AOMedia AV1 Codec
tpl_model.h
1/*
2 * Copyright (c) 2019, Alliance for Open Media. All rights reserved
3 *
4 * This source code is subject to the terms of the BSD 2 Clause License and
5 * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
6 * was not distributed with this source code in the LICENSE file, you can
7 * obtain it at www.aomedia.org/license/software. If the Alliance for Open
8 * Media Patent License 1.0 was not distributed with this source code in the
9 * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
10 */
11
12#ifndef AOM_AV1_ENCODER_TPL_MODEL_H_
13#define AOM_AV1_ENCODER_TPL_MODEL_H_
14
15#ifdef __cplusplus
16extern "C" {
17#endif
18
21struct AV1_PRIMARY;
22struct AV1_COMP;
23struct AV1_SEQ_CODING_TOOLS;
25struct EncodeFrameInput;
26struct GF_GROUP;
27struct TPL_INFO;
28
29#include "config/aom_config.h"
30
31#include "aom_scale/yv12config.h"
32
33#include "av1/common/mv.h"
34#include "av1/common/scale.h"
35#include "av1/encoder/block.h"
37#include "av1/encoder/ratectrl.h"
38
39static INLINE BLOCK_SIZE convert_length_to_bsize(int length) {
40 switch (length) {
41 case 64: return BLOCK_64X64;
42 case 32: return BLOCK_32X32;
43 case 16: return BLOCK_16X16;
44 case 8: return BLOCK_8X8;
45 case 4: return BLOCK_4X4;
46 default:
47 assert(0 && "Invalid block size for tpl model");
48 return BLOCK_16X16;
49 }
50}
51
52typedef struct AV1TplRowMultiThreadSync {
53#if CONFIG_MULTITHREAD
54 // Synchronization objects for top-right dependency.
55 pthread_mutex_t *mutex_;
56 pthread_cond_t *cond_;
57#endif
58 // Buffer to store the macroblock whose encoding is complete.
59 // num_finished_cols[i] stores the number of macroblocks which finished
60 // encoding in the ith macroblock row.
61 int *num_finished_cols;
62 // Number of extra macroblocks of the top row to be complete for encoding
63 // of the current macroblock to start. A value of 1 indicates top-right
64 // dependency.
65 int sync_range;
66 // Number of macroblock rows.
67 int rows;
68 // Number of threads processing the current tile.
69 int num_threads_working;
70} AV1TplRowMultiThreadSync;
71
72typedef struct AV1TplRowMultiThreadInfo {
73 // Row synchronization related function pointers.
74 void (*sync_read_ptr)(AV1TplRowMultiThreadSync *tpl_mt_sync, int r, int c);
75 void (*sync_write_ptr)(AV1TplRowMultiThreadSync *tpl_mt_sync, int r, int c,
76 int cols);
77} AV1TplRowMultiThreadInfo;
78
79// TODO(jingning): This needs to be cleaned up next.
80
81// TPL stats buffers are prepared for every frame in the GOP,
82// including (internal) overlays and (internal) arfs.
83// In addition, frames in the lookahead that are outside of the GOP
84// are also used.
85// Thus it should use
86// (gop_length) + (# overlays) + (MAX_LAG_BUFFERS - gop_len) =
87// MAX_LAG_BUFFERS + (# overlays)
88// 2 * MAX_LAG_BUFFERS is therefore a safe estimate.
89// TODO(bohanli): test setting it to 1.5 * MAX_LAG_BUFFER
90#define MAX_TPL_FRAME_IDX (2 * MAX_LAG_BUFFERS)
91// The first REF_FRAMES + 1 buffers are reserved.
