[FFmpeg-devel] [PATCH 1/2] checkasm/takdsp: add decorrelate_sf test

[FFmpeg-devel] [PATCH 1/2] checkasm/takdsp: add decorrelate_sf test

[James Almer]

Signed-off-by: James Almer <jamrial@gmail.com>
---
 tests/checkasm/takdsp.c | 36 +++++++++++++++++++++++++++++++++---
 1 file changed, 33 insertions(+), 3 deletions(-)

diff --git a/tests/checkasm/takdsp.c b/tests/checkasm/takdsp.c
index 495b7242c5..8df93cfd52 100644
--- a/tests/checkasm/takdsp.c
+++ b/tests/checkasm/takdsp.c
@@ -24,6 +24,7 @@
 #include "libavutil/mem_internal.h"
 
 #include "libavcodec/takdsp.h"
+#include "libavcodec/mathops.h"
 
 #include "checkasm.h"
 
@@ -33,8 +34,9 @@
             buf[i] = rnd(); \
     } while (0)
 
-static void test_decorrelate_ls(TAKDSPContext *s) {
 #define BUF_SIZE 1024
+
+static void test_decorrelate_ls(TAKDSPContext *s) {
     declare_func(void, int32_t *, int32_t *, int);
 
     if (check_func(s->decorrelate_ls, "decorrelate_ls")) {
@@ -60,7 +62,6 @@ static void test_decorrelate_ls(TAKDSPContext *s) {
 }
 
 static void test_decorrelate_sr(TAKDSPContext *s) {
-#define BUF_SIZE 1024
     declare_func(void, int32_t *, int32_t *, int);
 
     if (check_func(s->decorrelate_sr, "decorrelate_sr")) {
@@ -86,7 +87,6 @@ static void test_decorrelate_sr(TAKDSPContext *s) {
 }
 
 static void test_decorrelate_sm(TAKDSPContext *s) {
-#define BUF_SIZE 1024
     declare_func(void, int32_t *, int32_t *, int);
 
     if (check_func(s->decorrelate_sm, "decorrelate_sm")) {
@@ -113,6 +113,35 @@ static void test_decorrelate_sm(TAKDSPContext *s) {
     report("decorrelate_sm");
 }
 
+static void test_decorrelate_sf(TAKDSPContext *s) {
+    declare_func(void, int32_t *, int32_t *, int, int, int);
+
+    if (check_func(s->decorrelate_sf, "decorrelate_sf")) {
+        LOCAL_ALIGNED_32(int32_t, p1, [BUF_SIZE]);
+        LOCAL_ALIGNED_32(int32_t, p1_2, [BUF_SIZE]);
+        LOCAL_ALIGNED_32(int32_t, p2, [BUF_SIZE]);
+        LOCAL_ALIGNED_32(int32_t, p2_2, [BUF_SIZE]);
+        int dshift, dfactor;
+
+        randomize(p1, BUF_SIZE);
+        memcpy(p1, p1_2, BUF_SIZE);
+        randomize(p2, BUF_SIZE);
+        memcpy(p2_2, p2, BUF_SIZE);
+        dshift = (rnd() & 0xF) + 1;
+        dfactor = sign_extend(rnd(), 10);
+        call_ref(p1, p2, BUF_SIZE, dshift, dfactor);
+        call_new(p1_2, p2_2, BUF_SIZE, dshift, dfactor);
+
+        if (memcmp(p2, p2_2, BUF_SIZE) != 0){
+            fail();
+        }
+
+        bench_new(p1, p2, BUF_SIZE, dshift, dfactor);
+    }
+
+    report("decorrelate_sf");
+}
+
 void checkasm_check_takdsp(void)
 {
     TAKDSPContext s = { 0 };
@@ -121,4 +150,5 @@ void checkasm_check_takdsp(void)
     test_decorrelate_ls(&s);
     test_decorrelate_sr(&s);
     test_decorrelate_sm(&s);
+    test_decorrelate_sf(&s);
 }
-- 
2.43.0

