* be found in the AUTHORS file in the root of the source tree.
*/
-
#ifndef VP9_ENCODER_VP9_ONYX_INT_H_
#define VP9_ENCODER_VP9_ONYX_INT_H_
#include <stdio.h>
+
#include "./vpx_config.h"
+#include "vpx_ports/mem.h"
+#include "vpx/internal/vpx_codec_internal.h"
+
+#include "vp9/common/vp9_entropy.h"
+#include "vp9/common/vp9_entropymode.h"
#include "vp9/common/vp9_onyx.h"
-#include "vp9/encoder/vp9_treewriter.h"
-#include "vp9/encoder/vp9_tokenize.h"
#include "vp9/common/vp9_onyxc_int.h"
-#include "vp9/encoder/vp9_variance.h"
+
#include "vp9/encoder/vp9_encodemb.h"
-#include "vp9/encoder/vp9_quantize.h"
-#include "vp9/common/vp9_entropy.h"
-#include "vp9/common/vp9_entropymode.h"
-#include "vpx_ports/mem.h"
-#include "vpx/internal/vpx_codec_internal.h"
-#include "vp9/encoder/vp9_mcomp.h"
+#include "vp9/encoder/vp9_firstpass.h"
#include "vp9/encoder/vp9_lookahead.h"
+#include "vp9/encoder/vp9_mbgraph.h"
+#include "vp9/encoder/vp9_mcomp.h"
+#include "vp9/encoder/vp9_quantize.h"
+#include "vp9/encoder/vp9_ratectrl.h"
+#include "vp9/encoder/vp9_tokenize.h"
+#include "vp9/encoder/vp9_treewriter.h"
+#include "vp9/encoder/vp9_variance.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
-#define DISABLE_RC_LONG_TERM_MEM 0
// #define MODE_TEST_HIT_STATS
-// #define SPEEDSTATS 1
#if CONFIG_MULTIPLE_ARF
// Set MIN_GF_INTERVAL to 1 for the full decomposition.
#define MIN_GF_INTERVAL 2
#else
#define MIN_GF_INTERVAL 4
#endif
-#define DEFAULT_GF_INTERVAL 7
+#define DEFAULT_GF_INTERVAL 10
+#define DEFAULT_KF_BOOST 2000
+#define DEFAULT_GF_BOOST 2000
#define KEY_FRAME_CONTEXT 5
// 0 = ZERO_MV, MV
signed char last_mode_lf_deltas[MAX_MODE_LF_DELTAS];
- int inter_mode_counts[INTER_MODE_CONTEXTS][INTER_MODES - 1][2];
FRAME_CONTEXT fc;
} CODING_CONTEXT;
-typedef struct {
- double frame;
- double intra_error;
- double coded_error;
- double sr_coded_error;
- double ssim_weighted_pred_err;
- double pcnt_inter;
- double pcnt_motion;
- double pcnt_second_ref;
- double pcnt_neutral;
- double MVr;
- double mvr_abs;
- double MVc;
- double mvc_abs;
- double MVrv;
- double MVcv;
- double mv_in_out_count;
- double new_mv_count;
- double duration;
- double count;
-} FIRSTPASS_STATS;
-
-typedef struct {
- int frames_so_far;
- double frame_intra_error;
- double frame_coded_error;
- double frame_pcnt_inter;
- double frame_pcnt_motion;
- double frame_mvr;
- double frame_mvr_abs;
- double frame_mvc;
- double frame_mvc_abs;
-} ONEPASS_FRAMESTATS;
-
-typedef struct {
- struct {
- int err;
- union {
- int_mv mv;
- MB_PREDICTION_MODE mode;
- } m;
- } ref[MAX_REF_FRAMES];
-} MBGRAPH_MB_STATS;
-
-typedef struct {
- MBGRAPH_MB_STATS *mb_stats;
-} MBGRAPH_FRAME_STATS;
-
// 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.
typedef enum {
} TX_SIZE_SEARCH_METHOD;
typedef enum {
+ NOT_IN_USE = 0,
+ RELAXED_NEIGHBORING_MIN_MAX = 1,
+ STRICT_NEIGHBORING_MIN_MAX = 2
+} AUTO_MIN_MAX_MODE;
+
+typedef enum {
// Values should be powers of 2 so that they can be selected as bits of
// an integer flags field
} MODE_SEARCH_SKIP_LOGIC;
typedef enum {
- SUBPEL_ITERATIVE = 0,
- SUBPEL_TREE = 1,
+ SUBPEL_TREE = 0,
// Other methods to come
} SUBPEL_SEARCH_METHODS;
LAST_FRAME_PARTITION_ALL = 2
} LAST_FRAME_PARTITION_METHOD;
+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;
+
+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;
+
typedef struct {
- int RD;
+ // Frame level coding parameter update
+ int frame_parameter_update;
+
+ // Motion search method (Diamond, NSTEP, Hex, Big Diamond, Square, etc).
