Palette: Use inverse_color_order to find color index faster.
Cherry-picked from aomedia/master: b1c3bb5
Change-Id: Icfc16070160fd9763abb1dbf5545103e62b4b9ff
diff --git a/av1/common/entropymode.c b/av1/common/entropymode.c
index 20e8904..78f4ffe 100644
--- a/av1/common/entropymode.c
+++ b/av1/common/entropymode.c
@@ -892,13 +892,14 @@
#if CONFIG_PALETTE
int av1_get_palette_color_context(const uint8_t *color_map, int cols, int r,
- int c, int n, int *color_order) {
+ int c, int n, uint8_t *color_order,
+ int *color_idx) {
int i, j, max, max_idx, temp;
int scores[PALETTE_MAX_SIZE + 10];
int weights[4] = { 3, 2, 3, 2 };
int color_ctx = 0;
int color_neighbors[4];
-
+ int inverse_color_order[PALETTE_MAX_SIZE];
assert(n <= PALETTE_MAX_SIZE);
if (c - 1 >= 0)
@@ -918,7 +919,10 @@
else
color_neighbors[3] = -1;
- for (i = 0; i < PALETTE_MAX_SIZE; ++i) color_order[i] = i;
+ for (i = 0; i < PALETTE_MAX_SIZE; ++i) {
+ color_order[i] = i;
+ inverse_color_order[i] = i;
+ }
memset(scores, 0, PALETTE_MAX_SIZE * sizeof(scores[0]));
for (i = 0; i < 4; ++i) {
if (color_neighbors[i] >= 0) scores[color_neighbors[i]] += weights[i];
@@ -944,6 +948,8 @@
temp = color_order[i];
color_order[i] = color_order[max_idx];
color_order[max_idx] = temp;
+ inverse_color_order[color_order[i]] = i;
+ inverse_color_order[color_order[max_idx]] = max_idx;
}
}
@@ -956,7 +962,9 @@
}
if (color_ctx >= PALETTE_COLOR_CONTEXTS) color_ctx = 0;
-
+ if (color_idx != NULL) {
+ *color_idx = inverse_color_order[color_map[r * cols + c]];
+ }
return color_ctx;
}
#endif // CONFIG_PALETTE
diff --git a/av1/common/entropymode.h b/av1/common/entropymode.h
index d0ad807..49dbbb0 100644
--- a/av1/common/entropymode.h
+++ b/av1/common/entropymode.h
@@ -320,7 +320,8 @@
#if CONFIG_PALETTE
int av1_get_palette_color_context(const uint8_t *color_map, int cols, int r,
- int c, int n, int *color_order);
+ int c, int n, uint8_t *color_order,
+ int *color_idx);
#endif // CONFIG_PALETTE
#ifdef __cplusplus
diff --git a/av1/decoder/detokenize.c b/av1/decoder/detokenize.c
index 7077788..b5f26d2 100644
--- a/av1/decoder/detokenize.c
+++ b/av1/decoder/detokenize.c
@@ -338,15 +338,15 @@
#if CONFIG_PALETTE
void av1_decode_palette_tokens(MACROBLOCKD *const xd, int plane,
aom_reader *r) {
- MODE_INFO *const mi = xd->mi[0];
- MB_MODE_INFO *const mbmi = &mi->mbmi;
+ const MODE_INFO *const mi = xd->mi[0];
+ const MB_MODE_INFO *const mbmi = &mi->mbmi;
const BLOCK_SIZE bsize = mbmi->sb_type;
const int rows = (4 * num_4x4_blocks_high_lookup[bsize]) >>
(xd->plane[plane != 0].subsampling_y);
const int cols = (4 * num_4x4_blocks_wide_lookup[bsize]) >>
(xd->plane[plane != 0].subsampling_x);
- int color_idx, color_ctx, color_order[PALETTE_MAX_SIZE];
- int n = mbmi->palette_mode_info.palette_size[plane != 0];
+ uint8_t color_order[PALETTE_MAX_SIZE];
