Google Git
Sign in
aomedia / avm / 4f53aeef236fe6c515d4e0b36f4fa9dce52c298c / . / av1 / encoder / partition_search.c
blob: 94b8eb322183d8ab800ebfbfc34f4aa40f4a1224 [file] [log] [blame]
/*
* Copyright (c) 2021, Alliance for Open Media. All rights reserved
*
* This source code is subject to the terms of the BSD 3-Clause Clear License
* and the Alliance for Open Media Patent License 1.0. If the BSD 3-Clause Clear
* License was not distributed with this source code in the LICENSE file, you
* can obtain it at aomedia.org/license/software-license/bsd-3-c-c/. If the
* Alliance for Open Media Patent License 1.0 was not distributed with this
* source code in the PATENTS file, you can obtain it at
* aomedia.org/license/patent-license/.
*/
#include "aom_ports/system_state.h"
#include "av1/common/blockd.h"
#include "av1/common/enums.h"
#include "av1/common/reconintra.h"
#include "av1/encoder/aq_complexity.h"
#include "av1/encoder/aq_variance.h"
#include "av1/encoder/context_tree.h"
#include "av1/encoder/encoder.h"
#include "av1/encoder/encodeframe.h"
#include "av1/encoder/encodeframe_utils.h"
#include "av1/encoder/encodemv.h"
#include "av1/encoder/motion_search_facade.h"
#include "av1/encoder/partition_search.h"
#include "av1/encoder/reconinter_enc.h"
#include "av1/encoder/tokenize.h"
#if CONFIG_ADAPTIVE_MVD
#include "av1/common/reconinter.h"
#endif // CONFIG_ADAPTIVE_MVD
#include "aom_util/debug_util.h"
#if CONFIG_TUNE_VMAF
#include "av1/encoder/tune_vmaf.h"
#endif
#if CONFIG_NEW_TX_PARTITION
static void update_partition_cdfs_and_counts(MACROBLOCKD *xd, int blk_col,
int blk_row, TX_SIZE max_tx_size,
int allow_update_cdf,
FRAME_COUNTS *counts) {
(void)counts;
MB_MODE_INFO *mbmi = xd->mi[0];
const BLOCK_SIZE bsize = mbmi->sb_type[xd->tree_type == CHROMA_PART];
const int is_inter = is_inter_block(mbmi, xd->tree_type);
const int txb_size_index =
is_inter ? av1_get_txb_size_index(bsize, blk_row, blk_col) : 0;
const int is_rect = is_rect_tx(max_tx_size);
const TX_PARTITION_TYPE partition = mbmi->tx_partition_type[txb_size_index];
const int allow_horz = allow_tx_horz_split(max_tx_size);
const int allow_vert = allow_tx_vert_split(max_tx_size);
const int allow_horz4 = allow_tx_horz4_split(max_tx_size);
const int allow_vert4 = allow_tx_vert4_split(max_tx_size);
if (allow_horz && allow_vert) {
const TX_PARTITION_TYPE split4_partition = get_split4_partition(partition);
const int split4_ctx =
is_inter ? txfm_partition_split4_inter_context(
xd->above_txfm_context + blk_col,
xd->left_txfm_context + blk_row, bsize, max_tx_size)
: get_tx_size_context(xd);
aom_cdf_prob *split4_cdf =
is_inter
? xd->tile_ctx->inter_4way_txfm_partition_cdf[is_rect][split4_ctx]
: xd->tile_ctx->intra_4way_txfm_partition_cdf[is_rect][split4_ctx];
if (allow_update_cdf) {
update_cdf(split4_cdf, split4_partition, 4);
}
#if CONFIG_ENTROPY_STATS
if (is_inter)
++counts
->inter_4way_txfm_partition[is_rect][split4_ctx][split4_partition];
else
++counts
->intra_4way_txfm_partition[is_rect][split4_ctx][split4_partition];
#endif // CONFIG_ENTROPY_STATS
if (((split4_partition == TX_PARTITION_VERT) && allow_vert4) ||
((split4_partition == TX_PARTITION_HORZ) && allow_horz4)) {
const int has_split = (partition == TX_PARTITION_HORZ4) ||
(partition == TX_PARTITION_VERT4);
if (allow_update_cdf) {
aom_cdf_prob *split2_rect_cdf =
is_inter ? xd->tile_ctx->inter_2way_rect_txfm_partition_cdf
: xd->tile_ctx->intra_2way_rect_txfm_partition_cdf;
update_cdf(split2_rect_cdf, has_split, 2);
}
#if CONFIG_ENTROPY_STATS
if (is_inter)
++counts->inter_2way_rect_txfm_partition[has_split];
else
++counts->intra_2way_rect_txfm_partition[has_split];
#endif // CONFIG_ENTROPY_STATS
}
} else if (allow_horz || allow_vert) {
const int has_first_split = partition != TX_PARTITION_NONE;
if (allow_update_cdf) {
aom_cdf_prob *split2_cdf =
is_inter ? xd->tile_ctx->inter_2way_txfm_partition_cdf
: xd->tile_ctx->intra_2way_txfm_partition_cdf;
update_cdf(split2_cdf, has_first_split, 2);
}
#if CONFIG_ENTROPY_STATS
if (is_inter)
++counts->inter_2way_txfm_partition[has_first_split];
else
++counts->intra_2way_txfm_partition[has_first_split];
#endif // CONFIG_ENTROPY_STATS
if (has_first_split && (allow_horz4 || allow_vert4)) {
const int has_second_split = (partition == TX_PARTITION_VERT4) ||
(partition == TX_PARTITION_HORZ4);
if (allow_update_cdf) {
aom_cdf_prob *split2_rect_cdf =
is_inter ? xd->tile_ctx->inter_2way_rect_txfm_partition_cdf
: xd->tile_ctx->intra_2way_rect_txfm_partition_cdf;
update_cdf(split2_rect_cdf, has_second_split, 2);
}
#if CONFIG_ENTROPY_STATS
if (is_inter)
++counts->inter_2way_txfm_partition[has_second_split];
else
++counts->intra_2way_txfm_partition[has_second_split];
#endif // CONFIG_ENTROPY_STATS
}
} else {
assert(!allow_horz && !allow_vert);
assert(partition == PARTITION_NONE);
}
}
#endif // CONFIG_NEW_TX_PARTITION
static void update_txfm_count(MACROBLOCK *x, MACROBLOCKD *xd,
FRAME_COUNTS *counts, TX_SIZE tx_size, int depth,
int blk_row, int blk_col,
uint8_t allow_update_cdf) {
MB_MODE_INFO *mbmi = xd->mi[0];
const BLOCK_SIZE bsize = mbmi->sb_type[xd->tree_type == CHROMA_PART];
const int max_blocks_high = max_block_high(xd, bsize, 0);
const int max_blocks_wide = max_block_wide(xd, bsize, 0);
const int txb_size_index = av1_get_txb_size_index(bsize, blk_row, blk_col);
if (blk_row >= max_blocks_high || blk_col >= max_blocks_wide) return;
assert(tx_size > TX_4X4);
#if CONFIG_NEW_TX_PARTITION
(void)depth;
TX_SIZE sub_txs[MAX_TX_PARTITIONS] = { 0 };
get_tx_partition_sizes(mbmi->tx_partition_type[txb_size_index], tx_size,
sub_txs);
// TODO(sarahparker) This assumes all of the tx sizes in the partition scheme
// are the same size. This will need to be adjusted to deal with the case
// where they can be different.
TX_SIZE this_size = sub_txs[0];
assert(mbmi->inter_tx_size[txb_size_index] == this_size);
if (mbmi->tx_partition_type[txb_size_index] != TX_PARTITION_NONE)
++x->txfm_search_info.txb_split_count;
update_partition_cdfs_and_counts(xd, blk_col, blk_row, tx_size,
allow_update_cdf, counts);
mbmi->tx_size = this_size;
txfm_partition_update(xd->above_txfm_context + blk_col,
xd->left_txfm_context + blk_row, this_size, tx_size);
#else // CONFIG_NEW_TX_PARTITION
int ctx = txfm_partition_context(
xd->above_txfm_context + blk_col, xd->left_txfm_context + blk_row,
mbmi->sb_type[xd->tree_type == CHROMA_PART], tx_size);
const TX_SIZE plane_tx_size = mbmi->inter_tx_size[txb_size_index];
if (depth == MAX_VARTX_DEPTH) {
// Don't add to counts in this case
mbmi->tx_size = tx_size;
txfm_partition_update(xd->above_txfm_context + blk_col,
xd->left_txfm_context + blk_row, tx_size, tx_size);
return;
}
if (tx_size == plane_tx_size) {
#if CONFIG_ENTROPY_STATS
++counts->txfm_partition[ctx][0];
#endif
if (allow_update_cdf)
update_cdf(xd->tile_ctx->txfm_partition_cdf[ctx], 0, 2);
mbmi->tx_size = tx_size;
txfm_partition_update(xd->above_txfm_context + blk_col,
xd->left_txfm_context + blk_row, tx_size, tx_size);
} else {
const TX_SIZE sub_txs = sub_tx_size_map[tx_size];
const int bsw = tx_size_wide_unit[sub_txs];
const int bsh = tx_size_high_unit[sub_txs];
#if CONFIG_ENTROPY_STATS
++counts->txfm_partition[ctx][1];
#endif
if (allow_update_cdf)
update_cdf(xd->tile_ctx->txfm_partition_cdf[ctx], 1, 2);
++x->txfm_search_info.txb_split_count;
if (sub_txs == TX_4X4) {
mbmi->inter_tx_size[txb_size_index] = TX_4X4;
mbmi->tx_size = TX_4X4;
txfm_partition_update(xd->above_txfm_context + blk_col,
xd->left_txfm_context + blk_row, TX_4X4, tx_size);
return;
}
for (int row = 0; row < tx_size_high_unit[tx_size]; row += bsh) {
for (int col = 0; col < tx_size_wide_unit[tx_size]; col += bsw) {
int offsetr = row;
int offsetc = col;
update_txfm_count(x, xd, counts, sub_txs, depth + 1, blk_row + offsetr,
blk_col + offsetc, allow_update_cdf);
}
}
}
#endif // CONFIG_NEW_TX_PARTITION
}
static void tx_partition_count_update(const AV1_COMMON *const cm, MACROBLOCK *x,
BLOCK_SIZE plane_bsize,
FRAME_COUNTS *td_counts,
uint8_t allow_update_cdf) {
MACROBLOCKD *xd = &x->e_mbd;
const int mi_width = mi_size_wide[plane_bsize];
const int mi_height = mi_size_high[plane_bsize];
const TX_SIZE max_tx_size = get_vartx_max_txsize(xd, plane_bsize, 0);
const int bh = tx_size_high_unit[max_tx_size];
const int bw = tx_size_wide_unit[max_tx_size];
xd->above_txfm_context =
cm->above_contexts.txfm[xd->tile.tile_row] + xd->mi_col;
xd->left_txfm_context =
xd->left_txfm_context_buffer + (xd->mi_row & MAX_MIB_MASK);
for (int idy = 0; idy < mi_height; idy += bh) {
for (int idx = 0; idx < mi_width; idx += bw) {
update_txfm_count(x, xd, td_counts, max_tx_size, 0, idy, idx,
allow_update_cdf);
}
}
}
static void set_txfm_context(MACROBLOCKD *xd, TX_SIZE tx_size, int blk_row,
int blk_col) {
MB_MODE_INFO *mbmi = xd->mi[0];
const BLOCK_SIZE bsize = mbmi->sb_type[xd->tree_type == CHROMA_PART];
const int max_blocks_high = max_block_high(xd, bsize, 0);
const int max_blocks_wide = max_block_wide(xd, bsize, 0);
const int txb_size_index = av1_get_txb_size_index(bsize, blk_row, blk_col);
const TX_SIZE plane_tx_size = mbmi->inter_tx_size[txb_size_index];
if (blk_row >= max_blocks_high || blk_col >= max_blocks_wide) return;
if (tx_size == plane_tx_size) {
mbmi->tx_size = tx_size;
txfm_partition_update(xd->above_txfm_context + blk_col,
xd->left_txfm_context + blk_row, tx_size, tx_size);
} else {
#if CONFIG_NEW_TX_PARTITION
TX_SIZE sub_txs[MAX_TX_PARTITIONS] = { 0 };
const int index = av1_get_txb_size_index(bsize, blk_row, blk_col);
get_tx_partition_sizes(mbmi->tx_partition_type[index], tx_size, sub_txs);
int cur_partition = 0;
int bsw = 0, bsh = 0;
for (int r = 0; r < tx_size_high_unit[tx_size]; r += bsh) {
for (int c = 0; c < tx_size_wide_unit[tx_size]; c += bsw) {
const TX_SIZE sub_tx = sub_txs[cur_partition];
bsw = tx_size_wide_unit[sub_tx];
bsh = tx_size_high_unit[sub_tx];
const int offsetr = blk_row + r;
const int offsetc = blk_col + c;
if (offsetr >= max_blocks_high || offsetc >= max_blocks_wide) continue;
mbmi->tx_size = sub_tx;
txfm_partition_update(xd->above_txfm_context + blk_col,
xd->left_txfm_context + blk_row, sub_tx, sub_tx);
cur_partition++;
}
}
#else
if (tx_size == TX_8X8) {
mbmi->inter_tx_size[txb_size_index] = TX_4X4;
mbmi->tx_size = TX_4X4;
txfm_partition_update(xd->above_txfm_context + blk_col,
xd->left_txfm_context + blk_row, TX_4X4, tx_size);
return;
}
const TX_SIZE sub_txs = sub_tx_size_map[tx_size];
const int bsw = tx_size_wide_unit[sub_txs];
const int bsh = tx_size_high_unit[sub_txs];
for (int row = 0; row < tx_size_high_unit[tx_size]; row += bsh) {
for (int col = 0; col < tx_size_wide_unit[tx_size]; col += bsw) {
const int offsetr = blk_row + row;
const int offsetc = blk_col + col;
if (offsetr >= max_blocks_high || offsetc >= max_blocks_wide) continue;
set_txfm_context(xd, sub_txs, offsetr, offsetc);
}
}
#endif // CONFIG_NEW_TX_PARTITION
}
}
static void tx_partition_set_contexts(const AV1_COMMON *const cm,
MACROBLOCKD *xd, BLOCK_SIZE plane_bsize) {
const int mi_width = mi_size_wide[plane_bsize];
const int mi_height = mi_size_high[plane_bsize];
const TX_SIZE max_tx_size = get_vartx_max_txsize(xd, plane_bsize, 0);
const int bh = tx_size_high_unit[max_tx_size];
const int bw = tx_size_wide_unit[max_tx_size];
xd->above_txfm_context =
cm->above_contexts.txfm[xd->tile.tile_row] + xd->mi_col;
xd->left_txfm_context =
xd->left_txfm_context_buffer + (xd->mi_row & MAX_MIB_MASK);
for (int idy = 0; idy < mi_height; idy += bh) {
for (int idx = 0; idx < mi_width; idx += bw) {
set_txfm_context(xd, max_tx_size, idy, idx);
}
}
}
static void encode_superblock(const AV1_COMP *const cpi, TileDataEnc *tile_data,
ThreadData *td, TokenExtra **t, RUN_TYPE dry_run,
BLOCK_SIZE bsize, int plane_start, int plane_end,
int *rate) {
const AV1_COMMON *const cm = &cpi->common;
const int num_planes = av1_num_planes(cm);
MACROBLOCK *const x = &td->mb;
MACROBLOCKD *const xd = &x->e_mbd;
MB_MODE_INFO **mi_4x4 = xd->mi;
MB_MODE_INFO *mbmi = mi_4x4[0];
const int seg_skip =
segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP);
const int mis = cm->mi_params.mi_stride;
const int mi_width = mi_size_wide[bsize];
const int mi_height = mi_size_high[bsize];
const int is_inter = is_inter_block(mbmi, xd->tree_type);
// Initialize tx_mode and tx_size_search_method
TxfmSearchParams *txfm_params = &x->txfm_search_params;
set_tx_size_search_method(
cm, &cpi->winner_mode_params, txfm_params,
cpi->sf.winner_mode_sf.enable_winner_mode_for_tx_size_srch, 1);
const int mi_row = xd->mi_row;
const int mi_col = xd->mi_col;
if (!is_inter) {
if (xd->tree_type != LUMA_PART) {
xd->cfl.store_y = store_cfl_required(cm, xd);
}
mbmi->skip_txfm[xd->tree_type == CHROMA_PART] = 1;
for (int plane = plane_start; plane < plane_end; ++plane) {
av1_encode_intra_block_plane(cpi, x, bsize, plane, dry_run,
cpi->optimize_seg_arr[mbmi->segment_id]);
}
// If there is at least one lossless segment, force the skip for intra
// block to be 0, in order to avoid the segment_id to be changed by in
// write_segment_id().
