pickrst.c 48.3 KB
Newer Older
1
/*
2
 * Copyright (c) 2016, Alliance for Open Media. All rights reserved
3
 *
4 5 6 7 8 9
 * 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.
10 11 12
 */

#include <assert.h>
13
#include <float.h>
14 15 16
#include <limits.h>
#include <math.h>

Yaowu Xu's avatar
Yaowu Xu committed
17
#include "./aom_scale_rtcd.h"
18

Yaowu Xu's avatar
Yaowu Xu committed
19
#include "aom_dsp/aom_dsp_common.h"
20 21
#include "aom_dsp/binary_codes_writer.h"
#include "aom_dsp/psnr.h"
Yaowu Xu's avatar
Yaowu Xu committed
22
#include "aom_mem/aom_mem.h"
23
#include "aom_ports/mem.h"
24
#include "aom_ports/system_state.h"
25

26 27
#include "av1/common/onyxc_int.h"
#include "av1/common/quant_common.h"
28
#include "av1/common/restoration.h"
29

30
#include "av1/encoder/av1_quantize.h"
31 32 33
#include "av1/encoder/encoder.h"
#include "av1/encoder/picklpf.h"
#include "av1/encoder/pickrst.h"
34

35 36 37
// When set to RESTORE_WIENER or RESTORE_SGRPROJ only those are allowed.
// When set to RESTORE_NONE (0) we allow switchable.
const RestorationType force_restore_type = RESTORE_NONE;
38 39 40 41

// Number of Wiener iterations
#define NUM_WIENER_ITERS 10

42
typedef double (*search_restore_type)(const YV12_BUFFER_CONFIG *src,
43 44
                                      AV1_COMP *cpi, int partial_frame,
                                      RestorationInfo *info,
45
                                      RestorationType *rest_level,
46 47
                                      double *best_tile_cost,
                                      YV12_BUFFER_CONFIG *dst_frame);
48

49
const int frame_level_restore_bits[RESTORE_TYPES] = { 2, 2, 2, 2 };
50 51

static int64_t sse_restoration_tile(const YV12_BUFFER_CONFIG *src,
52 53
                                    const YV12_BUFFER_CONFIG *dst,
                                    const AV1_COMMON *cm, int h_start,
54
                                    int width, int v_start, int height,
55 56
                                    int components_pattern) {
  int64_t filt_err = 0;
57 58 59 60
  (void)cm;
  // Y and UV components cannot be mixed
  assert(components_pattern == 1 || components_pattern == 2 ||
         components_pattern == 4 || components_pattern == 6);
61
#if CONFIG_HIGHBITDEPTH
62
  if (cm->use_highbitdepth) {
63 64 65 66 67
    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) {
68 69
      filt_err +=
          aom_highbd_get_u_sse_part(src, dst, h_start, width, v_start, height);
70 71
    }
    if ((components_pattern >> AOM_PLANE_V) & 1) {
72 73
      filt_err +=
          aom_highbd_get_v_sse_part(src, dst, h_start, width, v_start, height);
74 75
    }
    return filt_err;
76
  }
77
#endif  // CONFIG_HIGHBITDEPTH
78 79 80 81
  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) {
82
    filt_err += aom_get_u_sse_part(src, dst, h_start, width, v_start, height);
83 84
  }
  if ((components_pattern >> AOM_PLANE_V) & 1) {
85
    filt_err += aom_get_v_sse_part(src, dst, h_start, width, v_start, height);
86
  }
87 88 89
  return filt_err;
}

90 91
static int64_t sse_restoration_frame(AV1_COMMON *const cm,
                                     const YV12_BUFFER_CONFIG *src,
92 93 94
                                     const YV12_BUFFER_CONFIG *dst,
                                     int components_pattern) {
  int64_t filt_err = 0;
95
#if CONFIG_HIGHBITDEPTH
96 97 98 99 100 101 102 103 104 105 106 107
  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;
  }
108 109
#else
  (void)cm;
110
#endif  // CONFIG_HIGHBITDEPTH
111 112 113 114 115 116 117 118 119 120 121 122
  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;
}

