restoration.c 75.6 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
#include "av1/common/onyxc_int.h"
19
20
21
#if CONFIG_FRAME_SUPERRES
#include "av1/common/resize.h"
#endif
22
#include "av1/common/restoration.h"
Yaowu Xu's avatar
Yaowu Xu committed
23
24
#include "aom_dsp/aom_dsp_common.h"
#include "aom_mem/aom_mem.h"
25

26
#include "aom_ports/mem.h"
27

28
const sgr_params_type sgr_params[SGRPROJ_PARAMS] = {
29
// r1, eps1, r2, eps2
30
#if MAX_RADIUS == 2
31
32
  { 2, 12, 1, 4 },  { 2, 15, 1, 6 },  { 2, 18, 1, 8 },  { 2, 20, 1, 9 },
  { 2, 22, 1, 10 }, { 2, 25, 1, 11 }, { 2, 35, 1, 12 }, { 2, 45, 1, 13 },
33
  { 2, 55, 1, 14 }, { 2, 65, 1, 15 }, { 2, 75, 1, 16 }, { 2, 30, 1, 6 },
34
35
  { 2, 50, 1, 12 }, { 2, 60, 1, 13 }, { 2, 70, 1, 14 }, { 2, 80, 1, 15 },
#else
36
37
38
39
  { 2, 12, 1, 4 },  { 2, 15, 1, 6 },  { 2, 18, 1, 8 },  { 2, 20, 1, 9 },
  { 2, 22, 1, 10 }, { 2, 25, 1, 11 }, { 2, 35, 1, 12 }, { 2, 45, 1, 13 },
  { 2, 55, 1, 14 }, { 2, 65, 1, 15 }, { 2, 75, 1, 16 }, { 3, 30, 1, 10 },
  { 3, 50, 1, 12 }, { 3, 50, 2, 25 }, { 3, 60, 2, 35 }, { 3, 70, 2, 45 },
40
#endif  // MAX_RADIUS == 2
41
42
};

43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
#if CONFIG_MAX_TILE
static void tile_width_and_height(const AV1_COMMON *cm, int is_uv, int sb_w,
                                  int sb_h, int *px_w, int *px_h) {
  const int scaled_sb_w = sb_w << MAX_MIB_SIZE_LOG2;
  const int scaled_sb_h = sb_h << MAX_MIB_SIZE_LOG2;

  const int ss_x = is_uv && cm->subsampling_x;
  const int ss_y = is_uv && cm->subsampling_y;

  *px_w = (scaled_sb_w + ss_x) >> ss_x;
  *px_h = (scaled_sb_h + ss_y) >> ss_y;
#if CONFIG_FRAME_SUPERRES
  if (!av1_superres_unscaled(cm)) {
    av1_calculate_unscaled_superres_size(px_w, px_h,
                                         cm->superres_scale_denominator);
  }
#endif  // CONFIG_FRAME_SUPERRES
}
#endif  // CONFIG_MAX_TILE

63
64
65
66
67
68
69
70
71
72
73
74
75
// Similar to av1_get_tile_rect(), except that we extend the bottommost tile in
// each frame to a multiple of 8 luma pixels.
// This is done to help simplify the implementation of striped-loop-restoration,
// by avoiding nasty edge cases which would otherwise appear when the (cropped)
// frame height is 57 or 63 (mod 64).
static AV1PixelRect get_ext_tile_rect(const TileInfo *tile_info,
                                      const AV1_COMMON *cm, int is_uv) {
  int ss_y = is_uv && cm->subsampling_y;
  AV1PixelRect tile_rect = av1_get_tile_rect(tile_info, cm, is_uv);
  tile_rect.bottom = ALIGN_POWER_OF_TWO(tile_rect.bottom, 3 - ss_y);
  return tile_rect;
}

76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
// Count horizontal or vertical units per tile (use a width or height for
// tile_size, respectively). We basically want to divide the tile size by the
// size of a restoration unit. Rather than rounding up unconditionally as you
// might expect, we round to nearest, which models the way a right or bottom
// restoration unit can extend to up to 150% its normal width or height. The
// max with 1 is to deal with tiles that are smaller than half of a restoration
// unit.
static int count_units_in_tile(int unit_size, int tile_size) {
  return AOMMAX((tile_size + (unit_size >> 1)) / unit_size, 1);
}

void av1_alloc_restoration_struct(AV1_COMMON *cm, RestorationInfo *rsi,
                                  int is_uv) {
#if CONFIG_MAX_TILE
  // We need to allocate enough space for restoration units to cover the
  // largest tile. Without CONFIG_MAX_TILE, this is always the tile at the
92
  // top-left and we can use get_ext_tile_rect(). With CONFIG_MAX_TILE, we have
93
94
  // to do the computation ourselves, iterating over the tiles and keeping
  // track of the largest width and height, then upscaling.
95
96
  int max_sb_w = 0;
  int max_sb_h = 0;
97
  for (int i = 0; i < cm->tile_cols; ++i) {
98
99
    const int sb_w = cm->tile_col_start_sb[i + 1] - cm->tile_col_start_sb[i];
    max_sb_w = AOMMAX(max_sb_w, sb_w);
100
101
  }
  for (int i = 0; i < cm->tile_rows; ++i) {
102
103
    const int sb_h = cm->tile_row_start_sb[i + 1] - cm->tile_row_start_sb[i];
    max_sb_h = AOMMAX(max_sb_h, sb_h);
104
  }
105
106
107
108

  int max_tile_w, max_tile_h;
  tile_width_and_height(cm, is_uv, max_sb_w, max_sb_h, &max_tile_w,
                        &max_tile_h);
109
110
111
112
#else
  TileInfo tile_info;
  av1_tile_init(&tile_info, cm, 0, 0);

113
  const AV1PixelRect tile_rect = get_ext_tile_rect(&tile_info, cm, is_uv);
114
115
116
117
  assert(tile_rect.left == 0 && tile_rect.top == 0);

  const int max_tile_w = tile_rect.right;
  const int max_tile_h = tile_rect.bottom;
118
#endif  // CONFIG_MAX_TILE
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138

