rdopt.c 442 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
 */

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

Yaowu Xu's avatar
Yaowu Xu committed
15
#include "./aom_dsp_rtcd.h"
Jingning Han's avatar
Jingning Han committed
16
#include "./av1_rtcd.h"
Jingning Han's avatar
Jingning Han committed
17

Yaowu Xu's avatar
Yaowu Xu committed
18
#include "aom_dsp/aom_dsp_common.h"
19
#include "aom_dsp/blend.h"
Yaowu Xu's avatar
Yaowu Xu committed
20
#include "aom_mem/aom_mem.h"
21
22
#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
33
34
35
#include "av1/common/common.h"
#include "av1/common/common_data.h"
#include "av1/common/entropy.h"
#include "av1/common/entropymode.h"
#include "av1/common/idct.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/scan.h"
#include "av1/common/seg_common.h"
36
37
38
#if CONFIG_LV_MAP
#include "av1/common/txb_common.h"
#endif
Yue Chen's avatar
Yue Chen committed
39
40
41
#if CONFIG_WARPED_MOTION
#include "av1/common/warped_motion.h"
#endif  // CONFIG_WARPED_MOTION
Jingning Han's avatar
Jingning Han committed
42

Jingning Han's avatar
Jingning Han committed
43
#include "av1/encoder/aq_variance.h"
44
#include "av1/encoder/av1_quantize.h"
45
46
47
48
#include "av1/encoder/cost.h"
#include "av1/encoder/encodemb.h"
#include "av1/encoder/encodemv.h"
#include "av1/encoder/encoder.h"
49
50
51
#if CONFIG_LV_MAP
#include "av1/encoder/encodetxb.h"
#endif
52
53
#include "av1/encoder/hybrid_fwd_txfm.h"
#include "av1/encoder/mcomp.h"
54
#if CONFIG_PALETTE
55
#include "av1/encoder/palette.h"
56
#endif  // CONFIG_PALETTE
57
58
59
#include "av1/encoder/ratectrl.h"
#include "av1/encoder/rd.h"
#include "av1/encoder/rdopt.h"
60
#include "av1/encoder/tokenize.h"
61
62
#if CONFIG_PVQ
#include "av1/encoder/pvq_encoder.h"
63
#endif  // CONFIG_PVQ
Yushin Cho's avatar
Yushin Cho committed
64
65
#if CONFIG_PVQ || CONFIG_DAALA_DIST
#include "av1/common/pvq.h"
66
#endif  // CONFIG_PVQ || CONFIG_DAALA_DIST
67
#if CONFIG_DUAL_FILTER
Angie Chiang's avatar
Angie Chiang committed
68
#define DUAL_FILTER_SET_SIZE (SWITCHABLE_FILTERS * SWITCHABLE_FILTERS)
69
#if USE_EXTRA_FILTER
Angie Chiang's avatar
Angie Chiang committed
70
static const int filter_sets[DUAL_FILTER_SET_SIZE][2] = {
71
72
73
  { 0, 0 }, { 0, 1 }, { 0, 2 }, { 0, 3 }, { 1, 0 }, { 1, 1 },
  { 1, 2 }, { 1, 3 }, { 2, 0 }, { 2, 1 }, { 2, 2 }, { 2, 3 },
  { 3, 0 }, { 3, 1 }, { 3, 2 }, { 3, 3 },
74
};
75
76
77
78
79
80
#else   // USE_EXTRA_FILTER
static const int filter_sets[DUAL_FILTER_SET_SIZE][2] = {
  { 0, 0 }, { 0, 1 }, { 0, 2 }, { 1, 0 }, { 1, 1 },
  { 1, 2 }, { 2, 0 }, { 2, 1 }, { 2, 2 },
};
#endif  // USE_EXTRA_FILTER
Angie Chiang's avatar
Angie Chiang committed
81
#endif  // CONFIG_DUAL_FILTER
82

83
84
#if CONFIG_EXT_REFS

85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
#define LAST_FRAME_MODE_MASK                                      \
  ((1 << INTRA_FRAME) | (1 << LAST2_FRAME) | (1 << LAST3_FRAME) | \
   (1 << GOLDEN_FRAME) | (1 << BWDREF_FRAME) | (1 << ALTREF_FRAME))
#define LAST2_FRAME_MODE_MASK                                    \
  ((1 << INTRA_FRAME) | (1 << LAST_FRAME) | (1 << LAST3_FRAME) | \
   (1 << GOLDEN_FRAME) | (1 << BWDREF_FRAME) | (1 << ALTREF_FRAME))
#define LAST3_FRAME_MODE_MASK                                    \
  ((1 << INTRA_FRAME) | (1 << LAST_FRAME) | (1 << LAST2_FRAME) | \
   (1 << GOLDEN_FRAME) | (1 << BWDREF_FRAME) | (1 << ALTREF_FRAME))
#define GOLDEN_FRAME_MODE_MASK                                   \
  ((1 << INTRA_FRAME) | (1 << LAST_FRAME) | (1 << LAST2_FRAME) | \
   (1 << LAST3_FRAME) | (1 << BWDREF_FRAME) | (1 << ALTREF_FRAME))
#define BWDREF_FRAME_MODE_MASK                                   \
  ((1 << INTRA_FRAME) | (1 << LAST_FRAME) | (1 << LAST2_FRAME) | \
   (1 << LAST3_FRAME) | (1 << GOLDEN_FRAME) | (1 << ALTREF_FRAME))
#define ALTREF_FRAME_MODE_MASK                                   \
  ((1 << INTRA_FRAME) | (1 << LAST_FRAME) | (1 << LAST2_FRAME) | \
   (1 << LAST3_FRAME) | (1 << GOLDEN_FRAME) | (1 << BWDREF_FRAME))

#else

#define LAST_FRAME_MODE_MASK \
  ((1 << GOLDEN_FRAME) | (1 << ALTREF_FRAME) | (1 << INTRA_FRAME))
#define GOLDEN_FRAME_MODE_MASK \
  ((1 << LAST_FRAME) | (1 << ALTREF_FRAME) | (1 << INTRA_FRAME))
#define ALTREF_FRAME_MODE_MASK \
  ((1 << LAST_FRAME) | (1 << GOLDEN_FRAME) | (1 << INTRA_FRAME))
Jingning Han's avatar
Jingning Han committed
112

113
114
#endif  // CONFIG_EXT_REFS

115
#if CONFIG_EXT_REFS
Zoe Liu's avatar
Zoe Liu committed
116
117
118
119
120
#if CONFIG_EXT_COMP_REFS
#define SECOND_REF_FRAME_MASK                                        \
  ((1 << ALTREF_FRAME) | (1 << BWDREF_FRAME) | (1 << GOLDEN_FRAME) | \
   (1 << LAST2_FRAME) | 0x01)  // NOLINT
#else                          // !CONFIG_EXT_COMP_REFS
121
#define SECOND_REF_FRAME_MASK ((1 << ALTREF_FRAME) | (1 << BWDREF_FRAME) | 0x01)
Zoe Liu's avatar
Zoe Liu committed
122
123
#endif  // CONFIG_EXT_COMP_REFS
#else   // !CONFIG_EXT_REFS
124
#define SECOND_REF_FRAME_MASK ((1 << ALTREF_FRAME) | 0x01)
125
#endif  // CONFIG_EXT_REFS
Jingning Han's avatar
Jingning Han committed
126

127
128
#define MIN_EARLY_TERM_INDEX 3
#define NEW_MV_DISCOUNT_FACTOR 8
Jingning Han's avatar
Jingning Han committed
129

130
131
132
133
134
#if CONFIG_EXT_INTRA
#define ANGLE_SKIP_THRESH 10
#define FILTER_FAST_SEARCH 1
#endif  // CONFIG_EXT_INTRA

135
136
137
138
139
// Setting this to 1 will disable trellis optimization within the
// transform search. Trellis optimization will still be applied
// in the final encode.
#define DISABLE_TRELLISQ_SEARCH 0

140
141
const double ADST_FLIP_SVM[8] = { -6.6623, -2.8062, -3.2531, 3.1671,    // vert
                                  -7.7051, -3.2234, -3.6193, 3.4533 };  // horz
142

Jingning Han's avatar
Jingning Han committed
143
144
145
146
147
typedef struct {
  PREDICTION_MODE mode;
  MV_REFERENCE_FRAME ref_frame[2];
} MODE_DEFINITION;

148
typedef struct { MV_REFERENCE_FRAME ref_frame[2]; } REF_DEFINITION;
Jingning Han's avatar
Jingning Han committed
149
150

struct rdcost_block_args {
Yaowu Xu's avatar
Yaowu Xu committed
151
  const AV1_COMP *cpi;
Jingning Han's avatar
Jingning Han committed
152
  MACROBLOCK *x;
153
154
  ENTROPY_CONTEXT t_above[2 * MAX_MIB_SIZE];
  ENTROPY_CONTEXT t_left[2 * MAX_MIB_SIZE];
155
  RD_STATS rd_stats;
Jingning Han's avatar
Jingning Han committed
156
157
158
159
160
161
162
  int64_t this_rd;
  int64_t best_rd;
  int exit_early;
  int use_fast_coef_costing;
};

#define LAST_NEW_MV_INDEX 6
Yaowu Xu's avatar
Yaowu Xu committed
163
static const MODE_DEFINITION av1_mode_order[MAX_MODES] = {
Emil Keyder's avatar
Emil Keyder committed
164
  { NEARESTMV, { LAST_FRAME, NONE_FRAME } },
165
#if CONFIG_EXT_REFS
Emil Keyder's avatar
Emil Keyder committed
166
167
168
  { NEARESTMV, { LAST2_FRAME, NONE_FRAME } },
  { NEARESTMV, { LAST3_FRAME, NONE_FRAME } },
  { NEARESTMV, { BWDREF_FRAME, NONE_FRAME } },
169
#endif  // CONFIG_EXT_REFS
Emil Keyder's avatar
Emil Keyder committed
170
171
  { NEARESTMV, { ALTREF_FRAME, NONE_FRAME } },
  { NEARESTMV, { GOLDEN_FRAME, NONE_FRAME } },
Jingning Han's avatar
Jingning Han committed
172

Emil Keyder's avatar
Emil Keyder committed
173
  { DC_PRED, { INTRA_FRAME, NONE_FRAME } },
Jingning Han's avatar
Jingning Han committed
174

