Google Git
Sign in
aomedia / aom / c7101aa94a592dd708bcd37d69649c3f238b659c / . / av1 / encoder / encodeframe.c
blob: fe46061da1ee886108949ef3efeb7fc2dcd00c79 [file] [log] [blame]
/*
* Copyright (c) 2016, Alliance for Open Media. All rights reserved
*
* This source code is subject to the terms of the BSD 2 Clause License and
* the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
* was not distributed with this source code in the LICENSE file, you can
* obtain it at www.aomedia.org/license/software. 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 www.aomedia.org/license/patent.
*/
#include <limits.h>
#include <math.h>
#include <stdio.h>
#include "config/aom_config.h"
#include "config/aom_dsp_rtcd.h"
#include "config/av1_rtcd.h"
#include "aom_dsp/aom_dsp_common.h"
#include "aom_dsp/binary_codes_writer.h"
#include "aom_ports/mem.h"
#include "aom_ports/aom_timer.h"
#include "aom_ports/system_state.h"
#if CONFIG_MISMATCH_DEBUG
#include "aom_util/debug_util.h"
#endif // CONFIG_MISMATCH_DEBUG
#include "av1/common/cfl.h"
#include "av1/common/common.h"
#include "av1/common/entropy.h"
#include "av1/common/entropymode.h"
#include "av1/common/idct.h"
#include "av1/common/mv.h"
#include "av1/common/mvref_common.h"
#include "av1/common/pred_common.h"
#include "av1/common/quant_common.h"
#include "av1/common/reconintra.h"
#include "av1/common/reconinter.h"
#include "av1/common/seg_common.h"
#include "av1/common/tile_common.h"
#include "av1/encoder/ab_partition_model_weights.h"
#include "av1/encoder/aq_complexity.h"
#include "av1/encoder/aq_cyclicrefresh.h"
#include "av1/encoder/aq_variance.h"
#include "av1/common/warped_motion.h"
#include "av1/encoder/global_motion.h"
#include "av1/encoder/encodeframe.h"
#include "av1/encoder/encodemb.h"
#include "av1/encoder/encodemv.h"
#include "av1/encoder/encodetxb.h"
#include "av1/encoder/ethread.h"
#include "av1/encoder/extend.h"
#include "av1/encoder/ml.h"
#include "av1/encoder/rd.h"
#include "av1/encoder/rdopt.h"
#include "av1/encoder/segmentation.h"
#include "av1/encoder/tokenize.h"
static void encode_superblock(const AV1_COMP *const cpi, TileDataEnc *tile_data,
ThreadData *td, TOKENEXTRA **t, RUN_TYPE dry_run,
int mi_row, int mi_col, BLOCK_SIZE bsize,
int *rate);
// This is used as a reference when computing the source variance for the
// purposes of activity masking.
// Eventually this should be replaced by custom no-reference routines,
// which will be faster.
static const uint8_t AV1_VAR_OFFS[MAX_SB_SIZE] = {
128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
128, 128, 128, 128, 128, 128, 128, 128
};
static const uint16_t AV1_HIGH_VAR_OFFS_8[MAX_SB_SIZE] = {
128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
128, 128, 128, 128, 128, 128, 128, 128
};
static const uint16_t AV1_HIGH_VAR_OFFS_10[MAX_SB_SIZE] = {
128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4
};
static const uint16_t AV1_HIGH_VAR_OFFS_12[MAX_SB_SIZE] = {
128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
128 * 16, 128 * 16
};
#if CONFIG_FP_MB_STATS
static const uint8_t num_16x16_blocks_wide_lookup[BLOCK_SIZES_ALL] = {
1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 2, 4, 4, 4, 8, 8, 1, 1, 1, 2, 2, 4
};
static const uint8_t num_16x16_blocks_high_lookup[BLOCK_SIZES_ALL] = {
1, 1, 1, 1, 1, 1, 1, 2, 1, 2, 4, 2, 4, 8, 4, 8, 1, 1, 2, 1, 4, 2
};
#endif // CONFIG_FP_MB_STATS
unsigned int av1_get_sby_perpixel_variance(const AV1_COMP *cpi,
const struct buf_2d *ref,
BLOCK_SIZE bs) {
unsigned int sse;
const unsigned int var =
cpi->fn_ptr[bs].vf(ref->buf, ref->stride, AV1_VAR_OFFS, 0, &sse);
return ROUND_POWER_OF_TWO(var, num_pels_log2_lookup[bs]);
}
unsigned int av1_high_get_sby_perpixel_variance(const AV1_COMP *cpi,
const struct buf_2d *ref,
BLOCK_SIZE bs, int bd) {
unsigned int var, sse;
switch (bd) {
case 10:
var =
cpi->fn_ptr[bs].vf(ref->buf, ref->stride,
CONVERT_TO_BYTEPTR(AV1_HIGH_VAR_OFFS_10), 0, &sse);
break;
case 12:
var =
cpi->fn_ptr[bs].vf(ref->buf, ref->stride,
CONVERT_TO_BYTEPTR(AV1_HIGH_VAR_OFFS_12), 0, &sse);
break;
case 8:
default:
var =
cpi->fn_ptr[bs].vf(ref->buf, ref->stride,
CONVERT_TO_BYTEPTR(AV1_HIGH_VAR_OFFS_8), 0, &sse);
break;
}
return ROUND_POWER_OF_TWO(var, num_pels_log2_lookup[bs]);
}
static unsigned int get_sby_perpixel_diff_variance(const AV1_COMP *const cpi,
const struct buf_2d *ref,
int mi_row, int mi_col,
BLOCK_SIZE bs) {
unsigned int sse, var;
uint8_t *last_y;
const YV12_BUFFER_CONFIG *last = get_ref_frame_buffer(cpi, LAST_FRAME);
assert(last != NULL);
last_y =
&last->y_buffer[mi_row * MI_SIZE * last->y_stride + mi_col * MI_SIZE];
var = cpi->fn_ptr[bs].vf(ref->buf, ref->stride, last_y, last->y_stride, &sse);
return ROUND_POWER_OF_TWO(var, num_pels_log2_lookup[bs]);
}
static BLOCK_SIZE get_rd_var_based_fixed_partition(AV1_COMP *cpi, MACROBLOCK *x,
int mi_row, int mi_col) {
unsigned int var = get_sby_perpixel_diff_variance(
cpi, &x->plane[0].src, mi_row, mi_col, BLOCK_64X64);
if (var < 8)
return BLOCK_64X64;
else if (var < 128)
return BLOCK_32X32;
else if (var < 2048)
return BLOCK_16X16;
else
return BLOCK_8X8;
}
// Lighter version of set_offsets that only sets the mode info
// pointers.
static void set_mode_info_offsets(const AV1_COMP *const cpi,
MACROBLOCK *const x, MACROBLOCKD *const xd,
int mi_row, int mi_col) {
const AV1_COMMON *const cm = &cpi->common;
const int idx_str = xd->mi_stride * mi_row + mi_col;
xd->mi = cm->mi_grid_visible + idx_str;
xd->mi[0] = cm->mi + idx_str;
x->mbmi_ext = cpi->mbmi_ext_base + (mi_row * cm->mi_cols + mi_col);
}
static void 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;
const int mi_width = mi_size_wide[bsize];
const int mi_height = mi_size_high[bsize];
set_mode_info_offsets(cpi, x, xd, mi_row, mi_col);
set_skip_context(xd, mi_row, mi_col, num_planes);
xd->above_txfm_context = cm->above_txfm_context[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, get_frame_new_buffer(cm), 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.
x->mv_limits.row_min =
-(((mi_row + mi_height) * MI_SIZE) + AOM_INTERP_EXTEND);
x->mv_limits.col_min = -(((mi_col + mi_width) * MI_SIZE) + AOM_INTERP_EXTEND);
x->mv_limits.row_max = (cm->mi_rows - mi_row) * MI_SIZE + AOM_INTERP_EXTEND;
x->mv_limits.col_max = (cm->mi_cols - mi_col) * MI_SIZE + AOM_INTERP_EXTEND;
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_rows,
cm->mi_cols);
// Set up source buffers.
av1_setup_src_planes(x, cpi->source, mi_row, mi_col, num_planes);
// R/D setup.
x->rdmult = cpi->rd.RDMULT;
// required by av1_append_sub8x8_mvs_for_idx() and av1_find_best_ref_mvs()
xd->tile = *tile;
}
static void 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;
MACROBLOCKD *const xd = &x->e_mbd;
MB_MODE_INFO *mbmi;
const struct segmentation *const seg = &cm->seg;
set_offsets_without_segment_id(cpi, tile, x, mi_row, mi_col, bsize);
mbmi = xd->mi[0];
xd->cfl.mi_row = mi_row;
xd->cfl.mi_col = mi_col;
mbmi->segment_id = 0;
// Setup segment ID.
