Google Git
Sign in
aomedia / aom / df89ebb0d290ab2f900fac427249c90588749227 / . / av1 / encoder / encodeframe.c
blob: 5ef2260d6c2e21e795a1062109079331235165ac [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 <float.h>
#include <math.h>
#include <stdbool.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/common/warped_motion.h"
#include "av1/encoder/aq_complexity.h"
#include "av1/encoder/aq_cyclicrefresh.h"
#include "av1/encoder/aq_variance.h"
#include "av1/encoder/corner_detect.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/partition_strategy.h"
#if !CONFIG_REALTIME_ONLY
#include "av1/encoder/partition_model_weights.h"
#endif
#include "av1/encoder/rd.h"
#include "av1/encoder/rdopt.h"
#include "av1/encoder/reconinter_enc.h"
#include "av1/encoder/segmentation.h"
#include "av1/encoder/tokenize.h"
#include "av1/encoder/tpl_model.h"
#include "av1/encoder/var_based_part.h"
static AOM_INLINE void encode_superblock(const AV1_COMP *const cpi,
TileDataEnc *tile_data, ThreadData *td,
TOKENEXTRA **t, RUN_TYPE dry_run,
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.
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
};
enum { PICK_MODE_RD = 0, PICK_MODE_NONRD };
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;
assert(bd == 8 || bd == 10 || bd == 12);
const int off_index = (bd - 8) >> 1;
const uint16_t *high_var_offs[3] = { AV1_HIGH_VAR_OFFS_8,
AV1_HIGH_VAR_OFFS_10,
AV1_HIGH_VAR_OFFS_12 };
var =
cpi->fn_ptr[bs].vf(ref->buf, ref->stride,
CONVERT_TO_BYTEPTR(high_var_offs[off_index]), 0, &sse);
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_yv12_buf(&cpi->common, 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;
}
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 AOM_INLINE void set_ssim_rdmult(const AV1_COMP *const cpi,
MACROBLOCK *const x,
const BLOCK_SIZE bsize, const int mi_row,
const int mi_col, int *const rdmult) {
const AV1_COMMON *const cm = &cpi->common;
const int bsize_base = BLOCK_16X16;
const int num_mi_w = mi_size_wide[bsize_base];
const int num_mi_h = mi_size_high[bsize_base];
const int num_cols = (cm->mi_cols + num_mi_w - 1) / num_mi_w;
const int num_rows = (cm->mi_rows + num_mi_h - 1) / num_mi_h;
const int num_bcols = (mi_size_wide[bsize] + num_mi_w - 1) / num_mi_w;
const int num_brows = (mi_size_high[bsize] + num_mi_h - 1) / num_mi_h;
int row, col;
double num_of_mi = 0.0;
double geom_mean_of_scale = 0.0;
assert(cpi->oxcf.tuning == AOM_TUNE_SSIM);
aom_clear_system_state();
for (row = mi_row / num_mi_w;
row < num_rows && row < mi_row / num_mi_w + num_brows; ++row) {
for (col = mi_col / num_mi_h;
col < num_cols && col < mi_col / num_mi_h + num_bcols; ++col) {
const int index = row * num_cols + col;
geom_mean_of_scale += log(cpi->ssim_rdmult_scaling_factors[index]);
num_of_mi += 1.0;
}
}
geom_mean_of_scale = exp(geom_mean_of_scale / num_of_mi);
*rdmult = (int)((double)(*rdmult) * geom_mean_of_scale + 0.5);
*rdmult = AOMMAX(*rdmult, 0);
set_error_per_bit(x, *rdmult);
aom_clear_system_state();
}
static int get_hier_tpl_rdmult(const AV1_COMP *const cpi, MACROBLOCK *const x,
const BLOCK_SIZE bsize, const int mi_row,
const int mi_col, int orig_rdmult) {
const AV1_COMMON *const cm = &cpi->common;
assert(IMPLIES(cpi->gf_group.size > 0,
cpi->gf_group.index < cpi->gf_group.size));
const int tpl_idx = cpi->gf_group.index;
const TplDepFrame *tpl_frame = &cpi->tpl_frame[tpl_idx];
MACROBLOCKD *const xd = &x->e_mbd;
const int deltaq_rdmult = set_deltaq_rdmult(cpi, xd);
if (cpi->tpl_model_pass == 1) {
assert(cpi->oxcf.enable_tpl_model == 2);
return deltaq_rdmult;
}
if (tpl_frame->is_valid == 0) return deltaq_rdmult;
if (!is_frame_tpl_eligible((AV1_COMP *)cpi)) return deltaq_rdmult;