92// tpl_data->tpl_frame starts after REF_FRAMES + 1
93#define MAX_LENGTH_TPL_FRAME_STATS (MAX_TPL_FRAME_IDX + REF_FRAMES + 1)
94#define TPL_DEP_COST_SCALE_LOG2 4
95
96#define TPL_EPSILON 0.0000001
97
98typedef struct TplTxfmStats {
99 int ready; // Whether abs_coeff_mean is ready
100 double abs_coeff_sum[256]; // Assume we are using 16x16 transform block
101 double abs_coeff_mean[256];
102 int txfm_block_count;
103 int coeff_num;
104} TplTxfmStats;
105
106typedef struct TplDepStats {
107 int64_t srcrf_sse;
108 int64_t srcrf_dist;
109 int64_t recrf_sse;
110 int64_t recrf_dist;
111 int64_t intra_sse;
112 int64_t intra_dist;
113 int64_t cmp_recrf_dist[2];
114 int64_t mc_dep_rate;
115 int64_t mc_dep_dist;
116 int64_t pred_error[INTER_REFS_PER_FRAME];
117 int32_t intra_cost;
118 int32_t inter_cost;
119 int32_t srcrf_rate;
120 int32_t recrf_rate;
121 int32_t intra_rate;
122 int32_t cmp_recrf_rate[2];
123 int_mv mv[INTER_REFS_PER_FRAME];
124 int8_t ref_frame_index[2];
125} TplDepStats;
126
127typedef struct TplDepFrame {
128 uint8_t is_valid;
129 TplDepStats *tpl_stats_ptr;
130 const YV12_BUFFER_CONFIG *gf_picture;
131 YV12_BUFFER_CONFIG *rec_picture;
132 int ref_map_index[REF_FRAMES];
133 int stride;
134 int width;
135 int height;
136 int mi_rows;
137 int mi_cols;
138 int base_rdmult;
139 uint32_t frame_display_index;
140} TplDepFrame;
141
146typedef struct TplParams {
150 int ready;
151
156
161
167 TplDepFrame tpl_stats_buffer[MAX_LENGTH_TPL_FRAME_STATS];
168
174 TplDepStats *tpl_stats_pool[MAX_LAG_BUFFERS];
175
182 TplTxfmStats *txfm_stats_list;
183
189
193 TplDepFrame *tpl_frame;
194
198 struct scale_factors sf;
199
204
210 const YV12_BUFFER_CONFIG *src_ref_frame[INTER_REFS_PER_FRAME];
211
217 const YV12_BUFFER_CONFIG *ref_frame[INTER_REFS_PER_FRAME];
218
223 AV1TplRowMultiThreadSync tpl_mt_sync;
224
229
230} TplParams;
231
232#if CONFIG_BITRATE_ACCURACY || CONFIG_RATECTRL_LOG
233#define VBR_RC_INFO_MAX_FRAMES 500
234#endif // CONFIG_BITRATE_ACCURACY || CONFIG_RATECTRL_LOG
235
236#if CONFIG_BITRATE_ACCURACY
237
242typedef struct {
243 int ready;
244 double total_bit_budget; // The total bit budget of the entire video
245 int show_frame_count; // Number of show frames in the entire video
246
247 int gop_showframe_count; // The number of show frames in the current gop
248 double gop_bit_budget; // The bitbudget for the current gop
249 double scale_factors[FRAME_UPDATE_TYPES]; // Scale factors to improve the
250 // budget estimation
251 double mv_scale_factors[FRAME_UPDATE_TYPES]; // Scale factors to improve
252 // MV entropy estimation
253
254 // === Below this line are GOP related data that will be updated per GOP ===
255 int base_q_index; // Stores the base q index.
256 int q_index_list_ready;
257 int q_index_list[VBR_RC_INFO_MAX_FRAMES]; // q indices for the current
258 // GOP
259
260 // Array to store qstep_ratio for each frame in a GOP
261 double qstep_ratio_list[VBR_RC_INFO_MAX_FRAMES];
262
263#if CONFIG_THREE_PASS
264 TplTxfmStats txfm_stats_list[VBR_RC_INFO_MAX_FRAMES];
265 FRAME_UPDATE_TYPE update_type_list[VBR_RC_INFO_MAX_FRAMES];
266 int gop_start_idx_list[VBR_RC_INFO_MAX_FRAMES];
267 int gop_length_list[VBR_RC_INFO_MAX_FRAMES];
268 int cur_gop_idx;
269 int total_frame_count;
270 int gop_count;
271#endif // CONFIG_THREE_PASS
272} VBR_RATECTRL_INFO;
273
274static INLINE void vbr_rc_reset_gop_data(VBR_RATECTRL_INFO *vbr_rc_info) {
275 vbr_rc_info->q_index_list_ready = 0;
276 av1_zero(vbr_rc_info->q_index_list);
277}
278
279void av1_vbr_rc_init(VBR_RATECTRL_INFO *vbr_rc_info, double total_bit_budget,
280 int show_frame_count);
281
282int av1_vbr_rc_frame_coding_idx(const VBR_RATECTRL_INFO *vbr_rc_info,
283 int gf_frame_index);
284
285void av1_vbr_rc_append_tpl_info(VBR_RATECTRL_INFO *vbr_rc_info,
286 const struct TPL_INFO *tpl_info);
287
288void av1_vbr_rc_set_gop_bit_budget(VBR_RATECTRL_INFO *vbr_rc_info,
289 int gop_showframe_count);
290
291void av1_vbr_rc_compute_q_indices(int base_q_index, int frame_count,
292 const double *qstep_ratio_list,
293 aom_bit_depth_t bit_depth, int *q_index_list);
294
303void av1_vbr_rc_update_q_index_list(VBR_RATECTRL_INFO *vbr_rc_info,
304 const TplParams *tpl_data,
305 const struct GF_GROUP *gf_group,
306 aom_bit_depth_t bit_depth);
307/*
308 *!\brief Compute the number of bits needed to encode a GOP
309 *
310 * \param[in] base_q_index base layer q_index
311 * \param[in] bit_depth bit depth
312 * \param[in] update_type_scale_factors array of scale factors for each
313 * update_type
314 * \param[in] frame_count size of update_type_list,
315 * qstep_ratio_list stats_list,
316 * q_index_list and
317 * estimated_bitrate_byframe
318 * \param[in] update_type_list array of update_type, one per frame
319 * \param[in] qstep_ratio_list array of qstep_ratio, one per frame
320 * \param[in] stats_list array of transform stats, one per
321 * frame
322 * \param[out] q_index_list array of q_index, one per frame
323 * \param[out] estimated_bitrate_byframe array to keep track of frame
324 * bitrate
325 *
326 * \return The estimated GOP bitrate.
