Remove palette interleave
* Run 45 degree wavefront coding for palette index
with palette_throughput experiment.
* Remove palette index interleave.
Change-Id: Ibb57004401f817dec8b00bc2a941d70a26783ff9
diff --git a/av1/encoder/bitstream.c b/av1/encoder/bitstream.c
index b7cd4f4..7ecf9c6 100644
--- a/av1/encoder/bitstream.c
+++ b/av1/encoder/bitstream.c
@@ -799,7 +799,7 @@
aom_bit_depth_t bit_depth, const TX_SIZE tx_size,
TOKEN_STATS *token_stats) {
const TOKENEXTRA *p = *tp;
-#if CONFIG_VAR_TX || CONFIG_PALETTE_THROUGHPUT
+#if CONFIG_VAR_TX
int count = 0;
const int seg_eob = tx_size_2d[tx_size];
#endif
@@ -862,7 +862,7 @@
}
++p;
-#if CONFIG_VAR_TX || CONFIG_PALETTE_THROUGHPUT
+#if CONFIG_VAR_TX
++count;
if (token == EOB_TOKEN || count == seg_eob) break;
#endif
@@ -876,7 +876,7 @@
aom_bit_depth_t bit_depth, const TX_SIZE tx_size,
TOKEN_STATS *token_stats) {
const TOKENEXTRA *p = *tp;
-#if CONFIG_VAR_TX || CONFIG_PALETTE_THROUGHPUT
+#if CONFIG_VAR_TX
int count = 0;
const int seg_eob = tx_size_2d[tx_size];
#endif
@@ -967,7 +967,7 @@
}
++p;
-#if CONFIG_VAR_TX || CONFIG_PALETTE_THROUGHPUT
+#if CONFIG_VAR_TX
++count;
if (token == EOB_TOKEN || count == seg_eob) break;
#endif
@@ -2182,23 +2182,18 @@
#endif
#if CONFIG_PALETTE
-#if CONFIG_PALETTE_THROUGHPUT
- // when block is skipped, palette index is coded here
- // since there is no coeff to be interleaved.
- if (m->mbmi.skip)
-#endif // CONFIG_PALETTE_THROUGHPUT
- for (plane = 0; plane <= 1; ++plane) {
- const uint8_t palette_size_plane =
- m->mbmi.palette_mode_info.palette_size[plane];
- if (palette_size_plane > 0) {
- int rows, cols;
- av1_get_block_dimensions(m->mbmi.sb_type, plane, xd, NULL, NULL, &rows,
- &cols);
- assert(*tok < tok_end);
- pack_palette_tokens(w, tok, palette_size_plane, rows * cols - 1);
- assert(*tok < tok_end + m->mbmi.skip);
- }
+ for (plane = 0; plane <= 1; ++plane) {
+ const uint8_t palette_size_plane =
+ m->mbmi.palette_mode_info.palette_size[plane];
+ if (palette_size_plane > 0) {
+ int rows, cols;
+ av1_get_block_dimensions(m->mbmi.sb_type, plane, xd, NULL, NULL, &rows,
+ &cols);
+ assert(*tok < tok_end);
+ pack_palette_tokens(w, tok, palette_size_plane, rows * cols - 1);
+ assert(*tok < tok_end + m->mbmi.skip);
}
+ }
#endif // CONFIG_PALETTE
#if CONFIG_COEF_INTERLEAVE
@@ -2339,30 +2334,8 @@
TX_SIZE tx = get_tx_size(plane, xd);
const int bkw = tx_size_wide_unit[tx];
const int bkh = tx_size_high_unit[tx];
-#if CONFIG_PALETTE && CONFIG_PALETTE_THROUGHPUT
- const uint8_t palette_size_plane =
- m->mbmi.palette_mode_info.palette_size[plane > 0];
- const int bkw_in_pixel = bkw << tx_size_wide_log2[0];
- const int bkh_in_pixel = bkh << tx_size_wide_log2[0];
- int rows, cols;
- av1_get_block_dimensions(m->mbmi.sb_type, plane, xd, NULL, NULL, &rows,
- &cols);
-#endif // CONFIG_PALETTE && CONFIG_PALETTE_THROUGHPUT
for (row = 0; row < num_4x4_h; row += bkh) {
for (col = 0; col < num_4x4_w; col += bkw) {
-#if CONFIG_PALETTE && CONFIG_PALETTE_THROUGHPUT
- if (palette_size_plane > 0 && plane <= 1) {
