2016-02-09 23:24:33 +03:00
|
|
|
/*
|
2016-11-29 02:26:06 +03:00
|
|
|
* Copyright (c) 2016, Alliance for Open Media. All rights reserved
|
2016-02-09 23:24:33 +03:00
|
|
|
*
|
2016-11-29 02:26:06 +03:00
|
|
|
* This source code is subject to the terms of the BSD 2 Clause License and
|
|
|
|
* the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
|
|
|
|
* was not distributed with this source code in the LICENSE file, you can
|
|
|
|
* obtain it at www.aomedia.org/license/software. If the Alliance for Open
|
|
|
|
* Media Patent License 1.0 was not distributed with this source code in the
|
|
|
|
* PATENTS file, you can obtain it at www.aomedia.org/license/patent.
|
2016-02-09 23:24:33 +03:00
|
|
|
*/
|
|
|
|
|
|
|
|
#include <assert.h>
|
2016-03-11 09:56:24 +03:00
|
|
|
#include <float.h>
|
2016-02-09 23:24:33 +03:00
|
|
|
#include <limits.h>
|
|
|
|
#include <math.h>
|
|
|
|
|
2016-08-31 00:01:10 +03:00
|
|
|
#include "./aom_scale_rtcd.h"
|
2016-02-09 23:24:33 +03:00
|
|
|
|
2016-08-23 02:08:15 +03:00
|
|
|
#include "aom_dsp/psnr.h"
|
2016-08-31 00:01:10 +03:00
|
|
|
#include "aom_dsp/aom_dsp_common.h"
|
|
|
|
#include "aom_mem/aom_mem.h"
|
2016-08-23 02:08:15 +03:00
|
|
|
#include "aom_ports/mem.h"
|
2016-02-09 23:24:33 +03:00
|
|
|
|
2016-08-23 02:08:15 +03:00
|
|
|
#include "av1/common/onyxc_int.h"
|
|
|
|
#include "av1/common/quant_common.h"
|
2017-01-04 02:14:57 +03:00
|
|
|
#include "av1/common/restoration.h"
|
2016-02-09 23:24:33 +03:00
|
|
|
|
2016-08-23 02:08:15 +03:00
|
|
|
#include "av1/encoder/encoder.h"
|
|
|
|
#include "av1/encoder/picklpf.h"
|
|
|
|
#include "av1/encoder/pickrst.h"
|
|
|
|
#include "av1/encoder/quantize.h"
|
2016-02-09 23:24:33 +03:00
|
|
|
|
2016-09-17 23:16:58 +03:00
|
|
|
typedef double (*search_restore_type)(const YV12_BUFFER_CONFIG *src,
|
|
|
|
AV1_COMP *cpi, int filter_level,
|
|
|
|
int partial_frame, RestorationInfo *info,
|
2017-01-06 22:18:23 +03:00
|
|
|
RestorationType *rest_level,
|
2016-12-08 14:25:47 +03:00
|
|
|
double *best_tile_cost,
|
|
|
|
YV12_BUFFER_CONFIG *dst_frame);
|
2016-09-17 23:16:58 +03:00
|
|
|
|
2016-12-08 20:22:44 +03:00
|
|
|
const int frame_level_restore_bits[RESTORE_TYPES] = { 2, 2, 3, 3, 2 };
|
2016-09-09 01:15:17 +03:00
|
|
|
|
|
|
|
static int64_t sse_restoration_tile(const YV12_BUFFER_CONFIG *src,
|
2016-12-08 14:25:47 +03:00
|
|
|
const YV12_BUFFER_CONFIG *dst,
|
|
|
|
const AV1_COMMON *cm, int h_start,
|
2016-12-15 03:53:17 +03:00
|
|
|
int width, int v_start, int height,
|
2017-01-04 02:14:57 +03:00
|
|
|
int components_pattern) {
|
|
|
|
int64_t filt_err = 0;
|
2017-01-06 00:58:16 +03:00
|
|
|
(void)cm;
|
|
|
|
// Y and UV components cannot be mixed
|
|
|
|
assert(components_pattern == 1 || components_pattern == 2 ||
|
|
|
|
components_pattern == 4 || components_pattern == 6);
|
2016-09-09 01:15:17 +03:00
|
|
|
#if CONFIG_AOM_HIGHBITDEPTH
|
|
|
|
if (cm->use_highbitdepth) {
|
2017-01-04 02:14:57 +03:00
|
|
|
if ((components_pattern >> AOM_PLANE_Y) & 1) {
|
|
|
|
filt_err +=
|
|
|
|
aom_highbd_get_y_sse_part(src, dst, h_start, width, v_start, height);
|
|
|
|
}
|
|
|
|
if ((components_pattern >> AOM_PLANE_U) & 1) {
|
2017-01-06 00:58:16 +03:00
|
|
|
filt_err +=
|
|
|
|
aom_highbd_get_u_sse_part(src, dst, h_start, width, v_start, height);
|
2017-01-04 02:14:57 +03:00
|
|
|
}
|
|
|
|
if ((components_pattern >> AOM_PLANE_V) & 1) {
|
2017-01-06 00:58:16 +03:00
|
|
|
filt_err +=
|
|
|
|
aom_highbd_get_v_sse_part(src, dst, h_start, width, v_start, height);
|
2016-12-15 03:53:17 +03:00
|
|
|
}
|
|
|
|
return filt_err;
|
2016-09-09 01:15:17 +03:00
|
|
|
}
|
|
|
|
#endif // CONFIG_AOM_HIGHBITDEPTH
|
2017-01-04 02:14:57 +03:00
|
|
|
if ((components_pattern >> AOM_PLANE_Y) & 1) {
|
|
|
|
filt_err += aom_get_y_sse_part(src, dst, h_start, width, v_start, height);
|
|
|
|
}
|
|
|
|
if ((components_pattern >> AOM_PLANE_U) & 1) {
|
2017-01-06 00:58:16 +03:00
|
|
|
filt_err += aom_get_u_sse_part(src, dst, h_start, width, v_start, height);
|
2017-01-04 02:14:57 +03:00
|
|
|
}
|
|
|
|
if ((components_pattern >> AOM_PLANE_V) & 1) {
|
2017-01-06 00:58:16 +03:00
|
|
|
filt_err += aom_get_v_sse_part(src, dst, h_start, width, v_start, height);
|
2016-12-15 03:53:17 +03:00
|
|
|
}
|
2016-09-09 01:15:17 +03:00
|
|
|
return filt_err;
|
|
|
|
}
|
|
|
|
|
2017-01-04 02:14:57 +03:00
|
|
|
static int64_t sse_restoration_frame(const YV12_BUFFER_CONFIG *src,
|
|
|
|
const YV12_BUFFER_CONFIG *dst,
|
|
|
|
int components_pattern) {
|
|
|
|
int64_t filt_err = 0;
|
|
|
|
#if CONFIG_AOM_HIGHBITDEPTH
|
|
|
|
if (cm->use_highbitdepth) {
|
|
|
|
if ((components_pattern >> AOM_PLANE_Y) & 1) {
|
|
|
|
filt_err += aom_highbd_get_y_sse(src, dst);
|
|
|
|
}
|
|
|
|
if ((components_pattern >> AOM_PLANE_U) & 1) {
|
|
|
|
filt_err += aom_highbd_get_u_sse(src, dst);
|
|
|
|
}
|
|
|
|
if ((components_pattern >> AOM_PLANE_V) & 1) {
|
|
|
|
filt_err += aom_highbd_get_v_sse(src, dst);
|
|
|
|
}
|
|
|
|
return filt_err;
|
|
|
|
}
|
|
|
|
#endif // CONFIG_AOM_HIGHBITDEPTH
|
|
|
|
if ((components_pattern >> AOM_PLANE_Y) & 1) {
|
|
|
|
filt_err = aom_get_y_sse(src, dst);
|
|
|
|
}
|
|
|
|
if ((components_pattern >> AOM_PLANE_U) & 1) {
|
|
|
|
filt_err += aom_get_u_sse(src, dst);
|
|
|
|
}
|
|
|
|
if ((components_pattern >> AOM_PLANE_V) & 1) {
|
|
|
|
filt_err += aom_get_v_sse(src, dst);
|
|
|
|
}
|
|
|
|
return filt_err;
|
|
|
|
}
|
|
|
|
|
2016-09-09 01:15:17 +03:00
|
|
|
static int64_t try_restoration_tile(const YV12_BUFFER_CONFIG *src,
|
|
|
|
AV1_COMP *const cpi, RestorationInfo *rsi,
|
2017-01-04 02:14:57 +03:00
|
|
|
int components_pattern, int partial_frame,
|
|
|
|
int tile_idx, int subtile_idx,
|
|
|
|
int subtile_bits,
|
2016-12-08 14:25:47 +03:00
|
|
|
YV12_BUFFER_CONFIG *dst_frame) {
|
2016-09-09 01:15:17 +03:00
|
|
|
AV1_COMMON *const cm = &cpi->common;
|
|
|
|
int64_t filt_err;
|
|
|
|
int tile_width, tile_height, nhtiles, nvtiles;
|
|
|
|
int h_start, h_end, v_start, v_end;
|
2017-01-06 00:58:16 +03:00
|
|
|
int ntiles, width, height;
|
|
|
|
|
|
|
|
// Y and UV components cannot be mixed
|
|
|
|
assert(components_pattern == 1 || components_pattern == 2 ||
|
|
|
|
components_pattern == 4 || components_pattern == 6);
|
|
|
|
|
|
|
|
if (components_pattern == 1) { // Y only
|
|
|
|
width = src->y_crop_width;
|
|
|
|
height = src->y_crop_height;
|
|
|
|
} else { // Color
|
|
|
|
width = src->uv_crop_width;
|
|
|
|
height = src->uv_crop_height;
|
|
|
|
}
|
|
|
|
ntiles = av1_get_rest_ntiles(width, height, &tile_width, &tile_height,
|
|
|
|
&nhtiles, &nvtiles);
|
2016-09-09 01:15:17 +03:00
|
|
|
(void)ntiles;
|
|
|
|
|
2017-01-04 02:14:57 +03:00
|
|
|
av1_loop_restoration_frame(cm->frame_to_show, cm, rsi, components_pattern,
|
|
|
|
partial_frame, dst_frame);
|
2016-09-09 01:15:17 +03:00
|
|
|
av1_get_rest_tile_limits(tile_idx, subtile_idx, subtile_bits, nhtiles,
|
2017-01-06 00:58:16 +03:00
|
|
|
nvtiles, tile_width, tile_height, width, height, 0,
|
|
|
|
0, &h_start, &h_end, &v_start, &v_end);
|
2016-12-08 14:25:47 +03:00
|
|
|
filt_err = sse_restoration_tile(src, dst_frame, cm, h_start, h_end - h_start,
|
2017-01-04 02:14:57 +03:00
|
|
|
v_start, v_end - v_start, components_pattern);
|
2016-09-09 01:15:17 +03:00
|
|
|
|
|
|
|
return filt_err;
|
|
|
|
}
|
|
|
|
|
|
|
|
static int64_t try_restoration_frame(const YV12_BUFFER_CONFIG *src,
|
2016-08-31 00:01:10 +03:00
|
|
|
AV1_COMP *const cpi, RestorationInfo *rsi,
|
2017-01-04 02:14:57 +03:00
|
|
|
int components_pattern, int partial_frame,
|
2016-12-08 14:25:47 +03:00
|
|
|
YV12_BUFFER_CONFIG *dst_frame) {
|
2016-08-31 00:01:10 +03:00
|
|
|
AV1_COMMON *const cm = &cpi->common;
|
2016-03-11 09:56:24 +03:00
|
|
|
int64_t filt_err;
|
2017-01-04 02:14:57 +03:00
|
|
|
av1_loop_restoration_frame(cm->frame_to_show, cm, rsi, components_pattern,
|
|
|
|
partial_frame, dst_frame);
|
|
|
|
filt_err = sse_restoration_frame(src, dst_frame, components_pattern);
|
2016-02-09 23:24:33 +03:00
|
|
|
return filt_err;
|
|
|
|
}
|
|
|
|
|
2016-12-21 13:44:52 +03:00
|
|
|
static int64_t get_pixel_proj_error(uint8_t *src8, int width, int height,
|
|
|
|
int src_stride, uint8_t *dat8,
|
|
|
|
int dat_stride, int bit_depth,
|
|
|
|
int32_t *flt1, int flt1_stride,
|
|
|
|
int32_t *flt2, int flt2_stride, int *xqd) {
|
2016-10-12 20:47:01 +03:00
|
|
|
int i, j;
|
|
|
|
int64_t err = 0;
|
|
|
|
int xq[2];
|
|
|
|
decode_xq(xqd, xq);
|
2016-12-21 13:44:52 +03:00
|
|
|
if (bit_depth == 8) {
|
|
|
|
const uint8_t *src = src8;
|
|
|
|
const uint8_t *dat = dat8;
|
|
|
|
for (i = 0; i < height; ++i) {
|
|
|
|
for (j = 0; j < width; ++j) {
|
|
|
|
const int32_t u =
|
|
|
|
(int32_t)(dat[i * dat_stride + j] << SGRPROJ_RST_BITS);
|
|
|
|
const int32_t f1 = (int32_t)flt1[i * flt1_stride + j] - u;
|
|
|
|
const int32_t f2 = (int32_t)flt2[i * flt2_stride + j] - u;
|
|
|
|
const int64_t v = xq[0] * f1 + xq[1] * f2 + (u << SGRPROJ_PRJ_BITS);
|
|
|
|
const int32_t e =
|
|
|
|
ROUND_POWER_OF_TWO(v, SGRPROJ_RST_BITS + SGRPROJ_PRJ_BITS) -
|
|
|
|
src[i * src_stride + j];
|
|
|
|
