Commit dfc58584 authored by Ronald S. Bultje's avatar Ronald S. Bultje Committed by Michael Niedermayer

vf_ssim: x86 simd for ssim_4x4xN and ssim_endN.

Both are 2-2.5x faster than their C counterpart.
Reviewed-by: 's avatarPaul B Mahol <onemda@gmail.com>
Reviewed-by: 's avatarJames Almer <jamrial@gmail.com>
Signed-off-by: 's avatarMichael Niedermayer <michael@niedermayer.cc>
parent 39a04ebc
/*
* Copyright (c) 2015 Ronald S. Bultje <rsbultje@gmail.com>
*
* 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
*/
#ifndef LIBAVFILTER_SSIM_H
#define LIBAVFILTER_SSIM_H
#include <stddef.h>
#include <stdint.h>
typedef struct SSIMDSPContext {
void (*ssim_4x4_line)(const uint8_t *buf, ptrdiff_t buf_stride,
const uint8_t *ref, ptrdiff_t ref_stride,
int (*sums)[4], int w);
float (*ssim_end_line)(const int (*sum0)[4], const int (*sum1)[4], int w);
} SSIMDSPContext;
void ff_ssim_init_x86(SSIMDSPContext *dsp);
#endif /* LIBAVFILTER_SSIM_H */
......@@ -42,6 +42,7 @@
#include "drawutils.h"
#include "formats.h"
#include "internal.h"
#include "ssim.h"
#include "video.h"
typedef struct SSIMContext {
......@@ -59,6 +60,7 @@ typedef struct SSIMContext {
int planeheight[4];
int *temp;
int is_rgb;
SSIMDSPContext dsp;
} SSIMContext;
#define OFFSET(x) offsetof(SSIMContext, x)
......@@ -85,8 +87,8 @@ static void set_meta(AVDictionary **metadata, const char *key, char comp, float
}
}
static void ssim_4x4xn(const uint8_t *main, int main_stride,
const uint8_t *ref, int ref_stride,
static void ssim_4x4xn(const uint8_t *main, ptrdiff_t main_stride,
const uint8_t *ref, ptrdiff_t ref_stride,
int (*sums)[4], int width)
{
int x, y, z;
......@@ -132,7 +134,7 @@ static float ssim_end1(int s1, int s2, int ss, int s12)
/ ((float)(fs1 * fs1 + fs2 * fs2 + ssim_c1) * (float)(vars + ssim_c2));
}
static float ssim_endn(int (*sum0)[4], int (*sum1)[4], int width)
static float ssim_endn(const int (*sum0)[4], const int (*sum1)[4], int width)
{
float ssim = 0.0;
int i;
......@@ -145,7 +147,8 @@ static float ssim_endn(int (*sum0)[4], int (*sum1)[4], int width)
return ssim;
}
static float ssim_plane(uint8_t *main, int main_stride,
static float ssim_plane(SSIMDSPContext *dsp,
uint8_t *main, int main_stride,
uint8_t *ref, int ref_stride,
int width, int height, void *temp)
{
......@@ -160,12 +163,12 @@ static float ssim_plane(uint8_t *main, int main_stride,
for (y = 1; y < height; y++) {
for (; z <= y; z++) {
FFSWAP(void*, sum0, sum1);
ssim_4x4xn(&main[4 * z * main_stride], main_stride,
&ref[4 * z * ref_stride], ref_stride,
sum0, width);
dsp->ssim_4x4_line(&main[4 * z * main_stride], main_stride,
&ref[4 * z * ref_stride], ref_stride,
sum0, width);
}
ssim += ssim_endn(sum0, sum1, width - 1);
ssim += dsp->ssim_end_line(sum0, sum1, width - 1);
}
return ssim / ((height - 1) * (width - 1));
......@@ -187,7 +190,7 @@ static AVFrame *do_ssim(AVFilterContext *ctx, AVFrame *main,
s->nb_frames++;
for (i = 0; i < s->nb_components; i++) {
c[i] = ssim_plane(main->data[i], main->linesize[i],
c[i] = ssim_plane(&s->dsp, main->data[i], main->linesize[i],
ref->data[i], ref->linesize[i],
s->planewidth[i], s->planeheight[i], s->temp);
ssimv += s->coefs[i] * c[i];
......@@ -294,6 +297,11 @@ static int config_input_ref(AVFilterLink *inlink)
if (!s->temp)
return AVERROR(ENOMEM);
s->dsp.ssim_4x4_line = ssim_4x4xn;
s->dsp.ssim_end_line = ssim_endn;
if (ARCH_X86)
ff_ssim_init_x86(&s->dsp);
return 0;
}
......
