update RELEASE.md.

From: @shenwei41
Reviewed-by: @liucunwei,@xsmq
Signed-off-by: @liucunwei

.gitmodules

@@ -1,8 +1,8 @@
 [submodule "parser"]
 	path = parser
 	url = https://gitee.com/ascend/parser.git
-	branch = master
+	branch = r1.3.0
 [submodule "metadef"]
 	path = metadef
 	url = https://gitee.com/ascend/metadef.git
-	branch = master
+	branch = r1.3.0

RELEASE.md

@@ -1,3 +1,18 @@
+### Major Features and Improvements
+* Multiple parallel communication groups can be distinguished and the communication tasks of different parallel communication groups can be divided into different streams.
+* Parallel tasks are added for the entire map optimization. The execution sequence is optimized for parallel groups.
+* Dynamic shape is supported for single operator in inference scenarios.
+* Online inference supports concurrent execution of multiple threads.
+* Memory allocation supports address reuse in the buffer pool.
+* Supports Event resource reuse.
+* Supports the BF16 data type.
+
+## Thanks to our Contributors
+Thanks goes to these wonderful people: wuweikang，weiyang，yanghaorang，xutianchun，shibeiji，zhouchao, tanghuikang, zhoulili, liujunzhu, zhengyuanhua, taoxiangdong Contributions of any kind are welcome!
+
+Contributions of any kind are welcome!
+
+
 # Release 1.0.0
 
 ## Major Features and Improvements

build.sh

@@ -229,7 +229,7 @@ if [[ "X$ENABLE_GE_UT" = "Xon" || "X$ENABLE_GE_COV" = "Xon" ]]; then
     rm -rf ${BASEPATH}/cov
     mkdir ${BASEPATH}/cov
     lcov -c -d build/tests/ut/ge -d build/tests/ut/common/graph/ -o cov/tmp.info
-    lcov -r cov/tmp.info '*/output/*' '*/build/opensrc/*' '*/build/proto/*' '*/third_party/*' '*/tests/*' '/usr/local/*' -o cov/coverage.info
+    lcov -r cov/tmp.info '*/output/*' '*/build/opensrc/*' '*/build/proto/*' '*/third_party/*' '*/tests/*' '/usr/local/*' '/usr/include/*' '*/metadef/*' '*/parser/*' -o cov/coverage.info
     cd ${BASEPATH}/cov
     genhtml coverage.info
 fi

ge/CMakeLists.txt

@@ -31,6 +31,7 @@ set(PROTO_HEADER_LIST
 protobuf_generate(ge PROTO_SRCS PROTO_HDRS ${PROTO_LIST})
 protobuf_generate(ge PROTO_CLIENT_SRCS PROTO_CLIENT_HDRS ${PROTO_CLIENT_LIST})
 protobuf_generate(ge PROTO_HEADER_SRCS PROTO_HEADER_HDRS ${PROTO_HEADER_LIST})
+protobuf_generate(ge_client PROTO_CLIENT_HEADER_SRCS PROTO_CLIENT_HEADER_HDRS ${PROTO_HEADER_LIST})
 
 if (NOT ENABLE_D AND NOT ENABLE_ACL AND NOT ENABLE_MS_TESTCASES)
 ############ libge_proto_common.a ############
@@ -56,7 +57,7 @@ target_link_libraries(ge_proto_common PRIVATE
 
 ############ libge_proto_client.a ############
 add_library(ge_proto_client STATIC
-    ${PROTO_HEADER_HDRS}
+    ${PROTO_CLIENT_HEADER_HDRS}
     ${PROTO_CLIENT_SRCS}
 )
 
@@ -65,6 +66,11 @@ target_compile_definitions(ge_proto_client PRIVATE
     google=ascend_private
 )
 
+target_include_directories(ge_proto_client PRIVATE
+    ${CMAKE_BINARY_DIR}/proto/ge_client
+    ${CMAKE_BINARY_DIR}/proto/ge_client/proto
+)
+
 target_compile_options(ge_proto_client PRIVATE
     -O2
     -fno-common
@@ -937,6 +943,10 @@ add_library(atc_stub_ge_compiler SHARED
 
 add_dependencies(atc_stub_ge_compiler ge_stub)
 
+target_compile_options(atc_stub_ge_compiler PRIVATE
+    -fno-common
+)
+
 target_link_libraries(atc_stub_ge_compiler PRIVATE
     $<BUILD_INTERFACE:intf_pub>
 )

ge/analyzer/analyzer.cc

@@ -221,7 +221,10 @@ ge::Status Analyzer::SaveAnalyzerDataToFile(uint64_t session_id, uint64_t graph_
   try {
     json_file_ << jsn.dump(kJsonDumpLevel) << std::endl;
   } catch (nlohmann::detail::type_error &e) {
-    GELOGE(FAILED, "[Json.dump][GraphInfo]json.dump to analyze file [%s] failed because [%s], session_id:%lu, graph_id:%lu", json_file_name_.c_str(), e.what(), session_id, graph_id);
+    GELOGE(FAILED,
+           "[Json.dump][GraphInfo]json.dump to analyze file [%s] failed because [%s],"
+	         "session_id:%lu, graph_id:%lu",
+           json_file_name_.c_str(), e.what(), session_id, graph_id);
     ret_failed = true;
   }
   json_file_.close();
@@ -241,7 +244,9 @@ ge::Status Analyzer::DoAnalyze(DataInfo &data_info) {
   GE_CHECK_NOTNULL(graph_info);
   auto status = SaveOpInfo(desc, data_info, graph_info);
   if (status != SUCCESS) {
-    GELOGE(status, "[Check][SaveOpInfo]save op info: desc_name [%s] desc_type [%s] failed!", desc->GetName().c_str(), desc->GetType().c_str());
+    GELOGE(status,
+           "[Check][SaveOpInfo]save op info: desc_name [%s] desc_type [%s] failed!",
+           desc->GetName().c_str(), desc->GetType().c_str());
     return FAILED;
   }
   // create json file

ge/client/ge_api.cc

@@ -171,17 +171,17 @@ Status GEInitialize(const std::map<AscendString, AscendString> &options) {
 
 // GE finalize, releasing all resources
 Status GEFinalize() {
-  ErrorManager::GetInstance().SetStage(ErrorMessage::kFinalize, ErrorMessage::kFinalize);
-  GELOGT(TRACE_INIT, "GEFinalize start");
-
-  ErrorManager::GetInstance().GenWorkStreamIdDefault();
+  std::lock_guard<std::mutex> lock(g_ge_release_mutex);
   // check init status
   if (!g_ge_initialized) {
-    GELOGW("GEFinalize is called before GEInitialize");
+    GELOGW("[FINAL][FINAL]GEFinalize is called before GEInitialize");
     return SUCCESS;
   }
 
-  std::lock_guard<std::mutex> lock(g_ge_release_mutex);
+  ErrorManager::GetInstance().SetStage(ErrorMessage::kFinalize, ErrorMessage::kFinalize);
+  ErrorManager::GetInstance().GenWorkStreamIdDefault();
+  GELOGT(TRACE_INIT, "GEFinalize start");
+
   // call Finalize
   Status ret = SUCCESS;
   Status middle_ret;

ge/common/CMakeLists.txt

@@ -16,6 +16,7 @@ set(PROTO_LIST
 )
 
 protobuf_generate(ge PROTO_SRCS PROTO_HDRS ${PROTO_LIST})
+protobuf_generate(ge_static PROTO_STATIC_SRCS PROTO_STATIC_HDRS ${PROTO_LIST})
 
 set(SRC_LIST
     "context/ctx.cc"
@@ -127,7 +128,7 @@ target_link_libraries(ge_common PRIVATE
 )
 
 ############ libge_common.a ############
-add_library(ge_common_static STATIC ${SRC_LIST} ${PROTO_HDRS})
+add_library(ge_common_static STATIC ${SRC_LIST} ${PROTO_STATIC_HDRS})
 target_compile_definitions(ge_common_static PRIVATE
     PROTOBUF_INLINE_NOT_IN_HEADERS=0
     HOST_VISIBILITY
@@ -158,7 +159,7 @@ target_include_directories(ge_common_static PRIVATE
     ${METADEF_DIR}/inc/external/graph
     ${METADEF_DIR}/inc/graph
     ${CMAKE_BINARY_DIR}
-    ${CMAKE_BINARY_DIR}/proto/ge
+    ${CMAKE_BINARY_DIR}/proto/ge_static
     #### yellow zone ####
     ${GE_DEPEND_DIR}/inc
     ${GE_DEPEND_DIR}/inc/cce

ge/common/dump/dump_manager.cc

@@ -96,7 +96,7 @@ FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY Status DumpManager::SetDumpConf
   dump_mode = dump_config.dump_mode;
   GELOGI("Dump mode is %s", dump_mode.c_str());
   dump_properties.SetDumpMode(dump_mode);
-  dump_properties_map_.emplace(kInferSessionId, dump_properties);
+  dump_properties_map_[kInferSessionId] = dump_properties;
 
   return SUCCESS;
 }

ge/common/dump/dump_op.cc

@@ -20,6 +20,7 @@
 #include "common/ge/datatype_util.h"
 #include "framework/common/debug/ge_log.h"
 #include "framework/common/util.h"
+#include "framework/common/types.h"
 #include "graph/anchor.h"
 #include "graph/ge_tensor.h"
 #include "graph/op_desc.h"
@@ -55,8 +56,10 @@ void DumpOp::SetLoopAddr(void *global_step, void *loop_per_iter, void *loop_cond
   loop_cond_ = reinterpret_cast<uintptr_t>(loop_cond);
 }
 
-void DumpOp::SetDynamicModelInfo(const string &dynamic_model_name, uint32_t dynamic_model_id) {
+void DumpOp::SetDynamicModelInfo(const string &dynamic_model_name, const string &dynamic_om_name,
+                                 uint32_t dynamic_model_id) {
   dynamic_model_name_ = dynamic_model_name;
+  dynamic_om_name_ = dynamic_om_name;
   dynamic_model_id_ = dynamic_model_id;
 }
 
@@ -200,6 +203,32 @@ Status DumpOp::ExecutorDumpOp(aicpu::dump::OpMappingInfo &op_mapping_info) {
   return SUCCESS;
 }
 
+Status DumpOp::SetDumpModelName(aicpu::dump::OpMappingInfo &op_mapping_info) {
+  if (dynamic_model_name_.empty() && dynamic_om_name_.empty()) {
+    GELOGI("Single op dump, no need set model name");
+    return SUCCESS;
+  }
+  std::set<std::string> model_list = dump_properties_.GetAllDumpModel();
+  bool not_find_by_omname = model_list.find(dynamic_om_name_) == model_list.end();
+  bool not_find_by_modelname = model_list.find(dynamic_model_name_) == model_list.end();
+  std::string dump_model_name = not_find_by_omname ? dynamic_model_name_ : dynamic_om_name_;
+  if (model_list.find(DUMP_ALL_MODEL) == model_list.end()) {
+    if (not_find_by_omname && not_find_by_modelname) {
+      std::string model_list_str;
+      for (auto &model : model_list) {
+        model_list_str += "[" + model + "].";
+      }
+      GELOGW("Model %s will not be set to dump, dump list: %s", dump_model_name.c_str(), model_list_str.c_str());
+      return FAILED;
+    }
+  }
+  if (!dump_model_name.empty() && dump_properties_.IsDumpOpen()) {
+    GELOGD("Dump model name is %s", dump_model_name.c_str());
+    op_mapping_info.set_model_name(dump_model_name);
+  }
+  return SUCCESS;
+}
+
 Status DumpOp::LaunchDumpOp() {
   GELOGI("Start to launch dump op %s", op_desc_->GetName().c_str());
   int32_t device_id = 0;
@@ -209,8 +238,7 @@ Status DumpOp::LaunchDumpOp() {
     return RT_ERROR_TO_GE_STATUS(rt_ret);
   }
   if (device_id < 0) {
-    GELOGE(ACL_ERROR_GE_INTERNAL_ERROR,
-           "Check device_id failed, device_id = %d, which should be not less than 0.",
+    GELOGE(ACL_ERROR_GE_INTERNAL_ERROR, "Check device_id failed, device_id = %d, which should be not less than 0.",
            device_id);
     return ACL_ERROR_GE_INTERNAL_ERROR;
   }
@@ -220,11 +248,12 @@ Status DumpOp::LaunchDumpOp() {
   op_mapping_info.set_flag(kAicpuLoadFlag);
   op_mapping_info.set_dump_step(dump_properties_.GetDumpStep());
   op_mapping_info.set_model_id(dynamic_model_id_);
-  if (!dynamic_model_name_.empty() && dump_properties_.IsDumpOpen()) {
-    op_mapping_info.set_model_name(dynamic_model_name_);
+
+  if (SetDumpModelName(op_mapping_info) != SUCCESS) {
+    return SUCCESS;
   }
   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(),
+  GELOGI("Dump step is %s ,dump path is %s in Launch dump op", dump_properties_.GetDumpStep().c_str(),
          dump_path.c_str());
   uint32_t task_id = 0;
   uint32_t stream_id = 0;
@@ -273,4 +302,4 @@ Status DumpOp::LaunchDumpOp() {
   }
   return SUCCESS;
 }
-}  // namesapce ge
+}  // namespace ge

ge/common/dump/dump_op.h

@@ -34,12 +34,13 @@ class DumpOp {
                    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);
+  void SetDynamicModelInfo(const string &dynamic_model_name, const string &dynamic_om_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);
+  Status SetDumpModelName(aicpu::dump::OpMappingInfo &op_mapping_info);
 
   DumpProperties dump_properties_;
   OpDescPtr op_desc_;
@@ -54,6 +55,7 @@ class DumpOp {
   uintptr_t loop_cond_;
 
   std::string dynamic_model_name_;
+  std::string dynamic_om_name_;
   std::uint32_t dynamic_model_id_;
 };
 }  // namespace ge

ge/common/dump/dump_properties.cc

@@ -35,14 +35,14 @@ const std::string kDumpStatusOpen = "on";
 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) {
+  const DumpProperties &other) {
   CopyFrom(other);
   return *this;
 }
@@ -97,7 +97,7 @@ FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY void DumpProperties::InitByOpti
 
 // 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) {
+  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());
   }
@@ -138,7 +138,7 @@ FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY std::set<std::string> DumpPrope
 }
 
 FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY std::set<std::string> DumpProperties::GetPropertyValue(
-    const std::string &model) const {
+  const std::string &model) const {
   auto iter = model_dump_properties_map_.find(model);
   if (iter != model_dump_properties_map_.end()) {
     return iter->second;
@@ -147,8 +147,9 @@ FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY std::set<std::string> DumpPrope
 }
 
 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 {
+  const std::string &model, const std::string &om_name, const std::string &op_name) const {
   // if dump all
+  GELOGD("model name is %s om name is %s op is %s in layer need dump", model.c_str(), om_name.c_str(), op_name.c_str());
   if (model_dump_properties_map_.find(DUMP_ALL_MODEL) != model_dump_properties_map_.end()) {
     return true;
   }
@@ -203,7 +204,7 @@ FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY const std::string &DumpProperti
 }
 
 FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY void DumpProperties::SetDumpOpSwitch(
-    const std::string &dump_op_switch) {
+  const std::string &dump_op_switch) {
   dump_op_switch_ = dump_op_switch;
 }
 
@@ -270,4 +271,4 @@ void DumpProperties::SetDumpDebugOptions() {
     GELOGI("ge.exec.enableDumpDebug is false or is not set.");
   }
 }
-}  // namespace
+}  // namespace ge

ge/common/dump/opdebug_register.cc

@@ -80,13 +80,11 @@ Status OpdebugRegister::RegisterDebugForStream(rtStream_t stream, uint32_t op_de
 
   uint32_t debug_stream_id = 0;
   uint32_t debug_task_id = 0;
-#ifdef ONLY_COMPILE_OPEN_SRC
   auto rt_ret = rtDebugRegisterForStream(stream, op_debug_mode, op_debug_addr_, &debug_stream_id, &debug_task_id);
   if (rt_ret != RT_ERROR_NONE) {
     GELOGE(RT_FAILED, "rtDebugRegisterForStream error, ret: 0x%X", rt_ret);
     return RT_ERROR_TO_GE_STATUS(rt_ret);
   }
-#endif
   GELOGD("debug_task_id:%u, debug_stream_id:%u in stream overflow.", debug_task_id, debug_stream_id);
   data_dumper.SaveOpDebugId(debug_task_id, debug_stream_id, p2p_debug_addr_, true);
   return SUCCESS;
@@ -94,7 +92,6 @@ Status OpdebugRegister::RegisterDebugForStream(rtStream_t stream, uint32_t op_de
 
 void OpdebugRegister::UnregisterDebugForStream(rtStream_t stream) {
   rtError_t rt_ret = RT_ERROR_NONE;
-#ifdef ONLY_COMPILE_OPEN_SRC
   if (stream != nullptr) {
     GELOGD("start call rtDebugUnRegisterForStream in unknown shape over flow.");
     rt_ret = rtDebugUnRegisterForStream(stream);
@@ -102,8 +99,6 @@ void OpdebugRegister::UnregisterDebugForStream(rtStream_t stream) {
       GELOGW("rtDebugUnRegisterForStream failed, ret: 0x%X", rt_ret);
     }
   }
-#endif
-
   if (op_debug_addr_ != nullptr) {
     rt_ret = rtFree(op_debug_addr_);
     if (rt_ret != RT_ERROR_NONE) {

ge/common/formats/format_transfers/datatype_transfer.cc

@@ -154,7 +154,8 @@ Status DataTypeTransfer::TransDataType(const CastArgs &args, TransResult &result
 
   std::shared_ptr<uint8_t> dst(new (std::nothrow) uint8_t[total_size], std::default_delete<uint8_t[]>());
   if (dst == nullptr) {
-    GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION, "Failed to alloc the memory for dst buf %zu, data size %zu", total_size, args.src_data_size);
+    GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION,
+           "Failed to alloc the memory for dst buf %zu, data size %zu", total_size, args.src_data_size);
     return ACL_ERROR_GE_MEMORY_ALLOCATION;
   }
 

ge/common/formats/format_transfers/format_transfer_c1hwncoc0_hwcn.cc

@@ -73,7 +73,8 @@ Status CheckArgsForC1hwncoc0ToHwcn(const TransArgs &args) {
 Status GetDstDataAfterTrans(const TransArgs &args, TransResult &result, int size, int64_t total_size) {
   std::shared_ptr<uint8_t> dst(new (std::nothrow) uint8_t[total_size], std::default_delete<uint8_t[]>());
   if (dst == nullptr) {
-    GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION, "Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld, shape %s",
+    GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION,
+           "Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld, shape %s",
            TypeUtils::FormatToSerialString(args.src_format).c_str(),
            TypeUtils::FormatToSerialString(args.dst_format).c_str(), total_size, ShapeToString(args.dst_shape).c_str());
     return ACL_ERROR_GE_MEMORY_ALLOCATION;

ge/common/formats/format_transfers/format_transfer_dhwcn_fracz3D.cc

@@ -94,7 +94,8 @@ Status TransFormatDhwckToFz3D(const TransArgs &args, TransResult &result) {
 
   std::shared_ptr<uint8_t> dst(new (std::nothrow) uint8_t[dst_size], std::default_delete<uint8_t[]>());
   if (dst == nullptr) {
-    GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION, "Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld",
+    GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION,
+           "Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld",
            TypeUtils::FormatToSerialString(args.src_format).c_str(),
            TypeUtils::FormatToSerialString(args.dst_format).c_str(), dst_size);
     return ACL_ERROR_GE_MEMORY_ALLOCATION;
@@ -122,7 +123,8 @@ Status TransFormatDhwckToFz3D(const TransArgs &args, TransResult &result) {
                                args.data + src_idx * data_size, static_cast<size_t>(data_size));
               }
               if (ret != EOK) {
-                GELOGE(ACL_ERROR_GE_MEMORY_OPERATE_FAILED, "Failed to operate the dst memory at offset %ld, error-code %d, pad mode %d",
+                GELOGE(ACL_ERROR_GE_MEMORY_OPERATE_FAILED,
+                       "Failed to operate the dst memory at offset %ld, error-code %d, pad mode %d",
                        dst_offset, ret, pad_zero);
                 return ACL_ERROR_GE_MEMORY_OPERATE_FAILED;
               }

ge/common/formats/format_transfers/format_transfer_dhwnc_fracz3D_transpose.cc

@@ -95,7 +95,8 @@ Status TransFormatDhwncToFz3DTranspose(const TransArgs &args, TransResult &resul
 
   std::shared_ptr<uint8_t> dst(new (std::nothrow) uint8_t[dst_size], std::default_delete<uint8_t[]>());
   if (dst == nullptr) {
-    GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION, "Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld",
+    GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION,
+           "Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld",
            TypeUtils::FormatToSerialString(args.src_format).c_str(),
            TypeUtils::FormatToSerialString(args.dst_format).c_str(), dst_size);
     return ACL_ERROR_GE_MEMORY_ALLOCATION;
@@ -123,7 +124,8 @@ Status TransFormatDhwncToFz3DTranspose(const TransArgs &args, TransResult &resul
                                args.data + src_idx * data_size, static_cast<size_t>(data_size));
               }
               if (ret != EOK) {
-                GELOGE(ACL_ERROR_GE_MEMORY_OPERATE_FAILED, "Failed to operate the dst memory at offset %ld, error-code %d, pad mode %d",
+                GELOGE(ACL_ERROR_GE_MEMORY_OPERATE_FAILED,
+                       "Failed to operate the dst memory at offset %ld, error-code %d, pad mode %d",
                        dst_offset, ret, pad_zero);
                 return ACL_ERROR_GE_MEMORY_OPERATE_FAILED;
               }

ge/common/formats/format_transfers/format_transfer_fractal_nz.cc

@@ -139,7 +139,8 @@ Status TransFormatFromNdToFracNz(const TransArgs &args, TransResult &result, con
 
   std::shared_ptr<uint8_t> dst(new (std::nothrow) uint8_t[dst_size](), std::default_delete<uint8_t[]>());
   if (dst == nullptr) {
-    GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION, "Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld",
+    GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION,
+           "Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld",
            TypeUtils::FormatToSerialString(args.src_format).c_str(),
            TypeUtils::FormatToSerialString(args.dst_format).c_str(), dst_size);
     return ACL_ERROR_GE_MEMORY_ALLOCATION;
@@ -175,7 +176,8 @@ Status TransFormatFromNdToFracNz(const TransArgs &args, TransResult &result, con
         auto ret = memcpy_s(dst.get() + dst_offset, static_cast<size_t>(protected_size), args.data + src_offset,
                             static_cast<size_t>(size * w0));
         if (ret != EOK) {
-          GELOGE(ACL_ERROR_GE_MEMORY_OPERATE_FAILED, "Failed to operate the dst memory at offset %ld, error-code %d", dst_offset, ret);
+          GELOGE(ACL_ERROR_GE_MEMORY_OPERATE_FAILED,
+                 "Failed to operate the dst memory at offset %ld, error-code %d", dst_offset, ret);
           return ACL_ERROR_GE_MEMORY_OPERATE_FAILED;
         }
       }
@@ -189,7 +191,8 @@ Status TransFormatFromNdToFracNz(const TransArgs &args, TransResult &result, con
         auto ret = memcpy_s(dst.get() + dst_offset, static_cast<size_t>(protected_size), args.data + src_offset,
                             static_cast<size_t>(size));
         if (ret != EOK) {
-          GELOGE(ACL_ERROR_GE_MEMORY_OPERATE_FAILED, "Failed to operate the dst memory at offset %ld, error-code %d", dst_offset, ret);
+          GELOGE(ACL_ERROR_GE_MEMORY_OPERATE_FAILED,
+                 "Failed to operate the dst memory at offset %ld, error-code %d", dst_offset, ret);
           return ACL_ERROR_GE_MEMORY_OPERATE_FAILED;
         }
       }
@@ -210,7 +213,8 @@ Status TransFormatFromFracNzToNd(const TransArgs &args, TransResult &result, con
 
   std::shared_ptr<uint8_t> dst(new (std::nothrow) uint8_t[dst_size], std::default_delete<uint8_t[]>());
   if (dst == nullptr) {
-    GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION, "Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld",
+    GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION,
+           "Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld",
            TypeUtils::FormatToSerialString(args.src_format).c_str(),
            TypeUtils::FormatToSerialString(args.dst_format).c_str(), dst_size);
     return ACL_ERROR_GE_MEMORY_ALLOCATION;
@@ -246,7 +250,8 @@ Status TransFormatFromFracNzToNd(const TransArgs &args, TransResult &result, con
         ret = memcpy_s(dst.get() + dst_offset, static_cast<size_t>(protected_size), args.data + src_offset,
                        static_cast<size_t>(size * w0));
         if (ret != EOK) {
-          GELOGE(ACL_ERROR_GE_MEMORY_OPERATE_FAILED, "Failed to operate the dst memory at offset %ld, error-code %d", dst_offset, ret);
+          GELOGE(ACL_ERROR_GE_MEMORY_OPERATE_FAILED,
+                 "Failed to operate the dst memory at offset %ld, error-code %d", dst_offset, ret);
           return ACL_ERROR_GE_MEMORY_OPERATE_FAILED;
         }
       }
@@ -260,7 +265,8 @@ Status TransFormatFromFracNzToNd(const TransArgs &args, TransResult &result, con
         ret = memcpy_s(dst.get() + dst_offset, static_cast<size_t>(protected_size), args.data + src_offset,
                        static_cast<size_t>(size));
         if (ret != EOK) {
-          GELOGE(ACL_ERROR_GE_MEMORY_OPERATE_FAILED, "Failed to operate the dst memory at offset %ld, error-code %d", dst_offset, ret);
+          GELOGE(ACL_ERROR_GE_MEMORY_OPERATE_FAILED,
+                 "Failed to operate the dst memory at offset %ld, error-code %d", dst_offset, ret);
           return ACL_ERROR_GE_MEMORY_OPERATE_FAILED;
         }
       }
@@ -274,14 +280,16 @@ Status TransFormatFromFracNzToNd(const TransArgs &args, TransResult &result, con
 
 Status FormatTransferFractalNz::TransFormat(const TransArgs &args, TransResult &result) {
   if (!IsDataTypeSupport(args.src_data_type)) {
-    GELOGE(ACL_ERROR_GE_DATATYPE_INVALID, "Trans format from %s to %s, src shape %s, dst shape %s, data type %s is not supported",
+    GELOGE(ACL_ERROR_GE_DATATYPE_INVALID,
+           "Trans format from %s to %s, src shape %s, dst shape %s, data type %s is not supported",
            TypeUtils::FormatToSerialString(args.src_format).c_str(),
            TypeUtils::FormatToSerialString(args.dst_format).c_str(), ShapeToString(args.src_shape).c_str(),
            ShapeToString(args.dst_shape).c_str(), TypeUtils::DataTypeToSerialString(args.src_data_type).c_str());
     return ACL_ERROR_GE_DATATYPE_INVALID;
   }
   if (!CheckShape(args.src_format, args.src_shape) || !IsShapeValid(args.dst_shape)) {
-    GELOGE(ACL_ERROR_GE_SHAPE_INVALID, "Trans format from %s to %s, src shape %s, dst shape %s, data type %s is not supported",
+    GELOGE(ACL_ERROR_GE_SHAPE_INVALID,
+           "Trans format from %s to %s, src shape %s, dst shape %s, data type %s is not supported",
            TypeUtils::FormatToSerialString(args.src_format).c_str(),
            TypeUtils::FormatToSerialString(args.dst_format).c_str(), ShapeToString(args.src_shape).c_str(),
            ShapeToString(args.dst_shape).c_str(), TypeUtils::DataTypeToSerialString(args.src_data_type).c_str());
@@ -325,7 +333,8 @@ Status FormatTransferFractalNz::TransShape(Format src_format, const ShapeVector
 
 Status FormatTransferFractalNzND::TransFormat(const TransArgs &args, TransResult &result) {
   if (!IsDataTypeSupport(args.src_data_type)) {
-    GELOGE(ACL_ERROR_GE_DATATYPE_INVALID, "Trans format from %s to %s, src shape %s, dst shape %s, data type %s is not supported",
+    GELOGE(ACL_ERROR_GE_DATATYPE_INVALID,
+           "Trans format from %s to %s, src shape %s, dst shape %s, data type %s is not supported",
            TypeUtils::FormatToSerialString(args.src_format).c_str(),
            TypeUtils::FormatToSerialString(args.dst_format).c_str(), ShapeToString(args.src_shape).c_str(),
            ShapeToString(args.dst_shape).c_str(), TypeUtils::DataTypeToSerialString(args.src_data_type).c_str());
@@ -333,7 +342,8 @@ Status FormatTransferFractalNzND::TransFormat(const TransArgs &args, TransResult
   }
 
   if (!IsShapeValid(args.src_shape) || !CheckShape(args.dst_format, args.dst_shape)) {
-    GELOGE(ACL_ERROR_GE_SHAPE_INVALID, "Trans format from %s to %s, src shape %s, dst shape %s, data type %s is not supported",
+    GELOGE(ACL_ERROR_GE_SHAPE_INVALID,
+           "Trans format from %s to %s, src shape %s, dst shape %s, data type %s is not supported",
            TypeUtils::FormatToSerialString(args.src_format).c_str(),
            TypeUtils::FormatToSerialString(args.dst_format).c_str(), ShapeToString(args.src_shape).c_str(),
            ShapeToString(args.dst_shape).c_str(), TypeUtils::DataTypeToSerialString(args.src_data_type).c_str());

ge/common/formats/format_transfers/format_transfer_fractal_z.cc

@@ -127,7 +127,8 @@ Status TransFormatFromNchwToFz(const TransArgs &args, TransResult &result) {
   std::shared_ptr<uint8_t> dst(new (std::nothrow) uint8_t[dst_size], std::default_delete<uint8_t[]>());
   GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(
       dst == nullptr,
-      GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION, "Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld",
+      GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION,
+             "Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld",
              TypeUtils::FormatToSerialString(args.src_format).c_str(),
              TypeUtils::FormatToSerialString(args.dst_format).c_str(), dst_size);
       return ACL_ERROR_GE_MEMORY_ALLOCATION;);
@@ -173,8 +174,9 @@ Status TransFormatFromNchwToFz(const TransArgs &args, TransResult &result) {
             }
           }
           if (ret != EOK) {
-            GELOGE(ACL_ERROR_GE_MEMORY_OPERATE_FAILED, "Failed to operate the dst memory at offset %ld, error-code %d pad mode %d", offset,
-                   ret, need_pad_zero);
+            GELOGE(ACL_ERROR_GE_MEMORY_OPERATE_FAILED,
+                   "Failed to operate the dst memory at offset %ld, error-code %d pad mode %d",
+                   offset, ret, need_pad_zero);
             return ACL_ERROR_GE_MEMORY_OPERATE_FAILED;
           }
         }
@@ -213,7 +215,8 @@ Status TransFormatHwcnToFz(const TransArgs &args, TransResult &result) {
   std::shared_ptr<uint8_t> dst(new (std::nothrow) uint8_t[dst_size], std::default_delete<uint8_t[]>());
   GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(
       dst == nullptr,
-      GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION, "Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld",
+      GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION,
+             "Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld",
              TypeUtils::FormatToSerialString(args.src_format).c_str(),
              TypeUtils::FormatToSerialString(args.dst_format).c_str(), dst_size);
       return ACL_ERROR_GE_MEMORY_ALLOCATION;);
@@ -235,7 +238,8 @@ Status TransFormatHwcnToFz(const TransArgs &args, TransResult &result) {
                              static_cast<size_t>(data_size));
             } else {
               if (protected_size < data_size) {
-                GELOGE(ACL_ERROR_GE_PARAM_INVALID, "Failed to operate the dst memory, protected_size is %ld and size is %ld",
+                GELOGE(ACL_ERROR_GE_PARAM_INVALID,
+                       "Failed to operate the dst memory, protected_size is %ld and size is %ld",
                        protected_size, data_size);
                 return ACL_ERROR_GE_PARAM_INVALID;
               }
@@ -247,7 +251,8 @@ Status TransFormatHwcnToFz(const TransArgs &args, TransResult &result) {
               }
             }
             if (ret != EOK) {
-              GELOGE(ACL_ERROR_GE_MEMORY_OPERATE_FAILED, "Failed to operate the dst memory at offset %ld, error-code %d, pad mode %d",
+              GELOGE(ACL_ERROR_GE_MEMORY_OPERATE_FAILED,
+                     "Failed to operate the dst memory at offset %ld, error-code %d, pad mode %d",
                      dst_offset, ret, pad_zero);
               return ACL_ERROR_GE_MEMORY_OPERATE_FAILED;
             }
@@ -288,7 +293,8 @@ Status TransFormatNhwcToFz(const TransArgs &args, TransResult &result) {
   std::shared_ptr<uint8_t> dst(new (std::nothrow) uint8_t[dst_size], std::default_delete<uint8_t[]>());
   GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(
       dst == nullptr,
-      GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION, "Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld",
+      GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION,
+             "Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld",
              TypeUtils::FormatToSerialString(args.src_format).c_str(),
              TypeUtils::FormatToSerialString(args.dst_format).c_str(), dst_size);
       return ACL_ERROR_GE_MEMORY_ALLOCATION;);
@@ -310,7 +316,8 @@ Status TransFormatNhwcToFz(const TransArgs &args, TransResult &result) {
                              static_cast<size_t>(data_size));
             } else {
               if (protected_size < data_size) {
-                GELOGE(ACL_ERROR_GE_PARAM_INVALID, "Failed to operate the dst memory, protected_size is %ld and size is %ld",
+                GELOGE(ACL_ERROR_GE_PARAM_INVALID,
+                       "Failed to operate the dst memory, protected_size is %ld and size is %ld",
                        protected_size, data_size);
                 return ACL_ERROR_GE_PARAM_INVALID;
               }
@@ -322,7 +329,8 @@ Status TransFormatNhwcToFz(const TransArgs &args, TransResult &result) {
               }
             }
             if (ret != EOK) {
-              GELOGE(ACL_ERROR_GE_MEMORY_OPERATE_FAILED, "Failed to operate the dst memory at offset %ld, error-code %d, pad mode %d",
+              GELOGE(ACL_ERROR_GE_MEMORY_OPERATE_FAILED,
+                     "Failed to operate the dst memory at offset %ld, error-code %d, pad mode %d",
                      dst_offset, ret, pad_zero);
               return ACL_ERROR_GE_MEMORY_OPERATE_FAILED;
             }