err += e * e;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
const uint16_t *src = CONVERT_TO_SHORTPTR(src8);
|
|
|
|
const uint16_t *dat = CONVERT_TO_SHORTPTR(dat8);
|
|
|
|
for (i = 0; i < height; ++i) {
|
|
|
|
for (j = 0; j < width; ++j) {
|
|
|
|
const int32_t u =
|
|
|
|
(int32_t)(dat[i * dat_stride + j] << SGRPROJ_RST_BITS);
|
|
|
|
const int32_t f1 = (int32_t)flt1[i * flt1_stride + j] - u;
|
|
|
|
const int32_t f2 = (int32_t)flt2[i * flt2_stride + j] - u;
|
|
|
|
const int64_t v = xq[0] * f1 + xq[1] * f2 + (u << SGRPROJ_PRJ_BITS);
|
|
|
|
const int32_t e =
|
|
|
|
ROUND_POWER_OF_TWO(v, SGRPROJ_RST_BITS + SGRPROJ_PRJ_BITS) -
|
|
|
|
src[i * src_stride + j];
|
|
|
|
err += e * e;
|
|
|
|
}
|
2016-10-12 20:47:01 +03:00
|
|
|
}
|
|
|
|
}
|
|
|
|
return err;
|
|
|
|
}
|
|
|
|
|
2016-12-21 13:44:52 +03:00
|
|
|
static void get_proj_subspace(uint8_t *src8, int width, int height,
|
|
|
|
int src_stride, uint8_t *dat8, int dat_stride,
|
|
|
|
int bit_depth, int32_t *flt1, int flt1_stride,
|
|
|
|
int32_t *flt2, int flt2_stride, int *xq) {
|
2016-10-12 20:47:01 +03:00
|
|
|
int i, j;
|
|
|
|
double H[2][2] = { { 0, 0 }, { 0, 0 } };
|
|
|
|
double C[2] = { 0, 0 };
|
|
|
|
double Det;
|
|
|
|
double x[2];
|
|
|
|
const int size = width * height;
|
|
|
|
|
|
|
|
xq[0] = -(1 << SGRPROJ_PRJ_BITS) / 4;
|
|
|
|
xq[1] = (1 << SGRPROJ_PRJ_BITS) - xq[0];
|
2016-12-21 13:44:52 +03:00
|
|
|
if (bit_depth == 8) {
|
|
|
|
const uint8_t *src = src8;
|
|
|
|
const uint8_t *dat = dat8;
|
|
|
|
for (i = 0; i < height; ++i) {
|
|
|
|
for (j = 0; j < width; ++j) {
|
|
|
|
const double u = (double)(dat[i * dat_stride + j] << SGRPROJ_RST_BITS);
|
|
|
|
const double s =
|
|
|
|
(double)(src[i * src_stride + j] << SGRPROJ_RST_BITS) - u;
|
|
|
|
const double f1 = (double)flt1[i * flt1_stride + j] - u;
|
|
|
|
const double f2 = (double)flt2[i * flt2_stride + j] - u;
|
|
|
|
H[0][0] += f1 * f1;
|
|
|
|
H[1][1] += f2 * f2;
|
|
|
|
H[0][1] += f1 * f2;
|
|
|
|
C[0] += f1 * s;
|
|
|
|
C[1] += f2 * s;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
const uint16_t *src = CONVERT_TO_SHORTPTR(src8);
|
|
|
|
const uint16_t *dat = CONVERT_TO_SHORTPTR(dat8);
|
|
|
|
for (i = 0; i < height; ++i) {
|
|
|
|
for (j = 0; j < width; ++j) {
|
|
|
|
const double u = (double)(dat[i * dat_stride + j] << SGRPROJ_RST_BITS);
|
|
|
|
const double s =
|
|
|
|
(double)(src[i * src_stride + j] << SGRPROJ_RST_BITS) - u;
|
|
|
|
const double f1 = (double)flt1[i * flt1_stride + j] - u;
|
|
|
|
const double f2 = (double)flt2[i * flt2_stride + j] - u;
|
|
|
|
H[0][0] += f1 * f1;
|
|
|
|
H[1][1] += f2 * f2;
|
|
|
|
H[0][1] += f1 * f2;
|
|
|
|
C[0] += f1 * s;
|
|
|
|
C[1] += f2 * s;
|
|
|
|
}
|
2016-10-12 20:47:01 +03:00
|
|
|
}
|
|
|
|
}
|
|
|
|
H[0][0] /= size;
|
|
|
|
H[0][1] /= size;
|
|
|
|
H[1][1] /= size;
|
|
|
|
H[1][0] = H[0][1];
|
|
|
|
C[0] /= size;
|
|
|
|
C[1] /= size;
|
|
|
|
Det = (H[0][0] * H[1][1] - H[0][1] * H[1][0]);
|
|
|
|
if (Det < 1e-8) return; // ill-posed, return default values
|
|
|
|
x[0] = (H[1][1] * C[0] - H[0][1] * C[1]) / Det;
|
|
|
|
x[1] = (H[0][0] * C[1] - H[1][0] * C[0]) / Det;
|
|
|
|
xq[0] = (int)rint(x[0] * (1 << SGRPROJ_PRJ_BITS));
|
|
|
|
xq[1] = (int)rint(x[1] * (1 << SGRPROJ_PRJ_BITS));
|
|
|
|
}
|
|
|
|
|
|
|
|
void encode_xq(int *xq, int *xqd) {
|
|
|
|
xqd[0] = -xq[0];
|
|
|
|
xqd[0] = clamp(xqd[0], SGRPROJ_PRJ_MIN0, SGRPROJ_PRJ_MAX0);
|
|
|
|
xqd[1] = (1 << SGRPROJ_PRJ_BITS) + xqd[0] - xq[1];
|
|
|
|
xqd[1] = clamp(xqd[1], SGRPROJ_PRJ_MIN1, SGRPROJ_PRJ_MAX1);
|
|
|
|
}
|
|
|
|
|
|
|
|
static void search_selfguided_restoration(uint8_t *dat8, int width, int height,
|
|
|
|
int dat_stride, uint8_t *src8,
|
|
|
|
int src_stride, int bit_depth,
|
2016-12-21 13:44:52 +03:00
|
|
|
int *eps, int *xqd, int32_t *rstbuf) {
|
|
|
|
int32_t *flt1 = rstbuf;
|
2016-12-16 14:13:02 +03:00
|
|
|
int32_t *flt2 = flt1 + RESTORATION_TILEPELS_MAX;
|
2016-12-21 13:44:52 +03:00
|
|
|
int32_t *tmpbuf2 = flt2 + RESTORATION_TILEPELS_MAX;
|
2016-10-12 20:47:01 +03:00
|
|
|
int i, j, ep, bestep = 0;
|
|
|
|
int64_t err, besterr = -1;
|
|
|
|
int exqd[2], bestxqd[2] = { 0, 0 };
|
2016-12-21 13:44:52 +03:00
|
|
|
|
2016-10-12 20:47:01 +03:00
|
|
|
for (ep = 0; ep < SGRPROJ_PARAMS; ep++) {
|
|
|
|
int exq[2];
|
|
|
|
if (bit_depth > 8) {
|
|
|
|
uint16_t *dat = CONVERT_TO_SHORTPTR(dat8);
|
|
|
|
for (i = 0; i < height; ++i) {
|
|
|
|
for (j = 0; j < width; ++j) {
|
2016-12-16 14:13:02 +03:00
|
|
|
flt1[i * width + j] = (int32_t)dat[i * dat_stride + j];
|
|
|
|
flt2[i * width + j] = (int32_t)dat[i * dat_stride + j];
|
2016-10-12 20:47:01 +03:00
|
|
|
}
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
uint8_t *dat = dat8;
|
|
|
|
for (i = 0; i < height; ++i) {
|
|
|
|
for (j = 0; j < width; ++j) {
|
|
|
|
const int k = i * width + j;
|
|
|
|
const int l = i * dat_stride + j;
|
2016-12-16 14:13:02 +03:00
|
|
|
flt1[k] = (int32_t)dat[l];
|
|
|
|
flt2[k] = (int32_t)dat[l];
|
2016-10-12 20:47:01 +03:00
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
av1_selfguided_restoration(flt1, width, height, width, bit_depth,
|
|
|
|
sgr_params[ep].r1, sgr_params[ep].e1, tmpbuf2);
|
|
|
|
av1_selfguided_restoration(flt2, width, height, width, bit_depth,
|
|
|
|
sgr_params[ep].r2, sgr_params[ep].e2, tmpbuf2);
|
2016-12-21 13:44:52 +03:00
|
|
|
get_proj_subspace(src8, width, height, src_stride, dat8, dat_stride,
|
|
|
|
bit_depth, flt1, width, flt2, width, exq);
|
2016-10-12 20:47:01 +03:00
|
|
|
encode_xq(exq, exqd);
|
2016-12-21 13:44:52 +03:00
|
|
|
err =
|
|
|
|
get_pixel_proj_error(src8, width, height, src_stride, dat8, dat_stride,
|
|
|
|
bit_depth, flt1, width, flt2, width, exqd);
|
2016-10-12 20:47:01 +03:00
|
|
|
if (besterr == -1 || err < besterr) {
|
|
|
|
bestep = ep;
|
|
|
|
besterr = err;
|
|
|
|
bestxqd[0] = exqd[0];
|
|
|
|
bestxqd[1] = exqd[1];
|
|
|
|
}
|
|
|
|
}
|
|
|
|
*eps = bestep;
|
|
|
|
xqd[0] = bestxqd[0];
|
|
|
|
xqd[1] = bestxqd[1];
|
|
|
|
}
|
|
|
|
|
|
|
|
static double search_sgrproj(const YV12_BUFFER_CONFIG *src, AV1_COMP *cpi,
|
|
|
|
int filter_level, int partial_frame,
|
2017-01-06 22:18:23 +03:00
|
|
|
RestorationInfo *info, RestorationType *type,
|
|
|
|
double *best_tile_cost,
|
2016-12-08 14:25:47 +03:00
|
|
|
YV12_BUFFER_CONFIG *dst_frame) {
|
2016-10-12 20:47:01 +03:00
|
|
|
SgrprojInfo *sgrproj_info = info->sgrproj_info;
|
|
|
|
double err, cost_norestore, cost_sgrproj;
|
|
|
|
int bits;
|
|
|
|
MACROBLOCK *x = &cpi->td.mb;
|
|
|
|
AV1_COMMON *const cm = &cpi->common;
|
|
|
|
const YV12_BUFFER_CONFIG *dgd = cm->frame_to_show;
|
2017-01-04 02:14:57 +03:00
|
|
|
RestorationInfo *rsi = &cpi->rst_search[0];
|
2016-10-12 20:47:01 +03:00
|
|
|
int tile_idx, tile_width, tile_height, nhtiles, nvtiles;
|
|
|
|
int h_start, h_end, v_start, v_end;
|
2016-12-15 03:53:17 +03:00
|
|
|
// Allocate for the src buffer at high precision
|
2016-10-12 20:47:01 +03:00
|
|
|
const int ntiles = av1_get_rest_ntiles(cm->width, cm->height, &tile_width,
|
|
|
|
&tile_height, &nhtiles, &nvtiles);
|
|
|
|
// Make a copy of the unfiltered / processed recon buffer
|
|
|
|
aom_yv12_copy_y(cm->frame_to_show, &cpi->last_frame_uf);
|
|
|
|
av1_loop_filter_frame(cm->frame_to_show, cm, &cpi->td.mb.e_mbd, filter_level,
|
|
|
|
1, partial_frame);
|
|
|
|
aom_yv12_copy_y(cm->frame_to_show, &cpi->last_frame_db);
|
|
|
|
|
2016-12-16 00:23:21 +03:00
|
|
|
rsi->frame_restoration_type = RESTORE_SGRPROJ;
|
2016-10-12 20:47:01 +03:00
|
|
|
|
2017-01-06 22:18:23 +03:00
|
|
|
for (tile_idx = 0; tile_idx < ntiles; ++tile_idx) {
|
|
|
|
rsi->restoration_type[tile_idx] = RESTORE_NONE;
|
|
|
|
}
|
2016-10-12 20:47:01 +03:00
|
|
|
// Compute best Sgrproj filters for each tile
|
|
|
|
for (tile_idx = 0; tile_idx < ntiles; ++tile_idx) {
|
|
|
|
av1_get_rest_tile_limits(tile_idx, 0, 0, nhtiles, nvtiles, tile_width,
|
|
|
|
tile_height, cm->width, cm->height, 0, 0, &h_start,
|
|
|
|
&h_end, &v_start, &v_end);
|
2016-12-08 14:25:47 +03:00
|
|
|
err = sse_restoration_tile(src, cm->frame_to_show, cm, h_start,
|
2016-12-15 03:53:17 +03:00
|
|
|
h_end - h_start, v_start, v_end - v_start, 1);
|
2016-10-12 20:47:01 +03:00
|
|
|
// #bits when a tile is not restored
|
|
|
|
bits = av1_cost_bit(RESTORE_NONE_SGRPROJ_PROB, 0);
|
|
|
|
cost_norestore = RDCOST_DBL(x->rdmult, x->rddiv, (bits >> 4), err);
|
|
|
|
best_tile_cost[tile_idx] = DBL_MAX;
|
|
|
|
search_selfguided_restoration(
|
|
|
|
dgd->y_buffer + v_start * dgd->y_stride + h_start, h_end - h_start,
|
|
|
|
v_end - v_start, dgd->y_stride,
|
|
|
|
src->y_buffer + v_start * src->y_stride + h_start, src->y_stride,
|
|
|
|
#if CONFIG_AOM_HIGHBITDEPTH
|
|
|
|
cm->bit_depth,
|
|
|
|
#else
|
|
|
|
8,
|
|
|
|
#endif // CONFIG_AOM_HIGHBITDEPTH
|
2016-12-16 00:23:21 +03:00
|
|
|
&rsi->sgrproj_info[tile_idx].ep, rsi->sgrproj_info[tile_idx].xqd,
|
2016-12-21 13:44:52 +03:00
|
|
|
cm->rst_internal.tmpbuf);
|
2017-01-06 22:18:23 +03:00
|
|
|
rsi->restoration_type[tile_idx] = RESTORE_SGRPROJ;
|
2016-12-16 00:23:21 +03:00
|
|
|
err = try_restoration_tile(src, cpi, rsi, 1, partial_frame, tile_idx, 0, 0,
|
