2 * MMX optimized DSP utils
3 * Copyright (c) 2000, 2001 Fabrice Bellard.
4 * Copyright (c) 2002-2004 Michael Niedermayer <michaelni@gmx.at>
6 * This library is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU Lesser General Public
8 * License as published by the Free Software Foundation; either
9 * version 2 of the License, or (at your option) any later version.
11 * This library is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
14 * Lesser General Public License for more details.
16 * You should have received a copy of the GNU Lesser General Public
17 * License along with this library; if not, write to the Free Software
18 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
20 * MMX optimization by Nick Kurshev <nickols_k@mail.ru>
23 #include "../dsputil.h"
24 #include "../simple_idct.h"
25 #include "../mpegvideo.h"
32 extern const uint8_t ff_h263_loop_filter_strength[32];
33 extern void ff_idct_xvid_mmx(short *block);
34 extern void ff_idct_xvid_mmx2(short *block);
36 int mm_flags; /* multimedia extension flags */
38 /* pixel operations */
39 static const uint64_t mm_bone attribute_used __attribute__ ((aligned(8))) = 0x0101010101010101ULL;
40 static const uint64_t mm_wone attribute_used __attribute__ ((aligned(8))) = 0x0001000100010001ULL;
41 static const uint64_t mm_wtwo attribute_used __attribute__ ((aligned(8))) = 0x0002000200020002ULL;
43 static const uint64_t ff_pdw_80000000[2] attribute_used __attribute__ ((aligned(16))) =
44 {0x8000000080000000ULL, 0x8000000080000000ULL};
46 static const uint64_t ff_pw_20 attribute_used __attribute__ ((aligned(8))) = 0x0014001400140014ULL;
47 static const uint64_t ff_pw_3 attribute_used __attribute__ ((aligned(8))) = 0x0003000300030003ULL;
48 static const uint64_t ff_pw_4 attribute_used __attribute__ ((aligned(8))) = 0x0004000400040004ULL;
49 static const uint64_t ff_pw_5 attribute_used __attribute__ ((aligned(8))) = 0x0005000500050005ULL;
50 static const uint64_t ff_pw_8 attribute_used __attribute__ ((aligned(8))) = 0x0008000800080008ULL;
51 static const uint64_t ff_pw_16 attribute_used __attribute__ ((aligned(8))) = 0x0010001000100010ULL;
52 static const uint64_t ff_pw_32 attribute_used __attribute__ ((aligned(8))) = 0x0020002000200020ULL;
53 static const uint64_t ff_pw_64 attribute_used __attribute__ ((aligned(8))) = 0x0040004000400040ULL;
54 static const uint64_t ff_pw_15 attribute_used __attribute__ ((aligned(8))) = 0x000F000F000F000FULL;
56 static const uint64_t ff_pb_3F attribute_used __attribute__ ((aligned(8))) = 0x3F3F3F3F3F3F3F3FULL;
57 static const uint64_t ff_pb_FC attribute_used __attribute__ ((aligned(8))) = 0xFCFCFCFCFCFCFCFCULL;
59 #define JUMPALIGN() __asm __volatile (".balign 8"::)
60 #define MOVQ_ZERO(regd) __asm __volatile ("pxor %%" #regd ", %%" #regd ::)
62 #define MOVQ_WONE(regd) \
64 "pcmpeqd %%" #regd ", %%" #regd " \n\t" \
65 "psrlw $15, %%" #regd ::)
67 #define MOVQ_BFE(regd) \
69 "pcmpeqd %%" #regd ", %%" #regd " \n\t"\
70 "paddb %%" #regd ", %%" #regd " \n\t" ::)
73 #define MOVQ_BONE(regd) __asm __volatile ("movq %0, %%" #regd " \n\t" ::"m"(mm_bone))
74 #define MOVQ_WTWO(regd) __asm __volatile ("movq %0, %%" #regd " \n\t" ::"m"(mm_wtwo))
76 // for shared library it's better to use this way for accessing constants
78 #define MOVQ_BONE(regd) \
80 "pcmpeqd %%" #regd ", %%" #regd " \n\t" \
81 "psrlw $15, %%" #regd " \n\t" \
82 "packuswb %%" #regd ", %%" #regd " \n\t" ::)
84 #define MOVQ_WTWO(regd) \
86 "pcmpeqd %%" #regd ", %%" #regd " \n\t" \
87 "psrlw $15, %%" #regd " \n\t" \
88 "psllw $1, %%" #regd " \n\t"::)
92 // using regr as temporary and for the output result
93 // first argument is unmodifed and second is trashed
94 // regfe is supposed to contain 0xfefefefefefefefe
95 #define PAVGB_MMX_NO_RND(rega, regb, regr, regfe) \
96 "movq " #rega ", " #regr " \n\t"\
97 "pand " #regb ", " #regr " \n\t"\
98 "pxor " #rega ", " #regb " \n\t"\
99 "pand " #regfe "," #regb " \n\t"\
100 "psrlq $1, " #regb " \n\t"\
101 "paddb " #regb ", " #regr " \n\t"
103 #define PAVGB_MMX(rega, regb, regr, regfe) \
104 "movq " #rega ", " #regr " \n\t"\
105 "por " #regb ", " #regr " \n\t"\
106 "pxor " #rega ", " #regb " \n\t"\
107 "pand " #regfe "," #regb " \n\t"\
108 "psrlq $1, " #regb " \n\t"\
109 "psubb " #regb ", " #regr " \n\t"
111 // mm6 is supposed to contain 0xfefefefefefefefe
112 #define PAVGBP_MMX_NO_RND(rega, regb, regr, regc, regd, regp) \
113 "movq " #rega ", " #regr " \n\t"\
114 "movq " #regc ", " #regp " \n\t"\
115 "pand " #regb ", " #regr " \n\t"\
116 "pand " #regd ", " #regp " \n\t"\
117 "pxor " #rega ", " #regb " \n\t"\
118 "pxor " #regc ", " #regd " \n\t"\
119 "pand %%mm6, " #regb " \n\t"\
120 "pand %%mm6, " #regd " \n\t"\
121 "psrlq $1, " #regb " \n\t"\
122 "psrlq $1, " #regd " \n\t"\
123 "paddb " #regb ", " #regr " \n\t"\
124 "paddb " #regd ", " #regp " \n\t"
126 #define PAVGBP_MMX(rega, regb, regr, regc, regd, regp) \
127 "movq " #rega ", " #regr " \n\t"\
128 "movq " #regc ", " #regp " \n\t"\
129 "por " #regb ", " #regr " \n\t"\
130 "por " #regd ", " #regp " \n\t"\
131 "pxor " #rega ", " #regb " \n\t"\
132 "pxor " #regc ", " #regd " \n\t"\
133 "pand %%mm6, " #regb " \n\t"\
134 "pand %%mm6, " #regd " \n\t"\
135 "psrlq $1, " #regd " \n\t"\
136 "psrlq $1, " #regb " \n\t"\
137 "psubb " #regb ", " #regr " \n\t"\
138 "psubb " #regd ", " #regp " \n\t"
140 /***********************************/
141 /* MMX no rounding */
142 #define DEF(x, y) x ## _no_rnd_ ## y ##_mmx
143 #define SET_RND MOVQ_WONE
144 #define PAVGBP(a, b, c, d, e, f) PAVGBP_MMX_NO_RND(a, b, c, d, e, f)
145 #define PAVGB(a, b, c, e) PAVGB_MMX_NO_RND(a, b, c, e)
147 #include "dsputil_mmx_rnd.h"
153 /***********************************/
156 #define DEF(x, y) x ## _ ## y ##_mmx
157 #define SET_RND MOVQ_WTWO
158 #define PAVGBP(a, b, c, d, e, f) PAVGBP_MMX(a, b, c, d, e, f)
159 #define PAVGB(a, b, c, e) PAVGB_MMX(a, b, c, e)
161 #include "dsputil_mmx_rnd.h"
168 /***********************************/
171 #define DEF(x) x ## _3dnow
172 /* for Athlons PAVGUSB is prefered */
173 #define PAVGB "pavgusb"
175 #include "dsputil_mmx_avg.h"
180 /***********************************/
183 #define DEF(x) x ## _mmx2
185 /* Introduced only in MMX2 set */
186 #define PAVGB "pavgb"
188 #include "dsputil_mmx_avg.h"
193 #define SBUTTERFLY(a,b,t,n)\
194 "movq " #a ", " #t " \n\t" /* abcd */\
195 "punpckl" #n " " #b ", " #a " \n\t" /* aebf */\
196 "punpckh" #n " " #b ", " #t " \n\t" /* cgdh */\
198 /***********************************/
201 #ifdef CONFIG_ENCODERS
202 static void get_pixels_mmx(DCTELEM *block, const uint8_t *pixels, int line_size)
205 "mov $-128, %%"REG_a" \n\t"
206 "pxor %%mm7, %%mm7 \n\t"
209 "movq (%0), %%mm0 \n\t"
210 "movq (%0, %2), %%mm2 \n\t"
211 "movq %%mm0, %%mm1 \n\t"
212 "movq %%mm2, %%mm3 \n\t"
213 "punpcklbw %%mm7, %%mm0 \n\t"
214 "punpckhbw %%mm7, %%mm1 \n\t"
215 "punpcklbw %%mm7, %%mm2 \n\t"
216 "punpckhbw %%mm7, %%mm3 \n\t"
217 "movq %%mm0, (%1, %%"REG_a") \n\t"
218 "movq %%mm1, 8(%1, %%"REG_a") \n\t"
219 "movq %%mm2, 16(%1, %%"REG_a") \n\t"
220 "movq %%mm3, 24(%1, %%"REG_a") \n\t"
222 "add $32, %%"REG_a" \n\t"
225 : "r" (block+64), "r" ((long)line_size), "r" ((long)line_size*2)
230 static inline void diff_pixels_mmx(DCTELEM *block, const uint8_t *s1, const uint8_t *s2, int stride)
233 "pxor %%mm7, %%mm7 \n\t"
234 "mov $-128, %%"REG_a" \n\t"
237 "movq (%0), %%mm0 \n\t"
238 "movq (%1), %%mm2 \n\t"
239 "movq %%mm0, %%mm1 \n\t"
240 "movq %%mm2, %%mm3 \n\t"
241 "punpcklbw %%mm7, %%mm0 \n\t"
242 "punpckhbw %%mm7, %%mm1 \n\t"
243 "punpcklbw %%mm7, %%mm2 \n\t"
244 "punpckhbw %%mm7, %%mm3 \n\t"
245 "psubw %%mm2, %%mm0 \n\t"
246 "psubw %%mm3, %%mm1 \n\t"
247 "movq %%mm0, (%2, %%"REG_a") \n\t"
248 "movq %%mm1, 8(%2, %%"REG_a") \n\t"
251 "add $16, %%"REG_a" \n\t"
253 : "+r" (s1), "+r" (s2)
254 : "r" (block+64), "r" ((long)stride)
258 #endif //CONFIG_ENCODERS
260 void put_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size)
265 /* read the pixels */
270 "movq %3, %%mm0 \n\t"
271 "movq 8%3, %%mm1 \n\t"
272 "movq 16%3, %%mm2 \n\t"
273 "movq 24%3, %%mm3 \n\t"
274 "movq 32%3, %%mm4 \n\t"
275 "movq 40%3, %%mm5 \n\t"
276 "movq 48%3, %%mm6 \n\t"
277 "movq 56%3, %%mm7 \n\t"
278 "packuswb %%mm1, %%mm0 \n\t"
279 "packuswb %%mm3, %%mm2 \n\t"
280 "packuswb %%mm5, %%mm4 \n\t"
281 "packuswb %%mm7, %%mm6 \n\t"
282 "movq %%mm0, (%0) \n\t"
283 "movq %%mm2, (%0, %1) \n\t"
284 "movq %%mm4, (%0, %1, 2) \n\t"
285 "movq %%mm6, (%0, %2) \n\t"
286 ::"r" (pix), "r" ((long)line_size), "r" ((long)line_size*3), "m"(*p)
291 // if here would be an exact copy of the code above
292 // compiler would generate some very strange code
295 "movq (%3), %%mm0 \n\t"
296 "movq 8(%3), %%mm1 \n\t"
297 "movq 16(%3), %%mm2 \n\t"
298 "movq 24(%3), %%mm3 \n\t"
299 "movq 32(%3), %%mm4 \n\t"
300 "movq 40(%3), %%mm5 \n\t"
301 "movq 48(%3), %%mm6 \n\t"
302 "movq 56(%3), %%mm7 \n\t"
303 "packuswb %%mm1, %%mm0 \n\t"
304 "packuswb %%mm3, %%mm2 \n\t"
305 "packuswb %%mm5, %%mm4 \n\t"
306 "packuswb %%mm7, %%mm6 \n\t"
307 "movq %%mm0, (%0) \n\t"
308 "movq %%mm2, (%0, %1) \n\t"
309 "movq %%mm4, (%0, %1, 2) \n\t"
310 "movq %%mm6, (%0, %2) \n\t"
311 ::"r" (pix), "r" ((long)line_size), "r" ((long)line_size*3), "r"(p)
315 static DECLARE_ALIGNED_8(const unsigned char, vector128[8]) =
316 { 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80 };
318 void put_signed_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size)
322 movq_m2r(*vector128, mm1);
323 for (i = 0; i < 8; i++) {
324 movq_m2r(*(block), mm0);
325 packsswb_m2r(*(block + 4), mm0);
328 movq_r2m(mm0, *pixels);
333 void add_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size)
339 /* read the pixels */
346 "movq (%2), %%mm0 \n\t"
347 "movq 8(%2), %%mm1 \n\t"
348 "movq 16(%2), %%mm2 \n\t"
349 "movq 24(%2), %%mm3 \n\t"
350 "movq %0, %%mm4 \n\t"
351 "movq %1, %%mm6 \n\t"
352 "movq %%mm4, %%mm5 \n\t"
353 "punpcklbw %%mm7, %%mm4 \n\t"
354 "punpckhbw %%mm7, %%mm5 \n\t"
355 "paddsw %%mm4, %%mm0 \n\t"
356 "paddsw %%mm5, %%mm1 \n\t"
357 "movq %%mm6, %%mm5 \n\t"
358 "punpcklbw %%mm7, %%mm6 \n\t"
359 "punpckhbw %%mm7, %%mm5 \n\t"
360 "paddsw %%mm6, %%mm2 \n\t"
361 "paddsw %%mm5, %%mm3 \n\t"
362 "packuswb %%mm1, %%mm0 \n\t"
363 "packuswb %%mm3, %%mm2 \n\t"
364 "movq %%mm0, %0 \n\t"
365 "movq %%mm2, %1 \n\t"
366 :"+m"(*pix), "+m"(*(pix+line_size))
374 static void put_pixels4_mmx(uint8_t *block, const uint8_t *pixels, int line_size, int h)
377 "lea (%3, %3), %%"REG_a" \n\t"
380 "movd (%1), %%mm0 \n\t"
381 "movd (%1, %3), %%mm1 \n\t"
382 "movd %%mm0, (%2) \n\t"
383 "movd %%mm1, (%2, %3) \n\t"
384 "add %%"REG_a", %1 \n\t"
385 "add %%"REG_a", %2 \n\t"
386 "movd (%1), %%mm0 \n\t"
387 "movd (%1, %3), %%mm1 \n\t"
388 "movd %%mm0, (%2) \n\t"
389 "movd %%mm1, (%2, %3) \n\t"
390 "add %%"REG_a", %1 \n\t"
391 "add %%"REG_a", %2 \n\t"
394 : "+g"(h), "+r" (pixels), "+r" (block)
395 : "r"((long)line_size)
400 static void put_pixels8_mmx(uint8_t *block, const uint8_t *pixels, int line_size, int h)
403 "lea (%3, %3), %%"REG_a" \n\t"
406 "movq (%1), %%mm0 \n\t"
407 "movq (%1, %3), %%mm1 \n\t"
408 "movq %%mm0, (%2) \n\t"
409 "movq %%mm1, (%2, %3) \n\t"
410 "add %%"REG_a", %1 \n\t"
411 "add %%"REG_a", %2 \n\t"
412 "movq (%1), %%mm0 \n\t"
413 "movq (%1, %3), %%mm1 \n\t"
414 "movq %%mm0, (%2) \n\t"
415 "movq %%mm1, (%2, %3) \n\t"
416 "add %%"REG_a", %1 \n\t"
417 "add %%"REG_a", %2 \n\t"
420 : "+g"(h), "+r" (pixels), "+r" (block)
421 : "r"((long)line_size)
426 static void put_pixels16_mmx(uint8_t *block, const uint8_t *pixels, int line_size, int h)
429 "lea (%3, %3), %%"REG_a" \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"
454 : "+g"(h), "+r" (pixels), "+r" (block)
455 : "r"((long)line_size)
460 static void clear_blocks_mmx(DCTELEM *blocks)
463 "pxor %%mm7, %%mm7 \n\t"
464 "mov $-128*6, %%"REG_a" \n\t"
466 "movq %%mm7, (%0, %%"REG_a") \n\t"
467 "movq %%mm7, 8(%0, %%"REG_a") \n\t"
468 "movq %%mm7, 16(%0, %%"REG_a") \n\t"
469 "movq %%mm7, 24(%0, %%"REG_a") \n\t"
470 "add $32, %%"REG_a" \n\t"
472 : : "r" (((uint8_t *)blocks)+128*6)
477 #ifdef CONFIG_ENCODERS
478 static int pix_sum16_mmx(uint8_t * pix, int line_size){
481 long index= -line_size*h;
484 "pxor %%mm7, %%mm7 \n\t"
485 "pxor %%mm6, %%mm6 \n\t"
487 "movq (%2, %1), %%mm0 \n\t"
488 "movq (%2, %1), %%mm1 \n\t"
489 "movq 8(%2, %1), %%mm2 \n\t"
490 "movq 8(%2, %1), %%mm3 \n\t"
491 "punpcklbw %%mm7, %%mm0 \n\t"
492 "punpckhbw %%mm7, %%mm1 \n\t"
493 "punpcklbw %%mm7, %%mm2 \n\t"
494 "punpckhbw %%mm7, %%mm3 \n\t"
495 "paddw %%mm0, %%mm1 \n\t"
496 "paddw %%mm2, %%mm3 \n\t"
497 "paddw %%mm1, %%mm3 \n\t"
498 "paddw %%mm3, %%mm6 \n\t"
501 "movq %%mm6, %%mm5 \n\t"
502 "psrlq $32, %%mm6 \n\t"
503 "paddw %%mm5, %%mm6 \n\t"
504 "movq %%mm6, %%mm5 \n\t"
505 "psrlq $16, %%mm6 \n\t"
506 "paddw %%mm5, %%mm6 \n\t"
507 "movd %%mm6, %0 \n\t"
508 "andl $0xFFFF, %0 \n\t"
509 : "=&r" (sum), "+r" (index)
510 : "r" (pix - index), "r" ((long)line_size)
515 #endif //CONFIG_ENCODERS
517 static void add_bytes_mmx(uint8_t *dst, uint8_t *src, int w){
521 "movq (%1, %0), %%mm0 \n\t"
522 "movq (%2, %0), %%mm1 \n\t"
523 "paddb %%mm0, %%mm1 \n\t"
524 "movq %%mm1, (%2, %0) \n\t"
525 "movq 8(%1, %0), %%mm0 \n\t"
526 "movq 8(%2, %0), %%mm1 \n\t"
527 "paddb %%mm0, %%mm1 \n\t"
528 "movq %%mm1, 8(%2, %0) \n\t"
533 : "r"(src), "r"(dst), "r"((long)w-15)
536 dst[i+0] += src[i+0];
539 #define H263_LOOP_FILTER \
540 "pxor %%mm7, %%mm7 \n\t"\
541 "movq %0, %%mm0 \n\t"\
542 "movq %0, %%mm1 \n\t"\
543 "movq %3, %%mm2 \n\t"\
544 "movq %3, %%mm3 \n\t"\
545 "punpcklbw %%mm7, %%mm0 \n\t"\
546 "punpckhbw %%mm7, %%mm1 \n\t"\
547 "punpcklbw %%mm7, %%mm2 \n\t"\
548 "punpckhbw %%mm7, %%mm3 \n\t"\
549 "psubw %%mm2, %%mm0 \n\t"\
550 "psubw %%mm3, %%mm1 \n\t"\
551 "movq %1, %%mm2 \n\t"\
552 "movq %1, %%mm3 \n\t"\
553 "movq %2, %%mm4 \n\t"\
554 "movq %2, %%mm5 \n\t"\
555 "punpcklbw %%mm7, %%mm2 \n\t"\
556 "punpckhbw %%mm7, %%mm3 \n\t"\
557 "punpcklbw %%mm7, %%mm4 \n\t"\
558 "punpckhbw %%mm7, %%mm5 \n\t"\
559 "psubw %%mm2, %%mm4 \n\t"\
560 "psubw %%mm3, %%mm5 \n\t"\
561 "psllw $2, %%mm4 \n\t"\
562 "psllw $2, %%mm5 \n\t"\
563 "paddw %%mm0, %%mm4 \n\t"\
564 "paddw %%mm1, %%mm5 \n\t"\
565 "pxor %%mm6, %%mm6 \n\t"\
566 "pcmpgtw %%mm4, %%mm6 \n\t"\
567 "pcmpgtw %%mm5, %%mm7 \n\t"\
568 "pxor %%mm6, %%mm4 \n\t"\
569 "pxor %%mm7, %%mm5 \n\t"\
570 "psubw %%mm6, %%mm4 \n\t"\
571 "psubw %%mm7, %%mm5 \n\t"\
572 "psrlw $3, %%mm4 \n\t"\
573 "psrlw $3, %%mm5 \n\t"\
574 "packuswb %%mm5, %%mm4 \n\t"\
575 "packsswb %%mm7, %%mm6 \n\t"\
576 "pxor %%mm7, %%mm7 \n\t"\
577 "movd %4, %%mm2 \n\t"\
578 "punpcklbw %%mm2, %%mm2 \n\t"\
579 "punpcklbw %%mm2, %%mm2 \n\t"\
580 "punpcklbw %%mm2, %%mm2 \n\t"\
581 "psubusb %%mm4, %%mm2 \n\t"\
582 "movq %%mm2, %%mm3 \n\t"\
583 "psubusb %%mm4, %%mm3 \n\t"\
584 "psubb %%mm3, %%mm2 \n\t"\
585 "movq %1, %%mm3 \n\t"\
586 "movq %2, %%mm4 \n\t"\
587 "pxor %%mm6, %%mm3 \n\t"\
588 "pxor %%mm6, %%mm4 \n\t"\
589 "paddusb %%mm2, %%mm3 \n\t"\
590 "psubusb %%mm2, %%mm4 \n\t"\
591 "pxor %%mm6, %%mm3 \n\t"\
592 "pxor %%mm6, %%mm4 \n\t"\
593 "paddusb %%mm2, %%mm2 \n\t"\
594 "packsswb %%mm1, %%mm0 \n\t"\
595 "pcmpgtb %%mm0, %%mm7 \n\t"\
596 "pxor %%mm7, %%mm0 \n\t"\
597 "psubb %%mm7, %%mm0 \n\t"\
598 "movq %%mm0, %%mm1 \n\t"\
599 "psubusb %%mm2, %%mm0 \n\t"\
600 "psubb %%mm0, %%mm1 \n\t"\
601 "pand %5, %%mm1 \n\t"\
602 "psrlw $2, %%mm1 \n\t"\
603 "pxor %%mm7, %%mm1 \n\t"\
604 "psubb %%mm7, %%mm1 \n\t"\
605 "movq %0, %%mm5 \n\t"\
606 "movq %3, %%mm6 \n\t"\
607 "psubb %%mm1, %%mm5 \n\t"\
608 "paddb %%mm1, %%mm6 \n\t"
610 static void h263_v_loop_filter_mmx(uint8_t *src, int stride, int qscale){
611 const int strength= ff_h263_loop_filter_strength[qscale];
617 "movq %%mm3, %1 \n\t"
618 "movq %%mm4, %2 \n\t"
619 "movq %%mm5, %0 \n\t"
620 "movq %%mm6, %3 \n\t"
621 : "+m" (*(uint64_t*)(src - 2*stride)),
622 "+m" (*(uint64_t*)(src - 1*stride)),
623 "+m" (*(uint64_t*)(src + 0*stride)),
624 "+m" (*(uint64_t*)(src + 1*stride))
625 : "g" (2*strength), "m"(ff_pb_FC)
629 static inline void transpose4x4(uint8_t *dst, uint8_t *src, int dst_stride, int src_stride){
630 asm volatile( //FIXME could save 1 instruction if done as 8x4 ...
