!1384 Performance optimization and fix memory leak.

From: @tangqunzhang
Reviewed-by: @ji_chen,@xchu42
Signed-off-by: @ji_chen

ge/generator/ge_generator.cc

@@ -36,6 +36,7 @@
 #include "graph/utils/type_utils.h"
 #include "init/gelib.h"
 #include "model/ge_model.h"
+#include "analyzer/analyzer.h"
 
 using std::map;
 using std::string;
@@ -1007,13 +1008,13 @@ Status GeGenerator::Impl::BuildModel(const Graph &graph, const vector<GeTensor>
   ErrorManager::GetInstance().SetStage(ErrorMessage::kModelCompile, ErrorMessage::kOther);
   if (ret != SUCCESS) {
     GELOGE(GE_GENERATOR_GRAPH_MANAGER_BUILD_GRAPH_FAILED, "GraphManager build graph fail, graph id: %u", graph_id);
-    VarManagerPool::Instance().RemoveVarManager(session_id);
-    return GE_GENERATOR_GRAPH_MANAGER_BUILD_GRAPH_FAILED;
+    ret = GE_GENERATOR_GRAPH_MANAGER_BUILD_GRAPH_FAILED;
   }
 
+  RtContextUtil::GetInstance().DestroyRtContexts(session_id);
+  Analyzer::GetInstance()->DestroySessionJsonObject(session_id);
   VarManagerPool::Instance().RemoveVarManager(session_id);
-
-  return SUCCESS;
+  return ret;
 }
 
 Status GeGenerator::Impl::GenerateInfershapeGraph(const Graph &graph) {

ge/graph/build/memory/block_mem_assigner.cc

@@ -1735,7 +1735,7 @@ Status BlockMemAssigner::AssignOutputMemoryWithReuse(const NodePtr &node, vector
 ///
 void BlockMemAssigner::AssignMemoryWithReuse(vector<int64_t> &ranges) {
   (void)ge::GetContext().GetOption(OPTION_EXEC_DISABLE_REUSED_MEMORY, ge_disable_reuse_mem_env_);
-  GELOGD("Reuse memory %s", ge_disable_reuse_mem_env_ == "1" ? "close" : "open");
+  GEEVENT("Reuse memory %s", ge_disable_reuse_mem_env_ == "1" ? "close" : "open");
   string op_no_reuse_mem_str;
   const char *op_no_reuse_mem = std::getenv(OP_NO_REUSE_MEM);
   GE_IF_BOOL_EXEC(op_no_reuse_mem != nullptr, op_no_reuse_mem_str = string(op_no_reuse_mem);
@@ -2125,7 +2125,7 @@ void SetBlockOpMemOffset(MemoryBlock *block, int32_t child_block_level) {
 
   child_block_level++;
   for (MemoryBlock *child_block : block->ChildBlockList()) {
-      SetBlockOpMemOffset(child_block, child_block_level);
+    SetBlockOpMemOffset(child_block, child_block_level);
   }
 }
 

ge/graph/build/memory/var_mem_assign_util.cc

@@ -311,6 +311,7 @@ Status VarMemAssignUtil::SetOutTransNodeToAssign(const ge::NodePtr &node, const
 }
 
 Status VarMemAssignUtil::AssignMemory2HasRefAttrNode(ge::ComputeGraphPtr &compute_graph) {
+  GraphToNodeMap graph_to_node;
   for (const ge::NodePtr &n : compute_graph->GetAllNodes()) {
     string ref_var_src_var_name;
     auto op_desc = n->GetOpDesc();
@@ -318,7 +319,8 @@ Status VarMemAssignUtil::AssignMemory2HasRefAttrNode(ge::ComputeGraphPtr &comput
     for (uint32_t idx = 0; idx < op_desc->GetOutputsSize(); idx += 1) {
       const auto out_desc = op_desc->MutableOutputDesc(idx);
       if (ge::AttrUtils::GetStr(out_desc, REF_VAR_SRC_VAR_NAME, ref_var_src_var_name)) {
-        GE_CHK_STATUS_RET(AssignData2VarRef(n, ref_var_src_var_name, compute_graph->GetSessionID(), idx));
+        GE_CHK_STATUS_RET(
+          AssignData2VarRef(n, ref_var_src_var_name, compute_graph->GetSessionID(), idx, graph_to_node));
       }
     }
   }
@@ -326,16 +328,37 @@ Status VarMemAssignUtil::AssignMemory2HasRefAttrNode(ge::ComputeGraphPtr &comput
 }
 
