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

ge/CMakeLists.txt

@@ -144,6 +144,7 @@ set(TRAIN_SRC_LIST
     "graph/passes/atomic_addr_clean_pass.cc"
     "graph/passes/mark_same_addr_pass.cc"
     "graph/passes/mark_graph_unknown_status_pass.cc"
+    "graph/passes/dynamic_single_op_reset_shape_pass.cc"
     "graph/passes/mark_agnostic_pass.cc"
     "graph/partition/dynamic_shape_partition.cc"
     "graph/partition/stage_partition.cc"
@@ -434,6 +435,7 @@ set(INFER_SRC_LIST
     "graph/passes/atomic_addr_clean_pass.cc"
     "graph/passes/mark_same_addr_pass.cc"
     "graph/passes/mark_graph_unknown_status_pass.cc"
+    "graph/passes/dynamic_single_op_reset_shape_pass.cc"
     "graph/passes/mark_agnostic_pass.cc"
     "graph/common/omg_util.cc"
     "graph/common/bcast.cc"

ge/common/CMakeLists.txt

@@ -130,6 +130,7 @@ target_compile_definitions(ge_common_static PRIVATE
     google=ascend_private
     $<IF:$<STREQUAL:${TARGET_SYSTEM_NAME},Windows>,OS_TYPE=WIN,OS_TYPE=0>
     $<$<STREQUAL:${TARGET_SYSTEM_NAME},Windows>:SECUREC_USING_STD_SECURE_LIB=0 NOMINMAX>
+    LOG_CPP
 )
 
 target_compile_options(ge_common_static PRIVATE

ge/common/dump/dump_op.cc

@@ -94,6 +94,9 @@ Status DumpOp::DumpOutput(aicpu::dump::Task &task) {
     for (auto dim : output_descs.at(i).GetShape().GetDims()) {
       output.mutable_shape()->add_dim(dim);
     }
+    for (auto dim : output_descs.at(i).GetOriginShape().GetDims()) {
+      output.mutable_origin_shape()->add_dim(dim);
+    }
     int64_t output_size = 0;
     if (TensorUtils::GetTensorSizeInBytes(output_descs.at(i), output_size) != SUCCESS) {
       GELOGE(PARAM_INVALID, "Get output size filed");
@@ -118,6 +121,9 @@ Status DumpOp::DumpInput(aicpu::dump::Task &task) {
     for (auto dim : input_descs.at(i).GetShape().GetDims()) {
       input.mutable_shape()->add_dim(dim);
     }
+    for (auto dim : input_descs.at(i).GetOriginShape().GetDims()) {
+      input.mutable_origin_shape()->add_dim(dim);
+    }
     int64_t input_size = 0;
     if (TensorUtils::GetTensorSizeInBytes(input_descs.at(i), input_size) != SUCCESS) {
       GELOGE(PARAM_INVALID, "Get output size filed");
@@ -214,8 +220,15 @@ Status DumpOp::LaunchDumpOp() {
   SetOpMappingLoopAddr(global_step_, loop_per_iter_, loop_cond_, op_mapping_info);
   GELOGI("Dump step is %s ,dump path is %s ,in Launch dump op", dump_properties_.GetDumpStep().c_str(),
          dump_path.c_str());
-
+  uint32_t task_id = 0;
+  uint32_t stream_id = 0;
+  rt_ret = rtGetTaskIdAndStreamID(&task_id, &stream_id);
+  if (rt_ret != RT_ERROR_NONE) {
+    GELOGW("call rtGetTaskIdAndStreamID failed, ret = 0x%X", rt_ret);
+  }
   aicpu::dump::Task task;
+  task.set_task_id(task_id);
+  task.set_stream_id(stream_id);
   task.mutable_op()->set_op_name(op_desc_->GetName());
   task.mutable_op()->set_op_type(op_desc_->GetType());
   if (dump_properties_.GetDumpMode() == kDumpOutput) {

ge/common/ge/tbe_plugin_manager.cc

@@ -181,12 +181,19 @@ void TBEPluginManager::GetCustomOpPath(std::string &customop_path) {
 void TBEPluginManager::LoadCustomOpLib() {
   LoadPluginSo(options_);
 
+  std::string fmk_type = std::to_string(domi::TENSORFLOW);
+  auto it = options_.find(ge::FRAMEWORK_TYPE);
+  if (it != options_.end()) {
+   fmk_type = it->second;
+  }
   std::vector<OpRegistrationData> registration_datas = domi::OpRegistry::Instance()->registrationDatas;
   GELOGI("The size of registration_datas is: %zu", registration_datas.size());
   for (OpRegistrationData reg_data : registration_datas) {
-    GELOGD("Begin to register optype: %s, imply_type: %s", reg_data.GetOmOptype().c_str(),
-           TypeUtils::ImplyTypeToSerialString(reg_data.GetImplyType()).c_str());
-    domi::OpRegistry::Instance()->Register(reg_data);
+    if (std::to_string(reg_data.GetFrameworkType()) == fmk_type) {
+      GELOGD("Begin to register optype: %s, imply_type: %s", reg_data.GetOmOptype().c_str(),
+             TypeUtils::ImplyTypeToSerialString(reg_data.GetImplyType()).c_str());
+      (void)domi::OpRegistry::Instance()->Register(reg_data);
+    }
   }
 }
 

ge/common/profiling/ge_profiling.cc

@@ -112,7 +112,6 @@ ge::Status RegProfCtrlCallback(MsprofCtrlCallback func) {
   if (ge::ProfilingManager::Instance().GetMsprofCallback().msprofCtrlCallback != nullptr) {
     GELOGW("Msprof ctrl callback is exist, just ignore it.");
   } else {
-    GELOGI("GE register Msprof ctrl callback.");
     ge::ProfilingManager::Instance().SetMsprofCtrlCallback(func);
   }
   return ge::SUCCESS;
@@ -124,7 +123,6 @@ ge::Status RegProfSetDeviceCallback(MsprofSetDeviceCallback func) {
     return ge::PARAM_INVALID;
   }
   // Pass MsprofSetDeviceCallback to runtime
-  GELOGI("GE pass setdevice callback to runtime.");
   ge::Status rt_ret = rtRegDeviceStateCallback(kRtSetDeviceRegName.c_str(), static_cast<rtDeviceStateCallback>(func));
   if (rt_ret != ge::SUCCESS) {
     GELOGE(rt_ret, "Pass MsprofSetDeviceCallback to runtime failed!");
@@ -158,7 +156,7 @@ ge::Status ProfCommandHandle(ProfCommandHandleType type, void *data, uint32_t le
   if (type != kProfCommandhandleFinalize) {
     GE_CHECK_NOTNULL(data);
   }
-  ProfCommandHandleData *prof_config_param = (ProfCommandHandleData *)data;
+  ProfCommandHandleData *prof_config_param = reinterpret_cast<ProfCommandHandleData *>(data);
   auto iter = kProfCommandTypeMap.find(type);
   if (iter == kProfCommandTypeMap.end()) {
     GELOGW("The prof comand type is invalid.");
@@ -183,7 +181,8 @@ ge::Status ProfCommandHandle(ProfCommandHandleType type, void *data, uint32_t le
   if (type != kProfCommandhandleFinalize) {
     command.module_index = prof_config_param->profSwitch;
   }
-  GELOGI("GE commandhandle execute, Command Type: %d, data type config: 0x%llx", type, command.module_index);
+  GELOGI("GE commandhandle execute, Command Type: %s, data type config: 0x%llx", iter->second.c_str(),
+         command.module_index);
   if (type == kProfCommandhandleStart || type == kProfCommandhandleStop) {
     GELOGI("Profiling device nums:%s , deviceID:[%s]", prof_params[0].c_str(), prof_params[kDeviceListIndex].c_str());
   }

ge/common/profiling/profiling_manager.cc

@@ -38,10 +38,8 @@ const std::string kProfModelUnsubscribe = "prof_model_cancel_subscribe";
 }  // namespace
 
 namespace ge {
-ProfilingManager::ProfilingManager() : is_load_profiling_(false),
-                                       is_execute_profiling_(false),
-                                       is_training_trace_(false),
-                                       subscribe_count_(0) {
+ProfilingManager::ProfilingManager()
+    : is_load_profiling_(false), is_execute_profiling_(false), is_training_trace_(false), subscribe_count_(0) {
   prof_cb_.msprofCtrlCallback = nullptr;
   prof_cb_.msprofReporterCallback = nullptr;
 }
@@ -102,8 +100,8 @@ ge::Status ProfilingManager::InitFromOptions(const Options &options, MsprofGeOpt
       return INTERNAL_ERROR;
     }
     is_execute_profiling_ = true;
-    GELOGI("The profiling in options is %s, %s. origin option: %s", options.profiling_mode.c_str(),
-          prof_conf.options, options.profiling_options.c_str());
+    GELOGI("The profiling in options is %s, %s. origin option: %s", options.profiling_mode.c_str(), prof_conf.options,
+           options.profiling_options.c_str());
   } else {
     (void)mmGetEnv("PROFILING_MODE", env_profiling_mode, MMPA_MAX_PATH);
     (void)mmGetEnv("PROFILING_OPTIONS", prof_conf.options, MSPROF_OPTIONS_DEF_LEN_MAX);
@@ -215,7 +213,7 @@ FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY void ProfilingManager::Profilin
     uint32_t task_id = task.task_id;
     uint32_t stream_id = task.stream_id;
     std::string shape_type = task.shape_type;
-    uint64_t cur_iter_num = task.cur_iter_num;
+    int64_t cur_iter_num = task.cur_iter_num;
     data = model_name.append(" ")
                      .append(op_name).append(" ")
                      .append(std::to_string(block_dim)).append(" ")
@@ -809,32 +807,46 @@ FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY void ProfilingManager::GetFpBpP
   if (!fp_point_.empty() && !bp_point_.empty()) {
     fp_point = fp_point_;
     bp_point = bp_point_;
-    GELOGI("Bp Fp have been initialized in env or options. bp_point: %s, fp_point: %s", bp_point.c_str(), fp_point.c_str());
+    GELOGI("Bp Fp have been initialized in env or options. bp_point: %s, fp_point: %s", bp_point.c_str(),
+           fp_point.c_str());
     return;
   }
   // ProfApi mode and training trace is set
-  try {
-    char env_profiling_options[MSPROF_OPTIONS_DEF_LEN_MAX] = { 0x00 };
+  // Parse options first
+  char env_profiling_options[MSPROF_OPTIONS_DEF_LEN_MAX] = { 0x00 };
+  bool is_profiling_valid = false;
+  std::string profiling_options;
+  if (ge::GetContext().GetOption(OPTION_EXEC_PROFILING_OPTIONS, profiling_options) == SUCCESS &&
+      !profiling_options.empty()) {
+    is_profiling_valid = true;
+  } else {
     INT32 ret = mmGetEnv("PROFILING_OPTIONS", env_profiling_options, MSPROF_OPTIONS_DEF_LEN_MAX);
     if (ret != EN_OK) {
       GELOGI("PROFILING_OPTIONS env is not exist.");
       return;
     }
     GELOGI("Parse env PROFILING_OPTIONS:%s.", env_profiling_options);
-    Json prof_options = Json::parse(env_profiling_options);
+    profiling_options = env_profiling_options;
+    is_profiling_valid = true;
+  }
+  if (is_profiling_valid) {
+    try {
+      Json prof_options = Json::parse(profiling_options);
 
-    fp_point_ = prof_options[kFpPoint];
-    bp_point_ = prof_options[kBpPoint];
+      fp_point_ = prof_options[kFpPoint];
+      bp_point_ = prof_options[kBpPoint];
 
-    fp_point = fp_point_;
-    bp_point = bp_point_;
-    if (!fp_point_.empty() && !bp_point_.empty()) {
-      GELOGI("Training trace bp fp is set, bp_point:%s, fp_point:%s.", bp_point_.c_str(), fp_point_.c_str());
+      fp_point = fp_point_;
+      bp_point = bp_point_;
+      if (!fp_point_.empty() && !bp_point_.empty()) {
+        GELOGI("Training trace bp fp is set, bp_point:%s, fp_point:%s.", bp_point_.c_str(), fp_point_.c_str());
+      }
+    } catch (...) {
+      GELOGW("Json prof options is invalid.");
+      return;
     }
-  } catch (...) {
-    GELOGE(FAILED, "Json prof options is invalid.");
-    return;
   }
+  
   return;
 }
 

ge/common/profiling/profiling_manager.h

@@ -36,21 +36,21 @@ using Json = nlohmann::json;
 namespace {
   const std::string GE_PROFILING_MODULE = "Framework";
   // DataTypeConfig MASK
-  #define PROF_ACL_API_MASK                0x0001
-  #define PROF_TASK_TIME_MASK              0x0002
-  #define PROF_AICORE_METRICS_MASK         0x0004
-  #define PROF_AICPU_TRACE_MASK            0x0008
-  #define PROF_MODEL_EXECUTE_MASK          0x0010
-  #define PROF_RUNTIME_API_MASK            0x0020
-  #define PROF_RUNTIME_TRACE_MASK          0x0040
-  #define PROF_SCHEDULE_TIMELINE_MASK      0x0080
-  #define PROF_SCHEDULE_TRACE_MASK         0x0100
-  #define PROF_AIVECTORCORE_METRICS_MASK   0x0200
-  #define PROF_SUBTASK_TIME_MASK           0x0400
-  #define PROF_TRAINING_TRACE_MASK         0x0800
-  #define PROF_HCCL_TRACE_MASK             0x1000
-  #define PROF_DATA_PROCESS_MASK           0x2000
-  #define PROF_MODEL_LOAD_MASK             0x8000000000000000
+  const uint64_t PROF_ACL_API_MASK = 0x0001;
+  const uint64_t PROF_TASK_TIME_MASK = 0x0002;
+  const uint64_t PROF_AICORE_METRICS_MASK = 0x0004;
+  const uint64_t PROF_AICPU_TRACE_MASK = 0x0008;
+  const uint64_t PROF_MODEL_EXECUTE_MASK = 0x0010;
+  const uint64_t PROF_RUNTIME_API_MASK = 0x0020;
+  const uint64_t PROF_RUNTIME_TRACE_MASK = 0x0040;
+  const uint64_t PROF_SCHEDULE_TIMELINE_MASK = 0x0080;
+  const uint64_t PROF_SCHEDULE_TRACE_MASK = 0x0100;
+  const uint64_t PROF_AIVECTORCORE_METRICS_MASK = 0x0200;
+  const uint64_t PROF_SUBTASK_TIME_MASK = 0x0400;
+  const uint64_t PROF_TRAINING_TRACE_MASK = 0x0800;
+  const uint64_t PROF_HCCL_TRACE_MASK = 0x1000;
+  const uint64_t PROF_DATA_PROCESS_MASK = 0x2000;
+  const uint64_t PROF_MODEL_LOAD_MASK = 0x8000000000000000;
 
 }  // namespace
 namespace ge {
@@ -80,7 +80,8 @@ class FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY ProfilingManager {
   bool ProfilingTrainingTraceOn() const { return is_training_trace_; }
   bool ProfilingModelLoadOn() const { return is_load_profiling_; }
   bool ProfilingModelExecuteOn() const;
-  bool ProfilingOn() const { return is_load_profiling_ && is_execute_profiling_; } // is_execute_profiling_ only used by ge option and env
+  // is_execute_profiling_ only used by ge option and env
+  bool ProfilingOn() const { return is_load_profiling_ && is_execute_profiling_; }
   void ReportProfilingData(uint32_t model_id, const std::vector<TaskDescInfo> &task_desc_info,
                            const std::vector<ComputeGraphDescInfo> &compute_graph_desc_info);
   void ProfilingTaskDescInfo(uint32_t model_id, const std::vector<TaskDescInfo> &task_desc_info,

ge/common/proto/op_mapping_info.proto

@@ -15,6 +15,7 @@ message Output {
     int32 original_output_data_type = 7;
     int32 original_output_format = 8;
     uint64 size = 9;
+    Shape origin_shape = 10;
 }
 
 message Input {
@@ -23,6 +24,7 @@ message Input {
     Shape shape = 3;
     uint64 address = 4;
     uint64 size = 5;
+    Shape origin_shape = 6;
 }
 
 enum BufferType {

ge/executor/CMakeLists.txt

@@ -173,6 +173,7 @@ target_compile_definitions(ge_executor PRIVATE
     google=ascend_private
     $<IF:$<STREQUAL:${TARGET_SYSTEM_NAME},Windows>,OS_TYPE=WIN,OS_TYPE=0>
     $<$<STREQUAL:${TARGET_SYSTEM_NAME},Windows>:SECUREC_USING_STD_SECURE_LIB=0 NOMINMAX>
+    LOG_CPP
 )
 
 target_include_directories(ge_executor PRIVATE

ge/executor/ge_executor.cc

@@ -209,19 +209,6 @@ bool IsDynmaicDimsSizeMatchModel(const vector<uint64_t> cur_dynamic_dims,
 
 namespace ge {
 bool GeExecutor::isInit_ = false;
-class ModelListenerAdapter : public ModelListener {
- public:
-  domi::Status OnComputeDone(uint32_t model_id, uint32_t dataIndex, uint32_t resultCode,
-                             std::vector<ge::OutputTensorInfo> &outputs) {
-    if (listener == nullptr) {
-      GELOGE(ge::FAILED, "listener is null.");
-      return FAILED;
-    }
-    return listener->OnComputeDone(model_id, dataIndex, resultCode, outputs);
-  }
-
-  std::shared_ptr<ge::ModelListener> listener;
-};
 
 static void InitOpsProtoManger() {
   string opsproto_path;
@@ -573,60 +560,6 @@ Status GeExecutor::SetDynamicAippData(uint32_t model_id, void *dynamic_input_add
   return SUCCESS;
 }
 
