Commit 0a81d36a authored by Jingning Han's avatar Jingning Han
Browse files

Replace vp10_ssim functions with vpx_ssim versions

Remove the ssim files from vp10 folder.

Change-Id: Icc1d9d61a0878d53584328b278dfa090aa070035
parent 54d66ef1
......@@ -261,17 +261,6 @@ if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
specialize qw/vp10_fdct8x8_quant sse2 ssse3 neon/;
}
#
# Structured Similarity (SSIM)
#
if (vpx_config("CONFIG_INTERNAL_STATS") eq "yes") {
add_proto qw/void vp10_ssim_parms_8x8/, "uint8_t *s, int sp, uint8_t *r, int rp, unsigned long *sum_s, unsigned long *sum_r, unsigned long *sum_sq_s, unsigned long *sum_sq_r, unsigned long *sum_sxr";
specialize qw/vp10_ssim_parms_8x8/, "$sse2_x86_64";
add_proto qw/void vp10_ssim_parms_16x16/, "uint8_t *s, int sp, uint8_t *r, int rp, unsigned long *sum_s, unsigned long *sum_r, unsigned long *sum_sq_s, unsigned long *sum_sq_r, unsigned long *sum_sxr";
specialize qw/vp10_ssim_parms_16x16/, "$sse2_x86_64";
}
# fdct functions
if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
......@@ -330,14 +319,6 @@ if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
add_proto qw/void vp10_highbd_quantize_fp_32x32/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
specialize qw/vp10_highbd_quantize_fp_32x32/;
#
# Structured Similarity (SSIM)
#
if (vpx_config("CONFIG_INTERNAL_STATS") eq "yes") {
add_proto qw/void vp10_highbd_ssim_parms_8x8/, "uint16_t *s, int sp, uint16_t *r, int rp, uint32_t *sum_s, uint32_t *sum_r, uint32_t *sum_sq_s, uint32_t *sum_sq_r, uint32_t *sum_sxr";
specialize qw/vp10_highbd_ssim_parms_8x8/;
}
# fdct functions
add_proto qw/void vp10_highbd_fht4x4/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
specialize qw/vp10_highbd_fht4x4/;
......
......@@ -54,8 +54,8 @@ static int variance(int sum, int sum_squared, int size) {
// var_1 = (r0^2+r1^2+r2^2+r3^2) - ((r0 + r1 + r2 + r3) / 4 )^2
// The returned blockiness is the scaled value
// Reconstructed blockiness / ( 1 + var_0 + var_1 ) ;
int blockiness_vertical(const uint8_t *s, int sp, const uint8_t *r, int rp,
int size) {
static int blockiness_vertical(const uint8_t *s, int sp, const uint8_t *r,
int rp, int size) {
int s_blockiness = 0;
int r_blockiness = 0;
int sum_0 = 0;
......@@ -86,8 +86,8 @@ int blockiness_vertical(const uint8_t *s, int sp, const uint8_t *r, int rp,
// Calculate a blockiness level for a horizontal block edge
// same as above.
int blockiness_horizontal(const uint8_t *s, int sp, const uint8_t *r, int rp,
int size) {
static int blockiness_horizontal(const uint8_t *s, int sp, const uint8_t *r,
int rp, int size) {
int s_blockiness = 0;
int r_blockiness = 0;
int sum_0 = 0;
......
