1483                                                                         \
1484 static void OPNAME ## qpel16_mc21_ ## MMX(uint8_t *dst, uint8_t *src,   \
1485                                           int stride)                   \
1486 {                                                                       \
1487     uint64_t half[16 * 2 + 17 * 2];                                     \
1488     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1489     uint8_t * const halfHV = ((uint8_t*)half);                          \
1490     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,        \
1491                                                  stride, 17);           \
1492     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,         \
1493                                                  16, 16);               \
1494     OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV, stride, 16, 16);  \
1495 }                                                                       \
1496                                                                         \
1497 static void OPNAME ## qpel16_mc23_ ## MMX(uint8_t *dst, uint8_t *src,   \
1498                                           int stride)                   \
1499 {                                                                       \
1500     uint64_t half[16 * 2 + 17 * 2];                                     \
1501     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1502     uint8_t * const halfHV = ((uint8_t*)half);                          \
1503     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,        \
1504                                                  stride, 17);           \
1505     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,         \
1506                                                  16, 16);               \
1507     OPNAME ## pixels16_l2_ ## MMX(dst, halfH + 16, halfHV, stride,      \
1508                                   16, 16);                              \
1509 }                                                                       \
1510                                                                         \
1511 static void OPNAME ## qpel16_mc12_ ## MMX(uint8_t *dst, uint8_t *src,   \
1512                                           int stride)                   \
1513 {                                                                       \
1514     uint64_t half[17 * 2];                                              \
1515     uint8_t * const halfH = ((uint8_t*)half);                           \
1516     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,        \
1517                                                  stride, 17);           \
1518     put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16,            \
1519                                       stride, 17);                      \
1520     OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH, stride, 16);   \
1521 }                                                                       \
1522                                                                         \
1523 static void OPNAME ## qpel16_mc32_ ## MMX(uint8_t *dst, uint8_t *src,   \
1524                                           int stride)                   \
1525 {                                                                       \
1526     uint64_t half[17 * 2];                                              \
1527     uint8_t * const halfH = ((uint8_t*)half);                           \
1528     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,        \
1529                                                  stride, 17);           \
1530     put ## RND ## pixels16_l2_ ## MMX(halfH, src + 1, halfH, 16,        \
1531                                       stride, 17);                      \
1532     OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH, stride, 16);   \
1533 }                                                                       \
1534                                                                         \
1535 static void OPNAME ## qpel16_mc22_ ## MMX(uint8_t *dst, uint8_t *src,   \
1536                                           int stride)                   \
1537 {                                                                       \
1538     uint64_t half[17 * 2];                                              \
1539     uint8_t * const halfH = ((uint8_t*)half);                           \
1540     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,        \
1541                                                  stride, 17);           \
1542     OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH, stride, 16);   \
1543 }
1544
1545 #define PUT_OP(a, b, temp, size)                \
1546     "mov"#size"        "#a", "#b"       \n\t"
1547
1548 #define AVG_MMXEXT_OP(a, b, temp, size)         \
1549     "mov"#size"        "#b", "#temp"    \n\t"   \
1550     "pavgb          "#temp", "#a"       \n\t"   \
1551     "mov"#size"        "#a", "#b"       \n\t"
1552
1553 QPEL_BASE(put_,        ff_pw_16, _,        PUT_OP)
