* [FFmpeg-devel] [PATCH] avcodec/x86/audiodsp: add scalarproduct avx2
@ 2022-09-12 18:29 Paul B Mahol
2022-09-12 18:39 ` James Almer
2022-09-13 9:25 ` Paul B Mahol
0 siblings, 2 replies; 6+ messages in thread
From: Paul B Mahol @ 2022-09-12 18:29 UTC (permalink / raw)
To: FFmpeg development discussions and patches
[-- Attachment #1: Type: text/plain, Size: 40 bytes --]
Patch attached.
How to get more speed?
[-- Attachment #2: 0001-avcodec-x86-audiodsp-add-scalarproduct-avx2.patch --]
[-- Type: text/x-patch, Size: 2187 bytes --]
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
[-- Attachment #3: Type: text/plain, Size: 251 bytes --]
_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
To unsubscribe, visit link above, or email
ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".
^ permalink raw reply [flat|nested] 6+ messages in thread
* Re: [FFmpeg-devel] [PATCH] avcodec/x86/audiodsp: add scalarproduct avx2
2022-09-12 18:29 [FFmpeg-devel] [PATCH] avcodec/x86/audiodsp: add scalarproduct avx2 Paul B Mahol
@ 2022-09-12 18:39 ` James Almer
2022-09-12 19:00 ` James Almer
2022-09-13 9:25 ` Paul B Mahol
1 sibling, 1 reply; 6+ messages in thread
From: James Almer @ 2022-09-12 18:39 UTC (permalink / raw)
To: ffmpeg-devel
> 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
>
_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
To unsubscribe, visit link above, or email
ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".
^ permalink raw reply [flat|nested] 6+ messages in thread
* Re: [FFmpeg-devel] [PATCH] avcodec/x86/audiodsp: add scalarproduct avx2
2022-09-12 18:39 ` James Almer
@ 2022-09-12 19:00 ` James Almer
0 siblings, 0 replies; 6+ messages in thread
From: James Almer @ 2022-09-12 19:00 UTC (permalink / raw)
To: ffmpeg-devel
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
>>
_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
To unsubscribe, visit link above, or email
ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".
^ permalink raw reply [flat|nested] 6+ messages in thread
* Re: [FFmpeg-devel] [PATCH] avcodec/x86/audiodsp: add scalarproduct avx2
2022-09-12 18:29 [FFmpeg-devel] [PATCH] avcodec/x86/audiodsp: add scalarproduct avx2 Paul B Mahol
2022-09-12 18:39 ` James Almer
@ 2022-09-13 9:25 ` Paul B Mahol
2022-09-13 9:28 ` Andreas Rheinhardt
2022-09-13 9:35 ` Paul B Mahol
1 sibling, 2 replies; 6+ messages in thread
From: Paul B Mahol @ 2022-09-13 9:25 UTC (permalink / raw)
To: FFmpeg development discussions and patches
[-- Attachment #1: Type: text/plain, Size: 142 bytes --]
On 9/12/22, Paul B Mahol <onemda@gmail.com> wrote:
> Patch attached.
>
> How to get more speed?
>
Thanks for review, updated patch attached.
[-- Attachment #2: 0001-avcodec-x86-audiodsp-add-scalarproduct-avx2.patch --]
[-- Type: text/x-patch, Size: 1988 bytes --]
From 2a5bab90b863c7cc9dbbb4d3d3021bec8ab09843 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 | 16 ++++++++++++++++
libavcodec/x86/audiodsp_init.c | 6 ++++++
2 files changed, 22 insertions(+)
diff --git a/libavcodec/x86/audiodsp.asm b/libavcodec/x86/audiodsp.asm
index b604b0443c..88e3fefef2 100644
--- a/libavcodec/x86/audiodsp.asm
+++ b/libavcodec/x86/audiodsp.asm
@@ -44,6 +44,22 @@ cglobal scalarproduct_int16, 3,3,3, v1, v2, order
movd eax, m2
RET
+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
;-----------------------------------------------------------------------------
; 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
[-- Attachment #3: Type: text/plain, Size: 251 bytes --]
_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
To unsubscribe, visit link above, or email
ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".