631 "movd %4, %%mm0 \n\t"
632 "movd %5, %%mm1 \n\t"
633 "movd %6, %%mm2 \n\t"
634 "movd %7, %%mm3 \n\t"
635 "punpcklbw %%mm1, %%mm0 \n\t"
636 "punpcklbw %%mm3, %%mm2 \n\t"
637 "movq %%mm0, %%mm1 \n\t"
638 "punpcklwd %%mm2, %%mm0 \n\t"
639 "punpckhwd %%mm2, %%mm1 \n\t"
640 "movd %%mm0, %0 \n\t"
641 "punpckhdq %%mm0, %%mm0 \n\t"
642 "movd %%mm0, %1 \n\t"
643 "movd %%mm1, %2 \n\t"
644 "punpckhdq %%mm1, %%mm1 \n\t"
645 "movd %%mm1, %3 \n\t"
647 : "=m" (*(uint32_t*)(dst + 0*dst_stride)),
648 "=m" (*(uint32_t*)(dst + 1*dst_stride)),
649 "=m" (*(uint32_t*)(dst + 2*dst_stride)),
650 "=m" (*(uint32_t*)(dst + 3*dst_stride))
651 : "m" (*(uint32_t*)(src + 0*src_stride)),
652 "m" (*(uint32_t*)(src + 1*src_stride)),
653 "m" (*(uint32_t*)(src + 2*src_stride)),
654 "m" (*(uint32_t*)(src + 3*src_stride))
658 static void h263_h_loop_filter_mmx(uint8_t *src, int stride, int qscale){
659 const int strength= ff_h263_loop_filter_strength[qscale];
660 uint64_t temp[4] __attribute__ ((aligned(8)));
661 uint8_t *btemp= (uint8_t*)temp;
665 transpose4x4(btemp , src , 8, stride);
666 transpose4x4(btemp+4, src + 4*stride, 8, stride);
668 H263_LOOP_FILTER // 5 3 4 6
674 : "g" (2*strength), "m"(ff_pb_FC)
678 "movq %%mm5, %%mm1 \n\t"
679 "movq %%mm4, %%mm0 \n\t"
680 "punpcklbw %%mm3, %%mm5 \n\t"
681 "punpcklbw %%mm6, %%mm4 \n\t"
682 "punpckhbw %%mm3, %%mm1 \n\t"
683 "punpckhbw %%mm6, %%mm0 \n\t"
684 "movq %%mm5, %%mm3 \n\t"
685 "movq %%mm1, %%mm6 \n\t"
686 "punpcklwd %%mm4, %%mm5 \n\t"
687 "punpcklwd %%mm0, %%mm1 \n\t"
688 "punpckhwd %%mm4, %%mm3 \n\t"
689 "punpckhwd %%mm0, %%mm6 \n\t"
690 "movd %%mm5, (%0) \n\t"
691 "punpckhdq %%mm5, %%mm5 \n\t"
692 "movd %%mm5, (%0,%2) \n\t"
693 "movd %%mm3, (%0,%2,2) \n\t"
694 "punpckhdq %%mm3, %%mm3 \n\t"
695 "movd %%mm3, (%0,%3) \n\t"
696 "movd %%mm1, (%1) \n\t"
697 "punpckhdq %%mm1, %%mm1 \n\t"
698 "movd %%mm1, (%1,%2) \n\t"
699 "movd %%mm6, (%1,%2,2) \n\t"
700 "punpckhdq %%mm6, %%mm6 \n\t"
701 "movd %%mm6, (%1,%3) \n\t"
703 "r" (src + 4*stride),
704 "r" ((long) stride ),
705 "r" ((long)(3*stride))
709 #ifdef CONFIG_ENCODERS
710 static int pix_norm1_mmx(uint8_t *pix, int line_size) {
717 "movq (%0),%%mm2\n" /* mm2 = pix[0-7] */
718 "movq 8(%0),%%mm3\n" /* mm3 = pix[8-15] */
720 "movq %%mm2,%%mm1\n" /* mm1 = mm2 = pix[0-7] */
722 "punpckhbw %%mm0,%%mm1\n" /* mm1 = [pix4-7] */
723 "punpcklbw %%mm0,%%mm2\n" /* mm2 = [pix0-3] */
725 "movq %%mm3,%%mm4\n" /* mm4 = mm3 = pix[8-15] */
726 "punpckhbw %%mm0,%%mm3\n" /* mm3 = [pix12-15] */
727 "punpcklbw %%mm0,%%mm4\n" /* mm4 = [pix8-11] */
729 "pmaddwd %%mm1,%%mm1\n" /* mm1 = (pix0^2+pix1^2,pix2^2+pix3^2) */
730 "pmaddwd %%mm2,%%mm2\n" /* mm2 = (pix4^2+pix5^2,pix6^2+pix7^2) */
732 "pmaddwd %%mm3,%%mm3\n"
733 "pmaddwd %%mm4,%%mm4\n"
735 "paddd %%mm1,%%mm2\n" /* mm2 = (pix0^2+pix1^2+pix4^2+pix5^2,
736 pix2^2+pix3^2+pix6^2+pix7^2) */
737 "paddd %%mm3,%%mm4\n"
738 "paddd %%mm2,%%mm7\n"
741 "paddd %%mm4,%%mm7\n"
746 "psrlq $32, %%mm7\n" /* shift hi dword to lo */
747 "paddd %%mm7,%%mm1\n"
749 : "+r" (pix), "=r"(tmp) : "r" ((long)line_size) : "%ecx" );
753 static int sse8_mmx(void *v, uint8_t * pix1, uint8_t * pix2, int line_size, int h) {
758 "pxor %%mm0,%%mm0\n" /* mm0 = 0 */
759 "pxor %%mm7,%%mm7\n" /* mm7 holds the sum */
761 "movq (%0),%%mm1\n" /* mm1 = pix1[0][0-7] */
762 "movq (%1),%%mm2\n" /* mm2 = pix2[0][0-7] */
763 "movq (%0,%3),%%mm3\n" /* mm3 = pix1[1][0-7] */
764 "movq (%1,%3),%%mm4\n" /* mm4 = pix2[1][0-7] */
766 /* todo: mm1-mm2, mm3-mm4 */
767 /* algo: substract mm1 from mm2 with saturation and vice versa */
768 /* OR the results to get absolute difference */
771 "psubusb %%mm2,%%mm1\n"
772 "psubusb %%mm4,%%mm3\n"
773 "psubusb %%mm5,%%mm2\n"
774 "psubusb %%mm6,%%mm4\n"
779 /* now convert to 16-bit vectors so we can square them */
783 "punpckhbw %%mm0,%%mm2\n"
784 "punpckhbw %%mm0,%%mm4\n"
785 "punpcklbw %%mm0,%%mm1\n" /* mm1 now spread over (mm1,mm2) */
786 "punpcklbw %%mm0,%%mm3\n" /* mm4 now spread over (mm3,mm4) */
788 "pmaddwd %%mm2,%%mm2\n"
789 "pmaddwd %%mm4,%%mm4\n"
790 "pmaddwd %%mm1,%%mm1\n"
791 "pmaddwd %%mm3,%%mm3\n"
793 "lea (%0,%3,2), %0\n" /* pix1 += 2*line_size */
794 "lea (%1,%3,2), %1\n" /* pix2 += 2*line_size */
796 "paddd %%mm2,%%mm1\n"
797 "paddd %%mm4,%%mm3\n"
798 "paddd %%mm1,%%mm7\n"
799 "paddd %%mm3,%%mm7\n"
805 "psrlq $32, %%mm7\n" /* shift hi dword to lo */
806 "paddd %%mm7,%%mm1\n"
808 : "+r" (pix1), "+r" (pix2), "=r"(tmp)
809 : "r" ((long)line_size) , "m" (h)
814 static int sse16_mmx(void *v, uint8_t * pix1, uint8_t * pix2, int line_size, int h) {
818 "pxor %%mm0,%%mm0\n" /* mm0 = 0 */
819 "pxor %%mm7,%%mm7\n" /* mm7 holds the sum */
821 "movq (%0),%%mm1\n" /* mm1 = pix1[0-7] */
822 "movq (%1),%%mm2\n" /* mm2 = pix2[0-7] */
823 "movq 8(%0),%%mm3\n" /* mm3 = pix1[8-15] */
824 "movq 8(%1),%%mm4\n" /* mm4 = pix2[8-15] */
826 /* todo: mm1-mm2, mm3-mm4 */
827 /* algo: substract mm1 from mm2 with saturation and vice versa */
828 /* OR the results to get absolute difference */
831 "psubusb %%mm2,%%mm1\n"
832 "psubusb %%mm4,%%mm3\n"
833 "psubusb %%mm5,%%mm2\n"
834 "psubusb %%mm6,%%mm4\n"
839 /* now convert to 16-bit vectors so we can square them */
843 "punpckhbw %%mm0,%%mm2\n"
844 "punpckhbw %%mm0,%%mm4\n"
845 "punpcklbw %%mm0,%%mm1\n" /* mm1 now spread over (mm1,mm2) */
846 "punpcklbw %%mm0,%%mm3\n" /* mm4 now spread over (mm3,mm4) */
848 "pmaddwd %%mm2,%%mm2\n"
849 "pmaddwd %%mm4,%%mm4\n"
850 "pmaddwd %%mm1,%%mm1\n"
851 "pmaddwd %%mm3,%%mm3\n"
856 "paddd %%mm2,%%mm1\n"
857 "paddd %%mm4,%%mm3\n"
858 "paddd %%mm1,%%mm7\n"
859 "paddd %%mm3,%%mm7\n"
865 "psrlq $32, %%mm7\n" /* shift hi dword to lo */
866 "paddd %%mm7,%%mm1\n"
868 : "+r" (pix1), "+r" (pix2), "=r"(tmp)
869 : "r" ((long)line_size) , "m" (h)
874 static int sse16_sse2(void *v, uint8_t * pix1, uint8_t * pix2, int line_size, int h) {
878 "pxor %%xmm0,%%xmm0\n" /* mm0 = 0 */
879 "pxor %%xmm7,%%xmm7\n" /* mm7 holds the sum */
881 "movdqu (%0),%%xmm1\n" /* mm1 = pix1[0][0-15] */
882 "movdqu (%1),%%xmm2\n" /* mm2 = pix2[0][0-15] */
883 "movdqu (%0,%4),%%xmm3\n" /* mm3 = pix1[1][0-15] */
884 "movdqu (%1,%4),%%xmm4\n" /* mm4 = pix2[1][0-15] */
886 /* todo: mm1-mm2, mm3-mm4 */
887 /* algo: substract mm1 from mm2 with saturation and vice versa */
888 /* OR the results to get absolute difference */
889 "movdqa %%xmm1,%%xmm5\n"
890 "movdqa %%xmm3,%%xmm6\n"
891 "psubusb %%xmm2,%%xmm1\n"
892 "psubusb %%xmm4,%%xmm3\n"
893 "psubusb %%xmm5,%%xmm2\n"
894 "psubusb %%xmm6,%%xmm4\n"
896 "por %%xmm1,%%xmm2\n"
897 "por %%xmm3,%%xmm4\n"
899 /* now convert to 16-bit vectors so we can square them */
900 "movdqa %%xmm2,%%xmm1\n"
901 "movdqa %%xmm4,%%xmm3\n"
903 "punpckhbw %%xmm0,%%xmm2\n"
904 "punpckhbw %%xmm0,%%xmm4\n"
905 "punpcklbw %%xmm0,%%xmm1\n" /* mm1 now spread over (mm1,mm2) */
906 "punpcklbw %%xmm0,%%xmm3\n" /* mm4 now spread over (mm3,mm4) */
908 "pmaddwd %%xmm2,%%xmm2\n"
909 "pmaddwd %%xmm4,%%xmm4\n"
910 "pmaddwd %%xmm1,%%xmm1\n"
911 "pmaddwd %%xmm3,%%xmm3\n"
913 "lea (%0,%4,2), %0\n" /* pix1 += 2*line_size */
914 "lea (%1,%4,2), %1\n" /* pix2 += 2*line_size */
916 "paddd %%xmm2,%%xmm1\n"
917 "paddd %%xmm4,%%xmm3\n"
918 "paddd %%xmm1,%%xmm7\n"
919 "paddd %%xmm3,%%xmm7\n"
924 "movdqa %%xmm7,%%xmm1\n"
925 "psrldq $8, %%xmm7\n" /* shift hi qword to lo */
926 "paddd %%xmm1,%%xmm7\n"
927 "movdqa %%xmm7,%%xmm1\n"
928 "psrldq $4, %%xmm7\n" /* shift hi dword to lo */
929 "paddd %%xmm1,%%xmm7\n"
931 : "+r" (pix1), "+r" (pix2), "+r"(h), "=r"(tmp)
932 : "r" ((long)line_size));
936 static int hf_noise8_mmx(uint8_t * pix1, int line_size, int h) {
944 "movq %%mm0, %%mm1\n"
948 "movq %%mm0, %%mm2\n"
949 "movq %%mm1, %%mm3\n"
950 "punpcklbw %%mm7,%%mm0\n"
951 "punpcklbw %%mm7,%%mm1\n"
952 "punpckhbw %%mm7,%%mm2\n"
953 "punpckhbw %%mm7,%%mm3\n"
954 "psubw %%mm1, %%mm0\n"
955 "psubw %%mm3, %%mm2\n"
960 "movq %%mm4, %%mm1\n"
964 "movq %%mm4, %%mm5\n"
965 "movq %%mm1, %%mm3\n"
966 "punpcklbw %%mm7,%%mm4\n"
967 "punpcklbw %%mm7,%%mm1\n"
968 "punpckhbw %%mm7,%%mm5\n"
969 "punpckhbw %%mm7,%%mm3\n"
970 "psubw %%mm1, %%mm4\n"
971 "psubw %%mm3, %%mm5\n"
972 "psubw %%mm4, %%mm0\n"
973 "psubw %%mm5, %%mm2\n"
974 "pxor %%mm3, %%mm3\n"
975 "pxor %%mm1, %%mm1\n"
976 "pcmpgtw %%mm0, %%mm3\n\t"
977 "pcmpgtw %%mm2, %%mm1\n\t"
978 "pxor %%mm3, %%mm0\n"
979 "pxor %%mm1, %%mm2\n"
980 "psubw %%mm3, %%mm0\n"
981 "psubw %%mm1, %%mm2\n"
982 "paddw %%mm0, %%mm2\n"
983 "paddw %%mm2, %%mm6\n"
989 "movq %%mm0, %%mm1\n"
993 "movq %%mm0, %%mm2\n"
994 "movq %%mm1, %%mm3\n"
995 "punpcklbw %%mm7,%%mm0\n"
996 "punpcklbw %%mm7,%%mm1\n"
997 "punpckhbw %%mm7,%%mm2\n"
998 "punpckhbw %%mm7,%%mm3\n"
999 "psubw %%mm1, %%mm0\n"
1000 "psubw %%mm3, %%mm2\n"
1001 "psubw %%mm0, %%mm4\n"
1002 "psubw %%mm2, %%mm5\n"
1003 "pxor %%mm3, %%mm3\n"
1004 "pxor %%mm1, %%mm1\n"
1005 "pcmpgtw %%mm4, %%mm3\n\t"
1006 "pcmpgtw %%mm5, %%mm1\n\t"
1007 "pxor %%mm3, %%mm4\n"
1008 "pxor %%mm1, %%mm5\n"
1009 "psubw %%mm3, %%mm4\n"
1010 "psubw %%mm1, %%mm5\n"
1011 "paddw %%mm4, %%mm5\n"
1012 "paddw %%mm5, %%mm6\n"
1017 "movq %%mm4, %%mm1\n"
1021 "movq %%mm4, %%mm5\n"
1022 "movq %%mm1, %%mm3\n"
1023 "punpcklbw %%mm7,%%mm4\n"
1024 "punpcklbw %%mm7,%%mm1\n"
1025 "punpckhbw %%mm7,%%mm5\n"
1026 "punpckhbw %%mm7,%%mm3\n"
1027 "psubw %%mm1, %%mm4\n"
1028 "psubw %%mm3, %%mm5\n"
1029 "psubw %%mm4, %%mm0\n"
1030 "psubw %%mm5, %%mm2\n"
1031 "pxor %%mm3, %%mm3\n"
1032 "pxor %%mm1, %%mm1\n"
1033 "pcmpgtw %%mm0, %%mm3\n\t"
1034 "pcmpgtw %%mm2, %%mm1\n\t"
1035 "pxor %%mm3, %%mm0\n"
1036 "pxor %%mm1, %%mm2\n"
1037 "psubw %%mm3, %%mm0\n"
1038 "psubw %%mm1, %%mm2\n"
1039 "paddw %%mm0, %%mm2\n"
1040 "paddw %%mm2, %%mm6\n"
1046 "movq %%mm6, %%mm0\n"
1047 "punpcklwd %%mm7,%%mm0\n"
1048 "punpckhwd %%mm7,%%mm6\n"
1049 "paddd %%mm0, %%mm6\n"
1051 "movq %%mm6,%%mm0\n"
1052 "psrlq $32, %%mm6\n"
1053 "paddd %%mm6,%%mm0\n"
1055 : "+r" (pix1), "=r"(tmp)
1056 : "r" ((long)line_size) , "g" (h-2)
1061 static int hf_noise16_mmx(uint8_t * pix1, int line_size, int h) {
1063 uint8_t * pix= pix1;
1066 "pxor %%mm7,%%mm7\n"
1067 "pxor %%mm6,%%mm6\n"
1070 "movq 1(%0),%%mm1\n"
1071 "movq %%mm0, %%mm2\n"
1072 "movq %%mm1, %%mm3\n"
1073 "punpcklbw %%mm7,%%mm0\n"
1074 "punpcklbw %%mm7,%%mm1\n"
1075 "punpckhbw %%mm7,%%mm2\n"
1076 "punpckhbw %%mm7,%%mm3\n"
1077 "psubw %%mm1, %%mm0\n"
1078 "psubw %%mm3, %%mm2\n"
1083 "movq 1(%0),%%mm1\n"
1084 "movq %%mm4, %%mm5\n"
1085 "movq %%mm1, %%mm3\n"
1086 "punpcklbw %%mm7,%%mm4\n"
1087 "punpcklbw %%mm7,%%mm1\n"
1088 "punpckhbw %%mm7,%%mm5\n"
1089 "punpckhbw %%mm7,%%mm3\n"
1090 "psubw %%mm1, %%mm4\n"
1091 "psubw %%mm3, %%mm5\n"
1092 "psubw %%mm4, %%mm0\n"
1093 "psubw %%mm5, %%mm2\n"
1094 "pxor %%mm3, %%mm3\n"
1095 "pxor %%mm1, %%mm1\n"
1096 "pcmpgtw %%mm0, %%mm3\n\t"
1097 "pcmpgtw %%mm2, %%mm1\n\t"
1098 "pxor %%mm3, %%mm0\n"
1099 "pxor %%mm1, %%mm2\n"
1100 "psubw %%mm3, %%mm0\n"
1101 "psubw %%mm1, %%mm2\n"
1102 "paddw %%mm0, %%mm2\n"
1103 "paddw %%mm2, %%mm6\n"
1109 "movq 1(%0),%%mm1\n"
1110 "movq %%mm0, %%mm2\n"
1111 "movq %%mm1, %%mm3\n"
1112 "punpcklbw %%mm7,%%mm0\n"
1113 "punpcklbw %%mm7,%%mm1\n"
1114 "punpckhbw %%mm7,%%mm2\n"
1115 "punpckhbw %%mm7,%%mm3\n"
1116 "psubw %%mm1, %%mm0\n"
1117 "psubw %%mm3, %%mm2\n"
1118 "psubw %%mm0, %%mm4\n"
1119 "psubw %%mm2, %%mm5\n"
1120 "pxor %%mm3, %%mm3\n"
1121 "pxor %%mm1, %%mm1\n"
1122 "pcmpgtw %%mm4, %%mm3\n\t"
1123 "pcmpgtw %%mm5, %%mm1\n\t"
1124 "pxor %%mm3, %%mm4\n"
1125 "pxor %%mm1, %%mm5\n"
1126 "psubw %%mm3, %%mm4\n"
1127 "psubw %%mm1, %%mm5\n"
1128 "paddw %%mm4, %%mm5\n"
1129 "paddw %%mm5, %%mm6\n"
1134 "movq 1(%0),%%mm1\n"
1135 "movq %%mm4, %%mm5\n"
1136 "movq %%mm1, %%mm3\n"
1137 "punpcklbw %%mm7,%%mm4\n"
1138 "punpcklbw %%mm7,%%mm1\n"
1139 "punpckhbw %%mm7,%%mm5\n"
1140 "punpckhbw %%mm7,%%mm3\n"
1141 "psubw %%mm1, %%mm4\n"
1142 "psubw %%mm3, %%mm5\n"
1143 "psubw %%mm4, %%mm0\n"
1144 "psubw %%mm5, %%mm2\n"
1145 "pxor %%mm3, %%mm3\n"
1146 "pxor %%mm1, %%mm1\n"
1147 "pcmpgtw %%mm0, %%mm3\n\t"
1148 "pcmpgtw %%mm2, %%mm1\n\t"
1149 "pxor %%mm3, %%mm0\n"
1150 "pxor %%mm1, %%mm2\n"
1151 "psubw %%mm3, %%mm0\n"
1152 "psubw %%mm1, %%mm2\n"
1153 "paddw %%mm0, %%mm2\n"
1154 "paddw %%mm2, %%mm6\n"
1160 "movq %%mm6, %%mm0\n"
1161 "punpcklwd %%mm7,%%mm0\n"
1162 "punpckhwd %%mm7,%%mm6\n"
1163 "paddd %%mm0, %%mm6\n"
1165 "movq %%mm6,%%mm0\n"
1166 "psrlq $32, %%mm6\n"
1167 "paddd %%mm6,%%mm0\n"
1169 : "+r" (pix1), "=r"(tmp)
1170 : "r" ((long)line_size) , "g" (h-2)
1172 return tmp + hf_noise8_mmx(pix+8, line_size, h);
1175 static int nsse16_mmx(void *p, uint8_t * pix1, uint8_t * pix2, int line_size, int h) {
1176 MpegEncContext *c = p;
1179 if(c) score1 = c->dsp.sse[0](c, pix1, pix2, line_size, h);
1180 else score1 = sse16_mmx(c, pix1, pix2, line_size, h);
1181 score2= hf_noise16_mmx(pix1, line_size, h) - hf_noise16_mmx(pix2, line_size, h);
