rd.c 61.6 KB
Newer Older
Jingning Han's avatar
Jingning Han committed
1
/*
Yaowu Xu's avatar
Yaowu Xu committed
2
 * Copyright (c) 2016, Alliance for Open Media. All rights reserved
Jingning Han's avatar
Jingning Han committed
3
 *
Yaowu Xu's avatar
Yaowu Xu committed
4
5
6
7
8
9
 * This source code is subject to the terms of the BSD 2 Clause License and
 * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
 * was not distributed with this source code in the LICENSE file, you can
 * obtain it at www.aomedia.org/license/software. If the Alliance for Open
 * Media Patent License 1.0 was not distributed with this source code in the
 * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
Jingning Han's avatar
Jingning Han committed
10
11
12
13
14
15
 */

#include <assert.h>
#include <math.h>
#include <stdio.h>

Yaowu Xu's avatar
Yaowu Xu committed
16
#include "./av1_rtcd.h"
Jingning Han's avatar
Jingning Han committed
17

Yaowu Xu's avatar
Yaowu Xu committed
18
19
#include "aom_dsp/aom_dsp_common.h"
#include "aom_mem/aom_mem.h"
20
21
22
#include "aom_ports/bitops.h"
#include "aom_ports/mem.h"
#include "aom_ports/system_state.h"
Jingning Han's avatar
Jingning Han committed
23

24
25
26
27
28
29
30
31
32
#include "av1/common/common.h"
#include "av1/common/entropy.h"
#include "av1/common/entropymode.h"
#include "av1/common/mvref_common.h"
#include "av1/common/pred_common.h"
#include "av1/common/quant_common.h"
#include "av1/common/reconinter.h"
#include "av1/common/reconintra.h"
#include "av1/common/seg_common.h"
33

34
#include "av1/encoder/av1_quantize.h"
35
36
37
38
#include "av1/encoder/cost.h"
#include "av1/encoder/encodemb.h"
#include "av1/encoder/encodemv.h"
#include "av1/encoder/encoder.h"
39
40
41
#if CONFIG_LV_MAP
#include "av1/encoder/encodetxb.h"
#endif
42
43
44
45
#include "av1/encoder/mcomp.h"
#include "av1/encoder/ratectrl.h"
#include "av1/encoder/rd.h"
#include "av1/encoder/tokenize.h"
Jingning Han's avatar
Jingning Han committed
46

47
#define RD_THRESH_POW 1.25
Jingning Han's avatar
Jingning Han committed
48
49
50
51
52
53
54
55

// Factor to weigh the rate for switchable interp filters.
#define SWITCHABLE_INTERP_RATE_FACTOR 1

// The baseline rd thresholds for breaking out of the rd loop for
// certain modes are assumed to be based on 8x8 blocks.
// This table is used to correct for block size.
// The factors here are << 2 (2 = x0.5, 32 = x8 etc).
56
static const uint8_t rd_thresh_block_size_factor[BLOCK_SIZES_ALL] = {
57
#if CONFIG_CHROMA_2X2 || CONFIG_CHROMA_SUB8X8
Jingning Han's avatar
Jingning Han committed
58
59
  2,  2,  2,
#endif
60
  2,  3,  3,  4, 6,  6,  8, 12, 12, 16, 24, 24, 32,
61
#if CONFIG_EXT_PARTITION
62
  48, 48, 64,
63
#endif  // CONFIG_EXT_PARTITION
64
65
66
67
  4,  4,  8,  8, 16, 16,
#if CONFIG_EXT_PARTITION
  32, 32
#endif  // CONFIG_EXT_PARTITION
Jingning Han's avatar
Jingning Han committed
68
69
};

Hui Su's avatar
Hui Su committed
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
#if CONFIG_EXT_TX
static const int use_intra_ext_tx_for_txsize[EXT_TX_SETS_INTRA][EXT_TX_SIZES] =
    {
#if CONFIG_CHROMA_2X2
      { 1, 1, 1, 1, 1 },  // unused
      { 0, 1, 1, 0, 0 },
      { 0, 0, 0, 1, 0 },
#if CONFIG_MRC_TX
      { 0, 0, 0, 0, 1 },
#endif  // CONFIG_MRC_TX
#else   // CONFIG_CHROMA_2X2
      { 1, 1, 1, 1 },  // unused
      { 1, 1, 0, 0 },
      { 0, 0, 1, 0 },
#if CONFIG_MRC_TX
      { 0, 0, 0, 1 },
#endif  // CONFIG_MRC_TX
#endif  // CONFIG_CHROMA_2X2
    };

static const int use_inter_ext_tx_for_txsize[EXT_TX_SETS_INTER][EXT_TX_SIZES] =
    {
#if CONFIG_CHROMA_2X2
      { 1, 1, 1, 1, 1 },  // unused
      { 0, 1, 1, 0, 0 }, { 0, 0, 0, 1, 0 }, { 0, 0, 0, 0, 1 },
#if CONFIG_MRC_TX
      { 0, 0, 0, 0, 1 },
#endif  // CONFIG_MRC_TX
#else   // CONFIG_CHROMA_2X2
      { 1, 1, 1, 1 },  // unused
      { 1, 1, 0, 0 }, { 0, 0, 1, 0 }, { 0, 0, 0, 1 },
#if CONFIG_MRC_TX
      { 0, 0, 0, 1 },
#endif  // CONFIG_MRC_TX
#endif  // CONFIG_CHROMA_2X2
    };
#endif  // CONFIG_EXT_TX

108
109
void av1_fill_mode_rates(AV1_COMMON *const cm, MACROBLOCK *x,
                         FRAME_CONTEXT *fc) {
Jingning Han's avatar
Jingning Han committed
110
111
  int i, j;

112
113
  if (cm->frame_type == KEY_FRAME) {
    for (i = 0; i < PARTITION_CONTEXTS_PRIMARY; ++i)
Yue Chen's avatar
Yue Chen committed
114
115
      av1_cost_tokens_from_cdf(x->partition_cost[i], fc->partition_cdf[i],
                               NULL);
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
#if CONFIG_UNPOISON_PARTITION_CTX
    for (; i < PARTITION_CONTEXTS_PRIMARY + PARTITION_BLOCK_SIZES; ++i) {
      aom_prob p = fc->partition_prob[i][PARTITION_VERT];
      assert(p > 0);
      x->partition_cost[i][PARTITION_NONE] = INT_MAX;
      x->partition_cost[i][PARTITION_HORZ] = INT_MAX;
      x->partition_cost[i][PARTITION_VERT] = av1_cost_bit(p, 0);
      x->partition_cost[i][PARTITION_SPLIT] = av1_cost_bit(p, 1);
    }
    for (; i < PARTITION_CONTEXTS_PRIMARY + 2 * PARTITION_BLOCK_SIZES; ++i) {
      aom_prob p = fc->partition_prob[i][PARTITION_HORZ];
      assert(p > 0);
      x->partition_cost[i][PARTITION_NONE] = INT_MAX;
      x->partition_cost[i][PARTITION_HORZ] = av1_cost_bit(p, 0);
      x->partition_cost[i][PARTITION_VERT] = INT_MAX;
      x->partition_cost[i][PARTITION_SPLIT] = av1_cost_bit(p, 1);
    }
    x->partition_cost[PARTITION_CONTEXTS][PARTITION_NONE] = INT_MAX;
    x->partition_cost[PARTITION_CONTEXTS][PARTITION_HORZ] = INT_MAX;
    x->partition_cost[PARTITION_CONTEXTS][PARTITION_VERT] = INT_MAX;
    x->partition_cost[PARTITION_CONTEXTS][PARTITION_SPLIT] = 0;
#endif  // CONFIG_UNPOISON_PARTITION_CTX
  }

140
141
142
143
144
#if CONFIG_KF_CTX
  for (i = 0; i < KF_MODE_CONTEXTS; ++i)
    for (j = 0; j < KF_MODE_CONTEXTS; ++j)
      av1_cost_tokens_from_cdf(x->y_mode_costs[i][j], fc->kf_y_cdf[i][j], NULL);
#else
Jingning Han's avatar
Jingning Han committed
145
146
  for (i = 0; i < INTRA_MODES; ++i)
    for (j = 0; j < INTRA_MODES; ++j)
Hui Su's avatar
Hui Su committed
147
      av1_cost_tokens_from_cdf(x->y_mode_costs[i][j], fc->kf_y_cdf[i][j], NULL);
148
#endif
Jingning Han's avatar
Jingning Han committed
149

150
  for (i = 0; i < BLOCK_SIZE_GROUPS; ++i)
Hui Su's avatar
Hui Su committed
151
    av1_cost_tokens_from_cdf(x->mbmode_cost[i], fc->y_mode_cdf[i], NULL);
152
  for (i = 0; i < INTRA_MODES; ++i)
153
    av1_cost_tokens_from_cdf(x->intra_uv_mode_cost[i], fc->uv_mode_cdf[i],
Hui Su's avatar
Hui Su committed
154
                             NULL);
Jingning Han's avatar
Jingning Han committed
155