......@@ -8,6 +8,7 @@ OBJS-$(CONFIG_NOISE_FILTER) += x86/vf_noise.o
OBJS-$(CONFIG_PP7_FILTER) += x86/vf_pp7_init.o
OBJS-$(CONFIG_PULLUP_FILTER) += x86/vf_pullup_init.o
OBJS-$(CONFIG_SPP_FILTER) += x86/vf_spp.o
OBJS-$(CONFIG_SSIM_FILTER) += x86/vf_ssim_init.o
OBJS-$(CONFIG_TINTERLACE_FILTER) += x86/vf_tinterlace_init.o
OBJS-$(CONFIG_VOLUME_FILTER) += x86/af_volume_init.o
OBJS-$(CONFIG_YADIF_FILTER) += x86/vf_yadif_init.o
......@@ -19,6 +20,7 @@ YASM-OBJS-$(CONFIG_IDET_FILTER) += x86/vf_idet.o
YASM-OBJS-$(CONFIG_INTERLACE_FILTER) += x86/vf_interlace.o
YASM-OBJS-$(CONFIG_PP7_FILTER) += x86/vf_pp7.o
YASM-OBJS-$(CONFIG_PULLUP_FILTER) += x86/vf_pullup.o
YASM-OBJS-$(CONFIG_SSIM_FILTER) += x86/vf_ssim.o
YASM-OBJS-$(CONFIG_TINTERLACE_FILTER) += x86/vf_interlace.o
YASM-OBJS-$(CONFIG_VOLUME_FILTER) += x86/af_volume.o
YASM-OBJS-$(CONFIG_YADIF_FILTER) += x86/vf_yadif.o x86/yadif-16.o x86/yadif-10.o
;*****************************************************************************
;* x86-optimized functions for ssim filter
;*
;* Copyright (C) 2015 Ronald S. Bultje <rsbultje@gmail.com>
;*
;* 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_RODATA
pw_1: times 8 dw 1
ssim_c1: times 4 dd 416 ;(.01*.01*255*255*64 + .5)
ssim_c2: times 4 dd 235963 ;(.03*.03*255*255*64*63 + .5)
SECTION .text
%if ARCH_X86_64
INIT_XMM ssse3
cglobal ssim_4x4_line, 6, 8, 16, buf, buf_stride, ref, ref_stride, sums, w, buf_stride3, ref_stride3
lea ref_stride3q, [ref_strideq*3]
lea buf_stride3q, [buf_strideq*3]
pxor m7, m7
mova m15, [pw_1]
.loop:
movh m0, [bufq+buf_strideq*0] ; a1
movh m1, [refq+ref_strideq*0] ; b1
movh m2, [bufq+buf_strideq*1] ; a2
movh m3, [refq+ref_strideq*1] ; b2
punpcklbw m0, m7 ; s1 [word]
punpcklbw m1, m7 ; s2 [word]
punpcklbw m2, m7 ; s1 [word]
punpcklbw m3, m7 ; s2 [word]
pmaddwd m4, m0, m0 ; a1 * a1
pmaddwd m5, m1, m1 ; b1 * b1
pmaddwd m8, m2, m2 ; a2 * a2
pmaddwd m9, m3, m3 ; b2 * b2
paddd m4, m5 ; ss
paddd m8, m9 ; ss
pmaddwd m6, m0, m1 ; a1 * b1 = ss12
pmaddwd m5, m2, m3 ; a2 * b2 = ss12
paddw m0, m2
paddw m1, m3
paddd m6, m5 ; s12
paddd m4, m8 ; ss
movh m2, [bufq+buf_strideq*2] ; a3
movh m3, [refq+ref_strideq*2] ; b3
movh m5, [bufq+buf_stride3q] ; a4
movh m8, [refq+ref_stride3q] ; b4
punpcklbw m2, m7 ; s1 [word]
punpcklbw m3, m7 ; s2 [word]
punpcklbw m5, m7 ; s1 [word]
punpcklbw m8, m7 ; s2 [word]
pmaddwd m9, m2, m2 ; a3 * a3
pmaddwd m10, m3, m3 ; b3 * b3
pmaddwd m12, m5, m5 ; a4 * a4
pmaddwd m13, m8, m8 ; b4 * b4
pmaddwd m11, m2, m3 ; a3 * b3 = ss12
pmaddwd m14, m5, m8 ; a4 * b4 = ss12
paddd m9, m10
paddd m12, m13
paddw m0, m2
paddw m1, m3
paddw m0, m5
paddw m1, m8
paddd m6, m11
paddd m4, m9
paddd m6, m14
paddd m4, m12
; m0 = [word] s1 a,a,a,a,b,b,b,b
; m1 = [word] s2 a,a,a,a,b,b,b,b
