blob: 71c6e27bfa56cbcf2eae864149e14ab216629a99 [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 <assert.h>
#include <float.h>
#include <limits.h>
#include <math.h>
#include <stdbool.h>
#include <stdio.h>
#include <stdlib.h>
#include <time.h>
#include "av1/common/scale.h"
#include "config/aom_config.h"
#include "config/aom_dsp_rtcd.h"
#include "aom/aomcx.h"
#if CONFIG_DENOISE
#include "aom_dsp/grain_table.h"
#include "aom_dsp/noise_util.h"
#include "aom_dsp/noise_model.h"
#endif
#include "aom_dsp/flow_estimation/corner_detect.h"
#include "aom_dsp/psnr.h"
#if CONFIG_INTERNAL_STATS
#include "aom_dsp/ssim.h"
#endif
#include "aom_ports/aom_timer.h"
#include "aom_ports/mem.h"
#include "aom_scale/aom_scale.h"
#include "aom_util/aom_pthread.h"
#if CONFIG_BITSTREAM_DEBUG
#include "aom_util/debug_util.h"
#endif // CONFIG_BITSTREAM_DEBUG
#include "av1/common/alloccommon.h"
#include "av1/common/debugmodes.h"
#include "av1/common/filter.h"
#include "av1/common/idct.h"
#include "av1/common/reconinter.h"
#include "av1/common/reconintra.h"
#include "av1/common/resize.h"
#include "av1/common/tile_common.h"
#include "av1/encoder/allintra_vis.h"
#include "av1/encoder/aq_complexity.h"
#include "av1/encoder/aq_cyclicrefresh.h"
#include "av1/encoder/aq_variance.h"
#include "av1/encoder/bitstream.h"
#if CONFIG_INTERNAL_STATS
#include "av1/encoder/blockiness.h"
#endif
#include "av1/encoder/context_tree.h"
#include "av1/encoder/dwt.h"
#include "av1/encoder/encodeframe.h"
#include "av1/encoder/encodemv.h"
#include "av1/encoder/encode_strategy.h"
#include "av1/encoder/encoder.h"
#include "av1/encoder/encoder_alloc.h"
#include "av1/encoder/encoder_utils.h"
#include "av1/encoder/encodetxb.h"
#include "av1/encoder/ethread.h"
#include "av1/encoder/firstpass.h"
#include "av1/encoder/hash_motion.h"
#include "av1/encoder/hybrid_fwd_txfm.h"
#include "av1/encoder/intra_mode_search.h"
#include "av1/encoder/mv_prec.h"
#include "av1/encoder/pass2_strategy.h"
#include "av1/encoder/pickcdef.h"
#include "av1/encoder/picklpf.h"
#include "av1/encoder/pickrst.h"
#include "av1/encoder/random.h"
#include "av1/encoder/ratectrl.h"
#include "av1/encoder/rc_utils.h"
#include "av1/encoder/rd.h"
#include "av1/encoder/rdopt.h"
#if CONFIG_SALIENCY_MAP
#include "av1/encoder/saliency_map.h"
#endif
#include "av1/encoder/segmentation.h"
#include "av1/encoder/speed_features.h"
#include "av1/encoder/superres_scale.h"
#include "av1/encoder/thirdpass.h"
#include "av1/encoder/tpl_model.h"
#include "av1/encoder/reconinter_enc.h"
#include "av1/encoder/var_based_part.h"
#define DEFAULT_EXPLICIT_ORDER_HINT_BITS 7
// #define OUTPUT_YUV_REC
#ifdef OUTPUT_YUV_REC
FILE *yuv_rec_file;
#define FILE_NAME_LEN 100
#endif
#ifdef OUTPUT_YUV_DENOISED
FILE *yuv_denoised_file = NULL;
#endif
static inline void Scale2Ratio(AOM_SCALING_MODE mode, int *hr, int *hs) {
switch (mode) {
case AOME_NORMAL:
*hr = 1;
*hs = 1;
break;
case AOME_FOURFIVE:
*hr = 4;
*hs = 5;
break;
case AOME_THREEFIVE:
*hr = 3;
*hs = 5;
break;
case AOME_THREEFOUR:
*hr = 3;
*hs = 4;
break;
case AOME_ONEFOUR:
*hr = 1;
*hs = 4;
break;
case AOME_ONEEIGHT:
*hr = 1;
*hs = 8;
break;
case AOME_ONETWO:
*hr = 1;
*hs = 2;
break;
case AOME_TWOTHREE:
*hr = 2;
*hs = 3;
break;
case AOME_ONETHREE:
*hr = 1;
*hs = 3;
break;
default:
*hr = 1;
*hs = 1;
assert(0);
break;
}
}
int av1_set_active_map(AV1_COMP *cpi, unsigned char *new_map_16x16, int rows,
int cols) {
const CommonModeInfoParams *const mi_params = &cpi->common.mi_params;
if (rows == mi_params->mb_rows && cols == mi_params->mb_cols) {
unsigned char *const active_map_4x4 = cpi->active_map.map;
const int mi_rows = mi_params->mi_rows;
const int mi_cols = mi_params->mi_cols;
cpi->active_map.update = 0;
cpi->rc.percent_blocks_inactive = 0;
assert(mi_rows % 2 == 0 && mi_rows > 0);
assert(mi_cols % 2 == 0 && mi_cols > 0);
if (new_map_16x16) {
int num_samples = 0;
int num_blocks_inactive = 0;
for (int r = 0; r < mi_rows; r += 4) {
for (int c = 0; c < mi_cols; c += 4) {
const uint8_t val = new_map_16x16[(r >> 2) * cols + (c >> 2)]
? AM_SEGMENT_ID_ACTIVE
: AM_SEGMENT_ID_INACTIVE;
num_samples++;
if (val == AM_SEGMENT_ID_INACTIVE) num_blocks_inactive++;
const int row_max = AOMMIN(4, mi_rows - r);
const int col_max = AOMMIN(4, mi_cols - c);
for (int x = 0; x < row_max; ++x) {
for (int y = 0; y < col_max; ++y) {
active_map_4x4[(r + x) * mi_cols + (c + y)] = val;
}
}
}
}
cpi->active_map.enabled = 1;
cpi->active_map.update = 1;
assert(num_samples);
cpi->rc.percent_blocks_inactive =
(num_blocks_inactive * 100) / num_samples;
}
return 0;
}
return -1;
}
int av1_get_active_map(AV1_COMP *cpi, unsigned char *new_map_16x16, int rows,
int cols) {
const CommonModeInfoParams *const mi_params = &cpi->common.mi_params;
if (rows == mi_params->mb_rows && cols == mi_params->mb_cols &&
new_map_16x16) {
unsigned char *const seg_map_8x8 = cpi->enc_seg.map;
const int mi_rows = mi_params->mi_rows;
const int mi_cols = mi_params->mi_cols;
const int row_scale = mi_size_high_log2[BLOCK_16X16];
const int col_scale = mi_size_wide_log2[BLOCK_16X16];
assert(mi_rows % 2 == 0);
assert(mi_cols % 2 == 0);
memset(new_map_16x16, !cpi->active_map.enabled, rows * cols);
if (cpi->active_map.enabled) {
for (int r = 0; r < (mi_rows >> row_scale); ++r) {
for (int c = 0; c < (mi_cols >> col_scale); ++c) {
// Cyclic refresh segments are considered active despite not having
// AM_SEGMENT_ID_ACTIVE
uint8_t temp = 0;
temp |= seg_map_8x8[(2 * r + 0) * mi_cols + (2 * c + 0)] !=
AM_SEGMENT_ID_INACTIVE;
temp |= seg_map_8x8[(2 * r + 0) * mi_cols + (2 * c + 1)] !=
AM_SEGMENT_ID_INACTIVE;
temp |= seg_map_8x8[(2 * r + 1) * mi_cols + (2 * c + 0)] !=
AM_SEGMENT_ID_INACTIVE;
temp |= seg_map_8x8[(2 * r + 1) * mi_cols + (2 * c + 1)] !=
AM_SEGMENT_ID_INACTIVE;
new_map_16x16[r * cols + c] |= temp;
}
}
}
return 0;
}
return -1;
}
void av1_initialize_enc(unsigned int usage, enum aom_rc_mode end_usage) {
bool is_allintra = usage == ALLINTRA;
av1_rtcd();
aom_dsp_rtcd();
aom_scale_rtcd();
av1_init_intra_predictors();
av1_init_me_luts();
if (!is_allintra) av1_init_wedge_masks();
if (!is_allintra || end_usage != AOM_Q) av1_rc_init_minq_luts();
}
void av1_new_framerate(AV1_COMP *cpi, double framerate) {
cpi->framerate = framerate < 0.1 ? 30 : framerate;
av1_rc_update_framerate(cpi, cpi->common.width, cpi->common.height);
}
double av1_get_compression_ratio(const AV1_COMMON *const cm,
size_t encoded_frame_size) {
const int upscaled_width = cm->superres_upscaled_width;
const int height = cm->height;
const int64_t luma_pic_size = (int64_t)upscaled_width * height;
const SequenceHeader *const seq_params = cm->seq_params;
const BITSTREAM_PROFILE profile = seq_params->profile;
const int pic_size_profile_factor =
profile == PROFILE_0 ? 15 : (profile == PROFILE_1 ? 30 : 36);
encoded_frame_size =
(encoded_frame_size > 129 ? encoded_frame_size - 128 : 1);
const int64_t uncompressed_frame_size =
(luma_pic_size * pic_size_profile_factor) >> 3;
return (double)uncompressed_frame_size / encoded_frame_size;
}
static void auto_tile_size_balancing(AV1_COMMON *const cm, int num_sbs,
int num_tiles_lg, int tile_col_row) {
CommonTileParams *const tiles = &cm->tiles;
int i, start_sb;
int size_sb = num_sbs >> num_tiles_lg;
int res_sbs = num_sbs - (size_sb << num_tiles_lg);
int num_tiles = 1 << num_tiles_lg;
int inc_index = num_tiles - res_sbs;
tiles->uniform_spacing = 0;
for (i = 0, start_sb = 0; start_sb < num_sbs && i < MAX_TILE_COLS; ++i) {
if (i == inc_index) ++size_sb;
if (tile_col_row)
tiles->col_start_sb[i] = start_sb;
else
tiles->row_start_sb[i] = start_sb;
start_sb += AOMMIN(size_sb, tiles->max_width_sb);
}
if (tile_col_row) {
tiles->cols = i;
tiles->col_start_sb[i] = num_sbs;
} else {
tiles->rows = i;
tiles->row_start_sb[i] = num_sbs;
}
}
static void set_tile_info(AV1_COMMON *const cm,
const TileConfig *const tile_cfg) {
const CommonModeInfoParams *const mi_params = &cm->mi_params;
const SequenceHeader *const seq_params = cm->seq_params;
CommonTileParams *const tiles = &cm->tiles;
int i, start_sb;
av1_get_tile_limits(cm);
int sb_cols =
CEIL_POWER_OF_TWO(mi_params->mi_cols, seq_params->mib_size_log2);
// configure tile columns
if (tile_cfg->tile_width_count == 0 || tile_cfg->tile_height_count == 0) {
tiles->uniform_spacing = 1;
tiles->log2_cols = AOMMAX(tile_cfg->tile_columns, tiles->min_log2_cols);
// Add a special case to handle super resolution
sb_cols = coded_to_superres_mi(sb_cols, cm->superres_scale_denominator);
int min_log2_cols = 0;
for (; (tiles->max_width_sb << min_log2_cols) <= sb_cols; ++min_log2_cols) {
}
tiles->log2_cols = AOMMAX(tiles->log2_cols, min_log2_cols);
tiles->log2_cols = AOMMIN(tiles->log2_cols, tiles->max_log2_cols);
} else if (tile_cfg->tile_widths[0] < 0) {
auto_tile_size_balancing(cm, sb_cols, tile_cfg->tile_columns, 1);
} else {
int size_sb, j = 0;
tiles->uniform_spacing = 0;
for (i = 0, start_sb = 0; start_sb < sb_cols && i < MAX_TILE_COLS; i++) {
tiles->col_start_sb[i] = start_sb;
size_sb = tile_cfg->tile_widths[j++];
if (j >= tile_cfg->tile_width_count) j = 0;
start_sb += AOMMIN(size_sb, tiles->max_width_sb);
}
tiles->cols = i;
tiles->col_start_sb[i] = sb_cols;
}
av1_calculate_tile_cols(seq_params, mi_params->mi_rows, mi_params->mi_cols,
tiles);
// configure tile rows
int sb_rows =
CEIL_POWER_OF_TWO(mi_params->mi_rows, seq_params->mib_size_log2);
if (tiles->uniform_spacing) {
tiles->log2_rows = AOMMAX(tile_cfg->tile_rows, tiles->min_log2_rows);
tiles->log2_rows = AOMMIN(tiles->log2_rows, tiles->max_log2_rows);
} else if (tile_cfg->tile_heights[0] < 0) {
auto_tile_size_balancing(cm, sb_rows, tile_cfg->tile_rows, 0);
} else {
int size_sb, j = 0;
for (i = 0, start_sb = 0; start_sb < sb_rows && i < MAX_TILE_ROWS; i++) {
tiles->row_start_sb[i] = start_sb;
size_sb = tile_cfg->tile_heights[j++];
if (j >= tile_cfg->tile_height_count) j = 0;
start_sb += AOMMIN(size_sb, tiles->max_height_sb);
}
tiles->rows = i;
tiles->row_start_sb[i] = sb_rows;
}
av1_calculate_tile_rows(seq_params, mi_params->mi_rows, tiles);
}
void av1_update_frame_size(AV1_COMP *cpi) {
AV1_COMMON *const cm = &cpi->common;
MACROBLOCKD *const xd = &cpi->td.mb.e_mbd;
// Setup mi_params here in case we need more mi's.
CommonModeInfoParams *const mi_params = &cm->mi_params;
mi_params->set_mb_mi(mi_params, cm->width, cm->height,
cpi->sf.part_sf.default_min_partition_size);
av1_init_macroblockd(cm, xd);
if (!cpi->ppi->seq_params_locked)
set_sb_size(cm->seq_params,
av1_select_sb_size(&cpi->oxcf, cm->width, cm->height,
cpi->ppi->number_spatial_layers));
set_tile_info(cm, &cpi->oxcf.tile_cfg);
}
static inline int does_level_match(int width, int height, double fps,
int lvl_width, int lvl_height,
double lvl_fps, int lvl_dim_mult) {
const int64_t lvl_luma_pels = (int64_t)lvl_width * lvl_height;
const double lvl_display_sample_rate = lvl_luma_pels * lvl_fps;
const int64_t luma_pels = (int64_t)width * height;
const double display_sample_rate = luma_pels * fps;
return luma_pels <= lvl_luma_pels &&
display_sample_rate <= lvl_display_sample_rate &&
width <= lvl_width * lvl_dim_mult &&
height <= lvl_height * lvl_dim_mult;
}
static void set_bitstream_level_tier(AV1_PRIMARY *const ppi, int width,
int height, double init_framerate) {
SequenceHeader *const seq_params = &ppi->seq_params;
const AV1LevelParams *const level_params = &ppi->level_params;
// TODO(any): This is a placeholder function that only addresses dimensions
// and max display sample rates.
