Reset hash logic before winner mode processing
Hash logic is reset before winner mode evaluation
to avoid incorrect usage.
Instruction count
cpu-used Reduction BD rate impact
3 0.0% -0.07%
4 0.3% -0.01%
STATS_CHANGED
Change-Id: Icd4537eac3e100b00d8fd29bcaef992d9a6e0d52
diff --git a/av1/encoder/rd.h b/av1/encoder/rd.h
index 1c8de01..b50af0f 100644
--- a/av1/encoder/rd.h
+++ b/av1/encoder/rd.h
@@ -282,6 +282,35 @@
return coeff_opt_thresh;
}
+// Used to reset the state of tx/mb rd hash information
+static INLINE void reset_hash_records(MACROBLOCK *const x) {
+ int32_t record_idx;
+
+ // Reset the state for use_inter_txb_hash
+ for (record_idx = 0; record_idx < ((MAX_MIB_SIZE >> 1) * (MAX_MIB_SIZE >> 1));
+ record_idx++)
+ x->txb_rd_record_8X8[record_idx].num =
+ x->txb_rd_record_8X8[record_idx].index_start = 0;
+ for (record_idx = 0; record_idx < ((MAX_MIB_SIZE >> 2) * (MAX_MIB_SIZE >> 2));
+ record_idx++)
+ x->txb_rd_record_16X16[record_idx].num =
+ x->txb_rd_record_16X16[record_idx].index_start = 0;
+ for (record_idx = 0; record_idx < ((MAX_MIB_SIZE >> 3) * (MAX_MIB_SIZE >> 3));
+ record_idx++)
+ x->txb_rd_record_32X32[record_idx].num =
+ x->txb_rd_record_32X32[record_idx].index_start = 0;
+ for (record_idx = 0; record_idx < ((MAX_MIB_SIZE >> 4) * (MAX_MIB_SIZE >> 4));
+ record_idx++)
+ x->txb_rd_record_64X64[record_idx].num =
+ x->txb_rd_record_64X64[record_idx].index_start = 0;
+
+ // Reset the state for use_intra_txb_hash
+ x->txb_rd_record_intra.num = x->txb_rd_record_intra.index_start = 0;
+
+ // Reset the state for use_mb_rd_hash
+ x->mb_rd_record.num = x->mb_rd_record.index_start = 0;
+}
+
void av1_setup_pred_block(const MACROBLOCKD *xd,
struct buf_2d dst[MAX_MB_PLANE],
const YV12_BUFFER_CONFIG *src,
diff --git a/av1/encoder/rdopt.c b/av1/encoder/rdopt.c
index e1d92d4..00f0a3e 100644
--- a/av1/encoder/rdopt.c
+++ b/av1/encoder/rdopt.c
@@ -11732,81 +11732,83 @@
MB_MODE_INFO *const mbmi = xd->mi[0];
const int num_planes = av1_num_planes(cm);
- if (xd->lossless[mbmi->segment_id] == 0 && best_mode_index != THR_INVALID &&
- is_winner_mode_processing_enabled(cpi, mbmi, best_mbmode->mode)) {
- int skip_blk = 0;
- RD_STATS rd_stats_y, rd_stats_uv;
- const int skip_ctx = av1_get_skip_context(xd);
-
+ if (is_winner_mode_processing_enabled(cpi, mbmi, best_mbmode->mode)) {
// Set params for winner mode evaluation
set_mode_eval_params(cpi, x, WINNER_MODE_EVAL);
- *mbmi = *best_mbmode;
+ if (xd->lossless[mbmi->segment_id] == 0 && best_mode_index != THR_INVALID) {
+ int skip_blk = 0;
+ RD_STATS rd_stats_y, rd_stats_uv;
+ const int skip_ctx = av1_get_skip_context(xd);
- set_ref_ptrs(cm, xd, mbmi->ref_frame[0], mbmi->ref_frame[1]);
+ *mbmi = *best_mbmode;
- // Select prediction reference frames.