1554 QPEL_BASE(avg_,        ff_pw_16, _,        AVG_MMXEXT_OP)
1555 QPEL_BASE(put_no_rnd_, ff_pw_15, _no_rnd_, PUT_OP)
1556 QPEL_OP(put_,          ff_pw_16, _,        PUT_OP,        mmxext)
1557 QPEL_OP(avg_,          ff_pw_16, _,        AVG_MMXEXT_OP, mmxext)
1558 QPEL_OP(put_no_rnd_,   ff_pw_15, _no_rnd_, PUT_OP,        mmxext)
1559
1560 void ff_put_rv40_qpel8_mc33_mmx(uint8_t *dst, uint8_t *src, int stride)
1561 {
1562   put_pixels8_xy2_mmx(dst, src, stride, 8);
1563 }
1564 void ff_put_rv40_qpel16_mc33_mmx(uint8_t *dst, uint8_t *src, int stride)
1565 {
1566   put_pixels16_xy2_mmx(dst, src, stride, 16);
1567 }
1568 void ff_avg_rv40_qpel8_mc33_mmx(uint8_t *dst, uint8_t *src, int stride)
1569 {
1570   avg_pixels8_xy2_mmx(dst, src, stride, 8);
1571 }
1572 void ff_avg_rv40_qpel16_mc33_mmx(uint8_t *dst, uint8_t *src, int stride)
1573 {
1574   avg_pixels16_xy2_mmx(dst, src, stride, 16);
1575 }
1576
1577 static void gmc_mmx(uint8_t *dst, uint8_t *src,
1578                     int stride, int h, int ox, int oy,
1579                     int dxx, int dxy, int dyx, int dyy,
1580                     int shift, int r, int width, int height)
1581 {
1582     const int w    = 8;
1583     const int ix   = ox  >> (16 + shift);
1584     const int iy   = oy  >> (16 + shift);
1585     const int oxs  = ox  >> 4;
1586     const int oys  = oy  >> 4;
1587     const int dxxs = dxx >> 4;
1588     const int dxys = dxy >> 4;
1589     const int dyxs = dyx >> 4;
1590     const int dyys = dyy >> 4;
1591     const uint16_t r4[4]   = { r, r, r, r };
1592     const uint16_t dxy4[4] = { dxys, dxys, dxys, dxys };
1593     const uint16_t dyy4[4] = { dyys, dyys, dyys, dyys };
1594     const uint64_t shift2 = 2 * shift;
1595     int x, y;
1596
1597     const int dxw = (dxx - (1 << (16 + shift))) * (w - 1);
1598     const int dyh = (dyy - (1 << (16 + shift))) * (h - 1);
1599     const int dxh = dxy * (h - 1);
1600     const int dyw = dyx * (w - 1);
1601     if ( // non-constant fullpel offset (3% of blocks)
1602         ((ox ^ (ox + dxw)) | (ox ^ (ox + dxh)) | (ox ^ (ox + dxw + dxh)) |
1603          (oy ^ (oy + dyw)) | (oy ^ (oy + dyh)) | (oy ^ (oy + dyw + dyh))) >> (16 + shift)
1604         // uses more than 16 bits of subpel mv (only at huge resolution)
1605         || (dxx | dxy | dyx | dyy) & 15 ||
1606         (unsigned)ix >= width  - w ||
1607         (unsigned)iy >= height - h) {
1608         // FIXME could still use mmx for some of the rows
1609         ff_gmc_c(dst, src, stride, h, ox, oy, dxx, dxy, dyx, dyy,
1610                  shift, r, width, height);
1611         return;
1612     }
1613
1614     src += ix + iy * stride;
1615
1616     __asm__ volatile (
1617         "movd         %0, %%mm6         \n\t"
1618         "pxor      %%mm7, %%mm7         \n\t"
1619         "punpcklwd %%mm6, %%mm6         \n\t"
1620         "punpcklwd %%mm6, %%mm6         \n\t"
1621         :: "r"(1<<shift)
1622     );
1623
1624     for (x = 0; x < w; x += 4) {
1625         uint16_t dx4[4] = { oxs - dxys + dxxs * (x + 0),
1626                             oxs - dxys + dxxs * (x + 1),
1627                             oxs - dxys + dxxs * (x + 2),
1628                             oxs - dxys + dxxs * (x + 3) };
1629         uint16_t dy4[4] = { oys - dyys + dyxs * (x + 0),
1630                             oys - dyys + dyxs * (x + 1),
1631                             oys - dyys + dyxs * (x + 2),
1632                             oys - dyys + dyxs * (x + 3) };
1633
1634         for (y = 0; y < h; y++) {
1635             __asm__ volatile (
1636                 "movq      %0, %%mm4    \n\t"
1637                 "movq      %1, %%mm5    \n\t"
1638                 "paddw     %2, %%mm4    \n\t"
1639                 "paddw     %3, %%mm5    \n\t"
1640                 "movq   %%mm4, %0       \n\t"
1641                 "movq   %%mm5, %1       \n\t"
1642                 "psrlw    $12, %%mm4    \n\t"
1643                 "psrlw    $12, %%mm5    \n\t"
1644                 : "+m"(*dx4), "+m"(*dy4)
1645                 : "m"(*dxy4), "m"(*dyy4)
1646             );
1647
1648             __asm__ volatile (
1649                 "movq      %%mm6, %%mm2 \n\t"
1650                 "movq      %%mm6, %%mm1 \n\t"
1651                 "psubw     %%mm4, %%mm2 \n\t"
1652                 "psubw     %%mm5, %%mm1 \n\t"
1653                 "movq      %%mm2, %%mm0 \n\t"
1654                 "movq      %%mm4, %%mm3 \n\t"
1655                 "pmullw    %%mm1, %%mm0 \n\t" // (s - dx) * (s - dy)
1656                 "pmullw    %%mm5, %%mm3 \n\t" // dx * dy
1657                 "pmullw    %%mm5, %%mm2 \n\t" // (s - dx) * dy
1658                 "pmullw    %%mm4, %%mm1 \n\t" // dx * (s - dy)
1659
1660                 "movd         %4, %%mm5 \n\t"
1661                 "movd         %3, %%mm4 \n\t"
1662                 "punpcklbw %%mm7, %%mm5 \n\t"
1663                 "punpcklbw %%mm7, %%mm4 \n\t"
1664                 "pmullw    %%mm5, %%mm3 \n\t" // src[1, 1] * dx * dy
1665                 "pmullw    %%mm4, %%mm2 \n\t" // src[0, 1] * (s - dx) * dy
1666
1667                 "movd         %2, %%mm5 \n\t"