// Need to add checks for max bit rate, max decoded luma sample rate, header
// rate, etc. that are not covered by this function.
AV1_LEVEL level = SEQ_LEVEL_MAX;
if (does_level_match(width, height, init_framerate, 512, 288, 30.0, 4)) {
level = SEQ_LEVEL_2_0;
} else if (does_level_match(width, height, init_framerate, 704, 396, 30.0,
4)) {
level = SEQ_LEVEL_2_1;
} else if (does_level_match(width, height, init_framerate, 1088, 612, 30.0,
4)) {
level = SEQ_LEVEL_3_0;
} else if (does_level_match(width, height, init_framerate, 1376, 774, 30.0,
4)) {
level = SEQ_LEVEL_3_1;
} else if (does_level_match(width, height, init_framerate, 2048, 1152, 30.0,
3)) {
level = SEQ_LEVEL_4_0;
} else if (does_level_match(width, height, init_framerate, 2048, 1152, 60.0,
3)) {
level = SEQ_LEVEL_4_1;
} else if (does_level_match(width, height, init_framerate, 4096, 2176, 30.0,
2)) {
level = SEQ_LEVEL_5_0;
} else if (does_level_match(width, height, init_framerate, 4096, 2176, 60.0,
2)) {
level = SEQ_LEVEL_5_1;
} else if (does_level_match(width, height, init_framerate, 4096, 2176, 120.0,
2)) {
level = SEQ_LEVEL_5_2;
} else if (does_level_match(width, height, init_framerate, 8192, 4352, 30.0,
2)) {
level = SEQ_LEVEL_6_0;
} else if (does_level_match(width, height, init_framerate, 8192, 4352, 60.0,
2)) {
level = SEQ_LEVEL_6_1;
} else if (does_level_match(width, height, init_framerate, 8192, 4352, 120.0,
2)) {
level = SEQ_LEVEL_6_2;
}
#if CONFIG_CWG_C013
// TODO(bohanli): currently target level is only working for the 0th operating
// point, so scalable coding is not supported.
else if (level_params->target_seq_level_idx[0] >= SEQ_LEVEL_7_0 &&
level_params->target_seq_level_idx[0] <= SEQ_LEVEL_8_3) {
// Only use level 7.x to 8.x when explicitly asked to.
if (does_level_match(width, height, init_framerate, 16384, 8704, 30.0, 2)) {
level = SEQ_LEVEL_7_0;
} else if (does_level_match(width, height, init_framerate, 16384, 8704,
60.0, 2)) {
level = SEQ_LEVEL_7_1;
} else if (does_level_match(width, height, init_framerate, 16384, 8704,
120.0, 2)) {
level = SEQ_LEVEL_7_2;
} else if (does_level_match(width, height, init_framerate, 32768, 17408,
30.0, 2)) {
level = SEQ_LEVEL_8_0;
} else if (does_level_match(width, height, init_framerate, 32768, 17408,
60.0, 2)) {
level = SEQ_LEVEL_8_1;
} else if (does_level_match(width, height, init_framerate, 32768, 17408,
120.0, 2)) {
level = SEQ_LEVEL_8_2;
}
}
#endif
for (int i = 0; i < MAX_NUM_OPERATING_POINTS; ++i) {
assert(is_valid_seq_level_idx(level_params->target_seq_level_idx[i]) ||
level_params->target_seq_level_idx[i] == SEQ_LEVEL_KEEP_STATS);
// If a higher target level is specified, it is then used rather than the
// inferred one from resolution and framerate.
seq_params->seq_level_idx[i] =
level_params->target_seq_level_idx[i] < SEQ_LEVELS &&
level_params->target_seq_level_idx[i] > level
? level_params->target_seq_level_idx[i]
: level;
// Set the maximum parameters for bitrate and buffer size for this profile,
// level, and tier
seq_params->op_params[i].bitrate = av1_max_level_bitrate(
seq_params->profile, seq_params->seq_level_idx[i], seq_params->tier[i]);
// Level with seq_level_idx = 31 returns a high "dummy" bitrate to pass the
// check
if (seq_params->op_params[i].bitrate == 0)
aom_internal_error(
&ppi->error, AOM_CODEC_UNSUP_BITSTREAM,
"AV1 does not support this combination of profile, level, and tier.");
// Buffer size in bits/s is bitrate in bits/s * 1 s
seq_params->op_params[i].buffer_size = seq_params->op_params[i].bitrate;
}
}
void av1_init_seq_coding_tools(AV1_PRIMARY *const ppi,
const AV1EncoderConfig *oxcf,
int disable_frame_id_numbers) {
SequenceHeader *const seq = &ppi->seq_params;
const FrameDimensionCfg *const frm_dim_cfg = &oxcf->frm_dim_cfg;
const ToolCfg *const tool_cfg = &oxcf->tool_cfg;
seq->still_picture =
!tool_cfg->force_video_mode && (oxcf->input_cfg.limit == 1);
seq->reduced_still_picture_hdr =
seq->still_picture && !tool_cfg->full_still_picture_hdr;
seq->force_screen_content_tools = 2;
seq->force_integer_mv = 2;
seq->order_hint_info.enable_order_hint = tool_cfg->enable_order_hint;
seq->frame_id_numbers_present_flag =
!seq->reduced_still_picture_hdr &&
!oxcf->tile_cfg.enable_large_scale_tile &&
tool_cfg->error_resilient_mode && !disable_frame_id_numbers;
if (seq->reduced_still_picture_hdr) {
seq->order_hint_info.enable_order_hint = 0;
seq->force_screen_content_tools = 2;
seq->force_integer_mv = 2;
}
seq->order_hint_info.order_hint_bits_minus_1 =
seq->order_hint_info.enable_order_hint
? DEFAULT_EXPLICIT_ORDER_HINT_BITS - 1
: -1;
seq->max_frame_width = frm_dim_cfg->forced_max_frame_width
? frm_dim_cfg->forced_max_frame_width
: frm_dim_cfg->width;
seq->max_frame_height = frm_dim_cfg->forced_max_frame_height
? frm_dim_cfg->forced_max_frame_height
: frm_dim_cfg->height;
seq->num_bits_width =
(seq->max_frame_width > 1) ? get_msb(seq->max_frame_width - 1) + 1 : 1;
seq->num_bits_height =
(seq->max_frame_height > 1) ? get_msb(seq->max_frame_height - 1) + 1 : 1;
assert(seq->num_bits_width <= 16);
assert(seq->num_bits_height <= 16);
seq->frame_id_length = FRAME_ID_LENGTH;
seq->delta_frame_id_length = DELTA_FRAME_ID_LENGTH;
seq->enable_dual_filter = tool_cfg->enable_dual_filter;
seq->order_hint_info.enable_dist_wtd_comp =
oxcf->comp_type_cfg.enable_dist_wtd_comp;
seq->order_hint_info.enable_dist_wtd_comp &=
seq->order_hint_info.enable_order_hint;
seq->order_hint_info.enable_ref_frame_mvs = tool_cfg->ref_frame_mvs_present;
seq->order_hint_info.enable_ref_frame_mvs &=
seq->order_hint_info.enable_order_hint;
seq->enable_superres = oxcf->superres_cfg.enable_superres;
seq->enable_cdef = tool_cfg->cdef_control != CDEF_NONE ? 1 : 0;
seq->enable_restoration = tool_cfg->enable_restoration;
seq->enable_warped_motion = oxcf->motion_mode_cfg.enable_warped_motion;
seq->enable_interintra_compound = tool_cfg->enable_interintra_comp;
seq->enable_masked_compound = oxcf->comp_type_cfg.enable_masked_comp;
seq->enable_intra_edge_filter = oxcf->intra_mode_cfg.enable_intra_edge_filter;
seq->enable_filter_intra = oxcf->intra_mode_cfg.enable_filter_intra;
set_bitstream_level_tier(ppi, frm_dim_cfg->width, frm_dim_cfg->height,
oxcf->input_cfg.init_framerate);
if (seq->operating_points_cnt_minus_1 == 0) {
seq->operating_point_idc[0] = 0;
seq->has_nonzero_operating_point_idc = false;
} else {
// Set operating_point_idc[] such that the i=0 point corresponds to the
// highest quality operating point (all layers), and subsequent
// operarting points (i > 0) are lower quality corresponding to
// skip decoding enhancement layers (temporal first).
int i = 0;
assert(seq->operating_points_cnt_minus_1 ==
(int)(ppi->number_spatial_layers * ppi->number_temporal_layers - 1));
for (unsigned int sl = 0; sl < ppi->number_spatial_layers; sl++) {
for (unsigned int tl = 0; tl < ppi->number_temporal_layers; tl++) {
seq->operating_point_idc[i] =
(~(~0u << (ppi->number_spatial_layers - sl)) << 8) |
~(~0u << (ppi->number_temporal_layers - tl));
assert(seq->operating_point_idc[i] != 0);
i++;
}
}
seq->has_nonzero_operating_point_idc = true;
}
}
static void init_config_sequence(struct AV1_PRIMARY *ppi,
const AV1EncoderConfig *oxcf) {
SequenceHeader *const seq_params = &ppi->seq_params;
const DecoderModelCfg *const dec_model_cfg = &oxcf->dec_model_cfg;
const ColorCfg *const color_cfg = &oxcf->color_cfg;
ppi->use_svc = 0;
ppi->number_spatial_layers = 1;
ppi->number_temporal_layers = 1;
seq_params->profile = oxcf->profile;
seq_params->bit_depth = oxcf->tool_cfg.bit_depth;
seq_params->use_highbitdepth = oxcf->use_highbitdepth;
seq_params->color_primaries = color_cfg->color_primaries;
seq_params->transfer_characteristics = color_cfg->transfer_characteristics;
seq_params->matrix_coefficients = color_cfg->matrix_coefficients;
seq_params->monochrome = oxcf->tool_cfg.enable_monochrome;
seq_params->chroma_sample_position = color_cfg->chroma_sample_position;
seq_params->color_range = color_cfg->color_range;
seq_params->timing_info_present = dec_model_cfg->timing_info_present;
seq_params->timing_info.num_units_in_display_tick =
dec_model_cfg->timing_info.num_units_in_display_tick;
seq_params->timing_info.time_scale = dec_model_cfg->timing_info.time_scale;
seq_params->timing_info.equal_picture_interval =
dec_model_cfg->timing_info.equal_picture_interval;
seq_params->timing_info.num_ticks_per_picture =
dec_model_cfg->timing_info.num_ticks_per_picture;
seq_params->display_model_info_present_flag =
dec_model_cfg->display_model_info_present_flag;
seq_params->decoder_model_info_present_flag =
dec_model_cfg->decoder_model_info_present_flag;
if (dec_model_cfg->decoder_model_info_present_flag) {
// set the decoder model parameters in schedule mode
seq_params->decoder_model_info.num_units_in_decoding_tick =
dec_model_cfg->num_units_in_decoding_tick;
ppi->buffer_removal_time_present = 1;
av1_set_aom_dec_model_info(&seq_params->decoder_model_info);
av1_set_dec_model_op_parameters(&seq_params->op_params[0]);
} else if (seq_params->timing_info_present &&
seq_params->timing_info.equal_picture_interval &&
!seq_params->decoder_model_info_present_flag) {
// set the decoder model parameters in resource availability mode
av1_set_resource_availability_parameters(&seq_params->op_params[0]);
} else {
seq_params->op_params[0].initial_display_delay =
10; // Default value (not signaled)
}
if (seq_params->monochrome) {
seq_params->subsampling_x = 1;
seq_params->subsampling_y = 1;
} else if (seq_params->color_primaries == AOM_CICP_CP_BT_709 &&
seq_params->transfer_characteristics == AOM_CICP_TC_SRGB &&
seq_params->matrix_coefficients == AOM_CICP_MC_IDENTITY) {
seq_params->subsampling_x = 0;
seq_params->subsampling_y = 0;
} else {
if (seq_params->profile == 0) {
seq_params->subsampling_x = 1;
seq_params->subsampling_y = 1;
} else if (seq_params->profile == 1) {
seq_params->subsampling_x = 0;
seq_params->subsampling_y = 0;
} else {
if (seq_params->bit_depth == AOM_BITS_12) {
seq_params->subsampling_x = oxcf->input_cfg.chroma_subsampling_x;
seq_params->subsampling_y = oxcf->input_cfg.chroma_subsampling_y;
} else {
seq_params->subsampling_x = 1;
seq_params->subsampling_y = 0;
}
}
}
av1_change_config_seq(ppi, oxcf, NULL);
}
static void init_config(struct AV1_COMP *cpi, const AV1EncoderConfig *oxcf) {
AV1_COMMON *const cm = &cpi->common;
ResizePendingParams *resize_pending_params = &cpi->resize_pending_params;
cpi->oxcf = *oxcf;
cpi->framerate = oxcf->input_cfg.init_framerate;
cm->width = oxcf->frm_dim_cfg.width;
cm->height = oxcf->frm_dim_cfg.height;
cpi->is_dropped_frame = false;
alloc_compressor_data(cpi);
cpi->data_alloc_width = cm->width;
cpi->data_alloc_height = cm->height;
cpi->frame_size_related_setup_done = false;
// Single thread case: use counts in common.
cpi->td.counts = &cpi->counts;
// Init SVC parameters.
cpi->svc.number_spatial_layers = 1;
cpi->svc.number_temporal_layers = 1;
cm->spatial_layer_id = 0;
cm->temporal_layer_id = 0;
// Init rtc_ref parameters.
cpi->ppi->rtc_ref.set_ref_frame_config = 0;
cpi->ppi->rtc_ref.non_reference_frame = 0;
cpi->ppi->rtc_ref.ref_frame_comp[0] = 0;
cpi->ppi->rtc_ref.ref_frame_comp[1] = 0;
cpi->ppi->rtc_ref.ref_frame_comp[2] = 0;
// change includes all joint functionality
av1_change_config(cpi, oxcf, false);
cpi->ref_frame_flags = 0;
// Reset resize pending flags
resize_pending_params->width = 0;
resize_pending_params->height = 0;
// Setup identity scale factor
av1_setup_scale_factors_for_frame(&cm->sf_identity, 1, 1, 1, 1);
init_buffer_indices(&cpi->force_intpel_info, cm->remapped_ref_idx);
av1_noise_estimate_init(&cpi->noise_estimate, cm->width, cm->height);
}
void av1_change_config_seq(struct AV1_PRIMARY *ppi,
const AV1EncoderConfig *oxcf,
bool *is_sb_size_changed) {
SequenceHeader *const seq_params = &ppi->seq_params;
const FrameDimensionCfg *const frm_dim_cfg = &oxcf->frm_dim_cfg;
const DecoderModelCfg *const dec_model_cfg = &oxcf->dec_model_cfg;
const ColorCfg *const color_cfg = &oxcf->color_cfg;
if (seq_params->profile != oxcf->profile) seq_params->profile = oxcf->profile;
seq_params->bit_depth = oxcf->tool_cfg.bit_depth;
seq_params->color_primaries = color_cfg->color_primaries;
seq_params->transfer_characteristics = color_cfg->transfer_characteristics;
seq_params->matrix_coefficients = color_cfg->matrix_coefficients;
seq_params->monochrome = oxcf->tool_cfg.enable_monochrome;
seq_params->chroma_sample_position = color_cfg->chroma_sample_position;
seq_params->color_range = color_cfg->color_range;
assert(IMPLIES(seq_params->profile <= PROFILE_1,
seq_params->bit_depth <= AOM_BITS_10));
seq_params->timing_info_present = dec_model_cfg->timing_info_present;
seq_params->timing_info.num_units_in_display_tick =
dec_model_cfg->timing_info.num_units_in_display_tick;
seq_params->timing_info.time_scale = dec_model_cfg->timing_info.time_scale;
seq_params->timing_info.equal_picture_interval =
dec_model_cfg->timing_info.equal_picture_interval;
seq_params->timing_info.num_ticks_per_picture =
dec_model_cfg->timing_info.num_ticks_per_picture;
seq_params->display_model_info_present_flag =
dec_model_cfg->display_model_info_present_flag;
seq_params->decoder_model_info_present_flag =
dec_model_cfg->decoder_model_info_present_flag;
if (dec_model_cfg->decoder_model_info_present_flag) {
// set the decoder model parameters in schedule mode
seq_params->decoder_model_info.num_units_in_decoding_tick =
dec_model_cfg->num_units_in_decoding_tick;
ppi->buffer_removal_time_present = 1;
av1_set_aom_dec_model_info(&seq_params->decoder_model_info);
av1_set_dec_model_op_parameters(&seq_params->op_params[0]);
} else if (seq_params->timing_info_present &&
seq_params->timing_info.equal_picture_interval &&
!seq_params->decoder_model_info_present_flag) {
// set the decoder model parameters in resource availability mode
av1_set_resource_availability_parameters(&seq_params->op_params[0]);
} else {
seq_params->op_params[0].initial_display_delay =
10; // Default value (not signaled)
}
av1_update_film_grain_parameters_seq(ppi, oxcf);
int sb_size = seq_params->sb_size;
// Superblock size should not be updated after the first key frame.