+ const int n = mbmi->palette_mode_info.palette_size[plane != 0];
int i, j;
uint8_t *color_map = xd->plane[plane != 0].color_index_map;
const aom_prob(*const prob)[PALETTE_COLOR_CONTEXTS][PALETTE_COLORS - 1] =
@@ -355,10 +355,10 @@
for (i = 0; i < rows; ++i) {
for (j = (i == 0 ? 1 : 0); j < cols; ++j) {
- color_ctx =
- av1_get_palette_color_context(color_map, cols, i, j, n, color_order);
- color_idx = aom_read_tree(r, av1_palette_color_tree[n - 2],
- prob[n - 2][color_ctx], ACCT_STR);
+ const int color_ctx = av1_get_palette_color_context(color_map, cols, i, j,
+ n, color_order, NULL);
+ const int color_idx = aom_read_tree(r, av1_palette_color_tree[n - 2],
+ prob[n - 2][color_ctx], ACCT_STR);
assert(color_idx >= 0 && color_idx < n);
color_map[i * cols + j] = color_order[color_idx];
}
diff --git a/av1/encoder/rdopt.c b/av1/encoder/rdopt.c
index 99da9ee..fd71a1e 100644
--- a/av1/encoder/rdopt.c
+++ b/av1/encoder/rdopt.c
@@ -1768,8 +1768,7 @@
if (colors > 1 && colors <= 64) {
int r, c, i, j, k;
const int max_itr = 50;
- int color_ctx, color_idx = 0;
- int color_order[PALETTE_MAX_SIZE];
+ uint8_t color_order[PALETTE_MAX_SIZE];
float *const data = x->palette_buffer->kmeans_data_buf;
float centroids[PALETTE_MAX_SIZE];
uint8_t *const color_map = xd->plane[0].color_index_map;
@@ -1853,13 +1852,9 @@
1);
for (i = 0; i < rows; ++i) {
for (j = (i == 0 ? 1 : 0); j < cols; ++j) {
- color_ctx = av1_get_palette_color_context(color_map, cols, i, j, k,
- color_order);
- for (r = 0; r < k; ++r)
- if (color_map[i * cols + j] == color_order[r]) {
- color_idx = r;
- break;
- }
+ int color_idx;
+ const int color_ctx = av1_get_palette_color_context(
+ color_map, cols, i, j, k, color_order, &color_idx);
assert(color_idx >= 0 && color_idx < k);
this_rate += cpi->palette_y_color_cost[k - 2][color_ctx][color_idx];
}
@@ -3647,8 +3642,7 @@
if (colors > 1 && colors <= 64) {
int r, c, n, i, j;
const int max_itr = 50;
- int color_ctx, color_idx = 0;
- int color_order[PALETTE_MAX_SIZE];
+ uint8_t color_order[PALETTE_MAX_SIZE];
int64_t this_sse;
float lb_u, ub_u, val_u;
float lb_v, ub_v, val_v;
@@ -3741,13 +3735,9 @@
for (i = 0; i < rows; ++i) {
for (j = (i == 0 ? 1 : 0); j < cols; ++j) {
- color_ctx = av1_get_palette_color_context(color_map, cols, i, j, n,
- color_order);
- for (r = 0; r < n; ++r)
- if (color_map[i * cols + j] == color_order[r]) {
- color_idx = r;
- break;
- }
+ int color_idx;
+ const int color_ctx = av1_get_palette_color_context(
+ color_map, cols, i, j, n, color_order, &color_idx);
assert(color_idx >= 0 && color_idx < n);
this_rate += cpi->palette_uv_color_cost[n - 2][color_ctx][color_idx];
}
@@ -9377,7 +9367,7 @@
int best_rate_nocoef;
#endif
int64_t distortion2 = 0, distortion_y = 0, dummy_rd = best_rd, this_rd;
- int skippable = 0;
+ int skippable = 0, rate_overhead = 0;