ge/common/formats/format_transfers/format_transfer_fractal_zz.cc

@@ -140,7 +140,8 @@ Status TransFormatFromNdToFracZz(const TransArgs &args, TransResult &result, con
 
   std::shared_ptr<uint8_t> dst(new (std::nothrow) uint8_t[dst_size](), std::default_delete<uint8_t[]>());
   if (dst == nullptr) {
-    GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION, "Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld",
+    GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION,
+           "Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld",
            TypeUtils::FormatToSerialString(args.src_format).c_str(),
            TypeUtils::FormatToSerialString(args.dst_format).c_str(), dst_size);
     return ACL_ERROR_GE_MEMORY_ALLOCATION;
@@ -179,7 +180,8 @@ Status TransFormatFromNdToFracZz(const TransArgs &args, TransResult &result, con
           auto ret = memcpy_s(dst.get() + dst_offset, static_cast<size_t>(protected_size), args.data + src_offset,
                               static_cast<size_t>(size * w0));
           if (ret != EOK) {
-            GELOGE(ACL_ERROR_GE_MEMORY_OPERATE_FAILED, "Failed to operate the dst memory at offset %ld, error-code %d", dst_offset, ret);
+            GELOGE(ACL_ERROR_GE_MEMORY_OPERATE_FAILED,
+                   "Failed to operate the dst memory at offset %ld, error-code %d", dst_offset, ret);
             return ACL_ERROR_GE_MEMORY_OPERATE_FAILED;
           }
         }
@@ -195,7 +197,8 @@ Status TransFormatFromNdToFracZz(const TransArgs &args, TransResult &result, con
           auto ret = memcpy_s(dst.get() + dst_offset, static_cast<size_t>(protected_size), args.data + src_offset,
                               static_cast<size_t>(size));
           if (ret != EOK) {
-            GELOGE(ACL_ERROR_GE_MEMORY_OPERATE_FAILED, "Failed to operate the dst memory at offset %ld, error-code %d", dst_offset, ret);
+            GELOGE(ACL_ERROR_GE_MEMORY_OPERATE_FAILED,
+                   "Failed to operate the dst memory at offset %ld, error-code %d", dst_offset, ret);
             return ACL_ERROR_GE_MEMORY_OPERATE_FAILED;
           }
         }
@@ -217,7 +220,8 @@ Status TransFormatFromFracZzToNd(const TransArgs &args, TransResult &result, con
 
   std::shared_ptr<uint8_t> dst(new (std::nothrow) uint8_t[dst_size](), std::default_delete<uint8_t[]>());
   if (dst == nullptr) {
-    GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION, "Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld",
+    GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION,
+           "Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld",
            TypeUtils::FormatToSerialString(args.src_format).c_str(),
            TypeUtils::FormatToSerialString(args.dst_format).c_str(), dst_size);
     return ACL_ERROR_GE_MEMORY_ALLOCATION;
@@ -257,7 +261,8 @@ Status TransFormatFromFracZzToNd(const TransArgs &args, TransResult &result, con
           auto ret = memcpy_s(dst.get() + dst_offset, static_cast<size_t>(protected_size), args.data + src_offset,
                               static_cast<size_t>(size * w0));
           if (ret != EOK) {
-            GELOGE(ACL_ERROR_GE_MEMORY_OPERATE_FAILED, "Failed to operate the dst memory at offset %ld, error-code %d", dst_offset, ret);
+            GELOGE(ACL_ERROR_GE_MEMORY_OPERATE_FAILED,
+                   "Failed to operate the dst memory at offset %ld, error-code %d", dst_offset, ret);
             return ACL_ERROR_GE_MEMORY_OPERATE_FAILED;
           }
         }
@@ -273,7 +278,8 @@ Status TransFormatFromFracZzToNd(const TransArgs &args, TransResult &result, con
           auto ret = memcpy_s(dst.get() + dst_offset, static_cast<size_t>(protected_size), args.data + src_offset,
                               static_cast<size_t>(size));
           if (ret != EOK) {
-            GELOGE(ACL_ERROR_GE_MEMORY_OPERATE_FAILED, "Failed to operate the dst memory at offset %ld, error-code %d", dst_offset, ret);
+            GELOGE(ACL_ERROR_GE_MEMORY_OPERATE_FAILED,
+                   "Failed to operate the dst memory at offset %ld, error-code %d", dst_offset, ret);
             return ACL_ERROR_GE_MEMORY_OPERATE_FAILED;
           }
         }
@@ -288,14 +294,16 @@ Status TransFormatFromFracZzToNd(const TransArgs &args, TransResult &result, con
 
 Status FormatTransferFractalZz::TransFormat(const TransArgs &args, TransResult &result) {
   if (!IsDataTypeSupport(args.src_data_type)) {
-    GELOGE(ACL_ERROR_GE_DATATYPE_INVALID, "Not support trans format from %s to %s, src shape %s, dst shape %s, data type %s",
+    GELOGE(ACL_ERROR_GE_DATATYPE_INVALID,
+           "Not support trans format from %s to %s, src shape %s, dst shape %s, data type %s",
            TypeUtils::FormatToSerialString(args.src_format).c_str(),
            TypeUtils::FormatToSerialString(args.dst_format).c_str(), ShapeToString(args.src_shape).c_str(),
            ShapeToString(args.dst_shape).c_str(), TypeUtils::DataTypeToSerialString(args.src_data_type).c_str());
     return ACL_ERROR_GE_DATATYPE_INVALID;
   }
   if (!CheckShape(args.src_format, args.src_shape) || !IsShapeValid(args.dst_shape)) {
-    GELOGE(ACL_ERROR_GE_SHAPE_INVALID, "Not support trans format from %s to %s, src shape %s, dst shape %s, data type %s",
+    GELOGE(ACL_ERROR_GE_SHAPE_INVALID,
+           "Not support trans format from %s to %s, src shape %s, dst shape %s, data type %s",
            TypeUtils::FormatToSerialString(args.src_format).c_str(),
            TypeUtils::FormatToSerialString(args.dst_format).c_str(), ShapeToString(args.src_shape).c_str(),
            ShapeToString(args.dst_shape).c_str(), TypeUtils::DataTypeToSerialString(args.src_data_type).c_str());
@@ -339,7 +347,8 @@ Status FormatTransferFractalZz::TransShape(Format src_format, const ShapeVector
 
 Status FormatTransferFractalZzND::TransFormat(const TransArgs &args, TransResult &result) {
   if (!IsDataTypeSupport(args.src_data_type)) {
-    GELOGE(ACL_ERROR_GE_DATATYPE_INVALID, "Not support trans format from %s to %s, src shape %s, dst shape %s, data type %s",
+    GELOGE(ACL_ERROR_GE_DATATYPE_INVALID,
+           "Not support trans format from %s to %s, src shape %s, dst shape %s, data type %s",
            TypeUtils::FormatToSerialString(args.src_format).c_str(),
            TypeUtils::FormatToSerialString(args.dst_format).c_str(), ShapeToString(args.src_shape).c_str(),
            ShapeToString(args.dst_shape).c_str(), TypeUtils::DataTypeToSerialString(args.src_data_type).c_str());
@@ -347,7 +356,8 @@ Status FormatTransferFractalZzND::TransFormat(const TransArgs &args, TransResult
   }
 
   if (!IsShapeValid(args.src_shape) || !CheckShape(args.dst_format, args.dst_shape)) {
-    GELOGE(ACL_ERROR_GE_SHAPE_INVALID, "Not support trans format from %s to %s, src shape %s, dst shape %s, data type %s",
+    GELOGE(ACL_ERROR_GE_SHAPE_INVALID,
+           "Not support trans format from %s to %s, src shape %s, dst shape %s, data type %s",
            TypeUtils::FormatToSerialString(args.src_format).c_str(),
            TypeUtils::FormatToSerialString(args.dst_format).c_str(), ShapeToString(args.src_shape).c_str(),
            ShapeToString(args.dst_shape).c_str(), TypeUtils::DataTypeToSerialString(args.src_data_type).c_str());

ge/common/formats/format_transfers/format_transfer_fracz_hwcn.cc

@@ -66,7 +66,7 @@ Status CheckArgsForFracZToHwcn(const TransArgs &args) {
         FmtToStr(ShapeToString(dst_shape));
     GE_ERRORLOG_AND_ERRORMSG(ACL_ERROR_GE_SHAPE_INVALID, error.c_str());
     return ACL_ERROR_GE_SHAPE_INVALID;
-  } 
+  }
 
   return SUCCESS;
 }
@@ -74,7 +74,8 @@ Status CheckArgsForFracZToHwcn(const TransArgs &args) {
 Status GetDstDataAfterTrans(const TransArgs &args, TransResult &result, const int size, const int64_t total_size) {
   std::shared_ptr<uint8_t> dst(new (std::nothrow) uint8_t[total_size], std::default_delete<uint8_t[]>());
   if (dst == nullptr) {
-    GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION, "Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld, shape %s",
+    GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION,
+           "Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld, shape %s",
            TypeUtils::FormatToSerialString(args.src_format).c_str(),
            TypeUtils::FormatToSerialString(args.dst_format).c_str(), total_size, ShapeToString(args.dst_shape).c_str());
     return ACL_ERROR_GE_MEMORY_ALLOCATION;

ge/common/formats/format_transfers/format_transfer_fracz_nchw.cc

@@ -59,9 +59,10 @@ Status CheckArgsForFracZToNchw(const TransArgs &args) {
   }
   int64_t c1 = Ceil(dst_shape.at(kNchwC), c0);
   int64_t n0 = Ceil(dst_shape.at(kNchwN), static_cast<int64_t>(kNiSize));
-  if (src_shape.at(kFracZHWC1) != dst_shape.at(kNchwH) * dst_shape.at(kNchwW) * c1 || src_shape.at(kFracZC0) != c0 ||
-      src_shape.at(kFracZNi) != kNiSize || src_shape.at(kFracZN0) != n0) {
-    GELOGE(ACL_ERROR_GE_SHAPE_INVALID, "Failed to check relationship between src and dst shape, src shape %s, dst shape %s",
+  if (src_shape.at(kFracZHWC1) != dst_shape.at(kNchwH) * dst_shape.at(kNchwW) * c1 ||
+      src_shape.at(kFracZC0) != c0 || src_shape.at(kFracZNi) != kNiSize || src_shape.at(kFracZN0) != n0) {
+    GELOGE(ACL_ERROR_GE_SHAPE_INVALID,
+           "Failed to check relationship between src and dst shape, src shape %s, dst shape %s",
            ShapeToString(src_shape).c_str(), ShapeToString(dst_shape).c_str());
     return ACL_ERROR_GE_SHAPE_INVALID;
   }
@@ -72,7 +73,8 @@ Status CheckArgsForFracZToNchw(const TransArgs &args) {
 Status GetDstDataAfterTrans(const TransArgs &args, TransResult &result, const int size, const int64_t total_size) {
   std::shared_ptr<uint8_t> dst(new (std::nothrow) uint8_t[total_size], std::default_delete<uint8_t[]>());
   if (dst == nullptr) {
-    GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION, "Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld, shape %s",
+    GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION,
+           "Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld, shape %s",
            TypeUtils::FormatToSerialString(args.src_format).c_str(),
            TypeUtils::FormatToSerialString(args.dst_format).c_str(), total_size, ShapeToString(args.dst_shape).c_str());
     return ACL_ERROR_GE_MEMORY_ALLOCATION;

ge/common/formats/format_transfers/format_transfer_fracz_nhwc.cc

@@ -59,9 +59,10 @@ Status CheckArgsForFracZToNhwc(const TransArgs &args) {
   }
   int64_t c1 = Ceil(dst_shape.at(kNhwcC), c0);
   int64_t n0 = Ceil(dst_shape.at(kNhwcN), static_cast<int64_t>(kNiSize));
-  if (src_shape.at(kFracZHWC1) != dst_shape.at(kNhwcH) * dst_shape.at(kNhwcW) * c1 || src_shape.at(kFracZC0) != c0 ||
-      src_shape.at(kFracZNi) != kNiSize || src_shape.at(kFracZN0) != n0) {
-    GELOGE(PARAM_INVALID, "Failed to check relationship between src and dst shape, src shape %s, dst shape %s",
+  if (src_shape.at(kFracZHWC1) != dst_shape.at(kNhwcH) * dst_shape.at(kNhwcW) * c1 ||
+      src_shape.at(kFracZC0) != c0 || src_shape.at(kFracZNi) != kNiSize || src_shape.at(kFracZN0) != n0) {
+    GELOGE(PARAM_INVALID,
+           "Failed to check relationship between src and dst shape, src shape %s, dst shape %s",
            ShapeToString(src_shape).c_str(), ShapeToString(dst_shape).c_str());
     return PARAM_INVALID;
   }
@@ -72,7 +73,8 @@ Status CheckArgsForFracZToNhwc(const TransArgs &args) {
 Status GetDstDataAfterTrans(const TransArgs &args, TransResult &result, int size, int64_t total_size) {
   std::shared_ptr<uint8_t> dst(new (std::nothrow) uint8_t[total_size], std::default_delete<uint8_t[]>());
   if (dst == nullptr) {
-    GELOGE(OUT_OF_MEMORY, "Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld, shape %s",
+    GELOGE(OUT_OF_MEMORY,
+           "Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld, shape %s",
            TypeUtils::FormatToSerialString(args.src_format).c_str(),
            TypeUtils::FormatToSerialString(args.dst_format).c_str(), total_size, ShapeToString(args.dst_shape).c_str());
     return OUT_OF_MEMORY;
@@ -140,7 +142,7 @@ Status FormatTransferFracZNhwc::TransFormat(const TransArgs &args, TransResult &
     }
 
     GELOGE(INTERNAL_ERROR, "Get %ld total size from dst shape %s, src shape %s", total_size,
-        ShapeToString(args.dst_shape).c_str(), ShapeToString(args.src_shape).c_str());
+           ShapeToString(args.dst_shape).c_str(), ShapeToString(args.src_shape).c_str());
     return PARAM_INVALID;
   }
   GELOGD("Begin to trans format from FracZ to NHWC, src shape %s, data type %s, dst shape %s, memory size %ld",

ge/common/formats/format_transfers/format_transfer_hwcn_c1hwncoc0.cc

@@ -91,7 +91,8 @@ Status CheckArgsForHwcnToC1hwncoc0(const TransArgs &args) {
 Status GetDstDataAfterTrans(const TransArgs &args, TransResult &result, const int size, const int64_t total_size) {
   std::shared_ptr<uint8_t> dst(new (std::nothrow) uint8_t[total_size], std::default_delete<uint8_t[]>());
   if (dst == nullptr) {
-    GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION, "Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld, shape %s",
+    GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION,
+           "Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld, shape %s",
            TypeUtils::FormatToSerialString(args.src_format).c_str(),
            TypeUtils::FormatToSerialString(args.dst_format).c_str(), total_size, ShapeToString(args.dst_shape).c_str());
     return ACL_ERROR_GE_MEMORY_ALLOCATION;

ge/common/formats/format_transfers/format_transfer_nc1hwc0_nchw.cc

@@ -72,7 +72,8 @@ Status CheckArgsForNc1hwc0ToNchw(const TransArgs &args) {
 Status GetDstDataAfterTrans(const TransArgs &args, TransResult &result, const int size, const int64_t total_size) {
   std::shared_ptr<uint8_t> dst(new (std::nothrow) uint8_t[total_size], std::default_delete<uint8_t[]>());
   if (dst == nullptr) {
-    GELOGE(OUT_OF_MEMORY, "Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld, shape %s",
+    GELOGE(OUT_OF_MEMORY,
+           "Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld, shape %s",
            TypeUtils::FormatToSerialString(args.src_format).c_str(),
            TypeUtils::FormatToSerialString(args.dst_format).c_str(), total_size, ShapeToString(args.dst_shape).c_str());
     return OUT_OF_MEMORY;

ge/common/formats/format_transfers/format_transfer_nc1hwc0_nhwc.cc

@@ -61,7 +61,8 @@ Status CheckArgsForNc1hwc0ToNhwc(const TransArgs &args) {
   if (src_shape.at(kNc1hwc0H) != dst_shape.at(kNhwcH) || src_shape.at(kNc1hwc0W) != dst_shape.at(kNhwcW) ||
       src_shape.at(kNc1hwc0N) != dst_shape.at(kNhwcN) || src_shape.at(kNc1hwc0C0) != c0 ||
       src_shape.at(kNc1hwc0C1) != (Ceil(dst_shape.at(kNhwcC), c0))) {
-    GELOGE(ACL_ERROR_GE_SHAPE_INVALID, "Failed to check relationship between src and dst shape, src shape %s, dst shape %s",
+    GELOGE(ACL_ERROR_GE_SHAPE_INVALID,
+           "Failed to check relationship between src and dst shape, src shape %s, dst shape %s",
            ShapeToString(src_shape).c_str(), ShapeToString(dst_shape).c_str());
     return ACL_ERROR_GE_SHAPE_INVALID;
   }
@@ -72,7 +73,8 @@ Status CheckArgsForNc1hwc0ToNhwc(const TransArgs &args) {
 Status GetDstDataAfterTrans(const TransArgs &args, TransResult &result, const int size, const int64_t total_size) {
   std::shared_ptr<uint8_t> dst(new (std::nothrow) uint8_t[total_size], std::default_delete<uint8_t[]>());
   if (dst == nullptr) {
-    GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION, "Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld, shape %s",
+    GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION,
+           "Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld, shape %s",
            TypeUtils::FormatToSerialString(args.src_format).c_str(),
            TypeUtils::FormatToSerialString(args.dst_format).c_str(), total_size, ShapeToString(args.dst_shape).c_str());
     return ACL_ERROR_GE_MEMORY_ALLOCATION;

ge/common/formats/format_transfers/format_transfer_nchw_fz_c04.cc

@@ -125,7 +125,8 @@ Status TransFormatFromNchwToFzC04(const TransArgs &args, TransResult &result) {
                   return ACL_ERROR_GE_INTERNAL_ERROR);
   auto t1 = h_o * w_o;
   auto t2 = n_o * c_o;
-  GE_IF_BOOL_EXEC(!CheckInt64MulOverflow(t1, t2), GELOGE(INTERNAL_ERROR, "int64 mul overflow.A[%ld], B[%ld]", t1, t2);
+  GE_IF_BOOL_EXEC(!CheckInt64MulOverflow(t1, t2),
+                  GELOGE(INTERNAL_ERROR, "int64 mul overflow.A[%ld], B[%ld]", t1, t2);
                   return ACL_ERROR_GE_INTERNAL_ERROR);
 
   int64_t total_ele_cnt = n_o * c_o * h_o * w_o;
@@ -140,7 +141,8 @@ Status TransFormatFromNchwToFzC04(const TransArgs &args, TransResult &result) {
 
   std::shared_ptr<uint8_t> dst(new (std::nothrow) uint8_t[dst_size], std::default_delete<uint8_t[]>());
   if (dst == nullptr) {
-    GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION, "Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld",
+    GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION,
+           "Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld",
            TypeUtils::FormatToSerialString(args.src_format).c_str(),
            TypeUtils::FormatToSerialString(args.dst_format).c_str(), dst_size);
     return ACL_ERROR_GE_MEMORY_ALLOCATION;
@@ -212,7 +214,8 @@ Status PaddingNC(const TransArgs &args, TransArgs &args_tmp, std::shared_ptr<uin
                   return ACL_ERROR_GE_INTERNAL_ERROR);
   auto t1 = h_o * w_o;
   auto t2 = n_o * c_o;
-  GE_IF_BOOL_EXEC(!CheckInt64MulOverflow(t1, t2), GELOGE(ACL_ERROR_GE_INTERNAL_ERROR, "int64 mul overflow.A[%ld], B[%ld]", t1, t2);
+  GE_IF_BOOL_EXEC(!CheckInt64MulOverflow(t1, t2),
+                  GELOGE(ACL_ERROR_GE_INTERNAL_ERROR,"int64 mul overflow.A[%ld], B[%ld]", t1, t2);
                   return ACL_ERROR_GE_INTERNAL_ERROR);
 
   int64_t total_ele_cnt = n_o * c_o * h_o * w_o;
@@ -228,7 +231,8 @@ Status PaddingNC(const TransArgs &args, TransArgs &args_tmp, std::shared_ptr<uin
 
   dst.reset(new (std::nothrow) uint8_t[dst_size], std::default_delete<uint8_t[]>());
   if (dst == nullptr) {
-    GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION, "Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld",
+    GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION,
+           "Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld",
            TypeUtils::FormatToSerialString(args.src_format).c_str(),
            TypeUtils::FormatToSerialString(args.dst_format).c_str(), dst_size);
     return ACL_ERROR_GE_MEMORY_ALLOCATION;
@@ -275,7 +279,8 @@ Status FormatTransferNchwToFZC04::TransFormat(const TransArgs &args, TransResult
   }
 
   std::vector<int64_t> expect_shape;
-  ret = TransShape(args_tmp.src_format, args_tmp.src_shape, args_tmp.src_data_type, args_tmp.dst_format, expect_shape);
+  ret = TransShape(args_tmp.src_format, args_tmp.src_shape, args_tmp.src_data_type,
+                   args_tmp.dst_format, expect_shape);
   if (ret != SUCCESS) {
     return ret;
   }

ge/common/formats/format_transfers/format_transfer_nhwc_nc1hwc0.cc

@@ -92,7 +92,8 @@ Status CheckArgsForNhwcToNc1hwc0(const TransArgs &args) {
 Status GetDstDataAfterTrans(const TransArgs &args, TransResult &result, const int size, const int64_t total_size) {
   std::shared_ptr<uint8_t> dst(new (std::nothrow) uint8_t[total_size], std::default_delete<uint8_t[]>());
   if (dst == nullptr) {
-    GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION, "Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld, shape %s",
+    GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION,
+           "Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld, shape %s",
            TypeUtils::FormatToSerialString(args.src_format).c_str(),
            TypeUtils::FormatToSerialString(args.dst_format).c_str(), total_size, ShapeToString(args.dst_shape).c_str());
     return ACL_ERROR_GE_MEMORY_ALLOCATION;

ge/common/helper/model_helper.cc

@@ -87,12 +87,13 @@ Status ModelHelper::SaveSizeToModelDef(const GeModelPtr &ge_model) {
 
   std::shared_ptr<ModelTaskDef> model_task_def = ge_model->GetModelTaskDefPtr();
   if (model_task_def == nullptr) {
-    GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION, "Create model task def ptr failed");
-    return ACL_ERROR_GE_MEMORY_ALLOCATION;
+    GELOGD("SaveSizeToModelDef task_info_size is 0.");
+    om_info.push_back(0);
+  } else {
+    size_t partition_task_size = model_task_def->ByteSizeLong();
+    GELOGD("SaveSizeToModelDef task_info_size is %zu", partition_task_size);
+    om_info.push_back(partition_task_size);
   }
-  size_t partition_task_size = model_task_def->ByteSizeLong();
-  GELOGD("SaveSizeToModelDef task_info_size is %zu", partition_task_size);
-  om_info.push_back(partition_task_size);
 
   GE_CHK_BOOL_EXEC(ge::AttrUtils::SetListInt(*(ge_model.get()), "om_info_list", om_info),
                    GELOGE(FAILED, "SetListInt of om_info_list failed.");
@@ -598,6 +599,7 @@ Status ModelHelper::GenerateGeRootModel(OmFileLoadHelper &om_load_helper) {
       is_first_model = false;
       root_model_->SetRootGraph(GraphUtils::GetComputeGraph(cur_model->GetGraph()));
       root_model_->SetModelId(cur_model->GetModelId());
+      root_model_->SetModelName(cur_model->GetName());
       model_ = cur_model;
       continue;
     }

ge/common/profiling/profiling_manager.cc

@@ -31,7 +31,7 @@ const char *const kFpPoint = "fp_point";
 const char *const kBpPoint = "bp_point";
 
 #ifdef DAVINCI_SUPPORT_PROFILING
-const size_t kReportMaxLen = 2048;
+const size_t kReportMaxLen = 1024;
 const int32_t kMaxDeviceNum = 256;
 const uint32_t kInteval = 2;
 const std::string kConfigNumsdev = "devNums";

ge/common/tbe_kernel_store.cc

@@ -15,6 +15,8 @@
  */
 
 #include "common/tbe_kernel_store.h"
+#include "graph/utils/attr_utils.h"
+#include "graph/debug/ge_attr_define.h"
 
 namespace ge {
 
@@ -31,6 +33,15 @@ void TBEKernelStore::LoadTBEKernelBinToOpDesc(const std::shared_ptr<ge::OpDesc>
       GE_IF_BOOL_EXEC(!op_desc->SetExtAttr(ge::OP_EXTATTR_NAME_TBE_KERNEL, kernel_bin),
                       GELOGW("LoadKernelTBEBinToOpDesc: SetExtAttr for kernel_bin failed");)
       GELOGI("Load tbe kernel:%s, %zu", kernel_bin->GetName().c_str(), kernel_bin->GetBinDataSize());
+
+      std::string atomic_kernel_name;
+      (void) AttrUtils::GetStr(op_desc, ATOMIC_ATTR_TBE_KERNEL_NAME, atomic_kernel_name);
+      if (!atomic_kernel_name.empty()) {
+        GELOGI("Get atomic kernel name is %s.", atomic_kernel_name.c_str());
+        auto atomic_kernel_bin = FindKernel(atomic_kernel_name);
+        GE_IF_BOOL_EXEC(!op_desc->SetExtAttr(EXT_ATTR_ATOMIC_TBE_KERNEL, atomic_kernel_bin),
+                        GELOGW("LoadKernelTBEBinToOpDesc: SetExtAttr for atomic kernel_bin failed");)
+      }
     }
   }
 }

ge/executor/CMakeLists.txt

@@ -8,6 +8,7 @@ set(PROTO_LIST
 )
 
 protobuf_generate(ge PROTO_SRCS PROTO_HDRS ${PROTO_LIST})
+protobuf_generate(ge_static PROTO_STATIC_SRCS PROTO_STATIC_HDRS ${PROTO_LIST})
 
 set(SRC_LIST
     "ge_executor.cc"
@@ -162,7 +163,7 @@ set(SRC_LIST
 )
 
 ######## libge_executor.a ########
-add_library(ge_executor STATIC ${SRC_LIST} ${PROTO_HDRS})
+add_library(ge_executor STATIC ${SRC_LIST} ${PROTO_STATIC_HDRS})
 
 target_compile_options(ge_executor PRIVATE
     $<$<OR:$<STREQUAL:${TARGET_SYSTEM_NAME},Linux>,$<STREQUAL:${TARGET_SYSTEM_NAME},Android>>:-fvisibility=hidden -O2 -Werror -Wno-deprecated-declarations -fno-common>
@@ -191,7 +192,7 @@ target_include_directories(ge_executor SYSTEM PRIVATE
     ${METADEF_DIR}/inc/external/graph
     ${METADEF_DIR}/inc/graph
     ${CMAKE_BINARY_DIR}
-    ${CMAKE_BINARY_DIR}/proto/ge
+    ${CMAKE_BINARY_DIR}/proto/ge_static
     #### yellow zone ####
     ${GE_CODE_DIR}/../inc
     ${GE_CODE_DIR}/../inc/cce
@@ -212,6 +213,7 @@ target_link_libraries(ge_executor PRIVATE
 add_library(ge_executor_shared SHARED ${SRC_LIST} ${PROTO_HDRS})
 
 target_compile_options(ge_executor_shared PRIVATE
+    -fno-common
     -Werror
     -O2
     -Wno-deprecated-declarations

ge/executor/ge_executor.cc

@@ -30,6 +30,8 @@
 #include "single_op/single_op_manager.h"
 #include "graph/load/model_manager/davinci_model.h"
 #include "opskernel_manager/ops_kernel_builder_manager.h"
+#include "graph/opsproto_manager.h"
+#include "ge_local_engine/engine/host_cpu_engine.h"
 
 using std::string;
 using std::vector;
@@ -199,6 +201,33 @@ bool IsDynmaicDimsSizeMatchModel(const vector<uint64_t> cur_dynamic_dims,
 namespace ge {
 bool GeExecutor::isInit_ = false;
 
+static void InitOpsProtoManager() {
+  string opsproto_path;
+  const char *path_env = std::getenv("ASCEND_OPP_PATH");
+  if (path_env != nullptr) {
+    string path = path_env;
+    string file_path = RealPath(path.c_str());
+    if (file_path.empty()) {
+      GELOGE(FAILED, "[Check][EnvPath]ASCEND_OPP_PATH path [%s] is invalid.", path.c_str());
+      REPORT_INPUT_ERROR("E68016", {"ASCEND_OPP_PATH", path}); 
+      return;
+    }
+    opsproto_path = (path + "/op_proto/custom/" + ":") + (path + "/op_proto/built-in/");
+    GELOGI("Get opsproto so path from env : %s", path.c_str());
+  } else {
+    string path_base = PluginManager::GetPath();
+    GELOGI("path_base is %s", path_base.c_str());
+    path_base = path_base.substr(0, path_base.rfind('/'));
+    path_base = path_base.substr(0, path_base.rfind('/') + 1);
+    opsproto_path = (path_base + "ops/op_proto/custom/" + ":") + (path_base + "ops/op_proto/built-in/");
+  }
+  GELOGI("Get opsproto path is %s", opsproto_path.c_str());
+  OpsProtoManager *manager = OpsProtoManager::Instance();
+  map<string, string> option_tmp;
+  option_tmp.emplace(std::pair<string, string>(string("ge.opsProtoLibPath"), opsproto_path));
+  (void)manager->Initialize(option_tmp);
+}
+
 GeExecutor::GeExecutor() {}
 
 Status GeExecutor::Initialize() {
@@ -208,6 +237,16 @@ Status GeExecutor::Initialize() {
     return ge::SUCCESS;
   }
 
+  OpTilingManager::GetInstance().LoadSo();
+
+  Status init_hostcpu_engine_status = HostCpuEngine::GetInstance().Initialize();
+  if (init_hostcpu_engine_status != SUCCESS) {
+    GELOGE(init_hostcpu_engine_status, "Failed to initialize HostCpuEngine");
+    return init_hostcpu_engine_status;
+  }
+
+  InitOpsProtoManager();
+
   std::vector<rtMemType_t> mem_type(1, RT_MEMORY_HBM);
   mem_type.push_back(RT_MEMORY_P2P_DDR);
   auto ret = MemManager::Instance().Initialize(mem_type);

ge/ge_local_engine/CMakeLists.txt

@@ -20,6 +20,8 @@ set(OPS_KERNEL_SRC_LIST
 )
 
 protobuf_generate(ge PROTO_SRCS PROTO_HDRS ${PROTO_LIST})
+protobuf_generate(ge_ops_shared PROTO_OPS_SHARED_SRCS PROTO_OPS_SHARED_HDRS ${PROTO_LIST})
+protobuf_generate(ge_ops_static PROTO_OPS_STATIC_SRCS PROTO_OPS_STATIC_HDRS ${PROTO_LIST})
 
 ############ libge_local_engine.so ############
 add_library(ge_local_engine SHARED ${SRC_LIST} ${PROTO_HDRS})
@@ -119,7 +121,7 @@ set_target_properties(atc_ge_local_engine PROPERTIES
 )
 
 ############ libge_local_opskernel_builder.so ############
-add_library(ge_local_opskernel_builder SHARED ${OPS_KERNEL_SRC_LIST} ${PROTO_HDRS})
+add_library(ge_local_opskernel_builder SHARED ${OPS_KERNEL_SRC_LIST} ${PROTO_OPS_SHARED_HDRS})
 
 target_compile_options(ge_local_opskernel_builder PRIVATE
     -Werror
@@ -143,7 +145,7 @@ target_include_directories(ge_local_opskernel_builder PRIVATE
     ${METADEF_DIR}/inc/external/graph
     ${METADEF_DIR}/inc/graph
     ${CMAKE_BINARY_DIR}
-    ${CMAKE_BINARY_DIR}/proto/ge
+    ${CMAKE_BINARY_DIR}/proto/ge_ops_shared
     #### yellow zone ####
     ${GE_CODE_DIR}/../inc
     #### blue zone ####
@@ -166,7 +168,7 @@ target_link_libraries(ge_local_opskernel_builder PRIVATE
 )
 
 ############ atclib/libge_local_opskernel_builder.so ############
-add_library(atc_ge_local_opskernel_builder SHARED ${OPS_KERNEL_SRC_LIST} ${PROTO_HDRS})
+add_library(atc_ge_local_opskernel_builder SHARED ${OPS_KERNEL_SRC_LIST} ${PROTO_OPS_SHARED_HDRS})
 
 target_compile_options(atc_ge_local_opskernel_builder PRIVATE
     -Werror
@@ -190,7 +192,7 @@ target_include_directories(atc_ge_local_opskernel_builder PRIVATE
     ${METADEF_DIR}/inc/external/graph
     ${METADEF_DIR}/inc/graph
     ${CMAKE_BINARY_DIR}
-    ${CMAKE_BINARY_DIR}/proto/ge
+    ${CMAKE_BINARY_DIR}/proto/ge_ops_shared
     #### yellow zone ####
     ${GE_CODE_DIR}/../inc
     #### blue zone ####
@@ -218,7 +220,7 @@ set_target_properties(atc_ge_local_opskernel_builder PROPERTIES
 )
 
