/*
* Copyright (c) 2016, Alliance for Open Media. All rights reserved.
*
* This source code is subject to the terms of the BSD 2 Clause License and
* the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
* was not distributed with this source code in the LICENSE file, you can
* obtain it at www.aomedia.org/license/software. If the Alliance for Open
* Media Patent License 1.0 was not distributed with this source code in the
* PATENTS file, you can obtain it at www.aomedia.org/license/patent.
*/
#include <limits.h>
#include "av1/common/reconintra.h"
#include "av1/encoder/encoder.h"
#include "av1/encoder/speed_features.h"
#include "av1/encoder/rdopt.h"
#include "aom_dsp/aom_dsp_common.h"
#define MAX_MESH_SPEED 5 // Max speed setting for mesh motion method
// Max speed setting for tx domain evaluation
#define MAX_TX_DOMAIN_EVAL_SPEED 5
static const MESH_PATTERN
good_quality_mesh_patterns[MAX_MESH_SPEED + 1][MAX_MESH_STEP] = {
{ { 64, 8 }, { 28, 4 }, { 15, 1 }, { 7, 1 } },
{ { 64, 8 }, { 28, 4 }, { 15, 1 }, { 7, 1 } },
{ { 64, 8 }, { 14, 2 }, { 7, 1 }, { 7, 1 } },
{ { 64, 16 }, { 24, 8 }, { 12, 4 }, { 7, 1 } },
{ { 64, 16 }, { 24, 8 }, { 12, 4 }, { 7, 1 } },
{ { 64, 16 }, { 24, 8 }, { 12, 4 }, { 7, 1 } },
};
// TODO(huisu@google.com): These settings are pretty relaxed, tune them for
// each speed setting
static const MESH_PATTERN
intrabc_mesh_patterns[MAX_MESH_SPEED + 1][MAX_MESH_STEP] = {
{ { 256, 1 }, { 256, 1 }, { 0, 0 }, { 0, 0 } },
{ { 256, 1 }, { 256, 1 }, { 0, 0 }, { 0, 0 } },
{ { 64, 1 }, { 64, 1 }, { 0, 0 }, { 0, 0 } },
{ { 64, 1 }, { 64, 1 }, { 0, 0 }, { 0, 0 } },
{ { 64, 4 }, { 16, 1 }, { 0, 0 }, { 0, 0 } },
{ { 64, 4 }, { 16, 1 }, { 0, 0 }, { 0, 0 } },
};
// Threshold values to be used for pruning the txfm_domain_distortion
// based on block MSE
// Index 0: Default mode evaluation, Winner mode processing is not
// applicable (Eg : IntraBc). Index 1: Mode evaluation.
// Index 2: Winner mode evaluation. Index 1 and 2 are applicable when
// enable_winner_mode_for_use_tx_domain_dist speed feature is ON
// TODO(any): Experiment the threshold logic based on variance metric
static const unsigned int tx_domain_dist_thresholds[4][MODE_EVAL_TYPES] = {
{ UINT_MAX, UINT_MAX, UINT_MAX },
{ 22026, 22026, 22026 },
{ 1377, 1377, 1377 },
{ 0, 0, 0 }
};
// Number of different levels of aggressiveness in using transform domain
// distortion during the R-D evaluation based on the speed feature
// tx_domain_dist_level.
#define TX_DOMAIN_DIST_LEVELS 4
// Transform domain distortion type to be used for default, mode and winner mode
// evaluation Index 0: Default mode evaluation, Winner mode processing is not
// applicable (Eg : IntraBc). Index 1: Mode evaluation. Index 2: Winner mode
// evaluation. Index 1 and 2 are applicable when
// enable_winner_mode_for_use_tx_domain_dist speed feature is ON
static const unsigned int
tx_domain_dist_types[TX_DOMAIN_DIST_LEVELS][MODE_EVAL_TYPES] = {
{ 0, 2, 0 }, { 1, 2, 0 }, { 2, 2, 0 }, { 2, 2, 2 }
};
// Threshold values to be used for disabling coeff RD-optimization
// based on block MSE / qstep^2.
// TODO(any): Experiment the threshold logic based on variance metric.
// Table has satd and dist threshold value index 0 : dist,index 1: satd
// For each row, the indices are as follows.
// Index 0: Default mode evaluation, Winner mode processing is not applicable
// (Eg : IntraBc)
// Index 1: Mode evaluation.
// Index 2: Winner mode evaluation.
// Index 1 and 2 are applicable when enable_winner_mode_for_coeff_opt speed
// feature is ON
// There are 7 levels with increasing speed, mapping to vertical indices.
static const unsigned int coeff_opt_thresholds[9][MODE_EVAL_TYPES][2] = {
{ { UINT_MAX, UINT_MAX }, { UINT_MAX, UINT_MAX }, { UINT_MAX, UINT_MAX } },
{ { 3200, UINT_MAX }, { 250, UINT_MAX }, { UINT_MAX, UINT_MAX } },
{ { 1728, UINT_MAX }, { 142, UINT_MAX }, { UINT_MAX, UINT_MAX } },
{ { 864, UINT_MAX }, { 142, UINT_MAX }, { UINT_MAX, UINT_MAX } },
{ { 432, UINT_MAX }, { 86, UINT_MAX }, { UINT_MAX, UINT_MAX } },
{ { 864, 97 }, { 142, 16 }, { UINT_MAX, UINT_MAX } },
{ { 432, 97 }, { 86, 16 }, { UINT_MAX, UINT_MAX } },
{ { 216, 25 }, { 86, 10 }, { UINT_MAX, UINT_MAX } },
{ { 216, 25 }, { 0, 10 }, { UINT_MAX, UINT_MAX } }
};
// Transform size to be used for default, mode and winner mode evaluation
// Index 0: Default mode evaluation, Winner mode processing is not applicable
// (Eg : IntraBc) Index 1: Mode evaluation. Index 2: Winner mode evaluation.
// Index 1 and 2 are applicable when enable_winner_mode_for_tx_size_srch speed
// feature is ON
static const TX_SIZE_SEARCH_METHOD
tx_size_search_methods[4][MODE_EVAL_TYPES] = {
{ USE_FULL_RD, USE_LARGESTALL, USE_FULL_RD },
{ USE_FAST_RD, USE_LARGESTALL, USE_FULL_RD },
{ USE_LARGESTALL, USE_LARGESTALL, USE_FULL_RD },
{ USE_LARGESTALL, USE_LARGESTALL, USE_LARGESTALL }
};
// Predict transform skip levels to be used for default, mode and winner mode
// evaluation. Index 0: Default mode evaluation, Winner mode processing is not
// applicable. Index 1: Mode evaluation, Index 2: Winner mode evaluation
// Values indicate the aggressiveness of skip flag prediction.
// 0 : no early skip prediction
// 1 : conservative early skip prediction using DCT_DCT
// 2 : early skip prediction based on SSE
static const unsigned int predict_skip_levels[3][MODE_EVAL_TYPES] = {
{ 0, 0, 0 }, { 1, 1, 1 }, { 1, 2, 1 }
};
// Predict skip or DC block level used during transform type search. It is
// indexed using the following:
// First index : Speed feature 'dc_blk_pred_level' (0 to 3)
// Second index : Mode evaluation type (DEFAULT_EVAL, MODE_EVAL and
// WINNER_MODE_EVAL).
//
// The values of predict_dc_levels[][] indicate the aggressiveness of predicting
// a block as transform skip or DC only.
// Type 0 : No skip block or DC only block prediction
// Type 1 : Prediction of skip block based on residual mean and variance
// Type 2 : Prediction of skip block or DC only block based on residual mean and
// variance
static const unsigned int predict_dc_levels[4][MODE_EVAL_TYPES] = {
{ 0, 0, 0 }, { 1, 1, 0 }, { 2, 2, 0 }, { 2, 2, 2 }
};
#if !CONFIG_FPMT_TEST
// This table holds the maximum number of reference frames for global motion.
// The table is indexed as per the speed feature 'gm_search_type'.
// 0 : All reference frames are allowed.
// 1 : All reference frames except L2 and L3 are allowed.
// 2 : All reference frames except L2, L3 and ARF2 are allowed.
// 3 : No reference frame is allowed.
static const int gm_available_reference_frames[GM_DISABLE_SEARCH + 1] = {
INTER_REFS_PER_FRAME, INTER_REFS_PER_FRAME - 2, INTER_REFS_PER_FRAME - 3, 0
};
#endif
// Qindex threshold levels used for selecting full-pel motion search.
// ms_qthresh[i][j][k] indicates the qindex boundary value for 'k'th qindex band
// for resolution index 'j' for aggressiveness level 'i'.
// Aggressiveness increases from i = 0 to 2.
// j = 0: lower than 720p resolution, j = 1: 720p or larger resolution.
// Currently invoked only for speed 0, 1 and 2.
static const int ms_qindex_thresh[3][2][2] = { { { 200, 70 }, { MAXQ, 200 } },
{ { 170, 50 }, { MAXQ, 200 } },
{ { 170, 40 }, { 200, 40 } } };
// Full-pel search methods for aggressive search based on qindex.