- const int col_in_pixel = col << tx_size_wide_log2[0];
- const int row_in_pixel = row << tx_size_high_log2[0];
- const int txbkw = AOMMIN(cols - col_in_pixel, bkw_in_pixel);
- const int txbkh = AOMMIN(rows - row_in_pixel, bkh_in_pixel);
- // first palette index is not coded here but in header instead
- const int num_palette_indexes =
- txbkw * txbkh - ((row == 0 && col == 0) ? 1 : 0);
- pack_palette_tokens(w, tok, palette_size_plane,
- num_palette_indexes);
- }
-#endif // CONFIG_PALETTE && CONFIG_PALETTE_THROUGHPUT
#if !CONFIG_PVQ
pack_mb_tokens(w, tok, tok_end, cm->bit_depth, tx, &token_stats);
#else
@@ -2374,46 +2347,6 @@
#else
TX_SIZE tx = get_tx_size(plane, xd);
TOKEN_STATS token_stats;
-#if CONFIG_PALETTE && CONFIG_PALETTE_THROUGHPUT
- const struct macroblockd_plane *const pd = &xd->plane[plane];
- BLOCK_SIZE bsize = mbmi->sb_type;
-#if CONFIG_CB4X4
- const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, pd);
-#else
- const BLOCK_SIZE plane_bsize =
- get_plane_block_size(AOMMAX(bsize, BLOCK_8X8), pd);
-#endif
- const int num_4x4_w =
- block_size_wide[plane_bsize] >> tx_size_wide_log2[0];
- const int num_4x4_h =
- block_size_high[plane_bsize] >> tx_size_wide_log2[0];
- int row, col;
- const int bkw = tx_size_wide_unit[tx];
- const int bkh = tx_size_high_unit[tx];
- const uint8_t palette_size_plane =
- m->mbmi.palette_mode_info.palette_size[plane > 0];
- const int bkw_in_pixel = bkw << tx_size_wide_log2[0];
- const int bkh_in_pixel = bkh << tx_size_wide_log2[0];
- int rows, cols;
- av1_get_block_dimensions(m->mbmi.sb_type, plane, xd, NULL, NULL, &rows,
- &cols);
- for (row = 0; row < num_4x4_h; row += bkh) {
- for (col = 0; col < num_4x4_w; col += bkw) {
- if (!is_inter_block(mbmi) && palette_size_plane > 0 && plane <= 1) {
- const int col_in_pixel = col << tx_size_wide_log2[0];
- const int row_in_pixel = row << tx_size_high_log2[0];
- const int txbkw = AOMMIN(cols - col_in_pixel, bkw_in_pixel);
- const int txbkh = AOMMIN(rows - row_in_pixel, bkh_in_pixel);
- // first palette index is not coded here but in header instead
- const int num_palette_indexes =
- txbkw * txbkh - ((row == 0 && col == 0) ? 1 : 0);
- pack_palette_tokens(w, tok, palette_size_plane,
- num_palette_indexes);
- }
- pack_mb_tokens(w, tok, tok_end, cm->bit_depth, tx, &token_stats);
- }
- }
-#else
#if !CONFIG_PVQ
init_token_stats(&token_stats);
pack_mb_tokens(w, tok, tok_end, cm->bit_depth, tx, &token_stats);
@@ -2421,7 +2354,6 @@
(void)token_stats;
pack_pvq_tokens(w, x, xd, plane, mbmi->sb_type, tx);
#endif
-#endif // CONFIG_PALETTE && CONFIG_PALETTE_THROUGHPUT
#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)) {
diff --git a/av1/encoder/encodeframe.c b/av1/encoder/encodeframe.c
index 174e796..065ea0e 100644
--- a/av1/encoder/encodeframe.c
+++ b/av1/encoder/encodeframe.c
@@ -5680,11 +5680,9 @@
if (mbmi->palette_mode_info.palette_size[plane] > 0) {
mbmi->palette_mode_info.palette_first_color_idx[plane] =
xd->plane[plane].color_index_map[0];
-// TODO(huisu): this increases the use of token buffer. Needs stretch
-// test to verify.