TX_SIZE best_tx_size, uv_tx;
TX_TYPE best_tx_type;
PALETTE_MODE_INFO palette_mode_info;
diff --git a/av1/encoder/tokenize.c b/av1/encoder/tokenize.c
index 561beed..86ab57f 100644
--- a/av1/encoder/tokenize.c
+++ b/av1/encoder/tokenize.c
@@ -409,18 +409,19 @@
}
#if CONFIG_PALETTE
-void av1_tokenize_palette_sb(const AV1_COMP *cpi, struct ThreadData *const td,
- int plane, TOKENEXTRA **t, RUN_TYPE dry_run,
- BLOCK_SIZE bsize, int *rate) {
- MACROBLOCK *const x = &td->mb;
- MACROBLOCKD *const xd = &x->e_mbd;
- MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
- uint8_t *color_map = xd->plane[plane != 0].color_index_map;
- PALETTE_MODE_INFO *pmi = &mbmi->palette_mode_info;
- int n = pmi->palette_size[plane != 0];
- int i, j, k;
+void av1_tokenize_palette_sb(const AV1_COMP *cpi,
+ const struct ThreadData *const td, int plane,
+ TOKENEXTRA **t, RUN_TYPE dry_run, BLOCK_SIZE bsize,
+ int *rate) {
+ const MACROBLOCK *const x = &td->mb;
+ const MACROBLOCKD *const xd = &x->e_mbd;
+ const MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+ const uint8_t *const color_map = xd->plane[plane != 0].color_index_map;
+ const PALETTE_MODE_INFO *const pmi = &mbmi->palette_mode_info;
+ const int n = pmi->palette_size[plane != 0];
+ int i, j;
int this_rate = 0;
- int color_idx = -1, color_ctx, color_order[PALETTE_MAX_SIZE];
+ uint8_t color_order[PALETTE_MAX_SIZE];
const int rows = (4 * num_4x4_blocks_high_lookup[bsize]) >>
(xd->plane[plane != 0].subsampling_y);
const int cols = (4 * num_4x4_blocks_wide_lookup[bsize]) >>
@@ -431,17 +432,13 @@
for (i = 0; i < rows; ++i) {
for (j = (i == 0 ? 1 : 0); j < cols; ++j) {
- color_ctx =
- av1_get_palette_color_context(color_map, cols, i, j, n, color_order);
- for (k = 0; k < n; ++k)
- if (color_map[i * cols + j] == color_order[k]) {
- color_idx = k;
- break;
- }
- assert(color_idx >= 0 && color_idx < n);
+ int color_new_idx;
+ const int color_ctx = av1_get_palette_color_context(
+ color_map, cols, i, j, n, color_order, &color_new_idx);
+ assert(color_new_idx >= 0 && color_new_idx < n);
if (dry_run == DRY_RUN_COSTCOEFFS)
- this_rate += cpi->palette_y_color_cost[n - 2][color_ctx][color_idx];
- (*t)->token = color_idx;
+ this_rate += cpi->palette_y_color_cost[n - 2][color_ctx][color_new_idx];
+ (*t)->token = color_new_idx;
(*t)->context_tree = probs[n - 2][color_ctx];
(*t)->skip_eob_node = 0;
++(*t);
diff --git a/av1/encoder/tokenize.h b/av1/encoder/tokenize.h
index ae896a6..89610df 100644
--- a/av1/encoder/tokenize.h
+++ b/av1/encoder/tokenize.h
@@ -72,7 +72,7 @@
#endif
#if CONFIG_PALETTE
void av1_tokenize_palette_sb(const struct AV1_COMP *cpi,
- struct ThreadData *const td, int plane,
+ const struct ThreadData *const td, int plane,
TOKENEXTRA **t, RUN_TYPE dry_run, BLOCK_SIZE bsize,
int *rate);
#endif // CONFIG_PALETTE
