/*
* Copyright (c) 2019, 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.
*/
/*!\defgroup gf_group_algo Golden Frame Group
* \ingroup high_level_algo
* Algorithms regarding determining the length of GF groups and defining GF
* group structures.
* @{
*/
/*! @} - end defgroup gf_group_algo */
#include <assert.h>
#include <limits.h>
#include <stdint.h>
#include "aom_dsp/aom_dsp_common.h"
#include "aom_mem/aom_mem.h"
#include "config/aom_config.h"
#include "config/aom_scale_rtcd.h"
#include "aom/aom_codec.h"
#include "aom/aom_encoder.h"
#include "av1/common/av1_common_int.h"
#include "av1/encoder/encoder.h"
#include "av1/encoder/firstpass.h"
#include "av1/encoder/gop_structure.h"
#include "av1/encoder/pass2_strategy.h"
#include "av1/encoder/ratectrl.h"
#include "av1/encoder/rc_utils.h"
#include "av1/encoder/temporal_filter.h"
#if CONFIG_THREE_PASS
#include "av1/encoder/thirdpass.h"
#endif
#include "av1/encoder/tpl_model.h"
#include "av1/encoder/encode_strategy.h"
#define DEFAULT_KF_BOOST 2300
#define DEFAULT_GF_BOOST 2000
#define GROUP_ADAPTIVE_MAXQ 1
static void init_gf_stats(GF_GROUP_STATS *gf_stats);
#if CONFIG_THREE_PASS
static int define_gf_group_pass3(AV1_COMP *cpi, EncodeFrameParams *frame_params,
int is_final_pass);
#endif
// Calculate an active area of the image that discounts formatting
// bars and partially discounts other 0 energy areas.
#define MIN_ACTIVE_AREA 0.5
#define MAX_ACTIVE_AREA 1.0
static double calculate_active_area(
const FRAME_INFO *frame_info,
const FIRSTPASS_STATS *this_frame) {
const double active_pct =
1.0 -
((this_frame->intra_skip_pct / 2) +
((this_frame->inactive_zone_rows * 2) / (
double)frame_info->mb_rows));
return fclamp(active_pct, MIN_ACTIVE_AREA, MAX_ACTIVE_AREA);
}
// Calculate a modified Error used in distributing bits between easier and
// harder frames.
#define ACT_AREA_CORRECTION 0.5
static double calculate_modified_err_new(
const FRAME_INFO *frame_info,
const FIRSTPASS_STATS *total_stats,
const FIRSTPASS_STATS *this_stats,
int vbrbias,
double modified_error_min,
double modified_error_max) {
if (total_stats == NULL) {
return 0;
}
const double av_weight = total_stats->weight / total_stats->count;
const double av_err =
(total_stats->coded_error * av_weight) / total_stats->count;
double modified_error =
av_err * pow(this_stats->coded_error * this_stats->weight /
DOUBLE_DIVIDE_CHECK(av_err),
vbrbias / 100.0);
// Correction for active area. Frames with a reduced active area
// (eg due to formatting bars) have a higher error per mb for the
// remaining active MBs. The correction here assumes that coding
// 0.5N blocks of complexity 2X is a little easier than coding N
// blocks of complexity X.
modified_error *=
pow(calculate_active_area(frame_info, this_stats), ACT_AREA_CORRECTION);
return fclamp(modified_error, modified_error_min, modified_error_max);
}
static double calculate_modified_err(
const FRAME_INFO *frame_info,
const TWO_PASS *twopass,
const AV1EncoderConfig *oxcf,
const FIRSTPASS_STATS *this_frame) {
const FIRSTPASS_STATS *total_stats = twopass->stats_buf_ctx->total_stats;
return calculate_modified_err_new(
frame_info, total_stats, this_frame, oxcf->rc_cfg.vbrbias,
twopass->modified_error_min, twopass->modified_error_max);
}
// Resets the first pass file to the given position using a relative seek from
// the current position.
static void reset_fpf_position(TWO_PASS_FRAME *p_frame,
const FIRSTPASS_STATS *position) {
p_frame->stats_in = position;
}
static int input_stats(TWO_PASS *p, TWO_PASS_FRAME *p_frame,
FIRSTPASS_STATS *fps) {
if (p_frame->stats_in >= p->stats_buf_ctx->stats_in_end)
return EOF;
*fps = *p_frame->stats_in;
++p_frame->stats_in;
return 1;
}
static int input_stats_lap(TWO_PASS *p, TWO_PASS_FRAME *p_frame,
FIRSTPASS_STATS *fps) {
if (p_frame->stats_in >= p->stats_buf_ctx->stats_in_end)
return EOF;
*fps = *p_frame->stats_in;
/* Move old stats[0] out to accommodate for next frame stats */
memmove(p->frame_stats_arr[0], p->frame_stats_arr[1],
(p->stats_buf_ctx->stats_in_end - p_frame->stats_in - 1) *
sizeof(FIRSTPASS_STATS));
p->stats_buf_ctx->stats_in_end--;
return 1;
}
// Read frame stats at an offset from the current position.
static const FIRSTPASS_STATS *read_frame_stats(
const TWO_PASS *p,
const TWO_PASS_FRAME *p_frame,
int offset) {
if ((offset >= 0 &&
p_frame->stats_in + offset >= p->stats_buf_ctx->stats_in_end) ||
(offset < 0 &&
p_frame->stats_in + offset < p->stats_buf_ctx->stats_in_start)) {
return NULL;
}
return &p_frame->stats_in[offset];
}
// This function returns the maximum target rate per frame.
static int frame_max_bits(
const RATE_CONTROL *rc,
const AV1EncoderConfig *oxcf) {
int64_t max_bits = ((int64_t)rc->avg_frame_bandwidth *
(int64_t)oxcf->rc_cfg.vbrmax_section) /
100;
if (max_bits < 0)
max_bits = 0;
else if (max_bits > rc->max_frame_bandwidth)
max_bits = rc->max_frame_bandwidth;
return (
int)max_bits;
}
// Based on history adjust expectations of bits per macroblock.
static void twopass_update_bpm_factor(AV1_COMP *cpi,
int rate_err_tol) {
TWO_PASS *
const twopass = &cpi->ppi->twopass;
const PRIMARY_RATE_CONTROL *
const p_rc = &cpi->ppi->p_rc;
// Based on recent history adjust expectations of bits per macroblock.
double rate_err_factor = 1.0;
const double adj_limit = AOMMAX(0.2, (
double)(100 - rate_err_tol) / 200.0);
const double min_fac = 1.0 - adj_limit;
const double max_fac = 1.0 + adj_limit;
#if CONFIG_THREE_PASS
if (cpi->third_pass_ctx && cpi->third_pass_ctx->frame_info_count > 0) {
int64_t actual_bits = 0;
int64_t target_bits = 0;
double factor = 0.0;
int count = 0;
for (
int i = 0; i < cpi->third_pass_ctx->frame_info_count; i++) {
actual_bits += cpi->third_pass_ctx->frame_info[i].actual_bits;
target_bits += cpi->third_pass_ctx->frame_info[i].bits_allocated;
factor += cpi->third_pass_ctx->frame_info[i].bpm_factor;
count++;
}
if (count == 0) {
factor = 1.0;
}
else {
factor /= (
double)count;
}
factor *= (
double)actual_bits / DOUBLE_DIVIDE_CHECK((
double)target_bits);
if ((twopass->bpm_factor <= 1 && factor < twopass->bpm_factor) ||
(twopass->bpm_factor >= 1 && factor > twopass->bpm_factor)) {
twopass->bpm_factor = factor;
twopass->bpm_factor =
AOMMAX(min_fac, AOMMIN(max_fac, twopass->bpm_factor));
}
}
#endif // CONFIG_THREE_PASS
int err_estimate = p_rc->rate_error_estimate;
int64_t total_actual_bits = p_rc->total_actual_bits;
double rolling_arf_group_actual_bits =
(
double)twopass->rolling_arf_group_actual_bits;
double rolling_arf_group_target_bits =
(
double)twopass->rolling_arf_group_target_bits;
#if CONFIG_FPMT_TEST
const int is_parallel_frame =
cpi->ppi->gf_group.frame_parallel_level[cpi->gf_frame_index] > 0 ? 1 : 0;
const int simulate_parallel_frame =
cpi->ppi->fpmt_unit_test_cfg == PARALLEL_SIMULATION_ENCODE
? is_parallel_frame
: 0;
total_actual_bits = simulate_parallel_frame ? p_rc->temp_total_actual_bits
: p_rc->total_actual_bits;
rolling_arf_group_target_bits =
(
double)(simulate_parallel_frame
? p_rc->temp_rolling_arf_group_target_bits
: twopass->rolling_arf_group_target_bits);
rolling_arf_group_actual_bits =
(
double)(simulate_parallel_frame
? p_rc->temp_rolling_arf_group_actual_bits
: twopass->rolling_arf_group_actual_bits);
err_estimate = simulate_parallel_frame ? p_rc->temp_rate_error_estimate
: p_rc->rate_error_estimate;
#endif
if ((p_rc->bits_off_target && total_actual_bits > 0) &&
(rolling_arf_group_target_bits >= 1.0)) {
if (rolling_arf_group_actual_bits > rolling_arf_group_target_bits) {
double error_fraction =
(rolling_arf_group_actual_bits - rolling_arf_group_target_bits) /
rolling_arf_group_target_bits;
error_fraction = (error_fraction > 1.0) ? 1.0 : error_fraction;
rate_err_factor = 1.0 + error_fraction;
}
else {
double error_fraction =
(rolling_arf_group_target_bits - rolling_arf_group_actual_bits) /
rolling_arf_group_target_bits;
rate_err_factor = 1.0 - error_fraction;
}
rate_err_factor = AOMMAX(min_fac, AOMMIN(max_fac, rate_err_factor));
}
// Is the rate control trending in the right direction. Only make
// an adjustment if things are getting worse.
