2 * software RGB to RGB converter
3 * pluralize by software PAL8 to RGB converter
4 * software YUV to YUV converter
5 * software YUV to RGB converter
6 * Written by Nick Kurshev.
7 * palette & YUV & runtime CPU stuff by Michael (michaelni@gmx.at)
8 * lot of big-endian byte order fixes by Alex Beregszaszi
10 * This file is part of FFmpeg.
12 * FFmpeg is free software; you can redistribute it and/or
13 * modify it under the terms of the GNU Lesser General Public
14 * License as published by the Free Software Foundation; either
15 * version 2.1 of the License, or (at your option) any later version.
17 * FFmpeg is distributed in the hope that it will be useful,
18 * but WITHOUT ANY WARRANTY; without even the implied warranty of
19 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
20 * Lesser General Public License for more details.
22 * You should have received a copy of the GNU Lesser General Public
23 * License along with FFmpeg; if not, write to the Free Software
24 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
35 #if COMPILE_TEMPLATE_AMD3DNOW
36 #define PREFETCH "prefetch"
37 #define PAVGB "pavgusb"
38 #elif COMPILE_TEMPLATE_MMXEXT
39 #define PREFETCH "prefetchnta"
42 #define PREFETCH " # nop"
45 #if COMPILE_TEMPLATE_AMD3DNOW
46 /* On K6 femms is faster than emms. On K7 femms is directly mapped to emms. */
52 #if COMPILE_TEMPLATE_MMXEXT
53 #define MOVNTQ "movntq"
54 #define SFENCE "sfence"
57 #define SFENCE " # nop"
60 #if !COMPILE_TEMPLATE_SSE2
62 #if !COMPILE_TEMPLATE_AMD3DNOW
64 static inline void RENAME(rgb24tobgr32)(const uint8_t *src, uint8_t *dst, int src_size)
67 const uint8_t *s = src;
69 const uint8_t *mm_end;
71 __asm__ volatile(PREFETCH" %0"::"m"(*s):"memory");
73 __asm__ volatile("movq %0, %%mm7"::"m"(mask32a):"memory");
76 PREFETCH" 32(%1) \n\t"
77 "movd (%1), %%mm0 \n\t"
78 "punpckldq 3(%1), %%mm0 \n\t"
79 "movd 6(%1), %%mm1 \n\t"
80 "punpckldq 9(%1), %%mm1 \n\t"
81 "movd 12(%1), %%mm2 \n\t"
82 "punpckldq 15(%1), %%mm2 \n\t"
83 "movd 18(%1), %%mm3 \n\t"
84 "punpckldq 21(%1), %%mm3 \n\t"
85 "por %%mm7, %%mm0 \n\t"
86 "por %%mm7, %%mm1 \n\t"
87 "por %%mm7, %%mm2 \n\t"
88 "por %%mm7, %%mm3 \n\t"
89 MOVNTQ" %%mm0, (%0) \n\t"
90 MOVNTQ" %%mm1, 8(%0) \n\t"
91 MOVNTQ" %%mm2, 16(%0) \n\t"
92 MOVNTQ" %%mm3, 24(%0)"
98 __asm__ volatile(SFENCE:::"memory");
99 __asm__ volatile(EMMS:::"memory");
108 #define STORE_BGR24_MMX \
109 "psrlq $8, %%mm2 \n\t" \
110 "psrlq $8, %%mm3 \n\t" \
111 "psrlq $8, %%mm6 \n\t" \
112 "psrlq $8, %%mm7 \n\t" \
113 "pand "MANGLE(mask24l)", %%mm0\n\t" \
114 "pand "MANGLE(mask24l)", %%mm1\n\t" \
115 "pand "MANGLE(mask24l)", %%mm4\n\t" \
116 "pand "MANGLE(mask24l)", %%mm5\n\t" \
117 "pand "MANGLE(mask24h)", %%mm2\n\t" \
118 "pand "MANGLE(mask24h)", %%mm3\n\t" \
119 "pand "MANGLE(mask24h)", %%mm6\n\t" \
120 "pand "MANGLE(mask24h)", %%mm7\n\t" \
121 "por %%mm2, %%mm0 \n\t" \
122 "por %%mm3, %%mm1 \n\t" \
123 "por %%mm6, %%mm4 \n\t" \
124 "por %%mm7, %%mm5 \n\t" \
126 "movq %%mm1, %%mm2 \n\t" \
127 "movq %%mm4, %%mm3 \n\t" \
128 "psllq $48, %%mm2 \n\t" \
129 "psllq $32, %%mm3 \n\t" \
130 "pand "MANGLE(mask24hh)", %%mm2\n\t" \
131 "pand "MANGLE(mask24hhh)", %%mm3\n\t" \
132 "por %%mm2, %%mm0 \n\t" \
133 "psrlq $16, %%mm1 \n\t" \
134 "psrlq $32, %%mm4 \n\t" \
135 "psllq $16, %%mm5 \n\t" \
136 "por %%mm3, %%mm1 \n\t" \
137 "pand "MANGLE(mask24hhhh)", %%mm5\n\t" \
138 "por %%mm5, %%mm4 \n\t" \
140 MOVNTQ" %%mm0, (%0) \n\t" \
141 MOVNTQ" %%mm1, 8(%0) \n\t" \
142 MOVNTQ" %%mm4, 16(%0)"
145 static inline void RENAME(rgb32tobgr24)(const uint8_t *src, uint8_t *dst, int src_size)
148 const uint8_t *s = src;
150 const uint8_t *mm_end;
152 __asm__ volatile(PREFETCH" %0"::"m"(*s):"memory");
156 PREFETCH" 32(%1) \n\t"
157 "movq (%1), %%mm0 \n\t"
158 "movq 8(%1), %%mm1 \n\t"
159 "movq 16(%1), %%mm4 \n\t"
160 "movq 24(%1), %%mm5 \n\t"
161 "movq %%mm0, %%mm2 \n\t"
162 "movq %%mm1, %%mm3 \n\t"
163 "movq %%mm4, %%mm6 \n\t"
164 "movq %%mm5, %%mm7 \n\t"
171 __asm__ volatile(SFENCE:::"memory");
172 __asm__ volatile(EMMS:::"memory");
182 original by Strepto/Astral
183 ported to gcc & bugfixed: A'rpi
184 MMX2, 3DNOW optimization by Nick Kurshev
185 32-bit C version, and and&add trick by Michael Niedermayer
187 static inline void RENAME(rgb15to16)(const uint8_t *src, uint8_t *dst, int src_size)
189 register const uint8_t* s=src;
190 register uint8_t* d=dst;
191 register const uint8_t *end;
192 const uint8_t *mm_end;
194 __asm__ volatile(PREFETCH" %0"::"m"(*s));
195 __asm__ volatile("movq %0, %%mm4"::"m"(mask15s));
199 PREFETCH" 32(%1) \n\t"
200 "movq (%1), %%mm0 \n\t"
201 "movq 8(%1), %%mm2 \n\t"
202 "movq %%mm0, %%mm1 \n\t"
203 "movq %%mm2, %%mm3 \n\t"
204 "pand %%mm4, %%mm0 \n\t"
205 "pand %%mm4, %%mm2 \n\t"
206 "paddw %%mm1, %%mm0 \n\t"
207 "paddw %%mm3, %%mm2 \n\t"
208 MOVNTQ" %%mm0, (%0) \n\t"
209 MOVNTQ" %%mm2, 8(%0)"
215 __asm__ volatile(SFENCE:::"memory");
216 __asm__ volatile(EMMS:::"memory");
219 register unsigned x= *((const uint32_t *)s);
220 *((uint32_t *)d) = (x&0x7FFF7FFF) + (x&0x7FE07FE0);
225 register unsigned short x= *((const uint16_t *)s);
226 *((uint16_t *)d) = (x&0x7FFF) + (x&0x7FE0);
230 static inline void RENAME(rgb16to15)(const uint8_t *src, uint8_t *dst, int src_size)
232 register const uint8_t* s=src;
233 register uint8_t* d=dst;
234 register const uint8_t *end;
235 const uint8_t *mm_end;
237 __asm__ volatile(PREFETCH" %0"::"m"(*s));
238 __asm__ volatile("movq %0, %%mm7"::"m"(mask15rg));
239 __asm__ volatile("movq %0, %%mm6"::"m"(mask15b));
243 PREFETCH" 32(%1) \n\t"
244 "movq (%1), %%mm0 \n\t"
245 "movq 8(%1), %%mm2 \n\t"
246 "movq %%mm0, %%mm1 \n\t"
247 "movq %%mm2, %%mm3 \n\t"
248 "psrlq $1, %%mm0 \n\t"
249 "psrlq $1, %%mm2 \n\t"
250 "pand %%mm7, %%mm0 \n\t"
251 "pand %%mm7, %%mm2 \n\t"
252 "pand %%mm6, %%mm1 \n\t"
253 "pand %%mm6, %%mm3 \n\t"
254 "por %%mm1, %%mm0 \n\t"
255 "por %%mm3, %%mm2 \n\t"
256 MOVNTQ" %%mm0, (%0) \n\t"
257 MOVNTQ" %%mm2, 8(%0)"
263 __asm__ volatile(SFENCE:::"memory");
264 __asm__ volatile(EMMS:::"memory");
267 register uint32_t x= *((const uint32_t*)s);
268 *((uint32_t *)d) = ((x>>1)&0x7FE07FE0) | (x&0x001F001F);
273 register uint16_t x= *((const uint16_t*)s);
274 *((uint16_t *)d) = ((x>>1)&0x7FE0) | (x&0x001F);
278 static inline void RENAME(rgb32to16)(const uint8_t *src, uint8_t *dst, int src_size)
280 const uint8_t *s = src;
282 const uint8_t *mm_end;
283 uint16_t *d = (uint16_t *)dst;
287 "movq %3, %%mm5 \n\t"
288 "movq %4, %%mm6 \n\t"
289 "movq %5, %%mm7 \n\t"
293 PREFETCH" 32(%1) \n\t"
294 "movd (%1), %%mm0 \n\t"
295 "movd 4(%1), %%mm3 \n\t"
296 "punpckldq 8(%1), %%mm0 \n\t"
297 "punpckldq 12(%1), %%mm3 \n\t"
298 "movq %%mm0, %%mm1 \n\t"
299 "movq %%mm3, %%mm4 \n\t"
300 "pand %%mm6, %%mm0 \n\t"
301 "pand %%mm6, %%mm3 \n\t"
302 "pmaddwd %%mm7, %%mm0 \n\t"
303 "pmaddwd %%mm7, %%mm3 \n\t"
304 "pand %%mm5, %%mm1 \n\t"
305 "pand %%mm5, %%mm4 \n\t"
306 "por %%mm1, %%mm0 \n\t"
307 "por %%mm4, %%mm3 \n\t"
308 "psrld $5, %%mm0 \n\t"
309 "pslld $11, %%mm3 \n\t"
310 "por %%mm3, %%mm0 \n\t"
311 MOVNTQ" %%mm0, (%0) \n\t"
318 : "r" (mm_end), "m" (mask3216g), "m" (mask3216br), "m" (mul3216)
320 __asm__ volatile(SFENCE:::"memory");
321 __asm__ volatile(EMMS:::"memory");
323 register int rgb = *(const uint32_t*)s; s += 4;
324 *d++ = ((rgb&0xFF)>>3) + ((rgb&0xFC00)>>5) + ((rgb&0xF80000)>>8);
328 static inline void RENAME(rgb32tobgr16)(const uint8_t *src, uint8_t *dst, int src_size)
330 const uint8_t *s = src;
332 const uint8_t *mm_end;
333 uint16_t *d = (uint16_t *)dst;
335 __asm__ volatile(PREFETCH" %0"::"m"(*src):"memory");
337 "movq %0, %%mm7 \n\t"
338 "movq %1, %%mm6 \n\t"
339 ::"m"(red_16mask),"m"(green_16mask));
343 PREFETCH" 32(%1) \n\t"
344 "movd (%1), %%mm0 \n\t"
345 "movd 4(%1), %%mm3 \n\t"
346 "punpckldq 8(%1), %%mm0 \n\t"
347 "punpckldq 12(%1), %%mm3 \n\t"
348 "movq %%mm0, %%mm1 \n\t"
349 "movq %%mm0, %%mm2 \n\t"
350 "movq %%mm3, %%mm4 \n\t"
351 "movq %%mm3, %%mm5 \n\t"
352 "psllq $8, %%mm0 \n\t"
353 "psllq $8, %%mm3 \n\t"
354 "pand %%mm7, %%mm0 \n\t"
355 "pand %%mm7, %%mm3 \n\t"
356 "psrlq $5, %%mm1 \n\t"
357 "psrlq $5, %%mm4 \n\t"
358 "pand %%mm6, %%mm1 \n\t"
359 "pand %%mm6, %%mm4 \n\t"
360 "psrlq $19, %%mm2 \n\t"
361 "psrlq $19, %%mm5 \n\t"
362 "pand %2, %%mm2 \n\t"
363 "pand %2, %%mm5 \n\t"
364 "por %%mm1, %%mm0 \n\t"
365 "por %%mm4, %%mm3 \n\t"
