rdopt.c 479 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
#if CONFIG_CFL
#include "av1/common/cfl.h"
#endif
27
28
29
30
31
32
#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"
33
#include "av1/common/obmc.h"
34
35
36
37
38
39
#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"
40
41
42
#if CONFIG_LV_MAP
#include "av1/common/txb_common.h"
#endif
Yue Chen's avatar
Yue Chen committed
43
44
45
#if CONFIG_WARPED_MOTION
#include "av1/common/warped_motion.h"
#endif  // CONFIG_WARPED_MOTION
Jingning Han's avatar
Jingning Han committed
46

Jingning Han's avatar
Jingning Han committed
47
#include "av1/encoder/aq_variance.h"
48
#include "av1/encoder/av1_quantize.h"
49
50
51
52
#include "av1/encoder/cost.h"
#include "av1/encoder/encodemb.h"
#include "av1/encoder/encodemv.h"
#include "av1/encoder/encoder.h"
53
54
55
#if CONFIG_LV_MAP
#include "av1/encoder/encodetxb.h"
#endif
56
57
58
59
60
61
#include "av1/encoder/hybrid_fwd_txfm.h"
#include "av1/encoder/mcomp.h"
#include "av1/encoder/palette.h"
#include "av1/encoder/ratectrl.h"
#include "av1/encoder/rd.h"
#include "av1/encoder/rdopt.h"
62
#include "av1/encoder/tokenize.h"
63
64
#if CONFIG_PVQ
#include "av1/encoder/pvq_encoder.h"
Yushin Cho's avatar
Yushin Cho committed
65
#include "av1/common/pvq.h"
66
#endif  // CONFIG_PVQ
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

Zoe Liu's avatar
Zoe Liu committed
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
112
113
114
#define LAST_FRAME_MODE_MASK                                          \
  ((1 << INTRA_FRAME) | (1 << LAST2_FRAME) | (1 << LAST3_FRAME) |     \
   (1 << GOLDEN_FRAME) | (1 << BWDREF_FRAME) | (1 << ALTREF2_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 << ALTREF2_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 << ALTREF2_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 << ALTREF2_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 << ALTREF2_FRAME) | \
   (1 << ALTREF_FRAME))
#define ALTREF2_FRAME_MODE_MASK                                     \
  ((1 << INTRA_FRAME) | (1 << LAST_FRAME) | (1 << LAST2_FRAME) |    \
   (1 << LAST3_FRAME) | (1 << GOLDEN_FRAME) | (1 << BWDREF_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) | \
   (1 << ALTREF2_FRAME))

#else  // !CONFIG_EXT_REFS
115
116
117
118
119
120
121

#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
122

123
124
#endif  // CONFIG_EXT_REFS

125
#if CONFIG_EXT_REFS
Zoe Liu's avatar
Zoe Liu committed
126
#if CONFIG_EXT_COMP_REFS
127
128
129
#define SECOND_REF_FRAME_MASK                                         \
  ((1 << ALTREF_FRAME) | (1 << ALTREF2_FRAME) | (1 << BWDREF_FRAME) | \
   (1 << GOLDEN_FRAME) | (1 << LAST2_FRAME) | 0x01)
130
#else  // !CONFIG_EXT_COMP_REFS
Zoe Liu's avatar
Zoe Liu committed
131
132
#define SECOND_REF_FRAME_MASK \
  ((1 << ALTREF_FRAME) | (1 << ALTREF2_FRAME) | (1 << BWDREF_FRAME) | 0x01)
Zoe Liu's avatar
Zoe Liu committed
133
134
#endif  // CONFIG_EXT_COMP_REFS
#else   // !CONFIG_EXT_REFS
135
#define SECOND_REF_FRAME_MASK ((1 << ALTREF_FRAME) | 0x01)
136
#endif  // CONFIG_EXT_REFS
Jingning Han's avatar
Jingning Han committed
137

138
139
#define MIN_EARLY_TERM_INDEX 3
#define NEW_MV_DISCOUNT_FACTOR 8
Jingning Han's avatar
Jingning Han committed
140

141
142
143
144
145
#if CONFIG_EXT_INTRA
#define ANGLE_SKIP_THRESH 10
#define FILTER_FAST_SEARCH 1
#endif  // CONFIG_EXT_INTRA

146
147
148
// Setting this to 1 will disable trellis optimization within the
// transform search. Trellis optimization will still be applied
// in the final encode.
149
#ifndef DISABLE_TRELLISQ_SEARCH
150
#define DISABLE_TRELLISQ_SEARCH 0
151
#endif
152

153
154
155
156
157
158
static const double ADST_FLIP_SVM[8] = {
  /* vertical */
  -6.6623, -2.8062, -3.2531, 3.1671,
  /* horizontal */
  -7.7051, -3.2234, -3.6193, 3.4533
};
159

Jingning Han's avatar
Jingning Han committed
160
161
162
163
164
typedef struct {
  PREDICTION_MODE mode;
  MV_REFERENCE_FRAME ref_frame[2];
} MODE_DEFINITION;

165
typedef struct { MV_REFERENCE_FRAME ref_frame[2]; } REF_DEFINITION;
Jingning Han's avatar
Jingning Han committed
166
167

struct rdcost_block_args {
Yaowu Xu's avatar
Yaowu Xu committed
168
  const AV1_COMP *cpi;
Jingning Han's avatar
Jingning Han committed
169
  MACROBLOCK *x;
170
171
  ENTROPY_CONTEXT t_above[2 * MAX_MIB_SIZE];
  ENTROPY_CONTEXT t_left[2 * MAX_MIB_SIZE];
172
  RD_STATS rd_stats;
Jingning Han's avatar
Jingning Han committed
173
174
175
176
177
178
179
  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
180
static const MODE_DEFINITION av1_mode_order[MAX_MODES] = {
Emil Keyder's avatar
Emil Keyder committed
181
  { NEARESTMV, { LAST_FRAME, NONE_FRAME } },
182
#if CONFIG_EXT_REFS
Emil Keyder's avatar
Emil Keyder committed
183
184
185
  { NEARESTMV, { LAST2_FRAME, NONE_FRAME } },
  { NEARESTMV, { LAST3_FRAME, NONE_FRAME } },
  { NEARESTMV, { BWDREF_FRAME, NONE_FRAME } },
Zoe Liu's avatar
Zoe Liu committed
186
  { NEARESTMV, { ALTREF2_FRAME, NONE_FRAME } },
187
#endif  // CONFIG_EXT_REFS
Emil Keyder's avatar
Emil Keyder committed
188
189
  { NEARESTMV, { ALTREF_FRAME, NONE_FRAME } },
  { NEARESTMV, { GOLDEN_FRAME, NONE_FRAME } },
Jingning Han's avatar
Jingning Han committed
190