123 124
static int64_t try_restoration_tile(const YV12_BUFFER_CONFIG *src,
                                    AV1_COMP *const cpi, RestorationInfo *rsi,
125 126 127
                                    int components_pattern, int partial_frame,
                                    int tile_idx, int subtile_idx,
                                    int subtile_bits,
128
                                    YV12_BUFFER_CONFIG *dst_frame) {
129 130 131 132
  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;
133 134 135 136 137 138 139 140 141 142 143 144 145
  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;
  }
146 147 148
  ntiles = av1_get_rest_ntiles(
      width, height, cm->rst_info[components_pattern > 1].restoration_tilesize,
      &tile_width, &tile_height, &nhtiles, &nvtiles);
149 150
  (void)ntiles;

151 152
  av1_loop_restoration_frame(cm->frame_to_show, cm, rsi, components_pattern,
                             partial_frame, dst_frame);
153
  av1_get_rest_tile_limits(tile_idx, subtile_idx, subtile_bits, nhtiles,
154 155
                           nvtiles, tile_width, tile_height, width, height, 0,
                           0, &h_start, &h_end, &v_start, &v_end);
156
  filt_err = sse_restoration_tile(src, dst_frame, cm, h_start, h_end - h_start,
157
                                  v_start, v_end - v_start, components_pattern);
158 159 160 161 162

  return filt_err;
}

static int64_t try_restoration_frame(const YV12_BUFFER_CONFIG *src,
Yaowu Xu's avatar
Yaowu Xu committed
163
                                     AV1_COMP *const cpi, RestorationInfo *rsi,
164
                                     int components_pattern, int partial_frame,
165
                                     YV12_BUFFER_CONFIG *dst_frame) {
Yaowu Xu's avatar
Yaowu Xu committed
166
  AV1_COMMON *const cm = &cpi->common;
167
  int64_t filt_err;
168 169
  av1_loop_restoration_frame(cm->frame_to_show, cm, rsi, components_pattern,
                             partial_frame, dst_frame);
170
  filt_err = sse_restoration_frame(cm, src, dst_frame, components_pattern);
171 172 173
  return filt_err;
}

174 175 176 177 178
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) {
179 180 181 182
  int i, j;
  int64_t err = 0;
  int xq[2];
  decode_xq(xqd, xq);
183 184 185 186 187 188 189 190 191
  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;
David Barker's avatar
David Barker committed
192
        const int32_t v = xq[0] * f1 + xq[1] * f2 + (u << SGRPROJ_PRJ_BITS);
193 194 195 196 197 198 199 200 201 202 203 204 205 206 207
        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;
David Barker's avatar
David Barker committed
208
        const int32_t v = xq[0] * f1 + xq[1] * f2 + (u << SGRPROJ_PRJ_BITS);
209 210 211 212 213
        const int32_t e =
            ROUND_POWER_OF_TWO(v, SGRPROJ_RST_BITS + SGRPROJ_PRJ_BITS) -
            src[i * src_stride + j];
        err += e * e;
      }
214 215 216 217 218
    }
  }
  return err;
}

219 220 221 222
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) {
223 224 225 226 227 228 229
  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;

230 231
  aom_clear_system_state();

232 233 234
  // Default
  xq[0] = 0;
  xq[1] = 0;
235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267
  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;
      }
268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284
    }
  }
  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) {
285
  xqd[0] = xq[0];
286
  xqd[0] = clamp(xqd[0], SGRPROJ_PRJ_MIN0, SGRPROJ_PRJ_MAX0);
287
  xqd[1] = (1 << SGRPROJ_PRJ_BITS) - xqd[0] - xq[1];
288 289 290 291 292 293
  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,
294 295
                                          int *eps, int *xqd, int32_t *rstbuf) {
  int32_t *flt1 = rstbuf;
296
  int32_t *flt2 = flt1 + RESTORATION_TILEPELS_MAX;
297
  int32_t *tmpbuf2 = flt2 + RESTORATION_TILEPELS_MAX;
298
  int ep, bestep = 0;
299 300
  int64_t err, besterr = -1;
  int exqd[2], bestxqd[2] = { 0, 0 };
301

