Message ID | CAPYw7P5YuCyxft1WyU3SsoFxzu3qOnhwtzWHFTkPx9_Ax9df_A@mail.gmail.com |
---|---|
State | New |
Headers | show |
Series | [FFmpeg-devel] avcodec/x86/audiodsp: add scalarproduct avx2 | expand |
Context | Check | Description |
---|---|---|
andriy/make_x86 | success | Make finished |
andriy/make_fate_x86 | success | Make fate finished |
> From 55eb5a18b4bf029f52f9d9108a750c576ba780ee Mon Sep 17 00:00:00 2001 > From: Paul B Mahol <onemda@gmail.com> > Date: Mon, 12 Sep 2022 18:53:31 +0200 > Subject: [PATCH] avcodec/x86/audiodsp: add scalarproduct avx2 > > Signed-off-by: Paul B Mahol <onemda@gmail.com> > --- > libavcodec/x86/audiodsp.asm | 24 ++++++++++++++++++++++++ > libavcodec/x86/audiodsp_init.c | 6 ++++++ > 2 files changed, 30 insertions(+) > > diff --git a/libavcodec/x86/audiodsp.asm b/libavcodec/x86/audiodsp.asm > index b604b0443c..55051f6aa7 100644 > --- a/libavcodec/x86/audiodsp.asm > +++ b/libavcodec/x86/audiodsp.asm > @@ -44,6 +44,30 @@ cglobal scalarproduct_int16, 3,3,3, v1, v2, order > movd eax, m2 > RET > > +INIT_YMM avx2 > +cglobal scalarproduct_int16, 3,4,3, v1, v2, order, offset > + xor offsetq, offsetq > + add orderd, orderd > + pxor m1, m1 > + cmp orderd, 32 This parameter needs to be multiple of 16. What will happen below if it's for example 48? Are both buffers padded enough to handle 16 bytes of overread? > + jl .l16 > +.loop: > + movu m0, [v1q + offsetq] > + pmaddwd m0, [v2q + offsetq] > + paddd m1, m0 > + add offsetq, mmsize > + cmp offsetq, orderq You should use the neg trick from the sse2 version so you can remove the cmp from this loop. > + jl .loop > + HADDD m1, m0 > + movd eax, xm1 > + RET > +.l16: > + movu xm0, [v1q + offsetq] > + pmaddwd xm0, [v2q + offsetq] > + paddd xm1, xm0 > + HADDD xm1, xm0 > + movd eax, xm1 > + RET > > ;----------------------------------------------------------------------------- > ; void ff_vector_clip_int32(int32_t *dst, const int32_t *src, int32_t min, > diff --git a/libavcodec/x86/audiodsp_init.c b/libavcodec/x86/audiodsp_init.c > index aa5e43e570..77d5948442 100644 > --- a/libavcodec/x86/audiodsp_init.c > +++ b/libavcodec/x86/audiodsp_init.c > @@ -24,6 +24,9 @@ > #include "libavutil/x86/cpu.h" > #include "libavcodec/audiodsp.h" > > +int32_t ff_scalarproduct_int16_avx2(const int16_t *v1, const int16_t *v2, > + int order); > + > int32_t ff_scalarproduct_int16_sse2(const int16_t *v1, const int16_t *v2, > int order); > > @@ -53,4 +56,7 @@ av_cold void ff_audiodsp_init_x86(AudioDSPContext *c) > > if (EXTERNAL_SSE4(cpu_flags)) > c->vector_clip_int32 = ff_vector_clip_int32_sse4; > + > + if (EXTERNAL_AVX2(cpu_flags)) > + c->scalarproduct_int16 = ff_scalarproduct_int16_avx2; > } > -- > 2.37.2 >
