Mercurial > mplayer.hg

comparison libswscale/rgb2rgb_template.c @ 22991:59671a52cc82

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
New implementation of rgb32tobgr32 The previous implementation segfaulted with MMX enabled when fed an image smaller than the size of the units the MMX code processed. The new code: - is faster for MMX, MMX2 and plain C - processes small images correctly - is LGPL
author ivo
date Mon, 16 Apr 2007 21:41:03 +0000
parents 2d1ad4285df4
children ac77d9ef8c83
comparison
equal deleted inserted replaced
22990:a76748e71a18 22991:59671a52cc82
1362 } 1362 }
1363 } 1363 }
1364 1364
1365 static inline void RENAME(rgb32tobgr32)(const uint8_t *src, uint8_t *dst, long src_size) 1365 static inline void RENAME(rgb32tobgr32)(const uint8_t *src, uint8_t *dst, long src_size)
1366 { 1366 {
1367 #ifdef HAVE_MMX 1367 uint8_t *d = dst, *s = (uint8_t *) src;
1368 /* TODO: unroll this loop */ 1368 const uint8_t *end = s + src_size;
1369 asm volatile ( 1369 #ifdef HAVE_MMX
1370 "xor %%"REG_a", %%"REG_a" \n\t" 1370 __asm __volatile(
1371 ASMALIGN(4) 1371 " "PREFETCH" (%1) \n"
1372 "1: \n\t" 1372 " movq %3, %%mm7 \n"
1373 PREFETCH" 32(%0, %%"REG_a") \n\t" 1373 " pxor %4, %%mm7 \n"
1374 "movq (%0, %%"REG_a"), %%mm0 \n\t" 1374 " movq %%mm7, %%mm6 \n"
1375 "movq %%mm0, %%mm1 \n\t" 1375 " pxor %5, %%mm7 \n"
1376 "movq %%mm0, %%mm2 \n\t" 1376 " jmp 2f \n"
1377 "pslld $16, %%mm0 \n\t" 1377 ASMALIGN(4)
1378 "psrld $16, %%mm1 \n\t" 1378 "1: \n"
1379 "pand "MANGLE(mask32r)", %%mm0 \n\t" 1379 " "PREFETCH" 32(%1) \n"
1380 "pand "MANGLE(mask32g)", %%mm2 \n\t" 1380 " movq (%1), %%mm0 \n"
1381 "pand "MANGLE(mask32b)", %%mm1 \n\t" 1381 " movq 8(%1), %%mm1 \n"
1382 "por %%mm0, %%mm2 \n\t" 1382 # ifdef HAVE_MMX2
1383 "por %%mm1, %%mm2 \n\t" 1383 " pshufw $177, %%mm0, %%mm3 \n"
1384 MOVNTQ" %%mm2, (%1, %%"REG_a") \n\t" 1384 " pshufw $177, %%mm1, %%mm5 \n"
1385 "add $8, %%"REG_a" \n\t" 1385 " pand %%mm7, %%mm0 \n"
1386 "cmp %2, %%"REG_a" \n\t" 1386 " pand %%mm6, %%mm3 \n"
1387 " jb 1b \n\t" 1387 " pand %%mm7, %%mm1 \n"
1388 :: "r" (src), "r"(dst), "r" (src_size-7) 1388 " pand %%mm6, %%mm5 \n"
1389 : "%"REG_a 1389 " por %%mm3, %%mm0 \n"
1390 ); 1390 " por %%mm5, %%mm1 \n"
1391 1391 # else
1392 __asm __volatile(SFENCE:::"memory"); 1392 " movq %%mm0, %%mm2 \n"
1393 __asm __volatile(EMMS:::"memory"); 1393 " movq %%mm1, %%mm4 \n"
1394 #else 1394 " pand %%mm7, %%mm0 \n"
1395 unsigned i; 1395 " pand %%mm6, %%mm2 \n"
1396 unsigned num_pixels = src_size >> 2; 1396 " pand %%mm7, %%mm1 \n"
1397 for(i=0; i<num_pixels; i++) 1397 " pand %%mm6, %%mm4 \n"
1398 { 1398 " movq %%mm2, %%mm3 \n"
1399 #ifdef WORDS_BIGENDIAN 1399 " movq %%mm4, %%mm5 \n"
1400 dst[4*i + 1] = src[4*i + 3]; 1400 " pslld $16, %%mm2 \n"
1401 dst[4*i + 2] = src[4*i + 2]; 1401 " psrld $16, %%mm3 \n"
1402 dst[4*i + 3] = src[4*i + 1]; 1402 " pslld $16, %%mm4 \n"
1403 #else 1403 " psrld $16, %%mm5 \n"
1404 dst[4*i + 0] = src[4*i + 2]; 1404 " por %%mm2, %%mm0 \n"
1405 dst[4*i + 1] = src[4*i + 1]; 1405 " por %%mm4, %%mm1 \n"
1406 dst[4*i + 2] = src[4*i + 0]; 1406 " por %%mm3, %%mm0 \n"
1407 #endif 1407 " por %%mm5, %%mm1 \n"
1408 } 1408 # endif
1409 #endif 1409 " "MOVNTQ" %%mm0, (%0) \n"
1410 " "MOVNTQ" %%mm1, 8(%0) \n"
1411 " add $16, %0 \n"
1412 " add $16, %1 \n"
1413 "2: \n"
1414 " cmp %1, %2 \n"
1415 " ja 1b \n"
1416 " "SFENCE" \n"
1417 " "EMMS" \n"
1418 : "+r"(d), "+r"(s)
1419 : "r" (end-15), "m" (mask32b), "m" (mask32r), "m" (mmx_one)
1420 : "memory");
1421 #endif
1422 for (; s<end; s+=4, d+=4) {
1423 int v = *(uint32_t *)s, g = v & 0xff00;
1424 v &= 0xff00ff;
1425 *(uint32_t *)d = (v>>16) + g + (v<<16);
1426 }
1410 } 1427 }
1411 1428
1412 static inline void RENAME(rgb24tobgr24)(const uint8_t *src, uint8_t *dst, long src_size) 1429 static inline void RENAME(rgb24tobgr24)(const uint8_t *src, uint8_t *dst, long src_size)
1413 { 1430 {
1414 unsigned i; 1431 unsigned i;