; m4 = [dword] ss a,a,b,b
; m6 = [dword] s12 a,a,b,b
pmaddwd m0, m15 ; [dword] s1 a,a,b,b
pmaddwd m1, m15 ; [dword] s2 a,a,b,b
phaddd m0, m4 ; [dword] s1 a, b, ss a, b
phaddd m1, m6 ; [dword] s2 a, b, s12 a, b
punpckhdq m2, m0, m1 ; [dword] ss a, s12 a, ss b, s12 b
punpckldq m0, m1 ; [dword] s1 a, s2 a, s1 b, s2 b
punpckhqdq m1, m0, m2 ; [dword] a s1, s2, ss, s12
punpcklqdq m0, m2 ; [dword] b s1, s2, ss, s12
mova [sumsq+ 0], m0
mova [sumsq+mmsize], m1
add bufq, mmsize/2
add refq, mmsize/2
add sumsq, mmsize*2
sub wd, mmsize/8
jg .loop
RET
%endif
INIT_XMM sse4
cglobal ssim_end_line, 3, 3, 6, sum0, sum1, w
pxor m0, m0
.loop:
mova m1, [sum0q+mmsize*0]
mova m2, [sum0q+mmsize*1]
mova m3, [sum0q+mmsize*2]
mova m4, [sum0q+mmsize*3]
paddd m1, [sum1q+mmsize*0]
paddd m2, [sum1q+mmsize*1]
paddd m3, [sum1q+mmsize*2]
paddd m4, [sum1q+mmsize*3]
paddd m1, m2
paddd m2, m3
paddd m3, m4
paddd m4, [sum0q+mmsize*4]
paddd m4, [sum1q+mmsize*4]
TRANSPOSE4x4D 1, 2, 3, 4, 5
; m1 = fs1, m2 = fs2, m3 = fss, m4 = fs12
pslld m3, 6
pslld m4, 6
pmulld m5, m1, m2 ; fs1 * fs2
pmulld m1, m1 ; fs1 * fs1
pmulld m2, m2 ; fs2 * fs2
psubd m3, m1
psubd m4, m5 ; covariance
psubd m3, m2 ; variance
; m1 = fs1 * fs1, m2 = fs2 * fs2, m3 = variance, m4 = covariance, m5 = fs1 * fs2
paddd m4, m4 ; 2 * covariance
paddd m5, m5 ; 2 * fs1 * fs2
paddd m1, m2 ; fs1 * fs1 + fs2 * fs2
paddd m3, [ssim_c2] ; variance + ssim_c2
paddd m4, [ssim_c2] ; 2 * covariance + ssim_c2
paddd m5, [ssim_c1] ; 2 * fs1 * fs2 + ssim_c1
paddd m1, [ssim_c1] ; fs1 * fs1 + fs2 * fs2 + ssim_c1
; convert to float
cvtdq2ps m3, m3
cvtdq2ps m4, m4
cvtdq2ps m5, m5
cvtdq2ps m1, m1
mulps m4, m5
mulps m3, m1
divps m4, m3 ; ssim_endl
addps m0, m4 ; ssim
add sum0q, mmsize*4
add sum1q, mmsize*4
sub wd, 4
jg .loop
; subps the ones we added too much
test wd, wd
jz .end
add wd, 4
test wd, 2
jz .skip2
psrldq m4, 8
.skip2:
test wd, 1
jz .skip1
psrldq m4, 4
.skip1:
subps m0, m4
.end:
movhlps m4, m0
addps m0, m4
movss m4, m0
shufps m0, m0, 1
addss m0, m4
%if ARCH_X86_32
movss r0m, m0
fld r0mp
%endif
RET
/*
* Copyright (c) 2015 Ronald S. Bultje <rsbultje@gmail.com>
*
* 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/cpu.h"
#include "libavfilter/ssim.h"
void ff_ssim_4x4_line_ssse3(const uint8_t *buf, ptrdiff_t buf_stride,
const uint8_t *ref, ptrdiff_t ref_stride,
int (*sums)[4], int w);
float ff_ssim_end_line_sse4(const int (*sum0)[4], const int (*sum1)[4], int w);
void ff_ssim_init_x86(SSIMDSPContext *dsp)
{
int cpu_flags = av_get_cpu_flags();
if (ARCH_X86_64 && EXTERNAL_SSSE3(cpu_flags))
dsp->ssim_4x4_line = ff_ssim_4x4_line_ssse3;
if (EXTERNAL_SSE4(cpu_flags))
dsp->ssim_end_line = ff_ssim_end_line_sse4;
}
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment