libavfilter/x86/vf_convolution: add sobel filter optimization and unit test with intel AVX512 VNNI

This commit enabled assembly code with intel AVX512 VNNI and added unit test for sobel filter

sobel_c: 4537
sobel_avx512icl 2136

Signed-off-by: bwang30 <bin.wang@intel.com>
Signed-off-by: Haihao Xiang <haihao.xiang@intel.com>

4 files changed, 252 insertions, 78 deletions

diff --git a/libavfilter/convolution.h b/libavfilter/convolution.h
index 88aabe9a20..e44bfb5da8 100644
--- a/libavfilter/convolution.h
+++ b/libavfilter/convolution.h
@@ -21,6 +21,7 @@
 #ifndef AVFILTER_CONVOLUTION_H
 #define AVFILTER_CONVOLUTION_H
 #include "avfilter.h"
+#include "libavutil/intreadwrite.h"
 
 enum MatrixMode {
     MATRIX_SQUARE,
@@ -61,4 +62,77 @@ typedef struct ConvolutionContext {
 } ConvolutionContext;
 
 void ff_convolution_init_x86(ConvolutionContext *s);
+void ff_sobel_init_x86(ConvolutionContext *s, int depth, int nb_planes);
+
+static void setup_3x3(int radius, const uint8_t *c[], const uint8_t *src, int stride,
+                      int x, int w, int y, int h, int bpc)
+{
+    int i;
+
+    for (i = 0; i < 9; i++) {
+        int xoff = FFABS(x + ((i % 3) - 1));
+        int yoff = FFABS(y + (i / 3) - 1);
+
+        xoff = xoff >= w ? 2 * w - 1 - xoff : xoff;
+        yoff = yoff >= h ? 2 * h - 1 - yoff : yoff;
+
+        c[i] = src + xoff * bpc + yoff * stride;
+    }
+}
+
+static void filter_sobel(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)
+{
+    const uint8_t *c0 = c[0], *c1 = c[1], *c2 = c[2];
+    const uint8_t *c3 = c[3], *c5 = c[5];
+    const uint8_t *c6 = c[6], *c7 = c[7], *c8 = c[8];
+    int x;
+
+    for (x = 0; x < width; x++) {
+        float suma = c0[x] * -1 + c1[x] * -2 + c2[x] * -1 +
+                     c6[x] *  1 + c7[x] *  2 + c8[x] *  1;
+        float sumb = c0[x] * -1 + c2[x] *  1 + c3[x] * -2 +
+                     c5[x] *  2 + c6[x] * -1 + c8[x] *  1;
+
+        dst[x] = av_clip_uint8(sqrtf(suma*suma + sumb*sumb) * scale + delta);
+    }
+}
+
+static void filter16_sobel(uint8_t *dstp, int width,
+                           float scale, float delta, const int *const matrix,
+                           const uint8_t *c[], int peak, int radius,
+                           int dstride, int stride, int size)
+{
+    uint16_t *dst = (uint16_t *)dstp;
+    int x;
+
+    for (x = 0; x < width; x++) {
+        float suma = AV_RN16A(&c[0][2 * x]) * -1 + AV_RN16A(&c[1][2 * x]) * -2 + AV_RN16A(&c[2][2 * x]) * -1 +
+                     AV_RN16A(&c[6][2 * x]) *  1 + AV_RN16A(&c[7][2 * x]) *  2 + AV_RN16A(&c[8][2 * x]) *  1;
+        float sumb = AV_RN16A(&c[0][2 * x]) * -1 + AV_RN16A(&c[2][2 * x]) *  1 + AV_RN16A(&c[3][2 * x]) * -2 +
+                     AV_RN16A(&c[5][2 * x]) *  2 + AV_RN16A(&c[6][2 * x]) * -1 + AV_RN16A(&c[8][2 * x]) *  1;
+
+        dst[x] = av_clip(sqrtf(suma*suma + sumb*sumb) * scale + delta, 0, peak);
+    }
+}
+
+static av_unused void ff_sobel_init(ConvolutionContext *s, int depth, int nb_planes)
+{
+    for (int i = 0; i < 4; i++) {
+        s->filter[i] = filter_sobel;
+        s->copy[i] = !((1 << i) & s->planes);
+        s->size[i] = 3;
+        s->setup[i] = setup_3x3;
+        s->rdiv[i] = s->scale;
+        s->bias[i] = s->delta;
+    }
+    if (s->depth > 8)
+        for (int i = 0; i < 4; i++)
+            s->filter[i] = filter16_sobel;
+#if ARCH_X86_64
+    ff_sobel_init_x86(s, depth, nb_planes);
+#endif
+}
 #endif
diff --git a/libavfilter/vf_convolution.c b/libavfilter/vf_convolution.c
index 9a9c099e6d..7762fa2a05 100644
--- a/libavfilter/vf_convolution.c
+++ b/libavfilter/vf_convolution.c
@@ -139,24 +139,6 @@ static void filter16_roberts(uint8_t *dstp, int width,
     }
 }
 