_______________________________________________
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".

[FFmpeg-devel] [PATCH 2/2] x86/takdsp: add avx2 versions of all functions

[James Almer]

On an Intel Core i7 12700k:

decorrelate_ls_c: 814.3
decorrelate_ls_sse2: 165.8
decorrelate_ls_avx2: 101.3
decorrelate_sf_c: 1602.6
decorrelate_sf_sse4: 640.1
decorrelate_sf_avx2: 324.6
decorrelate_sm_c: 1564.8
decorrelate_sm_sse2: 379.3
decorrelate_sm_avx2: 203.3
decorrelate_sr_c: 785.3
decorrelate_sr_sse2: 176.3
decorrelate_sr_avx2: 99.8

Signed-off-by: James Almer <jamrial@gmail.com>
---
 libavcodec/x86/takdsp.asm    | 36 ++++++++++++++++++++++--------------
 libavcodec/x86/takdsp_init.c | 11 +++++++++++
 2 files changed, 33 insertions(+), 14 deletions(-)

diff --git a/libavcodec/x86/takdsp.asm b/libavcodec/x86/takdsp.asm
index be8e1ab553..a5501cc285 100644
--- a/libavcodec/x86/takdsp.asm
+++ b/libavcodec/x86/takdsp.asm
@@ -28,7 +28,7 @@ pd_128: times 4 dd 128
 
 SECTION .text
 
-INIT_XMM sse2
+%macro TAK_DECORRELATE 0
 cglobal tak_decorrelate_ls, 3, 3, 2, p1, p2, length
     shl                     lengthd, 2
     add                         p1q, lengthq
@@ -73,10 +73,8 @@ cglobal tak_decorrelate_sm, 3, 3, 6, p1, p2, length
     mova                         m1, [p2q+lengthq]
     mova                         m3, [p1q+lengthq+mmsize]
     mova                         m4, [p2q+lengthq+mmsize]
-    mova                         m2, m1
-    mova                         m5, m4
-    psrad                        m2, 1
-    psrad                        m5, 1
+    psrad                        m2, m1, 1
+    psrad                        m5, m4, 1
     psubd                        m0, m2
     psubd                        m3, m5
     paddd                        m1, m0
@@ -88,29 +86,39 @@ cglobal tak_decorrelate_sm, 3, 3, 6, p1, p2, length
     add                     lengthq, mmsize*2
     jl .loop
     RET
+%endmacro
 
-INIT_XMM sse4
+INIT_XMM sse2
+TAK_DECORRELATE
+INIT_YMM avx2
+TAK_DECORRELATE
+
+%macro TAK_DECORRELATE_SF 0
 cglobal tak_decorrelate_sf, 3, 3, 5, p1, p2, length, dshift, dfactor
     shl             lengthd, 2
     add                 p1q, lengthq
     add                 p2q, lengthq
     neg             lengthq
 
-    movd                 m2, dshiftm
-    movd                 m3, dfactorm
-    pshufd               m3, m3, 0
-    mova                 m4, [pd_128]
+    movd                xm2, dshiftm
+    VPBROADCASTD         m3, dfactorm
+    VBROADCASTI128       m4, [pd_128]
 
 .loop:
-    mova                 m0, [p1q+lengthq]
     mova                 m1, [p2q+lengthq]
-    psrad                m1, m2
+    psrad                m1, xm2
     pmulld               m1, m3
     paddd                m1, m4
     psrad                m1, 8
-    pslld                m1, m2
-    psubd                m1, m0
+    pslld                m1, xm2
+    psubd                m1, [p1q+lengthq]
     mova      [p1q+lengthq], m1
     add             lengthq, mmsize
     jl .loop
     RET
+%endmacro
+
+INIT_XMM sse4
+TAK_DECORRELATE_SF
+INIT_YMM avx2
+TAK_DECORRELATE_SF
diff --git a/libavcodec/x86/takdsp_init.c b/libavcodec/x86/takdsp_init.c
index b2e6e639ee..c99a057b24 100644
--- a/libavcodec/x86/takdsp_init.c
+++ b/libavcodec/x86/takdsp_init.c
@@ -24,9 +24,13 @@
 #include "config.h"
 