1668                 "movd         %1, %%mm4 \n\t"
1669                 "punpcklbw %%mm7, %%mm5 \n\t"
1670                 "punpcklbw %%mm7, %%mm4 \n\t"
1671                 "pmullw    %%mm5, %%mm1 \n\t" // src[1, 0] * dx * (s - dy)
1672                 "pmullw    %%mm4, %%mm0 \n\t" // src[0, 0] * (s - dx) * (s - dy)
1673                 "paddw        %5, %%mm1 \n\t"
1674                 "paddw     %%mm3, %%mm2 \n\t"
1675                 "paddw     %%mm1, %%mm0 \n\t"
1676                 "paddw     %%mm2, %%mm0 \n\t"
1677
1678                 "psrlw        %6, %%mm0 \n\t"
1679                 "packuswb  %%mm0, %%mm0 \n\t"
1680                 "movd      %%mm0, %0    \n\t"
1681
1682                 : "=m"(dst[x + y * stride])
1683                 : "m"(src[0]), "m"(src[1]),
1684                   "m"(src[stride]), "m"(src[stride + 1]),
1685                   "m"(*r4), "m"(shift2)
1686             );
1687             src += stride;
1688         }
1689         src += 4 - h * stride;
1690     }
1691 }
1692 #endif /* HAVE_INLINE_ASM */
1693
1694 #include "h264_qpel.c"
1695
1696 void ff_put_h264_chroma_mc8_rnd_mmx  (uint8_t *dst, uint8_t *src,
1697                                       int stride, int h, int x, int y);
1698 void ff_avg_h264_chroma_mc8_rnd_mmxext(uint8_t *dst, uint8_t *src,
1699                                        int stride, int h, int x, int y);
1700 void ff_avg_h264_chroma_mc8_rnd_3dnow(uint8_t *dst, uint8_t *src,
1701                                       int stride, int h, int x, int y);
1702
1703 void ff_put_h264_chroma_mc4_mmx      (uint8_t *dst, uint8_t *src,
1704                                       int stride, int h, int x, int y);
1705 void ff_avg_h264_chroma_mc4_mmxext   (uint8_t *dst, uint8_t *src,
1706                                       int stride, int h, int x, int y);
1707 void ff_avg_h264_chroma_mc4_3dnow    (uint8_t *dst, uint8_t *src,
1708                                       int stride, int h, int x, int y);
1709
1710 void ff_put_h264_chroma_mc2_mmxext   (uint8_t *dst, uint8_t *src,
1711                                       int stride, int h, int x, int y);
1712 void ff_avg_h264_chroma_mc2_mmxext   (uint8_t *dst, uint8_t *src,
1713                                       int stride, int h, int x, int y);
1714
1715 void ff_put_h264_chroma_mc8_rnd_ssse3(uint8_t *dst, uint8_t *src,
1716                                       int stride, int h, int x, int y);
1717 void ff_put_h264_chroma_mc4_ssse3    (uint8_t *dst, uint8_t *src,
1718                                       int stride, int h, int x, int y);
1719
1720 void ff_avg_h264_chroma_mc8_rnd_ssse3(uint8_t *dst, uint8_t *src,
1721                                       int stride, int h, int x, int y);
1722 void ff_avg_h264_chroma_mc4_ssse3    (uint8_t *dst, uint8_t *src,
1723                                       int stride, int h, int x, int y);
1724
1725 #define CHROMA_MC(OP, NUM, DEPTH, OPT)                                  \
1726 void ff_ ## OP ## _h264_chroma_mc ## NUM ## _ ## DEPTH ## _ ## OPT      \
1727                                       (uint8_t *dst, uint8_t *src,      \
1728                                        int stride, int h, int x, int y);
1729
1730 CHROMA_MC(put, 2, 10, mmxext)
1731 CHROMA_MC(avg, 2, 10, mmxext)
1732 CHROMA_MC(put, 4, 10, mmxext)
1733 CHROMA_MC(avg, 4, 10, mmxext)
1734 CHROMA_MC(put, 8, 10, sse2)
1735 CHROMA_MC(avg, 8, 10, sse2)
1736 CHROMA_MC(put, 8, 10, avx)
1737 CHROMA_MC(avg, 8, 10, avx)
1738
1739 #if HAVE_INLINE_ASM
1740
1741 /* CAVS-specific */
1742 void ff_put_cavs_qpel8_mc00_mmxext(uint8_t *dst, uint8_t *src, int stride)
1743 {
1744     put_pixels8_mmx(dst, src, stride, 8);
1745 }
1746
1747 void ff_avg_cavs_qpel8_mc00_mmxext(uint8_t *dst, uint8_t *src, int stride)
1748 {
1749     avg_pixels8_mmx(dst, src, stride, 8);
1750 }
1751
1752 void ff_put_cavs_qpel16_mc00_mmxext(uint8_t *dst, uint8_t *src, int stride)
1753 {
1754     put_pixels16_mmx(dst, src, stride, 16);
1755 }
1756
1757 void ff_avg_cavs_qpel16_mc00_mmxext(uint8_t *dst, uint8_t *src, int stride)
1758 {
1759     avg_pixels16_mmx(dst, src, stride, 16);
1760 }
1761
1762 /* VC-1-specific */
1763 void ff_put_vc1_mspel_mc00_mmx(uint8_t *dst, const uint8_t *src,
1764                                int stride, int rnd)
1765 {
1766     put_pixels8_mmx(dst, src, stride, 8);
1767 }
1768
1769 void ff_avg_vc1_mspel_mc00_mmxext(uint8_t *dst, const uint8_t *src,
1770                                   int stride, int rnd)
1771 {
1772     avg_pixels8_mmxext(dst, src, stride, 8);
1773 }
1774
1775 static void vector_clipf_sse(float *dst, const float *src,
1776                              float min, float max, int len)
1777 {
1778     x86_reg i = (len - 16) * 4;
1779     __asm__ volatile (
1780         "movss          %3, %%xmm4      \n\t"
1781         "movss          %4, %%xmm5      \n\t"
1782         "shufps $0, %%xmm4, %%xmm4      \n\t"
1783         "shufps $0, %%xmm5, %%xmm5      \n\t"
1784         "1:                             \n\t"
1785         "movaps   (%2, %0), %%xmm0      \n\t" // 3/1 on intel
