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_quantize.h"
39 #include "vp9/encoder/vp9_ratectrl.h"
40 #include "vp9/encoder/vp9_rd.h"
41 #include "vp9/encoder/vp9_speed_features.h"
42 #include "vp9/encoder/vp9_svc_layercontext.h"
43 #include "vp9/encoder/vp9_tokenize.h"
45 #if CONFIG_VP9_TEMPORAL_DENOISING
46 #include "vp9/encoder/vp9_denoiser.h"
54 int nmvjointcost[MV_JOINTS];
55 int nmvcosts[2][MV_VALS];
56 int nmvcosts_hp[2][MV_VALS];
58 vpx_prob segment_pred_probs[PREDICTION_PROBS];
60 unsigned char *last_frame_seg_map_copy;
62 // 0 = Intra, Last, GF, ARF
63 signed char last_ref_lf_deltas[MAX_REF_LF_DELTAS];
65 signed char last_mode_lf_deltas[MAX_MODE_LF_DELTAS];
72 // encode_breakout is disabled.
73 ENCODE_BREAKOUT_DISABLED = 0,
74 // encode_breakout is enabled.
75 ENCODE_BREAKOUT_ENABLED = 1,
76 // encode_breakout is enabled with small max_thresh limit.
77 ENCODE_BREAKOUT_LIMITED = 2
78 } ENCODE_BREAKOUT_TYPE;
88 // Good Quality Fast Encoding. The encoder balances quality with the amount of
89 // time it takes to encode the output. Speed setting controls how fast.
92 // The encoder places priority on the quality of the output over encoding
93 // speed. The output is compressed at the highest possible quality. This
94 // option takes the longest amount of time to encode. Speed setting ignored.
97 // Realtime/Live Encoding. This mode is optimized for realtime encoding (for
98 // example, capturing a television signal or feed from a live camera). Speed
99 // setting controls how fast.
104 FRAMEFLAGS_KEY = 1 << 0,
105 FRAMEFLAGS_GOLDEN = 1 << 1,
106 FRAMEFLAGS_ALTREF = 1 << 2,
113 CYCLIC_REFRESH_AQ = 3,
114 AQ_MODE_COUNT // This should always be the last member of the enum
118 RESIZE_NONE = 0, // No frame resizing allowed (except for SVC).
119 RESIZE_FIXED = 1, // All frames are coded at the specified dimension.
120 RESIZE_DYNAMIC = 2 // Coded size of each frame is determined by the codec.
123 typedef struct VP9EncoderConfig {
124 BITSTREAM_PROFILE profile;
125 vpx_bit_depth_t bit_depth; // Codec bit-depth.
126 int width; // width of data passed to the compressor
127 int height; // height of data passed to the compressor
128 unsigned int input_bit_depth; // Input bit depth.
129 double init_framerate; // set to passed in framerate
130 int64_t target_bandwidth; // bandwidth to be used in kilobits per second
132 int noise_sensitivity; // pre processing blur: recommendation 0
133 int sharpness; // sharpening output: recommendation 0:
135 // maximum allowed bitrate for any intra frame in % of bitrate target.
136 unsigned int rc_max_intra_bitrate_pct;
137 // maximum allowed bitrate for any inter frame in % of bitrate target.
138 unsigned int rc_max_inter_bitrate_pct;
139 // percent of rate boost for golden frame in CBR mode.
140 unsigned int gf_cbr_boost_pct;
145 // Key Framing Operations
146 int auto_key; // autodetect cut scenes and set the keyframes
147 int key_freq; // maximum distance to key frame.
149 int lag_in_frames; // how many frames lag before we start encoding
151 // ----------------------------------------------------------------
152 // DATARATE CONTROL OPTIONS
154 // vbr, cbr, constrained quality or constant quality
155 enum vpx_rc_mode rc_mode;
157 // buffer targeting aggressiveness
161 // buffering parameters
162 int64_t starting_buffer_level_ms;
163 int64_t optimal_buffer_level_ms;
164 int64_t maximum_buffer_size_ms;
166 // Frame drop threshold.
167 int drop_frames_water_mark;
169 // controlling quality
174 AQ_MODE aq_mode; // Adaptive Quantization mode
176 // Internal frame size scaling.
