RT: Pick cdef filter strengths based on q for screen content
Train a new model to predict cdef filter strengths for screen
content based on the frame q.
Add a qindex threshold in the speed feature to achieve a better
tradeoff for different speed settings.
Performance:
speed_10 ovr_psnr avg_psnr ssim vmaf
rtc_screen -3.485% -1.362% -4.240% 3.724%
The average slowdown is 2.4%.
For speed 9, a lower threshold is used to retain higher coding gain.
Change-Id: Icf0690429e0739b9c72be90630cb6f4cdb7a670c
diff --git a/av1/encoder/encoder.c b/av1/encoder/encoder.c
index 32d0496..2542076 100644
--- a/av1/encoder/encoder.c
+++ b/av1/encoder/encoder.c
@@ -2144,11 +2144,16 @@
start_timing(cpi, cdef_time);
#endif
const int num_workers = cpi->mt_info.num_mod_workers[MOD_CDEF];
+ const int use_screen_content_model =
+ cm->quant_params.base_qindex >
+ AOMMAX(cpi->sf.rt_sf.screen_content_cdef_filter_qindex_thresh,
+ cpi->rc.best_quality + 5) &&
+ cpi->oxcf.tune_cfg.content == AOM_CONTENT_SCREEN;
// Find CDEF parameters
av1_cdef_search(&cpi->mt_info, &cm->cur_frame->buf, cpi->source, cm, xd,
cpi->sf.lpf_sf.cdef_pick_method, cpi->td.mb.rdmult,
cpi->sf.rt_sf.skip_cdef_sb, cpi->rc.frames_since_key,
- cpi->oxcf.tool_cfg.cdef_control,
+ cpi->oxcf.tool_cfg.cdef_control, use_screen_content_model,
cpi->svc.non_reference_frame);
// Apply the filter
diff --git a/av1/encoder/pickcdef.c b/av1/encoder/pickcdef.c
index 557c9ee..adbb07c 100644
--- a/av1/encoder/pickcdef.c
+++ b/av1/encoder/pickcdef.c
@@ -508,7 +508,7 @@
}
static void pick_cdef_from_qp(AV1_COMMON *const cm, int skip_cdef,
- int frames_since_key) {
+ int frames_since_key, int is_screen_content) {
const int bd = cm->seq_params->bit_depth;
const int q =
av1_ac_quant_QTX(cm->quant_params.base_qindex, 0, bd) >> (bd - 8);
@@ -527,32 +527,47 @@
int predicted_y_f2 = 0;
int predicted_uv_f1 = 0;
int predicted_uv_f2 = 0;
- if (!frame_is_intra_only(cm)) {
- predicted_y_f1 = clamp((int)roundf(q * q * -0.0000023593946f +
- q * 0.0068615186f + 0.02709886f),
- 0, 15);
- predicted_y_f2 = clamp((int)roundf(q * q * -0.00000057629734f +
- q * 0.0013993345f + 0.03831067f),
- 0, 3);
- predicted_uv_f1 = clamp((int)roundf(q * q * -0.0000007095069f +
- q * 0.0034628846f + 0.00887099f),
- 0, 15);
- predicted_uv_f2 = clamp((int)roundf(q * q * 0.00000023874085f +
- q * 0.00028223585f + 0.05576307f),
- 0, 3);
+ if (is_screen_content) {
+ predicted_y_f1 =
+ (int)(5.88217781e-06 * q * q + 6.10391455e-03 * q + 9.95043102e-02);
+ predicted_y_f2 =
+ (int)(-7.79934857e-06 * q * q + 6.58957830e-03 * q + 8.81045025e-01);
+ predicted_uv_f1 =
+ (int)(-6.79500136e-06 * q * q + 1.02695586e-02 * q + 1.36126802e-01);
+ predicted_uv_f2 =
+ (int)(-9.99613695e-08 * q * q - 1.79361339e-05 * q + 1.17022324e+0);
+ predicted_y_f1 = clamp(predicted_y_f1, 0, 15);
+ predicted_y_f2 = clamp(predicted_y_f2, 0, 3);
+ predicted_uv_f1 = clamp(predicted_uv_f1, 0, 15);
+ predicted_uv_f2 = clamp(predicted_uv_f2, 0, 3);
} else {
- predicted_y_f1 = clamp(
- (int)roundf(q * q * 0.0000033731974f + q * 0.008070594f + 0.0187634f),
- 0, 15);
- predicted_y_f2 = clamp(
- (int)roundf(q * q * 0.0000029167343f + q * 0.0027798624f + 0.0079405f),
- 0, 3);
- predicted_uv_f1 = clamp(
- (int)roundf(q * q * -0.0000130790995f + q * 0.012892405f - 0.00748388f),
- 0, 15);
- predicted_uv_f2 = clamp((int)roundf(q * q * 0.0000032651783f +
- q * 0.00035520183f + 0.00228092f),
- 0, 3);
+ if (!frame_is_intra_only(cm)) {
+ predicted_y_f1 = clamp((int)roundf(q * q * -0.0000023593946f +
+ q * 0.0068615186f + 0.02709886f),
+ 0, 15);
+ predicted_y_f2 = clamp((int)roundf(q * q * -0.00000057629734f +
+ q * 0.0013993345f + 0.03831067f),
+ 0, 3);
+ predicted_uv_f1 = clamp((int)roundf(q * q * -0.0000007095069f +