-// Load model
-Status GeExecutor::LoadModelOffline(uint32_t &model_id, const std::string &path, const std::string &key,
-                                    int32_t priority, std::shared_ptr<ge::ModelListener> listener) {
-  GELOGI("load model offline begin.");
-  if (!isInit_) {
-    GELOGE(ACL_ERROR_GE_EXEC_NOT_INIT, "GeExecutor has not been initialized!");
-    return ACL_ERROR_GE_EXEC_NOT_INIT;
-  }
-
-  string filePath = RealPath(path.c_str());
-  if (filePath.empty()) {
-    GELOGE(ACL_ERROR_GE_EXEC_MODEL_PATH_INVALID,
-           "File path is invalid. please check your text file '%s'.", path.c_str());
-    return ACL_ERROR_GE_EXEC_MODEL_PATH_INVALID;
-  }
-
-  std::shared_ptr<ModelListenerAdapter> listener_adapter = MakeShared<ModelListenerAdapter>();
-  if (listener_adapter == nullptr) {
-    GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION, "ModelListenerAdapter make shared failed!");
-    return ACL_ERROR_GE_MEMORY_ALLOCATION;
-  }
-  listener_adapter->listener = listener;
-
-  Status ret = GraphLoader::LoadModelFromFile(path, key, priority, listener_adapter, model_id);
-  if (ret != SUCCESS) {
-    GELOGE(ret, "[GeExecutor] LoadModelFromFile failed");
-    return ACL_ERROR_GE_LOAD_MODEL;
-  }
-  return SUCCESS;
-}
-
-Status GeExecutor::LoadModel(uint32_t &model_id, const ModelData &model_data,
-                             std::shared_ptr<ge::ModelListener> listener) {
-  GELOGI("Load model begin.");
-  if (!isInit_) {
-    GELOGE(ACL_ERROR_GE_EXEC_NOT_INIT, "GeExecutor has not been initialized!");
-    return ACL_ERROR_GE_EXEC_NOT_INIT;
-  }
-
-  std::shared_ptr<ModelListenerAdapter> listener_adapter = MakeShared<ModelListenerAdapter>();
-  if (listener_adapter == nullptr) {
-    GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION, "ModelListenerAdapter make shared failed!");
-    return ACL_ERROR_GE_MEMORY_ALLOCATION;
-  }
-  listener_adapter->listener = listener;
-
-  Status ret = GraphLoader::LoadModel(model_data, listener_adapter, model_id);
-  if (ret != SUCCESS) {
-    GELOGE(ret, "[GeExecutor] LoadModel failed.");
-    return ACL_ERROR_GE_LOAD_MODEL;
-  }
-  return ret;
-}
-
 Status GeExecutor::UnloadModel(uint32_t model_id) {
   GELOGD("unload model %u begin.", model_id);
   if (!isInit_) {
@@ -659,21 +592,6 @@ Status GeExecutor::UnloadModel(uint32_t model_id) {
   return SUCCESS;
 }
 
-Status GeExecutor::RunModel(const ge::RunModelData &input_data, ge::RunModelData &output_data) {
-  GELOGI("run model begin.");
-  if (!isInit_) {
-    GELOGE(ACL_ERROR_GE_EXEC_NOT_INIT, "GeExecutor has not been initialized!");
-    return ACL_ERROR_GE_EXEC_NOT_INIT;
-  }
-
-  InputData inputs;
-  GetDomiInputData(input_data, inputs);
-  OutputData outputs;
-  GetDomiOutputData(output_data, outputs);
-
-  return GraphExecutor::DataInput(inputs, outputs);
-}
-
 // Get input and output descriptor
 Status GeExecutor::GetModelDescInfo(uint32_t model_id, std::vector<ge::TensorDesc> &input_desc,
                                     std::vector<ge::TensorDesc> &output_desc, bool new_model_desc) {

ge/executor/proto/op_mapping_info.proto

@@ -15,6 +15,7 @@ message Output {
     int32 original_output_data_type = 7;
     int32 original_output_format = 8;
     uint64 size = 9;
+    Shape origin_shape = 10;
 }
 
 message Input {
@@ -23,6 +24,7 @@ message Input {
     Shape shape = 3;
     uint64 address = 4;
     uint64 size = 5;
+    Shape origin_shape = 6;
 }
 
 enum BufferType {

ge/ge_inference.mk

@@ -109,6 +109,7 @@ OMG_HOST_SRC_FILES := \
     graph/passes/atomic_addr_clean_pass.cc \
     graph/passes/mark_same_addr_pass.cc \
     graph/passes/mark_graph_unknown_status_pass.cc \
+    graph/passes/dynamic_single_op_reset_shape_pass.cc \
     graph/passes/mark_agnostic_pass.cc \
     graph/common/omg_util.cc \
     graph/common/bcast.cc \

ge/ge_local_engine/CMakeLists.txt

@@ -203,6 +203,7 @@ target_compile_options(ge_local_opskernel_builder_static PRIVATE
 
 target_compile_definitions(ge_local_opskernel_builder_static PRIVATE
     google=ascend_private
+    LOG_CPP
 )
 
 target_include_directories(ge_local_opskernel_builder_static PRIVATE

ge/ge_runner.mk

@@ -113,6 +113,7 @@ LIBGE_LOCAL_SRC_FILES := \
     graph/passes/atomic_addr_clean_pass.cc \
     graph/passes/mark_same_addr_pass.cc \
     graph/passes/mark_graph_unknown_status_pass.cc \
+    graph/passes/dynamic_single_op_reset_shape_pass.cc \
     graph/passes/mark_agnostic_pass.cc \
     graph/partition/dynamic_shape_partition.cc \
     graph/partition/stage_partition.cc \

ge/generator/ge_generator.cc

@@ -47,6 +47,8 @@ const char *const kEngineNameDefault = "default";
 const char *const kVectorEngine = "VectorEngine";
 const char *const kAIcoreEngine = "AIcoreEngine";
 const char *const kFileNameSuffix = "online";
+const size_t kDynamicDimSize = 1;
+const int64_t kDynamicDimValue = -2;
 
 std::map<ge::OpEngineType, std::string> engine_type_map{
     {ge::ENGINE_SYS, kEngineNameDefault}, {ge::ENGINE_AICORE, kAIcoreEngine}, {ge::ENGINE_VECTOR, kVectorEngine}};
@@ -248,6 +250,43 @@ static void GetOpsProtoPath(string &opsproto_path) {
   opsproto_path = (path_base + "ops/op_proto/custom/" + ":") + (path_base + "ops/op_proto/built-in/");
 }
 
+static Status CheckShapeReset(const OpDescPtr &op_desc, bool &change_shape_flag) {
+  GE_CHECK_NOTNULL_EXEC(op_desc, return PARAM_INVALID);
+  change_shape_flag = false;
+  for (size_t i = 0; i < op_desc->GetAllInputsDesc().size(); i++) {
+    auto input_desc = op_desc->MutableInputDesc(static_cast<uint32_t>(i));
+    GE_CHECK_NOTNULL(input_desc);
+    // pass scalar input desc
+    auto dims = input_desc->GetShape().GetDims();
+    if (dims.size() == kDynamicDimSize && dims[0] == kDynamicDimValue) {
+      change_shape_flag = true;
+    }
+  }
+  return SUCCESS;
+}
+
+static void ResetTensorVecShape(const vector<GeTensor> &inputs, vector<GeTensor> &inputs_dynamic) {
+  for (auto input : inputs) {
+    auto input_desc = input.GetTensorDesc();
+    GeShape shape_ori = input_desc.GetShape();
+
+    std::vector<int64_t> dynamic_shape_dims = {kDynamicDimValue};
+    GeShape dynamic_shape(dynamic_shape_dims);
+
+    ge::GeTensor inputTensor;
+    ge::GeTensorDesc desc(input_desc);
+
+    bool is_const = false;
+    (void)AttrUtils::GetBool(input_desc, CONST_ATTR_NAME_INPUT, is_const);
+    if (!is_const && shape_ori.GetDims().size() > 0) {
+      desc.SetShape(dynamic_shape);
+    }
+
+    inputTensor.SetTensorDesc(desc);
+    inputs_dynamic.push_back(inputTensor);
+  }
+}
+
 class GeGenerator::Impl {
  public:
   Impl(OmgContext &omg_context) : omg_context_(omg_context) {}
@@ -638,7 +677,18 @@ Status GeGenerator::BuildSingleOp(OpDescPtr &op_desc, const vector<GeTensor> &in
   }
   GeModelPtr &ge_model = name_to_ge_model.begin()->second;
   GELOGD("The opType in op_desc_tmp is [%s]", op_desc_tmp->GetType().c_str());
-  GE_CHK_STATUS_RET_NOLOG(impl_->SaveParams(ge_model, op_desc_tmp->GetType(), op_attrs, inputs, outputs));
+
+  bool dynamic_flag = false;
+  if (CheckShapeReset(op_desc, dynamic_flag) == SUCCESS && dynamic_flag) {
+    vector<GeTensor> inputs_dynamic;
+    vector<GeTensor> outputs_dynamic;
+    ResetTensorVecShape(inputs, inputs_dynamic);
+    ResetTensorVecShape(outputs, outputs_dynamic);
+    GE_CHK_STATUS_RET_NOLOG(
+      impl_->SaveParams(ge_model, op_desc_tmp->GetType(), op_attrs, inputs_dynamic, outputs_dynamic));
+  } else {
+    GE_CHK_STATUS_RET_NOLOG(impl_->SaveParams(ge_model, op_desc_tmp->GetType(), op_attrs, inputs, outputs));
+  }
   GE_CHK_STATUS_RET_NOLOG(impl_->SaveModel(model_file_name, ge_model, model_buff));
   return SUCCESS;
 }

ge/graph/build/memory/CMakeLists.txt

@@ -18,6 +18,7 @@ target_compile_options(ge_memory PRIVATE
 
 target_compile_definitions(ge_memory PRIVATE
     google=ascend_private
+    LOG_CPP
 )
 
 target_link_libraries(ge_memory PRIVATE

ge/graph/build/memory/graph_mem_assigner.cc

@@ -99,7 +99,7 @@ Status GraphMemoryAssigner::AssignMemory() {
   MemoryOffset memory_offset(RT_MEMORY_HBM, mem_assigner->GetMemOffset());
   memory_offset_.emplace(RT_MEMORY_HBM, memory_offset);
 
-  if (mem_assigner->GetP2PMemOffset() > 0) {
+  if (mem_assigner->GetP2PMemOffset() >= 0) {
     MemoryOffset p2p_memory_offset(RT_MEMORY_P2P_DDR, mem_assigner->GetP2PMemOffset());
     memory_offset_.emplace(RT_MEMORY_P2P_DDR, p2p_memory_offset);
   }

ge/graph/build/stream_graph_optimizer.cc

@@ -48,26 +48,41 @@ void StreamGraphOptimizer::RefreshNodeId(const ComputeGraphPtr &comp_graph, Grap
   }
 }
 
-bool StreamGraphOptimizer::IsSameStreamId(const ComputeGraphPtr &comp_graph) {
+bool StreamGraphOptimizer::IsSameStreamIdOrBatchLabel(const ComputeGraphPtr &comp_graph) {
   if (comp_graph == nullptr) {
     return false;
   }
   std::set<int64_t> stream_set;
+  std::set<std::string> label_set;
   for (const ge::NodePtr &cur_node : comp_graph->GetDirectNode()) {
     GE_IF_BOOL_EXEC(cur_node->GetOpDesc() == nullptr, continue);
     int64_t stream_id = cur_node->GetOpDesc()->GetStreamId();
     if (stream_id == kInvalidStream) {
       continue;
     }
-    GELOGD("Node %s in subgraph %s stream id is: %ld, node num: %zu", cur_node->GetName().c_str(),
-           comp_graph->GetName().c_str(), stream_id, comp_graph->GetDirectNodesSize());
     stream_set.insert(stream_id);
+
+    std::string batch_label;
+    if (AttrUtils::GetStr(cur_node->GetOpDesc(), ATTR_NAME_BATCH_LABEL, batch_label)) {
+      label_set.insert(batch_label);
+    } else {
+      GELOGD("Node %s[%s] has no batch label, subgraph %s, stream id: %ld", cur_node->GetName().c_str(),
+             cur_node->GetType().c_str(), comp_graph->GetName().c_str(), stream_id);
+      continue;
+    }
+
+    GELOGD("Node %s in subgraph %s stream id: %ld, node num: %zu", cur_node->GetName().c_str(),
+           comp_graph->GetName().c_str(), stream_id, comp_graph->GetDirectNodesSize());
   }
-  if (stream_set.size() > 1) {
-    GELOGI("Nodes of graph: %s have different stream id, node num: %zu, different stream num: %zu.",
+  if (stream_set.size() > 1 || label_set.size() > 1) {
+    GELOGI("Nodes of graph: %s have different stream id or batch_label, node num: %zu, different stream num: %zu.",
            comp_graph->GetName().c_str(), comp_graph->GetDirectNodesSize(), stream_set.size());
     return false;
   }
+
+  if (!label_set.empty()) {
+    (void)AttrUtils::SetStr(comp_graph, ATTR_NAME_BATCH_LABEL, *label_set.begin());
+  }
   return true;
 }
 
@@ -99,8 +114,8 @@ Status StreamGraphOptimizer::OptimizeStreamedSubGraph(const ComputeGraphPtr &com
           continue;
         }
 
-        if (!IsSameStreamId(subgraph)) {
-          GELOGI("There are more than one stream in subgraph %s", subgraph->GetName().c_str());
+        if (!IsSameStreamIdOrBatchLabel(subgraph)) {
+          GELOGI("There are more than one stream or batch_label in subgraph %s", subgraph->GetName().c_str());
           continue;
         }
         OpDescPtr op_desc = nodes.at(0)->GetOpDesc();
@@ -112,9 +127,11 @@ Status StreamGraphOptimizer::OptimizeStreamedSubGraph(const ComputeGraphPtr &com
           return FAILED;
         }
         run_context.stream = run_context.graphStreamList[stream_id];
-        GELOGD("Subgraph has same stream id, subgraph: %s, engine_name: %s, stream_id: %ld, rtstream: %lu.",
-               subgraph->GetName().c_str(), engine_name.c_str(), stream_id,
-               static_cast<uint64_t>(reinterpret_cast<uintptr_t>(run_context.stream)));
+	std::string batch_label;
+	(void)AttrUtils::GetStr(subgraph, ATTR_NAME_BATCH_LABEL, batch_label);
+        GELOGD("Subgraph has same stream id, subgraph: %s, engine_name: %s, stream_id: %ld, rtstream: %lu, "
+	       "batch_label: %s", subgraph->GetName().c_str(), engine_name.c_str(), stream_id,
+               static_cast<uint64_t>(reinterpret_cast<uintptr_t>(run_context.stream)), batch_label.c_str());
         for (auto iter = graph_optimizers.begin(); iter != graph_optimizers.end(); ++iter) {
           GE_CHECK_NOTNULL(*iter);
           Status ret = (*iter)->OptimizeStreamGraph(*subgraph, run_context);

ge/graph/build/stream_graph_optimizer.h

@@ -41,7 +41,7 @@ class StreamGraphOptimizer {
  private:
   void RefreshNodeId(const ComputeGraphPtr &comp_graph, Graph2SubGraphInfoList &subgraph_map);
 