 ############ libge_local_opskernel_builder.a ############
-add_library(ge_local_opskernel_builder_static STATIC ${OPS_KERNEL_SRC_LIST} ${PROTO_HDRS})
+add_library(ge_local_opskernel_builder_static STATIC ${OPS_KERNEL_SRC_LIST} ${PROTO_OPS_STATIC_HDRS})
 
 target_compile_options(ge_local_opskernel_builder_static PRIVATE
     -Werror
@@ -243,7 +245,7 @@ target_include_directories(ge_local_opskernel_builder_static PRIVATE
     ${METADEF_DIR}/inc/external/graph
     ${METADEF_DIR}/inc/graph
     ${CMAKE_BINARY_DIR}
-    ${CMAKE_BINARY_DIR}/proto/ge
+    ${CMAKE_BINARY_DIR}/proto/ge_ops_static
     #### yellow zone ####
     ${GE_CODE_DIR}/../inc
     #### blue zone ####

ge/ge_local_engine/ops_kernel_store/op/ge_deleted_op.cc

@@ -38,6 +38,7 @@ REGISTER_OP_CREATOR(ExpandDims, GeDeletedOp);
 REGISTER_OP_CREATOR(Reshape, GeDeletedOp);
 REGISTER_OP_CREATOR(ReFormat, GeDeletedOp);
 REGISTER_OP_CREATOR(Squeeze, GeDeletedOp);
+REGISTER_OP_CREATOR(Unsqueeze, GeDeletedOp);
 REGISTER_OP_CREATOR(Size, GeDeletedOp);
 REGISTER_OP_CREATOR(Shape, GeDeletedOp);
 REGISTER_OP_CREATOR(ShapeN, GeDeletedOp);

ge/ge_runtime/CMakeLists.txt

@@ -16,6 +16,7 @@ set(GE_SRC_LIST
     "task/label_goto_task.cc"
     "task/label_set_task.cc"
     "task/label_switch_task.cc"
+    "task/label_manager.cc"
 )
 
 add_library(ge_runtime SHARED ${GE_SRC_LIST})

ge/ge_runtime/task/label_goto_task.cc

@@ -23,26 +23,40 @@ LabelGotoTask::LabelGotoTask(const ModelContext &model_context, const std::share
     : TaskRepeater<LabelGotoTaskInfo>(model_context, task_info),
       task_info_(task_info),
       stream_(nullptr),
-      label_(nullptr) {
+      index_value_(nullptr) {
   if (task_info_ == nullptr) {
     GELOGW("task_info_ is null!");
     return;
   }
   auto stream_list = model_context.stream_list();
   auto label_list = model_context.label_list();
+  rt_model_handle_ = model_context.rt_model_handle();
   uint32_t stream_id = task_info->stream_id();
-  uint32_t label_id = task_info->label_id();
+  label_id_ = task_info->label_id();
   GELOGI("Stream list size:%zu, stream id:%u.", stream_list.size(), stream_id);
-  GELOGI("Label list size:%zu, label id:%u.", label_list.size(), label_id);
-  if (stream_id >= stream_list.size() || label_id >= label_list.size()) {
+  GELOGI("Label list size:%zu, label id:%u.", label_list.size(), label_id_);
+  if (stream_id >= stream_list.size() || label_id_ >= label_list.size()) {
     GELOGW("Stream/Label id invalid.");
     return;
   }
   stream_ = stream_list[stream_id];
-  label_ = label_list[label_id];
+  label_manager_ = LabelManager::GetInstance();
+  if (label_manager_ == nullptr) {
+    GELOGW("Get label manager instance failed.");
+    return;
+  }
+  label_info_ = label_manager_->GetLabelInfo(rt_model_handle_, {label_id_}, label_list);
 }
 
-LabelGotoTask::~LabelGotoTask() {}
+LabelGotoTask::~LabelGotoTask() {
+  if (index_value_ != nullptr) {
+    rtError_t rt_ret = rtFree(index_value_);
+    if (rt_ret != RT_ERROR_NONE) {
+      GELOGE(RT_FAILED, "rtFree index_value_ failed! ret: 0x%X.", rt_ret);
+    }
+    index_value_ = nullptr;
+  }
+}
 
 bool LabelGotoTask::Distribute() {
   GELOGI("LabelGotoTask Distribute start.");
@@ -50,11 +64,29 @@ bool LabelGotoTask::Distribute() {
     GELOGE(PARAM_INVALID, "stream is null!");
     return false;
   }
-  if (label_ == nullptr) {
-    GELOGE(PARAM_INVALID, "label is null!");
+
+  if (label_info_ == nullptr) {
+    GELOGE(PARAM_INVALID, "label info is null!");
     return false;
   }
-  rtError_t rt_ret = rtLabelGotoEx(label_, stream_);
+
+  if (index_value_ == nullptr) {
+    rtError_t rt_ret = rtMalloc(&index_value_, sizeof(uint64_t), RT_MEMORY_HBM);
+    if (rt_ret != RT_ERROR_NONE) {
+      GELOGE(RT_FAILED, "Call rt api failed, ret: 0x%X", rt_ret);
+      return false;
+    }
+
+    uint64_t index = 0;
+    rt_ret = rtMemcpy(index_value_, sizeof(uint64_t), &index, sizeof(index), RT_MEMCPY_HOST_TO_DEVICE);
+    if (rt_ret != RT_ERROR_NONE) {
+      GELOGE(RT_FAILED, "Call rt api failed, ret: 0x%X", rt_ret);
+      return false;
+    }
+  }
+
+  void *label_info = label_info_->GetLabelInfo();
+  rtError_t rt_ret = rtLabelSwitchByIndex(index_value_, 1, label_info, stream_);
   if (rt_ret != RT_ERROR_NONE) {
     GELOGE(RT_FAILED, "Call rt api failed, ret: 0x%X", rt_ret);
     return false;
@@ -65,6 +97,5 @@ bool LabelGotoTask::Distribute() {
 }
 
 REGISTER_TASK(TaskInfoType::LABEL_GOTO, LabelGotoTask, LabelGotoTaskInfo);
-
 }  // namespace model_runner
 }  // namespace ge

ge/ge_runtime/task/label_goto_task.h

@@ -18,7 +18,11 @@
 #define GE_GE_RUNTIME_TASK_LABEL_GOTO_TASK_H_
 
 #include <memory>
+#include <vector>
+#include <map>
+#include <mutex>
 #include "ge_runtime/task/task.h"
+#include "ge_runtime/task/label_manager.h"
 
 namespace ge {
 namespace model_runner {
@@ -33,7 +37,11 @@ class LabelGotoTask : public TaskRepeater<LabelGotoTaskInfo> {
  private:
   std::shared_ptr<LabelGotoTaskInfo> task_info_;
   void *stream_;
-  void *label_;
+  std::shared_ptr<LabelGuard> label_info_;
+  void *index_value_;
+  uint32_t label_id_;
+  rtModel_t rt_model_handle_;
+  std::shared_ptr<LabelManager> label_manager_;
 };
 }  // namespace model_runner
 }  // namespace ge

ge/ge_runtime/task/label_manager.cc

@@ -0,0 +1,119 @@
+/**
+ * Copyright 2021 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 "ge_runtime/task/label_manager.h"
+#include <algorithm>
+#include <string>
+#include "runtime/mem.h"
+#include "runtime/rt_model.h"
+#include "common/ge_inner_error_codes.h"
+#include "framework/common/debug/ge_log.h"
+
+namespace ge {
+namespace model_runner {
+std::weak_ptr<LabelManager> LabelManager::instance_;
+std::mutex LabelManager::instance_mutex_;
+
+template <class T>
+static std::string GetVectorString(const std::vector<T> &vec) {
+  std::string ret;
+  for (size_t i = 0; i < vec.size(); ++i) {
+    if (i != 0) {
+      ret.push_back(',');
+    }
+    ret += std::to_string(vec[i]);
+  }
+  return ret;
+}
+
+LabelGuard::~LabelGuard() {
+  void *label_info = GetLabelInfo();
+  if (label_info != nullptr) {
+    rtError_t rt_ret = rtFree(label_info);
+    if (rt_ret != RT_ERROR_NONE) {
+      GELOGE(RT_FAILED, "rtFree label_info failed! ret: 0x%X.", rt_ret);
+    }
+  }
+}
+
+std::shared_ptr<LabelManager> LabelManager::GetInstance() {
+  std::lock_guard<std::mutex> lock(instance_mutex_);
+  auto instance = instance_.lock();
+  if (instance != nullptr) {
+    return instance;
+  }
+
+  instance = std::make_shared<LabelManager>();
+  instance_ = instance;
+  return instance;
+}
+
+std::shared_ptr<LabelGuard> LabelManager::GetLabelInfo(rtModel_t model, const std::vector<uint32_t> &label_ids,
+                                                       const std::vector<void *> &all_label) {
+  std::lock_guard<std::mutex> lock(model_info_mapping_mutex_);
+  rtError_t rt_ret;
+  auto model_iter = model_info_mapping_.find(model);
+  if (model_iter == model_info_mapping_.end()) {
+    model_info_mapping_.emplace(model, std::map<std::string, std::weak_ptr<LabelGuard>>());
+    model_iter = model_info_mapping_.find(model);
+  }
+
+  std::string label_id_str = GetVectorString(label_ids);
+  auto &label_map = model_iter->second;
+  auto label_iter = label_map.find(label_id_str);
+  if (label_iter != label_map.end()) {
+    auto label_guard = label_iter->second.lock();
+    if (label_guard != nullptr) {
+      GELOGI("model %p find same label id %s.", model, label_id_str.c_str());
+      return label_guard;
+    }
+  }
+
+  GELOGI("Alloc label id %s for model %p.", label_id_str.c_str(), model);
+  void *label_info;
+  std::vector<void *> label_list;
+  bool status = true;
+  std::transform(label_ids.begin(), label_ids.end(), std::back_inserter(label_list),
+                 [&all_label, &status](uint32_t idx) -> void * {
+                   if (idx >= all_label.size()) {
+                     GELOGE(PARAM_INVALID, "Invalid label id %u, all label list size %zu.", idx, all_label.size());
+                     status = false;
+                     return nullptr;
+                   }
+                   return all_label[idx];
+                 });
+  if (!status) {
+    GELOGE(PARAM_INVALID, "Get label info failed.");
+    return nullptr;
+  }
+  uint32_t label_info_size = sizeof(rtLabelDevInfo) * label_list.size();
+  rt_ret = rtMalloc(&label_info, label_info_size, RT_MEMORY_HBM);
+  if (rt_ret != RT_ERROR_NONE) {
+    GELOGE(RT_FAILED, "Call rt api failed, ret: 0x%X", rt_ret);
+    return nullptr;
+  }
+
+  rt_ret = rtLabelListCpy(label_list.data(), label_list.size(), label_info, label_info_size);
+  if (rt_ret != RT_ERROR_NONE) {
+    GELOGE(RT_FAILED, "Call rt api failed, ret: 0x%X", rt_ret);
+    return nullptr;
+  }
+
+  auto label_guard = std::make_shared<LabelGuard>(label_info);
+  label_map.emplace(label_id_str, label_guard);
+  return label_guard;
+}
+}  // namespace model_runner
+}  // namespace ge

ge/ge_runtime/task/label_manager.h

@@ -0,0 +1,54 @@
+/**
+ * Copyright 2021 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_GE_RUNTIME_TASK_LABEL_MANAGER_H_
+#define GE_GE_RUNTIME_TASK_LABEL_MANAGER_H_
+
+#include <vector>
+#include <memory>
+#include <mutex>
+#include <map>
+#include <runtime/base.h>
+
+namespace ge {
+namespace model_runner {
+class LabelGuard {
+ public:
+  explicit LabelGuard(void *label_info) : label_info_(reinterpret_cast<uintptr_t>(label_info)) {}
+  ~LabelGuard();
+  void *GetLabelInfo() { return reinterpret_cast<void *>(label_info_); }
+
+ private:
+  uintptr_t label_info_;
+};
+
+class LabelManager {
+ public:
+  static std::shared_ptr<LabelManager> GetInstance();
+  std::shared_ptr<LabelGuard> GetLabelInfo(rtModel_t model, const std::vector<uint32_t> &label_ids,
+                                           const std::vector<void *> &all_label);
+
+ private:
+  std::mutex model_info_mapping_mutex_;
+  std::map<rtModel_t, std::map<std::string, std::weak_ptr<LabelGuard>>> model_info_mapping_;
+
+  static std::weak_ptr<LabelManager> instance_;
+  static std::mutex instance_mutex_;
+};
+
+
+}  // namespace model_runner
+}  // namespace ge
+#endif  // GE_GE_RUNTIME_TASK_LABEL_MANAGER_H_

ge/ge_runtime/task/label_switch_task.cc

@@ -24,14 +24,14 @@ LabelSwitchTask::LabelSwitchTask(const ModelContext &model_context,
     : TaskRepeater<LabelSwitchTaskInfo>(model_context, task_info),
       task_info_(task_info),
       stream_(nullptr),
-      all_label_resource_(),
       label_info_(nullptr) {
   if (task_info_ == nullptr) {
     GELOGW("task_info_ is null!");
     return;
   }
 
-  all_label_resource_ = model_context.label_list();
+  rt_model_handle_ = model_context.rt_model_handle();
+  auto all_label_resource = model_context.label_list();
   auto stream_list = model_context.stream_list();
   uint32_t stream_id = task_info->stream_id();
   GELOGI("Stream list size:%zu, stream id:%u.", stream_list.size(), stream_id);
@@ -40,52 +40,24 @@ LabelSwitchTask::LabelSwitchTask(const ModelContext &model_context,
     return;
   }
   stream_ = stream_list[stream_id];
-}
-
-LabelSwitchTask::~LabelSwitchTask() {
-  if (label_info_ != nullptr) {
-    rtError_t rt_ret = rtFree(label_info_);
-    if (rt_ret != RT_ERROR_NONE) {
-      GELOGE(RT_FAILED, "rtFree fwkOpBuf failed! ret: 0x%X.", rt_ret);
-    }
-    label_info_ = nullptr;
+  label_manager_ = LabelManager::GetInstance();
+  if (label_manager_ == nullptr) {
+    GELOGW("Get label manager instance failed.");
+    return;
   }
+  label_info_ = label_manager_->GetLabelInfo(rt_model_handle_, task_info_->label_list(), all_label_resource);
 }
 
+LabelSwitchTask::~LabelSwitchTask() {}
+
 bool LabelSwitchTask::Distribute() {
   GELOGI("LabelSwitchTask Distribute start.");
   if (!CheckParamValid()) {
     return false;
   }
 
-  const std::vector<uint32_t> &label_index_list = task_info_->label_list();
-  std::vector<void *> label_list(task_info_->label_size(), nullptr);
-
-  for (size_t i = 0; i < task_info_->label_size(); ++i) {
-    uint32_t label_index = label_index_list[i];
-    if (label_index >= all_label_resource_.size()) {
-      GELOGE(PARAM_INVALID, "label %zu index is %u, but there are %zu labels in total.", i, label_index,
-             all_label_resource_.size());
-      return false;
-    }
-    label_list[i] = all_label_resource_[label_index];
-    GELOGI("Case %zu: label id %zu.", i, label_index);
-  }
-
-  uint32_t label_info_size = sizeof(rtLabelDevInfo) * task_info_->label_size();
-  rtError_t rt_ret = rtMalloc(&label_info_, label_info_size, RT_MEMORY_HBM);
-  if (rt_ret != RT_ERROR_NONE) {
-    GELOGE(RT_FAILED, "Call rt api failed, ret: 0x%X", rt_ret);
-    return false;
-  }
-
-  rt_ret = rtLabelListCpy(label_list.data(), label_list.size(), label_info_, label_info_size);
-  if (rt_ret != RT_ERROR_NONE) {
-    GELOGE(RT_FAILED, "Call rt api failed, ret: 0x%X", rt_ret);
-    return false;
-  }
-
-  rt_ret = rtLabelSwitchByIndex(task_info_->cond(), label_list.size(), label_info_, stream_);
+  void *label_info = label_info_->GetLabelInfo();
+  rtError_t rt_ret = rtLabelSwitchByIndex(task_info_->cond(), task_info_->label_size(), label_info, stream_);
   if (rt_ret != RT_ERROR_NONE) {
     GELOGE(RT_FAILED, "Call rt api failed, ret: 0x%X", rt_ret);
     return false;
@@ -117,8 +89,8 @@ bool LabelSwitchTask::CheckParamValid() {
     return false;
   }
 
-  if (label_info_ != nullptr) {
-    GELOGE(PARAM_INVALID, "label_info_ has dirty data.");
+  if (label_info_ == nullptr) {
+    GELOGE(PARAM_INVALID, "CopyLabelList failed, label info is null.");
     return false;
   }
 
@@ -126,6 +98,5 @@ bool LabelSwitchTask::CheckParamValid() {
 }
 
 REGISTER_TASK(TaskInfoType::LABEL_SWITCH, LabelSwitchTask, LabelSwitchTaskInfo);
-
 }  // namespace model_runner
 }  // namespace ge

ge/ge_runtime/task/label_switch_task.h

@@ -19,6 +19,7 @@
 
 #include <memory>
 #include "ge_runtime/task/task.h"
+#include "ge_runtime/task/label_manager.h"
 
 namespace ge {
 namespace model_runner {
@@ -35,8 +36,9 @@ class LabelSwitchTask : public TaskRepeater<LabelSwitchTaskInfo> {
 
   std::shared_ptr<LabelSwitchTaskInfo> task_info_;
   void *stream_;
-  std::vector<void *> all_label_resource_;
-  void *label_info_;
+  rtModel_t rt_model_handle_;
+  std::shared_ptr<LabelGuard> label_info_;
+  std::shared_ptr<LabelManager> label_manager_;
 };
 }  // namespace model_runner
 }  // namespace ge

ge/generator/ge_generator.cc

@@ -50,9 +50,13 @@ const char *const kFileNameSuffix = "online";
 const char *const kAicpuAllshape = "_AllShape";
 constexpr char const *kAttrSupportDynamicShape = "support_dynamicshape";
 const int64_t kDynamicDimValue = -2;
+const int kDefaultDeviceId = 0;
+const int kDefaultJobId = 0;
 
 std::map<ge::OpEngineType, std::string> engine_type_map{
-    {ge::ENGINE_SYS, kEngineNameDefault}, {ge::ENGINE_AICORE, kAIcoreEngine}, {ge::ENGINE_VECTOR, kVectorEngine}};
+    {ge::ENGINE_SYS, kEngineNameDefault},
+    {ge::ENGINE_AICORE, kAIcoreEngine},
+    {ge::ENGINE_VECTOR, kVectorEngine}};
 
 bool ContainsDynamicInpus(const ge::OpDesc &op_desc) {
   for (auto &tensor_desc : op_desc.GetAllInputsDescPtr()) {
@@ -63,6 +67,9 @@ bool ContainsDynamicInpus(const ge::OpDesc &op_desc) {
   }
   return false;
 }
+bool IsOptional(const ge::GeTensorDesc &tensor_desc) {
+  return tensor_desc.GetFormat() == ge::FORMAT_RESERVED && tensor_desc.GetDataType() == ge::DT_UNDEFINED;
+}
 }  // namespace
 
 namespace ge {
@@ -83,8 +90,9 @@ static Status CheckEngineTypeSupport(const NodePtr &node, OpEngineType engine_ty
   } else {
     ErrorManager::GetInstance().ATCReportErrMessage("E14001", {"opname", "optype", "value", "reason"},
         {op_desc->GetName(), op_desc->GetType(), "engine type",
-        "it only support kEngineNameDefault/kAIcoreEngine/kVectorEngine"});
-    GELOGE(FAILED, "CheckEngineType: engine type: %d not support", static_cast<int>(engine_type));
+        "it only support default/AIcoreEngine/VectorEngine"});
+    GELOGE(FAILED, "[Check][EngineType]value:%d not support, "
+           "only support default/AIcoreEngine/VectorEngine now", static_cast<int>(engine_type));
     return FAILED;
   }
 
@@ -149,7 +157,7 @@ static Status CheckEngineTypeSupport(const NodePtr &node, OpEngineType engine_ty
 }
 
 static Status AddInputs(const ComputeGraphPtr &graph, const NodePtr &node, const GeTensorDesc &tensor, int32_t index,
-                        bool attr) {
+                        bool attr, int32_t &data_index) {
   GE_CHECK_NOTNULL_EXEC(graph, return PARAM_INVALID);
   GE_CHECK_NOTNULL_EXEC(node, return PARAM_INVALID);
 
@@ -188,17 +196,21 @@ static Status AddInputs(const ComputeGraphPtr &graph, const NodePtr &node, const
 
   (void)AttrUtils::SetBool(data_op, "_is_single_op", true);
 
-  GE_CHK_BOOL_EXEC(data_op->AddInputDesc(tensor) == GRAPH_SUCCESS, return FAILED, "Add input desc fail.");
-  GE_CHK_BOOL_EXEC(data_op->AddOutputDesc(tensor) == GRAPH_SUCCESS, return FAILED, "Add output desc fail.");
-  if (attr) {
-    GE_CHK_BOOL_EXEC(AttrUtils::SetInt(data_op, ATTR_NAME_INDEX, index), return FAILED, "Set index fail.");
+  GE_CHK_BOOL_EXEC(data_op->AddInputDesc(tensor) == GRAPH_SUCCESS, return FAILED,
+                   "[Add][InputDesc]fail for node:%s", data_op->GetName().c_str());
+  GE_CHK_BOOL_EXEC(data_op->AddOutputDesc(tensor) == GRAPH_SUCCESS, return FAILED,
+                   "[Add][OutputDesc]fail for node:%s", data_op->GetName().c_str());
+  if (attr && !is_const) {
+    GE_CHK_BOOL_EXEC(AttrUtils::SetInt(data_op, ATTR_NAME_INDEX, data_index), return FAILED,
+                     "[Set][Attr:%s]fail for node:%s", ATTR_NAME_INDEX.c_str(), data_op->GetName().c_str());
+    ++data_index;
   }
 
   ge::NodePtr arg_node = graph->AddNode(data_op);
-  GE_CHK_BOOL_EXEC(arg_node != nullptr, return FAILED, "Insert Data node fail.");
+  GE_CHK_BOOL_EXEC(arg_node != nullptr, return FAILED, "Insert Data node fail");
 
   GE_CHK_STATUS(GraphUtils::AddEdge(arg_node->GetOutDataAnchor(0), node->GetInDataAnchor(index)),
-                "Add edge[%s->%s] fail.", data_op->GetName().c_str(), node->GetName().c_str());
+                "[Add][Edge]fail from node:%s to node:%s", data_op->GetName().c_str(), node->GetName().c_str());
 
   return SUCCESS;
 }
@@ -213,20 +225,23 @@ static Status AddOutputs(const ComputeGraphPtr &graph, const NodePtr &node, cons
   for (const auto &out_desc : outputs) {
     GeTensorDesc tensor = out_desc.GetTensorDesc();
     TensorUtils::SetInputTensor(tensor, true);
-    GE_CHK_BOOL_EXEC(op_desc->AddInputDesc(tensor) == GRAPH_SUCCESS, return FAILED, "Add input desc fail");
+    GE_CHK_BOOL_EXEC(op_desc->AddInputDesc(tensor) == GRAPH_SUCCESS, return FAILED,
+                     "[Add][InputDesc]fail for node:%s", op_desc->GetName().c_str());
 
     TensorUtils::SetInputTensor(tensor, false);
     TensorUtils::SetOutputTensor(tensor, true);
-    GE_CHK_BOOL_EXEC(op_desc->AddOutputDesc(tensor) == GRAPH_SUCCESS, return FAILED, "Add output desc fail");
+    GE_CHK_BOOL_EXEC(op_desc->AddOutputDesc(tensor) == GRAPH_SUCCESS, return FAILED,
+                     "[Add][OutputDesc]fail for node:%s", op_desc->GetName().c_str());
     count++;
   }
   GE_CHECK_NOTNULL_EXEC(graph, return PARAM_INVALID);
   ge::NodePtr out_node = graph->AddNode(op_desc);
-  GE_CHK_BOOL_EXEC(out_node != nullptr, return FAILED, "Insert Output node fail.");
+  GE_CHK_BOOL_EXEC(out_node != nullptr, return FAILED,
+                   "[Add][Node:%s]fail in graph:%u", op_desc->GetName().c_str(), graph->GetGraphID());
   GE_CHECK_NOTNULL_EXEC(node, return PARAM_INVALID);
   for (int32_t i = 0; i < count; ++i) {
     GE_CHK_STATUS(GraphUtils::AddEdge(node->GetOutDataAnchor(i), out_node->GetInDataAnchor(i)),
-                  "Add edge[%s->%s] fail.", node->GetName().c_str(), out_node->GetName().c_str());
+                  "[Add][Edge]fail from node:%s to node:%s", node->GetName().c_str(), out_node->GetName().c_str());
   }
 
   return SUCCESS;
@@ -554,6 +569,44 @@ bool GeGenerator::Impl::SetOmSystemInfo(AttrHolder &obj) {
   return true;
 }
 
+Status GeGenerator::SetModelNameForDump(const GeRootModelPtr &ge_root_model) {
+  bool is_unknown_shape = false;
+  Status ret = ge_root_model->CheckIsUnknownShape(is_unknown_shape);
+  if (ret != SUCCESS) {
+    GELOGE(FAILED, "[Check][IsUnknownShape]Check root model is unknown shape failed, model id:%u",
+           ge_root_model->GetModelId());
+    REPORT_CALL_ERROR("E19999", "Check root model is unknown shape failed, model id:%zu",
+                      ge_root_model->GetModelId());
+    return FAILED;
+  }
+  GeModelPtr model_root = nullptr;
+  if (is_unknown_shape) {
+    model_root = MakeShared<GeModel>();
+    GE_CHECK_NOTNULL(model_root);
+    model_root->SetGraph(GraphUtils::CreateGraphFromComputeGraph(ge_root_model->GetRootGraph()));
+    ge_root_model->SetSubgraphInstanceNameToModel(ge_root_model->GetRootGraph()->GetName(), model_root);
+  }
+
+  ModelHelper model_helper;
+  string model_name;
+  GE_CHECK_NOTNULL(ge_root_model->GetRootGraph());
+  Status name_ret = model_helper.GetModelNameFromMergedGraphName(ge_root_model->GetRootGraph()->GetName(),
+                                                                 model_name);
+  if (name_ret != SUCCESS) {
+    ErrorManager::GetInstance().ATCReportErrMessage("E10000", {"parameter"}, {"output"});
+    GELOGE(FAILED, "[Check][GetModelNameStep]Get model_name failed. Param --output is invalid, root graph name: %s",
+           ge_root_model->GetRootGraph()->GetName().c_str());
+    REPORT_CALL_ERROR("E19999", "Get model_name failed. Param --output is invalid, root graph name: %s",
+                      ge_root_model->GetRootGraph()->GetName().c_str());
+    return PARAM_INVALID;
+  }
+  map<string, GeModelPtr> name_to_ge_model = ge_root_model->GetSubgraphInstanceNameToModel();
+  GeModelPtr &ge_model = name_to_ge_model[ge_root_model->GetRootGraph()->GetName()];
+  GE_CHECK_NOTNULL(ge_model);
+  ge_model->SetName(model_name);
+  return SUCCESS;
+}
+
 Status GeGenerator::GenerateModel(const Graph &graph, const string &file_name_prefix, const vector<GeTensor> &inputs,
                                   ModelBufferData &model, bool is_offline) {
   rtContext_t ctx = nullptr;
@@ -588,20 +641,10 @@ Status GeGenerator::GenerateModel(const Graph &graph, const string &file_name_pr
   }
 
   GE_CHECK_NOTNULL(ge_root_model);
-  GE_CHECK_NOTNULL(ge_root_model->GetRootGraph());
-  ModelHelper model_helper;
-  string model_name = "";
-  Status name_ret = model_helper.GetModelNameFromMergedGraphName(ge_root_model->GetRootGraph()->GetName(),
-                                                                 model_name);
-  if (name_ret != SUCCESS) {
-    ErrorManager::GetInstance().ATCReportErrMessage("E10000", {"parameter"}, {"output"});
-    GELOGE(FAILED, "Get model_name failed. Param --output is invalid.");
-    return PARAM_INVALID;
+  ret = SetModelNameForDump(ge_root_model);
+  if (ret != SUCCESS) {
+    return ret;
   }
-  map<string, GeModelPtr> name_to_ge_model = ge_root_model->GetSubgraphInstanceNameToModel();
-  GeModelPtr &ge_model = name_to_ge_model[ge_root_model->GetRootGraph()->GetName()];
-  GE_RETURN_WITH_LOG_IF_FALSE(ge_model != nullptr, "ge_model cannot be null");
-  ge_model->SetName(model_name);
   ret = impl_->SaveRootModel(file_name_prefix, ge_root_model, model);
   if (ret != SUCCESS) {
     GELOGE(ret, "Save model failed");
@@ -652,6 +695,34 @@ namespace {
     }
     return SUCCESS;
   }
+
+  bool CheckNoAicore(const ComputeGraphPtr &graph) {
+    for (const auto &node : graph->GetDirectNode()) {
+      if (node == nullptr) {
+        continue;
+      }
+      auto op_desc = node->GetOpDesc();
+      if (op_desc == nullptr) {
+        continue;
+      }
+      if (op_desc->GetOpEngineName() == kAIcoreEngine) {
+        return false;
+      }
+    }
+    return true;
+  }
+}
+
+void GeGenerator::RemoveConst(const vector<GeTensor> &inputs, vector<GeTensor> &outputs) {
+  for (auto &input : inputs) {
+    GeTensorDesc input_desc = input.GetTensorDesc();
+    bool is_const = false;
+    (void)AttrUtils::GetBool(input_desc, CONST_ATTR_NAME_INPUT, is_const);
+    bool is_optional = IsOptional(input_desc);
+    if (!is_optional && !is_const) {
+      outputs.emplace_back(input);
+    }
+  }
 }
 
 Status GeGenerator::CheckForSingleOp(OpDescPtr &op_desc, const vector<GeTensor> &inputs,
@@ -710,7 +781,7 @@ Status GeGenerator::BuildSingleOp(OpDescPtr &op_desc, const vector<GeTensor> &in
     auto node = comp_graph->FindNode(op_desc->GetName());
     Status ret = CheckEngineTypeSupport(node, engine_type);
     if (ret != SUCCESS) {
-      GELOGE(ret, "check engine type failed.");
+      GELOGE(ret, "[Check][EngineType]value:%d for node:%s not support", engine_type, node->GetName().c_str());
       return ret;
     }
   }
@@ -718,7 +789,9 @@ Status GeGenerator::BuildSingleOp(OpDescPtr &op_desc, const vector<GeTensor> &in
   GELOGI("ATC parser success in single op build.");
 
   GeRootModelPtr ge_root_model = nullptr;
-  GE_CHK_STATUS_RET_NOLOG(impl_->BuildModel(graph, inputs, ge_root_model));
+  vector<GeTensor> data_inputs;
+  RemoveConst(inputs, data_inputs);
+  GE_CHK_STATUS_RET_NOLOG(impl_->BuildModel(graph, data_inputs, ge_root_model));
   map<string, GeAttrValue> op_attrs = op_desc_tmp->GetAllAttrs();
   GE_CHECK_NOTNULL(ge_root_model);
   GE_CHECK_NOTNULL(ge_root_model->GetRootGraph());
@@ -734,7 +807,7 @@ Status GeGenerator::BuildSingleOp(OpDescPtr &op_desc, const vector<GeTensor> &in
 
   bool all_shape = false;
   (void)AttrUtils::GetBool(op_desc, kAicpuAllshape, all_shape);
-  if (all_shape) {
+  if (all_shape && CheckNoAicore(root_graph)) {
     GELOGD("Get aicpu all_shape kernel!");
     vector<GeTensor> inputs_dynamic;
     vector<GeTensor> outputs_dynamic;
@@ -801,18 +874,19 @@ Status GeGenerator::BuildSingleOpGraph(OpDescPtr &op_desc, const vector<GeTensor
 
