restoration.c 58.4 KB
Newer Older
1
/*
Yaowu Xu's avatar
Yaowu Xu committed
2
3
4
5
6
7
8
9
 * Copyright (c) 2016, Alliance for Open Media. All rights reserved
 *
 * 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
13
14
 *
 */

#include <math.h>

Yaowu Xu's avatar
Yaowu Xu committed
15
16
#include "./aom_config.h"
#include "./aom_dsp_rtcd.h"
17
#include "./aom_scale_rtcd.h"
18
19
#include "av1/common/onyxc_int.h"
#include "av1/common/restoration.h"
Yaowu Xu's avatar
Yaowu Xu committed
20
21
#include "aom_dsp/aom_dsp_common.h"
#include "aom_mem/aom_mem.h"
22
#include "aom_ports/mem.h"
23

24
25
26
static int domaintxfmrf_vtable[DOMAINTXFMRF_ITERS][DOMAINTXFMRF_PARAMS][256];

static const int domaintxfmrf_params[DOMAINTXFMRF_PARAMS] = {
27
  32,  40,  48,  56,  64,  68,  72,  76,  80,  82,  84,  86,  88,
28
29
30
31
32
33
  90,  92,  94,  96,  97,  98,  99,  100, 101, 102, 103, 104, 105,
  106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118,
  119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 130, 132, 134,
  136, 138, 140, 142, 146, 150, 154, 158, 162, 166, 170, 174
};

34
35
const sgr_params_type sgr_params[SGRPROJ_PARAMS] = {
  // r1, eps1, r2, eps2
36
37
  { 2, 25, 1, 11 }, { 2, 35, 1, 12 }, { 2, 45, 1, 13 }, { 2, 55, 1, 14 },
  { 2, 65, 1, 15 }, { 3, 50, 2, 25 }, { 3, 60, 2, 35 }, { 3, 70, 2, 45 },
38
39
};

clang-format's avatar
clang-format committed
40
41
typedef void (*restore_func_type)(uint8_t *data8, int width, int height,
                                  int stride, RestorationInternal *rst,
42
                                  uint8_t *dst8, int dst_stride);
Yaowu Xu's avatar
Yaowu Xu committed
43
#if CONFIG_AOM_HIGHBITDEPTH
clang-format's avatar
clang-format committed
44
45
typedef void (*restore_func_highbd_type)(uint8_t *data8, int width, int height,
                                         int stride, RestorationInternal *rst,
46
47
                                         int bit_depth, uint8_t *dst8,
                                         int dst_stride);
Yaowu Xu's avatar
Yaowu Xu committed
48
#endif  // CONFIG_AOM_HIGHBITDEPTH
49

50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
int av1_alloc_restoration_struct(RestorationInfo *rst_info, int width,
                                 int height) {
  const int ntiles = av1_get_rest_ntiles(width, height, NULL, NULL, NULL, NULL);
  rst_info->restoration_type = (RestorationType *)aom_realloc(
      rst_info->restoration_type, sizeof(*rst_info->restoration_type) * ntiles);
  rst_info->wiener_info = (WienerInfo *)aom_realloc(
      rst_info->wiener_info, sizeof(*rst_info->wiener_info) * ntiles);
  assert(rst_info->wiener_info != NULL);
  rst_info->sgrproj_info = (SgrprojInfo *)aom_realloc(
      rst_info->sgrproj_info, sizeof(*rst_info->sgrproj_info) * ntiles);
  assert(rst_info->sgrproj_info != NULL);
  rst_info->domaintxfmrf_info = (DomaintxfmrfInfo *)aom_realloc(
      rst_info->domaintxfmrf_info,
      sizeof(*rst_info->domaintxfmrf_info) * ntiles);
  assert(rst_info->domaintxfmrf_info != NULL);
  return ntiles;
}

void av1_free_restoration_struct(RestorationInfo *rst_info) {
  aom_free(rst_info->restoration_type);
  rst_info->restoration_type = NULL;
  aom_free(rst_info->wiener_info);
  rst_info->wiener_info = NULL;
  aom_free(rst_info->sgrproj_info);
  rst_info->sgrproj_info = NULL;
  aom_free(rst_info->domaintxfmrf_info);
  rst_info->domaintxfmrf_info = NULL;
}

79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
static void GenDomainTxfmRFVtable() {
  int i, j;
  const double sigma_s = sqrt(2.0);
  for (i = 0; i < DOMAINTXFMRF_ITERS; ++i) {
    const int nm = (1 << (DOMAINTXFMRF_ITERS - i - 1));
    const double A = exp(-DOMAINTXFMRF_MULT / (sigma_s * nm));
    for (j = 0; j < DOMAINTXFMRF_PARAMS; ++j) {
      const double sigma_r =
          (double)domaintxfmrf_params[j] / DOMAINTXFMRF_SIGMA_SCALE;
      const double scale = sigma_s / sigma_r;
      int k;
      for (k = 0; k < 256; ++k) {
        domaintxfmrf_vtable[i][j][k] =
            RINT(DOMAINTXFMRF_VTABLE_PREC * pow(A, 1.0 + k * scale));
      }
    }
  }
}

98
void av1_loop_restoration_precal() { GenDomainTxfmRFVtable(); }
99

100
static void loop_restoration_init(RestorationInternal *rst, int kf) {
101
  rst->keyframe = kf;
102
103
}

104
void extend_frame(uint8_t *data, int width, int height, int stride) {
105
106
107
108
  uint8_t *data_p;
  int i;
  for (i = 0; i < height; ++i) {
    data_p = data + i * stride;
109
110
    memset(data_p - WIENER_HALFWIN, data_p[0], WIENER_HALFWIN);
    memset(data_p + width, data_p[width - 1], WIENER_HALFWIN);
111
  }
112
113
114
  data_p = data - WIENER_HALFWIN;
  for (i = -WIENER_HALFWIN; i < 0; ++i) {
    memcpy(data_p + i * stride, data_p, width + 2 * WIENER_HALFWIN);
115
  }
116
  for (i = height; i < height + WIENER_HALFWIN; ++i) {
117
    memcpy(data_p + i * stride, data_p + (height - 1) * stride,
118
           width + 2 * WIENER_HALFWIN);
119
120
121
  }
}

122
123
124
125
static void loop_copy_tile(uint8_t *data, int tile_idx, int subtile_idx,
                           int subtile_bits, int width, int height, int stride,
                           RestorationInternal *rst, uint8_t *dst,
                           int dst_stride) {
126
127
  const int tile_width = rst->tile_width;
  const int tile_height = rst->tile_height;
128
129
130
131
132
133
134
135
136
137
  int i;
  int h_start, h_end, v_start, v_end;
  av1_get_rest_tile_limits(tile_idx, subtile_idx, subtile_bits, rst->nhtiles,
                           rst->nvtiles, tile_width, tile_height, width, height,
                           0, 0, &h_start, &h_end, &v_start, &v_end);
  for (i = v_start; i < v_end; ++i)
    memcpy(dst + i * dst_stride + h_start, data + i * stride + h_start,
           h_end - h_start);
}

