code sync for runpackage C75B150

inc/common/optimizer/graph_optimizer.h

@@ -42,6 +42,9 @@ class GraphOptimizer {
   // optimize original graph for FE quant optimize
   virtual Status OptimizeGraphPrepare(ComputeGraph &graph) { return SUCCESS; }
 
+  // optimize graph before build for RTS
+  virtual Status OptimizeGraphBeforeBuild(ComputeGraph &graph) { return SUCCESS; }
+
   // optimize original graph, using in graph preparation stage
   virtual Status OptimizeOriginalGraph(ComputeGraph &graph) = 0;
 

inc/common/util/ai_core/common/aicore_util_attr_define.h

@@ -0,0 +1,39 @@
+/**
+ * Copyright 2019-2020 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 INC_COMMON_UTILS_AI_CORE_COMMON_ATTR_DEFINE_H_
+#define INC_COMMON_UTILS_AI_CORE_COMMON_ATTR_DEFINE_H_
+
+#include <string>
+
+namespace fe {
+static const std::string SCOPE_ID_ATTR = "fusion_scope";
+
+static const std::string FE_IMPLY_TYPE = "_fe_imply_type";
+
+static const std::string PARENT_OP_TYPE = "parentOpType";
+
+static const std::string ATTR_NAME_TASK_L2_FUSION_INFO_EXTEND_PTR = "task_l2_fusion_info_extend_content";
+
+static const std::string ATTR_DATA_DUMP_REF = "_datadump_ref";
+
+static const std::string ATTR_NAME_L2_FUSION_EXTEND_PTR = "l2_fusion_extend_content";
+
+static const std::string L1_OPTIMIZED = "l1_optimized";
+
+static const std::string L2_OPTIMIZED = "l2_optimized";
+}  // namespace fe
+#endif

inc/common/util/ai_core/common/aicore_util_types.h

@@ -0,0 +1,118 @@
+/**
+ * Copyright 2019-2020 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 INC_COMMON_UTILS_AI_CORE_COMMON_TYPES_H_
+#define INC_COMMON_UTILS_AI_CORE_COMMON_TYPES_H_
+
+#include "graph/anchor.h"
+#include "graph/types.h"
+#include "runtime/kernel.h"
+#include <map>
+#include <string>
+#include <vector>
+
+namespace fe {
+struct FusionOpSrc {
+  uint32_t src_op_id;
+  ge::AnchorPtr src_anchor;
+  int32_t fusion_src_index;
+  int32_t fusion_dst_index;
+};
+
+struct FusionOpDst {
+  uint32_t dst_op_id;
+  ge::AnchorPtr dst_anchor;
+};
+
+struct FusionDataFlow {
+  std::pair<ge::AnchorPtr, ge::AnchorPtr> edge;
+  std::pair<std::string, ge::AnchorPtr> node_dataindex_pair;
+};
+
+typedef struct tagL2FusionData {
+  uint32_t l2Index;
+  uint64_t l2Addr;
+  uint64_t l2PageNum;
+} L2FusionData_t;
+typedef std::map<uint64_t, L2FusionData_t> L2FusionDataMap_t;
+
+typedef struct tagFeSmDesc {
+  rtL2Ctrl_t l2ctrl;
+  std::string nodeName[8];
+  uint8_t outputIndex[8];
+} feSmDesc_t;
+
+typedef struct TagTaskL2FusionInfo {
+  std::string nodeName;
+  feSmDesc_t l2Info;
+  L2FusionDataMap_t input;
+  L2FusionDataMap_t output;
+  uint32_t isUsed;
+} TaskL2FusionInfo_t;
+
+using L2FusionInfoPtr = std::shared_ptr<TaskL2FusionInfo_t>;
+
+typedef struct ToOpStruct {
+  int64_t opL1Space = 0;
+  std::vector<int64_t> opL1FusionType;
+  int64_t opL1WorkspaceFlag = 0;  // for workspace flag
+  int64_t opL1WorkspaceSize = 0;
+  std::vector<std::vector<int64_t>> validInputShape;
+  std::vector<std::vector<int64_t>> validOutputShape;
+  std::vector<std::vector<int64_t>> sliceInputOffset;   // conv & pooling & ReadSelect
+  std::vector<std::vector<int64_t>> sliceOutputOffset;  // WriteSelect
+  std::vector<uint32_t> totalShape;
+  uint32_t splitIndex = 0;
+  ToOpStruct() {
+    // set invalid value for essential variable
+    opL1Space = -1;
+    opL1WorkspaceSize = -1;
+  }
+} ToOpStruct_t;
+
+enum OpImplType {
+  EN_IMPL_CUSTOM_CONSTANT_CCE = 0,    // custom constant op
+  EN_IMPL_CUSTOM_TIK,                 // custom tik op
+  EN_IMPL_CUSTOM_TBE,                 // custom tbe op
+  EN_IMPL_HW_CONSTANT_CCE,            // Huawei built-in constant op
+  EN_IMPL_HW_GENERAL_CCE,             // Huawei built-in cce op
+  EN_IMPL_HW_TIK,                     // Huawei built-in tik op
+  EN_IMPL_HW_TBE,                     // Huawei built-in tbe op
+  EN_IMPL_RL,                         // RL op
+  EN_IMPL_PLUGIN_TBE,                 // Huawei built-in tbe plugin op
+  EN_IMPL_VECTOR_CORE_HW_TBE,         // Huawei built-in tbe op
+  EN_IMPL_VECTOR_CORE_CUSTOM_TBE,     // custom tbe op
+  EN_IMPL_NON_PERSISTENT_CUSTOM_TBE,  // custom tbe op
+  EN_RESERVED                         // reserved value
+};
+
+static const std::map<ge::DataType, uint32_t> DATATYPE_SIZE_MAP{{ge::DT_FLOAT, sizeof(float)},
+                                                                {ge::DT_FLOAT16, sizeof(int16_t)},
+                                                                {ge::DT_INT8, sizeof(int8_t)},
+                                                                {ge::DT_INT32, sizeof(int32_t)},
+                                                                {ge::DT_UINT8, sizeof(uint8_t)},
+                                                                {ge::DT_UINT32, sizeof(uint32_t)},
+                                                                {ge::DT_INT16, sizeof(int16_t)},
+                                                                {ge::DT_UINT16, sizeof(uint16_t)},
+                                                                {ge::DT_INT64, sizeof(int64_t)},
+                                                                {ge::DT_UINT64, sizeof(uint64_t)},
+                                                                {ge::DT_DOUBLE, sizeof(double)},
+                                                                {ge::DT_BOOL, sizeof(bool)},
+                                                                {ge::DT_DUAL, sizeof(float) + sizeof(int8_t)},
+                                                                {ge::DT_DUAL_SUB_UINT8, sizeof(int8_t)},
+                                                                {ge::DT_DUAL_SUB_INT8, sizeof(int8_t)}};
+}  // namespace fe
+#endif

inc/common/util/ai_core/common/graph_comm.h

@@ -0,0 +1,107 @@
+/**
+ * Copyright 2019-2020 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 INC_COMMON_UTILS_AI_CORE_COMMON_GRAPH_COMMON_H_
+#define INC_COMMON_UTILS_AI_CORE_COMMON_GRAPH_COMMON_H_
+
+#include "graph/compute_graph.h"
+#include "common/aicore_util_types.h"
+#include "register/graph_optimizer/graph_optimize_register_error_codes.h"
+
+#include <map>
+#include <string>
+#include <utility>
+#include <vector>
+
+namespace fe {
+
+using kScopeNodeMap_t = std::map<int64_t, std::vector<ge::NodePtr>>;
+using kScopeNodePair_t = std::pair<int64_t, std::vector<ge::NodePtr>>;
+
+class GraphCommImpl;
+using GraphCommImplPtr = std::unique_ptr<GraphCommImpl>;
+
+class GraphComm {
+ public:
+  GraphComm(const string &engineName);
+  virtual ~GraphComm();
+  GraphComm(const GraphComm &in) = delete;
+  GraphComm &operator=(const GraphComm &in) = delete;
+
+  Status GetscopeNodeMap(ge::ComputeGraph &graph, kScopeNodeMap_t &fusionMap);
+
+  Status CopyFusionOpNodes(vector<FusionDataFlow> &fusInputEdgeList, vector<FusionDataFlow> &fusOutputEdgeList,
+                           vector<ge::NodePtr> &fusNodelist, ge::OpDescPtr fusionOpDesc,
+                           ge::ComputeGraphPtr fusionGraph);
+
+  Status CopyFusionOpEdges(ge::OpDescPtr fusionOpDesc, ge::ComputeGraph &origGraph, ge::ComputeGraphPtr fusionGraph);
+
+  Status GetNodeDataFlowMap(const ge::NodePtr &fusNode,
+                            std::map<ge::NodePtr, std::map<ge::AnchorPtr, ge::AnchorPtr>> &fusionOpAnchorsMap,
+                            ge::kFusionDataFlowVec_t &fusDataflowList, const int &mapType);
+
+  Status GetFusionNodeEdgeList(std::vector<ge::NodePtr> &fusNodelist, std::vector<FusionDataFlow> &fusInputEdgeList,
+                               std::vector<FusionDataFlow> &fusOutputEdgeList);
+  void ClearFusionSrc();
+
+  void ClearFusionDst();
+
+  void AddFusionOutputSrc(const uint32_t &src_op_id, const ge::AnchorPtr &src_anchor, const int32_t &fusion_src_index,
+                          std::pair<string, ge::AnchorPtr> &node_dataindex_pair);
+
+  void AddFusionInputSrc(const uint32_t &src_op_id, const ge::AnchorPtr &src_anchor, const int32_t &fusion_dst_index,
+                         std::pair<string, ge::AnchorPtr> &node_dataindex_pair);
+
+  void SaveFusionDst(const uint32_t &dst_op_id, ge::AnchorPtr dst_anchor);
+
+  bool IsFusionDstExist(const uint32_t &dst_op_id, const ge::AnchorPtr &dst_anchor);
+
+  bool GetFusionSrc(const uint32_t &src_op_id, const ge::AnchorPtr &src_anchor, int32_t &fusion_src_index,
+                    int32_t &fusion_dst_index);
+
+  Status GetFusionNodeCtrlEdgeList(vector<ge::NodePtr> &fusNodelist, vector<FusionDataFlow> &fusInputCtrlEdgeList,
+                                   vector<FusionDataFlow> &fusOutputCtrlEdgeList);
+
+  Status MergeFusionNodeEdgeList(ge::NodePtr &fusNode, vector<ge::NodePtr> &fusNodelist,
+                                 vector<FusionDataFlow> &fusInputEdgeList, vector<FusionDataFlow> &fusOutputEdgeList);
+
+  Status MergeFusionNodeCtrlEdgeList(ge::NodePtr &fusNode, vector<ge::NodePtr> &fusNodelist,
+                                     vector<FusionDataFlow> &fusInputEdgeList,
+                                     vector<FusionDataFlow> &fusOutputEdgeList);
+
+  string GetEngineName();
+
+ private:
+  Status MergeFusionNodeInputEdgeList(ge::NodePtr fusNode, std::vector<ge::NodePtr> &fusNodelist,
+                                      std::vector<FusionDataFlow> &fusInputEdgeList);
+  Status MergeFusionNodeOutputEdgeList(ge::NodePtr fusNode, std::vector<ge::NodePtr> &fusNodelist,
+                                       std::vector<FusionDataFlow> &fusOutputEdgeList);
+
+  string engineName_;
+
+  std::vector<FusionOpSrc> exist_fusion_src_list_;
+  std::vector<FusionOpDst> exist_fusion_dst_list_;
+
+  // std::vector<std::multimap<std::string, uint32_t>>
+  ge::kFusionDataFlowVec_t fusion_input_dataflow_list_;
+
+  // std::vector<std::multimap<std::string, ge::AnchorPtr>>
+  ge::kFusionDataFlowVec_t fusion_output_dataflow_list_;
+
+  GraphCommImplPtr graphCommImplPtr_;
+};
+}  // namespace fe
+#endif

inc/common/util/ai_core/common/scope_allocator.h

@@ -0,0 +1,42 @@
+/**
+ * Copyright 2019-2020 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 INC_COMMON_UTILS_AI_CORE_COMMON_SCOPE_ALLOCATOR_H_
+#define INC_COMMON_UTILS_AI_CORE_COMMON_SCOPE_ALLOCATOR_H_
+
+#include "graph/op_desc.h"
+
+namespace fe {
+class ScopeAllocator {
+ public:
+  ScopeAllocator();
+  virtual ~ScopeAllocator();
+  ScopeAllocator(const ScopeAllocator& in) = delete;
+  ScopeAllocator& operator=(const ScopeAllocator& in) = delete;
+
+ public:
+  void Init();
+  int64_t GetCurrentScopeId();
+  int64_t AllocateScopeId(void);
+  bool HasScopeAttr(ge::ConstOpDescPtr opdef);
+  bool GetScopeAttr(ge::ConstOpDescPtr opdef, int64_t& scopeId);
+  bool SetScopeAttr(ge::OpDescPtr opdef, int64_t scopeId);
+
+ private:
+  int64_t scopeId;
+};
+}  // namespace fe
+#endif

src/ge/graph/passes/switch_split_pass.h → inc/common/util/ai_core/param_calculate/aicore_param_calculator.h

@@ -14,15 +14,20 @@
  * limitations under the License.
  */
 
-#ifndef GE_GRAPH_PASSES_SWITCH_SPLIT_PASS_H_
-#define GE_GRAPH_PASSES_SWITCH_SPLIT_PASS_H_
+#ifndef AICORE_PARAM_CALCULATOR
+#define AICORE_PARAM_CALCULATOR
 
-#include <set>
-#include "graph/passes/base_pass.h"
-namespace ge {
-class SwitchSplitPass : public BaseNodePass {
+#include "graph/node.h"
+#include "graph_optimizer/graph_optimize_register_error_codes.h"
+
+namespace fe {
+class AICoreParamCalculator {
  public:
-  Status Run(NodePtr &node) override;
+  AICoreParamCalculator();
+
+  ~AICoreParamCalculator();
+
+  Status CalcOpRunningParam(ge::Node &node);
 };
-}  // namespace ge
-#endif  // GE_GRAPH_PASSES_SWITCH_SPLIT_PASS_H_
+}  // namespace fe
+#endif  // AICORE_PARAM_CALCULATOR

inc/common/util/ai_core/param_calculate/tensorsize_calculator.h

@@ -0,0 +1,45 @@
+/**
+ * Copyright 2019-2020 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 TENSORSIZE_CALCULATOR_H
+#define TENSORSIZE_CALCULATOR_H
+
+#include "graph_optimizer/graph_optimize_register_error_codes.h"
+
+#include <map>
+#include <string>
+#include "graph/compute_graph.h"
+#include "graph/op_desc.h"
+
+namespace fe {
+class TensorSizeCalculator {
+ public:
+  /**
+   * Calculate the tensor size of input and output of each opdesc
+   * @param opDesc opdesc object
+   * @param opImplType op impl type
+   * @return status SUCCESS or FAILED
+   */
+  static Status CalculateOpTensorSize(ge::OpDesc &opDesc);
+
+ private:
+  static Status CalcInputOpTensorSize(ge::OpDesc &opDesc, int32_t &outputRealCalcFlag);
+
+  static Status CalcOutputOpTensorSize(ge::OpDesc &opDesc, int32_t &outputRealCalcFlag);
+};
+}  // namespace fe
+
+#endif  // TENSORSIZE_CALCULATOR_H

inc/external/ge/ge_api.h

@@ -98,6 +98,15 @@ class GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY Session {
   ///
   Status RunGraphAsync(uint32_t graphId, const std::vector<ge::InputTensorInfo> &inputs, RunAsyncCallback callback);
 
+  ///
+  /// @ingroup ge_graph
+  /// @brief get variables in the session with specific session id
+  /// @param [in] var_names: variable names
+  /// @param [out] var_values: variable values
+  /// @return Status result of function
+  ///
+  Status GetVariables(const std::vector<std::string> &var_names, std::vector<Tensor> &var_values);
+
   ///
   /// @ingroup ge_graph
   /// @brief register callback func with specific summary or checkpoint by users

inc/external/ge/ge_ir_build.h

@@ -23,6 +23,12 @@
 #include "graph/graph.h"
 #include "graph/ge_error_codes.h"
 
+namespace {
+#define IR_MAJOR_VERSION (int(1))
+#define IR_MINOR_VERSION (int(0))
+#define IR_PATCH_VERSION (int(0))
+}  // namespace
+
 namespace ge {
 
 struct ModelBufferData {
@@ -71,5 +77,17 @@ graphStatus aclgrphBuildModel(const ge::Graph &graph, const std::map<std::string
  */
 graphStatus aclgrphSaveModel(const string &output_file, const ModelBufferData &model);
 
+/**
+ * @ingroup AscendCL
+ * @brief query IR interface version
+ *
+ * @param major_version[OUT] IR interface major version
+ * @param minor_version[OUT] IR interface minor version
+ * @param patch_version[OUT] IR interface patch version
+ * @retval GRAPH_SUCCESS The function is successfully executed.
+ * @retval OtherValues Failure
+ */
+graphStatus aclgrphGetIRVersion(int *major_version, int *minor_version, int *patch_version);
+
 };  // namespace ge
 #endif

inc/external/graph/operator.h

@@ -45,9 +45,11 @@
 namespace ge {
 class Operator;
 class OperatorImpl;
+class NodeUtils;
 class NamedAttrs;
 class Graph;
 class AttrValue;
+class Node;
 
 using SubgraphBuilder = std::function<Graph()>;
 using OperatorImplPtr = std::shared_ptr<OperatorImpl>;
@@ -65,8 +67,8 @@ using std::string;
 class GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY Operator {
  public:
   friend class OperatorImpl;
-
   friend class GraphBuilderImpl;
+  friend class NodeUtils;
 
   using OpInt = int64_t;
   using OpFloat = float;
@@ -104,6 +106,8 @@ class GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY Operator {
 
   Operator &SetInput(const string &dst_name, const Operator &src_oprt, const string &name);  // lint !e148
 
+  Operator &SetInput(const string &dst_name, const Operator &src_oprt, uint32_t index);
+
   Operator &AddControlInput(const Operator &src_oprt);
 
   graphStatus GetInputConstData(const string &dst_name, Tensor &data) const;
@@ -269,11 +273,15 @@ class GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY Operator {
 
   OutHandler GetOutput(const string &name) const;
 
+  OutHandler GetOutput(uint32_t index) const;
+
   OperatorImplPtr GetOperatorImplPtr() const;
 
   OperatorImplPtr operator_impl_{nullptr};
 
   graphStatus GetInputConstDataOut(const string &dst_name, Tensor &data) const;
+
+  std::shared_ptr<const Node> GetNode() const;
 };
 /*lint +e148*/
 }  // namespace ge

inc/external/graph/operator_reg.h

@@ -130,6 +130,10 @@ class OpReg {
     Operator::SetInput(#x, v, srcName);                                        \
     return *this;                                                              \
   }                                                                            \
+  _THIS_TYPE &set_input_##x(Operator &v, uint32_t index) {                     \
+    Operator::SetInput(#x, v, index);                                          \
+    return *this;                                                              \
+  }                                                                            \
   _THIS_TYPE &set_input_##x(Operator &v) {                                     \
     Operator::SetInput(#x, v);                                                 \
     return *this;                                                              \
@@ -159,6 +163,10 @@ class OpReg {
     Operator::SetInput(#x, v, srcName);                                        \
     return *this;                                                              \
   }                                                                            \
+  _THIS_TYPE &set_input_##x(Operator &v, uint32_t index) {                     \
+    Operator::SetInput(#x, v, index);                                          \
+    return *this;                                                              \
+  }                                                                            \
   TensorDesc get_input_desc_##x() const { return Operator::GetInputDesc(#x); } \
   graphStatus update_input_desc_##x(const TensorDesc &tensorDesc) {            \
     return Operator::UpdateInputDesc(#x, tensorDesc);                          \

inc/external/register/scope/scope_fusion_pass_register.h

@@ -0,0 +1,331 @@
+/**
+ * Copyright 2019-2020 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 EXTERNAL_REGISTER_SCOPE_SCOPE_FUSION_PASS_REGISTER_H_
+#define EXTERNAL_REGISTER_SCOPE_SCOPE_FUSION_PASS_REGISTER_H_
+
+#include <memory>
+#include <string>
+#include <vector>
+#include <map>
+#include "ge/ge_api_error_codes.h"
+#include "register/register_error_codes.h"
+#include "register/register_types.h"
+#include "graph/operator.h"
+
+#define CHECK_INNER_NODE_CONDITION(cond, fusion_rlt)  \
+  do {                                                \
+    if (!(cond)) {                                    \
+      if ((fusion_rlt) != nullptr) {                  \
+        (fusion_rlt)->SetType(ge::kScopeInvalidType); \
+      }                                               \
+      return;                                         \
+    }                                                 \
+  } while (0)
+
+namespace domi {
+class TensorFlowModelParser;
+}  // namespace domi
+namespace ge {
+const int32_t kFusionDisableIndex = 99999;
+const char *const kScopeToMultiNodes = "ScopeToMultiNodes";
+const char *const kScopeInvalidType = "ScopeInvalidType";
+const char *const kInputFromFusionScope = "InputFromFusionScope";
+const char *const kOutputToFusionScope = "OutputToFusionScope";
+class ScopePattern;
+using ScopeFusionPatterns = std::vector<std::vector<ScopePattern *>>;
+
+class ScopePassManager;
+
+class GE_FUNC_HOST_VISIBILITY GE_FUNC_DEV_VISIBILITY Scope {
+ public:
+  explicit Scope(const std::string &name, const std::string &sub_type = "", Scope *father_scope = nullptr);
+  ~Scope();
+
+  std::string Name() const;
+  std::string SubType() const;
+  std::map<std::string, ge::OperatorPtr> AllNodesMap() const;
+  Scope *GetSubScope(const std::string &scope_name) const;
+  std::string LastName() const;
+  std::vector<Scope *> GetAllSubScopes() const;
+  const Scope *GetFatherScope() const;
+
+ private:
+  class ScopeImpl;
+  std::unique_ptr<ScopeImpl> impl_;
+  friend class ScopeBasePass;
+  friend class ScopeTree;
+  friend class NodeOpTypeFeature;
+  friend class NodeAttrFeature;
+  friend class ScopeFeature;
+};
+
+class GE_FUNC_HOST_VISIBILITY GE_FUNC_DEV_VISIBILITY FusionScopesResult {
+ public:
+  FusionScopesResult();
+  ~FusionScopesResult();
+  void SetName(const std::string &name);
+  void SetType(const std::string &type);
+  void SetDescription(const std::string &description);
+  std::string Name() const;
+  std::vector<ge::OperatorPtr> Nodes() const;
+  void InsertInputs(const std::string &inner_op_name, const std::vector<int32_t> &index_map);
+  void InsertOutputs(const std::string &inner_op_name, const std::vector<int32_t> &index_map);
+
+  class InnerNodeInfo {
+   public:
+    explicit InnerNodeInfo(const std::string &fusion_node_name);
+    InnerNodeInfo(const std::string &fusion_node_name, const std::string &name, const std::string &type);
+    InnerNodeInfo(InnerNodeInfo &&other) noexcept;
+    InnerNodeInfo &operator=(InnerNodeInfo &&other) noexcept;
+    InnerNodeInfo(const InnerNodeInfo &) = delete;
+    InnerNodeInfo &operator=(const InnerNodeInfo &) = delete;
+    ~InnerNodeInfo();
+    InnerNodeInfo &SetName(const std::string &name);
+    InnerNodeInfo &SetType(const std::string &type);
+    InnerNodeInfo &InsertInput(const std::string &input_node, int32_t peer_out_idx);
+    InnerNodeInfo &InsertOutput(const std::string &output_node, int32_t peer_in_idx);
+    ge::graphStatus BuildInnerNode();
+    ge::graphStatus SetInputFormat(const std::string &input_name, const std::string &format);
+    ge::graphStatus SetOutputFormat(const std::string &output_name, const std::string &format);
+    ge::graphStatus SetDynamicInputFormat(const std::string &input_name, uint32_t index, const std::string &format);
+    ge::graphStatus SetDynamicOutputFormat(const std::string &output_name, uint32_t index, const std::string &format);
+    ge::Operator *MutableOperator();
+
+    std::string GetName() const;
+    std::string GetType() const;
+    std::vector<std::pair<std::string, int32_t>> GetInputs() const;
+    std::vector<std::pair<std::string, int32_t>> GetOutputs() const;
+
+   private:
+    class InnerNodeInfoImpl;
+    std::unique_ptr<InnerNodeInfoImpl> impl_;
+  };
+
+  InnerNodeInfo *AddInnerNode(const std::string &name, const std::string &type);
+  InnerNodeInfo *MutableRecentInnerNode();
+  InnerNodeInfo *MutableInnerNode(uint32_t index);
+  ge::graphStatus CheckInnerNodesInfo();
+
+ private:
+  class FusionScopesResultImpl;
+  std::unique_ptr<FusionScopesResultImpl> impl_;
+  friend class ScopeGraph;
+  friend class ScopeBasePass;
+  friend class TensorFlowModelParser;
+};
+
+class GE_FUNC_HOST_VISIBILITY GE_FUNC_DEV_VISIBILITY ScopeTree {
+ public:
+  ScopeTree();
+  Status Init();
+  ScopeTree(const ScopeTree &scopetree) = delete;
+  ScopeTree &operator=(const ScopeTree &scopetree) = delete;
+  ~ScopeTree();
+
+  std::vector<Scope *> GetAllScopes() const;
+
+ private:
+  class ScopeTreeImpl;
+  std::unique_ptr<ScopeTreeImpl> impl_;
+  friend class ScopeGraph;
+  friend class ScopeBasePass;
+};
+
+class GE_FUNC_HOST_VISIBILITY GE_FUNC_DEV_VISIBILITY ScopeGraph {
+ public:
+  ScopeGraph();
+  Status Init();
+  ScopeGraph(const ScopeGraph &scope_graph) = delete;
+  ScopeGraph &operator=(const ScopeGraph &scope_graph) = delete;
+  ~ScopeGraph();
+
+  const ScopeTree *GetScopeTree() const;
+  std::map<std::string, ge::OperatorPtr> GetNodesMap() const;
+
+ private:
+  class ScopeGraphImpl;
+  std::unique_ptr<ScopeGraphImpl> impl_;
+  friend class ScopePassManager;
+  friend class ScopeBasePass;
+  friend class TensorFlowModelParser;
+};
+
+class GE_FUNC_HOST_VISIBILITY GE_FUNC_DEV_VISIBILITY ScopeAttrValue {
+ public:
+  ScopeAttrValue();
+  ScopeAttrValue(ScopeAttrValue const &attr_value);
+  ScopeAttrValue &operator=(ScopeAttrValue const &attr_value);
+  ~ScopeAttrValue();
+
+  void SetIntValue(int64_t value);
+  void SetFloatValue(float value);
+  void SetStringValue(std::string value);
+  void SetBoolValue(bool value);
+
+ private:
+  class ScopeAttrValueImpl;
+  std::unique_ptr<ScopeAttrValueImpl> impl_;
+  friend class NodeAttrFeature;
+};
+
+class GE_FUNC_HOST_VISIBILITY GE_FUNC_DEV_VISIBILITY ScopeBaseFeature {
+ public:
+  virtual bool Match(const Scope *scope) = 0;
+  virtual ~ScopeBaseFeature(){};
+};
+
+class GE_FUNC_HOST_VISIBILITY GE_FUNC_DEV_VISIBILITY NodeOpTypeFeature : ScopeBaseFeature {
+ public:
+  NodeOpTypeFeature(std::string nodeType, int num, int step = 0);
+  NodeOpTypeFeature(NodeOpTypeFeature const &feature);
+  NodeOpTypeFeature &operator=(NodeOpTypeFeature const &feature);
+  ~NodeOpTypeFeature();
+  bool Match(const Scope *scope) override;
+
+ private:
+  class NodeOpTypeFeatureImpl;
+  std::unique_ptr<NodeOpTypeFeatureImpl> impl_;
+};
+
+class GE_FUNC_HOST_VISIBILITY GE_FUNC_DEV_VISIBILITY NodeAttrFeature : ScopeBaseFeature {
+ public:
+  NodeAttrFeature(std::string nodeType, std::string attr_name, ge::DataType datatype, ScopeAttrValue attr_value);
+  NodeAttrFeature(NodeAttrFeature const &feature);
+  NodeAttrFeature &operator=(NodeAttrFeature const &feature);
+  ~NodeAttrFeature();
+  bool Match(const Scope *scope) override;
+
+ private:
+  class NodeAttrFeatureImpl;
+  std::unique_ptr<NodeAttrFeatureImpl> impl_;
+};
+
+class GE_FUNC_HOST_VISIBILITY GE_FUNC_DEV_VISIBILITY ScopeFeature : ScopeBaseFeature {
+ public:
+  ScopeFeature(std::string sub_type, int32_t num, std::string suffix = "", std::string sub_scope_mask = "",
+               int step = 0);
+  ScopeFeature(ScopeFeature const &feature);
+  ScopeFeature &operator=(ScopeFeature const &feature);
+  ~ScopeFeature();
+  bool Match(const Scope *scope) override;
+
+ private:
+  class ScopeFeatureImpl;
+  std::unique_ptr<ScopeFeatureImpl> impl_;
+};
+
+class GE_FUNC_HOST_VISIBILITY GE_FUNC_DEV_VISIBILITY ScopePattern {
+ public:
+  ScopePattern();
+  ~ScopePattern();
+
+  ScopePattern &SetSubType(const std::string &sub_type);
+  ScopePattern &AddNodeOpTypeFeature(NodeOpTypeFeature feature);
+  ScopePattern &AddNodeAttrFeature(NodeAttrFeature feature);
+  ScopePattern &AddScopeFeature(ScopeFeature feature);
+
+ private:
+  class ScopePatternImpl;
+  std::unique_ptr<ScopePatternImpl> impl_;
+  friend class ScopeBasePass;
+};
+
+class GE_FUNC_HOST_VISIBILITY GE_FUNC_DEV_VISIBILITY ScopesResult {
+ public:
+  ScopesResult();
+  ScopesResult(ScopesResult const &result);
+  ScopesResult &operator=(ScopesResult const &result);
+  ~ScopesResult();
+
+  void SetScopes(std::vector<Scope *> &scopes);
+  void SetNodes(std::vector<ge::OperatorPtr> &nodes);
+
+ private:
+  class ScopesResultImpl;
+  std::unique_ptr<ScopesResultImpl> impl_;
+  friend class ScopeBasePass;
+};
+
+class GE_FUNC_HOST_VISIBILITY GE_FUNC_DEV_VISIBILITY ScopeBasePass {
+ public:
+  ScopeBasePass();
+  virtual ~ScopeBasePass();
+
+ protected:
+  // Subclasses implement respective fusion strategies and build the Patterns
+  virtual std::vector<ScopeFusionPatterns> DefinePatterns() = 0;
+  // Define the name of the scope pass
+  virtual std::string PassName() = 0;
+  // Subclasses implement respective multi-scope or operator fusion methods across scopes
+  virtual Status LastMatchScopesAndOPs(std::shared_ptr<ScopeGraph> &scope_graph,
+                                       std::vector<ScopesResult> &results) = 0;
+  // Subclasses implement their own results and set the input and output of the final fusion operator
+  virtual void GenerateFusionResult(const std::vector<Scope *> &scopes, FusionScopesResult *fusion_rlt) = 0;
+
+ private:
+  class ScopeBasePassImpl;
+  std::unique_ptr<ScopeBasePassImpl> impl_;
+  friend class ge::ScopePassManager;
+  friend class ScopeBasePassImpl;
+};
+
+class GE_FUNC_HOST_VISIBILITY GE_FUNC_DEV_VISIBILITY ScopeFusionPassRegistry {
+ public:
+  using CreateFn = ScopeBasePass *(*)();
+  ~ScopeFusionPassRegistry();
+
+  static ScopeFusionPassRegistry &GetInstance() {
+    static ScopeFusionPassRegistry instance;
+    return instance;
+  }
+
+  void RegisterScopeFusionPass(const std::string &pass_name, CreateFn create_fn, bool is_general);
+
+ private:
+  ScopeFusionPassRegistry();
+  class ScopeFusionPassRegistryImpl;
+  /*lint -e148*/
+  std::unique_ptr<ScopeFusionPassRegistryImpl> impl_;
+  friend class TensorFlowModelParser;
+};
+
+class GE_FUNC_HOST_VISIBILITY GE_FUNC_DEV_VISIBILITY ScopeUtil {
+ public:
+  static std::string StringReplaceAll(std::string str, const std::string &old_value, const std::string &new_value);
+  static void FreeScopePatterns(ScopeFusionPatterns &patterns);
+  static void FreeOneBatchPattern(std::vector<ScopePattern *> &one_batch_pattern);
+};
+
+class GE_FUNC_HOST_VISIBILITY GE_FUNC_DEV_VISIBILITY ScopeFusionPassRegistrar {
+ public:
+  ScopeFusionPassRegistrar(const char *pass_name, ScopeBasePass *(*create_fn)(), bool is_general);
+  ~ScopeFusionPassRegistrar() {}
+};
+
+#define REGISTER_SCOPE_FUSION_PASS(pass_name, scope_pass, is_general) \
+  REGISTER_SCOPE_FUSION_PASS_UNIQ_HELPER(__COUNTER__, pass_name, scope_pass, is_general)
+
+#define REGISTER_SCOPE_FUSION_PASS_UNIQ_HELPER(ctr, pass_name, scope_pass, is_general) \
+  REGISTER_SCOPE_FUSION_PASS_UNIQ(ctr, pass_name, scope_pass, is_general)
+
+#define REGISTER_SCOPE_FUSION_PASS_UNIQ(ctr, pass_name, scope_pass, is_general)                   \
+  static ::ge::ScopeFusionPassRegistrar register_scope_fusion_pass##ctr __attribute__((unused)) = \
+    ::ge::ScopeFusionPassRegistrar(                                                               \
+      pass_name, []() -> ::ge::ScopeBasePass * { return new (std::nothrow) scope_pass(); }, is_general)
+}  // namespace ge
+
+#endif  // EXTERNAL_REGISTER_SCOPE_SCOPE_FUSION_PASS_REGISTER_H_

inc/framework/common/ge_types.h

@@ -22,7 +22,7 @@
 #include <string>
 #include <vector>
 
-#include "common/fmk_error_codes.h"
+#include "framework/common/fmk_error_codes.h"
 #include "ge/ge_api_error_codes.h"
 #include "external/graph/types.h"
 #include "external/ge/ge_api_types.h"
@@ -49,6 +49,7 @@ enum OpEngineType {
 };
 
 const char *const GE_ENGINE_ATTR_MEM_TYPE_HBM = "HBM";
+const char *const GE_OPTION_EXEC_PLACEMENT = "ge.exec.placement";
 
 // Data cache, including data address and length
 struct DataBuffer {
@@ -128,6 +129,7 @@ struct OriginInputInfo {
 
 // The structure of AIPP info
 struct AippConfigInfo {
+  int8_t aipp_mode;
   int8_t input_format;
   int32_t src_image_size_w;
   int32_t src_image_size_h;
@@ -175,6 +177,9 @@ struct AippConfigInfo {
   float var_reci_chn_1;
   float var_reci_chn_2;
   float var_reci_chn_3;
+  int8_t support_rotation;
+  uint32_t related_input_rank;
+  uint32_t max_src_image_size;
 };
 
 // The structure of offline Modeldata
@@ -250,5 +255,31 @@ struct ComputeGraphDescInfo {
   std::vector<std::vector<int64_t>> output_shape;
   std::vector<DataType> output_data_type;
 };
+
+struct OpDescInfo {
+  std::string op_name;
+  uint32_t task_id;
+  uint32_t stream_id;
+  std::vector<Format> input_format;
+  std::vector<std::vector<int64_t>> input_shape;
+  std::vector<DataType> input_data_type;
+  std::vector<void *> input_addrs;
+  std::vector<Format> output_format;
+  std::vector<std::vector<int64_t>> output_shape;
+  std::vector<DataType> output_data_type;
+  std::vector<void *> output_addrs;
+};
+struct ModelDumpConfig {
+  std::string model_name;
+  std::vector<std::string> layers;
+};
+
+struct DumpConfig {
+  std::string dump_path;
+  std::string dump_mode;
+  std::string dump_status;
+  std::string dump_op_switch;
+  std::vector<ModelDumpConfig> dump_list;
+};
 }  // namespace ge
 #endif  // INC_FRAMEWORK_COMMON_GE_TYPES_H_

inc/framework/common/types.h

@@ -606,6 +606,7 @@ static constexpr uint32_t MODEL_FILE_RESERVED_LENGTH = 79;
 /// @brief INPUT node type
 ///
 FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY extern const std::string INPUT_TYPE;
+FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY extern const std::string DUMMY_DATA;
 
 ///
 /// @ingroup domi_omg

inc/framework/common/util.h

@@ -347,6 +347,14 @@ std::string ToString(const google::protobuf::RepeatedField<T> &rpd_field) {
 ///
 uint64_t GetCurrentTimestap();
 
+///
+/// @ingroup domi_common
+/// @brief Obtains the absolute time (timestamp) of the current system.
+/// @return Timestamp, in seconds (US)
+///
+///
+uint32_t GetCurrentSecondTimestap();
+
 ///
 /// @ingroup domi_common
 /// @brief Check whether the product of two int64 numbers exceeds the int64 range.

inc/framework/engine/dnnengine.h

@@ -31,6 +31,7 @@ enum PriorityEnum {
   COST_1,
   COST_2,
   COST_9 = 9,
+  COST_10 = 10,
 };
 
 struct DNNEngineAttribute {

inc/framework/executor/ge_executor.h

@@ -135,6 +135,15 @@ class GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY GeExecutor {
   ///
   ge::Status GetCombinedDynamicDims(uint32_t model_id, std::vector<std::vector<int64_t>> &batch_info);
 
+  ///
+  /// @ingroup ge
+  /// @brief Get user designeate shape order
+  /// @param [in] model_id
+  /// @param [out] user_designate_shape_order
+  /// @return execute result
+  ///
+  ge::Status GetUserDesignateShapeOrder(uint32_t model_id, std::vector<std::string> &user_designate_shape_order);
+
   ge::Status GetCurShape(const uint32_t model_id, std::vector<int64_t> &batch_info, int32_t &dynamic_type);
 
   ///
@@ -162,6 +171,8 @@ class GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY GeExecutor {
 
   ge::Status CommandHandle(const ge::Command &command);
 
+  ge::Status SetDump(const DumpConfig &dump_config);
+
   ///
   /// @ingroup ge
   /// @brief Query model memory consuming interface
@@ -261,6 +272,7 @@ class GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY GeExecutor {
   ge::Status GetOrigInputInfo(uint32_t model_id, uint32_t index, OriginInputInfo &orig_input_info);
   ge::Status GetAllAippInputOutputDims(uint32_t model_id, uint32_t index, std::vector<InputOutputDims> &input_dims,
                                        std::vector<InputOutputDims> &output_dims);
+  ge::Status GetOpDescInfo(uint32_t device_id, uint32_t stream_id, uint32_t task_id, OpDescInfo &op_desc_info);
 
  private:
   static bool isInit_;

inc/framework/memory/memory_api.h

@@ -27,6 +27,7 @@ namespace ge {
 enum MemStorageType {
   HBM = 0,
   RDMA_HBM,
+  HOST_DDR,
 };
 
 struct HostVarInfo {

inc/framework/omg/omg.h

@@ -96,6 +96,10 @@ Status CheckCustomAiCpuOpLib();
 
 Status DumpInfershapeJson(const ge::Graph &graph, const char *json_file);
 
+Status SetOutputNodeInfo(ge::Graph &graph, const std::string &output_type, const std::string &output_format);
+
+Status GetOutputLeaf(ge::NodePtr node, std::vector<std::pair<ge::NodePtr, int32_t>> &output_nodes_info);
+
 void GetOutputNodesNameAndIndex(std::vector<std::pair<ge::NodePtr, int32_t>> &output_nodes_info,
                                 std::vector<std::string> &output_nodes_name);
 

inc/graph/debug/ge_attr_define.h

@@ -883,6 +883,7 @@ GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY extern const std::string REF_VAR_
 
 // Assign
 GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY extern const std::string ASSIGN_VALIDATE_SHAPE;
+GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY extern const std::string ASSIGN_VAR_NAME;
 
 // ShapeN
 GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY extern const std::string SHAPEN_ATTR_N;
@@ -939,6 +940,7 @@ GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY extern const std::string ATTR_NAM
 GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY extern const std::string ATTR_NAME_BATCH_NUM;
 GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY extern const std::string ATTR_NAME_BATCH_LABEL;
 GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY extern const std::string ATTR_NAME_COMBINED_BATCH;
+GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY extern const std::string ATTR_USER_DESIGNEATE_SHAPE_ORDER;
 
 // Control flow
 GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY extern const std::string ATTR_NAME_STREAM_SWITCH_COND;
@@ -957,7 +959,6 @@ GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY extern const std::string ATTR_NAM
 
 // Function Op
 GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY extern const std::string ATTR_NAME_PARENT_NODE_INDEX;
-GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY extern const std::string ATTR_NAME_PARENT_CONST_TYPE;
 
 // Used for mark the active node is for loop, type:bool
 GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY extern const std::string ATTR_NAME_IS_LOOP_ACTIVE;
@@ -968,6 +969,8 @@ GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY extern const std::string ATTR_NAM
 
 GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY extern const std::string ATTR_NAME_MEMORY_TYPE_WORKSPACE;
 
+GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY extern const std::string ATTR_NAME_MEMORY_TYPE_RANGE;
+
 // Atomic addr clean attrs
 GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY extern const std::string ATOMIC_ATTR_INPUT_INDEX;
 GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY extern const std::string ATOMIC_ATTR_OUTPUT_INDEX;

inc/graph/ge_context.h

@@ -24,6 +24,7 @@ namespace ge {
 class GEContext {
  public:
   graphStatus GetOption(const std::string &key, std::string &option);
+  bool GetHostExecFlag();
   uint64_t SessionId();
   uint32_t DeviceId();
   uint64_t TraceId();

inc/graph/op_desc.h

@@ -153,9 +153,6 @@ class OpDesc : public std::enable_shared_from_this<OpDesc>, public AttrHolder {
 
   graphStatus AddDynamicOutputDesc(const string &name, const unsigned int num, bool isPushBack = true);
 
