rdopt.c 446 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"
Yue Chen's avatar
Yue Chen committed
36 37 38
#if CONFIG_WARPED_MOTION
#include "av1/common/warped_motion.h"
#endif  // CONFIG_WARPED_MOTION
Jingning Han's avatar
Jingning Han committed
39

Jingning Han's avatar
Jingning Han committed
40
#include "av1/encoder/aq_variance.h"
41 42 43 44 45 46
#include "av1/encoder/cost.h"
#include "av1/encoder/encodemb.h"
#include "av1/encoder/encodemv.h"
#include "av1/encoder/encoder.h"
#include "av1/encoder/hybrid_fwd_txfm.h"
#include "av1/encoder/mcomp.h"
47
#if CONFIG_PALETTE
48
#include "av1/encoder/palette.h"
49
#endif  // CONFIG_PALETTE
50 51 52 53
#include "av1/encoder/quantize.h"
#include "av1/encoder/ratectrl.h"
#include "av1/encoder/rd.h"
#include "av1/encoder/rdopt.h"
54
#include "av1/encoder/tokenize.h"
55 56 57
#if CONFIG_PVQ
#include "av1/encoder/pvq_encoder.h"
#endif
Yushin Cho's avatar
Yushin Cho committed
58 59 60
#if CONFIG_PVQ || CONFIG_DAALA_DIST
#include "av1/common/pvq.h"
#endif
61
#if CONFIG_DUAL_FILTER
Angie Chiang's avatar
Angie Chiang committed
62 63
#define DUAL_FILTER_SET_SIZE (SWITCHABLE_FILTERS * SWITCHABLE_FILTERS)
static const int filter_sets[DUAL_FILTER_SET_SIZE][2] = {
64 65 66
  { 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 },
67
};
Angie Chiang's avatar
Angie Chiang committed
68
#endif  // CONFIG_DUAL_FILTER
69

70 71
#if CONFIG_EXT_REFS

72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98
#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
99

100 101
#endif  // CONFIG_EXT_REFS

102
#if CONFIG_EXT_REFS
103
#define SECOND_REF_FRAME_MASK ((1 << ALTREF_FRAME) | (1 << BWDREF_FRAME) | 0x01)
104
#else
105
#define SECOND_REF_FRAME_MASK ((1 << ALTREF_FRAME) | 0x01)
106
#endif  // CONFIG_EXT_REFS
Jingning Han's avatar
Jingning Han committed
107

108 109
#define MIN_EARLY_TERM_INDEX 3
#define NEW_MV_DISCOUNT_FACTOR 8
Jingning Han's avatar
Jingning Han committed
110

111 112 113 114 115
#if CONFIG_EXT_INTRA
#define ANGLE_SKIP_THRESH 10
#define FILTER_FAST_SEARCH 1
#endif  // CONFIG_EXT_INTRA

116 117
const double ADST_FLIP_SVM[8] = { -6.6623, -2.8062, -3.2531, 3.1671,    // vert
                                  -7.7051, -3.2234, -3.6193, 3.4533 };  // horz
118

Jingning Han's avatar
Jingning Han committed
119 120 121 122 123
typedef struct {
  PREDICTION_MODE mode;
  MV_REFERENCE_FRAME ref_frame[2];
} MODE_DEFINITION;

124
typedef struct { MV_REFERENCE_FRAME ref_frame[2]; } REF_DEFINITION;
Jingning Han's avatar
Jingning Han committed
125 126

struct rdcost_block_args {
Yaowu Xu's avatar
Yaowu Xu committed
127
  const AV1_COMP *cpi;
Jingning Han's avatar
Jingning Han committed
128
  MACROBLOCK *x;
129 130
  ENTROPY_CONTEXT t_above[2 * MAX_MIB_SIZE];
  ENTROPY_CONTEXT t_left[2 * MAX_MIB_SIZE];
131
  RD_STATS rd_stats;
Jingning Han's avatar
Jingning Han committed
132 133 134 135
  int64_t this_rd;
  int64_t best_rd;
  int exit_early;
  int use_fast_coef_costing;
136
  const SCAN_ORDER *scan_order;
Jingning Han's avatar
Jingning Han committed
137 138 139
};

#define LAST_NEW_MV_INDEX 6
Yaowu Xu's avatar
Yaowu Xu committed
140
static const MODE_DEFINITION av1_mode_order[MAX_MODES] = {
Emil Keyder's avatar
Emil Keyder committed
141
  { NEARESTMV, { LAST_FRAME, NONE_FRAME } },
142
#if CONFIG_EXT_REFS
Emil Keyder's avatar
Emil Keyder committed
143 144 145
  { NEARESTMV, { LAST2_FRAME, NONE_FRAME } },
  { NEARESTMV, { LAST3_FRAME, NONE_FRAME } },
  { NEARESTMV, { BWDREF_FRAME, NONE_FRAME } },
146
#endif  // CONFIG_EXT_REFS
Emil Keyder's avatar
Emil Keyder committed
147 148
  { NEARESTMV, { ALTREF_FRAME, NONE_FRAME } },
  { NEARESTMV, { GOLDEN_FRAME, NONE_FRAME } },
Jingning Han's avatar
Jingning Han committed
149

Emil Keyder's avatar
Emil Keyder committed
150
  { DC_PRED, { INTRA_FRAME, NONE_FRAME } },
Jingning Han's avatar
Jingning Han committed
151

Emil Keyder's avatar
Emil Keyder committed
152
  { NEWMV, { LAST_FRAME, NONE_FRAME } },
153
#if CONFIG_EXT_REFS
Emil Keyder's avatar
Emil Keyder committed
154 155 156
  { NEWMV, { LAST2_FRAME, NONE_FRAME } },
  { NEWMV, { LAST3_FRAME, NONE_FRAME } },
  { NEWMV, { BWDREF_FRAME, NONE_FRAME } },
157
#endif  // CONFIG_EXT_REFS
Emil Keyder's avatar
Emil Keyder committed
158 159
  { NEWMV, { ALTREF_FRAME, NONE_FRAME } },
  { NEWMV, { GOLDEN_FRAME, NONE_FRAME } },
Jingning Han's avatar
Jingning Han committed
160

