dsputil_mmx: fix incorrect assembly code
[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_cpu.h"
27 #include "libavcodec/dsputil.h"
28 #include "libavcodec/h264dsp.h"
29 #include "libavcodec/mpegvideo.h"
30 #include "libavcodec/simple_idct.h"
31 #include "libavcodec/ac3dec.h"
32 #include "dsputil_mmx.h"
33 #include "idct_xvid.h"
34
35 //#undef NDEBUG
36 //#include <assert.h>
37
38 /* pixel operations */
39 DECLARE_ALIGNED(8,  const uint64_t, ff_bone) = 0x0101010101010101ULL;
40 DECLARE_ALIGNED(8,  const uint64_t, ff_wtwo) = 0x0002000200020002ULL;
41
42 DECLARE_ALIGNED(16, const uint64_t, ff_pdw_80000000)[2] =
43     { 0x8000000080000000ULL, 0x8000000080000000ULL };
44
45 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_1)    = { 0x0001000100010001ULL, 0x0001000100010001ULL };
46 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_2)    = { 0x0002000200020002ULL, 0x0002000200020002ULL };
47 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_3)    = { 0x0003000300030003ULL, 0x0003000300030003ULL };
48 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_4)    = { 0x0004000400040004ULL, 0x0004000400040004ULL };
49 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_5)    = { 0x0005000500050005ULL, 0x0005000500050005ULL };
50 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_8)    = { 0x0008000800080008ULL, 0x0008000800080008ULL };
51 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_9)    = { 0x0009000900090009ULL, 0x0009000900090009ULL };
52 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_15)   =   0x000F000F000F000FULL;
53 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_16)   = { 0x0010001000100010ULL, 0x0010001000100010ULL };
54 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_17)   = { 0x0011001100110011ULL, 0x0011001100110011ULL };
55 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_18)   = { 0x0012001200120012ULL, 0x0012001200120012ULL };
56 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_20)   =   0x0014001400140014ULL;
57 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_27)   = { 0x001B001B001B001BULL, 0x001B001B001B001BULL };
58 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_28)   = { 0x001C001C001C001CULL, 0x001C001C001C001CULL };
59 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_32)   = { 0x0020002000200020ULL, 0x0020002000200020ULL };
60 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_42)   =   0x002A002A002A002AULL;
61 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_53)   =   0x0035003500350035ULL;
62 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_63)   = { 0x003F003F003F003FULL, 0x003F003F003F003FULL };
63 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_64)   = { 0x0040004000400040ULL, 0x0040004000400040ULL };
64 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_96)   =   0x0060006000600060ULL;
65 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_128)  =   0x0080008000800080ULL;
66 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_255)  =   0x00ff00ff00ff00ffULL;
67 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_512)  = { 0x0200020002000200ULL, 0x0200020002000200ULL };
68 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_1019) = { 0x03FB03FB03FB03FBULL, 0x03FB03FB03FB03FBULL };
69
70 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_0)    = { 0x0000000000000000ULL, 0x0000000000000000ULL };
71 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_1)    = { 0x0101010101010101ULL, 0x0101010101010101ULL };
72 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_3)    = { 0x0303030303030303ULL, 0x0303030303030303ULL };
73 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_4)    = { 0x0404040404040404ULL, 0x0404040404040404ULL };
74 DECLARE_ALIGNED(8,  const uint64_t, ff_pb_7)    =   0x0707070707070707ULL;
75 DECLARE_ALIGNED(8,  const uint64_t, ff_pb_1F)   =   0x1F1F1F1F1F1F1F1FULL;
76 DECLARE_ALIGNED(8,  const uint64_t, ff_pb_3F)   =   0x3F3F3F3F3F3F3F3FULL;
77 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_80)   = { 0x8080808080808080ULL, 0x8080808080808080ULL };
78 DECLARE_ALIGNED(8,  const uint64_t, ff_pb_81)   =   0x8181818181818181ULL;
79 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_A1)   = { 0xA1A1A1A1A1A1A1A1ULL, 0xA1A1A1A1A1A1A1A1ULL };
80 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_F8)   = { 0xF8F8F8F8F8F8F8F8ULL, 0xF8F8F8F8F8F8F8F8ULL };
81 DECLARE_ALIGNED(8,  const uint64_t, ff_pb_FC)   =   0xFCFCFCFCFCFCFCFCULL;
82 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_FE)   = { 0xFEFEFEFEFEFEFEFEULL, 0xFEFEFEFEFEFEFEFEULL };
83
84 DECLARE_ALIGNED(16, const double, ff_pd_1)[2] = { 1.0, 1.0 };
85 DECLARE_ALIGNED(16, const double, ff_pd_2)[2] = { 2.0, 2.0 };
86
87 #define JUMPALIGN()     __asm__ volatile (".p2align 3"::)
88 #define MOVQ_ZERO(regd) __asm__ volatile ("pxor %%"#regd", %%"#regd ::)
89
90 #define MOVQ_BFE(regd)                                  \
91     __asm__ volatile (                                  \
92         "pcmpeqd %%"#regd", %%"#regd"   \n\t"           \
93         "paddb   %%"#regd", %%"#regd"   \n\t" ::)
94
95 #ifndef PIC
96 #define MOVQ_BONE(regd) __asm__ volatile ("movq %0, %%"#regd" \n\t" :: "m"(ff_bone))
97 #define MOVQ_WTWO(regd) __asm__ volatile ("movq %0, %%"#regd" \n\t" :: "m"(ff_wtwo))
98 #else
99 // for shared library it's better to use this way for accessing constants
100 // pcmpeqd -> -1
101 #define MOVQ_BONE(regd)                                 \
102     __asm__ volatile (                                  \
103         "pcmpeqd  %%"#regd", %%"#regd"  \n\t"           \
104         "psrlw          $15, %%"#regd"  \n\t"           \
105         "packuswb %%"#regd", %%"#regd"  \n\t" ::)
106
107 #define MOVQ_WTWO(regd)                                 \
108     __asm__ volatile (                                  \
109         "pcmpeqd %%"#regd", %%"#regd"   \n\t"           \
110         "psrlw         $15, %%"#regd"   \n\t"           \
111         "psllw          $1, %%"#regd"   \n\t"::)
112
113 #endif
114
115 // using regr as temporary and for the output result
116 // first argument is unmodifed and second is trashed
117 // regfe is supposed to contain 0xfefefefefefefefe
118 #define PAVGB_MMX_NO_RND(rega, regb, regr, regfe)                \
119     "movq   "#rega", "#regr"            \n\t"                    \
120     "pand   "#regb", "#regr"            \n\t"                    \
121     "pxor   "#rega", "#regb"            \n\t"                    \
122     "pand  "#regfe", "#regb"            \n\t"                    \
123     "psrlq       $1, "#regb"            \n\t"                    \
124     "paddb  "#regb", "#regr"            \n\t"
125
126 #define PAVGB_MMX(rega, regb, regr, regfe)                       \
127     "movq   "#rega", "#regr"            \n\t"                    \
128     "por    "#regb", "#regr"            \n\t"                    \
129     "pxor   "#rega", "#regb"            \n\t"                    \
130     "pand  "#regfe", "#regb"            \n\t"                    \
131     "psrlq       $1, "#regb"            \n\t"                    \
132     "psubb  "#regb", "#regr"            \n\t"
133
134 // mm6 is supposed to contain 0xfefefefefefefefe
135 #define PAVGBP_MMX_NO_RND(rega, regb, regr,  regc, regd, regp)   \
136     "movq  "#rega", "#regr"             \n\t"                    \
137     "movq  "#regc", "#regp"             \n\t"                    \
138     "pand  "#regb", "#regr"             \n\t"                    \
139     "pand  "#regd", "#regp"             \n\t"                    \
140     "pxor  "#rega", "#regb"             \n\t"                    \
141     "pxor  "#regc", "#regd"             \n\t"                    \
142     "pand    %%mm6, "#regb"             \n\t"                    \
143     "pand    %%mm6, "#regd"             \n\t"                    \
144     "psrlq      $1, "#regb"             \n\t"                    \
145     "psrlq      $1, "#regd"             \n\t"                    \
146     "paddb "#regb", "#regr"             \n\t"                    \
147     "paddb "#regd", "#regp"             \n\t"
148
149 #define PAVGBP_MMX(rega, regb, regr, regc, regd, regp)           \
150     "movq  "#rega", "#regr"             \n\t"                    \
151     "movq  "#regc", "#regp"             \n\t"                    \
152     "por   "#regb", "#regr"             \n\t"                    \
153     "por   "#regd", "#regp"             \n\t"                    \
154     "pxor  "#rega", "#regb"             \n\t"                    \
155     "pxor  "#regc", "#regd"             \n\t"                    \
156     "pand    %%mm6, "#regb"             \n\t"                    \
157     "pand    %%mm6, "#regd"             \n\t"                    \
158     "psrlq      $1, "#regd"             \n\t"                    \
159     "psrlq      $1, "#regb"             \n\t"                    \
160     "psubb "#regb", "#regr"             \n\t"                    \
161     "psubb "#regd", "#regp"             \n\t"
162
163 /***********************************/
164 /* MMX no rounding */
165 #define DEF(x, y) x ## _no_rnd_ ## y ## _mmx
166 #define SET_RND  MOVQ_WONE
167 #define PAVGBP(a, b, c, d, e, f)        PAVGBP_MMX_NO_RND(a, b, c, d, e, f)
168 #define PAVGB(a, b, c, e)               PAVGB_MMX_NO_RND(a, b, c, e)
169 #define OP_AVG(a, b, c, e)              PAVGB_MMX(a, b, c, e)
170
171 #include "dsputil_mmx_rnd_template.c"
172
173 #undef DEF
174 #undef SET_RND
175 #undef PAVGBP
176 #undef PAVGB
177 /***********************************/
178 /* MMX rounding */
179
180 #define DEF(x, y) x ## _ ## y ## _mmx
181 #define SET_RND  MOVQ_WTWO
182 #define PAVGBP(a, b, c, d, e, f)        PAVGBP_MMX(a, b, c, d, e, f)
183 #define PAVGB(a, b, c, e)               PAVGB_MMX(a, b, c, e)
184
185 #include "dsputil_mmx_rnd_template.c"
186
187 #undef DEF
188 #undef SET_RND
189 #undef PAVGBP
190 #undef PAVGB
191 #undef OP_AVG
192
193 /***********************************/
194 /* 3Dnow specific */
195
196 #define DEF(x) x ## _3dnow
197 #define PAVGB "pavgusb"
198 #define OP_AVG PAVGB
199
200 #include "dsputil_mmx_avg_template.c"
201
202 #undef DEF
203 #undef PAVGB
204 #undef OP_AVG
205
206 /***********************************/
207 /* MMX2 specific */
208
209 #define DEF(x) x ## _mmx2
210
211 /* Introduced only in MMX2 set */
212 #define PAVGB "pavgb"
213 #define OP_AVG PAVGB
214
215 #include "dsputil_mmx_avg_template.c"
216
217 #undef DEF
218 #undef PAVGB
219 #undef OP_AVG
220
221 #define put_no_rnd_pixels16_mmx put_pixels16_mmx
222 #define put_no_rnd_pixels8_mmx put_pixels8_mmx
223 #define put_pixels16_mmx2 put_pixels16_mmx
224 #define put_pixels8_mmx2 put_pixels8_mmx
225 #define put_pixels4_mmx2 put_pixels4_mmx
226 #define put_no_rnd_pixels16_mmx2 put_no_rnd_pixels16_mmx
227 #define put_no_rnd_pixels8_mmx2 put_no_rnd_pixels8_mmx
228 #define put_pixels16_3dnow put_pixels16_mmx
229 #define put_pixels8_3dnow put_pixels8_mmx
230 #define put_pixels4_3dnow put_pixels4_mmx
231 #define put_no_rnd_pixels16_3dnow put_no_rnd_pixels16_mmx
232 #define put_no_rnd_pixels8_3dnow put_no_rnd_pixels8_mmx
233
234 /***********************************/
235 /* standard MMX */
236
237 void ff_put_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels,
238                                int line_size)
239 {
240     const DCTELEM *p;
241     uint8_t *pix;
242
243     /* read the pixels */
244     p   = block;
245     pix = pixels;
246     /* unrolled loop */
247     __asm__ volatile (
248         "movq      (%3), %%mm0          \n\t"
249         "movq     8(%3), %%mm1          \n\t"
250         "movq    16(%3), %%mm2          \n\t"
251         "movq    24(%3), %%mm3          \n\t"
252         "movq    32(%3), %%mm4          \n\t"
253         "movq    40(%3), %%mm5          \n\t"
254         "movq    48(%3), %%mm6          \n\t"
255         "movq    56(%3), %%mm7          \n\t"
256         "packuswb %%mm1, %%mm0          \n\t"
257         "packuswb %%mm3, %%mm2          \n\t"
258         "packuswb %%mm5, %%mm4          \n\t"
259         "packuswb %%mm7, %%mm6          \n\t"
260         "movq     %%mm0, (%0)           \n\t"
261         "movq     %%mm2, (%0, %1)       \n\t"
262         "movq     %%mm4, (%0, %1, 2)    \n\t"
263         "movq     %%mm6, (%0, %2)       \n\t"
264         :: "r"(pix), "r"((x86_reg)line_size), "r"((x86_reg)line_size * 3),
265            "r"(p)
266         : "memory");
267     pix += line_size * 4;
268     p   += 32;
269
270     // if here would be an exact copy of the code above
271     // compiler would generate some very strange code
272     // thus using "r"
273     __asm__ volatile (
274         "movq       (%3), %%mm0         \n\t"
275         "movq      8(%3), %%mm1         \n\t"
276         "movq     16(%3), %%mm2         \n\t"
277         "movq     24(%3), %%mm3         \n\t"
278         "movq     32(%3), %%mm4         \n\t"
279         "movq     40(%3), %%mm5         \n\t"
280         "movq     48(%3), %%mm6         \n\t"
281         "movq     56(%3), %%mm7         \n\t"
282         "packuswb  %%mm1, %%mm0         \n\t"
283         "packuswb  %%mm3, %%mm2         \n\t"
284         "packuswb  %%mm5, %%mm4         \n\t"
285         "packuswb  %%mm7, %%mm6         \n\t"
286         "movq      %%mm0, (%0)          \n\t"
287         "movq      %%mm2, (%0, %1)      \n\t"
288         "movq      %%mm4, (%0, %1, 2)   \n\t"
289         "movq      %%mm6, (%0, %2)      \n\t"
290         :: "r"(pix), "r"((x86_reg)line_size), "r"((x86_reg)line_size * 3), "r"(p)
291         : "memory");
292 }
293
294 #define put_signed_pixels_clamped_mmx_half(off)             \
295     "movq          "#off"(%2), %%mm1        \n\t"           \
296     "movq     16 + "#off"(%2), %%mm2        \n\t"           \
297     "movq     32 + "#off"(%2), %%mm3        \n\t"           \
298     "movq     48 + "#off"(%2), %%mm4        \n\t"           \
299     "packsswb  8 + "#off"(%2), %%mm1        \n\t"           \
300     "packsswb 24 + "#off"(%2), %%mm2        \n\t"           \
301     "packsswb 40 + "#off"(%2), %%mm3        \n\t"           \
302     "packsswb 56 + "#off"(%2), %%mm4        \n\t"           \