  // To calculate hpertile and vpertile (horizontal and vertical units per
  // tile), we basically want to divide the largest tile width or height by the
  // size of a restoration unit. Rather than rounding up unconditionally as you
  // might expect, we round to nearest, which models the way a right or bottom
  // restoration unit can extend to up to 150% its normal width or height. The
  // max with 1 is to deal with tiles that are smaller than half of a
  // restoration unit.
  const int unit_size = rsi->restoration_unit_size;
  const int hpertile = count_units_in_tile(unit_size, max_tile_w);
  const int vpertile = count_units_in_tile(unit_size, max_tile_h);

  rsi->units_per_tile = hpertile * vpertile;
  rsi->horz_units_per_tile = hpertile;
  rsi->vert_units_per_tile = vpertile;

  const int ntiles = cm->tile_rows * cm->tile_cols;
  const int nunits = ntiles * rsi->units_per_tile;

  aom_free(rsi->unit_info);
139
140
141
  CHECK_MEM_ERROR(cm, rsi->unit_info,
                  (RestorationUnitInfo *)aom_memalign(
                      16, sizeof(*rsi->unit_info) * nunits));
142
143
144
}

void av1_free_restoration_struct(RestorationInfo *rst_info) {
145
146
  aom_free(rst_info->unit_info);
  rst_info->unit_info = NULL;
147
}
148
149
150

// TODO(debargha): This table can be substantially reduced since only a few
// values are actually used.
David Barker's avatar
David Barker committed
151
int sgrproj_mtable[MAX_EPS][MAX_NELEM];
152
153
154
155
156
157
158
159
160
161

static void GenSgrprojVtable() {
  int e, n;
  for (e = 1; e <= MAX_EPS; ++e)
    for (n = 1; n <= MAX_NELEM; ++n) {
      const int n2e = n * n * e;
      sgrproj_mtable[e - 1][n - 1] =
          (((1 << SGRPROJ_MTABLE_BITS) + n2e / 2) / n2e);
    }
}
162
163

void av1_loop_restoration_precal() { GenSgrprojVtable(); }
164

165
166
static void extend_frame_lowbd(uint8_t *data, int width, int height, int stride,
                               int border_horz, int border_vert) {
167
168
169
170
  uint8_t *data_p;
  int i;
  for (i = 0; i < height; ++i) {
    data_p = data + i * stride;
171
172
    memset(data_p - border_horz, data_p[0], border_horz);
    memset(data_p + width, data_p[width - 1], border_horz);
173
  }
174
175
176
  data_p = data - border_horz;
  for (i = -border_vert; i < 0; ++i) {
    memcpy(data_p + i * stride, data_p, width + 2 * border_horz);
177
  }
178
  for (i = height; i < height + border_vert; ++i) {
179
    memcpy(data_p + i * stride, data_p + (height - 1) * stride,
180
           width + 2 * border_horz);
181
182
183
  }
}

184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
#if CONFIG_HIGHBITDEPTH
static void extend_frame_highbd(uint16_t *data, int width, int height,
                                int stride, int border_horz, int border_vert) {
  uint16_t *data_p;
  int i, j;
  for (i = 0; i < height; ++i) {
    data_p = data + i * stride;
    for (j = -border_horz; j < 0; ++j) data_p[j] = data_p[0];
    for (j = width; j < width + border_horz; ++j) data_p[j] = data_p[width - 1];
  }
  data_p = data - border_horz;
  for (i = -border_vert; i < 0; ++i) {
    memcpy(data_p + i * stride, data_p,
           (width + 2 * border_horz) * sizeof(uint16_t));
  }
  for (i = height; i < height + border_vert; ++i) {
    memcpy(data_p + i * stride, data_p + (height - 1) * stride,
           (width + 2 * border_horz) * sizeof(uint16_t));
  }
}
#endif

void extend_frame(uint8_t *data, int width, int height, int stride,
                  int border_horz, int border_vert, int highbd) {
#if !CONFIG_HIGHBITDEPTH
  assert(highbd == 0);
  (void)highbd;
#else
  if (highbd)
    extend_frame_highbd(CONVERT_TO_SHORTPTR(data), width, height, stride,
                        border_horz, border_vert);
  else
#endif
  extend_frame_lowbd(data, width, height, stride, border_horz, border_vert);
}

220
221
222
223
static void copy_tile_lowbd(int width, int height, const uint8_t *src,
                            int src_stride, uint8_t *dst, int dst_stride) {
  for (int i = 0; i < height; ++i)
    memcpy(dst + i * dst_stride, src + i * src_stride, width);
224
225
226
}

#if CONFIG_HIGHBITDEPTH
227
228
229
230
static void copy_tile_highbd(int width, int height, const uint16_t *src,
                             int src_stride, uint16_t *dst, int dst_stride) {
  for (int i = 0; i < height; ++i)
    memcpy(dst + i * dst_stride, src + i * src_stride, width * sizeof(*dst));
231
232
233
}
#endif

234
235
static void copy_tile(int width, int height, const uint8_t *src, int src_stride,
                      uint8_t *dst, int dst_stride, int highbd) {
236
237
238
239
240
#if !CONFIG_HIGHBITDEPTH
  assert(highbd == 0);
  (void)highbd;
#else
  if (highbd)
241
    copy_tile_highbd(width, height, CONVERT_TO_SHORTPTR(src), src_stride,
242
243
244
                     CONVERT_TO_SHORTPTR(dst), dst_stride);
  else
#endif
245
  copy_tile_lowbd(width, height, src, src_stride, dst, dst_stride);
246
}
247

248
249
250
#if CONFIG_STRIPED_LOOP_RESTORATION
#define REAL_PTR(hbd, d) ((hbd) ? (uint8_t *)CONVERT_TO_SHORTPTR(d) : (d))