if (seg->enabled) {
if (seg->enabled && !cpi->vaq_refresh) {
const uint8_t *const map =
seg->update_map ? cpi->segmentation_map : cm->last_frame_seg_map;
mbmi->segment_id =
map ? get_segment_id(cm, map, bsize, mi_row, mi_col) : 0;
}
av1_init_plane_quantizers(cpi, x, mbmi->segment_id);
}
}
static void reset_intmv_filter_type(MB_MODE_INFO *mbmi) {
InterpFilter filters[2];
for (int dir = 0; dir < 2; ++dir) {
filters[dir] = av1_extract_interp_filter(mbmi->interp_filters, dir);
}
mbmi->interp_filters = av1_make_interp_filters(filters[0], filters[1]);
}
static void update_filter_type_count(uint8_t allow_update_cdf,
FRAME_COUNTS *counts,
const MACROBLOCKD *xd,
const MB_MODE_INFO *mbmi) {
int dir;
for (dir = 0; dir < 2; ++dir) {
const int ctx = av1_get_pred_context_switchable_interp(xd, dir);
InterpFilter filter = av1_extract_interp_filter(mbmi->interp_filters, dir);
++counts->switchable_interp[ctx][filter];
if (allow_update_cdf) {
update_cdf(xd->tile_ctx->switchable_interp_cdf[ctx], filter,
SWITCHABLE_FILTERS);
}
}
}
static void update_global_motion_used(PREDICTION_MODE mode, BLOCK_SIZE bsize,
const MB_MODE_INFO *mbmi,
RD_COUNTS *rdc) {
if (mode == GLOBALMV || mode == GLOBAL_GLOBALMV) {
const int num_4x4s = mi_size_wide[bsize] * mi_size_high[bsize];
int ref;
for (ref = 0; ref < 1 + has_second_ref(mbmi); ++ref) {
rdc->global_motion_used[mbmi->ref_frame[ref]] += num_4x4s;
}
}
}
static void reset_tx_size(MACROBLOCK *x, MB_MODE_INFO *mbmi,
const TX_MODE tx_mode) {
MACROBLOCKD *const xd = &x->e_mbd;
if (xd->lossless[mbmi->segment_id]) {
mbmi->tx_size = TX_4X4;
} else if (tx_mode != TX_MODE_SELECT) {
mbmi->tx_size = tx_size_from_tx_mode(mbmi->sb_type, tx_mode);
} else {
BLOCK_SIZE bsize = mbmi->sb_type;
TX_SIZE min_tx_size = depth_to_tx_size(MAX_TX_DEPTH, bsize);
mbmi->tx_size = (TX_SIZE)TXSIZEMAX(mbmi->tx_size, min_tx_size);
}
if (is_inter_block(mbmi)) {
memset(mbmi->inter_tx_size, mbmi->tx_size, sizeof(mbmi->inter_tx_size));
}
memset(mbmi->txk_type, DCT_DCT, sizeof(mbmi->txk_type[0]) * TXK_TYPE_BUF_LEN);
av1_zero(x->blk_skip);
x->skip = 0;
}
static void update_state(const AV1_COMP *const cpi, TileDataEnc *tile_data,
ThreadData *td, PICK_MODE_CONTEXT *ctx, int mi_row,
int mi_col, BLOCK_SIZE bsize, RUN_TYPE dry_run) {
int i, x_idx, y;
const AV1_COMMON *const cm = &cpi->common;
const int num_planes = av1_num_planes(cm);
RD_COUNTS *const rdc = &td->rd_counts;
MACROBLOCK *const x = &td->mb;
MACROBLOCKD *const xd = &x->e_mbd;
struct macroblock_plane *const p = x->plane;
struct macroblockd_plane *const pd = xd->plane;
MB_MODE_INFO *mi = &ctx->mic;
MB_MODE_INFO *const mi_addr = xd->mi[0];
const struct segmentation *const seg = &cm->seg;
const int bw = mi_size_wide[mi->sb_type];
const int bh = mi_size_high[mi->sb_type];
const int mis = cm->mi_stride;
const int mi_width = mi_size_wide[bsize];
const int mi_height = mi_size_high[bsize];
assert(mi->sb_type == bsize);
*mi_addr = *mi;
*x->mbmi_ext = ctx->mbmi_ext;
reset_intmv_filter_type(mi_addr);
memcpy(x->blk_skip, ctx->blk_skip, sizeof(x->blk_skip[0]) * ctx->num_4x4_blk);
x->skip = ctx->skip;
// If segmentation in use
if (seg->enabled) {
// For in frame complexity AQ copy the segment id from the segment map.
if (cpi->oxcf.aq_mode == COMPLEXITY_AQ) {
const uint8_t *const map =
seg->update_map ? cpi->segmentation_map : cm->last_frame_seg_map;
mi_addr->segment_id =
map ? get_segment_id(cm, map, bsize, mi_row, mi_col) : 0;
reset_tx_size(x, mi_addr, cm->tx_mode);
}
// Else for cyclic refresh mode update the segment map, set the segment id
// and then update the quantizer.
if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ) {
av1_cyclic_refresh_update_segment(cpi, mi_addr, mi_row, mi_col, bsize,
ctx->rate, ctx->dist, x->skip);
reset_tx_size(x, mi_addr, cm->tx_mode);
}
if (mi_addr->uv_mode == UV_CFL_PRED && !is_cfl_allowed(xd))
mi_addr->uv_mode = UV_DC_PRED;
}
for (i = 0; i < num_planes; ++i) {
p[i].coeff = ctx->coeff[i];
p[i].qcoeff = ctx->qcoeff[i];
pd[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];
// Restore the coding context of the MB to that that was in place
// when the mode was picked for it
for (y = 0; y < mi_height; y++)
for (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) {
xd->mi[x_idx + y * mis] = mi_addr;
}
if (cpi->oxcf.aq_mode) av1_init_plane_quantizers(cpi, x, mi_addr->segment_id);
if (dry_run) return;
#if CONFIG_INTERNAL_STATS
{
unsigned int *const mode_chosen_counts =
(unsigned int *)cpi->mode_chosen_counts; // Cast const away.
if (frame_is_intra_only(cm)) {
static const int kf_mode_index[] = {
THR_DC /*DC_PRED*/,
THR_V_PRED /*V_PRED*/,
THR_H_PRED /*H_PRED*/,
THR_D45_PRED /*D45_PRED*/,
THR_D135_PRED /*D135_PRED*/,
THR_D113_PRED /*D113_PRED*/,
THR_D157_PRED /*D157_PRED*/,
THR_D203_PRED /*D203_PRED*/,
THR_D67_PRED /*D67_PRED*/,
THR_SMOOTH, /*SMOOTH_PRED*/
THR_SMOOTH_V, /*SMOOTH_V_PRED*/
THR_SMOOTH_H, /*SMOOTH_H_PRED*/
THR_PAETH /*PAETH_PRED*/,
};
++mode_chosen_counts[kf_mode_index[mi_addr->mode]];
} else {
// Note how often each mode chosen as best
++mode_chosen_counts[ctx->best_mode_index];
}
}
#endif
if (!frame_is_intra_only(cm)) {
if (is_inter_block(mi_addr)) {
// TODO(sarahparker): global motion stats need to be handled per-tile
// to be compatible with tile-based threading.
update_global_motion_used(mi_addr->mode, bsize, mi_addr, rdc);
}
if (cm->interp_filter == SWITCHABLE &&
mi_addr->motion_mode != WARPED_CAUSAL &&
!is_nontrans_global_motion(xd, xd->mi[0])) {
update_filter_type_count(tile_data->allow_update_cdf, td->counts, xd,
mi_addr);
}
rdc->comp_pred_diff[SINGLE_REFERENCE] += ctx->single_pred_diff;
rdc->comp_pred_diff[COMPOUND_REFERENCE] += ctx->comp_pred_diff;
rdc->comp_pred_diff[REFERENCE_MODE_SELECT] += ctx->hybrid_pred_diff;
}
const int x_mis = AOMMIN(bw, cm->mi_cols - mi_col);
const int y_mis = AOMMIN(bh, cm->mi_rows - mi_row);
av1_copy_frame_mvs(cm, mi, mi_row, mi_col, x_mis, y_mis);
}
void av1_setup_src_planes(MACROBLOCK *x, const YV12_BUFFER_CONFIG *src,
int mi_row, int mi_col, const int num_planes) {
// Set current frame pointer.
x->e_mbd.cur_buf = src;
// We use AOMMIN(num_planes, MAX_MB_PLANE) instead of num_planes to quiet
// the static analysis warnings.
for (int i = 0; i < AOMMIN(num_planes, MAX_MB_PLANE); i++) {
const int is_uv = i > 0;
setup_pred_plane(&x->plane[i].src, x->e_mbd.mi[0]->sb_type, src->buffers[i],
src->crop_widths[is_uv], src->crop_heights[is_uv],
src->strides[is_uv], mi_row, mi_col, NULL,
x->e_mbd.plane[i].subsampling_x,
x->e_mbd.plane[i].subsampling_y);
}
}
static int set_segment_rdmult(const AV1_COMP *const cpi, MACROBLOCK *const x,
int8_t segment_id) {
const AV1_COMMON *const cm = &cpi->common;
av1_init_plane_quantizers(cpi, x, segment_id);
aom_clear_system_state();
int segment_qindex = av1_get_qindex(&cm->seg, segment_id, cm->base_qindex);
return av1_compute_rd_mult(cpi, segment_qindex + cm->y_dc_delta_q);
}
static int set_deltaq_rdmult(const AV1_COMP *const cpi, MACROBLOCKD *const xd) {
const AV1_COMMON *const cm = &cpi->common;
return av1_compute_rd_mult(
cpi, cm->base_qindex + xd->delta_qindex + cm->y_dc_delta_q);
}
static void rd_pick_sb_modes(const 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,
int64_t best_rd) {
const AV1_COMMON *const cm = &cpi->common;
const int num_planes = av1_num_planes(cm);
TileInfo *const tile_info = &tile_data->tile_info;
MACROBLOCKD *const xd = &x->e_mbd;
MB_MODE_INFO *mbmi;
MB_MODE_INFO *ctx_mbmi = &ctx->mic;
struct macroblock_plane *const p = x->plane;
struct macroblockd_plane *const pd = xd->plane;
const AQ_MODE aq_mode = cpi->oxcf.aq_mode;
const DELTAQ_MODE deltaq_mode = cpi->oxcf.deltaq_mode;
int i, orig_rdmult;
aom_clear_system_state();
set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
mbmi = xd->mi[0];
if (ctx->rd_mode_is_ready) {
assert(ctx_mbmi->sb_type == bsize);
assert(ctx_mbmi->partition == partition);
*mbmi = *ctx_mbmi;
rd_cost->rate = ctx->rate;
rd_cost->dist = ctx->dist;
rd_cost->rdcost = ctx->rdcost;
} else {
mbmi->sb_type = bsize;
mbmi->partition = partition;
}
#if CONFIG_RD_DEBUG
mbmi->mi_row = mi_row;
mbmi->mi_col = mi_col;
#endif
for (i = 0; i < num_planes; ++i) {
p[i].coeff = ctx->coeff[i];
p[i].qcoeff = ctx->qcoeff[i];
pd[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];
if (!ctx->rd_mode_is_ready) {
ctx->skippable = 0;
// Set to zero to make sure we do not use the previous encoded frame stats
mbmi->skip = 0;
// Reset skip mode flag.
mbmi->skip_mode = 0;
}
x->skip_chroma_rd =
!is_chroma_reference(mi_row, mi_col, bsize, xd->plane[1].subsampling_x,
xd->plane[1].subsampling_y);
if (ctx->rd_mode_is_ready) {
x->skip = ctx->skip;
*x->mbmi_ext = ctx->mbmi_ext;
return;
}
if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
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);
}
// Save rdmult before it might be changed, so it can be restored later.
