avfilter/vf_gblur: add x86 SIMD optimizations

The horizontal pass get ~2x performance with the patch
under single thread.

Tested overall performance using the command(avx2 enabled):
./ffmpeg -i 1080p.mp4 -vf gblur -f null /dev/null
./ffmpeg -i 1080p.mp4 -vf gblur=threads=1 -f null /dev/null
For single thread, the fps improves from 43 to 60, about 40%.
For multi-thread, the fps improves from 110 to 130, about 20%.
Signed-off-by: Ruiling Song <ruiling.song@intel.com>

libavfilter/gblur.h

+/*
+ * Copyright (c) 2011 Pascal Getreuer
+ * Copyright (c) 2016 Paul B Mahol
+ *
+ * Redistribution and use in source and binary forms, with or without modification,
+ * are permitted provided that the following conditions are met:
+ *
+ *  * Redistributions of source code must retain the above copyright
+ *    notice, this list of conditions and the following disclaimer.
+ *  * Redistributions in binary form must reproduce the above
+ *    copyright notice, this list of conditions and the following
+ *    disclaimer in the documentation and/or other materials provided
+ *    with the distribution.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ * HOLDER BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+ * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+ * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+ * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+ * LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+ * NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+ * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+#ifndef AVFILTER_GBLUR_H
+#define AVFILTER_GBLUR_H
+#include "avfilter.h"
+
+typedef struct GBlurContext {
+    const AVClass *class;
+
+    float sigma;
+    float sigmaV;
+    int steps;
+    int planes;
+
+    int depth;
+    int planewidth[4];
+    int planeheight[4];
+    float *buffer;
+    float boundaryscale;
+    float boundaryscaleV;
+    float postscale;
+    float postscaleV;
+    float nu;
+    float nuV;
+    int nb_planes;
+    void (*horiz_slice)(float *buffer, int width, int height, int steps, float nu, float bscale);
+} GBlurContext;
+void ff_gblur_init(GBlurContext *s);
+void ff_gblur_init_x86(GBlurContext *s);
+#endif

libavfilter/vf_gblur.c

@@ -30,30 +30,10 @@
 #include "libavutil/pixdesc.h"
 #include "avfilter.h"
 #include "formats.h"
+#include "gblur.h"
 #include "internal.h"
 #include "video.h"
 
-typedef struct GBlurContext {
-    const AVClass *class;
-
-    float sigma;
-    float sigmaV;
-    int steps;
-    int planes;
-
-    int depth;
-    int planewidth[4];
-    int planeheight[4];
-    float *buffer;
-    float boundaryscale;
-    float boundaryscaleV;
-    float postscale;
-    float postscaleV;
-    float nu;
-    float nuV;
-    int nb_planes;
-} GBlurContext;
-
 #define OFFSET(x) offsetof(GBlurContext, x)
 #define FLAGS AV_OPT_FLAG_VIDEO_PARAM|AV_OPT_FLAG_FILTERING_PARAM
 
@@ -72,39 +52,44 @@ typedef struct ThreadData {
     int width;
 } ThreadData;
 
-static int filter_horizontally(AVFilterContext *ctx, void *arg, int jobnr, int nb_jobs)
+static void horiz_slice_c(float *buffer, int width, int height, int steps,
+                          float nu, float bscale)
 {
-    GBlurContext *s = ctx->priv;
-    ThreadData *td = arg;
-    const int height = td->height;
-    const int width = td->width;
-    const int slice_start = (height *  jobnr   ) / nb_jobs;
-    const int slice_end   = (height * (jobnr+1)) / nb_jobs;
-    const float boundaryscale = s->boundaryscale;
-    const int steps = s->steps;
-    const float nu = s->nu;
-    float *buffer = s->buffer;
-    int y, x, step;
+    int step, x, y;
     float *ptr;
-
-    /* Filter horizontally along each row */
-    for (y = slice_start; y < slice_end; y++) {
+    for (y = 0; y < height; y++) {
         for (step = 0; step < steps; step++) {
             ptr = buffer + width * y;
-            ptr[0] *= boundaryscale;
+            ptr[0] *= bscale;
 
