Reduce memory usage of inter_tx_size[] in MB_MODE_INFO
Reduce the length of inter_tx_size[] from 1024 to 16.
On a cif test sequence,
encoder memory consumption decreases by 18% (380MB -> 312MB);
decoder memory consumption decreases by 56% (21.4MB -> 9.4MB).
Change-Id: I42928eb9312748f96f4393c8d8040791f38f98b6
diff --git a/av1/decoder/decodeframe.c b/av1/decoder/decodeframe.c
index 212d953..aaeebb0f 100644
--- a/av1/decoder/decodeframe.c
+++ b/av1/decoder/decodeframe.c
@@ -176,11 +176,10 @@
(void)mi_row;
(void)mi_col;
const struct macroblockd_plane *const pd = &xd->plane[plane];
- const int tx_row = blk_row >> (1 - pd->subsampling_y);
- const int tx_col = blk_col >> (1 - pd->subsampling_x);
const TX_SIZE plane_tx_size =
plane ? av1_get_uv_tx_size(mbmi, pd->subsampling_x, pd->subsampling_y)
- : mbmi->inter_tx_size[tx_row][tx_col];
+ : mbmi->inter_tx_size[av1_get_txb_size_index(plane_bsize, blk_row,
+ blk_col)];
// Scale to match transform block unit.
const int max_blocks_high = max_block_high(xd, plane_bsize, plane);
const int max_blocks_wide = max_block_wide(xd, plane_bsize, plane);
diff --git a/av1/decoder/decodemv.c b/av1/decoder/decodemv.c
index 0661451..1ef9021 100644
--- a/av1/decoder/decodemv.c
+++ b/av1/decoder/decodemv.c
@@ -353,22 +353,20 @@
FRAME_CONTEXT *ec_ctx = xd->tile_ctx;
(void)cm;
int is_split = 0;
- const int tx_row = blk_row >> 1;
- const int tx_col = blk_col >> 1;
- const int max_blocks_high = max_block_high(xd, mbmi->sb_type, 0);
- const int max_blocks_wide = max_block_wide(xd, mbmi->sb_type, 0);
- TX_SIZE(*const inter_tx_size)
- [MAX_MIB_SIZE] =
- (TX_SIZE(*)[MAX_MIB_SIZE]) & mbmi->inter_tx_size[tx_row][tx_col];
+ const BLOCK_SIZE bsize = mbmi->sb_type;
+ const int max_blocks_high = max_block_high(xd, bsize, 0);
+ const int max_blocks_wide = max_block_wide(xd, bsize, 0);
if (blk_row >= max_blocks_high || blk_col >= max_blocks_wide) return;
assert(tx_size > TX_4X4);
if (depth == MAX_VARTX_DEPTH) {
- int idx, idy;
- inter_tx_size[0][0] = tx_size;
- for (idy = 0; idy < AOMMAX(1, tx_size_high_unit[tx_size] / 2); ++idy)
- for (idx = 0; idx < AOMMAX(1, tx_size_wide_unit[tx_size] / 2); ++idx)
- inter_tx_size[idy][idx] = tx_size;
+ for (int idy = 0; idy < tx_size_high_unit[tx_size]; ++idy) {
+ for (int idx = 0; idx < tx_size_wide_unit[tx_size]; ++idx) {
+ const int index =
+ av1_get_txb_size_index(bsize, blk_row + idy, blk_col + idx);
+ mbmi->inter_tx_size[index] = tx_size;
+ }
+ }
mbmi->tx_size = tx_size;
mbmi->min_tx_size = TXSIZEMIN(mbmi->min_tx_size, tx_size);
txfm_partition_update(xd->above_txfm_context + blk_col,
@@ -376,9 +374,9 @@
return;
}
- int ctx = txfm_partition_context(xd->above_txfm_context + blk_col,
- xd->left_txfm_context + blk_row,