orig_rdmult = x->rdmult;
if (aq_mode == VARIANCE_AQ) {
if (cpi->vaq_refresh) {
const int energy =
bsize <= BLOCK_16X16 ? x->mb_energy : av1_block_energy(cpi, x, bsize);
mbmi->segment_id = av1_vaq_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);
}
if (deltaq_mode > 0) x->rdmult = set_deltaq_rdmult(cpi, xd);
// 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)) {
av1_rd_pick_intra_mode_sb(cpi, x, mi_row, mi_col, rd_cost, bsize, ctx,
best_rd);
} else {
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);
} else {
av1_rd_pick_inter_mode_sb(cpi, tile_data, x, mi_row, mi_col, rd_cost,
bsize, ctx, best_rd);
}
}
// 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) &&
(cm->frame_type == KEY_FRAME || cpi->refresh_alt_ref_frame ||
cpi->refresh_alt2_ref_frame ||
(cpi->refresh_golden_frame && !cpi->rc.is_src_frame_alt_ref))) {
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->rate = rd_cost->rate;
ctx->dist = rd_cost->dist;
ctx->rdcost = rd_cost->rdcost;
}
static void update_inter_mode_stats(FRAME_CONTEXT *fc, FRAME_COUNTS *counts,
PREDICTION_MODE mode, int16_t mode_context,
uint8_t allow_update_cdf) {
(void)counts;
int16_t mode_ctx = mode_context & NEWMV_CTX_MASK;
if (mode == NEWMV) {
#if CONFIG_ENTROPY_STATS
++counts->newmv_mode[mode_ctx][0];
#endif
if (allow_update_cdf) update_cdf(fc->newmv_cdf[mode_ctx], 0, 2);
return;
} else {
#if CONFIG_ENTROPY_STATS
++counts->newmv_mode[mode_ctx][1];
#endif
if (allow_update_cdf) update_cdf(fc->newmv_cdf[mode_ctx], 1, 2);
mode_ctx = (mode_context >> GLOBALMV_OFFSET) & GLOBALMV_CTX_MASK;
if (mode == GLOBALMV) {
#if CONFIG_ENTROPY_STATS
++counts->zeromv_mode[mode_ctx][0];
#endif
if (allow_update_cdf) update_cdf(fc->zeromv_cdf[mode_ctx], 0, 2);
return;
} else {
#if CONFIG_ENTROPY_STATS
++counts->zeromv_mode[mode_ctx][1];
#endif
if (allow_update_cdf) update_cdf(fc->zeromv_cdf[mode_ctx], 1, 2);
mode_ctx = (mode_context >> REFMV_OFFSET) & REFMV_CTX_MASK;
#if CONFIG_ENTROPY_STATS
++counts->refmv_mode[mode_ctx][mode != NEARESTMV];
#endif
if (allow_update_cdf)
update_cdf(fc->refmv_cdf[mode_ctx], mode != NEARESTMV, 2);
}
}
}
static void update_palette_cdf(MACROBLOCKD *xd, const MB_MODE_INFO *const mbmi,
FRAME_COUNTS *counts, uint8_t allow_update_cdf) {
FRAME_CONTEXT *fc = xd->tile_ctx;
const BLOCK_SIZE bsize = mbmi->sb_type;
const PALETTE_MODE_INFO *const pmi = &mbmi->palette_mode_info;
const int palette_bsize_ctx = av1_get_palette_bsize_ctx(bsize);
(void)counts;
if (mbmi->mode == DC_PRED) {
const int n = pmi->palette_size[0];
const int palette_mode_ctx = av1_get_palette_mode_ctx(xd);
#if CONFIG_ENTROPY_STATS
++counts->palette_y_mode[palette_bsize_ctx][palette_mode_ctx][n > 0];
#endif
if (allow_update_cdf)
update_cdf(fc->palette_y_mode_cdf[palette_bsize_ctx][palette_mode_ctx],
n > 0, 2);
if (n > 0) {
#if CONFIG_ENTROPY_STATS
++counts->palette_y_size[palette_bsize_ctx][n - PALETTE_MIN_SIZE];
#endif
if (allow_update_cdf) {
update_cdf(fc->palette_y_size_cdf[palette_bsize_ctx],
n - PALETTE_MIN_SIZE, PALETTE_SIZES);
}
}
}
if (mbmi->uv_mode == UV_DC_PRED) {
const int n = pmi->palette_size[1];
const int palette_uv_mode_ctx = (pmi->palette_size[0] > 0);
#if CONFIG_ENTROPY_STATS
++counts->palette_uv_mode[palette_uv_mode_ctx][n > 0];
#endif
if (allow_update_cdf)
update_cdf(fc->palette_uv_mode_cdf[palette_uv_mode_ctx], n > 0, 2);
if (n > 0) {
#if CONFIG_ENTROPY_STATS
++counts->palette_uv_size[palette_bsize_ctx][n - PALETTE_MIN_SIZE];
#endif
if (allow_update_cdf) {
update_cdf(fc->palette_uv_size_cdf[palette_bsize_ctx],
n - PALETTE_MIN_SIZE, PALETTE_SIZES);
}
}
}
}
static void sum_intra_stats(const AV1_COMMON *const cm, FRAME_COUNTS *counts,
MACROBLOCKD *xd, const MB_MODE_INFO *const mbmi,
const MB_MODE_INFO *above_mi,
const MB_MODE_INFO *left_mi, const int intraonly,
const int mi_row, const int mi_col,
uint8_t allow_update_cdf) {
FRAME_CONTEXT *fc = xd->tile_ctx;
const PREDICTION_MODE y_mode = mbmi->mode;
const UV_PREDICTION_MODE uv_mode = mbmi->uv_mode;
(void)counts;
const BLOCK_SIZE bsize = mbmi->sb_type;
if (intraonly) {
#if CONFIG_ENTROPY_STATS
const PREDICTION_MODE above = av1_above_block_mode(above_mi);
const PREDICTION_MODE left = av1_left_block_mode(left_mi);
const int above_ctx = intra_mode_context[above];
const int left_ctx = intra_mode_context[left];
++counts->kf_y_mode[above_ctx][left_ctx][y_mode];
#endif // CONFIG_ENTROPY_STATS
if (allow_update_cdf)
update_cdf(get_y_mode_cdf(fc, above_mi, left_mi), y_mode, INTRA_MODES);
} else {
#if CONFIG_ENTROPY_STATS
++counts->y_mode[size_group_lookup[bsize]][y_mode];
#endif // CONFIG_ENTROPY_STATS
if (allow_update_cdf)
update_cdf(fc->y_mode_cdf[size_group_lookup[bsize]], y_mode, INTRA_MODES);
}
if (av1_filter_intra_allowed(cm, mbmi)) {
const int use_filter_intra_mode =
mbmi->filter_intra_mode_info.use_filter_intra;
#if CONFIG_ENTROPY_STATS
++counts->filter_intra[mbmi->sb_type][use_filter_intra_mode];
if (use_filter_intra_mode) {
++counts
->filter_intra_mode[mbmi->filter_intra_mode_info.filter_intra_mode];
}
#endif // CONFIG_ENTROPY_STATS
if (allow_update_cdf) {
update_cdf(fc->filter_intra_cdfs[mbmi->sb_type], use_filter_intra_mode,
2);
if (use_filter_intra_mode) {
update_cdf(fc->filter_intra_mode_cdf,
mbmi->filter_intra_mode_info.filter_intra_mode,
FILTER_INTRA_MODES);
}
}
}
if (av1_is_directional_mode(mbmi->mode) && av1_use_angle_delta(bsize)) {
#if CONFIG_ENTROPY_STATS
++counts->angle_delta[mbmi->mode - V_PRED]
[mbmi->angle_delta[PLANE_TYPE_Y] + MAX_ANGLE_DELTA];
#endif
if (allow_update_cdf) {
update_cdf(fc->angle_delta_cdf[mbmi->mode - V_PRED],
mbmi->angle_delta[PLANE_TYPE_Y] + MAX_ANGLE_DELTA,
2 * MAX_ANGLE_DELTA + 1);
}
}
if (!is_chroma_reference(mi_row, mi_col, bsize,
xd->plane[AOM_PLANE_U].subsampling_x,
xd->plane[AOM_PLANE_U].subsampling_y))
return;
#if CONFIG_ENTROPY_STATS
++counts->uv_mode[is_cfl_allowed(xd)][y_mode][uv_mode];
#endif // CONFIG_ENTROPY_STATS
if (allow_update_cdf) {
const CFL_ALLOWED_TYPE cfl_allowed = is_cfl_allowed(xd);
update_cdf(fc->uv_mode_cdf[cfl_allowed][y_mode], uv_mode,
UV_INTRA_MODES - !cfl_allowed);
}
if (uv_mode == UV_CFL_PRED) {
const int joint_sign = mbmi->cfl_alpha_signs;
const int idx = mbmi->cfl_alpha_idx;
#if CONFIG_ENTROPY_STATS
++counts->cfl_sign[joint_sign];
#endif
if (allow_update_cdf)
update_cdf(fc->cfl_sign_cdf, joint_sign, CFL_JOINT_SIGNS);
if (CFL_SIGN_U(joint_sign) != CFL_SIGN_ZERO) {
aom_cdf_prob *cdf_u = fc->cfl_alpha_cdf[CFL_CONTEXT_U(joint_sign)];
#if CONFIG_ENTROPY_STATS
++counts->cfl_alpha[CFL_CONTEXT_U(joint_sign)][CFL_IDX_U(idx)];
#endif
if (allow_update_cdf)
update_cdf(cdf_u, CFL_IDX_U(idx), CFL_ALPHABET_SIZE);
}
if (CFL_SIGN_V(joint_sign) != CFL_SIGN_ZERO) {
aom_cdf_prob *cdf_v = fc->cfl_alpha_cdf[CFL_CONTEXT_V(joint_sign)];
#if CONFIG_ENTROPY_STATS
++counts->cfl_alpha[CFL_CONTEXT_V(joint_sign)][CFL_IDX_V(idx)];
#endif
if (allow_update_cdf)
update_cdf(cdf_v, CFL_IDX_V(idx), CFL_ALPHABET_SIZE);
}
}
if (av1_is_directional_mode(get_uv_mode(uv_mode)) &&
av1_use_angle_delta(bsize)) {
#if CONFIG_ENTROPY_STATS
++counts->angle_delta[uv_mode - UV_V_PRED]
[mbmi->angle_delta[PLANE_TYPE_UV] + MAX_ANGLE_DELTA];
#endif
if (allow_update_cdf) {
update_cdf(fc->angle_delta_cdf[uv_mode - UV_V_PRED],
mbmi->angle_delta[PLANE_TYPE_UV] + MAX_ANGLE_DELTA,
2 * MAX_ANGLE_DELTA + 1);
}
}
if (av1_allow_palette(cm->allow_screen_content_tools, bsize))
update_palette_cdf(xd, mbmi, counts, allow_update_cdf);
}
static void update_stats(const AV1_COMMON *const cm, TileDataEnc *tile_data,
ThreadData *td, int mi_row, int mi_col) {
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 BLOCK_SIZE bsize = mbmi->sb_type;
FRAME_CONTEXT *fc = xd->tile_ctx;
const uint8_t allow_update_cdf = tile_data->allow_update_cdf;
// 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 int seg_ref_active =
segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_REF_FRAME);
if (cm->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
if (allow_update_cdf)
update_cdf(fc->skip_mode_cdfs[skip_mode_ctx], mbmi->skip_mode, 2);
}
if (!mbmi->skip_mode) {
if (!seg_ref_active) {
const int skip_ctx = av1_get_skip_context(xd);
#if CONFIG_ENTROPY_STATS
td->counts->skip[skip_ctx][mbmi->skip]++;
#endif