-  void RemoveInputDesc(uint32_t index);
-  void RemoveOutputDesc(uint32_t index);
-
   bool IsOptionalInput(const string &name) const;
 
   bool IsOptionalInput(uint32_t index) const;

inc/graph/utils/node_utils.h

@@ -20,6 +20,7 @@
 #include <set>
 #include <map>
 #include <vector>
+#include "external/graph/operator.h"
 #include "graph/node.h"
 
 namespace ge {
@@ -63,8 +64,11 @@ class NodeUtils {
   static void UnlinkAll(const Node &node);
   static graphStatus UpdatePeerNodeInputDesc(const NodePtr &node_ptr);
 
-  static graphStatus AppendInputAnchor(const NodePtr &node, uint32_t index);
-  static graphStatus RemoveInputAnchor(const NodePtr &node, uint32_t index);
+  static graphStatus AppendInputAnchor(const NodePtr &node, uint32_t num);
+  static graphStatus RemoveInputAnchor(const NodePtr &node, uint32_t num);
+
+  static graphStatus AppendOutputAnchor(const NodePtr &node, uint32_t num);
+  static graphStatus RemoveOutputAnchor(const NodePtr &node, uint32_t num);
 
   static bool IsInNodesEmpty(const Node &node);
   static GeTensorDesc GetOutputDesc(const Node &node, uint32_t index);
@@ -77,6 +81,7 @@ class NodeUtils {
   static graphStatus GetNodeUnknownShapeStatus(const Node &node, bool &is_unknow);
 
   static std::string GetNodeType(const Node &node);
+  static std::string GetNodeType(const NodePtr &node);
 
   static ComputeGraphPtr GetSubgraph(const Node &node, uint32_t index);
   static graphStatus SetSubgraph(Node &node, uint32_t index, const ComputeGraphPtr &subgraph);
@@ -100,8 +105,17 @@ class NodeUtils {
   /// @param [in] node
   /// @return Node
   ///
+  static NodePtr GetParentInput(const Node &node);
   static NodePtr GetParentInput(const NodePtr &node);
 
+  ///
+  /// @brief Get is dynamic shape graph from node.
+  /// @param [in] node
+  /// @return bool
+  ///
+  static bool IsDynamicShape(const Node &node);
+  static bool IsDynamicShape(const NodePtr &node);
+
   ///
   /// @brief Check is varying_input for while node
   /// @param [in] node: Data node for subgraph
@@ -115,7 +129,7 @@ class NodeUtils {
   /// @param [out] string
   /// @return bool
   ///
-  static bool GetConstOpType(const NodePtr &in_node, std::string &op_type);
+  static bool GetConstOpType(const NodePtr &node, std::string &type);
 
   ///
   /// @brief Remove node-related subgraphs, including subgraphs of nodes in the subgraph.
@@ -138,9 +152,15 @@ class NodeUtils {
   ///
   static vector<NodePtr> GetSubgraphOutputNodes(const Node &node);
 
-  static NodePtr GetInDataNodeByIndex(const Node &node, int index);
+  static NodePtr GetInDataNodeByIndex(const Node &node, const int index);
+
+  static vector<pair<InDataAnchorPtr, NodePtr>> GetOutDataNodesWithAnchorByIndex(const Node &node, const int index);
+
+  static ge::ConstNodePtr GetNodeFromOperator(const Operator &oprt);
+
+  static graphStatus GetInputConstData(const ConstNodePtr &node_ptr, const string &dst_name, GeTensorPtr &ge_tensor);
 
-  static vector<NodePtr> GetOutDataNodesByIndex(const Node &node, int index);
+  static graphStatus GetInputConstData(const Node &node, const string &dst_name, GeTensorPtr &ge_tensor);
 
  private:
   static std::map<NodePtr, std::vector<uint32_t>> map_send_info_;

inc/graph/utils/type_utils.h

@@ -34,6 +34,7 @@ class TypeUtils {
   static bool IsFormatValid(Format format);
   static bool IsInternalFormat(Format format);
 
+  static std::string ImplyTypeToSerialString(domi::ImplyType imply_type);
   static std::string DataTypeToSerialString(DataType data_type);
   static DataType SerialStringToDataType(const std::string &str);
   static std::string FormatToSerialString(Format format);

src/common/graph/ge_attr_define.cc

@@ -830,6 +830,7 @@ const std::string REF_VAR_PRE_PEER_OUT_INDEX = "ref_var_pre_peer_out_index";
 
 // Assign
 const std::string ASSIGN_VALIDATE_SHAPE = "validate_shape";
+const std::string ASSIGN_VAR_NAME = "_assign_var_name";
 
 //  space2bacth batch2space
 const std::string BATCH_SPACE_ATTR_BLOCK = "block";
@@ -931,7 +932,6 @@ const std::string ATTR_NAME_NEXT_ITERATION = "_next_iteration_node";
 
 // Function Op
 const std::string ATTR_NAME_PARENT_NODE_INDEX = "_parent_node_index";
-const std::string ATTR_NAME_PARENT_CONST_TYPE = "_parent_const_type";
 
 // Used for mark the active node is for loop, type:bool
 const std::string ATTR_NAME_IS_LOOP_ACTIVE = "is_loop_active";
@@ -942,6 +942,8 @@ const std::string ATTR_NAME_MEMORY_TYPE_OUTPUT = "memory_type_output";
 
 const std::string ATTR_NAME_MEMORY_TYPE_WORKSPACE = "memory_type_workspace";
 
+const std::string ATTR_NAME_MEMORY_TYPE_RANGE = "_memory_type_range";
+
 const std::string MODEL_ATTR_SESSION_ID = "session_id";
 
 // lx fusion
@@ -991,6 +993,8 @@ const std::string ATTR_MBATCH_ORIGIN_INPUT_DIMS = "_mbatch_origin_input_dims";
 
 const std::string ATTR_DYNAMIC_TYPE = "mbatch_dynamic_type";
 
+const std::string ATTR_USER_DESIGNEATE_SHAPE_ORDER = "user_designate_shape_order";
+
 // For inserted op
 const std::string ATTR_INSERTED_BY_GE = "_inserted_by_ge";
 

src/common/graph/op_desc.cc

@@ -684,18 +684,6 @@ graphStatus OpDesc::AddDynamicOutputDesc(const string &name, const unsigned int
   return GRAPH_SUCCESS;
 }
 
-void OpDesc::RemoveInputDesc(uint32_t index) {
-  while (inputs_desc_.size() > index) {
-    inputs_desc_.pop_back();
-  }
-}
-
-void OpDesc::RemoveOutputDesc(uint32_t index) {
-  while (outputs_desc_.size() > index) {
-    outputs_desc_.pop_back();
-  }
-}
-
 bool OpDesc::IsOptionalInput(const string &name) const {
   return optional_input_names_.find(name) != optional_input_names_.end();
 }

src/common/graph/operator.cc

@@ -277,6 +277,22 @@ class OperatorImpl : public std::enable_shared_from_this<OperatorImpl> {
     return output_ptr;
   }
 
+  OutHandler GetOutput(uint32_t index) {
+    GE_CHK_BOOL_EXEC(op_desc_ != nullptr, return nullptr, "op_desc_ is nullptr.");
+
+    string name = op_desc_->GetOutputNameByIndex(index);
+    if (name.empty()) {
+      GELOGE(GRAPH_FAILED, "Find src name by index failed. index[%u]", index);
+      return nullptr;
+    }
+    shared_ptr<OpIO> output_ptr = ComGraphMakeShared<OpIO>(name, index, shared_from_this());
+    if (output_ptr == nullptr) {
+      GELOGE(GRAPH_FAILED, "OpIO make shared failed");
+      return nullptr;
+    }
+    return output_ptr;
+  }
+
   GeTensorDesc GetOutputDesc(const string &name) const {
     GE_CHK_BOOL_EXEC(op_desc_ != nullptr, return GeTensorDesc(), "op_desc_ is nullptr.");
 
@@ -540,6 +556,13 @@ Operator &Operator::SetInput(const std::string &dst_name, const ge::Operator &sr
   return *this;
 }
 
+Operator &Operator::SetInput(const std::string &dst_name, const ge::Operator &src_oprt, uint32_t index) {
+  auto out_handler = src_oprt.GetOutput(index);
+  GE_CHK_BOOL_EXEC(out_handler != nullptr, return *this, "out_handler is nullptr.");
+  (void)SetInput(dst_name, out_handler);
+  return *this;
+}
+
 Operator &Operator::AddControlInput(const Operator &src_oprt) {
   if (operator_impl_ == nullptr) {
     GELOGE(GRAPH_FAILED, "operator impl is nullptr.");
@@ -621,6 +644,11 @@ graphStatus Operator::GetInputConstDataOut(const string &dst_name, Tensor &data)
   return GRAPH_FAILED;
 }
 
+std::shared_ptr<const Node> Operator::GetNode() const {
+  GE_CHK_BOOL_EXEC(operator_impl_ != nullptr, return nullptr, "operator impl is nullptr.");
+  return operator_impl_->GetNode();
+}
+
 TensorDesc Operator::GetInputDesc(const std::string &name) const {
   GE_CHK_BOOL_EXEC(operator_impl_ != nullptr, return TensorDesc(), "operator impl is nullptr.");
   return TensorAdapter::GeTensorDesc2TensorDesc(operator_impl_->GetInputDesc(name));
@@ -657,6 +685,11 @@ OutHandler Operator::GetOutput(const string &name) const {
   return operator_impl_->GetOutput(name);
 }
 
+OutHandler Operator::GetOutput(uint32_t index) const {
+  GE_CHK_BOOL_EXEC(operator_impl_ != nullptr, return nullptr, "operator impl is nullptr.");
+  return operator_impl_->GetOutput(index);
+}
+
 TensorDesc Operator::GetOutputDesc(const std::string &name) const {
   GE_CHK_BOOL_EXEC(operator_impl_ != nullptr, return TensorDesc(), "operator impl is nullptr.");
   return TensorAdapter::GeTensorDesc2TensorDesc(operator_impl_->GetOutputDesc(name));
@@ -1540,6 +1573,7 @@ void GraphUtils::BreakConnect(const std::map<OperatorImplPtr, NodePtr> &all_node
     }
     op_impl->ClearOutputLinks();
     op_impl->ClearInputLinks();
+    OperatorKeeper::GetInstance().CheckOutOperator(op_impl);
   }
 }
 }  // namespace ge

src/common/graph/option/ge_context.cc

@@ -17,12 +17,14 @@
 #include "./ge_context.h"
 #include "./ge_global_options.h"
 #include "./ge_local_context.h"
+#include "framework/common/ge_types.h"
 #include "framework/common/debug/ge_log.h"
 
 namespace ge {
 namespace {
 const int64_t kMinTrainingTraceJobId = 256;
 const int kDecimal = 10;
+const char *kHostExecPlacement = "HOST";
 }  // namespace
 GEContext &GetContext() {
   static GEContext ge_context{};
@@ -33,6 +35,16 @@ graphStatus GEContext::GetOption(const std::string &key, std::string &option) {
   return GetThreadLocalContext().GetOption(key, option);
 }
 
+bool GEContext::GetHostExecFlag() {
+  std::string exec_placement;
+  if (GetThreadLocalContext().GetOption(GE_OPTION_EXEC_PLACEMENT, exec_placement) != GRAPH_SUCCESS) {
+    GELOGW("get option OPTION_EXEC_PLACEMENT failed.");
+    return false;
+  }
+  GELOGD("Option ge.exec.placement is %s.", exec_placement.c_str());
+  return exec_placement == kHostExecPlacement;
+}
+
 std::map<std::string, std::string> &GetMutableGlobalOptions() {
   static std::map<std::string, std::string> global_options{};
   return global_options;

src/common/graph/ref_relation.cc

@@ -243,8 +243,8 @@ graphStatus RefRelations::Impl::BuildRefRelationsForWhile(
     }
     auto in_data_anchor_idx = in_anchor->GetIdx();
     auto net_in_desc = netoutput->GetOpDesc()->MutableInputDesc(static_cast<uint32_t>(in_data_anchor_idx));
-    int ref_d;
-    int ref_n;
+    int ref_d = 0;
+    int ref_n = 0;
     (void)AttrUtils::GetInt(peer_out_data_node->GetOpDesc(), kRefIndex, ref_d);
     (void)AttrUtils::GetInt(net_in_desc, kRefIndex, ref_n);
 

src/common/graph/shape_refiner.cc

@@ -351,6 +351,66 @@ graphStatus UpdateParentNodeOutTensor(const ConstNodePtr &node) {
   }
   return UpdateParentNodeForBranch(node, ref_out_tensors);
 }
+
+string Serial(const vector<int64_t> &dims) {
+  string serial_string;
+  serial_string += "[";
+  for (int64_t dim : dims) {
+    serial_string += std::to_string(dim) + " ";
+  }
+  serial_string += "]";
+  return serial_string;
+}
+
+graphStatus UpdateOpInputDesc(const ConstNodePtr &node_ptr) {
+  GE_IF_BOOL_EXEC(node_ptr == nullptr, GELOGE(GRAPH_FAILED, "node is null."); return GRAPH_FAILED);
+  GE_IF_BOOL_EXEC(node_ptr->GetOpDesc() == nullptr, GELOGE(GRAPH_FAILED, "op_desc is null."); return GRAPH_FAILED);
+  for (const auto &in_anchor : node_ptr->GetAllInDataAnchors()) {
+    auto in_idx = in_anchor->GetIdx();
+    auto peer_out_data_anchor = in_anchor->GetPeerOutAnchor();
+    if (peer_out_data_anchor == nullptr) {
+      continue;
+    }
+    auto peer_out_data_node = peer_out_data_anchor->GetOwnerNode();
+    if (peer_out_data_node == nullptr || peer_out_data_node->GetOpDesc() == nullptr) {
+      continue;
+    }
+    int peer_out_idx = peer_out_data_anchor->GetIdx();
+    auto in_desc = node_ptr->GetOpDesc()->MutableInputDesc(static_cast<uint32_t>(in_idx));
+    auto peer_out_desc = peer_out_data_node->GetOpDesc()->MutableOutputDesc(static_cast<uint32_t>(peer_out_idx));
+
+    // check shape and dtype continuity. do not stop process
+    auto in_shape = in_desc->GetShape().GetDims();
+    auto in_dtype = in_desc->GetDataType();
+    auto peer_out_shape = peer_out_desc->GetShape().GetDims();
+    auto peer_out_dtype = peer_out_desc->GetDataType();
+    if (peer_out_dtype != in_dtype) {
+      GELOGW(
+        "current node [%s] [%d]\'th out_dtype is [%s].peer output node [%s] [%d]\'th "
+        "output_dtype is [%s].The two dtype should be same! Please check graph and fix it",
+        node_ptr->GetName().c_str(), in_idx, TypeUtils::DataTypeToSerialString(in_dtype).c_str(),
+        peer_out_data_node->GetName().c_str(), peer_out_idx, TypeUtils::DataTypeToSerialString(peer_out_dtype).c_str());
+    } else if ((!in_shape.empty()) && (in_shape != peer_out_shape)) {
+      string in_shape_str = Serial(in_shape);
+      string peer_out_shape_str = Serial(peer_out_shape);
+      GELOGW(
+        "current node [%s] [%d]\'th out_shape is [%s].peer input node [%s] [%d]\'th "
+        "input_shape is [%s].The two shape should be same! Please check graph and fix it",
+        node_ptr->GetName().c_str(), in_idx, in_shape_str.c_str(), peer_out_data_node->GetName().c_str(), peer_out_idx,
+        peer_out_shape_str.c_str());
+    }
+    // refresh current node input desc
+    in_desc->SetOriginShape(peer_out_desc->GetOriginShape());
+    in_desc->SetShape(peer_out_desc->GetShape());
+    in_desc->SetDataType(peer_out_desc->GetDataType());
+    in_desc->SetOriginDataType(peer_out_desc->GetOriginDataType());
+    std::vector<std::pair<int64_t, int64_t>> shape_range;
+    (void)peer_out_desc->GetShapeRange(shape_range);
+    in_desc->SetShapeRange(shape_range);
+    ge::TensorUtils::SetRealDimCnt(*in_desc, static_cast<uint32_t>(peer_out_desc->GetShape().GetDims().size()));
+  }
+  return GRAPH_SUCCESS;
+}
 }  // namespace
 void ShapeRefiner::PrintInOutTensorShape(const ge::NodePtr &node, const std::string &phase) {
   if (!IsLogEnable(GE, DLOG_DEBUG)) {
@@ -427,9 +487,7 @@ graphStatus ShapeRefiner::InferShapeAndType(const ConstNodePtr &node, Operator &
   return InferShapeAndType(node, op, true);
 }
 graphStatus ShapeRefiner::InferShapeAndType(const ConstNodePtr &node, Operator &op, bool before_subgraph) {
-  GE_IF_BOOL_EXEC(node == nullptr, GELOGE(GRAPH_FAILED, "node is null."); return GRAPH_FAILED);
   auto op_desc = node->GetOpDesc();
-  GE_IF_BOOL_EXEC(op_desc == nullptr, GELOGE(GRAPH_FAILED, "op_desc is null."); return GRAPH_FAILED);
   const auto &op_type = op_desc->GetType();
 
   graphStatus ret;
@@ -554,6 +612,19 @@ GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY graphStatus ShapeRefiner::InferSh
 GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY graphStatus ShapeRefiner::InferShapeAndType(const NodePtr &node,
                                                                                            bool before_subgraph) {
   GE_IF_BOOL_EXEC(node == nullptr, GELOGE(GRAPH_FAILED, "node is null."); return GRAPH_FAILED);
+  bool is_unknown_graph = node->GetOwnerComputeGraph()->GetGraphUnknownFlag();
+  auto opdesc = node->GetOpDesc();
+  GE_IF_BOOL_EXEC(opdesc == nullptr, GELOGE(GRAPH_FAILED, "op_desc is null."); return GRAPH_FAILED);
+  // some op can not infershape twice such as aipp
+  bool need_update_input = !is_unknown_graph && !opdesc->HasAttr("has_infered_verified");
+  if (need_update_input) {
+    auto status = UpdateOpInputDesc(node);
+    if (status != GRAPH_SUCCESS) {
+      GELOGE(GRAPH_FAILED, "update op input_desc failed!");
+      return status;
+    }
+  }
+
   if (node->Verify() != GRAPH_SUCCESS) {
     GELOGE(GRAPH_FAILED, "Verifying %s failed.", node->GetName().c_str());
     return GRAPH_FAILED;
@@ -561,7 +632,6 @@ GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY graphStatus ShapeRefiner::InferSh
   PrintInOutTensorShape(node, "before_infershape");
   Operator op = OpDescUtils::CreateOperatorFromNode(node);
 
-  bool is_unknown_graph = node->GetOwnerComputeGraph()->GetGraphUnknownFlag();
   if (!is_unknown_graph) {
     auto inference_context = CreateInferenceContext(context_map, node);
     if (inference_context == nullptr) {
@@ -574,7 +644,21 @@ GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY graphStatus ShapeRefiner::InferSh
 
   graphStatus status = InferShapeAndType(node, op, before_subgraph);
   if (status == GRAPH_PARAM_INVALID || status == GRAPH_SUCCESS) {
-    (void)ge::NodeUtils::UpdatePeerNodeInputDesc(node);
+    if (is_unknown_graph) {
+      return GRAPH_SUCCESS;
+    }
+    auto op_desc = node->GetOpDesc();
+    for (const auto &out_anchor : node->GetAllOutDataAnchors()) {
+      auto output_tensor = op_desc->MutableOutputDesc(out_anchor->GetIdx());
+      ge::TensorUtils::SetRealDimCnt(*output_tensor, static_cast<uint32_t>(output_tensor->GetShape().GetDims().size()));
+      output_tensor->SetOriginShape(output_tensor->GetShape());
+      output_tensor->SetOriginDataType(output_tensor->GetDataType());
+
+      GELOGD("node name is %s, origin shape is %ld, origin format is %s, origin data type is %s",
+             node->GetName().c_str(), output_tensor->GetOriginShape().GetShapeSize(),
+             TypeUtils::FormatToSerialString(output_tensor->GetOriginFormat()).c_str(),
+             TypeUtils::DataTypeToSerialString(output_tensor->GetOriginDataType()).c_str());
+    }
   } else {
     GELOGE(GRAPH_FAILED, "%s call infer function failed.", node->GetName().c_str());
     return GRAPH_FAILED;

src/common/graph/utils/node_utils.cc

@@ -15,6 +15,7 @@
  */
 
 #include "utils/node_utils.h"
+#include "utils/op_desc_utils.h"
 #include "graph/utils/graph_utils.h"
 #include "debug/ge_op_types.h"
 #include "debug/ge_util.h"
@@ -301,6 +302,8 @@ GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY graphStatus NodeUtils::UpdatePeer
   }
   for (const auto &out_anchor : node_ptr->GetAllOutDataAnchors()) {
     auto output_tensor = op_desc->MutableOutputDesc(out_anchor->GetIdx());
+    auto out_dims = output_tensor->GetShape().GetDims();
+    auto out_dtype = output_tensor->GetDataType();
     ge::TensorUtils::SetRealDimCnt(*output_tensor, static_cast<uint32_t>(output_tensor->GetShape().GetDims().size()));
     output_tensor->SetOriginShape(output_tensor->GetShape());
     output_tensor->SetOriginDataType(output_tensor->GetDataType());
@@ -320,6 +323,35 @@ GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY graphStatus NodeUtils::UpdatePeer
         GELOGE(GRAPH_FAILED, "peer_input_desc is nullptr");
         continue;
       }
+      // check shape and dtype continuity. do not stop process
+      auto peer_input_dims = peer_input_desc->GetShape().GetDims();
+      auto peer_input_dtype = peer_input_desc->GetDataType();
+      if (out_dtype != peer_input_dtype) {
+        GELOGW(
+          "current node [%s] [%d]\'th out_dtype is [%s].peer input node [%s] [%d]\'th "
+          "input_dtype is [%s].The two dtype should be same! Please check graph and fix it",
+          node_ptr->GetName().c_str(), out_anchor->GetIdx(), TypeUtils::DataTypeToSerialString(out_dtype).c_str(),
+          peer_anchor->GetOwnerNode()->GetName().c_str(), peer_anchor->GetIdx(),
+          TypeUtils::DataTypeToSerialString(peer_input_dtype).c_str());
+      } else if ((!peer_input_dims.empty()) && (out_dims != peer_input_dims)) {
+        string out_shape_str, peer_in_shape_str;
+        out_shape_str += "[";
+        for (int64_t dim : out_dims) {
+          out_shape_str += std::to_string(dim) + " ";
+        }
+        out_shape_str += "]";
+        peer_in_shape_str += "[";
+        for (int64_t dim : peer_input_dims) {
+          peer_in_shape_str += std::to_string(dim) + " ";
+        }
+        peer_in_shape_str += "]";
+
+        GELOGW(
+          "current node [%s] [%d]\'th out_shape is [%s].peer input node [%s] [%d]\'th "
+          "input_shape is [%s].The two shape should be same! Please check graph and fix it",
+          node_ptr->GetName().c_str(), out_anchor->GetIdx(), out_shape_str.c_str(),
+          peer_anchor->GetOwnerNode()->GetName().c_str(), peer_anchor->GetIdx(), peer_in_shape_str.c_str());
+      }
       GELOGI("Peer input opdesc name is %s, need to flush: shape size is %zu, datatype is %d, original datatype is %d",
              peer_anchor->GetOwnerNode()->GetOpDesc()->GetName().c_str(), output_tensor->GetShape().GetDimNum(),
              output_tensor->GetDataType(), output_tensor->GetOriginDataType());
@@ -341,15 +373,15 @@ GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY graphStatus NodeUtils::UpdatePeer
 }
 
 GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY graphStatus NodeUtils::AppendInputAnchor(const NodePtr &node,
-                                                                                        uint32_t index) {
+                                                                                        uint32_t num) {
   if (node == nullptr) {
-    GELOGE(GRAPH_FAILED, "Nodeptr is nullptr");
+    GELOGE(GRAPH_FAILED, "Input node is null");
     return GRAPH_FAILED;
   }
 
   GeTensorDesc data_desc(GeShape(), FORMAT_ND, DT_FLOAT);
-  OpDescPtr op_desc = node->op_;
-  for (size_t i = op_desc->GetInputsSize(); i < index; ++i) {
+  const auto &op_desc = node->GetOpDesc();
+  for (size_t i = op_desc->GetInputsSize(); i < num; ++i) {
     if (op_desc->AddInputDesc(data_desc) != GRAPH_SUCCESS) {
       GELOGE(GRAPH_FAILED, "Add input desc failed");
       return GRAPH_FAILED;
@@ -357,7 +389,7 @@ GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY graphStatus NodeUtils::AppendInpu
 
     auto anchor = ComGraphMakeShared<InDataAnchor>(node, i);
     if (anchor == nullptr) {
-      GELOGE(GRAPH_FAILED, "Current in_data_anchor is null, malloc shared_ptr failed.");
+      GELOGE(OUT_OF_MEMORY, "Current in data anchor is null, make shared_ptr failed.");
       return GRAPH_FAILED;
     }
     node->in_data_anchors_.push_back(anchor);
@@ -367,22 +399,81 @@ GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY graphStatus NodeUtils::AppendInpu
 }
 
 GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY graphStatus NodeUtils::RemoveInputAnchor(const NodePtr &node,
-                                                                                        uint32_t index) {
+                                                                                        uint32_t num) {
   if (node == nullptr) {
-    GELOGE(GRAPH_FAILED, "Nodeptr is nullptr");
+    GELOGE(GRAPH_FAILED, "Input node is null");
     return GRAPH_FAILED;
   }
 
-  OpDescPtr op_desc = node->op_;
-  op_desc->RemoveInputDesc(index);
+  const auto &op_desc = node->GetOpDesc();
+  while (op_desc->GetInputsSize() > num) {
+    if (!OpDescUtils::ClearInputDesc(op_desc, num)) {
+      return GRAPH_FAILED;
+    }
+  }
 
-  while (node->in_data_anchors_.size() > index) {
+  auto input_names = op_desc->GetAllInputName();
+  (void)op_desc->UpdateInputName(input_names);
+  auto is_input_const = op_desc->GetIsInputConst();
+  is_input_const.resize(num);
+  op_desc->SetIsInputConst(is_input_const);
+
+  while (node->in_data_anchors_.size() > num) {
     node->in_data_anchors_.pop_back();
   }
 
   return GRAPH_SUCCESS;
 }
 
+GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY graphStatus NodeUtils::AppendOutputAnchor(const NodePtr &node,
+                                                                                         uint32_t num) {
+  if (node == nullptr) {
+    GELOGE(GRAPH_FAILED, "Input node is null");
+    return GRAPH_FAILED;
+  }
+
+  GeTensorDesc data_desc(GeShape(), FORMAT_ND, DT_FLOAT);
+  const OpDescPtr &op_desc = node->GetOpDesc();
+  for (size_t i = op_desc->GetOutputsSize(); i < num; ++i) {
+    if (op_desc->AddOutputDesc(data_desc) != GRAPH_SUCCESS) {
+      GELOGE(GRAPH_FAILED, "Add output desc failed");
+      return GRAPH_FAILED;
+    }
+
+    auto anchor = ComGraphMakeShared<OutDataAnchor>(node, i);
+    if (anchor == nullptr) {
+      GELOGE(OUT_OF_MEMORY, "Current out data anchor is null, make shared_ptr failed.");
+      return GRAPH_FAILED;
+    }
+    node->out_data_anchors_.push_back(anchor);
+  }
+
+  return GRAPH_SUCCESS;
+}
+
+GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY graphStatus NodeUtils::RemoveOutputAnchor(const NodePtr &node,
+                                                                                         uint32_t num) {
+  if (node == nullptr) {
+    GELOGE(GRAPH_FAILED, "Input node is null");
+    return GRAPH_FAILED;
+  }
+
+  const auto &op_desc = node->GetOpDesc();
+  auto output_names = op_desc->GetAllOutputName();
+  while (op_desc->GetOutputsSize() > num) {
+    if (!OpDescUtils::ClearOutputDesc(op_desc, num)) {
+      return GRAPH_FAILED;
+    }
+  }
+  (void)op_desc->UpdateOutputName(output_names);
+
+  while (node->out_data_anchors_.size() > num) {
+    node->out_data_anchors_.pop_back();
+  }
+
+  return GRAPH_SUCCESS;
+}
+
 bool NodeUtils::IsInNodesEmpty(const Node &node) {
   for (const auto &in_anchor : node.in_data_anchors_) {
     if (in_anchor != nullptr) {
@@ -488,11 +579,22 @@ std::string NodeUtils::GetNodeType(const Node &node) {
   if (node.GetType() != FRAMEWORKOP) {
     return node.GetType();
   }
+
   std::string type;
   (void)AttrUtils::GetStr(node.GetOpDesc(), ATTR_NAME_FRAMEWORK_ORIGINAL_TYPE, type);
   return type;
 }
 
+std::string NodeUtils::GetNodeType(const NodePtr &node) { return node == nullptr ? "" : GetNodeType(*node); }
+
+graphStatus NodeUtils::GetInputConstData(const ConstNodePtr &node_ptr, const string &dst_name, GeTensorPtr &ge_tensor) {
+  return GRAPH_SUCCESS;
+}
+
+graphStatus NodeUtils::GetInputConstData(const Node &node, const string &dst_name, GeTensorPtr &ge_tensor) {
+  return GRAPH_SUCCESS;
+}
+
 ComputeGraphPtr NodeUtils::GetSubgraph(const Node &node, uint32_t index) {
   auto op_desc = node.GetOpDesc();
   if (op_desc == nullptr) {
@@ -544,16 +646,17 @@ bool NodeUtils::IsSubgraphInput(const NodePtr &node) {
   if (parent_op_desc == nullptr) {
     return false;
   }
-  if (AttrUtils::HasAttr(parent_op_desc, ATTR_NAME_IS_UNKNOWN_SHAPE)) {
-    bool is_unknown_shape = false;
-    (void)AttrUtils::GetBool(parent_op_desc, ATTR_NAME_IS_UNKNOWN_SHAPE, is_unknown_shape);
-    if (is_unknown_shape) return false;
-  }
 
-  if (AttrUtils::HasAttr(parent_op_desc, ATTR_NAME_IS_UNKNOWN_SHAPE) &&
-      kCaseOpTypes.count(parent_op_desc->GetType()) == 0 && kWhileOpTypes.count(parent_op_desc->GetType()) == 0 &&
-      kForOpTypes.count(parent_op_desc->GetType()) == 0 && kIfOpTypes.count(parent_op_desc->GetType()) == 0) {
-    return false;
+  // dynamic shape unknown graph false
+  // dynamic shape known graph with functional subgraph maybe true
+  if (AttrUtils::HasAttr(parent_op_desc, ATTR_NAME_IS_UNKNOWN_SHAPE)) {
+    if (node->GetOwnerComputeGraph()->GetParentGraph()->GetGraphUnknownFlag()) {
+      return false;
+    } else {
+      if (node->GetOwnerComputeGraph()->GetParentNode()->GetOwnerComputeGraph()->GetParentNode() == nullptr) {
+        return false;
+      }
+    }
   }
 
   return node->GetOpDesc()->HasAttr(ATTR_NAME_PARENT_NODE_INDEX);
@@ -576,15 +679,13 @@ bool NodeUtils::IsSubgraphOutput(const NodePtr &node) {
   }
 
   if (AttrUtils::HasAttr(parent_op_desc, ATTR_NAME_IS_UNKNOWN_SHAPE)) {
-    bool is_unknown_shape = false;
-    (void)AttrUtils::GetBool(parent_op_desc, ATTR_NAME_IS_UNKNOWN_SHAPE, is_unknown_shape);
-    if (is_unknown_shape) return false;
-  }
-
-  if (AttrUtils::HasAttr(parent_op_desc, ATTR_NAME_IS_UNKNOWN_SHAPE) &&
-      kCaseOpTypes.count(parent_op_desc->GetType()) == 0 && kWhileOpTypes.count(parent_op_desc->GetType()) == 0 &&
-      kForOpTypes.count(parent_op_desc->GetType()) == 0 && kIfOpTypes.count(parent_op_desc->GetType()) == 0) {
-    return false;
+    if (node->GetOwnerComputeGraph()->GetParentGraph()->GetGraphUnknownFlag()) {
+      return false;
+    } else {
+      if (node->GetOwnerComputeGraph()->GetParentNode()->GetOwnerComputeGraph()->GetParentNode() == nullptr) {
+        return false;
+      }
+    }
   }
 
   for (GeTensorDesc &tensor : node->GetOpDesc()->GetAllInputsDesc()) {
@@ -601,16 +702,14 @@ bool NodeUtils::IsSubgraphOutput(const NodePtr &node) {
 /// @param [in] node
 /// @return Node
 ///
-NodePtr NodeUtils::GetParentInput(const NodePtr &node) {
-  GE_CHECK_NOTNULL_EXEC(node, return nullptr);
-
+NodePtr NodeUtils::GetParentInput(const Node &node) {
   uint32_t parent_index = 0;
-  if (!AttrUtils::GetInt(node->GetOpDesc(), ATTR_NAME_PARENT_NODE_INDEX, parent_index)) {
+  if (!AttrUtils::GetInt(node.GetOpDesc(), ATTR_NAME_PARENT_NODE_INDEX, parent_index)) {
     return nullptr;
   }
 
   // Subgraph Data Node, check for constant input.
-  const ComputeGraphPtr &graph = node->GetOwnerComputeGraph();
+  const ComputeGraphPtr &graph = node.GetOwnerComputeGraph();
   GE_CHECK_NOTNULL_EXEC(graph, return nullptr);
 
   const NodePtr &parent_node = graph->GetParentNode();
@@ -625,6 +724,26 @@ NodePtr NodeUtils::GetParentInput(const NodePtr &node) {
   return peer_out_anchor->GetOwnerNode();
 }
 
+NodePtr NodeUtils::GetParentInput(const NodePtr &node) { return node == nullptr ? node : GetParentInput(*node); }
+
+///
+/// @brief Get is dynamic shape graph from node.
+/// @param [in] node
+/// @return bool
+///
+bool NodeUtils::IsDynamicShape(const Node &node) {
+  const auto graph = GraphUtils::FindRootGraph(node.GetOwnerComputeGraph());
+  if (graph == nullptr) {
+    return false;
+  }
+
+  bool is_dynamic_shape = false;
+  (void)AttrUtils::GetBool(graph, ATTR_NAME_DYNAMIC_SHAPE_PARTITIONED, is_dynamic_shape);
+  return is_dynamic_shape;
+}
+
+bool NodeUtils::IsDynamicShape(const NodePtr &node) { return node == nullptr ? false : IsDynamicShape(*node); }
+
 ///
 /// @brief Check is varying_input for while node
 /// @param [in] node: Data node for subgraph
@@ -678,27 +797,22 @@ bool NodeUtils::IsWhileVaryingInput(const ge::NodePtr &node) {
 /// @param [out] string
 /// @return bool
 ///
-bool NodeUtils::GetConstOpType(const NodePtr &in_node, std::string &op_type) {
-  GE_CHECK_NOTNULL_EXEC(in_node, return false);
+bool NodeUtils::GetConstOpType(const NodePtr &node, std::string &type) {
+  if (node == nullptr) {
+    return false;
+  }
 
-  if ((in_node->GetType() == CONSTANT) || (in_node->GetType() == CONSTANTOP)) {
-    op_type = in_node->GetType();
+  if ((node->GetType() == CONSTANT) || (node->GetType() == CONSTANTOP)) {
+    type = node->GetType();
     return true;
   }
 
-  if (in_node->GetType() == DATA) {
-    std::string const_type;
-    if (!AttrUtils::GetStr(in_node->GetOpDesc(), ATTR_NAME_PARENT_CONST_TYPE, const_type)) {
-      return false;
-    }
-
-    if ((const_type == CONSTANT) || (const_type == CONSTANTOP)) {
-      op_type = const_type;
-      return true;
-    }
+  if (node->GetType() != DATA) {
+    return false;  // not subgraph input node
   }
 
-  return false;
+  const auto &parent = GetParentInput(node);
+  return GetConstOpType(parent, type);
 }
 
 ///
@@ -809,7 +923,7 @@ vector<NodePtr> NodeUtils::GetSubgraphOutputNodes(const Node &node) {
   return out_data_node_vec;
 }
 
-NodePtr NodeUtils::GetInDataNodeByIndex(const Node &node, int index) {
+NodePtr NodeUtils::GetInDataNodeByIndex(const Node &node, const int index) {
   if (node.GetInDataAnchor(index) == nullptr) {
     return nullptr;
   }
@@ -819,12 +933,13 @@ NodePtr NodeUtils::GetInDataNodeByIndex(const Node &node, int index) {
   return node.GetInDataAnchor(index)->GetPeerOutAnchor()->GetOwnerNode();
 }
 
-vector<NodePtr> NodeUtils::GetOutDataNodesByIndex(const Node &node, int index) {
-  vector<NodePtr> out_data_nodes;
+vector<pair<InDataAnchorPtr, NodePtr>> NodeUtils::GetOutDataNodesWithAnchorByIndex(const Node &node, const int index) {
+  vector<pair<InDataAnchorPtr, NodePtr>> out_data_nodes;
   auto out_data_anchor = node.GetOutDataAnchor(index);
   if (out_data_anchor == nullptr) {
     return out_data_nodes;
   }
+
   for (const auto peer_in_anchor : out_data_anchor->GetPeerInDataAnchors()) {
     if (peer_in_anchor == nullptr) {
       continue;
@@ -832,8 +947,10 @@ vector<NodePtr> NodeUtils::GetOutDataNodesByIndex(const Node &node, int index) {
     if (peer_in_anchor->GetOwnerNode() == nullptr) {
       continue;
     }
-    out_data_nodes.emplace_back(peer_in_anchor->GetOwnerNode());
+    out_data_nodes.emplace_back(std::make_pair(peer_in_anchor, peer_in_anchor->GetOwnerNode()));
   }
   return out_data_nodes;
 }
+
+ConstNodePtr NodeUtils::GetNodeFromOperator(const Operator &oprt) { return oprt.GetNode(); }
 }  // namespace ge

src/common/graph/utils/op_desc_utils.cc

@@ -438,6 +438,11 @@ GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY vector<ge::NodePtr> OpDescUtils::
       if (switch_input.size() > 0) {
         ret.insert(ret.end(), switch_input.begin(), switch_input.end());
       }
+    } else if (in_node->GetType() == DATA) {
+      auto parent = NodeUtils::GetParentInput(in_node);
+      if ((parent != nullptr) && (parent->GetType() == CONSTANT)) {
+        ret.push_back(parent);
+      }
     }
   }
   return ret;

src/common/graph/utils/type_utils.cc

@@ -244,6 +244,21 @@ static const std::map<domi::FrameworkType, std::string> kFmkTypeToString = {
   {domi::ANDROID_NN, "android_nn"}, {domi::ONNX, "onnx"},           {domi::FRAMEWORK_RESERVED, "framework_reserved"},
 };
 
