rdopt.c 422 KB
Newer Older
1
/*
2
 * Copyright (c) 2016, Alliance for Open Media. All rights reserved
3
 *
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.
10 11 12 13 14
 */

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

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

18
#include "aom_dsp/aom_dsp_common.h"
19
#include "aom_dsp/blend.h"
20
#include "aom_mem/aom_mem.h"
Angie Chiang's avatar
Angie Chiang committed
21
#include "aom_ports/aom_timer.h"
22 23
#include "aom_ports/mem.h"
#include "aom_ports/system_state.h"
24

25 26 27
#if CONFIG_CFL
#include "av1/common/cfl.h"
#endif
28 29 30 31 32 33
#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"
34
#include "av1/common/obmc.h"
35 36 37 38 39 40
#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"
41 42 43
#if CONFIG_LV_MAP
#include "av1/common/txb_common.h"
#endif
Yue Chen's avatar
Yue Chen committed
44
#include "av1/common/warped_motion.h"
45

Jingning Han's avatar
Jingning Han committed
46
#include "av1/encoder/aq_variance.h"
47
#include "av1/encoder/av1_quantize.h"
48 49 50 51
#include "av1/encoder/cost.h"
#include "av1/encoder/encodemb.h"
#include "av1/encoder/encodemv.h"
#include "av1/encoder/encoder.h"
52 53 54
#if CONFIG_LV_MAP
#include "av1/encoder/encodetxb.h"
#endif
55 56 57 58 59 60
#include "av1/encoder/hybrid_fwd_txfm.h"
#include "av1/encoder/mcomp.h"
#include "av1/encoder/palette.h"
#include "av1/encoder/ratectrl.h"
#include "av1/encoder/rd.h"
#include "av1/encoder/rdopt.h"
61
#include "av1/encoder/tokenize.h"
62
#include "av1/encoder/tx_prune_model_weights.h"
Yushin Cho's avatar
Yushin Cho committed
63

64
#if CONFIG_DUAL_FILTER
Angie Chiang's avatar
Angie Chiang committed
65
#define DUAL_FILTER_SET_SIZE (SWITCHABLE_FILTERS * SWITCHABLE_FILTERS)
66 67 68 69
static const int filter_sets[DUAL_FILTER_SET_SIZE][2] = {
  { 0, 0 }, { 0, 1 }, { 0, 2 }, { 1, 0 }, { 1, 1 },
  { 1, 2 }, { 2, 0 }, { 2, 1 }, { 2, 2 },
};
Angie Chiang's avatar
Angie Chiang committed
70
#endif  // CONFIG_DUAL_FILTER
71

Zoe Liu's avatar
Zoe Liu committed
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 99 100
#define LAST_FRAME_MODE_MASK                                          \
  ((1 << INTRA_FRAME) | (1 << LAST2_FRAME) | (1 << LAST3_FRAME) |     \
   (1 << GOLDEN_FRAME) | (1 << BWDREF_FRAME) | (1 << ALTREF2_FRAME) | \
   (1 << ALTREF_FRAME))
#define LAST2_FRAME_MODE_MASK                                         \
  ((1 << INTRA_FRAME) | (1 << LAST_FRAME) | (1 << LAST3_FRAME) |      \
   (1 << GOLDEN_FRAME) | (1 << BWDREF_FRAME) | (1 << ALTREF2_FRAME) | \
   (1 << ALTREF_FRAME))
#define LAST3_FRAME_MODE_MASK                                         \
  ((1 << INTRA_FRAME) | (1 << LAST_FRAME) | (1 << LAST2_FRAME) |      \
   (1 << GOLDEN_FRAME) | (1 << BWDREF_FRAME) | (1 << ALTREF2_FRAME) | \
   (1 << ALTREF_FRAME))
#define GOLDEN_FRAME_MODE_MASK                                       \
  ((1 << INTRA_FRAME) | (1 << LAST_FRAME) | (1 << LAST2_FRAME) |     \
   (1 << LAST3_FRAME) | (1 << BWDREF_FRAME) | (1 << ALTREF2_FRAME) | \
   (1 << ALTREF_FRAME))
#define BWDREF_FRAME_MODE_MASK                                       \
  ((1 << INTRA_FRAME) | (1 << LAST_FRAME) | (1 << LAST2_FRAME) |     \
   (1 << LAST3_FRAME) | (1 << GOLDEN_FRAME) | (1 << ALTREF2_FRAME) | \
   (1 << ALTREF_FRAME))
#define ALTREF2_FRAME_MODE_MASK                                     \
  ((1 << INTRA_FRAME) | (1 << LAST_FRAME) | (1 << LAST2_FRAME) |    \
   (1 << LAST3_FRAME) | (1 << GOLDEN_FRAME) | (1 << BWDREF_FRAME) | \
   (1 << ALTREF_FRAME))
#define ALTREF_FRAME_MODE_MASK                                      \
  ((1 << INTRA_FRAME) | (1 << LAST_FRAME) | (1 << LAST2_FRAME) |    \
   (1 << LAST3_FRAME) | (1 << GOLDEN_FRAME) | (1 << BWDREF_FRAME) | \
   (1 << ALTREF2_FRAME))

101
#if CONFIG_EXT_COMP_REFS
102 103 104
#define SECOND_REF_FRAME_MASK                                         \
  ((1 << ALTREF_FRAME) | (1 << ALTREF2_FRAME) | (1 << BWDREF_FRAME) | \
   (1 << GOLDEN_FRAME) | (1 << LAST2_FRAME) | 0x01)
105
#else  // !CONFIG_EXT_COMP_REFS
Zoe Liu's avatar
Zoe Liu committed
106 107
#define SECOND_REF_FRAME_MASK \
  ((1 << ALTREF_FRAME) | (1 << ALTREF2_FRAME) | (1 << BWDREF_FRAME) | 0x01)
108
#endif  // CONFIG_EXT_COMP_REFS
109

110
#define NEW_MV_DISCOUNT_FACTOR 8
111

112 113
#define ANGLE_SKIP_THRESH 10

114 115 116 117 118 119
static const double ADST_FLIP_SVM[8] = {
  /* vertical */
  -6.6623, -2.8062, -3.2531, 3.1671,
  /* horizontal */
  -7.7051, -3.2234, -3.6193, 3.4533
};
120

121 122 123 124 125
typedef struct {
  PREDICTION_MODE mode;
  MV_REFERENCE_FRAME ref_frame[2];
} MODE_DEFINITION;

126
typedef struct { MV_REFERENCE_FRAME ref_frame[2]; } REF_DEFINITION;
127 128

struct rdcost_block_args {
129
  const AV1_COMP *cpi;
130
  MACROBLOCK *x;
131 132
  ENTROPY_CONTEXT t_above[2 * MAX_MIB_SIZE];
  ENTROPY_CONTEXT t_left[2 * MAX_MIB_SIZE];
133
  RD_STATS rd_stats;
134 135 136 137 138 139 140
  int64_t this_rd;
  int64_t best_rd;
  int exit_early;
  int use_fast_coef_costing;
};

#define LAST_NEW_MV_INDEX 6
141
static const MODE_DEFINITION av1_mode_order[MAX_MODES] = {
Emil Keyder's avatar
Emil Keyder committed
142 143 144 145
  { NEARESTMV, { LAST_FRAME, NONE_FRAME } },
  { NEARESTMV, { LAST2_FRAME, NONE_FRAME } },
  { NEARESTMV, { LAST3_FRAME, NONE_FRAME } },
  { NEARESTMV, { BWDREF_FRAME, NONE_FRAME } },
Zoe Liu's avatar
Zoe Liu committed
146
  { NEARESTMV, { ALTREF2_FRAME, NONE_FRAME } },
Emil Keyder's avatar
Emil Keyder committed
147 148
  { NEARESTMV, { ALTREF_FRAME, NONE_FRAME } },
  { NEARESTMV, { GOLDEN_FRAME, NONE_FRAME } },
149

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

Emil Keyder's avatar
Emil Keyder committed
152 153 154 155
  { NEWMV, { LAST_FRAME, NONE_FRAME } },
  { NEWMV, { LAST2_FRAME, NONE_FRAME } },
  { NEWMV, { LAST3_FRAME, NONE_FRAME } },
  { NEWMV, { BWDREF_FRAME, NONE_FRAME } },
Zoe Liu's avatar
Zoe Liu committed
156
  { NEWMV, { ALTREF2_FRAME, NONE_FRAME } },
Emil Keyder's avatar
Emil Keyder committed
157 158
  { NEWMV, { ALTREF_FRAME, NONE_FRAME } },
  { NEWMV, { GOLDEN_FRAME, NONE_FRAME } },
159