if (!ppi->seq_params_locked) {
set_sb_size(seq_params, av1_select_sb_size(oxcf, frm_dim_cfg->width,
frm_dim_cfg->height,
ppi->number_spatial_layers));
for (int i = 0; i < MAX_NUM_OPERATING_POINTS; ++i)
seq_params->tier[i] = (oxcf->tier_mask >> i) & 1;
}
if (is_sb_size_changed != NULL && sb_size != seq_params->sb_size)
*is_sb_size_changed = true;
// Init sequence level coding tools
// This should not be called after the first key frame.
if (!ppi->seq_params_locked) {
seq_params->operating_points_cnt_minus_1 =
(ppi->number_spatial_layers > 1 || ppi->number_temporal_layers > 1)
? ppi->number_spatial_layers * ppi->number_temporal_layers - 1
: 0;
av1_init_seq_coding_tools(
ppi, oxcf, ppi->use_svc || ppi->rtc_ref.set_ref_frame_config);
}
seq_params->timing_info_present &= !seq_params->reduced_still_picture_hdr;
#if CONFIG_AV1_HIGHBITDEPTH
highbd_set_var_fns(ppi);
#endif
set_primary_rc_buffer_sizes(oxcf, ppi);
}
void av1_change_config(struct AV1_COMP *cpi, const AV1EncoderConfig *oxcf,
bool is_sb_size_changed) {
AV1_COMMON *const cm = &cpi->common;
SequenceHeader *const seq_params = cm->seq_params;
RATE_CONTROL *const rc = &cpi->rc;
PRIMARY_RATE_CONTROL *const p_rc = &cpi->ppi->p_rc;
MACROBLOCK *const x = &cpi->td.mb;
AV1LevelParams *const level_params = &cpi->ppi->level_params;
RefreshFrameInfo *const refresh_frame = &cpi->refresh_frame;
const FrameDimensionCfg *const frm_dim_cfg = &cpi->oxcf.frm_dim_cfg;
const RateControlCfg *const rc_cfg = &oxcf->rc_cfg;
FeatureFlags *const features = &cm->features;
// in case of LAP, lag in frames is set according to number of lap buffers
// calculated at init time. This stores and restores LAP's lag in frames to
// prevent override by new cfg.
int lap_lag_in_frames = -1;
if (cpi->ppi->lap_enabled && cpi->compressor_stage == LAP_STAGE) {
lap_lag_in_frames = cpi->oxcf.gf_cfg.lag_in_frames;
}
cpi->oxcf = *oxcf;
av1_update_film_grain_parameters(cpi, oxcf);
// When user provides superres_mode = AOM_SUPERRES_AUTO, we still initialize
// superres mode for current encoding = AOM_SUPERRES_NONE. This is to ensure
// that any analysis (e.g. TPL) happening outside the main encoding loop still
// happens at full resolution.
// This value will later be set appropriately just before main encoding loop.
cpi->superres_mode = oxcf->superres_cfg.superres_mode == AOM_SUPERRES_AUTO
? AOM_SUPERRES_NONE
: oxcf->superres_cfg.superres_mode; // default
x->e_mbd.bd = (int)seq_params->bit_depth;
x->e_mbd.global_motion = cm->global_motion;
memcpy(level_params->target_seq_level_idx, cpi->oxcf.target_seq_level_idx,
sizeof(level_params->target_seq_level_idx));
level_params->keep_level_stats = 0;
for (int i = 0; i < MAX_NUM_OPERATING_POINTS; ++i) {
if (level_params->target_seq_level_idx[i] < SEQ_LEVELS ||
level_params->target_seq_level_idx[i] == SEQ_LEVEL_KEEP_STATS) {
level_params->keep_level_stats |= 1u << i;
if (!level_params->level_info[i]) {
CHECK_MEM_ERROR(cm, level_params->level_info[i],
aom_calloc(1, sizeof(*level_params->level_info[i])));
}
}
}
// TODO(huisu@): level targeting currently only works for the 0th operating
// point, so scalable coding is not supported yet.
if (level_params->target_seq_level_idx[0] < SEQ_LEVELS) {
// Adjust encoder config in order to meet target level.
config_target_level(cpi, level_params->target_seq_level_idx[0],
seq_params->tier[0]);
}
if (has_no_stats_stage(cpi) && (rc_cfg->mode == AOM_Q)) {
p_rc->baseline_gf_interval = FIXED_GF_INTERVAL;
} else if (!is_one_pass_rt_params(cpi) ||
cm->current_frame.frame_number == 0) {
// For rtc mode: logic for setting the baseline_gf_interval is done
// in av1_get_one_pass_rt_params(), and it should not be reset here in
// change_config(), unless after init_config (first frame).
p_rc->baseline_gf_interval = (MIN_GF_INTERVAL + MAX_GF_INTERVAL) / 2;
}
refresh_frame->golden_frame = false;
refresh_frame->bwd_ref_frame = false;
features->refresh_frame_context =
(oxcf->tool_cfg.frame_parallel_decoding_mode)
? REFRESH_FRAME_CONTEXT_DISABLED
: REFRESH_FRAME_CONTEXT_BACKWARD;
if (oxcf->tile_cfg.enable_large_scale_tile)
features->refresh_frame_context = REFRESH_FRAME_CONTEXT_DISABLED;
if (x->palette_buffer == NULL) {
CHECK_MEM_ERROR(cm, x->palette_buffer,
aom_memalign(16, sizeof(*x->palette_buffer)));
}
if (x->tmp_conv_dst == NULL) {
CHECK_MEM_ERROR(
cm, x->tmp_conv_dst,
aom_memalign(32, MAX_SB_SIZE * MAX_SB_SIZE * sizeof(*x->tmp_conv_dst)));
x->e_mbd.tmp_conv_dst = x->tmp_conv_dst;
}
// The buffers 'tmp_pred_bufs[]' and 'comp_rd_buffer' are used in inter frames
// to store intermediate inter mode prediction results and are not required
// for allintra encoding mode. Hence, the memory allocations for these buffers
// are avoided for allintra encoding mode.
if (cpi->oxcf.kf_cfg.key_freq_max != 0) {
if (x->comp_rd_buffer.pred0 == NULL)
alloc_compound_type_rd_buffers(cm->error, &x->comp_rd_buffer);
for (int i = 0; i < 2; ++i) {
if (x->tmp_pred_bufs[i] == NULL) {
CHECK_MEM_ERROR(cm, x->tmp_pred_bufs[i],
aom_memalign(32, 2 * MAX_MB_PLANE * MAX_SB_SQUARE *
sizeof(*x->tmp_pred_bufs[i])));
x->e_mbd.tmp_obmc_bufs[i] = x->tmp_pred_bufs[i];
}
}
}
av1_reset_segment_features(cm);
av1_set_high_precision_mv(cpi, 1, 0);
// Under a configuration change, where maximum_buffer_size may change,
// keep buffer level clipped to the maximum allowed buffer size.
p_rc->bits_off_target =
AOMMIN(p_rc->bits_off_target, p_rc->maximum_buffer_size);
p_rc->buffer_level = AOMMIN(p_rc->buffer_level, p_rc->maximum_buffer_size);
// Set up frame rate and related parameters rate control values.
av1_new_framerate(cpi, cpi->framerate);
// Set absolute upper and lower quality limits
rc->worst_quality = rc_cfg->worst_allowed_q;
rc->best_quality = rc_cfg->best_allowed_q;
// If lossless has been requested make sure average Q accumulators are reset.
if (is_lossless_requested(&cpi->oxcf.rc_cfg)) {
int i;
for (i = 0; i < FRAME_TYPES; ++i) {
p_rc->avg_frame_qindex[i] = 0;
}
}
features->interp_filter =
oxcf->tile_cfg.enable_large_scale_tile ? EIGHTTAP_REGULAR : SWITCHABLE;
features->switchable_motion_mode = is_switchable_motion_mode_allowed(
features->allow_warped_motion, oxcf->motion_mode_cfg.enable_obmc);
if (frm_dim_cfg->render_width > 0 && frm_dim_cfg->render_height > 0) {
cm->render_width = frm_dim_cfg->render_width;
cm->render_height = frm_dim_cfg->render_height;
} else {
cm->render_width = frm_dim_cfg->width;
cm->render_height = frm_dim_cfg->height;
}
cm->width = frm_dim_cfg->width;
cm->height = frm_dim_cfg->height;
if (cm->width > cpi->data_alloc_width ||
cm->height > cpi->data_alloc_height || is_sb_size_changed) {
av1_free_context_buffers(cm);
av1_free_shared_coeff_buffer(&cpi->td.shared_coeff_buf);
av1_free_sms_tree(&cpi->td);
av1_free_pmc(cpi->td.firstpass_ctx, av1_num_planes(cm));
cpi->td.firstpass_ctx = NULL;
alloc_compressor_data(cpi);
realloc_segmentation_maps(cpi);
cpi->data_alloc_width = cm->width;
cpi->data_alloc_height = cm->height;
cpi->frame_size_related_setup_done = false;
}
av1_update_frame_size(cpi);
rc->is_src_frame_alt_ref = 0;
if (!cpi->ppi->rtc_ref.set_ref_frame_config)
cpi->ext_flags.refresh_frame.update_pending = 0;
cpi->ext_flags.refresh_frame_context_pending = 0;
if (cpi->ppi->use_svc)
av1_update_layer_context_change_config(cpi, rc_cfg->target_bandwidth);
check_reset_rc_flag(cpi);
// restore the value of lag_in_frame for LAP stage.
if (lap_lag_in_frames != -1) {
cpi->oxcf.gf_cfg.lag_in_frames = lap_lag_in_frames;
}
#if CONFIG_REALTIME_ONLY
assert(!oxcf->tool_cfg.enable_global_motion);
cpi->alloc_pyramid = false;
#else
cpi->alloc_pyramid = oxcf->tool_cfg.enable_global_motion;
#endif // CONFIG_REALTIME_ONLY
}
static inline void init_frame_info(FRAME_INFO *frame_info,
const AV1_COMMON *const cm) {
const CommonModeInfoParams *const mi_params = &cm->mi_params;
const SequenceHeader *const seq_params = cm->seq_params;
frame_info->frame_width = cm->width;
frame_info->frame_height = cm->height;
frame_info->mi_cols = mi_params->mi_cols;
frame_info->mi_rows = mi_params->mi_rows;
frame_info->mb_cols = mi_params->mb_cols;
frame_info->mb_rows = mi_params->mb_rows;
frame_info->num_mbs = mi_params->MBs;
frame_info->bit_depth = seq_params->bit_depth;
frame_info->subsampling_x = seq_params->subsampling_x;
frame_info->subsampling_y = seq_params->subsampling_y;
}
static inline void init_frame_index_set(FRAME_INDEX_SET *frame_index_set) {
frame_index_set->show_frame_count = 0;
}
static inline void update_counters_for_show_frame(AV1_COMP *const cpi) {
assert(cpi->common.show_frame);
cpi->frame_index_set.show_frame_count++;
cpi->common.current_frame.frame_number++;
}
AV1_PRIMARY *av1_create_primary_compressor(
struct aom_codec_pkt_list *pkt_list_head, int num_lap_buffers,
const AV1EncoderConfig *oxcf) {
AV1_PRIMARY *volatile const ppi = aom_memalign(32, sizeof(AV1_PRIMARY));
if (!ppi) return NULL;
av1_zero(*ppi);
// The jmp_buf is valid only for the duration of the function that calls
// setjmp(). Therefore, this function must reset the 'setjmp' field to 0
// before it returns.
if (setjmp(ppi->error.jmp)) {
ppi->error.setjmp = 0;
av1_remove_primary_compressor(ppi);
return 0;
}
ppi->error.setjmp = 1;
ppi->seq_params_locked = 0;
ppi->lap_enabled = num_lap_buffers > 0;
ppi->output_pkt_list = pkt_list_head;
ppi->b_calculate_psnr = CONFIG_INTERNAL_STATS;
ppi->frames_left = oxcf->input_cfg.limit;
ppi->num_fp_contexts = 1;
init_config_sequence(ppi, oxcf);
#if CONFIG_ENTROPY_STATS
av1_zero(ppi->aggregate_fc);
#endif // CONFIG_ENTROPY_STATS
av1_primary_rc_init(oxcf, &ppi->p_rc);
// For two pass and lag_in_frames > 33 in LAP.
ppi->p_rc.enable_scenecut_detection = ENABLE_SCENECUT_MODE_2;
if (ppi->lap_enabled) {
if ((num_lap_buffers <
(MAX_GF_LENGTH_LAP + SCENE_CUT_KEY_TEST_INTERVAL + 1)) &&
num_lap_buffers >= (MAX_GF_LENGTH_LAP + 3)) {
/*
* For lag in frames >= 19 and <33, enable scenecut
* with limited future frame prediction.
*/
ppi->p_rc.enable_scenecut_detection = ENABLE_SCENECUT_MODE_1;
} else if (num_lap_buffers < (MAX_GF_LENGTH_LAP + 3)) {
// Disable scenecut when lag_in_frames < 19.
ppi->p_rc.enable_scenecut_detection = DISABLE_SCENECUT;
}
}
#define BFP(BT, SDF, SDAF, VF, SVF, SVAF, SDX4DF, SDX3DF, JSDAF, JSVAF) \
ppi->fn_ptr[BT].sdf = SDF; \
ppi->fn_ptr[BT].sdaf = SDAF; \
ppi->fn_ptr[BT].vf = VF; \
ppi->fn_ptr[BT].svf = SVF; \
ppi->fn_ptr[BT].svaf = SVAF; \
ppi->fn_ptr[BT].sdx4df = SDX4DF; \
ppi->fn_ptr[BT].jsdaf = JSDAF; \
ppi->fn_ptr[BT].jsvaf = JSVAF; \
ppi->fn_ptr[BT].sdx3df = SDX3DF;
// Realtime mode doesn't use 4x rectangular blocks.