- mbmi->sb_type, tx_size);
+ const int ctx = txfm_partition_context(xd->above_txfm_context + blk_col,
+ xd->left_txfm_context + blk_row,
+ mbmi->sb_type, tx_size);
is_split = aom_read_symbol(r, ec_ctx->txfm_partition_cdf[ctx], 2, ACCT_STR);
if (is_split) {
@@ -387,11 +385,13 @@
const int bsh = tx_size_high_unit[sub_txs];
if (sub_txs == TX_4X4) {
- int idx, idy;
- inter_tx_size[0][0] = sub_txs;
- for (idy = 0; idy < AOMMAX(1, tx_size_high_unit[tx_size] / 2); ++idy)
- for (idx = 0; idx < AOMMAX(1, tx_size_wide_unit[tx_size] / 2); ++idx)
- inter_tx_size[idy][idx] = inter_tx_size[0][0];
+ for (int idy = 0; idy < tx_size_high_unit[tx_size]; ++idy) {
+ for (int idx = 0; idx < tx_size_wide_unit[tx_size]; ++idx) {
+ const int index =
+ av1_get_txb_size_index(bsize, blk_row + idy, blk_col + idx);
+ mbmi->inter_tx_size[index] = sub_txs;
+ }
+ }
mbmi->tx_size = sub_txs;
mbmi->min_tx_size = mbmi->tx_size;
txfm_partition_update(xd->above_txfm_context + blk_col,
@@ -409,11 +409,13 @@
}
}
} else {
- int idx, idy;
- inter_tx_size[0][0] = tx_size;
- for (idy = 0; idy < AOMMAX(1, tx_size_high_unit[tx_size] / 2); ++idy)
- for (idx = 0; idx < AOMMAX(1, tx_size_wide_unit[tx_size] / 2); ++idx)
- inter_tx_size[idy][idx] = tx_size;
+ for (int idy = 0; idy < tx_size_high_unit[tx_size]; ++idy) {
+ for (int idx = 0; idx < tx_size_wide_unit[tx_size]; ++idx) {
+ const int index =
+ av1_get_txb_size_index(bsize, blk_row + idy, blk_col + idx);
+ mbmi->inter_tx_size[index] = tx_size;
+ }
+ }
mbmi->tx_size = tx_size;
mbmi->min_tx_size = TXSIZEMIN(mbmi->min_tx_size, tx_size);
txfm_partition_update(xd->above_txfm_context + blk_col,
@@ -933,9 +935,7 @@
}
} else {
mbmi->tx_size = read_tx_size(cm, xd, 1, !mbmi->skip, r);
- for (int idy = 0; idy < height; ++idy)
- for (int idx = 0; idx < width; ++idx)
- mbmi->inter_tx_size[idy >> 1][idx >> 1] = mbmi->tx_size;
+ memset(mbmi->inter_tx_size, mbmi->tx_size, sizeof(mbmi->inter_tx_size));
mbmi->min_tx_size = mbmi->tx_size;
set_txfm_ctxs(mbmi->tx_size, xd->n8_w, xd->n8_h, mbmi->skip, xd);
}
@@ -2188,14 +2188,8 @@
read_tx_size_vartx(cm, xd, mbmi, max_tx_size, 0, idy, idx, r);
} else {
mbmi->tx_size = read_tx_size(cm, xd, inter_block, !mbmi->skip, r);
-
- if (inter_block) {
- const int width = block_size_wide[bsize] >> tx_size_wide_log2[0];
- const int height = block_size_high[bsize] >> tx_size_high_log2[0];
- for (int idy = 0; idy < height; ++idy)
- for (int idx = 0; idx < width; ++idx)
- mbmi->inter_tx_size[idy >> 1][idx >> 1] = mbmi->tx_size;
- }
+ if (inter_block)
+ memset(mbmi->inter_tx_size, mbmi->tx_size, sizeof(mbmi->inter_tx_size));
mbmi->min_tx_size = mbmi->tx_size;
set_txfm_ctxs(mbmi->tx_size, xd->n8_w, xd->n8_h, mbmi->skip, xd);
}
