Re: [FFmpeg-devel] [PATCH] libavfilter/x86/vf_convolution: add sobel filter optimization and unit test with intel AVX512 VNNI
-Original Message-
From: ffmpeg-devel On Behalf Of Andreas
Rheinhardt
Sent: Friday, August 12, 2022 5:11 PM
To: ffmpeg-devel@ffmpeg.org
Subject: Re: [FFmpeg-devel] [PATCH] libavfilter/x86/vf_convolution: add sobel
filter optimization and unit test with intel AVX512 VNNI
bin.wang-at-intel@ffmpeg.org:
> From: bwang30
>
> This commit enabled assembly code with intel AVX512 VNNI and added
> unit test for sobel filter
>
> sobel_c: 4537
> sobel_avx512icl 2470
>
> Signed-off-by: bwang30
> ---
> libavfilter/convolution.h | 2 +
> libavfilter/vf_convolution.c | 8 ++
> libavfilter/x86/vf_convolution.asm| 162 ++
> libavfilter/x86/vf_convolution_init.c | 18 +++
> tests/checkasm/Makefile | 1 +
> tests/checkasm/checkasm.c | 3 +
> tests/checkasm/checkasm.h | 1 +
> tests/checkasm/vf_convolution.c | 116 ++
> 8 files changed, 311 insertions(+)
> create mode 100644 tests/checkasm/vf_convolution.c
>
> diff --git a/libavfilter/convolution.h b/libavfilter/convolution.h
> index 88aabe9a20..143b0fb2d9 100644
> --- a/libavfilter/convolution.h
> +++ b/libavfilter/convolution.h
> @@ -61,4 +61,6 @@ typedef struct ConvolutionContext { }
> ConvolutionContext;
>
> void ff_convolution_init_x86(ConvolutionContext *s);
> +void ff_sobel_init_x86(ConvolutionContext *s); int
> +ff_filter_param_init(AVFilterContext *ctx);
> #endif
> diff --git a/libavfilter/vf_convolution.c
> b/libavfilter/vf_convolution.c index 9a9c099e6d..98aa952258 100644
> --- a/libavfilter/vf_convolution.c
> +++ b/libavfilter/vf_convolution.c
> @@ -874,6 +874,9 @@ static int param_init(AVFilterContext *ctx)
> if (s->depth > 8)
> for (p = 0; p < s->nb_planes; p++)
> s->filter[p] = filter16_sobel;
> +#if CONFIG_CONVOLUTION_FILTER && ARCH_X86_64
> +ff_sobel_init_x86(s);
> +#endif
> } else if (!strcmp(ctx->filter->name, "kirsch")) {
> if (s->depth > 8)
> for (p = 0; p < s->nb_planes; p++) @@ -887,6 +890,11 @@
> static int param_init(AVFilterContext *ctx)
> return 0;
> }
>
> +int ff_filter_param_init(AVFilterContext *ctx) {
> +return param_init(ctx);
> +}
> +
> static int config_input(AVFilterLink *inlink) {
> AVFilterContext *ctx = inlink->dst; diff --git
> a/libavfilter/x86/vf_convolution.asm
> b/libavfilter/x86/vf_convolution.asm
> index 754d4d1064..59c807b218 100644
> --- a/libavfilter/x86/vf_convolution.asm
> +++ b/libavfilter/x86/vf_convolution.asm
> @@ -22,6 +22,10 @@
>
> SECTION_RODATA
> half: dd 0.5
> +data_p1: dd 1
> +data_n1: dd -1
> +data_p2: dd 2
> +data_n2: dd -2
>
> SECTION .text
>