Emil Keyder's avatar
Emil Keyder committed
175
  { NEWMV, { LAST_FRAME, NONE_FRAME } },
176
#if CONFIG_EXT_REFS
Emil Keyder's avatar
Emil Keyder committed
177
178
179
  { NEWMV, { LAST2_FRAME, NONE_FRAME } },
  { NEWMV, { LAST3_FRAME, NONE_FRAME } },
  { NEWMV, { BWDREF_FRAME, NONE_FRAME } },
180
#endif  // CONFIG_EXT_REFS
Emil Keyder's avatar
Emil Keyder committed
181
182
  { NEWMV, { ALTREF_FRAME, NONE_FRAME } },
  { NEWMV, { GOLDEN_FRAME, NONE_FRAME } },
Jingning Han's avatar
Jingning Han committed
183

Emil Keyder's avatar
Emil Keyder committed
184
  { NEARMV, { LAST_FRAME, NONE_FRAME } },
185
#if CONFIG_EXT_REFS
Emil Keyder's avatar
Emil Keyder committed
186
187
188
  { NEARMV, { LAST2_FRAME, NONE_FRAME } },
  { NEARMV, { LAST3_FRAME, NONE_FRAME } },
  { NEARMV, { BWDREF_FRAME, NONE_FRAME } },
189
#endif  // CONFIG_EXT_REFS
Emil Keyder's avatar
Emil Keyder committed
190
191
  { NEARMV, { ALTREF_FRAME, NONE_FRAME } },
  { NEARMV, { GOLDEN_FRAME, NONE_FRAME } },
Jingning Han's avatar
Jingning Han committed
192

Emil Keyder's avatar
Emil Keyder committed
193
  { ZEROMV, { LAST_FRAME, NONE_FRAME } },
194
#if CONFIG_EXT_REFS
Emil Keyder's avatar
Emil Keyder committed
195
196
197
  { ZEROMV, { LAST2_FRAME, NONE_FRAME } },
  { ZEROMV, { LAST3_FRAME, NONE_FRAME } },
  { ZEROMV, { BWDREF_FRAME, NONE_FRAME } },
198
#endif  // CONFIG_EXT_REFS
Emil Keyder's avatar
Emil Keyder committed
199
200
  { ZEROMV, { GOLDEN_FRAME, NONE_FRAME } },
  { ZEROMV, { ALTREF_FRAME, NONE_FRAME } },
Jingning Han's avatar
Jingning Han committed
201

202
// TODO(zoeliu): May need to reconsider the order on the modes to check
203

204
#if CONFIG_EXT_INTER
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254

#if CONFIG_COMPOUND_SINGLEREF
  // Single ref comp mode
  { SR_NEAREST_NEARMV, { LAST_FRAME, NONE_FRAME } },
#if CONFIG_EXT_REFS
  { SR_NEAREST_NEARMV, { LAST2_FRAME, NONE_FRAME } },
  { SR_NEAREST_NEARMV, { LAST3_FRAME, NONE_FRAME } },
  { SR_NEAREST_NEARMV, { BWDREF_FRAME, NONE_FRAME } },
#endif  // CONFIG_EXT_REFS
  { SR_NEAREST_NEARMV, { GOLDEN_FRAME, NONE_FRAME } },
  { SR_NEAREST_NEARMV, { ALTREF_FRAME, NONE_FRAME } },

  /*
  { SR_NEAREST_NEWMV, { LAST_FRAME, NONE_FRAME } },
#if CONFIG_EXT_REFS
  { SR_NEAREST_NEWMV, { LAST2_FRAME, NONE_FRAME } },
  { SR_NEAREST_NEWMV, { LAST3_FRAME, NONE_FRAME } },
  { SR_NEAREST_NEWMV, { BWDREF_FRAME, NONE_FRAME } },
#endif  // CONFIG_EXT_REFS
  { SR_NEAREST_NEWMV, { GOLDEN_FRAME, NONE_FRAME } },
  { SR_NEAREST_NEWMV, { ALTREF_FRAME, NONE_FRAME } },*/

  { SR_NEAR_NEWMV, { LAST_FRAME, NONE_FRAME } },
#if CONFIG_EXT_REFS
  { SR_NEAR_NEWMV, { LAST2_FRAME, NONE_FRAME } },
  { SR_NEAR_NEWMV, { LAST3_FRAME, NONE_FRAME } },
  { SR_NEAR_NEWMV, { BWDREF_FRAME, NONE_FRAME } },
#endif  // CONFIG_EXT_REFS
  { SR_NEAR_NEWMV, { GOLDEN_FRAME, NONE_FRAME } },
  { SR_NEAR_NEWMV, { ALTREF_FRAME, NONE_FRAME } },

  { SR_ZERO_NEWMV, { LAST_FRAME, NONE_FRAME } },
#if CONFIG_EXT_REFS
  { SR_ZERO_NEWMV, { LAST2_FRAME, NONE_FRAME } },
  { SR_ZERO_NEWMV, { LAST3_FRAME, NONE_FRAME } },
  { SR_ZERO_NEWMV, { BWDREF_FRAME, NONE_FRAME } },
#endif  // CONFIG_EXT_REFS
  { SR_ZERO_NEWMV, { GOLDEN_FRAME, NONE_FRAME } },
  { SR_ZERO_NEWMV, { ALTREF_FRAME, NONE_FRAME } },

  { SR_NEW_NEWMV, { LAST_FRAME, NONE_FRAME } },
#if CONFIG_EXT_REFS
  { SR_NEW_NEWMV, { LAST2_FRAME, NONE_FRAME } },
  { SR_NEW_NEWMV, { LAST3_FRAME, NONE_FRAME } },
  { SR_NEW_NEWMV, { BWDREF_FRAME, NONE_FRAME } },
#endif  // CONFIG_EXT_REFS
  { SR_NEW_NEWMV, { GOLDEN_FRAME, NONE_FRAME } },
  { SR_NEW_NEWMV, { ALTREF_FRAME, NONE_FRAME } },
#endif  // CONFIG_COMPOUND_SINGLEREF

255
  { NEAREST_NEARESTMV, { LAST_FRAME, ALTREF_FRAME } },
256
#if CONFIG_EXT_REFS
257
258
  { NEAREST_NEARESTMV, { LAST2_FRAME, ALTREF_FRAME } },
  { NEAREST_NEARESTMV, { LAST3_FRAME, ALTREF_FRAME } },
259
#endif  // CONFIG_EXT_REFS
260
  { NEAREST_NEARESTMV, { GOLDEN_FRAME, ALTREF_FRAME } },
261
#if CONFIG_EXT_REFS
262
263
264
265
  { NEAREST_NEARESTMV, { LAST_FRAME, BWDREF_FRAME } },
  { NEAREST_NEARESTMV, { LAST2_FRAME, BWDREF_FRAME } },
  { NEAREST_NEARESTMV, { LAST3_FRAME, BWDREF_FRAME } },
  { NEAREST_NEARESTMV, { GOLDEN_FRAME, BWDREF_FRAME } },
Zoe Liu's avatar
Zoe Liu committed
266
267
268
269
270
271

#if CONFIG_EXT_COMP_REFS
  { NEAREST_NEARESTMV, { LAST_FRAME, LAST2_FRAME } },
  { NEAREST_NEARESTMV, { LAST_FRAME, GOLDEN_FRAME } },
  { NEAREST_NEARESTMV, { BWDREF_FRAME, ALTREF_FRAME } },
#endif  // CONFIG_EXT_COMP_REFS
272
#endif  // CONFIG_EXT_REFS
273

274
#else  // CONFIG_EXT_INTER
275

276
  { NEARESTMV, { LAST_FRAME, ALTREF_FRAME } },
277
#if CONFIG_EXT_REFS
278
279
  { NEARESTMV, { LAST2_FRAME, ALTREF_FRAME } },
  { NEARESTMV, { LAST3_FRAME, ALTREF_FRAME } },
280
#endif  // CONFIG_EXT_REFS
281
  { NEARESTMV, { GOLDEN_FRAME, ALTREF_FRAME } },
282
#if CONFIG_EXT_REFS
283
284
285
286
  { NEARESTMV, { LAST_FRAME, BWDREF_FRAME } },
  { NEARESTMV, { LAST2_FRAME, BWDREF_FRAME } },
  { NEARESTMV, { LAST3_FRAME, BWDREF_FRAME } },
  { NEARESTMV, { GOLDEN_FRAME, BWDREF_FRAME } },
Zoe Liu's avatar
Zoe Liu committed
287
288
289
290
291
292

#if CONFIG_EXT_COMP_REFS
  { NEARESTMV, { LAST_FRAME, LAST2_FRAME } },
  { NEARESTMV, { LAST_FRAME, GOLDEN_FRAME } },
  { NEARESTMV, { BWDREF_FRAME, ALTREF_FRAME } },
#endif  // CONFIG_EXT_COMP_REFS
293
#endif  // CONFIG_EXT_REFS
294
#endif  // CONFIG_EXT_INTER
Jingning Han's avatar
Jingning Han committed
295

Emil Keyder's avatar
Emil Keyder committed
296
  { TM_PRED, { INTRA_FRAME, NONE_FRAME } },
Jingning Han's avatar
Jingning Han committed
297

298
#if CONFIG_ALT_INTRA
Emil Keyder's avatar
Emil Keyder committed
299
  { SMOOTH_PRED, { INTRA_FRAME, NONE_FRAME } },
Urvang Joshi's avatar
Urvang Joshi committed
300
301
302
303
#if CONFIG_SMOOTH_HV
  { SMOOTH_V_PRED, { INTRA_FRAME, NONE_FRAME } },
  { SMOOTH_H_PRED, { INTRA_FRAME, NONE_FRAME } },
#endif  // CONFIG_SMOOTH_HV
304
305
#endif  // CONFIG_ALT_INTRA

306
#if CONFIG_EXT_INTER
307
308
309
310
311
312
313
  { NEAR_NEARMV, { LAST_FRAME, ALTREF_FRAME } },
  { NEW_NEARESTMV, { LAST_FRAME, ALTREF_FRAME } },
  { NEAREST_NEWMV, { LAST_FRAME, ALTREF_FRAME } },
  { NEW_NEARMV, { LAST_FRAME, ALTREF_FRAME } },
  { NEAR_NEWMV, { LAST_FRAME, ALTREF_FRAME } },
  { NEW_NEWMV, { LAST_FRAME, ALTREF_FRAME } },
  { ZERO_ZEROMV, { LAST_FRAME, ALTREF_FRAME } },
314