if (!cpi->common.seg.segid_preskip && cpi->common.seg.update_map &&
cpi->enc_seg.has_lossless_segment)
mbmi->skip_txfm[xd->tree_type == CHROMA_PART] = 0;
xd->cfl.store_y = 0;
if (av1_allow_palette(cm->features.allow_screen_content_tools, bsize)) {
for (int plane = plane_start; plane < AOMMIN(2, plane_end); ++plane) {
if (mbmi->palette_mode_info.palette_size[plane] > 0) {
if (!dry_run) {
av1_tokenize_color_map(x, plane, t, bsize, mbmi->tx_size,
PALETTE_MAP, tile_data->allow_update_cdf,
td->counts);
} else if (dry_run == DRY_RUN_COSTCOEFFS) {
rate +=
av1_cost_color_map(x, plane, bsize, mbmi->tx_size, PALETTE_MAP);
}
}
}
}
av1_update_intra_mb_txb_context(cpi, td, dry_run, bsize,
tile_data->allow_update_cdf);
} else {
int ref;
const int is_compound = has_second_ref(mbmi);
set_ref_ptrs(cm, xd, mbmi->ref_frame[0], mbmi->ref_frame[1]);
for (ref = 0; ref < 1 + is_compound; ++ref) {
const YV12_BUFFER_CONFIG *cfg =
get_ref_frame_yv12_buf(cm, mbmi->ref_frame[ref]);
assert(IMPLIES(!is_intrabc_block(mbmi, xd->tree_type), cfg));
av1_setup_pre_planes(xd, ref, cfg, mi_row, mi_col,
xd->block_ref_scale_factors[ref], num_planes);
}
int start_plane = 0;
av1_enc_build_inter_predictor(cm, xd, mi_row, mi_col, NULL, bsize,
start_plane, av1_num_planes(cm) - 1);
if (mbmi->motion_mode == OBMC_CAUSAL) {
assert(cpi->oxcf.motion_mode_cfg.enable_obmc);
av1_build_obmc_inter_predictors_sb(cm, xd);
}
#if CONFIG_MISMATCH_DEBUG
if (dry_run == OUTPUT_ENABLED) {
for (int plane = plane_start; plane < plane_end; ++plane) {
const struct macroblockd_plane *pd = &xd->plane[plane];
int pixel_c, pixel_r;
mi_to_pixel_loc(&pixel_c, &pixel_r, mi_col, mi_row, 0, 0,
pd->subsampling_x, pd->subsampling_y);
if (!is_chroma_reference(mi_row, mi_col, bsize, pd->subsampling_x,
pd->subsampling_y))
continue;
mismatch_record_block_pre(pd->dst.buf, pd->dst.stride,
cm->current_frame.order_hint, plane, pixel_c,
pixel_r, pd->width, pd->height,
xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH);
}
}
#else
(void)num_planes;
#endif
av1_encode_sb(cpi, x, bsize, dry_run, plane_start, plane_end);
av1_tokenize_sb_vartx(cpi, td, dry_run, bsize, rate,
tile_data->allow_update_cdf, plane_start, plane_end);
}
if (!dry_run) {
if (av1_allow_intrabc(cm) && is_intrabc_block(mbmi, xd->tree_type))
td->intrabc_used = 1;
if (txfm_params->tx_mode_search_type == TX_MODE_SELECT &&
!xd->lossless[mbmi->segment_id] &&
mbmi->sb_type[xd->tree_type == CHROMA_PART] > BLOCK_4X4 &&
!(is_inter &&
(mbmi->skip_txfm[xd->tree_type == CHROMA_PART] || seg_skip))) {
if (is_inter) {
tx_partition_count_update(cm, x, bsize, td->counts,
tile_data->allow_update_cdf);
} else {
if (mbmi->tx_size != max_txsize_rect_lookup[bsize] &&
xd->tree_type != CHROMA_PART)
++x->txfm_search_info.txb_split_count;
if (block_signals_txsize(bsize) && xd->tree_type != CHROMA_PART) {
#if CONFIG_NEW_TX_PARTITION
const TX_SIZE max_tx_size = max_txsize_rect_lookup[bsize];
update_partition_cdfs_and_counts(
xd, 0, 0, max_tx_size, tile_data->allow_update_cdf, td->counts);
#else // CONFIG_NEW_TX_PARTITION
const int tx_size_ctx = get_tx_size_context(xd);
const int32_t tx_size_cat = bsize_to_tx_size_cat(bsize);
const int depth = tx_size_to_depth(mbmi->tx_size, bsize);
const int max_depths = bsize_to_max_depth(bsize);
if (tile_data->allow_update_cdf)
update_cdf(xd->tile_ctx->tx_size_cdf[tx_size_cat][tx_size_ctx],
depth, max_depths + 1);
#if CONFIG_ENTROPY_STATS
++td->counts->intra_tx_size[tx_size_cat][tx_size_ctx][depth];
#endif
#endif // CONFIG_NEW_TX_PARTITION
}
}
if (xd->tree_type != CHROMA_PART)
assert(
IMPLIES(is_rect_tx(mbmi->tx_size), is_rect_tx_allowed(xd, mbmi)));
} else {
int i, j;
TX_SIZE intra_tx_size;
// The new intra coding scheme requires no change of transform size
if (is_inter) {
if (xd->lossless[mbmi->segment_id]) {
intra_tx_size = TX_4X4;
} else {
intra_tx_size =
tx_size_from_tx_mode(bsize, txfm_params->tx_mode_search_type);
}
} else {
intra_tx_size = mbmi->tx_size;
}
for (j = 0; j < mi_height; j++)
for (i = 0; i < mi_width; i++)
if (mi_col + i < cm->mi_params.mi_cols &&
mi_row + j < cm->mi_params.mi_rows)
mi_4x4[mis * j + i]->tx_size = intra_tx_size;
if (intra_tx_size != max_txsize_rect_lookup[bsize])
++x->txfm_search_info.txb_split_count;
}
#if CONFIG_REF_MV_BANK
#if CONFIG_IBC_SR_EXT && !CONFIG_BVP_IMPROVEMENT
if (cm->seq_params.enable_refmvbank && is_inter &&
!is_intrabc_block(mbmi, xd->tree_type))
#else
if (cm->seq_params.enable_refmvbank && is_inter)
#endif // CONFIG_IBC_SR_EXT && !CONFIG_BVP_IMPROVEMENT
av1_update_ref_mv_bank(cm, xd, mbmi);
#endif // CONFIG_REF_MV_BANK
}
if (txfm_params->tx_mode_search_type == TX_MODE_SELECT &&
block_signals_txsize(mbmi->sb_type[xd->tree_type == CHROMA_PART]) &&
is_inter &&
!(mbmi->skip_txfm[xd->tree_type == CHROMA_PART] || seg_skip) &&
!xd->lossless[mbmi->segment_id]) {
if (dry_run) tx_partition_set_contexts(cm, xd, bsize);
} else {
TX_SIZE tx_size = mbmi->tx_size;
// The new intra coding scheme requires no change of transform size
if (is_inter) {
if (xd->lossless[mbmi->segment_id]) {
tx_size = TX_4X4;
} else {
tx_size = tx_size_from_tx_mode(bsize, txfm_params->tx_mode_search_type);
}
} else {
tx_size = (bsize > BLOCK_4X4) ? tx_size : TX_4X4;
}
mbmi->tx_size = tx_size;
set_txfm_ctxs(tx_size, xd->width, xd->height,
(mbmi->skip_txfm[xd->tree_type == CHROMA_PART] || seg_skip) &&
is_inter_block(mbmi, xd->tree_type),
xd);
}
if (is_inter_block(mbmi, xd->tree_type) && !xd->is_chroma_ref &&
is_cfl_allowed(xd)) {
cfl_store_block(xd, mbmi->sb_type[xd->tree_type == CHROMA_PART],
mbmi->tx_size);
}
if (xd->tree_type == LUMA_PART) {
const CommonModeInfoParams *const mi_params = &cm->mi_params;
for (int y = 0; y < mi_height; y++) {
for (int x_idx = 0; x_idx < mi_width; x_idx++) {
if ((xd->mb_to_right_edge >> (3 + MI_SIZE_LOG2)) + mi_width > x_idx &&
(xd->mb_to_bottom_edge >> (3 + MI_SIZE_LOG2)) + mi_height > y) {
if (y == 0 && x_idx == 0) continue;
const int mi_idx =
get_alloc_mi_idx(mi_params, mi_row + y, mi_col + x_idx);
xd->mi[x_idx + y * mis] = &mi_params->mi_alloc[mi_idx];
xd->mi[x_idx + y * mis]->skip_txfm[PLANE_TYPE_Y] =
xd->mi[0]->skip_txfm[PLANE_TYPE_Y];
}
}
}
}
#if CONFIG_IBC_SR_EXT
av1_mark_block_as_coded(xd, mi_row, mi_col, bsize, cm->seq_params.sb_size);
#endif // CONFIG_IBC_SR_EXT
}
static void setup_block_rdmult(const AV1_COMP *const cpi, MACROBLOCK *const x,
int mi_row, int mi_col, BLOCK_SIZE bsize,
AQ_MODE aq_mode, MB_MODE_INFO *mbmi) {
x->rdmult = cpi->rd.RDMULT;
MACROBLOCKD *const xd = &x->e_mbd;
if (aq_mode != NO_AQ && xd->tree_type == SHARED_PART) {
assert(mbmi != NULL);
if (aq_mode == VARIANCE_AQ) {
if (cpi->vaq_refresh) {
const int energy = bsize <= BLOCK_16X16
? x->mb_energy
: av1_log_block_var(cpi, x, bsize);
mbmi->segment_id = energy;
}
x->rdmult = set_segment_rdmult(cpi, x, mbmi->segment_id);
} else if (aq_mode == COMPLEXITY_AQ) {
x->rdmult = set_segment_rdmult(cpi, x, mbmi->segment_id);
} else if (aq_mode == CYCLIC_REFRESH_AQ) {
// If segment is boosted, use rdmult for that segment.
if (cyclic_refresh_segment_id_boosted(mbmi->segment_id))
x->rdmult = av1_cyclic_refresh_get_rdmult(cpi->cyclic_refresh);
}
}
const AV1_COMMON *const cm = &cpi->common;
if (cm->delta_q_info.delta_q_present_flag) {
x->rdmult =
av1_get_hier_tpl_rdmult(cpi, x, bsize, mi_row, mi_col, x->rdmult);
}
if (cpi->oxcf.tune_cfg.tuning == AOM_TUNE_SSIM) {
av1_set_ssim_rdmult(cpi, &x->mv_costs, bsize, mi_row, mi_col, &x->rdmult);
}
#if CONFIG_TUNE_VMAF
if (cpi->oxcf.tune_cfg.tuning == AOM_TUNE_VMAF_WITHOUT_PREPROCESSING ||
cpi->oxcf.tune_cfg.tuning == AOM_TUNE_VMAF_MAX_GAIN ||
cpi->oxcf.tune_cfg.tuning == AOM_TUNE_VMAF_NEG_MAX_GAIN) {
av1_set_vmaf_rdmult(cpi, x, bsize, mi_row, mi_col, &x->rdmult);
}
#endif
}
void av1_set_offsets_without_segment_id(const AV1_COMP *const cpi,
const TileInfo *const tile,
MACROBLOCK *const x, int mi_row,
int mi_col, BLOCK_SIZE bsize) {
const AV1_COMMON *const cm = &cpi->common;
const int num_planes = av1_num_planes(cm);
MACROBLOCKD *const xd = &x->e_mbd;
assert(bsize < BLOCK_SIZES_ALL);
const int mi_width = mi_size_wide[bsize];
const int mi_height = mi_size_high[bsize];
set_mode_info_offsets(&cpi->common.mi_params, &cpi->mbmi_ext_info, x, xd,
mi_row, mi_col);
set_entropy_context(xd, mi_row, mi_col, num_planes);
xd->above_txfm_context = cm->above_contexts.txfm[tile->tile_row] + mi_col;
xd->left_txfm_context =
xd->left_txfm_context_buffer + (mi_row & MAX_MIB_MASK);
// Set up destination pointers.
av1_setup_dst_planes(xd->plane, bsize, &cm->cur_frame->buf, mi_row, mi_col, 0,
num_planes);
// Set up limit values for MV components.
// Mv beyond the range do not produce new/different prediction block.
av1_set_mv_limits(&cm->mi_params, &x->mv_limits, mi_row, mi_col, mi_height,
mi_width, cpi->oxcf.border_in_pixels);
set_plane_n4(xd, mi_width, mi_height, num_planes);
// Set up distance of MB to edge of frame in 1/8th pel units.
assert(!(mi_col & (mi_width - 1)) && !(mi_row & (mi_height - 1)));
set_mi_row_col(xd, tile, mi_row, mi_height, mi_col, mi_width,
cm->mi_params.mi_rows, cm->mi_params.mi_cols);
// Set up source buffers.
av1_setup_src_planes(x, cpi->source, mi_row, mi_col, num_planes, bsize);
// required by av1_append_sub8x8_mvs_for_idx() and av1_find_best_ref_mvs()
xd->tile = *tile;
}
void av1_set_offsets(const AV1_COMP *const cpi, const TileInfo *const tile,
MACROBLOCK *const x, int mi_row, int mi_col,
BLOCK_SIZE bsize) {
const AV1_COMMON *const cm = &cpi->common;
const struct segmentation *const seg = &cm->seg;
MACROBLOCKD *const xd = &x->e_mbd;
MB_MODE_INFO *mbmi;
av1_set_offsets_without_segment_id(cpi, tile, x, mi_row, mi_col, bsize);
// Setup segment ID.
mbmi = xd->mi[0];
mbmi->segment_id = 0;
if (seg->enabled) {
if (seg->enabled && !cpi->vaq_refresh) {
const uint8_t *const map =
seg->update_map ? cpi->enc_seg.map : cm->last_frame_seg_map;
mbmi->segment_id =
map ? get_segment_id(&cm->mi_params, map, bsize, mi_row, mi_col) : 0;
}
av1_init_plane_quantizers(cpi, x, mbmi->segment_id);
}
}
/*!\brief Interface for AV1 mode search for an individual coding block
*
* \ingroup partition_search
* \callgraph
* \callergraph
* Searches prediction modes, transform, and coefficient coding modes for an
* individual coding block. This function is the top-level interface that
* directs the encoder to the proper mode search function, among these
* implemented for inter/intra + rd/non-rd + non-skip segment/skip segment.
