vp9_onyx_int.h 28.8 KB
Newer Older
John Koleszar's avatar
John Koleszar committed
1
/*
2
 *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
John Koleszar's avatar
John Koleszar committed
3
 *
4
 *  Use of this source code is governed by a BSD-style license
5
6
 *  that can be found in the LICENSE file in the root of the source
 *  tree. An additional intellectual property rights grant can be found
7
 *  in the file PATENTS.  All contributing project authors may
8
 *  be found in the AUTHORS file in the root of the source tree.
John Koleszar's avatar
John Koleszar committed
9
10
 */

11
12
#ifndef VP9_ENCODER_VP9_ONYX_INT_H_
#define VP9_ENCODER_VP9_ONYX_INT_H_
John Koleszar's avatar
John Koleszar committed
13
14

#include <stdio.h>
15

16
#include "./vpx_config.h"
17
18
#include "vpx_ports/mem.h"
#include "vpx/internal/vpx_codec_internal.h"
19
#include "vpx/vp8cx.h"
20

21
#include "vp9/common/vp9_ppflags.h"
22
23
#include "vp9/common/vp9_entropy.h"
#include "vp9/common/vp9_entropymode.h"
24
#include "vp9/common/vp9_onyxc_int.h"
25

Marco Paniconi's avatar
Marco Paniconi committed
26
#include "vp9/encoder/vp9_aq_cyclicrefresh.h"
27
#include "vp9/encoder/vp9_encodemb.h"
28
#include "vp9/encoder/vp9_firstpass.h"
29
#include "vp9/encoder/vp9_lookahead.h"
30
#include "vp9/encoder/vp9_mbgraph.h"
31
32
33
#include "vp9/encoder/vp9_mcomp.h"
#include "vp9/encoder/vp9_quantize.h"
#include "vp9/encoder/vp9_ratectrl.h"
34
#include "vp9/encoder/vp9_svc_layercontext.h"
35
36
#include "vp9/encoder/vp9_tokenize.h"
#include "vp9/encoder/vp9_variance.h"
John Koleszar's avatar
John Koleszar committed
37

38
39
40
41
#ifdef __cplusplus
extern "C" {
#endif

Paul Wilkins's avatar
Paul Wilkins committed
42
// #define MODE_TEST_HIT_STATS
43

44
#define DEFAULT_GF_INTERVAL         10
John Koleszar's avatar
John Koleszar committed
45

46
47
#define MAX_MODES 30
#define MAX_REFS  6
48

John Koleszar's avatar
John Koleszar committed
49
typedef struct {
50
51
52
  int nmvjointcost[MV_JOINTS];
  int nmvcosts[2][MV_VALS];
  int nmvcosts_hp[2][MV_VALS];
John Koleszar's avatar
John Koleszar committed
53

54
  vp9_prob segment_pred_probs[PREDICTION_PROBS];
John Koleszar's avatar
John Koleszar committed
55

John Koleszar's avatar
John Koleszar committed
56
  unsigned char *last_frame_seg_map_copy;
57

John Koleszar's avatar
John Koleszar committed
58
59
  // 0 = Intra, Last, GF, ARF
  signed char last_ref_lf_deltas[MAX_REF_LF_DELTAS];
60
  // 0 = ZERO_MV, MV
John Koleszar's avatar
John Koleszar committed
61
  signed char last_mode_lf_deltas[MAX_MODE_LF_DELTAS];
John Koleszar's avatar
John Koleszar committed
62

63
  FRAME_CONTEXT fc;
John Koleszar's avatar
John Koleszar committed
64
65
} CODING_CONTEXT;

66
67
// This enumerator type needs to be kept aligned with the mode order in
// const MODE_DEFINITION vp9_mode_order[MAX_MODES] used in the rd code.
John Koleszar's avatar
John Koleszar committed
68
69
typedef enum {
  THR_NEARESTMV,
Paul Wilkins's avatar
Paul Wilkins committed
70
71
  THR_NEARESTA,
  THR_NEARESTG,
John Koleszar's avatar
John Koleszar committed
72

73
74
75
  THR_DC,

  THR_NEWMV,
Paul Wilkins's avatar
Paul Wilkins committed
76
  THR_NEWA,
77
78
79
  THR_NEWG,

  THR_NEARMV,
John Koleszar's avatar
John Koleszar committed
80
  THR_NEARA,
81
82
  THR_COMP_NEARESTLA,
  THR_COMP_NEARESTGA,
John Koleszar's avatar
John Koleszar committed
83
84

  THR_TM,
85

Paul Wilkins's avatar
Paul Wilkins committed
86
87
88
89
90
  THR_COMP_NEARLA,
  THR_COMP_NEWLA,
  THR_NEARG,
  THR_COMP_NEARGA,
  THR_COMP_NEWGA,
91

Paul Wilkins's avatar
Paul Wilkins committed
92
93
94
  THR_ZEROMV,
  THR_ZEROG,
  THR_ZEROA,
John Koleszar's avatar
John Koleszar committed
95
96
  THR_COMP_ZEROLA,
  THR_COMP_ZEROGA,
97

Paul Wilkins's avatar
Paul Wilkins committed
98
99
100
  THR_H_PRED,
  THR_V_PRED,
  THR_D135_PRED,
Dmitry Kovalev's avatar
Dmitry Kovalev committed
101
  THR_D207_PRED,
Paul Wilkins's avatar
Paul Wilkins committed
102
103
104
105
  THR_D153_PRED,
  THR_D63_PRED,
  THR_D117_PRED,
  THR_D45_PRED,
Ronald S. Bultje's avatar
Ronald S. Bultje committed
106
} THR_MODES;
John Koleszar's avatar
John Koleszar committed
107

108
109
110
111
112
113
114
115
116
typedef enum {
  THR_LAST,
  THR_GOLD,
  THR_ALTR,
  THR_COMP_LA,
  THR_COMP_GA,
  THR_INTRA,
} THR_MODES_SUB8X8;