if (tpl_idx >= MAX_LAG_BUFFERS) return deltaq_rdmult;
if (cpi->oxcf.superres_mode != SUPERRES_NONE) return deltaq_rdmult;
if (cpi->oxcf.aq_mode != NO_AQ) return deltaq_rdmult;
const int bsize_base = BLOCK_16X16;
const int num_mi_w = mi_size_wide[bsize_base];
const int num_mi_h = mi_size_high[bsize_base];
const int num_cols = (cm->mi_cols + num_mi_w - 1) / num_mi_w;
const int num_rows = (cm->mi_rows + num_mi_h - 1) / num_mi_h;
const int num_bcols = (mi_size_wide[bsize] + num_mi_w - 1) / num_mi_w;
const int num_brows = (mi_size_high[bsize] + num_mi_h - 1) / num_mi_h;
int row, col;
double base_block_count = 0.0;
double geom_mean_of_scale = 0.0;
aom_clear_system_state();
for (row = mi_row / num_mi_w;
row < num_rows && row < mi_row / num_mi_w + num_brows; ++row) {
for (col = mi_col / num_mi_h;
col < num_cols && col < mi_col / num_mi_h + num_bcols; ++col) {
const int index = row * num_cols + col;
geom_mean_of_scale += log(cpi->tpl_sb_rdmult_scaling_factors[index]);
base_block_count += 1.0;
}
}
geom_mean_of_scale = exp(geom_mean_of_scale / base_block_count);
int rdmult = (int)((double)orig_rdmult * geom_mean_of_scale + 0.5);
rdmult = AOMMAX(rdmult, 0);
set_error_per_bit(x, rdmult);
aom_clear_system_state();
if (bsize == cm->seq_params.sb_size) {
const int rdmult_sb = set_deltaq_rdmult(cpi, xd);
assert(rdmult_sb == rdmult);
(void)rdmult_sb;
}
return rdmult;
}
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 AOM_INLINE 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;
if (aq_mode != NO_AQ) {
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 = get_hier_tpl_rdmult(cpi, x, bsize, mi_row, mi_col, x->rdmult);
}
if (cpi->oxcf.tuning == AOM_TUNE_SSIM) {
set_ssim_rdmult(cpi, x, bsize, mi_row, mi_col, &x->rdmult);
}
}
static AOM_INLINE 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;
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, 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, &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.
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, bsize);
// required by av1_append_sub8x8_mvs_for_idx() and av1_find_best_ref_mvs()
xd->tile = *tile;
xd->cfl.mi_row = mi_row;
xd->cfl.mi_col = mi_col;
}
static AOM_INLINE 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;
const struct segmentation *const seg = &cm->seg;
MACROBLOCKD *const xd = &x->e_mbd;
MB_MODE_INFO *mbmi;
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->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 AOM_INLINE void update_filter_type_count(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];
}
}
static AOM_INLINE void update_filter_type_cdf(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);
update_cdf(xd->tile_ctx->switchable_interp_cdf[ctx], filter,
SWITCHABLE_FILTERS);
}
}
static AOM_INLINE 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 AOM_INLINE 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));
}
const int stride = xd->tx_type_map_stride;
const int bw = mi_size_wide[mbmi->sb_type];
for (int row = 0; row < mi_size_high[mbmi->sb_type]; ++row) {
memset(xd->tx_type_map + row * stride, DCT_DCT,
bw * sizeof(xd->tx_type_map[0]));
}
av1_zero(x->blk_skip);
x->skip = 0;
}
static AOM_INLINE void update_state(const AV1_COMP *const cpi, ThreadData *td,
const PICK_MODE_CONTEXT *const 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;
const MB_MODE_INFO *const 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;
memcpy(x->blk_skip, ctx->blk_skip, sizeof(x->blk_skip[0]) * ctx->num_4x4_blk);
x->skip = ctx->rd_stats.skip;
xd->tx_type_map = ctx->tx_type_map;
xd->tx_type_map_stride = mi_size_wide[bsize];
// If not dry_run, copy the transform type data into the frame level buffer.