366 "por %%mm2, %%mm0 \n\t"
367 "por %%mm5, %%mm3 \n\t"
368 "psllq $16, %%mm3 \n\t"
369 "por %%mm3, %%mm0 \n\t"
370 MOVNTQ" %%mm0, (%0) \n\t"
371 :: "r"(d),"r"(s),"m"(blue_16mask):"memory");
375 __asm__ volatile(SFENCE:::"memory");
376 __asm__ volatile(EMMS:::"memory");
378 register int rgb = *(const uint32_t*)s; s += 4;
379 *d++ = ((rgb&0xF8)<<8) + ((rgb&0xFC00)>>5) + ((rgb&0xF80000)>>19);
383 static inline void RENAME(rgb32to15)(const uint8_t *src, uint8_t *dst, int src_size)
385 const uint8_t *s = src;
387 const uint8_t *mm_end;
388 uint16_t *d = (uint16_t *)dst;
392 "movq %3, %%mm5 \n\t"
393 "movq %4, %%mm6 \n\t"
394 "movq %5, %%mm7 \n\t"
398 PREFETCH" 32(%1) \n\t"
399 "movd (%1), %%mm0 \n\t"
400 "movd 4(%1), %%mm3 \n\t"
401 "punpckldq 8(%1), %%mm0 \n\t"
402 "punpckldq 12(%1), %%mm3 \n\t"
403 "movq %%mm0, %%mm1 \n\t"
404 "movq %%mm3, %%mm4 \n\t"
405 "pand %%mm6, %%mm0 \n\t"
406 "pand %%mm6, %%mm3 \n\t"
407 "pmaddwd %%mm7, %%mm0 \n\t"
408 "pmaddwd %%mm7, %%mm3 \n\t"
409 "pand %%mm5, %%mm1 \n\t"
410 "pand %%mm5, %%mm4 \n\t"
411 "por %%mm1, %%mm0 \n\t"
412 "por %%mm4, %%mm3 \n\t"
413 "psrld $6, %%mm0 \n\t"
414 "pslld $10, %%mm3 \n\t"
415 "por %%mm3, %%mm0 \n\t"
416 MOVNTQ" %%mm0, (%0) \n\t"
423 : "r" (mm_end), "m" (mask3215g), "m" (mask3216br), "m" (mul3215)
425 __asm__ volatile(SFENCE:::"memory");
426 __asm__ volatile(EMMS:::"memory");
428 register int rgb = *(const uint32_t*)s; s += 4;
429 *d++ = ((rgb&0xFF)>>3) + ((rgb&0xF800)>>6) + ((rgb&0xF80000)>>9);
433 static inline void RENAME(rgb32tobgr15)(const uint8_t *src, uint8_t *dst, int src_size)
435 const uint8_t *s = src;
437 const uint8_t *mm_end;
438 uint16_t *d = (uint16_t *)dst;
440 __asm__ volatile(PREFETCH" %0"::"m"(*src):"memory");
442 "movq %0, %%mm7 \n\t"
443 "movq %1, %%mm6 \n\t"
444 ::"m"(red_15mask),"m"(green_15mask));
448 PREFETCH" 32(%1) \n\t"
449 "movd (%1), %%mm0 \n\t"
450 "movd 4(%1), %%mm3 \n\t"
451 "punpckldq 8(%1), %%mm0 \n\t"
452 "punpckldq 12(%1), %%mm3 \n\t"
453 "movq %%mm0, %%mm1 \n\t"
454 "movq %%mm0, %%mm2 \n\t"
455 "movq %%mm3, %%mm4 \n\t"
456 "movq %%mm3, %%mm5 \n\t"
457 "psllq $7, %%mm0 \n\t"
458 "psllq $7, %%mm3 \n\t"
459 "pand %%mm7, %%mm0 \n\t"
460 "pand %%mm7, %%mm3 \n\t"
461 "psrlq $6, %%mm1 \n\t"
462 "psrlq $6, %%mm4 \n\t"
463 "pand %%mm6, %%mm1 \n\t"
464 "pand %%mm6, %%mm4 \n\t"
465 "psrlq $19, %%mm2 \n\t"
466 "psrlq $19, %%mm5 \n\t"
467 "pand %2, %%mm2 \n\t"
468 "pand %2, %%mm5 \n\t"
469 "por %%mm1, %%mm0 \n\t"
470 "por %%mm4, %%mm3 \n\t"
471 "por %%mm2, %%mm0 \n\t"
472 "por %%mm5, %%mm3 \n\t"
473 "psllq $16, %%mm3 \n\t"
474 "por %%mm3, %%mm0 \n\t"
475 MOVNTQ" %%mm0, (%0) \n\t"
476 ::"r"(d),"r"(s),"m"(blue_15mask):"memory");
480 __asm__ volatile(SFENCE:::"memory");
481 __asm__ volatile(EMMS:::"memory");
483 register int rgb = *(const uint32_t*)s; s += 4;
484 *d++ = ((rgb&0xF8)<<7) + ((rgb&0xF800)>>6) + ((rgb&0xF80000)>>19);
488 static inline void RENAME(rgb24tobgr16)(const uint8_t *src, uint8_t *dst, int src_size)
490 const uint8_t *s = src;
492 const uint8_t *mm_end;
493 uint16_t *d = (uint16_t *)dst;
495 __asm__ volatile(PREFETCH" %0"::"m"(*src):"memory");
497 "movq %0, %%mm7 \n\t"
498 "movq %1, %%mm6 \n\t"
499 ::"m"(red_16mask),"m"(green_16mask));
503 PREFETCH" 32(%1) \n\t"
504 "movd (%1), %%mm0 \n\t"
505 "movd 3(%1), %%mm3 \n\t"
506 "punpckldq 6(%1), %%mm0 \n\t"
507 "punpckldq 9(%1), %%mm3 \n\t"
508 "movq %%mm0, %%mm1 \n\t"
509 "movq %%mm0, %%mm2 \n\t"
510 "movq %%mm3, %%mm4 \n\t"
511 "movq %%mm3, %%mm5 \n\t"
512 "psrlq $3, %%mm0 \n\t"
513 "psrlq $3, %%mm3 \n\t"
514 "pand %2, %%mm0 \n\t"
515 "pand %2, %%mm3 \n\t"
516 "psrlq $5, %%mm1 \n\t"
517 "psrlq $5, %%mm4 \n\t"
518 "pand %%mm6, %%mm1 \n\t"
519 "pand %%mm6, %%mm4 \n\t"
520 "psrlq $8, %%mm2 \n\t"
521 "psrlq $8, %%mm5 \n\t"
522 "pand %%mm7, %%mm2 \n\t"
523 "pand %%mm7, %%mm5 \n\t"
524 "por %%mm1, %%mm0 \n\t"
525 "por %%mm4, %%mm3 \n\t"
526 "por %%mm2, %%mm0 \n\t"
527 "por %%mm5, %%mm3 \n\t"
528 "psllq $16, %%mm3 \n\t"
529 "por %%mm3, %%mm0 \n\t"
530 MOVNTQ" %%mm0, (%0) \n\t"
531 ::"r"(d),"r"(s),"m"(blue_16mask):"memory");
535 __asm__ volatile(SFENCE:::"memory");
536 __asm__ volatile(EMMS:::"memory");
541 *d++ = (b>>3) | ((g&0xFC)<<3) | ((r&0xF8)<<8);
545 static inline void RENAME(rgb24to16)(const uint8_t *src, uint8_t *dst, int src_size)
547 const uint8_t *s = src;
549 const uint8_t *mm_end;
550 uint16_t *d = (uint16_t *)dst;
552 __asm__ volatile(PREFETCH" %0"::"m"(*src):"memory");
554 "movq %0, %%mm7 \n\t"
555 "movq %1, %%mm6 \n\t"
556 ::"m"(red_16mask),"m"(green_16mask));
560 PREFETCH" 32(%1) \n\t"
561 "movd (%1), %%mm0 \n\t"
562 "movd 3(%1), %%mm3 \n\t"
563 "punpckldq 6(%1), %%mm0 \n\t"
564 "punpckldq 9(%1), %%mm3 \n\t"
565 "movq %%mm0, %%mm1 \n\t"
566 "movq %%mm0, %%mm2 \n\t"
567 "movq %%mm3, %%mm4 \n\t"
568 "movq %%mm3, %%mm5 \n\t"
569 "psllq $8, %%mm0 \n\t"
570 "psllq $8, %%mm3 \n\t"
571 "pand %%mm7, %%mm0 \n\t"
572 "pand %%mm7, %%mm3 \n\t"
573 "psrlq $5, %%mm1 \n\t"
574 "psrlq $5, %%mm4 \n\t"
575 "pand %%mm6, %%mm1 \n\t"
576 "pand %%mm6, %%mm4 \n\t"
577 "psrlq $19, %%mm2 \n\t"
578 "psrlq $19, %%mm5 \n\t"
579 "pand %2, %%mm2 \n\t"
580 "pand %2, %%mm5 \n\t"
581 "por %%mm1, %%mm0 \n\t"
582 "por %%mm4, %%mm3 \n\t"
583 "por %%mm2, %%mm0 \n\t"
584 "por %%mm5, %%mm3 \n\t"
585 "psllq $16, %%mm3 \n\t"
586 "por %%mm3, %%mm0 \n\t"
587 MOVNTQ" %%mm0, (%0) \n\t"
588 ::"r"(d),"r"(s),"m"(blue_16mask):"memory");
592 __asm__ volatile(SFENCE:::"memory");
593 __asm__ volatile(EMMS:::"memory");
598 *d++ = (b>>3) | ((g&0xFC)<<3) | ((r&0xF8)<<8);
602 static inline void RENAME(rgb24tobgr15)(const uint8_t *src, uint8_t *dst, int src_size)
604 const uint8_t *s = src;
606 const uint8_t *mm_end;
607 uint16_t *d = (uint16_t *)dst;
609 __asm__ volatile(PREFETCH" %0"::"m"(*src):"memory");
611 "movq %0, %%mm7 \n\t"
612 "movq %1, %%mm6 \n\t"
613 ::"m"(red_15mask),"m"(green_15mask));
617 PREFETCH" 32(%1) \n\t"
618 "movd (%1), %%mm0 \n\t"
619 "movd 3(%1), %%mm3 \n\t"
620 "punpckldq 6(%1), %%mm0 \n\t"
621 "punpckldq 9(%1), %%mm3 \n\t"
622 "movq %%mm0, %%mm1 \n\t"
623 "movq %%mm0, %%mm2 \n\t"
624 "movq %%mm3, %%mm4 \n\t"
625 "movq %%mm3, %%mm5 \n\t"
626 "psrlq $3, %%mm0 \n\t"
627 "psrlq $3, %%mm3 \n\t"
628 "pand %2, %%mm0 \n\t"
629 "pand %2, %%mm3 \n\t"
630 "psrlq $6, %%mm1 \n\t"
631 "psrlq $6, %%mm4 \n\t"
632 "pand %%mm6, %%mm1 \n\t"
633 "pand %%mm6, %%mm4 \n\t"
634 "psrlq $9, %%mm2 \n\t"
635 "psrlq $9, %%mm5 \n\t"
636 "pand %%mm7, %%mm2 \n\t"
637 "pand %%mm7, %%mm5 \n\t"
638 "por %%mm1, %%mm0 \n\t"
639 "por %%mm4, %%mm3 \n\t"
640 "por %%mm2, %%mm0 \n\t"
641 "por %%mm5, %%mm3 \n\t"
642 "psllq $16, %%mm3 \n\t"
643 "por %%mm3, %%mm0 \n\t"
644 MOVNTQ" %%mm0, (%0) \n\t"
645 ::"r"(d),"r"(s),"m"(blue_15mask):"memory");
649 __asm__ volatile(SFENCE:::"memory");
650 __asm__ volatile(EMMS:::"memory");
655 *d++ = (b>>3) | ((g&0xF8)<<2) | ((r&0xF8)<<7);
659 static inline void RENAME(rgb24to15)(const uint8_t *src, uint8_t *dst, int src_size)
661 const uint8_t *s = src;
663 const uint8_t *mm_end;
664 uint16_t *d = (uint16_t *)dst;
666 __asm__ volatile(PREFETCH" %0"::"m"(*src):"memory");
668 "movq %0, %%mm7 \n\t"
669 "movq %1, %%mm6 \n\t"
670 ::"m"(red_15mask),"m"(green_15mask));
674 PREFETCH" 32(%1) \n\t"
675 "movd (%1), %%mm0 \n\t"
676 "movd 3(%1), %%mm3 \n\t"
677 "punpckldq 6(%1), %%mm0 \n\t"
678 "punpckldq 9(%1), %%mm3 \n\t"
679 "movq %%mm0, %%mm1 \n\t"
680 "movq %%mm0, %%mm2 \n\t"
681 "movq %%mm3, %%mm4 \n\t"
682 "movq %%mm3, %%mm5 \n\t"
683 "psllq $7, %%mm0 \n\t"
684 "psllq $7, %%mm3 \n\t"
685 "pand %%mm7, %%mm0 \n\t"
686 "pand %%mm7, %%mm3 \n\t"
687 "psrlq $6, %%mm1 \n\t"
688 "psrlq $6, %%mm4 \n\t"
689 "pand %%mm6, %%mm1 \n\t"
690 "pand %%mm6, %%mm4 \n\t"
691 "psrlq $19, %%mm2 \n\t"
692 "psrlq $19, %%mm5 \n\t"
693 "pand %2, %%mm2 \n\t"
694 "pand %2, %%mm5 \n\t"
695 "por %%mm1, %%mm0 \n\t"
696 "por %%mm4, %%mm3 \n\t"
697 "por %%mm2, %%mm0 \n\t"
698 "por %%mm5, %%mm3 \n\t"
699 "psllq $16, %%mm3 \n\t"
700 "por %%mm3, %%mm0 \n\t"
701 MOVNTQ" %%mm0, (%0) \n\t"
702 ::"r"(d),"r"(s),"m"(blue_15mask):"memory");
706 __asm__ volatile(SFENCE:::"memory");
707 __asm__ volatile(EMMS:::"memory");
712 *d++ = (b>>3) | ((g&0xF8)<<2) | ((r&0xF8)<<7);
716 static inline void RENAME(rgb15tobgr24)(const uint8_t *src, uint8_t *dst, int src_size)
719 const uint16_t *mm_end;
721 const uint16_t *s = (const uint16_t*)src;
722 end = s + src_size/2;