   // 2. Create InputData node.
   int32_t arg_index = 0;
+  int32_t data_index = 0;
   if (inputs.empty()) {
     for (const auto &input_desc : op_desc->GetAllInputsDescPtr()) {
       GE_CHECK_NOTNULL_EXEC(input_desc, return INTERNAL_ERROR);
       if (!IsNeedConnectInputOpForSingleOp(*input_desc)) {
         continue;
       }
-      GE_CHK_STATUS_RET_NOLOG(AddInputs(compute_graph, op_node, *input_desc, arg_index, false));
+      GE_CHK_STATUS_RET_NOLOG(AddInputs(compute_graph, op_node, *input_desc, arg_index, false, data_index));
       arg_index++;
     }
   } else {
     for (const auto &in_desc : inputs) {
-      GE_CHK_STATUS_RET_NOLOG(AddInputs(compute_graph, op_node, in_desc.GetTensorDesc(), arg_index, true));
+      GE_CHK_STATUS_RET_NOLOG(AddInputs(compute_graph, op_node, in_desc.GetTensorDesc(), arg_index, true, data_index));
       arg_index++;
     }
   }
@@ -871,13 +945,12 @@ Status GeGenerator::Impl::SaveRootModel(const string &file_name_prefix, GeRootMo
                    "ge root model has no sub model")
   GeModelPtr model_root = nullptr;
   if (is_unknown_shape) {
-    model_root = make_shared<GeModel>();
-    model_root->SetGraph(GraphUtils::CreateGraphFromComputeGraph(ge_root_model->GetRootGraph()));
-    ge_root_model->SetSubgraphInstanceNameToModel(ge_root_model->GetRootGraph()->GetName(), model_root);
-    model_root->SetName(ge_root_model->GetRootGraph()->GetName());
+    auto name_to_ge_model = ge_root_model->GetSubgraphInstanceNameToModel();
+    model_root = name_to_ge_model[ge_root_model->GetRootGraph()->GetName()];
   } else {
     model_root = ge_root_model->GetSubgraphInstanceNameToModel().begin()->second;
   }
+  GE_CHECK_NOTNULL(model_root);
   // set atc version
   if (!SetAtcVersionInfo(*(model_root.get()))) {
     GELOGW("SetPackageVersionInfo of atc failed!");
@@ -915,6 +988,13 @@ Status GeGenerator::Impl::BuildModel(const Graph &graph, const vector<GeTensor>
 
   static std::atomic<uint64_t> atomic_session_id(0);
   auto session_id = atomic_session_id.fetch_add(1);
+  // This is a temporary add for graph with variable
+  auto version = static_cast<int32_t>(SessionVersion::ClOUD_VERSION);
+  ret = VarManager::Instance(session_id)->Init(version, session_id, kDefaultDeviceId, kDefaultJobId);
+  GELOGI("Start init var instance, session_id %lu", session_id);
+  if (ret != SUCCESS) {
+    GELOGW("Failed init var instance, session_id %lu", session_id);
+  }
   if (is_singleop_unregistered_) {
     ret = graph_manager_.BuildGraphForUnregisteredOp(graph_id, inputs, ge_root_model, session_id);
   } else {

ge/graph/build/graph_builder.cc

@@ -382,54 +382,6 @@ Status GraphBuilder::BuildForHostCpuGraph(ComputeGraphPtr &comp_graph, GeModelPt
   return BuildForUnknownShapeGraph(comp_graph, ge_model_ptr, session_id);
 }
 
-static Status InsertMemcpyNode(const ComputeGraphPtr &graph, const OutDataAnchorPtr &out_anchor,
-                               const std::vector<InDataAnchorPtr> &in_anchors, const std::string &name) {
-  GE_CHECK_NOTNULL(out_anchor);
-  NodePtr in_node = out_anchor->GetOwnerNode();
-  GE_CHECK_NOTNULL(in_node);
-  OpDescBuilder op_desc_builder(name, MEMCPYADDRASYNC);
-  OpDescPtr op_desc = op_desc_builder.AddInput("x", in_node->GetOpDesc()->GetOutputDesc(0))
-                                     .AddOutput("y", in_node->GetOpDesc()->GetOutputDesc(0))
-                                     .Build();
-  (void)AttrUtils::SetBool(op_desc, ATTR_NO_NEED_CONSTANT_FOLDING, false);
-  if (GraphUtils::InsertNodeAfter(out_anchor, in_anchors, graph->AddNode(op_desc)) != GRAPH_SUCCESS) {
-    GELOGE(FAILED, "Insert IDENTITY node %s after %s failed.", name.c_str(), in_node->GetName().c_str());
-    return FAILED;
-  }
-  return SUCCESS;
-}
-
-static Status GenerateTaskForConstant(const std::shared_ptr<ComputeGraph> &graph) {
-  for (auto &node : graph->GetDirectNode()) {
-    // CONSTANT not generate task, so insert IDENTITY between CONSTANT and NETOUTPUT
-    auto op_desc = node->GetOpDesc();
-    if (op_desc == nullptr) {
-      continue;
-    }
-    auto op_type = op_desc->GetType();
-    if (op_type == NETOUTPUT) {
-      for (InDataAnchorPtr &in_data_anchor : node->GetAllInDataAnchors()) {
-        const OutDataAnchorPtr &peer_out_anchor = in_data_anchor->GetPeerOutAnchor();
-        GE_IF_BOOL_EXEC(peer_out_anchor == nullptr, continue);
-        NodePtr in_node = peer_out_anchor->GetOwnerNode();
-        GE_CHECK_NOTNULL(in_node);
-
-        std::string in_node_op_type = in_node->GetType();
-        if (in_node_op_type == CONSTANT) {
-          GELOGD("Insert MemcpyAsync node between %s and %s.", in_node->GetName().c_str(), node->GetName().c_str());
-          std::string name = node->GetName() + "_input_" + std::to_string(in_data_anchor->GetIdx()) + "_Memcpy";
-          if (InsertMemcpyNode(graph, peer_out_anchor, {in_data_anchor}, name) != SUCCESS) {
-            GELOGE(FAILED, "Insert memcpy between %s and %s failed.",
-                   in_node->GetName().c_str(), node->GetName().c_str());
-            return FAILED;
-          }
-        }
-      }
-    }
-  }
-  return SUCCESS;
-}
-
 Status GraphBuilder::MarkFpBpProfilingTaskAttr(ComputeGraphPtr &com_graph) {
   bool original_unknown_shape_flag = com_graph->GetGraphUnknownFlag();
   com_graph->SetGraphUnknownFlag(false);
@@ -512,9 +464,6 @@ Status GraphBuilder::BuildForDynamicShapeGraph(ComputeGraphPtr &comp_graph,
         !sub_graph->GetParentGraph()->GetGraphUnknownFlag()) {
       continue;
     }
-
-    GE_CHK_STATUS_RET(GenerateTaskForConstant(sub_graph), "Generate task For constant node in subgraph failed.");
-
     if (sub_graph->GetGraphUnknownFlag()) {
       // unknown shape build flow
       GE_CHK_STATUS_RET(BuildForUnknownShapeGraph(sub_graph, ge_model_ptr, session_id),

ge/graph/build/logical_stream_allocator.cc

@@ -33,13 +33,21 @@ using std::queue;
 namespace ge {
 LogicalStreamPass::LogicalStreamPass(const string &name) : name_(name) {}
 
-const string &LogicalStreamPass::GetName() const { return name_; }
+const string &LogicalStreamPass::GetName() const {
+  return name_;
+}
 
-bool LogicalStreamPass::IsEngineSkip(const Subgraph &subgraph) const { return subgraph.engine_conf.skip_assign_stream; }
+bool LogicalStreamPass::IsEngineSkip(const Subgraph &subgraph) const {
+  return subgraph.engine_conf.skip_assign_stream;
+}
 
-bool LogicalStreamPass::IsEngineAttach(const Subgraph &subgraph) const { return subgraph.engine_conf.attach; }
+bool LogicalStreamPass::IsEngineAttach(const Subgraph &subgraph) const {
+  return subgraph.engine_conf.attach;
+}
 
-bool LogicalStreamPass::IsEngineIndependent(const Subgraph &subgraph) const { return subgraph.engine_conf.independent; }
+bool LogicalStreamPass::IsEngineIndependent(const Subgraph &subgraph) const {
+  return subgraph.engine_conf.independent;
+}
 
 bool LogicalStreamPass::HasStreamLabel(const Subgraph &subgraph) const {
   return !subgraph.subgraph_info.GetStreamLabel().empty();
@@ -60,14 +68,14 @@ Status AssignByLabelPass::Run(ComputeGraphPtr graph, const vector<SubgraphPtr> &
       // Subgraphs of the same stream_label are assigned to the same stream,
       // and different stream_labels are assigned new streams.
       auto iter = label_streams.find(stream_label);
-      if (iter != label_streams.end()) {
-        subgraph->stream_id = iter->second;
-      } else {
+      if (iter == label_streams.end()) {
         subgraph->stream_id = next_stream;
         GELOGI("Assign new stream %ld for label %s.", next_stream, stream_label.c_str());
 
         label_streams.emplace(stream_label, next_stream);
-        ++next_stream;
+        next_stream++;
+      } else {
+        subgraph->stream_id = iter->second;
       }
       changed = true;
     }
@@ -92,15 +100,15 @@ Status IndependentStreamPass::Run(ComputeGraphPtr graph, const vector<SubgraphPt
     const string &stream_label = subgraph->subgraph_info.GetStreamLabel();
     auto &label_streams = engine_streams[engine];
     auto iter = label_streams.find(stream_label);
-    if (iter != label_streams.end()) {
-      subgraph->stream_id = iter->second;
-    } else {
+    if (iter == label_streams.end()) {
       subgraph->stream_id = next_stream;
       GELOGI("Assign new independent stream %ld for engine %s (label: %s).", next_stream, engine.c_str(),
              stream_label.c_str());
 
       label_streams.emplace(stream_label, next_stream);
-      ++next_stream;
+      next_stream++;
+    } else {
+      subgraph->stream_id = iter->second;
     }
     changed = true;
   }
@@ -121,7 +129,9 @@ Status AssignByDependencyPass::Run(ComputeGraphPtr graph, const vector<SubgraphP
     }
 
     SubgraphPtr reusable_subgraph = GetReusableSubgraph(subgraph, end_subgraph_map, pld_subgraph_map);
-    if (reusable_subgraph != nullptr) {
+    if (reusable_subgraph == nullptr) {
+      (void)AssignNewStream(subgraph);
+    } else {
       if (HasAssignedStream(*reusable_subgraph)) {
         subgraph->stream_id = reusable_subgraph->stream_id;
       } else {
@@ -140,8 +150,6 @@ Status AssignByDependencyPass::Run(ComputeGraphPtr graph, const vector<SubgraphP
       GELOGI("Subgraph %s of engine %s reuses stream of subgraph %s of engine %s.", subgraph->name.c_str(),
              subgraph->engine_conf.id.c_str(), reusable_subgraph->name.c_str(),
              reusable_subgraph->engine_conf.id.c_str());
-    } else {
-      (void)AssignNewStream(subgraph);
     }
     changed = true;
   }
@@ -191,13 +199,15 @@ bool AssignByDependencyPass::CouldReuse(const SubgraphPtr &subgraph, const Subgr
     auto iter = pld_subgraph_map.find(end_pld_pair.second);
     if (iter != pld_subgraph_map.end()) {
       const SubgraphPtr &pred_subgraph_succ = iter->second;
-      if (pred_subgraph_succ != subgraph && pred_subgraph_succ->engine_conf.id == pred_subgraph->engine_conf.id) {
+      if ((pred_subgraph_succ != subgraph) &&
+          (pred_subgraph_succ->engine_conf.id == pred_subgraph->engine_conf.id)) {
         return false;
       }
     }
   }
 
-  if ((subgraph->engine_conf.id == pred_subgraph->engine_conf.id) || IsEngineAttach(*subgraph)) {
+  if ((subgraph->engine_conf.id == pred_subgraph->engine_conf.id) ||
+      IsEngineAttach(*subgraph)) {
     return true;
   }
 
@@ -406,7 +416,7 @@ Status UpdateForSkippedEnginePass::Run(ComputeGraphPtr graph, const vector<Subgr
         auto op_desc = node->GetOpDesc();
         GE_CHECK_NOTNULL(op_desc);
         auto stream_id = op_desc->GetStreamId();
-        if (stream_id != kInvalidStream && !HasStreamLabel(*subgraph)) {
+        if ((stream_id != kInvalidStream) && !HasStreamLabel(*subgraph)) {
           ops_without_label.emplace(op_desc);
         }
       }
@@ -463,7 +473,7 @@ Status AllReduceParallelPass::Run(ComputeGraphPtr graph, const vector<SubgraphPt
 
   for (const NodePtr &node : graph->GetDirectNode()) {
     if (!IsHcomNode(node->GetType()) ||
-        node->GetInDataNodes().size() <= 1) {
+        (node->GetInDataNodes().size() <= 1)) {
       continue;
     }
 
@@ -575,7 +585,7 @@ Status LogicalStreamAllocator::DoAssign(const ComputeGraphPtr &graph, const Grap
   GE_CHECK_NOTNULL(graph);
 
   NodePtr parent_node = graph->GetParentNode();
-  if (parent_node == nullptr || parent_node->GetOpDesc() == nullptr) {
+  if ((parent_node == nullptr) || (parent_node->GetOpDesc() == nullptr)) {
     context_.default_stream = kInvalidStream;
   } else {
     context_.default_stream = parent_node->GetOpDesc()->GetStreamId();
@@ -597,7 +607,7 @@ Status LogicalStreamAllocator::DoAssign(const ComputeGraphPtr &graph, const Grap
     return status;
   }
 
-  GELOGD("Subgraphs of graph %s:", graph->GetName().c_str());
+  GELOGD("Subgraphs of graph %s", graph->GetName().c_str());
   for (const auto &subgraph : subgraphs) {
     if (subgraph != nullptr) {
       GELOGD("subgraph: %s", subgraph->name.c_str());
@@ -686,7 +696,7 @@ void LogicalStreamAllocator::RefreshContinuousStreams(const ComputeGraphPtr &gra
       auto op_desc = node->GetOpDesc();
       if (op_desc != nullptr) {
         int64_t stream_id = op_desc->GetStreamId();
-        if (stream_id != kInvalidStream && stream_id < stream_num) {
+        if ((stream_id != kInvalidStream) && (stream_id < stream_num)) {
           stream_has_node[stream_id] = true;
         }
       }
@@ -695,10 +705,10 @@ void LogicalStreamAllocator::RefreshContinuousStreams(const ComputeGraphPtr &gra
 
   context_.next_stream = 0;
   vector<int64_t> old_to_new_streams(stream_num, kInvalidStream);
-  for (size_t old_stream = 0; old_stream < stream_has_node.size(); ++old_stream) {
+  for (size_t old_stream = 0; old_stream < stream_has_node.size(); old_stream++) {
     if (stream_has_node[old_stream]) {
       old_to_new_streams[old_stream] = context_.next_stream;
-      ++context_.next_stream;
+      context_.next_stream++;
     }
   }
 
@@ -706,7 +716,7 @@ void LogicalStreamAllocator::RefreshContinuousStreams(const ComputeGraphPtr &gra
     auto op_desc = node->GetOpDesc();
     if (op_desc != nullptr) {
       int64_t stream_id = op_desc->GetStreamId();
-      if (stream_id != kInvalidStream && stream_id < stream_num) {
+      if ((stream_id != kInvalidStream) && (stream_id < stream_num)) {
         op_desc->SetStreamId(old_to_new_streams[stream_id]);
       }
     }

ge/graph/build/memory/binary_block_mem_assigner.cc

@@ -70,7 +70,10 @@ Status BinaryBlockMemAssigner::GetMemoryRanges(vector<int64_t> &range_ceils) {
     return SUCCESS;
   }
   if ((all_memory_size.front() <= 0) || (log(kLogBase) == 0)) {
-    GELOGE(FAILED, "Memory size:%ld is invalid.", all_memory_size.front());
+    GELOGE(FAILED, "[Check][MemRangeStep]first mem_range_step:%ld less than 0,invalid,"
+          "maybe has dynamic shape in graph", all_memory_size.front());
+    REPORT_INNER_ERROR("E19999", "first mem_range_step:%ld less than 0,invalid,"
+          "maybe has dynamic shape in graph", all_memory_size.front());
     return FAILED;
   }
   // Memory size is 512 aligned, so it is not necessary to take less than 512
@@ -81,12 +84,18 @@ Status BinaryBlockMemAssigner::GetMemoryRanges(vector<int64_t> &range_ceils) {
   GELOGD("Range number: %zu", range_number);
 
   vector<vector<int64_t>> ranges(range_number);
-  GE_CHK_BOOL_EXEC((range_number != 0), return PARAM_INVALID, "range_number can't be 0.");
+  GE_CHK_BOOL_EXEC((range_number != 0),
+    REPORT_INNER_ERROR("E19999", "inner data[range_number] is 0, judge invalid");
+    return PARAM_INVALID,
+    "[Check][RangeNumber]inner data is 0, judge invalid.");
   size_t range_number_limit = all_memory_size.size() / range_number;
   int64_t range_ceil = min_memory_size;
   for (size_t i = 1; i <= range_number; i++) {
     GE_IF_BOOL_EXEC(TypeUtils::CheckUint64MulOverflow(static_cast<uint64_t>(range_ceil), kRangeCeilInterval),
-                    GELOGE(FAILED, "Multiply result is out of range.");
+                    GELOGE(FAILED, "[Check][MemRangeCeil]Multiply result is out of range,"
+                      "range_ceil:%ld, interval:%u", range_ceil, kRangeCeilInterval);
+                    REPORT_INNER_ERROR("E19999", "process mem_range_ceil,multiply result out of range,"
+                      "range_ceil:%ld, interval:%u", range_ceil, kRangeCeilInterval);
                     return FAILED);
     range_ceil *= kRangeCeilInterval;  // The block size of each interval is doubled every time.
     for (auto iter = all_memory_size.begin(); iter != all_memory_size.end();) {

ge/graph/build/memory/block_mem_assigner.cc

@@ -30,6 +30,7 @@
 #include "graph/utils/node_utils.h"
 #include "graph/utils/op_desc_utils.h"
 #include "graph/utils/tensor_utils.h"
+#include "graph/utils/type_utils.h"
 
 #include "graph/debug/ge_attr_define.h"
 
@@ -457,7 +458,16 @@ Status GetNoAlignSize(const ge::OpDesc &desc, uint32_t index, size_t &size) {
   DataType data_type = output_op_desc->GetDataType();
   graphStatus graph_status = TensorUtils::CalcTensorMemSize(shape, format, data_type, tensor_size);
   if (graph_status != GRAPH_SUCCESS) {
-    GELOGE(graph_status, "CalcTensorMemSize failed!");
+    GELOGE(graph_status, "[Calculate][TensorSize]shape:%s, format:%s, data_type:%s, op:%s, out_index:%u",
+           shape.ToString().c_str(),
+           TypeUtils::FormatToSerialString(format).c_str(),
+           TypeUtils::DataTypeToSerialString(data_type).c_str(),
+           desc.GetName().c_str(), index);
+    REPORT_CALL_ERROR("E19999", "CalcTensorMemSize fail, shape:%s, format:%s, data_type:%s, op:%s, out_index:%u",
+                      shape.ToString().c_str(),
+                      TypeUtils::FormatToSerialString(format).c_str(),
+                      TypeUtils::DataTypeToSerialString(data_type).c_str(),
+                      desc.GetName().c_str(), index);
     return FAILED;
   }
   size = static_cast<size_t>(tensor_size);
@@ -586,8 +596,13 @@ void BlockMemAssigner::GetOutAndWorkSpaceMem(vector<int64_t> &all_memory_size) {
       GeTensorDesc output_desc = node_op_desc->GetOutputDesc(out_anchor->GetIdx());
       int64_t size = 0;
       GE_IF_BOOL_EXEC(ge::TensorUtils::GetSize(output_desc, size) != SUCCESS, GELOGI("Get size failed"));
-      GE_IF_BOOL_EXEC(size < 0, GELOGE(FAILED, "Node:%s size:%ld is invalid, maybe it is unknown shape node.",
-                                       node_op_desc->GetName().c_str(), size);
+      GE_IF_BOOL_EXEC(size < 0,
+                      GELOGE(FAILED, "[Check][TensorSize]tensor_size:%ld is invalid, "
+                             "maybe it is unknown shape node, Node_name:%s",
+                             size, node_op_desc->GetName().c_str());
+                      REPORT_INNER_ERROR("E19999", "tensor_size:%ld is invalid, "
+                                         "maybe it is unknown shape node, Node_name:%s",
+                                         size, node_op_desc->GetName().c_str());
                       return;);
       batch_all_memory_size[batch_label].emplace_back(size);
       if (batch_total_size.find(batch_label) == batch_total_size.end()) {
@@ -678,22 +693,34 @@ bool BlockMemAssigner::IsOutNodeSetContinuousInput(const NodePtr &n, uint32_t ou
   if (static_cast<size_t>(out_index) < n->GetAllOutDataAnchors().size()) {
     auto out_anchor = n->GetOutDataAnchor(out_index);
     GE_IF_BOOL_EXEC(out_anchor == nullptr,
-                    GELOGE(FAILED, "Node[%s] output[%u] anchor is null.", n->GetName().c_str(), out_index);
+                    GELOGE(FAILED, "[Check][Anchor]Node[%s] output[%u] anchor is null.",
+                           n->GetName().c_str(), out_index);
+                    REPORT_INNER_ERROR("E19999", "output anchor is null, node_name: %s output_index: %u.",
+                                       n->GetName().c_str(), out_index);
                     return false;);
     for (auto const &peer_in_anchor : out_anchor->GetPeerInDataAnchors()) {
       GE_IF_BOOL_EXEC(peer_in_anchor == nullptr,
-                      GELOGE(FAILED, "Node[%s] output[%u] peer_in_anchor 0 is null.", n->GetName().c_str(), out_index);
+                      GELOGE(FAILED, "[Check][Anchor]Node[%s] output[%u] peer_in_anchor 0 is null.",
+                             n->GetName().c_str(), out_index);
+                      REPORT_INNER_ERROR("E19999", "output anchor peer is null, node_name: %s output_index: %u.",
+                                         n->GetName().c_str(), out_index);
                       return false;);
       auto peer_node = peer_in_anchor->GetOwnerNode();
       GE_IF_BOOL_EXEC(peer_node == nullptr,
-                      GELOGE(FAILED, "Node[%s] output[%u] node is null.", n->GetName().c_str(), out_index);
+                      GELOGE(FAILED, "[Check][Node]Node[%s] output[%u] peer node is null.",
+                             n->GetName().c_str(), out_index);
+                      REPORT_INNER_ERROR("E19999", "output anchor peer node is null, node_name: %s output_index: %u.",
+                                         n->GetName().c_str(), out_index);
                       return false;);
 
       // Get the continuous input type of the node, default is false
       bool is_input_continuous = false;
       auto peer_in_node_desc = peer_node->GetOpDesc();
       GE_IF_BOOL_EXEC(peer_in_node_desc == nullptr,
-                      GELOGE(FAILED, "Node[%s] output[%u] nodedesc is null.", n->GetName().c_str(), out_index);
+                      GELOGE(FAILED, "[Check][OpDesc]Node[%s] output[%u] nodedesc is null.",
+                             n->GetName().c_str(), out_index);
+                      REPORT_INNER_ERROR("E19999", "output anchor peer op_desc is null, node_name:%s output_index:%u.",
+                                         n->GetName().c_str(), out_index);
                       return false;);
 
       // If GetBool fail, is_input_continuous is false.
@@ -793,7 +820,10 @@ bool BlockMemAssigner::IsContinuousMemoryReuse(const NodePtr &n, const NodePtr &
     if ((in_anchor == nullptr) || (in_anchor->GetPeerOutAnchor() == nullptr) ||
         (in_anchor->GetPeerOutAnchor()->GetOwnerNode() == nullptr) ||
         (in_anchor->GetPeerOutAnchor()->GetOwnerNode()->GetOpDesc() == nullptr)) {
-      GELOGE(FAILED, "Node[%s] output[%u] peer input node desc is null.", n->GetName().c_str(), out_index);
+      GELOGE(FAILED, "[Check][OpDesc]Node[%s] output[%u] peer input node desc is null.",
+             n->GetName().c_str(), out_index);
+      REPORT_INNER_ERROR("E19999", "get output anchor peer op_desc fail, node_name: %s output_index: %u.",
+                         n->GetName().c_str(), out_index);
       return false;
     }
     auto peer_out_node_desc = in_anchor->GetPeerOutAnchor()->GetOwnerNode()->GetOpDesc();
@@ -1077,7 +1107,10 @@ MemoryBlock *BlockMemAssigner::ApplyMemory(size_t block_size, size_t real_size,
                                            OpMemoryType mem_type, const NodePtr &n, uint32_t out_index,
                                            const vector<bool> &workspace_reuse_flag, const bool is_op_reuse_mem,
                                            const bool continuous, int64_t memory_type) {
-  GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(n == nullptr, return nullptr, "Input parameter n is null.");
+  GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(
+      n == nullptr,
+      REPORT_INNER_ERROR("E19999", "Input parameter n(type:node_ptr) is null, apply memory failed");
+      return nullptr, "[Check][Param]Input parameter n(type:node_ptr) is null.");
   auto node_op_desc = n->GetOpDesc();
   GE_IF_BOOL_EXEC(node_op_desc == nullptr, return nullptr);
   std::string batch_label;
@@ -1129,7 +1162,12 @@ MemoryBlock *BlockMemAssigner::ApplyMemory(size_t block_size, size_t real_size,
   }
 
   auto block = new (std::nothrow) MemoryBlock(block_size, node_op_desc->GetStreamId(), is_reuse_memory, memory_type);
-  GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(block == nullptr, return nullptr, "new an object failed.");
+  GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(
+      block == nullptr,
+      REPORT_INNER_ERROR("E19999", "new a memoryblock object failed. node_name:%s out_index:%u",
+                         n->GetName().c_str(), out_index);
+      return nullptr,
+      "[New][Object]new MemoryBlock failed, node_name:%s out_index:%u", n->GetName().c_str(), out_index);
 
   // Data and netoutput need zero copy block
   block->is_zero_copy_ = IsZeroCopyBlock(n, continuous);
@@ -1188,9 +1226,15 @@ void BlockMemAssigner::ContinuousOutRefCheck(bool &isAllOutputRef, bool &isOutpu
 
 Status BlockMemAssigner::ApplyContinuousMemory(const NodePtr &n, const vector<int64_t> &ranges,
                                                const bool is_op_reuse_mem) {
-  GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(n == nullptr, return INTERNAL_ERROR, "input node is null.");
+  GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(
+      n == nullptr,
+      REPORT_INNER_ERROR("E19999", "Input parameter n(type:node_ptr) is null");
+      return INTERNAL_ERROR, "[check][param]Input parameter n(type:NodePtr) is null.");
   auto node_op_desc = n->GetOpDesc();
-  GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(node_op_desc == nullptr, return INTERNAL_ERROR, "node_op_desc is null.");
+  GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(
+      node_op_desc == nullptr,
+      REPORT_INNER_ERROR("E19999", "Input parameter n(type:OpDescPtr) is null");
+      return INTERNAL_ERROR, "[Check][Param]Input parameter n(type:OpDescPtr) is null");
 
   // continuous output support ref only when all output ref input
   bool isAllOutputRef = true;
@@ -1204,7 +1248,9 @@ Status BlockMemAssigner::ApplyContinuousMemory(const NodePtr &n, const vector<in
   }
 
   if (!isAllOutputRef && isOutputHasRef) {
-    GELOGE(INTERNAL_ERROR, "continuous output node ref part input, not support this situation, node_name:%s",
+    REPORT_INNER_ERROR("E19999", "continuous output node ref part input, not support now. node_name:%s",
+                       n->GetName().c_str());
+    GELOGE(INTERNAL_ERROR, "[Check][OutRefStatus]continuous output node ref part input, not support, node_name:%s",
            n->GetName().c_str());
     return INTERNAL_ERROR;
   }
@@ -1215,7 +1261,9 @@ Status BlockMemAssigner::ApplyContinuousMemory(const NodePtr &n, const vector<in
   for (uint32_t index = 0; index < static_cast<uint32_t>(node_op_desc->GetOutputsSize()); index++) {
     auto output_op_desc = node_op_desc->GetOutputDescPtr(index);
     if (output_op_desc == nullptr) {
-      GELOGE(INTERNAL_ERROR, "Get output desc failed, node_name:%s, output_index:%u", n->GetName().c_str(), index);
+      REPORT_INNER_ERROR("E19999", "get output_desc failed, node_name:%s, output_index:%u",
+                         n->GetName().c_str(), index);
+      GELOGE(INTERNAL_ERROR, "[Get][OutputDesc]node_name:%s, output_index:%u", n->GetName().c_str(), index);
       return INTERNAL_ERROR;
     }
 
@@ -1226,7 +1274,9 @@ Status BlockMemAssigner::ApplyContinuousMemory(const NodePtr &n, const vector<in
 
     int64_t size = 0;
     if (ge::TensorUtils::GetSize(*output_op_desc, size) != SUCCESS) {
-      GELOGE(INTERNAL_ERROR, "Get size failed, node_name:%s, output_index:%u", n->GetName().c_str(), index);
+      REPORT_CALL_ERROR("E19999", "get tensor_size failed, node_name:%s, output_index:%u",
+                        n->GetName().c_str(), index);
+      GELOGE(INTERNAL_ERROR, "[Get][TensorSize]node_name:%s, output_index:%u", n->GetName().c_str(), index);
       return INTERNAL_ERROR;
     }
     size_t align_size = static_cast<size_t>(size);
@@ -1266,7 +1316,9 @@ Status BlockMemAssigner::ApplyContinuousMemory(const NodePtr &n, const vector<in
     block->last_continuous_block_ = true;
     ++(block->ref_count_);
   } else {
-    GELOGE(INTERNAL_ERROR, "node apply continuous output memory failed. node_name:%s", n->GetName().c_str());
+    REPORT_CALL_ERROR("E19999", "apply continuousMemory failed, node_name:%s, total_size:%ld",
+                      n->GetName().c_str(), total_size);
+    GELOGE(INTERNAL_ERROR, "[Apply][ContinuousMemory]node_name:%s, total_size:%ld", n->GetName().c_str(), total_size);
     return INTERNAL_ERROR;
   }
   return SUCCESS;
@@ -1274,25 +1326,44 @@ Status BlockMemAssigner::ApplyContinuousMemory(const NodePtr &n, const vector<in
 
 MemoryBlock *BlockMemAssigner::ApplyOutMemory(const NodePtr &n, uint32_t index, const vector<int64_t> &ranges,
                                               const bool is_op_reuse_mem, const bool continuous) {
-  GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(n == nullptr, return nullptr, "input node is null.");
+  GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(
+      n == nullptr,
+      REPORT_INNER_ERROR("E19999", "Input parameter n(type:NodePtr) is null");
+      return nullptr, "[Check][Param]Input parameter n(type:NodePtr) is null");
   auto node_op_desc = n->GetOpDesc();
-  GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(node_op_desc == nullptr, return nullptr, "node_op_desc is null.");
+  GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(
+      node_op_desc == nullptr,
+      REPORT_INNER_ERROR("E19999", "Input parameter n(type:OpDescPtr) is null");
+      return nullptr, "[Check][Param]Input parameter n(type:OpDescPtr) is null");
   MemoryBlock *block = nullptr;
   NodeIndexIO node_index_io(n, index, kOut);
   int64_t size = 0;
   auto output_op_desc = node_op_desc->GetOutputDescPtr(index);
-  GE_IF_BOOL_EXEC(output_op_desc == nullptr, return nullptr);
+  GE_IF_BOOL_EXEC(
+      output_op_desc == nullptr,
+      REPORT_INNER_ERROR("E19999", "get output_desc failed, node_name:%s, output_index:%u",
+                         n->GetName().c_str(), index);
+      GELOGE(FAILED, "[Get][OutputDesc]node_name:%s, output_index:%u", n->GetName().c_str(), index);
+      return nullptr);
   GE_IF_BOOL_EXEC(ge::TensorUtils::GetSize(*output_op_desc, size) != SUCCESS, GELOGI("Get size failed"));
   size_t no_align_size = 0;
-  GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(GetNoAlignSize(*node_op_desc, index, no_align_size) != SUCCESS,
-                                 return nullptr, "Get no align size failed");
+  GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(
+      GetNoAlignSize(*node_op_desc, index, no_align_size) != SUCCESS,
+      REPORT_CALL_ERROR("E19999", "Get no align size failed, node_name:%s, output_index:%u",
+                        n->GetName().c_str(), index);
+      return nullptr,
+      "[Get][TensorSize]Get no align size, node_name:%s, output_index:%u", n->GetName().c_str(), index);
 
   std::string symbol;
   bool reuse_input = false;
   if (IsSymbolExist(node_index_io, symbol)) {
     block = symbol_blocks_[symbol];
-    GE_IF_BOOL_EXEC(block == nullptr, GELOGE(FAILED, "Node %s ref block is nullptr.", node_op_desc->GetName().c_str());
-        return nullptr);
+    GE_IF_BOOL_EXEC(block == nullptr,
+      REPORT_INNER_ERROR("E19999", "get ref block failed, node_name:%s, symbol:%s",
+                         node_op_desc->GetName().c_str(), node_index_io.ToString().c_str());
+      GELOGE(FAILED, "[Get][RefBlock]node_name:%s, symbol:%s",
+             node_op_desc->GetName().c_str(), node_index_io.ToString().c_str());
+      return nullptr);
     // reduce old size
     size_t align_size = block->Size();
     AlignMemOffset(align_size);
@@ -1335,12 +1406,28 @@ MemoryBlock *BlockMemAssigner::ApplyOutMemory(const NodePtr &n, uint32_t index,
     vector<bool> workspace_reuse_flag;
     block = ApplyMemory(block_size, size, no_align_size, kOutput, n, index,
                         workspace_reuse_flag, is_op_reuse_mem, continuous, memory_type);
+    GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(
+        block == nullptr,
+        REPORT_CALL_ERROR("E19999", "apply out Memory failed, node_name:%s, block_size:%ld, out_index:%u",
+                          n->GetName().c_str(), block_size, index);
+        return nullptr,
+        "[Apply][Memory]node_name:%s, block_size:%ld, out_index:%u",
+        n->GetName().c_str(), block_size, index);
   }
-  GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(block == nullptr, return nullptr, "Block is nullptr.");
   int out_count = 0;
-  GE_IF_BOOL_EXEC(index >= n->GetAllOutDataAnchors().size(), GELOGE(FAILED, "index is out of range."); return nullptr);
+  GE_IF_BOOL_EXEC(
+      index >= n->GetAllOutDataAnchors().size(),
+      REPORT_INNER_ERROR("E19999", "out index:%u exceed out_size:%lu, node_name:%s",
+                         index, n->GetAllOutDataAnchors().size(), n->GetName().c_str());
+      GELOGE(FAILED, "[Check][OutIndex]index:%u exceed out_size:%lu, node_name:%s",
+             index, n->GetAllOutDataAnchors().size(), n->GetName().c_str());
+      return nullptr);
   auto out_data_anchor = n->GetOutDataAnchor(index);
-  GE_IF_BOOL_EXEC(out_data_anchor == nullptr, GELOGE(FAILED, "Out data anchor is nullptr."); return nullptr);
+  GE_IF_BOOL_EXEC(
+      out_data_anchor == nullptr,
+      REPORT_INNER_ERROR("E19999", "out anchor is null, index:%u, node_name:%s", index, n->GetName().c_str());
+      GELOGE(FAILED, "[Check][OutAnchor]is null, index:%u, node_name:%s", index, n->GetName().c_str());
+      return nullptr);
   for (const auto &in_anchor : out_data_anchor->GetPeerInDataAnchors()) {
     auto owner_node = in_anchor->GetOwnerNode();
     auto op_desc = owner_node->GetOpDesc();
@@ -1546,8 +1633,14 @@ Status BlockMemAssigner::AssignOutputMemoryWithReuse(const NodePtr &node, vector
   GELOGD("Assign memory node[%s], output size[%zu], output memory type size[%zu]", op_desc->GetName().c_str(),
          op_desc->GetOutputsSize(), memorys_type.size());
   if (has_mem_type_attr && (memorys_type.size() != op_desc->GetOutputsSize())) {
-    GELOGE(INTERNAL_ERROR, "fusion: node[%s], output memory size err[outputsize:%zu, memorysize:%zu]",
-           op_desc->GetName().c_str(), op_desc->GetOutputsSize(), memorys_type.size());
+    REPORT_INNER_ERROR("E19999", "Attr[%s] size:%zu not equal to node output size:%zu, node_name:%s",
+                       ATTR_NAME_OUTPUT_MEM_TYPE_LIST.c_str(), memorys_type.size(),
+                       op_desc->GetOutputsSize(), op_desc->GetName().c_str());
+    GELOGE(
+        INTERNAL_ERROR,
+        "[Check][MemTypeAttr]Attr %s size:%zu not equal to node output size:%zu, node_name:%s",
+        ATTR_NAME_OUTPUT_MEM_TYPE_LIST.c_str(), memorys_type.size(),
+        op_desc->GetOutputsSize(), op_desc->GetName().c_str());
     return INTERNAL_ERROR;
   }
 
