Use propagated error to determine frame qp in 2-pass tpl
Note the arf rdmult for first-pass tpl is changed back to normal
rdmult, since it is found that using rdmult corresponding to qp=5
boosts the frames too much.
This CL provides considerable gains:
hd2 (%):
avgpsnr ovrpsnr ssim vmaf
-2.360 -2.422 -3.401 -1.453
qmode (%):
avgpsnr ovrpsnr ssim vmaf
-3.362 -3.392 -3.698 -1.198
Change-Id: If5379cb1fbc636c7f41e815d56f5df2ffb4dd398
diff --git a/av1/qmode_rc/ratectrl_qmode.cc b/av1/qmode_rc/ratectrl_qmode.cc
index 2c7f101..5bd9277 100644
--- a/av1/qmode_rc/ratectrl_qmode.cc
+++ b/av1/qmode_rc/ratectrl_qmode.cc
@@ -1458,6 +1458,50 @@
return tpl_gop_dep_stats;
}
+static double GetFrameImportanceTwoPassTpl(
+ const TplFrameDepStats &frame_dep_stats, int frame_width,
+ int frame_height) {
+ const int sb_size = 64;
+ const int num_unit_per_sb = sb_size / frame_dep_stats.unit_size;
+ const int sb_rows = (frame_height + sb_size - 1) / sb_size;
+ const int sb_cols = (frame_width + sb_size - 1) / sb_size;
+ const int unit_rows = (frame_height + frame_dep_stats.unit_size - 1) /
+ frame_dep_stats.unit_size;
+ const int unit_cols =
+ (frame_width + frame_dep_stats.unit_size - 1) / frame_dep_stats.unit_size;
+ std::vector<uint8_t> superblock_q_indices;
+
+ // Cumulate frame level stats
+ double cum_inter_cost = 0;
+ double cum_rdcost_diff = 0;
+ for (int sb_row = 0; sb_row < sb_rows; ++sb_row) {
+ for (int sb_col = 0; sb_col < sb_cols; ++sb_col) {
+ const int unit_row_start = sb_row * num_unit_per_sb;
+ const int unit_row_end =
+ std::min((sb_row + 1) * num_unit_per_sb, unit_rows);
+ const int unit_col_start = sb_col * num_unit_per_sb;
+ const int unit_col_end =
+ std::min((sb_col + 1) * num_unit_per_sb, unit_cols);
+ // A simplified version of av1_get_q_for_deltaq_objective()
+ for (int unit_row = unit_row_start; unit_row < unit_row_end; ++unit_row) {
+ for (int unit_col = unit_col_start; unit_col < unit_col_end;
+ ++unit_col) {
+ const TplUnitDepStats &unit_dep_stats =
+ frame_dep_stats.unit_stats[unit_row][unit_col];
+ const TplUnitDepStats &alt_unit_dep_stats =
+ frame_dep_stats.alt_unit_stats[unit_row][unit_col];
+ cum_inter_cost += unit_dep_stats.inter_cost;
+ cum_rdcost_diff += std::max(unit_dep_stats.propagation_cost -
+ alt_unit_dep_stats.propagation_cost,
+ 0.0);
+ }
+ }
+ }
+ }
+ cum_inter_cost = std::max(cum_inter_cost, 1.0);
+ return (cum_rdcost_diff + cum_inter_cost) / cum_inter_cost;
+}
+
static std::vector<uint8_t> SetupDeltaQ(const TplFrameDepStats &frame_dep_stats,
int frame_width, int frame_height,
int base_qindex,
@@ -1475,39 +1519,6 @@
(frame_width + frame_dep_stats.unit_size - 1) / frame_dep_stats.unit_size;
std::vector<uint8_t> superblock_q_indices;
- if (use_twopass_data) {
- // Cumulate frame level stats
- double cum_inter_cost = 0;
- double cum_rdcost_diff = 0;
- for (int sb_row = 0; sb_row < sb_rows; ++sb_row) {
- for (int sb_col = 0; sb_col < sb_cols; ++sb_col) {
- const int unit_row_start = sb_row * num_unit_per_sb;
- const int unit_row_end =
- std::min((sb_row + 1) * num_unit_per_sb, unit_rows);
- const int unit_col_start = sb_col * num_unit_per_sb;
- const int unit_col_end =
- std::min((sb_col + 1) * num_unit_per_sb, unit_cols);
- // A simplified version of av1_get_q_for_deltaq_objective()
- for (int unit_row = unit_row_start; unit_row < unit_row_end;
- ++unit_row) {
- for (int unit_col = unit_col_start; unit_col < unit_col_end;
- ++unit_col) {
- const TplUnitDepStats &unit_dep_stats =
- frame_dep_stats.unit_stats[unit_row][unit_col];
- const TplUnitDepStats &alt_unit_dep_stats =
- frame_dep_stats.alt_unit_stats[unit_row][unit_col];
- cum_inter_cost += unit_dep_stats.inter_cost;
- cum_rdcost_diff += std::max(unit_dep_stats.propagation_cost -
- alt_unit_dep_stats.propagation_cost,
- 0.0);
- }
- }
- }
- }
- cum_inter_cost = std::max(cum_inter_cost, 1.0);
- frame_importance = (cum_rdcost_diff + cum_inter_cost) / cum_inter_cost;
- }
-
// Calculate delta_q offset for each superblock.