2016-12-08 14:25:47 +03:00
|
|
|
dst_frame);
|
2016-10-12 20:47:01 +03:00
|
|
|
bits = SGRPROJ_BITS << AV1_PROB_COST_SHIFT;
|
|
|
|
bits += av1_cost_bit(RESTORE_NONE_SGRPROJ_PROB, 1);
|
|
|
|
cost_sgrproj = RDCOST_DBL(x->rdmult, x->rddiv, (bits >> 4), err);
|
|
|
|
if (cost_sgrproj >= cost_norestore) {
|
2017-01-06 22:18:23 +03:00
|
|
|
type[tile_idx] = RESTORE_NONE;
|
2016-10-12 20:47:01 +03:00
|
|
|
} else {
|
2017-01-06 22:18:23 +03:00
|
|
|
type[tile_idx] = RESTORE_SGRPROJ;
|
2016-12-16 00:23:21 +03:00
|
|
|
memcpy(&sgrproj_info[tile_idx], &rsi->sgrproj_info[tile_idx],
|
2016-10-12 20:47:01 +03:00
|
|
|
sizeof(sgrproj_info[tile_idx]));
|
|
|
|
bits = SGRPROJ_BITS << AV1_PROB_COST_SHIFT;
|
|
|
|
best_tile_cost[tile_idx] = RDCOST_DBL(
|
|
|
|
x->rdmult, x->rddiv,
|
|
|
|
(bits + cpi->switchable_restore_cost[RESTORE_SGRPROJ]) >> 4, err);
|
|
|
|
}
|
2017-01-06 22:18:23 +03:00
|
|
|
rsi->restoration_type[tile_idx] = RESTORE_NONE;
|
2016-10-12 20:47:01 +03:00
|
|
|
}
|
|
|
|
// Cost for Sgrproj filtering
|
2016-12-16 00:23:21 +03:00
|
|
|
bits = frame_level_restore_bits[rsi->frame_restoration_type]
|
2016-10-12 20:47:01 +03:00
|
|
|
<< AV1_PROB_COST_SHIFT;
|
|
|
|
for (tile_idx = 0; tile_idx < ntiles; ++tile_idx) {
|
|
|
|
bits +=
|
2017-01-06 22:18:23 +03:00
|
|
|
av1_cost_bit(RESTORE_NONE_SGRPROJ_PROB, type[tile_idx] != RESTORE_NONE);
|
2016-12-16 00:23:21 +03:00
|
|
|
memcpy(&rsi->sgrproj_info[tile_idx], &sgrproj_info[tile_idx],
|
2016-10-12 20:47:01 +03:00
|
|
|
sizeof(sgrproj_info[tile_idx]));
|
2017-01-06 22:18:23 +03:00
|
|
|
if (type[tile_idx] == RESTORE_SGRPROJ) {
|
2016-10-12 20:47:01 +03:00
|
|
|
bits += (SGRPROJ_BITS << AV1_PROB_COST_SHIFT);
|
|
|
|
}
|
2017-01-06 22:18:23 +03:00
|
|
|
rsi->restoration_type[tile_idx] = type[tile_idx];
|
2016-10-12 20:47:01 +03:00
|
|
|
}
|
2016-12-16 00:23:21 +03:00
|
|
|
err = try_restoration_frame(src, cpi, rsi, 1, partial_frame, dst_frame);
|
2016-10-12 20:47:01 +03:00
|
|
|
cost_sgrproj = RDCOST_DBL(x->rdmult, x->rddiv, (bits >> 4), err);
|
|
|
|
|
|
|
|
aom_yv12_copy_y(&cpi->last_frame_uf, cm->frame_to_show);
|
|
|
|
return cost_sgrproj;
|
|
|
|
}
|
|
|
|
|
2016-11-21 20:35:44 +03:00
|
|
|
static int64_t compute_sse(uint8_t *dgd, int width, int height, int dgd_stride,
|
|
|
|
uint8_t *src, int src_stride) {
|
|
|
|
int64_t sse = 0;
|
|
|
|
int i, j;
|
|
|
|
for (i = 0; i < height; ++i) {
|
|
|
|
for (j = 0; j < width; ++j) {
|
|
|
|
const int diff =
|
|
|
|
(int)dgd[i * dgd_stride + j] - (int)src[i * src_stride + j];
|
|
|
|
sse += diff * diff;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
return sse;
|
|
|
|
}
|
|
|
|
|
|
|
|
#if CONFIG_AOM_HIGHBITDEPTH
|
|
|
|
static int64_t compute_sse_highbd(uint16_t *dgd, int width, int height,
|
|
|
|
int dgd_stride, uint16_t *src,
|
|
|
|
int src_stride) {
|
|
|
|
int64_t sse = 0;
|
|
|
|
int i, j;
|
|
|
|
for (i = 0; i < height; ++i) {
|
|
|
|
for (j = 0; j < width; ++j) {
|
|
|
|
const int diff =
|
|
|
|
(int)dgd[i * dgd_stride + j] - (int)src[i * src_stride + j];
|
|
|
|
sse += diff * diff;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
return sse;
|
|
|
|
}
|
|
|
|
#endif // CONFIG_AOM_HIGHBITDEPTH
|
|
|
|
|
|
|
|
static void search_domaintxfmrf_restoration(uint8_t *dgd8, int width,
|
|
|
|
int height, int dgd_stride,
|
|
|
|
uint8_t *src8, int src_stride,
|
2016-12-16 00:23:21 +03:00
|
|
|
int bit_depth, int *sigma_r,
|
2016-12-21 13:44:52 +03:00
|
|
|
uint8_t *fltbuf, int32_t *tmpbuf) {
|
2016-11-21 20:35:44 +03:00
|
|
|
const int first_p_step = 8;
|
|
|
|
const int second_p_range = first_p_step >> 1;
|
|
|
|
const int second_p_step = 2;
|
|
|
|
const int third_p_range = second_p_step >> 1;
|
|
|
|
const int third_p_step = 1;
|
2016-12-08 14:25:47 +03:00
|
|
|
int p, best_p0, best_p = -1;
|
2016-11-21 20:35:44 +03:00
|
|
|
int64_t best_sse = INT64_MAX, sse;
|
|
|
|
if (bit_depth == 8) {
|
2016-12-21 13:44:52 +03:00
|
|
|
uint8_t *flt = fltbuf;
|
2016-11-21 20:35:44 +03:00
|
|
|
uint8_t *dgd = dgd8;
|
|
|
|
uint8_t *src = src8;
|
|
|
|
// First phase
|
|
|
|
for (p = first_p_step / 2; p < DOMAINTXFMRF_PARAMS; p += first_p_step) {
|
2016-12-16 00:23:21 +03:00
|
|
|
av1_domaintxfmrf_restoration(dgd, width, height, dgd_stride, p, flt,
|
2016-12-15 06:20:46 +03:00
|
|
|
width, tmpbuf);
|
2016-12-16 00:23:21 +03:00
|
|
|
sse = compute_sse(flt, width, height, width, src, src_stride);
|
2016-11-21 20:35:44 +03:00
|
|
|
if (sse < best_sse || best_p == -1) {
|
|
|
|
best_p = p;
|
|
|
|
best_sse = sse;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
// Second Phase
|
|
|
|
best_p0 = best_p;
|
|
|
|
for (p = best_p0 - second_p_range; p <= best_p0 + second_p_range;
|
|
|
|
p += second_p_step) {
|
|
|
|
if (p < 0 || p == best_p || p >= DOMAINTXFMRF_PARAMS) continue;
|
2016-12-16 00:23:21 +03:00
|
|
|
av1_domaintxfmrf_restoration(dgd, width, height, dgd_stride, p, flt,
|
2016-12-15 06:20:46 +03:00
|
|
|
width, tmpbuf);
|
2016-12-16 00:23:21 +03:00
|
|
|
sse = compute_sse(flt, width, height, width, src, src_stride);
|
2016-11-21 20:35:44 +03:00
|
|
|
if (sse < best_sse) {
|
|
|
|
best_p = p;
|
|
|
|
best_sse = sse;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
// Third Phase
|
|
|
|
best_p0 = best_p;
|
|
|
|
for (p = best_p0 - third_p_range; p <= best_p0 + third_p_range;
|
|
|
|
p += third_p_step) {
|
|
|
|
if (p < 0 || p == best_p || p >= DOMAINTXFMRF_PARAMS) continue;
|
2016-12-16 00:23:21 +03:00
|
|
|
av1_domaintxfmrf_restoration(dgd, width, height, dgd_stride, p, flt,
|
2016-12-15 06:20:46 +03:00
|
|
|
width, tmpbuf);
|
2016-12-16 00:23:21 +03:00
|
|
|
sse = compute_sse(flt, width, height, width, src, src_stride);
|
2016-11-21 20:35:44 +03:00
|
|
|
if (sse < best_sse) {
|
|
|
|
best_p = p;
|
|
|
|
best_sse = sse;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
#if CONFIG_AOM_HIGHBITDEPTH
|
2016-12-16 00:23:21 +03:00
|
|
|
uint16_t *flt = (uint16_t *)fltbuf;
|
2016-11-21 20:35:44 +03:00
|
|
|
uint16_t *dgd = CONVERT_TO_SHORTPTR(dgd8);
|
|
|
|
uint16_t *src = CONVERT_TO_SHORTPTR(src8);
|
|
|
|
// First phase
|
|
|
|
for (p = first_p_step / 2; p < DOMAINTXFMRF_PARAMS; p += first_p_step) {
|
2016-12-08 14:25:47 +03:00
|
|
|
av1_domaintxfmrf_restoration_highbd(dgd, width, height, dgd_stride, p,
|
2016-12-16 00:23:21 +03:00
|
|
|
bit_depth, flt, width, tmpbuf);
|
|
|
|
sse = compute_sse_highbd(flt, width, height, width, src, src_stride);
|
2016-11-21 20:35:44 +03:00
|
|
|
if (sse < best_sse || best_p == -1) {
|
|
|
|
best_p = p;
|
|
|
|
best_sse = sse;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
// Second Phase
|
|
|
|
best_p0 = best_p;
|
|
|
|
for (p = best_p0 - second_p_range; p <= best_p0 + second_p_range;
|
|
|
|
p += second_p_step) {
|
|
|
|
if (p < 0 || p == best_p || p >= DOMAINTXFMRF_PARAMS) continue;
|
2016-12-08 14:25:47 +03:00
|
|
|
av1_domaintxfmrf_restoration_highbd(dgd, width, height, dgd_stride, p,
|
2016-12-16 00:23:21 +03:00
|
|
|
bit_depth, flt, width, tmpbuf);
|
|
|
|
sse = compute_sse_highbd(flt, width, height, width, src, src_stride);
|
2016-11-21 20:35:44 +03:00
|
|
|
if (sse < best_sse) {
|
|
|
|
best_p = p;
|
|
|
|
best_sse = sse;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
// Third Phase
|
|
|
|
best_p0 = best_p;
|
|
|
|
for (p = best_p0 - third_p_range; p <= best_p0 + third_p_range;
|
|
|
|
p += third_p_step) {
|
|
|
|
if (p < 0 || p == best_p || p >= DOMAINTXFMRF_PARAMS) continue;
|
2016-12-08 14:25:47 +03:00
|
|
|
av1_domaintxfmrf_restoration_highbd(dgd, width, height, dgd_stride, p,
|
2016-12-16 00:23:21 +03:00
|
|
|
bit_depth, flt, width, tmpbuf);
|
|
|
|
sse = compute_sse_highbd(flt, width, height, width, src, src_stride);
|
2016-11-21 20:35:44 +03:00
|
|
|
if (sse < best_sse) {
|
|
|
|
best_p = p;
|
|
|
|
best_sse = sse;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
|
|
assert(0);
|
|
|
|
#endif // CONFIG_AOM_HIGHBITDEPTH
|
|
|
|
}
|
|
|
|
*sigma_r = best_p;
|
|
|
|
}
|
|
|
|
|
|
|
|
static double search_domaintxfmrf(const YV12_BUFFER_CONFIG *src, AV1_COMP *cpi,
|
|
|
|
int filter_level, int partial_frame,
|
2017-01-06 22:18:23 +03:00
|
|
|
RestorationInfo *info, RestorationType *type,
|
|
|
|
double *best_tile_cost,
|
2016-12-08 14:25:47 +03:00
|
|
|
YV12_BUFFER_CONFIG *dst_frame) {
|
2016-11-21 20:35:44 +03:00
|
|
|
DomaintxfmrfInfo *domaintxfmrf_info = info->domaintxfmrf_info;
|
2016-12-08 20:22:44 +03:00
|
|
|
double cost_norestore, cost_domaintxfmrf;
|
|
|
|
int64_t err;
|
2016-11-21 20:35:44 +03:00
|
|
|
int bits;
|
|
|
|
MACROBLOCK *x = &cpi->td.mb;
|
|
|
|
AV1_COMMON *const cm = &cpi->common;
|
|
|
|
const YV12_BUFFER_CONFIG *dgd = cm->frame_to_show;
|
2017-01-04 02:14:57 +03:00
|
|
|
RestorationInfo *rsi = &cpi->rst_search[0];
|
2016-11-21 20:35:44 +03:00
|
|
|
int tile_idx, tile_width, tile_height, nhtiles, nvtiles;
|
|
|
|
int h_start, h_end, v_start, v_end;
|
|
|
|
const int ntiles = av1_get_rest_ntiles(cm->width, cm->height, &tile_width,
|
|
|
|
&tile_height, &nhtiles, &nvtiles);
|
|
|
|
// Make a copy of the unfiltered / processed recon buffer
|
|
|
|