1183 if(c) return score1 + ABS(score2)*c->avctx->nsse_weight;
1184 else return score1 + ABS(score2)*8;
1187 static int nsse8_mmx(void *p, uint8_t * pix1, uint8_t * pix2, int line_size, int h) {
1188 MpegEncContext *c = p;
1189 int score1= sse8_mmx(c, pix1, pix2, line_size, h);
1190 int score2= hf_noise8_mmx(pix1, line_size, h) - hf_noise8_mmx(pix2, line_size, h);
1192 if(c) return score1 + ABS(score2)*c->avctx->nsse_weight;
1193 else return score1 + ABS(score2)*8;
1196 static int vsad_intra16_mmx(void *v, uint8_t * pix, uint8_t * dummy, int line_size, int h) {
1199 assert( (((int)pix) & 7) == 0);
1200 assert((line_size &7) ==0);
1202 #define SUM(in0, in1, out0, out1) \
1203 "movq (%0), %%mm2\n"\
1204 "movq 8(%0), %%mm3\n"\
1206 "movq %%mm2, " #out0 "\n"\
1207 "movq %%mm3, " #out1 "\n"\
1208 "psubusb " #in0 ", %%mm2\n"\
1209 "psubusb " #in1 ", %%mm3\n"\
1210 "psubusb " #out0 ", " #in0 "\n"\
1211 "psubusb " #out1 ", " #in1 "\n"\
1212 "por %%mm2, " #in0 "\n"\
1213 "por %%mm3, " #in1 "\n"\
1214 "movq " #in0 ", %%mm2\n"\
1215 "movq " #in1 ", %%mm3\n"\
1216 "punpcklbw %%mm7, " #in0 "\n"\
1217 "punpcklbw %%mm7, " #in1 "\n"\
1218 "punpckhbw %%mm7, %%mm2\n"\
1219 "punpckhbw %%mm7, %%mm3\n"\
1220 "paddw " #in1 ", " #in0 "\n"\
1221 "paddw %%mm3, %%mm2\n"\
1222 "paddw %%mm2, " #in0 "\n"\
1223 "paddw " #in0 ", %%mm6\n"
1228 "pxor %%mm6,%%mm6\n"
1229 "pxor %%mm7,%%mm7\n"
1231 "movq 8(%0),%%mm1\n"
1234 SUM(%%mm0, %%mm1, %%mm4, %%mm5)
1237 SUM(%%mm4, %%mm5, %%mm0, %%mm1)
1239 SUM(%%mm0, %%mm1, %%mm4, %%mm5)
1244 "movq %%mm6,%%mm0\n"
1245 "psrlq $32, %%mm6\n"
1246 "paddw %%mm6,%%mm0\n"
1247 "movq %%mm0,%%mm6\n"
1248 "psrlq $16, %%mm0\n"
1249 "paddw %%mm6,%%mm0\n"
1251 : "+r" (pix), "=r"(tmp)
1252 : "r" ((long)line_size) , "m" (h)
1254 return tmp & 0xFFFF;
1258 static int vsad_intra16_mmx2(void *v, uint8_t * pix, uint8_t * dummy, int line_size, int h) {
1261 assert( (((int)pix) & 7) == 0);
1262 assert((line_size &7) ==0);
1264 #define SUM(in0, in1, out0, out1) \
1265 "movq (%0), " #out0 "\n"\
1266 "movq 8(%0), " #out1 "\n"\
1268 "psadbw " #out0 ", " #in0 "\n"\
1269 "psadbw " #out1 ", " #in1 "\n"\
1270 "paddw " #in1 ", " #in0 "\n"\
1271 "paddw " #in0 ", %%mm6\n"
1275 "pxor %%mm6,%%mm6\n"
1276 "pxor %%mm7,%%mm7\n"
1278 "movq 8(%0),%%mm1\n"
1281 SUM(%%mm0, %%mm1, %%mm4, %%mm5)
1284 SUM(%%mm4, %%mm5, %%mm0, %%mm1)
1286 SUM(%%mm0, %%mm1, %%mm4, %%mm5)
1292 : "+r" (pix), "=r"(tmp)
1293 : "r" ((long)line_size) , "m" (h)
1299 static int vsad16_mmx(void *v, uint8_t * pix1, uint8_t * pix2, int line_size, int h) {
1302 assert( (((int)pix1) & 7) == 0);
1303 assert( (((int)pix2) & 7) == 0);
1304 assert((line_size &7) ==0);
1306 #define SUM(in0, in1, out0, out1) \
1307 "movq (%0),%%mm2\n"\
1308 "movq (%1)," #out0 "\n"\
1309 "movq 8(%0),%%mm3\n"\
1310 "movq 8(%1)," #out1 "\n"\
1313 "psubb " #out0 ", %%mm2\n"\
1314 "psubb " #out1 ", %%mm3\n"\
1315 "pxor %%mm7, %%mm2\n"\
1316 "pxor %%mm7, %%mm3\n"\
1317 "movq %%mm2, " #out0 "\n"\
1318 "movq %%mm3, " #out1 "\n"\
1319 "psubusb " #in0 ", %%mm2\n"\
1320 "psubusb " #in1 ", %%mm3\n"\
1321 "psubusb " #out0 ", " #in0 "\n"\
1322 "psubusb " #out1 ", " #in1 "\n"\
1323 "por %%mm2, " #in0 "\n"\
1324 "por %%mm3, " #in1 "\n"\
1325 "movq " #in0 ", %%mm2\n"\
1326 "movq " #in1 ", %%mm3\n"\
1327 "punpcklbw %%mm7, " #in0 "\n"\
1328 "punpcklbw %%mm7, " #in1 "\n"\
1329 "punpckhbw %%mm7, %%mm2\n"\
1330 "punpckhbw %%mm7, %%mm3\n"\
1331 "paddw " #in1 ", " #in0 "\n"\
1332 "paddw %%mm3, %%mm2\n"\
1333 "paddw %%mm2, " #in0 "\n"\
1334 "paddw " #in0 ", %%mm6\n"
1339 "pxor %%mm6,%%mm6\n"
1340 "pcmpeqw %%mm7,%%mm7\n"
1341 "psllw $15, %%mm7\n"
1342 "packsswb %%mm7, %%mm7\n"
1345 "movq 8(%0),%%mm1\n"
1346 "movq 8(%1),%%mm3\n"
1350 "psubb %%mm2, %%mm0\n"
1351 "psubb %%mm3, %%mm1\n"
1352 "pxor %%mm7, %%mm0\n"
1353 "pxor %%mm7, %%mm1\n"
1354 SUM(%%mm0, %%mm1, %%mm4, %%mm5)
1357 SUM(%%mm4, %%mm5, %%mm0, %%mm1)
1359 SUM(%%mm0, %%mm1, %%mm4, %%mm5)
1364 "movq %%mm6,%%mm0\n"
1365 "psrlq $32, %%mm6\n"
1366 "paddw %%mm6,%%mm0\n"
1367 "movq %%mm0,%%mm6\n"
1368 "psrlq $16, %%mm0\n"
1369 "paddw %%mm6,%%mm0\n"
1371 : "+r" (pix1), "+r" (pix2), "=r"(tmp)
1372 : "r" ((long)line_size) , "m" (h)
1374 return tmp & 0x7FFF;
1378 static int vsad16_mmx2(void *v, uint8_t * pix1, uint8_t * pix2, int line_size, int h) {
1381 assert( (((int)pix1) & 7) == 0);
1382 assert( (((int)pix2) & 7) == 0);
1383 assert((line_size &7) ==0);
1385 #define SUM(in0, in1, out0, out1) \
1386 "movq (%0)," #out0 "\n"\
1387 "movq (%1),%%mm2\n"\
1388 "movq 8(%0)," #out1 "\n"\
1389 "movq 8(%1),%%mm3\n"\
1392 "psubb %%mm2, " #out0 "\n"\
1393 "psubb %%mm3, " #out1 "\n"\
1394 "pxor %%mm7, " #out0 "\n"\
1395 "pxor %%mm7, " #out1 "\n"\
1396 "psadbw " #out0 ", " #in0 "\n"\
1397 "psadbw " #out1 ", " #in1 "\n"\
1398 "paddw " #in1 ", " #in0 "\n"\
1399 "paddw " #in0 ", %%mm6\n"
1403 "pxor %%mm6,%%mm6\n"
1404 "pcmpeqw %%mm7,%%mm7\n"
1405 "psllw $15, %%mm7\n"
1406 "packsswb %%mm7, %%mm7\n"
1409 "movq 8(%0),%%mm1\n"
1410 "movq 8(%1),%%mm3\n"
1414 "psubb %%mm2, %%mm0\n"
1415 "psubb %%mm3, %%mm1\n"
1416 "pxor %%mm7, %%mm0\n"
1417 "pxor %%mm7, %%mm1\n"
1418 SUM(%%mm0, %%mm1, %%mm4, %%mm5)
1421 SUM(%%mm4, %%mm5, %%mm0, %%mm1)
1423 SUM(%%mm0, %%mm1, %%mm4, %%mm5)
1429 : "+r" (pix1), "+r" (pix2), "=r"(tmp)
1430 : "r" ((long)line_size) , "m" (h)
1436 static void diff_bytes_mmx(uint8_t *dst, uint8_t *src1, uint8_t *src2, int w){
1440 "movq (%2, %0), %%mm0 \n\t"
1441 "movq (%1, %0), %%mm1 \n\t"
1442 "psubb %%mm0, %%mm1 \n\t"
1443 "movq %%mm1, (%3, %0) \n\t"
1444 "movq 8(%2, %0), %%mm0 \n\t"
1445 "movq 8(%1, %0), %%mm1 \n\t"
1446 "psubb %%mm0, %%mm1 \n\t"
1447 "movq %%mm1, 8(%3, %0) \n\t"
1452 : "r"(src1), "r"(src2), "r"(dst), "r"((long)w-15)
1455 dst[i+0] = src1[i+0]-src2[i+0];
1458 static void sub_hfyu_median_prediction_mmx2(uint8_t *dst, uint8_t *src1, uint8_t *src2, int w, int *left, int *left_top){
1464 "movq -1(%1, %0), %%mm0 \n\t" // LT
1465 "movq (%1, %0), %%mm1 \n\t" // T
1466 "movq -1(%2, %0), %%mm2 \n\t" // L
1467 "movq (%2, %0), %%mm3 \n\t" // X
1468 "movq %%mm2, %%mm4 \n\t" // L
1469 "psubb %%mm0, %%mm2 \n\t"
1470 "paddb %%mm1, %%mm2 \n\t" // L + T - LT
1471 "movq %%mm4, %%mm5 \n\t" // L
1472 "pmaxub %%mm1, %%mm4 \n\t" // max(T, L)
1473 "pminub %%mm5, %%mm1 \n\t" // min(T, L)
1474 "pminub %%mm2, %%mm4 \n\t"
1475 "pmaxub %%mm1, %%mm4 \n\t"
1476 "psubb %%mm4, %%mm3 \n\t" // dst - pred
1477 "movq %%mm3, (%3, %0) \n\t"
1482 : "r"(src1), "r"(src2), "r"(dst), "r"((long)w)
1488 dst[0]= src2[0] - mid_pred(l, src1[0], (l + src1[0] - lt)&0xFF);
1490 *left_top= src1[w-1];
1494 #define LBUTTERFLY2(a1,b1,a2,b2)\
1495 "paddw " #b1 ", " #a1 " \n\t"\
1496 "paddw " #b2 ", " #a2 " \n\t"\
1497 "paddw " #b1 ", " #b1 " \n\t"\
1498 "paddw " #b2 ", " #b2 " \n\t"\
1499 "psubw " #a1 ", " #b1 " \n\t"\
1500 "psubw " #a2 ", " #b2 " \n\t"
1503 LBUTTERFLY2(%%mm0, %%mm1, %%mm2, %%mm3)\
1504 LBUTTERFLY2(%%mm4, %%mm5, %%mm6, %%mm7)\
1505 LBUTTERFLY2(%%mm0, %%mm2, %%mm1, %%mm3)\
1506 LBUTTERFLY2(%%mm4, %%mm6, %%mm5, %%mm7)\
1507 LBUTTERFLY2(%%mm0, %%mm4, %%mm1, %%mm5)\
1508 LBUTTERFLY2(%%mm2, %%mm6, %%mm3, %%mm7)\
1511 "pxor " #z ", " #z " \n\t"\
1512 "pcmpgtw " #a ", " #z " \n\t"\
1513 "pxor " #z ", " #a " \n\t"\
1514 "psubw " #z ", " #a " \n\t"
1516 #define MMABS_SUM(a,z, sum)\
1517 "pxor " #z ", " #z " \n\t"\
1518 "pcmpgtw " #a ", " #z " \n\t"\
1519 "pxor " #z ", " #a " \n\t"\
1520 "psubw " #z ", " #a " \n\t"\
1521 "paddusw " #a ", " #sum " \n\t"
1523 #define MMABS_MMX2(a,z)\
1524 "pxor " #z ", " #z " \n\t"\
1525 "psubw " #a ", " #z " \n\t"\
1526 "pmaxsw " #z ", " #a " \n\t"
1528 #define MMABS_SUM_MMX2(a,z, sum)\
1529 "pxor " #z ", " #z " \n\t"\
1530 "psubw " #a ", " #z " \n\t"\
1531 "pmaxsw " #z ", " #a " \n\t"\
1532 "paddusw " #a ", " #sum " \n\t"
1534 #define TRANSPOSE4(a,b,c,d,t)\
1535 SBUTTERFLY(a,b,t,wd) /* a=aebf t=cgdh */\
1536 SBUTTERFLY(c,d,b,wd) /* c=imjn b=kolp */\
1537 SBUTTERFLY(a,c,d,dq) /* a=aeim d=bfjn */\
1538 SBUTTERFLY(t,b,c,dq) /* t=cgko c=dhlp */
1540 #define LOAD4(o, a, b, c, d)\
1541 "movq "#o"(%1), " #a " \n\t"\
1542 "movq "#o"+16(%1), " #b " \n\t"\
1543 "movq "#o"+32(%1), " #c " \n\t"\
1544 "movq "#o"+48(%1), " #d " \n\t"
1546 #define STORE4(o, a, b, c, d)\
1547 "movq "#a", "#o"(%1) \n\t"\
1548 "movq "#b", "#o"+16(%1) \n\t"\
1549 "movq "#c", "#o"+32(%1) \n\t"\
1550 "movq "#d", "#o"+48(%1) \n\t"\
1552 static int hadamard8_diff_mmx(void *s, uint8_t *src1, uint8_t *src2, int stride, int h){
1553 DECLARE_ALIGNED_8(uint64_t, temp[16]);
1558 diff_pixels_mmx((DCTELEM*)temp, src1, src2, stride);
1561 LOAD4(0 , %%mm0, %%mm1, %%mm2, %%mm3)
1562 LOAD4(64, %%mm4, %%mm5, %%mm6, %%mm7)
1566 "movq %%mm7, 112(%1) \n\t"
1568 TRANSPOSE4(%%mm0, %%mm1, %%mm2, %%mm3, %%mm7)
1569 STORE4(0 , %%mm0, %%mm3, %%mm7, %%mm2)
1571 "movq 112(%1), %%mm7 \n\t"
1572 TRANSPOSE4(%%mm4, %%mm5, %%mm6, %%mm7, %%mm0)
1573 STORE4(64, %%mm4, %%mm7, %%mm0, %%mm6)
1575 LOAD4(8 , %%mm0, %%mm1, %%mm2, %%mm3)
1576 LOAD4(72, %%mm4, %%mm5, %%mm6, %%mm7)
1580 "movq %%mm7, 120(%1) \n\t"
1582 TRANSPOSE4(%%mm0, %%mm1, %%mm2, %%mm3, %%mm7)
1583 STORE4(8 , %%mm0, %%mm3, %%mm7, %%mm2)
1585 "movq 120(%1), %%mm7 \n\t"
1586 TRANSPOSE4(%%mm4, %%mm5, %%mm6, %%mm7, %%mm0)
1587 "movq %%mm7, %%mm5 \n\t"//FIXME remove
1588 "movq %%mm6, %%mm7 \n\t"
1589 "movq %%mm0, %%mm6 \n\t"
1590 // STORE4(72, %%mm4, %%mm7, %%mm0, %%mm6) //FIXME remove
1592 LOAD4(64, %%mm0, %%mm1, %%mm2, %%mm3)
1593 // LOAD4(72, %%mm4, %%mm5, %%mm6, %%mm7)
1596 "movq %%mm7, 64(%1) \n\t"
1598 MMABS_SUM(%%mm1, %%mm7, %%mm0)
1599 MMABS_SUM(%%mm2, %%mm7, %%mm0)
1600 MMABS_SUM(%%mm3, %%mm7, %%mm0)
1601 MMABS_SUM(%%mm4, %%mm7, %%mm0)
1602 MMABS_SUM(%%mm5, %%mm7, %%mm0)
1603 MMABS_SUM(%%mm6, %%mm7, %%mm0)
1604 "movq 64(%1), %%mm1 \n\t"
1605 MMABS_SUM(%%mm1, %%mm7, %%mm0)
1606 "movq %%mm0, 64(%1) \n\t"
1608 LOAD4(0 , %%mm0, %%mm1, %%mm2, %%mm3)
1609 LOAD4(8 , %%mm4, %%mm5, %%mm6, %%mm7)
1612 "movq %%mm7, (%1) \n\t"
1614 MMABS_SUM(%%mm1, %%mm7, %%mm0)
1615 MMABS_SUM(%%mm2, %%mm7, %%mm0)
1616 MMABS_SUM(%%mm3, %%mm7, %%mm0)
1617 MMABS_SUM(%%mm4, %%mm7, %%mm0)
1618 MMABS_SUM(%%mm5, %%mm7, %%mm0)
1619 MMABS_SUM(%%mm6, %%mm7, %%mm0)
1620 "movq (%1), %%mm1 \n\t"
1621 MMABS_SUM(%%mm1, %%mm7, %%mm0)
1622 "movq 64(%1), %%mm1 \n\t"
1623 MMABS_SUM(%%mm1, %%mm7, %%mm0)
1625 "movq %%mm0, %%mm1 \n\t"
1626 "psrlq $32, %%mm0 \n\t"
1627 "paddusw %%mm1, %%mm0 \n\t"
1628 "movq %%mm0, %%mm1 \n\t"
1629 "psrlq $16, %%mm0 \n\t"
1630 "paddusw %%mm1, %%mm0 \n\t"
1631 "movd %%mm0, %0 \n\t"
1639 static int hadamard8_diff_mmx2(void *s, uint8_t *src1, uint8_t *src2, int stride, int h){
1640 DECLARE_ALIGNED_8(uint64_t, temp[16]);
1645 diff_pixels_mmx((DCTELEM*)temp, src1, src2, stride);
1648 LOAD4(0 , %%mm0, %%mm1, %%mm2, %%mm3)
1649 LOAD4(64, %%mm4, %%mm5, %%mm6, %%mm7)
1653 "movq %%mm7, 112(%1) \n\t"
1655 TRANSPOSE4(%%mm0, %%mm1, %%mm2, %%mm3, %%mm7)
1656 STORE4(0 , %%mm0, %%mm3, %%mm7, %%mm2)
1658 "movq 112(%1), %%mm7 \n\t"
1659 TRANSPOSE4(%%mm4, %%mm5, %%mm6, %%mm7, %%mm0)
1660 STORE4(64, %%mm4, %%mm7, %%mm0, %%mm6)
1662 LOAD4(8 , %%mm0, %%mm1, %%mm2, %%mm3)
1663 LOAD4(72, %%mm4, %%mm5, %%mm6, %%mm7)
1667 "movq %%mm7, 120(%1) \n\t"
1669 TRANSPOSE4(%%mm0, %%mm1, %%mm2, %%mm3, %%mm7)
1670 STORE4(8 , %%mm0, %%mm3, %%mm7, %%mm2)
1672 "movq 120(%1), %%mm7 \n\t"
1673 TRANSPOSE4(%%mm4, %%mm5, %%mm6, %%mm7, %%mm0)
1674 "movq %%mm7, %%mm5 \n\t"//FIXME remove
1675 "movq %%mm6, %%mm7 \n\t"
1676 "movq %%mm0, %%mm6 \n\t"
1677 // STORE4(72, %%mm4, %%mm7, %%mm0, %%mm6) //FIXME remove
1679 LOAD4(64, %%mm0, %%mm1, %%mm2, %%mm3)
1680 // LOAD4(72, %%mm4, %%mm5, %%mm6, %%mm7)
1683 "movq %%mm7, 64(%1) \n\t"
1684 MMABS_MMX2(%%mm0, %%mm7)
1685 MMABS_SUM_MMX2(%%mm1, %%mm7, %%mm0)
1686 MMABS_SUM_MMX2(%%mm2, %%mm7, %%mm0)
1687 MMABS_SUM_MMX2(%%mm3, %%mm7, %%mm0)
1688 MMABS_SUM_MMX2(%%mm4, %%mm7, %%mm0)
1689 MMABS_SUM_MMX2(%%mm5, %%mm7, %%mm0)
1690 MMABS_SUM_MMX2(%%mm6, %%mm7, %%mm0)
1691 "movq 64(%1), %%mm1 \n\t"
1692 MMABS_SUM_MMX2(%%mm1, %%mm7, %%mm0)
1693 "movq %%mm0, 64(%1) \n\t"
1695 LOAD4(0 , %%mm0, %%mm1, %%mm2, %%mm3)
1696 LOAD4(8 , %%mm4, %%mm5, %%mm6, %%mm7)
1699 "movq %%mm7, (%1) \n\t"
1700 MMABS_MMX2(%%mm0, %%mm7)
1701 MMABS_SUM_MMX2(%%mm1, %%mm7, %%mm0)
1702 MMABS_SUM_MMX2(%%mm2, %%mm7, %%mm0)
1703 MMABS_SUM_MMX2(%%mm3, %%mm7, %%mm0)
1704 MMABS_SUM_MMX2(%%mm4, %%mm7, %%mm0)
1705 MMABS_SUM_MMX2(%%mm5, %%mm7, %%mm0)
1706 MMABS_SUM_MMX2(%%mm6, %%mm7, %%mm0)
1707 "movq (%1), %%mm1 \n\t"
1708 MMABS_SUM_MMX2(%%mm1, %%mm7, %%mm0)
1709 "movq 64(%1), %%mm1 \n\t"
1710 MMABS_SUM_MMX2(%%mm1, %%mm7, %%mm0)
1712 "pshufw $0x0E, %%mm0, %%mm1 \n\t"
1713 "paddusw %%mm1, %%mm0 \n\t"
1714 "pshufw $0x01, %%mm0, %%mm1 \n\t"
1715 "paddusw %%mm1, %%mm0 \n\t"
1716 "movd %%mm0, %0 \n\t"