315
#if CONFIG_EXT_REFS
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
  { NEAR_NEARMV, { LAST2_FRAME, ALTREF_FRAME } },
  { NEW_NEARESTMV, { LAST2_FRAME, ALTREF_FRAME } },
  { NEAREST_NEWMV, { LAST2_FRAME, ALTREF_FRAME } },
  { NEW_NEARMV, { LAST2_FRAME, ALTREF_FRAME } },
  { NEAR_NEWMV, { LAST2_FRAME, ALTREF_FRAME } },
  { NEW_NEWMV, { LAST2_FRAME, ALTREF_FRAME } },
  { ZERO_ZEROMV, { LAST2_FRAME, ALTREF_FRAME } },

  { NEAR_NEARMV, { LAST3_FRAME, ALTREF_FRAME } },
  { NEW_NEARESTMV, { LAST3_FRAME, ALTREF_FRAME } },
  { NEAREST_NEWMV, { LAST3_FRAME, ALTREF_FRAME } },
  { NEW_NEARMV, { LAST3_FRAME, ALTREF_FRAME } },
  { NEAR_NEWMV, { LAST3_FRAME, ALTREF_FRAME } },
  { NEW_NEWMV, { LAST3_FRAME, ALTREF_FRAME } },
  { ZERO_ZEROMV, { LAST3_FRAME, ALTREF_FRAME } },
331
#endif  // CONFIG_EXT_REFS
332

333
334
335
336
337
338
339
  { NEAR_NEARMV, { GOLDEN_FRAME, ALTREF_FRAME } },
  { NEW_NEARESTMV, { GOLDEN_FRAME, ALTREF_FRAME } },
  { NEAREST_NEWMV, { GOLDEN_FRAME, ALTREF_FRAME } },
  { NEW_NEARMV, { GOLDEN_FRAME, ALTREF_FRAME } },
  { NEAR_NEWMV, { GOLDEN_FRAME, ALTREF_FRAME } },
  { NEW_NEWMV, { GOLDEN_FRAME, ALTREF_FRAME } },
  { ZERO_ZEROMV, { GOLDEN_FRAME, ALTREF_FRAME } },
340
341

#if CONFIG_EXT_REFS
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
  { NEAR_NEARMV, { LAST_FRAME, BWDREF_FRAME } },
  { NEW_NEARESTMV, { LAST_FRAME, BWDREF_FRAME } },
  { NEAREST_NEWMV, { LAST_FRAME, BWDREF_FRAME } },
  { NEW_NEARMV, { LAST_FRAME, BWDREF_FRAME } },
  { NEAR_NEWMV, { LAST_FRAME, BWDREF_FRAME } },
  { NEW_NEWMV, { LAST_FRAME, BWDREF_FRAME } },
  { ZERO_ZEROMV, { LAST_FRAME, BWDREF_FRAME } },

  { NEAR_NEARMV, { LAST2_FRAME, BWDREF_FRAME } },
  { NEW_NEARESTMV, { LAST2_FRAME, BWDREF_FRAME } },
  { NEAREST_NEWMV, { LAST2_FRAME, BWDREF_FRAME } },
  { NEW_NEARMV, { LAST2_FRAME, BWDREF_FRAME } },
  { NEAR_NEWMV, { LAST2_FRAME, BWDREF_FRAME } },
  { NEW_NEWMV, { LAST2_FRAME, BWDREF_FRAME } },
  { ZERO_ZEROMV, { LAST2_FRAME, BWDREF_FRAME } },

  { NEAR_NEARMV, { LAST3_FRAME, BWDREF_FRAME } },
  { NEW_NEARESTMV, { LAST3_FRAME, BWDREF_FRAME } },
  { NEAREST_NEWMV, { LAST3_FRAME, BWDREF_FRAME } },
  { NEW_NEARMV, { LAST3_FRAME, BWDREF_FRAME } },
  { NEAR_NEWMV, { LAST3_FRAME, BWDREF_FRAME } },
  { NEW_NEWMV, { LAST3_FRAME, BWDREF_FRAME } },
  { ZERO_ZEROMV, { LAST3_FRAME, BWDREF_FRAME } },

  { NEAR_NEARMV, { GOLDEN_FRAME, BWDREF_FRAME } },
  { NEW_NEARESTMV, { GOLDEN_FRAME, BWDREF_FRAME } },
  { NEAREST_NEWMV, { GOLDEN_FRAME, BWDREF_FRAME } },
  { NEW_NEARMV, { GOLDEN_FRAME, BWDREF_FRAME } },
  { NEAR_NEWMV, { GOLDEN_FRAME, BWDREF_FRAME } },
  { NEW_NEWMV, { GOLDEN_FRAME, BWDREF_FRAME } },
  { ZERO_ZEROMV, { GOLDEN_FRAME, BWDREF_FRAME } },
Zoe Liu's avatar
Zoe Liu committed
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398

#if CONFIG_EXT_COMP_REFS
  { NEAR_NEARMV, { LAST_FRAME, LAST2_FRAME } },
  { NEW_NEARESTMV, { LAST_FRAME, LAST2_FRAME } },
  { NEAREST_NEWMV, { LAST_FRAME, LAST2_FRAME } },
  { NEW_NEARMV, { LAST_FRAME, LAST2_FRAME } },
  { NEAR_NEWMV, { LAST_FRAME, LAST2_FRAME } },
  { NEW_NEWMV, { LAST_FRAME, LAST2_FRAME } },
  { ZERO_ZEROMV, { LAST_FRAME, LAST2_FRAME } },

  { NEAR_NEARMV, { LAST_FRAME, GOLDEN_FRAME } },
  { NEW_NEARESTMV, { LAST_FRAME, GOLDEN_FRAME } },
  { NEAREST_NEWMV, { LAST_FRAME, GOLDEN_FRAME } },
  { NEW_NEARMV, { LAST_FRAME, GOLDEN_FRAME } },
  { NEAR_NEWMV, { LAST_FRAME, GOLDEN_FRAME } },
  { NEW_NEWMV, { LAST_FRAME, GOLDEN_FRAME } },
  { ZERO_ZEROMV, { LAST_FRAME, GOLDEN_FRAME } },

  { NEAR_NEARMV, { BWDREF_FRAME, ALTREF_FRAME } },
  { NEW_NEARESTMV, { BWDREF_FRAME, ALTREF_FRAME } },
  { NEAREST_NEWMV, { BWDREF_FRAME, ALTREF_FRAME } },
  { NEW_NEARMV, { BWDREF_FRAME, ALTREF_FRAME } },
  { NEAR_NEWMV, { BWDREF_FRAME, ALTREF_FRAME } },
  { NEW_NEWMV, { BWDREF_FRAME, ALTREF_FRAME } },
  { ZERO_ZEROMV, { BWDREF_FRAME, ALTREF_FRAME } },
#endif  // CONFIG_EXT_COMP_REFS
399
#endif  // CONFIG_EXT_REFS
400

Zoe Liu's avatar
Zoe Liu committed
401
#else  // !CONFIG_EXT_INTER
402

403
404
  { NEARMV, { LAST_FRAME, ALTREF_FRAME } },
  { NEWMV, { LAST_FRAME, ALTREF_FRAME } },
405
#if CONFIG_EXT_REFS
406
407
408
409
  { NEARMV, { LAST2_FRAME, ALTREF_FRAME } },
  { NEWMV, { LAST2_FRAME, ALTREF_FRAME } },
  { NEARMV, { LAST3_FRAME, ALTREF_FRAME } },
  { NEWMV, { LAST3_FRAME, ALTREF_FRAME } },
410
#endif  // CONFIG_EXT_REFS
411
412
  { NEARMV, { GOLDEN_FRAME, ALTREF_FRAME } },
  { NEWMV, { GOLDEN_FRAME, ALTREF_FRAME } },
413
414

#if CONFIG_EXT_REFS
415
416
417
418
419
420
421
422
  { NEARMV, { LAST_FRAME, BWDREF_FRAME } },
  { NEWMV, { LAST_FRAME, BWDREF_FRAME } },
  { NEARMV, { LAST2_FRAME, BWDREF_FRAME } },
  { NEWMV, { LAST2_FRAME, BWDREF_FRAME } },
  { NEARMV, { LAST3_FRAME, BWDREF_FRAME } },
  { NEWMV, { LAST3_FRAME, BWDREF_FRAME } },
  { NEARMV, { GOLDEN_FRAME, BWDREF_FRAME } },
  { NEWMV, { GOLDEN_FRAME, BWDREF_FRAME } },
Zoe Liu's avatar
Zoe Liu committed
423
424
425
426
427
428
429
430
431

#if CONFIG_EXT_COMP_REFS
  { NEARMV, { LAST_FRAME, LAST2_FRAME } },
  { NEWMV, { LAST_FRAME, LAST2_FRAME } },
  { NEARMV, { LAST_FRAME, GOLDEN_FRAME } },
  { NEWMV, { LAST_FRAME, GOLDEN_FRAME } },
  { NEARMV, { BWDREF_FRAME, ALTREF_FRAME } },
  { NEWMV, { BWDREF_FRAME, ALTREF_FRAME } },
#endif  // CONFIG_EXT_COMP_REFS
432
#endif  // CONFIG_EXT_REFS
Jingning Han's avatar
Jingning Han committed
433

434
  { ZEROMV, { LAST_FRAME, ALTREF_FRAME } },
435
#if CONFIG_EXT_REFS
436
437
  { ZEROMV, { LAST2_FRAME, ALTREF_FRAME } },
  { ZEROMV, { LAST3_FRAME, ALTREF_FRAME } },
438
#endif  // CONFIG_EXT_REFS
439
  { ZEROMV, { GOLDEN_FRAME, ALTREF_FRAME } },
440
441

#if CONFIG_EXT_REFS
442
443
444
445
  { ZEROMV, { LAST_FRAME, BWDREF_FRAME } },
  { ZEROMV, { LAST2_FRAME, BWDREF_FRAME } },
  { ZEROMV, { LAST3_FRAME, BWDREF_FRAME } },
  { ZEROMV, { GOLDEN_FRAME, BWDREF_FRAME } },
Zoe Liu's avatar
Zoe Liu committed
446
447
448
449
450
451

#if CONFIG_EXT_COMP_REFS
  { ZEROMV, { LAST_FRAME, LAST2_FRAME } },
  { ZEROMV, { LAST_FRAME, GOLDEN_FRAME } },
  { ZEROMV, { BWDREF_FRAME, ALTREF_FRAME } },
#endif  // CONFIG_EXT_COMP_REFS
452
#endif  // CONFIG_EXT_REFS
453