if (allow_update_cdf) update_cdf(fc->skip_cdfs[skip_ctx], mbmi->skip, 2);
}
}
if (cm->delta_q_present_flag &&
(bsize != cm->seq_params.sb_size || !mbmi->skip) &&
super_block_upper_left) {
#if CONFIG_ENTROPY_STATS
const int dq =
(mbmi->current_qindex - xd->current_qindex) / cm->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]++;
#endif
xd->current_qindex = mbmi->current_qindex;
if (cm->delta_lf_present_flag) {
if (cm->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) {
#if CONFIG_ENTROPY_STATS
const int delta_lf =
(mbmi->delta_lf[lf_id] - xd->delta_lf[lf_id]) / cm->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]++;
#endif
xd->delta_lf[lf_id] = mbmi->delta_lf[lf_id];
}
} else {
#if CONFIG_ENTROPY_STATS
const int delta_lf =
(mbmi->delta_lf_from_base - xd->delta_lf_from_base) /
cm->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
xd->delta_lf_from_base = mbmi->delta_lf_from_base;
}
}
}
if (!is_inter_block(mbmi)) {
sum_intra_stats(cm, td->counts, xd, mbmi, xd->above_mbmi, xd->left_mbmi,
frame_is_intra_only(cm), mi_row, mi_col,
tile_data->allow_update_cdf);
}
if (av1_allow_intrabc(cm)) {
if (allow_update_cdf)
update_cdf(fc->intrabc_cdf, is_intrabc_block(mbmi), 2);
#if CONFIG_ENTROPY_STATS
++td->counts->intrabc[is_intrabc_block(mbmi)];
#endif // CONFIG_ENTROPY_STATS
}
if (!frame_is_intra_only(cm)) {
RD_COUNTS *rdc = &td->rd_counts;
FRAME_COUNTS *const counts = td->counts;
if (mbmi->skip_mode) {
rdc->skip_mode_used_flag = 1;
if (cm->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]);
return;
}
const int inter_block = is_inter_block(mbmi);
if (!seg_ref_active) {
#if CONFIG_ENTROPY_STATS
counts->intra_inter[av1_get_intra_inter_context(xd)][inter_block]++;
#endif
if (allow_update_cdf) {
update_cdf(fc->intra_inter_cdf[av1_get_intra_inter_context(xd)],
inter_block, 2);
}
// 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];
av1_collect_neighbors_ref_counts(xd);
if (cm->reference_mode == REFERENCE_MODE_SELECT) {
if (has_second_ref(mbmi))
// This flag is also updated for 4x4 blocks
rdc->compound_ref_used_flag = 1;
if (is_comp_ref_allowed(bsize)) {
#if CONFIG_ENTROPY_STATS
counts->comp_inter[av1_get_reference_mode_context(xd)]
[has_second_ref(mbmi)]++;
#endif // CONFIG_ENTROPY_STATS
if (allow_update_cdf) {
update_cdf(av1_get_reference_mode_cdf(xd), has_second_ref(mbmi),
2);
}
}
}
if (has_second_ref(mbmi)) {
const COMP_REFERENCE_TYPE comp_ref_type = has_uni_comp_refs(mbmi)
? UNIDIR_COMP_REFERENCE
: BIDIR_COMP_REFERENCE;
if (allow_update_cdf) {
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);
if (allow_update_cdf)
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);
if (allow_update_cdf)
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) {
if (allow_update_cdf) {
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);
if (allow_update_cdf)
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) {
if (allow_update_cdf) {
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 {
if (allow_update_cdf) {
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
}
if (allow_update_cdf) {
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) {
if (allow_update_cdf) {
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
}
}
} else {
const int bit = (ref0 >= BWDREF_FRAME);
if (allow_update_cdf)
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);
if (allow_update_cdf) {
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) {
if (allow_update_cdf) {
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);
if (allow_update_cdf)
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) {
if (allow_update_cdf) {
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 {
if (allow_update_cdf) {
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
}
}
}
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
if (allow_update_cdf)
update_cdf(fc->interintra_cdf[bsize_group], 1, 2);
#if CONFIG_ENTROPY_STATS
counts->interintra_mode[bsize_group][mbmi->interintra_mode]++;
#endif
if (allow_update_cdf) {
update_cdf(fc->interintra_mode_cdf[bsize_group],
mbmi->interintra_mode, INTERINTRA_MODES);
}
if (is_interintra_wedge_used(bsize)) {
#if CONFIG_ENTROPY_STATS
counts->wedge_interintra[bsize][mbmi->use_wedge_interintra]++;
#endif
if (allow_update_cdf) {
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
if (allow_update_cdf) {
update_cdf(fc->wedge_idx_cdf[bsize],
mbmi->interintra_wedge_index, 16);
}
}
}
} else {
#if CONFIG_ENTROPY_STATS
counts->interintra[bsize_group][0]++;
#endif
if (allow_update_cdf)
update_cdf(fc->interintra_cdf[bsize_group], 0, 2);
}
}
set_ref_ptrs(cm, xd, mbmi->ref_frame[0], mbmi->ref_frame[1]);
const MOTION_MODE motion_allowed =
cm->switchable_motion_mode
? motion_mode_allowed(xd->global_motion, xd, mbmi,
cm->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
if (allow_update_cdf) {
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
if (allow_update_cdf) {
update_cdf(fc->obmc_cdf[bsize], mbmi->motion_mode == OBMC_CAUSAL,
2);
}
}
}
if (has_second_ref(mbmi)) {
assert(cm->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(xd);
#if CONFIG_ENTROPY_STATS
++counts->comp_group_idx[comp_group_idx_ctx][mbmi->comp_group_idx];
#endif
if (allow_update_cdf) {
update_cdf(fc->comp_group_idx_cdf[comp_group_idx_ctx],
mbmi->comp_group_idx, 2);
}
}
if (mbmi->comp_group_idx == 0) {
const int comp_index_ctx = get_comp_index_context(cm, xd);
#if CONFIG_ENTROPY_STATS
++counts->compound_index[comp_index_ctx][mbmi->compound_idx];
#endif
if (allow_update_cdf) {
update_cdf(fc->compound_index_cdf[comp_index_ctx],
mbmi->compound_idx, 2);
}
} else {
assert(masked_compound_used);
if (is_interinter_compound_used(COMPOUND_WEDGE, bsize)) {
#if CONFIG_ENTROPY_STATS
++counts->compound_type[bsize][mbmi->interinter_comp.type - 1];
#endif
if (allow_update_cdf) {
update_cdf(fc->compound_type_cdf[bsize],
mbmi->interinter_comp.type - 1, COMPOUND_TYPES - 1);
}
}
}
}
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
if (allow_update_cdf) {
update_cdf(fc->wedge_idx_cdf[bsize],
mbmi->interinter_comp.wedge_index, 16);
}
}
}
}
}
if (inter_block &&
!segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP)) {
int16_t mode_ctx;
const PREDICTION_MODE mode = mbmi->mode;
mode_ctx =
av1_mode_context_analyzer(mbmi_ext->mode_context, mbmi->ref_frame);
if (has_second_ref(mbmi)) {
#if CONFIG_ENTROPY_STATS
++counts->inter_compound_mode[mode_ctx][INTER_COMPOUND_OFFSET(mode)];
#endif
if (allow_update_cdf)
update_cdf(fc->inter_compound_mode_cdf[mode_ctx],
INTER_COMPOUND_OFFSET(mode), INTER_COMPOUND_MODES);
} else {
update_inter_mode_stats(fc, counts, mode, mode_ctx, allow_update_cdf);
}
int mode_allowed = (mbmi->mode == NEWMV);
mode_allowed |= (mbmi->mode == NEW_NEWMV);
if (mode_allowed) {
uint8_t ref_frame_type = av1_ref_frame_type(mbmi->ref_frame);
int idx;
for (idx = 0; idx < 2; ++idx) {
if (mbmi_ext->ref_mv_count[ref_frame_type] > idx + 1) {
#if CONFIG_ENTROPY_STATS
uint8_t drl_ctx =
av1_drl_ctx(mbmi_ext->ref_mv_stack[ref_frame_type], idx);
++counts->drl_mode[drl_ctx][mbmi->ref_mv_idx != idx];
#endif
if (mbmi->ref_mv_idx == idx) break;
}
}
}
if (have_nearmv_in_inter_mode(mbmi->mode)) {
uint8_t ref_frame_type = av1_ref_frame_type(mbmi->ref_frame);
int idx;
for (idx = 1; idx < 3; ++idx) {
if (mbmi_ext->ref_mv_count[ref_frame_type] > idx + 1) {
#if CONFIG_ENTROPY_STATS
uint8_t drl_ctx =
av1_drl_ctx(mbmi_ext->ref_mv_stack[ref_frame_type], idx);
++counts->drl_mode[drl_ctx][mbmi->ref_mv_idx != idx - 1];
#endif
if (mbmi->ref_mv_idx == idx - 1) break;
}
}
}
}
}
}
typedef struct {
ENTROPY_CONTEXT a[MAX_MIB_SIZE * MAX_MB_PLANE];
ENTROPY_CONTEXT l[MAX_MIB_SIZE * MAX_MB_PLANE];
PARTITION_CONTEXT sa[MAX_MIB_SIZE];
PARTITION_CONTEXT sl[MAX_MIB_SIZE];
TXFM_CONTEXT *p_ta;
TXFM_CONTEXT *p_tl;
TXFM_CONTEXT ta[MAX_MIB_SIZE];
TXFM_CONTEXT tl[MAX_MIB_SIZE];
} RD_SEARCH_MACROBLOCK_CONTEXT;
static void restore_context(MACROBLOCK *x,
const RD_SEARCH_MACROBLOCK_CONTEXT *ctx, int mi_row,
int mi_col, BLOCK_SIZE bsize,
const int num_planes) {
MACROBLOCKD *xd = &x->e_mbd;
int p;
const int num_4x4_blocks_wide =
block_size_wide[bsize] >> tx_size_wide_log2[0];
const int num_4x4_blocks_high =
block_size_high[bsize] >> tx_size_high_log2[0];
int mi_width = mi_size_wide[bsize];