177 RESIZE_TYPE resize_mode;
178 int scaled_frame_width;
179 int scaled_frame_height;
181 // Enable feature to reduce the frame quantization every x frames.
182 int frame_periodic_boost;
184 // two pass datarate control
185 int two_pass_vbrbias; // two pass datarate control tweaks
186 int two_pass_vbrmin_section;
187 int two_pass_vbrmax_section;
188 // END DATARATE CONTROL OPTIONS
189 // ----------------------------------------------------------------
191 // Spatial and temporal scalability.
192 int ss_number_layers; // Number of spatial layers.
193 int ts_number_layers; // Number of temporal layers.
194 // Bitrate allocation for spatial layers.
195 int layer_target_bitrate[VPX_MAX_LAYERS];
196 int ss_target_bitrate[VPX_SS_MAX_LAYERS];
197 int ss_enable_auto_arf[VPX_SS_MAX_LAYERS];
198 // Bitrate allocation (CBR mode) and framerate factor, for temporal layers.
199 int ts_rate_decimator[VPX_TS_MAX_LAYERS];
203 int encode_breakout; // early breakout : for video conf recommend 800
205 /* Bitfield defining the error resiliency features to enable.
206 * Can provide decodable frames after losses in previous
207 * frames and decodable partitions after losses in the same frame.
209 unsigned int error_resilient_mode;
211 /* Bitfield defining the parallel decoding mode where the
212 * decoding in successive frames may be conducted in parallel
213 * just by decoding the frame headers.
215 unsigned int frame_parallel_decoding_mode;
228 vpx_fixed_buf_t two_pass_stats_in;
229 struct vpx_codec_pkt_list *output_pkt_list;
231 #if CONFIG_FP_MB_STATS
232 vpx_fixed_buf_t firstpass_mb_stats_in;
236 vp9e_tune_content content;
237 #if CONFIG_VP9_HIGHBITDEPTH
238 int use_highbitdepth;
240 vpx_color_space_t color_space;
242 VP9E_TEMPORAL_LAYERING_MODE temporal_layering_mode;
245 static INLINE int is_lossless_requested(const VP9EncoderConfig *cfg) {
246 return cfg->best_allowed_q == 0 && cfg->worst_allowed_q == 0;
249 // TODO(jingning) All spatially adaptive variables should go to TileDataEnc.
250 typedef struct TileDataEnc {
252 int thresh_freq_fact[BLOCK_SIZES][MAX_MODES];
253 int mode_map[BLOCK_SIZES][MAX_MODES];
256 typedef struct RD_COUNTS {
257 vp9_coeff_count coef_counts[TX_SIZES][PLANE_TYPES];
258 int64_t comp_pred_diff[REFERENCE_MODES];
259 int64_t filter_diff[SWITCHABLE_FILTER_CONTEXTS];
262 typedef struct ThreadData {
265 FRAME_COUNTS *counts;
267 PICK_MODE_CONTEXT *leaf_tree;
272 struct EncWorkerData;
274 typedef struct ActiveMap {
287 typedef struct IMAGE_STAT {
292 typedef struct VP9_COMP {
295 MB_MODE_INFO_EXT *mbmi_ext_base;
296 DECLARE_ALIGNED(16, int16_t, y_dequant[QINDEX_RANGE][8]);
297 DECLARE_ALIGNED(16, int16_t, uv_dequant[QINDEX_RANGE][8]);
299 VP9EncoderConfig oxcf;
300 struct lookahead_ctx *lookahead;
301 struct lookahead_entry *alt_ref_source;
303 YV12_BUFFER_CONFIG *Source;
304 YV12_BUFFER_CONFIG *Last_Source; // NULL for first frame and alt_ref frames
305 YV12_BUFFER_CONFIG *un_scaled_source;
306 YV12_BUFFER_CONFIG scaled_source;
307 YV12_BUFFER_CONFIG *unscaled_last_source;
308 YV12_BUFFER_CONFIG scaled_last_source;
310 TileDataEnc *tile_data;
311 int allocated_tiles; // Keep track of memory allocated for tiles.
313 // For a still frame, this flag is set to 1 to skip partition search.