251
252
253
254
255
256
257
258
259
260
261
262
// Helper function: Save one column of left/right context to the appropriate
// column buffers, then extend the edge of the current tile into that column.
//
// Note: The code to deal with above/below boundaries may have filled out
// the corners of the border with data from the tiles to our left or right,
// which isn't allowed. To fix that up, we need to include the top and
// bottom context regions in the area which we extend.
// But note that we don't need to store the pixels we overwrite in the
// corners of the context area - those have already been overwritten once,
// so their original values are already in rlbs->tmp_save_{above,below}.
#if CONFIG_LOOPFILTERING_ACROSS_TILES
static void setup_boundary_column(const uint8_t *src8, int src_stride,
263
264
                                  uint8_t *dst8, int dst_stride, uint16_t *buf,
                                  int h, int use_highbd) {
265
266
267
268
269
  if (use_highbd) {
    const uint16_t *src16 = CONVERT_TO_SHORTPTR(src8);
    uint16_t *dst16 = CONVERT_TO_SHORTPTR(dst8);
    for (int i = -RESTORATION_BORDER; i < 0; i++)
      dst16[i * dst_stride] = src16[i * src_stride];
270
271
    for (int i = 0; i < h; i++) {
      buf[i] = dst16[i * dst_stride];
272
273
      dst16[i * dst_stride] = src16[i * src_stride];
    }
274
    for (int i = h; i < h + RESTORATION_BORDER; i++)
275
276
277
278
      dst16[i * dst_stride] = src16[i * src_stride];
  } else {
    for (int i = -RESTORATION_BORDER; i < 0; i++)
      dst8[i * dst_stride] = src8[i * src_stride];
279
280
    for (int i = 0; i < h; i++) {
      buf[i] = dst8[i * dst_stride];
281
282
      dst8[i * dst_stride] = src8[i * src_stride];
    }
283
    for (int i = h; i < h + RESTORATION_BORDER; i++)
284
285
286
      dst8[i * dst_stride] = src8[i * src_stride];
  }
}
287
288
289
290
291
292
293
294
295
296
297

static void restore_boundary_column(uint8_t *dst8, int dst_stride,
                                    const uint16_t *buf, int h,
                                    int use_highbd) {
  if (use_highbd) {
    uint16_t *dst16 = CONVERT_TO_SHORTPTR(dst8);
    for (int i = 0; i < h; i++) dst16[i * dst_stride] = buf[i];
  } else {
    for (int i = 0; i < h; i++) dst8[i * dst_stride] = buf[i];
  }
}
298
299
#endif  // CONFIG_LOOPFILTERING_ACROSS_TILES

300
// With striped loop restoration, the filtering for each 64-pixel stripe gets
301
302
303
304
// most of its input from the output of CDEF (stored in data8), but we need to
// fill out a border of 3 pixels above/below the stripe according to the
// following
// rules:
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
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
// * At a frame boundary, we copy the outermost row of CDEF pixels three times.
//   This extension is done by a call to extend_frame() at the start of the loop
//   restoration process, so the value of copy_above/copy_below doesn't strictly
//   matter.
//   However, by setting *copy_above = *copy_below = 1 whenever loop filtering
//   across tiles is disabled, we can allow
//   {setup,restore}_processing_stripe_boundary to assume that the top/bottom
//   data has always been copied, simplifying the behaviour at the left and
//   right edges of tiles.
//
// * If we're at a tile boundary and loop filtering across tiles is enabled,
//   then there is a logical stripe which is 64 pixels high, but which is split
//   into an 8px high and a 56px high stripe so that the processing (and
//   coefficient set usage) can be aligned to tiles.
//   In this case, we use the 3 rows of CDEF output across the boundary for
//   context; this corresponds to leaving the frame buffer as-is.
//
// * If we're at a tile boundary and loop filtering across tiles is disabled,
//   then we take the outermost row of CDEF pixels *within the current tile*
//   and copy it three times. Thus we behave exactly as if the tile were a full
//   frame.
//
// * Otherwise, we're at a stripe boundary within a tile. In that case, we
//   take 2 rows of deblocked pixels and extend them to 3 rows of context.
//
// The distinction between the latter two cases is handled by the
// av1_loop_restoration_save_boundary_lines() function, so here we just need
// to decide if we're overwriting the above/below boundary pixels or not.
static void get_stripe_boundary_info(const RestorationTileLimits *limits,
                                     const AV1PixelRect *tile_rect, int ss_y,
#if CONFIG_LOOPFILTERING_ACROSS_TILES
                                     int loop_filter_across_tiles_enabled,
#endif  // CONFIG_LOOPFILTERING_ACROSS_TILES
                                     int *copy_above, int *copy_below) {
  *copy_above = 1;
  *copy_below = 1;

#if CONFIG_LOOPFILTERING_ACROSS_TILES
  if (loop_filter_across_tiles_enabled) {
#endif
    const int full_stripe_height = RESTORATION_PROC_UNIT_SIZE >> ss_y;
    const int rtile_offset = RESTORATION_TILE_OFFSET >> ss_y;

    const int first_stripe_in_tile = (limits->v_start == tile_rect->top);
    const int this_stripe_height =
        full_stripe_height - (first_stripe_in_tile ? rtile_offset : 0);
    const int last_stripe_in_tile =
        (limits->v_start + this_stripe_height >= tile_rect->bottom);

    if (first_stripe_in_tile) *copy_above = 0;
    if (last_stripe_in_tile) *copy_below = 0;
#if CONFIG_LOOPFILTERING_ACROSS_TILES
  }
#endif
}

// Overwrite the border pixels around a processing stripe so that the conditions
// listed above get_stripe_boundary_info() are preserved.
// We save the pixels which get overwritten into a temporary buffer, so that
// they can be restored by restore_processing_stripe_boundary() after we've
// processed the stripe.
367
368
//
// limits gives the rectangular limits of the remaining stripes for the current
369
370
// restoration unit. rsb is the stored stripe boundaries (taken from either
// deblock or CDEF output as necessary).
371
372
373
374
//
// tile_rect is the limits of the current tile and tile_stripe0 is the index of
// the first stripe in this tile (needed to convert the tile-relative stripe
// index we get from limits into something we can look up in rsb).
375
static void setup_processing_stripe_boundary(
376
    const RestorationTileLimits *limits, const RestorationStripeBoundaries *rsb,
377
    int rsb_row, int use_highbd, int h,
378
#if CONFIG_LOOPFILTERING_ACROSS_TILES
379
    const AV1PixelRect *tile_rect, int loop_filter_across_tiles_enabled,
380
#endif  // CONFIG_LOOPFILTERING_ACROSS_TILES
381
382
    uint8_t *data8, int data_stride, RestorationLineBuffers *rlbs,
    int copy_above, int copy_below) {
383
  assert(CONFIG_HIGHBITDEPTH || !use_highbd);
384