On 9/12/2022 3:39 PM, James Almer wrote: >> From 55eb5a18b4bf029f52f9d9108a750c576ba780ee Mon Sep 17 00:00:00 2001 >> From: Paul B Mahol <onemda@gmail.com> >> Date: Mon, 12 Sep 2022 18:53:31 +0200 >> Subject: [PATCH] avcodec/x86/audiodsp: add scalarproduct avx2 >> >> Signed-off-by: Paul B Mahol <onemda@gmail.com> >> --- >> libavcodec/x86/audiodsp.asm | 24 ++++++++++++++++++++++++ >> libavcodec/x86/audiodsp_init.c | 6 ++++++ >> 2 files changed, 30 insertions(+) >> >> diff --git a/libavcodec/x86/audiodsp.asm b/libavcodec/x86/audiodsp.asm >> index b604b0443c..55051f6aa7 100644 >> --- a/libavcodec/x86/audiodsp.asm >> +++ b/libavcodec/x86/audiodsp.asm >> @@ -44,6 +44,30 @@ cglobal scalarproduct_int16, 3,3,3, v1, v2, order >> movd eax, m2 >> RET >> >> +INIT_YMM avx2 >> +cglobal scalarproduct_int16, 3,4,3, v1, v2, order, offset >> + xor offsetq, offsetq >> + add orderd, orderd >> + pxor m1, m1 >> + cmp orderd, 32 > > This parameter needs to be multiple of 16. What will happen below if > it's for example 48? Are both buffers padded enough to handle 16 bytes > of overread? Nevermind, it's int16_t* buffers. You can simplify this as: INIT_YMM avx2 cglobal scalarproduct_int16, 3,3,3, v1, v2, order add orderd, orderd add v1q, orderq add v2q, orderq neg orderq pxor m1, m1 .loop: movu m0, [v1q + orderq] pmaddwd m0, [v2q + orderq] paddd m1, m0 add orderq, mmsize jl .loop HADDD m1, m0 movd eax, xm1 RET > >> + jl .l16 >> +.loop: >> + movu m0, [v1q + offsetq] >> + pmaddwd m0, [v2q + offsetq] >> + paddd m1, m0 >> + add offsetq, mmsize >> + cmp offsetq, orderq > > You should use the neg trick from the sse2 version so you can remove the > cmp from this loop. > >> + jl .loop >> + HADDD m1, m0 >> + movd eax, xm1 >> + RET >> +.l16: >> + movu xm0, [v1q + offsetq] >> + pmaddwd xm0, [v2q + offsetq] >> + paddd xm1, xm0 >> + HADDD xm1, xm0 >> + movd eax, xm1 >> + RET >> >> ;----------------------------------------------------------------------------- >> >> ; void ff_vector_clip_int32(int32_t *dst, const int32_t *src, int32_t >> min, >> diff --git a/libavcodec/x86/audiodsp_init.c >> b/libavcodec/x86/audiodsp_init.c >> index aa5e43e570..77d5948442 100644 >> --- a/libavcodec/x86/audiodsp_init.c >> +++ b/libavcodec/x86/audiodsp_init.c >> @@ -24,6 +24,9 @@ >> #include "libavutil/x86/cpu.h" >> #include "libavcodec/audiodsp.h" >> >> +int32_t ff_scalarproduct_int16_avx2(const int16_t *v1, const int16_t >> *v2, >> + int order); >> + >> int32_t ff_scalarproduct_int16_sse2(const int16_t *v1, const int16_t >> *v2, >> int order); >> >> @@ -53,4 +56,7 @@ av_cold void ff_audiodsp_init_x86(AudioDSPContext *c) >> >> if (EXTERNAL_SSE4(cpu_flags)) >> c->vector_clip_int32 = ff_vector_clip_int32_sse4; >> + >> + if (EXTERNAL_AVX2(cpu_flags)) >> + c->scalarproduct_int16 = ff_scalarproduct_int16_avx2; >> } >> -- >> 2.37.2 >>
On 9/12/22, Paul B Mahol <onemda@gmail.com> wrote: > Patch attached. > > How to get more speed? > Thanks for review, updated patch attached.
