From ccdde79060a37d4694eafb9e39ce02a4eb026fd0 Mon Sep 17 00:00:00 2001 From: Christophe Gisquet Date: Fri, 8 Jan 2016 19:55:43 +0100 Subject: [PATCH] x86: float_dsp: avx version of butterflies_float No gain. --- libavutil/x86/float_dsp.asm | 33 ++++++++++++++++++++++++++++----- libavutil/x86/float_dsp_init.c | 2 ++ 2 files changed, 30 insertions(+), 5 deletions(-) diff --git a/libavutil/x86/float_dsp.asm b/libavutil/x86/float_dsp.asm index 021ff03..5948c95 100644 --- a/libavutil/x86/float_dsp.asm +++ b/libavutil/x86/float_dsp.asm @@ -357,19 +357,42 @@ cglobal scalarproduct_float, 3,3,2, v1, v2, offset ;----------------------------------------------------------------------------- ; void ff_butterflies_float(float *src0, float *src1, int len); ;----------------------------------------------------------------------------- -INIT_XMM sse +%macro BUTTERFLIES 0 cglobal butterflies_float, 3,3,3, src0, src1, len shl lend, 2 +%if cpuflag(avx) + test lend, mmsize-1 + jz .normal + sub lend, 16 + movaps xm0, [src0q + lenq] + movaps xm1, [src1q + lenq] + subps xm2, xm0, xm1 + addps xm0, xm0, xm1 + movaps [src1q + lenq], xm2 + movaps [src0q + lenq], xm0 + jz .end +.normal: +%define MOVR movu +%else +%define MOVR mova +%endif add src0q, lenq add src1q, lenq neg lenq .loop: - mova m0, [src0q + lenq] - mova m1, [src1q + lenq] + MOVR m0, [src0q + lenq] + MOVR m1, [src1q + lenq] subps m2, m0, m1 addps m0, m0, m1 - mova [src1q + lenq], m2 - mova [src0q + lenq], m0 + MOVR [src1q + lenq], m2 + MOVR [src0q + lenq], m0 add lenq, mmsize jl .loop +.end: REP_RET +%endmacro + +INIT_XMM sse +BUTTERFLIES +INIT_YMM avx +BUTTERFLIES diff --git a/libavutil/x86/float_dsp_init.c b/libavutil/x86/float_dsp_init.c index f211f23..f8b645a 100644 --- a/libavutil/x86/float_dsp_init.c +++ b/libavutil/x86/float_dsp_init.c @@ -64,6 +64,7 @@ void ff_vector_fmul_reverse_avx(float *dst, const float *src0, float ff_scalarproduct_float_sse(const float *v1, const float *v2, int order); void ff_butterflies_float_sse(float *src0, float *src1, int len); +void ff_butterflies_float_avx(float *src0, float *src1, int len); av_cold void ff_float_dsp_init_x86(AVFloatDSPContext *fdsp) { @@ -91,6 +92,7 @@ av_cold void ff_float_dsp_init_x86(AVFloatDSPContext *fdsp) fdsp->vector_dmul_scalar = ff_vector_dmul_scalar_avx; fdsp->vector_fmul_add = ff_vector_fmul_add_avx; fdsp->vector_fmul_reverse = ff_vector_fmul_reverse_avx; + fdsp->butterflies_float = ff_butterflies_float_avx; } if (EXTERNAL_FMA3(cpu_flags) && !(cpu_flags & AV_CPU_FLAG_AVXSLOW)) { fdsp->vector_fmac_scalar = ff_vector_fmac_scalar_fma3; -- 2.6.3