1786         "movaps 16(%2, %0), %%xmm1      \n\t"
1787         "movaps 32(%2, %0), %%xmm2      \n\t"
1788         "movaps 48(%2, %0), %%xmm3      \n\t"
1789         "maxps      %%xmm4, %%xmm0      \n\t"
1790         "maxps      %%xmm4, %%xmm1      \n\t"
1791         "maxps      %%xmm4, %%xmm2      \n\t"
1792         "maxps      %%xmm4, %%xmm3      \n\t"
1793         "minps      %%xmm5, %%xmm0      \n\t"
1794         "minps      %%xmm5, %%xmm1      \n\t"
1795         "minps      %%xmm5, %%xmm2      \n\t"
1796         "minps      %%xmm5, %%xmm3      \n\t"
1797         "movaps     %%xmm0,   (%1, %0)  \n\t"
1798         "movaps     %%xmm1, 16(%1, %0)  \n\t"
1799         "movaps     %%xmm2, 32(%1, %0)  \n\t"
1800         "movaps     %%xmm3, 48(%1, %0)  \n\t"
1801         "sub           $64, %0          \n\t"
1802         "jge            1b              \n\t"
1803         : "+&r"(i)
1804         : "r"(dst), "r"(src), "m"(min), "m"(max)
1805         : "memory"
1806     );
1807 }
1808
1809 #endif /* HAVE_INLINE_ASM */
1810
1811 int32_t ff_scalarproduct_int16_mmxext(const int16_t *v1, const int16_t *v2,
1812                                       int order);
1813 int32_t ff_scalarproduct_int16_sse2(const int16_t *v1, const int16_t *v2,
1814                                     int order);
1815 int32_t ff_scalarproduct_and_madd_int16_mmxext(int16_t *v1, const int16_t *v2,
1816                                                const int16_t *v3,
1817                                                int order, int mul);
1818 int32_t ff_scalarproduct_and_madd_int16_sse2(int16_t *v1, const int16_t *v2,
1819                                              const int16_t *v3,
1820                                              int order, int mul);
1821 int32_t ff_scalarproduct_and_madd_int16_ssse3(int16_t *v1, const int16_t *v2,
1822                                               const int16_t *v3,
1823                                               int order, int mul);
1824
1825 void ff_apply_window_int16_round_mmxext(int16_t *output, const int16_t *input,
1826                                         const int16_t *window, unsigned int len);
1827 void ff_apply_window_int16_round_sse2(int16_t *output, const int16_t *input,
1828                                       const int16_t *window, unsigned int len);
1829 void ff_apply_window_int16_mmxext(int16_t *output, const int16_t *input,
1830                                   const int16_t *window, unsigned int len);
1831 void ff_apply_window_int16_sse2(int16_t *output, const int16_t *input,
1832                                 const int16_t *window, unsigned int len);
1833 void ff_apply_window_int16_ssse3(int16_t *output, const int16_t *input,
1834                                  const int16_t *window, unsigned int len);
1835 void ff_apply_window_int16_ssse3_atom(int16_t *output, const int16_t *input,
1836                                       const int16_t *window, unsigned int len);
1837
1838 void ff_bswap32_buf_ssse3(uint32_t *dst, const uint32_t *src, int w);
1839 void ff_bswap32_buf_sse2(uint32_t *dst, const uint32_t *src, int w);
1840
1841 void ff_add_hfyu_median_prediction_mmxext(uint8_t *dst, const uint8_t *top,
1842                                           const uint8_t *diff, int w,
1843                                           int *left, int *left_top);
1844 int  ff_add_hfyu_left_prediction_ssse3(uint8_t *dst, const uint8_t *src,
1845                                        int w, int left);
1846 int  ff_add_hfyu_left_prediction_sse4(uint8_t *dst, const uint8_t *src,
1847                                       int w, int left);
1848
1849 float ff_scalarproduct_float_sse(const float *v1, const float *v2, int order);
1850
1851 void ff_vector_clip_int32_mmx     (int32_t *dst, const int32_t *src,
1852                                    int32_t min, int32_t max, unsigned int len);
1853 void ff_vector_clip_int32_sse2    (int32_t *dst, const int32_t *src,
1854                                    int32_t min, int32_t max, unsigned int len);
1855 void ff_vector_clip_int32_int_sse2(int32_t *dst, const int32_t *src,
1856                                    int32_t min, int32_t max, unsigned int len);
1857 void ff_vector_clip_int32_sse4    (int32_t *dst, const int32_t *src,
1858                                    int32_t min, int32_t max, unsigned int len);
1859
1860 #define SET_QPEL_FUNCS(PFX, IDX, SIZE, CPU, PREFIX)                          \
1861     do {                                                                     \
1862     c->PFX ## _pixels_tab[IDX][ 0] = PREFIX ## PFX ## SIZE ## _mc00_ ## CPU; \
1863     c->PFX ## _pixels_tab[IDX][ 1] = PREFIX ## PFX ## SIZE ## _mc10_ ## CPU; \
1864     c->PFX ## _pixels_tab[IDX][ 2] = PREFIX ## PFX ## SIZE ## _mc20_ ## CPU; \
1865     c->PFX ## _pixels_tab[IDX][ 3] = PREFIX ## PFX ## SIZE ## _mc30_ ## CPU; \
1866     c->PFX ## _pixels_tab[IDX][ 4] = PREFIX ## PFX ## SIZE ## _mc01_ ## CPU; \
1867     c->PFX ## _pixels_tab[IDX][ 5] = PREFIX ## PFX ## SIZE ## _mc11_ ## CPU; \
1868     c->PFX ## _pixels_tab[IDX][ 6] = PREFIX ## PFX ## SIZE ## _mc21_ ## CPU; \
