rd.c 45.6 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
39
40
41
42
#include "av1/encoder/cost.h"
#include "av1/encoder/encodemb.h"
#include "av1/encoder/encodemv.h"
#include "av1/encoder/encoder.h"
#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
43

44
#define RD_THRESH_POW 1.25
Jingning Han's avatar
Jingning Han committed
45
46
47
48
49
50
51
52
53

// 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).
static const uint8_t rd_thresh_block_size_factor[BLOCK_SIZES] = {
Jingning Han's avatar
Jingning Han committed
54
55
56
#if CONFIG_CB4X4
  2,  2,  2,
#endif
57
  2,  3,  3, 4, 6, 6, 8, 12, 12, 16, 24, 24, 32,
58
59
60
#if CONFIG_EXT_PARTITION
  48, 48, 64
#endif  // CONFIG_EXT_PARTITION
Jingning Han's avatar
Jingning Han committed
61
62
};

Yaowu Xu's avatar
Yaowu Xu committed
63
static void fill_mode_costs(AV1_COMP *cpi) {
Jingning Han's avatar
Jingning Han committed
64
65
66
67
68
  const FRAME_CONTEXT *const fc = cpi->common.fc;
  int i, j;

  for (i = 0; i < INTRA_MODES; ++i)
    for (j = 0; j < INTRA_MODES; ++j)
Yaowu Xu's avatar
Yaowu Xu committed
69
70
      av1_cost_tokens(cpi->y_mode_costs[i][j], av1_kf_y_mode_prob[i][j],
                      av1_intra_mode_tree);
Jingning Han's avatar
Jingning Han committed
71

72
  for (i = 0; i < BLOCK_SIZE_GROUPS; ++i)
Yaowu Xu's avatar
Yaowu Xu committed
73
74
    av1_cost_tokens(cpi->mbmode_cost[i], fc->y_mode_prob[i],
                    av1_intra_mode_tree);
75

76
  for (i = 0; i < INTRA_MODES; ++i)
Yaowu Xu's avatar
Yaowu Xu committed
77
78
    av1_cost_tokens(cpi->intra_uv_mode_cost[i], fc->uv_mode_prob[i],
                    av1_intra_mode_tree);
Jingning Han's avatar
Jingning Han committed
79
80

  for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
Yaowu Xu's avatar
Yaowu Xu committed
81
82
    av1_cost_tokens(cpi->switchable_interp_costs[i],
                    fc->switchable_interp_prob[i], av1_switchable_interp_tree);
hui su's avatar
hui su committed
83

84
#if CONFIG_PALETTE
hui su's avatar
hui su committed
85
  for (i = 0; i < PALETTE_BLOCK_SIZES; ++i) {
Yaowu Xu's avatar
Yaowu Xu committed
86
87
88
89
    av1_cost_tokens(cpi->palette_y_size_cost[i],
                    av1_default_palette_y_size_prob[i], av1_palette_size_tree);
    av1_cost_tokens(cpi->palette_uv_size_cost[i],
                    av1_default_palette_uv_size_prob[i], av1_palette_size_tree);
hui su's avatar
hui su committed
90
91
  }

92
  for (i = 0; i < PALETTE_SIZES; ++i) {
93
    for (j = 0; j < PALETTE_COLOR_INDEX_CONTEXTS; ++j) {
Yaowu Xu's avatar
Yaowu Xu committed
94
      av1_cost_tokens(cpi->palette_y_color_cost[i][j],
95
96
                      av1_default_palette_y_color_index_prob[i][j],
                      av1_palette_color_index_tree[i]);
Yaowu Xu's avatar
Yaowu Xu committed
97
      av1_cost_tokens(cpi->palette_uv_color_cost[i][j],
98
99
                      av1_default_palette_uv_color_index_prob[i][j],
                      av1_palette_color_index_tree[i]);
hui su's avatar
hui su committed
100
    }
101
102
  }
#endif  // CONFIG_PALETTE
103

Jingning Han's avatar
Jingning Han committed
104
  for (i = 0; i < MAX_TX_DEPTH; ++i)
105
    for (j = 0; j < TX_SIZE_CONTEXTS; ++j)
Yaowu Xu's avatar
Yaowu Xu committed
106
107
      av1_cost_tokens(cpi->tx_size_cost[i][j], fc->tx_size_probs[i][j],
                      av1_tx_size_tree[i]);
108

109
#if CONFIG_EXT_TX
110
111
112
  for (i = TX_4X4; i < EXT_TX_SIZES; ++i) {
    int s;
    for (s = 1; s < EXT_TX_SETS_INTER; ++s) {
113
      if (use_inter_ext_tx_for_txsize[s][i]) {
Yaowu Xu's avatar
Yaowu Xu committed
114
115
        av1_cost_tokens(cpi->inter_tx_type_costs[s][i],
                        fc->inter_ext_tx_prob[s][i], av1_ext_tx_inter_tree[s]);
116
117
118
      }
    }
    for (s = 1; s < EXT_TX_SETS_INTRA; ++s) {
119
      if (use_intra_ext_tx_for_txsize[s][i]) {
120
        for (j = 0; j < INTRA_MODES; ++j)
Yaowu Xu's avatar
Yaowu Xu committed
121
122
123
          av1_cost_tokens(cpi->intra_tx_type_costs[s][i][j],
                          fc->intra_ext_tx_prob[s][i][j],
                          av1_ext_tx_intra_tree[s]);
124
125
      }
    }
126
  }
127
#else
128
129
  for (i = TX_4X4; i < EXT_TX_SIZES; ++i) {
    for (j = 0; j < TX_TYPES; ++j)
Yaowu Xu's avatar
Yaowu Xu committed
130
131
      av1_cost_tokens(cpi->intra_tx_type_costs[i][j],
                      fc->intra_ext_tx_prob[i][j], av1_ext_tx_tree);
132
133
  }
  for (i = TX_4X4; i < EXT_TX_SIZES; ++i) {
Yaowu Xu's avatar
Yaowu Xu committed
134
135
    av1_cost_tokens(cpi->inter_tx_type_costs[i], fc->inter_ext_tx_prob[i],
                    av1_ext_tx_tree);
136
  }
137
#endif  // CONFIG_EXT_TX
138
#if CONFIG_EXT_INTRA
hui su's avatar
hui su committed
139
#if CONFIG_INTRA_INTERP
140
  for (i = 0; i < INTRA_FILTERS + 1; ++i)
Yaowu Xu's avatar
Yaowu Xu committed
141
142
    av1_cost_tokens(cpi->intra_filter_cost[i], fc->intra_filter_probs[i],
                    av1_intra_filter_tree);
hui su's avatar
hui su committed
143
#endif  // CONFIG_INTRA_INTERP
144
#endif  // CONFIG_EXT_INTRA
145
146
147
148
#if CONFIG_LOOP_RESTORATION
  av1_cost_tokens(cpi->switchable_restore_cost, fc->switchable_restore_prob,
                  av1_switchable_restore_tree);
#endif  // CONFIG_LOOP_RESTORATION
149
150
151
152
#if CONFIG_GLOBAL_MOTION
  av1_cost_tokens(cpi->gmtype_cost, fc->global_motion_types_prob,
                  av1_global_motion_types_tree);
#endif  // CONFIG_GLOBAL_MOTION
Jingning Han's avatar
Jingning Han committed
153
154
}

