Clean up and turn on rd_debug for non-sub8x8 inter block
Change-Id: Ib56273c23b7c7cb7c44d5ed95078156f1b09144c
diff --git a/av1/common/blockd.h b/av1/common/blockd.h
index 84de8e4..1766a7c 100644
--- a/av1/common/blockd.h
+++ b/av1/common/blockd.h
@@ -225,9 +225,11 @@
int skip;
#if CONFIG_RD_DEBUG
int txb_coeff_cost[MAX_MB_PLANE];
+#if CONFIG_VAR_TX
int txb_coeff_cost_map[MAX_MB_PLANE][TXB_COEFF_COST_MAP_SIZE]
[TXB_COEFF_COST_MAP_SIZE];
-#endif
+#endif // CONFIG_VAR_TX
+#endif // CONFIG_RD_DEBUG
} RD_STATS;
// This structure now relates to 8x8 block regions.
diff --git a/av1/encoder/bitstream.c b/av1/encoder/bitstream.c
index 7030210..4192801 100644
--- a/av1/encoder/bitstream.c
+++ b/av1/encoder/bitstream.c
@@ -1742,7 +1742,9 @@
static int rd_token_stats_mismatch(RD_STATS *rd_stats, TOKEN_STATS *token_stats,
int plane) {
if (rd_stats->txb_coeff_cost[plane] != token_stats->cost) {
+#if CONFIG_VAR_TX
int r, c;
+#endif
printf("\nplane %d rd_stats->txb_coeff_cost %d token_stats->cost %d\n",
plane, rd_stats->txb_coeff_cost[plane], token_stats->cost);
#if CONFIG_VAR_TX
@@ -1794,9 +1796,6 @@
MODE_INFO *m;
int plane;
int bh, bw;
-#if CONFIG_RD_DEBUG
- int64_t txb_coeff_cost[MAX_MB_PLANE] = { 0 };
-#endif
#if CONFIG_PVQ
MB_MODE_INFO *mbmi;
BLOCK_SIZE bsize;
@@ -1888,9 +1887,9 @@
if (!m->mbmi.skip) {
assert(*tok < tok_end);
for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
+ MB_MODE_INFO *mbmi = &m->mbmi;
#if CONFIG_VAR_TX
const struct macroblockd_plane *const pd = &xd->plane[plane];
- MB_MODE_INFO *mbmi = &m->mbmi;
BLOCK_SIZE bsize = mbmi->sb_type;
const BLOCK_SIZE plane_bsize =
get_plane_block_size(AOMMAX(bsize, BLOCK_8X8), pd);
@@ -1930,7 +1929,8 @@
}
}
#if CONFIG_RD_DEBUG
- if (rd_token_stats_mismatch(&m->mbmi.rd_stats, &token_stats, plane)) {
+ if (mbmi->sb_type >= BLOCK_8X8 &&
+ rd_token_stats_mismatch(&m->mbmi.rd_stats, &token_stats, plane)) {
dump_mode_info(m);
assert(0);
}
@@ -1949,8 +1949,17 @@
TX_SIZE tx =
plane ? get_uv_tx_size(&m->mbmi, &xd->plane[plane]) : m->mbmi.tx_size;
TOKEN_STATS token_stats;
- token_stats.cost = 0;
+ init_token_stats(&token_stats);
pack_mb_tokens(w, tok, tok_end, cm->bit_depth, tx, &token_stats);
+#if CONFIG_RD_DEBUG
+ if (is_inter_block(mbmi) && mbmi->sb_type >= BLOCK_8X8 &&
+ rd_token_stats_mismatch(&m->mbmi.rd_stats, &token_stats, plane)) {
+ dump_mode_info(m);
+ assert(0);
+ }
+#else
+ (void)mbmi;
+#endif
#endif // CONFIG_VAR_TX
assert(*tok < tok_end && (*tok)->token == EOSB_TOKEN);
diff --git a/av1/encoder/rdopt.c b/av1/encoder/rdopt.c
index f293a04..0567de5 100644
--- a/av1/encoder/rdopt.c
+++ b/av1/encoder/rdopt.c
@@ -1229,10 +1229,11 @@
}
#if !CONFIG_PVQ
this_rd_stats.rate = rate_block(plane, block, coeff_ctx, tx_size, args);
-#if RD_DEBUG
+#if CONFIG_RD_DEBUG
av1_update_txb_coeff_cost(&this_rd_stats, plane, tx_size, blk_row, blk_col,
this_rd_stats.rate);
#endif
+
args->t_above[blk_col] = (x->plane[plane].eobs[block] > 0);
args->t_left[blk_row] = (x->plane[plane].eobs[block] > 0);
#else
@@ -3608,10 +3609,6 @@
mbmi->inter_tx_size[idy][idx] = best_tx_size[idy][idx];
mbmi->tx_size = best_tx;
mbmi->min_tx_size = best_min_tx_size;
-#if CONFIG_RD_DEBUG
- // record plane y's transform block coefficient cost
- mbmi->rd_stats = *rd_stats;
-#endif
memcpy(x->blk_skip[0], best_blk_skip, sizeof(best_blk_skip[0]) * n4);
}
@@ -7833,10 +7830,6 @@
is_cost_valid_uv =
super_block_uvrd(cpi, x, rd_stats_uv, bsize, ref_best_rd - rdcosty);
#endif // CONFIG_VAR_TX
-#if CONFIG_RD_DEBUG
- // record uv planes' transform block coefficient cost
- if (is_cost_valid_uv) av1_merge_rd_stats(&mbmi->rd_stats, rd_stats_uv);
-#endif
if (!is_cost_valid_uv) {
#if CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
continue;
@@ -7847,6 +7840,14 @@
}
/* clang-format on */
av1_merge_rd_stats(rd_stats, rd_stats_uv);
+#if CONFIG_RD_DEBUG
+ // record transform block coefficient cost
+ // TODO(angiebird): So far rd_debug tool only detects descrepancy of
+ // coefficient cost. Therefore, it is fine to copy rd_stats into mbmi
+ // here because we already collect the coefficient cost. Move this part to
+ // other place when we need to compare non-coefficient cost.
