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_MMX2
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_MMX2
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");
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"
99 __asm__ volatile(SFENCE:::"memory");
100 __asm__ volatile(EMMS:::"memory");
109 #define STORE_BGR24_MMX \
110 "psrlq $8, %%mm2 \n\t" \
111 "psrlq $8, %%mm3 \n\t" \
112 "psrlq $8, %%mm6 \n\t" \
113 "psrlq $8, %%mm7 \n\t" \
114 "pand "MANGLE(mask24l)", %%mm0\n\t" \
115 "pand "MANGLE(mask24l)", %%mm1\n\t" \
116 "pand "MANGLE(mask24l)", %%mm4\n\t" \
117 "pand "MANGLE(mask24l)", %%mm5\n\t" \
118 "pand "MANGLE(mask24h)", %%mm2\n\t" \
119 "pand "MANGLE(mask24h)", %%mm3\n\t" \
120 "pand "MANGLE(mask24h)", %%mm6\n\t" \
121 "pand "MANGLE(mask24h)", %%mm7\n\t" \
122 "por %%mm2, %%mm0 \n\t" \
123 "por %%mm3, %%mm1 \n\t" \
124 "por %%mm6, %%mm4 \n\t" \
125 "por %%mm7, %%mm5 \n\t" \
127 "movq %%mm1, %%mm2 \n\t" \
128 "movq %%mm4, %%mm3 \n\t" \
129 "psllq $48, %%mm2 \n\t" \
130 "psllq $32, %%mm3 \n\t" \
131 "pand "MANGLE(mask24hh)", %%mm2\n\t" \
132 "pand "MANGLE(mask24hhh)", %%mm3\n\t" \
133 "por %%mm2, %%mm0 \n\t" \
134 "psrlq $16, %%mm1 \n\t" \
135 "psrlq $32, %%mm4 \n\t" \
136 "psllq $16, %%mm5 \n\t" \
137 "por %%mm3, %%mm1 \n\t" \
138 "pand "MANGLE(mask24hhhh)", %%mm5\n\t" \
139 "por %%mm5, %%mm4 \n\t" \
141 MOVNTQ" %%mm0, %0 \n\t" \
142 MOVNTQ" %%mm1, 8%0 \n\t" \
146 static inline void RENAME(rgb32tobgr24)(const uint8_t *src, uint8_t *dst, int src_size)
149 const uint8_t *s = src;
151 const uint8_t *mm_end;
153 __asm__ volatile(PREFETCH" %0"::"m"(*s):"memory");
158 "movq %1, %%mm0 \n\t"
159 "movq 8%1, %%mm1 \n\t"
160 "movq 16%1, %%mm4 \n\t"
161 "movq 24%1, %%mm5 \n\t"
162 "movq %%mm0, %%mm2 \n\t"
163 "movq %%mm1, %%mm3 \n\t"
164 "movq %%mm4, %%mm6 \n\t"
165 "movq %%mm5, %%mm7 \n\t"
173 __asm__ volatile(SFENCE:::"memory");
174 __asm__ volatile(EMMS:::"memory");
184 original by Strepto/Astral
185 ported to gcc & bugfixed: A'rpi
186 MMX2, 3DNOW optimization by Nick Kurshev
187 32-bit C version, and and&add trick by Michael Niedermayer
189 static inline void RENAME(rgb15to16)(const uint8_t *src, uint8_t *dst, int src_size)
191 register const uint8_t* s=src;
192 register uint8_t* d=dst;
193 register const uint8_t *end;
194 const uint8_t *mm_end;
196 __asm__ volatile(PREFETCH" %0"::"m"(*s));
197 __asm__ volatile("movq %0, %%mm4"::"m"(mask15s));
202 "movq %1, %%mm0 \n\t"
203 "movq 8%1, %%mm2 \n\t"
204 "movq %%mm0, %%mm1 \n\t"
205 "movq %%mm2, %%mm3 \n\t"
206 "pand %%mm4, %%mm0 \n\t"
207 "pand %%mm4, %%mm2 \n\t"
208 "paddw %%mm1, %%mm0 \n\t"
209 "paddw %%mm3, %%mm2 \n\t"
210 MOVNTQ" %%mm0, %0 \n\t"
218 __asm__ volatile(SFENCE:::"memory");
219 __asm__ volatile(EMMS:::"memory");
222 register unsigned x= *((const uint32_t *)s);
223 *((uint32_t *)d) = (x&0x7FFF7FFF) + (x&0x7FE07FE0);
228 register unsigned short x= *((const uint16_t *)s);
229 *((uint16_t *)d) = (x&0x7FFF) + (x&0x7FE0);
233 static inline void RENAME(rgb16to15)(const uint8_t *src, uint8_t *dst, int src_size)
235 register const uint8_t* s=src;
236 register uint8_t* d=dst;
237 register const uint8_t *end;
238 const uint8_t *mm_end;
240 __asm__ volatile(PREFETCH" %0"::"m"(*s));
241 __asm__ volatile("movq %0, %%mm7"::"m"(mask15rg));
242 __asm__ volatile("movq %0, %%mm6"::"m"(mask15b));
247 "movq %1, %%mm0 \n\t"
248 "movq 8%1, %%mm2 \n\t"
249 "movq %%mm0, %%mm1 \n\t"
250 "movq %%mm2, %%mm3 \n\t"
251 "psrlq $1, %%mm0 \n\t"
252 "psrlq $1, %%mm2 \n\t"
253 "pand %%mm7, %%mm0 \n\t"
254 "pand %%mm7, %%mm2 \n\t"
255 "pand %%mm6, %%mm1 \n\t"
256 "pand %%mm6, %%mm3 \n\t"
257 "por %%mm1, %%mm0 \n\t"
258 "por %%mm3, %%mm2 \n\t"
259 MOVNTQ" %%mm0, %0 \n\t"
267 __asm__ volatile(SFENCE:::"memory");
268 __asm__ volatile(EMMS:::"memory");
271 register uint32_t x= *((const uint32_t*)s);
272 *((uint32_t *)d) = ((x>>1)&0x7FE07FE0) | (x&0x001F001F);
277 register uint16_t x= *((const uint16_t*)s);
278 *((uint16_t *)d) = ((x>>1)&0x7FE0) | (x&0x001F);
282 static inline void RENAME(rgb32to16)(const uint8_t *src, uint8_t *dst, int src_size)
284 const uint8_t *s = src;
286 const uint8_t *mm_end;
287 uint16_t *d = (uint16_t *)dst;
290 #if 1 //is faster only if multiplies are reasonably fast (FIXME figure out on which CPUs this is faster, on Athlon it is slightly faster)
292 "movq %3, %%mm5 \n\t"
293 "movq %4, %%mm6 \n\t"
294 "movq %5, %%mm7 \n\t"
298 PREFETCH" 32(%1) \n\t"
299 "movd (%1), %%mm0 \n\t"
300 "movd 4(%1), %%mm3 \n\t"
301 "punpckldq 8(%1), %%mm0 \n\t"
302 "punpckldq 12(%1), %%mm3 \n\t"
303 "movq %%mm0, %%mm1 \n\t"
304 "movq %%mm3, %%mm4 \n\t"
305 "pand %%mm6, %%mm0 \n\t"
306 "pand %%mm6, %%mm3 \n\t"
307 "pmaddwd %%mm7, %%mm0 \n\t"
308 "pmaddwd %%mm7, %%mm3 \n\t"
309 "pand %%mm5, %%mm1 \n\t"
310 "pand %%mm5, %%mm4 \n\t"
311 "por %%mm1, %%mm0 \n\t"
312 "por %%mm4, %%mm3 \n\t"
313 "psrld $5, %%mm0 \n\t"
314 "pslld $11, %%mm3 \n\t"
315 "por %%mm3, %%mm0 \n\t"
316 MOVNTQ" %%mm0, (%0) \n\t"
323 : "r" (mm_end), "m" (mask3216g), "m" (mask3216br), "m" (mul3216)
326 __asm__ volatile(PREFETCH" %0"::"m"(*src):"memory");
328 "movq %0, %%mm7 \n\t"
329 "movq %1, %%mm6 \n\t"
330 ::"m"(red_16mask),"m"(green_16mask));
334 "movd %1, %%mm0 \n\t"
335 "movd 4%1, %%mm3 \n\t"
336 "punpckldq 8%1, %%mm0 \n\t"
337 "punpckldq 12%1, %%mm3 \n\t"
338 "movq %%mm0, %%mm1 \n\t"
339 "movq %%mm0, %%mm2 \n\t"
340 "movq %%mm3, %%mm4 \n\t"
341 "movq %%mm3, %%mm5 \n\t"
342 "psrlq $3, %%mm0 \n\t"
343 "psrlq $3, %%mm3 \n\t"
344 "pand %2, %%mm0 \n\t"
345 "pand %2, %%mm3 \n\t"
346 "psrlq $5, %%mm1 \n\t"
347 "psrlq $5, %%mm4 \n\t"
348 "pand %%mm6, %%mm1 \n\t"
349 "pand %%mm6, %%mm4 \n\t"
350 "psrlq $8, %%mm2 \n\t"
351 "psrlq $8, %%mm5 \n\t"
352 "pand %%mm7, %%mm2 \n\t"
353 "pand %%mm7, %%mm5 \n\t"
354 "por %%mm1, %%mm0 \n\t"
355 "por %%mm4, %%mm3 \n\t"
356 "por %%mm2, %%mm0 \n\t"
357 "por %%mm5, %%mm3 \n\t"
358 "psllq $16, %%mm3 \n\t"
359 "por %%mm3, %%mm0 \n\t"
360 MOVNTQ" %%mm0, %0 \n\t"
361 :"=m"(*d):"m"(*s),"m"(blue_16mask):"memory");
366 __asm__ volatile(SFENCE:::"memory");
367 __asm__ volatile(EMMS:::"memory");
369 register int rgb = *(const uint32_t*)s; s += 4;
370 *d++ = ((rgb&0xFF)>>3) + ((rgb&0xFC00)>>5) + ((rgb&0xF80000)>>8);
374 static inline void RENAME(rgb32tobgr16)(const uint8_t *src, uint8_t *dst, int src_size)
376 const uint8_t *s = src;
378 const uint8_t *mm_end;
379 uint16_t *d = (uint16_t *)dst;
381 __asm__ volatile(PREFETCH" %0"::"m"(*src):"memory");
383 "movq %0, %%mm7 \n\t"
384 "movq %1, %%mm6 \n\t"
385 ::"m"(red_16mask),"m"(green_16mask));
390 "movd %1, %%mm0 \n\t"
391 "movd 4%1, %%mm3 \n\t"
392 "punpckldq 8%1, %%mm0 \n\t"
393 "punpckldq 12%1, %%mm3 \n\t"
394 "movq %%mm0, %%mm1 \n\t"
395 "movq %%mm0, %%mm2 \n\t"
396 "movq %%mm3, %%mm4 \n\t"
397 "movq %%mm3, %%mm5 \n\t"
398 "psllq $8, %%mm0 \n\t"
399 "psllq $8, %%mm3 \n\t"
400 "pand %%mm7, %%mm0 \n\t"
401 "pand %%mm7, %%mm3 \n\t"
402 "psrlq $5, %%mm1 \n\t"
403 "psrlq $5, %%mm4 \n\t"
404 "pand %%mm6, %%mm1 \n\t"
405 "pand %%mm6, %%mm4 \n\t"
406 "psrlq $19, %%mm2 \n\t"
407 "psrlq $19, %%mm5 \n\t"
408 "pand %2, %%mm2 \n\t"
409 "pand %2, %%mm5 \n\t"
410 "por %%mm1, %%mm0 \n\t"
411 "por %%mm4, %%mm3 \n\t"
412 "por %%mm2, %%mm0 \n\t"
413 "por %%mm5, %%mm3 \n\t"
414 "psllq $16, %%mm3 \n\t"
415 "por %%mm3, %%mm0 \n\t"
416 MOVNTQ" %%mm0, %0 \n\t"
417 :"=m"(*d):"m"(*s),"m"(blue_16mask):"memory");
421 __asm__ volatile(SFENCE:::"memory");
422 __asm__ volatile(EMMS:::"memory");
424 register int rgb = *(const uint32_t*)s; s += 4;
425 *d++ = ((rgb&0xF8)<<8) + ((rgb&0xFC00)>>5) + ((rgb&0xF80000)>>19);
429 static inline void RENAME(rgb32to15)(const uint8_t *src, uint8_t *dst, int src_size)
431 const uint8_t *s = src;
433 const uint8_t *mm_end;
434 uint16_t *d = (uint16_t *)dst;
437 #if 1 //is faster only if multiplies are reasonably fast (FIXME figure out on which CPUs this is faster, on Athlon it is slightly faster)
439 "movq %3, %%mm5 \n\t"
440 "movq %4, %%mm6 \n\t"
441 "movq %5, %%mm7 \n\t"
445 PREFETCH" 32(%1) \n\t"
446 "movd (%1), %%mm0 \n\t"
447 "movd 4(%1), %%mm3 \n\t"
448 "punpckldq 8(%1), %%mm0 \n\t"
449 "punpckldq 12(%1), %%mm3 \n\t"
450 "movq %%mm0, %%mm1 \n\t"
451 "movq %%mm3, %%mm4 \n\t"
452 "pand %%mm6, %%mm0 \n\t"
453 "pand %%mm6, %%mm3 \n\t"
454 "pmaddwd %%mm7, %%mm0 \n\t"