 void ff_tak_decorrelate_ls_sse2(int32_t *p1, int32_t *p2, int length);
+void ff_tak_decorrelate_ls_avx2(int32_t *p1, int32_t *p2, int length);
 void ff_tak_decorrelate_sr_sse2(int32_t *p1, int32_t *p2, int length);
+void ff_tak_decorrelate_sr_avx2(int32_t *p1, int32_t *p2, int length);
 void ff_tak_decorrelate_sm_sse2(int32_t *p1, int32_t *p2, int length);
+void ff_tak_decorrelate_sm_avx2(int32_t *p1, int32_t *p2, int length);
 void ff_tak_decorrelate_sf_sse4(int32_t *p1, int32_t *p2, int length, int dshift, int dfactor);
+void ff_tak_decorrelate_sf_avx2(int32_t *p1, int32_t *p2, int length, int dshift, int dfactor);
 
 av_cold void ff_takdsp_init_x86(TAKDSPContext *c)
 {
@@ -42,5 +46,12 @@ av_cold void ff_takdsp_init_x86(TAKDSPContext *c)
     if (EXTERNAL_SSE4(cpu_flags)) {
         c->decorrelate_sf = ff_tak_decorrelate_sf_sse4;
     }
+
+    if (EXTERNAL_AVX2_FAST(cpu_flags)) {
+        c->decorrelate_ls = ff_tak_decorrelate_ls_avx2;
+        c->decorrelate_sr = ff_tak_decorrelate_sr_avx2;
+        c->decorrelate_sm = ff_tak_decorrelate_sm_avx2;
+        c->decorrelate_sf = ff_tak_decorrelate_sf_avx2;
+    }
 #endif
 }
-- 
2.43.0

_______________________________________________
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".

Re: [FFmpeg-devel] [PATCH 2/2] x86/takdsp: add avx2 versions of all functions

[Michael Niedermayer]

[-- Attachment #1.1: Type: text/plain, Size: 1786 bytes --]

On Thu, Dec 21, 2023 at 10:15:49PM -0300, James Almer wrote:
> On an Intel Core i7 12700k:
> 
> decorrelate_ls_c: 814.3
> decorrelate_ls_sse2: 165.8
> decorrelate_ls_avx2: 101.3
> decorrelate_sf_c: 1602.6
> decorrelate_sf_sse4: 640.1
> decorrelate_sf_avx2: 324.6
> decorrelate_sm_c: 1564.8
> decorrelate_sm_sse2: 379.3
> decorrelate_sm_avx2: 203.3
> decorrelate_sr_c: 785.3
> decorrelate_sr_sse2: 176.3
> decorrelate_sr_avx2: 99.8
> 
> Signed-off-by: James Almer <jamrial@gmail.com>

on AMD Ryzen 9 3950X 16-Core Processor

Illegal instruction (core dumped)
threads=1
tests/Makefile:308: recipe for target 'fate-lossless-tak' failed
make: *** [fate-lossless-tak] Error 132