723 __asm__ volatile(PREFETCH" %0"::"m"(*s):"memory");
727 PREFETCH" 32(%1) \n\t"
728 "movq (%1), %%mm0 \n\t"
729 "movq (%1), %%mm1 \n\t"
730 "movq (%1), %%mm2 \n\t"
731 "pand %2, %%mm0 \n\t"
732 "pand %3, %%mm1 \n\t"
733 "pand %4, %%mm2 \n\t"
734 "psllq $5, %%mm0 \n\t"
735 "pmulhw "MANGLE(mul15_mid)", %%mm0 \n\t"
736 "pmulhw "MANGLE(mul15_mid)", %%mm1 \n\t"
737 "pmulhw "MANGLE(mul15_hi)", %%mm2 \n\t"
738 "movq %%mm0, %%mm3 \n\t"
739 "movq %%mm1, %%mm4 \n\t"
740 "movq %%mm2, %%mm5 \n\t"
741 "punpcklwd %5, %%mm0 \n\t"
742 "punpcklwd %5, %%mm1 \n\t"
743 "punpcklwd %5, %%mm2 \n\t"
744 "punpckhwd %5, %%mm3 \n\t"
745 "punpckhwd %5, %%mm4 \n\t"
746 "punpckhwd %5, %%mm5 \n\t"
747 "psllq $8, %%mm1 \n\t"
748 "psllq $16, %%mm2 \n\t"
749 "por %%mm1, %%mm0 \n\t"
750 "por %%mm2, %%mm0 \n\t"
751 "psllq $8, %%mm4 \n\t"
752 "psllq $16, %%mm5 \n\t"
753 "por %%mm4, %%mm3 \n\t"
754 "por %%mm5, %%mm3 \n\t"
756 "movq %%mm0, %%mm6 \n\t"
757 "movq %%mm3, %%mm7 \n\t"
759 "movq 8(%1), %%mm0 \n\t"
760 "movq 8(%1), %%mm1 \n\t"
761 "movq 8(%1), %%mm2 \n\t"
762 "pand %2, %%mm0 \n\t"
763 "pand %3, %%mm1 \n\t"
764 "pand %4, %%mm2 \n\t"
765 "psllq $5, %%mm0 \n\t"
766 "pmulhw "MANGLE(mul15_mid)", %%mm0 \n\t"
767 "pmulhw "MANGLE(mul15_mid)", %%mm1 \n\t"
768 "pmulhw "MANGLE(mul15_hi)", %%mm2 \n\t"
769 "movq %%mm0, %%mm3 \n\t"
770 "movq %%mm1, %%mm4 \n\t"
771 "movq %%mm2, %%mm5 \n\t"
772 "punpcklwd %5, %%mm0 \n\t"
773 "punpcklwd %5, %%mm1 \n\t"
774 "punpcklwd %5, %%mm2 \n\t"
775 "punpckhwd %5, %%mm3 \n\t"
776 "punpckhwd %5, %%mm4 \n\t"
777 "punpckhwd %5, %%mm5 \n\t"
778 "psllq $8, %%mm1 \n\t"
779 "psllq $16, %%mm2 \n\t"
780 "por %%mm1, %%mm0 \n\t"
781 "por %%mm2, %%mm0 \n\t"
782 "psllq $8, %%mm4 \n\t"
783 "psllq $16, %%mm5 \n\t"
784 "por %%mm4, %%mm3 \n\t"
785 "por %%mm5, %%mm3 \n\t"
788 :"r"(s),"m"(mask15b),"m"(mask15g),"m"(mask15r), "m"(mmx_null)
790 /* borrowed 32 to 24 */
792 "movq %%mm0, %%mm4 \n\t"
793 "movq %%mm3, %%mm5 \n\t"
794 "movq %%mm6, %%mm0 \n\t"
795 "movq %%mm7, %%mm1 \n\t"
797 "movq %%mm4, %%mm6 \n\t"
798 "movq %%mm5, %%mm7 \n\t"
799 "movq %%mm0, %%mm2 \n\t"
800 "movq %%mm1, %%mm3 \n\t"
809 __asm__ volatile(SFENCE:::"memory");
810 __asm__ volatile(EMMS:::"memory");
812 register uint16_t bgr;
814 *d++ = ((bgr&0x1F)<<3) | ((bgr&0x1F)>>2);
815 *d++ = ((bgr&0x3E0)>>2) | ((bgr&0x3E0)>>7);
816 *d++ = ((bgr&0x7C00)>>7) | ((bgr&0x7C00)>>12);
820 static inline void RENAME(rgb16tobgr24)(const uint8_t *src, uint8_t *dst, int src_size)
823 const uint16_t *mm_end;
824 uint8_t *d = (uint8_t *)dst;
825 const uint16_t *s = (const uint16_t *)src;
826 end = s + src_size/2;
827 __asm__ volatile(PREFETCH" %0"::"m"(*s):"memory");
831 PREFETCH" 32(%1) \n\t"
832 "movq (%1), %%mm0 \n\t"
833 "movq (%1), %%mm1 \n\t"
834 "movq (%1), %%mm2 \n\t"
835 "pand %2, %%mm0 \n\t"
836 "pand %3, %%mm1 \n\t"
837 "pand %4, %%mm2 \n\t"
838 "psllq $5, %%mm0 \n\t"
839 "psrlq $1, %%mm2 \n\t"
840 "pmulhw "MANGLE(mul15_mid)", %%mm0 \n\t"
841 "pmulhw "MANGLE(mul16_mid)", %%mm1 \n\t"
842 "pmulhw "MANGLE(mul15_hi)", %%mm2 \n\t"
843 "movq %%mm0, %%mm3 \n\t"
844 "movq %%mm1, %%mm4 \n\t"
845 "movq %%mm2, %%mm5 \n\t"
846 "punpcklwd %5, %%mm0 \n\t"
847 "punpcklwd %5, %%mm1 \n\t"
848 "punpcklwd %5, %%mm2 \n\t"
849 "punpckhwd %5, %%mm3 \n\t"
850 "punpckhwd %5, %%mm4 \n\t"
851 "punpckhwd %5, %%mm5 \n\t"
852 "psllq $8, %%mm1 \n\t"
853 "psllq $16, %%mm2 \n\t"
854 "por %%mm1, %%mm0 \n\t"
855 "por %%mm2, %%mm0 \n\t"
856 "psllq $8, %%mm4 \n\t"
857 "psllq $16, %%mm5 \n\t"
858 "por %%mm4, %%mm3 \n\t"
859 "por %%mm5, %%mm3 \n\t"
861 "movq %%mm0, %%mm6 \n\t"
862 "movq %%mm3, %%mm7 \n\t"
864 "movq 8(%1), %%mm0 \n\t"
865 "movq 8(%1), %%mm1 \n\t"
866 "movq 8(%1), %%mm2 \n\t"
867 "pand %2, %%mm0 \n\t"
868 "pand %3, %%mm1 \n\t"
869 "pand %4, %%mm2 \n\t"
870 "psllq $5, %%mm0 \n\t"
871 "psrlq $1, %%mm2 \n\t"
872 "pmulhw "MANGLE(mul15_mid)", %%mm0 \n\t"
873 "pmulhw "MANGLE(mul16_mid)", %%mm1 \n\t"
874 "pmulhw "MANGLE(mul15_hi)", %%mm2 \n\t"
875 "movq %%mm0, %%mm3 \n\t"
876 "movq %%mm1, %%mm4 \n\t"
877 "movq %%mm2, %%mm5 \n\t"
878 "punpcklwd %5, %%mm0 \n\t"
879 "punpcklwd %5, %%mm1 \n\t"
880 "punpcklwd %5, %%mm2 \n\t"
881 "punpckhwd %5, %%mm3 \n\t"
882 "punpckhwd %5, %%mm4 \n\t"
883 "punpckhwd %5, %%mm5 \n\t"
884 "psllq $8, %%mm1 \n\t"
885 "psllq $16, %%mm2 \n\t"
886 "por %%mm1, %%mm0 \n\t"
887 "por %%mm2, %%mm0 \n\t"
888 "psllq $8, %%mm4 \n\t"
889 "psllq $16, %%mm5 \n\t"
890 "por %%mm4, %%mm3 \n\t"
891 "por %%mm5, %%mm3 \n\t"
893 :"r"(s),"m"(mask16b),"m"(mask16g),"m"(mask16r),"m"(mmx_null)
895 /* borrowed 32 to 24 */
897 "movq %%mm0, %%mm4 \n\t"
898 "movq %%mm3, %%mm5 \n\t"
899 "movq %%mm6, %%mm0 \n\t"
900 "movq %%mm7, %%mm1 \n\t"
902 "movq %%mm4, %%mm6 \n\t"
903 "movq %%mm5, %%mm7 \n\t"
904 "movq %%mm0, %%mm2 \n\t"
905 "movq %%mm1, %%mm3 \n\t"
914 __asm__ volatile(SFENCE:::"memory");
915 __asm__ volatile(EMMS:::"memory");
917 register uint16_t bgr;
919 *d++ = ((bgr&0x1F)<<3) | ((bgr&0x1F)>>2);
920 *d++ = ((bgr&0x7E0)>>3) | ((bgr&0x7E0)>>9);
921 *d++ = ((bgr&0xF800)>>8) | ((bgr&0xF800)>>13);
926 * mm0 = 00 B3 00 B2 00 B1 00 B0
927 * mm1 = 00 G3 00 G2 00 G1 00 G0
928 * mm2 = 00 R3 00 R2 00 R1 00 R0
929 * mm6 = FF FF FF FF FF FF FF FF
930 * mm7 = 00 00 00 00 00 00 00 00
933 "packuswb %%mm7, %%mm0 \n\t" /* 00 00 00 00 B3 B2 B1 B0 */ \
934 "packuswb %%mm7, %%mm1 \n\t" /* 00 00 00 00 G3 G2 G1 G0 */ \
935 "packuswb %%mm7, %%mm2 \n\t" /* 00 00 00 00 R3 R2 R1 R0 */ \
936 "punpcklbw %%mm1, %%mm0 \n\t" /* G3 B3 G2 B2 G1 B1 G0 B0 */ \
937 "punpcklbw %%mm6, %%mm2 \n\t" /* FF R3 FF R2 FF R1 FF R0 */ \
938 "movq %%mm0, %%mm3 \n\t" \
939 "punpcklwd %%mm2, %%mm0 \n\t" /* FF R1 G1 B1 FF R0 G0 B0 */ \
940 "punpckhwd %%mm2, %%mm3 \n\t" /* FF R3 G3 B3 FF R2 G2 B2 */ \
941 MOVNTQ" %%mm0, (%0) \n\t" \
942 MOVNTQ" %%mm3, 8(%0) \n\t" \
944 static inline void RENAME(rgb15to32)(const uint8_t *src, uint8_t *dst, int src_size)
947 const uint16_t *mm_end;
949 const uint16_t *s = (const uint16_t *)src;
950 end = s + src_size/2;
951 __asm__ volatile(PREFETCH" %0"::"m"(*s):"memory");
952 __asm__ volatile("pxor %%mm7,%%mm7 \n\t":::"memory");
953 __asm__ volatile("pcmpeqd %%mm6,%%mm6 \n\t":::"memory");
957 PREFETCH" 32(%1) \n\t"
958 "movq (%1), %%mm0 \n\t"
959 "movq (%1), %%mm1 \n\t"
960 "movq (%1), %%mm2 \n\t"
961 "pand %2, %%mm0 \n\t"
962 "pand %3, %%mm1 \n\t"
963 "pand %4, %%mm2 \n\t"
964 "psllq $5, %%mm0 \n\t"
965 "pmulhw %5, %%mm0 \n\t"
966 "pmulhw %5, %%mm1 \n\t"
967 "pmulhw "MANGLE(mul15_hi)", %%mm2 \n\t"
969 ::"r"(d),"r"(s),"m"(mask15b),"m"(mask15g),"m"(mask15r) ,"m"(mul15_mid)
974 __asm__ volatile(SFENCE:::"memory");
975 __asm__ volatile(EMMS:::"memory");
977 register uint16_t bgr;
979 *d++ = ((bgr&0x1F)<<3) | ((bgr&0x1F)>>2);
980 *d++ = ((bgr&0x3E0)>>2) | ((bgr&0x3E0)>>7);
981 *d++ = ((bgr&0x7C00)>>7) | ((bgr&0x7C00)>>12);
986 static inline void RENAME(rgb16to32)(const uint8_t *src, uint8_t *dst, int src_size)
989 const uint16_t *mm_end;
991 const uint16_t *s = (const uint16_t*)src;
992 end = s + src_size/2;
993 __asm__ volatile(PREFETCH" %0"::"m"(*s):"memory");
994 __asm__ volatile("pxor %%mm7,%%mm7 \n\t":::"memory");
995 __asm__ volatile("pcmpeqd %%mm6,%%mm6 \n\t":::"memory");
999 PREFETCH" 32(%1) \n\t"
1000 "movq (%1), %%mm0 \n\t"
1001 "movq (%1), %%mm1 \n\t"
1002 "movq (%1), %%mm2 \n\t"
1003 "pand %2, %%mm0 \n\t"
1004 "pand %3, %%mm1 \n\t"
1005 "pand %4, %%mm2 \n\t"
1006 "psllq $5, %%mm0 \n\t"
1007 "psrlq $1, %%mm2 \n\t"
1008 "pmulhw %5, %%mm0 \n\t"
1009 "pmulhw "MANGLE(mul16_mid)", %%mm1 \n\t"
1010 "pmulhw "MANGLE(mul15_hi)", %%mm2 \n\t"
1012 ::"r"(d),"r"(s),"m"(mask16b),"m"(mask16g),"m"(mask16r),"m"(mul15_mid)
1017 __asm__ volatile(SFENCE:::"memory");
1018 __asm__ volatile(EMMS:::"memory");
1020 register uint16_t bgr;
1022 *d++ = ((bgr&0x1F)<<3) | ((bgr&0x1F)>>2);
1023 *d++ = ((bgr&0x7E0)>>3) | ((bgr&0x7E0)>>9);
1024 *d++ = ((bgr&0xF800)>>8) | ((bgr&0xF800)>>13);
1029 static inline void RENAME(shuffle_bytes_2103)(const uint8_t *src, uint8_t *dst, int src_size)
1031 x86_reg idx = 15 - src_size;
1032 const uint8_t *s = src-idx;
1033 uint8_t *d = dst-idx;
1037 PREFETCH" (%1, %0) \n\t"
1038 "movq %3, %%mm7 \n\t"
1039 "pxor %4, %%mm7 \n\t"
1040 "movq %%mm7, %%mm6 \n\t"
1041 "pxor %5, %%mm7 \n\t"
1044 PREFETCH" 32(%1, %0) \n\t"
1045 "movq (%1, %0), %%mm0 \n\t"
1046 "movq 8(%1, %0), %%mm1 \n\t"
1047 # if COMPILE_TEMPLATE_MMXEXT