-  bool IsSameStreamId(const ComputeGraphPtr &comp_graph);
+  bool IsSameStreamIdOrBatchLabel(const ComputeGraphPtr &comp_graph);
 };
 }  // namespace ge
 #endif  // GE_GRAPH_BUILD_OPTIMIZE_STREAM_GRAPH_H_

ge/graph/build/task_generator.cc

@@ -567,7 +567,7 @@ Status TaskGenerator::MarkFirstAndLastOps(const vector<OpDescPtr> &ops, bool is_
       continue;
     }
     string op_type = op_desc->GetType();
-    if (!is_single_stream && (!op_desc->GetSubgraphInstanceNames().empty() || separator_types.count(op_type) != 0)) {
+    if (!op_desc->GetSubgraphInstanceNames().empty() || separator_types.count(op_type) != 0) {
       continuous_op_lists.emplace_back(vector<OpDescPtr>());
     } else {
       continuous_op_lists.back().emplace_back(op_desc);

ge/graph/load/graph_loader.cc

@@ -122,14 +122,14 @@ Status GraphLoader::LoadDataFromFile(const std::string &path, const std::string
                                      ModelData &model_data) {
   Status ret;
   if (!CheckInputPathValid(path)) {
-    GELOGE(GE_EXEC_MODEL_PATH_INVALID, "model path is invalid: %s", path.c_str());
-    return GE_EXEC_MODEL_PATH_INVALID;
+    GELOGE(ACL_ERROR_GE_EXEC_MODEL_PATH_INVALID, "model path is invalid: %s", path.c_str());
+    return ACL_ERROR_GE_EXEC_MODEL_PATH_INVALID;
   }
 
   GELOGI("Load model begin, model path is: %s", path.c_str());
   if (!key_path.empty() && !CheckInputPathValid(key_path)) {
-    GELOGE(GE_EXEC_MODEL_KEY_PATH_INVALID, "decrypt_key path is invalid: %s", key_path.c_str());
-    return GE_EXEC_MODEL_KEY_PATH_INVALID;
+    GELOGE(ACL_ERROR_GE_PARAM_INVALID, "decrypt_key path is invalid: %s", key_path.c_str());
+    return ACL_ERROR_GE_PARAM_INVALID;
   }
 
   ret = DavinciModelParser::LoadFromFile(path.c_str(), key_path.c_str(), priority, model_data);
@@ -144,63 +144,6 @@ Status GraphLoader::LoadDataFromFile(const std::string &path, const std::string
     return SUCCESS;
 }
 
-Status GraphLoader::LoadModelFromFile(const std::string &path, const std::string &key_path, int32_t priority,
-                                      const std::shared_ptr<ModelListener> &listener, uint32_t &model_id) {
-  Status ret;
-  ModelData model_data;
-  ret = LoadDataFromFile(path, key_path, priority, model_data);
-  if (ret != SUCCESS) {
-    GELOGE(ret, "LoadModelFromFile: Load failed. ret = %u", ret);
-    if (model_data.model_data != nullptr) {
-      delete[] static_cast<char *>(model_data.model_data);
-      model_data.model_data = nullptr;
-    }
-    return ret;
-  }
-
-  ret = LoadModel(model_data, listener, model_id);
-  if (ret != SUCCESS) {
-    GELOGE(ret, "LoadModel: Load failed. ret = %u", ret);
-    if (model_data.model_data != nullptr) {
-      delete[] static_cast<char *>(model_data.model_data);
-      model_data.model_data = nullptr;
-    }
-  }
-
-  if (model_data.model_data != nullptr) {
-    delete[] static_cast<char *>(model_data.model_data);
-    model_data.model_data = nullptr;
-  }
-
-  return ret;
-}
-
-Status GraphLoader::LoadModel(const ModelData &model_data, const std::shared_ptr<ModelListener> &listener,
-                              uint32_t &model_id) {
-  GELOGI("Load model begin, model_id:%u.", model_id);
-
-  // For GeOp, Open Device 0 here.
-  GE_CHK_RT_RET(rtSetDevice(0));
-  auto model_manager = ModelManager::GetInstance();
-  GE_CHECK_NOTNULL(model_manager);
-  Status ret = model_manager->LoadModelOffline(model_id, model_data, listener);
-  if (ret != SUCCESS) {
-    GE_CHK_RT(rtDeviceReset(0));
-    GELOGE(ret, "LoadModel: Load failed.");
-    return ret;
-  }
-  ret = model_manager->Start(model_id);
-  if (ret != SUCCESS) {
-    if (model_manager->Unload(model_id) != SUCCESS) {
-      GELOGE(FAILED, "LoadModel: Unload failed while trying to unload after a failed start.");
-    }
-    GELOGE(ret, "LoadModel: Start failed.");
-    return ret;
-  }
-  GELOGI("LoadModel: Start model success, model_id:%u.", model_id);
-  return SUCCESS;
-}
-
 Status GraphLoader::CommandHandle(const Command &command) {
   try {
     auto model_manager = ModelManager::GetInstance();
@@ -225,16 +168,16 @@ Status GraphLoader::CommandHandle(const Command &command) {
 }
 
 Status GraphLoader::LoadModelFromData(uint32_t &model_id, const ModelData &model_data, void *dev_ptr,
-                                      size_t memsize, void *weight_ptr, size_t weightsize) {
+                                      size_t mem_size, void *weight_ptr, size_t weight_size) {
   GELOGI("Load model begin, model_id:%u.", model_id);
   // For ACL, Open Device from App.
   auto model_manager = ModelManager::GetInstance();
   GE_CHECK_NOTNULL(model_manager);
   Status ret = model_manager->LoadModelOffline(
-      model_id, model_data, nullptr, dev_ptr, memsize, weight_ptr, weightsize);
+      model_id, model_data, nullptr, dev_ptr, mem_size, weight_ptr, weight_size);
   if (ret != SUCCESS) {
-    GELOGE(ret, "Load model failed, model_id:%u.", model_id);
-    return ret;
+    GELOGE(ACL_ERROR_GE_LOAD_MODEL, "Load model failed, model_id:%u.", model_id);
+    return ACL_ERROR_GE_LOAD_MODEL;
   }
   GELOGI("Load model success, model_id:%u.", model_id);
   return SUCCESS;
@@ -259,8 +202,8 @@ Status GraphLoader::LoadModelWithQ(uint32_t &model_id, const ModelData &model_da
   GE_CHECK_NOTNULL(model_manager);
   Status ret = model_manager->LoadModelWithQ(model_id, model_data, input_queue_ids, output_queue_ids);
   if (ret != SUCCESS) {
-    GELOGE(ret, "Load model with queue failed, model_id:%u.", model_id);
-    return ret;
+    GELOGE(ACL_ERROR_GE_LOAD_MODEL, "Load model with queue failed, model_id:%u.", model_id);
+    return ACL_ERROR_GE_LOAD_MODEL;
   }
 
   GELOGI("Load model with queue success, model_id:%u.", model_id);

ge/graph/load/graph_loader.h

@@ -44,12 +44,6 @@ class GraphLoader {
 
   static Status GetMaxUsedMemory(uint32_t model_id, uint64_t &max_size);
 
-  static Status LoadModel(const ModelData &model_data, const std::shared_ptr<ModelListener> &listener,
-                          uint32_t &model_id);
-
-  static Status LoadModelFromFile(const std::string &path, const std::string &key_path, int32_t priority,
-                                  const std::shared_ptr<ModelListener> &listener, uint32_t &model_id);
-
   static Status CommandHandle(const Command &command);
 
   static Status GetMemoryInfo(int64_t &free);

ge/graph/load/new_model_manager/data_dumper.cc

@@ -319,6 +319,9 @@ Status DataDumper::GenerateOutput(aicpu::dump::Output &output, const OpDesc::Vis
   for (auto dim : tensor_descs.at(index).GetShape().GetDims()) {
     output.mutable_shape()->add_dim(dim);
   }
+  for (auto dim : tensor_descs.at(index).GetOriginShape().GetDims()) {
+    output.mutable_origin_shape()->add_dim(dim);
+  }
   int64_t output_size = 0;
   if (TensorUtils::GetTensorSizeInBytes(tensor_descs.at(index), output_size) != SUCCESS) {
     GELOGE(PARAM_INVALID, "Get output size filed");
@@ -476,6 +479,9 @@ Status DataDumper::GenerateInput(aicpu::dump::Input &input, const OpDesc::Vistor
   for (auto dim : tensor_descs.at(index).GetShape().GetDims()) {
     input.mutable_shape()->add_dim(dim);
   }
+  for (auto dim : tensor_descs.at(index).GetOriginShape().GetDims()) {
+    input.mutable_origin_shape()->add_dim(dim);
+  }
   int64_t input_size = 0;
   if (AttrUtils::GetInt(tensor_descs.at(index), ATTR_NAME_INPUT_ORIGIN_SIZE, input_size)) {
     GELOGI("Get aipp input size according to attr is %ld", input_size);

ge/graph/load/new_model_manager/davinci_model.cc

@@ -289,8 +289,8 @@ Status DavinciModel::InitWeightMem(void *dev_ptr, void *weight_ptr, size_t weigh
     if (weight_ptr == nullptr) {
       weights_mem_base_ = MallocWeightsMem(weights_size);
       if (weights_mem_base_ == nullptr) {
-        GELOGE(GE_EXEC_ALLOC_WEIGHT_MEM_FAILED, "Alloc weight memory failed. size: %zu", weights_size);
-        return GE_EXEC_ALLOC_WEIGHT_MEM_FAILED;
+        GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION, "Alloc weight memory failed. size: %zu", weights_size);
+        return ACL_ERROR_GE_MEMORY_ALLOCATION;
       }
       is_inner_weight_base_ = true;
     }
@@ -307,8 +307,8 @@ Status DavinciModel::InitWeightMem(void *dev_ptr, void *weight_ptr, size_t weigh
 
 Status DavinciModel::InitFeatureMapAndP2PMem(void *dev_ptr, size_t mem_size) {
   if (is_feature_map_mem_has_inited_) {
-    GELOGE(FAILED, "call InitFeatureMapMem more than once .");
-    return FAILED;
+    GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION, "call InitFeatureMapMem more than once .");
+    return ACL_ERROR_GE_MEMORY_ALLOCATION;
   }
   is_feature_map_mem_has_inited_ = true;
 
@@ -316,8 +316,8 @@ Status DavinciModel::InitFeatureMapAndP2PMem(void *dev_ptr, size_t mem_size) {
   std::size_t p2p_data_size = P2PMemInfos().at(RT_MEMORY_P2P_DDR).memory_size;
 
   if ((dev_ptr != nullptr) && (mem_size < TotalMemSize())) {
-    GELOGE(FAILED, "Invalid mem param: mem_size=%zu totalsize=%zu.", mem_size, TotalMemSize());
-    return FAILED;
+    GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION, "Invalid mem param: mem_size=%zu totalsize=%zu.", mem_size, TotalMemSize());
+    return ACL_ERROR_GE_MEMORY_ALLOCATION;
   }
 
   mem_base_ = static_cast<uint8_t *>(dev_ptr);
@@ -327,8 +327,8 @@ Status DavinciModel::InitFeatureMapAndP2PMem(void *dev_ptr, size_t mem_size) {
   if (TotalMemSize() && mem_base_ == nullptr) {
     mem_base_ = MallocFeatureMapMem(data_size);
     if (mem_base_ == nullptr) {
-      GELOGE(GE_EXEC_ALLOC_FEATURE_MAP_MEM_FAILED, "Alloc feature map memory failed. size: %zu", data_size);
-      return GE_EXEC_ALLOC_FEATURE_MAP_MEM_FAILED;
+      GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION, "Alloc feature map memory failed. size: %zu", data_size);
+      return ACL_ERROR_GE_MEMORY_ALLOCATION;
     }
     GEEVENT("[IMAS]InitFeatureMapAndP2PMem graph_%u MallocMemory type[F] memaddr[%p] mem_size[%zu]",
             runtime_param_.graph_id, mem_base_, data_size);
@@ -343,8 +343,8 @@ Status DavinciModel::InitFeatureMapAndP2PMem(void *dev_ptr, size_t mem_size) {
   if (p2p_data_size != 0) {
     p2p_mem_base_ = MallocP2PMem(p2p_data_size);
     if (p2p_mem_base_ == nullptr) {
-      GELOGE(GE_EXEC_ALLOC_P2P_MEM_FAILED, "Alloc p2p memory failed,size: %zu", p2p_data_size);
-      return GE_EXEC_ALLOC_P2P_MEM_FAILED;
+      GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION, "Alloc p2p memory failed,size: %zu", p2p_data_size);
+      return ACL_ERROR_GE_MEMORY_ALLOCATION;
     }
     GELOGI("InitFeatureMapAndP2PMem graph_%u MallocMemory type[F] memaddr[%p] mem_size[%zu]", runtime_param_.graph_id,
            p2p_mem_base_, p2p_data_size);
@@ -710,6 +710,7 @@ Status DavinciModel::Init(void *dev_ptr, size_t mem_size, void *weight_ptr, size
   }
 
   // collect profiling for ge
+  GE_CHK_STATUS_RET(InitModelProfile(), "Init model profile failed");
   auto &profiling_manager = ProfilingManager::Instance();
   if (profiling_manager.ProfilingModelLoadOn()) {
     Status p_ret = ReportProfilingData();
@@ -970,7 +971,7 @@ Status DavinciModel::InitDataOp(const NodePtr &node, uint32_t &data_op_index, ma
   uint32_t parent_index = 0;  // Ignore subgraph Data Node.
   if (AttrUtils::GetInt(op_desc, ATTR_NAME_PARENT_NODE_INDEX, parent_index)) {
     GELOGI("Init zero copy by subgraph Data node: %s.", op_desc->GetName().c_str());
-    return InitInputBatchLabel(node);
+    return SUCCESS;
   }
 
   data_op_list_.push_back(op_desc);
@@ -1011,10 +1012,6 @@ Status DavinciModel::InitDataOp(const NodePtr &node, uint32_t &data_op_index, ma
   }
 
   data_op_index++;
-  if (InitInputZeroCopyNodes(node) != SUCCESS) {
-    GELOGE(PARAM_INVALID, "Input zero copy nodes init failed!");
-    return PARAM_INVALID;
-  }
   return SUCCESS;
 }
 
@@ -1036,39 +1033,6 @@ void DavinciModel::AdjustDataOpList(const map<uint32_t, OpDescPtr> &data_by_inde
   }
 }
 
-///
-/// @ingroup ge
-/// @brief input zero copy node Initialize.
-/// @param [in] NodePtr: Data Op.
-/// @return Status
-///
-Status DavinciModel::InitInputZeroCopyNodes(const NodePtr &node) {
-  auto out_data_anchor = node->GetOutDataAnchor(kDataIndex);
-  if (out_data_anchor == nullptr) {
-    GELOGE(FAILED, "Out data anchor is nullptr");
-    return FAILED;
-  }
-  for (auto &peer_in_data_anchor : out_data_anchor->GetPeerInDataAnchors()) {
-    auto node = peer_in_data_anchor->GetOwnerNode();
-    auto op_desc = node->GetOpDesc();
-    if (op_desc == nullptr) {
-      GELOGE(FAILED, "Op desc is nullptr");
-      return FAILED;
-    }
-    string batch_label;
-    (void)ge::AttrUtils::GetStr(op_desc, ATTR_NAME_BATCH_LABEL, batch_label);
-    if (batch_label.empty()) {
-      batch_label = kDefaultBatchLable;
-    }
-    if (zero_copy_op_id_batch_label_.find(op_desc->GetId()) == zero_copy_op_id_batch_label_.end()) {
-      zero_copy_op_id_batch_label_.emplace(pair<int64_t, string>(op_desc->GetId(), batch_label));
-      GELOGD("Init input zero copy nodes success, op name:%s, op id: %ld, batch label: %s.", op_desc->GetName().c_str(),
-             op_desc->GetId(), batch_label.c_str());
-    }
-  }
-  return SUCCESS;
-}
-
 bool DavinciModel::IsGetNextSinkDynamic(const OpDescPtr &op_desc) {
   bool getnext_sink_dynamic = false;
   if (ge::AttrUtils::GetBool(op_desc, ATTR_GETNEXT_SINK_DYNMAIC, getnext_sink_dynamic) && getnext_sink_dynamic) {
@@ -1094,7 +1058,7 @@ Status DavinciModel::InitNetOutput(const NodePtr &node) {
   if (owner_graph->GetParentGraph() != nullptr) {
     GELOGI("Init zero copy by subgraph NetOutput node: %s.", op_desc->GetName().c_str());
     op_list_.erase(op_desc->GetId());
-    return InitOutputBatchLabel(node);
+    return SUCCESS;
   }
 
   output_op_list_.push_back(op_desc);
@@ -1146,8 +1110,6 @@ Status DavinciModel::InitNetOutput(const NodePtr &node) {
     }
   }
 
-  GE_IF_BOOL_EXEC(InitOutputZeroCopyNodes(node) != SUCCESS,
-                  GELOGE(PARAM_INVALID, "Output zero copy nodes init failed!"); return PARAM_INVALID;);
   GetAllGearsInfo(node);
   if (is_getnext_sink_dynamic_) {
     GE_IF_BOOL_EXEC(GetGetDynamicDimsNodeInfo(node) != SUCCESS,
@@ -1343,121 +1305,6 @@ void DavinciModel::ParseDynamicOutShape(const std::vector<std::string> &str_info
   }
 }
 