+ q * 0.0034628846f + 0.00887099f),
+ 0, 15);
+ predicted_uv_f2 = clamp((int)roundf(q * q * 0.00000023874085f +
+ q * 0.00028223585f + 0.05576307f),
+ 0, 3);
+ } else {
+ predicted_y_f1 = clamp(
+ (int)roundf(q * q * 0.0000033731974f + q * 0.008070594f + 0.0187634f),
+ 0, 15);
+ predicted_y_f2 = clamp((int)roundf(q * q * 0.0000029167343f +
+ q * 0.0027798624f + 0.0079405f),
+ 0, 3);
+ predicted_uv_f1 = clamp((int)roundf(q * q * -0.0000130790995f +
+ q * 0.012892405f - 0.00748388f),
+ 0, 15);
+ predicted_uv_f2 = clamp((int)roundf(q * q * 0.0000032651783f +
+ q * 0.00035520183f + 0.00228092f),
+ 0, 3);
+ }
}
cdef_info->cdef_strengths[0] =
predicted_y_f1 * CDEF_SEC_STRENGTHS + predicted_y_f2;
@@ -584,7 +599,8 @@
const YV12_BUFFER_CONFIG *ref, AV1_COMMON *cm,
MACROBLOCKD *xd, CDEF_PICK_METHOD pick_method, int rdmult,
int skip_cdef_feature, int frames_since_key,
- CDEF_CONTROL cdef_control, int non_reference_frame) {
+ CDEF_CONTROL cdef_control, const int is_screen_content,
+ int non_reference_frame) {
assert(cdef_control != CDEF_NONE);
if (cdef_control == CDEF_REFERENCE && non_reference_frame) {
CdefInfo *const cdef_info = &cm->cdef_info;
@@ -596,7 +612,8 @@
}
if (pick_method == CDEF_PICK_FROM_Q) {
- pick_cdef_from_qp(cm, skip_cdef_feature, frames_since_key);
+ pick_cdef_from_qp(cm, skip_cdef_feature, frames_since_key,
+ is_screen_content);
return;
}
const CommonModeInfoParams *const mi_params = &cm->mi_params;
diff --git a/av1/encoder/pickcdef.h b/av1/encoder/pickcdef.h
index d52cb4b..e070a8a 100644
--- a/av1/encoder/pickcdef.h
+++ b/av1/encoder/pickcdef.h
@@ -228,6 +228,7 @@
* \param[in] skip_cdef_feature Speed feature to skip cdef
* \param[in] frames_since_key Number of frames since key frame
* \param[in] cdef_control Parameter that controls CDEF application
+ * \param[in] is_screen_content Whether it is screen content type
* \param[in] non_reference_frame Indicates if current frame is
* non-reference
*
@@ -247,7 +248,8 @@
const YV12_BUFFER_CONFIG *ref, AV1_COMMON *cm,
MACROBLOCKD *xd, CDEF_PICK_METHOD pick_method, int rdmult,
int skip_cdef_feature, int frames_since_key,
- CDEF_CONTROL cdef_control, int non_reference_frame);
+ CDEF_CONTROL cdef_control, const int is_screen_content,
+ int non_reference_frame);
#ifdef __cplusplus
} // extern "C"
diff --git a/av1/encoder/speed_features.c b/av1/encoder/speed_features.c
index 3b0e151..270f891 100644
--- a/av1/encoder/speed_features.c
+++ b/av1/encoder/speed_features.c
@@ -1666,6 +1666,7 @@
if (speed >= 9) {
sf->rt_sf.sse_early_term_inter_search = EARLY_TERM_IDX_3;
sf->lpf_sf.cdef_pick_method = CDEF_PICK_FROM_Q;
+ sf->rt_sf.screen_content_cdef_filter_qindex_thresh = 20;
sf->rt_sf.estimate_motion_for_var_based_partition = 0;
sf->rt_sf.force_large_partition_blocks = 1;
sf->rt_sf.skip_intra_pred = 2;
@@ -1682,6 +1683,7 @@
sf->mv_sf.subpel_search_method = SUBPEL_TREE_PRUNED_MORE;
sf->rt_sf.force_half_pel_block = 1;
sf->rt_sf.reduce_zeromv_mvres = true;
+ sf->rt_sf.screen_content_cdef_filter_qindex_thresh = 80;
}
}
diff --git a/av1/encoder/speed_features.h b/av1/encoder/speed_features.h
index bbf17d5..8bb428f 100644
--- a/av1/encoder/speed_features.h
+++ b/av1/encoder/speed_features.h
@@ -1531,6 +1531,13 @@
// by 8.44% for speed 9 on a typical image dataset with coding performance
// gain of 0.78%.
bool vbp_prune_16x16_split_using_min_max_sub_blk_var;
+
+ // A qindex threshold that determines whether to use qindex based
+ // CDEF filter strength estimation for screen content types.
+ // This speed feature has a substantial gain on coding metrics,
+ // with moderate increased encoding time.
+ // Set to zero to turn off this speed feature.
+ int screen_content_cdef_filter_qindex_thresh;
} REAL_TIME_SPEED_FEATURES;
/*!\endcond */