             /* Filter rightwards */
             for (x = 1; x < width; x++)
                 ptr[x] += nu * ptr[x - 1];
-
-            ptr[x = width - 1] *= boundaryscale;
+            ptr[x = width - 1] *= bscale;
 
             /* Filter leftwards */
             for (; x > 0; x--)
                 ptr[x - 1] += nu * ptr[x];
         }
     }
+}
 
+static int filter_horizontally(AVFilterContext *ctx, void *arg, int jobnr, int nb_jobs)
+{
+    GBlurContext *s = ctx->priv;
+    ThreadData *td = arg;
+    const int height = td->height;
+    const int width = td->width;
+    const int slice_start = (height *  jobnr   ) / nb_jobs;
+    const int slice_end   = (height * (jobnr+1)) / nb_jobs;
+    const float boundaryscale = s->boundaryscale;
+    const int steps = s->steps;
+    const float nu = s->nu;
+    float *buffer = s->buffer;
+
+    s->horiz_slice(buffer + width * slice_start, width, slice_end - slice_start,
+                   steps, nu, boundaryscale);
+    emms_c();
     return 0;
 }
 
@@ -231,6 +216,13 @@ static int query_formats(AVFilterContext *ctx)
     return ff_set_common_formats(ctx, ff_make_format_list(pix_fmts));
 }
 
+void ff_gblur_init(GBlurContext *s)
+{
+    s->horiz_slice = horiz_slice_c;
+    if (ARCH_X86_64)
+        ff_gblur_init_x86(s);
+}
+
 static int config_input(AVFilterLink *inlink)
 {
     const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(inlink->format);
@@ -251,6 +243,7 @@ static int config_input(AVFilterLink *inlink)
     if (s->sigmaV < 0) {
         s->sigmaV = s->sigma;
     }
+    ff_gblur_init(s);
 
     return 0;
 }

libavfilter/x86/Makefile

@@ -7,6 +7,7 @@ OBJS-$(CONFIG_BWDIF_FILTER)                  += x86/vf_bwdif_init.o
 OBJS-$(CONFIG_COLORSPACE_FILTER)             += x86/colorspacedsp_init.o
 OBJS-$(CONFIG_EQ_FILTER)                     += x86/vf_eq.o
 OBJS-$(CONFIG_FSPP_FILTER)                   += x86/vf_fspp_init.o
+OBJS-$(CONFIG_GBLUR_FILTER)                  += x86/vf_gblur_init.o
 OBJS-$(CONFIG_GRADFUN_FILTER)                += x86/vf_gradfun_init.o
 OBJS-$(CONFIG_FRAMERATE_FILTER)              += x86/vf_framerate_init.o
 OBJS-$(CONFIG_HFLIP_FILTER)                  += x86/vf_hflip_init.o
@@ -41,6 +42,7 @@ X86ASM-OBJS-$(CONFIG_BWDIF_FILTER)           += x86/vf_bwdif.o
 X86ASM-OBJS-$(CONFIG_COLORSPACE_FILTER)      += x86/colorspacedsp.o
 X86ASM-OBJS-$(CONFIG_FRAMERATE_FILTER)       += x86/vf_framerate.o
 X86ASM-OBJS-$(CONFIG_FSPP_FILTER)            += x86/vf_fspp.o
+X86ASM-OBJS-$(CONFIG_GBLUR_FILTER)           += x86/vf_gblur.o
 X86ASM-OBJS-$(CONFIG_GRADFUN_FILTER)         += x86/vf_gradfun.o
 X86ASM-OBJS-$(CONFIG_HFLIP_FILTER)           += x86/vf_hflip.o
 X86ASM-OBJS-$(CONFIG_HQDN3D_FILTER)          += x86/vf_hqdn3d.o