(gdb) disassemble $rip-32, $rip+32
Dump of assembler code from 0x55555651a580 to 0x55555651a5c0:
   0x000055555651a580:	or     $0x17,%al
   0x000055555651a582:	movdqa %xmm1,(%rdi,%rdx,1)
   0x000055555651a587:	add    $0x10,%rdx
   0x000055555651a58b:	jl     0x55555651a562
   0x000055555651a58d:	retq
   0x000055555651a58e:	nop
   0x000055555651a58f:	nop
   0x000055555651a590:	shl    $0x2,%edx
   0x000055555651a593:	add    %rdx,%rdi
   0x000055555651a596:	add    %rdx,%rsi
   0x000055555651a599:	neg    %rdx
   0x000055555651a59c:	vmovd  %ecx,%xmm2
=> 0x000055555651a5a0:	vpbroadcastd %r8d,%ymm3
   0x000055555651a5a6:	vbroadcasti128 0x4bc751(%rip),%ymm4        # 0x5555569d6d00
   0x000055555651a5af:	vmovdqa (%rsi,%rdx,1),%ymm1
   0x000055555651a5b4:	vpsrad %xmm2,%ymm1,%ymm1
   0x000055555651a5b8:	vpmulld %ymm3,%ymm1,%ymm1
   0x000055555651a5bd:	vpaddd %ymm4,%ymm1,%ymm1
End of assembler dump.


[...]
-- 
Michael     GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB

Why not whip the teacher when the pupil misbehaves? -- Diogenes of Sinope

[-- Attachment #1.2: signature.asc --]
[-- Type: application/pgp-signature, Size: 195 bytes --]

[-- Attachment #2: 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".

Re: [FFmpeg-devel] [PATCH 2/2] x86/takdsp: add avx2 versions of all functions

[James Almer]

On 12/22/2023 8:08 PM, Michael Niedermayer wrote:
> On Thu, Dec 21, 2023 at 10:15:49PM -0300, James Almer wrote:
>> On an Intel Core i7 12700k:
>>
>> decorrelate_ls_c: 814.3
>> decorrelate_ls_sse2: 165.8
>> decorrelate_ls_avx2: 101.3
>> decorrelate_sf_c: 1602.6
>> decorrelate_sf_sse4: 640.1
>> decorrelate_sf_avx2: 324.6
>> decorrelate_sm_c: 1564.8
>> decorrelate_sm_sse2: 379.3
>> decorrelate_sm_avx2: 203.3
>> decorrelate_sr_c: 785.3
>> decorrelate_sr_sse2: 176.3
>> decorrelate_sr_avx2: 99.8
>>
>> Signed-off-by: James Almer <jamrial@gmail.com>
> 
> on AMD Ryzen 9 3950X 16-Core Processor
> 
> Illegal instruction (core dumped)
> threads=1
> tests/Makefile:308: recipe for target 'fate-lossless-tak' failed
> make: *** [fate-lossless-tak] Error 132
> 
> (gdb) disassemble $rip-32, $rip+32
> Dump of assembler code from 0x55555651a580 to 0x55555651a5c0:
>     0x000055555651a580:	or     $0x17,%al
>     0x000055555651a582:	movdqa %xmm1,(%rdi,%rdx,1)
>     0x000055555651a587:	add    $0x10,%rdx
>     0x000055555651a58b:	jl     0x55555651a562
>     0x000055555651a58d:	retq
>     0x000055555651a58e:	nop
>     0x000055555651a58f:	nop
>     0x000055555651a590:	shl    $0x2,%edx
>     0x000055555651a593:	add    %rdx,%rdi
>     0x000055555651a596:	add    %rdx,%rsi
>     0x000055555651a599:	neg    %rdx
>     0x000055555651a59c:	vmovd  %ecx,%xmm2
> => 0x000055555651a5a0:	vpbroadcastd %r8d,%ymm3

Right, on linux the fifth argument is on a gpr, and vpbroadcastd with 
gpr source is avx512.

Will fix and resend.

>     0x000055555651a5a6:	vbroadcasti128 0x4bc751(%rip),%ymm4        # 0x5555569d6d00
>     0x000055555651a5af:	vmovdqa (%rsi,%rdx,1),%ymm1
>     0x000055555651a5b4:	vpsrad %xmm2,%ymm1,%ymm1
>     0x000055555651a5b8:	vpmulld %ymm3,%ymm1,%ymm1
>     0x000055555651a5bd:	vpaddd %ymm4,%ymm1,%ymm1
> End of assembler dump.
> 
> 
> [...]
> 
> 
> _______________________________________________
> 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".
_______________________________________________
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".