// Index 0 is for resolutions lower than 720p, index 1 for 720p or larger
// resolutions. Currently invoked only for speed 1 and 2.
static const SEARCH_METHODS motion_search_method[2] = { CLAMPED_DIAMOND,
DIAMOND };
// Intra only frames, golden frames (except alt ref overlays) and
// alt ref frames tend to be coded at a higher than ambient quality
static int frame_is_boosted(const AV1_COMP *cpi) {
return frame_is_kf_gf_arf(cpi);
}
// Set transform rd gate level for all transform search cases.
static inline void set_txfm_rd_gate_level(
int txfm_rd_gate_level[TX_SEARCH_CASES], int level) {
assert(level <= MAX_TX_RD_GATE_LEVEL);
for (int idx = 0; idx < TX_SEARCH_CASES; idx++)
txfm_rd_gate_level[idx] = level;
}
static void set_allintra_speed_feature_framesize_dependent(
const AV1_COMP *const cpi, SPEED_FEATURES *const sf, int speed) {
const AV1_COMMON *const cm = &cpi->common;
const int is_480p_or_larger = AOMMIN(cm->width, cm->height) >= 480;
const int is_720p_or_larger = AOMMIN(cm->width, cm->height) >= 720;
const int is_1080p_or_larger = AOMMIN(cm->width, cm->height) >= 1080;
const int is_4k_or_larger = AOMMIN(cm->width, cm->height) >= 2160;
const bool use_hbd = cpi->oxcf.use_highbitdepth;
if (is_480p_or_larger) {
sf->part_sf.use_square_partition_only_threshold = BLOCK_128X128;
if (is_720p_or_larger)
sf->part_sf.auto_max_partition_based_on_simple_motion = ADAPT_PRED;
else
sf->part_sf.auto_max_partition_based_on_simple_motion = RELAXED_PRED;
} else {
sf->part_sf.use_square_partition_only_threshold = BLOCK_64X64;
sf->part_sf.auto_max_partition_based_on_simple_motion = DIRECT_PRED;
if (use_hbd) sf->tx_sf.prune_tx_size_level = 1;
}
if (is_4k_or_larger) {
sf->part_sf.default_min_partition_size = BLOCK_8X8;
}
// TODO(huisu@google.com): train models for 720P and above.
if (!is_720p_or_larger) {
sf->part_sf.ml_partition_search_breakout_thresh[0] = 200; // BLOCK_8X8
sf->part_sf.ml_partition_search_breakout_thresh[1] = 250; // BLOCK_16X16
sf->part_sf.ml_partition_search_breakout_thresh[2] = 300; // BLOCK_32X32
sf->part_sf.ml_partition_search_breakout_thresh[3] = 500; // BLOCK_64X64
sf->part_sf.ml_partition_search_breakout_thresh[4] = -1; // BLOCK_128X128
sf->part_sf.ml_early_term_after_part_split_level = 1;
}
if (is_720p_or_larger) {
// TODO(chiyotsai@google.com): make this speed feature adaptive based on
// current block's vertical texture instead of hardcoded with resolution
sf->mv_sf.use_downsampled_sad = 2;
}
if (speed >= 1) {
if (is_720p_or_larger) {
sf->part_sf.use_square_partition_only_threshold = BLOCK_128X128;
} else if (is_480p_or_larger) {
sf->part_sf.use_square_partition_only_threshold = BLOCK_64X64;
} else {
sf->part_sf.use_square_partition_only_threshold = BLOCK_32X32;
}
if (!is_720p_or_larger) {
sf->part_sf.ml_partition_search_breakout_thresh[0] = 200; // BLOCK_8X8
sf->part_sf.ml_partition_search_breakout_thresh[1] = 250; // BLOCK_16X16
sf->part_sf.ml_partition_search_breakout_thresh[2] = 300; // BLOCK_32X32
sf->part_sf.ml_partition_search_breakout_thresh[3] = 300; // BLOCK_64X64
sf->part_sf.ml_partition_search_breakout_thresh[4] = -1; // BLOCK_128X128
}
sf->part_sf.ml_early_term_after_part_split_level = 2;
}
if (speed >= 2) {
if (is_720p_or_larger) {
sf->part_sf.use_square_partition_only_threshold = BLOCK_64X64;
} else if (is_480p_or_larger) {
sf->part_sf.use_square_partition_only_threshold = BLOCK_32X32;
} else {
sf->part_sf.use_square_partition_only_threshold = BLOCK_32X32;
}
if (is_720p_or_larger) {
sf->part_sf.partition_search_breakout_dist_thr = (1 << 24);
sf->part_sf.partition_search_breakout_rate_thr = 120;
} else {
sf->part_sf.partition_search_breakout_dist_thr = (1 << 22);
sf->part_sf.partition_search_breakout_rate_thr = 100;
}
if (is_480p_or_larger) {
sf->tx_sf.tx_type_search.prune_tx_type_using_stats = 1;
if (use_hbd) sf->tx_sf.prune_tx_size_level = 2;
} else {
if (use_hbd) sf->tx_sf.prune_tx_size_level = 3;
}
}
if (speed >= 3) {
sf->part_sf.ml_early_term_after_part_split_level = 0;
if (is_720p_or_larger) {
sf->part_sf.partition_search_breakout_dist_thr = (1 << 25);
sf->part_sf.partition_search_breakout_rate_thr = 200;
} else {
sf->part_sf.max_intra_bsize = BLOCK_32X32;
sf->part_sf.partition_search_breakout_dist_thr = (1 << 23);
sf->part_sf.partition_search_breakout_rate_thr = 120;
}
if (use_hbd) sf->tx_sf.prune_tx_size_level = 3;
}
if (speed >= 4) {
if (is_720p_or_larger) {
sf->part_sf.partition_search_breakout_dist_thr = (1 << 26);
} else {
sf->part_sf.partition_search_breakout_dist_thr = (1 << 24);
}
if (is_480p_or_larger) {
sf->tx_sf.tx_type_search.prune_tx_type_using_stats = 2;
}
}
if (speed >= 6) {
if (is_720p_or_larger) {
sf->part_sf.auto_max_partition_based_on_simple_motion = NOT_IN_USE;
} else if (is_480p_or_larger) {
sf->part_sf.auto_max_partition_based_on_simple_motion = DIRECT_PRED;
}
if (is_1080p_or_larger) {
sf->part_sf.default_min_partition_size = BLOCK_8X8;
}
sf->part_sf.use_square_partition_only_threshold = BLOCK_16X16;
}
if (speed >= 7) {
// TODO(kyslov): add more speed features to control speed/quality
}
if (speed >= 8) {
if (!is_480p_or_larger) {
sf->rt_sf.nonrd_check_partition_merge_mode = 2;
}
if (is_720p_or_larger) {
sf->rt_sf.force_large_partition_blocks_intra = 1;
}
}
if (speed >= 9) {
// TODO(kyslov): add more speed features to control speed/quality
if (!is_4k_or_larger) {
// In av1_select_sb_size(), superblock size is set to 64x64 only for
// resolutions less than 4k in speed>=9, to improve the multithread
// performance. If cost update levels are set to INTERNAL_COST_UPD_OFF
// for resolutions >= 4k, the SB size setting can be modified for these
// resolutions as well.
sf->inter_sf.coeff_cost_upd_level = INTERNAL_COST_UPD_OFF;
sf->inter_sf.mode_cost_upd_level = INTERNAL_COST_UPD_OFF;
}
}
}
static void set_allintra_speed_features_framesize_independent(
const AV1_COMP *const cpi, SPEED_FEATURES *const sf, int speed) {
const AV1_COMMON *const cm = &cpi->common;
const int allow_screen_content_tools =
cm->features.allow_screen_content_tools;
const int use_hbd = cpi->oxcf.use_highbitdepth;
sf->part_sf.less_rectangular_check_level = 1;
sf->part_sf.ml_prune_partition = 1;
sf->part_sf.prune_ext_partition_types_search_level = 1;
sf->part_sf.prune_part4_search = 2;
sf->part_sf.simple_motion_search_prune_rect = 1;
sf->part_sf.ml_predict_breakout_level = use_hbd ? 1 : 3;
sf->part_sf.reuse_prev_rd_results_for_part_ab = 1;
sf->part_sf.use_best_rd_for_pruning = 1;
sf->intra_sf.intra_pruning_with_hog = 1;
sf->intra_sf.prune_luma_palette_size_search_level = 1;
sf->intra_sf.dv_cost_upd_level = INTERNAL_COST_UPD_OFF;
sf->intra_sf.early_term_chroma_palette_size_search = 1;
sf->tx_sf.adaptive_txb_search_level = 1;
sf->tx_sf.intra_tx_size_search_init_depth_sqr = 1;
sf->tx_sf.model_based_prune_tx_search_level = 1;
sf->tx_sf.tx_type_search.use_reduced_intra_txset = 1;
sf->rt_sf.use_nonrd_pick_mode = 0;
sf->rt_sf.use_real_time_ref_set = 0;
if (cpi->twopass_frame.fr_content_type == FC_GRAPHICS_ANIMATION ||
cpi->use_screen_content_tools) {
sf->mv_sf.exhaustive_searches_thresh = (1 << 20);
} else {
sf->mv_sf.exhaustive_searches_thresh = (1 << 25);
}
sf->rd_sf.perform_coeff_opt = 1;
sf->hl_sf.superres_auto_search_type = SUPERRES_AUTO_DUAL;
if (speed >= 1) {
sf->part_sf.intra_cnn_based_part_prune_level =
allow_screen_content_tools ? 0 : 2;
sf->part_sf.simple_motion_search_early_term_none = 1;
// TODO(Venkat): Clean-up frame type dependency for
// simple_motion_search_split in partition search function and set the
// speed feature accordingly
sf->part_sf.simple_motion_search_split = allow_screen_content_tools ? 1 : 2;
sf->part_sf.ml_predict_breakout_level = use_hbd ? 2 : 3;
sf->part_sf.reuse_best_prediction_for_part_ab = 1;
sf->mv_sf.exhaustive_searches_thresh <<= 1;
sf->intra_sf.prune_palette_search_level = 1;
sf->intra_sf.prune_luma_palette_size_search_level = 2;
sf->intra_sf.top_intra_model_count_allowed = 3;
sf->tx_sf.adaptive_txb_search_level = 2;
sf->tx_sf.inter_tx_size_search_init_depth_rect = 1;
sf->tx_sf.inter_tx_size_search_init_depth_sqr = 1;
sf->tx_sf.intra_tx_size_search_init_depth_rect = 1;
sf->tx_sf.model_based_prune_tx_search_level = 0;
sf->tx_sf.tx_type_search.ml_tx_split_thresh = 4000;
sf->tx_sf.tx_type_search.prune_2d_txfm_mode = TX_TYPE_PRUNE_2;
sf->tx_sf.tx_type_search.skip_tx_search = 1;
sf->rd_sf.perform_coeff_opt = 2;
sf->rd_sf.tx_domain_dist_level = 1;
sf->rd_sf.tx_domain_dist_thres_level = 1;
sf->lpf_sf.cdef_pick_method = CDEF_FAST_SEARCH_LVL1;
sf->lpf_sf.dual_sgr_penalty_level = 1;
sf->lpf_sf.enable_sgr_ep_pruning = 1;
}
if (speed >= 2) {
sf->mv_sf.auto_mv_step_size = 1;
sf->intra_sf.disable_smooth_intra = 1;
sf->intra_sf.intra_pruning_with_hog = 2;
sf->intra_sf.prune_filter_intra_level = 1;
sf->rd_sf.perform_coeff_opt = 3;
sf->lpf_sf.prune_wiener_based_on_src_var = 1;
sf->lpf_sf.prune_sgr_based_on_wiener = 1;
}
if (speed >= 3) {
sf->hl_sf.high_precision_mv_usage = CURRENT_Q;
sf->hl_sf.recode_loop = ALLOW_RECODE_KFARFGF;
sf->part_sf.less_rectangular_check_level = 2;
sf->part_sf.simple_motion_search_prune_agg = SIMPLE_AGG_LVL1;
sf->part_sf.prune_ext_part_using_split_info = 1;
sf->mv_sf.full_pixel_search_level = 1;
sf->mv_sf.search_method = DIAMOND;
// TODO(chiyotsai@google.com): the thresholds chosen for intra hog are
// inherited directly from luma hog with some minor tweaking. Eventually we
// should run this with a bayesian optimizer to find the Pareto frontier.
sf->intra_sf.chroma_intra_pruning_with_hog = 2;
sf->intra_sf.intra_pruning_with_hog = 3;
sf->intra_sf.prune_palette_search_level = 2;
sf->tx_sf.adaptive_txb_search_level = 2;
sf->tx_sf.tx_type_search.use_skip_flag_prediction = 2;
sf->tx_sf.use_rd_based_breakout_for_intra_tx_search = true ;
// TODO(any): evaluate if these lpf features can be moved to speed 2.