> @@ -154,3 +158,161 @@ cglobal filter_3x3, 4, 15, 7, dst, width, rdiv,
> bias, matrix, ptr, c0, c1, c2, c INIT_XMM sse4
> FILTER_3X3
> %endif
> +
> +
> +%macro SOBEL_MUL_16 3
> +movd xmm2, [%2]
> +VPBROADCASTD m2, xmm2
> +movdqu xmm3, [c%1q + xq]
> +vpmovzxbd m3, xmm3
> +vpdpbusd m%3, m3, m2
> +%endmacro
> +
> +%macro SOBEL_ADD_16 2
> +movdqu xmm3, [c%1q + xq]
> +vpmovzxbd m3, xmm3
> +vpaddd m%2, m3
> +%endmacro
> +
> +
> +%macro SOBEL_MUL 2
> +movzx ptrd, byte [c%1q + xq]
> +imul ptrd, [%2]
> +add rd, ptrd
> +%endmacro
> +
> +%macro SOBEL_ADD 1
> +movzx ptrd, byte [c%1q + xq]
> +add rd, ptrd
> +%endmacro
> +
> +; void filter_sobel_avx512(uint8_t *dst, int width,
> +; float scale, float delta, const int *const matrix,
> +; const uint8_t *c[], int peak, int radius,
> +; int dstride, int stride)
> +%macro FILTER_SOBEL 0
> +%if UNIX64
> +cglobal filter_sobel, 4, 15, 7, dst, width, matrix, ptr, c0, c1, c2,
> +c3, c4, c5, c6, c7, c8, r, x %else cglobal filter_sobel, 4, 15, 7,
> +dst, width, rdiv, bias, matrix, ptr, c0, c1, c2, c3, c4, c5, c6, c7,
> +c8, r, x %endif %if WIN64
> +SWAP xmm0, xmm2
> +SWAP xmm1, xmm3
> +mov r2q, matrixmp
> +mov r3q, ptrmp
> +DEFINE_ARGS dst, width, matrix, ptr, c0, c1, c2, c3, c4, c5, c6,
> +c7, c8, r, x %endif
> +movsxdifnidn widthq, widthd
> +VBROADCASTSS m0, xmm0
> +VBROADCASTSS m1, xmm1
> +pxor m6, m6
> +mov c0q, [ptrq + 0*gprsize]
> +mov c1q, [ptrq + 1*gprsize]
> +mov c2q, [ptrq + 2*gprsize]
> +mov c3q, [ptrq + 3*gprsize]
> +mov c4q, [ptrq + 4*gprsize]
> +mov c5q, [ptrq + 5*gprsize]
> +mov c6q, [ptrq + 6*gprsize]
> +mov c7q, [pt
Re: [FFmpeg-devel] [PATCH] libavfilter/x86/vf_convolution: add sobel filter optimization and unit test with intel AVX512 VNNI
bin.wang-at-intel@ffmpeg.org:
> From: bwang30
>
> This commit enabled assembly code with intel AVX512 VNNI and added unit test
> for sobel filter
>
> sobel_c: 4537
> sobel_avx512icl 2470
>
> Signed-off-by: bwang30
> ---
> libavfilter/convolution.h | 2 +
> libavfilter/vf_convolution.c | 8 ++
> libavfilter/x86/vf_convolution.asm| 162 ++
> libavfilter/x86/vf_convolution_init.c | 18 +++
> tests/checkasm/Makefile | 1 +
> tests/checkasm/checkasm.c | 3 +
> tests/checkasm/checkasm.h | 1 +
> tests/checkasm/vf_convolution.c | 116 ++
> 8 files changed, 311 insertions(+)
> create mode 100644 tests/checkasm/vf_convolution.c
>
> diff --git a/libavfilter/convolution.h b/libavfilter/convolution.h
> index 88aabe9a20..143b0fb2d9 100644
> --- a/libavfilter/convolution.h
> +++ b/libavfilter/convolution.h
> @@ -61,4 +61,6 @@ typedef struct ConvolutionContext {
> } ConvolutionContext;
>
> void ff_convolution_init_x86(ConvolutionContext *s);
