commit 677cc01e6d9cfccb644beec95a8e59f20d7628e3
author Fyodor Kyslov <kyslov@google.com> Wed Jan 27 18:35:07 2021 -0800
committer Fyodor Kyslov <kyslov@google.com> Fri Jan 29 18:00:50 2021 +0000
tree 9363a598860f1e0f218558f1b3f8ba21e8480d4f
parent d289ce97f9cbb479ff42614d2f2d4d48563fe71e

AV1 RT: Make TXB RD Records dynamically allocated

TXB RD Records take about 3.2MB and are part of MACROBLOCK structure.
For RealTime multithreading this leads to huge overhead when copying
MACROBLOCK structure for each worker thread. Allocating this dynamically
helps to reduce overhead - ~15% speed up for 2 threads encoding on 360p
content of speed 8

Change-Id: I1b88fbf41bec67fe2654c88f41f3ef04ca2d9153

av1/encoder/ethread.c

diff --git a/av1/encoder/ethread.c b/av1/encoder/ethread.c
index 6f6ffde..2aa3dd0 100644
--- a/av1/encoder/ethread.c
+++ b/av1/encoder/ethread.c
@@ -483,8 +483,10 @@
                            &td->mb.e_mbd);
 
     cfl_init(&td->mb.e_mbd.cfl, &cm->seq_params);
-    av1_crc32c_calculator_init(
-        &td->mb.txfm_search_info.mb_rd_record.crc_calculator);
+    if (td->mb.txfm_search_info.txb_rd_records != NULL) {
+      av1_crc32c_calculator_init(
+          &td->mb.txfm_search_info.txb_rd_records->mb_rd_record.crc_calculator);
+    }
 
     av1_encode_sb_row(cpi, td, tile_row, tile_col, current_mi_row);
 #if CONFIG_MULTITHREAD
@@ -773,6 +775,10 @@
     EncWorkerData *const thread_data = (EncWorkerData *)worker->data1;
     cpi->intrabc_used |= thread_data->td->intrabc_used;
     cpi->deltaq_used |= thread_data->td->deltaq_used;
+    if (thread_data->td->mb.txfm_search_info.txb_rd_records) {
+      aom_free(thread_data->td->mb.txfm_search_info.txb_rd_records);
+      thread_data->td->mb.txfm_search_info.txb_rd_records = NULL;
+    }
 
     // Accumulate counters.
     if (i > 0) {
@@ -825,6 +831,11 @@
         }
       }
     }
+    if (!cpi->sf.rt_sf.use_nonrd_pick_mode) {
+      thread_data->td->mb.txfm_search_info.txb_rd_records =
+          (TxbRdRecords *)aom_malloc(sizeof(TxbRdRecords));
+    }
+
     if (thread_data->td->counts != &cpi->counts) {
       memcpy(thread_data->td->counts, &cpi->counts, sizeof(cpi->counts));
     }
@@ -868,6 +879,10 @@
     if (thread_data->td != &cpi->td) {
       thread_data->td->mb = cpi->td.mb;
     }
+    if (!cpi->sf.rt_sf.use_nonrd_pick_mode) {
+      thread_data->td->mb.txfm_search_info.txb_rd_records =
+          (TxbRdRecords *)aom_malloc(sizeof(TxbRdRecords));
+    }
   }
 }
 #endif
@@ -1148,6 +1163,12 @@
   fp_prepare_enc_workers(cpi, fp_enc_row_mt_worker_hook, num_workers);
   launch_workers(&cpi->mt_info, num_workers);
   sync_enc_workers(&cpi->mt_info, cm, num_workers);
+  for (int i = num_workers - 1; i >= 0; i--) {
+    EncWorkerData *const thread_data = &cpi->mt_info.tile_thr_data[i];
+    if (thread_data->td->mb.txfm_search_info.txb_rd_records) {
+      aom_free(thread_data->td->mb.txfm_search_info.txb_rd_records);
+    }
+  }
 }
 
 void av1_tpl_row_mt_sync_read_dummy(AV1TplRowMultiThreadSync *tpl_mt_sync,