454
#endif  // CONFIG_EXT_INTER
Jingning Han's avatar
Jingning Han committed
455

Emil Keyder's avatar
Emil Keyder committed
456
457
458
459
460
461
462
463
  { H_PRED, { INTRA_FRAME, NONE_FRAME } },
  { V_PRED, { INTRA_FRAME, NONE_FRAME } },
  { D135_PRED, { INTRA_FRAME, NONE_FRAME } },
  { D207_PRED, { INTRA_FRAME, NONE_FRAME } },
  { D153_PRED, { INTRA_FRAME, NONE_FRAME } },
  { D63_PRED, { INTRA_FRAME, NONE_FRAME } },
  { D117_PRED, { INTRA_FRAME, NONE_FRAME } },
  { D45_PRED, { INTRA_FRAME, NONE_FRAME } },
464
465

#if CONFIG_EXT_INTER
466
467
468
469
  { ZEROMV, { LAST_FRAME, INTRA_FRAME } },
  { NEARESTMV, { LAST_FRAME, INTRA_FRAME } },
  { NEARMV, { LAST_FRAME, INTRA_FRAME } },
  { NEWMV, { LAST_FRAME, INTRA_FRAME } },
470
471

#if CONFIG_EXT_REFS
472
473
474
475
476
477
478
479
480
  { ZEROMV, { LAST2_FRAME, INTRA_FRAME } },
  { NEARESTMV, { LAST2_FRAME, INTRA_FRAME } },
  { NEARMV, { LAST2_FRAME, INTRA_FRAME } },
  { NEWMV, { LAST2_FRAME, INTRA_FRAME } },

  { ZEROMV, { LAST3_FRAME, INTRA_FRAME } },
  { NEARESTMV, { LAST3_FRAME, INTRA_FRAME } },
  { NEARMV, { LAST3_FRAME, INTRA_FRAME } },
  { NEWMV, { LAST3_FRAME, INTRA_FRAME } },
481
482
#endif  // CONFIG_EXT_REFS

483
484
485
486
  { ZEROMV, { GOLDEN_FRAME, INTRA_FRAME } },
  { NEARESTMV, { GOLDEN_FRAME, INTRA_FRAME } },
  { NEARMV, { GOLDEN_FRAME, INTRA_FRAME } },
  { NEWMV, { GOLDEN_FRAME, INTRA_FRAME } },
487

488
#if CONFIG_EXT_REFS
489
490
491
492
  { ZEROMV, { BWDREF_FRAME, INTRA_FRAME } },
  { NEARESTMV, { BWDREF_FRAME, INTRA_FRAME } },
  { NEARMV, { BWDREF_FRAME, INTRA_FRAME } },
  { NEWMV, { BWDREF_FRAME, INTRA_FRAME } },
493
494
#endif  // CONFIG_EXT_REFS

495
496
497
498
  { ZEROMV, { ALTREF_FRAME, INTRA_FRAME } },
  { NEARESTMV, { ALTREF_FRAME, INTRA_FRAME } },
  { NEARMV, { ALTREF_FRAME, INTRA_FRAME } },
  { NEWMV, { ALTREF_FRAME, INTRA_FRAME } },
499
#endif  // CONFIG_EXT_INTER
Jingning Han's avatar
Jingning Han committed
500
501
};

hui su's avatar
hui su committed
502
503
504
505
506
507
508
509
510
511
512
513
static const PREDICTION_MODE intra_rd_search_mode_order[INTRA_MODES] = {
  DC_PRED,       H_PRED,        V_PRED,
#if CONFIG_ALT_INTRA
  SMOOTH_PRED,
#endif  // CONFIG_ALT_INTRA
  TM_PRED,
#if CONFIG_ALT_INTRA && CONFIG_SMOOTH_HV
  SMOOTH_V_PRED, SMOOTH_H_PRED,
#endif  // CONFIG_ALT_INTRA && CONFIG_SMOOTH_HV
  D135_PRED,     D207_PRED,     D153_PRED, D63_PRED, D117_PRED, D45_PRED,
};

514
#if CONFIG_EXT_INTRA || CONFIG_FILTER_INTRA || CONFIG_PALETTE
hui su's avatar
hui su committed
515
static INLINE int write_uniform_cost(int n, int v) {
516
517
  const int l = get_unsigned_bits(n);
  const int m = (1 << l) - n;
518
  if (l == 0) return 0;
hui su's avatar
hui su committed
519
  if (v < m)
Yaowu Xu's avatar
Yaowu Xu committed
520
    return (l - 1) * av1_cost_bit(128, 0);
hui su's avatar
hui su committed
521
  else
Yaowu Xu's avatar
Yaowu Xu committed
522
    return l * av1_cost_bit(128, 0);
hui su's avatar
hui su committed
523
}
524
#endif  // CONFIG_EXT_INTRA || CONFIG_FILTER_INTRA || CONFIG_PALETTE
hui su's avatar
hui su committed
525

526
527
528
// constants for prune 1 and prune 2 decision boundaries
#define FAST_EXT_TX_CORR_MID 0.0
#define FAST_EXT_TX_EDST_MID 0.1
529
530
531
#define FAST_EXT_TX_CORR_MARGIN 0.5
#define FAST_EXT_TX_EDST_MARGIN 0.3

Yushin Cho's avatar
Yushin Cho committed
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
#if CONFIG_DAALA_DIST
static int od_compute_var_4x4(od_coeff *x, int stride) {
  int sum;
  int s2;
  int i;
  sum = 0;
  s2 = 0;
  for (i = 0; i < 4; i++) {
    int j;
    for (j = 0; j < 4; j++) {
      int t;

      t = x[i * stride + j];
      sum += t;
      s2 += t * t;
    }
  }
  // TODO(yushin) : Check wheter any changes are required for high bit depth.
  return (s2 - (sum * sum >> 4)) >> 4;
}

553
554
555
556
557
558
559
/* OD_DIST_LP_MID controls the frequency weighting filter used for computing
   the distortion. For a value X, the filter is [1 X 1]/(X + 2) and
   is applied both horizontally and vertically. For X=5, the filter is
   a good approximation for the OD_QM8_Q4_HVS quantization matrix. */
#define OD_DIST_LP_MID (5)
#define OD_DIST_LP_NORM (OD_DIST_LP_MID + 2)

Yushin Cho's avatar
Yushin Cho committed
560
static double od_compute_dist_8x8(int qm, int use_activity_masking, od_coeff *x,
561
                                  od_coeff *y, od_coeff *e_lp, int stride) {
Yushin Cho's avatar
Yushin Cho committed
562
563
564
565
566
567
568
569
570
571
572
573
  double sum;
  int min_var;
  double mean_var;
  double var_stat;
  double activity;
  double calibration;
  int i;
  int j;
  double vardist;

  vardist = 0;
  OD_ASSERT(qm != OD_FLAT_QM);
574
  (void)qm;
Yushin Cho's avatar
Yushin Cho committed
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
#if 1
  min_var = INT_MAX;
  mean_var = 0;
  for (i = 0; i < 3; i++) {
    for (j = 0; j < 3; j++) {
      int varx;
      int vary;
      varx = od_compute_var_4x4(x + 2 * i * stride + 2 * j, stride);
      vary = od_compute_var_4x4(y + 2 * i * stride + 2 * j, stride);
      min_var = OD_MINI(min_var, varx);
      mean_var += 1. / (1 + varx);
      /* The cast to (double) is to avoid an overflow before the sqrt.*/
      vardist += varx - 2 * sqrt(varx * (double)vary) + vary;
    }
  }
  /* We use a different variance statistic depending on whether activity
     masking is used, since the harmonic mean appeared slghtly worse with
     masking off. The calibration constant just ensures that we preserve the
     rate compared to activity=1. */
  if (use_activity_masking) {
    calibration = 1.95;
    var_stat = 9. / mean_var;
  } else {
    calibration = 1.62;
    var_stat = min_var;
  }
  /* 1.62 is a calibration constant, 0.25 is a noise floor and 1/6 is the
     activity masking constant. */
  activity = calibration * pow(.25 + var_stat, -1. / 6);
#else
  activity = 1;
606
#endif  // 1
Yushin Cho's avatar
Yushin Cho committed
607
608
609
  sum = 0;
  for (i = 0; i < 8; i++) {
    for (j = 0; j < 8; j++)
610
      sum += e_lp[i * stride + j] * (double)e_lp[i * stride + j];
Yushin Cho's avatar
Yushin Cho committed
611
  }
612
613
614
  /* Normalize the filter to unit DC response. */
  sum *= 1. / (OD_DIST_LP_NORM * OD_DIST_LP_NORM * OD_DIST_LP_NORM *
               OD_DIST_LP_NORM);
Yushin Cho's avatar
Yushin Cho committed
615
616
617
618
  return activity * activity * (sum + vardist);
}

// Note : Inputs x and y are in a pixel domain
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
static double od_compute_dist_common(int qm, int activity_masking, od_coeff *x,
                                     od_coeff *y, int bsize_w, int bsize_h,
                                     int qindex, od_coeff *tmp,
                                     od_coeff *e_lp) {
  int i, j;
  double sum = 0;
  const int mid = OD_DIST_LP_MID;

  for (j = 0; j < bsize_w; j++) {
    e_lp[j] = mid * tmp[j] + 2 * tmp[bsize_w + j];
    e_lp[(bsize_h - 1) * bsize_w + j] = mid * tmp[(bsize_h - 1) * bsize_w + j] +
                                        2 * tmp[(bsize_h - 2) * bsize_w + j];
  }
  for (i = 1; i < bsize_h - 1; i++) {
    for (j = 0; j < bsize_w; j++) {
      e_lp[i * bsize_w + j] = mid * tmp[i * bsize_w + j] +
                              tmp[(i - 1) * bsize_w + j] +
                              tmp[(i + 1) * bsize_w + j];
    }
  }
  for (i = 0; i < bsize_h; i += 8) {
    for (j = 0; j < bsize_w; j += 8) {
      sum += od_compute_dist_8x8(qm, activity_masking, &x[i * bsize_w + j],
                                 &y[i * bsize_w + j], &e_lp[i * bsize_w + j],
                                 bsize_w);
    }
  }
  /* Scale according to linear regression against SSE, for 8x8 blocks. */
  if (activity_masking) {
    sum *= 2.2 + (1.7 - 2.2) * (qindex - 99) / (210 - 99) +
           (qindex < 99 ? 2.5 * (qindex - 99) / 99 * (qindex - 99) / 99 : 0);
  } else {
    sum *= qindex >= 128
               ? 1.4 + (0.9 - 1.4) * (qindex - 128) / (209 - 128)
               : qindex <= 43 ? 1.5 + (2.0 - 1.5) * (qindex - 43) / (16 - 43)
                              : 1.5 + (1.4 - 1.5) * (qindex - 43) / (128 - 43);
  }