John Koleszar's avatar
John Koleszar committed
117
118
119
typedef enum {
  DIAMOND = 0,
  NSTEP = 1,
120
121
  HEX = 2,
  BIGDIA = 3,
122
  SQUARE = 4,
123
124
  FAST_HEX = 5,
  FAST_DIAMOND = 6
John Koleszar's avatar
John Koleszar committed
125
126
} SEARCH_METHODS;

Deb Mukherjee's avatar
Deb Mukherjee committed
127
128
129
130
131
132
133
typedef enum {
  USE_FULL_RD = 0,
  USE_LARGESTINTRA,
  USE_LARGESTINTRA_MODELINTER,
  USE_LARGESTALL
} TX_SIZE_SEARCH_METHOD;

134
135
136
137
138
139
typedef enum {
  NOT_IN_USE = 0,
  RELAXED_NEIGHBORING_MIN_MAX = 1,
  STRICT_NEIGHBORING_MIN_MAX = 2
} AUTO_MIN_MAX_MODE;

140
typedef enum {
141
  // Terminate search early based on distortion so far compared to
142
  // qp step, distortion in the neighborhood of the frame, etc.
143
  FLAG_EARLY_TERMINATE = 1 << 0,
144

145
146
  // Skips comp inter modes if the best so far is an intra mode.
  FLAG_SKIP_COMP_BESTINTRA = 1 << 1,
147

148
  // Skips comp inter modes if the best single intermode so far does
149
  // not have the same reference as one of the two references being
150
151
  // tested.
  FLAG_SKIP_COMP_REFMISMATCH = 1 << 2,
152

153
154
  // Skips oblique intra modes if the best so far is an inter mode.
  FLAG_SKIP_INTRA_BESTINTER = 1 << 3,
155

156
157
158
  // Skips oblique intra modes  at angles 27, 63, 117, 153 if the best
  // intra so far is not one of the neighboring directions.
  FLAG_SKIP_INTRA_DIRMISMATCH = 1 << 4,
159

160
161
  // Skips intra modes other than DC_PRED if the source variance is small
  FLAG_SKIP_INTRA_LOWVAR = 1 << 5,
162
163
} MODE_SEARCH_SKIP_LOGIC;

164
typedef enum {
165
  SUBPEL_TREE = 0,
166
167
168
  // Other methods to come
} SUBPEL_SEARCH_METHODS;

Paul Wilkins's avatar
Paul Wilkins committed
169
170
171
172
173
174
typedef enum {
  LAST_FRAME_PARTITION_OFF = 0,
  LAST_FRAME_PARTITION_LOW_MOTION = 1,
  LAST_FRAME_PARTITION_ALL = 2
} LAST_FRAME_PARTITION_METHOD;

175
176
177
178
179
180
181
182
183
184
185
typedef enum {
  // No recode.
  DISALLOW_RECODE = 0,
  // Allow recode for KF and exceeding maximum frame bandwidth.
  ALLOW_RECODE_KFMAXBW = 1,
  // Allow recode only for KF/ARF/GF frames.
  ALLOW_RECODE_KFARFGF = 2,
  // Allow recode for all frames based on bitrate constraints.
  ALLOW_RECODE = 3,
} RECODE_LOOP_TYPE;

186
187
188
189
190
191
192
193
194
typedef enum {
  // encode_breakout is disabled.
  ENCODE_BREAKOUT_DISABLED = 0,
  // encode_breakout is enabled.
  ENCODE_BREAKOUT_ENABLED = 1,
  // encode_breakout is enabled with small max_thresh limit.
  ENCODE_BREAKOUT_LIMITED = 2
} ENCODE_BREAKOUT_TYPE;

195
196
197
198
199
200
201
202
203
204
205
typedef enum {
  // Search partitions using RD/NONRD criterion
  SEARCH_PARTITION = 0,

  // Always use a fixed size partition
  FIXED_PARTITION = 1,

  // Use a fixed size partition in every 64X64 SB, where the size is
  // determined based on source variance
  VAR_BASED_FIXED_PARTITION = 2,

206
207
  REFERENCE_PARTITION = 3,

208
209
210
211
212
  // Use an arbitrary partitioning scheme based on source variance within
  // a 64X64 SB
  VAR_BASED_PARTITION
} PARTITION_SEARCH_TYPE;

213
214
215
216
217
218
219
220
221
typedef enum {
  // Try the full image with different values.
  LPF_PICK_FROM_FULL_IMAGE,
  // Try a small portion of the image with different values.
  LPF_PICK_FROM_SUBIMAGE,
  // Estimate the level based on quantizer and frame type
  LPF_PICK_FROM_Q,
} LPF_PICK_METHOD;

John Koleszar's avatar
John Koleszar committed
222
typedef struct {
223
224
  // Frame level coding parameter update
  int frame_parameter_update;
Jim Bankoski's avatar
Jim Bankoski committed
225
226

  // Motion search method (Diamond, NSTEP, Hex, Big Diamond, Square, etc).
John Koleszar's avatar
John Koleszar committed
227
  SEARCH_METHODS search_method;
Jim Bankoski's avatar
Jim Bankoski committed
228

229
  RECODE_LOOP_TYPE recode_loop;
Jim Bankoski's avatar
Jim Bankoski committed
230
231
232
233
234

  // Subpel_search_method can only be subpel_tree which does a subpixel
  // logarithmic search that keeps stepping at 1/2 pixel units until
  // you stop getting a gain, and then goes on to 1/4 and repeats
  // the same process. Along the way it skips many diagonals.
235
  SUBPEL_SEARCH_METHODS subpel_search_method;
Jim Bankoski's avatar
Jim Bankoski committed
236
237

  // Maximum number of steps in logarithmic subpel search before giving up.
238
  int subpel_iters_per_step;
Jim Bankoski's avatar
Jim Bankoski committed
239

Yaowu Xu's avatar
Yaowu Xu committed
240
241
242
  // Control when to stop subpel search
  int subpel_force_stop;