302 303
  for (ep = 0; ep < SGRPROJ_PARAMS; ep++) {
    int exq[2];
304
#if CONFIG_HIGHBITDEPTH
305 306
    if (bit_depth > 8) {
      uint16_t *dat = CONVERT_TO_SHORTPTR(dat8);
307 308
#if USE_HIGHPASS_IN_SGRPROJ
      av1_highpass_filter_highbd(dat, width, height, dat_stride, flt1, width,
309
                                 sgr_params[ep].corner, sgr_params[ep].edge);
310
#else
311 312 313
      av1_selfguided_restoration_highbd(dat, width, height, dat_stride, flt1,
                                        width, bit_depth, sgr_params[ep].r1,
                                        sgr_params[ep].e1, tmpbuf2);
314
#endif  // USE_HIGHPASS_IN_SGRPROJ
315 316 317
      av1_selfguided_restoration_highbd(dat, width, height, dat_stride, flt2,
                                        width, bit_depth, sgr_params[ep].r2,
                                        sgr_params[ep].e2, tmpbuf2);
318
    } else {
319
#endif
320 321
#if USE_HIGHPASS_IN_SGRPROJ
      av1_highpass_filter(dat8, width, height, dat_stride, flt1, width,
322
                          sgr_params[ep].corner, sgr_params[ep].edge);
323 324 325 326
#else
    av1_selfguided_restoration(dat8, width, height, dat_stride, flt1, width,
                               sgr_params[ep].r1, sgr_params[ep].e1, tmpbuf2);
#endif  // USE_HIGHPASS_IN_SGRPROJ
327
      av1_selfguided_restoration(dat8, width, height, dat_stride, flt2, width,
328
                                 sgr_params[ep].r2, sgr_params[ep].e2, tmpbuf2);
329
#if CONFIG_HIGHBITDEPTH
330
    }
331
#endif
332 333
    get_proj_subspace(src8, width, height, src_stride, dat8, dat_stride,
                      bit_depth, flt1, width, flt2, width, exq);
334
    encode_xq(exq, exqd);
335 336 337
    err =
        get_pixel_proj_error(src8, width, height, src_stride, dat8, dat_stride,
                             bit_depth, flt1, width, flt2, width, exqd);
338 339 340 341 342 343 344 345 346 347 348 349
    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];
}

350 351 352 353 354 355 356 357 358 359 360 361 362 363
static int count_sgrproj_bits(SgrprojInfo *sgrproj_info,
                              SgrprojInfo *ref_sgrproj_info) {
  int bits = SGRPROJ_PARAMS_BITS;
  bits += aom_count_primitive_refsubexpfin(
      SGRPROJ_PRJ_MAX0 - SGRPROJ_PRJ_MIN0 + 1, SGRPROJ_PRJ_SUBEXP_K,
      ref_sgrproj_info->xqd[0] - SGRPROJ_PRJ_MIN0,
      sgrproj_info->xqd[0] - SGRPROJ_PRJ_MIN0);
  bits += aom_count_primitive_refsubexpfin(
      SGRPROJ_PRJ_MAX1 - SGRPROJ_PRJ_MIN1 + 1, SGRPROJ_PRJ_SUBEXP_K,
      ref_sgrproj_info->xqd[1] - SGRPROJ_PRJ_MIN1,
      sgrproj_info->xqd[1] - SGRPROJ_PRJ_MIN1);
  return bits;
}

364
static double search_sgrproj(const YV12_BUFFER_CONFIG *src, AV1_COMP *cpi,
365 366
                             int partial_frame, RestorationInfo *info,
                             RestorationType *type, double *best_tile_cost,
367
                             YV12_BUFFER_CONFIG *dst_frame) {
368 369 370 371 372 373
  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;
374
  RestorationInfo *rsi = &cpi->rst_search[0];
375 376
  int tile_idx, tile_width, tile_height, nhtiles, nvtiles;
  int h_start, h_end, v_start, v_end;
377
  // Allocate for the src buffer at high precision
378 379 380
  const int ntiles = av1_get_rest_ntiles(
      cm->width, cm->height, cm->rst_info[0].restoration_tilesize, &tile_width,
      &tile_height, &nhtiles, &nvtiles);
381 382 383
  SgrprojInfo ref_sgrproj_info;
  set_default_sgrproj(&ref_sgrproj_info);