  return sum;
}

Yushin Cho's avatar
Yushin Cho committed
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
static double od_compute_dist(int qm, int activity_masking, od_coeff *x,
                              od_coeff *y, int bsize_w, int bsize_h,
                              int qindex) {
  int i;
  double sum;
  sum = 0;

  assert(bsize_w >= 8 && bsize_h >= 8);

  if (qm == OD_FLAT_QM) {
    for (i = 0; i < bsize_w * bsize_h; i++) {
      double tmp;
      tmp = x[i] - y[i];
      sum += tmp * tmp;
    }
  } else {
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
    int j;
    DECLARE_ALIGNED(16, od_coeff, e[MAX_TX_SQUARE]);
    DECLARE_ALIGNED(16, od_coeff, tmp[MAX_TX_SQUARE]);
    DECLARE_ALIGNED(16, od_coeff, e_lp[MAX_TX_SQUARE]);
    int mid = OD_DIST_LP_MID;
    for (i = 0; i < bsize_h; i++) {
      for (j = 0; j < bsize_w; j++) {
        e[i * bsize_w + j] = x[i * bsize_w + j] - y[i * bsize_w + j];
      }
    }
    for (i = 0; i < bsize_h; i++) {
      tmp[i * bsize_w] = mid * e[i * bsize_w] + 2 * e[i * bsize_w + 1];
      tmp[i * bsize_w + bsize_w - 1] =
          mid * e[i * bsize_w + bsize_w - 1] + 2 * e[i * bsize_w + bsize_w - 2];
      for (j = 1; j < bsize_w - 1; j++) {
        tmp[i * bsize_w + j] = mid * e[i * bsize_w + j] +
                               e[i * bsize_w + j - 1] + e[i * bsize_w + j + 1];
      }
    }
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
    sum = od_compute_dist_common(qm, activity_masking, x, y, bsize_w, bsize_h,
                                 qindex, tmp, e_lp);
  }
  return sum;
}

static double od_compute_dist_diff(int qm, int activity_masking, od_coeff *x,
                                   od_coeff *e, int bsize_w, int bsize_h,
                                   int qindex) {
  int i;
  double sum;
  sum = 0;

  assert(bsize_w >= 8 && bsize_h >= 8);

  if (qm == OD_FLAT_QM) {
    for (i = 0; i < bsize_w * bsize_h; i++) {
      double tmp;
      tmp = e[i];
      sum += tmp * tmp;
715
    }
716
717
718
719
720
721
722
  } else {
    int j;
    DECLARE_ALIGNED(16, od_coeff, y[MAX_TX_SQUARE]);
    DECLARE_ALIGNED(16, od_coeff, tmp[MAX_TX_SQUARE]);
    DECLARE_ALIGNED(16, od_coeff, e_lp[MAX_TX_SQUARE]);
    int mid = OD_DIST_LP_MID;
    for (i = 0; i < bsize_h; i++) {
723
      for (j = 0; j < bsize_w; j++) {
724
        y[i * bsize_w + j] = x[i * bsize_w + j] - e[i * bsize_w + j];
725
726
      }
    }
727
728
729
730
731
732
733
    for (i = 0; i < bsize_h; i++) {
      tmp[i * bsize_w] = mid * e[i * bsize_w] + 2 * e[i * bsize_w + 1];
      tmp[i * bsize_w + bsize_w - 1] =
          mid * e[i * bsize_w + bsize_w - 1] + 2 * e[i * bsize_w + bsize_w - 2];
      for (j = 1; j < bsize_w - 1; j++) {
        tmp[i * bsize_w + j] = mid * e[i * bsize_w + j] +
                               e[i * bsize_w + j - 1] + e[i * bsize_w + j + 1];
Yushin Cho's avatar
Yushin Cho committed
734
735
      }
    }
736
737
738

    sum = od_compute_dist_common(qm, activity_masking, x, y, bsize_w, bsize_h,
                                 qindex, tmp, e_lp);
Yushin Cho's avatar
Yushin Cho committed
739
740
741
742
  }
  return sum;
}

743
744
745
746
int64_t av1_daala_dist(const MACROBLOCKD *xd, const uint8_t *src,
                       int src_stride, const uint8_t *dst, int dst_stride,
                       int bsw, int bsh, int visible_w, int visible_h, int qm,
                       int use_activity_masking, int qindex) {
Yushin Cho's avatar
Yushin Cho committed
747
748
749
750
  int i, j;
  int64_t d;
  DECLARE_ALIGNED(16, od_coeff, orig[MAX_TX_SQUARE]);
  DECLARE_ALIGNED(16, od_coeff, rec[MAX_TX_SQUARE]);
751
752
753
#if !CONFIG_HIGHBITDEPTH
  (void)xd;
#endif
Yushin Cho's avatar
Yushin Cho committed
754
755
  assert(qm == OD_HVS_QM);

756
757
#if CONFIG_HIGHBITDEPTH
  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
758
    for (j = 0; j < bsh; j++)
759
760
      for (i = 0; i < bsw; i++)
        orig[j * bsw + i] = CONVERT_TO_SHORTPTR(src)[j * src_stride + i];
761

762
    if ((bsw == visible_w) && (bsh == visible_h)) {
763
      for (j = 0; j < bsh; j++)
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
        for (i = 0; i < bsw; i++)
          rec[j * bsw + i] = CONVERT_TO_SHORTPTR(dst)[j * dst_stride + i];
    } else {
      for (j = 0; j < visible_h; j++)
        for (i = 0; i < visible_w; i++)
          rec[j * bsw + i] = CONVERT_TO_SHORTPTR(dst)[j * dst_stride + i];

      if (visible_w < bsw) {
        for (j = 0; j < bsh; j++)
          for (i = visible_w; i < bsw; i++)
            rec[j * bsw + i] = CONVERT_TO_SHORTPTR(src)[j * src_stride + i];
      }

      if (visible_h < bsh) {
        for (j = visible_h; j < bsh; j++)
          for (i = 0; i < bsw; i++)
            rec[j * bsw + i] = CONVERT_TO_SHORTPTR(src)[j * src_stride + i];
      }
782
    }
783
784
785
786
  } else {
#endif
    for (j = 0; j < bsh; j++)
      for (i = 0; i < bsw; i++) orig[j * bsw + i] = src[j * src_stride + i];
Yushin Cho's avatar
Yushin Cho committed
787

788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
    if ((bsw == visible_w) && (bsh == visible_h)) {
      for (j = 0; j < bsh; j++)
        for (i = 0; i < bsw; i++) rec[j * bsw + i] = dst[j * dst_stride + i];
    } else {
      for (j = 0; j < visible_h; j++)
        for (i = 0; i < visible_w; i++)
          rec[j * bsw + i] = dst[j * dst_stride + i];

      if (visible_w < bsw) {
        for (j = 0; j < bsh; j++)
          for (i = visible_w; i < bsw; i++)
            rec[j * bsw + i] = src[j * src_stride + i];
      }

      if (visible_h < bsh) {
        for (j = visible_h; j < bsh; j++)
          for (i = 0; i < bsw; i++) rec[j * bsw + i] = src[j * src_stride + i];
      }
806
    }
807
#if CONFIG_HIGHBITDEPTH
808
  }
809
810
#endif  // CONFIG_HIGHBITDEPTH

Yushin Cho's avatar
Yushin Cho committed
811
812
813
814
  d = (int64_t)od_compute_dist(qm, use_activity_masking, orig, rec, bsw, bsh,
                               qindex);
  return d;
}
815

816
817
818
819
820
static int64_t av1_daala_dist_diff(const MACROBLOCKD *xd, const uint8_t *src,
                                   int src_stride, const int16_t *diff,
                                   int dst_stride, int bsw, int bsh,
                                   int visible_w, int visible_h, int qm,
                                   int use_activity_masking, int qindex) {
821
822
823
824
  int i, j;
  int64_t d;
  DECLARE_ALIGNED(16, od_coeff, orig[MAX_TX_SQUARE]);
  DECLARE_ALIGNED(16, od_coeff, diff32[MAX_TX_SQUARE]);
825
826
827
#if !CONFIG_HIGHBITDEPTH
  (void)xd;
#endif
828
829
  assert(qm == OD_HVS_QM);

830
831
832
833
834
835
836
837
838
839
840
841
#if CONFIG_HIGHBITDEPTH
  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
    for (j = 0; j < bsh; j++)
      for (i = 0; i < bsw; i++)
        orig[j * bsw + i] = CONVERT_TO_SHORTPTR(src)[j * src_stride + i];
  } else {
#endif
    for (j = 0; j < bsh; j++)
      for (i = 0; i < bsw; i++) orig[j * bsw + i] = src[j * src_stride + i];
#if CONFIG_HIGHBITDEPTH
  }
#endif  // CONFIG_HIGHBITDEPTH
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860

  if ((bsw == visible_w) && (bsh == visible_h)) {
    for (j = 0; j < bsh; j++)
      for (i = 0; i < bsw; i++) diff32[j * bsw + i] = diff[j * dst_stride + i];
  } else {
    for (j = 0; j < visible_h; j++)
      for (i = 0; i < visible_w; i++)
        diff32[j * bsw + i] = diff[j * dst_stride + i];

    if (visible_w < bsw) {
      for (j = 0; j < bsh; j++)
        for (i = visible_w; i < bsw; i++) diff32[j * bsw + i] = 0;
    }

    if (visible_h < bsh) {
      for (j = visible_h; j < bsh; j++)
        for (i = 0; i < bsw; i++) diff32[j * bsw + i] = 0;
    }
  }
861

862
863
864
865
866
  d = (int64_t)od_compute_dist_diff(qm, use_activity_masking, orig, diff32, bsw,
                                    bsh, qindex);

  return d;
}
867
#endif  // CONFIG_DAALA_DIST
Yushin Cho's avatar
Yushin Cho committed
868

Yaowu Xu's avatar
Yaowu Xu committed
869
static void get_energy_distribution_fine(const AV1_COMP *cpi, BLOCK_SIZE bsize,
870
871
                                         const uint8_t *src, int src_stride,
                                         const uint8_t *dst, int dst_stride,
872
                                         double *hordist, double *verdist) {
873
874
  const int bw = block_size_wide[bsize];
  const int bh = block_size_high[bsize];
875
  unsigned int esq[16] = { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 };
876