......@@ -12,15 +12,7 @@
#include <stdio.h>
#include <limits.h>
#include "./vp10_rtcd.h"
#include "./vpx_config.h"
#include "./vpx_dsp_rtcd.h"
#include "./vpx_scale_rtcd.h"
#include "vpx/internal/vpx_psnr.h"
#include "vpx_dsp/vpx_filter.h"
#include "vpx_ports/mem.h"
#include "vpx_ports/vpx_timer.h"
#include "vpx_scale/vpx_scale.h"
#include "vp10/common/alloccommon.h"
#include "vp10/common/filter.h"
......@@ -51,12 +43,21 @@
#include "vp10/encoder/segmentation.h"
#include "vp10/encoder/skin_detection.h"
#include "vp10/encoder/speed_features.h"
#if CONFIG_INTERNAL_STATS
#include "vp10/encoder/ssim.h"
#endif
#include "vp10/encoder/svc_layercontext.h"
#include "vp10/encoder/temporal_filter.h"
#include "./vp10_rtcd.h"
#include "./vpx_dsp_rtcd.h"
#include "./vpx_scale_rtcd.h"
#include "vpx/internal/vpx_psnr.h"
#if CONFIG_INTERNAL_STATS
#include "vpx_dsp/ssim.h"
#endif
#include "vpx_dsp/vpx_filter.h"
#include "vpx_ports/mem.h"
#include "vpx_ports/vpx_timer.h"
#include "vpx_scale/vpx_scale.h"
#define AM_SEGMENT_ID_INACTIVE 7
#define AM_SEGMENT_ID_ACTIVE 0
......@@ -4416,13 +4417,13 @@ int vp10_get_compressed_data(VP9_COMP *cpi, unsigned int *frame_flags,
#if CONFIG_VP9_HIGHBITDEPTH
if (cm->use_highbitdepth) {
frame_ssim2 = vp10_highbd_calc_ssim(orig, recon, &weight,
frame_ssim2 = vpx_highbd_calc_ssim(orig, recon, &weight,
(int)cm->bit_depth);
} else {
frame_ssim2 = vp10_calc_ssim(orig, recon, &weight);
frame_ssim2 = vpx_calc_ssim(orig, recon, &weight);
}
#else
frame_ssim2 = vp10_calc_ssim(orig, recon, &weight);
frame_ssim2 = vpx_calc_ssim(orig, recon, &weight);
#endif // CONFIG_VP9_HIGHBITDEPTH
cpi->worst_ssim= MIN(cpi->worst_ssim, frame_ssim2);
......@@ -4431,13 +4432,13 @@ int vp10_get_compressed_data(VP9_COMP *cpi, unsigned int *frame_flags,
#if CONFIG_VP9_HIGHBITDEPTH
if (cm->use_highbitdepth) {
frame_ssim2 = vp10_highbd_calc_ssim(
frame_ssim2 = vpx_highbd_calc_ssim(
orig, &cm->post_proc_buffer, &weight, (int)cm->bit_depth);
} else {
frame_ssim2 = vp10_calc_ssim(orig, &cm->post_proc_buffer, &weight);
frame_ssim2 = vpx_calc_ssim(orig, &cm->post_proc_buffer, &weight);
}
#else
frame_ssim2 = vp10_calc_ssim(orig, &cm->post_proc_buffer, &weight);
frame_ssim2 = vpx_calc_ssim(orig, &cm->post_proc_buffer, &weight);
#endif // CONFIG_VP9_HIGHBITDEPTH
cpi->summedp_quality += frame_ssim2 * weight;
......@@ -4472,7 +4473,7 @@ int vp10_get_compressed_data(VP9_COMP *cpi, unsigned int *frame_flags,
if (!cm->use_highbitdepth)
#endif
{
double this_inconsistency = vp10_get_ssim_metrics(
double this_inconsistency = vpx_get_ssim_metrics(
cpi->Source->y_buffer, cpi->Source->y_stride,
cm->frame_to_show->y_buffer, cm->frame_to_show->y_stride,
cpi->Source->y_width, cpi->Source->y_height, cpi->ssim_vars,
......@@ -4492,14 +4493,14 @@ int vp10_get_compressed_data(VP9_COMP *cpi, unsigned int *frame_flags,
double y, u, v, frame_all;
#if CONFIG_VP9_HIGHBITDEPTH
if (cm->use_highbitdepth) {
frame_all = vp10_highbd_calc_ssimg(cpi->Source, cm->frame_to_show, &y,
frame_all = vpx_highbd_calc_ssimg(cpi->Source, cm->frame_to_show, &y,
&u, &v, (int)cm->bit_depth);
} else {
frame_all = vp10_calc_ssimg(cpi->Source, cm->frame_to_show, &y, &u,
frame_all = vpx_calc_ssimg(cpi->Source, cm->frame_to_show, &y, &u,
&v);
}
#else
frame_all = vp10_calc_ssimg(cpi->Source, cm->frame_to_show, &y, &u, &v);
frame_all = vpx_calc_ssimg(cpi->Source, cm->frame_to_show, &y, &u, &v);
#endif // CONFIG_VP9_HIGHBITDEPTH
adjust_image_stat(y, u, v, frame_all, &cpi->ssimg);
}
......@@ -4508,7 +4509,7 @@ int vp10_get_compressed_data(VP9_COMP *cpi, unsigned int *frame_flags,
#endif
{
double y, u, v, frame_all;
frame_all = vp10_calc_fastssim(cpi->Source, cm->frame_to_show, &y, &u,
frame_all = vpx_calc_fastssim(cpi->Source, cm->frame_to_show, &y, &u,
&v);
adjust_image_stat(y, u, v, frame_all, &cpi->fastssim);
/* TODO(JBB): add 10/12 bit support */
......@@ -4518,12 +4519,11 @@ int vp10_get_compressed_data(VP9_COMP *cpi, unsigned int *frame_flags,
#endif
{
double y, u, v, frame_all;
frame_all = vp10_psnrhvs(cpi->Source, cm->frame_to_show, &y, &u, &v);
frame_all = vpx_psnrhvs(cpi->Source, cm->frame_to_show, &y, &u, &v);
adjust_image_stat(y, u, v, frame_all, &cpi->psnrhvs);
}
}
}
#endif
if (is_two_pass_svc(cpi)) {
......
