rtc: Early termination of NEWMV mode evaluation
As part of this CL, introduced a speed feature to early
terminate NEWMV mode evaluation based on sse of best mode
so far.
Instruction Count BD-Rate Loss(%)
cpu-used Reduction(%) avg.psnr ovr.psnr ssim
5 1.652 -0.0222 -0.0280 -0.0763
6 1.311 0.0274 0.0276 0.0162
STATS_CHANGED for rt speed 5 and 6
Change-Id: Ia5ee51329a2addb2676dc03c75b3644cbfbb4904
diff --git a/av1/encoder/interp_search.h b/av1/encoder/interp_search.h
index 902b699..e25d44d 100644
--- a/av1/encoder/interp_search.h
+++ b/av1/encoder/interp_search.h
@@ -149,6 +149,11 @@
* MACROBLOCK::pred_sse due to different interpolation filter used.
*/
unsigned int best_single_sse_in_refs[REF_FRAMES];
+ /*!
+ * Holds the sse of best mode so far in the mode evaluation process. This is
+ * used in intermediate termination of NEWMV mode evaluation.
+ */
+ unsigned int best_pred_sse;
} HandleInterModeArgs;
/*!\cond */
diff --git a/av1/encoder/rdopt.c b/av1/encoder/rdopt.c
index 0da6596..c346c92 100644
--- a/av1/encoder/rdopt.c
+++ b/av1/encoder/rdopt.c
@@ -2757,6 +2757,15 @@
if (this_sse < args->best_single_sse_in_refs[ref]) {
args->best_single_sse_in_refs[ref] = this_sse;
}
+
+ if (cpi->sf.rt_sf.skip_newmv_mode_based_on_sse) {
+ const double scale_factor[11] = { 0.7, 0.7, 0.7, 0.7, 0.7, 0.8,
+ 0.8, 0.9, 0.9, 0.9, 0.9 };
+ assert(num_pels_log2_lookup[bsize] >= 4);
+ if (args->best_pred_sse <
+ scale_factor[num_pels_log2_lookup[bsize] - 4] * this_sse)
+ continue;
+ }
}
rd_stats->rate += rate_mv;
@@ -5399,7 +5408,8 @@
-1,
-1,
{ 0 },
- { 0 } };
+ { 0 },
+ UINT_MAX };
for (i = 0; i < MODE_CTX_REF_FRAMES; ++i) args.cmp_mode[i] = -1;
// Indicates the appropriate number of simple translation winner modes for
// exhaustive motion mode evaluation
@@ -5618,6 +5628,7 @@
args.single_newmv_valid = search_state.single_newmv_valid;
args.single_comp_cost = real_compmode_cost;
args.ref_frame_cost = ref_frame_cost;
+ args.best_pred_sse = search_state.best_pred_sse;
int64_t skip_rd[2] = { search_state.best_skip_rd[0],
search_state.best_skip_rd[1] };
diff --git a/av1/encoder/speed_features.c b/av1/encoder/speed_features.c
index 9d5b4de..4b2f8e7 100644
--- a/av1/encoder/speed_features.c
+++ b/av1/encoder/speed_features.c
@@ -1415,6 +1415,7 @@
cm->width * cm->height > 640 * 480)
sf->rt_sf.use_temporal_noise_estimate = 1;
sf->rt_sf.skip_tx_no_split_var_based_partition = 1;
+ sf->rt_sf.skip_newmv_mode_based_on_sse = 1;
// For SVC: use better mv search on base temporal layers, and only
// on base spatial layer if highest resolution is above 640x360.
@@ -1840,6 +1841,7 @@
rt_sf->skip_cdef_sb = 0;
rt_sf->force_large_partition_blocks_intra = 0;
rt_sf->skip_tx_no_split_var_based_partition = 0;
+ rt_sf->skip_newmv_mode_based_on_sse = 0;
}
void av1_set_speed_features_framesize_dependent(AV1_COMP *cpi, int speed) {
diff --git a/av1/encoder/speed_features.h b/av1/encoder/speed_features.h
index 954c512..77a2d01 100644
--- a/av1/encoder/speed_features.h
+++ b/av1/encoder/speed_features.h
@@ -1305,6 +1305,10 @@
// Skip evaluation of no split in tx size selection for merge partition
int skip_tx_no_split_var_based_partition;
+
+ // Intermediate termination of newMV mode evaluation based on so far best mode
+ // sse
+ int skip_newmv_mode_based_on_sse;
} REAL_TIME_SPEED_FEATURES;
/*!\endcond */
