Tune txfm_rd_gate_level for motion mode rd
Increase the aggressiveness of skipping transform search based on
skip rd in mode evaluation stage of motion mode.
Instruction Count BD-Rate Loss(%)
cpu Reduction(%) avg.psnr ovr.psnr ssim
5 1.325 0.0496 0.0479 0.0488
6 1.411 0.0404 0.0431 0.0583
STATS_CHANGED for Good preset, speed 5 and 6
Change-Id: I344c5e15bcf165d7e52b64236d0300aa7ac2377b
diff --git a/av1/encoder/rdopt.c b/av1/encoder/rdopt.c
index ee7486d..67f5e41 100644
--- a/av1/encoder/rdopt.c
+++ b/av1/encoder/rdopt.c
@@ -1321,6 +1321,16 @@
const int mi_row = xd->mi_row;
const int mi_col = xd->mi_col;
int mode_index_start, mode_index_end;
+ int txfm_rd_gate_level = cpi->sf.inter_sf.txfm_rd_gate_level;
+
+ // Set aggressive level of transform rd gating for mode evaluation stage.
+ if (cpi->sf.inter_sf.motion_mode_txfm_rd_gating_offset &&
+ txfm_rd_gate_level && !eval_motion_mode &&
+ num_pels_log2_lookup[bsize] > 8) {
+ txfm_rd_gate_level += cpi->sf.inter_sf.motion_mode_txfm_rd_gating_offset;
+ txfm_rd_gate_level = AOMMIN(txfm_rd_gate_level, MAX_TX_RD_GATE_LEVEL);
+ }
+
// Modify the start and end index according to speed features. For example,
// if SIMPLE_TRANSLATION has already been searched according to
// the motion_mode_for_winner_cand speed feature, update the mode_index_start
@@ -1524,7 +1534,7 @@
if (rd_stats->rdcost < *best_est_rd) {
*best_est_rd = rd_stats->rdcost;
assert(sse_y >= 0);
- ref_skip_rd[1] = cpi->sf.inter_sf.txfm_rd_gate_level
+ ref_skip_rd[1] = txfm_rd_gate_level
? RDCOST(x->rdmult, mode_rate, (sse_y << 4))
: INT64_MAX;
}
@@ -1546,14 +1556,14 @@
// Perform full transform search
int64_t skip_rd = INT64_MAX;
int64_t skip_rdy = INT64_MAX;
- if (cpi->sf.inter_sf.txfm_rd_gate_level) {
+ if (txfm_rd_gate_level) {
// Check if the mode is good enough based on skip RD
int64_t sse_y = INT64_MAX;
int64_t curr_sse = get_sse(cpi, x, &sse_y);
skip_rd = RDCOST(x->rdmult, rd_stats->rate, curr_sse);
skip_rdy = RDCOST(x->rdmult, rd_stats->rate, (sse_y << 4));
int eval_txfm = check_txfm_eval(x, bsize, ref_skip_rd[0], skip_rd,
- cpi->sf.inter_sf.txfm_rd_gate_level, 0);
+ txfm_rd_gate_level, 0);
if (!eval_txfm) continue;
}
diff --git a/av1/encoder/rdopt_utils.h b/av1/encoder/rdopt_utils.h
index 91823d8..17e76fb 100644
--- a/av1/encoder/rdopt_utils.h
+++ b/av1/encoder/rdopt_utils.h
@@ -23,6 +23,7 @@
#endif
#define MAX_REF_MV_SEARCH 3
+#define MAX_TX_RD_GATE_LEVEL 5
#define INTER_INTRA_RD_THRESH_SCALE 9
#define INTER_INTRA_RD_THRESH_SHIFT 4
@@ -352,10 +353,12 @@
// Derive aggressiveness factor for gating the transform search
// Lower value indicates more aggressiveness. Be more conservative (high
// value) for (i) low quantizers (ii) regions where prediction is poor
- const int scale[5] = { INT_MAX, 4, 3, 2, 2 };
+ const int scale[MAX_TX_RD_GATE_LEVEL + 1] = { INT_MAX, 4, 3, 2, 2, 1 };
const int qslope = 2 * (!is_luma_only);
- const int level_to_qindex_map[5] = { 0, 0, 0, 80, 100 };
+ const int level_to_qindex_map[MAX_TX_RD_GATE_LEVEL + 1] = { 0, 0, 0,
+ 80, 100, 140 };
int aggr_factor = 4;
+ assert(level <= MAX_TX_RD_GATE_LEVEL);
const int pred_qindex_thresh = level_to_qindex_map[level];
if (!is_luma_only && level <= 2) {
aggr_factor = 4 * AOMMAX(1, ROUND_POWER_OF_TWO((MAXQ - x->qindex) * qslope,
@@ -374,7 +377,9 @@
// since best_skip_rd is computed after and skip_rd is computed (with 8-bit
// prediction signals blended for WEDGE/DIFFWTD rather than 16-bit) before
// interpolation filter search
- const int luma_mul[5] = { INT_MAX, 32, 29, 17, 17 };
+ const int luma_mul[MAX_TX_RD_GATE_LEVEL + 1] = {
+ INT_MAX, 32, 29, 17, 17, 17
+ };
int mul_factor = is_luma_only ? luma_mul[level] : 16;
int64_t rd_thresh =
(best_skip_rd == INT64_MAX)
diff --git a/av1/encoder/speed_features.c b/av1/encoder/speed_features.c
index 36bc645..9bef51c 100644
--- a/av1/encoder/speed_features.c
+++ b/av1/encoder/speed_features.c
@@ -1210,6 +1210,7 @@
sf->inter_sf.prune_inter_modes_if_skippable = 1;
sf->inter_sf.txfm_rd_gate_level = boosted ? 0 : 4;
+ sf->inter_sf.motion_mode_txfm_rd_gating_offset = boosted ? 0 : 1;
sf->inter_sf.enable_fast_compound_mode_search = 2;
sf->intra_sf.chroma_intra_pruning_with_hog = 3;
@@ -1998,6 +1999,7 @@
inter_sf->prune_warped_prob_thresh = 0;
inter_sf->reuse_compound_type_decision = 0;
inter_sf->txfm_rd_gate_level = 0;
+ inter_sf->motion_mode_txfm_rd_gating_offset = 0;
inter_sf->prune_inter_modes_if_skippable = 0;
inter_sf->disable_masked_comp = 0;
inter_sf->enable_fast_compound_mode_search = 0;
diff --git a/av1/encoder/speed_features.h b/av1/encoder/speed_features.h
index 6ce7e35..f351ed5 100644
--- a/av1/encoder/speed_features.h
+++ b/av1/encoder/speed_features.h
@@ -843,6 +843,10 @@
// Bypass transform search based on skip rd
int txfm_rd_gate_level;
+ // Set transform rd gating offset used in mode evaluation stage of motion
+ // mode. This sf is applicable only if txfm_rd_gate_level is enabled.
+ int motion_mode_txfm_rd_gating_offset;
+
// Limit the inter mode tested in the RD loop
int reduce_inter_modes;
