PR #20788 opened by mkver URL: https://code.ffmpeg.org/FFmpeg/FFmpeg/pulls/20788 Patch URL: https://code.ffmpeg.org/FFmpeg/FFmpeg/pulls/20788.patch
Practically no change in benchmarks (and in codesize). hevc_idct_4x4_dc_8_c: 7.8 ( 1.00x) hevc_idct_4x4_dc_8_mmxext: 6.9 ( 1.14x) hevc_idct_4x4_dc_8_sse2: 6.8 ( 1.15x) hevc_idct_4x4_dc_10_c: 7.9 ( 1.00x) hevc_idct_4x4_dc_10_mmxext: 6.9 ( 1.16x) hevc_idct_4x4_dc_10_sse2: 6.8 ( 1.16x) hevc_idct_4x4_dc_12_c: 7.8 ( 1.00x) hevc_idct_4x4_dc_12_mmxext: 7.0 ( 1.13x) hevc_idct_4x4_dc_12_sse2: 6.8 ( 1.15x) >From e81edabc4d6dfb825e6432da48a0a827b69e6ade Mon Sep 17 00:00:00 2001 From: Andreas Rheinhardt <[email protected]> Date: Thu, 30 Oct 2025 00:07:42 +0100 Subject: [PATCH] avcodec/x86/hevc/idct: Port ff_hevc_idct_4x4_dc_{8,10,12}_mmxext to SSE2 Practically no change in benchmarks (and in codesize). hevc_idct_4x4_dc_8_c: 7.8 ( 1.00x) hevc_idct_4x4_dc_8_mmxext: 6.9 ( 1.14x) hevc_idct_4x4_dc_8_sse2: 6.8 ( 1.15x) hevc_idct_4x4_dc_10_c: 7.9 ( 1.00x) hevc_idct_4x4_dc_10_mmxext: 6.9 ( 1.16x) hevc_idct_4x4_dc_10_sse2: 6.8 ( 1.16x) hevc_idct_4x4_dc_12_c: 7.8 ( 1.00x) hevc_idct_4x4_dc_12_mmxext: 7.0 ( 1.13x) hevc_idct_4x4_dc_12_sse2: 6.8 ( 1.15x) Signed-off-by: Andreas Rheinhardt <[email protected]> --- libavcodec/x86/hevc/dsp_init.c | 11 ++++------- libavcodec/x86/hevc/idct.asm | 19 ++++++------------- tests/checkasm/hevc_idct.c | 2 +- 3 files changed, 11 insertions(+), 21 deletions(-) diff --git a/libavcodec/x86/hevc/dsp_init.c b/libavcodec/x86/hevc/dsp_init.c index ba921e7299..6966340c42 100644 --- a/libavcodec/x86/hevc/dsp_init.c +++ b/libavcodec/x86/hevc/dsp_init.c @@ -65,7 +65,7 @@ void ff_hevc_idct_ ## W ## _dc_8_ ## opt(int16_t *coeffs); \ void ff_hevc_idct_ ## W ## _dc_10_ ## opt(int16_t *coeffs); \ void ff_hevc_idct_ ## W ## _dc_12_ ## opt(int16_t *coeffs) -IDCT_DC_FUNCS(4x4, mmxext); +IDCT_DC_FUNCS(4x4, sse2); IDCT_DC_FUNCS(8x8, sse2); IDCT_DC_FUNCS(16x16, sse2); IDCT_DC_FUNCS(32x32, sse2); @@ -816,8 +816,6 @@ void ff_hevc_dsp_init_x86(HEVCDSPContext *c, const int bit_depth) if (bit_depth == 8) { if (EXTERNAL_MMXEXT(cpu_flags)) { - c->idct_dc[0] = ff_hevc_idct_4x4_dc_8_mmxext; - c->add_residual[0] = ff_hevc_add_residual_4_8_mmxext; } if (EXTERNAL_SSE2(cpu_flags)) { @@ -832,6 +830,7 @@ void ff_hevc_dsp_init_x86(HEVCDSPContext *c, const int bit_depth) } SAO_BAND_INIT(8, sse2); + c->idct_dc[0] = ff_hevc_idct_4x4_dc_8_sse2; c->idct_dc[1] = ff_hevc_idct_8x8_dc_8_sse2; c->idct_dc[2] = ff_hevc_idct_16x16_dc_8_sse2; c->idct_dc[3] = ff_hevc_idct_32x32_dc_8_sse2; @@ -998,7 +997,6 @@ void ff_hevc_dsp_init_x86(HEVCDSPContext *c, const int bit_depth) } else if (bit_depth == 10) { if (EXTERNAL_MMXEXT(cpu_flags)) { c->add_residual[0] = ff_hevc_add_residual_4_10_mmxext; - c->idct_dc[0] = ff_hevc_idct_4x4_dc_10_mmxext; } if (EXTERNAL_SSE2(cpu_flags)) { c->hevc_v_loop_filter_chroma = ff_hevc_v_loop_filter_chroma_10_sse2; @@ -1013,6 +1011,7 @@ void ff_hevc_dsp_init_x86(HEVCDSPContext *c, const int bit_depth) SAO_BAND_INIT(10, sse2); SAO_EDGE_INIT(10, sse2); + c->idct_dc[0] = ff_hevc_idct_4x4_dc_10_sse2; c->idct_dc[1] = ff_hevc_idct_8x8_dc_10_sse2; c->idct_dc[2] = ff_hevc_idct_16x16_dc_10_sse2; c->idct_dc[3] = ff_hevc_idct_32x32_dc_10_sse2; @@ -1218,9 +1217,6 @@ void ff_hevc_dsp_init_x86(HEVCDSPContext *c, const int bit_depth) } #endif /* HAVE_AVX2_EXTERNAL */ } else if (bit_depth == 12) { - if (EXTERNAL_MMXEXT(cpu_flags)) { - c->idct_dc[0] = ff_hevc_idct_4x4_dc_12_mmxext; - } if (EXTERNAL_SSE2(cpu_flags)) { c->hevc_v_loop_filter_chroma = ff_hevc_v_loop_filter_chroma_12_sse2; c->hevc_h_loop_filter_chroma = ff_hevc_h_loop_filter_chroma_12_sse2; @@ -1231,6 +1227,7 @@ void ff_hevc_dsp_init_x86(HEVCDSPContext *c, const int bit_depth) SAO_BAND_INIT(12, sse2); SAO_EDGE_INIT(12, sse2); + c->idct_dc[0] = ff_hevc_idct_4x4_dc_12_sse2; c->idct_dc[1] = ff_hevc_idct_8x8_dc_12_sse2; c->idct_dc[2] = ff_hevc_idct_16x16_dc_12_sse2; c->idct_dc[3] = ff_hevc_idct_32x32_dc_12_sse2; diff --git a/libavcodec/x86/hevc/idct.asm b/libavcodec/x86/hevc/idct.asm index 021e5dab14..088144171d 100644 --- a/libavcodec/x86/hevc/idct.asm +++ b/libavcodec/x86/hevc/idct.asm @@ -273,16 +273,11 @@ cglobal hevc_idct_%1x%1_dc_%2, 1, 2, 1, coeff, tmp sar tmpd, (15 - %2) movd m0, tmpd SPLATW m0, xm0 - mova [coeffq+mmsize*0], m0 - mova [coeffq+mmsize*1], m0 - mova [coeffq+mmsize*2], m0 - mova [coeffq+mmsize*3], m0 -%if mmsize == 16 - mova [coeffq+mmsize*4], m0 - mova [coeffq+mmsize*5], m0 - mova [coeffq+mmsize*6], m0 - mova [coeffq+mmsize*7], m0 -%endif +%assign %%offset 0 +%rep 2*%1*%1/mmsize + mova [coeffq+%%offset], m0 + %assign %%offset %%offset+mmsize +%endrep RET %endmacro @@ -809,10 +804,8 @@ cglobal hevc_idct_32x32_%1, 1, 6, 16, 256, coeffs %endmacro %macro INIT_IDCT_DC 1 -INIT_MMX mmxext -IDCT_DC_NL 4, %1 - INIT_XMM sse2 +IDCT_DC_NL 4, %1 IDCT_DC_NL 8, %1 IDCT_DC 16, 4, %1 IDCT_DC 32, 16, %1 diff --git a/tests/checkasm/hevc_idct.c b/tests/checkasm/hevc_idct.c index 2bd7ae9409..139ae81727 100644 --- a/tests/checkasm/hevc_idct.c +++ b/tests/checkasm/hevc_idct.c @@ -69,7 +69,7 @@ static void check_idct_dc(HEVCDSPContext *h, int bit_depth) for (i = 2; i <= 5; i++) { int block_size = 1 << i; int size = block_size * block_size; - declare_func_emms(AV_CPU_FLAG_MMXEXT, void, int16_t *coeffs); + declare_func(void, int16_t *coeffs); randomize_buffers(coeffs0, size); memcpy(coeffs1, coeffs0, sizeof(*coeffs0) * size); -- 2.49.1 _______________________________________________ ffmpeg-devel mailing list -- [email protected] To unsubscribe send an email to [email protected]