138
139
static void loop_wiener_filter_tile(uint8_t *data, int tile_idx, int width,
                                    int height, int stride,
140
                                    RestorationInternal *rst, uint8_t *dst,
141
                                    int dst_stride) {
142
143
  const int tile_width = rst->tile_width;
  const int tile_height = rst->tile_height;
144
145
  int i, j;
  int h_start, h_end, v_start, v_end;
146
147
  DECLARE_ALIGNED(16, InterpKernel, hkernel);
  DECLARE_ALIGNED(16, InterpKernel, vkernel);
148

149
  if (rst->rsi->restoration_type[tile_idx] == RESTORE_NONE) {
150
151
152
153
    loop_copy_tile(data, tile_idx, 0, 0, width, height, stride, rst, dst,
                   dst_stride);
    return;
  }
154
155
  // TODO(david.barker): Store hfilter/vfilter as an InterpKernel
  // instead of the current format. Then this can be removed.
156
157
  assert(WIENER_WIN == SUBPEL_TAPS - 1);
  for (i = 0; i < WIENER_WIN; ++i) {
158
159
    hkernel[i] = rst->rsi->wiener_info[tile_idx].hfilter[i];
    vkernel[i] = rst->rsi->wiener_info[tile_idx].vfilter[i];
160
  }
161
162
  hkernel[WIENER_WIN] = 0;
  vkernel[WIENER_WIN] = 0;
163
164
  hkernel[3] -= 128;
  vkernel[3] -= 128;
165
  av1_get_rest_tile_limits(tile_idx, 0, 0, rst->nhtiles, rst->nvtiles,
166
                           tile_width, tile_height, width, height, 0, 0,
167
                           &h_start, &h_end, &v_start, &v_end);
168
169
170
171
172
173
174
175
  // Convolve the whole tile (done in blocks here to match the requirements
  // of the vectorized convolve functions, but the result is equivalent)
  for (i = v_start; i < v_end; i += MAX_SB_SIZE)
    for (j = h_start; j < h_end; j += MAX_SB_SIZE) {
      int w = AOMMIN(MAX_SB_SIZE, (h_end - j + 15) & ~15);
      int h = AOMMIN(MAX_SB_SIZE, (v_end - i + 15) & ~15);
      const uint8_t *data_p = data + i * stride + j;
      uint8_t *dst_p = dst + i * dst_stride + j;
176
177
      aom_convolve8_add_src(data_p, stride, dst_p, dst_stride, hkernel, 16,
                            vkernel, 16, w, h);
178
179
180
    }
}

clang-format's avatar
clang-format committed
181
static void loop_wiener_filter(uint8_t *data, int width, int height, int stride,
182
183
184
185
                               RestorationInternal *rst, uint8_t *dst,
                               int dst_stride) {
  int tile_idx;
  extend_frame(data, width, height, stride);
186
  for (tile_idx = 0; tile_idx < rst->ntiles; ++tile_idx) {
187
188
    loop_wiener_filter_tile(data, tile_idx, width, height, stride, rst, dst,
                            dst_stride);
189
  }
190
}
191

192
193
/* Calculate windowed sums (if sqr=0) or sums of squares (if sqr=1)
   over the input. The window is of size (2r + 1)x(2r + 1), and we
194
   specialize to r = 1, 2, 3. A default function is used for r > 3.
195
196
197
198
199
200
201
202
203
204
205
206
207
208

   Each loop follows the same format: We keep a window's worth of input
   in individual variables and select data out of that as appropriate.
*/
static void boxsum1(int32_t *src, int width, int height, int src_stride,
                    int sqr, int32_t *dst, int dst_stride) {
  int i, j, a, b, c;

  // Vertical sum over 3-pixel regions, from src into dst.
  if (!sqr) {
    for (j = 0; j < width; ++j) {
      a = src[j];
      b = src[src_stride + j];
      c = src[2 * src_stride + j];
209

210
211
212
213
214
215
216
217
218
219
220
221
222
223
      dst[j] = a + b;
      for (i = 1; i < height - 2; ++i) {
        // Loop invariant: At the start of each iteration,
        // a = src[(i - 1) * src_stride + j]
        // b = src[(i    ) * src_stride + j]
        // c = src[(i + 1) * src_stride + j]
        dst[i * dst_stride + j] = a + b + c;
        a = b;
        b = c;
        c = src[(i + 2) * src_stride + j];
      }
      dst[i * dst_stride + j] = a + b + c;
      dst[(i + 1) * dst_stride + j] = b + c;
    }
224
  } else {
225
226
227
228
229
230
231
232
233
234
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
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
    for (j = 0; j < width; ++j) {
      a = src[j] * src[j];
      b = src[src_stride + j] * src[src_stride + j];
      c = src[2 * src_stride + j] * src[2 * src_stride + j];

      dst[j] = a + b;
      for (i = 1; i < height - 2; ++i) {
        dst[i * dst_stride + j] = a + b + c;
        a = b;
        b = c;
        c = src[(i + 2) * src_stride + j] * src[(i + 2) * src_stride + j];
      }
      dst[i * dst_stride + j] = a + b + c;
      dst[(i + 1) * dst_stride + j] = b + c;
    }
  }

  // Horizontal sum over 3-pixel regions of dst
  for (i = 0; i < height; ++i) {
    a = dst[i * dst_stride];
    b = dst[i * dst_stride + 1];
    c = dst[i * dst_stride + 2];

    dst[i * dst_stride] = a + b;
    for (j = 1; j < width - 2; ++j) {
      // Loop invariant: At the start of each iteration,
      // a = src[i * src_stride + (j - 1)]
      // b = src[i * src_stride + (j    )]
      // c = src[i * src_stride + (j + 1)]
      dst[i * dst_stride + j] = a + b + c;
      a = b;
      b = c;
      c = dst[i * dst_stride + (j + 2)];
    }
    dst[i * dst_stride + j] = a + b + c;
    dst[i * dst_stride + (j + 1)] = b + c;
  }
}

static void boxsum2(int32_t *src, int width, int height, int src_stride,
                    int sqr, int32_t *dst, int dst_stride) {
  int i, j, a, b, c, d, e;

  // Vertical sum over 5-pixel regions, from src into dst.
  if (!sqr) {
    for (j = 0; j < width; ++j) {
      a = src[j];
      b = src[src_stride + j];
      c = src[2 * src_stride + j];
      d = src[3 * src_stride + j];
      e = src[4 * src_stride + j];

      dst[j] = a + b + c;
      dst[dst_stride + j] = a + b + c + d;
      for (i = 2; i < height - 3; ++i) {
        // Loop invariant: At the start of each iteration,
        // a = src[(i - 2) * src_stride + j]
        // b = src[(i - 1) * src_stride + j]
        // c = src[(i    ) * src_stride + j]
        // d = src[(i + 1) * src_stride + j]
        // e = src[(i + 2) * src_stride + j]
        dst[i * dst_stride + j] = a + b + c + d + e;
        a = b;
        b = c;
        c = d;
        d = e;
        e = src[(i + 3) * src_stride + j];
      }
      dst[i * dst_stride + j] = a + b + c + d + e;
      dst[(i + 1) * dst_stride + j] = b + c + d + e;
      dst[(i + 2) * dst_stride + j] = c + d + e;
    }
  } else {
    for (j = 0; j < width; ++j) {
      a = src[j] * src[j];
      b = src[src_stride + j] * src[src_stride + j];
      c = src[2 * src_stride + j] * src[2 * src_stride + j];
      d = src[3 * src_stride + j] * src[3 * src_stride + j];
      e = src[4 * src_stride + j] * src[4 * src_stride + j];

      dst[j] = a + b + c;
      dst[dst_stride + j] = a + b + c + d;
      for (i = 2; i < height - 3; ++i) {
        dst[i * dst_stride + j] = a + b + c + d + e;
        a = b;
        b = c;
        c = d;
        d = e;
        e = src[(i + 3) * src_stride + j] * src[(i + 3) * src_stride + j];
      }
      dst[i * dst_stride + j] = a + b + c + d + e;
      dst[(i + 1) * dst_stride + j] = b + c + d + e;
      dst[(i + 2) * dst_stride + j] = c + d + e;
    }
  }