Jim Bankoski's avatar
Jim Bankoski committed
243
244
  // This parameter controls the number of steps we'll do in a diamond
  // search.
John Koleszar's avatar
John Koleszar committed
245
  int max_step_search_steps;
Jim Bankoski's avatar
Jim Bankoski committed
246
247
248

  // This parameter controls which step in the n-step process we start at.
  // It's changed adaptively based on circumstances.
249
  int reduce_first_step_size;
Jim Bankoski's avatar
Jim Bankoski committed
250
251
252

  // If this is set to 1, we limit the motion search range to 2 times the
  // largest motion vector found in the last frame.
Paul Wilkins's avatar
Paul Wilkins committed
253
  int auto_mv_step_size;
Jim Bankoski's avatar
Jim Bankoski committed
254
255

  // Trellis (dynamic programming) optimization of quantized values (+1, 0).
John Koleszar's avatar
John Koleszar committed
256
  int optimize_coefficients;
Jim Bankoski's avatar
Jim Bankoski committed
257
258
259
260
261
262

  // Always set to 0. If on it enables 0 cost background transmission
  // (except for the initial transmission of the segmentation). The feature is
  // disabled because the addition of very large block sizes make the
  // backgrounds very to cheap to encode, and the segmentation we have
  // adds overhead.
263
  int static_segmentation;
Jim Bankoski's avatar
Jim Bankoski committed
264
265
266
267
268

  // If 1 we iterate finding a best reference for 2 ref frames together - via
  // a log search that iterates 4 times (check around mv for last for best
  // error of combined predictor then check around mv for alt). If 0 we
  // we just use the best motion vector found for each frame by itself.
269
  int comp_inter_joint_search_thresh;
Jim Bankoski's avatar
Jim Bankoski committed
270
271
272

  // This variable is used to cap the maximum number of times we skip testing a
  // mode to be evaluated. A high value means we will be faster.
273
  int adaptive_rd_thresh;
Jim Bankoski's avatar
Jim Bankoski committed
274
275
276
277

  // Enables skipping the reconstruction step (idct, recon) in the
  // intermediate steps assuming the last frame didn't have too many intra
  // blocks and the q is less than a threshold.
278
  int skip_encode_sb;
279
  int skip_encode_frame;
280
281
282
  // Speed feature to allow or disallow skipping of recode at block
  // level within a frame.
  int allow_skip_recode;
Jim Bankoski's avatar
Jim Bankoski committed
283
284
285
286
287
288
289
290

  // This variable allows us to reuse the last frames partition choices
  // (64x64 v 32x32 etc) for this frame. It can be set to only use the last
  // frame as a starting point in low motion scenes or always use it. If set
  // we use last partitioning_redo frequency to determine how often to redo
  // the partitioning from scratch. Adjust_partitioning_from_last_frame
  // enables us to adjust up or down one partitioning from the last frames
  // partitioning.
Paul Wilkins's avatar
Paul Wilkins committed
291
  LAST_FRAME_PARTITION_METHOD use_lastframe_partitioning;
Jim Bankoski's avatar
Jim Bankoski committed
292
293
294
295

  // Determine which method we use to determine transform size. We can choose
  // between options like full rd, largest for prediction size, largest
  // for intra and model coefs for the rest.
Deb Mukherjee's avatar
Deb Mukherjee committed
296
  TX_SIZE_SEARCH_METHOD tx_size_search_method;
Jim Bankoski's avatar
Jim Bankoski committed
297
298
299

  // Low precision 32x32 fdct keeps everything in 16 bits and thus is less
  // precise but significantly faster than the non lp version.
300
  int use_lp32x32fdct;
Jim Bankoski's avatar
Jim Bankoski committed
301
302
303
304
305
306

  // TODO(JBB): remove this as its no longer used.

  // After looking at the first set of modes (set by index here), skip
  // checking modes for reference frames that don't match the reference frame
  // of the best so far.
307
  int mode_skip_start;
Jim Bankoski's avatar
Jim Bankoski committed
308
309

  // TODO(JBB): Remove this.
Paul Wilkins's avatar
Paul Wilkins committed
310
  int reference_masking;
Jim Bankoski's avatar
Jim Bankoski committed
311

312
313
314
  PARTITION_SEARCH_TYPE partition_search_type;

  // Used if partition_search_type = FIXED_SIZE_PARTITION
315
  BLOCK_SIZE always_this_block_size;
Jim Bankoski's avatar
Jim Bankoski committed
316

317
318
319
320
321
322
323
  // Skip rectangular partition test when partition type none gives better
  // rd than partition type split.
  int less_rectangular_check;

  // Disable testing non square partitions. (eg 16x32)
  int use_square_partition_only;

Jim Bankoski's avatar
Jim Bankoski committed
324
  // Sets min and max partition sizes for this 64x64 region based on the
325
326
  // same 64x64 in last encoded frame, and the left and above neighbor.
  AUTO_MIN_MAX_MODE auto_min_max_partition_size;
Jim Bankoski's avatar
Jim Bankoski committed
327
328
329

  // Min and max partition size we enable (block_size) as per auto
  // min max, but also used by adjust partitioning, and pick_partitioning.
330
331
  BLOCK_SIZE min_partition_size;
  BLOCK_SIZE max_partition_size;
Jim Bankoski's avatar
Jim Bankoski committed
332
333
334

  // Whether or not we allow partitions one smaller or one greater than the last
  // frame's partitioning. Only used if use_lastframe_partitioning is set.
335
  int adjust_partitioning_from_last_frame;
Jim Bankoski's avatar
Jim Bankoski committed
336
337
338

  // How frequently we re do the partitioning from scratch. Only used if
  // use_lastframe_partitioning is set.
339
  int last_partitioning_redo_frequency;
Jim Bankoski's avatar
Jim Bankoski committed
340
341
342
343