156
157
158
159
160
161
162
#if CONFIG_FILTER_INTRA
  for (i = 0; i < 2; ++i) {
    av1_cost_tokens_from_cdf(x->filter_intra_mode_cost[i],
                             fc->filter_intra_mode_cdf[i], NULL);
  }
#endif

Jingning Han's avatar
Jingning Han committed
163
  for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
Yue Chen's avatar
Yue Chen committed
164
    av1_cost_tokens_from_cdf(x->switchable_interp_costs[i],
165
                             fc->switchable_interp_cdf[i], NULL);
hui su's avatar
hui su committed
166
167

  for (i = 0; i < PALETTE_BLOCK_SIZES; ++i) {
168
    av1_cost_tokens_from_cdf(x->palette_y_size_cost[i],
169
                             fc->palette_y_size_cdf[i], NULL);
170
    av1_cost_tokens_from_cdf(x->palette_uv_size_cost[i],
171
                             fc->palette_uv_size_cdf[i], NULL);
hui su's avatar
hui su committed
172
173
  }

174
  for (i = 0; i < PALETTE_SIZES; ++i) {
175
    for (j = 0; j < PALETTE_COLOR_INDEX_CONTEXTS; ++j) {
176
      av1_cost_tokens_from_cdf(x->palette_y_color_cost[i][j],
177
                               fc->palette_y_color_index_cdf[i][j], NULL);
178
      av1_cost_tokens_from_cdf(x->palette_uv_color_cost[i][j],
179
                               fc->palette_uv_color_index_cdf[i][j], NULL);
hui su's avatar
hui su committed
180
    }
181
  }
182
183
184
185
186
187
188
189
190
191
#if CONFIG_MRC_TX
  for (i = 0; i < PALETTE_SIZES; ++i) {
    for (j = 0; j < PALETTE_COLOR_INDEX_CONTEXTS; ++j) {
      av1_cost_tokens_from_cdf(x->mrc_mask_inter_cost[i][j],
                               fc->mrc_mask_inter_cdf[i][j], NULL);
      av1_cost_tokens_from_cdf(x->mrc_mask_intra_cost[i][j],
                               fc->mrc_mask_intra_cdf[i][j], NULL);
    }
  }
#endif  // CONFIG_MRC_TX
192

193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
#if CONFIG_CFL
  int sign_cost[CFL_JOINT_SIGNS];
  av1_cost_tokens_from_cdf(sign_cost, fc->cfl_sign_cdf, NULL);
  for (int joint_sign = 0; joint_sign < CFL_JOINT_SIGNS; joint_sign++) {
    const aom_cdf_prob *cdf_u = fc->cfl_alpha_cdf[CFL_CONTEXT_U(joint_sign)];
    const aom_cdf_prob *cdf_v = fc->cfl_alpha_cdf[CFL_CONTEXT_V(joint_sign)];
    int *cost_u = x->cfl_cost[joint_sign][CFL_PRED_U];
    int *cost_v = x->cfl_cost[joint_sign][CFL_PRED_V];
    if (CFL_SIGN_U(joint_sign) == CFL_SIGN_ZERO)
      memset(cost_u, 0, CFL_ALPHABET_SIZE * sizeof(*cost_u));
    else
      av1_cost_tokens_from_cdf(cost_u, cdf_u, NULL);
    if (CFL_SIGN_V(joint_sign) == CFL_SIGN_ZERO)
      memset(cost_v, 0, CFL_ALPHABET_SIZE * sizeof(*cost_v));
    else
      av1_cost_tokens_from_cdf(cost_v, cdf_v, NULL);
    for (int u = 0; u < CFL_ALPHABET_SIZE; u++)
      cost_u[u] += sign_cost[joint_sign];
  }
#endif  // CONFIG_CFL

Jingning Han's avatar
Jingning Han committed
214
  for (i = 0; i < MAX_TX_DEPTH; ++i)
215
    for (j = 0; j < TX_SIZE_CONTEXTS; ++j)
Yue Chen's avatar
Yue Chen committed
216
217
      av1_cost_tokens_from_cdf(x->tx_size_cost[i][j], fc->tx_size_cdf[i][j],
                               NULL);
218

219
#if CONFIG_EXT_TX
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
#if CONFIG_LGT_FROM_PRED
  if (LGT_FROM_PRED_INTRA) {
    for (i = 0; i < LGT_SIZES; ++i) {
      for (j = 0; j < INTRA_MODES; ++j) {
        x->intra_lgt_cost[i][j][0] = av1_cost_bit(fc->intra_lgt_prob[i][j], 0);
        x->intra_lgt_cost[i][j][1] = av1_cost_bit(fc->intra_lgt_prob[i][j], 1);
      }
    }
  }
  if (LGT_FROM_PRED_INTER) {
    for (i = 0; i < LGT_SIZES; ++i) {
      x->inter_lgt_cost[i][0] = av1_cost_bit(fc->inter_lgt_prob[i], 0);
      x->inter_lgt_cost[i][1] = av1_cost_bit(fc->inter_lgt_prob[i], 1);
    }
  }
#endif  // CONFIG_LGT_FROM_PRED
236
237
238
  for (i = TX_4X4; i < EXT_TX_SIZES; ++i) {
    int s;
    for (s = 1; s < EXT_TX_SETS_INTER; ++s) {
239
      if (use_inter_ext_tx_for_txsize[s][i]) {
240
241
242
        av1_cost_tokens_from_cdf(
            x->inter_tx_type_costs[s][i], fc->inter_ext_tx_cdf[s][i],
            av1_ext_tx_inv[av1_ext_tx_set_idx_to_type[1][s]]);
243
244
245
      }
    }
    for (s = 1; s < EXT_TX_SETS_INTRA; ++s) {
246
      if (use_intra_ext_tx_for_txsize[s][i]) {
Hui Su's avatar
Hui Su committed
247
248
249
        for (j = 0; j < INTRA_MODES; ++j) {
          av1_cost_tokens_from_cdf(
              x->intra_tx_type_costs[s][i][j], fc->intra_ext_tx_cdf[s][i][j],
250
              av1_ext_tx_inv[av1_ext_tx_set_idx_to_type[0][s]]);
Hui Su's avatar
Hui Su committed
251
        }
252
253
      }
    }
254
  }
255
#else
256
257
  for (i = TX_4X4; i < EXT_TX_SIZES; ++i) {
    for (j = 0; j < TX_TYPES; ++j)
Yue Chen's avatar
Yue Chen committed
258
259
      av1_cost_tokens_from_cdf(x->intra_tx_type_costs[i][j],
                               fc->intra_ext_tx_cdf[i][j], av1_ext_tx_inv);
260
261
  }
  for (i = TX_4X4; i < EXT_TX_SIZES; ++i) {
Yue Chen's avatar
Yue Chen committed
262
263
    av1_cost_tokens_from_cdf(x->inter_tx_type_costs[i], fc->inter_ext_tx_cdf[i],
                             av1_ext_tx_inv);
264
  }
265
#endif  // CONFIG_EXT_TX
266
#if CONFIG_LOOP_RESTORATION
267
268
269
270
271
272
273
274
275
276
277
278
279
  av1_cost_tokens_from_cdf(x->switchable_restore_cost,
                           fc->switchable_restore_cdf, NULL);
#if CONFIG_NEW_MULTISYMBOL
  av1_cost_tokens_from_cdf(x->wiener_restore_cost, fc->wiener_restore_cdf,
                           NULL);
  av1_cost_tokens_from_cdf(x->sgrproj_restore_cost, fc->sgrproj_restore_cdf,
                           NULL);
#else
  x->wiener_restore_cost[0] = av1_cost_bit(RESTORE_NONE_WIENER_PROB, 0);
  x->wiener_restore_cost[1] = av1_cost_bit(RESTORE_NONE_WIENER_PROB, 1);
  x->sgrproj_restore_cost[0] = av1_cost_bit(RESTORE_NONE_SGRPROJ_PROB, 0);
  x->sgrproj_restore_cost[1] = av1_cost_bit(RESTORE_NONE_SGRPROJ_PROB, 1);
#endif  // CONFIG_NEW_MULTISYMBOL
280
#endif  // CONFIG_LOOP_RESTORATION
Hui Su's avatar
Hui Su committed
281
282
283
#if CONFIG_INTRABC
  av1_cost_tokens_from_cdf(x->intrabc_cost, fc->intrabc_cdf, NULL);
#endif  // CONFIG_INTRABC
284
285
286

  if (!frame_is_intra_only(cm)) {
    for (i = 0; i < NEWMV_MODE_CONTEXTS; ++i) {
Yue Chen's avatar
Yue Chen committed
287
288
289
#if CONFIG_NEW_MULTISYMBOL
      av1_cost_tokens_from_cdf(x->newmv_mode_cost[i], fc->newmv_cdf[i], NULL);
#else
290
291
      x->newmv_mode_cost[i][0] = av1_cost_bit(fc->newmv_prob[i], 0);
      x->newmv_mode_cost[i][1] = av1_cost_bit(fc->newmv_prob[i], 1);
Yue Chen's avatar
Yue Chen committed
292
#endif
293
294
295
    }

