| @@ -386,8 +386,10 @@ cglobal deblock_h_luma_8_%1, 5,7 | |||||
| INIT_XMM | INIT_XMM | ||||
| DEBLOCK_LUMA sse2 | DEBLOCK_LUMA sse2 | ||||
| %ifdef HAVE_AVX | |||||
| INIT_AVX | INIT_AVX | ||||
| DEBLOCK_LUMA avx | DEBLOCK_LUMA avx | ||||
| %endif | |||||
| %else | %else | ||||
| @@ -506,8 +508,10 @@ INIT_MMX | |||||
| DEBLOCK_LUMA mmxext, v8, 8 | DEBLOCK_LUMA mmxext, v8, 8 | ||||
| INIT_XMM | INIT_XMM | ||||
| DEBLOCK_LUMA sse2, v, 16 | DEBLOCK_LUMA sse2, v, 16 | ||||
| %ifdef HAVE_AVX | |||||
| INIT_AVX | INIT_AVX | ||||
| DEBLOCK_LUMA avx, v, 16 | DEBLOCK_LUMA avx, v, 16 | ||||
| %endif | |||||
| %endif ; ARCH | %endif ; ARCH | ||||
| @@ -778,8 +782,10 @@ cglobal deblock_h_luma_intra_8_%1, 2,4 | |||||
| INIT_XMM | INIT_XMM | ||||
| DEBLOCK_LUMA_INTRA sse2, v | DEBLOCK_LUMA_INTRA sse2, v | ||||
| %ifdef HAVE_AVX | |||||
| INIT_AVX | INIT_AVX | ||||
| DEBLOCK_LUMA_INTRA avx , v | DEBLOCK_LUMA_INTRA avx , v | ||||
| %endif | |||||
| %ifndef ARCH_X86_64 | %ifndef ARCH_X86_64 | ||||
| INIT_MMX | INIT_MMX | ||||
| DEBLOCK_LUMA_INTRA mmxext, v8 | DEBLOCK_LUMA_INTRA mmxext, v8 | ||||
| @@ -419,9 +419,11 @@ cglobal deblock_h_luma_10_%1, 5,7,15 | |||||
| INIT_XMM | INIT_XMM | ||||
| DEBLOCK_LUMA_64 sse2 | DEBLOCK_LUMA_64 sse2 | ||||
| %ifdef HAVE_AVX | |||||
| INIT_AVX | INIT_AVX | ||||
| DEBLOCK_LUMA_64 avx | DEBLOCK_LUMA_64 avx | ||||
| %endif | %endif | ||||
| %endif | |||||
| %macro SWAPMOVA 2 | %macro SWAPMOVA 2 | ||||
| %ifid %1 | %ifid %1 | ||||
| @@ -714,8 +716,10 @@ cglobal deblock_h_luma_intra_10_%1, 4,7,16 | |||||
| INIT_XMM | INIT_XMM | ||||
| DEBLOCK_LUMA_INTRA_64 sse2 | DEBLOCK_LUMA_INTRA_64 sse2 | ||||
| %ifdef HAVE_AVX | |||||
| INIT_AVX | INIT_AVX | ||||
| DEBLOCK_LUMA_INTRA_64 avx | DEBLOCK_LUMA_INTRA_64 avx | ||||
| %endif | |||||
| %endif | %endif | ||||
| @@ -799,10 +803,12 @@ DEBLOCK_LUMA_INTRA mmxext | |||||
| INIT_XMM | INIT_XMM | ||||
| DEBLOCK_LUMA sse2 | DEBLOCK_LUMA sse2 | ||||
| DEBLOCK_LUMA_INTRA sse2 | DEBLOCK_LUMA_INTRA sse2 | ||||
| %ifdef HAVE_AVX | |||||
| INIT_AVX | INIT_AVX | ||||
| DEBLOCK_LUMA avx | DEBLOCK_LUMA avx | ||||
| DEBLOCK_LUMA_INTRA avx | DEBLOCK_LUMA_INTRA avx | ||||
| %endif | %endif | ||||
| %endif | |||||
| ; in: %1=p0, %2=q0, %3=p1, %4=q1, %5=mask, %6=tmp, %7=tmp | ; in: %1=p0, %2=q0, %3=p1, %4=q1, %5=mask, %6=tmp, %7=tmp | ||||
| ; out: %1=p0', %2=q0' | ; out: %1=p0', %2=q0' | ||||
| @@ -913,5 +919,7 @@ DEBLOCK_CHROMA mmxext | |||||
| %endif | %endif | ||||
| INIT_XMM | INIT_XMM | ||||
| DEBLOCK_CHROMA sse2 | DEBLOCK_CHROMA sse2 | ||||
| %ifdef HAVE_AVX | |||||
| INIT_AVX | INIT_AVX | ||||
| DEBLOCK_CHROMA avx | DEBLOCK_CHROMA avx | ||||
| %endif | |||||
| @@ -418,7 +418,7 @@ void ff_h264dsp_init_x86(H264DSPContext *c, const int bit_depth) | |||||
| c->biweight_h264_pixels_tab[3]= ff_h264_biweight_8x8_ssse3; | c->biweight_h264_pixels_tab[3]= ff_h264_biweight_8x8_ssse3; | ||||
| c->biweight_h264_pixels_tab[4]= ff_h264_biweight_8x4_ssse3; | c->biweight_h264_pixels_tab[4]= ff_h264_biweight_8x4_ssse3; | ||||
| } | } | ||||
| if (mm_flags&AV_CPU_FLAG_AVX) { | |||||
| if (HAVE_AVX && mm_flags&AV_CPU_FLAG_AVX) { | |||||
| #if HAVE_ALIGNED_STACK | #if HAVE_ALIGNED_STACK | ||||
| c->h264_v_loop_filter_luma = ff_deblock_v_luma_8_avx; | c->h264_v_loop_filter_luma = ff_deblock_v_luma_8_avx; | ||||
| c->h264_h_loop_filter_luma = ff_deblock_h_luma_8_avx; | c->h264_h_loop_filter_luma = ff_deblock_h_luma_8_avx; | ||||