// Encoder will fetch tx types when writing bitstream.
if (!dry_run) {
const int grid_idx = get_mi_grid_idx(cm, mi_row, mi_col);
uint8_t *const tx_type_map = cm->tx_type_map + grid_idx;
const int mi_stride = cm->mi_stride;
for (int blk_row = 0; blk_row < bh; ++blk_row) {
av1_copy_array(tx_type_map + blk_row * mi_stride,
xd->tx_type_map + blk_row * xd->tx_type_map_stride, bw);
}
xd->tx_type_map = tx_type_map;
xd->tx_type_map_stride = mi_stride;
}
// 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, x->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->rd_stats.rate, ctx->rd_stats.dist,
x->skip);
}
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(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);
if (cm->seq_params.order_hint_info.enable_ref_frame_mvs)
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,
BLOCK_SIZE bsize) {
// 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, bsize, 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 EdgeInfo edge_info(const struct buf_2d *ref, const BLOCK_SIZE bsize,
const bool high_bd, const int bd) {
const int width = block_size_wide[bsize];
const int height = block_size_high[bsize];
// Implementation requires width to be a multiple of 8. It also requires
// height to be a multiple of 4, but this is always the case.
assert(height % 4 == 0);
if (width % 8 != 0) {
EdgeInfo ei = { .magnitude = 0, .x = 0, .y = 0 };
return ei;
}
return av1_edge_exists(ref->buf, ref->stride, width, height, high_bd, bd);
}
static int use_pb_simple_motion_pred_sse(const AV1_COMP *const cpi) {
// TODO(debargha, yuec): Not in use, need to implement a speed feature
// utilizing this data point, and replace '0' by the corresponding speed
// feature flag.
return 0 && !frame_is_intra_only(&cpi->common);
}
// This function will copy the winner reference mode information from block
// level (x->mbmi_ext) to frame level (cpi->mbmi_ext_frame_base). This frame
// level buffer (cpi->mbmi_ext_frame_base) will be used during bitstream
// preparation.
static INLINE void copy_winner_ref_mode_from_mbmi_ext(MACROBLOCK *const x) {
MACROBLOCKD *const xd = &x->e_mbd;
MB_MODE_INFO *mbmi = xd->mi[0];
uint8_t ref_frame_type = av1_ref_frame_type(mbmi->ref_frame);
memcpy(x->mbmi_ext_frame->ref_mv_stack,
x->mbmi_ext->ref_mv_stack[ref_frame_type],
sizeof(x->mbmi_ext->ref_mv_stack[USABLE_REF_MV_STACK_SIZE]));
memcpy(x->mbmi_ext_frame->weight, x->mbmi_ext->weight[ref_frame_type],
sizeof(x->mbmi_ext->weight[USABLE_REF_MV_STACK_SIZE]));
x->mbmi_ext_frame->mode_context = x->mbmi_ext->mode_context[ref_frame_type];
x->mbmi_ext_frame->ref_mv_count = x->mbmi_ext->ref_mv_count[ref_frame_type];
memcpy(x->mbmi_ext_frame->global_mvs, x->mbmi_ext->global_mvs,
sizeof(x->mbmi_ext->global_mvs));
}
static AOM_INLINE 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,
int pick_mode_type) {
if (best_rd.rdcost < 0) {
ctx->rd_stats.rdcost = INT64_MAX;
ctx->rd_stats.skip = 0;
av1_invalid_rd_stats(rd_cost);
return;
}
set_offsets(cpi, &tile_data->tile_info, x, mi_row, mi_col, bsize);
if (ctx->rd_mode_is_ready) {
assert(ctx->mic.sb_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;
}
AV1_COMMON *const cm = &cpi->common;
const int num_planes = av1_num_planes(cm);
MACROBLOCKD *const xd = &x->e_mbd;
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.aq_mode;
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 = bsize;
mbmi->partition = partition;
#if CONFIG_RD_DEBUG
mbmi->mi_row = mi_row;
mbmi->mi_col = mi_col;
#endif
xd->tx_type_map = x->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];
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];
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 (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);
}
if (use_pb_simple_motion_pred_sse(cpi)) {
const MV ref_mv_full = { .row = 0, .col = 0 };
unsigned int var = 0;
av1_simple_motion_sse_var(cpi, x, mi_row, mi_col, bsize, ref_mv_full, 0,
&x->simple_motion_pred_sse, &var);
}
// If the threshold for disabling wedge search is zero, it means the feature
// should not be used. Use a value that will always succeed in the check.