SEARCH_METHODS search_method;
- int auto_filter;
- int recode_loop;
+
+ RECODE_LOOP_TYPE recode_loop;
+
+ // 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.
SUBPEL_SEARCH_METHODS subpel_search_method;
+
+ // Maximum number of steps in logarithmic subpel search before giving up.
int subpel_iters_per_step;
+
+ // Control when to stop subpel search
+ int subpel_force_stop;
+
+ // 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 thresh_mult[MAX_MODES];
int thresh_mult_sub8x8[MAX_REFS];
+
+ // This parameter controls the number of steps we'll do in a diamond
+ // search.
int max_step_search_steps;
+
+ // This parameter controls which step in the n-step process we start at.
+ // It's changed adaptively based on circumstances.
int reduce_first_step_size;
+
+ // If this is set to 1, we limit the motion search range to 2 times the
+ // largest motion vector found in the last frame.
int auto_mv_step_size;
+
+ // Trellis (dynamic programming) optimization of quantized values (+1, 0).
int optimize_coefficients;
+
+ // 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.
int static_segmentation;
+
+ // 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.
int comp_inter_joint_search_thresh;
+
+ // 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.
int adaptive_rd_thresh;
+
+ // 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.
int skip_encode_sb;
int skip_encode_frame;
+
+ // 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.
LAST_FRAME_PARTITION_METHOD use_lastframe_partitioning;
+
+ // 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.
TX_SIZE_SEARCH_METHOD tx_size_search_method;
+
+ // Low precision 32x32 fdct keeps everything in 16 bits and thus is less
+ // precise but significantly faster than the non lp version.
int use_lp32x32fdct;
- int use_avoid_tested_higherror;
+
+ // TODO(JBB): remove this as its no longer used.
+
+ // If set partition size will always be always_this_block_size.
int use_one_partition_size_always;
+
+ // 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;
+
+ // 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.
int mode_skip_start;
+
+ // TODO(JBB): Remove this.
int reference_masking;
+
+ // Used in conjunction with use_one_partition_size_always.
BLOCK_SIZE always_this_block_size;
- int auto_min_max_partition_size;
+
+ // Sets min and max partition sizes for this 64x64 region based on the
+ // same 64x64 in last encoded frame, and the left and above neighbor.
+ AUTO_MIN_MAX_MODE auto_min_max_partition_size;
+
+ // Min and max partition size we enable (block_size) as per auto
+ // min max, but also used by adjust partitioning, and pick_partitioning.
BLOCK_SIZE min_partition_size;
BLOCK_SIZE max_partition_size;
+
+ // 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.
int adjust_partitioning_from_last_frame;
+
+ // How frequently we re do the partitioning from scratch. Only used if
+ // use_lastframe_partitioning is set.
int last_partitioning_redo_frequency;
+
+ // 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.
int disable_split_mask;
- int using_small_partition_info;
+
// TODO(jingning): combine the related motion search speed features
+ // 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.
int adaptive_motion_search;
- int adaptive_pred_filter_type;
+
+ // 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.
+ int adaptive_pred_interp_filter;
// 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;
+
// A source variance threshold below which the split mode is disabled
unsigned int disable_split_var_thresh;
+
// 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;
+
+ // These bit masks allow you to enable or disable intra modes for each
+ // transform size separately.
int intra_y_mode_mask[TX_SIZES];
int intra_uv_mode_mask[TX_SIZES];
+
+ // 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.
int use_rd_breakout;
+
+ // 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.
int use_uv_intra_rd_estimate;
+
+ // This feature controls how the loop filter level is determined:
+ // 0: Try the full image with different values.
+ // 1: Try a small portion of the image with different values.
+ // 2: Estimate the level based on quantizer and frame type
int use_fast_lpf_pick;
+
+ // This feature limits the number of coefficients updates we actually do
+ // by only looking at counts from 1/2 the bands.
int use_fast_coef_updates; // 0: 2-loop, 1: 1-loop, 2: 1-loop reduced
+
+ // This flag controls the use of non-RD mode decision.
+ int use_pick_mode;
+
+ // This variable sets the encode_breakout threshold. Currently, it is only
+ // enabled in real time mode.