455 "pmaddwd %%mm7, %%mm3 \n\t"
456 "pand %%mm5, %%mm1 \n\t"
457 "pand %%mm5, %%mm4 \n\t"
458 "por %%mm1, %%mm0 \n\t"
459 "por %%mm4, %%mm3 \n\t"
460 "psrld $6, %%mm0 \n\t"
461 "pslld $10, %%mm3 \n\t"
462 "por %%mm3, %%mm0 \n\t"
463 MOVNTQ" %%mm0, (%0) \n\t"
470 : "r" (mm_end), "m" (mask3215g), "m" (mask3216br), "m" (mul3215)
473 __asm__ volatile(PREFETCH" %0"::"m"(*src):"memory");
475 "movq %0, %%mm7 \n\t"
476 "movq %1, %%mm6 \n\t"
477 ::"m"(red_15mask),"m"(green_15mask));
481 "movd %1, %%mm0 \n\t"
482 "movd 4%1, %%mm3 \n\t"
483 "punpckldq 8%1, %%mm0 \n\t"
484 "punpckldq 12%1, %%mm3 \n\t"
485 "movq %%mm0, %%mm1 \n\t"
486 "movq %%mm0, %%mm2 \n\t"
487 "movq %%mm3, %%mm4 \n\t"
488 "movq %%mm3, %%mm5 \n\t"
489 "psrlq $3, %%mm0 \n\t"
490 "psrlq $3, %%mm3 \n\t"
491 "pand %2, %%mm0 \n\t"
492 "pand %2, %%mm3 \n\t"
493 "psrlq $6, %%mm1 \n\t"
494 "psrlq $6, %%mm4 \n\t"
495 "pand %%mm6, %%mm1 \n\t"
496 "pand %%mm6, %%mm4 \n\t"
497 "psrlq $9, %%mm2 \n\t"
498 "psrlq $9, %%mm5 \n\t"
499 "pand %%mm7, %%mm2 \n\t"
500 "pand %%mm7, %%mm5 \n\t"
501 "por %%mm1, %%mm0 \n\t"
502 "por %%mm4, %%mm3 \n\t"
503 "por %%mm2, %%mm0 \n\t"
504 "por %%mm5, %%mm3 \n\t"
505 "psllq $16, %%mm3 \n\t"
506 "por %%mm3, %%mm0 \n\t"
507 MOVNTQ" %%mm0, %0 \n\t"
508 :"=m"(*d):"m"(*s),"m"(blue_15mask):"memory");
513 __asm__ volatile(SFENCE:::"memory");
514 __asm__ volatile(EMMS:::"memory");
516 register int rgb = *(const uint32_t*)s; s += 4;
517 *d++ = ((rgb&0xFF)>>3) + ((rgb&0xF800)>>6) + ((rgb&0xF80000)>>9);
521 static inline void RENAME(rgb32tobgr15)(const uint8_t *src, uint8_t *dst, int src_size)
523 const uint8_t *s = src;
525 const uint8_t *mm_end;
526 uint16_t *d = (uint16_t *)dst;
528 __asm__ volatile(PREFETCH" %0"::"m"(*src):"memory");
530 "movq %0, %%mm7 \n\t"
531 "movq %1, %%mm6 \n\t"
532 ::"m"(red_15mask),"m"(green_15mask));
537 "movd %1, %%mm0 \n\t"
538 "movd 4%1, %%mm3 \n\t"
539 "punpckldq 8%1, %%mm0 \n\t"
540 "punpckldq 12%1, %%mm3 \n\t"
541 "movq %%mm0, %%mm1 \n\t"
542 "movq %%mm0, %%mm2 \n\t"
543 "movq %%mm3, %%mm4 \n\t"
544 "movq %%mm3, %%mm5 \n\t"
545 "psllq $7, %%mm0 \n\t"
546 "psllq $7, %%mm3 \n\t"
547 "pand %%mm7, %%mm0 \n\t"
548 "pand %%mm7, %%mm3 \n\t"
549 "psrlq $6, %%mm1 \n\t"
550 "psrlq $6, %%mm4 \n\t"
551 "pand %%mm6, %%mm1 \n\t"
552 "pand %%mm6, %%mm4 \n\t"
553 "psrlq $19, %%mm2 \n\t"
554 "psrlq $19, %%mm5 \n\t"
555 "pand %2, %%mm2 \n\t"
556 "pand %2, %%mm5 \n\t"
557 "por %%mm1, %%mm0 \n\t"
558 "por %%mm4, %%mm3 \n\t"
559 "por %%mm2, %%mm0 \n\t"
560 "por %%mm5, %%mm3 \n\t"
561 "psllq $16, %%mm3 \n\t"
562 "por %%mm3, %%mm0 \n\t"
563 MOVNTQ" %%mm0, %0 \n\t"
564 :"=m"(*d):"m"(*s),"m"(blue_15mask):"memory");
568 __asm__ volatile(SFENCE:::"memory");
569 __asm__ volatile(EMMS:::"memory");
571 register int rgb = *(const uint32_t*)s; s += 4;
572 *d++ = ((rgb&0xF8)<<7) + ((rgb&0xF800)>>6) + ((rgb&0xF80000)>>19);
576 static inline void RENAME(rgb24tobgr16)(const uint8_t *src, uint8_t *dst, int src_size)
578 const uint8_t *s = src;
580 const uint8_t *mm_end;
581 uint16_t *d = (uint16_t *)dst;
583 __asm__ volatile(PREFETCH" %0"::"m"(*src):"memory");
585 "movq %0, %%mm7 \n\t"
586 "movq %1, %%mm6 \n\t"
587 ::"m"(red_16mask),"m"(green_16mask));
592 "movd %1, %%mm0 \n\t"
593 "movd 3%1, %%mm3 \n\t"
594 "punpckldq 6%1, %%mm0 \n\t"
595 "punpckldq 9%1, %%mm3 \n\t"
596 "movq %%mm0, %%mm1 \n\t"
597 "movq %%mm0, %%mm2 \n\t"
598 "movq %%mm3, %%mm4 \n\t"
599 "movq %%mm3, %%mm5 \n\t"
600 "psrlq $3, %%mm0 \n\t"
601 "psrlq $3, %%mm3 \n\t"
602 "pand %2, %%mm0 \n\t"
603 "pand %2, %%mm3 \n\t"
604 "psrlq $5, %%mm1 \n\t"
605 "psrlq $5, %%mm4 \n\t"
606 "pand %%mm6, %%mm1 \n\t"
607 "pand %%mm6, %%mm4 \n\t"
608 "psrlq $8, %%mm2 \n\t"
609 "psrlq $8, %%mm5 \n\t"
610 "pand %%mm7, %%mm2 \n\t"
611 "pand %%mm7, %%mm5 \n\t"
612 "por %%mm1, %%mm0 \n\t"
613 "por %%mm4, %%mm3 \n\t"
614 "por %%mm2, %%mm0 \n\t"
615 "por %%mm5, %%mm3 \n\t"
616 "psllq $16, %%mm3 \n\t"
617 "por %%mm3, %%mm0 \n\t"
618 MOVNTQ" %%mm0, %0 \n\t"
619 :"=m"(*d):"m"(*s),"m"(blue_16mask):"memory");
623 __asm__ volatile(SFENCE:::"memory");
624 __asm__ volatile(EMMS:::"memory");
629 *d++ = (b>>3) | ((g&0xFC)<<3) | ((r&0xF8)<<8);
633 static inline void RENAME(rgb24to16)(const uint8_t *src, uint8_t *dst, int src_size)
635 const uint8_t *s = src;
637 const uint8_t *mm_end;
638 uint16_t *d = (uint16_t *)dst;
640 __asm__ volatile(PREFETCH" %0"::"m"(*src):"memory");
642 "movq %0, %%mm7 \n\t"
643 "movq %1, %%mm6 \n\t"
644 ::"m"(red_16mask),"m"(green_16mask));
649 "movd %1, %%mm0 \n\t"
650 "movd 3%1, %%mm3 \n\t"
651 "punpckldq 6%1, %%mm0 \n\t"
652 "punpckldq 9%1, %%mm3 \n\t"
653 "movq %%mm0, %%mm1 \n\t"
654 "movq %%mm0, %%mm2 \n\t"
655 "movq %%mm3, %%mm4 \n\t"
656 "movq %%mm3, %%mm5 \n\t"
657 "psllq $8, %%mm0 \n\t"
658 "psllq $8, %%mm3 \n\t"
659 "pand %%mm7, %%mm0 \n\t"
660 "pand %%mm7, %%mm3 \n\t"
661 "psrlq $5, %%mm1 \n\t"
662 "psrlq $5, %%mm4 \n\t"
663 "pand %%mm6, %%mm1 \n\t"
664 "pand %%mm6, %%mm4 \n\t"
665 "psrlq $19, %%mm2 \n\t"
666 "psrlq $19, %%mm5 \n\t"
667 "pand %2, %%mm2 \n\t"
668 "pand %2, %%mm5 \n\t"
669 "por %%mm1, %%mm0 \n\t"
670 "por %%mm4, %%mm3 \n\t"
671 "por %%mm2, %%mm0 \n\t"
672 "por %%mm5, %%mm3 \n\t"
673 "psllq $16, %%mm3 \n\t"
674 "por %%mm3, %%mm0 \n\t"
675 MOVNTQ" %%mm0, %0 \n\t"
676 :"=m"(*d):"m"(*s),"m"(blue_16mask):"memory");
680 __asm__ volatile(SFENCE:::"memory");
681 __asm__ volatile(EMMS:::"memory");
686 *d++ = (b>>3) | ((g&0xFC)<<3) | ((r&0xF8)<<8);
690 static inline void RENAME(rgb24tobgr15)(const uint8_t *src, uint8_t *dst, int src_size)
692 const uint8_t *s = src;
694 const uint8_t *mm_end;
695 uint16_t *d = (uint16_t *)dst;
697 __asm__ volatile(PREFETCH" %0"::"m"(*src):"memory");
699 "movq %0, %%mm7 \n\t"
700 "movq %1, %%mm6 \n\t"
701 ::"m"(red_15mask),"m"(green_15mask));
706 "movd %1, %%mm0 \n\t"
707 "movd 3%1, %%mm3 \n\t"
708 "punpckldq 6%1, %%mm0 \n\t"
709 "punpckldq 9%1, %%mm3 \n\t"
710 "movq %%mm0, %%mm1 \n\t"
711 "movq %%mm0, %%mm2 \n\t"
712 "movq %%mm3, %%mm4 \n\t"
713 "movq %%mm3, %%mm5 \n\t"
714 "psrlq $3, %%mm0 \n\t"
715 "psrlq $3, %%mm3 \n\t"
716 "pand %2, %%mm0 \n\t"
717 "pand %2, %%mm3 \n\t"
718 "psrlq $6, %%mm1 \n\t"
719 "psrlq $6, %%mm4 \n\t"
720 "pand %%mm6, %%mm1 \n\t"
721 "pand %%mm6, %%mm4 \n\t"
722 "psrlq $9, %%mm2 \n\t"
723 "psrlq $9, %%mm5 \n\t"
724 "pand %%mm7, %%mm2 \n\t"
725 "pand %%mm7, %%mm5 \n\t"
726 "por %%mm1, %%mm0 \n\t"
727 "por %%mm4, %%mm3 \n\t"
728 "por %%mm2, %%mm0 \n\t"
729 "por %%mm5, %%mm3 \n\t"
730 "psllq $16, %%mm3 \n\t"
731 "por %%mm3, %%mm0 \n\t"
732 MOVNTQ" %%mm0, %0 \n\t"
733 :"=m"(*d):"m"(*s),"m"(blue_15mask):"memory");
737 __asm__ volatile(SFENCE:::"memory");
738 __asm__ volatile(EMMS:::"memory");
743 *d++ = (b>>3) | ((g&0xF8)<<2) | ((r&0xF8)<<7);
747 static inline void RENAME(rgb24to15)(const uint8_t *src, uint8_t *dst, int src_size)
749 const uint8_t *s = src;
751 const uint8_t *mm_end;
752 uint16_t *d = (uint16_t *)dst;
754 __asm__ volatile(PREFETCH" %0"::"m"(*src):"memory");
756 "movq %0, %%mm7 \n\t"
757 "movq %1, %%mm6 \n\t"
758 ::"m"(red_15mask),"m"(green_15mask));
763 "movd %1, %%mm0 \n\t"
764 "movd 3%1, %%mm3 \n\t"
765 "punpckldq 6%1, %%mm0 \n\t"
766 "punpckldq 9%1, %%mm3 \n\t"
767 "movq %%mm0, %%mm1 \n\t"
768 "movq %%mm0, %%mm2 \n\t"
769 "movq %%mm3, %%mm4 \n\t"
770 "movq %%mm3, %%mm5 \n\t"
771 "psllq $7, %%mm0 \n\t"
772 "psllq $7, %%mm3 \n\t"
773 "pand %%mm7, %%mm0 \n\t"
774 "pand %%mm7, %%mm3 \n\t"
775 "psrlq $6, %%mm1 \n\t"
776 "psrlq $6, %%mm4 \n\t"
777 "pand %%mm6, %%mm1 \n\t"
778 "pand %%mm6, %%mm4 \n\t"
779 "psrlq $19, %%mm2 \n\t"
780 "psrlq $19, %%mm5 \n\t"
781 "pand %2, %%mm2 \n\t"
782 "pand %2, %%mm5 \n\t"
783 "por %%mm1, %%mm0 \n\t"
784 "por %%mm4, %%mm3 \n\t"
785 "por %%mm2, %%mm0 \n\t"
786 "por %%mm5, %%mm3 \n\t"
787 "psllq $16, %%mm3 \n\t"
788 "por %%mm3, %%mm0 \n\t"
789 MOVNTQ" %%mm0, %0 \n\t"
790 :"=m"(*d):"m"(*s),"m"(blue_15mask):"memory");
794 __asm__ volatile(SFENCE:::"memory");
795 __asm__ volatile(EMMS:::"memory");
800 *d++ = (b>>3) | ((g&0xF8)<<2) | ((r&0xF8)<<7);
804 static inline void RENAME(rgb15tobgr24)(const uint8_t *src, uint8_t *dst, int src_size)
807 const uint16_t *mm_end;
809 const uint16_t *s = (const uint16_t*)src;
810 end = s + src_size/2;
811 __asm__ volatile(PREFETCH" %0"::"m"(*s):"memory");
816 "movq %1, %%mm0 \n\t"