-static void filter16_sobel(uint8_t *dstp, int width,
-                           float scale, float delta, const int *const matrix,
-                           const uint8_t *c[], int peak, int radius,
-                           int dstride, int stride, int size)
-{
-    uint16_t *dst = (uint16_t *)dstp;
-    int x;
-
-    for (x = 0; x < width; x++) {
-        float suma = AV_RN16A(&c[0][2 * x]) * -1 + AV_RN16A(&c[1][2 * x]) * -2 + AV_RN16A(&c[2][2 * x]) * -1 +
-                     AV_RN16A(&c[6][2 * x]) *  1 + AV_RN16A(&c[7][2 * x]) *  2 + AV_RN16A(&c[8][2 * x]) *  1;
-        float sumb = AV_RN16A(&c[0][2 * x]) * -1 + AV_RN16A(&c[2][2 * x]) *  1 + AV_RN16A(&c[3][2 * x]) * -2 +
-                     AV_RN16A(&c[5][2 * x]) *  2 + AV_RN16A(&c[6][2 * x]) * -1 + AV_RN16A(&c[8][2 * x]) *  1;
-
-        dst[x] = av_clip(sqrtf(suma*suma + sumb*sumb) * scale + delta, 0, peak);
-    }
-}
-
 static void filter16_scharr(uint8_t *dstp, int width,
                             float scale, float delta, const int *const matrix,
                             const uint8_t *c[], int peak, int radius,
@@ -261,26 +243,6 @@ static void filter_roberts(uint8_t *dst, int width,
     }
 }
 
-static void filter_sobel(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)
-{
-    const uint8_t *c0 = c[0], *c1 = c[1], *c2 = c[2];
-    const uint8_t *c3 = c[3], *c5 = c[5];
-    const uint8_t *c6 = c[6], *c7 = c[7], *c8 = c[8];
-    int x;
-
-    for (x = 0; x < width; x++) {
-        float suma = c0[x] * -1 + c1[x] * -2 + c2[x] * -1 +
-                     c6[x] *  1 + c7[x] *  2 + c8[x] *  1;
-        float sumb = c0[x] * -1 + c2[x] *  1 + c3[x] * -2 +
-                     c5[x] *  2 + c6[x] * -1 + c8[x] *  1;
-
-        dst[x] = av_clip_uint8(sqrtf(suma*suma + sumb*sumb) * scale + delta);
-    }
-}
-
 static void filter_scharr(uint8_t *dst, int width,
                           float scale, float delta, const int *const matrix,
                           const uint8_t *c[], int peak, int radius,
@@ -552,22 +514,6 @@ static void filter_column(uint8_t *dst, int height,
     }
 }
 
-static void setup_3x3(int radius, const uint8_t *c[], const uint8_t *src, int stride,
-                      int x, int w, int y, int h, int bpc)
-{
-    int i;
-
-    for (i = 0; i < 9; i++) {
-        int xoff = FFABS(x + ((i % 3) - 1));
-        int yoff = FFABS(y + (i / 3) - 1);
-
-        xoff = xoff >= w ? 2 * w - 1 - xoff : xoff;
-        yoff = yoff >= h ? 2 * h - 1 - yoff : yoff;
-
-        c[i] = src + xoff * bpc + yoff * stride;
-    }
-}
-
 static void setup_5x5(int radius, const uint8_t *c[], const uint8_t *src, int stride,
                       int x, int w, int y, int h, int bpc)
 {
@@ -708,6 +654,18 @@ static int param_init(AVFilterContext *ctx)
     const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(inlink->format);
     int p, i;
 