aom_yv12_copy_y(cm->frame_to_show, &cpi->last_frame_uf);
|
|
|
|
av1_loop_filter_frame(cm->frame_to_show, cm, &cpi->td.mb.e_mbd, filter_level,
|
|
|
|
1, partial_frame);
|
|
|
|
aom_yv12_copy_y(cm->frame_to_show, &cpi->last_frame_db);
|
|
|
|
|
2016-12-16 00:23:21 +03:00
|
|
|
rsi->frame_restoration_type = RESTORE_DOMAINTXFMRF;
|
2016-11-21 20:35:44 +03:00
|
|
|
|
2017-01-06 22:18:23 +03:00
|
|
|
for (tile_idx = 0; tile_idx < ntiles; ++tile_idx) {
|
|
|
|
rsi->restoration_type[tile_idx] = RESTORE_NONE;
|
|
|
|
}
|
2016-11-21 20:35:44 +03:00
|
|
|
// Compute best Domaintxfm filters for each tile
|
|
|
|
for (tile_idx = 0; tile_idx < ntiles; ++tile_idx) {
|
|
|
|
av1_get_rest_tile_limits(tile_idx, 0, 0, nhtiles, nvtiles, tile_width,
|
|
|
|
tile_height, cm->width, cm->height, 0, 0, &h_start,
|
|
|
|
&h_end, &v_start, &v_end);
|
2016-12-08 14:25:47 +03:00
|
|
|
err = sse_restoration_tile(src, cm->frame_to_show, cm, h_start,
|
2016-12-15 03:53:17 +03:00
|
|
|
h_end - h_start, v_start, v_end - v_start, 1);
|
2016-11-21 20:35:44 +03:00
|
|
|
// #bits when a tile is not restored
|
|
|
|
bits = av1_cost_bit(RESTORE_NONE_DOMAINTXFMRF_PROB, 0);
|
|
|
|
cost_norestore = RDCOST_DBL(x->rdmult, x->rddiv, (bits >> 4), err);
|
|
|
|
best_tile_cost[tile_idx] = DBL_MAX;
|
|
|
|
|
|
|
|
search_domaintxfmrf_restoration(
|
|
|
|
dgd->y_buffer + v_start * dgd->y_stride + h_start, h_end - h_start,
|
|
|
|
v_end - v_start, dgd->y_stride,
|
|
|
|
src->y_buffer + v_start * src->y_stride + h_start, src->y_stride,
|
|
|
|
#if CONFIG_AOM_HIGHBITDEPTH
|
|
|
|
cm->bit_depth,
|
|
|
|
#else
|
|
|
|
8,
|
|
|
|
#endif // CONFIG_AOM_HIGHBITDEPTH
|
2016-12-16 00:23:21 +03:00
|
|
|
&rsi->domaintxfmrf_info[tile_idx].sigma_r, cpi->extra_rstbuf,
|
|
|
|
cm->rst_internal.tmpbuf);
|
2016-11-21 20:35:44 +03:00
|
|
|
|
2017-01-06 22:18:23 +03:00
|
|
|
rsi->restoration_type[tile_idx] = RESTORE_DOMAINTXFMRF;
|
2016-12-16 00:23:21 +03:00
|
|
|
err = try_restoration_tile(src, cpi, rsi, 1, partial_frame, tile_idx, 0, 0,
|
2016-12-08 14:25:47 +03:00
|
|
|
dst_frame);
|
2016-11-21 20:35:44 +03:00
|
|
|
bits = DOMAINTXFMRF_PARAMS_BITS << AV1_PROB_COST_SHIFT;
|
|
|
|
bits += av1_cost_bit(RESTORE_NONE_DOMAINTXFMRF_PROB, 1);
|
|
|
|
cost_domaintxfmrf = RDCOST_DBL(x->rdmult, x->rddiv, (bits >> 4), err);
|
|
|
|
if (cost_domaintxfmrf >= cost_norestore) {
|
2017-01-06 22:18:23 +03:00
|
|
|
type[tile_idx] = RESTORE_NONE;
|
2016-11-21 20:35:44 +03:00
|
|
|
} else {
|
2017-01-06 22:18:23 +03:00
|
|
|
type[tile_idx] = RESTORE_DOMAINTXFMRF;
|
2016-12-16 00:23:21 +03:00
|
|
|
memcpy(&domaintxfmrf_info[tile_idx], &rsi->domaintxfmrf_info[tile_idx],
|
2016-11-21 20:35:44 +03:00
|
|
|
sizeof(domaintxfmrf_info[tile_idx]));
|
|
|
|
bits = DOMAINTXFMRF_PARAMS_BITS << AV1_PROB_COST_SHIFT;
|
|
|
|
best_tile_cost[tile_idx] = RDCOST_DBL(
|
|
|
|
x->rdmult, x->rddiv,
|
|
|
|
(bits + cpi->switchable_restore_cost[RESTORE_DOMAINTXFMRF]) >> 4,
|
|
|
|
err);
|
|
|
|
}
|
2017-01-06 22:18:23 +03:00
|
|
|
rsi->restoration_type[tile_idx] = RESTORE_NONE;
|
2016-11-21 20:35:44 +03:00
|
|
|
}
|
|
|
|
// Cost for Domaintxfmrf filtering
|
2016-12-16 00:23:21 +03:00
|
|
|
bits = frame_level_restore_bits[rsi->frame_restoration_type]
|
2016-11-21 20:35:44 +03:00
|
|
|
<< AV1_PROB_COST_SHIFT;
|
|
|
|
for (tile_idx = 0; tile_idx < ntiles; ++tile_idx) {
|
|
|
|
bits += av1_cost_bit(RESTORE_NONE_DOMAINTXFMRF_PROB,
|
2017-01-06 22:18:23 +03:00
|
|
|
type[tile_idx] != RESTORE_NONE);
|
2016-12-16 00:23:21 +03:00
|
|
|
memcpy(&rsi->domaintxfmrf_info[tile_idx], &domaintxfmrf_info[tile_idx],
|
2016-11-21 20:35:44 +03:00
|
|
|
sizeof(domaintxfmrf_info[tile_idx]));
|
2017-01-06 22:18:23 +03:00
|
|
|
if (type[tile_idx] == RESTORE_DOMAINTXFMRF) {
|
2016-11-21 20:35:44 +03:00
|
|
|
bits += (DOMAINTXFMRF_PARAMS_BITS << AV1_PROB_COST_SHIFT);
|
|
|
|
}
|
2017-01-06 22:18:23 +03:00
|
|
|
rsi->restoration_type[tile_idx] = type[tile_idx];
|
2016-11-21 20:35:44 +03:00
|
|
|
}
|
2016-12-16 00:23:21 +03:00
|
|
|
err = try_restoration_frame(src, cpi, rsi, 1, partial_frame, dst_frame);
|
2016-11-21 20:35:44 +03:00
|
|
|
cost_domaintxfmrf = RDCOST_DBL(x->rdmult, x->rddiv, (bits >> 4), err);
|
|
|
|
|
|
|
|
aom_yv12_copy_y(&cpi->last_frame_uf, cm->frame_to_show);
|
|
|
|
return cost_domaintxfmrf;
|
|
|
|
}
|
|
|
|
|
2016-07-21 07:29:45 +03:00
|
|
|
static double find_average(uint8_t *src, int h_start, int h_end, int v_start,
|
|
|
|
int v_end, int stride) {
|
2016-02-09 23:24:33 +03:00
|
|
|
uint64_t sum = 0;
|
|
|
|
double avg = 0;
|
|
|
|
int i, j;
|
2016-07-21 07:29:45 +03:00
|
|
|
for (i = v_start; i < v_end; i++)
|
|
|
|
for (j = h_start; j < h_end; j++) sum += src[i * stride + j];
|
|
|
|
avg = (double)sum / ((v_end - v_start) * (h_end - h_start));
|
2016-02-09 23:24:33 +03:00
|
|
|
return avg;
|
|
|
|
}
|
|
|
|
|
2016-07-21 07:29:45 +03:00
|
|
|
static void compute_stats(uint8_t *dgd, uint8_t *src, int h_start, int h_end,
|
|
|
|
int v_start, int v_end, int dgd_stride,
|
|
|
|
int src_stride, double *M, double *H) {
|
2016-02-09 23:24:33 +03:00
|
|
|
int i, j, k, l;
|
2016-12-16 00:23:21 +03:00
|
|
|
double Y[WIENER_WIN2];
|
2016-07-21 07:29:45 +03:00
|
|
|
const double avg =
|
|
|
|
find_average(dgd, h_start, h_end, v_start, v_end, dgd_stride);
|
2016-02-09 23:24:33 +03:00
|
|
|
|
2016-12-16 00:23:21 +03:00
|
|
|
memset(M, 0, sizeof(*M) * WIENER_WIN2);
|
|
|
|
memset(H, 0, sizeof(*H) * WIENER_WIN2 * WIENER_WIN2);
|
2016-07-21 07:29:45 +03:00
|
|
|
for (i = v_start; i < v_end; i++) {
|
|
|
|
for (j = h_start; j < h_end; j++) {
|
2016-02-09 23:24:33 +03:00
|
|
|
const double X = (double)src[i * src_stride + j] - avg;
|
|
|
|
int idx = 0;
|
2016-12-16 00:23:21 +03:00
|
|
|
for (k = -WIENER_HALFWIN; k <= WIENER_HALFWIN; k++) {
|
|
|
|
for (l = -WIENER_HALFWIN; l <= WIENER_HALFWIN; l++) {
|
2016-02-09 23:24:33 +03:00
|
|
|
Y[idx] = (double)dgd[(i + l) * dgd_stride + (j + k)] - avg;
|
|
|
|
idx++;
|
|
|
|
}
|
|
|
|
}
|
2016-12-16 00:23:21 +03:00
|
|
|
for (k = 0; k < WIENER_WIN2; ++k) {
|
2016-02-09 23:24:33 +03:00
|
|
|
M[k] += Y[k] * X;
|
2016-12-16 00:23:21 +03:00
|
|
|
H[k * WIENER_WIN2 + k] += Y[k] * Y[k];
|
|
|
|
for (l = k + 1; l < WIENER_WIN2; ++l) {
|
2017-01-06 18:34:50 +03:00
|
|
|
// H is a symmetric matrix, so we only need to fill out the upper
|
|
|
|
// triangle here. We can copy it down to the lower triangle outside
|
|
|
|
// the (i, j) loops.
|
|
|
|
H[k * WIENER_WIN2 + l] += Y[k] * Y[l];
|
2016-02-09 23:24:33 +03:00
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
2017-01-06 18:34:50 +03:00
|
|
|
for (k = 0; k < WIENER_WIN2; ++k) {
|
|
|
|
for (l = k + 1; l < WIENER_WIN2; ++l) {
|
|
|
|
H[l * WIENER_WIN2 + k] = H[k * WIENER_WIN2 + l];
|
|
|
|
}
|
|
|
|
}
|
2016-02-09 23:24:33 +03:00
|
|
|
}
|
|
|
|
|
2016-08-31 00:01:10 +03:00
|
|
|
#if CONFIG_AOM_HIGHBITDEPTH
|
2016-07-21 07:29:45 +03:00
|
|
|
static double find_average_highbd(uint16_t *src, int h_start, int h_end,
|
|
|
|
int v_start, int v_end, int stride) {
|
2016-02-09 23:24:33 +03:00
|
|
|
uint64_t sum = 0;
|
|
|
|
double avg = 0;
|
|
|
|
int i, j;
|
2016-07-21 07:29:45 +03:00
|
|
|
for (i = v_start; i < v_end; i++)
|
|
|
|
for (j = h_start; j < h_end; j++) sum += src[i * stride + j];
|
|
|
|
avg = (double)sum / ((v_end - v_start) * (h_end - h_start));
|
2016-02-09 23:24:33 +03:00
|
|
|
return avg;
|
|
|
|
}
|
|
|
|
|
2016-07-21 07:29:45 +03:00
|
|
|
static void compute_stats_highbd(uint8_t *dgd8, uint8_t *src8, int h_start,
|
|
|
|
int h_end, int v_start, int v_end,
|
|
|
|
int dgd_stride, int src_stride, double *M,
|
|
|
|
double *H) {
|
2016-02-09 23:24:33 +03:00
|
|
|
int i, j, k, l;
|
2016-12-16 00:23:21 +03:00
|
|
|
double Y[WIENER_WIN2];
|
2016-02-09 23:24:33 +03:00
|
|
|
uint16_t *src = CONVERT_TO_SHORTPTR(src8);
|
|
|
|
uint16_t *dgd = CONVERT_TO_SHORTPTR(dgd8);
|
2016-07-21 07:29:45 +03:00
|
|
|
const double avg =
|
|
|
|
find_average_highbd(dgd, h_start, h_end, v_start, v_end, dgd_stride);
|
2016-02-09 23:24:33 +03:00
|
|
|
|
2016-12-16 00:23:21 +03:00
|
|
|
memset(M, 0, sizeof(*M) * WIENER_WIN2);
|
|
|
|
memset(H, 0, sizeof(*H) * WIENER_WIN2 * WIENER_WIN2);
|
2016-07-21 07:29:45 +03:00
|
|
|
for (i = v_start; i < v_end; i++) {
|
|
|
|
for (j = h_start; j < h_end; j++) {
|
2016-02-09 23:24:33 +03:00
|
|
|
const double X = (double)src[i * src_stride + j] - avg;
|
|
|
|
int idx = 0;
|
2016-12-16 00:23:21 +03:00
|
|
|
for (k = -WIENER_HALFWIN; k <= WIENER_HALFWIN; k++) {
|
|
|
|
for (l = -WIENER_HALFWIN; l <= WIENER_HALFWIN; l++) {
|
2016-02-09 23:24:33 +03:00
|
|
|
Y[idx] = (double)dgd[(i + l) * dgd_stride + (j + k)] - avg;
|
|
|
|
idx++;
|
|
|
|
}
|
|
|
|
}
|
2016-12-16 00:23:21 +03:00
|
|
|
for (k = 0; k < WIENER_WIN2; ++k) {
|
2016-02-09 23:24:33 +03:00
|
|
|
M[k] += Y[k] * X;
|
2016-12-16 00:23:21 +03:00
|
|
|
H[k * WIENER_WIN2 + k] += Y[k] * Y[k];
|
|
|
|
for (l = k + 1; l < WIENER_WIN2; ++l) {
|
2017-01-06 18:34:50 +03:00
|
|
|
// H is a symmetric matrix, so we only need to fill out the upper
|
|
|
|
// triangle here. We can copy it down to the lower triangle outside
|
|
|
|
// the (i, j) loops.