877
  const int f_index = bsize - BLOCK_16X16;
878
  if (f_index < 0) {
879
880
    const int w_shift = bw == 8 ? 1 : 2;
    const int h_shift = bh == 8 ? 1 : 2;
881
#if CONFIG_HIGHBITDEPTH
882
    if (cpi->common.use_highbitdepth) {
883
884
885
886
887
      const uint16_t *src16 = CONVERT_TO_SHORTPTR(src);
      const uint16_t *dst16 = CONVERT_TO_SHORTPTR(dst);
      for (int i = 0; i < bh; ++i)
        for (int j = 0; j < bw; ++j) {
          const int index = (j >> w_shift) + ((i >> h_shift) << 2);
888
889
890
          esq[index] +=
              (src16[j + i * src_stride] - dst16[j + i * dst_stride]) *
              (src16[j + i * src_stride] - dst16[j + i * dst_stride]);
891
892
        }
    } else {
893
#endif  // CONFIG_HIGHBITDEPTH
894

895
896
897
      for (int i = 0; i < bh; ++i)
        for (int j = 0; j < bw; ++j) {
          const int index = (j >> w_shift) + ((i >> h_shift) << 2);
898
899
900
          esq[index] += (src[j + i * src_stride] - dst[j + i * dst_stride]) *
                        (src[j + i * src_stride] - dst[j + i * dst_stride]);
        }
901
#if CONFIG_HIGHBITDEPTH
902
    }
903
#endif  // CONFIG_HIGHBITDEPTH
904
  } else {
905
906
907
908
909
910
911
    cpi->fn_ptr[f_index].vf(src, src_stride, dst, dst_stride, &esq[0]);
    cpi->fn_ptr[f_index].vf(src + bw / 4, src_stride, dst + bw / 4, dst_stride,
                            &esq[1]);
    cpi->fn_ptr[f_index].vf(src + bw / 2, src_stride, dst + bw / 2, dst_stride,
                            &esq[2]);
    cpi->fn_ptr[f_index].vf(src + 3 * bw / 4, src_stride, dst + 3 * bw / 4,
                            dst_stride, &esq[3]);
912
913
914
    src += bh / 4 * src_stride;
    dst += bh / 4 * dst_stride;

915
916
917
918
919
920
921
    cpi->fn_ptr[f_index].vf(src, src_stride, dst, dst_stride, &esq[4]);
    cpi->fn_ptr[f_index].vf(src + bw / 4, src_stride, dst + bw / 4, dst_stride,
                            &esq[5]);
    cpi->fn_ptr[f_index].vf(src + bw / 2, src_stride, dst + bw / 2, dst_stride,
                            &esq[6]);
    cpi->fn_ptr[f_index].vf(src + 3 * bw / 4, src_stride, dst + 3 * bw / 4,
                            dst_stride, &esq[7]);
922
923
924
    src += bh / 4 * src_stride;
    dst += bh / 4 * dst_stride;

925
926
927
928
929
930
931
    cpi->fn_ptr[f_index].vf(src, src_stride, dst, dst_stride, &esq[8]);
    cpi->fn_ptr[f_index].vf(src + bw / 4, src_stride, dst + bw / 4, dst_stride,
                            &esq[9]);
    cpi->fn_ptr[f_index].vf(src + bw / 2, src_stride, dst + bw / 2, dst_stride,
                            &esq[10]);
    cpi->fn_ptr[f_index].vf(src + 3 * bw / 4, src_stride, dst + 3 * bw / 4,
                            dst_stride, &esq[11]);
932
933
934
    src += bh / 4 * src_stride;
    dst += bh / 4 * dst_stride;

935
936
937
938
939
940
941
    cpi->fn_ptr[f_index].vf(src, src_stride, dst, dst_stride, &esq[12]);
    cpi->fn_ptr[f_index].vf(src + bw / 4, src_stride, dst + bw / 4, dst_stride,
                            &esq[13]);
    cpi->fn_ptr[f_index].vf(src + bw / 2, src_stride, dst + bw / 2, dst_stride,
                            &esq[14]);
    cpi->fn_ptr[f_index].vf(src + 3 * bw / 4, src_stride, dst + 3 * bw / 4,
                            dst_stride, &esq[15]);
942
943
  }

944
945
946
  double total = (double)esq[0] + esq[1] + esq[2] + esq[3] + esq[4] + esq[5] +
                 esq[6] + esq[7] + esq[8] + esq[9] + esq[10] + esq[11] +
                 esq[12] + esq[13] + esq[14] + esq[15];
947
948
  if (total > 0) {
    const double e_recip = 1.0 / total;
949
950
951
952
953
954
    hordist[0] = ((double)esq[0] + esq[4] + esq[8] + esq[12]) * e_recip;
    hordist[1] = ((double)esq[1] + esq[5] + esq[9] + esq[13]) * e_recip;
    hordist[2] = ((double)esq[2] + esq[6] + esq[10] + esq[14]) * e_recip;
    verdist[0] = ((double)esq[0] + esq[1] + esq[2] + esq[3]) * e_recip;
    verdist[1] = ((double)esq[4] + esq[5] + esq[6] + esq[7]) * e_recip;
    verdist[2] = ((double)esq[8] + esq[9] + esq[10] + esq[11]) * e_recip;
955
956
957
958
959
960
961
  } else {
    hordist[0] = verdist[0] = 0.25;
    hordist[1] = verdist[1] = 0.25;
    hordist[2] = verdist[2] = 0.25;
  }
}

Urvang Joshi's avatar
Urvang Joshi committed
962
963
964
static int adst_vs_flipadst(const AV1_COMP *cpi, BLOCK_SIZE bsize,
                            const uint8_t *src, int src_stride,
                            const uint8_t *dst, int dst_stride) {
965
966
  int prune_bitmask = 0;
  double svm_proj_h = 0, svm_proj_v = 0;
Alex Converse's avatar
Alex Converse committed
967
  double hdist[3] = { 0, 0, 0 }, vdist[3] = { 0, 0, 0 };
968
969
  get_energy_distribution_fine(cpi, bsize, src, src_stride, dst, dst_stride,
                               hdist, vdist);
970

971
  svm_proj_v = vdist[0] * ADST_FLIP_SVM[0] + vdist[1] * ADST_FLIP_SVM[1] +
972
               vdist[2] * ADST_FLIP_SVM[2] + ADST_FLIP_SVM[3];
973
  svm_proj_h = hdist[0] * ADST_FLIP_SVM[4] + hdist[1] * ADST_FLIP_SVM[5] +
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
               hdist[2] * ADST_FLIP_SVM[6] + ADST_FLIP_SVM[7];
  if (svm_proj_v > FAST_EXT_TX_EDST_MID + FAST_EXT_TX_EDST_MARGIN)
    prune_bitmask |= 1 << FLIPADST_1D;
  else if (svm_proj_v < FAST_EXT_TX_EDST_MID - FAST_EXT_TX_EDST_MARGIN)
    prune_bitmask |= 1 << ADST_1D;

  if (svm_proj_h > FAST_EXT_TX_EDST_MID + FAST_EXT_TX_EDST_MARGIN)
    prune_bitmask |= 1 << (FLIPADST_1D + 8);
  else if (svm_proj_h < FAST_EXT_TX_EDST_MID - FAST_EXT_TX_EDST_MARGIN)
    prune_bitmask |= 1 << (ADST_1D + 8);

  return prune_bitmask;
}

#if CONFIG_EXT_TX
Alex Converse's avatar
Alex Converse committed
989
990
static void get_horver_correlation(const int16_t *diff, int stride, int w,
                                   int h, double *hcorr, double *vcorr) {
991
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
  // Returns hor/ver correlation coefficient
  const int num = (h - 1) * (w - 1);
  double num_r;
  int i, j;
  int64_t xy_sum = 0, xz_sum = 0;
  int64_t x_sum = 0, y_sum = 0, z_sum = 0;
  int64_t x2_sum = 0, y2_sum = 0, z2_sum = 0;
  double x_var_n, y_var_n, z_var_n, xy_var_n, xz_var_n;
  *hcorr = *vcorr = 1;

  assert(num > 0);
  num_r = 1.0 / num;
  for (i = 1; i < h; ++i) {
    for (j = 1; j < w; ++j) {
      const int16_t x = diff[i * stride + j];
      const int16_t y = diff[i * stride + j - 1];
      const int16_t z = diff[(i - 1) * stride + j];
      xy_sum += x * y;
      xz_sum += x * z;
      x_sum += x;
      y_sum += y;
      z_sum += z;
      x2_sum += x * x;
      y2_sum += y * y;
      z2_sum += z * z;
    }
  }
1018
1019
1020
  x_var_n = x2_sum - (x_sum * x_sum) * num_r;
  y_var_n = y2_sum - (y_sum * y_sum) * num_r;
  z_var_n = z2_sum - (z_sum * z_sum) * num_r;
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
  xy_var_n = xy_sum - (x_sum * y_sum) * num_r;
  xz_var_n = xz_sum - (x_sum * z_sum) * num_r;
  if (x_var_n > 0 && y_var_n > 0) {
    *hcorr = xy_var_n / sqrt(x_var_n * y_var_n);
    *hcorr = *hcorr < 0 ? 0 : *hcorr;
  }
  if (x_var_n > 0 && z_var_n > 0) {
    *vcorr = xz_var_n / sqrt(x_var_n * z_var_n);
    *vcorr = *vcorr < 0 ? 0 : *vcorr;
  }
}

Alex Converse's avatar
Alex Converse committed
1033
1034
int dct_vs_idtx(const int16_t *diff, int stride, int w, int h) {
  double hcorr, vcorr;
1035
  int prune_bitmask = 0;
Alex Converse's avatar
Alex Converse committed
1036
  get_horver_correlation(diff, stride, w, h, &hcorr, &vcorr);
1037

Alex Converse's avatar
Alex Converse committed
1038
  if (vcorr > FAST_EXT_TX_CORR_MID + FAST_EXT_TX_CORR_MARGIN)
1039
    prune_bitmask |= 1 << IDTX_1D;
Alex Converse's avatar
Alex Converse committed
1040
  else if (vcorr < FAST_EXT_TX_CORR_MID - FAST_EXT_TX_CORR_MARGIN)
1041
1042
    prune_bitmask |= 1 << DCT_1D;

