!818 Memory optimization during model loading

From: @li-lei0106
Reviewed-by: @ji_chen,@xchu42
Signed-off-by: @ji_chen

ge/graph/load/new_model_manager/data_dumper.cc

@@ -830,6 +830,13 @@ Status DataDumper::UnloadDumpInfo() {
   return SUCCESS;
 }
 
+void DataDumper::DumpShrink() {
+  compute_graph_.reset();
+  input_map_.clear();
+  ref_info_.clear();
+  op_list_.clear();
+}
+
 void DataDumper::PrintCheckLog(string &dump_list_key) {
   std::set<std::string> model_list = dump_properties_.GetAllDumpModel();
   if (model_list.empty()) {

ge/graph/load/new_model_manager/data_dumper.h

@@ -83,6 +83,8 @@ class DataDumper {
 
   Status UnloadDumpInfo();
 
+  void DumpShrink();
+
   void SetDumpProperties(const DumpProperties &dump_properties) { dump_properties_ = dump_properties; }
   const DumpProperties &GetDumpProperties() const { return dump_properties_; }
   bool GetOpDescInfo(uint32_t stream_id, uint32_t task_id, OpDescInfo &op_desc_info) const;
@@ -112,18 +114,18 @@ class DataDumper {
   struct InnerInputMapping;
 
   std::vector<OpDescInfo> op_desc_info_;
-  std::vector<InnerDumpInfo> op_list_;
+  std::vector<InnerDumpInfo> op_list_;  // release after DavinciModel::Init
   uint32_t end_graph_task_id_ = 0;
   uint32_t end_graph_stream_id_ = 0;
   bool is_end_graph_ = false;
-  std::multimap<std::string, InnerInputMapping> input_map_;
+  std::multimap<std::string, InnerInputMapping> input_map_;  // release after DavinciModel::Init
   bool load_flag_;
   uint32_t device_id_;
   uintptr_t global_step_;
   uintptr_t loop_per_iter_;
   uintptr_t loop_cond_;
-  ComputeGraphPtr compute_graph_;
-  std::map<OpDescPtr, void *> ref_info_;
+  ComputeGraphPtr compute_graph_;  // release after DavinciModel::Init
+  std::map<OpDescPtr, void *> ref_info_;  // release after DavinciModel::Init
   void *l1_fusion_addr_ = nullptr;
 
 

ge/graph/load/new_model_manager/davinci_model.cc

@@ -150,14 +150,7 @@ DavinciModel::~DavinciModel() {
       GELOGW("UnloadDumpInfo failed, ret: %u.", ret);
     }
 
-    for (const auto &op_and_addr : saved_task_addrs_) {
-      auto addr = op_and_addr.second;
-      if (addr != nullptr) {
-        GE_CHK_RT(rtFree(addr));
-      }
-      addr = nullptr;
-    }
-    saved_task_addrs_.clear();
+    ClearTaskAddrs();
 
     GE_CHK_STATUS(ModelRunStop());
 
@@ -221,6 +214,17 @@ DavinciModel::~DavinciModel() {
   }
 }
 
+void DavinciModel::ClearTaskAddrs() {
+  for (const auto &op_and_addr : saved_task_addrs_) {
+    auto addr = op_and_addr.second;
+    if (addr != nullptr) {
+      GE_CHK_RT(rtFree(addr));
+    }
+    addr = nullptr;
+  }
+  saved_task_addrs_.clear();
+}
+
 void DavinciModel::UnbindHcomStream() {
   if (!all_hccl_stream_list_.empty()) {
     for (size_t i = 0; i < all_hccl_stream_list_.size(); i++) {
@@ -263,7 +267,10 @@ Status DavinciModel::Assign(const GeModelPtr &ge_model) {
 ///
 void DavinciModel::Shrink() {
   skt_info_ = {0, 0, 0, 0, nullptr, nullptr, {}, {}, {}, {}, {}, RT_KERNEL_DEFAULT, -1, 0, nullptr};
+  DumperShrink();
   ge_model_.reset();  // delete object.
+  op_list_.clear();
+  ClearTaskAddrs();
 }
 