    for (i = 0; i < ZEROMV_MODE_CONTEXTS; ++i) {
Yue Chen's avatar
Yue Chen committed
296
297
298
#if CONFIG_NEW_MULTISYMBOL
      av1_cost_tokens_from_cdf(x->zeromv_mode_cost[i], fc->zeromv_cdf[i], NULL);
#else
299
300
      x->zeromv_mode_cost[i][0] = av1_cost_bit(fc->zeromv_prob[i], 0);
      x->zeromv_mode_cost[i][1] = av1_cost_bit(fc->zeromv_prob[i], 1);
Yue Chen's avatar
Yue Chen committed
301
#endif
302
303
304
    }

    for (i = 0; i < REFMV_MODE_CONTEXTS; ++i) {
Yue Chen's avatar
Yue Chen committed
305
306
307
#if CONFIG_NEW_MULTISYMBOL
      av1_cost_tokens_from_cdf(x->refmv_mode_cost[i], fc->refmv_cdf[i], NULL);
#else
308
309
      x->refmv_mode_cost[i][0] = av1_cost_bit(fc->refmv_prob[i], 0);
      x->refmv_mode_cost[i][1] = av1_cost_bit(fc->refmv_prob[i], 1);
Yue Chen's avatar
Yue Chen committed
310
#endif
311
312
313
    }

    for (i = 0; i < DRL_MODE_CONTEXTS; ++i) {
Yue Chen's avatar
Yue Chen committed
314
315
316
#if CONFIG_NEW_MULTISYMBOL
      av1_cost_tokens_from_cdf(x->drl_mode_cost0[i], fc->drl_cdf[i], NULL);
#else
317
318
      x->drl_mode_cost0[i][0] = av1_cost_bit(fc->drl_prob[i], 0);
      x->drl_mode_cost0[i][1] = av1_cost_bit(fc->drl_prob[i], 1);
Yue Chen's avatar
Yue Chen committed
319
#endif
320
321
    }
    for (i = 0; i < INTER_MODE_CONTEXTS; ++i)
Yue Chen's avatar
Yue Chen committed
322
323
      av1_cost_tokens_from_cdf(x->inter_compound_mode_cost[i],
                               fc->inter_compound_mode_cdf[i], NULL);
324
#if CONFIG_WEDGE || CONFIG_COMPOUND_SEGMENT
325
326
327
    for (i = 0; i < BLOCK_SIZES_ALL; ++i)
      av1_cost_tokens_from_cdf(x->compound_type_cost[i],
                               fc->compound_type_cdf[i], NULL);
328
#endif  // CONFIG_WEDGE || CONFIG_COMPOUND_SEGMENT
329
330
#if CONFIG_COMPOUND_SINGLEREF
    for (i = 0; i < INTER_MODE_CONTEXTS; ++i)
Yue Chen's avatar
Yue Chen committed
331
332
      av1_cost_tokens_from_cdf(x->inter_singleref_comp_mode_cost[i],
                               fc->inter_singleref_comp_mode_cdf[i], NULL);
333
334
335
#endif  // CONFIG_COMPOUND_SINGLEREF
#if CONFIG_INTERINTRA
    for (i = 0; i < BLOCK_SIZE_GROUPS; ++i)
Yue Chen's avatar
Yue Chen committed
336
337
      av1_cost_tokens_from_cdf(x->interintra_mode_cost[i],
                               fc->interintra_mode_cdf[i], NULL);
338
339
340
#endif  // CONFIG_INTERINTRA
#if CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
    for (i = BLOCK_8X8; i < BLOCK_SIZES_ALL; i++) {
341
342
      av1_cost_tokens_from_cdf(x->motion_mode_cost[i], fc->motion_mode_cdf[i],
                               NULL);
343
344
345
    }
#if CONFIG_MOTION_VAR && CONFIG_WARPED_MOTION
    for (i = BLOCK_8X8; i < BLOCK_SIZES_ALL; i++) {
346
347
348
349
350
#if CONFIG_NCOBMC_ADAPT_WEIGHT
      av1_cost_tokens_from_cdf(x->motion_mode_cost2[i], fc->ncobmc_cdf[i],
                               NULL);
#endif
#if CONFIG_NEW_MULTISYMBOL || CONFIG_NCOBMC_ADAPT_WEIGHT
Yue Chen's avatar
Yue Chen committed
351
352
      av1_cost_tokens_from_cdf(x->motion_mode_cost1[i], fc->obmc_cdf[i], NULL);
#else
353
354
      x->motion_mode_cost1[i][0] = av1_cost_bit(fc->obmc_prob[i], 0);
      x->motion_mode_cost1[i][1] = av1_cost_bit(fc->obmc_prob[i], 1);
Yue Chen's avatar
Yue Chen committed
355
#endif
356
357
358
359
    }
#endif  // CONFIG_MOTION_VAR && CONFIG_WARPED_MOTION
#if CONFIG_MOTION_VAR && CONFIG_NCOBMC_ADAPT_WEIGHT
    for (i = ADAPT_OVERLAP_BLOCK_8X8; i < ADAPT_OVERLAP_BLOCKS; ++i) {
360
361
      av1_cost_tokens_from_cdf(x->ncobmc_mode_cost[i], fc->ncobmc_mode_cdf[i],
                               NULL);
362
363
364
365
    }
#endif  // CONFIG_MOTION_VAR && CONFIG_NCOBMC_ADAPT_WEIGHT
#endif  // CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
  }
Jingning Han's avatar
Jingning Han committed
366
367
368
369
370
371
}

// Values are now correlated to quantizer.
static int sad_per_bit16lut_8[QINDEX_RANGE];
static int sad_per_bit4lut_8[QINDEX_RANGE];

372
#if CONFIG_HIGHBITDEPTH
Jingning Han's avatar
Jingning Han committed
373
374
375
376
377
378
379
static int sad_per_bit16lut_10[QINDEX_RANGE];
static int sad_per_bit4lut_10[QINDEX_RANGE];
static int sad_per_bit16lut_12[QINDEX_RANGE];
static int sad_per_bit4lut_12[QINDEX_RANGE];
#endif

static void init_me_luts_bd(int *bit16lut, int *bit4lut, int range,
Yaowu Xu's avatar
Yaowu Xu committed
380
                            aom_bit_depth_t bit_depth) {
Jingning Han's avatar
Jingning Han committed
381
382
383
384
385
  int i;
  // Initialize the sad lut tables using a formulaic calculation for now.
  // This is to make it easier to resolve the impact of experimental changes
  // to the quantizer tables.
  for (i = 0; i < range; i++) {
Yaowu Xu's avatar
Yaowu Xu committed
386
    const double q = av1_convert_qindex_to_q(i, bit_depth);
Jingning Han's avatar
Jingning Han committed
387
388
389
390
391
    bit16lut[i] = (int)(0.0418 * q + 2.4107);
    bit4lut[i] = (int)(0.063 * q + 2.742);
  }
}

Yaowu Xu's avatar
Yaowu Xu committed
392
void av1_init_me_luts(void) {
Jingning Han's avatar
Jingning Han committed
393
  init_me_luts_bd(sad_per_bit16lut_8, sad_per_bit4lut_8, QINDEX_RANGE,
Yaowu Xu's avatar
Yaowu Xu committed
394
                  AOM_BITS_8);
395
#if CONFIG_HIGHBITDEPTH
Jingning Han's avatar
Jingning Han committed
396
  init_me_luts_bd(sad_per_bit16lut_10, sad_per_bit4lut_10, QINDEX_RANGE,
Yaowu Xu's avatar
Yaowu Xu committed
397
                  AOM_BITS_10);
Jingning Han's avatar
Jingning Han committed
398
  init_me_luts_bd(sad_per_bit16lut_12, sad_per_bit4lut_12, QINDEX_RANGE,
Yaowu Xu's avatar
Yaowu Xu committed
399
                  AOM_BITS_12);
Jingning Han's avatar
Jingning Han committed
400
401
402
#endif
}

403
404
static const int rd_boost_factor[16] = { 64, 32, 32, 32, 24, 16, 12, 12,
                                         8,  8,  4,  4,  2,  2,  1,  0 };
Jingning Han's avatar
Jingning Han committed
405
static const int rd_frame_type_factor[FRAME_UPDATE_TYPES] = {
406
  128, 144, 128, 128, 144,
407
#if CONFIG_EXT_REFS
408
  // TODO(zoeliu): To adjust further following factor values.
409
  128, 128, 128,
410
411
  // TODO(weitinglin): We should investigate if the values should be the same
  //                   as the value used by OVERLAY frame
412
  144,  // INTNL_OVERLAY_UPDATE
Zoe Liu's avatar
Zoe Liu committed
413
  128   // INTNL_ARF_UPDATE
414
#endif  // CONFIG_EXT_REFS
Jingning Han's avatar
Jingning Han committed
415
416
};