384
  rsi->frame_restoration_type = RESTORE_SGRPROJ;
385

386 387 388
  for (tile_idx = 0; tile_idx < ntiles; ++tile_idx) {
    rsi->restoration_type[tile_idx] = RESTORE_NONE;
  }
389 390 391 392 393
  // 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);
394
    err = sse_restoration_tile(src, cm->frame_to_show, cm, h_start,
395
                               h_end - h_start, v_start, v_end - v_start, 1);
396 397 398 399 400 401 402 403
    // #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,
404
#if CONFIG_HIGHBITDEPTH
405 406 407
        cm->bit_depth,
#else
        8,
408
#endif  // CONFIG_HIGHBITDEPTH
409
        &rsi->sgrproj_info[tile_idx].ep, rsi->sgrproj_info[tile_idx].xqd,
410
        cm->rst_internal.tmpbuf);
411
    rsi->restoration_type[tile_idx] = RESTORE_SGRPROJ;
412
    err = try_restoration_tile(src, cpi, rsi, 1, partial_frame, tile_idx, 0, 0,
413
                               dst_frame);
414 415
    bits = count_sgrproj_bits(&rsi->sgrproj_info[tile_idx], &ref_sgrproj_info)
           << AV1_PROB_COST_SHIFT;
416 417 418
    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) {
419
      type[tile_idx] = RESTORE_NONE;
420
    } else {
421
      type[tile_idx] = RESTORE_SGRPROJ;
422
      memcpy(&sgrproj_info[tile_idx], &rsi->sgrproj_info[tile_idx],
423
             sizeof(sgrproj_info[tile_idx]));
424 425 426 427 428
      bits = count_sgrproj_bits(&rsi->sgrproj_info[tile_idx], &ref_sgrproj_info)
             << AV1_PROB_COST_SHIFT;
      memcpy(&ref_sgrproj_info, &sgrproj_info[tile_idx],
             sizeof(ref_sgrproj_info));
      best_tile_cost[tile_idx] = err;
429
    }
430
    rsi->restoration_type[tile_idx] = RESTORE_NONE;
431 432
  }
  // Cost for Sgrproj filtering
433
  set_default_sgrproj(&ref_sgrproj_info);
434
  bits = frame_level_restore_bits[rsi->frame_restoration_type]
435 436 437
         << AV1_PROB_COST_SHIFT;
  for (tile_idx = 0; tile_idx < ntiles; ++tile_idx) {
    bits +=
438
        av1_cost_bit(RESTORE_NONE_SGRPROJ_PROB, type[tile_idx] != RESTORE_NONE);
439
    memcpy(&rsi->sgrproj_info[tile_idx], &sgrproj_info[tile_idx],
440
           sizeof(sgrproj_info[tile_idx]));
441
    if (type[tile_idx] == RESTORE_SGRPROJ) {
442 443 444 445 446
      bits +=
          count_sgrproj_bits(&rsi->sgrproj_info[tile_idx], &ref_sgrproj_info)
          << AV1_PROB_COST_SHIFT;
      memcpy(&ref_sgrproj_info, &rsi->sgrproj_info[tile_idx],
             sizeof(ref_sgrproj_info));
447
    }
448
    rsi->restoration_type[tile_idx] = type[tile_idx];
449
  }
450
  err = try_restoration_frame(src, cpi, rsi, 1, partial_frame, dst_frame);
451 452 453 454 455
  cost_sgrproj = RDCOST_DBL(x->rdmult, x->rddiv, (bits >> 4), err);

  return cost_sgrproj;
}

456 457
static double find_average(uint8_t *src, int h_start, int h_end, int v_start,
                           int v_end, int stride) {
458 459 460
  uint64_t sum = 0;
  double avg = 0;
  int i, j;
461
  aom_clear_system_state();
462 463 464
  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));
465 466 467
  return avg;
}

