rdopt.c 400 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 69
#define DUAL_FILTER_SET_SIZE (SWITCHABLE_FILTERS * SWITCHABLE_FILTERS)
static const int filter_sets[DUAL_FILTER_SET_SIZE][2] = {
70 71 72
  { 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 },
73
};
Angie Chiang's avatar
Angie Chiang committed
74
#endif  // CONFIG_DUAL_FILTER
75

76 77
#if CONFIG_EXT_REFS

78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104
#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
105

106 107
#endif  // CONFIG_EXT_REFS

108
#if CONFIG_EXT_REFS
109
#define SECOND_REF_FRAME_MASK ((1 << ALTREF_FRAME) | (1 << BWDREF_FRAME) | 0x01)
110
#else
111
#define SECOND_REF_FRAME_MASK ((1 << ALTREF_FRAME) | 0x01)
112
#endif  // CONFIG_EXT_REFS
Jingning Han's avatar
Jingning Han committed
113

114 115
#define MIN_EARLY_TERM_INDEX 3
#define NEW_MV_DISCOUNT_FACTOR 8
Jingning Han's avatar
Jingning Han committed
116

117 118 119 120 121
#if CONFIG_EXT_INTRA
#define ANGLE_SKIP_THRESH 10
#define FILTER_FAST_SEARCH 1
#endif  // CONFIG_EXT_INTRA

122 123
const double ADST_FLIP_SVM[8] = { -6.6623, -2.8062, -3.2531, 3.1671,    // vert
                                  -7.7051, -3.2234, -3.6193, 3.4533 };  // horz
124

Jingning Han's avatar
Jingning Han committed
125 126 127 128 129
typedef struct {
  PREDICTION_MODE mode;
  MV_REFERENCE_FRAME ref_frame[2];
} MODE_DEFINITION;

130
typedef struct { MV_REFERENCE_FRAME ref_frame[2]; } REF_DEFINITION;
Jingning Han's avatar
Jingning Han committed
131 132

struct rdcost_block_args {
Yaowu Xu's avatar
Yaowu Xu committed
133
  const AV1_COMP *cpi;
Jingning Han's avatar
Jingning Han committed
134
  MACROBLOCK *x;
135 136
  ENTROPY_CONTEXT t_above[2 * MAX_MIB_SIZE];
  ENTROPY_CONTEXT t_left[2 * MAX_MIB_SIZE];
137
  RD_STATS rd_stats;
Jingning Han's avatar
Jingning Han committed
138 139 140 141 142 143 144
  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
145
static const MODE_DEFINITION av1_mode_order[MAX_MODES] = {
Emil Keyder's avatar
Emil Keyder committed
146
  { NEARESTMV, { LAST_FRAME, NONE_FRAME } },
147
#if CONFIG_EXT_REFS
Emil Keyder's avatar
Emil Keyder committed
148 149 150
  { NEARESTMV, { LAST2_FRAME, NONE_FRAME } },
  { NEARESTMV, { LAST3_FRAME, NONE_FRAME } },
  { NEARESTMV, { BWDREF_FRAME, NONE_FRAME } },
151
#endif  // CONFIG_EXT_REFS
Emil Keyder's avatar
Emil Keyder committed
152 153
  { NEARESTMV, { ALTREF_FRAME, NONE_FRAME } },
  { NEARESTMV, { GOLDEN_FRAME, NONE_FRAME } },
Jingning Han's avatar
Jingning Han committed
154

Emil Keyder's avatar
Emil Keyder committed
155
  { DC_PRED, { INTRA_FRAME, NONE_FRAME } },
Jingning Han's avatar
Jingning Han committed
156

Emil Keyder's avatar
Emil Keyder committed
157
  { NEWMV, { LAST_FRAME, NONE_FRAME } },
158
#if CONFIG_EXT_REFS
Emil Keyder's avatar
Emil Keyder committed
159 160 161
  { NEWMV, { LAST2_FRAME, NONE_FRAME } },
  { NEWMV, { LAST3_FRAME, NONE_FRAME } },
  { NEWMV, { BWDREF_FRAME, NONE_FRAME } },
162
#endif  // CONFIG_EXT_REFS
Emil Keyder's avatar
Emil Keyder committed
163 164
  { NEWMV, { ALTREF_FRAME, NONE_FRAME } },
  { NEWMV, { GOLDEN_FRAME, NONE_FRAME } },
Jingning Han's avatar
Jingning Han committed
165