#if !CONFIG_REALTIME_ONLY
BFP(BLOCK_4X16, aom_sad4x16, aom_sad4x16_avg, aom_variance4x16,
aom_sub_pixel_variance4x16, aom_sub_pixel_avg_variance4x16,
aom_sad4x16x4d, aom_sad4x16x3d, aom_dist_wtd_sad4x16_avg,
aom_dist_wtd_sub_pixel_avg_variance4x16)
BFP(BLOCK_16X4, aom_sad16x4, aom_sad16x4_avg, aom_variance16x4,
aom_sub_pixel_variance16x4, aom_sub_pixel_avg_variance16x4,
aom_sad16x4x4d, aom_sad16x4x3d, aom_dist_wtd_sad16x4_avg,
aom_dist_wtd_sub_pixel_avg_variance16x4)
BFP(BLOCK_8X32, aom_sad8x32, aom_sad8x32_avg, aom_variance8x32,
aom_sub_pixel_variance8x32, aom_sub_pixel_avg_variance8x32,
aom_sad8x32x4d, aom_sad8x32x3d, aom_dist_wtd_sad8x32_avg,
aom_dist_wtd_sub_pixel_avg_variance8x32)
BFP(BLOCK_32X8, aom_sad32x8, aom_sad32x8_avg, aom_variance32x8,
aom_sub_pixel_variance32x8, aom_sub_pixel_avg_variance32x8,
aom_sad32x8x4d, aom_sad32x8x3d, aom_dist_wtd_sad32x8_avg,
aom_dist_wtd_sub_pixel_avg_variance32x8)
BFP(BLOCK_16X64, aom_sad16x64, aom_sad16x64_avg, aom_variance16x64,
aom_sub_pixel_variance16x64, aom_sub_pixel_avg_variance16x64,
aom_sad16x64x4d, aom_sad16x64x3d, aom_dist_wtd_sad16x64_avg,
aom_dist_wtd_sub_pixel_avg_variance16x64)
BFP(BLOCK_64X16, aom_sad64x16, aom_sad64x16_avg, aom_variance64x16,
aom_sub_pixel_variance64x16, aom_sub_pixel_avg_variance64x16,
aom_sad64x16x4d, aom_sad64x16x3d, aom_dist_wtd_sad64x16_avg,
aom_dist_wtd_sub_pixel_avg_variance64x16)
#endif // !CONFIG_REALTIME_ONLY
BFP(BLOCK_128X128, aom_sad128x128, aom_sad128x128_avg, aom_variance128x128,
aom_sub_pixel_variance128x128, aom_sub_pixel_avg_variance128x128,
aom_sad128x128x4d, aom_sad128x128x3d, aom_dist_wtd_sad128x128_avg,
aom_dist_wtd_sub_pixel_avg_variance128x128)
BFP(BLOCK_128X64, aom_sad128x64, aom_sad128x64_avg, aom_variance128x64,
aom_sub_pixel_variance128x64, aom_sub_pixel_avg_variance128x64,
aom_sad128x64x4d, aom_sad128x64x3d, aom_dist_wtd_sad128x64_avg,
aom_dist_wtd_sub_pixel_avg_variance128x64)
BFP(BLOCK_64X128, aom_sad64x128, aom_sad64x128_avg, aom_variance64x128,
aom_sub_pixel_variance64x128, aom_sub_pixel_avg_variance64x128,
aom_sad64x128x4d, aom_sad64x128x3d, aom_dist_wtd_sad64x128_avg,
aom_dist_wtd_sub_pixel_avg_variance64x128)
BFP(BLOCK_32X16, aom_sad32x16, aom_sad32x16_avg, aom_variance32x16,
aom_sub_pixel_variance32x16, aom_sub_pixel_avg_variance32x16,
aom_sad32x16x4d, aom_sad32x16x3d, aom_dist_wtd_sad32x16_avg,
aom_dist_wtd_sub_pixel_avg_variance32x16)
BFP(BLOCK_16X32, aom_sad16x32, aom_sad16x32_avg, aom_variance16x32,
aom_sub_pixel_variance16x32, aom_sub_pixel_avg_variance16x32,
aom_sad16x32x4d, aom_sad16x32x3d, aom_dist_wtd_sad16x32_avg,
aom_dist_wtd_sub_pixel_avg_variance16x32)
BFP(BLOCK_64X32, aom_sad64x32, aom_sad64x32_avg, aom_variance64x32,
aom_sub_pixel_variance64x32, aom_sub_pixel_avg_variance64x32,
aom_sad64x32x4d, aom_sad64x32x3d, aom_dist_wtd_sad64x32_avg,
aom_dist_wtd_sub_pixel_avg_variance64x32)
BFP(BLOCK_32X64, aom_sad32x64, aom_sad32x64_avg, aom_variance32x64,
aom_sub_pixel_variance32x64, aom_sub_pixel_avg_variance32x64,
aom_sad32x64x4d, aom_sad32x64x3d, aom_dist_wtd_sad32x64_avg,
aom_dist_wtd_sub_pixel_avg_variance32x64)
BFP(BLOCK_32X32, aom_sad32x32, aom_sad32x32_avg, aom_variance32x32,
aom_sub_pixel_variance32x32, aom_sub_pixel_avg_variance32x32,
aom_sad32x32x4d, aom_sad32x32x3d, aom_dist_wtd_sad32x32_avg,
aom_dist_wtd_sub_pixel_avg_variance32x32)
BFP(BLOCK_64X64, aom_sad64x64, aom_sad64x64_avg, aom_variance64x64,
aom_sub_pixel_variance64x64, aom_sub_pixel_avg_variance64x64,
aom_sad64x64x4d, aom_sad64x64x3d, aom_dist_wtd_sad64x64_avg,
aom_dist_wtd_sub_pixel_avg_variance64x64)
BFP(BLOCK_16X16, aom_sad16x16, aom_sad16x16_avg, aom_variance16x16,
aom_sub_pixel_variance16x16, aom_sub_pixel_avg_variance16x16,
aom_sad16x16x4d, aom_sad16x16x3d, aom_dist_wtd_sad16x16_avg,
aom_dist_wtd_sub_pixel_avg_variance16x16)
BFP(BLOCK_16X8, aom_sad16x8, aom_sad16x8_avg, aom_variance16x8,
aom_sub_pixel_variance16x8, aom_sub_pixel_avg_variance16x8,
aom_sad16x8x4d, aom_sad16x8x3d, aom_dist_wtd_sad16x8_avg,
aom_dist_wtd_sub_pixel_avg_variance16x8)
BFP(BLOCK_8X16, aom_sad8x16, aom_sad8x16_avg, aom_variance8x16,
aom_sub_pixel_variance8x16, aom_sub_pixel_avg_variance8x16,
aom_sad8x16x4d, aom_sad8x16x3d, aom_dist_wtd_sad8x16_avg,
aom_dist_wtd_sub_pixel_avg_variance8x16)
BFP(BLOCK_8X8, aom_sad8x8, aom_sad8x8_avg, aom_variance8x8,
aom_sub_pixel_variance8x8, aom_sub_pixel_avg_variance8x8, aom_sad8x8x4d,
aom_sad8x8x3d, aom_dist_wtd_sad8x8_avg,
aom_dist_wtd_sub_pixel_avg_variance8x8)
BFP(BLOCK_8X4, aom_sad8x4, aom_sad8x4_avg, aom_variance8x4,
aom_sub_pixel_variance8x4, aom_sub_pixel_avg_variance8x4, aom_sad8x4x4d,
aom_sad8x4x3d, aom_dist_wtd_sad8x4_avg,
aom_dist_wtd_sub_pixel_avg_variance8x4)
BFP(BLOCK_4X8, aom_sad4x8, aom_sad4x8_avg, aom_variance4x8,
aom_sub_pixel_variance4x8, aom_sub_pixel_avg_variance4x8, aom_sad4x8x4d,
aom_sad4x8x3d, aom_dist_wtd_sad4x8_avg,
aom_dist_wtd_sub_pixel_avg_variance4x8)
BFP(BLOCK_4X4, aom_sad4x4, aom_sad4x4_avg, aom_variance4x4,
aom_sub_pixel_variance4x4, aom_sub_pixel_avg_variance4x4, aom_sad4x4x4d,
aom_sad4x4x3d, aom_dist_wtd_sad4x4_avg,
aom_dist_wtd_sub_pixel_avg_variance4x4)
#if !CONFIG_REALTIME_ONLY
#define OBFP(BT, OSDF, OVF, OSVF) \
ppi->fn_ptr[BT].osdf = OSDF; \
ppi->fn_ptr[BT].ovf = OVF; \
ppi->fn_ptr[BT].osvf = OSVF;
OBFP(BLOCK_128X128, aom_obmc_sad128x128, aom_obmc_variance128x128,
aom_obmc_sub_pixel_variance128x128)
OBFP(BLOCK_128X64, aom_obmc_sad128x64, aom_obmc_variance128x64,
aom_obmc_sub_pixel_variance128x64)
OBFP(BLOCK_64X128, aom_obmc_sad64x128, aom_obmc_variance64x128,
aom_obmc_sub_pixel_variance64x128)
OBFP(BLOCK_64X64, aom_obmc_sad64x64, aom_obmc_variance64x64,
aom_obmc_sub_pixel_variance64x64)
OBFP(BLOCK_64X32, aom_obmc_sad64x32, aom_obmc_variance64x32,
aom_obmc_sub_pixel_variance64x32)
OBFP(BLOCK_32X64, aom_obmc_sad32x64, aom_obmc_variance32x64,
aom_obmc_sub_pixel_variance32x64)
OBFP(BLOCK_32X32, aom_obmc_sad32x32, aom_obmc_variance32x32,
aom_obmc_sub_pixel_variance32x32)
OBFP(BLOCK_32X16, aom_obmc_sad32x16, aom_obmc_variance32x16,
aom_obmc_sub_pixel_variance32x16)
OBFP(BLOCK_16X32, aom_obmc_sad16x32, aom_obmc_variance16x32,
aom_obmc_sub_pixel_variance16x32)
OBFP(BLOCK_16X16, aom_obmc_sad16x16, aom_obmc_variance16x16,
aom_obmc_sub_pixel_variance16x16)
OBFP(BLOCK_16X8, aom_obmc_sad16x8, aom_obmc_variance16x8,
aom_obmc_sub_pixel_variance16x8)
OBFP(BLOCK_8X16, aom_obmc_sad8x16, aom_obmc_variance8x16,
aom_obmc_sub_pixel_variance8x16)
OBFP(BLOCK_8X8, aom_obmc_sad8x8, aom_obmc_variance8x8,
aom_obmc_sub_pixel_variance8x8)
OBFP(BLOCK_4X8, aom_obmc_sad4x8, aom_obmc_variance4x8,
aom_obmc_sub_pixel_variance4x8)
OBFP(BLOCK_8X4, aom_obmc_sad8x4, aom_obmc_variance8x4,
aom_obmc_sub_pixel_variance8x4)
OBFP(BLOCK_4X4, aom_obmc_sad4x4, aom_obmc_variance4x4,
aom_obmc_sub_pixel_variance4x4)
OBFP(BLOCK_4X16, aom_obmc_sad4x16, aom_obmc_variance4x16,
aom_obmc_sub_pixel_variance4x16)
OBFP(BLOCK_16X4, aom_obmc_sad16x4, aom_obmc_variance16x4,
aom_obmc_sub_pixel_variance16x4)
OBFP(BLOCK_8X32, aom_obmc_sad8x32, aom_obmc_variance8x32,
aom_obmc_sub_pixel_variance8x32)
OBFP(BLOCK_32X8, aom_obmc_sad32x8, aom_obmc_variance32x8,
aom_obmc_sub_pixel_variance32x8)
OBFP(BLOCK_16X64, aom_obmc_sad16x64, aom_obmc_variance16x64,
aom_obmc_sub_pixel_variance16x64)
OBFP(BLOCK_64X16, aom_obmc_sad64x16, aom_obmc_variance64x16,
aom_obmc_sub_pixel_variance64x16)
#endif // !CONFIG_REALTIME_ONLY
#define MBFP(BT, MCSDF, MCSVF) \
ppi->fn_ptr[BT].msdf = MCSDF; \
ppi->fn_ptr[BT].msvf = MCSVF;
MBFP(BLOCK_128X128, aom_masked_sad128x128,
aom_masked_sub_pixel_variance128x128)
MBFP(BLOCK_128X64, aom_masked_sad128x64, aom_masked_sub_pixel_variance128x64)
MBFP(BLOCK_64X128, aom_masked_sad64x128, aom_masked_sub_pixel_variance64x128)
MBFP(BLOCK_64X64, aom_masked_sad64x64, aom_masked_sub_pixel_variance64x64)
MBFP(BLOCK_64X32, aom_masked_sad64x32, aom_masked_sub_pixel_variance64x32)
MBFP(BLOCK_32X64, aom_masked_sad32x64, aom_masked_sub_pixel_variance32x64)
MBFP(BLOCK_32X32, aom_masked_sad32x32, aom_masked_sub_pixel_variance32x32)
MBFP(BLOCK_32X16, aom_masked_sad32x16, aom_masked_sub_pixel_variance32x16)
MBFP(BLOCK_16X32, aom_masked_sad16x32, aom_masked_sub_pixel_variance16x32)
MBFP(BLOCK_16X16, aom_masked_sad16x16, aom_masked_sub_pixel_variance16x16)
MBFP(BLOCK_16X8, aom_masked_sad16x8, aom_masked_sub_pixel_variance16x8)
MBFP(BLOCK_8X16, aom_masked_sad8x16, aom_masked_sub_pixel_variance8x16)
MBFP(BLOCK_8X8, aom_masked_sad8x8, aom_masked_sub_pixel_variance8x8)
MBFP(BLOCK_4X8, aom_masked_sad4x8, aom_masked_sub_pixel_variance4x8)
MBFP(BLOCK_8X4, aom_masked_sad8x4, aom_masked_sub_pixel_variance8x4)
MBFP(BLOCK_4X4, aom_masked_sad4x4, aom_masked_sub_pixel_variance4x4)
#if !CONFIG_REALTIME_ONLY
MBFP(BLOCK_4X16, aom_masked_sad4x16, aom_masked_sub_pixel_variance4x16)
MBFP(BLOCK_16X4, aom_masked_sad16x4, aom_masked_sub_pixel_variance16x4)
MBFP(BLOCK_8X32, aom_masked_sad8x32, aom_masked_sub_pixel_variance8x32)
MBFP(BLOCK_32X8, aom_masked_sad32x8, aom_masked_sub_pixel_variance32x8)
MBFP(BLOCK_16X64, aom_masked_sad16x64, aom_masked_sub_pixel_variance16x64)
MBFP(BLOCK_64X16, aom_masked_sad64x16, aom_masked_sub_pixel_variance64x16)
#endif
#define SDSFP(BT, SDSF, SDSX4DF) \
ppi->fn_ptr[BT].sdsf = SDSF; \
ppi->fn_ptr[BT].sdsx4df = SDSX4DF;
SDSFP(BLOCK_128X128, aom_sad_skip_128x128, aom_sad_skip_128x128x4d)
SDSFP(BLOCK_128X64, aom_sad_skip_128x64, aom_sad_skip_128x64x4d)
SDSFP(BLOCK_64X128, aom_sad_skip_64x128, aom_sad_skip_64x128x4d)
SDSFP(BLOCK_64X64, aom_sad_skip_64x64, aom_sad_skip_64x64x4d)
SDSFP(BLOCK_64X32, aom_sad_skip_64x32, aom_sad_skip_64x32x4d)
SDSFP(BLOCK_32X64, aom_sad_skip_32x64, aom_sad_skip_32x64x4d)
SDSFP(BLOCK_32X32, aom_sad_skip_32x32, aom_sad_skip_32x32x4d)
SDSFP(BLOCK_32X16, aom_sad_skip_32x16, aom_sad_skip_32x16x4d)
SDSFP(BLOCK_16X32, aom_sad_skip_16x32, aom_sad_skip_16x32x4d)
SDSFP(BLOCK_16X16, aom_sad_skip_16x16, aom_sad_skip_16x16x4d)
SDSFP(BLOCK_16X8, aom_sad_skip_16x8, aom_sad_skip_16x8x4d)
SDSFP(BLOCK_8X16, aom_sad_skip_8x16, aom_sad_skip_8x16x4d)
SDSFP(BLOCK_8X8, aom_sad_skip_8x8, aom_sad_skip_8x8x4d)
SDSFP(BLOCK_4X8, aom_sad_skip_4x8, aom_sad_skip_4x8x4d)
#if !CONFIG_REALTIME_ONLY
SDSFP(BLOCK_64X16, aom_sad_skip_64x16, aom_sad_skip_64x16x4d)
SDSFP(BLOCK_16X64, aom_sad_skip_16x64, aom_sad_skip_16x64x4d)
SDSFP(BLOCK_32X8, aom_sad_skip_32x8, aom_sad_skip_32x8x4d)
SDSFP(BLOCK_8X32, aom_sad_skip_8x32, aom_sad_skip_8x32x4d)
SDSFP(BLOCK_4X16, aom_sad_skip_4x16, aom_sad_skip_4x16x4d)
#endif
#undef SDSFP
#if CONFIG_AV1_HIGHBITDEPTH
highbd_set_var_fns(ppi);
#endif
{
// As cm->mi_params is a part of the frame level context (cpi), it is
// unavailable at this point. mi_params is created as a local temporary
// variable, to be passed into the functions used for allocating tpl
// buffers. The values in this variable are populated according to initial
// width and height of the frame.