817 "movq %1, %%mm1 \n\t"
818 "movq %1, %%mm2 \n\t"
819 "pand %2, %%mm0 \n\t"
820 "pand %3, %%mm1 \n\t"
821 "pand %4, %%mm2 \n\t"
822 "psllq $5, %%mm0 \n\t"
823 "pmulhw %6, %%mm0 \n\t"
824 "pmulhw %6, %%mm1 \n\t"
825 "pmulhw %7, %%mm2 \n\t"
826 "movq %%mm0, %%mm3 \n\t"
827 "movq %%mm1, %%mm4 \n\t"
828 "movq %%mm2, %%mm5 \n\t"
829 "punpcklwd %5, %%mm0 \n\t"
830 "punpcklwd %5, %%mm1 \n\t"
831 "punpcklwd %5, %%mm2 \n\t"
832 "punpckhwd %5, %%mm3 \n\t"
833 "punpckhwd %5, %%mm4 \n\t"
834 "punpckhwd %5, %%mm5 \n\t"
835 "psllq $8, %%mm1 \n\t"
836 "psllq $16, %%mm2 \n\t"
837 "por %%mm1, %%mm0 \n\t"
838 "por %%mm2, %%mm0 \n\t"
839 "psllq $8, %%mm4 \n\t"
840 "psllq $16, %%mm5 \n\t"
841 "por %%mm4, %%mm3 \n\t"
842 "por %%mm5, %%mm3 \n\t"
844 "movq %%mm0, %%mm6 \n\t"
845 "movq %%mm3, %%mm7 \n\t"
847 "movq 8%1, %%mm0 \n\t"
848 "movq 8%1, %%mm1 \n\t"
849 "movq 8%1, %%mm2 \n\t"
850 "pand %2, %%mm0 \n\t"
851 "pand %3, %%mm1 \n\t"
852 "pand %4, %%mm2 \n\t"
853 "psllq $5, %%mm0 \n\t"
854 "pmulhw %6, %%mm0 \n\t"
855 "pmulhw %6, %%mm1 \n\t"
856 "pmulhw %7, %%mm2 \n\t"
857 "movq %%mm0, %%mm3 \n\t"
858 "movq %%mm1, %%mm4 \n\t"
859 "movq %%mm2, %%mm5 \n\t"
860 "punpcklwd %5, %%mm0 \n\t"
861 "punpcklwd %5, %%mm1 \n\t"
862 "punpcklwd %5, %%mm2 \n\t"
863 "punpckhwd %5, %%mm3 \n\t"
864 "punpckhwd %5, %%mm4 \n\t"
865 "punpckhwd %5, %%mm5 \n\t"
866 "psllq $8, %%mm1 \n\t"
867 "psllq $16, %%mm2 \n\t"
868 "por %%mm1, %%mm0 \n\t"
869 "por %%mm2, %%mm0 \n\t"
870 "psllq $8, %%mm4 \n\t"
871 "psllq $16, %%mm5 \n\t"
872 "por %%mm4, %%mm3 \n\t"
873 "por %%mm5, %%mm3 \n\t"
876 :"m"(*s),"m"(mask15b),"m"(mask15g),"m"(mask15r),"m"(mmx_null),"m"(mul15_mid),"m"(mul15_hi)
878 /* borrowed 32 to 24 */
880 "movq %%mm0, %%mm4 \n\t"
881 "movq %%mm3, %%mm5 \n\t"
882 "movq %%mm6, %%mm0 \n\t"
883 "movq %%mm7, %%mm1 \n\t"
885 "movq %%mm4, %%mm6 \n\t"
886 "movq %%mm5, %%mm7 \n\t"
887 "movq %%mm0, %%mm2 \n\t"
888 "movq %%mm1, %%mm3 \n\t"
898 __asm__ volatile(SFENCE:::"memory");
899 __asm__ volatile(EMMS:::"memory");
901 register uint16_t bgr;
903 *d++ = ((bgr&0x1F)<<3) | ((bgr&0x1F)>>2);
904 *d++ = ((bgr&0x3E0)>>2) | ((bgr&0x3E0)>>7);
905 *d++ = ((bgr&0x7C00)>>7) | ((bgr&0x7C00)>>12);
909 static inline void RENAME(rgb16tobgr24)(const uint8_t *src, uint8_t *dst, int src_size)
912 const uint16_t *mm_end;
913 uint8_t *d = (uint8_t *)dst;
914 const uint16_t *s = (const uint16_t *)src;
915 end = s + src_size/2;
916 __asm__ volatile(PREFETCH" %0"::"m"(*s):"memory");
921 "movq %1, %%mm0 \n\t"
922 "movq %1, %%mm1 \n\t"
923 "movq %1, %%mm2 \n\t"
924 "pand %2, %%mm0 \n\t"
925 "pand %3, %%mm1 \n\t"
926 "pand %4, %%mm2 \n\t"
927 "psllq $5, %%mm0 \n\t"
928 "psrlq $1, %%mm2 \n\t"
929 "pmulhw %6, %%mm0 \n\t"
930 "pmulhw %8, %%mm1 \n\t"
931 "pmulhw %7, %%mm2 \n\t"
932 "movq %%mm0, %%mm3 \n\t"
933 "movq %%mm1, %%mm4 \n\t"
934 "movq %%mm2, %%mm5 \n\t"
935 "punpcklwd %5, %%mm0 \n\t"
936 "punpcklwd %5, %%mm1 \n\t"
937 "punpcklwd %5, %%mm2 \n\t"
938 "punpckhwd %5, %%mm3 \n\t"
939 "punpckhwd %5, %%mm4 \n\t"
940 "punpckhwd %5, %%mm5 \n\t"
941 "psllq $8, %%mm1 \n\t"
942 "psllq $16, %%mm2 \n\t"
943 "por %%mm1, %%mm0 \n\t"
944 "por %%mm2, %%mm0 \n\t"
945 "psllq $8, %%mm4 \n\t"
946 "psllq $16, %%mm5 \n\t"
947 "por %%mm4, %%mm3 \n\t"
948 "por %%mm5, %%mm3 \n\t"
950 "movq %%mm0, %%mm6 \n\t"
951 "movq %%mm3, %%mm7 \n\t"
953 "movq 8%1, %%mm0 \n\t"
954 "movq 8%1, %%mm1 \n\t"
955 "movq 8%1, %%mm2 \n\t"
956 "pand %2, %%mm0 \n\t"
957 "pand %3, %%mm1 \n\t"
958 "pand %4, %%mm2 \n\t"
959 "psllq $5, %%mm0 \n\t"
960 "psrlq $1, %%mm2 \n\t"
961 "pmulhw %6, %%mm0 \n\t"
962 "pmulhw %8, %%mm1 \n\t"
963 "pmulhw %7, %%mm2 \n\t"
964 "movq %%mm0, %%mm3 \n\t"
965 "movq %%mm1, %%mm4 \n\t"
966 "movq %%mm2, %%mm5 \n\t"
967 "punpcklwd %5, %%mm0 \n\t"
968 "punpcklwd %5, %%mm1 \n\t"
969 "punpcklwd %5, %%mm2 \n\t"
970 "punpckhwd %5, %%mm3 \n\t"
971 "punpckhwd %5, %%mm4 \n\t"
972 "punpckhwd %5, %%mm5 \n\t"
973 "psllq $8, %%mm1 \n\t"
974 "psllq $16, %%mm2 \n\t"
975 "por %%mm1, %%mm0 \n\t"
976 "por %%mm2, %%mm0 \n\t"
977 "psllq $8, %%mm4 \n\t"
978 "psllq $16, %%mm5 \n\t"
979 "por %%mm4, %%mm3 \n\t"
980 "por %%mm5, %%mm3 \n\t"
982 :"m"(*s),"m"(mask16b),"m"(mask16g),"m"(mask16r),"m"(mmx_null),"m"(mul15_mid),"m"(mul15_hi),"m"(mul16_mid)
984 /* borrowed 32 to 24 */
986 "movq %%mm0, %%mm4 \n\t"
987 "movq %%mm3, %%mm5 \n\t"
988 "movq %%mm6, %%mm0 \n\t"
989 "movq %%mm7, %%mm1 \n\t"
991 "movq %%mm4, %%mm6 \n\t"
992 "movq %%mm5, %%mm7 \n\t"
993 "movq %%mm0, %%mm2 \n\t"
994 "movq %%mm1, %%mm3 \n\t"
1004 __asm__ volatile(SFENCE:::"memory");
1005 __asm__ volatile(EMMS:::"memory");
1007 register uint16_t bgr;
1009 *d++ = ((bgr&0x1F)<<3) | ((bgr&0x1F)>>2);
1010 *d++ = ((bgr&0x7E0)>>3) | ((bgr&0x7E0)>>9);
1011 *d++ = ((bgr&0xF800)>>8) | ((bgr&0xF800)>>13);
1016 * mm0 = 00 B3 00 B2 00 B1 00 B0
1017 * mm1 = 00 G3 00 G2 00 G1 00 G0
1018 * mm2 = 00 R3 00 R2 00 R1 00 R0
1019 * mm6 = FF FF FF FF FF FF FF FF
1020 * mm7 = 00 00 00 00 00 00 00 00
1022 #define PACK_RGB32 \
1023 "packuswb %%mm7, %%mm0 \n\t" /* 00 00 00 00 B3 B2 B1 B0 */ \
1024 "packuswb %%mm7, %%mm1 \n\t" /* 00 00 00 00 G3 G2 G1 G0 */ \
1025 "packuswb %%mm7, %%mm2 \n\t" /* 00 00 00 00 R3 R2 R1 R0 */ \
1026 "punpcklbw %%mm1, %%mm0 \n\t" /* G3 B3 G2 B2 G1 B1 G0 B0 */ \
1027 "punpcklbw %%mm6, %%mm2 \n\t" /* FF R3 FF R2 FF R1 FF R0 */ \
1028 "movq %%mm0, %%mm3 \n\t" \
1029 "punpcklwd %%mm2, %%mm0 \n\t" /* FF R1 G1 B1 FF R0 G0 B0 */ \
1030 "punpckhwd %%mm2, %%mm3 \n\t" /* FF R3 G3 B3 FF R2 G2 B2 */ \
1031 MOVNTQ" %%mm0, %0 \n\t" \
1032 MOVNTQ" %%mm3, 8%0 \n\t" \
1034 static inline void RENAME(rgb15to32)(const uint8_t *src, uint8_t *dst, int src_size)
1036 const uint16_t *end;
1037 const uint16_t *mm_end;
1039 const uint16_t *s = (const uint16_t *)src;
1040 end = s + src_size/2;
1041 __asm__ volatile(PREFETCH" %0"::"m"(*s):"memory");
1042 __asm__ volatile("pxor %%mm7,%%mm7 \n\t":::"memory");
1043 __asm__ volatile("pcmpeqd %%mm6,%%mm6 \n\t":::"memory");
1045 while (s < mm_end) {
1047 PREFETCH" 32%1 \n\t"
1048 "movq %1, %%mm0 \n\t"
1049 "movq %1, %%mm1 \n\t"
1050 "movq %1, %%mm2 \n\t"
1051 "pand %2, %%mm0 \n\t"
1052 "pand %3, %%mm1 \n\t"
1053 "pand %4, %%mm2 \n\t"
1054 "psllq $5, %%mm0 \n\t"
1055 "pmulhw %5, %%mm0 \n\t"
1056 "pmulhw %5, %%mm1 \n\t"
1057 "pmulhw %6, %%mm2 \n\t"
1060 :"m"(*s),"m"(mask15b),"m"(mask15g),"m"(mask15r),"m"(mul15_mid),"m"(mul15_hi)
1065 __asm__ volatile(SFENCE:::"memory");
1066 __asm__ volatile(EMMS:::"memory");
1068 register uint16_t bgr;
1070 *d++ = ((bgr&0x1F)<<3) | ((bgr&0x1F)>>2);
1071 *d++ = ((bgr&0x3E0)>>2) | ((bgr&0x3E0)>>7);
1072 *d++ = ((bgr&0x7C00)>>7) | ((bgr&0x7C00)>>12);
1077 static inline void RENAME(rgb16to32)(const uint8_t *src, uint8_t *dst, int src_size)
1079 const uint16_t *end;
1080 const uint16_t *mm_end;
1082 const uint16_t *s = (const uint16_t*)src;
1083 end = s + src_size/2;
1084 __asm__ volatile(PREFETCH" %0"::"m"(*s):"memory");
1085 __asm__ volatile("pxor %%mm7,%%mm7 \n\t":::"memory");
1086 __asm__ volatile("pcmpeqd %%mm6,%%mm6 \n\t":::"memory");
1088 while (s < mm_end) {
1090 PREFETCH" 32%1 \n\t"
1091 "movq %1, %%mm0 \n\t"
1092 "movq %1, %%mm1 \n\t"
1093 "movq %1, %%mm2 \n\t"
1094 "pand %2, %%mm0 \n\t"
1095 "pand %3, %%mm1 \n\t"
1096 "pand %4, %%mm2 \n\t"
1097 "psllq $5, %%mm0 \n\t"
1098 "psrlq $1, %%mm2 \n\t"
1099 "pmulhw %5, %%mm0 \n\t"
1100 "pmulhw %7, %%mm1 \n\t"
1101 "pmulhw %6, %%mm2 \n\t"
1104 :"m"(*s),"m"(mask16b),"m"(mask16g),"m"(mask16r),"m"(mul15_mid),"m"(mul15_hi),"m"(mul16_mid)
1109 __asm__ volatile(SFENCE:::"memory");
1110 __asm__ volatile(EMMS:::"memory");
1112 register uint16_t bgr;
1114 *d++ = ((bgr&0x1F)<<3) | ((bgr&0x1F)>>2);
1115 *d++ = ((bgr&0x7E0)>>3) | ((bgr&0x7E0)>>9);
1116 *d++ = ((bgr&0xF800)>>8) | ((bgr&0xF800)>>13);
1121 static inline void RENAME(shuffle_bytes_2103)(const uint8_t *src, uint8_t *dst, int src_size)
1123 x86_reg idx = 15 - src_size;
1124 const uint8_t *s = src-idx;
1125 uint8_t *d = dst-idx;
1129 PREFETCH" (%1, %0) \n\t"
1130 "movq %3, %%mm7 \n\t"
1131 "pxor %4, %%mm7 \n\t"
1132 "movq %%mm7, %%mm6 \n\t"
1133 "pxor %5, %%mm7 \n\t"
1136 PREFETCH" 32(%1, %0) \n\t"
1137 "movq (%1, %0), %%mm0 \n\t"
1138 "movq 8(%1, %0), %%mm1 \n\t"
1139 # if COMPILE_TEMPLATE_MMX2
1140 "pshufw $177, %%mm0, %%mm3 \n\t"
1141 "pshufw $177, %%mm1, %%mm5 \n\t"
