2 * Copyright (c) 2010 The WebM project authors. All Rights Reserved.
4 * Use of this source code is governed by a BSD-style license
5 * that can be found in the LICENSE file in the root of the source
6 * 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.
11 #ifndef VP9_ENCODER_VP9_ENCODER_H_
12 #define VP9_ENCODER_VP9_ENCODER_H_
16 #include "./vpx_config.h"
17 #include "vpx/internal/vpx_codec_internal.h"
18 #include "vpx/vp8cx.h"
19 #if CONFIG_INTERNAL_STATS
20 #include "vpx_dsp/ssim.h"
22 #include "vpx_dsp/variance.h"
23 #include "vpx_util/vpx_thread.h"
25 #include "vp9/common/vp9_alloccommon.h"
26 #include "vp9/common/vp9_ppflags.h"
27 #include "vp9/common/vp9_entropymode.h"
28 #include "vp9/common/vp9_thread_common.h"
29 #include "vp9/common/vp9_onyxc_int.h"
31 #include "vp9/encoder/vp9_aq_cyclicrefresh.h"
32 #include "vp9/encoder/vp9_context_tree.h"
33 #include "vp9/encoder/vp9_encodemb.h"
34 #include "vp9/encoder/vp9_firstpass.h"
35 #include "vp9/encoder/vp9_lookahead.h"
36 #include "vp9/encoder/vp9_mbgraph.h"
37 #include "vp9/encoder/vp9_mcomp.h"
38 #include "vp9/encoder/vp9_noise_estimate.h"
39 #include "vp9/encoder/vp9_quantize.h"
40 #include "vp9/encoder/vp9_ratectrl.h"
41 #include "vp9/encoder/vp9_rd.h"
42 #include "vp9/encoder/vp9_speed_features.h"
43 #include "vp9/encoder/vp9_svc_layercontext.h"
44 #include "vp9/encoder/vp9_tokenize.h"
46 #if CONFIG_VP9_TEMPORAL_DENOISING
47 #include "vp9/encoder/vp9_denoiser.h"
55 int nmvjointcost[MV_JOINTS];
56 int nmvcosts[2][MV_VALS];
57 int nmvcosts_hp[2][MV_VALS];
59 vpx_prob segment_pred_probs[PREDICTION_PROBS];
61 unsigned char *last_frame_seg_map_copy;
63 // 0 = Intra, Last, GF, ARF
64 signed char last_ref_lf_deltas[MAX_REF_LF_DELTAS];
66 signed char last_mode_lf_deltas[MAX_MODE_LF_DELTAS];
73 // encode_breakout is disabled.
74 ENCODE_BREAKOUT_DISABLED = 0,
75 // encode_breakout is enabled.
76 ENCODE_BREAKOUT_ENABLED = 1,
77 // encode_breakout is enabled with small max_thresh limit.
78 ENCODE_BREAKOUT_LIMITED = 2
79 } ENCODE_BREAKOUT_TYPE;
89 // Good Quality Fast Encoding. The encoder balances quality with the amount of
90 // time it takes to encode the output. Speed setting controls how fast.
93 // The encoder places priority on the quality of the output over encoding
94 // speed. The output is compressed at the highest possible quality. This
95 // option takes the longest amount of time to encode. Speed setting ignored.
98 // Realtime/Live Encoding. This mode is optimized for realtime encoding (for
99 // example, capturing a television signal or feed from a live camera). Speed
100 // setting controls how fast.
105 FRAMEFLAGS_KEY = 1 << 0,
106 FRAMEFLAGS_GOLDEN = 1 << 1,
107 FRAMEFLAGS_ALTREF = 1 << 2,
114 CYCLIC_REFRESH_AQ = 3,
116 AQ_MODE_COUNT // This should always be the last member of the enum
120 RESIZE_NONE = 0, // No frame resizing allowed (except for SVC).
121 RESIZE_FIXED = 1, // All frames are coded at the specified dimension.
122 RESIZE_DYNAMIC = 2 // Coded size of each frame is determined by the codec.
125 typedef struct VP9EncoderConfig {
126 BITSTREAM_PROFILE profile;
127 vpx_bit_depth_t bit_depth; // Codec bit-depth.
128 int width; // width of data passed to the compressor
129 int height; // height of data passed to the compressor
130 unsigned int input_bit_depth; // Input bit depth.
131 double init_framerate; // set to passed in framerate
132 int64_t target_bandwidth; // bandwidth to be used in kilobits per second
134 int noise_sensitivity; // pre processing blur: recommendation 0
135 int sharpness; // sharpening output: recommendation 0:
137 // maximum allowed bitrate for any intra frame in % of bitrate target.
