ref: 42e5073f94ffa15cc5bd6284b9b1daed5c093cf1
parent: 629279a45c1be8b5daa22b39e9c8f85503831e24
author: paulwilkins <paulwilkins@google.com>
date: Wed Apr 5 06:35:54 EDT 2017
Experiment. Store first pass errors as per MB values. Most existing first pass stats are stored in a form normalized to a macro-block scale. However the error scores for intra / inter etc were stored as frame level values but mainly used as MB level values. This change fixes that. Normalized per MB values make comparisons between different formats easier and in any case this is usually what is wanted. An change in results should be limited to slight differences in rounding. *** Change after patch 8 +2 requiring new approval. Final pre-submit testing showed one 4K clip with above expected change. Investigation showed this was due to a value used to test for ultra low intra complexity in key frame detection. This was a per frame not per MB value but also did not scale with frame size. Replacement with a small per MB value (based on original per frame value and cif frame size) resolved the KF detection problem. Also converted kf_group_error_left to a double in line with other error values to reduce rounding problems in KF group bit allocation All clips and sets now show nominal (or 0) change as expected. Change-Id: Ic2d57980398c99ade2b7380e3e6ca6b32186901f
--- a/vp9/encoder/vp9_firstpass.c
+++ b/vp9/encoder/vp9_firstpass.c
@@ -707,9 +707,14 @@
fps->frame = cm->current_video_frame;
fps->spatial_layer_id = cpi->svc.spatial_layer_id;
- fps->coded_error = (double)(fp_acc_data->coded_error >> 8) + min_err;
- fps->sr_coded_error = (double)(fp_acc_data->sr_coded_error >> 8) + min_err;
- fps->intra_error = (double)(fp_acc_data->intra_error >> 8) + min_err;
+
+ fps->coded_error =
+ ((double)(fp_acc_data->coded_error >> 8) + min_err) / num_mbs;
+ fps->sr_coded_error =
+ ((double)(fp_acc_data->sr_coded_error >> 8) + min_err) / num_mbs;
+ fps->intra_error =
+ ((double)(fp_acc_data->intra_error >> 8) + min_err) / num_mbs;
+
fps->frame_noise_energy =
(double)(fp_acc_data->frame_noise_energy) / (double)num_mbs;
fps->count = 1.0;
@@ -1565,8 +1570,9 @@
const int num_mbs = (cpi->oxcf.resize_mode != RESIZE_NONE)
? cpi->initial_mbs
: cpi->common.MBs;
- const int active_mbs = VPXMAX(1, num_mbs - (int)(num_mbs * inactive_zone));
- const double av_err_per_mb = section_err / active_mbs;
+ const double active_pct = VPXMAX(0.01, 1.0 - inactive_zone);
+ const int active_mbs = (int)VPXMAX(1, (double)num_mbs * active_pct);
+ const double av_err_per_mb = section_err / active_pct;
const double speed_term = 1.0 + 0.04 * oxcf->speed;
double last_group_rate_err;
const int target_norm_bits_per_mb =
@@ -1727,9 +1733,7 @@
static double get_sr_decay_rate(const VP9_COMP *cpi,
const FIRSTPASS_STATS *frame) {- const int num_mbs = (cpi->oxcf.resize_mode != RESIZE_NONE) ? cpi->initial_mbs
- : cpi->common.MBs;
- double sr_diff = (frame->sr_coded_error - frame->coded_error) / num_mbs;
+ double sr_diff = (frame->sr_coded_error - frame->coded_error);
double sr_decay = 1.0;
double modified_pct_inter;
double modified_pcnt_intra;
@@ -1738,7 +1742,7 @@
(cpi->initial_height + cpi->initial_width));
modified_pct_inter = frame->pcnt_inter;
- if (((frame->coded_error / num_mbs) > LOW_CODED_ERR_PER_MB) &&
+ 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 =
@@ -1860,20 +1864,16 @@
const double lq = vp9_convert_qindex_to_q(
cpi->rc.avg_frame_qindex[INTER_FRAME], cpi->common.bit_depth);
const double boost_q_correction = VPXMIN((0.5 + (lq * 0.015)), 1.5);
- int num_mbs = (cpi->oxcf.resize_mode != RESIZE_NONE) ? cpi->initial_mbs
- : cpi->common.MBs;
+ const double active_area = calculate_active_area(cpi, this_frame);
- // Correct for any inactive region in the image
- num_mbs = (int)VPXMAX(1, num_mbs * calculate_active_area(cpi, this_frame));
-
// Underlying boost factor is based on inter error ratio.