*
* \param[in] cpi Top-level encoder structure
* \param[in] tile_data Pointer to struct holding adaptive
* data/contexts/models for the tile during
* encoding
* \param[in] x Pointer to structure holding all the data for
* the current macroblock
* \param[in] mi_row Row coordinate of the block in a step size of
* MI_SIZE
* \param[in] mi_col Column coordinate of the block in a step size of
* MI_SIZE
* \param[in] rd_cost Pointer to structure holding rate and distortion
* stats for the current block
* \param[in] partition Partition mode of the parent block
* \param[in] bsize Current block size
* \param[in] ctx Pointer to structure holding coding contexts and
* chosen modes for the current block
* \param[in] best_rd Upper bound of rd cost of a valid partition
*
* Nothing is returned. Instead, the chosen modes and contexts necessary
* for reconstruction are stored in ctx, the rate-distortion stats are stored in
* rd_cost. If no valid mode leading to rd_cost <= best_rd, the status will be
* signalled by an INT64_MAX rd_cost->rdcost.
*/
static void pick_sb_modes(AV1_COMP *const cpi, TileDataEnc *tile_data,
MACROBLOCK *const x, int mi_row, int mi_col,
RD_STATS *rd_cost, PARTITION_TYPE partition,
BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx,
RD_STATS best_rd) {
if (best_rd.rdcost < 0) {
ctx->rd_stats.rdcost = INT64_MAX;
ctx->rd_stats.skip_txfm = 0;
av1_invalid_rd_stats(rd_cost);
return;
}
AV1_COMMON *const cm = &cpi->common;
const int num_planes = av1_num_planes(cm);
MACROBLOCKD *const xd = &x->e_mbd;
int plane_type = (xd->tree_type == CHROMA_PART);
av1_set_offsets(cpi, &tile_data->tile_info, x, mi_row, mi_col, bsize);
if (ctx->rd_mode_is_ready) {
assert(ctx->mic.sb_type[plane_type] == bsize);
assert(ctx->mic.partition == partition);
rd_cost->rate = ctx->rd_stats.rate;
rd_cost->dist = ctx->rd_stats.dist;
rd_cost->rdcost = ctx->rd_stats.rdcost;
return;
}
MB_MODE_INFO *mbmi;
struct macroblock_plane *const p = x->plane;
struct macroblockd_plane *const pd = xd->plane;
const AQ_MODE aq_mode = cpi->oxcf.q_cfg.aq_mode;
TxfmSearchInfo *txfm_info = &x->txfm_search_info;
int i;
#if CONFIG_COLLECT_COMPONENT_TIMING
start_timing(cpi, rd_pick_sb_modes_time);
#endif
aom_clear_system_state();
mbmi = xd->mi[0];
mbmi->sb_type[plane_type] = bsize;
if (xd->tree_type == SHARED_PART) mbmi->sb_type[PLANE_TYPE_UV] = bsize;
mbmi->partition = partition;
#if CONFIG_RD_DEBUG
mbmi->mi_row = mi_row;
mbmi->mi_col = mi_col;
#endif
// Sets up the tx_type_map buffer in MACROBLOCKD.
xd->tx_type_map = txfm_info->tx_type_map_;
xd->tx_type_map_stride = mi_size_wide[bsize];
for (i = 0; i < num_planes; ++i) {
p[i].coeff = ctx->coeff[i];
p[i].qcoeff = ctx->qcoeff[i];
p[i].dqcoeff = ctx->dqcoeff[i];
p[i].eobs = ctx->eobs[i];
p[i].txb_entropy_ctx = ctx->txb_entropy_ctx[i];
}
for (i = 0; i < 2; ++i) pd[i].color_index_map = ctx->color_index_map[i];
ctx->skippable = 0;
// Set to zero to make sure we do not use the previous encoded frame stats
mbmi->skip_txfm[xd->tree_type == CHROMA_PART] = 0;
// Reset skip mode flag.
mbmi->skip_mode = 0;
if (is_cur_buf_hbd(xd)) {
x->source_variance = av1_high_get_sby_perpixel_variance(
cpi, &x->plane[0].src, bsize, xd->bd);
} else {
x->source_variance =
av1_get_sby_perpixel_variance(cpi, &x->plane[0].src, bsize);
}
// Initialize default mode evaluation params
set_mode_eval_params(cpi, x, DEFAULT_EVAL);
// Save rdmult before it might be changed, so it can be restored later.
const int orig_rdmult = x->rdmult;
setup_block_rdmult(cpi, x, mi_row, mi_col, bsize, aq_mode, mbmi);
// Set error per bit for current rdmult
av1_set_error_per_bit(&x->mv_costs, x->rdmult);
av1_rd_cost_update(x->rdmult, &best_rd);
// Find best coding mode & reconstruct the MB so it is available
// as a predictor for MBs that follow in the SB
if (frame_is_intra_only(cm)) {
#if CONFIG_COLLECT_COMPONENT_TIMING
start_timing(cpi, av1_rd_pick_intra_mode_sb_time);
#endif
av1_rd_pick_intra_mode_sb(cpi, x, rd_cost, bsize, ctx, best_rd.rdcost);
#if CONFIG_COLLECT_COMPONENT_TIMING
end_timing(cpi, av1_rd_pick_intra_mode_sb_time);
#endif
} else {
#if CONFIG_COLLECT_COMPONENT_TIMING
start_timing(cpi, av1_rd_pick_inter_mode_sb_time);
#endif
if (segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP)) {
av1_rd_pick_inter_mode_sb_seg_skip(cpi, tile_data, x, mi_row, mi_col,
rd_cost, bsize, ctx, best_rd.rdcost);
} else {
av1_rd_pick_inter_mode_sb(cpi, tile_data, x, rd_cost, bsize, ctx,
best_rd.rdcost);
}
#if CONFIG_COLLECT_COMPONENT_TIMING
end_timing(cpi, av1_rd_pick_inter_mode_sb_time);
#endif
}
// Examine the resulting rate and for AQ mode 2 make a segment choice.
if (rd_cost->rate != INT_MAX && aq_mode == COMPLEXITY_AQ &&
bsize >= BLOCK_16X16) {
av1_caq_select_segment(cpi, x, bsize, mi_row, mi_col, rd_cost->rate);
}
x->rdmult = orig_rdmult;
// TODO(jingning) The rate-distortion optimization flow needs to be
// refactored to provide proper exit/return handle.
if (rd_cost->rate == INT_MAX) rd_cost->rdcost = INT64_MAX;
ctx->rd_stats.rate = rd_cost->rate;
ctx->rd_stats.dist = rd_cost->dist;
ctx->rd_stats.rdcost = rd_cost->rdcost;
#if CONFIG_COLLECT_COMPONENT_TIMING
end_timing(cpi, rd_pick_sb_modes_time);
#endif
}
static void update_drl_index_stats(int max_drl_bits, const int16_t mode_ctx,
FRAME_CONTEXT *fc, FRAME_COUNTS *counts,
const MB_MODE_INFO *mbmi,
const MB_MODE_INFO_EXT *mbmi_ext) {
#if !CONFIG_ENTROPY_STATS
(void)counts;
#endif // !CONFIG_ENTROPY_STATS
assert(have_drl_index(mbmi->mode));
#if IMPROVED_AMVD
if (mbmi->mode == AMVDNEWMV) max_drl_bits = AOMMIN(max_drl_bits, 1);
#endif // IMPROVED_AMVD
uint8_t ref_frame_type = av1_ref_frame_type(mbmi->ref_frame);
assert(mbmi->ref_mv_idx < max_drl_bits + 1);
for (int idx = 0; idx < max_drl_bits; ++idx) {
aom_cdf_prob *drl_cdf =
av1_get_drl_cdf(fc, mbmi_ext->weight[ref_frame_type], mode_ctx, idx);
#if CONFIG_ENTROPY_STATS
int drl_ctx = av1_drl_ctx(mode_ctx);
switch (idx) {
case 0: counts->drl_mode[0][drl_ctx][mbmi->ref_mv_idx != idx]++; break;
case 1: counts->drl_mode[1][drl_ctx][mbmi->ref_mv_idx != idx]++; break;
default: counts->drl_mode[2][drl_ctx][mbmi->ref_mv_idx != idx]++; break;
}
#endif // CONFIG_ENTROPY_STATS
update_cdf(drl_cdf, mbmi->ref_mv_idx != idx, 2);
if (mbmi->ref_mv_idx == idx) break;
}
}
#if CONFIG_BVP_IMPROVEMENT
static void update_intrabc_drl_idx_stats(int max_ref_bv_num, FRAME_CONTEXT *fc,
FRAME_COUNTS *counts,
const MB_MODE_INFO *mbmi) {
#if !CONFIG_ENTROPY_STATS
(void)counts;
#endif // !CONFIG_ENTROPY_STATS
assert(mbmi->intrabc_drl_idx < max_ref_bv_num);
int bit_cnt = 0;
for (int idx = 0; idx < max_ref_bv_num - 1; ++idx) {
#if CONFIG_ENTROPY_STATS
counts->intrabc_drl_idx[bit_cnt][mbmi->intrabc_drl_idx != idx]++;
#endif // CONFIG_ENTROPY_STATS
update_cdf(fc->intrabc_drl_idx_cdf[bit_cnt], mbmi->intrabc_drl_idx != idx,
2);
if (mbmi->intrabc_drl_idx == idx) break;
++bit_cnt;
}
}
#endif // CONFIG_BVP_IMPROVEMENT
static void update_stats(const AV1_COMMON *const cm, ThreadData *td) {
MACROBLOCK *x = &td->mb;
MACROBLOCKD *const xd = &x->e_mbd;
const MB_MODE_INFO *const mbmi = xd->mi[0];
const MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
const CurrentFrame *const current_frame = &cm->current_frame;
const BLOCK_SIZE bsize = mbmi->sb_type[xd->tree_type == CHROMA_PART];
FRAME_CONTEXT *fc = xd->tile_ctx;
#if CONFIG_NEW_REF_SIGNALING
const int seg_ref_active = 0;
#else
const int seg_ref_active =
segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_REF_FRAME);
#endif // CONFIG_NEW_REF_SIGNALING
if (current_frame->skip_mode_info.skip_mode_flag && !seg_ref_active &&
is_comp_ref_allowed(bsize)) {
const int skip_mode_ctx = av1_get_skip_mode_context(xd);
#if CONFIG_ENTROPY_STATS
td->counts->skip_mode[skip_mode_ctx][mbmi->skip_mode]++;
#endif
update_cdf(fc->skip_mode_cdfs[skip_mode_ctx], mbmi->skip_mode, 2);
}
#if CONFIG_SKIP_MODE_ENHANCEMENT
if (!seg_ref_active) {
#else
if (!mbmi->skip_mode && !seg_ref_active) {
#endif // CONFIG_SKIP_MODE_ENHANCEMENT
const int skip_ctx = av1_get_skip_txfm_context(xd);
#if CONFIG_ENTROPY_STATS
td->counts
->skip_txfm[skip_ctx][mbmi->skip_txfm[xd->tree_type == CHROMA_PART]]++;
#endif
update_cdf(fc->skip_txfm_cdfs[skip_ctx],
mbmi->skip_txfm[xd->tree_type == CHROMA_PART], 2);
}
#if CONFIG_ENTROPY_STATS
// delta quant applies to both intra and inter
const int super_block_upper_left =
((xd->mi_row & (cm->seq_params.mib_size - 1)) == 0) &&
((xd->mi_col & (cm->seq_params.mib_size - 1)) == 0);
const DeltaQInfo *const delta_q_info = &cm->delta_q_info;
if (delta_q_info->delta_q_present_flag &&
(bsize != cm->seq_params.sb_size || !mbmi->skip_txfm) &&
super_block_upper_left) {
const int dq = (mbmi->current_qindex - xd->current_base_qindex) /
delta_q_info->delta_q_res;
const int absdq = abs(dq);
for (int i = 0; i < AOMMIN(absdq, DELTA_Q_SMALL); ++i) {
td->counts->delta_q[i][1]++;
}
if (absdq < DELTA_Q_SMALL) td->counts->delta_q[absdq][0]++;
if (delta_q_info->delta_lf_present_flag) {
if (delta_q_info->delta_lf_multi) {
const int frame_lf_count =
av1_num_planes(cm) > 1 ? FRAME_LF_COUNT : FRAME_LF_COUNT - 2;
for (int lf_id = 0; lf_id < frame_lf_count; ++lf_id) {
const int delta_lf = (mbmi->delta_lf[lf_id] - xd->delta_lf[lf_id]) /
delta_q_info->delta_lf_res;
const int abs_delta_lf = abs(delta_lf);
for (int i = 0; i < AOMMIN(abs_delta_lf, DELTA_LF_SMALL); ++i) {
td->counts->delta_lf_multi[lf_id][i][1]++;
}
if (abs_delta_lf < DELTA_LF_SMALL)
td->counts->delta_lf_multi[lf_id][abs_delta_lf][0]++;
}
} else {
const int delta_lf =
(mbmi->delta_lf_from_base - xd->delta_lf_from_base) /
delta_q_info->delta_lf_res;
const int abs_delta_lf = abs(delta_lf);
for (int i = 0; i < AOMMIN(abs_delta_lf, DELTA_LF_SMALL); ++i) {
td->counts->delta_lf[i][1]++;
}
if (abs_delta_lf < DELTA_LF_SMALL)
td->counts->delta_lf[abs_delta_lf][0]++;
}
}
}
#endif
if (!is_inter_block(mbmi, xd->tree_type)) {
av1_sum_intra_stats(cm, td->counts, xd, mbmi
#if !CONFIG_AIMC
,
xd->above_mbmi, xd->left_mbmi, frame_is_intra_only(cm)
#endif // !CONFIG_AIMC
);
}
if (av1_allow_intrabc(cm) && xd->tree_type != CHROMA_PART) {
update_cdf(fc->intrabc_cdf, is_intrabc_block(mbmi, xd->tree_type), 2);
#if CONFIG_BVCOST_UPDATE
if (is_intrabc_block(mbmi, xd->tree_type)) {
const int_mv ref_mv = mbmi_ext->ref_mv_stack[INTRA_FRAME][0].this_mv;
av1_update_mv_stats(&mbmi->mv[0].as_mv, &ref_mv.as_mv, &fc->ndvc,
#if CONFIG_ADAPTIVE_MVD
0,
#endif // CONFIG_ADAPTIVE_MVD
MV_SUBPEL_NONE);
}
#endif // CONFIG_BVCOST_UPDATE
#if CONFIG_BVP_IMPROVEMENT
if (is_intrabc_block(mbmi, xd->tree_type)) {
update_cdf(fc->intrabc_mode_cdf, mbmi->intrabc_mode, 2);
#if CONFIG_ENTROPY_STATS
++td->counts->intrabc_mode[mbmi->intrabc_mode];
#endif // CONFIG_ENTROPY_STATS
update_intrabc_drl_idx_stats(MAX_REF_BV_STACK_SIZE, fc, td->counts, mbmi);
}
#endif // CONFIG_BVP_IMPROVEMENT
#if CONFIG_ENTROPY_STATS
++td->counts->intrabc[is_intrabc_block(mbmi, xd->tree_type)];
#endif // CONFIG_ENTROPY_STATS
}
#if CONFIG_SKIP_MODE_ENHANCEMENT
if (mbmi->skip_mode && have_drl_index(mbmi->mode)) {
FRAME_COUNTS *const counts = td->counts;
const int16_t mode_ctx_pristine =
av1_mode_context_pristine(mbmi_ext->mode_context, mbmi->ref_frame);
update_drl_index_stats(cm->features.max_drl_bits, mode_ctx_pristine, fc,
counts, mbmi, mbmi_ext);
}
#endif // CONFIG_SKIP_MODE_ENHANCEMENT
if (frame_is_intra_only(cm) || mbmi->skip_mode) return;
FRAME_COUNTS *const counts = td->counts;
const int inter_block = is_inter_block(mbmi, xd->tree_type);
if (!seg_ref_active) {
#if CONFIG_ENTROPY_STATS && !CONFIG_CONTEXT_DERIVATION
counts->intra_inter[av1_get_intra_inter_context(xd)][inter_block]++;
#endif // CONFIG_ENTROPY_STATS && !CONFIG_CONTEXT_DERIVATION
#if CONFIG_CONTEXT_DERIVATION
const int skip_txfm = mbmi->skip_txfm[xd->tree_type == CHROMA_PART];
#if CONFIG_ENTROPY_STATS
counts->intra_inter[skip_txfm][av1_get_intra_inter_context(xd)]
[inter_block]++;
#endif // CONFIG_ENTROPY_STATS
update_cdf(fc->intra_inter_cdf[skip_txfm][av1_get_intra_inter_context(xd)],
inter_block, 2);
#else
update_cdf(fc->intra_inter_cdf[av1_get_intra_inter_context(xd)],
inter_block, 2);
#endif // CONFIG_CONTEXT_DERIVATION
// If the segment reference feature is enabled we have only a single
// reference frame allowed for the segment so exclude it from
// the reference frame counts used to work out probabilities.