CommonModeInfoParams mi_params;
enc_set_mb_mi(&mi_params, oxcf->frm_dim_cfg.width, oxcf->frm_dim_cfg.height,
BLOCK_4X4);
const BLOCK_SIZE bsize = BLOCK_16X16;
const int w = mi_size_wide[bsize];
const int h = mi_size_high[bsize];
const int num_cols = (mi_params.mi_cols + w - 1) / w;
const int num_rows = (mi_params.mi_rows + h - 1) / h;
AOM_CHECK_MEM_ERROR(
&ppi->error, ppi->tpl_sb_rdmult_scaling_factors,
aom_calloc(num_rows * num_cols,
sizeof(*ppi->tpl_sb_rdmult_scaling_factors)));
#if CONFIG_INTERNAL_STATS
ppi->b_calculate_blockiness = 1;
ppi->b_calculate_consistency = 1;
for (int i = 0; i <= STAT_ALL; i++) {
ppi->psnr[0].stat[i] = 0;
ppi->psnr[1].stat[i] = 0;
ppi->fastssim.stat[i] = 0;
ppi->psnrhvs.stat[i] = 0;
}
ppi->psnr[0].worst = 100.0;
ppi->psnr[1].worst = 100.0;
ppi->worst_ssim = 100.0;
ppi->worst_ssim_hbd = 100.0;
ppi->count[0] = 0;
ppi->count[1] = 0;
ppi->total_bytes = 0;
if (ppi->b_calculate_psnr) {
ppi->total_sq_error[0] = 0;
ppi->total_samples[0] = 0;
ppi->total_sq_error[1] = 0;
ppi->total_samples[1] = 0;
ppi->total_recode_hits = 0;
ppi->summed_quality = 0;
ppi->summed_weights = 0;
ppi->summed_quality_hbd = 0;
ppi->summed_weights_hbd = 0;
}
ppi->fastssim.worst = 100.0;
ppi->psnrhvs.worst = 100.0;
if (ppi->b_calculate_blockiness) {
ppi->total_blockiness = 0;
ppi->worst_blockiness = 0.0;
}
ppi->total_inconsistency = 0;
ppi->worst_consistency = 100.0;
if (ppi->b_calculate_consistency) {
AOM_CHECK_MEM_ERROR(&ppi->error, ppi->ssim_vars,
aom_malloc(sizeof(*ppi->ssim_vars) * 4 *
mi_params.mi_rows * mi_params.mi_cols));
}
#endif
}
ppi->error.setjmp = 0;
return ppi;
}
AV1_COMP *av1_create_compressor(AV1_PRIMARY *ppi, const AV1EncoderConfig *oxcf,
BufferPool *const pool, COMPRESSOR_STAGE stage,
int lap_lag_in_frames) {
AV1_COMP *volatile const cpi = aom_memalign(32, sizeof(AV1_COMP));
if (!cpi) return NULL;
av1_zero(*cpi);
cpi->ppi = ppi;
AV1_COMMON *volatile const cm = &cpi->common;
cm->seq_params = &ppi->seq_params;
cm->error =
(struct aom_internal_error_info *)aom_calloc(1, sizeof(*cm->error));
if (!cm->error) {
aom_free(cpi);
return NULL;
}
// The jmp_buf is valid only for the duration of the function that calls
// setjmp(). Therefore, this function must reset the 'setjmp' field to 0
// before it returns.
if (setjmp(cm->error->jmp)) {
cm->error->setjmp = 0;
av1_remove_compressor(cpi);
return NULL;
}
cm->error->setjmp = 1;
cpi->compressor_stage = stage;
cpi->do_frame_data_update = true;
CommonModeInfoParams *const mi_params = &cm->mi_params;
mi_params->free_mi = enc_free_mi;
mi_params->setup_mi = enc_setup_mi;
mi_params->set_mb_mi =
(oxcf->pass == AOM_RC_FIRST_PASS || cpi->compressor_stage == LAP_STAGE)
? stat_stage_set_mb_mi
: enc_set_mb_mi;
mi_params->mi_alloc_bsize = BLOCK_4X4;
CHECK_MEM_ERROR(cm, cm->fc,
(FRAME_CONTEXT *)aom_memalign(32, sizeof(*cm->fc)));
CHECK_MEM_ERROR(
cm, cm->default_frame_context,
(FRAME_CONTEXT *)aom_memalign(32, sizeof(*cm->default_frame_context)));
memset(cm->fc, 0, sizeof(*cm->fc));
memset(cm->default_frame_context, 0, sizeof(*cm->default_frame_context));
cpi->common.buffer_pool = pool;
init_config(cpi, oxcf);
if (cpi->compressor_stage == LAP_STAGE) {
cpi->oxcf.gf_cfg.lag_in_frames = lap_lag_in_frames;
}
av1_rc_init(&cpi->oxcf, &cpi->rc);
init_frame_info(&cpi->frame_info, cm);
init_frame_index_set(&cpi->frame_index_set);
cm->current_frame.frame_number = 0;
cpi->rc.frame_number_encoded = 0;
cpi->rc.prev_frame_is_dropped = 0;
cpi->rc.max_consec_drop = INT_MAX;
cpi->rc.drop_count_consec = 0;
cm->current_frame_id = -1;
cpi->tile_data = NULL;
cpi->last_show_frame_buf = NULL;
realloc_segmentation_maps(cpi);
cpi->refresh_frame.alt_ref_frame = false;
#if CONFIG_SPEED_STATS
cpi->tx_search_count = 0;
#endif // CONFIG_SPEED_STATS
cpi->time_stamps.first_ts_start = INT64_MAX;
#ifdef OUTPUT_YUV_REC
yuv_rec_file = fopen("rec.yuv", "wb");
#endif
#ifdef OUTPUT_YUV_DENOISED
yuv_denoised_file = fopen("denoised.yuv", "wb");
#endif
#if !CONFIG_REALTIME_ONLY
if (is_stat_consumption_stage(cpi)) {
const size_t packet_sz = sizeof(FIRSTPASS_STATS);
const int packets = (int)(oxcf->twopass_stats_in.sz / packet_sz);
if (!cpi->ppi->lap_enabled) {
/*Re-initialize to stats buffer, populated by application in the case of
* two pass*/
cpi->ppi->twopass.stats_buf_ctx->stats_in_start =
oxcf->twopass_stats_in.buf;
cpi->twopass_frame.stats_in =
cpi->ppi->twopass.stats_buf_ctx->stats_in_start;
cpi->ppi->twopass.stats_buf_ctx->stats_in_end =
&cpi->ppi->twopass.stats_buf_ctx->stats_in_start[packets - 1];
// The buffer size is packets - 1 because the last packet is total_stats.
av1_firstpass_info_init(&cpi->ppi->twopass.firstpass_info,
oxcf->twopass_stats_in.buf, packets - 1);
av1_init_second_pass(cpi);
} else {
av1_firstpass_info_init(&cpi->ppi->twopass.firstpass_info, NULL, 0);
av1_init_single_pass_lap(cpi);
}
}
#endif
// The buffer "obmc_buffer" is used in inter frames for fast obmc search.
// Hence, the memory allocation for the same is avoided for allintra encoding
// mode.
if (cpi->oxcf.kf_cfg.key_freq_max != 0)
alloc_obmc_buffers(&cpi->td.mb.obmc_buffer, cm->error);
for (int x = 0; x < 2; x++)
for (int y = 0; y < 2; y++)
CHECK_MEM_ERROR(
cm, cpi->td.mb.intrabc_hash_info.hash_value_buffer[x][y],
(uint32_t *)aom_malloc(
AOM_BUFFER_SIZE_FOR_BLOCK_HASH *
sizeof(*cpi->td.mb.intrabc_hash_info.hash_value_buffer[0][0])));
cpi->td.mb.intrabc_hash_info.g_crc_initialized = 0;
av1_set_speed_features_framesize_independent(cpi, oxcf->speed);
av1_set_speed_features_framesize_dependent(cpi, oxcf->speed);
int max_mi_cols = mi_params->mi_cols;
int max_mi_rows = mi_params->mi_rows;
if (oxcf->frm_dim_cfg.forced_max_frame_width) {
max_mi_cols = size_in_mi(oxcf->frm_dim_cfg.forced_max_frame_width);
}
if (oxcf->frm_dim_cfg.forced_max_frame_height) {
max_mi_rows = size_in_mi(oxcf->frm_dim_cfg.forced_max_frame_height);
}
const int consec_zero_mv_alloc_size = (max_mi_rows * max_mi_cols) >> 2;
CHECK_MEM_ERROR(
cm, cpi->consec_zero_mv,
aom_calloc(consec_zero_mv_alloc_size, sizeof(*cpi->consec_zero_mv)));
cpi->consec_zero_mv_alloc_size = consec_zero_mv_alloc_size;
cpi->mb_weber_stats = NULL;
cpi->mb_delta_q = NULL;
cpi->palette_pixel_num = 0;
cpi->scaled_last_source_available = 0;
{
const BLOCK_SIZE bsize = BLOCK_16X16;
const int w = mi_size_wide[bsize];
const int h = mi_size_high[bsize];
const int num_cols = (max_mi_cols + w - 1) / w;
const int num_rows = (max_mi_rows + h - 1) / h;
CHECK_MEM_ERROR(cm, cpi->ssim_rdmult_scaling_factors,
aom_calloc(num_rows * num_cols,
sizeof(*cpi->ssim_rdmult_scaling_factors)));
CHECK_MEM_ERROR(cm, cpi->tpl_rdmult_scaling_factors,
aom_calloc(num_rows * num_cols,
sizeof(*cpi->tpl_rdmult_scaling_factors)));
}
#if CONFIG_TUNE_VMAF
{
const BLOCK_SIZE bsize = BLOCK_64X64;
const int w = mi_size_wide[bsize];
const int h = mi_size_high[bsize];
const int num_cols = (mi_params->mi_cols + w - 1) / w;
const int num_rows = (mi_params->mi_rows + h - 1) / h;
CHECK_MEM_ERROR(cm, cpi->vmaf_info.rdmult_scaling_factors,
aom_calloc(num_rows * num_cols,
sizeof(*cpi->vmaf_info.rdmult_scaling_factors)));
for (int i = 0; i < MAX_ARF_LAYERS; i++) {
cpi->vmaf_info.last_frame_unsharp_amount[i] = -1.0;
cpi->vmaf_info.last_frame_ysse[i] = -1.0;
cpi->vmaf_info.last_frame_vmaf[i] = -1.0;
}
cpi->vmaf_info.original_qindex = -1;
cpi->vmaf_info.vmaf_model = NULL;
}
#endif
#if CONFIG_TUNE_BUTTERAUGLI
{
const int w = mi_size_wide[butteraugli_rdo_bsize];
const int h = mi_size_high[butteraugli_rdo_bsize];
const int num_cols = (mi_params->mi_cols + w - 1) / w;
const int num_rows = (mi_params->mi_rows + h - 1) / h;
CHECK_MEM_ERROR(
cm, cpi->butteraugli_info.rdmult_scaling_factors,
aom_malloc(num_rows * num_cols *
sizeof(*cpi->butteraugli_info.rdmult_scaling_factors)));
memset(&cpi->butteraugli_info.source, 0,
sizeof(cpi->butteraugli_info.source));
memset(&cpi->butteraugli_info.resized_source, 0,
sizeof(cpi->butteraugli_info.resized_source));
cpi->butteraugli_info.recon_set = false;
}
#endif
#if CONFIG_SALIENCY_MAP
{
CHECK_MEM_ERROR(cm, cpi->saliency_map,
(uint8_t *)aom_calloc(cm->height * cm->width,
sizeof(*cpi->saliency_map)));
// Buffer initialization based on MIN_MIB_SIZE_LOG2 to ensure that
// cpi->sm_scaling_factor buffer is allocated big enough, since we have no
// idea of the actual superblock size we are going to use yet.
const int min_mi_w_sb = (1 << MIN_MIB_SIZE_LOG2);
const int min_mi_h_sb = (1 << MIN_MIB_SIZE_LOG2);
const int max_sb_cols =
(cm->mi_params.mi_cols + min_mi_w_sb - 1) / min_mi_w_sb;
const int max_sb_rows =
(cm->mi_params.mi_rows + min_mi_h_sb - 1) / min_mi_h_sb;
CHECK_MEM_ERROR(cm, cpi->sm_scaling_factor,
(double *)aom_calloc(max_sb_rows * max_sb_cols,
sizeof(*cpi->sm_scaling_factor)));
}
#endif
#if CONFIG_COLLECT_PARTITION_STATS
av1_zero(cpi->partition_stats);
#endif // CONFIG_COLLECT_PARTITION_STATS
// Initialize the members of DeltaQuantParams with INT_MAX to ensure that
// the quantizer tables are correctly initialized using the default deltaq
// parameters when av1_init_quantizer is called for the first time.