1048 "pshufw $177, %%mm0, %%mm3 \n\t"
1049 "pshufw $177, %%mm1, %%mm5 \n\t"
1050 "pand %%mm7, %%mm0 \n\t"
1051 "pand %%mm6, %%mm3 \n\t"
1052 "pand %%mm7, %%mm1 \n\t"
1053 "pand %%mm6, %%mm5 \n\t"
1054 "por %%mm3, %%mm0 \n\t"
1055 "por %%mm5, %%mm1 \n\t"
1057 "movq %%mm0, %%mm2 \n\t"
1058 "movq %%mm1, %%mm4 \n\t"
1059 "pand %%mm7, %%mm0 \n\t"
1060 "pand %%mm6, %%mm2 \n\t"
1061 "pand %%mm7, %%mm1 \n\t"
1062 "pand %%mm6, %%mm4 \n\t"
1063 "movq %%mm2, %%mm3 \n\t"
1064 "movq %%mm4, %%mm5 \n\t"
1065 "pslld $16, %%mm2 \n\t"
1066 "psrld $16, %%mm3 \n\t"
1067 "pslld $16, %%mm4 \n\t"
1068 "psrld $16, %%mm5 \n\t"
1069 "por %%mm2, %%mm0 \n\t"
1070 "por %%mm4, %%mm1 \n\t"
1071 "por %%mm3, %%mm0 \n\t"
1072 "por %%mm5, %%mm1 \n\t"
1074 MOVNTQ" %%mm0, (%2, %0) \n\t"
1075 MOVNTQ" %%mm1, 8(%2, %0) \n\t"
1082 : "r" (s), "r" (d), "m" (mask32b), "m" (mask32r), "m" (mmx_one)
1084 for (; idx<15; idx+=4) {
1085 register int v = *(const uint32_t *)&s[idx], g = v & 0xff00ff00;
1087 *(uint32_t *)&d[idx] = (v>>16) + g + (v<<16);
1091 static inline void RENAME(rgb24tobgr24)(const uint8_t *src, uint8_t *dst, int src_size)
1094 x86_reg mmx_size= 23 - src_size;
1096 "test %%"REG_a", %%"REG_a" \n\t"
1098 "movq "MANGLE(mask24r)", %%mm5 \n\t"
1099 "movq "MANGLE(mask24g)", %%mm6 \n\t"
1100 "movq "MANGLE(mask24b)", %%mm7 \n\t"
1103 PREFETCH" 32(%1, %%"REG_a") \n\t"
1104 "movq (%1, %%"REG_a"), %%mm0 \n\t" // BGR BGR BG
1105 "movq (%1, %%"REG_a"), %%mm1 \n\t" // BGR BGR BG
1106 "movq 2(%1, %%"REG_a"), %%mm2 \n\t" // R BGR BGR B
1107 "psllq $16, %%mm0 \n\t" // 00 BGR BGR
1108 "pand %%mm5, %%mm0 \n\t"
1109 "pand %%mm6, %%mm1 \n\t"
1110 "pand %%mm7, %%mm2 \n\t"
1111 "por %%mm0, %%mm1 \n\t"
1112 "por %%mm2, %%mm1 \n\t"
1113 "movq 6(%1, %%"REG_a"), %%mm0 \n\t" // BGR BGR BG
1114 MOVNTQ" %%mm1, (%2, %%"REG_a") \n\t" // RGB RGB RG
1115 "movq 8(%1, %%"REG_a"), %%mm1 \n\t" // R BGR BGR B
1116 "movq 10(%1, %%"REG_a"), %%mm2 \n\t" // GR BGR BGR
1117 "pand %%mm7, %%mm0 \n\t"
1118 "pand %%mm5, %%mm1 \n\t"
1119 "pand %%mm6, %%mm2 \n\t"
1120 "por %%mm0, %%mm1 \n\t"
1121 "por %%mm2, %%mm1 \n\t"
1122 "movq 14(%1, %%"REG_a"), %%mm0 \n\t" // R BGR BGR B
1123 MOVNTQ" %%mm1, 8(%2, %%"REG_a") \n\t" // B RGB RGB R
1124 "movq 16(%1, %%"REG_a"), %%mm1 \n\t" // GR BGR BGR
1125 "movq 18(%1, %%"REG_a"), %%mm2 \n\t" // BGR BGR BG
1126 "pand %%mm6, %%mm0 \n\t"
1127 "pand %%mm7, %%mm1 \n\t"
1128 "pand %%mm5, %%mm2 \n\t"
1129 "por %%mm0, %%mm1 \n\t"
1130 "por %%mm2, %%mm1 \n\t"
1131 MOVNTQ" %%mm1, 16(%2, %%"REG_a") \n\t"
1132 "add $24, %%"REG_a" \n\t"
1136 : "r" (src-mmx_size), "r"(dst-mmx_size)
1139 __asm__ volatile(SFENCE:::"memory");
1140 __asm__ volatile(EMMS:::"memory");
1142 if (mmx_size==23) return; //finished, was multiple of 8
1146 src_size= 23-mmx_size;
1149 for (i=0; i<src_size; i+=3) {
1152 dst[i + 1] = src[i + 1];
1153 dst[i + 2] = src[i + 0];
1158 static inline void RENAME(yuvPlanartoyuy2)(const uint8_t *ysrc, const uint8_t *usrc, const uint8_t *vsrc, uint8_t *dst,
1159 int width, int height,
1160 int lumStride, int chromStride, int dstStride, int vertLumPerChroma)
1163 const x86_reg chromWidth= width>>1;
1164 for (y=0; y<height; y++) {
1165 //FIXME handle 2 lines at once (fewer prefetches, reuse some chroma, but very likely memory-limited anyway)
1167 "xor %%"REG_a", %%"REG_a" \n\t"
1170 PREFETCH" 32(%1, %%"REG_a", 2) \n\t"
1171 PREFETCH" 32(%2, %%"REG_a") \n\t"
1172 PREFETCH" 32(%3, %%"REG_a") \n\t"
1173 "movq (%2, %%"REG_a"), %%mm0 \n\t" // U(0)
1174 "movq %%mm0, %%mm2 \n\t" // U(0)
1175 "movq (%3, %%"REG_a"), %%mm1 \n\t" // V(0)
1176 "punpcklbw %%mm1, %%mm0 \n\t" // UVUV UVUV(0)
1177 "punpckhbw %%mm1, %%mm2 \n\t" // UVUV UVUV(8)
1179 "movq (%1, %%"REG_a",2), %%mm3 \n\t" // Y(0)
1180 "movq 8(%1, %%"REG_a",2), %%mm5 \n\t" // Y(8)
1181 "movq %%mm3, %%mm4 \n\t" // Y(0)
1182 "movq %%mm5, %%mm6 \n\t" // Y(8)
1183 "punpcklbw %%mm0, %%mm3 \n\t" // YUYV YUYV(0)
1184 "punpckhbw %%mm0, %%mm4 \n\t" // YUYV YUYV(4)
1185 "punpcklbw %%mm2, %%mm5 \n\t" // YUYV YUYV(8)
1186 "punpckhbw %%mm2, %%mm6 \n\t" // YUYV YUYV(12)
1188 MOVNTQ" %%mm3, (%0, %%"REG_a", 4) \n\t"
1189 MOVNTQ" %%mm4, 8(%0, %%"REG_a", 4) \n\t"
1190 MOVNTQ" %%mm5, 16(%0, %%"REG_a", 4) \n\t"
1191 MOVNTQ" %%mm6, 24(%0, %%"REG_a", 4) \n\t"
1193 "add $8, %%"REG_a" \n\t"
1194 "cmp %4, %%"REG_a" \n\t"
1196 ::"r"(dst), "r"(ysrc), "r"(usrc), "r"(vsrc), "g" (chromWidth)
1199 if ((y&(vertLumPerChroma-1)) == vertLumPerChroma-1) {
1200 usrc += chromStride;
1201 vsrc += chromStride;
1212 * Height should be a multiple of 2 and width should be a multiple of 16.
1213 * (If this is a problem for anyone then tell me, and I will fix it.)
1215 static inline void RENAME(yv12toyuy2)(const uint8_t *ysrc, const uint8_t *usrc, const uint8_t *vsrc, uint8_t *dst,
1216 int width, int height,
1217 int lumStride, int chromStride, int dstStride)
1219 //FIXME interpolate chroma
1220 RENAME(yuvPlanartoyuy2)(ysrc, usrc, vsrc, dst, width, height, lumStride, chromStride, dstStride, 2);
1223 static inline void RENAME(yuvPlanartouyvy)(const uint8_t *ysrc, const uint8_t *usrc, const uint8_t *vsrc, uint8_t *dst,
1224 int width, int height,
1225 int lumStride, int chromStride, int dstStride, int vertLumPerChroma)
1228 const x86_reg chromWidth= width>>1;
1229 for (y=0; y<height; y++) {
1230 //FIXME handle 2 lines at once (fewer prefetches, reuse some chroma, but very likely memory-limited anyway)
1232 "xor %%"REG_a", %%"REG_a" \n\t"
1235 PREFETCH" 32(%1, %%"REG_a", 2) \n\t"
1236 PREFETCH" 32(%2, %%"REG_a") \n\t"
1237 PREFETCH" 32(%3, %%"REG_a") \n\t"
1238 "movq (%2, %%"REG_a"), %%mm0 \n\t" // U(0)
1239 "movq %%mm0, %%mm2 \n\t" // U(0)
1240 "movq (%3, %%"REG_a"), %%mm1 \n\t" // V(0)
1241 "punpcklbw %%mm1, %%mm0 \n\t" // UVUV UVUV(0)
1242 "punpckhbw %%mm1, %%mm2 \n\t" // UVUV UVUV(8)
1244 "movq (%1, %%"REG_a",2), %%mm3 \n\t" // Y(0)
1245 "movq 8(%1, %%"REG_a",2), %%mm5 \n\t" // Y(8)
1246 "movq %%mm0, %%mm4 \n\t" // Y(0)
1247 "movq %%mm2, %%mm6 \n\t" // Y(8)
1248 "punpcklbw %%mm3, %%mm0 \n\t" // YUYV YUYV(0)
1249 "punpckhbw %%mm3, %%mm4 \n\t" // YUYV YUYV(4)
1250 "punpcklbw %%mm5, %%mm2 \n\t" // YUYV YUYV(8)
1251 "punpckhbw %%mm5, %%mm6 \n\t" // YUYV YUYV(12)
1253 MOVNTQ" %%mm0, (%0, %%"REG_a", 4) \n\t"
1254 MOVNTQ" %%mm4, 8(%0, %%"REG_a", 4) \n\t"
1255 MOVNTQ" %%mm2, 16(%0, %%"REG_a", 4) \n\t"
1256 MOVNTQ" %%mm6, 24(%0, %%"REG_a", 4) \n\t"
1258 "add $8, %%"REG_a" \n\t"
1259 "cmp %4, %%"REG_a" \n\t"
1261 ::"r"(dst), "r"(ysrc), "r"(usrc), "r"(vsrc), "g" (chromWidth)
1264 if ((y&(vertLumPerChroma-1)) == vertLumPerChroma-1) {
1265 usrc += chromStride;
1266 vsrc += chromStride;
1277 * Height should be a multiple of 2 and width should be a multiple of 16
1278 * (If this is a problem for anyone then tell me, and I will fix it.)
1280 static inline void RENAME(yv12touyvy)(const uint8_t *ysrc, const uint8_t *usrc, const uint8_t *vsrc, uint8_t *dst,
1281 int width, int height,
1282 int lumStride, int chromStride, int dstStride)
1284 //FIXME interpolate chroma
1285 RENAME(yuvPlanartouyvy)(ysrc, usrc, vsrc, dst, width, height, lumStride, chromStride, dstStride, 2);
1289 * Width should be a multiple of 16.
1291 static inline void RENAME(yuv422ptouyvy)(const uint8_t *ysrc, const uint8_t *usrc, const uint8_t *vsrc, uint8_t *dst,
1292 int width, int height,
1293 int lumStride, int chromStride, int dstStride)
1295 RENAME(yuvPlanartouyvy)(ysrc, usrc, vsrc, dst, width, height, lumStride, chromStride, dstStride, 1);
1299 * Width should be a multiple of 16.
1301 static inline void RENAME(yuv422ptoyuy2)(const uint8_t *ysrc, const uint8_t *usrc, const uint8_t *vsrc, uint8_t *dst,
1302 int width, int height,
1303 int lumStride, int chromStride, int dstStride)
1305 RENAME(yuvPlanartoyuy2)(ysrc, usrc, vsrc, dst, width, height, lumStride, chromStride, dstStride, 1);
1309 * Height should be a multiple of 2 and width should be a multiple of 16.
1310 * (If this is a problem for anyone then tell me, and I will fix it.)
1312 static inline void RENAME(yuy2toyv12)(const uint8_t *src, uint8_t *ydst, uint8_t *udst, uint8_t *vdst,
1313 int width, int height,
1314 int lumStride, int chromStride, int srcStride)
1317 const x86_reg chromWidth= width>>1;
1318 for (y=0; y<height; y+=2) {
1320 "xor %%"REG_a", %%"REG_a" \n\t"
1321 "pcmpeqw %%mm7, %%mm7 \n\t"
1322 "psrlw $8, %%mm7 \n\t" // FF,00,FF,00...