1869     c->PFX ## _pixels_tab[IDX][ 7] = PREFIX ## PFX ## SIZE ## _mc31_ ## CPU; \
1870     c->PFX ## _pixels_tab[IDX][ 8] = PREFIX ## PFX ## SIZE ## _mc02_ ## CPU; \
1871     c->PFX ## _pixels_tab[IDX][ 9] = PREFIX ## PFX ## SIZE ## _mc12_ ## CPU; \
1872     c->PFX ## _pixels_tab[IDX][10] = PREFIX ## PFX ## SIZE ## _mc22_ ## CPU; \
1873     c->PFX ## _pixels_tab[IDX][11] = PREFIX ## PFX ## SIZE ## _mc32_ ## CPU; \
1874     c->PFX ## _pixels_tab[IDX][12] = PREFIX ## PFX ## SIZE ## _mc03_ ## CPU; \
1875     c->PFX ## _pixels_tab[IDX][13] = PREFIX ## PFX ## SIZE ## _mc13_ ## CPU; \
1876     c->PFX ## _pixels_tab[IDX][14] = PREFIX ## PFX ## SIZE ## _mc23_ ## CPU; \
1877     c->PFX ## _pixels_tab[IDX][15] = PREFIX ## PFX ## SIZE ## _mc33_ ## CPU; \
1878     } while (0)
1879
1880 #define SET_HPEL_FUNCS(PFX, IDX, SIZE, CPU)                                     \
1881     do {                                                                        \
1882         c->PFX ## _pixels_tab[IDX][0] = PFX ## _pixels ## SIZE ## _     ## CPU; \
1883         c->PFX ## _pixels_tab[IDX][1] = PFX ## _pixels ## SIZE ## _x2_  ## CPU; \
1884         c->PFX ## _pixels_tab[IDX][2] = PFX ## _pixels ## SIZE ## _y2_  ## CPU; \
1885         c->PFX ## _pixels_tab[IDX][3] = PFX ## _pixels ## SIZE ## _xy2_ ## CPU; \
1886     } while (0)
1887
1888 #define H264_QPEL_FUNCS(x, y, CPU)                                                            \
1889     do {                                                                                      \
1890         c->put_h264_qpel_pixels_tab[0][x + y * 4] = put_h264_qpel16_mc ## x ## y ## _ ## CPU; \
1891         c->put_h264_qpel_pixels_tab[1][x + y * 4] = put_h264_qpel8_mc  ## x ## y ## _ ## CPU; \
1892         c->avg_h264_qpel_pixels_tab[0][x + y * 4] = avg_h264_qpel16_mc ## x ## y ## _ ## CPU; \
1893         c->avg_h264_qpel_pixels_tab[1][x + y * 4] = avg_h264_qpel8_mc  ## x ## y ## _ ## CPU; \
1894     } while (0)
1895
1896 #define H264_QPEL_FUNCS_10(x, y, CPU)                                                               \
1897     do {                                                                                            \
1898         c->put_h264_qpel_pixels_tab[0][x + y * 4] = ff_put_h264_qpel16_mc ## x ## y ## _10_ ## CPU; \
1899         c->put_h264_qpel_pixels_tab[1][x + y * 4] = ff_put_h264_qpel8_mc  ## x ## y ## _10_ ## CPU; \
1900         c->avg_h264_qpel_pixels_tab[0][x + y * 4] = ff_avg_h264_qpel16_mc ## x ## y ## _10_ ## CPU; \
1901         c->avg_h264_qpel_pixels_tab[1][x + y * 4] = ff_avg_h264_qpel8_mc  ## x ## y ## _10_ ## CPU; \
1902     } while (0)
1903
1904 static void dsputil_init_mmx(DSPContext *c, AVCodecContext *avctx, int mm_flags)
1905 {
1906     const int high_bit_depth = avctx->bits_per_raw_sample > 8;
1907
1908 #if HAVE_INLINE_ASM
1909     c->put_pixels_clamped        = ff_put_pixels_clamped_mmx;
1910     c->put_signed_pixels_clamped = ff_put_signed_pixels_clamped_mmx;
1911     c->add_pixels_clamped        = ff_add_pixels_clamped_mmx;
1912
1913     if (!high_bit_depth) {
1914         c->clear_block  = clear_block_mmx;
1915         c->clear_blocks = clear_blocks_mmx;
1916         c->draw_edges   = draw_edges_mmx;
1917
1918         SET_HPEL_FUNCS(put,        0, 16, mmx);
1919         SET_HPEL_FUNCS(put_no_rnd, 0, 16, mmx);
1920         SET_HPEL_FUNCS(avg,        0, 16, mmx);
1921         SET_HPEL_FUNCS(avg_no_rnd, 0, 16, mmx);
1922         SET_HPEL_FUNCS(put,        1,  8, mmx);
1923         SET_HPEL_FUNCS(put_no_rnd, 1,  8, mmx);
1924         SET_HPEL_FUNCS(avg,        1,  8, mmx);
1925         SET_HPEL_FUNCS(avg_no_rnd, 1,  8, mmx);
1926
1927         switch (avctx->idct_algo) {
1928         case FF_IDCT_AUTO:
1929         case FF_IDCT_SIMPLEMMX:
1930             c->idct_put              = ff_simple_idct_put_mmx;
1931             c->idct_add              = ff_simple_idct_add_mmx;
1932             c->idct                  = ff_simple_idct_mmx;
1933             c->idct_permutation_type = FF_SIMPLE_IDCT_PERM;
1934             break;
1935         case FF_IDCT_XVIDMMX:
1936             c->idct_put              = ff_idct_xvid_mmx_put;
1937             c->idct_add              = ff_idct_xvid_mmx_add;
1938             c->idct                  = ff_idct_xvid_mmx;
1939             break;
1940         }
1941     }
1942
1943     c->gmc = gmc_mmx;
1944
1945     c->add_bytes = add_bytes_mmx;
1946
1947     if (CONFIG_H263_DECODER || CONFIG_H263_ENCODER) {
1948         c->h263_v_loop_filter = h263_v_loop_filter_mmx;
1949         c->h263_h_loop_filter = h263_h_loop_filter_mmx;
1950     }
1951 #endif /* HAVE_INLINE_ASM */
1952
1953 #if HAVE_YASM
1954     if (!high_bit_depth && CONFIG_H264CHROMA) {
1955         c->put_h264_chroma_pixels_tab[0] = ff_put_h264_chroma_mc8_rnd_mmx;
1956         c->put_h264_chroma_pixels_tab[1] = ff_put_h264_chroma_mc4_mmx;
1957     }
1958
1959     c->vector_clip_int32 = ff_vector_clip_int32_mmx;
1960 #endif
1961
1962 }
1963
1964 static void dsputil_init_mmxext(DSPContext *c, AVCodecContext *avctx,
1965                                 int mm_flags)
1966 {