int mi_height = mi_size_high[bsize];
for (p = 0; p < num_planes; p++) {
int tx_col = mi_col;
int tx_row = mi_row & MAX_MIB_MASK;
memcpy(xd->above_context[p] + (tx_col >> xd->plane[p].subsampling_x),
ctx->a + num_4x4_blocks_wide * p,
(sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_wide) >>
xd->plane[p].subsampling_x);
memcpy(xd->left_context[p] + (tx_row >> xd->plane[p].subsampling_y),
ctx->l + num_4x4_blocks_high * p,
(sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_high) >>
xd->plane[p].subsampling_y);
}
memcpy(xd->above_seg_context + mi_col, ctx->sa,
sizeof(*xd->above_seg_context) * mi_width);
memcpy(xd->left_seg_context + (mi_row & MAX_MIB_MASK), ctx->sl,
sizeof(xd->left_seg_context[0]) * mi_height);
xd->above_txfm_context = ctx->p_ta;
xd->left_txfm_context = ctx->p_tl;
memcpy(xd->above_txfm_context, ctx->ta,
sizeof(*xd->above_txfm_context) * mi_width);
memcpy(xd->left_txfm_context, ctx->tl,
sizeof(*xd->left_txfm_context) * mi_height);
}
static void save_context(const MACROBLOCK *x, RD_SEARCH_MACROBLOCK_CONTEXT *ctx,
int mi_row, int mi_col, BLOCK_SIZE bsize,
const int num_planes) {
const MACROBLOCKD *xd = &x->e_mbd;
int p;
const int num_4x4_blocks_wide =
block_size_wide[bsize] >> tx_size_wide_log2[0];
const int num_4x4_blocks_high =
block_size_high[bsize] >> tx_size_high_log2[0];
int mi_width = mi_size_wide[bsize];
int mi_height = mi_size_high[bsize];
// buffer the above/left context information of the block in search.
for (p = 0; p < num_planes; ++p) {
int tx_col = mi_col;
int tx_row = mi_row & MAX_MIB_MASK;
memcpy(ctx->a + num_4x4_blocks_wide * p,
xd->above_context[p] + (tx_col >> xd->plane[p].subsampling_x),
(sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_wide) >>
xd->plane[p].subsampling_x);
memcpy(ctx->l + num_4x4_blocks_high * p,
xd->left_context[p] + (tx_row >> xd->plane[p].subsampling_y),
(sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_high) >>
xd->plane[p].subsampling_y);
}
memcpy(ctx->sa, xd->above_seg_context + mi_col,
sizeof(*xd->above_seg_context) * mi_width);
memcpy(ctx->sl, xd->left_seg_context + (mi_row & MAX_MIB_MASK),
sizeof(xd->left_seg_context[0]) * mi_height);
memcpy(ctx->ta, xd->above_txfm_context,
sizeof(*xd->above_txfm_context) * mi_width);
memcpy(ctx->tl, xd->left_txfm_context,
sizeof(*xd->left_txfm_context) * mi_height);
ctx->p_ta = xd->above_txfm_context;
ctx->p_tl = xd->left_txfm_context;
}
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 *ctx,
int *rate) {
TileInfo *const tile = &tile_data->tile_info;
MACROBLOCK *const x = &td->mb;
MACROBLOCKD *xd = &x->e_mbd;
set_offsets(cpi, tile, x, mi_row, mi_col, bsize);
MB_MODE_INFO *mbmi = xd->mi[0];
mbmi->partition = partition;
update_state(cpi, tile_data, td, ctx, mi_row, mi_col, bsize, dry_run);
if (!dry_run) av1_set_coeff_buffer(cpi, x, mi_row, mi_col);
encode_superblock(cpi, tile_data, td, tp, dry_run, mi_row, mi_col, bsize,
rate);
if (dry_run == 0)
x->cb_offset += block_size_wide[bsize] * block_size_high[bsize];
if (!dry_run) {
if (bsize == cpi->common.seq_params.sb_size && mbmi->skip == 1 &&
cpi->common.delta_lf_present_flag) {
const int frame_lf_count = av1_num_planes(&cpi->common) > 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->compound_idx == 0 ||
mbmi->interinter_comp.type == COMPOUND_AVERAGE)
mbmi->comp_group_idx = 0;
else
mbmi->comp_group_idx = 1;
}
update_stats(&cpi->common, tile_data, td, mi_row, mi_col);
}
}
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) {
const AV1_COMMON *const cm = &cpi->common;
MACROBLOCK *const x = &td->mb;
MACROBLOCKD *const xd = &x->e_mbd;
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 >= cm->mi_rows || mi_col >= cm->mi_cols) return;
if (!dry_run && ctx >= 0) {
const int has_rows = (mi_row + hbs) < cm->mi_rows;
const int has_cols = (mi_col + hbs) < cm->mi_cols;
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;
update_cdf(fc->partition_cdf[ctx], partition,
partition_cdf_length(bsize));
}
}
}
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 < cm->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 < cm->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 < 4; ++i) {
int this_mi_row = mi_row + i * quarter_step;
if (i > 0 && this_mi_row >= cm->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 < 4; ++i) {
int this_mi_col = mi_col + i * quarter_step;
if (i > 0 && this_mi_col >= cm->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);
}
// Check to see if the given partition size is allowed for a specified number
// of mi block rows and columns remaining in the image.
// If not then return the largest allowed partition size
static BLOCK_SIZE find_partition_size(BLOCK_SIZE bsize, int rows_left,
int cols_left, int *bh, int *bw) {
if (rows_left <= 0 || cols_left <= 0) {
return AOMMIN(bsize, BLOCK_8X8);
} else {
for (; bsize > 0; bsize -= 3) {
*bh = mi_size_high[bsize];
*bw = mi_size_wide[bsize];
if ((*bh <= rows_left) && (*bw <= cols_left)) {
break;
}
}
}
return bsize;
}
static void set_partial_sb_partition(const AV1_COMMON *const cm,
MB_MODE_INFO *mi, int bh_in, int bw_in,
int mi_rows_remaining,
int mi_cols_remaining, BLOCK_SIZE bsize,
MB_MODE_INFO **mib) {
int bh = bh_in;
int r, c;
for (r = 0; r < cm->seq_params.mib_size; r += bh) {
int bw = bw_in;
for (c = 0; c < cm->seq_params.mib_size; c += bw) {
const int index = r * cm->mi_stride + c;
mib[index] = mi + index;
mib[index]->sb_type = find_partition_size(
bsize, mi_rows_remaining - r, mi_cols_remaining - c, &bh, &bw);
}
}
}
// This function attempts to set all mode info entries in a given superblock
// to the same block partition size.
// However, at the bottom and right borders of the image the requested size
// may not be allowed in which case this code attempts to choose the largest
// allowable partition.
static void set_fixed_partitioning(AV1_COMP *cpi, const TileInfo *const tile,
MB_MODE_INFO **mib, int mi_row, int mi_col,
BLOCK_SIZE bsize) {
AV1_COMMON *const cm = &cpi->common;
const int mi_rows_remaining = tile->mi_row_end - mi_row;
const int mi_cols_remaining = tile->mi_col_end - mi_col;
int block_row, block_col;
MB_MODE_INFO *const mi_upper_left = cm->mi + mi_row * cm->mi_stride + mi_col;
int bh = mi_size_high[bsize];
int bw = mi_size_wide[bsize];
assert((mi_rows_remaining > 0) && (mi_cols_remaining > 0));
// Apply the requested partition size to the SB if it is all "in image"
if ((mi_cols_remaining >= cm->seq_params.mib_size) &&
(mi_rows_remaining >= cm->seq_params.mib_size)) {
for (block_row = 0; block_row < cm->seq_params.mib_size; block_row += bh) {
for (block_col = 0; block_col < cm->seq_params.mib_size;
block_col += bw) {
int index = block_row * cm->mi_stride + block_col;
mib[index] = mi_upper_left + index;
mib[index]->sb_type = bsize;
}
}
} else {
// Else this is a partial SB.
set_partial_sb_partition(cm, mi_upper_left, bh, bw, mi_rows_remaining,
mi_cols_remaining, bsize, mib);
}
}
static void 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 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 int bs = mi_size_wide[bsize];
const int hbs = bs / 2;
int i;
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, 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, none_rdc, chosen_rdc;
BLOCK_SIZE sub_subsize = BLOCK_4X4;
int splits_below = 0;
BLOCK_SIZE bs_type = mib[0]->sb_type;
int do_partition_search = 1;
PICK_MODE_CONTEXT *ctx_none = &pc_tree->none;
if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;
assert(mi_size_wide[bsize] == mi_size_high[bsize]);
av1_invalid_rd_stats(&last_part_rdc);
av1_invalid_rd_stats(&none_rdc);
av1_invalid_rd_stats(&chosen_rdc);
pc_tree->partitioning = partition;
xd->above_txfm_context = cm->above_txfm_context[tile_info->tile_row] + mi_col;
xd->left_txfm_context =
xd->left_txfm_context_buffer + (mi_row & MAX_MIB_MASK);
save_context(x, &x_ctx, mi_row, mi_col, bsize, num_planes);
if (bsize == BLOCK_16X16 && cpi->vaq_refresh) {
set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
x->mb_energy = av1_block_energy(cpi, x, bsize);
}
if (do_partition_search &&
cpi->sf.partition_search_type == SEARCH_PARTITION &&
cpi->sf.adjust_partitioning_from_last_frame) {
// Check if any of the sub blocks are further split.
