[FFmpeg-devel] [PATCH v3] libavfilter/x86/vf_convolution: add sobel filter optimization and unit test with intel AVX512 VNNI
Wang, Bin
bin.wang at intel.com
Wed Sep 7 15:33:06 EEST 2022
-----Original Message-----
From: ffmpeg-devel <ffmpeg-devel-bounces at ffmpeg.org> On Behalf Of Henrik Gramner
Sent: Tuesday, September 6, 2022 9:02 PM
To: FFmpeg development discussions and patches <ffmpeg-devel at ffmpeg.org>
Subject: Re: [FFmpeg-devel] [PATCH v3] libavfilter/x86/vf_convolution: add sobel filter optimization and unit test with intel AVX512 VNNI
On Tue, Aug 23, 2022 at 10:43 AM <bin.wang-at-intel.com at ffmpeg.org> wrote:
> +.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
> +
> [...]
> +
> + ;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
This could probably be improved by doing something along the lines of (untested, but you get the idea):
ALIGN 64
sobel_perm: db 0, 16, 32, 48, 1, 17, 33, 49, 2, 18, 34, 50, 3, 19, 35, 51
db 4, 20, 36, 52, 5, 21, 37, 53, 6, 22, 38, 54, 7, 23, 39, 55
db 8, 24, 40, 56, 9, 25, 41, 57, 10, 26, 42, 58, 11, 27, 43, 59
db 12, 28, 44, 60, 13, 29, 45, 61, 14, 30, 46, 62, 15, 31, 47, 63
sobel_mulA: db -1, 1, -2, 2
sobel_mulB: db 1, -1, 2, -2
[...]
mova m6, [sobel_perm]
.loop1:
movu xm3, [c2q + xq]
pmovzxbd m5, [c0q + xq]
vinserti32x4 ym3, [c6q + xq], 1
pmovzxbd m4, [c8q + xq]
vinserti32x4 m2, m3, [c1q + xq], 2
vinserti32x4 m3, [c5q + xq], 2
vinserti32x4 m2, [c7q + xq], 3
vinserti32x4 m3, [c3q + xq], 3
vpermb m2, m6, m2
psubd m4, m5
vpermb m3, m6, m3
mova m5, m4
vpdpbusd m4, m2, [sobel_mulA] {1to16}
vpdpbusd m5, m3, [sobel_mulB] {1to16}
> + mulps m4, m0 ; sum *= scale
> + addps m4, m1 ; sum += delta
fmaddps m4, m4, m0, m1
> + vpmovusdb xmm4, m4
> + movdqu [dstq + xq], xmm4
vpmovusdb [dstq + xq], m4
> + mulss xmm5, xmm5 ; b1 * b1
> + addss xmm4, xmm5
fmaddss xmm4, xmm5, xmm5, xmm4
> + mulss xmm4, xmm0 ; sum *= rdiv
> + addss xmm4, xmm1 ; sum += bias
fmaddss xmm4, xmm4, xmm0, xmm1
> +
> + 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
> +
> + cvtdq2ps m5, m5
> + VFMADD231PS m4, m5, m5
> +
> + sqrtps m4, 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 rd, rd
> + pxor m4, m4
> +
> + ;Gx
> + SOBEL_MUL 0, data_n1
> + SOBEL_MUL 1, data_n2
> + SOBEL_MUL 2, data_n1
> + SOBEL_ADD 6
> + SOBEL_MUL 7, data_p2
> + SOBEL_ADD 8
> +
> + cvtsi2ss xmm4, rd
> + mulss xmm4, xmm4
> +
> + xor rd, rd
> + ;Gy
> + SOBEL_MUL 0, data_n1
> + SOBEL_ADD 2
> + SOBEL_MUL 3, data_n2
> + SOBEL_MUL 5, data_p2
> + SOBEL_MUL 6, data_n1
> + SOBEL_ADD 8
> +
> + cvtsi2ss xmm5, rd
> + mulss xmm5, xmm5 ; b1 * b1
> + addss xmm4, xmm5
> +
> + sqrtps xmm4, xmm4
> + mulss xmm4, xmm0 ; sum *= rdiv
> + addss xmm4, xmm1 ; sum += bias
> + cvttps2dq xmm4, xmm4 ; trunc to integer
> + packssdw xmm4, xmm4
> + packuswb xmm4, xmm4
> + movd rd, xmm4
> + mov [dstq + xq], rb
> +
> + add xq, 1
> + cmp xq, widthq
> + jl .loop2
> +.end:
> + RET
> +%endmacro
> +
> +%if ARCH_X86_64
> +%if HAVE_AVX512ICL_EXTERNAL
> +INIT_ZMM avx512icl
> +FILTER_SOBEL
> +%endif
