commit 759b320464b501cfe996c7504a282e2dbe7f9455
author Jayasanker J <jayasanker.j@ittiam.com> Wed Mar 24 19:12:32 2021 +0530
committer Wan-Teh Chang <wtc@google.com> Wed Mar 31 21:40:00 2021 +0000
tree 6d333e5c25b902bf60aee2edb86de90d9fa44040
parent c6626020aa74f829ba701ed82274fe0d888420d9

Compute gradient data once at superblock level

This CL enables the generation of gradient data at superblock level
and the consumption of the same at block level when
partition_search_type is SEARCH_PARTITION. This optimization is
enabled only for intra frames. This design avoids repetitive
calculations of gradient data at pixel level as it is calculated
based on source pixels.

For AVIF still image encode,

           Instruction Count
  speed      Reduction(%)
    0           0.863
    1           1.500
    2           1.217
    3           1.812
    4           2.844
    5           3.242
    6           0.470

BUG=aomedia:2996
BUG=aomedia:2959

Change-Id: I0cf14b50c5598654dd6419b2793908c810ec7cbe

av1/encoder/ethread.c

diff --git a/av1/encoder/ethread.c b/av1/encoder/ethread.c
index 0d5d383..6cdc2c5 100644
--- a/av1/encoder/ethread.c
+++ b/av1/encoder/ethread.c
@@ -633,6 +633,10 @@
                                  sizeof(*thread_data->td->tmp_pred_bufs[j])));
       }
 
+      CHECK_MEM_ERROR(cm, thread_data->td->pixel_gradient_info,
+                      aom_malloc(sizeof(*thread_data->td->pixel_gradient_info) *
+                                 PLANE_TYPES * MAX_SB_SQUARE));
+
       if (cpi->sf.part_sf.partition_search_type == VAR_BASED_PARTITION) {
         const int num_64x64_blocks =
             (cm->seq_params.sb_size == BLOCK_64X64) ? 1 : 4;
@@ -873,6 +877,8 @@
         thread_data->td->mb.tmp_pred_bufs[j] =
             thread_data->td->tmp_pred_bufs[j];
       }
+      thread_data->td->mb.pixel_gradient_info =
+          thread_data->td->pixel_gradient_info;
 
       thread_data->td->mb.e_mbd.tmp_conv_dst = thread_data->td->mb.tmp_conv_dst;
       for (int j = 0; j < 2; ++j) {