138 unsigned int rc_max_intra_bitrate_pct;
139 // maximum allowed bitrate for any inter frame in % of bitrate target.
140 unsigned int rc_max_inter_bitrate_pct;
141 // percent of rate boost for golden frame in CBR mode.
142 unsigned int gf_cbr_boost_pct;
147 // Key Framing Operations
148 int auto_key; // autodetect cut scenes and set the keyframes
149 int key_freq; // maximum distance to key frame.
151 int lag_in_frames; // how many frames lag before we start encoding
153 // ----------------------------------------------------------------
154 // DATARATE CONTROL OPTIONS
156 // vbr, cbr, constrained quality or constant quality
157 enum vpx_rc_mode rc_mode;
159 // buffer targeting aggressiveness
163 // buffering parameters
164 int64_t starting_buffer_level_ms;
165 int64_t optimal_buffer_level_ms;
166 int64_t maximum_buffer_size_ms;
168 // Frame drop threshold.
169 int drop_frames_water_mark;
171 // controlling quality
176 AQ_MODE aq_mode; // Adaptive Quantization mode
178 // Internal frame size scaling.
179 RESIZE_TYPE resize_mode;
180 int scaled_frame_width;
181 int scaled_frame_height;
183 // Enable feature to reduce the frame quantization every x frames.
184 int frame_periodic_boost;
186 // two pass datarate control
187 int two_pass_vbrbias; // two pass datarate control tweaks
188 int two_pass_vbrmin_section;
189 int two_pass_vbrmax_section;
190 // END DATARATE CONTROL OPTIONS
191 // ----------------------------------------------------------------
193 // Spatial and temporal scalability.
194 int ss_number_layers; // Number of spatial layers.
195 int ts_number_layers; // Number of temporal layers.
196 // Bitrate allocation for spatial layers.
197 int layer_target_bitrate[VPX_MAX_LAYERS];
198 int ss_target_bitrate[VPX_SS_MAX_LAYERS];
199 int ss_enable_auto_arf[VPX_SS_MAX_LAYERS];
200 // Bitrate allocation (CBR mode) and framerate factor, for temporal layers.
201 int ts_rate_decimator[VPX_TS_MAX_LAYERS];
205 int encode_breakout; // early breakout : for video conf recommend 800
207 /* Bitfield defining the error resiliency features to enable.
208 * Can provide decodable frames after losses in previous
209 * frames and decodable partitions after losses in the same frame.
211 unsigned int error_resilient_mode;
213 /* Bitfield defining the parallel decoding mode where the
214 * decoding in successive frames may be conducted in parallel
215 * just by decoding the frame headers.
217 unsigned int frame_parallel_decoding_mode;
230 vpx_fixed_buf_t two_pass_stats_in;
231 struct vpx_codec_pkt_list *output_pkt_list;
233 #if CONFIG_FP_MB_STATS
234 vpx_fixed_buf_t firstpass_mb_stats_in;
238 vp9e_tune_content content;
239 #if CONFIG_VP9_HIGHBITDEPTH
240 int use_highbitdepth;
242 vpx_color_space_t color_space;
243 vpx_color_range_t color_range;
246 VP9E_TEMPORAL_LAYERING_MODE temporal_layering_mode;
249 static INLINE int is_lossless_requested(const VP9EncoderConfig *cfg) {
250 return cfg->best_allowed_q == 0 && cfg->worst_allowed_q == 0;
253 // TODO(jingning) All spatially adaptive variables should go to TileDataEnc.
254 typedef struct TileDataEnc {
256 int thresh_freq_fact[BLOCK_SIZES][MAX_MODES];
257 int mode_map[BLOCK_SIZES][MAX_MODES];
260 typedef struct RD_COUNTS {
261 vp9_coeff_count coef_counts[TX_SIZES][PLANE_TYPES];
262 int64_t comp_pred_diff[REFERENCE_MODES];
263 int64_t filter_diff[SWITCHABLE_FILTER_CONTEXTS];
268 typedef struct ThreadData {
271 FRAME_COUNTS *counts;
273 PICK_MODE_CONTEXT *leaf_tree;
278 struct EncWorkerData;
280 typedef struct ActiveMap {
293 typedef struct IMAGE_STAT {
298 typedef struct VP9_COMP {
301 MB_MODE_INFO_EXT *mbmi_ext_base;
302 DECLARE_ALIGNED(16, int16_t, y_dequant[QINDEX_RANGE][8]);
303 DECLARE_ALIGNED(16, int16_t, uv_dequant[QINDEX_RANGE][8]);
305 VP9EncoderConfig oxcf;
306 struct lookahead_ctx *lookahead;
307 struct lookahead_entry *alt_ref_source;
309 YV12_BUFFER_CONFIG *Source;
310 YV12_BUFFER_CONFIG *Last_Source; // NULL for first frame and alt_ref frames
311 YV12_BUFFER_CONFIG *un_scaled_source;
312 YV12_BUFFER_CONFIG scaled_source;
313 YV12_BUFFER_CONFIG *unscaled_last_source;
314 YV12_BUFFER_CONFIG scaled_last_source;
316 TileDataEnc *tile_data;
317 int allocated_tiles; // Keep track of memory allocated for tiles.