Emil Keyder's avatar
Emil Keyder committed
191
  { DC_PRED, { INTRA_FRAME, NONE_FRAME } },
Jingning Han's avatar
Jingning Han committed
192

Emil Keyder's avatar
Emil Keyder committed
193
  { NEWMV, { LAST_FRAME, NONE_FRAME } },
194
#if CONFIG_EXT_REFS
Emil Keyder's avatar
Emil Keyder committed
195
196
197
  { NEWMV, { LAST2_FRAME, NONE_FRAME } },
  { NEWMV, { LAST3_FRAME, NONE_FRAME } },
  { NEWMV, { BWDREF_FRAME, NONE_FRAME } },
Zoe Liu's avatar
Zoe Liu committed
198
  { NEWMV, { ALTREF2_FRAME, NONE_FRAME } },
199
#endif  // CONFIG_EXT_REFS
Emil Keyder's avatar
Emil Keyder committed
200
201
  { NEWMV, { ALTREF_FRAME, NONE_FRAME } },
  { NEWMV, { GOLDEN_FRAME, NONE_FRAME } },
Jingning Han's avatar
Jingning Han committed
202

Emil Keyder's avatar
Emil Keyder committed
203
  { NEARMV, { LAST_FRAME, NONE_FRAME } },
204
#if CONFIG_EXT_REFS
Emil Keyder's avatar
Emil Keyder committed
205
206
207
  { NEARMV, { LAST2_FRAME, NONE_FRAME } },
  { NEARMV, { LAST3_FRAME, NONE_FRAME } },
  { NEARMV, { BWDREF_FRAME, NONE_FRAME } },
Zoe Liu's avatar
Zoe Liu committed
208
  { NEARMV, { ALTREF2_FRAME, NONE_FRAME } },
209
#endif  // CONFIG_EXT_REFS
Emil Keyder's avatar
Emil Keyder committed
210
211
  { NEARMV, { ALTREF_FRAME, NONE_FRAME } },
  { NEARMV, { GOLDEN_FRAME, NONE_FRAME } },
Jingning Han's avatar
Jingning Han committed
212

Emil Keyder's avatar
Emil Keyder committed
213
  { ZEROMV, { LAST_FRAME, NONE_FRAME } },
214
#if CONFIG_EXT_REFS
Emil Keyder's avatar
Emil Keyder committed
215
216
217
  { ZEROMV, { LAST2_FRAME, NONE_FRAME } },
  { ZEROMV, { LAST3_FRAME, NONE_FRAME } },
  { ZEROMV, { BWDREF_FRAME, NONE_FRAME } },
Zoe Liu's avatar
Zoe Liu committed
218
  { ZEROMV, { ALTREF2_FRAME, NONE_FRAME } },
219
#endif  // CONFIG_EXT_REFS
Emil Keyder's avatar
Emil Keyder committed
220
221
  { ZEROMV, { GOLDEN_FRAME, NONE_FRAME } },
  { ZEROMV, { ALTREF_FRAME, NONE_FRAME } },
Jingning Han's avatar
Jingning Han committed
222

223
// TODO(zoeliu): May need to reconsider the order on the modes to check
224

225
#if CONFIG_EXT_INTER
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275

#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

276
  { NEAREST_NEARESTMV, { LAST_FRAME, ALTREF_FRAME } },
277
#if CONFIG_EXT_REFS
278
279
  { NEAREST_NEARESTMV, { LAST2_FRAME, ALTREF_FRAME } },
  { NEAREST_NEARESTMV, { LAST3_FRAME, ALTREF_FRAME } },
280
#endif  // CONFIG_EXT_REFS
281
  { NEAREST_NEARESTMV, { GOLDEN_FRAME, ALTREF_FRAME } },
282
#if CONFIG_EXT_REFS
283
284
285
286
  { 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
287
288
289
290
  { NEAREST_NEARESTMV, { LAST_FRAME, ALTREF2_FRAME } },
  { NEAREST_NEARESTMV, { LAST2_FRAME, ALTREF2_FRAME } },
  { NEAREST_NEARESTMV, { LAST3_FRAME, ALTREF2_FRAME } },
  { NEAREST_NEARESTMV, { GOLDEN_FRAME, ALTREF2_FRAME } },
Zoe Liu's avatar
Zoe Liu committed
291
292
293

#if CONFIG_EXT_COMP_REFS
  { NEAREST_NEARESTMV, { LAST_FRAME, LAST2_FRAME } },
294
  { NEAREST_NEARESTMV, { LAST_FRAME, LAST3_FRAME } },
Zoe Liu's avatar
Zoe Liu committed
295
296
297
  { NEAREST_NEARESTMV, { LAST_FRAME, GOLDEN_FRAME } },
  { NEAREST_NEARESTMV, { BWDREF_FRAME, ALTREF_FRAME } },
#endif  // CONFIG_EXT_COMP_REFS
298
#endif  // CONFIG_EXT_REFS
299

300
#else  // CONFIG_EXT_INTER
301

302
  { NEARESTMV, { LAST_FRAME, ALTREF_FRAME } },
303
#if CONFIG_EXT_REFS
304
305
  { NEARESTMV, { LAST2_FRAME, ALTREF_FRAME } },
  { NEARESTMV, { LAST3_FRAME, ALTREF_FRAME } },
306
#endif  // CONFIG_EXT_REFS
307
  { NEARESTMV, { GOLDEN_FRAME, ALTREF_FRAME } },
308
#if CONFIG_EXT_REFS
309
310
311
312
  { 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
313
314
315
316
  { NEARESTMV, { LAST_FRAME, ALTREF2_FRAME } },
  { NEARESTMV, { LAST2_FRAME, ALTREF2_FRAME } },
  { NEARESTMV, { LAST3_FRAME, ALTREF2_FRAME } },
  { NEARESTMV, { GOLDEN_FRAME, ALTREF2_FRAME } },
Zoe Liu's avatar
Zoe Liu committed
317
318
319

#if CONFIG_EXT_COMP_REFS
  { NEARESTMV, { LAST_FRAME, LAST2_FRAME } },
320
  { NEARESTMV, { LAST_FRAME, LAST3_FRAME } },
Zoe Liu's avatar
Zoe Liu committed
321
322
323
  { NEARESTMV, { LAST_FRAME, GOLDEN_FRAME } },
  { NEARESTMV, { BWDREF_FRAME, ALTREF_FRAME } },
#endif  // CONFIG_EXT_COMP_REFS
324
#endif  // CONFIG_EXT_REFS
325
#endif  // CONFIG_EXT_INTER
Jingning Han's avatar
Jingning Han committed
326