Yaowu Xu's avatar
Yaowu Xu committed
417
418
int av1_compute_rd_mult(const AV1_COMP *cpi, int qindex) {
  const int64_t q = av1_dc_quant(qindex, 0, cpi->common.bit_depth);
419
#if CONFIG_HIGHBITDEPTH
Jingning Han's avatar
Jingning Han committed
420
421
  int64_t rdmult = 0;
  switch (cpi->common.bit_depth) {
Yaowu Xu's avatar
Yaowu Xu committed
422
423
424
    case AOM_BITS_8: rdmult = 88 * q * q / 24; break;
    case AOM_BITS_10: rdmult = ROUND_POWER_OF_TWO(88 * q * q / 24, 4); break;
    case AOM_BITS_12: rdmult = ROUND_POWER_OF_TWO(88 * q * q / 24, 8); break;
Jingning Han's avatar
Jingning Han committed
425
    default:
Yaowu Xu's avatar
Yaowu Xu committed
426
      assert(0 && "bit_depth should be AOM_BITS_8, AOM_BITS_10 or AOM_BITS_12");
Jingning Han's avatar
Jingning Han committed
427
428
429
430
      return -1;
  }
#else
  int64_t rdmult = 88 * q * q / 24;
431
#endif  // CONFIG_HIGHBITDEPTH
Jingning Han's avatar
Jingning Han committed
432
433
434
  if (cpi->oxcf.pass == 2 && (cpi->common.frame_type != KEY_FRAME)) {
    const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
    const FRAME_UPDATE_TYPE frame_type = gf_group->update_type[gf_group->index];
Yaowu Xu's avatar
Yaowu Xu committed
435
    const int boost_index = AOMMIN(15, (cpi->rc.gfu_boost / 100));
Jingning Han's avatar
Jingning Han committed
436
437
438
439

    rdmult = (rdmult * rd_frame_type_factor[frame_type]) >> 7;
    rdmult += ((rdmult * rd_boost_factor[boost_index]) >> 7);
  }
440
  if (rdmult < 1) rdmult = 1;
Jingning Han's avatar
Jingning Han committed
441
442
443
  return (int)rdmult;
}

Yaowu Xu's avatar
Yaowu Xu committed
444
static int compute_rd_thresh_factor(int qindex, aom_bit_depth_t bit_depth) {
Jingning Han's avatar
Jingning Han committed
445
  double q;
446
#if CONFIG_HIGHBITDEPTH
Jingning Han's avatar
Jingning Han committed
447
  switch (bit_depth) {
Yaowu Xu's avatar
Yaowu Xu committed
448
449
450
    case AOM_BITS_8: q = av1_dc_quant(qindex, 0, AOM_BITS_8) / 4.0; break;
    case AOM_BITS_10: q = av1_dc_quant(qindex, 0, AOM_BITS_10) / 16.0; break;
    case AOM_BITS_12: q = av1_dc_quant(qindex, 0, AOM_BITS_12) / 64.0; break;
Jingning Han's avatar
Jingning Han committed
451
    default:
Yaowu Xu's avatar
Yaowu Xu committed
452
      assert(0 && "bit_depth should be AOM_BITS_8, AOM_BITS_10 or AOM_BITS_12");
Jingning Han's avatar
Jingning Han committed
453
454
455
      return -1;
  }
#else
456
  (void)bit_depth;
Yaowu Xu's avatar
Yaowu Xu committed
457
  q = av1_dc_quant(qindex, 0, AOM_BITS_8) / 4.0;
458
#endif  // CONFIG_HIGHBITDEPTH
Jingning Han's avatar
Jingning Han committed
459
  // TODO(debargha): Adjust the function below.
Yaowu Xu's avatar
Yaowu Xu committed
460
  return AOMMAX((int)(pow(q, RD_THRESH_POW) * 5.12), 8);
Jingning Han's avatar
Jingning Han committed
461
462
}

Yaowu Xu's avatar
Yaowu Xu committed
463
void av1_initialize_me_consts(const AV1_COMP *cpi, MACROBLOCK *x, int qindex) {
464
#if CONFIG_HIGHBITDEPTH
Jingning Han's avatar
Jingning Han committed
465
  switch (cpi->common.bit_depth) {
Yaowu Xu's avatar
Yaowu Xu committed
466
    case AOM_BITS_8:
Jingning Han's avatar
Jingning Han committed
467
468
469
      x->sadperbit16 = sad_per_bit16lut_8[qindex];
      x->sadperbit4 = sad_per_bit4lut_8[qindex];
      break;
Yaowu Xu's avatar
Yaowu Xu committed
470
    case AOM_BITS_10:
Jingning Han's avatar
Jingning Han committed
471
472
473
      x->sadperbit16 = sad_per_bit16lut_10[qindex];
      x->sadperbit4 = sad_per_bit4lut_10[qindex];
      break;
Yaowu Xu's avatar
Yaowu Xu committed
474
    case AOM_BITS_12:
Jingning Han's avatar
Jingning Han committed
475
476
477
478
      x->sadperbit16 = sad_per_bit16lut_12[qindex];
      x->sadperbit4 = sad_per_bit4lut_12[qindex];
      break;
    default:
Yaowu Xu's avatar
Yaowu Xu committed
479
      assert(0 && "bit_depth should be AOM_BITS_8, AOM_BITS_10 or AOM_BITS_12");
Jingning Han's avatar
Jingning Han committed
480
481
482
483
484
  }
#else
  (void)cpi;
  x->sadperbit16 = sad_per_bit16lut_8[qindex];
  x->sadperbit4 = sad_per_bit4lut_8[qindex];
485
#endif  // CONFIG_HIGHBITDEPTH
Jingning Han's avatar
Jingning Han committed
486
487
}

Yaowu Xu's avatar
Yaowu Xu committed
488
static void set_block_thresholds(const AV1_COMMON *cm, RD_OPT *rd) {
Jingning Han's avatar
Jingning Han committed
489
490
491
492
  int i, bsize, segment_id;

  for (segment_id = 0; segment_id < MAX_SEGMENTS; ++segment_id) {
    const int qindex =
Yaowu Xu's avatar
Yaowu Xu committed
493
        clamp(av1_get_qindex(&cm->seg, segment_id, cm->base_qindex) +
494
495
                  cm->y_dc_delta_q,
              0, MAXQ);
Jingning Han's avatar
Jingning Han committed
496
497
    const int q = compute_rd_thresh_factor(qindex, cm->bit_depth);

498
    for (bsize = 0; bsize < BLOCK_SIZES_ALL; ++bsize) {
Jingning Han's avatar
Jingning Han committed
499
500
501
502
503
      // Threshold here seems unnecessarily harsh but fine given actual
      // range of values used for cpi->sf.thresh_mult[].
      const int t = q * rd_thresh_block_size_factor[bsize];
      const int thresh_max = INT_MAX / t;

504
505
506
507
508
509
#if CONFIG_CB4X4
      for (i = 0; i < MAX_MODES; ++i)
        rd->threshes[segment_id][bsize][i] = rd->thresh_mult[i] < thresh_max
                                                 ? rd->thresh_mult[i] * t / 4
                                                 : INT_MAX;
#else
Jingning Han's avatar
Jingning Han committed
510
511
      if (bsize >= BLOCK_8X8) {
        for (i = 0; i < MAX_MODES; ++i)
512
513
514
          rd->threshes[segment_id][bsize][i] = rd->thresh_mult[i] < thresh_max
                                                   ? rd->thresh_mult[i] * t / 4
                                                   : INT_MAX;
Jingning Han's avatar
Jingning Han committed
515
516
517
518
519
520
521
      } else {
        for (i = 0; i < MAX_REFS; ++i)
          rd->threshes[segment_id][bsize][i] =
              rd->thresh_mult_sub8x8[i] < thresh_max
                  ? rd->thresh_mult_sub8x8[i] * t / 4
                  : INT_MAX;
      }
522
#endif
Jingning Han's avatar
Jingning Han committed
523
524
525
526
    }
  }
}

Yaowu Xu's avatar
Yaowu Xu committed
527
528
void av1_set_mvcost(MACROBLOCK *x, MV_REFERENCE_FRAME ref_frame, int ref,
                    int ref_mv_idx) {
529
  MB_MODE_INFO_EXT *mbmi_ext = x->mbmi_ext;
Yaowu Xu's avatar
Yaowu Xu committed
530
531
532
533
  int8_t rf_type = av1_ref_frame_type(x->e_mbd.mi[0]->mbmi.ref_frame);
  int nmv_ctx = av1_nmv_ctx(mbmi_ext->ref_mv_count[rf_type],
                            mbmi_ext->ref_mv_stack[rf_type], ref, ref_mv_idx);
  (void)ref_frame;
534
535
536
537
  x->mvcost = x->mv_cost_stack[nmv_ctx];
  x->nmvjointcost = x->nmv_vec_cost[nmv_ctx];
}

538
#if CONFIG_LV_MAP
539
#if !LV_MAP_PROB
540
541
542
543
static void get_rate_cost(aom_prob p, int cost[2]) {
  cost[0] = av1_cost_bit(p, 0);
  cost[1] = av1_cost_bit(p, 1);
}
544
#endif  // !LV_MAP_PROB
545
546

void av1_fill_coeff_costs(MACROBLOCK *x, FRAME_CONTEXT *fc) {
547
  for (int tx_size = 0; tx_size < TX_SIZES; ++tx_size) {
548
549
550
    for (int plane = 0; plane < PLANE_TYPES; ++plane) {
      LV_MAP_COEFF_COST *pcost = &x->coeff_costs[tx_size][plane];

