From ac4b32df71bd932838043a4838b86d11e169707f Mon Sep 17 00:00:00 2001 From: Peter Ross Date: Tue, 1 Apr 2014 14:21:51 +0200 Subject: On2 VP7 decoder Further performance improvements and security fixes by Vittorio Giovara, Luca Barbato and Diego Biurrun. Signed-off-by: Vittorio Giovara Signed-off-by: Luca Barbato Signed-off-by: Diego Biurrun --- libavcodec/x86/vp8dsp_init.c | 77 ++++++++++++++++++++++++++++---------------- 1 file changed, 50 insertions(+), 27 deletions(-) (limited to 'libavcodec/x86/vp8dsp_init.c') diff --git a/libavcodec/x86/vp8dsp_init.c b/libavcodec/x86/vp8dsp_init.c index 69460aa73f..b3b32b7314 100644 --- a/libavcodec/x86/vp8dsp_init.c +++ b/libavcodec/x86/vp8dsp_init.c @@ -315,25 +315,69 @@ DECLARE_LOOP_FILTER(sse4) c->put_vp8_bilinear_pixels_tab[IDX][2][2] = ff_put_vp8_bilinear ## SIZE ## _hv_ ## OPT -av_cold void ff_vp8dsp_init_x86(VP8DSPContext* c) +av_cold void ff_vp78dsp_init_x86(VP8DSPContext *c) { #if HAVE_YASM int cpu_flags = av_get_cpu_flags(); if (EXTERNAL_MMX(cpu_flags)) { - c->vp8_idct_dc_add = ff_vp8_idct_dc_add_mmx; - c->vp8_idct_dc_add4uv = ff_vp8_idct_dc_add4uv_mmx; #if ARCH_X86_32 - c->vp8_idct_dc_add4y = ff_vp8_idct_dc_add4y_mmx; - c->vp8_idct_add = ff_vp8_idct_add_mmx; - c->vp8_luma_dc_wht = ff_vp8_luma_dc_wht_mmx; c->put_vp8_epel_pixels_tab[0][0][0] = c->put_vp8_bilinear_pixels_tab[0][0][0] = ff_put_vp8_pixels16_mmx; #endif c->put_vp8_epel_pixels_tab[1][0][0] = c->put_vp8_bilinear_pixels_tab[1][0][0] = ff_put_vp8_pixels8_mmx; + } + + /* note that 4-tap width=16 functions are missing because w=16 + * is only used for luma, and luma is always a copy or sixtap. */ + if (EXTERNAL_MMXEXT(cpu_flags)) { + VP8_MC_FUNC(2, 4, mmxext); + VP8_BILINEAR_MC_FUNC(2, 4, mmxext); +#if ARCH_X86_32 + VP8_LUMA_MC_FUNC(0, 16, mmxext); + VP8_MC_FUNC(1, 8, mmxext); + VP8_BILINEAR_MC_FUNC(0, 16, mmxext); + VP8_BILINEAR_MC_FUNC(1, 8, mmxext); +#endif + } + + if (EXTERNAL_SSE(cpu_flags)) { + c->put_vp8_epel_pixels_tab[0][0][0] = + c->put_vp8_bilinear_pixels_tab[0][0][0] = ff_put_vp8_pixels16_sse; + } + + if (EXTERNAL_SSE2(cpu_flags) && (cpu_flags & AV_CPU_FLAG_SSE2SLOW)) { + VP8_LUMA_MC_FUNC(0, 16, sse2); + VP8_MC_FUNC(1, 8, sse2); + VP8_BILINEAR_MC_FUNC(0, 16, sse2); + VP8_BILINEAR_MC_FUNC(1, 8, sse2); + } + + if (EXTERNAL_SSSE3(cpu_flags)) { + VP8_LUMA_MC_FUNC(0, 16, ssse3); + VP8_MC_FUNC(1, 8, ssse3); + VP8_MC_FUNC(2, 