Paul B Mahol: > > +INIT_YMM avx2 > +cglobal scalarproduct_int16, 3,3,2, v1, v2, order > + add orderd, orderd > + add v1q, orderq > + add v2q, orderq > + neg orderq > + pxor m1, m1 > +.loop: > + movu m0, [v1q + orderq] > + pmaddwd m0, [v2q + orderq] > + paddd m1, m0 > + add orderq, mmsize > + jl .loop > + HADDD m1, m0 > + movd eax, xm1 > + RET > avx2 asm functions are guarded by %if HAVE_AVX2_EXTERNAL (presumably not all assemblers can handle them?). - Andreas
On 9/13/22, Paul B Mahol <onemda@gmail.com> wrote: > On 9/12/22, Paul B Mahol <onemda@gmail.com> wrote: >> Patch attached. >> >> How to get more speed? >> > > Thanks for review, updated patch attached. > Fixed patch from Andreas review attached.
From 55eb5a18b4bf029f52f9d9108a750c576ba780ee Mon Sep 17 00:00:00 2001 From: Paul B Mahol <onemda@gmail.com> Date: Mon, 12 Sep 2022 18:53:31 +0200 Subject: [PATCH] avcodec/x86/audiodsp: add scalarproduct avx2 Signed-off-by: Paul B Mahol <onemda@gmail.com> --- libavcodec/x86/audiodsp.asm | 24 ++++++++++++++++++++++++ libavcodec/x86/audiodsp_init.c | 6 ++++++ 2 files changed, 30 insertions(+) diff --git a/libavcodec/x86/audiodsp.asm b/libavcodec/x86/audiodsp.asm index b604b0443c..55051f6aa7 100644 --- a/libavcodec/x86/audiodsp.asm +++ b/libavcodec/x86/audiodsp.asm @@ -44,6 +44,30 @@ cglobal scalarproduct_int16, 3,3,3, v1, v2, order movd eax, m2 RET +INIT_YMM avx2 +cglobal scalarproduct_int16, 3,4,3, v1, v2, order, offset + xor offsetq, offsetq + add orderd, orderd + pxor m1, m1 + cmp orderd, 32 + jl .l16 +.loop: + movu m0, [v1q + offsetq] + pmaddwd m0, [v2q + offsetq] + paddd m1, m0 + add offsetq, mmsize + cmp offsetq, orderq + jl .loop + HADDD m1, m0 + movd eax, xm1 + RET +.l16: + movu xm0, [v1q + offsetq] + pmaddwd xm0, [v2q + offsetq] + paddd xm1, xm0 + HADDD xm1, xm0 + movd eax, xm1 + RET ;----------------------------------------------------------------------------- ; void ff_vector_clip_int32(int32_t *dst, const int32_t *src, int32_t min, diff --git a/libavcodec/x86/audiodsp_init.c b/libavcodec/x86/audiodsp_init.c index aa5e43e570..77d5948442 100644 --- a/libavcodec/x86/audiodsp_init.c +++ b/libavcodec/x86/audiodsp_init.c @@ -24,6 +24,9 @@ #include "libavutil/x86/cpu.h" #include "libavcodec/audiodsp.h" +int32_t ff_scalarproduct_int16_avx2(const int16_t *v1, const int16_t *v2, + int order); + int32_t ff_scalarproduct_int16_sse2(const int16_t *v1, const int16_t *v2, int order); @@ -53,4 +56,7 @@ av_cold void ff_audiodsp_init_x86(AudioDSPContext *c) if (EXTERNAL_SSE4(cpu_flags)) c->vector_clip_int32 = ff_vector_clip_int32_sse4; + + if (EXTERNAL_AVX2(cpu_flags)) + c->scalarproduct_int16 = ff_scalarproduct_int16_avx2; } -- 2.37.2