1725 WARPER8_16_SQ(hadamard8_diff_mmx, hadamard8_diff16_mmx)
1726 WARPER8_16_SQ(hadamard8_diff_mmx2, hadamard8_diff16_mmx2)
1727 #endif //CONFIG_ENCODERS
1729 #define put_no_rnd_pixels8_mmx(a,b,c,d) put_pixels8_mmx(a,b,c,d)
1730 #define put_no_rnd_pixels16_mmx(a,b,c,d) put_pixels16_mmx(a,b,c,d)
1732 #define QPEL_V_LOW(m3,m4,m5,m6, pw_20, pw_3, rnd, in0, in1, in2, in7, out, OP)\
1733 "paddw " #m4 ", " #m3 " \n\t" /* x1 */\
1734 "movq "MANGLE(ff_pw_20)", %%mm4 \n\t" /* 20 */\
1735 "pmullw " #m3 ", %%mm4 \n\t" /* 20x1 */\
1736 "movq "#in7", " #m3 " \n\t" /* d */\
1737 "movq "#in0", %%mm5 \n\t" /* D */\
1738 "paddw " #m3 ", %%mm5 \n\t" /* x4 */\
1739 "psubw %%mm5, %%mm4 \n\t" /* 20x1 - x4 */\
1740 "movq "#in1", %%mm5 \n\t" /* C */\
1741 "movq "#in2", %%mm6 \n\t" /* B */\
1742 "paddw " #m6 ", %%mm5 \n\t" /* x3 */\
1743 "paddw " #m5 ", %%mm6 \n\t" /* x2 */\
1744 "paddw %%mm6, %%mm6 \n\t" /* 2x2 */\
1745 "psubw %%mm6, %%mm5 \n\t" /* -2x2 + x3 */\
1746 "pmullw "MANGLE(ff_pw_3)", %%mm5 \n\t" /* -6x2 + 3x3 */\
1747 "paddw " #rnd ", %%mm4 \n\t" /* x2 */\
1748 "paddw %%mm4, %%mm5 \n\t" /* 20x1 - 6x2 + 3x3 - x4 */\
1749 "psraw $5, %%mm5 \n\t"\
1750 "packuswb %%mm5, %%mm5 \n\t"\
1751 OP(%%mm5, out, %%mm7, d)
1753 #define QPEL_BASE(OPNAME, ROUNDER, RND, OP_MMX2, OP_3DNOW)\
1754 static void OPNAME ## mpeg4_qpel16_h_lowpass_mmx2(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\
1758 "pxor %%mm7, %%mm7 \n\t"\
1760 "movq (%0), %%mm0 \n\t" /* ABCDEFGH */\
1761 "movq %%mm0, %%mm1 \n\t" /* ABCDEFGH */\
1762 "movq %%mm0, %%mm2 \n\t" /* ABCDEFGH */\
1763 "punpcklbw %%mm7, %%mm0 \n\t" /* 0A0B0C0D */\
1764 "punpckhbw %%mm7, %%mm1 \n\t" /* 0E0F0G0H */\
1765 "pshufw $0x90, %%mm0, %%mm5 \n\t" /* 0A0A0B0C */\
1766 "pshufw $0x41, %%mm0, %%mm6 \n\t" /* 0B0A0A0B */\
1767 "movq %%mm2, %%mm3 \n\t" /* ABCDEFGH */\
1768 "movq %%mm2, %%mm4 \n\t" /* ABCDEFGH */\
1769 "psllq $8, %%mm2 \n\t" /* 0ABCDEFG */\
1770 "psllq $16, %%mm3 \n\t" /* 00ABCDEF */\
1771 "psllq $24, %%mm4 \n\t" /* 000ABCDE */\
1772 "punpckhbw %%mm7, %%mm2 \n\t" /* 0D0E0F0G */\
1773 "punpckhbw %%mm7, %%mm3 \n\t" /* 0C0D0E0F */\
1774 "punpckhbw %%mm7, %%mm4 \n\t" /* 0B0C0D0E */\
1775 "paddw %%mm3, %%mm5 \n\t" /* b */\
1776 "paddw %%mm2, %%mm6 \n\t" /* c */\
1777 "paddw %%mm5, %%mm5 \n\t" /* 2b */\
1778 "psubw %%mm5, %%mm6 \n\t" /* c - 2b */\
1779 "pshufw $0x06, %%mm0, %%mm5 \n\t" /* 0C0B0A0A */\
1780 "pmullw "MANGLE(ff_pw_3)", %%mm6 \n\t" /* 3c - 6b */\
1781 "paddw %%mm4, %%mm0 \n\t" /* a */\
1782 "paddw %%mm1, %%mm5 \n\t" /* d */\
1783 "pmullw "MANGLE(ff_pw_20)", %%mm0 \n\t" /* 20a */\
1784 "psubw %%mm5, %%mm0 \n\t" /* 20a - d */\
1785 "paddw %6, %%mm6 \n\t"\
1786 "paddw %%mm6, %%mm0 \n\t" /* 20a - 6b + 3c - d */\
1787 "psraw $5, %%mm0 \n\t"\
1788 "movq %%mm0, %5 \n\t"\
1789 /* mm1=EFGH, mm2=DEFG, mm3=CDEF, mm4=BCDE, mm7=0 */\
1791 "movq 5(%0), %%mm0 \n\t" /* FGHIJKLM */\
1792 "movq %%mm0, %%mm5 \n\t" /* FGHIJKLM */\
1793 "movq %%mm0, %%mm6 \n\t" /* FGHIJKLM */\
1794 "psrlq $8, %%mm0 \n\t" /* GHIJKLM0 */\
1795 "psrlq $16, %%mm5 \n\t" /* HIJKLM00 */\
1796 "punpcklbw %%mm7, %%mm0 \n\t" /* 0G0H0I0J */\
1797 "punpcklbw %%mm7, %%mm5 \n\t" /* 0H0I0J0K */\
1798 "paddw %%mm0, %%mm2 \n\t" /* b */\
1799 "paddw %%mm5, %%mm3 \n\t" /* c */\
1800 "paddw %%mm2, %%mm2 \n\t" /* 2b */\
1801 "psubw %%mm2, %%mm3 \n\t" /* c - 2b */\
1802 "movq %%mm6, %%mm2 \n\t" /* FGHIJKLM */\
1803 "psrlq $24, %%mm6 \n\t" /* IJKLM000 */\
1804 "punpcklbw %%mm7, %%mm2 \n\t" /* 0F0G0H0I */\
1805 "punpcklbw %%mm7, %%mm6 \n\t" /* 0I0J0K0L */\
1806 "pmullw "MANGLE(ff_pw_3)", %%mm3 \n\t" /* 3c - 6b */\
1807 "paddw %%mm2, %%mm1 \n\t" /* a */\
1808 "paddw %%mm6, %%mm4 \n\t" /* d */\
1809 "pmullw "MANGLE(ff_pw_20)", %%mm1 \n\t" /* 20a */\
1810 "psubw %%mm4, %%mm3 \n\t" /* - 6b +3c - d */\
1811 "paddw %6, %%mm1 \n\t"\
1812 "paddw %%mm1, %%mm3 \n\t" /* 20a - 6b +3c - d */\
1813 "psraw $5, %%mm3 \n\t"\
1814 "movq %5, %%mm1 \n\t"\
1815 "packuswb %%mm3, %%mm1 \n\t"\
1816 OP_MMX2(%%mm1, (%1),%%mm4, q)\
1817 /* mm0= GHIJ, mm2=FGHI, mm5=HIJK, mm6=IJKL, mm7=0 */\
1819 "movq 9(%0), %%mm1 \n\t" /* JKLMNOPQ */\
1820 "movq %%mm1, %%mm4 \n\t" /* JKLMNOPQ */\
1821 "movq %%mm1, %%mm3 \n\t" /* JKLMNOPQ */\
1822 "psrlq $8, %%mm1 \n\t" /* KLMNOPQ0 */\
1823 "psrlq $16, %%mm4 \n\t" /* LMNOPQ00 */\
1824 "punpcklbw %%mm7, %%mm1 \n\t" /* 0K0L0M0N */\
1825 "punpcklbw %%mm7, %%mm4 \n\t" /* 0L0M0N0O */\
1826 "paddw %%mm1, %%mm5 \n\t" /* b */\
1827 "paddw %%mm4, %%mm0 \n\t" /* c */\
1828 "paddw %%mm5, %%mm5 \n\t" /* 2b */\
1829 "psubw %%mm5, %%mm0 \n\t" /* c - 2b */\
1830 "movq %%mm3, %%mm5 \n\t" /* JKLMNOPQ */\
1831 "psrlq $24, %%mm3 \n\t" /* MNOPQ000 */\
1832 "pmullw "MANGLE(ff_pw_3)", %%mm0 \n\t" /* 3c - 6b */\
1833 "punpcklbw %%mm7, %%mm3 \n\t" /* 0M0N0O0P */\
1834 "paddw %%mm3, %%mm2 \n\t" /* d */\
1835 "psubw %%mm2, %%mm0 \n\t" /* -6b + 3c - d */\
1836 "movq %%mm5, %%mm2 \n\t" /* JKLMNOPQ */\
1837 "punpcklbw %%mm7, %%mm2 \n\t" /* 0J0K0L0M */\
1838 "punpckhbw %%mm7, %%mm5 \n\t" /* 0N0O0P0Q */\
1839 "paddw %%mm2, %%mm6 \n\t" /* a */\
1840 "pmullw "MANGLE(ff_pw_20)", %%mm6 \n\t" /* 20a */\
1841 "paddw %6, %%mm0 \n\t"\
1842 "paddw %%mm6, %%mm0 \n\t" /* 20a - 6b + 3c - d */\
1843 "psraw $5, %%mm0 \n\t"\
1844 /* mm1=KLMN, mm2=JKLM, mm3=MNOP, mm4=LMNO, mm5=NOPQ mm7=0 */\
1846 "paddw %%mm5, %%mm3 \n\t" /* a */\
1847 "pshufw $0xF9, %%mm5, %%mm6 \n\t" /* 0O0P0Q0Q */\
1848 "paddw %%mm4, %%mm6 \n\t" /* b */\
1849 "pshufw $0xBE, %%mm5, %%mm4 \n\t" /* 0P0Q0Q0P */\
1850 "pshufw $0x6F, %%mm5, %%mm5 \n\t" /* 0Q0Q0P0O */\
1851 "paddw %%mm1, %%mm4 \n\t" /* c */\
1852 "paddw %%mm2, %%mm5 \n\t" /* d */\
1853 "paddw %%mm6, %%mm6 \n\t" /* 2b */\
1854 "psubw %%mm6, %%mm4 \n\t" /* c - 2b */\
1855 "pmullw "MANGLE(ff_pw_20)", %%mm3 \n\t" /* 20a */\
1856 "pmullw "MANGLE(ff_pw_3)", %%mm4 \n\t" /* 3c - 6b */\
1857 "psubw %%mm5, %%mm3 \n\t" /* -6b + 3c - d */\
1858 "paddw %6, %%mm4 \n\t"\
1859 "paddw %%mm3, %%mm4 \n\t" /* 20a - 6b + 3c - d */\
1860 "psraw $5, %%mm4 \n\t"\
1861 "packuswb %%mm4, %%mm0 \n\t"\
1862 OP_MMX2(%%mm0, 8(%1), %%mm4, q)\
1868 : "+a"(src), "+c"(dst), "+m"(h)\
1869 : "d"((long)srcStride), "S"((long)dstStride), /*"m"(ff_pw_20), "m"(ff_pw_3),*/ "m"(temp), "m"(ROUNDER)\
1874 static void OPNAME ## mpeg4_qpel16_h_lowpass_3dnow(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\
1877 /* quick HACK, XXX FIXME MUST be optimized */\
1880 temp[ 0]= (src[ 0]+src[ 1])*20 - (src[ 0]+src[ 2])*6 + (src[ 1]+src[ 3])*3 - (src[ 2]+src[ 4]);\
1881 temp[ 1]= (src[ 1]+src[ 2])*20 - (src[ 0]+src[ 3])*6 + (src[ 0]+src[ 4])*3 - (src[ 1]+src[ 5]);\
1882 temp[ 2]= (src[ 2]+src[ 3])*20 - (src[ 1]+src[ 4])*6 + (src[ 0]+src[ 5])*3 - (src[ 0]+src[ 6]);\
1883 temp[ 3]= (src[ 3]+src[ 4])*20 - (src[ 2]+src[ 5])*6 + (src[ 1]+src[ 6])*3 - (src[ 0]+src[ 7]);\
1884 temp[ 4]= (src[ 4]+src[ 5])*20 - (src[ 3]+src[ 6])*6 + (src[ 2]+src[ 7])*3 - (src[ 1]+src[ 8]);\
1885 temp[ 5]= (src[ 5]+src[ 6])*20 - (src[ 4]+src[ 7])*6 + (src[ 3]+src[ 8])*3 - (src[ 2]+src[ 9]);\
1886 temp[ 6]= (src[ 6]+src[ 7])*20 - (src[ 5]+src[ 8])*6 + (src[ 4]+src[ 9])*3 - (src[ 3]+src[10]);\
1887 temp[ 7]= (src[ 7]+src[ 8])*20 - (src[ 6]+src[ 9])*6 + (src[ 5]+src[10])*3 - (src[ 4]+src[11]);\
1888 temp[ 8]= (src[ 8]+src[ 9])*20 - (src[ 7]+src[10])*6 + (src[ 6]+src[11])*3 - (src[ 5]+src[12]);\
1889 temp[ 9]= (src[ 9]+src[10])*20 - (src[ 8]+src[11])*6 + (src[ 7]+src[12])*3 - (src[ 6]+src[13]);\
1890 temp[10]= (src[10]+src[11])*20 - (src[ 9]+src[12])*6 + (src[ 8]+src[13])*3 - (src[ 7]+src[14]);\
1891 temp[11]= (src[11]+src[12])*20 - (src[10]+src[13])*6 + (src[ 9]+src[14])*3 - (src[ 8]+src[15]);\
1892 temp[12]= (src[12]+src[13])*20 - (src[11]+src[14])*6 + (src[10]+src[15])*3 - (src[ 9]+src[16]);\
1893 temp[13]= (src[13]+src[14])*20 - (src[12]+src[15])*6 + (src[11]+src[16])*3 - (src[10]+src[16]);\
1894 temp[14]= (src[14]+src[15])*20 - (src[13]+src[16])*6 + (src[12]+src[16])*3 - (src[11]+src[15]);\
1895 temp[15]= (src[15]+src[16])*20 - (src[14]+src[16])*6 + (src[13]+src[15])*3 - (src[12]+src[14]);\
1897 "movq (%0), %%mm0 \n\t"\
1898 "movq 8(%0), %%mm1 \n\t"\
1899 "paddw %2, %%mm0 \n\t"\
1900 "paddw %2, %%mm1 \n\t"\
1901 "psraw $5, %%mm0 \n\t"\
1902 "psraw $5, %%mm1 \n\t"\
1903 "packuswb %%mm1, %%mm0 \n\t"\
1904 OP_3DNOW(%%mm0, (%1), %%mm1, q)\
1905 "movq 16(%0), %%mm0 \n\t"\
1906 "movq 24(%0), %%mm1 \n\t"\
1907 "paddw %2, %%mm0 \n\t"\
1908 "paddw %2, %%mm1 \n\t"\
1909 "psraw $5, %%mm0 \n\t"\
1910 "psraw $5, %%mm1 \n\t"\
1911 "packuswb %%mm1, %%mm0 \n\t"\
1912 OP_3DNOW(%%mm0, 8(%1), %%mm1, q)\
1913 :: "r"(temp), "r"(dst), "m"(ROUNDER)\
1921 static void OPNAME ## mpeg4_qpel8_h_lowpass_mmx2(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\
1925 "pxor %%mm7, %%mm7 \n\t"\
1927 "movq (%0), %%mm0 \n\t" /* ABCDEFGH */\
1928 "movq %%mm0, %%mm1 \n\t" /* ABCDEFGH */\
1929 "movq %%mm0, %%mm2 \n\t" /* ABCDEFGH */\
1930 "punpcklbw %%mm7, %%mm0 \n\t" /* 0A0B0C0D */\
1931 "punpckhbw %%mm7, %%mm1 \n\t" /* 0E0F0G0H */\
1932 "pshufw $0x90, %%mm0, %%mm5 \n\t" /* 0A0A0B0C */\
1933 "pshufw $0x41, %%mm0, %%mm6 \n\t" /* 0B0A0A0B */\
1934 "movq %%mm2, %%mm3 \n\t" /* ABCDEFGH */\
1935 "movq %%mm2, %%mm4 \n\t" /* ABCDEFGH */\
1936 "psllq $8, %%mm2 \n\t" /* 0ABCDEFG */\
1937 "psllq $16, %%mm3 \n\t" /* 00ABCDEF */\
1938 "psllq $24, %%mm4 \n\t" /* 000ABCDE */\
1939 "punpckhbw %%mm7, %%mm2 \n\t" /* 0D0E0F0G */\
1940 "punpckhbw %%mm7, %%mm3 \n\t" /* 0C0D0E0F */\
1941 "punpckhbw %%mm7, %%mm4 \n\t" /* 0B0C0D0E */\
1942 "paddw %%mm3, %%mm5 \n\t" /* b */\
1943 "paddw %%mm2, %%mm6 \n\t" /* c */\
1944 "paddw %%mm5, %%mm5 \n\t" /* 2b */\
1945 "psubw %%mm5, %%mm6 \n\t" /* c - 2b */\
1946 "pshufw $0x06, %%mm0, %%mm5 \n\t" /* 0C0B0A0A */\
1947 "pmullw "MANGLE(ff_pw_3)", %%mm6 \n\t" /* 3c - 6b */\
1948 "paddw %%mm4, %%mm0 \n\t" /* a */\
1949 "paddw %%mm1, %%mm5 \n\t" /* d */\
1950 "pmullw "MANGLE(ff_pw_20)", %%mm0 \n\t" /* 20a */\
1951 "psubw %%mm5, %%mm0 \n\t" /* 20a - d */\
1952 "paddw %6, %%mm6 \n\t"\
1953 "paddw %%mm6, %%mm0 \n\t" /* 20a - 6b + 3c - d */\
1954 "psraw $5, %%mm0 \n\t"\
1955 /* mm1=EFGH, mm2=DEFG, mm3=CDEF, mm4=BCDE, mm7=0 */\
1957 "movd 5(%0), %%mm5 \n\t" /* FGHI */\
1958 "punpcklbw %%mm7, %%mm5 \n\t" /* 0F0G0H0I */\
1959 "pshufw $0xF9, %%mm5, %%mm6 \n\t" /* 0G0H0I0I */\
1960 "paddw %%mm5, %%mm1 \n\t" /* a */\
1961 "paddw %%mm6, %%mm2 \n\t" /* b */\
1962 "pshufw $0xBE, %%mm5, %%mm6 \n\t" /* 0H0I0I0H */\
1963 "pshufw $0x6F, %%mm5, %%mm5 \n\t" /* 0I0I0H0G */\
1964 "paddw %%mm6, %%mm3 \n\t" /* c */\
1965 "paddw %%mm5, %%mm4 \n\t" /* d */\
1966 "paddw %%mm2, %%mm2 \n\t" /* 2b */\
1967 "psubw %%mm2, %%mm3 \n\t" /* c - 2b */\
1968 "pmullw "MANGLE(ff_pw_20)", %%mm1 \n\t" /* 20a */\
1969 "pmullw "MANGLE(ff_pw_3)", %%mm3 \n\t" /* 3c - 6b */\
1970 "psubw %%mm4, %%mm3 \n\t" /* -6b + 3c - d */\
1971 "paddw %6, %%mm1 \n\t"\
1972 "paddw %%mm1, %%mm3 \n\t" /* 20a - 6b + 3c - d */\
1973 "psraw $5, %%mm3 \n\t"\
1974 "packuswb %%mm3, %%mm0 \n\t"\
1975 OP_MMX2(%%mm0, (%1), %%mm4, q)\
1981 : "+a"(src), "+c"(dst), "+m"(h)\
1982 : "S"((long)srcStride), "D"((long)dstStride), /*"m"(ff_pw_20), "m"(ff_pw_3),*/ "m"(temp), "m"(ROUNDER)\
1987 static void OPNAME ## mpeg4_qpel8_h_lowpass_3dnow(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\
1990 /* quick HACK, XXX FIXME MUST be optimized */\
1993 temp[ 0]= (src[ 0]+src[ 1])*20 - (src[ 0]+src[ 2])*6 + (src[ 1]+src[ 3])*3 - (src[ 2]+src[ 4]);\
1994 temp[ 1]= (src[ 1]+src[ 2])*20 - (src[ 0]+src[ 3])*6 + (src[ 0]+src[ 4])*3 - (src[ 1]+src[ 5]);\
1995 temp[ 2]= (src[ 2]+src[ 3])*20 - (src[ 1]+src[ 4])*6 + (src[ 0]+src[ 5])*3 - (src[ 0]+src[ 6]);\
1996 temp[ 3]= (src[ 3]+src[ 4])*20 - (src[ 2]+src[ 5])*6 + (src[ 1]+src[ 6])*3 - (src[ 0]+src[ 7]);\
1997 temp[ 4]= (src[ 4]+src[ 5])*20 - (src[ 3]+src[ 6])*6 + (src[ 2]+src[ 7])*3 - (src[ 1]+src[ 8]);\
1998 temp[ 5]= (src[ 5]+src[ 6])*20 - (src[ 4]+src[ 7])*6 + (src[ 3]+src[ 8])*3 - (src[ 2]+src[ 8]);\
1999 temp[ 6]= (src[ 6]+src[ 7])*20 - (src[ 5]+src[ 8])*6 + (src[ 4]+src[ 8])*3 - (src[ 3]+src[ 7]);\
2000 temp[ 7]= (src[ 7]+src[ 8])*20 - (src[ 6]+src[ 8])*6 + (src[ 5]+src[ 7])*3 - (src[ 4]+src[ 6]);\
2002 "movq (%0), %%mm0 \n\t"\
2003 "movq 8(%0), %%mm1 \n\t"\
2004 "paddw %2, %%mm0 \n\t"\
2005 "paddw %2, %%mm1 \n\t"\
2006 "psraw $5, %%mm0 \n\t"\
2007 "psraw $5, %%mm1 \n\t"\
2008 "packuswb %%mm1, %%mm0 \n\t"\
2009 OP_3DNOW(%%mm0, (%1), %%mm1, q)\
2010 :: "r"(temp), "r"(dst), "m"(ROUNDER)\
2018 #define QPEL_OP(OPNAME, ROUNDER, RND, OP, MMX)\
2020 static void OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
2021 uint64_t temp[17*4];\
2022 uint64_t *temp_ptr= temp;\
2027 "pxor %%mm7, %%mm7 \n\t"\
2029 "movq (%0), %%mm0 \n\t"\
2030 "movq (%0), %%mm1 \n\t"\
2031 "movq 8(%0), %%mm2 \n\t"\
2032 "movq 8(%0), %%mm3 \n\t"\
2033 "punpcklbw %%mm7, %%mm0 \n\t"\
