diff options
| author | Michael Niedermayer <michaelni@gmx.at> | 2013-01-16 12:38:41 +0100 |
|---|---|---|
| committer | Michael Niedermayer <michaelni@gmx.at> | 2013-01-16 12:38:41 +0100 |
| commit | 5c7e9e16c961f1f7258734426afac3cee4349580 (patch) | |
| tree | 20ea7bec723ef262bffb8f2447adcc1dc3d98069 /libavutil/x86/float_dsp_init.c | |
| parent | 0e79fe37e5c5500db2e65ce6b7ea0bbdb3f24665 (diff) | |
| parent | e034cc6c60c77dce390b1ac31141b1862bdf8999 (diff) | |
| download | ffmpeg-streaming-5c7e9e16c961f1f7258734426afac3cee4349580.zip ffmpeg-streaming-5c7e9e16c961f1f7258734426afac3cee4349580.tar.gz | |
Merge remote-tracking branch 'qatar/master'
* qatar/master:
lavc: Move vector_fmul_window to AVFloatDSPContext
rtpdec_mpeg4: Check the remaining amount of data before reading
Conflicts:
libavcodec/dsputil.c
Merged-by: Michael Niedermayer <michaelni@gmx.at>
Diffstat (limited to 'libavutil/x86/float_dsp_init.c')
| -rw-r--r-- | libavutil/x86/float_dsp_init.c | 75 |
1 files changed, 75 insertions, 0 deletions
diff --git a/libavutil/x86/float_dsp_init.c b/libavutil/x86/float_dsp_init.c index e5112e7..93cff9f 100644 --- a/libavutil/x86/float_dsp_init.c +++ b/libavutil/x86/float_dsp_init.c @@ -21,6 +21,7 @@ #include "libavutil/cpu.h" #include "libavutil/float_dsp.h" #include "cpu.h" +#include "asm.h" extern void ff_vector_fmul_sse(float *dst, const float *src0, const float *src1, int len); @@ -40,10 +41,84 @@ extern void ff_vector_dmul_scalar_sse2(double *dst, const double *src, extern void ff_vector_dmul_scalar_avx(double *dst, const double *src, double mul, int len); +#if HAVE_6REGS +static void vector_fmul_window_3dnowext(float *dst, const float *src0, + const float *src1, const float *win, + int len) +{ + x86_reg i = -len * 4; + x86_reg j = len * 4 - 8; + __asm__ volatile ( + "1: \n" + "pswapd (%5, %1), %%mm1 \n" + "movq (%5, %0), %%mm0 \n" + "pswapd (%4, %1), %%mm5 \n" + "movq (%3, %0), %%mm4 \n" + "movq %%mm0, %%mm2 \n" + "movq %%mm1, %%mm3 \n" + "pfmul %%mm4, %%mm2 \n" // src0[len + i] * win[len + i] + "pfmul %%mm5, %%mm3 \n" // src1[j] * win[len + j] + "pfmul %%mm4, %%mm1 \n" // src0[len + i] * win[len + j] + "pfmul %%mm5, %%mm0 \n" // src1[j] * win[len + i] + "pfadd %%mm3, %%mm2 \n" + "pfsub %%mm0, %%mm1 \n" + "pswapd %%mm2, %%mm2 \n" + "movq %%mm1, (%2, %0) \n" + "movq %%mm2, (%2, %1) \n" + "sub $8, %1 \n" + "add $8, %0 \n" + "jl 1b \n" + "femms \n" + : "+r"(i), "+r"(j) + : "r"(dst + len), "r"(src0 + len), "r"(src1), "r"(win + len) + ); +} + +static void vector_fmul_window_sse(float *dst, const float *src0, + const float *src1, const float *win, int len) +{ + x86_reg i = -len * 4; + x86_reg j = len * 4 - 16; + __asm__ volatile ( + "1: \n" + "movaps (%5, %1), %%xmm1 \n" + "movaps (%5, %0), %%xmm0 \n" + "movaps (%4, %1), %%xmm5 \n" + "movaps (%3, %0), %%xmm4 \n" + "shufps $0x1b, %%xmm1, %%xmm1 \n" + "shufps $0x1b, %%xmm5, %%xmm5 \n" + "movaps %%xmm0, %%xmm2 \n" + "movaps %%xmm1, %%xmm3 \n" + "mulps %%xmm4, %%xmm2 \n" // src0[len + i] * win[len + i] + "mulps %%xmm5, %%xmm3 \n" // src1[j] * win[len + j] + "mulps %%xmm4, %%xmm1 \n" // src0[len + i] * win[len + j] + "mulps %%xmm5, %%xmm0 \n" // src1[j] * win[len + i] + "addps %%xmm3, %%xmm2 \n" + "subps %%xmm0, %%xmm1 \n" + "shufps $0x1b, %%xmm2, %%xmm2 \n" + "movaps %%xmm1, (%2, %0) \n" + "movaps %%xmm2, (%2, %1) \n" + "sub $16, %1 \n" + "add $16, %0 \n" + "jl 1b \n" + : "+r"(i), "+r"(j) + : "r"(dst + len), "r"(src0 + len), "r"(src1), "r"(win + len) + ); +} +#endif /* HAVE_6REGS */ + void ff_float_dsp_init_x86(AVFloatDSPContext *fdsp) { int mm_flags = av_get_cpu_flags(); +#if HAVE_6REGS + if (INLINE_AMD3DNOWEXT(mm_flags)) { + fdsp->vector_fmul_window = vector_fmul_window_3dnowext; + } + if (INLINE_SSE(mm_flags)) { + fdsp->vector_fmul_window = vector_fmul_window_sse; + } +#endif if (EXTERNAL_SSE(mm_flags)) { fdsp->vector_fmul = ff_vector_fmul_sse; fdsp->vector_fmac_scalar = ff_vector_fmac_scalar_sse; |