> +void ff_sobel_init_x86(ConvolutionContext *s);
> +int ff_filter_param_init(AVFilterContext *ctx);
> #endif
> diff --git a/libavfilter/vf_convolution.c b/libavfilter/vf_convolution.c
> index 9a9c099e6d..98aa952258 100644
> --- a/libavfilter/vf_convolution.c
> +++ b/libavfilter/vf_convolution.c
> @@ -874,6 +874,9 @@ static int param_init(AVFilterContext *ctx)
> if (s->depth > 8)
> for (p = 0; p < s->nb_planes; p++)
> s->filter[p] = filter16_sobel;
> +#if CONFIG_CONVOLUTION_FILTER && ARCH_X86_64
> +ff_sobel_init_x86(s);
> +#endif
> } else if (!strcmp(ctx->filter->name, "kirsch")) {
> if (s->depth > 8)
> for (p = 0; p < s->nb_planes; p++)
> @@ -887,6 +890,11 @@ static int param_init(AVFilterContext *ctx)
> return 0;
> }
>
> +int ff_filter_param_init(AVFilterContext *ctx)
> +{
> +return param_init(ctx);
> +}
> +
> static int config_input(AVFilterLink *inlink)
> {
> AVFilterContext *ctx = inlink->dst;
> diff --git a/libavfilter/x86/vf_convolution.asm
> b/libavfilter/x86/vf_convolution.asm
> index 754d4d1064..59c807b218 100644
> --- a/libavfilter/x86/vf_convolution.asm
> +++ b/libavfilter/x86/vf_convolution.asm
> @@ -22,6 +22,10 @@
>
> SECTION_RODATA
> half: dd 0.5
> +data_p1: dd 1
> +data_n1: dd -1
> +data_p2: dd 2
> +data_n2: dd -2
>
> SECTION .text
>
> @@ -154,3 +158,161 @@ cglobal filter_3x3, 4, 15, 7, dst, width, rdiv, bias,
> matrix, ptr, c0, c1, c2, c
> INIT_XMM sse4
> FILTER_3X3
> %endif
> +
> +
> +%macro SOBEL_MUL_16 3
> +movd xmm2, [%2]
> +VPBROADCASTD m2, xmm2
> +movdqu xmm3, [c%1q + xq]
> +vpmovzxbd m3, xmm3
> +vpdpbusd m%3, m3, m2
> +%endmacro
> +
> +%macro SOBEL_ADD_16 2
> +movdqu xmm3, [c%1q + xq]
> +vpmovzxbd m3, xmm3
> +vpaddd m%2, m3
> +%endmacro
> +
> +
> +%macro SOBEL_MUL 2
> +movzx ptrd, byte [c%1q + xq]
> +imul ptrd, [%2]
> +add rd, ptrd
> +%endmacro
> +
> +%macro SOBEL_ADD 1
> +movzx ptrd, byte [c%1q + xq]
> +add rd, ptrd
> +%endmacro
> +
> +; void filter_sobel_avx512(uint8_t *dst, int width,
> +; float scale, float delta, const int *const matrix,
> +; const uint8_t *c[], int peak, int radius,
> +; int dstride, int stride)
> +%macro FILTER_SOBEL 0
> +%if UNIX64
> +cglobal filter_sobel, 4, 15, 7, dst, width, matrix, ptr, c0, c1, c2, c3, c4,
> c5, c6, c7, c8, r, x
> +%else
> +cglobal filter_sobel, 4, 15, 7, dst, width, rdiv, bias, matrix, ptr, c0, c1,
> c2, c3, c4, c5, c6, c7, c8, r, x
> +%endif
> +%if WIN64
> +SWAP xmm0, xmm2
> +SWAP xmm1, xmm3
> +mov r2q, matrixmp
> +mov r3q, ptrmp
> +DEFINE_ARGS dst, width, matrix, ptr, c0, c1, c2, c3, c4, c5, c6, c7, c8,
> r, x
> +%endif
> +movsxdifnidn widthq, widthd
> +VBROADCASTSS m0, xmm0
> +VBROADCASTSS m1, xmm1
> +pxor m6, m6