if (inter_block) {
const MV_REFERENCE_FRAME ref0 = mbmi->ref_frame[0];
const MV_REFERENCE_FRAME ref1 = mbmi->ref_frame[1];
#if CONFIG_TIP
if (cm->features.tip_frame_mode && is_tip_allowed_bsize(bsize)) {
const int tip_ctx = get_tip_ctx(xd);
update_cdf(fc->tip_cdf[tip_ctx], is_tip_ref_frame(ref0), 2);
#if CONFIG_ENTROPY_STATS
++counts->tip_ref[tip_ctx][is_tip_ref_frame(ref0)];
#endif
}
#endif // CONFIG_TIP
if (current_frame->reference_mode == REFERENCE_MODE_SELECT
#if CONFIG_TIP
&& !is_tip_ref_frame(ref0)
#endif // CONFIG_TIP
) {
if (is_comp_ref_allowed(bsize)) {
#if CONFIG_ENTROPY_STATS
counts->comp_inter[av1_get_reference_mode_context(cm, xd)]
[has_second_ref(mbmi)]++;
#endif // CONFIG_ENTROPY_STATS
update_cdf(av1_get_reference_mode_cdf(cm, xd), has_second_ref(mbmi),
2);
}
}
if (has_second_ref(mbmi)) {
#if CONFIG_NEW_REF_SIGNALING
const int n_refs = cm->ref_frames_info.num_total_refs;
int n_bits = 0;
assert(ref0 < ref1);
for (int i = 0; i < n_refs + n_bits - 2 && n_bits < 2; i++) {
const int bit = ref0 == i || ref1 == i;
const int bit_type = n_bits == 0 ? -1
: av1_get_compound_ref_bit_type(
&cm->ref_frames_info, ref0, i);
if (n_bits > 0 || i < RANKED_REF0_TO_PRUNE - 1)
update_cdf(
av1_get_pred_cdf_compound_ref(xd, i, n_bits, bit_type, n_refs),
bit, 2);
#if CONFIG_ENTROPY_STATS
if (n_bits == 0) {
if (i < RANKED_REF0_TO_PRUNE - 1)
counts->comp_ref0[av1_get_ref_pred_context(xd, i, n_refs)][i]
[bit]++;
} else {
counts->comp_ref1[av1_get_ref_pred_context(xd, i, n_refs)][bit_type]
[i - 1][bit]++;
}
#endif // CONFIG_ENTROPY_STATS
n_bits += bit;
}
#else
const COMP_REFERENCE_TYPE comp_ref_type = has_uni_comp_refs(mbmi)
? UNIDIR_COMP_REFERENCE
: BIDIR_COMP_REFERENCE;
update_cdf(av1_get_comp_reference_type_cdf(xd), comp_ref_type,
COMP_REFERENCE_TYPES);
#if CONFIG_ENTROPY_STATS
counts->comp_ref_type[av1_get_comp_reference_type_context(xd)]
[comp_ref_type]++;
#endif // CONFIG_ENTROPY_STATS
if (comp_ref_type == UNIDIR_COMP_REFERENCE) {
const int bit = (ref0 == BWDREF_FRAME);
update_cdf(av1_get_pred_cdf_uni_comp_ref_p(xd), bit, 2);
#if CONFIG_ENTROPY_STATS
counts
->uni_comp_ref[av1_get_pred_context_uni_comp_ref_p(xd)][0][bit]++;
#endif // CONFIG_ENTROPY_STATS
if (!bit) {
const int bit1 = (ref1 == LAST3_FRAME || ref1 == GOLDEN_FRAME);
update_cdf(av1_get_pred_cdf_uni_comp_ref_p1(xd), bit1, 2);
#if CONFIG_ENTROPY_STATS
counts->uni_comp_ref[av1_get_pred_context_uni_comp_ref_p1(xd)][1]
[bit1]++;
#endif // CONFIG_ENTROPY_STATS
if (bit1) {
update_cdf(av1_get_pred_cdf_uni_comp_ref_p2(xd),
ref1 == GOLDEN_FRAME, 2);
#if CONFIG_ENTROPY_STATS
counts->uni_comp_ref[av1_get_pred_context_uni_comp_ref_p2(xd)][2]
[ref1 == GOLDEN_FRAME]++;
#endif // CONFIG_ENTROPY_STATS
}
}
} else {
const int bit = (ref0 == GOLDEN_FRAME || ref0 == LAST3_FRAME);
update_cdf(av1_get_pred_cdf_comp_ref_p(xd), bit, 2);
#if CONFIG_ENTROPY_STATS
counts->comp_ref[av1_get_pred_context_comp_ref_p(xd)][0][bit]++;
#endif // CONFIG_ENTROPY_STATS
if (!bit) {
update_cdf(av1_get_pred_cdf_comp_ref_p1(xd), ref0 == LAST2_FRAME,
2);
#if CONFIG_ENTROPY_STATS
counts->comp_ref[av1_get_pred_context_comp_ref_p1(xd)][1]
[ref0 == LAST2_FRAME]++;
#endif // CONFIG_ENTROPY_STATS
} else {
update_cdf(av1_get_pred_cdf_comp_ref_p2(xd), ref0 == GOLDEN_FRAME,
2);
#if CONFIG_ENTROPY_STATS
counts->comp_ref[av1_get_pred_context_comp_ref_p2(xd)][2]
[ref0 == GOLDEN_FRAME]++;
#endif // CONFIG_ENTROPY_STATS
}
update_cdf(av1_get_pred_cdf_comp_bwdref_p(xd), ref1 == ALTREF_FRAME,
2);
#if CONFIG_ENTROPY_STATS
counts->comp_bwdref[av1_get_pred_context_comp_bwdref_p(xd)][0]
[ref1 == ALTREF_FRAME]++;
#endif // CONFIG_ENTROPY_STATS
if (ref1 != ALTREF_FRAME) {
update_cdf(av1_get_pred_cdf_comp_bwdref_p1(xd),
ref1 == ALTREF2_FRAME, 2);
#if CONFIG_ENTROPY_STATS
counts->comp_bwdref[av1_get_pred_context_comp_bwdref_p1(xd)][1]
[ref1 == ALTREF2_FRAME]++;
#endif // CONFIG_ENTROPY_STATS
}
}
#endif // CONFIG_NEW_REF_SIGNALING
#if CONFIG_TIP
} else if (!is_tip_ref_frame(ref0)) {
#else
} else {
#endif // CONFIG_TIP
#if CONFIG_NEW_REF_SIGNALING
const int n_refs = cm->ref_frames_info.num_total_refs;
const MV_REFERENCE_FRAME ref0_nrs = mbmi->ref_frame[0];
for (int i = 0; i < n_refs - 1; i++) {
const int bit = ref0_nrs == i;
update_cdf(av1_get_pred_cdf_single_ref(xd, i, n_refs), bit, 2);
#if CONFIG_ENTROPY_STATS
counts->single_ref[av1_get_ref_pred_context(xd, i, n_refs)][i][bit]++;
#endif // CONFIG_ENTROPY_STATS
if (bit) break;
}
#else
const int bit = (ref0 >= BWDREF_FRAME);
update_cdf(av1_get_pred_cdf_single_ref_p1(xd), bit, 2);
#if CONFIG_ENTROPY_STATS
counts->single_ref[av1_get_pred_context_single_ref_p1(xd)][0][bit]++;
#endif // CONFIG_ENTROPY_STATS
if (bit) {
assert(ref0 <= ALTREF_FRAME);
update_cdf(av1_get_pred_cdf_single_ref_p2(xd), ref0 == ALTREF_FRAME,
2);
#if CONFIG_ENTROPY_STATS
counts->single_ref[av1_get_pred_context_single_ref_p2(xd)][1]
[ref0 == ALTREF_FRAME]++;
#endif // CONFIG_ENTROPY_STATS
if (ref0 != ALTREF_FRAME) {
update_cdf(av1_get_pred_cdf_single_ref_p6(xd),
ref0 == ALTREF2_FRAME, 2);
#if CONFIG_ENTROPY_STATS
counts->single_ref[av1_get_pred_context_single_ref_p6(xd)][5]
[ref0 == ALTREF2_FRAME]++;
#endif // CONFIG_ENTROPY_STATS
}
} else {
const int bit1 = !(ref0 == LAST2_FRAME || ref0 == LAST_FRAME);
update_cdf(av1_get_pred_cdf_single_ref_p3(xd), bit1, 2);
#if CONFIG_ENTROPY_STATS
counts->single_ref[av1_get_pred_context_single_ref_p3(xd)][2][bit1]++;
#endif // CONFIG_ENTROPY_STATS
if (!bit1) {
update_cdf(av1_get_pred_cdf_single_ref_p4(xd), ref0 != LAST_FRAME,
2);
#if CONFIG_ENTROPY_STATS
counts->single_ref[av1_get_pred_context_single_ref_p4(xd)][3]
[ref0 != LAST_FRAME]++;
#endif // CONFIG_ENTROPY_STATS
} else {
update_cdf(av1_get_pred_cdf_single_ref_p5(xd), ref0 != LAST3_FRAME,
2);
#if CONFIG_ENTROPY_STATS
counts->single_ref[av1_get_pred_context_single_ref_p5(xd)][4]
[ref0 != LAST3_FRAME]++;
#endif // CONFIG_ENTROPY_STATS
}
}
#endif // CONFIG_NEW_REF_SIGNALING
}
if (cm->seq_params.enable_interintra_compound &&
is_interintra_allowed(mbmi)) {
const int bsize_group = size_group_lookup[bsize];
if (mbmi->ref_frame[1] == INTRA_FRAME) {
#if CONFIG_ENTROPY_STATS
counts->interintra[bsize_group][1]++;
#endif
update_cdf(fc->interintra_cdf[bsize_group], 1, 2);
#if CONFIG_ENTROPY_STATS
counts->interintra_mode[bsize_group][mbmi->interintra_mode]++;
#endif
update_cdf(fc->interintra_mode_cdf[bsize_group],
mbmi->interintra_mode, INTERINTRA_MODES);
if (av1_is_wedge_used(bsize)) {
#if CONFIG_ENTROPY_STATS
counts->wedge_interintra[bsize][mbmi->use_wedge_interintra]++;
#endif
update_cdf(fc->wedge_interintra_cdf[bsize],
mbmi->use_wedge_interintra, 2);
if (mbmi->use_wedge_interintra) {
#if CONFIG_ENTROPY_STATS
counts->wedge_idx[bsize][mbmi->interintra_wedge_index]++;
#endif
update_cdf(fc->wedge_idx_cdf[bsize], mbmi->interintra_wedge_index,
16);
}
}
} else {
#if CONFIG_ENTROPY_STATS
counts->interintra[bsize_group][0]++;
#endif
update_cdf(fc->interintra_cdf[bsize_group], 0, 2);
}
}
const MOTION_MODE motion_allowed =
cm->features.switchable_motion_mode
? motion_mode_allowed(xd->global_motion, xd, mbmi,
cm->features.allow_warped_motion)
: SIMPLE_TRANSLATION;
if (mbmi->ref_frame[1] != INTRA_FRAME) {
if (motion_allowed == WARPED_CAUSAL) {
#if CONFIG_ENTROPY_STATS
counts->motion_mode[bsize][mbmi->motion_mode]++;
#endif
update_cdf(fc->motion_mode_cdf[bsize], mbmi->motion_mode,
MOTION_MODES);
} else if (motion_allowed == OBMC_CAUSAL) {
#if CONFIG_ENTROPY_STATS
counts->obmc[bsize][mbmi->motion_mode == OBMC_CAUSAL]++;
#endif
update_cdf(fc->obmc_cdf[bsize], mbmi->motion_mode == OBMC_CAUSAL, 2);
}
}
if (has_second_ref(mbmi)
#if CONFIG_OPTFLOW_REFINEMENT
&& mbmi->mode < NEAR_NEARMV_OPTFLOW
#endif // CONFIG_OPTFLOW_REFINEMENT
#if IMPROVED_AMVD && CONFIG_JOINT_MVD
&& !is_joint_amvd_coding_mode(mbmi->adaptive_mvd_flag)
#endif // IMPROVED_AMVD && CONFIG_JOINT_MVD
) {
assert(current_frame->reference_mode != SINGLE_REFERENCE &&
is_inter_compound_mode(mbmi->mode) &&
mbmi->motion_mode == SIMPLE_TRANSLATION);
const int masked_compound_used = is_any_masked_compound_used(bsize) &&
cm->seq_params.enable_masked_compound;
if (masked_compound_used) {
const int comp_group_idx_ctx = get_comp_group_idx_context(cm, xd);
#if CONFIG_ENTROPY_STATS
++counts->comp_group_idx[comp_group_idx_ctx][mbmi->comp_group_idx];
#endif
update_cdf(fc->comp_group_idx_cdf[comp_group_idx_ctx],
mbmi->comp_group_idx, 2);
}
if (mbmi->comp_group_idx == 1) {
assert(masked_compound_used);
if (is_interinter_compound_used(COMPOUND_WEDGE, bsize)) {
#if CONFIG_ENTROPY_STATS
++counts->compound_type[bsize][mbmi->interinter_comp.type -
COMPOUND_WEDGE];
#endif
update_cdf(fc->compound_type_cdf[bsize],
mbmi->interinter_comp.type - COMPOUND_WEDGE,
MASKED_COMPOUND_TYPES);
}
}
}
if (mbmi->interinter_comp.type == COMPOUND_WEDGE) {
if (is_interinter_compound_used(COMPOUND_WEDGE, bsize)) {
#if CONFIG_ENTROPY_STATS
counts->wedge_idx[bsize][mbmi->interinter_comp.wedge_index]++;
#endif
update_cdf(fc->wedge_idx_cdf[bsize],
mbmi->interinter_comp.wedge_index, 16);
}
}
}
}
if (inter_block && cm->features.interp_filter == SWITCHABLE &&
mbmi->motion_mode != WARPED_CAUSAL &&
!is_nontrans_global_motion(xd, mbmi)) {
update_filter_type_cdf(xd, mbmi);
}
if (inter_block &&
!segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP)) {
const PREDICTION_MODE mode = mbmi->mode;
const int16_t mode_ctx =
av1_mode_context_analyzer(mbmi_ext->mode_context, mbmi->ref_frame);
if (has_second_ref(mbmi)) {
#if CONFIG_OPTFLOW_REFINEMENT
if (cm->features.opfl_refine_type == REFINE_SWITCHABLE &&
is_opfl_refine_allowed(cm, mbmi)) {
const int use_optical_flow = mode >= NEAR_NEARMV_OPTFLOW;
#if CONFIG_ENTROPY_STATS
++counts->use_optflow[mode_ctx][use_optical_flow];
#endif
update_cdf(fc->use_optflow_cdf[mode_ctx], use_optical_flow, 2);
}
const int comp_mode_idx = opfl_get_comp_idx(mode);
#if CONFIG_ENTROPY_STATS
++counts->inter_compound_mode[mode_ctx][comp_mode_idx];