Emil Keyder's avatar
Emil Keyder committed
327
  { TM_PRED, { INTRA_FRAME, NONE_FRAME } },
Jingning Han's avatar
Jingning Han committed
328

Emil Keyder's avatar
Emil Keyder committed
329
  { SMOOTH_PRED, { INTRA_FRAME, NONE_FRAME } },
Urvang Joshi's avatar
Urvang Joshi committed
330
331
332
333
#if CONFIG_SMOOTH_HV
  { SMOOTH_V_PRED, { INTRA_FRAME, NONE_FRAME } },
  { SMOOTH_H_PRED, { INTRA_FRAME, NONE_FRAME } },
#endif  // CONFIG_SMOOTH_HV
334

335
#if CONFIG_EXT_INTER
336
337
338
339
340
341
342
  { 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 } },
343

344
#if CONFIG_EXT_REFS
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
  { 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 } },
360
#endif  // CONFIG_EXT_REFS
361

362
363
364
365
366
367
368
  { 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 } },
369
370

#if CONFIG_EXT_REFS
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
  { 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
402

Zoe Liu's avatar
Zoe Liu committed
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
  { NEAR_NEARMV, { LAST_FRAME, ALTREF2_FRAME } },
  { NEW_NEARESTMV, { LAST_FRAME, ALTREF2_FRAME } },
  { NEAREST_NEWMV, { LAST_FRAME, ALTREF2_FRAME } },
  { NEW_NEARMV, { LAST_FRAME, ALTREF2_FRAME } },
  { NEAR_NEWMV, { LAST_FRAME, ALTREF2_FRAME } },
  { NEW_NEWMV, { LAST_FRAME, ALTREF2_FRAME } },
  { ZERO_ZEROMV, { LAST_FRAME, ALTREF2_FRAME } },

  { NEAR_NEARMV, { LAST2_FRAME, ALTREF2_FRAME } },
  { NEW_NEARESTMV, { LAST2_FRAME, ALTREF2_FRAME } },
  { NEAREST_NEWMV, { LAST2_FRAME, ALTREF2_FRAME } },
  { NEW_NEARMV, { LAST2_FRAME, ALTREF2_FRAME } },
  { NEAR_NEWMV, { LAST2_FRAME, ALTREF2_FRAME } },
  { NEW_NEWMV, { LAST2_FRAME, ALTREF2_FRAME } },
  { ZERO_ZEROMV, { LAST2_FRAME, ALTREF2_FRAME } },

  { NEAR_NEARMV, { LAST3_FRAME, ALTREF2_FRAME } },
  { NEW_NEARESTMV, { LAST3_FRAME, ALTREF2_FRAME } },
  { NEAREST_NEWMV, { LAST3_FRAME, ALTREF2_FRAME } },
  { NEW_NEARMV, { LAST3_FRAME, ALTREF2_FRAME } },
  { NEAR_NEWMV, { LAST3_FRAME, ALTREF2_FRAME } },
  { NEW_NEWMV, { LAST3_FRAME, ALTREF2_FRAME } },
  { ZERO_ZEROMV, { LAST3_FRAME, ALTREF2_FRAME } },

  { NEAR_NEARMV, { GOLDEN_FRAME, ALTREF2_FRAME } },
  { NEW_NEARESTMV, { GOLDEN_FRAME, ALTREF2_FRAME } },
  { NEAREST_NEWMV, { GOLDEN_FRAME, ALTREF2_FRAME } },
  { NEW_NEARMV, { GOLDEN_FRAME, ALTREF2_FRAME } },
  { NEAR_NEWMV, { GOLDEN_FRAME, ALTREF2_FRAME } },
  { NEW_NEWMV, { GOLDEN_FRAME, ALTREF2_FRAME } },
  { ZERO_ZEROMV, { GOLDEN_FRAME, ALTREF2_FRAME } },

Zoe Liu's avatar
Zoe Liu committed
435
436
437
438
439
440
441
442
443
#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 } },

444
445
446
447
448
449
450
451
  { NEAR_NEARMV, { LAST_FRAME, LAST3_FRAME } },
  { NEW_NEARESTMV, { LAST_FRAME, LAST3_FRAME } },
  { NEAREST_NEWMV, { LAST_FRAME, LAST3_FRAME } },
  { NEW_NEARMV, { LAST_FRAME, LAST3_FRAME } },
  { NEAR_NEWMV, { LAST_FRAME, LAST3_FRAME } },
  { NEW_NEWMV, { LAST_FRAME, LAST3_FRAME } },
  { ZERO_ZEROMV, { LAST_FRAME, LAST3_FRAME } },

Zoe Liu's avatar
Zoe Liu committed
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
  { 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
468
#endif  // CONFIG_EXT_REFS
469

Zoe Liu's avatar
Zoe Liu committed
470
#else  // !CONFIG_EXT_INTER
471

472
473
  { NEARMV, { LAST_FRAME, ALTREF_FRAME } },
  { NEWMV, { LAST_FRAME, ALTREF_FRAME } },
474
#if CONFIG_EXT_REFS
475
476
477
478
  { NEARMV, { LAST2_FRAME, ALTREF_FRAME } },
  { NEWMV, { LAST2_FRAME, ALTREF_FRAME } },
  { NEARMV, { LAST3_FRAME, ALTREF_FRAME } },
  { NEWMV, { LAST3_FRAME, ALTREF_FRAME } },
479
#endif  // CONFIG_EXT_REFS
480
481
  { NEARMV, { GOLDEN_FRAME, ALTREF_FRAME } },
  { NEWMV, { GOLDEN_FRAME, ALTREF_FRAME } },
482
483

#if CONFIG_EXT_REFS
484
485
486
487
488
489
490
491
  { 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
492

Zoe Liu's avatar
Zoe Liu committed
493
494
495
496
497
498
499
500
501
  { NEARMV, { LAST_FRAME, ALTREF2_FRAME } },
  { NEWMV, { LAST_FRAME, ALTREF2_FRAME } },
  { NEARMV, { LAST2_FRAME, ALTREF2_FRAME } },
  { NEWMV, { LAST2_FRAME, ALTREF2_FRAME } },
  { NEARMV, { LAST3_FRAME, ALTREF2_FRAME } },
  { NEWMV, { LAST3_FRAME, ALTREF2_FRAME } },
  { NEARMV, { GOLDEN_FRAME, ALTREF2_FRAME } },
  { NEWMV, { GOLDEN_FRAME, ALTREF2_FRAME } },