Emil Keyder's avatar
Emil Keyder committed
160 161 162 163
  { NEARMV, { LAST_FRAME, NONE_FRAME } },
  { NEARMV, { LAST2_FRAME, NONE_FRAME } },
  { NEARMV, { LAST3_FRAME, NONE_FRAME } },
  { NEARMV, { BWDREF_FRAME, NONE_FRAME } },
Zoe Liu's avatar
Zoe Liu committed
164
  { NEARMV, { ALTREF2_FRAME, NONE_FRAME } },
Emil Keyder's avatar
Emil Keyder committed
165 166
  { NEARMV, { ALTREF_FRAME, NONE_FRAME } },
  { NEARMV, { GOLDEN_FRAME, NONE_FRAME } },
167

Sarah Parker's avatar
Sarah Parker committed
168 169 170 171 172 173 174
  { GLOBALMV, { LAST_FRAME, NONE_FRAME } },
  { GLOBALMV, { LAST2_FRAME, NONE_FRAME } },
  { GLOBALMV, { LAST3_FRAME, NONE_FRAME } },
  { GLOBALMV, { BWDREF_FRAME, NONE_FRAME } },
  { GLOBALMV, { ALTREF2_FRAME, NONE_FRAME } },
  { GLOBALMV, { GOLDEN_FRAME, NONE_FRAME } },
  { GLOBALMV, { ALTREF_FRAME, NONE_FRAME } },
175

176
  // TODO(zoeliu): May need to reconsider the order on the modes to check
177

178 179 180 181 182 183 184 185
  { NEAREST_NEARESTMV, { LAST_FRAME, ALTREF_FRAME } },
  { NEAREST_NEARESTMV, { LAST2_FRAME, ALTREF_FRAME } },
  { NEAREST_NEARESTMV, { LAST3_FRAME, ALTREF_FRAME } },
  { NEAREST_NEARESTMV, { GOLDEN_FRAME, ALTREF_FRAME } },
  { NEAREST_NEARESTMV, { LAST_FRAME, BWDREF_FRAME } },
  { NEAREST_NEARESTMV, { LAST2_FRAME, BWDREF_FRAME } },
  { NEAREST_NEARESTMV, { LAST3_FRAME, BWDREF_FRAME } },
  { NEAREST_NEARESTMV, { GOLDEN_FRAME, BWDREF_FRAME } },
Zoe Liu's avatar
Zoe Liu committed
186 187 188 189
  { NEAREST_NEARESTMV, { LAST_FRAME, ALTREF2_FRAME } },
  { NEAREST_NEARESTMV, { LAST2_FRAME, ALTREF2_FRAME } },
  { NEAREST_NEARESTMV, { LAST3_FRAME, ALTREF2_FRAME } },
  { NEAREST_NEARESTMV, { GOLDEN_FRAME, ALTREF2_FRAME } },
190 191 192

#if CONFIG_EXT_COMP_REFS
  { NEAREST_NEARESTMV, { LAST_FRAME, LAST2_FRAME } },
193
  { NEAREST_NEARESTMV, { LAST_FRAME, LAST3_FRAME } },
194 195 196
  { NEAREST_NEARESTMV, { LAST_FRAME, GOLDEN_FRAME } },
  { NEAREST_NEARESTMV, { BWDREF_FRAME, ALTREF_FRAME } },
#endif  // CONFIG_EXT_COMP_REFS
197

198
  { PAETH_PRED, { INTRA_FRAME, NONE_FRAME } },
199

Emil Keyder's avatar
Emil Keyder committed
200
  { SMOOTH_PRED, { INTRA_FRAME, NONE_FRAME } },
201 202
  { SMOOTH_V_PRED, { INTRA_FRAME, NONE_FRAME } },
  { SMOOTH_H_PRED, { INTRA_FRAME, NONE_FRAME } },
203

204 205 206 207 208 209
  { 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 } },
Sarah Parker's avatar
Sarah Parker committed
210
  { GLOBAL_GLOBALMV, { LAST_FRAME, ALTREF_FRAME } },
211

212 213 214 215 216 217
  { 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 } },
Sarah Parker's avatar
Sarah Parker committed
218
  { GLOBAL_GLOBALMV, { LAST2_FRAME, ALTREF_FRAME } },
219 220 221 222 223 224 225

  { 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 } },
Sarah Parker's avatar
Sarah Parker committed
226
  { GLOBAL_GLOBALMV, { LAST3_FRAME, ALTREF_FRAME } },
227

228 229 230 231 232 233
  { 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 } },
Sarah Parker's avatar
Sarah Parker committed
234
  { GLOBAL_GLOBALMV, { GOLDEN_FRAME, ALTREF_FRAME } },
235

236 237 238 239 240 241
  { 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 } },
Sarah Parker's avatar
Sarah Parker committed
242
  { GLOBAL_GLOBALMV, { LAST_FRAME, BWDREF_FRAME } },
243 244 245 246 247 248 249

  { 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 } },
Sarah Parker's avatar
Sarah Parker committed
250
  { GLOBAL_GLOBALMV, { LAST2_FRAME, BWDREF_FRAME } },
251 252 253 254 255 256 257

  { 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 } },
Sarah Parker's avatar
Sarah Parker committed
258
  { GLOBAL_GLOBALMV, { LAST3_FRAME, BWDREF_FRAME } },
259 260 261 262 263 264 265

  { 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 } },
Sarah Parker's avatar
Sarah Parker committed
266
  { GLOBAL_GLOBALMV, { GOLDEN_FRAME, BWDREF_FRAME } },
267

Zoe Liu's avatar
Zoe Liu committed
268 269 270 271 272 273
  { NEAR_NEARMV, { LAST_FRAME, ALTREF2_FRAME } },
  { NEW_NEARESTMV, { LAST_FRAME, ALTREF2_FRAME } },
  { NEAREST_NEWMV, { LAST_FRAME, ALTREF2_FRAME } },
  { NEW_NEARMV, { LAST_FRAME, ALTREF2_FRAME } },
  { NEAR_NEWMV, { LAST_FRAME, ALTREF2_FRAME } },
  { NEW_NEWMV, { LAST_FRAME, ALTREF2_FRAME } },
Sarah Parker's avatar
Sarah Parker committed
274
  { GLOBAL_GLOBALMV, { LAST_FRAME, ALTREF2_FRAME } },
Zoe Liu's avatar
Zoe Liu committed
275 276 277 278 279 280 281

  { NEAR_NEARMV, { LAST2_FRAME, ALTREF2_FRAME } },
  { NEW_NEARESTMV, { LAST2_FRAME, ALTREF2_FRAME } },
  { NEAREST_NEWMV, { LAST2_FRAME, ALTREF2_FRAME } },
  { NEW_NEARMV, { LAST2_FRAME, ALTREF2_FRAME } },
  { NEAR_NEWMV, { LAST2_FRAME, ALTREF2_FRAME } },
  { NEW_NEWMV, { LAST2_FRAME, ALTREF2_FRAME } },
Sarah Parker's avatar
Sarah Parker committed
282
  { GLOBAL_GLOBALMV, { LAST2_FRAME, ALTREF2_FRAME } },
Zoe Liu's avatar
Zoe Liu committed
283 284 285 286 287 288 289

  { NEAR_NEARMV, { LAST3_FRAME, ALTREF2_FRAME } },
  { NEW_NEARESTMV, { LAST3_FRAME, ALTREF2_FRAME } },
  { NEAREST_NEWMV, { LAST3_FRAME, ALTREF2_FRAME } },
  { NEW_NEARMV, { LAST3_FRAME, ALTREF2_FRAME } },
  { NEAR_NEWMV, { LAST3_FRAME, ALTREF2_FRAME } },
  { NEW_NEWMV, { LAST3_FRAME, ALTREF2_FRAME } },
Sarah Parker's avatar
Sarah Parker committed
290
  { GLOBAL_GLOBALMV, { LAST3_FRAME, ALTREF2_FRAME } },
Zoe Liu's avatar
Zoe Liu committed
291 292 293 294 295 296 297