1142 "pand %%mm7, %%mm0 \n\t"
1143 "pand %%mm6, %%mm3 \n\t"
1144 "pand %%mm7, %%mm1 \n\t"
1145 "pand %%mm6, %%mm5 \n\t"
1146 "por %%mm3, %%mm0 \n\t"
1147 "por %%mm5, %%mm1 \n\t"
1149 "movq %%mm0, %%mm2 \n\t"
1150 "movq %%mm1, %%mm4 \n\t"
1151 "pand %%mm7, %%mm0 \n\t"
1152 "pand %%mm6, %%mm2 \n\t"
1153 "pand %%mm7, %%mm1 \n\t"
1154 "pand %%mm6, %%mm4 \n\t"
1155 "movq %%mm2, %%mm3 \n\t"
1156 "movq %%mm4, %%mm5 \n\t"
1157 "pslld $16, %%mm2 \n\t"
1158 "psrld $16, %%mm3 \n\t"
1159 "pslld $16, %%mm4 \n\t"
1160 "psrld $16, %%mm5 \n\t"
1161 "por %%mm2, %%mm0 \n\t"
1162 "por %%mm4, %%mm1 \n\t"
1163 "por %%mm3, %%mm0 \n\t"
1164 "por %%mm5, %%mm1 \n\t"
1166 MOVNTQ" %%mm0, (%2, %0) \n\t"
1167 MOVNTQ" %%mm1, 8(%2, %0) \n\t"
1174 : "r" (s), "r" (d), "m" (mask32b), "m" (mask32r), "m" (mmx_one)
1176 for (; idx<15; idx+=4) {
1177 register int v = *(const uint32_t *)&s[idx], g = v & 0xff00ff00;
1179 *(uint32_t *)&d[idx] = (v>>16) + g + (v<<16);
1183 static inline void RENAME(rgb24tobgr24)(const uint8_t *src, uint8_t *dst, int src_size)
1186 x86_reg mmx_size= 23 - src_size;
1188 "test %%"REG_a", %%"REG_a" \n\t"
1190 "movq "MANGLE(mask24r)", %%mm5 \n\t"
1191 "movq "MANGLE(mask24g)", %%mm6 \n\t"
1192 "movq "MANGLE(mask24b)", %%mm7 \n\t"
1195 PREFETCH" 32(%1, %%"REG_a") \n\t"
1196 "movq (%1, %%"REG_a"), %%mm0 \n\t" // BGR BGR BG
1197 "movq (%1, %%"REG_a"), %%mm1 \n\t" // BGR BGR BG
1198 "movq 2(%1, %%"REG_a"), %%mm2 \n\t" // R BGR BGR B
1199 "psllq $16, %%mm0 \n\t" // 00 BGR BGR
1200 "pand %%mm5, %%mm0 \n\t"
1201 "pand %%mm6, %%mm1 \n\t"
1202 "pand %%mm7, %%mm2 \n\t"
1203 "por %%mm0, %%mm1 \n\t"
1204 "por %%mm2, %%mm1 \n\t"
1205 "movq 6(%1, %%"REG_a"), %%mm0 \n\t" // BGR BGR BG
1206 MOVNTQ" %%mm1, (%2, %%"REG_a") \n\t" // RGB RGB RG
1207 "movq 8(%1, %%"REG_a"), %%mm1 \n\t" // R BGR BGR B
1208 "movq 10(%1, %%"REG_a"), %%mm2 \n\t" // GR BGR BGR
1209 "pand %%mm7, %%mm0 \n\t"
1210 "pand %%mm5, %%mm1 \n\t"
1211 "pand %%mm6, %%mm2 \n\t"
1212 "por %%mm0, %%mm1 \n\t"
1213 "por %%mm2, %%mm1 \n\t"
1214 "movq 14(%1, %%"REG_a"), %%mm0 \n\t" // R BGR BGR B
1215 MOVNTQ" %%mm1, 8(%2, %%"REG_a") \n\t" // B RGB RGB R
1216 "movq 16(%1, %%"REG_a"), %%mm1 \n\t" // GR BGR BGR
1217 "movq 18(%1, %%"REG_a"), %%mm2 \n\t" // BGR BGR BG
1218 "pand %%mm6, %%mm0 \n\t"
1219 "pand %%mm7, %%mm1 \n\t"
1220 "pand %%mm5, %%mm2 \n\t"
1221 "por %%mm0, %%mm1 \n\t"
1222 "por %%mm2, %%mm1 \n\t"
1223 MOVNTQ" %%mm1, 16(%2, %%"REG_a") \n\t"
1224 "add $24, %%"REG_a" \n\t"
1228 : "r" (src-mmx_size), "r"(dst-mmx_size)
1231 __asm__ volatile(SFENCE:::"memory");
1232 __asm__ volatile(EMMS:::"memory");
1234 if (mmx_size==23) return; //finished, was multiple of 8
1238 src_size= 23-mmx_size;
1241 for (i=0; i<src_size; i+=3) {
1244 dst[i + 1] = src[i + 1];
1245 dst[i + 2] = src[i + 0];
1250 static inline void RENAME(yuvPlanartoyuy2)(const uint8_t *ysrc, const uint8_t *usrc, const uint8_t *vsrc, uint8_t *dst,
1251 int width, int height,
1252 int lumStride, int chromStride, int dstStride, int vertLumPerChroma)
1255 const x86_reg chromWidth= width>>1;
1256 for (y=0; y<height; y++) {
1257 //FIXME handle 2 lines at once (fewer prefetches, reuse some chroma, but very likely memory-limited anyway)
1259 "xor %%"REG_a", %%"REG_a" \n\t"
1262 PREFETCH" 32(%1, %%"REG_a", 2) \n\t"
1263 PREFETCH" 32(%2, %%"REG_a") \n\t"
1264 PREFETCH" 32(%3, %%"REG_a") \n\t"
1265 "movq (%2, %%"REG_a"), %%mm0 \n\t" // U(0)
1266 "movq %%mm0, %%mm2 \n\t" // U(0)
1267 "movq (%3, %%"REG_a"), %%mm1 \n\t" // V(0)
1268 "punpcklbw %%mm1, %%mm0 \n\t" // UVUV UVUV(0)
1269 "punpckhbw %%mm1, %%mm2 \n\t" // UVUV UVUV(8)
1271 "movq (%1, %%"REG_a",2), %%mm3 \n\t" // Y(0)
1272 "movq 8(%1, %%"REG_a",2), %%mm5 \n\t" // Y(8)
1273 "movq %%mm3, %%mm4 \n\t" // Y(0)
1274 "movq %%mm5, %%mm6 \n\t" // Y(8)
1275 "punpcklbw %%mm0, %%mm3 \n\t" // YUYV YUYV(0)
1276 "punpckhbw %%mm0, %%mm4 \n\t" // YUYV YUYV(4)
1277 "punpcklbw %%mm2, %%mm5 \n\t" // YUYV YUYV(8)
1278 "punpckhbw %%mm2, %%mm6 \n\t" // YUYV YUYV(12)
1280 MOVNTQ" %%mm3, (%0, %%"REG_a", 4) \n\t"
1281 MOVNTQ" %%mm4, 8(%0, %%"REG_a", 4) \n\t"
1282 MOVNTQ" %%mm5, 16(%0, %%"REG_a", 4) \n\t"
1283 MOVNTQ" %%mm6, 24(%0, %%"REG_a", 4) \n\t"
1285 "add $8, %%"REG_a" \n\t"
1286 "cmp %4, %%"REG_a" \n\t"
1288 ::"r"(dst), "r"(ysrc), "r"(usrc), "r"(vsrc), "g" (chromWidth)
1291 if ((y&(vertLumPerChroma-1)) == vertLumPerChroma-1) {
1292 usrc += chromStride;
1293 vsrc += chromStride;
1304 * Height should be a multiple of 2 and width should be a multiple of 16.
1305 * (If this is a problem for anyone then tell me, and I will fix it.)
1307 static inline void RENAME(yv12toyuy2)(const uint8_t *ysrc, const uint8_t *usrc, const uint8_t *vsrc, uint8_t *dst,
1308 int width, int height,
1309 int lumStride, int chromStride, int dstStride)
1311 //FIXME interpolate chroma
1312 RENAME(yuvPlanartoyuy2)(ysrc, usrc, vsrc, dst, width, height, lumStride, chromStride, dstStride, 2);
1315 static inline void RENAME(yuvPlanartouyvy)(const uint8_t *ysrc, const uint8_t *usrc, const uint8_t *vsrc, uint8_t *dst,
1316 int width, int height,
1317 int lumStride, int chromStride, int dstStride, int vertLumPerChroma)
1320 const x86_reg chromWidth= width>>1;
1321 for (y=0; y<height; y++) {
1322 //FIXME handle 2 lines at once (fewer prefetches, reuse some chroma, but very likely memory-limited anyway)
1324 "xor %%"REG_a", %%"REG_a" \n\t"
1327 PREFETCH" 32(%1, %%"REG_a", 2) \n\t"
1328 PREFETCH" 32(%2, %%"REG_a") \n\t"
1329 PREFETCH" 32(%3, %%"REG_a") \n\t"
1330 "movq (%2, %%"REG_a"), %%mm0 \n\t" // U(0)
1331 "movq %%mm0, %%mm2 \n\t" // U(0)
1332 "movq (%3, %%"REG_a"), %%mm1 \n\t" // V(0)
1333 "punpcklbw %%mm1, %%mm0 \n\t" // UVUV UVUV(0)
1334 "punpckhbw %%mm1, %%mm2 \n\t" // UVUV UVUV(8)
1336 "movq (%1, %%"REG_a",2), %%mm3 \n\t" // Y(0)
1337 "movq 8(%1, %%"REG_a",2), %%mm5 \n\t" // Y(8)
1338 "movq %%mm0, %%mm4 \n\t" // Y(0)
1339 "movq %%mm2, %%mm6 \n\t" // Y(8)
1340 "punpcklbw %%mm3, %%mm0 \n\t" // YUYV YUYV(0)
1341 "punpckhbw %%mm3, %%mm4 \n\t" // YUYV YUYV(4)
1342 "punpcklbw %%mm5, %%mm2 \n\t" // YUYV YUYV(8)
1343 "punpckhbw %%mm5, %%mm6 \n\t" // YUYV YUYV(12)
1345 MOVNTQ" %%mm0, (%0, %%"REG_a", 4) \n\t"
1346 MOVNTQ" %%mm4, 8(%0, %%"REG_a", 4) \n\t"
1347 MOVNTQ" %%mm2, 16(%0, %%"REG_a", 4) \n\t"
1348 MOVNTQ" %%mm6, 24(%0, %%"REG_a", 4) \n\t"
1350 "add $8, %%"REG_a" \n\t"
1351 "cmp %4, %%"REG_a" \n\t"
1353 ::"r"(dst), "r"(ysrc), "r"(usrc), "r"(vsrc), "g" (chromWidth)
1356 if ((y&(vertLumPerChroma-1)) == vertLumPerChroma-1) {
1357 usrc += chromStride;
1358 vsrc += chromStride;
1369 * Height should be a multiple of 2 and width should be a multiple of 16
1370 * (If this is a problem for anyone then tell me, and I will fix it.)
1372 static inline void RENAME(yv12touyvy)(const uint8_t *ysrc, const uint8_t *usrc, const uint8_t *vsrc, uint8_t *dst,
1373 int width, int height,
1374 int lumStride, int chromStride, int dstStride)
1376 //FIXME interpolate chroma
1377 RENAME(yuvPlanartouyvy)(ysrc, usrc, vsrc, dst, width, height, lumStride, chromStride, dstStride, 2);
1381 * Width should be a multiple of 16.
1383 static inline void RENAME(yuv422ptouyvy)(const uint8_t *ysrc, const uint8_t *usrc, const uint8_t *vsrc, uint8_t *dst,
1384 int width, int height,
1385 int lumStride, int chromStride, int dstStride)
1387 RENAME(yuvPlanartouyvy)(ysrc, usrc, vsrc, dst, width, height, lumStride, chromStride, dstStride, 1);
1391 * Width should be a multiple of 16.
1393 static inline void RENAME(yuv422ptoyuy2)(const uint8_t *ysrc, const uint8_t *usrc, const uint8_t *vsrc, uint8_t *dst,
1394 int width, int height,
1395 int lumStride, int chromStride, int dstStride)
1397 RENAME(yuvPlanartoyuy2)(ysrc, usrc, vsrc, dst, width, height, lumStride, chromStride, dstStride, 1);
1401 * Height should be a multiple of 2 and width should be a multiple of 16.
1402 * (If this is a problem for anyone then tell me, and I will fix it.)
1404 static inline void RENAME(yuy2toyv12)(const uint8_t *src, uint8_t *ydst, uint8_t *udst, uint8_t *vdst,
1405 int width, int height,
1406 int lumStride, int chromStride, int srcStride)
1409 const x86_reg chromWidth= width>>1;
1410 for (y=0; y<height; y+=2) {
1412 "xor %%"REG_a", %%"REG_a" \n\t"
1413 "pcmpeqw %%mm7, %%mm7 \n\t"
1414 "psrlw $8, %%mm7 \n\t" // FF,00,FF,00...