-#if !CONFIG_PALETTE_THROUGHPUT
+ // TODO(huisu): this increases the use of token buffer. Needs stretch
+ // test to verify.
av1_tokenize_palette_sb(cpi, td, plane, t, dry_run, bsize, rate);
-#endif
}
}
}
diff --git a/av1/encoder/tokenize.c b/av1/encoder/tokenize.c
index 9df6f93..e22cf3a 100644
--- a/av1/encoder/tokenize.c
+++ b/av1/encoder/tokenize.c
@@ -399,7 +399,7 @@
#endif
#endif // !CONFIG_PVQ
-#if CONFIG_PALETTE && !CONFIG_PALETTE_THROUGHPUT
+#if CONFIG_PALETTE
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,
@@ -422,8 +422,15 @@
&cols);
assert(plane == 0 || plane == 1);
+#if CONFIG_PALETTE_THROUGHPUT
+ int k;
+ for (k = 1; k < rows + cols - 1; ++k) {
+ for (j = AOMMIN(k, cols - 1); j >= AOMMAX(0, k - rows + 1); --j) {
+ i = k - j;
+#else
for (i = 0; i < rows; ++i) {
for (j = (i == 0 ? 1 : 0); j < cols; ++j) {
+#endif // CONFIG_PALETTE_THROUGHPUT
int color_new_idx;
const int color_ctx = av1_get_palette_color_index_context(
color_map, plane_block_width, i, j, n, color_order, &color_new_idx);
@@ -441,71 +448,6 @@
}
#endif // CONFIG_PALETTE
-#if CONFIG_PALETTE && CONFIG_PALETTE_THROUGHPUT
-void tokenize_palette_b(int plane, int block, int blk_row, int blk_col,
- BLOCK_SIZE plane_bsize, TX_SIZE tx_size, void *arg) {
- struct tokenize_b_args *const args = arg;
- ThreadData *const td = args->td;
- MACROBLOCK *const x = &td->mb;
- MACROBLOCKD *const xd = &x->e_mbd;
- struct macroblockd_plane *pd = &xd->plane[plane];
- TOKENEXTRA **t = args->tp;
- MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
- const uint8_t *const color_map = xd->plane[plane].color_index_map;
- const PALETTE_MODE_INFO *const pmi = &mbmi->palette_mode_info;
- const int n = pmi->palette_size[plane];
- int i, j;
- uint8_t color_order[PALETTE_MAX_SIZE];
- const aom_prob(
- *const probs)[PALETTE_COLOR_INDEX_CONTEXTS][PALETTE_COLORS - 1] =
- plane == 0 ? av1_default_palette_y_color_index_prob
- : av1_default_palette_uv_color_index_prob;
- int bsize = txsize_to_bsize[tx_size];
- int plane_block_width, plane_block_height, rows, cols;
- int block_width, block_height, tx_block_width, tx_block_height;
- (void)block;
-
- if (n == 0) return;
-
- plane_block_width = block_size_wide[plane_bsize];
- plane_block_height = block_size_high[plane_bsize];
- rows = (xd->mb_to_bottom_edge >= 0)
- ? plane_block_height
- : (xd->mb_to_bottom_edge >> (3 + pd->subsampling_y)) +
- plane_block_height;
- cols = (xd->mb_to_right_edge >= 0)
- ? plane_block_width
- : (xd->mb_to_right_edge >> (3 + pd->subsampling_x)) +
- plane_block_width;
- assert(plane_block_width >= cols);
- assert(plane_block_height >= rows);
- tx_block_width = 1 << tx_size_wide_log2[0];
- tx_block_height = 1 << tx_size_high_log2[0];
- block_width = AOMMIN(cols - blk_col * tx_block_width, block_size_wide[bsize]);
- block_height =
- AOMMIN(rows - blk_row * tx_block_height, block_size_high[bsize]);
-
- assert(plane == 0 || plane == 1);
-
- // run wavefront on the palette map index encoding per transform block
- for (i = ((blk_row == 0 && blk_col == 0) ? 1 : 0);
- i < block_width + block_height - 1; ++i) {
- for (j = AOMMIN(i, block_width - 1); j >= AOMMAX(0, i - block_height + 1);