Zoe Liu's avatar
Zoe Liu committed
502
503
504
#if CONFIG_EXT_COMP_REFS
  { NEARMV, { LAST_FRAME, LAST2_FRAME } },
  { NEWMV, { LAST_FRAME, LAST2_FRAME } },
505
506
  { NEARMV, { LAST_FRAME, LAST3_FRAME } },
  { NEWMV, { LAST_FRAME, LAST3_FRAME } },
Zoe Liu's avatar
Zoe Liu committed
507
508
509
510
511
  { 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
512
#endif  // CONFIG_EXT_REFS
Jingning Han's avatar
Jingning Han committed
513

514
  { ZEROMV, { LAST_FRAME, ALTREF_FRAME } },
515
#if CONFIG_EXT_REFS
516
517
  { ZEROMV, { LAST2_FRAME, ALTREF_FRAME } },
  { ZEROMV, { LAST3_FRAME, ALTREF_FRAME } },
518
#endif  // CONFIG_EXT_REFS
519
  { ZEROMV, { GOLDEN_FRAME, ALTREF_FRAME } },
520
521

#if CONFIG_EXT_REFS
522
523
524
525
  { 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
526

Zoe Liu's avatar
Zoe Liu committed
527
528
529
530
531
  { ZEROMV, { LAST_FRAME, ALTREF2_FRAME } },
  { ZEROMV, { LAST2_FRAME, ALTREF2_FRAME } },
  { ZEROMV, { LAST3_FRAME, ALTREF2_FRAME } },
  { ZEROMV, { GOLDEN_FRAME, ALTREF2_FRAME } },

Zoe Liu's avatar
Zoe Liu committed
532
533
#if CONFIG_EXT_COMP_REFS
  { ZEROMV, { LAST_FRAME, LAST2_FRAME } },
534
  { ZEROMV, { LAST_FRAME, LAST3_FRAME } },
Zoe Liu's avatar
Zoe Liu committed
535
536
537
  { ZEROMV, { LAST_FRAME, GOLDEN_FRAME } },
  { ZEROMV, { BWDREF_FRAME, ALTREF_FRAME } },
#endif  // CONFIG_EXT_COMP_REFS
538
#endif  // CONFIG_EXT_REFS
539

540
#endif  // CONFIG_EXT_INTER
Jingning Han's avatar
Jingning Han committed
541

Emil Keyder's avatar
Emil Keyder committed
542
543
544
545
546
547
548
549
  { 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 } },
550
551

#if CONFIG_EXT_INTER
552
553
554
555
  { ZEROMV, { LAST_FRAME, INTRA_FRAME } },
  { NEARESTMV, { LAST_FRAME, INTRA_FRAME } },
  { NEARMV, { LAST_FRAME, INTRA_FRAME } },
  { NEWMV, { LAST_FRAME, INTRA_FRAME } },
556
557

#if CONFIG_EXT_REFS
558
559
560
561
562
563
564
565
566
  { 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 } },
567
568
#endif  // CONFIG_EXT_REFS

569
570
571
572
  { ZEROMV, { GOLDEN_FRAME, INTRA_FRAME } },
  { NEARESTMV, { GOLDEN_FRAME, INTRA_FRAME } },
  { NEARMV, { GOLDEN_FRAME, INTRA_FRAME } },
  { NEWMV, { GOLDEN_FRAME, INTRA_FRAME } },
573

574
#if CONFIG_EXT_REFS
575
576
577
578
  { ZEROMV, { BWDREF_FRAME, INTRA_FRAME } },
  { NEARESTMV, { BWDREF_FRAME, INTRA_FRAME } },
  { NEARMV, { BWDREF_FRAME, INTRA_FRAME } },
  { NEWMV, { BWDREF_FRAME, INTRA_FRAME } },
Zoe Liu's avatar
Zoe Liu committed
579
580
581
582
583

  { ZEROMV, { ALTREF2_FRAME, INTRA_FRAME } },
  { NEARESTMV, { ALTREF2_FRAME, INTRA_FRAME } },
  { NEARMV, { ALTREF2_FRAME, INTRA_FRAME } },
  { NEWMV, { ALTREF2_FRAME, INTRA_FRAME } },
584
585
#endif  // CONFIG_EXT_REFS

586
587
588
589
  { ZEROMV, { ALTREF_FRAME, INTRA_FRAME } },
  { NEARESTMV, { ALTREF_FRAME, INTRA_FRAME } },
  { NEARMV, { ALTREF_FRAME, INTRA_FRAME } },
  { NEWMV, { ALTREF_FRAME, INTRA_FRAME } },
590
#endif  // CONFIG_EXT_INTER
Jingning Han's avatar
Jingning Han committed
591
592
};

hui su's avatar
hui su committed
593
static const PREDICTION_MODE intra_rd_search_mode_order[INTRA_MODES] = {
Urvang Joshi's avatar
Urvang Joshi committed
594
595
  DC_PRED,       H_PRED,        V_PRED,    SMOOTH_PRED, TM_PRED,
#if CONFIG_SMOOTH_HV
hui su's avatar
hui su committed
596
  SMOOTH_V_PRED, SMOOTH_H_PRED,
Urvang Joshi's avatar
Urvang Joshi committed
597
598
#endif  // CONFIG_SMOOTH_HV
  D135_PRED,     D207_PRED,     D153_PRED, D63_PRED,    D117_PRED, D45_PRED,
hui su's avatar
hui su committed
599
600
};

Luc Trudeau's avatar
Luc Trudeau committed
601
602
#if CONFIG_CFL
static const UV_PREDICTION_MODE uv_rd_search_mode_order[UV_INTRA_MODES] = {
Urvang Joshi's avatar
Urvang Joshi committed
603
604
605
  UV_DC_PRED,       UV_CFL_PRED,      UV_H_PRED,
  UV_V_PRED,        UV_SMOOTH_PRED,   UV_TM_PRED,
#if CONFIG_SMOOTH_HV
Luc Trudeau's avatar
Luc Trudeau committed
606
  UV_SMOOTH_V_PRED, UV_SMOOTH_H_PRED,
Urvang Joshi's avatar
Urvang Joshi committed
607
608
609
#endif  // CONFIG_SMOOTH_HV
  UV_D135_PRED,     UV_D207_PRED,     UV_D153_PRED,
  UV_D63_PRED,      UV_D117_PRED,     UV_D45_PRED,
Luc Trudeau's avatar
Luc Trudeau committed
610
611
612
613
614
};
#else
#define uv_rd_search_mode_order intra_rd_search_mode_order
#endif  // CONFIG_CFL

hui su's avatar
hui su committed
615
static INLINE int write_uniform_cost(int n, int v) {
616
617
  const int l = get_unsigned_bits(n);
  const int m = (1 << l) - n;
618
  if (l == 0) return 0;
hui su's avatar
hui su committed
619
  if (v < m)
Yaowu Xu's avatar
Yaowu Xu committed
620
    return (l - 1) * av1_cost_bit(128, 0);
hui su's avatar
hui su committed
621
  else
Yaowu Xu's avatar
Yaowu Xu committed
622
    return l * av1_cost_bit(128, 0);
hui su's avatar
hui su committed
623
624
}