Emil Keyder's avatar
Emil Keyder committed
166
  { NEARMV, { LAST_FRAME, NONE_FRAME } },
167
#if CONFIG_EXT_REFS
Emil Keyder's avatar
Emil Keyder committed
168 169 170
  { NEARMV, { LAST2_FRAME, NONE_FRAME } },
  { NEARMV, { LAST3_FRAME, NONE_FRAME } },
  { NEARMV, { BWDREF_FRAME, NONE_FRAME } },
171
#endif  // CONFIG_EXT_REFS
Emil Keyder's avatar
Emil Keyder committed
172 173
  { NEARMV, { ALTREF_FRAME, NONE_FRAME } },
  { NEARMV, { GOLDEN_FRAME, NONE_FRAME } },
Jingning Han's avatar
Jingning Han committed
174

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

184
// TODO(zoeliu): May need to reconsider the order on the modes to check
185

186
#if CONFIG_EXT_INTER
187
  { NEAREST_NEARESTMV, { LAST_FRAME, ALTREF_FRAME } },
188
#if CONFIG_EXT_REFS
189 190
  { NEAREST_NEARESTMV, { LAST2_FRAME, ALTREF_FRAME } },
  { NEAREST_NEARESTMV, { LAST3_FRAME, ALTREF_FRAME } },
191
#endif  // CONFIG_EXT_REFS
192
  { NEAREST_NEARESTMV, { GOLDEN_FRAME, ALTREF_FRAME } },
193
#if CONFIG_EXT_REFS
194 195 196 197
  { NEAREST_NEARESTMV, { LAST_FRAME, BWDREF_FRAME } },
  { NEAREST_NEARESTMV, { LAST2_FRAME, BWDREF_FRAME } },
  { NEAREST_NEARESTMV, { LAST3_FRAME, BWDREF_FRAME } },
  { NEAREST_NEARESTMV, { GOLDEN_FRAME, BWDREF_FRAME } },
198
#endif  // CONFIG_EXT_REFS
199

200
#else  // CONFIG_EXT_INTER
201

202
  { NEARESTMV, { LAST_FRAME, ALTREF_FRAME } },
203
#if CONFIG_EXT_REFS
204 205
  { NEARESTMV, { LAST2_FRAME, ALTREF_FRAME } },
  { NEARESTMV, { LAST3_FRAME, ALTREF_FRAME } },
206
#endif  // CONFIG_EXT_REFS
207
  { NEARESTMV, { GOLDEN_FRAME, ALTREF_FRAME } },
208
#if CONFIG_EXT_REFS
209 210 211 212
  { NEARESTMV, { LAST_FRAME, BWDREF_FRAME } },
  { NEARESTMV, { LAST2_FRAME, BWDREF_FRAME } },
  { NEARESTMV, { LAST3_FRAME, BWDREF_FRAME } },
  { NEARESTMV, { GOLDEN_FRAME, BWDREF_FRAME } },
213
#endif  // CONFIG_EXT_REFS
214
#endif  // CONFIG_EXT_INTER
Jingning Han's avatar
Jingning Han committed
215

Emil Keyder's avatar
Emil Keyder committed
216
  { TM_PRED, { INTRA_FRAME, NONE_FRAME } },
Jingning Han's avatar
Jingning Han committed
217

218
#if CONFIG_ALT_INTRA
Emil Keyder's avatar
Emil Keyder committed
219
  { SMOOTH_PRED, { INTRA_FRAME, NONE_FRAME } },
Urvang Joshi's avatar
Urvang Joshi committed
220 221 222 223
#if CONFIG_SMOOTH_HV
  { SMOOTH_V_PRED, { INTRA_FRAME, NONE_FRAME } },
  { SMOOTH_H_PRED, { INTRA_FRAME, NONE_FRAME } },
#endif  // CONFIG_SMOOTH_HV
224 225
#endif  // CONFIG_ALT_INTRA

226
#if CONFIG_EXT_INTER
227 228 229 230 231 232 233
  { 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 } },
234

235
#if CONFIG_EXT_REFS
236 237 238 239 240 241 242 243 244 245 246 247 248 249 250
  { 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 } },
251
#endif  // CONFIG_EXT_REFS
252

253 254 255 256 257 258 259
  { 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 } },
260 261

#if CONFIG_EXT_REFS
262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292
  { 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 } },
293
#endif  // CONFIG_EXT_REFS
294 295 296