  // Horizontal sum over 5-pixel regions of dst
  for (i = 0; i < height; ++i) {
    a = dst[i * dst_stride];
    b = dst[i * dst_stride + 1];
    c = dst[i * dst_stride + 2];
    d = dst[i * dst_stride + 3];
    e = dst[i * dst_stride + 4];

    dst[i * dst_stride] = a + b + c;
    dst[i * dst_stride + 1] = a + b + c + d;
    for (j = 2; j < width - 3; ++j) {
      // Loop invariant: At the start of each iteration,
      // a = src[i * src_stride + (j - 2)]
      // b = src[i * src_stride + (j - 1)]
      // c = src[i * src_stride + (j    )]
      // d = src[i * src_stride + (j + 1)]
      // e = src[i * src_stride + (j + 2)]
      dst[i * dst_stride + j] = a + b + c + d + e;
      a = b;
      b = c;
      c = d;
      d = e;
      e = dst[i * dst_stride + (j + 3)];
    }
    dst[i * dst_stride + j] = a + b + c + d + e;
    dst[i * dst_stride + (j + 1)] = b + c + d + e;
    dst[i * dst_stride + (j + 2)] = c + d + e;
  }
}

351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
static void boxsum3(int32_t *src, int width, int height, int src_stride,
                    int sqr, int32_t *dst, int dst_stride) {
  int i, j, a, b, c, d, e, f, g;

  // Vertical sum over 7-pixel regions, from src into dst.
  if (!sqr) {
    for (j = 0; j < width; ++j) {
      a = src[j];
      b = src[1 * src_stride + j];
      c = src[2 * src_stride + j];
      d = src[3 * src_stride + j];
      e = src[4 * src_stride + j];
      f = src[5 * src_stride + j];
      g = src[6 * src_stride + j];

      dst[j] = a + b + c + d;
      dst[dst_stride + j] = a + b + c + d + e;
      dst[2 * dst_stride + j] = a + b + c + d + e + f;
      for (i = 3; i < height - 4; ++i) {
        dst[i * dst_stride + j] = a + b + c + d + e + f + g;
        a = b;
        b = c;
        c = d;
        d = e;
        e = f;
        f = g;
        g = src[(i + 4) * src_stride + j];
      }
      dst[i * dst_stride + j] = a + b + c + d + e + f + g;
      dst[(i + 1) * dst_stride + j] = b + c + d + e + f + g;
      dst[(i + 2) * dst_stride + j] = c + d + e + f + g;
      dst[(i + 3) * dst_stride + j] = d + e + f + g;
    }
  } else {
    for (j = 0; j < width; ++j) {
      a = src[j] * src[j];
      b = src[1 * src_stride + j] * src[1 * src_stride + j];
      c = src[2 * src_stride + j] * src[2 * src_stride + j];
      d = src[3 * src_stride + j] * src[3 * src_stride + j];
      e = src[4 * src_stride + j] * src[4 * src_stride + j];
      f = src[5 * src_stride + j] * src[5 * src_stride + j];
      g = src[6 * src_stride + j] * src[6 * src_stride + j];

      dst[j] = a + b + c + d;
      dst[dst_stride + j] = a + b + c + d + e;
      dst[2 * dst_stride + j] = a + b + c + d + e + f;
      for (i = 3; i < height - 4; ++i) {
        dst[i * dst_stride + j] = a + b + c + d + e + f + g;
        a = b;
        b = c;
        c = d;
        d = e;
        e = f;
        f = g;
        g = src[(i + 4) * src_stride + j] * src[(i + 4) * src_stride + j];
      }
      dst[i * dst_stride + j] = a + b + c + d + e + f + g;
      dst[(i + 1) * dst_stride + j] = b + c + d + e + f + g;
      dst[(i + 2) * dst_stride + j] = c + d + e + f + g;
      dst[(i + 3) * dst_stride + j] = d + e + f + g;
    }
  }

  // Horizontal sum over 7-pixel regions of dst
  for (i = 0; i < height; ++i) {
    a = dst[i * dst_stride];
    b = dst[i * dst_stride + 1];
    c = dst[i * dst_stride + 2];
    d = dst[i * dst_stride + 3];
    e = dst[i * dst_stride + 4];
    f = dst[i * dst_stride + 5];
    g = dst[i * dst_stride + 6];

    dst[i * dst_stride] = a + b + c + d;
    dst[i * dst_stride + 1] = a + b + c + d + e;
    dst[i * dst_stride + 2] = a + b + c + d + e + f;
    for (j = 3; j < width - 4; ++j) {
      dst[i * dst_stride + j] = a + b + c + d + e + f + g;
      a = b;
      b = c;
      c = d;
      d = e;
      e = f;
      f = g;
      g = dst[i * dst_stride + (j + 4)];
    }
    dst[i * dst_stride + j] = a + b + c + d + e + f + g;
    dst[i * dst_stride + (j + 1)] = b + c + d + e + f + g;
    dst[i * dst_stride + (j + 2)] = c + d + e + f + g;
    dst[i * dst_stride + (j + 3)] = d + e + f + g;
  }
}

// Generic version for any r. To be removed after experiments are done.
static void boxsumr(int32_t *src, int width, int height, int src_stride, int r,
                    int sqr, int32_t *dst, int dst_stride) {
  int32_t *tmp = aom_malloc(width * height * sizeof(*tmp));
  int tmp_stride = width;
  int i, j;
  if (sqr) {
    for (j = 0; j < width; ++j) tmp[j] = src[j] * src[j];
    for (j = 0; j < width; ++j)
      for (i = 1; i < height; ++i)
        tmp[i * tmp_stride + j] =
            tmp[(i - 1) * tmp_stride + j] +
            src[i * src_stride + j] * src[i * src_stride + j];
  } else {
    memcpy(tmp, src, sizeof(*tmp) * width);
    for (j = 0; j < width; ++j)
      for (i = 1; i < height; ++i)
        tmp[i * tmp_stride + j] =
            tmp[(i - 1) * tmp_stride + j] + src[i * src_stride + j];
  }
  for (i = 0; i <= r; ++i)
    memcpy(&dst[i * dst_stride], &tmp[(i + r) * tmp_stride],
           sizeof(*tmp) * width);
  for (i = r + 1; i < height - r; ++i)
    for (j = 0; j < width; ++j)
      dst[i * dst_stride + j] =
          tmp[(i + r) * tmp_stride + j] - tmp[(i - r - 1) * tmp_stride + j];
  for (i = height - r; i < height; ++i)
    for (j = 0; j < width; ++j)
      dst[i * dst_stride + j] = tmp[(height - 1) * tmp_stride + j] -
                                tmp[(i - r - 1) * tmp_stride + j];

