Skip motion search w.r.t. third reference in firstpass
The stats based on motion search w.r.t third reference
frame in the first pass is required in the ML model
which predicts if a flat GOP structure (golden-frame
only structure without ALT-REF and Internal-ARFs) is
better. As the ML model is enabled only for constant
quality mode under certain conditions, motion search
w.r.t third reference frame is enabled only for those
cases and disabled for others.
Instruction Count
cpu-used Reduction(%)
3 0.41
4 0.60
5 0.91
6 1.21
Change-Id: I73aff59ed4ce55104fc6cafa6331e40579b6ec42
diff --git a/av1/encoder/encoder.h b/av1/encoder/encoder.h
index 98bcb6e..538cc46 100644
--- a/av1/encoder/encoder.h
+++ b/av1/encoder/encoder.h
@@ -3188,6 +3188,16 @@
return lag_in_frames >= ALT_MIN_LAG && enable_auto_arf;
}
+static AOM_INLINE int can_disable_altref(const GFConfig *gf_cfg) {
+ return is_altref_enabled(gf_cfg->lag_in_frames, gf_cfg->enable_auto_arf) &&
+ (gf_cfg->gf_min_pyr_height == 0);
+}
+
+static AOM_INLINE int use_ml_model_to_decide_flat_gop(
+ const RateControlCfg *rc_cfg) {
+ return (rc_cfg->mode == AOM_Q && rc_cfg->cq_level <= 200);
+}
+
// Check if statistics generation stage
static INLINE int is_stat_generation_stage(const AV1_COMP *const cpi) {
assert(IMPLIES(cpi->compressor_stage == LAP_STAGE,
diff --git a/av1/encoder/firstpass.c b/av1/encoder/firstpass.c
index dd0f8af..52dc514 100644
--- a/av1/encoder/firstpass.c
+++ b/av1/encoder/firstpass.c
@@ -54,6 +54,8 @@
#define NCOUNT_INTRA_THRESH 8192
#define NCOUNT_INTRA_FACTOR 3
+#define INVALID_FP_STATS_TO_PREDICT_FLAT_GOP -1
+
static AOM_INLINE void output_stats(FIRSTPASS_STATS *stats,
struct aom_codec_pkt_list *pktlist) {
struct aom_codec_cx_pkt pkt;
@@ -121,9 +123,11 @@
section->frame_avg_wavelet_energy += frame->frame_avg_wavelet_energy;
section->coded_error += frame->coded_error;
section->sr_coded_error += frame->sr_coded_error;
+ section->tr_coded_error += frame->tr_coded_error;
section->pcnt_inter += frame->pcnt_inter;
section->pcnt_motion += frame->pcnt_motion;
section->pcnt_second_ref += frame->pcnt_second_ref;
+ section->pcnt_third_ref += frame->pcnt_third_ref;
section->pcnt_neutral += frame->pcnt_neutral;
section->intra_skip_pct += frame->intra_skip_pct;
section->inactive_zone_rows += frame->inactive_zone_rows;
@@ -356,6 +360,11 @@
return raw_err_stdev;
}
+static AOM_INLINE int do_third_ref_motion_search(const RateControlCfg *rc_cfg,
+ const GFConfig *gf_cfg) {
+ return use_ml_model_to_decide_flat_gop(rc_cfg) && can_disable_altref(gf_cfg);
+}
+
#define UL_INTRA_THRESH 50
#define INVALID_ROW -1
// Computes and returns the intra pred error of a block.
@@ -682,14 +691,22 @@
// Motion search in 3rd reference frame.
int alt_motion_error = motion_error;
- if (alt_ref_frame != NULL) {
- FULLPEL_MV tmp_mv = kZeroFullMv;
- xd->plane[0].pre[0].buf = alt_ref_frame->y_buffer + alt_ref_frame_yoffset;
- xd->plane[0].pre[0].stride = alt_ref_frame->y_stride;
- alt_motion_error =
- get_prediction_error_bitdepth(is_high_bitdepth, bitdepth, bsize,
- &x->plane[0].src, &xd->plane[0].pre[0]);
- first_pass_motion_search(cpi, x, &kZeroMv, &tmp_mv, &alt_motion_error);
+ // The ML model to predict if a flat structure (golden-frame only structure
+ // without ALT-REF and Internal-ARFs) is better requires stats based on
+ // motion search w.r.t 3rd reference frame in the first pass. As the ML
+ // model is enabled under certain conditions, motion search in 3rd reference
+ // frame is also enabled for those cases.