 Status VarMemAssignUtil::AssignData2VarRef(const ge::NodePtr &has_ref_attr_node, const string &src_var_name,
-                                           uint64_t session_id, uint32_t out_index) {
+                                           uint64_t session_id, uint32_t out_index,
+                                           GraphToNodeMap &graph_to_node) {
   // Get ref_var_src_var address
   auto root_graph = GraphUtils::FindRootGraph(has_ref_attr_node->GetOwnerComputeGraph());
   GE_CHECK_NOTNULL(root_graph);
-  ge::NodePtr var_ref_src_var = root_graph->FindNode(src_var_name);
-  if (var_ref_src_var == nullptr) {
+  // Cache mapping (name to nodeptr) simproves query performance
+  auto &name_to_node = graph_to_node[root_graph];
+  if (name_to_node.empty()) {
+    for (const ge::NodePtr &n : root_graph->GetDirectNode()) {
+      name_to_node.emplace(n->GetName(), n);
+    }
+    for (auto sub_graph : root_graph->GetAllSubgraphs()) {
+      auto &name_to_node_sub = graph_to_node[sub_graph];
+      if (name_to_node_sub.empty()) {
+        for (const ge::NodePtr &n : sub_graph->GetDirectNode()) {
+          name_to_node_sub.emplace(n->GetName(), n);
+        }
+      }
+    }
+  }
+
+  ge::NodePtr var_ref_src_var = nullptr;
+  auto it = name_to_node.find(src_var_name);
+  if ((it != name_to_node.end()) && (it->second != nullptr)) {
+    var_ref_src_var = it->second;
+  } else {
     for (auto sub_graph : root_graph->GetAllSubgraphs()) {
-        auto node_ptr = sub_graph->FindNode(src_var_name);
-        if (node_ptr != nullptr) {
-          var_ref_src_var = node_ptr;
+      auto &name_to_node_sub = graph_to_node[sub_graph];
+      it = name_to_node_sub.find(src_var_name);
+      if ((it != name_to_node_sub.end()) && (it->second != nullptr)) {
+          var_ref_src_var = it->second;
           break;
         }
     }

ge/graph/build/memory/var_mem_assign_util.h

@@ -22,6 +22,8 @@
 #include "graph/utils/node_utils.h"
 
 namespace ge {
+using GraphToNodeMap = std::map<ge::ComputeGraphPtr, std::map<std::string, ge::NodePtr>>;
+
 class VarMemAssignUtil {
  public:
   static Status AssignVarMemory(ge::ComputeGraphPtr &compute_graph);
@@ -47,7 +49,7 @@ class VarMemAssignUtil {
   static Status DealTransNode(const ge::NodePtr &final_trans_node);
   static Status DealExportTransNode(const ge::NodePtr &node, const ge::NodePtr &final_trans_node);
   static Status AssignData2VarRef(const ge::NodePtr &variable_ref, const std::string &src_var_name, uint64_t session_id,
-                                  uint32_t out_index);
+                                  uint32_t out_index, GraphToNodeMap &graph_to_node);
 
   static Status SetOutTransNodeToAssign(const ge::NodePtr &node, const ge::NodePtr &final_trans_node, size_t index);
 };

ge/graph/load/model_manager/davinci_model.cc

@@ -2137,7 +2137,6 @@ Status DavinciModel::CopyInputData(const InputData &input_data, bool device_data
 
 Status DavinciModel::SyncVarData() {
   GELOGI("Sync var data, model id:%u", model_id_);
-  Status ret = SUCCESS;
 
   if (global_step_addr_ != nullptr && global_step_size_ != 0) {
     const vector<uint64_t> v_step = { iterator_count_ };
@@ -2145,7 +2144,7 @@ Status DavinciModel::SyncVarData() {
                            RT_MEMCPY_HOST_TO_DEVICE));
   }
 
-  return ret;
+  return SUCCESS;
 }
 
 Status DavinciModel::InitModelProfile() {
@@ -3262,11 +3261,9 @@ Status DavinciModel::CopyModelData(const InputData &input_data, OutputData &outp
 ///
 Status DavinciModel::UpdateIoTaskArgs(const std::map<uint32_t, ZeroCopyOffset> &data_info, bool is_input,
                                       const vector<DataBuffer> &blobs, bool is_dynamic, const string &batch_label) {
-  string input_or_output;
-  is_input ? input_or_output = "input" : input_or_output = "output";
   if (blobs.size() != data_info.size()) {
     GELOGE(ACL_ERROR_GE_PARAM_INVALID, "Verify %s data num failed: model requires %zu, but user actually feeds %zu",
-           input_or_output.c_str(), data_info.size(), blobs.size());
+           is_input ? "input" : "output", data_info.size(), blobs.size());
     return ACL_ERROR_GE_PARAM_INVALID;
   }
 