> +%endif
> diff --git a/libavfilter/x86/vf_convolution_init.c
> b/libavfilter/x86/vf_convolution_init.c
> index b78a47d02b..52a3d28991 100644
> --- a/libavfilter/x86/vf_convolution_init.c
> +++ b/libavfilter/x86/vf_convolution_init.c
> @@ -29,6 +29,11 @@ void ff_filter_3x3_sse4(uint8_t *dst, int width,
> const uint8_t *c[], int peak, int radius,
> int dstride, int stride, int size);
>
> +void ff_filter_sobel_avx512icl(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, int size);
> +
> av_cold void ff_convolution_init_x86(ConvolutionContext *s) { #if
> ARCH_X86_64 @@ -44,3 +49,16 @@ av_cold void
> ff_convolution_init_x86(ConvolutionContext *s)
> }
> #endif
> }
> +
> +av_cold void ff_sobel_init_x86(ConvolutionContext *s) { #if
> +ARCH_X86_64
> + int cpu_flags = av_get_cpu_flags();
> + for (int i = 0; i < s->nb_planes; i++) {
> + if (s->depth == 8) {
> + if (EXTERNAL_AVX512ICL(cpu_flags))
> + s->filter[i] = ff_filter_sobel_avx512icl;
> + }
> + }
> +#endif
> +}
> diff --git a/tests/checkasm/Makefile b/tests/checkasm/Makefile index
> 1ac170491b..4e91547fde 100644
> --- a/tests/checkasm/Makefile
> +++ b/tests/checkasm/Makefile
> @@ -44,6 +44,7 @@ AVFILTEROBJS-$(CONFIG_GBLUR_FILTER) += vf_gblur.o
> AVFILTEROBJS-$(CONFIG_HFLIP_FILTER) += vf_hflip.o
> AVFILTEROBJS-$(CONFIG_THRESHOLD_FILTER) += vf_threshold.o
> AVFILTEROBJS-$(CONFIG_NLMEANS_FILTER) += vf_nlmeans.o
> +AVFILTEROBJS-$(CONFIG_CONVOLUTION_FILTER) += vf_convolution.o
>
> CHECKASMOBJS-$(CONFIG_AVFILTER) += $(AVFILTEROBJS-yes)
>
> diff --git a/tests/checkasm/checkasm.c b/tests/checkasm/checkasm.c
> index e56fd3850e..ae5e9d1143 100644
> --- a/tests/checkasm/checkasm.c
> +++ b/tests/checkasm/checkasm.c
> @@ -191,6 +191,9 @@ static const struct {
> #if CONFIG_THRESHOLD_FILTER
> { "vf_threshold", checkasm_check_vf_threshold },
> #endif
> + #if CONFIG_CONVOLUTION_FILTER
> + { "vf_convolution", checkasm_check_vf_convolution },
> + #endif
> #endif
> #if CONFIG_SWSCALE
> { "sw_gbrp", checkasm_check_sw_gbrp }, diff --git
> a/tests/checkasm/checkasm.h b/tests/checkasm/checkasm.h index
> d7645d3730..71c722cf77 100644
> --- a/tests/checkasm/checkasm.h
> +++ b/tests/checkasm/checkasm.h
> @@ -85,6 +85,7 @@ void checkasm_check_vf_eq(void); void
> checkasm_check_vf_gblur(void); void checkasm_check_vf_hflip(void);
> void checkasm_check_vf_threshold(void);
> +void checkasm_check_vf_convolution(void);
> void checkasm_check_vp8dsp(void);
> void checkasm_check_vp9dsp(void);
> void checkasm_check_videodsp(void);
> diff --git a/tests/checkasm/vf_convolution.c
> b/tests/checkasm/vf_convolution.c new file mode 100644 index
> 0000000000..a10da8b45e
> --- /dev/null
> +++ b/tests/checkasm/vf_convolution.c
> @@ -0,0 +1,103 @@
> +/*
> + * This file is part of FFmpeg.
> + *
> + * FFmpeg is free software; you can redistribute it and/or modify
> + * it under the terms of the GNU General Public License as published
> +by
> + * the Free Software Foundation; either version 2 of the License, or
> + * (at your option) any later version.