327 *
328 */
329double av1_vbr_rc_info_estimate_gop_bitrate(
330 int base_q_index, aom_bit_depth_t bit_depth,
331 const double *update_type_scale_factors, int frame_count,
332 const FRAME_UPDATE_TYPE *update_type_list, const double *qstep_ratio_list,
333 const TplTxfmStats *stats_list, int *q_index_list,
334 double *estimated_bitrate_byframe);
335
357int av1_vbr_rc_info_estimate_base_q(
358 double bit_budget, aom_bit_depth_t bit_depth,
359 const double *update_type_scale_factors, int frame_count,
360 const FRAME_UPDATE_TYPE *update_type_list, const double *qstep_ratio_list,
361 const TplTxfmStats *stats_list, int *q_index_list,
362 double *estimated_bitrate_byframe);
363
364#endif // CONFIG_BITRATE_ACCURACY
365
366#if CONFIG_RD_COMMAND
367typedef enum {
368 RD_OPTION_NONE,
369 RD_OPTION_SET_Q,
370 RD_OPTION_SET_Q_RDMULT
371} RD_OPTION;
372
373typedef struct RD_COMMAND {
374 RD_OPTION option_ls[MAX_LENGTH_TPL_FRAME_STATS];
375 int q_index_ls[MAX_LENGTH_TPL_FRAME_STATS];
376 int rdmult_ls[MAX_LENGTH_TPL_FRAME_STATS];
377 int frame_count;
378 int frame_index;
379} RD_COMMAND;
380
381void av1_read_rd_command(const char *filepath, RD_COMMAND *rd_command);
382#endif // CONFIG_RD_COMMAND
383
392void av1_setup_tpl_buffers(struct AV1_PRIMARY *const ppi,
393 CommonModeInfoParams *const mi_params, int width,
394 int height, int byte_alignment, int lag_in_frames);
395
407int av1_tpl_setup_stats(struct AV1_COMP *cpi, int gop_eval,
408 const struct EncodeFrameParams *const frame_params);
409
412void av1_tpl_preload_rc_estimate(
413 struct AV1_COMP *cpi, const struct EncodeFrameParams *const frame_params);
414
415int av1_tpl_ptr_pos(int mi_row, int mi_col, int stride, uint8_t right_shift);
416
417void av1_init_tpl_stats(TplParams *const tpl_data);
418
419int av1_tpl_stats_ready(const TplParams *tpl_data, int gf_frame_index);
420
421void av1_tpl_rdmult_setup(struct AV1_COMP *cpi);
422
423void av1_tpl_rdmult_setup_sb(struct AV1_COMP *cpi, MACROBLOCK *const x,
424 BLOCK_SIZE sb_size, int mi_row, int mi_col);
425
426void av1_mc_flow_dispenser_row(struct AV1_COMP *cpi,
427 TplTxfmStats *tpl_txfm_stats, MACROBLOCK *x,
428 int mi_row, BLOCK_SIZE bsize, TX_SIZE tx_size);
429
442double av1_exponential_entropy(double q_step, double b);
443
457double av1_laplace_entropy(double q_step, double b, double zero_bin_ratio);
458
476double av1_laplace_estimate_frame_rate(int q_index, int block_count,
477 const double *abs_coeff_mean,
478 int coeff_num);
479
480/*
481 *!\brief Init TplTxfmStats
482 *
483 * \param[in] tpl_txfm_stats a structure for storing transform stats
484 *
485 */
486void av1_init_tpl_txfm_stats(TplTxfmStats *tpl_txfm_stats);
487
488#if CONFIG_BITRATE_ACCURACY
489/*
490 *!\brief Accumulate TplTxfmStats
491 *
492 * \param[in] sub_stats a structure for storing sub transform stats
493 * \param[out] accumulated_stats a structure for storing accumulated
494 *transform stats
495 *
496 */
497void av1_accumulate_tpl_txfm_stats(const TplTxfmStats *sub_stats,
498 TplTxfmStats *accumulated_stats);
499
500/*
501 *!\brief Record a transform block into TplTxfmStats
502 *
503 * \param[in] tpl_txfm_stats A structure for storing transform stats
504 * \param[out] coeff An array of transform coefficients. Its size
505 * should equal to tpl_txfm_stats.coeff_num.