303     "paddb              %%mm0, %%mm1        \n\t"           \
304     "paddb              %%mm0, %%mm2        \n\t"           \
305     "paddb              %%mm0, %%mm3        \n\t"           \
306     "paddb              %%mm0, %%mm4        \n\t"           \
307     "movq               %%mm1, (%0)         \n\t"           \
308     "movq               %%mm2, (%0, %3)     \n\t"           \
309     "movq               %%mm3, (%0, %3, 2)  \n\t"           \
310     "movq               %%mm4, (%0, %1)     \n\t"
311
312 void ff_put_signed_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels,
313                                       int line_size)
314 {
315     x86_reg line_skip = line_size;
316     x86_reg line_skip3;
317
318     __asm__ volatile (
319         "movq "MANGLE(ff_pb_80)", %%mm0     \n\t"
320         "lea         (%3, %3, 2), %1        \n\t"
321         put_signed_pixels_clamped_mmx_half(0)
322         "lea         (%0, %3, 4), %0        \n\t"
323         put_signed_pixels_clamped_mmx_half(64)
324         : "+&r"(pixels), "=&r"(line_skip3)
325         : "r"(block), "r"(line_skip)
326         : "memory");
327 }
328
329 void ff_add_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels,
330                                int line_size)
331 {
332     const DCTELEM *p;
333     uint8_t *pix;
334     int i;
335
336     /* read the pixels */
337     p   = block;
338     pix = pixels;
339     MOVQ_ZERO(mm7);
340     i = 4;
341     do {
342         __asm__ volatile (
343             "movq        (%2), %%mm0    \n\t"
344             "movq       8(%2), %%mm1    \n\t"
345             "movq      16(%2), %%mm2    \n\t"
346             "movq      24(%2), %%mm3    \n\t"
347             "movq          %0, %%mm4    \n\t"
348             "movq          %1, %%mm6    \n\t"
349             "movq       %%mm4, %%mm5    \n\t"
350             "punpcklbw  %%mm7, %%mm4    \n\t"
351             "punpckhbw  %%mm7, %%mm5    \n\t"
352             "paddsw     %%mm4, %%mm0    \n\t"
353             "paddsw     %%mm5, %%mm1    \n\t"
354             "movq       %%mm6, %%mm5    \n\t"
355             "punpcklbw  %%mm7, %%mm6    \n\t"
356             "punpckhbw  %%mm7, %%mm5    \n\t"
357             "paddsw     %%mm6, %%mm2    \n\t"
358             "paddsw     %%mm5, %%mm3    \n\t"
359             "packuswb   %%mm1, %%mm0    \n\t"
360             "packuswb   %%mm3, %%mm2    \n\t"
361             "movq       %%mm0, %0       \n\t"
362             "movq       %%mm2, %1       \n\t"
363             : "+m"(*pix), "+m"(*(pix + line_size))
364             : "r"(p)
365             : "memory");
366         pix += line_size * 2;
367         p   += 16;
368     } while (--i);
369 }
370
371 static void put_pixels4_mmx(uint8_t *block, const uint8_t *pixels,
372                             int line_size, int h)
373 {
374     __asm__ volatile (
375         "lea   (%3, %3), %%"REG_a"      \n\t"
376         ".p2align     3                 \n\t"
377         "1:                             \n\t"
378         "movd  (%1    ), %%mm0          \n\t"
379         "movd  (%1, %3), %%mm1          \n\t"
380         "movd     %%mm0, (%2)           \n\t"
381         "movd     %%mm1, (%2, %3)       \n\t"
382         "add  %%"REG_a", %1             \n\t"
383         "add  %%"REG_a", %2             \n\t"
384         "movd  (%1    ), %%mm0          \n\t"
385         "movd  (%1, %3), %%mm1          \n\t"
386         "movd     %%mm0, (%2)           \n\t"
387         "movd     %%mm1, (%2, %3)       \n\t"
388         "add  %%"REG_a", %1             \n\t"
389         "add  %%"REG_a", %2             \n\t"
390         "subl        $4, %0             \n\t"
391         "jnz         1b                 \n\t"
392         : "+g"(h), "+r"(pixels),  "+r"(block)
393         : "r"((x86_reg)line_size)
394         : "%"REG_a, "memory"
395         );
396 }
397
398 static void put_pixels8_mmx(uint8_t *block, const uint8_t *pixels,
399                             int line_size, int h)
400 {
401     __asm__ volatile (
402         "lea   (%3, %3), %%"REG_a"      \n\t"
403         ".p2align     3                 \n\t"
404         "1:                             \n\t"
405         "movq  (%1    ), %%mm0          \n\t"
406         "movq  (%1, %3), %%mm1          \n\t"
407         "movq     %%mm0, (%2)           \n\t"
408         "movq     %%mm1, (%2, %3)       \n\t"
409         "add  %%"REG_a", %1             \n\t"
410         "add  %%"REG_a", %2             \n\t"
411         "movq  (%1    ), %%mm0          \n\t"
412         "movq  (%1, %3), %%mm1          \n\t"
413         "movq     %%mm0, (%2)           \n\t"
414         "movq     %%mm1, (%2, %3)       \n\t"
415         "add  %%"REG_a", %1             \n\t"
416         "add  %%"REG_a", %2             \n\t"
417         "subl        $4, %0             \n\t"
418         "jnz         1b                 \n\t"
419         : "+g"(h), "+r"(pixels),  "+r"(block)
420         : "r"((x86_reg)line_size)
421         : "%"REG_a, "memory"
422         );
423 }
424
425 static void put_pixels16_mmx(uint8_t *block, const uint8_t *pixels,
426                              int line_size, int h)
427 {
428     __asm__ volatile (
429         "lea   (%3, %3), %%"REG_a"      \n\t"
430         ".p2align     3                 \n\t"
431         "1:                             \n\t"
432         "movq  (%1    ), %%mm0          \n\t"
433         "movq 8(%1    ), %%mm4          \n\t"
434         "movq  (%1, %3), %%mm1          \n\t"
435         "movq 8(%1, %3), %%mm5          \n\t"
436         "movq     %%mm0,  (%2)          \n\t"
437         "movq     %%mm4, 8(%2)          \n\t"
438         "movq     %%mm1,  (%2, %3)      \n\t"
439         "movq     %%mm5, 8(%2, %3)      \n\t"
440         "add  %%"REG_a", %1             \n\t"
441         "add  %%"REG_a", %2             \n\t"
442         "movq  (%1    ), %%mm0          \n\t"
443         "movq 8(%1    ), %%mm4          \n\t"
444         "movq  (%1, %3), %%mm1          \n\t"
445         "movq 8(%1, %3), %%mm5          \n\t"
446         "movq     %%mm0,  (%2)          \n\t"
447         "movq     %%mm4, 8(%2)          \n\t"
448         "movq     %%mm1,  (%2, %3)      \n\t"
449         "movq     %%mm5, 8(%2, %3)      \n\t"
450         "add  %%"REG_a", %1             \n\t"
451         "add  %%"REG_a", %2             \n\t"
452         "subl        $4, %0             \n\t"
453         "jnz         1b                 \n\t"
454         : "+g"(h), "+r"(pixels),  "+r"(block)
455         : "r"((x86_reg)line_size)
456         : "%"REG_a, "memory"
457         );
458 }
459
460 static void put_pixels16_sse2(uint8_t *block, const uint8_t *pixels,
461                               int line_size, int h)
462 {
463     __asm__ volatile (
464         "1:                              \n\t"
465         "movdqu (%1       ), %%xmm0      \n\t"
466         "movdqu (%1, %3   ), %%xmm1      \n\t"
467         "movdqu (%1, %3, 2), %%xmm2      \n\t"
468         "movdqu (%1, %4   ), %%xmm3      \n\t"
469         "lea    (%1, %3, 4), %1          \n\t"
470         "movdqa      %%xmm0, (%2)        \n\t"
471         "movdqa      %%xmm1, (%2, %3)    \n\t"
472         "movdqa      %%xmm2, (%2, %3, 2) \n\t"
473         "movdqa      %%xmm3, (%2, %4)    \n\t"
474         "subl            $4, %0          \n\t"
475         "lea    (%2, %3, 4), %2          \n\t"
476         "jnz             1b              \n\t"
477         : "+g"(h), "+r"(pixels),  "+r"(block)
478         : "r"((x86_reg)line_size), "r"((x86_reg)3L * line_size)
479         : "memory"
480         );
481 }
482
483 static void avg_pixels16_sse2(uint8_t *block, const uint8_t *pixels,
484                               int line_size, int h)
485 {
486     __asm__ volatile (
487         "1:                                 \n\t"
488         "movdqu (%1       ), %%xmm0         \n\t"
489         "movdqu (%1, %3   ), %%xmm1         \n\t"
490         "movdqu (%1, %3, 2), %%xmm2         \n\t"
491         "movdqu (%1, %4   ), %%xmm3         \n\t"
492         "lea    (%1, %3, 4), %1             \n\t"
493         "pavgb  (%2       ), %%xmm0         \n\t"
494         "pavgb  (%2, %3   ), %%xmm1         \n\t"
495         "pavgb  (%2, %3, 2), %%xmm2         \n\t"
496         "pavgb     (%2, %4), %%xmm3         \n\t"
497         "movdqa      %%xmm0, (%2)           \n\t"
498         "movdqa      %%xmm1, (%2, %3)       \n\t"
499         "movdqa      %%xmm2, (%2, %3, 2)    \n\t"
500         "movdqa      %%xmm3, (%2, %4)       \n\t"
501         "subl            $4, %0             \n\t"
502         "lea    (%2, %3, 4), %2             \n\t"
503         "jnz             1b                 \n\t"
504         : "+g"(h), "+r"(pixels),  "+r"(block)
505         : "r"((x86_reg)line_size), "r"((x86_reg)3L * line_size)
506         : "memory"
507         );
508 }
509
510 #define CLEAR_BLOCKS(name, n)                           \
511 static void name(DCTELEM *blocks)                       \
512 {                                                       \
513     __asm__ volatile (                                  \
514         "pxor %%mm7, %%mm7              \n\t"           \
515         "mov     %1,        %%"REG_a"   \n\t"           \
516         "1:                             \n\t"           \
517         "movq %%mm7,   (%0, %%"REG_a")  \n\t"           \
518         "movq %%mm7,  8(%0, %%"REG_a")  \n\t"           \
519         "movq %%mm7, 16(%0, %%"REG_a")  \n\t"           \
520         "movq %%mm7, 24(%0, %%"REG_a")  \n\t"           \
521         "add    $32, %%"REG_a"          \n\t"           \
522         "js      1b                     \n\t"           \
523         :: "r"(((uint8_t *)blocks) + 128 * n),          \
524            "i"(-128 * n)                                \
525         : "%"REG_a                                      \
526         );                                              \
527 }
528 CLEAR_BLOCKS(clear_blocks_mmx, 6)
529 CLEAR_BLOCKS(clear_block_mmx, 1)
530
531 static void clear_block_sse(DCTELEM *block)
532 {
533     __asm__ volatile (
534         "xorps  %%xmm0, %%xmm0          \n"
535         "movaps %%xmm0,    (%0)         \n"
536         "movaps %%xmm0,  16(%0)         \n"
537         "movaps %%xmm0,  32(%0)         \n"
538         "movaps %%xmm0,  48(%0)         \n"
539         "movaps %%xmm0,  64(%0)         \n"
540         "movaps %%xmm0,  80(%0)         \n"
541         "movaps %%xmm0,  96(%0)         \n"
542         "movaps %%xmm0, 112(%0)         \n"
543         :: "r"(block)
544         : "memory"
545     );
546 }
547
548 static void clear_blocks_sse(DCTELEM *blocks)
549 {
550     __asm__ volatile (
551         "xorps  %%xmm0, %%xmm0              \n"
552         "mov        %1,         %%"REG_a"   \n"
553         "1:                                 \n"
554         "movaps %%xmm0,    (%0, %%"REG_a")  \n"
555         "movaps %%xmm0,  16(%0, %%"REG_a")  \n"
556         "movaps %%xmm0,  32(%0, %%"REG_a")  \n"
557         "movaps %%xmm0,  48(%0, %%"REG_a")  \n"
558         "movaps %%xmm0,  64(%0, %%"REG_a")  \n"
559         "movaps %%xmm0,  80(%0, %%"REG_a")  \n"
560         "movaps %%xmm0,  96(%0, %%"REG_a")  \n"
561         "movaps %%xmm0, 112(%0, %%"REG_a")  \n"
562         "add      $128,         %%"REG_a"   \n"
563         "js         1b                      \n"
564         :: "r"(((uint8_t *)blocks) + 128 * 6),
565            "i"(-128 * 6)
566         : "%"REG_a
567     );
568 }
569
570 static void add_bytes_mmx(uint8_t *dst, uint8_t *src, int w)
571 {
572     x86_reg i = 0;
573     __asm__ volatile (
574         "jmp          2f                \n\t"
575         "1:                             \n\t"
576         "movq   (%1, %0), %%mm0         \n\t"
577         "movq   (%2, %0), %%mm1         \n\t"
578         "paddb     %%mm0, %%mm1         \n\t"
579         "movq      %%mm1, (%2, %0)      \n\t"
580         "movq  8(%1, %0), %%mm0         \n\t"
581         "movq  8(%2, %0), %%mm1         \n\t"
582         "paddb     %%mm0, %%mm1         \n\t"
583         "movq      %%mm1, 8(%2, %0)     \n\t"
584         "add         $16, %0            \n\t"
585         "2:                             \n\t"
586         "cmp          %3, %0            \n\t"
587         "js           1b                \n\t"
588         : "+r"(i)
589         : "r"(src), "r"(dst), "r"((x86_reg)w - 15)
590     );
591     for ( ; i < w; i++)
592         dst[i + 0] += src[i + 0];
593 }
594
595 #if HAVE_7REGS
596 static void add_hfyu_median_prediction_cmov(uint8_t *dst, const uint8_t *top,
597                                             const uint8_t *diff, int w,
598                                             int *left, int *left_top)
599 {
600     x86_reg w2 = -w;
601     x86_reg x;
602     int l  = *left     & 0xff;
603     int tl = *left_top & 0xff;
604     int t;
605     __asm__ volatile (
606         "mov          %7, %3            \n"
607         "1:                             \n"
608         "movzbl (%3, %4), %2            \n"
609         "mov          %2, %k3           \n"
610         "sub         %b1, %b3           \n"
611         "add         %b0, %b3           \n"
612         "mov          %2, %1            \n"
613         "cmp          %0, %2            \n"
614         "cmovg        %0, %2            \n"
615         "cmovg        %1, %0            \n"
616         "cmp         %k3, %0            \n"
617         "cmovg       %k3, %0            \n"
618         "mov          %7, %3            \n"
619         "cmp          %2, %0            \n"
620         "cmovl        %2, %0            \n"
621         "add    (%6, %4), %b0           \n"
622         "mov         %b0, (%5, %4)      \n"
623         "inc          %4                \n"
624         "jl           1b                \n"
625         : "+&q"(l), "+&q"(tl), "=&r"(t), "=&q"(x), "+&r"(w2)
626         : "r"(dst + w), "r"(diff + w), "rm"(top + w)
627     );
628     *left     = l;
629     *left_top = tl;
630 }
631 #endif
632
633 static inline void transpose4x4(uint8_t *dst, uint8_t *src, x86_reg dst_stride, x86_reg src_stride){
634     __asm__ volatile( //FIXME could save 1 instruction if done as 8x4 ...
