Merge remote-tracking branch 'upstream/development' into development

ge/analyzer/analyzer.cc

@@ -145,6 +145,7 @@ void Analyzer::DestroyGraphJsonObject(uint64_t session_id, uint64_t graph_id) {
     if (iter1 == (iter->second).end()) {
       GELOGW("Can not find the graph json object by session_id[%lu] and graph_id[%lu]. Do nothing.", session_id,
              graph_id);
+      return;
     }
     (iter->second).erase(iter1);
   }

ge/hybrid/node_executor/aicore/aicore_node_executor.cc

@@ -159,9 +159,13 @@ Status AiCoreNodeTask::ExecuteAsync(TaskContext &context, std::function<void()>
   auto op_desc = context.GetNodeItem().op_desc;
   GE_CHECK_NOTNULL(op_desc);
   GELOGI("[%s] ExecuteAsync Start.", op_desc->GetName().c_str());
-  for (auto &task : tasks_) {
+  for (auto it = tasks_.begin(); it != tasks_.end(); ++it) {
+    // AtomicAddrClean has 2 tasks
+    if (tasks_.size() == 2 && it == tasks_.begin() && !(*(tasks_.rbegin()))->GetClearAtomic()) {
+      continue;
+    }
     RECORD_EXECUTION_EVENT(context.GetExecutionContext(), context.GetNodeName(), "[AiCoreNodeLaunchKernel] Start");
-    GE_CHK_STATUS_RET_NOLOG(task->LaunchKernel(context.GetStream()));
+    GE_CHK_STATUS_RET_NOLOG((*it)->LaunchKernel(context.GetStream()));
     RECORD_EXECUTION_EVENT(context.GetExecutionContext(), context.GetNodeName(), "[AiCoreNodeLaunchKernel] End");
     RECORD_EXECUTION_EVENT(context.GetExecutionContext(), context.GetNodeName(), "[AiCoreNodeLaunchKernel] End");
   }
@@ -181,8 +185,12 @@ Status AiCoreNodeTask::UpdateArgs(TaskContext &context) {
   auto op_desc = context.GetNodeItem().op_desc;
   GE_CHECK_NOTNULL(op_desc);
   GELOGI("[%s] AiCoreNodeTask UpdateArgs Start.", op_desc->GetName().c_str());
-  for (auto &task : tasks_) {
-    GE_CHK_STATUS_RET_NOLOG(task->UpdateArgs(context));
+  for (auto it = tasks_.rbegin(); it != tasks_.rend(); ++it) {
+    GE_CHK_STATUS_RET_NOLOG((*it)->UpdateArgs(context));
+    // AtomicAddrClean has 2 tasks
+    if (tasks_.size() == 2 && it == tasks_.rbegin() && !(*it)->GetClearAtomic()) {
+      break;
+    }
   }
   GELOGI("[%s] AiCoreNodeTask UpdateArgs End.", op_desc->GetName().c_str());
   return SUCCESS;
@@ -190,8 +198,12 @@ Status AiCoreNodeTask::UpdateArgs(TaskContext &context) {
 
 Status AiCoreNodeTask::UpdateTilingData(TaskContext &context) {
   GELOGD("[%s] PrepareWithShape started", context.GetNodeName());
-  for (auto &task : tasks_) {
-    GE_CHK_STATUS_RET_NOLOG(task->PrepareWithShape(context));
+  for (auto it = tasks_.rbegin(); it != tasks_.rend(); ++it) {
+    GE_CHK_STATUS_RET_NOLOG((*it)->PrepareWithShape(context));
+    // AtomicAddrClean has 2 tasks
+    if (tasks_.size() == 2 && it == tasks_.rbegin() && !(*it)->GetClearAtomic()) {
+      break;
+    }
   }
   GELOGD("[%s] Done PrepareWithShape successfully.", context.GetNodeName());
   return SUCCESS;

ge/hybrid/node_executor/aicore/aicore_op_task.cc

@@ -121,6 +121,7 @@ Status AiCoreOpTask::UpdateTilingInfo(TaskContext &context) {
   GELOGD("[%s] Start to update tiling info for task: [%s]", node->GetName().c_str(), stub_name_.c_str());
   OpRunInfo tiling_info;
   tiling_info.block_dim = -1; // codex: Using uninitialized value
+  tiling_info.clear_atomic = true;
 
   auto execution_context = context.GetExecutionContext();
   RECORD_EXECUTION_EVENT(execution_context, context.GetNodeName(), "[CalcTilingInfo] Start");
@@ -130,6 +131,7 @@ Status AiCoreOpTask::UpdateTilingInfo(TaskContext &context) {
   // update op args by tiling info
   block_dim_ = static_cast<uint32_t>(tiling_info.block_dim);
   op_desc->SetWorkspaceBytes(tiling_info.workspaces);
+  clear_atomic_ = tiling_info.clear_atomic;
 
   tiling_data_ = tiling_info.tiling_data.str();
   if (tiling_data_.empty()) {

ge/hybrid/node_executor/aicore/aicore_op_task.h

@@ -46,6 +46,8 @@ class AiCoreOpTask {
 
   const std::string& GetName() const;
 