// For screen content, "prune_sgr_based_on_wiener = 2" cause large quality
// loss.
sf->lpf_sf.prune_sgr_based_on_wiener = allow_screen_content_tools ? 1 : 2;
sf->lpf_sf.disable_loop_restoration_chroma = 0;
sf->lpf_sf.reduce_wiener_window_size = 1;
sf->lpf_sf.prune_wiener_based_on_src_var = 2;
}
if (speed >= 4) {
sf->mv_sf.subpel_search_method = SUBPEL_TREE_PRUNED_MORE;
sf->part_sf.simple_motion_search_prune_agg = SIMPLE_AGG_LVL2;
sf->part_sf.simple_motion_search_reduce_search_steps = 4;
sf->part_sf.prune_ext_part_using_split_info = 2;
sf->part_sf.early_term_after_none_split = 1;
sf->part_sf.ml_predict_breakout_level = 3;
sf->intra_sf.prune_chroma_modes_using_luma_winner = 1;
sf->mv_sf.simple_motion_subpel_force_stop = HALF_PEL;
sf->tpl_sf.prune_starting_mv = 2;
sf->tpl_sf.subpel_force_stop = HALF_PEL;
sf->tpl_sf.search_method = FAST_BIGDIA;
sf->tx_sf.tx_type_search.winner_mode_tx_type_pruning = 2;
sf->tx_sf.tx_type_search.fast_intra_tx_type_search = 1;
sf->tx_sf.tx_type_search.prune_2d_txfm_mode = TX_TYPE_PRUNE_3;
sf->tx_sf.tx_type_search.prune_tx_type_est_rd = 1;
sf->rd_sf.perform_coeff_opt = 5;
sf->rd_sf.tx_domain_dist_thres_level = 3;
sf->lpf_sf.lpf_pick = LPF_PICK_FROM_FULL_IMAGE_NON_DUAL;
sf->lpf_sf.cdef_pick_method = CDEF_FAST_SEARCH_LVL3;
sf->mv_sf.reduce_search_range = 1;
sf->winner_mode_sf.enable_winner_mode_for_coeff_opt = 1;
sf->winner_mode_sf.enable_winner_mode_for_use_tx_domain_dist = 1;
sf->winner_mode_sf.multi_winner_mode_type = MULTI_WINNER_MODE_DEFAULT;
sf->winner_mode_sf.enable_winner_mode_for_tx_size_srch = 1;
}
if (speed >= 5) {
sf->part_sf.simple_motion_search_prune_agg = SIMPLE_AGG_LVL3;
sf->part_sf.ext_partition_eval_thresh =
allow_screen_content_tools ? BLOCK_8X8 : BLOCK_16X16;
sf->part_sf.intra_cnn_based_part_prune_level =
allow_screen_content_tools ? 1 : 2;
sf->intra_sf.chroma_intra_pruning_with_hog = 3;
sf->lpf_sf.use_coarse_filter_level_search = 0;
// Disable Wiener and Self-guided Loop restoration filters.
sf->lpf_sf.disable_wiener_filter = true ;
sf->lpf_sf.disable_sgr_filter = true ;
sf->mv_sf.prune_mesh_search = PRUNE_MESH_SEARCH_LVL_2;
sf->winner_mode_sf.multi_winner_mode_type = MULTI_WINNER_MODE_FAST;
}
if (speed >= 6) {
sf->intra_sf.prune_smooth_intra_mode_for_chroma = 1;
sf->intra_sf.prune_filter_intra_level = 2;
sf->intra_sf.chroma_intra_pruning_with_hog = 4;
sf->intra_sf.intra_pruning_with_hog = 4;
sf->intra_sf.cfl_search_range = 1;
sf->intra_sf.top_intra_model_count_allowed = 2;
sf->intra_sf.adapt_top_model_rd_count_using_neighbors = 1;
sf->intra_sf.prune_luma_odd_delta_angles_in_intra = 1;
sf->part_sf.prune_rectangular_split_based_on_qidx =
allow_screen_content_tools ? 0 : 2;
sf->part_sf.prune_rect_part_using_4x4_var_deviation = true ;
sf->part_sf.prune_rect_part_using_none_pred_mode = true ;
sf->part_sf.prune_sub_8x8_partition_level =
allow_screen_content_tools ? 0 : 1;
sf->part_sf.prune_part4_search = 3;
// TODO(jingning): This might not be a good trade off if the
// target image quality is very low.
sf->part_sf.default_max_partition_size = BLOCK_32X32;
sf->mv_sf.use_bsize_dependent_search_method = 1;
sf->tx_sf.tx_type_search.winner_mode_tx_type_pruning = 3;
sf->tx_sf.tx_type_search.prune_tx_type_est_rd = 0;
sf->tx_sf.prune_intra_tx_depths_using_nn = true ;
sf->rd_sf.perform_coeff_opt = 6;
sf->rd_sf.tx_domain_dist_level = 3;
sf->lpf_sf.cdef_pick_method = CDEF_FAST_SEARCH_LVL4;
sf->lpf_sf.lpf_pick = LPF_PICK_FROM_Q;
sf->winner_mode_sf.multi_winner_mode_type = MULTI_WINNER_MODE_OFF;
sf->winner_mode_sf.prune_winner_mode_eval_level = 1;
sf->winner_mode_sf.dc_blk_pred_level = 1;
}
// The following should make all-intra mode speed 7 approximately equal
// to real-time speed 6,
// all-intra speed 8 close to real-time speed 7, and all-intra speed 9
// close to real-time speed 8
if (speed >= 7) {
sf->part_sf.default_min_partition_size = BLOCK_8X8;
sf->part_sf.partition_search_type = VAR_BASED_PARTITION;
sf->lpf_sf.cdef_pick_method = CDEF_PICK_FROM_Q;
sf->rt_sf.mode_search_skip_flags |= FLAG_SKIP_INTRA_DIRMISMATCH;
sf->rt_sf.var_part_split_threshold_shift = 7;
}
if (speed >= 8) {
sf->rt_sf.hybrid_intra_pickmode = 1;
sf->rt_sf.use_nonrd_pick_mode = 1;
sf->rt_sf.nonrd_check_partition_merge_mode = 1;
sf->rt_sf.var_part_split_threshold_shift = 8;
// Set mask for intra modes.
for (int i = 0; i < BLOCK_SIZES; ++i)
if (i >= BLOCK_32X32)
sf->rt_sf.intra_y_mode_bsize_mask_nrd[i] = INTRA_DC;
else
// Use DC, H, V intra mode for block sizes < 32X32.
sf->rt_sf.intra_y_mode_bsize_mask_nrd[i] = INTRA_DC_H_V;
}
if (speed >= 9) {
sf->inter_sf.coeff_cost_upd_level = INTERNAL_COST_UPD_SBROW;
sf->inter_sf.mode_cost_upd_level = INTERNAL_COST_UPD_SBROW;
sf->rt_sf.nonrd_check_partition_merge_mode = 0;
sf->rt_sf.hybrid_intra_pickmode = 0;
sf->rt_sf.var_part_split_threshold_shift = 9;
sf->rt_sf.vbp_prune_16x16_split_using_min_max_sub_blk_var = true ;
sf->rt_sf.prune_h_pred_using_best_mode_so_far = true ;
sf->rt_sf.enable_intra_mode_pruning_using_neighbors = true ;
sf->rt_sf.prune_intra_mode_using_best_sad_so_far = true ;
}
// As the speed feature prune_chroma_modes_using_luma_winner already
// constrains the number of chroma directional mode evaluations to a maximum
// of 1, the HOG computation and the associated pruning logic does not seem to
// help speed-up the chroma mode evaluations. Hence disable the speed feature
// chroma_intra_pruning_with_hog when prune_chroma_modes_using_luma_winner is
// enabled.
if (sf->intra_sf.prune_chroma_modes_using_luma_winner)
sf->intra_sf.chroma_intra_pruning_with_hog = 0;
}
static void set_good_speed_feature_framesize_dependent(
const AV1_COMP *const cpi, SPEED_FEATURES *const sf, int speed) {
const AV1_COMMON *const cm = &cpi->common;
const int is_480p_or_lesser = AOMMIN(cm->width, cm->height) <= 480;
const int is_480p_or_larger = AOMMIN(cm->width, cm->height) >= 480;
const int is_720p_or_larger = AOMMIN(cm->width, cm->height) >= 720;
const int is_1080p_or_larger = AOMMIN(cm->width, cm->height) >= 1080;
const int is_4k_or_larger = AOMMIN(cm->width, cm->height) >= 2160;
const bool use_hbd = cpi->oxcf.use_highbitdepth;
// Speed features applicable for temporal filtering and tpl modules may be
// changed based on frame type at places where the sf is applied (Example :
// use_downsampled_sad). This is because temporal filtering and tpl modules
// are called before this function (except for the first key frame).