+ if (do_third_ref_motion_search(&cpi->oxcf.rc_cfg, &cpi->oxcf.gf_cfg)) {
+ if (alt_ref_frame != NULL) {
+ FULLPEL_MV tmp_mv = kZeroFullMv;
+ xd->plane[0].pre[0].buf =
+ alt_ref_frame->y_buffer + alt_ref_frame_yoffset;
+ xd->plane[0].pre[0].stride = alt_ref_frame->y_stride;
+ alt_motion_error = get_prediction_error_bitdepth(
+ is_high_bitdepth, bitdepth, bsize, &x->plane[0].src,
+ &xd->plane[0].pre[0]);
+ first_pass_motion_search(cpi, x, &kZeroMv, &tmp_mv, &alt_motion_error);
+ }
}
if (alt_motion_error < motion_error && alt_motion_error < gf_motion_error &&
alt_motion_error < this_intra_error) {
@@ -851,6 +868,13 @@
// cpi->source_time_stamp.
fps.duration = (double)ts_duration;
+ // Invalidate the stats related to third ref motion search if not valid.
+ // This helps to print a warning in second pass encoding.
+ if (do_third_ref_motion_search(&cpi->oxcf.rc_cfg, &cpi->oxcf.gf_cfg) == 0) {
+ fps.pcnt_third_ref = INVALID_FP_STATS_TO_PREDICT_FLAT_GOP;
+ fps.tr_coded_error = INVALID_FP_STATS_TO_PREDICT_FLAT_GOP;
+ }
+
// We will store the stats inside the persistent twopass struct (and NOT the
// local variable 'fps'), and then cpi->output_pkt_list will point to it.
*this_frame_stats = fps;
diff --git a/av1/encoder/pass2_strategy.c b/av1/encoder/pass2_strategy.c
index f7e03db..9dbdfb2 100644
--- a/av1/encoder/pass2_strategy.c
+++ b/av1/encoder/pass2_strategy.c
@@ -19,6 +19,7 @@
#include <stdint.h>
+#include "common/tools_common.h"
#include "config/aom_config.h"
#include "config/aom_scale_rtcd.h"
@@ -43,6 +44,10 @@
#define DEFAULT_KF_BOOST 2300
#define DEFAULT_GF_BOOST 2000
#define GROUP_ADAPTIVE_MAXQ 1
+
+#define IS_FP_STATS_TO_PREDICT_FLAT_GOP_INVALID(fp_stats) \
+ (((fp_stats)->tr_coded_error < 0) || ((fp_stats)->pcnt_third_ref < 0))
+
static void init_gf_stats(GF_GROUP_STATS *gf_stats);
// Calculate an active area of the image that discounts formatting
@@ -2425,8 +2430,10 @@
p_rc->use_arf_in_this_kf_group && (i < gf_cfg->lag_in_frames) &&
(i >= MIN_GF_INTERVAL);
+ FIRSTPASS_STATS *total_stats = twopass->stats_buf_ctx->total_stats;
// TODO(urvang): Improve and use model for VBR, CQ etc as well.
- if (use_alt_ref && rc_cfg->mode == AOM_Q && rc_cfg->cq_level <= 200) {
+ if (use_alt_ref && use_ml_model_to_decide_flat_gop(rc_cfg) &&
+ !IS_FP_STATS_TO_PREDICT_FLAT_GOP_INVALID(total_stats)) {
aom_clear_system_state();
float features[21];
get_features_from_gf_stats(
@@ -3384,18 +3391,28 @@
CurrentFrame *const current_frame = &cm->current_frame;
RATE_CONTROL *const rc = &cpi->rc;
TWO_PASS *const twopass = &cpi->ppi->twopass;
+ FIRSTPASS_STATS *total_stats = twopass->stats_buf_ctx->total_stats;
+
+ if (current_frame->frame_number == 0) {
+ const GFConfig *const gf_cfg = &cpi->oxcf.gf_cfg;
+ const RateControlCfg *const rc_cfg = &cpi->oxcf.rc_cfg;
+ if (use_ml_model_to_decide_flat_gop(rc_cfg) && can_disable_altref(gf_cfg) &&
+ IS_FP_STATS_TO_PREDICT_FLAT_GOP_INVALID(total_stats))
+ warn(
+ "First pass stats required in the ML model to predict a flat GOP "
+ "structure is invalid. Continuing encoding by disabling the ML "
+ "model.\n");
+ }
if (cpi->oxcf.rc_cfg.mode != AOM_Q && current_frame->frame_number == 0 &&
- cpi->gf_frame_index == 0 &&
- cpi->ppi->twopass.stats_buf_ctx->total_stats &&
+ cpi->gf_frame_index == 0 && total_stats &&
cpi->ppi->twopass.stats_buf_ctx->total_left_stats) {
if (cpi->ppi->lap_enabled) {
/*
* Accumulate total_stats using available limited number of stats,
* and assign it to total_left_stats.
*/
- *cpi->ppi->twopass.stats_buf_ctx->total_left_stats =
- *cpi->ppi->twopass.stats_buf_ctx->total_stats;
+ *cpi->ppi->twopass.stats_buf_ctx->total_left_stats = *total_stats;
}
// Special case code for first frame.
const int section_target_bandwidth = get_section_target_bandwidth(cpi);