385
386
387
  // Offsets within the line buffers. The buffer logically starts at column
  // -RESTORATION_EXTRA_HORZ so the 1st column (at x0 - RESTORATION_EXTRA_HORZ)
  // has column x0 in the buffer.
388
  const int buf_stride = rsb->stripe_boundary_stride;
389
390
391
392
  const int buf_x0_off = limits->h_start;
  const int line_width =
      (limits->h_end - limits->h_start) + 2 * RESTORATION_EXTRA_HORZ;
  const int line_size = line_width << use_highbd;
393

394
  const int data_x0 = limits->h_start - RESTORATION_EXTRA_HORZ;
395

396
397
398
399
400
  // Replace RESTORATION_BORDER pixels above the top of the stripe
  // We expand RESTORATION_CTX_VERT=2 lines from rsb->stripe_boundary_above
  // to fill RESTORATION_BORDER=3 lines of above pixels. This is done by
  // duplicating the topmost of the 2 lines (see the AOMMAX call when
  // calculating src_row, which gets the values 0, 0, 1 for i = -3, -2, -1).
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
  //
  // Special case: If we're at the top of a tile, which isn't on the topmost
  // tile row, and we're allowed to loop filter across tiles, then we have a
  // logical 64-pixel-high stripe which has been split into an 8-pixel high
  // stripe and a 56-pixel high stripe (the current one). So, in this case,
  // we want to leave the boundary alone!
  if (copy_above) {
    uint8_t *data8_tl = data8 + data_x0 + limits->v_start * data_stride;

    for (int i = -RESTORATION_BORDER; i < 0; ++i) {
      const int buf_row = rsb_row + AOMMAX(i + RESTORATION_CTX_VERT, 0);
      const int buf_off = buf_x0_off + buf_row * buf_stride;
      const uint8_t *buf = rsb->stripe_boundary_above + (buf_off << use_highbd);
      uint8_t *dst8 = data8_tl + i * data_stride;
      // Save old pixels, then replace with data from stripe_boundary_above
      memcpy(rlbs->tmp_save_above[i + RESTORATION_BORDER],
             REAL_PTR(use_highbd, dst8), line_size);
      memcpy(REAL_PTR(use_highbd, dst8), buf, line_size);
    }
420
  }
421

422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
  // Replace RESTORATION_BORDER pixels below the bottom of the stripe.
  // The second buffer row is repeated, so src_row gets the values 0, 1, 1
  // for i = 0, 1, 2.
  if (copy_below) {
    const int stripe_end = limits->v_start + h;
    uint8_t *data8_bl = data8 + data_x0 + stripe_end * data_stride;

    for (int i = 0; i < RESTORATION_BORDER; ++i) {
      const int buf_row = rsb_row + AOMMIN(i, RESTORATION_CTX_VERT - 1);
      const int buf_off = buf_x0_off + buf_row * buf_stride;
      const uint8_t *src = rsb->stripe_boundary_below + (buf_off << use_highbd);

      uint8_t *dst8 = data8_bl + i * data_stride;
      // Save old pixels, then replace with data from stripe_boundary_below
      memcpy(rlbs->tmp_save_below[i], REAL_PTR(use_highbd, dst8), line_size);
      memcpy(REAL_PTR(use_highbd, dst8), src, line_size);
    }
439
  }
440

441
442
443
444
445
446
#if CONFIG_LOOPFILTERING_ACROSS_TILES
  if (!loop_filter_across_tiles_enabled) {
    // If loopfiltering across tiles is disabled, we need to check if we're at
    // the edge of the current tile column. If we are, we need to extend the
    // leftmost/rightmost column within the tile by 3 pixels, so that the output
    // doesn't depend on pixels from the next column over.
447
448
    // This applies to the top and bottom borders too, since those may have
    // been filled out with data from the tile to the top-left (etc.) of us.
449
450
    const int at_tile_left_border = (limits->h_start == tile_rect->left);
    const int at_tile_right_border = (limits->h_end == tile_rect->right);
451

452
453
454
455
    if (at_tile_left_border) {
      uint8_t *dst8 = data8 + limits->h_start + limits->v_start * data_stride;
      for (int j = -RESTORATION_BORDER; j < 0; j++)
        setup_boundary_column(dst8, data_stride, dst8 + j, data_stride,
456
457
                              rlbs->tmp_save_left[j + RESTORATION_BORDER], h,
                              use_highbd);
458
459
460
461
462
463
    }

    if (at_tile_right_border) {
      uint8_t *dst8 = data8 + limits->h_end + limits->v_start * data_stride;
      for (int j = 0; j < RESTORATION_BORDER; j++)
        setup_boundary_column(dst8 - 1, data_stride, dst8 + j, data_stride,
464
                              rlbs->tmp_save_right[j], h, use_highbd);
465
466
467
    }
  }
#endif  // CONFIG_LOOPFILTERING_ACROSS_TILES
468
469
470
471
}

// This function restores the boundary lines modified by
// setup_processing_stripe_boundary.
472
static void restore_processing_stripe_boundary(
473
    const RestorationTileLimits *limits, const RestorationLineBuffers *rlbs,
474
    int use_highbd, int h,
475
476
477
#if CONFIG_LOOPFILTERING_ACROSS_TILES
    const AV1PixelRect *tile_rect, int loop_filter_across_tiles_enabled,
#endif  // CONFIG_LOOPFILTERING_ACROSS_TILES
478
    uint8_t *data8, int data_stride, int copy_above, int copy_below) {
479
  assert(CONFIG_HIGHBITDEPTH || !use_highbd);
480
481
482
483

  const int line_width =
      (limits->h_end - limits->h_start) + 2 * RESTORATION_EXTRA_HORZ;
  const int line_size = line_width << use_highbd;
484

485
486
  const int data_x0 = limits->h_start - RESTORATION_EXTRA_HORZ;

487
488
489
490
491
492
493
  if (copy_above) {
    uint8_t *data8_tl = data8 + data_x0 + limits->v_start * data_stride;
    for (int i = -RESTORATION_BORDER; i < 0; ++i) {
      uint8_t *dst8 = data8_tl + i * data_stride;
      memcpy(REAL_PTR(use_highbd, dst8),
             rlbs->tmp_save_above[i + RESTORATION_BORDER], line_size);
    }
494
  }
495