#else  // CONFIG_EXT_INTER

297 298
  { NEARMV, { LAST_FRAME, ALTREF_FRAME } },
  { NEWMV, { LAST_FRAME, ALTREF_FRAME } },
299
#if CONFIG_EXT_REFS
300 301 302 303
  { NEARMV, { LAST2_FRAME, ALTREF_FRAME } },
  { NEWMV, { LAST2_FRAME, ALTREF_FRAME } },
  { NEARMV, { LAST3_FRAME, ALTREF_FRAME } },
  { NEWMV, { LAST3_FRAME, ALTREF_FRAME } },
304
#endif  // CONFIG_EXT_REFS
305 306
  { NEARMV, { GOLDEN_FRAME, ALTREF_FRAME } },
  { NEWMV, { GOLDEN_FRAME, ALTREF_FRAME } },
307 308

#if CONFIG_EXT_REFS
309 310 311 312 313 314 315 316
  { 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 } },
317
#endif  // CONFIG_EXT_REFS
Jingning Han's avatar
Jingning Han committed
318

319
  { ZEROMV, { LAST_FRAME, ALTREF_FRAME } },
320
#if CONFIG_EXT_REFS
321 322
  { ZEROMV, { LAST2_FRAME, ALTREF_FRAME } },
  { ZEROMV, { LAST3_FRAME, ALTREF_FRAME } },
323
#endif  // CONFIG_EXT_REFS
324
  { ZEROMV, { GOLDEN_FRAME, ALTREF_FRAME } },
325 326

#if CONFIG_EXT_REFS
327 328 329 330
  { ZEROMV, { LAST_FRAME, BWDREF_FRAME } },
  { ZEROMV, { LAST2_FRAME, BWDREF_FRAME } },
  { ZEROMV, { LAST3_FRAME, BWDREF_FRAME } },
  { ZEROMV, { GOLDEN_FRAME, BWDREF_FRAME } },
331
#endif  // CONFIG_EXT_REFS
332

333
#endif  // CONFIG_EXT_INTER
Jingning Han's avatar
Jingning Han committed
334

Emil Keyder's avatar
Emil Keyder committed
335 336 337 338 339 340 341 342
  { 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 } },
343 344

#if CONFIG_EXT_INTER
345 346 347 348
  { ZEROMV, { LAST_FRAME, INTRA_FRAME } },
  { NEARESTMV, { LAST_FRAME, INTRA_FRAME } },
  { NEARMV, { LAST_FRAME, INTRA_FRAME } },
  { NEWMV, { LAST_FRAME, INTRA_FRAME } },
349 350

#if CONFIG_EXT_REFS
351 352 353 354 355 356 357 358 359
  { 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 } },
360 361
#endif  // CONFIG_EXT_REFS

362 363 364 365
  { ZEROMV, { GOLDEN_FRAME, INTRA_FRAME } },
  { NEARESTMV, { GOLDEN_FRAME, INTRA_FRAME } },
  { NEARMV, { GOLDEN_FRAME, INTRA_FRAME } },
  { NEWMV, { GOLDEN_FRAME, INTRA_FRAME } },
366

367
#if CONFIG_EXT_REFS
368 369 370 371
  { ZEROMV, { BWDREF_FRAME, INTRA_FRAME } },
  { NEARESTMV, { BWDREF_FRAME, INTRA_FRAME } },
  { NEARMV, { BWDREF_FRAME, INTRA_FRAME } },
  { NEWMV, { BWDREF_FRAME, INTRA_FRAME } },
372 373
#endif  // CONFIG_EXT_REFS

374 375 376 377
  { ZEROMV, { ALTREF_FRAME, INTRA_FRAME } },
  { NEARESTMV, { ALTREF_FRAME, INTRA_FRAME } },
  { NEARMV, { ALTREF_FRAME, INTRA_FRAME } },
  { NEWMV, { ALTREF_FRAME, INTRA_FRAME } },
378
#endif  // CONFIG_EXT_INTER
Jingning Han's avatar
Jingning Han committed
379 380
};

381
#if CONFIG_EXT_INTRA || CONFIG_FILTER_INTRA || CONFIG_PALETTE
hui su's avatar
hui su committed
382
static INLINE int write_uniform_cost(int n, int v) {
383 384
  const int l = get_unsigned_bits(n);
  const int m = (1 << l) - n;
385
  if (l == 0) return 0;
hui su's avatar
hui su committed
386
  if (v < m)
Yaowu Xu's avatar
Yaowu Xu committed
387
    return (l - 1) * av1_cost_bit(128, 0);
hui su's avatar
hui su committed
388
  else
Yaowu Xu's avatar
Yaowu Xu committed
389
    return l * av1_cost_bit(128, 0);
hui su's avatar
hui su committed
390
}
391
#endif  // CONFIG_EXT_INTRA || CONFIG_FILTER_INTRA || CONFIG_PALETTE
hui su's avatar
hui su committed
392

