diff options
| author | Ronald S. Bultje <rsbultje@gmail.com> | 2011-11-05 17:31:40 -0700 |
|---|---|---|
| committer | Ronald S. Bultje <rsbultje@gmail.com> | 2011-11-05 20:48:14 -0700 |
| commit | c435653627529e22d74214c2266f571255e404d6 (patch) | |
| tree | b18c4b74a98a28d7b887f2d0aaee401e168e0fe9 /libswscale/x86/swscale_mmx.c | |
| parent | 2f7f2e4b41ec682f5145175f16ba9e97c156d5fe (diff) | |
| download | ffmpeg-c435653627529e22d74214c2266f571255e404d6.tar.gz | |
swscale: write yuv2plane1 MMX/SSE2/SSE4/AVX functions.
Diffstat (limited to 'libswscale/x86/swscale_mmx.c')
| -rw-r--r-- | libswscale/x86/swscale_mmx.c | 28 |
1 files changed, 28 insertions, 0 deletions
diff --git a/libswscale/x86/swscale_mmx.c b/libswscale/x86/swscale_mmx.c index a923e47538..b3f1deae60 100644 --- a/libswscale/x86/swscale_mmx.c +++ b/libswscale/x86/swscale_mmx.c @@ -228,6 +228,22 @@ VSCALEX_FUNCS(sse4, sse4); VSCALEX_FUNC(16, sse4); VSCALEX_FUNCS(avx, avx); +#define VSCALE_FUNC(size, opt) \ +extern void ff_yuv2plane1_ ## size ## _ ## opt(const int16_t *src, uint8_t *dst, int dstW, \ + const uint8_t *dither, int offset) +#define VSCALE_FUNCS(opt1, opt2) \ + VSCALE_FUNC(8, opt1); \ + VSCALE_FUNC(9, opt2); \ + VSCALE_FUNC(10, opt2); \ + VSCALE_FUNC(16, opt1) + +#if ARCH_X86_32 +VSCALE_FUNCS(mmx, mmx2); +#endif +VSCALE_FUNCS(sse2, sse2); +VSCALE_FUNC(16, sse4); +VSCALE_FUNCS(avx, avx); + void ff_sws_init_swScale_mmx(SwsContext *c) { int cpu_flags = av_get_cpu_flags(); @@ -268,11 +284,19 @@ switch(c->dstBpc){ \ case 9: if (!isBE(c->dstFormat) && opt2chk) vscalefn = ff_yuv2planeX_9_ ## opt2; break; \ default: vscalefn = ff_yuv2planeX_8_ ## opt1; break; \ } +#define ASSIGN_VSCALE_FUNC(vscalefn, opt1, opt2, opt2chk) \ + switch(c->dstBpc){ \ + case 16: if (!isBE(c->dstFormat)) vscalefn = ff_yuv2plane1_16_ ## opt1; break; \ + case 10: if (!isBE(c->dstFormat) && opt2chk) vscalefn = ff_yuv2plane1_10_ ## opt2; break; \ + case 9: if (!isBE(c->dstFormat) && opt2chk) vscalefn = ff_yuv2plane1_9_ ## opt2; break; \ + default: vscalefn = ff_yuv2plane1_8_ ## opt1; break; \ + } #if ARCH_X86_32 if (cpu_flags & AV_CPU_FLAG_MMX) { ASSIGN_MMX_SCALE_FUNC(c->hyScale, c->hLumFilterSize, mmx, mmx); ASSIGN_MMX_SCALE_FUNC(c->hcScale, c->hChrFilterSize, mmx, mmx); ASSIGN_VSCALEX_FUNC(c->yuv2planeX, mmx, mmx2, cpu_flags & AV_CPU_FLAG_MMX2,); + ASSIGN_VSCALE_FUNC(c->yuv2plane1, mmx, mmx2, cpu_flags & AV_CPU_FLAG_MMX2); } #endif #define ASSIGN_SSE_SCALE_FUNC(hscalefn, filtersize, opt1, opt2) \ @@ -287,6 +311,7 @@ switch(c->dstBpc){ \ ASSIGN_SSE_SCALE_FUNC(c->hyScale, c->hLumFilterSize, sse2, sse2); ASSIGN_SSE_SCALE_FUNC(c->hcScale, c->hChrFilterSize, sse2, sse2); ASSIGN_VSCALEX_FUNC(c->yuv2planeX, sse2, sse2, 1,); + ASSIGN_VSCALE_FUNC(c->yuv2plane1, sse2, sse2, 1); } if (cpu_flags & AV_CPU_FLAG_SSSE3) { ASSIGN_SSE_SCALE_FUNC(c->hyScale, c->hLumFilterSize, ssse3, ssse3); @@ -298,10 +323,13 @@ switch(c->dstBpc){ \ ASSIGN_SSE_SCALE_FUNC(c->hcScale, c->hChrFilterSize, sse4, ssse3); ASSIGN_VSCALEX_FUNC(c->yuv2planeX, sse4, sse4, 1, if (!isBE(c->dstFormat)) c->yuv2planeX = ff_yuv2planeX_16_sse4); + if (c->dstBpc == 16 && !isBE(c->dstFormat)) + c->yuv2plane1 = ff_yuv2plane1_16_sse4; } if (cpu_flags & AV_CPU_FLAG_AVX) { ASSIGN_VSCALEX_FUNC(c->yuv2planeX, avx, avx, 1,); + ASSIGN_VSCALE_FUNC(c->yuv2plane1, avx, avx, 1); } #endif } |