// TODO(deepa.kg@ittiam.com): For the speed features applicable to temporal
// filtering and tpl modules, modify the sf initialization appropriately
// before calling the modules.
const int boosted = frame_is_boosted(cpi);
const int is_boosted_arf2_bwd_type =
boosted ||
cpi->ppi->gf_group.update_type[cpi->gf_frame_index] == INTNL_ARF_UPDATE;
const int is_lf_frame =
cpi->ppi->gf_group.update_type[cpi->gf_frame_index] == LF_UPDATE;
const int allow_screen_content_tools =
cm->features.allow_screen_content_tools;
if (is_480p_or_larger) {
sf->part_sf.use_square_partition_only_threshold = BLOCK_128X128;
if (is_720p_or_larger)
sf->part_sf.auto_max_partition_based_on_simple_motion = ADAPT_PRED;
else
sf->part_sf.auto_max_partition_based_on_simple_motion = RELAXED_PRED;
} else {
sf->part_sf.use_square_partition_only_threshold = BLOCK_64X64;
sf->part_sf.auto_max_partition_based_on_simple_motion = DIRECT_PRED;
if (use_hbd) sf->tx_sf.prune_tx_size_level = 1;
}
if (is_4k_or_larger) {
sf->part_sf.default_min_partition_size = BLOCK_8X8;
}
// TODO(huisu@google.com): train models for 720P and above.
if (!is_720p_or_larger) {
sf->part_sf.ml_partition_search_breakout_thresh[0] = 200; // BLOCK_8X8
sf->part_sf.ml_partition_search_breakout_thresh[1] = 250; // BLOCK_16X16
sf->part_sf.ml_partition_search_breakout_thresh[2] = 300; // BLOCK_32X32
sf->part_sf.ml_partition_search_breakout_thresh[3] = 500; // BLOCK_64X64
sf->part_sf.ml_partition_search_breakout_thresh[4] = -1; // BLOCK_128X128
sf->part_sf.ml_early_term_after_part_split_level = 1;
}
if (is_720p_or_larger) {
// TODO(chiyotsai@google.com): make this speed feature adaptive based on
// current block's vertical texture instead of hardcoded with resolution
sf->mv_sf.use_downsampled_sad = 2;
}
if (!is_720p_or_larger) {
const RateControlCfg *const rc_cfg = &cpi->oxcf.rc_cfg;
const int rate_tolerance =
AOMMIN(rc_cfg->under_shoot_pct, rc_cfg->over_shoot_pct);
sf->hl_sf.recode_tolerance = 25 + (rate_tolerance >> 2);
}
if (speed >= 1) {
if (is_480p_or_lesser) sf->inter_sf.skip_newmv_in_drl = 1;
if (is_720p_or_larger) {
sf->part_sf.use_square_partition_only_threshold = BLOCK_128X128;
} else if (is_480p_or_larger) {
sf->part_sf.use_square_partition_only_threshold = BLOCK_64X64;
} else {
sf->part_sf.use_square_partition_only_threshold = BLOCK_32X32;
}
if (!is_720p_or_larger) {
sf->part_sf.ml_partition_search_breakout_thresh[0] = 200; // BLOCK_8X8
sf->part_sf.ml_partition_search_breakout_thresh[1] = 250; // BLOCK_16X16
sf->part_sf.ml_partition_search_breakout_thresh[2] = 300; // BLOCK_32X32
sf->part_sf.ml_partition_search_breakout_thresh[3] = 300; // BLOCK_64X64
sf->part_sf.ml_partition_search_breakout_thresh[4] = -1; // BLOCK_128X128
}
sf->part_sf.ml_early_term_after_part_split_level = 2;
sf->lpf_sf.cdef_pick_method = CDEF_FAST_SEARCH_LVL1;
}
if (speed >= 2) {
if (is_720p_or_larger) {
sf->part_sf.use_square_partition_only_threshold = BLOCK_64X64;
} else if (is_480p_or_larger) {
sf->part_sf.use_square_partition_only_threshold = BLOCK_32X32;
} else {
sf->part_sf.use_square_partition_only_threshold = BLOCK_32X32;
}
if (is_720p_or_larger) {
sf->part_sf.partition_search_breakout_dist_thr = (1 << 24);
sf->part_sf.partition_search_breakout_rate_thr = 120;
} else {
sf->part_sf.partition_search_breakout_dist_thr = (1 << 22);
sf->part_sf.partition_search_breakout_rate_thr = 100;
}
if (is_720p_or_larger) {
sf->inter_sf.prune_obmc_prob_thresh = 16;
} else {
sf->inter_sf.prune_obmc_prob_thresh = 8;
}
if (is_480p_or_larger) {
sf->inter_sf.disable_interintra_wedge_var_thresh = 100;
} else {
sf->inter_sf.disable_interintra_wedge_var_thresh = UINT_MAX;
}
if (is_480p_or_lesser) sf->inter_sf.skip_ext_comp_nearmv_mode = 1;
if (is_720p_or_larger) {
sf->inter_sf.limit_inter_mode_cands = is_lf_frame ? 1 : 0;
} else {
sf->inter_sf.limit_inter_mode_cands = is_lf_frame ? 2 : 0;
}
if (is_480p_or_larger) {
sf->tx_sf.tx_type_search.prune_tx_type_using_stats = 1;
if (use_hbd) sf->tx_sf.prune_tx_size_level = 2;
} else {
if (use_hbd) sf->tx_sf.prune_tx_size_level = 3;
sf->tx_sf.tx_type_search.winner_mode_tx_type_pruning = boosted ? 0 : 1;
sf->winner_mode_sf.enable_winner_mode_for_tx_size_srch = boosted ? 0 : 1;
}
if (!is_720p_or_larger) {
sf->mv_sf.disable_second_mv = 1;
sf->mv_sf.auto_mv_step_size = 2;
} else {
sf->mv_sf.disable_second_mv = boosted ? 0 : 2;
sf->mv_sf.auto_mv_step_size = 1;
}
if (!is_720p_or_larger) {
sf->hl_sf.recode_tolerance = 50;
sf->inter_sf.disable_interinter_wedge_newmv_search =
is_boosted_arf2_bwd_type ? 0 : 1;
sf->inter_sf.enable_fast_wedge_mask_search = 1;
}
}
if (speed >= 3) {
sf->inter_sf.enable_fast_wedge_mask_search = 1;
sf->inter_sf.skip_newmv_in_drl = 2;
sf->inter_sf.skip_ext_comp_nearmv_mode = 1;
sf->inter_sf.limit_inter_mode_cands = is_lf_frame ? 3 : 0;
sf->inter_sf.disable_interinter_wedge_newmv_search = boosted ? 0 : 1;
sf->tx_sf.tx_type_search.winner_mode_tx_type_pruning = 1;
sf->winner_mode_sf.enable_winner_mode_for_tx_size_srch =
frame_is_intra_only(&cpi->common) ? 0 : 1;
sf->part_sf.ml_early_term_after_part_split_level = 0;
if (is_720p_or_larger) {
sf->part_sf.partition_search_breakout_dist_thr = (1 << 25);
sf->part_sf.partition_search_breakout_rate_thr = 200;
sf->part_sf.skip_non_sq_part_based_on_none = is_lf_frame ? 2 : 0;
} else {
sf->part_sf.max_intra_bsize = BLOCK_32X32;
sf->part_sf.partition_search_breakout_dist_thr = (1 << 23);
sf->part_sf.partition_search_breakout_rate_thr = 120;
sf->part_sf.skip_non_sq_part_based_on_none = is_lf_frame ? 1 : 0;
}
if (use_hbd) sf->tx_sf.prune_tx_size_level = 3;
if (is_480p_or_larger) {
sf->part_sf.early_term_after_none_split = 1;
} else {
sf->part_sf.early_term_after_none_split = 0;
}
if (is_720p_or_larger) {
sf->intra_sf.skip_intra_in_interframe = boosted ? 1 : 2;
} else {
sf->intra_sf.skip_intra_in_interframe = boosted ? 1 : 3;
}
if (is_720p_or_larger) {
sf->inter_sf.disable_interinter_wedge_var_thresh = 100;
sf->inter_sf.limit_txfm_eval_per_mode = boosted ? 0 : 1;
} else {
sf->inter_sf.disable_interinter_wedge_var_thresh = UINT_MAX;
sf->inter_sf.limit_txfm_eval_per_mode = boosted ? 0 : 2;
sf->lpf_sf.cdef_pick_method = CDEF_FAST_SEARCH_LVL2;
}
sf->inter_sf.disable_interintra_wedge_var_thresh = UINT_MAX;
}
if (speed >= 4) {
sf->tx_sf.tx_type_search.winner_mode_tx_type_pruning = 2;
sf->winner_mode_sf.enable_winner_mode_for_tx_size_srch = 1;
if (is_720p_or_larger) {
sf->part_sf.partition_search_breakout_dist_thr = (1 << 26);
} else {
sf->part_sf.partition_search_breakout_dist_thr = (1 << 24);
}
sf->part_sf.early_term_after_none_split = 1;
if (is_480p_or_larger) {
sf->tx_sf.tx_type_search.prune_tx_type_using_stats = 2;
} else {
sf->mv_sf.skip_fullpel_search_using_startmv = boosted ? 0 : 1;
}
sf->inter_sf.disable_interinter_wedge_var_thresh = UINT_MAX;
sf->inter_sf.prune_obmc_prob_thresh = INT_MAX;
sf->inter_sf.limit_txfm_eval_per_mode = boosted ? 0 : 2;
if (is_480p_or_lesser) sf->inter_sf.skip_newmv_in_drl = 3;
if (is_720p_or_larger) {
sf->inter_sf.prune_comp_ref_frames = 1;
} else if (is_480p_or_larger) {
sf->inter_sf.prune_comp_ref_frames = is_boosted_arf2_bwd_type ? 0 : 1;
}
if (is_720p_or_larger)
sf->hl_sf.recode_tolerance = 32;
else
sf->hl_sf.recode_tolerance = 55;
sf->intra_sf.skip_intra_in_interframe = 4;
sf->lpf_sf.cdef_pick_method = CDEF_FAST_SEARCH_LVL3;
}
if (speed >= 5) {
if (is_720p_or_larger) {
sf->inter_sf.prune_warped_prob_thresh = 16;
} else if (is_480p_or_larger) {
sf->inter_sf.prune_warped_prob_thresh = 8;
}
if (is_720p_or_larger) sf->hl_sf.recode_tolerance = 40;
sf->inter_sf.skip_newmv_in_drl = 4;
sf->inter_sf.prune_comp_ref_frames = 1;