468 469 470
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) {
471
  int i, j, k, l;
472
  double Y[WIENER_WIN2];
473 474
  const double avg =
      find_average(dgd, h_start, h_end, v_start, v_end, dgd_stride);
475

476 477
  memset(M, 0, sizeof(*M) * WIENER_WIN2);
  memset(H, 0, sizeof(*H) * WIENER_WIN2 * WIENER_WIN2);
478 479
  for (i = v_start; i < v_end; i++) {
    for (j = h_start; j < h_end; j++) {
480 481
      const double X = (double)src[i * src_stride + j] - avg;
      int idx = 0;
482 483
      for (k = -WIENER_HALFWIN; k <= WIENER_HALFWIN; k++) {
        for (l = -WIENER_HALFWIN; l <= WIENER_HALFWIN; l++) {
484 485 486 487
          Y[idx] = (double)dgd[(i + l) * dgd_stride + (j + k)] - avg;
          idx++;
        }
      }
488
      for (k = 0; k < WIENER_WIN2; ++k) {
489
        M[k] += Y[k] * X;
490 491
        H[k * WIENER_WIN2 + k] += Y[k] * Y[k];
        for (l = k + 1; l < WIENER_WIN2; ++l) {
492 493 494 495
          // 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];
496 497 498 499
        }
      }
    }
  }
500 501 502 503 504
  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];
    }
  }
505 506
}

507
#if CONFIG_HIGHBITDEPTH
508 509
static double find_average_highbd(uint16_t *src, int h_start, int h_end,
                                  int v_start, int v_end, int stride) {
510 511 512
  uint64_t sum = 0;
  double avg = 0;
  int i, j;
513
  aom_clear_system_state();
514 515 516
  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));
517 518 519
  return avg;
}

520 521 522 523
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) {
524
  int i, j, k, l;
525
  double Y[WIENER_WIN2];
526 527
  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
  uint16_t *dgd = CONVERT_TO_SHORTPTR(dgd8);
528 529
  const double avg =
      find_average_highbd(dgd, h_start, h_end, v_start, v_end, dgd_stride);
530

531 532
  memset(M, 0, sizeof(*M) * WIENER_WIN2);
  memset(H, 0, sizeof(*H) * WIENER_WIN2 * WIENER_WIN2);
533 534
  for (i = v_start; i < v_end; i++) {
    for (j = h_start; j < h_end; j++) {
535 536
      const double X = (double)src[i * src_stride + j] - avg;
      int idx = 0;
537 538
      for (k = -WIENER_HALFWIN; k <= WIENER_HALFWIN; k++) {
        for (l = -WIENER_HALFWIN; l <= WIENER_HALFWIN; l++) {
539 540 541 542
          Y[idx] = (double)dgd[(i + l) * dgd_stride + (j + k)] - avg;
          idx++;
        }
      }
543
      for (k = 0; k < WIENER_WIN2; ++k) {
544
        M[k] += Y[k] * X;
545 546
        H[k * WIENER_WIN2 + k] += Y[k] * Y[k];
        for (l = k + 1; l < WIENER_WIN2; ++l) {
547 548 549 550
          // 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];
551 552 553 554
        }
      }
    }
  }
555 556 557 558 559
  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];
    }
  }
560
}
561
#endif  // CONFIG_HIGHBITDEPTH
562 563 564 565 566

// 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;
567 568 569

  aom_clear_system_state();