Yaowu Xu's avatar
Yaowu Xu committed
155
156
void av1_fill_token_costs(av1_coeff_cost *c,
                          av1_coeff_probs_model (*p)[PLANE_TYPES]) {
Jingning Han's avatar
Jingning Han committed
157
158
  int i, j, k, l;
  TX_SIZE t;
159
  for (t = 0; t < TX_SIZES; ++t)
Jingning Han's avatar
Jingning Han committed
160
161
162
163
    for (i = 0; i < PLANE_TYPES; ++i)
      for (j = 0; j < REF_TYPES; ++j)
        for (k = 0; k < COEF_BANDS; ++k)
          for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l) {
Yaowu Xu's avatar
Yaowu Xu committed
164
165
166
167
168
            aom_prob probs[ENTROPY_NODES];
            av1_model_to_full_probs(p[t][i][j][k][l], probs);
            av1_cost_tokens((int *)c[t][i][j][k][0][l], probs, av1_coef_tree);
            av1_cost_tokens_skip((int *)c[t][i][j][k][1][l], probs,
                                 av1_coef_tree);
Jingning Han's avatar
Jingning Han committed
169
170
171
172
173
174
175
176
177
            assert(c[t][i][j][k][0][l][EOB_TOKEN] ==
                   c[t][i][j][k][1][l][EOB_TOKEN]);
          }
}

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

178
#if CONFIG_HIGHBITDEPTH
Jingning Han's avatar
Jingning Han committed
179
180
181
182
183
184
185
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
186
                            aom_bit_depth_t bit_depth) {
Jingning Han's avatar
Jingning Han committed
187
188
189
190
191
  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
192
    const double q = av1_convert_qindex_to_q(i, bit_depth);
Jingning Han's avatar
Jingning Han committed
193
194
195
196
197
    bit16lut[i] = (int)(0.0418 * q + 2.4107);
    bit4lut[i] = (int)(0.063 * q + 2.742);
  }
}

Yaowu Xu's avatar
Yaowu Xu committed
198
void av1_init_me_luts(void) {
Jingning Han's avatar
Jingning Han committed
199
  init_me_luts_bd(sad_per_bit16lut_8, sad_per_bit4lut_8, QINDEX_RANGE,
Yaowu Xu's avatar
Yaowu Xu committed
200
                  AOM_BITS_8);
201
#if CONFIG_HIGHBITDEPTH
Jingning Han's avatar
Jingning Han committed
202
  init_me_luts_bd(sad_per_bit16lut_10, sad_per_bit4lut_10, QINDEX_RANGE,
Yaowu Xu's avatar
Yaowu Xu committed
203
                  AOM_BITS_10);
Jingning Han's avatar
Jingning Han committed
204
  init_me_luts_bd(sad_per_bit16lut_12, sad_per_bit4lut_12, QINDEX_RANGE,
Yaowu Xu's avatar
Yaowu Xu committed
205
                  AOM_BITS_12);
Jingning Han's avatar
Jingning Han committed
206
207
208
#endif
}

209
210
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
211
static const int rd_frame_type_factor[FRAME_UPDATE_TYPES] = {
212
  128, 144, 128, 128, 144,
213
#if CONFIG_EXT_REFS
214
215
  // TODO(zoeliu): To adjust further following factor values.
  128, 128, 128
216
217
  // TODO(weitinglin): We should investigate if the values should be the same
  //                   as the value used by OVERLAY frame
Yaowu Xu's avatar
Yaowu Xu committed
218
219
  ,
  144
220
#endif  // CONFIG_EXT_REFS
Jingning Han's avatar
Jingning Han committed
221
222
};

Yaowu Xu's avatar
Yaowu Xu committed
223
224
int av1_compute_rd_mult(const AV1_COMP *cpi, int qindex) {
  const int64_t q = av1_dc_quant(qindex, 0, cpi->common.bit_depth);
225
#if CONFIG_HIGHBITDEPTH
Jingning Han's avatar
Jingning Han committed
226
227
  int64_t rdmult = 0;
  switch (cpi->common.bit_depth) {
Yaowu Xu's avatar
Yaowu Xu committed
228
229
230
    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
231
    default:
Yaowu Xu's avatar
Yaowu Xu committed
232
      assert(0 && "bit_depth should be AOM_BITS_8, AOM_BITS_10 or AOM_BITS_12");
Jingning Han's avatar
Jingning Han committed
233
234
235
236
      return -1;
  }
#else
  int64_t rdmult = 88 * q * q / 24;
237
#endif  // CONFIG_HIGHBITDEPTH
Jingning Han's avatar
Jingning Han committed
238
239
240
  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
241
    const int boost_index = AOMMIN(15, (cpi->rc.gfu_boost / 100));
Jingning Han's avatar
Jingning Han committed
242
243
244
245

    rdmult = (rdmult * rd_frame_type_factor[frame_type]) >> 7;
    rdmult += ((rdmult * rd_boost_factor[boost_index]) >> 7);
  }
246
  if (rdmult < 1) rdmult = 1;
Jingning Han's avatar
Jingning Han committed
247
248
249
  return (int)rdmult;
}