|
|
|
|
H[k * WIENER_WIN2 + l] += Y[k] * Y[l];
|
2016-02-09 23:24:33 +03:00
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
2017-01-06 18:34:50 +03:00
|
|
|
for (k = 0; k < WIENER_WIN2; ++k) {
|
|
|
|
for (l = k + 1; l < WIENER_WIN2; ++l) {
|
|
|
|
H[l * WIENER_WIN2 + k] = H[k * WIENER_WIN2 + l];
|
|
|
|
}
|
|
|
|
}
|
2016-02-09 23:24:33 +03:00
|
|
|
}
|
2016-08-31 00:01:10 +03:00
|
|
|
#endif // CONFIG_AOM_HIGHBITDEPTH
|
2016-02-09 23:24:33 +03:00
|
|
|
|
|
|
|
// Solves Ax = b, where x and b are column vectors
|
|
|
|
static int linsolve(int n, double *A, int stride, double *b, double *x) {
|
|
|
|
int i, j, k;
|
|
|
|
double c;
|
|
|
|
// Partial pivoting
|
|
|
|
for (i = n - 1; i > 0; i--) {
|
|
|
|
if (A[(i - 1) * stride] < A[i * stride]) {
|
|
|
|
for (j = 0; j < n; j++) {
|
|
|
|
c = A[i * stride + j];
|
|
|
|
A[i * stride + j] = A[(i - 1) * stride + j];
|
|
|
|
A[(i - 1) * stride + j] = c;
|
|
|
|
}
|
|
|
|
c = b[i];
|
|
|
|
b[i] = b[i - 1];
|
|
|
|
b[i - 1] = c;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
// Forward elimination
|
|
|
|
for (k = 0; k < n - 1; k++) {
|
|
|
|
for (i = k; i < n - 1; i++) {
|
|
|
|
c = A[(i + 1) * stride + k] / A[k * stride + k];
|
2016-08-12 06:13:14 +03:00
|
|
|
for (j = 0; j < n; j++) A[(i + 1) * stride + j] -= c * A[k * stride + j];
|
2016-02-09 23:24:33 +03:00
|
|
|
b[i + 1] -= c * b[k];
|
|
|
|
}
|
|
|
|
}
|
|
|
|
// Backward substitution
|
|
|
|
for (i = n - 1; i >= 0; i--) {
|
2016-08-12 06:13:14 +03:00
|
|
|
if (fabs(A[i * stride + i]) < 1e-10) return 0;
|
2016-02-09 23:24:33 +03:00
|
|
|
c = 0;
|
2016-08-12 06:13:14 +03:00
|
|
|
for (j = i + 1; j <= n - 1; j++) c += A[i * stride + j] * x[j];
|
2016-02-09 23:24:33 +03:00
|
|
|
x[i] = (b[i] - c) / A[i * stride + i];
|
|
|
|
}
|
|
|
|
return 1;
|
|
|
|
}
|
|
|
|
|
|
|
|
static INLINE int wrap_index(int i) {
|
2016-12-16 00:23:21 +03:00
|
|
|
return (i >= WIENER_HALFWIN1 ? WIENER_WIN - 1 - i : i);
|
2016-02-09 23:24:33 +03:00
|
|
|
}
|
|
|
|
|
|
|
|
// Fix vector b, update vector a
|
|
|
|
static void update_a_sep_sym(double **Mc, double **Hc, double *a, double *b) {
|
|
|
|
int i, j;
|
2016-12-16 00:23:21 +03:00
|
|
|
double S[WIENER_WIN];
|
|
|
|
double A[WIENER_WIN], B[WIENER_WIN2];
|
2016-06-01 08:50:24 +03:00
|
|
|
int w, w2;
|
2016-02-09 23:24:33 +03:00
|
|
|
memset(A, 0, sizeof(A));
|
|
|
|
memset(B, 0, sizeof(B));
|
2016-12-16 00:23:21 +03:00
|
|
|
for (i = 0; i < WIENER_WIN; i++) {
|
|
|
|
for (j = 0; j < WIENER_WIN; ++j) {
|
2016-02-09 23:24:33 +03:00
|
|
|
const int jj = wrap_index(j);
|
|
|
|
A[jj] += Mc[i][j] * b[i];
|
|
|
|
}
|
|
|
|
}
|
2016-12-16 00:23:21 +03:00
|
|
|
for (i = 0; i < WIENER_WIN; i++) {
|
|
|
|
for (j = 0; j < WIENER_WIN; j++) {
|
2016-02-09 23:24:33 +03:00
|
|
|
int k, l;
|
2016-12-16 00:23:21 +03:00
|
|
|
for (k = 0; k < WIENER_WIN; ++k)
|
|
|
|
for (l = 0; l < WIENER_WIN; ++l) {
|
2016-02-09 23:24:33 +03:00
|
|
|
const int kk = wrap_index(k);
|
|
|
|
const int ll = wrap_index(l);
|
2016-12-16 00:23:21 +03:00
|
|
|
B[ll * WIENER_HALFWIN1 + kk] +=
|
|
|
|
Hc[j * WIENER_WIN + i][k * WIENER_WIN2 + l] * b[i] * b[j];
|
2016-02-09 23:24:33 +03:00
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
2016-06-01 08:50:24 +03:00
|
|
|
// Normalization enforcement in the system of equations itself
|
2016-12-16 00:23:21 +03:00
|
|
|
w = WIENER_WIN;
|
2016-06-01 08:50:24 +03:00
|
|
|
w2 = (w >> 1) + 1;
|
|
|
|
for (i = 0; i < w2 - 1; ++i)
|
2016-08-12 06:13:14 +03:00
|
|
|
A[i] -=
|
|
|
|
A[w2 - 1] * 2 + B[i * w2 + w2 - 1] - 2 * B[(w2 - 1) * w2 + (w2 - 1)];
|
2016-06-01 08:50:24 +03:00
|
|
|
for (i = 0; i < w2 - 1; ++i)
|
|
|
|
for (j = 0; j < w2 - 1; ++j)
|
|
|
|
B[i * w2 + j] -= 2 * (B[i * w2 + (w2 - 1)] + B[(w2 - 1) * w2 + j] -
|
|
|
|
2 * B[(w2 - 1) * w2 + (w2 - 1)]);
|
|
|
|
if (linsolve(w2 - 1, B, w2, A, S)) {
|
|
|
|
S[w2 - 1] = 1.0;
|
|
|
|
for (i = w2; i < w; ++i) {
|
|
|
|
S[i] = S[w - 1 - i];
|
|
|
|
S[w2 - 1] -= 2 * S[i];
|
2016-02-09 23:24:33 +03:00
|
|
|
}
|
2016-06-01 08:50:24 +03:00
|
|
|
memcpy(a, S, w * sizeof(*a));
|
2016-02-09 23:24:33 +03:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// Fix vector a, update vector b
|
|
|
|
static void update_b_sep_sym(double **Mc, double **Hc, double *a, double *b) {
|
|
|
|
int i, j;
|
2016-12-16 00:23:21 +03:00
|
|
|
double S[WIENER_WIN];
|
|
|
|
double A[WIENER_WIN], B[WIENER_WIN2];
|
2016-06-01 08:50:24 +03:00
|
|
|
int w, w2;
|
2016-02-09 23:24:33 +03:00
|
|
|
memset(A, 0, sizeof(A));
|
|
|
|
memset(B, 0, sizeof(B));
|
2016-12-16 00:23:21 +03:00
|
|
|
for (i = 0; i < WIENER_WIN; i++) {
|
2016-02-09 23:24:33 +03:00
|
|
|
const int ii = wrap_index(i);
|
2016-12-16 00:23:21 +03:00
|
|
|
for (j = 0; j < WIENER_WIN; j++) A[ii] += Mc[i][j] * a[j];
|
2016-02-09 23:24:33 +03:00
|
|
|
}
|
|
|
|
|
2016-12-16 00:23:21 +03:00
|
|
|
for (i = 0; i < WIENER_WIN; i++) {
|
|
|
|
for (j = 0; j < WIENER_WIN; j++) {
|
2016-02-09 23:24:33 +03:00
|
|
|
const int ii = wrap_index(i);
|
|
|
|
const int jj = wrap_index(j);
|
|
|
|
int k, l;
|
2016-12-16 00:23:21 +03:00
|
|
|
for (k = 0; k < WIENER_WIN; ++k)
|
|
|
|
for (l = 0; l < WIENER_WIN; ++l)
|
|
|
|
B[jj * WIENER_HALFWIN1 + ii] +=
|
|
|
|
Hc[i * WIENER_WIN + j][k * WIENER_WIN2 + l] * a[k] * a[l];
|
2016-02-09 23:24:33 +03:00
|
|
|
}
|
|
|
|
}
|
2016-06-01 08:50:24 +03:00
|
|
|
// Normalization enforcement in the system of equations itself
|
2016-12-16 00:23:21 +03:00
|
|
|
w = WIENER_WIN;
|
|
|
|
w2 = WIENER_HALFWIN1;
|
2016-06-01 08:50:24 +03:00
|
|
|
for (i = 0; i < w2 - 1; ++i)
|
2016-08-12 06:13:14 +03:00
|
|
|
A[i] -=
|
|
|
|
A[w2 - 1] * 2 + B[i * w2 + w2 - 1] - 2 * B[(w2 - 1) * w2 + (w2 - 1)];
|
2016-06-01 08:50:24 +03:00
|
|
|
for (i = 0; i < w2 - 1; ++i)
|
|
|
|
for (j = 0; j < w2 - 1; ++j)
|
|
|
|
B[i * w2 + j] -= 2 * (B[i * w2 + (w2 - 1)] + B[(w2 - 1) * w2 + j] -
|
|
|
|
2 * B[(w2 - 1) * w2 + (w2 - 1)]);
|
|
|
|
if (linsolve(w2 - 1, B, w2, A, S)) {
|
|
|
|
S[w2 - 1] = 1.0;
|
|
|
|
for (i = w2; i < w; ++i) {
|
|
|
|
S[i] = S[w - 1 - i];
|
|
|
|
S[w2 - 1] -= 2 * S[i];
|
2016-02-09 23:24:33 +03:00
|
|
|
}
|
2016-06-01 08:50:24 +03:00
|
|
|
memcpy(b, S, w * sizeof(*b));
|
2016-02-09 23:24:33 +03:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2016-08-12 06:13:14 +03:00
|
|
|
static int wiener_decompose_sep_sym(double *M, double *H, double *a,
|
|
|
|
double *b) {
|
2016-12-16 00:23:21 +03:00
|
|
|
static const double init_filt[WIENER_WIN] = {
|
2016-08-12 06:13:14 +03:00
|
|
|
0.035623, -0.127154, 0.211436, 0.760190, 0.211436, -0.127154, 0.035623,
|
2016-02-09 23:24:33 +03:00
|
|
|
};
|
|
|
|
int i, j, iter;
|
2016-12-16 00:23:21 +03:00
|
|
|
double *Hc[WIENER_WIN2];
|
|
|
|
double *Mc[WIENER_WIN];
|
|
|
|
for (i = 0; i < WIENER_WIN; i++) {
|
|
|
|
Mc[i] = M + i * WIENER_WIN;
|
|
|
|
for (j = 0; j < WIENER_WIN; j++) {
|
|
|
|
Hc[i * WIENER_WIN + j] =
|
|
|
|
H + i * WIENER_WIN * WIENER_WIN2 + j * WIENER_WIN;
|
2016-02-09 23:24:33 +03:00
|
|
|
}
|
|
|
|
}
|
2016-12-16 00:23:21 +03:00
|
|
|
memcpy(a, init_filt, sizeof(*a) * WIENER_WIN);
|
|
|
|
memcpy(b, init_filt, sizeof(*b) * WIENER_WIN);
|
2016-02-09 23:24:33 +03:00
|
|
|
|
|
|
|
iter = 1;
|
|
|
|
while (iter < 10) {
|
|
|
|
update_a_sep_sym(Mc, Hc, a, b);
|
|
|
|
update_b_sep_sym(Mc, Hc, a, b);
|
|
|
|
iter++;
|
|
|
|
}
|
2016-03-11 09:56:24 +03:00
|
|
|
return 1;
|
2016-02-09 23:24:33 +03:00
|
|
|
}
|
|
|
|
|
2017-01-04 02:14:57 +03:00
|
|
|
// Computes the function x'*H*x - x'*M for the learned 2D filter x, and compares
|
2016-06-01 08:50:24 +03:00
|
|
|
// against identity filters; Final score is defined as the difference between
|
|
|
|
// the function values
|
2016-08-12 06:13:14 +03:00
|
|
|
static double compute_score(double *M, double *H, int *vfilt, int *hfilt) {
|
2016-12-16 00:23:21 +03:00
|
|
|
double ab[WIENER_WIN * WIENER_WIN];
|
2016-06-01 08:50:24 +03:00
|
|
|
int i, k, l;
|
|
|
|
double P = 0, Q = 0;
|
|
|
|
double iP = 0, iQ = 0;
|
|
|
|
double Score, iScore;
|
2016-12-16 00:23:21 +03:00
|
|
|
double a[WIENER_WIN], b[WIENER_WIN];
|
|
|
|
a[WIENER_HALFWIN] = b[WIENER_HALFWIN] = 1.0;
|
|
|
|
for (i = 0; i < WIENER_HALFWIN; ++i) {
|
|
|
|
a[i] = a[WIENER_WIN - i - 1] = (double)vfilt[i] / WIENER_FILT_STEP;
|
|
|
|
b[i] = b[WIENER_WIN - i - 1] = (double)hfilt[i] / WIENER_FILT_STEP;
|
|
|
|
a[WIENER_HALFWIN] -= 2 * a[i];
|
|
|
|
b[WIENER_HALFWIN] -= 2 * b[i];
|
2016-06-01 08:50:24 +03:00
|
|
|
}
|
2017-01-04 02:14:57 +03:00
|
|
|
for (k = 0; k < WIENER_WIN; ++k) {
|
|
|
|
for (l = 0; l < WIENER_WIN; ++l) ab[k * WIENER_WIN + l] = a[l] * b[k];
|
2016-06-01 08:50:24 +03:00
|
|
|
}
|
2017-01-04 02:14:57 +03:00
|
|
|
for (k = 0; k < WIENER_WIN2; ++k) {
|
2016-06-01 08:50:24 +03:00
|
|
|
P += ab[k] * M[k];
|
2017-01-04 02:14:57 +03:00
|
|
|
for (l = 0; l < WIENER_WIN2; ++l)
|
|
|
|
Q += ab[k] * H[k * WIENER_WIN2 + l] * ab[l];
|
2016-06-01 08:50:24 +03:00
|
|
|
}
|
|
|
|
Score = Q - 2 * P;
|
|
|
|
|
2017-01-04 02:14:57 +03:00
|
|
|
iP = M[WIENER_WIN2 >> 1];
|
|
|
|
iQ = H[(WIENER_WIN2 >> 1) * WIENER_WIN2 + (WIENER_WIN2 >> 1)];
|
2016-06-01 08:50:24 +03:00
|
|
|
iScore = iQ - 2 * iP;
|
|
|
|
|
|
|
|
