commit 1f58dd84770c7108f288e9002168e21b70c2a287
author Ravi Chaudhary <ravi.chaudhary@ittiam.com> Fri Dec 07 17:24:15 2018 +0530
committer Yunqing Wang <yunqingwang@google.com> Sat Dec 08 20:38:21 2018 +0000
tree 0040b83c1bc8d17d640c6399aefceedc6a8e755d
parent 98dae942edd3032c89b0e8bae9b4d50398931405

Thread0 as master thread in enc multi-threading

When multi-threading of encoder is enabled, last thread will
act as the master thread. This patch will enable Thread0 to
be the master thread instead of the last thread.

This change will facilitate switching between tile-mt and
row-mt at frame level.

Change-Id: I674f234a3cda5b83162a61a7f2b517ff9cd92075

av1/encoder/encoder.c

diff --git a/av1/encoder/encoder.c b/av1/encoder/encoder.c
index c3f6917..2836160 100644
--- a/av1/encoder/encoder.c
+++ b/av1/encoder/encoder.c
@@ -3103,7 +3103,7 @@
     cpi->tpl_stats[frame].is_valid = 0;
   }
 
-  for (t = 0; t < cpi->num_workers; ++t) {
+  for (t = cpi->num_workers - 1; t >= 0; --t) {
     AVxWorker *const worker = &cpi->workers[t];
     EncWorkerData *const thread_data = &cpi->tile_thr_data[t];
 
@@ -3112,7 +3112,7 @@
 
     // Deallocate allocated thread data.
     if (cpi->row_mt == 1) aom_free(thread_data->td->tctx);
-    if (t < cpi->num_workers - 1) {
+    if (t > 0) {
       aom_free(thread_data->td->palette_buffer);
       aom_free(thread_data->td->tmp_conv_dst);
       for (int j = 0; j < 2; ++j) {

av1/encoder/ethread.c

diff --git a/av1/encoder/ethread.c b/av1/encoder/ethread.c
index a3fb93e..7b85f7e 100644
--- a/av1/encoder/ethread.c
+++ b/av1/encoder/ethread.c
@@ -386,7 +386,7 @@
   }
 #endif
 
-  for (int i = 0; i < num_workers; i++) {
+  for (int i = num_workers - 1; i >= 0; i--) {
     AVxWorker *const worker = &cpi->workers[i];
     EncWorkerData *const thread_data = &cpi->tile_thr_data[i];
 
@@ -397,7 +397,7 @@
     thread_data->cpi = cpi;
     thread_data->thread_id = i;
 
-    if (i < num_workers - 1) {
+    if (i > 0) {
       // Allocate thread data.
       CHECK_MEM_ERROR(cm, thread_data->td,
                       aom_memalign(32, sizeof(*thread_data->td)));
@@ -478,14 +478,14 @@
 static void launch_enc_workers(AV1_COMP *cpi, int num_workers) {
   const AVxWorkerInterface *const winterface = aom_get_worker_interface();
   // Encode a frame
-  for (int i = 0; i < num_workers; i++) {
+  for (int i = num_workers - 1; i >= 0; i--) {
     AVxWorker *const worker = &cpi->workers[i];
     EncWorkerData *const thread_data = (EncWorkerData *)worker->data1;
 
     // Set the starting tile for each thread.
     thread_data->start = i;
 
-    if (i == cpi->num_workers - 1)
+    if (i == 0)
       winterface->execute(worker);
     else
       winterface->launch(worker);
@@ -497,7 +497,7 @@
   int had_error = 0;
 
   // Encoding ends.
-  for (int i = 0; i < num_workers; i++) {
+  for (int i = num_workers - 1; i >= 0; i--) {
     AVxWorker *const worker = &cpi->workers[i];
     had_error |= !winterface->sync(worker);
   }
@@ -508,12 +508,12 @@
 }
 
 static void accumulate_counters_enc_workers(AV1_COMP *cpi, int num_workers) {
-  for (int i = 0; i < num_workers; i++) {
+  for (int i = num_workers - 1; i >= 0; i--) {
     AVxWorker *const worker = &cpi->workers[i];
     EncWorkerData *const thread_data = (EncWorkerData *)worker->data1;
     cpi->intrabc_used |= thread_data->td->intrabc_used;
     // Accumulate counters.
-    if (i < cpi->num_workers - 1) {
+    if (i > 0) {
       av1_accumulate_frame_counts(&cpi->counts, thread_data->td->counts);
       accumulate_rd_opt(&cpi->td, thread_data->td);
       cpi->td.mb.txb_split_count += thread_data->td->mb.txb_split_count;
@@ -523,7 +523,7 @@
 
 static void prepare_enc_workers(AV1_COMP *cpi, AVxWorkerHook hook,
                                 int num_workers) {
-  for (int i = 0; i < num_workers; i++) {
+  for (int i = num_workers - 1; i >= 0; i--) {
     AVxWorker *const worker = &cpi->workers[i];
     EncWorkerData *const thread_data = &cpi->tile_thr_data[i];
 
@@ -560,7 +560,7 @@
       memcpy(thread_data->td->counts, &cpi->counts, sizeof(cpi->counts));
     }
 
-    if (i < num_workers - 1) {
+    if (i > 0) {
       thread_data->td->mb.palette_buffer = thread_data->td->palette_buffer;
       thread_data->td->mb.tmp_conv_dst = thread_data->td->tmp_conv_dst;
       for (int j = 0; j < 2; ++j) {