2034 "punpckhbw %%mm7, %%mm1 \n\t"\
2035 "punpcklbw %%mm7, %%mm2 \n\t"\
2036 "punpckhbw %%mm7, %%mm3 \n\t"\
2037 "movq %%mm0, (%1) \n\t"\
2038 "movq %%mm1, 17*8(%1) \n\t"\
2039 "movq %%mm2, 2*17*8(%1) \n\t"\
2040 "movq %%mm3, 3*17*8(%1) \n\t"\
2045 : "+r" (src), "+r" (temp_ptr), "+r"(count)\
2046 : "r" ((long)srcStride)\
2053 /*FIXME reorder for speed */\
2055 /*"pxor %%mm7, %%mm7 \n\t"*/\
2057 "movq (%0), %%mm0 \n\t"\
2058 "movq 8(%0), %%mm1 \n\t"\
2059 "movq 16(%0), %%mm2 \n\t"\
2060 "movq 24(%0), %%mm3 \n\t"\
2061 QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 16(%0), 8(%0), (%0), 32(%0), (%1), OP)\
2062 QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 8(%0), (%0), (%0), 40(%0), (%1, %3), OP)\
2064 QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, (%0), (%0), 8(%0), 48(%0), (%1), OP)\
2066 QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, (%0), 8(%0), 16(%0), 56(%0), (%1, %3), OP)\
2068 QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 8(%0), 16(%0), 24(%0), 64(%0), (%1), OP)\
2069 QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 16(%0), 24(%0), 32(%0), 72(%0), (%1, %3), OP)\
2071 QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 24(%0), 32(%0), 40(%0), 80(%0), (%1), OP)\
2072 QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 32(%0), 40(%0), 48(%0), 88(%0), (%1, %3), OP)\
2074 QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 40(%0), 48(%0), 56(%0), 96(%0), (%1), OP)\
2075 QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 48(%0), 56(%0), 64(%0),104(%0), (%1, %3), OP)\
2077 QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 56(%0), 64(%0), 72(%0),112(%0), (%1), OP)\
2078 QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 64(%0), 72(%0), 80(%0),120(%0), (%1, %3), OP)\
2080 QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 72(%0), 80(%0), 88(%0),128(%0), (%1), OP)\
2082 QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 80(%0), 88(%0), 96(%0),128(%0), (%1, %3), OP)\
2084 QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 88(%0), 96(%0),104(%0),120(%0), (%1), OP)\
2085 QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 96(%0),104(%0),112(%0),112(%0), (%1, %3), OP)\
2087 "add $136, %0 \n\t"\
2092 : "+r"(temp_ptr), "+r"(dst), "+g"(count)\
2093 : "r"((long)dstStride), "r"(2*(long)dstStride), /*"m"(ff_pw_20), "m"(ff_pw_3),*/ "m"(ROUNDER), "g"(4-14*(long)dstStride)\
2098 static void OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
2099 uint64_t temp[9*2];\
2100 uint64_t *temp_ptr= temp;\
2105 "pxor %%mm7, %%mm7 \n\t"\
2107 "movq (%0), %%mm0 \n\t"\
2108 "movq (%0), %%mm1 \n\t"\
2109 "punpcklbw %%mm7, %%mm0 \n\t"\
2110 "punpckhbw %%mm7, %%mm1 \n\t"\
2111 "movq %%mm0, (%1) \n\t"\
2112 "movq %%mm1, 9*8(%1) \n\t"\
2117 : "+r" (src), "+r" (temp_ptr), "+r"(count)\
2118 : "r" ((long)srcStride)\
2125 /*FIXME reorder for speed */\
2127 /*"pxor %%mm7, %%mm7 \n\t"*/\
2129 "movq (%0), %%mm0 \n\t"\
2130 "movq 8(%0), %%mm1 \n\t"\
2131 "movq 16(%0), %%mm2 \n\t"\
2132 "movq 24(%0), %%mm3 \n\t"\
2133 QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 16(%0), 8(%0), (%0), 32(%0), (%1), OP)\
2134 QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 8(%0), (%0), (%0), 40(%0), (%1, %3), OP)\
2136 QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, (%0), (%0), 8(%0), 48(%0), (%1), OP)\
2138 QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, (%0), 8(%0), 16(%0), 56(%0), (%1, %3), OP)\
2140 QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 8(%0), 16(%0), 24(%0), 64(%0), (%1), OP)\
2142 QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 16(%0), 24(%0), 32(%0), 64(%0), (%1, %3), OP)\
2144 QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 24(%0), 32(%0), 40(%0), 56(%0), (%1), OP)\
2145 QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 32(%0), 40(%0), 48(%0), 48(%0), (%1, %3), OP)\
2152 : "+r"(temp_ptr), "+r"(dst), "+g"(count)\
2153 : "r"((long)dstStride), "r"(2*(long)dstStride), /*"m"(ff_pw_20), "m"(ff_pw_3),*/ "m"(ROUNDER), "g"(4-6*(long)dstStride)\
2158 static void OPNAME ## qpel8_mc00_ ## MMX (uint8_t *dst, uint8_t *src, int stride){\
2159 OPNAME ## pixels8_mmx(dst, src, stride, 8);\
2162 static void OPNAME ## qpel8_mc10_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
2164 uint8_t * const half= (uint8_t*)temp;\
2165 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(half, src, 8, stride, 8);\
2166 OPNAME ## pixels8_l2_ ## MMX(dst, src, half, stride, stride, 8);\
2169 static void OPNAME ## qpel8_mc20_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
2170 OPNAME ## mpeg4_qpel8_h_lowpass_ ## MMX(dst, src, stride, stride, 8);\
2173 static void OPNAME ## qpel8_mc30_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
2175 uint8_t * const half= (uint8_t*)temp;\
2176 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(half, src, 8, stride, 8);\
2177 OPNAME ## pixels8_l2_ ## MMX(dst, src+1, half, stride, stride, 8);\
2180 static void OPNAME ## qpel8_mc01_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
2182 uint8_t * const half= (uint8_t*)temp;\
2183 put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(half, src, 8, stride);\
2184 OPNAME ## pixels8_l2_ ## MMX(dst, src, half, stride, stride, 8);\
2187 static void OPNAME ## qpel8_mc02_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
2188 OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, src, stride, stride);\
2191 static void OPNAME ## qpel8_mc03_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
2193 uint8_t * const half= (uint8_t*)temp;\
2194 put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(half, src, 8, stride);\
2195 OPNAME ## pixels8_l2_ ## MMX(dst, src+stride, half, stride, stride, 8);\
2197 static void OPNAME ## qpel8_mc11_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
2198 uint64_t half[8 + 9];\
2199 uint8_t * const halfH= ((uint8_t*)half) + 64;\
2200 uint8_t * const halfHV= ((uint8_t*)half);\
2201 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\
2202 put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8, stride, 9);\
2203 put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
2204 OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV, stride, 8, 8);\
2206 static void OPNAME ## qpel8_mc31_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
2207 uint64_t half[8 + 9];\
2208 uint8_t * const halfH= ((uint8_t*)half) + 64;\
2209 uint8_t * const halfHV= ((uint8_t*)half);\
2210 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\
2211 put ## RND ## pixels8_l2_ ## MMX(halfH, src+1, halfH, 8, stride, 9);\
2212 put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
2213 OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV, stride, 8, 8);\
2215 static void OPNAME ## qpel8_mc13_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
2216 uint64_t half[8 + 9];\
2217 uint8_t * const halfH= ((uint8_t*)half) + 64;\
2218 uint8_t * const halfHV= ((uint8_t*)half);\
2219 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\
2220 put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8, stride, 9);\
2221 put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
2222 OPNAME ## pixels8_l2_ ## MMX(dst, halfH+8, halfHV, stride, 8, 8);\
2224 static void OPNAME ## qpel8_mc33_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
2225 uint64_t half[8 + 9];\
2226 uint8_t * const halfH= ((uint8_t*)half) + 64;\
2227 uint8_t * const halfHV= ((uint8_t*)half);\
2228 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\
2229 put ## RND ## pixels8_l2_ ## MMX(halfH, src+1, halfH, 8, stride, 9);\
2230 put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
2231 OPNAME ## pixels8_l2_ ## MMX(dst, halfH+8, halfHV, stride, 8, 8);\
2233 static void OPNAME ## qpel8_mc21_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
2234 uint64_t half[8 + 9];\
2235 uint8_t * const halfH= ((uint8_t*)half) + 64;\
2236 uint8_t * const halfHV= ((uint8_t*)half);\
2237 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\
2238 put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
2239 OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV, stride, 8, 8);\
2241 static void OPNAME ## qpel8_mc23_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
2242 uint64_t half[8 + 9];\
2243 uint8_t * const halfH= ((uint8_t*)half) + 64;\
2244 uint8_t * const halfHV= ((uint8_t*)half);\
2245 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\
2246 put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
2247 OPNAME ## pixels8_l2_ ## MMX(dst, halfH+8, halfHV, stride, 8, 8);\
2249 static void OPNAME ## qpel8_mc12_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
2250 uint64_t half[8 + 9];\
2251 uint8_t * const halfH= ((uint8_t*)half);\
2252 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\
2253 put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8, stride, 9);\
2254 OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH, stride, 8);\
2256 static void OPNAME ## qpel8_mc32_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
2257 uint64_t half[8 + 9];\
2258 uint8_t * const halfH= ((uint8_t*)half);\
2259 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\
2260 put ## RND ## pixels8_l2_ ## MMX(halfH, src+1, halfH, 8, stride, 9);\
2261 OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH, stride, 8);\
2263 static void OPNAME ## qpel8_mc22_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
2265 uint8_t * const halfH= ((uint8_t*)half);\
2266 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\
2267 OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH, stride, 8);\
2269 static void OPNAME ## qpel16_mc00_ ## MMX (uint8_t *dst, uint8_t *src, int stride){\
2270 OPNAME ## pixels16_mmx(dst, src, stride, 16);\
2273 static void OPNAME ## qpel16_mc10_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
2275 uint8_t * const half= (uint8_t*)temp;\
2276 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(half, src, 16, stride, 16);\
2277 OPNAME ## pixels16_l2_ ## MMX(dst, src, half, stride, stride, 16);\
2280 static void OPNAME ## qpel16_mc20_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
2281 OPNAME ## mpeg4_qpel16_h_lowpass_ ## MMX(dst, src, stride, stride, 16);\
2284 static void OPNAME ## qpel16_mc30_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
2286 uint8_t * const half= (uint8_t*)temp;\
2287 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(half, src, 16, stride, 16);\
2288 OPNAME ## pixels16_l2_ ## MMX(dst, src+1, half, stride, stride, 16);\
2291 static void OPNAME ## qpel16_mc01_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
2293 uint8_t * const half= (uint8_t*)temp;\
2294 put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(half, src, 16, stride);\
2295 OPNAME ## pixels16_l2_ ## MMX(dst, src, half, stride, stride, 16);\
2298 static void OPNAME ## qpel16_mc02_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
2299 OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, src, stride, stride);\
2302 static void OPNAME ## qpel16_mc03_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
2304 uint8_t * const half= (uint8_t*)temp;\
2305 put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(half, src, 16, stride);\
2306 OPNAME ## pixels16_l2_ ## MMX(dst, src+stride, half, stride, stride, 16);\
2308 static void OPNAME ## qpel16_mc11_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
2309 uint64_t half[16*2 + 17*2];\
2310 uint8_t * const halfH= ((uint8_t*)half) + 256;\
2311 uint8_t * const halfHV= ((uint8_t*)half);\
2312 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\
2313 put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16, stride, 17);\
2314 put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\
2315 OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV, stride, 16, 16);\
2317 static void OPNAME ## qpel16_mc31_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
2318 uint64_t half[16*2 + 17*2];\
2319 uint8_t * const halfH= ((uint8_t*)half) + 256;\
2320 uint8_t * const halfHV= ((uint8_t*)half);\
2321 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\
2322 put ## RND ## pixels16_l2_ ## MMX(halfH, src+1, halfH, 16, stride, 17);\
2323 put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\
2324 OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV, stride, 16, 16);\
2326 static void OPNAME ## qpel16_mc13_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
2327 uint64_t half[16*2 + 17*2];\
2328 uint8_t * const halfH= ((uint8_t*)half) + 256;\
2329 uint8_t * const halfHV= ((uint8_t*)half);\
2330 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\
2331 put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16, stride, 17);\