496
497
498
  if (copy_below) {
    const int stripe_bottom = limits->v_start + h;
    uint8_t *data8_bl = data8 + data_x0 + stripe_bottom * data_stride;
499

500
501
    for (int i = 0; i < RESTORATION_BORDER; ++i) {
      if (stripe_bottom + i >= limits->v_end + RESTORATION_BORDER) break;
502

503
504
505
      uint8_t *dst8 = data8_bl + i * data_stride;
      memcpy(REAL_PTR(use_highbd, dst8), rlbs->tmp_save_below[i], line_size);
    }
506
  }
507
508
509
510
511
512

#if CONFIG_LOOPFILTERING_ACROSS_TILES
  if (!loop_filter_across_tiles_enabled) {
    // Restore any pixels we overwrote at the left/right edge of this
    // processing unit
    // Note: We don't need to restore the corner pixels, even if we overwrote
513
514
515
    // them in the equivalent place in setup_processing_stripe_boundary:
    // Because !loop_filter_across_tiles_enabled => copy_above = copy_below = 1,
    // the corner pixels will already have been restored before we get here.
516
517
518
519
520
521
    const int at_tile_left_border = (limits->h_start == tile_rect->left);
    const int at_tile_right_border = (limits->h_end == tile_rect->right);

    if (at_tile_left_border) {
      uint8_t *dst8 = data8 + limits->h_start + limits->v_start * data_stride;
      for (int j = -RESTORATION_BORDER; j < 0; j++)
522
523
524
        restore_boundary_column(dst8 + j, data_stride,
                                rlbs->tmp_save_left[j + RESTORATION_BORDER], h,
                                use_highbd);
525
526
527
528
529
    }

    if (at_tile_right_border) {
      uint8_t *dst8 = data8 + limits->h_end + limits->v_start * data_stride;
      for (int j = 0; j < RESTORATION_BORDER; j++)
530
531
        restore_boundary_column(dst8 + j, data_stride, rlbs->tmp_save_right[j],
                                h, use_highbd);
532
533
534
    }
  }
#endif  // CONFIG_LOOPFILTERING_ACROSS_TILES
535
536
537
}
#endif

538
539
#if USE_WIENER_HIGH_INTERMEDIATE_PRECISION
#define wiener_convolve8_add_src aom_convolve8_add_src_hip
540
#else
541
#define wiener_convolve8_add_src aom_convolve8_add_src
542
543
#endif

544
545
546
547
548
549
550
551
552
553
554
static void wiener_filter_stripe(const RestorationUnitInfo *rui,
                                 int stripe_width, int stripe_height,
                                 int procunit_width, const uint8_t *src,
                                 int src_stride, uint8_t *dst, int dst_stride,
                                 int32_t *tmpbuf, int bit_depth) {
  (void)tmpbuf;
  (void)bit_depth;
  assert(bit_depth == 8);

  for (int j = 0; j < stripe_width; j += procunit_width) {
    int w = AOMMIN(procunit_width, (stripe_width - j + 15) & ~15);
555
556
557
    const uint8_t *src_p = src + j;
    uint8_t *dst_p = dst + j;
    wiener_convolve8_add_src(src_p, src_stride, dst_p, dst_stride,
558
                             rui->wiener_info.hfilter, 16,
559
                             rui->wiener_info.vfilter, 16, w, stripe_height);
560
  }
561
}
562

563
564
/* 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
565
   specialize to r = 1, 2, 3. A default function is used for r > 3.
566
567
568
569
570
571
572
573
574
575
576
577
578
579

   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];
580

581
582
583
584
585
586
587
588
589
590
591
592
593
594
      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;
    }
595
  } else {
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
    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;
  }
}

722
#if MAX_RADIUS > 2
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
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
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
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);
}
867
#endif  // MAX_RADIUS > 2
868

869
870
871
872
873
874
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);
875
#if MAX_RADIUS > 2
876
877
  else if (r == 3)
    boxsum3(src, width, height, src_stride, sqr, dst, dst_stride);
878
  else if (r > 3)
879
    boxsumr(src, width, height, src_stride, r, sqr, dst, dst_stride);
880
881
882
#endif  // MAX_RADIUS > 2
  else
    assert(0 && "Invalid value of r in self-guided filter");
883
884
}

885
#if MAX_RADIUS > 2
886
887
static void boxnum(int width, int height, int r, int8_t *num, int num_stride) {
  int i, j;
888
889
890
  for (i = 0; i <= r; ++i) {
    for (j = 0; j <= r; ++j) {
      num[i * num_stride + j] = (r + 1 + i) * (r + 1 + j);
891
892
893
894
895
896
      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];
    }
  }
897
898
  for (j = 0; j <= r; ++j) {
    const int val = (2 * r + 1) * (r + 1 + j);
899
900
901
902
903
    for (i = r + 1; i < height - r; ++i) {
      num[i * num_stride + j] = val;
      num[i * num_stride + (width - 1 - j)] = val;
    }
  }
904
905
  for (i = 0; i <= r; ++i) {
    const int val = (2 * r + 1) * (r + 1 + i);
906
907
908
909
910
911
912
    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) {
913
      num[i * num_stride + j] = (2 * r + 1) * (2 * r + 1);
914
915
916
    }
  }
}
917
#endif  // MAX_RADIUS > 2
918

919
void decode_xq(const int *xqd, int *xq) {
920
  xq[0] = xqd[0];
921
922
923
  xq[1] = (1 << SGRPROJ_PRJ_BITS) - xq[0] - xqd[1];
}

David Barker's avatar
David Barker committed
924
const int32_t x_by_xplus1[256] = {
925
926
927
  // Special case: Map 0 -> 1 (corresponding to a value of 1/256)
  // instead of 0. See comments in av1_selfguided_restoration_internal() for why
  1,   128, 171, 192, 205, 213, 219, 224, 228, 230, 233, 235, 236, 238, 239,
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
  240, 241, 242, 243, 243, 244, 244, 245, 245, 246, 246, 247, 247, 247, 247,
  248, 248, 248, 248, 249, 249, 249, 249, 249, 250, 250, 250, 250, 250, 250,
  250, 251, 251, 251, 251, 251, 251, 251, 251, 251, 251, 252, 252, 252, 252,
  252, 252, 252, 252, 252, 252, 252, 252, 252, 252, 252, 252, 252, 253, 253,
  253, 253, 253, 253, 253, 253, 253, 253, 253, 253, 253, 253, 253, 253, 253,
  253, 253, 253, 253, 253, 253, 253, 253, 253, 253, 253, 253, 254, 254, 254,
  254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254,
  254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254,
  254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254,
  254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254,
  254, 254, 254, 254, 254, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
  255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
  255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
  255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
  255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
  255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
  256,
};