1325 PREFETCH" 64(%0, %%"REG_a", 4) \n\t"
1326 "movq (%0, %%"REG_a", 4), %%mm0 \n\t" // YUYV YUYV(0)
1327 "movq 8(%0, %%"REG_a", 4), %%mm1 \n\t" // YUYV YUYV(4)
1328 "movq %%mm0, %%mm2 \n\t" // YUYV YUYV(0)
1329 "movq %%mm1, %%mm3 \n\t" // YUYV YUYV(4)
1330 "psrlw $8, %%mm0 \n\t" // U0V0 U0V0(0)
1331 "psrlw $8, %%mm1 \n\t" // U0V0 U0V0(4)
1332 "pand %%mm7, %%mm2 \n\t" // Y0Y0 Y0Y0(0)
1333 "pand %%mm7, %%mm3 \n\t" // Y0Y0 Y0Y0(4)
1334 "packuswb %%mm1, %%mm0 \n\t" // UVUV UVUV(0)
1335 "packuswb %%mm3, %%mm2 \n\t" // YYYY YYYY(0)
1337 MOVNTQ" %%mm2, (%1, %%"REG_a", 2) \n\t"
1339 "movq 16(%0, %%"REG_a", 4), %%mm1 \n\t" // YUYV YUYV(8)
1340 "movq 24(%0, %%"REG_a", 4), %%mm2 \n\t" // YUYV YUYV(12)
1341 "movq %%mm1, %%mm3 \n\t" // YUYV YUYV(8)
1342 "movq %%mm2, %%mm4 \n\t" // YUYV YUYV(12)
1343 "psrlw $8, %%mm1 \n\t" // U0V0 U0V0(8)
1344 "psrlw $8, %%mm2 \n\t" // U0V0 U0V0(12)
1345 "pand %%mm7, %%mm3 \n\t" // Y0Y0 Y0Y0(8)
1346 "pand %%mm7, %%mm4 \n\t" // Y0Y0 Y0Y0(12)
1347 "packuswb %%mm2, %%mm1 \n\t" // UVUV UVUV(8)
1348 "packuswb %%mm4, %%mm3 \n\t" // YYYY YYYY(8)
1350 MOVNTQ" %%mm3, 8(%1, %%"REG_a", 2) \n\t"
1352 "movq %%mm0, %%mm2 \n\t" // UVUV UVUV(0)
1353 "movq %%mm1, %%mm3 \n\t" // UVUV UVUV(8)
1354 "psrlw $8, %%mm0 \n\t" // V0V0 V0V0(0)
1355 "psrlw $8, %%mm1 \n\t" // V0V0 V0V0(8)
1356 "pand %%mm7, %%mm2 \n\t" // U0U0 U0U0(0)
1357 "pand %%mm7, %%mm3 \n\t" // U0U0 U0U0(8)
1358 "packuswb %%mm1, %%mm0 \n\t" // VVVV VVVV(0)
1359 "packuswb %%mm3, %%mm2 \n\t" // UUUU UUUU(0)
1361 MOVNTQ" %%mm0, (%3, %%"REG_a") \n\t"
1362 MOVNTQ" %%mm2, (%2, %%"REG_a") \n\t"
1364 "add $8, %%"REG_a" \n\t"
1365 "cmp %4, %%"REG_a" \n\t"
1367 ::"r"(src), "r"(ydst), "r"(udst), "r"(vdst), "g" (chromWidth)
1368 : "memory", "%"REG_a
1375 "xor %%"REG_a", %%"REG_a" \n\t"
1378 PREFETCH" 64(%0, %%"REG_a", 4) \n\t"
1379 "movq (%0, %%"REG_a", 4), %%mm0 \n\t" // YUYV YUYV(0)
1380 "movq 8(%0, %%"REG_a", 4), %%mm1 \n\t" // YUYV YUYV(4)
1381 "movq 16(%0, %%"REG_a", 4), %%mm2 \n\t" // YUYV YUYV(8)
1382 "movq 24(%0, %%"REG_a", 4), %%mm3 \n\t" // YUYV YUYV(12)
1383 "pand %%mm7, %%mm0 \n\t" // Y0Y0 Y0Y0(0)
1384 "pand %%mm7, %%mm1 \n\t" // Y0Y0 Y0Y0(4)
1385 "pand %%mm7, %%mm2 \n\t" // Y0Y0 Y0Y0(8)
1386 "pand %%mm7, %%mm3 \n\t" // Y0Y0 Y0Y0(12)
1387 "packuswb %%mm1, %%mm0 \n\t" // YYYY YYYY(0)
1388 "packuswb %%mm3, %%mm2 \n\t" // YYYY YYYY(8)
1390 MOVNTQ" %%mm0, (%1, %%"REG_a", 2) \n\t"
1391 MOVNTQ" %%mm2, 8(%1, %%"REG_a", 2) \n\t"
1393 "add $8, %%"REG_a" \n\t"
1394 "cmp %4, %%"REG_a" \n\t"
1397 ::"r"(src), "r"(ydst), "r"(udst), "r"(vdst), "g" (chromWidth)
1398 : "memory", "%"REG_a
1400 udst += chromStride;
1401 vdst += chromStride;
1405 __asm__ volatile(EMMS" \n\t"
1409 #endif /* !COMPILE_TEMPLATE_AMD3DNOW */
1411 #if COMPILE_TEMPLATE_MMXEXT || COMPILE_TEMPLATE_AMD3DNOW
1412 static inline void RENAME(planar2x)(const uint8_t *src, uint8_t *dst, int srcWidth, int srcHeight, int srcStride, int dstStride)
1419 for (x=0; x<srcWidth-1; x++) {
1420 dst[2*x+1]= (3*src[x] + src[x+1])>>2;
1421 dst[2*x+2]= ( src[x] + 3*src[x+1])>>2;
1423 dst[2*srcWidth-1]= src[srcWidth-1];
1427 for (y=1; y<srcHeight; y++) {
1428 const x86_reg mmxSize= srcWidth&~15;
1430 "mov %4, %%"REG_a" \n\t"
1431 "movq "MANGLE(mmx_ff)", %%mm0 \n\t"
1432 "movq (%0, %%"REG_a"), %%mm4 \n\t"
1433 "movq %%mm4, %%mm2 \n\t"
1434 "psllq $8, %%mm4 \n\t"
1435 "pand %%mm0, %%mm2 \n\t"
1436 "por %%mm2, %%mm4 \n\t"
1437 "movq (%1, %%"REG_a"), %%mm5 \n\t"
1438 "movq %%mm5, %%mm3 \n\t"
1439 "psllq $8, %%mm5 \n\t"
1440 "pand %%mm0, %%mm3 \n\t"
1441 "por %%mm3, %%mm5 \n\t"
1443 "movq (%0, %%"REG_a"), %%mm0 \n\t"
1444 "movq (%1, %%"REG_a"), %%mm1 \n\t"
1445 "movq 1(%0, %%"REG_a"), %%mm2 \n\t"
1446 "movq 1(%1, %%"REG_a"), %%mm3 \n\t"
1447 PAVGB" %%mm0, %%mm5 \n\t"
1448 PAVGB" %%mm0, %%mm3 \n\t"
1449 PAVGB" %%mm0, %%mm5 \n\t"
1450 PAVGB" %%mm0, %%mm3 \n\t"
1451 PAVGB" %%mm1, %%mm4 \n\t"
1452 PAVGB" %%mm1, %%mm2 \n\t"
1453 PAVGB" %%mm1, %%mm4 \n\t"
1454 PAVGB" %%mm1, %%mm2 \n\t"
1455 "movq %%mm5, %%mm7 \n\t"
1456 "movq %%mm4, %%mm6 \n\t"
1457 "punpcklbw %%mm3, %%mm5 \n\t"
1458 "punpckhbw %%mm3, %%mm7 \n\t"
1459 "punpcklbw %%mm2, %%mm4 \n\t"
1460 "punpckhbw %%mm2, %%mm6 \n\t"
1461 MOVNTQ" %%mm5, (%2, %%"REG_a", 2) \n\t"
1462 MOVNTQ" %%mm7, 8(%2, %%"REG_a", 2) \n\t"
1463 MOVNTQ" %%mm4, (%3, %%"REG_a", 2) \n\t"
1464 MOVNTQ" %%mm6, 8(%3, %%"REG_a", 2) \n\t"
1465 "add $8, %%"REG_a" \n\t"
1466 "movq -1(%0, %%"REG_a"), %%mm4 \n\t"
1467 "movq -1(%1, %%"REG_a"), %%mm5 \n\t"
1469 :: "r" (src + mmxSize ), "r" (src + srcStride + mmxSize ),
1470 "r" (dst + mmxSize*2), "r" (dst + dstStride + mmxSize*2),
1475 for (x=mmxSize-1; x<srcWidth-1; x++) {
1476 dst[2*x +1]= (3*src[x+0] + src[x+srcStride+1])>>2;
1477 dst[2*x+dstStride+2]= ( src[x+0] + 3*src[x+srcStride+1])>>2;
1478 dst[2*x+dstStride+1]= ( src[x+1] + 3*src[x+srcStride ])>>2;
1479 dst[2*x +2]= (3*src[x+1] + src[x+srcStride ])>>2;
1481 dst[srcWidth*2 -1 ]= (3*src[srcWidth-1] + src[srcWidth-1 + srcStride])>>2;
1482 dst[srcWidth*2 -1 + dstStride]= ( src[srcWidth-1] + 3*src[srcWidth-1 + srcStride])>>2;
1491 for (x=0; x<srcWidth-1; x++) {
1492 dst[2*x+1]= (3*src[x] + src[x+1])>>2;
1493 dst[2*x+2]= ( src[x] + 3*src[x+1])>>2;
1495 dst[2*srcWidth-1]= src[srcWidth-1];
1497 __asm__ volatile(EMMS" \n\t"
1501 #endif /* COMPILE_TEMPLATE_MMXEXT || COMPILE_TEMPLATE_AMD3DNOW */
1503 #if !COMPILE_TEMPLATE_AMD3DNOW
1505 * Height should be a multiple of 2 and width should be a multiple of 16.
1506 * (If this is a problem for anyone then tell me, and I will fix it.)
1507 * Chrominance data is only taken from every second line, others are ignored.
1508 * FIXME: Write HQ version.
1510 static inline void RENAME(uyvytoyv12)(const uint8_t *src, uint8_t *ydst, uint8_t *udst, uint8_t *vdst,
1511 int width, int height,
1512 int lumStride, int chromStride, int srcStride)
1515 const x86_reg chromWidth= width>>1;
1516 for (y=0; y<height; y+=2) {
1518 "xor %%"REG_a", %%"REG_a" \n\t"
1519 "pcmpeqw %%mm7, %%mm7 \n\t"
1520 "psrlw $8, %%mm7 \n\t" // FF,00,FF,00...
1523 PREFETCH" 64(%0, %%"REG_a", 4) \n\t"
1524 "movq (%0, %%"REG_a", 4), %%mm0 \n\t" // UYVY UYVY(0)
1525 "movq 8(%0, %%"REG_a", 4), %%mm1 \n\t" // UYVY UYVY(4)
1526 "movq %%mm0, %%mm2 \n\t" // UYVY UYVY(0)
1527 "movq %%mm1, %%mm3 \n\t" // UYVY UYVY(4)
1528 "pand %%mm7, %%mm0 \n\t" // U0V0 U0V0(0)
1529 "pand %%mm7, %%mm1 \n\t" // U0V0 U0V0(4)
1530 "psrlw $8, %%mm2 \n\t" // Y0Y0 Y0Y0(0)
1531 "psrlw $8, %%mm3 \n\t" // Y0Y0 Y0Y0(4)
1532 "packuswb %%mm1, %%mm0 \n\t" // UVUV UVUV(0)
1533 "packuswb %%mm3, %%mm2 \n\t" // YYYY YYYY(0)
1535 MOVNTQ" %%mm2, (%1, %%"REG_a", 2) \n\t"
1537 "movq 16(%0, %%"REG_a", 4), %%mm1 \n\t" // UYVY UYVY(8)
1538 "movq 24(%0, %%"REG_a", 4), %%mm2 \n\t" // UYVY UYVY(12)
1539 "movq %%mm1, %%mm3 \n\t" // UYVY UYVY(8)
1540 "movq %%mm2, %%mm4 \n\t" // UYVY UYVY(12)
1541 "pand %%mm7, %%mm1 \n\t" // U0V0 U0V0(8)
1542 "pand %%mm7, %%mm2 \n\t" // U0V0 U0V0(12)
1543 "psrlw $8, %%mm3 \n\t" // Y0Y0 Y0Y0(8)
1544 "psrlw $8, %%mm4 \n\t" // Y0Y0 Y0Y0(12)
1545 "packuswb %%mm2, %%mm1 \n\t" // UVUV UVUV(8)
1546 "packuswb %%mm4, %%mm3 \n\t" // YYYY YYYY(8)
1548 MOVNTQ" %%mm3, 8(%1, %%"REG_a", 2) \n\t"
1550 "movq %%mm0, %%mm2 \n\t" // UVUV UVUV(0)
1551 "movq %%mm1, %%mm3 \n\t" // UVUV UVUV(8)
1552 "psrlw $8, %%mm0 \n\t" // V0V0 V0V0(0)
1553 "psrlw $8, %%mm1 \n\t" // V0V0 V0V0(8)
1554 "pand %%mm7, %%mm2 \n\t" // U0U0 U0U0(0)
1555 "pand %%mm7, %%mm3 \n\t" // U0U0 U0U0(8)
1556 "packuswb %%mm1, %%mm0 \n\t" // VVVV VVVV(0)
1557 "packuswb %%mm3, %%mm2 \n\t" // UUUU UUUU(0)
1559 MOVNTQ" %%mm0, (%3, %%"REG_a") \n\t"
1560 MOVNTQ" %%mm2, (%2, %%"REG_a") \n\t"
1562 "add $8, %%"REG_a" \n\t"
1563 "cmp %4, %%"REG_a" \n\t"
1565 ::"r"(src), "r"(ydst), "r"(udst), "r"(vdst), "g" (chromWidth)
1566 : "memory", "%"REG_a
1573 "xor %%"REG_a", %%"REG_a" \n\t"
1576 PREFETCH" 64(%0, %%"REG_a", 4) \n\t"
1577 "movq (%0, %%"REG_a", 4), %%mm0 \n\t" // YUYV YUYV(0)
1578 "movq 8(%0, %%"REG_a", 4), %%mm1 \n\t" // YUYV YUYV(4)
1579 "movq 16(%0, %%"REG_a", 4), %%mm2 \n\t" // YUYV YUYV(8)
1580 "movq 24(%0, %%"REG_a", 4), %%mm3 \n\t" // YUYV YUYV(12)
1581 "psrlw $8, %%mm0 \n\t" // Y0Y0 Y0Y0(0)
1582 "psrlw $8, %%mm1 \n\t" // Y0Y0 Y0Y0(4)
1583 "psrlw $8, %%mm2 \n\t" // Y0Y0 Y0Y0(8)
1584 "psrlw $8, %%mm3 \n\t" // Y0Y0 Y0Y0(12)
1585 "packuswb %%mm1, %%mm0 \n\t" // YYYY YYYY(0)
1586 "packuswb %%mm3, %%mm2 \n\t" // YYYY YYYY(8)
1588 MOVNTQ" %%mm0, (%1, %%"REG_a", 2) \n\t"
1589 MOVNTQ" %%mm2, 8(%1, %%"REG_a", 2) \n\t"
1591 "add $8, %%"REG_a" \n\t"
1592 "cmp %4, %%"REG_a" \n\t"
1595 ::"r"(src), "r"(ydst), "r"(udst), "r"(vdst), "g" (chromWidth)
1596 : "memory", "%"REG_a
1598 udst += chromStride;
1599 vdst += chromStride;
1603 __asm__ volatile(EMMS" \n\t"
1607 #endif /* !COMPILE_TEMPLATE_AMD3DNOW */
1610 * Height should be a multiple of 2 and width should be a multiple of 2.