  // Disables sub 8x8 blocksizes in different scenarios: Choices are to disable
  // it always, to allow it for only Last frame and Intra, disable it for all
  // inter modes or to enable it always.
344
  int disable_split_mask;
Jim Bankoski's avatar
Jim Bankoski committed
345

Jingning Han's avatar
Jingning Han committed
346
  // TODO(jingning): combine the related motion search speed features
Jim Bankoski's avatar
Jim Bankoski committed
347
348
  // This allows us to use motion search at other sizes as a starting
  // point for this motion search and limits the search range around it.
Jingning Han's avatar
Jingning Han committed
349
  int adaptive_motion_search;
Jim Bankoski's avatar
Jim Bankoski committed
350
351
352
353
354

  // Allows sub 8x8 modes to use the prediction filter that was determined
  // best for 8x8 mode. If set to 0 we always re check all the filters for
  // sizes less than 8x8, 1 means we check all filter modes if no 8x8 filter
  // was selected, and 2 means we use 8 tap if no 8x8 filter mode was selected.
355
  int adaptive_pred_interp_filter;
356

357
358
359
360
  // Search through variable block partition types in non-RD mode decision
  // encoding process for RTC.
  int partition_check;

361
362
363
364
  // Implements various heuristics to skip searching modes
  // The heuristics selected are based on  flags
  // defined in the MODE_SEARCH_SKIP_HEURISTICS enum
  unsigned int mode_search_skip_flags;
Jim Bankoski's avatar
Jim Bankoski committed
365

366
367
  // A source variance threshold below which the split mode is disabled
  unsigned int disable_split_var_thresh;
Jim Bankoski's avatar
Jim Bankoski committed
368

369
370
371
  // A source variance threshold below which filter search is disabled
  // Choose a very large value (UINT_MAX) to use 8-tap always
  unsigned int disable_filter_search_var_thresh;
Jim Bankoski's avatar
Jim Bankoski committed
372
373
374

  // These bit masks allow you to enable or disable intra modes for each
  // transform size separately.
375
376
  int intra_y_mode_mask[TX_SIZES];
  int intra_uv_mode_mask[TX_SIZES];
Jim Bankoski's avatar
Jim Bankoski committed
377
378
379
380

  // This variable enables an early break out of mode testing if the model for
  // rd built from the prediction signal indicates a value that's much
  // higher than the best rd we've seen so far.
381
  int use_rd_breakout;
Jim Bankoski's avatar
Jim Bankoski committed
382
383
384
385

  // This enables us to use an estimate for intra rd based on dc mode rather
  // than choosing an actual uv mode in the stage of encoding before the actual
  // final encode.
386
  int use_uv_intra_rd_estimate;
Jim Bankoski's avatar
Jim Bankoski committed
387

388
389
  // This feature controls how the loop filter level is determined.
  LPF_PICK_METHOD lpf_pick;
Jim Bankoski's avatar
Jim Bankoski committed
390
391
392

  // This feature limits the number of coefficients updates we actually do
  // by only looking at counts from 1/2 the bands.
393
  int use_fast_coef_updates;  // 0: 2-loop, 1: 1-loop, 2: 1-loop reduced
394

395
  // This flag controls the use of non-RD mode decision.
396
  int use_nonrd_pick_mode;
397
398
399
400

  // This variable sets the encode_breakout threshold. Currently, it is only
  // enabled in real time mode.
  int encode_breakout_thresh;
401
402
403
404

  // A binary mask indicating if NEARESTMV, NEARMV, ZEROMV, NEWMV
  // modes are disabled in order from LSB to MSB for each BLOCK_SIZE.
  int disable_inter_mode_mask[BLOCK_SIZES];
405
406
407
408

  // This feature controls whether we do the expensive context update and
  // calculation in the rd coefficient costing loop.
  int use_fast_coef_costing;
409
410
411
412
413

  // This variable controls the maximum block size where intra blocks can be
  // used in inter frames.
  // TODO(aconverse): Fold this into one of the other many mode skips
  BLOCK_SIZE max_intra_bsize;
John Koleszar's avatar
John Koleszar committed
414
415
} SPEED_FEATURES;

416
417
418
419
420
421
422
423
typedef enum {
  NORMAL      = 0,
  FOURFIVE    = 1,
  THREEFIVE   = 2,
  ONETWO      = 3
} VPX_SCALING;

typedef enum {
424
425
426
427
  USAGE_LOCAL_FILE_PLAYBACK = 0,
  USAGE_STREAM_FROM_SERVER  = 1,
  USAGE_CONSTRAINED_QUALITY = 2,
  USAGE_CONSTANT_QUALITY    = 3,
428
429
430
} END_USAGE;

typedef enum {
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
  // Good Quality Fast Encoding. The encoder balances quality with the
  // amount of time it takes to encode the output. (speed setting
  // controls how fast)
  MODE_GOODQUALITY = 1,

  // One Pass - Best Quality. The encoder places priority on the
  // quality of the output over encoding speed. The output is compressed
  // at the highest possible quality. This option takes the longest
  // amount of time to encode. (speed setting ignored)
  MODE_BESTQUALITY = 2,

  // Two Pass - First Pass. The encoder generates a file of statistics
  // for use in the second encoding pass. (speed setting controls how fast)
  MODE_FIRSTPASS = 3,

  // Two Pass - Second Pass. The encoder uses the statistics that were
  // generated in the first encoding pass to create the compressed
  // output. (speed setting controls how fast)
  MODE_SECONDPASS = 4,

  // Two Pass - Second Pass Best.  The encoder uses the statistics that
  // were generated in the first encoding pass to create the compressed
  // output using the highest possible quality, and taking a
  // longer amount of time to encode. (speed setting ignored)
  MODE_SECONDPASS_BEST = 5,