Yaowu Xu's avatar
Yaowu Xu committed
250
static int compute_rd_thresh_factor(int qindex, aom_bit_depth_t bit_depth) {
Jingning Han's avatar
Jingning Han committed
251
  double q;
252
#if CONFIG_HIGHBITDEPTH
Jingning Han's avatar
Jingning Han committed
253
  switch (bit_depth) {
Yaowu Xu's avatar
Yaowu Xu committed
254
255
256
    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
257
    default:
Yaowu Xu's avatar
Yaowu Xu committed
258
      assert(0 && "bit_depth should be AOM_BITS_8, AOM_BITS_10 or AOM_BITS_12");
Jingning Han's avatar
Jingning Han committed
259
260
261
      return -1;
  }
#else
262
  (void)bit_depth;
Yaowu Xu's avatar
Yaowu Xu committed
263
  q = av1_dc_quant(qindex, 0, AOM_BITS_8) / 4.0;
264
#endif  // CONFIG_HIGHBITDEPTH
Jingning Han's avatar
Jingning Han committed
265
  // TODO(debargha): Adjust the function below.
Yaowu Xu's avatar
Yaowu Xu committed
266
  return AOMMAX((int)(pow(q, RD_THRESH_POW) * 5.12), 8);
Jingning Han's avatar
Jingning Han committed
267
268
}

Yaowu Xu's avatar
Yaowu Xu committed
269
void av1_initialize_me_consts(const AV1_COMP *cpi, MACROBLOCK *x, int qindex) {
270
#if CONFIG_HIGHBITDEPTH
Jingning Han's avatar
Jingning Han committed
271
  switch (cpi->common.bit_depth) {
Yaowu Xu's avatar
Yaowu Xu committed
272
    case AOM_BITS_8:
Jingning Han's avatar
Jingning Han committed
273
274
275
      x->sadperbit16 = sad_per_bit16lut_8[qindex];
      x->sadperbit4 = sad_per_bit4lut_8[qindex];
      break;
Yaowu Xu's avatar
Yaowu Xu committed
276
    case AOM_BITS_10:
Jingning Han's avatar
Jingning Han committed
277
278
279
      x->sadperbit16 = sad_per_bit16lut_10[qindex];
      x->sadperbit4 = sad_per_bit4lut_10[qindex];
      break;
Yaowu Xu's avatar
Yaowu Xu committed
280
    case AOM_BITS_12:
Jingning Han's avatar
Jingning Han committed
281
282
283
284
      x->sadperbit16 = sad_per_bit16lut_12[qindex];
      x->sadperbit4 = sad_per_bit4lut_12[qindex];
      break;
    default:
Yaowu Xu's avatar
Yaowu Xu committed
285
      assert(0 && "bit_depth should be AOM_BITS_8, AOM_BITS_10 or AOM_BITS_12");
Jingning Han's avatar
Jingning Han committed
286
287
288
289
290
  }
#else
  (void)cpi;
  x->sadperbit16 = sad_per_bit16lut_8[qindex];
  x->sadperbit4 = sad_per_bit4lut_8[qindex];
291
#endif  // CONFIG_HIGHBITDEPTH
Jingning Han's avatar
Jingning Han committed
292
293
}

Yaowu Xu's avatar
Yaowu Xu committed
294
static void set_block_thresholds(const AV1_COMMON *cm, RD_OPT *rd) {
Jingning Han's avatar
Jingning Han committed
295
296
297
298
  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
299
        clamp(av1_get_qindex(&cm->seg, segment_id, cm->base_qindex) +
300
301
                  cm->y_dc_delta_q,
              0, MAXQ);
Jingning Han's avatar
Jingning Han committed
302
303
304
305
306
307
308
309
    const int q = compute_rd_thresh_factor(qindex, cm->bit_depth);

    for (bsize = 0; bsize < BLOCK_SIZES; ++bsize) {
      // 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;

310
311
312
313
314
315
#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
316
317
      if (bsize >= BLOCK_8X8) {
        for (i = 0; i < MAX_MODES; ++i)
318
319
320
          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
321
322
323
324
325
326
327
      } 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;
      }
328
#endif
Jingning Han's avatar
Jingning Han committed
329
330
331
332
    }
  }
}

Yaowu Xu's avatar
Yaowu Xu committed
333
334
void av1_set_mvcost(MACROBLOCK *x, MV_REFERENCE_FRAME ref_frame, int ref,
                    int ref_mv_idx) {
335
  MB_MODE_INFO_EXT *mbmi_ext = x->mbmi_ext;
Yaowu Xu's avatar
Yaowu Xu committed
336
337
338
339
  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;
340
341
342
343
  x->mvcost = x->mv_cost_stack[nmv_ctx];
  x->nmvjointcost = x->nmv_vec_cost[nmv_ctx];
}

Yaowu Xu's avatar
Yaowu Xu committed
344
345
void av1_initialize_rd_consts(AV1_COMP *cpi) {
  AV1_COMMON *const cm = &cpi->common;
Jingning Han's avatar
Jingning Han committed
346
347
348
  MACROBLOCK *const x = &cpi->td.mb;
  RD_OPT *const rd = &cpi->rd;
  int i;
Jingning Han's avatar
Jingning Han committed
349
  int nmv_ctx;
Jingning Han's avatar
Jingning Han committed
350