if (partition == PARTITION_SPLIT && subsize > BLOCK_8X8) {
sub_subsize = get_partition_subsize(subsize, PARTITION_SPLIT);
splits_below = 1;
for (i = 0; i < 4; i++) {
int jj = i >> 1, ii = i & 0x01;
MB_MODE_INFO *this_mi = mib[jj * hbs * cm->mi_stride + ii * hbs];
if (this_mi && this_mi->sb_type >= sub_subsize) {
splits_below = 0;
}
}
}
// If partition is not none try none unless each of the 4 splits are split
// even further..
if (partition != PARTITION_NONE && !splits_below &&
mi_row + hbs < cm->mi_rows && mi_col + hbs < cm->mi_cols) {
pc_tree->partitioning = PARTITION_NONE;
rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &none_rdc,
PARTITION_NONE, bsize, ctx_none, INT64_MAX);
if (none_rdc.rate < INT_MAX) {
none_rdc.rate += x->partition_cost[pl][PARTITION_NONE];
none_rdc.rdcost = RDCOST(x->rdmult, none_rdc.rate, none_rdc.dist);
}
restore_context(x, &x_ctx, mi_row, mi_col, bsize, num_planes);
mib[0]->sb_type = bs_type;
pc_tree->partitioning = partition;
}
}
switch (partition) {
case PARTITION_NONE:
rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &last_part_rdc,
PARTITION_NONE, bsize, ctx_none, INT64_MAX);
break;
case PARTITION_HORZ:
rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &last_part_rdc,
PARTITION_HORZ, subsize, &pc_tree->horizontal[0],
INT64_MAX);
if (last_part_rdc.rate != INT_MAX && bsize >= BLOCK_8X8 &&
mi_row + hbs < cm->mi_rows) {
RD_STATS tmp_rdc;
PICK_MODE_CONTEXT *ctx_h = &pc_tree->horizontal[0];
av1_init_rd_stats(&tmp_rdc);
update_state(cpi, tile_data, td, ctx_h, mi_row, mi_col, subsize, 1);
encode_superblock(cpi, tile_data, td, tp, DRY_RUN_NORMAL, mi_row,
mi_col, subsize, NULL);
rd_pick_sb_modes(cpi, tile_data, x, mi_row + hbs, mi_col, &tmp_rdc,
PARTITION_HORZ, subsize, &pc_tree->horizontal[1],
INT64_MAX);
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:
rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &last_part_rdc,
PARTITION_VERT, subsize, &pc_tree->vertical[0],
INT64_MAX);
if (last_part_rdc.rate != INT_MAX && bsize >= BLOCK_8X8 &&
mi_col + hbs < cm->mi_cols) {
RD_STATS tmp_rdc;
PICK_MODE_CONTEXT *ctx_v = &pc_tree->vertical[0];
av1_init_rd_stats(&tmp_rdc);
update_state(cpi, tile_data, td, ctx_v, mi_row, mi_col, subsize, 1);
encode_superblock(cpi, tile_data, td, tp, DRY_RUN_NORMAL, mi_row,
mi_col, subsize, NULL);
rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col + hbs, &tmp_rdc,
PARTITION_VERT, subsize,
&pc_tree->vertical[bsize > BLOCK_8X8], INT64_MAX);
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 (i = 0; i < 4; 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 >= cm->mi_rows) || (mi_col + x_idx >= cm->mi_cols))
continue;
av1_init_rd_stats(&tmp_rdc);
rd_use_partition(cpi, td, tile_data,
mib + jj * hbs * cm->mi_stride + ii * hbs, tp,
mi_row + y_idx, mi_col + x_idx, subsize, &tmp_rdc.rate,
&tmp_rdc.dist, i != 3, 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 partiton types");
default: assert(0); break;
}
if (last_part_rdc.rate < INT_MAX) {
last_part_rdc.rate += x->partition_cost[pl][partition];
last_part_rdc.rdcost =
RDCOST(x->rdmult, last_part_rdc.rate, last_part_rdc.dist);
}
if (do_partition_search && cpi->sf.adjust_partitioning_from_last_frame &&
cpi->sf.partition_search_type == SEARCH_PARTITION &&
partition != PARTITION_SPLIT && bsize > BLOCK_8X8 &&
(mi_row + bs < cm->mi_rows || mi_row + hbs == cm->mi_rows) &&
(mi_col + bs < cm->mi_cols || mi_col + hbs == cm->mi_cols)) {
BLOCK_SIZE split_subsize = get_partition_subsize(bsize, PARTITION_SPLIT);
chosen_rdc.rate = 0;
chosen_rdc.dist = 0;
restore_context(x, &x_ctx, mi_row, mi_col, bsize, num_planes);
pc_tree->partitioning = PARTITION_SPLIT;
// Split partition.
for (i = 0; i < 4; i++) {
int x_idx = (i & 1) * hbs;
int y_idx = (i >> 1) * hbs;
RD_STATS tmp_rdc;
if ((mi_row + y_idx >= cm->mi_rows) || (mi_col + x_idx >= cm->mi_cols))
continue;
save_context(x, &x_ctx, mi_row, mi_col, bsize, num_planes);
pc_tree->split[i]->partitioning = PARTITION_NONE;
rd_pick_sb_modes(cpi, tile_data, x, mi_row + y_idx, mi_col + x_idx,
&tmp_rdc, PARTITION_SPLIT, split_subsize,
&pc_tree->split[i]->none, INT64_MAX);
restore_context(x, &x_ctx, mi_row, mi_col, bsize, num_planes);
if (tmp_rdc.rate == INT_MAX || tmp_rdc.dist == INT64_MAX) {
av1_invalid_rd_stats(&chosen_rdc);
break;
}
chosen_rdc.rate += tmp_rdc.rate;
chosen_rdc.dist += tmp_rdc.dist;
if (i != 3)
encode_sb(cpi, td, tile_data, tp, mi_row + y_idx, mi_col + x_idx,
OUTPUT_ENABLED, split_subsize, pc_tree->split[i], NULL);
chosen_rdc.rate += x->partition_cost[pl][PARTITION_NONE];
}
if (chosen_rdc.rate < INT_MAX) {
chosen_rdc.rate += x->partition_cost[pl][PARTITION_SPLIT];
chosen_rdc.rdcost = RDCOST(x->rdmult, chosen_rdc.rate, chosen_rdc.dist);
}
}
// If last_part is better set the partitioning to that.
if (last_part_rdc.rdcost < chosen_rdc.rdcost) {
mib[0]->sb_type = bsize;
if (bsize >= BLOCK_8X8) pc_tree->partitioning = partition;
chosen_rdc = last_part_rdc;
}
// If none was better set the partitioning to that.
if (none_rdc.rdcost < chosen_rdc.rdcost) {
if (bsize >= BLOCK_8X8) pc_tree->partitioning = PARTITION_NONE;
chosen_rdc = none_rdc;
}
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(chosen_rdc.rate < INT_MAX && chosen_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 = 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 = chosen_rdc.rate;
*dist = chosen_rdc.dist;
}
/* clang-format off */
static const BLOCK_SIZE min_partition_size[BLOCK_SIZES_ALL] = {
BLOCK_4X4, // 4x4
BLOCK_4X4, BLOCK_4X4, BLOCK_4X4, // 4x8, 8x4, 8x8
BLOCK_4X4, BLOCK_4X4, BLOCK_8X8, // 8x16, 16x8, 16x16
BLOCK_8X8, BLOCK_8X8, BLOCK_16X16, // 16x32, 32x16, 32x32
BLOCK_16X16, BLOCK_16X16, BLOCK_16X16, // 32x64, 64x32, 64x64
BLOCK_16X16, BLOCK_16X16, BLOCK_16X16, // 64x128, 128x64, 128x128
BLOCK_4X4, BLOCK_4X4, BLOCK_8X8, // 4x16, 16x4, 8x32
BLOCK_8X8, BLOCK_16X16, BLOCK_16X16, // 32x8, 16x64, 64x16
};
static const BLOCK_SIZE max_partition_size[BLOCK_SIZES_ALL] = {
BLOCK_8X8, // 4x4
BLOCK_16X16, BLOCK_16X16, BLOCK_16X16, // 4x8, 8x4, 8x8
BLOCK_32X32, BLOCK_32X32, BLOCK_32X32, // 8x16, 16x8, 16x16
BLOCK_64X64, BLOCK_64X64, BLOCK_64X64, // 16x32, 32x16, 32x32
BLOCK_LARGEST, BLOCK_LARGEST, BLOCK_LARGEST, // 32x64, 64x32, 64x64
BLOCK_LARGEST, BLOCK_LARGEST, BLOCK_LARGEST, // 64x128, 128x64, 128x128
BLOCK_16X16, BLOCK_16X16, BLOCK_32X32, // 4x16, 16x4, 8x32
BLOCK_32X32, BLOCK_LARGEST, BLOCK_LARGEST, // 32x8, 16x64, 64x16
};
// Next square block size less or equal than current block size.
static const BLOCK_SIZE next_square_size[BLOCK_SIZES_ALL] = {
BLOCK_4X4, // 4x4
BLOCK_4X4, BLOCK_4X4, BLOCK_8X8, // 4x8, 8x4, 8x8
BLOCK_8X8, BLOCK_8X8, BLOCK_16X16, // 8x16, 16x8, 16x16
BLOCK_16X16, BLOCK_16X16, BLOCK_32X32, // 16x32, 32x16, 32x32
BLOCK_32X32, BLOCK_32X32, BLOCK_64X64, // 32x64, 64x32, 64x64
BLOCK_64X64, BLOCK_64X64, BLOCK_128X128, // 64x128, 128x64, 128x128
BLOCK_4X4, BLOCK_4X4, BLOCK_8X8, // 4x16, 16x4, 8x32
BLOCK_8X8, BLOCK_16X16, BLOCK_16X16, // 32x8, 16x64, 64x16
};
/* clang-format on */
// Look at all the mode_info entries for blocks that are part of this
// partition and find the min and max values for sb_type.