sf->mv_sf.skip_fullpel_search_using_startmv = boosted ? 0 : 1;
if (!is_720p_or_larger) {
sf->inter_sf.mv_cost_upd_level = INTERNAL_COST_UPD_SBROW_SET;
sf->inter_sf.prune_nearest_near_mv_using_refmv_weight =
(boosted || allow_screen_content_tools) ? 0 : 1;
sf->mv_sf.use_downsampled_sad = 1;
}
if (!is_480p_or_larger) {
sf->part_sf.partition_search_breakout_dist_thr = (1 << 26);
}
if (is_480p_or_lesser) {
sf->inter_sf.prune_nearmv_using_neighbors = PRUNE_NEARMV_LEVEL1;
} else {
sf->inter_sf.prune_nearmv_using_neighbors = PRUNE_NEARMV_LEVEL2;
}
if (is_720p_or_larger)
sf->part_sf.ext_part_eval_based_on_cur_best =
(allow_screen_content_tools || frame_is_intra_only(cm)) ? 0 : 1;
if (is_480p_or_larger) {
sf->tpl_sf.reduce_num_frames = 1;
}
}
if (speed >= 6) {
sf->tx_sf.tx_type_search.winner_mode_tx_type_pruning = 4;
sf->inter_sf.prune_nearmv_using_neighbors = PRUNE_NEARMV_LEVEL3;
sf->inter_sf.prune_comp_ref_frames = 2;
sf->inter_sf.prune_nearest_near_mv_using_refmv_weight =
(boosted || allow_screen_content_tools) ? 0 : 1;
sf->mv_sf.skip_fullpel_search_using_startmv = boosted ? 0 : 2;
if (is_720p_or_larger) {
sf->part_sf.auto_max_partition_based_on_simple_motion = NOT_IN_USE;
} else if (is_480p_or_larger) {
sf->part_sf.auto_max_partition_based_on_simple_motion = DIRECT_PRED;
}
if (is_480p_or_larger) {
sf->hl_sf.allow_sub_blk_me_in_tf = 1;
}
if (is_1080p_or_larger) {
sf->part_sf.default_min_partition_size = BLOCK_8X8;
}
if (is_720p_or_larger) {
sf->inter_sf.disable_masked_comp = 1;
}
if (!is_720p_or_larger) {
sf->inter_sf.coeff_cost_upd_level = INTERNAL_COST_UPD_SBROW;
sf->inter_sf.mode_cost_upd_level = INTERNAL_COST_UPD_SBROW;
}
if (is_720p_or_larger) {
sf->part_sf.use_square_partition_only_threshold = BLOCK_32X32;
sf->part_sf.partition_search_breakout_dist_thr = (1 << 28);
} else {
sf->part_sf.use_square_partition_only_threshold = BLOCK_16X16;
sf->part_sf.partition_search_breakout_dist_thr = (1 << 26);
}
if (is_720p_or_larger) {
sf->inter_sf.prune_ref_mv_idx_search = 2;
} else {
sf->inter_sf.prune_ref_mv_idx_search = 1;
}
if (!is_720p_or_larger) {
sf->tx_sf.tx_type_search.fast_inter_tx_type_prob_thresh =
is_boosted_arf2_bwd_type ? 450 : 150;
}
sf->lpf_sf.cdef_pick_method = CDEF_FAST_SEARCH_LVL4;
sf->hl_sf.recode_tolerance = 55;
}
}
static void set_good_speed_features_framesize_independent(
const AV1_COMP *const cpi, SPEED_FEATURES *const sf, int speed) {
const AV1_COMMON *const cm = &cpi->common;
const GF_GROUP *const gf_group = &cpi->ppi->gf_group;
const int boosted = frame_is_boosted(cpi);
const int is_boosted_arf2_bwd_type =
boosted || gf_group->update_type[cpi->gf_frame_index] == INTNL_ARF_UPDATE;
const int is_inter_frame =
gf_group->frame_type[cpi->gf_frame_index] == INTER_FRAME;
const int allow_screen_content_tools =
cm->features.allow_screen_content_tools;
const int use_hbd = cpi->oxcf.use_highbitdepth;
if (!cpi->oxcf.tile_cfg.enable_large_scale_tile) {
sf->hl_sf.high_precision_mv_usage = LAST_MV_DATA;
}
// Speed 0 for all speed features that give neutral coding performance change.
sf->gm_sf.gm_search_type = boosted ? GM_REDUCED_REF_SEARCH_SKIP_L2_L3_ARF2
: GM_SEARCH_CLOSEST_REFS_ONLY;
sf->gm_sf.prune_ref_frame_for_gm_search = boosted ? 0 : 1;
sf->gm_sf.disable_gm_search_based_on_stats = 1;
sf->part_sf.less_rectangular_check_level = 1;
sf->part_sf.ml_prune_partition = 1;
sf->part_sf.prune_ext_partition_types_search_level = 1;
sf->part_sf.prune_part4_search = 2;
sf->part_sf.simple_motion_search_prune_rect = 1;
sf->part_sf.ml_predict_breakout_level = use_hbd ? 1 : 3;
sf->part_sf.reuse_prev_rd_results_for_part_ab = 1;
sf->part_sf.use_best_rd_for_pruning = 1;
sf->part_sf.simple_motion_search_prune_agg =
allow_screen_content_tools ? NO_PRUNING : SIMPLE_AGG_LVL0;
// TODO(debargha): Test, tweak and turn on either 1 or 2
sf->inter_sf.inter_mode_rd_model_estimation = 1;
sf->inter_sf.model_based_post_interp_filter_breakout = 1;
sf->inter_sf.prune_compound_using_single_ref = 1;
sf->inter_sf.prune_mode_search_simple_translation = 1;
sf->inter_sf.prune_ref_frame_for_rect_partitions =
(boosted || (allow_screen_content_tools))
? 0
: (is_boosted_arf2_bwd_type ? 1 : 2);
sf->inter_sf.reduce_inter_modes = boosted ? 1 : 2;
sf->inter_sf.selective_ref_frame = 1;
sf->inter_sf.use_dist_wtd_comp_flag = DIST_WTD_COMP_SKIP_MV_SEARCH;
sf->interp_sf.use_fast_interpolation_filter_search = 1;
sf->intra_sf.intra_pruning_with_hog = 1;
sf->tx_sf.adaptive_txb_search_level = 1;
sf->tx_sf.intra_tx_size_search_init_depth_sqr = 1;
sf->tx_sf.model_based_prune_tx_search_level = 1;
sf->tx_sf.tx_type_search.use_reduced_intra_txset = 1;
sf->tpl_sf.search_method = NSTEP_8PT;
sf->rt_sf.use_nonrd_pick_mode = 0;
sf->rt_sf.use_real_time_ref_set = 0;
if (cpi->twopass_frame.fr_content_type == FC_GRAPHICS_ANIMATION ||
cpi->use_screen_content_tools) {
sf->mv_sf.exhaustive_searches_thresh = (1 << 20);
} else {
sf->mv_sf.exhaustive_searches_thresh = (1 << 25);
}
sf->rd_sf.perform_coeff_opt = 1;
sf->hl_sf.superres_auto_search_type = SUPERRES_AUTO_DUAL;
if (speed >= 1) {
sf->hl_sf.adjust_num_frames_for_arf_filtering =
allow_screen_content_tools ? 0 : 1;
sf->part_sf.intra_cnn_based_part_prune_level =
allow_screen_content_tools ? 0 : 2;
sf->part_sf.simple_motion_search_early_term_none = 1;
// TODO(Venkat): Clean-up frame type dependency for
// simple_motion_search_split in partition search function and set the
// speed feature accordingly
sf->part_sf.simple_motion_search_split = allow_screen_content_tools ? 1 : 2;
sf->part_sf.ml_predict_breakout_level = use_hbd ? 2 : 3;
sf->mv_sf.exhaustive_searches_thresh <<= 1;
sf->mv_sf.obmc_full_pixel_search_level = 1;
sf->mv_sf.use_accurate_subpel_search = USE_4_TAPS;
sf->mv_sf.disable_extensive_joint_motion_search = 1;
sf->inter_sf.prune_comp_search_by_single_result = boosted ? 2 : 1;
sf->inter_sf.prune_comp_type_by_comp_avg = 1;
sf->inter_sf.prune_comp_type_by_model_rd = boosted ? 0 : 1;
sf->inter_sf.prune_ref_frame_for_rect_partitions =
(frame_is_intra_only(&cpi->common) || (allow_screen_content_tools))
? 0
: (boosted ? 1 : 2);
sf->inter_sf.reduce_inter_modes = boosted ? 1 : 3;
sf->inter_sf.reuse_inter_intra_mode = 1;
sf->inter_sf.selective_ref_frame = 2;
sf->inter_sf.skip_arf_compound = 1;
sf->interp_sf.use_interp_filter = 1;
sf->intra_sf.prune_palette_search_level = 1;
sf->tx_sf.adaptive_txb_search_level = 2;
sf->tx_sf.inter_tx_size_search_init_depth_rect = 1;
sf->tx_sf.inter_tx_size_search_init_depth_sqr = 1;
sf->tx_sf.intra_tx_size_search_init_depth_rect = 1;
sf->tx_sf.model_based_prune_tx_search_level = 0;
sf->tx_sf.tx_type_search.ml_tx_split_thresh = 4000;
sf->tx_sf.tx_type_search.prune_2d_txfm_mode = TX_TYPE_PRUNE_2;
sf->tx_sf.tx_type_search.skip_tx_search = 1;
sf->rd_sf.perform_coeff_opt = boosted ? 2 : 3;
sf->rd_sf.tx_domain_dist_level = boosted ? 1 : 2;
sf->rd_sf.tx_domain_dist_thres_level = 1;
sf->lpf_sf.dual_sgr_penalty_level = 1;
sf->lpf_sf.enable_sgr_ep_pruning = 1;
// TODO(any, yunqing): move this feature to speed 0.
sf->tpl_sf.skip_alike_starting_mv = 1;
}
if (speed >= 2) {
sf->hl_sf.recode_loop = ALLOW_RECODE_KFARFGF;
sf->fp_sf.skip_motion_search_threshold = 25;
sf->gm_sf.num_refinement_steps = 2;
sf->part_sf.reuse_best_prediction_for_part_ab =
!frame_is_intra_only(&cpi->common);
sf->mv_sf.simple_motion_subpel_force_stop = QUARTER_PEL;
sf->mv_sf.subpel_iters_per_step = 1;
sf->mv_sf.reduce_search_range = 1;
// TODO(chiyotsai@google.com): We can get 10% speed up if we move
// adaptive_rd_thresh to speed 1. But currently it performs poorly on some
// clips (e.g. 5% loss on dinner_1080p). We need to examine the sequence a
// bit more closely to figure out why.