@@ -1673,8 +1766,12 @@ void BlockMemAssigner::AssignMemoryWithReuse(vector<int64_t> &ranges) {
            temp.size(), tvm_workspace_memory_type.size());
 
     if (has_tvm_workspace_mem_type_attr && (temp.size() != tvm_workspace_memory_type.size())) {
-      GELOGE(INTERNAL_ERROR, "fusion: node[%s], tvm workspace memory size error![v_temp:%zu, workspace:%zu]",
-             n->GetName().c_str(), temp.size(), tvm_workspace_memory_type.size());
+      REPORT_INNER_ERROR("E19999", "Attr[%s]size:%zu is not equal to workspace size:%zu, node_name:%s",
+                         TVM_ATTR_NAME_WORKSPACE_TYPE.c_str(), tvm_workspace_memory_type.size(),
+                         temp.size(), n->GetName().c_str());
+      GELOGE(INTERNAL_ERROR, "[Check][Attr]Attr %s size:%zu is not equal to workspace size:%zu, node_name:%s",
+             TVM_ATTR_NAME_WORKSPACE_TYPE.c_str(), tvm_workspace_memory_type.size(),
+             temp.size(), n->GetName().c_str());
       return;
     }
     for (size_t i = 0; i < temp.size(); i++) {
@@ -2083,8 +2180,12 @@ bool BlockMemAssigner::GetWorkSpaceMemoryType(const NodePtr &node, size_t index,
   bool has_workspace_mem_type_attr =
       ge::AttrUtils::GetListInt(op_desc, TVM_ATTR_NAME_WORKSPACE_TYPE, workspace_memory_type);
   if (has_workspace_mem_type_attr && (workspace_memory_type.size() <= index)) {
-    GELOGE(INTERNAL_ERROR, "node[%s], workspace_memory size error![index:%zu, workspace:%zu]",
-           node->GetName().c_str(), index, workspace_memory_type.size());
+    REPORT_INNER_ERROR("E19999", "get workspace mem_type failed, "
+                       "index %zu invalid, bigger than attr %s size:%zu, node_name:%s",
+                       index, TVM_ATTR_NAME_WORKSPACE_TYPE.c_str(),
+                       workspace_memory_type.size(), node->GetName().c_str());
+    GELOGE(INTERNAL_ERROR, "[Get][WorkspaceMemType]index %zu invalid, bigger than attr %s size:%zu, node_name:%s",
+           index, TVM_ATTR_NAME_WORKSPACE_TYPE.c_str(), workspace_memory_type.size(), node->GetName().c_str());
     return false;
   }
   memory_type = has_workspace_mem_type_attr ? workspace_memory_type[index] : RT_MEMORY_HBM;

ge/graph/build/memory/graph_mem_assigner.cc

@@ -99,7 +99,8 @@ Status VariableMemoryAssigner::AssignMemory2HasRefAttrNode() {
 Status GraphMemoryAssigner::AssignMemory() {
   ge::HybridMemAssignerPtr mem_assigner(new(std::nothrow) HybridMemAssigner(compute_graph_));
   if (mem_assigner->Assign() != ge::SUCCESS) {
-    GELOGE(ge::FAILED, "Memory assigner failed");
+    GELOGE(ge::FAILED, "[Assign][GraphMem]graph_id:%u, graph_name:%s",
+           compute_graph_->GetGraphID(), compute_graph_->GetName().c_str());
     return ge::FAILED;
   }
   MemoryOffset memory_offset(RT_MEMORY_HBM, mem_assigner->GetMemOffset());
@@ -115,7 +116,10 @@ Status GraphMemoryAssigner::AssignMemory() {
   auto variable_assigner =
       std::unique_ptr<ge::VariableMemoryAssigner>(new(std::nothrow) ge::VariableMemoryAssigner(compute_graph_));
   if (variable_assigner == nullptr) {
-    GELOGE(ge::FAILED, "Alloc VariableMemoryAssigner failed.");
+    GELOGE(ge::FAILED, "[New][Object:VariableMemoryAssigner]graph_id:%u, graph_name:%s",
+           compute_graph_->GetGraphID(), compute_graph_->GetName().c_str());
+    REPORT_INNER_ERROR("E19999", "New Object:VariableMemoryAssigner failed when assign graph memory, "
+                       "graph_id:%u, graph_name:%s", compute_graph_->GetGraphID(), compute_graph_->GetName().c_str());
     return ge::FAILED;
   }
 
@@ -134,7 +138,10 @@ ge::Status GraphMemoryAssigner::AssignVarAttr2Nodes() {
   auto variable_assigner =
       std::unique_ptr<ge::VariableMemoryAssigner>(new(std::nothrow) ge::VariableMemoryAssigner(compute_graph_));
   if (variable_assigner == nullptr) {
-    GELOGE(ge::FAILED, "Alloc VariableMemoryAssigner failed.");
+    GELOGE(ge::FAILED, "[New][Object:VariableMemoryAssigner]graph_id:%u, graph_name:%s",
+           compute_graph_->GetGraphID(), compute_graph_->GetName().c_str());
+    REPORT_INNER_ERROR("E19999", "New Object:VariableMemoryAssigner failed when assign graph memory, "
+                       "graph_id:%u, graph_name:%s", compute_graph_->GetGraphID(), compute_graph_->GetName().c_str());
     return ge::FAILED;
   }
   if (variable_assigner->AssignVarAttr2Nodes() != ge::SUCCESS) {
@@ -147,8 +154,10 @@ ge::Status GraphMemoryAssigner::AssignMemory2HasRefAttrNode() {
   auto variable_assigner =
       std::unique_ptr<ge::VariableMemoryAssigner>(new(std::nothrow) ge::VariableMemoryAssigner(compute_graph_));
   if (variable_assigner == nullptr) {
-    GELOGE(ge::FAILED, "Alloc VariableMemoryAssigner failed.");
-    return ge::FAILED;
+    GELOGE(ge::FAILED, "[New][Object:VariableMemoryAssigner]graph_id:%u, graph_name:%s",
+           compute_graph_->GetGraphID(), compute_graph_->GetName().c_str());
+    REPORT_INNER_ERROR("E19999", "New Object:VariableMemoryAssigner failed when assign graph memory, "
+                       "graph_id:%u, graph_name:%s", compute_graph_->GetGraphID(), compute_graph_->GetName().c_str());
   }
   if (variable_assigner->AssignMemory2HasRefAttrNode() != ge::SUCCESS) {
     return ge::FAILED;
@@ -161,17 +170,18 @@ ge::Status CalculateTensorRealSizeAndOutSize(const ge::ConstGeTensorDescPtr &out
                                              int64_t &batch_dim_num, int64_t &out_size) {
   graphStatus graph_status = ge::TensorUtils::GetSize(*output_desc, out_size);
   if (graph_status != GRAPH_SUCCESS) {
-    GELOGE(FAILED, "Opdesc GetSize failed!");
+    GELOGE(FAILED, "[Get][TensorSize]");
+    REPORT_INNER_ERROR("E19999", "New Object:VariableMemoryAssigner failed when assign graph memory");
     return FAILED;
   }
 
   GeShape output_shape = output_desc->GetShape();
   std::vector<int64_t> output_dims = output_shape.GetDims();
   if (dim_index >= static_cast<int64_t>(output_dims.size())) {
-    std::string error = "Invaild value" + FmtToStr(dim_index) +
-        " of attr _reuse_input_on_dim_index, which is out of data range [0,"
-        + std::to_string(output_dims.size()) + ")";
-    GE_ERRORLOG_AND_ERRORMSG(FAILED, error.c_str());
+    REPORT_INNER_ERROR("E19999", "Inner param dim_index value:%ld invalid, bigger than dim size:%lu in shape:%s",
+                       dim_index, output_dims.size(), output_shape.ToString().c_str());
+    GELOGE(FAILED, "[Check][Param:dim_index]value:%ld invalid, bigger than dim size:%lu in shape:%s",
+           dim_index, output_dims.size(), output_shape.ToString().c_str());
     return FAILED;
   }
 
@@ -187,14 +197,23 @@ ge::Status CalculateTensorRealSizeAndOutSize(const ge::ConstGeTensorDescPtr &out
 
   graph_status = ge::TensorUtils::CalcTensorMemSize(output_shape, out_format, data_type, output_mem_size);
   if (graph_status != GRAPH_SUCCESS) {
-    GELOGE(graph_status, "Opdesc CalcTensorMemSize failed!");
+    GELOGE(graph_status, "[Calc][TensorSize]");
     return FAILED;
   }
 
   if (output_mem_size < 0) {
-    std::string error = "After calculating tensor memory size, output_mem_size" + FmtToStr(output_mem_size) +
-        " is out of data range [0," + std::to_string(INT64_MAX) + "]";
-    GE_ERRORLOG_AND_ERRORMSG(FAILED, error.c_str());
+    REPORT_INNER_ERROR("E19999", "After calculating, tensor memory size:%ld invalid, less than 0. "
+                       "shape:%s, format:%s, dtype:%s, maybe has dynamic shape",
+                       output_mem_size,
+                       output_shape.ToString().c_str(),
+                       TypeUtils::FormatToSerialString(out_format).c_str(),
+                       TypeUtils::DataTypeToSerialString(data_type).c_str());
+    GELOGE(FAILED, "[Check][TensorSize]value:%ld invalid after calc, less than 0. shape:%s, format:%s, dtype:%s, "
+           "maybe has dynamic shape",
+           output_mem_size,
+           output_shape.ToString().c_str(),
+           TypeUtils::FormatToSerialString(out_format).c_str(),
+           TypeUtils::DataTypeToSerialString(data_type).c_str());
     return FAILED;
   }
 
@@ -203,7 +222,10 @@ ge::Status CalculateTensorRealSizeAndOutSize(const ge::ConstGeTensorDescPtr &out
 
 Status GraphMemoryAssigner::ReAssignMemory(bool is_loop_graph, map<int64_t, size_t> &mem_type_to_offset) {
   if (memory_offset_.empty()) {
-    GELOGE(FAILED, "memory_offset_ is empty.");
+    REPORT_INNER_ERROR("E19999", "InnerData memory_offset_ empty, not expected when ReAssignMemory, "
+                       "graph_id:%u, graph_name:%s", compute_graph_->GetGraphID(), compute_graph_->GetName().c_str());
+    GELOGE(FAILED, "[Check][InnerData:memory_offset_]empty is not expected, "
+           "graph_id:%u, graph_name:%s", compute_graph_->GetGraphID(), compute_graph_->GetName().c_str());
     return ge::FAILED;
   }
 
@@ -218,8 +240,10 @@ Status GraphMemoryAssigner::ReAssignMemory(bool is_loop_graph, map<int64_t, size
 
   auto session_id = compute_graph_->GetSessionID();
   if (total_mem_offset > VarManager::Instance(session_id)->GetGraphMemoryMaxSize()) {
-    GELOGE(ge::FAILED, "Current memoffset %zu is greater than memory manager malloc max size %zu", total_mem_offset,
-           VarManager::Instance(session_id)->GetGraphMemoryMaxSize());
+    GELOGE(ge::FAILED, "[Check][TotalMemOffset] %zu is greater than memory manager malloc max size %zu, "
+           "graph_id:%u, graph_name:%s, reduce your batchsize or scale your model may solve problem",
+           total_mem_offset, VarManager::Instance(session_id)->GetGraphMemoryMaxSize(),
+           compute_graph_->GetGraphID(), compute_graph_->GetName().c_str());
     for (auto iter : mem_type_to_offset) {
       ErrorManager::GetInstance().ATCReportErrMessage("E19022", {"memType", "size", "item", "maxsize"},
         {std::to_string(iter.first), std::to_string(iter.second), "featuremap",
@@ -234,7 +258,13 @@ Status GraphMemoryAssigner::ReAssignMemory(bool is_loop_graph, map<int64_t, size
 
 Status GraphMemoryAssigner::AssignZeroCopyMemory(map<int64_t, size_t> &mem_offset, size_t &zero_mem_copy_size) {
   BlockMemAssignerPtr priority_assigner = std::move(mem_assigner_->GetPriorityAssinger());
-  GE_IF_BOOL_EXEC(priority_assigner == nullptr, GELOGE(FAILED, "Get priority_assigner failed."); return ge::FAILED;);
+  if (priority_assigner == nullptr) {
+    REPORT_INNER_ERROR("E19999", "InnerData priority_assigner nullptr, not expected when AssignZeroCopyMemory, "
+                       "graph_id:%u, graph_name:%s", compute_graph_->GetGraphID(), compute_graph_->GetName().c_str());
+    GELOGE(FAILED, "[Check][InnerData:priority_assigner]nullptr is invalid, "
+           "graph_id:%u, graph_name:%s", compute_graph_->GetGraphID(), compute_graph_->GetName().c_str());
+    return ge::FAILED;
+  }
 
   size_t mem_offset_tmp = mem_offset[RT_MEMORY_HBM];
 
@@ -254,8 +284,11 @@ Status GraphMemoryAssigner::AssignZeroCopyMemory(map<int64_t, size_t> &mem_offse
   zero_mem_copy_size = mem_offset[RT_MEMORY_HBM] - mem_offset_tmp;
   auto iter = memory_offset_.find(RT_MEMORY_HBM);
   if (iter == memory_offset_.end()) {
-    std::string error = "Memory offset does not have memory type[HBM]";
-    GE_ERRORLOG_AND_ERRORMSG(FAILED, error.c_str());
+    REPORT_INNER_ERROR("E19999", "InnerData memory_offset_ does not have type[HBM], "
+                       "not expected when AssignZeroCopyMemory, "
+                       "graph_id:%u, graph_name:%s", compute_graph_->GetGraphID(), compute_graph_->GetName().c_str());
+    GELOGE(FAILED, "[Check][InnerData]memory_offset_ does not have memory type[HBM]"
+           "graph_id:%u, graph_name:%s", compute_graph_->GetGraphID(), compute_graph_->GetName().c_str());
     return FAILED;
   }
   iter->second.mem_offset_ = mem_offset[RT_MEMORY_HBM];
@@ -304,7 +337,7 @@ uint32_t GetContinuousMemoryType(const OpDescPtr &op_desc) {
   }
 
   if (continuous_type != 0) {
-    GELOGI("Current node %s continuous type %d.", op_desc->GetName().c_str(), continuous_type);
+    GELOGI("Current node %s continuous type %d", op_desc->GetName().c_str(), continuous_type);
   }
   return continuous_type;
 }
@@ -312,8 +345,9 @@ uint32_t GetContinuousMemoryType(const OpDescPtr &op_desc) {
 Status GetMemorySize(const OpDescPtr &op_desc, const ge::ConstGeTensorDescPtr &output_desc, uint32_t continuous_type,
                      int64_t &tensor_size, int64_t &nopadding_size) {
   if ((op_desc == nullptr) || (output_desc == nullptr)) {
-    GELOGE(FAILED, "Input para is nullptr.");
-    return FAILED;
+    REPORT_INNER_ERROR("E19999", "InnerData param op_desc or output_desc is nullptr, "
+                       "not expected when GetMemorySize");
+    GELOGE(FAILED, "[Check][Param]op_desc or output_desc is nullptr");
   }
   tensor_size = 0;
   nopadding_size = 0;
@@ -322,7 +356,10 @@ Status GetMemorySize(const OpDescPtr &op_desc, const ge::ConstGeTensorDescPtr &o
     int64_t attr_dim_index;
     bool get_attr_dim_flag = ge::AttrUtils::GetInt(op_desc, ATTR_NAME_REUSE_INPUT_ON_DIM_INDEX, attr_dim_index);
     if (!get_attr_dim_flag) {
-      GELOGE(FAILED, "Get attr _reuse_input_on_dim_index failed.");
+      REPORT_INNER_ERROR("E19999", "Get Attr:%s failed when GetMemorySize, op_name:%s",
+                         ATTR_NAME_REUSE_INPUT_ON_DIM_INDEX.c_str(), op_desc->GetName().c_str());
+      GELOGE(FAILED, "[Get][Attr:%s]fail for op_name:%s",
+             ATTR_NAME_REUSE_INPUT_ON_DIM_INDEX.c_str(), op_desc->GetName().c_str());
       return FAILED;
     }
 
@@ -330,17 +367,25 @@ Status GetMemorySize(const OpDescPtr &op_desc, const ge::ConstGeTensorDescPtr &o
     int64_t batch_dim_num = 1;
     if (CalculateTensorRealSizeAndOutSize(output_desc, attr_dim_index, nopadding_size, batch_dim_num, tensor_size) !=
         SUCCESS) {
-      GELOGE(FAILED, "CalculateTensorRealSizeAndOutSize failed for node %s.", op_desc->GetName().c_str());
+      REPORT_CALL_ERROR("E19999", "CalculateTensorRealSizeAndOutSize failed, attr_dim_index:%ld, op_name:%s",
+                        attr_dim_index, op_desc->GetName().c_str());
+      GELOGE(FAILED, "[Calculate][NopaddingSize]failed for node %s, attr_dim_index:%ld",
+             op_desc->GetName().c_str(), attr_dim_index);
       return FAILED;
     }
   } else {
     if (ge::TensorUtils::GetSize(*output_desc, tensor_size) != ge::SUCCESS) {
-      GELOGE(FAILED, "GetSize failed.");
+      REPORT_INNER_ERROR("E19999", "Get Tensor Size failed, op_name:%s", op_desc->GetName().c_str());
+      GELOGE(FAILED, "[Get][TensorSize]failed in padding case, op_name:%s", op_desc->GetName().c_str());
       return FAILED;
     }
   }
   if ((tensor_size < 0) || (nopadding_size < 0)) {
-    GELOGE(FAILED, "GetMemorySize for node %s failed.", op_desc->GetName().c_str());
+    REPORT_INNER_ERROR("E19999", "GetMemorySize fail, "
+                       "tensor_size:%ld or nopadding_size:%ld less than 0, invalid, op_name:%s",
+                       tensor_size, nopadding_size, op_desc->GetName().c_str());
+    GELOGE(FAILED, "[Get][MemorySize]tensor_size:%ld or nopadding_size:%ld less than 0, invalid, op_name:%s",
+           tensor_size, nopadding_size, op_desc->GetName().c_str());
     return FAILED;
   }
   return SUCCESS;
@@ -374,7 +419,7 @@ bool IsContinuousInputConflict(const ge::NodePtr &node, const OpDescPtr &peer_op
   // If GetBool fail, is_peer_reference is false.
   (void) AttrUtils::GetBool(peer_op_desc, ATTR_NAME_REFERENCE, is_peer_reference);
   GE_IF_BOOL_EXEC(is_peer_reference,
-                  std::string warning = "Current op" + FmtToStr(node->GetOpDesc()->GetName()) +
+                  std::string warning = "[Check][Continuous]Current op" + FmtToStr(node->GetOpDesc()->GetName()) +
                       " requires continuous input, while the previous op" + FmtToStr(peer_op_desc->GetName()) +
                       " is ref. There may be conflict between the two.";
                   GELOGW("%s", warning.c_str());
@@ -404,7 +449,7 @@ Status GraphMemoryAssigner::ReAssignContinuousMemory(bool is_loop_graph) {
     if (continuous_input) {
       if (AssignContinuousInputMemoryWithAtomicProcessDirectly(node, node_2_continuous_type)) {
         GE_CHK_STATUS_RET(AssignContinuousInputMemoryWithAtomicProcess(node, continuous_type),
-                          "Assign node %s continuous input memory failed.", node->GetName().c_str())
+                          "[Assign][Memory:Continuous:Input]fail for node:%s", node->GetName().c_str())
       } else {
         nodes_stack.push_back(node);
       }
@@ -413,10 +458,11 @@ Status GraphMemoryAssigner::ReAssignContinuousMemory(bool is_loop_graph) {
     int64_t memory_type = RT_MEMORY_HBM;
     bool continuous_output = ((continuous_type & kTypeOutput) != 0) || ((continuous_type & kTypeOutputNoPadding) != 0);
     if (continuous_output) {
-      GE_CHK_STATUS_RET(GetNodeMemoryType(node, memory_type, "output"), "Get node memory type failed.");
+      GE_CHK_STATUS_RET(GetNodeMemoryType(node, memory_type, "output"),
+                        "[Get][MemType]fail for node:%s", node->GetName().c_str());
       ret = AssignContinuousOutputMemory(node, memory_type, continuous_type);
       if (ret != ge::SUCCESS) {
-        GELOGE(ret, "Assign continuous output memory failed!");
+        GELOGE(ret, "[Assign][Memory:Continuous:Ouput]fail for node:%s", node->GetName().c_str());
         return ret;
       }
     }
@@ -427,14 +473,16 @@ Status GraphMemoryAssigner::ReAssignContinuousMemory(bool is_loop_graph) {
     nodes_stack.pop_back();
     auto iter = node_2_continuous_type.find(node);
     if (iter == node_2_continuous_type.end()) {
-      GELOGE(FAILED, "node %s has no continuous type!", node->GetName().c_str());
+      REPORT_INNER_ERROR("E19999", "Inner data error when process continuous memory alloc for node:%s, "
+                         "but has no continuous type", node->GetName().c_str());
+      GELOGE(FAILED, "[Get][ContinuousType] find fail for node:%s", node->GetName().c_str());
       return FAILED;
     }
     GE_CHK_STATUS_RET(AssignContinuousInputMemoryWithAtomicProcess(node, iter->second, true),
-                      "Assign node %s continuous input memory failed.", node->GetName().c_str())
+                      "[Assign][Memory:Continuous:Input]fail for node:%s.", node->GetName().c_str())
   }
   for (auto pair : memory_offset_) {
-    GELOGD("After reassign continuous memory, memory type = %ld, mem_offset = %zu.", pair.first,
+    GELOGD("After reassign continuous memory, memory type = %ld, mem offset = %zu.", pair.first,
            pair.second.mem_offset_);
   }
   return ge::SUCCESS;
@@ -442,11 +490,13 @@ Status GraphMemoryAssigner::ReAssignContinuousMemory(bool is_loop_graph) {
 
 Status GraphMemoryAssigner::AssignContinuousInputMemory(const ge::NodePtr &node, int64_t &continuous_mem_start,
     int64_t &continuous_mem_size, int64_t memory_type, uint32_t continuous_type, bool reverse_refresh) {
-  GELOGI("Current node %s needs continuous input.", node->GetName().c_str());
+  GELOGI("Current node %s needs continuous input", node->GetName().c_str());
   auto iter = memory_offset_.find(memory_type);
   if (iter == memory_offset_.end()) {
-    std::string error = "Memory offset does not have memory type" + FmtToStr(memory_type);
-    GE_ERRORLOG_AND_ERRORMSG(FAILED, error.c_str());
+    REPORT_INNER_ERROR("E19999", "find memory offset fail for mem_type:%ld, "
+                       "when assign continuous input memory for node:%s, ", memory_type, node->GetName().c_str());
+    GELOGE(FAILED, "[Find][MemOffset]fail for mem_type:%ld, when AssignContinuousInputMemory for node:%s",
+           memory_type, node->GetName().c_str());
     return FAILED;
   }
   // The head and tail of hcom continuous input should be added 512
@@ -459,8 +509,9 @@ Status GraphMemoryAssigner::AssignContinuousInputMemory(const ge::NodePtr &node,
   GE_CHECK_NOTNULL(op_desc);
   vector<int64_t> output_list_this = op_desc->GetOutputOffset();
   if (output_list_this.empty()) {
-    std::string error = "node:" + FmtToStr(op_desc->GetName()) + "has no output offset";
-    GE_ERRORLOG_AND_ERRORMSG(FAILED, error.c_str());
+    REPORT_INNER_ERROR("E19999", "No output offset in node :%s, not expected when assign continuous input memory",
+                       node->GetName().c_str());
+    GELOGE(FAILED, "[Get][OutputOffset] empty is invalid, node:%s", node->GetName().c_str());
     return FAILED;
   }
   (void) ge::AttrUtils::GetBool(op_desc, ATTR_NAME_CONTINUOUS_INPUT_ALLOC, is_continuous_input_allocated);
@@ -480,8 +531,9 @@ Status GraphMemoryAssigner::AssignContinuousInputMemory(const ge::NodePtr &node,
     lx_fusion = lx_fusion && !offsets_of_fusion.empty();
     if (lx_fusion) {
       if (peer_out_data_anchor->GetIdx() >= static_cast<int>(offsets_of_fusion.size())) {
-        std::string error = "fusion: peer node" + FmtToStr(peer_op_desc->GetName()) +
-            " index" + FmtToStr(peer_out_data_anchor->GetIdx()) + " is out of range.";
+        std::string error = "fusion: peer node:" + FmtToStr(peer_op_desc->GetName()) +
+            " anchor_index:" + FmtToStr(peer_out_data_anchor->GetIdx()) +
+            " is out of range:" + FmtToStr(offsets_of_fusion.size());
         GE_ERRORLOG_AND_ERRORMSG(FAILED, error.c_str());
         return FAILED;
       }
@@ -497,7 +549,9 @@ Status GraphMemoryAssigner::AssignContinuousInputMemory(const ge::NodePtr &node,
     bool is_nopadding = ((continuous_type & kTypeInputNoPadding) != 0) || lx_fusion;
     vector<int64_t> output_list = peer_op_desc->GetOutputOffset();
     if (peer_out_data_anchor->GetIdx() >= static_cast<int>(output_list.size())) {
-      std::string error = "index" + FmtToStr(peer_out_data_anchor->GetIdx()) + " is out of range.";
+      std::string error = "peer node:" + FmtToStr(peer_op_desc->GetName()) +
+          " anchor_index:" + FmtToStr(peer_out_data_anchor->GetIdx()) +
+          " is out of range:" + FmtToStr(output_list.size());
       GE_ERRORLOG_AND_ERRORMSG(FAILED, error.c_str());
       return FAILED;
     }
@@ -506,13 +560,13 @@ Status GraphMemoryAssigner::AssignContinuousInputMemory(const ge::NodePtr &node,
     bool is_allocated_first_input = is_continuous_input_allocated && (in_data_anchor->GetIdx() == 0);
     if (is_allocated_first_input) {
       std::map<int32_t, int32_t> out2ins;
-      GE_CHK_STATUS_RET(GetAllRef(node, out2ins), "Node: %s get all ref failed", node->GetName().c_str());
+      GE_CHK_STATUS_RET(GetAllRef(node, out2ins), "[Get][AllRef]fail for node: %s", node->GetName().c_str());
       // output is beginning offset, set offset for input; only support this case now
       if ((out2ins.size() == 1) && (out2ins.begin()->second == 0) && (reverse_refresh)) {
         auto peer_output_offset = output_list.at(peer_out_data_anchor->GetIdx());
         output_list.at(peer_out_data_anchor->GetIdx()) = output_list_this.at(out2ins.begin()->first);
         peer_op_desc->SetOutputOffset(output_list);
-        GELOGI("Node %s out %d ref in %d input node %s, use output offset %ld update %ld.", node->GetName().c_str(),
+        GELOGI("Node %s out %d ref in %d input node %s, use output offset %ld update %ld", node->GetName().c_str(),
                out2ins.begin()->first, out2ins.begin()->second, peer_op_desc->GetName().c_str(),
                output_list_this.at(out2ins.begin()->first), peer_output_offset);
       } else {
@@ -542,7 +596,7 @@ Status GraphMemoryAssigner::AssignContinuousInputMemory(const ge::NodePtr &node,
     }
 
     GELOGI("[IMAS]Continuous input : Set %s name[%s] optype[%s] output[%d] offset to [%zu] stream_id[%ld] memtype[%ld] "
-        "size[%zu] realsize[%ld] nopadding size[%d].", node->GetOwnerComputeGraph()->GetName().c_str(),
+        "size[%zu] realsize[%ld] nopadding size[%d]", node->GetOwnerComputeGraph()->GetName().c_str(),
         peer_op_desc->GetName().c_str(), node->GetType().c_str(), peer_out_data_anchor->GetIdx(),
         output_list.at(peer_out_data_anchor->GetIdx()), peer_op_desc->GetStreamId(), memory_type,
         is_continuous_input_allocated ? 0UL : align_size, real_size, is_nopadding);
@@ -563,17 +617,32 @@ Status GraphMemoryAssigner::AssignContinuousInputMemory(const ge::NodePtr &node,
 Status GetFirstInputPeerOutOutputOffset(const ge::NodePtr &node, int64_t &mem_offset) {
   auto in_data_anchor_list = node->GetAllInDataAnchors();
   if (in_data_anchor_list.empty()) {
-    GELOGE(FAILED, "Node %s's in data anchor is empty.", node->GetName().c_str());
+    REPORT_INNER_ERROR("E19999", "InAnchor list empty in node:%s, not expect when GetFirstInputPeerOutOutputOffset",
+                       node->GetName().c_str());
+    GELOGE(FAILED, "[Get][InAnchor]empty is invalid, node:%s", node->GetName().c_str());
     return FAILED;
   }
   auto peer_out_data_anchor = in_data_anchor_list.at(0)->GetPeerOutAnchor();
-  GE_IF_BOOL_EXEC(peer_out_data_anchor == nullptr, GELOGE(ge::FAILED, "peer_out_data_anchor is null.");
+  GE_IF_BOOL_EXEC(peer_out_data_anchor == nullptr,
+                  REPORT_INNER_ERROR("E19999", "PeerAcnhor is null, "
+                                     "not expect when GetFirstInputPeerOutOutputOffset for node:%s",
+                                     node->GetName().c_str());
+                  GELOGE(ge::FAILED, "[Check][PeerAnchor]null is invalid, node:%s", node->GetName().c_str());
                   return ge::FAILED);
   auto peer_op_desc = peer_out_data_anchor->GetOwnerNode()->GetOpDesc();
-  GE_IF_BOOL_EXEC(peer_op_desc == nullptr, GELOGE(ge::FAILED, "peer_op_desc is null."); return ge::FAILED);
+  GE_IF_BOOL_EXEC(peer_op_desc == nullptr,
+                  REPORT_INNER_ERROR("E19999", "PeerOpDesc is null, "
+                                     "not expect when GetFirstInputPeerOutOutputOffset for node:%s",
+                                     node->GetName().c_str());
+                  GELOGE(ge::FAILED, "[Check][PeerOpDesc]null is invalid, node:%s",  node->GetName().c_str());
+                  return ge::FAILED);
   vector<int64_t> in_node_output_offsets = peer_op_desc->GetOutputOffset();
   if (peer_out_data_anchor->GetIdx() >= static_cast<int>(in_node_output_offsets.size())) {
-    GELOGE(FAILED, "Index : %d is out of range.", peer_out_data_anchor->GetIdx());
+    REPORT_INNER_ERROR("E19999", "PeerAnchorIndex:%d bigger than in_offset size:%lu, "
+                       "judge invalid when GetFirstInputPeerOutOutputOffset for node:%s",
+                       peer_out_data_anchor->GetIdx(), in_node_output_offsets.size(), node->GetName().c_str());
+    GELOGE(FAILED, "[Check][Index:PeerOutDataAnchor]PeerIndex:%d bigger than in_offset size:%lu, node:%s",
+           peer_out_data_anchor->GetIdx(), in_node_output_offsets.size(), node->GetName().c_str());
     return FAILED;
   }
   mem_offset = in_node_output_offsets.at(peer_out_data_anchor->GetIdx());
@@ -584,11 +653,18 @@ Status GraphMemoryAssigner::AssignContinuousOutputMemory(const ge::NodePtr &node
                                                          uint32_t continuous_type) {
   GELOGI("Current node %s needs continuous output.", node->GetName().c_str());
   auto out_op_desc = node->GetOpDesc();
-  GE_IF_BOOL_EXEC(out_op_desc == nullptr, GELOGE(ge::FAILED, "out_op_desc is null."); return ge::FAILED);
+  GE_IF_BOOL_EXEC(out_op_desc == nullptr,
+                  REPORT_INNER_ERROR("E19999", "OpDesc is null, "
+                                     "not expect when AssignContinuousOutputMemory for node:%s",
+                                     node->GetName().c_str());
+                  GELOGE(ge::FAILED, "[Check][OpDesc]null is invalid, node:%s",  node->GetName().c_str()));
   vector<int64_t> output_list = out_op_desc->GetOutputOffset();
   if ((out_op_desc->GetOutputsSize() > output_list.size()) || (output_list.size() == 0)) {
-    GELOGE(ge::FAILED, "The size %zu of node output desc is more than output_list's size %zu.",
-           out_op_desc->GetOutputsSize(), output_list.size());
+    REPORT_INNER_ERROR("E19999", "Output size:%zu more than output offset size:%zu, invalid in node:%s, "
+                       "when AssignContinuousOutputMemory",
+                       out_op_desc->GetOutputsSize(), output_list.size(), node->GetName().c_str());
+    GELOGE(ge::FAILED, "[Check][InnerData]Output size:%zu more than output offset size:%zu, invalid in node:%s",
+           out_op_desc->GetOutputsSize(), output_list.size(), node->GetName().c_str());
     return ge::FAILED;
   }
 