// At the moment this is designed to work on a superblock but could be
// adjusted to use a size parameter.
//
// The min and max are assumed to have been initialized prior to calling this
// function so repeat calls can accumulate a min and max of more than one
// superblock.
static void get_sb_partition_size_range(const AV1_COMMON *const cm,
MACROBLOCKD *xd, MB_MODE_INFO **mib,
BLOCK_SIZE *min_block_size,
BLOCK_SIZE *max_block_size) {
int i, j;
int index = 0;
// Check the sb_type for each block that belongs to this region.
for (i = 0; i < cm->seq_params.mib_size; ++i) {
for (j = 0; j < cm->seq_params.mib_size; ++j) {
MB_MODE_INFO *mi = mib[index + j];
BLOCK_SIZE sb_type = mi ? mi->sb_type : BLOCK_4X4;
*min_block_size = AOMMIN(*min_block_size, sb_type);
*max_block_size = AOMMAX(*max_block_size, sb_type);
}
index += xd->mi_stride;
}
}
// Checks to see if a super block is on a horizontal image edge.
// In most cases this is the "real" edge unless there are formatting
// bars embedded in the stream.
static int active_h_edge(const AV1_COMP *cpi, int mi_row, int mi_step) {
int top_edge = 0;
int bottom_edge = cpi->common.mi_rows;
int is_active_h_edge = 0;
// For two pass account for any formatting bars detected.
if (cpi->oxcf.pass == 2) {
const TWO_PASS *const twopass = &cpi->twopass;
// The inactive region is specified in MBs not mi units.
// The image edge is in the following MB row.
top_edge += (int)(twopass->this_frame_stats.inactive_zone_rows * 2);
bottom_edge -= (int)(twopass->this_frame_stats.inactive_zone_rows * 2);
bottom_edge = AOMMAX(top_edge, bottom_edge);
}
if (((top_edge >= mi_row) && (top_edge < (mi_row + mi_step))) ||
((bottom_edge >= mi_row) && (bottom_edge < (mi_row + mi_step)))) {
is_active_h_edge = 1;
}
return is_active_h_edge;
}
// Checks to see if a super block is on a vertical image edge.
// In most cases this is the "real" edge unless there are formatting
// bars embedded in the stream.
static int active_v_edge(const AV1_COMP *cpi, int mi_col, int mi_step) {
int left_edge = 0;
int right_edge = cpi->common.mi_cols;
int is_active_v_edge = 0;
// For two pass account for any formatting bars detected.
if (cpi->oxcf.pass == 2) {
const TWO_PASS *const twopass = &cpi->twopass;
// The inactive region is specified in MBs not mi units.
// The image edge is in the following MB row.
left_edge += (int)(twopass->this_frame_stats.inactive_zone_cols * 2);
right_edge -= (int)(twopass->this_frame_stats.inactive_zone_cols * 2);
right_edge = AOMMAX(left_edge, right_edge);
}
if (((left_edge >= mi_col) && (left_edge < (mi_col + mi_step))) ||
((right_edge >= mi_col) && (right_edge < (mi_col + mi_step)))) {
is_active_v_edge = 1;
}
return is_active_v_edge;
}
// Checks to see if a super block is at the edge of the active image.
// In most cases this is the "real" edge unless there are formatting
// bars embedded in the stream.
static int active_edge_sb(const AV1_COMP *cpi, int mi_row, int mi_col) {
return active_h_edge(cpi, mi_row, cpi->common.seq_params.mib_size) ||
active_v_edge(cpi, mi_col, cpi->common.seq_params.mib_size);
}
// Look at neighboring blocks and set a min and max partition size based on
// what they chose.
static void rd_auto_partition_range(AV1_COMP *cpi, const TileInfo *const tile,
MACROBLOCKD *const xd, int mi_row,
int mi_col, BLOCK_SIZE *min_block_size,
BLOCK_SIZE *max_block_size) {
AV1_COMMON *const cm = &cpi->common;
MB_MODE_INFO **mi = xd->mi;
const int left_in_image = xd->left_available && mi[-1];
const int above_in_image = xd->up_available && mi[-xd->mi_stride];
const int mi_rows_remaining = tile->mi_row_end - mi_row;
const int mi_cols_remaining = tile->mi_col_end - mi_col;
int bh, bw;
BLOCK_SIZE min_size = BLOCK_4X4;
BLOCK_SIZE max_size = BLOCK_LARGEST;
// Trap case where we do not have a prediction.
if (left_in_image || above_in_image || cm->frame_type != KEY_FRAME) {
// Default "min to max" and "max to min"
min_size = BLOCK_LARGEST;
max_size = BLOCK_4X4;
// NOTE: each call to get_sb_partition_size_range() uses the previous
// passed in values for min and max as a starting point.
// Find the min and max partition used in previous frame at this location
if (cm->frame_type != KEY_FRAME) {
MB_MODE_INFO **prev_mi =
&cm->prev_mi_grid_visible[mi_row * xd->mi_stride + mi_col];
get_sb_partition_size_range(cm, xd, prev_mi, &min_size, &max_size);
}
// Find the min and max partition sizes used in the left superblock
if (left_in_image) {
MB_MODE_INFO **left_sb_mi = &mi[-cm->seq_params.mib_size];
get_sb_partition_size_range(cm, xd, left_sb_mi, &min_size, &max_size);
}
// Find the min and max partition sizes used in the above suprblock.
if (above_in_image) {
MB_MODE_INFO **above_sb_mi =
&mi[-xd->mi_stride * cm->seq_params.mib_size];
get_sb_partition_size_range(cm, xd, above_sb_mi, &min_size, &max_size);
}
// Adjust observed min and max for "relaxed" auto partition case.
if (cpi->sf.auto_min_max_partition_size == RELAXED_NEIGHBORING_MIN_MAX) {
min_size = min_partition_size[min_size];
max_size = max_partition_size[max_size];
}
}
// Check border cases where max and min from neighbors may not be legal.
max_size = find_partition_size(max_size, mi_rows_remaining, mi_cols_remaining,
&bh, &bw);
min_size = AOMMIN(min_size, max_size);
// Test for blocks at the edge of the active image.
// This may be the actual edge of the image or where there are formatting
// bars.
if (active_edge_sb(cpi, mi_row, mi_col)) {
min_size = BLOCK_4X4;
} else {
min_size = AOMMIN(cpi->sf.rd_auto_partition_min_limit, min_size);
}
// When use_square_partition_only is true, make sure at least one square
// partition is allowed by selecting the next smaller square size as
// *min_block_size.
if (cpi->sf.use_square_partition_only) {
min_size = AOMMIN(min_size, next_square_size[max_size]);
}
*min_block_size = AOMMIN(min_size, cm->seq_params.sb_size);
*max_block_size = AOMMIN(max_size, cm->seq_params.sb_size);
}
// TODO(jingning) refactor functions setting partition search range
static void set_partition_range(const AV1_COMMON *const cm,
const MACROBLOCKD *const xd, int mi_row,
int mi_col, BLOCK_SIZE bsize,
BLOCK_SIZE *const min_bs,
BLOCK_SIZE *const max_bs) {
const int mi_width = mi_size_wide[bsize];
const int mi_height = mi_size_high[bsize];
int idx, idy;
const int idx_str = cm->mi_stride * mi_row + mi_col;
MB_MODE_INFO **const prev_mi = &cm->prev_mi_grid_visible[idx_str];
BLOCK_SIZE min_size = cm->seq_params.sb_size; // default values
BLOCK_SIZE max_size = BLOCK_4X4;
if (prev_mi) {
for (idy = 0; idy < mi_height; ++idy) {
for (idx = 0; idx < mi_width; ++idx) {
const MB_MODE_INFO *const mi = prev_mi[idy * cm->mi_stride + idx];
const BLOCK_SIZE bs = mi ? mi->sb_type : bsize;
min_size = AOMMIN(min_size, bs);
max_size = AOMMAX(max_size, bs);
}
}
}
if (xd->left_available) {
for (idy = 0; idy < mi_height; ++idy) {
const MB_MODE_INFO *const mi = xd->mi[idy * cm->mi_stride - 1];
const BLOCK_SIZE bs = mi ? mi->sb_type : bsize;
min_size = AOMMIN(min_size, bs);
max_size = AOMMAX(max_size, bs);
}
}
if (xd->up_available) {
for (idx = 0; idx < mi_width; ++idx) {
const MB_MODE_INFO *const mi = xd->mi[idx - cm->mi_stride];
const BLOCK_SIZE bs = mi ? mi->sb_type : bsize;
min_size = AOMMIN(min_size, bs);
max_size = AOMMAX(max_size, bs);
}
}
if (min_size == max_size) {
min_size = min_partition_size[min_size];
max_size = max_partition_size[max_size];
}
*min_bs = AOMMIN(min_size, cm->seq_params.sb_size);
*max_bs = AOMMIN(max_size, cm->seq_params.sb_size);
}
static INLINE void store_pred_mv(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx) {
memcpy(ctx->pred_mv, x->pred_mv, sizeof(x->pred_mv));
}
static INLINE void load_pred_mv(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx) {
memcpy(x->pred_mv, ctx->pred_mv, sizeof(x->pred_mv));
}
#if CONFIG_FP_MB_STATS
const int qindex_skip_threshold_lookup[BLOCK_SIZES] = {
0, 10, 10, 30, 40, 40, 60, 80, 80, 90, 100, 100, 120,
// TODO(debargha): What are the correct numbers here?
130, 130, 150
};
const int qindex_split_threshold_lookup[BLOCK_SIZES] = {
0, 3, 3, 7, 15, 15, 30, 40, 40, 60, 80, 80, 120,
// TODO(debargha): What are the correct numbers here?
160, 160, 240
};
const int complexity_16x16_blocks_threshold[BLOCK_SIZES] = {
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 4, 4, 6,
// TODO(debargha): What are the correct numbers here?