Alex Converse's avatar
Alex Converse committed
1043
  if (hcorr > FAST_EXT_TX_CORR_MID + FAST_EXT_TX_CORR_MARGIN)
1044
    prune_bitmask |= 1 << (IDTX_1D + 8);
Alex Converse's avatar
Alex Converse committed
1045
  else if (hcorr < FAST_EXT_TX_CORR_MID - FAST_EXT_TX_CORR_MARGIN)
1046
1047
1048
1049
1050
    prune_bitmask |= 1 << (DCT_1D + 8);
  return prune_bitmask;
}

// Performance drop: 0.5%, Speed improvement: 24%
Yaowu Xu's avatar
Yaowu Xu committed
1051
static int prune_two_for_sby(const AV1_COMP *cpi, BLOCK_SIZE bsize,
Alex Converse's avatar
Alex Converse committed
1052
1053
                             MACROBLOCK *x, const MACROBLOCKD *xd,
                             int adst_flipadst, int dct_idtx) {
1054
  int prune = 0;
1055

Alex Converse's avatar
Alex Converse committed
1056
1057
1058
  if (adst_flipadst) {
    const struct macroblock_plane *const p = &x->plane[0];
    const struct macroblockd_plane *const pd = &xd->plane[0];
1059
    prune |= adst_vs_flipadst(cpi, bsize, p->src.buf, p->src.stride,
Alex Converse's avatar
Alex Converse committed
1060
1061
1062
1063
1064
1065
1066
1067
1068
                              pd->dst.buf, pd->dst.stride);
  }
  if (dct_idtx) {
    av1_subtract_plane(x, bsize, 0);
    const struct macroblock_plane *const p = &x->plane[0];
    const int bw = 4 << (b_width_log2_lookup[bsize]);
    const int bh = 4 << (b_height_log2_lookup[bsize]);
    prune |= dct_vs_idtx(p->src_diff, bw, bw, bh);
  }
1069
1070
1071

  return prune;
}
1072
1073
#endif  // CONFIG_EXT_TX

1074
// Performance drop: 0.3%, Speed improvement: 5%
Yaowu Xu's avatar
Yaowu Xu committed
1075
static int prune_one_for_sby(const AV1_COMP *cpi, BLOCK_SIZE bsize,
Alex Converse's avatar
Alex Converse committed
1076
1077
1078
                             const MACROBLOCK *x, const MACROBLOCKD *xd) {
  const struct macroblock_plane *const p = &x->plane[0];
  const struct macroblockd_plane *const pd = &xd->plane[0];
1079
  return adst_vs_flipadst(cpi, bsize, p->src.buf, p->src.stride, pd->dst.buf,
Alex Converse's avatar
Alex Converse committed
1080
                          pd->dst.stride);
1081
1082
}

Yaowu Xu's avatar
Yaowu Xu committed
1083
static int prune_tx_types(const AV1_COMP *cpi, BLOCK_SIZE bsize, MACROBLOCK *x,
Urvang Joshi's avatar
Urvang Joshi committed
1084
                          const MACROBLOCKD *const xd, int tx_set) {
1085
#if CONFIG_EXT_TX
Yaowu Xu's avatar
Yaowu Xu committed
1086
  const int *tx_set_1D = tx_set >= 0 ? ext_tx_used_inter_1D[tx_set] : NULL;
1087
#else
1088
  const int tx_set_1D[TX_TYPES_1D] = { 0 };
1089
#endif  // CONFIG_EXT_TX
1090

1091
  switch (cpi->sf.tx_type_search.prune_mode) {
1092
1093
    case NO_PRUNE: return 0; break;
    case PRUNE_ONE:
1094
      if ((tx_set >= 0) && !(tx_set_1D[FLIPADST_1D] & tx_set_1D[ADST_1D]))
1095
        return 0;
1096
1097
      return prune_one_for_sby(cpi, bsize, x, xd);
      break;
1098
1099
#if CONFIG_EXT_TX
    case PRUNE_TWO:
1100
      if ((tx_set >= 0) && !(tx_set_1D[FLIPADST_1D] & tx_set_1D[ADST_1D])) {
1101
        if (!(tx_set_1D[DCT_1D] & tx_set_1D[IDTX_1D])) return 0;
1102
1103
        return prune_two_for_sby(cpi, bsize, x, xd, 0, 1);
      }
1104
      if ((tx_set >= 0) && !(tx_set_1D[DCT_1D] & tx_set_1D[IDTX_1D]))
1105
1106
        return prune_two_for_sby(cpi, bsize, x, xd, 1, 0);
      return prune_two_for_sby(cpi, bsize, x, xd, 1, 1);
1107
      break;
1108
#endif  // CONFIG_EXT_TX
1109
1110
1111
1112
1113
  }
  assert(0);
  return 0;
}

1114
static int do_tx_type_search(TX_TYPE tx_type, int prune) {
1115
1116
1117
// TODO(sarahparker) implement for non ext tx
#if CONFIG_EXT_TX
  return !(((prune >> vtx_tab[tx_type]) & 1) |
1118
           ((prune >> (htx_tab[tx_type] + 8)) & 1));
1119
1120
#else
  // temporary to avoid compiler warnings
1121
1122
1123
1124
  (void)vtx_tab;
  (void)htx_tab;
  (void)tx_type;
  (void)prune;
1125
  return 1;
1126
#endif  // CONFIG_EXT_TX
1127
1128
}

Yaowu Xu's avatar
Yaowu Xu committed
1129
static void model_rd_from_sse(const AV1_COMP *const cpi,
1130
1131
                              const MACROBLOCKD *const xd, BLOCK_SIZE bsize,
                              int plane, int64_t sse, int *rate,
Geza Lore's avatar
Geza Lore committed
1132
1133
1134
                              int64_t *dist) {
  const struct macroblockd_plane *const pd = &xd->plane[plane];
  const int dequant_shift =
1135
#if CONFIG_HIGHBITDEPTH
1136
      (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) ? xd->bd - 5 :
1137
#endif  // CONFIG_HIGHBITDEPTH
1138
                                                    3;
Geza Lore's avatar
Geza Lore committed
1139
1140
1141
1142
1143
1144
1145

  // Fast approximate the modelling function.
  if (cpi->sf.simple_model_rd_from_var) {
    const int64_t square_error = sse;
    int quantizer = (pd->dequant[1] >> dequant_shift);

    if (quantizer < 120)
Yaowu Xu's avatar
Yaowu Xu committed
1146
      *rate = (int)((square_error * (280 - quantizer)) >>
Yaowu Xu's avatar
Yaowu Xu committed
1147
                    (16 - AV1_PROB_COST_SHIFT));
Geza Lore's avatar
Geza Lore committed
1148
1149
1150
1151
    else
      *rate = 0;
    *dist = (square_error * quantizer) >> 8;
  } else {
Yaowu Xu's avatar
Yaowu Xu committed
1152
1153
    av1_model_rd_from_var_lapndz(sse, num_pels_log2_lookup[bsize],
                                 pd->dequant[1] >> dequant_shift, rate, dist);
Geza Lore's avatar
Geza Lore committed
1154
1155
1156
1157
1158
  }

  *dist <<= 4;
}

Yaowu Xu's avatar
Yaowu Xu committed
1159
static void model_rd_for_sb(const AV1_COMP *const cpi, BLOCK_SIZE bsize,
1160
1161
1162
1163
                            MACROBLOCK *x, MACROBLOCKD *xd, int plane_from,
                            int plane_to, int *out_rate_sum,
                            int64_t *out_dist_sum, int *skip_txfm_sb,
                            int64_t *skip_sse_sb) {
Jingning Han's avatar
Jingning Han committed
1164
1165
1166
  // Note our transform coeffs are 8 times an orthogonal transform.
  // Hence quantizer step is also 8 times. To get effective quantizer
  // we need to divide by 8 before sending to modeling function.
Geza Lore's avatar
Geza Lore committed
1167
1168
1169
  int plane;
  const int ref = xd->mi[0]->mbmi.ref_frame[0];

Jingning Han's avatar
Jingning Han committed
1170
1171
1172
1173
1174
1175
  int64_t rate_sum = 0;
  int64_t dist_sum = 0;
  int64_t total_sse = 0;

  x->pred_sse[ref] = 0;

Geza Lore's avatar
Geza Lore committed
1176
1177
1178
  for (plane = plane_from; plane <= plane_to; ++plane) {
    struct macroblock_plane *const p = &x->plane[plane];
    struct macroblockd_plane *const pd = &xd->plane[plane];
1179
#if CONFIG_CHROMA_SUB8X8
1180
1181
    const BLOCK_SIZE bs = AOMMAX(BLOCK_4X4, get_plane_block_size(bsize, pd));
#else
Jingning Han's avatar
Jingning Han committed
1182
    const BLOCK_SIZE bs = get_plane_block_size(bsize, pd);
1183
#endif  // CONFIG_CHROMA_SUB8X8
Jingning Han's avatar
Jingning Han committed
1184

Geza Lore's avatar
Geza Lore committed
1185
1186
1187
    unsigned int sse;
    int rate;
    int64_t dist;
Jingning Han's avatar
Jingning Han committed
1188

1189
1190
#if CONFIG_CB4X4
    if (x->skip_chroma_rd && plane) continue;
1191
#endif  // CONFIG_CB4X4
1192

Geza Lore's avatar
Geza Lore committed
1193
1194
    // TODO(geza): Write direct sse functions that do not compute
    // variance as well.
1195
1196
    cpi->fn_ptr[bs].vf(p->src.buf, p->src.stride, pd->dst.buf, pd->dst.stride,
                       &sse);
Jingning Han's avatar
Jingning Han committed
1197

1198
    if (plane == 0) x->pred_sse[ref] = sse;
Jingning Han's avatar
Jingning Han committed
1199

Geza Lore's avatar
Geza Lore committed
1200
    total_sse += sse;
Jingning Han's avatar
Jingning Han committed
1201

Geza Lore's avatar
Geza Lore committed
1202
    model_rd_from_sse(cpi, xd, bs, plane, sse, &rate, &dist);
Geza Lore's avatar
Geza Lore committed
1203
1204
1205

    rate_sum += rate;
    dist_sum += dist;
Jingning Han's avatar
Jingning Han committed
1206
1207
  }

Geza Lore's avatar
Geza Lore committed
1208
  *skip_txfm_sb = total_sse == 0;
Jingning Han's avatar
Jingning Han committed
1209
1210
  *skip_sse_sb = total_sse << 4;
  *out_rate_sum = (int)rate_sum;
Geza Lore's avatar
Geza Lore committed
1211
  *out_dist_sum = dist_sum;
Jingning Han's avatar
Jingning Han committed
1212
1213
}