2332 put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\
2333 OPNAME ## pixels16_l2_ ## MMX(dst, halfH+16, halfHV, stride, 16, 16);\
2335 static void OPNAME ## qpel16_mc33_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
2336 uint64_t half[16*2 + 17*2];\
2337 uint8_t * const halfH= ((uint8_t*)half) + 256;\
2338 uint8_t * const halfHV= ((uint8_t*)half);\
2339 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\
2340 put ## RND ## pixels16_l2_ ## MMX(halfH, src+1, halfH, 16, stride, 17);\
2341 put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\
2342 OPNAME ## pixels16_l2_ ## MMX(dst, halfH+16, halfHV, stride, 16, 16);\
2344 static void OPNAME ## qpel16_mc21_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
2345 uint64_t half[16*2 + 17*2];\
2346 uint8_t * const halfH= ((uint8_t*)half) + 256;\
2347 uint8_t * const halfHV= ((uint8_t*)half);\
2348 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\
2349 put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\
2350 OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV, stride, 16, 16);\
2352 static void OPNAME ## qpel16_mc23_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
2353 uint64_t half[16*2 + 17*2];\
2354 uint8_t * const halfH= ((uint8_t*)half) + 256;\
2355 uint8_t * const halfHV= ((uint8_t*)half);\
2356 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\
2357 put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\
2358 OPNAME ## pixels16_l2_ ## MMX(dst, halfH+16, halfHV, stride, 16, 16);\
2360 static void OPNAME ## qpel16_mc12_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
2361 uint64_t half[17*2];\
2362 uint8_t * const halfH= ((uint8_t*)half);\
2363 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\
2364 put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16, stride, 17);\
2365 OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH, stride, 16);\
2367 static void OPNAME ## qpel16_mc32_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
2368 uint64_t half[17*2];\
2369 uint8_t * const halfH= ((uint8_t*)half);\
2370 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\
2371 put ## RND ## pixels16_l2_ ## MMX(halfH, src+1, halfH, 16, stride, 17);\
2372 OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH, stride, 16);\
2374 static void OPNAME ## qpel16_mc22_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
2375 uint64_t half[17*2];\
2376 uint8_t * const halfH= ((uint8_t*)half);\
2377 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\
2378 OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH, stride, 16);\
2381 #define PUT_OP(a,b,temp, size) "mov" #size " " #a ", " #b " \n\t"
2382 #define AVG_3DNOW_OP(a,b,temp, size) \
2383 "mov" #size " " #b ", " #temp " \n\t"\
2384 "pavgusb " #temp ", " #a " \n\t"\
2385 "mov" #size " " #a ", " #b " \n\t"
2386 #define AVG_MMX2_OP(a,b,temp, size) \
2387 "mov" #size " " #b ", " #temp " \n\t"\
2388 "pavgb " #temp ", " #a " \n\t"\
2389 "mov" #size " " #a ", " #b " \n\t"
2391 QPEL_BASE(put_ , ff_pw_16, _ , PUT_OP, PUT_OP)
2392 QPEL_BASE(avg_ , ff_pw_16, _ , AVG_MMX2_OP, AVG_3DNOW_OP)
2393 QPEL_BASE(put_no_rnd_, ff_pw_15, _no_rnd_, PUT_OP, PUT_OP)
2394 QPEL_OP(put_ , ff_pw_16, _ , PUT_OP, 3dnow)
2395 QPEL_OP(avg_ , ff_pw_16, _ , AVG_3DNOW_OP, 3dnow)
2396 QPEL_OP(put_no_rnd_, ff_pw_15, _no_rnd_, PUT_OP, 3dnow)
2397 QPEL_OP(put_ , ff_pw_16, _ , PUT_OP, mmx2)
2398 QPEL_OP(avg_ , ff_pw_16, _ , AVG_MMX2_OP, mmx2)
2399 QPEL_OP(put_no_rnd_, ff_pw_15, _no_rnd_, PUT_OP, mmx2)
2402 static void just_return() { return; }
2405 #define SET_QPEL_FUNC(postfix1, postfix2) \
2406 c->put_ ## postfix1 = put_ ## postfix2;\
2407 c->put_no_rnd_ ## postfix1 = put_no_rnd_ ## postfix2;\
2408 c->avg_ ## postfix1 = avg_ ## postfix2;
2410 static void gmc_mmx(uint8_t *dst, uint8_t *src, int stride, int h, int ox, int oy,
2411 int dxx, int dxy, int dyx, int dyy, int shift, int r, int width, int height){
2413 const int ix = ox>>(16+shift);
2414 const int iy = oy>>(16+shift);
2415 const int oxs = ox>>4;
2416 const int oys = oy>>4;
2417 const int dxxs = dxx>>4;
2418 const int dxys = dxy>>4;
2419 const int dyxs = dyx>>4;
2420 const int dyys = dyy>>4;
2421 const uint16_t r4[4] = {r,r,r,r};
2422 const uint16_t dxy4[4] = {dxys,dxys,dxys,dxys};
2423 const uint16_t dyy4[4] = {dyys,dyys,dyys,dyys};
2424 const uint64_t shift2 = 2*shift;
2425 uint8_t edge_buf[(h+1)*stride];
2428 const int dxw = (dxx-(1<<(16+shift)))*(w-1);
2429 const int dyh = (dyy-(1<<(16+shift)))*(h-1);
2430 const int dxh = dxy*(h-1);
2431 const int dyw = dyx*(w-1);
2432 if( // non-constant fullpel offset (3% of blocks)
2433 (ox^(ox+dxw) | ox^(ox+dxh) | ox^(ox+dxw+dxh) |
2434 oy^(oy+dyw) | oy^(oy+dyh) | oy^(oy+dyw+dyh)) >> (16+shift)
2435 // uses more than 16 bits of subpel mv (only at huge resolution)
2436 || (dxx|dxy|dyx|dyy)&15 )
2438 //FIXME could still use mmx for some of the rows
2439 ff_gmc_c(dst, src, stride, h, ox, oy, dxx, dxy, dyx, dyy, shift, r, width, height);
2443 src += ix + iy*stride;
2444 if( (unsigned)ix >= width-w ||
2445 (unsigned)iy >= height-h )
2447 ff_emulated_edge_mc(edge_buf, src, stride, w+1, h+1, ix, iy, width, height);
2452 "movd %0, %%mm6 \n\t"
2453 "pxor %%mm7, %%mm7 \n\t"
2454 "punpcklwd %%mm6, %%mm6 \n\t"
2455 "punpcklwd %%mm6, %%mm6 \n\t"
2459 for(x=0; x<w; x+=4){
2460 uint16_t dx4[4] = { oxs - dxys + dxxs*(x+0),
2461 oxs - dxys + dxxs*(x+1),
2462 oxs - dxys + dxxs*(x+2),
2463 oxs - dxys + dxxs*(x+3) };
2464 uint16_t dy4[4] = { oys - dyys + dyxs*(x+0),
2465 oys - dyys + dyxs*(x+1),
2466 oys - dyys + dyxs*(x+2),
2467 oys - dyys + dyxs*(x+3) };
2471 "movq %0, %%mm4 \n\t"
2472 "movq %1, %%mm5 \n\t"
2473 "paddw %2, %%mm4 \n\t"
2474 "paddw %3, %%mm5 \n\t"
2475 "movq %%mm4, %0 \n\t"
2476 "movq %%mm5, %1 \n\t"
2477 "psrlw $12, %%mm4 \n\t"
2478 "psrlw $12, %%mm5 \n\t"
2479 : "+m"(*dx4), "+m"(*dy4)
2480 : "m"(*dxy4), "m"(*dyy4)
2484 "movq %%mm6, %%mm2 \n\t"
2485 "movq %%mm6, %%mm1 \n\t"
2486 "psubw %%mm4, %%mm2 \n\t"
2487 "psubw %%mm5, %%mm1 \n\t"
2488 "movq %%mm2, %%mm0 \n\t"
2489 "movq %%mm4, %%mm3 \n\t"
2490 "pmullw %%mm1, %%mm0 \n\t" // (s-dx)*(s-dy)
2491 "pmullw %%mm5, %%mm3 \n\t" // dx*dy
2492 "pmullw %%mm5, %%mm2 \n\t" // (s-dx)*dy
2493 "pmullw %%mm4, %%mm1 \n\t" // dx*(s-dy)
2495 "movd %4, %%mm5 \n\t"
2496 "movd %3, %%mm4 \n\t"
2497 "punpcklbw %%mm7, %%mm5 \n\t"
2498 "punpcklbw %%mm7, %%mm4 \n\t"
2499 "pmullw %%mm5, %%mm3 \n\t" // src[1,1] * dx*dy
2500 "pmullw %%mm4, %%mm2 \n\t" // src[0,1] * (s-dx)*dy
2502 "movd %2, %%mm5 \n\t"
2503 "movd %1, %%mm4 \n\t"
2504 "punpcklbw %%mm7, %%mm5 \n\t"
2505 "punpcklbw %%mm7, %%mm4 \n\t"
2506 "pmullw %%mm5, %%mm1 \n\t" // src[1,0] * dx*(s-dy)
2507 "pmullw %%mm4, %%mm0 \n\t" // src[0,0] * (s-dx)*(s-dy)
2508 "paddw %5, %%mm1 \n\t"
2509 "paddw %%mm3, %%mm2 \n\t"
2510 "paddw %%mm1, %%mm0 \n\t"
2511 "paddw %%mm2, %%mm0 \n\t"
2513 "psrlw %6, %%mm0 \n\t"
2514 "packuswb %%mm0, %%mm0 \n\t"
2515 "movd %%mm0, %0 \n\t"
2517 : "=m"(dst[x+y*stride])
2518 : "m"(src[0]), "m"(src[1]),
2519 "m"(src[stride]), "m"(src[stride+1]),
2520 "m"(*r4), "m"(shift2)
2528 static int try_8x8basis_mmx(int16_t rem[64], int16_t weight[64], int16_t basis[64], int scale){
2531 assert(ABS(scale) < 256);
2532 scale<<= 16 + 1 - BASIS_SHIFT + RECON_SHIFT;
2535 "pcmpeqw %%mm6, %%mm6 \n\t" // -1w
2536 "psrlw $15, %%mm6 \n\t" // 1w
2537 "pxor %%mm7, %%mm7 \n\t"
2538 "movd %4, %%mm5 \n\t"
2539 "punpcklwd %%mm5, %%mm5 \n\t"
2540 "punpcklwd %%mm5, %%mm5 \n\t"
2542 "movq (%1, %0), %%mm0 \n\t"
2543 "movq 8(%1, %0), %%mm1 \n\t"
2544 "pmulhw %%mm5, %%mm0 \n\t"
2545 "pmulhw %%mm5, %%mm1 \n\t"
2546 "paddw %%mm6, %%mm0 \n\t"
2547 "paddw %%mm6, %%mm1 \n\t"
2548 "psraw $1, %%mm0 \n\t"
2549 "psraw $1, %%mm1 \n\t"
2550 "paddw (%2, %0), %%mm0 \n\t"
2551 "paddw 8(%2, %0), %%mm1 \n\t"
2552 "psraw $6, %%mm0 \n\t"
2553 "psraw $6, %%mm1 \n\t"
2554 "pmullw (%3, %0), %%mm0 \n\t"
2555 "pmullw 8(%3, %0), %%mm1 \n\t"
2556 "pmaddwd %%mm0, %%mm0 \n\t"
2557 "pmaddwd %%mm1, %%mm1 \n\t"
2558 "paddd %%mm1, %%mm0 \n\t"
2559 "psrld $4, %%mm0 \n\t"
2560 "paddd %%mm0, %%mm7 \n\t"
2562 "cmp $128, %0 \n\t" //FIXME optimize & bench
2564 "movq %%mm7, %%mm6 \n\t"
2565 "psrlq $32, %%mm7 \n\t"
2566 "paddd %%mm6, %%mm7 \n\t"
2567 "psrld $2, %%mm7 \n\t"
2568 "movd %%mm7, %0 \n\t"
2571 : "r"(basis), "r"(rem), "r"(weight), "g"(scale)
2576 static void add_8x8basis_mmx(int16_t rem[64], int16_t basis[64], int scale){
2579 if(ABS(scale) < 256){
2580 scale<<= 16 + 1 - BASIS_SHIFT + RECON_SHIFT;
2582 "pcmpeqw %%mm6, %%mm6 \n\t" // -1w
2583 "psrlw $15, %%mm6 \n\t" // 1w
2584 "movd %3, %%mm5 \n\t"
2585 "punpcklwd %%mm5, %%mm5 \n\t"
2586 "punpcklwd %%mm5, %%mm5 \n\t"
2588 "movq (%1, %0), %%mm0 \n\t"
2589 "movq 8(%1, %0), %%mm1 \n\t"
2590 "pmulhw %%mm5, %%mm0 \n\t"
2591 "pmulhw %%mm5, %%mm1 \n\t"
2592 "paddw %%mm6, %%mm0 \n\t"
2593 "paddw %%mm6, %%mm1 \n\t"
2594 "psraw $1, %%mm0 \n\t"
2595 "psraw $1, %%mm1 \n\t"
2596 "paddw (%2, %0), %%mm0 \n\t"
2597 "paddw 8(%2, %0), %%mm1 \n\t"
2598 "movq %%mm0, (%2, %0) \n\t"
2599 "movq %%mm1, 8(%2, %0) \n\t"
2601 "cmp $128, %0 \n\t" //FIXME optimize & bench
2605 : "r"(basis), "r"(rem), "g"(scale)
2608 for(i=0; i<8*8; i++){
2609 rem[i] += (basis[i]*scale + (1<<(BASIS_SHIFT - RECON_SHIFT-1)))>>(BASIS_SHIFT - RECON_SHIFT);
2614 #define PREFETCH(name, op) \
2615 void name(void *mem, int stride, int h){\
2616 const uint8_t *p= mem;\
2618 asm volatile(#op" %0" :: "m"(*p));\
2622 PREFETCH(prefetch_mmx2, prefetcht0)
2623 PREFETCH(prefetch_3dnow, prefetch)
2626 #include "h264dsp_mmx.c"
2629 void ff_cavsdsp_init_mmx2(DSPContext* c, AVCodecContext *avctx);
2631 void ff_put_cavs_qpel8_mc00_mmx2(uint8_t *dst, uint8_t *src, int stride) {
2632 put_pixels8_mmx(dst, src, stride, 8);
2634 void ff_avg_cavs_qpel8_mc00_mmx2(uint8_t *dst, uint8_t *src, int stride) {
2635 avg_pixels8_mmx(dst, src, stride, 8);
2637 void ff_put_cavs_qpel16_mc00_mmx2(uint8_t *dst, uint8_t *src, int stride) {
2638 put_pixels16_mmx(dst, src, stride, 16);
2640 void ff_avg_cavs_qpel16_mc00_mmx2(uint8_t *dst, uint8_t *src, int stride) {
2641 avg_pixels16_mmx(dst, src, stride, 16);
2644 /* external functions, from idct_mmx.c */
2645 void ff_mmx_idct(DCTELEM *block);
2646 void ff_mmxext_idct(DCTELEM *block);
2648 void ff_vp3_idct_sse2(int16_t *input_data);
2649 void ff_vp3_idct_mmx(int16_t *data);
2650 void ff_vp3_dsp_init_mmx(void);
2652 /* XXX: those functions should be suppressed ASAP when all IDCTs are
2654 static void ff_libmpeg2mmx_idct_put(uint8_t *dest, int line_size, DCTELEM *block)
2656 ff_mmx_idct (block);
2657 put_pixels_clamped_mmx(block, dest, line_size);
2659 static void ff_libmpeg2mmx_idct_add(uint8_t *dest, int line_size, DCTELEM *block)
2661 ff_mmx_idct (block);
2662 add_pixels_clamped_mmx(block, dest, line_size);
2664 static void ff_libmpeg2mmx2_idct_put(uint8_t *dest, int line_size, DCTELEM *block)
2666 ff_mmxext_idct (block);
2667 put_pixels_clamped_mmx(block, dest, line_size);
2669 static void ff_libmpeg2mmx2_idct_add(uint8_t *dest, int line_size, DCTELEM *block)
2671 ff_mmxext_idct (block);
2672 add_pixels_clamped_mmx(block, dest, line_size);
2674 static void ff_vp3_idct_put_sse2(uint8_t *dest, int line_size, DCTELEM *block)
2676 ff_vp3_idct_sse2(block);
2677 put_signed_pixels_clamped_mmx(block, dest, line_size);
2679 static void ff_vp3_idct_add_sse2(uint8_t *dest, int line_size, DCTELEM *block)
2681 ff_vp3_idct_sse2(block);
2682 add_pixels_clamped_mmx(block, dest, line_size);
2684 static void ff_vp3_idct_put_mmx(uint8_t *dest, int line_size, DCTELEM *block)
2686 ff_vp3_idct_mmx(block);
2687 put_signed_pixels_clamped_mmx(block, dest, line_size);
2689 static void ff_vp3_idct_add_mmx(uint8_t *dest, int line_size, DCTELEM *block)
2691 ff_vp3_idct_mmx(block);
2692 add_pixels_clamped_mmx(block, dest, line_size);
2695 static void ff_idct_xvid_mmx_put(uint8_t *dest, int line_size, DCTELEM *block)
2697 ff_idct_xvid_mmx (block);
2698 put_pixels_clamped_mmx(block, dest, line_size);
2700 static void ff_idct_xvid_mmx_add(uint8_t *dest, int line_size, DCTELEM *block)
2702 ff_idct_xvid_mmx (block);
2703 add_pixels_clamped_mmx(block, dest, line_size);
2705 static void ff_idct_xvid_mmx2_put(uint8_t *dest, int line_size, DCTELEM *block)
2707 ff_idct_xvid_mmx2 (block);
2708 put_pixels_clamped_mmx(block, dest, line_size);
2710 static void ff_idct_xvid_mmx2_add(uint8_t *dest, int line_size, DCTELEM *block)
2712 ff_idct_xvid_mmx2 (block);
2713 add_pixels_clamped_mmx(block, dest, line_size);
2717 static void vorbis_inverse_coupling_3dnow(float *mag, float *ang, int blocksize)
2720 asm volatile("pxor %%mm7, %%mm7":);
2721 for(i=0; i<blocksize; i+=2) {
2723 "movq %0, %%mm0 \n\t"
2724 "movq %1, %%mm1 \n\t"
2725 "movq %%mm0, %%mm2 \n\t"
2726 "movq %%mm1, %%mm3 \n\t"