[FFmpeg-devel] [PATCH 2/2 v2] x86/takdsp: add avx2 versions of all functions

[James Almer]

On an Intel Core i7 12700k:

decorrelate_ls_c: 814.3
decorrelate_ls_sse2: 165.8
decorrelate_ls_avx2: 101.3
decorrelate_sf_c: 1602.6
decorrelate_sf_sse4: 640.1
decorrelate_sf_avx2: 324.6
decorrelate_sm_c: 1564.8
decorrelate_sm_sse2: 379.3
decorrelate_sm_avx2: 203.3
decorrelate_sr_c: 785.3
decorrelate_sr_sse2: 176.3
decorrelate_sr_avx2: 99.8

Signed-off-by: James Almer <jamrial@gmail.com>
---
 libavcodec/x86/takdsp.asm    | 41 ++++++++++++++++++++++++------------
 libavcodec/x86/takdsp_init.c | 11 ++++++++++
 2 files changed, 38 insertions(+), 14 deletions(-)

diff --git a/libavcodec/x86/takdsp.asm b/libavcodec/x86/takdsp.asm
index be8e1ab553..d55c5f39aa 100644
--- a/libavcodec/x86/takdsp.asm
+++ b/libavcodec/x86/takdsp.asm
@@ -28,7 +28,7 @@ pd_128: times 4 dd 128
 
 SECTION .text
 
-INIT_XMM sse2
+%macro TAK_DECORRELATE 0
 cglobal tak_decorrelate_ls, 3, 3, 2, p1, p2, length
     shl                     lengthd, 2
     add                         p1q, lengthq
@@ -73,10 +73,8 @@ cglobal tak_decorrelate_sm, 3, 3, 6, p1, p2, length
     mova                         m1, [p2q+lengthq]
     mova                         m3, [p1q+lengthq+mmsize]
     mova                         m4, [p2q+lengthq+mmsize]
-    mova                         m2, m1
-    mova                         m5, m4
-    psrad                        m2, 1
-    psrad                        m5, 1
+    psrad                        m2, m1, 1
+    psrad                        m5, m4, 1
     psubd                        m0, m2
     psubd                        m3, m5
     paddd                        m1, m0
@@ -88,29 +86,44 @@ cglobal tak_decorrelate_sm, 3, 3, 6, p1, p2, length
     add                     lengthq, mmsize*2
     jl .loop
     RET
+%endmacro
 
-INIT_XMM sse4
+INIT_XMM sse2
+TAK_DECORRELATE
+INIT_YMM avx2
+TAK_DECORRELATE
+
+%macro TAK_DECORRELATE_SF 0
 cglobal tak_decorrelate_sf, 3, 3, 5, p1, p2, length, dshift, dfactor
     shl             lengthd, 2
     add                 p1q, lengthq
     add                 p2q, lengthq
     neg             lengthq
 
-    movd                 m2, dshiftm
-    movd                 m3, dfactorm
-    pshufd               m3, m3, 0
-    mova                 m4, [pd_128]
+    movd                xm2, dshiftm
+%if UNIX64
+    movd                xm3, dfactorm
+    VPBROADCASTD         m3, xm3
+%else
+    VPBROADCASTD         m3, dfactorm
+%endif
+    VBROADCASTI128       m4, [pd_128]
 