return Score - iScore;
|
|
|
|
}
|
|
|
|
|
2016-02-09 23:24:33 +03:00
|
|
|
static void quantize_sym_filter(double *f, int *fi) {
|
|
|
|
int i;
|
2016-12-16 00:23:21 +03:00
|
|
|
for (i = 0; i < WIENER_HALFWIN; ++i) {
|
|
|
|
fi[i] = RINT(f[i] * WIENER_FILT_STEP);
|
2016-02-09 23:24:33 +03:00
|
|
|
}
|
|
|
|
// Specialize for 7-tap filter
|
|
|
|
fi[0] = CLIP(fi[0], WIENER_FILT_TAP0_MINV, WIENER_FILT_TAP0_MAXV);
|
|
|
|
fi[1] = CLIP(fi[1], WIENER_FILT_TAP1_MINV, WIENER_FILT_TAP1_MAXV);
|
|
|
|
fi[2] = CLIP(fi[2], WIENER_FILT_TAP2_MINV, WIENER_FILT_TAP2_MAXV);
|
2017-01-04 02:14:57 +03:00
|
|
|
// Satisfy filter constraints
|
|
|
|
fi[WIENER_WIN - 1] = fi[0];
|
|
|
|
fi[WIENER_WIN - 2] = fi[1];
|
|
|
|
fi[WIENER_WIN - 3] = fi[2];
|
|
|
|
fi[3] = WIENER_FILT_STEP - 2 * (fi[0] + fi[1] + fi[2]);
|
|
|
|
}
|
|
|
|
|
|
|
|
static double search_wiener_uv(const YV12_BUFFER_CONFIG *src, AV1_COMP *cpi,
|
|
|
|
int filter_level, int partial_frame, int plane,
|
2017-01-06 22:18:23 +03:00
|
|
|
RestorationInfo *info, RestorationType *type,
|
2017-01-04 02:14:57 +03:00
|
|
|
YV12_BUFFER_CONFIG *dst_frame) {
|
|
|
|
WienerInfo *wiener_info = info->wiener_info;
|
|
|
|
AV1_COMMON *const cm = &cpi->common;
|
|
|
|
RestorationInfo *rsi = cpi->rst_search;
|
|
|
|
int64_t err;
|
|
|
|
int bits;
|
2017-01-06 22:18:23 +03:00
|
|
|
double cost_wiener, cost_norestore, cost_wiener_frame, cost_norestore_frame;
|
2017-01-04 02:14:57 +03:00
|
|
|
MACROBLOCK *x = &cpi->td.mb;
|
|
|
|
double M[WIENER_WIN2];
|
|
|
|
double H[WIENER_WIN2 * WIENER_WIN2];
|
|
|
|
double vfilterd[WIENER_WIN], hfilterd[WIENER_WIN];
|
|
|
|
const YV12_BUFFER_CONFIG *dgd = cm->frame_to_show;
|
|
|
|
const int width = src->uv_crop_width;
|
|
|
|
const int height = src->uv_crop_height;
|
|
|
|
const int src_stride = src->uv_stride;
|
|
|
|
const int dgd_stride = dgd->uv_stride;
|
|
|
|
double score;
|
|
|
|
int tile_idx, tile_width, tile_height, nhtiles, nvtiles;
|
2017-01-06 22:18:23 +03:00
|
|
|
int h_start, h_end, v_start, v_end;
|
2017-01-06 00:58:16 +03:00
|
|
|
const int ntiles = av1_get_rest_ntiles(width, height, &tile_width,
|
2017-01-04 02:14:57 +03:00
|
|
|
&tile_height, &nhtiles, &nvtiles);
|
|
|
|
assert(width == dgd->uv_crop_width);
|
|
|
|
assert(height == dgd->uv_crop_height);
|
|
|
|
|
|
|
|
// Make a copy of the unfiltered / processed recon buffer
|
2017-01-07 00:02:57 +03:00
|
|
|
aom_yv12_copy_y(cm->frame_to_show, &cpi->last_frame_uf);
|
|
|
|
aom_yv12_copy_u(cm->frame_to_show, &cpi->last_frame_uf);
|
|
|
|
aom_yv12_copy_v(cm->frame_to_show, &cpi->last_frame_uf);
|
2017-01-04 02:14:57 +03:00
|
|
|
av1_loop_filter_frame(cm->frame_to_show, cm, &cpi->td.mb.e_mbd, filter_level,
|
|
|
|
0, partial_frame);
|
2017-01-07 00:02:57 +03:00
|
|
|
aom_yv12_copy_y(cm->frame_to_show, &cpi->last_frame_db);
|
|
|
|
aom_yv12_copy_u(cm->frame_to_show, &cpi->last_frame_db);
|
|
|
|
aom_yv12_copy_v(cm->frame_to_show, &cpi->last_frame_db);
|
2017-01-04 02:14:57 +03:00
|
|
|
|
|
|
|
rsi[plane].frame_restoration_type = RESTORE_NONE;
|
2017-01-07 00:02:57 +03:00
|
|
|
err = sse_restoration_frame(src, cm->frame_to_show, (1 << plane));
|
2017-01-04 02:14:57 +03:00
|
|
|
bits = 0;
|
2017-01-06 22:18:23 +03:00
|
|
|
cost_norestore_frame = RDCOST_DBL(x->rdmult, x->rddiv, (bits >> 4), err);
|
2017-01-04 02:14:57 +03:00
|
|
|
|
|
|
|
rsi[plane].frame_restoration_type = RESTORE_WIENER;
|
2017-01-06 18:34:50 +03:00
|
|
|
|
2017-01-06 22:18:23 +03:00
|
|
|
for (tile_idx = 0; tile_idx < ntiles; ++tile_idx) {
|
|
|
|
rsi[plane].restoration_type[tile_idx] = RESTORE_NONE;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Compute best Wiener filters for each tile
|
|
|
|
for (tile_idx = 0; tile_idx < ntiles; ++tile_idx) {
|
|
|
|
av1_get_rest_tile_limits(tile_idx, 0, 0, nhtiles, nvtiles, tile_width,
|
|
|
|
tile_height, width, height, 0, 0, &h_start, &h_end,
|
|
|
|
&v_start, &v_end);
|
|
|
|
err = sse_restoration_tile(src, cm->frame_to_show, cm, h_start,
|
|
|
|
h_end - h_start, v_start, v_end - v_start,
|
|
|
|
1 << plane);
|
|
|
|
// #bits when a tile is not restored
|
|
|
|
bits = av1_cost_bit(RESTORE_NONE_WIENER_PROB, 0);
|
|
|
|
cost_norestore = RDCOST_DBL(x->rdmult, x->rddiv, (bits >> 4), err);
|
|
|
|
// best_tile_cost[tile_idx] = DBL_MAX;
|
|
|
|
|
|
|
|
av1_get_rest_tile_limits(tile_idx, 0, 0, nhtiles, nvtiles, tile_width,
|
|
|
|
tile_height, width, height, WIENER_HALFWIN,
|
|
|
|
WIENER_HALFWIN, &h_start, &h_end, &v_start,
|
|
|
|
&v_end);
|
2017-01-06 18:34:50 +03:00
|
|
|
if (plane == AOM_PLANE_U) {
|
2017-01-06 22:18:23 +03:00
|
|
|
#if CONFIG_AOM_HIGHBITDEPTH
|
|
|
|
if (cm->use_highbitdepth)
|
|
|
|
compute_stats_highbd(dgd->u_buffer, src->u_buffer, h_start, h_end,
|
|
|
|
v_start, v_end, dgd_stride, src_stride, M, H);
|
|
|
|
else
|
|
|
|
#endif // CONFIG_AOM_HIGHBITDEPTH
|
|
|
|
compute_stats(dgd->u_buffer, src->u_buffer, h_start, h_end, v_start,
|
|
|
|
v_end, dgd_stride, src_stride, M, H);
|
2017-01-06 18:34:50 +03:00
|
|
|
} else if (plane == AOM_PLANE_V) {
|
2017-01-06 22:18:23 +03:00
|
|
|
#if CONFIG_AOM_HIGHBITDEPTH
|
|
|
|
if (cm->use_highbitdepth)
|
|
|
|
compute_stats_highbd(dgd->v_buffer, src->v_buffer, h_start, h_end,
|
|
|
|
v_start, v_end, dgd_stride, src_stride, M, H);
|
|
|
|
else
|
|
|
|
#endif // CONFIG_AOM_HIGHBITDEPTH
|
|
|
|
compute_stats(dgd->v_buffer, src->v_buffer, h_start, h_end, v_start,
|
|
|
|
v_end, dgd_stride, src_stride, M, H);
|
2017-01-06 18:34:50 +03:00
|
|
|
} else {
|
|
|
|
assert(0);
|
|
|
|
}
|
2017-01-06 22:18:23 +03:00
|
|
|
|
|
|
|
type[tile_idx] = RESTORE_WIENER;
|
|
|
|
|
|
|
|
if (!wiener_decompose_sep_sym(M, H, vfilterd, hfilterd)) {
|
|
|
|
type[tile_idx] = RESTORE_NONE;
|
|
|
|
continue;
|
2017-01-06 18:34:50 +03:00
|
|
|
}
|
2017-01-06 22:18:23 +03:00
|
|
|
quantize_sym_filter(vfilterd, rsi[plane].wiener_info[tile_idx].vfilter);
|
|
|
|
quantize_sym_filter(hfilterd, rsi[plane].wiener_info[tile_idx].hfilter);
|
2017-01-06 18:34:50 +03:00
|
|
|
|
2017-01-06 22:18:23 +03:00
|
|
|
// Filter score computes the value of the function x'*A*x - x'*b for the
|
|
|
|
// learned filter and compares it against identity filer. If there is no
|
|
|
|
// reduction in the function, the filter is reverted back to identity
|
|
|
|
score = compute_score(M, H, rsi[plane].wiener_info[tile_idx].vfilter,
|
|
|
|
rsi[plane].wiener_info[tile_idx].hfilter);
|
|
|
|
if (score > 0.0) {
|
|
|
|
type[tile_idx] = RESTORE_NONE;
|
|
|
|
continue;
|
|
|
|
}
|
2017-01-04 02:14:57 +03:00
|
|
|
|
2017-01-06 22:18:23 +03:00
|
|
|
rsi[plane].restoration_type[tile_idx] = RESTORE_WIENER;
|
|
|
|
err = try_restoration_tile(src, cpi, rsi, 1 << plane, partial_frame,
|
|
|
|
tile_idx, 0, 0, dst_frame);
|
|
|
|
bits = WIENER_FILT_BITS << AV1_PROB_COST_SHIFT;
|
|
|
|
bits += av1_cost_bit(RESTORE_NONE_WIENER_PROB, 1);
|
|
|
|
cost_wiener = RDCOST_DBL(x->rdmult, x->rddiv, (bits >> 4), err);
|
|
|
|
if (cost_wiener >= cost_norestore) {
|
|
|
|
type[tile_idx] = RESTORE_NONE;
|
|
|
|
} else {
|
|
|
|
type[tile_idx] = RESTORE_WIENER;
|
|
|
|
memcpy(&wiener_info[tile_idx], &rsi[plane].wiener_info[tile_idx],
|
|
|
|
sizeof(wiener_info[tile_idx]));
|
|
|
|
}
|
|
|
|
rsi[plane].restoration_type[tile_idx] = RESTORE_NONE;
|
|
|
|
}
|
|
|
|
aom_yv12_copy_frame(&cpi->last_frame_db, cm->frame_to_show);
|
|
|
|
// Cost for Wiener filtering
|
|
|
|
bits = 0;
|
|
|
|
for (tile_idx = 0; tile_idx < ntiles; ++tile_idx) {
|
|
|
|
bits +=
|
|
|
|
av1_cost_bit(RESTORE_NONE_WIENER_PROB, type[tile_idx] != RESTORE_NONE);
|
|
|
|
memcpy(&rsi[plane].wiener_info[tile_idx], &wiener_info[tile_idx],
|
|
|
|
sizeof(wiener_info[tile_idx]));
|
|
|
|
if (type[tile_idx] == RESTORE_WIENER) {
|
|
|
|
bits += (WIENER_FILT_BITS << AV1_PROB_COST_SHIFT);
|
|
|
|
}
|
|
|
|
rsi[plane].restoration_type[tile_idx] = type[tile_idx];
|
2017-01-04 02:14:57 +03:00
|
|
|
}
|
2017-01-06 22:18:23 +03:00
|
|
|
err = try_restoration_frame(src, cpi, rsi, 1 << plane, partial_frame,
|
2017-01-04 02:14:57 +03:00
|
|
|
dst_frame);
|
2017-01-06 22:18:23 +03:00
|
|
|
cost_wiener_frame = RDCOST_DBL(x->rdmult, x->rddiv, (bits >> 4), err);
|
|
|
|
|
|
|
|
if (cost_wiener_frame < cost_norestore_frame) {
|
|
|
|
info->frame_restoration_type = RESTORE_WIENER;
|
|
|
|
} else {
|
2017-01-04 02:14:57 +03:00
|
|
|
info->frame_restoration_type = RESTORE_NONE;
|
|
|
|
}
|
|
|
|
|
2017-01-07 00:02:57 +03:00
|
|
|
aom_yv12_copy_y(&cpi->last_frame_uf, cm->frame_to_show);
|
|
|
|
aom_yv12_copy_u(&cpi->last_frame_uf, cm->frame_to_show);
|
|
|
|
aom_yv12_copy_v(&cpi->last_frame_uf, cm->frame_to_show);
|
2017-01-06 22:18:23 +03:00
|
|
|
return info->frame_restoration_type == RESTORE_WIENER ? cost_wiener_frame
|
|
|
|
: cost_norestore_frame;
|
2016-02-09 23:24:33 +03:00
|
|
|
}
|
|
|
|
|
2016-09-17 23:16:58 +03:00
|
|
|
static double search_wiener(const YV12_BUFFER_CONFIG *src, AV1_COMP *cpi,
|
|
|
|
int filter_level, int partial_frame,
|
2017-01-06 22:18:23 +03:00
|
|
|
RestorationInfo *info, RestorationType *type,
|
|
|
|
double *best_tile_cost,
|
2016-12-08 14:25:47 +03:00
|
|
|
YV12_BUFFER_CONFIG *dst_frame) {
|
2016-09-17 23:16:58 +03:00