393 394 395
// 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
396 397 398
#define FAST_EXT_TX_CORR_MARGIN 0.5
#define FAST_EXT_TX_EDST_MARGIN 0.3

Yushin Cho's avatar
Yushin Cho committed
399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419
#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;
}

420 421 422 423 424 425 426
/* 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
427
static double od_compute_dist_8x8(int qm, int use_activity_masking, od_coeff *x,
428
                                  od_coeff *y, od_coeff *e_lp, int stride) {
Yushin Cho's avatar
Yushin Cho committed
429 430 431 432 433 434 435 436 437 438 439 440
  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);
441
  (void)qm;
Yushin Cho's avatar
Yushin Cho committed
442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472
#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;
473
#endif  // 1
Yushin Cho's avatar
Yushin Cho committed
474 475 476
  sum = 0;
  for (i = 0; i < 8; i++) {
    for (j = 0; j < 8; j++)
477
      sum += e_lp[i * stride + j] * (double)e_lp[i * stride + j];
Yushin Cho's avatar
Yushin Cho committed
478
  }
479 480 481
  /* 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
482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501
  return activity * activity * (sum + vardist);
}

// Note : Inputs x and y are in a pixel domain
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 {
502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533
    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];
      }
    }
    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];
      }
    }
Yushin Cho's avatar
Yushin Cho committed
534 535 536
    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],
537 538
                                   &y[i * bsize_w + j], &e_lp[i * bsize_w + j],
                                   bsize_w);
Yushin Cho's avatar
Yushin Cho committed
539 540
      }
    }
541 542 543 544 545 546 547 548 549 550 551
    /* 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);
    }
Yushin Cho's avatar
Yushin Cho committed
552 553 554 555
  }
  return sum;
}

Yushin Cho's avatar
Yushin Cho committed
556 557 558
int64_t av1_daala_dist(const uint8_t *src, int src_stride, const uint8_t *dst,
                       int dst_stride, int bsw, int bsh, int qm,
                       int use_activity_masking, int qindex) {
Yushin Cho's avatar
Yushin Cho committed
559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575
  int i, j;
  int64_t d;
  DECLARE_ALIGNED(16, od_coeff, orig[MAX_TX_SQUARE]);
  DECLARE_ALIGNED(16, od_coeff, rec[MAX_TX_SQUARE]);

  assert(qm == OD_HVS_QM);

  for (j = 0; j < bsh; j++)
    for (i = 0; i < bsw; i++) orig[j * bsw + i] = src[j * src_stride + i];

  for (j = 0; j < bsh; j++)
    for (i = 0; i < bsw; i++) rec[j * bsw + i] = dst[j * dst_stride + i];

  d = (int64_t)od_compute_dist(qm, use_activity_masking, orig, rec, bsw, bsh,
                               qindex);
  return d;
}
576
#endif  // CONFIG_DAALA_DIST
Yushin Cho's avatar
Yushin Cho committed
577

Yaowu Xu's avatar
Yaowu Xu committed
578
static void get_energy_distribution_fine(const AV1_COMP *cpi, BLOCK_SIZE bsize,
579 580
                                         const uint8_t *src, int src_stride,
                                         const uint8_t *dst, int dst_stride,
581
                                         double *hordist, double *verdist) {
582 583
  const int bw = block_size_wide[bsize];
  const int bh = block_size_high[bsize];
584
  unsigned int esq[16] = { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 };
585

586
  const int f_index = bsize - BLOCK_16X16;
587
  if (f_index < 0) {
588 589
    const int w_shift = bw == 8 ? 1 : 2;
    const int h_shift = bh == 8 ? 1 : 2;
590
#if CONFIG_HIGHBITDEPTH
591
    if (cpi->common.use_highbitdepth) {
592 593 594 595 596
      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);
597 598 599
          esq[index] +=
              (src16[j + i * src_stride] - dst16[j + i * dst_stride]) *
              (src16[j + i * src_stride] - dst16[j + i * dst_stride]);
600 601
        }
    } else {
602
#endif  // CONFIG_HIGHBITDEPTH
603