 .loop:
-    mova                 m0, [p1q+lengthq]
     mova                 m1, [p2q+lengthq]
-    psrad                m1, m2
+    psrad                m1, xm2
     pmulld               m1, m3
     paddd                m1, m4
     psrad                m1, 8
-    pslld                m1, m2
-    psubd                m1, m0
+    pslld                m1, xm2
+    psubd                m1, [p1q+lengthq]
     mova      [p1q+lengthq], m1
     add             lengthq, mmsize
     jl .loop
     RET
+%endmacro
+
+INIT_XMM sse4
+TAK_DECORRELATE_SF
+INIT_YMM avx2
+TAK_DECORRELATE_SF
diff --git a/libavcodec/x86/takdsp_init.c b/libavcodec/x86/takdsp_init.c
index 12b62b8247..9553f8442c 100644
--- a/libavcodec/x86/takdsp_init.c
+++ b/libavcodec/x86/takdsp_init.c
@@ -24,9 +24,13 @@
 #include "config.h"
 
 void ff_tak_decorrelate_ls_sse2(const int32_t *p1, int32_t *p2, int length);
+void ff_tak_decorrelate_ls_avx2(const int32_t *p1, int32_t *p2, int length);
 void ff_tak_decorrelate_sr_sse2(int32_t *p1, const int32_t *p2, int length);
+void ff_tak_decorrelate_sr_avx2(int32_t *p1, const int32_t *p2, int length);
 void ff_tak_decorrelate_sm_sse2(int32_t *p1, int32_t *p2, int length);
+void ff_tak_decorrelate_sm_avx2(int32_t *p1, int32_t *p2, int length);
 void ff_tak_decorrelate_sf_sse4(int32_t *p1, const int32_t *p2, int length, int dshift, int dfactor);
+void ff_tak_decorrelate_sf_avx2(int32_t *p1, const int32_t *p2, int length, int dshift, int dfactor);
 
 av_cold void ff_takdsp_init_x86(TAKDSPContext *c)
 {
@@ -42,5 +46,12 @@ av_cold void ff_takdsp_init_x86(TAKDSPContext *c)
     if (EXTERNAL_SSE4(cpu_flags)) {
         c->decorrelate_sf = ff_tak_decorrelate_sf_sse4;
     }
+
+    if (EXTERNAL_AVX2_FAST(cpu_flags)) {
+        c->decorrelate_ls = ff_tak_decorrelate_ls_avx2;
+        c->decorrelate_sr = ff_tak_decorrelate_sr_avx2;
+        c->decorrelate_sm = ff_tak_decorrelate_sm_avx2;
+        c->decorrelate_sf = ff_tak_decorrelate_sf_avx2;
+    }
 #endif
 }
-- 
2.43.0

_______________________________________________
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".

Re: [FFmpeg-devel] [PATCH 2/2 v2] x86/takdsp: add avx2 versions of all functions

[Lynne]

Dec 23, 2023, 00:53 by jamrial@gmail.com:

> On an Intel Core i7 12700k:
>
> decorrelate_ls_c: 814.3
> decorrelate_ls_sse2: 165.8
> decorrelate_ls_avx2: 101.3
> decorrelate_sf_c: 1602.6
> decorrelate_sf_sse4: 640.1
> decorrelate_sf_avx2: 324.6
> decorrelate_sm_c: 1564.8
> decorrelate_sm_sse2: 379.3
> decorrelate_sm_avx2: 203.3
> decorrelate_sr_c: 785.3
> decorrelate_sr_sse2: 176.3
> decorrelate_sr_avx2: 99.8
>
> Signed-off-by: James Almer <jamrial@gmail.com>
>

Even better on a Zen3:
checkasm: all 8 tests passed
decorrelate_ls_c: 111.1
decorrelate_ls_sse2: 272.6
decorrelate_ls_avx2: 94.1
decorrelate_sf_c: 170.6
decorrelate_sf_sse4: 400.1
decorrelate_sf_avx2: 196.1
decorrelate_sm_c: 187.6
decorrelate_sm_sse2: 383.1
decorrelate_sm_avx2: 179.1
decorrelate_sr_c: 102.6
decorrelate_sr_sse2: 272.6
decorrelate_sr_avx2: 94.1

Tested, decoding works fine too, LGTM
_______________________________________________
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".