|
|
|
WienerInfo *wiener_info = info->wiener_info;
|
2016-08-31 00:01:10 +03:00
|
|
|
AV1_COMMON *const cm = &cpi->common;
|
2017-01-04 02:14:57 +03:00
|
|
|
RestorationInfo *rsi = cpi->rst_search;
|
2016-03-11 09:56:24 +03:00
|
|
|
int64_t err;
|
|
|
|
int bits;
|
2016-10-12 20:47:01 +03:00
|
|
|
double cost_wiener, cost_norestore;
|
2016-02-09 23:24:33 +03:00
|
|
|
MACROBLOCK *x = &cpi->td.mb;
|
2016-12-16 00:23:21 +03:00
|
|
|
double M[WIENER_WIN2];
|
|
|
|
double H[WIENER_WIN2 * WIENER_WIN2];
|
|
|
|
double vfilterd[WIENER_WIN], hfilterd[WIENER_WIN];
|
2016-02-09 23:24:33 +03:00
|
|
|
const YV12_BUFFER_CONFIG *dgd = cm->frame_to_show;
|
|
|
|
const int width = cm->width;
|
|
|
|
const int height = cm->height;
|
|
|
|
const int src_stride = src->y_stride;
|
|
|
|
const int dgd_stride = dgd->y_stride;
|
2016-06-01 08:50:24 +03:00
|
|
|
double score;
|
2016-09-09 01:15:17 +03:00
|
|
|
int tile_idx, tile_width, tile_height, nhtiles, nvtiles;
|
2016-07-21 07:29:45 +03:00
|
|
|
int h_start, h_end, v_start, v_end;
|
2016-09-09 01:15:17 +03:00
|
|
|
const int ntiles = av1_get_rest_ntiles(width, height, &tile_width,
|
|
|
|
&tile_height, &nhtiles, &nvtiles);
|
2016-02-09 23:24:33 +03:00
|
|
|
assert(width == dgd->y_crop_width);
|
|
|
|
assert(height == dgd->y_crop_height);
|
|
|
|
assert(width == src->y_crop_width);
|
|
|
|
assert(height == src->y_crop_height);
|
|
|
|
|
|
|
|
// Make a copy of the unfiltered / processed recon buffer
|
2016-08-31 00:01:10 +03:00
|
|
|
aom_yv12_copy_y(cm->frame_to_show, &cpi->last_frame_uf);
|
|
|
|
av1_loop_filter_frame(cm->frame_to_show, cm, &cpi->td.mb.e_mbd, filter_level,
|
|
|
|
1, partial_frame);
|
|
|
|
aom_yv12_copy_y(cm->frame_to_show, &cpi->last_frame_db);
|
2016-02-09 23:24:33 +03:00
|
|
|
|
2016-12-16 00:23:21 +03:00
|
|
|
rsi->frame_restoration_type = RESTORE_WIENER;
|
2016-09-17 23:16:58 +03:00
|
|
|
|
2017-01-06 22:18:23 +03:00
|
|
|
for (tile_idx = 0; tile_idx < ntiles; ++tile_idx) {
|
|
|
|
rsi->restoration_type[tile_idx] = RESTORE_NONE;
|
|
|
|
}
|
2016-07-21 07:29:45 +03:00
|
|
|
|
2017-01-06 18:34:50 +03:00
|
|
|
// Construct a (WIENER_HALFWIN)-pixel border around the frame
|
|
|
|
#if CONFIG_AOM_HIGHBITDEPTH
|
|
|
|
if (cm->use_highbitdepth)
|
|
|
|
extend_frame_highbd(CONVERT_TO_SHORTPTR(dgd->y_buffer), width, height,
|
|
|
|
dgd_stride);
|
|
|
|
else
|
|
|
|
#endif
|
|
|
|
extend_frame(dgd->y_buffer, width, height, dgd_stride);
|
|
|
|
|
2016-07-21 07:29:45 +03:00
|
|
|
// Compute best Wiener filters for each tile
|
|
|
|
for (tile_idx = 0; tile_idx < ntiles; ++tile_idx) {
|
2016-09-09 01:15:17 +03:00
|
|
|
av1_get_rest_tile_limits(tile_idx, 0, 0, nhtiles, nvtiles, tile_width,
|
|
|
|
tile_height, width, height, 0, 0, &h_start, &h_end,
|
|
|
|
&v_start, &v_end);
|
2016-12-08 14:25:47 +03:00
|
|
|
err = sse_restoration_tile(src, cm->frame_to_show, cm, h_start,
|
2016-12-15 03:53:17 +03:00
|
|
|
h_end - h_start, v_start, v_end - v_start, 1);
|
2016-09-09 01:15:17 +03:00
|
|
|
// #bits when a tile is not restored
|
|
|
|
bits = av1_cost_bit(RESTORE_NONE_WIENER_PROB, 0);
|
2016-10-12 20:47:01 +03:00
|
|
|
cost_norestore = RDCOST_DBL(x->rdmult, x->rddiv, (bits >> 4), err);
|
2016-09-17 23:16:58 +03:00
|
|
|
best_tile_cost[tile_idx] = DBL_MAX;
|
2016-09-09 01:15:17 +03:00
|
|
|
|
|
|
|
av1_get_rest_tile_limits(tile_idx, 0, 0, nhtiles, nvtiles, tile_width,
|
2017-01-06 18:34:50 +03:00
|
|
|
tile_height, width, height, 0, 0, &h_start, &h_end,
|
|
|
|
&v_start, &v_end);
|
2016-08-31 00:01:10 +03:00
|
|
|
#if CONFIG_AOM_HIGHBITDEPTH
|
2016-07-21 07:29:45 +03:00
|
|
|
if (cm->use_highbitdepth)
|
|
|
|
compute_stats_highbd(dgd->y_buffer, src->y_buffer, h_start, h_end,
|
|
|
|
v_start, v_end, dgd_stride, src_stride, M, H);
|
|
|
|
else
|
2016-08-31 00:01:10 +03:00
|
|
|
#endif // CONFIG_AOM_HIGHBITDEPTH
|
2016-07-21 07:29:45 +03:00
|
|
|
compute_stats(dgd->y_buffer, src->y_buffer, h_start, h_end, v_start,
|
|
|
|
v_end, dgd_stride, src_stride, M, H);
|
|
|
|
|
2017-01-06 22:18:23 +03:00
|
|
|
type[tile_idx] = RESTORE_WIENER;
|
|
|
|
|
2016-07-21 07:29:45 +03:00
|
|
|
if (!wiener_decompose_sep_sym(M, H, vfilterd, hfilterd)) {
|
2017-01-06 22:18:23 +03:00
|
|
|
type[tile_idx] = RESTORE_NONE;
|
2016-07-21 07:29:45 +03:00
|
|
|
continue;
|
|
|
|
}
|
2016-12-16 00:23:21 +03:00
|
|
|
quantize_sym_filter(vfilterd, rsi->wiener_info[tile_idx].vfilter);
|
|
|
|
quantize_sym_filter(hfilterd, rsi->wiener_info[tile_idx].hfilter);
|
2016-07-21 07:29:45 +03:00
|
|
|
|
|
|
|
// Filter score computes the value of the function x'*A*x - x'*b for the
|
|
|
|
// learned filter and compares it against identity filer. If there is no
|
|
|
|
// reduction in the function, the filter is reverted back to identity
|
2016-12-16 00:23:21 +03:00
|
|
|
score = compute_score(M, H, rsi->wiener_info[tile_idx].vfilter,
|
|
|
|
rsi->wiener_info[tile_idx].hfilter);
|
2016-07-21 07:29:45 +03:00
|
|
|
if (score > 0.0) {
|
2017-01-06 22:18:23 +03:00
|
|
|
type[tile_idx] = RESTORE_NONE;
|
2016-07-21 07:29:45 +03:00
|
|
|
continue;
|
|
|
|
}
|
2016-02-09 23:24:33 +03:00
|
|
|
|
2017-01-06 22:18:23 +03:00
|
|
|
rsi->restoration_type[tile_idx] = RESTORE_WIENER;
|
2016-12-16 00:23:21 +03:00
|
|
|
err = try_restoration_tile(src, cpi, rsi, 1, partial_frame, tile_idx, 0, 0,
|
2016-12-08 14:25:47 +03:00
|
|
|
dst_frame);
|
2016-09-09 01:15:17 +03:00
|
|
|
bits = WIENER_FILT_BITS << AV1_PROB_COST_SHIFT;
|
|
|
|
bits += av1_cost_bit(RESTORE_NONE_WIENER_PROB, 1);
|
|
|
|
cost_wiener = RDCOST_DBL(x->rdmult, x->rddiv, (bits >> 4), err);
|
2016-10-12 20:47:01 +03:00
|
|
|
if (cost_wiener >= cost_norestore) {
|
2017-01-06 22:18:23 +03:00
|
|
|
type[tile_idx] = RESTORE_NONE;
|
2016-09-17 23:16:58 +03:00
|
|
|
} else {
|
2017-01-06 22:18:23 +03:00
|
|
|
type[tile_idx] = RESTORE_WIENER;
|
|
|
|
memcpy(&wiener_info[tile_idx], &rsi->wiener_info[tile_idx],
|
|
|
|
sizeof(wiener_info[tile_idx]));
|
2016-09-09 01:15:17 +03:00
|
|
|
bits = WIENER_FILT_BITS << AV1_PROB_COST_SHIFT;
|
|
|
|
best_tile_cost[tile_idx] = RDCOST_DBL(
|
|
|
|
x->rdmult, x->rddiv,
|
|
|
|
(bits + cpi->switchable_restore_cost[RESTORE_WIENER]) >> 4, err);
|
|
|
|
}
|
2017-01-06 22:18:23 +03:00
|
|
|
rsi->restoration_type[tile_idx] = RESTORE_NONE;
|
2016-03-11 09:56:24 +03:00
|
|
|
}
|
2016-07-21 07:29:45 +03:00
|
|
|
// Cost for Wiener filtering
|
2016-12-16 00:23:21 +03:00
|
|
|
bits = frame_level_restore_bits[rsi->frame_restoration_type]
|
2016-09-09 01:15:17 +03:00
|
|
|
<< AV1_PROB_COST_SHIFT;
|
2016-07-21 07:29:45 +03:00
|
|
|
for (tile_idx = 0; tile_idx < ntiles; ++tile_idx) {
|
2017-01-06 22:18:23 +03:00
|
|
|
bits +=
|
|
|
|
av1_cost_bit(RESTORE_NONE_WIENER_PROB, type[tile_idx] != RESTORE_NONE);
|
|
|
|
memcpy(&rsi->wiener_info[tile_idx], &wiener_info[tile_idx],
|
|
|
|
sizeof(wiener_info[tile_idx]));
|
|
|
|
if (type[tile_idx] == RESTORE_WIENER) {
|
2016-09-09 01:15:17 +03:00
|
|
|
bits += (WIENER_FILT_BITS << AV1_PROB_COST_SHIFT);
|
2016-09-17 23:16:58 +03:00
|
|
|
}
|
2017-01-06 22:18:23 +03:00
|
|
|
rsi->restoration_type[tile_idx] = type[tile_idx];
|
2016-06-01 08:50:24 +03:00
|
|
|
}
|
2016-12-16 00:23:21 +03:00
|
|
|
err = try_restoration_frame(src, cpi, rsi, 1, partial_frame, dst_frame);
|
2016-09-09 01:15:17 +03:00
|
|
|
cost_wiener = RDCOST_DBL(x->rdmult, x->rddiv, (bits >> 4), err);
|
2016-10-12 20:47:01 +03:00
|
|
|
|
2016-08-31 00:01:10 +03:00
|
|
|
aom_yv12_copy_y(&cpi->last_frame_uf, cm->frame_to_show);
|
2016-09-17 23:16:58 +03:00
|
|
|
return cost_wiener;
|
2016-02-09 23:24:33 +03:00
|
|
|
}
|
|
|
|
|
2016-09-17 23:16:58 +03:00
|
|
|
static double search_norestore(const YV12_BUFFER_CONFIG *src, AV1_COMP *cpi,
|
|
|
|
int filter_level, int partial_frame,
|
2017-01-06 22:18:23 +03:00
|
|
|
RestorationInfo *info, RestorationType *type,
|
|
|
|
double *best_tile_cost,
|
2016-12-08 14:25:47 +03:00
|
|
|
YV12_BUFFER_CONFIG *dst_frame) {
|
2016-09-17 23:16:58 +03:00
|
|
|
double err, cost_norestore;
|
|
|
|
int bits;
|
2016-09-09 01:15:17 +03:00
|
|
|
MACROBLOCK *x = &cpi->td.mb;
|
2016-09-17 23:16:58 +03:00
|
|
|
AV1_COMMON *const cm = &cpi->common;
|
|
|
|
int tile_idx, tile_width, tile_height, nhtiles, nvtiles;
|
2016-09-09 01:15:17 +03:00
|
|
|
int h_start, h_end, v_start, v_end;
|
|
|
|
const int ntiles = av1_get_rest_ntiles(cm->width, cm->height, &tile_width,
|
|
|
|
&tile_height, &nhtiles, &nvtiles);
|
2016-09-17 23:16:58 +03:00
|
|
|
(void)info;
|
2016-12-08 14:25:47 +03:00
|
|
|
(void)dst_frame;
|
2016-09-09 01:15:17 +03:00
|
|
|
|
|
|
|
// Make a copy of the unfiltered / processed recon buffer
|
|
|
|
aom_yv12_copy_y(cm->frame_to_show, &cpi->last_frame_uf);
|
|
|
|
av1_loop_filter_frame(cm->frame_to_show, cm, &cpi->td.mb.e_mbd, filter_level,
|
|
|
|
1, partial_frame);
|
|
|
|
aom_yv12_copy_y(cm->frame_to_show, &cpi->last_frame_db);
|
|
|
|
|
|
|
|
for (tile_idx = 0; tile_idx < ntiles; ++tile_idx) {
|
|
|
|
av1_get_rest_tile_limits(tile_idx, 0, 0, nhtiles, nvtiles, tile_width,
|
|
|
|
tile_height, cm->width, cm->height, 0, 0, &h_start,
|
|
|
|
&h_end, &v_start, &v_end);
|
2016-12-08 14:25:47 +03:00
|
|
|