319 // For a still frame, this flag is set to 1 to skip partition search.
320 int partition_search_skippable_frame;
322 int scaled_ref_idx[MAX_REF_FRAMES];
327 int refresh_last_frame;
328 int refresh_golden_frame;
329 int refresh_alt_ref_frame;
331 int ext_refresh_frame_flags_pending;
332 int ext_refresh_last_frame;
333 int ext_refresh_golden_frame;
334 int ext_refresh_alt_ref_frame;
336 int ext_refresh_frame_context_pending;
337 int ext_refresh_frame_context;
339 YV12_BUFFER_CONFIG last_frame_uf;
341 TOKENEXTRA *tile_tok[4][1 << 6];
342 unsigned int tok_count[4][1 << 6];
344 // Ambient reconstruction err target for force key frames
349 CODING_CONTEXT coding_context;
354 int *nmvsadcosts_hp[2];
356 int64_t last_time_stamp_seen;
357 int64_t last_end_time_stamp_seen;
358 int64_t first_time_stamp_ever;
363 int interp_filter_selected[MAX_REF_FRAMES][SWITCHABLE];
365 struct vpx_codec_pkt_list *output_pkt_list;
367 MBGRAPH_FRAME_STATS mbgraph_stats[MAX_LAG_BUFFERS];
368 int mbgraph_n_frames; // number of frames filled in the above
369 int static_mb_pct; // % forced skip mbs by segmentation
374 unsigned int max_mv_magnitude;
377 int allow_comp_inter_inter;
379 // Default value is 1. From first pass stats, encode_breakout may be disabled.
380 ENCODE_BREAKOUT_TYPE allow_encode_breakout;
382 // Get threshold from external input. A suggested threshold is 800 for HD
383 // clips, and 300 for < HD clips.
386 unsigned char *segmentation_map;
388 // segment threashold for encode breakout
389 int segment_encode_breakout[MAX_SEGMENTS];
391 CYCLIC_REFRESH *cyclic_refresh;
392 ActiveMap active_map;
394 fractional_mv_step_fp *find_fractional_mv_step;
395 vp9_full_search_fn_t full_search_sad;
396 vp9_diamond_search_fn_t diamond_search_sad;
397 vp9_variance_fn_ptr_t fn_ptr[BLOCK_SIZES];
398 uint64_t time_receive_data;
399 uint64_t time_compress_data;
400 uint64_t time_pick_lpf;
401 uint64_t time_encode_sb_row;
403 #if CONFIG_FP_MB_STATS
409 YV12_BUFFER_CONFIG alt_ref_buffer;
412 #if CONFIG_INTERNAL_STATS
413 unsigned int mode_chosen_counts[MAX_MODES];
416 uint64_t total_sq_error;
417 uint64_t total_samples;
420 uint64_t totalp_sq_error;
421 uint64_t totalp_samples;
424 double total_blockiness;
425 double worst_blockiness;
428 double summed_quality;
429 double summed_weights;
430 double summedp_quality;
431 double summedp_weights;
432 unsigned int tot_recode_hits;
439 int b_calculate_ssimg;
440 int b_calculate_blockiness;
442 int b_calculate_consistency;
444 double total_inconsistency;
445 double worst_consistency;
449 int b_calculate_psnr;
455 int initial_mbs; // Number of MBs in the full-size frame; to be used to
456 // normalize the firstpass stats. This will differ from the
457 // number of MBs in the current frame when the frame is
464 // Store frame variance info in SOURCE_VAR_BASED_PARTITION search type.
465 diff *source_diff_var;
466 // The threshold used in SOURCE_VAR_BASED_PARTITION search type.