1611 * (If this is a problem for anyone then tell me, and I will fix it.)
1612 * Chrominance data is only taken from every second line,
1613 * others are ignored in the C version.
1614 * FIXME: Write HQ version.
1616 static inline void RENAME(rgb24toyv12)(const uint8_t *src, uint8_t *ydst, uint8_t *udst, uint8_t *vdst,
1617 int width, int height,
1618 int lumStride, int chromStride, int srcStride)
1621 const x86_reg chromWidth= width>>1;
1622 for (y=0; y<height-2; y+=2) {
1624 for (i=0; i<2; i++) {
1626 "mov %2, %%"REG_a" \n\t"
1627 "movq "MANGLE(ff_bgr2YCoeff)", %%mm6 \n\t"
1628 "movq "MANGLE(ff_w1111)", %%mm5 \n\t"
1629 "pxor %%mm7, %%mm7 \n\t"
1630 "lea (%%"REG_a", %%"REG_a", 2), %%"REG_d" \n\t"
1633 PREFETCH" 64(%0, %%"REG_d") \n\t"
1634 "movd (%0, %%"REG_d"), %%mm0 \n\t"
1635 "movd 3(%0, %%"REG_d"), %%mm1 \n\t"
1636 "punpcklbw %%mm7, %%mm0 \n\t"
1637 "punpcklbw %%mm7, %%mm1 \n\t"
1638 "movd 6(%0, %%"REG_d"), %%mm2 \n\t"
1639 "movd 9(%0, %%"REG_d"), %%mm3 \n\t"
1640 "punpcklbw %%mm7, %%mm2 \n\t"
1641 "punpcklbw %%mm7, %%mm3 \n\t"
1642 "pmaddwd %%mm6, %%mm0 \n\t"
1643 "pmaddwd %%mm6, %%mm1 \n\t"
1644 "pmaddwd %%mm6, %%mm2 \n\t"
1645 "pmaddwd %%mm6, %%mm3 \n\t"
1646 #ifndef FAST_BGR2YV12
1647 "psrad $8, %%mm0 \n\t"
1648 "psrad $8, %%mm1 \n\t"
1649 "psrad $8, %%mm2 \n\t"
1650 "psrad $8, %%mm3 \n\t"
1652 "packssdw %%mm1, %%mm0 \n\t"
1653 "packssdw %%mm3, %%mm2 \n\t"
1654 "pmaddwd %%mm5, %%mm0 \n\t"
1655 "pmaddwd %%mm5, %%mm2 \n\t"
1656 "packssdw %%mm2, %%mm0 \n\t"
1657 "psraw $7, %%mm0 \n\t"
1659 "movd 12(%0, %%"REG_d"), %%mm4 \n\t"
1660 "movd 15(%0, %%"REG_d"), %%mm1 \n\t"
1661 "punpcklbw %%mm7, %%mm4 \n\t"
1662 "punpcklbw %%mm7, %%mm1 \n\t"
1663 "movd 18(%0, %%"REG_d"), %%mm2 \n\t"
1664 "movd 21(%0, %%"REG_d"), %%mm3 \n\t"
1665 "punpcklbw %%mm7, %%mm2 \n\t"
1666 "punpcklbw %%mm7, %%mm3 \n\t"
1667 "pmaddwd %%mm6, %%mm4 \n\t"
1668 "pmaddwd %%mm6, %%mm1 \n\t"
1669 "pmaddwd %%mm6, %%mm2 \n\t"
1670 "pmaddwd %%mm6, %%mm3 \n\t"
1671 #ifndef FAST_BGR2YV12
1672 "psrad $8, %%mm4 \n\t"
1673 "psrad $8, %%mm1 \n\t"
1674 "psrad $8, %%mm2 \n\t"
1675 "psrad $8, %%mm3 \n\t"
1677 "packssdw %%mm1, %%mm4 \n\t"
1678 "packssdw %%mm3, %%mm2 \n\t"
1679 "pmaddwd %%mm5, %%mm4 \n\t"
1680 "pmaddwd %%mm5, %%mm2 \n\t"
1681 "add $24, %%"REG_d" \n\t"
1682 "packssdw %%mm2, %%mm4 \n\t"
1683 "psraw $7, %%mm4 \n\t"
1685 "packuswb %%mm4, %%mm0 \n\t"
1686 "paddusb "MANGLE(ff_bgr2YOffset)", %%mm0 \n\t"
1688 MOVNTQ" %%mm0, (%1, %%"REG_a") \n\t"
1689 "add $8, %%"REG_a" \n\t"
1691 : : "r" (src+width*3), "r" (ydst+width), "g" ((x86_reg)-width)
1692 : "%"REG_a, "%"REG_d
1699 "mov %4, %%"REG_a" \n\t"
1700 "movq "MANGLE(ff_w1111)", %%mm5 \n\t"
1701 "movq "MANGLE(ff_bgr2UCoeff)", %%mm6 \n\t"
1702 "pxor %%mm7, %%mm7 \n\t"
1703 "lea (%%"REG_a", %%"REG_a", 2), %%"REG_d" \n\t"
1704 "add %%"REG_d", %%"REG_d" \n\t"
1707 PREFETCH" 64(%0, %%"REG_d") \n\t"
1708 PREFETCH" 64(%1, %%"REG_d") \n\t"
1709 #if COMPILE_TEMPLATE_MMXEXT || COMPILE_TEMPLATE_AMD3DNOW
1710 "movq (%0, %%"REG_d"), %%mm0 \n\t"
1711 "movq (%1, %%"REG_d"), %%mm1 \n\t"
1712 "movq 6(%0, %%"REG_d"), %%mm2 \n\t"
1713 "movq 6(%1, %%"REG_d"), %%mm3 \n\t"
1714 PAVGB" %%mm1, %%mm0 \n\t"
1715 PAVGB" %%mm3, %%mm2 \n\t"
1716 "movq %%mm0, %%mm1 \n\t"
1717 "movq %%mm2, %%mm3 \n\t"
1718 "psrlq $24, %%mm0 \n\t"
1719 "psrlq $24, %%mm2 \n\t"
1720 PAVGB" %%mm1, %%mm0 \n\t"
1721 PAVGB" %%mm3, %%mm2 \n\t"
1722 "punpcklbw %%mm7, %%mm0 \n\t"
1723 "punpcklbw %%mm7, %%mm2 \n\t"
1725 "movd (%0, %%"REG_d"), %%mm0 \n\t"
1726 "movd (%1, %%"REG_d"), %%mm1 \n\t"
1727 "movd 3(%0, %%"REG_d"), %%mm2 \n\t"
1728 "movd 3(%1, %%"REG_d"), %%mm3 \n\t"
1729 "punpcklbw %%mm7, %%mm0 \n\t"
1730 "punpcklbw %%mm7, %%mm1 \n\t"
1731 "punpcklbw %%mm7, %%mm2 \n\t"
1732 "punpcklbw %%mm7, %%mm3 \n\t"
1733 "paddw %%mm1, %%mm0 \n\t"
1734 "paddw %%mm3, %%mm2 \n\t"
1735 "paddw %%mm2, %%mm0 \n\t"
1736 "movd 6(%0, %%"REG_d"), %%mm4 \n\t"
1737 "movd 6(%1, %%"REG_d"), %%mm1 \n\t"
1738 "movd 9(%0, %%"REG_d"), %%mm2 \n\t"
1739 "movd 9(%1, %%"REG_d"), %%mm3 \n\t"
1740 "punpcklbw %%mm7, %%mm4 \n\t"
1741 "punpcklbw %%mm7, %%mm1 \n\t"
1742 "punpcklbw %%mm7, %%mm2 \n\t"
1743 "punpcklbw %%mm7, %%mm3 \n\t"
1744 "paddw %%mm1, %%mm4 \n\t"
1745 "paddw %%mm3, %%mm2 \n\t"
1746 "paddw %%mm4, %%mm2 \n\t"
1747 "psrlw $2, %%mm0 \n\t"
1748 "psrlw $2, %%mm2 \n\t"
1750 "movq "MANGLE(ff_bgr2VCoeff)", %%mm1 \n\t"
1751 "movq "MANGLE(ff_bgr2VCoeff)", %%mm3 \n\t"
1753 "pmaddwd %%mm0, %%mm1 \n\t"
1754 "pmaddwd %%mm2, %%mm3 \n\t"
1755 "pmaddwd %%mm6, %%mm0 \n\t"
1756 "pmaddwd %%mm6, %%mm2 \n\t"
1757 #ifndef FAST_BGR2YV12
1758 "psrad $8, %%mm0 \n\t"
1759 "psrad $8, %%mm1 \n\t"
1760 "psrad $8, %%mm2 \n\t"
1761 "psrad $8, %%mm3 \n\t"
1763 "packssdw %%mm2, %%mm0 \n\t"
1764 "packssdw %%mm3, %%mm1 \n\t"
1765 "pmaddwd %%mm5, %%mm0 \n\t"
1766 "pmaddwd %%mm5, %%mm1 \n\t"
1767 "packssdw %%mm1, %%mm0 \n\t" // V1 V0 U1 U0
1768 "psraw $7, %%mm0 \n\t"
1770 #if COMPILE_TEMPLATE_MMXEXT || COMPILE_TEMPLATE_AMD3DNOW
1771 "movq 12(%0, %%"REG_d"), %%mm4 \n\t"
1772 "movq 12(%1, %%"REG_d"), %%mm1 \n\t"
1773 "movq 18(%0, %%"REG_d"), %%mm2 \n\t"
1774 "movq 18(%1, %%"REG_d"), %%mm3 \n\t"
1775 PAVGB" %%mm1, %%mm4 \n\t"
1776 PAVGB" %%mm3, %%mm2 \n\t"
1777 "movq %%mm4, %%mm1 \n\t"
1778 "movq %%mm2, %%mm3 \n\t"
1779 "psrlq $24, %%mm4 \n\t"
1780 "psrlq $24, %%mm2 \n\t"
1781 PAVGB" %%mm1, %%mm4 \n\t"
1782 PAVGB" %%mm3, %%mm2 \n\t"
1783 "punpcklbw %%mm7, %%mm4 \n\t"
1784 "punpcklbw %%mm7, %%mm2 \n\t"
1786 "movd 12(%0, %%"REG_d"), %%mm4 \n\t"
1787 "movd 12(%1, %%"REG_d"), %%mm1 \n\t"
1788 "movd 15(%0, %%"REG_d"), %%mm2 \n\t"
1789 "movd 15(%1, %%"REG_d"), %%mm3 \n\t"
1790 "punpcklbw %%mm7, %%mm4 \n\t"
1791 "punpcklbw %%mm7, %%mm1 \n\t"
1792 "punpcklbw %%mm7, %%mm2 \n\t"
1793 "punpcklbw %%mm7, %%mm3 \n\t"
1794 "paddw %%mm1, %%mm4 \n\t"
1795 "paddw %%mm3, %%mm2 \n\t"
1796 "paddw %%mm2, %%mm4 \n\t"
1797 "movd 18(%0, %%"REG_d"), %%mm5 \n\t"
1798 "movd 18(%1, %%"REG_d"), %%mm1 \n\t"
1799 "movd 21(%0, %%"REG_d"), %%mm2 \n\t"
1800 "movd 21(%1, %%"REG_d"), %%mm3 \n\t"
1801 "punpcklbw %%mm7, %%mm5 \n\t"
1802 "punpcklbw %%mm7, %%mm1 \n\t"
1803 "punpcklbw %%mm7, %%mm2 \n\t"
1804 "punpcklbw %%mm7, %%mm3 \n\t"
1805 "paddw %%mm1, %%mm5 \n\t"
1806 "paddw %%mm3, %%mm2 \n\t"
1807 "paddw %%mm5, %%mm2 \n\t"
1808 "movq "MANGLE(ff_w1111)", %%mm5 \n\t"
1809 "psrlw $2, %%mm4 \n\t"
1810 "psrlw $2, %%mm2 \n\t"
1812 "movq "MANGLE(ff_bgr2VCoeff)", %%mm1 \n\t"
1813 "movq "MANGLE(ff_bgr2VCoeff)", %%mm3 \n\t"
1815 "pmaddwd %%mm4, %%mm1 \n\t"
1816 "pmaddwd %%mm2, %%mm3 \n\t"
1817 "pmaddwd %%mm6, %%mm4 \n\t"
1818 "pmaddwd %%mm6, %%mm2 \n\t"
1819 #ifndef FAST_BGR2YV12
1820 "psrad $8, %%mm4 \n\t"
1821 "psrad $8, %%mm1 \n\t"
1822 "psrad $8, %%mm2 \n\t"
1823 "psrad $8, %%mm3 \n\t"
1825 "packssdw %%mm2, %%mm4 \n\t"
1826 "packssdw %%mm3, %%mm1 \n\t"
1827 "pmaddwd %%mm5, %%mm4 \n\t"
1828 "pmaddwd %%mm5, %%mm1 \n\t"
1829 "add $24, %%"REG_d" \n\t"
1830 "packssdw %%mm1, %%mm4 \n\t" // V3 V2 U3 U2
1831 "psraw $7, %%mm4 \n\t"
1833 "movq %%mm0, %%mm1 \n\t"
1834 "punpckldq %%mm4, %%mm0 \n\t"
1835 "punpckhdq %%mm4, %%mm1 \n\t"
1836 "packsswb %%mm1, %%mm0 \n\t"
1837 "paddb "MANGLE(ff_bgr2UVOffset)", %%mm0 \n\t"
1838 "movd %%mm0, (%2, %%"REG_a") \n\t"
1839 "punpckhdq %%mm0, %%mm0 \n\t"
1840 "movd %%mm0, (%3, %%"REG_a") \n\t"
1841 "add $4, %%"REG_a" \n\t"
1843 : : "r" (src+chromWidth*6), "r" (src+srcStride+chromWidth*6), "r" (udst+chromWidth), "r" (vdst+chromWidth), "g" (-chromWidth)
1844 : "%"REG_a, "%"REG_d
1847 udst += chromStride;
1848 vdst += chromStride;
1852 __asm__ volatile(EMMS" \n\t"
1856 rgb24toyv12_c(src, ydst, udst, vdst, width, height-y, lumStride, chromStride, srcStride);