Emil Keyder's avatar
Emil Keyder committed
161
  { NEARMV, { LAST_FRAME, NONE_FRAME } },
162
#if CONFIG_EXT_REFS
Emil Keyder's avatar
Emil Keyder committed
163 164 165
  { NEARMV, { LAST2_FRAME, NONE_FRAME } },
  { NEARMV, { LAST3_FRAME, NONE_FRAME } },
  { NEARMV, { BWDREF_FRAME, NONE_FRAME } },
166
#endif  // CONFIG_EXT_REFS
Emil Keyder's avatar
Emil Keyder committed
167 168
  { NEARMV, { ALTREF_FRAME, NONE_FRAME } },
  { NEARMV, { GOLDEN_FRAME, NONE_FRAME } },
Jingning Han's avatar
Jingning Han committed
169

Yue Chen's avatar
Yue Chen committed
170
#if CONFIG_EXT_INTER
Emil Keyder's avatar
Emil Keyder committed
171
  { NEWFROMNEARMV, { LAST_FRAME, NONE_FRAME } },
172
#if CONFIG_EXT_REFS
Emil Keyder's avatar
Emil Keyder committed
173 174 175
  { NEWFROMNEARMV, { LAST2_FRAME, NONE_FRAME } },
  { NEWFROMNEARMV, { LAST3_FRAME, NONE_FRAME } },
  { NEWFROMNEARMV, { BWDREF_FRAME, NONE_FRAME } },
176
#endif  // CONFIG_EXT_REFS
Emil Keyder's avatar
Emil Keyder committed
177 178
  { NEWFROMNEARMV, { ALTREF_FRAME, NONE_FRAME } },
  { NEWFROMNEARMV, { GOLDEN_FRAME, NONE_FRAME } },
Yue Chen's avatar
Yue Chen committed
179 180
#endif  // CONFIG_EXT_INTER

Emil Keyder's avatar
Emil Keyder committed
181
  { ZEROMV, { LAST_FRAME, NONE_FRAME } },
182
#if CONFIG_EXT_REFS
Emil Keyder's avatar
Emil Keyder committed
183 184 185
  { ZEROMV, { LAST2_FRAME, NONE_FRAME } },
  { ZEROMV, { LAST3_FRAME, NONE_FRAME } },
  { ZEROMV, { BWDREF_FRAME, NONE_FRAME } },
186
#endif  // CONFIG_EXT_REFS
Emil Keyder's avatar
Emil Keyder committed
187 188
  { ZEROMV, { GOLDEN_FRAME, NONE_FRAME } },
  { ZEROMV, { ALTREF_FRAME, NONE_FRAME } },
Jingning Han's avatar
Jingning Han committed
189

190
// TODO(zoeliu): May need to reconsider the order on the modes to check
191

192
#if CONFIG_EXT_INTER
193
  { NEAREST_NEARESTMV, { LAST_FRAME, ALTREF_FRAME } },
194
#if CONFIG_EXT_REFS
195 196
  { NEAREST_NEARESTMV, { LAST2_FRAME, ALTREF_FRAME } },
  { NEAREST_NEARESTMV, { LAST3_FRAME, ALTREF_FRAME } },
197
#endif  // CONFIG_EXT_REFS
198
  { NEAREST_NEARESTMV, { GOLDEN_FRAME, ALTREF_FRAME } },
199
#if CONFIG_EXT_REFS
200 201 202 203
  { NEAREST_NEARESTMV, { LAST_FRAME, BWDREF_FRAME } },
  { NEAREST_NEARESTMV, { LAST2_FRAME, BWDREF_FRAME } },
  { NEAREST_NEARESTMV, { LAST3_FRAME, BWDREF_FRAME } },
  { NEAREST_NEARESTMV, { GOLDEN_FRAME, BWDREF_FRAME } },
204
#endif  // CONFIG_EXT_REFS
205

206
#else  // CONFIG_EXT_INTER
207

208
  { NEARESTMV, { LAST_FRAME, ALTREF_FRAME } },
209
#if CONFIG_EXT_REFS
210 211
  { NEARESTMV, { LAST2_FRAME, ALTREF_FRAME } },
  { NEARESTMV, { LAST3_FRAME, ALTREF_FRAME } },
212
#endif  // CONFIG_EXT_REFS
213
  { NEARESTMV, { GOLDEN_FRAME, ALTREF_FRAME } },
214
#if CONFIG_EXT_REFS
215 216 217 218
  { NEARESTMV, { LAST_FRAME, BWDREF_FRAME } },
  { NEARESTMV, { LAST2_FRAME, BWDREF_FRAME } },
  { NEARESTMV, { LAST3_FRAME, BWDREF_FRAME } },
  { NEARESTMV, { GOLDEN_FRAME, BWDREF_FRAME } },
219
#endif  // CONFIG_EXT_REFS
220
#endif  // CONFIG_EXT_INTER
Jingning Han's avatar
Jingning Han committed
221

Emil Keyder's avatar
Emil Keyder committed
222
  { TM_PRED, { INTRA_FRAME, NONE_FRAME } },
Jingning Han's avatar
Jingning Han committed
223

224
#if CONFIG_ALT_INTRA
Emil Keyder's avatar
Emil Keyder committed
225
  { SMOOTH_PRED, { INTRA_FRAME, NONE_FRAME } },
226 227
#endif  // CONFIG_ALT_INTRA

228
#if CONFIG_EXT_INTER
229 230 231 232 233 234 235 236 237
  { NEAR_NEARESTMV, { LAST_FRAME, ALTREF_FRAME } },
  { NEAREST_NEARMV, { LAST_FRAME, ALTREF_FRAME } },
  { 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 } },
238

239
#if CONFIG_EXT_REFS
240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258
  { NEAR_NEARESTMV, { LAST2_FRAME, ALTREF_FRAME } },
  { NEAREST_NEARMV, { LAST2_FRAME, ALTREF_FRAME } },
  { 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_NEARESTMV, { LAST3_FRAME, ALTREF_FRAME } },
  { NEAREST_NEARMV, { LAST3_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 } },
259
#endif  // CONFIG_EXT_REFS
260