467 unsigned int source_var_thresh;
468 int frames_till_next_var_check;
472 search_site_config ss_cfg;
474 int mbmode_cost[INTRA_MODES];
475 unsigned int inter_mode_cost[INTER_MODE_CONTEXTS][INTER_MODES];
476 int intra_uv_mode_cost[FRAME_TYPES][INTRA_MODES][INTRA_MODES];
477 int y_mode_costs[INTRA_MODES][INTRA_MODES][INTRA_MODES];
478 int switchable_interp_costs[SWITCHABLE_FILTER_CONTEXTS][SWITCHABLE_FILTERS];
479 int partition_cost[PARTITION_CONTEXTS][PARTITION_TYPES];
481 int multi_arf_allowed;
482 int multi_arf_enabled;
483 int multi_arf_last_grp_enabled;
485 #if CONFIG_VP9_TEMPORAL_DENOISING
486 VP9_DENOISER denoiser;
491 int resize_scale_num;
492 int resize_scale_den;
494 int resize_buffer_underflow;
497 int use_skin_detection;
499 NOISE_ESTIMATE noise_estimate;
501 // VAR_BASED_PARTITION thresholds
502 // 0 - threshold_64x64; 1 - threshold_32x32;
503 // 2 - threshold_16x16; 3 - vbp_threshold_8x8;
504 int64_t vbp_thresholds[4];
505 int64_t vbp_threshold_minmax;
506 int64_t vbp_threshold_sad;
507 BLOCK_SIZE vbp_bsize_min;
512 struct EncWorkerData *tile_thr_data;
513 VP9LfSync lf_row_sync;
516 void vp9_initialize_enc(void);
518 struct VP9_COMP *vp9_create_compressor(VP9EncoderConfig *oxcf,
519 BufferPool *const pool);
520 void vp9_remove_compressor(VP9_COMP *cpi);
522 void vp9_change_config(VP9_COMP *cpi, const VP9EncoderConfig *oxcf);
524 // receive a frames worth of data. caller can assume that a copy of this
525 // frame is made and not just a copy of the pointer..
526 int vp9_receive_raw_frame(VP9_COMP *cpi, unsigned int frame_flags,
527 YV12_BUFFER_CONFIG *sd, int64_t time_stamp,
528 int64_t end_time_stamp);
530 int vp9_get_compressed_data(VP9_COMP *cpi, unsigned int *frame_flags,
531 size_t *size, uint8_t *dest,
532 int64_t *time_stamp, int64_t *time_end, int flush);
534 int vp9_get_preview_raw_frame(VP9_COMP *cpi, YV12_BUFFER_CONFIG *dest,
535 vp9_ppflags_t *flags);
537 int vp9_use_as_reference(VP9_COMP *cpi, int ref_frame_flags);
539 void vp9_update_reference(VP9_COMP *cpi, int ref_frame_flags);
541 int vp9_copy_reference_enc(VP9_COMP *cpi, VP9_REFFRAME ref_frame_flag,
542 YV12_BUFFER_CONFIG *sd);
544 int vp9_set_reference_enc(VP9_COMP *cpi, VP9_REFFRAME ref_frame_flag,
545 YV12_BUFFER_CONFIG *sd);
547 int vp9_update_entropy(VP9_COMP *cpi, int update);
549 int vp9_set_active_map(VP9_COMP *cpi, unsigned char *map, int rows, int cols);
551 int vp9_get_active_map(VP9_COMP *cpi, unsigned char *map, int rows, int cols);
553 int vp9_set_internal_size(VP9_COMP *cpi,
554 VPX_SCALING horiz_mode, VPX_SCALING vert_mode);
556 int vp9_set_size_literal(VP9_COMP *cpi, unsigned int width,
557 unsigned int height);
559 void vp9_set_svc(VP9_COMP *cpi, int use_svc);
561 int vp9_get_quantizer(struct VP9_COMP *cpi);
563 static INLINE int frame_is_kf_gf_arf(const VP9_COMP *cpi) {
564 return frame_is_intra_only(&cpi->common) ||
565 cpi->refresh_alt_ref_frame ||
566 (cpi->refresh_golden_frame && !cpi->rc.is_src_frame_alt_ref);
569 static INLINE int get_ref_frame_map_idx(const VP9_COMP *cpi,
570 MV_REFERENCE_FRAME ref_frame) {
571 if (ref_frame == LAST_FRAME) {
572 return cpi->lst_fb_idx;
573 } else if (ref_frame == GOLDEN_FRAME) {
574 return cpi->gld_fb_idx;
576 return cpi->alt_fb_idx;
580 static INLINE int get_ref_frame_buf_idx(const VP9_COMP *const cpi,
582 const VP9_COMMON *const cm = &cpi->common;
583 const int map_idx = get_ref_frame_map_idx(cpi, ref_frame);
584 return (map_idx != INVALID_IDX) ? cm->ref_frame_map[map_idx] : INVALID_IDX;
587 static INLINE YV12_BUFFER_CONFIG *get_ref_frame_buffer(
588 VP9_COMP *cpi, MV_REFERENCE_FRAME ref_frame) {
589 VP9_COMMON *const cm = &cpi->common;
590 const int buf_idx = get_ref_frame_buf_idx(cpi, ref_frame);
592 buf_idx != INVALID_IDX ? &cm->buffer_pool->frame_bufs[buf_idx].buf : NULL;
595 static INLINE int get_token_alloc(int mb_rows, int mb_cols) {
596 // TODO(JBB): double check we can't exceed this token count if we have a
597 // 32x32 transform crossing a boundary at a multiple of 16.