- for (int i = 0; i < num_planes; i++) {
- xd->plane[i].pre[0] = yv12_mb[mbmi->ref_frame[0]][i];
- if (has_second_ref(mbmi))
- xd->plane[i].pre[1] = yv12_mb[mbmi->ref_frame[1]][i];
- }
+ set_ref_ptrs(cm, xd, mbmi->ref_frame[0], mbmi->ref_frame[1]);
- if (is_inter_mode(mbmi->mode)) {
- const int mi_row = xd->mi_row;
- const int mi_col = xd->mi_col;
- av1_enc_build_inter_predictor(cm, xd, mi_row, mi_col, NULL, bsize, 0,
- av1_num_planes(cm) - 1);
- if (mbmi->motion_mode == OBMC_CAUSAL)
- av1_build_obmc_inter_predictors_sb(cm, xd);
+ // Select prediction reference frames.
+ for (int i = 0; i < num_planes; i++) {
+ xd->plane[i].pre[0] = yv12_mb[mbmi->ref_frame[0]][i];
+ if (has_second_ref(mbmi))
+ xd->plane[i].pre[1] = yv12_mb[mbmi->ref_frame[1]][i];
+ }
- av1_subtract_plane(x, bsize, 0);
- if (x->tx_mode == TX_MODE_SELECT && !xd->lossless[mbmi->segment_id]) {
- pick_tx_size_type_yrd(cpi, x, &rd_stats_y, bsize, INT64_MAX);
- assert(rd_stats_y.rate != INT_MAX);
+ if (is_inter_mode(mbmi->mode)) {
+ const int mi_row = xd->mi_row;
+ const int mi_col = xd->mi_col;
+ av1_enc_build_inter_predictor(cm, xd, mi_row, mi_col, NULL, bsize, 0,
+ av1_num_planes(cm) - 1);
+ if (mbmi->motion_mode == OBMC_CAUSAL)
+ av1_build_obmc_inter_predictors_sb(cm, xd);
+
+ av1_subtract_plane(x, bsize, 0);
+ if (x->tx_mode == TX_MODE_SELECT && !xd->lossless[mbmi->segment_id]) {
+ pick_tx_size_type_yrd(cpi, x, &rd_stats_y, bsize, INT64_MAX);
+ assert(rd_stats_y.rate != INT_MAX);
+ } else {
+ super_block_yrd(cpi, x, &rd_stats_y, bsize, INT64_MAX);
+ memset(mbmi->inter_tx_size, mbmi->tx_size,
+ sizeof(mbmi->inter_tx_size));
+ for (int i = 0; i < xd->n4_h * xd->n4_w; ++i)
+ set_blk_skip(x, 0, i, rd_stats_y.skip);
+ }
} else {
super_block_yrd(cpi, x, &rd_stats_y, bsize, INT64_MAX);
- memset(mbmi->inter_tx_size, mbmi->tx_size, sizeof(mbmi->inter_tx_size));
- for (int i = 0; i < xd->n4_h * xd->n4_w; ++i)
- set_blk_skip(x, 0, i, rd_stats_y.skip);
}
- } else {
- super_block_yrd(cpi, x, &rd_stats_y, bsize, INT64_MAX);
- }
- if (num_planes > 1) {
- super_block_uvrd(cpi, x, &rd_stats_uv, bsize, INT64_MAX);
- } else {
- av1_init_rd_stats(&rd_stats_uv);
- }
+ if (num_planes > 1) {
+ super_block_uvrd(cpi, x, &rd_stats_uv, bsize, INT64_MAX);
+ } else {
+ av1_init_rd_stats(&rd_stats_uv);
+ }
- if (RDCOST(x->rdmult,
- x->skip_cost[skip_ctx][0] + rd_stats_y.rate + rd_stats_uv.rate,
- (rd_stats_y.dist + rd_stats_uv.dist)) >
- RDCOST(x->rdmult, x->skip_cost[skip_ctx][1],
- (rd_stats_y.sse + rd_stats_uv.sse))) {
- skip_blk = 1;
- rd_stats_y.rate = x->skip_cost[skip_ctx][1];
- rd_stats_uv.rate = 0;
- rd_stats_y.dist = rd_stats_y.sse;
- rd_stats_uv.dist = rd_stats_uv.sse;
- } else {
- skip_blk = 0;
- rd_stats_y.rate += x->skip_cost[skip_ctx][0];
- }
+ if (RDCOST(x->rdmult,
+ x->skip_cost[skip_ctx][0] + rd_stats_y.rate + rd_stats_uv.rate,
+ (rd_stats_y.dist + rd_stats_uv.dist)) >
+ RDCOST(x->rdmult, x->skip_cost[skip_ctx][1],
+ (rd_stats_y.sse + rd_stats_uv.sse))) {
+ skip_blk = 1;
+ rd_stats_y.rate = x->skip_cost[skip_ctx][1];
+ rd_stats_uv.rate = 0;