#endif
update_cdf(fc->inter_compound_mode_cdf[mode_ctx], comp_mode_idx,
INTER_COMPOUND_REF_TYPES);
#else
#if CONFIG_ENTROPY_STATS
++counts->inter_compound_mode[mode_ctx][INTER_COMPOUND_OFFSET(mode)];
#endif
update_cdf(fc->inter_compound_mode_cdf[mode_ctx],
INTER_COMPOUND_OFFSET(mode), INTER_COMPOUND_MODES);
#endif // CONFIG_OPTFLOW_REFINEMENT
#if IMPROVED_AMVD && CONFIG_JOINT_MVD
if (is_joint_mvd_coding_mode(mbmi->mode) &&
cm->seq_params.enable_adaptive_mvd)
update_cdf(fc->adaptive_mvd_cdf, mbmi->adaptive_mvd_flag, 2);
#endif // IMPROVED_AMVD && CONFIG_JOINT_MVD
} else {
av1_update_inter_mode_stats(fc, counts, mode, mode_ctx);
}
const int new_mv = have_newmv_in_each_reference(mbmi->mode);
#if CONFIG_JOINT_MVD
const int jmvd_base_ref_list = get_joint_mvd_base_ref_list(cm, mbmi);
#endif // CONFIG_JOINT_MVD
#if CONFIG_ADAPTIVE_MVD
const int is_adaptive_mvd = enable_adaptive_mvd_resolution(cm, mbmi);
#endif // CONFIG_ADAPTIVE_MVD
if (have_drl_index(mbmi->mode)) {
const int16_t mode_ctx_pristine =
av1_mode_context_pristine(mbmi_ext->mode_context, mbmi->ref_frame);
update_drl_index_stats(cm->features.max_drl_bits, mode_ctx_pristine, fc,
counts, mbmi, mbmi_ext);
}
if (have_newmv_in_inter_mode(mbmi->mode) && xd->tree_type != CHROMA_PART) {
const int allow_hp = cm->features.cur_frame_force_integer_mv
? MV_SUBPEL_NONE
: cm->features.allow_high_precision_mv;
if (new_mv) {
for (int ref = 0; ref < 1 + has_second_ref(mbmi); ++ref) {
const int_mv ref_mv = av1_get_ref_mv(x, ref);
av1_update_mv_stats(&mbmi->mv[ref].as_mv, &ref_mv.as_mv, &fc->nmvc,
#if CONFIG_ADAPTIVE_MVD
is_adaptive_mvd,
#endif // CONFIG_ADAPTIVE_MVD
allow_hp);
}
} else if (have_nearmv_newmv_in_inter_mode(mbmi->mode)) {
const int ref =
#if CONFIG_OPTFLOW_REFINEMENT
mbmi->mode == NEAR_NEWMV_OPTFLOW ||
#endif // CONFIG_OPTFLOW_REFINEMENT
#if CONFIG_JOINT_MVD
jmvd_base_ref_list ||
#endif // CONFIG_JOINT_MVD
mbmi->mode == NEAR_NEWMV;
const int_mv ref_mv = av1_get_ref_mv(x, ref);
av1_update_mv_stats(&mbmi->mv[ref].as_mv, &ref_mv.as_mv, &fc->nmvc,
#if CONFIG_ADAPTIVE_MVD
is_adaptive_mvd,
#endif // CONFIG_ADAPTIVE_MVD
allow_hp);
}
}
}
}
/*!\brief Reconstructs an individual coding block
*
* \ingroup partition_search
* Reconstructs an individual coding block by applying the chosen modes stored
* in ctx, also updates mode counts and entropy models.
*
* This function works on planes determined by get_partition_plane_start() and
* get_partition_plane_end() based on xd->tree_type.
*
* \param[in] cpi Top-level encoder structure
* \param[in] tile_data Pointer to struct holding adaptive
* data/contexts/models for the tile during encoding
* \param[in] td Pointer to thread data
* \param[in] tp Pointer to the starting token
* \param[in] mi_row Row coordinate of the block in a step size of MI_SIZE
* \param[in] mi_col Column coordinate of the block in a step size of
* MI_SIZE
* \param[in] dry_run A code indicating whether it is part of the final
* pass for reconstructing the superblock
* \param[in] bsize Current block size
* \param[in] partition Partition mode of the parent block
* \param[in] ctx Pointer to structure holding coding contexts and the
* chosen modes for the current block
* \param[in] rate Pointer to the total rate for the current block
*
* Nothing is returned. Instead, reconstructions (w/o in-loop filters)
* will be updated in the pixel buffers in td->mb.e_mbd. Also, the chosen modes
* will be stored in the MB_MODE_INFO buffer td->mb.e_mbd.mi[0].
*/
static void encode_b(const AV1_COMP *const cpi, TileDataEnc *tile_data,
ThreadData *td, TokenExtra **tp, int mi_row, int mi_col,
RUN_TYPE dry_run, BLOCK_SIZE bsize,
PARTITION_TYPE partition, PICK_MODE_CONTEXT *const ctx,
int *rate) {
TileInfo *const tile = &tile_data->tile_info;
MACROBLOCK *const x = &td->mb;
MACROBLOCKD *xd = &x->e_mbd;
const AV1_COMMON *const cm = &cpi->common;
av1_set_offsets_without_segment_id(cpi, tile, x, mi_row, mi_col, bsize);
const int origin_mult = x->rdmult;
setup_block_rdmult(cpi, x, mi_row, mi_col, bsize, NO_AQ, NULL);
MB_MODE_INFO *mbmi = xd->mi[0];
mbmi->partition = partition;
av1_update_state(cpi, td, ctx, mi_row, mi_col, bsize, dry_run);
int plane_type = (xd->tree_type == CHROMA_PART);
const int plane_start = get_partition_plane_start(xd->tree_type);
const int plane_end =
get_partition_plane_end(xd->tree_type, av1_num_planes(cm));
if (!dry_run) {
x->mbmi_ext_frame->cb_offset[plane_type] = x->cb_offset[plane_type];
assert(x->cb_offset[plane_type] <
(1 << num_pels_log2_lookup[cpi->common.seq_params.sb_size]));
}
encode_superblock(cpi, tile_data, td, tp, dry_run, bsize, plane_start,
plane_end, rate);
if (!dry_run) {
x->cb_offset[plane_type] += block_size_wide[bsize] * block_size_high[bsize];
if (bsize == cpi->common.seq_params.sb_size &&
mbmi->skip_txfm[xd->tree_type == CHROMA_PART] == 1 &&
cm->delta_q_info.delta_lf_present_flag) {
const int frame_lf_count =
av1_num_planes(cm) > 1 ? FRAME_LF_COUNT : FRAME_LF_COUNT - 2;
for (int lf_id = 0; lf_id < frame_lf_count; ++lf_id)
mbmi->delta_lf[lf_id] = xd->delta_lf[lf_id];
mbmi->delta_lf_from_base = xd->delta_lf_from_base;
}
if (has_second_ref(mbmi)) {
if (mbmi->interinter_comp.type == COMPOUND_AVERAGE)
mbmi->comp_group_idx = 0;
else
mbmi->comp_group_idx = 1;
}
// delta quant applies to both intra and inter
const int super_block_upper_left =
((mi_row & (cm->seq_params.mib_size - 1)) == 0) &&
((mi_col & (cm->seq_params.mib_size - 1)) == 0);
const DeltaQInfo *const delta_q_info = &cm->delta_q_info;
if (delta_q_info->delta_q_present_flag &&
(bsize != cm->seq_params.sb_size ||
!mbmi->skip_txfm[xd->tree_type == CHROMA_PART]) &&
super_block_upper_left) {
xd->current_base_qindex = mbmi->current_qindex;
if (delta_q_info->delta_lf_present_flag) {
if (delta_q_info->delta_lf_multi) {
const int frame_lf_count =
av1_num_planes(cm) > 1 ? FRAME_LF_COUNT : FRAME_LF_COUNT - 2;
for (int lf_id = 0; lf_id < frame_lf_count; ++lf_id) {
xd->delta_lf[lf_id] = mbmi->delta_lf[lf_id];
}
} else {
xd->delta_lf_from_base = mbmi->delta_lf_from_base;
}
}
}
RD_COUNTS *rdc = &td->rd_counts;
if (mbmi->skip_mode) {
assert(!frame_is_intra_only(cm));
rdc->skip_mode_used_flag = 1;
if (cm->current_frame.reference_mode == REFERENCE_MODE_SELECT) {
assert(has_second_ref(mbmi));
rdc->compound_ref_used_flag = 1;
}
set_ref_ptrs(cm, xd, mbmi->ref_frame[0], mbmi->ref_frame[1]);
} else {
#if CONFIG_NEW_REF_SIGNALING
const int seg_ref_active = 0;
#else
const int seg_ref_active =
segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_REF_FRAME);
#endif // CONFIG_NEW_REF_SIGNALING
if (!seg_ref_active) {
// If the segment reference feature is enabled we have only a single
// reference frame allowed for the segment so exclude it from
// the reference frame counts used to work out probabilities.
if (is_inter_block(mbmi, xd->tree_type)) {
av1_collect_neighbors_ref_counts(xd);
if (cm->current_frame.reference_mode == REFERENCE_MODE_SELECT) {
if (has_second_ref(mbmi)) {
// This flag is also updated for 4x4 blocks
rdc->compound_ref_used_flag = 1;
}
}
set_ref_ptrs(cm, xd, mbmi->ref_frame[0], mbmi->ref_frame[1]);
}
}
}
if (tile_data->allow_update_cdf) update_stats(&cpi->common, td);
// Gather obmc and warped motion count to update the probability.
if ((!cpi->sf.inter_sf.disable_obmc &&
cpi->sf.inter_sf.prune_obmc_prob_thresh > 0) ||
(cm->features.allow_warped_motion &&
cpi->sf.inter_sf.prune_warped_prob_thresh > 0)) {
const int inter_block = is_inter_block(mbmi, xd->tree_type);
#if CONFIG_NEW_REF_SIGNALING
const int seg_ref_active = 0;
#else
const int seg_ref_active =
segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_REF_FRAME);
#endif // CONFIG_NEW_REF_SIGNALING
if (!seg_ref_active && inter_block) {
const MOTION_MODE motion_allowed =
cm->features.switchable_motion_mode
? motion_mode_allowed(xd->global_motion, xd, mbmi,
cm->features.allow_warped_motion)
: SIMPLE_TRANSLATION;
if (mbmi->ref_frame[1] != INTRA_FRAME) {
if (motion_allowed >= OBMC_CAUSAL) {
td->rd_counts.obmc_used[bsize][mbmi->motion_mode == OBMC_CAUSAL]++;
}
if (motion_allowed == WARPED_CAUSAL) {
td->rd_counts.warped_used[mbmi->motion_mode == WARPED_CAUSAL]++;
}
}
}
}
}
// TODO(Ravi/Remya): Move this copy function to a better logical place
// This function will copy the best mode information from block
// level (x->mbmi_ext) to frame level (cpi->mbmi_ext_info.frame_base). This
// frame level buffer (cpi->mbmi_ext_info.frame_base) will be used during
// bitstream preparation.
if (xd->tree_type != CHROMA_PART)
av1_copy_mbmi_ext_to_mbmi_ext_frame(
x->mbmi_ext_frame, x->mbmi_ext,
av1_ref_frame_type(xd->mi[0]->ref_frame));
x->rdmult = origin_mult;
}
/*!\brief Reconstructs a partition (may contain multiple coding blocks)
*
* \ingroup partition_search
* Reconstructs a sub-partition of the superblock by applying the chosen modes
* and partition trees stored in pc_tree.
*
* This function works on planes determined by get_partition_plane_start() and
* get_partition_plane_end() based on xd->tree_type.
*
* \param[in] cpi Top-level encoder structure
* \param[in] td Pointer to thread data
* \param[in] tile_data Pointer to struct holding adaptive
* data/contexts/models for the tile during encoding
* \param[in] tp Pointer to the starting token
* \param[in] mi_row Row coordinate of the block in a step size of MI_SIZE
* \param[in] mi_col Column coordinate of the block in a step size of
* MI_SIZE
* \param[in] dry_run A code indicating whether it is part of the final
* pass for reconstructing the superblock
* \param[in] bsize Current block size
* \param[in] pc_tree Pointer to the PC_TREE node storing the picked
* partitions and mode info for the current block
* \param[in] rate Pointer to the total rate for the current block
*
* Nothing is returned. Instead, reconstructions (w/o in-loop filters)
* will be updated in the pixel buffers in td->mb.e_mbd.