+ int encode_breakout_thresh;
} SPEED_FEATURES;
typedef struct {
- // Rate targetting variables
- int this_frame_target;
- int projected_frame_size;
- int sb64_target_rate;
- int last_q[3]; // Separate values for Intra/Inter/ARF-GF
- int last_boosted_qindex; // Last boosted GF/KF/ARF q
-
- int gfu_boost;
- int last_boost;
- int kf_boost;
-
- double rate_correction_factor;
- double key_frame_rate_correction_factor;
- double gf_rate_correction_factor;
-
- unsigned int frames_since_golden;
- unsigned int frames_till_gf_update_due; // Count down till next GF
- unsigned int max_gf_interval;
- unsigned int baseline_gf_interval;
- unsigned int frames_to_key;
- unsigned int frames_since_key;
- unsigned int this_key_frame_forced;
- unsigned int next_key_frame_forced;
- unsigned int source_alt_ref_pending;
- unsigned int source_alt_ref_active;
- unsigned int is_src_frame_alt_ref;
-
- int per_frame_bandwidth; // Current section per frame bandwidth target
- int av_per_frame_bandwidth; // Average frame size target for clip
- int min_frame_bandwidth; // Minimum allocation used for any frame
-
- int ni_av_qi;
- int ni_tot_qi;
- int ni_frames;
- int avg_frame_qindex[3]; // 0 - KEY, 1 - INTER, 2 - ARF/GF
- double tot_q;
- double avg_q;
-
- int buffer_level;
- int bits_off_target;
-
- int decimation_factor;
- int decimation_count;
-
- int rolling_target_bits;
- int rolling_actual_bits;
-
- int long_rolling_target_bits;
- int long_rolling_actual_bits;
-
- int64_t total_actual_bits;
- int total_target_vs_actual; // debug stats
-
- int worst_quality;
- int active_worst_quality;
- int best_quality;
- // int active_best_quality;
-} RATE_CONTROL;
+ RATE_CONTROL rc;
+ int target_bandwidth;
+ int64_t starting_buffer_level;
+ int64_t optimal_buffer_level;
+ int64_t maximum_buffer_size;
+ double framerate;
+ int avg_frame_size;
+} LAYER_CONTEXT;
typedef struct VP9_COMP {
DECLARE_ALIGNED(16, int16_t, y_quant[QINDEX_RANGE][8]);
MACROBLOCK mb;
VP9_COMMON common;
VP9_CONFIG oxcf;
- struct rdcost_block_args rdcost_stack;
struct lookahead_ctx *lookahead;
struct lookahead_entry *source;
#if CONFIG_MULTIPLE_ARF
int gld_fb_idx;
int alt_fb_idx;
- int current_layer;
- int use_svc;
-
#if CONFIG_MULTIPLE_ARF
int alt_ref_fb_idx[REF_FRAMES - 3];
#endif
int refresh_last_frame;
int refresh_golden_frame;
int refresh_alt_ref_frame;
+
+ 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;
+
YV12_BUFFER_CONFIG last_frame_uf;
TOKENEXTRA *tok;
vp9_coeff_probs_model frame_coef_probs[TX_SIZES][PLANE_TYPES];
vp9_coeff_stats frame_branch_ct[TX_SIZES][PLANE_TYPES];
- int64_t target_bandwidth;
struct vpx_codec_pkt_list *output_pkt_list;
-#if 0
- // Experimental code for lagged and one pass
- ONEPASS_FRAMESTATS one_pass_frame_stats[MAX_LAG_BUFFERS];
- int one_pass_frame_index;
-#endif
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
// for real time encoding
int speed;
- int compressor_speed;
int cpu_used;
int pass;
- vp9_prob last_skip_false_probs[3][MBSKIP_CONTEXTS];
+ vp9_prob last_skip_false_probs[3][SKIP_CONTEXTS];
int last_skip_probs_q[3];
int ref_frame_flags;
SPEED_FEATURES sf;
- int error_bins[1024];
unsigned int max_mv_magnitude;
int mv_step_param;
+ // 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;
+
unsigned char *segmentation_map;
// segment threashold for encode breakout
uint64_t time_pick_lpf;
uint64_t time_encode_sb_row;
- struct twopass_rc {
- unsigned int section_intra_rating;
- unsigned int next_iiratio;
- unsigned int this_iiratio;
- FIRSTPASS_STATS total_stats;
- FIRSTPASS_STATS this_frame_stats;
- FIRSTPASS_STATS *stats_in, *stats_in_end, *stats_in_start;
- FIRSTPASS_STATS total_left_stats;
- int first_pass_done;
- int64_t bits_left;
