diff options
| author | Ronald S. Bultje <rsbultje@gmail.com> | 2012-01-08 16:37:43 -0800 |
|---|---|---|
| committer | Ronald S. Bultje <rsbultje@gmail.com> | 2012-01-08 16:37:43 -0800 |
| commit | 2170a0e6add6bb0a6fbdf689b82361c21d9b72be (patch) | |
| tree | 8eb78b7d78b23e0e8b9d9b04fc5c07f954f0a000 /libswscale/x86/swscale_mmx.c | |
| parent | 599855748a93dc8bdaf5a5b3ac681e9a7026bc04 (diff) | |
| download | ffmpeg-2170a0e6add6bb0a6fbdf689b82361c21d9b72be.tar.gz | |
swscale: convert yuy2/uyvy/nv12/nv21ToY/UV from inline asm to yasm.
Also implement SSE2/AVX variants.
Diffstat (limited to 'libswscale/x86/swscale_mmx.c')
| -rw-r--r-- | libswscale/x86/swscale_mmx.c | 83 |
1 files changed, 83 insertions, 0 deletions
diff --git a/libswscale/x86/swscale_mmx.c b/libswscale/x86/swscale_mmx.c index 4305cef41d..867a9f1244 100644 --- a/libswscale/x86/swscale_mmx.c +++ b/libswscale/x86/swscale_mmx.c @@ -244,6 +244,26 @@ VSCALE_FUNCS(sse2, sse2); VSCALE_FUNC(16, sse4); VSCALE_FUNCS(avx, avx); +#define INPUT_UV_FUNC(fmt, opt) \ +extern void ff_ ## fmt ## ToUV_ ## opt(uint8_t *dstU, uint8_t *dstV, \ + const uint8_t *src, const uint8_t *unused1, \ + int w, uint32_t *unused2) +#define INPUT_FUNC(fmt, opt) \ +extern void ff_ ## fmt ## ToY_ ## opt(uint8_t *dst, const uint8_t *src, \ + int w, uint32_t *unused); \ + INPUT_UV_FUNC(fmt, opt) +#define INPUT_FUNCS(opt) \ + INPUT_FUNC(uyvy, opt); \ + INPUT_FUNC(yuyv, opt); \ + INPUT_UV_FUNC(nv12, opt); \ + INPUT_UV_FUNC(nv21, opt) + +#if ARCH_X86_32 +INPUT_FUNCS(mmx); +#endif +INPUT_FUNCS(sse2); +INPUT_FUNCS(avx); + void ff_sws_init_swScale_mmx(SwsContext *c) { int cpu_flags = av_get_cpu_flags(); @@ -296,6 +316,30 @@ switch(c->dstBpc){ \ ASSIGN_MMX_SCALE_FUNC(c->hyScale, c->hLumFilterSize, mmx, mmx); ASSIGN_MMX_SCALE_FUNC(c->hcScale, c->hChrFilterSize, mmx, mmx); ASSIGN_VSCALE_FUNC(c->yuv2plane1, mmx, mmx2, cpu_flags & AV_CPU_FLAG_MMX2); + + switch (c->srcFormat) { + case PIX_FMT_Y400A: + c->lumToYV12 = ff_yuyvToY_mmx; + if (c->alpPixBuf) + c->alpToYV12 = ff_uyvyToY_mmx; + break; + case PIX_FMT_YUYV422: + c->lumToYV12 = ff_yuyvToY_mmx; + c->chrToYV12 = ff_yuyvToUV_mmx; + break; + case PIX_FMT_UYVY422: + c->lumToYV12 = ff_uyvyToY_mmx; + c->chrToYV12 = ff_uyvyToUV_mmx; + break; + case PIX_FMT_NV12: + c->chrToYV12 = ff_nv12ToUV_mmx; + break; + case PIX_FMT_NV21: + c->chrToYV12 = ff_nv21ToUV_mmx; + break; + default: + break; + } } if (cpu_flags & AV_CPU_FLAG_MMX2) { ASSIGN_VSCALEX_FUNC(c->yuv2planeX, mmx2,); @@ -314,6 +358,28 @@ switch(c->dstBpc){ \ ASSIGN_SSE_SCALE_FUNC(c->hcScale, c->hChrFilterSize, sse2, sse2); ASSIGN_VSCALEX_FUNC(c->yuv2planeX, sse2,); ASSIGN_VSCALE_FUNC(c->yuv2plane1, sse2, sse2, 1); + + switch (c->srcFormat) { + case PIX_FMT_Y400A: + c->lumToYV12 = ff_yuyvToY_sse2; + if (c->alpPixBuf) + c->alpToYV12 = ff_uyvyToY_sse2; + break; + case PIX_FMT_YUYV422: + c->lumToYV12 = ff_yuyvToY_sse2; + c->chrToYV12 = ff_yuyvToUV_sse2; + break; + case PIX_FMT_UYVY422: + c->lumToYV12 = ff_uyvyToY_sse2; + c->chrToYV12 = ff_uyvyToUV_sse2; + break; + case PIX_FMT_NV12: + c->chrToYV12 = ff_nv12ToUV_sse2; + break; + case PIX_FMT_NV21: + c->chrToYV12 = ff_nv21ToUV_sse2; + break; + } } if (cpu_flags & AV_CPU_FLAG_SSSE3) { ASSIGN_SSE_SCALE_FUNC(c->hyScale, c->hLumFilterSize, ssse3, ssse3); @@ -332,6 +398,23 @@ switch(c->dstBpc){ \ if (cpu_flags & AV_CPU_FLAG_AVX) { ASSIGN_VSCALEX_FUNC(c->yuv2planeX, avx,); ASSIGN_VSCALE_FUNC(c->yuv2plane1, avx, avx, 1); + + switch (c->srcFormat) { + case PIX_FMT_YUYV422: + c->chrToYV12 = ff_yuyvToUV_avx; + break; + case PIX_FMT_UYVY422: + c->chrToYV12 = ff_uyvyToUV_avx; + break; + case PIX_FMT_NV12: + c->chrToYV12 = ff_nv12ToUV_avx; + break; + case PIX_FMT_NV21: + c->chrToYV12 = ff_nv21ToUV_avx; + break; + default: + break; + } } #endif } |