commit 55ba17089951380ec2938038cd88245a8a15922d
author Jingning Han <jingning@google.com> Tue Jan 14 11:43:04 2020 -0800
committer Jingning Han <jingning@google.com> Wed Jan 15 17:01:05 2020 +0000
tree fcc28de2595e0a2fb6273f252eabc9252294189a
parent c2635258e9da5a06b45e3b8831f3b54c48a6308a

Avoid repeated motion search pattern assignment

This also avoids breaking the search pattern in higher speed
settings. Tested in speed 5 on BasketballDrill_832x480_50.y4m, the
ins count drops from 322490 M to 296797 M.

STATS_CHANGED

Change-Id: Id0ba41752ee710c385a58b5085c16e836b2e5de5

av1/encoder/encoder.c

diff --git a/av1/encoder/encoder.c b/av1/encoder/encoder.c
index 9e85e2b..e412218 100644
--- a/av1/encoder/encoder.c
+++ b/av1/encoder/encoder.c
@@ -1045,15 +1045,13 @@
   AV1_COMMON *cm = &cpi->common;
   const SequenceHeader *const seq_params = &cm->seq_params;
   const AV1EncoderConfig *oxcf = &cpi->oxcf;
-  int is_scale = (oxcf->resize_mode || oxcf->superres_mode);
 
   // TODO(agrange) Check if ARF is enabled and skip allocation if not.
   // (yunqing)Here use same border as lookahead buffers.
   if (aom_realloc_frame_buffer(
           &cpi->alt_ref_buffer, oxcf->width, oxcf->height,
           seq_params->subsampling_x, seq_params->subsampling_y,
-          seq_params->use_highbitdepth,
-          is_scale ? oxcf->border_in_pixels : AOM_ENC_LOOKAHEAD_BORDER,
+          seq_params->use_highbitdepth, AOM_ENC_NO_SCALE_BORDER,
           cm->byte_alignment, NULL, NULL, NULL))
     aom_internal_error(&cm->error, AOM_CODEC_MEM_ERROR,
                        "Failed to allocate altref buffer");
@@ -3183,7 +3181,7 @@
             &cpi->tpl_stats_buffer[frame].rec_picture_buf, cm->width,
             cm->height, cm->seq_params.subsampling_x,
             cm->seq_params.subsampling_y, cm->seq_params.use_highbitdepth,
-            cpi->oxcf.border_in_pixels, cm->byte_alignment))
+            AOM_ENC_NO_SCALE_BORDER, cm->byte_alignment))
       aom_internal_error(&cm->error, AOM_CODEC_MEM_ERROR,
                          "Failed to allocate frame buffer");
   }

av1/encoder/tpl_model.c

diff --git a/av1/encoder/tpl_model.c b/av1/encoder/tpl_model.c
index 6ccfd77..11aa2da 100644
--- a/av1/encoder/tpl_model.c
+++ b/av1/encoder/tpl_model.c
@@ -135,12 +135,6 @@
   uint32_t sse;
   int cost_list[5];
   const MvLimits tmp_mv_limits = x->mv_limits;
-  // We hash the ss_cfgs based on the reference's stride to avoid having to to
-  // compute ss_cfg everytime this function is called.
-  // TODO(chiyotsai@google.com): Make this non-static to prepare for tpl
-  // multi-threading.
-  static search_site_config ss_cfgs[11];
-
   MV best_ref_mv1_full; /* full-pixel value of best_ref_mv1 */
 
   best_ref_mv1_full.col = center_mv.col >> 3;
@@ -157,10 +151,11 @@
 
   av1_set_mv_search_range(&x->mv_limits, &center_mv);
 
-  search_site_config *ss_cfg = &ss_cfgs[stride_ref % 11];
-  if (ss_cfg->stride != stride_ref) {
-    av1_init3smotion_compensation(ss_cfg, stride_ref);
-  }
+  search_site_config *ss_cfg = &cpi->ss_cfg[SS_CFG_SRC];
+  if (ss_cfg->stride != stride_ref) ss_cfg = &cpi->ss_cfg[SS_CFG_LOOKAHEAD];
+
+  assert(ss_cfg->stride == stride_ref);
+
   av1_full_pixel_search(cpi, x, bsize, &best_ref_mv1_full, step_param, 1,
                         search_method, 0, sadpb, cond_cost_list(cpi, cost_list),
                         &center_mv, INT_MAX, 0, (MI_SIZE * mi_col),
@@ -1014,6 +1009,8 @@
   int bottom_index, top_index;
   EncodeFrameParams this_frame_params = *frame_params;
 
+  if (cpi->oxcf.superres_mode != SUPERRES_NONE) return;
+
   cm->current_frame.frame_type = frame_params->frame_type;
   for (int gf_index = gf_group->index; gf_index < gf_group->size; ++gf_index) {
     av1_configure_buffer_updates(cpi, &this_frame_params,