314 int partition_search_skippable_frame;
316 int scaled_ref_idx[MAX_REF_FRAMES];
321 int refresh_last_frame;
322 int refresh_golden_frame;
323 int refresh_alt_ref_frame;
325 int ext_refresh_frame_flags_pending;
326 int ext_refresh_last_frame;
327 int ext_refresh_golden_frame;
328 int ext_refresh_alt_ref_frame;
330 int ext_refresh_frame_context_pending;
331 int ext_refresh_frame_context;
333 YV12_BUFFER_CONFIG last_frame_uf;
335 TOKENEXTRA *tile_tok[4][1 << 6];
336 unsigned int tok_count[4][1 << 6];
338 // Ambient reconstruction err target for force key frames
343 CODING_CONTEXT coding_context;
348 int *nmvsadcosts_hp[2];
350 int64_t last_time_stamp_seen;
351 int64_t last_end_time_stamp_seen;
352 int64_t first_time_stamp_ever;
357 int interp_filter_selected[MAX_REF_FRAMES][SWITCHABLE];
359 struct vpx_codec_pkt_list *output_pkt_list;
361 MBGRAPH_FRAME_STATS mbgraph_stats[MAX_LAG_BUFFERS];
362 int mbgraph_n_frames; // number of frames filled in the above
363 int static_mb_pct; // % forced skip mbs by segmentation
368 unsigned int max_mv_magnitude;
371 int allow_comp_inter_inter;
373 // Default value is 1. From first pass stats, encode_breakout may be disabled.
374 ENCODE_BREAKOUT_TYPE allow_encode_breakout;
376 // Get threshold from external input. A suggested threshold is 800 for HD
377 // clips, and 300 for < HD clips.
380 unsigned char *segmentation_map;
382 // segment threashold for encode breakout
383 int segment_encode_breakout[MAX_SEGMENTS];
385 CYCLIC_REFRESH *cyclic_refresh;
386 ActiveMap active_map;
388 fractional_mv_step_fp *find_fractional_mv_step;
389 vp9_full_search_fn_t full_search_sad;
390 vp9_diamond_search_fn_t diamond_search_sad;
391 vp9_variance_fn_ptr_t fn_ptr[BLOCK_SIZES];
392 uint64_t time_receive_data;
393 uint64_t time_compress_data;
394 uint64_t time_pick_lpf;
395 uint64_t time_encode_sb_row;
397 #if CONFIG_FP_MB_STATS
403 YV12_BUFFER_CONFIG alt_ref_buffer;
406 #if CONFIG_INTERNAL_STATS
407 unsigned int mode_chosen_counts[MAX_MODES];
410 uint64_t total_sq_error;
411 uint64_t total_samples;
414 uint64_t totalp_sq_error;
415 uint64_t totalp_samples;
418 double total_blockiness;
419 double worst_blockiness;
422 double summed_quality;
423 double summed_weights;
424 double summedp_quality;
425 double summedp_weights;
426 unsigned int tot_recode_hits;
433 int b_calculate_ssimg;
434 int b_calculate_blockiness;
436 int b_calculate_consistency;
438 double total_inconsistency;
439 double worst_consistency;
443 int b_calculate_psnr;
449 int initial_mbs; // Number of MBs in the full-size frame; to be used to
450 // normalize the firstpass stats. This will differ from the
451 // number of MBs in the current frame when the frame is
458 // Store frame variance info in SOURCE_VAR_BASED_PARTITION search type.
459 diff *source_diff_var;
460 // The threshold used in SOURCE_VAR_BASED_PARTITION search type.