- int64_t clip_bits_total;
- double avg_iiratio;
- double modified_error_total;
- double modified_error_left;
- double kf_intra_err_min;
- double gf_intra_err_min;
- int maxq_max_limit;
- int maxq_min_limit;
- int static_scene_max_gf_interval;
- int kf_bits;
- // Remaining error from uncoded frames in a gf group. Two pass use only
- int64_t gf_group_error_left;
-
- // Projected total bits available for a key frame group of frames
- int64_t kf_group_bits;
-
- // Error score of frames still to be coded in kf group
- int64_t kf_group_error_left;
-
- // Projected Bits available for a group of frames including 1 GF or ARF
- int64_t gf_group_bits;
- // Bits for the golden frame or ARF - 2 pass only
- int gf_bits;
- int alt_extra_bits;
-
- int sr_update_lag;
-
- int kf_zeromotion_pct;
- int gf_zeromotion_pct;
- } twopass;
+ struct twopass_rc twopass;
YV12_BUFFER_CONFIG alt_ref_buffer;
YV12_BUFFER_CONFIG *frames[MAX_LAG_BUFFERS];
double total_u;
double total_v;
double total;
- double total_sq_error;
+ uint64_t total_sq_error;
+ uint64_t total_samples;
+
double totalp_y;
double totalp_u;
double totalp_v;
double totalp;
- double total_sq_error2;
+ uint64_t totalp_sq_error;
+ uint64_t totalp_samples;
+
int bytes;
double summed_quality;
double summed_weights;
int *mb_norm_activity_map;
int output_partition;
- /* force next frame to intra when kf_auto says so */
+ // Force next frame to intra when kf_auto says so.
int force_next_frame_intra;
int droppable;
int dummy_packing; /* flag to indicate if packing is dummy */
- unsigned int switchable_interp_count[SWITCHABLE_FILTER_CONTEXTS]
- [SWITCHABLE_FILTERS];
-
unsigned int tx_stepdown_count[TX_SIZES];
int initial_width;
int initial_height;
- int number_spatial_layers;
- int enable_encode_breakout; // Default value is 1. From first pass stats,
- // encode_breakout may be disabled.
+ int use_svc;
+
+ struct svc {
+ int spatial_layer_id;
+ int temporal_layer_id;
+ int number_spatial_layers;
+ int number_temporal_layers;
+ // Layer context used for rate control in CBR mode, only defined for
+ // temporal layers for now.
+ LAYER_CONTEXT layer_context[VPX_TS_MAX_LAYERS];
+ } svc;
#if CONFIG_MULTIPLE_ARF
// ARF tracking variables.
int64_t mode_test_hits[BLOCK_SIZES];
#endif
- /* Y,U,V,(A) */
+ // Y,U,V,(A)
ENTROPY_CONTEXT *above_context[MAX_MB_PLANE];
ENTROPY_CONTEXT left_context[MAX_MB_PLANE][16];
PARTITION_CONTEXT left_seg_context[8];
} VP9_COMP;
-static int get_ref_frame_idx(VP9_COMP *cpi, MV_REFERENCE_FRAME ref_frame) {
+static int get_ref_frame_idx(const VP9_COMP *cpi,
+ MV_REFERENCE_FRAME ref_frame) {
if (ref_frame == LAST_FRAME) {
return cpi->lst_fb_idx;
} else if (ref_frame == GOLDEN_FRAME) {
}
}
-static int get_scale_ref_frame_idx(VP9_COMP *cpi,
- MV_REFERENCE_FRAME ref_frame) {
- if (ref_frame == LAST_FRAME) {
- return 0;
- } else if (ref_frame == GOLDEN_FRAME) {
- return 1;
- } else {
- return 2;
- }
+static 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;
}
void vp9_encode_frame(VP9_COMP *cpi);
void vp9_pack_bitstream(VP9_COMP *cpi, uint8_t *dest, size_t *size);
-void vp9_activity_masking(VP9_COMP *cpi, MACROBLOCK *x);
-
void vp9_set_speed_features(VP9_COMP *cpi);
-int vp9_calc_ss_err(YV12_BUFFER_CONFIG *source, YV12_BUFFER_CONFIG *dest);
+int vp9_calc_ss_err(const YV12_BUFFER_CONFIG *source,
+ const YV12_BUFFER_CONFIG *reference);
void vp9_alloc_compressor_data(VP9_COMP *cpi);
return mb_rows * mb_cols * (48 * 16 + 4);
}
+static void set_ref_ptrs(VP9_COMMON *cm, MACROBLOCKD *xd,
+ MV_REFERENCE_FRAME ref0, MV_REFERENCE_FRAME ref1) {
+ 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];
+}
+
+#ifdef __cplusplus
+} // extern "C"
+#endif
+
#endif // VP9_ENCODER_VP9_ONYX_INT_H_