if ((rate_err_factor < 1.0 && err_estimate >= 0) ||
(rate_err_factor > 1.0 && err_estimate <= 0)) {
twopass->bpm_factor *= rate_err_factor;
twopass->bpm_factor = AOMMAX(min_fac, AOMMIN(max_fac, twopass->bpm_factor));
}
}
static const double q_div_term[(QINDEX_RANGE >> 4) + 1] = {
18.0, 30.0, 38.0, 44.0, 47.0, 50.0, 52.0, 54.0, 56.0,
58.0, 60.0, 62.0, 64.0, 66.0, 68.0, 70.0, 72.0
};
#define EPMB_SCALER 1250000
static double calc_correction_factor(
double err_per_mb,
int q) {
double power_term = 0.90;
const int index = q >> 4;
const double divisor =
q_div_term[index] +
(((q_div_term[index + 1] - q_div_term[index]) * (q % 16)) / 16.0);
double error_term = EPMB_SCALER * pow(err_per_mb, power_term);
return error_term / divisor;
}
// Similar to find_qindex_by_rate() function in ratectrl.c, but includes
// calculation of a correction_factor.
static int find_qindex_by_rate_with_correction(uint64_t desired_bits_per_mb,
aom_bit_depth_t bit_depth,
double error_per_mb,
double group_weight_factor,
int best_qindex,
int worst_qindex) {
assert(best_qindex <= worst_qindex);
int low = best_qindex;
int high = worst_qindex;
while (low < high) {
const int mid = (low + high) >> 1;
const double q_factor = calc_correction_factor(error_per_mb, mid);
const double q = av1_convert_qindex_to_q(mid, bit_depth);
const uint64_t mid_bits_per_mb =
(uint64_t)((q_factor * group_weight_factor) / q);
if (mid_bits_per_mb > desired_bits_per_mb) {
low = mid + 1;
}
else {
high = mid;
}
}
return low;
}
/*!\brief Choose a target maximum Q for a group of frames
*
* \ingroup rate_control
*
* This function is used to estimate a suitable maximum Q for a
* group of frames. Inititally it is called to get a crude estimate
* for the whole clip. It is then called for each ARF/GF group to get
* a revised estimate for that group.
*
* \param[in] cpi Top-level encoder structure
* \param[in] av_frame_err The average per frame coded error score
* for frames making up this section/group.
* \param[in] inactive_zone Used to mask off /ignore part of the
* frame. The most common use case is where
* a wide format video (e.g. 16:9) is
* letter-boxed into a more square format.
* Here we want to ignore the bands at the
* top and bottom.
* \param[in] av_target_bandwidth The target bits per frame
*
* \return The maximum Q for frames in the group.
*/
static int get_twopass_worst_quality(AV1_COMP *cpi,
const double av_frame_err,
double inactive_zone,
int av_target_bandwidth) {
const RATE_CONTROL *
const rc = &cpi->rc;
const AV1EncoderConfig *
const oxcf = &cpi->oxcf;
const RateControlCfg *
const rc_cfg = &oxcf->rc_cfg;
inactive_zone = fclamp(inactive_zone, 0.0, 0.9999);
if (av_target_bandwidth <= 0) {
return rc->worst_quality;
// Highest value allowed
}
else {
const int num_mbs = (oxcf->resize_cfg.resize_mode != RESIZE_NONE)
? cpi->initial_mbs
: cpi->common.mi_params.MBs;
const int active_mbs = AOMMAX(1, num_mbs - (
int)(num_mbs * inactive_zone));
const double av_err_per_mb = av_frame_err / (1.0 - inactive_zone);
const uint64_t target_norm_bits_per_mb =
((uint64_t)av_target_bandwidth << BPER_MB_NORMBITS) / active_mbs;
int rate_err_tol = AOMMIN(rc_cfg->under_shoot_pct, rc_cfg->over_shoot_pct);
const double size_factor =
(active_mbs < 500) ? 0.925 : ((active_mbs > 3000) ? 1.05 : 1.0);
const double speed_factor =
AOMMIN(1.02, (0.975 + (0.005 * cpi->oxcf.speed)));
// Update bpm correction factor based on previous GOP rate error.
twopass_update_bpm_factor(cpi, rate_err_tol);
// Try and pick a max Q that will be high enough to encode the
// content at the given rate.
int q = find_qindex_by_rate_with_correction(
target_norm_bits_per_mb, cpi->common.seq_params->bit_depth,
av_err_per_mb,
cpi->ppi->twopass.bpm_factor * speed_factor * size_factor,
rc->best_quality, rc->worst_quality);
// Restriction on active max q for constrained quality mode.
if (rc_cfg->mode == AOM_CQ) q = AOMMAX(q, rc_cfg->cq_level);
return q;
}
}
#define INTRA_PART 0.005
#define DEFAULT_DECAY_LIMIT 0.75
#define LOW_SR_DIFF_TRHESH 0.01
#define NCOUNT_FRAME_II_THRESH 5.0
#define LOW_CODED_ERR_PER_MB 0.01
/* This function considers how the quality of prediction may be deteriorating
* with distance. It comapres the coded error for the last frame and the
* second reference frame (usually two frames old) and also applies a factor
* based on the extent of INTRA coding.
*
* The decay factor is then used to reduce the contribution of frames further
* from the alt-ref or golden frame, to the bitframe boost calculation for that
* alt-ref or golden frame.
*/
static double get_sr_decay_rate(
const FIRSTPASS_STATS *frame) {
double sr_diff = (frame->sr_coded_error - frame->coded_error);
double sr_decay = 1.0;
double modified_pct_inter;
double modified_pcnt_intra;
modified_pct_inter = frame->pcnt_inter;
if ((frame->coded_error > LOW_CODED_ERR_PER_MB) &&
((frame->intra_error / DOUBLE_DIVIDE_CHECK(frame->coded_error)) <
(
double)NCOUNT_FRAME_II_THRESH)) {
modified_pct_inter = frame->pcnt_inter - frame->pcnt_neutral;
}
modified_pcnt_intra = 100 * (1.0 - modified_pct_inter);
if ((sr_diff > LOW_SR_DIFF_TRHESH)) {
double sr_diff_part = ((sr_diff * 0.25) / frame->intra_error);
sr_decay = 1.0 - sr_diff_part - (INTRA_PART * modified_pcnt_intra);
}
return AOMMAX(sr_decay, DEFAULT_DECAY_LIMIT);
}
// This function gives an estimate of how badly we believe the prediction
// quality is decaying from frame to frame.
static double get_zero_motion_factor(
const FIRSTPASS_STATS *frame) {
const double zero_motion_pct = frame->pcnt_inter - frame->pcnt_motion;
double sr_decay = get_sr_decay_rate(frame);
return AOMMIN(sr_decay, zero_motion_pct);
}
#define DEFAULT_ZM_FACTOR 0.5
static double get_prediction_decay_rate(
const FIRSTPASS_STATS *frame_stats) {
const double sr_decay_rate = get_sr_decay_rate(frame_stats);
double zero_motion_factor =
DEFAULT_ZM_FACTOR * (frame_stats->pcnt_inter - frame_stats->pcnt_motion);
// Clamp value to range 0.0 to 1.0
// This should happen anyway if input values are sensibly clamped but checked
// here just in case.
if (zero_motion_factor > 1.0)
zero_motion_factor = 1.0;
else if (zero_motion_factor < 0.0)
zero_motion_factor = 0.0;
return AOMMAX(zero_motion_factor,
(sr_decay_rate + ((1.0 - sr_decay_rate) * zero_motion_factor)));
}
// Function to test for a condition where a complex transition is followed
// by a static section. For example in slide shows where there is a fade
// between slides. This is to help with more optimal kf and gf positioning.
static int detect_transition_to_still(
const FIRSTPASS_INFO *firstpass_info,
int next_stats_index,
const int min_gf_interval,
const int frame_interval,
const int still_interval,
const double loop_decay_rate,
const double last_decay_rate) {
// Break clause to detect very still sections after motion
// For example a static image after a fade or other transition
// instead of a clean scene cut.
if (frame_interval > min_gf_interval && loop_decay_rate >= 0.999 &&
last_decay_rate < 0.9) {
int stats_left =
av1_firstpass_info_future_count(firstpass_info, next_stats_index);
if (stats_left >= still_interval) {
int j;
// Look ahead a few frames to see if static condition persists...
for (j = 0; j < still_interval; ++j) {
const FIRSTPASS_STATS *stats =
av1_firstpass_info_peek(firstpass_info, next_stats_index + j);
if (stats->pcnt_inter - stats->pcnt_motion < 0.999)
break;
}
// Only if it does do we signal a transition to still.
return j == still_interval;
}
}
return 0;
}
// This function detects a flash through the high relative pcnt_second_ref
// score in the frame following a flash frame. The offset passed in should
// reflect this.
static int detect_flash(
const TWO_PASS *twopass,
const TWO_PASS_FRAME *twopass_frame,
const int offset) {
const FIRSTPASS_STATS *
const next_frame =
read_frame_stats(twopass, twopass_frame, offset);
// What we are looking for here is a situation where there is a
// brief break in prediction (such as a flash) but subsequent frames
// are reasonably well predicted by an earlier (pre flash) frame.
// The recovery after a flash is indicated by a high pcnt_second_ref
// compared to pcnt_inter.
return next_frame != NULL &&
next_frame->pcnt_second_ref > next_frame->pcnt_inter &&
next_frame->pcnt_second_ref >= 0.5;
}
// Update the motion related elements to the GF arf boost calculation.
static void accumulate_frame_motion_stats(
const FIRSTPASS_STATS *stats,
GF_GROUP_STATS *gf_stats,
double f_w,
double f_h) {
const double pct = stats->pcnt_motion;
// Accumulate Motion In/Out of frame stats.
gf_stats->this_frame_mv_in_out = stats->mv_in_out_count * pct;
gf_stats->mv_in_out_accumulator += gf_stats->this_frame_mv_in_out;
gf_stats->abs_mv_in_out_accumulator += fabs(gf_stats->this_frame_mv_in_out);
// Accumulate a measure of how uniform (or conversely how random) the motion
// field is (a ratio of abs(mv) / mv).
if (pct > 0.05) {
const double mvr_ratio =
fabs(stats->mvr_abs) / DOUBLE_DIVIDE_CHECK(fabs(stats->MVr));
const double mvc_ratio =
fabs(stats->mvc_abs) / DOUBLE_DIVIDE_CHECK(fabs(stats->MVc));
gf_stats->mv_ratio_accumulator +=
pct *
(mvr_ratio < stats->mvr_abs * f_h ? mvr_ratio : stats->mvr_abs * f_h);
gf_stats->mv_ratio_accumulator +=
pct *
(mvc_ratio < stats->mvc_abs * f_w ? mvc_ratio : stats->mvc_abs * f_w);
}
}
static void accumulate_this_frame_stats(
const FIRSTPASS_STATS *stats,
const double mod_frame_err,
GF_GROUP_STATS *gf_stats) {
gf_stats->gf_group_err += mod_frame_err;
#if GROUP_ADAPTIVE_MAXQ
gf_stats->gf_group_raw_error += stats->coded_error;
#endif
gf_stats->gf_group_skip_pct += stats->intra_skip_pct;
gf_stats->gf_group_inactive_zone_rows += stats->inactive_zone_rows;
}
static void accumulate_next_frame_stats(
const FIRSTPASS_STATS *stats,
const int flash_detected,
const int frames_since_key,
const int cur_idx,
GF_GROUP_STATS *gf_stats,
int f_w,
int f_h) {
accumulate_frame_motion_stats(stats, gf_stats, f_w, f_h);
// sum up the metric values of current gf group
gf_stats->avg_sr_coded_error += stats->sr_coded_error;
gf_stats->avg_pcnt_second_ref += stats->pcnt_second_ref;
gf_stats->avg_new_mv_count += stats->new_mv_count;
gf_stats->avg_wavelet_energy += stats->frame_avg_wavelet_energy;
if (fabs(stats->raw_error_stdev) > 0.000001) {
gf_stats->non_zero_stdev_count++;
gf_stats->avg_raw_err_stdev += stats->raw_error_stdev;
}
// Accumulate the effect of prediction quality decay
if (!flash_detected) {
gf_stats->last_loop_decay_rate = gf_stats->loop_decay_rate;
gf_stats->loop_decay_rate = get_prediction_decay_rate(stats);
gf_stats->decay_accumulator =
gf_stats->decay_accumulator * gf_stats->loop_decay_rate;
// Monitor for static sections.
if ((frames_since_key + cur_idx - 1) > 1) {
gf_stats->zero_motion_accumulator = AOMMIN(
gf_stats->zero_motion_accumulator, get_zero_motion_factor(stats));
}
}
}
static void average_gf_stats(
const int total_frame, GF_GROUP_STATS *gf_stats) {
if (total_frame) {
gf_stats->avg_sr_coded_error /= total_frame;
gf_stats->avg_pcnt_second_ref /= total_frame;
gf_stats->avg_new_mv_count /= total_frame;
gf_stats->avg_wavelet_energy /= total_frame;
}
if (gf_stats->non_zero_stdev_count)
gf_stats->avg_raw_err_stdev /= gf_stats->non_zero_stdev_count;
}
#define BOOST_FACTOR 12.5
static double baseline_err_per_mb(
const FRAME_INFO *frame_info) {
unsigned int screen_area = frame_info->frame_height * frame_info->frame_width;
// Use a different error per mb factor for calculating boost for
// different formats.
if (screen_area <= 640 * 360) {
return 500.0;
}
else {
return 1000.0;
}
}
static double calc_frame_boost(
const PRIMARY_RATE_CONTROL *p_rc,
const FRAME_INFO *frame_info,
const FIRSTPASS_STATS *this_frame,
double this_frame_mv_in_out,
double max_boost) {
double frame_boost;
const double lq = av1_convert_qindex_to_q(p_rc->avg_frame_qindex[INTER_FRAME],
frame_info->bit_depth);
const double boost_q_correction = AOMMIN((0.5 + (lq * 0.015)), 1.5);
const double active_area = calculate_active_area(frame_info, this_frame);
// Underlying boost factor is based on inter error ratio.
frame_boost = AOMMAX(baseline_err_per_mb(frame_info) * active_area,
this_frame->intra_error * active_area) /
DOUBLE_DIVIDE_CHECK(this_frame->coded_error);
frame_boost = frame_boost * BOOST_FACTOR * boost_q_correction;
// Increase boost for frames where new data coming into frame (e.g. zoom out).