  // Realtime/Live Encoding. This mode is optimized for realtime
  // encoding (for example, capturing a television signal or feed from
  // a live camera). (speed setting controls how fast)
  MODE_REALTIME = 6,
461
462
463
} MODE;

typedef enum {
464
465
466
  FRAMEFLAGS_KEY    = 1 << 0,
  FRAMEFLAGS_GOLDEN = 1 << 1,
  FRAMEFLAGS_ALTREF = 1 << 2,
467
468
469
470
471
472
} FRAMETYPE_FLAGS;

typedef enum {
  NO_AQ = 0,
  VARIANCE_AQ = 1,
  COMPLEXITY_AQ = 2,
473
  CYCLIC_REFRESH_AQ = 3,
474
475
  AQ_MODE_COUNT  // This should always be the last member of the enum
} AQ_MODE;
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490

typedef struct {
  int version;  // 4 versions of bitstream defined:
                //   0 - best quality/slowest decode,
                //   3 - lowest quality/fastest decode
  int width;  // width of data passed to the compressor
  int height;  // height of data passed to the compressor
  double framerate;  // set to passed in framerate
  int64_t target_bandwidth;  // bandwidth to be used in kilobits per second

  int noise_sensitivity;  // pre processing blur: recommendation 0
  int sharpness;  // sharpening output: recommendation 0:
  int cpu_used;
  unsigned int rc_max_intra_bitrate_pct;

491
  MODE mode;
492
493
494
495
496
497
498
499
500
501

  // Key Framing Operations
  int auto_key;  // autodetect cut scenes and set the keyframes
  int key_freq;  // maximum distance to key frame.

  int lag_in_frames;  // how many frames lag before we start encoding

  // ----------------------------------------------------------------
  // DATARATE CONTROL OPTIONS

502
  END_USAGE end_usage;  // vbr or cbr
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521

  // buffer targeting aggressiveness
  int under_shoot_pct;
  int over_shoot_pct;

  // buffering parameters
  int64_t starting_buffer_level;  // in seconds
  int64_t optimal_buffer_level;
  int64_t maximum_buffer_size;

  // Frame drop threshold.
  int drop_frames_water_mark;

  // controlling quality
  int fixed_q;
  int worst_allowed_q;
  int best_allowed_q;
  int cq_level;
  int lossless;
522
  AQ_MODE aq_mode;  // Adaptive Quantization mode
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

  // two pass datarate control
  int two_pass_vbrbias;        // two pass datarate control tweaks
  int two_pass_vbrmin_section;
  int two_pass_vbrmax_section;
  // END DATARATE CONTROL OPTIONS
  // ----------------------------------------------------------------

  // Spatial and temporal scalability.
  int ss_number_layers;  // Number of spatial layers.
  int ts_number_layers;  // Number of temporal layers.
  // Bitrate allocation for spatial layers.
  int ss_target_bitrate[VPX_SS_MAX_LAYERS];
  // Bitrate allocation (CBR mode) and framerate factor, for temporal layers.
  int ts_target_bitrate[VPX_TS_MAX_LAYERS];
  int ts_rate_decimator[VPX_TS_MAX_LAYERS];

  // these parameters aren't to be used in final build don't use!!!
  int play_alternate;
  int alt_freq;

  int encode_breakout;  // early breakout : for video conf recommend 800

  /* Bitfield defining the error resiliency features to enable.
   * Can provide decodable frames after losses in previous
   * frames and decodable partitions after losses in the same frame.
   */
  unsigned int error_resilient_mode;

  /* Bitfield defining the parallel decoding mode where the
   * decoding in successive frames may be conducted in parallel
   * just by decoding the frame headers.
   */
  unsigned int frame_parallel_decoding_mode;

  int arnr_max_frames;
  int arnr_strength;
  int arnr_type;

  int tile_columns;
  int tile_rows;

  struct vpx_fixed_buf         two_pass_stats_in;
  struct vpx_codec_pkt_list  *output_pkt_list;

  vp8e_tuning tuning;
} VP9_CONFIG;

571
typedef struct VP9_COMP {
572
573
574
575
  DECLARE_ALIGNED(16, int16_t, y_quant[QINDEX_RANGE][8]);
  DECLARE_ALIGNED(16, int16_t, y_quant_shift[QINDEX_RANGE][8]);
  DECLARE_ALIGNED(16, int16_t, y_zbin[QINDEX_RANGE][8]);
  DECLARE_ALIGNED(16, int16_t, y_round[QINDEX_RANGE][8]);
John Koleszar's avatar
John Koleszar committed
576

577
578
579
580
  DECLARE_ALIGNED(16, int16_t, uv_quant[QINDEX_RANGE][8]);
  DECLARE_ALIGNED(16, int16_t, uv_quant_shift[QINDEX_RANGE][8]);
  DECLARE_ALIGNED(16, int16_t, uv_zbin[QINDEX_RANGE][8]);
  DECLARE_ALIGNED(16, int16_t, uv_round[QINDEX_RANGE][8]);
John Koleszar's avatar
John Koleszar committed
581

582
#if CONFIG_ALPHA
583
584
585
586
  DECLARE_ALIGNED(16, int16_t, a_quant[QINDEX_RANGE][8]);
  DECLARE_ALIGNED(16, int16_t, a_quant_shift[QINDEX_RANGE][8]);
  DECLARE_ALIGNED(16, int16_t, a_zbin[QINDEX_RANGE][8]);
  DECLARE_ALIGNED(16, int16_t, a_round[QINDEX_RANGE][8]);
587
#endif
John Koleszar's avatar
John Koleszar committed
588

John Koleszar's avatar
John Koleszar committed
589
  MACROBLOCK mb;
590
  VP9_COMMON common;
591
  VP9_CONFIG oxcf;
John Koleszar's avatar
John Koleszar committed
592
593
  struct lookahead_ctx    *lookahead;
  struct lookahead_entry  *source;
594
#if CONFIG_MULTIPLE_ARF
Dmitry Kovalev's avatar
Dmitry Kovalev committed
595
  struct lookahead_entry  *alt_ref_source[REF_FRAMES];
596
#else
John Koleszar's avatar
John Koleszar committed
597
  struct lookahead_entry  *alt_ref_source;
598
#endif
599
  struct lookahead_entry  *last_source;
John Koleszar's avatar
John Koleszar committed
600