261 262 263 264 265 266 267 268 269
  { NEAR_NEARESTMV, { GOLDEN_FRAME, ALTREF_FRAME } },
  { NEAREST_NEARMV, { GOLDEN_FRAME, ALTREF_FRAME } },
  { 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 } },
270 271

#if CONFIG_EXT_REFS
272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310
  { NEAR_NEARESTMV, { LAST_FRAME, BWDREF_FRAME } },
  { NEAREST_NEARMV, { LAST_FRAME, BWDREF_FRAME } },
  { 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_NEARESTMV, { LAST2_FRAME, BWDREF_FRAME } },
  { NEAREST_NEARMV, { LAST2_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_NEARESTMV, { LAST3_FRAME, BWDREF_FRAME } },
  { NEAREST_NEARMV, { LAST3_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_NEARESTMV, { GOLDEN_FRAME, BWDREF_FRAME } },
  { NEAREST_NEARMV, { GOLDEN_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 } },
311
#endif  // CONFIG_EXT_REFS
312 313 314

#else  // CONFIG_EXT_INTER

315 316
  { NEARMV, { LAST_FRAME, ALTREF_FRAME } },
  { NEWMV, { LAST_FRAME, ALTREF_FRAME } },
317
#if CONFIG_EXT_REFS
318 319 320 321
  { NEARMV, { LAST2_FRAME, ALTREF_FRAME } },
  { NEWMV, { LAST2_FRAME, ALTREF_FRAME } },
  { NEARMV, { LAST3_FRAME, ALTREF_FRAME } },
  { NEWMV, { LAST3_FRAME, ALTREF_FRAME } },
322
#endif  // CONFIG_EXT_REFS
323 324
  { NEARMV, { GOLDEN_FRAME, ALTREF_FRAME } },
  { NEWMV, { GOLDEN_FRAME, ALTREF_FRAME } },
325 326

#if CONFIG_EXT_REFS
327 328 329 330 331 332 333 334
  { 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 } },
335
#endif  // CONFIG_EXT_REFS
Jingning Han's avatar
Jingning Han committed
336

337
  { ZEROMV, { LAST_FRAME, ALTREF_FRAME } },
338
#if CONFIG_EXT_REFS
339 340
  { ZEROMV, { LAST2_FRAME, ALTREF_FRAME } },
  { ZEROMV, { LAST3_FRAME, ALTREF_FRAME } },
341
#endif  // CONFIG_EXT_REFS
342
  { ZEROMV, { GOLDEN_FRAME, ALTREF_FRAME } },
343 344

#if CONFIG_EXT_REFS
345 346 347 348
  { ZEROMV, { LAST_FRAME, BWDREF_FRAME } },
  { ZEROMV, { LAST2_FRAME, BWDREF_FRAME } },
  { ZEROMV, { LAST3_FRAME, BWDREF_FRAME } },
  { ZEROMV, { GOLDEN_FRAME, BWDREF_FRAME } },
349
#endif  // CONFIG_EXT_REFS
350

351
#endif  // CONFIG_EXT_INTER
Jingning Han's avatar
Jingning Han committed
352

Emil Keyder's avatar
Emil Keyder committed
353 354 355 356 357 358 359 360
  { 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 } },
361 362

#if CONFIG_EXT_INTER
363 364 365 366
  { ZEROMV, { LAST_FRAME, INTRA_FRAME } },
  { NEARESTMV, { LAST_FRAME, INTRA_FRAME } },
  { NEARMV, { LAST_FRAME, INTRA_FRAME } },
  { NEWMV, { LAST_FRAME, INTRA_FRAME } },
367 368

#if CONFIG_EXT_REFS
369 370 371 372 373 374 375 376 377
  { 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 } },
378 379
#endif  // CONFIG_EXT_REFS

380 381 382 383
  { ZEROMV, { GOLDEN_FRAME, INTRA_FRAME } },
  { NEARESTMV, { GOLDEN_FRAME, INTRA_FRAME } },
  { NEARMV, { GOLDEN_FRAME, INTRA_FRAME } },
  { NEWMV, { GOLDEN_FRAME, INTRA_FRAME } },
384

385
#if CONFIG_EXT_REFS
386 387 388 389
  { ZEROMV, { BWDREF_FRAME, INTRA_FRAME } },
  { NEARESTMV, { BWDREF_FRAME, INTRA_FRAME } },
  { NEARMV, { BWDREF_FRAME, INTRA_FRAME } },
  { NEWMV, { BWDREF_FRAME, INTRA_FRAME } },
390 391
#endif  // CONFIG_EXT_REFS

392 393 394 395
  { ZEROMV, { ALTREF_FRAME, INTRA_FRAME } },
  { NEARESTMV, { ALTREF_FRAME, INTRA_FRAME } },
  { NEARMV, { ALTREF_FRAME, INTRA_FRAME } },
  { NEWMV, { ALTREF_FRAME, INTRA_FRAME } },
396
#endif  // CONFIG_EXT_INTER
Jingning Han's avatar
Jingning Han committed
397 398
};

Yaowu Xu's avatar
Yaowu Xu committed
399
static const REF_DEFINITION av1_ref_order[MAX_REFS] = {
Emil Keyder's avatar
Emil Keyder committed
400
  { { LAST_FRAME, NONE_FRAME } },
401
#if CONFIG_EXT_REFS
Emil Keyder's avatar
Emil Keyder committed
402 403
  { { LAST2_FRAME, NONE_FRAME } },    { { LAST3_FRAME, NONE_FRAME } },
  { { BWDREF_FRAME, NONE_FRAME } },
404
#endif  // CONFIG_EXT_REFS
Emil Keyder's avatar
Emil Keyder committed
405
  { { GOLDEN_FRAME, NONE_FRAME } },   { { ALTREF_FRAME, NONE_FRAME } },
406

407
  { { LAST_FRAME, ALTREF_FRAME } },
408
#if CONFIG_EXT_REFS
409
  { { LAST2_FRAME, ALTREF_FRAME } },  { { LAST3_FRAME, ALTREF_FRAME } },
410
#endif  // CONFIG_EXT_REFS
411
  { { GOLDEN_FRAME, ALTREF_FRAME } },
412 413