DeltaQuantParams *const prev_deltaq_params =
&cpi->enc_quant_dequant_params.prev_deltaq_params;
prev_deltaq_params->y_dc_delta_q = INT_MAX;
prev_deltaq_params->u_dc_delta_q = INT_MAX;
prev_deltaq_params->v_dc_delta_q = INT_MAX;
prev_deltaq_params->u_ac_delta_q = INT_MAX;
prev_deltaq_params->v_ac_delta_q = INT_MAX;
av1_init_quantizer(&cpi->enc_quant_dequant_params, &cm->quant_params,
cm->seq_params->bit_depth);
av1_qm_init(&cm->quant_params, av1_num_planes(cm));
av1_loop_filter_init(cm);
cm->superres_scale_denominator = SCALE_NUMERATOR;
cm->superres_upscaled_width = oxcf->frm_dim_cfg.width;
cm->superres_upscaled_height = oxcf->frm_dim_cfg.height;
#if !CONFIG_REALTIME_ONLY
av1_loop_restoration_precal();
#endif
cpi->third_pass_ctx = NULL;
if (cpi->oxcf.pass == AOM_RC_THIRD_PASS) {
av1_init_thirdpass_ctx(cm, &cpi->third_pass_ctx, NULL);
}
cpi->second_pass_log_stream = NULL;
cpi->use_ducky_encode = 0;
cm->error->setjmp = 0;
return cpi;
}
#if CONFIG_INTERNAL_STATS
#define SNPRINT(H, T) snprintf((H) + strlen(H), sizeof(H) - strlen(H), (T))
#define SNPRINT2(H, T, V) \
snprintf((H) + strlen(H), sizeof(H) - strlen(H), (T), (V))
#endif // CONFIG_INTERNAL_STATS
void av1_remove_primary_compressor(AV1_PRIMARY *ppi) {
if (!ppi) return;
#if !CONFIG_REALTIME_ONLY
av1_tf_info_free(&ppi->tf_info);
#endif // !CONFIG_REALTIME_ONLY
for (int i = 0; i < MAX_NUM_OPERATING_POINTS; ++i) {
aom_free(ppi->level_params.level_info[i]);
}
av1_lookahead_destroy(ppi->lookahead);
aom_free(ppi->tpl_sb_rdmult_scaling_factors);
ppi->tpl_sb_rdmult_scaling_factors = NULL;
TplParams *const tpl_data = &ppi->tpl_data;
aom_free(tpl_data->txfm_stats_list);
for (int frame = 0; frame < MAX_LAG_BUFFERS; ++frame) {
aom_free(tpl_data->tpl_stats_pool[frame]);
aom_free_frame_buffer(&tpl_data->tpl_rec_pool[frame]);
tpl_data->tpl_stats_pool[frame] = NULL;
}
#if !CONFIG_REALTIME_ONLY
av1_tpl_dealloc(&tpl_data->tpl_mt_sync);
#endif
av1_terminate_workers(ppi);
free_thread_data(ppi);
aom_free(ppi->p_mt_info.tile_thr_data);
ppi->p_mt_info.tile_thr_data = NULL;
aom_free(ppi->p_mt_info.workers);
ppi->p_mt_info.workers = NULL;
ppi->p_mt_info.num_workers = 0;
aom_free(ppi);
}
void av1_remove_compressor(AV1_COMP *cpi) {
if (!cpi) return;
#if CONFIG_RATECTRL_LOG
if (cpi->oxcf.pass == 3) {
rc_log_show(&cpi->rc_log);
}
#endif // CONFIG_RATECTRL_LOG
AV1_COMMON *cm = &cpi->common;
if (cm->current_frame.frame_number > 0) {
#if CONFIG_SPEED_STATS
if (!is_stat_generation_stage(cpi)) {
fprintf(stdout, "tx_search_count = %d\n", cpi->tx_search_count);
}
#endif // CONFIG_SPEED_STATS
#if CONFIG_COLLECT_PARTITION_STATS == 2
if (!is_stat_generation_stage(cpi)) {
av1_print_fr_partition_timing_stats(&cpi->partition_stats,
"fr_part_timing_data.csv");
}
#endif
}
#if CONFIG_AV1_TEMPORAL_DENOISING
av1_denoiser_free(&(cpi->denoiser));
#endif
if (cm->error) {
// Help detect use after free of the error detail string.
memset(cm->error->detail, 'A', sizeof(cm->error->detail) - 1);
cm->error->detail[sizeof(cm->error->detail) - 1] = '\0';
aom_free(cm->error);
}
aom_free(cpi->td.tctx);
MultiThreadInfo *const mt_info = &cpi->mt_info;
#if CONFIG_MULTITHREAD
pthread_mutex_t *const enc_row_mt_mutex_ = mt_info->enc_row_mt.mutex_;
pthread_cond_t *const enc_row_mt_cond_ = mt_info->enc_row_mt.cond_;
pthread_mutex_t *const gm_mt_mutex_ = mt_info->gm_sync.mutex_;
pthread_mutex_t *const tpl_error_mutex_ = mt_info->tpl_row_mt.mutex_;
pthread_mutex_t *const pack_bs_mt_mutex_ = mt_info->pack_bs_sync.mutex_;
if (enc_row_mt_mutex_ != NULL) {
pthread_mutex_destroy(enc_row_mt_mutex_);
aom_free(enc_row_mt_mutex_);
}
if (enc_row_mt_cond_ != NULL) {
pthread_cond_destroy(enc_row_mt_cond_);
aom_free(enc_row_mt_cond_);
}
if (gm_mt_mutex_ != NULL) {
pthread_mutex_destroy(gm_mt_mutex_);
aom_free(gm_mt_mutex_);
}
if (tpl_error_mutex_ != NULL) {
pthread_mutex_destroy(tpl_error_mutex_);
aom_free(tpl_error_mutex_);
}
if (pack_bs_mt_mutex_ != NULL) {
pthread_mutex_destroy(pack_bs_mt_mutex_);
aom_free(pack_bs_mt_mutex_);
}
#endif
av1_row_mt_mem_dealloc(cpi);
if (mt_info->num_workers > 1) {
av1_row_mt_sync_mem_dealloc(&cpi->ppi->intra_row_mt_sync);
av1_loop_filter_dealloc(&mt_info->lf_row_sync);
av1_cdef_mt_dealloc(&mt_info->cdef_sync);
#if !CONFIG_REALTIME_ONLY
av1_loop_restoration_dealloc(&mt_info->lr_row_sync);
av1_tf_mt_dealloc(&mt_info->tf_sync);
#endif
}
av1_free_thirdpass_ctx(cpi->third_pass_ctx);
av1_close_second_pass_log(cpi);
dealloc_compressor_data(cpi);
av1_ext_part_delete(&cpi->ext_part_controller);
av1_remove_common(cm);
aom_free(cpi);
#ifdef OUTPUT_YUV_REC
fclose(yuv_rec_file);
#endif
#ifdef OUTPUT_YUV_DENOISED
fclose(yuv_denoised_file);
#endif
}
static void generate_psnr_packet(AV1_COMP *cpi) {
struct aom_codec_cx_pkt pkt;
int i;
PSNR_STATS psnr;
#if CONFIG_AV1_HIGHBITDEPTH
const uint32_t in_bit_depth = cpi->oxcf.input_cfg.input_bit_depth;
const uint32_t bit_depth = cpi->td.mb.e_mbd.bd;
aom_calc_highbd_psnr(cpi->source, &cpi->common.cur_frame->buf, &psnr,
bit_depth, in_bit_depth);
#else
aom_calc_psnr(cpi->source, &cpi->common.cur_frame->buf, &psnr);
#endif
for (i = 0; i < 4; ++i) {
pkt.data.psnr.samples[i] = psnr.samples[i];
pkt.data.psnr.sse[i] = psnr.sse[i];
pkt.data.psnr.psnr[i] = psnr.psnr[i];
}
#if CONFIG_AV1_HIGHBITDEPTH
if ((cpi->source->flags & YV12_FLAG_HIGHBITDEPTH) &&
(in_bit_depth < bit_depth)) {
for (i = 0; i < 4; ++i) {
pkt.data.psnr.samples_hbd[i] = psnr.samples_hbd[i];
pkt.data.psnr.sse_hbd[i] = psnr.sse_hbd[i];
pkt.data.psnr.psnr_hbd[i] = psnr.psnr_hbd[i];
}
}
#endif
pkt.kind = AOM_CODEC_PSNR_PKT;
aom_codec_pkt_list_add(cpi->ppi->output_pkt_list, &pkt);
}
int av1_use_as_reference(int *ext_ref_frame_flags, int ref_frame_flags) {
if (ref_frame_flags > ((1 << INTER_REFS_PER_FRAME) - 1)) return -1;
*ext_ref_frame_flags = ref_frame_flags;
return 0;
}
int av1_copy_reference_enc(AV1_COMP *cpi, int idx, YV12_BUFFER_CONFIG *sd) {
AV1_COMMON *const cm = &cpi->common;
const int num_planes = av1_num_planes(cm);
YV12_BUFFER_CONFIG *cfg = get_ref_frame(cm, idx);
if (cfg) {
aom_yv12_copy_frame(cfg, sd, num_planes);
return 0;
} else {
return -1;
}
}
int av1_set_reference_enc(AV1_COMP *cpi, int idx, YV12_BUFFER_CONFIG *sd) {
AV1_COMMON *const cm = &cpi->common;
const int num_planes = av1_num_planes(cm);
YV12_BUFFER_CONFIG *cfg = get_ref_frame(cm, idx);
if (cfg) {
aom_yv12_copy_frame(sd, cfg, num_planes);
return 0;
} else {
return -1;
}
}
#ifdef OUTPUT_YUV_REC
void aom_write_one_yuv_frame(AV1_COMMON *cm, YV12_BUFFER_CONFIG *s) {
uint8_t *src = s->y_buffer;
int h = cm->height;
if (yuv_rec_file == NULL) return;
if (s->flags & YV12_FLAG_HIGHBITDEPTH) {
uint16_t *src16 = CONVERT_TO_SHORTPTR(s->y_buffer);
do {
fwrite(src16, s->y_width, 2, yuv_rec_file);
src16 += s->y_stride;
} while (--h);
src16 = CONVERT_TO_SHORTPTR(s->u_buffer);
h = s->uv_height;
do {
fwrite(src16, s->uv_width, 2, yuv_rec_file);
src16 += s->uv_stride;
} while (--h);
src16 = CONVERT_TO_SHORTPTR(s->v_buffer);
h = s->uv_height;
do {
fwrite(src16, s->uv_width, 2, yuv_rec_file);
src16 += s->uv_stride;
} while (--h);
fflush(yuv_rec_file);
return;
}
do {
fwrite(src, s->y_width, 1, yuv_rec_file);
src += s->y_stride;
} while (--h);
src = s->u_buffer;
h = s->uv_height;
do {
fwrite(src, s->uv_width, 1, yuv_rec_file);
src += s->uv_stride;
} while (--h);
src = s->v_buffer;
h = s->uv_height;
do {
fwrite(src, s->uv_width, 1, yuv_rec_file);
src += s->uv_stride;
} while (--h);
fflush(yuv_rec_file);
}
#endif // OUTPUT_YUV_REC
void av1_set_mv_search_params(AV1_COMP *cpi) {
const AV1_COMMON *const cm = &cpi->common;
MotionVectorSearchParams *const mv_search_params = &cpi->mv_search_params;
const int max_mv_def = AOMMAX(cm->width, cm->height);
// Default based on max resolution.
mv_search_params->mv_step_param = av1_init_search_range(max_mv_def);
if (cpi->sf.mv_sf.auto_mv_step_size) {
if (frame_is_intra_only(cm)) {
// Initialize max_mv_magnitude for use in the first INTER frame
// after a key/intra-only frame.
mv_search_params->max_mv_magnitude = max_mv_def;
} else {
// Use adaptive mv steps based on previous frame stats for show frames and
// internal arfs.
FRAME_UPDATE_TYPE cur_update_type =
cpi->ppi->gf_group.update_type[cpi->gf_frame_index];
int use_auto_mv_step =
(cm->show_frame || cur_update_type == INTNL_ARF_UPDATE) &&
mv_search_params->max_mv_magnitude != -1 &&
cpi->sf.mv_sf.auto_mv_step_size >= 2;
if (use_auto_mv_step) {
// Allow mv_steps to correspond to twice the max mv magnitude found
// in the previous frame, capped by the default max_mv_magnitude based
// on resolution.
mv_search_params->mv_step_param = av1_init_search_range(
AOMMIN(max_mv_def, 2 * mv_search_params->max_mv_magnitude));
}
// Reset max_mv_magnitude based on update flag.
if (cpi->do_frame_data_update) mv_search_params->max_mv_magnitude = -1;
}
}
}
void av1_set_screen_content_options(AV1_COMP *cpi, FeatureFlags *features) {
const AV1_COMMON *const cm = &cpi->common;
const MACROBLOCKD *const xd = &cpi->td.mb.e_mbd;
if (cm->seq_params->force_screen_content_tools != 2) {
features->allow_screen_content_tools = features->allow_intrabc =
cm->seq_params->force_screen_content_tools;
return;
}
if (cpi->oxcf.tune_cfg.content == AOM_CONTENT_SCREEN) {
features->allow_screen_content_tools = 1;
features->allow_intrabc = cpi->oxcf.mode == REALTIME ? 0 : 1;
cpi->is_screen_content_type = 1;
cpi->use_screen_content_tools = 1;
return;
}
if (cpi->oxcf.mode == REALTIME) {
features->allow_screen_content_tools = features->allow_intrabc = 0;
return;
}
// Screen content tools are not evaluated in non-RD encoding mode unless
// content type is not set explicitly, i.e., when
// cpi->oxcf.tune_cfg.content != AOM_CONTENT_SCREEN, use_nonrd_pick_mode = 1
// and hybrid_intra_pickmode = 0. Hence, screen content detection is
// disabled.
if (cpi->sf.rt_sf.use_nonrd_pick_mode &&
!cpi->sf.rt_sf.hybrid_intra_pickmode) {
features->allow_screen_content_tools = features->allow_intrabc = 0;
return;
}
// Estimate if the source frame is screen content, based on the portion of
// blocks that have few luma colors.
const uint8_t *src = cpi->unfiltered_source->y_buffer;
assert(src != NULL);
const int use_hbd = cpi->unfiltered_source->flags & YV12_FLAG_HIGHBITDEPTH;
const int stride = cpi->unfiltered_source->y_stride;
const int width = cpi->unfiltered_source->y_width;
const int height = cpi->unfiltered_source->y_height;
const int64_t area = (int64_t)width * height;
const int bd = cm->seq_params->bit_depth;
const int blk_w = 16;
const int blk_h = 16;
// These threshold values are selected experimentally.
const int color_thresh = 4;
const unsigned int var_thresh = 0;
// Counts of blocks with no more than color_thresh colors.
int64_t counts_1 = 0;
// Counts of blocks with no more than color_thresh colors and variance larger
// than var_thresh.