+static const std::map<domi::ImplyType, std::string> kImplyTypeToString = {
+  {domi::ImplyType::BUILDIN, "buildin"}, {domi::ImplyType::TVM, "tvm"},        {domi::ImplyType::CUSTOM, "custom"},
+  {domi::ImplyType::AI_CPU, "ai_cpu"},   {domi::ImplyType::CCE, "cce"},        {domi::ImplyType::GELOCAL, "gelocal"},
+  {domi::ImplyType::HCCL, "hccl"},       {domi::ImplyType::INVALID, "invalid"}};
+
+std::string TypeUtils::ImplyTypeToSerialString(domi::ImplyType imply_type) {
+  auto it = kImplyTypeToString.find(imply_type);
+  if (it != kImplyTypeToString.end()) {
+    return it->second;
+  } else {
+    GELOGE(GRAPH_FAILED, "ImplyTypeToSerialString: imply_type not support %u", imply_type);
+    return "UNDEFINED";
+  }
+}
+
 bool TypeUtils::IsDataTypeValid(DataType dt) {
   uint32_t num = static_cast<uint32_t>(dt);
   GE_CHK_BOOL_EXEC((num <= DT_UNDEFINED), return false, "The DataType is invalid");

src/ge/CMakeLists.txt

@@ -56,6 +56,9 @@ include_directories(${CMAKE_BINARY_DIR}/proto/ge)
 # need to remove dependencies on pb files later
 file(GLOB TRAIN_SRC_LIST RELATIVE ${CMAKE_CURRENT_LIST_DIR}
         "client/ge_api.cc"
+        "common/dump/dump_manager.cc"
+        "common/dump/dump_properties.cc"
+        "common/dump/dump_op.cc"
         "common/formats/format_transfers/*.cc"
         "common/formats/formats.cc"
         "common/formats/utils/formats_trans_utils.cc"
@@ -124,6 +127,7 @@ file(GLOB TRAIN_SRC_LIST RELATIVE ${CMAKE_CURRENT_LIST_DIR}
         "graph/preprocess/insert_op/ge_aipp_op.cc"
         "graph/preprocess/insert_op/util_insert_aipp_op.cc"
         "graph/preprocess/multi_batch_copy_graph.cc"
+        "graph/preprocess/multi_batch_options.cc"
         "host_kernels/add_kernel.cc"
         "host_kernels/broadcast_args_kernel.cc"
         "host_kernels/broadcast_gradient_args_kernel.cc"
@@ -138,6 +142,7 @@ file(GLOB TRAIN_SRC_LIST RELATIVE ${CMAKE_CURRENT_LIST_DIR}
         "host_kernels/floormod_kernel.cc"
         "host_kernels/gather_v2_kernel.cc"
         "host_kernels/greater_kernel.cc"
+        "host_kernels/identity_kernel.cc"
         "host_kernels/kernel_utils.cc"
         "host_kernels/maximum_kernel.cc"
         "host_kernels/mul_kernel.cc"
@@ -172,10 +177,18 @@ file(GLOB TRAIN_SRC_LIST RELATIVE ${CMAKE_CURRENT_LIST_DIR}
         "hybrid/node_executor/aicpu/aicpu_node_executor.cc"
         "hybrid/node_executor/compiledsubgraph/known_node_executor.cc"
         "hybrid/node_executor/controlop/control_op_executor.cc"
+        "hybrid/node_executor/ge_local/ge_local_node_executor.cc"
         "hybrid/node_executor/hccl/hccl_node_executor.cc"
         "hybrid/node_executor/hostcpu/ge_local_node_executor.cc"
+        "hybrid/node_executor/host_cpu/host_cpu_node_executor.cc"
+        "hybrid/node_executor/host_cpu/kernel_factory.cc"
+        "hybrid/node_executor/host_cpu/kernel/no_op_kernel.cc"
+        "hybrid/node_executor/host_cpu/kernel/variable_kernel.cc"
+        "hybrid/node_executor/host_cpu/kernel/assign_kernel.cc"
+        "hybrid/node_executor/host_cpu/kernel/random_uniform_kernel.cc"
         "hybrid/node_executor/node_executor.cc"
         "hybrid/node_executor/partitioned_call/partitioned_call_node_executor.cc"
+        "hybrid/node_executor/rts/rts_node_executor.cc"
         "hybrid/node_executor/task_context.cc"
         "init/gelib.cc"
         "model/ge_model.cc"
@@ -215,6 +228,9 @@ target_link_libraries(ge_runner
 ######### libge_compiler.so #############
 # need to remove dependencies on pb files later
 file(GLOB INFER_SRC_LIST RELATIVE ${CMAKE_CURRENT_LIST_DIR}
+        "common/dump/dump_properties.cc"
+        "common/dump/dump_manager.cc"
+        "common/dump/dump_op.cc"
         "common/formats/format_transfers/*.cc"
         "common/formats/formats.cc"
         "common/formats/utils/formats_trans_utils.cc"
@@ -274,6 +290,7 @@ file(GLOB INFER_SRC_LIST RELATIVE ${CMAKE_CURRENT_LIST_DIR}
         "graph/preprocess/insert_op/ge_aipp_op.cc"
         "graph/preprocess/insert_op/util_insert_aipp_op.cc"
         "graph/preprocess/multi_batch_copy_graph.cc"
+        "graph/preprocess/multi_batch_options.cc"
         "host_kernels/add_kernel.cc"
         "host_kernels/broadcast_args_kernel.cc"
         "host_kernels/broadcast_gradient_args_kernel.cc"
@@ -288,6 +305,7 @@ file(GLOB INFER_SRC_LIST RELATIVE ${CMAKE_CURRENT_LIST_DIR}
         "host_kernels/floormod_kernel.cc"
         "host_kernels/gather_v2_kernel.cc"
         "host_kernels/greater_kernel.cc"
+        "host_kernels/identity_kernel.cc"
         "host_kernels/kernel_utils.cc"
         "host_kernels/maximum_kernel.cc"
         "host_kernels/mul_kernel.cc"

src/ge/client/ge_api.cc

@@ -390,6 +390,22 @@ Status Session::RunGraphAsync(uint32_t graph_id, const std::vector<InputTensorIn
   }
   return SUCCESS;
 }
+
+Status Session::GetVariables(const std::vector<std::string> &var_names, std::vector<Tensor> &var_values) {
+  auto instance_ptr = ge::GELib::GetInstance();
+  if (instance_ptr == nullptr || !instance_ptr->InitFlag()) {
+    GELOGE(GE_CLI_GE_NOT_INITIALIZED, "SessionConstructor failed");
+    return FAILED;
+  }
+  GELOGT(TRACE_RUNNING, "Get Variables");
+  Status ret = ge::GELib::GetInstance()->SessionManagerObj().GetVariables(sessionId_, var_names, var_values);
+  if (ret != SUCCESS) {
+    GELOGE(ret, "SessionManager RunGraphAsync failed");
+    return FAILED;
+  }
+  return SUCCESS;
+}
+
 bool Session::IsGraphNeedRebuild(uint32_t graph_id) {
   return ge::GELib::GetInstance()->SessionManagerObj().IsGraphNeedRebuild(sessionId_, graph_id);
 }

src/ge/common/dump/dump_manager.cc

@@ -0,0 +1,120 @@
+/**
+ * Copyright 2019-2020 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 "common/dump/dump_manager.h"
+#include "framework/common/debug/ge_log.h"
+#include "framework/common/debug/log.h"
+
+namespace {
+const char *const kDumpOFF = "OFF";
+const char *const kDumpoff = "off";
+const char *const kDumpOn = "on";
+}  // namespace
+namespace ge {
+FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY DumpManager &DumpManager::GetInstance() {
+  static DumpManager instance;
+  return instance;
+}
+
+FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY Status DumpManager::SetDumpConf(const DumpConfig &dump_config) {
+  std::lock_guard<std::mutex> lock(mutex_);
+  dump_properties_.ClearDumpPropertyValue();
+  dump_properties_.ClearDumpInfo();
+  std::string dump_status;
+  std::string dump_path;
+  std::string dump_mode;
+  std::string dump_op_switch;
+
+  if (dump_config.dump_status.empty()) {
+    GELOGI("Dump does not open");
+    return SUCCESS;
+  }
+
+  dump_status = dump_config.dump_status;
+  GELOGI("Dump status is %s", dump_status.c_str());
+  if (dump_config.dump_status == kDumpoff || dump_config.dump_status == kDumpOFF) {
+    dump_properties_.ClearDumpPropertyValue();
+    return SUCCESS;
+  }
+  dump_op_switch = dump_config.dump_op_switch;
+  if (dump_op_switch == kDumpoff && dump_config.dump_list.empty()) {
+    GELOGE(PARAM_INVALID, "Dump list is invalid,dump_op_switch is %s", dump_op_switch.c_str());
+    return PARAM_INVALID;
+  }
+
+  if (!dump_config.dump_list.empty()) {
+    for (auto model_dump : dump_config.dump_list) {
+      std::string model_name = model_dump.model_name;
+      GELOGI("Dump model is %s", model_name.c_str());
+      std::set<std::string> dump_layers;
+      for (auto layer : model_dump.layers) {
+        GELOGI("Dump layer is %s in model", layer.c_str());
+        dump_layers.insert(layer);
+      }
+      dump_properties_.AddPropertyValue(model_name, dump_layers);
+    }
+    if (dump_op_switch == kDumpOn) {
+      GELOGI("Start to dump model and single op,dumo op switch is %s", dump_op_switch.c_str());
+    } else {
+      GELOGI("Only dump model,dump op switch is %s", dump_op_switch.c_str());
+    }
+  } else {
+    GELOGI("Only dump single op,dumo op switch is %s", dump_op_switch.c_str());
+  }
+
+  dump_path = dump_config.dump_path;
+  if (dump_path.empty()) {
+    GELOGE(PARAM_INVALID, "Dump path is empty");
+    return PARAM_INVALID;
+  }
+
+  if (dump_path[dump_path.size() - 1] != '/') {
+    dump_path = dump_path + "/";
+  }
+  dump_path = dump_path + CurrentTimeInStr() + "/";
+  GELOGI("Dump path is %s", dump_path.c_str());
+  dump_properties_.SetDumpPath(dump_path);
+
+  dump_mode = dump_config.dump_mode;
+  GELOGI("Dump mode is %s", dump_mode.c_str());
+  dump_properties_.SetDumpMode(dump_mode);
+
+  return SUCCESS;
+}
+
+FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY bool DumpManager::IsDumpOpen() {
+  std::lock_guard<std::mutex> lock(mutex_);
+  if (!dump_properties_.GetDumpPath().empty()) {
+    return true;
+  }
+  return false;
+}
+
+FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY const DumpProperties &DumpManager::GetDumpProperties() {
+  std::lock_guard<std::mutex> lock(mutex_);
+  return dump_properties_;
+}
+
+FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY void DumpManager::SetModelName(const std::string &model_name) {
+  std::lock_guard<std::mutex> lock(mutex_);
+  model_name_ = model_name;
+}
+
+FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY const std::string &DumpManager::GetModelName() {
+  std::lock_guard<std::mutex> lock(mutex_);
+  return model_name_;
+}
+}  // namespace ge

src/ge/common/dump/dump_manager.h

@@ -0,0 +1,42 @@
+/**
+ * Copyright 2019-2020 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_COMMON_DUMP_DUMP_MANAGER_H_
+#define GE_COMMON_DUMP_DUMP_MANAGER_H_
+
+#include <mutex>
+
+#include "common/dump/dump_properties.h"
+#include "common/ge_types.h"
+
+namespace ge {
+class DumpManager {
+ public:
+  static DumpManager &GetInstance();
+
+  Status SetDumpConf(const DumpConfig &dump_config);
+  bool IsDumpOpen();
+  const DumpProperties &GetDumpProperties();
+  void SetModelName(const std::string &model_name);
+  const std::string &GetModelName();
+
+ private:
+  DumpProperties dump_properties_;
+  std::mutex mutex_;
+  std::string model_name_;
+};
+}  // namespace ge
+#endif  // GE_COMMON_DUMP_DUMP_MANAGER_H_

src/ge/common/dump/dump_op.cc

@@ -0,0 +1,255 @@
+/**
+ * Copyright 2019-2020 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 "common/dump/dump_op.h"
+
+#include "aicpu/common/aicpu_task_struct.h"
+#include "common/dump/dump_manager.h"
+#include "common/ge/datatype_util.h"
+#include "framework/common/debug/ge_log.h"
+#include "framework/common/util.h"
+#include "graph/anchor.h"
+#include "graph/ge_tensor.h"
+#include "graph/op_desc.h"
+#include "graph/utils/tensor_utils.h"
+#include "proto/ge_ir.pb.h"
+#include "proto/op_mapping_info.pb.h"
+#include "runtime/mem.h"
+
+namespace {
+const uint32_t kAicpuLoadFlag = 1;
+const char *const kDumpOutput = "output";
+const char *const kDumpInput = "input";
+const char *const kDumpAll = "all";
+const char *const kDumpKernelsDumpOp = "DumpDataInfo";
+}  // namespace
+
+namespace ge {
+DumpOp::~DumpOp() {
+  if (proto_dev_mem_ != nullptr) {
+    (void)rtFree(proto_dev_mem_);
+  }
+  if (proto_size_dev_mem_ != nullptr) {
+    (void)rtFree(proto_size_dev_mem_);
+  }
+  proto_dev_mem_ = nullptr;
+  proto_size_dev_mem_ = nullptr;
+}
+
+void DumpOp::SetLoopAddr(void *global_step, void *loop_per_iter, void *loop_cond) {
+  global_step_ = reinterpret_cast<uintptr_t>(global_step);
+  loop_per_iter_ = reinterpret_cast<uintptr_t>(loop_per_iter);
+  loop_cond_ = reinterpret_cast<uintptr_t>(loop_cond);
+}
+
+void DumpOp::SetDynamicModelInfo(const string &dynamic_model_name, uint32_t dynamic_model_id) {
+  dynamic_model_name_ = dynamic_model_name;
+  dynamic_model_id_ = dynamic_model_id;
+}
+
+static void SetOpMappingLoopAddr(uintptr_t step_id, uintptr_t loop_per_iter, uintptr_t loop_cond,
+                                 aicpu::dump::OpMappingInfo &op_mapping_info) {
+  if (step_id != 0) {
+    GELOGI("step_id exists.");
+    op_mapping_info.set_step_id_addr(static_cast<uint64_t>(step_id));
+  } else {
+    GELOGI("step_id is null.");
+  }
+
+  if (loop_per_iter != 0) {
+    GELOGI("loop_per_iter exists.");
+    op_mapping_info.set_iterations_per_loop_addr(static_cast<uint64_t>(loop_per_iter));
+  } else {
+    GELOGI("loop_per_iter is null.");
+  }
+
+  if (loop_cond != 0) {
+    GELOGI("loop_cond exists.");
+    op_mapping_info.set_loop_cond_addr(static_cast<uint64_t>(loop_cond));
+  } else {
+    GELOGI("loop_cond is null.");
+  }
+}
+
+Status DumpOp::DumpOutput(aicpu::dump::Task &task) {
+  GELOGI("Start dump output in Launch dump op");
+  const auto &output_descs = op_desc_->GetAllOutputsDesc();
+  for (size_t i = 0; i < output_descs.size(); ++i) {
+    aicpu::dump::Output output;
+    output.set_data_type(static_cast<int32_t>(DataTypeUtil::GetIrDataType(output_descs.at(i).GetDataType())));
+    output.set_format(static_cast<int32_t>(output_descs.at(i).GetFormat()));
+    for (auto dim : output_descs.at(i).GetShape().GetDims()) {
+      output.mutable_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");
+      return PARAM_INVALID;
+    }
+    GELOGD("Get output size in lanch dump op is %ld", output_size);
+    output.set_size(output_size);
+    output.set_address(static_cast<uint64_t>(output_addrs_[i]));
+    task.mutable_output()->Add(std::move(output));
+  }
+  return SUCCESS;
+}
+
+Status DumpOp::DumpInput(aicpu::dump::Task &task) {
+  GELOGI("Start dump input in Launch dump op");
+  const auto &input_descs = op_desc_->GetAllInputsDesc();
+  for (size_t i = 0; i < input_descs.size(); ++i) {
+    aicpu::dump::Input input;
+    input.set_data_type(static_cast<int32_t>(DataTypeUtil::GetIrDataType(input_descs.at(i).GetDataType())));
+    input.set_format(static_cast<int32_t>(input_descs.at(i).GetFormat()));
+
+    for (auto dim : input_descs.at(i).GetShape().GetDims()) {
+      input.mutable_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");
+      return PARAM_INVALID;
+    }
+    GELOGD("Get input size in lanch dump op is %ld", input_size);
+    input.set_size(input_size);
+    input.set_address(static_cast<uint64_t>(input_addrs_[i]));
+    task.mutable_input()->Add(std::move(input));
+  }
+  return SUCCESS;
+}
+
+void DumpOp::SetDumpInfo(const DumpProperties &dump_properties, const OpDescPtr &op_desc, vector<uintptr_t> input_addrs,
+                         vector<uintptr_t> output_addrs, rtStream_t stream) {
+  dump_properties_ = dump_properties;
+  op_desc_ = op_desc;
+  input_addrs_ = input_addrs;
+  output_addrs_ = output_addrs;
+  stream_ = stream;
+}
+
+Status DumpOp::ExecutorDumpOp(aicpu::dump::OpMappingInfo &op_mapping_info) {
+  std::string proto_msg;
+  size_t proto_size = op_mapping_info.ByteSizeLong();
+  bool ret = op_mapping_info.SerializeToString(&proto_msg);
+  if (!ret || proto_size == 0) {
+    GELOGE(FAILED, "Protobuf serialize failed,proto_size is %zu", proto_size);
+    return FAILED;
+  }
+
+  rtError_t rt_ret = rtMalloc(&proto_dev_mem_, proto_size, RT_MEMORY_HBM);
+  if (rt_ret != RT_ERROR_NONE) {
+    GELOGE(RT_FAILED, "Call rtMalloc failed, ret: 0x%X", rt_ret);
+    return RT_FAILED;
+  }
+
+  rt_ret = rtMemcpy(proto_dev_mem_, proto_size, proto_msg.c_str(), proto_size, RT_MEMCPY_HOST_TO_DEVICE);
+  if (rt_ret != RT_ERROR_NONE) {
+    GELOGE(RT_FAILED, "Call rtMemcpy failed, ret: 0x%X", rt_ret);
+    return RT_FAILED;
+  }
+
+  rt_ret = rtMalloc(&proto_size_dev_mem_, sizeof(size_t), RT_MEMORY_HBM);
+  if (rt_ret != RT_ERROR_NONE) {
+    GELOGE(RT_FAILED, "Call rtMalloc failed, ret: 0x%X", rt_ret);
+    return RT_FAILED;
+  }
+  rt_ret = rtMemcpy(proto_size_dev_mem_, sizeof(size_t), &proto_size, sizeof(size_t), RT_MEMCPY_HOST_TO_DEVICE);
+  if (rt_ret != RT_ERROR_NONE) {
+    GELOGE(RT_FAILED, "Call rtMemcpy failed, ret: 0x%X", rt_ret);
+    return RT_FAILED;
+  }
+
+  constexpr int32_t ioAddrNum = 2;
+  constexpr uint32_t argsSize = sizeof(aicpu::AicpuParamHead) + ioAddrNum * sizeof(uint64_t);
+  char args[argsSize] = {0};
+  auto paramHead = reinterpret_cast<aicpu::AicpuParamHead *>(args);
+  paramHead->length = argsSize;
+  paramHead->ioAddrNum = ioAddrNum;
+  auto ioAddr = reinterpret_cast<uint64_t *>(args + sizeof(aicpu::AicpuParamHead));
+  ioAddr[0] = reinterpret_cast<uintptr_t>(proto_dev_mem_);
+  ioAddr[1] = reinterpret_cast<uintptr_t>(proto_size_dev_mem_);
+  rt_ret = rtCpuKernelLaunch(nullptr, kDumpKernelsDumpOp,
+                             1,  // blockDim default 1
+                             args, argsSize,
+                             nullptr,  // no need smDesc
+                             stream_);
+  if (rt_ret != RT_ERROR_NONE) {
+    GELOGE(RT_FAILED, "Call rtCpuKernelLaunch failed,rt_ret:0x%X", rt_ret);
+    return rt_ret;
+  }
+  GELOGI("Kernel launch dump op success");
+  return SUCCESS;
+}
+
+Status DumpOp::LaunchDumpOp() {
+  GELOGI("Start to launch dump op %s", op_desc_->GetName().c_str());
+  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);
+    return RT_FAILED;
+  }
+  aicpu::dump::OpMappingInfo op_mapping_info;
+  auto dump_path = dump_properties_.GetDumpPath() + std::to_string(device_id) + "/";
+  op_mapping_info.set_dump_path(dump_path);
+  op_mapping_info.set_flag(kAicpuLoadFlag);
+  op_mapping_info.set_dump_step(dump_properties_.GetDumpStep());
+  if (!dynamic_model_name_.empty()) {
+    op_mapping_info.set_model_name(dynamic_model_name_);
+    op_mapping_info.set_model_id(dynamic_model_id_);
+  }
+  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());
+
+  aicpu::dump::Task task;
+  task.mutable_op()->set_op_name(op_desc_->GetName());
+  task.mutable_op()->set_op_type(op_desc_->GetType());
+  if (dump_properties_.GetDumpMode() == kDumpOutput) {
+    if (DumpOutput(task) != SUCCESS) {
+      GELOGE(FAILED, "Dump output failed");
+      return FAILED;
+    }
+    op_mapping_info.mutable_task()->Add(std::move(task));
+  }
+  if (dump_properties_.GetDumpMode() == kDumpInput) {
+    if (DumpInput(task) != SUCCESS) {
+      GELOGE(FAILED, "Dump input failed");
+      return FAILED;
+    }
+    op_mapping_info.mutable_task()->Add(std::move(task));
+  }
+  if (dump_properties_.GetDumpMode() == kDumpAll) {
+    auto ret = DumpOutput(task);
+    if (ret != SUCCESS) {
+      GELOGE(FAILED, "Dump output failed when in dumping all");
+      return FAILED;
+    }
+    ret = DumpInput(task);
+    if (ret != SUCCESS) {
+      GELOGE(FAILED, "Dump input failed when in dumping all");
+      return FAILED;
+    }
+    op_mapping_info.mutable_task()->Add(std::move(task));
+  }
+  auto ret = ExecutorDumpOp(op_mapping_info);
+  if (ret != SUCCESS) {
+    GELOGE(ret, "Executor dump op failed");
+    return ret;
+  }
+  return SUCCESS;
+}
+}  // namespace ge

src/ge/common/dump/dump_op.h

@@ -0,0 +1,61 @@
+/**
+ * Copyright 2019-2020 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_COMMON_DUMP_DUMP_OP_H_
+#define GE_COMMON_DUMP_DUMP_OP_H_
+
+#include <string>
+
+#include "common/ge_inner_error_codes.h"
+#include "common/properties_manager.h"
+#include "proto/op_mapping_info.pb.h"
+#include "runtime/stream.h"
+
+namespace ge {
+class DumpOp {
+ public:
+  DumpOp() = default;
+  ~DumpOp();
+
+  void SetDumpInfo(const DumpProperties &dump_properties, const OpDescPtr &op_desc, vector<uintptr_t> input_addrs,
+                   vector<uintptr_t> output_addrs, rtStream_t stream);
+  Status LaunchDumpOp();
+  void SetLoopAddr(void *global_step, void *loop_per_iter, void *loop_cond);
+  void SetDynamicModelInfo(const string &dynamic_model_name, uint32_t dynamic_model_id);
+
+ private:
+  Status ExecutorDumpOp(aicpu::dump::OpMappingInfo &op_mapping_info);
+  Status DumpOutput(aicpu::dump::Task &task);
+  Status DumpInput(aicpu::dump::Task &task);
+
+  DumpProperties dump_properties_;
+  OpDescPtr op_desc_;
+  std::vector<uintptr_t> input_addrs_;
+  std::vector<uintptr_t> output_addrs_;
+
+  void *proto_dev_mem_ = nullptr;
+  void *proto_size_dev_mem_ = nullptr;
+  rtStream_t stream_;
+  uintptr_t global_step_;
+  uintptr_t loop_per_iter_;
+  uintptr_t loop_cond_;
+
+  std::string dynamic_model_name_;
+  std::uint32_t dynamic_model_id_;
+};
+}  // namespace ge
+
+#endif  // GE_COMMON_DUMP_DUMP_OP_H_

src/ge/common/dump/dump_properties.cc

@@ -0,0 +1,238 @@
+/**
+ * Copyright 2019-2020 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 "common/dump/dump_properties.h"
+
+#include <cstdio>
+#include <string>
+
+#include "common/ge/ge_util.h"
+#include "common/util.h"
+#include "framework/common/debug/ge_log.h"
+#include "framework/common/debug/log.h"
+#include "framework/common/ge_types.h"
+#include "framework/common/types.h"
+#include "graph/debug/ge_attr_define.h"
+#include "graph/ge_context.h"
+#include "graph/utils/attr_utils.h"
+
+namespace {
+const std::string kEnableFlag = "1";
+
+const uint32_t kAicoreOverflow = (0x1 << 0);
+const uint32_t kAtomicOverflow = (0x1 << 1);
+const uint32_t kAllOverflow = (kAicoreOverflow | kAtomicOverflow);
+}  // namespace
+namespace ge {
+FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY DumpProperties::DumpProperties(const DumpProperties &other) {
+  CopyFrom(other);
+}
+
+FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY DumpProperties &DumpProperties::operator=(
+  const DumpProperties &other) {
+  CopyFrom(other);
+  return *this;
+}
+
+FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY void DumpProperties::InitByOptions() {
+  enable_dump_.clear();
+  enable_dump_debug_.clear();
+  dump_path_.clear();
+  dump_step_.clear();
+  dump_mode_.clear();
+  is_op_debug_ = false;
+  op_debug_mode_ = 0;
+
+  std::string enable_dump;
+  (void)GetContext().GetOption(OPTION_EXEC_ENABLE_DUMP, enable_dump);
+  enable_dump_ = enable_dump;
+
+  std::string enable_dump_debug;
+  (void)GetContext().GetOption(OPTION_EXEC_ENABLE_DUMP_DEBUG, enable_dump_debug);
+  enable_dump_debug_ = enable_dump_debug;
+
+  if ((enable_dump_ == kEnableFlag) || (enable_dump_debug_ == kEnableFlag)) {
+    std::string dump_path;
+    if (GetContext().GetOption(OPTION_EXEC_DUMP_PATH, dump_path) == GRAPH_SUCCESS) {
+      if (!dump_path.empty() && dump_path[dump_path.size() - 1] != '/') {
+        dump_path = dump_path + "/";
+      }
+      dump_path = dump_path + CurrentTimeInStr() + "/";
+      GELOGI("Get dump path %s successfully", dump_path.c_str());
+      SetDumpPath(dump_path);
+    } else {
+      GELOGW("Dump path is not set");
+    }
+  }
+
+  if (enable_dump_ == kEnableFlag) {
+    std::string dump_step;
+    if (GetContext().GetOption(OPTION_EXEC_DUMP_STEP, dump_step) == GRAPH_SUCCESS) {
+      GELOGD("Get dump step %s successfully", dump_step.c_str());
+      SetDumpStep(dump_step);
+    }
+    string dump_mode;
+    if (GetContext().GetOption(OPTION_EXEC_DUMP_MODE, dump_mode) == GRAPH_SUCCESS) {
+      GELOGD("Get dump mode %s successfully", dump_mode.c_str());
+      SetDumpMode(dump_mode);
+    }
+    AddPropertyValue(DUMP_ALL_MODEL, {});
+  }
+
+  SetDumpDebugOptions();
+}
+
+// The following is the new dump scenario of the fusion operator
+FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY void DumpProperties::AddPropertyValue(
+  const std::string &model, const std::set<std::string> &layers) {
+  for (const std::string &layer : layers) {
+    GELOGI("This model %s config to dump layer %s", model.c_str(), layer.c_str());
+  }
+
+  model_dump_properties_map_[model] = layers;
+}
+
+FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY void DumpProperties::DeletePropertyValue(const std::string &model) {
+  auto iter = model_dump_properties_map_.find(model);
+  if (iter != model_dump_properties_map_.end()) {
+    model_dump_properties_map_.erase(iter);
+  }
+}
+
+FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY void DumpProperties::ClearDumpPropertyValue() {
+  model_dump_properties_map_.clear();
+}
+
+FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY void DumpProperties::ClearDumpInfo() {
+  enable_dump_.clear();
+  enable_dump_debug_.clear();
+  dump_path_.clear();
+  dump_step_.clear();
+  dump_mode_.clear();
+  is_op_debug_ = false;
+  op_debug_mode_ = 0;
+}
+
+FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY std::set<std::string> DumpProperties::GetAllDumpModel() const {
+  std::set<std::string> model_list;
+  for (auto &iter : model_dump_properties_map_) {
+    model_list.insert(iter.first);
+  }
+
+  return model_list;
+}
+
+FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY std::set<std::string> DumpProperties::GetPropertyValue(
+  const std::string &model) const {
+  auto iter = model_dump_properties_map_.find(model);
+  if (iter != model_dump_properties_map_.end()) {
+    return iter->second;
+  }
+  return {};
+}
+
+FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY bool DumpProperties::IsLayerNeedDump(
+  const std::string &model, const std::string &om_name, const std::string &op_name) const {
+  // if dump all
+  if (model_dump_properties_map_.find(DUMP_ALL_MODEL) != model_dump_properties_map_.end()) {
+    return true;
+  }
+
+  // if this model need dump
+  auto om_name_iter = model_dump_properties_map_.find(om_name);
+  auto model_name_iter = model_dump_properties_map_.find(model);
+  if (om_name_iter != model_dump_properties_map_.end() || model_name_iter != model_dump_properties_map_.end()) {
+    // if no dump layer info, dump all layer in this model
+    auto model_iter = om_name_iter != model_dump_properties_map_.end() ? om_name_iter : model_name_iter;
+    if (model_iter->second.empty()) {
+      return true;
+    }
+
+    return model_iter->second.find(op_name) != model_iter->second.end();
+  }
+
+  GELOGD("Model %s is not seated to be dump.", model.c_str());
+  return false;
+}
+
+FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY void DumpProperties::SetDumpPath(const std::string &path) {
+  dump_path_ = path;
+}
+
+FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY const std::string &DumpProperties::GetDumpPath() const {
+  return dump_path_;
+}
+
+FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY void DumpProperties::SetDumpStep(const std::string &step) {
+  dump_step_ = step;
+}
+
+FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY const std::string &DumpProperties::GetDumpStep() const {
+  return dump_step_;
+}
+
+FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY void DumpProperties::SetDumpMode(const std::string &mode) {
+  dump_mode_ = mode;
+}
+
+FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY const std::string &DumpProperties::GetDumpMode() const {
+  return dump_mode_;
+}
+
+void DumpProperties::CopyFrom(const DumpProperties &other) {
+  if (&other != this) {
+    enable_dump_ = other.enable_dump_;
+    enable_dump_debug_ = other.enable_dump_debug_;
+    dump_path_ = other.dump_path_;
+    dump_step_ = other.dump_step_;
+    dump_mode_ = other.dump_mode_;
+
+    model_dump_properties_map_ = other.model_dump_properties_map_;
+    is_op_debug_ = other.is_op_debug_;
+    op_debug_mode_ = other.op_debug_mode_;
+  }
+}
+
+void DumpProperties::SetDumpDebugOptions() {
+  if (enable_dump_debug_ == kEnableFlag) {
+    std::string dump_debug_mode;
+    if (GetContext().GetOption(OPTION_EXEC_DUMP_DEBUG_MODE, dump_debug_mode) == GRAPH_SUCCESS) {
+      GELOGD("Get dump debug mode %s successfully", dump_debug_mode.c_str());
+    } else {
+      GELOGW("Dump debug mode is not set.");
+      return;
+    }
+
+    if (dump_debug_mode == OP_DEBUG_AICORE) {
+      GELOGD("ge.exec.dumpDebugMode=aicore_overflow, op debug is open.");
+      is_op_debug_ = true;
+      op_debug_mode_ = kAicoreOverflow;
+    } else if (dump_debug_mode == OP_DEBUG_ATOMIC) {
+      GELOGD("ge.exec.dumpDebugMode=atomic_overflow, op debug is open.");
+      is_op_debug_ = true;
+      op_debug_mode_ = kAtomicOverflow;
+    } else if (dump_debug_mode == OP_DEBUG_ALL) {
+      GELOGD("ge.exec.dumpDebugMode=all, op debug is open.");
+      is_op_debug_ = true;
+      op_debug_mode_ = kAllOverflow;
+    } else {
+      GELOGW("ge.exec.dumpDebugMode is invalid.");
+    }
+  } else {
+    GELOGI("ge.exec.enableDumpDebug is false or is not set.");
+  }
+}
+}  // namespace ge

src/ge/common/dump/dump_properties.h

@@ -0,0 +1,86 @@
+/**
+ * Copyright 2019-2020 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_COMMON_DUMP_DUMP_PROPERTIES_H_
+#define GE_COMMON_DUMP_DUMP_PROPERTIES_H_
+
+#include <map>
+#include <set>
+#include <string>
+#include <vector>
+
+namespace ge {
+class DumpProperties {
+ public:
+  DumpProperties() = default;
+
+  ~DumpProperties() = default;
+
+  DumpProperties(const DumpProperties &dump);
+
+  DumpProperties &operator=(const DumpProperties &dump);
+
+  void InitByOptions();
+
+  void AddPropertyValue(const std::string &model, const std::set<std::string> &layers);
+
+  void DeletePropertyValue(const std::string &model);
+
+  void ClearDumpPropertyValue();
+
+  void ClearDumpInfo();
+
+  std::set<std::string> GetAllDumpModel() const;
+
+  std::set<std::string> GetPropertyValue(const std::string &model) const;
+
+  bool IsLayerNeedDump(const std::string &model, const std::string &om_name, const std::string &op_name) const;
+
+  void SetDumpPath(const std::string &path);
+
+  const std::string &GetDumpPath() const;
+
+  void SetDumpStep(const std::string &step);
+
+  const std::string &GetDumpStep() const;
+
+  void SetDumpMode(const std::string &mode);
+
+  const std::string &GetDumpMode() const;
+
+  bool IsOpDebugOpen() const { return is_op_debug_; }
+
+  uint32_t GetOpDebugMode() const { return op_debug_mode_; }
+
+ private:
+  void CopyFrom(const DumpProperties &other);
+
+  void SetDumpDebugOptions();
+
+  std::string enable_dump_;
+  std::string enable_dump_debug_;
+
+  std::string dump_path_;
+  std::string dump_step_;
+  std::string dump_mode_;
+  std::map<std::string, std::set<std::string>> model_dump_properties_map_;
+
+  bool is_op_debug_ = false;
+  uint32_t op_debug_mode_ = 0;
+};
+}  // namespace ge
+
+#endif  // GE_COMMON_DUMP_DUMP_PROPERTIES_H_

src/ge/common/ge/datatype_util.cc

@@ -15,23 +15,54 @@
  */
 
 #include "common/ge/datatype_util.h"
+#include "proto/ge_ir.pb.h"
 
 #include <map>
 
 namespace {
 const std::vector<ge::DataType> kEmptyDatatypeVector;
 std::map<ge::DataType, std::vector<ge::DataType>> g_translatable_data_type = {
-    // key:src datatype, value:dst datatype
-    {ge::DT_FLOAT, {ge::DT_FLOAT16, ge::DT_FLOAT}},
-    {ge::DT_BOOL, {ge::DT_INT32}},
-    {ge::DT_FLOAT16, {ge::DT_FLOAT, ge::DT_FLOAT16}},
-    {ge::DT_INT64, {ge::DT_INT32}}};
+  // key:src datatype, value:dst datatype
+  {ge::DT_FLOAT, {ge::DT_FLOAT16, ge::DT_FLOAT}},
+  {ge::DT_BOOL, {ge::DT_INT32}},
+  {ge::DT_FLOAT16, {ge::DT_FLOAT, ge::DT_FLOAT16}},
+  {ge::DT_INT64, {ge::DT_INT32}}};
 
 std::map<ge::DataType, std::vector<ge::DataType>> g_reverse_translatable_data_type = {
-    // key:dst datatype,value:src datatype
-    {ge::DT_FLOAT16, {ge::DT_FLOAT, ge::DT_FLOAT16}},
-    {ge::DT_INT32, {ge::DT_BOOL, ge::DT_INT64}},
-    {ge::DT_FLOAT, {ge::DT_FLOAT16, ge::DT_FLOAT}}};
+  // key:dst datatype,value:src datatype
+  {ge::DT_FLOAT16, {ge::DT_FLOAT, ge::DT_FLOAT16}},
+  {ge::DT_INT32, {ge::DT_BOOL, ge::DT_INT64}},
+  {ge::DT_FLOAT, {ge::DT_FLOAT16, ge::DT_FLOAT}}};
+
+static const std::map<ge::DataType, ge::proto::DataType> g_dump_data_type_map = {
+  // key:ge datatype,value:proto datatype
+  {ge::DT_UNDEFINED, ge::proto::DT_UNDEFINED},
+  {ge::DT_FLOAT, ge::proto::DT_FLOAT},
+  {ge::DT_FLOAT16, ge::proto::DT_FLOAT16},
+  {ge::DT_INT8, ge::proto::DT_INT8},
+  {ge::DT_UINT8, ge::proto::DT_UINT8},
+  {ge::DT_INT16, ge::proto::DT_INT16},
+  {ge::DT_UINT16, ge::proto::DT_UINT16},
+  {ge::DT_INT32, ge::proto::DT_INT32},
+  {ge::DT_INT64, ge::proto::DT_INT64},
+  {ge::DT_UINT32, ge::proto::DT_UINT32},
+  {ge::DT_UINT64, ge::proto::DT_UINT64},
+  {ge::DT_BOOL, ge::proto::DT_BOOL},
+  {ge::DT_DOUBLE, ge::proto::DT_DOUBLE},
+  {ge::DT_DUAL, ge::proto::DT_DUAL},
+  {ge::DT_DUAL_SUB_INT8, ge::proto::DT_DUAL_SUB_INT8},
+  {ge::DT_DUAL_SUB_UINT8, ge::proto::DT_DUAL_SUB_UINT8},
+  {ge::DT_COMPLEX64, ge::proto::DT_COMPLEX64},
+  {ge::DT_COMPLEX128, ge::proto::DT_COMPLEX128},
+  {ge::DT_QINT8, ge::proto::DT_QINT8},
+  {ge::DT_QINT16, ge::proto::DT_QINT16},
+  {ge::DT_QINT32, ge::proto::DT_QINT32},
+  {ge::DT_QUINT8, ge::proto::DT_QUINT8},
+  {ge::DT_QUINT16, ge::proto::DT_QUINT16},
+  {ge::DT_RESOURCE, ge::proto::DT_RESOURCE},
+  {ge::DT_STRING_REF, ge::proto::DT_STRING_REF},
+  {ge::DT_STRING, ge::proto::DT_STRING},
+};
 }  // namespace
 
 namespace ge {
@@ -67,4 +98,13 @@ const std::vector<ge::DataType> &DataTypeUtil::GetTranslatableDataTypesByDst(con
 
   return search->second;
 }
+
+int32_t DataTypeUtil::GetIrDataType(ge::DataType data_type) {
+  auto iter = g_dump_data_type_map.find(data_type);
+  if (iter == g_dump_data_type_map.end()) {
+    return static_cast<int32_t>(ge::proto::DT_UNDEFINED);
+  }
+
+  return static_cast<int32_t>(iter->second);
+}
 }  // namespace ge

src/ge/common/ge/datatype_util.h

@@ -37,16 +37,17 @@ static const int32_t kGeSizeUint16 = sizeof(uint16_t);
 static const int32_t kGeSizeUint32 = sizeof(uint32_t);
 
 static std::map<ge::DataType, int32_t> CONST_OPDATA_TYPE_SIZE_MAP = {
-    {ge::DT_FLOAT, kGeSizeFloat},   {ge::DT_FLOAT16, kGeSizeHalfFloat}, {ge::DT_INT8, kGeSizeInt8},
-    {ge::DT_INT16, kGeSizeInt16},   {ge::DT_INT32, kGeSizeInt32},        {ge::DT_INT64, kGeSizeInt64},
-    {ge::DT_UINT8, kGeSizeUint8},   {ge::DT_UINT16, kGeSizeUint16},      {ge::DT_UINT32, kGeSizeUint32},
-    {ge::DT_UINT64, kGeSizeUint64}, {ge::DT_DOUBLE, kGeSizeDouble},      {ge::DT_BOOL, kGeSizeBool}};
+  {ge::DT_FLOAT, kGeSizeFloat},   {ge::DT_FLOAT16, kGeSizeHalfFloat}, {ge::DT_INT8, kGeSizeInt8},
+  {ge::DT_INT16, kGeSizeInt16},   {ge::DT_INT32, kGeSizeInt32},       {ge::DT_INT64, kGeSizeInt64},
+  {ge::DT_UINT8, kGeSizeUint8},   {ge::DT_UINT16, kGeSizeUint16},     {ge::DT_UINT32, kGeSizeUint32},
+  {ge::DT_UINT64, kGeSizeUint64}, {ge::DT_DOUBLE, kGeSizeDouble},     {ge::DT_BOOL, kGeSizeBool}};
 
 class GE_FUNC_HOST_VISIBILITY GE_FUNC_DEV_VISIBILITY DataTypeUtil {
  public:
   static bool DataTypeTranslatable(const ge::DataType &src_out_data_type, const ge::DataType &dst_in_data_type);
   static const std::vector<ge::DataType> &GetTranslatableDataTypesBySrc(const ge::DataType &src_out_data_type);
   static const std::vector<ge::DataType> &GetTranslatableDataTypesByDst(const ge::DataType &dst_in_data_type);
+  static int32_t GetIrDataType(ge::DataType data_type);
 };
 }  // namespace ge
 #endif  // GE_COMMON_GE_DATATYPE_UTIL_H_

src/ge/common/ge/tbe_plugin_manager.cc

@@ -187,8 +187,8 @@ void TBEPluginManager::LoadCustomOpLib() {
   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: %u", reg_data.GetOmOptype().c_str(),
-           static_cast<uint32_t>(reg_data.GetImplyType()));
+    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);
   }
 }

src/ge/common/ge_common.mk

@@ -36,7 +36,6 @@ GE_COMMON_LOCAL_SRC_FILES := \
     properties_manager.cc \
     types.cc\
     model_parser/base.cc \
-    model_parser/graph_parser_util.cc \
     tbe_kernel_store.cc \
     op/attr_value_util.cc \
     op/ge_op_utils.cc \

src/ge/common/math/math_util.h

@@ -562,7 +562,6 @@ inline Status CheckUint64MulOverflow(uint64_t a, uint64_t b) {
 /// @return Status
 inline Status CheckFp16MulOverflow(fp16_t a, fp16_t b) {
   fp16_t result = static_cast<fp16_t>(a) * static_cast<fp16_t>(b);
-  printf("result: %u, 0x%x\n", result.val, result.val);
   if (FP16_IS_INVALID(result.val)) {
     return FAILED;
   }
@@ -885,6 +884,23 @@ inline Status CheckInt32DivOverflow(int32_t a, int32_t b) {
            static_cast<uint32_t>(b));                                                                                  \
     return INTERNAL_ERROR;                                                                                             \
   }
-}  // namespace ge
 