*/
static void encode_sb(const AV1_COMP *const cpi, ThreadData *td,
TileDataEnc *tile_data, TokenExtra **tp, int mi_row,
int mi_col, RUN_TYPE dry_run, BLOCK_SIZE bsize,
PC_TREE *pc_tree, int *rate) {
assert(bsize < BLOCK_SIZES_ALL);
const AV1_COMMON *const cm = &cpi->common;
const CommonModeInfoParams *const mi_params = &cm->mi_params;
MACROBLOCK *const x = &td->mb;
MACROBLOCKD *const xd = &x->e_mbd;
assert(bsize < BLOCK_SIZES_ALL);
const int hbs = mi_size_wide[bsize] / 2;
const int is_partition_root = bsize >= BLOCK_8X8;
const int ctx = is_partition_root
? partition_plane_context(xd, mi_row, mi_col, bsize)
: -1;
const PARTITION_TYPE partition = pc_tree->partitioning;
const BLOCK_SIZE subsize = get_partition_subsize(bsize, partition);
int quarter_step = mi_size_wide[bsize] / 4;
int i;
BLOCK_SIZE bsize2 = get_partition_subsize(bsize, PARTITION_SPLIT);
if (mi_row >= mi_params->mi_rows || mi_col >= mi_params->mi_cols) return;
if (subsize == BLOCK_INVALID) return;
if (!dry_run && ctx >= 0) {
const int has_rows = (mi_row + hbs) < mi_params->mi_rows;
const int has_cols = (mi_col + hbs) < mi_params->mi_cols;
const int plane_index = xd->tree_type == CHROMA_PART;
if (has_rows && has_cols) {
#if CONFIG_ENTROPY_STATS
td->counts->partition[ctx][partition]++;
#endif
if (tile_data->allow_update_cdf) {
FRAME_CONTEXT *fc = xd->tile_ctx;
int luma_split_flag = 0;
int parent_block_width = block_size_wide[bsize];
if (xd->tree_type == CHROMA_PART &&
parent_block_width >= SHARED_PART_SIZE) {
luma_split_flag =
get_luma_split_flag(bsize, mi_params, mi_row, mi_col);
}
if (luma_split_flag <= 3) {
update_cdf(fc->partition_cdf[plane_index][ctx], partition,
partition_cdf_length(bsize));
} else {
// if luma blocks uses smaller blocks, then chroma will also split
assert(partition == PARTITION_SPLIT);
}
}
}
}
switch (partition) {
case PARTITION_NONE:
encode_b(cpi, tile_data, td, tp, mi_row, mi_col, dry_run, subsize,
partition, pc_tree->none, rate);
break;
case PARTITION_VERT:
encode_b(cpi, tile_data, td, tp, mi_row, mi_col, dry_run, subsize,
partition, pc_tree->vertical[0], rate);
if (mi_col + hbs < mi_params->mi_cols) {
encode_b(cpi, tile_data, td, tp, mi_row, mi_col + hbs, dry_run, subsize,
partition, pc_tree->vertical[1], rate);
}
break;
case PARTITION_HORZ:
encode_b(cpi, tile_data, td, tp, mi_row, mi_col, dry_run, subsize,
partition, pc_tree->horizontal[0], rate);
if (mi_row + hbs < mi_params->mi_rows) {
encode_b(cpi, tile_data, td, tp, mi_row + hbs, mi_col, dry_run, subsize,
partition, pc_tree->horizontal[1], rate);
}
break;
case PARTITION_SPLIT:
encode_sb(cpi, td, tile_data, tp, mi_row, mi_col, dry_run, subsize,
pc_tree->split[0], rate);
encode_sb(cpi, td, tile_data, tp, mi_row, mi_col + hbs, dry_run, subsize,
pc_tree->split[1], rate);
encode_sb(cpi, td, tile_data, tp, mi_row + hbs, mi_col, dry_run, subsize,
pc_tree->split[2], rate);
encode_sb(cpi, td, tile_data, tp, mi_row + hbs, mi_col + hbs, dry_run,
subsize, pc_tree->split[3], rate);
break;
case PARTITION_HORZ_A:
encode_b(cpi, tile_data, td, tp, mi_row, mi_col, dry_run, bsize2,
partition, pc_tree->horizontala[0], rate);
encode_b(cpi, tile_data, td, tp, mi_row, mi_col + hbs, dry_run, bsize2,
partition, pc_tree->horizontala[1], rate);
encode_b(cpi, tile_data, td, tp, mi_row + hbs, mi_col, dry_run, subsize,
partition, pc_tree->horizontala[2], rate);
break;
case PARTITION_HORZ_B:
encode_b(cpi, tile_data, td, tp, mi_row, mi_col, dry_run, subsize,
partition, pc_tree->horizontalb[0], rate);
encode_b(cpi, tile_data, td, tp, mi_row + hbs, mi_col, dry_run, bsize2,
partition, pc_tree->horizontalb[1], rate);
encode_b(cpi, tile_data, td, tp, mi_row + hbs, mi_col + hbs, dry_run,
bsize2, partition, pc_tree->horizontalb[2], rate);
break;
case PARTITION_VERT_A:
encode_b(cpi, tile_data, td, tp, mi_row, mi_col, dry_run, bsize2,
partition, pc_tree->verticala[0], rate);
encode_b(cpi, tile_data, td, tp, mi_row + hbs, mi_col, dry_run, bsize2,
partition, pc_tree->verticala[1], rate);
encode_b(cpi, tile_data, td, tp, mi_row, mi_col + hbs, dry_run, subsize,
partition, pc_tree->verticala[2], rate);
break;
case PARTITION_VERT_B:
encode_b(cpi, tile_data, td, tp, mi_row, mi_col, dry_run, subsize,
partition, pc_tree->verticalb[0], rate);
encode_b(cpi, tile_data, td, tp, mi_row, mi_col + hbs, dry_run, bsize2,
partition, pc_tree->verticalb[1], rate);
encode_b(cpi, tile_data, td, tp, mi_row + hbs, mi_col + hbs, dry_run,
bsize2, partition, pc_tree->verticalb[2], rate);
break;
case PARTITION_HORZ_4:
for (i = 0; i < SUB_PARTITIONS_PART4; ++i) {
int this_mi_row = mi_row + i * quarter_step;
if (i > 0 && this_mi_row >= mi_params->mi_rows) break;
encode_b(cpi, tile_data, td, tp, this_mi_row, mi_col, dry_run, subsize,
partition, pc_tree->horizontal4[i], rate);
}
break;
case PARTITION_VERT_4:
for (i = 0; i < SUB_PARTITIONS_PART4; ++i) {
int this_mi_col = mi_col + i * quarter_step;
if (i > 0 && this_mi_col >= mi_params->mi_cols) break;
encode_b(cpi, tile_data, td, tp, mi_row, this_mi_col, dry_run, subsize,
partition, pc_tree->vertical4[i], rate);
}
break;
default: assert(0 && "Invalid partition type."); break;
}
update_ext_partition_context(xd, mi_row, mi_col, subsize, bsize, partition);
}
/*!\brief AV1 block partition search (partition estimation and partial search).
*
* \ingroup partition_search
* Encode the block by applying pre-calculated partition patterns that are
* represented by coding block sizes stored in the mbmi array. Minor partition
* adjustments are tested and applied if they lead to lower rd costs. The
* partition types are limited to a basic set: none, horz, vert, and split.
*
* \param[in] cpi Top-level encoder structure
* \param[in] td Pointer to thread data
* \param[in] tile_data Pointer to struct holding adaptive
data/contexts/models for the tile during encoding
* \param[in] mib Array representing MB_MODE_INFO pointers for mi
blocks starting from the first pixel of the current
block
* \param[in] tp Pointer to the starting token
* \param[in] mi_row Row coordinate of the block in a step size of MI_SIZE
* \param[in] mi_col Column coordinate of the block in a step size of
MI_SIZE
* \param[in] bsize Current block size
* \param[in] rate Pointer to the final rate for encoding the current
block
* \param[in] dist Pointer to the final distortion of the current block
* \param[in] do_recon Whether the reconstruction function needs to be run,
either for finalizing a superblock or providing
reference for future sub-partitions
* \param[in] pc_tree Pointer to the PC_TREE node holding the picked
partitions and mode info for the current block
*
* Nothing is returned. The pc_tree struct is modified to store the
* picked partition and modes. The rate and dist are also updated with those
* corresponding to the best partition found.
*/
void av1_rd_use_partition(AV1_COMP *cpi, ThreadData *td, TileDataEnc *tile_data,
MB_MODE_INFO **mib, TokenExtra **tp, int mi_row,
int mi_col, BLOCK_SIZE bsize, int *rate,
int64_t *dist, int do_recon, PC_TREE *pc_tree) {
AV1_COMMON *const cm = &cpi->common;
const CommonModeInfoParams *const mi_params = &cm->mi_params;
const int num_planes = av1_num_planes(cm);
TileInfo *const tile_info = &tile_data->tile_info;
MACROBLOCK *const x = &td->mb;
MACROBLOCKD *const xd = &x->e_mbd;
const ModeCosts *mode_costs = &x->mode_costs;
const int bs = mi_size_wide[bsize];
const int hbs = bs / 2;
const int pl = (bsize >= BLOCK_8X8)
? partition_plane_context(xd, mi_row, mi_col, bsize)
: 0;
const PARTITION_TYPE partition =
(bsize >= BLOCK_8X8) ? get_partition(cm, xd->tree_type == CHROMA_PART,
mi_row, mi_col, bsize)
: PARTITION_NONE;
const BLOCK_SIZE subsize = get_partition_subsize(bsize, partition);
RD_SEARCH_MACROBLOCK_CONTEXT x_ctx;
RD_STATS last_part_rdc, invalid_rdc;
int plane_type = (xd->tree_type == CHROMA_PART);
const int plane_start = get_partition_plane_start(xd->tree_type);
const int plane_end = get_partition_plane_end(xd->tree_type, num_planes);
if (pc_tree->none == NULL) {
pc_tree->none = av1_alloc_pmc(cm, bsize, &td->shared_coeff_buf);
}
PICK_MODE_CONTEXT *ctx_none = pc_tree->none;
if (mi_row >= mi_params->mi_rows || mi_col >= mi_params->mi_cols) return;
assert(mi_size_wide[bsize] == mi_size_high[bsize]);
av1_invalid_rd_stats(&last_part_rdc);
av1_invalid_rd_stats(&invalid_rdc);
pc_tree->partitioning = partition;
xd->above_txfm_context =
cm->above_contexts.txfm[tile_info->tile_row] + mi_col;
xd->left_txfm_context =
xd->left_txfm_context_buffer + (mi_row & MAX_MIB_MASK);
av1_save_context(x, &x_ctx, mi_row, mi_col, bsize, num_planes);
if (bsize == BLOCK_16X16 && cpi->vaq_refresh) {
av1_set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
x->mb_energy = av1_log_block_var(cpi, x, bsize);
}
// Save rdmult before it might be changed, so it can be restored later.
const int orig_rdmult = x->rdmult;
setup_block_rdmult(cpi, x, mi_row, mi_col, bsize, NO_AQ, NULL);
for (int i = 0; i < SUB_PARTITIONS_SPLIT; ++i) {
pc_tree->split[i] = av1_alloc_pc_tree_node(subsize);
pc_tree->split[i]->index = i;
}
switch (partition) {
case PARTITION_NONE:
pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &last_part_rdc,
PARTITION_NONE, bsize, ctx_none, invalid_rdc);
break;
case PARTITION_HORZ:
for (int i = 0; i < SUB_PARTITIONS_RECT; ++i) {
if (pc_tree->horizontal[i] == NULL) {
pc_tree->horizontal[i] =
av1_alloc_pmc(cm, subsize, &td->shared_coeff_buf);
}
}
pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &last_part_rdc,
PARTITION_HORZ, subsize, pc_tree->horizontal[0],
invalid_rdc);
if (last_part_rdc.rate != INT_MAX && bsize >= BLOCK_8X8 &&
mi_row + hbs < mi_params->mi_rows) {
RD_STATS tmp_rdc;
const PICK_MODE_CONTEXT *const ctx_h = pc_tree->horizontal[0];
av1_init_rd_stats(&tmp_rdc);
av1_update_state(cpi, td, ctx_h, mi_row, mi_col, subsize, 1);
encode_superblock(cpi, tile_data, td, tp, DRY_RUN_NORMAL, subsize,
plane_start, plane_end, NULL);
pick_sb_modes(cpi, tile_data, x, mi_row + hbs, mi_col, &tmp_rdc,
PARTITION_HORZ, subsize, pc_tree->horizontal[1],
invalid_rdc);
if (tmp_rdc.rate == INT_MAX || tmp_rdc.dist == INT64_MAX) {
av1_invalid_rd_stats(&last_part_rdc);
break;
}
last_part_rdc.rate += tmp_rdc.rate;
last_part_rdc.dist += tmp_rdc.dist;
last_part_rdc.rdcost += tmp_rdc.rdcost;
}
break;
case PARTITION_VERT:
for (int i = 0; i < SUB_PARTITIONS_RECT; ++i) {
if (pc_tree->vertical[i] == NULL) {
pc_tree->vertical[i] =
av1_alloc_pmc(cm, subsize, &td->shared_coeff_buf);
}
}
pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &last_part_rdc,
PARTITION_VERT, subsize, pc_tree->vertical[0], invalid_rdc);
if (last_part_rdc.rate != INT_MAX && bsize >= BLOCK_8X8 &&
mi_col + hbs < mi_params->mi_cols) {
RD_STATS tmp_rdc;
const PICK_MODE_CONTEXT *const ctx_v = pc_tree->vertical[0];
av1_init_rd_stats(&tmp_rdc);
av1_update_state(cpi, td, ctx_v, mi_row, mi_col, subsize, 1);
encode_superblock(cpi, tile_data, td, tp, DRY_RUN_NORMAL, subsize,
plane_start, plane_end, NULL);
pick_sb_modes(cpi, tile_data, x, mi_row, mi_col + hbs, &tmp_rdc,
PARTITION_VERT, subsize,
pc_tree->vertical[bsize > BLOCK_8X8], invalid_rdc);
if (tmp_rdc.rate == INT_MAX || tmp_rdc.dist == INT64_MAX) {
av1_invalid_rd_stats(&last_part_rdc);
break;
}
last_part_rdc.rate += tmp_rdc.rate;
last_part_rdc.dist += tmp_rdc.dist;
last_part_rdc.rdcost += tmp_rdc.rdcost;
}
break;
case PARTITION_SPLIT:
last_part_rdc.rate = 0;
last_part_rdc.dist = 0;
last_part_rdc.rdcost = 0;
for (int i = 0; i < SUB_PARTITIONS_SPLIT; i++) {
int x_idx = (i & 1) * hbs;
int y_idx = (i >> 1) * hbs;
int jj = i >> 1, ii = i & 0x01;
RD_STATS tmp_rdc;
if ((mi_row + y_idx >= mi_params->mi_rows) ||
(mi_col + x_idx >= mi_params->mi_cols))
continue;
av1_init_rd_stats(&tmp_rdc);
av1_rd_use_partition(
cpi, td, tile_data,
mib + jj * hbs * mi_params->mi_stride + ii * hbs, tp,
mi_row + y_idx, mi_col + x_idx, subsize, &tmp_rdc.rate,
&tmp_rdc.dist, i != (SUB_PARTITIONS_SPLIT - 1), pc_tree->split[i]);
if (tmp_rdc.rate == INT_MAX || tmp_rdc.dist == INT64_MAX) {
av1_invalid_rd_stats(&last_part_rdc);
break;
}
last_part_rdc.rate += tmp_rdc.rate;
last_part_rdc.dist += tmp_rdc.dist;
}
break;
case PARTITION_VERT_A:
case PARTITION_VERT_B:
case PARTITION_HORZ_A:
case PARTITION_HORZ_B:
case PARTITION_HORZ_4:
case PARTITION_VERT_4:
assert(0 && "Cannot handle extended partition types");
default: assert(0); break;
}
if (last_part_rdc.rate < INT_MAX) {
last_part_rdc.rate +=
mode_costs->partition_cost[xd->tree_type == CHROMA_PART][pl][partition];
last_part_rdc.rdcost =
RDCOST(x->rdmult, last_part_rdc.rate, last_part_rdc.dist);
}
// If last_part is better set the partitioning to that.
mib[0]->sb_type[plane_type] = bsize;
if (bsize >= BLOCK_8X8) pc_tree->partitioning = partition;
av1_restore_context(x, &x_ctx, mi_row, mi_col, bsize, num_planes);
// We must have chosen a partitioning and encoding or we'll fail later on.