if (cpi->sf.disable_wedge_search_edge_thresh == 0) {
x->edge_strength = UINT16_MAX;
x->edge_strength_x = UINT16_MAX;
x->edge_strength_y = UINT16_MAX;
} else {
EdgeInfo ei =
edge_info(&x->plane[0].src, bsize, is_cur_buf_hbd(xd), xd->bd);
x->edge_strength = ei.magnitude;
x->edge_strength_x = ei.x;
x->edge_strength_y = ei.y;
}
// 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
set_error_per_bit(x, 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 {
// TODO(kyslov): do the same for pick_intra_mode and
// pick_inter_mode_sb_seg_skip
switch (pick_mode_type) {
case PICK_MODE_RD:
av1_rd_pick_inter_mode_sb(cpi, tile_data, x, rd_cost, bsize, ctx,
best_rd.rdcost);
break;
case PICK_MODE_NONRD:
av1_nonrd_pick_inter_mode_sb(cpi, tile_data, x, rd_cost, bsize, ctx,
best_rd.rdcost);
break;
default: assert(0 && "Unknown pick mode type.");
}
}
#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) &&
(cm->current_frame.frame_type == KEY_FRAME ||
cpi->refresh_alt_ref_frame || cpi->refresh_bwd_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->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 AOM_INLINE void update_inter_mode_stats(FRAME_CONTEXT *fc,
FRAME_COUNTS *counts,
PREDICTION_MODE mode,
int16_t mode_context) {
(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
update_cdf(fc->newmv_cdf[mode_ctx], 0, 2);
return;
}
#if CONFIG_ENTROPY_STATS
++counts->newmv_mode[mode_ctx][1];
#endif
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
update_cdf(fc->zeromv_cdf[mode_ctx], 0, 2);
return;
}
#if CONFIG_ENTROPY_STATS
++counts->zeromv_mode[mode_ctx][1];
#endif
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
update_cdf(fc->refmv_cdf[mode_ctx], mode != NEARESTMV, 2);
}
static AOM_INLINE void update_palette_cdf(MACROBLOCKD *xd,
const MB_MODE_INFO *const mbmi,
FRAME_COUNTS *counts) {
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
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
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
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
update_cdf(fc->palette_uv_size_cdf[palette_bsize_ctx],
n - PALETTE_MIN_SIZE, PALETTE_SIZES);
}
}
}
static AOM_INLINE 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) {
FRAME_CONTEXT *fc = xd->tile_ctx;
const PREDICTION_MODE y_mode = mbmi->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
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
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
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
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;
const UV_PREDICTION_MODE uv_mode = mbmi->uv_mode;
const CFL_ALLOWED_TYPE cfl_allowed = is_cfl_allowed(xd);
#if CONFIG_ENTROPY_STATS
++counts->uv_mode[cfl_allowed][y_mode][uv_mode];
#endif // CONFIG_ENTROPY_STATS
update_cdf(fc->uv_mode_cdf[cfl_allowed][y_mode], uv_mode,
UV_INTRA_MODES - !cfl_allowed);
if (uv_mode == UV_CFL_PRED) {
const int8_t joint_sign = mbmi->cfl_alpha_signs;
const uint8_t idx = mbmi->cfl_alpha_idx;
#if CONFIG_ENTROPY_STATS
++counts->cfl_sign[joint_sign];
#endif
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
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
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
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);
}
}
static AOM_INLINE void update_stats(const AV1_COMMON *const cm, 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 CurrentFrame *const current_frame = &cm->current_frame;
const BLOCK_SIZE bsize = mbmi->sb_type;
FRAME_CONTEXT *fc = xd->tile_ctx;
const int seg_ref_active =
segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_REF_FRAME);
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 (!mbmi->skip_mode && !seg_ref_active) {
const int skip_ctx = av1_get_skip_context(xd);
#if CONFIG_ENTROPY_STATS
td->counts->skip[skip_ctx][mbmi->skip]++;
#endif
update_cdf(fc->skip_cdfs[skip_ctx], mbmi->skip, 2);
}
#if CONFIG_ENTROPY_STATS
// 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) &&
super_block_upper_left) {
const int dq =
(mbmi->current_qindex - xd->current_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)) {