#if CONFIG_EXT_REFS
414 415
  { { LAST_FRAME, BWDREF_FRAME } },   { { LAST2_FRAME, BWDREF_FRAME } },
  { { LAST3_FRAME, BWDREF_FRAME } },  { { GOLDEN_FRAME, BWDREF_FRAME } },
416
#endif  // CONFIG_EXT_REFS
417

Emil Keyder's avatar
Emil Keyder committed
418
  { { INTRA_FRAME, NONE_FRAME } },
Jingning Han's avatar
Jingning Han committed
419 420
};

421
#if CONFIG_EXT_INTRA || CONFIG_FILTER_INTRA || CONFIG_PALETTE
hui su's avatar
hui su committed
422 423
static INLINE int write_uniform_cost(int n, int v) {
  int l = get_unsigned_bits(n), m = (1 << l) - n;
424
  if (l == 0) return 0;
hui su's avatar
hui su committed
425
  if (v < m)
Yaowu Xu's avatar
Yaowu Xu committed
426
    return (l - 1) * av1_cost_bit(128, 0);
hui su's avatar
hui su committed
427
  else
Yaowu Xu's avatar
Yaowu Xu committed
428
    return l * av1_cost_bit(128, 0);
hui su's avatar
hui su committed
429
}
430
#endif  // CONFIG_EXT_INTRA || CONFIG_FILTER_INTRA || CONFIG_PALETTE
hui su's avatar
hui su committed
431

432 433 434
// 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
435 436 437 438
#define FAST_EXT_TX_CORR_MARGIN 0.5
#define FAST_EXT_TX_EDST_MARGIN 0.3

static const TX_TYPE_1D vtx_tab[TX_TYPES] = {
439
  DCT_1D,      ADST_1D, DCT_1D,      ADST_1D,
440
#if CONFIG_EXT_TX
441 442
  FLIPADST_1D, DCT_1D,  FLIPADST_1D, ADST_1D, FLIPADST_1D, IDTX_1D,
  DCT_1D,      IDTX_1D, ADST_1D,     IDTX_1D, FLIPADST_1D, IDTX_1D,
443 444 445 446
#endif  // CONFIG_EXT_TX
};

static const TX_TYPE_1D htx_tab[TX_TYPES] = {
447
  DCT_1D,  DCT_1D,      ADST_1D,     ADST_1D,
448
#if CONFIG_EXT_TX
449 450
  DCT_1D,  FLIPADST_1D, FLIPADST_1D, FLIPADST_1D, ADST_1D, IDTX_1D,
  IDTX_1D, DCT_1D,      IDTX_1D,     ADST_1D,     IDTX_1D, FLIPADST_1D,
451 452
#endif  // CONFIG_EXT_TX
};
453

Yushin Cho's avatar
Yushin Cho committed
454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617
#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;
}

static double od_compute_dist_8x8(int qm, int use_activity_masking, od_coeff *x,
                                  od_coeff *y, int stride) {
  od_coeff e[8 * 8];
  od_coeff et[8 * 8];
  int16_t src[8 * 8];
  tran_low_t coeff[8 * 8];
  double sum;
  int min_var;
  double mean_var;
  double var_stat;
  double activity;
  double calibration;
  int i;
  int j;
  double vardist;
  FWD_TXFM_PARAM fwd_txfm_param;

  vardist = 0;
  OD_ASSERT(qm != OD_FLAT_QM);
#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;
#endif
  sum = 0;
  for (i = 0; i < 8; i++) {
    for (j = 0; j < 8; j++)
      e[8 * i + j] = x[i * stride + j] - y[i * stride + j];
  }

  for (i = 0; i < 8; i++)
    for (j = 0; j < 8; j++) src[8 * i + j] = e[8 * i + j];

  fwd_txfm_param.tx_type = DCT_DCT;
  fwd_txfm_param.tx_size = TX_8X8;
  fwd_txfm_param.lossless = 0;

  fwd_txfm(&src[0], &coeff[0], 8, &fwd_txfm_param);

  for (i = 0; i < 8; i++)
    for (j = 0; j < 8; j++) et[8 * i + j] = coeff[8 * i + j] >> 3;

  sum = 0;
  for (i = 0; i < 8; i++)
    for (j = 0; j < 8; j++)
      sum += et[8 * i + j] * (double)et[8 * i + j] * 16. /
             OD_QM8_Q4_HVS[i * 8 + j];

  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;

  (void)qindex;

  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 {
    for (i = 0; i < bsize_h; i += 8) {
      int j;
      for (j = 0; j < bsize_w; j += 8) {
        sum += od_compute_dist_8x8(qm, activity_masking, &x[i * bsize_w + j],
                                   &y[i * bsize_w + j], bsize_w);
      }
    }
    /* Compensate for the fact that the quantization matrix lowers the
       distortion value. We tried a half-dozen values and picked the one where
       we liked the ntt-short1 curves best. The tuning is approximate since
       the different metrics go in different directions. */
    /*Start interpolation at coded_quantizer 1.7=f(36) and end it at 1.2=f(47)*/
    // TODO(yushin): Check whether qindex of AV1 work here, replacing daala's
    // coded_quantizer.
    /*sum *= qindex >= 47 ? 1.2 :
        qindex <= 36 ? 1.7 :
     1.7 + (1.2 - 1.7)*(qindex - 36)/(47 - 36);*/
  }
  return sum;
}

static int64_t av1_daala_dist(const uint8_t *src, int src_stride,
                              const uint8_t *dst, int dst_stride, int tx_size,
                              int qm, int use_activity_masking, int qindex) {
  int i, j;
  int64_t d;
  const BLOCK_SIZE tx_bsize = txsize_to_bsize[tx_size];
  const int bsw = block_size_wide[tx_bsize];
  const int bsh = block_size_high[tx_bsize];
  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;
}
#endif  // #if CONFIG_DAALA_DIST

Yaowu Xu's avatar
Yaowu Xu committed
618
static void get_energy_distribution_fine(const AV1_COMP *cpi, BLOCK_SIZE bsize,
619 620 621
                                         uint8_t *src, int src_stride,
                                         uint8_t *dst, int dst_stride,
                                         double *hordist, double *verdist) {
622 623
  int bw = block_size_wide[bsize];
  int bh = block_size_high[bsize];
624
  unsigned int esq[16] = { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 };
625 626
  unsigned int var[16];
  double total = 0;
627