sf->inter_sf.adaptive_rd_thresh = 1;
sf->inter_sf.disable_interinter_wedge_var_thresh = 100;
sf->inter_sf.fast_interintra_wedge_search = 1;
sf->inter_sf.prune_comp_search_by_single_result = boosted ? 4 : 1;
sf->inter_sf.prune_ext_comp_using_neighbors = 1;
sf->inter_sf.prune_comp_using_best_single_mode_ref = 2;
sf->inter_sf.prune_comp_type_by_comp_avg = 2;
sf->inter_sf.selective_ref_frame = 3;
sf->inter_sf.use_dist_wtd_comp_flag = DIST_WTD_COMP_DISABLED;
sf->inter_sf.enable_fast_compound_mode_search = 1;
sf->inter_sf.reuse_mask_search_results = 1;
set_txfm_rd_gate_level(sf->inter_sf.txfm_rd_gate_level, boosted ? 0 : 1);
sf->inter_sf.inter_mode_txfm_breakout = boosted ? 0 : 1;
sf->inter_sf.alt_ref_search_fp = 1;
sf->interp_sf.adaptive_interp_filter_search = 1;
sf->interp_sf.disable_dual_filter = 1;
sf->intra_sf.disable_smooth_intra =
!frame_is_intra_only(&cpi->common) || (cpi->rc.frames_to_key > 1);
sf->intra_sf.intra_pruning_with_hog = 2;
sf->intra_sf.skip_intra_in_interframe = is_inter_frame ? 2 : 1;
sf->intra_sf.skip_filter_intra_in_inter_frames = 1;
sf->tpl_sf.prune_starting_mv = 1;
sf->tpl_sf.search_method = DIAMOND;
sf->rd_sf.perform_coeff_opt = is_boosted_arf2_bwd_type ? 3 : 4;
sf->rd_sf.use_mb_rd_hash = 1;
sf->lpf_sf.prune_wiener_based_on_src_var = 1;
sf->lpf_sf.prune_sgr_based_on_wiener = 1;
sf->lpf_sf.disable_loop_restoration_chroma = boosted ? 0 : 1;
sf->lpf_sf.reduce_wiener_window_size = boosted ? 0 : 1;
// TODO(any): Re-evaluate this feature set to 1 in speed 2.
sf->tpl_sf.allow_compound_pred = 0;
sf->tpl_sf.prune_ref_frames_in_tpl = 1;
}
if (speed >= 3) {
sf->hl_sf.high_precision_mv_usage = CURRENT_Q;
sf->gm_sf.prune_ref_frame_for_gm_search = 1;
sf->gm_sf.prune_zero_mv_with_sse = 1;
sf->gm_sf.num_refinement_steps = 0;
sf->part_sf.less_rectangular_check_level = 2;
sf->part_sf.simple_motion_search_prune_agg =
allow_screen_content_tools
? SIMPLE_AGG_LVL0
: (boosted ? SIMPLE_AGG_LVL1 : QIDX_BASED_AGG_LVL1);
sf->part_sf.prune_ext_part_using_split_info = 1;
sf->part_sf.simple_motion_search_rect_split = 1;
sf->mv_sf.full_pixel_search_level = 1;
sf->mv_sf.subpel_search_method = SUBPEL_TREE_PRUNED;
sf->mv_sf.search_method = DIAMOND;
sf->mv_sf.disable_second_mv = 2;
sf->mv_sf.prune_mesh_search = PRUNE_MESH_SEARCH_LVL_1;
sf->mv_sf.use_intrabc = 0;
sf->inter_sf.disable_interinter_wedge_newmv_search = boosted ? 0 : 1;
sf->inter_sf.mv_cost_upd_level = INTERNAL_COST_UPD_SBROW;
sf->inter_sf.disable_onesided_comp = 1;
sf->inter_sf.disable_interintra_wedge_var_thresh = UINT_MAX;
// TODO(any): Experiment with the early exit mechanism for speeds 0, 1 and 2
// and clean-up the speed feature
sf->inter_sf.perform_best_rd_based_gating_for_chroma = 1;
sf->inter_sf.prune_inter_modes_based_on_tpl = boosted ? 0 : 1;
sf->inter_sf.prune_comp_search_by_single_result = boosted ? 4 : 2;
sf->inter_sf.selective_ref_frame = 5;
sf->inter_sf.reuse_compound_type_decision = 1;
set_txfm_rd_gate_level(sf->inter_sf.txfm_rd_gate_level,
boosted ? 0 : (is_boosted_arf2_bwd_type ? 1 : 2));
sf->inter_sf.inter_mode_txfm_breakout = boosted ? 0 : 2;
sf->interp_sf.adaptive_interp_filter_search = 2;
// TODO(chiyotsai@google.com): the thresholds chosen for intra hog are
// inherited directly from luma hog with some minor tweaking. Eventually we
// should run this with a bayesian optimizer to find the Pareto frontier.
sf->intra_sf.chroma_intra_pruning_with_hog = 2;
sf->intra_sf.intra_pruning_with_hog = 3;
sf->intra_sf.prune_palette_search_level = 2;
sf->intra_sf.top_intra_model_count_allowed = 2;
sf->tpl_sf.prune_starting_mv = 2;
sf->tpl_sf.skip_alike_starting_mv = 2;
sf->tpl_sf.prune_intra_modes = 1;
sf->tpl_sf.reduce_first_step_size = 6;
sf->tpl_sf.subpel_force_stop = QUARTER_PEL;
sf->tpl_sf.gop_length_decision_method = 1;
sf->tx_sf.adaptive_txb_search_level = boosted ? 2 : 3;
sf->tx_sf.tx_type_search.use_skip_flag_prediction = 2;
sf->tx_sf.tx_type_search.prune_2d_txfm_mode = TX_TYPE_PRUNE_3;
// TODO(any): Refactor the code related to following winner mode speed
// features
sf->winner_mode_sf.enable_winner_mode_for_coeff_opt = 1;
sf->winner_mode_sf.enable_winner_mode_for_use_tx_domain_dist = 1;
sf->winner_mode_sf.motion_mode_for_winner_cand =
boosted ? 0
: gf_group->update_type[cpi->gf_frame_index] == INTNL_ARF_UPDATE ? 1
: 2;
sf->winner_mode_sf.prune_winner_mode_eval_level = boosted ? 0 : 4;
// For screen content, "prune_sgr_based_on_wiener = 2" cause large quality
// loss.
sf->lpf_sf.prune_sgr_based_on_wiener = allow_screen_content_tools ? 1 : 2;
sf->lpf_sf.prune_wiener_based_on_src_var = 2;
sf->lpf_sf.use_coarse_filter_level_search =
frame_is_intra_only(&cpi->common) ? 0 : 1;
sf->lpf_sf.use_downsampled_wiener_stats = 1;
}
if (speed >= 4) {
sf->mv_sf.subpel_search_method = SUBPEL_TREE_PRUNED_MORE;
sf->gm_sf.prune_zero_mv_with_sse = 2;
sf->gm_sf.downsample_level = 1;
sf->part_sf.simple_motion_search_prune_agg =
allow_screen_content_tools ? SIMPLE_AGG_LVL0 : SIMPLE_AGG_LVL2;
sf->part_sf.simple_motion_search_reduce_search_steps = 4;
sf->part_sf.prune_ext_part_using_split_info = 2;
sf->part_sf.ml_predict_breakout_level = 3;
sf->part_sf.prune_rectangular_split_based_on_qidx =
(allow_screen_content_tools || frame_is_intra_only(&cpi->common)) ? 0
: 1;
sf->inter_sf.alt_ref_search_fp = 2;
sf->inter_sf.txfm_rd_gate_level[TX_SEARCH_DEFAULT] = boosted ? 0 : 3;
sf->inter_sf.txfm_rd_gate_level[TX_SEARCH_MOTION_MODE] = boosted ? 0 : 5;
sf->inter_sf.txfm_rd_gate_level[TX_SEARCH_COMP_TYPE_MODE] = boosted ? 0 : 3;
sf->inter_sf.prune_inter_modes_based_on_tpl = boosted ? 0 : 2;
sf->inter_sf.prune_ext_comp_using_neighbors = 2;
sf->inter_sf.prune_obmc_prob_thresh = INT_MAX;
sf->inter_sf.disable_interinter_wedge_var_thresh = UINT_MAX;
sf->interp_sf.cb_pred_filter_search = 1;
sf->interp_sf.skip_sharp_interp_filter_search = 1;
sf->interp_sf.use_interp_filter = 2;
sf->intra_sf.intra_uv_mode_mask[TX_16X16] = UV_INTRA_DC_H_V_CFL;
sf->intra_sf.intra_uv_mode_mask[TX_32X32] = UV_INTRA_DC_H_V_CFL;
sf->intra_sf.intra_uv_mode_mask[TX_64X64] = UV_INTRA_DC_H_V_CFL;
// TODO(any): "intra_y_mode_mask" doesn't help much at speed 4.