551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
#if LV_MAP_PROB
      for (int ctx = 0; ctx < TXB_SKIP_CONTEXTS; ++ctx)
        av1_cost_tokens_from_cdf(pcost->txb_skip_cost[ctx],
                                 fc->txb_skip_cdf[tx_size][ctx], NULL);

      for (int ctx = 0; ctx < SIG_COEF_CONTEXTS; ++ctx)
        av1_cost_tokens_from_cdf(pcost->nz_map_cost[ctx],
                                 fc->nz_map_cdf[tx_size][plane][ctx], NULL);

      for (int ctx = 0; ctx < EOB_COEF_CONTEXTS; ++ctx)
        av1_cost_tokens_from_cdf(pcost->eob_cost[ctx],
                                 fc->eob_flag_cdf[tx_size][plane][ctx], NULL);

      for (int ctx = 0; ctx < DC_SIGN_CONTEXTS; ++ctx)
        av1_cost_tokens_from_cdf(pcost->dc_sign_cost[ctx],
                                 fc->dc_sign_cdf[plane][ctx], NULL);

      for (int layer = 0; layer < NUM_BASE_LEVELS; ++layer)
        for (int ctx = 0; ctx < COEFF_BASE_CONTEXTS; ++ctx)
          av1_cost_tokens_from_cdf(
              pcost->base_cost[layer][ctx],
              fc->coeff_base_cdf[tx_size][plane][layer][ctx], NULL);

574
575
576
577
578
579
580
#if BR_NODE
      for (int br = 0; br < BASE_RANGE_SETS; ++br)
        for (int ctx = 0; ctx < LEVEL_CONTEXTS; ++ctx)
          av1_cost_tokens_from_cdf(pcost->br_cost[br][ctx],
                                   fc->coeff_br_cdf[tx_size][plane][br][ctx],
                                   NULL);

581
582
583
      for (int ctx = 0; ctx < LEVEL_CONTEXTS; ++ctx) {
        int lps_rate[2];
        av1_cost_tokens_from_cdf(lps_rate,
584
                                 fc->coeff_lps_cdf[tx_size][plane][ctx], NULL);
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616

        for (int base_range = 0; base_range < COEFF_BASE_RANGE + 1;
             ++base_range) {
          int br_set_idx = base_range < COEFF_BASE_RANGE
                               ? coeff_to_br_index[base_range]
                               : BASE_RANGE_SETS;

          pcost->lps_cost[ctx][base_range] = 0;

          for (int idx = 0; idx < BASE_RANGE_SETS; ++idx) {
            if (idx == br_set_idx) {
              pcost->lps_cost[ctx][base_range] += pcost->br_cost[idx][ctx][1];

              int br_base = br_index_to_coeff[br_set_idx];
              int br_offset = base_range - br_base;
              int extra_bits = (1 << br_extra_bits[idx]) - 1;
              for (int tok = 0; tok < extra_bits; ++tok) {
                if (tok == br_offset) {
                  pcost->lps_cost[ctx][base_range] += lps_rate[1];
                  break;
                } else {
                  pcost->lps_cost[ctx][base_range] += lps_rate[0];
                }
              }
              break;
            } else {
              pcost->lps_cost[ctx][base_range] += pcost->br_cost[idx][ctx][0];
            }
          }
          // load the base range cost
        }
      }
Angie Chiang's avatar
Angie Chiang committed
617
618
619
620
621
#else   // BR_NODE
      for (int ctx = 0; ctx < LEVEL_CONTEXTS; ++ctx)
        av1_cost_tokens_from_cdf(pcost->lps_cost[ctx],
                                 fc->coeff_lps_cdf[tx_size][plane][ctx], NULL);
#endif  // BR_NODE
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
#if CONFIG_CTX1D
      for (int tx_class = 0; tx_class < TX_CLASSES; ++tx_class)
        av1_cost_tokens_from_cdf(pcost->eob_mode_cost[tx_class],
                                 fc->eob_mode_cdf[tx_size][plane][tx_class],
                                 NULL);

      for (int tx_class = 0; tx_class < TX_CLASSES; ++tx_class)
        for (int ctx = 0; ctx < EMPTY_LINE_CONTEXTS; ++ctx)
          av1_cost_tokens_from_cdf(
              pcost->empty_line_cost[tx_class][ctx],
              fc->empty_line_cdf[tx_size][plane][tx_class][ctx], NULL);

      for (int tx_class = 0; tx_class < TX_CLASSES; ++tx_class)
        for (int ctx = 0; ctx < HV_EOB_CONTEXTS; ++ctx)
          av1_cost_tokens_from_cdf(
              pcost->hv_eob_cost[tx_class][ctx],
              fc->hv_eob_cdf[tx_size][plane][tx_class][ctx], NULL);
#endif  // CONFIG_CTX1D
640
#else   // LV_MAP_PROB
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
      for (int ctx = 0; ctx < TXB_SKIP_CONTEXTS; ++ctx)
        get_rate_cost(fc->txb_skip[tx_size][ctx], pcost->txb_skip_cost[ctx]);

      for (int ctx = 0; ctx < SIG_COEF_CONTEXTS; ++ctx)
        get_rate_cost(fc->nz_map[tx_size][plane][ctx], pcost->nz_map_cost[ctx]);

      for (int ctx = 0; ctx < EOB_COEF_CONTEXTS; ++ctx)
        get_rate_cost(fc->eob_flag[tx_size][plane][ctx], pcost->eob_cost[ctx]);

      for (int ctx = 0; ctx < DC_SIGN_CONTEXTS; ++ctx)
        get_rate_cost(fc->dc_sign[plane][ctx], pcost->dc_sign_cost[ctx]);

      for (int layer = 0; layer < NUM_BASE_LEVELS; ++layer)
        for (int ctx = 0; ctx < COEFF_BASE_CONTEXTS; ++ctx)
          get_rate_cost(fc->coeff_base[tx_size][plane][layer][ctx],
                        pcost->base_cost[layer][ctx]);

      for (int ctx = 0; ctx < LEVEL_CONTEXTS; ++ctx)
        get_rate_cost(fc->coeff_lps[tx_size][plane][ctx], pcost->lps_cost[ctx]);
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675

#if CONFIG_CTX1D
      for (int tx_class = 0; tx_class < TX_CLASSES; ++tx_class)
        get_rate_cost(fc->eob_mode[tx_size][plane][tx_class],
                      pcost->eob_mode_cost[tx_class]);

      for (int tx_class = 0; tx_class < TX_CLASSES; ++tx_class)
        for (int ctx = 0; ctx < EMPTY_LINE_CONTEXTS; ++ctx)
          get_rate_cost(fc->empty_line[tx_size][plane][tx_class][ctx],
                        pcost->empty_line_cost[tx_class][ctx]);

      for (int tx_class = 0; tx_class < TX_CLASSES; ++tx_class)
        for (int ctx = 0; ctx < HV_EOB_CONTEXTS; ++ctx)
          get_rate_cost(fc->hv_eob[tx_size][plane][tx_class][ctx],
                        pcost->hv_eob_cost[tx_class][ctx]);
#endif  // CONFIG_CTX1D
676
#endif  // LV_MAP_PROB
677
678
679
    }
  }
}
680
#endif  // CONFIG_LV_MAP
681

682
683
void av1_fill_token_costs_from_cdf(av1_coeff_cost *cost,
                                   coeff_cdf_model (*cdf)[PLANE_TYPES]) {
hui su's avatar
hui su committed
684
685
686
687
688
689
690
691
692
693
694
695
696
697
  for (int tx = 0; tx < TX_SIZES; ++tx) {
    for (int pt = 0; pt < PLANE_TYPES; ++pt) {
      for (int rt = 0; rt < REF_TYPES; ++rt) {
        for (int band = 0; band < COEF_BANDS; ++band) {
          for (int ctx = 0; ctx < BAND_COEFF_CONTEXTS(band); ++ctx) {
            av1_cost_tokens_from_cdf(cost[tx][pt][rt][band][ctx],
                                     cdf[tx][pt][rt][band][ctx], NULL);
          }
        }
      }
    }
  }
}

Yaowu Xu's avatar
Yaowu Xu committed
698
699
void av1_initialize_rd_consts(AV1_COMP *cpi) {
  AV1_COMMON *const cm = &cpi->common;
Jingning Han's avatar
Jingning Han committed
700
701
  MACROBLOCK *const x = &cpi->td.mb;
  RD_OPT *const rd = &cpi->rd;
702
  int nmv_ctx;
Jingning Han's avatar
Jingning Han committed
703

Yaowu Xu's avatar
Yaowu Xu committed
704
  aom_clear_system_state();
Jingning Han's avatar
Jingning Han committed
705

Yaowu Xu's avatar
Yaowu Xu committed
706
  rd->RDMULT = av1_compute_rd_mult(cpi, cm->base_qindex + cm->y_dc_delta_q);
Jingning Han's avatar
Jingning Han committed
707

708
  set_error_per_bit(x, rd->RDMULT);
Jingning Han's avatar
Jingning Han committed
709
710
711

  set_block_thresholds(cm, rd);