+#define FMK_FP16_ZEROCHECK(a)                               \
+  if (fabs(a) < DBL_EPSILON) {                              \
+    GELOGE(INTERNAL_ERROR, "fp16 %f can not be zero !", a); \
+    return INTERNAL_ERROR;                                  \
+  }
+
+#define FMK_FLOAT_ZEROCHECK(a)                               \
+  if (fabs(a) < FLT_EPSILON) {                               \
+    GELOGE(INTERNAL_ERROR, "float %f can not be zero !", a); \
+    return INTERNAL_ERROR;                                   \
+  }
+
+#define FMK_DOUBLE_ZEROCHECK(a)                                \
+  if (fabs(a) < DBL_EPSILON) {                                 \
+    GELOGE(INTERNAL_ERROR, "double %lf can not be zero !", a); \
+    return INTERNAL_ERROR;                                     \
+  }
+}  // namespace ge
 #endif  // GE_COMMON_MATH_MATH_UTIL_H_

src/ge/common/model_parser/graph_parser_util.cc

@@ -1,501 +0,0 @@
-/**
- * Copyright 2019-2020 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_parser_util.h"
-#include <memory>
-#include "common/auth/file_saver.h"
-#include "common/convert/pb2json.h"
-#include "common/debug/log.h"
-#include "common/debug/memory_dumper.h"
-#include "common/model_parser/base.h"
-#include "common/model_saver.h"
-#include "common/properties_manager.h"
-#include "common/string_util.h"
-#include "common/types.h"
-#include "common/util.h"
-#include "common/util/error_manager/error_manager.h"
-#include "external/register/register_types.h"
-#include "framework/common/debug/ge_log.h"
-#include "framework/omg/parser/parser_inner_ctx.h"
-#include "graph/compute_graph.h"
-#include "graph/debug/ge_attr_define.h"
-#include "graph/debug/ge_attr_define.h"
-#include "graph/optimize/common/params.h"
-#include "graph/utils/type_utils.h"
-#include "omg/omg_inner_types.h"
-#include "omg/parser/model_parser.h"
-#include "omg/parser/parser_factory.h"
-#include "omg/parser/weights_parser.h"
-#include "parser/common/pre_checker.h"
-#include "proto/ge_ir.pb.h"
-#include "register/op_registry.h"
-
-namespace ge {
-namespace {
-// The function is incomplete. Currently, only l2_optimize, off_optimize is supported.
-const char *const kInputShapeSample1 = "\"input_name1:n1,c1,h1,w1\"";
-const char *const kInputShapeSample2 = "\"input_name1:1,3,224,224\"";
-const char *const kSplitError1 = "size not equal to 2 split by \":\"";
-const char *const kEmptyError = "can not be empty";
-const char *const kFloatNumError = "exist float number";
-const char *const kDigitError = "is not digit";
-const char *const kOutputTypeSample = "correct sample is \"opname:index:dtype\"";
-const char *const kOutputTypeSupport = "only support FP32, FP16, UINT8";
-const char *const kOutputTypeError = "The multiple out nodes set in output_type must be found in out_nodes.";
-
-vector<string> SplitInputShape(const std::string &input_shape) {
-  vector<string> shape_pair_vec;
-  size_t pos = input_shape.rfind(":");
-  if (pos != std::string::npos) {
-    shape_pair_vec.emplace_back(input_shape.substr(0, pos));
-    shape_pair_vec.emplace_back(input_shape.substr(pos + 1, input_shape.size() - pos));
-  }
-  return shape_pair_vec;
-}
-
-static std::map<std::string, ge::DataType> output_type_str_to_datatype = {
-  {"FP32", ge::DT_FLOAT}, {"FP16", ge::DT_FLOAT16}, {"UINT8", ge::DT_UINT8}};
-
-static bool CheckInputTrueOrFalse(const std::string &s, const std::string &atc_param) {
-  if ((s == "true") || (s == "false")) {
-    return true;
-  } else {
-    ErrorManager::GetInstance().ATCReportErrMessage("E10033", {"parameter", "value"}, {atc_param, s});
-    GELOGE(PARAM_INVALID, "Input parameter[--%s]'s value[%s] must be true or false.", atc_param.c_str(), s.c_str());
-    return false;
-  }
-}
-
-bool CheckDigitStr(std::string &str) {
-  for (char c : str) {
-    if (!isdigit(c)) {
-      GELOGE(domi::FAILED, "value[%s] is not positive integer", str.c_str());
-      return false;
-    }
-  }
-  return true;
-}
-
-Status StringToInt(std::string &str, int32_t &value) {
-  try {
-    if (!CheckDigitStr(str)) {
-      GELOGE(PARAM_INVALID, "Invalid of digit string: %s ", str.c_str());
-      ErrorManager::GetInstance().ATCReportErrMessage("E10001", {"parameter", "value", "reason"},
-                                                      {"--output_type", str, "is not positive integer"});
-      return PARAM_INVALID;
-    }
-    value = stoi(str);
-  } catch (std::invalid_argument &) {
-    GELOGE(PARAM_INVALID, "Invalid of digit string: %s, catch invalid_argument.", str.c_str());
-    ErrorManager::GetInstance().ATCReportErrMessage("E10014", {"parameter", "value"}, {"output_type", str});
-    return PARAM_INVALID;
-  } catch (std::out_of_range &) {
-    GELOGE(PARAM_INVALID, "Invalid of digit string: %s, catch out_of_range.", str.c_str());
-    ErrorManager::GetInstance().ATCReportErrMessage("E10013", {"parameter", "value"}, {"output_type", str});
-    return PARAM_INVALID;
-  }
-  return SUCCESS;
-}
-
-Status VerifyOutputTypeAndOutNodes(std::vector<std::string> &out_type_vec) {
-  std::vector<std::pair<std::string, int32_t>> user_out_nodes = domi::GetContext().user_out_nodes;
-  std::set<std::string> out_nodes_info;
-  for (uint32_t i = 0; i < user_out_nodes.size(); ++i) {
-    // out_nodes set should include output_type and output_format
-    std::string tmp = user_out_nodes[i].first + ":" + to_string(user_out_nodes[i].second);
-    out_nodes_info.emplace(tmp);
-  }
-  for (uint32_t i = 0; i < out_type_vec.size(); ++i) {
-    if (out_nodes_info.find(out_type_vec[i]) == out_nodes_info.end()) {
-      ErrorManager::GetInstance().ATCReportErrMessage("E10001", {"parameter", "value", "reason"},
-                                                      {"--output_type", out_type_vec[i], kOutputTypeError});
-      GELOGE(domi::FAILED, "Invalid value for --output_type[%s], %s.", out_type_vec[i].c_str(), kOutputTypeError);
-      return domi::FAILED;
-    }
-  }
-  return domi::SUCCESS;
-}
-
-Status ParseOutputType(const std::string &output_type, std::map<std::string, vector<uint32_t>> &out_type_index_map,
-                       std::map<std::string, vector<ge::DataType>> &out_type_dt_map) {
-  if (output_type.find(':') == std::string::npos) {
-    GELOGI("output_type is not multiple nodes, means all out nodes");
-    auto it = output_type_str_to_datatype.find(output_type);
-    if (it == output_type_str_to_datatype.end()) {
-      ErrorManager::GetInstance().ATCReportErrMessage("E10001", {"parameter", "value", "reason"},
-                                                      {"--output_type", output_type, kOutputTypeSupport});
-      GELOGE(PARAM_INVALID, "Invalid value for --output_type[%s], %s.", output_type.c_str(), kOutputTypeSupport);
-      return domi::FAILED;
-    }
-    return domi::SUCCESS;
-  }
-  std::vector<std::string> out_type_vec;
-  vector<string> nodes_v = StringUtils::Split(output_type, ';');
-  for (const string &node : nodes_v) {
-    vector<string> node_index_type_v = StringUtils::Split(node, ':');
-    if (node_index_type_v.size() != 3) {  // The size must be 3.
-      ErrorManager::GetInstance().ATCReportErrMessage("E10001", {"parameter", "value", "reason"},
-                                                      {"--output_type", node, kOutputTypeSample});
-      GELOGE(PARAM_INVALID, "Invalid value for --output_type[%s], %s.", node.c_str(), kOutputTypeSample);
-      return domi::FAILED;
-    }
-    ge::DataType tmp_dt;
-    std::string node_name = StringUtils::Trim(node_index_type_v[0]);
-    std::string index_str = StringUtils::Trim(node_index_type_v[1]);
-    int32_t index;
-    if (StringToInt(index_str, index) != SUCCESS) {
-      GELOGE(PARAM_INVALID, "This str must be digit string, while the actual input is %s.", index_str.c_str());
-      return domi::FAILED;
-    }
-    std::string dt_value = StringUtils::Trim(node_index_type_v[2]);
-    auto it = output_type_str_to_datatype.find(dt_value);
-    if (it == output_type_str_to_datatype.end()) {
-      ErrorManager::GetInstance().ATCReportErrMessage("E10001", {"parameter", "value", "reason"},
-                                                      {"--output_type", dt_value, kOutputTypeSupport});
-      GELOGE(ge::PARAM_INVALID, "Invalid value for --output_type[%s], %s.", dt_value.c_str(), kOutputTypeSupport);
-      return domi::FAILED;
-    } else {
-      tmp_dt = it->second;
-    }
-    out_type_vec.push_back(node_name + ":" + index_str);
-    auto it_index = out_type_index_map.find(node_name);
-    if (it_index == out_type_index_map.end()) {
-      vector<uint32_t> tmp_vec;
-      tmp_vec.push_back(index);
-      out_type_index_map.emplace(node_name, tmp_vec);
-    } else {
-      it_index->second.push_back(index);
-    }
-
-    auto it_dt = out_type_dt_map.find(node_name);
-    if (it_dt == out_type_dt_map.end()) {
-      vector<ge::DataType> tmp_vec;
-      tmp_vec.push_back(tmp_dt);
-      out_type_dt_map.emplace(node_name, tmp_vec);
-    } else {
-      it_dt->second.push_back(tmp_dt);
-    }
-  }
-  return VerifyOutputTypeAndOutNodes(out_type_vec);
-}
-
-Status CheckOutNode(ge::OpDescPtr op_desc, int32_t index) {
-  int32_t out_size = op_desc->GetOutputsSize();
-  if (index < 0 || index >= out_size) {
-    GELOGE(domi::FAILED,
-           "out_node [%s] output index:%d must be smaller "
-           "than node output size:%d and can not be negative!",
-           op_desc->GetName().c_str(), index, out_size);
-    std::string fail_reason = "output index:" + to_string(index) +
-                              " must be smaller than output size:" + to_string(out_size) + " and can not be negative!";
-    ErrorManager::GetInstance().ATCReportErrMessage("E10003", {"parameter", "value", "reason"},
-                                                    {"out_nodes", op_desc->GetName(), fail_reason});
-    return domi::FAILED;
-  }
-  return domi::SUCCESS;
-}
-
-Status GetOutputLeaf(NodePtr node, std::vector<std::pair<ge::NodePtr, int32_t>> &output_nodes_info) {
-  ge::OpDescPtr tmpDescPtr = node->GetOpDesc();
-  if (tmpDescPtr == nullptr) {
-    GELOGE(domi::FAILED, "Get outnode op desc fail.");
-    return domi::FAILED;
-  }
-  size_t size = tmpDescPtr->GetOutputsSize();
-  if (node->GetType() != NETOUTPUT) {
-    for (size_t index = 0; index < size; ++index) {
-      output_nodes_info.push_back(std::make_pair(node, index));
-    }
-  } else {
-    const auto in_anchors = node->GetAllInDataAnchors();
-    for (auto in_anchor : in_anchors) {
-      auto out_anchor = in_anchor->GetPeerOutAnchor();
-      if (out_anchor == nullptr) {
-        GELOGE(domi::FAILED, "Get leaf node op desc fail.");
-        return domi::FAILED;
-      }
-      auto out_node = out_anchor->GetOwnerNode();
-      output_nodes_info.push_back(std::make_pair(out_node, out_anchor->GetIdx()));
-    }
-  }
-  return SUCCESS;
-}
-
-void GetOutputNodesNameAndIndex(std::vector<std::pair<ge::NodePtr, int32_t>> &output_nodes_info,
-                                std::vector<std::string> &output_nodes_name) {
-  output_nodes_name.clear();
-  if (domi::GetContext().out_top_names.empty()) {
-    // tf process, no top name.
-    for (const auto output_node_info : output_nodes_info) {
-      std::string node_name = output_node_info.first->GetName();
-      int32_t index = output_node_info.second;
-      output_nodes_name.push_back(node_name + ":" + std::to_string(index));
-    }
-    return;
-  }
-  // caffe process, need add top name after node_name:index
-  for (size_t i = 0; i < output_nodes_info.size(); ++i) {
-    std::string node_name = output_nodes_info[i].first->GetName();
-    int32_t index = output_nodes_info[i].second;
-    if (i < domi::GetContext().out_top_names.size()) {
-      output_nodes_name.push_back(node_name + ":" + std::to_string(index) + ":" + domi::GetContext().out_top_names[i]);
-    } else {
-      GELOGW("Get top name of node [%s] fail.", node_name.c_str());
-      output_nodes_name.push_back(node_name + ":" + std::to_string(index));
-    }
-  }
-}
-}  // namespace
-
-FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY Status ParseOutputFp16NodesFormat(const string &is_output_fp16) {
-  if (is_output_fp16.empty()) {
-    return SUCCESS;
-  }
-
-  vector<domiTensorFormat_t> &output_formats = domi::GetContext().output_formats;
-  output_formats.clear();
-  vector<string> node_format_vec = StringUtils::Split(is_output_fp16, ',');
-  for (auto &is_fp16 : node_format_vec) {
-    StringUtils::Trim(is_fp16);
-    if (!CheckInputTrueOrFalse(is_fp16, "is_output_adjust_hw_layout")) {
-      GELOGE(PARAM_INVALID, "Invalid Param, is_output_adjust_hw_layout only support true/false: but is [%s]",
-             is_output_fp16.c_str());
-      return PARAM_INVALID;
-    }
-    if (is_fp16 == "false") {
-      output_formats.push_back(DOMI_TENSOR_ND);
-    } else if (is_fp16 == "true") {
-      output_formats.push_back(domi::DOMI_TENSOR_NC1HWC0);
-    }
-  }
-  return SUCCESS;
-}
-
-FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY Status SetOutputNodeInfo(ge::Graph &graph,
-                                                                          const std::string &output_type,
-                                                                          const std::string &output) {
-  ge::ComputeGraphPtr compute_graph = ge::GraphUtils::GetComputeGraph(graph);
-  GE_CHECK_NOTNULL(compute_graph);
-
-  std::vector<std::pair<std::string, int32_t>> user_out_nodes = domi::GetContext().user_out_nodes;
-  std::vector<domiTensorFormat_t> output_formats = domi::GetContext().output_formats;
-  std::vector<std::pair<ge::NodePtr, int32_t>> output_nodes_info;
-  std::vector<std::string> output_nodes_name;
-  std::map<std::string, vector<uint32_t>> out_type_index_map;
-  std::map<std::string, vector<ge::DataType>> out_type_dt_map;
-  if (!output_type.empty()) {
-    if (ParseOutputType(output_type, out_type_index_map, out_type_dt_map) != SUCCESS) {
-      GELOGE(domi::FAILED, "Parse output_type failed.");
-      return domi::FAILED;
-    }
-  }
-
-  // User declared outputs
-  for (uint32_t i = 0; i < user_out_nodes.size(); ++i) {
-    ge::NodePtr out_node = compute_graph->FindNode(user_out_nodes[i].first);
-    if (out_node == nullptr) {
-      GELOGE(domi::FAILED, "Can not find src node (%s) in graph.", user_out_nodes[i].first.c_str());
-      return domi::FAILED;
-    }
-    auto op_desc = out_node->GetOpDesc();
-    GE_CHECK_NOTNULL(op_desc);
-    if (CheckOutNode(op_desc, user_out_nodes[i].second) != SUCCESS) {
-      GELOGE(domi::FAILED, "Check out node (%s) fail.", user_out_nodes[i].first.c_str());
-      return domi::FAILED;
-    }
-    if (i < output_formats.size()) {
-      if (output_formats[i] == domi::DOMI_TENSOR_NC1HWC0) {
-        GELOGI("The output node [%s] should be set NC1HWC0", user_out_nodes[i].first.c_str());
-        if (!ge::AttrUtils::SetBool(op_desc, "output_set_fp16_nc1hwc0", true)) {
-          GELOGW("The output node [%s] set NC1HWC0 failed", user_out_nodes[i].first.c_str());
-        }
-      }
-    }
-    auto it_index = out_type_index_map.find(user_out_nodes[i].first);
-    auto it_dt = out_type_dt_map.find(user_out_nodes[i].first);
-    if ((it_index != out_type_index_map.end()) && (it_dt != out_type_dt_map.end())) {
-      GELOGI("The output node [%s] need to be set output_type", user_out_nodes[i].first.c_str());
-      (void)ge::AttrUtils::SetListDataType(op_desc, "_output_dt_list", it_dt->second);
-      (void)ge::AttrUtils::SetListInt(op_desc, "_output_dt_index", it_index->second);
-    }
-    output_nodes_info.push_back(std::make_pair(out_node, user_out_nodes[i].second));
-  }
-  // default output node (leaf)
-  if (user_out_nodes.empty()) {
-    for (ge::NodePtr node : compute_graph->GetDirectNode()) {
-      if (!node->GetInDataNodes().empty() && node->GetOutDataNodes().empty()) {
-        Status ret = GetOutputLeaf(node, output_nodes_info);
-        GE_CHK_BOOL_RET_STATUS(ret == SUCCESS, ret, "find leaf fail.");
-      }
-    }
-  }
-  GetOutputNodesNameAndIndex(output_nodes_info, output_nodes_name);
-  compute_graph->SetGraphOutNodesInfo(output_nodes_info);
-  domi::GetContext().net_out_nodes = output_nodes_name;
-  return domi::SUCCESS;
-}
-
-FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY bool ParseInputShape(
-  const string &input_shape, unordered_map<string, vector<int64_t>> &shape_map,
-  vector<pair<string, vector<int64_t>>> &user_shape_map, bool is_dynamic_input) {
-  vector<string> shape_vec = StringUtils::Split(input_shape, ';');
-  const int DEFAULT_SHAPE_PAIR_SIZE = 2;
-  for (const auto &shape : shape_vec) {
-    vector<string> shape_pair_vec = SplitInputShape(shape);
-    if (shape_pair_vec.size() != DEFAULT_SHAPE_PAIR_SIZE) {
-      ErrorManager::GetInstance().ATCReportErrMessage("E10002", {"shape", "reason", "sample"},
-                                                      {shape, kSplitError1, kInputShapeSample1});
-      GELOGW("Parse input parameter [--input_shape]'s shape[%s] failed, reason: %s, correct sample is %s.",
-             shape.c_str(), kSplitError1, kInputShapeSample1);
-      return false;
-    }
-    if (shape_pair_vec[1].empty()) {
-      ErrorManager::GetInstance().ATCReportErrMessage("E10002", {"shape", "reason", "sample"},
-                                                      {shape, kEmptyError, kInputShapeSample1});
-      GELOGW("Parse input parameter [--input_shape]'s shape[%s] failed, reason: %s, correct sample is %s.",
-             shape.c_str(), kEmptyError, kInputShapeSample1);
-      return false;
-    }
-
-    vector<string> shape_value_strs = StringUtils::Split(shape_pair_vec[1], ',');
-    vector<int64_t> shape_values;
-    for (auto &shape_value_str : shape_value_strs) {
-      // stoul: The method may throw an exception: invalid_argument/out_of_range
-      if (std::string::npos != shape_value_str.find('.')) {
-        ErrorManager::GetInstance().ATCReportErrMessage("E10002", {"shape", "reason", "sample"},
-                                                        {shape, kFloatNumError, kInputShapeSample2});
-        GELOGW("Parse input parameter [--input_shape]'s shape[%s] failed, reason: %s, correct sample is %s.",
-               shape.c_str(), kFloatNumError, kInputShapeSample2);
-        return false;
-      }
-
-      long left_result = 0;
-      try {
-        left_result = stol(StringUtils::Trim(shape_value_str));
-        if (!shape_value_str.empty() && (shape_value_str.front() == '-')) {
-          // The value maybe dynamic shape [-1], need substr it and verify isdigit.
-          shape_value_str = shape_value_str.substr(1);
-        }
-        for (char c : shape_value_str) {
-          if (!isdigit(c)) {
-            ErrorManager::GetInstance().ATCReportErrMessage("E10002", {"shape", "reason", "sample"},
-                                                            {shape, kDigitError, kInputShapeSample2});
-            GELOGE(PARAM_INVALID, "--input_shape's shape value[%s] is not digit", shape_value_str.c_str());
-            return false;
-          }
-        }
-      } catch (const std::out_of_range &) {
-        ErrorManager::GetInstance().ATCReportErrMessage("E10013", {"parameter", "value"},
-                                                        {"input_shape", shape_value_str});
-        GELOGW("Input parameter[--input_shape]’s value[%s] cause out of range execption!", shape_value_str.c_str());
-        return false;
-      } catch (const std::invalid_argument &) {
-        ErrorManager::GetInstance().ATCReportErrMessage("E10014", {"parameter", "value"},
-                                                        {"input_shape", shape_value_str});
-        GELOGW("Input parameter[--input_shape]’s value[%s] cause invalid argument!", shape_value_str.c_str());
-        return false;
-      } catch (...) {
-        ErrorManager::GetInstance().ATCReportErrMessage("E10015", {"parameter", "value"},
-                                                        {"input_shape", shape_value_str});
-        GELOGW("Input parameter[--input_shape]’s value[%s] cause unkown execption!", shape_value_str.c_str());
-        return false;
-      }
-      int64_t result = left_result;
-      // - 1 is not currently supported
-      if (!is_dynamic_input && result <= 0) {
-        ErrorManager::GetInstance().ATCReportErrMessage("E10011", {"shape", "result"}, {shape, std::to_string(result)});
-        GELOGW(
-          "Input parameter[--input_shape]’s shape value[%s] is invalid, "
-          "expect positive integer, but value is %ld.",
-          shape.c_str(), result);
-        return false;
-      }
-      shape_values.push_back(result);
-    }
-
-    shape_map.emplace(make_pair(StringUtils::Trim(shape_pair_vec[0]), shape_values));
-    user_shape_map.push_back(make_pair(StringUtils::Trim(shape_pair_vec[0]), shape_values));
-  }
-
-  return true;
-}
-
-FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY Status ParseOutputNodes(const string &out_nodes) {
-  try {
-    // parse output node
-    if (!out_nodes.empty()) {
-      domi::GetContext().out_nodes_map.clear();
-      domi::GetContext().user_out_nodes.clear();
-
-      vector<string> nodes_v = StringUtils::Split(out_nodes, ';');
-      for (const string &node : nodes_v) {
-        vector<string> key_value_v = StringUtils::Split(node, ':');
-        if (key_value_v.size() != 2) {  // The size must be 2.
-          ErrorManager::GetInstance().ATCReportErrMessage(
-            "E10001", {"parameter", "value", "reason"},
-            {"--out_nodes", node, "the correct format is \"node_name1:0;node_name1:1;node_name2:0\""});
-          GELOGE(PARAM_INVALID,
-                 "The input format of --out_nodes is invalid, the correct format is "
-                 "\"node_name1:0;node_name1:1;node_name2:0\", while the actual input is %s.",
-                 node.c_str());
-          return PARAM_INVALID;
-        }
-        auto iter = domi::GetContext().out_nodes_map.find(key_value_v[0]);
-        // stoi: The method may throw an exception: invalid_argument/out_of_range
-        if (!CheckDigitStr(key_value_v[1])) {
-          ErrorManager::GetInstance().ATCReportErrMessage("E10001", {"parameter", "value", "reason"},
-                                                          {"--out_nodes", out_nodes, "is not positive integer"});
-          GELOGE(PARAM_INVALID, "This str must be digit string, while the actual input is %s", out_nodes.c_str());
-          return PARAM_INVALID;
-        }
-        int32_t index = stoi(StringUtils::Trim(key_value_v[1]));
-        if (iter != domi::GetContext().out_nodes_map.end()) {
-          iter->second.emplace_back(index);
-        } else {
-          std::vector<int32_t> index_v;
-          index_v.emplace_back(index);
-          domi::GetContext().out_nodes_map.emplace(key_value_v[0], index_v);
-        }
-        domi::GetContext().user_out_nodes.push_back(std::make_pair(key_value_v[0], index));
-      }
-    }
-  } catch (std::invalid_argument &) {
-    GELOGE(PARAM_INVALID, "Invalid of out_nodes: %s ", out_nodes.c_str());
-    ErrorManager::GetInstance().ATCReportErrMessage("E10014", {"parameter", "value"}, {"out_nodes", out_nodes});
-    return PARAM_INVALID;
-  } catch (std::out_of_range &) {
-    GELOGE(PARAM_INVALID, "Invalid of out_nodes: %s ", out_nodes.c_str());
-    ErrorManager::GetInstance().ATCReportErrMessage("E10013", {"parameter", "value"}, {"out_nodes", out_nodes});
-    return PARAM_INVALID;
-  }
-  return SUCCESS;
-}
-
-FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY Status ParseOpConf(const char *op_conf) {
-  if (op_conf != nullptr && *op_conf != '\0') {
-    // divided by ":"
-    PropertiesManager::Instance().SetPropertyDelimiter(OP_CONF_DELIMITER);
-    // Parsing the op_conf configuration item file
-    if (!PropertiesManager::Instance().Init(op_conf)) {
-      GELOGE(FAILED, "op_name_map init failed!");
-      return FAILED;
-    }
-    // Return map and put it into ATC global variable
-    domi::GetContext().op_conf_map = PropertiesManager::Instance().GetPropertyMap();
-  }
-  return SUCCESS;
-}
-}  // namespace ge

src/ge/common/model_parser/graph_parser_util.h

@@ -1,62 +0,0 @@
-/**
- * Copyright 2019-2020 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_COMMON_GRAPH_PARSER_UTIL_H_
-#define GE_COMMON_GRAPH_PARSER_UTIL_H_
-
-#include <google/protobuf/message.h>
-#include <string>
-#include <unordered_map>
-#include <vector>
-#include "framework/common/types.h"
-#include "framework/omg/omg_inner_types.h"
-#include "proto/ge_ir.pb.h"
-#include "proto/om.pb.h"
-
-#include "graph/compute_graph.h"
-#include "graph/graph.h"
-#include "graph/model.h"
-#include "runtime/kernel.h"
-
-using domi::Status;
-using std::pair;
-using std::string;
-using std::unordered_map;
-using std::vector;
-
-namespace ge {
-Status SetOutputNodeInfo(ge::Graph &graph, const std::string &output_type, const std::string &output_format);
-
-Status ParseOutputFp16NodesFormat(const string &is_output_fp16);
-
-Status ParseOutputNodes(const string &out_nodes);
-
-bool ParseInputShape(const string &input_shape, unordered_map<string, vector<int64_t>> &shape_map,
-                     vector<pair<string, vector<int64_t>>> &user_shape_map, bool is_dynamic_input);
-
-Status ParseOpConf(const char *op_conf);
-}  // namespace ge
-
-namespace domi {
-/**
- * @ingroup domi_omg
- * @brief get omg context
- * @return reference of OmgContext
- */
-ge::OmgContext &GetContext();
-}  // namespace domi
-
-#endif  // GE_COMMON_GRAPH_PARSER_UTIL_H_

src/ge/common/profiling/profiling_manager.cc

@@ -76,8 +76,8 @@ FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY ge::Status ProfilingManager::In
     for (size_t i = 0; i < device_id_.size(); ++i) {
       ret = StartProfiling(0, device_id_[i]);
       if (ret != SUCCESS) {
-        GELOGE(ret, "Profiling start failed on device %d.", device_id_[i]);
-        return FAILED;
+        GELOGW("Profiling start failed on device %d.", device_id_[i]);
+        continue;
       }
       GELOGI("Profiling init succ on device %d.", device_id_[i]);
     }
@@ -316,7 +316,7 @@ FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY ge::Status ProfilingManager::St
     ProfMgrCfg prof_cfg = {send_profiling_config_};
     void *prof_handle = ProfMgrStartUp(&prof_cfg);
     if (prof_handle == nullptr) {
-      GELOGW("ProfMgrStartUp failed.");
+      GELOGW("ProfMgrStartUp failed on device %d ", device_id);
       return FAILED;
     }
     GELOGD("StartProfiling, prof_handle: %p", prof_handle);

src/ge/common/properties_manager.cc

@@ -31,193 +31,6 @@
 #include "graph/utils/attr_utils.h"
 
 namespace ge {
-namespace {
-const string kEnableFlag = "1";
-
-const uint32_t kAicoreOverflow = (0x1 << 0);
-const uint32_t kAtomicOverflow = (0x1 << 1);
-const uint32_t kAllOverflow = (kAicoreOverflow | kAtomicOverflow);
-}  // namespace
-
-FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY DumpProperties::DumpProperties(const DumpProperties &other) {
-  CopyFrom(other);
-}
-
-FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY DumpProperties &DumpProperties::operator=(
-  const DumpProperties &other) {
-  CopyFrom(other);
-  return *this;
-}
-
-FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY void DumpProperties::InitByOptions() {
-  enable_dump_.clear();
-  enable_dump_debug_.clear();
-  dump_path_.clear();
-  dump_step_.clear();
-  dump_mode_.clear();
-  is_op_debug_ = false;
-  op_debug_mode_ = 0;
-
-  string enable_dump;
-  (void)GetContext().GetOption(OPTION_EXEC_ENABLE_DUMP, enable_dump);
-  enable_dump_ = enable_dump;
-
-  string enable_dump_debug;
-  (void)GetContext().GetOption(OPTION_EXEC_ENABLE_DUMP_DEBUG, enable_dump_debug);
-  enable_dump_debug_ = enable_dump_debug;
-
-  if ((enable_dump_ == kEnableFlag) || (enable_dump_debug_ == kEnableFlag)) {
-    string dump_path;
-    if (GetContext().GetOption(OPTION_EXEC_DUMP_PATH, dump_path) == GRAPH_SUCCESS) {
-      if (!dump_path.empty() && dump_path[dump_path.size() - 1] != '/') {
-        dump_path = dump_path + "/";
-      }
-      dump_path = dump_path + CurrentTimeInStr() + "/";
-      GELOGI("Get dump path %s successfully", dump_path.c_str());
-      SetDumpPath(dump_path);
-    } else {
-      GELOGW("DUMP_PATH is not set");
-    }
-  }
-
-  if (enable_dump_ == kEnableFlag) {
-    string dump_step;
-    if (GetContext().GetOption(OPTION_EXEC_DUMP_STEP, dump_step) == GRAPH_SUCCESS) {
-      GELOGD("Get dump step %s successfully", dump_step.c_str());
-      SetDumpStep(dump_step);
-    }
-    string dump_mode;
-    if (GetContext().GetOption(OPTION_EXEC_DUMP_MODE, dump_mode) == GRAPH_SUCCESS) {
-      GELOGD("Get dump mode %s successfully", dump_mode.c_str());
-      SetDumpMode(dump_mode);
-    }
-    AddPropertyValue(DUMP_ALL_MODEL, {});
-  }
-
-  SetDumpDebugOptions();
-}
-
-// The following is the new dump scenario of the fusion operator
-FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY void DumpProperties::AddPropertyValue(
-  const std::string &model, const std::set<std::string> &layers) {
-  for (const std::string &layer : layers) {
-    GELOGI("This model %s config to dump layer %s", model.c_str(), layer.c_str());
-  }
-
-  model_dump_properties_map_[model] = layers;
-}
-
-FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY void DumpProperties::DeletePropertyValue(const std::string &model) {
-  auto iter = model_dump_properties_map_.find(model);
-  if (iter != model_dump_properties_map_.end()) {
-    model_dump_properties_map_.erase(iter);
-  }
-}
-
-FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY std::set<std::string> DumpProperties::GetAllDumpModel() const {
-  std::set<std::string> model_list;
-  for (auto &iter : model_dump_properties_map_) {
-    model_list.insert(iter.first);
-  }
-
-  return model_list;
-}
-
-FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY std::set<std::string> DumpProperties::GetPropertyValue(
-  const std::string &model) const {
-  auto iter = model_dump_properties_map_.find(model);
-  if (iter != model_dump_properties_map_.end()) {
-    return iter->second;
-  }
-  return {};
-}
-
-FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY bool DumpProperties::IsLayerNeedDump(
-  const std::string &model, const std::string &om_name, const std::string &op_name) const {
-  // if dump all
-  if (model_dump_properties_map_.find(DUMP_ALL_MODEL) != model_dump_properties_map_.end()) {
-    return true;
-  }
-
-  // if this model need dump
-  auto om_name_iter = model_dump_properties_map_.find(om_name);
-  auto model_name_iter = model_dump_properties_map_.find(model);
-  if (om_name_iter != model_dump_properties_map_.end() || model_name_iter != model_dump_properties_map_.end()) {
-    // if no dump layer info, dump all layer in this model
-    auto model_iter = om_name_iter != model_dump_properties_map_.end() ? om_name_iter : model_name_iter;
-    if (model_iter->second.empty()) {
-      return true;
-    }
-
-    return model_iter->second.find(op_name) != model_iter->second.end();
-  }
-
-  GELOGD("Model %s is not seated to be dump.", model.c_str());
-  return false;
-}
-
-FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY void DumpProperties::SetDumpPath(const std::string &path) {
-  dump_path_ = path;
-}
-
-FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY std::string DumpProperties::GetDumpPath() const { return dump_path_; }
-
-FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY void DumpProperties::SetDumpStep(const std::string &step) {
-  dump_step_ = step;
-}
-
-FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY std::string DumpProperties::GetDumpStep() const { return dump_step_; }
-
-FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY void DumpProperties::SetDumpMode(const std::string &mode) {
-  dump_mode_ = mode;
-}
-
-FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY std::string DumpProperties::GetDumpMode() const { return dump_mode_; }
-
-void DumpProperties::CopyFrom(const DumpProperties &other) {
-  if (&other != this) {
-    enable_dump_ = other.enable_dump_;
-    enable_dump_debug_ = other.enable_dump_debug_;
-    dump_path_ = other.dump_path_;
-    dump_step_ = other.dump_step_;
-    dump_mode_ = other.dump_mode_;
-
-    model_dump_properties_map_ = other.model_dump_properties_map_;
-    is_op_debug_ = other.is_op_debug_;
-    op_debug_mode_ = other.op_debug_mode_;
-  }
-}
-
-void DumpProperties::SetDumpDebugOptions() {
-  if (enable_dump_debug_ == kEnableFlag) {
-    string dump_debug_mode;
-    if (GetContext().GetOption(OPTION_EXEC_DUMP_DEBUG_MODE, dump_debug_mode) == GRAPH_SUCCESS) {
-      GELOGD("Get dump debug mode %s successfully", dump_debug_mode.c_str());
-    } else {
-      GELOGW("Dump debug mode is not set.");
-      return;
-    }
-
-    if (dump_debug_mode == OP_DEBUG_AICORE) {
-      GELOGD("ge.exec.dumpDebugMode=aicore_overflow, op debug is open.");
-      is_op_debug_ = true;
-      op_debug_mode_ = kAicoreOverflow;
-    } else if (dump_debug_mode == OP_DEBUG_ATOMIC) {
-      GELOGD("ge.exec.dumpDebugMode=atomic_overflow, op debug is open.");
-      is_op_debug_ = true;
-      op_debug_mode_ = kAtomicOverflow;
-    } else if (dump_debug_mode == OP_DEBUG_ALL) {
-      GELOGD("ge.exec.dumpDebugMode=all, op debug is open.");
-      is_op_debug_ = true;
-      op_debug_mode_ = kAllOverflow;
-    } else {
-      GELOGW("ge.exec.dumpDebugMode is invalid.");
-    }
-  } else {
-    GELOGI("ge.exec.enableDumpDebug is false or is not set.");
-  }
-}
-
 PropertiesManager::PropertiesManager() : is_inited_(false), delimiter("=") {}
 PropertiesManager::~PropertiesManager() {}
 

src/ge/common/properties_manager.h

@@ -24,6 +24,7 @@
 #include <vector>
 
 #include "graph/op_desc.h"
+#include "common/dump/dump_properties.h"
 
 namespace ge {
 // Configuration property management
@@ -32,50 +33,6 @@ static const char *USE_FUSION __attribute__((unused)) = "FMK_USE_FUSION";
 static const char *TIMESTAT_ENABLE __attribute__((unused)) = "DAVINCI_TIMESTAT_ENABLE";
 static const char *ANNDROID_DEBUG __attribute__((unused)) = "ANNDROID_DEBUG";
 
-class DumpProperties {
- public:
-  DumpProperties() = default;
-  ~DumpProperties() = default;
-  DumpProperties(const DumpProperties &dump);
-  DumpProperties &operator=(const DumpProperties &dump);
-
-  void InitByOptions();
-
-  void AddPropertyValue(const std::string &model, const std::set<std::string> &layers);
-  void DeletePropertyValue(const std::string &model);
-
-  std::set<std::string> GetAllDumpModel() const;
-  std::set<std::string> GetPropertyValue(const std::string &model) const;
-  bool IsLayerNeedDump(const std::string &model, const std::string &om_name, const std::string &op_name) const;
-
-  void SetDumpPath(const std::string &path);
-  std::string GetDumpPath() const;
-
-  void SetDumpStep(const std::string &step);
-  std::string GetDumpStep() const;
-
-  void SetDumpMode(const std::string &mode);
-  std::string GetDumpMode() const;
-
-  bool IsOpDebugOpen() const { return is_op_debug_; }
-  uint32_t GetOpDebugMode() const { return op_debug_mode_; }
-
- private:
-  void CopyFrom(const DumpProperties &other);
-  void SetDumpDebugOptions();
-
-  string enable_dump_;
-  string enable_dump_debug_;
-
-  std::string dump_path_;
-  std::string dump_step_;
-  std::string dump_mode_;
-  std::map<std::string, std::set<std::string>> model_dump_properties_map_;
-
-  bool is_op_debug_ = false;
-  uint32_t op_debug_mode_ = 0;
-};
-
 class PropertiesManager {
  public:
   // Singleton

src/ge/common/types.cc

@@ -502,6 +502,7 @@ const uint32_t MODEL_FILE_HEAD_LEN = 256;
 /// @brief Input node type
 ///
 const std::string INPUT_TYPE = "Input";
+const std::string DUMMY_DATA = "DummyData";
 
 ///
 /// @ingroup domi_omg

src/ge/common/util.cc

@@ -57,7 +57,7 @@ const int kWarningThreshold = 536870912 * 2;  // 536870912 represent 512M
 /// Based on the security coding specification and the current actual (protobuf) model size, it is determined as 2G-1
 const int kMaxFileSizeLimit = INT_MAX;
 const int kMaxBuffSize = 256;
-const char *const kPathValidReason = "The path can only contains 'a-z' 'A-Z' '0-9' '-' '.' '_' and chinese character";
+const char *const kPathValidReason = "The path can only contain 'a-z' 'A-Z' '0-9' '-' '.' '_' and chinese character";
 }  // namespace
 
 namespace ge {
@@ -311,6 +311,14 @@ FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY uint64_t GetCurrentTimestap() {
   return static_cast<uint64_t>(total_use_time);
 }
 
+FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY uint32_t GetCurrentSecondTimestap() {
+  struct timeval tv {};
+  int ret = gettimeofday(&tv, nullptr);
+  GE_LOGE_IF(ret != 0, "Func gettimeofday may failed: ret=%d", ret);
+  auto total_use_time = tv.tv_sec;  // seconds
+  return static_cast<uint32_t>(total_use_time);
+}
+
 FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY bool CheckInt64MulOverflow(int64_t a, int64_t b) {
   if (a > 0) {
     if (b > 0) {
@@ -372,10 +380,9 @@ FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY bool CheckInputPathValid(const
   }
 
   // A regular matching expression to verify the validity of the input file path
-  // ^(/|./|(../)+|)([.]?[\u4e00-\u9fa5A-Za-z0-9_.-]+/)*[\u4e00-\u9fa5A-Za-z0-9_+.-]+$
   // Path section: Support upper and lower case letters, numbers dots(.) chinese and underscores
   // File name section: Support upper and lower case letters, numbers, underscores chinese and dots(.)
-  std::string mode = "^(/+|./+|(../+)+|)(../|([.]?[\u4e00-\u9fa5A-Za-z0-9_.-]+)/+)*[\u4e00-\u9fa5A-Za-z0-9_+.-]+$";
+  std::string mode = "^[\u4e00-\u9fa5A-Za-z0-9./_-]+$";
 
   GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(
     !ValidateStr(real_path, mode),
@@ -408,10 +415,9 @@ FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY bool CheckOutputPathValid(const
     return "", "Path[%s] len is too long, it must be less than %d", file_path.c_str(), PATH_MAX);
 
   // A regular matching expression to verify the validity of the input file path
-  // ^(/|./|(../)+|)([.]?[\u4e00-\u9fa5A-Za-z0-9_-]+/)*[\u4e00-\u9fa5A-Za-z0-9_+.-]+$
   // Path section: Support upper and lower case letters, numbers dots(.) chinese and underscores
   // File name section: Support upper and lower case letters, numbers, underscores chinese and dots(.)
-  std::string mode = "^(/+|./+|(../+)+|)(../|([.]?[\u4e00-\u9fa5A-Za-z0-9_.-]+)/+)*[\u4e00-\u9fa5A-Za-z0-9_+.-]+$";
+  std::string mode = "^[\u4e00-\u9fa5A-Za-z0-9./_-]+$";
 
   GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(
     !ValidateStr(file_path, mode),
@@ -460,9 +466,9 @@ FMK_FUNC_HOST_VISIBILITY bool ValidateStr(const std::string &str, const std::str
   int ret = regcomp(&reg, mode.c_str(), cflags);
   if (ret) {
     regerror(ret, &reg, ebuff, kMaxBuffSize);
-    GELOGE(ge::PARAM_INVALID, "regcomp failed, reason: %s", ebuff);
+    GELOGW("regcomp failed, reason: %s", ebuff);
     regfree(&reg);
-    return false;
+    return true;
   }
 
   ret = regexec(&reg, str.c_str(), 0, nullptr, 0);

src/ge/engine_manager/dnnengine_manager.cc

@@ -42,6 +42,8 @@ const char *const kVectorCore = "VectorCore";
 const char *const kVectorEngine = "VectorEngine";
 const char *const kAIcoreEngine = "AIcoreEngine";
 const char *const kCustomOpFlag = "_custom_op_flag";
+const char *const kHostCpuEngineName = "DNN_VM_HOST_CPU";
+const char *const kHostCpuOpKernelLibName = "DNN_VM_HOST_CPU_OP_STORE";
 }  // namespace
 
 namespace ge {
@@ -181,6 +183,7 @@ std::string DNNEngineManager::GetDNNEngineName(const OpDescPtr &op_desc) {
     GELOGI("DNNEngineManager: Can not get op info by op type %s", op_desc->GetType().c_str());
     return "";
   }
+  GE_IF_BOOL_EXEC(ge::GetContext().GetHostExecFlag(), return GetHostCpuEngineName(op_infos, op_desc));
   std::string ge_core_type;
   Status ret = ge::GetContext().GetOption(ge::CORE_TYPE, ge_core_type);
   GE_IF_BOOL_EXEC(ret != SUCCESS, GELOGD("get the option CORE_TYPE fail, set it to default value VECTOR_ENGINE"));
@@ -245,6 +248,22 @@ std::string DNNEngineManager::GetDNNEngineName(const OpDescPtr &op_desc) {
   return "";
 }
 
+std::string DNNEngineManager::GetHostCpuEngineName(const std::vector<OpInfo> &op_infos,
+                                                   const OpDescPtr &op_desc) const {
+  for (const auto &it : op_infos) {
+    if ((it.engine == kHostCpuEngineName) && (it.opKernelLib == kHostCpuOpKernelLibName)) {
+      op_desc->SetOpEngineName(kHostCpuEngineName);
+      op_desc->SetOpKernelLibName(kHostCpuOpKernelLibName);
+      GELOGI("DNNEngineManager: Set OpKernelLibName %s and OpEngineName %s to %s", kHostCpuOpKernelLibName,
+             kHostCpuEngineName, op_desc->GetName().c_str());
+      return kHostCpuEngineName;
+    }
+  }
+  GELOGE(FAILED, "DNNEngineManager: HostCpuEngine not support [%s, %s].", op_desc->GetName().c_str(),
+         op_desc->GetType().c_str());
+  return "";
+}
+
 const std::map<std::string, SchedulerConf> &DNNEngineManager::GetSchedulers() const { return schedulers_; }
 