@@ -3274,7 +3271,7 @@ Status DavinciModel::UpdateIoTaskArgs(const std::map<uint32_t, ZeroCopyOffset> &
     if (data.first >= blobs.size()) {  // check data index.
       GELOGE(ACL_ERROR_GE_PARAM_INVALID,
              "Verify %s data num failed: can not find No.%u data, because user only feeds %zu",
-             input_or_output.c_str(), data.first, blobs.size());
+             is_input ? "input" : "output", data.first, blobs.size());
       return ACL_ERROR_GE_PARAM_INVALID;
     }
 
@@ -3306,21 +3303,20 @@ Status DavinciModel::UpdateIoTaskArgs(const std::map<uint32_t, ZeroCopyOffset> &
     }
 
     for (size_t count = 0; count < data.second.GetDataCount(); ++count) {
-      int64_t size = data.second.GetDataInfo().at(count).first;
       void *addr = data.second.GetDataInfo().at(count).second;
       void *buffer_addr = reinterpret_cast<void *>(reinterpret_cast<uintptr_t>(buffer.data) +
                                                    data.second.GetRelativeOffset().at(count));
       GELOGI("[ZCPY] Copy %s blobs_index %u, virtual_addr: %p, size: %ld, user_data_addr: %p, batch_label: %s",
-             input_or_output.c_str(), data.first, addr, size, buffer_addr, batch_label.c_str());
+             is_input ? "input" : "output", data.first, addr, data.second.GetDataInfo().at(count).first,
+             buffer_addr, batch_label.c_str());
       // For input data, just copy for rts task.
-      for (ZeroCopyTask &task : zero_copy_tasks_) {
-        if (task.GetBatchLabel() != kDefaultBatchLable && task.GetBatchLabel() != batch_label) {
+      for (auto &task : zero_copy_tasks_) {
+        bool not_same_batch = (task.GetBatchLabel() != kDefaultBatchLable && task.GetBatchLabel() != batch_label);
+        if (not_same_batch) {
           continue;
         }
         uintptr_t addr_val = reinterpret_cast<uintptr_t>(addr);
-        if (task.UpdateTaskParam(addr_val, buffer_addr) != SUCCESS) {
-          return ACL_ERROR_GE_PARAM_INVALID;
-        }
+        (void)task.UpdateTaskParam(addr_val, buffer_addr);
       }
     }
   }
@@ -3980,7 +3976,7 @@ Status DavinciModel::InitOrigInputInfo(uint32_t index, const OpDescPtr &op_desc)
 Status DavinciModel::GetOrigInputInfo(uint32_t index, OriginInputInfo &orig_input_info) const {
   const auto it = orig_input_info_.find(index);
   if (it == orig_input_info_.end()) {
-    GELOGE(ACL_ERROR_GE_AIPP_NOT_EXIST, "there is not AIPP related with index %u.", index);
+    GELOGE(ACL_ERROR_GE_AIPP_NOT_EXIST, "There is not AIPP related with index %u.", index);
     return ACL_ERROR_GE_AIPP_NOT_EXIST;
   }
 
@@ -4014,7 +4010,7 @@ void DavinciModel::ParseAIPPInfo(std::string in_out_info, InputOutputDims &dims_
 
 Status DavinciModel::InitAippInputOutputDims(uint32_t index, const OpDescPtr &op_desc) {
   if (!op_desc->HasAttr(ATTR_NAME_AIPP_INPUTS) || !op_desc->HasAttr(ATTR_NAME_AIPP_OUTPUTS)) {
-    GELOGI("there is not AIPP related with index %u.", index);
+    GELOGI("There is not AIPP related with index %u.", index);
     return SUCCESS;
   }
 