-///
-/// @ingroup ge
-/// @brief output zero copy node Initialize.
-/// @param [in] NodePtr: netoutput Op.
-/// @return Status
-///
-Status DavinciModel::InitOutputZeroCopyNodes(const NodePtr &node) {
-  set<NodePtr> nodes_need_record;
-  for (auto &in_data_anchor : node->GetAllInDataAnchors()) {
-    auto peer_out_data_anchor = in_data_anchor->GetPeerOutAnchor();
-    if (peer_out_data_anchor == nullptr) {
-      continue;
-    }
-    auto peer_node = peer_out_data_anchor->GetOwnerNode();
-    nodes_need_record.emplace(peer_node);
-
-    // Merge node output multiplexed input, upstream nodes need to be considered in multiple batch scenarios
-    if (peer_node->GetType() == MERGE) {
-      for (const auto &merge_peer_in_data_anchor : peer_node->GetAllInDataAnchors()) {
-        auto merge_peer_out_data_anchor = merge_peer_in_data_anchor->GetPeerOutAnchor();
-        if (merge_peer_out_data_anchor == nullptr) {
-          continue;
-        }
-        auto merge_peer_node = merge_peer_out_data_anchor->GetOwnerNode();
-        nodes_need_record.emplace(merge_peer_node);
-      }
-    } else {
-      for (const auto &other_in_data_anchor : peer_out_data_anchor->GetPeerInDataAnchors()) {
-        auto other_in_node = other_in_data_anchor->GetOwnerNode();
-        if (other_in_node->GetType() != NETOUTPUT) {
-          nodes_need_record.emplace(other_in_node);
-        }
-      }
-    }
-  }
-
-  for (const auto &node_need_record : nodes_need_record) {
-    auto op_desc = node_need_record->GetOpDesc();
-    GE_CHECK_NOTNULL(op_desc);
-    string batch_label;
-    (void)ge::AttrUtils::GetStr(op_desc, ATTR_NAME_BATCH_LABEL, batch_label);
-    if (batch_label.empty()) {
-      batch_label = kDefaultBatchLable;
-    }
-    if (zero_copy_op_id_batch_label_.find(op_desc->GetId()) == zero_copy_op_id_batch_label_.end()) {
-      zero_copy_op_id_batch_label_.emplace(pair<int64_t, string>(op_desc->GetId(), batch_label));
-      GELOGD("Init Output zero copy nodes success, op name:%s, op id: %ld, batch label: %s.",
-             op_desc->GetName().c_str(), op_desc->GetId(), batch_label.c_str());
-    }
-  }
-  return SUCCESS;
-}
-
-///
-/// @ingroup ge
-/// @brief input zero copy node Initialize.
-/// @param [in] NodePtr: Data Op.
-/// @return Status
-///
-Status DavinciModel::InitInputBatchLabel(const NodePtr &node) {
-  string batch_label;
-  if (!AttrUtils::GetStr(node->GetOpDesc(), ATTR_NAME_BATCH_LABEL, batch_label)) {
-    return SUCCESS;  // Not Multi-batch.
-  }
-
-  const auto &out_data_anchor = node->GetOutDataAnchor(kDataIndex);
-  GE_CHECK_NOTNULL(out_data_anchor);
-
-  for (const auto &peer_in_data_anchor : out_data_anchor->GetPeerInDataAnchors()) {
-    const auto &node = peer_in_data_anchor->GetOwnerNode();
-    const auto &op_desc = node->GetOpDesc();
-    GE_CHECK_NOTNULL(op_desc);
-
-    if (zero_copy_op_id_batch_label_.find(op_desc->GetId()) == zero_copy_op_id_batch_label_.end()) {
-      zero_copy_op_id_batch_label_[op_desc->GetId()] = batch_label;
-      GELOGD("Init input zero copy nodes success, op name: %s, op id: %ld, batch label: %s", op_desc->GetName().c_str(),
-             op_desc->GetId(), batch_label.c_str());
-    }
-  }
-
-  return SUCCESS;
-}
-
-///
-/// @ingroup ge
-/// @brief output zero copy node Initialize for Case.
-/// @param [in] NodePtr: netoutput Op.
-/// @return Status
-///
-Status DavinciModel::InitOutputBatchLabel(const NodePtr &node) {
-  string batch_label;
-  if (!AttrUtils::GetStr(node->GetOpDesc(), ATTR_NAME_BATCH_LABEL, batch_label)) {
-    return SUCCESS;  // Not Multi-batch.
-  }
-
-  for (const auto &in_data_anchor : node->GetAllInDataAnchors()) {
-    const auto &peer_out_data_anchor = in_data_anchor->GetPeerOutAnchor();
-    if (peer_out_data_anchor == nullptr) {
-      continue;
-    }
-
-    const auto &peer_node = peer_out_data_anchor->GetOwnerNode();
-    const auto &op_desc = peer_node->GetOpDesc();
-    GE_CHECK_NOTNULL(op_desc);
-
-    if (zero_copy_op_id_batch_label_.find(op_desc->GetId()) == zero_copy_op_id_batch_label_.end()) {
-      zero_copy_op_id_batch_label_[op_desc->GetId()] = batch_label;
-      GELOGD("Init Output zero copy nodes success, op name: %s, op id: %ld, batch label: %s",
-             op_desc->GetName().c_str(), op_desc->GetId(), batch_label.c_str());
-    }
-  }
-
-  return SUCCESS;
-}
-
 /// @ingroup ge
 /// @brief LabelSet Op Initialize.
 /// @param [in] op_desc: LabelSet Op descriptor.
@@ -2240,12 +2087,61 @@ Status DavinciModel::SyncVarData() {
   return ret;
 }
 
-inline int64_t SumSize(const vector<int64_t> &size_list) {
-  int64_t sum_size = 0;
-  for (const int64_t &size : size_list) {
-    sum_size += size;
+Status DavinciModel::InitModelProfile() {
+  for (const auto &task : task_list_) {
+    GE_CHECK_NOTNULL(task);
+    const FusionOpInfo *fusion_op_info = task->GetFusionOpInfo();
+    // when type is RT_MODEL_TASK_KERNEL, ctx is not null
+    if ((fusion_op_info == nullptr) || fusion_op_info->original_op_names.empty()) {
+      continue;
+    }
+
+    GELOGI("task.id = %u, opNum = %zu", task->GetTaskID(), fusion_op_info->original_op_names.size());
+    op_id_map_.insert(std::make_pair(fusion_op_info->op_index, task->GetTaskID()));
+  }
+
+  std::set<uint32_t> task_id_set;
+  using CIT = std::multimap<uint32_t, uint32_t>::const_iterator;
+  using Range = std::pair<CIT, CIT>;
+  for (const auto &task : task_list_) {
+    GE_CHECK_NOTNULL(task);
+    const FusionOpInfo *fusion_op_info = task->GetFusionOpInfo();
+    if ((fusion_op_info == nullptr) || fusion_op_info->original_op_names.empty()) {
+      continue;
+    }
+
+    if (task_id_set.count(task->GetTaskID()) > 0) {
+      continue;
+    }
+
+    const auto &op_desc = GetOpByIndex(fusion_op_info->op_index);
+    GE_CHK_BOOL_EXEC(op_desc != nullptr, return FAILED, "index: %u out of range", fusion_op_info->op_index);
+
+    ProfileInfo profile;
+    profile.fusion_info = *fusion_op_info;
+    Range range = op_id_map_.equal_range(fusion_op_info->op_index);
+    for (CIT range_idx = range.first; range_idx != range.second; ++range_idx) {
+      profile.task_count++;
+      task_id_set.insert(range_idx->second);
+    }
+
+    // memory info
+    TaskMemInfo &mem_info = profile.memory_info;
+    const auto input_size = ModelUtils::GetInputSize(op_desc);
+    const auto output_size = ModelUtils::GetOutputSize(op_desc);
+    const auto workspace_size = ModelUtils::GetWorkspaceSize(op_desc);
+    const auto weight_size = ModelUtils::GetWeightSize(op_desc);
+    mem_info.input_size = std::accumulate(input_size.begin(), input_size.end(), 0);
+    mem_info.output_size = std::accumulate(output_size.begin(), output_size.end(), 0);
+    mem_info.workspace_size = std::accumulate(workspace_size.begin(), workspace_size.end(), 0);
+    mem_info.weight_size = std::accumulate(weight_size.begin(), weight_size.end(), 0);
+    mem_info.total_size = mem_info.weight_size + mem_info.input_size + mem_info.output_size + mem_info.workspace_size;
+
+    profile_list_.emplace_back(profile);
   }
-  return sum_size;
+
+  GELOGI("fusion task size: %zu, profile info size: %zu", op_id_map_.size(), profile_list_.size());
+  return SUCCESS;
 }
 
 Status DavinciModel::SinkModelProfile() {
@@ -2253,18 +2149,12 @@ Status DavinciModel::SinkModelProfile() {
   auto &prof_mgr = ProfilingManager::Instance();
   ReporterData reporter_data{};
   // report model data tag name
-  std::string tag_name;
-  tag_name.append("model_load_info_").append(std::to_string(this->Id()));
+  std::string tag_name("model_load_info_" + std::to_string(this->Id()));
   GE_CHK_BOOL_EXEC(memcpy_s(reporter_data.tag, MSPROF_ENGINE_MAX_TAG_LEN, tag_name.c_str(), tag_name.size()) == EOK,
                    return FAILED, "Sink model tag memcpy error.");
 
   // Model Header
-  string name;
-  if (!om_name_.empty()) {
-    name = om_name_;
-  } else {
-    name = name_;
-  }
+  std::string name = om_name_.empty() ? name_ : om_name_;
   size_t name_len = name.size();
   reporter_data.deviceId = device_id_;
   reporter_data.data = (unsigned char *)&name_len;
@@ -2296,128 +2186,71 @@ Status DavinciModel::SinkModelProfile() {
   GE_CHK_BOOL_EXEC(prof_mgr.CallMsprofReport(reporter_data) == 0, return FAILED,
                    "Reporter data fail, model id:%u.", this->Id());
 
-  int32_t task_num = task_list_.size();
-  std::multimap<uint32_t, uint32_t> op_id_map;
-  std::set<uint32_t> task_id_set;
-  for (int32_t i = 0; i < task_num; i++) {
-    auto task = task_list_[i];
-    GE_CHECK_NOTNULL(task);
-    auto fusion_op_info = task->GetFusionOpInfo();
-    // when type is RT_MODEL_TASK_KERNEL, ctx is not null
-    if (fusion_op_info != nullptr) {
-      uint32_t op_num = fusion_op_info->original_op_names.size();
-      uint32_t task_id = task->GetTaskID();
-      if (op_num > 0) {
-        GELOGI("task.id = %u, opNum = %u", task_id, op_num);
-        op_id_map.insert(std::make_pair(fusion_op_info->op_index, task_id));
-      }
-    }
-  }
-
-  struct memoryInfo {
-    int64_t input_size;
-    int64_t output_size;
-    int64_t weight_size;
-    int64_t workspace_size;
-    int64_t total_size;
-
-    memoryInfo() : input_size(0), output_size(0), weight_size(0), workspace_size(0), total_size(0) {}
-  };
-
   using CIT = std::multimap<uint32_t, uint32_t>::const_iterator;
   using Range = std::pair<CIT, CIT>;
-  for (int32_t i = 0; i < task_num; i++) {
-    auto task = task_list_[i];
-    GE_CHECK_NOTNULL(task);
-    auto fusion_op_info = task->GetFusionOpInfo();
-    if (fusion_op_info != nullptr && fusion_op_info->original_op_names.size() > 0) {
-      uint32_t task_id = task->GetTaskID();
-      uint32_t op_num = fusion_op_info->original_op_names.size();
-      uint32_t task_count = 0;
-      if (task_id_set.count(task_id) != 0) {
-        continue;
-      }
-
-      uint32_t op_id = fusion_op_info->op_index;
-      Range range = op_id_map.equal_range(op_id);
-      for (CIT range_idx = range.first; range_idx != range.second; ++range_idx) {
-        task_count++;
-        uint32_t task_id = range_idx->second;
-        task_id_set.insert(task_id);
-      }
-
-      // op name after fusion
-      string fusion_op_name = fusion_op_info->op_name;
-      int32_t fusion_op_name_len = fusion_op_name.size() == 0 ? 1 : fusion_op_name.size();
-      reporter_data.data = (unsigned char *)&fusion_op_name_len;
-      reporter_data.dataLen = sizeof(int32_t);
-      GE_CHK_BOOL_EXEC(prof_mgr.CallMsprofReport(reporter_data) == 0, return FAILED,
-                       "Reporter data fail, model id:%u.", this->Id());
-
-      reporter_data.data = (unsigned char *)fusion_op_name.c_str();
-      reporter_data.dataLen = fusion_op_name_len;
-      GE_CHK_BOOL_EXEC(prof_mgr.CallMsprofReport(reporter_data) == 0, return FAILED,
-                       "Reporter data fail, model id:%u.", this->Id());
-
-      // original op name before fusion
-      reporter_data.data = (unsigned char *)&op_num;
-      reporter_data.dataLen = sizeof(int32_t);
-      GE_CHK_BOOL_EXEC(prof_mgr.CallMsprofReport(reporter_data) == 0, return FAILED,
-                       "Reporter data fail, model id:%u.", this->Id());
-
-      for (uint32_t k = 0; k < op_num; k++) {
-        std::string op_name = fusion_op_info->original_op_names[k];
-        int32_t op_name_len = op_name.size() == 0 ? 1 : op_name.size();
-        reporter_data.data = (unsigned char *)&op_name_len;
-        reporter_data.dataLen = sizeof(int32_t);
-        GE_CHK_BOOL_EXEC(prof_mgr.CallMsprofReport(reporter_data) == 0, return FAILED,
-                         "Reporter data fail, model id:%u.", this->Id());
-        reporter_data.data = (unsigned char *)op_name.c_str();
-        reporter_data.dataLen = op_name_len;
-        GE_CHK_BOOL_EXEC(prof_mgr.CallMsprofReport(reporter_data) == 0, return FAILED,
-                         "Reporter data fail, model id:%u.", this->Id());
-      }
-
-      // stream id info
-      uint32_t streamId = task->GetStreamId();
-      reporter_data.data = (unsigned char *)&streamId;
+  for (const ProfileInfo &profile : profile_list_) {
+    // op name after fusion
+    string fusion_op_name = profile.fusion_info.op_name;
+    int32_t fusion_op_name_len = fusion_op_name.size() == 0 ? 1 : fusion_op_name.size();
+    reporter_data.data = (unsigned char *)&fusion_op_name_len;
+    reporter_data.dataLen = sizeof(int32_t);
+    GE_CHK_BOOL_EXEC(prof_mgr.CallMsprofReport(reporter_data) == 0, return FAILED,
+                     "Reporter data fail, model id:%u.", this->Id());
+
+    reporter_data.data = (unsigned char *)fusion_op_name.c_str();
+    reporter_data.dataLen = fusion_op_name_len;
+    GE_CHK_BOOL_EXEC(prof_mgr.CallMsprofReport(reporter_data) == 0, return FAILED,
+                     "Reporter data fail, model id:%u.", this->Id());
+
+    // original op name before fusion
+    uint32_t op_num = profile.fusion_info.original_op_names.size();
+    reporter_data.data = (unsigned char *)&op_num;
+    reporter_data.dataLen = sizeof(int32_t);
+    GE_CHK_BOOL_EXEC(prof_mgr.CallMsprofReport(reporter_data) == 0, return FAILED,
+                     "Reporter data fail, model id:%u.", this->Id());
+
+    for (uint32_t k = 0; k < op_num; k++) {
+      std::string op_name = profile.fusion_info.original_op_names[k];
+      int32_t op_name_len = op_name.size() == 0 ? 1 : op_name.size();
+      reporter_data.data = (unsigned char *)&op_name_len;
       reporter_data.dataLen = sizeof(int32_t);
       GE_CHK_BOOL_EXEC(prof_mgr.CallMsprofReport(reporter_data) == 0, return FAILED,
                        "Reporter data fail, model id:%u.", this->Id());
-
-      // memory info
-      struct memoryInfo memory_info;
-      uint32_t op_index = fusion_op_info->op_index;
-      auto iter = op_list_.find(op_index);
-      GE_CHK_BOOL_EXEC(iter != op_list_.end(), return FAILED, "index is out of range, index: %u", op_index);
-      auto op_desc = iter->second;
-      memory_info.input_size = SumSize(ModelUtils::GetInputSize(op_desc));
-      memory_info.output_size = SumSize(ModelUtils::GetOutputSize(op_desc));
-      memory_info.workspace_size = SumSize(ModelUtils::GetWorkspaceSize(op_desc));
-      memory_info.weight_size = SumSize(ModelUtils::GetWeightSize(op_desc));
-      memory_info.total_size =
-          memory_info.weight_size + memory_info.input_size + memory_info.output_size + memory_info.workspace_size;
-      reporter_data.data = (unsigned char *)&memory_info;
-      reporter_data.dataLen = sizeof(struct memoryInfo);
+      reporter_data.data = (unsigned char *)op_name.c_str();
+      reporter_data.dataLen = op_name_len;
       GE_CHK_BOOL_EXEC(prof_mgr.CallMsprofReport(reporter_data) == 0, return FAILED,
                        "Reporter data fail, model id:%u.", this->Id());
+    }
 