  for (i = 0; i < height; ++i) tmp[i * tmp_stride] = dst[i * dst_stride];
  for (i = 0; i < height; ++i)
    for (j = 1; j < width; ++j)
      tmp[i * tmp_stride + j] =
          tmp[i * tmp_stride + j - 1] + dst[i * src_stride + j];

  for (j = 0; j <= r; ++j)
    for (i = 0; i < height; ++i)
      dst[i * dst_stride + j] = tmp[i * tmp_stride + j + r];
  for (j = r + 1; j < width - r; ++j)
    for (i = 0; i < height; ++i)
      dst[i * dst_stride + j] =
          tmp[i * tmp_stride + j + r] - tmp[i * tmp_stride + j - r - 1];
  for (j = width - r; j < width; ++j)
    for (i = 0; i < height; ++i)
      dst[i * dst_stride + j] =
          tmp[i * tmp_stride + width - 1] - tmp[i * tmp_stride + j - r - 1];
  aom_free(tmp);
}

496
497
498
499
500
501
static void boxsum(int32_t *src, int width, int height, int src_stride, int r,
                   int sqr, int32_t *dst, int dst_stride) {
  if (r == 1)
    boxsum1(src, width, height, src_stride, sqr, dst, dst_stride);
  else if (r == 2)
    boxsum2(src, width, height, src_stride, sqr, dst, dst_stride);
502
503
504
505
  else if (r == 3)
    boxsum3(src, width, height, src_stride, sqr, dst, dst_stride);
  else
    boxsumr(src, width, height, src_stride, r, sqr, dst, dst_stride);
506
507
508
509
}

static void boxnum(int width, int height, int r, int8_t *num, int num_stride) {
  int i, j;
510
511
512
  for (i = 0; i <= r; ++i) {
    for (j = 0; j <= r; ++j) {
      num[i * num_stride + j] = (r + 1 + i) * (r + 1 + j);
513
514
515
516
517
518
      num[i * num_stride + (width - 1 - j)] = num[i * num_stride + j];
      num[(height - 1 - i) * num_stride + j] = num[i * num_stride + j];
      num[(height - 1 - i) * num_stride + (width - 1 - j)] =
          num[i * num_stride + j];
    }
  }
519
520
  for (j = 0; j <= r; ++j) {
    const int val = (2 * r + 1) * (r + 1 + j);
521
522
523
524
525
    for (i = r + 1; i < height - r; ++i) {
      num[i * num_stride + j] = val;
      num[i * num_stride + (width - 1 - j)] = val;
    }
  }
526
527
  for (i = 0; i <= r; ++i) {
    const int val = (2 * r + 1) * (r + 1 + i);
528
529
530
531
532
533
534
    for (j = r + 1; j < width - r; ++j) {
      num[i * num_stride + j] = val;
      num[(height - 1 - i) * num_stride + j] = val;
    }
  }
  for (i = r + 1; i < height - r; ++i) {
    for (j = r + 1; j < width - r; ++j) {
535
      num[i * num_stride + j] = (2 * r + 1) * (2 * r + 1);
536
537
538
539
540
541
542
543
544
545
    }
  }
}

void decode_xq(int *xqd, int *xq) {
  xq[0] = -xqd[0];
  xq[1] = (1 << SGRPROJ_PRJ_BITS) - xq[0] - xqd[1];
}

#define APPROXIMATE_SGR 1
546
void av1_selfguided_restoration(int32_t *dgd, int width, int height, int stride,
547
548
549
                                int bit_depth, int r, int eps,
                                int32_t *tmpbuf) {
  int32_t *A = tmpbuf;
550
  int32_t *B = A + RESTORATION_TILEPELS_MAX;
551
552
553
554
  int8_t num[RESTORATION_TILEPELS_MAX];
  int i, j;
  eps <<= 2 * (bit_depth - 8);

555
556
557
558
559
  // Don't filter tiles with dimensions < 5 on any axis
  if ((width < 5) || (height < 5)) return;