Yaowu Xu's avatar
Yaowu Xu committed
351
  aom_clear_system_state();
Jingning Han's avatar
Jingning Han committed
352
353

  rd->RDDIV = RDDIV_BITS;  // In bits (to multiply D by 128).
Yaowu Xu's avatar
Yaowu Xu committed
354
  rd->RDMULT = av1_compute_rd_mult(cpi, cm->base_qindex + cm->y_dc_delta_q);
Jingning Han's avatar
Jingning Han committed
355

356
  set_error_per_bit(x, rd->RDMULT);
Jingning Han's avatar
Jingning Han committed
357
358
359

  set_block_thresholds(cm, rd);

Jingning Han's avatar
Jingning Han committed
360
  for (nmv_ctx = 0; nmv_ctx < NMV_CONTEXTS; ++nmv_ctx) {
Yaowu Xu's avatar
Yaowu Xu committed
361
    av1_build_nmv_cost_table(
Jingning Han's avatar
Jingning Han committed
362
363
364
365
        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);
366
  }
Jingning Han's avatar
Jingning Han committed
367
368
369
  x->mvcost = x->mv_cost_stack[0];
  x->nmvjointcost = x->nmv_vec_cost[0];

370
  if (cpi->oxcf.pass != 1) {
371
    av1_fill_token_costs(x->token_costs, cm->fc->coef_probs);
372

Thomas Daede's avatar
Thomas Daede committed
373
    if (cm->frame_type == KEY_FRAME) {
374
#if CONFIG_EXT_PARTITION_TYPES
375
      for (i = 0; i < PARTITION_PLOFFSET; ++i)
376
377
        av1_cost_tokens(cpi->partition_cost[i], cm->fc->partition_prob[i],
                        av1_partition_tree);
378
      for (; i < PARTITION_CONTEXTS_PRIMARY; ++i)
379
380
        av1_cost_tokens(cpi->partition_cost[i], cm->fc->partition_prob[i],
                        av1_ext_partition_tree);
381
#else
382
      for (i = 0; i < PARTITION_CONTEXTS_PRIMARY; ++i)
383
384
        av1_cost_tokens(cpi->partition_cost[i], cm->fc->partition_prob[i],
                        av1_partition_tree);
385
#endif  // CONFIG_EXT_PARTITION_TYPES
386
387
388
389
#if CONFIG_UNPOISON_PARTITION_CTX
      for (; i < PARTITION_CONTEXTS_PRIMARY + PARTITION_BLOCK_SIZES; ++i) {
        aom_prob p = cm->fc->partition_prob[i][PARTITION_VERT];
        assert(p > 0);
390
391
392
393
        cpi->partition_cost[i][PARTITION_NONE] = INT_MAX;
        cpi->partition_cost[i][PARTITION_HORZ] = INT_MAX;
        cpi->partition_cost[i][PARTITION_VERT] = av1_cost_bit(p, 0);
        cpi->partition_cost[i][PARTITION_SPLIT] = av1_cost_bit(p, 1);
394
395
396
397
      }
      for (; i < PARTITION_CONTEXTS_PRIMARY + 2 * PARTITION_BLOCK_SIZES; ++i) {
        aom_prob p = cm->fc->partition_prob[i][PARTITION_HORZ];
        assert(p > 0);
398
399
400
401
        cpi->partition_cost[i][PARTITION_NONE] = INT_MAX;
        cpi->partition_cost[i][PARTITION_HORZ] = av1_cost_bit(p, 0);
        cpi->partition_cost[i][PARTITION_VERT] = INT_MAX;
        cpi->partition_cost[i][PARTITION_SPLIT] = av1_cost_bit(p, 1);
402
      }
403
404
405
406
407
      cpi->partition_cost[PARTITION_CONTEXTS][PARTITION_NONE] = INT_MAX;
      cpi->partition_cost[PARTITION_CONTEXTS][PARTITION_HORZ] = INT_MAX;
      cpi->partition_cost[PARTITION_CONTEXTS][PARTITION_VERT] = INT_MAX;
      cpi->partition_cost[PARTITION_CONTEXTS][PARTITION_SPLIT] = 0;
#endif  // CONFIG_UNPOISON_PARTITION_CTX
408
409
410
    }

    fill_mode_costs(cpi);
411

412
413
    if (!frame_is_intra_only(cm)) {
      for (i = 0; i < NEWMV_MODE_CONTEXTS; ++i) {
Yaowu Xu's avatar
Yaowu Xu committed
414
415
        cpi->newmv_mode_cost[i][0] = av1_cost_bit(cm->fc->newmv_prob[i], 0);
        cpi->newmv_mode_cost[i][1] = av1_cost_bit(cm->fc->newmv_prob[i], 1);
416
      }
417

418
      for (i = 0; i < ZEROMV_MODE_CONTEXTS; ++i) {
Yaowu Xu's avatar
Yaowu Xu committed
419
420
        cpi->zeromv_mode_cost[i][0] = av1_cost_bit(cm->fc->zeromv_prob[i], 0);
        cpi->zeromv_mode_cost[i][1] = av1_cost_bit(cm->fc->zeromv_prob[i], 1);
421
      }
Jingning Han's avatar
Jingning Han committed
422

423
      for (i = 0; i < REFMV_MODE_CONTEXTS; ++i) {
Yaowu Xu's avatar
Yaowu Xu committed
424
425
        cpi->refmv_mode_cost[i][0] = av1_cost_bit(cm->fc->refmv_prob[i], 0);
        cpi->refmv_mode_cost[i][1] = av1_cost_bit(cm->fc->refmv_prob[i], 1);
426
      }
427