- --j) {
- int color_new_idx;
- const int color_ctx = av1_get_palette_color_index_context(
- color_map, plane_block_width, blk_row * tx_block_width + (i - j),
- blk_col * tx_block_height + j, n, color_order, &color_new_idx);
- assert(color_new_idx >= 0 && color_new_idx < n);
- (*t)->token = color_new_idx;
- (*t)->context_tree = probs[n - 2][color_ctx];
- (*t)->skip_eob_node = 0;
- ++(*t);
- }
- }
-}
-#endif // CONFIG_PALETTE && CONFIG_PALETTE_THROUGHPUT
-
#if CONFIG_PVQ
static void add_pvq_block(AV1_COMMON *const cm, MACROBLOCK *const x,
PVQ_INFO *pvq) {
@@ -709,16 +651,6 @@
#endif // !CONFIG_PVQ
}
-#if CONFIG_PALETTE && CONFIG_PALETTE_THROUGHPUT
-void tokenize_joint_b(int plane, int block, int blk_row, int blk_col,
- BLOCK_SIZE plane_bsize, TX_SIZE tx_size, void *arg) {
- if (plane < 2)
- tokenize_palette_b(plane, block, blk_row, blk_col, plane_bsize, tx_size,
- arg);
- tokenize_b(plane, block, blk_row, blk_col, plane_bsize, tx_size, arg);
-}
-#endif // CONFIG_PALETTE && CONFIG_PALETTE_THROUGHPUT
-
struct is_skippable_args {
uint16_t *eobs;
int *skippable;
@@ -865,15 +797,6 @@
!segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP);
struct tokenize_b_args arg = { cpi, td, t, 0 };
if (mbmi->skip) {
-#if CONFIG_PALETTE && CONFIG_PALETTE_THROUGHPUT
- if (!dry_run) {
- int plane;
- for (plane = 0; plane < 2; ++plane) {
- av1_foreach_transformed_block_in_plane(xd, bsize, plane,
- tokenize_palette_b, &arg);
- }
- }
-#endif // CONFIG_PALETTE && CONFIG_PALETTE_THROUGHPUT
if (!dry_run) td->counts->skip[ctx][1] += skip_inc;
reset_skip_context(xd, bsize);
return;
@@ -882,12 +805,7 @@
if (!dry_run) {
#if CONFIG_COEF_INTERLEAVE
td->counts->skip[ctx][0] += skip_inc;
-#if CONFIG_PALETTE && CONFIG_PALETTE_THROUGHPUT
- av1_foreach_transformed_block_in_plane(xd, bsize, plane, tokenize_joint_b,
- &arg);
-#else
av1_foreach_transformed_block_interleave(xd, bsize, tokenize_b, &arg);
-#endif // CONFIG_PALETTE && CONFIG_PALETTE_THROUGHPUT
#else
int plane;
@@ -905,13 +823,8 @@
(void)mi_row;
(void)mi_col;
#endif
-#if CONFIG_PALETTE && CONFIG_PALETTE_THROUGHPUT
- av1_foreach_transformed_block_in_plane(xd, bsize, plane, tokenize_joint_b,
- &arg);
-#else
av1_foreach_transformed_block_in_plane(xd, bsize, plane, tokenize_b,
&arg);
-#endif // CONFIG_PALETTE && CONFIG_PALETTE_THROUGHPUT
#if !CONFIG_PVQ
(*t)->token = EOSB_TOKEN;
(*t)++;
diff --git a/av1/encoder/tokenize.h b/av1/encoder/tokenize.h
index dbfb6a8..46d3eff 100644
--- a/av1/encoder/tokenize.h
+++ b/av1/encoder/tokenize.h
@@ -74,15 +74,10 @@
int mi_col, BLOCK_SIZE bsize, int *rate);
#endif
#if CONFIG_PALETTE
-#if CONFIG_PALETTE_THROUGHPUT
-void tokenize_palette_b(int plane, int block, int blk_row, int blk_col,
- BLOCK_SIZE plane_bsize, TX_SIZE tx_size, void *arg);
-#else
void av1_tokenize_palette_sb(const struct AV1_COMP *cpi,
const struct ThreadData *const td, int plane,
TOKENEXTRA **t, RUN_TYPE dry_run, BLOCK_SIZE bsize,
int *rate);
-#endif // CONFIG_PALETTE_THROUGHPUT
#endif // CONFIG_PALETTE
void av1_tokenize_sb(const struct AV1_COMP *cpi, struct ThreadData *td,
TOKENEXTRA **t, RUN_TYPE dry_run, BLOCK_SIZE bsize,