 Status DNNEngineManager::ParserJsonFile() {

src/ge/engine_manager/dnnengine_manager.h

@@ -76,6 +76,7 @@ class DNNEngineManager {
   Status ParserEngineMessage(const json engines_json, const string &scheduler_mark,
                              map<string, EngineConfPtr> &engines);
   Status CheckJsonFile();
+  std::string GetHostCpuEngineName(const std::vector<OpInfo> &op_infos, const OpDescPtr &op_desc) const;
   PluginManager plugin_mgr_;
   std::map<std::string, DNNEnginePtr> engines_map_;
   std::map<std::string, ge::DNNEngineAttribute> engines_attrs_map_;

src/ge/engine_manager/engine_conf.json

@@ -5,6 +5,13 @@
       "name": "1980_hwts",
       "ex_attrs": "",
       "cal_engines": [
+        {
+          "id": "DNN_VM_HOST_CPU",
+          "name": "HOST_CPU",
+          "independent": false,
+          "skip_assign_stream": true,
+          "attach": true
+        },
         {
           "id": "DNN_VM_GE_LOCAL",
           "name": "GE_LOCAL",

src/ge/executor/CMakeLists.txt

@@ -26,6 +26,9 @@ file(GLOB PROTO_LIST RELATIVE ${CMAKE_CURRENT_LIST_DIR}
 
 file(GLOB SRC_LIST RELATIVE ${CMAKE_CURRENT_LIST_DIR}
         "ge_executor.cc"
+        "../common/dump/dump_properties.cc"
+        "../common/dump/dump_manager.cc"
+        "../common/dump/dump_op.cc"
         "../common/ge/op_tiling_manager.cc"
         "../common/ge/plugin_manager.cc"
         "../common/profiling/profiling_manager.cc"

src/ge/executor/ge_executor.cc

@@ -23,6 +23,7 @@
 #include "common/ge/ge_util.h"
 #include "common/helper/model_helper.h"
 #include "common/profiling/profiling_manager.h"
+#include "common/dump/dump_manager.h"
 #include "common/util.h"
 #include "framework/common/debug/ge_log.h"
 #include "framework/common/util.h"
@@ -35,6 +36,8 @@
 #include "graph/utils/graph_utils.h"
 #include "mmpa/mmpa_api.h"
 #include "single_op/single_op_manager.h"
+#include "graph/manager/graph_var_manager.h"
+#include "graph/load/new_model_manager/davinci_model.h"
 
 using std::string;
 using std::vector;
@@ -348,18 +351,46 @@ Status GeExecutor::SetDynamicDims(uint32_t model_id, void *dynamic_input_addr, u
   }
 
   vector<uint64_t> cur_dynamic_dims;
-  if (GetCurDynamicDims(model_id, dynamic_dims, cur_dynamic_dims) != SUCCESS) {
-    GELOGE(FAILED, "GetCurDynamicDims failed.");
+  std::vector<ge::TensorDesc> input_desc;
+  std::vector<ge::TensorDesc> output_desc;
+  ret = GetModelDescInfo(model_id, input_desc, output_desc);
+  if (ret != ge::SUCCESS) {
+    GELOGE(FAILED, "GetModelDescInfo failed.");
     return FAILED;
   }
-
+  vector<string> user_designate_shape_order;
+  vector<int64_t> all_data_dims;
+  ret = GetUserDesignateShapeOrder(model_id, user_designate_shape_order);
+  if (ret != ge::SUCCESS) {
+    GELOGE(FAILED, "GetUserDesignateShapeOrder failed.");
+    return FAILED;
+  }
+  for (auto &data_name : user_designate_shape_order) {
+    for (size_t j = 0; j < input_desc.size(); ++j) {
+      if (input_desc.at(j).GetName() == data_name) {
+        for (auto dim : input_desc.at(j).GetShape().GetDims()) {
+          all_data_dims.push_back(dim);
+        }
+        break;
+      }
+    }
+  }
+  if (dynamic_dims.size() != all_data_dims.size()) {
+    GELOGE(FAILED, "Dynamic input size [%lu] is not equal with all data dims size [%lu]!", dynamic_dims.size(),
+           all_data_dims.size());
+    return FAILED;
+  }
+  for (std::size_t i = 0; i < all_data_dims.size(); ++i) {
+    if (all_data_dims[i] < 0) {
+      cur_dynamic_dims.push_back(dynamic_dims[i]);
+    }
+  }
   size_t dynamic_dim_num = cur_dynamic_dims.size();
   uint64_t dynamic_input_size = static_cast<uint64_t>(dynamic_dim_num * sizeof(uint64_t));
   if (length < dynamic_input_size) {
     GELOGE(FAILED, "Dynamic input size [%lu] is less than [%lu]!", length, dynamic_input_size);
     return FAILED;
   }
-
   for (uint32_t i = 0; i < dynamic_dim_num; ++i) {
     // Memcpy dynamic dim[i] from host to device
     if (rtMemcpy(reinterpret_cast<void *>(reinterpret_cast<uint8_t *>(dynamic_input_addr) + sizeof(uint64_t) * i),
@@ -549,6 +580,12 @@ Status GeExecutor::UnloadModel(uint32_t model_id) {
     GELOGE(ret, "[GraphLoader] DestroyAicpuSessionForInfer failed. model id: %u", model_id);
     return FAILED;
   }
+
+  std::shared_ptr<DavinciModel> davinci_model = ModelManager::GetInstance()->GetModel(model_id);
+  if (davinci_model != nullptr) {
+    uint64_t session_id = davinci_model->GetSessionId();
+    VarManagerPool::Instance().RemoveVarManager(session_id);
+  }
   return GraphLoader::UnloadModel(model_id);
 }
 
@@ -658,6 +695,30 @@ Status GeExecutor::GetCombinedDynamicDims(uint32_t model_id, vector<vector<int64
   return SUCCESS;
 }
 
+///
+/// @ingroup ge
+/// @brief Get user designeate shape order
+/// @param [in] model_id
+/// @param [out] user_designate_shape_order
+/// @return execute result
+///
+Status GeExecutor::GetUserDesignateShapeOrder(uint32_t model_id, vector<string> &user_designate_shape_order) {
+  GELOGI("Begin to get user designate shape info.");
+  if (!isInit_) {
+    GELOGE(GE_EXEC_NOT_INIT, "GeExecutor has not been initialized!");
+    return GE_EXEC_NOT_INIT;
+  }
+
+  Status ret = GraphExecutor::GetUserDesignateShapeOrder(model_id, user_designate_shape_order);
+  if (ret != SUCCESS) {
+    GELOGE(ret, "GetUserDesignateShapeOrder failed.");
+    return ret;
+  }
+
+  GELOGI("Get user designate shape order succ.");
+  return SUCCESS;
+}
+
 ///
 /// @ingroup ge
 /// @brief Get AIPP input format
@@ -674,7 +735,7 @@ Status GeExecutor::GetAIPPInfo(uint32_t model_id, uint32_t index, AippConfigInfo
   }
   Status ret = GraphExecutor::GetAIPPInfo(model_id, index, aipp_info);
   if (ret != SUCCESS) {
-    GELOGE(ret, "GetAIPPInfo failed.");
+    GELOGW("GetAIPPInfo is not success.");
     return ret;
   }
   GELOGI("GetAIPPInfo succ.");
@@ -1020,4 +1081,26 @@ Status GeExecutor::GetAllAippInputOutputDims(uint32_t model_id, uint32_t index,
   GELOGI("GetAllAippInputOutputDims succ.");
   return SUCCESS;
 }
+
+Status GeExecutor::GetOpDescInfo(uint32_t device_id, uint32_t stream_id, uint32_t task_id, OpDescInfo &op_desc_info) {
+  GELOGI("Begin to GetOpDescInfo.");
+  Status ret = GraphExecutor::GetOpDescInfo(device_id, stream_id, task_id, op_desc_info);
+  if (ret != SUCCESS) {
+    GELOGE(ret, "GetOpDescInfo failed.");
+    return ret;
+  }
+  GELOGI("GetOpDescInfo succ.");
+  return SUCCESS;
+}
+
+Status GeExecutor::SetDump(const DumpConfig &dump_config) {
+  GELOGI("Start to set dump config");
+  auto ret = DumpManager::GetInstance().SetDumpConf(dump_config);
+  if (ret != SUCCESS) {
+    GELOGE(ret, "Set dump conf failed");
+    return ret;
+  }
+  GELOGI("Set dump config succ.");
+  return SUCCESS;
+}
 }  // namespace ge

src/ge/executor/module.mk

@@ -3,6 +3,9 @@ LOCAL_PATH := $(call my-dir)
 local_ge_executor_src_files :=  \
     ge_executor.cc \
     ../common/profiling/profiling_manager.cc \
+    ../common/dump/dump_properties.cc \
+    ../common/dump/dump_manager.cc \
+    ../common/dump/dump_op.cc \
     ../common/ge/plugin_manager.cc \
     ../common/ge/op_tiling_manager.cc \
     ../graph/load/graph_loader.cc \

src/ge/ge_inference.mk

@@ -26,6 +26,9 @@ COMMON_LOCAL_SRC_FILES := \
     common/formats/format_transfers/format_transfer_nchw_fz_c04.cc \
     common/formats/formats.cc \
     common/profiling/profiling_manager.cc \
+    common/dump/dump_properties.cc \
+    common/dump/dump_manager.cc \
+    common/dump/dump_op.cc \
     common/helper/model_cache_helper.cc \
     ge_local_engine/engine/host_cpu_engine.cc \
 
@@ -42,6 +45,7 @@ GRAPH_MANAGER_LOCAL_SRC_FILES := \
     graph/manager/graph_manager_utils.cc \
     graph/manager/graph_context.cc \
     graph/preprocess/graph_preprocess.cc \
+    graph/preprocess/multi_batch_options.cc \
     graph/preprocess/multi_batch_copy_graph.cc \
     graph/execute/graph_execute.cc \
     graph/load/graph_loader.cc \
@@ -149,6 +153,7 @@ OMG_HOST_SRC_FILES := \
     host_kernels/slice_kernel.cc \
     host_kernels/slice_d_kernel.cc \
     host_kernels/dynamic_stitch_kernel.cc \
+    host_kernels/identity_kernel.cc \
     graph/passes/stop_gradient_pass.cc \
     graph/passes/prevent_gradient_pass.cc \
     graph/passes/identity_pass.cc \
@@ -165,12 +170,16 @@ OMG_HOST_SRC_FILES := \
     graph/passes/switch_to_stream_switch_pass.cc \
     graph/passes/attach_stream_label_pass.cc \
     graph/passes/multi_batch_pass.cc \
+    graph/passes/multi_batch_clone_pass.cc \
+    graph/passes/subexpression_migration_pass.cc \
+    graph/passes/unused_args_clean_pass.cc \
     graph/passes/next_iteration_pass.cc \
     graph/passes/control_trigger_pass.cc \
     graph/passes/cond_pass.cc \
     graph/passes/cond_remove_pass.cc \
     graph/passes/for_pass.cc \
     graph/passes/enter_pass.cc \
+    graph/passes/assign_pass.cc \
     graph/passes/addn_pass.cc \
     graph/passes/common_subexpression_elimination_pass.cc \
     graph/passes/transop_symmetry_elimination_pass.cc \
@@ -185,11 +194,10 @@ OMG_HOST_SRC_FILES := \
     graph/passes/transpose_transdata_pass.cc \
     graph/passes/hccl_memcpy_pass.cc \
     graph/passes/flow_ctrl_pass.cc \
+    graph/passes/global_step_insert_pass.cc \
     graph/passes/link_gen_mask_nodes_pass.cc \
     graph/passes/replace_with_empty_const_pass.cc \
     graph/passes/hccl_group_pass.cc \
-    graph/passes/switch_fusion_pass.cc \
-    graph/passes/switch_split_pass.cc \
     graph/passes/memcpy_addr_async_pass.cc \
     graph/passes/set_input_output_offset_pass.cc \
 

src/ge/ge_runner.mk

@@ -26,6 +26,9 @@ LIBGE_LOCAL_SRC_FILES := \
     common/ge/op_tiling_manager.cc\
     common/helper/model_cache_helper.cc \
     common/profiling/profiling_manager.cc \
+    common/dump/dump_manager.cc \
+    common/dump/dump_properties.cc \
+    common/dump/dump_op.cc \
     engine_manager/dnnengine_manager.cc \
     ge_local_engine/engine/host_cpu_engine.cc \
     generator/ge_generator.cc \
@@ -93,7 +96,6 @@ LIBGE_LOCAL_SRC_FILES := \
     graph/manager/util/variable_accelerate_ctrl.cc               \
     graph/optimize/graph_optimize.cc \
     graph/optimize/mem_rw_conflict_optimize.cc \
-    graph/optimize/optimizer/allreduce_fusion_pass.cc \
     graph/optimize/summary_optimize.cc \
     graph/partition/engine_place.cc \
     graph/partition/graph_partition.cc \
@@ -119,10 +121,10 @@ LIBGE_LOCAL_SRC_FILES := \
     graph/passes/dimension_compute_pass.cc \
     graph/passes/dropout_pass.cc \
     graph/passes/hccl_group_pass.cc \
-    graph/passes/switch_fusion_pass.cc \
-    graph/passes/switch_split_pass.cc \
     graph/passes/enter_pass.cc \
+    graph/passes/assign_pass.cc \
     graph/passes/flow_ctrl_pass.cc \
+    graph/passes/global_step_insert_pass.cc \
     host_kernels/transpose_kernel.cc \
     host_kernels/add_kernel.cc \
     host_kernels/broadcast_args_kernel.cc \
@@ -131,6 +133,7 @@ LIBGE_LOCAL_SRC_FILES := \
     host_kernels/concat_offset_kernel.cc \
     host_kernels/concat_v2_kernel.cc \
     host_kernels/dynamic_stitch_kernel.cc \
+    host_kernels/identity_kernel.cc \
     host_kernels/empty_kernel.cc \
     host_kernels/expanddims_kernel.cc \
     host_kernels/fill_kernel.cc \
@@ -172,6 +175,9 @@ LIBGE_LOCAL_SRC_FILES := \
     graph/passes/link_gen_mask_nodes_pass.cc \
     graph/passes/merge_pass.cc \
     graph/passes/multi_batch_pass.cc \
+    graph/passes/multi_batch_clone_pass.cc \
+    graph/passes/subexpression_migration_pass.cc \
+    graph/passes/unused_args_clean_pass.cc \
     graph/passes/net_output_pass.cc \
     graph/passes/next_iteration_pass.cc \
     graph/passes/no_use_reshape_remove_pass.cc \
@@ -225,6 +231,7 @@ LIBGE_LOCAL_SRC_FILES := \
     graph/preprocess/graph_preprocess.cc \
     graph/preprocess/insert_op/ge_aipp_op.cc \
     graph/preprocess/insert_op/util_insert_aipp_op.cc \
+    graph/preprocess/multi_batch_options.cc \
     graph/preprocess/multi_batch_copy_graph.cc \
     init/gelib.cc \
     model/ge_model.cc \
@@ -267,10 +274,17 @@ LIBGE_LOCAL_SRC_FILES := \
     hybrid/node_executor/aicpu/aicpu_ext_info.cc                         \
     hybrid/node_executor/aicpu/aicpu_node_executor.cc                    \
     hybrid/node_executor/compiledsubgraph/known_node_executor.cc         \
-    hybrid/node_executor/hostcpu/ge_local_node_executor.cc               \
+    hybrid/node_executor/ge_local/ge_local_node_executor.cc              \
+    hybrid/node_executor/host_cpu/host_cpu_node_executor.cc              \
+    hybrid/node_executor/host_cpu/kernel_factory.cc                      \
+    hybrid/node_executor/host_cpu/kernel/no_op_kernel.cc                 \
+    hybrid/node_executor/host_cpu/kernel/variable_kernel.cc              \
+    hybrid/node_executor/host_cpu/kernel/assign_kernel.cc                \
+    hybrid/node_executor/host_cpu/kernel/random_uniform_kernel.cc        \
     hybrid/node_executor/controlop/control_op_executor.cc                \
     hybrid/node_executor/partitioned_call/partitioned_call_node_executor.cc \
     hybrid/node_executor/hccl/hccl_node_executor.cc                      \
+    hybrid/node_executor/rts/rts_node_executor.cc                        \
     hybrid/node_executor/node_executor.cc                                \
     hybrid/node_executor/task_context.cc                                 \
     hybrid/hybrid_davinci_model.cc                                       \
@@ -343,7 +357,6 @@ LOCAL_SHARED_LIBRARIES := \
     libgraph \
     libregister \
     libge_common \
-    libhccl \
     libmsprof \
     liberror_manager \
 
@@ -425,7 +438,6 @@ LOCAL_SHARED_LIBRARIES := \
     libc_sec \
     libslog \
     libmmpa \
-    libhccl \
     libmsprof \
 
 LOCAL_LDFLAGS := -lrt -ldl
@@ -457,7 +469,6 @@ LOCAL_SHARED_LIBRARIES := \
     libc_sec \
     libslog \
     libmmpa \
-    libhccl \
     libmsprof \
 
 LOCAL_LDFLAGS := -lrt -ldl

src/ge/generator/ge_generator.cc

@@ -658,10 +658,13 @@ Status GeGenerator::Impl::BuildModel(const Graph &graph, const vector<GeTensor>
 
   if (ret != SUCCESS) {
     GELOGE(GE_GENERATOR_GRAPH_MANAGER_BUILD_GRAPH_FAILED, "GraphManager build graph fail, graph id: %u", id);
+    VarManagerPool::Instance().RemoveVarManager(session_id);
     return GE_GENERATOR_GRAPH_MANAGER_BUILD_GRAPH_FAILED;
   }
   id += 1;
 
+  VarManagerPool::Instance().RemoveVarManager(session_id);
+
   return SUCCESS;
 }
 

src/ge/graph/build/graph_builder.cc

@@ -28,6 +28,7 @@
 #include "graph/common/ge_call_wrapper.h"
 #include "init/gelib.h"
 #include "model/ge_model.h"
+#include "graph/ge_context.h"
 
 using domi::BuildMode;
 
@@ -166,11 +167,15 @@ Status GraphBuilder::Build(ComputeGraphPtr &comp_graph, std::vector<SubGraphInfo
   return SUCCESS;
 }
 
-Status GraphBuilder::BuildForKnownShapeGraph(ComputeGraphPtr &comp_graph,
-                                             std::vector<SubGraphInfoPtr> &subgraph_ptr_list, GeModelPtr &ge_model_ptr,
-                                             uint64_t session_id) {
+Status GraphBuilder::BuildForKnownShapeGraph(ComputeGraphPtr &comp_graph, std::vector<SubGraphInfoPtr> &subgraph_list,
+                                             GeModelPtr &ge_model_ptr, uint64_t session_id) {
+  if (ge::GetContext().GetHostExecFlag()) {
+    GE_CHK_STATUS_RET(BuildForHostCpuGraph(comp_graph, ge_model_ptr, session_id), "Build for host-cpu graph failed.");
+    return SUCCESS;
+  }
+
   GELOGI("Begin to build known shape graph[%s].", comp_graph->GetName().c_str());
-  Status ret = SecondPartition(comp_graph, subgraph_ptr_list);
+  Status ret = SecondPartition(comp_graph, subgraph_list);
   GE_CHK_STATUS_RET(ret, "Graph[%s] second partition Failed.", comp_graph->GetName().c_str());
   auto subgraph_map = graph_partitioner_.GetSubGraphMap();
 
@@ -257,6 +262,10 @@ Status GraphBuilder::BuildForUnknownShapeGraph(ComputeGraphPtr &comp_graph, GeMo
   return SUCCESS;
 }
 
+Status GraphBuilder::BuildForHostCpuGraph(ComputeGraphPtr &comp_graph, GeModelPtr &ge_model_ptr, uint64_t session_id) {
+  return BuildForUnknownShapeGraph(comp_graph, ge_model_ptr, session_id);
+}
+
 Status GraphBuilder::BuildForDynamicShapeGraph(ComputeGraphPtr &comp_graph,
                                                std::vector<SubGraphInfoPtr> &subgraph_ptr_list,
                                                GeRootModelPtr &ge_root_model_ptr, GeModelPtr &ge_model_ptr,

src/ge/graph/build/graph_builder.h

@@ -63,10 +63,12 @@ class GraphBuilder {
   Status BuildForDynamicShapeGraph(ComputeGraphPtr &comp_graph, std::vector<SubGraphInfoPtr> &subgraph_ptr_list,
                                    GeRootModelPtr &ge_root_model_ptr, GeModelPtr &ge_model_ptr,
                                    uint64_t session_id = INVALID_SESSION_ID);
-  Status BuildForKnownShapeGraph(ComputeGraphPtr &comp_graph, std::vector<SubGraphInfoPtr> &subgraph_ptr_list,
+  Status BuildForKnownShapeGraph(ComputeGraphPtr &comp_graph, std::vector<SubGraphInfoPtr> &subgraph_list,
                                  GeModelPtr &ge_model_ptr, uint64_t session_id = INVALID_SESSION_ID);
   Status BuildForUnknownShapeGraph(ComputeGraphPtr &comp_graph, GeModelPtr &ge_model_ptr,
                                    uint64_t session_id = INVALID_SESSION_ID);
+  Status BuildForHostCpuGraph(ComputeGraphPtr &comp_graph, GeModelPtr &ge_model_ptr,
+                              uint64_t session_id = INVALID_SESSION_ID);
   int build_mode_;
 
   std::map<std::string, int> stream_max_parallel_num_;

src/ge/graph/build/memory/block_mem_assigner.cc

@@ -745,6 +745,23 @@ bool BlockMemAssigner::IsContinuousOutput(const NodePtr &n) {
   return false;
 }
 
+bool BlockMemAssigner::IsZeroCopyBlock(const NodePtr &node, bool continuous) {
+  if (NodeUtils::IsDynamicShape(node)) {
+    return ((node->GetType() == DATA_TYPE) && !continuous) || (node->GetType() == NETOUTPUT);
+  }
+
+  if ((node->GetType() == DATA_TYPE) && !continuous) {
+    return !node->GetOpDesc()->HasAttr(ATTR_NAME_PARENT_NODE_INDEX);
+  }
+
+  if (node->GetType() == NETOUTPUT) {
+    const auto &owner = node->GetOwnerComputeGraph();
+    return owner->GetParentGraph() == nullptr;
+  }
+
+  return false;
+}
+
 MemoryBlock *BlockMemAssigner::ApplyMemory(size_t block_size, size_t real_size, size_t no_align_size,
                                            MemoryType mem_type, const NodePtr &n, uint32_t out_index,
                                            const vector<bool> &workspace_reuse_flag, const bool is_op_reuse_mem,
@@ -793,9 +810,7 @@ MemoryBlock *BlockMemAssigner::ApplyMemory(size_t block_size, size_t real_size,
   GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(block == nullptr, return nullptr, "new an object failed.");
 
   // Data and netoutput need zero copy block
-  if ((node_op_desc->GetType() == DATA_TYPE && !continuous) || (node_op_desc->GetType() == NETOUTPUT)) {
-    block->is_zero_copy_ = true;
-  }
+  block->is_zero_copy_ = IsZeroCopyBlock(n, continuous);
 
   block->Init(real_size, mem_type, n, out_index, no_align_size);
   block->stream_id_ = node_op_desc->GetStreamId();
@@ -970,6 +985,14 @@ bool IsAtomicOutputMemory(const ge::NodePtr &node, uint32_t output_index, bool i
   return false;
 }
 
+bool IsKnownSubgraphData(const NodePtr &node) {
+  if (NodeUtils::IsDynamicShape(node)) {
+    return false;
+  }
+
+  return node->GetOpDesc()->HasAttr(ATTR_NAME_PARENT_NODE_INDEX);
+}
+
 void BlockMemAssigner::ReleaseMemory(MemoryBlock *to_release, vector<MemoryBlock *> &reusable_memory) {
   GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(to_release == nullptr, return, "Input parameter to_release is null.");
   GE_CHK_TRUE_EXEC_INFO(to_release->ref_count_ <= 0, return, "Release memory");
@@ -1092,7 +1115,7 @@ Status BlockMemAssigner::AssignOutputMemoryWithReuse(const NodePtr &node, vector
   (void)ge::AttrUtils::GetBool(op_desc, ATOMIC_ATTR_IS_ATOMIC_NODE, is_atomic);
   // Allocate memory for the current node and release node memory of the same size in the workspace
   GE_IF_BOOL_EXEC(ge_disable_reuse_mem_env_ != "1",
-                  ReleaseMemorys(stream_workspace_blocks_[stream_id], reusable_blocks_[stream_id]);)
+                  ReleaseMemorys(stream_workspace_blocks_[stream_id], reusable_blocks_[stream_id]));
   if (IsContinuousOutput(node)) {
     (void)ApplyContinuousMemory(node, ranges, is_op_reuse_mem_);
     return SUCCESS;
@@ -1118,6 +1141,7 @@ Status BlockMemAssigner::AssignOutputMemoryWithReuse(const NodePtr &node, vector
       out_node_set_continuous_input = IsOutNodeSetContinuousInput(node, i, peer_name, peer_input_index);
       no_need_assign_memory = IsAtomicOutputMemory(node, i, is_atomic, out_node_set_continuous_input);
     }
+    no_need_assign_memory = (no_need_assign_memory || IsKnownSubgraphData(node));
     if (no_need_assign_memory) {
       zero_memory_list_.emplace_back(node, kOutput, i, false);
       continue;
@@ -1474,8 +1498,8 @@ void SetOffsetSize(const NodeTypeIndex &node_type, const MemoryBlock *block, siz
       return;
     }
 
-    if ((op_desc->GetType() == DATA) || (op_desc->GetType() == AIPP_DATA_TYPE) || (op_desc->GetType() == MULTISHAPE) ||
-        (op_desc->GetType() == NETOUTPUT)) {
+    static const set<string> kSetOffsetTypes = {DATA_TYPE, AIPP_DATA_TYPE, MULTISHAPE, NETOUTPUT};
+    if ((kSetOffsetTypes.count(op_desc->GetType()) > 0) && !IsKnownSubgraphData(node_type.node)) {
       if ((output_list[node_type.index] == kInvalidOffset) || (output_list[node_type.index] < offset)) {
         output_list.at(node_type.index) = offset;
       }

src/ge/graph/build/memory/block_mem_assigner.h

@@ -352,6 +352,8 @@ class BlockMemAssigner : public MemAssigner {
 
   void AssignContinuousBlocks();
 
+  bool IsZeroCopyBlock(const NodePtr &node, bool continuous);
+
   bool IsOutNodeSetContinuousInput(const NodePtr &n, uint32_t out_index, std::string &peer_name,
                                    uint32_t &peer_input_index);
 

src/ge/graph/build/memory/graph_mem_assigner.cc

@@ -1227,6 +1227,18 @@ ge::Status GraphMemoryAssigner::SetInputOffset() {
   return ge::SUCCESS;
 }
 
+NodePtr GraphMemoryAssigner::GetKnownInputNode(const NodePtr &node) const {
+  if (!node->GetOpDesc()->HasAttr(ATTR_NAME_PARENT_NODE_INDEX)) {
+    return node;
+  }
+
+  if (NodeUtils::IsDynamicShape(node)) {
+    return node;
+  }
+
+  return NodeUtils::GetParentInput(node);
+}
+
 ge::Status GraphMemoryAssigner::UpdateConstArgsOffset(const NodePtr &node, vector<int64_t> &input_list) const {
   uint32_t parent_index = 0;
   if (!AttrUtils::GetInt(node->GetOpDesc(), ATTR_NAME_PARENT_NODE_INDEX, parent_index)) {
@@ -1235,13 +1247,29 @@ ge::Status GraphMemoryAssigner::UpdateConstArgsOffset(const NodePtr &node, vecto
 
   // Subgraph Data Node, check for constant input.
   std::string op_type;
-  NodePtr in_node = NodeUtils::GetParentInput(node);
-  if (!NodeUtils::GetConstOpType(in_node, op_type)) {
-    return SUCCESS;  // not constant input.
+  const auto &in_node = NodeUtils::GetParentInput(node);
+  if (NodeUtils::GetConstOpType(in_node, op_type)) {
+    input_list = in_node->GetOpDesc()->GetOutputOffset();
+    node->GetOpDesc()->SetOutputOffset(input_list);  // Set Data output same as const output.
+    return SUCCESS;                                  // Constant input.
+  }
+
+  // Memory allocated for dynamic shape subgraph Data.
+  if (NodeUtils::IsDynamicShape(node)) {
+    return SUCCESS;
+  }
+
+  const auto &owner = node->GetOwnerComputeGraph();
+  const auto &parent_desc = owner->GetParentNode()->GetOpDesc();
+  const auto parent_inputs = parent_desc->GetInputOffset();
+  if (parent_inputs.size() <= parent_index) {
+    GELOGE(FAILED, "Get Parent input offset failed, node: %s, input size: %zu, parent index: %u",
+           node->GetName().c_str(), parent_inputs.size(), parent_index);
+    return FAILED;
   }
 
-  vector<int64_t> const_input_list = in_node->GetOpDesc()->GetOutputOffset();
-  node->GetOpDesc()->SetOutputOffset(const_input_list);  // Set Data output same as const output.
+  input_list = {parent_inputs[parent_index]};
+  node->GetOpDesc()->SetOutputOffset(input_list);  // Set Data output same as parent input.
   return SUCCESS;
 }
 
@@ -1287,7 +1315,8 @@ ge::Status GraphMemoryAssigner::UpdateOpInputOffset(const NodePtr &node, vector<
                input_list.back());
       } else {
         int64_t output_offset = output_list.at(peer_out_anchor->GetIdx());
-        if (peer_out_anchor->GetOwnerNode()->GetType() == CONSTANT) {
+        const auto &in_node = GetKnownInputNode(peer_out_anchor->GetOwnerNode());
+        if (in_node->GetType() == CONSTANT) {
           GeTensorDesc tensor_desc = tmp_op_desc->GetInputDesc(input_index);
           GE_CHK_STATUS(TensorUtils::GetDataOffset(tensor_desc, output_offset));
         }

src/ge/graph/build/memory/graph_mem_assigner.h

@@ -181,6 +181,8 @@ class GraphMemoryAssigner {
 
   ge::Status UpdateConstArgsOffset(const NodePtr &node, vector<int64_t> &input_list) const;
 
+  NodePtr GetKnownInputNode(const NodePtr &node) const;
+
   MemoryOffsetList memory_offset_;
   ge::ComputeGraphPtr compute_graph_;
   HybridMemAssignerPtr mem_assigner_;

src/ge/graph/build/model_builder.cc

@@ -182,38 +182,26 @@ void ModelBuilder::SetInputIsConst(const ge::NodePtr &n) {
   for (size_t i = 0; i < is_input_const.size(); i++) {
     is_input_const[i] = false;
   }
+
+  std::string const_type;
   auto in_data_anchors = n->GetAllInDataAnchors();
   for (size_t index = 0; index < in_data_anchors.size(); index++) {
     auto in_data_anchor = in_data_anchors.at(index);
     const auto &peer_out_anchor = in_data_anchor->GetPeerOutAnchor();
     GE_IF_BOOL_EXEC(peer_out_anchor == nullptr, continue);
     const auto &src_node = peer_out_anchor->GetOwnerNode();
-    if (src_node->GetType() == CONSTANT) {
+    if (!NodeUtils::GetConstOpType(src_node, const_type)) {
+      continue;
+    }
+
+    if (const_type == CONSTANT) {
       if (!SetInputConst(node_op_desc, src_node, index, is_input_const)) {
         return;
       }
-    } else if (src_node->GetType() == CONSTANTOP) {
+    } else {
       if ((index < is_input_const.size()) && is_input_const[index]) {
         is_input_const[index] = false;
       }
-    } else if (src_node->GetType() == DATA) {
-      uint32_t parent_index = 0;
-      if (!AttrUtils::GetInt(src_node->GetOpDesc(), ATTR_NAME_PARENT_NODE_INDEX, parent_index)) {
-        continue;
-      }
-
-      // Subgraph Data Node, check for constant input.
-      std::string op_type;
-      const NodePtr in_node = NodeUtils::GetParentInput(src_node);
-      if (!NodeUtils::GetConstOpType(in_node, op_type)) {
-        continue;  // not constant input.
-      }
-
-      if (op_type == CONSTANT) {
-        if (!SetInputConst(node_op_desc, in_node, index, is_input_const)) {
-          return;
-        }
-      }
     }
   }
 

src/ge/graph/build/stream_allocator.cc

@@ -16,6 +16,7 @@
 
 #include "graph/build/stream_allocator.h"
 #include <memory>
+#include <algorithm>
 #include "common/ge/ge_util.h"
 #include "framework/common/debug/ge_log.h"
 #include "framework/common/fmk_error_codes.h"
@@ -374,8 +375,8 @@ Status StreamAllocator::InsertOneEventInTwoNodes(const NodePtr &cur_node, const
     return SUCCESS;
   }
 
-  if ((cur_node->GetType() == ENTER) || (cur_node->GetType() == REFENTER)) {
-    GELOGD("No need to insert event after enter_node %s.", cur_node->GetName().c_str());
+  if (((cur_node->GetType() == ENTER) || (cur_node->GetType() == REFENTER)) && (next_node->GetType() != STREAMACTIVE)) {
+    GELOGD("No need to insert event between %s and %s.", cur_node->GetName().c_str(), next_node->GetName().c_str());
     return SUCCESS;
   }
 
@@ -721,6 +722,7 @@ Status StreamAllocator::SplitStreams(vector<set<int64_t>> &split_streams) {
       GELOGE(FAILED, "SplitStreams:streamid(%ld) > last_stream_id(%ld)", stream_id, last_stream_id);
       return FAILED;
     }
+    bool is_stream_first_node = (stream_node_num_vec[stream_id] == 0);
     AddNodeNum(cur_node, stream_node_num_vec[stream_id]);
     stream_2_nodes_map[stream_id].push_back(cur_node);
     // The maximum number of tasks per stream.
@@ -737,7 +739,7 @@ Status StreamAllocator::SplitStreams(vector<set<int64_t>> &split_streams) {
       stream_continuous_2_nodes_map[continuous_stream_label].push_back(cur_node);
     }
     // Split the stream if it exceeds the maximum number of nodes in the stream.
-    if (NeedSpiltNewStream(stream_node_num_vec[stream_id], max_node_num_one_stream, op_desc)) {
+    if (NeedSpiltNewStream(stream_node_num_vec[stream_id], max_node_num_one_stream, op_desc, is_stream_first_node)) {
       last_stream_id++;
       GELOGI(
         "stream_node_num_vec[%ld]= %ld > max_node_num_one_stream : %ld, "
@@ -801,7 +803,11 @@ Status StreamAllocator::SplitStreams(vector<set<int64_t>> &split_streams) {
 }
 
 bool StreamAllocator::NeedSpiltNewStream(int64_t stream_node_num, int64_t max_node_num_one_stream,
-                                         const OpDescPtr &op_desc) const {
+                                         const OpDescPtr &op_desc, bool is_stream_first_node) const {
+  if (is_stream_first_node) {
+    GELOGD("First node of stream does not need to split new stream");
+    return false;
+  }
   const set<string> label_op_types({LABELSET, LABELGOTO, LABELGOTOEX, LABELSWITCH, LABELSWITCHBYINDEX});
   bool is_first_active_node = false;
   (void)AttrUtils::GetBool(op_desc, ATTR_NAME_SUBGRAPH_FIRST_ACTIVE, is_first_active_node);
@@ -1019,6 +1025,18 @@ Status StreamAllocator::SetActiveStreamsForLoop() {
       loop_active_streams.emplace_back(static_cast<uint32_t>(stream_id));
     }
   }
+  map<int64_t, NodePtr> stream_id_to_last_node;
+  set<int64_t> streams_skip_iterator_event;
+  for (const auto &node : whole_graph_->GetNodes(whole_graph_->GetGraphUnknownFlag())) {
+    int64_t stream_id = node->GetOpDesc()->GetStreamId();
+    if (find(loop_active_streams.begin(), loop_active_streams.end(), stream_id) != loop_active_streams.end()) {
+      stream_id_to_last_node[stream_id] = node;
+      // last node in stream which has streamswitch or IF may be not execute, it will cause block if add event on them
+      if (node->GetOpDesc()->GetType() == STREAMSWITCH) {
+        streams_skip_iterator_event.insert(stream_id);
+      }
+    }
+  }
   // Set the stream that needs to be activated
   for (const auto &node : whole_graph_->GetNodes(whole_graph_->GetGraphUnknownFlag())) {
     GE_CHECK_NOTNULL(node->GetOpDesc());
@@ -1031,7 +1049,31 @@ Status StreamAllocator::SetActiveStreamsForLoop() {
                          GELOGE(FAILED, "SetListInt failed.");
                          return FAILED);
         for (const auto &stream_id : loop_active_streams) {
-          GELOGI("Active stream %u for node: %s", stream_id, node->GetName().c_str());
+          GELOGI("Active stream %u for node: %s.", stream_id, node->GetName().c_str());
+        }
+
+        // In switch group optimze case, some data input branch may exec slowly.
+        // when condition input branch judge false and some switch has no false branch,
+        // In this condition, data branch has no synchronize point,
+        // it may cause some stream actived by iterator next step when this stream still alive.
+        // If above situation happen, active message will lose, cause process block in next iteration.
+        // In order to avoid this abnormal happen,
+        // add event between each last node and iterator active node in target active stream
+        GELOGI("there are %zu next iterator target streams has streamswitch node.", streams_skip_iterator_event.size());
+        for (auto iter : stream_id_to_last_node) {
+          if (streams_skip_iterator_event.find(iter.first) != streams_skip_iterator_event.end()) {
+            GELOGI("skip stream %ld which has streamswitch node when add event to next iterator active node",
+                   iter.first);
+            continue;
+          }
+          if (iter.second->GetOwnerComputeGraph()->GetParentGraph() != nullptr) {
+            GELOGI("skip stream %ld which last node in subgraph when add event to next iterator active node",
+                   iter.first);
+            continue;
+          }
+          AddSendEventId(iter.second, event_num_);
+          AddRecvEventId(node, event_num_);
+          event_num_++;
         }
 
         break;
@@ -1132,7 +1174,7 @@ Status StreamAllocator::InsertSyncEventNodes() {
         return status;
       }
 
-      GELOGI("Insert recv event %u before node: %s", event_id, node->GetName().c_str());
+      GELOGI("Insert recv event %u before node: %s.", event_id, node->GetName().c_str());
     }
 
     // Add the node corresponding to the send event
@@ -1160,7 +1202,7 @@ Status StreamAllocator::InsertSyncEventNodes() {
         return status;
       }
 
-      GELOGI("Insert send event %u after node: %s", event_id, node->GetName().c_str());
+      GELOGI("Insert send event %u after node: %s.", event_id, node->GetName().c_str());
     }
   }
 

src/ge/graph/build/stream_allocator.h

@@ -58,7 +58,8 @@ class StreamAllocator {
   bool IsActiveAfterNextIteration(const NodePtr &active_node_ptr) const;
 
   Status SplitStreams(std::vector<std::set<int64_t>> &split_streams);
-  bool NeedSpiltNewStream(int64_t stream_node_num, int64_t max_node_num_one_stream, const OpDescPtr &op_desc) const;
+  bool NeedSpiltNewStream(int64_t stream_node_num, int64_t max_node_num_one_stream, const OpDescPtr &op_desc,
+                          bool is_stream_first_node) const;
 
   Status UpdateActiveStreams(const std::vector<std::set<int64_t>> &split_streams);
   void UpdateLabelStreams(const std::vector<std::set<int64_t>> &split_streams);

src/ge/graph/build/task_generator.cc

@@ -95,8 +95,8 @@ Status TaskGenerator::GetTaskInfo(Model &model, ComputeGraphPtr &graph, uint64_t
                    GELOGE(FAILED, "SetListStr failed.");
                    return FAILED);
 
-  GELOGI("Call GenerateTask Success, task_def_list.size:%zu, op_name_map.size:%zu", task_def_list.size(),
-         op_name_map.size());
+  GELOGI("GenerateTask Success, task list:%zu, op map:%zu, logic mem base:%p, logic weight base:%p, logic var base:%p",
+         task_def_list.size(), op_name_map.size(), run_context.dataMemBase, run_context.weightMemBase, var_mem_base_);
 
   // Init and serialize model_task_def
   ModelTaskDef model_task_def;
@@ -260,7 +260,7 @@ Status TaskGenerator::GenerateTask(RunContext &run_context, ComputeGraphPtr &gra
   int64_t group_key;
   uint32_t node_index = 0;
   rtStream_t stream = nullptr;
-  bool is_unknown_shape = graph->GetGraphUnknownFlag();
+  bool is_unknown_shape = graph->GetGraphUnknownFlag() || GetContext().GetHostExecFlag();
   if (is_unknown_shape) {
     GE_CHK_STATUS_RET(SetUnknownShapeStream(run_context, stream), "Set unknown shape stream failed.");
   }
@@ -479,7 +479,12 @@ Status TaskGenerator::UpdateAnchorStatus(const NodePtr &node) {
         GELOGE(INTERNAL_ERROR, "AnchorUtils::SetStatus failed.");
         return INTERNAL_ERROR;
       }
-    } else if (peer_anchor->GetOwnerNode()->GetType() == CONSTANT) {
+      continue;
+    }
+
+    std::string const_type;
+    bool is_const = NodeUtils::GetConstOpType(peer_anchor->GetOwnerNode(), const_type);
+    if (is_const && (const_type == CONSTANT)) {
       if (AnchorUtils::SetStatus(anchor, ANCHOR_CONST) != GRAPH_SUCCESS) {
         GELOGE(INTERNAL_ERROR, "AnchorUtils::SetStatus failed.");
         return INTERNAL_ERROR;

src/ge/graph/common/transop_util.cc

@@ -17,9 +17,13 @@
 #include "graph/common/transop_util.h"
 
 #include "common/types.h"
+#include "graph/utils/type_utils.h"
+#include "framework/common/debug/ge_log.h"
 
 namespace {
 const int kInvalidTransopDataIndex = -1;
+const int kTransOpOutIndex = 0;
+std::map<ge::DataType, ge::DataType> precision_loss_transfer_map = {{ge::DT_FLOAT, ge::DT_BOOL}};
 }  // namespace
 
 namespace ge {
@@ -60,4 +64,20 @@ int TransOpUtil::GetTransOpDataIndex(const std::string &type) {
   }
   return kInvalidTransopDataIndex;
 }
+
+bool TransOpUtil::CheckPrecisionLoss(const ge::NodePtr &src_node) {
+  auto idx = TransOpUtil::GetTransOpDataIndex(src_node);
+  auto input_desc = src_node->GetOpDesc()->GetInputDesc(idx);
+  auto output_desc = src_node->GetOpDesc()->GetOutputDesc(kTransOpOutIndex);
+  auto src_dtype = input_desc.GetDataType();
+  auto dst_dtype = output_desc.GetDataType();
+  auto iter = precision_loss_transfer_map.find(src_dtype);
+  if (iter != precision_loss_transfer_map.end() && iter->second == dst_dtype) {
+    GELOGW("Node %s transfer data type from %s to %s ,it will cause precision loss. ignore pass.",
+           src_node->GetName().c_str(), TypeUtils::DataTypeToSerialString(src_dtype).c_str(),
+           TypeUtils::DataTypeToSerialString(dst_dtype).c_str());
+    return false;
+  }
+  return true;
+}
 }  // namespace ge

src/ge/graph/common/transop_util.h

@@ -33,6 +33,8 @@ class GE_FUNC_HOST_VISIBILITY GE_FUNC_DEV_VISIBILITY TransOpUtil {
 
   static int GetTransOpDataIndex(const std::string &type);
 