1417 PREFETCH" 64(%0, %%"REG_a", 4) \n\t"
1418 "movq (%0, %%"REG_a", 4), %%mm0 \n\t" // YUYV YUYV(0)
1419 "movq 8(%0, %%"REG_a", 4), %%mm1 \n\t" // YUYV YUYV(4)
1420 "movq %%mm0, %%mm2 \n\t" // YUYV YUYV(0)
1421 "movq %%mm1, %%mm3 \n\t" // YUYV YUYV(4)
1422 "psrlw $8, %%mm0 \n\t" // U0V0 U0V0(0)
1423 "psrlw $8, %%mm1 \n\t" // U0V0 U0V0(4)
1424 "pand %%mm7, %%mm2 \n\t" // Y0Y0 Y0Y0(0)
1425 "pand %%mm7, %%mm3 \n\t" // Y0Y0 Y0Y0(4)
1426 "packuswb %%mm1, %%mm0 \n\t" // UVUV UVUV(0)
1427 "packuswb %%mm3, %%mm2 \n\t" // YYYY YYYY(0)
1429 MOVNTQ" %%mm2, (%1, %%"REG_a", 2) \n\t"
1431 "movq 16(%0, %%"REG_a", 4), %%mm1 \n\t" // YUYV YUYV(8)
1432 "movq 24(%0, %%"REG_a", 4), %%mm2 \n\t" // YUYV YUYV(12)
1433 "movq %%mm1, %%mm3 \n\t" // YUYV YUYV(8)
1434 "movq %%mm2, %%mm4 \n\t" // YUYV YUYV(12)
1435 "psrlw $8, %%mm1 \n\t" // U0V0 U0V0(8)
1436 "psrlw $8, %%mm2 \n\t" // U0V0 U0V0(12)
1437 "pand %%mm7, %%mm3 \n\t" // Y0Y0 Y0Y0(8)
1438 "pand %%mm7, %%mm4 \n\t" // Y0Y0 Y0Y0(12)
1439 "packuswb %%mm2, %%mm1 \n\t" // UVUV UVUV(8)
1440 "packuswb %%mm4, %%mm3 \n\t" // YYYY YYYY(8)
1442 MOVNTQ" %%mm3, 8(%1, %%"REG_a", 2) \n\t"
1444 "movq %%mm0, %%mm2 \n\t" // UVUV UVUV(0)
1445 "movq %%mm1, %%mm3 \n\t" // UVUV UVUV(8)
1446 "psrlw $8, %%mm0 \n\t" // V0V0 V0V0(0)
1447 "psrlw $8, %%mm1 \n\t" // V0V0 V0V0(8)
1448 "pand %%mm7, %%mm2 \n\t" // U0U0 U0U0(0)
1449 "pand %%mm7, %%mm3 \n\t" // U0U0 U0U0(8)
1450 "packuswb %%mm1, %%mm0 \n\t" // VVVV VVVV(0)
1451 "packuswb %%mm3, %%mm2 \n\t" // UUUU UUUU(0)
1453 MOVNTQ" %%mm0, (%3, %%"REG_a") \n\t"
1454 MOVNTQ" %%mm2, (%2, %%"REG_a") \n\t"
1456 "add $8, %%"REG_a" \n\t"
1457 "cmp %4, %%"REG_a" \n\t"
1459 ::"r"(src), "r"(ydst), "r"(udst), "r"(vdst), "g" (chromWidth)
1460 : "memory", "%"REG_a
1467 "xor %%"REG_a", %%"REG_a" \n\t"
1470 PREFETCH" 64(%0, %%"REG_a", 4) \n\t"
1471 "movq (%0, %%"REG_a", 4), %%mm0 \n\t" // YUYV YUYV(0)
1472 "movq 8(%0, %%"REG_a", 4), %%mm1 \n\t" // YUYV YUYV(4)
1473 "movq 16(%0, %%"REG_a", 4), %%mm2 \n\t" // YUYV YUYV(8)
1474 "movq 24(%0, %%"REG_a", 4), %%mm3 \n\t" // YUYV YUYV(12)
1475 "pand %%mm7, %%mm0 \n\t" // Y0Y0 Y0Y0(0)
1476 "pand %%mm7, %%mm1 \n\t" // Y0Y0 Y0Y0(4)
1477 "pand %%mm7, %%mm2 \n\t" // Y0Y0 Y0Y0(8)
1478 "pand %%mm7, %%mm3 \n\t" // Y0Y0 Y0Y0(12)
1479 "packuswb %%mm1, %%mm0 \n\t" // YYYY YYYY(0)
1480 "packuswb %%mm3, %%mm2 \n\t" // YYYY YYYY(8)
1482 MOVNTQ" %%mm0, (%1, %%"REG_a", 2) \n\t"
1483 MOVNTQ" %%mm2, 8(%1, %%"REG_a", 2) \n\t"
1485 "add $8, %%"REG_a" \n\t"
1486 "cmp %4, %%"REG_a" \n\t"
1489 ::"r"(src), "r"(ydst), "r"(udst), "r"(vdst), "g" (chromWidth)
1490 : "memory", "%"REG_a
1492 udst += chromStride;
1493 vdst += chromStride;
1497 __asm__ volatile(EMMS" \n\t"
1501 #endif /* !COMPILE_TEMPLATE_AMD3DNOW */
1503 #if COMPILE_TEMPLATE_MMX2 || COMPILE_TEMPLATE_AMD3DNOW
1504 static inline void RENAME(planar2x)(const uint8_t *src, uint8_t *dst, int srcWidth, int srcHeight, int srcStride, int dstStride)
1511 for (x=0; x<srcWidth-1; x++) {
1512 dst[2*x+1]= (3*src[x] + src[x+1])>>2;
1513 dst[2*x+2]= ( src[x] + 3*src[x+1])>>2;
1515 dst[2*srcWidth-1]= src[srcWidth-1];
1519 for (y=1; y<srcHeight; y++) {
1520 const x86_reg mmxSize= srcWidth&~15;
1522 "mov %4, %%"REG_a" \n\t"
1523 "movq "MANGLE(mmx_ff)", %%mm0 \n\t"
1524 "movq (%0, %%"REG_a"), %%mm4 \n\t"
1525 "movq %%mm4, %%mm2 \n\t"
1526 "psllq $8, %%mm4 \n\t"
1527 "pand %%mm0, %%mm2 \n\t"
1528 "por %%mm2, %%mm4 \n\t"
1529 "movq (%1, %%"REG_a"), %%mm5 \n\t"
1530 "movq %%mm5, %%mm3 \n\t"
1531 "psllq $8, %%mm5 \n\t"
1532 "pand %%mm0, %%mm3 \n\t"
1533 "por %%mm3, %%mm5 \n\t"
1535 "movq (%0, %%"REG_a"), %%mm0 \n\t"
1536 "movq (%1, %%"REG_a"), %%mm1 \n\t"
1537 "movq 1(%0, %%"REG_a"), %%mm2 \n\t"
1538 "movq 1(%1, %%"REG_a"), %%mm3 \n\t"
1539 PAVGB" %%mm0, %%mm5 \n\t"
1540 PAVGB" %%mm0, %%mm3 \n\t"
1541 PAVGB" %%mm0, %%mm5 \n\t"
1542 PAVGB" %%mm0, %%mm3 \n\t"
1543 PAVGB" %%mm1, %%mm4 \n\t"
1544 PAVGB" %%mm1, %%mm2 \n\t"
1545 PAVGB" %%mm1, %%mm4 \n\t"
1546 PAVGB" %%mm1, %%mm2 \n\t"
1547 "movq %%mm5, %%mm7 \n\t"
1548 "movq %%mm4, %%mm6 \n\t"
1549 "punpcklbw %%mm3, %%mm5 \n\t"
1550 "punpckhbw %%mm3, %%mm7 \n\t"
1551 "punpcklbw %%mm2, %%mm4 \n\t"
1552 "punpckhbw %%mm2, %%mm6 \n\t"
1553 MOVNTQ" %%mm5, (%2, %%"REG_a", 2) \n\t"
1554 MOVNTQ" %%mm7, 8(%2, %%"REG_a", 2) \n\t"
1555 MOVNTQ" %%mm4, (%3, %%"REG_a", 2) \n\t"
1556 MOVNTQ" %%mm6, 8(%3, %%"REG_a", 2) \n\t"
1557 "add $8, %%"REG_a" \n\t"
1558 "movq -1(%0, %%"REG_a"), %%mm4 \n\t"
1559 "movq -1(%1, %%"REG_a"), %%mm5 \n\t"
1561 :: "r" (src + mmxSize ), "r" (src + srcStride + mmxSize ),
1562 "r" (dst + mmxSize*2), "r" (dst + dstStride + mmxSize*2),
1567 for (x=mmxSize-1; x<srcWidth-1; x++) {
1568 dst[2*x +1]= (3*src[x+0] + src[x+srcStride+1])>>2;
1569 dst[2*x+dstStride+2]= ( src[x+0] + 3*src[x+srcStride+1])>>2;
1570 dst[2*x+dstStride+1]= ( src[x+1] + 3*src[x+srcStride ])>>2;
1571 dst[2*x +2]= (3*src[x+1] + src[x+srcStride ])>>2;
1573 dst[srcWidth*2 -1 ]= (3*src[srcWidth-1] + src[srcWidth-1 + srcStride])>>2;
1574 dst[srcWidth*2 -1 + dstStride]= ( src[srcWidth-1] + 3*src[srcWidth-1 + srcStride])>>2;
1583 for (x=0; x<srcWidth-1; x++) {
1584 dst[2*x+1]= (3*src[x] + src[x+1])>>2;
1585 dst[2*x+2]= ( src[x] + 3*src[x+1])>>2;
1587 dst[2*srcWidth-1]= src[srcWidth-1];
1589 __asm__ volatile(EMMS" \n\t"
1593 #endif /* COMPILE_TEMPLATE_MMX2 || COMPILE_TEMPLATE_AMD3DNOW */
1595 #if !COMPILE_TEMPLATE_AMD3DNOW
1597 * Height should be a multiple of 2 and width should be a multiple of 16.
1598 * (If this is a problem for anyone then tell me, and I will fix it.)
1599 * Chrominance data is only taken from every second line, others are ignored.
1600 * FIXME: Write HQ version.
1602 static inline void RENAME(uyvytoyv12)(const uint8_t *src, uint8_t *ydst, uint8_t *udst, uint8_t *vdst,
1603 int width, int height,
1604 int lumStride, int chromStride, int srcStride)
1607 const x86_reg chromWidth= width>>1;
1608 for (y=0; y<height; y+=2) {
1610 "xor %%"REG_a", %%"REG_a" \n\t"
1611 "pcmpeqw %%mm7, %%mm7 \n\t"
1612 "psrlw $8, %%mm7 \n\t" // FF,00,FF,00...
1615 PREFETCH" 64(%0, %%"REG_a", 4) \n\t"
1616 "movq (%0, %%"REG_a", 4), %%mm0 \n\t" // UYVY UYVY(0)
1617 "movq 8(%0, %%"REG_a", 4), %%mm1 \n\t" // UYVY UYVY(4)
1618 "movq %%mm0, %%mm2 \n\t" // UYVY UYVY(0)
1619 "movq %%mm1, %%mm3 \n\t" // UYVY UYVY(4)
1620 "pand %%mm7, %%mm0 \n\t" // U0V0 U0V0(0)
1621 "pand %%mm7, %%mm1 \n\t" // U0V0 U0V0(4)
1622 "psrlw $8, %%mm2 \n\t" // Y0Y0 Y0Y0(0)
1623 "psrlw $8, %%mm3 \n\t" // Y0Y0 Y0Y0(4)
1624 "packuswb %%mm1, %%mm0 \n\t" // UVUV UVUV(0)
1625 "packuswb %%mm3, %%mm2 \n\t" // YYYY YYYY(0)
1627 MOVNTQ" %%mm2, (%1, %%"REG_a", 2) \n\t"
1629 "movq 16(%0, %%"REG_a", 4), %%mm1 \n\t" // UYVY UYVY(8)
1630 "movq 24(%0, %%"REG_a", 4), %%mm2 \n\t" // UYVY UYVY(12)
1631 "movq %%mm1, %%mm3 \n\t" // UYVY UYVY(8)
1632 "movq %%mm2, %%mm4 \n\t" // UYVY UYVY(12)
1633 "pand %%mm7, %%mm1 \n\t" // U0V0 U0V0(8)
1634 "pand %%mm7, %%mm2 \n\t" // U0V0 U0V0(12)
1635 "psrlw $8, %%mm3 \n\t" // Y0Y0 Y0Y0(8)
1636 "psrlw $8, %%mm4 \n\t" // Y0Y0 Y0Y0(12)
1637 "packuswb %%mm2, %%mm1 \n\t" // UVUV UVUV(8)
1638 "packuswb %%mm4, %%mm3 \n\t" // YYYY YYYY(8)
1640 MOVNTQ" %%mm3, 8(%1, %%"REG_a", 2) \n\t"
1642 "movq %%mm0, %%mm2 \n\t" // UVUV UVUV(0)
1643 "movq %%mm1, %%mm3 \n\t" // UVUV UVUV(8)
1644 "psrlw $8, %%mm0 \n\t" // V0V0 V0V0(0)
1645 "psrlw $8, %%mm1 \n\t" // V0V0 V0V0(8)
1646 "pand %%mm7, %%mm2 \n\t" // U0U0 U0U0(0)
1647 "pand %%mm7, %%mm3 \n\t" // U0U0 U0U0(8)
1648 "packuswb %%mm1, %%mm0 \n\t" // VVVV VVVV(0)
1649 "packuswb %%mm3, %%mm2 \n\t" // UUUU UUUU(0)
1651 MOVNTQ" %%mm0, (%3, %%"REG_a") \n\t"
1652 MOVNTQ" %%mm2, (%2, %%"REG_a") \n\t"
1654 "add $8, %%"REG_a" \n\t"
1655 "cmp %4, %%"REG_a" \n\t"
1657 ::"r"(src), "r"(ydst), "r"(udst), "r"(vdst), "g" (chromWidth)
1658 : "memory", "%"REG_a
1665 "xor %%"REG_a", %%"REG_a" \n\t"
1668 PREFETCH" 64(%0, %%"REG_a", 4) \n\t"
1669 "movq (%0, %%"REG_a", 4), %%mm0 \n\t" // YUYV YUYV(0)
1670 "movq 8(%0, %%"REG_a", 4), %%mm1 \n\t" // YUYV YUYV(4)
1671 "movq 16(%0, %%"REG_a", 4), %%mm2 \n\t" // YUYV YUYV(8)
1672 "movq 24(%0, %%"REG_a", 4), %%mm3 \n\t" // YUYV YUYV(12)
1673 "psrlw $8, %%mm0 \n\t" // Y0Y0 Y0Y0(0)
1674 "psrlw $8, %%mm1 \n\t" // Y0Y0 Y0Y0(4)
1675 "psrlw $8, %%mm2 \n\t" // Y0Y0 Y0Y0(8)
1676 "psrlw $8, %%mm3 \n\t" // Y0Y0 Y0Y0(12)
1677 "packuswb %%mm1, %%mm0 \n\t" // YYYY YYYY(0)
1678 "packuswb %%mm3, %%mm2 \n\t" // YYYY YYYY(8)
1680 MOVNTQ" %%mm0, (%1, %%"REG_a", 2) \n\t"
1681 MOVNTQ" %%mm2, 8(%1, %%"REG_a", 2) \n\t"
1683 "add $8, %%"REG_a" \n\t"
1684 "cmp %4, %%"REG_a" \n\t"
1687 ::"r"(src), "r"(ydst), "r"(udst), "r"(vdst), "g" (chromWidth)
1688 : "memory", "%"REG_a
1690 udst += chromStride;
1691 vdst += chromStride;
1695 __asm__ volatile(EMMS" \n\t"
1699 #endif /* !COMPILE_TEMPLATE_AMD3DNOW */
1702 * Height should be a multiple of 2 and width should be a multiple of 2.