598 // mb_rows, cols are in units of 16 pixels. We assume 3 planes all at full
599 // resolution. We assume up to 1 token per pixel, and then allow
601 return mb_rows * mb_cols * (16 * 16 * 3 + 4);
604 // Get the allocated token size for a tile. It does the same calculation as in
605 // the frame token allocation.
606 static INLINE int allocated_tokens(TileInfo tile) {
607 int tile_mb_rows = (tile.mi_row_end - tile.mi_row_start + 1) >> 1;
608 int tile_mb_cols = (tile.mi_col_end - tile.mi_col_start + 1) >> 1;
610 return get_token_alloc(tile_mb_rows, tile_mb_cols);
613 int64_t vp9_get_y_sse(const YV12_BUFFER_CONFIG *a, const YV12_BUFFER_CONFIG *b);
614 #if CONFIG_VP9_HIGHBITDEPTH
615 int64_t vp9_highbd_get_y_sse(const YV12_BUFFER_CONFIG *a,
616 const YV12_BUFFER_CONFIG *b);
617 #endif // CONFIG_VP9_HIGHBITDEPTH
619 void vp9_scale_references(VP9_COMP *cpi);
621 void vp9_update_reference_frames(VP9_COMP *cpi);
623 void vp9_set_high_precision_mv(VP9_COMP *cpi, int allow_high_precision_mv);
625 YV12_BUFFER_CONFIG *vp9_scale_if_required(VP9_COMMON *cm,
626 YV12_BUFFER_CONFIG *unscaled,
627 YV12_BUFFER_CONFIG *scaled,
628 int use_normative_scaler);
630 void vp9_apply_encoding_flags(VP9_COMP *cpi, vpx_enc_frame_flags_t flags);
632 static INLINE int is_two_pass_svc(const struct VP9_COMP *const cpi) {
633 return cpi->use_svc && cpi->oxcf.pass != 0;
636 static INLINE int is_one_pass_cbr_svc(const struct VP9_COMP *const cpi) {
637 return (cpi->use_svc && cpi->oxcf.pass == 0);
640 static INLINE int is_altref_enabled(const VP9_COMP *const cpi) {
641 return cpi->oxcf.mode != REALTIME && cpi->oxcf.lag_in_frames > 0 &&
642 (cpi->oxcf.enable_auto_arf &&
643 (!is_two_pass_svc(cpi) ||
644 cpi->oxcf.ss_enable_auto_arf[cpi->svc.spatial_layer_id]));
647 static INLINE void set_ref_ptrs(VP9_COMMON *cm, MACROBLOCKD *xd,
648 MV_REFERENCE_FRAME ref0,
649 MV_REFERENCE_FRAME ref1) {
650 xd->block_refs[0] = &cm->frame_refs[ref0 >= LAST_FRAME ? ref0 - LAST_FRAME
652 xd->block_refs[1] = &cm->frame_refs[ref1 >= LAST_FRAME ? ref1 - LAST_FRAME
656 static INLINE int get_chessboard_index(const int frame_index) {
657 return frame_index & 0x1;
660 static INLINE int *cond_cost_list(const struct VP9_COMP *cpi, int *cost_list) {
661 return cpi->sf.mv.subpel_search_method != SUBPEL_TREE ? cost_list : NULL;
664 void vp9_new_framerate(VP9_COMP *cpi, double framerate);
666 #define LAYER_IDS_TO_IDX(sl, tl, num_tl) ((sl) * (num_tl) + (tl))
672 #endif // VP9_ENCODER_VP9_ENCODER_H_