Yaowu Xu's avatar
Yaowu Xu committed
1214
1215
int64_t av1_block_error_c(const tran_low_t *coeff, const tran_low_t *dqcoeff,
                          intptr_t block_size, int64_t *ssz) {
Jingning Han's avatar
Jingning Han committed
1216
1217
1218
1219
1220
  int i;
  int64_t error = 0, sqcoeff = 0;

  for (i = 0; i < block_size; i++) {
    const int diff = coeff[i] - dqcoeff[i];
1221
    error += diff * diff;
Jingning Han's avatar
Jingning Han committed
1222
1223
1224
1225
1226
1227
1228
    sqcoeff += coeff[i] * coeff[i];
  }

  *ssz = sqcoeff;
  return error;
}

Yaowu Xu's avatar
Yaowu Xu committed
1229
1230
int64_t av1_block_error_fp_c(const int16_t *coeff, const int16_t *dqcoeff,
                             int block_size) {
Jingning Han's avatar
Jingning Han committed
1231
1232
1233
1234
1235
  int i;
  int64_t error = 0;

  for (i = 0; i < block_size; i++) {
    const int diff = coeff[i] - dqcoeff[i];
1236
    error += diff * diff;
Jingning Han's avatar
Jingning Han committed
1237
1238
1239
1240
1241
  }

  return error;
}

1242
#if CONFIG_HIGHBITDEPTH
Yaowu Xu's avatar
Yaowu Xu committed
1243
1244
1245
int64_t av1_highbd_block_error_c(const tran_low_t *coeff,
                                 const tran_low_t *dqcoeff, intptr_t block_size,
                                 int64_t *ssz, int bd) {
Jingning Han's avatar
Jingning Han committed
1246
1247
1248
1249
1250
1251
1252
  int i;
  int64_t error = 0, sqcoeff = 0;
  int shift = 2 * (bd - 8);
  int rounding = shift > 0 ? 1 << (shift - 1) : 0;

  for (i = 0; i < block_size; i++) {
    const int64_t diff = coeff[i] - dqcoeff[i];
1253
    error += diff * diff;
Jingning Han's avatar
Jingning Han committed
1254
1255
1256
1257
1258
1259
1260
1261
1262
    sqcoeff += (int64_t)coeff[i] * (int64_t)coeff[i];
  }
  assert(error >= 0 && sqcoeff >= 0);
  error = (error + rounding) >> shift;
  sqcoeff = (sqcoeff + rounding) >> shift;

  *ssz = sqcoeff;
  return error;
}
1263
#endif  // CONFIG_HIGHBITDEPTH
Jingning Han's avatar
Jingning Han committed
1264

Thomas Daede's avatar
Thomas Daede committed
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
#if CONFIG_PVQ
// Without PVQ, av1_block_error_c() return two kind of errors,
// 1) reconstruction (i.e. decoded) error and
// 2) Squared sum of transformed residue (i.e. 'coeff')
// However, if PVQ is enabled, coeff does not keep the transformed residue
// but instead a transformed original is kept.
// Hence, new parameter ref vector (i.e. transformed predicted signal)
// is required to derive the residue signal,
// i.e. coeff - ref = residue (all transformed).

1275
#if CONFIG_HIGHBITDEPTH
Thomas Daede's avatar
Thomas Daede committed
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
static int64_t av1_highbd_block_error2_c(const tran_low_t *coeff,
                                         const tran_low_t *dqcoeff,
                                         const tran_low_t *ref,
                                         intptr_t block_size, int64_t *ssz,
                                         int bd) {
  int64_t error;
  int64_t sqcoeff;
  int shift = 2 * (bd - 8);
  int rounding = shift > 0 ? 1 << (shift - 1) : 0;
  // Use the existing sse codes for calculating distortion of decoded signal:
  // i.e. (orig - decoded)^2
  // For high bit depth, throw away ssz until a 32-bit version of
  // av1_block_error_fp is written.
  int64_t ssz_trash;
  error = av1_block_error(coeff, dqcoeff, block_size, &ssz_trash);
  // prediction residue^2 = (orig - ref)^2
  sqcoeff = av1_block_error(coeff, ref, block_size, &ssz_trash);
  error = (error + rounding) >> shift;
  sqcoeff = (sqcoeff + rounding) >> shift;
  *ssz = sqcoeff;
  return error;
}
#else
// TODO(yushin) : Since 4x4 case does not need ssz, better to refactor into
// a separate function that does not do the extra computations for ssz.
static int64_t av1_block_error2_c(const tran_low_t *coeff,
                                  const tran_low_t *dqcoeff,
                                  const tran_low_t *ref, intptr_t block_size,
                                  int64_t *ssz) {
  int64_t error;
  // Use the existing sse codes for calculating distortion of decoded signal:
  // i.e. (orig - decoded)^2
  error = av1_block_error_fp(coeff, dqcoeff, block_size);
  // prediction residue^2 = (orig - ref)^2
  *ssz = av1_block_error_fp(coeff, ref, block_size);
  return error;
}
1313
#endif  // CONFIG_HIGHBITDEPTH
Thomas Daede's avatar
Thomas Daede committed
1314
1315
#endif  // CONFIG_PVQ

1316
#if !CONFIG_PVQ || CONFIG_VAR_TX
1317
/* The trailing '0' is a terminator which is used inside av1_cost_coeffs() to
Jingning Han's avatar
Jingning Han committed
1318
1319
1320
1321
 * decide whether to include cost of a trailing EOB node or not (i.e. we
 * can skip this if the last coefficient in this transform block, e.g. the
 * 16th coefficient in a 4x4 block or the 64th coefficient in a 8x8 block,
 * were non-zero). */
1322
1323
1324
1325
1326
#if !CONFIG_LV_MAP
static int cost_coeffs(const AV1_COMMON *const cm, MACROBLOCK *x, int plane,
                       int block, TX_SIZE tx_size, const SCAN_ORDER *scan_order,
                       const ENTROPY_CONTEXT *a, const ENTROPY_CONTEXT *l,
                       int use_fast_coef_costing) {
Jingning Han's avatar
Jingning Han committed
1327
1328
1329
1330
1331
  MACROBLOCKD *const xd = &x->e_mbd;
  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
  const struct macroblock_plane *p = &x->plane[plane];
  const struct macroblockd_plane *pd = &xd->plane[plane];
  const PLANE_TYPE type = pd->plane_type;
1332
  const uint16_t *band_count = &band_count_table[tx_size][1];
Jingning Han's avatar
Jingning Han committed
1333
1334
  const int eob = p->eobs[block];
  const tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
1335
  const int tx_size_ctx = txsize_sqr_map[tx_size];
1336
1337
  unsigned int(*token_costs)[2][COEFF_CONTEXTS][ENTROPY_TOKENS] =
      x->token_costs[tx_size_ctx][type][is_inter_block(mbmi)];
1338
  uint8_t token_cache[MAX_TX_SQUARE];
1339
  int pt = combine_entropy_contexts(*a, *l);
Jingning Han's avatar
Jingning Han committed
1340
  int c, cost;
1341
1342
  const int16_t *scan = scan_order->scan;
  const int16_t *nb = scan_order->neighbors;
1343
1344
1345
  const int ref = is_inter_block(mbmi);
  aom_prob *blockz_probs =
      cm->fc->blockzero_probs[txsize_sqr_map[tx_size]][type][ref];
Thomas Davies's avatar
Thomas Davies committed
1346

1347
#if CONFIG_HIGHBITDEPTH
1348
  const int cat6_bits = av1_get_cat6_extrabits_size(tx_size, xd->bd);
Jingning Han's avatar
Jingning Han committed
1349
#else
1350
  const int cat6_bits = av1_get_cat6_extrabits_size(tx_size, 8);
1351
#endif  // CONFIG_HIGHBITDEPTH
Jingning Han's avatar
Jingning Han committed
1352

1353
#if !CONFIG_VAR_TX && !CONFIG_SUPERTX
Jingning Han's avatar
Jingning Han committed
1354
  // Check for consistency of tx_size with mode info
Angie Chiang's avatar
Angie Chiang committed
1355
  assert(tx_size == get_tx_size(plane, xd));
1356
#endif  // !CONFIG_VAR_TX && !CONFIG_SUPERTX
1357
  (void)cm;
Jingning Han's avatar
Jingning Han committed
1358
1359
1360

  if (eob == 0) {
    // single eob token
1361
    cost = av1_cost_bit(blockz_probs[pt], 0);
Jingning Han's avatar
Jingning Han committed
1362
  } else {
Julia Robson's avatar
Julia Robson committed
1363
1364
1365
1366
1367
1368
    if (use_fast_coef_costing) {
      int band_left = *band_count++;

      // dc token
      int v = qcoeff[0];
      int16_t prev_t;
1369
      cost = av1_get_token_cost(v, &prev_t, cat6_bits);
Thomas Davies's avatar
Thomas Davies committed
1370
      cost += (*token_costs)[!prev_t][pt][prev_t];
Julia Robson's avatar
Julia Robson committed
1371

Yaowu Xu's avatar
Yaowu Xu committed
1372
      token_cache[0] = av1_pt_energy_class[prev_t];
Julia Robson's avatar
Julia Robson committed
1373
1374
1375
1376
1377
1378
1379
1380
      ++token_costs;

      // ac tokens
      for (c = 1; c < eob; c++) {
        const int rc = scan[c];
        int16_t t;

        v = qcoeff[rc];
1381
        cost += av1_get_token_cost(v, &t, cat6_bits);
Thomas Davies's avatar
Thomas Davies committed
1382
        cost += (*token_costs)[!t][!prev_t][t];
Julia Robson's avatar
Julia Robson committed
1383
1384
1385
1386
1387
        prev_t = t;
        if (!--band_left) {
          band_left = *band_count++;
          ++token_costs;
        }
Jingning Han's avatar
Jingning Han committed
1388
1389
      }

Julia Robson's avatar
Julia Robson committed
1390
      // eob token
1391
      cost += (*token_costs)[0][!prev_t][EOB_TOKEN];
Julia Robson's avatar
Julia Robson committed
1392
1393
1394
1395
1396
1397
1398

    } else {  // !use_fast_coef_costing
      int band_left = *band_count++;

      // dc token
      int v = qcoeff[0];
      int16_t tok;
1399
      cost = av1_get_token_cost(v, &tok, cat6_bits);
Thomas Davies's avatar
Thomas Davies committed