8, 8, 10
};
typedef enum {
MV_ZERO = 0,
MV_LEFT = 1,
MV_UP = 2,
MV_RIGHT = 3,
MV_DOWN = 4,
MV_INVALID
} MOTION_DIRECTION;
static INLINE MOTION_DIRECTION get_motion_direction_fp(uint8_t fp_byte) {
if (fp_byte & FPMB_MOTION_ZERO_MASK) {
return MV_ZERO;
} else if (fp_byte & FPMB_MOTION_LEFT_MASK) {
return MV_LEFT;
} else if (fp_byte & FPMB_MOTION_RIGHT_MASK) {
return MV_RIGHT;
} else if (fp_byte & FPMB_MOTION_UP_MASK) {
return MV_UP;
} else {
return MV_DOWN;
}
}
static INLINE int get_motion_inconsistency(MOTION_DIRECTION this_mv,
MOTION_DIRECTION that_mv) {
if (this_mv == that_mv) {
return 0;
} else {
return abs(this_mv - that_mv) == 2 ? 2 : 1;
}
}
#endif
// 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(const AV1_COMP *const cpi, ThreadData *td,
TileDataEnc *tile_data, TOKENEXTRA **tp,
int is_first, int is_last, int mi_row, int mi_col,
BLOCK_SIZE subsize, RD_STATS *best_rdc,
RD_STATS *sum_rdc, RD_STATS *this_rdc,
PARTITION_TYPE partition,
PICK_MODE_CONTEXT *prev_ctx,
PICK_MODE_CONTEXT *this_ctx) {
#define RTS_X_RATE_NOCOEF_ARG
#define RTS_MAX_RDCOST best_rdc->rdcost
MACROBLOCK *const x = &td->mb;
if (cpi->sf.adaptive_motion_search) load_pred_mv(x, prev_ctx);
// On the first time around, write the rd stats straight to sum_rdc. Also, we
// should treat sum_rdc as containing zeros (even if it doesn't) to avoid
// having to zero it at the start.
if (is_first) this_rdc = sum_rdc;
const int64_t spent_rdcost = is_first ? 0 : sum_rdc->rdcost;
const int64_t rdcost_remaining = best_rdc->rdcost - spent_rdcost;
rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, this_rdc,
RTS_X_RATE_NOCOEF_ARG partition, subsize, this_ctx,
rdcost_remaining);
if (!is_first) {
if (this_rdc->rate == INT_MAX) {
sum_rdc->rdcost = INT64_MAX;
} else {
sum_rdc->rate += this_rdc->rate;
sum_rdc->dist += this_rdc->dist;
sum_rdc->rdcost += this_rdc->rdcost;
}
}
if (sum_rdc->rdcost >= RTS_MAX_RDCOST) return 0;
if (!is_last) {
update_state(cpi, tile_data, td, this_ctx, mi_row, mi_col, subsize, 1);
encode_superblock(cpi, tile_data, td, tp, DRY_RUN_NORMAL, mi_row, mi_col,
subsize, NULL);
}
return 1;
#undef RTS_X_RATE_NOCOEF_ARG
#undef RTS_MAX_RDCOST
}
static void rd_test_partition3(const AV1_COMP *const cpi, ThreadData *td,
TileDataEnc *tile_data, TOKENEXTRA **tp,
PC_TREE *pc_tree, RD_STATS *best_rdc,
PICK_MODE_CONTEXT ctxs[3],
PICK_MODE_CONTEXT *ctx, int mi_row, int mi_col,
BLOCK_SIZE bsize, PARTITION_TYPE partition,
int mi_row0, int mi_col0, BLOCK_SIZE subsize0,
int mi_row1, int mi_col1, BLOCK_SIZE subsize1,
int mi_row2, int mi_col2, BLOCK_SIZE subsize2) {
MACROBLOCK *const x = &td->mb;
MACROBLOCKD *const xd = &x->e_mbd;
RD_STATS sum_rdc, this_rdc;
#define RTP_STX_TRY_ARGS
if (!rd_try_subblock(cpi, td, tile_data, tp, 1, 0, mi_row0, mi_col0, subsize0,
best_rdc, &sum_rdc, &this_rdc,
RTP_STX_TRY_ARGS partition, ctx, &ctxs[0]))
return;
if (!rd_try_subblock(cpi, td, tile_data, tp, 0, 0, mi_row1, mi_col1, subsize1,
best_rdc, &sum_rdc, &this_rdc,
RTP_STX_TRY_ARGS partition, &ctxs[0], &ctxs[1]))
return;
// With the new layout of mixed partitions for PARTITION_HORZ_B and
// PARTITION_VERT_B, the last subblock might start past halfway through the
// main block, so we might signal it even though the subblock lies strictly
// outside the image. In that case, we won't spend any bits coding it and the
// difference (obviously) doesn't contribute to the error.
const int try_block2 = 1;
if (try_block2 &&
!rd_try_subblock(cpi, td, tile_data, tp, 0, 1, mi_row2, mi_col2, subsize2,
best_rdc, &sum_rdc, &this_rdc,
RTP_STX_TRY_ARGS partition, &ctxs[1], &ctxs[2]))
return;
if (sum_rdc.rdcost >= best_rdc->rdcost) return;
int pl = partition_plane_context(xd, mi_row, mi_col, bsize);
sum_rdc.rate += x->partition_cost[pl][partition];
sum_rdc.rdcost = RDCOST(x->rdmult, sum_rdc.rate, sum_rdc.dist);
if (sum_rdc.rdcost >= best_rdc->rdcost) return;
*best_rdc = sum_rdc;
pc_tree->partitioning = partition;
#undef RTP_STX_TRY_ARGS
}
#if CONFIG_DIST_8X8
static int64_t dist_8x8_yuv(const AV1_COMP *const cpi, MACROBLOCK *const x,
uint8_t *src_plane_8x8[MAX_MB_PLANE],
uint8_t *dst_plane_8x8[MAX_MB_PLANE]) {
const AV1_COMMON *const cm = &cpi->common;
const int num_planes = av1_num_planes(cm);
MACROBLOCKD *const xd = &x->e_mbd;
int64_t dist_8x8, dist_8x8_uv, total_dist;
const int src_stride = x->plane[0].src.stride;
int plane;
const int dst_stride = xd->plane[0].dst.stride;
dist_8x8 =
av1_dist_8x8(cpi, x, src_plane_8x8[0], src_stride, dst_plane_8x8[0],
dst_stride, BLOCK_8X8, 8, 8, 8, 8, x->qindex)
<< 4;
// Compute chroma distortion for a luma 8x8 block
dist_8x8_uv = 0;
if (num_planes > 1) {
for (plane = 1; plane < MAX_MB_PLANE; ++plane) {
unsigned sse;
const int src_stride_uv = x->plane[plane].src.stride;
const int dst_stride_uv = xd->plane[plane].dst.stride;
const int ssx = xd->plane[plane].subsampling_x;
const int ssy = xd->plane[plane].subsampling_y;
const BLOCK_SIZE plane_bsize = get_plane_block_size(BLOCK_8X8, ssx, ssy);
cpi->fn_ptr[plane_bsize].vf(src_plane_8x8[plane], src_stride_uv,
dst_plane_8x8[plane], dst_stride_uv, &sse);
dist_8x8_uv += (int64_t)sse << 4;
}
}
return total_dist = dist_8x8 + dist_8x8_uv;
}
#endif // CONFIG_DIST_8X8
static void reset_partition(PC_TREE *pc_tree, BLOCK_SIZE bsize) {
pc_tree->partitioning = PARTITION_NONE;
pc_tree->cb_search_range = SEARCH_FULL_PLANE;
if (bsize >= BLOCK_8X8) {
BLOCK_SIZE subsize = get_partition_subsize(bsize, PARTITION_SPLIT);
for (int idx = 0; idx < 4; ++idx)
reset_partition(pc_tree->split[idx], subsize);
}
}
static void rd_pick_sqr_partition(const AV1_COMP *const cpi, ThreadData *td,
TileDataEnc *tile_data, TOKENEXTRA **tp,
int mi_row, int mi_col, BLOCK_SIZE bsize,
RD_STATS *rd_cost, int64_t best_rd,
PC_TREE *pc_tree, int64_t *none_rd) {
const AV1_COMMON *const cm = &cpi->common;
TileInfo *const tile_info = &tile_data->tile_info;
MACROBLOCK *const x = &td->mb;
MACROBLOCKD *const xd = &x->e_mbd;
const int mi_step = mi_size_wide[bsize] / 2;
RD_SEARCH_MACROBLOCK_CONTEXT x_ctx;
const TOKENEXTRA *const tp_orig = *tp;
PICK_MODE_CONTEXT *ctx_none = &pc_tree->none;
int tmp_partition_cost[PARTITION_TYPES];
BLOCK_SIZE subsize;
RD_STATS this_rdc, sum_rdc, best_rdc, pn_rdc;
const int bsize_at_least_8x8 = (bsize >= BLOCK_8X8);
int do_square_split = bsize_at_least_8x8;
const int pl = bsize_at_least_8x8
? partition_plane_context(xd, mi_row, mi_col, bsize)
: 0;
const int *partition_cost =
pl >= 0 ? x->partition_cost[pl] : x->partition_cost[0];
const int num_planes = av1_num_planes(cm);
int64_t split_rd[4] = { 0, 0, 0, 0 };
// Override skipping rectangular partition operations for edge blocks
const int has_rows = (mi_row + mi_step < cm->mi_rows);
const int has_cols = (mi_col + mi_step < cm->mi_cols);
if (none_rd) *none_rd = 0;
int partition_none_allowed = has_rows && has_cols;
(void)*tp_orig;
(void)split_rd;
av1_zero(pc_tree->pc_tree_stats);
pc_tree->pc_tree_stats.valid = 1;
// Override partition costs at the edges of the frame in the same
// way as in read_partition (see decodeframe.c)
if (!(has_rows && has_cols)) {
assert(bsize_at_least_8x8 && pl >= 0);
const aom_cdf_prob *partition_cdf = cm->fc->partition_cdf[pl];
for (int i = 0; i < PARTITION_TYPES; ++i) tmp_partition_cost[i] = INT_MAX;
if (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, bsize);
static const int bot_inv_map[2] = { PARTITION_HORZ, PARTITION_SPLIT };
av1_cost_tokens_from_cdf(tmp_partition_cost, bot_cdf, bot_inv_map);
} else if (has_rows) {