@@ -4031,7 +4027,7 @@ Status DavinciModel::InitAippInputOutputDims(uint32_t index, const OpDescPtr &op
       ConstGeTensorDescPtr data_input_desc = op_desc->GetInputDescPtr(kDataIndex);
       int64_t data_input_size;
       (void)TensorUtils::GetSize(*(op_desc->GetInputDescPtr(kDataIndex)), data_input_size);
-      GELOGD("related Data[%d]: tensor_name: %s, dim_num: %zu, tensor_size: %zu, format: %s, data_type: %s, shape: %s.",
+      GELOGD("Related Data[%d]: tensor_name: %s, dim_num: %zu, tensor_size: %zu, format: %s, data_type: %s, shape: %s.",
           index, op_desc->GetName().c_str(), data_input_desc->GetShape().GetDimNum(), data_input_size,
           TypeUtils::FormatToSerialString(data_input_desc->GetFormat()).c_str(),
           TypeUtils::DataTypeToSerialString(data_input_desc->GetDataType()).c_str(),
@@ -4058,7 +4054,7 @@ Status DavinciModel::GetAllAippInputOutputDims(uint32_t index, vector<InputOutpu
                                                vector<InputOutputDims> &output_dims) const {
   const auto it = aipp_dims_info_.find(index);
   if (it == aipp_dims_info_.end()) {
-    GELOGE(ACL_ERROR_GE_AIPP_NOT_EXIST, "there is not AIPP related with index %u.", index);
+    GELOGE(ACL_ERROR_GE_AIPP_NOT_EXIST, "There is not AIPP related with index %u.", index);
     return ACL_ERROR_GE_AIPP_NOT_EXIST;
   }
 

tests/ut/ge/generator/ge_generator_unittest.cc

@@ -155,4 +155,17 @@ TEST_F(UtestGeGenerator, test_remove_const) {
   vector<GeTensor> outputs;
   generator.RemoveConst(inputs, outputs);
 }
+
+TEST_F(UtestGeGenerator, test_generate_online_model) {
+  GeTensorDesc tensor_desc;
+  GeTensor tensor(tensor_desc);
+  const vector<GeTensor> inputs = { tensor, tensor };
+  auto compute_graph = MakeGraph();
+  compute_graph->TopologicalSorting();
+  Graph graph = ge::GraphUtils::CreateGraphFromComputeGraph(compute_graph);
+  GeGenerator generator;
+  generator.Initialize({});
+  std::string name;
+  EXPECT_NE(generator.GenerateOfflineModel(graph, name, inputs), SUCCESS);
+}
 }  // namespace ge

tests/ut/ge/graph/build/mem_assigner_unittest.cc

@@ -33,6 +33,7 @@
 #include "graph/build/memory/graph_mem_assigner.h"
 #include "graph/build/memory/hybrid_mem_assigner.h"
 #include "graph/build/memory/max_block_mem_assigner.h"
+#include "graph/manager/graph_var_manager.h"
 #undef protected
 #undef private
 
@@ -77,8 +78,8 @@ class UtestMemoryAssignerTest : public testing::Test {
     op_def->SetWorkspaceBytes(workspace_bytes);
     return op_def;
   }
-  void MakeGraph(ge::ComputeGraphPtr &graph) {
-    ge::OpDescPtr op_def_a = CreateOpWithWsSize("A", 6000);
+  void MakeGraph(ge::ComputeGraphPtr &graph, const string &type = "some") {
+    ge::OpDescPtr op_def_a = CreateOpWithWsSize("A", 6000, type);
     op_def_a->SetStreamId(0);
     ge::OpDescPtr op_def_b = CreateOpWithWsSize("B", 120000);
     op_def_b->SetStreamId(0);
@@ -263,3 +264,38 @@ TEST_F(UtestMemoryAssignerTest, graph_memory_set_last_used_attr) {
   (void) ge::AttrUtils::GetInt(node_f->GetOpDesc()->GetInputDesc(0), ATTR_NAME_IS_END_OF_INPUTMEM_LIFECYCLE, flag);
   EXPECT_EQ(flag, 1);
 }
+
+TEST_F(UtestMemoryAssignerTest, graph_memory_assign_ref_var) {
+  ge::ComputeGraphPtr graph = make_shared<ge::ComputeGraph>("");
+  MakeGraph(graph, VARIABLE);
+  auto node_a = graph->FindNode("A");
+  auto node_b = graph->FindNode("B");
+  std::string value = "A";
+  (void) ge::AttrUtils::SetStr(node_b->GetOpDesc()->MutableOutputDesc(0), REF_VAR_SRC_VAR_NAME, value);
+  MemoryAssigner memory_assigner(graph);
+  map<int64_t, size_t> mem_offset;
+  size_t zero_memory_size = 0;
+  VarManager::Instance(0)->Init(0, 0, 0, 0);
+  EXPECT_EQ(memory_assigner.AssignMemory(false, mem_offset, zero_memory_size), GRAPH_SUCCESS);
+
+  EXPECT_EQ(node_b->GetOpDesc()->GetOutputOffset()[0], node_a->GetOpDesc()->GetOutputOffset()[0]);
+}
+
+TEST_F(UtestMemoryAssignerTest, graph_memory_assign_ref_var_not_found) {
+  ge::ComputeGraphPtr graph = make_shared<ge::ComputeGraph>("");
+  MakeGraph(graph, VARIABLE);
+
+  ge::ComputeGraphPtr sub_graph = make_shared<ge::ComputeGraph>("");
+  MakeReuseGraph(sub_graph);
+  graph->AddSubGraph(sub_graph);
+
+  auto node_a = graph->FindNode("A");
+  auto node_b = graph->FindNode("B");
+  std::string value = "M";
+  (void) ge::AttrUtils::SetStr(node_b->GetOpDesc()->MutableOutputDesc(0), REF_VAR_SRC_VAR_NAME, value);
+  MemoryAssigner memory_assigner(graph);
+  map<int64_t, size_t> mem_offset;
+  size_t zero_memory_size = 0;
+  VarManager::Instance(0)->Init(0, 0, 0, 0);
+  EXPECT_NE(memory_assigner.AssignMemory(false, mem_offset, zero_memory_size), GRAPH_SUCCESS);
+}