......@@ -14,9 +14,7 @@
#include <stdio.h>
#include "./vpx_config.h"
#include "vpx/internal/vpx_codec_internal.h"
#include "vpx/vp8cx.h"
#include "vpx_util/vpx_thread.h"
#include "vp10/common/alloccommon.h"
#include "vp10/common/ppflags.h"
......@@ -34,18 +32,21 @@
#include "vp10/encoder/quantize.h"
#include "vp10/encoder/ratectrl.h"
#include "vp10/encoder/rd.h"
#if CONFIG_INTERNAL_STATS
#include "vp10/encoder/ssim.h"
#endif
#include "vp10/encoder/speed_features.h"
#include "vp10/encoder/svc_layercontext.h"
#include "vp10/encoder/tokenize.h"
#include "vpx_dsp/variance.h"
#if CONFIG_VP9_TEMPORAL_DENOISING
#include "vp10/encoder/denoiser.h"
#endif
#if CONFIG_INTERNAL_STATS
#include "vpx_dsp/ssim.h"
#endif
#include "vpx_dsp/variance.h"
#include "vpx/internal/vpx_codec_internal.h"
#include "vpx_util/vpx_thread.h"
#ifdef __cplusplus
extern "C" {
#endif
......
/*
* Copyright (c) 2010 The WebM project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*
* This code was originally written by: Nathan E. Egge, at the Daala
* project.
*/
#include <math.h>
#include <string.h>
#include "./vpx_config.h"
#include "./vp10_rtcd.h"
#include "vp10/encoder/ssim.h"
/* TODO(jbb): High bit depth version of this code needed */
typedef struct fs_level fs_level;
typedef struct fs_ctx fs_ctx;
#define SSIM_C1 (255 * 255 * 0.01 * 0.01)
#define SSIM_C2 (255 * 255 * 0.03 * 0.03)
#define FS_MINI(_a, _b) ((_a) < (_b) ? (_a) : (_b))
#define FS_MAXI(_a, _b) ((_a) > (_b) ? (_a) : (_b))
struct fs_level {
uint16_t *im1;
uint16_t *im2;
double *ssim;
int w;
int h;
};
struct fs_ctx {
fs_level *level;
int nlevels;
unsigned *col_buf;
};
static void fs_ctx_init(fs_ctx *_ctx, int _w, int _h, int _nlevels) {
unsigned char *data;
size_t data_size;
int lw;
int lh;
int l;
lw = (_w + 1) >> 1;
lh = (_h + 1) >> 1;
data_size = _nlevels * sizeof(fs_level)
+ 2 * (lw + 8) * 8 * sizeof(*_ctx->col_buf);
for (l = 0; l < _nlevels; l++) {
size_t im_size;
size_t level_size;
im_size = lw * (size_t) lh;
level_size = 2 * im_size * sizeof(*_ctx->level[l].im1);
level_size += sizeof(*_ctx->level[l].ssim) - 1;
level_size /= sizeof(*_ctx->level[l].ssim);
level_size += im_size;
level_size *= sizeof(*_ctx->level[l].ssim);
data_size += level_size;
lw = (lw + 1) >> 1;
lh = (lh + 1) >> 1;
}
data = (unsigned char *) malloc(data_size);
_ctx->level = (fs_level *) data;
_ctx->nlevels = _nlevels;
data += _nlevels * sizeof(*_ctx->level);
lw = (_w + 1) >> 1;
lh = (_h + 1) >> 1;
for (l = 0; l < _nlevels; l++) {
size_t im_size;
size_t level_size;
_ctx->level[l].w = lw;
_ctx->level[l].h = lh;
im_size = lw * (size_t) lh;