// Slightly reduce boost if there is a net balance of motion out of the frame
// (zoom in). The range for this_frame_mv_in_out is -1.0 to +1.0.
if (this_frame_mv_in_out > 0.0)
frame_boost += frame_boost * (this_frame_mv_in_out * 2.0);
// In the extreme case the boost is halved.
else
frame_boost += frame_boost * (this_frame_mv_in_out / 2.0);
return AOMMIN(frame_boost, max_boost * boost_q_correction);
}
static double calc_kf_frame_boost(
const PRIMARY_RATE_CONTROL *p_rc,
const FRAME_INFO *frame_info,
const FIRSTPASS_STATS *this_frame,
double *sr_accumulator,
double max_boost) {
double frame_boost;
const double lq = av1_convert_qindex_to_q(p_rc->avg_frame_qindex[INTER_FRAME],
frame_info->bit_depth);
const double boost_q_correction = AOMMIN((0.50 + (lq * 0.015)), 2.00);
const double active_area = calculate_active_area(frame_info, this_frame);
// Underlying boost factor is based on inter error ratio.
frame_boost = AOMMAX(baseline_err_per_mb(frame_info) * active_area,
this_frame->intra_error * active_area) /
DOUBLE_DIVIDE_CHECK(
(this_frame->coded_error + *sr_accumulator) * active_area);
// Update the accumulator for second ref error difference.
// This is intended to give an indication of how much the coded error is
// increasing over time.
*sr_accumulator += (this_frame->sr_coded_error - this_frame->coded_error);
*sr_accumulator = AOMMAX(0.0, *sr_accumulator);
// Q correction and scaling
// The 40.0 value here is an experimentally derived baseline minimum.
// This value is in line with the minimum per frame boost in the alt_ref
// boost calculation.
frame_boost = ((frame_boost + 40.0) * boost_q_correction);
return AOMMIN(frame_boost, max_boost * boost_q_correction);
}
static int get_projected_gfu_boost(
const PRIMARY_RATE_CONTROL *p_rc,
int gfu_boost,
int frames_to_project,
int num_stats_used_for_gfu_boost) {
/*
* If frames_to_project is equal to num_stats_used_for_gfu_boost,
* it means that gfu_boost was calculated over frames_to_project to
* begin with(ie; all stats required were available), hence return
* the original boost.
*/
if (num_stats_used_for_gfu_boost >= frames_to_project)
return gfu_boost;
double min_boost_factor = sqrt(p_rc->baseline_gf_interval);
// Get the current tpl factor (number of frames = frames_to_project).
double tpl_factor = av1_get_gfu_boost_projection_factor(
min_boost_factor, MAX_GFUBOOST_FACTOR, frames_to_project);
// Get the tpl factor when number of frames = num_stats_used_for_prior_boost.
double tpl_factor_num_stats = av1_get_gfu_boost_projection_factor(
min_boost_factor, MAX_GFUBOOST_FACTOR, num_stats_used_for_gfu_boost);
int projected_gfu_boost =
(
int)rint((tpl_factor * gfu_boost) / tpl_factor_num_stats);
return projected_gfu_boost;
}
#define GF_MAX_BOOST 90.0
#define GF_MIN_BOOST 50
#define MIN_DECAY_FACTOR 0.01
int av1_calc_arf_boost(
const TWO_PASS *twopass,
const TWO_PASS_FRAME *twopass_frame,
const PRIMARY_RATE_CONTROL *p_rc, FRAME_INFO *frame_info,
int offset,
int f_frames,
int b_frames,
int *num_fpstats_used,
int *num_fpstats_required,
int project_gfu_boost) {
int i;
GF_GROUP_STATS gf_stats;
init_gf_stats(&gf_stats);
double boost_score = (
double)NORMAL_BOOST;
int arf_boost;
int flash_detected = 0;
if (num_fpstats_used) *num_fpstats_used = 0;
// Search forward from the proposed arf/next gf position.
for (i = 0; i < f_frames; ++i) {
const FIRSTPASS_STATS *this_frame =
read_frame_stats(twopass, twopass_frame, i + offset);
if (this_frame == NULL)
break;
// Update the motion related elements to the boost calculation.
accumulate_frame_motion_stats(this_frame, &gf_stats,
frame_info->frame_width,
frame_info->frame_height);
// We want to discount the flash frame itself and the recovery
// frame that follows as both will have poor scores.
flash_detected = detect_flash(twopass, twopass_frame, i + offset) ||
detect_flash(twopass, twopass_frame, i + offset + 1);
// Accumulate the effect of prediction quality decay.
if (!flash_detected) {
gf_stats.decay_accumulator *= get_prediction_decay_rate(this_frame);
gf_stats.decay_accumulator = gf_stats.decay_accumulator < MIN_DECAY_FACTOR
? MIN_DECAY_FACTOR
: gf_stats.decay_accumulator;
}
boost_score +=
gf_stats.decay_accumulator *
calc_frame_boost(p_rc, frame_info, this_frame,
gf_stats.this_frame_mv_in_out, GF_MAX_BOOST);
if (num_fpstats_used) (*num_fpstats_used)++;
}
arf_boost = (
int)boost_score;
// Reset for backward looking loop.
boost_score = 0.0;
init_gf_stats(&gf_stats);
// Search backward towards last gf position.
for (i = -1; i >= -b_frames; --i) {
const FIRSTPASS_STATS *this_frame =
read_frame_stats(twopass, twopass_frame, i + offset);
if (this_frame == NULL)
break;
// Update the motion related elements to the boost calculation.
accumulate_frame_motion_stats(this_frame, &gf_stats,
frame_info->frame_width,
frame_info->frame_height);
// We want to discount the the flash frame itself and the recovery
// frame that follows as both will have poor scores.
flash_detected = detect_flash(twopass, twopass_frame, i + offset) ||
detect_flash(twopass, twopass_frame, i + offset + 1);
// Cumulative effect of prediction quality decay.
if (!flash_detected) {
gf_stats.decay_accumulator *= get_prediction_decay_rate(this_frame);
gf_stats.decay_accumulator = gf_stats.decay_accumulator < MIN_DECAY_FACTOR
? MIN_DECAY_FACTOR
: gf_stats.decay_accumulator;
}
boost_score +=
gf_stats.decay_accumulator *
calc_frame_boost(p_rc, frame_info, this_frame,
gf_stats.this_frame_mv_in_out, GF_MAX_BOOST);
if (num_fpstats_used) (*num_fpstats_used)++;
}
arf_boost += (
int)boost_score;
if (project_gfu_boost) {
assert(num_fpstats_required != NULL);
assert(num_fpstats_used != NULL);
*num_fpstats_required = f_frames + b_frames;
arf_boost = get_projected_gfu_boost(p_rc, arf_boost, *num_fpstats_required,
*num_fpstats_used);
}
if (arf_boost < ((b_frames + f_frames) * GF_MIN_BOOST))
arf_boost = ((b_frames + f_frames) * GF_MIN_BOOST);
return arf_boost;
}
// Calculate a section intra ratio used in setting max loop filter.
static int calculate_section_intra_ratio(
const FIRSTPASS_STATS *begin,
const FIRSTPASS_STATS *end,
int section_length) {
const FIRSTPASS_STATS *s = begin;
double intra_error = 0.0;
double coded_error = 0.0;
int i = 0;
while (s < end && i < section_length) {
intra_error += s->intra_error;
coded_error += s->coded_error;
++s;
++i;
}
return (
int)(intra_error / DOUBLE_DIVIDE_CHECK(coded_error));
}
/*!\brief Calculates the bit target for this GF/ARF group
*
* \ingroup rate_control
*
* Calculates the total bits to allocate in this GF/ARF group.
*
* \param[in] cpi Top-level encoder structure
* \param[in] gf_group_err Cumulative coded error score for the
* frames making up this group.
*
* \return The target total number of bits for this GF/ARF group.