> +mov c0q, [ptrq + 0*gprsize]
> +mov c1q, [ptrq + 1*gprsize]
> +mov c2q, [ptrq + 2*gprsize]
> +mov c3q, [ptrq + 3*gprsize]
> +mov c4q, [ptrq + 4*gprsize]
> +mov c5q, [ptrq + 5*gprsize]
> +mov c6q, [ptrq + 6*gprsize]
> +mov c7q, [ptrq + 7*gprsize]
> +mov c8q, [ptrq + 8*gprsize]
> +
> +xor xq, xq
> +cmp widthq, mmsize/4
> +jl .loop2
> +
> +mov rq, widthq
> +and rq, mmsize/4-1
> +sub widthq, rq
> +
> +.loop1:
> +pxor m4, m4
> +pxor m5, m5
> +
> +;Gx
> +SOBEL_MUL_16 0, data_n1, 4
> +SOBEL_MUL_16 1, data_n2, 4
> +SOBEL_MUL_16 2, data_n1, 4
> +SOBEL_ADD_16 6, 4
> +SOBEL_MUL_16 7, data_p2, 4
> +SOBEL_ADD_16 8, 4
> +
> +cvtdq2ps m4, m4
> +mulps m4, m4
> +
> +;Gy
> +SOBEL_MUL_16 0, data_n1, 5
> +SOBEL_ADD_16 2, 5
> +SOBEL_MUL_16 3, data_n2, 5
> +SOBEL_MUL_16 5, data_p2, 5
> +SOBEL_MUL_16 6, data_n1, 5
> +SOBEL_ADD_16 8, 5
>
[FFmpeg-devel] [PATCH] libavfilter/x86/vf_convolution: add sobel filter optimization and unit test with intel AVX512 VNNI
From: bwang30
This commit enabled assembly code with intel AVX512 VNNI and added unit test
for sobel filter
sobel_c: 4537
sobel_avx512icl 2470
Signed-off-by: bwang30
---
libavfilter/convolution.h | 2 +
libavfilter/vf_convolution.c | 8 ++
libavfilter/x86/vf_convolution.asm| 162 ++
libavfilter/x86/vf_convolution_init.c | 18 +++
tests/checkasm/Makefile | 1 +
tests/checkasm/checkasm.c | 3 +
tests/checkasm/checkasm.h | 1 +
tests/checkasm/vf_convolution.c | 116 ++
8 files changed, 311 insertions(+)
create mode 100644 tests/checkasm/vf_convolution.c
diff --git a/libavfilter/convolution.h b/libavfilter/convolution.h
index 88aabe9a20..143b0fb2d9 100644
--- a/libavfilter/convolution.h
+++ b/libavfilter/convolution.h
@@ -61,4 +61,6 @@ typedef struct ConvolutionContext {
} ConvolutionContext;
void ff_convolution_init_x86(ConvolutionContext *s);
+void ff_sobel_init_x86(ConvolutionContext *s);
+int ff_filter_param_init(AVFilterContext *ctx);
#endif
diff --git a/libavfilter/vf_convolution.c b/libavfilter/vf_convolution.c
index 9a9c099e6d..98aa952258 100644
--- a/libavfilter/vf_convolution.c
+++ b/libavfilter/vf_convolution.c
@@ -874,6 +874,9 @@ static int param_init(AVFilterContext *ctx)
if (s->depth > 8)
for (p = 0; p < s->nb_planes; p++)
s->filter[p] = filter16_sobel;
+#if CONFIG_CONVOLUTION_FILTER && ARCH_X86_64
+ff_sobel_init_x86(s);
+#endif
} else if (!strcmp(ctx->filter->name, "kirsch")) {
if (s->depth > 8)
for (p = 0; p < s->nb_planes; p++)
@@ -887,6 +890,11 @@ static int param_init(AVFilterContext *ctx)
return 0;
}
+int ff_filter_param_init(AVFilterContext *ctx)
+{
+return param_init(ctx);
+}
+
static int config_input(AVFilterLink *inlink)
{