  { NEAR_NEARMV, { GOLDEN_FRAME, ALTREF2_FRAME } },
  { NEW_NEARESTMV, { GOLDEN_FRAME, ALTREF2_FRAME } },
  { NEAREST_NEWMV, { GOLDEN_FRAME, ALTREF2_FRAME } },
  { NEW_NEARMV, { GOLDEN_FRAME, ALTREF2_FRAME } },
  { NEAR_NEWMV, { GOLDEN_FRAME, ALTREF2_FRAME } },
  { NEW_NEWMV, { GOLDEN_FRAME, ALTREF2_FRAME } },
Sarah Parker's avatar
Sarah Parker committed
298
  { GLOBAL_GLOBALMV, { GOLDEN_FRAME, ALTREF2_FRAME } },
Zoe Liu's avatar
Zoe Liu committed
299

Emil Keyder's avatar
Emil Keyder committed
300 301 302 303 304 305 306 307
  { 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 } },
308

309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341
#if CONFIG_EXT_COMP_REFS
  { NEAR_NEARMV, { LAST_FRAME, LAST2_FRAME } },
  { NEW_NEARESTMV, { LAST_FRAME, LAST2_FRAME } },
  { NEAREST_NEWMV, { LAST_FRAME, LAST2_FRAME } },
  { NEW_NEARMV, { LAST_FRAME, LAST2_FRAME } },
  { NEAR_NEWMV, { LAST_FRAME, LAST2_FRAME } },
  { NEW_NEWMV, { LAST_FRAME, LAST2_FRAME } },
  { GLOBAL_GLOBALMV, { LAST_FRAME, LAST2_FRAME } },

  { NEAR_NEARMV, { LAST_FRAME, LAST3_FRAME } },
  { NEW_NEARESTMV, { LAST_FRAME, LAST3_FRAME } },
  { NEAREST_NEWMV, { LAST_FRAME, LAST3_FRAME } },
  { NEW_NEARMV, { LAST_FRAME, LAST3_FRAME } },
  { NEAR_NEWMV, { LAST_FRAME, LAST3_FRAME } },
  { NEW_NEWMV, { LAST_FRAME, LAST3_FRAME } },
  { GLOBAL_GLOBALMV, { LAST_FRAME, LAST3_FRAME } },

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

  { NEAR_NEARMV, { BWDREF_FRAME, ALTREF_FRAME } },
  { NEW_NEARESTMV, { BWDREF_FRAME, ALTREF_FRAME } },
  { NEAREST_NEWMV, { BWDREF_FRAME, ALTREF_FRAME } },
  { NEW_NEARMV, { BWDREF_FRAME, ALTREF_FRAME } },
  { NEAR_NEWMV, { BWDREF_FRAME, ALTREF_FRAME } },
  { NEW_NEWMV, { BWDREF_FRAME, ALTREF_FRAME } },
  { GLOBAL_GLOBALMV, { BWDREF_FRAME, ALTREF_FRAME } },
#endif  // CONFIG_EXT_COMP_REFS
342 343
};

344
static const PREDICTION_MODE intra_rd_search_mode_order[INTRA_MODES] = {
345
  DC_PRED,       H_PRED,        V_PRED,    SMOOTH_PRED, PAETH_PRED,
346 347
  SMOOTH_V_PRED, SMOOTH_H_PRED, D135_PRED, D207_PRED,   D153_PRED,
  D63_PRED,      D117_PRED,     D45_PRED,
348 349
};

Luc Trudeau's avatar
Luc Trudeau committed
350 351
#if CONFIG_CFL
static const UV_PREDICTION_MODE uv_rd_search_mode_order[UV_INTRA_MODES] = {
352 353 354 355
  UV_DC_PRED,     UV_CFL_PRED,   UV_H_PRED,        UV_V_PRED,
  UV_SMOOTH_PRED, UV_PAETH_PRED, UV_SMOOTH_V_PRED, UV_SMOOTH_H_PRED,
  UV_D135_PRED,   UV_D207_PRED,  UV_D153_PRED,     UV_D63_PRED,
  UV_D117_PRED,   UV_D45_PRED,
Luc Trudeau's avatar
Luc Trudeau committed
356 357 358 359 360
};
#else
#define uv_rd_search_mode_order intra_rd_search_mode_order
#endif  // CONFIG_CFL

hui su's avatar
hui su committed
361
static INLINE int write_uniform_cost(int n, int v) {
362 363
  const int l = get_unsigned_bits(n);
  const int m = (1 << l) - n;
364
  if (l == 0) return 0;
hui su's avatar
hui su committed
365
  if (v < m)
Hui Su's avatar
Hui Su committed
366
    return av1_cost_literal(l - 1);
hui su's avatar
hui su committed
367
  else
Hui Su's avatar
Hui Su committed
368
    return av1_cost_literal(l);
hui su's avatar
hui su committed
369 370
}

371 372 373
// 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
374 375 376
#define FAST_EXT_TX_CORR_MARGIN 0.5
#define FAST_EXT_TX_EDST_MARGIN 0.3

377 378 379 380 381
int inter_block_yrd(const AV1_COMP *cpi, MACROBLOCK *x, RD_STATS *rd_stats,
                    BLOCK_SIZE bsize, int64_t ref_best_rd, int fast);
int inter_block_uvrd(const AV1_COMP *cpi, MACROBLOCK *x, RD_STATS *rd_stats,
                     BLOCK_SIZE bsize, int64_t ref_best_rd, int fast);

382 383 384 385 386 387 388
static unsigned pixel_dist_visible_only(
    const AV1_COMP *const cpi, const MACROBLOCK *x, const uint8_t *src,
    const int src_stride, const uint8_t *dst, const int dst_stride,
    const BLOCK_SIZE tx_bsize, int txb_rows, int txb_cols, int visible_rows,
    int visible_cols) {
  unsigned sse;

389
  if (txb_rows == visible_rows && txb_cols == visible_cols) {
390 391 392 393 394 395 396 397 398 399 400 401 402 403 404
    cpi->fn_ptr[tx_bsize].vf(src, src_stride, dst, dst_stride, &sse);
    return sse;
  }
  const MACROBLOCKD *xd = &x->e_mbd;

  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
    uint64_t sse64 = aom_highbd_sse_odd_size(src, src_stride, dst, dst_stride,
                                             visible_cols, visible_rows);
    return (unsigned int)ROUND_POWER_OF_TWO(sse64, (xd->bd - 8) * 2);
  }
  sse = aom_sse_odd_size(src, src_stride, dst, dst_stride, visible_cols,
                         visible_rows);
  return sse;
}

405
#if CONFIG_DIST_8X8
Yushin Cho's avatar
Yushin Cho committed
406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449
static uint64_t cdef_dist_8x8_16bit(uint16_t *dst, int dstride, uint16_t *src,
                                    int sstride, int coeff_shift) {
  uint64_t svar = 0;
  uint64_t dvar = 0;
  uint64_t sum_s = 0;
  uint64_t sum_d = 0;
  uint64_t sum_s2 = 0;
  uint64_t sum_d2 = 0;
  uint64_t sum_sd = 0;
  uint64_t dist = 0;

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

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

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

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

  return dist;
}

Yushin Cho's avatar
Yushin Cho committed
450
static int od_compute_var_4x4(uint16_t *x, int stride) {
451 452 453 454 455 456 457 458 459 460 461 462 463 464 465
  int sum;
  int s2;
  int i;
  sum = 0;
  s2 = 0;
  for (i = 0; i < 4; i++) {
    int j;
    for (j = 0; j < 4; j++) {
      int t;

      t = x[i * stride + j];
      sum += t;
      s2 += t * t;
    }
  }
Yushin Cho's avatar
Yushin Cho committed
466

467 468 469
  return (s2 - (sum * sum >> 4)) >> 4;
}

470 471 472 473 474 475 476
/* 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
477 478
static double od_compute_dist_8x8(int use_activity_masking, uint16_t *x,
                                  uint16_t *y, od_coeff *e_lp, int stride) {
479 480 481 482 483 484 485 486 487 488 489
  double sum;
  int min_var;
  double mean_var;
  double var_stat;
  double activity;
  double calibration;
  int i;
  int j;
  double vardist;

  vardist = 0;
Yushin Cho's avatar
Yushin Cho committed
490

491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506
#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
507
     masking is used, since the harmonic mean appeared slightly worse with
508 509 510 511 512 513 514 515 516 517 518 519 520 521
     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;
522
#endif  // 1
523 524 525
  sum = 0;
  for (i = 0; i < 8; i++) {
    for (j = 0; j < 8; j++)
526
      sum += e_lp[i * stride + j] * (double)e_lp[i * stride + j];
527
  }
528 529 530
  /* 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);
531 532 533 534
  return activity * activity * (sum + vardist);
}