625
626
627
// 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
628
629
630
#define FAST_EXT_TX_CORR_MARGIN 0.5
#define FAST_EXT_TX_EDST_MARGIN 0.3

631
#if CONFIG_DIST_8X8
Yushin Cho's avatar
Yushin Cho committed
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
static uint64_t cdef_dist_8x8_16bit(uint16_t *dst, int dstride, uint16_t *src,
                                    int sstride, int coeff_shift) {
  uint64_t svar = 0;
  uint64_t dvar = 0;
  uint64_t sum_s = 0;
  uint64_t sum_d = 0;
  uint64_t sum_s2 = 0;
  uint64_t sum_d2 = 0;
  uint64_t sum_sd = 0;
  uint64_t dist = 0;

  int i, j;
  for (i = 0; i < 8; i++) {
    for (j = 0; j < 8; j++) {
      sum_s += src[i * sstride + j];
      sum_d += dst[i * dstride + j];
      sum_s2 += src[i * sstride + j] * src[i * sstride + j];
      sum_d2 += dst[i * dstride + j] * dst[i * dstride + j];
      sum_sd += src[i * sstride + j] * dst[i * dstride + j];
    }
  }
  /* Compute the variance -- the calculation cannot go negative. */
  svar = sum_s2 - ((sum_s * sum_s + 32) >> 6);
  dvar = sum_d2 - ((sum_d * sum_d + 32) >> 6);

  // Tuning of jm's original dering distortion metric used in CDEF tool,
  // suggested by jm
  const uint64_t a = 4;
  const uint64_t b = 2;
  const uint64_t c1 = (400 * a << 2 * coeff_shift);
  const uint64_t c2 = (b * 20000 * a * a << 4 * coeff_shift);

  dist =
      (uint64_t)floor(.5 +
                      (sum_d2 + sum_s2 - 2 * sum_sd) * .5 * (svar + dvar + c1) /
                          (sqrt(svar * (double)dvar + c2)));

  // Calibrate dist to have similar rate for the same QP with MSE only
  // distortion (as in master branch)
  dist = (uint64_t)((float)dist * 0.75);

  return dist;
}

Yushin Cho's avatar
Yushin Cho committed
676
static int od_compute_var_4x4(uint16_t *x, int stride) {
Yushin Cho's avatar
Yushin Cho committed
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
  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;
    }
  }
Yushin Cho's avatar
Yushin Cho committed
692

Yushin Cho's avatar
Yushin Cho committed
693
694
695
  return (s2 - (sum * sum >> 4)) >> 4;
}

696
697
698
699
700
701
702
/* 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
703
704
static double od_compute_dist_8x8(int use_activity_masking, uint16_t *x,
                                  uint16_t *y, od_coeff *e_lp, int stride) {
Yushin Cho's avatar
Yushin Cho committed
705
706
707
708
709
710
711
712
713
714
715
  double sum;
  int min_var;
  double mean_var;
  double var_stat;
  double activity;
  double calibration;
  int i;
  int j;
  double vardist;

  vardist = 0;
Yushin Cho's avatar
Yushin Cho committed
716

Yushin Cho's avatar
Yushin Cho committed
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
#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
James Zern's avatar
James Zern committed
733
     masking is used, since the harmonic mean appeared slightly worse with
Yushin Cho's avatar
Yushin Cho committed
734
735
736
737
738
739
740
741
742
743
744
745
746
747
     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;
748
#endif  // 1
Yushin Cho's avatar
Yushin Cho committed
749
750
751
  sum = 0;
  for (i = 0; i < 8; i++) {
    for (j = 0; j < 8; j++)
752
      sum += e_lp[i * stride + j] * (double)e_lp[i * stride + j];
Yushin Cho's avatar
Yushin Cho committed
753
  }
754
755
756
  /* 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
757
758
759
760
  return activity * activity * (sum + vardist);
}

// Note : Inputs x and y are in a pixel domain
Yushin Cho's avatar
Yushin Cho committed
761
762
static double od_compute_dist_common(int activity_masking, uint16_t *x,
                                     uint16_t *y, int bsize_w, int bsize_h,
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
                                     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) {
Yushin Cho's avatar
Yushin Cho committed
783
      sum += od_compute_dist_8x8(activity_masking, &x[i * bsize_w + j],
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
                                 &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
802
803
static double od_compute_dist(uint16_t *x, uint16_t *y, int bsize_w,
                              int bsize_h, int qindex) {
Yushin Cho's avatar
Yushin Cho committed
804
  assert(bsize_w >= 8 && bsize_h >= 8);
Yushin Cho's avatar
Yushin Cho committed
805
806
807
808
809
#if CONFIG_PVQ
  int activity_masking = 1;
#else
  int activity_masking = 0;
#endif
Luc Trudeau's avatar
Luc Trudeau committed
810
811
812
813
814
815
816
  int i, 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]);
  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];
817
    }
Luc Trudeau's avatar
Luc Trudeau committed
818
819
820
821
822
823
824
825
826
  }
  int mid = OD_DIST_LP_MID;
  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];
827
    }
828
  }
Luc Trudeau's avatar
Luc Trudeau committed
829
830
  return od_compute_dist_common(activity_masking, x, y, bsize_w, bsize_h,
                                qindex, tmp, e_lp);
831
832
}

Yushin Cho's avatar
Yushin Cho committed
833
834
static double od_compute_dist_diff(uint16_t *x, int16_t *e, int bsize_w,
                                   int bsize_h, int qindex) {
835
  assert(bsize_w >= 8 && bsize_h >= 8);
Yushin Cho's avatar
Yushin Cho committed
836
837
838
839
840
#if CONFIG_PVQ
  int activity_masking = 1;
#else
  int activity_masking = 0;
#endif
Luc Trudeau's avatar
Luc Trudeau committed
841
842
843
844
845
846
847
  DECLARE_ALIGNED(16, uint16_t, y[MAX_TX_SQUARE]);
  DECLARE_ALIGNED(16, od_coeff, tmp[MAX_TX_SQUARE]);
  DECLARE_ALIGNED(16, od_coeff, e_lp[MAX_TX_SQUARE]);
  int i, j;
  for (i = 0; i < bsize_h; i++) {
    for (j = 0; j < bsize_w; j++) {
      y[i * bsize_w + j] = x[i * bsize_w + j] - e[i * bsize_w + j];
848
    }
Luc Trudeau's avatar
Luc Trudeau committed
849
850
851
852
853
854
855
856
857
  }
  int mid = OD_DIST_LP_MID;
  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
858
859
    }
  }
Luc Trudeau's avatar
Luc Trudeau committed
860
861
  return od_compute_dist_common(activity_masking, x, y, bsize_w, bsize_h,
                                qindex, tmp, e_lp);
Yushin Cho's avatar
Yushin Cho committed
862
863
}