@@ -647,14 +723,18 @@ Status GraphMemoryAssigner::ReAssignAtomicMemory(bool is_loop_graph) {
   map<string, vector<NodePtr>> connecting_output_atomic_nodes;
   Status status = FilterAtomicNodesForMemoryAssign(normal_atomic_and_clean_nodes_map, connecting_output_atomic_nodes);
   if (status != SUCCESS) {
-    GELOGE(status, "Failed to filter atomic nodes for memory assignment.");
+    GELOGE(status, "[Filter][AtomicNode]failed in graph_id:%u, graph_name:%s",
+           compute_graph_->GetGraphID(), compute_graph_->GetName().c_str());
     return status;
   }
 
   auto mem_iter = memory_offset_.find(RT_MEMORY_HBM);
   if (mem_iter == memory_offset_.end()) {
-    std::string error = "Memory offset does not have memory type" + FmtToStr(RT_MEMORY_HBM);
-    GE_ERRORLOG_AND_ERRORMSG(FAILED, error.c_str());
+    REPORT_INNER_ERROR("E19999", "InnerData memory_offset_ does not have type[HBM], "
+                       "not expected when ReAssignAtomicMemory, "
+                       "graph_id:%u, graph_name:%s", compute_graph_->GetGraphID(), compute_graph_->GetName().c_str());
+    GELOGE(FAILED, "[Check][InnerData]memory_offset_ does not have memory type[HBM]"
+           "graph_id:%u, graph_name:%s", compute_graph_->GetGraphID(), compute_graph_->GetName().c_str());
     return FAILED;
   }
 
@@ -670,7 +750,7 @@ Status GraphMemoryAssigner::ReAssignAtomicMemory(bool is_loop_graph) {
         vector<int64_t> mem_offset_end;
         status = AssignAtomicOutputAndWorkspaceMemory(atomic_node, mem_offset_end);
         if (status != SUCCESS) {
-          GELOGE(status, "Assign atomic output and workspace memory failed, node name is %s.",
+          GELOGE(status, "[Assign][Memory]output atomic mem and workspace mem, fail for node name is %s.",
                  atomic_node->GetName().c_str());
           return status;
         }
@@ -679,7 +759,7 @@ Status GraphMemoryAssigner::ReAssignAtomicMemory(bool is_loop_graph) {
       int64_t atomic_mem_size = static_cast<int64_t>(mem_iter->second.mem_offset_) - atomic_mem_start;
       if (atomic_mem_size != 0) {
         GE_CHK_STATUS_RET(SetAtomicCleanAttr(iter.first, {atomic_mem_start}, {atomic_mem_size}, RT_MEMORY_HBM),
-                          "Failed to set attr for atomic addr clean node %s.", iter.first->GetName().c_str());
+                          "[Set][Attr]fail for atomic addr clean node %s.", iter.first->GetName().c_str());
       }
     }
     batch_max_mem_offset = std::max(batch_max_mem_offset, static_cast<int64_t>(mem_iter->second.mem_offset_));
@@ -690,7 +770,8 @@ Status GraphMemoryAssigner::ReAssignAtomicMemory(bool is_loop_graph) {
   for (auto &iter_batch : connecting_output_atomic_nodes) {
     mem_iter->second.mem_offset_ = batch_atomic_mem_start;
     if (AssignConnectNetOutputAtomicMemory(iter_batch.second) != SUCCESS) {
-      GELOGE(FAILED, "Failed to assign memory of nodes that connect to netoutput.");
+      GELOGE(FAILED, "[Assign][Memory]for nodes that connect to netoutput failed."
+             "graph_id:%u, graph_name:%s", compute_graph_->GetGraphID(), compute_graph_->GetName().c_str());
       return FAILED;
     }
     batch_max_mem_offset = std::max(batch_max_mem_offset, static_cast<int64_t>(mem_iter->second.mem_offset_));
@@ -721,9 +802,10 @@ Status GraphMemoryAssigner::FilterAtomicNodesForMemoryAssign(
               // If GetBool fail, is_reference is false.
               (void) ge::AttrUtils::GetBool(peer_in_node_desc, ATTR_NAME_REFERENCE, is_reference);
               if (is_reference) {
-                std::string error = "Op" + FmtToStr(peer_in_node_desc->GetName()) +
-                    " cannot have both atomic and is_reference attribute.";
-                GE_ERRORLOG_AND_ERRORMSG(FAILED, error.c_str());
+                REPORT_INNER_ERROR("E19999", "Op:%s cannot have both atomic and is_reference attribute, "
+                                   "not support now", peer_in_node_desc->GetName().c_str());
+                GELOGE(FAILED, "[Check][Attr]Op:%s cannot have both atomic and is_reference attribute, "
+                               "not support now", peer_in_node_desc->GetName().c_str());
                 return ge::PARAM_INVALID;
               }
 
@@ -761,7 +843,7 @@ Status GraphMemoryAssigner::AssignAtomicOutputAndWorkspaceMemory(const ge::NodeP
   // Assign atomic node output memory
   Status ret = AssignAtomicOutputMemory(node, mem_offset_end);
   if (ret != SUCCESS) {
-    GELOGE(ret, "Failed to assign atomic output memory, node is %s.", node_op_desc->GetName().c_str());
+    GELOGE(ret, "[Assign][Memory:Ouput:Atomic]Failed for node:%s.", node_op_desc->GetName().c_str());
     return ret;
   }
 
@@ -781,7 +863,7 @@ Status GraphMemoryAssigner::AssignAtomicOutputAndWorkspaceMemory(const ge::NodeP
       ret = AssignOrdinaryAtomicWorkspaceMemory(node_op_desc, atomic_workspace_info, mem_offset_end);
     }
     if (ret != SUCCESS) {
-      GELOGE(ret, "Assign atomic workspace memory failed, node is %s.", node_op_desc->GetName().c_str());
+      GELOGE(ret, "[Assign][Memory:Atomic:Workspace]fail for node:%s.", node_op_desc->GetName().c_str());
       return ret;
     }
   } else {
@@ -794,8 +876,11 @@ Status GraphMemoryAssigner::AssignAtomicOutputAndWorkspaceMemory(const ge::NodeP
 Status GraphMemoryAssigner::AssignConnectNetOutputAtomicMemory(vector<NodePtr> &connect_netoutput_nodes) {
   auto iter = memory_offset_.find(RT_MEMORY_HBM);
   if (iter == memory_offset_.end()) {
-    std::string error = "Memory offset does not have memory type" + FmtToStr(RT_MEMORY_HBM);
-    GE_ERRORLOG_AND_ERRORMSG(FAILED, error.c_str());
+    REPORT_INNER_ERROR("E19999", "InnerData memory_offset_ does not have type[HBM], "
+                       "not expected when AssignConnectNetOutputAtomicMemory, "
+                       "graph_id:%u, graph_name:%s", compute_graph_->GetGraphID(), compute_graph_->GetName().c_str());
+    GELOGE(FAILED, "[Check][InnerData]memory_offset_ does not have memory type[HBM]"
+           "graph_id:%u, graph_name:%s", compute_graph_->GetGraphID(), compute_graph_->GetName().c_str());
     return FAILED;
   }
   for (auto &node : connect_netoutput_nodes) {
@@ -811,13 +896,14 @@ Status GraphMemoryAssigner::AssignConnectNetOutputAtomicMemory(vector<NodePtr> &
            node->GetName().c_str(), node->GetOpDesc()->GetType().c_str(), original_atomic_mem_start);
     vector<int64_t> mem_offset_end;
     if (AssignAtomicOutputAndWorkspaceMemory(node, mem_offset_end) != SUCCESS) {
-      GELOGE(FAILED, "Assign atomic output and workspace memory failed, node is %s.", node->GetName().c_str());
+      GELOGE(FAILED, "[Assign][Memory]output atomic mem and workspace mem, fail for node name is %s.",
+             node->GetName().c_str());
       return FAILED;
     }
 
     // All atomic nodes use atomic_addr_clean op independently, so we need to set the attr separately.
     if (SetIndependentAtomicAttr(node, original_atomic_mem_start, mem_offset_end, RT_MEMORY_HBM) != SUCCESS) {
-      GELOGE(FAILED, "Failed to set atomic attr separately.");
+      GELOGE(FAILED, "[Set][Attr:IndependentAtomic]fail for node:%s", node->GetName().c_str());
       return FAILED;
     }
   }
@@ -842,8 +928,11 @@ Status GraphMemoryAssigner::AssignReferenceMemory() {
     vector<int64_t> output_list = out_op_desc->GetOutputOffset();
 
     if (out_op_desc->GetOutputsSize() > output_list.size()) {
-      GELOGE(ge::FAILED, "The size %zu of node output desc is more than output_list's size %zu.",
-             out_op_desc->GetOutputsSize(), output_list.size());
+      REPORT_INNER_ERROR("E19999", "Output size:%zu more than output offset size:%zu, judge invalid in node:%s "
+                         "when AssignReferenceMemory",
+                         out_op_desc->GetOutputsSize(), output_list.size(), node->GetName().c_str());
+      GELOGE(ge::FAILED, "[Check][InnerData]Output size:%zu more than output offset size:%zu, invalid in node:%s",
+             out_op_desc->GetOutputsSize(), output_list.size(), node->GetName().c_str());
       return ge::FAILED;
     }
 
@@ -896,9 +985,12 @@ bool GraphMemoryAssigner::CheckInputIsSupportAtomic(const ge::NodePtr &node) {
     }
     if ((peer_op_desc->GetType() == CONSTANTOP) || (peer_op_desc->GetType() == AIPP_DATA_TYPE) ||
         (peer_op_desc->GetType() == VARIABLE)) {
-      std::string error = "Op" + FmtToStr(node->GetName()) + "'s peer out node" +
-          FmtToStr(peer_op_desc->GetName()) + " is invalid, Constant/AippData/Variable is not supported";
-      GE_ERRORLOG_AND_ERRORMSG(FAILED, error.c_str());
+      REPORT_INNER_ERROR("E19999", "node(type:%s, name:%s) link to atomic node(name:%s), "
+                         "this situation not supported now",
+                         peer_op_desc->GetType().c_str(), peer_op_desc->GetName().c_str(), node->GetName().c_str());
+      GELOGE(ge::FAILED, "[Check][Link]node(type:%s, name:%s) link to atomic node(name:%s), "
+             "this situation not supported now",
+             peer_op_desc->GetType().c_str(), peer_op_desc->GetName().c_str(), node->GetName().c_str());
       return false;
     }
   }
@@ -918,22 +1010,27 @@ Status GraphMemoryAssigner::AssignAtomicOutputMemory(const ge::NodePtr &node, ve
   // Check atomic output
   vector<int64_t> output_list = op_desc->GetOutputOffset();
   if (atomic_output_index.size() > output_list.size()) {
-    std::string error = "Op" + FmtToStr(node->GetName()) +
-        "'s size of atomic_output_index is more than the size of output_list";
+    std::string error =
+        "Op:" + FmtToStr(node->GetName()) + "'s size:" + FmtToStr(atomic_output_index.size()) +
+        " of atomic_output_index is more than the size:" + FmtToStr(output_list.size()) + " of output_list";
     GE_ERRORLOG_AND_ERRORMSG(FAILED, error.c_str());
     return ge::FAILED;
   }
   auto output_list_size = static_cast<int64_t>(output_list.size());
   auto iter = memory_offset_.find(RT_MEMORY_HBM);
   if (iter == memory_offset_.end()) {
-    std::string error = "Memory offset does not have memory type" + FmtToStr(RT_MEMORY_HBM);
-    GE_ERRORLOG_AND_ERRORMSG(FAILED, error.c_str());
+    REPORT_INNER_ERROR("E19999", "InnerData memory_offset_ does not have type[HBM], "
+                       "not expected when AssignAtomicOutputMemory, "
+                       "graph_id:%u, graph_name:%s", compute_graph_->GetGraphID(), compute_graph_->GetName().c_str());
+    GELOGE(FAILED, "[Check][InnerData]memory_offset_ does not have memory type[HBM]"
+           "graph_id:%u, graph_name:%s", compute_graph_->GetGraphID(), compute_graph_->GetName().c_str());
     return FAILED;
   }
   for (auto &output_index : atomic_output_index) {
     if (output_index >= output_list_size) {
-      std::string error = "Op" + FmtToStr(node->GetName()) + "'s output index" + FmtToStr(output_index) +
-          " is more than the size" + FmtToStr(output_list_size) + " of output_list.";
+      std::string error =
+          "Op:" + FmtToStr(node->GetName()) + "'s atomic_output index:" + FmtToStr(output_index) +
+          " is more than the size:" + FmtToStr(output_list_size) + " of output_list.";
       GE_ERRORLOG_AND_ERRORMSG(ge::PARAM_INVALID, error.c_str());
       return ge::PARAM_INVALID;
     }
@@ -941,7 +1038,8 @@ Status GraphMemoryAssigner::AssignAtomicOutputMemory(const ge::NodePtr &node, ve
     // If the input of the cascade op needs to clear the atomic addr, there is no need to clear it separately here
     bool is_assigned_mem = false;
     if (GetMemoryAssignmentStatus(node, output_index, is_assigned_mem) != SUCCESS) {
-      GELOGE(ge::FAILED, "Failed to get memory assignment of node %s.", node->GetName().c_str());
+      GELOGE(ge::FAILED, "[Get][MemoryAssignmentStatus]fail for node %s, out_index:%ld",
+             node->GetName().c_str(), output_index);
       return ge::FAILED;
     }
 
@@ -981,8 +1079,9 @@ Status GraphMemoryAssigner::AssignAtomicOutputMemory(const ge::NodePtr &node, ve
 Status GraphMemoryAssigner::GetMemoryAssignmentStatus(const ge::NodePtr &node, int64_t output_index,
                                                       bool &is_mem_assigned) {
   if (static_cast<size_t>(output_index) >= node->GetAllOutDataAnchors().size()) {
-    std::string error = "Op" + FmtToStr(node->GetName()) + "'s output index" + FmtToStr(output_index) +
-        " is more than the size of node's AllOutDataAnchors.";
+    std::string error =
+        "Op:" + FmtToStr(node->GetName()) + "'s output index:" + FmtToStr(output_index) +
+        " is more than the size:" + FmtToStr(node->GetAllOutDataAnchors().size()) + " of node's AllOutDataAnchors.";
     GE_ERRORLOG_AND_ERRORMSG(ge::PARAM_INVALID, error.c_str());
     return ge::PARAM_INVALID;
   }
@@ -1010,8 +1109,11 @@ Status GraphMemoryAssigner::AssignOrdinaryAtomicWorkspaceMemory(const ge::OpDesc
   GELOGI("Begin to reassign normal atomic memory, node = %s.", op_desc->GetName().c_str());
   auto mem_type_iter = memory_offset_.find(RT_MEMORY_HBM);
   if (mem_type_iter == memory_offset_.end()) {
-    std::string error = "Memory offset does not have memory type" + FmtToStr(RT_MEMORY_HBM);
-    GE_ERRORLOG_AND_ERRORMSG(FAILED, error.c_str());
+    REPORT_INNER_ERROR("E19999", "InnerData memory_offset_ does not have type[HBM], "
+                       "not expected when AssignOrdinaryAtomicWorkspaceMemory, "
+                       "graph_id:%u, graph_name:%s", compute_graph_->GetGraphID(), compute_graph_->GetName().c_str());
+    GELOGE(FAILED, "[Check][InnerData]memory_offset_ does not have memory type[HBM]"
+           "graph_id:%u, graph_name:%s", compute_graph_->GetGraphID(), compute_graph_->GetName().c_str());
     return FAILED;
   }
   vector<int64_t> workspace_vector = op_desc->GetWorkspace();
@@ -1032,8 +1134,9 @@ Status GraphMemoryAssigner::AssignOrdinaryAtomicWorkspaceMemory(const ge::OpDesc
       auto workspace_index = static_cast<uint64_t>(info_iter.first);
       auto workspace_size = info_iter.second;
       if (workspace_index >= workspace_vector.size()) {
-        std::string error = "The workspace index" + FmtToStr(workspace_index) +
-            " is more than the size" + FmtToStr(workspace_vector.size()) + " of workspace vector.";
+        std::string error = "The workspace index:" + FmtToStr(workspace_index) +
+            " is more than the size:" + FmtToStr(workspace_vector.size()) + " of workspace vector in op:" +
+            op_desc->GetName().c_str();
         GE_ERRORLOG_AND_ERRORMSG(ge::PARAM_INVALID, error.c_str());
         return ge::PARAM_INVALID;
       }
@@ -1063,8 +1166,11 @@ Status GraphMemoryAssigner::AssignFusionAtomicWorkspaceMemory(const ge::OpDescPt
   GELOGI("Begin to reassign fusion atomic memory, node = %s.", op_desc->GetName().c_str());
   auto mem_type_iter = memory_offset_.find(RT_MEMORY_HBM);
   if (mem_type_iter == memory_offset_.end()) {
-    std::string error = "Memory offset does not have memory type" + FmtToStr(RT_MEMORY_HBM);
-    GE_ERRORLOG_AND_ERRORMSG(FAILED, error.c_str());
+    REPORT_INNER_ERROR("E19999", "InnerData memory_offset_ does not have type[HBM], "
+                       "not expected when AssignFusionAtomicWorkspaceMemory, "
+                       "graph_id:%u, graph_name:%s", compute_graph_->GetGraphID(), compute_graph_->GetName().c_str());
+    GELOGE(FAILED, "[Check][InnerData]memory_offset_ does not have memory type[HBM]"
+           "graph_id:%u, graph_name:%s", compute_graph_->GetGraphID(), compute_graph_->GetName().c_str());
     return FAILED;
   }
   map<string, map<int64_t, int64_t>> sub_node_workspace_offset;
@@ -1095,7 +1201,10 @@ Status GraphMemoryAssigner::AssignFusionAtomicWorkspaceMemory(const ge::OpDescPt
     sub_node_workspace_offset.insert(std::make_pair(iter.first, index_offset));
   }
   if (!(op_desc->SetExtAttr(EXT_ATTR_ATOMIC_WORKSPACE_OFFSET, sub_node_workspace_offset))) {
-    GELOGE(FAILED, "Set EXT_ATTR_ATOMIC_WORKSPACE_OFFSET failed, op name:%s.", op_desc->GetName().c_str());
+    REPORT_INNER_ERROR("E19999", "Set Attr:%s fail for node:%s when AssignFusionAtomicWorkspaceMemory",
+                       EXT_ATTR_ATOMIC_WORKSPACE_OFFSET.c_str(), op_desc->GetName().c_str());
+    GELOGE(FAILED, "[Set][Attr:%s]fail for node:%s.",
+           EXT_ATTR_ATOMIC_WORKSPACE_OFFSET.c_str(), op_desc->GetName().c_str());
     return FAILED;
   }
 
@@ -1106,7 +1215,7 @@ Status GraphMemoryAssigner::CheckOffset() {
   std::map<std::string, std::string> anchor_to_symbol;
   std::map<std::string, std::list<NodeIndexIO>> symbol_to_anchors;
   if (GraphUtils::GetRefMapping(compute_graph_, symbol_to_anchors, anchor_to_symbol) != GRAPH_SUCCESS) {
-    GELOGE(FAILED, "Get ref-mapping for graph %s failed.", compute_graph_->GetName().c_str());
+    GELOGE(FAILED, "[Get][RefMapping]fail for graph %s", compute_graph_->GetName().c_str());
     return FAILED;
   }
   for (const ge::NodePtr &node : compute_graph_->GetAllNodes()) {
@@ -1148,7 +1257,6 @@ Status GraphMemoryAssigner::CheckOffset() {
         std::string error = "Invalid workspace" + FmtToStr(ge::kInvalidOffset) +
             + " in node" + FmtToStr(node->GetName());
         GE_ERRORLOG_AND_ERRORMSG(FAILED, error.c_str());
-        GELOGE(FAILED, "Invalid workspace in node: %s workspace: %ld.", node->GetName().c_str(), ge::kInvalidOffset);
         return FAILED;
       }
     }
@@ -1158,8 +1266,10 @@ Status GraphMemoryAssigner::CheckOffset() {
 
 ge::Status GraphMemoryAssigner::SetInputOffset() {
   if (memory_offset_.empty()) {
-    GELOGE(FAILED, "memory_offset_ is empty.");
-    return FAILED;
+    REPORT_INNER_ERROR("E19999", "InnerData memory_offset_ empty, not expected when SetInputOffset, "
+                       "graph_id:%u, graph_name:%s", compute_graph_->GetGraphID(), compute_graph_->GetName().c_str());
+    GELOGE(FAILED, "[Check][InnerData:memory_offset_]empty is not expected, "
+           "graph_id:%u, graph_name:%s", compute_graph_->GetGraphID(), compute_graph_->GetName().c_str());
   }
   for (auto pair : memory_offset_) {
     GEEVENT("[IMAS]AfterAssignMemory : %s memoffset[%zu], memtype[%ld]", compute_graph_->GetName().c_str(),
@@ -1168,7 +1278,7 @@ ge::Status GraphMemoryAssigner::SetInputOffset() {
 
   for (const ge::NodePtr &node : compute_graph_->GetAllNodes()) {
     if (UpdateOpInputOffset(node) != ge::SUCCESS) {
-      GELOGE(ge::FAILED, "Update op input offset failed");
+      GELOGE(ge::FAILED, "[Update][Offset:Input]fail for op:%s", node->GetName().c_str());
       return ge::FAILED;
     }
   }
@@ -1316,12 +1426,12 @@ ge::Status GraphMemoryAssigner::UpdateOpInputOffset(const NodePtr &node) const {
     }
   } else if (node->GetType() == DATA_TYPE) {
     if (UpdateConstArgsOffset(node, input_list) != SUCCESS) {
-      GELOGE(FAILED, "Update data: %s args offset failed.", node->GetName().c_str());
+      GELOGE(FAILED, "[Update][Offset:Input:Const]fail for node:%s ", node->GetName().c_str());
       return FAILED;
     }
   } else {
     if (UpdateOpInputOffset(node, input_list) != SUCCESS) {
-      GELOGE(FAILED, "Update node: %s input offset failed.", node->GetName().c_str());
+      GELOGE(FAILED, "[Update][Offset:Input]fail for node:%s", node->GetName().c_str());
       return FAILED;
     }
   }
@@ -1361,7 +1471,7 @@ Status GraphMemoryAssigner::SetIndependentAtomicAttr(const ge::NodePtr &node, in
              peer_out_node_desc->GetName().c_str(), peer_out_node_desc->GetType().c_str());
       if (peer_out_node_desc->GetType() == ATOMICADDRCLEAN) {
         if (SetAtomicCleanAttr(peer_out_node, memory_offset_start, memory_offset_size, memory_type) != SUCCESS) {
-          GELOGE(FAILED, "Set atomic clean attr failed.");
+          GELOGE(FAILED, "[Set][AtomicCleanAttr]fail for node:%s", peer_out_node->GetName().c_str());
           return FAILED;
         }
       }
@@ -1387,7 +1497,10 @@ ge::Status GraphMemoryAssigner::SetAtomicCleanAttr(const NodePtr &node, const ve
     (void) ge::AttrUtils::GetListInt(node_op_desc, ATTR_NAME_AUTOMIC_ADD_START, mem_start_vector);
     mem_start_vector.insert(mem_start_vector.end(), atomic_mem_start.begin(), atomic_mem_start.end());
     GE_CHK_BOOL_EXEC(ge::AttrUtils::SetListInt(node_op_desc, ATTR_NAME_AUTOMIC_ADD_START, mem_start_vector),
-                     GELOGE(FAILED, "SetListInt failed.");
+                     REPORT_INNER_ERROR("E19999", "Set Attr:%s failed when SetAtomicCleanAttr, op_name:%s",
+                                        ATTR_NAME_AUTOMIC_ADD_START.c_str(), node_op_desc->GetName().c_str());
+                     GELOGE(FAILED, "[Set][Attr:%s]fail for op_name:%s",
+                            ATTR_NAME_AUTOMIC_ADD_START.c_str(), node_op_desc->GetName().c_str());
                      return FAILED);
 
     std::vector<int64_t> mem_size_vector;
@@ -1395,7 +1508,10 @@ ge::Status GraphMemoryAssigner::SetAtomicCleanAttr(const NodePtr &node, const ve
     (void) ge::AttrUtils::GetListInt(node_op_desc, ATTR_NAME_AUTOMIC_ADD_MEM_SIZE, mem_size_vector);
     mem_size_vector.insert(mem_size_vector.end(), atomic_mem_size.begin(), atomic_mem_size.end());
     GE_CHK_BOOL_EXEC(ge::AttrUtils::SetListInt(node_op_desc, ATTR_NAME_AUTOMIC_ADD_MEM_SIZE, mem_size_vector),
-                     GELOGE(FAILED, "SetListInt failed.");
+                     REPORT_INNER_ERROR("E19999", "Set Attr:%s failed when SetAtomicCleanAttr, op_name:%s",
+                                        ATTR_NAME_AUTOMIC_ADD_MEM_SIZE.c_str(), node_op_desc->GetName().c_str());
+                     GELOGE(FAILED, "[Set][Attr:%s]fail for op_name:%s",
+                            ATTR_NAME_AUTOMIC_ADD_MEM_SIZE.c_str(), node_op_desc->GetName().c_str());
                      return FAILED);
 
     std::stringstream ss;
@@ -1437,12 +1553,14 @@ ge::Status GraphMemoryAssigner::GetNodeListMemoryType(const vector<NodePtr> &nod
   // In the dynamic batch scenario, the memory attributes of nodes are the same.
   for (auto &n : nodes) {
     if (mem_reuse_model == kVirtualInputNodeMemoryReuse) {
-      GE_CHK_STATUS_RET(GetNodeMemoryType(n, memory_type, "input"), "Get node memory type failed.")
+      GE_CHK_STATUS_RET(GetNodeMemoryType(n, memory_type, "input"),
+                        "[Get][MemType:input]fail for node:%s", n->GetName().c_str())
       break;
     }
 
     if (mem_reuse_model == kVirtualOutputNodeMemoryReuse) {
-      GE_CHK_STATUS_RET(GetNodeMemoryType(n, memory_type, "output"), "Get node memory type failed.");
+      GE_CHK_STATUS_RET(GetNodeMemoryType(n, memory_type, "output"),
+                        "[Get][MemType:output]fail for node:%s", n->GetName().c_str())
       break;
     }
   }
@@ -1478,7 +1596,7 @@ ge::Status GraphMemoryAssigner::GetNodeMemoryType(const NodePtr &node, int64_t &
   }
 
   if (!CheckContinuousMemType(mem_type_list)) {
-    GELOGE(FAILED, "Check continuous memory type failed.");
+    GELOGE(FAILED, "[Check][MemType:Continuous]fail for node:%s", node->GetName().c_str());
     return FAILED;
   }
   // It is continuous memory and memory type is the same, so use the first memory.
@@ -1526,7 +1644,11 @@ ge::Status GraphMemoryAssigner::GetAllRef(const NodePtr &node, map<int32_t, int3
       if (node->GetInDataAnchor(reuse_in_index) != nullptr) {
         out2ins.emplace(out_data_anchor->GetIdx(), reuse_in_index);
       } else {
-        GELOGE(FAILED, "Invalid reuse_input value %d on output %d of node %s, please check attr reuse_input",
+        REPORT_INNER_ERROR("E19999", "Invalid reuse_input value %d on output %d of node %s, "
+                           "please check attr reuse_input",
+                           reuse_in_index, out_data_anchor->GetIdx(), node->GetName().c_str());
+        GELOGE(FAILED, "[Check][Attr]Invalid reuse_input value %d on output %d of node %s, "
+               "please check attr reuse_input",
                reuse_in_index, out_data_anchor->GetIdx(), node->GetName().c_str());
         return FAILED;
       }
@@ -1549,7 +1671,7 @@ bool GraphMemoryAssigner::AssignContinuousInputMemoryWithAtomicProcessDirectly(
     auto continuous_type = iter->second;
     bool continuous_input = ((continuous_type & kTypeInput) != 0) || ((continuous_type & kTypeInputNoPadding) != 0);
     if (continuous_input) {
-      GELOGI("Node %s 's precursor node %s need assign continuous input memory, store node firstly.",
+      GELOGI("Node %s 's precursor node %s need assign continuous input memory, store node firstly",
              input_continuous_node->GetName().c_str(), in_node->GetName().c_str());
       return false;
     }
@@ -1559,7 +1681,7 @@ bool GraphMemoryAssigner::AssignContinuousInputMemoryWithAtomicProcessDirectly(
     node_2_continuous_type.emplace(out_node, continuous_type);
     bool continuous_input = ((continuous_type & kTypeInput) != 0) || ((continuous_type & kTypeInputNoPadding) != 0);
     if (continuous_input) {
-      GELOGI("Node %s 's succeed node %s need assign continuous input memory, store node firstly.",
+      GELOGI("Node %s 's succeed node %s need assign continuous input memory, store node firstly",
              input_continuous_node->GetName().c_str(), out_node->GetName().c_str());
       return false;
     }
@@ -1575,11 +1697,12 @@ ge::Status GraphMemoryAssigner::AssignContinuousInputMemoryWithAtomicProcess(con
   int64_t mem_clean_size = 0;
   int64_t memory_type = RT_MEMORY_HBM;
 
-  GE_CHK_STATUS_RET(GetNodeMemoryType(input_continuous_node, memory_type, "input"), "Get node memory type failed.");
+  GE_CHK_STATUS_RET(GetNodeMemoryType(input_continuous_node, memory_type, "input"),
+                    "[Get][MemType]fail for node:%s", input_continuous_node->GetName().c_str());
   auto ret = AssignContinuousInputMemory(input_continuous_node, mem_clean_start, mem_clean_size, memory_type,
                                          continuous_type, reverse_refresh);
   if (ret != ge::SUCCESS) {
-    GELOGE(ret, "Assign continuous input memory failed!");
+    GELOGE(ret, "[Assign][Memory:Input:continuous]fail for node:%s", input_continuous_node->GetName().c_str());
     return ret;
   }
 
@@ -1590,7 +1713,6 @@ ge::Status GraphMemoryAssigner::AssignContinuousInputMemoryWithAtomicProcess(con
   if (!input_indexes.empty() && input_indexes[0] == kAllInputAddrIsAtomic) {
     // check whether there is an atomic conflict between the current node and the peer out node
     if (!CheckInputIsSupportAtomic(input_continuous_node)) {
-      GELOGE(ge::FAILED, "There is an atomic conflict between the current node and the peer out node, not supported!");
       return ge::FAILED;
     }
 