-      // task info
-      reporter_data.data = (unsigned char *)&task_count;
+    // stream id info
+    uint32_t streamId = profile.fusion_info.stream_id;
+    reporter_data.data = (unsigned char *)&streamId;
+    reporter_data.dataLen = sizeof(int32_t);
+    GE_CHK_BOOL_EXEC(prof_mgr.CallMsprofReport(reporter_data) == 0, return FAILED,
+                     "Reporter data fail, model id:%u.", this->Id());
+
+    // memory info
+    reporter_data.data = (unsigned char *)&profile.memory_info;
+    reporter_data.dataLen = sizeof(profile.memory_info);
+    GE_CHK_BOOL_EXEC(prof_mgr.CallMsprofReport(reporter_data) == 0, return FAILED,
+                     "Reporter data fail, model id:%u.", this->Id());
+
+    // task info
+    reporter_data.data = (unsigned char *)&profile.task_count;
+    reporter_data.dataLen = sizeof(uint32_t);
+    GE_CHK_BOOL_EXEC(prof_mgr.CallMsprofReport(reporter_data) == 0, return FAILED,
+                     "Reporter data fail, model id:%u.", this->Id());
+
+    Range task_range = op_id_map_.equal_range(profile.fusion_info.op_index);
+    for (CIT idx = task_range.first; idx != task_range.second; ++idx) {
+      uint32_t task_id = idx->second;
+      reporter_data.data = (unsigned char *)&task_id;
       reporter_data.dataLen = sizeof(uint32_t);
       GE_CHK_BOOL_EXEC(prof_mgr.CallMsprofReport(reporter_data) == 0, return FAILED,
                        "Reporter data fail, model id:%u.", this->Id());
-
-      Range task_range = op_id_map.equal_range(op_id);
-      for (CIT idx = task_range.first; idx != task_range.second; ++idx) {
-        uint32_t task_id = idx->second;
-        reporter_data.data = (unsigned char *)&task_id;
-        reporter_data.dataLen = sizeof(uint32_t);
-        GE_CHK_BOOL_EXEC(prof_mgr.CallMsprofReport(reporter_data) == 0, return FAILED,
-                         "Reporter data fail, model id:%u.", this->Id());
-      }
     }
   }
+
   return SUCCESS;
 }
 
@@ -3264,27 +3097,20 @@ void DavinciModel::SetZeroCopyAddr(const OpDescPtr &op_desc, const std::vector<v
 
     for (auto &input_outside_addrs : new_input_outside_addrs_) {
       ZeroCopyOffset &input_outside = input_outside_addrs.second;
-      bool ret = input_outside.SetOutsideAddrsValue(zero_copy_task, outside_addrs[i], args, offset + i * kAddrLen);
-      if (ret) {
-        void *args_val = static_cast<uint8_t *>(args) + offset + i * kAddrLen;
-        SetBatchLabelAddr(op_desc, reinterpret_cast<uintptr_t>(args_val));
-      }
+      input_outside.SetOutsideAddrsValue(zero_copy_task, outside_addrs[i], args, offset + i * kAddrLen);
     }
 
     for (auto &output_outside_addrs : new_output_outside_addrs_) {
       ZeroCopyOffset &output_outside = output_outside_addrs.second;
-      bool ret = output_outside.SetOutsideAddrsValue(zero_copy_task, outside_addrs[i], args, offset + i * kAddrLen);
-      if (ret) {
-        void *args_val = static_cast<uint8_t *>(args) + offset + i * kAddrLen;
-        SetBatchLabelAddr(op_desc, reinterpret_cast<uintptr_t>(args_val));
-      }
+      output_outside.SetOutsideAddrsValue(zero_copy_task, outside_addrs[i], args, offset + i * kAddrLen);
     }
   }
-  auto it = zero_copy_op_id_batch_label_.find(op_desc->GetId());
-  if (it == zero_copy_op_id_batch_label_.end()) {
+
+  string batch_label;
+  if (!AttrUtils::GetStr(op_desc, ATTR_NAME_BATCH_LABEL, batch_label) || batch_label.empty()) {
     zero_copy_task.SetBatchLabel(kDefaultBatchLable);
   } else {
-    zero_copy_task.SetBatchLabel(it->second);
+    zero_copy_task.SetBatchLabel(batch_label);
   }
 
   std::lock_guard<std::mutex> lock(outside_addrs_mutex_);
@@ -3294,27 +3120,6 @@ void DavinciModel::SetZeroCopyAddr(const OpDescPtr &op_desc, const std::vector<v
   }
 }
 
-void DavinciModel::SetBatchLabelAddr(const OpDescPtr &op_desc, uintptr_t addr) {
-  // Establish a mapping between batch label and zero copy address for multi-batch scenes
-  auto it = zero_copy_op_id_batch_label_.find(op_desc->GetId());
-  if (it == zero_copy_op_id_batch_label_.end()) {
-    return;
-  }
-
-  const string &batch_label = it->second;
-  auto iter = zero_copy_batch_label_addrs_.find(batch_label);
-  if (iter != zero_copy_batch_label_addrs_.end()) {
-    iter->second.insert(addr);
-    GELOGD("[ZCPY] Set zero copy batch label and addrs success, batch label: %s, op name:%s.", batch_label.c_str(),
-           op_desc->GetName().c_str());
-  } else {
-    set<uintptr_t> addrs = {addr};
-    zero_copy_batch_label_addrs_.emplace(pair<string, set<uintptr_t>>(batch_label, addrs));
-    GELOGD("[ZCPY] New added zero copy batch label and addrs success, batch label: %s, op name:%s.",
-           batch_label.c_str(), op_desc->GetName().c_str());
-  }
-}
-
 ///
 /// @ingroup ge
 /// @brief Copy Check input size and model op size.
@@ -3448,15 +3253,15 @@ Status DavinciModel::UpdateIoTaskArgs(const std::map<uint32_t, ZeroCopyOffset> &
       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", input_or_output.c_str(),
-             data.first, addr, size, buffer_addr);
+      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());
       // For input data, just copy for rts task.
       for (ZeroCopyTask &task : zero_copy_tasks_) {
         if (task.GetBatchLabel() != kDefaultBatchLable && task.GetBatchLabel() != batch_label) {
           continue;
         }
         uintptr_t addr_val = reinterpret_cast<uintptr_t>(addr);
-        if (task.UpdateTaskParam(addr_val, buffer_addr, zero_copy_batch_label_addrs_, batch_label) != SUCCESS) {
+        if (task.UpdateTaskParam(addr_val, buffer_addr) != SUCCESS) {
           return FAILED;
         }
       }
@@ -3818,9 +3623,6 @@ Status DavinciModel::NnExecute(rtStream_t stream, bool async_mode, const InputDa
   GELOGD("Model Run begin, model id:%u, data index:%u, flag:%d.", model_id_, input_data.index, is_async_mode_);
   GE_CHK_STATUS_RET(InitModelStream(stream), "Init model stream failed.");
   is_dynamic_ = input_data.is_dynamic_batch;
-  if (!is_dynamic_) {
-    zero_copy_batch_label_addrs_.clear();
-  }
 
   GE_IF_BOOL_EXEC(ProfilingManager::Instance().ProfilingModelExecuteOn(), SetProfileTime(MODEL_PRE_PROC_START));
   Status ret = CopyModelData(input_data, output_data, is_dynamic_);

ge/graph/load/new_model_manager/davinci_model.h

@@ -76,6 +76,20 @@ struct timeInfo {
   int64_t dumpEndTime;
 };
 
+struct TaskMemInfo {
+  int64_t input_size{0};
+  int64_t output_size{0};
+  int64_t weight_size{0};
+  int64_t workspace_size{0};
+  int64_t total_size{0};
+};
+
+struct ProfileInfo {
+  FusionOpInfo fusion_info;
+  TaskMemInfo memory_info;
+  uint32_t task_count{0};
+};
+
 enum ExecuteMode {
   INITIALIZATION,
   SYNCHRONIZATION,
@@ -226,8 +240,6 @@ class DavinciModel {
   const vector<OpDescPtr> &GetDataList() const { return data_op_list_; }
 
   // get Op
-  const map<uint32_t, OpDescPtr> &GetOpList() const { return op_list_; }
-
   OpDescPtr GetOpByIndex(uint32_t index) const {
     if (op_list_.find(index) == op_list_.end()) {
       return nullptr;
@@ -436,10 +448,6 @@ class DavinciModel {
 
   int64_t GetLoadEndTime() { return load_end_time_; }
 
-  Status SinkModelProfile();
-
-  Status SinkTimeProfile(const InputData &current_data);
-
   Status ReportProfilingData();
 
   void SaveDumpOpInfo(const RuntimeParam &model_param, const OpDescPtr &op, uint32_t task_id, uint32_t stream_id) {
@@ -537,15 +545,6 @@ class DavinciModel {
   struct timeInfo time_info_;
   int32_t dataInputTid;
 
-  ///
-  /// @ingroup ge
-  /// @brief Save Batch label Info.
-  /// @param [in] const OpDescPtr &op_desc
-  /// @param [in] uintptr_t addr: address value in args block.
-  /// @return None.
-  ///
-  void SetBatchLabelAddr(const OpDescPtr &op_desc, uintptr_t addr);
-
   ///
   /// @ingroup ge
   /// @brief Copy Check input size and model op size.
@@ -657,14 +656,6 @@ class DavinciModel {
   ///
   void AdjustDataOpList(const map<uint32_t, OpDescPtr> &data_by_index);
 
-  ///
-  /// @ingroup ge
-  /// @brief input zero copy node Initialize.
-  /// @param [in] NodePtr: Data Op.
-  /// @return Status
-  ///
-  Status InitInputZeroCopyNodes(const NodePtr &node);
-
   ///
   /// @ingroup ge
   /// @brief NetOutput Op Initialize.
@@ -673,30 +664,6 @@ class DavinciModel {
   ///
   Status InitNetOutput(const NodePtr &node);
 
-  ///
-  /// @ingroup ge
-  /// @brief output zero copy node Initialize.
-  /// @param [in] NodePtr: Data Op.
-  /// @return Status
-  ///
-  Status InitOutputZeroCopyNodes(const NodePtr &node);
-
-  ///
-  /// @ingroup ge
-  /// @brief input zero copy node Initialize for Case.
-  /// @param [in] NodePtr: Data Op.
-  /// @return Status
-  ///
-  Status InitInputBatchLabel(const NodePtr &node);
-
-  ///
-  /// @ingroup ge
-  /// @brief output zero copy node Initialize for Case.
-  /// @param [in] NodePtr: netoutput Op.
-  /// @return Status
-  ///
-  Status InitOutputBatchLabel(const NodePtr &node);
-
   ///
   /// @ingroup ge
   /// @brief Constant Op Init.
@@ -845,6 +812,11 @@ class DavinciModel {
 
   void SetDataDumperArgs(const ComputeGraphPtr &compute_graph);
 
+  Status InitModelProfile();
+  Status SinkModelProfile();
+
+  Status SinkTimeProfile(const InputData &current_data);
+
   Status GenOutputTensorInfo(const OpDescPtr &op_desc, uint32_t data_index, OutputData *output_data,
                              std::vector<ge::OutputTensorInfo> &outputs);
 
@@ -922,11 +894,6 @@ class DavinciModel {
   std::vector<ZeroCopyTask> zero_copy_tasks_;  // Task used Data or NetOutput addr.
   std::set<const void *> copy_only_addrs_;     // Address need copy to original place.
 
-  // {op_id, batch_label}
-  std::map<int64_t, std::string> zero_copy_op_id_batch_label_;
-  // {batch_label, addrs}
-  std::map<std::string, std::set<uintptr_t>> zero_copy_batch_label_addrs_;
-
   std::vector<TaskInfoPtr> task_list_;
   // rt_moodel_handle
   rtModel_t rt_model_handle_;
@@ -1026,6 +993,9 @@ class DavinciModel {
   // key: input_index: input is merge node; value: each gear info and each output shape
   std::map<size_t, std::map<vector<int64_t>, vector<int64_t>>> merge_nodes_gear_and_real_out_shape_info_;
   std::vector<std::vector<int64_t>> all_gears_info_;
+
+  std::multimap<uint32_t, uint32_t> op_id_map_;
+  std::vector<ProfileInfo> profile_list_;
 };
 }  // namespace ge
 #endif  // GE_GRAPH_LOAD_NEW_MODEL_MANAGER_DAVINCI_MODEL_H_

ge/graph/load/new_model_manager/model_manager.cc

@@ -89,6 +89,7 @@ Status ModelManager::KernelLaunchEx(aicpu::FWKAdapter::FWKOperateType op_type, u
   if (op_type == aicpu::FWKAdapter::FWKOperateType::FWK_ADPT_KERNEL_DESTROY) {
     std::vector<uint64_t> v_aicpu_kernel;
     std::string model_key = std::to_string(session_id) + "_" + std::to_string(model_id);
+    std::lock_guard<std::recursive_mutex> lock(map_mutex_);
     auto iter = model_aicpu_kernel_.find(model_key);
     if (iter != model_aicpu_kernel_.end()) {
       GELOGD("kernel destroy session_id %lu, model_id %u.", session_id, model_id);
@@ -176,7 +177,7 @@ Status ModelManager::KernelLaunchEx(aicpu::FWKAdapter::FWKOperateType op_type, u
 }
 
 void ModelManager::DestroyAicpuSession(uint64_t session_id) {
-  std::lock_guard<std::mutex> lock(sess_ids_mutex_);
+  std::lock_guard<std::recursive_mutex> lock(map_mutex_);
   auto it = sess_ids_.find(session_id);
   if (it == sess_ids_.end()) {
     GELOGI("The session: %lu not created.", session_id);
@@ -205,7 +206,7 @@ void ModelManager::DestroyAicpuSession(uint64_t session_id) {
 }
 
 ge::Status ModelManager::DestroyAicpuSessionForInfer(uint32_t model_id) {
-  std::lock_guard<std::mutex> lock(map_mutex_);
+  std::lock_guard<std::recursive_mutex> lock(map_mutex_);
   auto hybrid_davinci_model = hybrid_model_map_.find(model_id);
   if (hybrid_davinci_model != hybrid_model_map_.end()) {
     uint64_t session_id = hybrid_davinci_model->second->GetSessionId();
@@ -215,8 +216,8 @@ ge::Status ModelManager::DestroyAicpuSessionForInfer(uint32_t model_id) {
 
   auto it = model_map_.find(model_id);
   if (it == model_map_.end()) {
-    GELOGE(GE_EXEC_MODEL_ID_INVALID, "model id %u does not exists.", model_id);
-    return GE_EXEC_MODEL_ID_INVALID;
+    GELOGE(ACL_ERROR_GE_EXEC_MODEL_ID_INVALID, "model id %u does not exists.", model_id);
+    return ACL_ERROR_GE_EXEC_MODEL_ID_INVALID;
   }
   uint64_t session_id = it->second->GetSessionId();
   DestroyAicpuSession(session_id);
@@ -225,7 +226,7 @@ ge::Status ModelManager::DestroyAicpuSessionForInfer(uint32_t model_id) {
 
 ge::Status ModelManager::DestroyAicpuKernel(uint64_t session_id, uint32_t model_id) {
   GELOGD("destroy aicpu kernel in session_id %lu, model_id %u.", session_id, model_id);
-  std::lock_guard<std::mutex> lock(map_mutex_);
+  std::lock_guard<std::recursive_mutex> lock(map_mutex_);
   std::string model_key = std::to_string(session_id) + "_" + std::to_string(model_id);
   if (model_aicpu_kernel_.find(model_key) != model_aicpu_kernel_.end()) {
     Status ret = KernelLaunchEx(aicpu::FWKAdapter::FWKOperateType::FWK_ADPT_KERNEL_DESTROY, session_id, model_id);
@@ -238,7 +239,7 @@ ge::Status ModelManager::DestroyAicpuKernel(uint64_t session_id, uint32_t model_
 }
 
 ge::Status ModelManager::CreateAicpuKernel(uint64_t session_id, uint32_t model_id, uint64_t kernel_id) {
-  std::lock_guard<std::mutex> lock(map_mutex_);
+  std::lock_guard<std::recursive_mutex> lock(map_mutex_);
   std::vector<uint64_t> v_aicpu_kernel;
   std::string model_key = std::to_string(session_id) + "_" + std::to_string(model_id);
   if (model_aicpu_kernel_.find(model_key) != model_aicpu_kernel_.end()) {
@@ -250,7 +251,7 @@ ge::Status ModelManager::CreateAicpuKernel(uint64_t session_id, uint32_t model_i
 }
 
