Annotation of /trunk/xine-lib/patches/xine-lib-1.1.0-pic-fb-asm.patch
Parent Directory
| 
Revision Log
Revision 153 -
(hide annotations)
(download)
Tue May 8 20:52:56 2007 UTC (17 years, 1 month ago) by niro
File size: 2106 byte(s)
Tue May 8 20:52:56 2007 UTC (17 years, 1 month ago) by niro
File size: 2106 byte(s)
-import
| 1 | niro | 153 | diff -ur xine-lib-1.1.0-orig/src/libffmpeg/libavcodec/i386/dsputil_mmx.c xine-lib-1.1.0/src/libffmpeg/libavcodec/i386/dsputil_mmx.c |
| 2 | --- xine-lib-1.1.0-orig/src/libffmpeg/libavcodec/i386/dsputil_mmx.c 2005-09-07 16:55:46.000000000 -0400 | ||
| 3 | +++ xine-lib-1.1.0/src/libffmpeg/libavcodec/i386/dsputil_mmx.c 2005-09-07 16:56:31.000000000 -0400@@ -615,31 +615,32 @@ | ||
| 4 | @@ -615,31 +615,32 @@ | ||
| 5 | } | ||
| 6 | |||
| 7 | static inline void transpose4x4(uint8_t *dst, uint8_t *src, int dst_stride, int src_stride){ | ||
| 8 | + void *dst_reg = dst, *src_reg = src; | ||
| 9 | + | ||
| 10 | asm volatile( //FIXME could save 1 instruction if done as 8x4 ... | ||
| 11 | - "movd %4, %%mm0 \n\t" | ||
| 12 | - "movd %5, %%mm1 \n\t" | ||
| 13 | - "movd %6, %%mm2 \n\t" | ||
| 14 | - "movd %7, %%mm3 \n\t" | ||
| 15 | + "movd (%1), %%mm0 \n\t" | ||
| 16 | + "movd (%1,%5), %%mm1 \n\t" | ||
| 17 | + "lea (%1, %5, 2), %1 \n\t" | ||
| 18 | + "movd (%1), %%mm2 \n\t" | ||
| 19 | + "movd (%1,%5), %%mm3 \n\t" | ||
| 20 | "punpcklbw %%mm1, %%mm0 \n\t" | ||
| 21 | "punpcklbw %%mm3, %%mm2 \n\t" | ||
| 22 | "movq %%mm0, %%mm1 \n\t" | ||
| 23 | "punpcklwd %%mm2, %%mm0 \n\t" | ||
| 24 | "punpckhwd %%mm2, %%mm1 \n\t" | ||
| 25 | - "movd %%mm0, %0 \n\t" | ||
| 26 | + "movd %%mm0, (%0) \n\t" | ||
| 27 | "punpckhdq %%mm0, %%mm0 \n\t" | ||
| 28 | - "movd %%mm0, %1 \n\t" | ||
| 29 | - "movd %%mm1, %2 \n\t" | ||
| 30 | + "movd %%mm0, (%0,%4) \n\t" | ||
| 31 | + "lea (%0, %4, 2), %0 \n\t" | ||
| 32 | + "movd %%mm1, (%0) \n\t" | ||
| 33 | "punpckhdq %%mm1, %%mm1 \n\t" | ||
| 34 | - "movd %%mm1, %3 \n\t" | ||
| 35 | - | ||
| 36 | - : "=m" (*(uint32_t*)(dst + 0*dst_stride)), | ||
| 37 | - "=m" (*(uint32_t*)(dst + 1*dst_stride)), | ||
| 38 | - "=m" (*(uint32_t*)(dst + 2*dst_stride)), | ||
| 39 | - "=m" (*(uint32_t*)(dst + 3*dst_stride)) | ||
| 40 | - : "m" (*(uint32_t*)(src + 0*src_stride)), | ||
| 41 | - "m" (*(uint32_t*)(src + 1*src_stride)), | ||
| 42 | - "m" (*(uint32_t*)(src + 2*src_stride)), | ||
| 43 | - "m" (*(uint32_t*)(src + 3*src_stride)) | ||
| 44 | + "movd %%mm1, (%0,%4) \n\t" | ||
| 45 | + : "=&r" (dst_reg), | ||
| 46 | + "=&r" (src_reg) | ||
| 47 | + : "0" (dst_reg), | ||
| 48 | + "1" (src_reg), | ||
| 49 | + "r" (dst_stride), | ||
| 50 | + "r" (src_stride) | ||
| 51 | ); | ||
| 52 | } | ||
| 53 |