sum_intra_stats(cm, td->counts, xd, mbmi, xd->above_mbmi, xd->left_mbmi,
frame_is_intra_only(cm), mi_row, mi_col);
}
if (av1_allow_intrabc(cm)) {
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) || mbmi->skip_mode) return;
FRAME_COUNTS *const counts = td->counts;
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
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];
if (current_frame->reference_mode == REFERENCE_MODE_SELECT) {
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
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;
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
}
}
} 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
}
}
}
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 (is_interintra_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->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
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)) {
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(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 == 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
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 -
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->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_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);
} else {
update_inter_mode_stats(fc, counts, mode, mode_ctx);
}
const int new_mv = mbmi->mode == NEWMV || mbmi->mode == NEW_NEWMV;
if (new_mv) {
const uint8_t ref_frame_type = av1_ref_frame_type(mbmi->ref_frame);
for (int idx = 0; idx < 2; ++idx) {
if (mbmi_ext->ref_mv_count[ref_frame_type] > idx + 1) {
const uint8_t drl_ctx =
av1_drl_ctx(mbmi_ext->weight[ref_frame_type], idx);
update_cdf(fc->drl_cdf[drl_ctx], mbmi->ref_mv_idx != idx, 2);
#if CONFIG_ENTROPY_STATS
++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)) {
const uint8_t ref_frame_type = av1_ref_frame_type(mbmi->ref_frame);
for (int idx = 1; idx < 3; ++idx) {
if (mbmi_ext->ref_mv_count[ref_frame_type] > idx + 1) {
const uint8_t drl_ctx =
av1_drl_ctx(mbmi_ext->weight[ref_frame_type], idx);
update_cdf(fc->drl_cdf[drl_ctx], mbmi->ref_mv_idx != idx - 1, 2);
#if CONFIG_ENTROPY_STATS
++counts->drl_mode[drl_ctx][mbmi->ref_mv_idx != idx - 1];
#endif
if (mbmi->ref_mv_idx == idx - 1) break;
}
}
}
if (have_newmv_in_inter_mode(mbmi->mode)) {
const int allow_hp = cm->cur_frame_force_integer_mv
? MV_SUBPEL_NONE
: cm->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,
allow_hp);
}
} else if (mbmi->mode == NEAREST_NEWMV || mbmi->mode == NEAR_NEWMV) {
const int ref = 1;
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,
allow_hp);
} else if (mbmi->mode == NEW_NEARESTMV || mbmi->mode == NEW_NEARMV) {
const int ref = 0;
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,
allow_hp);
}
}
}
}
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 AOM_INLINE 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 AOM_INLINE 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 AOM_INLINE 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;
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;
update_state(cpi, td, ctx, mi_row, mi_col, bsize, dry_run);
if (!dry_run) {
x->mbmi_ext_frame->cb_offset = x->cb_offset;
assert(x->cb_offset <
(1 << num_pels_log2_lookup[cpi->common.seq_params.sb_size]));
}
encode_superblock(cpi, tile_data, td, tp, dry_run, bsize, rate);
if (!dry_run) {
const AV1_COMMON *const cm = &cpi->common;
x->cb_offset += block_size_wide[bsize] * block_size_high[bsize];
if (bsize == cpi->common.seq_params.sb_size && mbmi->skip == 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->compound_idx == 0 ||
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) &&
super_block_upper_left) {
xd->current_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 {
const int seg_ref_active =
segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_REF_FRAME);
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)) {
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, mi_row, mi_col);
}
// Gather obmc count to update the probability.