604 605 606
      for (int i = 0; i < bh; ++i)
        for (int j = 0; j < bw; ++j) {
          const int index = (j >> w_shift) + ((i >> h_shift) << 2);
607 608 609
          esq[index] += (src[j + i * src_stride] - dst[j + i * dst_stride]) *
                        (src[j + i * src_stride] - dst[j + i * dst_stride]);
        }
610
#if CONFIG_HIGHBITDEPTH
611
    }
612
#endif  // CONFIG_HIGHBITDEPTH
613
  } else {
614 615 616 617 618 619 620
    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]);
621 622 623
    src += bh / 4 * src_stride;
    dst += bh / 4 * dst_stride;

624 625 626 627 628 629 630
    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]);
631 632 633
    src += bh / 4 * src_stride;
    dst += bh / 4 * dst_stride;

634 635 636 637 638 639 640
    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]);
641 642 643
    src += bh / 4 * src_stride;
    dst += bh / 4 * dst_stride;

644 645 646 647 648 649 650
    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]);
651 652
  }

653 654 655
  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];
656 657
  if (total > 0) {
    const double e_recip = 1.0 / total;
658 659 660 661 662 663
    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;
664 665 666 667 668 669 670
  } 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
671 672 673
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) {
674 675
  int prune_bitmask = 0;
  double svm_proj_h = 0, svm_proj_v = 0;
Alex Converse's avatar
Alex Converse committed
676
  double hdist[3] = { 0, 0, 0 }, vdist[3] = { 0, 0, 0 };
677 678
  get_energy_distribution_fine(cpi, bsize, src, src_stride, dst, dst_stride,
                               hdist, vdist);
679

680
  svm_proj_v = vdist[0] * ADST_FLIP_SVM[0] + vdist[1] * ADST_FLIP_SVM[1] +
681
               vdist[2] * ADST_FLIP_SVM[2] + ADST_FLIP_SVM[3];
682
  svm_proj_h = hdist[0] * ADST_FLIP_SVM[4] + hdist[1] * ADST_FLIP_SVM[5] +
683 684 685 686 687 688 689 690 691 692 693 694 695 696 697
               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
698 699
static void get_horver_correlation(const int16_t *diff, int stride, int w,
                                   int h, double *hcorr, double *vcorr) {
700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726
  // 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;
    }
  }
727 728 729
  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;
730 731 732 733 734 735 736 737 738 739 740 741
  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
742 743
int dct_vs_idtx(const int16_t *diff, int stride, int w, int h) {
  double hcorr, vcorr;
744
  int prune_bitmask = 0;
Alex Converse's avatar
Alex Converse committed
745
  get_horver_correlation(diff, stride, w, h, &hcorr, &vcorr);
746

Alex Converse's avatar
Alex Converse committed
747
  if (vcorr > FAST_EXT_TX_CORR_MID + FAST_EXT_TX_CORR_MARGIN)
748
    prune_bitmask |= 1 << IDTX_1D;
Alex Converse's avatar
Alex Converse committed
749
  else if (vcorr < FAST_EXT_TX_CORR_MID - FAST_EXT_TX_CORR_MARGIN)
750 751
    prune_bitmask |= 1 << DCT_1D;

Alex Converse's avatar
Alex Converse committed
752
  if (hcorr > FAST_EXT_TX_CORR_MID + FAST_EXT_TX_CORR_MARGIN)
753
    prune_bitmask |= 1 << (IDTX_1D + 8);
Alex Converse's avatar
Alex Converse committed
754
  else if (hcorr < FAST_EXT_TX_CORR_MID - FAST_EXT_TX_CORR_MARGIN)
755 756 757 758 759
    prune_bitmask |= 1 << (DCT_1D + 8);
  return prune_bitmask;
}

// Performance drop: 0.5%, Speed improvement: 24%
Yaowu Xu's avatar
Yaowu Xu committed
760
static int prune_two_for_sby(const AV1_COMP *cpi, BLOCK_SIZE bsize,
Alex Converse's avatar
Alex Converse committed
761 762
                             MACROBLOCK *x, const MACROBLOCKD *xd,
                             int adst_flipadst, int dct_idtx) {
763
  int prune = 0;
764