*/
static int64_t calculate_total_gf_group_bits(AV1_COMP *cpi,
double gf_group_err) {
const RATE_CONTROL *
const rc = &cpi->rc;
const PRIMARY_RATE_CONTROL *
const p_rc = &cpi->ppi->p_rc;
const TWO_PASS *
const twopass = &cpi->ppi->twopass;
const int max_bits = frame_max_bits(rc, &cpi->oxcf);
int64_t total_group_bits;
// Calculate the bits to be allocated to the group as a whole.
if ((twopass->kf_group_bits > 0) && (twopass->kf_group_error_left > 0)) {
total_group_bits = (int64_t)(twopass->kf_group_bits *
(gf_group_err / twopass->kf_group_error_left));
}
else {
total_group_bits = 0;
}
// Clamp odd edge cases.
total_group_bits = (total_group_bits < 0) ? 0
: (total_group_bits > twopass->kf_group_bits)
? twopass->kf_group_bits
: total_group_bits;
// Clip based on user supplied data rate variability limit.
if (total_group_bits > (int64_t)max_bits * p_rc->baseline_gf_interval)
total_group_bits = (int64_t)max_bits * p_rc->baseline_gf_interval;
return total_group_bits;
}
// Calculate the number of bits to assign to boosted frames in a group.
static int calculate_boost_bits(
int frame_count,
int boost,
int64_t total_group_bits) {
int allocation_chunks;
// return 0 for invalid inputs (could arise e.g. through rounding errors)
if (!boost || (total_group_bits <= 0))
return 0;
if (frame_count <= 0)
return (
int)(AOMMIN(total_group_bits, INT_MAX));
allocation_chunks = (frame_count * 100) + boost;
// Prevent overflow.
if (boost > 1023) {
int divisor = boost >> 10;
boost /= divisor;
allocation_chunks /= divisor;
}
// Calculate the number of extra bits for use in the boosted frame or frames.
return AOMMAX((
int)(((int64_t)boost * total_group_bits) / allocation_chunks),
0);
}
// Calculate the boost factor based on the number of bits assigned, i.e. the
// inverse of calculate_boost_bits().
static int calculate_boost_factor(
int frame_count,
int bits,
int64_t total_group_bits) {
return (
int)(100.0 * frame_count * bits / (total_group_bits - bits));
}
// Reduce the number of bits assigned to keyframe or arf if necessary, to
// prevent bitrate spikes that may break level constraints.
// frame_type: 0: keyframe; 1: arf.
static int adjust_boost_bits_for_target_level(
const AV1_COMP *
const cpi,
RATE_CONTROL *
const rc,
int bits_assigned,
int64_t group_bits,
int frame_type) {
const AV1_COMMON *
const cm = &cpi->common;
const SequenceHeader *
const seq_params = cm->seq_params;
PRIMARY_RATE_CONTROL *
const p_rc = &cpi->ppi->p_rc;
const int temporal_layer_id = cm->temporal_layer_id;
const int spatial_layer_id = cm->spatial_layer_id;
for (
int index = 0; index < seq_params->operating_points_cnt_minus_1 + 1;
++index) {
if (!is_in_operating_point(seq_params->operating_point_idc[index],
temporal_layer_id, spatial_layer_id)) {
continue;
}
const AV1_LEVEL target_level =
cpi->ppi->level_params.target_seq_level_idx[index];
if (target_level >= SEQ_LEVELS)
continue;
assert(is_valid_seq_level_idx(target_level));
const double level_bitrate_limit = av1_get_max_bitrate_for_level(
target_level, seq_params->tier[0], seq_params->profile);
const int target_bits_per_frame =
(
int)(level_bitrate_limit / cpi->framerate);
if (frame_type == 0) {
// Maximum bits for keyframe is 8 times the target_bits_per_frame.
const int level_enforced_max_kf_bits = target_bits_per_frame * 8;
if (bits_assigned > level_enforced_max_kf_bits) {
const int frames = rc->frames_to_key - 1;
p_rc->kf_boost = calculate_boost_factor(
frames, level_enforced_max_kf_bits, group_bits);
bits_assigned =
calculate_boost_bits(frames, p_rc->kf_boost, group_bits);
}
}
else if (frame_type == 1) {
// Maximum bits for arf is 4 times the target_bits_per_frame.
const int level_enforced_max_arf_bits = target_bits_per_frame * 4;
if (bits_assigned > level_enforced_max_arf_bits) {
p_rc->gfu_boost =
calculate_boost_factor(p_rc->baseline_gf_interval,
level_enforced_max_arf_bits, group_bits);
bits_assigned = calculate_boost_bits(p_rc->baseline_gf_interval,
p_rc->gfu_boost, group_bits);
}
}
else {
assert(0);
}
}
return bits_assigned;
}
// Allocate bits to each frame in a GF / ARF group
static void allocate_gf_group_bits(GF_GROUP *gf_group,
PRIMARY_RATE_CONTROL *
const p_rc,
RATE_CONTROL *
const rc,
int64_t gf_group_bits,
int gf_arf_bits,
int key_frame,
int use_arf) {
static const double layer_fraction[MAX_ARF_LAYERS + 1] = { 1.0, 0.70, 0.55,
0.60, 0.60, 1.0,
1.0 };
int64_t total_group_bits = gf_group_bits;
int base_frame_bits;
const int gf_group_size = gf_group->size;
int layer_frames[MAX_ARF_LAYERS + 1] = { 0 };
// For key frames the frame target rate is already set and it
// is also the golden frame.
// === [frame_index == 0] ===
int frame_index = !!key_frame;
// Subtract the extra bits set aside for ARF frames from the Group Total
if (use_arf) total_group_bits -= gf_arf_bits;
int num_frames =
AOMMAX(1, p_rc->baseline_gf_interval - (rc->frames_since_key == 0));
base_frame_bits = (
int)(total_group_bits / num_frames);
// Check the number of frames in each layer in case we have a
// non standard group length.
int max_arf_layer = gf_group->max_layer_depth - 1;
for (
int idx = frame_index; idx < gf_group_size; ++idx) {
if ((gf_group->update_type[idx] == ARF_UPDATE) ||
(gf_group->update_type[idx] == INTNL_ARF_UPDATE)) {
layer_frames[gf_group->layer_depth[idx]]++;
}
}
// Allocate extra bits to each ARF layer
int i;
int layer_extra_bits[MAX_ARF_LAYERS + 1] = { 0 };
assert(max_arf_layer <= MAX_ARF_LAYERS);
for (i = 1; i <= max_arf_layer; ++i) {
double fraction = (i == max_arf_layer) ? 1.0 : layer_fraction[i];
layer_extra_bits[i] =
(
int)((gf_arf_bits * fraction) / AOMMAX(1, layer_frames[i]));
gf_arf_bits -= (
int)(gf_arf_bits * fraction);
}
// Now combine ARF layer and baseline bits to give total bits for each frame.
int arf_extra_bits;
for (
int idx = frame_index; idx < gf_group_size; ++idx) {
switch (gf_group->update_type[idx]) {
case ARF_UPDATE:
case INTNL_ARF_UPDATE:
arf_extra_bits = layer_extra_bits[gf_group->layer_depth[idx]];
gf_group->bit_allocation[idx] =
(base_frame_bits > INT_MAX - arf_extra_bits)
? INT_MAX
: (base_frame_bits + arf_extra_bits);
break;
case INTNL_OVERLAY_UPDATE:
case OVERLAY_UPDATE: gf_group->bit_allocation[idx] = 0;
break;
default: gf_group->bit_allocation[idx] = base_frame_bits;
break;
}
}
// Set the frame following the current GOP to 0 bit allocation. For ARF
// groups, this next frame will be overlay frame, which is the first frame
// in the next GOP. For GF group, next GOP will overwrite the rate allocation.
// Setting this frame to use 0 bit (of out the current GOP budget) will
// simplify logics in reference frame management.
if (gf_group_size < MAX_STATIC_GF_GROUP_LENGTH)
gf_group->bit_allocation[gf_group_size] = 0;
}
// Returns true if KF group and GF group both are almost completely static.
static inline int is_almost_static(
double gf_zero_motion,
int kf_zero_motion,
int is_lap_enabled) {
if (is_lap_enabled) {
/*
* when LAP enabled kf_zero_motion is not reliable, so use strict
* constraint on gf_zero_motion.