+  static bool CheckPrecisionLoss(const NodePtr &src_node);
+
  private:
   TransOpUtil();
 

src/ge/graph/execute/graph_execute.cc

@@ -519,6 +519,25 @@ Status GraphExecutor::GetCombinedDynamicDims(uint32_t model_id, std::vector<std:
   return SUCCESS;
 }
 
+///
+/// @ingroup ge
+/// @brief Get user designate shape order
+/// @param [in] model_id
+/// @param [out] user_input_shape_order
+/// @return execute result
+///
+ge::Status GraphExecutor::GetUserDesignateShapeOrder(uint32_t model_id,
+                                                     std::vector<std::string> &user_input_shape_order) {
+  auto model_manager = ge::ModelManager::GetInstance();
+  GE_CHECK_NOTNULL(model_manager);
+  Status ret = model_manager->GetUserDesignateShapeOrder(model_id, user_input_shape_order);
+  if (ret != SUCCESS) {
+    GELOGE(ret, "GetUserDesignateShapeOrder failed.");
+    return ret;
+  }
+  return SUCCESS;
+}
+
 Status GraphExecutor::GetCurShape(const uint32_t model_id, std::vector<int64_t> &batch_info, int32_t &dynamic_type) {
   auto model_manager = ge::ModelManager::GetInstance();
   GE_CHECK_NOTNULL(model_manager);
@@ -570,7 +589,7 @@ Status GraphExecutor::GetAIPPInfo(uint32_t model_id, uint32_t index, AippConfigI
   GE_CHECK_NOTNULL(model_manager);
   Status ret = model_manager->GetAIPPInfo(model_id, index, aipp_info);
   if (ret != SUCCESS) {
-    GELOGE(ret, "GetAIPPInfo failed.");
+    GELOGW("GetAIPPInfo is not success.");
     return ret;
   }
 
@@ -602,4 +621,16 @@ Status GraphExecutor::GetAllAippInputOutputDims(uint32_t model_id, uint32_t inde
 
   return SUCCESS;
 }
+
+Status GraphExecutor::GetOpDescInfo(uint32_t device_id, uint32_t stream_id, uint32_t task_id,
+                                    OpDescInfo &op_desc_info) {
+  auto model_manager = ge::ModelManager::GetInstance();
+  GE_CHECK_NOTNULL(model_manager);
+  Status ret = model_manager->GetOpDescInfo(device_id, stream_id, task_id, op_desc_info);
+  if (ret != SUCCESS) {
+    GELOGE(ret, "GetOpDescInfo failed.");
+    return ret;
+  }
+  return SUCCESS;
+}
 }  // namespace ge

src/ge/graph/execute/graph_execute.h

@@ -95,6 +95,15 @@ class GraphExecutor {
   ///
   static Status GetCombinedDynamicDims(uint32_t model_id, std::vector<std::vector<int64_t>> &batch_info);
 
+  ///
+  /// @ingroup ge
+  /// @brief Get user designate shape order
+  /// @param [in] model_id
+  /// @param [out] user_input_shape_order
+  /// @return execute result
+  ///
+  static Status GetUserDesignateShapeOrder(uint32_t model_id, std::vector<std::string> &user_input_shape_order);
+
   static Status GetCurShape(const uint32_t model_id, std::vector<int64_t> &batch_info, int32_t &dynamic_type);
 
   static Status GetModelAttr(uint32_t model_id, std::vector<string> &dynamic_output_shape_info);
@@ -107,6 +116,8 @@ class GraphExecutor {
   static Status GetAllAippInputOutputDims(uint32_t model_id, uint32_t index, std::vector<InputOutputDims> &input_dims,
                                           std::vector<InputOutputDims> &output_dims);
 
+  static Status GetOpDescInfo(uint32_t device_id, uint32_t stream_id, uint32_t task_id, OpDescInfo &op_desc_info);
+
  private:
   Status PrepareInputData(const std::vector<GeTensor> &input_tensor, InputData &graph_input_data,
                           OutputData &graph_output_data, std::vector<InputOutputDescInfo> &output_desc);

src/ge/graph/load/new_model_manager/aipp_utils.cc

@@ -38,7 +38,9 @@ namespace ge {
 
 Status AippUtils::ConvertAippParams2AippInfo(domi::AippOpParams *aipp_params, AippConfigInfo &aipp_info) {
   GE_CHECK_NOTNULL(aipp_params);
+  AIPP_CONVERT_TO_AIPP_INFO(aipp_mode);
   AIPP_CONVERT_TO_AIPP_INFO(input_format);
+  AIPP_CONVERT_TO_AIPP_INFO(related_input_rank);
   AIPP_CONVERT_TO_AIPP_INFO(src_image_size_w);
   AIPP_CONVERT_TO_AIPP_INFO(src_image_size_h);
   AIPP_CONVERT_TO_AIPP_INFO(crop);
@@ -85,6 +87,8 @@ Status AippUtils::ConvertAippParams2AippInfo(domi::AippOpParams *aipp_params, Ai
   AIPP_CONVERT_TO_AIPP_INFO_WITH_INDEX(var_reci_chn_1, 0);
   AIPP_CONVERT_TO_AIPP_INFO_WITH_INDEX(var_reci_chn_2, 0);
   AIPP_CONVERT_TO_AIPP_INFO_WITH_INDEX(var_reci_chn_3, 0);
+  AIPP_CONVERT_TO_AIPP_INFO(support_rotation);
+  AIPP_CONVERT_TO_AIPP_INFO(max_src_image_size);
   return SUCCESS;
 }
 }  // namespace ge

src/ge/graph/load/new_model_manager/data_dumper.cc

@@ -171,6 +171,44 @@ void DataDumper::SaveOpDebugId(uint32_t task_id, uint32_t stream_id, void *op_de
   is_op_debug_ = is_op_debug;
 }
 
+void DataDumper::SaveDumpOpInfo(const RuntimeParam &model_param, const OpDescPtr &op, uint32_t task_id,
+                                uint32_t stream_id) {
+  GELOGD("Start SaveDumpOpInfo of task_id: %u, stream_id: %u", task_id, stream_id);
+  OpDescInfo op_desc_info;
+  op_desc_info.op_name = op->GetName();
+  op_desc_info.task_id = task_id;
+  op_desc_info.stream_id = stream_id;
+  for (size_t i = 0; i < op->GetInputsSize(); ++i) {
+    GeTensorDesc input_desc = op->GetInputDesc(i);
+    op_desc_info.input_format.emplace_back(input_desc.GetFormat());
+    op_desc_info.input_shape.emplace_back(input_desc.GetShape().GetDims());
+    op_desc_info.input_data_type.emplace_back(input_desc.GetDataType());
+  }
+  for (size_t j = 0; j < op->GetOutputsSize(); ++j) {
+    GeTensorDesc output_desc = op->GetOutputDesc(j);
+    op_desc_info.output_format.emplace_back(output_desc.GetFormat());
+    op_desc_info.output_shape.emplace_back(output_desc.GetShape().GetDims());
+    op_desc_info.output_data_type.emplace_back(output_desc.GetDataType());
+  }
+  op_desc_info.input_addrs = ModelUtils::GetInputDataAddrs(model_param, op);
+  op_desc_info.output_addrs = ModelUtils::GetOutputDataAddrs(model_param, op);
+
+  op_desc_info_.emplace_back(op_desc_info);
+}
+
+bool DataDumper::GetOpDescInfo(uint32_t stream_id, uint32_t task_id, OpDescInfo &op_desc_info) const {
+  GELOGI("There are %zu op need to dump.", op_desc_info_.size());
+  for (size_t index = 0; index < op_desc_info_.size(); ++index) {
+    OpDescInfo dump_op_info = op_desc_info_.at(index);
+    if (dump_op_info.task_id == task_id && dump_op_info.stream_id == stream_id) {
+      GELOGI("find exception op of task_id: %u, stream_id: %u.", task_id, stream_id);
+      op_desc_info = dump_op_info;
+      return true;
+    }
+  }
+  return false;
+}
+
 void DataDumper::SaveDumpTask(uint32_t task_id, uint32_t stream_id, const std::shared_ptr<OpDesc> &op_desc,
                               uintptr_t args) {
   if (op_desc == nullptr) {
@@ -325,17 +363,24 @@ Status DataDumper::DumpOutputWithTask(const InnerDumpInfo &inner_dump_info, aicp
     // check dump output tensor desc is redirected by attr ATTR_DATA_DUMP_REF
     if (AttrUtils::GetStr(&output_desc, ATTR_DATA_DUMP_REF, node_name_index)) {
       GE_CHK_STATUS_RET(DumpRefOutput(inner_dump_info, output, i, node_name_index), "DumpRefOutput failed");
+      task.mutable_output()->Add(std::move(output));
     } else {
-      GE_IF_BOOL_EXEC(
-        IsTensorDescWithSkipDumpAddrType(has_mem_type_attr, v_memory_type, i),
-        GELOGD("DumpOutputWithTask[%s] output[%zu] is l1 addr, skip it", inner_dump_info.op->GetName().c_str(), i);
-        continue;);
-
-      const auto input_size = inner_dump_info.op->GetInputsSize();
-      auto addr = inner_dump_info.args + (i + input_size) * kAddrLen;
-      GE_CHK_STATUS_RET(GenerateOutput(output, output_descs, addr, i), "Generate output failed");
+      if (IsTensorDescWithSkipDumpAddrType(has_mem_type_attr, v_memory_type, i)) {
+        GELOGI("[L1Fusion] DumpOutputWithTask[%s] output[%zu] is l1 addr.", inner_dump_info.op->GetName().c_str(), i);
+        int64_t output_size = 0;
+        if (TensorUtils::GetTensorSizeInBytes(output_descs.at(i), output_size) != SUCCESS) {
+          GELOGE(PARAM_INVALID, "Get output size failed.");
+          return PARAM_INVALID;
+        }
+        GELOGI("Get output size of l1_fusion_dump is %ld", output_size);
+        GenerateOpBuffer(output_size, task);
+      } else {
+        const auto input_size = inner_dump_info.op->GetInputsSize();
+        auto addr = inner_dump_info.args + (i + input_size) * kAddrLen;
+        GE_CHK_STATUS_RET(GenerateOutput(output, output_descs, addr, i), "Generate output failed");
+        task.mutable_output()->Add(std::move(output));
+      }
     }
-    task.mutable_output()->Add(std::move(output));
   }
   return SUCCESS;
 }
@@ -468,20 +513,38 @@ Status DataDumper::DumpInput(const InnerDumpInfo &inner_dump_info, aicpu::dump::
     // check dump input tensor desc is redirected by attr ATTR_DATA_DUMP_REF
     if (AttrUtils::GetStr(&input_descs.at(i), ATTR_DATA_DUMP_REF, node_name_index)) {
       GE_CHK_STATUS_RET(DumpRefInput(inner_dump_info, input, i, node_name_index), "DumpRefInput failed");
+      task.mutable_input()->Add(std::move(input));
       // normal dump without attr
     } else {
-      GE_IF_BOOL_EXEC(IsTensorDescWithSkipDumpAddrType(has_mem_type_attr, v_memory_type, i),
-                      GELOGD("DumpInput[%s] input[%zu] is l1 addr, skip it", inner_dump_info.op->GetName().c_str(), i);
-                      continue;);
-
-      auto addr = inner_dump_info.args + kAddrLen * i;
-      GE_CHK_STATUS_RET(GenerateInput(input, input_descs, addr, i), "Generate input failed");
+      if (IsTensorDescWithSkipDumpAddrType(has_mem_type_attr, v_memory_type, i)) {
+        GELOGI("[L1Fusion] DumpInput[%s] input[%zu] is l1 addr", inner_dump_info.op->GetName().c_str(), i);
+        int64_t input_size = 0;
+        if (AttrUtils::GetInt(input_descs.at(i), ATTR_NAME_INPUT_ORIGIN_SIZE, input_size)) {
+          GELOGI("Get aipp input size according to attr is %ld", input_size);
+        } else if (TensorUtils::GetTensorSizeInBytes(input_descs.at(i), input_size) != SUCCESS) {
+          GELOGE(PARAM_INVALID, "Get input size failed.");
+          return PARAM_INVALID;
+        }
+        GELOGI("Get input size of l1_fusion_dump is %ld", input_size);
+        GenerateOpBuffer(input_size, task);
+      } else {
+        auto addr = inner_dump_info.args + kAddrLen * i;
+        GE_CHK_STATUS_RET(GenerateInput(input, input_descs, addr, i), "Generate input failed");
+        task.mutable_input()->Add(std::move(input));
+      }
     }
-    task.mutable_input()->Add(std::move(input));
   }
   return SUCCESS;
 }
 
+void DataDumper::GenerateOpBuffer(const int64_t &size, aicpu::dump::Task &task) {
+  aicpu::dump::OpBuffer op_buffer;
+  op_buffer.set_buffer_type(aicpu::dump::BufferType::L1);
+  op_buffer.set_address(reinterpret_cast<uintptr_t>(l1_fusion_addr_));
+  op_buffer.set_size(size);
+  task.mutable_buffer()->Add(std::move(op_buffer));
+}
+
 Status DataDumper::ExecuteLoadDumpInfo(aicpu::dump::OpMappingInfo &op_mapping_info) {
   std::string proto_str;
   size_t proto_size = op_mapping_info.ByteSizeLong();
@@ -720,7 +783,7 @@ void DataDumper::PrintCheckLog(string &dump_list_key) {
   bool not_find_by_omname = model_list.find(om_name_) == model_list.end();
   bool not_find_by_modelname = model_list.find(model_name_) == model_list.end();
   dump_list_key = not_find_by_omname ? model_name_ : om_name_;
-  GELOGI("%zu op need dump in %s.", op_list_.size(), dump_list_key.c_str());
+  GELOGI("%zu op need dump in known shape model %s.", op_list_.size(), dump_list_key.c_str());
 
   if (model_list.find(DUMP_ALL_MODEL) == model_list.end()) {
     if (not_find_by_omname && not_find_by_modelname) {

src/ge/graph/load/new_model_manager/data_dumper.h

@@ -30,6 +30,7 @@
 #include "proto/op_mapping_info.pb.h"
 #include "runtime/mem.h"
 #include "task_info/task_info.h"
+#include "framework/common/ge_types.h"
 
 namespace ge {
 class DataDumper {
@@ -64,10 +65,14 @@ class DataDumper {
 
   void SetRefInfo(const std::map<OpDescPtr, void *> &ref_info) { ref_info_ = ref_info; };
 
+  void SetL1FusionAddr(void *addr) { l1_fusion_addr_ = addr; };
+
   void SetLoopAddr(void *global_step, void *loop_per_iter, void *loop_cond);
 
   void SaveDumpInput(const std::shared_ptr<Node> &node);
 
+  void SaveDumpOpInfo(const RuntimeParam &model_param, const OpDescPtr &op, uint32_t task_id, uint32_t stream_id);
+
   // args is device memory stored first output addr
   void SaveDumpTask(uint32_t task_id, uint32_t stream_id, const std::shared_ptr<OpDesc> &op_desc, uintptr_t args);
   void SaveEndGraphId(uint32_t task_id, uint32_t stream_id);
@@ -81,6 +86,7 @@ class DataDumper {
 
   void SetDumpProperties(const DumpProperties &dump_properties) { dump_properties_ = dump_properties; }
   const DumpProperties &GetDumpProperties() const { return dump_properties_; }
+  bool GetOpDescInfo(uint32_t stream_id, uint32_t task_id, OpDescInfo &op_desc_info) const;
 
  private:
   void ReleaseDevMem(void **ptr) noexcept;
@@ -100,6 +106,7 @@ class DataDumper {
   struct InnerDumpInfo;
   struct InnerInputMapping;
 
+  std::vector<OpDescInfo> op_desc_info_;
   std::vector<InnerDumpInfo> op_list_;
   uint32_t end_graph_task_id_ = 0;
   uint32_t end_graph_stream_id_ = 0;
@@ -111,6 +118,7 @@ class DataDumper {
   uintptr_t loop_cond_;
   ComputeGraphPtr compute_graph_;
   std::map<OpDescPtr, void *> ref_info_;
+  void *l1_fusion_addr_ = nullptr;
 
   uint32_t op_debug_task_id_ = 0;
   uint32_t op_debug_stream_id_ = 0;
@@ -135,6 +143,7 @@ class DataDumper {
                        const uintptr_t &addr, size_t index);
   Status GenerateOutput(aicpu::dump::Output &output, const OpDesc::Vistor<GeTensorDesc> &tensor_descs,
                         const uintptr_t &addr, size_t index);
+  void GenerateOpBuffer(const int64_t &size, aicpu::dump::Task &task);
 };
 struct DataDumper::InnerDumpInfo {
   uint32_t task_id;

src/ge/graph/load/new_model_manager/davinci_model.cc

@@ -84,6 +84,8 @@ const uint32_t kAddrLen = sizeof(void *);
 const int kDecimal = 10;
 const int kBytes = 8;
 const uint32_t kDataMemAlignSizeCompare = 64;
+const uint32_t kDumpL1FusionOpMByteSize = 2 * 1024 * 1024;
+const uint32_t kDumpFlagOfL1Fusion = 0;
 const char *const kDefaultBatchLable = "Batch_default";
 
 inline bool IsDataOp(const std::string &node_type) {
@@ -97,7 +99,6 @@ inline bool IsNoTaskAndDumpNeeded(const OpDescPtr &op_desc) {
 }  // namespace
 
 std::mutex DavinciModel::tvm_bin_mutex_;
-std::set<std::string> DavinciModel::tvm_bin_kernel_;
 
 DavinciModel::DavinciModel(int32_t priority, const std::shared_ptr<ModelListener> &listener)
     : weights_mem_base_(nullptr),
@@ -179,6 +180,10 @@ DavinciModel::~DavinciModel() {
 
       FreeFeatureMapMem();
 
+      if (l1_fusion_addr_ != nullptr) {
+        GE_CHK_RT(rtFree(l1_fusion_addr_));
+      }
+
       if (rt_model_handle_ != nullptr) {
         GE_CHK_RT(rtModelDestroy(rt_model_handle_));
         rt_model_handle_ = nullptr;
@@ -305,7 +310,7 @@ Status DavinciModel::InitModelMem(void *dev_ptr, size_t mem_size, void *weight_p
     if (weight_ptr == nullptr) {
       weights_mem_base_ = MallocWeightsMem(weights_size);
       if (weights_mem_base_ == nullptr) {
-        GELOGE(GE_EXEC_ALLOC_FEATURE_MAP_MEM_FAILED, "Alloc weight memory failed. size: %zu", weights_size);
+        GELOGE(GE_EXEC_ALLOC_WEIGHT_MEM_FAILED, "Alloc weight memory failed. size: %zu", weights_size);
         return GE_EXEC_ALLOC_WEIGHT_MEM_FAILED;
       }
       is_inner_weight_base_ = true;
@@ -367,7 +372,7 @@ void DavinciModel::InitRuntimeParams() {
   session_id_ = runtime_param_.session_id;
 
   GELOGI(
-    "InitRuntimeParams(), session_id:%u, stream_num:%lu, event_num:%u, label_num:%u, "
+    "InitRuntimeParams(), session_id:%lu, stream_num:%u, event_num:%u, label_num:%u, "
     "logic_mem_base:0x%lx, logic_weight_base:0x%lx, logic_var_base:0x%lx, "
     "memory_size:%lu, weight_size:%lu, var_size:%lu",
     runtime_param_.session_id, runtime_param_.stream_num, runtime_param_.event_num, runtime_param_.label_num,
@@ -401,6 +406,7 @@ void DavinciModel::CheckHasHcomOp() {
 ///
 Status DavinciModel::BindModelStream() {
   // Stream not in active_stream_indication_ is active stream.
+  is_stream_list_bind_ = false;
   if ((!input_queue_ids_.empty() || !output_queue_ids_.empty()) || (deploy_type_ == AICPU_DEPLOY_CROSS_THREAD)) {
     for (size_t i = 0; i < stream_list_.size(); ++i) {
       if (active_stream_indication_.count(i) == 0) {
@@ -419,7 +425,7 @@ Status DavinciModel::BindModelStream() {
       GE_CHK_RT_RET(rtModelBindStream(rt_model_handle_, stream_list_[i], RT_HEAD_STREAM));
     }
   }
-
+  is_stream_list_bind_ = true;
   return SUCCESS;
 }
 
@@ -600,6 +606,12 @@ Status DavinciModel::Init(void *dev_ptr, size_t mem_size, void *weight_ptr, size
   // create model_handle to load model
   GE_CHK_RT_RET(rtModelCreate(&rt_model_handle_, 0));
   GE_CHK_RT_RET(rtModelGetId(rt_model_handle_, &runtime_model_id_));
+  // malloc 2M for dump l1fusion op
+  GE_CHK_RT_RET(rtMalloc(&l1_fusion_addr_, kDumpL1FusionOpMByteSize, RT_MEMORY_DDR));
+
+  // send l1fusion dump addr to rts
+  GE_CHK_RT_RET(rtDumpAddrSet(rt_model_handle_, l1_fusion_addr_, kDumpL1FusionOpMByteSize, kDumpFlagOfL1Fusion));
+
   // inference will use default graph_id 0;
   runtime_param_.graph_id = compute_graph->GetGraphID();
 
@@ -748,11 +760,18 @@ Status DavinciModel::InitNodes(const ComputeGraphPtr &compute_graph) {
 
   typedef Status (DavinciModel::*OpDescCall)(const OpDescPtr &);
   static std::map<std::string, OpDescCall> op_desc_handle = {
-    {VARIABLE, &DavinciModel::InitVariable},           {CONSTANTOP, &DavinciModel::InitConstant},
-    {STREAMACTIVE, &DavinciModel::InitStreamActive},   {STREAMSWITCH, &DavinciModel::InitStreamSwitch},
-    {STREAMSWITCHN, &DavinciModel::InitStreamSwitchN}, {LABELSET, &DavinciModel::InitLabelSet},
+    {VARIABLE, &DavinciModel::InitVariable},
+    {CONSTANTOP, &DavinciModel::InitConstant},
+    {STREAMACTIVE, &DavinciModel::InitStreamActive},
+    {STREAMSWITCH, &DavinciModel::InitStreamSwitch},
+    {STREAMSWITCHN, &DavinciModel::InitStreamSwitchN},
+    {LABELSET, &DavinciModel::InitLabelSet},
+    {CASE, &DavinciModel::InitCase},
   };
 
+  GE_CHK_STATUS_RET(InitInputOutputForDynamic(compute_graph), "InitInputOutputForDynamic failed.");
+
+  map<uint32_t, OpDescPtr> data_by_index;
   auto nodes = compute_graph->GetAllNodes();
   const TBEKernelStore &tbekernel_store = ge_model_->GetTBEKernelStore();
   for (size_t i = 0; i < nodes.size(); i++) {
@@ -770,7 +789,7 @@ Status DavinciModel::InitNodes(const ComputeGraphPtr &compute_graph) {
     GE_TIMESTAMP_ADD(LoadTBEKernelBinToOpDesc);
 
     if (IsDataOp(op_desc->GetType())) {
-      if (InitDataOp(node, data_op_index) != SUCCESS) {
+      if (InitDataOp(node, data_op_index, data_by_index) != SUCCESS) {
         GELOGE(PARAM_INVALID, "Data init failed, Name: %s", op_desc->GetName().c_str());
         return PARAM_INVALID;
       }
@@ -839,21 +858,44 @@ Status DavinciModel::InitNodes(const ComputeGraphPtr &compute_graph) {
     GE_TIMESTAMP_ADD(InitTbeHandle);
   }
 
+  AdjustDataOpList(data_by_index);
   GE_TIMESTAMP_CALLNUM_END(LoadTBEKernelBinToOpDesc, "GraphLoader::LoadTBEKernelBinToOpDesc.");
   GE_TIMESTAMP_CALLNUM_END(InitTbeHandle, "GraphLoader::InitTbeHandle.");
   return SUCCESS;
 }
 
+Status DavinciModel::InitInputOutputForDynamic(const ComputeGraphPtr &compute_graph) {
+  if (!known_node_) return SUCCESS;
+  // for dynamic shape
+  auto direct_nodes = compute_graph->GetDirectNode();
+  for (size_t i = 0; i < direct_nodes.size(); i++) {
+    auto node = direct_nodes.at(i);
+    auto op_desc = node->GetOpDesc();
+    if (op_desc == nullptr) {
+      GELOGE(PARAM_INVALID, "op_desc is null.");
+      return PARAM_INVALID;
+    }
+    if (IsDataOp(op_desc->GetType())) {
+      GELOGD("init data op %s", op_desc->GetName().c_str());
+      data_op_list_.push_back(op_desc);
+    }
+    if (op_desc->GetType() == NETOUTPUT) {
+      GELOGD("init netouput op %s", op_desc->GetName().c_str());
+      output_op_list_.push_back(op_desc);
+    }
+  }
+  return SUCCESS;
+}
+
 /// @ingroup ge
 /// @brief Data Op Initialize.
 /// @param [in] NodePtr: Data Op.
 /// @param [in/out] data_op_index: NetOutput addr size info.
 /// @return Status
-Status DavinciModel::InitDataOp(const NodePtr &node, uint32_t &data_op_index) {
+Status DavinciModel::InitDataOp(const NodePtr &node, uint32_t &data_op_index, map<uint32_t, OpDescPtr> &data_by_index) {
   // op_desc Checked by Init: Data, valid.
   auto op_desc = node->GetOpDesc();
   if (known_node_) {
-    data_op_list_.push_back(op_desc);
     return SUCCESS;
   }
   uint32_t parent_index = 0;  // Ignore subgraph Data Node.
@@ -885,6 +927,7 @@ Status DavinciModel::InitDataOp(const NodePtr &node, uint32_t &data_op_index) {
     return PARAM_INVALID;
   }
   new_input_data_info_[data_index] = zero_copy_offset;
+  data_by_index[data_index] = op_desc;
 
   for (size_t index = 0; index < virtual_addr_list.size(); ++index) {
     void *addr = virtual_addr_list.at(index);
@@ -904,6 +947,24 @@ Status DavinciModel::InitDataOp(const NodePtr &node, uint32_t &data_op_index) {
   return SUCCESS;
 }
 
+///
+/// @ingroup ge
+/// @brief Sort Data op list by index.
+/// @param [in] data_by_index: map of Data Op.
+/// @return
+///
+void DavinciModel::AdjustDataOpList(const map<uint32_t, OpDescPtr> &data_by_index) {
+  if (data_by_index.size() != data_op_list_.size()) {
+    GELOGW("Data map size: %zu, Data list size: %zu.", data_by_index.size(), data_op_list_.size());
+    return;
+  }
+
+  data_op_list_.clear();
+  for (auto &item : data_by_index) {
+    data_op_list_.emplace_back(item.second);
+  }
+}
+
 ///
 /// @ingroup ge
 /// @brief input zero copy node Initialize.
@@ -946,7 +1007,6 @@ Status DavinciModel::InitNetOutput(const NodePtr &node) {
   auto op_desc = node->GetOpDesc();
   // excludes the function op sub graph, e.g. case,if
   if (known_node_) {
-    output_op_list_.push_back(op_desc);
     return SUCCESS;
   }
   ComputeGraphPtr owner_graph = node->GetOwnerComputeGraph();
@@ -989,9 +1049,6 @@ Status DavinciModel::InitNetOutput(const NodePtr &node) {
     new_output_data_info_[num + idx] = zero_copy_offset;
     void *addr = virtual_addr_list.at(idx);
     int64_t input_offset = input_offset_list.at(idx);
-    if (new_output_outside_addrs_.find(addr) != new_output_outside_addrs_.end()) {
-      continue;
-    }
     vector<void *> tensor_addrs;
     zero_copy_offset.SetOutputOutsideAddrs(input_offset, fusion_flag, addr, tensor_addrs);
     auto rslt = new_output_outside_addrs_.insert(std::pair<void *, ZeroCopyOffset>(addr, zero_copy_offset));
@@ -1464,6 +1521,17 @@ void DavinciModel::GetCombinedDynamicDims(std::vector<std::vector<int64_t>> &bat
   batch_info = combined_batch_info_;
 }
 
+///
+/// @ingroup ge
+/// @brief Get user designate shape order
+/// @param [out] user_input_shape_order
+/// @return None
+///
+void DavinciModel::GetUserDesignateShapeOrder(std::vector<std::string> &user_input_shape_order) const {
+  user_input_shape_order.clear();
+  user_input_shape_order = user_designate_shape_order_;
+}
+
 ///
 /// @ingroup ge
 /// @brief Get AIPP input info
@@ -1475,7 +1543,7 @@ Status DavinciModel::GetAIPPInfo(uint32_t index, AippConfigInfo &aipp_info) {
   GE_CHK_BOOL_RET_STATUS(index < data_op_list_.size(), PARAM_INVALID, "Index %u is invalid.", index);
   OpDescPtr data_op = data_op_list_[index];
   if (!data_op->HasAttr(ATTR_NAME_AIPP)) {
-    GELOGE(GE_AIPP_NOT_EXIST, "GetAIPPInfo: there is not AIPP related with index %u.", index);
+    GELOGW("GetAIPPInfo: there is not AIPP related with index %u.", index);
     return GE_AIPP_NOT_EXIST;
   }
 
@@ -1488,10 +1556,6 @@ Status DavinciModel::GetAIPPInfo(uint32_t index, AippConfigInfo &aipp_info) {
   GE_CHK_STATUS_RET(OpUtils::ConvertAippParams(aipp_attr, aipp_params.get()), "get aipp params failed");
   GELOGI("GetAIPPInfo: node data: %s, type: %s, current index: %u, current node related input rank: %u",
          data_op->GetName().c_str(), data_op->GetType().c_str(), index, aipp_params->related_input_rank());
-  if (aipp_params->aipp_mode() == domi::AippOpParams::dynamic) {
-    GELOGI("GetAIPPInfo,  dynamic Aipp is not support to query temporarily.");
-    return GE_DYNAMIC_AIPP_NOT_SUPPORT_QUERY;
-  }
 
   GE_CHK_STATUS_RET(AippUtils::ConvertAippParams2AippInfo(aipp_params.get(), aipp_info),
                     "convert aipp params to aipp config info failed");
@@ -1563,51 +1627,51 @@ Status DavinciModel::GetInputOutputDescInfoForZeroCopy(vector<InputOutputDescInf
   return SUCCESS;
 }
 
-void DavinciModel::CreateInputDimsInfo(const OpDescPtr &op_desc, Format format, InputOutputDescInfo &input) {
+void DavinciModel::SetInputDimsInfo(const vector<int64_t> &model_input_dims, Format &format,
+                                    InputOutputDescInfo &input) {
   uint32_t n, c, h, w;
   n = format == FORMAT_NHWC ? NHWC_DIM_N : NCHW_DIM_N;
   c = format == FORMAT_NHWC ? NHWC_DIM_C : NCHW_DIM_C;
   h = format == FORMAT_NHWC ? NHWC_DIM_H : NCHW_DIM_H;
   w = format == FORMAT_NHWC ? NHWC_DIM_W : NCHW_DIM_W;
 
+  if (model_input_dims.size() == static_cast<size_t>(NORMAL_TENSOR_SIZE)) {
+    input.shape_info.num = model_input_dims[n];
+    input.shape_info.height = model_input_dims[h];
+    input.shape_info.width = model_input_dims[w];
+    input.shape_info.channel = model_input_dims[c];
+  }
+  for (size_t k = 0; k < model_input_dims.size(); ++k) {
+    input.shape_info.dims.push_back(model_input_dims[k]);
+  }
+  return;
+}
+
+void DavinciModel::CreateInputDimsInfo(const OpDescPtr &op_desc, Format format, InputOutputDescInfo &input) {
   if (is_new_model_desc_ && op_desc->HasAttr(ATTR_NAME_INPUT_DIMS)) {
     // When static aipp is set, need to get the model input dims which processed by aipp
     vector<int64_t> model_input_dims;
     (void)AttrUtils::GetListInt(op_desc, ATTR_NAME_INPUT_DIMS, model_input_dims);
-    if (model_input_dims.size() == static_cast<size_t>(NORMAL_TENSOR_SIZE)) {
-      input.shape_info.num = model_input_dims[n];
-      input.shape_info.height = model_input_dims[h];
-      input.shape_info.width = model_input_dims[w];
-      input.shape_info.channel = model_input_dims[c];
-    }
-    for (size_t k = 0; k < model_input_dims.size(); ++k) {
-      input.shape_info.dims.push_back(model_input_dims[k]);
-    }
-    is_new_model_desc_ = false;
+    SetInputDimsInfo(model_input_dims, format, input);
     return;
   }
-
-  if (!op_desc->HasAttr(ATTR_MBATCH_ORIGIN_INPUT_DIMS)) {
-    if (op_desc->GetInputDescPtr(0)->GetShape().GetDimNum() == static_cast<size_t>(NORMAL_TENSOR_SIZE)) {
-      input.shape_info.num = op_desc->GetInputDescPtr(0)->GetShape().GetDim(n);
-      input.shape_info.height = op_desc->GetInputDescPtr(0)->GetShape().GetDim(h);
-      input.shape_info.width = op_desc->GetInputDescPtr(0)->GetShape().GetDim(w);
-      input.shape_info.channel = op_desc->GetInputDescPtr(0)->GetShape().GetDim(c);
-    }
-    for (size_t k = 0; k < op_desc->GetInputDescPtr(0)->GetShape().GetDimNum(); k++) {
-      input.shape_info.dims.push_back(op_desc->GetInputDescPtr(0)->GetShape().GetDim(k));
-    }
+  // judge if this data is linked dynamic aipp first, multiply batch has been considered
+  if (op_desc->HasAttr("_dynamic_aipp_input_dims")) {
+    vector<int64_t> dynamic_aipp_input_dims;
+    (void)AttrUtils::GetListInt(op_desc, "_dynamic_aipp_input_dims", dynamic_aipp_input_dims);
+    SetInputDimsInfo(dynamic_aipp_input_dims, format, input);
+    return;
   } else {
-    vector<int64_t> origin_input_dims;
-    (void)AttrUtils::GetListInt(op_desc, ATTR_MBATCH_ORIGIN_INPUT_DIMS, origin_input_dims);
-    if (origin_input_dims.size() == static_cast<size_t>(NORMAL_TENSOR_SIZE)) {
-      input.shape_info.num = origin_input_dims[n];
-      input.shape_info.height = origin_input_dims[h];
-      input.shape_info.width = origin_input_dims[w];
-      input.shape_info.channel = origin_input_dims[c];
-    }
-    for (size_t k = 0; k < origin_input_dims.size(); ++k) {
-      input.shape_info.dims.push_back(origin_input_dims[k]);
+    // judge if this data is multiply batch
+    if (!op_desc->HasAttr(ATTR_MBATCH_ORIGIN_INPUT_DIMS)) {
+      vector<int64_t> input_dims = op_desc->GetInputDescPtr(0)->GetShape().GetDims();
+      SetInputDimsInfo(input_dims, format, input);
+      return;
+    } else {
+      vector<int64_t> origin_input_dims;
+      (void)AttrUtils::GetListInt(op_desc, ATTR_MBATCH_ORIGIN_INPUT_DIMS, origin_input_dims);
+      SetInputDimsInfo(origin_input_dims, format, input);
+      return;
     }
   }
 }
@@ -1630,6 +1694,8 @@ Status DavinciModel::GetInputDescInfo(vector<InputOutputDescInfo> &input_desc, s
     formats.push_back(format);
     input_desc.push_back(input);
   }
+  // cause GetInputDescInfo called not only once, set is_new_model_desc_ to false after calc the model input dims
+  is_new_model_desc_ = false;
   return SUCCESS;
 }
 
@@ -2106,22 +2172,24 @@ Status DavinciModel::CopyOutputData(uint32_t data_id, OutputData &output_data, r
       return FAILED;
     }
 
+    if ((kind == RT_MEMCPY_DEVICE_TO_DEVICE) && (copy_only_addrs_.count(output.second.GetBasicAddr()) == 0)) {
+      continue;  // Skip: Feed by zero copy.
+    }
+
     DataBuffer &buffer = blobs[output.first];
     uint64_t mem_size = static_cast<uint64_t>(output.second.GetDataSize());
     if ((buffer.length == 0) || (mem_size == 0)) {
       GELOGI("Length of data is zero, No need copy. output tensor index=%u", output.first);
       continue;
     }
-    if (buffer.length < mem_size) {
+    if (is_dynamic_) {
+      GELOGI("No need to check output data size.");
+    } else if (buffer.length < mem_size) {
       GELOGE(FAILED, "Tensor data size=%lu, buffer size=%u", mem_size, buffer.length);
       return FAILED;
     } else if (buffer.length > mem_size) {
       GELOGW("Tensor data size=%lu, buffer size=%u", mem_size, buffer.length);
     }
-
-    if ((kind == RT_MEMCPY_DEVICE_TO_DEVICE) && (copy_only_addrs_.count(output.second.GetBasicAddr()) == 0)) {
-      continue;  // Skip: Feed by zero copy.
-    }
     uint64_t data_size = output.second.GetDataSize();
     uint64_t buffer_length = buffer.length;
     void *buffer_addr = reinterpret_cast<void *>(reinterpret_cast<uintptr_t>(buffer.data));
@@ -2564,10 +2632,12 @@ Status DavinciModel::ModelRunStop() {
 void DavinciModel::UnbindTaskSinkStream() {
   // unbinding hcom stream
   UnbindHcomStream();
-  for (size_t i = 0; i < stream_list_.size(); i++) {
-    // unbind rt_model_handle and streams
-    GE_LOGW_IF(rtModelUnbindStream(rt_model_handle_, stream_list_[i]) != RT_ERROR_NONE,
-               "Unbind stream from model failed! Index: %zu", i);
+  if (is_stream_list_bind_) {
+    for (size_t i = 0; i < stream_list_.size(); i++) {
+      // unbind rt_model_handle and streams
+      GE_LOGW_IF(rtModelUnbindStream(rt_model_handle_, stream_list_[i]) != RT_ERROR_NONE,
+                 "Unbind stream from model failed! Index: %zu", i);
+    }
   }
 
   if (is_inner_model_stream_) {
@@ -2610,11 +2680,7 @@ Status DavinciModel::CreateKnownZeroCopyMap(const vector<void *> &inputs, const
     return SUCCESS;
   }
   const vector<void *> addr_list = ModelUtils::GetInputDataAddrs(runtime_param_, output_op_list_[kDataIndex]);
-  if (outputs.size() > addr_list.size()) {
-    GELOGE(FAILED, "output data addr %u should less than output op number %u.", outputs.size(), addr_list.size());
-    return FAILED;
-  }
-  for (size_t i = 0; i < addr_list.size(); ++i) {
+  for (size_t i = 0; i < addr_list.size() && i < outputs.size(); ++i) {
     knonw_output_data_info_[addr_list[i]] = outputs[i];
     GELOGI("DavinciModel::CreateKnownZeroCopyMap output %d,v addr %p,p addr %p .", i, addr_list[i], outputs[i]);
   }
@@ -2755,19 +2821,21 @@ Status DavinciModel::DistributeTask() {
   }
 
   const auto &model_task_def = ge_model_->GetModelTaskDefPtr();
+  GELOGI("there are %zu task need to save.", task_list_.size());
   for (size_t task_index = 0; task_index < task_list_.size(); ++task_index) {
     auto &task = task_list_.at(task_index);
     GE_CHK_STATUS_RET(task->Distribute(), "Task[%zu] distribute fail", task_index);
     // for data dump
-    if (reinterpret_cast<void *>(task->GetDumpArgs()) != nullptr) {
-      auto op_index = std::max(model_task_def->task(task_index).kernel().context().op_index(),
-                               model_task_def->task(task_index).kernel_ex().op_index());
-      OpDescPtr op = GetOpByIndex(op_index);
-      if (op == nullptr) {
-        GELOGE(PARAM_INVALID, "Op index %u is null, op list size %zu.", op_index, op_list_.size());
-        return PARAM_INVALID;
-      }
+    auto op_index = std::max(model_task_def->task(task_index).kernel().context().op_index(),
+                             model_task_def->task(task_index).kernel_ex().op_index());
+    OpDescPtr op = GetOpByIndex(op_index);
+    if (op == nullptr) {
+      GELOGE(PARAM_INVALID, "Op index %u is null, op list size %zu.", op_index, op_list_.size());
+      return PARAM_INVALID;
+    }
 
+    SaveDumpOpInfo(runtime_param_, op, task->GetTaskID(), task->GetStreamId());
+    if (reinterpret_cast<void *>(task->GetDumpArgs()) != nullptr) {
       bool call_dump = GetDumpProperties().IsLayerNeedDump(name_, om_name_, op->GetName()) && task->CallSaveDumpInfo();
       if (call_dump) {
         SaveDumpTask(task->GetTaskID(), task->GetStreamId(), op, task->GetDumpArgs());
@@ -2873,7 +2941,7 @@ void DavinciModel::DisableZeroCopy(const void *addr) {
 void DavinciModel::SetZeroCopyAddr(const OpDescPtr &op_desc, const std::vector<void *> &outside_addrs, const void *info,
                                    void *args, size_t size, size_t offset) {
   // Internal call has ensured that op_desc is not nullptr
-  GELOGI("[ZCPY] SetZeroCopyAddr for %s.", op_desc->GetName().c_str());
+  GELOGD("[ZCPY] SetZeroCopyAddr for %s.", op_desc->GetName().c_str());
   size_t nums = outside_addrs.size();
   ZeroCopyTask zero_copy_task(op_desc->GetName(), static_cast<uint8_t *>(args), size);
   for (size_t i = 0; i < nums; ++i) {
@@ -2994,7 +3062,7 @@ Status DavinciModel::CopyModelData(const InputData &input_data, OutputData &outp
   }
 
   for (ZeroCopyTask &task : zero_copy_tasks_) {
-    GE_CHK_STATUS_RET(task.DistributeParam(is_async_mode_ ? rt_model_stream_ : nullptr), "[ZCPY] Update args failed.");
+    GE_CHK_STATUS_RET(task.DistributeParam(is_async_mode_, rt_model_stream_), "[ZCPY] Update args failed.");
   }
 
   output_data.index = input_data.index;
@@ -3106,7 +3174,6 @@ const char *DavinciModel::GetRegisterStub(const string &binfile, const string &s
   } else {
     binfile_key = session_graph_id + "_" + binfile;
   }
-  std::lock_guard<std::mutex> lock(tvm_bin_mutex_);
   auto it = tvm_bin_kernel_.find(binfile_key);
   if (it != tvm_bin_kernel_.end()) {
     return it->c_str();
@@ -3242,7 +3309,6 @@ void DavinciModel::StoreTbeHandle(const std::string &handle_key) {
   // Online mode FE may call rtFunctionRegister.
   TBEHandleStore &kernel_store = TBEHandleStore::GetInstance();
 