628
  const int f_index = bsize - BLOCK_16X16;
629 630 631 632
  if (f_index < 0) {
    int i, j, index;
    int w_shift = bw == 8 ? 1 : 2;
    int h_shift = bh == 8 ? 1 : 2;
Yaowu Xu's avatar
Yaowu Xu committed
633
#if CONFIG_AOM_HIGHBITDEPTH
634 635 636 637 638 639
    if (cpi->common.use_highbitdepth) {
      uint16_t *src16 = CONVERT_TO_SHORTPTR(src);
      uint16_t *dst16 = CONVERT_TO_SHORTPTR(dst);
      for (i = 0; i < bh; ++i)
        for (j = 0; j < bw; ++j) {
          index = (j >> w_shift) + ((i >> h_shift) << 2);
640 641 642
          esq[index] +=
              (src16[j + i * src_stride] - dst16[j + i * dst_stride]) *
              (src16[j + i * src_stride] - dst16[j + i * dst_stride]);
643 644
        }
    } else {
Yaowu Xu's avatar
Yaowu Xu committed
645
#endif  // CONFIG_AOM_HIGHBITDEPTH
646 647 648 649 650 651 652

      for (i = 0; i < bh; ++i)
        for (j = 0; j < bw; ++j) {
          index = (j >> w_shift) + ((i >> h_shift) << 2);
          esq[index] += (src[j + i * src_stride] - dst[j + i * dst_stride]) *
                        (src[j + i * src_stride] - dst[j + i * dst_stride]);
        }
Yaowu Xu's avatar
Yaowu Xu committed
653
#if CONFIG_AOM_HIGHBITDEPTH
654
    }
Yaowu Xu's avatar
Yaowu Xu committed
655
#endif  // CONFIG_AOM_HIGHBITDEPTH
656
  } else {
657 658 659 660 661
    var[0] = cpi->fn_ptr[f_index].vf(src, src_stride, dst, dst_stride, &esq[0]);
    var[1] = cpi->fn_ptr[f_index].vf(src + bw / 4, src_stride, dst + bw / 4,
                                     dst_stride, &esq[1]);
    var[2] = cpi->fn_ptr[f_index].vf(src + bw / 2, src_stride, dst + bw / 2,
                                     dst_stride, &esq[2]);
662 663 664 665 666
    var[3] = cpi->fn_ptr[f_index].vf(src + 3 * bw / 4, src_stride,
                                     dst + 3 * bw / 4, dst_stride, &esq[3]);
    src += bh / 4 * src_stride;
    dst += bh / 4 * dst_stride;

667 668 669 670 671
    var[4] = cpi->fn_ptr[f_index].vf(src, src_stride, dst, dst_stride, &esq[4]);
    var[5] = cpi->fn_ptr[f_index].vf(src + bw / 4, src_stride, dst + bw / 4,
                                     dst_stride, &esq[5]);
    var[6] = cpi->fn_ptr[f_index].vf(src + bw / 2, src_stride, dst + bw / 2,
                                     dst_stride, &esq[6]);
672 673 674 675 676
    var[7] = cpi->fn_ptr[f_index].vf(src + 3 * bw / 4, src_stride,
                                     dst + 3 * bw / 4, dst_stride, &esq[7]);
    src += bh / 4 * src_stride;
    dst += bh / 4 * dst_stride;

677 678 679 680 681
    var[8] = cpi->fn_ptr[f_index].vf(src, src_stride, dst, dst_stride, &esq[8]);
    var[9] = cpi->fn_ptr[f_index].vf(src + bw / 4, src_stride, dst + bw / 4,
                                     dst_stride, &esq[9]);
    var[10] = cpi->fn_ptr[f_index].vf(src + bw / 2, src_stride, dst + bw / 2,
                                      dst_stride, &esq[10]);
682 683 684 685 686
    var[11] = cpi->fn_ptr[f_index].vf(src + 3 * bw / 4, src_stride,
                                      dst + 3 * bw / 4, dst_stride, &esq[11]);
    src += bh / 4 * src_stride;
    dst += bh / 4 * dst_stride;

687 688 689 690 691 692
    var[12] =
        cpi->fn_ptr[f_index].vf(src, src_stride, dst, dst_stride, &esq[12]);
    var[13] = cpi->fn_ptr[f_index].vf(src + bw / 4, src_stride, dst + bw / 4,
                                      dst_stride, &esq[13]);
    var[14] = cpi->fn_ptr[f_index].vf(src + bw / 2, src_stride, dst + bw / 2,
                                      dst_stride, &esq[14]);
693 694 695 696
    var[15] = cpi->fn_ptr[f_index].vf(src + 3 * bw / 4, src_stride,
                                      dst + 3 * bw / 4, dst_stride, &esq[15]);
  }

697 698 699
  total = 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];
700 701
  if (total > 0) {
    const double e_recip = 1.0 / total;
702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719
    hordist[0] =
        ((double)esq[0] + (double)esq[4] + (double)esq[8] + (double)esq[12]) *
        e_recip;
    hordist[1] =
        ((double)esq[1] + (double)esq[5] + (double)esq[9] + (double)esq[13]) *
        e_recip;
    hordist[2] =
        ((double)esq[2] + (double)esq[6] + (double)esq[10] + (double)esq[14]) *
        e_recip;
    verdist[0] =
        ((double)esq[0] + (double)esq[1] + (double)esq[2] + (double)esq[3]) *
        e_recip;
    verdist[1] =
        ((double)esq[4] + (double)esq[5] + (double)esq[6] + (double)esq[7]) *
        e_recip;
    verdist[2] =
        ((double)esq[8] + (double)esq[9] + (double)esq[10] + (double)esq[11]) *
        e_recip;
720 721 722 723 724
  } else {
    hordist[0] = verdist[0] = 0.25;
    hordist[1] = verdist[1] = 0.25;
    hordist[2] = verdist[2] = 0.25;
  }
725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740
  (void)var[0];
  (void)var[1];
  (void)var[2];
  (void)var[3];
  (void)var[4];
  (void)var[5];
  (void)var[6];
  (void)var[7];
  (void)var[8];
  (void)var[9];
  (void)var[10];
  (void)var[11];
  (void)var[12];
  (void)var[13];
  (void)var[14];
  (void)var[15];
741 742
}