428
      for (i = 0; i < DRL_MODE_CONTEXTS; ++i) {
Yaowu Xu's avatar
Yaowu Xu committed
429
430
        cpi->drl_mode_cost0[i][0] = av1_cost_bit(cm->fc->drl_prob[i], 0);
        cpi->drl_mode_cost0[i][1] = av1_cost_bit(cm->fc->drl_prob[i], 1);
431
      }
432
#if CONFIG_EXT_INTER
433
      for (i = 0; i < INTER_MODE_CONTEXTS; ++i)
Yaowu Xu's avatar
Yaowu Xu committed
434
435
436
        av1_cost_tokens((int *)cpi->inter_compound_mode_cost[i],
                        cm->fc->inter_compound_mode_probs[i],
                        av1_inter_compound_mode_tree);
437
#if CONFIG_INTERINTRA
438
      for (i = 0; i < BLOCK_SIZE_GROUPS; ++i)
Yaowu Xu's avatar
Yaowu Xu committed
439
440
441
        av1_cost_tokens((int *)cpi->interintra_mode_cost[i],
                        cm->fc->interintra_mode_prob[i],
                        av1_interintra_mode_tree);
442
#endif  // CONFIG_INTERINTRA
443
#endif  // CONFIG_EXT_INTER
Yue Chen's avatar
Yue Chen committed
444
#if CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
445
      for (i = BLOCK_8X8; i < BLOCK_SIZES; i++) {
Yue Chen's avatar
Yue Chen committed
446
447
        av1_cost_tokens((int *)cpi->motion_mode_cost[i],
                        cm->fc->motion_mode_prob[i], av1_motion_mode_tree);
448
      }
Yue Chen's avatar
Yue Chen committed
449
450
451
452
453
454
#if CONFIG_MOTION_VAR && CONFIG_WARPED_MOTION
      for (i = BLOCK_8X8; i < BLOCK_SIZES; i++) {
        cpi->motion_mode_cost1[i][0] = av1_cost_bit(cm->fc->obmc_prob[i], 0);
        cpi->motion_mode_cost1[i][1] = av1_cost_bit(cm->fc->obmc_prob[i], 1);
      }
#endif  // CONFIG_MOTION_VAR && CONFIG_WARPED_MOTION
Yue Chen's avatar
Yue Chen committed
455
#endif  // CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
456
    }
Jingning Han's avatar
Jingning Han committed
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
  }
}

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[] = {
474
475
476
477
478
479
480
481
482
    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
483
484
485
486
487
488
489
490
491
  };
  // 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[] = {
492
493
494
495
496
497
498
499
500
    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
501
502
  };
  static const int xsq_iq_q10[] = {
503
504
505
506
507
508
509
510
511
512
513
514
    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
515
516
517
518
519
520
521
522
523
524
525
  };
  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
526
527
528
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
529
530
531
532
533
534
535
536
537
538
539
540
541
542
  // 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
543
    const int xsq_q10 = (int)AOMMIN(xsq_q10_64, MAX_XSQ_Q10);
Jingning Han's avatar
Jingning Han committed
544
    model_rd_norm(xsq_q10, &r_q10, &d_q10);
Yaowu Xu's avatar
Yaowu Xu committed
545
    *rate = ROUND_POWER_OF_TWO(r_q10 << n_log2, 10 - AV1_PROB_COST_SHIFT);
Jingning Han's avatar
Jingning Han committed
546
547
548
549
    *dist = (var * (int64_t)d_q10 + 512) >> 10;
  }
}

550
551
552
553
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]) {
554
555
  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
556
557
558
  const ENTROPY_CONTEXT *const above = pd->above_context;
  const ENTROPY_CONTEXT *const left = pd->left_context;

559
560
561
562
563
564
#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
565
  int i;
566

567
#if CONFIG_CHROMA_2X2
568
  switch (tx_size) {
569
    case TX_2X2:
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
      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;
599
600
601
602
603
604
605
606
607
608
609
610
611
612
#if CONFIG_TX64X64
    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
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
    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];
650
      break;
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
#if CONFIG_EXT_TX && CONFIG_RECT_TX && CONFIG_RECT_TX_EXT
    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;
#endif  // CONFIG_EXT_TX && CONFIG_RECT_TX && CONFIG_RECT_TX_EXT
679
680
681
682

    default: assert(0 && "Invalid transform size."); break;
  }
  return;
683
#endif  // CONFIG_CHROMA_2X2
684
685

  switch (tx_size) {
Jingning Han's avatar
Jingning Han committed
686
687
688
689
    case TX_4X4:
      memcpy(t_above, above, sizeof(ENTROPY_CONTEXT) * num_4x4_w);
      memcpy(t_left, left, sizeof(ENTROPY_CONTEXT) * num_4x4_h);
      break;
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
    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;
708
709
710
711
712
713
#if CONFIG_TX64X64
    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)
714
715
        t_left[i] =
            !!(*(const uint64_t *)&left[i] | *(const uint64_t *)&left[i + 8]);
716
717
      break;
#endif  // CONFIG_TX64X64
718
719
720
721
722
723
724
725
726
727
    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;
728
    case TX_8X16:
Jingning Han's avatar
Jingning Han committed
729
730
      for (i = 0; i < num_4x4_w; i += 2)
        t_above[i] = !!*(const uint16_t *)&above[i];
731
732
733
734
735
736
      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
737
738
739
      for (i = 0; i < num_4x4_h; i += 2)
        t_left[i] = !!*(const uint16_t *)&left[i];
      break;
740
    case TX_16X32:
Jingning Han's avatar
Jingning Han committed
741
742
      for (i = 0; i < num_4x4_w; i += 4)
        t_above[i] = !!*(const uint32_t *)&above[i];
743
744
      for (i = 0; i < num_4x4_h; i += 8)
        t_left[i] = !!*(const uint64_t *)&left[i];
Jingning Han's avatar
Jingning Han committed
745
      break;
746
    case TX_32X16:
Jingning Han's avatar
Jingning Han committed
747
748
      for (i = 0; i < num_4x4_w; i += 8)
        t_above[i] = !!*(const uint64_t *)&above[i];
749
750
751
      for (i = 0; i < num_4x4_h; i += 4)
        t_left[i] = !!*(const uint32_t *)&left[i];
      break;
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
#if CONFIG_EXT_TX && CONFIG_RECT_TX && CONFIG_RECT_TX_EXT
    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;
#endif  // CONFIG_EXT_TX && CONFIG_RECT_TX && CONFIG_RECT_TX_EXT
clang-format's avatar
clang-format committed
776
    default: assert(0 && "Invalid transform size."); break;
Jingning Han's avatar
Jingning Han committed
777
778
779
  }
}