461 unsigned int source_var_thresh;
462 int frames_till_next_var_check;
466 search_site_config ss_cfg;
468 int mbmode_cost[INTRA_MODES];
469 unsigned int inter_mode_cost[INTER_MODE_CONTEXTS][INTER_MODES];
470 int intra_uv_mode_cost[FRAME_TYPES][INTRA_MODES];
471 int y_mode_costs[INTRA_MODES][INTRA_MODES][INTRA_MODES];
472 int switchable_interp_costs[SWITCHABLE_FILTER_CONTEXTS][SWITCHABLE_FILTERS];
473 int partition_cost[PARTITION_CONTEXTS][PARTITION_TYPES];
475 int multi_arf_allowed;
476 int multi_arf_enabled;
477 int multi_arf_last_grp_enabled;
479 #if CONFIG_VP9_TEMPORAL_DENOISING
480 VP9_DENOISER denoiser;
485 int resize_scale_num;
486 int resize_scale_den;
488 int resize_buffer_underflow;
491 // VAR_BASED_PARTITION thresholds
492 // 0 - threshold_64x64; 1 - threshold_32x32;
493 // 2 - threshold_16x16; 3 - vbp_threshold_8x8;
494 int64_t vbp_thresholds[4];
495 int64_t vbp_threshold_minmax;
496 int64_t vbp_threshold_sad;
497 BLOCK_SIZE vbp_bsize_min;
502 struct EncWorkerData *tile_thr_data;
503 VP9LfSync lf_row_sync;
506 void vp9_initialize_enc(void);
508 struct VP9_COMP *vp9_create_compressor(VP9EncoderConfig *oxcf,
509 BufferPool *const pool);
510 void vp9_remove_compressor(VP9_COMP *cpi);
512 void vp9_change_config(VP9_COMP *cpi, const VP9EncoderConfig *oxcf);
514 // receive a frames worth of data. caller can assume that a copy of this
515 // frame is made and not just a copy of the pointer..
516 int vp9_receive_raw_frame(VP9_COMP *cpi, unsigned int frame_flags,
517 YV12_BUFFER_CONFIG *sd, int64_t time_stamp,
518 int64_t end_time_stamp);
520 int vp9_get_compressed_data(VP9_COMP *cpi, unsigned int *frame_flags,
521 size_t *size, uint8_t *dest,
522 int64_t *time_stamp, int64_t *time_end, int flush);
524 int vp9_get_preview_raw_frame(VP9_COMP *cpi, YV12_BUFFER_CONFIG *dest,
525 vp9_ppflags_t *flags);
527 int vp9_use_as_reference(VP9_COMP *cpi, int ref_frame_flags);
529 void vp9_update_reference(VP9_COMP *cpi, int ref_frame_flags);
531 int vp9_copy_reference_enc(VP9_COMP *cpi, VP9_REFFRAME ref_frame_flag,
532 YV12_BUFFER_CONFIG *sd);
534 int vp9_set_reference_enc(VP9_COMP *cpi, VP9_REFFRAME ref_frame_flag,
535 YV12_BUFFER_CONFIG *sd);
537 int vp9_update_entropy(VP9_COMP *cpi, int update);
539 int vp9_set_active_map(VP9_COMP *cpi, unsigned char *map, int rows, int cols);
541 int vp9_get_active_map(VP9_COMP *cpi, unsigned char *map, int rows, int cols);
543 int vp9_set_internal_size(VP9_COMP *cpi,
544 VPX_SCALING horiz_mode, VPX_SCALING vert_mode);
546 int vp9_set_size_literal(VP9_COMP *cpi, unsigned int width,
547 unsigned int height);
549 void vp9_set_svc(VP9_COMP *cpi, int use_svc);
551 int vp9_get_quantizer(struct VP9_COMP *cpi);
553 static INLINE int frame_is_kf_gf_arf(const VP9_COMP *cpi) {
554 return frame_is_intra_only(&cpi->common) ||
555 cpi->refresh_alt_ref_frame ||
556 (cpi->refresh_golden_frame && !cpi->rc.is_src_frame_alt_ref);
559 static INLINE int get_ref_frame_map_idx(const VP9_COMP *cpi,
560 MV_REFERENCE_FRAME ref_frame) {
561 if (ref_frame == LAST_FRAME) {
562 return cpi->lst_fb_idx;
563 } else if (ref_frame == GOLDEN_FRAME) {
564 return cpi->gld_fb_idx;
566 return cpi->alt_fb_idx;
570 static INLINE int get_ref_frame_buf_idx(const VP9_COMP *const cpi,
572 const VP9_COMMON *const cm = &cpi->common;
573 const int map_idx = get_ref_frame_map_idx(cpi, ref_frame);
574 return (map_idx != INVALID_IDX) ? cm->ref_frame_map[map_idx] : INVALID_IDX;
577 static INLINE YV12_BUFFER_CONFIG *get_ref_frame_buffer(
578 VP9_COMP *cpi, MV_REFERENCE_FRAME ref_frame) {
579 VP9_COMMON *const cm = &cpi->common;
580 const int buf_idx = get_ref_frame_buf_idx(cpi, ref_frame);
582 buf_idx != INVALID_IDX ? &cm->buffer_pool->frame_bufs[buf_idx].buf : NULL;
585 static INLINE int get_token_alloc(int mb_rows, int mb_cols) {
586 // TODO(JBB): double check we can't exceed this token count if we have a
587 // 32x32 transform crossing a boundary at a multiple of 16.