*/
return (gf_zero_motion >= 0.999);
}
else {
return (gf_zero_motion >= 0.995) &&
(kf_zero_motion >= STATIC_KF_GROUP_THRESH);
}
}
#define ARF_ABS_ZOOM_THRESH 4.4
static inline int detect_gf_cut(AV1_COMP *cpi,
int frame_index,
int cur_start,
int flash_detected,
int active_max_gf_interval,
int active_min_gf_interval,
GF_GROUP_STATS *gf_stats) {
RATE_CONTROL *
const rc = &cpi->rc;
TWO_PASS *
const twopass = &cpi->ppi->twopass;
AV1_COMMON *
const cm = &cpi->common;
// Motion breakout threshold for loop below depends on image size.
const double mv_ratio_accumulator_thresh = (cm->height + cm->width) / 4.0;
if (!flash_detected) {
// Break clause to detect very still sections after motion. For example,
// a static image after a fade or other transition.
// TODO(angiebird): This is a temporary change, we will avoid using
// twopass_frame.stats_in in the follow-up CL
int index = (
int)(cpi->twopass_frame.stats_in -
twopass->stats_buf_ctx->stats_in_start);
if (detect_transition_to_still(&twopass->firstpass_info, index,
rc->min_gf_interval, frame_index - cur_start,
5, gf_stats->loop_decay_rate,
gf_stats->last_loop_decay_rate)) {
return 1;
}
}
// Some conditions to breakout after min interval.
if (frame_index - cur_start >= active_min_gf_interval &&
// If possible don't break very close to a kf
(rc->frames_to_key - frame_index >= rc->min_gf_interval) &&
((frame_index - cur_start) & 0x01) && !flash_detected &&
(gf_stats->mv_ratio_accumulator > mv_ratio_accumulator_thresh ||
gf_stats->abs_mv_in_out_accumulator > ARF_ABS_ZOOM_THRESH)) {
return 1;
}
// If almost totally static, we will not use the the max GF length later,
// so we can continue for more frames.
if (((frame_index - cur_start) >= active_max_gf_interval + 1) &&
!is_almost_static(gf_stats->zero_motion_accumulator,
twopass->kf_zeromotion_pct, cpi->ppi->lap_enabled)) {
return 1;
}
return 0;
}
static int is_shorter_gf_interval_better(
AV1_COMP *cpi,
const EncodeFrameParams *frame_params) {
const RATE_CONTROL *
const rc = &cpi->rc;
PRIMARY_RATE_CONTROL *
const p_rc = &cpi->ppi->p_rc;
int gop_length_decision_method = cpi->sf.tpl_sf.gop_length_decision_method;
int shorten_gf_interval;
av1_tpl_preload_rc_estimate(cpi, frame_params);
if (gop_length_decision_method == 2) {
// GF group length is decided based on GF boost and tpl stats of ARFs from
// base layer, (base+1) layer.
shorten_gf_interval =
(p_rc->gfu_boost <
p_rc->num_stats_used_for_gfu_boost * GF_MIN_BOOST * 1.4) &&
!av1_tpl_setup_stats(cpi, 3, frame_params);
}
else {
int do_complete_tpl = 1;
GF_GROUP *
const gf_group = &cpi->ppi->gf_group;
int is_temporal_filter_enabled =
(rc->frames_since_key > 0 && gf_group->arf_index > -1);
if (gop_length_decision_method == 1) {
// Check if tpl stats of ARFs from base layer, (base+1) layer,
// (base+2) layer can decide the GF group length.
int gop_length_eval = av1_tpl_setup_stats(cpi, 2, frame_params);
if (gop_length_eval != 2) {
do_complete_tpl = 0;
shorten_gf_interval = !gop_length_eval;
}
}
if (do_complete_tpl) {
// Decide GF group length based on complete tpl stats.
shorten_gf_interval = !av1_tpl_setup_stats(cpi, 1, frame_params);
// Tpl stats is reused when the ARF is temporally filtered and GF
// interval is not shortened.
if (is_temporal_filter_enabled && !shorten_gf_interval) {
cpi->skip_tpl_setup_stats = 1;
#if CONFIG_BITRATE_ACCURACY && !CONFIG_THREE_PASS
assert(cpi->gf_frame_index == 0);
av1_vbr_rc_update_q_index_list(&cpi->vbr_rc_info, &cpi->ppi->tpl_data,
gf_group,
cpi->common.seq_params->bit_depth);
#endif // CONFIG_BITRATE_ACCURACY
}
}
}
return shorten_gf_interval;
}
#define MIN_SHRINK_LEN 6
// the minimum length of gf if we are shrinking
#define SMOOTH_FILT_LEN 7
#define HALF_FILT_LEN (SMOOTH_FILT_LEN / 2)
#define WINDOW_SIZE 7
#define HALF_WIN (WINDOW_SIZE / 2)
// Smooth filter intra_error and coded_error in firstpass stats.
// If stats[i].is_flash==1, the ith element should not be used in the filtering.
static void smooth_filter_stats(
const FIRSTPASS_STATS *stats,
int start_idx,
int last_idx,
double *filt_intra_err,
double *filt_coded_err) {
// A 7-tap gaussian smooth filter
static const double smooth_filt[SMOOTH_FILT_LEN] = { 0.006, 0.061, 0.242,
0.383, 0.242, 0.061,
0.006 };
int i, j;
for (i = start_idx; i <= last_idx; i++) {
double total_wt = 0;
for (j = -HALF_FILT_LEN; j <= HALF_FILT_LEN; j++) {
int idx = AOMMIN(AOMMAX(i + j, start_idx), last_idx);
if (stats[idx].is_flash)
continue;
filt_intra_err[i] +=
smooth_filt[j + HALF_FILT_LEN] * stats[idx].intra_error;
total_wt += smooth_filt[j + HALF_FILT_LEN];
}
if (total_wt > 0.01) {
filt_intra_err[i] /= total_wt;
}
else {
filt_intra_err[i] = stats[i].intra_error;
}
}
for (i = start_idx; i <= last_idx; i++) {
double total_wt = 0;
for (j = -HALF_FILT_LEN; j <= HALF_FILT_LEN; j++) {
int idx = AOMMIN(AOMMAX(i + j, start_idx), last_idx);
// Coded error involves idx and idx - 1.
if (stats[idx].is_flash || (idx > 0 && stats[idx - 1].is_flash))
continue;
filt_coded_err[i] +=
smooth_filt[j + HALF_FILT_LEN] * stats[idx].coded_error;
total_wt += smooth_filt[j + HALF_FILT_LEN];
}
if (total_wt > 0.01) {
filt_coded_err[i] /= total_wt;
}
else {
filt_coded_err[i] = stats[i].coded_error;
}
}
}
// Calculate gradient
static void get_gradient(
const double *values,
int start,
int last,
double *grad) {
if (start == last) {
grad[start] = 0;
return;
}
for (
int i = start; i <= last; i++) {
int prev = AOMMAX(i - 1, start);
int next = AOMMIN(i + 1, last);
grad[i] = (values[next] - values[prev]) / (next - prev);
}
}
static int find_next_scenecut(
const FIRSTPASS_STATS *
const stats_start,
int first,
int last) {
// Identify unstable areas caused by scenecuts.
// Find the max and 2nd max coded error, and the average of the rest frames.
// If there is only one frame that yields a huge coded error, it is likely a
// scenecut.