Yaowu Xu's avatar
Yaowu Xu committed
780
781
782
783
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]) {
784
785
786
787
#if CONFIG_CB4X4 && !CONFIG_CHROMA_2X2
  const BLOCK_SIZE plane_bsize =
      AOMMAX(BLOCK_4X4, get_plane_block_size(bsize, pd));
#else
788
  const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, pd);
789
#endif
790
  get_entropy_contexts_plane(plane_bsize, tx_size, pd, t_above, t_left);
791
792
}

793
void av1_mv_pred(const AV1_COMP *cpi, MACROBLOCK *x, uint8_t *ref_y_buffer,
Yaowu Xu's avatar
Yaowu Xu committed
794
                 int ref_y_stride, int ref_frame, BLOCK_SIZE block_size) {
Jingning Han's avatar
Jingning Han committed
795
796
797
798
799
800
801
802
  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;
803
  MV pred_mv[MAX_MV_REF_CANDIDATES + 1];
804
805
806
807
808
809
  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;
810
  }
811
812
813
  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
814
  assert(num_mv_refs <= (int)(sizeof(pred_mv) / sizeof(pred_mv[0])));
815

Jingning Han's avatar
Jingning Han committed
816
817
818
819
820
821
  // 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
822
    max_mv = AOMMAX(max_mv, AOMMAX(abs(this_mv->row), abs(this_mv->col)) >> 3);
Jingning Han's avatar
Jingning Han committed
823

824
825
    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
826

827
    ref_y_ptr = &ref_y_buffer[ref_y_stride * fp_row + fp_col];
Jingning Han's avatar
Jingning Han committed
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
    // 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;
}

Yaowu Xu's avatar
Yaowu Xu committed
844
845
846
847
848
void av1_setup_pred_block(const MACROBLOCKD *xd,
                          struct buf_2d dst[MAX_MB_PLANE],
                          const YV12_BUFFER_CONFIG *src, int mi_row, int mi_col,
                          const struct scale_factors *scale,
                          const struct scale_factors *scale_uv) {
Jingning Han's avatar
Jingning Han committed
849
850
851
852
853
854
855
856
857
  int i;

  dst[0].buf = src->y_buffer;
  dst[0].stride = src->y_stride;
  dst[1].buf = src->u_buffer;
  dst[2].buf = src->v_buffer;
  dst[1].stride = dst[2].stride = src->uv_stride;

  for (i = 0; i < MAX_MB_PLANE; ++i) {
858
    setup_pred_plane(dst + i, xd->mi[0]->mbmi.sb_type, dst[i].buf,
859
860
                     i ? src->uv_crop_width : src->y_crop_width,
                     i ? src->uv_crop_height : src->y_crop_height,
861
                     dst[i].stride, mi_row, mi_col, i ? scale_uv : scale,
Jingning Han's avatar
Jingning Han committed
862
863
864
865
                     xd->plane[i].subsampling_x, xd->plane[i].subsampling_y);
  }
}

Yaowu Xu's avatar
Yaowu Xu committed
866
867
int av1_raster_block_offset(BLOCK_SIZE plane_bsize, int raster_block,
                            int stride) {
Jingning Han's avatar
Jingning Han committed
868
869
870
871
872
873
  const int bw = b_width_log2_lookup[plane_bsize];
  const int y = 4 * (raster_block >> bw);
  const int x = 4 * (raster_block & ((1 << bw) - 1));
  return y * stride + x;
}

Yaowu Xu's avatar
Yaowu Xu committed
874
875
int16_t *av1_raster_block_offset_int16(BLOCK_SIZE plane_bsize, int raster_block,
                                       int16_t *base) {
876
  const int stride = block_size_wide[plane_bsize];
Yaowu Xu's avatar
Yaowu Xu committed
877
  return base + av1_raster_block_offset(plane_bsize, raster_block, stride);
Jingning Han's avatar
Jingning Han committed
878
879
}

Yaowu Xu's avatar
Yaowu Xu committed
880
881
882
YV12_BUFFER_CONFIG *av1_get_scaled_ref_frame(const AV1_COMP *cpi,
                                             int ref_frame) {
  const AV1_COMMON *const cm = &cpi->common;
Jingning Han's avatar
Jingning Han committed
883
884
  const int scaled_idx = cpi->scaled_ref_idx[ref_frame - 1];
  const int ref_idx = get_ref_frame_buf_idx(cpi, ref_frame);
885
886
887
  return (scaled_idx != ref_idx && scaled_idx != INVALID_IDX)
             ? &cm->buffer_pool->frame_bufs[scaled_idx].buf
             : NULL;
Jingning Han's avatar
Jingning Han committed
888
889
}