570 571
  // Forward elimination
  for (k = 0; k < n - 1; k++) {
572 573 574 575 576 577 578 579 580 581 582 583 584
    // Bring the largest magitude to the diagonal position
    for (i = n - 1; i > k; i--) {
      if (fabs(A[(i - 1) * stride + k]) < fabs(A[i * stride + k])) {
        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;
      }
    }
585
    for (i = k; i < n - 1; i++) {
586
      if (fabs(A[k * stride + k]) < 1e-10) return 0;
587
      c = A[(i + 1) * stride + k] / A[k * stride + k];
588
      for (j = 0; j < n; j++) A[(i + 1) * stride + j] -= c * A[k * stride + j];
589 590 591 592 593
      b[i + 1] -= c * b[k];
    }
  }
  // Backward substitution
  for (i = n - 1; i >= 0; i--) {
594
    if (fabs(A[i * stride + i]) < 1e-10) return 0;
595
    c = 0;
596
    for (j = i + 1; j <= n - 1; j++) c += A[i * stride + j] * x[j];
597 598
    x[i] = (b[i] - c) / A[i * stride + i];
  }
599

600 601 602 603
  return 1;
}

static INLINE int wrap_index(int i) {
604
  return (i >= WIENER_HALFWIN1 ? WIENER_WIN - 1 - i : i);
605 606 607 608 609
}

// Fix vector b, update vector a
static void update_a_sep_sym(double **Mc, double **Hc, double *a, double *b) {
  int i, j;
610
  double S[WIENER_WIN];
611
  double A[WIENER_HALFWIN1], B[WIENER_HALFWIN1 * WIENER_HALFWIN1];
Aamir Anis's avatar
Aamir Anis committed
612
  int w, w2;
613 614
  memset(A, 0, sizeof(A));
  memset(B, 0, sizeof(B));
615 616
  for (i = 0; i < WIENER_WIN; i++) {
    for (j = 0; j < WIENER_WIN; ++j) {
617 618 619 620
      const int jj = wrap_index(j);
      A[jj] += Mc[i][j] * b[i];
    }
  }
621 622
  for (i = 0; i < WIENER_WIN; i++) {
    for (j = 0; j < WIENER_WIN; j++) {
623
      int k, l;
624 625
      for (k = 0; k < WIENER_WIN; ++k)
        for (l = 0; l < WIENER_WIN; ++l) {
626 627
          const int kk = wrap_index(k);
          const int ll = wrap_index(l);
628 629
          B[ll * WIENER_HALFWIN1 + kk] +=
              Hc[j * WIENER_WIN + i][k * WIENER_WIN2 + l] * b[i] * b[j];
630 631 632
        }
    }
  }
Aamir Anis's avatar
Aamir Anis committed
633
  // Normalization enforcement in the system of equations itself
634
  w = WIENER_WIN;
Aamir Anis's avatar
Aamir Anis committed
635 636
  w2 = (w >> 1) + 1;
  for (i = 0; i < w2 - 1; ++i)
637 638
    A[i] -=
        A[w2 - 1] * 2 + B[i * w2 + w2 - 1] - 2 * B[(w2 - 1) * w2 + (w2 - 1)];
Aamir Anis's avatar
Aamir Anis committed
639 640 641 642 643 644 645 646 647
  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];
648
    }
Aamir Anis's avatar
Aamir Anis committed
649
    memcpy(a, S, w * sizeof(*a));
650 651 652 653 654 655
  }
}

// Fix vector a, update vector b
static void update_b_sep_sym(double **Mc, double **Hc, double *a, double *b) {
  int i, j;
656
  double S[WIENER_WIN];
657
  double A[WIENER_HALFWIN1], B[WIENER_HALFWIN1 * WIENER_HALFWIN1];
Aamir Anis's avatar
Aamir Anis committed
658
  int w, w2;
659 660
  memset(A, 0, sizeof(A));
  memset(B, 0, sizeof(B));
661
  for (i = 0; i < WIENER_WIN; i++) {
662
    const int ii = wrap_index(i);
663
    for (j = 0; j < WIENER_WIN; j++) A[ii] += Mc[i][j] * a[j];
664 665
  }