double this_ratio, max_prev_ratio, max_next_ratio, max_prev_coded,
max_next_coded;
if (last - first == 0)
return -1;
for (
int i = first; i <= last; i++) {
if (stats_start[i].is_flash || (i > 0 && stats_start[i - 1].is_flash))
continue;
double temp_intra = AOMMAX(stats_start[i].intra_error, 0.01);
this_ratio = stats_start[i].coded_error / temp_intra;
// find the avg ratio in the preceding neighborhood
max_prev_ratio = 0;
max_prev_coded = 0;
for (
int j = AOMMAX(first, i - HALF_WIN); j < i; j++) {
if (stats_start[j].is_flash || (j > 0 && stats_start[j - 1].is_flash))
continue;
temp_intra = AOMMAX(stats_start[j].intra_error, 0.01);
double temp_ratio = stats_start[j].coded_error / temp_intra;
if (temp_ratio > max_prev_ratio) {
max_prev_ratio = temp_ratio;
}
if (stats_start[j].coded_error > max_prev_coded) {
max_prev_coded = stats_start[j].coded_error;
}
}
// find the avg ratio in the following neighborhood
max_next_ratio = 0;
max_next_coded = 0;
for (
int j = i + 1; j <= AOMMIN(i + HALF_WIN, last); j++) {
if (stats_start[i].is_flash || (i > 0 && stats_start[i - 1].is_flash))
continue;
temp_intra = AOMMAX(stats_start[j].intra_error, 0.01);
double temp_ratio = stats_start[j].coded_error / temp_intra;
if (temp_ratio > max_next_ratio) {
max_next_ratio = temp_ratio;
}
if (stats_start[j].coded_error > max_next_coded) {
max_next_coded = stats_start[j].coded_error;
}
}
if (max_prev_ratio < 0.001 && max_next_ratio < 0.001) {
// the ratios are very small, only check a small fixed threshold
if (this_ratio < 0.02)
continue;
}
else {
// check if this frame has a larger ratio than the neighborhood
double max_sr = stats_start[i].sr_coded_error;
if (i < last) max_sr = AOMMAX(max_sr, stats_start[i + 1].sr_coded_error);
double max_sr_fr_ratio =
max_sr / AOMMAX(stats_start[i].coded_error, 0.01);
if (max_sr_fr_ratio > 1.2)
continue;
if (this_ratio < 2 * AOMMAX(max_prev_ratio, max_next_ratio) &&
stats_start[i].coded_error <
2 * AOMMAX(max_prev_coded, max_next_coded)) {
continue;
}
}
return i;
}
return -1;
}
// Remove the region with index next_region.
// parameter merge: 0: merge with previous; 1: merge with next; 2:
// merge with both, take type from previous if possible
// After removing, next_region will be the index of the next region.
static void remove_region(
int merge, REGIONS *regions,
int *num_regions,
int *next_region) {
int k = *next_region;
assert(k < *num_regions);
if (*num_regions == 1) {
*num_regions = 0;
return;
}
if (k == 0) {
merge = 1;
}
else if (k == *num_regions - 1) {
merge = 0;
}
int num_merge = (merge == 2) ? 2 : 1;
switch (merge) {
case 0:
regions[k - 1].last = regions[k].last;
*next_region = k;
break;
case 1:
regions[k + 1].start = regions[k].start;
*next_region = k + 1;
break;
case 2:
regions[k - 1].last = regions[k + 1].last;
*next_region = k;
break;
default: assert(0);
}
*num_regions -= num_merge;
for (k = *next_region - (merge == 1); k < *num_regions; k++) {
regions[k] = regions[k + num_merge];
}
}
// Insert a region in the cur_region_idx. The start and last should both be in
// the current region. After insertion, the cur_region_idx will point to the
// last region that was splitted from the original region.
static void insert_region(
int start,
int last, REGION_TYPES type,
REGIONS *regions,
int *num_regions,
int *cur_region_idx) {
int k = *cur_region_idx;
REGION_TYPES this_region_type = regions[k].type;
int this_region_last = regions[k].last;
int num_add = (start != regions[k].start) + (last != regions[k].last);
// move the following regions further to the back
for (
int r = *num_regions - 1; r > k; r--) {
regions[r + num_add] = regions[r];
}
*num_regions += num_add;
if (start > regions[k].start) {
regions[k].last = start - 1;
k++;
regions[k].start = start;
}
regions[k].type = type;
if (last < this_region_last) {
regions[k].last = last;
k++;
regions[k].start = last + 1;
regions[k].last = this_region_last;
regions[k].type = this_region_type;
}
else {
regions[k].last = this_region_last;
}
*cur_region_idx = k;
}
// Get the average of stats inside a region.
static void analyze_region(
const FIRSTPASS_STATS *stats,
int k,
REGIONS *regions) {
int i;
regions[k].avg_cor_coeff = 0;
regions[k].avg_sr_fr_ratio = 0;
regions[k].avg_intra_err = 0;
regions[k].avg_coded_err = 0;
int check_first_sr = (k != 0);
for (i = regions[k].start; i <= regions[k].last; i++) {
if (i > regions[k].start || check_first_sr) {
double num_frames =
(
double)(regions[k].last - regions[k].start + check_first_sr);
double max_coded_error =
AOMMAX(stats[i].coded_error, stats[i - 1].coded_error);
double this_ratio =
stats[i].sr_coded_error / AOMMAX(max_coded_error, 0.001);
regions[k].avg_sr_fr_ratio += this_ratio / num_frames;
}
regions[k].avg_intra_err +=
stats[i].intra_error / (
double)(regions[k].last - regions[k].start + 1);
regions[k].avg_coded_err +=
stats[i].coded_error / (
double)(regions[k].last - regions[k].start + 1);
regions[k].avg_cor_coeff +=
AOMMAX(stats[i].cor_coeff, 0.001) /
(
double)(regions[k].last - regions[k].start + 1);
regions[k].avg_noise_var +=
AOMMAX(stats[i].noise_var, 0.001) /
(
double)(regions[k].last - regions[k].start + 1);
}
}
// Calculate the regions stats of every region.
static void get_region_stats(
const FIRSTPASS_STATS *stats, REGIONS *regions,
int num_regions) {
for (
int k = 0; k < num_regions; k++) {
analyze_region(stats, k, regions);
}
}
// Find tentative stable regions
static int find_stable_regions(
const FIRSTPASS_STATS *stats,
const double *grad_coded,
int this_start,
int this_last, REGIONS *regions) {
int i, j, k = 0;
regions[k].start = this_start;
for (i = this_start; i <= this_last; i++) {
// Check mean and variance of stats in a window
double mean_intra = 0.001, var_intra = 0.001;
double mean_coded = 0.001, var_coded = 0.001;
int count = 0;
for (j = -HALF_WIN; j <= HALF_WIN; j++) {
int idx = AOMMIN(AOMMAX(i + j, this_start), this_last);
if (stats[idx].is_flash || (idx > 0 && stats[idx - 1].is_flash))
continue;
mean_intra += stats[idx].intra_error;
var_intra += stats[idx].intra_error * stats[idx].intra_error;
mean_coded += stats[idx].coded_error;
var_coded += stats[idx].coded_error * stats[idx].coded_error;
count++;
}
REGION_TYPES cur_type;
if (count > 0) {
mean_intra /= (
double)count;
var_intra /= (
double)count;
mean_coded /= (
double)count;
var_coded /= (
double)count;
int is_intra_stable = (var_intra / (mean_intra * mean_intra) < 1.03);
int is_coded_stable = (var_coded / (mean_coded * mean_coded) < 1.04 &&
fabs(grad_coded[i]) / mean_coded < 0.05) ||
mean_coded / mean_intra < 0.05;
int is_coded_small = mean_coded < 0.5 * mean_intra;
cur_type = (is_intra_stable && is_coded_stable && is_coded_small)
? STABLE_REGION
: HIGH_VAR_REGION;
}
else {
cur_type = HIGH_VAR_REGION;
}
// mark a new region if type changes
if (i == regions[k].start) {
// first frame in the region
regions[k].type = cur_type;
}
else if (cur_type != regions[k].type) {
// Append a new region
regions[k].last = i - 1;
regions[k + 1].start = i;
regions[k + 1].type = cur_type;
k++;
}
}
regions[k].last = this_last;
return k + 1;
}
// Clean up regions that should be removed or merged.
static void cleanup_regions(REGIONS *regions,
int *num_regions) {
int k = 0;
while (k < *num_regions) {
if ((k > 0 && regions[k - 1].type == regions[k].type &&
regions[k].type != SCENECUT_REGION) ||
regions[k].last < regions[k].start) {
remove_region(0, regions, num_regions, &k);
}
else {
k++;
}
}
}
// Remove regions that are of type and shorter than length.
// Merge it with its neighboring regions.
static void remove_short_regions(REGIONS *regions,
int *num_regions,
REGION_TYPES type,
int length) {
int k = 0;
while (k < *num_regions && (*num_regions) > 1) {
if ((regions[k].last - regions[k].start + 1 < length &&
regions[k].type == type)) {
// merge current region with the previous and next regions
remove_region(2, regions, num_regions, &k);
}
else {
k++;
}
}
cleanup_regions(regions, num_regions);
}
static void adjust_unstable_region_bounds(
const FIRSTPASS_STATS *stats,
REGIONS *regions,
int *num_regions) {
int i, j, k;
// Remove regions that are too short. Likely noise.
remove_short_regions(regions, num_regions, STABLE_REGION, HALF_WIN);
remove_short_regions(regions, num_regions, HIGH_VAR_REGION, HALF_WIN);
get_region_stats(stats, regions, *num_regions);
// Adjust region boundaries. The thresholds are empirically obtained, but
// overall the performance is not very sensitive to small changes to them.
for (k = 0; k < *num_regions; k++) {
if (regions[k].type == STABLE_REGION)
continue;
if (k > 0) {
// Adjust previous boundary.