+  bool GetClearAtomic() const {return clear_atomic_;}
+
  protected:
   Status UpdateTilingInfo(TaskContext &context);
   virtual std::string GetKeyForOpParamSize() const;
@@ -66,6 +68,7 @@ class AiCoreOpTask {
   std::unique_ptr<uint8_t[]> args_ = nullptr;
   uint32_t args_size_ = 0;
   uint32_t block_dim_ = 1;
+  bool clear_atomic_ = true;
 };
 
 class AtomicAddrCleanOpTask : public AiCoreOpTask {

ge/ir_build/ge_ir_build.cc

@@ -171,7 +171,7 @@ class Impl {
   graphStatus InitDomiOmgContext(const string &input_shape, const string &input_format, const string &net_format,
                                  bool is_dynamic_input);
   void SetRtSocVersion();
-
+  void UpdateThreadContext();
  public:
   ge::GeGenerator generator_;
   std::map<std::string, std::string> options_;
@@ -225,8 +225,6 @@ graphStatus Impl::Init(const std::map<std::string, std::string> &options) {
     return ret;
   }
 
-  GetThreadLocalContext().SetGlobalOption(GetMutableGlobalOptions());
-  GetThreadLocalContext().SetGraphOption(options_);
   std::string build_mode = (options_.find(BUILD_MODE) == options_.end() || options_[BUILD_MODE] == BUILD_MODE_NORMAL)
                            ? "" : options_[BUILD_MODE];
   options_[BUILD_MODE] = build_mode;
@@ -286,7 +284,7 @@ graphStatus Impl::Init(const std::map<std::string, std::string> &options) {
   ge::PrintOptionMap(options_, "ge option");
 
   SetRtSocVersion();
-
+  UpdateThreadContext();
   // 3. init generator with options_
   ret = generator_.Initialize(options_, omg_context_);
   if (ret != GRAPH_SUCCESS) {
@@ -310,6 +308,11 @@ void Impl::SetRtSocVersion() {
   }
 }
 
+void Impl::UpdateThreadContext() {
+  GetThreadLocalContext().SetGlobalOption(GetMutableGlobalOptions());
+  GetThreadLocalContext().SetGraphOption(options_);
+}
+
 graphStatus Impl::CreateInputsForIRBuild(const ge::Graph &graph, vector<ge::GeTensor> &inputs) {
   auto compute_graph = ge::GraphUtils::GetComputeGraph(graph);
   GE_CHECK_NOTNULL(compute_graph);
@@ -333,13 +336,15 @@ graphStatus Impl::CreateInputsForIRBuild(const ge::Graph &graph, vector<ge::GeTe
         data_shape = tensor.GetShape();
         GELOGI("Data op get shape from InputDesc in ge ir graph.");
       }
-
+      // If user point input format, do work for all data ops; else do according to tensor_desc
+      auto data_format = omg_context_.format != domi::DOMI_TENSOR_ND ?
+        ge::TypeUtils::DomiFormatToFormat(omg_context_.format) : tensor.GetFormat();
       ge::DataType data_type = tensor.GetDataType();
       string data_type_str = ge::TypeUtils::DataTypeToSerialString(data_type);
       GELOGI("Data op get data type:%s from InputDesc in ge ir graph.", data_type_str.c_str());
 
       ge::GeTensor inputTensor;
-      ge::GeTensorDesc desc(data_shape, ge::Format(omg_context_.format), data_type);
+      ge::GeTensorDesc desc(data_shape, ge::Format(data_format), data_type);
       inputTensor.SetTensorDesc(desc);
       inputs.push_back(inputTensor);
     }

ge/single_op/task/aicpu_task_builder.cc

@@ -112,14 +112,11 @@ namespace ge {
 
   Status AiCpuTaskBuilder::BuildTask(ge::AiCpuTask &task, const SingleOpModelParam &param,
                                      bool dynamic_flag, uint64_t session_id) {
-    void *io_addr = nullptr;
-    void *kernel_workspace = nullptr;
-    GE_CHK_STATUS_RET_NOLOG(InitWorkspaceAndIO(&io_addr, &kernel_workspace, param, dynamic_flag));
+    GE_CHK_STATUS_RET_NOLOG(InitWorkspaceAndIO(&task.io_addr_, &task.workspace_addr_, param, dynamic_flag));
 
     STR_FWK_OP_KERNEL fwk_op_kernel = {0};
-    auto ret = SetFmkOpKernel(io_addr, kernel_workspace, fwk_op_kernel);
+    auto ret = SetFmkOpKernel(task.io_addr_, task.workspace_addr_, fwk_op_kernel);
     if (ret != SUCCESS) {
-      (void)rtFree(io_addr);
       return ret;
     }
 
@@ -150,15 +147,12 @@ namespace ge {
       return FAILED;)
     ret = SetKernelArgs(&task.args_, fwk_op_kernel);
     if (ret != SUCCESS) {
-      (void)rtFree(io_addr);
       return ret;
     }
 
     task.arg_size_ = sizeof(STR_FWK_OP_KERNEL);
     task.op_type_ = op_desc_->GetName();
-    task.io_addr_ = io_addr;
     task.task_info_ = kernel_def_.task_info();
-    task.workspace_addr_ = kernel_workspace;
     task.dynamic_flag_ = dynamic_flag;
 
     auto debug_info = BuildTaskUtils::GetTaskInfo(op_desc_);