 ModelManager::~ModelManager() {
-  std::lock_guard<std::mutex> lock(map_mutex_);
+  std::lock_guard<std::recursive_mutex> lock(map_mutex_);
   model_map_.clear();
   model_aicpu_kernel_.clear();
   cust_aicpu_so_.clear();
@@ -358,18 +359,18 @@ Status ModelManager::LoadModelOnline(uint32_t &model_id, const shared_ptr<ge::Ge
 
 void ModelManager::InsertModel(uint32_t id, std::shared_ptr<DavinciModel> &davinci_model) {
   GE_CHK_BOOL_EXEC(davinci_model != nullptr, return, "davinci_model ptr is null, id: %u", id);
-  std::lock_guard<std::mutex> lock(map_mutex_);
+  std::lock_guard<std::recursive_mutex> lock(map_mutex_);
   model_map_[id] = davinci_model;
 }
 
 void ModelManager::InsertModel(uint32_t id, shared_ptr<hybrid::HybridDavinciModel> &hybrid_model) {
   GE_CHK_BOOL_EXEC(hybrid_model != nullptr, return, "hybrid_model ptr is null, id: %u", id);
-  std::lock_guard<std::mutex> lock(map_mutex_);
+  std::lock_guard<std::recursive_mutex> lock(map_mutex_);
   hybrid_model_map_[id] = hybrid_model;
 }
 
 Status ModelManager::DeleteModel(uint32_t id) {
-  std::lock_guard<std::mutex> lock(map_mutex_);
+  std::lock_guard<std::recursive_mutex> lock(map_mutex_);
 
   auto it = model_map_.find(id);
   auto hybrid_model_it = hybrid_model_map_.find(id);
@@ -384,22 +385,22 @@ Status ModelManager::DeleteModel(uint32_t id) {
   } else if (hybrid_model_it != hybrid_model_map_.end()) {
     (void)hybrid_model_map_.erase(hybrid_model_it);
   } else {
-    GELOGE(GE_EXEC_MODEL_ID_INVALID, "model id %u does not exists.", id);
-    return GE_EXEC_MODEL_ID_INVALID;
+    GELOGE(ACL_ERROR_GE_EXEC_MODEL_ID_INVALID, "model id %u does not exists.", id);
+    return ACL_ERROR_GE_EXEC_MODEL_ID_INVALID;
   }
 
   return SUCCESS;
 }
 
 std::shared_ptr<DavinciModel> ModelManager::GetModel(uint32_t id) {
-  std::lock_guard<std::mutex> lock(map_mutex_);
+  std::lock_guard<std::recursive_mutex> lock(map_mutex_);
 
   auto it = model_map_.find(id);
   return (it == model_map_.end()) ? nullptr : it->second;
 }
 
 std::shared_ptr<hybrid::HybridDavinciModel> ModelManager::GetHybridModel(uint32_t id) {
-  std::lock_guard<std::mutex> lock(map_mutex_);
+  std::lock_guard<std::recursive_mutex> lock(map_mutex_);
 
   auto it = hybrid_model_map_.find(id);
   return (it == hybrid_model_map_.end()) ? nullptr : it->second;
@@ -902,7 +903,7 @@ Status ModelManager::GetInputOutputDescInfo(const uint32_t model_id, vector<Inpu
   }
 
   std::shared_ptr<DavinciModel> davinci_model = GetModel(model_id);
-  GE_CHK_BOOL_RET_STATUS(davinci_model != nullptr, GE_EXEC_MODEL_ID_INVALID,
+  GE_CHK_BOOL_RET_STATUS(davinci_model != nullptr, ACL_ERROR_GE_EXEC_MODEL_ID_INVALID,
                          "GetInputOutputDescInfo Failed, Invalid model id %u!", model_id);
 
   davinci_model->SetModelDescVersion(new_model_desc);
@@ -970,8 +971,9 @@ Status ModelManager::GetUserDesignateShapeOrder(const uint32_t model_id,
 }
 
 Status ModelManager::GetCurShape(const uint32_t model_id, std::vector<int64_t> &batch_info, int32_t &dynamic_type) {
-  std::shared_ptr<DavinciModel> davinci_model = GetModel(model_id);
-  GE_CHECK_NOTNULL(davinci_model);
+  auto davinci_model = GetModel(model_id);
+  GE_CHK_BOOL_RET_STATUS(davinci_model != nullptr, ACL_ERROR_GE_EXEC_MODEL_ID_INVALID,
+                         "GetCurShape Failed, Invalid Model ID %u!", model_id);
   davinci_model->GetCurShape(batch_info, dynamic_type);
   return SUCCESS;
 }
@@ -984,7 +986,8 @@ Status ModelManager::GetModelAttr(uint32_t model_id, std::vector<string> &dynami
   }
 
   std::shared_ptr<DavinciModel> davinci_model = GetModel(model_id);
-  GE_CHECK_NOTNULL(davinci_model);
+  GE_CHK_BOOL_RET_STATUS(davinci_model != nullptr, ACL_ERROR_GE_EXEC_MODEL_ID_INVALID,
+                         "GetModelAttr Failed, Invalid Model ID %u!", model_id);
   davinci_model->GetModelAttr(dynamic_output_shape_info);
   return SUCCESS;
 }
@@ -994,9 +997,8 @@ Status ModelManager::GetInputOutputDescInfoForZeroCopy(const uint32_t model_id,
                                                        std::vector<uint32_t> &inputFormats,
                                                        std::vector<uint32_t> &outputFormats) {
   std::shared_ptr<DavinciModel> davinci_model = GetModel(model_id);
-  GE_CHK_BOOL_RET_STATUS(davinci_model != nullptr, PARAM_INVALID, "GetInputOutputDescInfo Failed, Invalid model id %u!",
-                         model_id);
-
+  GE_CHK_BOOL_RET_STATUS(davinci_model != nullptr, ACL_ERROR_GE_EXEC_MODEL_ID_INVALID,
+      "GetInputOutputDescInfo Failed, Invalid model id %u!", model_id);
   return davinci_model->GetInputOutputDescInfoForZeroCopy(input_desc, output_desc, inputFormats, outputFormats);
 }
 
@@ -1011,18 +1013,14 @@ Status ModelManager::GetInputOutputDescInfoForZeroCopy(const uint32_t model_id,
 Status ModelManager::GetAIPPInfo(const uint32_t model_id, uint32_t index, AippConfigInfo &aipp_info) {
   std::shared_ptr<DavinciModel> davinci_model = GetModel(model_id);
   GE_CHK_BOOL_RET_STATUS(davinci_model != nullptr, ACL_ERROR_GE_EXEC_MODEL_ID_INVALID,
-                         "GetAIPPInfo failed, invalid model_id is %u.",
-                         model_id);
-
+      "GetAIPPInfo failed, invalid model_id is %u.", model_id);
   return davinci_model->GetAIPPInfo(index, aipp_info);
 }
 
 Status ModelManager::GetAippType(uint32_t model_id, uint32_t index, InputAippType &type, size_t &aipp_index) {
   std::shared_ptr<DavinciModel> davinci_model = GetModel(model_id);
   GE_CHK_BOOL_RET_STATUS(davinci_model != nullptr, ACL_ERROR_GE_EXEC_MODEL_ID_INVALID,
-                         "GetAIPPInfo failed, invalid model_id is %u.",
-                         model_id);
-
+      "GetAIPPInfo failed, invalid model_id is %u.", model_id);
   return davinci_model->GetAippType(index, type, aipp_index);
 }
 
@@ -1059,8 +1057,7 @@ Status ModelManager::LoadModelOffline(uint32_t &model_id, const ModelData &model
   if (model_helper.GetModelType()) {
     bool is_shape_unknown = false;
     GE_CHK_STATUS_RET(model_helper.GetGeRootModel()->CheckIsUnknownShape(is_shape_unknown),
-                      "CheckIsUnknownShape failed, model id:%u",
-                      model_id);
+                      "CheckIsUnknownShape failed, model id:%u", model_id);
     if (is_shape_unknown || GetContext().GetHostExecFlag()) {
       return DoLoadHybridModelOnline(model_id, model_helper.GetGeRootModel(), listener);
     }
@@ -1078,8 +1075,8 @@ Status ModelManager::LoadModelOffline(uint32_t &model_id, const ModelData &model
       GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION, "Make shared failed");
       return ACL_ERROR_GE_MEMORY_ALLOCATION;
     } catch (...) {
-      GELOGE(INTERNAL_ERROR, "Make shared failed since other exception raise");
-      return INTERNAL_ERROR;
+      GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION, "Make shared failed since other exception raise");
+      return ACL_ERROR_GE_MEMORY_ALLOCATION;
     }
     ret = davinci_model->Assign(ge_model);
     if (ret != SUCCESS) {
@@ -1091,7 +1088,7 @@ Status ModelManager::LoadModelOffline(uint32_t &model_id, const ModelData &model
     int32_t device_id = 0;
     rtError_t rt_ret = rtGetDevice(&device_id);
     if (rt_ret != RT_ERROR_NONE || device_id < 0) {
-      GELOGE(RT_FAILED, "Call rtGetDevice failed, ret = 0x%X, device_id = %d.", rt_ret, device_id);
+      GELOGE(rt_ret, "Call rtGetDevice failed, ret = 0x%X, device_id = %d.", rt_ret, device_id);
       return RT_ERROR_TO_GE_STATUS(rt_ret);
     }
     davinci_model->SetDeviceId(device_id);
@@ -1246,7 +1243,7 @@ Status ModelManager::ExecuteModel(uint32_t model_id, rtStream_t stream, bool asy
 }
 
 Status ModelManager::CreateAicpuSession(uint64_t session_id) {
-  std::lock_guard<std::mutex> lock(sess_ids_mutex_);
+  std::lock_guard<std::recursive_mutex> lock(map_mutex_);
   auto it = sess_ids_.find(session_id);
   // never been created by any model
   if (it == sess_ids_.end()) {
@@ -1465,8 +1462,7 @@ void ModelManager::GenModelId(uint32_t *id) {
   if (id == nullptr) {
     return;
   }
-
-  std::lock_guard<std::mutex> lock(map_mutex_);
+  std::lock_guard<std::recursive_mutex> lock(map_mutex_);
   *id = ++max_model_id_;
 }
 

ge/graph/load/new_model_manager/model_manager.h

@@ -353,8 +353,7 @@ class FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY ModelManager {
   std::map<uint32_t, std::shared_ptr<hybrid::HybridDavinciModel>> hybrid_model_map_;
   std::map<std::string, std::vector<uint64_t>> model_aicpu_kernel_;
   uint32_t max_model_id_;
-  std::mutex map_mutex_;
-  std::mutex sess_ids_mutex_;
+  std::recursive_mutex map_mutex_;
   std::mutex session_id_create_mutex_;
   static::std::mutex exeception_infos_mutex_;
   uint64_t session_id_bias_;

ge/graph/load/new_model_manager/task_info/kernel_task_info.cc

@@ -90,20 +90,18 @@ Status KernelTaskInfo::Init(const domi::TaskDef &task_def, DavinciModel *davinci
                   fusion_op_info_.op_index = context.op_index(); fusion_op_info_.original_op_names = original_op_names;
                   fusion_op_info_.op_name = op_desc_->GetName());
 
-  string session_graph_model_id;
-  davinci_model_->GetUniqueId(op_desc_, session_graph_model_id);
-  // get bin_file_key
-  const char *bin_file_key = davinci_model_->GetRegisterStub(op_desc_->GetName(), session_graph_model_id);
   // new aicpu kernel(rtCpuKernelLaunch) no need to check function
   if (kernel_type_ == ccKernelType::CCE_AI_CORE) {
-    rtError_t rt_ret;
-    rt_ret = rtGetFunctionByName(const_cast<char *>(kernel_def.stub_func().c_str()), &stub_func_);
+    rtError_t rt_ret = rtGetFunctionByName(const_cast<char *>(kernel_def.stub_func().c_str()), &stub_func_);
     GE_IF_BOOL_EXEC(rt_ret != RT_ERROR_NONE, GELOGE(RT_FAILED, "execute rtGetFunctionByName failed. stub_func: %s",
                                                     kernel_def.stub_func().c_str());
                     return RT_ERROR_TO_GE_STATUS(rt_ret););
   } else if (kernel_type_ == ccKernelType::TE) {
-    rtError_t rt_ret;
-    rt_ret = rtGetFunctionByName(bin_file_key, &stub_func_);
+    // get bin_file_key
+    string session_graph_model_id;
+    davinci_model_->GetUniqueId(op_desc_, session_graph_model_id);
+    const char *bin_file_key = davinci_model_->GetRegisterStub(op_desc_->GetName(), session_graph_model_id);
+    rtError_t rt_ret = rtGetFunctionByName(bin_file_key, &stub_func_);
     GE_IF_BOOL_EXEC(rt_ret != RT_ERROR_NONE,
                     GELOGE(RT_FAILED, "execute rtGetFunctionByName failed. bin_file_key: %s", bin_file_key);
                     return RT_ERROR_TO_GE_STATUS(rt_ret););

ge/graph/load/new_model_manager/zero_copy_offset.cc

@@ -183,22 +183,18 @@ void ZeroCopyOffset::SetOutputOutsideAddrs(const int64_t &input_offset, const bo
   addr_count_ = out_count;
 }
 
-bool ZeroCopyOffset::SetOutsideAddrsValue(ZeroCopyTask &zero_copy_task, void *outside_addr, void *args, size_t offset) {
+void ZeroCopyOffset::SetOutsideAddrsValue(ZeroCopyTask &zero_copy_task, void *outside_addr, void *args, size_t offset) {
   const auto addr_val = reinterpret_cast<uintptr_t>(outside_addr);
-  bool set_batch_label_flag = false;
   for (uint32_t out_count = 0; out_count < GetAddrCount(); ++out_count) {
-    auto &addrs_mapping_list = GetOutsideAddrs();
-    auto args_addrs = addrs_mapping_list[out_count].find(outside_addr);
-    if (args_addrs != addrs_mapping_list[out_count].end()) {
+    auto args_addrs = outside_addrs_[out_count].find(outside_addr);
+    if (args_addrs != outside_addrs_[out_count].end()) {
       GE_CHK_STATUS(zero_copy_task.SetTaskArgsOffset(addr_val, offset), "Input args invalid.");
       void *args_val = static_cast<uint8_t *>(args) + offset;
       args_addrs->second.push_back(args_val);
       GELOGD("[ZCPY] set copy input: virtual_addr: 0x%lx, task_addr: %p, args: %p, offset: %zu.", addr_val, args_val,
              args, offset);
-      set_batch_label_flag = true;
     }
   }
-  return set_batch_label_flag;
 }
 
 }  // namespace ge

ge/graph/load/new_model_manager/zero_copy_offset.h

@@ -51,7 +51,7 @@ class ZeroCopyOffset {
                             const OpDescPtr &op_desc, const size_t &idx, bool &fusion_flag);
   void SetOutputOutsideAddrs(const int64_t &input_offset, const bool &fusion_flag, void *addr,
                              std::vector<void *> &tensor_addrs);
-  bool SetOutsideAddrsValue(ZeroCopyTask &zero_copy_task, void *outside_addr, void *args, size_t offset);
+  void SetOutsideAddrsValue(ZeroCopyTask &zero_copy_task, void *outside_addr, void *args, size_t offset);
 
   // basic_addr of l2-fusion
   void *GetBasicAddr() const { return basic_addr_; }

ge/graph/load/new_model_manager/zero_copy_task.cc

@@ -22,8 +22,6 @@
 #include "common/ge_compiler_options.h"
 
 namespace ge {
-const char *const kDefaultBatchLable = "Batch_default";
-
 ZeroCopyTask::ZeroCopyTask(const string &name, uint8_t *args, size_t size)
     : name_(name), args_addr_(args), args_size_(size), is_updated_(false) {}
 
@@ -66,68 +64,23 @@ void ZeroCopyTask::SetOriginalArgs(const void *info, size_t size) {
   const uint8_t *data = static_cast<const uint8_t *>(info);
   args_info_.assign(data, data + size);
 
-  GELOGI("[ZCPY] %s set info from virtual_addr: %p, args_addr: %p, args size: %zu, info size: %zu", name_.c_str(), info,
+  GELOGI("[ZCPY] %s set original args info: %p, args_addr: %p, args size: %zu, info size: %zu", name_.c_str(), info,
          args_addr_, args_size_, size);
 }
 