// Note : Inputs x and y are in a pixel domain
Yushin Cho's avatar
Yushin Cho committed
535 536
static double od_compute_dist_common(int activity_masking, uint16_t *x,
                                     uint16_t *y, int bsize_w, int bsize_h,
537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556
                                     int qindex, od_coeff *tmp,
                                     od_coeff *e_lp) {
  int i, j;
  double sum = 0;
  const int mid = OD_DIST_LP_MID;

  for (j = 0; j < bsize_w; j++) {
    e_lp[j] = mid * tmp[j] + 2 * tmp[bsize_w + j];
    e_lp[(bsize_h - 1) * bsize_w + j] = mid * tmp[(bsize_h - 1) * bsize_w + j] +
                                        2 * tmp[(bsize_h - 2) * bsize_w + j];
  }
  for (i = 1; i < bsize_h - 1; i++) {
    for (j = 0; j < bsize_w; j++) {
      e_lp[i * bsize_w + j] = mid * tmp[i * bsize_w + j] +
                              tmp[(i - 1) * bsize_w + j] +
                              tmp[(i + 1) * bsize_w + j];
    }
  }
  for (i = 0; i < bsize_h; i += 8) {
    for (j = 0; j < bsize_w; j += 8) {
Yushin Cho's avatar
Yushin Cho committed
557
      sum += od_compute_dist_8x8(activity_masking, &x[i * bsize_w + j],
558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575
                                 &y[i * bsize_w + j], &e_lp[i * bsize_w + j],
                                 bsize_w);
    }
  }
  /* Scale according to linear regression against SSE, for 8x8 blocks. */
  if (activity_masking) {
    sum *= 2.2 + (1.7 - 2.2) * (qindex - 99) / (210 - 99) +
           (qindex < 99 ? 2.5 * (qindex - 99) / 99 * (qindex - 99) / 99 : 0);
  } else {
    sum *= qindex >= 128
               ? 1.4 + (0.9 - 1.4) * (qindex - 128) / (209 - 128)
               : qindex <= 43 ? 1.5 + (2.0 - 1.5) * (qindex - 43) / (16 - 43)
                              : 1.5 + (1.4 - 1.5) * (qindex - 43) / (128 - 43);
  }

  return sum;
}

Yushin Cho's avatar
Yushin Cho committed
576 577
static double od_compute_dist(uint16_t *x, uint16_t *y, int bsize_w,
                              int bsize_h, int qindex) {
578
  assert(bsize_w >= 8 && bsize_h >= 8);
Yushin Cho's avatar
Yushin Cho committed
579

Yushin Cho's avatar
Yushin Cho committed
580
  int activity_masking = 0;
Yushin Cho's avatar
Yushin Cho committed
581

Luc Trudeau's avatar
Luc Trudeau committed
582 583 584 585 586 587 588
  int i, j;
  DECLARE_ALIGNED(16, od_coeff, e[MAX_TX_SQUARE]);
  DECLARE_ALIGNED(16, od_coeff, tmp[MAX_TX_SQUARE]);
  DECLARE_ALIGNED(16, od_coeff, e_lp[MAX_TX_SQUARE]);
  for (i = 0; i < bsize_h; i++) {
    for (j = 0; j < bsize_w; j++) {
      e[i * bsize_w + j] = x[i * bsize_w + j] - y[i * bsize_w + j];
589
    }
Luc Trudeau's avatar
Luc Trudeau committed
590 591 592 593 594 595 596 597 598
  }
  int mid = OD_DIST_LP_MID;
  for (i = 0; i < bsize_h; i++) {
    tmp[i * bsize_w] = mid * e[i * bsize_w] + 2 * e[i * bsize_w + 1];
    tmp[i * bsize_w + bsize_w - 1] =
        mid * e[i * bsize_w + bsize_w - 1] + 2 * e[i * bsize_w + bsize_w - 2];
    for (j = 1; j < bsize_w - 1; j++) {
      tmp[i * bsize_w + j] = mid * e[i * bsize_w + j] + e[i * bsize_w + j - 1] +
                             e[i * bsize_w + j + 1];
599
    }
600
  }
Luc Trudeau's avatar
Luc Trudeau committed
601 602
  return od_compute_dist_common(activity_masking, x, y, bsize_w, bsize_h,
                                qindex, tmp, e_lp);
603 604
}

Yushin Cho's avatar
Yushin Cho committed
605 606
static double od_compute_dist_diff(uint16_t *x, int16_t *e, int bsize_w,
                                   int bsize_h, int qindex) {
607
  assert(bsize_w >= 8 && bsize_h >= 8);
Yushin Cho's avatar
Yushin Cho committed
608

Yushin Cho's avatar
Yushin Cho committed
609
  int activity_masking = 0;
Yushin Cho's avatar
Yushin Cho committed
610

Luc Trudeau's avatar
Luc Trudeau committed
611 612 613 614 615 616 617
  DECLARE_ALIGNED(16, uint16_t, y[MAX_TX_SQUARE]);
  DECLARE_ALIGNED(16, od_coeff, tmp[MAX_TX_SQUARE]);
  DECLARE_ALIGNED(16, od_coeff, e_lp[MAX_TX_SQUARE]);
  int i, j;
  for (i = 0; i < bsize_h; i++) {
    for (j = 0; j < bsize_w; j++) {
      y[i * bsize_w + j] = x[i * bsize_w + j] - e[i * bsize_w + j];
618
    }
Luc Trudeau's avatar
Luc Trudeau committed
619 620 621 622 623 624 625 626 627
  }
  int mid = OD_DIST_LP_MID;
  for (i = 0; i < bsize_h; i++) {
    tmp[i * bsize_w] = mid * e[i * bsize_w] + 2 * e[i * bsize_w + 1];
    tmp[i * bsize_w + bsize_w - 1] =
        mid * e[i * bsize_w + bsize_w - 1] + 2 * e[i * bsize_w + bsize_w - 2];
    for (j = 1; j < bsize_w - 1; j++) {
      tmp[i * bsize_w + j] = mid * e[i * bsize_w + j] + e[i * bsize_w + j - 1] +
                             e[i * bsize_w + j + 1];
628 629
    }
  }
Luc Trudeau's avatar
Luc Trudeau committed
630 631
  return od_compute_dist_common(activity_masking, x, y, bsize_w, bsize_h,
                                qindex, tmp, e_lp);
632 633
}

634
int64_t av1_dist_8x8(const AV1_COMP *const cpi, const MACROBLOCK *x,
Yushin Cho's avatar
Yushin Cho committed
635 636 637 638
                     const uint8_t *src, int src_stride, const uint8_t *dst,
                     int dst_stride, const BLOCK_SIZE tx_bsize, int bsw,
                     int bsh, int visible_w, int visible_h, int qindex) {
  int64_t d = 0;
639
  int i, j;
640
  const MACROBLOCKD *xd = &x->e_mbd;
Yushin Cho's avatar
Yushin Cho committed
641 642 643 644

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

645 646 647 648 649
  assert(bsw >= 8);
  assert(bsh >= 8);
  assert((bsw & 0x07) == 0);
  assert((bsh & 0x07) == 0);

650 651 652
  if (x->tune_metric == AOM_TUNE_CDEF_DIST ||
      x->tune_metric == AOM_TUNE_DAALA_DIST) {
    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
653
      for (j = 0; j < bsh; j++)
654
        for (i = 0; i < bsw; i++)
655
          orig[j * bsw + i] = CONVERT_TO_SHORTPTR(src)[j * src_stride + i];
656

657
      if ((bsw == visible_w) && (bsh == visible_h)) {
658 659
        for (j = 0; j < bsh; j++)
          for (i = 0; i < bsw; i++)
660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676
            rec[j * bsw + i] = CONVERT_TO_SHORTPTR(dst)[j * dst_stride + i];
      } else {
        for (j = 0; j < visible_h; j++)
          for (i = 0; i < visible_w; i++)
            rec[j * bsw + i] = CONVERT_TO_SHORTPTR(dst)[j * dst_stride + i];

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

        if (visible_h < bsh) {
          for (j = visible_h; j < bsh; j++)
            for (i = 0; i < bsw; i++)
              rec[j * bsw + i] = CONVERT_TO_SHORTPTR(src)[j * src_stride + i];
        }
677
      }
678
    } else {
679
      for (j = 0; j < bsh; j++)
680
        for (i = 0; i < bsw; i++) orig[j * bsw + i] = src[j * src_stride + i];
681