1703 * (If this is a problem for anyone then tell me, and I will fix it.)
1704 * Chrominance data is only taken from every second line,
1705 * others are ignored in the C version.
1706 * FIXME: Write HQ version.
1708 static inline void RENAME(rgb24toyv12)(const uint8_t *src, uint8_t *ydst, uint8_t *udst, uint8_t *vdst,
1709 int width, int height,
1710 int lumStride, int chromStride, int srcStride)
1713 const x86_reg chromWidth= width>>1;
1714 for (y=0; y<height-2; y+=2) {
1716 for (i=0; i<2; i++) {
1718 "mov %2, %%"REG_a" \n\t"
1719 "movq "MANGLE(ff_bgr2YCoeff)", %%mm6 \n\t"
1720 "movq "MANGLE(ff_w1111)", %%mm5 \n\t"
1721 "pxor %%mm7, %%mm7 \n\t"
1722 "lea (%%"REG_a", %%"REG_a", 2), %%"REG_d" \n\t"
1725 PREFETCH" 64(%0, %%"REG_d") \n\t"
1726 "movd (%0, %%"REG_d"), %%mm0 \n\t"
1727 "movd 3(%0, %%"REG_d"), %%mm1 \n\t"
1728 "punpcklbw %%mm7, %%mm0 \n\t"
1729 "punpcklbw %%mm7, %%mm1 \n\t"
1730 "movd 6(%0, %%"REG_d"), %%mm2 \n\t"
1731 "movd 9(%0, %%"REG_d"), %%mm3 \n\t"
1732 "punpcklbw %%mm7, %%mm2 \n\t"
1733 "punpcklbw %%mm7, %%mm3 \n\t"
1734 "pmaddwd %%mm6, %%mm0 \n\t"
1735 "pmaddwd %%mm6, %%mm1 \n\t"
1736 "pmaddwd %%mm6, %%mm2 \n\t"
1737 "pmaddwd %%mm6, %%mm3 \n\t"
1738 #ifndef FAST_BGR2YV12
1739 "psrad $8, %%mm0 \n\t"
1740 "psrad $8, %%mm1 \n\t"
1741 "psrad $8, %%mm2 \n\t"
1742 "psrad $8, %%mm3 \n\t"
1744 "packssdw %%mm1, %%mm0 \n\t"
1745 "packssdw %%mm3, %%mm2 \n\t"
1746 "pmaddwd %%mm5, %%mm0 \n\t"
1747 "pmaddwd %%mm5, %%mm2 \n\t"
1748 "packssdw %%mm2, %%mm0 \n\t"
1749 "psraw $7, %%mm0 \n\t"
1751 "movd 12(%0, %%"REG_d"), %%mm4 \n\t"
1752 "movd 15(%0, %%"REG_d"), %%mm1 \n\t"
1753 "punpcklbw %%mm7, %%mm4 \n\t"
1754 "punpcklbw %%mm7, %%mm1 \n\t"
1755 "movd 18(%0, %%"REG_d"), %%mm2 \n\t"
1756 "movd 21(%0, %%"REG_d"), %%mm3 \n\t"
1757 "punpcklbw %%mm7, %%mm2 \n\t"
1758 "punpcklbw %%mm7, %%mm3 \n\t"
1759 "pmaddwd %%mm6, %%mm4 \n\t"
1760 "pmaddwd %%mm6, %%mm1 \n\t"
1761 "pmaddwd %%mm6, %%mm2 \n\t"
1762 "pmaddwd %%mm6, %%mm3 \n\t"
1763 #ifndef FAST_BGR2YV12
1764 "psrad $8, %%mm4 \n\t"
1765 "psrad $8, %%mm1 \n\t"
1766 "psrad $8, %%mm2 \n\t"
1767 "psrad $8, %%mm3 \n\t"
1769 "packssdw %%mm1, %%mm4 \n\t"
1770 "packssdw %%mm3, %%mm2 \n\t"
1771 "pmaddwd %%mm5, %%mm4 \n\t"
1772 "pmaddwd %%mm5, %%mm2 \n\t"
1773 "add $24, %%"REG_d" \n\t"
1774 "packssdw %%mm2, %%mm4 \n\t"
1775 "psraw $7, %%mm4 \n\t"
1777 "packuswb %%mm4, %%mm0 \n\t"
1778 "paddusb "MANGLE(ff_bgr2YOffset)", %%mm0 \n\t"
1780 MOVNTQ" %%mm0, (%1, %%"REG_a") \n\t"
1781 "add $8, %%"REG_a" \n\t"
1783 : : "r" (src+width*3), "r" (ydst+width), "g" ((x86_reg)-width)
1784 : "%"REG_a, "%"REG_d
1791 "mov %4, %%"REG_a" \n\t"
1792 "movq "MANGLE(ff_w1111)", %%mm5 \n\t"
1793 "movq "MANGLE(ff_bgr2UCoeff)", %%mm6 \n\t"
1794 "pxor %%mm7, %%mm7 \n\t"
1795 "lea (%%"REG_a", %%"REG_a", 2), %%"REG_d" \n\t"
1796 "add %%"REG_d", %%"REG_d" \n\t"
1799 PREFETCH" 64(%0, %%"REG_d") \n\t"
1800 PREFETCH" 64(%1, %%"REG_d") \n\t"
1801 #if COMPILE_TEMPLATE_MMX2 || COMPILE_TEMPLATE_AMD3DNOW
1802 "movq (%0, %%"REG_d"), %%mm0 \n\t"
1803 "movq (%1, %%"REG_d"), %%mm1 \n\t"
1804 "movq 6(%0, %%"REG_d"), %%mm2 \n\t"
1805 "movq 6(%1, %%"REG_d"), %%mm3 \n\t"
1806 PAVGB" %%mm1, %%mm0 \n\t"
1807 PAVGB" %%mm3, %%mm2 \n\t"
1808 "movq %%mm0, %%mm1 \n\t"
1809 "movq %%mm2, %%mm3 \n\t"
1810 "psrlq $24, %%mm0 \n\t"
1811 "psrlq $24, %%mm2 \n\t"
1812 PAVGB" %%mm1, %%mm0 \n\t"
1813 PAVGB" %%mm3, %%mm2 \n\t"
1814 "punpcklbw %%mm7, %%mm0 \n\t"
1815 "punpcklbw %%mm7, %%mm2 \n\t"
1817 "movd (%0, %%"REG_d"), %%mm0 \n\t"
1818 "movd (%1, %%"REG_d"), %%mm1 \n\t"
1819 "movd 3(%0, %%"REG_d"), %%mm2 \n\t"
1820 "movd 3(%1, %%"REG_d"), %%mm3 \n\t"
1821 "punpcklbw %%mm7, %%mm0 \n\t"
1822 "punpcklbw %%mm7, %%mm1 \n\t"
1823 "punpcklbw %%mm7, %%mm2 \n\t"
1824 "punpcklbw %%mm7, %%mm3 \n\t"
1825 "paddw %%mm1, %%mm0 \n\t"
1826 "paddw %%mm3, %%mm2 \n\t"
1827 "paddw %%mm2, %%mm0 \n\t"
1828 "movd 6(%0, %%"REG_d"), %%mm4 \n\t"
1829 "movd 6(%1, %%"REG_d"), %%mm1 \n\t"
1830 "movd 9(%0, %%"REG_d"), %%mm2 \n\t"
1831 "movd 9(%1, %%"REG_d"), %%mm3 \n\t"
1832 "punpcklbw %%mm7, %%mm4 \n\t"
1833 "punpcklbw %%mm7, %%mm1 \n\t"
1834 "punpcklbw %%mm7, %%mm2 \n\t"
1835 "punpcklbw %%mm7, %%mm3 \n\t"
1836 "paddw %%mm1, %%mm4 \n\t"
1837 "paddw %%mm3, %%mm2 \n\t"
1838 "paddw %%mm4, %%mm2 \n\t"
1839 "psrlw $2, %%mm0 \n\t"
1840 "psrlw $2, %%mm2 \n\t"
1842 "movq "MANGLE(ff_bgr2VCoeff)", %%mm1 \n\t"
1843 "movq "MANGLE(ff_bgr2VCoeff)", %%mm3 \n\t"
1845 "pmaddwd %%mm0, %%mm1 \n\t"
1846 "pmaddwd %%mm2, %%mm3 \n\t"
1847 "pmaddwd %%mm6, %%mm0 \n\t"
1848 "pmaddwd %%mm6, %%mm2 \n\t"
1849 #ifndef FAST_BGR2YV12
1850 "psrad $8, %%mm0 \n\t"
1851 "psrad $8, %%mm1 \n\t"
1852 "psrad $8, %%mm2 \n\t"
1853 "psrad $8, %%mm3 \n\t"
1855 "packssdw %%mm2, %%mm0 \n\t"
1856 "packssdw %%mm3, %%mm1 \n\t"
1857 "pmaddwd %%mm5, %%mm0 \n\t"
1858 "pmaddwd %%mm5, %%mm1 \n\t"
1859 "packssdw %%mm1, %%mm0 \n\t" // V1 V0 U1 U0
1860 "psraw $7, %%mm0 \n\t"
1862 #if COMPILE_TEMPLATE_MMX2 || COMPILE_TEMPLATE_AMD3DNOW
1863 "movq 12(%0, %%"REG_d"), %%mm4 \n\t"
1864 "movq 12(%1, %%"REG_d"), %%mm1 \n\t"
1865 "movq 18(%0, %%"REG_d"), %%mm2 \n\t"
1866 "movq 18(%1, %%"REG_d"), %%mm3 \n\t"
1867 PAVGB" %%mm1, %%mm4 \n\t"
1868 PAVGB" %%mm3, %%mm2 \n\t"
1869 "movq %%mm4, %%mm1 \n\t"
1870 "movq %%mm2, %%mm3 \n\t"
1871 "psrlq $24, %%mm4 \n\t"
1872 "psrlq $24, %%mm2 \n\t"
1873 PAVGB" %%mm1, %%mm4 \n\t"
1874 PAVGB" %%mm3, %%mm2 \n\t"
1875 "punpcklbw %%mm7, %%mm4 \n\t"
1876 "punpcklbw %%mm7, %%mm2 \n\t"
1878 "movd 12(%0, %%"REG_d"), %%mm4 \n\t"
1879 "movd 12(%1, %%"REG_d"), %%mm1 \n\t"
1880 "movd 15(%0, %%"REG_d"), %%mm2 \n\t"
1881 "movd 15(%1, %%"REG_d"), %%mm3 \n\t"
1882 "punpcklbw %%mm7, %%mm4 \n\t"
1883 "punpcklbw %%mm7, %%mm1 \n\t"
1884 "punpcklbw %%mm7, %%mm2 \n\t"
1885 "punpcklbw %%mm7, %%mm3 \n\t"
1886 "paddw %%mm1, %%mm4 \n\t"
1887 "paddw %%mm3, %%mm2 \n\t"
1888 "paddw %%mm2, %%mm4 \n\t"
1889 "movd 18(%0, %%"REG_d"), %%mm5 \n\t"
1890 "movd 18(%1, %%"REG_d"), %%mm1 \n\t"
1891 "movd 21(%0, %%"REG_d"), %%mm2 \n\t"
1892 "movd 21(%1, %%"REG_d"), %%mm3 \n\t"
1893 "punpcklbw %%mm7, %%mm5 \n\t"
1894 "punpcklbw %%mm7, %%mm1 \n\t"
1895 "punpcklbw %%mm7, %%mm2 \n\t"
1896 "punpcklbw %%mm7, %%mm3 \n\t"
1897 "paddw %%mm1, %%mm5 \n\t"
1898 "paddw %%mm3, %%mm2 \n\t"
1899 "paddw %%mm5, %%mm2 \n\t"
1900 "movq "MANGLE(ff_w1111)", %%mm5 \n\t"
1901 "psrlw $2, %%mm4 \n\t"
1902 "psrlw $2, %%mm2 \n\t"
1904 "movq "MANGLE(ff_bgr2VCoeff)", %%mm1 \n\t"
1905 "movq "MANGLE(ff_bgr2VCoeff)", %%mm3 \n\t"
1907 "pmaddwd %%mm4, %%mm1 \n\t"
1908 "pmaddwd %%mm2, %%mm3 \n\t"
1909 "pmaddwd %%mm6, %%mm4 \n\t"
1910 "pmaddwd %%mm6, %%mm2 \n\t"
1911 #ifndef FAST_BGR2YV12
1912 "psrad $8, %%mm4 \n\t"
1913 "psrad $8, %%mm1 \n\t"
1914 "psrad $8, %%mm2 \n\t"
1915 "psrad $8, %%mm3 \n\t"
1917 "packssdw %%mm2, %%mm4 \n\t"
1918 "packssdw %%mm3, %%mm1 \n\t"
1919 "pmaddwd %%mm5, %%mm4 \n\t"
1920 "pmaddwd %%mm5, %%mm1 \n\t"
1921 "add $24, %%"REG_d" \n\t"
1922 "packssdw %%mm1, %%mm4 \n\t" // V3 V2 U3 U2
1923 "psraw $7, %%mm4 \n\t"
1925 "movq %%mm0, %%mm1 \n\t"
1926 "punpckldq %%mm4, %%mm0 \n\t"
1927 "punpckhdq %%mm4, %%mm1 \n\t"
1928 "packsswb %%mm1, %%mm0 \n\t"
1929 "paddb "MANGLE(ff_bgr2UVOffset)", %%mm0 \n\t"
1930 "movd %%mm0, (%2, %%"REG_a") \n\t"
1931 "punpckhdq %%mm0, %%mm0 \n\t"
1932 "movd %%mm0, (%3, %%"REG_a") \n\t"
1933 "add $4, %%"REG_a" \n\t"
1935 : : "r" (src+chromWidth*6), "r" (src+srcStride+chromWidth*6), "r" (udst+chromWidth), "r" (vdst+chromWidth), "g" (-chromWidth)
1936 : "%"REG_a, "%"REG_d
1939 udst += chromStride;
1940 vdst += chromStride;
1944 __asm__ volatile(EMMS" \n\t"