John Koleszar's avatar
John Koleszar committed
601
  YV12_BUFFER_CONFIG *Source;
602
  YV12_BUFFER_CONFIG *Last_Source;  // NULL for first frame and alt_ref frames
John Koleszar's avatar
John Koleszar committed
603
604
  YV12_BUFFER_CONFIG *un_scaled_source;
  YV12_BUFFER_CONFIG scaled_source;
605
606
  YV12_BUFFER_CONFIG *unscaled_last_source;
  YV12_BUFFER_CONFIG scaled_last_source;
John Koleszar's avatar
John Koleszar committed
607

608
  int key_frame_frequency;
John Koleszar's avatar
John Koleszar committed
609

610
611
  int gold_is_last;  // gold same as last frame ( short circuit gold searches)
  int alt_is_last;  // Alt same as last ( short circuit altref search)
John Koleszar's avatar
John Koleszar committed
612
  int gold_is_alt;  // don't do both alt and gold search ( just do gold).
John Koleszar's avatar
John Koleszar committed
613

614
  int scaled_ref_idx[3];
615
616
617
  int lst_fb_idx;
  int gld_fb_idx;
  int alt_fb_idx;
618

619
#if CONFIG_MULTIPLE_ARF
Dmitry Kovalev's avatar
Dmitry Kovalev committed
620
  int alt_ref_fb_idx[REF_FRAMES - 3];
621
#endif
622
623
624
  int refresh_last_frame;
  int refresh_golden_frame;
  int refresh_alt_ref_frame;
Deb Mukherjee's avatar
Deb Mukherjee committed
625
626
627
628
629
630
631
632
633

  int ext_refresh_frame_flags_pending;
  int ext_refresh_last_frame;
  int ext_refresh_golden_frame;
  int ext_refresh_alt_ref_frame;

  int ext_refresh_frame_context_pending;
  int ext_refresh_frame_context;

John Koleszar's avatar
John Koleszar committed
634
  YV12_BUFFER_CONFIG last_frame_uf;
John Koleszar's avatar
John Koleszar committed
635

John Koleszar's avatar
John Koleszar committed
636
  TOKENEXTRA *tok;
Ronald S. Bultje's avatar
Ronald S. Bultje committed
637
  unsigned int tok_count[4][1 << 6];
John Koleszar's avatar
John Koleszar committed
638

639
640
641
642
643
644
#if CONFIG_MULTIPLE_ARF
  // Position within a frame coding order (including any additional ARF frames).
  unsigned int sequence_number;
  // Next frame in naturally occurring order that has not yet been coded.
  int next_frame_in_order;
#endif
John Koleszar's avatar
John Koleszar committed
645

John Koleszar's avatar
John Koleszar committed
646
647
  // Ambient reconstruction err target for force key frames
  int ambient_err;
648

649
650
651
652
653
654
655
  // Thresh_mult is used to set a threshold for the rd score. A higher value
  // means that we will accept the best mode so far more often. This number
  // is used in combination with the current block size, and thresh_freq_fact
  // to pick a threshold.
  int rd_thresh_mult[MAX_MODES];
  int rd_thresh_mult_sub8x8[MAX_REFS];

656
  int rd_threshes[MAX_SEGMENTS][BLOCK_SIZES][MAX_MODES];
657
  int rd_thresh_freq_fact[BLOCK_SIZES][MAX_MODES];
658
  int rd_thresh_sub8x8[MAX_SEGMENTS][BLOCK_SIZES][MAX_REFS];
659
  int rd_thresh_freq_sub8x8[BLOCK_SIZES][MAX_REFS];
660

661
  int64_t rd_comp_pred_diff[REFERENCE_MODES];
662
  int64_t rd_prediction_type_threshes[MAX_REF_FRAMES][REFERENCE_MODES];
663
  int64_t rd_tx_select_diff[TX_MODES];
664
  // FIXME(rbultje) can this overflow?
665
  int rd_tx_select_threshes[MAX_REF_FRAMES][TX_MODES];
John Koleszar's avatar
John Koleszar committed
666

667
  int64_t rd_filter_diff[SWITCHABLE_FILTER_CONTEXTS];
668
  int64_t rd_filter_threshes[MAX_REF_FRAMES][SWITCHABLE_FILTER_CONTEXTS];
669
  int64_t rd_filter_cache[SWITCHABLE_FILTER_CONTEXTS];
670
  int64_t mask_filter_rd;
671

John Koleszar's avatar
John Koleszar committed
672
673
  int RDMULT;
  int RDDIV;
John Koleszar's avatar
John Koleszar committed
674

John Koleszar's avatar
John Koleszar committed
675
  CODING_CONTEXT coding_context;
John Koleszar's avatar
John Koleszar committed
676

677
678
  int zbin_mode_boost;
  int zbin_mode_boost_enabled;
John Koleszar's avatar
John Koleszar committed
679
  int active_arnr_frames;           // <= cpi->oxcf.arnr_max_frames
680
  int active_arnr_strength;         // <= cpi->oxcf.arnr_max_strength
John Koleszar's avatar
John Koleszar committed
681

James Zern's avatar
James Zern committed
682
  double output_framerate;
John Koleszar's avatar
John Koleszar committed
683
684
685
  int64_t last_time_stamp_seen;
  int64_t last_end_time_stamp_seen;
  int64_t first_time_stamp_ever;
John Koleszar's avatar
John Koleszar committed
686

687
  RATE_CONTROL rc;
John Koleszar's avatar
John Koleszar committed
688

John Koleszar's avatar
John Koleszar committed
689
  int cq_target_quality;
690

691
692
693
  vp9_coeff_count coef_counts[TX_SIZES][PLANE_TYPES];
  vp9_coeff_probs_model frame_coef_probs[TX_SIZES][PLANE_TYPES];
  vp9_coeff_stats frame_branch_ct[TX_SIZES][PLANE_TYPES];
694