if (!cpi->sf.disable_obmc && cpi->sf.prune_obmc_prob_thresh > 0) {
const int inter_block = is_inter_block(mbmi);
const int seg_ref_active =
segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_REF_FRAME);
if (!seg_ref_active && inter_block) {
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 &&
motion_allowed == OBMC_CAUSAL) {
td->rd_counts.obmc_used[bsize][mbmi->motion_mode == OBMC_CAUSAL]++;
}
}
}
}
// TODO(Ravi/Remya): Move this copy function to a better logical place
copy_winner_ref_mode_from_mbmi_ext(x);
x->rdmult = origin_mult;
}
static AOM_INLINE 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;
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 >= 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);
}
static AOM_INLINE 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 grid_index = get_mi_grid_idx(cm, r, c);
const int mi_index = get_alloc_mi_idx(cm, r, c);
mib[grid_index] = mi + mi_index;
mib[grid_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 AOM_INLINE 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 + get_alloc_mi_idx(cm, mi_row, mi_col);
int bh = mi_size_high[bsize];
int bw = mi_size_wide[bsize];
assert(bsize >= cm->mi_alloc_bsize &&
"Attempted to use bsize < cm->mi_alloc_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) {
const int grid_index = get_mi_grid_idx(cm, block_row, block_col);
const int mi_index = get_alloc_mi_idx(cm, block_row, block_col);
mib[grid_index] = mi_upper_left + mi_index;
mib[grid_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 AOM_INLINE 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, invalid_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);
av1_invalid_rd_stats(&invalid_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_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);
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;
pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &none_rdc,
PARTITION_NONE, bsize, ctx_none, invalid_rdc, PICK_MODE_RD);
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:
pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &last_part_rdc,
PARTITION_NONE, bsize, ctx_none, invalid_rdc, PICK_MODE_RD);
break;
case PARTITION_HORZ:
pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &last_part_rdc,
PARTITION_HORZ, subsize, &pc_tree->horizontal[0],
invalid_rdc, PICK_MODE_RD);
if (last_part_rdc.rate != INT_MAX && bsize >= BLOCK_8X8 &&
mi_row + hbs < cm->mi_rows) {
RD_STATS tmp_rdc;
const PICK_MODE_CONTEXT *const ctx_h = &pc_tree->horizontal[0];
av1_init_rd_stats(&tmp_rdc);
update_state(cpi, td, ctx_h, mi_row, mi_col, subsize, 1);
encode_superblock(cpi, tile_data, td, tp, DRY_RUN_NORMAL, subsize,
NULL);
pick_sb_modes(cpi, tile_data, x, mi_row + hbs, mi_col, &tmp_rdc,
PARTITION_HORZ, subsize, &pc_tree->horizontal[1],
invalid_rdc, PICK_MODE_RD);
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:
pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &last_part_rdc,
PARTITION_VERT, subsize, &pc_tree->vertical[0], invalid_rdc,
PICK_MODE_RD);
if (last_part_rdc.rate != INT_MAX && bsize >= BLOCK_8X8 &&
mi_col + hbs < cm->mi_cols) {
RD_STATS tmp_rdc;
const PICK_MODE_CONTEXT *const ctx_v = &pc_tree->vertical[0];
av1_init_rd_stats(&tmp_rdc);
update_state(cpi, td, ctx_v, mi_row, mi_col, subsize, 1);
encode_superblock(cpi, tile_data, td, tp, DRY_RUN_NORMAL, subsize,
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,
PICK_MODE_RD);
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 partition 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;
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,
invalid_rdc, PICK_MODE_RD);