for (int sb_row = 0; sb_row < sb_rows; ++sb_row) {
for (int sb_col = 0; sb_col < sb_cols; ++sb_col) {
@@ -1748,9 +1759,11 @@
if (gop_frame.update_type == GopFrameType::kRegularGolden ||
gop_frame.update_type == GopFrameType::kRegularKey ||
gop_frame.update_type == GopFrameType::kRegularArf) {
- if (rc_param_.tpl_pass_index) param.q_index = kSecondTplPassQp;
- param.rdmult = av1_compute_rd_mult_based_on_qindex(
- AOM_BITS_8, ARF_UPDATE, kSecondTplPassQp);
+ if (rc_param_.tpl_pass_index) {
+ param.q_index = kSecondTplPassQp;
+ param.rdmult = av1_compute_rd_mult_based_on_qindex(
+ AOM_BITS_8, ARF_UPDATE, kSecondTplPassQp);
+ }
}
}
gop_encode_info.param_list.push_back(param);
@@ -2011,33 +2024,6 @@
const int active_worst_quality = base_q_index;
int active_best_quality = base_q_index;
- double base_rdcost = 1.0; // baseline total rdcost
- double hqr_rdcost = 0; // high quality reference total rdcost
- double arf_rdcost_high = 1.0;
-
- bool kf_arf_seen = false;
-
- for (int i = 0; i < frame_count; ++i) {
- FrameEncodeParameters param;
- const GopFrame &gop_frame = gop_struct.gop_frame_list[i];
- const TplFrameDepStats &frame_dep_stats =
- gop_dep_stats->frame_dep_stats_list[i];
- if (gop_frame.update_type == GopFrameType::kRegularGolden ||
- gop_frame.update_type == GopFrameType::kRegularKey ||
- gop_frame.update_type == GopFrameType::kRegularArf) {
- if (!kf_arf_seen) {
- arf_rdcost_high += frame_dep_stats.rdcost;
- }
- kf_arf_seen = 1;
- } else {
- base_rdcost += frame_dep_stats.rdcost;
- hqr_rdcost += frame_dep_stats.alt_rdcost;
- }
- }
-
- double tp_frame_importance =
- 1.0 + fabs((base_rdcost - hqr_rdcost) / arf_rdcost_high);
-
for (int i = 0; i < frame_count; ++i) {
FrameEncodeParameters param;
const GopFrame &gop_frame = gop_struct.gop_frame_list[i];
@@ -2051,18 +2037,21 @@
gop_frame.update_type == GopFrameType::kRegularArf) {
const TplFrameDepStats &frame_dep_stats =
gop_dep_stats->frame_dep_stats_list[i];
- const double cost_without_propagation =
- TplFrameDepStatsAccumulateIntraCost(frame_dep_stats);
- const double cost_with_propagation =
- TplFrameDepStatsAccumulate(frame_dep_stats);
- double frame_importance =
- cost_with_propagation / cost_without_propagation;
+ double frame_importance;
// TODO(jingning): Temporarily make the switch between single and
// two TPL passes depending on the availability. This part of code
// needs further modifications to support SB level calculation.
- if (rc_param_.tpl_pass_count == TplPassCount::kTwoTplPasses)
- frame_importance = tp_frame_importance;
+ if (rc_param_.tpl_pass_count == TplPassCount::kTwoTplPasses) {
+ frame_importance = GetFrameImportanceTwoPassTpl(
+ frame_dep_stats, rc_param_.frame_width, rc_param_.frame_height);
+ } else {
+ const double cost_without_propagation =
+ TplFrameDepStatsAccumulateIntraCost(frame_dep_stats);
+ const double cost_with_propagation =
+ TplFrameDepStatsAccumulate(frame_dep_stats);
+ frame_importance = cost_with_propagation / cost_without_propagation;
+ }
// Imitate the behavior of av1_tpl_get_qstep_ratio()
const double qstep_ratio = sqrt(1 / frame_importance);