-  // Need protection of tvm_bin_mutex_.
   auto it = used_tbe_handle_map_.find(handle_key);
   if (it != used_tbe_handle_map_.end()) {
     // GE registered, increase reference.
@@ -3262,9 +3328,9 @@ void DavinciModel::StoreTbeHandle(const std::string &handle_key) {
 void DavinciModel::CleanTbeHandle() {
   TBEHandleStore &kernel_store = TBEHandleStore::GetInstance();
 
-  std::lock_guard<std::mutex> lock(tvm_bin_mutex_);
   kernel_store.EraseTBEHandle(used_tbe_handle_map_);
   used_tbe_handle_map_.clear();
+  tvm_bin_kernel_.clear();
 }
 
 ///
@@ -3315,21 +3381,26 @@ Status DavinciModel::InitStreamSwitchN(const OpDescPtr &op_desc) {
     GELOGI("StreamSwitchNOp node:%s, active_stream_id=%u.", op_desc->GetName().c_str(), active_stream_list[j]);
   }
 
-  (void)AttrUtils::GetInt(op_desc, ATTR_DYNAMIC_TYPE, dynamic_type_);
-
-  batch_info_.clear();
-  combined_batch_info_.clear();
   uint32_t batch_num = 0;
   if (!AttrUtils::GetInt(op_desc, ATTR_NAME_BATCH_NUM, batch_num)) {
     GELOGE(FAILED, "Failed to get attr ATTR_NAME_BATCH_NUM, StreamSwitchN: %s.", op_desc->GetName().c_str());
     return FAILED;
   }
 
-  for (uint32_t i = 0; i < batch_num; i++) {
+  return SetDynamicBatchInfo(op_desc, batch_num);
+}
+
+Status DavinciModel::SetDynamicBatchInfo(const OpDescPtr &op_desc, uint32_t batch_num) {
+  batch_info_.clear();
+  combined_batch_info_.clear();
+
+  (void)AttrUtils::GetInt(op_desc, ATTR_DYNAMIC_TYPE, dynamic_type_);
+  (void)AttrUtils::GetListStr(op_desc, ATTR_USER_DESIGNEATE_SHAPE_ORDER, user_designate_shape_order_);
+  for (uint32_t i = 0; i < batch_num; ++i) {
     std::vector<int64_t> batch_shape;
     const std::string attr_name = ATTR_NAME_PRED_VALUE + "_" + std::to_string(i);
     if (!AttrUtils::GetListInt(op_desc, attr_name, batch_shape)) {
-      GELOGE(FAILED, "Failed to get attr ATTR_NAME_PRED_VALUE, StreamSwitchN: %s.", op_desc->GetName().c_str());
+      GELOGE(FAILED, "Get attr ATTR_NAME_PRED_VALUE failed, Node: %s", op_desc->GetName().c_str());
       batch_info_.clear();
       return FAILED;
     }
@@ -3344,6 +3415,16 @@ Status DavinciModel::InitStreamSwitchN(const OpDescPtr &op_desc) {
   return SUCCESS;
 }
 
+Status DavinciModel::InitCase(const OpDescPtr &op_desc) {
+  uint32_t batch_num = 0;
+  if (!AttrUtils::GetInt(op_desc, ATTR_NAME_BATCH_NUM, batch_num)) {
+    GELOGI("Not multi-batch Node: %s", op_desc->GetName().c_str());
+    return SUCCESS;
+  }
+
+  return SetDynamicBatchInfo(op_desc, batch_num);
+}
+
 bool DavinciModel::IsBroadCastOpData(const ge::NodePtr &var_node) {
   for (auto out_anchor : var_node->GetAllOutDataAnchors()) {
     GE_RT_FALSE_CHECK_NOTNULL(out_anchor);
@@ -3406,12 +3487,13 @@ Status DavinciModel::NnExecute(rtStream_t stream, bool async_mode, const InputDa
   GELOGI("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.");
 
-  if (!input_data.is_dynamic_batch) {
+  is_dynamic_ = input_data.is_dynamic_batch;
+  if (!is_dynamic_) {
     zero_copy_batch_label_addrs_.clear();
   }
 
   GE_IF_BOOL_EXEC(ProfilingManager::Instance().ProfilingOn(), SetProfileTime(MODEL_PRE_PROC_START));
-  Status ret = CopyModelData(input_data, output_data, input_data.is_dynamic_batch);
+  Status ret = CopyModelData(input_data, output_data, is_dynamic_);
   GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(ret != SUCCESS, return ret, "Copy input data to model failed. model id: %u",
                                  model_id_);
 
@@ -3587,6 +3669,7 @@ void DavinciModel::SetDataDumperArgs(const ComputeGraphPtr &compute_graph) {
   data_dumper_.SetOmName(om_name_);
   data_dumper_.SetComputeGraph(compute_graph);
   data_dumper_.SetRefInfo(saved_task_addrs_);
+  data_dumper_.SetL1FusionAddr(l1_fusion_addr_);
 
   int32_t device_id = 0;
   rtError_t rt_ret = rtGetDevice(&device_id);
@@ -3627,19 +3710,9 @@ void DavinciModel::PushHcclStream(rtStream_t value) {
   all_hccl_stream_list_.push_back(value);
 }
 
-void DavinciModel::CreateHcclFollowStream(rtStream_t stream, int64_t remain_cap) {
+void DavinciModel::SaveHcclFollowStream(int64_t main_stream_id, rtStream_t stream) {
   std::lock_guard<std::mutex> lock(capacity_of_stream_mutex_);
-  capacity_of_stream_.emplace_back(make_pair(stream, remain_cap));
-}
-
-void DavinciModel::ReuseHcclFollowStream(int64_t remain_cap, int64_t &index) {
-  std::lock_guard<std::mutex> lock(capacity_of_stream_mutex_);
-  if (remain_cap == 0) {
-    capacity_of_stream_.erase(capacity_of_stream_.begin() + index);
-  } else {
-    capacity_of_stream_.at(index).second = remain_cap;
-    index++;
-  }
+  main_follow_stream_mapping_[main_stream_id].emplace_back(stream);
 }
 
 Status DavinciModel::GetComputeGraphInfo(const ComputeGraphPtr &graph, vector<ComputeGraphDescInfo> &graph_desc_info) {
@@ -3756,8 +3829,7 @@ Status DavinciModel::GetAllAippInputOutputDims(uint32_t index, std::vector<Input
       (void)TensorUtils::GetSize(*(data_op->GetInputDescPtr(kDataIndex)), data_input_size);
       GELOGD(
         "GetAllAippInputOutputDims related Data[%d]: tensor_name is %s, dim_num is %u, tensor_size: %zu, format: "
-        "%s, "
-        "data_type: %s, shape: %s .",
+        "%s, data_type: %s, shape: %s .",
         index, data_op->GetName().c_str(), data_input_desc->GetShape().GetDimNum(), data_input_size,
         TypeUtils::FormatToSerialString(data_input_desc->GetFormat()).c_str(),
         TypeUtils::DataTypeToSerialString(data_input_desc->GetDataType()).c_str(),

src/ge/graph/load/new_model_manager/davinci_model.h

@@ -184,10 +184,10 @@ class DavinciModel {
   size_t TotalMemSize() const { return runtime_param_.mem_size; }
 
   // model name
-  string Name() { return name_; }
+  string Name() const { return name_; }
 
   // om_name
-  string OmName() { return om_name_; }
+  string OmName() const { return om_name_; }
   // version
   uint32_t Version() const { return version_; }
 
@@ -268,7 +268,7 @@ class DavinciModel {
   /// @brief For TVM Op, avoid Addr Reuse.
   /// @return void*
   ///
-  static const char *GetRegisterStub(const string &tvm_binfile_key, const string &session_graph_model_id = "");
+  const char *GetRegisterStub(const string &tvm_binfile_key, const string &session_graph_model_id = "");
 
   ///
   /// @ingroup ge
@@ -299,6 +299,8 @@ class DavinciModel {
   ///
   void GetCombinedDynamicDims(std::vector<std::vector<int64_t>> &batch_info) const;
 
+  void GetUserDesignateShapeOrder(std::vector<std::string> &user_input_shape_order) const;
+
   void GetCurShape(std::vector<int64_t> &batch_info, int32_t &dynamic_type);
 
   void GetModelAttr(std::vector<std::string> &dynamic_output_shape_info);
@@ -440,6 +442,10 @@ class DavinciModel {
 
   Status SinkTimeProfile(const InputData &current_data);
 
+  void SaveDumpOpInfo(const RuntimeParam &model_param, const OpDescPtr &op, uint32_t task_id, uint32_t stream_id) {
+    data_dumper_.SaveDumpOpInfo(model_param, op, task_id, stream_id);
+  }
+
   void SaveDumpTask(uint32_t task_id, uint32_t stream_id, const std::shared_ptr<OpDesc> &op_desc, uintptr_t args) {
     data_dumper_.SaveDumpTask(task_id, stream_id, op_desc, args);
   }
@@ -449,9 +455,8 @@ class DavinciModel {
 
   DavinciModel(const DavinciModel &model) = delete;
 
-  const vector<std::pair<rtStream_t, int64_t>> &GetHcclFolowStream() { return capacity_of_stream_; }
-  void CreateHcclFollowStream(rtStream_t stream, int64_t remain_cap);
-  void ReuseHcclFollowStream(int64_t remain_cap, int64_t &index);
+  const map<int64_t, std::vector<rtStream_t>> &GetHcclFolowStream() { return main_follow_stream_mapping_; }
+  void SaveHcclFollowStream(int64_t main_stream_id, rtStream_t stream);
 
   void InitRuntimeParams();
   Status InitVariableMem();
@@ -500,6 +505,16 @@ class DavinciModel {
   void SetDumpProperties(const DumpProperties &dump_properties) { data_dumper_.SetDumpProperties(dump_properties); }
   const DumpProperties &GetDumpProperties() const { return data_dumper_.GetDumpProperties(); }
 
+  void SetMemcpyOffsetAndAddr(map<int64_t, void *> &memcpy_4g_offset_addr) {
+    memcpy_4g_offset_addr_.insert(memcpy_4g_offset_addr.begin(), memcpy_4g_offset_addr.end());
+  }
+  const map<int64_t, void *> &GetMemcpyOffsetAndAddr() const { return memcpy_4g_offset_addr_; }
+
+  bool GetOpDescInfo(uint32_t stream_id, uint32_t task_id, OpDescInfo &op_desc_info) const {
+    return data_dumper_.GetOpDescInfo(stream_id, task_id, op_desc_info);
+  }
+  Status InitInputOutputForDynamic(const ComputeGraphPtr &compute_graph);
+
  private:
   // memory address of weights
   uint8_t *weights_mem_base_;
@@ -575,6 +590,8 @@ class DavinciModel {
 
   void CreateInputDimsInfo(const OpDescPtr &op_desc, Format format, InputOutputDescInfo &input);
 
+  void SetInputDimsInfo(const vector<int64_t> &model_input_dims, Format &format, InputOutputDescInfo &input);
+
   Status GetInputDescInfo(vector<InputOutputDescInfo> &input_desc, std::vector<uint32_t> &formats);
 
   Status InitTaskInfo(domi::ModelTaskDef &modelTaskInfo);
@@ -619,7 +636,15 @@ class DavinciModel {
   /// @param [in/out] data_op_index: NetOutput addr size info.
   /// @return Status
   ///
-  Status InitDataOp(const NodePtr &node, uint32_t &data_op_index);
+  Status InitDataOp(const NodePtr &node, uint32_t &data_op_index, map<uint32_t, OpDescPtr> &data_by_index);
+
+  ///
+  /// @ingroup ge
+  /// @brief Sort Data op list by index.
+  /// @param [in] data_by_index: map of Data Op.
+  /// @return
+  ///
+  void AdjustDataOpList(const map<uint32_t, OpDescPtr> &data_by_index);
 
   ///
   /// @ingroup ge
@@ -666,6 +691,15 @@ class DavinciModel {
 
   Status InitStreamSwitchN(const OpDescPtr &op_desc);
 
+  ///
+  /// @ingroup ge
+  /// @brief Case Op Init.
+  /// @return Status
+  ///
+  Status InitCase(const OpDescPtr &op_desc);
+
+  Status SetDynamicBatchInfo(const OpDescPtr &op_desc, uint32_t batch_num);
+
   ///
   /// @ingroup ge
   /// @brief TVM Op Init.
@@ -840,7 +874,7 @@ class DavinciModel {
 
   // for reuse hccl_follow_stream
   std::mutex capacity_of_stream_mutex_;
-  std::vector<std::pair<rtStream_t, int64_t>> capacity_of_stream_;
+  std::map<int64_t, std::vector<rtStream_t>> main_follow_stream_mapping_;
 
   vector<rtEvent_t> event_list_;
 
@@ -866,6 +900,7 @@ class DavinciModel {
 
   bool is_async_mode_;  // For NN execute, Async mode use rtMemcpyAsync on rt_model_stream_.
 
+  bool is_stream_list_bind_{false};
   bool is_pure_head_stream_{false};
   rtStream_t rt_head_stream_{nullptr};
   rtStream_t rt_entry_stream_{nullptr};
@@ -891,8 +926,8 @@ class DavinciModel {
   std::set<uint32_t> hcom_streams_;
   RuntimeParam runtime_param_;
 
-  static std::mutex tvm_bin_mutex_;  // lock for tvm maps.
-  static std::set<std::string> tvm_bin_kernel_;
+  static std::mutex tvm_bin_mutex_;
+  std::set<std::string> tvm_bin_kernel_;
 
   std::map<std::string, uint32_t> used_tbe_handle_map_;
 
@@ -906,6 +941,7 @@ class DavinciModel {
   uint64_t iterator_count_;
   bool is_l1_fusion_enable_;
   std::map<OpDescPtr, void *> saved_task_addrs_;
+  void *l1_fusion_addr_ = nullptr;
 
   bool known_node_ = false;
   uint32_t total_args_size_ = 0;
@@ -921,7 +957,9 @@ class DavinciModel {
 
   vector<vector<int64_t>> batch_info_;
   std::vector<std::vector<int64_t>> combined_batch_info_;
+  vector<string> user_designate_shape_order_;
   int32_t dynamic_type_ = 0;
+  bool is_dynamic_ = false;
 
   vector<uint64_t> batch_size_;
   // key: input tensor name, generally rts op;
@@ -938,6 +976,8 @@ class DavinciModel {
   void *op_debug_addr_ = nullptr;
   void *p2p_debug_addr_ = nullptr;
   bool is_new_model_desc_{false};
+
+  std::map<int64_t, void *> memcpy_4g_offset_addr_;
 };
 }  // namespace ge
 #endif  // GE_GRAPH_LOAD_NEW_MODEL_MANAGER_DAVINCI_MODEL_H_

src/ge/graph/load/new_model_manager/model_manager.cc

@@ -20,6 +20,7 @@
 
 #include "common/l2_cache_optimize.h"
 #include "common/profiling/profiling_manager.h"
+#include "common/dump/dump_manager.h"
 #include "common/properties_manager.h"
 #include "framework/common/debug/ge_log.h"
 #include "framework/common/util.h"
@@ -172,7 +173,7 @@ ge::Status ModelManager::DestroyAicpuSessionForInfer(uint32_t model_id) {
     return GE_EXEC_MODEL_ID_INVALID;
   }
   uint64_t session_id = it->second->GetSessionId();
-  GELOGI("Destroy aicpu session for infer, session id is %u.", session_id);
+  GELOGI("Destroy aicpu session for infer, session id is %lu.", session_id);
   DestroyAicpuSession(session_id);
   return SUCCESS;
 }
@@ -259,7 +260,7 @@ Status ModelManager::LoadModelOnline(uint32_t &model_id, const shared_ptr<ge::Ge
   bool is_shape_unknown = false;
   GE_CHK_STATUS_RET(ge_root_model->CheckIsUnknownShape(is_shape_unknown), "CheckIsUnknownShape failed, model id:%u",
                     model_id);
-  if (is_shape_unknown) {
+  if (is_shape_unknown || GetContext().GetHostExecFlag()) {
     return DoLoadHybridModelOnline(model_id, ge_root_model, listener);
   }
 
@@ -729,6 +730,22 @@ Status ModelManager::GetCombinedDynamicDims(const uint32_t model_id, vector<vect
   return SUCCESS;
 }
 
+///
+/// @ingroup ge
+/// @brief Get user designate shape order
+/// @param [in] model_id
+/// @param [out] user_input_shape_order
+/// @return execute result
+///
+Status ModelManager::GetUserDesignateShapeOrder(const uint32_t model_id,
+                                                std::vector<std::string> &user_input_shape_order) {
+  auto davinci_model = GetModel(model_id);
+  GE_CHK_BOOL_RET_STATUS(davinci_model != nullptr, PARAM_INVALID,
+                         "GetUserDesignateShapeOrder Failed, Invalid Model ID %u!", model_id)
+  davinci_model->GetUserDesignateShapeOrder(user_input_shape_order);
+  return SUCCESS;
+}
+
 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);
@@ -831,7 +848,11 @@ Status ModelManager::LoadModelOffline(uint32_t &model_id, const ModelData &model
     }
     davinci_model->SetDeviceId(device_id);
     davinci_model->SetOmName(model.om_name);
-    davinci_model->SetDumpProperties(dump_properties_);
+    if (DumpManager::GetInstance().IsDumpOpen()) {
+      davinci_model->SetDumpProperties(DumpManager::GetInstance().GetDumpProperties());
+    } else {
+      davinci_model->SetDumpProperties(dump_properties_);
+    }
 
     /// In multi-threaded inference,  using the same session_id among multiple threads may cause some threads to fail.
     /// These session_ids come from the same model, so the values of session_id are the same.
@@ -1070,4 +1091,19 @@ ge::Status ModelManager::SyncExecuteModel(uint32_t model_id, const vector<GeTens
 
   return model->Execute(inputs, outputs);
 }
+
+Status ModelManager::GetOpDescInfo(uint32_t device_id, uint32_t stream_id, uint32_t task_id, OpDescInfo &op_desc_info) {
+  for (const auto &model : model_map_) {
+    auto davinci_model = model.second;
+    if (davinci_model->GetDeviceId() == device_id) {
+      GELOGI("Start to GetOpDescInfo of device_id: %u.", device_id);
+      if (davinci_model->GetOpDescInfo(stream_id, task_id, op_desc_info)) {
+        GELOGI("Find specific node of stream_id: %u, task_id: %u.", stream_id, task_id);
+        return SUCCESS;
+      }
+    }
+  }
+  return FAILED;
+}
+
 }  // namespace ge

src/ge/graph/load/new_model_manager/model_manager.h

@@ -201,6 +201,15 @@ class FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY ModelManager {
   ///
   ge::Status GetCombinedDynamicDims(const uint32_t model_id, std::vector<std::vector<int64_t>> &batch_info);
 
+  ///
+  /// @ingroup ge
+  /// @brief Get user designate shape order
+  /// @param [in] model_id
+  /// @param [out] user_input_shape_order
+  /// @return execute result
+  ///
+  Status GetUserDesignateShapeOrder(const uint32_t model_id, std::vector<std::string> &user_input_shape_order);
+
   ///
   /// @ingroup ge
   /// @brief Get AIPP info
@@ -263,6 +272,7 @@ class FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY ModelManager {
                                        std::vector<InputOutputDims> &output_dims);
 
   bool IsDynamicShape(uint32_t model_id);
+  ge::Status GetOpDescInfo(uint32_t device_id, uint32_t stream_id, uint32_t task_id, OpDescInfo &op_desc_info);
 
  private:
   ///

src/ge/graph/load/new_model_manager/task_info/hccl_task_info.cc

@@ -28,7 +28,6 @@ namespace {
 const uint32_t kMaxTaskOfStream = 200;
 }
 
-uint32_t HcclTaskInfo::max_node_of_hccl_stream_ = 0;
 std::mutex HcclTaskInfo::hccl_follow_stream_mutex_;
 
 HcclTaskInfo::~HcclTaskInfo() {
@@ -41,7 +40,6 @@ HcclTaskInfo::~HcclTaskInfo() {
   }
   davinci_model_ = nullptr;
   ops_kernel_store_ = nullptr;
-  max_node_of_hccl_stream_ = 0;
   args_ = nullptr;
 }
 Status HcclTaskInfo::Init(const domi::TaskDef &task_def, DavinciModel *davinci_model) {
@@ -133,45 +131,39 @@ Status HcclTaskInfo::SetFollowStream(const ge::ConstOpDescPtr &op_desc, DavinciM
   }
 
   std::lock_guard<std::mutex> lock(hccl_follow_stream_mutex_);
-  if (max_node_of_hccl_stream_ == 0) {
-    uint32_t max_stream_count;
-    uint32_t max_task_count;
-    ret = rtGetMaxStreamAndTask(RT_NORMAL_STREAM, &max_stream_count, &max_task_count);
-    if (ret != RT_ERROR_NONE) {
-      GELOGE(RT_FAILED, "Get max stream and task count by rts failed.");
-      return RT_ERROR_TO_GE_STATUS(ret);
-    }
-    max_node_of_hccl_stream_ = max_task_count / kMaxTaskOfStream;
-  }
+  int64_t main_stream_id = op_desc->GetStreamId();
+  const std::map<int64_t, std::vector<rtStream_t>> &main_follow_stream_mapping = davinci_model->GetHcclFolowStream();
 
-  if (static_cast<size_t>(hccl_stream_num) <= davinci_model->GetHcclFolowStream().size()) {
-    GELOGI("capacity of follow stream is enough to be reused.");
-    ReuseStream(hccl_stream_num, davinci_model);
+  if (main_follow_stream_mapping.find(main_stream_id) != main_follow_stream_mapping.end()) {
+    const std::vector<rtStream_t> &follow_stream_usage = main_follow_stream_mapping.at(main_stream_id);
+    if (static_cast<size_t>(hccl_stream_num) <= follow_stream_usage.size()) {
+      GELOGI("capacity of follow stream is enough to be reused.");
+      for (int64_t i = 0; i < hccl_stream_num; i++) {
+        hccl_stream_list_.emplace_back(follow_stream_usage.at(i));
+      }
+    } else {
+      GELOGI("need to reuse follow stream and create new follow stream.");
+      size_t created_stream_num = follow_stream_usage.size();
+      hccl_stream_list_ = follow_stream_usage;
+      ret = CreateStream(hccl_stream_num - created_stream_num, davinci_model, main_stream_id);
+      if (ret != SUCCESS) {
+        GELOGE(RT_FAILED, "Create hccl stream failed.");
+        return RT_ERROR_TO_GE_STATUS(ret);
+      }
+    }
+    GELOGI("Initialize hccl slave stream success, hcclStreamNum =%ld", hccl_stream_num);
   } else {
-    GELOGI("need to reuse follow stream and create new follow stream.");
-    size_t created_stream_num = davinci_model->GetHcclFolowStream().size();
-    ReuseStream(created_stream_num, davinci_model);
-    ret = CreateStream(hccl_stream_num - created_stream_num, davinci_model);
+    GELOGI("need to create follow stream for %s with new mainstream %ld.", op_desc->GetName().c_str(), main_stream_id);
+    ret = CreateStream(hccl_stream_num, davinci_model, main_stream_id);
     if (ret != SUCCESS) {
       GELOGE(RT_FAILED, "Create hccl stream failed.");
       return RT_ERROR_TO_GE_STATUS(ret);
     }
   }
-  GELOGI("Initialize hccl slave stream success, hcclStreamNum =%ld", hccl_stream_num);
   return SUCCESS;
 }
 
-void HcclTaskInfo::ReuseStream(int64_t stream_num, DavinciModel *davinci_model) {
-  GELOGI("Start to reuse %ld follow stream.", stream_num);
-  int64_t index = 0;
-  for (int64_t i = 0; i < stream_num; i++) {
-    hccl_stream_list_.emplace_back(davinci_model->GetHcclFolowStream().at(index).first);
-    int64_t remain_cap = davinci_model->GetHcclFolowStream().at(index).second - 1;
-    davinci_model->ReuseHcclFollowStream(remain_cap, index);
-  }
-}
-
-Status HcclTaskInfo::CreateStream(int64_t stream_num, DavinciModel *davinci_model) {
+Status HcclTaskInfo::CreateStream(int64_t stream_num, DavinciModel *davinci_model, int64_t main_stream_id) {
   GELOGI("Start to create %ld hccl stream.", stream_num);
   for (int64_t i = 0; i < stream_num; ++i) {
     rtStream_t stream = nullptr;
@@ -189,8 +181,7 @@ Status HcclTaskInfo::CreateStream(int64_t stream_num, DavinciModel *davinci_mode
       return RT_ERROR_TO_GE_STATUS(rt_ret);
     }
     GELOGD("hccl_stream addr is=%p", stream);
-    int64_t remain_cap = max_node_of_hccl_stream_ - 1;
-    davinci_model->CreateHcclFollowStream(stream, remain_cap);
+    davinci_model->SaveHcclFollowStream(main_stream_id, stream);
 
     hccl_stream_list_.emplace_back(stream);
     davinci_model->PushHcclStream(stream);

src/ge/graph/load/new_model_manager/task_info/hccl_task_info.h

@@ -60,9 +60,7 @@ class HcclTaskInfo : public TaskInfo {
 
   void GetPrivateDefByTaskDef(const domi::TaskDef &task);
 
-  void ReuseStream(int64_t stream_num, DavinciModel *davinci_model);
-
-  ge::Status CreateStream(int64_t stream_num, DavinciModel *davinci_model);
+  ge::Status CreateStream(int64_t stream_num, DavinciModel *davinci_model, int64_t main_stream_id);
 
   Status SetFollowStream(const ge::ConstOpDescPtr &op_desc, DavinciModel *davinci_model);
 
@@ -77,7 +75,6 @@ class HcclTaskInfo : public TaskInfo {
   void *private_def_;
   uint32_t private_def_len_;
   static std::mutex hccl_follow_stream_mutex_;
-  static uint32_t max_node_of_hccl_stream_;
   vector<GETaskKernelHcclInfo> kernel_hccl_infos_;
   vector<void *> input_data_addrs_;
   vector<void *> output_data_addrs_;

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

@@ -25,7 +25,6 @@
 #include "framework/common/debug/ge_log.h"
 #include "framework/common/l2_cache_optimize.h"
 #include "graph/debug/ge_attr_define.h"
-#include "graph/debug/ge_attr_define.h"
 #include "graph/load/new_model_manager/davinci_model.h"
 #include "graph/load/new_model_manager/model_utils.h"
 #include "runtime/kernel.h"
@@ -92,7 +91,7 @@ Status KernelTaskInfo::Init(const domi::TaskDef &task_def, DavinciModel *davinci
   string session_graph_model_id;
   davinci_model_->GetUniqueId(op_desc_, session_graph_model_id);
   // get bin_file_key
-  const char *bin_file_key = DavinciModel::GetRegisterStub(op_desc_->GetName(), session_graph_model_id);
+  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_ == cce::ccKernelType::CCE_AI_CORE) {
     rtError_t rt_ret;
@@ -395,7 +394,14 @@ Status KernelTaskInfo::Distribute() {
       "stubfunc:%p blockdim:%u stream:%p",
       call_skt, task_id_, skt_id_, skt_info_.last_task_id, stub_func_name_.c_str(), stub_func_, block_dim_, stream_);
     // l1 fusion enable and env flag open (kCloseSkt for skt debug)
-    if (call_skt && (env_flag != kCloseSkt)) {
+    bool open_dump = false;
+    auto all_dump_model = davinci_model_->GetDumpProperties().GetAllDumpModel();
+    if (all_dump_model.find(ge::DUMP_ALL_MODEL) != all_dump_model.end() ||
+        all_dump_model.find(davinci_model_->Name()) != all_dump_model.end() ||
+        all_dump_model.find(davinci_model_->OmName()) != all_dump_model.end()) {
+      open_dump = true;
+    }
+    if (call_skt && (env_flag != kCloseSkt) && !open_dump) {
       GE_RETURN_IF_ERROR(SuperKernelDistribute());
     } else {
       // call rtKernelLaunch for current task
@@ -577,7 +583,7 @@ Status KernelTaskInfo::InitTVMTask(uint16_t offset, const domi::KernelDef &kerne
   // When inferencing, stub_func_ is different from dynamic-registration to runtime, and needs to be modified.
   string session_graph_model_id;
   davinci_model_->GetUniqueId(op_desc, session_graph_model_id);
-  const char *bin_file_key = DavinciModel::GetRegisterStub(op_desc->GetName(), session_graph_model_id);
+  const char *bin_file_key = davinci_model_->GetRegisterStub(op_desc->GetName(), session_graph_model_id);
   rtError_t rt_ret = rtQueryFunctionRegistered(const_cast<char *>(bin_file_key));
   if (rt_ret != RT_ERROR_NONE) {
     stub_func_ = const_cast<char *>(bin_file_key);
@@ -634,7 +640,11 @@ Status KernelTaskInfo::InitTVMTask(uint16_t offset, const domi::KernelDef &kerne
   skt_dump_args_ = static_cast<char *>(args_) + offset;
   if (davinci_model_->GetDumpProperties().IsLayerNeedDump(davinci_model_->Name(), davinci_model_->OmName(),
                                                           op_desc->GetName())) {
-    dump_flag_ = RT_KERNEL_DUMPFLAG;
+    if (IsL1FusionOp(op_desc)) {
+      dump_flag_ = RT_FUSION_KERNEL_DUMPFLAG;
+    } else {
+      dump_flag_ = RT_KERNEL_DUMPFLAG;
+    }
     dump_args_ = static_cast<char *>(args_) + offset;
   }
 
@@ -653,6 +663,25 @@ Status KernelTaskInfo::InitTVMTask(uint16_t offset, const domi::KernelDef &kerne
   return SUCCESS;
 }
 
+bool KernelTaskInfo::IsL1FusionOp(const OpDescPtr &op_desc) {
+  std::vector<int64_t> input_memory_type;
+  (void)ge::AttrUtils::GetListInt(op_desc, ATTR_NAME_INPUT_MEM_TYPE_LIST, input_memory_type);
+  for (size_t i = 0; i < input_memory_type.size(); ++i) {
+    if (input_memory_type.at(i) == RT_MEMORY_L1) {
+      return true;
+    }
+  }
+
+  std::vector<int64_t> output_memory_type;
+  (void)ge::AttrUtils::GetListInt(op_desc, ATTR_NAME_OUTPUT_MEM_TYPE_LIST, output_memory_type);
+  for (size_t i = 0; i < output_memory_type.size(); ++i) {
+    if (output_memory_type.at(i) == RT_MEMORY_L1) {
+      return true;
+    }
+  }
+  return false;
+}
+
 Status KernelTaskInfo::InitAICPUCustomTask(uint32_t op_index, const domi::KernelDef &kernel_def) {
   GELOGI("Do InitAICPUCustomTask");
   OpDescPtr op_desc = davinci_model_->GetOpByIndex(op_index);
@@ -904,7 +933,11 @@ Status KernelTaskInfo::InitAicpuTask(uint32_t op_index, const domi::KernelDef &k
 
   if (davinci_model_->GetDumpProperties().IsLayerNeedDump(davinci_model_->Name(), davinci_model_->OmName(),
                                                           op_desc->GetName())) {
-    dump_flag_ = RT_KERNEL_DUMPFLAG;
+    if (IsL1FusionOp(op_desc)) {
+      dump_flag_ = RT_FUSION_KERNEL_DUMPFLAG;
+    } else {
+      dump_flag_ = RT_KERNEL_DUMPFLAG;
+    }
     dump_args_ = static_cast<char *>(args_) + sizeof(aicpu::AicpuParamHead);
   }
 

src/ge/graph/load/new_model_manager/task_info/kernel_task_info.h

@@ -127,6 +127,7 @@ class KernelTaskInfo : public TaskInfo {
   static void FreeRtMem(void **ptr);
 
   Status SuperKernelDistribute();
+  bool IsL1FusionOp(const OpDescPtr &op_desc);
 
   // For super kernel
   Status SaveSKTDumpInfo();

src/ge/graph/load/new_model_manager/task_info/memcpy_addr_async_task_info.cc

@@ -59,7 +59,12 @@ Status MemcpyAddrAsyncTaskInfo::Init(const domi::TaskDef &task_def, DavinciModel
 
   // malloc args memory
   size_t args_size = sizeof(void *) * io_addrs.size();
-  rtError_t rt_ret = rtMalloc(&args_, args_size + kAlignBytes, RT_MEMORY_HBM);
+  rtMemType_t memory_type = RT_MEMORY_HBM;
+  if (op_desc->HasAttr(ATTR_NAME_MEMORY_TYPE_RANGE)) {
+    memory_type = RT_MEMORY_TS_4G;
+  }
+  GELOGI("memory_type: %u", memory_type);
+  rtError_t rt_ret = rtMalloc(&args_, args_size + kAlignBytes, memory_type);
   if (rt_ret != RT_ERROR_NONE) {
     GELOGE(RT_FAILED, "Call rt api failed, ret: 0x%X", rt_ret);
     return RT_ERROR_TO_GE_STATUS(rt_ret);

src/ge/graph/load/new_model_manager/task_info/memcpy_async_task_info.cc

@@ -36,6 +36,12 @@ Status MemcpyAsyncTaskInfo::Init(const domi::TaskDef &task_def, DavinciModel *da
   count_ = memcpy_async.count();
   kind_ = memcpy_async.kind();
   dst_max_ = memcpy_async.dst_max();
+  OpDescPtr op_desc = davinci_model->GetOpByIndex(memcpy_async.op_index());
+  if (op_desc == nullptr) {
+    GELOGE(INTERNAL_ERROR, "Task op index:%u out of range", memcpy_async.op_index());
+    return INTERNAL_ERROR;
+  }
+
   if (davinci_model->IsKnownNode()) {
     src_ = reinterpret_cast<uint8_t *>(davinci_model_->GetCurrentArgsAddr(args_offset_));
     dst_ = reinterpret_cast<uint8_t *>(reinterpret_cast<uintptr_t>(src_) + sizeof(void *));
@@ -49,9 +55,17 @@ Status MemcpyAsyncTaskInfo::Init(const domi::TaskDef &task_def, DavinciModel *da
     return ret;
   }
 
-  ret = ModelUtils::GetRtAddress(davinci_model->GetRuntimeParam(), memcpy_async.dst(), dst_);
-  if (ret != SUCCESS) {
-    return ret;
+  // dst_ needs different address for different chips
+  if (op_desc->HasAttr(ATTR_NAME_MEMORY_TYPE_RANGE)) {
+    ret = AllocTsMemoryForMemcpy(op_desc, davinci_model);
+    if (ret != SUCCESS) {
+      return ret;
+    }
+  } else {
+    ret = ModelUtils::GetRtAddress(davinci_model->GetRuntimeParam(), memcpy_async.dst(), dst_);
+    if (ret != SUCCESS) {
+      return ret;
+    }
   }
 
   GELOGI("MemcpyAsyncTaskInfo Init Success, logic[0x%lx, 0x%lx], src:%p, dst:%p, max:%lu, count:%lu",
@@ -108,5 +122,33 @@ Status MemcpyAsyncTaskInfo::UpdateArgs() {
   return SUCCESS;
 }
 
+Status MemcpyAsyncTaskInfo::AllocTsMemoryForMemcpy(const OpDescPtr &op_desc, DavinciModel *davinci_model) {
+  int64_t size = 0;
+  auto tensor_desc = op_desc->GetOutputDescPtr(0);
+  if ((tensor_desc == nullptr) || (TensorUtils::GetTensorSizeInBytes(*tensor_desc, size) != GRAPH_SUCCESS)) {
+    GELOGE(FAILED, "GetTensorSizeInBytes failed!");
+    return FAILED;
+  }
+
+  rtError_t rt_ret = rtMalloc(&memory_4g_, size, RT_MEMORY_TS_4G);
+  if (rt_ret != RT_ERROR_NONE) {
+    GELOGE(RT_FAILED, "rtMalloc failed, ret: 0x%X", rt_ret);
+    return FAILED;
+  }
+
+  // map save the opdesc's offset and special address, for update the streamSwitchN's input address
+  std::map<int64_t, void *> memcpy_4g_offset_addr;
+  vector<int64_t> offsets = op_desc->GetOutputOffset();
+  if (offsets.empty()) {
+    GELOGE(FAILED, "GetOutputOffset failed!");
+    return FAILED;
+  }
+  memcpy_4g_offset_addr.insert(std::pair<int64_t, void *>(offsets[0], memory_4g_));
+  davinci_model->SetMemcpyOffsetAndAddr(memcpy_4g_offset_addr);
+
+  dst_ = reinterpret_cast<uint8_t *>(memory_4g_);
+  return SUCCESS;
+}
+
 REGISTER_TASK_INFO(RT_MODEL_TASK_MEMCPY_ASYNC, MemcpyAsyncTaskInfo);
 }  // namespace ge

src/ge/graph/load/new_model_manager/task_info/memcpy_async_task_info.h

@@ -18,15 +18,24 @@
 #define GE_GRAPH_LOAD_NEW_MODEL_MANAGER_TASK_INFO_MEMCPY_ASYNC_TASK_INFO_H_
 
 #include "graph/load/new_model_manager/task_info/task_info.h"
+#include "graph/op_desc.h"
 
 namespace ge {
 class MemcpyAsyncTaskInfo : public TaskInfo {
  public:
-  MemcpyAsyncTaskInfo() : dst_(nullptr), dst_max_(0), src_(nullptr), count_(0), kind_(0) {}
+  MemcpyAsyncTaskInfo() : dst_(nullptr), dst_max_(0), src_(nullptr), count_(0), kind_(0), memory_4g_(nullptr) {}
 
   ~MemcpyAsyncTaskInfo() override {
     src_ = nullptr;
     dst_ = nullptr;
+
+    if (memory_4g_ != nullptr) {
+      rtError_t ret = rtFree(memory_4g_);
+      if (ret != RT_ERROR_NONE) {
+        GELOGE(RT_FAILED, "Call rt api failed, ret: 0x%X", ret);
+      }
+      memory_4g_ = nullptr;
+    }
   }
 
   Status Init(const domi::TaskDef &task_def, DavinciModel *davinci_model) override;
@@ -38,6 +47,7 @@ class MemcpyAsyncTaskInfo : public TaskInfo {
   Status CalculateArgs(const domi::TaskDef &task_def, DavinciModel *davinci_model) override;
 
  private:
+  Status AllocTsMemoryForMemcpy(const OpDescPtr &op_desc, DavinciModel *davinci_model);
   uint8_t *dst_;
   uint64_t dst_max_;
   uint8_t *src_;
@@ -46,6 +56,7 @@ class MemcpyAsyncTaskInfo : public TaskInfo {
   DavinciModel *davinci_model_ = nullptr;
   uint32_t args_offset_ = 0;
   domi::MemcpyAsyncDef memcpy_async;
+  void *memory_4g_;
 };
 }  // namespace ge
 #endif  // GE_GRAPH_LOAD_NEW_MODEL_MANAGER_TASK_INFO_MEMCPY_ASYNC_TASK_INFO_H_

src/ge/graph/load/new_model_manager/task_info/stream_switchn_task_info.cc

@@ -66,16 +66,13 @@ Status StreamSwitchNTaskInfo::Init(const domi::TaskDef &task_def, DavinciModel *
     GELOGE(FAILED, "Get true stream ptr of switchN op failed.");
     return FAILED;
   }
-  if (davinci_model->IsKnownNode()) {
-    input_ptr_ = davinci_model->GetCurrentFixedAddr(args_offset_);
-  } else {
-    auto input_data_addr = ModelUtils::GetInputDataAddrs(davinci_model->GetRuntimeParam(), op_desc);
-    if (input_data_addr.empty()) {
-      GELOGE(FAILED, "Input data addr is nullptr.");
-      return FAILED;
-    }
-    input_ptr_ = input_data_addr[0];
+
+  // update StreamSwitchN's input_ptr_
+  Status ret = InputPtrUpdate(op_desc, davinci_model);
+  if (ret != SUCCESS) {
+    return ret;
   }
+
   davinci_model->DisableZeroCopy(input_ptr_);
   GELOGI("StreamSwitchNTaskInfo Init Success, inputSize:%u, elementSize:%d, trueStreamID:%ld.", input_size_,
          element_size_, op_desc->GetStreamId());
@@ -154,5 +151,36 @@ Status StreamSwitchNTaskInfo::CalculateArgs(const domi::TaskDef &task_def, Davin
   GELOGI("Calculate stream switchn task args , tensor_size %ld, args_offset %ld", tensor_size, args_offset_);
   return SUCCESS;
 }
+
+Status StreamSwitchNTaskInfo::InputPtrUpdate(const OpDescPtr &op_desc, DavinciModel *davinci_model) {
+  bool is_4g_mem = false;
+  const map<int64_t, void *> memcpy_4g_offset_addr = davinci_model->GetMemcpyOffsetAndAddr();
+  vector<int64_t> input_offset = op_desc->GetInputOffset();
+  if (input_offset.empty()) {
+    GELOGE(FAILED, "Get StreamSwitchN's input offset failed.");
+    return FAILED;
+  }
+
+  auto iter = memcpy_4g_offset_addr.find(input_offset[0]);
+  if (iter != memcpy_4g_offset_addr.end()) {
+    input_ptr_ = iter->second;
+    is_4g_mem = true;
+  }
+
+  if (is_4g_mem == false) {
+    if (davinci_model->IsKnownNode()) {
+      input_ptr_ = davinci_model->GetCurrentFixedAddr(args_offset_);
+    } else {
+      auto input_data_addr = ModelUtils::GetInputDataAddrs(davinci_model->GetRuntimeParam(), op_desc);
+      if (input_data_addr.empty()) {
+        return FAILED;
+      }
+      input_ptr_ = input_data_addr[0];
+    }
+  }
+
+  GELOGI("StreamSwitchN's input_ptr is %p, is_4g_mem: %d", input_ptr_, is_4g_mem);
+  return SUCCESS;
+}
 REGISTER_TASK_INFO(RT_MODEL_TASK_STREAM_SWITCH_N, StreamSwitchNTaskInfo);
-}  // namespace ge
+}  // namespace ge

src/ge/graph/load/new_model_manager/task_info/stream_switchn_task_info.h

@@ -42,6 +42,7 @@ class StreamSwitchNTaskInfo : public TaskInfo {
 
  private:
   Status GetTrueStreamPtr(const OpDescPtr &op_desc, DavinciModel *davinci_model);
+  Status InputPtrUpdate(const OpDescPtr &op_desc, DavinciModel *davinci_model);
   void *input_ptr_;
   uint32_t input_size_;
   void *value_ptr_;

src/ge/graph/load/new_model_manager/zero_copy_task.cc

@@ -143,10 +143,11 @@ Status ZeroCopyTask::UpdateTaskParam(uintptr_t addr, void *buffer_addr, const ma
 /**
  * @ingroup ge
  * @brief Update task param to device.
+ * @param [in] async_mode: true for asychronous mode.
  * @param [in] stream: Stream for asychronous update.
  * @return: 0 SUCCESS / others FAILED
  */
-Status ZeroCopyTask::DistributeParam(rtStream_t stream) {
+Status ZeroCopyTask::DistributeParam(bool async_mode, rtStream_t stream) {
   if (!is_updated_) {
     return SUCCESS;
   }
@@ -154,7 +155,7 @@ Status ZeroCopyTask::DistributeParam(rtStream_t stream) {
   is_updated_ = false;
   GE_CHECK_NOTNULL(args_addr_);
   rtError_t rt_err = RT_ERROR_NONE;
-  if (stream != nullptr) {
+  if (async_mode) {
     rt_err =
       rtMemcpyAsync(args_addr_, args_size_, args_info_.data(), args_info_.size(), RT_MEMCPY_HOST_TO_DEVICE_EX, stream);
   } else {

src/ge/graph/load/new_model_manager/zero_copy_task.h

@@ -77,10 +77,11 @@ class ZeroCopyTask {
   /**
    * @ingroup ge
    * @brief Update task param to device.
+   * @param [in] async_mode: true for asychronous mode.
    * @param [in] stream: Stream for asychronous update.
    * @return: 0 SUCCESS / others FAILED
    */
-  ge::Status DistributeParam(rtStream_t stream);
+  ge::Status DistributeParam(bool async_mode, rtStream_t stream);
 
  protected:
   bool CheckDynamicBatch(const map<string, set<uintptr_t>> &batch_addrs, const string &batch_label, uintptr_t addr);
@@ -97,4 +98,4 @@ class ZeroCopyTask {
   map<uintptr_t, vector<size_t>> task_addr_offset_;
 };
 }  // namespace ge
-#endif  // GE_GRAPH_LOAD_NEW_MODEL_MANAGER_ZERO_COPY_TASK_H_
+#endif  // GE_GRAPH_LOAD_NEW_MODEL_MANAGER_ZERO_COPY_TASK_H_

src/ge/graph/manager/graph_manager.cc

@@ -91,7 +91,13 @@
 #include "graph/passes/variable_ref_delete_op_pass.h"
 #include "graph/passes/variable_ref_useless_control_out_delete_pass.h"
 #include "graph/passes/end_of_sequence_add_control_pass.h"
+#include "graph/passes/subexpression_migration_pass.h"
+#include "graph/passes/unused_args_clean_pass.h"
+#include "graph/passes/global_step_insert_pass.h"
 #include "graph/utils/tensor_adapter.h"
+#include "graph/utils/type_utils.h"
+#include "graph/graph_util.h"
+#include "graph/types.h"
 #include "inc/pass_manager.h"
 #include "init/gelib.h"
 