1948 rgb24toyv12_c(src, ydst, udst, vdst, width, height-y, lumStride, chromStride, srcStride);
1950 #endif /* !COMPILE_TEMPLATE_SSE2 */
1952 #if !COMPILE_TEMPLATE_AMD3DNOW
1953 static void RENAME(interleaveBytes)(const uint8_t *src1, const uint8_t *src2, uint8_t *dest,
1954 int width, int height, int src1Stride,
1955 int src2Stride, int dstStride)
1959 for (h=0; h < height; h++) {
1962 #if COMPILE_TEMPLATE_SSE2
1964 "xor %%"REG_a", %%"REG_a" \n\t"
1966 PREFETCH" 64(%1, %%"REG_a") \n\t"
1967 PREFETCH" 64(%2, %%"REG_a") \n\t"
1968 "movdqa (%1, %%"REG_a"), %%xmm0 \n\t"
1969 "movdqa (%1, %%"REG_a"), %%xmm1 \n\t"
1970 "movdqa (%2, %%"REG_a"), %%xmm2 \n\t"
1971 "punpcklbw %%xmm2, %%xmm0 \n\t"
1972 "punpckhbw %%xmm2, %%xmm1 \n\t"
1973 "movntdq %%xmm0, (%0, %%"REG_a", 2) \n\t"
1974 "movntdq %%xmm1, 16(%0, %%"REG_a", 2) \n\t"
1975 "add $16, %%"REG_a" \n\t"
1976 "cmp %3, %%"REG_a" \n\t"
1978 ::"r"(dest), "r"(src1), "r"(src2), "r" ((x86_reg)width-15)
1979 : "memory", "%"REG_a""
1983 "xor %%"REG_a", %%"REG_a" \n\t"
1985 PREFETCH" 64(%1, %%"REG_a") \n\t"
1986 PREFETCH" 64(%2, %%"REG_a") \n\t"
1987 "movq (%1, %%"REG_a"), %%mm0 \n\t"
1988 "movq 8(%1, %%"REG_a"), %%mm2 \n\t"
1989 "movq %%mm0, %%mm1 \n\t"
1990 "movq %%mm2, %%mm3 \n\t"
1991 "movq (%2, %%"REG_a"), %%mm4 \n\t"
1992 "movq 8(%2, %%"REG_a"), %%mm5 \n\t"
1993 "punpcklbw %%mm4, %%mm0 \n\t"
1994 "punpckhbw %%mm4, %%mm1 \n\t"
1995 "punpcklbw %%mm5, %%mm2 \n\t"
1996 "punpckhbw %%mm5, %%mm3 \n\t"
1997 MOVNTQ" %%mm0, (%0, %%"REG_a", 2) \n\t"
1998 MOVNTQ" %%mm1, 8(%0, %%"REG_a", 2) \n\t"
1999 MOVNTQ" %%mm2, 16(%0, %%"REG_a", 2) \n\t"
2000 MOVNTQ" %%mm3, 24(%0, %%"REG_a", 2) \n\t"
2001 "add $16, %%"REG_a" \n\t"
2002 "cmp %3, %%"REG_a" \n\t"
2004 ::"r"(dest), "r"(src1), "r"(src2), "r" ((x86_reg)width-15)
2005 : "memory", "%"REG_a
2008 for (w= (width&(~15)); w < width; w++) {
2009 dest[2*w+0] = src1[w];
2010 dest[2*w+1] = src2[w];
2022 #endif /* !COMPILE_TEMPLATE_AMD3DNOW */
2024 #if !COMPILE_TEMPLATE_SSE2
2025 #if !COMPILE_TEMPLATE_AMD3DNOW
2026 static inline void RENAME(vu9_to_vu12)(const uint8_t *src1, const uint8_t *src2,
2027 uint8_t *dst1, uint8_t *dst2,
2028 int width, int height,
2029 int srcStride1, int srcStride2,
2030 int dstStride1, int dstStride2)
2034 w=width/2; h=height/2;
2038 ::"m"(*(src1+srcStride1)),"m"(*(src2+srcStride2)):"memory");
2040 const uint8_t* s1=src1+srcStride1*(y>>1);
2041 uint8_t* d=dst1+dstStride1*y;
2043 for (;x<w-31;x+=32) {
2045 PREFETCH" 32%1 \n\t"
2046 "movq %1, %%mm0 \n\t"
2047 "movq 8%1, %%mm2 \n\t"
2048 "movq 16%1, %%mm4 \n\t"
2049 "movq 24%1, %%mm6 \n\t"
2050 "movq %%mm0, %%mm1 \n\t"
2051 "movq %%mm2, %%mm3 \n\t"
2052 "movq %%mm4, %%mm5 \n\t"
2053 "movq %%mm6, %%mm7 \n\t"
2054 "punpcklbw %%mm0, %%mm0 \n\t"
2055 "punpckhbw %%mm1, %%mm1 \n\t"
2056 "punpcklbw %%mm2, %%mm2 \n\t"
2057 "punpckhbw %%mm3, %%mm3 \n\t"
2058 "punpcklbw %%mm4, %%mm4 \n\t"
2059 "punpckhbw %%mm5, %%mm5 \n\t"
2060 "punpcklbw %%mm6, %%mm6 \n\t"
2061 "punpckhbw %%mm7, %%mm7 \n\t"
2062 MOVNTQ" %%mm0, %0 \n\t"
2063 MOVNTQ" %%mm1, 8%0 \n\t"
2064 MOVNTQ" %%mm2, 16%0 \n\t"
2065 MOVNTQ" %%mm3, 24%0 \n\t"
2066 MOVNTQ" %%mm4, 32%0 \n\t"
2067 MOVNTQ" %%mm5, 40%0 \n\t"
2068 MOVNTQ" %%mm6, 48%0 \n\t"
2069 MOVNTQ" %%mm7, 56%0"
2074 for (;x<w;x++) d[2*x]=d[2*x+1]=s1[x];
2077 const uint8_t* s2=src2+srcStride2*(y>>1);
2078 uint8_t* d=dst2+dstStride2*y;
2080 for (;x<w-31;x+=32) {
2082 PREFETCH" 32%1 \n\t"
2083 "movq %1, %%mm0 \n\t"
2084 "movq 8%1, %%mm2 \n\t"
2085 "movq 16%1, %%mm4 \n\t"
2086 "movq 24%1, %%mm6 \n\t"
2087 "movq %%mm0, %%mm1 \n\t"
2088 "movq %%mm2, %%mm3 \n\t"
2089 "movq %%mm4, %%mm5 \n\t"
2090 "movq %%mm6, %%mm7 \n\t"
2091 "punpcklbw %%mm0, %%mm0 \n\t"
2092 "punpckhbw %%mm1, %%mm1 \n\t"
2093 "punpcklbw %%mm2, %%mm2 \n\t"
2094 "punpckhbw %%mm3, %%mm3 \n\t"
2095 "punpcklbw %%mm4, %%mm4 \n\t"
2096 "punpckhbw %%mm5, %%mm5 \n\t"
2097 "punpcklbw %%mm6, %%mm6 \n\t"
2098 "punpckhbw %%mm7, %%mm7 \n\t"
2099 MOVNTQ" %%mm0, %0 \n\t"
2100 MOVNTQ" %%mm1, 8%0 \n\t"
2101 MOVNTQ" %%mm2, 16%0 \n\t"
2102 MOVNTQ" %%mm3, 24%0 \n\t"
2103 MOVNTQ" %%mm4, 32%0 \n\t"
2104 MOVNTQ" %%mm5, 40%0 \n\t"
2105 MOVNTQ" %%mm6, 48%0 \n\t"
2106 MOVNTQ" %%mm7, 56%0"
2111 for (;x<w;x++) d[2*x]=d[2*x+1]=s2[x];
2120 static inline void RENAME(yvu9_to_yuy2)(const uint8_t *src1, const uint8_t *src2, const uint8_t *src3,
2122 int width, int height,
2123 int srcStride1, int srcStride2,
2124 int srcStride3, int dstStride)
2128 w=width/2; h=height;
2130 const uint8_t* yp=src1+srcStride1*y;
2131 const uint8_t* up=src2+srcStride2*(y>>2);
2132 const uint8_t* vp=src3+srcStride3*(y>>2);
2133 uint8_t* d=dst+dstStride*y;
2137 PREFETCH" 32(%1, %0) \n\t"
2138 PREFETCH" 32(%2, %0) \n\t"
2139 PREFETCH" 32(%3, %0) \n\t"
2140 "movq (%1, %0, 4), %%mm0 \n\t" /* Y0Y1Y2Y3Y4Y5Y6Y7 */
2141 "movq (%2, %0), %%mm1 \n\t" /* U0U1U2U3U4U5U6U7 */
2142 "movq (%3, %0), %%mm2 \n\t" /* V0V1V2V3V4V5V6V7 */
2143 "movq %%mm0, %%mm3 \n\t" /* Y0Y1Y2Y3Y4Y5Y6Y7 */
2144 "movq %%mm1, %%mm4 \n\t" /* U0U1U2U3U4U5U6U7 */
2145 "movq %%mm2, %%mm5 \n\t" /* V0V1V2V3V4V5V6V7 */
2146 "punpcklbw %%mm1, %%mm1 \n\t" /* U0U0 U1U1 U2U2 U3U3 */
2147 "punpcklbw %%mm2, %%mm2 \n\t" /* V0V0 V1V1 V2V2 V3V3 */
2148 "punpckhbw %%mm4, %%mm4 \n\t" /* U4U4 U5U5 U6U6 U7U7 */
2149 "punpckhbw %%mm5, %%mm5 \n\t" /* V4V4 V5V5 V6V6 V7V7 */
2151 "movq %%mm1, %%mm6 \n\t"
2152 "punpcklbw %%mm2, %%mm1 \n\t" /* U0V0 U0V0 U1V1 U1V1*/
2153 "punpcklbw %%mm1, %%mm0 \n\t" /* Y0U0 Y1V0 Y2U0 Y3V0*/
2154 "punpckhbw %%mm1, %%mm3 \n\t" /* Y4U1 Y5V1 Y6U1 Y7V1*/
2155 MOVNTQ" %%mm0, (%4, %0, 8) \n\t"
2156 MOVNTQ" %%mm3, 8(%4, %0, 8) \n\t"
2158 "punpckhbw %%mm2, %%mm6 \n\t" /* U2V2 U2V2 U3V3 U3V3*/
2159 "movq 8(%1, %0, 4), %%mm0 \n\t"
2160 "movq %%mm0, %%mm3 \n\t"
2161 "punpcklbw %%mm6, %%mm0 \n\t" /* Y U2 Y V2 Y U2 Y V2*/
2162 "punpckhbw %%mm6, %%mm3 \n\t" /* Y U3 Y V3 Y U3 Y V3*/
2163 MOVNTQ" %%mm0, 16(%4, %0, 8) \n\t"
2164 MOVNTQ" %%mm3, 24(%4, %0, 8) \n\t"
2166 "movq %%mm4, %%mm6 \n\t"
2167 "movq 16(%1, %0, 4), %%mm0 \n\t"
2168 "movq %%mm0, %%mm3 \n\t"
2169 "punpcklbw %%mm5, %%mm4 \n\t"
2170 "punpcklbw %%mm4, %%mm0 \n\t" /* Y U4 Y V4 Y U4 Y V4*/
2171 "punpckhbw %%mm4, %%mm3 \n\t" /* Y U5 Y V5 Y U5 Y V5*/
2172 MOVNTQ" %%mm0, 32(%4, %0, 8) \n\t"
2173 MOVNTQ" %%mm3, 40(%4, %0, 8) \n\t"
2175 "punpckhbw %%mm5, %%mm6 \n\t"
2176 "movq 24(%1, %0, 4), %%mm0 \n\t"
2177 "movq %%mm0, %%mm3 \n\t"
2178 "punpcklbw %%mm6, %%mm0 \n\t" /* Y U6 Y V6 Y U6 Y V6*/
2179 "punpckhbw %%mm6, %%mm3 \n\t" /* Y U7 Y V7 Y U7 Y V7*/
2180 MOVNTQ" %%mm0, 48(%4, %0, 8) \n\t"
2181 MOVNTQ" %%mm3, 56(%4, %0, 8) \n\t"
2184 : "r"(yp), "r" (up), "r"(vp), "r"(d)
2188 const int x2 = x<<2;
2191 d[8*x+2] = yp[x2+1];
2193 d[8*x+4] = yp[x2+2];
2195 d[8*x+6] = yp[x2+3];
2205 #endif /* !COMPILE_TEMPLATE_AMD3DNOW */
2207 static void RENAME(extract_even)(const uint8_t *src, uint8_t *dst, x86_reg count)
2216 "pcmpeqw %%mm7, %%mm7 \n\t"
2217 "psrlw $8, %%mm7 \n\t"
2219 "movq -30(%1, %0, 2), %%mm0 \n\t"
2220 "movq -22(%1, %0, 2), %%mm1 \n\t"
2221 "movq -14(%1, %0, 2), %%mm2 \n\t"
2222 "movq -6(%1, %0, 2), %%mm3 \n\t"
2223 "pand %%mm7, %%mm0 \n\t"
2224 "pand %%mm7, %%mm1 \n\t"
2225 "pand %%mm7, %%mm2 \n\t"
2226 "pand %%mm7, %%mm3 \n\t"
2227 "packuswb %%mm1, %%mm0 \n\t"
2228 "packuswb %%mm3, %%mm2 \n\t"
2229 MOVNTQ" %%mm0,-15(%2, %0) \n\t"
2230 MOVNTQ" %%mm2,- 7(%2, %0) \n\t"
2234 : "r"(src), "r"(dst)
2239 dst[count]= src[2*count];
2244 #if !COMPILE_TEMPLATE_AMD3DNOW
2245 static void RENAME(extract_even2)(const uint8_t *src, uint8_t *dst0, uint8_t *dst1, x86_reg count)
2254 "pcmpeqw %%mm7, %%mm7 \n\t"
2255 "psrlw $8, %%mm7 \n\t"
2257 "movq -28(%1, %0, 4), %%mm0 \n\t"
