From 86c69d116b587bf7e47e1ccfd0d80d2cbf908a55 Mon Sep 17 00:00:00 2001
From: limingqi107 <limingqi@huawei.com>
Date: Thu, 29 Jul 2021 11:02:28 +0800
Subject: [PATCH] unified runtime bug of host device and graph kernel

---
 mindspore/ccsrc/backend/session/kernel_graph.cc         | 6 ------
 mindspore/ccsrc/runtime/framework/actor/actor_common.cc | 6 +++++-
 mindspore/ccsrc/runtime/framework/actor/copy_actor.cc   | 7 +++++++
 mindspore/core/mindrt/src/thread/actor_threadpool.cc    | 2 +-
 4 files changed, 13 insertions(+), 8 deletions(-)

diff --git a/mindspore/ccsrc/backend/session/kernel_graph.cc b/mindspore/ccsrc/backend/session/kernel_graph.cc
index 04ee9c7e056..a108be59a6f 100644
--- a/mindspore/ccsrc/backend/session/kernel_graph.cc
+++ b/mindspore/ccsrc/backend/session/kernel_graph.cc
@@ -1216,7 +1216,6 @@ void KernelGraph::UpdateGraphOutputMap(const std::vector<AnfWithOutIndex> &old_o
     if (old_output == new_output) {
       continue;
     }
-
     // Update the graph output map.
     if (graph_output_to_front_node_map_.count(old_output) > 0) {
       MS_LOG(INFO) << "Replace backend output node " << old_output.first->fullname_with_scope() << " with index "
@@ -1226,11 +1225,6 @@ void KernelGraph::UpdateGraphOutputMap(const std::vector<AnfWithOutIndex> &old_o
       graph_output_to_front_node_map_.erase(old_output);
     }
 
-    // Update the internal output map.
-    if (IsInternalOutput(old_output.first, old_output.second)) {
-      ReplaceInternalOutput(old_output.first, new_output.first, old_output.second, new_output.second);
-    }
-
     if (old_output.first == new_output.first) {
       continue;
     }
diff --git a/mindspore/ccsrc/runtime/framework/actor/actor_common.cc b/mindspore/ccsrc/runtime/framework/actor/actor_common.cc
index 9af9b0216a6..fdbc11d7e50 100644
--- a/mindspore/ccsrc/runtime/framework/actor/actor_common.cc
+++ b/mindspore/ccsrc/runtime/framework/actor/actor_common.cc
@@ -32,7 +32,7 @@ void ComputeThreadNums(size_t *actor_thread_num, size_t *OMP_thread_num) {
   auto context_ptr = MsContext::GetInstance();
   MS_EXCEPTION_IF_NULL(context_ptr);
   // The pyNative mode is the step execution strategy, so only need the kActorThreadMinNum.
-  if (context_ptr->get_param<bool>(MS_CTX_SAVE_GRAPHS_FLAG) == kPynativeMode) {
+  if (context_ptr->get_param<int>(MS_CTX_EXECUTION_MODE) == kPynativeMode) {
     *actor_thread_num = kActorThreadMinNum;
   } else {
     *actor_thread_num = cpu_core_num < kActorThreadMinNum ? kActorThreadMinNum : cpu_core_num;
@@ -117,6 +117,10 @@ bool IsGatherActor(const AnfNodePtr &front_node,
 bool Copy(DeviceTensor *dst_device_tensor, const DeviceTensor *src_device_tensor) {
   MS_EXCEPTION_IF_NULL(dst_device_tensor);
   MS_EXCEPTION_IF_NULL(src_device_tensor);
+  if (src_device_tensor->GetSize() != dst_device_tensor->GetSize()) {
+    MS_LOG(WARNING) << " Copy size is not equal, input size:" << src_device_tensor->GetSize()
+                    << ", output size:" << dst_device_tensor->GetSize();
+  }
 
   // Exist the size alignment in some device, so get the min device size.
   size_t copy_size = std::min(src_device_tensor->GetSize(), dst_device_tensor->GetSize());
diff --git a/mindspore/ccsrc/runtime/framework/actor/copy_actor.cc b/mindspore/ccsrc/runtime/framework/actor/copy_actor.cc
index 10510227630..660c2213218 100644
--- a/mindspore/ccsrc/runtime/framework/actor/copy_actor.cc
+++ b/mindspore/ccsrc/runtime/framework/actor/copy_actor.cc
@@ -72,6 +72,13 @@ void CopyActor::SendMemoryFreeReq(OpContext<DeviceTensor> *context) {
 
 void CopyActor::OnMemoryAllocFinish(OpContext<DeviceTensor> *context) {
   MS_EXCEPTION_IF_NULL(context);
+  MS_EXCEPTION_IF_NULL(output_device_tensor_[0]);
+  MS_EXCEPTION_IF_NULL(input_device_tensor_[0]);
+
+  if (input_device_tensor_[0]->GetSize() != output_device_tensor_[0]->GetSize()) {
+    MS_LOG(WARNING) << GetAID().Name() << " copy size is not equal, input size:" << input_device_tensor_[0]->GetSize()
+                    << ", output size:" << output_device_tensor_[0]->GetSize();
+  }
 
   if (!Copy(output_device_tensor_[0], input_device_tensor_[0])) {
     std::string error_info = "Copy device tensor failed: " + GetAID().Name();
diff --git a/mindspore/core/mindrt/src/thread/actor_threadpool.cc b/mindspore/core/mindrt/src/thread/actor_threadpool.cc
index fb81db4e8cb..2427a84da48 100644
--- a/mindspore/core/mindrt/src/thread/actor_threadpool.cc
+++ b/mindspore/core/mindrt/src/thread/actor_threadpool.cc
@@ -18,7 +18,7 @@
 #include "thread/core_affinity.h"
 
 namespace mindspore {
-constexpr size_t MAX_READY_ACTOR_NR = 1024;
+constexpr size_t MAX_READY_ACTOR_NR = 4096;
 void ActorWorker::CreateThread(ActorThreadPool *pool) {
   THREAD_RETURN_IF_NULL(pool);
   pool_ = pool;