int64_t counts_2 = 0;
for (int r = 0; r + blk_h <= height; r += blk_h) {
for (int c = 0; c + blk_w <= width; c += blk_w) {
int count_buf[1 << 8]; // Maximum (1 << 8) bins for hbd path.
const uint8_t *const this_src = src + r * stride + c;
int n_colors;
if (use_hbd)
av1_count_colors_highbd(this_src, stride, blk_w, blk_h, bd, NULL,
count_buf, &n_colors, NULL);
else
av1_count_colors(this_src, stride, blk_w, blk_h, count_buf, &n_colors);
if (n_colors > 1 && n_colors <= color_thresh) {
++counts_1;
struct buf_2d buf;
buf.stride = stride;
buf.buf = (uint8_t *)this_src;
const unsigned int var = av1_get_perpixel_variance(
cpi, xd, &buf, BLOCK_16X16, AOM_PLANE_Y, use_hbd);
if (var > var_thresh) ++counts_2;
}
}
}
// The threshold values are selected experimentally.
features->allow_screen_content_tools = counts_1 * blk_h * blk_w * 10 > area;
// IntraBC would force loop filters off, so we use more strict rules that also
// requires that the block has high variance.
features->allow_intrabc = features->allow_screen_content_tools &&
counts_2 * blk_h * blk_w * 12 > area;
cpi->use_screen_content_tools = features->allow_screen_content_tools;
cpi->is_screen_content_type =
features->allow_intrabc || (counts_1 * blk_h * blk_w * 10 > area * 4 &&
counts_2 * blk_h * blk_w * 30 > area);
}
static void init_motion_estimation(AV1_COMP *cpi) {
AV1_COMMON *const cm = &cpi->common;
MotionVectorSearchParams *const mv_search_params = &cpi->mv_search_params;
const int aligned_width = (cm->width + 7) & ~7;
const int y_stride =
aom_calc_y_stride(aligned_width, cpi->oxcf.border_in_pixels);
const int y_stride_src = ((cpi->oxcf.frm_dim_cfg.width != cm->width ||
cpi->oxcf.frm_dim_cfg.height != cm->height) ||
av1_superres_scaled(cm))
? y_stride
: cpi->ppi->lookahead->buf->img.y_stride;
int fpf_y_stride =
cm->cur_frame != NULL ? cm->cur_frame->buf.y_stride : y_stride;
// Update if search_site_cfg is uninitialized or the current frame has a new
// stride
const int should_update =
!mv_search_params->search_site_cfg[SS_CFG_SRC][DIAMOND].stride ||
!mv_search_params->search_site_cfg[SS_CFG_LOOKAHEAD][DIAMOND].stride ||
(y_stride !=
mv_search_params->search_site_cfg[SS_CFG_SRC][DIAMOND].stride);
if (!should_update) {
return;
}
// Initialization of search_site_cfg for NUM_DISTINCT_SEARCH_METHODS.
for (SEARCH_METHODS i = DIAMOND; i < NUM_DISTINCT_SEARCH_METHODS; i++) {
const int level = ((i == NSTEP_8PT) || (i == CLAMPED_DIAMOND)) ? 1 : 0;
av1_init_motion_compensation[i](
&mv_search_params->search_site_cfg[SS_CFG_SRC][i], y_stride, level);
av1_init_motion_compensation[i](
&mv_search_params->search_site_cfg[SS_CFG_LOOKAHEAD][i], y_stride_src,
level);
}
// First pass search site config initialization.
av1_init_motion_fpf(&mv_search_params->search_site_cfg[SS_CFG_FPF][DIAMOND],
fpf_y_stride);
for (SEARCH_METHODS i = NSTEP; i < NUM_DISTINCT_SEARCH_METHODS; i++) {
memcpy(&mv_search_params->search_site_cfg[SS_CFG_FPF][i],
&mv_search_params->search_site_cfg[SS_CFG_FPF][DIAMOND],
sizeof(search_site_config));
}
}
static void init_ref_frame_bufs(AV1_COMP *cpi) {
AV1_COMMON *const cm = &cpi->common;
int i;
if (cm->cur_frame) {
cm->cur_frame->ref_count--;
cm->cur_frame = NULL;
}
for (i = 0; i < REF_FRAMES; ++i) {
if (cm->ref_frame_map[i]) {
cm->ref_frame_map[i]->ref_count--;
cm->ref_frame_map[i] = NULL;
}
}
#ifndef NDEBUG
BufferPool *const pool = cm->buffer_pool;
for (i = 0; i < pool->num_frame_bufs; ++i) {
assert(pool->frame_bufs[i].ref_count == 0);
}
#endif
}
// TODO(chengchen): consider renaming this function as it is necessary
// for the encoder to setup critical parameters, and it does not
// deal with initial width any longer.
aom_codec_err_t av1_check_initial_width(AV1_COMP *cpi, int use_highbitdepth,
int subsampling_x, int subsampling_y) {
AV1_COMMON *const cm = &cpi->common;
SequenceHeader *const seq_params = cm->seq_params;
if (!cpi->frame_size_related_setup_done ||
seq_params->use_highbitdepth != use_highbitdepth ||
seq_params->subsampling_x != subsampling_x ||
seq_params->subsampling_y != subsampling_y) {
seq_params->subsampling_x = subsampling_x;
seq_params->subsampling_y = subsampling_y;
seq_params->use_highbitdepth = use_highbitdepth;
av1_set_speed_features_framesize_independent(cpi, cpi->oxcf.speed);
av1_set_speed_features_framesize_dependent(cpi, cpi->oxcf.speed);
if (!is_stat_generation_stage(cpi)) {
#if !CONFIG_REALTIME_ONLY
if (!av1_tf_info_alloc(&cpi->ppi->tf_info, cpi))
return AOM_CODEC_MEM_ERROR;
#endif // !CONFIG_REALTIME_ONLY
}
init_ref_frame_bufs(cpi);
init_motion_estimation(cpi); // TODO(agrange) This can be removed.
cpi->initial_mbs = cm->mi_params.MBs;
cpi->frame_size_related_setup_done = true;
}
return AOM_CODEC_OK;
}
#if CONFIG_AV1_TEMPORAL_DENOISING
static void setup_denoiser_buffer(AV1_COMP *cpi) {
AV1_COMMON *const cm = &cpi->common;
if (cpi->oxcf.noise_sensitivity > 0 &&
!cpi->denoiser.frame_buffer_initialized) {
if (av1_denoiser_alloc(
cm, &cpi->svc, &cpi->denoiser, cpi->ppi->use_svc,
cpi->oxcf.noise_sensitivity, cm->width, cm->height,
cm->seq_params->subsampling_x, cm->seq_params->subsampling_y,
cm->seq_params->use_highbitdepth, AOM_BORDER_IN_PIXELS))
aom_internal_error(cm->error, AOM_CODEC_MEM_ERROR,
"Failed to allocate denoiser");
}
}
#endif
// Returns 1 if the assigned width or height was <= 0.
static int set_size_literal(AV1_COMP *cpi, int width, int height) {
AV1_COMMON *cm = &cpi->common;
aom_codec_err_t err = av1_check_initial_width(
cpi, cm->seq_params->use_highbitdepth, cm->seq_params->subsampling_x,
cm->seq_params->subsampling_y);
if (err != AOM_CODEC_OK) {
aom_internal_error(cm->error, err, "av1_check_initial_width() failed");
}
if (width <= 0 || height <= 0) return 1;
cm->width = width;
cm->height = height;
#if CONFIG_AV1_TEMPORAL_DENOISING
setup_denoiser_buffer(cpi);
#endif
if (cm->width > cpi->data_alloc_width ||
cm->height > cpi->data_alloc_height) {
av1_free_context_buffers(cm);
av1_free_shared_coeff_buffer(&cpi->td.shared_coeff_buf);
av1_free_sms_tree(&cpi->td);
av1_free_pmc(cpi->td.firstpass_ctx, av1_num_planes(cm));
cpi->td.firstpass_ctx = NULL;
alloc_compressor_data(cpi);
realloc_segmentation_maps(cpi);
cpi->data_alloc_width = cm->width;
cpi->data_alloc_height = cm->height;
cpi->frame_size_related_setup_done = false;
}
alloc_mb_mode_info_buffers(cpi);
av1_update_frame_size(cpi);
return 0;
}
void av1_set_frame_size(AV1_COMP *cpi, int width, int height) {
AV1_COMMON *const cm = &cpi->common;
const SequenceHeader *const seq_params = cm->seq_params;
const int num_planes = av1_num_planes(cm);
MACROBLOCKD *const xd = &cpi->td.mb.e_mbd;
int ref_frame;
if (width != cm->width || height != cm->height) {
// There has been a change in the encoded frame size
set_size_literal(cpi, width, height);
// Recalculate 'all_lossless' in case super-resolution was (un)selected.
cm->features.all_lossless =
cm->features.coded_lossless && !av1_superres_scaled(cm);
av1_noise_estimate_init(&cpi->noise_estimate, cm->width, cm->height);
#if CONFIG_AV1_TEMPORAL_DENOISING
// Reset the denoiser on the resized frame.
if (cpi->oxcf.noise_sensitivity > 0) {
av1_denoiser_free(&(cpi->denoiser));
setup_denoiser_buffer(cpi);
}
#endif
}
if (is_stat_consumption_stage(cpi)) {
av1_set_target_rate(cpi, cm->width, cm->height);
}
alloc_frame_mvs(cm, cm->cur_frame);
// Allocate above context buffers
CommonContexts *const above_contexts = &cm->above_contexts;
if (above_contexts->num_planes < av1_num_planes(cm) ||
above_contexts->num_mi_cols < cm->mi_params.mi_cols ||
above_contexts->num_tile_rows < cm->tiles.rows) {
av1_free_above_context_buffers(above_contexts);
if (av1_alloc_above_context_buffers(above_contexts, cm->tiles.rows,
cm->mi_params.mi_cols,
av1_num_planes(cm)))
aom_internal_error(cm->error, AOM_CODEC_MEM_ERROR,
"Failed to allocate context buffers");
}
AV1EncoderConfig *oxcf = &cpi->oxcf;
oxcf->border_in_pixels = av1_get_enc_border_size(
av1_is_resize_needed(oxcf), oxcf->kf_cfg.key_freq_max == 0,
cm->seq_params->sb_size);
// Reset the frame pointers to the current frame size.
if (aom_realloc_frame_buffer(
&cm->cur_frame->buf, cm->width, cm->height, seq_params->subsampling_x,
seq_params->subsampling_y, seq_params->use_highbitdepth,
cpi->oxcf.border_in_pixels, cm->features.byte_alignment, NULL, NULL,
NULL, cpi->alloc_pyramid, 0))
aom_internal_error(cm->error, AOM_CODEC_MEM_ERROR,
"Failed to allocate frame buffer");
if (!is_stat_generation_stage(cpi)) av1_init_cdef_worker(cpi);
#if !CONFIG_REALTIME_ONLY
if (is_restoration_used(cm)) {
for (int i = 0; i < num_planes; ++i)
cm->rst_info[i].frame_restoration_type = RESTORE_NONE;
const bool is_sgr_enabled = !cpi->sf.lpf_sf.disable_sgr_filter;
av1_alloc_restoration_buffers(cm, is_sgr_enabled);
// Store the allocated restoration buffers in MT object.
if (cpi->ppi->p_mt_info.num_workers > 1) {
av1_init_lr_mt_buffers(cpi);
}
}
#endif
init_motion_estimation(cpi);
int has_valid_ref_frame = 0;
for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
RefCntBuffer *const buf = get_ref_frame_buf(cm, ref_frame);
if (buf != NULL) {
struct scale_factors *sf = get_ref_scale_factors(cm, ref_frame);
av1_setup_scale_factors_for_frame(sf, buf->buf.y_crop_width,
buf->buf.y_crop_height, cm->width,
cm->height);
has_valid_ref_frame |= av1_is_valid_scale(sf);
if (av1_is_scaled(sf)) aom_extend_frame_borders(&buf->buf, num_planes);
}
}
if (!frame_is_intra_only(cm) && !has_valid_ref_frame) {
aom_internal_error(
cm->error, AOM_CODEC_CORRUPT_FRAME,
"Can't find at least one reference frame with valid size");
}
av1_setup_scale_factors_for_frame(&cm->sf_identity, cm->width, cm->height,
cm->width, cm->height);
set_ref_ptrs(cm, xd, LAST_FRAME, LAST_FRAME);
}
static inline int extend_borders_mt(const AV1_COMP *cpi,
MULTI_THREADED_MODULES stage, int plane) {
const AV1_COMMON *const cm = &cpi->common;
if (cpi->mt_info.num_mod_workers[stage] < 2) return 0;
switch (stage) {
// TODO(deepa.kg@ittiam.com): When cdef and loop-restoration are disabled,
// multi-thread frame border extension along with loop filter frame.
// As loop-filtering of a superblock row modifies the pixels of the
// above superblock row, border extension requires that loop filtering
// of the current and above superblock row is complete.
case MOD_LPF: return 0;
case MOD_CDEF:
return is_cdef_used(cm) && !cpi->ppi->rtc_ref.non_reference_frame &&
!is_restoration_used(cm) && !av1_superres_scaled(cm);
case MOD_LR:
return is_restoration_used(cm) &&
(cm->rst_info[plane].frame_restoration_type != RESTORE_NONE);
default: assert(0);
}
return 0;
}
/*!\brief Select and apply cdef filters and switchable restoration filters
*
* \ingroup high_level_algo
*/
static void cdef_restoration_frame(AV1_COMP *cpi, AV1_COMMON *cm,
MACROBLOCKD *xd, int use_restoration,
int use_cdef,
unsigned int skip_apply_postproc_filters) {
#if !CONFIG_REALTIME_ONLY
if (use_restoration)
av1_loop_restoration_save_boundary_lines(&cm->cur_frame->buf, cm, 0);
#else
(void)use_restoration;
#endif
if (use_cdef) {
#if CONFIG_COLLECT_COMPONENT_TIMING
start_timing(cpi, cdef_time);
#endif
const int num_workers = cpi->mt_info.num_mod_workers[MOD_CDEF];
// Find CDEF parameters
av1_cdef_search(cpi);
// Apply the filter
if ((skip_apply_postproc_filters & SKIP_APPLY_CDEF) == 0) {
assert(!cpi->ppi->rtc_ref.non_reference_frame);
if (num_workers > 1) {
// Extension of frame borders is multi-threaded along with cdef.
const int do_extend_border =
extend_borders_mt(cpi, MOD_CDEF, /* plane */ 0);
av1_cdef_frame_mt(cm, xd, cpi->mt_info.cdef_worker,
cpi->mt_info.workers, &cpi->mt_info.cdef_sync,
num_workers, av1_cdef_init_fb_row_mt,
do_extend_border);
} else {
av1_cdef_frame(&cm->cur_frame->buf, cm, xd, av1_cdef_init_fb_row);
}
}
#if CONFIG_COLLECT_COMPONENT_TIMING
end_timing(cpi, cdef_time);
#endif
}
const int use_superres = av1_superres_scaled(cm);
if (use_superres) {
if ((skip_apply_postproc_filters & SKIP_APPLY_SUPERRES) == 0) {
av1_superres_post_encode(cpi);
}
}
#if !CONFIG_REALTIME_ONLY
#if CONFIG_COLLECT_COMPONENT_TIMING
start_timing(cpi, loop_restoration_time);
#endif
if (use_restoration) {
MultiThreadInfo *const mt_info = &cpi->mt_info;
const int num_workers = mt_info->num_mod_workers[MOD_LR];
av1_loop_restoration_save_boundary_lines(&cm->cur_frame->buf, cm, 1);
av1_pick_filter_restoration(cpi->source, cpi);
if ((skip_apply_postproc_filters & SKIP_APPLY_RESTORATION) == 0 &&
(cm->rst_info[0].frame_restoration_type != RESTORE_NONE ||
cm->rst_info[1].frame_restoration_type != RESTORE_NONE ||
cm->rst_info[2].frame_restoration_type != RESTORE_NONE)) {
if (num_workers > 1) {
// Extension of frame borders is multi-threaded along with loop
// restoration filter.
const int do_extend_border = 1;
av1_loop_restoration_filter_frame_mt(
&cm->cur_frame->buf, cm, 0, mt_info->workers, num_workers,
&mt_info->lr_row_sync, &cpi->lr_ctxt, do_extend_border);
} else {
av1_loop_restoration_filter_frame(&cm->cur_frame->buf, cm, 0,
&cpi->lr_ctxt);
}
}
}
#if CONFIG_COLLECT_COMPONENT_TIMING
end_timing(cpi, loop_restoration_time);
#endif
#endif // !CONFIG_REALTIME_ONLY
}
static void extend_frame_borders(AV1_COMP *cpi) {
const AV1_COMMON *const cm = &cpi->common;
// TODO(debargha): Fix mv search range on encoder side
for (int plane = 0; plane < av1_num_planes(cm); ++plane) {
const bool extend_border_done = extend_borders_mt(cpi, MOD_CDEF, plane) ||
extend_borders_mt(cpi, MOD_LR, plane);
if (!extend_border_done) {
const YV12_BUFFER_CONFIG *const ybf = &cm->cur_frame->buf;
aom_extend_frame_borders_plane_row(ybf, plane, 0,
ybf->crop_heights[plane > 0]);
}
}
}
/*!\brief Select and apply deblocking filters, cdef filters, and restoration
* filters.