David Barker's avatar
David Barker committed
947
const int32_t one_by_x[MAX_NELEM] = {
948
  4096, 2048, 1365, 1024, 819, 683, 585, 512, 455, 410, 372, 341, 315,
949
950
951
952
953
  293,  273,  256,  241,  228, 216, 205, 195, 186, 178, 171, 164,
#if MAX_RADIUS > 2
  158,  152,  146,  141,  137, 132, 128, 124, 120, 117, 114, 111, 108,
  105,  102,  100,  98,   95,  93,  91,  89,  87,  85,  84
#endif  // MAX_RADIUS > 2
954
955
};

956
static void av1_selfguided_restoration_internal(int32_t *dgd, int width,
957
958
                                                int height, int dgd_stride,
                                                int32_t *dst, int dst_stride,
959
                                                int bit_depth, int r, int eps) {
960
961
  const int width_ext = width + 2 * SGRPROJ_BORDER_HORZ;
  const int height_ext = height + 2 * SGRPROJ_BORDER_VERT;
David Barker's avatar
David Barker committed
962
963
964
965
  // Adjusting the stride of A and B here appears to avoid bad cache effects,
  // leading to a significant speed improvement.
  // We also align the stride to a multiple of 16 bytes, for consistency
  // with the SIMD version of this function.
966
  int buf_stride = ((width_ext + 3) & ~3) + 16;
967
968
969
970
  int32_t A_[RESTORATION_PROC_UNIT_PELS];
  int32_t B_[RESTORATION_PROC_UNIT_PELS];
  int32_t *A = A_;
  int32_t *B = B_;
971
972
#if MAX_RADIUS > 2
  const int num_stride = width_ext;
973
  int8_t num_[RESTORATION_PROC_UNIT_PELS];
974
  int8_t *num = num_ + SGRPROJ_BORDER_VERT * num_stride + SGRPROJ_BORDER_HORZ;
975
#endif
976
  int i, j;
977

978
979
980
  assert(r <= MAX_RADIUS && "Need MAX_RADIUS >= r");
  assert(r <= SGRPROJ_BORDER_VERT - 1 && r <= SGRPROJ_BORDER_HORZ - 1 &&
         "Need SGRPROJ_BORDER_* >= r+1");
981

982
983
984
985
  boxsum(dgd - dgd_stride * SGRPROJ_BORDER_VERT - SGRPROJ_BORDER_HORZ,
         width_ext, height_ext, dgd_stride, r, 0, B, buf_stride);
  boxsum(dgd - dgd_stride * SGRPROJ_BORDER_VERT - SGRPROJ_BORDER_HORZ,
         width_ext, height_ext, dgd_stride, r, 1, A, buf_stride);
986
#if MAX_RADIUS > 2
987
  boxnum(width_ext, height_ext, r, num_, num_stride);
988
#endif
989
990
  A += SGRPROJ_BORDER_VERT * buf_stride + SGRPROJ_BORDER_HORZ;
  B += SGRPROJ_BORDER_VERT * buf_stride + SGRPROJ_BORDER_HORZ;
991
992
993
994
  // Calculate the eventual A[] and B[] arrays. Include a 1-pixel border - ie,
  // for a 64x64 processing unit, we calculate 66x66 pixels of A[] and B[].
  for (i = -1; i < height + 1; ++i) {
    for (j = -1; j < width + 1; ++j) {
David Barker's avatar
David Barker committed
995
      const int k = i * buf_stride + j;
996
#if MAX_RADIUS > 2
997
      const int n = num[i * num_stride + j];
998
999
1000
#else
      const int n = (2 * r + 1) * (2 * r + 1);
#endif
1001

1002
1003
1004
1005
1006
1007
1008
      // a < 2^16 * n < 2^22 regardless of bit depth
      uint32_t a = ROUND_POWER_OF_TWO(A[k], 2 * (bit_depth - 8));
      // b < 2^8 * n < 2^14 regardless of bit depth
      uint32_t b = ROUND_POWER_OF_TWO(B[k], bit_depth - 8);

      // Each term in calculating p = a * n - b * b is < 2^16 * n^2 < 2^28,
      // and p itself satisfies p < 2^14 * n^2 < 2^26.
1009
1010
1011
      // This bound on p is due to:
      // https://en.wikipedia.org/wiki/Popoviciu's_inequality_on_variances
      //
1012
1013
1014
1015
      // Note: Sometimes, in high bit depth, we can end up with a*n < b*b.
      // This is an artefact of rounding, and can only happen if all pixels
      // are (almost) identical, so in this case we saturate to p=0.
      uint32_t p = (a * n < b * b) ? 0 : a * n - b * b;
1016
1017
1018
1019
1020

      // Note: If MAX_RADIUS <= 2, then this 's' is a function only of
      // r and eps. Further, this is the only place we use 'eps', so we could
      // pre-calculate 's' for each parameter set and store that in place of
      // 'eps'.
1021
1022
1023
1024
1025
1026
1027
      uint32_t s = sgrproj_mtable[eps - 1][n - 1];

      // p * s < (2^14 * n^2) * round(2^20 / n^2 eps) < 2^34 / eps < 2^32
      // as long as eps >= 4. So p * s fits into a uint32_t, and z < 2^12
      // (this holds even after accounting for the rounding in s)
      const uint32_t z = ROUND_POWER_OF_TWO(p * s, SGRPROJ_MTABLE_BITS);