588 // mb_rows, cols are in units of 16 pixels. We assume 3 planes all at full
589 // resolution. We assume up to 1 token per pixel, and then allow
591 return mb_rows * mb_cols * (16 * 16 * 3 + 4);
594 // Get the allocated token size for a tile. It does the same calculation as in
595 // the frame token allocation.
596 static INLINE int allocated_tokens(TileInfo tile) {
597 int tile_mb_rows = (tile.mi_row_end - tile.mi_row_start + 1) >> 1;
598 int tile_mb_cols = (tile.mi_col_end - tile.mi_col_start + 1) >> 1;
600 return get_token_alloc(tile_mb_rows, tile_mb_cols);
603 int64_t vp9_get_y_sse(const YV12_BUFFER_CONFIG *a, const YV12_BUFFER_CONFIG *b);
604 #if CONFIG_VP9_HIGHBITDEPTH
605 int64_t vp9_highbd_get_y_sse(const YV12_BUFFER_CONFIG *a,
606 const YV12_BUFFER_CONFIG *b);
607 #endif // CONFIG_VP9_HIGHBITDEPTH
609 void vp9_scale_references(VP9_COMP *cpi);
611 void vp9_update_reference_frames(VP9_COMP *cpi);
613 void vp9_set_high_precision_mv(VP9_COMP *cpi, int allow_high_precision_mv);
615 YV12_BUFFER_CONFIG *vp9_scale_if_required(VP9_COMMON *cm,
616 YV12_BUFFER_CONFIG *unscaled,
617 YV12_BUFFER_CONFIG *scaled,
618 int use_normative_scaler);
620 void vp9_apply_encoding_flags(VP9_COMP *cpi, vpx_enc_frame_flags_t flags);
622 static INLINE int is_two_pass_svc(const struct VP9_COMP *const cpi) {
623 return cpi->use_svc && cpi->oxcf.pass != 0;
626 static INLINE int is_one_pass_cbr_svc(const struct VP9_COMP *const cpi) {
627 return (cpi->use_svc && cpi->oxcf.pass == 0);
630 static INLINE int is_altref_enabled(const VP9_COMP *const cpi) {
631 return cpi->oxcf.mode != REALTIME && cpi->oxcf.lag_in_frames > 0 &&
632 (cpi->oxcf.enable_auto_arf &&
633 (!is_two_pass_svc(cpi) ||
634 cpi->oxcf.ss_enable_auto_arf[cpi->svc.spatial_layer_id]));
637 static INLINE void set_ref_ptrs(VP9_COMMON *cm, MACROBLOCKD *xd,
638 MV_REFERENCE_FRAME ref0,
639 MV_REFERENCE_FRAME ref1) {
640 xd->block_refs[0] = &cm->frame_refs[ref0 >= LAST_FRAME ? ref0 - LAST_FRAME
642 xd->block_refs[1] = &cm->frame_refs[ref1 >= LAST_FRAME ? ref1 - LAST_FRAME
646 static INLINE int get_chessboard_index(const int frame_index) {
647 return frame_index & 0x1;
650 static INLINE int *cond_cost_list(const struct VP9_COMP *cpi, int *cost_list) {
651 return cpi->sf.mv.subpel_search_method != SUBPEL_TREE ? cost_list : NULL;
654 void vp9_new_framerate(VP9_COMP *cpi, double framerate);
656 #define LAYER_IDS_TO_IDX(sl, tl, num_tl) ((sl) * (num_tl) + (tl))
662 #endif // VP9_ENCODER_VP9_ENCODER_H_