712
  for (nmv_ctx = 0; nmv_ctx < NMV_CONTEXTS; ++nmv_ctx) {
RogerZhou's avatar
RogerZhou committed
713
714
715
716
717
718
719
720
721
722
723
724
725
#if CONFIG_AMVR
    if (cm->cur_frame_mv_precision_level) {
      av1_build_nmv_cost_table(x->nmv_vec_cost[nmv_ctx], x->nmvcost[nmv_ctx],
                               &cm->fc->nmvc[nmv_ctx], MV_SUBPEL_NONE);
    } else {
      av1_build_nmv_cost_table(
          x->nmv_vec_cost[nmv_ctx],
          cm->allow_high_precision_mv ? x->nmvcost_hp[nmv_ctx]
                                      : x->nmvcost[nmv_ctx],
          &cm->fc->nmvc[nmv_ctx], cm->allow_high_precision_mv);
    }

#else
726
727
728
729
730
    av1_build_nmv_cost_table(
        x->nmv_vec_cost[nmv_ctx],
        cm->allow_high_precision_mv ? x->nmvcost_hp[nmv_ctx]
                                    : x->nmvcost[nmv_ctx],
        &cm->fc->nmvc[nmv_ctx], cm->allow_high_precision_mv);
RogerZhou's avatar
RogerZhou committed
731
#endif
732
733
734
  }
  x->mvcost = x->mv_cost_stack[0];
  x->nmvjointcost = x->nmv_vec_cost[0];
Jingning Han's avatar
Jingning Han committed
735

Alex Converse's avatar
Alex Converse committed
736
737
738
739
740
741
742
743
744
745
#if CONFIG_INTRABC
  if (frame_is_intra_only(cm) && cm->allow_screen_content_tools &&
      cpi->oxcf.pass != 1) {
    av1_build_nmv_cost_table(
        x->nmv_vec_cost[0],
        cm->allow_high_precision_mv ? x->nmvcost_hp[0] : x->nmvcost[0],
        &cm->fc->ndvc, MV_SUBPEL_NONE);
  }
#endif

746
#if CONFIG_GLOBAL_MOTION
747
  if (cpi->oxcf.pass != 1) {
748
    for (int i = 0; i < TRANS_TYPES; ++i)
749
#if GLOBAL_TRANS_TYPES > 4
750
751
      cpi->gmtype_cost[i] = (1 + (i > 0 ? GLOBAL_TYPE_BITS : 0))
                            << AV1_PROB_COST_SHIFT;
752
753
754
755
756
757
758
759
#else
      // IDENTITY: 1 bit
      // TRANSLATION: 3 bits
      // ROTZOOM: 2 bits
      // AFFINE: 3 bits
      cpi->gmtype_cost[i] = (1 + (i > 0 ? (i == ROTZOOM ? 1 : 2) : 0))
                            << AV1_PROB_COST_SHIFT;
#endif  // GLOBAL_TRANS_TYPES > 4
Jingning Han's avatar
Jingning Han committed
760
  }
761
#endif  // CONFIG_GLOBAL_MOTION
Jingning Han's avatar
Jingning Han committed
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
}

static void model_rd_norm(int xsq_q10, int *r_q10, int *d_q10) {
  // NOTE: The tables below must be of the same size.

  // The functions described below are sampled at the four most significant
  // bits of x^2 + 8 / 256.

  // Normalized rate:
  // This table models the rate for a Laplacian source with given variance
  // when quantized with a uniform quantizer with given stepsize. The
  // closed form expression is:
  // Rn(x) = H(sqrt(r)) + sqrt(r)*[1 + H(r)/(1 - r)],
  // where r = exp(-sqrt(2) * x) and x = qpstep / sqrt(variance),
  // and H(x) is the binary entropy function.
  static const int rate_tab_q10[] = {
778
779
780
781
782
783
784
785
786
    65536, 6086, 5574, 5275, 5063, 4899, 4764, 4651, 4553, 4389, 4255, 4142,
    4044,  3958, 3881, 3811, 3748, 3635, 3538, 3453, 3376, 3307, 3244, 3186,
    3133,  3037, 2952, 2877, 2809, 2747, 2690, 2638, 2589, 2501, 2423, 2353,
    2290,  2232, 2179, 2130, 2084, 2001, 1928, 1862, 1802, 1748, 1698, 1651,
    1608,  1530, 1460, 1398, 1342, 1290, 1243, 1199, 1159, 1086, 1021, 963,
    911,   864,  821,  781,  745,  680,  623,  574,  530,  490,  455,  424,
    395,   345,  304,  269,  239,  213,  190,  171,  154,  126,  104,  87,
    73,    61,   52,   44,   38,   28,   21,   16,   12,   10,   8,    6,
    5,     3,    2,    1,    1,    1,    0,    0,
Jingning Han's avatar
Jingning Han committed
787
788
789
790
791
792
793
794
795
  };
  // Normalized distortion:
  // This table models the normalized distortion for a Laplacian source
  // with given variance when quantized with a uniform quantizer
  // with given stepsize. The closed form expression is:
  // Dn(x) = 1 - 1/sqrt(2) * x / sinh(x/sqrt(2))
  // where x = qpstep / sqrt(variance).
  // Note the actual distortion is Dn * variance.
  static const int dist_tab_q10[] = {
796
797
798
799
800
801
802
803
804
    0,    0,    1,    1,    1,    2,    2,    2,    3,    3,    4,    5,
    5,    6,    7,    7,    8,    9,    11,   12,   13,   15,   16,   17,
    18,   21,   24,   26,   29,   31,   34,   36,   39,   44,   49,   54,
    59,   64,   69,   73,   78,   88,   97,   106,  115,  124,  133,  142,
    151,  167,  184,  200,  215,  231,  245,  260,  274,  301,  327,  351,
    375,  397,  418,  439,  458,  495,  528,  559,  587,  613,  637,  659,
    680,  717,  749,  777,  801,  823,  842,  859,  874,  899,  919,  936,
    949,  960,  969,  977,  983,  994,  1001, 1006, 1010, 1013, 1015, 1017,
    1018, 1020, 1022, 1022, 1023, 1023, 1023, 1024,
Jingning Han's avatar
Jingning Han committed
805
806
  };
  static const int xsq_iq_q10[] = {
807
808
809
810
811
812
813
814
815
816
817
818
    0,      4,      8,      12,     16,     20,     24,     28,     32,
    40,     48,     56,     64,     72,     80,     88,     96,     112,
    128,    144,    160,    176,    192,    208,    224,    256,    288,
    320,    352,    384,    416,    448,    480,    544,    608,    672,
    736,    800,    864,    928,    992,    1120,   1248,   1376,   1504,
    1632,   1760,   1888,   2016,   2272,   2528,   2784,   3040,   3296,
    3552,   3808,   4064,   4576,   5088,   5600,   6112,   6624,   7136,
    7648,   8160,   9184,   10208,  11232,  12256,  13280,  14304,  15328,
    16352,  18400,  20448,  22496,  24544,  26592,  28640,  30688,  32736,
    36832,  40928,  45024,  49120,  53216,  57312,  61408,  65504,  73696,
    81888,  90080,  98272,  106464, 114656, 122848, 131040, 147424, 163808,
    180192, 196576, 212960, 229344, 245728,
Jingning Han's avatar
Jingning Han committed
819
820
821
822
823
824
825
826
827
828
829
  };
  const int tmp = (xsq_q10 >> 2) + 8;
  const int k = get_msb(tmp) - 3;
  const int xq = (k << 3) + ((tmp >> k) & 0x7);
  const int one_q10 = 1 << 10;
  const int a_q10 = ((xsq_q10 - xsq_iq_q10[xq]) << 10) >> (2 + k);
  const int b_q10 = one_q10 - a_q10;
  *r_q10 = (rate_tab_q10[xq] * b_q10 + rate_tab_q10[xq + 1] * a_q10) >> 10;
  *d_q10 = (dist_tab_q10[xq] * b_q10 + dist_tab_q10[xq + 1] * a_q10) >> 10;
}

Yaowu Xu's avatar
Yaowu Xu committed
830
831
832
void av1_model_rd_from_var_lapndz(int64_t var, unsigned int n_log2,
                                  unsigned int qstep, int *rate,
                                  int64_t *dist) {
Jingning Han's avatar
Jingning Han committed
833
834
835
836
837
838
839
840
841
842
843
844
845
846
  // This function models the rate and distortion for a Laplacian
  // source with given variance when quantized with a uniform quantizer
  // with given stepsize. The closed form expressions are in:
  // Hang and Chen, "Source Model for transform video coder and its
  // application - Part I: Fundamental Theory", IEEE Trans. Circ.
  // Sys. for Video Tech., April 1997.
  if (var == 0) {
    *rate = 0;
    *dist = 0;
  } else {
    int d_q10, r_q10;
    static const uint32_t MAX_XSQ_Q10 = 245727;
    const uint64_t xsq_q10_64 =
        (((uint64_t)qstep * qstep << (n_log2 + 10)) + (var >> 1)) / var;
Yaowu Xu's avatar
Yaowu Xu committed
847
    const int xsq_q10 = (int)AOMMIN(xsq_q10_64, MAX_XSQ_Q10);
Jingning Han's avatar
Jingning Han committed
848
    model_rd_norm(xsq_q10, &r_q10, &d_q10);
Yaowu Xu's avatar
Yaowu Xu committed
849
    *rate = ROUND_POWER_OF_TWO(r_q10 << n_log2, 10 - AV1_PROB_COST_SHIFT);
Jingning Han's avatar
Jingning Han committed
850
851
852
853
    *dist = (var * (int64_t)d_q10 + 512) >> 10;
  }
}