1967     const int bit_depth      = avctx->bits_per_raw_sample;
1968     const int high_bit_depth = bit_depth > 8;
1969
1970 #if HAVE_INLINE_ASM
1971     SET_QPEL_FUNCS(avg_qpel,        0, 16, mmxext, );
1972     SET_QPEL_FUNCS(avg_qpel,        1,  8, mmxext, );
1973
1974     SET_QPEL_FUNCS(put_qpel,        0, 16, mmxext, );
1975     SET_QPEL_FUNCS(put_qpel,        1,  8, mmxext, );
1976     SET_QPEL_FUNCS(put_no_rnd_qpel, 0, 16, mmxext, );
1977     SET_QPEL_FUNCS(put_no_rnd_qpel, 1,  8, mmxext, );
1978
1979     if (!high_bit_depth) {
1980         c->put_pixels_tab[0][1] = put_pixels16_x2_mmxext;
1981         c->put_pixels_tab[0][2] = put_pixels16_y2_mmxext;
1982
1983         c->avg_pixels_tab[0][0] = avg_pixels16_mmxext;
1984         c->avg_pixels_tab[0][1] = avg_pixels16_x2_mmxext;
1985         c->avg_pixels_tab[0][2] = avg_pixels16_y2_mmxext;
1986
1987         c->put_pixels_tab[1][1] = put_pixels8_x2_mmxext;
1988         c->put_pixels_tab[1][2] = put_pixels8_y2_mmxext;
1989
1990         c->avg_pixels_tab[1][0] = avg_pixels8_mmxext;
1991         c->avg_pixels_tab[1][1] = avg_pixels8_x2_mmxext;
1992         c->avg_pixels_tab[1][2] = avg_pixels8_y2_mmxext;
1993     }
1994
1995     if (!(avctx->flags & CODEC_FLAG_BITEXACT)) {
1996         if (!high_bit_depth) {
1997             c->put_no_rnd_pixels_tab[0][1] = put_no_rnd_pixels16_x2_mmxext;
1998             c->put_no_rnd_pixels_tab[0][2] = put_no_rnd_pixels16_y2_mmxext;
1999             c->put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_mmxext;
2000             c->put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_mmxext;
2001
2002             c->avg_pixels_tab[0][3] = avg_pixels16_xy2_mmxext;
2003             c->avg_pixels_tab[1][3] = avg_pixels8_xy2_mmxext;
2004         }
2005     }
2006
2007     if (!high_bit_depth && avctx->idct_algo == FF_IDCT_XVIDMMX) {
2008         c->idct_put = ff_idct_xvid_mmxext_put;
2009         c->idct_add = ff_idct_xvid_mmxext_add;
2010         c->idct     = ff_idct_xvid_mmxext;
2011     }
2012
2013     if (CONFIG_VP3_DECODER && (avctx->codec_id == AV_CODEC_ID_VP3 ||
2014                                avctx->codec_id == AV_CODEC_ID_THEORA)) {
2015         c->put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_exact_mmxext;
2016         c->put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_exact_mmxext;
2017     }
2018 #endif /* HAVE_INLINE_ASM */
2019
2020 #if HAVE_MMXEXT_EXTERNAL
2021     if (CONFIG_H264QPEL) {
2022         if (!high_bit_depth) {
2023             SET_QPEL_FUNCS(put_h264_qpel, 0, 16, mmxext, );
2024             SET_QPEL_FUNCS(put_h264_qpel, 1,  8, mmxext, );
2025             SET_QPEL_FUNCS(put_h264_qpel, 2,  4, mmxext, );
2026             SET_QPEL_FUNCS(avg_h264_qpel, 0, 16, mmxext, );
2027             SET_QPEL_FUNCS(avg_h264_qpel, 1,  8, mmxext, );
2028             SET_QPEL_FUNCS(avg_h264_qpel, 2,  4, mmxext, );
2029         } else if (bit_depth == 10) {
2030 #if !ARCH_X86_64
2031             SET_QPEL_FUNCS(avg_h264_qpel, 0, 16, 10_mmxext, ff_);
2032             SET_QPEL_FUNCS(put_h264_qpel, 0, 16, 10_mmxext, ff_);
2033             SET_QPEL_FUNCS(put_h264_qpel, 1,  8, 10_mmxext, ff_);
2034             SET_QPEL_FUNCS(avg_h264_qpel, 1,  8, 10_mmxext, ff_);
2035 #endif
2036             SET_QPEL_FUNCS(put_h264_qpel, 2, 4,  10_mmxext, ff_);
2037             SET_QPEL_FUNCS(avg_h264_qpel, 2, 4,  10_mmxext, ff_);
2038         }
2039     }
2040
2041     if (!high_bit_depth && CONFIG_H264CHROMA) {
2042         c->avg_h264_chroma_pixels_tab[0] = ff_avg_h264_chroma_mc8_rnd_mmxext;
2043         c->avg_h264_chroma_pixels_tab[1] = ff_avg_h264_chroma_mc4_mmxext;
2044         c->avg_h264_chroma_pixels_tab[2] = ff_avg_h264_chroma_mc2_mmxext;
2045         c->put_h264_chroma_pixels_tab[2] = ff_put_h264_chroma_mc2_mmxext;
2046     }
2047     if (bit_depth == 10 && CONFIG_H264CHROMA) {
2048         c->put_h264_chroma_pixels_tab[2] = ff_put_h264_chroma_mc2_10_mmxext;
2049         c->avg_h264_chroma_pixels_tab[2] = ff_avg_h264_chroma_mc2_10_mmxext;
2050         c->put_h264_chroma_pixels_tab[1] = ff_put_h264_chroma_mc4_10_mmxext;
2051         c->avg_h264_chroma_pixels_tab[1] = ff_avg_h264_chroma_mc4_10_mmxext;
2052     }
2053
2054     /* slower than cmov version on AMD */
2055     if (!(mm_flags & AV_CPU_FLAG_3DNOW))
2056         c->add_hfyu_median_prediction = ff_add_hfyu_median_prediction_mmxext;
2057
2058     c->scalarproduct_int16          = ff_scalarproduct_int16_mmxext;
2059     c->scalarproduct_and_madd_int16 = ff_scalarproduct_and_madd_int16_mmxext;
2060
2061     if (avctx->flags & CODEC_FLAG_BITEXACT) {
2062         c->apply_window_int16 = ff_apply_window_int16_mmxext;
2063     } else {
2064         c->apply_window_int16 = ff_apply_window_int16_round_mmxext;
2065     }
2066 #endif /* HAVE_MMXEXT_EXTERNAL */
2067 }
2068
2069 static void dsputil_init_3dnow(DSPContext *c, AVCodecContext *avctx,
2070                                int mm_flags)
2071 {
2072     const int high_bit_depth = avctx->bits_per_raw_sample > 8;
2073
2074 #if HAVE_INLINE_ASM
2075     if (!high_bit_depth) {