level_size = 2 * im_size * sizeof(*_ctx->level[l].im1);
level_size += sizeof(*_ctx->level[l].ssim) - 1;
level_size /= sizeof(*_ctx->level[l].ssim);
level_size *= sizeof(*_ctx->level[l].ssim);
_ctx->level[l].im1 = (uint16_t *) data;
_ctx->level[l].im2 = _ctx->level[l].im1 + im_size;
data += level_size;
_ctx->level[l].ssim = (double *) data;
data += im_size * sizeof(*_ctx->level[l].ssim);
lw = (lw + 1) >> 1;
lh = (lh + 1) >> 1;
}
_ctx->col_buf = (unsigned *) data;
}
static void fs_ctx_clear(fs_ctx *_ctx) {
free(_ctx->level);
}
static void fs_downsample_level(fs_ctx *_ctx, int _l) {
const uint16_t *src1;
const uint16_t *src2;
uint16_t *dst1;
uint16_t *dst2;
int w2;
int h2;
int w;
int h;
int i;
int j;
w = _ctx->level[_l].w;
h = _ctx->level[_l].h;
dst1 = _ctx->level[_l].im1;
dst2 = _ctx->level[_l].im2;
w2 = _ctx->level[_l - 1].w;
h2 = _ctx->level[_l - 1].h;
src1 = _ctx->level[_l - 1].im1;
src2 = _ctx->level[_l - 1].im2;
for (j = 0; j < h; j++) {
int j0offs;
int j1offs;
j0offs = 2 * j * w2;
j1offs = FS_MINI(2 * j + 1, h2) * w2;
for (i = 0; i < w; i++) {
int i0;
int i1;
i0 = 2 * i;
i1 = FS_MINI(i0 + 1, w2);
dst1[j * w + i] = src1[j0offs + i0] + src1[j0offs + i1]
+ src1[j1offs + i0] + src1[j1offs + i1];
dst2[j * w + i] = src2[j0offs + i0] + src2[j0offs + i1]
+ src2[j1offs + i0] + src2[j1offs + i1];
}
}
}
static void fs_downsample_level0(fs_ctx *_ctx, const unsigned char *_src1,
int _s1ystride, const unsigned char *_src2,
int _s2ystride, int _w, int _h) {
uint16_t *dst1;
uint16_t *dst2;
int w;
int h;
int i;
int j;
w = _ctx->level[0].w;
h = _ctx->level[0].h;
dst1 = _ctx->level[0].im1;
dst2 = _ctx->level[0].im2;
for (j = 0; j < h; j++) {
int j0;
int j1;
j0 = 2 * j;
j1 = FS_MINI(j0 + 1, _h);
for (i = 0; i < w; i++) {
int i0;
int i1;
i0 = 2 * i;
i1 = FS_MINI(i0 + 1, _w);
dst1[j * w + i] = _src1[j0 * _s1ystride + i0]
+ _src1[j0 * _s1ystride + i1] + _src1[j1 * _s1ystride + i0]
+ _src1[j1 * _s1ystride + i1];
dst2[j * w + i] = _src2[j0 * _s2ystride + i0]
+ _src2[j0 * _s2ystride + i1] + _src2[j1 * _s2ystride + i0]
+ _src2[j1 * _s2ystride + i1];
}
}
}
static void fs_apply_luminance(fs_ctx *_ctx, int _l) {
unsigned *col_sums_x;
unsigned *col_sums_y;
uint16_t *im1;
uint16_t *im2;
double *ssim;
double c1;
int w;
int h;
int j0offs;
int j1offs;
int i;
int j;
w = _ctx->level[_l].w;
h = _ctx->level[_l].h;
col_sums_x = _ctx->col_buf;
col_sums_y = col_sums_x + w;
im1 = _ctx->level[_l].im1;
im2 = _ctx->level[_l].im2;
for (i = 0; i < w; i++)
col_sums_x[i] = 5 * im1[i];
for (i = 0; i < w; i++)
col_sums_y[i] = 5 * im2[i];
for (j = 1; j < 4; j++) {