John Koleszar's avatar
John Koleszar committed
695
  struct vpx_codec_pkt_list  *output_pkt_list;
696

John Koleszar's avatar
John Koleszar committed
697
698
699
700
701
702
  MBGRAPH_FRAME_STATS mbgraph_stats[MAX_LAG_BUFFERS];
  int mbgraph_n_frames;             // number of frames filled in the above
  int static_mb_pct;                // % forced skip mbs by segmentation
  int seg0_progress, seg0_idx, seg0_cnt;

  // for real time encoding
703
  int speed;
John Koleszar's avatar
John Koleszar committed
704
705
706

  int cpu_used;
  int pass;
John Koleszar's avatar
John Koleszar committed
707

John Koleszar's avatar
John Koleszar committed
708
  int ref_frame_flags;
John Koleszar's avatar
John Koleszar committed
709

John Koleszar's avatar
John Koleszar committed
710
  SPEED_FEATURES sf;
John Koleszar's avatar
John Koleszar committed
711

Paul Wilkins's avatar
Paul Wilkins committed
712
  unsigned int max_mv_magnitude;
713
  int mv_step_param;
Paul Wilkins's avatar
Paul Wilkins committed
714

715
716
717
718
719
720
721
  // Default value is 1. From first pass stats, encode_breakout may be disabled.
  ENCODE_BREAKOUT_TYPE allow_encode_breakout;

  // Get threshold from external input. In real time mode, it can be
  // overwritten according to encoding speed.
  int encode_breakout;

John Koleszar's avatar
John Koleszar committed
722
  unsigned char *segmentation_map;
Paul Wilkins's avatar
Paul Wilkins committed
723

John Koleszar's avatar
John Koleszar committed
724
  // segment threashold for encode breakout
Paul Wilkins's avatar
Paul Wilkins committed
725
  int  segment_encode_breakout[MAX_SEGMENTS];
John Koleszar's avatar
John Koleszar committed
726

727
728
  unsigned char *complexity_map;

John Koleszar's avatar
John Koleszar committed
729
730
  unsigned char *active_map;
  unsigned int active_map_enabled;
John Koleszar's avatar
John Koleszar committed
731

732
  CYCLIC_REFRESH *cyclic_refresh;
733

John Koleszar's avatar
John Koleszar committed
734
  fractional_mv_step_fp *find_fractional_mv_step;
735
  fractional_mv_step_comp_fp *find_fractional_mv_step_comp;
736
737
738
  vp9_full_search_fn_t full_search_sad;
  vp9_refining_search_fn_t refining_search_sad;
  vp9_diamond_search_fn_t diamond_search_sad;
739
  vp9_variance_fn_ptr_t fn_ptr[BLOCK_SIZES];
Yaowu Xu's avatar
Yaowu Xu committed
740
741
742
  uint64_t time_receive_data;
  uint64_t time_compress_data;
  uint64_t time_pick_lpf;
743
  uint64_t time_encode_sb_row;
John Koleszar's avatar
John Koleszar committed
744

745
  struct twopass_rc twopass;
John Koleszar's avatar
John Koleszar committed
746

John Koleszar's avatar
John Koleszar committed
747
748
749
  YV12_BUFFER_CONFIG alt_ref_buffer;
  YV12_BUFFER_CONFIG *frames[MAX_LAG_BUFFERS];
  int fixed_divide[512];
John Koleszar's avatar
John Koleszar committed
750

751
#if CONFIG_INTERNAL_STATS
752
753
  unsigned int mode_chosen_counts[MAX_MODES];

John Koleszar's avatar
John Koleszar committed
754
755
756
757
758
  int    count;
  double total_y;
  double total_u;
  double total_v;
  double total;
759
760
761
  uint64_t total_sq_error;
  uint64_t total_samples;

John Koleszar's avatar
John Koleszar committed
762
763
764
765
  double totalp_y;
  double totalp_u;
  double totalp_v;
  double totalp;
766
767
768
  uint64_t totalp_sq_error;
  uint64_t totalp_samples;

John Koleszar's avatar
John Koleszar committed
769
770
771
  int    bytes;
  double summed_quality;
  double summed_weights;
772
773
  double summedp_quality;
  double summedp_weights;
John Koleszar's avatar
John Koleszar committed
774
775
776
777
778
779
780
781
782
  unsigned int tot_recode_hits;


  double total_ssimg_y;
  double total_ssimg_u;
  double total_ssimg_v;
  double total_ssimg_all;

  int b_calculate_ssimg;
John Koleszar's avatar
John Koleszar committed
783
#endif
John Koleszar's avatar
John Koleszar committed
784
  int b_calculate_psnr;
785

John Koleszar's avatar
John Koleszar committed
786
787
788
789
  // Per MB activity measurement
  unsigned int activity_avg;
  unsigned int *mb_activity_map;
  int *mb_norm_activity_map;
790

John Koleszar's avatar
John Koleszar committed
791
  int droppable;
Adrian Grange's avatar
Adrian Grange committed
792

John Koleszar's avatar
John Koleszar committed
793
  int dummy_packing;    /* flag to indicate if packing is dummy */
794

Dmitry Kovalev's avatar
Dmitry Kovalev committed
795
  unsigned int tx_stepdown_count[TX_SIZES];
Deb Mukherjee's avatar
Deb Mukherjee committed
796

797
798
  int initial_width;
  int initial_height;
799

800
801
  int use_svc;

802
  SVC svc;
803

804
805
806
807
808
809
810
811
812
813
814
815
#if CONFIG_MULTIPLE_ARF
  // ARF tracking variables.
  int multi_arf_enabled;
  unsigned int frame_coding_order_period;
  unsigned int new_frame_coding_order_period;
  int frame_coding_order[MAX_LAG_BUFFERS * 2];
  int arf_buffer_idx[MAX_LAG_BUFFERS * 3 / 2];
  int arf_weight[MAX_LAG_BUFFERS];
  int arf_buffered;
  int this_frame_weight;
  int max_arf_level;
#endif
Paul Wilkins's avatar
Paul Wilkins committed
816