err = sse_restoration_tile(src, cm->frame_to_show, cm, h_start,
|
2016-12-15 03:53:17 +03:00
|
|
|
h_end - h_start, v_start, v_end - v_start, 1);
|
2016-09-17 23:16:58 +03:00
|
|
|
best_tile_cost[tile_idx] =
|
2016-09-09 01:15:17 +03:00
|
|
|
RDCOST_DBL(x->rdmult, x->rddiv,
|
|
|
|
(cpi->switchable_restore_cost[RESTORE_NONE] >> 4), err);
|
2017-01-06 22:18:23 +03:00
|
|
|
type[tile_idx] = RESTORE_NONE;
|
2016-09-17 23:16:58 +03:00
|
|
|
}
|
|
|
|
// RD cost associated with no restoration
|
2016-12-08 14:25:47 +03:00
|
|
|
err = sse_restoration_tile(src, cm->frame_to_show, cm, 0, cm->width, 0,
|
2016-12-15 03:53:17 +03:00
|
|
|
cm->height, 1);
|
2016-09-17 23:16:58 +03:00
|
|
|
bits = frame_level_restore_bits[RESTORE_NONE] << AV1_PROB_COST_SHIFT;
|
|
|
|
cost_norestore = RDCOST_DBL(x->rdmult, x->rddiv, (bits >> 4), err);
|
|
|
|
aom_yv12_copy_y(&cpi->last_frame_uf, cm->frame_to_show);
|
|
|
|
return cost_norestore;
|
|
|
|
}
|
|
|
|
|
|
|
|
static double search_switchable_restoration(
|
|
|
|
AV1_COMP *cpi, int filter_level, int partial_frame, RestorationInfo *rsi,
|
|
|
|
double *tile_cost[RESTORE_SWITCHABLE_TYPES]) {
|
|
|
|
AV1_COMMON *const cm = &cpi->common;
|
|
|
|
MACROBLOCK *x = &cpi->td.mb;
|
|
|
|
double cost_switchable = 0;
|
|
|
|
int r, bits, tile_idx;
|
|
|
|
const int ntiles =
|
|
|
|
av1_get_rest_ntiles(cm->width, cm->height, NULL, NULL, NULL, NULL);
|
|
|
|
|
|
|
|
// Make a copy of the unfiltered / processed recon buffer
|
|
|
|
aom_yv12_copy_y(cm->frame_to_show, &cpi->last_frame_uf);
|
|
|
|
av1_loop_filter_frame(cm->frame_to_show, cm, &cpi->td.mb.e_mbd, filter_level,
|
|
|
|
1, partial_frame);
|
|
|
|
aom_yv12_copy_y(cm->frame_to_show, &cpi->last_frame_db);
|
|
|
|
|
|
|
|
rsi->frame_restoration_type = RESTORE_SWITCHABLE;
|
|
|
|
bits = frame_level_restore_bits[rsi->frame_restoration_type]
|
|
|
|
<< AV1_PROB_COST_SHIFT;
|
|
|
|
cost_switchable = RDCOST_DBL(x->rdmult, x->rddiv, bits >> 4, 0);
|
|
|
|
for (tile_idx = 0; tile_idx < ntiles; ++tile_idx) {
|
|
|
|
double best_cost = tile_cost[RESTORE_NONE][tile_idx];
|
|
|
|
rsi->restoration_type[tile_idx] = RESTORE_NONE;
|
|
|
|
for (r = 1; r < RESTORE_SWITCHABLE_TYPES; r++) {
|
|
|
|
if (tile_cost[r][tile_idx] < best_cost) {
|
|
|
|
rsi->restoration_type[tile_idx] = r;
|
|
|
|
best_cost = tile_cost[r][tile_idx];
|
2016-09-09 01:15:17 +03:00
|
|
|
}
|
|
|
|
}
|
2016-09-17 23:16:58 +03:00
|
|
|
cost_switchable += best_cost;
|
2016-09-09 01:15:17 +03:00
|
|
|
}
|
|
|
|
aom_yv12_copy_y(&cpi->last_frame_uf, cm->frame_to_show);
|
2016-09-17 23:16:58 +03:00
|
|
|
return cost_switchable;
|
2016-09-09 01:15:17 +03:00
|
|
|
}
|
|
|
|
|
|
|
|
void av1_pick_filter_restoration(const YV12_BUFFER_CONFIG *src, AV1_COMP *cpi,
|
2016-08-31 00:01:10 +03:00
|
|
|
LPF_PICK_METHOD method) {
|
2016-09-17 23:16:58 +03:00
|
|
|
static search_restore_type search_restore_fun[RESTORE_SWITCHABLE_TYPES] = {
|
2016-12-08 20:22:44 +03:00
|
|
|
search_norestore, search_wiener, search_sgrproj, search_domaintxfmrf,
|
2016-09-17 23:16:58 +03:00
|
|
|
};
|
2016-08-31 00:01:10 +03:00
|
|
|
AV1_COMMON *const cm = &cpi->common;
|
2016-02-09 23:24:33 +03:00
|
|
|
struct loopfilter *const lf = &cm->lf;
|
2016-09-17 23:16:58 +03:00
|
|
|
double cost_restore[RESTORE_TYPES];
|
|
|
|
double *tile_cost[RESTORE_SWITCHABLE_TYPES];
|
2017-01-06 22:18:23 +03:00
|
|
|
RestorationType *restore_types[RESTORE_SWITCHABLE_TYPES];
|
2016-09-17 23:16:58 +03:00
|
|
|
double best_cost_restore;
|
|
|
|
RestorationType r, best_restore;
|
|
|
|
|
2016-09-09 01:15:17 +03:00
|
|
|
const int ntiles =
|
|
|
|
av1_get_rest_ntiles(cm->width, cm->height, NULL, NULL, NULL, NULL);
|
2016-09-17 23:16:58 +03:00
|
|
|
|
2017-01-06 22:18:23 +03:00
|
|
|
for (r = 0; r < RESTORE_SWITCHABLE_TYPES; r++) {
|
2016-09-17 23:16:58 +03:00
|
|
|
tile_cost[r] = (double *)aom_malloc(sizeof(*tile_cost[0]) * ntiles);
|
2017-01-06 22:18:23 +03:00
|
|
|
restore_types[r] =
|
|
|
|
(RestorationType *)aom_malloc(sizeof(*restore_types[0]) * ntiles);
|
|
|
|
}
|
2016-02-09 23:24:33 +03:00
|
|
|
|
2016-08-12 06:13:14 +03:00
|
|
|
lf->sharpness_level = cm->frame_type == KEY_FRAME ? 0 : cpi->oxcf.sharpness;
|
2016-02-09 23:24:33 +03:00
|
|
|
|
|
|
|
if (method == LPF_PICK_MINIMAL_LPF && lf->filter_level) {
|
2016-08-12 06:13:14 +03:00
|
|
|
lf->filter_level = 0;
|
2017-01-04 02:14:57 +03:00
|
|
|
cm->rst_info[0].frame_restoration_type = RESTORE_NONE;
|
2016-02-09 23:24:33 +03:00
|
|
|
} else if (method >= LPF_PICK_FROM_Q) {
|
|
|
|
const int min_filter_level = 0;
|
2016-08-31 00:01:10 +03:00
|
|
|
const int max_filter_level = av1_get_max_filter_level(cpi);
|
|
|
|
const int q = av1_ac_quant(cm->base_qindex, 0, cm->bit_depth);
|
2016-08-12 06:13:14 +03:00
|
|
|
// These values were determined by linear fitting the result of the
|
|
|
|
// searched level, filt_guess = q * 0.316206 + 3.87252
|
2016-08-31 00:01:10 +03:00
|
|
|
#if CONFIG_AOM_HIGHBITDEPTH
|
2016-02-09 23:24:33 +03:00
|
|
|
int filt_guess;
|
|
|
|
switch (cm->bit_depth) {
|
2016-08-31 00:01:10 +03:00
|
|
|
case AOM_BITS_8:
|
2016-02-09 23:24:33 +03:00
|
|
|
filt_guess = ROUND_POWER_OF_TWO(q * 20723 + 1015158, 18);
|
|
|
|
break;
|
2016-08-31 00:01:10 +03:00
|
|
|
case AOM_BITS_10:
|
2016-02-09 23:24:33 +03:00
|
|
|
filt_guess = ROUND_POWER_OF_TWO(q * 20723 + 4060632, 20);
|
|
|
|
break;
|
2016-08-31 00:01:10 +03:00
|
|
|
case AOM_BITS_12:
|
2016-02-09 23:24:33 +03:00
|
|
|
filt_guess = ROUND_POWER_OF_TWO(q * 20723 + 16242526, 22);
|
|
|
|
break;
|
|
|
|
default:
|
2016-08-12 06:13:14 +03:00
|
|
|
assert(0 &&
|
2016-08-31 00:01:10 +03:00
|
|
|
"bit_depth should be AOM_BITS_8, AOM_BITS_10 "
|
|
|
|
"or AOM_BITS_12");
|
2016-02-09 23:24:33 +03:00
|
|
|
return;
|
|
|
|
}
|
|
|
|
#else
|
|
|
|
int filt_guess = ROUND_POWER_OF_TWO(q * 20723 + 1015158, 18);
|
2016-08-31 00:01:10 +03:00
|
|
|
#endif // CONFIG_AOM_HIGHBITDEPTH
|
2016-08-12 06:13:14 +03:00
|
|
|
if (cm->frame_type == KEY_FRAME) filt_guess -= 4;
|
2016-02-09 23:24:33 +03:00
|
|
|
lf->filter_level = clamp(filt_guess, min_filter_level, max_filter_level);
|
|
|
|
} else {
|
2016-09-17 23:16:58 +03:00
|
|
|
lf->filter_level =
|
|
|
|
av1_search_filter_level(src, cpi, method == LPF_PICK_FROM_SUBIMAGE,
|
|
|
|
&cost_restore[RESTORE_NONE]);
|
|
|
|
}
|
|
|
|
for (r = 0; r < RESTORE_SWITCHABLE_TYPES; ++r) {
|
2016-12-08 14:25:47 +03:00
|
|
|
cost_restore[r] = search_restore_fun[r](
|
|
|
|
src, cpi, lf->filter_level, method == LPF_PICK_FROM_SUBIMAGE,
|
2017-01-06 22:18:23 +03:00
|
|
|
&cm->rst_info[0], restore_types[r], tile_cost[r],
|
|
|
|
&cpi->trial_frame_rst);
|
2016-09-17 23:16:58 +03:00
|
|
|
}
|
|
|
|
cost_restore[RESTORE_SWITCHABLE] = search_switchable_restoration(
|
2017-01-04 02:14:57 +03:00
|
|
|
cpi, lf->filter_level, method == LPF_PICK_FROM_SUBIMAGE, &cm->rst_info[0],
|
2016-09-17 23:16:58 +03:00
|
|
|
tile_cost);
|
|
|
|
|
|
|
|
best_cost_restore = DBL_MAX;
|
|
|
|
best_restore = 0;
|
|
|
|
for (r = 0; r < RESTORE_TYPES; ++r) {
|
|
|
|
if (cost_restore[r] < best_cost_restore) {
|
|
|
|
best_restore = r;
|
|
|
|
best_cost_restore = cost_restore[r];
|
|
|
|
}
|
2016-02-09 23:24:33 +03:00
|
|
|
}
|
2017-01-04 02:14:57 +03:00
|
|
|
cm->rst_info[0].frame_restoration_type = best_restore;
|
2017-01-06 22:18:23 +03:00
|
|
|
if (best_restore != RESTORE_SWITCHABLE) {
|
|
|
|
memcpy(cm->rst_info[0].restoration_type, restore_types[best_restore],
|
|
|
|
ntiles * sizeof(restore_types[best_restore][0]));
|
|
|
|
}
|
2017-01-04 02:14:57 +03:00
|
|
|
|
|
|
|
// Color components
|
|
|
|
search_wiener_uv(src, cpi, lf->filter_level, method == LPF_PICK_FROM_SUBIMAGE,
|
|
|
|
AOM_PLANE_U, &cm->rst_info[AOM_PLANE_U],
|
2017-01-06 22:18:23 +03:00
|
|
|
cm->rst_info[AOM_PLANE_U].restoration_type,
|
2017-01-04 02:14:57 +03:00
|
|
|
&cpi->trial_frame_rst);
|
|
|
|
search_wiener_uv(src, cpi, lf->filter_level, method == LPF_PICK_FROM_SUBIMAGE,
|
|
|
|
AOM_PLANE_V, &cm->rst_info[AOM_PLANE_V],
|
2017-01-06 22:18:23 +03:00
|
|
|
cm->rst_info[AOM_PLANE_V].restoration_type,
|
2017-01-04 02:14:57 +03:00
|
|
|
&cpi->trial_frame_rst);
|
2016-09-17 23:16:58 +03:00
|
|
|
/*
|
2017-01-07 00:02:57 +03:00
|
|
|
printf("Frame %d/%d restore types: %d %d %d\n",
|
|
|
|
cm->current_video_frame, cm->show_frame,
|
2017-01-04 02:14:57 +03:00
|
|
|
cm->rst_info[0].frame_restoration_type,
|
|
|
|
cm->rst_info[1].frame_restoration_type,
|
|
|
|
cm->rst_info[2].frame_restoration_type);
|
2017-01-06 00:58:16 +03:00
|
|
|
*/
|
|
|
|
/*
|
2016-12-08 20:22:44 +03:00
|
|
|
printf("Frame %d/%d frame_restore_type %d : %f %f %f %f %f\n",
|
2016-09-17 23:16:58 +03:00
|
|
|
cm->current_video_frame, cm->show_frame,
|
2017-01-04 02:14:57 +03:00
|
|
|
cm->rst_info[0].frame_restoration_type, cost_restore[0],
|
|
|
|
cost_restore[1],
|
2016-12-08 20:22:44 +03:00
|
|
|
cost_restore[2], cost_restore[3], cost_restore[4]);
|
2016-09-17 23:16:58 +03:00
|
|
|
*/
|
2017-01-04 02:14:57 +03:00
|
|
|
|
2017-01-06 22:18:23 +03:00
|
|
|
for (r = 0; r < RESTORE_SWITCHABLE_TYPES; r++) {
|
|
|
|
aom_free(tile_cost[r]);
|
|
|
|
aom_free(restore_types[r]);
|
|
|
|
}
|
2016-02-09 23:24:33 +03:00
|
|
|
}
|