// First find the average intra/coded error in the previous
// neighborhood.
double avg_intra_err = 0;
const int starti = AOMMAX(regions[k - 1].last - WINDOW_SIZE + 1,
regions[k - 1].start + 1);
const int lasti = regions[k - 1].last;
int counti = 0;
for (i = starti; i <= lasti; i++) {
avg_intra_err += stats[i].intra_error;
counti++;
}
if (counti > 0) {
avg_intra_err = AOMMAX(avg_intra_err / (
double)counti, 0.001);
int count_coded = 0, count_grad = 0;
for (j = lasti + 1; j <= regions[k].last; j++) {
const int intra_close =
fabs(stats[j].intra_error - avg_intra_err) / avg_intra_err < 0.1;
const int coded_small = stats[j].coded_error / avg_intra_err < 0.1;
const int coeff_close = stats[j].cor_coeff > 0.995;
if (!coeff_close || !coded_small) count_coded--;
if (intra_close && count_coded >= 0 && count_grad >= 0) {
// this frame probably belongs to the previous stable region
regions[k - 1].last = j;
regions[k].start = j + 1;
}
else {
break;
}
}
}
}
// if k > 0
if (k < *num_regions - 1) {
// Adjust next boundary.
// First find the average intra/coded error in the next neighborhood.
double avg_intra_err = 0;
const int starti = regions[k + 1].start;
const int lasti = AOMMIN(regions[k + 1].last - 1,
regions[k + 1].start + WINDOW_SIZE - 1);
int counti = 0;
for (i = starti; i <= lasti; i++) {
avg_intra_err += stats[i].intra_error;
counti++;
}
if (counti > 0) {
avg_intra_err = AOMMAX(avg_intra_err / (
double)counti, 0.001);
// At the boundary, coded error is large, but still the frame is stable
int count_coded = 1, count_grad = 1;
for (j = starti - 1; j >= regions[k].start; j--) {
const int intra_close =
fabs(stats[j].intra_error - avg_intra_err) / avg_intra_err < 0.1;
const int coded_small =
stats[j + 1].coded_error / avg_intra_err < 0.1;
const int coeff_close = stats[j].cor_coeff > 0.995;
if (!coeff_close || !coded_small) count_coded--;
if (intra_close && count_coded >= 0 && count_grad >= 0) {
// this frame probably belongs to the next stable region
regions[k + 1].start = j;
regions[k].last = j - 1;
}
else {
break;
}
}
}
}
// if k < *num_regions - 1
}
// end of loop over all regions
cleanup_regions(regions, num_regions);
remove_short_regions(regions, num_regions, HIGH_VAR_REGION, HALF_WIN);
get_region_stats(stats, regions, *num_regions);
// If a stable regions has higher error than neighboring high var regions,
// or if the stable region has a lower average correlation,
// then it should be merged with them
k = 0;
while (k < *num_regions && (*num_regions) > 1) {
if (regions[k].type == STABLE_REGION &&
(regions[k].last - regions[k].start + 1) < 2 * WINDOW_SIZE &&
((k > 0 &&
// previous regions
(regions[k].avg_coded_err > regions[k - 1].avg_coded_err * 1.01 ||
regions[k].avg_cor_coeff < regions[k - 1].avg_cor_coeff * 0.999)) &&
(k < *num_regions - 1 &&
// next region
(regions[k].avg_coded_err > regions[k + 1].avg_coded_err * 1.01 ||
regions[k].avg_cor_coeff < regions[k + 1].avg_cor_coeff * 0.999)))) {
// merge current region with the previous and next regions
remove_region(2, regions, num_regions, &k);
analyze_region(stats, k - 1, regions);
}
else if (regions[k].type == HIGH_VAR_REGION &&
(regions[k].last - regions[k].start + 1) < 2 * WINDOW_SIZE &&
((k > 0 &&
// previous regions
(regions[k].avg_coded_err <
regions[k - 1].avg_coded_err * 0.99 ||
regions[k].avg_cor_coeff >
regions[k - 1].avg_cor_coeff * 1.001)) &&
(k < *num_regions - 1 &&
// next region
(regions[k].avg_coded_err <
regions[k + 1].avg_coded_err * 0.99 ||
regions[k].avg_cor_coeff >
regions[k + 1].avg_cor_coeff * 1.001)))) {
// merge current region with the previous and next regions
remove_region(2, regions, num_regions, &k);
analyze_region(stats, k - 1, regions);
}
else {
k++;
}
}
remove_short_regions(regions, num_regions, STABLE_REGION, WINDOW_SIZE);
remove_short_regions(regions, num_regions, HIGH_VAR_REGION, HALF_WIN);
}
// Identify blending regions.
static void find_blending_regions(
const FIRSTPASS_STATS *stats,
REGIONS *regions,
int *num_regions) {
int i, k = 0;
// Blending regions will have large content change, therefore will have a
// large consistent change in intra error.
int count_stable = 0;
while (k < *num_regions) {
if (regions[k].type == STABLE_REGION) {
k++;
count_stable++;
continue;
}
int dir = 0;
int start = 0, last;
for (i = regions[k].start; i <= regions[k].last; i++) {
// First mark the regions that has consistent large change of intra error.
if (k == 0 && i == regions[k].start)
continue;
if (stats[i].is_flash || (i > 0 && stats[i - 1].is_flash))
continue;
double grad = stats[i].intra_error - stats[i - 1].intra_error;
int large_change = fabs(grad) / AOMMAX(stats[i].intra_error, 0.01) > 0.05;
int this_dir = 0;
if (large_change) {
this_dir = (grad > 0) ? 1 : -1;
}
// the current trend continues
if (dir == this_dir)
continue;
if (dir != 0) {
// Mark the end of a new large change group and add it
last = i - 1;
insert_region(start, last, BLENDING_REGION, regions, num_regions, &k);
}
dir = this_dir;
if (k == 0 && i == regions[k].start + 1) {
start = i - 1;
}
else {
start = i;
}
}
if (dir != 0) {
last = regions[k].last;
insert_region(start, last, BLENDING_REGION, regions, num_regions, &k);
}
k++;
}
// If the blending region has very low correlation, mark it as high variance
// since we probably cannot benefit from it anyways.
get_region_stats(stats, regions, *num_regions);
for (k = 0; k < *num_regions; k++) {
if (regions[k].type != BLENDING_REGION)
continue;
if (regions[k].last == regions[k].start || regions[k].avg_cor_coeff < 0.6 ||
count_stable == 0)
regions[k].type = HIGH_VAR_REGION;
}
get_region_stats(stats, regions, *num_regions);
// It is possible for blending to result in a "dip" in intra error (first
// decrease then increase). Therefore we need to find the dip and combine the
// two regions.
k = 1;
while (k < *num_regions) {
if (k < *num_regions - 1 && regions[k].type == HIGH_VAR_REGION) {
// Check if this short high variance regions is actually in the middle of
// a blending region.
if (regions[k - 1].type == BLENDING_REGION &&
regions[k + 1].type == BLENDING_REGION &&
regions[k].last - regions[k].start < 3) {
int prev_dir = (stats[regions[k - 1].last].intra_error -
stats[regions[k - 1].last - 1].intra_error) > 0
? 1
: -1;
int next_dir = (stats[regions[k + 1].last].intra_error -
stats[regions[k + 1].last - 1].intra_error) > 0
? 1
: -1;
if (prev_dir < 0 && next_dir > 0) {
// This is possibly a mid region of blending. Check the ratios
double ratio_thres = AOMMIN(regions[k - 1].avg_sr_fr_ratio,
regions[k + 1].avg_sr_fr_ratio) *
0.95;
if (regions[k].avg_sr_fr_ratio > ratio_thres) {
--> --------------------
--> maximum size reached
--> --------------------