2727 "pfcmpge %%mm7, %%mm2 \n\t" // m <= 0.0
2728 "pfcmpge %%mm7, %%mm3 \n\t" // a <= 0.0
2729 "pslld $31, %%mm2 \n\t" // keep only the sign bit
2730 "pxor %%mm2, %%mm1 \n\t"
2731 "movq %%mm3, %%mm4 \n\t"
2732 "pand %%mm1, %%mm3 \n\t"
2733 "pandn %%mm1, %%mm4 \n\t"
2734 "pfadd %%mm0, %%mm3 \n\t" // a = m + ((a<0) & (a ^ sign(m)))
2735 "pfsub %%mm4, %%mm0 \n\t" // m = m + ((a>0) & (a ^ sign(m)))
2736 "movq %%mm3, %1 \n\t"
2737 "movq %%mm0, %0 \n\t"
2738 :"+m"(mag[i]), "+m"(ang[i])
2742 asm volatile("femms");
2744 static void vorbis_inverse_coupling_sse(float *mag, float *ang, int blocksize)
2749 "movaps %0, %%xmm5 \n\t"
2750 ::"m"(ff_pdw_80000000[0])
2752 for(i=0; i<blocksize; i+=4) {
2754 "movaps %0, %%xmm0 \n\t"
2755 "movaps %1, %%xmm1 \n\t"
2756 "xorps %%xmm2, %%xmm2 \n\t"
2757 "xorps %%xmm3, %%xmm3 \n\t"
2758 "cmpleps %%xmm0, %%xmm2 \n\t" // m <= 0.0
2759 "cmpleps %%xmm1, %%xmm3 \n\t" // a <= 0.0
2760 "andps %%xmm5, %%xmm2 \n\t" // keep only the sign bit
2761 "xorps %%xmm2, %%xmm1 \n\t"
2762 "movaps %%xmm3, %%xmm4 \n\t"
2763 "andps %%xmm1, %%xmm3 \n\t"
2764 "andnps %%xmm1, %%xmm4 \n\t"
2765 "addps %%xmm0, %%xmm3 \n\t" // a = m + ((a<0) & (a ^ sign(m)))
2766 "subps %%xmm4, %%xmm0 \n\t" // m = m + ((a>0) & (a ^ sign(m)))
2767 "movaps %%xmm3, %1 \n\t"
2768 "movaps %%xmm0, %0 \n\t"
2769 :"+m"(mag[i]), "+m"(ang[i])
2775 static void vector_fmul_3dnow(float *dst, const float *src, int len){
2778 for(i=0; i<len; i++) {
2780 "movq %0, %%mm0 \n\t"
2781 "pfmul %1, %%mm0 \n\t"
2782 "movq %%mm0, %0 \n\t"
2788 asm volatile("femms");
2790 static void vector_fmul_sse(float *dst, const float *src, int len){
2793 for(i=0; i<len; i++) {
2795 "movaps %0, %%xmm0 \n\t"
2796 "mulps %1, %%xmm0 \n\t"
2797 "movaps %%xmm0, %0 \n\t"
2805 static void vector_fmul_reverse_3dnow2(float *dst, const float *src0, const float *src1, int len){
2809 "pswapd 8(%1), %%mm0 \n\t"
2810 "pswapd (%1), %%mm1 \n\t"
2811 "pfmul (%3,%0), %%mm0 \n\t"
2812 "pfmul 8(%3,%0), %%mm1 \n\t"
2813 "movq %%mm0, (%2,%0) \n\t"
2814 "movq %%mm1, 8(%2,%0) \n\t"
2818 :"+r"(i), "+r"(src1)
2819 :"r"(dst), "r"(src0)
2821 asm volatile("femms");
2823 static void vector_fmul_reverse_sse2(float *dst, const float *src0, const float *src1, int len){
2827 "pshufd $0x1b, 16(%1), %%xmm0 \n\t"
2828 "pshufd $0x1b, (%1), %%xmm1 \n\t"
2829 "mulps (%3,%0), %%xmm0 \n\t"
2830 "mulps 16(%3,%0), %%xmm1 \n\t"
2831 "movaps %%xmm0, (%2,%0) \n\t"
2832 "movaps %%xmm1, 16(%2,%0) \n\t"
2836 :"+r"(i), "+r"(src1)
2837 :"r"(dst), "r"(src0)
2841 static void vector_fmul_add_add_3dnow(float *dst, const float *src0, const float *src1,
2842 const float *src2, int src3, int len, int step){
2844 if(step == 2 && src3 == 0){
2849 "movq (%2,%0), %%mm0 \n\t"
2850 "movq 8(%2,%0), %%mm1 \n\t"
2851 "pfmul (%3,%0), %%mm0 \n\t"
2852 "pfmul 8(%3,%0), %%mm1 \n\t"
2853 "pfadd (%4,%0), %%mm0 \n\t"
2854 "pfadd 8(%4,%0), %%mm1 \n\t"
2855 "movd %%mm0, (%1) \n\t"
2856 "movd %%mm1, 16(%1) \n\t"
2857 "psrlq $32, %%mm0 \n\t"
2858 "psrlq $32, %%mm1 \n\t"
2859 "movd %%mm0, 8(%1) \n\t"
2860 "movd %%mm1, 24(%1) \n\t"
2865 :"r"(src0), "r"(src1), "r"(src2)
2869 else if(step == 1 && src3 == 0){
2870 for(i=0; i<len; i+=2){
2872 "movq %1, %%mm0 \n\t"
2873 "pfmul %2, %%mm0 \n\t"
2874 "pfadd %3, %%mm0 \n\t"
2875 "movq %%mm0, %0 \n\t"
2877 :"m"(src0[i]), "m"(src1[i]), "m"(src2[i])
2882 ff_vector_fmul_add_add_c(dst, src0, src1, src2, src3, len, step);
2883 asm volatile("femms");
2885 static void vector_fmul_add_add_sse2(float *dst, const float *src0, const float *src1,
2886 const float *src2, float src3, int len, int step){
2888 if(step == 2 && src3 == 0){
2893 "movaps (%2,%0), %%xmm0 \n\t"
2894 "movaps 16(%2,%0), %%xmm1 \n\t"
2895 "mulps (%3,%0), %%xmm0 \n\t"
2896 "mulps 16(%3,%0), %%xmm1 \n\t"
2897 "addps (%4,%0), %%xmm0 \n\t"
2898 "addps 16(%4,%0), %%xmm1 \n\t"
2899 "movd %%xmm0, (%1) \n\t"
2900 "movd %%xmm1, 32(%1) \n\t"
2901 "psrldq $4, %%xmm0 \n\t"
2902 "psrldq $4, %%xmm1 \n\t"
2903 "movd %%xmm0, 8(%1) \n\t"
2904 "movd %%xmm1, 40(%1) \n\t"
2905 "psrldq $4, %%xmm0 \n\t"
2906 "psrldq $4, %%xmm1 \n\t"
2907 "movd %%xmm0, 16(%1) \n\t"
2908 "movd %%xmm1, 48(%1) \n\t"
2909 "psrldq $4, %%xmm0 \n\t"
2910 "psrldq $4, %%xmm1 \n\t"
2911 "movd %%xmm0, 24(%1) \n\t"
2912 "movd %%xmm1, 56(%1) \n\t"
2917 :"r"(src0), "r"(src1), "r"(src2)
2921 else if(step == 1 && src3 == 0){
2922 for(i=0; i<len; i+=4){
2924 "movaps %1, %%xmm0 \n\t"
2925 "mulps %2, %%xmm0 \n\t"
2926 "addps %3, %%xmm0 \n\t"
2927 "movaps %%xmm0, %0 \n\t"
2929 :"m"(src0[i]), "m"(src1[i]), "m"(src2[i])
2934 ff_vector_fmul_add_add_c(dst, src0, src1, src2, src3, len, step);
2937 void float_to_int16_3dnow(int16_t *dst, const float *src, int len){
2938 // not bit-exact: pf2id uses different rounding than C and SSE
2940 for(i=0; i<len; i+=4) {
2942 "pf2id %1, %%mm0 \n\t"
2943 "pf2id %2, %%mm1 \n\t"
2944 "packssdw %%mm1, %%mm0 \n\t"
2945 "movq %%mm0, %0 \n\t"
2947 :"m"(src[i]), "m"(src[i+2])
2950 asm volatile("femms");
2952 void float_to_int16_sse(int16_t *dst, const float *src, int len){
2954 for(i=0; i<len; i+=4) {
2956 "cvtps2pi %1, %%mm0 \n\t"
2957 "cvtps2pi %2, %%mm1 \n\t"
2958 "packssdw %%mm1, %%mm0 \n\t"
2959 "movq %%mm0, %0 \n\t"
2961 :"m"(src[i]), "m"(src[i+2])
2964 asm volatile("emms");
2967 #ifdef CONFIG_SNOW_ENCODER
2968 extern void ff_snow_horizontal_compose97i_sse2(DWTELEM *b, int width);
2969 extern void ff_snow_horizontal_compose97i_mmx(DWTELEM *b, int width);
2970 extern void ff_snow_vertical_compose97i_sse2(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2, DWTELEM *b3, DWTELEM *b4, DWTELEM *b5, int width);
2971 extern void ff_snow_vertical_compose97i_mmx(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2, DWTELEM *b3, DWTELEM *b4, DWTELEM *b5, int width);
2972 extern void ff_snow_inner_add_yblock_sse2(uint8_t *obmc, const int obmc_stride, uint8_t * * block, int b_w, int b_h,
2973 int src_x, int src_y, int src_stride, slice_buffer * sb, int add, uint8_t * dst8);
2974 extern void ff_snow_inner_add_yblock_mmx(uint8_t *obmc, const int obmc_stride, uint8_t * * block, int b_w, int b_h,
2975 int src_x, int src_y, int src_stride, slice_buffer * sb, int add, uint8_t * dst8);
2978 void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx)
2980 mm_flags = mm_support();
2982 if (avctx->dsp_mask) {
2983 if (avctx->dsp_mask & FF_MM_FORCE)
2984 mm_flags |= (avctx->dsp_mask & 0xffff);
2986 mm_flags &= ~(avctx->dsp_mask & 0xffff);
2990 av_log(avctx, AV_LOG_INFO, "libavcodec: CPU flags:");
2991 if (mm_flags & MM_MMX)
2992 av_log(avctx, AV_LOG_INFO, " mmx");
2993 if (mm_flags & MM_MMXEXT)
2994 av_log(avctx, AV_LOG_INFO, " mmxext");
2995 if (mm_flags & MM_3DNOW)
2996 av_log(avctx, AV_LOG_INFO, " 3dnow");
2997 if (mm_flags & MM_SSE)
2998 av_log(avctx, AV_LOG_INFO, " sse");
2999 if (mm_flags & MM_SSE2)
3000 av_log(avctx, AV_LOG_INFO, " sse2");
3001 av_log(avctx, AV_LOG_INFO, "\n");
3004 if (mm_flags & MM_MMX) {
3005 const int idct_algo= avctx->idct_algo;
3007 #ifdef CONFIG_ENCODERS
3008 const int dct_algo = avctx->dct_algo;
3009 if(dct_algo==FF_DCT_AUTO || dct_algo==FF_DCT_MMX){
3010 if(mm_flags & MM_SSE2){
3011 c->fdct = ff_fdct_sse2;
3012 }else if(mm_flags & MM_MMXEXT){
3013 c->fdct = ff_fdct_mmx2;
3015 c->fdct = ff_fdct_mmx;
3018 #endif //CONFIG_ENCODERS
3019 if(avctx->lowres==0){
3020 if(idct_algo==FF_IDCT_AUTO || idct_algo==FF_IDCT_SIMPLEMMX){
3021 c->idct_put= ff_simple_idct_put_mmx;
3022 c->idct_add= ff_simple_idct_add_mmx;
3023 c->idct = ff_simple_idct_mmx;
3024 c->idct_permutation_type= FF_SIMPLE_IDCT_PERM;
3025 }else if(idct_algo==FF_IDCT_LIBMPEG2MMX){
3026 if(mm_flags & MM_MMXEXT){
3027 c->idct_put= ff_libmpeg2mmx2_idct_put;
3028 c->idct_add= ff_libmpeg2mmx2_idct_add;
3029 c->idct = ff_mmxext_idct;
3031 c->idct_put= ff_libmpeg2mmx_idct_put;
3032 c->idct_add= ff_libmpeg2mmx_idct_add;
3033 c->idct = ff_mmx_idct;
3035 c->idct_permutation_type= FF_LIBMPEG2_IDCT_PERM;
3037 }else if(idct_algo==FF_IDCT_VP3){
3038 if(mm_flags & MM_SSE2){
3039 c->idct_put= ff_vp3_idct_put_sse2;
3040 c->idct_add= ff_vp3_idct_add_sse2;
3041 c->idct = ff_vp3_idct_sse2;
3042 c->idct_permutation_type= FF_TRANSPOSE_IDCT_PERM;
3044 ff_vp3_dsp_init_mmx();
3045 c->idct_put= ff_vp3_idct_put_mmx;
3046 c->idct_add= ff_vp3_idct_add_mmx;
3047 c->idct = ff_vp3_idct_mmx;
3048 c->idct_permutation_type= FF_PARTTRANS_IDCT_PERM;
3051 }else if(idct_algo==FF_IDCT_CAVS){
3052 c->idct_permutation_type= FF_TRANSPOSE_IDCT_PERM;
3054 }else if(idct_algo==FF_IDCT_XVIDMMX){
3055 if(mm_flags & MM_MMXEXT){
3056 c->idct_put= ff_idct_xvid_mmx2_put;
3057 c->idct_add= ff_idct_xvid_mmx2_add;
3058 c->idct = ff_idct_xvid_mmx2;
3060 c->idct_put= ff_idct_xvid_mmx_put;
3061 c->idct_add= ff_idct_xvid_mmx_add;
3062 c->idct = ff_idct_xvid_mmx;
3068 #ifdef CONFIG_ENCODERS
3069 c->get_pixels = get_pixels_mmx;
3070 c->diff_pixels = diff_pixels_mmx;
3071 #endif //CONFIG_ENCODERS
3072 c->put_pixels_clamped = put_pixels_clamped_mmx;
3073 c->put_signed_pixels_clamped = put_signed_pixels_clamped_mmx;
3074 c->add_pixels_clamped = add_pixels_clamped_mmx;
3075 c->clear_blocks = clear_blocks_mmx;
3076 #ifdef CONFIG_ENCODERS
3077 c->pix_sum = pix_sum16_mmx;
3078 #endif //CONFIG_ENCODERS
3080 c->put_pixels_tab[0][0] = put_pixels16_mmx;
3081 c->put_pixels_tab[0][1] = put_pixels16_x2_mmx;
3082 c->put_pixels_tab[0][2] = put_pixels16_y2_mmx;
3083 c->put_pixels_tab[0][3] = put_pixels16_xy2_mmx;
3085 c->put_no_rnd_pixels_tab[0][0] = put_pixels16_mmx;
3086 c->put_no_rnd_pixels_tab[0][1] = put_no_rnd_pixels16_x2_mmx;
3087 c->put_no_rnd_pixels_tab[0][2] = put_no_rnd_pixels16_y2_mmx;
3088 c->put_no_rnd_pixels_tab[0][3] = put_no_rnd_pixels16_xy2_mmx;
3090 c->avg_pixels_tab[0][0] = avg_pixels16_mmx;
3091 c->avg_pixels_tab[0][1] = avg_pixels16_x2_mmx;
3092 c->avg_pixels_tab[0][2] = avg_pixels16_y2_mmx;
3093 c->avg_pixels_tab[0][3] = avg_pixels16_xy2_mmx;
3095 c->avg_no_rnd_pixels_tab[0][0] = avg_no_rnd_pixels16_mmx;
3096 c->avg_no_rnd_pixels_tab[0][1] = avg_no_rnd_pixels16_x2_mmx;
3097 c->avg_no_rnd_pixels_tab[0][2] = avg_no_rnd_pixels16_y2_mmx;
3098 c->avg_no_rnd_pixels_tab[0][3] = avg_no_rnd_pixels16_xy2_mmx;
3100 c->put_pixels_tab[1][0] = put_pixels8_mmx;
3101 c->put_pixels_tab[1][1] = put_pixels8_x2_mmx;
3102 c->put_pixels_tab[1][2] = put_pixels8_y2_mmx;
3103 c->put_pixels_tab[1][3] = put_pixels8_xy2_mmx;
3105 c->put_no_rnd_pixels_tab[1][0] = put_pixels8_mmx;
3106 c->put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_mmx;
3107 c->put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_mmx;
3108 c->put_no_rnd_pixels_tab[1][3] = put_no_rnd_pixels8_xy2_mmx;
3110 c->avg_pixels_tab[1][0] = avg_pixels8_mmx;
3111 c->avg_pixels_tab[1][1] = avg_pixels8_x2_mmx;
3112 c->avg_pixels_tab[1][2] = avg_pixels8_y2_mmx;
3113 c->avg_pixels_tab[1][3] = avg_pixels8_xy2_mmx;
3115 c->avg_no_rnd_pixels_tab[1][0] = avg_no_rnd_pixels8_mmx;
3116 c->avg_no_rnd_pixels_tab[1][1] = avg_no_rnd_pixels8_x2_mmx;
3117 c->avg_no_rnd_pixels_tab[1][2] = avg_no_rnd_pixels8_y2_mmx;
3118 c->avg_no_rnd_pixels_tab[1][3] = avg_no_rnd_pixels8_xy2_mmx;
3122 c->add_bytes= add_bytes_mmx;
3123 #ifdef CONFIG_ENCODERS
3124 c->diff_bytes= diff_bytes_mmx;
3126 c->hadamard8_diff[0]= hadamard8_diff16_mmx;
3127 c->hadamard8_diff[1]= hadamard8_diff_mmx;
3129 c->pix_norm1 = pix_norm1_mmx;
3130 c->sse[0] = (mm_flags & MM_SSE2) ? sse16_sse2 : sse16_mmx;
3131 c->sse[1] = sse8_mmx;
3132 c->vsad[4]= vsad_intra16_mmx;
3134 c->nsse[0] = nsse16_mmx;
3135 c->nsse[1] = nsse8_mmx;
3136 if(!(avctx->flags & CODEC_FLAG_BITEXACT)){
3137 c->vsad[0] = vsad16_mmx;
3140 if(!(avctx->flags & CODEC_FLAG_BITEXACT)){
3141 c->try_8x8basis= try_8x8basis_mmx;
3143 c->add_8x8basis= add_8x8basis_mmx;
3145 #endif //CONFIG_ENCODERS
3147 c->h263_v_loop_filter= h263_v_loop_filter_mmx;
3148 c->h263_h_loop_filter= h263_h_loop_filter_mmx;
3149 c->put_h264_chroma_pixels_tab[0]= put_h264_chroma_mc8_mmx;
3150 c->put_h264_chroma_pixels_tab[1]= put_h264_chroma_mc4_mmx;
3152 c->h264_idct_dc_add=
3153 c->h264_idct_add= ff_h264_idct_add_mmx;
3154 c->h264_idct8_dc_add=
3155 c->h264_idct8_add= ff_h264_idct8_add_mmx;
3157 if (mm_flags & MM_MMXEXT) {
3158 c->prefetch = prefetch_mmx2;
3160 c->put_pixels_tab[0][1] = put_pixels16_x2_mmx2;
3161 c->put_pixels_tab[0][2] = put_pixels16_y2_mmx2;
3163 c->avg_pixels_tab[0][0] = avg_pixels16_mmx2;
3164 c->avg_pixels_tab[0][1] = avg_pixels16_x2_mmx2;
3165 c->avg_pixels_tab[0][2] = avg_pixels16_y2_mmx2;
3167 c->put_pixels_tab[1][1] = put_pixels8_x2_mmx2;
3168 c->put_pixels_tab[1][2] = put_pixels8_y2_mmx2;
3170 c->avg_pixels_tab[1][0] = avg_pixels8_mmx2;
3171 c->avg_pixels_tab[1][1] = avg_pixels8_x2_mmx2;
3172 c->avg_pixels_tab[1][2] = avg_pixels8_y2_mmx2;
3174 #ifdef CONFIG_ENCODERS