> + *
> + * FFmpeg is distributed in the hope that it will be useful,
> + * but WITHOUT ANY WARRANTY; without even the implied warranty of
> + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
> + * GNU General Public License for more details.
> + *
> + * You should have received a copy of the GNU General Public License
> +along
> + * with FFmpeg; if not, write to the Free Software Foundation, Inc.,
> + * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
> + */
> +
> +#include <string.h>
> +#include "checkasm.h"
> +#include "libavfilter/avfilter.h"
> +#include "libavfilter/convolution.h"
> +#include "libavutil/intreadwrite.h"
> +#include "libavutil/mem_internal.h"
> +
> +#define WIDTH 512
> +#define HEIGHT 512
> +#define SRC_STRIDE 512
> +#define PIXELS (WIDTH * HEIGHT)
> +
> +#define randomize_buffers(buf, size) \
> + do { \
> + int j; \
> + uint8_t *tmp_buf = (uint8_t *)buf;\
> + for (j = 0; j< size; j++) \
> + tmp_buf[j] = rnd() & 0xFF; \
> + } while (0)
> +
> +static void check_sobel(const char * report_name) {
> + LOCAL_ALIGNED_32(uint8_t, src, [PIXELS]);
> + LOCAL_ALIGNED_32(uint8_t, dst_ref, [PIXELS]);
> + LOCAL_ALIGNED_32(uint8_t, dst_new, [PIXELS]);
> + const int height = WIDTH;
> + const int width = HEIGHT;
> + const int stride = SRC_STRIDE;
> + const int dstride = SRC_STRIDE;
> + int mode = 0;
> + const uint8_t *c[49];
> + const int radius = 1;
> + const int bpc = 1;
> + const int step = mode == MATRIX_COLUMN ? 16 : 1;
> + const int slice_start = 0;
> + const int slice_end = height;
> + int y;
> + const int sizew = mode == MATRIX_COLUMN ? height : width;
> + float scale = 2;
> + float delta = 10;
> +
> + ConvolutionContext s;
> +
> + s.scale = scale;
> + s.delta = delta;
> + s.depth = 8;
> + s.nb_planes = 3;
> + ff_convolution_init(&s, "sobel");
> +
> + declare_func(void, 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, int size);
> +
> + memset(dst_ref, 0, PIXELS);
> + memset(dst_new, 0, PIXELS);
> + randomize_buffers(src, PIXELS);
> +
> + if (check_func(s.filter[0], "%s", report_name)) {
> + for (y = slice_start; y < slice_end; y += step) {
> + const int xoff = mode == MATRIX_COLUMN ? (y - slice_start) * bpc : radius * bpc;
> + const int yoff = mode == MATRIX_COLUMN ? radius * dstride
> + : 0;
> +
> + s.setup[0](radius, c, src, stride, radius, width, y, height, bpc);
> + call_ref(dst_ref + yoff + xoff, sizew - 2 * radius,
> + scale, delta, NULL, c, 0, radius,
> + dstride, stride, slice_end - step);
> + call_new(dst_new + yoff + xoff, sizew - 2 * radius,
> + scale, delta, NULL, c, 0, radius,
> + dstride, stride, slice_end - step);
> + if (memcmp(dst_ref + yoff + xoff, dst_new + yoff + xoff, slice_end - step))
> + fail();
> + bench_new(dst_new + yoff + xoff, sizew - 2 * radius,
> + scale, delta, NULL, c, 0, radius,
> + dstride, stride, slice_end - step);
> + if (mode != MATRIX_COLUMN)
> + dst_ref += dstride;
> + }
> + }
> +
> +}
> +
> +void checkasm_check_vf_convolution(void)
> +{
> + check_sobel("sobel");
> + report("convolution:sobel");
> +}
> --
> 2.27.0
>
> _______________________________________________
> ffmpeg-devel mailing list
> ffmpeg-devel at ffmpeg.org
> https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
>
> To unsubscribe, visit link above, or email
> ffmpeg-devel-request at ffmpeg.org with subject "unsubscribe".
Thanks for the great advice! After applying the optimized instructions, I got a nearly 13% improvement (from 2470 to 2136), patch v4 is here: https://patchwork.ffmpeg.org/project/ffmpeg/patch/20220907062641.23091-1-bin.wang@intel.com/
Thanks
Bin
More information about the ffmpeg-devel
mailing list