506 *
507 */
508void av1_record_tpl_txfm_block(TplTxfmStats *tpl_txfm_stats,
509 const tran_low_t *coeff);
510
511/*
512 *!\brief Update abs_coeff_mean and ready of txfm_stats
513 * If txfm_block_count > 0, this function will use abs_coeff_sum and
514 * txfm_block_count to compute abs_coeff_mean. Moreover, reday flag
515 * will be set to one.
516 *
517 * \param[in] txfm_stats A structure for storing transform stats
518 */
519void av1_tpl_txfm_stats_update_abs_coeff_mean(TplTxfmStats *txfm_stats);
520#endif // CONFIG_BITRATE_ACCURACY
521
537double av1_estimate_coeff_entropy(double q_step, double b,
538 double zero_bin_ratio, int qcoeff);
539
552double av1_estimate_txfm_block_entropy(int q_index,
553 const double *abs_coeff_mean,
554 int *qcoeff_arr, int coeff_num);
555
556// TODO(angiebird): Add doxygen description here.
557int64_t av1_delta_rate_cost(int64_t delta_rate, int64_t recrf_dist,
558 int64_t srcrf_dist, int pix_num);
559
575int av1_get_overlap_area(int row_a, int col_a, int row_b, int col_b, int width,
576 int height);
577
587int av1_tpl_get_q_index(const TplParams *tpl_data, int gf_frame_index,
588 int leaf_qindex, aom_bit_depth_t bit_depth);
589
597double av1_tpl_get_frame_importance(const TplParams *tpl_data,
598 int gf_frame_index);
599
610double av1_tpl_get_qstep_ratio(const TplParams *tpl_data, int gf_frame_index);
611
620int av1_get_q_index_from_qstep_ratio(int leaf_qindex, double qstep_ratio,
621 aom_bit_depth_t bit_depth);
622
637int_mv av1_compute_mv_difference(const TplDepFrame *tpl_frame, int row, int col,
638 int step, int tpl_stride, int right_shift);
639
647double av1_tpl_compute_frame_mv_entropy(const TplDepFrame *tpl_frame,
648 uint8_t right_shift);
649
650#if CONFIG_RATECTRL_LOG
651typedef struct {
652 int coding_frame_count;
653 int base_q_index;
654
655 // Encode decision
656 int q_index_list[VBR_RC_INFO_MAX_FRAMES];
657 double qstep_ratio_list[VBR_RC_INFO_MAX_FRAMES];
658 FRAME_UPDATE_TYPE update_type_list[VBR_RC_INFO_MAX_FRAMES];
659
660 // Frame stats
661 TplTxfmStats txfm_stats_list[VBR_RC_INFO_MAX_FRAMES];
662
663 // Estimated encode results
664 double est_coeff_rate_list[VBR_RC_INFO_MAX_FRAMES];
665
666 // Actual encode results
667 double act_rate_list[VBR_RC_INFO_MAX_FRAMES];
668 double act_coeff_rate_list[VBR_RC_INFO_MAX_FRAMES];
669} RATECTRL_LOG;
670
671static INLINE void rc_log_init(RATECTRL_LOG *rc_log) { av1_zero(*rc_log); }
672
673static INLINE void rc_log_frame_stats(RATECTRL_LOG *rc_log, int coding_index,
674 const TplTxfmStats *txfm_stats) {
675 rc_log->txfm_stats_list[coding_index] = *txfm_stats;
676}
677
678static INLINE void rc_log_frame_encode_param(RATECTRL_LOG *rc_log,
679 int coding_index,
680 double qstep_ratio, int q_index,
681 FRAME_UPDATE_TYPE update_type) {
682 rc_log->qstep_ratio_list[coding_index] = qstep_ratio;
683 rc_log->q_index_list[coding_index] = q_index;
684 rc_log->update_type_list[coding_index] = update_type;
685 const TplTxfmStats *txfm_stats = &rc_log->txfm_stats_list[coding_index];
686 rc_log->est_coeff_rate_list[coding_index] = 0;
687 if (txfm_stats->ready) {
688 rc_log->est_coeff_rate_list[coding_index] = av1_laplace_estimate_frame_rate(
689 q_index, txfm_stats->txfm_block_count, txfm_stats->abs_coeff_mean,
690 txfm_stats->coeff_num);
691 }
692}
693