Alex Converse's avatar
Alex Converse committed
765 766 767
  if (adst_flipadst) {
    const struct macroblock_plane *const p = &x->plane[0];
    const struct macroblockd_plane *const pd = &xd->plane[0];
768
    prune |= adst_vs_flipadst(cpi, bsize, p->src.buf, p->src.stride,
Alex Converse's avatar
Alex Converse committed
769 770 771 772 773 774 775 776 777
                              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);
  }
778 779 780

  return prune;
}
781 782
#endif  // CONFIG_EXT_TX

783
// Performance drop: 0.3%, Speed improvement: 5%
Yaowu Xu's avatar
Yaowu Xu committed
784
static int prune_one_for_sby(const AV1_COMP *cpi, BLOCK_SIZE bsize,
Alex Converse's avatar
Alex Converse committed
785 786 787
                             const MACROBLOCK *x, const MACROBLOCKD *xd) {
  const struct macroblock_plane *const p = &x->plane[0];
  const struct macroblockd_plane *const pd = &xd->plane[0];
788
  return adst_vs_flipadst(cpi, bsize, p->src.buf, p->src.stride, pd->dst.buf,
Alex Converse's avatar
Alex Converse committed
789
                          pd->dst.stride);
790 791
}

Yaowu Xu's avatar
Yaowu Xu committed
792
static int prune_tx_types(const AV1_COMP *cpi, BLOCK_SIZE bsize, MACROBLOCK *x,
Urvang Joshi's avatar
Urvang Joshi committed
793
                          const MACROBLOCKD *const xd, int tx_set) {
794
#if CONFIG_EXT_TX
Yaowu Xu's avatar
Yaowu Xu committed
795
  const int *tx_set_1D = tx_set >= 0 ? ext_tx_used_inter_1D[tx_set] : NULL;
796
#else
797
  const int tx_set_1D[TX_TYPES_1D] = { 0 };
798
#endif  // CONFIG_EXT_TX
799

800
  switch (cpi->sf.tx_type_search.prune_mode) {
801 802
    case NO_PRUNE: return 0; break;
    case PRUNE_ONE:
803
      if ((tx_set >= 0) && !(tx_set_1D[FLIPADST_1D] & tx_set_1D[ADST_1D]))
804
        return 0;
805 806
      return prune_one_for_sby(cpi, bsize, x, xd);
      break;
807 808
#if CONFIG_EXT_TX
    case PRUNE_TWO:
809
      if ((tx_set >= 0) && !(tx_set_1D[FLIPADST_1D] & tx_set_1D[ADST_1D])) {
810
        if (!(tx_set_1D[DCT_1D] & tx_set_1D[IDTX_1D])) return 0;
811 812
        return prune_two_for_sby(cpi, bsize, x, xd, 0, 1);
      }
813
      if ((tx_set >= 0) && !(tx_set_1D[DCT_1D] & tx_set_1D[IDTX_1D]))
814 815
        return prune_two_for_sby(cpi, bsize, x, xd, 1, 0);
      return prune_two_for_sby(cpi, bsize, x, xd, 1, 1);
816
      break;
817
#endif  // CONFIG_EXT_TX
818 819 820 821 822
  }
  assert(0);
  return 0;
}

823
static int do_tx_type_search(TX_TYPE tx_type, int prune) {
824 825 826
// TODO(sarahparker) implement for non ext tx
#if CONFIG_EXT_TX
  return !(((prune >> vtx_tab[tx_type]) & 1) |
827
           ((prune >> (htx_tab[tx_type] + 8)) & 1));
828 829
#else
  // temporary to avoid compiler warnings
830 831 832 833
  (void)vtx_tab;
  (void)htx_tab;
  (void)tx_type;
  (void)prune;
834
  return 1;
835
#endif  // CONFIG_EXT_TX
836 837
}

Yaowu Xu's avatar
Yaowu Xu committed
838
static void model_rd_from_sse(const AV1_COMP *const cpi,
839 840
                              const MACROBLOCKD *const xd, BLOCK_SIZE bsize,
                              int plane, int64_t sse, int *rate,
Geza Lore's avatar
Geza Lore committed
841 842 843
                              int64_t *dist) {
  const struct macroblockd_plane *const pd = &xd->plane[plane];
  const int dequant_shift =
844
#if CONFIG_HIGHBITDEPTH
845
      (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) ? xd->bd - 5 :
846
#endif  // CONFIG_HIGHBITDEPTH
847
                                                    3;
Geza Lore's avatar
Geza Lore committed
848 849 850 851 852 853 854

  // 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
855
      *rate = (int)((square_error * (280 - quantizer)) >>
Yaowu Xu's avatar
Yaowu Xu committed
856
                    (16 - AV1_PROB_COST_SHIFT));
Geza Lore's avatar
Geza Lore committed
857 858 859 860
    else
      *rate = 0;
    *dist = (square_error * quantizer) >> 8;
  } else {
Yaowu Xu's avatar
Yaowu Xu committed
861 862
    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
863 864 865 866 867
  }

