From 533688786799b22d0711eedcfe8f84deea014f30 Mon Sep 17 00:00:00 2001 From: James Darnley Date: Fri, 10 Feb 2017 23:17:57 +0100 Subject: avcodec/h264: sse2, avx h luma mbaff deblock/loop filter x86-64 only Yorkfield: - sse2: ~2.17x (434 vs. 200 cycles) Nehalem: - sse2: ~2.94x (409 vs. 139 cycles) Skylake: - sse2: ~3.10x (370 vs. 119 cycles) - avx: ~3.29x (370 vs. 112 cycles) --- libavcodec/x86/h264_deblock.asm | 89 +++++++++++++++++++++++++++++++++++++++++ libavcodec/x86/h264dsp_init.c | 10 +++++ libavutil/x86/x86util.asm | 15 +++++++ 3 files changed, 114 insertions(+) diff --git a/libavcodec/x86/h264_deblock.asm b/libavcodec/x86/h264_deblock.asm index 509a0dbe0c..93caa67c85 100644 --- a/libavcodec/x86/h264_deblock.asm +++ b/libavcodec/x86/h264_deblock.asm @@ -377,10 +377,99 @@ cglobal deblock_h_luma_8, 5,9,0,0x60+16*WIN64 RET %endmacro +%macro DEBLOCK_H_LUMA_MBAFF 0 + +cglobal deblock_h_luma_mbaff_8, 5, 9, 10, 8*16, pix_, stride_, alpha_, beta_, tc0_, base3_, stride3_ + movsxd stride_q, stride_d + dec alpha_d + dec beta_d + mov base3_q, pix_q + lea stride3_q, [3*stride_q] + add base3_q, stride3_q + + movq m0, [pix_q - 4] + movq m1, [pix_q + stride_q - 4] + movq m2, [pix_q + 2*stride_q - 4] + movq m3, [base3_q - 4] + movq m4, [base3_q + stride_q - 4] + movq m5, [base3_q + 2*stride_q - 4] + movq m6, [base3_q + stride3_q - 4] + movq m7, [base3_q + 4*stride_q - 4] + + TRANSPOSE_8X8B 0,1,2,3,4,5,6,7 + + %assign i 0 + %rep 8 + movq [rsp + 16*i], m %+ i + %assign i i+1 + %endrep + + ; p2 = m1 [rsp + 16] + ; p1 = m2 [rsp + 32] + ; p0 = m3 [rsp + 48] + ; q0 = m4 [rsp + 64] + ; q1 = m5 [rsp + 80] + ; q2 = m6 [rsp + 96] + + SWAP 0, 2 + SWAP 1, 3 + SWAP 2, 4 + SWAP 3, 5 + + LOAD_MASK alpha_d, beta_d + movd m8, [tc0_q] + punpcklbw m8, m8 + pcmpeqb m9, m9 + pcmpeqb m9, m8 + pandn m9, m7 + pand m8, m9 + + movdqa m3, [rsp + 16] ; p2 + DIFF_GT2 m1, m3, m5, m6, m7 ; |p2-p0| > beta-1 + pand m6, m9 + psubb m7, m8, m6 + pand m6, m8 + LUMA_Q1 m0, m3, [rsp + 16], [rsp + 32], m6, m4 + + movdqa m4, [rsp + 96] ; q2 + DIFF_GT2 m2, m4, m5, m6, m3 ; |q2-q0| > beta-1 + pand m6, m9 + pand m8, m6 + psubb m7, m6 + mova m3, [rsp + 80] + LUMA_Q1 m3, m4, [rsp + 96], [rsp + 80], m8, m6 + + DEBLOCK_P0_Q0 + SWAP 1, 3 + SWAP 2, 4 + movq m0, [rsp] + movq m1, [rsp + 16] + movq m2, [rsp + 32] + movq m5, [rsp + 80] + movq m6, [rsp + 96] + movq m7, [rsp + 112] + + TRANSPOSE_8X8B 0,1,2,3,4,5,6,7 + movq [pix_q - 4], m0 + movq [pix_q + stride_q - 4], m1 + movq [pix_q + 2*stride_q - 4], m2 + movq [base3_q - 4], m3 + movq [base3_q + stride_q - 4], m4 + movq [base3_q + 2*stride_q - 4], m5 + movq [base3_q + stride3_q - 4], m6 + movq [base3_q + 4*stride_q - 4], m7 + +RET + +%endmacro + INIT_XMM sse2 +DEBLOCK_H_LUMA_MBAFF DEBLOCK_LUMA + %if HAVE_AVX_EXTERNAL INIT_XMM avx +DEBLOCK_H_LUMA_MBAFF DEBLOCK_LUMA %endif diff --git a/libavcodec/x86/h264dsp_init.c b/libavcodec/x86/h264dsp_init.c index 7b3d17f971..10f19401ef 100644 --- a/libavcodec/x86/h264dsp_init.c +++ b/libavcodec/x86/h264dsp_init.c @@ -137,6 +137,9 @@ LF_IFUNC(h, chroma422_intra, depth, avx) \ LF_FUNC(v, chroma, depth, avx) \ LF_IFUNC(v, chroma_intra, depth, avx) +LF_FUNC(h, luma_mbaff, 8, sse2) +LF_FUNC(h, luma_mbaff, 8, avx) + LF_FUNCS(uint8_t, 8) LF_FUNCS(uint16_t, 10) @@ -297,6 +300,10 @@ av_cold void ff_h264dsp_init_x86(H264DSPContext *c, const int bit_depth, c->h264_h_loop_filter_luma = ff_deblock_h_luma_8_sse2; c->h264_v_loop_filter_luma_intra = ff_deblock_v_luma_intra_8_sse2; c->h264_h_loop_filter_luma_intra = ff_deblock_h_luma_intra_8_sse2; + +#if ARCH_X86_64 + c->h264_h_loop_filter_luma_mbaff = ff_deblock_h_luma_mbaff_8_sse2; +#endif } if (EXTERNAL_SSSE3(cpu_flags)) { c->biweight_h264_pixels_tab[0] = ff_h264_biweight_16_ssse3; @@ -307,6 +314,9 @@ av_cold void ff_h264dsp_init_x86(H264DSPContext *c, const int bit_depth, c->h264_h_loop_filter_luma = ff_deblock_h_luma_8_avx; c->h264_v_loop_filter_luma_intra = ff_deblock_v_luma_intra_8_avx; c->h264_h_loop_filter_luma_intra = ff_deblock_h_luma_intra_8_avx; +#if ARCH_X86_64 + c->h264_h_loop_filter_luma_mbaff = ff_deblock_h_luma_mbaff_8_avx; +#endif } } else if (bit_depth == 10) { if (EXTERNAL_MMXEXT(cpu_flags)) { diff --git a/libavutil/x86/x86util.asm b/libavutil/x86/x86util.asm index 1408f0a176..c50ddc6c54 100644 --- a/libavutil/x86/x86util.asm +++ b/libavutil/x86/x86util.asm @@ -265,6 +265,21 @@ SWAP %12, %15 %endmacro +%macro TRANSPOSE_8X8B 8 + %if mmsize == 8 + %error "This macro does not support mmsize == 8" + %endif + punpcklbw m%1, m%2 + punpcklbw m%3, m%4 + punpcklbw m%5, m%6 + punpcklbw m%7, m%8 + TRANSPOSE4x4W %1, %3, %5, %7, %2 + MOVHL m%2, m%1 + MOVHL m%4, m%3 + MOVHL m%6, m%5 + MOVHL m%8, m%7 +%endmacro + ; PABSW macro assumes %1 != %2, while ABS1/2 macros work in-place %macro PABSW 2 %if cpuflag(ssse3) -- cgit v1.2.3