635         "movd  (%1), %%mm0              \n\t"
636         "add   %3, %1                   \n\t"
637         "movd  (%1), %%mm1              \n\t"
638         "movd  (%1,%3,1), %%mm2         \n\t"
639         "movd  (%1,%3,2), %%mm3         \n\t"
640         "punpcklbw %%mm1, %%mm0         \n\t"
641         "punpcklbw %%mm3, %%mm2         \n\t"
642         "movq %%mm0, %%mm1              \n\t"
643         "punpcklwd %%mm2, %%mm0         \n\t"
644         "punpckhwd %%mm2, %%mm1         \n\t"
645         "movd  %%mm0, (%0)              \n\t"
646         "add   %2, %0                   \n\t"
647         "punpckhdq %%mm0, %%mm0         \n\t"
648         "movd  %%mm0, (%0)              \n\t"
649         "movd  %%mm1, (%0,%2,1)         \n\t"
650         "punpckhdq %%mm1, %%mm1         \n\t"
651         "movd  %%mm1, (%0,%2,2)         \n\t"
652
653         :  "+&r" (dst),
654            "+&r" (src)
655         :  "r" (dst_stride),
656            "r" (src_stride)
657         :  "memory"
658     );
659 }
660
661 #define H263_LOOP_FILTER                        \
662     "pxor      %%mm7, %%mm7             \n\t"   \
663     "movq         %0, %%mm0             \n\t"   \
664     "movq         %0, %%mm1             \n\t"   \
665     "movq         %3, %%mm2             \n\t"   \
666     "movq         %3, %%mm3             \n\t"   \
667     "punpcklbw %%mm7, %%mm0             \n\t"   \
668     "punpckhbw %%mm7, %%mm1             \n\t"   \
669     "punpcklbw %%mm7, %%mm2             \n\t"   \
670     "punpckhbw %%mm7, %%mm3             \n\t"   \
671     "psubw     %%mm2, %%mm0             \n\t"   \
672     "psubw     %%mm3, %%mm1             \n\t"   \
673     "movq         %1, %%mm2             \n\t"   \
674     "movq         %1, %%mm3             \n\t"   \
675     "movq         %2, %%mm4             \n\t"   \
676     "movq         %2, %%mm5             \n\t"   \
677     "punpcklbw %%mm7, %%mm2             \n\t"   \
678     "punpckhbw %%mm7, %%mm3             \n\t"   \
679     "punpcklbw %%mm7, %%mm4             \n\t"   \
680     "punpckhbw %%mm7, %%mm5             \n\t"   \
681     "psubw     %%mm2, %%mm4             \n\t"   \
682     "psubw     %%mm3, %%mm5             \n\t"   \
683     "psllw        $2, %%mm4             \n\t"   \
684     "psllw        $2, %%mm5             \n\t"   \
685     "paddw     %%mm0, %%mm4             \n\t"   \
686     "paddw     %%mm1, %%mm5             \n\t"   \
687     "pxor      %%mm6, %%mm6             \n\t"   \
688     "pcmpgtw   %%mm4, %%mm6             \n\t"   \
689     "pcmpgtw   %%mm5, %%mm7             \n\t"   \
690     "pxor      %%mm6, %%mm4             \n\t"   \
691     "pxor      %%mm7, %%mm5             \n\t"   \
692     "psubw     %%mm6, %%mm4             \n\t"   \
693     "psubw     %%mm7, %%mm5             \n\t"   \
694     "psrlw        $3, %%mm4             \n\t"   \
695     "psrlw        $3, %%mm5             \n\t"   \
696     "packuswb  %%mm5, %%mm4             \n\t"   \
697     "packsswb  %%mm7, %%mm6             \n\t"   \
698     "pxor      %%mm7, %%mm7             \n\t"   \
699     "movd         %4, %%mm2             \n\t"   \
700     "punpcklbw %%mm2, %%mm2             \n\t"   \
701     "punpcklbw %%mm2, %%mm2             \n\t"   \
702     "punpcklbw %%mm2, %%mm2             \n\t"   \
703     "psubusb   %%mm4, %%mm2             \n\t"   \
704     "movq      %%mm2, %%mm3             \n\t"   \
705     "psubusb   %%mm4, %%mm3             \n\t"   \
706     "psubb     %%mm3, %%mm2             \n\t"   \
707     "movq         %1, %%mm3             \n\t"   \
708     "movq         %2, %%mm4             \n\t"   \
709     "pxor      %%mm6, %%mm3             \n\t"   \
710     "pxor      %%mm6, %%mm4             \n\t"   \
711     "paddusb   %%mm2, %%mm3             \n\t"   \
712     "psubusb   %%mm2, %%mm4             \n\t"   \
713     "pxor      %%mm6, %%mm3             \n\t"   \
714     "pxor      %%mm6, %%mm4             \n\t"   \
715     "paddusb   %%mm2, %%mm2             \n\t"   \
716     "packsswb  %%mm1, %%mm0             \n\t"   \
717     "pcmpgtb   %%mm0, %%mm7             \n\t"   \
718     "pxor      %%mm7, %%mm0             \n\t"   \
719     "psubb     %%mm7, %%mm0             \n\t"   \
720     "movq      %%mm0, %%mm1             \n\t"   \
721     "psubusb   %%mm2, %%mm0             \n\t"   \
722     "psubb     %%mm0, %%mm1             \n\t"   \
723     "pand         %5, %%mm1             \n\t"   \
724     "psrlw        $2, %%mm1             \n\t"   \
725     "pxor      %%mm7, %%mm1             \n\t"   \
726     "psubb     %%mm7, %%mm1             \n\t"   \
727     "movq         %0, %%mm5             \n\t"   \
728     "movq         %3, %%mm6             \n\t"   \
729     "psubb     %%mm1, %%mm5             \n\t"   \
730     "paddb     %%mm1, %%mm6             \n\t"
731
732 static void h263_v_loop_filter_mmx(uint8_t *src, int stride, int qscale)
733 {
734     if (CONFIG_H263_DECODER || CONFIG_H263_ENCODER) {
735         const int strength = ff_h263_loop_filter_strength[qscale];
736
737         __asm__ volatile (
738             H263_LOOP_FILTER
739
740             "movq %%mm3, %1             \n\t"
741             "movq %%mm4, %2             \n\t"
742             "movq %%mm5, %0             \n\t"
743             "movq %%mm6, %3             \n\t"
744             : "+m"(*(uint64_t*)(src - 2 * stride)),
745               "+m"(*(uint64_t*)(src - 1 * stride)),
746               "+m"(*(uint64_t*)(src + 0 * stride)),
747               "+m"(*(uint64_t*)(src + 1 * stride))
748             : "g"(2 * strength), "m"(ff_pb_FC)
749             );
750     }
751 }
752
753 static void h263_h_loop_filter_mmx(uint8_t *src, int stride, int qscale)
754 {
755     if (CONFIG_H263_DECODER || CONFIG_H263_ENCODER) {
756         const int strength = ff_h263_loop_filter_strength[qscale];
757         DECLARE_ALIGNED(8, uint64_t, temp)[4];
758         uint8_t *btemp = (uint8_t*)temp;
759
760         src -= 2;
761
762         transpose4x4(btemp,     src,              8, stride);
763         transpose4x4(btemp + 4, src + 4 * stride, 8, stride);
764         __asm__ volatile (
765             H263_LOOP_FILTER // 5 3 4 6
766
767             : "+m"(temp[0]),
768               "+m"(temp[1]),
769               "+m"(temp[2]),
770               "+m"(temp[3])
771             : "g"(2 * strength), "m"(ff_pb_FC)
772             );
773
774         __asm__ volatile (
775             "movq      %%mm5, %%mm1         \n\t"
776             "movq      %%mm4, %%mm0         \n\t"
777             "punpcklbw %%mm3, %%mm5         \n\t"
778             "punpcklbw %%mm6, %%mm4         \n\t"
779             "punpckhbw %%mm3, %%mm1         \n\t"
780             "punpckhbw %%mm6, %%mm0         \n\t"
781             "movq      %%mm5, %%mm3         \n\t"
782             "movq      %%mm1, %%mm6         \n\t"
783             "punpcklwd %%mm4, %%mm5         \n\t"
784             "punpcklwd %%mm0, %%mm1         \n\t"
785             "punpckhwd %%mm4, %%mm3         \n\t"
786             "punpckhwd %%mm0, %%mm6         \n\t"
787             "movd      %%mm5, (%0)          \n\t"
788             "punpckhdq %%mm5, %%mm5         \n\t"
789             "movd      %%mm5, (%0, %2)      \n\t"
790             "movd      %%mm3, (%0, %2, 2)   \n\t"
791             "punpckhdq %%mm3, %%mm3         \n\t"
792             "movd      %%mm3, (%0, %3)      \n\t"
793             "movd      %%mm1, (%1)          \n\t"
794             "punpckhdq %%mm1, %%mm1         \n\t"
795             "movd      %%mm1, (%1, %2)      \n\t"
796             "movd      %%mm6, (%1, %2, 2)   \n\t"
797             "punpckhdq %%mm6, %%mm6         \n\t"
798             "movd      %%mm6, (%1, %3)      \n\t"
799             :: "r"(src),
800                "r"(src + 4 * stride),
801                "r"((x86_reg)stride),
802                "r"((x86_reg)(3 * stride))
803             );
804     }
805 }
806
807 /* Draw the edges of width 'w' of an image of size width, height
808  * this MMX version can only handle w == 8 || w == 16. */
809 static void draw_edges_mmx(uint8_t *buf, int wrap, int width, int height,
810                            int w, int h, int sides)
811 {
812     uint8_t *ptr, *last_line;
813     int i;
814
815     last_line = buf + (height - 1) * wrap;
816     /* left and right */
817     ptr = buf;
818     if (w == 8) {
819         __asm__ volatile (
820             "1:                             \n\t"
821             "movd            (%0), %%mm0    \n\t"
822             "punpcklbw      %%mm0, %%mm0    \n\t"
823             "punpcklwd      %%mm0, %%mm0    \n\t"
824             "punpckldq      %%mm0, %%mm0    \n\t"
825             "movq           %%mm0, -8(%0)   \n\t"
826             "movq      -8(%0, %2), %%mm1    \n\t"
827             "punpckhbw      %%mm1, %%mm1    \n\t"
828             "punpckhwd      %%mm1, %%mm1    \n\t"
829             "punpckhdq      %%mm1, %%mm1    \n\t"
830             "movq           %%mm1, (%0, %2) \n\t"
831             "add               %1, %0       \n\t"
832             "cmp               %3, %0       \n\t"
833             "jb                1b           \n\t"
834             : "+r"(ptr)
835             : "r"((x86_reg)wrap), "r"((x86_reg)width), "r"(ptr + wrap * height)
836             );
837     } else {
838         __asm__ volatile (
839             "1:                                 \n\t"
840             "movd            (%0), %%mm0        \n\t"
841             "punpcklbw      %%mm0, %%mm0        \n\t"
842             "punpcklwd      %%mm0, %%mm0        \n\t"
843             "punpckldq      %%mm0, %%mm0        \n\t"
844             "movq           %%mm0, -8(%0)       \n\t"
845             "movq           %%mm0, -16(%0)      \n\t"
846             "movq      -8(%0, %2), %%mm1        \n\t"
847             "punpckhbw      %%mm1, %%mm1        \n\t"
848             "punpckhwd      %%mm1, %%mm1        \n\t"
849             "punpckhdq      %%mm1, %%mm1        \n\t"
850             "movq           %%mm1,  (%0, %2)    \n\t"
851             "movq           %%mm1, 8(%0, %2)    \n\t"
852             "add               %1, %0           \n\t"
853             "cmp               %3, %0           \n\t"
854             "jb                1b               \n\t"
855             : "+r"(ptr)
856             : "r"((x86_reg)wrap), "r"((x86_reg)width), "r"(ptr + wrap * height)
857             );
858     }
859
860     /* top and bottom (and hopefully also the corners) */
861     if (sides & EDGE_TOP) {
862         for (i = 0; i < h; i += 4) {
863             ptr = buf - (i + 1) * wrap - w;
864             __asm__ volatile (
865                 "1:                             \n\t"
866                 "movq (%1, %0), %%mm0           \n\t"
867                 "movq    %%mm0, (%0)            \n\t"
868                 "movq    %%mm0, (%0, %2)        \n\t"
869                 "movq    %%mm0, (%0, %2, 2)     \n\t"
870                 "movq    %%mm0, (%0, %3)        \n\t"
871                 "add        $8, %0              \n\t"
872                 "cmp        %4, %0              \n\t"
873                 "jb         1b                  \n\t"
874                 : "+r"(ptr)
875                 : "r"((x86_reg)buf - (x86_reg)ptr - w), "r"((x86_reg) -wrap),
876                   "r"((x86_reg) -wrap * 3), "r"(ptr + width + 2 * w)
877                 );
878         }
879     }
880
881     if (sides & EDGE_BOTTOM) {
882         for (i = 0; i < h; i += 4) {
883             ptr = last_line + (i + 1) * wrap - w;
884             __asm__ volatile (
885                 "1:                             \n\t"
886                 "movq (%1, %0), %%mm0           \n\t"
887                 "movq    %%mm0, (%0)            \n\t"
888                 "movq    %%mm0, (%0, %2)        \n\t"
889                 "movq    %%mm0, (%0, %2, 2)     \n\t"
890                 "movq    %%mm0, (%0, %3)        \n\t"
891                 "add        $8, %0              \n\t"
892                 "cmp        %4, %0              \n\t"
893                 "jb         1b                  \n\t"
894                 : "+r"(ptr)
895                 : "r"((x86_reg)last_line - (x86_reg)ptr - w),
896                   "r"((x86_reg)wrap), "r"((x86_reg)wrap * 3),
897                   "r"(ptr + width + 2 * w)
898                 );
899         }
900     }
901 }
902
903 #define QPEL_V_LOW(m3, m4, m5, m6, pw_20, pw_3, rnd,                      \
904                    in0, in1, in2, in7, out, OP)                           \
905     "paddw               "#m4", "#m3"   \n\t" /* x1 */                    \
906     "movq   "MANGLE(ff_pw_20)", %%mm4   \n\t" /* 20 */                    \
907     "pmullw              "#m3", %%mm4   \n\t" /* 20x1 */                  \
908     "movq               "#in7", "#m3"   \n\t" /* d */                     \
909     "movq               "#in0", %%mm5   \n\t" /* D */                     \
910     "paddw               "#m3", %%mm5   \n\t" /* x4 */                    \
911     "psubw               %%mm5, %%mm4   \n\t" /* 20x1 - x4 */             \
912     "movq               "#in1", %%mm5   \n\t" /* C */                     \
913     "movq               "#in2", %%mm6   \n\t" /* B */                     \
914     "paddw               "#m6", %%mm5   \n\t" /* x3 */                    \
915     "paddw               "#m5", %%mm6   \n\t" /* x2 */                    \
916     "paddw               %%mm6, %%mm6   \n\t" /* 2x2 */                   \
917     "psubw               %%mm6, %%mm5   \n\t" /* -2x2 + x3 */             \
918     "pmullw  "MANGLE(ff_pw_3)", %%mm5   \n\t" /* -6x2 + 3x3 */            \
919     "paddw              "#rnd", %%mm4   \n\t" /* x2 */                    \
920     "paddw               %%mm4, %%mm5   \n\t" /* 20x1 - 6x2 + 3x3 - x4 */ \
921     "psraw                  $5, %%mm5   \n\t"                             \
922     "packuswb            %%mm5, %%mm5   \n\t"                             \
923     OP(%%mm5, out, %%mm7, d)
924
925 #define QPEL_BASE(OPNAME, ROUNDER, RND, OP_MMX2, OP_3DNOW)                \
926 static void OPNAME ## mpeg4_qpel16_h_lowpass_mmx2(uint8_t *dst,           \
927                                                   uint8_t *src,           \
928                                                   int dstStride,          \
929                                                   int srcStride,          \
930                                                   int h)                  \
931 {                                                                         \
932     uint64_t temp;                                                        \
933                                                                           \
934     __asm__ volatile (                                                    \
935         "pxor      %%mm7, %%mm7             \n\t"                         \
936         "1:                                 \n\t"                         \
937         "movq       (%0), %%mm0             \n\t" /* ABCDEFGH */          \
938         "movq      %%mm0, %%mm1             \n\t" /* ABCDEFGH */          \
939         "movq      %%mm0, %%mm2             \n\t" /* ABCDEFGH */          \
940         "punpcklbw %%mm7, %%mm0             \n\t" /* 0A0B0C0D */          \
941         "punpckhbw %%mm7, %%mm1             \n\t" /* 0E0F0G0H */          \
942         "pshufw    $0x90, %%mm0, %%mm5      \n\t" /* 0A0A0B0C */          \
943         "pshufw    $0x41, %%mm0, %%mm6      \n\t" /* 0B0A0A0B */          \
944         "movq      %%mm2, %%mm3             \n\t" /* ABCDEFGH */          \
945         "movq      %%mm2, %%mm4             \n\t" /* ABCDEFGH */          \
946         "psllq        $8, %%mm2             \n\t" /* 0ABCDEFG */          \
947         "psllq       $16, %%mm3             \n\t" /* 00ABCDEF */          \
948         "psllq       $24, %%mm4             \n\t" /* 000ABCDE */          \
949         "punpckhbw %%mm7, %%mm2             \n\t" /* 0D0E0F0G */          \
950         "punpckhbw %%mm7, %%mm3             \n\t" /* 0C0D0E0F */          \
951         "punpckhbw %%mm7, %%mm4             \n\t" /* 0B0C0D0E */          \
952         "paddw     %%mm3, %%mm5             \n\t" /* b */                 \
953         "paddw     %%mm2, %%mm6             \n\t" /* c */                 \
954         "paddw     %%mm5, %%mm5             \n\t" /* 2b */                \
955         "psubw     %%mm5, %%mm6             \n\t" /* c - 2b */            \
956         "pshufw    $0x06, %%mm0, %%mm5      \n\t" /* 0C0B0A0A */          \
957         "pmullw "MANGLE(ff_pw_3)", %%mm6    \n\t" /* 3c - 6b */           \
958         "paddw     %%mm4, %%mm0             \n\t" /* a */                 \
959         "paddw     %%mm1, %%mm5             \n\t" /* d */                 \
960         "pmullw "MANGLE(ff_pw_20)", %%mm0   \n\t" /* 20a */               \
961         "psubw     %%mm5, %%mm0             \n\t" /* 20a - d */           \
962         "paddw        %6, %%mm6             \n\t"                         \
963         "paddw     %%mm6, %%mm0             \n\t" /* 20a - 6b + 3c - d */ \
964         "psraw        $5, %%mm0             \n\t"                         \
965         "movq      %%mm0, %5                \n\t"                         \
966         /* mm1 = EFGH, mm2 = DEFG, mm3 = CDEF, mm4 = BCDE, mm7 = 0 */     \
967                                                                           \
968         "movq      5(%0), %%mm0             \n\t" /* FGHIJKLM */          \
969         "movq      %%mm0, %%mm5             \n\t" /* FGHIJKLM */          \
970         "movq      %%mm0, %%mm6             \n\t" /* FGHIJKLM */          \
971         "psrlq        $8, %%mm0             \n\t" /* GHIJKLM0 */          \
972         "psrlq       $16, %%mm5             \n\t" /* HIJKLM00 */          \
973         "punpcklbw %%mm7, %%mm0             \n\t" /* 0G0H0I0J */          \
974         "punpcklbw %%mm7, %%mm5             \n\t" /* 0H0I0J0K */          \
975         "paddw     %%mm0, %%mm2             \n\t" /* b */                 \
976         "paddw     %%mm5, %%mm3             \n\t" /* c */                 \
977         "paddw     %%mm2, %%mm2             \n\t" /* 2b */                \
978         "psubw     %%mm2, %%mm3             \n\t" /* c - 2b */            \
979         "movq      %%mm6, %%mm2             \n\t" /* FGHIJKLM */          \
980         "psrlq       $24, %%mm6             \n\t" /* IJKLM000 */          \
981         "punpcklbw %%mm7, %%mm2             \n\t" /* 0F0G0H0I */          \
982         "punpcklbw %%mm7, %%mm6             \n\t" /* 0I0J0K0L */          \
983         "pmullw "MANGLE(ff_pw_3)", %%mm3    \n\t" /* 3c - 6b */           \
984         "paddw     %%mm2, %%mm1             \n\t" /* a */                 \
985         "paddw     %%mm6, %%mm4             \n\t" /* d */                 \
986         "pmullw "MANGLE(ff_pw_20)", %%mm1   \n\t" /* 20a */               \
987         "psubw     %%mm4, %%mm3             \n\t" /* - 6b +3c - d */      \
988         "paddw        %6, %%mm1             \n\t"                         \
989         "paddw     %%mm1, %%mm3             \n\t" /* 20a - 6b +3c - d */  \
990         "psraw        $5, %%mm3             \n\t"                         \
991         "movq         %5, %%mm1             \n\t"                         \
992         "packuswb  %%mm3, %%mm1             \n\t"                         \
993         OP_MMX2(%%mm1, (%1), %%mm4, q)                                    \
994         /* mm0 = GHIJ, mm2 = FGHI, mm5 = HIJK, mm6 = IJKL, mm7 = 0 */     \
995                                                                           \
996         "movq      9(%0), %%mm1             \n\t" /* JKLMNOPQ */          \
997         "movq      %%mm1, %%mm4             \n\t" /* JKLMNOPQ */          \
998         "movq      %%mm1, %%mm3             \n\t" /* JKLMNOPQ */          \
999         "psrlq        $8, %%mm1             \n\t" /* KLMNOPQ0 */          \