  *dist <<= 4;
}

Yaowu Xu's avatar
Yaowu Xu committed
868
static void model_rd_for_sb(const AV1_COMP *const cpi, BLOCK_SIZE bsize,
869 870 871 872
                            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
873 874 875
  // 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
876 877 878
  int plane;
  const int ref = xd->mi[0]->mbmi.ref_frame[0];

Jingning Han's avatar
Jingning Han committed
879 880 881 882 883 884
  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
885 886 887
  for (plane = plane_from; plane <= plane_to; ++plane) {
    struct macroblock_plane *const p = &x->plane[plane];
    struct macroblockd_plane *const pd = &xd->plane[plane];
Jingning Han's avatar
Jingning Han committed
888
#if CONFIG_CB4X4 && !CONFIG_CHROMA_2X2
889 890
    const BLOCK_SIZE bs = AOMMAX(BLOCK_4X4, get_plane_block_size(bsize, pd));
#else
Jingning Han's avatar
Jingning Han committed
891
    const BLOCK_SIZE bs = get_plane_block_size(bsize, pd);
892
#endif  // CONFIG_CB4X4 && !CONFIG_CHROMA_2X2
Jingning Han's avatar
Jingning Han committed
893

Geza Lore's avatar
Geza Lore committed
894 895 896
    unsigned int sse;
    int rate;
    int64_t dist;
Jingning Han's avatar
Jingning Han committed
897

898 899
#if CONFIG_CB4X4
    if (x->skip_chroma_rd && plane) continue;
900
#endif  // CONFIG_CB4X4
901

Geza Lore's avatar
Geza Lore committed
902 903
    // TODO(geza): Write direct sse functions that do not compute
    // variance as well.
904 905
    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
906

907
    if (plane == 0) x->pred_sse[ref] = sse;
Jingning Han's avatar
Jingning Han committed
908

Geza Lore's avatar
Geza Lore committed
909
    total_sse += sse;
Jingning Han's avatar
Jingning Han committed
910

Geza Lore's avatar
Geza Lore committed
911
    model_rd_from_sse(cpi, xd, bs, plane, sse, &rate, &dist);
Geza Lore's avatar
Geza Lore committed
912 913 914

    rate_sum += rate;
    dist_sum += dist;
Jingning Han's avatar
Jingning Han committed
915 916
  }

Geza Lore's avatar
Geza Lore committed
917
  *skip_txfm_sb = total_sse == 0;
Jingning Han's avatar
Jingning Han committed
918 919
  *skip_sse_sb = total_sse << 4;
  *out_rate_sum = (int)rate_sum;
Geza Lore's avatar
Geza Lore committed
920
  *out_dist_sum = dist_sum;
Jingning Han's avatar
Jingning Han committed
921 922
}

Yaowu Xu's avatar
Yaowu Xu committed
923 924
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
925 926 927 928 929
  int i;
  int64_t error = 0, sqcoeff = 0;

  for (i = 0; i < block_size; i++) {
    const int diff = coeff[i] - dqcoeff[i];
930
    error += diff * diff;
Jingning Han's avatar
Jingning Han committed
931 932 933 934 935 936 937
    sqcoeff += coeff[i] * coeff[i];
  }

  *ssz = sqcoeff;
  return error;
}

Yaowu Xu's avatar
Yaowu Xu committed
938 939
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
940 941 942 943 944
  int i;
  int64_t error = 0;

  for (i = 0; i < block_size; i++) {
    const int diff = coeff[i] - dqcoeff[i];
945
    error += diff * diff;
Jingning Han's avatar
Jingning Han committed
946 947 948 949 950
  }

  return error;
}

951
#if CONFIG_HIGHBITDEPTH
Yaowu Xu's avatar
Yaowu Xu committed
952 953 954
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
955 956 957 958 959 960 961
  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];
962
    error += diff * diff;
Jingning Han's avatar
Jingning Han committed
963 964 965 966 967 968 969 970 971
    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;
}
972
#endif  // CONFIG_HIGHBITDEPTH
Jingning Han's avatar
Jingning Han committed
973

Thomas Daede's avatar
Thomas Daede committed
974 975 976 977 978 979 980 981 982 983
#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).