 Status DavinciModel::InitWeightMem(void *dev_ptr, void *weight_ptr, size_t weight_size) {
@@ -738,7 +745,6 @@ Status DavinciModel::ReportProfilingData() {
   }
   ProfilingManager::Instance().ReportProfilingData(model_id_, GetTaskDescInfo(), compute_graph_desc_info);
   GE_CHK_STATUS(SinkModelProfile(), "Sink model profiler failed.");
-  op_list_.clear();
 
   return SUCCESS;
 }
@@ -963,7 +969,9 @@ Status DavinciModel::InitDataOp(const ComputeGraphPtr &graph, const NodePtr &nod
   }
 
   data_by_index[data_index] = op_desc;
-  data_op_list_.push_back(op_desc);
+  auto data_op = AttrUtils::CopyOpDesc(op_desc);
+  GE_CHECK_NOTNULL(data_op);
+  data_op_list_.push_back(data_op);
   if (known_node_) {
     return SUCCESS;
   }
@@ -1019,7 +1027,9 @@ Status DavinciModel::OptInputOutputInfo(const map<uint32_t, OpDescPtr> &data_by_
 
   data_op_list_.clear();
   for (auto &item : data_by_index) {
-    data_op_list_.emplace_back(item.second);
+    auto data_op = AttrUtils::CopyOpDesc(item.second);
+    GE_CHECK_NOTNULL(data_op);
+    data_op_list_.emplace_back(data_op);
     auto output_addrs = ModelUtils::GetOutputDataAddrs(runtime_param_, item.second);
     GELOGD("Data node: %s, output addr size: %zu", item.second->GetName().c_str(), output_addrs.size());
     input_addrs_list_.emplace_back(output_addrs);

ge/graph/load/new_model_manager/davinci_model.h

@@ -481,6 +481,10 @@ class DavinciModel {
     data_dumper_.SaveDumpTask(task_id, stream_id, op_desc, args);
   }
 
+  void DumperShrink() {
+    data_dumper_.DumpShrink();
+  }
+
   void SetEndGraphId(uint32_t task_id, uint32_t stream_id);
   DavinciModel &operator=(const DavinciModel &model) = delete;
 
@@ -644,6 +648,8 @@ class DavinciModel {
 
   void ReleaseTask();
 
+  void ClearTaskAddrs();
+
   void UnbindTaskSinkStream();
 
   bool IsAicpuKernelConnectSpecifiedLayer();
@@ -875,12 +881,12 @@ class DavinciModel {
   string om_name_;
 
   uint32_t version_;
-  GeModelPtr ge_model_;
+  GeModelPtr ge_model_;  // release after DavinciModel::Init
 
   bool need_destroy_aicpu_kernel_{false};
   vector<string> out_node_name_;
 
-  map<uint32_t, OpDescPtr> op_list_;
+  map<uint32_t, OpDescPtr> op_list_;  // release after DavinciModel::Init
 
   // data op_desc
   vector<OpDescPtr> data_op_list_;
@@ -975,7 +981,7 @@ class DavinciModel {
   DataDumper data_dumper_;
   uint64_t iterator_count_;
   bool is_l1_fusion_enable_;
-  map<OpDescPtr, void *> saved_task_addrs_;
+  map<OpDescPtr, void *> saved_task_addrs_;  // release after DavinciModel::Init
   void *l1_fusion_addr_ = nullptr;
 
   bool known_node_ = false;

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

@@ -254,6 +254,17 @@ TEST_F(UtestModelManagerDavinciModel, eventlist_success) {
   delete model;
 }
 
+// test Shrink
+TEST_F(UtestModelManagerDavinciModel, shrink_success) {
+  DavinciModel model(0, g_label_call_back);
+  OpDescPtr op_desc_ptr = make_shared<OpDesc>("Cast", "Cast");
+  void *addr = nullptr;
+  rtMalloc(&addr, 128, RT_MEMORY_HBM);
+  model.saved_task_addrs_.emplace(op_desc_ptr, addr);
+  model.Shrink();
+  EXPECT_EQ(model.saved_task_addrs_.isEmpty(), true);
+}
+
 // test rtMalloc
 TEST_F(UtestModelManagerDavinciModel, failed_reset_device) {
   DavinciModel model(0, g_label_call_back);