1000         "psrlq       $16, %%mm4             \n\t" /* LMNOPQ00 */          \
1001         "punpcklbw %%mm7, %%mm1             \n\t" /* 0K0L0M0N */          \
1002         "punpcklbw %%mm7, %%mm4             \n\t" /* 0L0M0N0O */          \
1003         "paddw     %%mm1, %%mm5             \n\t" /* b */                 \
1004         "paddw     %%mm4, %%mm0             \n\t" /* c */                 \
1005         "paddw     %%mm5, %%mm5             \n\t" /* 2b */                \
1006         "psubw     %%mm5, %%mm0             \n\t" /* c - 2b */            \
1007         "movq      %%mm3, %%mm5             \n\t" /* JKLMNOPQ */          \
1008         "psrlq       $24, %%mm3             \n\t" /* MNOPQ000 */          \
1009         "pmullw "MANGLE(ff_pw_3)", %%mm0    \n\t" /* 3c - 6b */           \
1010         "punpcklbw %%mm7, %%mm3             \n\t" /* 0M0N0O0P */          \
1011         "paddw     %%mm3, %%mm2             \n\t" /* d */                 \
1012         "psubw     %%mm2, %%mm0             \n\t" /* -6b + 3c - d */      \
1013         "movq      %%mm5, %%mm2             \n\t" /* JKLMNOPQ */          \
1014         "punpcklbw %%mm7, %%mm2             \n\t" /* 0J0K0L0M */          \
1015         "punpckhbw %%mm7, %%mm5             \n\t" /* 0N0O0P0Q */          \
1016         "paddw     %%mm2, %%mm6             \n\t" /* a */                 \
1017         "pmullw "MANGLE(ff_pw_20)", %%mm6   \n\t" /* 20a */               \
1018         "paddw        %6, %%mm0             \n\t"                         \
1019         "paddw     %%mm6, %%mm0             \n\t" /* 20a - 6b + 3c - d */ \
1020         "psraw        $5, %%mm0             \n\t"                         \
1021         /* mm1 = KLMN, mm2 = JKLM, mm3 = MNOP, */                         \
1022         /* mm4 = LMNO, mm5 = NOPQ mm7 = 0 */                              \
1023                                                                           \
1024         "paddw    %%mm5, %%mm3              \n\t" /* a */                 \
1025         "pshufw   $0xF9, %%mm5, %%mm6       \n\t" /* 0O0P0Q0Q */          \
1026         "paddw    %%mm4, %%mm6              \n\t" /* b */                 \
1027         "pshufw   $0xBE, %%mm5, %%mm4       \n\t" /* 0P0Q0Q0P */          \
1028         "pshufw   $0x6F, %%mm5, %%mm5       \n\t" /* 0Q0Q0P0O */          \
1029         "paddw    %%mm1, %%mm4              \n\t" /* c */                 \
1030         "paddw    %%mm2, %%mm5              \n\t" /* d */                 \
1031         "paddw    %%mm6, %%mm6              \n\t" /* 2b */                \
1032         "psubw    %%mm6, %%mm4              \n\t" /* c - 2b */            \
1033         "pmullw "MANGLE(ff_pw_20)", %%mm3   \n\t" /* 20a */               \
1034         "pmullw  "MANGLE(ff_pw_3)", %%mm4   \n\t" /* 3c - 6b */           \
1035         "psubw    %%mm5, %%mm3              \n\t" /* -6b + 3c - d */      \
1036         "paddw       %6, %%mm4              \n\t"                         \
1037         "paddw    %%mm3, %%mm4              \n\t" /* 20a - 6b + 3c - d */ \
1038         "psraw       $5, %%mm4              \n\t"                         \
1039         "packuswb %%mm4, %%mm0              \n\t"                         \
1040         OP_MMX2(%%mm0, 8(%1), %%mm4, q)                                   \
1041                                                                           \
1042         "add         %3, %0                 \n\t"                         \
1043         "add         %4, %1                 \n\t"                         \
1044         "decl        %2                     \n\t"                         \
1045         "jnz         1b                     \n\t"                         \
1046         : "+a"(src), "+c"(dst), "+D"(h)                                   \
1047         : "d"((x86_reg)srcStride), "S"((x86_reg)dstStride),               \
1048           /* "m"(ff_pw_20), "m"(ff_pw_3), */ "m"(temp), "m"(ROUNDER)      \
1049         : "memory"                                                        \
1050         );                                                                \
1051 }                                                                         \
1052                                                                           \
1053 static void OPNAME ## mpeg4_qpel16_h_lowpass_3dnow(uint8_t *dst,          \
1054                                                    uint8_t *src,          \
1055                                                    int dstStride,         \
1056                                                    int srcStride,         \
1057                                                    int h)                 \
1058 {                                                                         \
1059     int i;                                                                \
1060     int16_t temp[16];                                                     \
1061     /* quick HACK, XXX FIXME MUST be optimized */                         \
1062     for (i = 0; i < h; i++) {                                             \
1063         temp[ 0] = (src[ 0] + src[ 1]) * 20 - (src[ 0] + src[ 2]) * 6 +   \
1064                    (src[ 1] + src[ 3]) *  3 - (src[ 2] + src[ 4]);        \
1065         temp[ 1] = (src[ 1] + src[ 2]) * 20 - (src[ 0] + src[ 3]) * 6 +   \
1066                    (src[ 0] + src[ 4]) *  3 - (src[ 1] + src[ 5]);        \
1067         temp[ 2] = (src[ 2] + src[ 3]) * 20 - (src[ 1] + src[ 4]) * 6 +   \
1068                    (src[ 0] + src[ 5]) *  3 - (src[ 0] + src[ 6]);        \
1069         temp[ 3] = (src[ 3] + src[ 4]) * 20 - (src[ 2] + src[ 5]) * 6 +   \
1070                    (src[ 1] + src[ 6]) *  3 - (src[ 0] + src[ 7]);        \
1071         temp[ 4] = (src[ 4] + src[ 5]) * 20 - (src[ 3] + src[ 6]) * 6 +   \
1072                    (src[ 2] + src[ 7]) *  3 - (src[ 1] + src[ 8]);        \
1073         temp[ 5] = (src[ 5] + src[ 6]) * 20 - (src[ 4] + src[ 7]) * 6 +   \
1074                    (src[ 3] + src[ 8]) *  3 - (src[ 2] + src[ 9]);        \
1075         temp[ 6] = (src[ 6] + src[ 7]) * 20 - (src[ 5] + src[ 8]) * 6 +   \
1076                    (src[ 4] + src[ 9]) *  3 - (src[ 3] + src[10]);        \
1077         temp[ 7] = (src[ 7] + src[ 8]) * 20 - (src[ 6] + src[ 9]) * 6 +   \
1078                    (src[ 5] + src[10]) *  3 - (src[ 4] + src[11]);        \
1079         temp[ 8] = (src[ 8] + src[ 9]) * 20 - (src[ 7] + src[10]) * 6 +   \
1080                    (src[ 6] + src[11]) *  3 - (src[ 5] + src[12]);        \
1081         temp[ 9] = (src[ 9] + src[10]) * 20 - (src[ 8] + src[11]) * 6 +   \
1082                    (src[ 7] + src[12]) *  3 - (src[ 6] + src[13]);        \
1083         temp[10] = (src[10] + src[11]) * 20 - (src[ 9] + src[12]) * 6 +   \
1084                    (src[ 8] + src[13]) *  3 - (src[ 7] + src[14]);        \
1085         temp[11] = (src[11] + src[12]) * 20 - (src[10] + src[13]) * 6 +   \
1086                    (src[ 9] + src[14]) *  3 - (src[ 8] + src[15]);        \
1087         temp[12] = (src[12] + src[13]) * 20 - (src[11] + src[14]) * 6 +   \
1088                    (src[10] + src[15]) *  3 - (src[ 9] + src[16]);        \
1089         temp[13] = (src[13] + src[14]) * 20 - (src[12] + src[15]) * 6 +   \
1090                    (src[11] + src[16]) *  3 - (src[10] + src[16]);        \
1091         temp[14] = (src[14] + src[15]) * 20 - (src[13] + src[16]) * 6 +   \
1092                    (src[12] + src[16]) *  3 - (src[11] + src[15]);        \
1093         temp[15] = (src[15] + src[16]) * 20 - (src[14] + src[16]) * 6 +   \
1094                    (src[13] + src[15]) *  3 - (src[12] + src[14]);        \
1095         __asm__ volatile (                                                \
1096             "movq      (%0), %%mm0          \n\t"                         \
1097             "movq     8(%0), %%mm1          \n\t"                         \
1098             "paddw       %2, %%mm0          \n\t"                         \
1099             "paddw       %2, %%mm1          \n\t"                         \
1100             "psraw       $5, %%mm0          \n\t"                         \
1101             "psraw       $5, %%mm1          \n\t"                         \
1102             "packuswb %%mm1, %%mm0          \n\t"                         \
1103             OP_3DNOW(%%mm0, (%1), %%mm1, q)                               \
1104             "movq    16(%0), %%mm0          \n\t"                         \
1105             "movq    24(%0), %%mm1          \n\t"                         \
1106             "paddw       %2, %%mm0          \n\t"                         \
1107             "paddw       %2, %%mm1          \n\t"                         \
1108             "psraw       $5, %%mm0          \n\t"                         \
1109             "psraw       $5, %%mm1          \n\t"                         \
1110             "packuswb %%mm1, %%mm0          \n\t"                         \
1111             OP_3DNOW(%%mm0, 8(%1), %%mm1, q)                              \
1112             :: "r"(temp), "r"(dst), "m"(ROUNDER)                          \
1113             : "memory"                                                    \
1114             );                                                            \
1115         dst += dstStride;                                                 \
1116         src += srcStride;                                                 \
1117     }                                                                     \
1118 }                                                                         \
1119                                                                           \
1120 static void OPNAME ## mpeg4_qpel8_h_lowpass_mmx2(uint8_t *dst,            \
1121                                                  uint8_t *src,            \
1122                                                  int dstStride,           \
1123                                                  int srcStride,           \
1124                                                  int h)                   \
1125 {                                                                         \
1126     __asm__ volatile (                                                    \
1127         "pxor      %%mm7, %%mm7             \n\t"                         \
1128         "1:                                 \n\t"                         \
1129         "movq       (%0), %%mm0             \n\t" /* ABCDEFGH */          \
1130         "movq      %%mm0, %%mm1             \n\t" /* ABCDEFGH */          \
1131         "movq      %%mm0, %%mm2             \n\t" /* ABCDEFGH */          \
1132         "punpcklbw %%mm7, %%mm0             \n\t" /* 0A0B0C0D */          \
1133         "punpckhbw %%mm7, %%mm1             \n\t" /* 0E0F0G0H */          \
1134         "pshufw    $0x90, %%mm0, %%mm5      \n\t" /* 0A0A0B0C */          \
1135         "pshufw    $0x41, %%mm0, %%mm6      \n\t" /* 0B0A0A0B */          \
1136         "movq      %%mm2, %%mm3             \n\t" /* ABCDEFGH */          \
1137         "movq      %%mm2, %%mm4             \n\t" /* ABCDEFGH */          \
1138         "psllq        $8, %%mm2             \n\t" /* 0ABCDEFG */          \
1139         "psllq       $16, %%mm3             \n\t" /* 00ABCDEF */          \
1140         "psllq       $24, %%mm4             \n\t" /* 000ABCDE */          \
1141         "punpckhbw %%mm7, %%mm2             \n\t" /* 0D0E0F0G */          \
1142         "punpckhbw %%mm7, %%mm3             \n\t" /* 0C0D0E0F */          \
1143         "punpckhbw %%mm7, %%mm4             \n\t" /* 0B0C0D0E */          \
1144         "paddw     %%mm3, %%mm5             \n\t" /* b */                 \
1145         "paddw     %%mm2, %%mm6             \n\t" /* c */                 \
1146         "paddw     %%mm5, %%mm5             \n\t" /* 2b */                \
1147         "psubw     %%mm5, %%mm6             \n\t" /* c - 2b */            \
1148         "pshufw    $0x06, %%mm0, %%mm5      \n\t" /* 0C0B0A0A */          \
1149         "pmullw "MANGLE(ff_pw_3)", %%mm6    \n\t" /* 3c - 6b */           \
1150         "paddw     %%mm4, %%mm0             \n\t" /* a */                 \
1151         "paddw     %%mm1, %%mm5             \n\t" /* d */                 \
1152         "pmullw "MANGLE(ff_pw_20)", %%mm0   \n\t" /* 20a */               \
1153         "psubw     %%mm5, %%mm0             \n\t" /* 20a - d */           \
1154         "paddw        %5, %%mm6             \n\t"                         \
1155         "paddw     %%mm6, %%mm0             \n\t" /* 20a - 6b + 3c - d */ \
1156         "psraw        $5, %%mm0             \n\t"                         \
1157         /* mm1 = EFGH, mm2 = DEFG, mm3 = CDEF, mm4 = BCDE, mm7 = 0 */     \
1158                                                                           \
1159         "movd      5(%0), %%mm5             \n\t" /* FGHI */              \
1160         "punpcklbw %%mm7, %%mm5             \n\t" /* 0F0G0H0I */          \
1161         "pshufw    $0xF9, %%mm5, %%mm6      \n\t" /* 0G0H0I0I */          \
1162         "paddw     %%mm5, %%mm1             \n\t" /* a */                 \
1163         "paddw     %%mm6, %%mm2             \n\t" /* b */                 \
1164         "pshufw    $0xBE, %%mm5, %%mm6      \n\t" /* 0H0I0I0H */          \
1165         "pshufw    $0x6F, %%mm5, %%mm5      \n\t" /* 0I0I0H0G */          \
1166         "paddw     %%mm6, %%mm3             \n\t" /* c */                 \
1167         "paddw     %%mm5, %%mm4             \n\t" /* d */                 \
1168         "paddw     %%mm2, %%mm2             \n\t" /* 2b */                \
1169         "psubw     %%mm2, %%mm3             \n\t" /* c - 2b */            \
1170         "pmullw "MANGLE(ff_pw_20)", %%mm1   \n\t" /* 20a */               \
1171         "pmullw  "MANGLE(ff_pw_3)", %%mm3   \n\t" /* 3c - 6b */           \
1172         "psubw     %%mm4, %%mm3             \n\t" /* -6b + 3c - d */      \
1173         "paddw        %5, %%mm1             \n\t"                         \
1174         "paddw     %%mm1, %%mm3             \n\t" /* 20a - 6b + 3c - d */ \
1175         "psraw        $5, %%mm3             \n\t"                         \
1176         "packuswb  %%mm3, %%mm0             \n\t"                         \
1177         OP_MMX2(%%mm0, (%1), %%mm4, q)                                    \
1178                                                                           \
1179         "add          %3, %0                \n\t"                         \
1180         "add          %4, %1                \n\t"                         \
1181         "decl         %2                    \n\t"                         \
1182         "jnz          1b                    \n\t"                         \
1183         : "+a"(src), "+c"(dst), "+d"(h)                                   \
1184         : "S"((x86_reg)srcStride), "D"((x86_reg)dstStride),               \
1185           /* "m"(ff_pw_20), "m"(ff_pw_3), */ "m"(ROUNDER)                 \
1186         : "memory"                                                        \
1187         );                                                                \
1188 }                                                                         \
1189                                                                           \
1190 static void OPNAME ## mpeg4_qpel8_h_lowpass_3dnow(uint8_t *dst,           \
1191                                                   uint8_t *src,           \
1192                                                   int dstStride,          \
1193                                                   int srcStride,          \
1194                                                   int h)                  \
1195 {                                                                         \
1196     int i;                                                                \
1197     int16_t temp[8];                                                      \
1198     /* quick HACK, XXX FIXME MUST be optimized */                         \
1199     for (i = 0; i < h; i++) {                                             \
1200         temp[0] = (src[0] + src[1]) * 20 - (src[0] + src[2]) * 6 +        \
1201                   (src[1] + src[3]) *  3 - (src[2] + src[4]);             \
1202         temp[1] = (src[1] + src[2]) * 20 - (src[0] + src[3]) * 6 +        \
1203                   (src[0] + src[4]) *  3 - (src[1] + src[5]);             \
1204         temp[2] = (src[2] + src[3]) * 20 - (src[1] + src[4]) * 6 +        \
1205                   (src[0] + src[5]) *  3 - (src[0] + src[6]);             \
1206         temp[3] = (src[3] + src[4]) * 20 - (src[2] + src[5]) * 6 +        \
1207                   (src[1] + src[6]) *  3 - (src[0] + src[7]);             \
1208         temp[4] = (src[4] + src[5]) * 20 - (src[3] + src[6]) * 6 +        \
1209                   (src[2] + src[7]) *  3 - (src[1] + src[8]);             \
1210         temp[5] = (src[5] + src[6]) * 20 - (src[4] + src[7]) * 6 +        \
1211                   (src[3] + src[8]) *  3 - (src[2] + src[8]);             \
1212         temp[6] = (src[6] + src[7]) * 20 - (src[5] + src[8]) * 6 +        \
1213                   (src[4] + src[8]) *  3 - (src[3] + src[7]);             \
1214         temp[7] = (src[7] + src[8]) * 20 - (src[6] + src[8]) * 6 +        \
1215                   (src[5] + src[7]) *  3 - (src[4] + src[6]);             \
1216         __asm__ volatile (                                                \
1217             "movq      (%0), %%mm0      \n\t"                             \
1218             "movq     8(%0), %%mm1      \n\t"                             \
1219             "paddw       %2, %%mm0      \n\t"                             \
1220             "paddw       %2, %%mm1      \n\t"                             \
1221             "psraw       $5, %%mm0      \n\t"                             \
1222             "psraw       $5, %%mm1      \n\t"                             \
1223             "packuswb %%mm1, %%mm0      \n\t"                             \
1224             OP_3DNOW(%%mm0, (%1), %%mm1, q)                               \
1225             :: "r"(temp), "r"(dst), "m"(ROUNDER)                          \
1226             : "memory"                                                    \
1227             );                                                            \
1228         dst += dstStride;                                                 \
1229         src += srcStride;                                                 \
1230     }                                                                     \
1231 }
1232
1233 #define QPEL_OP(OPNAME, ROUNDER, RND, OP, MMX)                          \
1234 static void OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(uint8_t *dst,      \
1235                                                      uint8_t *src,      \
1236                                                      int dstStride,     \
1237                                                      int srcStride)     \
1238 {                                                                       \
1239     uint64_t temp[17 * 4];                                              \
1240     uint64_t *temp_ptr = temp;                                          \
1241     int count = 17;                                                     \
1242                                                                         \
1243     /* FIXME unroll */                                                  \
1244     __asm__ volatile (                                                  \
1245         "pxor      %%mm7, %%mm7             \n\t"                       \
1246         "1:                                 \n\t"                       \
1247         "movq       (%0), %%mm0             \n\t"                       \
1248         "movq       (%0), %%mm1             \n\t"                       \
1249         "movq      8(%0), %%mm2             \n\t"                       \
1250         "movq      8(%0), %%mm3             \n\t"                       \
1251         "punpcklbw %%mm7, %%mm0             \n\t"                       \
1252         "punpckhbw %%mm7, %%mm1             \n\t"                       \
1253         "punpcklbw %%mm7, %%mm2             \n\t"                       \
1254         "punpckhbw %%mm7, %%mm3             \n\t"                       \
1255         "movq      %%mm0, (%1)              \n\t"                       \
1256         "movq      %%mm1, 17 * 8(%1)        \n\t"                       \
1257         "movq      %%mm2, 2 * 17 * 8(%1)    \n\t"                       \
1258         "movq      %%mm3, 3 * 17 * 8(%1)    \n\t"                       \
1259         "add          $8, %1                \n\t"                       \
1260         "add          %3, %0                \n\t"                       \
1261         "decl         %2                    \n\t"                       \
1262         "jnz          1b                    \n\t"                       \
1263         : "+r"(src), "+r"(temp_ptr), "+r"(count)                        \
1264         : "r"((x86_reg)srcStride)                                       \
1265         : "memory"                                                      \
1266         );                                                              \
1267                                                                         \
1268     temp_ptr = temp;                                                    \
1269     count    = 4;                                                       \
1270                                                                         \
1271     /* FIXME reorder for speed */                                       \
1272     __asm__ volatile (                                                  \
1273         /* "pxor  %%mm7, %%mm7            \n\t" */                      \
1274         "1:                             \n\t"                           \
1275         "movq    (%0), %%mm0            \n\t"                           \
1276         "movq   8(%0), %%mm1            \n\t"                           \
1277         "movq  16(%0), %%mm2            \n\t"                           \
1278         "movq  24(%0), %%mm3            \n\t"                           \
1279         QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 16(%0),   8(%0),    (%0),  32(%0), (%1),     OP) \
1280         QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5,  8(%0),    (%0),    (%0),  40(%0), (%1, %3), OP) \
1281         "add       %4, %1               \n\t"                           \
1282         QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5,   (%0),    (%0),   8(%0),  48(%0), (%1),     OP) \
1283                                                                         \
1284         QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5,   (%0),   8(%0),  16(%0),  56(%0), (%1, %3), OP) \
1285         "add       %4, %1               \n\t"                           \
1286         QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5,  8(%0),  16(%0),  24(%0),  64(%0), (%1),     OP) \
1287         QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 16(%0),  24(%0),  32(%0),  72(%0), (%1, %3), OP) \
1288         "add       %4, %1               \n\t"                           \
1289         QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 24(%0),  32(%0),  40(%0),  80(%0), (%1),     OP) \
1290         QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 32(%0),  40(%0),  48(%0),  88(%0), (%1, %3), OP) \
1291         "add       %4, %1               \n\t"                           \
1292         QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 40(%0),  48(%0),  56(%0),  96(%0), (%1),     OP) \
1293         QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 48(%0),  56(%0),  64(%0), 104(%0), (%1, %3), OP) \
1294         "add       %4, %1               \n\t"                           \
1295         QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 56(%0),  64(%0),  72(%0), 112(%0), (%1),     OP) \
1296         QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 64(%0),  72(%0),  80(%0), 120(%0), (%1, %3), OP) \
1297         "add       %4, %1               \n\t"                           \
1298         QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 72(%0),  80(%0),  88(%0), 128(%0), (%1),     OP) \
1299                                                                         \
1300         QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 80(%0),  88(%0),  96(%0), 128(%0), (%1, %3), OP) \
1301         "add       %4, %1               \n\t"                           \
1302         QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 88(%0),  96(%0), 104(%0), 120(%0), (%1),     OP) \
1303         QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 96(%0), 104(%0), 112(%0), 112(%0), (%1, %3), OP) \
1304                                                                         \
1305         "add     $136, %0               \n\t"                           \
1306         "add       %6, %1               \n\t"                           \
1307         "decl      %2                   \n\t"                           \
1308         "jnz       1b                   \n\t"                           \
1309                                                                         \
1310         : "+r"(temp_ptr), "+r"(dst), "+g"(count)                        \
1311         : "r"((x86_reg)dstStride), "r"(2 * (x86_reg)dstStride),         \
1312           /* "m"(ff_pw_20), "m"(ff_pw_3), */ "m"(ROUNDER),              \
1313           "g"(4 - 14 * (x86_reg)dstStride)                              \
1314         : "memory"                                                      \
1315         );                                                              \
1316 }                                                                       \
1317                                                                         \
1318 static void OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(uint8_t *dst,       \
1319                                                     uint8_t *src,       \
1320                                                     int dstStride,      \
1321                                                     int srcStride)      \
1322 {                                                                       \
1323     uint64_t temp[9 * 2];                                               \
1324     uint64_t *temp_ptr = temp;                                          \
1325     int count = 9;                                                      \
1326                                                                         \
1327     /* FIXME unroll */                                                  \
1328     __asm__ volatile (                                                  \
1329         "pxor      %%mm7, %%mm7         \n\t"                           \
1330         "1:                             \n\t"                           \
1331         "movq       (%0), %%mm0         \n\t"                           \
1332         "movq       (%0), %%mm1         \n\t"                           \
1333         "punpcklbw %%mm7, %%mm0         \n\t"                           \
1334         "punpckhbw %%mm7, %%mm1         \n\t"                           \
1335         "movq      %%mm0, (%1)          \n\t"                           \
1336         "movq      %%mm1, 9*8(%1)       \n\t"                           \
1337         "add          $8, %1            \n\t"                           \
1338         "add          %3, %0            \n\t"                           \
1339         "decl         %2                \n\t"                           \
1340         "jnz          1b                \n\t"                           \
1341         : "+r"(src), "+r"(temp_ptr), "+r"(count)                        \
1342         : "r"((x86_reg)srcStride)                                       \
1343         : "memory"                                                      \
1344         );                                                              \
1345                                                                         \
1346     temp_ptr = temp;                                                    \
1347     count    = 2;                                                       \
1348                                                                         \
1349     /* FIXME reorder for speed */                                       \
1350     __asm__ volatile (                                                  \
1351         /* "pxor  %%mm7, %%mm7            \n\t" */                      \
1352         "1:                             \n\t"                           \
1353         "movq    (%0), %%mm0            \n\t"                           \
1354         "movq   8(%0), %%mm1            \n\t"                           \
1355         "movq  16(%0), %%mm2            \n\t"                           \
1356         "movq  24(%0), %%mm3            \n\t"                           \
1357         QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 16(%0),  8(%0),   (%0), 32(%0), (%1), OP)     \
1358         QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5,  8(%0),   (%0),   (%0), 40(%0), (%1, %3), OP) \
1359         "add       %4, %1               \n\t"                           \
1360         QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5,   (%0),   (%0),  8(%0), 48(%0), (%1), OP)     \
1361                                                                         \
1362         QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5,   (%0),  8(%0), 16(%0), 56(%0), (%1, %3), OP) \
1363         "add       %4, %1               \n\t"                           \
1364         QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5,  8(%0), 16(%0), 24(%0), 64(%0), (%1), OP)     \
1365                                                                         \
1366         QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 16(%0), 24(%0), 32(%0), 64(%0), (%1, %3), OP) \
1367         "add       %4, %1               \n\t"                           \
1368         QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 24(%0), 32(%0), 40(%0), 56(%0), (%1), OP)     \
1369         QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 32(%0), 40(%0), 48(%0), 48(%0), (%1, %3), OP) \
1370                                                                         \
1371         "add      $72, %0               \n\t"                           \
1372         "add       %6, %1               \n\t"                           \
1373         "decl      %2                   \n\t"                           \
1374         "jnz       1b                   \n\t"                           \
1375                                                                         \
1376         : "+r"(temp_ptr), "+r"(dst), "+g"(count)                        \
1377         : "r"((x86_reg)dstStride), "r"(2 * (x86_reg)dstStride),         \
1378           /* "m"(ff_pw_20), "m"(ff_pw_3), */ "m"(ROUNDER),              \
1379           "g"(4 - 6 * (x86_reg)dstStride)                               \
1380         : "memory"                                                      \
1381         );                                                              \
1382 }                                                                       \
1383                                                                         \
1384 static void OPNAME ## qpel8_mc00_ ## MMX (uint8_t *dst, uint8_t *src,   \
1385                                           int stride)                   \
1386 {                                                                       \
1387     OPNAME ## pixels8_ ## MMX(dst, src, stride, 8);                     \
1388 }                                                                       \
1389                                                                         \
1390 static void OPNAME ## qpel8_mc10_ ## MMX(uint8_t *dst, uint8_t *src,    \
1391                                          int stride)                    \
1392 {                                                                       \
1393     uint64_t temp[8];                                                   \
1394     uint8_t * const half = (uint8_t*)temp;                              \
1395     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(half, src, 8,           \
1396                                                 stride, 8);             \
1397     OPNAME ## pixels8_l2_ ## MMX(dst, src, half, stride, stride, 8);    \
1398 }                                                                       \
1399                                                                         \
1400 static void OPNAME ## qpel8_mc20_ ## MMX(uint8_t *dst, uint8_t *src,    \
1401                                          int stride)                    \
1402 {                                                                       \
1403     OPNAME ## mpeg4_qpel8_h_lowpass_ ## MMX(dst, src, stride,           \
1404                                             stride, 8);                 \
1405 }                                                                       \
1406                                                                         \
1407 static void OPNAME ## qpel8_mc30_ ## MMX(uint8_t *dst, uint8_t *src,    \
1408                                          int stride)                    \
1409 {                                                                       \
1410     uint64_t temp[8];                                                   \
1411     uint8_t * const half = (uint8_t*)temp;                              \
1412     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(half, src, 8,           \
1413                                                 stride, 8);             \
1414     OPNAME ## pixels8_l2_ ## MMX(dst, src + 1, half, stride,            \
1415                                  stride, 8);                            \
1416 }                                                                       \
1417                                                                         \
1418 static void OPNAME ## qpel8_mc01_ ## MMX(uint8_t *dst, uint8_t *src,    \
1419                                          int stride)                    \
1420 {                                                                       \
1421     uint64_t temp[8];                                                   \
1422     uint8_t * const half = (uint8_t*)temp;                              \
1423     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(half, src, 8, stride);  \
1424     OPNAME ## pixels8_l2_ ## MMX(dst, src, half, stride, stride, 8);    \
1425 }                                                                       \
1426                                                                         \
1427 static void OPNAME ## qpel8_mc02_ ## MMX(uint8_t *dst, uint8_t *src,    \
1428                                          int stride)                    \
1429 {                                                                       \
1430     OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, src, stride, stride);  \
1431 }                                                                       \
1432                                                                         \
1433 static void OPNAME ## qpel8_mc03_ ## MMX(uint8_t *dst, uint8_t *src,    \
1434                                          int stride)                    \
1435 {                                                                       \
1436     uint64_t temp[8];                                                   \
1437     uint8_t * const half = (uint8_t*)temp;                              \
1438     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(half, src, 8, stride);  \
1439     OPNAME ## pixels8_l2_ ## MMX(dst, src + stride, half, stride,       \
1440                                  stride, 8);                            \
1441 }                                                                       \
1442                                                                         \
1443 static void OPNAME ## qpel8_mc11_ ## MMX(uint8_t *dst, uint8_t *src,    \
1444                                          int stride)                    \
1445 {                                                                       \
1446     uint64_t half[8 + 9];                                               \
1447     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
1448     uint8_t * const halfHV = ((uint8_t*)half);                          \
1449     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,          \
1450                                                 stride, 9);             \
1451     put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8, stride, 9);  \
1452     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);   \
1453     OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV, stride, 8, 8);     \
1454 }                                                                       \
1455                                                                         \
1456 static void OPNAME ## qpel8_mc31_ ## MMX(uint8_t *dst, uint8_t *src,    \
1457                                          int stride)                    \
1458 {                                                                       \
1459     uint64_t half[8 + 9];                                               \
1460     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
1461     uint8_t * const halfHV = ((uint8_t*)half);                          \
1462     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,          \
1463                                                 stride, 9);             \
1464     put ## RND ## pixels8_l2_ ## MMX(halfH, src + 1, halfH, 8,          \
1465                                      stride, 9);                        \
1466     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);   \
1467     OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV, stride, 8, 8);     \
1468 }                                                                       \
1469                                                                         \
1470 static void OPNAME ## qpel8_mc13_ ## MMX(uint8_t *dst, uint8_t *src,    \
1471                                          int stride)                    \
1472 {                                                                       \
1473     uint64_t half[8 + 9];                                               \
1474     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
1475     uint8_t * const halfHV = ((uint8_t*)half);                          \
1476     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,          \
1477                                                 stride, 9);             \
1478     put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8, stride, 9);  \
1479     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);   \
1480     OPNAME ## pixels8_l2_ ## MMX(dst, halfH + 8, halfHV, stride, 8, 8); \
1481 }                                                                       \
1482                                                                         \
1483 static void OPNAME ## qpel8_mc33_ ## MMX(uint8_t *dst, uint8_t *src,    \
1484                                          int stride)                    \
1485 {                                                                       \
1486     uint64_t half[8 + 9];                                               \
1487     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
1488     uint8_t * const halfHV = ((uint8_t*)half);                          \
1489     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,          \
1490                                                 stride, 9);             \
1491     put ## RND ## pixels8_l2_ ## MMX(halfH, src + 1, halfH, 8,          \
1492                                      stride, 9);                        \
1493     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);   \
1494     OPNAME ## pixels8_l2_ ## MMX(dst, halfH + 8, halfHV, stride, 8, 8); \
1495 }                                                                       \
1496                                                                         \
1497 static void OPNAME ## qpel8_mc21_ ## MMX(uint8_t *dst, uint8_t *src,    \
1498                                          int stride)                    \
1499 {                                                                       \
1500     uint64_t half[8 + 9];                                               \
1501     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
1502     uint8_t * const halfHV = ((uint8_t*)half);                          \
1503     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,          \
1504                                                 stride, 9);             \
1505     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);   \
1506     OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV, stride, 8, 8);     \
1507 }                                                                       \
1508                                                                         \
1509 static void OPNAME ## qpel8_mc23_ ## MMX(uint8_t *dst, uint8_t *src,    \
1510                                          int stride)                    \
1511 {                                                                       \
1512     uint64_t half[8 + 9];                                               \
1513     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
1514     uint8_t * const halfHV = ((uint8_t*)half);                          \
1515     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,          \
1516                                                 stride, 9);             \
1517     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);   \
1518     OPNAME ## pixels8_l2_ ## MMX(dst, halfH + 8, halfHV, stride, 8, 8); \
1519 }                                                                       \
1520                                                                         \
1521 static void OPNAME ## qpel8_mc12_ ## MMX(uint8_t *dst, uint8_t *src,    \
1522                                          int stride)                    \
1523 {                                                                       \
1524     uint64_t half[8 + 9];                                               \
1525     uint8_t * const halfH = ((uint8_t*)half);                           \
1526     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,          \
1527                                                 stride, 9);             \
1528     put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8, stride, 9);  \
1529     OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH, stride, 8);     \
1530 }                                                                       \
1531                                                                         \
1532 static void OPNAME ## qpel8_mc32_ ## MMX(uint8_t *dst, uint8_t *src,    \
1533                                          int stride)                    \
1534 {                                                                       \
1535     uint64_t half[8 + 9];                                               \
1536     uint8_t * const halfH = ((uint8_t*)half);                           \
1537     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,          \
1538                                                 stride, 9);             \
1539     put ## RND ## pixels8_l2_ ## MMX(halfH, src + 1, halfH, 8,          \
1540                                      stride, 9);                        \
1541     OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH, stride, 8);     \
1542 }                                                                       \
1543                                                                         \
1544 static void OPNAME ## qpel8_mc22_ ## MMX(uint8_t *dst, uint8_t *src,    \
1545                                          int stride)                    \
1546 {                                                                       \
1547     uint64_t half[9];                                                   \
1548     uint8_t * const halfH = ((uint8_t*)half);                           \
1549     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,          \
1550                                                 stride, 9);             \
1551     OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH, stride, 8);     \
1552 }                                                                       \
1553                                                                         \
1554 static void OPNAME ## qpel16_mc00_ ## MMX (uint8_t *dst, uint8_t *src,  \
1555                                            int stride)                  \
1556 {                                                                       \
1557     OPNAME ## pixels16_ ## MMX(dst, src, stride, 16);                   \
1558 }                                                                       \
1559                                                                         \
1560 static void OPNAME ## qpel16_mc10_ ## MMX(uint8_t *dst, uint8_t *src,   \
1561                                           int stride)                   \
1562 {                                                                       \
1563     uint64_t temp[32];                                                  \
1564     uint8_t * const half = (uint8_t*)temp;                              \
1565     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(half, src, 16,         \
1566                                                  stride, 16);           \
1567     OPNAME ## pixels16_l2_ ## MMX(dst, src, half, stride, stride, 16);  \
1568 }                                                                       \
1569                                                                         \
1570 static void OPNAME ## qpel16_mc20_ ## MMX(uint8_t *dst, uint8_t *src,   \
1571                                           int stride)                   \
1572 {                                                                       \
1573     OPNAME ## mpeg4_qpel16_h_lowpass_ ## MMX(dst, src,                  \
1574                                              stride, stride, 16);       \
1575 }                                                                       \
1576                                                                         \
1577 static void OPNAME ## qpel16_mc30_ ## MMX(uint8_t *dst, uint8_t *src,   \
1578                                           int stride)                   \
1579 {                                                                       \
1580     uint64_t temp[32];                                                  \
1581     uint8_t * const half = (uint8_t*)temp;                              \
1582     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(half, src, 16,         \
1583                                                  stride, 16);           \
1584     OPNAME ## pixels16_l2_ ## MMX(dst, src + 1, half,                   \
1585                                   stride, stride, 16);                  \
1586 }                                                                       \
1587                                                                         \
1588 static void OPNAME ## qpel16_mc01_ ## MMX(uint8_t *dst, uint8_t *src,   \
1589                                           int stride)                   \
1590 {                                                                       \
1591     uint64_t temp[32];                                                  \
1592     uint8_t * const half = (uint8_t*)temp;                              \
1593     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(half, src, 16,         \
1594                                                  stride);               \
1595     OPNAME ## pixels16_l2_ ## MMX(dst, src, half, stride, stride, 16);  \
1596 }                                                                       \
1597                                                                         \
1598 static void OPNAME ## qpel16_mc02_ ## MMX(uint8_t *dst, uint8_t *src,   \
1599                                           int stride)                   \
1600 {                                                                       \
1601     OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, src, stride, stride); \
1602 }                                                                       \
1603                                                                         \
1604 static void OPNAME ## qpel16_mc03_ ## MMX(uint8_t *dst, uint8_t *src,   \
1605                                           int stride)                   \
1606 {                                                                       \
1607     uint64_t temp[32];                                                  \
1608     uint8_t * const half = (uint8_t*)temp;                              \
1609     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(half, src, 16,         \
1610                                                  stride);               \
1611     OPNAME ## pixels16_l2_ ## MMX(dst, src+stride, half,                \
1612                                   stride, stride, 16);                  \
1613 }                                                                       \
1614                                                                         \
1615 static void OPNAME ## qpel16_mc11_ ## MMX(uint8_t *dst, uint8_t *src,   \
1616                                           int stride)                   \
1617 {                                                                       \
1618     uint64_t half[16 * 2 + 17 * 2];                                     \
1619     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1620     uint8_t * const halfHV = ((uint8_t*)half);                          \
1621     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,        \
1622                                                  stride, 17);           \
1623     put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16,            \
1624                                       stride, 17);                      \
1625     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,         \
1626                                                  16, 16);               \
1627     OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV, stride, 16, 16);  \
1628 }                                                                       \
1629                                                                         \
1630 static void OPNAME ## qpel16_mc31_ ## MMX(uint8_t *dst, uint8_t *src,   \
1631                                           int stride)                   \
1632 {                                                                       \
1633     uint64_t half[16 * 2 + 17 * 2];                                     \
1634     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1635     uint8_t * const halfHV = ((uint8_t*)half);                          \
1636     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,        \
1637                                                  stride, 17);           \
1638     put ## RND ## pixels16_l2_ ## MMX(halfH, src + 1, halfH, 16,        \
1639                                       stride, 17);                      \
1640     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,         \
1641                                                  16, 16);               \
1642     OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV, stride, 16, 16);  \
1643 }                                                                       \
1644                                                                         \
1645 static void OPNAME ## qpel16_mc13_ ## MMX(uint8_t *dst, uint8_t *src,   \
1646                                           int stride)                   \
1647 {                                                                       \
1648     uint64_t half[16 * 2 + 17 * 2];                                     \
1649     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1650     uint8_t * const halfHV = ((uint8_t*)half);                          \
1651     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,        \
1652                                                  stride, 17);           \
1653     put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16,            \
1654                                       stride, 17);                      \
1655     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,         \
1656                                                  16, 16);               \
1657     OPNAME ## pixels16_l2_ ## MMX(dst, halfH + 16, halfHV, stride,      \
1658                                   16, 16);                              \
1659 }                                                                       \
1660                                                                         \
1661 static void OPNAME ## qpel16_mc33_ ## MMX(uint8_t *dst, uint8_t *src,   \
1662                                           int stride)                   \
1663 {                                                                       \
1664     uint64_t half[16 * 2 + 17 * 2];                                     \
1665     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1666     uint8_t * const halfHV = ((uint8_t*)half);                          \
1667     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,        \
1668                                                  stride, 17);           \
1669     put ## RND ## pixels16_l2_ ## MMX(halfH, src + 1, halfH, 16,        \
1670                                       stride, 17);                      \
1671     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,         \
1672                                                  16, 16);               \
1673     OPNAME ## pixels16_l2_ ## MMX(dst, halfH + 16, halfHV, stride,      \
1674                                   16, 16);                              \
1675 }                                                                       \
1676                                                                         \
1677 static void OPNAME ## qpel16_mc21_ ## MMX(uint8_t *dst, uint8_t *src,   \
1678                                           int stride)                   \
1679 {                                                                       \
1680     uint64_t half[16 * 2 + 17 * 2];                                     \
1681     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1682     uint8_t * const halfHV = ((uint8_t*)half);                          \
1683     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,        \
1684                                                  stride, 17);           \
1685     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,         \
1686                                                  16, 16);               \
1687     OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV, stride, 16, 16);  \
1688 }                                                                       \
1689                                                                         \
1690 static void OPNAME ## qpel16_mc23_ ## MMX(uint8_t *dst, uint8_t *src,   \
1691                                           int stride)                   \
1692 {                                                                       \
1693     uint64_t half[16 * 2 + 17 * 2];                                     \
1694     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1695     uint8_t * const halfHV = ((uint8_t*)half);                          \
1696     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,        \
1697                                                  stride, 17);           \
1698     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,         \
1699                                                  16, 16);               \
1700     OPNAME ## pixels16_l2_ ## MMX(dst, halfH + 16, halfHV, stride,      \
1701                                   16, 16);                              \
1702 }                                                                       \
1703                                                                         \
1704 static void OPNAME ## qpel16_mc12_ ## MMX(uint8_t *dst, uint8_t *src,   \
1705                                           int stride)                   \
1706 {                                                                       \
1707     uint64_t half[17 * 2];                                              \
1708     uint8_t * const halfH = ((uint8_t*)half);                           \
1709     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,        \
1710                                                  stride, 17);           \
1711     put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16,            \
1712                                       stride, 17);                      \
1713     OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH, stride, 16);   \
1714 }                                                                       \
1715                                                                         \
1716 static void OPNAME ## qpel16_mc32_ ## MMX(uint8_t *dst, uint8_t *src,   \
1717                                           int stride)                   \
1718 {                                                                       \
1719     uint64_t half[17 * 2];                                              \
1720     uint8_t * const halfH = ((uint8_t*)half);                           \
1721     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,        \
1722                                                  stride, 17);           \
1723     put ## RND ## pixels16_l2_ ## MMX(halfH, src + 1, halfH, 16,        \
1724                                       stride, 17);                      \
1725     OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH, stride, 16);   \
1726 }                                                                       \
1727                                                                         \
1728 static void OPNAME ## qpel16_mc22_ ## MMX(uint8_t *dst, uint8_t *src,   \
1729                                           int stride)                   \
1730 {                                                                       \
1731     uint64_t half[17 * 2];                                              \
1732     uint8_t * const halfH = ((uint8_t*)half);                           \
1733     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,        \
1734                                                  stride, 17);           \
1735     OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH, stride, 16);   \
1736 }
1737
1738 #define PUT_OP(a, b, temp, size)                \
1739     "mov"#size"        "#a", "#b"       \n\t"
1740
1741 #define AVG_3DNOW_OP(a, b, temp, size)          \
1742     "mov"#size"        "#b", "#temp"    \n\t"   \
1743     "pavgusb        "#temp", "#a"       \n\t"   \
1744     "mov"#size"        "#a", "#b"       \n\t"
1745
1746 #define AVG_MMX2_OP(a, b, temp, size)           \
1747     "mov"#size"        "#b", "#temp"    \n\t"   \
1748     "pavgb          "#temp", "#a"       \n\t"   \
1749     "mov"#size"        "#a", "#b"       \n\t"
1750