@@ -102,6 +108,8 @@ const char *const kNetOutput = "NetOutput";
 const char *const kVariable = "Variable";
 const char *const kSend = "Send";
 const char *const kRecv = "Recv";
+const char *const kCheckPointForGetVar = "CheckPointGraphForGetVar";
+const char *const kCheckPointGraph = "checkpoint_graph";
 
 bool IsTailingOptimization() {
   string is_tailing_optimization_option;
@@ -380,6 +388,11 @@ Status GraphManager::PreRun(const GraphNodePtr &graph_node, const std::vector<Ge
 
   GM_RUN_AND_DUMP_PERF("PrepareRunningFormatRefiner", graph_preparer_.PrepareRunningFormatRefiner);
   GM_RUN_AND_DUMP_PERF("RefineRunningFormat", graph_optimize_.OptimizeOriginalGraphJudgeInsert, compute_graph);
+  if (std::getenv("AnalyzeMode")) {
+    GELOGI("Do return failed after refine_running_format when in analyze mode!");
+    return FAILED;
+  }
+  GM_RUN_AND_DUMP_PERF("SubexpressionMigration", SubexpressionMigration, compute_graph);
   GE_RUN(GraphManager, graph_preparer_.RecordAIPPInfo, compute_graph);
   if (IsTailingOptimization()) {
     GM_RUN_AND_DUMP_PERF("OptimizeSwitchOp", graph_preparer_.SwitchOpOptimize, compute_graph);
@@ -392,9 +405,11 @@ Status GraphManager::PreRun(const GraphNodePtr &graph_node, const std::vector<Ge
     GE_CHK_STATUS_RET(graph_pass.AddPass("PreRun::CtrlEdgeTransferPass", new (std::nothrow) CtrlEdgeTransferPass))
     GE_CHK_STATUS_RET(graph_pass.Run(compute_graph));
   }
+
   GE_CHK_STATUS_RET(graph_optimize_.IdentifyReference(compute_graph), "Identify reference failed.");
   GM_RUN_AND_DUMP_PERF("OptimizeSubgraph", OptimizeSubgraph, graph_node, compute_graph, session_id);
   GM_RUN_AND_DUMP_PERF("Optimize2", OptimizeStage2, compute_graph);
+  GM_RUN_AND_DUMP_PERF("OptimizeGraphBeforeBuildForRts", graph_optimize_.OptimizeGraphBeforeBuildForRts, compute_graph);
   GM_RUN_AND_DUMP_PERF("Build", Build, graph_node, compute_graph, ge_root_model, session_id);
 
   // when set incre build, save om model and var manager
@@ -403,12 +418,25 @@ Status GraphManager::PreRun(const GraphNodePtr &graph_node, const std::vector<Ge
   if (save_ret != SUCCESS) {
     GELOGW("Fail to save cache.");
   }
-  // release rts generate context
-  RtContextUtil::GetInstance().DestroyRtContexts(session_id);
   GEEVENT("[GEPERFTRACE] GE PreRun End");
   return SUCCESS;
 }
-#undef RUN_AND_DUMP
+
+Status GraphManager::SubexpressionMigration(ComputeGraphPtr &compute_graph) {
+  PassManager pass_manager;
+  GE_CHK_STATUS_RET(pass_manager.AddPass("SubexpressionMigrationPass", new (std::nothrow) SubexpressionMigrationPass));
+  GE_CHK_STATUS_RET(pass_manager.AddPass("UnusedArgsCleanPass", new (std::nothrow) UnusedArgsCleanPass));
+
+  GE_TIMESTAMP_START(SubexpressionMigrationPass);
+  auto ret = pass_manager.Run(compute_graph);
+  GE_TIMESTAMP_END(SubexpressionMigrationPass, "GraphManager::OptimizeStage1_1");
+  if (ret != SUCCESS && ret != NOT_CHANGED) {
+    GELOGE(ret, "Run SubexpressionMigrationPass failed, ret:%u.", ret);
+    return ret;
+  }
+
+  return SUCCESS;
+}
 
 Status GraphManager::StartForRunGraph(const GraphNodePtr &graph_node, const std::vector<GeTensor> &inputs,
                                       GeRootModelPtr &ge_root_model, uint64_t session_id) {
@@ -427,6 +455,8 @@ Status GraphManager::StartForRunGraph(const GraphNodePtr &graph_node, const std:
     ret = IncreBuild(graph_node, ge_model);
     if (ret != SUCCESS) {
       ret = PreRun(graph_node, inputs, ge_root_model, session_id);
+      // release rts generate context
+      RtContextUtil::GetInstance().DestroyRtContexts(session_id);
       if (ret != SUCCESS) {
         GELOGE(ret, "PreRun Failed.");
         return ret;
@@ -1388,6 +1418,9 @@ bool GraphManager::CheckNetOutputForCheckpointGraph(NodePtr &node) {
 }
 
 bool GraphManager::CheckVariableForCheckpointGraph(NodePtr &node) {
+  if (node->GetOpDesc()->HasAttr(kCheckPointForGetVar)) {
+    return false;
+  }
   auto out = node->GetOutDataAnchor(0);
   if (out == nullptr) {
     GELOGE(GE_GRAPH_PARAM_NULLPTR, "out is nullptr.");
@@ -1573,48 +1606,6 @@ Status GraphManager::RemoveIsolatedConst(ge::ComputeGraphPtr &compute_graph) {
   return SUCCESS;
 }
 
-Status GraphManager::NewOptimizeAfterMergeSubGraph(ge::ComputeGraphPtr &compute_graph) {
-  GELOGD("NewOptimizeAfterMergeSubGraph in");
-
-  GEPass ge_passes(compute_graph);
-  NamesToPass names_to_passes;
-  ConstantFoldingPass constant_folding_pass;
-  names_to_passes.emplace_back("ConstantFoldingPass", &constant_folding_pass);
-  GE_TIMESTAMP_START(names_to_passes);
-  auto ret = ge_passes.Run(names_to_passes);
-  GE_TIMESTAMP_END(names_to_passes, "GraphManager::ge_passes");
-  if (ret != SUCCESS) {
-    GELOGE(ret, "Run ge_passes optimize for OptimizeAfterMergeSubGraph failed, ret:%d.", ret);
-    return ret;
-  }
-
-  ret = RemoveIsolatedConst(compute_graph);
-  if (ret != SUCCESS) {
-    GELOGE(ret, "Remove isolated Constant failed, ret:%d.", ret);
-    return ret;
-  }
-
-  PassManager passes;
-  GE_CHK_STATUS_RET(passes.AddPass("MultiBatchPass", new (std::nothrow) MultiBatchPass));
-  GE_CHK_STATUS_RET(passes.AddPass("CompileNodesPass", new (std::nothrow) CompileNodesPass));
-  GE_CHK_STATUS_RET(passes.AddPass("AtomicAddrCleanPass", new (std::nothrow) AtomicAddrCleanPass));
-
-  GE_TIMESTAMP_START(passes);
-  ret = passes.Run(compute_graph);
-  GE_TIMESTAMP_END(passes, "GraphManager::passes");
-  if (ret != SUCCESS && ret != NOT_CHANGED) {
-    GELOGE(ret, "Run passes optimize for OptimizeAfterMergeSubGraph failed");
-    return ret;
-  }
-
-  ret = compute_graph->TopologicalSorting();
-  if (ret != SUCCESS) {
-    GELOGE(ret, "Graph topological sort failed, ret:%d.", ret);
-    return ret;
-  }
-  return SUCCESS;
-}
-
 Status GraphManager::OptimizeStage1(ge::ComputeGraphPtr &compute_graph) {
   string options = "default";
   if (GetContext().GetOption("ge.exec.variable_acc", options) != SUCCESS) {
@@ -1721,10 +1712,17 @@ Status GraphManager::OptimizeStage1(ge::ComputeGraphPtr &compute_graph) {
     graph_pass.AddPass("OptimizeStage1_3::SwitchToStreamSwitchPass", new (std::nothrow) SwitchToStreamSwitchPass))
   GE_CHK_STATUS_RET(
     graph_pass.AddPass("OptimizeStage1_3::AttachStreamLabelPass", new (std::nothrow) AttachStreamLabelPass))
+  GE_CHK_STATUS_RET(graph_pass.AddPass("OptimizeStage1_3::MultiBatchPass", new (std::nothrow) MultiBatchPass(true)))
   GE_CHK_STATUS_RET(graph_pass.AddPass("OptimizeStage1_3::IteratorOpPass", new (std::nothrow) IteratorOpPass))
   GE_CHK_STATUS_RET(graph_pass.AddPass("OptimizeStage1_3::VariableRefUselessControlOutDeletePass",
                                        new (std::nothrow) VariableRefUselessControlOutDeletePass))
   GE_CHK_STATUS_RET(graph_pass.AddPass("OptimizeStage1_3::ReshapeRecoveryPass", new (std::nothrow) ReshapeRecoveryPass))
+  if (options_.train_graph_flag) {
+    // Priority: The GlobalStepInsertPass should work before graph partitioner.
+    // Reason: Make sure that the var "global_step" can be partitioned to known sub graph and allocated memory
+    GE_CHK_STATUS_RET(
+      graph_pass.AddPass("OptimizeStage1_3::GlobalStepInsertPass", new (std::nothrow) GlobalStepInsertPass))
+  }
   GE_TIMESTAMP_START(graph_pass);
   ret = graph_pass.Run(compute_graph);
   GE_TIMESTAMP_END(graph_pass, "GraphManager::OptimizeStage1_3");
@@ -1787,11 +1785,8 @@ Status GraphManager::OptimizeStage2(ge::ComputeGraphPtr &compute_graph) {
 
   PassManager pass_for_control_attr_optimize;
   if (options_.train_graph_flag) {
-    const char *unknown_shape_skip = std::getenv("EXPERIMENTAL_DYNAMIC_PARTITION");
-    if (unknown_shape_skip == nullptr) {
-      GE_CHK_STATUS_RET(pass_for_control_attr_optimize.AddPass("OptimizeStage2::ControlAttrOptimize::FlowCtrlPass",
-                                                               new (std::nothrow) FlowCtrlPass))
-    }
+    GE_CHK_STATUS_RET(pass_for_control_attr_optimize.AddPass("OptimizeStage2::ControlAttrOptimize::FlowCtrlPass",
+                                                             new (std::nothrow) FlowCtrlPass))
   }
 
   GE_CHK_STATUS_RET(pass_for_control_attr_optimize.AddPass("OptimizeStage2::ControlAttrOptimize::MultiBatchPass",
@@ -1821,14 +1816,10 @@ Status GraphManager::OptimizeStage2(ge::ComputeGraphPtr &compute_graph) {
     pass_for_control_attr_optimize.AddPass("OptimizeStage2::AfterMergePasses::"
                                            "EndOfSequenceAddControlPass",
                                            new (std::nothrow) EndOfSequenceAddControlPass))
-
-  const char *unknown_shape_skip = std::getenv("EXPERIMENTAL_DYNAMIC_PARTITION");
-  if (unknown_shape_skip == nullptr) {
-    // SubgraphPass solves memory_assign_conflicts by insert MemcpyAsync node, which depends on multi attrs and
-    // graph-structure. So try not to add new pass after SubgraphPass.
-    GE_CHK_STATUS_RET(pass_for_control_attr_optimize.AddPass("OptimizeStage2::ControlAttrOptimize::SubgraphPass",
-                                                             new (std::nothrow) SubgraphPass))
-  }
+  // SubgraphPass solves memory_assign_conflicts by insert MemcpyAsync node, which depends on multi attrs and
+  // graph-structure. So try not to add new pass after SubgraphPass.
+  GE_CHK_STATUS_RET(pass_for_control_attr_optimize.AddPass("OptimizeStage2::ControlAttrOptimize::SubgraphPass",
+                                                           new (std::nothrow) SubgraphPass))
   // AttachStreamLabelPass modifies attr without changing structure of compute_graph
   GE_CHK_STATUS_RET(pass_for_control_attr_optimize.AddPass("OptimizeStage2::ControlAttrOptimize::AttachStreamLabelPass",
                                                            new (std::nothrow) AttachStreamLabelPass))
@@ -1870,120 +1861,6 @@ void GraphManager::ChangeConstTypeWhenTraining(const ComputeGraphPtr &compute_gr
     }
   }
 }
-Status GraphManager::OptimizeAfterMergeSubGraph(ge::ComputeGraphPtr &compute_graph) {
-  GELOGI("Start optimize after merge sub graph.");
-
-  GEPass ge_passes_for_shape(compute_graph);
-  NamesToPass names_to_passes_for_shape;
-  CastRemovePass cast_remove_pass;
-  names_to_passes_for_shape.emplace_back("CastRemovePass", &cast_remove_pass);
-  TransposeTransDataPass transpose_transdata_pass;
-  names_to_passes_for_shape.emplace_back("TransposeTransDataPass", &transpose_transdata_pass);
-  GE_TIMESTAMP_START(ge_passes_for_shape);
-  Status ret = ge_passes_for_shape.Run(names_to_passes_for_shape);
-  GE_TIMESTAMP_END(ge_passes_for_shape, "GraphManager::GePassesForShape");
-  GE_CHK_STATUS_RET(ret, "Run ge_passes_for_shape optimize for OptimizeAfterMergeSubGraph failed, ret:%d.", ret);
-
-  string options = "default";
-  if (GetContext().GetOption("ge.exec.variable_acc", options) != SUCCESS) {
-    GELOGI("get ge.exec.variable_acc failed. set default value.");
-  }
-  PassManager after_merge_passes;
-  GE_CHK_STATUS_RET(after_merge_passes.AddPass("PermutePass", new (std::nothrow) PermutePass));
-  GE_IF_BOOL_EXEC(options == "default" || options == "1", GELOGI("turn on variable accelerator"); GE_CHK_STATUS_RET(
-                    after_merge_passes.AddPass("VariableOpPass", new (std::nothrow) VariableOpPass(&var_acc_ctrl_))));
-  ret = after_merge_passes.Run(compute_graph);
-  if (ret != SUCCESS && ret != NOT_CHANGED) {
-    GELOGE(ret, "Run passes after merge sub graph failed, ret:%d.", ret);
-    return ret;
-  }
-
-  // reshape remove + symmetry_elimination_pass to replace transop depth fusion pass
-  GEPass ge_passes_symmetry(compute_graph);
-  NamesToPass names_to_passes_for_symmetry;
-  ReshapeRemovePass reshape_remove_pass;
-  names_to_passes_for_symmetry.emplace_back("ReshapeRemovePass", &reshape_remove_pass);
-  TransOpSymmetryEliminationPass symmetry_elimination_pass;
-  names_to_passes_for_symmetry.emplace_back("TransOpSymmetryEliminationPass", &symmetry_elimination_pass);
-  ret = ge_passes_symmetry.Run(names_to_passes_for_symmetry);
-  GE_CHK_STATUS_RET(ret, "Run ge_passes optimize for OptimizeAfterMergeSubGraph failed, ret:%d.", ret);
-
-  PassManager after_merge_fusion_passes;
-  GE_CHK_STATUS_RET(after_merge_fusion_passes.AddPass("TransOpWithoutReshapeFusionPass",
-                                                      new (std::nothrow) TransOpWithoutReshapeFusionPass));
-  GE_CHK_STATUS_RET(
-    after_merge_fusion_passes.AddPass("TransOpBreadthFusionPass", new (std::nothrow) TransOpBreadthFusionPass));
-  GE_CHK_STATUS_RET(
-    after_merge_fusion_passes.AddPass("VariableRefDeleteOpPass", new (std::nothrow) VariableRefDeleteOpPass));
-  GE_CHK_STATUS_RET(after_merge_fusion_passes.AddPass("SameTransdataBreadthFusionPass",
-                                                      new (std::nothrow) SameTransdataBreadthFusionPass));
-  GE_CHK_STATUS_RET(
-    after_merge_fusion_passes.AddPass("MarkGraphUnknownStatusPass", new (std::nothrow) MarkGraphUnknownStatusPass));
-  GE_CHK_STATUS_RET(after_merge_fusion_passes.AddPass("AtomicAddrCleanPass", new (std::nothrow) AtomicAddrCleanPass));
-  GE_CHK_STATUS_RET(after_merge_fusion_passes.AddPass(
-    "LinkGenMaskNodesPass", new (std::nothrow) LinkGenMaskNodesPass(options_.stream_max_parallel_num)));
-  GE_TIMESTAMP_START(after_merge_fusion_passes);
-  ret = after_merge_fusion_passes.Run(compute_graph);
-  GE_TIMESTAMP_END(after_merge_fusion_passes, "GraphManager::AfterMergePasses");
-  if (ret != SUCCESS && ret != NOT_CHANGED) {
-    GELOGE(ret, "Run passes after merge sub graph failed, ret:%d.", ret);
-    return ret;
-  }
-
-  // add variable attr for hccl broadcast,need to be removed after variable pass online
-  for (const ge::NodePtr &node : compute_graph->GetDirectNode()) {
-    if (node->GetOpDesc()->GetType() != VARIABLE) {
-      continue;
-    }
-
-    if (IsBroadCastOpData(node)) {
-      AdjustBroadCastOpData(node);
-    }
-    if (IsAssignOpData(node)) {
-      AdjustAssignOpData(node);
-    }
-  }
-
-  GEPass ge_passes(compute_graph);
-  NamesToPass names_to_passes;
-  TransOpNearbyAllreduceFusionPass trans_op_nearby_allreduce_fusion_pass;
-  names_to_passes.emplace_back("TransOpNearbyAllreduceFusionPass", &trans_op_nearby_allreduce_fusion_pass);
-  names_to_passes_for_shape.emplace_back("ReshapeRemovePass", &reshape_remove_pass);
-  ConstantFoldingPass constant_folding_pass;
-  names_to_passes.emplace_back("ConstantFoldingPass", &constant_folding_pass);
-  DimensionAdjustPass dimension_adjust_pass;
-  names_to_passes.emplace_back("DimensionAdjustPass", &dimension_adjust_pass);
-  CondRemovePass condition_remove_pass;
-  names_to_passes.emplace_back("CondRemovePass", &condition_remove_pass);
-  GE_TIMESTAMP_START(names_to_passes);
-  ret = ge_passes.Run(names_to_passes);
-  GE_TIMESTAMP_END(names_to_passes, "GraphManager::MergedGraphNameToPasses");
-  GE_CHK_STATUS_RET(ret, "Run ge_passes optimize for OptimizeAfterMergeSubGraph failed, ret:%d.", ret);
-
-  ret = RemoveIsolatedConst(compute_graph);
-  GE_CHK_STATUS_RET(ret, "Remove isolated Constant failed, ret:%d.", ret);
-
-  PassManager pass_for_optimize;
-  const char *unknown_shape_skip = std::getenv("EXPERIMENTAL_DYNAMIC_PARTITION");
-  if (unknown_shape_skip == nullptr) {
-    GE_CHK_STATUS_RET(pass_for_optimize.AddPass("SubgraphPass", new (std::nothrow) SubgraphPass));
-  }
-  GE_CHK_STATUS_RET(pass_for_optimize.AddPass("MultiBatchPass", new (std::nothrow) MultiBatchPass));
-  GE_CHK_STATUS_RET(pass_for_optimize.AddPass("CompileNodesPass", new (std::nothrow) CompileNodesPass));
-  GE_TIMESTAMP_START(pass_for_optimize);
-  ret = pass_for_optimize.Run(compute_graph);
-  GE_TIMESTAMP_END(pass_for_optimize, "GraphManager::OptimizePass");
-  if (ret != SUCCESS && ret != NOT_CHANGED) {
-    GELOGE(ret, "Run optimize pass failed");
-    return ret;
-  }
-
-  ret = compute_graph->TopologicalSorting();
-  GE_CHK_STATUS_RET(ret, "Graph topological sort failed, ret:%d.", ret);
-
-  GELOGI("End optimize after merge sub graph.");
-  return SUCCESS;
-}
 
 Status GraphManager::LoadGraphAsync(const GeRootModelPtr &ge_root_model, const GraphNodePtr &graph_node) {
   GELOGI("[LoadGraphAsync] run_graph_flag[%d], graph_id[%u]", options_.run_graph_flag, graph_node->GetGraphId());
@@ -2185,6 +2062,19 @@ Status GraphManager::IncreBuild(const GraphNodePtr &graph_node, GeModelPtr &ge_m
   return FAILED;
 }
 
+void GraphManager::ConstructGeInput(std::vector<ge::GeTensor> &ge_inputs, PreRunArgs &args) {
+  for (auto const &input : args.input_tensor) {
+    std::vector<int64_t> input_dims;
+    std::transform(input.dims.begin(), input.dims.end(), std::back_inserter(input_dims),
+                   [](int64_t x) -> int64_t { return x; });
+    GeShape input_shape(input_dims);
+    GeTensorDesc input_tensor_desc;
+    input_tensor_desc.SetShape(input_shape);
+    input_tensor_desc.SetDataType(static_cast<ge::DataType>(input.data_type));
+    ge_inputs.emplace_back(input_tensor_desc);
+  }
+}
+
 void GraphManager::PreRunThread(GraphManager *graph_manager) {
   if (prctl(PR_SET_NAME, ("GE_PreRun")) != 0) {
     GELOGW("Set thread name failed.");
@@ -2198,16 +2088,8 @@ void GraphManager::PreRunThread(GraphManager *graph_manager) {
     GetThreadLocalContext() = args.context;
     GELOGI("A new loop start.");
     std::vector<ge::GeTensor> ge_inputs;
-    for (auto const &input : args.input_tensor) {
-      std::vector<int64_t> input_dims;
-      std::transform(input.dims.begin(), input.dims.end(), std::back_inserter(input_dims),
-                     [](int64_t x) -> int64_t { return x; });
-      GeShape input_shape(input_dims);
-      GeTensorDesc input_tensor_desc;
-      input_tensor_desc.SetShape(input_shape);
-      input_tensor_desc.SetDataType(static_cast<ge::DataType>(input.data_type));
-      ge_inputs.emplace_back(input_tensor_desc);
-    }
+    ConstructGeInput(ge_inputs, args);
+
     // find graph
     GraphNodePtr graph_node = nullptr;
     Status ret = graph_manager->GetGraphNode(args.graph_id, graph_node);
@@ -2229,14 +2111,11 @@ void GraphManager::PreRunThread(GraphManager *graph_manager) {
     graph_node->SetRunFlag(true);
 
     ComputeGraphPtr compute_graph_tmp = GraphUtils::GetComputeGraph(*(graph_node->GetGraph()));
-
-    if (graph_manager->GetTrainFlag()) {
-      if (compute_graph_tmp == nullptr) {
-        ReturnError(graph_manager, args.callback, GE_GRAPH_GRAPH_NODE_NULL,
-                    "[RunGraph] compute_graph_tmp is NULL, graph id = %u.");
-        graph_node->Unlock();
-        return;
-      }
+    if (compute_graph_tmp == nullptr) {
+      ReturnError(graph_manager, args.callback, GE_GRAPH_GRAPH_NODE_NULL,
+                  "[RunGraph] compute_graph_tmp is NULL, graph id = %u.");
+      graph_node->Unlock();
+      return;
     }
     // when set incre build, save cache helper.
     graph_manager->AddModelCacheHelperToMap(args.graph_id, args.session_id, compute_graph_tmp);
@@ -2266,11 +2145,19 @@ void GraphManager::PreRunThread(GraphManager *graph_manager) {
       GeModelPtr ge_model = nullptr;
       if (graph_manager->IncreBuild(graph_node, ge_model) != SUCCESS) {
         ret = graph_manager->PreRun(graph_node, ge_inputs, ge_root_model, args.session_id);
+        // release rts generate context
+        RtContextUtil::GetInstance().DestroyRtContexts(args.session_id);
         if (ret != SUCCESS) {
           graph_node->SetRunFlag(false);
-          ReturnError(graph_manager, args.callback, ret, "PreRun Failed, thread exit..");
-          graph_node->Unlock();
-          return;
+          if (!std::getenv("AnalyzeMode")) {
+            ReturnError(graph_manager, args.callback, ret, "PreRun Failed, thread exit..");
+            graph_node->Unlock();
+            return;
+          } else {
+            ReturnError(graph_manager, graph_node, args.callback, ret, "PreRun Failed, keep geop continue!");
+            graph_node->Unlock();
+            continue;
+          }
         }
       }
       graph_node->SetBuildFlag(true);
@@ -2350,13 +2237,74 @@ void GraphManager::ReturnError(GraphManager *graph_manager, RunAsyncCallback cal
   if (graph_manager == nullptr) {
     return;
   }
-
-  GELOGE(ret, "%s.", log.c_str());
   StopQueue(graph_manager);
+  GELOGE(ret, "%s.", log.c_str());
   std::vector<ge::OutputTensorInfo> outputs;
   callback(ret, outputs);
 }
 
+void GraphManager::ReturnError(GraphManager *graph_manager, GraphNodePtr &graph_node, RunAsyncCallback callback,
+                               Status ret, const string &log) {
+  std::vector<ge::OutputTensorInfo> outputs;
+  auto compute_graph = GraphUtils::GetComputeGraph(*graph_node->GetGraph());
+  if (graph_manager == nullptr || compute_graph == nullptr) {
+    GELOGE(GRAPH_FAILED, "[Analyze Mode] compute graph is null!");
+    callback(GRAPH_FAILED, outputs);
+    return;
+  }
+
+  for (const auto &node : compute_graph->GetAllNodes()) {
+    if (node->GetType() != "NetOutput") {
+      continue;
+    }
+    for (size_t i = 0; i < node->GetAllInDataAnchorsSize(); i++) {
+      auto input_desc = node->GetOpDesc()->MutableInputDesc(i);
+      ge::OutputTensorInfo tensor;
+      tensor.dims = input_desc->GetShape().GetDims();
+      tensor.data_type = static_cast<uint32_t>(input_desc->GetDataType());
+      int64_t len = 1;
+      if (input_desc->GetShape().GetDims() != std::vector<int64_t>({})) {
+        len = input_desc->GetShape().GetShapeSize();
+      }
+      if (len < 0) {
+        GELOGE(GRAPH_FAILED, "Analyze Mode does not support GEOP output unknown shape!");
+        callback(GRAPH_FAILED, outputs);
+        return;
+      } else if (len == 0) {
+        GELOGI("getted shape size is 0.Do process as empty tensor!");
+        len = 1;
+      }
+      auto size = GetSizeByDataType(input_desc->GetDataType());
+      if (size <= 0) {
+        GELOGE(PARAM_INVALID, "Failed to get cube size, the data type %s is invalid",
+               ge::TypeUtils::DataTypeToSerialString(input_desc->GetDataType()).c_str());
+        callback(GRAPH_FAILED, outputs);
+        return;
+      }
+      if (CheckInt64MulOverflow(len, static_cast<int64_t>(size)) != true) {
+        GELOGE(MEMALLOC_FAILED, "int64 multiply happens overflow! a:%ld b:%d", len, size);
+        callback(GRAPH_FAILED, outputs);
+        return;
+      }
+      tensor.length = len * size;
+      auto pbuff = new (std::nothrow) uint8_t[tensor.length];
+      if (!pbuff) {
+        GELOGE(MEMALLOC_FAILED, "new buff failed!");
+        callback(GRAPH_FAILED, outputs);
+        return;
+      }
+      // To avoid global step too small and can not stop, totally set a bigger value
+      for (int64_t i = 0; i < tensor.length; i++) {
+        *(pbuff + i) = 0x7F;  // here stands for a positive max value
+      }
+      tensor.data.reset(pbuff);
+      outputs.emplace_back(std::move(tensor));
+    }
+  }
+  callback(SUCCESS, outputs);
+  return;
+}
+
 bool GraphManager::IsGraphNeedRebuild(uint32_t graph_id) {
   // find graph
   GraphNodePtr graph_node = nullptr;
@@ -2479,4 +2427,99 @@ Status GraphManager::Build(const GraphNodePtr &graph_node, ComputeGraphPtr &comp
   graph_node->SetGeRootModel(ge_root_model);
   return SUCCESS;
 }
+
+Status GraphManager::GenCheckPointGraph(const std::map<std::string, GeTensorDesc> &all_variables, Graph &graph) {
+  ge::ComputeGraphPtr compute_graph = MakeShared<ComputeGraph>(kCheckPointGraph);
+  GE_CHECK_NOTNULL(compute_graph);
+  OpDescPtr save_desc = MakeShared<ge::OpDesc>(compute_graph->GetName() + "_" + kSave, kSave);
+  GE_CHECK_NOTNULL(save_desc);
+  uint32_t save_index = 0;
+  for (auto iter = all_variables.begin(); iter != all_variables.end(); ++iter) {
+    GE_CHK_GRAPH_STATUS_RET(save_desc->AddInputDesc(save_index, iter->second));
+    save_index++;
+  }
+  NodePtr save_node = compute_graph->AddNode(save_desc);
+
+  uint32_t index = 0;
+  for (auto iter = all_variables.begin(); iter != all_variables.end(); ++iter) {
+    OpDescPtr var_desc = MakeShared<ge::OpDesc>(iter->first, VARIABLE);
+    GE_CHECK_NOTNULL(var_desc);
+    if (!AttrUtils::SetBool(var_desc, kCheckPointForGetVar, true)) {
+      GELOGW("Set check point graph attr failed.");
+    }
+    GE_CHK_GRAPH_STATUS_RET(var_desc->AddOutputDesc(iter->second));
+    NodePtr var_node = compute_graph->AddNode(var_desc);
+    GE_CHK_STATUS(GraphUtils::AddEdge(var_node->GetOutDataAnchor(0), save_node->GetInDataAnchor(index)),
+                  "Add edge[%s->%s] fail.", var_node->GetName().c_str(), save_node->GetName().c_str());
+    index++;
+  }
+  compute_graph->Dump();
+  graph = GraphUtils::CreateGraphFromComputeGraph(compute_graph);
+  return SUCCESS;
+}
+
+Status GraphManager::SaveVariables(const Graph &graph, const std::vector<std::string> &var_names,
+                                   const std::vector<Tensor> &outputs, std::vector<Tensor> &var_values) {
+  map<string, Tensor> var_results;
+  GE_CHK_STATUS_RET(SaveCheckPointResult(graph, outputs, var_results), "Save check point result failed.");
+  if (!var_names.empty()) {
+    for (const auto &var_name : var_names) {
+      if (var_results.count(var_name) == 0) {
+        GELOGE(FAILED, "Fetch var[%s] value failed.", var_name.c_str());
+        return FAILED;
+      } else {
+        var_values.emplace_back(var_results[var_name]);
+      }
+    }
+  } else {
+    for (auto iter = var_results.begin(); iter != var_results.end(); ++iter) {
+      var_values.emplace_back(iter->second);
+    }
+  }
+  return SUCCESS;
+}
+
+Status GraphManager::SaveCheckPointResult(const Graph &graph, const std::vector<Tensor> &outputs,
+                                          map<string, Tensor> &var_results) {
+  auto compute_graph = GraphUtils::GetComputeGraph(graph);
+  NodePtr netoutput_node = nullptr;
+  for (const auto &node : compute_graph->GetAllNodes()) {
+    if (node->GetType() == NETOUTPUT) {
+      netoutput_node = node;
+      break;
+    }
+  }
+  GE_CHECK_NOTNULL(netoutput_node);
+  for (const auto &in : netoutput_node->GetAllInDataAnchors()) {
+    auto out_anchor = in->GetPeerOutAnchor();
+    GE_CHECK_NOTNULL(out_anchor);
+    auto peer_node = out_anchor->GetOwnerNode();
+    while (peer_node->GetType() != VARIABLE) {
+      if (peer_node->GetAllInDataAnchors().size() != 1) {
+        GELOGE(FAILED, "peer_node [%s] has more than 1 input in checkpoint Graph.", peer_node->GetName().c_str());
+        return FAILED;
+      }
+      auto peer_node_in_anchor = peer_node->GetAllInDataAnchors().at(0);
+      auto peer_node_out_anchor = peer_node_in_anchor->GetPeerOutAnchor();
+      if (peer_node_out_anchor != nullptr) {
+        peer_node = peer_node_out_anchor->GetOwnerNode();
+        if (peer_node->GetType() == VARIABLE) {
+          break;
+        }
+      }
+    }
+    if (peer_node->GetType() != VARIABLE) {
+      GELOGE(FAILED, " peer_node %s is not variable in checkpoint Graph.", peer_node->GetName().c_str());
+      return FAILED;
+    }
+    auto var_name = peer_node->GetName();
+    GELOGI("[GraphManager] SaveVariables, varName is %s.", var_name.c_str());
+    if (in->GetIdx() >= static_cast<int>(outputs.size())) {
+      GELOGE(FAILED, "variable index[%d] out of range[%zu].", in->GetIdx(), outputs.size());
+      return FAILED;
+    }
+    var_results.emplace(var_name, outputs.at(in->GetIdx()));
+  }
+  return SUCCESS;
+}
 }  // namespace ge

src/ge/graph/manager/graph_manager.h

@@ -159,6 +159,13 @@ class GraphManager {
 
   void SetOptionsRunGraphFlag(bool run_graph_flag);
 
+  Status GenCheckPointGraph(const std::map<std::string, GeTensorDesc> &all_variables, Graph &graph);
+
+  Status SaveVariables(const Graph &graph, const std::vector<std::string> &var_names,
+                       const std::vector<Tensor> &outputs, std::vector<Tensor> &var_values);
+
+  Status SaveCheckPointResult(const Graph &graph, const std::vector<Tensor> &outputs, map<string, Tensor> &var_results);
+
  private:
   struct PreRunArgs {
     GraphId graph_id;
@@ -267,9 +274,8 @@ class GraphManager {
 
   Status OptimizeStage1(ComputeGraphPtr &compute_graph);
   Status OptimizeStage2(ComputeGraphPtr &compute_graph);
-  Status OptimizeAfterMergeSubGraph(ge::ComputeGraphPtr &compute_graph);
 
-  Status NewOptimizeAfterMergeSubGraph(ge::ComputeGraphPtr &compute_graph);
+  Status SubexpressionMigration(ComputeGraphPtr &compute_graph);
 
   Status LoadGraphAsync(const GeRootModelPtr &ge_root_model, const GraphNodePtr &graph_node);
 
@@ -288,10 +294,13 @@ class GraphManager {
   Status IncreBuild(const GraphNodePtr &graph_node, GeModelPtr &ge_model);
   void RemoveModelCacheHelper(const GraphId &graph_id);
 
+  static void ConstructGeInput(std::vector<ge::GeTensor> &ge_inputs, PreRunArgs &args);
   static void PreRunThread(GraphManager *graph_manager);
   static void RunThread(GraphManager *graph_manager);
   static void StopQueue(GraphManager *graph_manager);
   static void ReturnError(GraphManager *graph_manager, RunAsyncCallback callback, Status ret, const string &log);
+  static void ReturnError(GraphManager *graph_manager, GraphNodePtr &graph_node, RunAsyncCallback callback, Status ret,
+                          const string &log);
 
   void ChangeConstTypeWhenTraining(const ComputeGraphPtr &compute_graph);
 

src/ge/graph/manager/graph_var_manager.cc

@@ -855,6 +855,32 @@ void VarManager::RemoveAllocatedGraphId(const std::string &var_name) {
   var_resource_->RemoveAllocatedGraphId(var_name);
 }
 
+Status VarManager::GetAllVariables(std::map<std::string, GeTensorDesc> &all_variables) {
+  std::lock_guard<std::recursive_mutex> lock(mutex_);
+  if (var_resource_ == nullptr) {
+    GELOGW("VarManager has not been inited.");
+    return INTERNAL_ERROR;
+  }
+  auto new_variable_desc = var_resource_->GetAllVarDesc();
+  if (new_variable_desc.size() == 0) {
+    GELOGW("VarManager don't have variables.");
+    return INTERNAL_ERROR;
+  }
+
+  for (auto iter = new_variable_desc.begin(); iter != new_variable_desc.end(); ++iter) {
+    auto trans_road = var_resource_->GetTransRoad(iter->first);
+    if (trans_road == nullptr || trans_road->empty()) {
+      GELOGI("The variable %s does not have any trans road", iter->first.c_str());
+      all_variables[iter->first] = iter->second;
+      continue;
+    }
+    // get origin trans info : the first trans node info
+    auto origin_trans_node_info = trans_road->at(0);
+    all_variables[iter->first] = origin_trans_node_info.input;
+  }
+  return SUCCESS;
+}
+
 VarManagerPool::~VarManagerPool() { Destory(); }
 
 VarManagerPool &VarManagerPool::Instance() {
@@ -897,4 +923,22 @@ VarManager *VarManagerPool::GetVarManager(uint64_t session_id) {
   var_manager_map_[session_id] = var_manager;
   return var_manager;
 }
+
+void VarManagerPool::RemoveVarManager(uint64_t session_id) {
+  VarManager *var_manager = nullptr;
+  {
+    std::lock_guard<std::mutex> lock(var_manager_mutex_);
+    auto it = var_manager_map_.find(session_id);
+    if (it != var_manager_map_.end()) {
+      var_manager = it->second;
+      var_manager_map_.erase(it);
+    }
+  }
+
+  if (var_manager != nullptr) {
+    var_manager->Destory();
+    delete var_manager;
+    var_manager = nullptr;
+  }
+}
 }  // namespace ge

src/ge/graph/manager/graph_var_manager.h

@@ -157,6 +157,8 @@ class VarResource {
 
   bool IsVarAddr(const int64_t &offset);
 
+  std::unordered_map<std::string, ge::GeTensorDesc> GetAllVarDesc() const { return cur_var_tensor_desc_map_; }
+
  private:
   std::string VarKey(const std::string &var_name, const ge::GeTensorDesc &tensor_desc);
 
@@ -276,6 +278,8 @@ class FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY VarManager {
 
   uint8_t *GetVarMemoryAddr(uint8_t *logic_addr, rtMemType_t memory_type);
 
+  Status GetAllVariables(std::map<std::string, GeTensorDesc> &all_variables);
+
  private:
   uint32_t version_;
   uint64_t session_id_;
@@ -300,6 +304,8 @@ class VarManagerPool {
 
   VarManager *GetVarManager(uint64_t session_id);
 
+  void RemoveVarManager(uint64_t session_id);
+
   void Destory() noexcept;
 
   ge::Status Init() const;

src/ge/graph/manager/rdma_pool_allocator.cc

@@ -16,7 +16,6 @@
 
 #include "graph/manager/rdma_pool_allocator.h"
 #include "framework/common/debug/ge_log.h"
-#include "graph/manager/graph_mem_allocator.h"
 
 namespace {
 const size_t kAlignedSize = 512;

src/ge/graph/manager/rdma_pool_allocator.h

@@ -27,12 +27,11 @@
 
 #include "framework/common/ge_inner_error_codes.h"
 #include "graph/manager/block_memory.h"
+#include "graph/manager/graph_mem_allocator.h"
 #include "graph/node.h"
 #include "runtime/mem.h"
 
 namespace ge {
-class MemoryAllocator;
-
 class RdmaPoolAllocator {
  public:
   explicit RdmaPoolAllocator(rtMemType_t memory_type);