j1offs = FS_MINI(j, h - 1) * w;
for (i = 0; i < w; i++)
col_sums_x[i] += im1[j1offs + i];
for (i = 0; i < w; i++)
col_sums_y[i] += im2[j1offs + i];
}
ssim = _ctx->level[_l].ssim;
c1 = (double) (SSIM_C1 * 4096 * (1 << 4 * _l));
for (j = 0; j < h; j++) {
unsigned mux;
unsigned muy;
int i0;
int i1;
mux = 5 * col_sums_x[0];
muy = 5 * col_sums_y[0];
for (i = 1; i < 4; i++) {
i1 = FS_MINI(i, w - 1);
mux += col_sums_x[i1];
muy += col_sums_y[i1];
}
for (i = 0; i < w; i++) {
ssim[j * w + i] *= (2 * mux * (double) muy + c1)
/ (mux * (double) mux + muy * (double) muy + c1);
if (i + 1 < w) {
i0 = FS_MAXI(0, i - 4);
i1 = FS_MINI(i + 4, w - 1);
mux += col_sums_x[i1] - col_sums_x[i0];
muy += col_sums_x[i1] - col_sums_x[i0];
}
}
if (j + 1 < h) {
j0offs = FS_MAXI(0, j - 4) * w;
for (i = 0; i < w; i++)
col_sums_x[i] -= im1[j0offs + i];
for (i = 0; i < w; i++)
col_sums_y[i] -= im2[j0offs + i];
j1offs = FS_MINI(j + 4, h - 1) * w;
for (i = 0; i < w; i++)
col_sums_x[i] += im1[j1offs + i];
for (i = 0; i < w; i++)
col_sums_y[i] += im2[j1offs + i];
}
}
}
#define FS_COL_SET(_col, _joffs, _ioffs) \
do { \
unsigned gx; \
unsigned gy; \
gx = gx_buf[((j + (_joffs)) & 7) * stride + i + (_ioffs)]; \
gy = gy_buf[((j + (_joffs)) & 7) * stride + i + (_ioffs)]; \
col_sums_gx2[(_col)] = gx * (double)gx; \
col_sums_gy2[(_col)] = gy * (double)gy; \
col_sums_gxgy[(_col)] = gx * (double)gy; \
} \
while (0)
#define FS_COL_ADD(_col, _joffs, _ioffs) \
do { \
unsigned gx; \
unsigned gy; \
gx = gx_buf[((j + (_joffs)) & 7) * stride + i + (_ioffs)]; \
gy = gy_buf[((j + (_joffs)) & 7) * stride + i + (_ioffs)]; \
col_sums_gx2[(_col)] += gx * (double)gx; \
col_sums_gy2[(_col)] += gy * (double)gy; \
col_sums_gxgy[(_col)] += gx * (double)gy; \
} \
while (0)
#define FS_COL_SUB(_col, _joffs, _ioffs) \
do { \
unsigned gx; \
unsigned gy; \
gx = gx_buf[((j + (_joffs)) & 7) * stride + i + (_ioffs)]; \
gy = gy_buf[((j + (_joffs)) & 7) * stride + i + (_ioffs)]; \
col_sums_gx2[(_col)] -= gx * (double)gx; \
col_sums_gy2[(_col)] -= gy * (double)gy; \
col_sums_gxgy[(_col)] -= gx * (double)gy; \
} \
while (0)
#define FS_COL_COPY(_col1, _col2) \
do { \
col_sums_gx2[(_col1)] = col_sums_gx2[(_col2)]; \
col_sums_gy2[(_col1)] = col_sums_gy2[(_col2)]; \
col_sums_gxgy[(_col1)] = col_sums_gxgy[(_col2)]; \
} \
while (0)
#define FS_COL_HALVE(_col1, _col2) \
do { \
col_sums_gx2[(_col1)] = col_sums_gx2[(_col2)] * 0.5; \
col_sums_gy2[(_col1)] = col_sums_gy2[(_col2)] * 0.5; \
col_sums_gxgy[(_col1)] = col_sums_gxgy[(_col2)] * 0.5; \
} \
while (0)
#define FS_COL_DOUBLE(_col1, _col2) \
do { \