2076         c->put_pixels_tab[0][1] = put_pixels16_x2_3dnow;
2077         c->put_pixels_tab[0][2] = put_pixels16_y2_3dnow;
2078
2079         c->avg_pixels_tab[0][0] = avg_pixels16_3dnow;
2080         c->avg_pixels_tab[0][1] = avg_pixels16_x2_3dnow;
2081         c->avg_pixels_tab[0][2] = avg_pixels16_y2_3dnow;
2082
2083         c->put_pixels_tab[1][1] = put_pixels8_x2_3dnow;
2084         c->put_pixels_tab[1][2] = put_pixels8_y2_3dnow;
2085
2086         c->avg_pixels_tab[1][0] = avg_pixels8_3dnow;
2087         c->avg_pixels_tab[1][1] = avg_pixels8_x2_3dnow;
2088         c->avg_pixels_tab[1][2] = avg_pixels8_y2_3dnow;
2089
2090         if (!(avctx->flags & CODEC_FLAG_BITEXACT)){
2091             c->put_no_rnd_pixels_tab[0][1] = put_no_rnd_pixels16_x2_3dnow;
2092             c->put_no_rnd_pixels_tab[0][2] = put_no_rnd_pixels16_y2_3dnow;
2093             c->put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_3dnow;
2094             c->put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_3dnow;
2095
2096             c->avg_pixels_tab[0][3] = avg_pixels16_xy2_3dnow;
2097             c->avg_pixels_tab[1][3] = avg_pixels8_xy2_3dnow;
2098         }
2099     }
2100
2101     if (CONFIG_VP3_DECODER && (avctx->codec_id == AV_CODEC_ID_VP3 ||
2102                                avctx->codec_id == AV_CODEC_ID_THEORA)) {
2103         c->put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_exact_3dnow;
2104         c->put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_exact_3dnow;
2105     }
2106 #endif /* HAVE_INLINE_ASM */
2107
2108 #if HAVE_YASM
2109     if (!high_bit_depth && CONFIG_H264CHROMA) {
2110         c->avg_h264_chroma_pixels_tab[0] = ff_avg_h264_chroma_mc8_rnd_3dnow;
2111         c->avg_h264_chroma_pixels_tab[1] = ff_avg_h264_chroma_mc4_3dnow;
2112     }
2113 #endif /* HAVE_YASM */
2114 }
2115
2116 static void dsputil_init_sse(DSPContext *c, AVCodecContext *avctx, int mm_flags)
2117 {
2118     const int high_bit_depth = avctx->bits_per_raw_sample > 8;
2119
2120 #if HAVE_INLINE_ASM
2121     if (!high_bit_depth) {
2122         if (!(CONFIG_MPEG_XVMC_DECODER && avctx->xvmc_acceleration > 1)) {
2123             /* XvMCCreateBlocks() may not allocate 16-byte aligned blocks */
2124             c->clear_block  = clear_block_sse;
2125             c->clear_blocks = clear_blocks_sse;
2126         }
2127     }
2128
2129     c->vector_clipf = vector_clipf_sse;
2130 #endif /* HAVE_INLINE_ASM */
2131
2132 #if HAVE_YASM
2133     c->scalarproduct_float          = ff_scalarproduct_float_sse;
2134 #endif /* HAVE_YASM */
2135 }
2136
2137 static void dsputil_init_sse2(DSPContext *c, AVCodecContext *avctx,
2138                               int mm_flags)
2139 {
2140     const int bit_depth      = avctx->bits_per_raw_sample;
2141     const int high_bit_depth = bit_depth > 8;
2142
2143 #if HAVE_SSE2_INLINE
2144     if (!high_bit_depth && avctx->idct_algo == FF_IDCT_XVIDMMX) {
2145         c->idct_put              = ff_idct_xvid_sse2_put;
2146         c->idct_add              = ff_idct_xvid_sse2_add;
2147         c->idct                  = ff_idct_xvid_sse2;
2148         c->idct_permutation_type = FF_SSE2_IDCT_PERM;
2149     }
2150 #endif /* HAVE_SSE2_INLINE */
2151
2152 #if HAVE_SSE2_EXTERNAL
2153     if (!(mm_flags & AV_CPU_FLAG_SSE2SLOW)) {
2154         // these functions are slower than mmx on AMD, but faster on Intel
2155         if (!high_bit_depth) {
2156             c->put_pixels_tab[0][0]        = ff_put_pixels16_sse2;
2157             c->put_no_rnd_pixels_tab[0][0] = ff_put_pixels16_sse2;
2158             c->avg_pixels_tab[0][0]        = ff_avg_pixels16_sse2;
2159             if (CONFIG_H264QPEL)
2160                 H264_QPEL_FUNCS(0, 0, sse2);
2161         }
2162     }
2163
2164     if (!high_bit_depth && CONFIG_H264QPEL) {
2165         H264_QPEL_FUNCS(0, 1, sse2);
2166         H264_QPEL_FUNCS(0, 2, sse2);
2167         H264_QPEL_FUNCS(0, 3, sse2);
2168         H264_QPEL_FUNCS(1, 1, sse2);
2169         H264_QPEL_FUNCS(1, 2, sse2);
2170         H264_QPEL_FUNCS(1, 3, sse2);
2171         H264_QPEL_FUNCS(2, 1, sse2);
2172         H264_QPEL_FUNCS(2, 2, sse2);
2173         H264_QPEL_FUNCS(2, 3, sse2);
2174         H264_QPEL_FUNCS(3, 1, sse2);
2175         H264_QPEL_FUNCS(3, 2, sse2);
2176         H264_QPEL_FUNCS(3, 3, sse2);
2177     }
2178
2179     if (bit_depth == 10) {
2180         if (CONFIG_H264QPEL) {
2181             SET_QPEL_FUNCS(put_h264_qpel, 0, 16, 10_sse2, ff_);
2182             SET_QPEL_FUNCS(put_h264_qpel, 1,  8, 10_sse2, ff_);
2183             SET_QPEL_FUNCS(avg_h264_qpel, 0, 16, 10_sse2, ff_);
2184             SET_QPEL_FUNCS(avg_h264_qpel, 1,  8, 10_sse2, ff_);
2185             H264_QPEL_FUNCS_10(1, 0, sse2_cache64);
2186             H264_QPEL_FUNCS_10(2, 0, sse2_cache64);
2187             H264_QPEL_FUNCS_10(3, 0, sse2_cache64);
2188         }
2189         if (CONFIG_H264CHROMA) {
2190             c->put_h264_chroma_pixels_tab[0] = ff_put_h264_chroma_mc8_10_sse2;
2191             c->avg_h264_chroma_pixels_tab[0] = ff_avg_h264_chroma_mc8_10_sse2;
2192         }
2193     }
2194