- frame_boost = (BASELINE_ERR_PER_MB * num_mbs) /
+ frame_boost = (BASELINE_ERR_PER_MB * active_area) /
DOUBLE_DIVIDE_CHECK(this_frame->coded_error + *sr_accumulator);
// 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) / 1;
+ *sr_accumulator += (this_frame->sr_coded_error - this_frame->coded_error);
*sr_accumulator = VPXMAX(0.0, *sr_accumulator);
// Small adjustment for cases where there is a zoom out
@@ -1896,20 +1896,16 @@
const double lq = vp9_convert_qindex_to_q(
cpi->rc.avg_frame_qindex[INTER_FRAME], cpi->common.bit_depth);
const double boost_q_correction = VPXMIN((0.50 + (lq * 0.015)), 2.00);
- int num_mbs = (cpi->oxcf.resize_mode != RESIZE_NONE) ? cpi->initial_mbs
- : cpi->common.MBs;
+ const double active_area = calculate_active_area(cpi, this_frame);
- // Correct for any inactive region in the image
- num_mbs = (int)VPXMAX(1, num_mbs * calculate_active_area(cpi, this_frame));
-
// Underlying boost factor is based on inter error ratio.
- frame_boost = (KF_BASELINE_ERR_PER_MB * num_mbs) /
+ frame_boost = (KF_BASELINE_ERR_PER_MB * active_area) /
DOUBLE_DIVIDE_CHECK(this_frame->coded_error + *sr_accumulator);
// 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) / 1;
+ *sr_accumulator += (this_frame->sr_coded_error - this_frame->coded_error);
*sr_accumulator = VPXMAX(0.0, *sr_accumulator);
// Small adjustment for cases where there is a zoom out
@@ -2042,7 +2038,7 @@
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)) {+ if ((twopass->kf_group_bits > 0) && (twopass->kf_group_error_left > 0.0)) {total_group_bits = (int64_t)(twopass->kf_group_bits *
(gf_group_err / twopass->kf_group_error_left));
} else {@@ -2566,7 +2562,7 @@
gf_group_bits);
// Adjust KF group bits and error remaining.
- twopass->kf_group_error_left -= (int64_t)gf_group_err;
+ twopass->kf_group_error_left -= gf_group_err;
// Allocate bits to each of the frames in the GF group.
allocate_gf_group_bits(cpi, gf_group_bits, gf_arf_bits);
@@ -2616,6 +2612,9 @@
#define II_IMPROVEMENT_THRESHOLD 3.5
#define KF_II_MAX 128.0
#define II_FACTOR 12.5
+// Test for very low intra complexity which could cause false key frames
+#define V_LOW_INTRA 0.5
+
static int test_candidate_kf(TWO_PASS *twopass,
const FIRSTPASS_STATS *last_frame,
const FIRSTPASS_STATS *this_frame,
@@ -2674,7 +2673,7 @@
0.20) &&
(next_iiratio < 3.0)) ||
((boost_score - old_boost_score) < 3.0) ||
- (local_next_frame.intra_error < 200)) {+ (local_next_frame.intra_error < V_LOW_INTRA)) {break;
}
@@ -2750,8 +2749,8 @@
rc->frames_to_key = 1;
- twopass->kf_group_bits = 0; // Total bits available to kf group
- twopass->kf_group_error_left = 0; // Group modified error score.
+ twopass->kf_group_bits = 0; // Total bits available to kf group
+ twopass->kf_group_error_left = 0.0; // Group modified error score.
kf_mod_err = calculate_modified_err(cpi, twopass, oxcf, this_frame);
@@ -2935,7 +2934,7 @@
gf_group->rf_level[0] = KF_STD;
// Note the total error score of the kf group minus the key frame itself.
- twopass->kf_group_error_left = (int)(kf_group_err - kf_mod_err);
+ twopass->kf_group_error_left = (kf_group_err - kf_mod_err);
// Adjust the count of total modified error left.
// The count of bits left is adjusted elsewhere based on real coded frame
@@ -3172,16 +3171,10 @@
target_rate = gf_group->bit_allocation[gf_group->index];
rc->base_frame_target = target_rate;
- {- const int num_mbs = (cpi->oxcf.resize_mode != RESIZE_NONE)
- ? cpi->initial_mbs
- : cpi->common.MBs;
- // The multiplication by 256 reverses a scaling factor of (>> 8)
- // applied when combining MB error values for the frame.
- twopass->mb_av_energy =
- log(((this_frame.intra_error * 256.0) / num_mbs) + 1.0);
- twopass->mb_smooth_pct = this_frame.intra_smooth_pct;
- }
+ // The multiplication by 256 reverses a scaling factor of (>> 8)
+ // applied when combining MB error values for the frame.
+ twopass->mb_av_energy = log((this_frame.intra_error * 256.0) + 1.0);
+ twopass->mb_smooth_pct = this_frame.intra_smooth_pct;
// Update the total stats remaining structure.
subtract_stats(&twopass->total_left_stats, &this_frame);
--- a/vp9/encoder/vp9_firstpass.h
+++ b/vp9/encoder/vp9_firstpass.h
@@ -159,7 +159,7 @@
int64_t kf_group_bits;
// Error score of frames still to be coded in kf group
- int64_t kf_group_error_left;
+ double kf_group_error_left;
double bpm_factor;
int rolling_arf_group_target_bits;