commit ec169b58db068b3dce7746a849266739a5a732b0
author Ranjit Kumar Tulabandu <ranjit.tulabandu@ittiam.com> Mon Aug 27 10:15:01 2018 +0530
committer Yunqing Wang <yunqingwang@google.com> Tue Aug 28 22:50:42 2018 +0000
tree a262188f66c8cc769150aa81985ecea513e650b6
parent 02b2a849daa9e1cd01605aac4d3df7dd5ca7d9db

Refine luma transform rd evaluation

Luma transform rd cost is refined to be precise at block
level

For speed=1 preset, 6.4% encode time reduction is seen for
10 frames of parkrun 720p50 content and 0.01% BD-rate
improvement is seen in AWCY tests.

STATS_CHANGED

Change-Id: Id9e496587e54ff8cbf2522c8d75a17b4fa2b3ea1

av1/encoder/rdopt.c

diff --git a/av1/encoder/rdopt.c b/av1/encoder/rdopt.c
index 0b5f22f..5cf61cd 100644
--- a/av1/encoder/rdopt.c
+++ b/av1/encoder/rdopt.c
@@ -5080,7 +5080,7 @@
                                    TXB_RD_INFO_NODE *rd_info_tree) {
   MACROBLOCKD *const xd = &x->e_mbd;
   int is_cost_valid = 1;
-  int64_t this_rd = 0;
+  int64_t this_rd = 0, skip_rd = 0;
 
   if (ref_best_rd < 0) is_cost_valid = 0;
 
@@ -5111,40 +5111,40 @@
     av1_get_entropy_contexts(bsize, pd, ctxa, ctxl);
     memcpy(tx_above, xd->above_txfm_context, sizeof(TXFM_CONTEXT) * mi_width);
     memcpy(tx_left, xd->left_txfm_context, sizeof(TXFM_CONTEXT) * mi_height);
+    const int skip_ctx = av1_get_skip_context(xd);
+    const int s0 = x->skip_cost[skip_ctx][0];
+    const int s1 = x->skip_cost[skip_ctx][1];
 
+    skip_rd = RDCOST(x->rdmult, s1, 0);
+    this_rd = RDCOST(x->rdmult, s0, 0);
     for (idy = 0; idy < mi_height; idy += bh) {
       for (idx = 0; idx < mi_width; idx += bw) {
+        int64_t best_rd_sofar = (ref_best_rd - (AOMMIN(skip_rd, this_rd)));
         select_tx_block(cpi, x, idy, idx, block, max_tx_size, init_depth,
                         plane_bsize, ctxa, ctxl, tx_above, tx_left,
-                        &pn_rd_stats, ref_best_rd - this_rd, &is_cost_valid,
-                        ftxs_mode, rd_info_tree);
+                        &pn_rd_stats, best_rd_sofar, &is_cost_valid, ftxs_mode,
+                        rd_info_tree);
         if (!is_cost_valid || pn_rd_stats.rate == INT_MAX) {
           av1_invalid_rd_stats(rd_stats);
           return;
         }
         av1_merge_rd_stats(rd_stats, &pn_rd_stats);
-        this_rd +=
-            AOMMIN(RDCOST(x->rdmult, pn_rd_stats.rate, pn_rd_stats.dist),
-                   RDCOST(x->rdmult, pn_rd_stats.zero_rate, pn_rd_stats.sse));
+        skip_rd = RDCOST(x->rdmult, s1, rd_stats->sse);
+        this_rd = RDCOST(x->rdmult, rd_stats->rate + s0, rd_stats->dist);
         block += step;
         if (rd_info_tree != NULL) rd_info_tree += 1;
       }
     }
+    if (skip_rd <= this_rd) {
+      this_rd = skip_rd;
+      rd_stats->rate = 0;
+      rd_stats->dist = rd_stats->sse;
+      rd_stats->skip = 1;
+    } else {
+      rd_stats->skip = 0;
+    }
   }
 
-  const int skip_ctx = av1_get_skip_context(xd);
-  const int s0 = x->skip_cost[skip_ctx][0];
-  const int s1 = x->skip_cost[skip_ctx][1];
-  int64_t skip_rd = RDCOST(x->rdmult, s1, rd_stats->sse);
-  this_rd = RDCOST(x->rdmult, rd_stats->rate + s0, rd_stats->dist);
-  if (skip_rd <= this_rd) {
-    this_rd = skip_rd;
-    rd_stats->rate = 0;
-    rd_stats->dist = rd_stats->sse;
-    rd_stats->skip = 1;
-  } else {
-    rd_stats->skip = 0;
-  }
   if (this_rd > ref_best_rd) is_cost_valid = 0;
 
   if (!is_cost_valid) {