Yaowu Xu's avatar
Yaowu Xu committed
743 744 745
static int adst_vs_flipadst(const AV1_COMP *cpi, BLOCK_SIZE bsize, uint8_t *src,
                            int src_stride, uint8_t *dst, int dst_stride,
                            double *hdist, double *vdist) {
746 747
  int prune_bitmask = 0;
  double svm_proj_h = 0, svm_proj_v = 0;
748 749
  get_energy_distribution_fine(cpi, bsize, src, src_stride, dst, dst_stride,
                               hdist, vdist);
750

751
  svm_proj_v = vdist[0] * ADST_FLIP_SVM[0] + vdist[1] * ADST_FLIP_SVM[1] +
752
               vdist[2] * ADST_FLIP_SVM[2] + ADST_FLIP_SVM[3];
753
  svm_proj_h = hdist[0] * ADST_FLIP_SVM[4] + hdist[1] * ADST_FLIP_SVM[5] +
754 755 756 757 758 759 760 761 762 763 764 765 766 767 768
               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
769
static void get_horver_correlation(int16_t *diff, int stride, int w, int h,
770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797
                                   double *hcorr, double *vcorr) {
  // 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;
    }
  }
798 799 800
  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;
801 802 803 804 805 806 807 808 809 810 811 812
  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;
  }
}

813 814
int dct_vs_idtx(int16_t *diff, int stride, int w, int h, double *hcorr,
                double *vcorr) {
815 816 817 818
  int prune_bitmask = 0;
  get_horver_correlation(diff, stride, w, h, hcorr, vcorr);

  if (*vcorr > FAST_EXT_TX_CORR_MID + FAST_EXT_TX_CORR_MARGIN)
819
    prune_bitmask |= 1 << IDTX_1D;
820 821 822 823
  else if (*vcorr < FAST_EXT_TX_CORR_MID - FAST_EXT_TX_CORR_MARGIN)
    prune_bitmask |= 1 << DCT_1D;

  if (*hcorr > FAST_EXT_TX_CORR_MID + FAST_EXT_TX_CORR_MARGIN)
824
    prune_bitmask |= 1 << (IDTX_1D + 8);
825 826 827 828 829 830
  else if (*hcorr < FAST_EXT_TX_CORR_MID - FAST_EXT_TX_CORR_MARGIN)
    prune_bitmask |= 1 << (DCT_1D + 8);
  return prune_bitmask;
}

// Performance drop: 0.5%, Speed improvement: 24%
Yaowu Xu's avatar
Yaowu Xu committed
831
static int prune_two_for_sby(const AV1_COMP *cpi, BLOCK_SIZE bsize,
832
                             MACROBLOCK *x, MACROBLOCKD *xd, int adst_flipadst,
833
                             int dct_idtx) {
834 835 836 837 838
  struct macroblock_plane *const p = &x->plane[0];
  struct macroblockd_plane *const pd = &xd->plane[0];
  const BLOCK_SIZE bs = get_plane_block_size(bsize, pd);
  const int bw = 4 << (b_width_log2_lookup[bs]);
  const int bh = 4 << (b_height_log2_lookup[bs]);
839
  double hdist[3] = { 0, 0, 0 }, vdist[3] = { 0, 0, 0 };
840
  double hcorr, vcorr;
841
  int prune = 0;
Yaowu Xu's avatar
Yaowu Xu committed
842
  av1_subtract_plane(x, bsize, 0);
843

844 845 846
  if (adst_flipadst)
    prune |= adst_vs_flipadst(cpi, bsize, p->src.buf, p->src.stride,
                              pd->dst.buf, pd->dst.stride, hdist, vdist);
847
  if (dct_idtx) prune |= dct_vs_idtx(p->src_diff, bw, bw, bh, &hcorr, &vcorr);
848 849 850

  return prune;
}
851 852
#endif  // CONFIG_EXT_TX

853
// Performance drop: 0.3%, Speed improvement: 5%
Yaowu Xu's avatar
Yaowu Xu committed
854
static int prune_one_for_sby(const AV1_COMP *cpi, BLOCK_SIZE bsize,
855
                             MACROBLOCK *x, MACROBLOCKD *xd) {
856 857
  struct macroblock_plane *const p = &x->plane[0];
  struct macroblockd_plane *const pd = &xd->plane[0];
858
  double hdist[3] = { 0, 0, 0 }, vdist[3] = { 0, 0, 0 };
Yaowu Xu's avatar
Yaowu Xu committed
859
  av1_subtract_plane(x, bsize, 0);
860 861
  return adst_vs_flipadst(cpi, bsize, p->src.buf, p->src.stride, pd->dst.buf,
                          pd->dst.stride, hdist, vdist);
862 863
}

Yaowu Xu's avatar
Yaowu Xu committed
864
static int prune_tx_types(const AV1_COMP *cpi, BLOCK_SIZE bsize, MACROBLOCK *x,
865 866 867 868
                          MACROBLOCKD *xd, int tx_set) {
#if CONFIG_EXT_TX
  const int *tx_set_1D = ext_tx_used_inter_1D[tx_set];
#else
869
  const int tx_set_1D[TX_TYPES_1D] = { 0 };
870 871
#endif

872
  switch (cpi->sf.tx_type_search.prune_mode) {
873 874
    case NO_PRUNE: return 0; break;
    case PRUNE_ONE:
875
      if ((tx_set >= 0) && !(tx_set_1D[FLIPADST_1D] & tx_set_1D[ADST_1D]))
876
        return 0;
877 878
      return prune_one_for_sby(cpi, bsize, x, xd);
      break;
879 880
#if CONFIG_EXT_TX
    case PRUNE_TWO:
881
      if ((tx_set >= 0) && !(tx_set_1D[FLIPADST_1D] & tx_set_1D[ADST_1D])) {
882
        if (!(tx_set_1D[DCT_1D] & tx_set_1D[IDTX_1D])) return 0;
883 884
        return prune_two_for_sby(cpi, bsize, x, xd, 0, 1);
      }
885
      if ((tx_set >= 0) && !(tx_set_1D[DCT_1D] & tx_set_1D[IDTX_1D]))
886 887
        return prune_two_for_sby(cpi, bsize, x, xd, 1, 0);
      return prune_two_for_sby(cpi, bsize, x, xd, 1, 1);
888
      break;
889
#endif
890 891 892 893 894
  }
  assert(0);
  return 0;
}