682
      if ((bsw == visible_w) && (bsh == visible_h)) {
683
        for (j = 0; j < bsh; j++)
684 685 686 687 688 689 690 691 692 693 694
          for (i = 0; i < bsw; i++) rec[j * bsw + i] = dst[j * dst_stride + i];
      } else {
        for (j = 0; j < visible_h; j++)
          for (i = 0; i < visible_w; i++)
            rec[j * bsw + i] = dst[j * dst_stride + i];

        if (visible_w < bsw) {
          for (j = 0; j < bsh; j++)
            for (i = visible_w; i < bsw; i++)
              rec[j * bsw + i] = src[j * src_stride + i];
        }
695

696 697 698 699 700
        if (visible_h < bsh) {
          for (j = visible_h; j < bsh; j++)
            for (i = 0; i < bsw; i++)
              rec[j * bsw + i] = src[j * src_stride + i];
        }
701
      }
702 703
    }
  }
Yushin Cho's avatar
Yushin Cho committed
704

705 706 707
  if (x->tune_metric == AOM_TUNE_DAALA_DIST) {
    d = (int64_t)od_compute_dist(orig, rec, bsw, bsh, qindex);
  } else if (x->tune_metric == AOM_TUNE_CDEF_DIST) {
Yushin Cho's avatar
Yushin Cho committed
708 709 710 711 712 713 714 715 716 717
    int coeff_shift = AOMMAX(xd->bd - 8, 0);

    for (i = 0; i < bsh; i += 8) {
      for (j = 0; j < bsw; j += 8) {
        d += cdef_dist_8x8_16bit(&rec[i * bsw + j], bsw, &orig[i * bsw + j],
                                 bsw, coeff_shift);
      }
    }
    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH)
      d = ((uint64_t)d) >> 2 * coeff_shift;
718 719
  } else {
    // Otherwise, MSE by default
720 721
    d = pixel_dist_visible_only(cpi, x, src, src_stride, dst, dst_stride,
                                tx_bsize, bsh, bsw, visible_h, visible_w);
Yushin Cho's avatar
Yushin Cho committed
722
  }
723

724 725
  return d;
}
726

Cheng Chen's avatar
Cheng Chen committed
727 728 729 730
static int64_t dist_8x8_diff(const MACROBLOCK *x, const uint8_t *src,
                             int src_stride, const int16_t *diff,
                             int diff_stride, int bsw, int bsh, int visible_w,
                             int visible_h, int qindex) {
Yushin Cho's avatar
Yushin Cho committed
731
  int64_t d = 0;
732
  int i, j;
733
  const MACROBLOCKD *xd = &x->e_mbd;
Yushin Cho's avatar
Yushin Cho committed
734 735 736 737

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

738 739 740 741 742
  assert(bsw >= 8);
  assert(bsh >= 8);
  assert((bsw & 0x07) == 0);
  assert((bsh & 0x07) == 0);

743 744 745 746 747 748 749 750 751 752
  if (x->tune_metric == AOM_TUNE_CDEF_DIST ||
      x->tune_metric == AOM_TUNE_DAALA_DIST) {
    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
      for (j = 0; j < bsh; j++)
        for (i = 0; i < bsw; i++)
          orig[j * bsw + i] = CONVERT_TO_SHORTPTR(src)[j * src_stride + i];
    } else {
      for (j = 0; j < bsh; j++)
        for (i = 0; i < bsw; i++) orig[j * bsw + i] = src[j * src_stride + i];
    }
753

754
    if ((bsw == visible_w) && (bsh == visible_h)) {
755
      for (j = 0; j < bsh; j++)
756 757 758 759 760 761 762 763 764 765 766
        for (i = 0; i < bsw; i++)
          diff16[j * bsw + i] = diff[j * diff_stride + i];
    } else {
      for (j = 0; j < visible_h; j++)
        for (i = 0; i < visible_w; i++)
          diff16[j * bsw + i] = diff[j * diff_stride + i];

      if (visible_w < bsw) {
        for (j = 0; j < bsh; j++)
          for (i = visible_w; i < bsw; i++) diff16[j * bsw + i] = 0;
      }
767

768 769 770 771
      if (visible_h < bsh) {
        for (j = visible_h; j < bsh; j++)
          for (i = 0; i < bsw; i++) diff16[j * bsw + i] = 0;
      }
772 773
    }
  }
774

775 776 777
  if (x->tune_metric == AOM_TUNE_DAALA_DIST) {
    d = (int64_t)od_compute_dist_diff(orig, diff16, bsw, bsh, qindex);
  } else if (x->tune_metric == AOM_TUNE_CDEF_DIST) {
Yushin Cho's avatar
Yushin Cho committed
778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794
    int coeff_shift = AOMMAX(xd->bd - 8, 0);
    DECLARE_ALIGNED(16, uint16_t, dst16[MAX_TX_SQUARE]);

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

    for (i = 0; i < bsh; i += 8) {
      for (j = 0; j < bsw; j += 8) {
        d += cdef_dist_8x8_16bit(&dst16[i * bsw + j], bsw, &orig[i * bsw + j],
                                 bsw, coeff_shift);
      }
    }
    // Don't scale 'd' for HBD since it will be done by caller side for diff
    // input
795 796
  } else {
    // Otherwise, MSE by default
797
    d = aom_sum_squares_2d_i16(diff, diff_stride, visible_w, visible_h);
Yushin Cho's avatar
Yushin Cho committed
798
  }
799 800 801

  return d;
}
Yushin Cho's avatar
Yushin Cho committed
802
#endif  // CONFIG_DIST_8X8
803

804
static void get_energy_distribution_fine(const AV1_COMP *cpi, BLOCK_SIZE bsize,
805 806
                                         const uint8_t *src, int src_stride,
                                         const uint8_t *dst, int dst_stride,
807
                                         double *hordist, double *verdist) {
808 809
  const int bw = block_size_wide[bsize];
  const int bh = block_size_high[bsize];
810
  unsigned int esq[16] = { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 };
811

812
  const int f_index = bsize - BLOCK_16X16;
813
  if (f_index < 0) {
814 815
    const int w_shift = bw == 8 ? 1 : 2;
    const int h_shift = bh == 8 ? 1 : 2;
816
    if (cpi->common.use_highbitdepth) {
817 818 819 820 821
      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);
822 823 824
          esq[index] +=
              (src16[j + i * src_stride] - dst16[j + i * dst_stride]) *
              (src16[j + i * src_stride] - dst16[j + i * dst_stride]);
825 826
        }
    } else {
827 828 829
      for (int i = 0; i < bh; ++i)
        for (int j = 0; j < bw; ++j) {
          const int index = (j >> w_shift) + ((i >> h_shift) << 2);
830 831 832 833
          esq[index] += (src[j + i * src_stride] - dst[j + i * dst_stride]) *
                        (src[j + i * src_stride] - dst[j + i * dst_stride]);
        }
    }
834
  } else {
835 836 837 838 839 840 841
    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]);
842 843 844
    src += bh / 4 * src_stride;
    dst += bh / 4 * dst_stride;

845 846 847 848 849 850 851
    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]);
852 853 854
    src += bh / 4 * src_stride;
    dst += bh / 4 * dst_stride;

855 856 857 858 859 860 861
    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]);
862 863 864
    src += bh / 4 * src_stride;
    dst += bh / 4 * dst_stride;

865 866 867 868 869 870 871
    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]);
872 873
  }

874 875 876
  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];
877 878
  if (total > 0) {
    const double e_recip = 1.0 / total;
879 880 881 882 883 884
    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;
885 886 887 888 889 890 891
  } else {
    hordist[0] = verdist[0] = 0.25;
    hordist[1] = verdist[1] = 0.25;
    hordist[2] = verdist[2] = 0.25;
  }
}

892 893 894
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) {
895 896
  int prune_bitmask = 0;
  double svm_proj_h = 0, svm_proj_v = 0;
897
  double hdist[3] = { 0, 0, 0 }, vdist[3] = { 0, 0, 0 };
898 899
  get_energy_distribution_fine(cpi, bsize, src, src_stride, dst, dst_stride,
                               hdist, vdist);
900

901
  svm_proj_v = vdist[0] * ADST_FLIP_SVM[0] + vdist[1] * ADST_FLIP_SVM[1] +
902
               vdist[2] * ADST_FLIP_SVM[2] + ADST_FLIP_SVM[3];
903
  svm_proj_h = hdist[0] * ADST_FLIP_SVM[4] + hdist[1] * ADST_FLIP_SVM[5] +
904 905 906 907 908 909 910 911 912 913 914 915 916 917
               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;
}