+ rd_stats_y.dist = rd_stats_y.sse;
+ rd_stats_uv.dist = rd_stats_uv.sse;
+ } else {
+ skip_blk = 0;
+ rd_stats_y.rate += x->skip_cost[skip_ctx][0];
+ }
- if (RDCOST(x->rdmult, best_rate_y + best_rate_uv, rd_cost->dist) >
- RDCOST(x->rdmult, rd_stats_y.rate + rd_stats_uv.rate,
- (rd_stats_y.dist + rd_stats_uv.dist))) {
- best_mbmode->tx_size = mbmi->tx_size;
- av1_copy(best_mbmode->inter_tx_size, mbmi->inter_tx_size);
- av1_copy_array(ctx->blk_skip, x->blk_skip, ctx->num_4x4_blk);
- av1_copy_array(ctx->tx_type_map, xd->tx_type_map, ctx->num_4x4_blk);
- rd_cost->rate +=
- (rd_stats_y.rate + rd_stats_uv.rate - best_rate_y - best_rate_uv);
- rd_cost->dist = rd_stats_y.dist + rd_stats_uv.dist;
- rd_cost->rdcost = RDCOST(x->rdmult, rd_cost->rate, rd_cost->dist);
- *best_skip2 = skip_blk;
+ if (RDCOST(x->rdmult, best_rate_y + best_rate_uv, rd_cost->dist) >
+ RDCOST(x->rdmult, rd_stats_y.rate + rd_stats_uv.rate,
+ (rd_stats_y.dist + rd_stats_uv.dist))) {
+ best_mbmode->tx_size = mbmi->tx_size;
+ av1_copy(best_mbmode->inter_tx_size, mbmi->inter_tx_size);
+ av1_copy_array(ctx->blk_skip, x->blk_skip, ctx->num_4x4_blk);
+ av1_copy_array(ctx->tx_type_map, xd->tx_type_map, ctx->num_4x4_blk);
+ rd_cost->rate +=
+ (rd_stats_y.rate + rd_stats_uv.rate - best_rate_y - best_rate_uv);
+ rd_cost->dist = rd_stats_y.dist + rd_stats_uv.dist;
+ rd_cost->rdcost = RDCOST(x->rdmult, rd_cost->rate, rd_cost->dist);
+ *best_skip2 = skip_blk;
+ }
}
}
}
diff --git a/av1/encoder/rdopt.h b/av1/encoder/rdopt.h
index add4e41..b371c95 100644
--- a/av1/encoder/rdopt.h
+++ b/av1/encoder/rdopt.h
@@ -374,6 +374,14 @@
// Set default transform type prune mode for winner mode evaluation
set_tx_type_prune(
sf, x, sf->tx_type_search.enable_winner_mode_tx_type_pruning, 1);
+
+ // Reset hash state for winner mode processing. Winner mode and subsequent
+ // transform/mode evaluations (palette/IntraBC) cann't reuse old data as
+ // the decisions would have been sub-optimal
+ // TODO(any): Move the evaluation of palette/IntraBC modes before winner
+ // mode is processed and clean-up the code below
+ reset_hash_records(x);
+
break;
default: assert(0);
}
diff --git a/av1/encoder/speed_features.c b/av1/encoder/speed_features.c
index 6596384..38cf4a7 100644
--- a/av1/encoder/speed_features.c
+++ b/av1/encoder/speed_features.c
@@ -396,8 +396,6 @@
// TODO(any): Experiment with the early exit mechanism for speeds 0, 1 and 2
// and clean-up the speed feature
sf->perform_best_rd_based_gating_for_chroma = 1;
- // TODO(any): Experiment on the dependency of this speed feature with
- // use_intra_txb_hash, use_inter_txb_hash and use_mb_rd_hash speed features
// TODO(any): Refactor the code related to following winner mode speed
// features
sf->enable_winner_mode_for_coeff_opt = 1;
@@ -421,6 +419,8 @@
if (speed >= 4) {
sf->selective_ref_frame = 4;
+ // TODO(any): Experiment with enabling of this speed feature as hash state
+ // is reset during winner mode processing
sf->use_intra_txb_hash = 0;
sf->tx_type_search.fast_intra_tx_type_search = 1;
sf->disable_loop_restoration_chroma =