tests/ut/ge/graph/load/davinci_model_unittest.cc

@@ -22,6 +22,7 @@
 #include "graph/utils/graph_utils.h"
 #include "common/profiling/profiling_manager.h"
 #include "graph/load/model_manager/davinci_model.h"
+#include "graph/manager/graph_var_manager.h"
 
 using namespace std;
 
@@ -51,6 +52,10 @@ int32_t MsprofReport(uint32_t moduleId, uint32_t type, void *data, uint32_t len)
 
 TEST_F(UtestDavinciModel, init_success) {
   DavinciModel model(0, nullptr);
+  VarManager::Instance(0)->Init(0, 0, 0, 0);
+  map<string, string> options;
+  options[GRAPH_MEMORY_MAX_SIZE] = "1048576";
+  VarManager::Instance(0)->SetMemoryMallocSize(options);
   ComputeGraphPtr graph = make_shared<ComputeGraph>("default");
   ProfilingManager::Instance().is_load_profiling_ = true;
 
@@ -777,6 +782,10 @@ TEST_F(UtestDavinciModel, init_data_aipp_input_dims_normal) {
 
 // test label_set_task Init
 TEST_F(UtestDavinciModel, label_task_success) {
+  VarManager::Instance(0)->Init(0, 0, 0, 0);
+  map<string, string> options;
+  options[GRAPH_MEMORY_MAX_SIZE] = "1048576";
+  VarManager::Instance(0)->SetMemoryMallocSize(options);
   DavinciModel model(0, nullptr);
   ComputeGraphPtr graph = make_shared<ComputeGraph>("default");
 
@@ -944,6 +953,11 @@ TEST_F(UtestDavinciModel, simple_test_gmock) {
 }
 
 TEST_F(UtestDavinciModel, NnExecute) {
+  VarManager::Instance(0)->Init(0, 0, 0, 0);
+  map<string, string> options;
+  options[GRAPH_MEMORY_MAX_SIZE] = "1048576";
+  VarManager::Instance(0)->SetMemoryMallocSize(options);
+
   DavinciModel model(0, nullptr);
   ComputeGraphPtr graph = make_shared<ComputeGraph>("default");
   ProfilingManager::Instance().is_load_profiling_ = true;
@@ -967,6 +981,26 @@ TEST_F(UtestDavinciModel, NnExecute) {
     NodePtr node = graph->AddNode(op_desc);    // op_index = 0
   }
 
+  {
+    OpDescPtr op_desc = CreateOpDesc("memcpy", MEMCPYASYNC);
+    op_desc->AddInputDesc(tensor);
+    op_desc->AddOutputDesc(tensor);
+    op_desc->SetInputOffset({1024});
+    op_desc->SetOutputOffset({5120});
+    NodePtr node = graph->AddNode(op_desc);
+
+    domi::TaskDef *task_def = model_task_def->add_task();
+    task_def->set_stream_id(0);
+    task_def->set_type(RT_MODEL_TASK_MEMCPY_ASYNC);
+    domi::MemcpyAsyncDef *memcpy_async = task_def->mutable_memcpy_async();
+    memcpy_async->set_src(1024);
+    memcpy_async->set_dst(5120);
+    memcpy_async->set_dst_max(512);
+    memcpy_async->set_count(1);
+    memcpy_async->set_kind(RT_MEMCPY_DEVICE_TO_DEVICE);
+    memcpy_async->set_op_index(op_desc->GetId());
+  }
+
   {
     OpDescPtr op_desc = CreateOpDesc("output", NETOUTPUT);
     op_desc->AddInputDesc(tensor);