918 919
static void get_horver_correlation(const int16_t *diff, int stride, int w,
                                   int h, double *hcorr, double *vcorr) {
920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946
  // 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;
    }
  }
947 948 949
  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;
950 951 952 953 954 955 956 957 958 959 960 961
  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;
  }
}

Cheng Chen's avatar
Cheng Chen committed
962
static int dct_vs_idtx(const int16_t *diff, int stride, int w, int h) {
963
  double hcorr, vcorr;
964
  int prune_bitmask = 0;
965
  get_horver_correlation(diff, stride, w, h, &hcorr, &vcorr);
966

967
  if (vcorr > FAST_EXT_TX_CORR_MID + FAST_EXT_TX_CORR_MARGIN)
968
    prune_bitmask |= 1 << IDTX_1D;
969
  else if (vcorr < FAST_EXT_TX_CORR_MID - FAST_EXT_TX_CORR_MARGIN)
970 971
    prune_bitmask |= 1 << DCT_1D;

972
  if (hcorr > FAST_EXT_TX_CORR_MID + FAST_EXT_TX_CORR_MARGIN)
973
    prune_bitmask |= 1 << (IDTX_1D + 8);
974
  else if (hcorr < FAST_EXT_TX_CORR_MID - FAST_EXT_TX_CORR_MARGIN)
975 976 977 978 979
    prune_bitmask |= 1 << (DCT_1D + 8);
  return prune_bitmask;
}

// Performance drop: 0.5%, Speed improvement: 24%
980
static int prune_two_for_sby(const AV1_COMP *cpi, BLOCK_SIZE bsize,
981 982
                             MACROBLOCK *x, const MACROBLOCKD *xd,
                             int adst_flipadst, int dct_idtx) {
983
  int prune = 0;
984

985 986 987
  if (adst_flipadst) {
    const struct macroblock_plane *const p = &x->plane[0];
    const struct macroblockd_plane *const pd = &xd->plane[0];
988
    prune |= adst_vs_flipadst(cpi, bsize, p->src.buf, p->src.stride,
989 990 991 992 993 994 995 996 997
                              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);
  }
998 999 1000

  return prune;
}
1001

1002
// Performance drop: 0.3%, Speed improvement: 5%
1003
static int prune_one_for_sby(const AV1_COMP *cpi, BLOCK_SIZE bsize,
1004 1005 1006
                             const MACROBLOCK *x, const MACROBLOCKD *xd) {
  const struct macroblock_plane *const p = &x->plane[0];
  const struct macroblockd_plane *const pd = &xd->plane[0];
1007
  return adst_vs_flipadst(cpi, bsize, p->src.buf, p->src.stride, pd->dst.buf,
1008
                          pd->dst.stride);
1009 1010
}

1011 1012 1013 1014 1015
// 1D Transforms used in inter set, this needs to be changed if
// ext_tx_used_inter is changed
static const int ext_tx_used_inter_1D[EXT_TX_SETS_INTER][TX_TYPES_1D] = {
  { 1, 0, 0, 0 }, { 1, 1, 1, 1 }, { 1, 1, 1, 1 }, { 1, 0, 0, 1 },
};
1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068

static void get_energy_distribution_finer(const int16_t *diff, int stride,
                                          int bw, int bh, float *hordist,
                                          float *verdist) {
  // First compute downscaled block energy values (esq); downscale factors
  // are defined by w_shift and h_shift.
  unsigned int esq[256];
  const int w_shift = bw <= 8 ? 0 : 1;
  const int h_shift = bh <= 8 ? 0 : 1;
  const int esq_w = bw <= 8 ? bw : bw / 2;
  const int esq_h = bh <= 8 ? bh : bh / 2;
  const int esq_sz = esq_w * esq_h;
  int i, j;
  memset(esq, 0, esq_sz * sizeof(esq[0]));
  for (i = 0; i < bh; i++) {
    unsigned int *cur_esq_row = esq + (i >> h_shift) * esq_w;
    const int16_t *cur_diff_row = diff + i * stride;
    for (j = 0; j < bw; j++) {
      cur_esq_row[j >> w_shift] += cur_diff_row[j] * cur_diff_row[j];
    }
  }

  uint64_t total = 0;
  for (i = 0; i < esq_sz; i++) total += esq[i];

  // Output hordist and verdist arrays are normalized 1D projections of esq
  if (total == 0) {
    float hor_val = 1.0f / esq_w;
    for (j = 0; j < esq_w - 1; j++) hordist[j] = hor_val;
    float ver_val = 1.0f / esq_h;
    for (i = 0; i < esq_h - 1; i++) verdist[i] = ver_val;
    return;
  }

  const float e_recip = 1.0f / (float)total;
  memset(hordist, 0, (esq_w - 1) * sizeof(hordist[0]));
  memset(verdist, 0, (esq_h - 1) * sizeof(verdist[0]));
  const unsigned int *cur_esq_row;
  for (i = 0; i < esq_h - 1; i++) {
    cur_esq_row = esq + i * esq_w;
    for (j = 0; j < esq_w - 1; j++) {
      hordist[j] += (float)cur_esq_row[j];
      verdist[i] += (float)cur_esq_row[j];
    }
    verdist[i] += (float)cur_esq_row[j];
  }
  cur_esq_row = esq + i * esq_w;
  for (j = 0; j < esq_w - 1; j++) hordist[j] += (float)cur_esq_row[j];

  for (j = 0; j < esq_w - 1; j++) hordist[j] *= e_recip;
  for (i = 0; i < esq_h - 1; i++) verdist[i] *= e_recip;
}

1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096
// Instead of 1D projections of the block energy distribution computed by
// get_energy_distribution_finer() this function computes a full
// two-dimensional energy distribution of the input block.
static void get_2D_energy_distribution(const int16_t *diff, int stride, int bw,
                                       int bh, float *edist) {
  unsigned int esq[256] = { 0 };
  const int esq_w = bw >> 2;
  const int esq_h = bh >> 2;
  const int esq_sz = esq_w * esq_h;
  uint64_t total = 0;
  for (int i = 0; i < bh; i += 4) {
    for (int j = 0; j < bw; j += 4) {
      unsigned int cur_sum_energy = 0;
      for (int k = 0; k < 4; k++) {
        const int16_t *cur_diff = diff + (i + k) * stride + j;
        cur_sum_energy += cur_diff[0] * cur_diff[0] +
                          cur_diff[1] * cur_diff[1] +
                          cur_diff[2] * cur_diff[2] + cur_diff[3] * cur_diff[3];
      }
      esq[(i >> 2) * esq_w + (j >> 2)] = cur_sum_energy;
      total += cur_sum_energy;
    }
  }

  const float e_recip = 1.0f / (float)total;
  for (int i = 0; i < esq_sz - 1; i++) edist[i] = esq[i] * e_recip;
}

1097 1098 1099 1100
// Similar to get_horver_correlation, but also takes into account first
// row/column, when computing horizontal/vertical correlation.
static void get_horver_correlation_full(const int16_t *diff, int stride, int w,
                                        int h, float *hcorr, float *vcorr) {
1101 1102
  const float num_hor = (float)(h * (w - 1));
  const float num_ver = (float)((h - 1) * w);
1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207
  int i, j;

  // The following notation is used:
  // x - current pixel
  // y - left neighbor pixel
  // z - top neighbor pixel
  int64_t xy_sum = 0, xz_sum = 0;
  int64_t xhor_sum = 0, xver_sum = 0, y_sum = 0, z_sum = 0;
  int64_t x2hor_sum = 0, x2ver_sum = 0, y2_sum = 0, z2_sum = 0;

  int16_t x, y, z;
  for (j = 1; j < w; ++j) {
    x = diff[j];
    y = diff[j - 1];
    xy_sum += x * y;
    xhor_sum += x;
    y_sum += y;
    x2hor_sum += x * x;
    y2_sum += y * y;
  }
  for (i = 1; i < h; ++i) {
    x = diff[i * stride];
    z = diff[(i - 1) * stride];
    xz_sum += x * z;
    xver_sum += x;
    z_sum += z;
    x2ver_sum += x * x;
    z2_sum += z * z;
    for (j = 1; j < w; ++j) {
      x = diff[i * stride + j];
      y = diff[i * stride + j - 1];
      z = diff[(i - 1) * stride + j];
      xy_sum += x * y;
      xz_sum += x * z;
      xhor_sum += x;
      xver_sum += x;
      y_sum += y;
      z_sum += z;
      x2hor_sum += x * x;
      x2ver_sum += x * x;
      y2_sum += y * y;
      z2_sum += z * z;
    }
  }
  const float xhor_var_n = x2hor_sum - (xhor_sum * xhor_sum) / num_hor;
  const float y_var_n = y2_sum - (y_sum * y_sum) / num_hor;
  const float xy_var_n = xy_sum - (xhor_sum * y_sum) / num_hor;
  const float xver_var_n = x2ver_sum - (xver_sum * xver_sum) / num_ver;
  const float z_var_n = z2_sum - (z_sum * z_sum) / num_ver;
  const float xz_var_n = xz_sum - (xver_sum * z_sum) / num_ver;