895
static int do_tx_type_search(TX_TYPE tx_type, int prune) {
896 897 898
// TODO(sarahparker) implement for non ext tx
#if CONFIG_EXT_TX
  return !(((prune >> vtx_tab[tx_type]) & 1) |
899
           ((prune >> (htx_tab[tx_type] + 8)) & 1));
900 901
#else
  // temporary to avoid compiler warnings
902 903 904 905
  (void)vtx_tab;
  (void)htx_tab;
  (void)tx_type;
  (void)prune;
906 907 908 909
  return 1;
#endif
}

Yaowu Xu's avatar
Yaowu Xu committed
910
static void model_rd_from_sse(const AV1_COMP *const cpi,
911 912
                              const MACROBLOCKD *const xd, BLOCK_SIZE bsize,
                              int plane, int64_t sse, int *rate,
Geza Lore's avatar
Geza Lore committed
913 914 915
                              int64_t *dist) {
  const struct macroblockd_plane *const pd = &xd->plane[plane];
  const int dequant_shift =
Yaowu Xu's avatar
Yaowu Xu committed
916
#if CONFIG_AOM_HIGHBITDEPTH
917
      (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) ? xd->bd - 5 :
Yaowu Xu's avatar
Yaowu Xu committed
918
#endif  // CONFIG_AOM_HIGHBITDEPTH
919
                                                    3;
Geza Lore's avatar
Geza Lore committed
920 921 922 923 924 925 926

  // 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
927
      *rate = (int)((square_error * (280 - quantizer)) >>
Yaowu Xu's avatar
Yaowu Xu committed
928
                    (16 - AV1_PROB_COST_SHIFT));
Geza Lore's avatar
Geza Lore committed
929 930 931 932
    else
      *rate = 0;
    *dist = (square_error * quantizer) >> 8;
  } else {
Yaowu Xu's avatar
Yaowu Xu committed
933 934
    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
935 936 937 938 939
  }

  *dist <<= 4;
}

Yaowu Xu's avatar
Yaowu Xu committed
940
static void model_rd_for_sb(const AV1_COMP *const cpi, BLOCK_SIZE bsize,
941 942 943 944
                            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
945 946 947
  // 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
948 949 950
  int plane;
  const int ref = xd->mi[0]->mbmi.ref_frame[0];

Jingning Han's avatar
Jingning Han committed
951 952 953 954 955 956
  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
957 958 959
  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
960 961
    const BLOCK_SIZE bs = get_plane_block_size(bsize, pd);

Geza Lore's avatar
Geza Lore committed
962 963 964
    unsigned int sse;
    int rate;
    int64_t dist;
Jingning Han's avatar
Jingning Han committed
965

Geza Lore's avatar
Geza Lore committed
966 967
    // TODO(geza): Write direct sse functions that do not compute
    // variance as well.
968 969
    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
970

971
    if (plane == 0) x->pred_sse[ref] = sse;
Jingning Han's avatar
Jingning Han committed
972

Geza Lore's avatar
Geza Lore committed
973
    total_sse += sse;
Jingning Han's avatar
Jingning Han committed
974

Geza Lore's avatar
Geza Lore committed
975
    model_rd_from_sse(cpi, xd, bs, plane, sse, &rate, &dist);
Geza Lore's avatar
Geza Lore committed
976 977 978

    rate_sum += rate;
    dist_sum += dist;
Jingning Han's avatar
Jingning Han committed
979 980
  }

Geza Lore's avatar
Geza Lore committed
981
  *skip_txfm_sb = total_sse == 0;
Jingning Han's avatar
Jingning Han committed
982 983
  *skip_sse_sb = total_sse << 4;
  *out_rate_sum = (int)rate_sum;
Geza Lore's avatar
Geza Lore committed
984
  *out_dist_sum = dist_sum;
Jingning Han's avatar
Jingning Han committed
985 986
}

987 988 989 990 991 992 993 994 995 996 997 998
#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).

// 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.
Yushin Cho's avatar
Yushin Cho committed
999 1000 1001 1002
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) {
1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014
  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;
}
#endif

Yaowu Xu's avatar
Yaowu Xu committed
1015 1016
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
1017 1018 1019 1020 1021
  int i;
  int64_t error = 0, sqcoeff = 0;

  for (i = 0; i < block_size; i++) {
    const int diff = coeff[i] - dqcoeff[i];
1022
    error += diff * diff;
Jingning Han's avatar
Jingning Han committed
1023 1024 1025 1026 1027 1028 1029
    sqcoeff += coeff[i] * coeff[i];
  }

  *ssz = sqcoeff;
  return error;
}

Yaowu Xu's avatar
Yaowu Xu committed
1030 1031
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
1032 1033 1034 1035 1036
  int i;
  int64_t error = 0;

  for (i = 0; i < block_size; i++) {
    const int diff = coeff[i] - dqcoeff[i];
1037
    error += diff * diff;
Jingning Han's avatar
Jingning Han committed
1038 1039 1040 1041 1042
  }

  return error;
}

Yaowu Xu's avatar
Yaowu Xu committed
1043 1044 1045 1046
#if CONFIG_AOM_HIGHBITDEPTH
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
1047 1048 1049 1050 1051 1052 1053
  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];
1054
    error += diff * diff;
Jingning Han's avatar
Jingning Han committed
1055 1056 1057 1058 1059 1060 1061 1062 1063
    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;
}
Yaowu Xu's avatar
Yaowu Xu committed
1064
#endif  // CONFIG_AOM_HIGHBITDEPTH
Jingning Han's avatar
Jingning Han committed
1065