  boxsum(dgd, width, height, stride, r, 0, B, width);
  boxsum(dgd, width, height, stride, r, 1, A, width);
560
  boxnum(width, height, r, num, width);
561
562
  // The following loop is optimized assuming r <= 2. If we allow
  // r > 2, then the loop will need modifying.
563
  assert(r <= 3);
564
565
566
567
  for (i = 0; i < height; ++i) {
    for (j = 0; j < width; ++j) {
      const int k = i * width + j;
      const int n = num[k];
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
      // Assuming that we only allow up to 12-bit depth and r <= 2,
      // we calculate p = n^2 * Var(n-pixel block of original image)
      // (where n = 2 * r + 1 <= 5).
      //
      // There is an inequality which gives a bound on the variance:
      // https://en.wikipedia.org/wiki/Popoviciu's_inequality_on_variances
      // In this case, since each pixel is in the range [0, 2^12),
      // the variance is at most 1/4 * (2^12)^2 = 2^22.
      // Then p <= 25^2 * 2^22 < 2^32, and also q <= p + 25^2 * 68 < 2^32.
      //
      // The point of all this is to guarantee that q < 2^32, so that
      // platforms with a 64-bit by 32-bit divide unit (eg, x86)
      // can do the division by q more efficiently.
      const uint32_t p = (uint32_t)((uint64_t)A[k] * n - (uint64_t)B[k] * B[k]);
      const uint32_t q = (uint32_t)(p + n * n * eps);
      assert((uint64_t)A[k] * n - (uint64_t)B[k] * B[k] < (25 * 25U << 22));
      A[k] = (int32_t)(((uint64_t)p << SGRPROJ_SGR_BITS) + (q >> 1)) / q;
585
586
587
588
589
590
591
592
593
594
      B[k] = ((SGRPROJ_SGR - A[k]) * B[k] + (n >> 1)) / n;
    }
  }
#if APPROXIMATE_SGR
  i = 0;
  j = 0;
  {
    const int k = i * width + j;
    const int l = i * stride + j;
    const int nb = 3;
595
    const int32_t a =
596
        3 * A[k] + 2 * A[k + 1] + 2 * A[k + width] + A[k + width + 1];
597
    const int32_t b =
598
        3 * B[k] + 2 * B[k + 1] + 2 * B[k + width] + B[k + width + 1];
599
    const int32_t v =
600
601
602
603
604
605
606
607
608
        (((a * dgd[l] + b) << SGRPROJ_RST_BITS) + (1 << nb) / 2) >> nb;
    dgd[l] = ROUND_POWER_OF_TWO(v, SGRPROJ_SGR_BITS);
  }
  i = 0;
  j = width - 1;
  {
    const int k = i * width + j;
    const int l = i * stride + j;
    const int nb = 3;
609
    const int32_t a =
610
        3 * A[k] + 2 * A[k - 1] + 2 * A[k + width] + A[k + width - 1];
611
    const int32_t b =
612
        3 * B[k] + 2 * B[k - 1] + 2 * B[k + width] + B[k + width - 1];
613
    const int32_t v =
614
615
616
617
618
619
620
621
622
        (((a * dgd[l] + b) << SGRPROJ_RST_BITS) + (1 << nb) / 2) >> nb;
    dgd[l] = ROUND_POWER_OF_TWO(v, SGRPROJ_SGR_BITS);
  }
  i = height - 1;
  j = 0;
  {
    const int k = i * width + j;
    const int l = i * stride + j;
    const int nb = 3;
623
    const int32_t a =
624
        3 * A[k] + 2 * A[k + 1] + 2 * A[k - width] + A[k - width + 1];
625
    const int32_t b =
626
        3 * B[k] + 2 * B[k + 1] + 2 * B[k - width] + B[k - width + 1];
627
    const int32_t v =
628
629
630
631
632
633
634
635
636
        (((a * dgd[l] + b) << SGRPROJ_RST_BITS) + (1 << nb) / 2) >> nb;
    dgd[l] = ROUND_POWER_OF_TWO(v, SGRPROJ_SGR_BITS);
  }
  i = height - 1;
  j = width - 1;
  {
    const int k = i * width + j;
    const int l = i * stride + j;
    const int nb = 3;
637
    const int32_t a =
638
        3 * A[k] + 2 * A[k - 1] + 2 * A[k - width] + A[k - width - 1];
639
    const int32_t b =
640
        3 * B[k] + 2 * B[k - 1] + 2 * B[k - width] + B[k - width - 1];
641
    const int32_t v =
642
643
644
645
646
647
648
649
        (((a * dgd[l] + b) << SGRPROJ_RST_BITS) + (1 << nb) / 2) >> nb;
    dgd[l] = ROUND_POWER_OF_TWO(v, SGRPROJ_SGR_BITS);
  }
  i = 0;
  for (j = 1; j < width - 1; ++j) {
    const int k = i * width + j;
    const int l = i * stride + j;
    const int nb = 3;
650
    const int32_t a = A[k] + 2 * (A[k - 1] + A[k + 1]) + A[k + width] +
651
                      A[k + width - 1] + A[k + width + 1];
652
    const int32_t b = B[k] + 2 * (B[k - 1] + B[k + 1]) + B[k + width] +
653
                      B[k + width - 1] + B[k + width + 1];
654
    const int32_t v =
655
656
657
658
659
660
661
662
        (((a * dgd[l] + b) << SGRPROJ_RST_BITS) + (1 << nb) / 2) >> nb;
    dgd[l] = ROUND_POWER_OF_TWO(v, SGRPROJ_SGR_BITS);
  }
  i = height - 1;
  for (j = 1; j < width - 1; ++j) {
    const int k = i * width + j;
    const int l = i * stride + j;
    const int nb = 3;
663
    const int32_t a = A[k] + 2 * (A[k - 1] + A[k + 1]) + A[k - width] +
664
                      A[k - width - 1] + A[k - width + 1];
665
    const int32_t b = B[k] + 2 * (B[k - 1] + B[k + 1]) + B[k - width] +
666
                      B[k - width - 1] + B[k - width + 1];
667
    const int32_t v =
668
669
670
671
672
673
674
675
        (((a * dgd[l] + b) << SGRPROJ_RST_BITS) + (1 << nb) / 2) >> nb;
    dgd[l] = ROUND_POWER_OF_TWO(v, SGRPROJ_SGR_BITS);
  }
  j = 0;
  for (i = 1; i < height - 1; ++i) {
    const int k = i * width + j;
    const int l = i * stride + j;
    const int nb = 3;
676
    const int32_t a = A[k] + 2 * (A[k - width] + A[k + width]) + A[k + 1] +
677
                      A[k - width + 1] + A[k + width + 1];
678
    const int32_t b = B[k] + 2 * (B[k - width] + B[k + width]) + B[k + 1] +
679
                      B[k - width + 1] + B[k + width + 1];
680
    const int32_t v =
681
682
683
684
685
686
687
688
        (((a * dgd[l] + b) << SGRPROJ_RST_BITS) + (1 << nb) / 2) >> nb;
    dgd[l] = ROUND_POWER_OF_TWO(v, SGRPROJ_SGR_BITS);
  }
  j = width - 1;
  for (i = 1; i < height - 1; ++i) {
    const int k = i * width + j;
    const int l = i * stride + j;
    const int nb = 3;
689
    const int32_t a = A[k] + 2 * (A[k - width] + A[k + width]) + A[k - 1] +
690
                      A[k - width - 1] + A[k + width - 1];
691
    const int32_t b = B[k] + 2 * (B[k - width] + B[k + width]) + B[k - 1] +
692
                      B[k - width - 1] + B[k + width - 1];
693
    const int32_t v =
694
695
696
697
698
699
700
701
        (((a * dgd[l] + b) << SGRPROJ_RST_BITS) + (1 << nb) / 2) >> nb;
    dgd[l] = ROUND_POWER_OF_TWO(v, SGRPROJ_SGR_BITS);
  }
  for (i = 1; i < height - 1; ++i) {
    for (j = 1; j < width - 1; ++j) {
      const int k = i * width + j;
      const int l = i * stride + j;
      const int nb = 5;
702
      const int32_t a =
703
704
705
706
          (A[k] + A[k - 1] + A[k + 1] + A[k - width] + A[k + width]) * 4 +
          (A[k - 1 - width] + A[k - 1 + width] + A[k + 1 - width] +
           A[k + 1 + width]) *
              3;
707
      const int32_t b =
708
709
710
711
          (B[k] + B[k - 1] + B[k + 1] + B[k - width] + B[k + width]) * 4 +
          (B[k - 1 - width] + B[k - 1 + width] + B[k + 1 - width] +
           B[k + 1 + width]) *
              3;
712
      const int32_t v =
713
714
715
716
717
718
          (((a * dgd[l] + b) << SGRPROJ_RST_BITS) + (1 << nb) / 2) >> nb;
      dgd[l] = ROUND_POWER_OF_TWO(v, SGRPROJ_SGR_BITS);
    }
  }
#else
  if (r > 1) boxnum(width, height, r = 1, num, width);
719
720
  boxsum(A, width, height, width, r, 0, A, width);
  boxsum(B, width, height, width, r, 0, B, width);
721
722
723
724
725
  for (i = 0; i < height; ++i) {
    for (j = 0; j < width; ++j) {
      const int k = i * width + j;
      const int l = i * stride + j;
      const int n = num[k];
726
      const int32_t v =
727
728
729
730
731
732
733
          (((A[k] * dgd[l] + B[k]) << SGRPROJ_RST_BITS) + (n >> 1)) / n;
      dgd[l] = ROUND_POWER_OF_TWO(v, SGRPROJ_SGR_BITS);
    }
  }
#endif  // APPROXIMATE_SGR
}