666 667
  for (i = 0; i < WIENER_WIN; i++) {
    for (j = 0; j < WIENER_WIN; j++) {
668 669 670
      const int ii = wrap_index(i);
      const int jj = wrap_index(j);
      int k, l;
671 672 673 674
      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];
675 676
    }
  }
Aamir Anis's avatar
Aamir Anis committed
677
  // Normalization enforcement in the system of equations itself
678 679
  w = WIENER_WIN;
  w2 = WIENER_HALFWIN1;
Aamir Anis's avatar
Aamir Anis committed
680
  for (i = 0; i < w2 - 1; ++i)
681 682
    A[i] -=
        A[w2 - 1] * 2 + B[i * w2 + w2 - 1] - 2 * B[(w2 - 1) * w2 + (w2 - 1)];
Aamir Anis's avatar
Aamir Anis committed
683 684 685 686 687 688 689 690 691
  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];
692
    }
Aamir Anis's avatar
Aamir Anis committed
693
    memcpy(b, S, w * sizeof(*b));
694 695 696
  }
}

697 698
static int wiener_decompose_sep_sym(double *M, double *H, double *a,
                                    double *b) {
699
  static const double init_filt[WIENER_WIN] = {
700
    0.035623, -0.127154, 0.211436, 0.760190, 0.211436, -0.127154, 0.035623,
701 702
  };
  int i, j, iter;
703 704 705 706 707 708 709
  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;
710 711
    }
  }
712 713
  memcpy(a, init_filt, sizeof(*a) * WIENER_WIN);
  memcpy(b, init_filt, sizeof(*b) * WIENER_WIN);
714 715

  iter = 1;
716
  while (iter < NUM_WIENER_ITERS) {
717 718 719 720
    update_a_sep_sym(Mc, Hc, a, b);
    update_b_sep_sym(Mc, Hc, a, b);
    iter++;
  }
721
  return 1;
722 723
}

724
// Computes the function x'*H*x - x'*M for the learned 2D filter x, and compares
Aamir Anis's avatar
Aamir Anis committed
725 726
// against identity filters; Final score is defined as the difference between
// the function values
727 728
static double compute_score(double *M, double *H, InterpKernel vfilt,
                            InterpKernel hfilt) {
729
  double ab[WIENER_WIN * WIENER_WIN];
Aamir Anis's avatar
Aamir Anis committed
730 731 732 733
  int i, k, l;
  double P = 0, Q = 0;
  double iP = 0, iQ = 0;
  double Score, iScore;
734
  double a[WIENER_WIN], b[WIENER_WIN];
735 736 737

  aom_clear_system_state();

738 739 740 741 742 743
  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];
Aamir Anis's avatar
Aamir Anis committed
744
  }
745 746
  for (k = 0; k < WIENER_WIN; ++k) {
    for (l = 0; l < WIENER_WIN; ++l) ab[k * WIENER_WIN + l] = a[l] * b[k];
Aamir Anis's avatar
Aamir Anis committed
747
  }
748
  for (k = 0; k < WIENER_WIN2; ++k) {
Aamir Anis's avatar
Aamir Anis committed
749
    P += ab[k] * M[k];
750 751
    for (l = 0; l < WIENER_WIN2; ++l)
      Q += ab[k] * H[k * WIENER_WIN2 + l] * ab[l];
Aamir Anis's avatar
Aamir Anis committed
752 753 754
  }
  Score = Q - 2 * P;

755 756
  iP = M[WIENER_WIN2 >> 1];
  iQ = H[(WIENER_WIN2 >> 1) * WIENER_WIN2 + (WIENER_WIN2 >> 1)];
Aamir Anis's avatar
Aamir Anis committed
757 758 759 760 761
  iScore = iQ - 2 * iP;

  return Score - iScore;
}

762
static void quantize_sym_filter(double *f, InterpKernel fi) {
763
  int i;
764 765
  for (i = 0; i < WIENER_HALFWIN; ++i) {
    fi[i] = RINT(f[i] * WIENER_FILT_STEP);
766 767 768 769 770
  }
  // 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);
771 772 773 774
  // Satisfy filter constraints
  fi[WIENER_WIN - 1] = fi[0];
  fi[WIENER_WIN - 2] = fi[1];
  fi[WIENER_WIN - 3] = fi[2];
775 776
  // The central element has an implicit +WIENER_FILT_STEP
  fi[3] = -2 * (fi[0] + fi[1] + fi[2]);
777 778
}