1066
#if !CONFIG_PVQ
1067
/* The trailing '0' is a terminator which is used inside av1_cost_coeffs() to
Jingning Han's avatar
Jingning Han committed
1068 1069 1070 1071
 * 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). */
1072 1073 1074
int av1_cost_coeffs(const AV1_COMMON *const cm, MACROBLOCK *x, int plane,
                    int block, int coeff_ctx, TX_SIZE tx_size,
                    const int16_t *scan, const int16_t *nb,
1075
                    int use_fast_coef_costing) {
Jingning Han's avatar
Jingning Han committed
1076 1077 1078 1079 1080
  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;
1081
  const uint16_t *band_count = &band_count_table[tx_size][1];
Jingning Han's avatar
Jingning Han committed
1082 1083
  const int eob = p->eobs[block];
  const tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
1084
  const int tx_size_ctx = txsize_sqr_map[tx_size];
1085 1086
  unsigned int(*token_costs)[2][COEFF_CONTEXTS][ENTROPY_TOKENS] =
      x->token_costs[tx_size_ctx][type][is_inter_block(mbmi)];
1087
  uint8_t token_cache[MAX_TX_SQUARE];
1088
  int pt = coeff_ctx;
Jingning Han's avatar
Jingning Han committed
1089
  int c, cost;
Yaowu Xu's avatar
Yaowu Xu committed
1090 1091
#if CONFIG_AOM_HIGHBITDEPTH
  const int *cat6_high_cost = av1_get_high_cost_table(xd->bd);
Jingning Han's avatar
Jingning Han committed
1092
#else
Yaowu Xu's avatar
Yaowu Xu committed
1093
  const int *cat6_high_cost = av1_get_high_cost_table(8);
Jingning Han's avatar
Jingning Han committed
1094 1095
#endif

1096
#if !CONFIG_VAR_TX && !CONFIG_SUPERTX
Jingning Han's avatar
Jingning Han committed
1097 1098 1099
  // Check for consistency of tx_size with mode info
  assert(type == PLANE_TYPE_Y ? mbmi->tx_size == tx_size
                              : get_uv_tx_size(mbmi, pd) == tx_size);
1100
#endif  // !CONFIG_VAR_TX && !CONFIG_SUPERTX
1101
  (void)cm;
Jingning Han's avatar
Jingning Han committed
1102 1103 1104 1105 1106

  if (eob == 0) {
    // single eob token
    cost = token_costs[0][0][pt][EOB_TOKEN];
  } else {
Julia Robson's avatar
Julia Robson committed
1107 1108 1109 1110 1111 1112
    if (use_fast_coef_costing) {
      int band_left = *band_count++;

      // dc token
      int v = qcoeff[0];
      int16_t prev_t;
Yaowu Xu's avatar
Yaowu Xu committed
1113
      cost = av1_get_token_cost(v, &prev_t, cat6_high_cost);
Julia Robson's avatar
Julia Robson committed
1114 1115
      cost += (*token_costs)[0][pt][prev_t];

Yaowu Xu's avatar
Yaowu Xu committed
1116
      token_cache[0] = av1_pt_energy_class[prev_t];
Julia Robson's avatar
Julia Robson committed
1117 1118 1119 1120 1121 1122 1123 1124
      ++token_costs;

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

        v = qcoeff[rc];
Yaowu Xu's avatar
Yaowu Xu committed
1125
        cost += av1_get_token_cost(v, &t, cat6_high_cost);
Julia Robson's avatar
Julia Robson committed
1126 1127 1128 1129 1130 1131
        cost += (*token_costs)[!prev_t][!prev_t][t];
        prev_t = t;
        if (!--band_left) {
          band_left = *band_count++;
          ++token_costs;
        }
Jingning Han's avatar
Jingning Han committed
1132 1133
      }

Julia Robson's avatar
Julia Robson committed
1134
      // eob token
1135
      if (band_left) cost += (*token_costs)[0][!prev_t][EOB_TOKEN];
Julia Robson's avatar
Julia Robson committed
1136 1137 1138 1139 1140 1141 1142

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

      // dc token
      int v = qcoeff[0];
      int16_t tok;
1143
      unsigned int(*tok_cost_ptr)[COEFF_CONTEXTS][ENTROPY_TOKENS];
Yaowu Xu's avatar
Yaowu Xu committed
1144
      cost = av1_get_token_cost(v, &tok, cat6_high_cost);
Julia Robson's avatar
Julia Robson committed
1145 1146
      cost += (*token_costs)[0][pt][tok];

Yaowu Xu's avatar
Yaowu Xu committed
1147
      token_cache[0] = av1_pt_energy_class[tok];
Julia Robson's avatar
Julia Robson committed
1148 1149 1150 1151 1152 1153 1154 1155 1156
      ++token_costs;

      tok_cost_ptr = &((*token_costs)[!tok]);

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

        v = qcoeff[rc];
Yaowu Xu's avatar
Yaowu Xu committed
1157
        cost += av1_get_token_cost(v, &tok, cat6_high_cost);
Julia Robson's avatar
Julia Robson committed
1158 1159
        pt = get_coef_context(nb, token_cache, c);
        cost += (*tok_cost_ptr)[pt][tok];
Yaowu Xu's avatar
Yaowu Xu committed
1160
        token_cache[rc] = av1_pt_energy_class[tok];
Julia Robson's avatar
Julia Robson committed
1161 1162 1163 1164 1165 1166 1167 1168 1169
        if (!--band_left) {
          band_left = *band_count++;
          ++token_costs;
        }
        tok_cost_ptr = &((*token_costs)[!tok]);
      }

      // eob token
      if (band_left) {
Jingning Han's avatar
Jingning Han committed
1170 1171 1172 1173 1174 1175 1176 1177
        pt = get_coef_context(nb, token_cache, c);
        cost += (*token_costs)[0][pt][EOB_TOKEN];
      }
    }
  }

  return cost;
}
1178
#endif
Jingning Han's avatar
Jingning Han committed
1179

Yaowu Xu's avatar
Yaowu Xu committed
1180 1181
static void dist_block(const AV1_COMP *cpi, MACROBLOCK *x, int plane, int block,
                       int blk_row, int blk_col, TX_SIZE tx_size,