694static INLINE void rc_log_frame_entropy(RATECTRL_LOG *rc_log, int coding_index,
695 double act_rate,
696 double act_coeff_rate) {
697 rc_log->act_rate_list[coding_index] = act_rate;
698 rc_log->act_coeff_rate_list[coding_index] = act_coeff_rate;
699}
700
701static INLINE void rc_log_record_chunk_info(RATECTRL_LOG *rc_log,
702 int base_q_index,
703 int coding_frame_count) {
704 rc_log->base_q_index = base_q_index;
705 rc_log->coding_frame_count = coding_frame_count;
706}
707
708static INLINE void rc_log_show(const RATECTRL_LOG *rc_log) {
709 printf("= chunk 1\n");
710 printf("coding_frame_count %d base_q_index %d\n", rc_log->coding_frame_count,
711 rc_log->base_q_index);
712 printf("= frame %d\n", rc_log->coding_frame_count);
713 for (int coding_idx = 0; coding_idx < rc_log->coding_frame_count;
714 coding_idx++) {
715 printf(
716 "coding_idx %d update_type %d q %d qstep_ratio %f est_coeff_rate %f "
717 "act_coeff_rate %f act_rate %f\n",
718 coding_idx, rc_log->update_type_list[coding_idx],
719 rc_log->q_index_list[coding_idx], rc_log->qstep_ratio_list[coding_idx],
720 rc_log->est_coeff_rate_list[coding_idx],
721 rc_log->act_coeff_rate_list[coding_idx],
722 rc_log->act_rate_list[coding_idx]);
723 }
724}
725#endif // CONFIG_RATECTRL_LOG
726
728#ifdef __cplusplus
729} // extern "C"
730#endif
731
732#endif // AOM_AV1_ENCODER_TPL_MODEL_H_
enum aom_bit_depth aom_bit_depth_t
Bit depth for codecThis enumeration determines the bit depth of the codec.
int av1_tpl_setup_stats(struct AV1_COMP *cpi, int gop_eval, const struct EncodeFrameParams *const frame_params)
Implements temporal dependency modelling for a GOP (GF/ARF group) and selects between 16 and 32 frame...
Describes look ahead buffer operations.
Top level encoder structure.
Definition encoder.h:2815
Top level primary encoder structure.
Definition encoder.h:2519
Params related to MB_MODE_INFO arrays and related info.
Definition av1_common_int.h:503
Input frames and last input frame.
Definition encoder.h:3572
contains per-frame encoding parameters decided upon by av1_encode_strategy() and passed down to av1_e...
Definition encoder.h:3584
Data related to the current GF/ARF group and the individual frames within the group.
Definition firstpass.h:354
Params related to temporal dependency model.
Definition tpl_model.h:146
const YV12_BUFFER_CONFIG * src_ref_frame[INTER_REFS_PER_FRAME]
Definition tpl_model.h:210
struct scale_factors sf
Definition tpl_model.h:198
int ready
Definition tpl_model.h:150
TplDepFrame tpl_stats_buffer[MAX_LENGTH_TPL_FRAME_STATS]
Definition tpl_model.h:167
uint8_t tpl_bsize_1d
Definition tpl_model.h:160
AV1TplRowMultiThreadSync tpl_mt_sync
Definition tpl_model.h:223
TplDepFrame * tpl_frame
Definition tpl_model.h:193
int border_in_pixels
Definition tpl_model.h:228
TplDepStats * tpl_stats_pool[MAX_LAG_BUFFERS]
Definition tpl_model.h:174
TplTxfmStats * txfm_stats_list
Definition tpl_model.h:182
YV12_BUFFER_CONFIG tpl_rec_pool[MAX_LAG_BUFFERS]
Definition tpl_model.h:188
uint8_t tpl_stats_block_mis_log2
Definition tpl_model.h:155
int frame_idx
Definition tpl_model.h:203
const YV12_BUFFER_CONFIG * ref_frame[INTER_REFS_PER_FRAME]
Definition tpl_model.h:217
Encoder's parameters related to the current coding block.
Definition block.h:878
YV12 frame buffer data structure.
Definition yv12config.h:44