-/**
- * @ingroup ge
- * @brief Check is dynamic batch node.
- * @param [in] addr: virtual address value from Op.
- * @param [in] data: data buffer from user.
- * @param [in] batch_addrs: dynamic batch addr info.
- * @param [in] batch_label: batch label.
- * @return: true / false
- */
-bool ZeroCopyTask::CheckDynamicBatch(const map<string, set<uintptr_t>> &batch_addrs, const string &batch_label,
-                                     uintptr_t addr) {
-  // Used for dynamic batch / resolution scene
-  set<uintptr_t> dynamic_input_addrs;
-  auto dynamic_input_iter = batch_addrs.find(batch_label);
-  if (dynamic_input_iter != batch_addrs.end()) {
-    dynamic_input_addrs = dynamic_input_iter->second;
-  }
-
-  set<uintptr_t> fix_input_addrs;
-  auto fix_input_iter = batch_addrs.find(kDefaultBatchLable);
-  if (fix_input_iter != batch_addrs.end()) {
-    fix_input_addrs = fix_input_iter->second;
-  }
-
-  if (fix_input_addrs.empty()) {
-    if (!dynamic_input_addrs.empty() && dynamic_input_addrs.find(addr) == dynamic_input_addrs.end()) {
-      return false;
-    }
-  } else {
-    if (!dynamic_input_addrs.empty() && dynamic_input_addrs.find(addr) == dynamic_input_addrs.end() &&
-        fix_input_addrs.find(addr) == fix_input_addrs.end()) {
-      return false;
-    }
-  }
-
-  return true;
-}
-
 /**
  * @ingroup ge
  * @brief Set user data addr to Task param.
  * @param [in] addr: virtual address value from Op.
  * @param [in] buffer_addr: real_data_buffer_addr from user.
- * @param [in] batch_addrs: dynamic batch addr info.
- * @param [in] batch_label: batch label.
  * @return: void
  */
-Status ZeroCopyTask::UpdateTaskParam(uintptr_t addr, void *buffer_addr, const map<string, set<uintptr_t>> &batch_addrs,
-                                     const string &batch_label) {
+Status ZeroCopyTask::UpdateTaskParam(uintptr_t addr, void *buffer_addr) {
   auto iter = task_addr_offset_.find(addr);
   if (iter != task_addr_offset_.end()) {
     auto &cur_pair = *iter;
     uint8_t *args_info = args_info_.data();
     for (auto offset : cur_pair.second) {
-      if (!CheckDynamicBatch(batch_addrs, batch_label, reinterpret_cast<uintptr_t>(args_addr_ + offset))) {
-        continue;
-      }
-
       auto dst_addr = static_cast<uint8_t *>(buffer_addr);
       GELOGI("[ZCPY] %s update task, args_addr: %p, size: %zu, offset: %zu, virtual_addr: 0x%lx, user_data_addr: %p",
              name_.c_str(), args_addr_, args_size_, offset, addr, buffer_addr);

ge/graph/load/new_model_manager/zero_copy_task.h

@@ -67,12 +67,9 @@ class ZeroCopyTask {
    * @brief Set user data addr to Task param.
    * @param [in] addr: virtual address value from Op.
    * @param [in] buffer_addr: data buffer_addr from user.
-   * @param [in] batch_addrs: dynamic batch addr info.
-   * @param [in] batch_label: batch label.
    * @return: 0 SUCCESS / others FAILED
    */
-  ge::Status UpdateTaskParam(uintptr_t addr, void *buffer_addr, const map<string, set<uintptr_t>> &batch_addrs,
-                             const string &batch_label);
+  ge::Status UpdateTaskParam(uintptr_t addr, void *buffer_addr);
 
   /**
    * @ingroup ge
@@ -91,9 +88,6 @@ class ZeroCopyTask {
     return batch_label_;
   }
 
- protected:
-  bool CheckDynamicBatch(const map<string, set<uintptr_t>> &batch_addrs, const string &batch_label, uintptr_t addr);
-
  private:
   const string name_;
 

ge/graph/manager/graph_manager.cc

@@ -56,6 +56,7 @@
 #include "graph/passes/iterator_op_pass.h"
 #include "graph/passes/link_gen_mask_nodes_pass.h"
 #include "graph/passes/mark_graph_unknown_status_pass.h"
+#include "graph/passes/dynamic_single_op_reset_shape_pass.h"
 #include "graph/passes/merge_pass.h"
 #include "graph/passes/merge_input_memcpy_pass.h"
 #include "graph/passes/merge_to_stream_merge_pass.h"
@@ -631,11 +632,22 @@ Status GraphManager::ReplaceSubgraphWithOriGraph(const ComputeGraphPtr &compute_
 
 Status GraphManager::SetSubgraph(uint64_t session_id, ComputeGraphPtr compute_graph, GraphPartitioner &partitioner) {
   GE_CHECK_NOTNULL(compute_graph);
+  PassManager pass_for_dynamic_shape_reset_optimize;
+  GE_CHK_STATUS_RET(pass_for_dynamic_shape_reset_optimize.AddPass(
+    "SetSubgraph::AfterSetSubgraph::DynamicSingleOpResetShapePass", new (std::nothrow) DynamicSingleOpResetShapePass))
+  GE_TIMESTAMP_START(pass_for_dynamic_shape_reset_optimize);
+  Status ret = pass_for_dynamic_shape_reset_optimize.Run(compute_graph);
+  GE_TIMESTAMP_END(pass_for_dynamic_shape_reset_optimize, "SetSubgraph::AfterSetSubgraph");
+  if (ret != SUCCESS && ret != NOT_CHANGED) {
+    GELOGE(ret, "Run passes when optimize subgraph failed");
+    return ret;
+  }
+
   auto sub_graph_map = partitioner.GetSubGraphMap();
   GELOGD("Directly optimize subgraph with build mode:%s, and step:%s.",
          options_.build_mode.c_str(),
          options_.build_step.c_str());
-  Status ret = OptimizeSubGraphWithMultiThreads(compute_graph, sub_graph_map, session_id);
+  ret = OptimizeSubGraphWithMultiThreads(compute_graph, sub_graph_map, session_id);
   if (ret != SUCCESS) {
     GELOGE(ret, "Multiply optimize subgraph failed");
     return ret;

ge/graph/manager/memory_api.cc

@@ -63,7 +63,7 @@ Status RdmaRemoteRegister(const std::vector<HostVarInfo> &var_info, rtMemType_t
   });
 
   auto hcom_remote_mem_register =
-      (HcclResult(*)(const MemRegisterAddr *, uint32_t))dlsym(handle, "hcom_remote_access_mem_register");
+      (HcclResult(*)(const MemRegisterAddr *, uint32_t))dlsym(handle, "HcomRegRemoteAccessMem");
   if (hcom_remote_mem_register == nullptr) {
     GELOGE(FAILED, "Failed to invoke hcom_remote_mem_register function.");
     return FAILED;

ge/graph/passes/dynamic_single_op_reset_shape_pass.cc

@@ -0,0 +1,159 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "graph/passes/dynamic_single_op_reset_shape_pass.h"
+#include "common/ge_inner_error_codes.h"
+#include "graph/utils/node_utils.h"
+#include "graph/utils/graph_utils.h"
+#include "graph/utils/tensor_utils.h"
+#include "graph/utils/op_desc_utils.h"
+#include "graph/utils/type_utils.h"
+#include "graph/debug/ge_attr_define.h"
+
+namespace ge {
+namespace {
+const int64_t kDynamicShapeDim = -2;
+const char *const kEngineNameAiCpu = "DNN_VM_AICPU_ASCEND";
+const char *const kEngineNameAiCpuTf = "DNN_VM_AICPU";
+}  // namespace
+Status DynamicSingleOpResetShapePass::Run(ComputeGraphPtr graph) {
+  GE_CHECK_NOTNULL(graph);
+
+  std::shared_ptr<GELib> instance = ge::GELib::GetInstance();
+  if (instance == nullptr || !instance->InitFlag()) {
+    GELOGE(ge::GE_CLI_GE_NOT_INITIALIZED, "Run CompileNodesPass failed.");
+    return ge::GE_CLI_GE_NOT_INITIALIZED;
+  }
+
+  // pass if graph has not aicpu node.
+  bool is_not_aicpu = false;
+  if (CheckAllAicpuNodes(graph, is_not_aicpu) != SUCCESS) {
+    GELOGE(ge::GE_CLI_GE_NOT_INITIALIZED, "Check if graph has not aicpu node failed.");
+    return ge::GE_CLI_GE_NOT_INITIALIZED;
+  }
+  if (is_not_aicpu) {
+    GELOGI("The graph [%s] has not aicpu node, whose aicpu nodes would not be reset dynamic shape",
+           graph->GetName().c_str());
+    return SUCCESS;
+  }
+
+  for (const auto &node : graph->GetDirectNode()) {
+    GE_CHECK_NOTNULL(node->GetOpDesc());
+    // pass input and output node
+    if (node->GetType() == DATA || node->GetType() == CONSTANT || node->GetType() == CONSTANTOP ||
+        node->GetType() == NETOUTPUT) {
+      continue;
+    }
+
+    // pass node without attr: ATTR_DYNAMIC_SHAPE_SINGLE_AICPU
+    bool single_aicpu_unknown = false;
+    if (!AttrUtils::GetBool(node->GetOpDesc(), ATTR_DYNAMIC_SHAPE_SINGLE_AICPU, single_aicpu_unknown) ||
+        !single_aicpu_unknown) {
+      continue;
+    }
+
+    // reset aicpu shape to unknown shape
+    auto op_desc = node->GetOpDesc();
+    if (ResetOpShape(op_desc) != SUCCESS) {
+      GELOGE(ge::GE_CLI_GE_NOT_INITIALIZED, "Reset node[%s] dynamic shapr failed.", node->GetName().c_str());
+      return ge::GE_CLI_GE_NOT_INITIALIZED;
+    }
+    GELOGD("Reset dynamic aicpu node [%s] shape success!", node->GetName().c_str());
+  }
+
+  GELOGD("Reset dynamic aicpu nodes shape of graph [%s] success!", graph->GetName().c_str());
+  return SUCCESS;
+}
+
+Status DynamicSingleOpResetShapePass::CheckAllAicpuNodes(const ComputeGraphPtr &graph, bool &is_not_aicpu) {
+  is_not_aicpu = false;
+  for (const auto &node : graph->GetDirectNode()) {
+    GE_CHECK_NOTNULL(node->GetOpDesc());
+    // pass input and output node
+    if (node->GetType() == DATA || node->GetType() == CONSTANT || node->GetType() == CONSTANTOP ||
+        node->GetType() == NETOUTPUT) {
+      continue;
+    }
+
+    // find if there are aicpu nodes.
+    auto op_desc = node->GetOpDesc();
+    string engine_name = op_desc->GetOpEngineName();
+    if (engine_name.empty()) {
+      GELOGE(GRAPH_FAILED, "Get engine failed of node[%s].", node->GetName().c_str());
+      return GRAPH_FAILED;
+    }
+    if (engine_name != kEngineNameAiCpu && engine_name != kEngineNameAiCpuTf) {
+      is_not_aicpu = true;
+      return SUCCESS;
+    }
+  }
+  return SUCCESS;
+}
+
+bool DynamicSingleOpResetShapePass::CheckIfConstInput(const GeTensorDescPtr &input_tensor_desc) {
+  bool is_const = false;
+  (void)AttrUtils::GetBool(input_tensor_desc, CONST_ATTR_NAME_INPUT, is_const);
+  return is_const;
+}
+
+Status DynamicSingleOpResetShapePass::ResetOpShape(OpDescPtr &op_desc) {
+  GE_CHECK_NOTNULL(op_desc);
+  std::vector<int64_t> dynamic_shape_dims = {kDynamicShapeDim};
+  GeShape dynamic_shape(dynamic_shape_dims);
+  bool reset_shape_flag = false;
+  if (ResetInputTensorShape(op_desc, dynamic_shape, reset_shape_flag) == SUCCESS && reset_shape_flag) {
+    (void)ResetOutputTensorShape(op_desc, dynamic_shape);
+  }
+  return SUCCESS;
+}
+
+Status DynamicSingleOpResetShapePass::ResetInputTensorShape(OpDescPtr &op_desc, const GeShape &dynamic_shape,
+                                                            bool &reset_shape_flag) {
+  reset_shape_flag = false;
+  GE_CHECK_NOTNULL(op_desc);
+  for (size_t i = 0; i < op_desc->GetAllInputsDesc().size(); i++) {
+    auto input_desc = op_desc->MutableInputDesc(static_cast<uint32_t>(i));
+    GE_CHECK_NOTNULL(input_desc);
+    // pass scalar input desc
+    auto dims_ori = input_desc->GetShape().GetDims();
+    if (dims_ori.size() == 0) {
+      continue;
+    }
+    // pass const input
+    if (CheckIfConstInput(input_desc)) {
+      continue;
+    }
+    reset_shape_flag = true;
+    input_desc->SetShape(dynamic_shape);
+  }
+  return SUCCESS;
+}
+
+Status DynamicSingleOpResetShapePass::ResetOutputTensorShape(OpDescPtr &op_desc, const GeShape &dynamic_shape) {
+  GE_CHECK_NOTNULL(op_desc);
+  for (size_t i = 0; i < op_desc->GetAllOutputsDesc().size(); i++) {
+    auto output_desc = op_desc->MutableOutputDesc(static_cast<uint32_t>(i));
+    GE_CHECK_NOTNULL(output_desc);
+    // pass scalar input desc
+    auto output_dims_ori = output_desc->GetShape().GetDims();
+    if (output_dims_ori.size() == 0) {
+      continue;
+    }
+    output_desc->SetShape(dynamic_shape);
+  }
+  return SUCCESS;
+}
+}  // namespace ge

ge/graph/passes/dynamic_single_op_reset_shape_pass.h

@@ -0,0 +1,36 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef GE_GRAPH_PASSES_DYNAMIC_SINGLE_OP_RESET_SHAPE_PASS_H_
+#define GE_GRAPH_PASSES_DYNAMIC_SINGLE_OP_RESET_SHAPE_PASS_H_
+#include "graph/graph.h"
+#include "inc/graph_pass.h"
+#include "init/gelib.h"
+
+namespace ge {
+class DynamicSingleOpResetShapePass : public GraphPass {
+ public:
+  Status Run(ComputeGraphPtr graph) override;
+
+ private:
+  Status ResetOpShape(OpDescPtr &op_desc);
+  Status ResetInputTensorShape(OpDescPtr &op_desc, const GeShape &dynamic_shape, bool &reset_shape_flag);
+  Status ResetOutputTensorShape(OpDescPtr &op_desc, const GeShape &dynamic_shape);
+  Status CheckAllAicpuNodes(const ComputeGraphPtr &graph, bool &is_not_aicpu);
+  bool CheckIfConstInput(const GeTensorDescPtr &input_tensor_desc);
+};
+}  // namespace ge
+#endif  // GE_GRAPH_PASSES_DYNAMIC_SINGLE_OP_RESET_SHAPE_PASS_H_

ge/graph/preprocess/multi_batch_copy_graph.cc

@@ -1407,11 +1407,13 @@ Status MultiBatchGraphCopyer::InsertIdentityAfterSwitchN() {
 }
 
 Status ProcessMultiBatch(ComputeGraphPtr &graph) {
-  const char *multi_batch_with_case = std::getenv("MULTI_BATCH_WITH_CASE");
-  if (multi_batch_with_case != nullptr) {
-    PassManager pass_manager;
-    GE_CHK_STATUS_RET(pass_manager.AddPass("MultiBatchClonePass", new (std::nothrow) MultiBatchClonePass));
-    return pass_manager.Run(graph);
+  if (GetLocalOmgContext().dynamic_node_type.empty()) {
+    const char *multi_batch_with_switchn = std::getenv("MULTI_BATCH_WITH_SWITCHN");
+    if (multi_batch_with_switchn == nullptr) {
+      PassManager pass_manager;
+      GE_CHK_STATUS_RET(pass_manager.AddPass("MultiBatchClonePass", new (std::nothrow) MultiBatchClonePass));
+      return pass_manager.Run(graph);
+    }
   }
   if (!GetLocalOmgContext().need_multi_batch) {
     GELOGI("No need to process_multi for no_train graph.");

ge/host_cpu_engine/CMakeLists.txt

@@ -193,6 +193,7 @@ target_compile_options(host_cpu_opskernel_builder_static PRIVATE
 
 target_compile_definitions(host_cpu_opskernel_builder_static PRIVATE
     google=ascend_private
+    LOG_CPP
 )
 
 target_include_directories(host_cpu_opskernel_builder_static PRIVATE

ge/hybrid/node_executor/hccl/hccl_node_executor.cc

@@ -42,10 +42,10 @@ Status HcclNodeTask::ExecuteAsync(TaskContext &context, std::function<void()> do
     GELOGE(FAILED, "hccl handle is nullptr! ");
     return FAILED;
   }
-  auto EnqueueHcomOpertion = (HcclResult(*)(HcomOpertion, std::function<void(HcclResult status)>))dlsym(
-      context.handle_, "EnqueueHcomOpertion");
-  if (EnqueueHcomOpertion == nullptr) {
-    GELOGE(FAILED, "Failed to invoke EnqueueHcomOpertion hcom unknown node function.");
+  auto HcomExecEnqueueOperation = (HcclResult(*)(HcomOperation, std::function<void(HcclResult status)>))dlsym(
+      context.handle_, "HcomExecEnqueueOperation");
+  if (HcomExecEnqueueOperation == nullptr) {
+    GELOGE(FAILED, "Failed to invoke HcomExecEnqueueOperation hcom unknown node function.");
     if (dlclose(context.handle_) != 0) {
       GELOGW("Failed to close handle %s", dlerror());
     }
@@ -70,7 +70,7 @@ Status HcclNodeTask::ExecuteAsync(TaskContext &context, std::function<void()> do
   const OpDescPtr op_desc = node_item.GetOpDesc();
   GE_CHECK_NOTNULL(op_desc);
 