779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814
static int count_wiener_bits(WienerInfo *wiener_info,
                             WienerInfo *ref_wiener_info) {
  int bits = 0;
  bits += aom_count_primitive_refsubexpfin(
      WIENER_FILT_TAP0_MAXV - WIENER_FILT_TAP0_MINV + 1,
      WIENER_FILT_TAP0_SUBEXP_K,
      ref_wiener_info->vfilter[0] - WIENER_FILT_TAP0_MINV,
      wiener_info->vfilter[0] - WIENER_FILT_TAP0_MINV);
  bits += aom_count_primitive_refsubexpfin(
      WIENER_FILT_TAP1_MAXV - WIENER_FILT_TAP1_MINV + 1,
      WIENER_FILT_TAP1_SUBEXP_K,
      ref_wiener_info->vfilter[1] - WIENER_FILT_TAP1_MINV,
      wiener_info->vfilter[1] - WIENER_FILT_TAP1_MINV);
  bits += aom_count_primitive_refsubexpfin(
      WIENER_FILT_TAP2_MAXV - WIENER_FILT_TAP2_MINV + 1,
      WIENER_FILT_TAP2_SUBEXP_K,
      ref_wiener_info->vfilter[2] - WIENER_FILT_TAP2_MINV,
      wiener_info->vfilter[2] - WIENER_FILT_TAP2_MINV);
  bits += aom_count_primitive_refsubexpfin(
      WIENER_FILT_TAP0_MAXV - WIENER_FILT_TAP0_MINV + 1,
      WIENER_FILT_TAP0_SUBEXP_K,
      ref_wiener_info->hfilter[0] - WIENER_FILT_TAP0_MINV,
      wiener_info->hfilter[0] - WIENER_FILT_TAP0_MINV);
  bits += aom_count_primitive_refsubexpfin(
      WIENER_FILT_TAP1_MAXV - WIENER_FILT_TAP1_MINV + 1,
      WIENER_FILT_TAP1_SUBEXP_K,
      ref_wiener_info->hfilter[1] - WIENER_FILT_TAP1_MINV,
      wiener_info->hfilter[1] - WIENER_FILT_TAP1_MINV);
  bits += aom_count_primitive_refsubexpfin(
      WIENER_FILT_TAP2_MAXV - WIENER_FILT_TAP2_MINV + 1,
      WIENER_FILT_TAP2_SUBEXP_K,
      ref_wiener_info->hfilter[2] - WIENER_FILT_TAP2_MINV,
      wiener_info->hfilter[2] - WIENER_FILT_TAP2_MINV);
  return bits;
}

815
static double search_wiener_uv(const YV12_BUFFER_CONFIG *src, AV1_COMP *cpi,
816
                               int partial_frame, int plane,
817
                               RestorationInfo *info, RestorationType *type,
818 819 820 821 822 823
                               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;
824
  double cost_wiener, cost_norestore, cost_wiener_frame, cost_norestore_frame;
825 826 827 828 829 830 831 832 833 834 835
  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;
836
  int h_start, h_end, v_start, v_end;
837 838 839
  const int ntiles =
      av1_get_rest_ntiles(width, height, cm->rst_info[1].restoration_tilesize,
                          &tile_width, &tile_height, &nhtiles, &nvtiles);
840 841
  WienerInfo ref_wiener_info;
  set_default_wiener(&ref_wiener_info);
842 843 844 845
  assert(width == dgd->uv_crop_width);
  assert(height == dgd->uv_crop_height);

  rsi[plane].frame_restoration_type = RESTORE_NONE;
846
  err = sse_restoration_frame(cm, src, cm->frame_to_show, (1 << plane));
847
  bits = 0;
848
  cost_norestore_frame = RDCOST_DBL(x->rdmult, x->rddiv, (bits >> 4), err);
849 850

  rsi[plane].frame_restoration_type = RESTORE_WIENER;
851

852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872
  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);
873
    if (plane == AOM_PLANE_U) {
874
#if CONFIG_HIGHBITDEPTH
875 876 877 878
      if (cm->use_highbitdepth)
        compute_stats_highbd(dgd->u_buffer, src->u_buffer,