@@ -1602,7 +1724,7 @@ ge::Status GraphMemoryAssigner::AssignContinuousInputMemoryWithAtomicProcess(con
       if (peer_out_node->GetType() == ATOMICADDRCLEAN) {
         ret = SetAtomicCleanAttr(peer_out_node, {mem_clean_start}, {mem_clean_size}, memory_type);
         if (ret != SUCCESS) {
-          GELOGE(ret, "Failed to set attr for atomic addr clean node %s.", peer_out_node->GetName().c_str());
+          GELOGE(ret, "[Set][AtomicCleanAttr]fail for node:%s", peer_out_node->GetName().c_str());
           return ret;
         }
       }

ge/graph/build/model_builder.cc

@@ -574,6 +574,50 @@ Status ModelBuilder::MergeWeights() {
   return SUCCESS;
 }
 
+Status ModelBuilder::SaveAtomicTBEKernel(const OpDescPtr &op_desc) {
+  ge::NodePtr atomic_clean_node = nullptr;
+  atomic_clean_node = op_desc->TryGetExtAttr("atomic_clean_node_ptr", atomic_clean_node);
+  if (atomic_clean_node == nullptr) {
+    return SUCCESS;
+  }
+
+  ge::OpDescPtr atomic_op_desc = atomic_clean_node->GetOpDesc();
+  GE_CHECK_NOTNULL(atomic_op_desc);
+  TBEKernelPtr tbe_kernel = atomic_op_desc->TryGetExtAttr(ge::OP_EXTATTR_NAME_TBE_KERNEL, TBEKernelPtr());
+  if (tbe_kernel == nullptr) {
+    std::string kernel_name;
+    GeAttrValue::BYTES kernel_buffer;
+    (void) AttrUtils::GetStr(atomic_op_desc, ATTR_NAME_TBE_KERNEL_NAME, kernel_name);
+    (void) AttrUtils::GetBytes(atomic_op_desc, ATTR_NAME_TBE_KERNEL_BUFFER, kernel_buffer);
+    if (!kernel_name.empty() && (kernel_buffer.GetSize() > 0)) {
+      GE_CHECK_NOTNULL(kernel_buffer.GetData());
+      std::vector<char> data(kernel_buffer.GetData(), kernel_buffer.GetData() + kernel_buffer.GetSize());
+      tbe_kernel = MakeShared<OpKernelBin>(kernel_name, std::move(data));
+      GE_CHECK_NOTNULL(tbe_kernel);
+    }
+  }
+  if (tbe_kernel == nullptr) {
+    GELOGD("Atomic_clean_node doesn't have tbe_kernel.");
+    return SUCCESS;
+  }
+  tbe_kernel_store_.AddTBEKernel(tbe_kernel);
+  GELOGD("Atomic_clean_node tbe_kernel_name %s!", tbe_kernel->GetName().c_str());
+  (void) AttrUtils::SetStr(op_desc, ATOMIC_ATTR_TBE_KERNEL_NAME, tbe_kernel->GetName());
+
+  std::string kernel_name;
+  (void) AttrUtils::GetStr(atomic_op_desc, atomic_op_desc->GetName() + "_kernelname", kernel_name);
+  (void) AttrUtils::SetStr(op_desc, op_desc->GetName() + "_atomic_kernelname", kernel_name);
+
+  std::string meta_data;
+  (void) AttrUtils::GetStr(atomic_op_desc, TVM_ATTR_NAME_METADATA, meta_data);
+  (void) AttrUtils::SetStr(op_desc, ATOMIC_ATTR_TVM_METADATA, meta_data);
+
+  std::string json_string;
+  (void) AttrUtils::GetStr(atomic_op_desc, TVM_ATTR_NAME_MAGIC, json_string);
+  (void) AttrUtils::SetStr(op_desc, ATOMIC_ATTR_TVM_MAGIC, json_string);
+  return SUCCESS;
+}
+
 Status ModelBuilder::SaveDataToModel(ge::Model &model, ge::GeModel &ge_model) {
   // Add weight
   ge_model.SetWeight(weight_buffer_);
@@ -607,6 +651,8 @@ Status ModelBuilder::SaveDataToModel(ge::Model &model, ge::GeModel &ge_model) {
     }
     tbe_name_set.insert(tbe_kernel->GetName());
     tbe_kernel_store_.AddTBEKernel(tbe_kernel);
+
+    GE_CHK_STATUS_RET(SaveAtomicTBEKernel(node_op_desc), "[Save][TBEKernel] save atomic tbekernel failed!");
   }
 
   SetModelCheckAicpuAttr(model, aicpu_op_types, aicpu_tf_op_types);

ge/graph/build/model_builder.h

@@ -89,6 +89,8 @@ class ModelBuilder {
   void SetModelCheckAicpuAttr(ge::Model &model, std::set<std::string> &aicpu_op_types,
                                 std::set<std::string> &aicpu_tf_op_types);
 
+  Status SaveAtomicTBEKernel(const OpDescPtr &op_desc);
+
   uint64_t session_id_;
 
   map<int64_t, size_t> mem_type_to_mem_offset_;

ge/graph/build/task_generator.cc

@@ -49,6 +49,7 @@ const char *const kIsLastNode = "is_last_node";
 const char *const kIsInputVar = "INPUT_IS_VAR";
 const char *const kIsOutputVar = "OUTPUT_IS_VAR";
 const char *const kProfilingMode = "PROFILING_MODE";
+const char *const kIteratorV2 = "IteratorV2";
 const uint32_t kProfilingArStep = 2;
 const uint64_t kProfilingFpStartLogid = 1;
 const uint64_t kProfilingBpEndLogid = 2;
@@ -57,6 +58,7 @@ const uint64_t kProfilingArEndLogid = 4;
 const uint64_t kProfilingIterEndLogid = 65535;
 const int64_t kHashFactor = 100000;
 const int64_t kInvalidGroupId = -1;
+const std::set<std::string> kFpNodeTypes = {ge::DATA, ge::GETNEXT, kIteratorV2};
 }  // namespace
 namespace ge {
 TaskGenerator::TaskGenerator(uint8_t *var_mem_base, uint64_t var_mem_size) {
@@ -621,8 +623,10 @@ Status TaskGenerator::AutoFindFpOpIndex(const ComputeGraphPtr &graph, ProfilingP
     if (op_kernel_lib_name.empty()) {
       continue;
     }
-
-    if (op_desc->GetType() == GETNEXT || op_desc->GetType() == DATA) {
+    auto type = op_desc->GetType();
+    std::string original_type;
+    (void)AttrUtils::GetStr(op_desc, ATTR_NAME_FRAMEWORK_ORIGINAL_TYPE, original_type);
+    if (kFpNodeTypes.find(type) != kFpNodeTypes.end() || kFpNodeTypes.find(original_type) != kFpNodeTypes.end()) {
       auto out_anchor = node->GetOutDataAnchor(0);
       for (auto &peer_in_anchor : out_anchor->GetPeerInDataAnchors()) {
         GE_CHECK_NOTNULL(peer_in_anchor);

ge/graph/execute/graph_execute.cc

@@ -20,9 +20,12 @@
 #include <string>
 
 #include "graph/load/model_manager/model_manager.h"
+#include "graph/load/model_manager/davinci_model.h"
 #include "omm/csa_interact.h"
 
 namespace ge {
+using Uint32Pair = pair<uint32_t, uint32_t>;
+const uint32_t kInvalidModelId = UINT32_MAX;
 GraphExecutor::GraphExecutor()
     : init_flag_(false),
       train_graph_flag_(false),
@@ -358,7 +361,8 @@ Status GraphExecutor::ExecuteGraph(GraphId graph_id, const GeRootModelPtr &ge_ro
 }
 
 Status GraphExecutor::ExecuteGraphAsync(GraphId graph_id, const GeRootModelPtr &ge_root_model,
-                                        const std::vector<InputTensorInfo> &input_tensor) {
+                                        const std::vector<InputTensorInfo> &input_tensor,
+                                        const RunAsyncCallback& callback) {
   GELOGI("[GraphExecutor] Start to async execute graph, graph_id=%u", graph_id);
   if (graph_id != last_graph_id_) {
     auto ret = FreeExecuteMemory();
@@ -368,7 +372,7 @@ Status GraphExecutor::ExecuteGraphAsync(GraphId graph_id, const GeRootModelPtr &
   }
   last_graph_id_ = graph_id;
   GE_CHECK_NOTNULL_EXEC(ge_root_model, return FAILED);
-  Status ret = AsyncExecuteModel(ge_root_model->GetModelId(), input_tensor);
+  Status ret = AsyncExecuteModel(ge_root_model, input_tensor, callback);
   if (ret != SUCCESS) {
     GELOGE(GE_GRAPH_SYNC_MODEL_FAILED, "[GraphExecutor] AsyncExecuteModel Error!");
     return GE_GRAPH_SYNC_MODEL_FAILED;
@@ -378,11 +382,81 @@ Status GraphExecutor::ExecuteGraphAsync(GraphId graph_id, const GeRootModelPtr &
   return SUCCESS;
 }
 
-Status GraphExecutor::AsyncExecuteModel(uint32_t model_id, const std::vector<InputTensorInfo> &inputs) {
+bool CompareByLoad(const Uint32Pair &lhs, const Uint32Pair &rhs) {
+  return lhs.second < rhs.second;
+}
+
+uint32_t GraphExecutor::GetExecuteModelId(const GeRootModelPtr &ge_root_model) {
+  std::vector<uint32_t> model_ids = ge_root_model->GetAllModelId();
+  if (model_ids.empty()) {
+    return kInvalidModelId;
+  }
+  if (model_ids.size() == 1) {
+    return ge_root_model->GetModelId();
+  }
+  std::vector<Uint32Pair> model_id_to_loads;
+  auto model_manager = ModelManager::GetInstance();
+  GE_CHECK_NOTNULL(model_manager);
+  for (auto model_id : model_ids) {
+    auto davinci_model = model_manager->GetModel(model_id);
+    auto hybrid_model = model_manager->GetHybridModel(model_id);
+    if (hybrid_model == nullptr) {
+      GE_CHECK_NOTNULL(davinci_model);
+    }
+    uint32_t input_load = hybrid_model != nullptr ? hybrid_model->GetDataInputerSize() :
+                                                    davinci_model->GetDataInputerSize();
+    uint32_t running_load = hybrid_model != nullptr ? static_cast<uint32_t>(hybrid_model->GetRunningFlag()) :
+                                                      static_cast<uint32_t>(davinci_model->GetRunningFlag());
+    uint32_t load = input_load + running_load;
+    if (load == 0) {
+      return model_id;
+    }
+    model_id_to_loads.emplace_back(model_id, load);
+  }
+  sort(model_id_to_loads.begin(), model_id_to_loads.end(), CompareByLoad);
+  if (model_id_to_loads.empty()) {
+    return kInvalidModelId;
+  }
+  return model_id_to_loads.begin()->first;
+}
+
+Status GraphExecutor::SetCallback(uint32_t model_id, const GeRootModelPtr &ge_root_model,
+                                  const RunAsyncCallback &callback) {
+  auto model_manager = ge::ModelManager::GetInstance();
+  GE_CHECK_NOTNULL(model_manager);
+  if (model_manager->IsNeedHybridLoad(*ge_root_model)) {
+    auto model = model_manager->GetHybridModel(model_id);
+    GE_CHECK_NOTNULL(model);
+    if (model->SetRunAsyncListenerCallback(callback) != SUCCESS) {
+      GELOGE(FAILED, "SetRunAsyncListenerCallback failed.");
+      return FAILED;
+    }
+  } else {
+    auto model = model_manager->GetModel(model_id);
+    GE_CHECK_NOTNULL(model);
+    if (model->SetRunAsyncListenerCallback(callback) != SUCCESS) {
+      GELOGE(FAILED, "SetRunAsyncListenerCallback failed.");
+      return FAILED;
+    }
+  }
+  return SUCCESS;
+}
+
+Status GraphExecutor::AsyncExecuteModel(const GeRootModelPtr &ge_root_model, const std::vector<InputTensorInfo> &inputs,
+                                        const RunAsyncCallback &callback) {
+  uint32_t model_id = GetExecuteModelId(ge_root_model);
+  if (model_id == kInvalidModelId) {
+    GELOGE(INTERNAL_ERROR, "No valid model id.");
+    return INTERNAL_ERROR;
+  }
   try {
     auto model_manager = ge::ModelManager::GetInstance();
     GE_CHECK_NOTNULL(model_manager);
     GELOGI("RunAsync begin.model_id %u", model_id);
+    if (SetCallback(model_id, ge_root_model, callback) != SUCCESS) {
+      GELOGE(FAILED, "RunAsync: SetCallBack for model fail");
+      return FAILED;
+    }
 
     Status ret = model_manager->DataInputTensor(model_id, inputs);
     if (ret != SUCCESS) {

ge/graph/execute/graph_execute.h

@@ -50,7 +50,7 @@ class GraphExecutor {
                       std::vector<GeTensor> &output_tensor);
 
   ge::Status ExecuteGraphAsync(GraphId graph_id, const GeRootModelPtr &ge_root_model,
-                               const std::vector<InputTensorInfo> &input_tensor);
+                               const std::vector<InputTensorInfo> &input_tensor, const RunAsyncCallback &callback);
 
   Status SetCondition(std::mutex *mutex, std::condition_variable *cond, std::shared_ptr<GraphModelListener> listener);
 
@@ -116,6 +116,8 @@ class GraphExecutor {
 
   static Status GetOpDescInfo(uint32_t device_id, uint32_t stream_id, uint32_t task_id, OpDescInfo &op_desc_info);
 
+  uint32_t GetExecuteModelId(const GeRootModelPtr &ge_root_model);
+
  private:
   Status PrepareInputData(const std::vector<GeTensor> &input_tensor, InputData &graph_input_data,
                           OutputData &graph_output_data, std::vector<InputOutputDescInfo> &output_desc);
@@ -123,7 +125,8 @@ class GraphExecutor {
   Status SyncExecuteModel(uint32_t model_id, const std::vector<GeTensor> &input_tensor,
                           std::vector<GeTensor> &output_tensor);
 
-  Status AsyncExecuteModel(uint32_t model_id, const std::vector<InputTensorInfo> &input_tensor);
+  Status AsyncExecuteModel(const GeRootModelPtr &ge_root_model, const std::vector<InputTensorInfo> &input_tensor,
+                           const RunAsyncCallback &callback);
 
   void InitModelIdInfo(std::vector<uint32_t> &out_model_id_info, std::vector<SubGraphInfoPtr> &sub_graph_vec,
                        uint32_t output_size);
@@ -132,6 +135,9 @@ class GraphExecutor {
 
   Status MallocInOutBuffer(const std::vector<uint64_t> &buffer_size, std::vector<void *> &data_addr);
 
+  static Status SetCallback(uint32_t model_id, const GeRootModelPtr &ge_root_model,
+                            const RunAsyncCallback &callback);
+
   bool init_flag_;
 
   bool train_graph_flag_;

ge/graph/load/graph_loader.cc

@@ -60,7 +60,6 @@ Status GraphLoader::LoadModelOnline(uint32_t &model_id, const std::shared_ptr<ge
     GELOGE(GE_GRAPH_PARAM_NULLPTR, "[LoadGraph] GE load graph model_ptr is nullptr.");
     return GE_GRAPH_PARAM_NULLPTR;
   }
-  model_id = ge_root_model_ptr->GetModelId();
 
   auto model_manager = ModelManager::GetInstance();
   GE_CHECK_NOTNULL(model_manager);

ge/graph/load/model_manager/data_dumper.cc

@@ -385,7 +385,7 @@ Status DataDumper::DumpRefOutput(const DataDumper::InnerDumpInfo &inner_dump_inf
 
 Status DataDumper::DumpOutputWithTask(const InnerDumpInfo &inner_dump_info, aicpu::dump::Task &task) {
   const auto &output_descs = inner_dump_info.op->GetAllOutputsDesc();
-  const std::vector<void *> output_addrs = ModelUtils::GetOutputDataAddrs(runtime_param_, inner_dump_info.op);
+  const std::vector<void *> output_addrs = ModelUtils::GetOutputDataAddrs(*runtime_param_, inner_dump_info.op);
   if (output_descs.size() != output_addrs.size()) {
     GELOGE(PARAM_INVALID, "Invalid output desc addrs size %zu, op %s has %zu output desc.", output_addrs.size(),
            inner_dump_info.op->GetName().c_str(), output_descs.size());
@@ -436,7 +436,7 @@ Status DataDumper::DumpOutput(const InnerDumpInfo &inner_dump_info, aicpu::dump:
   // else data, const or variable op
   aicpu::dump::Output output;
   auto output_tensor = inner_dump_info.op->GetOutputDescPtr(inner_dump_info.output_anchor_index);
-  const std::vector<void *> output_addrs = ModelUtils::GetOutputDataAddrs(runtime_param_, inner_dump_info.op);
+  const std::vector<void *> output_addrs = ModelUtils::GetOutputDataAddrs(*runtime_param_, inner_dump_info.op);
   if (output_tensor == nullptr) {
     GELOGE(PARAM_INVALID, "output_tensor is null, index: %d, size: %zu.", inner_dump_info.output_anchor_index,
            inner_dump_info.op->GetOutputsSize());
@@ -540,7 +540,7 @@ Status DataDumper::DumpRefInput(const DataDumper::InnerDumpInfo &inner_dump_info
 Status DataDumper::DumpInput(const InnerDumpInfo &inner_dump_info, aicpu::dump::Task &task) {
   GELOGI("Start dump input");
   const auto &input_descs = inner_dump_info.op->GetAllInputsDesc();
-  const std::vector<void *> input_addrs = ModelUtils::GetInputDataAddrs(runtime_param_, inner_dump_info.op);
+  const std::vector<void *> input_addrs = ModelUtils::GetInputDataAddrs(*runtime_param_, inner_dump_info.op);
   if (input_descs.size() != input_addrs.size()) {
     GELOGE(PARAM_INVALID, "Invalid input desc addrs size %zu, op %s has %zu input desc.", input_addrs.size(),
            inner_dump_info.op->GetName().c_str(), input_descs.size());

ge/graph/load/model_manager/data_dumper.h

@@ -36,9 +36,21 @@
 namespace ge {
 class DataDumper {
  public:
-  DataDumper() : runtime_param_{} {}
-
-  explicit DataDumper(const RuntimeParam &rsh) : runtime_param_(rsh) {}
+  explicit DataDumper(RuntimeParam *rsh)
+      : model_name_(),
+        model_id_(0),
+        runtime_param_(rsh),
+        dev_mem_load_(nullptr),
+        dev_mem_unload_(nullptr),
+        op_list_(),
+        input_map_(),
+        load_flag_(false),
+        device_id_(0),
+        global_step_(0),
+        loop_per_iter_(0),
+        loop_cond_(0),
+        compute_graph_(nullptr),
+        ref_info_() {}
 
   ~DataDumper();
 
@@ -93,10 +105,10 @@ class DataDumper {
   // for inference data dump
   std::string om_name_;
 
-  uint32_t model_id_ = 0;
-  const RuntimeParam &runtime_param_;
-  void *dev_mem_load_ = nullptr;
-  void *dev_mem_unload_ = nullptr;
+  uint32_t model_id_;
+  RuntimeParam *runtime_param_;
+  void *dev_mem_load_;
+  void *dev_mem_unload_;
 
   struct InnerDumpInfo;
   struct InnerInputMapping;
@@ -107,12 +119,12 @@ class DataDumper {
   uint32_t end_graph_stream_id_ = 0;
   bool is_end_graph_ = false;
   std::multimap<std::string, InnerInputMapping> input_map_;  // release after DavinciModel::Init
-  bool load_flag_ = false;
-  uint32_t device_id_ = 0;
-  uintptr_t global_step_ = 0;
-  uintptr_t loop_per_iter_ = 0;
-  uintptr_t loop_cond_ = 0;
-  ComputeGraphPtr compute_graph_ = nullptr;  // release after DavinciModel::Init
+  bool load_flag_;
+  uint32_t device_id_;
+  uintptr_t global_step_;
+  uintptr_t loop_per_iter_;
+  uintptr_t loop_cond_;
+  ComputeGraphPtr compute_graph_;  // release after DavinciModel::Init
   std::map<OpDescPtr, void *> ref_info_;     // release after DavinciModel::Init
   void *l1_fusion_addr_ = nullptr;
 

ge/graph/load/model_manager/data_inputer.h

@@ -134,6 +134,8 @@ class DataInputer {
   ///
   void Stop() { queue_.Stop(); }
 
+  uint32_t Size() { return queue_.Size(); }
+
  private:
   ///
   /// @ingroup domi_ome

ge/graph/load/model_manager/davinci_model.cc

@@ -31,6 +31,7 @@
 #include "common/scope_guard.h"
 #include "common/thread_pool.h"
 #include "framework/common/debug/ge_log.h"
+#include "framework/common/util.h"
 #include "graph/common/ge_call_wrapper.h"
 #include "graph/compute_graph.h"
 #include "graph/debug/ge_attr_define.h"
@@ -184,7 +185,7 @@ DavinciModel::DavinciModel(int32_t priority, const std::shared_ptr<ModelListener
       last_execute_mode_(INITIALIZATION),
       session_id_(0),
       device_id_(0),
-      maxDumpOpNum_(0), data_dumper_(runtime_param_),
+      maxDumpOpNum_(0), data_dumper_(&runtime_param_),
       iterator_count_(0),
       is_l1_fusion_enable_(false),
       is_first_execute_(true) {
@@ -297,6 +298,11 @@ void DavinciModel::ReleaseTask() {
       GE_CHK_STATUS(task->Release(), "Release task failed.");
     }
   }
+
+  for (auto &item : label_goto_args_) {
+    GE_FREE_RT_LOG(item.second.first);
+  }
+  label_goto_args_.clear();
 }
 
 Status DavinciModel::Assign(const GeModelPtr &ge_model) {
@@ -654,12 +660,12 @@ Status DavinciModel::Init(void *dev_ptr, size_t mem_size, void *weight_ptr, size
   runtime_param_.graph_id = compute_graph->GetGraphID();
 
   // op debug register
-  GE_CHK_STATUS_RET(OpDebugRegister(), "OpDebugRegister failed.");
+  GE_CHK_STATUS_RET(OpDebugRegister(), "OpDebugRegister failed");
 
   GE_TIMESTAMP_START(TransAllVarData);
-  GE_CHK_STATUS_RET(TransAllVarData(compute_graph, runtime_param_.graph_id), "TransAllVarData failed.");
+  GE_CHK_STATUS_RET(TransAllVarData(compute_graph, runtime_param_.graph_id), "TransAllVarData failed");
   GE_TIMESTAMP_END(TransAllVarData, "GraphLoader::TransAllVarData");
-  GE_CHK_STATUS_RET(TransVarDataUtils::CopyVarData(compute_graph, session_id_, device_id_), "copy var data failed.");
+  GE_CHK_STATUS_RET(TransVarDataUtils::CopyVarData(compute_graph, session_id_, device_id_), "copy var data failed");
 
   GE_TIMESTAMP_START(InitModelMem);
   GELOGD("Known node is %d.", known_node_);
@@ -667,7 +673,7 @@ Status DavinciModel::Init(void *dev_ptr, size_t mem_size, void *weight_ptr, size
   if (!known_node_) {
     GE_CHK_STATUS_RET_NOLOG(InitFeatureMapAndP2PMem(dev_ptr, mem_size));
     data_inputer_ = new (std::nothrow) DataInputer();
-    GE_CHK_BOOL_RET_STATUS(data_inputer_ != nullptr, MEMALLOC_FAILED, "data_inputer_ is nullptr.");
+    GE_CHK_BOOL_RET_STATUS(data_inputer_ != nullptr, MEMALLOC_FAILED, "data_inputer_ is nullptr");
   }
   fixed_mem_base_ = reinterpret_cast<uintptr_t>(mem_base_);
   GE_TIMESTAMP_END(InitModelMem, "GraphLoader::InitModelMem");
@@ -1334,6 +1340,39 @@ void DavinciModel::ParseDynamicOutShape(const std::vector<std::string> &str_info
   }
 }
 
+Status DavinciModel::GetLabelGotoAddr(uint32_t label_index, rtMemType_t mem_type, void *&arg_addr, uint32_t &arg_size) {
+  std::lock_guard<std::mutex> lock(label_args_mutex_);
+  auto it = label_goto_args_.find(label_index);
+  if (it != label_goto_args_.end()) {
+    arg_addr = it->second.first;
+    arg_size = it->second.second;
+    return SUCCESS;
+  }
+
+  if (label_index >= label_list_.size()) {
+    GELOGE(INTERNAL_ERROR, "Invalid label id:%u, label size:%zu", label_index, label_list_.size());
+    return INTERNAL_ERROR;
+  }
+  GE_CHECK_NOTNULL(label_list_[label_index]);
+  vector<rtLabel_t> label_used = { label_list_[label_index] };
+
+  arg_size = label_used.size() * sizeof(rtLabelDevInfo);
+  rtError_t rt_ret = rtMalloc(&arg_addr, arg_size, mem_type);
+  if (rt_ret != RT_ERROR_NONE) {
+    GELOGE(RT_FAILED, "Call rtMalloc failed, error: %#x", rt_ret);
+    return RT_ERROR_TO_GE_STATUS(rt_ret);
+  }
+
+  label_goto_args_[label_index] = { arg_addr, arg_size };
+  rt_ret = rtLabelListCpy(label_used.data(), label_used.size(), arg_addr, arg_size);
+  if (rt_ret != RT_ERROR_NONE) {
+    GELOGE(RT_FAILED, "Call rtLabelListCpy failed, error: %#x", rt_ret);
+    return RT_ERROR_TO_GE_STATUS(rt_ret);
+  }
+
+  return SUCCESS;
+}
+
 /// @ingroup ge
 /// @brief LabelSet Op Initialize.
 /// @param [in] op_desc: LabelSet Op descriptor.
@@ -2547,6 +2586,8 @@ void *DavinciModel::Run(DavinciModel *model) {
 
   ErrorManager::GetInstance().SetStage(ErrorMessage::kModelExecute, ErrorMessage::kModelExecute);
   while (model->RunFlag()) {
+    // Model hasn't truly started runing before received data
+    model->SetRunningFlag(false);
     bool rslt_flg = true;
     if (model->GetDataInputer() == nullptr) {
       GELOGW("Data inputer is nullptr.");
@@ -2556,6 +2597,8 @@ void *DavinciModel::Run(DavinciModel *model) {
 
     std::shared_ptr<InputDataWrapper> data_wrapper;
     Status ret = model->GetDataInputer()->Pop(data_wrapper);
+    // Model run indeedly start after received data.
+    model->SetRunningFlag(true);
     if (data_wrapper == nullptr || ret != SUCCESS) {
       GELOGI("data_wrapper is null!");
       continue;
@@ -2642,7 +2685,9 @@ void *DavinciModel::Run(DavinciModel *model) {
 
     model->iterator_count_++;
     model->is_first_execute_ = false;
-    GELOGI("run iterator count is %lu", model->iterator_count_);
+    // model run finished
+    model->SetRunningFlag(false);
+    GELOGI("run iterator count is %lu, model_id:%u", model->iterator_count_, model->model_id_);
   }
 
   CsaInteract::GetInstance().WriteInternalErrorCode();
@@ -2700,7 +2745,7 @@ Status DavinciModel::ModelRunStart() {
 
   error_context_ = ErrorManager::GetInstance().GetErrorContext();
   CREATE_STD_THREAD(thread_id_, DavinciModel::Run, this);
-  GELOGI("model tread create success, model id:%u.", model_id_);
+  GELOGI("model thread create success, model id:%u.", model_id_);
   return SUCCESS;
 }
 
@@ -2836,23 +2881,16 @@ Status DavinciModel::UpdateKnownNodeArgs(const vector<void *> &inputs, const vec
   GELOGI("DavinciModel::UpdateKnownNodeArgs in");
   GE_CHK_STATUS_RET(CreateKnownZeroCopyMap(inputs, outputs),
                     "DavinciModel::UpdateKnownNodeArgs create map for input/output zero copy.");
-  if (!base_addr_not_changed_) {
-    total_io_addrs_.clear();
-    orig_total_io_addrs_.clear();
-    for (size_t task_index = 0; task_index < task_list_.size(); ++task_index) {
-      auto &task = task_list_[task_index];
-      if (task != nullptr) {
-        Status ret = task->UpdateArgs();
-        if (ret != SUCCESS) {
-          GELOGE(FAILED, "task %zu created by davinci model is nullptr.", task_index);
-          return FAILED;
-        }
+  total_io_addrs_.clear();
+  for (size_t task_index = 0; task_index < task_list_.size(); ++task_index) {
+    auto &task = task_list_[task_index];
+    if (task != nullptr) {
+      Status ret = task->UpdateArgs();
+      if (ret != SUCCESS) {
+        GELOGE(FAILED, "task %zu created by davinci model is nullptr.", task_index);
+        return FAILED;
       }
     }
-    // cache latest iterator io addr
-    orig_total_io_addrs_ = total_io_addrs_;
-  } else {
-    total_io_addrs_ = orig_total_io_addrs_;
   }
   GE_CHK_STATUS_RET(UpdateKnownZeroCopyAddr(total_io_addrs_, false), "DavinciModel::UpdateKnownZeroCopyAddr failed.");
 
@@ -2892,6 +2930,14 @@ Status DavinciModel::InitTaskInfo(domi::ModelTaskDef &model_task_def) {
   return SUCCESS;
 }
 
+Status DavinciModel::CheckCapability(rtFeatureType_t featureType, int32_t featureInfo, bool &is_support) const {
+  int64_t value = RT_CAPABILITY_SUPPORT;
+  auto rt_ret = rtGetRtCapability(featureType, featureInfo, &value);
+  GE_CHK_BOOL_RET_STATUS(rt_ret == RT_ERROR_NONE, FAILED, "call rtGetRtCapability failed!");
+  is_support = (value == RT_CAPABILITY_SUPPORT) ? true : false;
+  return SUCCESS;
+}
+
 Status DavinciModel::MallocKnownArgs() {
   GELOGI("DavinciModel::MallocKnownArgs in");
   const auto &model_task_def = ge_model_->GetModelTaskDefPtr();
@@ -2910,20 +2956,22 @@ Status DavinciModel::MallocKnownArgs() {
       return ret;
     }
   }
+  rtError_t rt_ret;
   // malloc args memory
-  if (total_args_size_ == 0) {
-    GELOGW("DavinciModel::MallocKnownArgs total_args_size_ equals to zero.");
-    return SUCCESS;
-  }
+  bool is_support = false;
+  GE_CHK_STATUS_RET_NOLOG(CheckCapability(FEATURE_TYPE_MEMORY, MEMORY_INFO_TS_4G_LIMITED, is_support));
+  auto mem_type = is_support ? RT_MEMORY_TS_4G : RT_MEMORY_HBM;
 
-  rtError_t rt_ret = rtMalloc(&args_, total_args_size_, RT_MEMORY_HBM);
-  if (rt_ret != RT_ERROR_NONE) {
-    GELOGE(RT_FAILED, "Call rtMalloc failed, ret: 0x%X", rt_ret);
-    return RT_ERROR_TO_GE_STATUS(rt_ret);
+  if (total_args_size_ != 0) {
+    rt_ret = rtMalloc(&args_, total_args_size_, mem_type);
+    if (rt_ret != RT_ERROR_NONE) {
+      GELOGE(RT_FAILED, "Call rtMalloc failed, ret: 0x%X", rt_ret);
+      return RT_ERROR_TO_GE_STATUS(rt_ret);
+    }
   }
   // malloc dynamic and static hybrid memory
   if (total_hybrid_args_size_ != 0) {
-    rt_ret = rtMalloc(&hybrid_addrs_, total_hybrid_args_size_, RT_MEMORY_HBM);
+    rt_ret = rtMalloc(&hybrid_addrs_, total_hybrid_args_size_, mem_type);
     if (rt_ret != RT_ERROR_NONE) {
       GELOGE(RT_FAILED, "Call rtMalloc failed, ret: 0x%X", rt_ret);
       return RT_ERROR_TO_GE_STATUS(rt_ret);
@@ -2932,7 +2980,7 @@ Status DavinciModel::MallocKnownArgs() {
   // malloc fixed addr memory, eg: rts op
   if (total_fixed_addr_size_ != 0) {
     GELOGI("Begin to allocate fixed addr.");
-    rt_ret = rtMalloc(&fixed_addrs_, total_fixed_addr_size_, RT_MEMORY_HBM);
+    rt_ret = rtMalloc(&fixed_addrs_, total_fixed_addr_size_, mem_type);
     if (rt_ret != RT_ERROR_NONE) {
       GELOGE(RT_FAILED, "Call rtMalloc failed, ret: 0x%X", rt_ret);
       return RT_ERROR_TO_GE_STATUS(rt_ret);
@@ -3025,9 +3073,8 @@ Status DavinciModel::DistributeTask() {
                              task_def.kernel_ex().op_index());
     OpDescPtr op = GetOpByIndex(op_index);
     GE_CHECK_NOTNULL(op);
-
     if (reinterpret_cast<void *>(task->GetDumpArgs()) != nullptr) {
-      bool call_dump = GetDumpProperties().IsLayerNeedDump(name_, om_name_, op->GetName()) && task->CallSaveDumpInfo();
+      bool call_dump = OpNeedDump(op->GetName()) && task->CallSaveDumpInfo();
       if (call_dump || is_op_debug_reg_) {
         SaveDumpTask(task->GetTaskID(), task->GetStreamId(), op, task->GetDumpArgs());
       }
@@ -3047,11 +3094,16 @@ Status DavinciModel::DistributeTask() {
   return SUCCESS;
 }
 
-void DavinciModel::SetEndGraphId(uint32_t task_id, uint32_t stream_id) {
+bool DavinciModel::ModelNeedDump() {
   auto all_dump_model = GetDumpProperties().GetAllDumpModel();
-  bool findByOmName = all_dump_model.find(om_name_) != all_dump_model.end();
-  bool findByModelName = all_dump_model.find(name_) != all_dump_model.end();
-  if (all_dump_model.find(ge::DUMP_ALL_MODEL) != all_dump_model.end() || findByOmName || findByModelName) {
+  bool ret = all_dump_model.find(ge::DUMP_ALL_MODEL) != all_dump_model.end() ||
+             all_dump_model.find(dump_model_name_) != all_dump_model.end() ||
+             all_dump_model.find(om_name_) != all_dump_model.end();
+  return ret;
+}
+
+void DavinciModel::SetEndGraphId(uint32_t task_id, uint32_t stream_id) {
+  if (ModelNeedDump()) {
     GELOGI("start save end_graph_info to dumper, task_id is %u, stream_id is %u", task_id, stream_id);
     data_dumper_.SaveEndGraphId(task_id, stream_id);
   }
@@ -3851,7 +3903,10 @@ Status DavinciModel::TransAllVarData(ComputeGraphPtr &graph, uint32_t graph_id)
 }
 
 void DavinciModel::SetDataDumperArgs(const ComputeGraphPtr &graph, const map<string, OpDescPtr> &variable_by_name) {
-  data_dumper_.SetModelName(name_);
+  if(dump_model_name_.empty()) {
+    dump_model_name_ = name_;
+  }
+  data_dumper_.SetModelName(dump_model_name_);
   data_dumper_.SetModelId(model_id_);
   data_dumper_.SetOmName(om_name_);
   data_dumper_.SetComputeGraph(graph);
@@ -4040,7 +4095,7 @@ int64_t DavinciModel::GetFixedAddrsSize(string tensor_name) {
 Status DavinciModel::InitL1DataDumperArgs() {
   auto all_dump_model = GetDumpProperties().GetAllDumpModel();
   bool find_by_om_name = all_dump_model.find(om_name_) != all_dump_model.end();
-  bool find_by_model_name = all_dump_model.find(name_) != all_dump_model.end();
+  bool find_by_model_name = all_dump_model.find(dump_model_name_) != all_dump_model.end();
   bool dump_l1fusion_op =
     (all_dump_model.find(ge::DUMP_ALL_MODEL) != all_dump_model.end()) || find_by_om_name || find_by_model_name;
   if (dump_l1fusion_op) {
@@ -4061,4 +4116,10 @@ Status DavinciModel::InitL1DataDumperArgs() {
   return SUCCESS;
 }
 