// sf->intra_sf.intra_y_mode_mask[TX_16X16] = INTRA_DC_H_V;
// sf->intra_sf.intra_y_mode_mask[TX_32X32] = INTRA_DC_H_V;
// sf->intra_sf.intra_y_mode_mask[TX_64X64] = INTRA_DC_H_V;
sf->intra_sf.skip_intra_in_interframe = 4;
sf->mv_sf.simple_motion_subpel_force_stop = HALF_PEL;
sf->mv_sf.prune_mesh_search = PRUNE_MESH_SEARCH_LVL_2;
sf->tpl_sf.subpel_force_stop = HALF_PEL;
sf->tpl_sf.search_method = FAST_BIGDIA;
sf->tpl_sf.use_sad_for_mode_decision = 1;
sf->tx_sf.tx_type_search.fast_intra_tx_type_search = 1;
sf->rd_sf.perform_coeff_opt = is_boosted_arf2_bwd_type ? 5 : 7;
// TODO(any): Extend multi-winner mode processing support for inter frames
sf->winner_mode_sf.multi_winner_mode_type =
frame_is_intra_only(&cpi->common) ? MULTI_WINNER_MODE_DEFAULT
: MULTI_WINNER_MODE_OFF;
sf->winner_mode_sf.dc_blk_pred_level = boosted ? 0 : 2;
sf->lpf_sf.lpf_pick = LPF_PICK_FROM_FULL_IMAGE_NON_DUAL;
}
if (speed >= 5) {
sf->hl_sf.weight_calc_level_in_tf = 1;
sf->hl_sf.adjust_num_frames_for_arf_filtering =
allow_screen_content_tools ? 0 : 2;
sf->fp_sf.reduce_mv_step_param = 4;
sf->part_sf.simple_motion_search_prune_agg =
allow_screen_content_tools ? SIMPLE_AGG_LVL0 : SIMPLE_AGG_LVL3;
sf->part_sf.ext_partition_eval_thresh =
allow_screen_content_tools ? BLOCK_8X8 : BLOCK_16X16;
sf->part_sf.prune_sub_8x8_partition_level =
allow_screen_content_tools ? 1 : 2;
sf->mv_sf.warp_search_method = WARP_SEARCH_DIAMOND;
sf->inter_sf.prune_inter_modes_if_skippable = 1;
sf->inter_sf.prune_single_ref = is_boosted_arf2_bwd_type ? 0 : 1;
sf->inter_sf.txfm_rd_gate_level[TX_SEARCH_DEFAULT] = boosted ? 0 : 4;
sf->inter_sf.txfm_rd_gate_level[TX_SEARCH_COMP_TYPE_MODE] = boosted ? 0 : 5;
sf->inter_sf.enable_fast_compound_mode_search = 2;
sf->interp_sf.skip_interp_filter_search = boosted ? 0 : 1;
sf->intra_sf.chroma_intra_pruning_with_hog = 3;
// TODO(any): Extend multi-winner mode processing support for inter frames
sf->winner_mode_sf.multi_winner_mode_type =
frame_is_intra_only(&cpi->common) ? MULTI_WINNER_MODE_FAST
: MULTI_WINNER_MODE_OFF;
// Disable Self-guided Loop restoration filter.
sf->lpf_sf.disable_sgr_filter = true ;
sf->lpf_sf.disable_wiener_coeff_refine_search = true ;
sf->tpl_sf.prune_starting_mv = 3;
sf->tpl_sf.use_y_only_rate_distortion = 1;
sf->tpl_sf.subpel_force_stop = FULL_PEL;
sf->tpl_sf.gop_length_decision_method = 2;
sf->tpl_sf.use_sad_for_mode_decision = 2;
sf->winner_mode_sf.dc_blk_pred_level = 2;
sf->fp_sf.disable_recon = 1;
}
if (speed >= 6) {
sf->hl_sf.disable_extra_sc_testing = 1;
sf->hl_sf.second_alt_ref_filtering = 0;
sf->gm_sf.downsample_level = 2;
sf->inter_sf.prune_inter_modes_based_on_tpl = boosted ? 0 : 3;
sf->inter_sf.selective_ref_frame = 6;
sf->inter_sf.prune_single_ref = is_boosted_arf2_bwd_type ? 0 : 2;
sf->inter_sf.prune_ext_comp_using_neighbors = 3;
sf->intra_sf.chroma_intra_pruning_with_hog = 4;
sf->intra_sf.intra_pruning_with_hog = 4;
sf->intra_sf.intra_uv_mode_mask[TX_32X32] = UV_INTRA_DC;
sf->intra_sf.intra_uv_mode_mask[TX_64X64] = UV_INTRA_DC;
sf->intra_sf.intra_y_mode_mask[TX_32X32] = INTRA_DC;
sf->intra_sf.intra_y_mode_mask[TX_64X64] = INTRA_DC;
sf->intra_sf.early_term_chroma_palette_size_search = 1;
sf->part_sf.prune_rectangular_split_based_on_qidx =
boosted || allow_screen_content_tools ? 0 : 2;
sf->part_sf.prune_part4_search = 3;
sf->mv_sf.simple_motion_subpel_force_stop = FULL_PEL;
sf->mv_sf.use_bsize_dependent_search_method = 1;
sf->tpl_sf.gop_length_decision_method = 3;
sf->rd_sf.perform_coeff_opt = is_boosted_arf2_bwd_type ? 6 : 8;
sf->winner_mode_sf.dc_blk_pred_level = 3;
sf->winner_mode_sf.multi_winner_mode_type = MULTI_WINNER_MODE_OFF;
sf->fp_sf.skip_zeromv_motion_search = 1;
}
}
static void set_rt_speed_feature_framesize_dependent(const AV1_COMP *const cpi,
SPEED_FEATURES *const sf,
int speed) {
const AV1_COMMON *const cm = &cpi->common;
const int boosted = frame_is_boosted(cpi);
const int is_1080p_or_larger = AOMMIN(cm->width, cm->height) >= 1080;
const int is_720p_or_larger = AOMMIN(cm->width, cm->height) >= 720;
const int is_480p_or_larger = AOMMIN(cm->width, cm->height) >= 480;
const int is_360p_or_larger = AOMMIN(cm->width, cm->height) >= 360;
if (!is_360p_or_larger) {
sf->rt_sf.prune_intra_mode_based_on_mv_range = 1;
sf->rt_sf.prune_inter_modes_wrt_gf_arf_based_on_sad = 1;
if (speed >= 6)
sf->winner_mode_sf.prune_winner_mode_eval_level = boosted ? 0 : 2;
if (speed == 7) sf->rt_sf.prefer_large_partition_blocks = 2;
if (speed >= 7) {
sf->lpf_sf.cdef_pick_method = CDEF_PICK_FROM_Q;
sf->rt_sf.check_only_zero_zeromv_on_large_blocks = true ;
sf->rt_sf.use_rtc_tf = 2;
}
if (speed == 8) sf->rt_sf.prefer_large_partition_blocks = 1;
if (speed >= 8) {
sf->rt_sf.use_nonrd_filter_search = 1;
sf->rt_sf.tx_size_level_based_on_qstep = 1;
}
if (speed >= 9) {
sf->rt_sf.use_comp_ref_nonrd = 0;
sf->rt_sf.nonrd_aggressive_skip = 1;
sf->rt_sf.skip_intra_pred = 1;
// Only turn on enable_ref_short_signaling for low resolution when only
// LAST and GOLDEN ref frames are used.
sf->rt_sf.enable_ref_short_signaling =
(!sf->rt_sf.use_nonrd_altref_frame &&
(!sf->rt_sf.use_comp_ref_nonrd ||
(!sf->rt_sf.ref_frame_comp_nonrd[1] &&
!sf->rt_sf.ref_frame_comp_nonrd[2])));
// TODO(kyslov) Re-enable when AV1 models are trained
#if 0
#if CONFIG_RT_ML_PARTITIONING
if (!frame_is_intra_only(cm)) {
sf->part_sf.partition_search_type = ML_BASED_PARTITION;
sf->rt_sf.reuse_inter_pred_nonrd = 0;
}
#endif
#endif
sf->rt_sf.use_adaptive_subpel_search = false ;
}
if (speed >= 10) {
// TODO(yunqingwang@google.com): To be conservative, disable
// sf->rt_sf.estimate_motion_for_var_based_partition = 3 for speed 10/qvga
// for now. May enable it in the future.
sf->rt_sf.estimate_motion_for_var_based_partition = 0;
sf->rt_sf.skip_intra_pred = 2;
sf->rt_sf.hybrid_intra_pickmode = 3;
sf->rt_sf.reduce_mv_pel_precision_lowcomplex = 1;
sf->rt_sf.reduce_mv_pel_precision_highmotion = 2;
sf->rt_sf.use_nonrd_filter_search = 0;
}
} else {
sf->rt_sf.prune_intra_mode_based_on_mv_range = 2;
sf->intra_sf.skip_filter_intra_in_inter_frames = 1;
if (speed <= 5) {
sf->tx_sf.tx_type_search.fast_inter_tx_type_prob_thresh =
boosted ? INT_MAX : 350;
sf->winner_mode_sf.prune_winner_mode_eval_level = boosted ? 0 : 2;
}
if (speed == 6) sf->part_sf.disable_8x8_part_based_on_qidx = 1;
if (speed >= 6) sf->rt_sf.skip_newmv_mode_based_on_sse = 2;
if (speed == 7) {
sf->rt_sf.prefer_large_partition_blocks = 1;
// Enable this feature for [360p, 720p] resolution range initially.
// Only enable for low bitdepth to mitigate issue: b/303023614.
if (!cpi->rc.rtc_external_ratectrl &&
AOMMIN(cm->width, cm->height) <= 720 && !cpi->oxcf.use_highbitdepth)
sf->hl_sf.accurate_bit_estimate = cpi->oxcf.q_cfg.aq_mode == NO_AQ;
}
if (speed >= 7) {
sf->rt_sf.use_rtc_tf = 1;
}
if (speed == 8 && !cpi->ppi->use_svc) {
sf->rt_sf.short_circuit_low_temp_var = 0;
sf->rt_sf.use_nonrd_altref_frame = 1;
}
if (speed >= 8) sf->rt_sf.tx_size_level_based_on_qstep = 2;
if (speed >= 9) {
sf->rt_sf.gf_length_lvl = 1;
sf->rt_sf.skip_cdef_sb = 1;
sf->rt_sf.sad_based_adp_altref_lag = 2;
sf->rt_sf.reduce_mv_pel_precision_highmotion = 2;
sf->rt_sf.use_adaptive_subpel_search = true ;
sf->interp_sf.cb_pred_filter_search = 1;
}
if (speed >= 10) {
sf->rt_sf.hybrid_intra_pickmode = 2;
sf->rt_sf.sad_based_adp_altref_lag = 4;
sf->rt_sf.tx_size_level_based_on_qstep = 0;
sf->rt_sf.reduce_mv_pel_precision_highmotion = 3;
sf->rt_sf.use_adaptive_subpel_search = false ;
sf->interp_sf.cb_pred_filter_search = 2;
}
}
if (!is_480p_or_larger) {
if (speed == 7) {
sf->rt_sf.nonrd_check_partition_merge_mode = 2;
}
}
if (!is_720p_or_larger) {
if (speed >= 9) {
sf->rt_sf.force_large_partition_blocks_intra = 1;
}
} else {
if (speed >= 6) sf->rt_sf.skip_newmv_mode_based_on_sse = 3;
if (speed == 7) sf->rt_sf.prefer_large_partition_blocks = 0;
if (speed >= 7) {
sf->rt_sf.reduce_mv_pel_precision_lowcomplex = 2;
sf->rt_sf.reduce_mv_pel_precision_highmotion = 1;
}
if (speed >= 9) {
sf->rt_sf.sad_based_adp_altref_lag = 1;
sf->rt_sf.reduce_mv_pel_precision_lowcomplex = 0;
sf->rt_sf.reduce_mv_pel_precision_highmotion = 2;
}
if (speed >= 10) {
sf->rt_sf.sad_based_adp_altref_lag = 3;
sf->rt_sf.reduce_mv_pel_precision_highmotion = 3;
}
}
// TODO(Any): Check/Tune settings of other sfs for 1080p.