864
int64_t av1_dist_8x8(const AV1_COMP *const cpi, const MACROBLOCK *x,
Yushin Cho's avatar
Yushin Cho committed
865
866
867
868
                     const uint8_t *src, int src_stride, const uint8_t *dst,
                     int dst_stride, const BLOCK_SIZE tx_bsize, int bsw,
                     int bsh, int visible_w, int visible_h, int qindex) {
  int64_t d = 0;
Yushin Cho's avatar
Yushin Cho committed
869
  int i, j;
870
  const MACROBLOCKD *xd = &x->e_mbd;
Yushin Cho's avatar
Yushin Cho committed
871
872
873
874

  DECLARE_ALIGNED(16, uint16_t, orig[MAX_TX_SQUARE]);
  DECLARE_ALIGNED(16, uint16_t, rec[MAX_TX_SQUARE]);

Yushin Cho's avatar
Yushin Cho committed
875
876
877
878
879
  assert(bsw >= 8);
  assert(bsh >= 8);
  assert((bsw & 0x07) == 0);
  assert((bsh & 0x07) == 0);

880
881
  if (x->tune_metric == AOM_TUNE_CDEF_DIST ||
      x->tune_metric == AOM_TUNE_DAALA_DIST) {
882
#if CONFIG_HIGHBITDEPTH
883
    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
884
      for (j = 0; j < bsh; j++)
885
        for (i = 0; i < bsw; i++)
886
          orig[j * bsw + i] = CONVERT_TO_SHORTPTR(src)[j * src_stride + i];
887

888
      if ((bsw == visible_w) && (bsh == visible_h)) {
889
890
        for (j = 0; j < bsh; j++)
          for (i = 0; i < bsw; i++)
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
            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];
        }
908
      }
909
    } else {
910
911
#endif
      for (j = 0; j < bsh; j++)
912
        for (i = 0; i < bsw; i++) orig[j * bsw + i] = src[j * src_stride + i];
913

914
      if ((bsw == visible_w) && (bsh == visible_h)) {
915
        for (j = 0; j < bsh; j++)
916
917
918
919
920
921
922
923
924
925
926
          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];
        }
927

928
929
930
931
932
        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];
        }
933
934
      }
#if CONFIG_HIGHBITDEPTH
935
    }
936
#endif  // CONFIG_HIGHBITDEPTH
937
  }
Yushin Cho's avatar
Yushin Cho committed
938

939
940
941
  if (x->tune_metric == AOM_TUNE_DAALA_DIST) {
    d = (int64_t)od_compute_dist(orig, rec, bsw, bsh, qindex);
  } else if (x->tune_metric == AOM_TUNE_CDEF_DIST) {
Yushin Cho's avatar
Yushin Cho committed
942
943
944
945
946
947
948
949
950
951
952
953
    int coeff_shift = AOMMAX(xd->bd - 8, 0);

    for (i = 0; i < bsh; i += 8) {
      for (j = 0; j < bsw; j += 8) {
        d += cdef_dist_8x8_16bit(&rec[i * bsw + j], bsw, &orig[i * bsw + j],
                                 bsw, coeff_shift);
      }
    }
#if CONFIG_HIGHBITDEPTH
    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH)
      d = ((uint64_t)d) >> 2 * coeff_shift;
#endif
954
955
956
957
958
959
  } else {
    // Otherwise, MSE by default
    unsigned sse;
    // TODO(Any): Use even faster function which does not calculate variance
    cpi->fn_ptr[tx_bsize].vf(src, src_stride, dst, dst_stride, &sse);
    d = sse;
Yushin Cho's avatar
Yushin Cho committed
960
  }
961

Yushin Cho's avatar
Yushin Cho committed
962
963
  return d;
}
964

965
static int64_t av1_dist_8x8_diff(const MACROBLOCK *x, const uint8_t *src,
Yushin Cho's avatar
Yushin Cho committed
966
967
968
969
                                 int src_stride, const int16_t *diff,
                                 int diff_stride, int bsw, int bsh,
                                 int visible_w, int visible_h, int qindex) {
  int64_t d = 0;
970
  int i, j;
971
  const MACROBLOCKD *xd = &x->e_mbd;
Yushin Cho's avatar
Yushin Cho committed
972
973
974
975

  DECLARE_ALIGNED(16, uint16_t, orig[MAX_TX_SQUARE]);
  DECLARE_ALIGNED(16, int16_t, diff16[MAX_TX_SQUARE]);

Yushin Cho's avatar
Yushin Cho committed
976
977
978
979
980
  assert(bsw >= 8);
  assert(bsh >= 8);
  assert((bsw & 0x07) == 0);
  assert((bsh & 0x07) == 0);

981
982
  if (x->tune_metric == AOM_TUNE_CDEF_DIST ||
      x->tune_metric == AOM_TUNE_DAALA_DIST) {
983
#if CONFIG_HIGHBITDEPTH
984
985
986
987
988
    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 {
989
#endif
990
991
      for (j = 0; j < bsh; j++)
        for (i = 0; i < bsw; i++) orig[j * bsw + i] = src[j * src_stride + i];
992
#if CONFIG_HIGHBITDEPTH
993
    }
994
#endif  // CONFIG_HIGHBITDEPTH
995

996
    if ((bsw == visible_w) && (bsh == visible_h)) {
997
      for (j = 0; j < bsh; j++)
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
        for (i = 0; i < bsw; i++)
          diff16[j * bsw + i] = diff[j * diff_stride + i];
    } else {
      for (j = 0; j < visible_h; j++)
        for (i = 0; i < visible_w; i++)
          diff16[j * bsw + i] = diff[j * diff_stride + i];

      if (visible_w < bsw) {
        for (j = 0; j < bsh; j++)
          for (i = visible_w; i < bsw; i++) diff16[j * bsw + i] = 0;
      }
1009

