Add speed feature for fast eob search in trellis
8% speedup for 0.5% drop in performance on lowres. Some
experiments need to be finalized before this is paired
with a speed setting.
Change-Id: Ie782a0aa5d5e7316276e736629d0779c96447e79
diff --git a/av1/encoder/encodemb.c b/av1/encoder/encodemb.c
index 6edb0f8..96cb135 100644
--- a/av1/encoder/encodemb.c
+++ b/av1/encoder/encodemb.c
@@ -111,9 +111,8 @@
return eob;
}
- (void)fast_mode;
return av1_optimize_txb_new(cpi, mb, plane, block, tx_size, tx_type, txb_ctx,
- rate_cost, cpi->oxcf.sharpness);
+ rate_cost, cpi->oxcf.sharpness, fast_mode);
}
enum {
@@ -232,8 +231,8 @@
tx_size, tx_type, AV1_XFORM_QUANT_FP);
TXB_CTX txb_ctx;
get_txb_ctx(plane_bsize, tx_size, plane, a, l, &txb_ctx);
- av1_optimize_b(args->cpi, x, plane, block, tx_size, tx_type, &txb_ctx, 1,
- &dummy_rate_cost);
+ av1_optimize_b(args->cpi, x, plane, block, tx_size, tx_type, &txb_ctx,
+ args->cpi->sf.trellis_eob_fast, &dummy_rate_cost);
} else {
av1_xform_quant(
cm, x, plane, block, blk_row, blk_col, plane_bsize, tx_size, tx_type,
@@ -584,8 +583,8 @@
tx_size, tx_type, AV1_XFORM_QUANT_FP);
TXB_CTX txb_ctx;
get_txb_ctx(plane_bsize, tx_size, plane, a, l, &txb_ctx);
- av1_optimize_b(args->cpi, x, plane, block, tx_size, tx_type, &txb_ctx, 1,
- &dummy_rate_cost);
+ av1_optimize_b(args->cpi, x, plane, block, tx_size, tx_type, &txb_ctx,
+ args->cpi->sf.trellis_eob_fast, &dummy_rate_cost);
} else {
av1_xform_quant(
cm, x, plane, block, blk_row, blk_col, plane_bsize, tx_size, tx_type,
diff --git a/av1/encoder/encodetxb.c b/av1/encoder/encodetxb.c
index 1cfd50b..2fe5552 100644
--- a/av1/encoder/encodetxb.c
+++ b/av1/encoder/encodetxb.c
@@ -1489,6 +1489,36 @@
}
}
+static INLINE void update_coeff_eob_fast(int *eob, int shift,
+ const int16_t *dequant_ptr,
+ const int16_t *scan,
+ const tran_low_t *coeff_ptr,
+ tran_low_t *qcoeff_ptr,
+ tran_low_t *dqcoeff_ptr) {
+ // TODO(sarahparker) make this work for aomqm
+ int eob_out = *eob;
+ int zbin[2] = { dequant_ptr[0] + ROUND_POWER_OF_TWO(dequant_ptr[0] * 70, 7),
+ dequant_ptr[1] + ROUND_POWER_OF_TWO(dequant_ptr[1] * 70, 7) };
+
+ for (int i = *eob - 1; i >= 0; i--) {
+ const int rc = scan[i];
+ const int qcoeff = qcoeff_ptr[rc];
+ const int coeff = coeff_ptr[rc];
+ const int coeff_sign = (coeff >> 31);
+ int64_t abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+
+ if (((abs_coeff << (1 + shift)) < zbin[rc != 0]) || (qcoeff == 0)) {
+ eob_out--;
+ qcoeff_ptr[rc] = 0;
+ dqcoeff_ptr[rc] = 0;
+ } else {
+ break;
+ }
+ }
+
+ *eob = eob_out;
+}
+
static AOM_FORCE_INLINE void update_coeff_eob(
int *accu_rate, int64_t *accu_dist, int *eob, int *nz_num, int *nz_ci,
int si, TX_SIZE tx_size, TX_CLASS tx_class, int bwl, int height,
@@ -1624,32 +1654,43 @@
int av1_optimize_txb_new(const struct AV1_COMP *cpi, MACROBLOCK *x, int plane,
int block, TX_SIZE tx_size, TX_TYPE tx_type,
const TXB_CTX *const txb_ctx, int *rate_cost,
- int sharpness) {
- const AV1_COMMON *cm = &cpi->common;
+ int sharpness, int fast_mode) {
MACROBLOCKD *xd = &x->e_mbd;
+ struct macroblockd_plane *pd = &xd->plane[plane];
+ const struct macroblock_plane *p = &x->plane[plane];
+ const SCAN_ORDER *scan_order = get_scan(tx_size, tx_type);
+ const int16_t *scan = scan_order->scan;
+ const int shift = av1_get_tx_scale(tx_size);
+ int eob = p->eobs[block];
+ const int16_t *dequant = p->dequant_QTX;
+ tran_low_t *qcoeff = BLOCK_OFFSET(p->qcoeff, block);
+ tran_low_t *dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
+ const tran_low_t *tcoeff = BLOCK_OFFSET(p->coeff, block);
+
+ if (fast_mode) {
+ update_coeff_eob_fast(&eob, shift, dequant, scan, tcoeff, qcoeff, dqcoeff);
+ p->eobs[block] = eob;
+ if (eob == 0) {
+ *rate_cost = av1_cost_skip_txb(x, txb_ctx, plane, tx_size);
+ return eob;
+ }
+ }
+
+ const AV1_COMMON *cm = &cpi->common;
const PLANE_TYPE plane_type = get_plane_type(plane);
const TX_SIZE txs_ctx = get_txsize_entropy_ctx(tx_size);
const TX_CLASS tx_class = tx_type_to_class[tx_type];
const MB_MODE_INFO *mbmi = xd->mi[0];