854
855
856
857
static void get_entropy_contexts_plane(
    BLOCK_SIZE plane_bsize, TX_SIZE tx_size, const struct macroblockd_plane *pd,
    ENTROPY_CONTEXT t_above[2 * MAX_MIB_SIZE],
    ENTROPY_CONTEXT t_left[2 * MAX_MIB_SIZE]) {
858
859
  const int num_4x4_w = block_size_wide[plane_bsize] >> tx_size_wide_log2[0];
  const int num_4x4_h = block_size_high[plane_bsize] >> tx_size_high_log2[0];
Jingning Han's avatar
Jingning Han committed
860
861
862
  const ENTROPY_CONTEXT *const above = pd->above_context;
  const ENTROPY_CONTEXT *const left = pd->left_context;

863
864
865
866
867
868
#if CONFIG_LV_MAP
  memcpy(t_above, above, sizeof(ENTROPY_CONTEXT) * num_4x4_w);
  memcpy(t_left, left, sizeof(ENTROPY_CONTEXT) * num_4x4_h);
  return;
#endif  // CONFIG_LV_MAP

Jingning Han's avatar
Jingning Han committed
869
  int i;
870

871
#if CONFIG_CHROMA_2X2
872
  switch (tx_size) {
873
    case TX_2X2:
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
      memcpy(t_above, above, sizeof(ENTROPY_CONTEXT) * num_4x4_w);
      memcpy(t_left, left, sizeof(ENTROPY_CONTEXT) * num_4x4_h);
      break;
    case TX_4X4:
      for (i = 0; i < num_4x4_w; i += 2)
        t_above[i] = !!*(const uint16_t *)&above[i];
      for (i = 0; i < num_4x4_h; i += 2)
        t_left[i] = !!*(const uint16_t *)&left[i];
      break;
    case TX_8X8:
      for (i = 0; i < num_4x4_w; i += 4)
        t_above[i] = !!*(const uint32_t *)&above[i];
      for (i = 0; i < num_4x4_h; i += 4)
        t_left[i] = !!*(const uint32_t *)&left[i];
      break;
    case TX_16X16:
      for (i = 0; i < num_4x4_w; i += 8)
        t_above[i] = !!*(const uint64_t *)&above[i];
      for (i = 0; i < num_4x4_h; i += 8)
        t_left[i] = !!*(const uint64_t *)&left[i];
      break;
    case TX_32X32:
      for (i = 0; i < num_4x4_w; i += 16)
        t_above[i] =
            !!(*(const uint64_t *)&above[i] | *(const uint64_t *)&above[i + 8]);
      for (i = 0; i < num_4x4_h; i += 16)
        t_left[i] =
            !!(*(const uint64_t *)&left[i] | *(const uint64_t *)&left[i + 8]);
      break;
903
#if CONFIG_TX64X64
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
    case TX_32X64:
      for (i = 0; i < num_4x4_w; i += 16)
        t_above[i] =
            !!(*(const uint64_t *)&above[i] | *(const uint64_t *)&above[i + 8]);
      for (i = 0; i < num_4x4_h; i += 32)
        t_left[i] =
            !!(*(const uint64_t *)&left[i] | *(const uint64_t *)&left[i + 8] |
               *(const uint64_t *)&left[i + 16] |
               *(const uint64_t *)&left[i + 24]);
      break;
    case TX_64X32:
      for (i = 0; i < num_4x4_w; i += 32)
        t_above[i] =
            !!(*(const uint64_t *)&above[i] | *(const uint64_t *)&above[i + 8] |
               *(const uint64_t *)&above[i + 16] |
               *(const uint64_t *)&above[i + 24]);
      for (i = 0; i < num_4x4_h; i += 16)
        t_left[i] =
            !!(*(const uint64_t *)&left[i] | *(const uint64_t *)&left[i + 8]);
      break;
924
925
926
927
928
929
930
931
932
933
934
935
936
    case TX_64X64:
      for (i = 0; i < num_4x4_w; i += 32)
        t_above[i] =
            !!(*(const uint64_t *)&above[i] | *(const uint64_t *)&above[i + 8] |
               *(const uint64_t *)&above[i + 16] |
               *(const uint64_t *)&above[i + 24]);
      for (i = 0; i < num_4x4_h; i += 32)
        t_left[i] =
            !!(*(const uint64_t *)&left[i] | *(const uint64_t *)&left[i + 8] |
               *(const uint64_t *)&left[i + 16] |
               *(const uint64_t *)&left[i + 24]);
      break;
#endif  // CONFIG_TX64X64
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
    case TX_4X8:
      for (i = 0; i < num_4x4_w; i += 2)
        t_above[i] = !!*(const uint16_t *)&above[i];
      for (i = 0; i < num_4x4_h; i += 4)
        t_left[i] = !!*(const uint32_t *)&left[i];
      break;
    case TX_8X4:
      for (i = 0; i < num_4x4_w; i += 4)
        t_above[i] = !!*(const uint32_t *)&above[i];
      for (i = 0; i < num_4x4_h; i += 2)
        t_left[i] = !!*(const uint16_t *)&left[i];
      break;
    case TX_8X16:
      for (i = 0; i < num_4x4_w; i += 4)
        t_above[i] = !!*(const uint32_t *)&above[i];
      for (i = 0; i < num_4x4_h; i += 8)
        t_left[i] = !!*(const uint64_t *)&left[i];
      break;
    case TX_16X8:
      for (i = 0; i < num_4x4_w; i += 8)
        t_above[i] = !!*(const uint64_t *)&above[i];
      for (i = 0; i < num_4x4_h; i += 4)
        t_left[i] = !!*(const uint32_t *)&left[i];
      break;
    case TX_16X32:
      for (i = 0; i < num_4x4_w; i += 8)
        t_above[i] = !!*(const uint64_t *)&above[i];
      for (i = 0; i < num_4x4_h; i += 16)
        t_left[i] =
            !!(*(const uint64_t *)&left[i] | *(const uint64_t *)&left[i + 8]);
      break;
    case TX_32X16:
      for (i = 0; i < num_4x4_w; i += 16)
        t_above[i] =
            !!(*(const uint64_t *)&above[i] | *(const uint64_t *)&above[i + 8]);
      for (i = 0; i < num_4x4_h; i += 8)
        t_left[i] = !!*(const uint64_t *)&left[i];
974
      break;
Yue Chen's avatar
Yue Chen committed
975
#if CONFIG_RECT_TX_EXT && (CONFIG_EXT_TX || CONFIG_VAR_TX)
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
    case TX_4X16:
      for (i = 0; i < num_4x4_w; i += 2)
        t_above[i] = !!*(const uint16_t *)&above[i];
      for (i = 0; i < num_4x4_h; i += 8)
        t_left[i] = !!*(const uint64_t *)&left[i];
      break;
    case TX_16X4:
      for (i = 0; i < num_4x4_w; i += 8)
        t_above[i] = !!*(const uint64_t *)&above[i];
      for (i = 0; i < num_4x4_h; i += 2)
        t_left[i] = !!*(const uint16_t *)&left[i];
      break;
    case TX_8X32:
      for (i = 0; i < num_4x4_w; i += 4)
        t_above[i] = !!*(const uint32_t *)&above[i];
      for (i = 0; i < num_4x4_h; i += 16)
        t_left[i] =
            !!(*(const uint64_t *)&left[i] | *(const uint64_t *)&left[i + 8]);
      break;
    case TX_32X8:
      for (i = 0; i < num_4x4_w; i += 16)
        t_above[i] =
            !!(*(const uint64_t *)&above[i] | *(const uint64_t *)&above[i + 8]);
      for (i = 0; i < num_4x4_h; i += 4)
        t_left[i] = !!*(const uint32_t *)&left[i];
      break;
Yue Chen's avatar
Yue Chen committed
1002
#endif
1003
1004
1005
1006

    default: assert(0 && "Invalid transform size."); break;
  }
  return;
1007
#endif  // CONFIG_CHROMA_2X2
1008
1009