AVFilterContext *ctx = inlink->dst;
diff --git a/libavfilter/x86/vf_convolution.asm
b/libavfilter/x86/vf_convolution.asm
index 754d4d1064..59c807b218 100644
--- a/libavfilter/x86/vf_convolution.asm
+++ b/libavfilter/x86/vf_convolution.asm
@@ -22,6 +22,10 @@
SECTION_RODATA
half: dd 0.5
+data_p1: dd 1
+data_n1: dd -1
+data_p2: dd 2
+data_n2: dd -2
SECTION .text
@@ -154,3 +158,161 @@ cglobal filter_3x3, 4, 15, 7, dst, width, rdiv, bias,
matrix, ptr, c0, c1, c2, c
INIT_XMM sse4
FILTER_3X3
%endif
+
+
+%macro SOBEL_MUL_16 3
+movd xmm2, [%2]
+VPBROADCASTD m2, xmm2
+movdqu xmm3, [c%1q + xq]
+vpmovzxbd m3, xmm3
+vpdpbusd m%3, m3, m2
+%endmacro
+
+%macro SOBEL_ADD_16 2
+movdqu xmm3, [c%1q + xq]
+vpmovzxbd m3, xmm3
+vpaddd m%2, m3
+%endmacro
+
+
+%macro SOBEL_MUL 2
+movzx ptrd, byte [c%1q + xq]
+imul ptrd, [%2]
+add rd, ptrd
+%endmacro
+
+%macro SOBEL_ADD 1
+movzx ptrd, byte [c%1q + xq]
+add rd, ptrd
+%endmacro
+
+; void filter_sobel_avx512(uint8_t *dst, int width,
+; float scale, float delta, const int *const matrix,
+; const uint8_t *c[], int peak, int radius,
+; int dstride, int stride)
+%macro FILTER_SOBEL 0
+%if UNIX64
+cglobal filter_sobel, 4, 15, 7, dst, width, matrix, ptr, c0, c1, c2, c3, c4,
c5, c6, c7, c8, r, x
+%else
+cglobal filter_sobel, 4, 15, 7, dst, width, rdiv, bias, matrix, ptr, c0, c1,
c2, c3, c4, c5, c6, c7, c8, r, x
+%endif
+%if WIN64
+SWAP xmm0, xmm2
+SWAP xmm1, xmm3
+mov r2q, matrixmp
+mov r3q, ptrmp
+DEFINE_ARGS dst, width, matrix, ptr, c0, c1, c2, c3, c4, c5, c6, c7, c8,
r, x
+%endif
+movsxdifnidn widthq, widthd
+VBROADCASTSS m0, xmm0
+VBROADCASTSS m1, xmm1
+pxor m6, m6
+mov c0q, [ptrq + 0*gprsize]
+mov c1q, [ptrq + 1*gprsize]
+mov c2q, [ptrq + 2*gprsize]
+mov c3q, [ptrq + 3*gprsize]
+mov c4q, [ptrq + 4*gprsize]
+mov c5q, [ptrq + 5*gprsize]
+mov c6q, [ptrq + 6*gprsize]
+mov c7q, [ptrq + 7*gprsize]
+mov c8q, [ptrq + 8*gprsize]
+
+xor xq, xq
+cmp widthq, mmsize/4
+jl .loop2
+
+mov rq, widthq
+and rq, mmsize/4-1
+sub widthq, rq
+
+.loop1:
+pxor m4, m4
+pxor m5, m5
+
+;Gx
+SOBEL_MUL_16 0, data_n1, 4
+SOBEL_MUL_16 1, data_n2, 4
+SOBEL_MUL_16 2, data_n1, 4
+SOBEL_ADD_16 6, 4
+SOBEL_MUL_16 7, data_p2, 4
+SOBEL_ADD_16 8, 4
+
+cvtdq2ps m4, m4
+mulps m4, m4
+
+;Gy
+SOBEL_MUL_16 0, data_n1, 5
+SOBEL_ADD_16 2, 5
+SOBEL_MUL_16 3, data_n2, 5
+SOBEL_MUL_16 5, data_p2, 5
+SOBEL_MUL_16 6, data_n1, 5
+SOBEL_ADD_16 8, 5
+
+cvtdq2psm5, m5
+VFMADD231PS m4, m5, m5
+
+sqrtpsm4, m4
+mulps m4, m0 ; sum *= scale
+addps m4, m1 ; sum += delta
+cvttps2dq m4, m4
+vpmovusdb xmm4, m4
+movdqu[dstq + xq], xmm4
+
+add xq, mmsize/4
+cmp xq, widthq
+jl .loop1
+
+add widthq, rq
+cmp xq, widthq
+jge .end
+
+.loop2:
+xor