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;
x->rdmult = orig_rdmult;
}
static int is_leaf_split_partition(AV1_COMMON *cm, int mi_row, int mi_col,
BLOCK_SIZE bsize) {
const int bs = mi_size_wide[bsize];
const int hbs = bs / 2;
assert(bsize >= BLOCK_8X8);
const BLOCK_SIZE subsize = get_partition_subsize(bsize, PARTITION_SPLIT);
for (int i = 0; i < 4; i++) {
int x_idx = (i & 1) * hbs;
int y_idx = (i >> 1) * hbs;
if ((mi_row + y_idx >= cm->mi_rows) || (mi_col + x_idx >= cm->mi_cols))
return 0;
if (get_partition(cm, mi_row + y_idx, mi_col + x_idx, subsize) !=
PARTITION_NONE)
return 0;
}
return 1;
}
static AOM_INLINE void nonrd_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, PC_TREE *pc_tree) {
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;
// Only square blocks from 8x8 to 128x128 are supported
assert(bsize >= BLOCK_8X8 && bsize <= BLOCK_128X128);
const int bs = mi_size_wide[bsize];
const int hbs = bs / 2;
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);
assert(subsize <= BLOCK_LARGEST);
const int pl = (bsize >= BLOCK_8X8)
? partition_plane_context(xd, mi_row, mi_col, bsize)
: 0;
RD_STATS dummy_cost;
av1_invalid_rd_stats(&dummy_cost);
RD_STATS invalid_rd;
av1_invalid_rd_stats(&invalid_rd);
if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;
assert(mi_size_wide[bsize] == mi_size_high[bsize]);
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);
switch (partition) {
case PARTITION_NONE:
pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &dummy_cost,
PARTITION_NONE, bsize, &pc_tree->none, invalid_rd,
PICK_MODE_NONRD);
encode_b(cpi, tile_data, td, tp, mi_row, mi_col, 0, bsize, partition,
&pc_tree->none, NULL);
break;
case PARTITION_VERT:
pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &dummy_cost,
PARTITION_VERT, subsize, &pc_tree->vertical[0], invalid_rd,
PICK_MODE_NONRD);
encode_b(cpi, tile_data, td, tp, mi_row, mi_col, 0, subsize,
PARTITION_VERT, &pc_tree->vertical[0], NULL);
if (mi_col + hbs < cm->mi_cols && bsize > BLOCK_8X8) {
pick_sb_modes(cpi, tile_data, x, mi_row, mi_col + hbs, &dummy_cost,
PARTITION_VERT, subsize, &pc_tree->vertical[1],
invalid_rd, PICK_MODE_NONRD);
encode_b(cpi, tile_data, td, tp, mi_row, mi_col + hbs, 0, subsize,
PARTITION_VERT, &pc_tree->vertical[1], NULL);
}
break;
case PARTITION_HORZ:
pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &dummy_cost,
PARTITION_HORZ, subsize, &pc_tree->horizontal[0],
invalid_rd, PICK_MODE_NONRD);
encode_b(cpi, tile_data, td, tp, mi_row, mi_col, 0, subsize,
PARTITION_HORZ, &pc_tree->horizontal[0], NULL);
if (mi_row + hbs < cm->mi_rows && bsize > BLOCK_8X8) {
pick_sb_modes(cpi, tile_data, x, mi_row + hbs, mi_col, &dummy_cost,
PARTITION_HORZ, subsize, &pc_tree->horizontal[1],
invalid_rd, PICK_MODE_NONRD);
encode_b(cpi, tile_data, td, tp, mi_row + hbs, mi_col, 0, subsize,
PARTITION_HORZ, &pc_tree->horizontal[1], NULL);
}
break;
case PARTITION_SPLIT:
if (cpi->sf.nonrd_merge_partition &&
is_leaf_split_partition(cm, mi_row, mi_col, bsize) &&
!frame_is_intra_only(cm)) {
RD_SEARCH_MACROBLOCK_CONTEXT x_ctx;
RD_STATS split_rdc, none_rdc;
av1_init_rd_stats(&split_rdc);
av1_invalid_rd_stats(&none_rdc);
save_context(x, &x_ctx, mi_row, mi_col, bsize, 3);
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);
pc_tree->partitioning = PARTITION_NONE;
pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &none_rdc,
PARTITION_NONE, bsize, &pc_tree->none, invalid_rd,
PICK_MODE_NONRD);
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, 3);
for (int i = 0; i < 4; i++) {
RD_STATS block_rdc;
av1_invalid_rd_stats(&block_rdc);
int x_idx = (i & 1) * hbs;