3175 c->hadamard8_diff[0]= hadamard8_diff16_mmx2;
3176 c->hadamard8_diff[1]= hadamard8_diff_mmx2;
3177 c->vsad[4]= vsad_intra16_mmx2;
3178 #endif //CONFIG_ENCODERS
3180 c->h264_idct_dc_add= ff_h264_idct_dc_add_mmx2;
3181 c->h264_idct8_dc_add= ff_h264_idct8_dc_add_mmx2;
3183 if(!(avctx->flags & CODEC_FLAG_BITEXACT)){
3184 c->put_no_rnd_pixels_tab[0][1] = put_no_rnd_pixels16_x2_mmx2;
3185 c->put_no_rnd_pixels_tab[0][2] = put_no_rnd_pixels16_y2_mmx2;
3186 c->put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_mmx2;
3187 c->put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_mmx2;
3188 c->avg_pixels_tab[0][3] = avg_pixels16_xy2_mmx2;
3189 c->avg_pixels_tab[1][3] = avg_pixels8_xy2_mmx2;
3190 #ifdef CONFIG_ENCODERS
3191 c->vsad[0] = vsad16_mmx2;
3192 #endif //CONFIG_ENCODERS
3196 SET_QPEL_FUNC(qpel_pixels_tab[0][ 0], qpel16_mc00_mmx2)
3197 SET_QPEL_FUNC(qpel_pixels_tab[0][ 1], qpel16_mc10_mmx2)
3198 SET_QPEL_FUNC(qpel_pixels_tab[0][ 2], qpel16_mc20_mmx2)
3199 SET_QPEL_FUNC(qpel_pixels_tab[0][ 3], qpel16_mc30_mmx2)
3200 SET_QPEL_FUNC(qpel_pixels_tab[0][ 4], qpel16_mc01_mmx2)
3201 SET_QPEL_FUNC(qpel_pixels_tab[0][ 5], qpel16_mc11_mmx2)
3202 SET_QPEL_FUNC(qpel_pixels_tab[0][ 6], qpel16_mc21_mmx2)
3203 SET_QPEL_FUNC(qpel_pixels_tab[0][ 7], qpel16_mc31_mmx2)
3204 SET_QPEL_FUNC(qpel_pixels_tab[0][ 8], qpel16_mc02_mmx2)
3205 SET_QPEL_FUNC(qpel_pixels_tab[0][ 9], qpel16_mc12_mmx2)
3206 SET_QPEL_FUNC(qpel_pixels_tab[0][10], qpel16_mc22_mmx2)
3207 SET_QPEL_FUNC(qpel_pixels_tab[0][11], qpel16_mc32_mmx2)
3208 SET_QPEL_FUNC(qpel_pixels_tab[0][12], qpel16_mc03_mmx2)
3209 SET_QPEL_FUNC(qpel_pixels_tab[0][13], qpel16_mc13_mmx2)
3210 SET_QPEL_FUNC(qpel_pixels_tab[0][14], qpel16_mc23_mmx2)
3211 SET_QPEL_FUNC(qpel_pixels_tab[0][15], qpel16_mc33_mmx2)
3212 SET_QPEL_FUNC(qpel_pixels_tab[1][ 0], qpel8_mc00_mmx2)
3213 SET_QPEL_FUNC(qpel_pixels_tab[1][ 1], qpel8_mc10_mmx2)
3214 SET_QPEL_FUNC(qpel_pixels_tab[1][ 2], qpel8_mc20_mmx2)
3215 SET_QPEL_FUNC(qpel_pixels_tab[1][ 3], qpel8_mc30_mmx2)
3216 SET_QPEL_FUNC(qpel_pixels_tab[1][ 4], qpel8_mc01_mmx2)
3217 SET_QPEL_FUNC(qpel_pixels_tab[1][ 5], qpel8_mc11_mmx2)
3218 SET_QPEL_FUNC(qpel_pixels_tab[1][ 6], qpel8_mc21_mmx2)
3219 SET_QPEL_FUNC(qpel_pixels_tab[1][ 7], qpel8_mc31_mmx2)
3220 SET_QPEL_FUNC(qpel_pixels_tab[1][ 8], qpel8_mc02_mmx2)
3221 SET_QPEL_FUNC(qpel_pixels_tab[1][ 9], qpel8_mc12_mmx2)
3222 SET_QPEL_FUNC(qpel_pixels_tab[1][10], qpel8_mc22_mmx2)
3223 SET_QPEL_FUNC(qpel_pixels_tab[1][11], qpel8_mc32_mmx2)
3224 SET_QPEL_FUNC(qpel_pixels_tab[1][12], qpel8_mc03_mmx2)
3225 SET_QPEL_FUNC(qpel_pixels_tab[1][13], qpel8_mc13_mmx2)
3226 SET_QPEL_FUNC(qpel_pixels_tab[1][14], qpel8_mc23_mmx2)
3227 SET_QPEL_FUNC(qpel_pixels_tab[1][15], qpel8_mc33_mmx2)
3231 #define dspfunc(PFX, IDX, NUM) \
3232 c->PFX ## _pixels_tab[IDX][ 0] = PFX ## NUM ## _mc00_mmx2; \
3233 c->PFX ## _pixels_tab[IDX][ 1] = PFX ## NUM ## _mc10_mmx2; \
3234 c->PFX ## _pixels_tab[IDX][ 2] = PFX ## NUM ## _mc20_mmx2; \
3235 c->PFX ## _pixels_tab[IDX][ 3] = PFX ## NUM ## _mc30_mmx2; \
3236 c->PFX ## _pixels_tab[IDX][ 4] = PFX ## NUM ## _mc01_mmx2; \
3237 c->PFX ## _pixels_tab[IDX][ 5] = PFX ## NUM ## _mc11_mmx2; \
3238 c->PFX ## _pixels_tab[IDX][ 6] = PFX ## NUM ## _mc21_mmx2; \
3239 c->PFX ## _pixels_tab[IDX][ 7] = PFX ## NUM ## _mc31_mmx2; \
3240 c->PFX ## _pixels_tab[IDX][ 8] = PFX ## NUM ## _mc02_mmx2; \
3241 c->PFX ## _pixels_tab[IDX][ 9] = PFX ## NUM ## _mc12_mmx2; \
3242 c->PFX ## _pixels_tab[IDX][10] = PFX ## NUM ## _mc22_mmx2; \
3243 c->PFX ## _pixels_tab[IDX][11] = PFX ## NUM ## _mc32_mmx2; \
3244 c->PFX ## _pixels_tab[IDX][12] = PFX ## NUM ## _mc03_mmx2; \
3245 c->PFX ## _pixels_tab[IDX][13] = PFX ## NUM ## _mc13_mmx2; \
3246 c->PFX ## _pixels_tab[IDX][14] = PFX ## NUM ## _mc23_mmx2; \
3247 c->PFX ## _pixels_tab[IDX][15] = PFX ## NUM ## _mc33_mmx2
3249 dspfunc(put_h264_qpel, 0, 16);
3250 dspfunc(put_h264_qpel, 1, 8);
3251 dspfunc(put_h264_qpel, 2, 4);
3252 dspfunc(avg_h264_qpel, 0, 16);
3253 dspfunc(avg_h264_qpel, 1, 8);
3254 dspfunc(avg_h264_qpel, 2, 4);
3257 c->avg_h264_chroma_pixels_tab[0]= avg_h264_chroma_mc8_mmx2;
3258 c->avg_h264_chroma_pixels_tab[1]= avg_h264_chroma_mc4_mmx2;
3259 c->avg_h264_chroma_pixels_tab[2]= avg_h264_chroma_mc2_mmx2;
3260 c->put_h264_chroma_pixels_tab[2]= put_h264_chroma_mc2_mmx2;
3261 c->h264_v_loop_filter_luma= h264_v_loop_filter_luma_mmx2;
3262 c->h264_h_loop_filter_luma= h264_h_loop_filter_luma_mmx2;
3263 c->h264_v_loop_filter_chroma= h264_v_loop_filter_chroma_mmx2;
3264 c->h264_h_loop_filter_chroma= h264_h_loop_filter_chroma_mmx2;
3265 c->h264_v_loop_filter_chroma_intra= h264_v_loop_filter_chroma_intra_mmx2;
3266 c->h264_h_loop_filter_chroma_intra= h264_h_loop_filter_chroma_intra_mmx2;
3268 c->weight_h264_pixels_tab[0]= ff_h264_weight_16x16_mmx2;
3269 c->weight_h264_pixels_tab[1]= ff_h264_weight_16x8_mmx2;
3270 c->weight_h264_pixels_tab[2]= ff_h264_weight_8x16_mmx2;
3271 c->weight_h264_pixels_tab[3]= ff_h264_weight_8x8_mmx2;
3272 c->weight_h264_pixels_tab[4]= ff_h264_weight_8x4_mmx2;
3273 c->weight_h264_pixels_tab[5]= ff_h264_weight_4x8_mmx2;
3274 c->weight_h264_pixels_tab[6]= ff_h264_weight_4x4_mmx2;
3275 c->weight_h264_pixels_tab[7]= ff_h264_weight_4x2_mmx2;
3277 c->biweight_h264_pixels_tab[0]= ff_h264_biweight_16x16_mmx2;
3278 c->biweight_h264_pixels_tab[1]= ff_h264_biweight_16x8_mmx2;
3279 c->biweight_h264_pixels_tab[2]= ff_h264_biweight_8x16_mmx2;
3280 c->biweight_h264_pixels_tab[3]= ff_h264_biweight_8x8_mmx2;
3281 c->biweight_h264_pixels_tab[4]= ff_h264_biweight_8x4_mmx2;
3282 c->biweight_h264_pixels_tab[5]= ff_h264_biweight_4x8_mmx2;
3283 c->biweight_h264_pixels_tab[6]= ff_h264_biweight_4x4_mmx2;
3284 c->biweight_h264_pixels_tab[7]= ff_h264_biweight_4x2_mmx2;
3286 #ifdef CONFIG_CAVS_DECODER
3287 ff_cavsdsp_init_mmx2(c, avctx);
3290 #ifdef CONFIG_ENCODERS
3291 c->sub_hfyu_median_prediction= sub_hfyu_median_prediction_mmx2;
3292 #endif //CONFIG_ENCODERS
3293 } else if (mm_flags & MM_3DNOW) {
3294 c->prefetch = prefetch_3dnow;
3296 c->put_pixels_tab[0][1] = put_pixels16_x2_3dnow;
3297 c->put_pixels_tab[0][2] = put_pixels16_y2_3dnow;
3299 c->avg_pixels_tab[0][0] = avg_pixels16_3dnow;
3300 c->avg_pixels_tab[0][1] = avg_pixels16_x2_3dnow;
3301 c->avg_pixels_tab[0][2] = avg_pixels16_y2_3dnow;
3303 c->put_pixels_tab[1][1] = put_pixels8_x2_3dnow;
3304 c->put_pixels_tab[1][2] = put_pixels8_y2_3dnow;
3306 c->avg_pixels_tab[1][0] = avg_pixels8_3dnow;
3307 c->avg_pixels_tab[1][1] = avg_pixels8_x2_3dnow;
3308 c->avg_pixels_tab[1][2] = avg_pixels8_y2_3dnow;
3310 if(!(avctx->flags & CODEC_FLAG_BITEXACT)){
3311 c->put_no_rnd_pixels_tab[0][1] = put_no_rnd_pixels16_x2_3dnow;
3312 c->put_no_rnd_pixels_tab[0][2] = put_no_rnd_pixels16_y2_3dnow;
3313 c->put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_3dnow;
3314 c->put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_3dnow;
3315 c->avg_pixels_tab[0][3] = avg_pixels16_xy2_3dnow;
3316 c->avg_pixels_tab[1][3] = avg_pixels8_xy2_3dnow;
3319 SET_QPEL_FUNC(qpel_pixels_tab[0][ 0], qpel16_mc00_3dnow)
3320 SET_QPEL_FUNC(qpel_pixels_tab[0][ 1], qpel16_mc10_3dnow)
3321 SET_QPEL_FUNC(qpel_pixels_tab[0][ 2], qpel16_mc20_3dnow)
3322 SET_QPEL_FUNC(qpel_pixels_tab[0][ 3], qpel16_mc30_3dnow)
3323 SET_QPEL_FUNC(qpel_pixels_tab[0][ 4], qpel16_mc01_3dnow)
3324 SET_QPEL_FUNC(qpel_pixels_tab[0][ 5], qpel16_mc11_3dnow)
3325 SET_QPEL_FUNC(qpel_pixels_tab[0][ 6], qpel16_mc21_3dnow)
3326 SET_QPEL_FUNC(qpel_pixels_tab[0][ 7], qpel16_mc31_3dnow)
3327 SET_QPEL_FUNC(qpel_pixels_tab[0][ 8], qpel16_mc02_3dnow)
3328 SET_QPEL_FUNC(qpel_pixels_tab[0][ 9], qpel16_mc12_3dnow)
3329 SET_QPEL_FUNC(qpel_pixels_tab[0][10], qpel16_mc22_3dnow)
3330 SET_QPEL_FUNC(qpel_pixels_tab[0][11], qpel16_mc32_3dnow)
3331 SET_QPEL_FUNC(qpel_pixels_tab[0][12], qpel16_mc03_3dnow)
3332 SET_QPEL_FUNC(qpel_pixels_tab[0][13], qpel16_mc13_3dnow)
3333 SET_QPEL_FUNC(qpel_pixels_tab[0][14], qpel16_mc23_3dnow)
3334 SET_QPEL_FUNC(qpel_pixels_tab[0][15], qpel16_mc33_3dnow)
3335 SET_QPEL_FUNC(qpel_pixels_tab[1][ 0], qpel8_mc00_3dnow)
3336 SET_QPEL_FUNC(qpel_pixels_tab[1][ 1], qpel8_mc10_3dnow)
3337 SET_QPEL_FUNC(qpel_pixels_tab[1][ 2], qpel8_mc20_3dnow)
3338 SET_QPEL_FUNC(qpel_pixels_tab[1][ 3], qpel8_mc30_3dnow)
3339 SET_QPEL_FUNC(qpel_pixels_tab[1][ 4], qpel8_mc01_3dnow)
3340 SET_QPEL_FUNC(qpel_pixels_tab[1][ 5], qpel8_mc11_3dnow)
3341 SET_QPEL_FUNC(qpel_pixels_tab[1][ 6], qpel8_mc21_3dnow)
3342 SET_QPEL_FUNC(qpel_pixels_tab[1][ 7], qpel8_mc31_3dnow)
3343 SET_QPEL_FUNC(qpel_pixels_tab[1][ 8], qpel8_mc02_3dnow)
3344 SET_QPEL_FUNC(qpel_pixels_tab[1][ 9], qpel8_mc12_3dnow)
3345 SET_QPEL_FUNC(qpel_pixels_tab[1][10], qpel8_mc22_3dnow)
3346 SET_QPEL_FUNC(qpel_pixels_tab[1][11], qpel8_mc32_3dnow)
3347 SET_QPEL_FUNC(qpel_pixels_tab[1][12], qpel8_mc03_3dnow)
3348 SET_QPEL_FUNC(qpel_pixels_tab[1][13], qpel8_mc13_3dnow)
3349 SET_QPEL_FUNC(qpel_pixels_tab[1][14], qpel8_mc23_3dnow)
3350 SET_QPEL_FUNC(qpel_pixels_tab[1][15], qpel8_mc33_3dnow)
3352 #define dspfunc(PFX, IDX, NUM) \
3353 c->PFX ## _pixels_tab[IDX][ 0] = PFX ## NUM ## _mc00_3dnow; \
3354 c->PFX ## _pixels_tab[IDX][ 1] = PFX ## NUM ## _mc10_3dnow; \
3355 c->PFX ## _pixels_tab[IDX][ 2] = PFX ## NUM ## _mc20_3dnow; \
3356 c->PFX ## _pixels_tab[IDX][ 3] = PFX ## NUM ## _mc30_3dnow; \
3357 c->PFX ## _pixels_tab[IDX][ 4] = PFX ## NUM ## _mc01_3dnow; \
3358 c->PFX ## _pixels_tab[IDX][ 5] = PFX ## NUM ## _mc11_3dnow; \
3359 c->PFX ## _pixels_tab[IDX][ 6] = PFX ## NUM ## _mc21_3dnow; \
3360 c->PFX ## _pixels_tab[IDX][ 7] = PFX ## NUM ## _mc31_3dnow; \
3361 c->PFX ## _pixels_tab[IDX][ 8] = PFX ## NUM ## _mc02_3dnow; \
3362 c->PFX ## _pixels_tab[IDX][ 9] = PFX ## NUM ## _mc12_3dnow; \
3363 c->PFX ## _pixels_tab[IDX][10] = PFX ## NUM ## _mc22_3dnow; \
3364 c->PFX ## _pixels_tab[IDX][11] = PFX ## NUM ## _mc32_3dnow; \
3365 c->PFX ## _pixels_tab[IDX][12] = PFX ## NUM ## _mc03_3dnow; \
3366 c->PFX ## _pixels_tab[IDX][13] = PFX ## NUM ## _mc13_3dnow; \
3367 c->PFX ## _pixels_tab[IDX][14] = PFX ## NUM ## _mc23_3dnow; \
3368 c->PFX ## _pixels_tab[IDX][15] = PFX ## NUM ## _mc33_3dnow
3370 dspfunc(put_h264_qpel, 0, 16);
3371 dspfunc(put_h264_qpel, 1, 8);
3372 dspfunc(put_h264_qpel, 2, 4);
3373 dspfunc(avg_h264_qpel, 0, 16);
3374 dspfunc(avg_h264_qpel, 1, 8);
3375 dspfunc(avg_h264_qpel, 2, 4);
3377 c->avg_h264_chroma_pixels_tab[0]= avg_h264_chroma_mc8_3dnow;
3378 c->avg_h264_chroma_pixels_tab[1]= avg_h264_chroma_mc4_3dnow;
3381 #ifdef CONFIG_SNOW_ENCODER
3382 if(mm_flags & MM_SSE2){
3383 c->horizontal_compose97i = ff_snow_horizontal_compose97i_sse2;
3384 c->vertical_compose97i = ff_snow_vertical_compose97i_sse2;
3385 c->inner_add_yblock = ff_snow_inner_add_yblock_sse2;
3388 c->horizontal_compose97i = ff_snow_horizontal_compose97i_mmx;
3389 c->vertical_compose97i = ff_snow_vertical_compose97i_mmx;
3390 c->inner_add_yblock = ff_snow_inner_add_yblock_mmx;
3394 if(mm_flags & MM_3DNOW){
3395 c->vorbis_inverse_coupling = vorbis_inverse_coupling_3dnow;
3396 c->vector_fmul = vector_fmul_3dnow;
3397 if(!(avctx->flags & CODEC_FLAG_BITEXACT))
3398 c->float_to_int16 = float_to_int16_3dnow;
3400 if(mm_flags & MM_3DNOWEXT)
3401 c->vector_fmul_reverse = vector_fmul_reverse_3dnow2;
3402 if(mm_flags & MM_SSE){
3403 c->vorbis_inverse_coupling = vorbis_inverse_coupling_sse;
3404 c->vector_fmul = vector_fmul_sse;
3405 c->float_to_int16 = float_to_int16_sse;
3407 if(mm_flags & MM_SSE2){
3408 c->vector_fmul_reverse = vector_fmul_reverse_sse2;
3409 c->vector_fmul_add_add = vector_fmul_add_add_sse2;
3411 if(mm_flags & MM_3DNOW)
3412 c->vector_fmul_add_add = vector_fmul_add_add_3dnow; // faster than sse2
3415 #ifdef CONFIG_ENCODERS
3416 dsputil_init_pix_mmx(c, avctx);
3417 #endif //CONFIG_ENCODERS
3419 // for speed testing
3420 get_pixels = just_return;
3421 put_pixels_clamped = just_return;
3422 add_pixels_clamped = just_return;
3424 pix_abs16x16 = just_return;
3425 pix_abs16x16_x2 = just_return;
3426 pix_abs16x16_y2 = just_return;
3427 pix_abs16x16_xy2 = just_return;
3429 put_pixels_tab[0] = just_return;
3430 put_pixels_tab[1] = just_return;
3431 put_pixels_tab[2] = just_return;
3432 put_pixels_tab[3] = just_return;
3434 put_no_rnd_pixels_tab[0] = just_return;
3435 put_no_rnd_pixels_tab[1] = just_return;
3436 put_no_rnd_pixels_tab[2] = just_return;
3437 put_no_rnd_pixels_tab[3] = just_return;
3439 avg_pixels_tab[0] = just_return;
3440 avg_pixels_tab[1] = just_return;
3441 avg_pixels_tab[2] = just_return;
3442 avg_pixels_tab[3] = just_return;
3444 avg_no_rnd_pixels_tab[0] = just_return;
3445 avg_no_rnd_pixels_tab[1] = just_return;
3446 avg_no_rnd_pixels_tab[2] = just_return;
3447 avg_no_rnd_pixels_tab[3] = just_return;
3449 //av_fdct = just_return;
3450 //ff_idct = just_return;