890
#if CONFIG_DUAL_FILTER
891
int av1_get_switchable_rate(const AV1_COMP *cpi, const MACROBLOCKD *xd) {
892
893
894
895
896
897
898
899
900
901
902
903
904
905
  const AV1_COMMON *const cm = &cpi->common;
  if (cm->interp_filter == SWITCHABLE) {
    const MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
    int inter_filter_cost = 0;
    int dir;

    for (dir = 0; dir < 2; ++dir) {
      if (has_subpel_mv_component(xd->mi[0], xd, dir) ||
          (mbmi->ref_frame[1] > INTRA_FRAME &&
           has_subpel_mv_component(xd->mi[0], xd, dir + 2))) {
        const int ctx = av1_get_pred_context_switchable_interp(xd, dir);
        inter_filter_cost +=
            cpi->switchable_interp_costs[ctx][mbmi->interp_filter[dir]];
      }
906
    }
907
908
909
    return SWITCHABLE_INTERP_RATE_FACTOR * inter_filter_cost;
  } else {
    return 0;
910
911
912
  }
}
#else
913
int av1_get_switchable_rate(const AV1_COMP *cpi, const MACROBLOCKD *xd) {
914
915
  const AV1_COMMON *const cm = &cpi->common;
  if (cm->interp_filter == SWITCHABLE) {
916
917
918
919
    const MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
    const int ctx = av1_get_pred_context_switchable_interp(xd);
    return SWITCHABLE_INTERP_RATE_FACTOR *
           cpi->switchable_interp_costs[ctx][mbmi->interp_filter];
920
921
  }
  return 0;
Jingning Han's avatar
Jingning Han committed
922
}
923
#endif
Jingning Han's avatar
Jingning Han committed
924

Yaowu Xu's avatar
Yaowu Xu committed
925
void av1_set_rd_speed_thresholds(AV1_COMP *cpi) {
Jingning Han's avatar
Jingning Han committed
926
927
928
929
930
  int i;
  RD_OPT *const rd = &cpi->rd;
  SPEED_FEATURES *const sf = &cpi->sf;

  // Set baseline threshold values.
931
  for (i = 0; i < MAX_MODES; ++i) rd->thresh_mult[i] = cpi->oxcf.mode == 0;
Jingning Han's avatar
Jingning Han committed
932
933
934

  if (sf->adaptive_rd_thresh) {
    rd->thresh_mult[THR_NEARESTMV] = 300;
935
936
937
#if CONFIG_EXT_REFS
    rd->thresh_mult[THR_NEARESTL2] = 300;
    rd->thresh_mult[THR_NEARESTL3] = 300;
938
    rd->thresh_mult[THR_NEARESTB] = 300;
939
#endif  // CONFIG_EXT_REFS
Jingning Han's avatar
Jingning Han committed
940
    rd->thresh_mult[THR_NEARESTA] = 300;
941
    rd->thresh_mult[THR_NEARESTG] = 300;
Jingning Han's avatar
Jingning Han committed
942
943
  } else {
    rd->thresh_mult[THR_NEARESTMV] = 0;
944
945
946
#if CONFIG_EXT_REFS
    rd->thresh_mult[THR_NEARESTL2] = 0;
    rd->thresh_mult[THR_NEARESTL3] = 0;
947
    rd->thresh_mult[THR_NEARESTB] = 0;
948
#endif  // CONFIG_EXT_REFS
Jingning Han's avatar
Jingning Han committed
949
    rd->thresh_mult[THR_NEARESTA] = 0;
950
    rd->thresh_mult[THR_NEARESTG] = 0;
Jingning Han's avatar
Jingning Han committed
951
952
953
954
955
  }

  rd->thresh_mult[THR_DC] += 1000;

  rd->thresh_mult[THR_NEWMV] += 1000;
956
957
958
#if CONFIG_EXT_REFS
  rd->thresh_mult[THR_NEWL2] += 1000;
  rd->thresh_mult[THR_NEWL3] += 1000;
959
  rd->thresh_mult[THR_NEWB] += 1000;
960
#endif  // CONFIG_EXT_REFS
Jingning Han's avatar
Jingning Han committed
961
962
963
964
  rd->thresh_mult[THR_NEWA] += 1000;
  rd->thresh_mult[THR_NEWG] += 1000;

  rd->thresh_mult[THR_NEARMV] += 1000;
965
966
967
#if CONFIG_EXT_REFS
  rd->thresh_mult[THR_NEARL2] += 1000;
  rd->thresh_mult[THR_NEARL3] += 1000;
968
  rd->thresh_mult[THR_NEARB] += 1000;
969
#endif  // CONFIG_EXT_REFS
Jingning Han's avatar
Jingning Han committed
970
  rd->thresh_mult[THR_NEARA] += 1000;
971
972
973
974
975
976
  rd->thresh_mult[THR_NEARG] += 1000;

  rd->thresh_mult[THR_ZEROMV] += 2000;
#if CONFIG_EXT_REFS
  rd->thresh_mult[THR_ZEROL2] += 2000;
  rd->thresh_mult[THR_ZEROL3] += 2000;
977
  rd->thresh_mult[THR_ZEROB] += 2000;
978
979
980
#endif  // CONFIG_EXT_REFS
  rd->thresh_mult[THR_ZEROG] += 2000;
  rd->thresh_mult[THR_ZEROA] += 2000;
Jingning Han's avatar
Jingning Han committed
981
982
983

  rd->thresh_mult[THR_TM] += 1000;

984
#if CONFIG_EXT_INTER
985

986
  rd->thresh_mult[THR_COMP_NEAREST_NEARESTLA] += 1000;
987
988
989
990
#if CONFIG_EXT_REFS
  rd->thresh_mult[THR_COMP_NEAREST_NEARESTL2A] += 1000;
  rd->thresh_mult[THR_COMP_NEAREST_NEARESTL3A] += 1000;
#endif  // CONFIG_EXT_REFS
991
  rd->thresh_mult[THR_COMP_NEAREST_NEARESTGA] += 1000;
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
#if CONFIG_EXT_REFS
  rd->thresh_mult[THR_COMP_NEAREST_NEARESTLB] += 1000;
  rd->thresh_mult[THR_COMP_NEAREST_NEARESTL2B] += 1000;
  rd->thresh_mult[THR_COMP_NEAREST_NEARESTL3B] += 1000;
  rd->thresh_mult[THR_COMP_NEAREST_NEARESTGB] += 1000;
#endif  // CONFIG_EXT_REFS