int y_idx = (i >> 1) * hbs;
if ((mi_row + y_idx >= cm->mi_rows) ||
(mi_col + x_idx >= cm->mi_cols))
continue;
xd->above_txfm_context =
cm->above_txfm_context[tile_info->tile_row] + mi_col + x_idx;
xd->left_txfm_context =
xd->left_txfm_context_buffer + ((mi_row + y_idx) & MAX_MIB_MASK);
pc_tree->split[i]->partitioning = PARTITION_NONE;
pick_sb_modes(cpi, tile_data, x, mi_row + y_idx, mi_col + x_idx,
&block_rdc, PARTITION_NONE, subsize,
&pc_tree->split[i]->none, invalid_rd, PICK_MODE_NONRD);
split_rdc.rate += block_rdc.rate;
split_rdc.dist += block_rdc.dist;
encode_b(cpi, tile_data, td, tp, mi_row + y_idx, mi_col + x_idx, 1,
subsize, PARTITION_NONE, &pc_tree->split[i]->none, NULL);
}
restore_context(x, &x_ctx, mi_row, mi_col, bsize, 3);
split_rdc.rate += x->partition_cost[pl][PARTITION_SPLIT];
split_rdc.rdcost = RDCOST(x->rdmult, split_rdc.rate, split_rdc.dist);
if (none_rdc.rdcost < split_rdc.rdcost) {
set_offsets_without_segment_id(cpi, &tile_data->tile_info, x, mi_row,
mi_col, bsize);
mib[0]->sb_type = bsize;
pc_tree->partitioning = PARTITION_NONE;
encode_b(cpi, tile_data, td, tp, mi_row, mi_col, 0, bsize, partition,
&pc_tree->none, NULL);
} else {
set_offsets_without_segment_id(cpi, &tile_data->tile_info, x, mi_row,
mi_col, bsize);
mib[0]->sb_type = subsize;
pc_tree->partitioning = PARTITION_SPLIT;
for (int i = 0; i < 4; i++) {
int x_idx = (i & 1) * hbs;
int y_idx = (i >> 1) * hbs;
if ((mi_row + y_idx >= cm->mi_rows) ||
(mi_col + x_idx >= cm->mi_cols))
continue;
set_offsets_without_segment_id(cpi, &tile_data->tile_info, x,
mi_row + y_idx, mi_col + x_idx,
subsize);
encode_b(cpi, tile_data, td, tp, mi_row + y_idx, mi_col + x_idx, 0,
subsize, PARTITION_NONE, &pc_tree->split[i]->none, NULL);
}
}
} else {
for (int i = 0; i < 4; i++) {
int x_idx = (i & 1) * hbs;
int y_idx = (i >> 1) * hbs;
int jj = i >> 1, ii = i & 0x01;
if ((mi_row + y_idx >= cm->mi_rows) ||
(mi_col + x_idx >= cm->mi_cols))
continue;
nonrd_use_partition(
cpi, td, tile_data, mib + jj * hbs * cm->mi_stride + ii * hbs, tp,
mi_row + y_idx, mi_col + x_idx, subsize, pc_tree->split[i]);
}
}
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 !CONFIG_REALTIME_ONLY
static const FIRSTPASS_STATS *read_one_frame_stats(const TWO_PASS *p, int frm) {
assert(frm >= 0);
if (frm < 0 || p->stats_in_start + frm > p->stats_in_end) {
return NULL;
}
return &p->stats_in_start[frm];
}
// 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 AV1_COMMON *const cm = &cpi->common;
const FIRSTPASS_STATS *const this_frame_stats = read_one_frame_stats(
&cpi->twopass, cm->current_frame.display_order_hint);
if (this_frame_stats == NULL) return AOM_CODEC_ERROR;
// The inactive region is specified in MBs not mi units.
// The image edge is in the following MB row.
top_edge += (int)(this_frame_stats->inactive_zone_rows * 4);
bottom_edge -= (int)(this_frame_stats->inactive_zone_rows * 4);
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 AV1_COMMON *const cm = &cpi->common;
const FIRSTPASS_STATS *const this_frame_stats = read_one_frame_stats(
&cpi->twopass, cm->current_frame.display_order_hint);
if (this_frame_stats == NULL) return AOM_CODEC_ERROR;
// The inactive region is specified in MBs not mi units.
// The image edge is in the following MB row.
left_edge += (int)(this_frame_stats->inactive_zone_cols * 4);
right_edge -= (int)(this_frame_stats->inactive_zone_cols * 4);
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;
}
#endif // !CONFIG_REALTIME_ONLY
static INLINE void store_pred_mv(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx) {
memcpy(ctx->pred_mv, x->pred_mv, sizeof(x->pred_mv