1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
      // Note: We have to be quite careful about the value of A[k].
      // This is used as a blend factor between individual pixel values and the
      // local mean. So it logically has a range of [0, 256], including both
      // endpoints.
      //
      // This is a pain for hardware, as we'd like something which can be stored
      // in exactly 8 bits.
      // Further, in the calculation of B[k] below, if z == 0 and r == 2,
      // then A[k] "should be" 0. But then we can end up setting B[k] to a value
      // slightly above 2^(8 + bit depth), due to rounding in the value of
      // one_by_x[25-1].
      //
      // Thus we saturate so that, when z == 0, A[k] is set to 1 instead of 0.
      // This fixes the above issues (256 - A[k] fits in a uint8, and we can't
      // overflow), without significantly affecting the final result: z == 0
      // implies that the image is essentially "flat", so the local mean and
      // individual pixel values are very similar.
      //
      // Note that saturating on the other side, ie. requring A[k] <= 255,
      // would be a bad idea, as that corresponds to the case where the image
      // is very variable, when we want to preserve the local pixel value as
      // much as possible.
      A[k] = x_by_xplus1[AOMMIN(z, 255)];  // in range [1, 256]

      // SGRPROJ_SGR - A[k] < 2^8 (from above), B[k] < 2^(bit_depth) * n,
1053
1054
1055
      // one_by_x[n - 1] = round(2^12 / n)
      // => the product here is < 2^(20 + bit_depth) <= 2^32,
      // and B[k] is set to a value < 2^(8 + bit depth)
1056
1057
      // This holds even with the rounding in one_by_x and in the overall
      // result, as long as SGRPROJ_SGR - A[k] is strictly less than 2^8.
1058
1059
1060
1061
      B[k] = (int32_t)ROUND_POWER_OF_TWO((uint32_t)(SGRPROJ_SGR - A[k]) *
                                             (uint32_t)B[k] *
                                             (uint32_t)one_by_x[n - 1],
                                         SGRPROJ_RECIP_BITS);
1062
1063
    }
  }
1064
1065
1066
  // Use the A[] and B[] arrays to calculate the filtered image
  for (i = 0; i < height; ++i) {
    for (j = 0; j < width; ++j) {
David Barker's avatar
David Barker committed
1067
      const int k = i * buf_stride + j;
1068
1069
      const int l = i * dgd_stride + j;
      const int m = i * dst_stride + j;
1070
      const int nb = 5;
1071
      const int32_t a =
David Barker's avatar
David Barker committed
1072
1073
1074
1075
          (A[k] + A[k - 1] + A[k + 1] + A[k - buf_stride] + A[k + buf_stride]) *
              4 +
          (A[k - 1 - buf_stride] + A[k - 1 + buf_stride] +
           A[k + 1 - buf_stride] + A[k + 1 + buf_stride]) *
1076
              3;
1077
      const int32_t b =
David Barker's avatar
David Barker committed
1078
1079
1080
1081
          (B[k] + B[k - 1] + B[k + 1] + B[k - buf_stride] + B[k + buf_stride]) *
              4 +
          (B[k - 1 - buf_stride] + B[k - 1 + buf_stride] +
           B[k + 1 - buf_stride] + B[k + 1 + buf_stride]) *
1082
              3;
1083
      const int32_t v = a * dgd[l] + b;
1084
      dst[m] = ROUND_POWER_OF_TWO(v, SGRPROJ_SGR_BITS + nb - SGRPROJ_RST_BITS);
1085
1086
1087
1088
    }
  }
}

1089
void av1_selfguided_restoration_c(const uint8_t *dgd8, int width, int height,
1090
1091
1092
                                  int dgd_stride, int32_t *flt1, int32_t *flt2,
                                  int flt_stride, const sgr_params_type *params,
                                  int bit_depth, int highbd) {
1093
  int32_t dgd32_[RESTORATION_PROC_UNIT_PELS];
1094
1095
  const int dgd32_stride = width + 2 * SGRPROJ_BORDER_HORZ;
  int32_t *dgd32 =
1096
      dgd32_ + dgd32_stride * SGRPROJ_BORDER_VERT + SGRPROJ_BORDER_HORZ;
1097
1098
1099
1100
1101

  if (highbd) {
    const uint16_t *dgd16 = CONVERT_TO_SHORTPTR(dgd8);
    for (int i = -SGRPROJ_BORDER_VERT; i < height + SGRPROJ_BORDER_VERT; ++i) {
      for (int j = -SGRPROJ_BORDER_HORZ; j < width + SGRPROJ_BORDER_HORZ; ++j) {
1102
        dgd32[i * dgd32_stride + j] = dgd16[i * dgd_stride + j];
1103
1104
1105
1106
1107
      }
    }
  } else {
    for (int i = -SGRPROJ_BORDER_VERT; i < height + SGRPROJ_BORDER_VERT; ++i) {
      for (int j = -SGRPROJ_BORDER_HORZ; j < width + SGRPROJ_BORDER_HORZ; ++j) {
1108
        dgd32[i * dgd32_stride + j] = dgd8[i * dgd_stride + j];
1109
      }
1110
1111
    }
  }
1112

1113
1114
1115
1116
1117
1118
  av1_selfguided_restoration_internal(dgd32, width, height, dgd32_stride, flt1,
                                      flt_stride, bit_depth, params->r1,
                                      params->e1);
  av1_selfguided_restoration_internal(dgd32, width, height, dgd32_stride, flt2,
                                      flt_stride, bit_depth, params->r2,
                                      params->e2);
1119
1120
}

1121
void apply_selfguided_restoration_c(const uint8_t *dat8, int width, int height,
1122
                                    int stride, int eps, const int *xqd,
1123
1124
1125
                                    uint8_t *dst8, int dst_stride,
                                    int32_t *tmpbuf, int bit_depth,
                                    int highbd) {
1126
  int xq[2];
1127
  int32_t *flt1 = tmpbuf;
1128
  int32_t *flt2 = flt1 + RESTORATION_TILEPELS_MAX;
1129
  assert(width * height <= RESTORATION_TILEPELS_MAX);
1130
1131
  av1_selfguided_restoration_c(dat8, width, height, stride, flt1, flt2, width,
                               &sgr_params[eps], bit_depth, highbd);
1132
  decode_xq(xqd, xq);
1133
1134
  for (int i = 0; i < height; ++i) {
    for (int j = 0; j < width; ++j) {
1135
      const int k = i * width + j;
1136
1137
1138
1139
1140
1141
1142
      uint8_t *dst8ij = dst8 + i * dst_stride + j;
      const uint8_t *dat8ij = dat8 + i * stride + j;

      const uint16_t pre_u = highbd ? *CONVERT_TO_SHORTPTR(dat8ij) : *dat8ij;
      const int32_t u = (int32_t)pre_u << SGRPROJ_RST_BITS;
      const int32_t f1 = flt1[k] - u;
      const int32_t f2 = flt2[k] - u;
David Barker's avatar
David Barker committed
1143
      const int32_t v = xq[0] * f1 + xq[1] * f2 + (u << SGRPROJ_PRJ_BITS);
1144
1145
      const int16_t w =
          (int16_t)ROUND_POWER_OF_TWO(v, SGRPROJ_PRJ_BITS + SGRPROJ_RST_BITS);
1146
1147
1148
1149
1150
1151

      const uint16_t out = clip_pixel_highbd(w, bit_depth);
      if (highbd)
        *CONVERT_TO_SHORTPTR(dst8ij) = out;
      else
        *dst8ij = out;
1152
1153
1154
1155
    }
  }
}

