Use source buffer from last gop for tpl
This should benefit tpl data for ext RC.
It also helps a little bit to libaom's built-in RC. Test results for q mode, speed 1, 48 lag in frames and 4 tiles:
PSNR1411 Vmaf-neg uvq-1.5
hdres2: mean: -0.264% -0.185% -0.622%
median: -0.309% -0.141% -0.609%
shorts_720p: mean: -0.098% -0.092% 13.157% (non-monotonic outliers)
median: -0.092% -0.114% -0.436%
STATS_CHANGED
Change-Id: I92c8c9e98598351e6d647f8460a1e99acbd3058d
diff --git a/av1/encoder/encoder.c b/av1/encoder/encoder.c
index ed4cdb4..aca5706 100644
--- a/av1/encoder/encoder.c
+++ b/av1/encoder/encoder.c
@@ -1750,6 +1750,8 @@
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]);
+ aom_free_frame_buffer(&tpl_data->prev_gop_arf_src);
+ tpl_data->prev_gop_arf_disp_order = -1;
tpl_data->tpl_stats_pool[frame] = NULL;
}
@@ -4543,6 +4545,46 @@
cpi->ppi->gf_group.layer_depth[cpi->gf_frame_index],
current_frame->display_order_hint, cpi->ppi->gf_group.max_layer_depth);
+ const GF_GROUP *gf_group = &cpi->ppi->gf_group;
+ // Check if this is the last frame in the gop. If so, make a copy of the
+ // source for TPL.
+ if (gf_group->update_type[cpi->gf_frame_index] != OVERLAY_UPDATE &&
+ gf_group->update_type[cpi->gf_frame_index] != INTNL_OVERLAY_UPDATE) {
+ int is_last = 1;
+ for (int i = 0; i < gf_group->size; ++i) {
+ if (gf_group->display_idx[i] >
+ (int64_t)current_frame->display_order_hint) {
+ is_last = 0;
+ break;
+ }
+ }
+ if (is_last) {
+ cpi->ppi->tpl_data.prev_gop_arf_disp_order = -1;
+ const AV1EncoderConfig *const oxcf = &cpi->oxcf;
+ int ret = aom_realloc_frame_buffer(
+ &cpi->ppi->tpl_data.prev_gop_arf_src, oxcf->frm_dim_cfg.width,
+ oxcf->frm_dim_cfg.height, cm->seq_params->subsampling_x,
+ cm->seq_params->subsampling_y, cm->seq_params->use_highbitdepth,
+ cpi->oxcf.border_in_pixels, cm->features.byte_alignment, NULL, NULL,
+ NULL, cpi->alloc_pyramid, 0);
+ if (ret)
+ aom_internal_error(cm->error, AOM_CODEC_MEM_ERROR,
+ "Failed to allocate tpl prev_gop_arf_src buf.");
+
+ // Currently it is not supported if source/refernece is resized.
+ if (cpi->source->y_width == cpi->ppi->tpl_data.prev_gop_arf_src.y_width &&
+ cpi->source->y_height ==
+ cpi->ppi->tpl_data.prev_gop_arf_src.y_height) {
+ // Copy the content from source to this buffer for next gop.
+ aom_yv12_copy_frame(cpi->source, &cpi->ppi->tpl_data.prev_gop_arf_src,
+ av1_num_planes(cm));
+
+ cpi->ppi->tpl_data.prev_gop_arf_disp_order =
+ current_frame->display_order_hint;
+ }
+ }
+ }
+
if (is_stat_generation_stage(cpi)) {
#if !CONFIG_REALTIME_ONLY
if (cpi->oxcf.q_cfg.use_fixed_qp_offsets)
diff --git a/av1/encoder/pass2_strategy.c b/av1/encoder/pass2_strategy.c
index 9cd5225..3c57d48 100644
--- a/av1/encoder/pass2_strategy.c
+++ b/av1/encoder/pass2_strategy.c
@@ -3837,6 +3837,8 @@
frame_params->frame_type = KEY_FRAME;
find_next_key_frame(cpi, &this_frame);
this_frame = this_frame_copy;
+ // Mark prev gop arf source as unusable
+ cpi->ppi->tpl_data.prev_gop_arf_disp_order = -1;
}
if (rc->frames_to_fwd_kf <= 0)
diff --git a/av1/encoder/tpl_model.c b/av1/encoder/tpl_model.c
index fca14fd..b299930 100644
--- a/av1/encoder/tpl_model.c
+++ b/av1/encoder/tpl_model.c
@@ -200,6 +200,16 @@
aom_internal_error(&ppi->error, AOM_CODEC_MEM_ERROR,
"Failed to allocate frame buffer");
}
+
+ if (aom_alloc_frame_buffer(
+ &tpl_data->prev_gop_arf_src, width, height, seq_params->subsampling_x,
+ seq_params->subsampling_y, seq_params->use_highbitdepth,
+ tpl_data->border_in_pixels, byte_alignment, false,
+ alloc_y_plane_only))
+ aom_internal_error(&ppi->error, AOM_CODEC_MEM_ERROR,
+ "Failed to allocate prev gop arf buffer");
+
+ tpl_data->prev_gop_arf_disp_order = -1;
}
static inline int32_t tpl_get_satd_cost(BitDepthInfo bd_info, int16_t *src_diff,
@@ -1591,7 +1601,12 @@
tpl_data->tpl_frame[-i - 1].rec_picture = NULL;
tpl_data->tpl_frame[-i - 1].frame_display_index = 0;
} else {
- tpl_data->tpl_frame[-i - 1].gf_picture = &cm->ref_frame_map[i]->buf;
+ if (cm->ref_frame_map[i]->display_order_hint ==
+ tpl_data->prev_gop_arf_disp_order) {
+ tpl_data->tpl_frame[-i - 1].gf_picture = &tpl_data->prev_gop_arf_src;
+ } else {
+ tpl_data->tpl_frame[-i - 1].gf_picture = &cm->ref_frame_map[i]->buf;
+ }
tpl_data->tpl_frame[-i - 1].rec_picture = &cm->ref_frame_map[i]->buf;
tpl_data->tpl_frame[-i - 1].frame_display_index =
cm->ref_frame_map[i]->display_order_hint;
diff --git a/av1/encoder/tpl_model.h b/av1/encoder/tpl_model.h
index 264bc09..8fba0cf 100644
--- a/av1/encoder/tpl_model.h
+++ b/av1/encoder/tpl_model.h
@@ -237,6 +237,16 @@
const YV12_BUFFER_CONFIG *ref_frame[INTER_REFS_PER_FRAME];
/*!
+ * The buffer for the past gop's last frame's src.
+ */
+ YV12_BUFFER_CONFIG prev_gop_arf_src;
+
+ /*!
+ * Display order of the past gop's last frame.
+ */
+ int64_t prev_gop_arf_disp_order;
+
+ /*!
* Parameters related to synchronization for top-right dependency in row based
* multi-threading of tpl
*/