  switch (tx_size) {
Jingning Han's avatar
Jingning Han committed
1010
1011
1012
1013
    case TX_4X4:
      memcpy(t_above, above, sizeof(ENTROPY_CONTEXT) * num_4x4_w);
      memcpy(t_left, left, sizeof(ENTROPY_CONTEXT) * num_4x4_h);
      break;
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
    case TX_8X8:
      for (i = 0; i < num_4x4_w; i += 2)
        t_above[i] = !!*(const uint16_t *)&above[i];
      for (i = 0; i < num_4x4_h; i += 2)
        t_left[i] = !!*(const uint16_t *)&left[i];
      break;
    case TX_16X16:
      for (i = 0; i < num_4x4_w; i += 4)
        t_above[i] = !!*(const uint32_t *)&above[i];
      for (i = 0; i < num_4x4_h; i += 4)
        t_left[i] = !!*(const uint32_t *)&left[i];
      break;
    case TX_32X32:
      for (i = 0; i < num_4x4_w; i += 8)
        t_above[i] = !!*(const uint64_t *)&above[i];
      for (i = 0; i < num_4x4_h; i += 8)
        t_left[i] = !!*(const uint64_t *)&left[i];
      break;
1032
#if CONFIG_TX64X64
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
    case TX_32X64:
      for (i = 0; i < num_4x4_w; i += 8)
        t_above[i] = !!*(const uint64_t *)&above[i];
      for (i = 0; i < num_4x4_h; i += 16)
        t_left[i] =
            !!(*(const uint64_t *)&left[i] | *(const uint64_t *)&left[i + 8]);
      break;
    case TX_64X32:
      for (i = 0; i < num_4x4_w; i += 16)
        t_above[i] =
            !!(*(const uint64_t *)&above[i] | *(const uint64_t *)&above[i + 8]);
      for (i = 0; i < num_4x4_h; i += 8)
        t_left[i] = !!*(const uint64_t *)&left[i];
      break;
1047
1048
1049
1050
1051
    case TX_64X64:
      for (i = 0; i < num_4x4_w; i += 16)
        t_above[i] =
            !!(*(const uint64_t *)&above[i] | *(const uint64_t *)&above[i + 8]);
      for (i = 0; i < num_4x4_h; i += 16)
1052
1053
        t_left[i] =
            !!(*(const uint64_t *)&left[i] | *(const uint64_t *)&left[i + 8]);
1054
1055
      break;
#endif  // CONFIG_TX64X64
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
    case TX_4X8:
      memcpy(t_above, above, sizeof(ENTROPY_CONTEXT) * num_4x4_w);
      for (i = 0; i < num_4x4_h; i += 2)
        t_left[i] = !!*(const uint16_t *)&left[i];
      break;
    case TX_8X4:
      for (i = 0; i < num_4x4_w; i += 2)
        t_above[i] = !!*(const uint16_t *)&above[i];
      memcpy(t_left, left, sizeof(ENTROPY_CONTEXT) * num_4x4_h);
      break;
1066
    case TX_8X16:
Jingning Han's avatar
Jingning Han committed
1067
1068
      for (i = 0; i < num_4x4_w; i += 2)
        t_above[i] = !!*(const uint16_t *)&above[i];
1069
1070
1071
1072
1073
1074
      for (i = 0; i < num_4x4_h; i += 4)
        t_left[i] = !!*(const uint32_t *)&left[i];
      break;
    case TX_16X8:
      for (i = 0; i < num_4x4_w; i += 4)
        t_above[i] = !!*(const uint32_t *)&above[i];
Jingning Han's avatar
Jingning Han committed
1075
1076
1077
      for (i = 0; i < num_4x4_h; i += 2)
        t_left[i] = !!*(const uint16_t *)&left[i];
      break;
1078
    case TX_16X32:
Jingning Han's avatar
Jingning Han committed
1079
1080
      for (i = 0; i < num_4x4_w; i += 4)
        t_above[i] = !!*(const uint32_t *)&above[i];
1081
1082
      for (i = 0; i < num_4x4_h; i += 8)
        t_left[i] = !!*(const uint64_t *)&left[i];
Jingning Han's avatar
Jingning Han committed
1083
      break;
1084
    case TX_32X16:
Jingning Han's avatar
Jingning Han committed
1085
1086
      for (i = 0; i < num_4x4_w; i += 8)
        t_above[i] = !!*(const uint64_t *)&above[i];
1087
1088
1089
      for (i = 0; i < num_4x4_h; i += 4)
        t_left[i] = !!*(const uint32_t *)&left[i];
      break;
Yue Chen's avatar
Yue Chen committed
1090
#if CONFIG_RECT_TX_EXT && (CONFIG_EXT_TX || CONFIG_VAR_TX)
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
    case TX_4X16:
      memcpy(t_above, above, sizeof(ENTROPY_CONTEXT) * num_4x4_w);
      for (i = 0; i < num_4x4_h; i += 4)
        t_left[i] = !!*(const uint32_t *)&left[i];
      break;
    case TX_16X4:
      for (i = 0; i < num_4x4_w; i += 4)
        t_above[i] = !!*(const uint32_t *)&above[i];
      memcpy(t_left, left, sizeof(ENTROPY_CONTEXT) * num_4x4_h);
      break;
    case TX_8X32:
      for (i = 0; i < num_4x4_w; i += 2)
        t_above[i] = !!*(const uint16_t *)&above[i];
      for (i = 0; i < num_4x4_h; i += 8)
        t_left[i] = !!*(const uint64_t *)&left[i];
      break;
    case TX_32X8:
      for (i = 0; i < num_4x4_w; i += 8)
        t_above[i] = !!*(const uint64_t *)&above[i];
      for (i = 0; i < num_4x4_h; i += 2)
        t_left[i] = !!*(const uint16_t *)&left[i];
      break;
Yue Chen's avatar
Yue Chen committed
1113
#endif
clang-format's avatar
clang-format committed
1114
    default: assert(0 && "Invalid transform size."); break;
Jingning Han's avatar
Jingning Han committed
1115
1116
1117
  }
}

Yaowu Xu's avatar
Yaowu Xu committed
1118
1119
1120
1121
void av1_get_entropy_contexts(BLOCK_SIZE bsize, TX_SIZE tx_size,
                              const struct macroblockd_plane *pd,
                              ENTROPY_CONTEXT t_above[2 * MAX_MIB_SIZE],
                              ENTROPY_CONTEXT t_left[2 * MAX_MIB_SIZE]) {
1122
#if CONFIG_CHROMA_SUB8X8
1123
1124
1125
  const BLOCK_SIZE plane_bsize =
      AOMMAX(BLOCK_4X4, get_plane_block_size(bsize, pd));
#else
1126
  const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, pd);
1127
#endif
1128
  get_entropy_contexts_plane(plane_bsize, tx_size, pd, t_above, t_left);
1129
1130
}

1131
void av1_mv_pred(const AV1_COMP *cpi, MACROBLOCK *x, uint8_t *ref_y_buffer,
Yaowu Xu's avatar
Yaowu Xu committed
1132
                 int ref_y_stride, int ref_frame, BLOCK_SIZE block_size) {
Jingning Han's avatar
Jingning Han committed
1133
1134
1135
1136
1137
1138
1139
1140
  int i;
  int zero_seen = 0;
  int best_index = 0;
  int best_sad = INT_MAX;
  int this_sad = INT_MAX;
  int max_mv = 0;
  uint8_t *src_y_ptr = x->plane[0].src.buf;
  uint8_t *ref_y_ptr;
1141
  MV pred_mv[MAX_MV_REF_CANDIDATES + 1];
1142
1143
1144
1145
1146
1147
  int num_mv_refs = 0;

  pred_mv[num_mv_refs++] = x->mbmi_ext->ref_mvs[ref_frame][0].as_mv;
  if (x->mbmi_ext->ref_mvs[ref_frame][0].as_int !=
      x->mbmi_ext->ref_mvs[ref_frame][1].as_int) {
    pred_mv[num_mv_refs++] = x->mbmi_ext->ref_mvs[ref_frame][1].as_mv;
1148
  }
1149
1150
1151
  if (cpi->sf.adaptive_motion_search && block_size < x->max_partition_size)
    pred_mv[num_mv_refs++] = x->pred_mv[ref_frame];

Jingning Han's avatar
Jingning Han committed
1152
  assert(num_mv_refs <= (int)(sizeof(pred_mv) / sizeof(pred_mv[0])));
1153

Jingning Han's avatar
Jingning Han committed
1154
1155
1156
1157
1158
1159
  // Get the sad for each candidate reference mv.
  for (i = 0; i < num_mv_refs; ++i) {
    const MV *this_mv = &pred_mv[i];
    int fp_row, fp_col;
    fp_row = (this_mv->row + 3 + (this_mv->row >= 0)) >> 3;
    fp_col = (this_mv->col + 3 + (this_mv->col >= 0)) >> 3;
Yaowu Xu's avatar
Yaowu Xu committed
1160
    max_mv = AOMMAX(max_mv, AOMMAX(abs(this_mv->row), abs(this_mv->col)) >> 3);
Jingning Han's avatar
Jingning Han committed
1161

1162
1163
    if (fp_row == 0 && fp_col == 0 && zero_seen) continue;
    zero_seen |= (fp_row == 0 && fp_col == 0);
Jingning Han's avatar
Jingning Han committed
1164

1165
    ref_y_ptr = &ref_y_buffer[ref_y_stride * fp_row + fp_col];
Jingning Han's avatar
Jingning Han committed
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
    // Find sad for current vector.
    this_sad = cpi->fn_ptr[block_size].sdf(src_y_ptr, x->plane[0].src.stride,
                                           ref_y_ptr, ref_y_stride);
    // Note if it is the best so far.
    if (this_sad < best_sad) {
      best_sad = this_sad;
      best_index = i;
    }
  }

  // Note the index of the mv that worked best in the reference list.
  x->mv_best_ref_index[ref_frame] = best_index;
  x->max_mv_context[ref_frame] = max_mv;
  x->pred_mv_sad[ref_frame] = best_sad;
}