  *hcorr = *vcorr = 1;
  if (xhor_var_n > 0 && y_var_n > 0) {
    *hcorr = xy_var_n / sqrtf(xhor_var_n * y_var_n);
    *hcorr = *hcorr < 0 ? 0 : *hcorr;
  }
  if (xver_var_n > 0 && z_var_n > 0) {
    *vcorr = xz_var_n / sqrtf(xver_var_n * z_var_n);
    *vcorr = *vcorr < 0 ? 0 : *vcorr;
  }
}

// Performs a forward pass through a neural network with 2 fully-connected
// layers, assuming ReLU as activation function. Number of output neurons
// is always equal to 4.
// fc1, fc2 - weight matrices of the respective layers.
// b1, b2 - bias vectors of the respective layers.
static void compute_1D_scores(float *features, int num_features,
                              const float *fc1, const float *b1,
                              const float *fc2, const float *b2,
                              int num_hidden_units, float *dst_scores) {
  assert(num_hidden_units <= 32);
  float hidden_layer[32];
  for (int i = 0; i < num_hidden_units; i++) {
    const float *cur_coef = fc1 + i * num_features;
    hidden_layer[i] = 0.0f;
    for (int j = 0; j < num_features; j++)
      hidden_layer[i] += cur_coef[j] * features[j];
    hidden_layer[i] = AOMMAX(hidden_layer[i] + b1[i], 0.0f);
  }
  for (int i = 0; i < 4; i++) {
    const float *cur_coef = fc2 + i * num_hidden_units;
    dst_scores[i] = 0.0f;
    for (int j = 0; j < num_hidden_units; j++)
      dst_scores[i] += cur_coef[j] * hidden_layer[j];
    dst_scores[i] += b2[i];
  }
}

// Transforms raw scores into a probability distribution across 16 TX types
static void score_2D_transform_pow8(float *scores_2D, float shift) {
  float sum = 0.0f;
  int i;

  for (i = 0; i < 16; i++) {
    float v, v2, v4;
    v = AOMMAX(scores_2D[i] + shift, 0.0f);
    v2 = v * v;
    v4 = v2 * v2;
    scores_2D[i] = v4 * v4;
    sum += scores_2D[i];
  }
  for (i = 0; i < 16; i++) scores_2D[i] /= sum;
}

1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257
// Similarly to compute_1D_scores() performs a forward pass through a
// neural network with two fully-connected layers. The only difference
// is that it assumes 1 output neuron, as required by the classifier used
// for TX size pruning.
static float compute_tx_split_prune_score(float *features, int num_features,
                                          const float *fc1, const float *b1,
                                          const float *fc2, float b2,
                                          int num_hidden_units) {
  assert(num_hidden_units <= 64);
  float hidden_layer[64];
  for (int i = 0; i < num_hidden_units; i++) {
    const float *cur_coef = fc1 + i * num_features;
    hidden_layer[i] = 0.0f;
    for (int j = 0; j < num_features; j++)
      hidden_layer[i] += cur_coef[j] * features[j];
    hidden_layer[i] = AOMMAX(hidden_layer[i] + b1[i], 0.0f);
  }
  float dst_score = 0.0f;
  for (int j = 0; j < num_hidden_units; j++)
    dst_score += fc2[j] * hidden_layer[j];
  dst_score += b2;
  return dst_score;
}

static int prune_tx_split(BLOCK_SIZE bsize, const int16_t *diff, float hcorr,
                          float vcorr) {
  if (bsize <= BLOCK_4X4 || bsize > BLOCK_16X16) return 0;

  float features[17];
  const int bw = block_size_wide[bsize], bh = block_size_high[bsize];
  const int feature_num = (bw / 4) * (bh / 4) + 1;
  assert(feature_num <= 17);

  get_2D_energy_distribution(diff, bw, bw, bh, features);
  features[feature_num - 2] = hcorr;
  features[feature_num - 1] = vcorr;

  const int bidx = bsize - BLOCK_4X4 - 1;
  const float *fc1 = av1_prune_tx_split_learned_weights[bidx];
  const float *b1 =
      fc1 + av1_prune_tx_split_num_hidden_units[bidx] * feature_num;
  const float *fc2 = b1 + av1_prune_tx_split_num_hidden_units[bidx];
  float b2 = *(fc2 + av1_prune_tx_split_num_hidden_units[bidx]);
  float score =
      compute_tx_split_prune_score(features, feature_num, fc1, b1, fc2, b2,
                                   av1_prune_tx_split_num_hidden_units[bidx]);

  return (score > av1_prune_tx_split_thresholds[bidx]);
}

1258 1259 1260
static void prune_tx_2D(BLOCK_SIZE bsize, MACROBLOCK *x,
                        TX_TYPE_PRUNE_MODE prune_mode, int use_tx_split_prune) {
  if (bsize >= BLOCK_32X32) return;
1261
  aom_clear_system_state();
1262 1263 1264 1265
  const struct macroblock_plane *const p = &x->plane[0];
  float hfeatures[16], vfeatures[16];
  float hscores[4], vscores[4];
  float scores_2D[16];
1266
  const int tx_type_table_2D[16] = {
1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281
    DCT_DCT,      DCT_ADST,      DCT_FLIPADST,      V_DCT,
    ADST_DCT,     ADST_ADST,     ADST_FLIPADST,     V_ADST,
    FLIPADST_DCT, FLIPADST_ADST, FLIPADST_FLIPADST, V_FLIPADST,
    H_DCT,        H_ADST,        H_FLIPADST,        IDTX
  };
  const int bw = block_size_wide[bsize], bh = block_size_high[bsize];
  const int hfeatures_num = bw <= 8 ? bw : bw / 2;
  const int vfeatures_num = bh <= 8 ? bh : bh / 2;
  assert(hfeatures_num <= 16);
  assert(vfeatures_num <= 16);

  get_energy_distribution_finer(p->src_diff, bw, bw, bh, hfeatures, vfeatures);
  get_horver_correlation_full(p->src_diff, bw, bw, bh,
                              &hfeatures[hfeatures_num - 1],
                              &vfeatures[vfeatures_num - 1]);
1282
  const int bidx = AOMMAX(bsize - BLOCK_4X4, 0);
1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311
  const float *fc1_hor = av1_prune_2D_learned_weights_hor[bidx];
  const float *b1_hor =
      fc1_hor + av1_prune_2D_num_hidden_units_hor[bidx] * hfeatures_num;
  const float *fc2_hor = b1_hor + av1_prune_2D_num_hidden_units_hor[bidx];
  const float *b2_hor = fc2_hor + av1_prune_2D_num_hidden_units_hor[bidx] * 4;
  compute_1D_scores(hfeatures, hfeatures_num, fc1_hor, b1_hor, fc2_hor, b2_hor,
                    av1_prune_2D_num_hidden_units_hor[bidx], hscores);

  const float *fc1_ver = av1_prune_2D_learned_weights_ver[bidx];
  const float *b1_ver =
      fc1_ver + av1_prune_2D_num_hidden_units_ver[bidx] * vfeatures_num;
  const float *fc2_ver = b1_ver + av1_prune_2D_num_hidden_units_ver[bidx];
  const float *b2_ver = fc2_ver + av1_prune_2D_num_hidden_units_ver[bidx] * 4;
  compute_1D_scores(vfeatures, vfeatures_num, fc1_ver, b1_ver, fc2_ver, b2_ver,
                    av1_prune_2D_num_hidden_units_ver[bidx], vscores);

  float score_2D_average = 0.0f;
  for (int i = 0; i < 4; i++) {
    float *cur_scores_2D = scores_2D + i * 4;
    cur_scores_2D[0] = vscores[i] * hscores[0];
    cur_scores_2D[1] = vscores[i] * hscores[1];
    cur_scores_2D[2] = vscores[i] * hscores[2];
    cur_scores_2D[3] = vscores[i] * hscores[3];
    score_2D_average += cur_scores_2D[0] + cur_scores_2D[1] + cur_scores_2D[2] +
                        cur_scores_2D[3];
  }
  score_2D_average /= 16;
  score_2D_transform_pow8(scores_2D, (20 - score_2D_average));