4, ssse3); + VP8_BILINEAR_MC_FUNC(0, 16, ssse3); + VP8_BILINEAR_MC_FUNC(1, 8, ssse3); + VP8_BILINEAR_MC_FUNC(2, 4, ssse3); + } +#endif /* HAVE_YASM */ +} +av_cold void ff_vp8dsp_init_x86(VP8DSPContext *c) +{ +#if HAVE_YASM + int cpu_flags = av_get_cpu_flags(); + + if (EXTERNAL_MMX(cpu_flags)) { + c->vp8_idct_dc_add = ff_vp8_idct_dc_add_mmx; + c->vp8_idct_dc_add4uv = ff_vp8_idct_dc_add4uv_mmx; #if ARCH_X86_32 + c->vp8_idct_dc_add4y = ff_vp8_idct_dc_add4y_mmx; + c->vp8_idct_add = ff_vp8_idct_add_mmx; + c->vp8_luma_dc_wht = ff_vp8_luma_dc_wht_mmx; + c->vp8_v_loop_filter_simple = ff_vp8_v_loop_filter_simple_mmx; c->vp8_h_loop_filter_simple = ff_vp8_h_loop_filter_simple_mmx; @@ -352,14 +396,7 @@ av_cold void ff_vp8dsp_init_x86(VP8DSPContext* c) /* note that 4-tap width=16 functions are missing because w=16 * is only used for luma, and luma is always a copy or sixtap. */ if (EXTERNAL_MMXEXT(cpu_flags)) { - VP8_MC_FUNC(2, 4, mmxext); - VP8_BILINEAR_MC_FUNC(2, 4, mmxext); #if ARCH_X86_32 - VP8_LUMA_MC_FUNC(0, 16, mmxext); - VP8_MC_FUNC(1, 8, mmxext); - VP8_BILINEAR_MC_FUNC(0, 16, mmxext); - VP8_BILINEAR_MC_FUNC(1, 8, mmxext); - c->vp8_v_loop_filter_simple = ff_vp8_v_loop_filter_simple_mmxext; c->vp8_h_loop_filter_simple = ff_vp8_h_loop_filter_simple_mmxext; @@ -378,16 +415,9 @@ av_cold void ff_vp8dsp_init_x86(VP8DSPContext* c) if (EXTERNAL_SSE(cpu_flags)) { c->vp8_idct_add = ff_vp8_idct_add_sse; c->vp8_luma_dc_wht = ff_vp8_luma_dc_wht_sse; - c->put_vp8_epel_pixels_tab[0][0][0] = - c->put_vp8_bilinear_pixels_tab[0][0][0] = ff_put_vp8_pixels16_sse; } if (EXTERNAL_SSE2(cpu_flags) && (cpu_flags & AV_CPU_FLAG_SSE2SLOW)) { - VP8_LUMA_MC_FUNC(0, 16, sse2); - VP8_MC_FUNC(1, 8, sse2); - VP8_BILINEAR_MC_FUNC(0, 16, sse2); - VP8_BILINEAR_MC_FUNC(1, 8, sse2); - c->vp8_v_loop_filter_simple = ff_vp8_v_loop_filter_simple_sse2; c->vp8_v_loop_filter16y_inner = ff_vp8_v_loop_filter16y_inner_sse2; @@ -410,13 +440,6 @@ av_cold void ff_vp8dsp_init_x86(VP8DSPContext* c) } if (EXTERNAL_SSSE3(cpu_flags)) { - VP8_LUMA_MC_FUNC(0, 16, ssse3); - VP8_MC_FUNC(1, 8, ssse3); - VP8_MC_FUNC(2, 4, ssse3); - VP8_BILINEAR_MC_FUNC(0, 16, ssse3); - VP8_BILINEAR_MC_FUNC(1, 8, ssse3); - VP8_BILINEAR_MC_FUNC(2, 4, ssse3); - c->vp8_v_loop_filter_simple = ff_vp8_v_loop_filter_simple_ssse3; c->vp8_h_loop_filter_simple = ff_vp8_h_loop_filter_simple_ssse3; -- cgit v1.2.3