+    s->depth = desc->comp[0].depth;
+    s->max = (1 << s->depth) - 1;
+
+    s->planewidth[1] = s->planewidth[2] = AV_CEIL_RSHIFT(inlink->w, desc->log2_chroma_w);
+    s->planewidth[0] = s->planewidth[3] = inlink->w;
+    s->planeheight[1] = s->planeheight[2] = AV_CEIL_RSHIFT(inlink->h, desc->log2_chroma_h);
+    s->planeheight[0] = s->planeheight[3] = inlink->h;
+
+    s->nb_planes = av_pix_fmt_count_planes(inlink->format);
+    s->nb_threads = ff_filter_get_nb_threads(ctx);
+    s->bpc = (s->depth + 7) / 8;
+
     if (!strcmp(ctx->filter->name, "convolution")) {
         for (i = 0; i < 4; i++) {
             int *matrix = (int *)s->matrix[i];
@@ -804,14 +762,7 @@ static int param_init(AVFilterContext *ctx)
             s->bias[i] = s->delta;
         }
     } else if (!strcmp(ctx->filter->name, "sobel")) {
-        for (i = 0; i < 4; i++) {
-            s->filter[i] = filter_sobel;
-            s->copy[i] = !((1 << i) & s->planes);
-            s->size[i] = 3;
-            s->setup[i] = setup_3x3;
-            s->rdiv[i] = s->scale;
-            s->bias[i] = s->delta;
-        }
+        ff_sobel_init(s, s->depth, s->nb_planes);
     } else if (!strcmp(ctx->filter->name, "kirsch")) {
         for (i = 0; i < 4; i++) {
             s->filter[i] = filter_kirsch;
@@ -832,18 +783,6 @@ static int param_init(AVFilterContext *ctx)
         }
     }
 
-    s->depth = desc->comp[0].depth;
-    s->max = (1 << s->depth) - 1;
-
-    s->planewidth[1] = s->planewidth[2] = AV_CEIL_RSHIFT(inlink->w, desc->log2_chroma_w);
-    s->planewidth[0] = s->planewidth[3] = inlink->w;
-    s->planeheight[1] = s->planeheight[2] = AV_CEIL_RSHIFT(inlink->h, desc->log2_chroma_h);
-    s->planeheight[0] = s->planeheight[3] = inlink->h;
-
-    s->nb_planes = av_pix_fmt_count_planes(inlink->format);
-    s->nb_threads = ff_filter_get_nb_threads(ctx);
-    s->bpc = (s->depth + 7) / 8;
-
     if (!strcmp(ctx->filter->name, "convolution")) {
         if (s->depth > 8) {
             for (p = 0; p < s->nb_planes; p++) {
@@ -870,10 +809,6 @@ static int param_init(AVFilterContext *ctx)
         if (s->depth > 8)
             for (p = 0; p < s->nb_planes; p++)
                 s->filter[p] = filter16_roberts;
-    } else if (!strcmp(ctx->filter->name, "sobel")) {
-        if (s->depth > 8)
-            for (p = 0; p < s->nb_planes; p++)
-                s->filter[p] = filter16_sobel;
     } else if (!strcmp(ctx->filter->name, "kirsch")) {
         if (s->depth > 8)
             for (p = 0; p < s->nb_planes; p++)
diff --git a/libavfilter/x86/vf_convolution.asm b/libavfilter/x86/vf_convolution.asm
index 754d4d1064..c912d56752 100644
--- a/libavfilter/x86/vf_convolution.asm
+++ b/libavfilter/x86/vf_convolution.asm
@@ -22,6 +22,18 @@
 
 SECTION_RODATA
 half:   dd 0.5
+data_p1: dd  1
+data_n1: dd -1
+data_p2: dd  2
+data_n2: dd -2
+
+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
 
 SECTION .text
 
@@ -154,3 +166,138 @@ 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 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
+
+    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}
+
+    cvtdq2ps  m4, m4
+    mulps     m4, m4
+
+    cvtdq2ps    m5, m5
+    VFMADD231PS m4, m5, m5
+
+    sqrtps    m4, m4
+    fmaddps m4, m4, m0, m1
+    cvttps2dq m4, m4
+    vpmovusdb [dstq + xq], m4
+
+    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
+    fmaddss xmm4, xmm5, xmm5, xmm4
+
+    sqrtps    xmm4, xmm4
+    fmaddss   xmm4, xmm4, xmm0, xmm1     ;sum = sum * rdiv + 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..bff10ca1a4 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, int depth, int nb_planes)
+{
+#if ARCH_X86_64
+    int cpu_flags = av_get_cpu_flags();
+    for (int i = 0; i < nb_planes; i++) {
+        if (depth == 8) {
+            if (EXTERNAL_AVX512ICL(cpu_flags))
+                s->filter[i] = ff_filter_sobel_avx512icl;
+        }
+    }
+#endif
+}