1858 #endif /* !COMPILE_TEMPLATE_SSE2 */
1860 #if !COMPILE_TEMPLATE_AMD3DNOW
1861 static void RENAME(interleaveBytes)(const uint8_t *src1, const uint8_t *src2, uint8_t *dest,
1862 int width, int height, int src1Stride,
1863 int src2Stride, int dstStride)
1867 for (h=0; h < height; h++) {
1870 #if COMPILE_TEMPLATE_SSE2
1872 "xor %%"REG_a", %%"REG_a" \n\t"
1874 PREFETCH" 64(%1, %%"REG_a") \n\t"
1875 PREFETCH" 64(%2, %%"REG_a") \n\t"
1876 "movdqa (%1, %%"REG_a"), %%xmm0 \n\t"
1877 "movdqa (%1, %%"REG_a"), %%xmm1 \n\t"
1878 "movdqa (%2, %%"REG_a"), %%xmm2 \n\t"
1879 "punpcklbw %%xmm2, %%xmm0 \n\t"
1880 "punpckhbw %%xmm2, %%xmm1 \n\t"
1881 "movntdq %%xmm0, (%0, %%"REG_a", 2) \n\t"
1882 "movntdq %%xmm1, 16(%0, %%"REG_a", 2) \n\t"
1883 "add $16, %%"REG_a" \n\t"
1884 "cmp %3, %%"REG_a" \n\t"
1886 ::"r"(dest), "r"(src1), "r"(src2), "r" ((x86_reg)width-15)
1887 : "memory", "%"REG_a""
1891 "xor %%"REG_a", %%"REG_a" \n\t"
1893 PREFETCH" 64(%1, %%"REG_a") \n\t"
1894 PREFETCH" 64(%2, %%"REG_a") \n\t"
1895 "movq (%1, %%"REG_a"), %%mm0 \n\t"
1896 "movq 8(%1, %%"REG_a"), %%mm2 \n\t"
1897 "movq %%mm0, %%mm1 \n\t"
1898 "movq %%mm2, %%mm3 \n\t"
1899 "movq (%2, %%"REG_a"), %%mm4 \n\t"
1900 "movq 8(%2, %%"REG_a"), %%mm5 \n\t"
1901 "punpcklbw %%mm4, %%mm0 \n\t"
1902 "punpckhbw %%mm4, %%mm1 \n\t"
1903 "punpcklbw %%mm5, %%mm2 \n\t"
1904 "punpckhbw %%mm5, %%mm3 \n\t"
1905 MOVNTQ" %%mm0, (%0, %%"REG_a", 2) \n\t"
1906 MOVNTQ" %%mm1, 8(%0, %%"REG_a", 2) \n\t"
1907 MOVNTQ" %%mm2, 16(%0, %%"REG_a", 2) \n\t"
1908 MOVNTQ" %%mm3, 24(%0, %%"REG_a", 2) \n\t"
1909 "add $16, %%"REG_a" \n\t"
1910 "cmp %3, %%"REG_a" \n\t"
1912 ::"r"(dest), "r"(src1), "r"(src2), "r" ((x86_reg)width-15)
1913 : "memory", "%"REG_a
1916 for (w= (width&(~15)); w < width; w++) {
1917 dest[2*w+0] = src1[w];
1918 dest[2*w+1] = src2[w];
1930 #endif /* !COMPILE_TEMPLATE_AMD3DNOW */
1932 #if !COMPILE_TEMPLATE_SSE2
1933 #if !COMPILE_TEMPLATE_AMD3DNOW
1934 static inline void RENAME(vu9_to_vu12)(const uint8_t *src1, const uint8_t *src2,
1935 uint8_t *dst1, uint8_t *dst2,
1936 int width, int height,
1937 int srcStride1, int srcStride2,
1938 int dstStride1, int dstStride2)
1942 w=width/2; h=height/2;
1946 ::"m"(*(src1+srcStride1)),"m"(*(src2+srcStride2)):"memory");
1948 const uint8_t* s1=src1+srcStride1*(y>>1);
1949 uint8_t* d=dst1+dstStride1*y;
1951 for (;x<w-31;x+=32) {
1953 PREFETCH" 32(%1,%2) \n\t"
1954 "movq (%1,%2), %%mm0 \n\t"
1955 "movq 8(%1,%2), %%mm2 \n\t"
1956 "movq 16(%1,%2), %%mm4 \n\t"
1957 "movq 24(%1,%2), %%mm6 \n\t"
1958 "movq %%mm0, %%mm1 \n\t"
1959 "movq %%mm2, %%mm3 \n\t"
1960 "movq %%mm4, %%mm5 \n\t"
1961 "movq %%mm6, %%mm7 \n\t"
1962 "punpcklbw %%mm0, %%mm0 \n\t"
1963 "punpckhbw %%mm1, %%mm1 \n\t"
1964 "punpcklbw %%mm2, %%mm2 \n\t"
1965 "punpckhbw %%mm3, %%mm3 \n\t"
1966 "punpcklbw %%mm4, %%mm4 \n\t"
1967 "punpckhbw %%mm5, %%mm5 \n\t"
1968 "punpcklbw %%mm6, %%mm6 \n\t"
1969 "punpckhbw %%mm7, %%mm7 \n\t"
1970 MOVNTQ" %%mm0, (%0,%2,2) \n\t"
1971 MOVNTQ" %%mm1, 8(%0,%2,2) \n\t"
1972 MOVNTQ" %%mm2, 16(%0,%2,2) \n\t"
1973 MOVNTQ" %%mm3, 24(%0,%2,2) \n\t"
1974 MOVNTQ" %%mm4, 32(%0,%2,2) \n\t"
1975 MOVNTQ" %%mm5, 40(%0,%2,2) \n\t"
1976 MOVNTQ" %%mm6, 48(%0,%2,2) \n\t"
1977 MOVNTQ" %%mm7, 56(%0,%2,2)"
1978 :: "r"(d), "r"(s1), "r"(x)
1981 for (;x<w;x++) d[2*x]=d[2*x+1]=s1[x];
1984 const uint8_t* s2=src2+srcStride2*(y>>1);
1985 uint8_t* d=dst2+dstStride2*y;
1987 for (;x<w-31;x+=32) {
1989 PREFETCH" 32(%1,%2) \n\t"
1990 "movq (%1,%2), %%mm0 \n\t"
1991 "movq 8(%1,%2), %%mm2 \n\t"
1992 "movq 16(%1,%2), %%mm4 \n\t"
1993 "movq 24(%1,%2), %%mm6 \n\t"
1994 "movq %%mm0, %%mm1 \n\t"
1995 "movq %%mm2, %%mm3 \n\t"
1996 "movq %%mm4, %%mm5 \n\t"
1997 "movq %%mm6, %%mm7 \n\t"
1998 "punpcklbw %%mm0, %%mm0 \n\t"
1999 "punpckhbw %%mm1, %%mm1 \n\t"
2000 "punpcklbw %%mm2, %%mm2 \n\t"
2001 "punpckhbw %%mm3, %%mm3 \n\t"
2002 "punpcklbw %%mm4, %%mm4 \n\t"
2003 "punpckhbw %%mm5, %%mm5 \n\t"
2004 "punpcklbw %%mm6, %%mm6 \n\t"
2005 "punpckhbw %%mm7, %%mm7 \n\t"
2006 MOVNTQ" %%mm0, (%0,%2,2) \n\t"
2007 MOVNTQ" %%mm1, 8(%0,%2,2) \n\t"
2008 MOVNTQ" %%mm2, 16(%0,%2,2) \n\t"
2009 MOVNTQ" %%mm3, 24(%0,%2,2) \n\t"
2010 MOVNTQ" %%mm4, 32(%0,%2,2) \n\t"
2011 MOVNTQ" %%mm5, 40(%0,%2,2) \n\t"
2012 MOVNTQ" %%mm6, 48(%0,%2,2) \n\t"
2013 MOVNTQ" %%mm7, 56(%0,%2,2)"
2014 :: "r"(d), "r"(s2), "r"(x)
2017 for (;x<w;x++) d[2*x]=d[2*x+1]=s2[x];
2026 static inline void RENAME(yvu9_to_yuy2)(const uint8_t *src1, const uint8_t *src2, const uint8_t *src3,
2028 int width, int height,
2029 int srcStride1, int srcStride2,
2030 int srcStride3, int dstStride)
2034 w=width/2; h=height;
2036 const uint8_t* yp=src1+srcStride1*y;
2037 const uint8_t* up=src2+srcStride2*(y>>2);
2038 const uint8_t* vp=src3+srcStride3*(y>>2);
2039 uint8_t* d=dst+dstStride*y;
2043 PREFETCH" 32(%1, %0) \n\t"
2044 PREFETCH" 32(%2, %0) \n\t"
2045 PREFETCH" 32(%3, %0) \n\t"
2046 "movq (%1, %0, 4), %%mm0 \n\t" /* Y0Y1Y2Y3Y4Y5Y6Y7 */
2047 "movq (%2, %0), %%mm1 \n\t" /* U0U1U2U3U4U5U6U7 */
2048 "movq (%3, %0), %%mm2 \n\t" /* V0V1V2V3V4V5V6V7 */
2049 "movq %%mm0, %%mm3 \n\t" /* Y0Y1Y2Y3Y4Y5Y6Y7 */
2050 "movq %%mm1, %%mm4 \n\t" /* U0U1U2U3U4U5U6U7 */
2051 "movq %%mm2, %%mm5 \n\t" /* V0V1V2V3V4V5V6V7 */
2052 "punpcklbw %%mm1, %%mm1 \n\t" /* U0U0 U1U1 U2U2 U3U3 */
2053 "punpcklbw %%mm2, %%mm2 \n\t" /* V0V0 V1V1 V2V2 V3V3 */
2054 "punpckhbw %%mm4, %%mm4 \n\t" /* U4U4 U5U5 U6U6 U7U7 */
2055 "punpckhbw %%mm5, %%mm5 \n\t" /* V4V4 V5V5 V6V6 V7V7 */
2057 "movq %%mm1, %%mm6 \n\t"
2058 "punpcklbw %%mm2, %%mm1 \n\t" /* U0V0 U0V0 U1V1 U1V1*/
2059 "punpcklbw %%mm1, %%mm0 \n\t" /* Y0U0 Y1V0 Y2U0 Y3V0*/
2060 "punpckhbw %%mm1, %%mm3 \n\t" /* Y4U1 Y5V1 Y6U1 Y7V1*/
2061 MOVNTQ" %%mm0, (%4, %0, 8) \n\t"
2062 MOVNTQ" %%mm3, 8(%4, %0, 8) \n\t"
2064 "punpckhbw %%mm2, %%mm6 \n\t" /* U2V2 U2V2 U3V3 U3V3*/
2065 "movq 8(%1, %0, 4), %%mm0 \n\t"
2066 "movq %%mm0, %%mm3 \n\t"
2067 "punpcklbw %%mm6, %%mm0 \n\t" /* Y U2 Y V2 Y U2 Y V2*/
2068 "punpckhbw %%mm6, %%mm3 \n\t" /* Y U3 Y V3 Y U3 Y V3*/
2069 MOVNTQ" %%mm0, 16(%4, %0, 8) \n\t"
2070 MOVNTQ" %%mm3, 24(%4, %0, 8) \n\t"
2072 "movq %%mm4, %%mm6 \n\t"
2073 "movq 16(%1, %0, 4), %%mm0 \n\t"
2074 "movq %%mm0, %%mm3 \n\t"
2075 "punpcklbw %%mm5, %%mm4 \n\t"
2076 "punpcklbw %%mm4, %%mm0 \n\t" /* Y U4 Y V4 Y U4 Y V4*/
2077 "punpckhbw %%mm4, %%mm3 \n\t" /* Y U5 Y V5 Y U5 Y V5*/
2078 MOVNTQ" %%mm0, 32(%4, %0, 8) \n\t"
2079 MOVNTQ" %%mm3, 40(%4, %0, 8) \n\t"
2081 "punpckhbw %%mm5, %%mm6 \n\t"
2082 "movq 24(%1, %0, 4), %%mm0 \n\t"
2083 "movq %%mm0, %%mm3 \n\t"
2084 "punpcklbw %%mm6, %%mm0 \n\t" /* Y U6 Y V6 Y U6 Y V6*/
2085 "punpckhbw %%mm6, %%mm3 \n\t" /* Y U7 Y V7 Y U7 Y V7*/
2086 MOVNTQ" %%mm0, 48(%4, %0, 8) \n\t"
2087 MOVNTQ" %%mm3, 56(%4, %0, 8) \n\t"
2090 : "r"(yp), "r" (up), "r"(vp), "r"(d)
2094 const int x2 = x<<2;
2097 d[8*x+2] = yp[x2+1];
2099 d[8*x+4] = yp[x2+2];
2101 d[8*x+6] = yp[x2+3];
2111 #endif /* !COMPILE_TEMPLATE_AMD3DNOW */
2113 static void RENAME(extract_even)(const uint8_t *src, uint8_t *dst, x86_reg count)
2122 "pcmpeqw %%mm7, %%mm7 \n\t"
2123 "psrlw $8, %%mm7 \n\t"
2125 "movq -30(%1, %0, 2), %%mm0 \n\t"
2126 "movq -22(%1, %0, 2), %%mm1 \n\t"
2127 "movq -14(%1, %0, 2), %%mm2 \n\t"
2128 "movq -6(%1, %0, 2), %%mm3 \n\t"
2129 "pand %%mm7, %%mm0 \n\t"
2130 "pand %%mm7, %%mm1 \n\t"
2131 "pand %%mm7, %%mm2 \n\t"
2132 "pand %%mm7, %%mm3 \n\t"
2133 "packuswb %%mm1, %%mm0 \n\t"
2134 "packuswb %%mm3, %%mm2 \n\t"
2135 MOVNTQ" %%mm0,-15(%2, %0) \n\t"
2136 MOVNTQ" %%mm2,- 7(%2, %0) \n\t"
2140 : "r"(src), "r"(dst)
2145 dst[count]= src[2*count];
2150 #if !COMPILE_TEMPLATE_AMD3DNOW
2151 static void RENAME(extract_even2)(const uint8_t *src, uint8_t *dst0, uint8_t *dst1, x86_reg count)