734
static void apply_selfguided_restoration(uint8_t *dat, int width, int height,
735
                                         int stride, int bit_depth, int eps,
736
                                         int *xqd, uint8_t *dst, int dst_stride,
737
                                         int32_t *tmpbuf) {
738
  int xq[2];
739
  int32_t *flt1 = tmpbuf;
740
  int32_t *flt2 = flt1 + RESTORATION_TILEPELS_MAX;
741
  int32_t *tmpbuf2 = flt2 + RESTORATION_TILEPELS_MAX;
742
  int i, j;
743
  assert(width * height <= RESTORATION_TILEPELS_MAX);
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
  for (i = 0; i < height; ++i) {
    for (j = 0; j < width; ++j) {
      flt1[i * width + j] = dat[i * stride + j];
      flt2[i * width + j] = dat[i * stride + j];
    }
  }
  av1_selfguided_restoration(flt1, width, height, width, bit_depth,
                             sgr_params[eps].r1, sgr_params[eps].e1, tmpbuf2);
  av1_selfguided_restoration(flt2, width, height, width, bit_depth,
                             sgr_params[eps].r2, sgr_params[eps].e2, tmpbuf2);
  decode_xq(xqd, xq);
  for (i = 0; i < height; ++i) {
    for (j = 0; j < width; ++j) {
      const int k = i * width + j;
      const int l = i * stride + j;
759
760
761
762
      const int m = i * dst_stride + j;
      const int32_t u = ((int32_t)dat[l] << SGRPROJ_RST_BITS);
      const int32_t f1 = (int32_t)flt1[k] - u;
      const int32_t f2 = (int32_t)flt2[k] - u;
763
764
765
      const int64_t v = xq[0] * f1 + xq[1] * f2 + (u << SGRPROJ_PRJ_BITS);
      const int16_t w =
          (int16_t)ROUND_POWER_OF_TWO(v, SGRPROJ_PRJ_BITS + SGRPROJ_RST_BITS);
766
      dst[m] = clip_pixel(w);
767
768
769
770
771
772
    }
  }
}

static void loop_sgrproj_filter_tile(uint8_t *data, int tile_idx, int width,
                                     int height, int stride,
773
774
                                     RestorationInternal *rst, uint8_t *dst,
                                     int dst_stride) {
775
776
  const int tile_width = rst->tile_width;
  const int tile_height = rst->tile_height;
777
  int h_start, h_end, v_start, v_end;
778
  uint8_t *data_p, *dst_p;
779

780
  if (rst->rsi->restoration_type[tile_idx] == RESTORE_NONE) {
781
782
783
784
    loop_copy_tile(data, tile_idx, 0, 0, width, height, stride, rst, dst,
                   dst_stride);
    return;
  }
785
786
787
788
  av1_get_rest_tile_limits(tile_idx, 0, 0, rst->nhtiles, rst->nvtiles,
                           tile_width, tile_height, width, height, 0, 0,
                           &h_start, &h_end, &v_start, &v_end);
  data_p = data + h_start + v_start * stride;
789
  dst_p = dst + h_start + v_start * dst_stride;
790
791
792
  apply_selfguided_restoration(data_p, h_end - h_start, v_end - v_start, stride,
                               8, rst->rsi->sgrproj_info[tile_idx].ep,
                               rst->rsi->sgrproj_info[tile_idx].xqd, dst_p,
793
                               dst_stride, rst->tmpbuf);
794
795
796
797
}

static void loop_sgrproj_filter(uint8_t *data, int width, int height,
                                int stride, RestorationInternal *rst,
798
                                uint8_t *dst, int dst_stride) {
799
800
  int tile_idx;
  for (tile_idx = 0; tile_idx < rst->ntiles; ++tile_idx) {
801
802
    loop_sgrproj_filter_tile(data, tile_idx, width, height, stride, rst, dst,
                             dst_stride);
803
804
805
  }
}

806
807
808
809
810
811
812
static void apply_domaintxfmrf(int iter, int param, uint8_t *diff_right,
                               uint8_t *diff_down, int width, int height,
                               int32_t *dat, int dat_stride) {
  int i, j, acc;
  // Do first row separately, to initialize the top to bottom filter
  i = 0;
  {
813
    // left to right
814
815
816
817
818
819
820
821
822
823
    acc = dat[i * dat_stride] * DOMAINTXFMRF_VTABLE_PREC;
    dat[i * dat_stride] = acc;
    for (j = 1; j < width; ++j) {
      const int in = dat[i * dat_stride + j];
      const int diff =
          diff_right[i * width + j - 1];  // Left absolute difference
      const int v = domaintxfmrf_vtable[iter][param][diff];
      acc = in * (DOMAINTXFMRF_VTABLE_PREC - v) +
            ROUND_POWER_OF_TWO(v * acc, DOMAINTXFMRF_VTABLE_PRECBITS);
      dat[i * dat_stride + j] = acc;
824
825
    }
    // right to left
826
827
828
829
830
831
832
    for (j = width - 2; j >= 0; --j) {
      const int in = dat[i * dat_stride + j];
      const int diff = diff_right[i * width + j];  // Right absolute difference
      const int v = domaintxfmrf_vtable[iter][param][diff];
      acc = ROUND_POWER_OF_TWO(in * (DOMAINTXFMRF_VTABLE_PREC - v) + acc * v,
                               DOMAINTXFMRF_VTABLE_PRECBITS);
      dat[i * dat_stride + j] = acc;
833
834
835
    }
  }

836
837
838
839
840
841
842
843
844
845
846
847
  for (i = 1; i < height; ++i) {
    // left to right
    acc = dat[i * dat_stride] * DOMAINTXFMRF_VTABLE_PREC;
    dat[i * dat_stride] = acc;
    for (j = 1; j < width; ++j) {
      const int in = dat[i * dat_stride + j];
      const int diff =
          diff_right[i * width + j - 1];  // Left absolute difference
      const int v = domaintxfmrf_vtable[iter][param][diff];
      acc = in * (DOMAINTXFMRF_VTABLE_PREC - v) +
            ROUND_POWER_OF_TWO(v * acc, DOMAINTXFMRF_VTABLE_PRECBITS);
      dat[i * dat_stride + j] = acc;
848
    }
849
850
851
852
853
854
855
856
    // right to left
    for (j = width - 2; j >= 0; --j) {
      const int in = dat[i * dat_stride + j];
      const int diff = diff_right[i * width + j];  // Right absolute difference
      const int v = domaintxfmrf_vtable[iter][param][diff];
      acc = ROUND_POWER_OF_TWO(in * (DOMAINTXFMRF_VTABLE_PREC - v) + acc * v,
                               DOMAINTXFMRF_VTABLE_PRECBITS);
      dat[i * dat_stride + j] = acc;
857
    }
858
    // top to bottom
859
    for (j = 0; j < width; ++j) {
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
      const int in = dat[i * dat_stride + j];
      const int in_above = dat[(i - 1) * dat_stride + j];
      const int diff =
          diff_down[(i - 1) * width + j];  // Upward absolute difference
      const int v = domaintxfmrf_vtable[iter][param][diff];
      acc =
          ROUND_POWER_OF_TWO(in * (DOMAINTXFMRF_VTABLE_PREC - v) + in_above * v,
                             DOMAINTXFMRF_VTABLE_PRECBITS);
      dat[i * dat_stride + j] = acc;
    }
  }
  for (j = 0; j < width; ++j) {
    // bottom to top + output rounding
    acc = dat[(height - 1) * dat_stride + j];
    dat[(height - 1) * dat_stride + j] =
        ROUND_POWER_OF_TWO(acc, DOMAINTXFMRF_VTABLE_PRECBITS);
    for (i = height - 2; i >= 0; --i) {
      const int in = dat[i * dat_stride + j];
      const int diff =
          diff_down[i * width + j];  // Downward absolute difference
      const int v = domaintxfmrf_vtable[iter][param][diff];
      acc = ROUND_POWER_OF_TWO(in * (DOMAINTXFMRF_VTABLE_PREC - v) + acc * v,
                               DOMAINTXFMRF_VTABLE_PRECBITS);
      dat[i * dat_stride + j] =
          ROUND_POWER_OF_TWO(acc, DOMAINTXFMRF_VTABLE_PRECBITS);
885
886
887
888
889
    }
  }
}