1010
1011
1012
1013
      if (visible_h < bsh) {
        for (j = visible_h; j < bsh; j++)
          for (i = 0; i < bsw; i++) diff16[j * bsw + i] = 0;
      }
1014
1015
    }
  }
1016

1017
1018
1019
  if (x->tune_metric == AOM_TUNE_DAALA_DIST) {
    d = (int64_t)od_compute_dist_diff(orig, diff16, bsw, bsh, qindex);
  } else if (x->tune_metric == AOM_TUNE_CDEF_DIST) {
Yushin Cho's avatar
Yushin Cho committed
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
    int coeff_shift = AOMMAX(xd->bd - 8, 0);
    DECLARE_ALIGNED(16, uint16_t, dst16[MAX_TX_SQUARE]);

    for (i = 0; i < bsh; i++) {
      for (j = 0; j < bsw; j++) {
        dst16[i * bsw + j] = orig[i * bsw + j] - diff16[i * bsw + j];
      }
    }

    for (i = 0; i < bsh; i += 8) {
      for (j = 0; j < bsw; j += 8) {
        d += cdef_dist_8x8_16bit(&dst16[i * bsw + j], bsw, &orig[i * bsw + j],
                                 bsw, coeff_shift);
      }
    }
    // Don't scale 'd' for HBD since it will be done by caller side for diff
    // input
1037
1038
1039
  } else {
    // Otherwise, MSE by default
    d = aom_sum_squares_2d_i16(diff, diff_stride, bsw, bsh);
Yushin Cho's avatar
Yushin Cho committed
1040
  }
1041
1042
1043

  return d;
}
Yushin Cho's avatar
Yushin Cho committed
1044
#endif  // CONFIG_DIST_8X8
Yushin Cho's avatar
Yushin Cho committed
1045

Yaowu Xu's avatar
Yaowu Xu committed
1046
static void get_energy_distribution_fine(const AV1_COMP *cpi, BLOCK_SIZE bsize,
1047
1048
                                         const uint8_t *src, int src_stride,
                                         const uint8_t *dst, int dst_stride,
1049
                                         double *hordist, double *verdist) {
1050
1051
  const int bw = block_size_wide[bsize];
  const int bh = block_size_high[bsize];
1052
  unsigned int esq[16] = { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 };
1053

1054
  const int f_index = bsize - BLOCK_16X16;
1055
  if (f_index < 0) {
1056
1057
    const int w_shift = bw == 8 ? 1 : 2;
    const int h_shift = bh == 8 ? 1 : 2;
1058
#if CONFIG_HIGHBITDEPTH
1059
    if (cpi->common.use_highbitdepth) {
1060
1061
1062
1063
1064
      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);
1065
1066
1067
          esq[index] +=
              (src16[j + i * src_stride] - dst16[j + i * dst_stride]) *
              (src16[j + i * src_stride] - dst16[j + i * dst_stride]);
1068
1069
        }
    } else {
1070
#endif  // CONFIG_HIGHBITDEPTH
1071

1072
1073
1074
      for (int i = 0; i < bh; ++i)
        for (int j = 0; j < bw; ++j) {
          const int index = (j >> w_shift) + ((i >> h_shift) << 2);
1075
1076
1077
          esq[index] += (src[j + i * src_stride] - dst[j + i * dst_stride]) *
                        (src[j + i * src_stride] - dst[j + i * dst_stride]);
        }
1078
#if CONFIG_HIGHBITDEPTH
1079
    }
1080
#endif  // CONFIG_HIGHBITDEPTH
1081
  } else {
1082
1083
1084
1085
1086
1087
1088
    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]);
1089
1090
1091
    src += bh / 4 * src_stride;
    dst += bh / 4 * dst_stride;

1092
1093
1094
1095
1096
1097
1098
    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]);
1099
1100
1101
    src += bh / 4 * src_stride;
    dst += bh / 4 * dst_stride;

1102
1103
1104
1105
1106
1107
1108
    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]);
1109
1110
1111
    src += bh / 4 * src_stride;
    dst += bh / 4 * dst_stride;

1112
1113
1114
1115
1116
1117
1118
    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]);
1119
1120
  }

1121
1122
1123
  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];
1124
1125
  if (total > 0) {
    const double e_recip = 1.0 / total;
1126
1127
1128
1129
1130
1131
    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;
1132
1133
1134
1135
1136
1137
1138
  } 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
1139
1140
1141
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) {
1142
1143
  int prune_bitmask = 0;
  double svm_proj_h = 0, svm_proj_v = 0;
Alex Converse's avatar
Alex Converse committed
1144
  double hdist[3] = { 0, 0, 0 }, vdist[3] = { 0, 0, 0 };
1145
1146
  get_energy_distribution_fine(cpi, bsize, src, src_stride, dst, dst_stride,
                               hdist, vdist);
1147

1148
  svm_proj_v = vdist[0] * ADST_FLIP_SVM[0] + vdist[1] * ADST_FLIP_SVM[1] +
1149
               vdist[2] * ADST_FLIP_SVM[2] + ADST_FLIP_SVM[3];
1150
  svm_proj_h = hdist[0] * ADST_FLIP_SVM[4] + hdist[1] * ADST_FLIP_SVM[5] +
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
               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
1166
1167
static void get_horver_correlation(const int16_t *diff, int stride, int w,
                                   int h, double *hcorr, double *vcorr) {
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
  // 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;
    }
  }
1195
1196
1197
  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;
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
  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
1210
1211
int dct_vs_idtx(const int16_t *diff, int stride, int w, int h) {
  double hcorr, vcorr;
1212
  int prune_bitmask = 0;
Alex Converse's avatar
Alex Converse committed
1213
  get_horver_correlation(diff, stride, w, h, &hcorr, &vcorr);
1214

Alex Converse's avatar
Alex Converse committed
1215
  if (vcorr > FAST_EXT_TX_CORR_MID + FAST_EXT_TX_CORR_MARGIN)
1216
    prune_bitmask |= 1 << IDTX_1D;
Alex Converse's avatar
Alex Converse committed
1217
  else if (vcorr < FAST_EXT_TX_CORR_MID - FAST_EXT_TX_CORR_MARGIN)
1218
1219
    prune_bitmask |= 1 << DCT_1D;

Alex Converse's avatar
Alex Converse committed
1220
  if (hcorr > FAST_EXT_TX_CORR_MID + FAST_EXT_TX_CORR_MARGIN)
1221
    prune_bitmask |= 1 << (IDTX_1D + 8);
Alex Converse's avatar
Alex Converse committed
1222
  else if (hcorr < FAST_EXT_TX_CORR_MID - FAST_EXT_TX_CORR_MARGIN)
1223
1224
1225
1226
1227
    prune_bitmask |= 1 << (DCT_1D + 8);
  return prune_bitmask;
}