- const struct macroblock_plane *p = &x->plane[plane];
- struct macroblockd_plane *pd = &xd->plane[plane];
- tran_low_t *qcoeff = BLOCK_OFFSET(p->qcoeff, block);
- tran_low_t *dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
- const tran_low_t *tcoeff = BLOCK_OFFSET(p->coeff, block);
- const int16_t *dequant = p->dequant_QTX;
const int bwl = get_txb_bwl(tx_size);
const int width = get_txb_wide(tx_size);
const int height = get_txb_high(tx_size);
assert(width == (1 << bwl));
const int is_inter = is_inter_block(mbmi);
- const SCAN_ORDER *scan_order = get_scan(tx_size, tx_type);
- const int16_t *scan = scan_order->scan;
const LV_MAP_COEFF_COST *txb_costs = &x->coeff_costs[txs_ctx][plane_type];
const int eob_multi_size = txsize_log2_minus4[tx_size];
const LV_MAP_EOB_COST *txb_eob_costs =
&x->eob_costs[eob_multi_size][plane_type];
- const int shift = av1_get_tx_scale(tx_size);
const int rshift =
(sharpness +
(cpi->oxcf.aq_mode == VARIANCE_AQ && mbmi->segment_id < 4
@@ -1667,7 +1708,6 @@
uint8_t levels_buf[TX_PAD_2D];
uint8_t *const levels = set_levels(levels_buf, width);
- int eob = p->eobs[block];
if (eob > 1) av1_txb_init_levels(qcoeff, width, height, levels);
@@ -1708,7 +1748,7 @@
#define UPDATE_COEFF_EOB_CASE(tx_class_literal) \
case tx_class_literal: \
- for (; si >= 0 && nz_num <= max_nz_num; --si) { \
+ for (; si >= 0 && nz_num <= max_nz_num && !fast_mode; --si) { \
update_coeff_eob(&accu_rate, &accu_dist, &eob, &nz_num, nz_ci, si, \
tx_size, tx_class_literal, bwl, height, \
txb_ctx->dc_sign_ctx, rdmult, shift, dequant, scan, \
diff --git a/av1/encoder/encodetxb.h b/av1/encoder/encodetxb.h
index bfc49de..0682590 100644
--- a/av1/encoder/encodetxb.h
+++ b/av1/encoder/encodetxb.h
@@ -78,7 +78,7 @@
int av1_optimize_txb_new(const struct AV1_COMP *cpi, MACROBLOCK *x, int plane,
int block, TX_SIZE tx_size, TX_TYPE tx_type,
const TXB_CTX *const txb_ctx, int *rate_cost,
- int sharpness);
+ int sharpness, int fast_mode);
// These numbers are empirically obtained.
static const int plane_rd_mult[REF_TYPES][PLANE_TYPES] = {
diff --git a/av1/encoder/rdopt.c b/av1/encoder/rdopt.c
index a80e65b..f114376 100644
--- a/av1/encoder/rdopt.c
+++ b/av1/encoder/rdopt.c
@@ -3259,8 +3259,8 @@
RDCOST(x->rdmult, 0, AOMMIN(this_rd_stats.dist, this_rd_stats.sse));
if (dist_cost_estimate - (dist_cost_estimate >> 3) > best_rd_) continue;
}
- av1_optimize_b(cpi, x, plane, block, tx_size, tx_type, txb_ctx, 1,
- &rate_cost);
+ av1_optimize_b(cpi, x, plane, block, tx_size, tx_type, txb_ctx,
+ cpi->sf.trellis_eob_fast, &rate_cost);
}
if (eobs_ptr[block] == 0) {
// When eob is 0, pixel domain distortion is more efficient and accurate.
@@ -3393,8 +3393,8 @@
} else {
av1_xform_quant(cm, x, plane, block, blk_row, blk_col, plane_bsize,
tx_size, best_tx_type, AV1_XFORM_QUANT_FP);
- av1_optimize_b(cpi, x, plane, block, tx_size, best_tx_type, txb_ctx, 1,
- &rate_cost);
+ av1_optimize_b(cpi, x, plane, block, tx_size, best_tx_type, txb_ctx,
+ cpi->sf.trellis_eob_fast, &rate_cost);
}
}
diff --git a/av1/encoder/speed_features.c b/av1/encoder/speed_features.c
index 21e299b..ff1018f 100644
--- a/av1/encoder/speed_features.c
+++ b/av1/encoder/speed_features.c
@@ -670,6 +670,8 @@
sf->mv.auto_mv_step_size = 0;
sf->comp_inter_joint_search_thresh = BLOCK_4X4;
sf->adaptive_rd_thresh = 0;
+ // TODO(sarahparker) Pair this with a speed setting once experiments are done
+ sf->trellis_eob_fast = 0;
sf->tx_size_search_method = USE_FULL_RD;
sf->inter_tx_size_search_init_depth_sqr = 0;
sf->inter_tx_size_search_init_depth_rect = 0;
diff --git a/av1/encoder/speed_features.h b/av1/encoder/speed_features.h
index dc1984f..7315bd7 100644
--- a/av1/encoder/speed_features.h
+++ b/av1/encoder/speed_features.h
@@ -668,6 +668,9 @@
// Flag used to control the extent of coeff R-D optimization
int perform_coeff_opt;
+ // Flag used to control the speed of the eob selection in trellis.
+ int trellis_eob_fast;
+
// This flag controls the use of non-RD mode decision.
int use_nonrd_pick_mode;