1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
static void sgrproj_filter_stripe(const RestorationUnitInfo *rui,
                                  int stripe_width, int stripe_height,
                                  int procunit_width, const uint8_t *src,
                                  int src_stride, uint8_t *dst, int dst_stride,
                                  int32_t *tmpbuf, int bit_depth) {
  (void)bit_depth;
  assert(bit_depth == 8);

  for (int j = 0; j < stripe_width; j += procunit_width) {
    int w = AOMMIN(procunit_width, stripe_width - j);
    apply_selfguided_restoration(src + j, w, stripe_height, src_stride,
                                 rui->sgrproj_info.ep, rui->sgrproj_info.xqd,
1168
                                 dst + j, dst_stride, tmpbuf, bit_depth, 0);
1169
1170
1171
  }
}

1172
#if CONFIG_HIGHBITDEPTH
1173
#if USE_WIENER_HIGH_INTERMEDIATE_PRECISION
1174
#define wiener_highbd_convolve8_add_src aom_highbd_convolve8_add_src_hip
1175
#else
1176
#define wiener_highbd_convolve8_add_src aom_highbd_convolve8_add_src
1177
#endif
1178

1179
1180
1181
static void wiener_filter_stripe_highbd(const RestorationUnitInfo *rui,
                                        int stripe_width, int stripe_height,
                                        int procunit_width, const uint8_t *src8,
1182
                                        int src_stride, uint8_t *dst8,
1183
1184
1185
1186
1187
1188
                                        int dst_stride, int32_t *tmpbuf,
                                        int bit_depth) {
  (void)tmpbuf;

  for (int j = 0; j < stripe_width; j += procunit_width) {
    int w = AOMMIN(procunit_width, (stripe_width - j + 15) & ~15);
1189
1190
    const uint8_t *src8_p = src8 + j;
    uint8_t *dst8_p = dst8 + j;
1191
1192
    wiener_highbd_convolve8_add_src(
        src8_p, src_stride, dst8_p, dst_stride, rui->wiener_info.hfilter, 16,
1193
        rui->wiener_info.vfilter, 16, w, stripe_height, bit_depth);
1194
  }
1195
1196
}

1197
1198
1199
1200
1201
1202
1203
1204
static void sgrproj_filter_stripe_highbd(const RestorationUnitInfo *rui,
                                         int stripe_width, int stripe_height,
                                         int procunit_width,
                                         const uint8_t *src8, int src_stride,
                                         uint8_t *dst8, int dst_stride,
                                         int32_t *tmpbuf, int bit_depth) {
  for (int j = 0; j < stripe_width; j += procunit_width) {
    int w = AOMMIN(procunit_width, stripe_width - j);
1205
1206
1207
    apply_selfguided_restoration(src8 + j, w, stripe_height, src_stride,
                                 rui->sgrproj_info.ep, rui->sgrproj_info.xqd,
                                 dst8 + j, dst_stride, tmpbuf, bit_depth, 1);
1208
1209
1210
1211
  }
}
#endif  // CONFIG_HIGHBITDEPTH

1212
1213
1214
1215
1216
typedef void (*stripe_filter_fun)(const RestorationUnitInfo *rui,
                                  int stripe_width, int stripe_height,
                                  int procunit_width, const uint8_t *src,
                                  int src_stride, uint8_t *dst, int dst_stride,
                                  int32_t *tmpbuf, int bit_depth);
1217
1218
1219

#if CONFIG_HIGHBITDEPTH
#define NUM_STRIPE_FILTERS 4
1220
#else
1221
#define NUM_STRIPE_FILTERS 2
1222
#endif
1223
1224
1225
1226
1227
1228
1229
1230

static const stripe_filter_fun stripe_filters[NUM_STRIPE_FILTERS] = {
  wiener_filter_stripe, sgrproj_filter_stripe,
#if CONFIG_HIGHBITDEPTH
  wiener_filter_stripe_highbd, sgrproj_filter_stripe_highbd
#endif  // CONFIG_HIGHBITDEPTH
};

1231
// Filter one restoration unit
1232
1233
void av1_loop_restoration_filter_unit(
    const RestorationTileLimits *limits, const RestorationUnitInfo *rui,
1234
#if CONFIG_STRIPED_LOOP_RESTORATION
1235
    const RestorationStripeBoundaries *rsb, RestorationLineBuffers *rlbs,
1236
    const AV1PixelRect *tile_rect, int tile_stripe0,
1237
1238
1239
1240
#if CONFIG_LOOPFILTERING_ACROSS_TILES
    int loop_filter_across_tiles_enabled,
#endif  // CONFIG_LOOPFILTERING_ACROSS_TILES
#endif  // CONFIG_STRIPED_LOOP_RESTORATION
1241
1242
    int ss_x, int ss_y, int highbd, int bit_depth, uint8_t *data8, int stride,
    uint8_t *dst8, int dst_stride, int32_t *tmpbuf) {
1243
1244
1245
1246
1247
1248
1249
1250
1251
  RestorationType unit_rtype = rui->restoration_type;

  int unit_h = limits->v_end - limits->v_start;
  int unit_w = limits->h_end - limits->h_start;
  uint8_t *data8_tl = data8 + limits->v_start * stride + limits->h_start;
  uint8_t *dst8_tl = dst8 + limits->v_start * dst_stride + limits->h_start;

  if (unit_rtype == RESTORE_NONE) {
    copy_tile(unit_w, unit_h, data8_tl, stride, dst8_tl, dst_stride, highbd);