817
818
819
820
#ifdef MODE_TEST_HIT_STATS
  // Debug / test stats
  int64_t mode_test_hits[BLOCK_SIZES];
#endif
821
} VP9_COMP;
John Koleszar's avatar
John Koleszar committed
822

823
824
825
826
827
void vp9_initialize_enc();

struct VP9_COMP *vp9_create_compressor(VP9_CONFIG *oxcf);
void vp9_remove_compressor(VP9_COMP *cpi);

828
void vp9_change_config(VP9_COMP *cpi, const VP9_CONFIG *oxcf);
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844

  // receive a frames worth of data. caller can assume that a copy of this
  // frame is made and not just a copy of the pointer..
int vp9_receive_raw_frame(VP9_COMP *cpi, unsigned int frame_flags,
                          YV12_BUFFER_CONFIG *sd, int64_t time_stamp,
                          int64_t end_time_stamp);

int vp9_get_compressed_data(VP9_COMP *cpi, unsigned int *frame_flags,
                            size_t *size, uint8_t *dest,
                            int64_t *time_stamp, int64_t *time_end, int flush);

int vp9_get_preview_raw_frame(VP9_COMP *cpi, YV12_BUFFER_CONFIG *dest,
                              vp9_ppflags_t *flags);

int vp9_use_as_reference(VP9_COMP *cpi, int ref_frame_flags);

845
void vp9_update_reference(VP9_COMP *cpi, int ref_frame_flags);
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876

int vp9_copy_reference_enc(VP9_COMP *cpi, VP9_REFFRAME ref_frame_flag,
                           YV12_BUFFER_CONFIG *sd);

int vp9_get_reference_enc(VP9_COMP *cpi, int index,
                          YV12_BUFFER_CONFIG **fb);

int vp9_set_reference_enc(VP9_COMP *cpi, VP9_REFFRAME ref_frame_flag,
                          YV12_BUFFER_CONFIG *sd);

int vp9_update_entropy(VP9_COMP *cpi, int update);

int vp9_set_roimap(VP9_COMP *cpi, unsigned char *map,
                   unsigned int rows, unsigned int cols,
                   int delta_q[MAX_SEGMENTS],
                   int delta_lf[MAX_SEGMENTS],
                   unsigned int threshold[MAX_SEGMENTS]);

int vp9_set_active_map(VP9_COMP *cpi, unsigned char *map,
                       unsigned int rows, unsigned int cols);

int vp9_set_internal_size(VP9_COMP *cpi,
                          VPX_SCALING horiz_mode, VPX_SCALING vert_mode);

int vp9_set_size_literal(VP9_COMP *cpi, unsigned int width,
                         unsigned int height);

void vp9_set_svc(VP9_COMP *cpi, int use_svc);

int vp9_get_quantizer(struct VP9_COMP *cpi);

877
878
static INLINE int get_ref_frame_idx(const VP9_COMP *cpi,
                                    MV_REFERENCE_FRAME ref_frame) {
879
880
881
882
883
884
885
886
887
  if (ref_frame == LAST_FRAME) {
    return cpi->lst_fb_idx;
  } else if (ref_frame == GOLDEN_FRAME) {
    return cpi->gld_fb_idx;
  } else {
    return cpi->alt_fb_idx;
  }
}

888
889
890
891
892
static INLINE YV12_BUFFER_CONFIG *get_ref_frame_buffer(
    VP9_COMP *cpi, MV_REFERENCE_FRAME ref_frame) {
  VP9_COMMON * const cm = &cpi->common;
  return &cm->frame_bufs[cm->ref_frame_map[get_ref_frame_idx(cpi, ref_frame)]]
      .buf;
893
894
}

895
void vp9_set_speed_features(VP9_COMP *cpi);
Yaowu Xu's avatar
Yaowu Xu committed
896

897
898
899
900
901
902
903
904
905
906
static INLINE int get_token_alloc(int mb_rows, int mb_cols) {
  // TODO(JBB): make this work for alpha channel and double check we can't
  // exceed this token count if we have a 32x32 transform crossing a boundary
  // at a multiple of 16.
  // mb_rows, cols are in units of 16 pixels. We assume 3 planes all at full
  // resolution. We assume up to 1 token per pixel, and then allow
  // a head room of 4.
  return mb_rows * mb_cols * (16 * 16 * 3 + 4);
}

907
908
int vp9_calc_ss_err(const YV12_BUFFER_CONFIG *source,
                    const YV12_BUFFER_CONFIG *reference);
909

910
void vp9_alloc_compressor_data(VP9_COMP *cpi);
911

Deb Mukherjee's avatar
Deb Mukherjee committed
912
int vp9_compute_qdelta(const VP9_COMP *cpi, double qstart, double qtarget);
913

914
915
916
int vp9_compute_qdelta_by_rate(VP9_COMP *cpi, int base_q_index,
                               double rate_target_ratio);

917
918
919
920
void vp9_scale_references(VP9_COMP *cpi);

void vp9_update_reference_frames(VP9_COMP *cpi);

921
922
923
924
extern const int q_trans[];

int64_t vp9_rescale(int64_t val, int64_t num, int denom);

925
926
927
static INLINE void set_ref_ptrs(VP9_COMMON *cm, MACROBLOCKD *xd,
                                MV_REFERENCE_FRAME ref0,
                                MV_REFERENCE_FRAME ref1) {
928
929
930
931
  xd->block_refs[0] = &cm->frame_refs[ref0 >= LAST_FRAME ? ref0 - LAST_FRAME
                                                         : 0];
  xd->block_refs[1] = &cm->frame_refs[ref1 >= LAST_FRAME ? ref1 - LAST_FRAME
                                                         : 0];
932
933
}

934
935
936
937
#ifdef __cplusplus
}  // extern "C"
#endif

938
#endif  // VP9_ENCODER_VP9_ONYX_INT_H_