commit f83d681765cd2aefa9a70ce771af48edd1dbf416
author Jingning Han <jingning@google.com> Thu Feb 27 13:08:19 2020 -0800
committer Jingning Han <jingning@google.com> Fri Feb 28 06:28:28 2020 +0000
tree 92433dd222f32769f7f2bb72b7e3b6b03deba972
parent 89223295465d8d6b4e2a8b117bc98431c37fe552

Reduce the tpl model memory usage

Reduce the tpl model memory usage to 35 frames. The memory footprint
before enabling 32 max gop length was 36 frames. The net change is
hence 1 frame less.

Change-Id: I847d6c7b16d0e3c166b44ee47fd9e2ee087a7932

av1/encoder/encoder.c

diff --git a/av1/encoder/encoder.c b/av1/encoder/encoder.c
index a4675ef..751e0dd 100644
--- a/av1/encoder/encoder.c
+++ b/av1/encoder/encoder.c
@@ -3162,6 +3162,7 @@
 #endif
 
   set_tpl_stats_block_size(cpi);
+
   for (int frame = 0; frame < MAX_LENGTH_TPL_FRAME_STATS; ++frame) {
     const int mi_cols = ALIGN_POWER_OF_TWO(cm->mi_cols, MAX_MIB_SIZE_LOG2);
     const int mi_rows = ALIGN_POWER_OF_TWO(cm->mi_rows, MAX_MIB_SIZE_LOG2);
@@ -3174,21 +3175,23 @@
     cpi->tpl_stats_buffer[frame].stride = cpi->tpl_stats_buffer[frame].width;
     cpi->tpl_stats_buffer[frame].mi_rows = cm->mi_rows;
     cpi->tpl_stats_buffer[frame].mi_cols = cm->mi_cols;
+  }
 
+  for (int frame = 0; frame < MAX_LAG_BUFFERS; ++frame) {
     CHECK_MEM_ERROR(
-        cm, cpi->tpl_stats_buffer[frame].tpl_stats_ptr,
+        cm, cpi->tpl_stats_pool[frame],
         aom_calloc(cpi->tpl_stats_buffer[frame].width *
                        cpi->tpl_stats_buffer[frame].height,
                    sizeof(*cpi->tpl_stats_buffer[frame].tpl_stats_ptr)));
-
-    if (aom_alloc_frame_buffer(
-            &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,
-            AOM_ENC_NO_SCALE_BORDER, cm->byte_alignment))
+    if (aom_alloc_frame_buffer(&cpi->tpl_rec_pool[frame], cm->width, cm->height,
+                               cm->seq_params.subsampling_x,
+                               cm->seq_params.subsampling_y,
+                               cm->seq_params.use_highbitdepth,
+                               AOM_ENC_NO_SCALE_BORDER, cm->byte_alignment))
       aom_internal_error(&cm->error, AOM_CODEC_MEM_ERROR,
                          "Failed to allocate frame buffer");
   }
+
   cpi->tpl_frame = &cpi->tpl_stats_buffer[REF_FRAMES + 1];
 
 #if CONFIG_COLLECT_PARTITION_STATS == 2
@@ -3530,12 +3533,9 @@
 #endif
   }
 
-  for (int frame = 0; frame < MAX_LENGTH_TPL_FRAME_STATS; ++frame) {
-    aom_free(cpi->tpl_stats_buffer[frame].tpl_stats_ptr);
-    cpi->tpl_stats_buffer[frame].is_valid = 0;
-
-    aom_free_frame_buffer(&cpi->tpl_stats_buffer[frame].rec_picture_buf);
-    cpi->tpl_stats_buffer[frame].rec_picture = NULL;
+  for (int frame = 0; frame < MAX_LAG_BUFFERS; ++frame) {
+    aom_free(cpi->tpl_stats_pool[frame]);
+    aom_free_frame_buffer(&cpi->tpl_rec_pool[frame]);
   }
 
   for (t = cpi->num_workers - 1; t >= 0; --t) {