Re: [FFmpeg-devel] [PATCH 2/2 v2] x86/takdsp: add avx2 versions of all functions

[Andreas Rheinhardt]

Lynne:
> Dec 23, 2023, 00:53 by jamrial@gmail.com:
> 
>> On an Intel Core i7 12700k:
>>
>> decorrelate_ls_c: 814.3
>> decorrelate_ls_sse2: 165.8
>> decorrelate_ls_avx2: 101.3
>> decorrelate_sf_c: 1602.6
>> decorrelate_sf_sse4: 640.1
>> decorrelate_sf_avx2: 324.6
>> decorrelate_sm_c: 1564.8
>> decorrelate_sm_sse2: 379.3
>> decorrelate_sm_avx2: 203.3
>> decorrelate_sr_c: 785.3
>> decorrelate_sr_sse2: 176.3
>> decorrelate_sr_avx2: 99.8
>>
>> Signed-off-by: James Almer <jamrial@gmail.com>
>>
> 
> Even better on a Zen3:
> checkasm: all 8 tests passed
> decorrelate_ls_c: 111.1
> decorrelate_ls_sse2: 272.6
> decorrelate_ls_avx2: 94.1
> decorrelate_sf_c: 170.6
> decorrelate_sf_sse4: 400.1
> decorrelate_sf_avx2: 196.1
> decorrelate_sm_c: 187.6
> decorrelate_sm_sse2: 383.1
> decorrelate_sm_avx2: 179.1
> decorrelate_sr_c: 102.6
> decorrelate_sr_sse2: 272.6
> decorrelate_sr_avx2: 94.1
> 

The SSE2 version is worse than the C version? Does this happen for more
DSP code?
(For decorrelate_sf_c, the C version is still the best and the gain of
AVX2 over C is not good for the other three either.)

- 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".

Re: [FFmpeg-devel] [PATCH 1/2] checkasm/takdsp: add decorrelate_sf test

[Martin Storsjö]

On Thu, 21 Dec 2023, James Almer wrote:

> Signed-off-by: James Almer <jamrial@gmail.com>
> ---
> tests/checkasm/takdsp.c | 36 +++++++++++++++++++++++++++++++++---
> 1 file changed, 33 insertions(+), 3 deletions(-)

> @@ -113,6 +113,35 @@ static void test_decorrelate_sm(TAKDSPContext *s) {
>     report("decorrelate_sm");
> }
>
> +static void test_decorrelate_sf(TAKDSPContext *s) {
> +    declare_func(void, int32_t *, int32_t *, int, int, int);
> +
> +    if (check_func(s->decorrelate_sf, "decorrelate_sf")) {
> +        LOCAL_ALIGNED_32(int32_t, p1, [BUF_SIZE]);
> +        LOCAL_ALIGNED_32(int32_t, p1_2, [BUF_SIZE]);
> +        LOCAL_ALIGNED_32(int32_t, p2, [BUF_SIZE]);
> +        LOCAL_ALIGNED_32(int32_t, p2_2, [BUF_SIZE]);
> +        int dshift, dfactor;
> +
> +        randomize(p1, BUF_SIZE);
> +        memcpy(p1, p1_2, BUF_SIZE);

The source/destination of the memcpy is flipped here. And it needs a 
*sizeof(*p1).

> +        randomize(p2, BUF_SIZE);
> +        memcpy(p2_2, p2, BUF_SIZE);
> +        dshift = (rnd() & 0xF) + 1;
> +        dfactor = sign_extend(rnd(), 10);
> +        call_ref(p1, p2, BUF_SIZE, dshift, dfactor);
> +        call_new(p1_2, p2_2, BUF_SIZE, dshift, dfactor);

This function only mutates p1, not p2, so the copy of p2_2 is not entirely 
necessary.

> +
> +        if (memcmp(p2, p2_2, BUF_SIZE) != 0){

As we're mutating p1, that one is the one that should be checked. Also 
*sizeof(*p1) in the memcmp. And space between ) and {.

// Martin

_______________________________________________
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".