+ mbmi->rd_stats = *rd_stats;
+#endif
#if CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
if (rd_stats->skip) {
rd_stats->rate -= rd_stats_uv->rate + rd_stats_y->rate;
diff --git a/av1/encoder/rdopt.h b/av1/encoder/rdopt.h
index 462274e..8d1df95 100644
--- a/av1/encoder/rdopt.h
+++ b/av1/encoder/rdopt.h
@@ -30,17 +30,23 @@
static INLINE void av1_update_txb_coeff_cost(RD_STATS *rd_stats, int plane,
TX_SIZE tx_size, int blk_row,
int blk_col, int txb_coeff_cost) {
- const int txb_h = tx_size_high_unit[tx_size];
- const int txb_w = tx_size_wide_unit[tx_size];
- int idx, idy;
-
+ (void)blk_row;
+ (void)blk_col;
+ (void)tx_size;
rd_stats->txb_coeff_cost[plane] += txb_coeff_cost;
- for (idy = 0; idy < txb_h; ++idy)
- for (idx = 0; idx < txb_w; ++idx)
- rd_stats->txb_coeff_cost_map[plane][blk_row + idy][blk_col + idx] = 0;
+#if CONFIG_VAR_TX
+ {
+ const int txb_h = tx_size_high_unit[tx_size];
+ const int txb_w = tx_size_wide_unit[tx_size];
+ int idx, idy;
+ for (idy = 0; idy < txb_h; ++idy)
+ for (idx = 0; idx < txb_w; ++idx)
+ rd_stats->txb_coeff_cost_map[plane][blk_row + idy][blk_col + idx] = 0;
- rd_stats->txb_coeff_cost_map[plane][blk_row][blk_col] = txb_coeff_cost;
+ rd_stats->txb_coeff_cost_map[plane][blk_row][blk_col] = txb_coeff_cost;
+ }
+#endif
assert(blk_row < 16);
assert(blk_col < 16);
@@ -57,11 +63,15 @@
rd_stats->skip = 1;
#if CONFIG_RD_DEBUG
for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
- int r, c;
rd_stats->txb_coeff_cost[plane] = 0;
- for (r = 0; r < TXB_COEFF_COST_MAP_SIZE; ++r)
- for (c = 0; c < TXB_COEFF_COST_MAP_SIZE; ++c)
- rd_stats->txb_coeff_cost_map[plane][r][c] = 0;
+#if CONFIG_VAR_TX
+ {
+ int r, c;
+ for (r = 0; r < TXB_COEFF_COST_MAP_SIZE; ++r)
+ for (c = 0; c < TXB_COEFF_COST_MAP_SIZE; ++c)
+ rd_stats->txb_coeff_cost_map[plane][r][c] = 0;
+ }
+#endif
}
#endif
}
@@ -76,11 +86,15 @@
rd_stats->skip = 0;
#if CONFIG_RD_DEBUG
for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
- int r, c;
rd_stats->txb_coeff_cost[plane] = INT_MAX;
- for (r = 0; r < TXB_COEFF_COST_MAP_SIZE; ++r)
- for (c = 0; c < TXB_COEFF_COST_MAP_SIZE; ++c)
- rd_stats->txb_coeff_cost_map[plane][r][c] = INT_MAX;
+#if CONFIG_VAR_TX
+ {
+ int r, c;
+ for (r = 0; r < TXB_COEFF_COST_MAP_SIZE; ++r)
+ for (c = 0; c < TXB_COEFF_COST_MAP_SIZE; ++c)
+ rd_stats->txb_coeff_cost_map[plane][r][c] = INT_MAX;
+ }
+#endif
}
#endif
}
@@ -96,17 +110,21 @@
rd_stats_dst->skip &= rd_stats_src->skip;
#if CONFIG_RD_DEBUG
for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
- int r, c;
- int ref_txb_coeff_cost = 0;
rd_stats_dst->txb_coeff_cost[plane] += rd_stats_src->txb_coeff_cost[plane];
- // TODO(angiebird): optimize this part
- for (r = 0; r < TXB_COEFF_COST_MAP_SIZE; ++r)
- for (c = 0; c < TXB_COEFF_COST_MAP_SIZE; ++c) {
- rd_stats_dst->txb_coeff_cost_map[plane][r][c] +=
- rd_stats_src->txb_coeff_cost_map[plane][r][c];
- ref_txb_coeff_cost += rd_stats_dst->txb_coeff_cost_map[plane][r][c];
- }
- assert(ref_txb_coeff_cost == rd_stats_dst->txb_coeff_cost[plane]);
+#if CONFIG_VAR_TX
+ {
+ // TODO(angiebird): optimize this part
+ int r, c;
+ int ref_txb_coeff_cost = 0;
+ for (r = 0; r < TXB_COEFF_COST_MAP_SIZE; ++r)
+ for (c = 0; c < TXB_COEFF_COST_MAP_SIZE; ++c) {
+ rd_stats_dst->txb_coeff_cost_map[plane][r][c] +=
+ rd_stats_src->txb_coeff_cost_map[plane][r][c];
+ ref_txb_coeff_cost += rd_stats_dst->txb_coeff_cost_map[plane][r][c];
+ }
+ assert(ref_txb_coeff_cost == rd_stats_dst->txb_coeff_cost[plane]);
+ }
+#endif
}
#endif
}