if (is_1080p_or_larger) {
if (speed >= 7) {
sf->rt_sf.reduce_mv_pel_precision_highmotion = 0;
sf->rt_sf.use_adaptive_subpel_search = 0;
}
if (speed >= 9) sf->interp_sf.cb_pred_filter_search = 0;
} else {
if (speed >= 9) sf->lpf_sf.cdef_pick_method = CDEF_PICK_FROM_Q;
if (speed >= 10) sf->rt_sf.nonrd_aggressive_skip = 1;
}
// TODO(marpan): Tune settings for speed 11 video mode,
if (speed >= 11 && cpi->oxcf.tune_cfg.content != AOM_CONTENT_SCREEN) {
sf->rt_sf.skip_cdef_sb = 1;
sf->rt_sf.force_only_last_ref = 1;
sf->rt_sf.selective_cdf_update = 1;
sf->rt_sf.use_nonrd_filter_search = 0;
if (is_360p_or_larger) {
sf->part_sf.fixed_partition_size = BLOCK_32X32;
sf->rt_sf.use_fast_fixed_part = 1;
sf->rt_sf.reduce_mv_pel_precision_lowcomplex = 2;
}
sf->rt_sf.increase_source_sad_thresh = 1;
sf->rt_sf.part_early_exit_zeromv = 2;
sf->rt_sf.set_zeromv_skip_based_on_source_sad = 2;
for (int i = 0; i < BLOCK_SIZES; ++i) {
sf->rt_sf.intra_y_mode_bsize_mask_nrd[i] = INTRA_DC;
}
sf->rt_sf.hybrid_intra_pickmode = 0;
}
// Setting for SVC, or when the ref_frame_config control is
// used to set the reference structure.
if (cpi->ppi->use_svc || cpi->ppi->rtc_ref.set_ref_frame_config) {
const RTC_REF *const rtc_ref = &cpi->ppi->rtc_ref;
// For SVC: for greater than 2 temporal layers, use better mv search on
// base temporal layers, and only on base spatial layer if highest
// resolution is above 640x360.
if (cpi->svc.number_temporal_layers >= 2 &&
cpi->svc.temporal_layer_id == 0 &&
(cpi->svc.spatial_layer_id == 0 ||
cpi->oxcf.frm_dim_cfg.width * cpi->oxcf.frm_dim_cfg.height <=
640 * 360)) {
sf->mv_sf.search_method = NSTEP;
sf->mv_sf.subpel_search_method = SUBPEL_TREE_PRUNED;
sf->rt_sf.fullpel_search_step_param = 10;
sf->rt_sf.reduce_mv_pel_precision_highmotion = 0;
if (cm->width * cm->height <= 352 * 288)
sf->rt_sf.nonrd_prune_ref_frame_search = 2;
sf->rt_sf.force_large_partition_blocks_intra = 0;
}
if (speed >= 8) {
if (cpi->svc.number_temporal_layers > 2)
sf->rt_sf.disable_cdf_update_non_reference_frame = true ;
sf->rt_sf.reduce_mv_pel_precision_highmotion = 3;
if (rtc_ref->non_reference_frame) {
sf->rt_sf.nonrd_aggressive_skip = 1;
sf->mv_sf.subpel_search_method = SUBPEL_TREE_PRUNED_MORE;
}
}
if (speed <= 9 && cpi->svc.number_temporal_layers > 2 &&
cpi->svc.temporal_layer_id == 0)
sf->rt_sf.check_only_zero_zeromv_on_large_blocks = false ;
else
sf->rt_sf.check_only_zero_zeromv_on_large_blocks = true ;
sf->rt_sf.frame_level_mode_cost_update = false ;
// Compound mode enabling.
if (rtc_ref->ref_frame_comp[0] || rtc_ref->ref_frame_comp[1] ||
rtc_ref->ref_frame_comp[2]) {
sf->rt_sf.use_comp_ref_nonrd = 1;
sf->rt_sf.ref_frame_comp_nonrd[0] =
rtc_ref->ref_frame_comp[0] && rtc_ref->reference[GOLDEN_FRAME - 1];
sf->rt_sf.ref_frame_comp_nonrd[1] =
rtc_ref->ref_frame_comp[1] && rtc_ref->reference[LAST2_FRAME - 1];
sf->rt_sf.ref_frame_comp_nonrd[2] =
rtc_ref->ref_frame_comp[2] && rtc_ref->reference[ALTREF_FRAME - 1];
} else {
sf->rt_sf.use_comp_ref_nonrd = 0;
}
if (cpi->svc.number_spatial_layers > 1 ||
cpi->svc.number_temporal_layers > 1)
sf->hl_sf.accurate_bit_estimate = 0;
sf->rt_sf.estimate_motion_for_var_based_partition = 1;
// For single layers RPS: bias/adjustment for recovery frame.
if (cpi->ppi->rtc_ref.bias_recovery_frame) {
sf->mv_sf.search_method = NSTEP;
sf->mv_sf.subpel_search_method = SUBPEL_TREE;
sf->rt_sf.fullpel_search_step_param = 8;
sf->rt_sf.nonrd_aggressive_skip = 0;
}
}
// Screen settings.
if (cpi->oxcf.tune_cfg.content == AOM_CONTENT_SCREEN) {
// TODO(marpan): Check settings for speed 7 and 8.
if (speed >= 7) {
sf->rt_sf.reduce_mv_pel_precision_highmotion = 1;
sf->mv_sf.use_bsize_dependent_search_method = 0;
sf->rt_sf.skip_cdef_sb = 1;
sf->rt_sf.increase_color_thresh_palette = 1;
if (!frame_is_intra_only(cm)) sf->rt_sf.dct_only_palette_nonrd = 1;
}
if (speed >= 8) {
sf->rt_sf.nonrd_check_partition_merge_mode = 3;
sf->rt_sf.nonrd_prune_ref_frame_search = 1;
sf->rt_sf.use_nonrd_filter_search = 0;
sf->rt_sf.prune_hv_pred_modes_using_src_sad = false ;
}
if (speed >= 9) {
sf->rt_sf.prune_idtx_nonrd = 1;
sf->rt_sf.part_early_exit_zeromv = 2;
sf->rt_sf.skip_lf_screen = 1;
sf->rt_sf.nonrd_prune_ref_frame_search = 3;
sf->rt_sf.var_part_split_threshold_shift = 10;
sf->mv_sf.subpel_search_method = SUBPEL_TREE_PRUNED_MORE;
sf->rt_sf.reduce_mv_pel_precision_highmotion = 3;
sf->rt_sf.reduce_mv_pel_precision_lowcomplex = 1;
sf->lpf_sf.cdef_pick_method = CDEF_PICK_FROM_Q;
sf->rt_sf.nonrd_check_partition_merge_mode = 0;
sf->interp_sf.cb_pred_filter_search = 0;
}
if (speed >= 10) {
if (cm->width * cm->height > 1920 * 1080)
sf->part_sf.disable_8x8_part_based_on_qidx = 1;
sf->rt_sf.screen_content_cdef_filter_qindex_thresh = 80;
sf->rt_sf.part_early_exit_zeromv = 1;
sf->rt_sf.nonrd_aggressive_skip = 1;
sf->rt_sf.thresh_active_maps_skip_lf_cdef = 90;
sf->rt_sf.hybrid_intra_pickmode = 0;
sf->rt_sf.dct_only_palette_nonrd = 1;
sf->rt_sf.prune_palette_search_nonrd = 1;
sf->rt_sf.prune_intra_mode_using_best_sad_so_far = true ;
sf->rt_sf.rc_faster_convergence_static = 1;
sf->rt_sf.rc_compute_spatial_var_sc = 1;
}
if (speed >= 11) {
sf->rt_sf.skip_lf_screen = 2;
sf->rt_sf.skip_cdef_sb = 2;
sf->rt_sf.prune_palette_search_nonrd = 2;
sf->rt_sf.increase_color_thresh_palette = 0;
sf->rt_sf.prune_h_pred_using_best_mode_so_far = true ;
sf->rt_sf.enable_intra_mode_pruning_using_neighbors = true ;
}
sf->rt_sf.skip_encoding_non_reference_slide_change =
cpi->oxcf.rc_cfg.drop_frames_water_mark > 0 ? 1 : 0;
sf->rt_sf.skip_newmv_flat_blocks_screen = 1;
sf->rt_sf.use_idtx_nonrd = 1;
sf->rt_sf.higher_thresh_scene_detection = 0;
sf->rt_sf.use_nonrd_altref_frame = 0;
sf->rt_sf.use_rtc_tf = 0;
sf->rt_sf.use_comp_ref_nonrd = 0;
sf->rt_sf.source_metrics_sb_nonrd = 1;
if (cpi->rc.high_source_sad == 1) {
sf->rt_sf.prefer_large_partition_blocks = 0;
sf->part_sf.max_intra_bsize = BLOCK_128X128;
for (int i = 0; i < BLOCK_SIZES; ++i) {
if (i > BLOCK_32X32)
sf->rt_sf.intra_y_mode_bsize_mask_nrd[i] = INTRA_DC;
else
sf->rt_sf.intra_y_mode_bsize_mask_nrd[i] = INTRA_DC_H_V;
}
}
if (speed >= 11 && cpi->rc.high_motion_content_screen_rtc) {
sf->rt_sf.higher_thresh_scene_detection = 1;
sf->rt_sf.force_only_last_ref = 1;
sf->rt_sf.use_nonrd_filter_search = 0;
sf->part_sf.fixed_partition_size = BLOCK_32X32;
sf->rt_sf.use_fast_fixed_part = 1;
sf->rt_sf.increase_source_sad_thresh = 1;
sf->rt_sf.selective_cdf_update = 1;
sf->mv_sf.search_method = FAST_DIAMOND;
--> --------------------
--> maximum size reached
--> --------------------
Messung V0.5 C=91 H=95 G=92
¤ Dauer der Verarbeitung: 0.15 Sekunden
¤
*© Formatika GbR, Deutschland