// No other opportunities for success.
if (bsize == cm->seq_params.sb_size)
assert(last_part_rdc.rate < INT_MAX && last_part_rdc.dist < INT64_MAX);
if (do_recon) {
if (bsize == cm->seq_params.sb_size) {
// NOTE: To get estimate for rate due to the tokens, use:
// int rate_coeffs = 0;
// encode_sb(cpi, td, tile_data, tp, mi_row, mi_col, DRY_RUN_COSTCOEFFS,
// bsize, pc_tree, &rate_coeffs);
x->cb_offset[plane_type] = 0;
encode_sb(cpi, td, tile_data, tp, mi_row, mi_col, OUTPUT_ENABLED, bsize,
pc_tree, NULL);
} else {
encode_sb(cpi, td, tile_data, tp, mi_row, mi_col, DRY_RUN_NORMAL, bsize,
pc_tree, NULL);
}
}
*rate = last_part_rdc.rate;
*dist = last_part_rdc.dist;
x->rdmult = orig_rdmult;
}
// Try searching for an encoding for the given subblock. Returns zero if the
// rdcost is already too high (to tell the caller not to bother searching for
// encodings of further subblocks).
static int rd_try_subblock(AV1_COMP *const cpi, ThreadData *td,
TileDataEnc *tile_data, TokenExtra **tp, int is_last,
int mi_row, int mi_col, BLOCK_SIZE subsize,
RD_STATS best_rdcost, RD_STATS *sum_rdc,
PARTITION_TYPE partition,
PICK_MODE_CONTEXT *this_ctx) {
MACROBLOCK *const x = &td->mb;
const int orig_mult = x->rdmult;
setup_block_rdmult(cpi, x, mi_row, mi_col, subsize, NO_AQ, NULL);
av1_rd_cost_update(x->rdmult, &best_rdcost);
RD_STATS rdcost_remaining;
av1_rd_stats_subtraction(x->rdmult, &best_rdcost, sum_rdc, &rdcost_remaining);
RD_STATS this_rdc;
pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &this_rdc, partition,
subsize, this_ctx, rdcost_remaining);
if (this_rdc.rate == INT_MAX) {
sum_rdc->rdcost = INT64_MAX;
} else {
sum_rdc->rate += this_rdc.rate;
sum_rdc->dist += this_rdc.dist;
av1_rd_cost_update(x->rdmult, sum_rdc);
}
if (sum_rdc->rdcost >= best_rdcost.rdcost) {
x->rdmult = orig_mult;
return 0;
}
MACROBLOCKD *xd = &x->e_mbd;
const AV1_COMMON *const cm = &cpi->common;
const int plane_start = get_partition_plane_start(xd->tree_type);
const int plane_end =
get_partition_plane_end(xd->tree_type, av1_num_planes(cm));
if (!is_last) {
av1_update_state(cpi, td, this_ctx, mi_row, mi_col, subsize, 1);
encode_superblock(cpi, tile_data, td, tp, DRY_RUN_NORMAL, subsize,
plane_start, plane_end, NULL);
}
x->rdmult = orig_mult;
return 1;
}
// Tests an AB partition, and updates the encoder status, the pick mode
// contexts, the best rdcost, and the best partition.
static bool rd_test_partition3(AV1_COMP *const cpi, ThreadData *td,
TileDataEnc *tile_data, TokenExtra **tp,
PC_TREE *pc_tree, RD_STATS *best_rdc,
PICK_MODE_CONTEXT *ctxs[SUB_PARTITIONS_AB],
int mi_row, int mi_col, BLOCK_SIZE bsize,
PARTITION_TYPE partition,
const BLOCK_SIZE ab_subsize[SUB_PARTITIONS_AB],
const int ab_mi_pos[SUB_PARTITIONS_AB][2]) {
const MACROBLOCK *const x = &td->mb;
const MACROBLOCKD *const xd = &x->e_mbd;
const int pl = partition_plane_context(xd, mi_row, mi_col, bsize);
RD_STATS sum_rdc;
av1_init_rd_stats(&sum_rdc);
sum_rdc.rate =
x->mode_costs.partition_cost[xd->tree_type == CHROMA_PART][pl][partition];
sum_rdc.rdcost = RDCOST(x->rdmult, sum_rdc.rate, 0);
// Loop over sub-partitions in AB partition type.
for (int i = 0; i < SUB_PARTITIONS_AB; i++) {
assert(ab_subsize[i] != BLOCK_INVALID);
if (!rd_try_subblock(cpi, td, tile_data, tp, i == SUB_PARTITIONS_AB - 1,
ab_mi_pos[i][0], ab_mi_pos[i][1], ab_subsize[i],
*best_rdc, &sum_rdc, partition, ctxs[i]))
return false;
}
av1_rd_cost_update(x->rdmult, &sum_rdc);
if (sum_rdc.rdcost >= best_rdc->rdcost) return false;
sum_rdc.rdcost = RDCOST(x->rdmult, sum_rdc.rate, sum_rdc.dist);
if (sum_rdc.rdcost >= best_rdc->rdcost) return false;
*best_rdc = sum_rdc;
pc_tree->partitioning = partition;
return true;
}
// Initialize state variables of partition search used in
// av1_rd_pick_partition().
static void init_partition_search_state_params(
MACROBLOCK *x, AV1_COMP *const cpi, PartitionSearchState *part_search_state,
int mi_row, int mi_col, BLOCK_SIZE bsize) {
MACROBLOCKD *const xd = &x->e_mbd;
const AV1_COMMON *const cm = &cpi->common;
PartitionBlkParams *blk_params = &part_search_state->part_blk_params;
const CommonModeInfoParams *const mi_params = &cpi->common.mi_params;
// Initialization of block size related parameters.
blk_params->mi_step = mi_size_wide[bsize] / 2;
blk_params->mi_row = mi_row;
blk_params->mi_col = mi_col;
blk_params->mi_row_edge = mi_row + blk_params->mi_step;
blk_params->mi_col_edge = mi_col + blk_params->mi_step;
blk_params->width = block_size_wide[bsize];
blk_params->min_partition_size_1d =
block_size_wide[x->sb_enc.min_partition_size];
blk_params->subsize = get_partition_subsize(bsize, PARTITION_SPLIT);
blk_params->split_bsize2 = blk_params->subsize;
blk_params->bsize_at_least_8x8 = (bsize >= BLOCK_8X8);
blk_params->bsize = bsize;
// Check if the partition corresponds to edge block.
blk_params->has_rows = (blk_params->mi_row_edge < mi_params->mi_rows);
blk_params->has_cols = (blk_params->mi_col_edge < mi_params->mi_cols);
// Update intra partitioning related info.
part_search_state->intra_part_info = &x->part_search_info;
// Prepare for segmentation CNN-based partitioning for intra-frame.
if (frame_is_intra_only(cm) && bsize == BLOCK_64X64) {
part_search_state->intra_part_info->quad_tree_idx = 0;
part_search_state->intra_part_info->cnn_output_valid = 0;
}
// Set partition plane context index.
part_search_state->pl_ctx_idx =
blk_params->bsize_at_least_8x8
? partition_plane_context(xd, mi_row, mi_col, bsize)
: 0;
// Partition cost buffer update
ModeCosts *mode_costs = &x->mode_costs;
part_search_state->partition_cost =
mode_costs->partition_cost[xd->tree_type == CHROMA_PART]
[part_search_state->pl_ctx_idx];
// Initialize HORZ and VERT win flags as true for all split partitions.
for (int i = 0; i < SUB_PARTITIONS_SPLIT; i++) {
part_search_state->split_part_rect_win[i].rect_part_win[HORZ] = true;
part_search_state->split_part_rect_win[i].rect_part_win[VERT] = true;
}
// Initialize the rd cost.
av1_init_rd_stats(&part_search_state->this_rdc);
// Initialize RD costs for partition types to 0.
part_search_state->none_rd = 0;
av1_zero(part_search_state->split_rd);
av1_zero(part_search_state->rect_part_rd);
// Initialize SPLIT partition to be not ready.
av1_zero(part_search_state->is_split_ctx_is_ready);
// Initialize HORZ and VERT partitions to be not ready.
av1_zero(part_search_state->is_rect_ctx_is_ready);
// Chroma subsampling.
part_search_state->ss_x = x->e_mbd.plane[1].subsampling_x;
part_search_state->ss_y = x->e_mbd.plane[1].subsampling_y;
// Initialize partition search flags to defaults.
part_search_state->terminate_partition_search = 0;
part_search_state->do_square_split =
blk_params->bsize_at_least_8x8 &&
(xd->tree_type != CHROMA_PART || bsize > BLOCK_8X8);
part_search_state->do_rectangular_split =
cpi->oxcf.part_cfg.enable_rect_partitions &&
(xd->tree_type != CHROMA_PART || bsize > BLOCK_8X8);
av1_zero(part_search_state->prune_rect_part);
// Initialize allowed partition types for the partition block.
part_search_state->partition_none_allowed =
blk_params->has_rows && blk_params->has_cols;
part_search_state->partition_rect_allowed[HORZ] =
blk_params->has_cols && blk_params->bsize_at_least_8x8 &&
cpi->oxcf.part_cfg.enable_rect_partitions &&
(xd->tree_type != CHROMA_PART || bsize > BLOCK_8X8) &&
get_plane_block_size(get_partition_subsize(bsize, PARTITION_HORZ),
part_search_state->ss_x,
part_search_state->ss_y) != BLOCK_INVALID;
part_search_state->partition_rect_allowed[VERT] =
blk_params->has_rows && blk_params->bsize_at_least_8x8 &&
(xd->tree_type != CHROMA_PART || bsize > BLOCK_8X8) &&
cpi->oxcf.part_cfg.enable_rect_partitions &&
get_plane_block_size(get_partition_subsize(bsize, PARTITION_VERT),
part_search_state->ss_x,
part_search_state->ss_y) != BLOCK_INVALID;
// Reset the flag indicating whether a partition leading to a rdcost lower
// than the bound best_rdc has been found.
part_search_state->found_best_partition = false;
}
// Override partition cost buffer for the edge blocks.
static void set_partition_cost_for_edge_blk(
AV1_COMMON const *cm, MACROBLOCKD *const xd,
PartitionSearchState *part_search_state) {
PartitionBlkParams blk_params = part_search_state->part_blk_params;
assert(blk_params.bsize_at_least_8x8 && part_search_state->pl_ctx_idx >= 0);
const int plane = xd->tree_type == CHROMA_PART;
const aom_cdf_prob *partition_cdf =
cm->fc->partition_cdf[plane][part_search_state->pl_ctx_idx];
const int max_cost = av1_cost_symbol(0);
for (PARTITION_TYPE i = 0; i < PARTITION_TYPES; ++i)
part_search_state->tmp_partition_cost[i] = max_cost;
if (blk_params.has_cols) {
// At the bottom, the two possibilities are HORZ and SPLIT.
aom_cdf_prob bot_cdf[2];
partition_gather_vert_alike(bot_cdf, partition_cdf, blk_params.bsize);
static const int bot_inv_map[2] = { PARTITION_HORZ, PARTITION_SPLIT };
av1_cost_tokens_from_cdf(part_search_state->tmp_partition_cost, bot_cdf,
bot_inv_map);
} else if (blk_params.has_rows) {
// At the right, the two possibilities are VERT and SPLIT.
aom_cdf_prob rhs_cdf[2];
partition_gather_horz_alike(rhs_cdf, partition_cdf, blk_params.bsize);
static const int rhs_inv_map[2] = { PARTITION_VERT, PARTITION_SPLIT };
av1_cost_tokens_from_cdf(part_search_state->tmp_partition_cost, rhs_cdf,
rhs_inv_map);
} else {
// At the bottom right, we always split.
part_search_state->tmp_partition_cost[PARTITION_SPLIT] = 0;
}
// Override the partition cost buffer.
part_search_state->partition_cost = part_search_state->tmp_partition_cost;
}
// Reset the partition search state flags when
// must_find_valid_partition is equal to 1.
static AOM_INLINE void reset_part_limitations(
AV1_COMP *const cpi, PartitionSearchState *part_search_state) {
PartitionBlkParams blk_params = part_search_state->part_blk_params;
const int is_rect_part_allowed =
blk_params.bsize_at_least_8x8 &&
cpi->oxcf.part_cfg.enable_rect_partitions &&
(blk_params.width > blk_params.min_partition_size_1d);
part_search_state->do_square_split =
blk_params.bsize_at_least_8x8 &&
(blk_params.width > blk_params.min_partition_size_1d);
part_search_state->partition_none_allowed =
blk_params.has_rows && blk_params.has_cols &&
(blk_params.width >= blk_params.min_partition_size_1d);
part_search_state->partition_rect_allowed[HORZ] =
blk_params.has_cols && is_rect_part_allowed &&
get_plane_block_size(
get_partition_subsize(blk_params.bsize, PARTITION_HORZ),
part_search_state->ss_x, part_search_state->ss_y) != BLOCK_INVALID;
part_search_state->partition_rect_allowed[VERT] =
blk_params.has_rows && is_rect_part_allowed &&
get_plane_block_size(
get_partition_subsize(blk_params.bsize, PARTITION_VERT),
part_search_state->ss_x, part_search_state->ss_y) != BLOCK_INVALID;
part_search_state->terminate_partition_search = 0;
}
// Rectangular partitions evaluation at sub-block level.
static void rd_pick_rect_partition(AV1_COMP *const cpi, TileDataEnc *tile_data,
MACROBLOCK *x,
PICK_MODE_CONTEXT *cur_partition_ctx,
PartitionSearchState *part_search_state,
RD_STATS *best_rdc, const int idx,
int mi_row, int mi_col, BLOCK_SIZE bsize,
PARTITION_TYPE partition_type) {
// Obtain the remainder from the best rd cost
// for further processing of partition.