void av1_domaintxfmrf_restoration(uint8_t *dgd, int width, int height,
890
                                  int stride, int param, uint8_t *dst,
891
892
                                  int dst_stride, int32_t *tmpbuf) {
  int32_t *dat = tmpbuf;
893
894
  uint8_t *diff_right = (uint8_t *)(tmpbuf + RESTORATION_TILEPELS_MAX);
  uint8_t *diff_down = diff_right + RESTORATION_TILEPELS_MAX;
895
  int i, j, t;
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912

  for (i = 0; i < height; ++i) {
    int cur_px = dgd[i * stride];
    for (j = 0; j < width - 1; ++j) {
      const int next_px = dgd[i * stride + j + 1];
      diff_right[i * width + j] = abs(cur_px - next_px);
      cur_px = next_px;
    }
  }
  for (j = 0; j < width; ++j) {
    int cur_px = dgd[j];
    for (i = 0; i < height - 1; ++i) {
      const int next_px = dgd[(i + 1) * stride + j];
      diff_down[i * width + j] = abs(cur_px - next_px);
      cur_px = next_px;
    }
  }
913
914
915
916
917
  for (i = 0; i < height; ++i) {
    for (j = 0; j < width; ++j) {
      dat[i * width + j] = dgd[i * stride + j];
    }
  }
918

919
  for (t = 0; t < DOMAINTXFMRF_ITERS; ++t) {
920
921
    apply_domaintxfmrf(t, param, diff_right, diff_down, width, height, dat,
                       width);
922
923
924
  }
  for (i = 0; i < height; ++i) {
    for (j = 0; j < width; ++j) {
925
      dst[i * dst_stride + j] = clip_pixel(dat[i * width + j]);
926
927
928
929
930
931
932
    }
  }
}

static void loop_domaintxfmrf_filter_tile(uint8_t *data, int tile_idx,
                                          int width, int height, int stride,
                                          RestorationInternal *rst,
933
                                          uint8_t *dst, int dst_stride) {
934
935
  const int tile_width = rst->tile_width;
  const int tile_height = rst->tile_height;
936
  int h_start, h_end, v_start, v_end;
937
  int32_t *tmpbuf = (int32_t *)rst->tmpbuf;
938

939
  if (rst->rsi->restoration_type[tile_idx] == RESTORE_NONE) {
940
941
942
943
    loop_copy_tile(data, tile_idx, 0, 0, width, height, stride, rst, dst,
                   dst_stride);
    return;
  }
944
945
946
  av1_get_rest_tile_limits(tile_idx, 0, 0, rst->nhtiles, rst->nvtiles,
                           tile_width, tile_height, width, height, 0, 0,
                           &h_start, &h_end, &v_start, &v_end);
947
948
949
  av1_domaintxfmrf_restoration(
      data + h_start + v_start * stride, h_end - h_start, v_end - v_start,
      stride, rst->rsi->domaintxfmrf_info[tile_idx].sigma_r,
950
      dst + h_start + v_start * dst_stride, dst_stride, tmpbuf);
951
952
953
954
}

static void loop_domaintxfmrf_filter(uint8_t *data, int width, int height,
                                     int stride, RestorationInternal *rst,
955
                                     uint8_t *dst, int dst_stride) {
956
957
958
  int tile_idx;
  for (tile_idx = 0; tile_idx < rst->ntiles; ++tile_idx) {
    loop_domaintxfmrf_filter_tile(data, tile_idx, width, height, stride, rst,
959
                                  dst, dst_stride);
960
961
962
  }
}

963
964
static void loop_switchable_filter(uint8_t *data, int width, int height,
                                   int stride, RestorationInternal *rst,
965
                                   uint8_t *dst, int dst_stride) {
966
967
  int tile_idx;
  extend_frame(data, width, height, stride);
968
  for (tile_idx = 0; tile_idx < rst->ntiles; ++tile_idx) {
969
970
971
972
    if (rst->rsi->restoration_type[tile_idx] == RESTORE_NONE) {
      loop_copy_tile(data, tile_idx, 0, 0, width, height, stride, rst, dst,
                     dst_stride);
    } else if (rst->rsi->restoration_type[tile_idx] == RESTORE_WIENER) {
973
974
      loop_wiener_filter_tile(data, tile_idx, width, height, stride, rst, dst,
                              dst_stride);
975
    } else if (rst->rsi->restoration_type[tile_idx] == RESTORE_SGRPROJ) {
976
977
      loop_sgrproj_filter_tile(data, tile_idx, width, height, stride, rst, dst,
                               dst_stride);
978
979
    } else if (rst->rsi->restoration_type[tile_idx] == RESTORE_DOMAINTXFMRF) {
      loop_domaintxfmrf_filter_tile(data, tile_idx, width, height, stride, rst,
980
                                    dst, dst_stride);
981
    }
982
983
984
  }
}

Yaowu Xu's avatar
Yaowu Xu committed
985
#if CONFIG_AOM_HIGHBITDEPTH
986
void extend_frame_highbd(uint16_t *data, int width, int height, int stride) {
987
988
989
990
  uint16_t *data_p;
  int i, j;
  for (i = 0; i < height; ++i) {
    data_p = data + i * stride;
991
992
    for (j = -WIENER_HALFWIN; j < 0; ++j) data_p[j] = data_p[0];
    for (j = width; j < width + WIENER_HALFWIN; ++j)
993
994
      data_p[j] = data_p[width - 1];
  }
995
996
  data_p = data - WIENER_HALFWIN;
  for (i = -WIENER_HALFWIN; i < 0; ++i) {
997
    memcpy(data_p + i * stride, data_p,
998
           (width + 2 * WIENER_HALFWIN) * sizeof(uint16_t));
999
  }
1000
  for (i = height; i < height + WIENER_HALFWIN; ++i) {
1001
    memcpy(data_p + i * stride, data_p + (height - 1) * stride,
1002
           (width + 2 * WIENER_HALFWIN) * sizeof(uint16_t));
1003
1004
1005
  }
}