libavfilter/x86/vf_gblur.asm

+;*****************************************************************************
+;* x86-optimized functions for gblur filter
+;*
+;* This file is part of FFmpeg.
+;*
+;* FFmpeg is free software; you can redistribute it and/or
+;* modify it under the terms of the GNU Lesser General Public
+;* License as published by the Free Software Foundation; either
+;* version 2.1 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
+;* Lesser General Public License for more details.
+;*
+;* You should have received a copy of the GNU Lesser 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 "libavutil/x86/x86util.asm"
+
+SECTION .text
+
+; void ff_horiz_slice_sse4(float *ptr, int width, int height, int steps,
+;                          float nu, float bscale)
+
+%macro HORIZ_SLICE 0
+%if UNIX64
+cglobal horiz_slice, 4, 9, 9, ptr, width, height, steps, x, y, step, stride, remain
+%else
+cglobal horiz_slice, 4, 9, 9, ptr, width, height, steps, nu, bscale, x, y, step, stride, remain
+%endif
+%if WIN64
+    movss m0, num
+    movss m1, bscalem
+    DEFINE_ARGS ptr, width, height, steps, x, y, step, stride, remain
+%endif
+    movsxdifnidn widthq, widthd
+
+    mulss m2, m0, m0 ; nu ^ 2
+    mulss m3, m2, m0 ; nu ^ 3
+    mulss m4, m3, m0 ; nu ^ 4
+    xor   xq, xq
+    xor   yd, yd
+    mov   strideq, widthq
+    ; stride = width * 4
+    shl   strideq, 2
+    ; w = w - ((w - 1) & 3)
+    mov   remainq, widthq
+    sub   remainq, 1
+    and   remainq, 3
+    sub   widthq, remainq
+
+    shufps m0, m0, 0
+    shufps m2, m2, 0
+    shufps m3, m3, 0
+    shufps m4, m4, 0
+
+.loop_y:
+    xor   stepd, stepd
+
+    .loop_step:
+        ; p0 *= bscale
+        mulss m5, m1, [ptrq + xq * 4]
+        movss [ptrq + xq * 4], m5
+        inc xq
+        ; filter rightwards
+        ; Here we are vectorizing the c version by 4
+        ;    for (x = 1; x < width; x++)
+        ;       ptr[x] += nu * ptr[x - 1];
+        ;   let p0 stands for ptr[x-1], the data from last loop
+        ;   and [p1,p2,p3,p4] be the vector data for this loop.
+        ; Unrolling the loop, we get:
+        ;   p1' = p1 + p0*nu
+        ;   p2' = p2 + p1*nu + p0*nu^2
+        ;   p3' = p3 + p2*nu + p1*nu^2 + p0*nu^3
+        ;   p4' = p4 + p3*nu + p2*nu^2 + p1*nu^3 + p0*nu^4
+        ; so we can do it in simd:
+        ; [p1',p2',p3',p4'] = [p1,p2,p3,p4] + [p0,p1,p2,p3]*nu +
+        ;                     [0,p0,p1,p2]*nu^2 + [0,0,p0,p1]*nu^3 +
+        ;                     [0,0,0,p0]*nu^4
+
+        .loop_x:
+            movu m6, [ptrq + xq * 4]         ; s  = [p1,p2,p3,p4]
+            pslldq m7, m6, 4                 ;      [0, p1,p2,p3]
+            movss  m7, m5                    ;      [p0,p1,p2,p3]
+            FMULADD_PS  m6, m7, m0, m6, m8   ; s += [p0,p1,p2,p3] * nu
+            pslldq m7, 4                     ;      [0,p0,p1,p2]
+            FMULADD_PS  m6, m7, m2, m6, m8   ; s += [0,p0,p1,p2]  * nu^2
+            pslldq m7, 4
+            FMULADD_PS  m6, m7, m3, m6, m8   ; s += [0,0,p0,p1]   * nu^3
+            pslldq m7, 4
+            FMULADD_PS  m6, m7, m4, m6, m8   ; s += [0,0,0,p0]    * nu^4
+            movu [ptrq + xq * 4], m6
+            shufps m5, m6, m6, q3333
+            add xq, 4
+            cmp xq, widthq
+            jl .loop_x
+
+        add widthq, remainq
+        cmp xq, widthq
+        je .end_scalar
+
+        .loop_scalar:
+            ; ptr[x] += nu * ptr[x-1]