-  HcomOpertion op_info;
+  HcomOperation op_info;
   op_info.hcclType = op_desc->GetType();
   op_info.inputPtr = inputs.empty() ? nullptr : inputs[0];
   op_info.outputPtr = outputs.empty() ? nullptr : outputs[0];
@@ -96,7 +96,7 @@ Status HcclNodeTask::ExecuteAsync(TaskContext &context, std::function<void()> do
   op_info.root = root_id;
   auto callback = [this, op_desc](HcclResult status) {
     if (status != HCCL_SUCCESS) {
-      GELOGE(HCCL_E_INTERNAL, "node %s call EnqueueHcomOpertion failed, ret: 0x%X", op_desc->GetName().c_str(), status);
+      GELOGE(HCCL_E_INTERNAL, "node %s call HcomExecEnqueueOperation failed, ret: 0x%X", op_desc->GetName().c_str(), status);
     }
     std::lock_guard<std::mutex> lock(this->hccl_mutex_);
     this->cond_.notify_all();
@@ -110,9 +110,9 @@ Status HcclNodeTask::ExecuteAsync(TaskContext &context, std::function<void()> do
          context.GetNodeName(), op_info.hcclType.c_str(), count, op_info.dataType, op_info.opType, op_info.root);
   op_info.count = count;
 
-  HcclResult hccl_ret = EnqueueHcomOpertion(op_info, callback);
+  HcclResult hccl_ret = HcomExecEnqueueOperation(op_info, callback);
   if (hccl_ret != HCCL_SUCCESS) {
-    GELOGE(HCCL_E_INTERNAL, "Call HcomExcutorInitialize failed, ret: 0x%X", hccl_ret);
+    GELOGE(HCCL_E_INTERNAL, "Call HcomExecInitialize failed, ret: 0x%X", hccl_ret);
     return HCCL_E_INTERNAL;
   }
 
@@ -213,11 +213,11 @@ Status RdmaNodeTask::ExtractTensor(TaskContext &context, vector<HcomRemoteAccess
 
 Status RdmaNodeTask::ExecuteAsync(TaskContext &context, std::function<void()> done_callback) {
   GELOGI("[%s] RdmaNodeTask::ExecuteAsync in.", context.GetNodeName());
-  auto EnqueueRemoteAccess =
+  auto HcomExecEnqueueRemoteAccess =
       (HcclResult(*)(const string &, const vector<HcomRemoteAccessAddrInfo> &,
-                     std::function<void(HcclResult status)>))dlsym(context.handle_, "EnqueueRemoteAccess");
-  if (EnqueueRemoteAccess == nullptr) {
-    GELOGE(FAILED, "Failed to invoke EnqueueRemoteAccess hcom unknown node function.");
+                     std::function<void(HcclResult status)>))dlsym(context.handle_, "HcomExecEnqueueRemoteAccess");
+  if (HcomExecEnqueueRemoteAccess == nullptr) {
+    GELOGE(FAILED, "Failed to invoke HcomExecEnqueueRemoteAccess hcom unknown node function.");
     if (dlclose(context.handle_) != 0) {
       GELOGW("Failed to close handle %s", dlerror());
     }
@@ -228,15 +228,15 @@ Status RdmaNodeTask::ExecuteAsync(TaskContext &context, std::function<void()> do
 
   auto callback = [this](HcclResult status) {
     if (status != HCCL_SUCCESS) {
-      GELOGE(HCCL_E_INTERNAL, "Call HcomExcutorInitialize failed, ret: 0x%X", status);
+      GELOGE(HCCL_E_INTERNAL, "Call HcomExecInitialize failed, ret: 0x%X", status);
     }
     std::lock_guard<std::mutex> lock(this->hccl_mutex_);
     this->cond_.notify_all();
     GELOGI("rdma callback success.");
   };
-  HcclResult hccl_ret = EnqueueRemoteAccess(context.GetNodeItem().NodeType(), addr_infos, callback);
+  HcclResult hccl_ret = HcomExecEnqueueRemoteAccess(context.GetNodeItem().NodeType(), addr_infos, callback);
   if (hccl_ret != HCCL_SUCCESS) {
-    GELOGE(HCCL_E_INTERNAL, "Call HcomExcutorInitialize failed, ret: 0x%X", hccl_ret);
+    GELOGE(HCCL_E_INTERNAL, "Call HcomExecInitialize failed, ret: 0x%X", hccl_ret);
     return HCCL_E_INTERNAL;
   }
 
@@ -307,32 +307,32 @@ Status HcclNodeExecutor::Initialize() {
     GELOGE(GE_PLGMGR_SO_NOT_EXIST, "Failed in dlopen %s! ", dlerror());
     return FAILED;
   }
-  auto HcomExcutorInitialize = (HcclResult(*)())dlsym(handle_, "HcomExcutorInitialize");
-  if (HcomExcutorInitialize == nullptr) {
-    GELOGE(FAILED, "Failed to invoke HcomExcutorInitialize hcom unknown node function.");
+  auto HcomExecInitialize = (HcclResult(*)())dlsym(handle_, "HcomExecInitialize");
+  if (HcomExecInitialize == nullptr) {
+    GELOGE(FAILED, "Failed to invoke HcomExecInitialize hcom unknown node function.");
     return FAILED;
   }
-  HcclResult hccl_ret = HcomExcutorInitialize();
+  HcclResult hccl_ret = HcomExecInitialize();
   if (hccl_ret == HCCL_E_PTR) {
     GELOGI("Hccl comm is null, hcom executor initialize is not required.");
   } else if (hccl_ret == HCCL_SUCCESS) {
     GELOGI("Hcom executor initialize success.");
   } else {
-    GELOGE(FAILED, "Call HcomExcutorInitialize failed, ret: 0x%X", hccl_ret);
+    GELOGE(FAILED, "Call HcomExecInitialize failed, ret: 0x%X", hccl_ret);
     return FAILED;
   }
   return SUCCESS;
 }
 
 Status HcclNodeExecutor::Finalize() {
-  auto HcomExcutorFinalize = (HcclResult(*)())dlsym(handle_, "HcomExcutorFinalize");
-  if (HcomExcutorFinalize == nullptr) {
-    GELOGE(FAILED, "Failed to invoke HcomExcutorFinalize hcom unknown node function.");
+  auto HcomExecFinalize = (HcclResult(*)())dlsym(handle_, "HcomExecFinalize");
+  if (HcomExecFinalize == nullptr) {
+    GELOGE(FAILED, "Failed to invoke HcomExecFinalize hcom unknown node function.");
     return FAILED;
   }
-  HcclResult hccl_ret = HcomExcutorFinalize();
+  HcclResult hccl_ret = HcomExecFinalize();
   if (hccl_ret != HCCL_SUCCESS) {
-    GELOGE(FAILED, "Call HcomExcutorFinalize failed, ret: 0x%X", hccl_ret);
+    GELOGE(FAILED, "Call HcomExecFinalize failed, ret: 0x%X", hccl_ret);
     return FAILED;
   }
   // dlclose file handle

ge/offline/CMakeLists.txt

@@ -27,6 +27,7 @@ target_compile_definitions(atc PRIVATE
     PROTOBUF_INLINE_NOT_IN_HEADERS=0
     COMPILE_OMG_PACKAGE
     google=ascend_private
+    LOG_CPP
 )
 
 target_include_directories(atc PRIVATE
@@ -87,6 +88,7 @@ target_compile_definitions(atc_atc.bin PRIVATE
     PROTOBUF_INLINE_NOT_IN_HEADERS=0
     COMPILE_OMG_PACKAGE
     google=ascend_private
+    LOG_CPP
 )
 
 target_include_directories(atc_atc.bin PRIVATE
@@ -152,6 +154,7 @@ target_compile_definitions(fwk_atc.bin PRIVATE
     PROTOBUF_INLINE_NOT_IN_HEADERS=0
     COMPILE_OMG_PACKAGE
     google=ascend_private
+    LOG_CPP
 )
 
 target_include_directories(fwk_atc.bin PRIVATE

ge/proto/op_mapping_info.proto

@@ -15,6 +15,7 @@ message Output {
     int32 original_output_data_type = 7;
     int32 original_output_format = 8;
     uint64 size = 9;
+    Shape origin_shape = 10;
 }
 
 message Input {
@@ -23,6 +24,7 @@ message Input {
     Shape shape = 3;
     uint64 address = 4;
     uint64 size = 5;
+    Shape origin_shape = 6;
 }
 
 enum BufferType {

inc/external/ge/ge_api_types.h

@@ -395,19 +395,9 @@ const std::set<std::string> ir_builder_suppported_options = {INPUT_FORMAT,
                                                              OP_BANK_UPDATE};
 
 // for interface: aclgrphParse
-const std::set<std::string> ir_parser_suppported_options = {INPUT_FORMAT,
-                                                            INPUT_SHAPE,
-                                                            OP_NAME_MAP,
-                                                            IS_DYNAMIC_INPUT,
-                                                            INPUT_FP16_NODES,
-                                                            IS_INPUT_ADJUST_HW_LAYOUT,
-                                                            IS_OUTPUT_ADJUST_HW_LAYOUT,
-                                                            OUTPUT,
-                                                            OUTPUT_TYPE,
-                                                            OUT_NODES,
-                                                            COMPRESS_WEIGHT_CONF,
-                                                            ENABLE_SCOPE_FUSION_PASSES,
-                                                            LOG_LEVEL};
+const std::set<std::string> ir_parser_suppported_options = {
+  INPUT_FP16_NODES, IS_INPUT_ADJUST_HW_LAYOUT, IS_OUTPUT_ADJUST_HW_LAYOUT, OUTPUT,
+  OUT_NODES,        COMPRESS_WEIGHT_CONF,      ENABLE_SCOPE_FUSION_PASSES};
 
 // for interface: aclgrphBuildInitialize
 const std::set<std::string> global_options = {CORE_TYPE,

inc/framework/common/ge_types.h

@@ -37,6 +37,9 @@ enum FrameworkType {
   MINDSPORE = 1,
   TENSORFLOW = 3,
   ANDROID_NN,
+#ifndef ONLY_COMPILE_OPEN_SRC
+  ONNX,
+#endif
   FRAMEWORK_RESERVED,
 };
 

inc/framework/common/profiling/ge_profiling.h

@@ -20,7 +20,8 @@
 #include "ge/ge_api_error_codes.h"
 #include "toolchain/prof_callback.h"
 
-#define MAX_DEV_NUM (64)
+const int MAX_DEV_NUM = 64;
+
 enum ProfCommandHandleType {
   kProfCommandhandleInit = 0,
   kProfCommandhandleStart,
@@ -32,7 +33,7 @@ enum ProfCommandHandleType {
 
 struct ProfCommandHandleData {
   uint64_t profSwitch;
-  uint32_t devNums; // length of device id list
+  uint32_t devNums;  // length of device id list
   uint32_t devIdList[MAX_DEV_NUM];
   uint32_t modelId;
 };

inc/framework/executor/ge_executor.h

@@ -30,8 +30,6 @@
 #include "runtime/base.h"
 
 namespace ge {
-class ModelListenerAdapter;
-
 class SingleOp;
 class DynamicSingleOp;
 
@@ -55,14 +53,8 @@ class GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY GeExecutor {
   ge::Status Initialize();
   ge::Status Finalize();
 
-  // Load model
-  ge::Status LoadModelOffline(uint32_t &model_id, const std::string &path, const std::string &key, int32_t priority,
-                              std::shared_ptr<ge::ModelListener> listener);
-
   ge::Status UnloadModel(uint32_t modelId);
 
-  ge::Status RunModel(const ge::RunModelData &input_data, ge::RunModelData &output_data);
-
   // Get input and output descriptor
   ge::Status GetModelDescInfo(uint32_t model_id, std::vector<ge::TensorDesc> &input_desc,
                               std::vector<ge::TensorDesc> &output_desc, bool new_model_desc = false);
@@ -168,9 +160,6 @@ class GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY GeExecutor {
   ge::Status GetModelDescInfoForZeroCopy(uint32_t model_id, std::vector<ge::TensorDesc> &input_desc,
                                          std::vector<ge::TensorDesc> &output_desc);
 
-  ge::Status LoadModel(uint32_t &model_id, const ge::ModelData &model_data,
-                       std::shared_ptr<ge::ModelListener> listener);
-
   ge::Status CommandHandle(const ge::Command &command);
 
   ge::Status SetDump(const DumpConfig &dump_config);
@@ -297,8 +286,6 @@ class GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY GeExecutor {
  private:
   static bool isInit_;
 };
-
-ge::Status ModelInfoParser(const ge::ModelData &model, ge::ModelInfo &model_info);
 }  // namespace ge
 
 #endif  // INC_FRAMEWORK_EXECUTOR_GE_EXECUTOR_H_

inc/framework/omg/parser/parser_inner_ctx.h

@@ -59,7 +59,7 @@ struct ParserContext {
   bool train_flag = false;
   domi::domiTensorFormat_t format = domi::DOMI_TENSOR_ND;
   domi::FrameworkType type = domi::FRAMEWORK_RESERVED;
-  RunMode run_mode = ONLY_PRE_CHECK;
+  RunMode run_mode = GEN_OM_MODEL;
   // save caffe custom proto path, used by caffe parse
   std::string custom_proto_path;
   // save caffe proto path, used by caffe parse

tests/ut/ge/CMakeLists.txt

@@ -182,6 +182,7 @@ set(COMMON_SRC_FILES
     "${GE_CODE_DIR}/ge/graph/passes/atomic_addr_clean_pass.cc"
     "${GE_CODE_DIR}/ge/graph/passes/mark_same_addr_pass.cc"
     "${GE_CODE_DIR}/ge/graph/passes/mark_graph_unknown_status_pass.cc"
+    "${GE_CODE_DIR}/ge/graph/passes/dynamic_single_op_reset_shape_pass.cc"
     "${GE_CODE_DIR}/ge/graph/passes/mark_agnostic_pass.cc"
     "${GE_CODE_DIR}/ge/graph/passes/dimension_compute_pass.cc"
     "${GE_CODE_DIR}/ge/graph/passes/dimension_adjust_pass.cc"

third_party/fwkacllib/inc/toolchain/slog.h

@@ -18,9 +18,7 @@
 #define D_SYSLOG_H_
 
 #ifdef __cplusplus
-#ifndef LOG_CPP
 extern "C" {
-#endif
 #endif // __cplusplus
 
 #ifndef LINUX
@@ -107,7 +105,6 @@ extern "C" {
 #define SECURITY_LOG_MASK   (0x00100000)
 #define RUN_LOG_MASK        (0x01000000)
 #define OPERATION_LOG_MASK  (0x10000000)
-#define RESERVERD_LENGTH 52
 
 typedef struct tagDCODE {
   const char *cName;
@@ -119,18 +116,6 @@ typedef struct tagKV {
   char *value;
 } KeyValue;
 
-typedef enum {
-    APPLICATION = 0,
-    SYSTEM
-} ProcessType;
-
-typedef struct {
-    ProcessType type;
-    unsigned int pid;
-    unsigned int deviceId;
-    char reserved[RESERVERD_LENGTH];
-} LogAttr;
-
 /**
  * @ingroup slog
  *
@@ -243,14 +228,6 @@ DLL_EXPORT int dlog_setlevel(int moduleId, int level, int enableEvent);
  */
 DLL_EXPORT int CheckLogLevel(int moduleId, int logLevel);
 
-/**
- * @ingroup slog
- * @brief DlogSetAttr: set log attr, default pid is 0, default device id is 0, default process type is APPLICATION
- * @param [in]logAttr: attr info, include pid(must be larger than 0), process type and device id(chip ID)
- * @return: 0: SUCCEED, others: FAILED
- */
-DLL_EXPORT int DlogSetAttr(LogAttr logAttr);
-
 /**
  * @ingroup slog
  * @brief dlog_error: print error log
@@ -390,8 +367,6 @@ void DlogInner(int moduleId, int level, const char *fmt, ...);
 void DlogWithKVInner(int moduleId, int level, KeyValue *pstKVArray, int kvNum, const char *fmt, ...);
 
 #ifdef __cplusplus
-#ifndef LOG_CPP
 }
-#endif // LOG_CPP
 #endif // __cplusplus
 #endif // D_SYSLOG_H_