^ permalink raw reply [flat|nested] 6+ messages in thread
* Re: [FFmpeg-devel] [PATCH] avcodec/x86/audiodsp: add scalarproduct avx2
2022-09-13 9:25 ` Paul B Mahol
@ 2022-09-13 9:28 ` Andreas Rheinhardt
2022-09-13 9:35 ` Paul B Mahol
1 sibling, 0 replies; 6+ messages in thread
From: Andreas Rheinhardt @ 2022-09-13 9:28 UTC (permalink / raw)
To: ffmpeg-devel
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
_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
To unsubscribe, visit link above, or email
ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".
^ permalink raw reply [flat|nested] 6+ messages in thread
* Re: [FFmpeg-devel] [PATCH] avcodec/x86/audiodsp: add scalarproduct avx2
2022-09-13 9:25 ` Paul B Mahol
2022-09-13 9:28 ` Andreas Rheinhardt
@ 2022-09-13 9:35 ` Paul B Mahol
1 sibling, 0 replies; 6+ messages in thread
From: Paul B Mahol @ 2022-09-13 9:35 UTC (permalink / raw)
To: FFmpeg development discussions and patches
[-- Attachment #1: Type: text/plain, Size: 247 bytes --]
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.
[-- Attachment #2: 0001-avcodec-x86-audiodsp-add-scalarproduct-avx2.patch --]
[-- Type: text/x-patch, Size: 2027 bytes --]
From aa3cbb9d792d18c1756d00138925382eab924804 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 | 18 ++++++++++++++++++
libavcodec/x86/audiodsp_init.c | 6 ++++++
2 files changed, 24 insertions(+)
diff --git a/libavcodec/x86/audiodsp.asm b/libavcodec/x86/audiodsp.asm
index b604b0443c..f64077cb13 100644
--- a/libavcodec/x86/audiodsp.asm
+++ b/libavcodec/x86/audiodsp.asm
@@ -44,6 +44,24 @@ cglobal scalarproduct_int16, 3,3,3, v1, v2, order
movd eax, m2
RET
+%if HAVE_AVX2_EXTERNAL
+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
+%endif
;-----------------------------------------------------------------------------
; 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..68aa3b2129 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_FAST(cpu_flags))
+ c->scalarproduct_int16 = ff_scalarproduct_int16_avx2;
}
--
2.37.2
[-- Attachment #3: Type: text/plain, Size: 251 bytes --]
_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
To unsubscribe, visit link above, or email
ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".
^ permalink raw reply [flat|nested] 6+ messages in thread
end of thread, other threads:[~2022-09-13 9:35 UTC | newest]
Thread overview: 6+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2022-09-12 18:29 [FFmpeg-devel] [PATCH] avcodec/x86/audiodsp: add scalarproduct avx2 Paul B Mahol
2022-09-12 18:39 ` James Almer
2022-09-12 19:00 ` James Almer
2022-09-13 9:25 ` Paul B Mahol
2022-09-13 9:28 ` Andreas Rheinhardt
2022-09-13 9:35 ` Paul B Mahol
Git Inbox Mirror of the ffmpeg-devel mailing list - see https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
This inbox may be cloned and mirrored by anyone:
git clone --mirror https://master.gitmailbox.com/ffmpegdev/0 ffmpegdev/git/0.git
# If you have public-inbox 1.1+ installed, you may
# initialize and index your mirror using the following commands:
public-inbox-init -V2 ffmpegdev ffmpegdev/ https://master.gitmailbox.com/ffmpegdev \
ffmpegdev@gitmailbox.com
public-inbox-index ffmpegdev
Example config snippet for mirrors.
AGPL code for this site: git clone https://public-inbox.org/public-inbox.git