*
* \ingroup high_level_algo
*/
static void loopfilter_frame(AV1_COMP *cpi, AV1_COMMON *cm) {
MultiThreadInfo *const mt_info = &cpi->mt_info;
const int num_workers = mt_info->num_mod_workers[MOD_LPF];
const int num_planes = av1_num_planes(cm);
MACROBLOCKD *xd = &cpi->td.mb.e_mbd;
cpi->td.mb.rdmult = cpi->rd.RDMULT;
assert(IMPLIES(is_lossless_requested(&cpi->oxcf.rc_cfg),
cm->features.coded_lossless && cm->features.all_lossless));
const int use_loopfilter =
is_loopfilter_used(cm) && !cpi->mt_info.pipeline_lpf_mt_with_enc;
const int use_cdef =
is_cdef_used(cm) && (!cpi->active_map.enabled ||
cpi->rc.percent_blocks_inactive <=
cpi->sf.rt_sf.thresh_active_maps_skip_lf_cdef);
const int use_superres = av1_superres_scaled(cm);
const int use_restoration = is_restoration_used(cm);
const unsigned int skip_apply_postproc_filters =
derive_skip_apply_postproc_filters(cpi, use_loopfilter, use_cdef,
use_superres, use_restoration);
#if CONFIG_COLLECT_COMPONENT_TIMING
start_timing(cpi, loop_filter_time);
#endif
if (use_loopfilter) {
av1_pick_filter_level(cpi->source, cpi, cpi->sf.lpf_sf.lpf_pick);
struct loopfilter *lf = &cm->lf;
if ((lf->filter_level[0] || lf->filter_level[1]) &&
(skip_apply_postproc_filters & SKIP_APPLY_LOOPFILTER) == 0) {
assert(!cpi->ppi->rtc_ref.non_reference_frame);
// lpf_opt_level = 1 : Enables dual/quad loop-filtering.
// lpf_opt_level is set to 1 if transform size search depth in inter
// blocks is limited to one as quad loop filtering assumes that all the
// transform blocks within a 16x8/8x16/16x16 prediction block are of the
// same size. lpf_opt_level = 2 : Filters both chroma planes together, in
// addition to enabling dual/quad loop-filtering. This is enabled when lpf
// pick method is LPF_PICK_FROM_Q as u and v plane filter levels are
// equal.
int lpf_opt_level = get_lpf_opt_level(&cpi->sf);
av1_loop_filter_frame_mt(&cm->cur_frame->buf, cm, xd, 0, num_planes, 0,
mt_info->workers, num_workers,
&mt_info->lf_row_sync, lpf_opt_level);
}
}
#if CONFIG_COLLECT_COMPONENT_TIMING
end_timing(cpi, loop_filter_time);
#endif
cdef_restoration_frame(cpi, cm, xd, use_restoration, use_cdef,
skip_apply_postproc_filters);
}
static void update_motion_stat(AV1_COMP *const cpi) {
AV1_COMMON *const cm = &cpi->common;
const CommonModeInfoParams *const mi_params = &cm->mi_params;
RATE_CONTROL *const rc = &cpi->rc;
SVC *const svc = &cpi->svc;
const int avg_cnt_zeromv =
100 * cpi->rc.cnt_zeromv / (mi_params->mi_rows * mi_params->mi_cols);
if (!cpi->ppi->use_svc ||
(cpi->ppi->use_svc &&
!cpi->svc.layer_context[cpi->svc.temporal_layer_id].is_key_frame &&
cpi->svc.spatial_layer_id == cpi->svc.number_spatial_layers - 1)) {
rc->avg_frame_low_motion =
(rc->avg_frame_low_motion == 0)
? avg_cnt_zeromv
: (3 * rc->avg_frame_low_motion + avg_cnt_zeromv) / 4;
// For SVC: set avg_frame_low_motion (only computed on top spatial layer)
// to all lower spatial layers.
if (cpi->ppi->use_svc &&
svc->spatial_layer_id == svc->number_spatial_layers - 1) {
for (int i = 0; i < svc->number_spatial_layers - 1; ++i) {
const int layer = LAYER_IDS_TO_IDX(i, svc->temporal_layer_id,
svc->number_temporal_layers);
LAYER_CONTEXT *const lc = &svc->layer_context[layer];
RATE_CONTROL *const lrc = &lc->rc;
lrc->avg_frame_low_motion = rc->avg_frame_low_motion;
}
}
}
}
/*!\brief Encode a frame without the recode loop, usually used in one-pass
* encoding and realtime coding.
*
* \ingroup high_level_algo
*
* \param[in] cpi Top-level encoder structure
*
* \return Returns a value to indicate if the encoding is done successfully.
* \retval #AOM_CODEC_OK
* \retval #AOM_CODEC_ERROR
*/
static int encode_without_recode(AV1_COMP *cpi) {
AV1_COMMON *const cm = &cpi->common;
const QuantizationCfg *const q_cfg = &cpi->oxcf.q_cfg;
SVC *const svc = &cpi->svc;
const int resize_pending = is_frame_resize_pending(cpi);
int top_index = 0, bottom_index = 0, q = 0;
YV12_BUFFER_CONFIG *unscaled = cpi->unscaled_source;
InterpFilter filter_scaler =
cpi->ppi->use_svc ? svc->downsample_filter_type[svc->spatial_layer_id]
: EIGHTTAP_SMOOTH;
int phase_scaler = cpi->ppi->use_svc
? svc->downsample_filter_phase[svc->spatial_layer_id]
: 0;
if (cpi->rc.postencode_drop && allow_postencode_drop_rtc(cpi))
av1_save_all_coding_context(cpi);
set_size_independent_vars(cpi);
av1_setup_frame_size(cpi);
cm->prev_frame = get_primary_ref_frame_buf(cm);
av1_set_size_dependent_vars(cpi, &q, &bottom_index, &top_index);
av1_set_mv_search_params(cpi);
if (cm->current_frame.frame_number == 0 &&
(cpi->ppi->use_svc || cpi->oxcf.rc_cfg.drop_frames_water_mark > 0) &&
cpi->svc.temporal_layer_id == 0) {
const SequenceHeader *seq_params = cm->seq_params;
if (aom_alloc_frame_buffer(
&cpi->svc.source_last_TL0, cpi->oxcf.frm_dim_cfg.width,
cpi->oxcf.frm_dim_cfg.height, seq_params->subsampling_x,
seq_params->subsampling_y, seq_params->use_highbitdepth,
cpi->oxcf.border_in_pixels, cm->features.byte_alignment, false,
0)) {
aom_internal_error(cm->error, AOM_CODEC_MEM_ERROR,
"Failed to allocate buffer for source_last_TL0");
}
}
if (!cpi->ppi->use_svc) {
phase_scaler = 8;
// 2:1 scaling.
if ((cm->width << 1) == unscaled->y_crop_width &&
(cm->height << 1) == unscaled->y_crop_height) {
filter_scaler = BILINEAR;
// For lower resolutions use eighttap_smooth.
if (cm->width * cm->height <= 320 * 180) filter_scaler = EIGHTTAP_SMOOTH;
} else if ((cm->width << 2) == unscaled->y_crop_width &&
(cm->height << 2) == unscaled->y_crop_height) {
// 4:1 scaling.
filter_scaler = EIGHTTAP_SMOOTH;
} else if ((cm->width << 2) == 3 * unscaled->y_crop_width &&
(cm->height << 2) == 3 * unscaled->y_crop_height) {
// 4:3 scaling.
filter_scaler = EIGHTTAP_REGULAR;
}
}
allocate_gradient_info_for_hog(cpi);
allocate_src_var_of_4x4_sub_block_buf(cpi);
const SPEED_FEATURES *sf = &cpi->sf;
if (sf->part_sf.partition_search_type == VAR_BASED_PARTITION)
variance_partition_alloc(cpi);
if (cm->current_frame.frame_type == KEY_FRAME ||
((sf->inter_sf.extra_prune_warped && cpi->refresh_frame.golden_frame)))
copy_frame_prob_info(cpi);
#if CONFIG_COLLECT_COMPONENT_TIMING
printf("\n Encoding a frame: \n");
#endif
#if CONFIG_TUNE_BUTTERAUGLI
if (cpi->oxcf.tune_cfg.tuning == AOM_TUNE_BUTTERAUGLI) {
av1_setup_butteraugli_rdmult(cpi);
}
#endif
cpi->source = av1_realloc_and_scale_if_required(
cm, unscaled, &cpi->scaled_source, filter_scaler, phase_scaler, true,
false, cpi->oxcf.border_in_pixels, cpi->alloc_pyramid);
if (frame_is_intra_only(cm) || resize_pending != 0) {
const int current_size =
(cm->mi_params.mi_rows * cm->mi_params.mi_cols) >> 2;
if (cpi->consec_zero_mv &&
(cpi->consec_zero_mv_alloc_size < current_size)) {
aom_free(cpi->consec_zero_mv);
cpi->consec_zero_mv_alloc_size = 0;
CHECK_MEM_ERROR(cm, cpi->consec_zero_mv,
aom_malloc(current_size * sizeof(*cpi->consec_zero_mv)));
cpi->consec_zero_mv_alloc_size = current_size;
}
assert(cpi->consec_zero_mv != NULL);
memset(cpi->consec_zero_mv, 0, current_size * sizeof(*cpi->consec_zero_mv));
}
if (cpi->scaled_last_source_available) {
cpi->last_source = &cpi->scaled_last_source;
cpi->scaled_last_source_available = 0;
} else if (cpi->unscaled_last_source != NULL) {
cpi->last_source = av1_realloc_and_scale_if_required(
cm, cpi->unscaled_last_source, &cpi->scaled_last_source, filter_scaler,
phase_scaler, true, false, cpi->oxcf.border_in_pixels,
cpi->alloc_pyramid);
}
if (cpi->sf.rt_sf.use_temporal_noise_estimate) {
av1_update_noise_estimate(cpi);
}
#if CONFIG_AV1_TEMPORAL_DENOISING
if (cpi->oxcf.noise_sensitivity > 0 && cpi->ppi->use_svc)
av1_denoiser_reset_on_first_frame(cpi);
#endif
// For 1 spatial layer encoding: if the (non-LAST) reference has different
// resolution from the source then disable that reference. This is to avoid
// significant increase in encode time from scaling the references in
// av1_scale_references. Note GOLDEN is forced to update on the (first/tigger)
// resized frame and ALTREF will be refreshed ~4 frames later, so both
// references become available again after few frames.
// For superres: don't disable golden reference.
if (svc->number_spatial_layers == 1) {
if (!cpi->oxcf.superres_cfg.enable_superres) {
if (cpi->ref_frame_flags & av1_ref_frame_flag_list[GOLDEN_FRAME]) {
const YV12_BUFFER_CONFIG *const ref =
get_ref_frame_yv12_buf(cm, GOLDEN_FRAME);
if (ref == NULL || ref->y_crop_width != cm->width ||
ref->y_crop_height != cm->height) {
cpi->ref_frame_flags ^= AOM_GOLD_FLAG;
}
}
}
if (cpi->ref_frame_flags & av1_ref_frame_flag_list[ALTREF_FRAME]) {
const YV12_BUFFER_CONFIG *const ref =
get_ref_frame_yv12_buf(cm, ALTREF_FRAME);
if (ref == NULL || ref->y_crop_width != cm->width ||
ref->y_crop_height != cm->height) {
cpi->ref_frame_flags ^= AOM_ALT_FLAG;
}
}
}
int scale_references = 0;
#if CONFIG_FPMT_TEST
scale_references =
cpi->ppi->fpmt_unit_test_cfg == PARALLEL_SIMULATION_ENCODE ? 1 : 0;
#endif // CONFIG_FPMT_TEST
if (scale_references ||
cpi->ppi->gf_group.frame_parallel_level[cpi->gf_frame_index] == 0) {
if (!frame_is_intra_only(cm)) {
av1_scale_references(cpi, filter_scaler, phase_scaler, 1);
}
}
av1_set_quantizer(cm, q_cfg->qm_minlevel, q_cfg->qm_maxlevel, q,
q_cfg->enable_chroma_deltaq, q_cfg->enable_hdr_deltaq);
av1_set_speed_features_qindex_dependent(cpi, cpi->oxcf.speed);
av1_init_quantizer(&cpi->enc_quant_dequant_params, &cm->quant_params,
cm->seq_params->bit_depth);
av1_set_variance_partition_thresholds(cpi, q, 0);
av1_setup_frame(cpi);
// Check if this high_source_sad (scene/slide change) frame should be
// encoded at high/max QP, and if so, set the q and adjust some rate
// control parameters.
if (cpi->sf.rt_sf.overshoot_detection_cbr == FAST_DETECTION_MAXQ &&
cpi->rc.high_source_sad) {
if (av1_encodedframe_overshoot_cbr(cpi, &q)) {
av1_set_quantizer(cm, q_cfg->qm_minlevel, q_cfg->qm_maxlevel, q,
q_cfg->enable_chroma_deltaq, q_cfg->enable_hdr_deltaq);
av1_set_speed_features_qindex_dependent(cpi, cpi->oxcf.speed);
av1_init_quantizer(&cpi->enc_quant_dequant_params, &cm->quant_params,
cm->seq_params->bit_depth);
av1_set_variance_partition_thresholds(cpi, q, 0);