1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325
  // TODO(huisu@google.com): support more tx set types.
  const int tx_set_types[2] = { EXT_TX_SET_ALL16, EXT_TX_SET_DTT9_IDTX_1DDCT };
  for (int tx_set_idx = 0; tx_set_idx < 2; ++tx_set_idx) {
    const int tx_set_type = tx_set_types[tx_set_idx];
    // Always keep the TX type with the highest score, prune all others with
    // score below score_thresh.
    int max_score_i = 0;
    float max_score = 0.0f;
    for (int i = 0; i < 16; i++) {
      if (scores_2D[i] > max_score &&
          av1_ext_tx_used[tx_set_type][tx_type_table_2D[i]]) {
        max_score = scores_2D[i];
        max_score_i = i;
      }
1326 1327
    }

1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348
    int pruning_aggressiveness = 0;
    if (prune_mode == PRUNE_2D_ACCURATE) {
      if (tx_set_type == EXT_TX_SET_ALL16)
        pruning_aggressiveness = 6;
      else if (tx_set_type == EXT_TX_SET_DTT9_IDTX_1DDCT)
        pruning_aggressiveness = 4;
    } else if (prune_mode == PRUNE_2D_FAST) {
      if (tx_set_type == EXT_TX_SET_ALL16)
        pruning_aggressiveness = 10;
      else if (tx_set_type == EXT_TX_SET_DTT9_IDTX_1DDCT)
        pruning_aggressiveness = 7;
    }
    const float score_thresh =
        av1_prune_2D_adaptive_thresholds[bidx][pruning_aggressiveness - 1];

    int prune_bitmask = 0;
    for (int i = 0; i < 16; i++) {
      if (scores_2D[i] < score_thresh && i != max_score_i)
        prune_bitmask |= (1 << tx_type_table_2D[i]);
    }
    x->tx_search_prune[tx_set_type] = prune_bitmask;
1349 1350
  }

1351 1352 1353 1354 1355 1356 1357 1358 1359 1360
  // Also apply TX size pruning if it's turned on. The value
  // of prune_tx_split_flag indicates whether we should do
  // full TX size search (flag=0) or use the largest available
  // TX size without performing any further search (flag=1).
  int prune_tx_split_flag = 0;
  if (use_tx_split_prune) {
    prune_tx_split_flag =
        prune_tx_split(bsize, p->src_diff, hfeatures[hfeatures_num - 1],
                       vfeatures[vfeatures_num - 1]);
  }
1361
  x->tx_search_prune[0] |= (prune_tx_split_flag << TX_TYPES);
1362
}
1363

1364 1365 1366
static void prune_tx(const AV1_COMP *cpi, BLOCK_SIZE bsize, MACROBLOCK *x,
                     const MACROBLOCKD *const xd, int tx_set_type,
                     int use_tx_split_prune) {
1367 1368 1369 1370 1371 1372
  av1_zero(x->tx_search_prune);
  const MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
  if (!is_inter_block(mbmi) || cpi->sf.tx_type_search.prune_mode == NO_PRUNE ||
      x->use_default_inter_tx_type || xd->lossless[mbmi->segment_id] ||
      x->cb_partition_scan)
    return;
1373 1374 1375
  int tx_set = ext_tx_set_index[1][tx_set_type];
  assert(tx_set >= 0);
  const int *tx_set_1D = ext_tx_used_inter_1D[tx_set];
1376
  switch (cpi->sf.tx_type_search.prune_mode) {
1377
    case NO_PRUNE: return;
1378
    case PRUNE_ONE:
1379 1380
      if (!(tx_set_1D[FLIPADST_1D] & tx_set_1D[ADST_1D])) return;
      x->tx_search_prune[tx_set_type] = prune_one_for_sby(cpi, bsize, x, xd);
1381
      break;
1382
    case PRUNE_TWO:
1383
      if (!(tx_set_1D[FLIPADST_1D] & tx_set_1D[ADST_1D])) {
1384 1385 1386 1387 1388 1389 1390
        if (!(tx_set_1D[DCT_1D] & tx_set_1D[IDTX_1D])) return;
        x->tx_search_prune[tx_set_type] =
            prune_two_for_sby(cpi, bsize, x, xd, 0, 1);
      }
      if (!(tx_set_1D[DCT_1D] & tx_set_1D[IDTX_1D])) {
        x->tx_search_prune[tx_set_type] =
            prune_two_for_sby(cpi, bsize, x, xd, 1, 0);
1391
      }
1392 1393
      x->tx_search_prune[tx_set_type] =
          prune_two_for_sby(cpi, bsize, x, xd, 1, 1);
1394
      break;
1395 1396
    case PRUNE_2D_ACCURATE:
    case PRUNE_2D_FAST:
1397 1398
      prune_tx_2D(bsize, x, cpi->sf.tx_type_search.prune_mode,
                  use_tx_split_prune);
1399
      break;
1400
    default: assert(0);
1401 1402 1403
  }
}

1404 1405
static int do_tx_type_search(TX_TYPE tx_type, int prune,
                             TX_TYPE_PRUNE_MODE mode) {
1406
  // TODO(sarahparker) implement for non ext tx
1407 1408 1409 1410 1411 1412
  if (mode >= PRUNE_2D_ACCURATE) {
    return !((prune >> tx_type) & 1);
  } else {
    return !(((prune >> vtx_tab[tx_type]) & 1) |
             ((prune >> (htx_tab[tx_type] + 8)) & 1));
  }
1413 1414
}

1415
static void model_rd_from_sse(const AV1_COMP *const cpi,
1416 1417
                              const MACROBLOCKD *const xd, BLOCK_SIZE bsize,
                              int plane, int64_t sse, int *rate,
Geza Lore's avatar
Geza Lore committed
1418 1419 1420
                              int64_t *dist) {
  const struct macroblockd_plane *const pd = &xd->plane[plane];
  const int dequant_shift =
1421
      (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) ? xd->bd - 5 : 3;
Geza Lore's avatar
Geza Lore committed
1422 1423 1424 1425

  // Fast approximate the modelling function.
  if (cpi->sf.simple_model_rd_from_var) {
    const int64_t square_error = sse;
1426
    int quantizer = (pd->dequant_Q3[1] >> dequant_shift);
Geza Lore's avatar
Geza Lore committed
1427 1428

    if (quantizer < 120)
Yaowu Xu's avatar
Yaowu Xu committed
1429
      *rate = (int)((square_error * (280 - quantizer)) >>
1430
                    (16 - AV1_PROB_COST_SHIFT));
Geza Lore's avatar
Geza Lore committed
1431 1432 1433 1434
    else
      *rate = 0;
    *dist = (square_error * quantizer) >> 8;
  } else {
1435
    av1_model_rd_from_var_lapndz(sse, num_pels_log2_lookup[bsize],
1436 1437
                                 pd->dequant_Q3[1] >> dequant_shift, rate,
                                 dist);
Geza Lore's avatar
Geza Lore committed
1438 1439 1440 1441 1442
  }

  *dist <<= 4;
}

1443
static void model_rd_for_sb(const AV1_COMP *const cpi, BLOCK_SIZE bsize,
1444 1445 1446 1447
                            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) {
1448 1449 1450
  // 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.
1451 1452 1453
  int plane;
  const int ref = xd->mi[0]->mbmi.ref_frame[0];

1454 1455 1456 1457 1458 1459
  int64_t rate_sum = 0;
  int64_t dist_sum = 0;
  int64_t total_sse = 0;

  x->pred_sse[ref] = 0;

1460 1461 1462
  for (plane = plane_from; plane <= plane_to; ++plane) {
    struct macroblock_plane *const p = &x->plane[plane];
    struct macroblockd_plane *const pd = &xd->plane[plane];
1463
    const BLOCK_SIZE bs = get_plane_block_size(bsize, pd);
1464 1465 1466
    unsigned int sse;
    int rate;
    int64_t dist;
1467