984
#if CONFIG_HIGHBITDEPTH
Thomas Daede's avatar
Thomas Daede committed
985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021
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;
}
1022
#endif  // CONFIG_HIGHBITDEPTH
Thomas Daede's avatar
Thomas Daede committed
1023 1024
#endif  // CONFIG_PVQ

1025
#if !CONFIG_PVQ || CONFIG_VAR_TX
1026
/* The trailing '0' is a terminator which is used inside av1_cost_coeffs() to
Jingning Han's avatar
Jingning Han committed
1027 1028 1029 1030
 * 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). */
1031 1032 1033 1034 1035
#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
1036 1037 1038 1039 1040
  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;
1041
  const uint16_t *band_count = &band_count_table[tx_size][1];
Jingning Han's avatar
Jingning Han committed
1042 1043
  const int eob = p->eobs[block];
  const tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
1044
  const int tx_size_ctx = txsize_sqr_map[tx_size];
1045 1046
  unsigned int(*token_costs)[2][COEFF_CONTEXTS][ENTROPY_TOKENS] =
      x->token_costs[tx_size_ctx][type][is_inter_block(mbmi)];
1047
  uint8_t token_cache[MAX_TX_SQUARE];
1048
  int pt = combine_entropy_contexts(*a, *l);
Jingning Han's avatar
Jingning Han committed
1049
  int c, cost;
1050 1051
  const int16_t *scan = scan_order->scan;
  const int16_t *nb = scan_order->neighbors;
1052 1053 1054
  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
1055

1056
#if CONFIG_HIGHBITDEPTH
1057
  const int cat6_bits = av1_get_cat6_extrabits_size(tx_size, xd->bd);
Jingning Han's avatar
Jingning Han committed
1058
#else
1059
  const int cat6_bits = av1_get_cat6_extrabits_size(tx_size, 8);
1060
#endif  // CONFIG_HIGHBITDEPTH
Jingning Han's avatar
Jingning Han committed
1061

1062
#if !CONFIG_VAR_TX && !CONFIG_SUPERTX
Jingning Han's avatar
Jingning Han committed
1063
  // Check for consistency of tx_size with mode info
Angie Chiang's avatar
Angie Chiang committed
1064
  assert(tx_size == get_tx_size(plane, xd));
1065
#endif  // !CONFIG_VAR_TX && !CONFIG_SUPERTX
1066
  (void)cm;
Jingning Han's avatar
Jingning Han committed
1067 1068 1069

  if (eob == 0) {
    // single eob token
1070
    cost = av1_cost_bit(blockz_probs[pt], 0);
Jingning Han's avatar
Jingning Han committed
1071
  } else {
Julia Robson's avatar
Julia Robson committed
1072 1073 1074 1075 1076 1077
    if (use_fast_coef_costing) {
      int band_left = *band_count++;

      // dc token
      int v = qcoeff[0];
      int16_t prev_t;
1078
      cost = av1_get_token_cost(v, &prev_t, cat6_bits);
Thomas Davies's avatar
Thomas Davies committed
1079
      cost += (*token_costs)[!prev_t][pt][prev_t];
Julia Robson's avatar
Julia Robson committed
1080

Yaowu Xu's avatar
Yaowu Xu committed
1081
      token_cache[0] = av1_pt_energy_class[prev_t];
Julia Robson's avatar
Julia Robson committed
1082 1083 1084 1085 1086 1087 1088 1089
      ++token_costs;

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

        v = qcoeff[rc];
1090
        cost += av1_get_token_cost(v, &t, cat6_bits);
Thomas Davies's avatar
Thomas Davies committed
1091
        cost += (*token_costs)[!t][!prev_t][t];
Julia Robson's avatar
Julia Robson committed
1092 1093 1094 1095 1096
        prev_t = t;
        if (!--band_left) {
          band_left = *band_count++;
          ++token_costs;
        }
Jingning Han's avatar
Jingning Han committed
1097 1098
      }

Julia Robson's avatar
Julia Robson committed
1099
      // eob token
1100
      cost += (*token_costs)[0][!prev_t][EOB_TOKEN];
Julia Robson's avatar
Julia Robson committed
1101 1102 1103 1104 1105 1106 1107

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

      // dc token
      int v = qcoeff[0];
      int16_t tok;
1108
      cost = av1_get_token_cost(v, &tok, cat6_bits);
Thomas Davies's avatar
Thomas Davies committed
1109
      cost += (*token_costs)[!tok][pt][tok];
Julia Robson's avatar
Julia Robson committed
1110

Yaowu Xu's avatar
Yaowu Xu committed
1111
      token_cache[0] = av1_pt_energy_class[tok];
Julia Robson's avatar
Julia Robson committed
1112 1113 1114 1115 1116 1117 1118
      ++token_costs;

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

        v = qcoeff[rc];