2195     c->scalarproduct_int16          = ff_scalarproduct_int16_sse2;
2196     c->scalarproduct_and_madd_int16 = ff_scalarproduct_and_madd_int16_sse2;
2197     if (mm_flags & AV_CPU_FLAG_ATOM) {
2198         c->vector_clip_int32 = ff_vector_clip_int32_int_sse2;
2199     } else {
2200         c->vector_clip_int32 = ff_vector_clip_int32_sse2;
2201     }
2202     if (avctx->flags & CODEC_FLAG_BITEXACT) {
2203         c->apply_window_int16 = ff_apply_window_int16_sse2;
2204     } else if (!(mm_flags & AV_CPU_FLAG_SSE2SLOW)) {
2205         c->apply_window_int16 = ff_apply_window_int16_round_sse2;
2206     }
2207     c->bswap_buf = ff_bswap32_buf_sse2;
2208 #endif /* HAVE_SSE2_EXTERNAL */
2209 }
2210
2211 static void dsputil_init_ssse3(DSPContext *c, AVCodecContext *avctx,
2212                                int mm_flags)
2213 {
2214 #if HAVE_SSSE3_EXTERNAL
2215     const int high_bit_depth = avctx->bits_per_raw_sample > 8;
2216     const int bit_depth      = avctx->bits_per_raw_sample;
2217
2218     if (!high_bit_depth && CONFIG_H264QPEL) {
2219         H264_QPEL_FUNCS(1, 0, ssse3);
2220         H264_QPEL_FUNCS(1, 1, ssse3);
2221         H264_QPEL_FUNCS(1, 2, ssse3);
2222         H264_QPEL_FUNCS(1, 3, ssse3);
2223         H264_QPEL_FUNCS(2, 0, ssse3);
2224         H264_QPEL_FUNCS(2, 1, ssse3);
2225         H264_QPEL_FUNCS(2, 2, ssse3);
2226         H264_QPEL_FUNCS(2, 3, ssse3);
2227         H264_QPEL_FUNCS(3, 0, ssse3);
2228         H264_QPEL_FUNCS(3, 1, ssse3);
2229         H264_QPEL_FUNCS(3, 2, ssse3);
2230         H264_QPEL_FUNCS(3, 3, ssse3);
2231     }
2232     if (bit_depth == 10 && CONFIG_H264QPEL) {
2233         H264_QPEL_FUNCS_10(1, 0, ssse3_cache64);
2234         H264_QPEL_FUNCS_10(2, 0, ssse3_cache64);
2235         H264_QPEL_FUNCS_10(3, 0, ssse3_cache64);
2236     }
2237     if (!high_bit_depth && CONFIG_H264CHROMA) {
2238         c->put_h264_chroma_pixels_tab[0] = ff_put_h264_chroma_mc8_rnd_ssse3;
2239         c->avg_h264_chroma_pixels_tab[0] = ff_avg_h264_chroma_mc8_rnd_ssse3;
2240         c->put_h264_chroma_pixels_tab[1] = ff_put_h264_chroma_mc4_ssse3;
2241         c->avg_h264_chroma_pixels_tab[1] = ff_avg_h264_chroma_mc4_ssse3;
2242     }
2243     c->add_hfyu_left_prediction = ff_add_hfyu_left_prediction_ssse3;
2244     if (mm_flags & AV_CPU_FLAG_SSE4) // not really sse4, just slow on Conroe
2245         c->add_hfyu_left_prediction = ff_add_hfyu_left_prediction_sse4;
2246
2247     if (mm_flags & AV_CPU_FLAG_ATOM)
2248         c->apply_window_int16 = ff_apply_window_int16_ssse3_atom;
2249     else
2250         c->apply_window_int16 = ff_apply_window_int16_ssse3;
2251     if (!(mm_flags & (AV_CPU_FLAG_SSE42|AV_CPU_FLAG_3DNOW))) // cachesplit
2252         c->scalarproduct_and_madd_int16 = ff_scalarproduct_and_madd_int16_ssse3;
2253     c->bswap_buf = ff_bswap32_buf_ssse3;
2254 #endif /* HAVE_SSSE3_EXTERNAL */
2255 }
2256
2257 static void dsputil_init_sse4(DSPContext *c, AVCodecContext *avctx,
2258                               int mm_flags)
2259 {
2260 #if HAVE_SSE4_EXTERNAL
2261     c->vector_clip_int32 = ff_vector_clip_int32_sse4;
2262 #endif /* HAVE_SSE4_EXTERNAL */
2263 }
2264
2265 static void dsputil_init_avx(DSPContext *c, AVCodecContext *avctx, int mm_flags)
2266 {
2267 #if HAVE_AVX_EXTERNAL
2268     const int bit_depth = avctx->bits_per_raw_sample;
2269
2270     if (bit_depth == 10) {
2271         // AVX implies !cache64.
2272         // TODO: Port cache(32|64) detection from x264.
2273         if (CONFIG_H264QPEL) {
2274             H264_QPEL_FUNCS_10(1, 0, sse2);
2275             H264_QPEL_FUNCS_10(2, 0, sse2);
2276             H264_QPEL_FUNCS_10(3, 0, sse2);
2277         }
2278
2279         if (CONFIG_H264CHROMA) {
2280             c->put_h264_chroma_pixels_tab[0] = ff_put_h264_chroma_mc8_10_avx;
2281             c->avg_h264_chroma_pixels_tab[0] = ff_avg_h264_chroma_mc8_10_avx;
2282         }
2283     }
2284 #endif /* HAVE_AVX_EXTERNAL */
2285 }
2286
2287 void ff_dsputil_init_mmx(DSPContext *c, AVCodecContext *avctx)
2288 {
2289     int mm_flags = av_get_cpu_flags();
2290
2291 #if HAVE_7REGS && HAVE_INLINE_ASM
2292     if (mm_flags & AV_CPU_FLAG_CMOV)
2293         c->add_hfyu_median_prediction = add_hfyu_median_prediction_cmov;
2294 #endif
2295
2296     if (mm_flags & AV_CPU_FLAG_MMX)
2297         dsputil_init_mmx(c, avctx, mm_flags);
2298
2299     if (mm_flags & AV_CPU_FLAG_MMXEXT)
2300         dsputil_init_mmxext(c, avctx, mm_flags);
2301
2302     if (mm_flags & AV_CPU_FLAG_3DNOW)
2303         dsputil_init_3dnow(c, avctx, mm_flags);
2304
2305     if (mm_flags & AV_CPU_FLAG_SSE)
2306         dsputil_init_sse(c, avctx, mm_flags);
2307
2308     if (mm_flags & AV_CPU_FLAG_SSE2)
2309         dsputil_init_sse2(c, avctx, mm_flags);
2310
2311     if (mm_flags & AV_CPU_FLAG_SSSE3)
2312         dsputil_init_ssse3(c, avctx, mm_flags);
2313
2314     if (mm_flags & AV_CPU_FLAG_SSE4)
2315         dsputil_init_sse4(c, avctx, mm_flags);
2316
2317     if (mm_flags & AV_CPU_FLAG_AVX)
2318         dsputil_init_avx(c, avctx, mm_flags);
2319
2320     if (CONFIG_ENCODERS)
2321         ff_dsputilenc_init_mmx(c, avctx);
2322 }