col_sums_gx2[(_col1)] = col_sums_gx2[(_col2)] * 2; \
col_sums_gy2[(_col1)] = col_sums_gy2[(_col2)] * 2; \
col_sums_gxgy[(_col1)] = col_sums_gxgy[(_col2)] * 2; \
} \
while (0)
static void fs_calc_structure(fs_ctx *_ctx, int _l) {
uint16_t *im1;
uint16_t *im2;
unsigned *gx_buf;
unsigned *gy_buf;
double *ssim;
double col_sums_gx2[8];
double col_sums_gy2[8];
double col_sums_gxgy[8];
double c2;
int stride;
int w;
int h;
int i;
int j;
w = _ctx->level[_l].w;
h = _ctx->level[_l].h;
im1 = _ctx->level[_l].im1;
im2 = _ctx->level[_l].im2;
ssim = _ctx->level[_l].ssim;
gx_buf = _ctx->col_buf;
stride = w + 8;
gy_buf = gx_buf + 8 * stride;
memset(gx_buf, 0, 2 * 8 * stride * sizeof(*gx_buf));
c2 = SSIM_C2 * (1 << 4 * _l) * 16 * 104;
for (j = 0; j < h + 4; j++) {
if (j < h - 1) {
for (i = 0; i < w - 1; i++) {
unsigned g1;
unsigned g2;
unsigned gx;
unsigned gy;
g1 = abs(im1[(j + 1) * w + i + 1] - im1[j * w + i]);
g2 = abs(im1[(j + 1) * w + i] - im1[j * w + i + 1]);
gx = 4 * FS_MAXI(g1, g2) + FS_MINI(g1, g2);
g1 = abs(im2[(j + 1) * w + i + 1] - im2[j * w + i]);
g2 = abs(im2[(j + 1) * w + i] - im2[j * w + i + 1]);
gy = 4 * FS_MAXI(g1, g2) + FS_MINI(g1, g2);
gx_buf[(j & 7) * stride + i + 4] = gx;
gy_buf[(j & 7) * stride + i + 4] = gy;
}
} else {
memset(gx_buf + (j & 7) * stride, 0, stride * sizeof(*gx_buf));
memset(gy_buf + (j & 7) * stride, 0, stride * sizeof(*gy_buf));
}
if (j >= 4) {
int k;
col_sums_gx2[3] = col_sums_gx2[2] = col_sums_gx2[1] = col_sums_gx2[0] = 0;
col_sums_gy2[3] = col_sums_gy2[2] = col_sums_gy2[1] = col_sums_gy2[0] = 0;
col_sums_gxgy[3] = col_sums_gxgy[2] = col_sums_gxgy[1] =
col_sums_gxgy[0] = 0;
for (i = 4; i < 8; i++) {
FS_COL_SET(i, -1, 0);
FS_COL_ADD(i, 0, 0);
for (k = 1; k < 8 - i; k++) {
FS_COL_DOUBLE(i, i);
FS_COL_ADD(i, -k - 1, 0);
FS_COL_ADD(i, k, 0);
}
}
for (i = 0; i < w; i++) {
double mugx2;
double mugy2;
double mugxgy;
mugx2 = col_sums_gx2[0];
for (k = 1; k < 8; k++)
mugx2 += col_sums_gx2[k];
mugy2 = col_sums_gy2[0];
for (k = 1; k < 8; k++)
mugy2 += col_sums_gy2[k];
mugxgy = col_sums_gxgy[0];
for (k = 1; k < 8; k++)
mugxgy += col_sums_gxgy[k];
ssim[(j - 4) * w + i] = (2 * mugxgy + c2) / (mugx2 + mugy2 + c2);
if (i + 1 < w) {
FS_COL_SET(0, -1, 1);
FS_COL_ADD(0, 0, 1);
FS_COL_SUB(2, -3, 2);
FS_COL_SUB(2, 2, 2);
FS_COL_HALVE(1, 2);
FS_COL_SUB(3, -4, 3);
FS_COL_SUB(3, 3, 3);
FS_COL_HALVE(2, 3);
FS_COL_COPY(3, 4);
FS_COL_DOUBLE(4, 5);
FS_COL_ADD(4, -4, 5);
FS_COL_ADD(4, 3, 5);
FS_COL_DOUBLE(5, 6);
FS_COL_ADD(5, -3, 6);
FS_COL_ADD(5, 2, 6);
FS_COL_DOUBLE(6, 7);
FS_COL_ADD(6, -2, 7);
FS_COL_ADD(6, 1, 7);
FS_COL_SET(7, -1, 8);
FS_COL_ADD(7, 0, 8);
}