1751 QPEL_BASE(put_,        ff_pw_16, _,        PUT_OP,       PUT_OP)
1752 QPEL_BASE(avg_,        ff_pw_16, _,        AVG_MMX2_OP,  AVG_3DNOW_OP)
1753 QPEL_BASE(put_no_rnd_, ff_pw_15, _no_rnd_, PUT_OP,       PUT_OP)
1754 QPEL_OP(put_,          ff_pw_16, _,        PUT_OP,       3dnow)
1755 QPEL_OP(avg_,          ff_pw_16, _,        AVG_3DNOW_OP, 3dnow)
1756 QPEL_OP(put_no_rnd_,   ff_pw_15, _no_rnd_, PUT_OP,       3dnow)
1757 QPEL_OP(put_,          ff_pw_16, _,        PUT_OP,       mmx2)
1758 QPEL_OP(avg_,          ff_pw_16, _,        AVG_MMX2_OP,  mmx2)
1759 QPEL_OP(put_no_rnd_,   ff_pw_15, _no_rnd_, PUT_OP,       mmx2)
1760
1761 /***********************************/
1762 /* bilinear qpel: not compliant to any spec, only for -lavdopts fast */
1763
1764 #define QPEL_2TAP_XY(OPNAME, SIZE, MMX, XY, HPEL)                              \
1765 static void OPNAME ## 2tap_qpel ## SIZE ## _mc ## XY ## _ ## MMX(uint8_t *dst, \
1766                                                                  uint8_t *src, \
1767                                                                  int stride)   \
1768 {                                                                              \
1769     OPNAME ## pixels ## SIZE ## HPEL(dst, src, stride, SIZE);                  \
1770 }
1771
1772 #define QPEL_2TAP_L3(OPNAME, SIZE, MMX, XY, S0, S1, S2)                        \
1773 static void OPNAME ## 2tap_qpel ## SIZE ## _mc ## XY ## _ ## MMX(uint8_t *dst, \
1774                                                                  uint8_t *src, \
1775                                                                  int stride)   \
1776 {                                                                              \
1777     OPNAME ## 2tap_qpel ## SIZE ## _l3_ ## MMX(dst, src + S0, stride, SIZE,    \
1778                                                S1, S2);                        \
1779 }
1780
1781 #define QPEL_2TAP(OPNAME, SIZE, MMX)                                        \
1782 QPEL_2TAP_XY(OPNAME, SIZE, MMX, 20, _x2_ ## MMX)                            \
1783 QPEL_2TAP_XY(OPNAME, SIZE, MMX, 02, _y2_ ## MMX)                            \
1784 QPEL_2TAP_XY(OPNAME, SIZE, MMX, 22, _xy2_mmx)                               \
1785 static const qpel_mc_func OPNAME ## 2tap_qpel ## SIZE ## _mc00_ ## MMX =    \
1786     OPNAME ## qpel ## SIZE ## _mc00_ ## MMX;                                \
1787 static const qpel_mc_func OPNAME ## 2tap_qpel ## SIZE ## _mc21_ ## MMX =    \
1788     OPNAME ## 2tap_qpel ## SIZE ## _mc20_ ## MMX;                           \
1789 static const qpel_mc_func OPNAME ## 2tap_qpel ## SIZE ## _mc12_ ## MMX =    \
1790     OPNAME ## 2tap_qpel ## SIZE ## _mc02_ ## MMX;                           \
1791 static void OPNAME ## 2tap_qpel ## SIZE ## _mc32_ ## MMX(uint8_t *dst,      \
1792                                                          uint8_t *src,      \
1793                                                          int stride)        \
1794 {                                                                           \
1795     OPNAME ## pixels ## SIZE ## _y2_ ## MMX(dst, src + 1, stride, SIZE);    \
1796 }                                                                           \
1797 static void OPNAME ## 2tap_qpel ## SIZE ## _mc23_ ## MMX(uint8_t *dst,      \
1798                                                          uint8_t *src,      \
1799                                                          int stride)        \
1800 {                                                                           \
1801     OPNAME ## pixels ## SIZE ## _x2_ ## MMX(dst, src + stride,              \
1802                                             stride, SIZE);                  \
1803 }                                                                           \
1804 QPEL_2TAP_L3(OPNAME, SIZE, MMX, 10, 0,           1,       0)                \
1805 QPEL_2TAP_L3(OPNAME, SIZE, MMX, 30, 1,          -1,       0)                \
1806 QPEL_2TAP_L3(OPNAME, SIZE, MMX, 01, 0,           stride,  0)                \
1807 QPEL_2TAP_L3(OPNAME, SIZE, MMX, 03, stride,     -stride,  0)                \
1808 QPEL_2TAP_L3(OPNAME, SIZE, MMX, 11, 0,           stride,  1)                \
1809 QPEL_2TAP_L3(OPNAME, SIZE, MMX, 31, 1,           stride, -1)                \
1810 QPEL_2TAP_L3(OPNAME, SIZE, MMX, 13, stride,     -stride,  1)                \
1811 QPEL_2TAP_L3(OPNAME, SIZE, MMX, 33, stride + 1, -stride, -1)                \
1812
1813 QPEL_2TAP(put_, 16, mmx2)
1814 QPEL_2TAP(avg_, 16, mmx2)
1815 QPEL_2TAP(put_,  8, mmx2)
1816 QPEL_2TAP(avg_,  8, mmx2)
1817 QPEL_2TAP(put_, 16, 3dnow)
1818 QPEL_2TAP(avg_, 16, 3dnow)
1819 QPEL_2TAP(put_,  8, 3dnow)
1820 QPEL_2TAP(avg_,  8, 3dnow)
1821
1822 void ff_put_rv40_qpel8_mc33_mmx(uint8_t *dst, uint8_t *src, int stride)
1823 {
1824   put_pixels8_xy2_mmx(dst, src, stride, 8);
1825 }
1826 void ff_put_rv40_qpel16_mc33_mmx(uint8_t *dst, uint8_t *src, int stride)
1827 {
1828   put_pixels16_xy2_mmx(dst, src, stride, 16);
1829 }
1830 void ff_avg_rv40_qpel8_mc33_mmx(uint8_t *dst, uint8_t *src, int stride)
1831 {
1832   avg_pixels8_xy2_mmx(dst, src, stride, 8);
1833 }
1834 void ff_avg_rv40_qpel16_mc33_mmx(uint8_t *dst, uint8_t *src, int stride)
1835 {
1836   avg_pixels16_xy2_mmx(dst, src, stride, 16);
1837 }
1838
1839 #if HAVE_YASM
1840 typedef void emu_edge_core_func(uint8_t *buf, const uint8_t *src,
1841                                 x86_reg linesize, x86_reg start_y,
1842                                 x86_reg end_y, x86_reg block_h,
1843                                 x86_reg start_x, x86_reg end_x,
1844                                 x86_reg block_w);
1845 extern emu_edge_core_func ff_emu_edge_core_mmx;
1846 extern emu_edge_core_func ff_emu_edge_core_sse;
1847
1848 static av_always_inline void emulated_edge_mc(uint8_t *buf, const uint8_t *src,
1849                                               int linesize,
1850                                               int block_w, int block_h,
1851                                               int src_x, int src_y,
1852                                               int w, int h,
1853                                               emu_edge_core_func *core_fn)
1854 {
1855     int start_y, start_x, end_y, end_x, src_y_add = 0;
1856
1857     if (src_y >= h) {
1858         src_y_add = h - 1 - src_y;
1859         src_y     = h - 1;
1860     } else if (src_y <= -block_h) {
1861         src_y_add = 1 - block_h - src_y;
1862         src_y     = 1 - block_h;
1863     }
1864     if (src_x >= w) {
1865         src   += w - 1 - src_x;
1866         src_x  = w - 1;
1867     } else if (src_x <= -block_w) {
1868         src   += 1 - block_w - src_x;
1869         src_x  = 1 - block_w;
1870     }
1871
1872     start_y = FFMAX(0, -src_y);
1873     start_x = FFMAX(0, -src_x);
1874     end_y   = FFMIN(block_h, h-src_y);
1875     end_x   = FFMIN(block_w, w-src_x);
1876     assert(start_x < end_x && block_w > 0);
1877     assert(start_y < end_y && block_h > 0);
1878
1879     // fill in the to-be-copied part plus all above/below
1880     src += (src_y_add + start_y) * linesize + start_x;
1881     buf += start_x;
1882     core_fn(buf, src, linesize, start_y, end_y,
1883             block_h, start_x, end_x, block_w);
1884 }
1885
1886 #if ARCH_X86_32
1887 static av_noinline void emulated_edge_mc_mmx(uint8_t *buf, const uint8_t *src,
1888                                              int linesize,
1889                                              int block_w, int block_h,
1890                                              int src_x, int src_y, int w, int h)
1891 {
1892     emulated_edge_mc(buf, src, linesize, block_w, block_h, src_x, src_y,
1893                      w, h, &ff_emu_edge_core_mmx);
1894 }
1895 #endif
1896
1897 static av_noinline void emulated_edge_mc_sse(uint8_t *buf, const uint8_t *src,
1898                                              int linesize,
1899                                              int block_w, int block_h,
1900                                              int src_x, int src_y, int w, int h)
1901 {
1902     emulated_edge_mc(buf, src, linesize, block_w, block_h, src_x, src_y,
1903                      w, h, &ff_emu_edge_core_sse);
1904 }
1905 #endif /* HAVE_YASM */
1906
1907 typedef void emulated_edge_mc_func(uint8_t *dst, const uint8_t *src,
1908                                    int linesize, int block_w, int block_h,
1909                                    int src_x, int src_y, int w, int h);
1910
1911 static av_always_inline void gmc(uint8_t *dst, uint8_t *src,
1912                                  int stride, int h, int ox, int oy,
1913                                  int dxx, int dxy, int dyx, int dyy,
1914                                  int shift, int r, int width, int height,
1915                                  emulated_edge_mc_func *emu_edge_fn)
1916 {
1917     const int w    = 8;
1918     const int ix   = ox  >> (16 + shift);
1919     const int iy   = oy  >> (16 + shift);
1920     const int oxs  = ox  >> 4;
1921     const int oys  = oy  >> 4;
1922     const int dxxs = dxx >> 4;
1923     const int dxys = dxy >> 4;
1924     const int dyxs = dyx >> 4;
1925     const int dyys = dyy >> 4;
1926     const uint16_t r4[4]   = { r, r, r, r };
1927     const uint16_t dxy4[4] = { dxys, dxys, dxys, dxys };
1928     const uint16_t dyy4[4] = { dyys, dyys, dyys, dyys };
1929     const uint64_t shift2 = 2 * shift;
1930     uint8_t edge_buf[(h + 1) * stride];
1931     int x, y;
1932
1933     const int dxw = (dxx - (1 << (16 + shift))) * (w - 1);
1934     const int dyh = (dyy - (1 << (16 + shift))) * (h - 1);
1935     const int dxh = dxy * (h - 1);
1936     const int dyw = dyx * (w - 1);
1937     if ( // non-constant fullpel offset (3% of blocks)
1938         ((ox ^ (ox + dxw)) | (ox ^ (ox + dxh)) | (ox ^ (ox + dxw + dxh)) |
1939          (oy ^ (oy + dyw)) | (oy ^ (oy + dyh)) | (oy ^ (oy + dyw + dyh))) >> (16 + shift)
1940         // uses more than 16 bits of subpel mv (only at huge resolution)
1941         || (dxx | dxy | dyx | dyy) & 15) {
1942         // FIXME could still use mmx for some of the rows
1943         ff_gmc_c(dst, src, stride, h, ox, oy, dxx, dxy, dyx, dyy,
1944                  shift, r, width, height);
1945         return;
1946     }
1947
1948     src += ix + iy * stride;
1949     if ((unsigned)ix >= width  - w ||
1950         (unsigned)iy >= height - h) {
1951         emu_edge_fn(edge_buf, src, stride, w + 1, h + 1, ix, iy, width, height);
1952         src = edge_buf;
1953     }
1954
1955     __asm__ volatile (
1956         "movd         %0, %%mm6         \n\t"
1957         "pxor      %%mm7, %%mm7         \n\t"
1958         "punpcklwd %%mm6, %%mm6         \n\t"
1959         "punpcklwd %%mm6, %%mm6         \n\t"
1960         :: "r"(1<<shift)
1961     );
1962
1963     for (x = 0; x < w; x += 4) {
1964         uint16_t dx4[4] = { oxs - dxys + dxxs * (x + 0),
1965                             oxs - dxys + dxxs * (x + 1),
1966                             oxs - dxys + dxxs * (x + 2),
1967                             oxs - dxys + dxxs * (x + 3) };
1968         uint16_t dy4[4] = { oys - dyys + dyxs * (x + 0),
1969                             oys - dyys + dyxs * (x + 1),
1970                             oys - dyys + dyxs * (x + 2),
1971                             oys - dyys + dyxs * (x + 3) };
1972
1973         for (y = 0; y < h; y++) {
1974             __asm__ volatile (
1975                 "movq      %0, %%mm4    \n\t"
1976                 "movq      %1, %%mm5    \n\t"
1977                 "paddw     %2, %%mm4    \n\t"
1978                 "paddw     %3, %%mm5    \n\t"
1979                 "movq   %%mm4, %0       \n\t"
1980                 "movq   %%mm5, %1       \n\t"
1981                 "psrlw    $12, %%mm4    \n\t"
1982                 "psrlw    $12, %%mm5    \n\t"
1983                 : "+m"(*dx4), "+m"(*dy4)
1984                 : "m"(*dxy4), "m"(*dyy4)
1985             );
1986
1987             __asm__ volatile (
1988                 "movq      %%mm6, %%mm2 \n\t"
1989                 "movq      %%mm6, %%mm1 \n\t"
1990                 "psubw     %%mm4, %%mm2 \n\t"
1991                 "psubw     %%mm5, %%mm1 \n\t"
1992                 "movq      %%mm2, %%mm0 \n\t"
1993                 "movq      %%mm4, %%mm3 \n\t"
1994                 "pmullw    %%mm1, %%mm0 \n\t" // (s - dx) * (s - dy)
1995                 "pmullw    %%mm5, %%mm3 \n\t" // dx * dy
1996                 "pmullw    %%mm5, %%mm2 \n\t" // (s - dx) * dy
1997                 "pmullw    %%mm4, %%mm1 \n\t" // dx * (s - dy)
1998
1999                 "movd         %4, %%mm5 \n\t"
2000                 "movd         %3, %%mm4 \n\t"
2001                 "punpcklbw %%mm7, %%mm5 \n\t"
2002                 "punpcklbw %%mm7, %%mm4 \n\t"
2003                 "pmullw    %%mm5, %%mm3 \n\t" // src[1, 1] * dx * dy
2004                 "pmullw    %%mm4, %%mm2 \n\t" // src[0, 1] * (s - dx) * dy
2005
2006                 "movd         %2, %%mm5 \n\t"
2007                 "movd         %1, %%mm4 \n\t"
2008                 "punpcklbw %%mm7, %%mm5 \n\t"
2009                 "punpcklbw %%mm7, %%mm4 \n\t"
2010                 "pmullw    %%mm5, %%mm1 \n\t" // src[1, 0] * dx * (s - dy)
2011                 "pmullw    %%mm4, %%mm0 \n\t" // src[0, 0] * (s - dx) * (s - dy)
2012                 "paddw        %5, %%mm1 \n\t"
2013                 "paddw     %%mm3, %%mm2 \n\t"
2014                 "paddw     %%mm1, %%mm0 \n\t"
2015                 "paddw     %%mm2, %%mm0 \n\t"
2016
2017                 "psrlw        %6, %%mm0 \n\t"
2018                 "packuswb  %%mm0, %%mm0 \n\t"
2019                 "movd      %%mm0, %0    \n\t"
2020
2021                 : "=m"(dst[x + y * stride])
2022                 : "m"(src[0]), "m"(src[1]),
2023                   "m"(src[stride]), "m"(src[stride + 1]),
2024                   "m"(*r4), "m"(shift2)
2025             );
2026             src += stride;
2027         }
2028         src += 4 - h * stride;
2029     }
2030 }
2031
2032 #if HAVE_YASM
2033 #if ARCH_X86_32
2034 static void gmc_mmx(uint8_t *dst, uint8_t *src,
2035                     int stride, int h, int ox, int oy,
2036                     int dxx, int dxy, int dyx, int dyy,
2037                     int shift, int r, int width, int height)
2038 {
2039     gmc(dst, src, stride, h, ox, oy, dxx, dxy, dyx, dyy, shift, r,
2040         width, height, &emulated_edge_mc_mmx);
2041 }
2042 #endif
2043 static void gmc_sse(uint8_t *dst, uint8_t *src,
2044                     int stride, int h, int ox, int oy,
2045                     int dxx, int dxy, int dyx, int dyy,
2046                     int shift, int r, int width, int height)
2047 {
2048     gmc(dst, src, stride, h, ox, oy, dxx, dxy, dyx, dyy, shift, r,
2049         width, height, &emulated_edge_mc_sse);
2050 }
2051 #else
2052 static void gmc_mmx(uint8_t *dst, uint8_t *src,
2053                     int stride, int h, int ox, int oy,
2054                     int dxx, int dxy, int dyx, int dyy,
2055                     int shift, int r, int width, int height)
2056 {
2057     gmc(dst, src, stride, h, ox, oy, dxx, dxy, dyx, dyy, shift, r,
2058         width, height, &ff_emulated_edge_mc_8);
2059 }
2060 #endif
2061
2062 #define PREFETCH(name, op)                      \
2063 static void name(void *mem, int stride, int h)  \
2064 {                                               \
2065     const uint8_t *p = mem;                     \
2066     do {                                        \
2067         __asm__ volatile (#op" %0" :: "m"(*p)); \
2068         p += stride;                            \
2069     } while (--h);                              \
2070 }
2071
2072 PREFETCH(prefetch_mmx2,  prefetcht0)
2073 PREFETCH(prefetch_3dnow, prefetch)
2074 #undef PREFETCH
2075
2076 #include "h264_qpel_mmx.c"
2077
2078 void ff_put_h264_chroma_mc8_mmx_rnd  (uint8_t *dst, uint8_t *src,
2079                                       int stride, int h, int x, int y);
2080 void ff_avg_h264_chroma_mc8_mmx2_rnd (uint8_t *dst, uint8_t *src,
2081                                       int stride, int h, int x, int y);
2082 void ff_avg_h264_chroma_mc8_3dnow_rnd(uint8_t *dst, uint8_t *src,
2083                                       int stride, int h, int x, int y);
2084
2085 void ff_put_h264_chroma_mc4_mmx      (uint8_t *dst, uint8_t *src,
2086                                       int stride, int h, int x, int y);
2087 void ff_avg_h264_chroma_mc4_mmx2     (uint8_t *dst, uint8_t *src,
2088                                       int stride, int h, int x, int y);
2089 void ff_avg_h264_chroma_mc4_3dnow    (uint8_t *dst, uint8_t *src,
2090                                       int stride, int h, int x, int y);
2091
2092 void ff_put_h264_chroma_mc2_mmx2     (uint8_t *dst, uint8_t *src,
2093                                       int stride, int h, int x, int y);
2094 void ff_avg_h264_chroma_mc2_mmx2     (uint8_t *dst, uint8_t *src,
2095                                       int stride, int h, int x, int y);
2096
2097 void ff_put_h264_chroma_mc8_ssse3_rnd(uint8_t *dst, uint8_t *src,
2098                                       int stride, int h, int x, int y);
2099 void ff_put_h264_chroma_mc4_ssse3    (uint8_t *dst, uint8_t *src,
2100                                       int stride, int h, int x, int y);
2101
2102 void ff_avg_h264_chroma_mc8_ssse3_rnd(uint8_t *dst, uint8_t *src,
2103                                       int stride, int h, int x, int y);
2104 void ff_avg_h264_chroma_mc4_ssse3    (uint8_t *dst, uint8_t *src,
2105                                       int stride, int h, int x, int y);
2106
2107 #define CHROMA_MC(OP, NUM, DEPTH, OPT)                                  \
2108 void ff_ ## OP ## _h264_chroma_mc ## NUM ## _ ## DEPTH ## _ ## OPT      \
2109                                       (uint8_t *dst, uint8_t *src,      \
2110                                        int stride, int h, int x, int y);
2111
2112 CHROMA_MC(put, 2, 10, mmxext)
2113 CHROMA_MC(avg, 2, 10, mmxext)
2114 CHROMA_MC(put, 4, 10, mmxext)
2115 CHROMA_MC(avg, 4, 10, mmxext)
2116 CHROMA_MC(put, 8, 10, sse2)
2117 CHROMA_MC(avg, 8, 10, sse2)
2118 CHROMA_MC(put, 8, 10, avx)
2119 CHROMA_MC(avg, 8, 10, avx)
2120
2121 /* CAVS-specific */
2122 void ff_put_cavs_qpel8_mc00_mmx2(uint8_t *dst, uint8_t *src, int stride)
2123 {
2124     put_pixels8_mmx(dst, src, stride, 8);
2125 }
2126
2127 void ff_avg_cavs_qpel8_mc00_mmx2(uint8_t *dst, uint8_t *src, int stride)
2128 {
2129     avg_pixels8_mmx(dst, src, stride, 8);
2130 }
2131
2132 void ff_put_cavs_qpel16_mc00_mmx2(uint8_t *dst, uint8_t *src, int stride)
2133 {
2134     put_pixels16_mmx(dst, src, stride, 16);
2135 }
2136
2137 void ff_avg_cavs_qpel16_mc00_mmx2(uint8_t *dst, uint8_t *src, int stride)
2138 {
2139     avg_pixels16_mmx(dst, src, stride, 16);
2140 }
2141
2142 /* VC-1-specific */
2143 void ff_put_vc1_mspel_mc00_mmx(uint8_t *dst, const uint8_t *src,
2144                                int stride, int rnd)
2145 {
2146     put_pixels8_mmx(dst, src, stride, 8);
2147 }
2148
2149 void ff_avg_vc1_mspel_mc00_mmx2(uint8_t *dst, const uint8_t *src,
2150                                 int stride, int rnd)
2151 {
2152     avg_pixels8_mmx2(dst, src, stride, 8);
2153 }
2154
2155 /* XXX: Those functions should be suppressed ASAP when all IDCTs are
2156  * converted. */
2157 #if CONFIG_GPL
2158 static void ff_libmpeg2mmx_idct_put(uint8_t *dest, int line_size,
2159                                     DCTELEM *block)
2160 {
2161     ff_mmx_idct(block);
2162     ff_put_pixels_clamped_mmx(block, dest, line_size);
2163 }
2164
2165 static void ff_libmpeg2mmx_idct_add(uint8_t *dest, int line_size,
2166                                     DCTELEM *block)
2167 {
2168     ff_mmx_idct(block);
2169     ff_add_pixels_clamped_mmx(block, dest, line_size);
2170 }
2171
2172 static void ff_libmpeg2mmx2_idct_put(uint8_t *dest, int line_size,
2173                                      DCTELEM *block)