// Performance drop: 0.5%, Speed improvement: 24%
Yaowu Xu's avatar
Yaowu Xu committed
1228
static int prune_two_for_sby(const AV1_COMP *cpi, BLOCK_SIZE bsize,
Alex Converse's avatar
Alex Converse committed
1229
1230
                             MACROBLOCK *x, const MACROBLOCKD *xd,
                             int adst_flipadst, int dct_idtx) {
1231
  int prune = 0;
1232

Alex Converse's avatar
Alex Converse committed
1233
1234
1235
  if (adst_flipadst) {
    const struct macroblock_plane *const p = &x->plane[0];
    const struct macroblockd_plane *const pd = &xd->plane[0];
1236
    prune |= adst_vs_flipadst(cpi, bsize, p->src.buf, p->src.stride,
Alex Converse's avatar
Alex Converse committed
1237
1238
1239
1240
1241
1242
1243
1244
1245
                              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);
  }
1246
1247
1248

  return prune;
}
1249
1250
#endif  // CONFIG_EXT_TX

1251
// Performance drop: 0.3%, Speed improvement: 5%
Yaowu Xu's avatar
Yaowu Xu committed
1252
static int prune_one_for_sby(const AV1_COMP *cpi, BLOCK_SIZE bsize,
Alex Converse's avatar
Alex Converse committed
1253
1254
1255
                             const MACROBLOCK *x, const MACROBLOCKD *xd) {
  const struct macroblock_plane *const p = &x->plane[0];
  const struct macroblockd_plane *const pd = &xd->plane[0];
1256
  return adst_vs_flipadst(cpi, bsize, p->src.buf, p->src.stride, pd->dst.buf,
Alex Converse's avatar
Alex Converse committed
1257
                          pd->dst.stride);
1258
1259
}

Hui Su's avatar
Hui Su committed
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
#if CONFIG_EXT_TX
// 1D Transforms used in inter set, this needs to be changed if
// ext_tx_used_inter is changed
static const int ext_tx_used_inter_1D[EXT_TX_SETS_INTER][TX_TYPES_1D] = {
  { 1, 0, 0, 0 }, { 1, 1, 1, 1 }, { 1, 1, 1, 1 }, { 1, 0, 0, 1 },
#if CONFIG_MRC_TX
  { 1, 0, 0, 1 },
#endif  // CONFIG_MRC_TX
};
#endif  // CONFIG_EXT_TX

Yaowu Xu's avatar
Yaowu Xu committed
1271
static int prune_tx_types(const AV1_COMP *cpi, BLOCK_SIZE bsize, MACROBLOCK *x,
Urvang Joshi's avatar
Urvang Joshi committed
1272
                          const MACROBLOCKD *const xd, int tx_set) {
1273
#if CONFIG_EXT_TX
Yaowu Xu's avatar
Yaowu Xu committed
1274
  const int *tx_set_1D = tx_set >= 0 ? ext_tx_used_inter_1D[tx_set] : NULL;
1275
#else
1276
  const int tx_set_1D[TX_TYPES_1D] = { 0 };
1277
#endif  // CONFIG_EXT_TX
1278

1279
  switch (cpi->sf.tx_type_search.prune_mode) {
1280
1281
    case NO_PRUNE: return 0; break;
    case PRUNE_ONE:
1282
      if ((tx_set >= 0) && !(tx_set_1D[FLIPADST_1D] & tx_set_1D[ADST_1D]))
1283
        return 0;
1284
1285
      return prune_one_for_sby(cpi, bsize, x, xd);
      break;
1286
1287
#if CONFIG_EXT_TX
    case PRUNE_TWO:
1288
      if ((tx_set >= 0) && !(tx_set_1D[FLIPADST_1D] & tx_set_1D[ADST_1D])) {
1289
        if (!(tx_set_1D[DCT_1D] & tx_set_1D[IDTX_1D])) return 0;
1290
1291
        return prune_two_for_sby(cpi, bsize, x, xd, 0, 1);
      }
1292
      if ((tx_set >= 0) && !(tx_set_1D[DCT_1D] & tx_set_1D[IDTX_1D]))
1293
1294
        return prune_two_for_sby(cpi, bsize, x, xd, 1, 0);
      return prune_two_for_sby(cpi, bsize, x, xd, 1, 1);
1295
      break;
1296
#endif  // CONFIG_EXT_TX
1297
1298
1299
1300
1301
  }
  assert(0);
  return 0;
}

1302
static int do_tx_type_search(TX_TYPE tx_type, int prune) {
1303
1304
1305
// TODO(sarahparker) implement for non ext tx
#if CONFIG_EXT_TX
  return !(((prune >> vtx_tab[tx_type]) & 1) |
1306
           ((prune >> (htx_tab[tx_type] + 8)) & 1));
1307
1308
#else
  // temporary to avoid compiler warnings
1309
1310
1311
1312
  (void)vtx_tab;
  (void)htx_tab;
  (void)tx_type;
  (void)prune;
1313
  return 1;
1314
#endif  // CONFIG_EXT_TX
1315
1316
}

Yaowu Xu's avatar
Yaowu Xu committed
1317
static void model_rd_from_sse(const AV1_COMP *const cpi,
1318
1319
                              const MACROBLOCKD *const xd, BLOCK_SIZE bsize,
                              int plane, int64_t sse, int *rate,
Geza Lore's avatar
Geza Lore committed
1320
1321
1322
                              int64_t *dist) {
  const struct macroblockd_plane *const pd = &xd->plane[plane];
  const int dequant_shift =
1323
#if CONFIG_HIGHBITDEPTH
1324
      (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) ? xd->bd - 5 :
1325
#endif  // CONFIG_HIGHBITDEPTH
1326
                                                    3;
Geza Lore's avatar
Geza Lore committed
1327
1328
1329
1330
1331
1332
1333

  // 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
1334
      *rate = (int)((square_error * (280 - quantizer)) >>
Yaowu Xu's avatar
Yaowu Xu committed
1335
                    (16 - AV1_PROB_COST_SHIFT));
Geza Lore's avatar
Geza Lore committed
1336
1337
1338
1339
    else
      *rate = 0;
    *dist = (square_error * quantizer) >> 8;
  } else {
Yaowu Xu's avatar
Yaowu Xu committed
1340
1341
    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
1342
1343
1344
1345
1346
  }

  *dist <<= 4;
}

Yaowu Xu's avatar
Yaowu Xu committed
1347
static void model_rd_for_sb(const AV1_COMP *const cpi, BLOCK_SIZE bsize,