1006
1007
1008
1009
static void loop_copy_tile_highbd(uint16_t *data, int tile_idx, int subtile_idx,
                                  int subtile_bits, int width, int height,
                                  int stride, RestorationInternal *rst,
                                  uint16_t *dst, int dst_stride) {
1010
1011
  const int tile_width = rst->tile_width;
  const int tile_height = rst->tile_height;
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
  int i;
  int h_start, h_end, v_start, v_end;
  av1_get_rest_tile_limits(tile_idx, subtile_idx, subtile_bits, rst->nhtiles,
                           rst->nvtiles, tile_width, tile_height, width, height,
                           0, 0, &h_start, &h_end, &v_start, &v_end);
  for (i = v_start; i < v_end; ++i)
    memcpy(dst + i * dst_stride + h_start, data + i * stride + h_start,
           (h_end - h_start) * sizeof(*dst));
}

1022
1023
1024
static void loop_wiener_filter_tile_highbd(uint16_t *data, int tile_idx,
                                           int width, int height, int stride,
                                           RestorationInternal *rst,
1025
1026
                                           int bit_depth, uint16_t *dst,
                                           int dst_stride) {
1027
1028
  const int tile_width = rst->tile_width;
  const int tile_height = rst->tile_height;
1029
1030
  int h_start, h_end, v_start, v_end;
  int i, j;
1031
1032
  DECLARE_ALIGNED(16, InterpKernel, hkernel);
  DECLARE_ALIGNED(16, InterpKernel, vkernel);
1033

1034
  if (rst->rsi->restoration_type[tile_idx] == RESTORE_NONE) {
1035
1036
1037
1038
    loop_copy_tile_highbd(data, tile_idx, 0, 0, width, height, stride, rst, dst,
                          dst_stride);
    return;
  }
1039
1040
  // TODO(david.barker): Store hfilter/vfilter as an InterpKernel
  // instead of the current format. Then this can be removed.
1041
1042
  assert(WIENER_WIN == SUBPEL_TAPS - 1);
  for (i = 0; i < WIENER_WIN; ++i) {
1043
1044
    hkernel[i] = rst->rsi->wiener_info[tile_idx].hfilter[i];
    vkernel[i] = rst->rsi->wiener_info[tile_idx].vfilter[i];
1045
  }
1046
1047
  hkernel[WIENER_WIN] = 0;
  vkernel[WIENER_WIN] = 0;
1048
1049
  hkernel[3] -= 128;
  vkernel[3] -= 128;
1050
  av1_get_rest_tile_limits(tile_idx, 0, 0, rst->nhtiles, rst->nvtiles,
1051
                           tile_width, tile_height, width, height, 0, 0,
1052
                           &h_start, &h_end, &v_start, &v_end);
1053
1054
1055
1056
1057
1058
1059
1060
  // Convolve the whole tile (done in blocks here to match the requirements
  // of the vectorized convolve functions, but the result is equivalent)
  for (i = v_start; i < v_end; i += MAX_SB_SIZE)
    for (j = h_start; j < h_end; j += MAX_SB_SIZE) {
      int w = AOMMIN(MAX_SB_SIZE, (h_end - j + 15) & ~15);
      int h = AOMMIN(MAX_SB_SIZE, (v_end - i + 15) & ~15);
      const uint16_t *data_p = data + i * stride + j;
      uint16_t *dst_p = dst + i * dst_stride + j;
1061
1062
1063
      aom_highbd_convolve8_add_src(CONVERT_TO_BYTEPTR(data_p), stride,
                                   CONVERT_TO_BYTEPTR(dst_p), dst_stride,
                                   hkernel, 16, vkernel, 16, w, h, bit_depth);
1064
1065
1066
    }
}

1067
1068
static void loop_wiener_filter_highbd(uint8_t *data8, int width, int height,
                                      int stride, RestorationInternal *rst,
1069
1070
                                      int bit_depth, uint8_t *dst8,
                                      int dst_stride) {
1071
  uint16_t *data = CONVERT_TO_SHORTPTR(data8);
1072
  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
1073
  int tile_idx;
1074
  extend_frame_highbd(data, width, height, stride);
1075
  for (tile_idx = 0; tile_idx < rst->ntiles; ++tile_idx) {
1076
    loop_wiener_filter_tile_highbd(data, tile_idx, width, height, stride, rst,
1077
                                   bit_depth, dst, dst_stride);
1078
  }
1079
1080
}

1081
1082
1083
static void apply_selfguided_restoration_highbd(
    uint16_t *dat, int width, int height, int stride, int bit_depth, int eps,
    int *xqd, uint16_t *dst, int dst_stride, int32_t *tmpbuf) {
1084
  int xq[2];
1085
  int32_t *flt1 = tmpbuf;
1086
  int32_t *flt2 = flt1 + RESTORATION_TILEPELS_MAX;
1087
  int32_t *tmpbuf2 = flt2 + RESTORATION_TILEPELS_MAX;
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
  int i, j;
  for (i = 0; i < height; ++i) {
    for (j = 0; j < width; ++j) {
      assert(i * width + j < RESTORATION_TILEPELS_MAX);
      flt1[i * width + j] = dat[i * stride + j];
      flt2[i * width + j] = dat[i * stride + j];
    }
  }
  av1_selfguided_restoration(flt1, width, height, width, bit_depth,
                             sgr_params[eps].r1, sgr_params[eps].e1, tmpbuf2);
  av1_selfguided_restoration(flt2, width, height, width, bit_depth,
                             sgr_params[eps].r2, sgr_params[eps].e2, tmpbuf2);
  decode_xq(xqd, xq);
  for (i = 0; i < height; ++i) {
    for (j = 0; j < width; ++j) {
      const int k = i * width + j;
      const int l = i * stride + j;
      const int m = i * dst_stride + j;
      const int32_t u = ((int32_t)dat[l] << SGRPROJ_RST_BITS);
      const int32_t f1 = (int32_t)flt1[k] - u;
      const int32_t f2 = (int32_t)flt2[k] - u;
      const int64_t v = xq[0] * f1 + xq[1] * f2 + (u << SGRPROJ_PRJ_BITS);
      const int16_t w =
          (int16_t)ROUND_POWER_OF_TWO(v, SGRPROJ_PRJ_BITS + SGRPROJ_RST_BITS);
      dst[m] = (uint16_t)clip_pixel_highbd(w, bit_depth);
    }
  }
}

1117
1118
1119
static void loop_sgrproj_filter_tile_highbd(uint16_t *data, int tile_idx,
                                            int width, int height, int stride,
                                            RestorationInternal *rst,
1120
1121
                                            int bit_depth, uint16_t *dst,
                                            int dst_stride) {
1122
1123
  const int tile_width = rst->tile_width;
  const int tile_height = rst->tile_height;
1124
  int h_start, h_end, v_start, v_end;
1125
  uint16_t *data_p, *dst_p;
1126

1127
  if (rst->rsi->restoration_type[tile_idx] == RESTORE_NONE) {
1128
1129
1130
1131
    loop_copy_tile_highbd(data, tile_idx, 0, 0, width, height, stride, rst, dst,
                          dst_stride);
    return;
  }
1132
1133
1134
1135
  av1_get_rest_tile_limits(tile_idx, 0, 0, rst->nhtiles, rst->nvtiles,
                           tile_width, tile_height, width, height, 0, 0,
                           &h_start, &h_end, &v_start, &v_end);
  data_p = data + h_start + v_start * stride;
1136
  dst_p = dst + h_start + v_start * dst_stride;