2174 {
2175     ff_mmxext_idct(block);
2176     ff_put_pixels_clamped_mmx(block, dest, line_size);
2177 }
2178
2179 static void ff_libmpeg2mmx2_idct_add(uint8_t *dest, int line_size,
2180                                      DCTELEM *block)
2181 {
2182     ff_mmxext_idct(block);
2183     ff_add_pixels_clamped_mmx(block, dest, line_size);
2184 }
2185 #endif
2186
2187 static void ff_idct_xvid_mmx_put(uint8_t *dest, int line_size, DCTELEM *block)
2188 {
2189     ff_idct_xvid_mmx(block);
2190     ff_put_pixels_clamped_mmx(block, dest, line_size);
2191 }
2192
2193 static void ff_idct_xvid_mmx_add(uint8_t *dest, int line_size, DCTELEM *block)
2194 {
2195     ff_idct_xvid_mmx(block);
2196     ff_add_pixels_clamped_mmx(block, dest, line_size);
2197 }
2198
2199 static void ff_idct_xvid_mmx2_put(uint8_t *dest, int line_size, DCTELEM *block)
2200 {
2201     ff_idct_xvid_mmx2(block);
2202     ff_put_pixels_clamped_mmx(block, dest, line_size);
2203 }
2204
2205 static void ff_idct_xvid_mmx2_add(uint8_t *dest, int line_size, DCTELEM *block)
2206 {
2207     ff_idct_xvid_mmx2(block);
2208     ff_add_pixels_clamped_mmx(block, dest, line_size);
2209 }
2210
2211 static void vorbis_inverse_coupling_3dnow(float *mag, float *ang, int blocksize)
2212 {
2213     int i;
2214     __asm__ volatile ("pxor %%mm7, %%mm7":);
2215     for (i = 0; i < blocksize; i += 2) {
2216         __asm__ volatile (
2217             "movq       %0, %%mm0   \n\t"
2218             "movq       %1, %%mm1   \n\t"
2219             "movq    %%mm0, %%mm2   \n\t"
2220             "movq    %%mm1, %%mm3   \n\t"
2221             "pfcmpge %%mm7, %%mm2   \n\t" // m <= 0.0
2222             "pfcmpge %%mm7, %%mm3   \n\t" // a <= 0.0
2223             "pslld     $31, %%mm2   \n\t" // keep only the sign bit
2224             "pxor    %%mm2, %%mm1   \n\t"
2225             "movq    %%mm3, %%mm4   \n\t"
2226             "pand    %%mm1, %%mm3   \n\t"
2227             "pandn   %%mm1, %%mm4   \n\t"
2228             "pfadd   %%mm0, %%mm3   \n\t" // a = m + ((a < 0) & (a ^ sign(m)))
2229             "pfsub   %%mm4, %%mm0   \n\t" // m = m + ((a > 0) & (a ^ sign(m)))
2230             "movq    %%mm3, %1      \n\t"
2231             "movq    %%mm0, %0      \n\t"
2232             : "+m"(mag[i]), "+m"(ang[i])
2233             :: "memory"
2234         );
2235     }
2236     __asm__ volatile ("femms");
2237 }
2238
2239 static void vorbis_inverse_coupling_sse(float *mag, float *ang, int blocksize)
2240 {
2241     int i;
2242
2243     __asm__ volatile (
2244         "movaps  %0, %%xmm5 \n\t"
2245         :: "m"(ff_pdw_80000000[0])
2246     );
2247     for (i = 0; i < blocksize; i += 4) {
2248         __asm__ volatile (
2249             "movaps      %0, %%xmm0 \n\t"
2250             "movaps      %1, %%xmm1 \n\t"
2251             "xorps   %%xmm2, %%xmm2 \n\t"
2252             "xorps   %%xmm3, %%xmm3 \n\t"
2253             "cmpleps %%xmm0, %%xmm2 \n\t" // m <= 0.0
2254             "cmpleps %%xmm1, %%xmm3 \n\t" // a <= 0.0
2255             "andps   %%xmm5, %%xmm2 \n\t" // keep only the sign bit
2256             "xorps   %%xmm2, %%xmm1 \n\t"
2257             "movaps  %%xmm3, %%xmm4 \n\t"
2258             "andps   %%xmm1, %%xmm3 \n\t"
2259             "andnps  %%xmm1, %%xmm4 \n\t"
2260             "addps   %%xmm0, %%xmm3 \n\t" // a = m + ((a < 0) & (a ^ sign(m)))
2261             "subps   %%xmm4, %%xmm0 \n\t" // m = m + ((a > 0) & (a ^ sign(m)))
2262             "movaps  %%xmm3, %1     \n\t"
2263             "movaps  %%xmm0, %0     \n\t"
2264             : "+m"(mag[i]), "+m"(ang[i])
2265             :: "memory"
2266         );
2267     }
2268 }
2269
2270 #define IF1(x) x
2271 #define IF0(x)
2272
2273 #define MIX5(mono, stereo)                                      \
2274     __asm__ volatile (                                          \
2275         "movss           0(%2), %%xmm5          \n"             \
2276         "movss           8(%2), %%xmm6          \n"             \
2277         "movss          24(%2), %%xmm7          \n"             \
2278         "shufps     $0, %%xmm5, %%xmm5          \n"             \
2279         "shufps     $0, %%xmm6, %%xmm6          \n"             \
2280         "shufps     $0, %%xmm7, %%xmm7          \n"             \
2281         "1:                                     \n"             \
2282         "movaps       (%0, %1), %%xmm0          \n"             \
2283         "movaps  0x400(%0, %1), %%xmm1          \n"             \
2284         "movaps  0x800(%0, %1), %%xmm2          \n"             \
2285         "movaps  0xc00(%0, %1), %%xmm3          \n"             \
2286         "movaps 0x1000(%0, %1), %%xmm4          \n"             \
2287         "mulps          %%xmm5, %%xmm0          \n"             \
2288         "mulps          %%xmm6, %%xmm1          \n"             \
2289         "mulps          %%xmm5, %%xmm2          \n"             \
2290         "mulps          %%xmm7, %%xmm3          \n"             \
2291         "mulps          %%xmm7, %%xmm4          \n"             \
2292  stereo("addps          %%xmm1, %%xmm0          \n")            \
2293         "addps          %%xmm1, %%xmm2          \n"             \
2294         "addps          %%xmm3, %%xmm0          \n"             \
2295         "addps          %%xmm4, %%xmm2          \n"             \
2296    mono("addps          %%xmm2, %%xmm0          \n")            \
2297         "movaps         %%xmm0, (%0, %1)        \n"             \
2298  stereo("movaps         %%xmm2, 0x400(%0, %1)   \n")            \
2299         "add               $16, %0              \n"             \
2300         "jl                 1b                  \n"             \
2301         : "+&r"(i)                                              \
2302         : "r"(samples[0] + len), "r"(matrix)                    \
2303         : XMM_CLOBBERS("%xmm0", "%xmm1", "%xmm2", "%xmm3",      \
2304                       "%xmm4", "%xmm5", "%xmm6", "%xmm7",)      \
2305          "memory"                                               \
2306     );
2307
2308 #define MIX_MISC(stereo)                                        \
2309     __asm__ volatile (                                          \
2310         "1:                                 \n"                 \
2311         "movaps     (%3, %0), %%xmm0        \n"                 \
2312  stereo("movaps       %%xmm0, %%xmm1        \n")                \
2313         "mulps        %%xmm4, %%xmm0        \n"                 \
2314  stereo("mulps        %%xmm5, %%xmm1        \n")                \
2315         "lea    1024(%3, %0), %1            \n"                 \
2316         "mov              %5, %2            \n"                 \
2317         "2:                                 \n"                 \
2318         "movaps         (%1), %%xmm2        \n"                 \
2319  stereo("movaps       %%xmm2, %%xmm3        \n")                \
2320         "mulps      (%4, %2), %%xmm2        \n"                 \
2321  stereo("mulps    16(%4, %2), %%xmm3        \n")                \
2322         "addps        %%xmm2, %%xmm0        \n"                 \
2323  stereo("addps        %%xmm3, %%xmm1        \n")                \
2324         "add           $1024, %1            \n"                 \
2325         "add             $32, %2            \n"                 \
2326         "jl               2b                \n"                 \
2327         "movaps       %%xmm0,     (%3, %0)  \n"                 \
2328  stereo("movaps       %%xmm1, 1024(%3, %0)  \n")                \
2329         "add             $16, %0            \n"                 \
2330         "jl               1b                \n"                 \
2331         : "+&r"(i), "=&r"(j), "=&r"(k)                          \
2332         : "r"(samples[0] + len), "r"(matrix_simd + in_ch),      \
2333           "g"((intptr_t) - 32 * (in_ch - 1))                    \
2334         : "memory"                                              \
2335     );
2336
2337 static void ac3_downmix_sse(float (*samples)[256], float (*matrix)[2],
2338                             int out_ch, int in_ch, int len)
2339 {
2340     int (*matrix_cmp)[2] = (int(*)[2])matrix;
2341     intptr_t i, j, k;
2342
2343     i = -len * sizeof(float);
2344     if (in_ch == 5 && out_ch == 2 &&
2345         !(matrix_cmp[0][1] | matrix_cmp[2][0]   |
2346           matrix_cmp[3][1] | matrix_cmp[4][0]   |
2347           (matrix_cmp[1][0] ^ matrix_cmp[1][1]) |
2348           (matrix_cmp[0][0] ^ matrix_cmp[2][1]))) {
2349         MIX5(IF0, IF1);
2350     } else if (in_ch == 5 && out_ch == 1 &&
2351                matrix_cmp[0][0] == matrix_cmp[2][0] &&
2352                matrix_cmp[3][0] == matrix_cmp[4][0]) {
2353         MIX5(IF1, IF0);
2354     } else {
2355         DECLARE_ALIGNED(16, float, matrix_simd)[AC3_MAX_CHANNELS][2][4];
2356         j = 2 * in_ch * sizeof(float);
2357         __asm__ volatile (
2358             "1:                                 \n"
2359             "sub             $8, %0             \n"
2360             "movss     (%2, %0), %%xmm4         \n"
2361             "movss    4(%2, %0), %%xmm5         \n"
2362             "shufps          $0, %%xmm4, %%xmm4 \n"
2363             "shufps          $0, %%xmm5, %%xmm5 \n"
2364             "movaps      %%xmm4,   (%1, %0, 4)  \n"
2365             "movaps      %%xmm5, 16(%1, %0, 4)  \n"
2366             "jg              1b                 \n"
2367             : "+&r"(j)
2368             : "r"(matrix_simd), "r"(matrix)
2369             : "memory"
2370         );
2371         if (out_ch == 2) {
2372             MIX_MISC(IF1);
2373         } else {
2374             MIX_MISC(IF0);
2375         }
2376     }
2377 }
2378
2379 #if HAVE_6REGS
2380 static void vector_fmul_window_3dnow2(float *dst, const float *src0,
2381                                       const float *src1, const float *win,
2382                                       int len)
2383 {
2384     x86_reg i = -len * 4;
2385     x86_reg j =  len * 4 - 8;
2386     __asm__ volatile (
2387         "1:                             \n"
2388         "pswapd (%5, %1), %%mm1         \n"
2389         "movq   (%5, %0), %%mm0         \n"
2390         "pswapd (%4, %1), %%mm5         \n"
2391         "movq   (%3, %0), %%mm4         \n"
2392         "movq      %%mm0, %%mm2         \n"
2393         "movq      %%mm1, %%mm3         \n"
2394         "pfmul     %%mm4, %%mm2         \n" // src0[len + i] * win[len + i]
2395         "pfmul     %%mm5, %%mm3         \n" // src1[j]       * win[len + j]
2396         "pfmul     %%mm4, %%mm1         \n" // src0[len + i] * win[len + j]
2397         "pfmul     %%mm5, %%mm0         \n" // src1[j]       * win[len + i]
2398         "pfadd     %%mm3, %%mm2         \n"
2399         "pfsub     %%mm0, %%mm1         \n"
2400         "pswapd    %%mm2, %%mm2         \n"
2401         "movq      %%mm1, (%2, %0)      \n"
2402         "movq      %%mm2, (%2, %1)      \n"
2403         "sub          $8, %1            \n"
2404         "add          $8, %0            \n"
2405         "jl           1b                \n"
2406         "femms                          \n"
2407         : "+r"(i), "+r"(j)
2408         : "r"(dst + len), "r"(src0 + len), "r"(src1), "r"(win + len)
2409     );
2410 }
2411
2412 static void vector_fmul_window_sse(float *dst, const float *src0,
2413                                    const float *src1, const float *win, int len)
2414 {
2415     x86_reg i = -len * 4;
2416     x86_reg j =  len * 4 - 16;
2417     __asm__ volatile (
2418         "1:                             \n"
2419         "movaps      (%5, %1), %%xmm1   \n"
2420         "movaps      (%5, %0), %%xmm0   \n"
2421         "movaps      (%4, %1), %%xmm5   \n"
2422         "movaps      (%3, %0), %%xmm4   \n"
2423         "shufps $0x1b, %%xmm1, %%xmm1   \n"
2424         "shufps $0x1b, %%xmm5, %%xmm5   \n"
2425         "movaps        %%xmm0, %%xmm2   \n"
2426         "movaps        %%xmm1, %%xmm3   \n"
2427         "mulps         %%xmm4, %%xmm2   \n" // src0[len + i] * win[len + i]
2428         "mulps         %%xmm5, %%xmm3   \n" // src1[j]       * win[len + j]
2429         "mulps         %%xmm4, %%xmm1   \n" // src0[len + i] * win[len + j]
2430         "mulps         %%xmm5, %%xmm0   \n" // src1[j]       * win[len + i]
2431         "addps         %%xmm3, %%xmm2   \n"
2432         "subps         %%xmm0, %%xmm1   \n"
2433         "shufps $0x1b, %%xmm2, %%xmm2   \n"
2434         "movaps        %%xmm1, (%2, %0) \n"
2435         "movaps        %%xmm2, (%2, %1) \n"
2436         "sub              $16, %1       \n"
2437         "add              $16, %0       \n"
2438         "jl                1b           \n"
2439         : "+r"(i), "+r"(j)
2440         : "r"(dst + len), "r"(src0 + len), "r"(src1), "r"(win + len)
2441     );
2442 }
2443 #endif /* HAVE_6REGS */
2444
2445 static void vector_clipf_sse(float *dst, const float *src,
2446                              float min, float max, int len)
2447 {
2448     x86_reg i = (len - 16) * 4;
2449     __asm__ volatile (
2450         "movss          %3, %%xmm4      \n\t"
2451         "movss          %4, %%xmm5      \n\t"
2452         "shufps $0, %%xmm4, %%xmm4      \n\t"
2453         "shufps $0, %%xmm5, %%xmm5      \n\t"
2454         "1:                             \n\t"
2455         "movaps   (%2, %0), %%xmm0      \n\t" // 3/1 on intel
2456         "movaps 16(%2, %0), %%xmm1      \n\t"
2457         "movaps 32(%2, %0), %%xmm2      \n\t"
2458         "movaps 48(%2, %0), %%xmm3      \n\t"
2459         "maxps      %%xmm4, %%xmm0      \n\t"
2460         "maxps      %%xmm4, %%xmm1      \n\t"
2461         "maxps      %%xmm4, %%xmm2      \n\t"
2462         "maxps      %%xmm4, %%xmm3      \n\t"
2463         "minps      %%xmm5, %%xmm0      \n\t"
2464         "minps      %%xmm5, %%xmm1      \n\t"
2465         "minps      %%xmm5, %%xmm2      \n\t"
2466         "minps      %%xmm5, %%xmm3      \n\t"
2467         "movaps     %%xmm0,   (%1, %0)  \n\t"
2468         "movaps     %%xmm1, 16(%1, %0)  \n\t"
2469         "movaps     %%xmm2, 32(%1, %0)  \n\t"
2470         "movaps     %%xmm3, 48(%1, %0)  \n\t"
2471         "sub           $64, %0          \n\t"
2472         "jge            1b              \n\t"
2473         : "+&r"(i)
2474         : "r"(dst), "r"(src), "m"(min), "m"(max)
2475         : "memory"
2476     );
2477 }
2478
2479 int32_t ff_scalarproduct_int16_mmx2(const int16_t *v1, const int16_t *v2,
2480                                     int order);
2481 int32_t ff_scalarproduct_int16_sse2(const int16_t *v1, const int16_t *v2,
2482                                     int order);
2483 int32_t ff_scalarproduct_and_madd_int16_mmx2(int16_t *v1, const int16_t *v2,
2484                                              const int16_t *v3,
2485                                              int order, int mul);
2486 int32_t ff_scalarproduct_and_madd_int16_sse2(int16_t *v1, const int16_t *v2,
2487                                              const int16_t *v3,
2488                                              int order, int mul);
2489 int32_t ff_scalarproduct_and_madd_int16_ssse3(int16_t *v1, const int16_t *v2,
2490                                               const int16_t *v3,
2491                                               int order, int mul);
2492
2493 void ff_apply_window_int16_mmxext    (int16_t *output, const int16_t *input,
2494                                       const int16_t *window, unsigned int len);
2495 void ff_apply_window_int16_mmxext_ba (int16_t *output, const int16_t *input,
2496                                       const int16_t *window, unsigned int len);
2497 void ff_apply_window_int16_sse2      (int16_t *output, const int16_t *input,
2498                                       const int16_t *window, unsigned int len);
2499 void ff_apply_window_int16_sse2_ba   (int16_t *output, const int16_t *input,
2500                                       const int16_t *window, unsigned int len);
2501 void ff_apply_window_int16_ssse3     (int16_t *output, const int16_t *input,
2502                                       const int16_t *window, unsigned int len);
2503 void ff_apply_window_int16_ssse3_atom(int16_t *output, const int16_t *input,
2504                                       const int16_t *window, unsigned int len);
2505
2506 void ff_bswap32_buf_ssse3(uint32_t *dst, const uint32_t *src, int w);
2507 void ff_bswap32_buf_sse2(uint32_t *dst, const uint32_t *src, int w);
2508
2509 void ff_add_hfyu_median_prediction_mmx2(uint8_t *dst, const uint8_t *top,
2510                                         const uint8_t *diff, int w,
2511                                         int *left, int *left_top);
2512 int  ff_add_hfyu_left_prediction_ssse3(uint8_t *dst, const uint8_t *src,
2513                                        int w, int left);
2514 int  ff_add_hfyu_left_prediction_sse4(uint8_t *dst, const uint8_t *src,
2515                                       int w, int left);
2516
2517 float ff_scalarproduct_float_sse(const float *v1, const float *v2, int order);
2518
2519 void ff_vector_fmul_reverse_sse(float *dst, const float *src0,
2520                                 const float *src1, int len);
2521 void ff_vector_fmul_reverse_avx(float *dst, const float *src0,
2522                                 const float *src1, int len);
2523
2524 void ff_vector_fmul_add_sse(float *dst, const float *src0, const float *src1,
2525                             const float *src2, int len);
2526 void ff_vector_fmul_add_avx(float *dst, const float *src0, const float *src1,
2527                             const float *src2, int len);
2528
2529 void ff_vector_clip_int32_mmx     (int32_t *dst, const int32_t *src,
2530                                    int32_t min, int32_t max, unsigned int len);
2531 void ff_vector_clip_int32_sse2    (int32_t *dst, const int32_t *src,
2532                                    int32_t min, int32_t max, unsigned int len);
2533 void ff_vector_clip_int32_int_sse2(int32_t *dst, const int32_t *src,
2534                                    int32_t min, int32_t max, unsigned int len);
2535 void ff_vector_clip_int32_sse4    (int32_t *dst, const int32_t *src,
2536                                    int32_t min, int32_t max, unsigned int len);
2537
2538 extern void ff_butterflies_float_interleave_sse(float *dst, const float *src0,
2539                                                 const float *src1, int len);
2540 extern void ff_butterflies_float_interleave_avx(float *dst, const float *src0,
2541                                                 const float *src1, int len);
2542
2543 #define SET_QPEL_FUNCS(PFX, IDX, SIZE, CPU, PREFIX)                          \
2544     do {                                                                     \
2545     c->PFX ## _pixels_tab[IDX][ 0] = PREFIX ## PFX ## SIZE ## _mc00_ ## CPU; \
2546     c->PFX ## _pixels_tab[IDX][ 1] = PREFIX ## PFX ## SIZE ## _mc10_ ## CPU; \
2547     c->PFX ## _pixels_tab[IDX][ 2] = PREFIX ## PFX ## SIZE ## _mc20_ ## CPU; \
2548     c->PFX ## _pixels_tab[IDX][ 3] = PREFIX ## PFX ## SIZE ## _mc30_ ## CPU; \
2549     c->PFX ## _pixels_tab[IDX][ 4] = PREFIX ## PFX ## SIZE ## _mc01_ ## CPU; \
2550     c->PFX ## _pixels_tab[IDX][ 5] = PREFIX ## PFX ## SIZE ## _mc11_ ## CPU; \
2551     c->PFX ## _pixels_tab[IDX][ 6] = PREFIX ## PFX ## SIZE ## _mc21_ ## CPU; \
2552     c->PFX ## _pixels_tab[IDX][ 7] = PREFIX ## PFX ## SIZE ## _mc31_ ## CPU; \
2553     c->PFX ## _pixels_tab[IDX][ 8] = PREFIX ## PFX ## SIZE ## _mc02_ ## CPU; \
2554     c->PFX ## _pixels_tab[IDX][ 9] = PREFIX ## PFX ## SIZE ## _mc12_ ## CPU; \
2555     c->PFX ## _pixels_tab[IDX][10] = PREFIX ## PFX ## SIZE ## _mc22_ ## CPU; \
2556     c->PFX ## _pixels_tab[IDX][11] = PREFIX ## PFX ## SIZE ## _mc32_ ## CPU; \
2557     c->PFX ## _pixels_tab[IDX][12] = PREFIX ## PFX ## SIZE ## _mc03_ ## CPU; \
2558     c->PFX ## _pixels_tab[IDX][13] = PREFIX ## PFX ## SIZE ## _mc13_ ## CPU; \