#else  // CONFIG_EXT_INTER

  rd->thresh_mult[THR_COMP_NEARESTLA] += 1000;
#if CONFIG_EXT_REFS
  rd->thresh_mult[THR_COMP_NEARESTL2A] += 1000;
  rd->thresh_mult[THR_COMP_NEARESTL3A] += 1000;
#endif  // CONFIG_EXT_REFS
  rd->thresh_mult[THR_COMP_NEARESTGA] += 1000;
#if CONFIG_EXT_REFS
  rd->thresh_mult[THR_COMP_NEARESTLB] += 1000;
  rd->thresh_mult[THR_COMP_NEARESTL2B] += 1000;
  rd->thresh_mult[THR_COMP_NEARESTL3B] += 1000;
  rd->thresh_mult[THR_COMP_NEARESTGB] += 1000;
#endif  // CONFIG_EXT_REFS

#endif  // CONFIG_EXT_INTER

#if CONFIG_EXT_INTER

1018
  rd->thresh_mult[THR_COMP_NEAR_NEARLA] += 1200;
1019
1020
1021
1022
1023
1024
1025
1026
  rd->thresh_mult[THR_COMP_NEAREST_NEWLA] += 1500;
  rd->thresh_mult[THR_COMP_NEW_NEARESTLA] += 1500;
  rd->thresh_mult[THR_COMP_NEAR_NEWLA] += 1700;
  rd->thresh_mult[THR_COMP_NEW_NEARLA] += 1700;
  rd->thresh_mult[THR_COMP_NEW_NEWLA] += 2000;
  rd->thresh_mult[THR_COMP_ZERO_ZEROLA] += 2500;

#if CONFIG_EXT_REFS
1027
  rd->thresh_mult[THR_COMP_NEAR_NEARL2A] += 1200;
1028
1029
1030
1031
1032
1033
1034
  rd->thresh_mult[THR_COMP_NEAREST_NEWL2A] += 1500;
  rd->thresh_mult[THR_COMP_NEW_NEARESTL2A] += 1500;
  rd->thresh_mult[THR_COMP_NEAR_NEWL2A] += 1700;
  rd->thresh_mult[THR_COMP_NEW_NEARL2A] += 1700;
  rd->thresh_mult[THR_COMP_NEW_NEWL2A] += 2000;
  rd->thresh_mult[THR_COMP_ZERO_ZEROL2A] += 2500;

1035
  rd->thresh_mult[THR_COMP_NEAR_NEARL3A] += 1200;
1036
1037
1038
1039
1040
1041
  rd->thresh_mult[THR_COMP_NEAREST_NEWL3A] += 1500;
  rd->thresh_mult[THR_COMP_NEW_NEARESTL3A] += 1500;
  rd->thresh_mult[THR_COMP_NEAR_NEWL3A] += 1700;
  rd->thresh_mult[THR_COMP_NEW_NEARL3A] += 1700;
  rd->thresh_mult[THR_COMP_NEW_NEWL3A] += 2000;
  rd->thresh_mult[THR_COMP_ZERO_ZEROL3A] += 2500;
1042
#endif  // CONFIG_EXT_REFS
1043

1044
1045
1046
1047
1048
1049
1050
  rd->thresh_mult[THR_COMP_NEAR_NEARGA] += 1200;
  rd->thresh_mult[THR_COMP_NEAREST_NEWGA] += 1500;
  rd->thresh_mult[THR_COMP_NEW_NEARESTGA] += 1500;
  rd->thresh_mult[THR_COMP_NEAR_NEWGA] += 1700;
  rd->thresh_mult[THR_COMP_NEW_NEARGA] += 1700;
  rd->thresh_mult[THR_COMP_NEW_NEWGA] += 2000;
  rd->thresh_mult[THR_COMP_ZERO_ZEROGA] += 2500;
1051

1052
1053
1054
1055
1056
1057
1058
1059
#if CONFIG_EXT_REFS
  rd->thresh_mult[THR_COMP_NEAR_NEARLB] += 1200;
  rd->thresh_mult[THR_COMP_NEAREST_NEWLB] += 1500;
  rd->thresh_mult[THR_COMP_NEW_NEARESTLB] += 1500;
  rd->thresh_mult[THR_COMP_NEAR_NEWLB] += 1700;
  rd->thresh_mult[THR_COMP_NEW_NEARLB] += 1700;
  rd->thresh_mult[THR_COMP_NEW_NEWLB] += 2000;
  rd->thresh_mult[THR_COMP_ZERO_ZEROLB] += 2500;
1060

1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
  rd->thresh_mult[THR_COMP_NEAR_NEARL2B] += 1200;
  rd->thresh_mult[THR_COMP_NEAREST_NEWL2B] += 1500;
  rd->thresh_mult[THR_COMP_NEW_NEARESTL2B] += 1500;
  rd->thresh_mult[THR_COMP_NEAR_NEWL2B] += 1700;
  rd->thresh_mult[THR_COMP_NEW_NEARL2B] += 1700;
  rd->thresh_mult[THR_COMP_NEW_NEWL2B] += 2000;
  rd->thresh_mult[THR_COMP_ZERO_ZEROL2B] += 2500;

  rd->thresh_mult[THR_COMP_NEAR_NEARL3B] += 1200;
  rd->thresh_mult[THR_COMP_NEAREST_NEWL3B] += 1500;
  rd->thresh_mult[THR_COMP_NEW_NEARESTL3B] += 1500;
  rd->thresh_mult[THR_COMP_NEAR_NEWL3B] += 1700;
  rd->thresh_mult[THR_COMP_NEW_NEARL3B] += 1700;
  rd->thresh_mult[THR_COMP_NEW_NEWL3B] += 2000;
  rd->thresh_mult[THR_COMP_ZERO_ZEROL3B] += 2500;