2160 "pcmpeqw %%mm7, %%mm7 \n\t"
2161 "psrlw $8, %%mm7 \n\t"
2163 "movq -28(%1, %0, 4), %%mm0 \n\t"
2164 "movq -20(%1, %0, 4), %%mm1 \n\t"
2165 "movq -12(%1, %0, 4), %%mm2 \n\t"
2166 "movq -4(%1, %0, 4), %%mm3 \n\t"
2167 "pand %%mm7, %%mm0 \n\t"
2168 "pand %%mm7, %%mm1 \n\t"
2169 "pand %%mm7, %%mm2 \n\t"
2170 "pand %%mm7, %%mm3 \n\t"
2171 "packuswb %%mm1, %%mm0 \n\t"
2172 "packuswb %%mm3, %%mm2 \n\t"
2173 "movq %%mm0, %%mm1 \n\t"
2174 "movq %%mm2, %%mm3 \n\t"
2175 "psrlw $8, %%mm0 \n\t"
2176 "psrlw $8, %%mm2 \n\t"
2177 "pand %%mm7, %%mm1 \n\t"
2178 "pand %%mm7, %%mm3 \n\t"
2179 "packuswb %%mm2, %%mm0 \n\t"
2180 "packuswb %%mm3, %%mm1 \n\t"
2181 MOVNTQ" %%mm0,- 7(%3, %0) \n\t"
2182 MOVNTQ" %%mm1,- 7(%2, %0) \n\t"
2186 : "r"(src), "r"(dst0), "r"(dst1)
2191 dst0[count]= src[4*count+0];
2192 dst1[count]= src[4*count+2];
2196 #endif /* !COMPILE_TEMPLATE_AMD3DNOW */
2198 static void RENAME(extract_even2avg)(const uint8_t *src0, const uint8_t *src1, uint8_t *dst0, uint8_t *dst1, x86_reg count)
2209 "pcmpeqw %%mm7, %%mm7 \n\t"
2210 "psrlw $8, %%mm7 \n\t"
2212 "movq -28(%1, %0, 4), %%mm0 \n\t"
2213 "movq -20(%1, %0, 4), %%mm1 \n\t"
2214 "movq -12(%1, %0, 4), %%mm2 \n\t"
2215 "movq -4(%1, %0, 4), %%mm3 \n\t"
2216 PAVGB" -28(%2, %0, 4), %%mm0 \n\t"
2217 PAVGB" -20(%2, %0, 4), %%mm1 \n\t"
2218 PAVGB" -12(%2, %0, 4), %%mm2 \n\t"
2219 PAVGB" - 4(%2, %0, 4), %%mm3 \n\t"
2220 "pand %%mm7, %%mm0 \n\t"
2221 "pand %%mm7, %%mm1 \n\t"
2222 "pand %%mm7, %%mm2 \n\t"
2223 "pand %%mm7, %%mm3 \n\t"
2224 "packuswb %%mm1, %%mm0 \n\t"
2225 "packuswb %%mm3, %%mm2 \n\t"
2226 "movq %%mm0, %%mm1 \n\t"
2227 "movq %%mm2, %%mm3 \n\t"
2228 "psrlw $8, %%mm0 \n\t"
2229 "psrlw $8, %%mm2 \n\t"
2230 "pand %%mm7, %%mm1 \n\t"
2231 "pand %%mm7, %%mm3 \n\t"
2232 "packuswb %%mm2, %%mm0 \n\t"
2233 "packuswb %%mm3, %%mm1 \n\t"
2234 MOVNTQ" %%mm0,- 7(%4, %0) \n\t"
2235 MOVNTQ" %%mm1,- 7(%3, %0) \n\t"
2239 : "r"(src0), "r"(src1), "r"(dst0), "r"(dst1)
2245 dst0[count]= (src0[4*count+0]+src1[4*count+0])>>1;
2246 dst1[count]= (src0[4*count+2]+src1[4*count+2])>>1;
2251 #if !COMPILE_TEMPLATE_AMD3DNOW
2252 static void RENAME(extract_odd2)(const uint8_t *src, uint8_t *dst0, uint8_t *dst1, x86_reg count)
2261 "pcmpeqw %%mm7, %%mm7 \n\t"
2262 "psrlw $8, %%mm7 \n\t"
2264 "movq -28(%1, %0, 4), %%mm0 \n\t"
2265 "movq -20(%1, %0, 4), %%mm1 \n\t"
2266 "movq -12(%1, %0, 4), %%mm2 \n\t"
2267 "movq -4(%1, %0, 4), %%mm3 \n\t"
2268 "psrlw $8, %%mm0 \n\t"
2269 "psrlw $8, %%mm1 \n\t"
2270 "psrlw $8, %%mm2 \n\t"
2271 "psrlw $8, %%mm3 \n\t"
2272 "packuswb %%mm1, %%mm0 \n\t"
2273 "packuswb %%mm3, %%mm2 \n\t"
2274 "movq %%mm0, %%mm1 \n\t"
2275 "movq %%mm2, %%mm3 \n\t"
2276 "psrlw $8, %%mm0 \n\t"
2277 "psrlw $8, %%mm2 \n\t"
2278 "pand %%mm7, %%mm1 \n\t"
2279 "pand %%mm7, %%mm3 \n\t"
2280 "packuswb %%mm2, %%mm0 \n\t"
2281 "packuswb %%mm3, %%mm1 \n\t"
2282 MOVNTQ" %%mm0,- 7(%3, %0) \n\t"
2283 MOVNTQ" %%mm1,- 7(%2, %0) \n\t"
2287 : "r"(src), "r"(dst0), "r"(dst1)
2293 dst0[count]= src[4*count+0];
2294 dst1[count]= src[4*count+2];
2298 #endif /* !COMPILE_TEMPLATE_AMD3DNOW */
2300 static void RENAME(extract_odd2avg)(const uint8_t *src0, const uint8_t *src1, uint8_t *dst0, uint8_t *dst1, x86_reg count)
2311 "pcmpeqw %%mm7, %%mm7 \n\t"
2312 "psrlw $8, %%mm7 \n\t"
2314 "movq -28(%1, %0, 4), %%mm0 \n\t"
2315 "movq -20(%1, %0, 4), %%mm1 \n\t"
2316 "movq -12(%1, %0, 4), %%mm2 \n\t"
2317 "movq -4(%1, %0, 4), %%mm3 \n\t"
2318 PAVGB" -28(%2, %0, 4), %%mm0 \n\t"
2319 PAVGB" -20(%2, %0, 4), %%mm1 \n\t"
2320 PAVGB" -12(%2, %0, 4), %%mm2 \n\t"
2321 PAVGB" - 4(%2, %0, 4), %%mm3 \n\t"
2322 "psrlw $8, %%mm0 \n\t"
2323 "psrlw $8, %%mm1 \n\t"
2324 "psrlw $8, %%mm2 \n\t"
2325 "psrlw $8, %%mm3 \n\t"
2326 "packuswb %%mm1, %%mm0 \n\t"
2327 "packuswb %%mm3, %%mm2 \n\t"
2328 "movq %%mm0, %%mm1 \n\t"
2329 "movq %%mm2, %%mm3 \n\t"
2330 "psrlw $8, %%mm0 \n\t"
2331 "psrlw $8, %%mm2 \n\t"
2332 "pand %%mm7, %%mm1 \n\t"
2333 "pand %%mm7, %%mm3 \n\t"
2334 "packuswb %%mm2, %%mm0 \n\t"
2335 "packuswb %%mm3, %%mm1 \n\t"
2336 MOVNTQ" %%mm0,- 7(%4, %0) \n\t"
2337 MOVNTQ" %%mm1,- 7(%3, %0) \n\t"
2341 : "r"(src0), "r"(src1), "r"(dst0), "r"(dst1)
2349 dst0[count]= (src0[4*count+0]+src1[4*count+0])>>1;
2350 dst1[count]= (src0[4*count+2]+src1[4*count+2])>>1;
2355 static void RENAME(yuyvtoyuv420)(uint8_t *ydst, uint8_t *udst, uint8_t *vdst, const uint8_t *src,
2356 int width, int height,
2357 int lumStride, int chromStride, int srcStride)
2360 const int chromWidth= -((-width)>>1);
2362 for (y=0; y<height; y++) {
2363 RENAME(extract_even)(src, ydst, width);
2365 RENAME(extract_odd2avg)(src-srcStride, src, udst, vdst, chromWidth);
2380 #if !COMPILE_TEMPLATE_AMD3DNOW
2381 static void RENAME(yuyvtoyuv422)(uint8_t *ydst, uint8_t *udst, uint8_t *vdst, const uint8_t *src,
2382 int width, int height,
2383 int lumStride, int chromStride, int srcStride)
2386 const int chromWidth= -((-width)>>1);
2388 for (y=0; y<height; y++) {
2389 RENAME(extract_even)(src, ydst, width);
2390 RENAME(extract_odd2)(src, udst, vdst, chromWidth);
2403 #endif /* !COMPILE_TEMPLATE_AMD3DNOW */
2405 static void RENAME(uyvytoyuv420)(uint8_t *ydst, uint8_t *udst, uint8_t *vdst, const uint8_t *src,
2406 int width, int height,
2407 int lumStride, int chromStride, int srcStride)
2410 const int chromWidth= -((-width)>>1);
2412 for (y=0; y<height; y++) {
2413 RENAME(extract_even)(src+1, ydst, width);
2415 RENAME(extract_even2avg)(src-srcStride, src, udst, vdst, chromWidth);
2430 #if !COMPILE_TEMPLATE_AMD3DNOW
2431 static void RENAME(uyvytoyuv422)(uint8_t *ydst, uint8_t *udst, uint8_t *vdst, const uint8_t *src,
2432 int width, int height,
2433 int lumStride, int chromStride, int srcStride)
2436 const int chromWidth= -((-width)>>1);
2438 for (y=0; y<height; y++) {
2439 RENAME(extract_even)(src+1, ydst, width);
2440 RENAME(extract_even2)(src, udst, vdst, chromWidth);
2453 #endif /* !COMPILE_TEMPLATE_AMD3DNOW */
2454 #endif /* !COMPILE_TEMPLATE_SSE2 */
2456 static inline void RENAME(rgb2rgb_init)(void)
2458 #if !COMPILE_TEMPLATE_SSE2
2459 #if !COMPILE_TEMPLATE_AMD3DNOW
2460 rgb15to16 = RENAME(rgb15to16);
2461 rgb15tobgr24 = RENAME(rgb15tobgr24);
2462 rgb15to32 = RENAME(rgb15to32);
2463 rgb16tobgr24 = RENAME(rgb16tobgr24);
2464 rgb16to32 = RENAME(rgb16to32);
2465 rgb16to15 = RENAME(rgb16to15);
2466 rgb24tobgr16 = RENAME(rgb24tobgr16);
2467 rgb24tobgr15 = RENAME(rgb24tobgr15);
2468 rgb24tobgr32 = RENAME(rgb24tobgr32);
2469 rgb32to16 = RENAME(rgb32to16);
2470 rgb32to15 = RENAME(rgb32to15);
2471 rgb32tobgr24 = RENAME(rgb32tobgr24);
2472 rgb24to15 = RENAME(rgb24to15);
2473 rgb24to16 = RENAME(rgb24to16);
2474 rgb24tobgr24 = RENAME(rgb24tobgr24);
2475 shuffle_bytes_2103 = RENAME(shuffle_bytes_2103);
2476 rgb32tobgr16 = RENAME(rgb32tobgr16);
2477 rgb32tobgr15 = RENAME(rgb32tobgr15);
2478 yv12toyuy2 = RENAME(yv12toyuy2);
2479 yv12touyvy = RENAME(yv12touyvy);
2480 yuv422ptoyuy2 = RENAME(yuv422ptoyuy2);
2481 yuv422ptouyvy = RENAME(yuv422ptouyvy);
2482 yuy2toyv12 = RENAME(yuy2toyv12);
2483 vu9_to_vu12 = RENAME(vu9_to_vu12);
2484 yvu9_to_yuy2 = RENAME(yvu9_to_yuy2);
2485 uyvytoyuv422 = RENAME(uyvytoyuv422);
2486 yuyvtoyuv422 = RENAME(yuyvtoyuv422);
2487 #endif /* !COMPILE_TEMPLATE_SSE2 */
2489 #if COMPILE_TEMPLATE_MMXEXT || COMPILE_TEMPLATE_AMD3DNOW
2490 planar2x = RENAME(planar2x);
2491 #endif /* COMPILE_TEMPLATE_MMXEXT || COMPILE_TEMPLATE_AMD3DNOW */
2492 rgb24toyv12 = RENAME(rgb24toyv12);
2494 yuyvtoyuv420 = RENAME(yuyvtoyuv420);
2495 uyvytoyuv420 = RENAME(uyvytoyuv420);
2496 #endif /* COMPILE_TEMPLATE_SSE2 */
2498 #if !COMPILE_TEMPLATE_AMD3DNOW
2499 interleaveBytes = RENAME(interleaveBytes);
2500 #endif /* !COMPILE_TEMPLATE_AMD3DNOW */