+            movss m5, [ptrq + 4*xq - 4]
+            mulss m5, m0
+            addss m5, [ptrq + 4*xq]
+            movss [ptrq + 4*xq], m5
+            inc xq
+            cmp xq, widthq
+            jl .loop_scalar
+        .end_scalar:
+            ; ptr[width - 1] *= bscale
+            dec xq
+            mulss m5, m1, [ptrq + 4*xq]
+            movss [ptrq + 4*xq], m5
+            shufps m5, m5, 0
+
+        ; filter leftwards
+        ;    for (; x > 0; x--)
+        ;        ptr[x - 1] += nu * ptr[x];
+        ; The idea here is basically the same as filter rightwards.
+        ; But we need to take care as the data layout is different.
+        ; Let p0 stands for the ptr[x], which is the data from last loop.
+        ; The way we do it in simd as below:
+        ; [p-4', p-3', p-2', p-1'] = [p-4, p-3, p-2, p-1]
+        ;                          + [p-3, p-2, p-1, p0] * nu
+        ;                          + [p-2, p-1, p0,  0]  * nu^2
+        ;                          + [p-1, p0,  0,   0]  * nu^3
+        ;                          + [p0,  0,   0,   0]  * nu^4
+        .loop_x_back:
+            sub xq, 4
+            movu m6, [ptrq + xq * 4]      ; s = [p-4, p-3, p-2, p-1]
+            psrldq m7, m6, 4              ;     [p-3, p-2, p-1, 0  ]
+            blendps m7, m5, 0x8           ;     [p-3, p-2, p-1, p0 ]
+            FMULADD_PS m6, m7, m0, m6, m8 ; s+= [p-3, p-2, p-1, p0 ] * nu
+            psrldq m7, 4                  ;
+            FMULADD_PS m6, m7, m2, m6, m8 ; s+= [p-2, p-1, p0,  0] * nu^2
+            psrldq m7, 4
+            FMULADD_PS m6, m7, m3, m6, m8 ; s+= [p-1, p0,   0,  0] * nu^3
+            psrldq m7, 4
+            FMULADD_PS m6, m7, m4, m6, m8 ; s+= [p0,  0,    0,  0] * nu^4
+            movu [ptrq + xq * 4], m6
+            shufps m5, m6, m6, 0          ; m5 = [p-4', p-4', p-4', p-4']
+            cmp xq, remainq
+            jg .loop_x_back
+
+        cmp xq, 0
+        je .end_scalar_back
+
+        .loop_scalar_back:
+            ; ptr[x-1] += nu * ptr[x]
+            movss m5, [ptrq + 4*xq]
+            mulss m5, m0
+            addss m5, [ptrq + 4*xq - 4]
+            movss [ptrq + 4*xq - 4], m5
+            dec xq
+            cmp xq, 0
+            jg .loop_scalar_back
+        .end_scalar_back:
+
+        ; reset aligned width for next line
+        sub widthq, remainq
+
+        inc stepd
+        cmp stepd, stepsd
+        jl .loop_step
+
+    add ptrq, strideq
+    inc yd
+    cmp yd, heightd
+    jl .loop_y
+
+    RET
+%endmacro
+
+%if ARCH_X86_64
+INIT_XMM sse4
+HORIZ_SLICE
+
+INIT_XMM avx2
+HORIZ_SLICE
+%endif

libavfilter/x86/vf_gblur_init.c

+/*
+ *
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 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
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser 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 "libavutil/attributes.h"
+#include "libavutil/cpu.h"
+#include "libavutil/x86/cpu.h"
+#include "libavfilter/gblur.h"
+
+void ff_horiz_slice_sse4(float *ptr, int width, int height, int steps, float nu, float bscale);
+void ff_horiz_slice_avx2(float *ptr, int width, int height, int steps, float nu, float bscale);
+
+av_cold void ff_gblur_init_x86(GBlurContext *s)
+{
+    int cpu_flags = av_get_cpu_flags();
+
+    if (EXTERNAL_SSE4(cpu_flags))
+        s->horiz_slice = ff_horiz_slice_sse4;
+    if (EXTERNAL_AVX2(cpu_flags))
+        s->horiz_slice = ff_horiz_slice_avx2;
+}