2258 "movq -20(%1, %0, 4), %%mm1 \n\t"
2259 "movq -12(%1, %0, 4), %%mm2 \n\t"
2260 "movq -4(%1, %0, 4), %%mm3 \n\t"
2261 "pand %%mm7, %%mm0 \n\t"
2262 "pand %%mm7, %%mm1 \n\t"
2263 "pand %%mm7, %%mm2 \n\t"
2264 "pand %%mm7, %%mm3 \n\t"
2265 "packuswb %%mm1, %%mm0 \n\t"
2266 "packuswb %%mm3, %%mm2 \n\t"
2267 "movq %%mm0, %%mm1 \n\t"
2268 "movq %%mm2, %%mm3 \n\t"
2269 "psrlw $8, %%mm0 \n\t"
2270 "psrlw $8, %%mm2 \n\t"
2271 "pand %%mm7, %%mm1 \n\t"
2272 "pand %%mm7, %%mm3 \n\t"
2273 "packuswb %%mm2, %%mm0 \n\t"
2274 "packuswb %%mm3, %%mm1 \n\t"
2275 MOVNTQ" %%mm0,- 7(%3, %0) \n\t"
2276 MOVNTQ" %%mm1,- 7(%2, %0) \n\t"
2280 : "r"(src), "r"(dst0), "r"(dst1)
2285 dst0[count]= src[4*count+0];
2286 dst1[count]= src[4*count+2];
2290 #endif /* !COMPILE_TEMPLATE_AMD3DNOW */
2292 static void RENAME(extract_even2avg)(const uint8_t *src0, const uint8_t *src1, uint8_t *dst0, uint8_t *dst1, x86_reg count)
2303 "pcmpeqw %%mm7, %%mm7 \n\t"
2304 "psrlw $8, %%mm7 \n\t"
2306 "movq -28(%1, %0, 4), %%mm0 \n\t"
2307 "movq -20(%1, %0, 4), %%mm1 \n\t"
2308 "movq -12(%1, %0, 4), %%mm2 \n\t"
2309 "movq -4(%1, %0, 4), %%mm3 \n\t"
2310 PAVGB" -28(%2, %0, 4), %%mm0 \n\t"
2311 PAVGB" -20(%2, %0, 4), %%mm1 \n\t"
2312 PAVGB" -12(%2, %0, 4), %%mm2 \n\t"
2313 PAVGB" - 4(%2, %0, 4), %%mm3 \n\t"
2314 "pand %%mm7, %%mm0 \n\t"
2315 "pand %%mm7, %%mm1 \n\t"
2316 "pand %%mm7, %%mm2 \n\t"
2317 "pand %%mm7, %%mm3 \n\t"
2318 "packuswb %%mm1, %%mm0 \n\t"
2319 "packuswb %%mm3, %%mm2 \n\t"
2320 "movq %%mm0, %%mm1 \n\t"
2321 "movq %%mm2, %%mm3 \n\t"
2322 "psrlw $8, %%mm0 \n\t"
2323 "psrlw $8, %%mm2 \n\t"
2324 "pand %%mm7, %%mm1 \n\t"
2325 "pand %%mm7, %%mm3 \n\t"
2326 "packuswb %%mm2, %%mm0 \n\t"
2327 "packuswb %%mm3, %%mm1 \n\t"
2328 MOVNTQ" %%mm0,- 7(%4, %0) \n\t"
2329 MOVNTQ" %%mm1,- 7(%3, %0) \n\t"
2333 : "r"(src0), "r"(src1), "r"(dst0), "r"(dst1)
2339 dst0[count]= (src0[4*count+0]+src1[4*count+0])>>1;
2340 dst1[count]= (src0[4*count+2]+src1[4*count+2])>>1;
2345 #if !COMPILE_TEMPLATE_AMD3DNOW
2346 static void RENAME(extract_odd2)(const uint8_t *src, uint8_t *dst0, uint8_t *dst1, x86_reg count)
2355 "pcmpeqw %%mm7, %%mm7 \n\t"
2356 "psrlw $8, %%mm7 \n\t"
2358 "movq -28(%1, %0, 4), %%mm0 \n\t"
2359 "movq -20(%1, %0, 4), %%mm1 \n\t"
2360 "movq -12(%1, %0, 4), %%mm2 \n\t"
2361 "movq -4(%1, %0, 4), %%mm3 \n\t"
2362 "psrlw $8, %%mm0 \n\t"
2363 "psrlw $8, %%mm1 \n\t"
2364 "psrlw $8, %%mm2 \n\t"
2365 "psrlw $8, %%mm3 \n\t"
2366 "packuswb %%mm1, %%mm0 \n\t"
2367 "packuswb %%mm3, %%mm2 \n\t"
2368 "movq %%mm0, %%mm1 \n\t"
2369 "movq %%mm2, %%mm3 \n\t"
2370 "psrlw $8, %%mm0 \n\t"
2371 "psrlw $8, %%mm2 \n\t"
2372 "pand %%mm7, %%mm1 \n\t"
2373 "pand %%mm7, %%mm3 \n\t"
2374 "packuswb %%mm2, %%mm0 \n\t"
2375 "packuswb %%mm3, %%mm1 \n\t"
2376 MOVNTQ" %%mm0,- 7(%3, %0) \n\t"
2377 MOVNTQ" %%mm1,- 7(%2, %0) \n\t"
2381 : "r"(src), "r"(dst0), "r"(dst1)
2387 dst0[count]= src[4*count+0];
2388 dst1[count]= src[4*count+2];
2392 #endif /* !COMPILE_TEMPLATE_AMD3DNOW */
2394 static void RENAME(extract_odd2avg)(const uint8_t *src0, const uint8_t *src1, uint8_t *dst0, uint8_t *dst1, x86_reg count)
2405 "pcmpeqw %%mm7, %%mm7 \n\t"
2406 "psrlw $8, %%mm7 \n\t"
2408 "movq -28(%1, %0, 4), %%mm0 \n\t"
2409 "movq -20(%1, %0, 4), %%mm1 \n\t"
2410 "movq -12(%1, %0, 4), %%mm2 \n\t"
2411 "movq -4(%1, %0, 4), %%mm3 \n\t"
2412 PAVGB" -28(%2, %0, 4), %%mm0 \n\t"
2413 PAVGB" -20(%2, %0, 4), %%mm1 \n\t"
2414 PAVGB" -12(%2, %0, 4), %%mm2 \n\t"
2415 PAVGB" - 4(%2, %0, 4), %%mm3 \n\t"
2416 "psrlw $8, %%mm0 \n\t"
2417 "psrlw $8, %%mm1 \n\t"
2418 "psrlw $8, %%mm2 \n\t"
2419 "psrlw $8, %%mm3 \n\t"
2420 "packuswb %%mm1, %%mm0 \n\t"
2421 "packuswb %%mm3, %%mm2 \n\t"
2422 "movq %%mm0, %%mm1 \n\t"
2423 "movq %%mm2, %%mm3 \n\t"
2424 "psrlw $8, %%mm0 \n\t"
2425 "psrlw $8, %%mm2 \n\t"
2426 "pand %%mm7, %%mm1 \n\t"
2427 "pand %%mm7, %%mm3 \n\t"
2428 "packuswb %%mm2, %%mm0 \n\t"
2429 "packuswb %%mm3, %%mm1 \n\t"
2430 MOVNTQ" %%mm0,- 7(%4, %0) \n\t"
2431 MOVNTQ" %%mm1,- 7(%3, %0) \n\t"
2435 : "r"(src0), "r"(src1), "r"(dst0), "r"(dst1)
2443 dst0[count]= (src0[4*count+0]+src1[4*count+0])>>1;
2444 dst1[count]= (src0[4*count+2]+src1[4*count+2])>>1;
2449 static void RENAME(yuyvtoyuv420)(uint8_t *ydst, uint8_t *udst, uint8_t *vdst, const uint8_t *src,
2450 int width, int height,
2451 int lumStride, int chromStride, int srcStride)
2454 const int chromWidth= -((-width)>>1);
2456 for (y=0; y<height; y++) {
2457 RENAME(extract_even)(src, ydst, width);
2459 RENAME(extract_odd2avg)(src-srcStride, src, udst, vdst, chromWidth);
2474 #if !COMPILE_TEMPLATE_AMD3DNOW
2475 static void RENAME(yuyvtoyuv422)(uint8_t *ydst, uint8_t *udst, uint8_t *vdst, const uint8_t *src,
2476 int width, int height,
2477 int lumStride, int chromStride, int srcStride)
2480 const int chromWidth= -((-width)>>1);
2482 for (y=0; y<height; y++) {
2483 RENAME(extract_even)(src, ydst, width);
2484 RENAME(extract_odd2)(src, udst, vdst, chromWidth);
2497 #endif /* !COMPILE_TEMPLATE_AMD3DNOW */
2499 static void RENAME(uyvytoyuv420)(uint8_t *ydst, uint8_t *udst, uint8_t *vdst, const uint8_t *src,
2500 int width, int height,
2501 int lumStride, int chromStride, int srcStride)
2504 const int chromWidth= -((-width)>>1);
2506 for (y=0; y<height; y++) {
2507 RENAME(extract_even)(src+1, ydst, width);
2509 RENAME(extract_even2avg)(src-srcStride, src, udst, vdst, chromWidth);
2524 #if !COMPILE_TEMPLATE_AMD3DNOW
2525 static void RENAME(uyvytoyuv422)(uint8_t *ydst, uint8_t *udst, uint8_t *vdst, const uint8_t *src,
2526 int width, int height,
2527 int lumStride, int chromStride, int srcStride)
2530 const int chromWidth= -((-width)>>1);
2532 for (y=0; y<height; y++) {
2533 RENAME(extract_even)(src+1, ydst, width);
2534 RENAME(extract_even2)(src, udst, vdst, chromWidth);
2547 #endif /* !COMPILE_TEMPLATE_AMD3DNOW */
2548 #endif /* !COMPILE_TEMPLATE_SSE2 */
2550 static inline void RENAME(rgb2rgb_init)(void)
2552 #if !COMPILE_TEMPLATE_SSE2
2553 #if !COMPILE_TEMPLATE_AMD3DNOW
2554 rgb15to16 = RENAME(rgb15to16);
2555 rgb15tobgr24 = RENAME(rgb15tobgr24);
2556 rgb15to32 = RENAME(rgb15to32);
2557 rgb16tobgr24 = RENAME(rgb16tobgr24);
2558 rgb16to32 = RENAME(rgb16to32);
2559 rgb16to15 = RENAME(rgb16to15);
2560 rgb24tobgr16 = RENAME(rgb24tobgr16);
2561 rgb24tobgr15 = RENAME(rgb24tobgr15);
2562 rgb24tobgr32 = RENAME(rgb24tobgr32);
2563 rgb32to16 = RENAME(rgb32to16);
2564 rgb32to15 = RENAME(rgb32to15);
2565 rgb32tobgr24 = RENAME(rgb32tobgr24);
2566 rgb24to15 = RENAME(rgb24to15);
2567 rgb24to16 = RENAME(rgb24to16);
2568 rgb24tobgr24 = RENAME(rgb24tobgr24);
2569 shuffle_bytes_2103 = RENAME(shuffle_bytes_2103);
2570 rgb32tobgr16 = RENAME(rgb32tobgr16);
2571 rgb32tobgr15 = RENAME(rgb32tobgr15);
2572 yv12toyuy2 = RENAME(yv12toyuy2);
2573 yv12touyvy = RENAME(yv12touyvy);
2574 yuv422ptoyuy2 = RENAME(yuv422ptoyuy2);
2575 yuv422ptouyvy = RENAME(yuv422ptouyvy);
2576 yuy2toyv12 = RENAME(yuy2toyv12);
2577 vu9_to_vu12 = RENAME(vu9_to_vu12);
2578 yvu9_to_yuy2 = RENAME(yvu9_to_yuy2);
2579 uyvytoyuv422 = RENAME(uyvytoyuv422);
2580 yuyvtoyuv422 = RENAME(yuyvtoyuv422);
2581 #endif /* !COMPILE_TEMPLATE_SSE2 */
2583 #if COMPILE_TEMPLATE_MMX2 || COMPILE_TEMPLATE_AMD3DNOW
2584 planar2x = RENAME(planar2x);
2585 #endif /* COMPILE_TEMPLATE_MMX2 || COMPILE_TEMPLATE_AMD3DNOW */
2586 rgb24toyv12 = RENAME(rgb24toyv12);
2588 yuyvtoyuv420 = RENAME(yuyvtoyuv420);
2589 uyvytoyuv420 = RENAME(uyvytoyuv420);
2590 #endif /* COMPILE_TEMPLATE_SSE2 */
2592 #if !COMPILE_TEMPLATE_AMD3DNOW
2593 interleaveBytes = RENAME(interleaveBytes);
2594 #endif /* !COMPILE_TEMPLATE_AMD3DNOW */