Introduce aggressive early txfm skip
Prediction of skip flag is done based only on SSE for normal mode in
presets 3 and 4. Default skip prediction is applicable for winner mode.
This logic is disabled for screen contents.
Instruction Count
cpu-used Reduction Quality Loss
3 0.98% -0.0073%
4 1.15% +0.0440%
STATS_CHANGED
Change-Id: Ia6ab667afcee33d18d0b53c88486509c1631fa5b
diff --git a/av1/encoder/block.h b/av1/encoder/block.h
index 461bdec..c99da91 100644
--- a/av1/encoder/block.h
+++ b/av1/encoder/block.h
@@ -452,6 +452,10 @@
int tx_size_search_method;
TX_MODE tx_mode;
+ // Used to control aggressiveness of skip flag prediction for mode processing
+ // (normal/winner mode)
+ unsigned int predict_skip_level;
+
// Copy out this SB's TPL block stats.
int valid_cost_b;
int64_t inter_cost_b[MAX_MC_FLOW_BLK_IN_SB * MAX_MC_FLOW_BLK_IN_SB];
diff --git a/av1/encoder/encoder.h b/av1/encoder/encoder.h
index 2e6cdf7..8c8474a 100644
--- a/av1/encoder/encoder.h
+++ b/av1/encoder/encoder.h
@@ -1044,6 +1044,11 @@
// speed feature is ON
unsigned int use_transform_domain_distortion[MODE_EVAL_TYPES];
+ // 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
+ unsigned int predict_skip_level[MODE_EVAL_TYPES];
+
AV1LfSync lf_row_sync;
AV1LrSync lr_row_sync;
AV1LrStruct lr_ctxt;
diff --git a/av1/encoder/rdopt.c b/av1/encoder/rdopt.c
index 4a024bc..0df6c44 100644
--- a/av1/encoder/rdopt.c
+++ b/av1/encoder/rdopt.c
@@ -4033,8 +4033,16 @@
// smaller than 32) into account.
const int16_t normalized_dc_q = dc_q >> 3;
const int64_t mse_thresh = (int64_t)normalized_dc_q * normalized_dc_q / 8;
- // Predict not to skip when mse is larger than threshold.
- if (mse > mse_thresh) return 0;
+ // For faster early skip decision, use dist to compare against threshold so
+ // that quality risk is less for the skip=1 decision. Otherwise, use mse
+ // since the fwd_txfm coeff checks will take care of quality
+ // TODO(any): Use dist to return 0 when predict_skip_level is 1
+ int64_t pred_err = (x->predict_skip_level >= 2) ? *dist : mse;
+ // Predict not to skip when error is larger than threshold.
+ if (pred_err > mse_thresh) return 0;
+ // Return as skip otherwise for aggressive early skip
+ else if (x->predict_skip_level >= 2)
+ return 1;
const int max_tx_size = max_predict_sf_tx_size[bsize];
const int tx_h = tx_size_high[max_tx_size];
@@ -4214,7 +4222,7 @@
// context and terminate early.
int64_t dist;
- if (cpi->sf.tx_type_search.use_skip_flag_prediction && is_inter &&
+ if (x->predict_skip_level && is_inter &&
(!xd->lossless[xd->mi[0]->segment_id]) &&
predict_skip_flag(x, bs, &dist, cpi->common.reduced_tx_set_used)) {
// Populate rdstats as per skip decision
@@ -6223,7 +6231,7 @@
// If we predict that skip is the optimal RD decision - set the respective
// context and terminate early.
int64_t dist;
- if (cpi->sf.tx_type_search.use_skip_flag_prediction &&
+ if (x->predict_skip_level &&
predict_skip_flag(x, bsize, &dist, cm->reduced_tx_set_used)) {
set_skip_flag(x, rd_stats, bsize, dist);
// Save the RD search results into tx_rd_record.
diff --git a/av1/encoder/rdopt.h b/av1/encoder/rdopt.h
index 971ce9d..0ca7384 100644
--- a/av1/encoder/rdopt.h
+++ b/av1/encoder/rdopt.h
@@ -305,6 +305,7 @@
case DEFAULT_EVAL:
x->use_default_inter_tx_type = 0;
x->use_default_intra_tx_type = 0;
+ x->predict_skip_level = cpi->predict_skip_level[DEFAULT_EVAL];
// Set default transform domain distortion type
set_tx_domain_dist_params(cpi, x, 0, 0);
@@ -320,6 +321,7 @@
cpi->oxcf.use_intra_default_tx_only);
x->use_default_inter_tx_type =
cpi->sf.tx_type_search.fast_inter_tx_type_search;
+ x->predict_skip_level = cpi->predict_skip_level[MODE_EVAL];
// Set transform domain distortion type for mode evaluation
set_tx_domain_dist_params(
@@ -337,6 +339,7 @@
case WINNER_MODE_EVAL:
x->use_default_inter_tx_type = 0;
x->use_default_intra_tx_type = 0;
+ x->predict_skip_level = cpi->predict_skip_level[WINNER_MODE_EVAL];
// Set transform domain distortion type for winner mode evaluation
set_tx_domain_dist_params(
diff --git a/av1/encoder/speed_features.c b/av1/encoder/speed_features.c
index 04783fd..5860806 100644
--- a/av1/encoder/speed_features.c
+++ b/av1/encoder/speed_features.c
@@ -94,6 +94,17 @@
{ USE_LARGESTALL, USE_LARGESTALL, USE_FULL_RD }
};
+// 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 unsigned int predict_skip_levels[3][MODE_EVAL_TYPES] = { { 0, 0, 0 },
+ { 1, 1, 1 },
+ { 1, 2, 1 } };
+
// scaling values to be used for gating wedge/compound segment based on best
// approximate rd
static int comp_type_rd_threshold_mul[3] = { 1, 11, 12 };
@@ -395,6 +406,8 @@
sf->mv.subpel_search_method = SUBPEL_TREE_PRUNED;
sf->simple_motion_search_prune_agg = 1;
sf->disable_sb_level_mv_cost_upd = 1;
+ sf->tx_type_search.use_skip_flag_prediction =
+ cm->allow_screen_content_tools ? 1 : 2;
}
if (speed >= 4) {
@@ -1006,6 +1019,13 @@
coeff_opt_dist_thresholds[cpi->sf.perform_coeff_opt],
sizeof(cpi->coeff_opt_dist_threshold));
+ // assert ensures that predict_skip_levels is accessed correctly
+ assert(cpi->sf.tx_type_search.use_skip_flag_prediction >= 0 &&
+ cpi->sf.tx_type_search.use_skip_flag_prediction < 3);
+ memcpy(cpi->predict_skip_level,
+ predict_skip_levels[cpi->sf.tx_type_search.use_skip_flag_prediction],
+ sizeof(cpi->predict_skip_level));
+
// Override speed feature setting for user config
if (cpi->oxcf.tx_size_search_method != USE_FULL_RD) {
cpi->sf.enable_winner_mode_for_tx_size_srch = 0;