RD_STATS best_remain_rdcost;
av1_rd_stats_subtraction(x->rdmult, best_rdc, &part_search_state->sum_rdc,
&best_remain_rdcost);
// Obtain the best mode for the partition sub-block.
pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &part_search_state->this_rdc,
partition_type, bsize, cur_partition_ctx, best_remain_rdcost);
av1_rd_cost_update(x->rdmult, &part_search_state->this_rdc);
// Update the partition rd cost with the current sub-block rd.
if (part_search_state->this_rdc.rate == INT_MAX) {
part_search_state->sum_rdc.rdcost = INT64_MAX;
} else {
part_search_state->sum_rdc.rate += part_search_state->this_rdc.rate;
part_search_state->sum_rdc.dist += part_search_state->this_rdc.dist;
av1_rd_cost_update(x->rdmult, &part_search_state->sum_rdc);
}
const RECT_PART_TYPE rect_part =
partition_type == PARTITION_HORZ ? HORZ : VERT;
part_search_state->rect_part_rd[rect_part][idx] =
part_search_state->this_rdc.rdcost;
}
typedef int (*active_edge_info)(const AV1_COMP *cpi, int mi_col, int mi_step);
// Checks if HORZ / VERT partition search is allowed.
static AOM_INLINE int is_rect_part_allowed(
const AV1_COMP *cpi, PartitionSearchState *part_search_state,
active_edge_info *active_edge, RECT_PART_TYPE rect_part, const int mi_pos) {
PartitionBlkParams blk_params = part_search_state->part_blk_params;
const int is_part_allowed =
(!part_search_state->terminate_partition_search &&
part_search_state->partition_rect_allowed[rect_part] &&
!part_search_state->prune_rect_part[rect_part] &&
(part_search_state->do_rectangular_split ||
active_edge[rect_part](cpi, mi_pos, blk_params.mi_step)));
return is_part_allowed;
}
// Rectangular partition types search function.
static void rectangular_partition_search(
AV1_COMP *const cpi, ThreadData *td, TileDataEnc *tile_data,
TokenExtra **tp, MACROBLOCK *x, PC_TREE *pc_tree,
RD_SEARCH_MACROBLOCK_CONTEXT *x_ctx,
PartitionSearchState *part_search_state, RD_STATS *best_rdc,
RD_RECT_PART_WIN_INFO *rect_part_win_info) {
const AV1_COMMON *const cm = &cpi->common;
PartitionBlkParams blk_params = part_search_state->part_blk_params;
RD_STATS *sum_rdc = &part_search_state->sum_rdc;
const int rect_partition_type[NUM_RECT_PARTS] = { PARTITION_HORZ,
PARTITION_VERT };
MACROBLOCKD *xd = &x->e_mbd;
const int plane_start = get_partition_plane_start(xd->tree_type);
const int plane_end =
get_partition_plane_end(xd->tree_type, av1_num_planes(cm));
// mi_pos_rect[NUM_RECT_PARTS][SUB_PARTITIONS_RECT][0]: mi_row postion of
// HORZ and VERT partition types.
// mi_pos_rect[NUM_RECT_PARTS][SUB_PARTITIONS_RECT][1]: mi_col postion of
// HORZ and VERT partition types.
const int mi_pos_rect[NUM_RECT_PARTS][SUB_PARTITIONS_RECT][2] = {
{ { blk_params.mi_row, blk_params.mi_col },
{ blk_params.mi_row_edge, blk_params.mi_col } },
{ { blk_params.mi_row, blk_params.mi_col },
{ blk_params.mi_row, blk_params.mi_col_edge } }
};
// Initialize active edge_type function pointer
// for HOZR and VERT partition types.
active_edge_info active_edge_type[NUM_RECT_PARTS] = { av1_active_h_edge,
av1_active_v_edge };
// Indicates edge blocks for HORZ and VERT partition types.
const int is_not_edge_block[NUM_RECT_PARTS] = { blk_params.has_rows,
blk_params.has_cols };
// Initialize pc tree context for HORZ and VERT partition types.
PICK_MODE_CONTEXT **cur_ctx[NUM_RECT_PARTS][SUB_PARTITIONS_RECT] = {
{ &pc_tree->horizontal[0], &pc_tree->horizontal[1] },
{ &pc_tree->vertical[0], &pc_tree->vertical[1] }
};
// Loop over rectangular partition types.
for (RECT_PART_TYPE i = HORZ; i < NUM_RECT_PARTS; i++) {
assert(IMPLIES(!cpi->oxcf.part_cfg.enable_rect_partitions,
!part_search_state->partition_rect_allowed[i]));
// Check if the HORZ / VERT partition search is to be performed.
if (!is_rect_part_allowed(cpi, part_search_state, active_edge_type, i,
mi_pos_rect[i][0][i]))
continue;
// Sub-partition idx.
int sub_part_idx = 0;
PARTITION_TYPE partition_type = rect_partition_type[i];
blk_params.subsize =
get_partition_subsize(blk_params.bsize, partition_type);
assert(blk_params.subsize <= BLOCK_LARGEST);
av1_init_rd_stats(sum_rdc);
for (int j = 0; j < SUB_PARTITIONS_RECT; j++) {
if (cur_ctx[i][j][0] == NULL) {
cur_ctx[i][j][0] =
av1_alloc_pmc(cm, blk_params.subsize, &td->shared_coeff_buf);
}
}
sum_rdc->rate = part_search_state->partition_cost[partition_type];
sum_rdc->rdcost = RDCOST(x->rdmult, sum_rdc->rate, 0);
#if CONFIG_COLLECT_PARTITION_STATS
if (best_rdc.rdcost - sum_rdc->rdcost >= 0) {
partition_attempts[partition_type] += 1;
aom_usec_timer_start(&partition_timer);
partition_timer_on = 1;
}
#endif
// First sub-partition evaluation in HORZ / VERT partition type.
rd_pick_rect_partition(
cpi, tile_data, x, cur_ctx[i][sub_part_idx][0], part_search_state,
best_rdc, 0, mi_pos_rect[i][sub_part_idx][0],
mi_pos_rect[i][sub_part_idx][1], blk_params.subsize, partition_type);
// Start of second sub-partition evaluation.
// Evaluate second sub-partition if the first sub-partition cost
// is less than the best cost and if it is not an edge block.
if (sum_rdc->rdcost < best_rdc->rdcost && is_not_edge_block[i]) {
const MB_MODE_INFO *const mbmi = &cur_ctx[i][sub_part_idx][0]->mic;
const PALETTE_MODE_INFO *const pmi = &mbmi->palette_mode_info;
// Neither palette mode nor cfl predicted.
if (pmi->palette_size[PLANE_TYPE_Y] == 0 &&
pmi->palette_size[PLANE_TYPE_UV] == 0) {
if (mbmi->uv_mode != UV_CFL_PRED)
part_search_state->is_rect_ctx_is_ready[i] = 1;
}
av1_update_state(cpi, td, cur_ctx[i][sub_part_idx][0], blk_params.mi_row,
blk_params.mi_col, blk_params.subsize, DRY_RUN_NORMAL);
encode_superblock(cpi, tile_data, td, tp, DRY_RUN_NORMAL,
blk_params.subsize, plane_start, plane_end, NULL);
// Second sub-partition evaluation in HORZ / VERT partition type.
sub_part_idx = 1;
rd_pick_rect_partition(
cpi, tile_data, x, cur_ctx[i][sub_part_idx][0], part_search_state,
best_rdc, 1, mi_pos_rect[i][sub_part_idx][0],
mi_pos_rect[i][sub_part_idx][1], blk_params.subsize, partition_type);
}
#if CONFIG_COLLECT_PARTITION_STATS
if (partition_timer_on) {
aom_usec_timer_mark(&partition_timer);
int64_t time = aom_usec_timer_elapsed(&partition_timer);
partition_times[partition_type] += time;
partition_timer_on = 0;
}
#endif
// Update HORZ / VERT best partition.
if (sum_rdc->rdcost < best_rdc->rdcost) {
sum_rdc->rdcost = RDCOST(x->rdmult, sum_rdc->rate, sum_rdc->dist);
if (sum_rdc->rdcost < best_rdc->rdcost) {
*best_rdc = *sum_rdc;
part_search_state->found_best_partition = true;
pc_tree->partitioning = partition_type;
}
} else {
// Update HORZ / VERT win flag.
if (rect_part_win_info != NULL)
rect_part_win_info->rect_part_win[i] = false;
}
av1_restore_context(x, x_ctx, blk_params.mi_row, blk_params.mi_col,
blk_params.bsize, av1_num_planes(cm));
}
}
// AB partition type evaluation.
static void rd_pick_ab_part(
AV1_COMP *const cpi, ThreadData *td, TileDataEnc *tile_data,
TokenExtra **tp, MACROBLOCK *x, RD_SEARCH_MACROBLOCK_CONTEXT *x_ctx,
PC_TREE *pc_tree, PICK_MODE_CONTEXT *dst_ctxs[SUB_PARTITIONS_AB],
PartitionSearchState *part_search_state, RD_STATS *best_rdc,
const BLOCK_SIZE ab_subsize[SUB_PARTITIONS_AB],
const int ab_mi_pos[SUB_PARTITIONS_AB][2], const PARTITION_TYPE part_type) {
const AV1_COMMON *const cm = &cpi->common;
PartitionBlkParams blk_params = part_search_state->part_blk_params;
const int mi_row = blk_params.mi_row;
const int mi_col = blk_params.mi_col;
const int bsize = blk_params.bsize;
#if CONFIG_COLLECT_PARTITION_STATS
{
RD_STATS tmp_sum_rdc;
av1_init_rd_stats(&tmp_sum_rdc);
tmp_sum_rdc.rate =
x->partition_cost[part_search_state->pl_ctx_idx][part_type];
tmp_sum_rdc.rdcost = RDCOST(x->rdmult, tmp_sum_rdc.rate, 0);
if (best_rdc->rdcost - tmp_sum_rdc.rdcost >= 0) {
partition_attempts[part_type] += 1;
aom_usec_timer_start(&partition_timer);
partition_timer_on = 1;
}
}
#endif
// Test this partition and update the best partition.
part_search_state->found_best_partition |= rd_test_partition3(
cpi, td, tile_data, tp, pc_tree, best_rdc, dst_ctxs, mi_row, mi_col,
bsize, part_type, ab_subsize, ab_mi_pos);
#if CONFIG_COLLECT_PARTITION_STATS
if (partition_timer_on) {
aom_usec_timer_mark(&partition_timer);
int64_t time = aom_usec_timer_elapsed(&partition_timer);
partition_times[part_type] += time;
partition_timer_on = 0;
}
#endif
av1_restore_context(x, x_ctx, mi_row, mi_col, bsize, av1_num_planes(cm));
}
// Check if AB partitions search is allowed.
static AOM_INLINE int is_ab_part_allowed(
PartitionSearchState *part_search_state,
const int ab_partitions_allowed[NUM_AB_PARTS], const int ab_part_type) {
const int is_horz_ab = (ab_part_type >> 1);
const int is_part_allowed =
(!part_search_state->terminate_partition_search &&
part_search_state->partition_rect_allowed[is_horz_ab] &&
ab_partitions_allowed[ab_part_type]);
return is_part_allowed;
}
// Set mode search context.
static AOM_INLINE void set_mode_search_ctx(
PC_TREE *pc_tree, const int is_ctx_ready[NUM_AB_PARTS][2],
PICK_MODE_CONTEXT **mode_srch_ctx[NUM_AB_PARTS][2]) {
mode_srch_ctx[HORZ_B][0] = &pc_tree->horizontal[0];
mode_srch_ctx[VERT_B][0] = &pc_tree->vertical[0];
if (is_ctx_ready[HORZ_A][0])
mode_srch_ctx[HORZ_A][0] = &pc_tree->split[0]->none;
if (is_ctx_ready[VERT_A][0])
mode_srch_ctx[VERT_A][0] = &pc_tree->split[0]->none;
if (is_ctx_ready[HORZ_A][1])
mode_srch_ctx[HORZ_A][1] = &pc_tree->split[1]->none;
}
// AB Partitions type search.
static void ab_partitions_search(
AV1_COMP *const cpi, ThreadData *td, TileDataEnc *tile_data,
TokenExtra **tp, MACROBLOCK *x, RD_SEARCH_MACROBLOCK_CONTEXT *x_ctx,
PC_TREE *pc_tree, PartitionSearchState *part_search_state,
RD_STATS *best_rdc, RD_RECT_PART_WIN_INFO *rect_part_win_info,
int pb_source_variance, int ext_partition_allowed) {
const AV1_COMMON *const cm = &cpi->common;
PartitionBlkParams blk_params = part_search_state->part_blk_params;
const int mi_row = blk_params.mi_row;
const int mi_col = blk_params.mi_col;
const int bsize = blk_params.bsize;
int ab_partitions_allowed[NUM_AB_PARTS] = { 1, 1, 1, 1 };
// Prune AB partitions
av1_prune_ab_partitions(
cpi, x, pc_tree, bsize, pb_source_variance, best_rdc->rdcost,
part_search_state->rect_part_rd, part_search_state->split_rd,
rect_part_win_info, ext_partition_allowed,
part_search_state->partition_rect_allowed[HORZ],
part_search_state->partition_rect_allowed[VERT],
&ab_partitions_allowed[HORZ_A], &ab_partitions_allowed[HORZ_B],
&ab_partitions_allowed[VERT_A], &ab_partitions_allowed[VERT_B]);
// Flags to indicate whether the mode search is done.
const int is_ctx_ready[NUM_AB_PARTS][2] = {
{ part_search_state->is_split_ctx_is_ready[0],
part_search_state->is_split_ctx_is_ready[1] },
{ part_search_state->is_rect_ctx_is_ready[HORZ], 0 },
{ part_search_state->is_split_ctx_is_ready[0], 0 },
{ part_search_state->is_rect_ctx_is_ready[VERT], 0 }
};
// Current partition context.
PICK_MODE_CONTEXT **cur_part_ctxs[NUM_AB_PARTS] = { pc_tree->horizontala,
pc_tree->horizontalb,
pc_tree->verticala,
pc_tree->verticalb };
// Context of already evaluted partition types.
PICK_MODE_CONTEXT **mode_srch_ctx[NUM_AB_PARTS][2];
// Set context of already evaluted partition types.
set_mode_search_ctx(pc_tree, is_ctx_ready, mode_srch_ctx);
// Array of sub-partition size of AB partition types.
const BLOCK_SIZE ab_subsize[NUM_AB_PARTS][SUB_PARTITIONS_AB] = {
{ blk_params.split_bsize2, blk_params.split_bsize2,
get_partition_subsize(bsize, PARTITION_HORZ_A) },
{ get_partition_subsize(bsize, PARTITION_HORZ_B), blk_params.split_bsize2,
blk_params.split_bsize2 },
{ blk_params.split_bsize2, blk_params.split_bsize2,
get_partition_subsize(bsize, PARTITION_VERT_A) },
{ get_partition_subsize(bsize, PARTITION_VERT_B), blk_params.split_bsize2,
blk_params.split_bsize2 }
};
// Array of mi_row, mi_col positions corresponds to each sub-partition in AB
// partition types.
const int ab_mi_pos[NUM_AB_PARTS][SUB_PARTITIONS_AB][2] = {
{ { mi_row, mi_col },
{ mi_row, blk_params.mi_col_edge },
{ blk_params.mi_row_edge, mi_col } },
{ { mi_row, mi_col },
{ blk_params.mi_row_edge, mi_col },
{ blk_params.mi_row_edge, blk_params.mi_col_edge } },
{ { mi_row, mi_col },
{ blk_params.mi_row_edge, mi_col },
{ mi_row, blk_params.mi_col_edge } },
{ { mi_row, mi_col },
{ mi_row, blk_params.mi_col_edge },
{ blk_params.mi_row_edge, blk_params.mi_col_edge } }
};
// Loop over AB partition types.
for (AB_PART_TYPE ab_part_type = 0; ab_part_type < NUM_AB_PARTS;
ab_part_type++) {
const PARTITION_TYPE part_type = ab_part_type + PARTITION_HORZ_A;
// Check if the AB partition search is to be performed.
if (!is_ab_part_allowed(part_search_state, ab_partitions_allowed,
ab_part_type))