+Status DavinciModel::SetRunAsyncListenerCallback(const RunAsyncCallback &callback) {
+  auto listener = dynamic_cast<RunAsyncListener *>(listener_.get());
+  GE_CHECK_NOTNULL(listener);
+  listener->SetCallback(callback);
+  return SUCCESS;
+}
 }  // namespace ge

ge/graph/load/model_manager/davinci_model.h

@@ -221,6 +221,11 @@ class DavinciModel {
   ///
   DataInputer *const GetDataInputer() const { return data_inputer_; }
 
+  uint32_t GetDataInputerSize() {
+    GE_CHECK_NOTNULL(data_inputer_);
+    return data_inputer_->Size();
+  }
+
   // get Stream number
   uint32_t StreamNum() const { return runtime_param_.stream_num; }
 
@@ -248,7 +253,10 @@ class DavinciModel {
   string Name() const { return name_; }
 
   // om_name
-  string OmName() const { return om_name_; }
+  const string &OmName() const { return om_name_; }
+
+  // dump_model_name
+  const string &DumpModelName() const { return dump_model_name_; }
 
   // version
   uint32_t Version() const { return version_; }
@@ -273,6 +281,8 @@ class DavinciModel {
 
   const vector<rtLabel_t> &GetLabelList() const { return label_list_; }
 
+  Status GetLabelGotoAddr(uint32_t label_index, rtMemType_t memory_type, void *&addr, uint32_t &size);
+
   Status DestroyThread();
 
   // get Op
@@ -481,6 +491,12 @@ class DavinciModel {
     data_dumper_.DumpShrink();
   }
 
+  bool OpNeedDump(const string &op_name) {
+    return GetDumpProperties().IsLayerNeedDump(dump_model_name_, om_name_, op_name);
+  }
+
+  bool ModelNeedDump();
+
   void SetEndGraphId(uint32_t task_id, uint32_t stream_id);
   DavinciModel &operator=(const DavinciModel &model) = delete;
 
@@ -528,11 +544,11 @@ class DavinciModel {
   }
   void SetKnownNode(bool known_node) { known_node_ = known_node; }
   bool IsKnownNode() { return known_node_; }
+  Status CheckCapability(rtFeatureType_t featureType, int32_t featureInfo, bool &is_support) const;
   Status MallocKnownArgs();
   Status UpdateKnownNodeArgs(const vector<void *> &inputs, const vector<void *> &outputs);
   Status CreateKnownZeroCopyMap(const vector<void *> &inputs, const vector<void *> &outputs);
   Status UpdateKnownZeroCopyAddr(vector<void *> &total_io_addrs, bool update_args = true);
-  void SetKnownNodeAddrNotChanged(bool base_addr_not_changed) { base_addr_not_changed_ = base_addr_not_changed; }
 
   Status GetOrigInputInfo(uint32_t index, OriginInputInfo &orig_input_info) const;
   Status GetAllAippInputOutputDims(uint32_t index, vector<InputOutputDims> &input_dims,
@@ -540,6 +556,7 @@ class DavinciModel {
 
   // om file name
   void SetOmName(const string &om_name) { om_name_ = om_name; }
+  void SetDumpModelName(const string &dump_model_name) { dump_model_name_ = dump_model_name; }
 
   void SetDumpProperties(const DumpProperties &dump_properties) { data_dumper_.SetDumpProperties(dump_properties); }
   const DumpProperties &GetDumpProperties() const { return data_dumper_.GetDumpProperties(); }
@@ -548,6 +565,10 @@ class DavinciModel {
     return data_dumper_.GetOpDescInfo(stream_id, task_id, op_desc_info);
   }
 
+  bool GetRunningFlag() const { return running_flg_; }
+  void SetRunningFlag(bool flag) { running_flg_ = flag; }
+  Status SetRunAsyncListenerCallback(const RunAsyncCallback &callback);
+
  private:
   // memory address of weights
   uint8_t *weights_mem_base_;
@@ -886,6 +907,7 @@ class DavinciModel {
 
   // used for inference data dump
   string om_name_;
+  string dump_model_name_;
 
   uint32_t version_;
   GeModelPtr ge_model_;  // release after DavinciModel::Init
@@ -911,6 +933,8 @@ class DavinciModel {
   shared_ptr<ModelListener> listener_;
 
   bool run_flg_;
+  // check whether model is running with data
+  bool running_flg_ = false;
 
   mutex mux_run_flg_;
 
@@ -930,6 +954,9 @@ class DavinciModel {
   vector<rtLabel_t> label_list_;
   set<uint32_t> label_id_indication_;
 
+  mutex label_args_mutex_;
+  map<uint32_t, pair<void *, uint32_t>> label_goto_args_;
+
   mutex outside_addrs_mutex_;
   vector<ZeroCopyTask> zero_copy_tasks_;  // Task used Data or NetOutput addr.
   set<const void *> copy_only_addrs_;     // Address need copy to original place.
@@ -1002,8 +1029,6 @@ class DavinciModel {
   map<const void *, void *> known_input_data_info_;
   map<const void *, void *> known_output_data_info_;
   vector<void *> total_io_addrs_;
-  vector<void *> orig_total_io_addrs_;
-  bool base_addr_not_changed_ = false;
 
   vector<vector<int64_t>> batch_info_;
   vector<vector<int64_t>> combined_batch_info_;

ge/graph/load/model_manager/model_manager.cc

@@ -271,7 +271,7 @@ ge::Status ModelManager::SetDynamicSize(uint32_t model_id, const std::vector<uin
   return SUCCESS;
 }
 
-ge::Status ModelManager::DoLoadHybridModelOnline(uint32_t model_id, const string &model_name,
+ge::Status ModelManager::DoLoadHybridModelOnline(uint32_t model_id, const string &om_name,
                                                  const shared_ptr<ge::GeRootModel> &ge_root_model,
                                                  const shared_ptr<ModelListener> &listener) {
   auto hybrid_model = hybrid::HybridDavinciModel::Create(ge_root_model);
@@ -279,13 +279,24 @@ ge::Status ModelManager::DoLoadHybridModelOnline(uint32_t model_id, const string
   hybrid_model->SetListener(listener);
   hybrid_model->SetModelId(model_id);
   hybrid_model->SetDeviceId(GetContext().DeviceId());
-  hybrid_model->SetModelName(model_name);
+  hybrid_model->SetOmName(om_name);
   GE_CHK_STATUS_RET(hybrid_model->Init(), "Failed to init hybrid model. model_id = %u", model_id);
   auto shared_model = std::shared_ptr<hybrid::HybridDavinciModel>(hybrid_model.release());
   InsertModel(model_id, shared_model);
   return SUCCESS;
 }
 
+bool ModelManager::IsNeedHybridLoad(ge::GeRootModel &ge_root_model) {
+  auto root_graph = ge_root_model.GetRootGraph();
+  if (root_graph == nullptr) {
+    GELOGE(FAILED, "no model on root model");
+    return false;
+  }
+  bool is_shape_unknown = root_graph->GetGraphUnknownFlag();
+  bool is_dsp_partitioned_graph = false;
+  (void)AttrUtils::GetBool(root_graph, ATTR_NAME_DYNAMIC_SHAPE_PARTITIONED, is_dsp_partitioned_graph);
+  return is_shape_unknown || is_dsp_partitioned_graph || GetContext().GetHostExecFlag();
+}
 ///
 /// @ingroup domi_ome
 /// @brief load model online
@@ -296,14 +307,12 @@ Status ModelManager::LoadModelOnline(uint32_t &model_id, const shared_ptr<ge::Ge
   GE_CHK_BOOL_RET_STATUS(listener.get() != nullptr, PARAM_INVALID, "Param incorrect, listener is null");
   if (model_id == INVALID_MODEL_ID) {
     GenModelId(&model_id);
+    GELOGD("Generate new model_id:%u", model_id);
   }
-
-  bool is_shape_unknown = false;
-  string model_name = "";
-  GE_CHK_STATUS_RET(ge_root_model->CheckIsUnknownShape(is_shape_unknown), "CheckIsUnknownShape failed, model id:%u",
-                    model_id);
-  if (is_shape_unknown || GetContext().GetHostExecFlag()) {
-    return DoLoadHybridModelOnline(model_id, model_name, ge_root_model, listener);
+  auto name_to_model = ge_root_model->GetSubgraphInstanceNameToModel();
+  string om_name;
+  if (IsNeedHybridLoad(*ge_root_model)) {
+    return DoLoadHybridModelOnline(model_id, om_name, ge_root_model, listener);
   }
 
   mmTimespec timespec = mmGetTickCount();
@@ -324,7 +333,6 @@ Status ModelManager::LoadModelOnline(uint32_t &model_id, const shared_ptr<ge::Ge
   auto root_graph = ge_root_model->GetRootGraph();
   GE_CHECK_NOTNULL(root_graph);
   string root_model_name = root_graph->GetName();
-  auto name_to_model = ge_root_model->GetSubgraphInstanceNameToModel();
   GeModelPtr ge_model = name_to_model[root_model_name];
   Status ret = SUCCESS;
   do {
@@ -332,7 +340,18 @@ Status ModelManager::LoadModelOnline(uint32_t &model_id, const shared_ptr<ge::Ge
     GE_IF_BOOL_EXEC(SUCCESS != (ret = davinci_model->Assign(ge_model)), GELOGW("assign model to modeldef failed.");
                     break;);
     GE_TIMESTAMP_END(Assign, "GraphLoader::ModelAssign");
-
+    /// 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.
+    /// Update session_id for infer in load model to avoid the same session_id.
+    if (!ge_root_model->GetTrainFlag()) {
+      uint64_t new_session_id;
+      ret = GenSessionId(new_session_id);
+      GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(ret != SUCCESS, return ret, "Generate session_id for infer failed.");
+      ret = davinci_model->UpdateSessionId(new_session_id);
+      GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(ret != SUCCESS, return ret, "Update session_id for infer failed.");
+      ge_model->InsertSessionMap(model_id, new_session_id);
+      GELOGD("Update new session id: %lu.", new_session_id);
+    }
     GE_TIMESTAMP_START(Init);
     GE_IF_BOOL_EXEC(SUCCESS != (ret = davinci_model->Init()), GELOGW("DavinciInit failed."); break;);
     GE_TIMESTAMP_END(Init, "GraphLoader::ModelInit");
@@ -345,16 +364,16 @@ Status ModelManager::LoadModelOnline(uint32_t &model_id, const shared_ptr<ge::Ge
   return ret;
 }
 
-void ModelManager::InsertModel(uint32_t id, std::shared_ptr<DavinciModel> &davinci_model) {
-  GE_CHK_BOOL_EXEC(davinci_model != nullptr, return, "davinci_model ptr is null, id: %u", id);
+void ModelManager::InsertModel(uint32_t model_id, std::shared_ptr<DavinciModel> &davinci_model) {
+  GE_CHK_BOOL_EXEC(davinci_model != nullptr, return, "davinci_model ptr is null, id: %u", model_id);
   std::lock_guard<std::recursive_mutex> lock(map_mutex_);
-  model_map_[id] = davinci_model;
+  model_map_[model_id] = davinci_model;
 }
 
-void ModelManager::InsertModel(uint32_t id, shared_ptr<hybrid::HybridDavinciModel> &hybrid_model) {
-  GE_CHK_BOOL_EXEC(hybrid_model != nullptr, return, "hybrid_model ptr is null, id: %u", id);
+void ModelManager::InsertModel(uint32_t model_id, shared_ptr<hybrid::HybridDavinciModel> &hybrid_model) {
+  GE_CHK_BOOL_EXEC(hybrid_model != nullptr, return, "hybrid_model ptr is null, id: %u", model_id);
   std::lock_guard<std::recursive_mutex> lock(map_mutex_);
-  hybrid_model_map_[id] = hybrid_model;
+  hybrid_model_map_[model_id] = hybrid_model;
 }
 
 Status ModelManager::DeleteModel(uint32_t id) {

ge/graph/load/model_manager/model_manager.h

@@ -294,6 +294,7 @@ class FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY ModelManager {
                                        std::vector<InputOutputDims> &output_dims);
 
   bool IsDynamicShape(uint32_t model_id);
+  bool IsNeedHybridLoad(ge::GeRootModel &ge_root_model);
   ge::Status GetOpDescInfo(uint32_t device_id, uint32_t stream_id, uint32_t task_id, OpDescInfo &op_desc_info);
 
   ge::Status EnableExceptionDump(const std::map<string, string> &options);
@@ -329,8 +330,8 @@ class FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY ModelManager {
   /// @ingroup domi_ome
   /// @brief insert new model into model manager set
   ///
-  void InsertModel(uint32_t id, std::shared_ptr<DavinciModel> &davinci_model);
-  void InsertModel(uint32_t id, std::shared_ptr<hybrid::HybridDavinciModel> &hybrid_model);
+  void InsertModel(uint32_t model_id, std::shared_ptr<DavinciModel> &davinci_model);
+  void InsertModel(uint32_t model_id, std::shared_ptr<hybrid::HybridDavinciModel> &hybrid_model);
 
   ///
   /// @ingroup domi_ome

ge/graph/load/model_manager/model_utils.cc

@@ -384,7 +384,8 @@ Status ModelUtils::GetVarAddr(const RuntimeParam &model_param, const ConstOpDesc
   switch (mem_type) {
     case RT_MEMORY_RDMA_HBM:
       if (offset < 0) {
-        GELOGE(PARAM_INVALID, "rdma var addr is invalid, addr=%p", reinterpret_cast<uint8_t *>(offset));
+        GELOGE(PARAM_INVALID, "rdma var addr is invalid, addr=%p",
+               reinterpret_cast<uint8_t *>(static_cast<uintptr_t>(offset)));
         return PARAM_INVALID;
       }
       var_addr = reinterpret_cast<uint8_t *>(static_cast<uintptr_t>(offset));

ge/graph/load/model_manager/task_info/end_graph_task_info.cc

@@ -45,10 +45,7 @@ Status EndGraphTaskInfo::Init(const domi::TaskDef &task_def, DavinciModel *davin
 Status EndGraphTaskInfo::Distribute() {
   GELOGI("EndGraphTaskInfo Distribute Start.");
   GE_CHECK_NOTNULL(davinci_model_);
-  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()) {
+  if (davinci_model_->ModelNeedDump()) {
     GELOGI("Start to call rtEndGraphEx");
     rtError_t rt_ret = rtEndGraphEx(model_, stream_, kDumpFlag);
     if (rt_ret != RT_ERROR_NONE) {

ge/graph/load/model_manager/task_info/kernel_ex_task_info.cc

@@ -238,8 +238,7 @@ Status KernelExTaskInfo::Init(const domi::TaskDef &task_def, DavinciModel *davin
 }
 
 void KernelExTaskInfo::InitDumpTask(void *addr, const OpDescPtr &op_desc) {
-  if (davinci_model_->GetDumpProperties().IsLayerNeedDump(davinci_model_->Name(), davinci_model_->OmName(),
-                                                          op_desc->GetName())) {
+  if (davinci_model_->OpNeedDump(op_desc->GetName())) {
     dump_flag_ = RT_KERNEL_DUMPFLAG;
     dump_args_ = addr;
   }

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

@@ -124,7 +124,8 @@ Status KernelTaskInfo::Init(const domi::TaskDef &task_def, DavinciModel *davinci
       return FAILED;
     }
 
-    ret = InitTVMTask(args_offset_tmp[0], kernel_def);
+    io_addr_offset_ = args_offset_tmp[0];
+    ret = InitTVMTask(io_addr_offset_, kernel_def);
   } else if (kernel_type_ == ccKernelType::CUSTOMIZED) {
     ret = InitAICPUCustomTask(context.op_index(), kernel_def);
   } else if (kernel_type_ == ccKernelType::AI_CPU || kernel_type_ == ccKernelType::CUST_AI_CPU) {
@@ -380,7 +381,8 @@ Status KernelTaskInfo::Distribute() {
   GELOGD("KernelTaskInfo Distribute Start.");
   if (davinci_model_->IsKnownNode()) {
     if (kernel_type_ == ccKernelType::TE) {
-      args_ = davinci_model_->GetCurrentArgsAddr(args_offset_);
+      args_ = l2_buffer_on_ ? davinci_model_->GetCurrentHybridArgsAddr(hybrid_args_offset_)
+                            : davinci_model_->GetCurrentArgsAddr(args_offset_);
     } else if (kernel_type_ == ccKernelType::AI_CPU || kernel_type_ == ccKernelType::CUST_AI_CPU) {
       args_ = davinci_model_->GetCurrentHybridArgsAddr(hybrid_args_offset_);
     }
@@ -407,10 +409,7 @@ Status KernelTaskInfo::Distribute() {
         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)
     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()) {
+    if (davinci_model_->ModelNeedDump()) {
       open_dump = true;
     }
     if (call_skt && (env_flag != kCloseSkt) && !open_dump) {
@@ -449,29 +448,41 @@ void KernelTaskInfo::SetIoAddrs(const OpDescPtr &op_desc) {
   }
 }
 
+Status KernelTaskInfo::CopyNoncontinuousArgs(uint16_t offset) {
+  GE_CHECK_NOTNULL(davinci_model_);
+  // copy new io addrs
+  vector<void *> io_addrs = io_addrs_;
+  davinci_model_->UpdateKnownZeroCopyAddr(io_addrs);
+  auto addr_size = kAddrLen * io_addrs.size();
+
+  // copy io addr
+  errno_t sec_ret = memcpy_s(args_addr.get() + offset, addr_size, io_addrs.data(), addr_size);
+  if (sec_ret != EOK) {
+    GELOGE(FAILED, "memcpy failed, ret: %d", sec_ret);
+    return FAILED;
+  }
+
+  // copy args to device
+  rtError_t rt_ret = rtMemcpy(args_, args_size_, args_addr.get(), args_size_, RT_MEMCPY_HOST_TO_DEVICE);
+  if (rt_ret != RT_ERROR_NONE) {
+    GELOGE(RT_FAILED, "Call rt api(rtMemcpy) failed, ret: 0x%X", rt_ret);
+    return RT_ERROR_TO_GE_STATUS(rt_ret);
+  }
+  GELOGD("Copy noncontinuous args success, kernel type %d.", kernel_type_);
+  return SUCCESS;
+}
+
 Status KernelTaskInfo::UpdateArgs() {
   GELOGI("KernelTaskInfo::UpdateArgs in.");
+  GE_CHECK_NOTNULL(davinci_model_);
   if (kernel_type_ == ccKernelType::TE) {
+    if (l2_buffer_on_) {
+      return CopyNoncontinuousArgs(io_addr_offset_);
+    }
     davinci_model_->SetTotalIOAddrs(io_addrs_);
   } else if (kernel_type_ == ccKernelType::AI_CPU || kernel_type_ == ccKernelType::CUST_AI_CPU) {
-    vector<void *> io_addrs = io_addrs_;
-    davinci_model_->UpdateKnownZeroCopyAddr(io_addrs);
-    uintptr_t io_addr = reinterpret_cast<uintptr_t>(args_addr.get()) + sizeof(aicpu::AicpuParamHead);
-    auto addrs_size = sizeof(uint64_t) * io_addrs.size();
-    errno_t sec_ret = memcpy_s(reinterpret_cast<void *>(io_addr), addrs_size, io_addrs.data(), addrs_size);
-    if (sec_ret != EOK) {
-      GELOGE(FAILED, "memcpy failed, ret: %d", sec_ret);
-      return FAILED;
-    }
-    // copy args to device
-    rtError_t rt_ret = rtMemcpy(args_, args_size_, args_addr.get(), args_size_, RT_MEMCPY_HOST_TO_DEVICE);
-    if (rt_ret != RT_ERROR_NONE) {
-      GELOGE(RT_FAILED, "Call rt api(rtMemcpy) failed, ret: 0x%X", rt_ret);
-      return RT_ERROR_TO_GE_STATUS(rt_ret);
-    }
+    return CopyNoncontinuousArgs(sizeof(aicpu::AicpuParamHead));
   }
-
-  GELOGI("KernelTaskInfo::UpdateArgs success.");
   return SUCCESS;
 }
 
@@ -516,8 +527,8 @@ Status KernelTaskInfo::UpdateL2Data(const domi::KernelDef &kernel_def) {
     return SUCCESS;
   }
 
-  char *sm_contrl = const_cast<char *>(sm_desc.data());
-  rtL2Ctrl_t *l2_ctrl_info = reinterpret_cast<rtL2Ctrl_t *>(sm_contrl);
+  char *sm_control = const_cast<char *>(sm_desc.data());
+  rtL2Ctrl_t *l2_ctrl_info = reinterpret_cast<rtL2Ctrl_t *>(sm_control);
   uint64_t gen_base_addr = davinci_model_->GetRtBaseAddr();
 
   // There is no weight for te op now. Update L2_mirror_addr by data memory base.
@@ -545,19 +556,31 @@ Status KernelTaskInfo::UpdateL2Data(const domi::KernelDef &kernel_def) {
   return SUCCESS;
 }
 
+void KernelTaskInfo::SetContinuousArgs(uint32_t args_size, DavinciModel *davinci_model) {
+  args_offset_ = davinci_model->GetTotalArgsSize();
+  davinci_model->SetTotalArgsSize(args_size);
+}
+
+void KernelTaskInfo::SetNoncontinuousArgs(uint32_t args_size, DavinciModel *davinci_model) {
+  hybrid_args_offset_ = davinci_model->GetHybridArgsSize();
+  davinci_model->SetHybridArgsSize(args_size);
+}
+
 Status KernelTaskInfo::CalculateArgs(const domi::TaskDef &task_def, DavinciModel *davinci_model) {
+  GE_CHECK_NOTNULL(davinci_model);
   const domi::KernelDef &kernel_def = task_def.kernel();
   const domi::KernelContext &context = kernel_def.context();
   kernel_type_ = static_cast<ccKernelType>(context.kernel_type());
+  uint32_t args_size = kernel_def.args_size();
   if (kernel_type_ == ccKernelType::TE) {
-    uint32_t args_size = kernel_def.args_size();
-    args_offset_ = davinci_model->GetTotalArgsSize();
-    davinci_model->SetTotalArgsSize(args_size);
-    GELOGI("kernel task name , args_size %u, args_offset %u", args_size, args_offset_);
+    if (kernel_def.sm_desc().empty()) {
+      SetContinuousArgs(args_size, davinci_model);
+      return SUCCESS;
+    }
+    l2_buffer_on_ = true;
+    SetNoncontinuousArgs(args_size, davinci_model);
   } else if (kernel_type_ == ccKernelType::AI_CPU || kernel_type_ == ccKernelType::CUST_AI_CPU) {
-    hybrid_args_offset_ = davinci_model->GetHybridArgsSize();
-    davinci_model->SetHybridArgsSize(kernel_def.args_size());
-    GELOGI("aicpu kernel task name , args_size %u, args_offset %u", kernel_def.args_size(), hybrid_args_offset_);
+    SetNoncontinuousArgs(args_size, davinci_model);
   }
   return SUCCESS;
 }
@@ -568,8 +591,23 @@ Status KernelTaskInfo::InitTVMTask(uint16_t offset, const domi::KernelDef &kerne
   // get tvm op desc
   OpDescPtr op_desc = davinci_model_->GetOpByIndex(ctx_.opIndex);
   GE_CHECK_NOTNULL(op_desc);
+
+  args_addr = std::unique_ptr<uint8_t[]>(new (std::nothrow) uint8_t[args_size_]);
+  errno_t sec_ret = memcpy_s(args_addr.get(), args_size_, kernel_def.args().data(), args_size_);
+  if (sec_ret != EOK) {
+    GELOGE(FAILED, "memcpy failed, ret: %d", sec_ret);
+    return FAILED;
+  }
+
+  Status ge_ret = UpdateL2Data(kernel_def);
+  // update origin l2 data
+  if (ge_ret != SUCCESS) {
+    return ge_ret;
+  }
+
   if (davinci_model_->IsKnownNode()) {
-    args_ = davinci_model_->GetCurrentArgsAddr(args_offset_);
+    args_ = l2_buffer_on_ ? davinci_model_->GetCurrentHybridArgsAddr(hybrid_args_offset_)
+                          : davinci_model_->GetCurrentArgsAddr(args_offset_);
     InitDumpTask(offset);
     return SUCCESS;
   }
@@ -609,12 +647,6 @@ Status KernelTaskInfo::InitTVMTask(uint16_t offset, const domi::KernelDef &kerne
     GELOGE(RT_FAILED, "Call rt api failed, ret: 0x%X", rt_ret);
     return RT_ERROR_TO_GE_STATUS(rt_ret);
   }
-  vector<uint8_t> args_info(args_size_);
-  errno_t sec_ret = memcpy_s(args_info.data(), args_size_, kernel_def.args().data(), args_size_);
-  if (sec_ret != EOK) {
-    GELOGE(FAILED, "memcpy failed, ret: %d", sec_ret);
-    return FAILED;
-  }
 
   if ((args_size_ <= offset) || (args_size_ - offset < kAddrLen * tensor_device_addrs.size())) {
     GELOGE(FAILED, "offset >= kernelInfo.argsSize or copy content beyond applied memory.");
@@ -628,7 +660,7 @@ Status KernelTaskInfo::InitTVMTask(uint16_t offset, const domi::KernelDef &kerne
     GELOGE(RT_FAILED, "Call rt api failed, ret: 0x%X", rt_ret);
     return RT_ERROR_TO_GE_STATUS(rt_ret);
   }
-  sec_ret = memcpy_s(args_info.data() + offset, args_size_ - offset, tensor_device_addrs.data(),
+  sec_ret = memcpy_s(args_addr.get() + offset, args_size_ - offset, tensor_device_addrs.data(),
                      kAddrLen * tensor_device_addrs.size());
   if (sec_ret != EOK) {
     GELOGE(FAILED, "memcpy failed, ret: %d", sec_ret);
@@ -640,19 +672,13 @@ Status KernelTaskInfo::InitTVMTask(uint16_t offset, const domi::KernelDef &kerne
   GE_CHK_BOOL_TRUE_EXEC_INFO(davinci_model_->GetOpDugReg(), dump_args_ = static_cast<char *>(args_) + offset,
                              "Op debug is open in TVM task info");
 
-  Status ge_ret = UpdateL2Data(kernel_def);
-  // update origin l2 data
-  if (ge_ret != SUCCESS) {
-    return ge_ret;
-  }
-
   vector<void *> virtual_io_addrs;  // use virtual address for zero copy key.
   virtual_io_addrs.insert(virtual_io_addrs.end(), input_data_addrs.begin(), input_data_addrs.end());
   virtual_io_addrs.insert(virtual_io_addrs.end(), output_data_addrs.begin(), output_data_addrs.end());
   if (op_desc->GetType() == ATOMICADDRCLEAN) {
     virtual_io_addrs.insert(virtual_io_addrs.end(), workspace_data_addrs.begin(), workspace_data_addrs.end());
   }
-  davinci_model_->SetZeroCopyAddr(op_desc, virtual_io_addrs, args_info.data(), args_, args_size_, offset);
+  davinci_model_->SetZeroCopyAddr(op_desc, virtual_io_addrs, args_addr.get(), args_, args_size_, offset);
 
   GELOGD("Do InitTVMTask end");
   return SUCCESS;
@@ -951,8 +977,7 @@ Status KernelTaskInfo::InitAicpuTask(uint32_t op_index, const domi::KernelDef &k
 }
 
 void KernelTaskInfo::InitDumpTask(uint32_t offset) {
-  if (davinci_model_->GetDumpProperties().IsLayerNeedDump(davinci_model_->Name(), davinci_model_->OmName(),
-                                                          op_desc_->GetName())) {
+  if (davinci_model_->OpNeedDump(op_desc_->GetName())) {
     if (IsL1FusionOp(op_desc_)) {
       dump_flag_ = RT_FUSION_KERNEL_DUMPFLAG;
     } else {

ge/graph/load/model_manager/task_info/kernel_task_info.h

@@ -129,6 +129,9 @@ class KernelTaskInfo : public TaskInfo {
   bool IsL1FusionOp(const OpDescPtr &op_desc);
   void SetIoAddrs(const OpDescPtr &op_desc);
   void InitDumpTask(uint32_t offset);
+  void SetContinuousArgs(uint32_t args_size, DavinciModel *davinci_model);
+  void SetNoncontinuousArgs(uint32_t args_size, DavinciModel *davinci_model);
+  Status CopyNoncontinuousArgs(uint16_t offset);
 
   // For super kernel
   Status SaveSKTDumpInfo();
@@ -163,6 +166,8 @@ class KernelTaskInfo : public TaskInfo {
   uint32_t hybrid_args_offset_ = 0;
   int64_t fixed_addr_offset_ = 0;
   std::unique_ptr<uint8_t[]> args_addr = nullptr;
+  uint16_t io_addr_offset_ = 0;
+  bool l2_buffer_on_ = false;
   bool call_save_dump_ = false;
 
   // aicpu ext_info device mem

ge/graph/load/model_manager/task_info/label_goto_ex_task_info.cc

@@ -17,9 +17,15 @@
 #include "graph/load/model_manager/task_info/label_goto_ex_task_info.h"
 
 #include "graph/load/model_manager/davinci_model.h"
-#include "graph/debug/ge_attr_define.h"
 
 namespace ge {
+constexpr uint8_t kGotoBranchMax = 1;
+
+LabelGotoExTaskInfo::~LabelGotoExTaskInfo() {
+  args_ = nullptr;
+  GE_FREE_RT_LOG(index_value_);
+}
+
 Status LabelGotoExTaskInfo::Init(const domi::TaskDef &task_def, DavinciModel *davinci_model) {
   GELOGI("LabelGotoExTaskInfo Init Start.");
   GE_CHECK_NOTNULL(davinci_model);
@@ -28,7 +34,7 @@ Status LabelGotoExTaskInfo::Init(const domi::TaskDef &task_def, DavinciModel *da
     return FAILED;
   }
 
-  // Get LabelGoto task def
+  // Get LabelGotoEx task def
   const domi::LabelGotoExDef &label_goto = task_def.label_goto_ex();
   OpDescPtr op_desc = davinci_model->GetOpByIndex(label_goto.op_index());
   if (op_desc == nullptr) {
@@ -43,20 +49,38 @@ Status LabelGotoExTaskInfo::Init(const domi::TaskDef &task_def, DavinciModel *da
     return INTERNAL_ERROR;
   }
 
-  const vector<rtLabel_t> &label_list = davinci_model->GetLabelList();
-  if (label_index >= label_list.size()) {
-    GELOGE(PARAM_INVALID, "LabelGotoExTaskInfo: Invalid label id:%u, label size:%zu", label_index, label_list.size());
-    return INTERNAL_ERROR;
+  rtMemType_t memory_type = op_desc->HasAttr(ATTR_NAME_MEMORY_TYPE_RANGE) ? RT_MEMORY_TS_4G : RT_MEMORY_HBM;
+  GELOGI("memory_type: %u", memory_type);
+
+  GE_CHK_STATUS_RET_NOLOG(davinci_model->GetLabelGotoAddr(label_index, memory_type, args_, args_size_));
+
+  rtError_t rt_ret = rtMalloc(&index_value_, sizeof(uint64_t), memory_type);
+  if (rt_ret != RT_ERROR_NONE) {
+    GELOGE(RT_FAILED, "Call rtMalloc failed, error: %#x", rt_ret);
+    return RT_ERROR_TO_GE_STATUS(rt_ret);
   }
-  label_ = label_list[label_index];
 
-  GELOGI("LabelGotoExTaskInfo Init Success, label id:%u, label:%p.", label_index, label_);
+  uint64_t branch_index = 0;
+  rt_ret = rtMemcpy(index_value_, sizeof(uint64_t), &branch_index, sizeof(uint64_t), RT_MEMCPY_HOST_TO_DEVICE);
+  if (rt_ret != RT_ERROR_NONE) {
+    GELOGE(RT_FAILED, "Call rtMemcpy failed, error: %#x", rt_ret);
+    return RT_ERROR_TO_GE_STATUS(rt_ret);
+  }
+
+  GELOGI("LabelGotoExTaskInfo Init Success, label id:%u", label_index);
   return SUCCESS;
 }
 
 Status LabelGotoExTaskInfo::Distribute() {
   GELOGI("LabelGotoExTaskInfo Distribute Start.");
-  rtError_t rt_ret = rtLabelGotoEx(label_, stream_);
+  GE_CHECK_NOTNULL(args_);
+  GE_CHECK_NOTNULL(index_value_);
+  if (args_size_ == 0) {
+    GELOGE(PARAM_INVALID, "branch max: %u, args size: %u invalid.", kGotoBranchMax, args_size_);
+    return PARAM_INVALID;
+  }
+
+  rtError_t rt_ret = rtLabelSwitchByIndex(index_value_, kGotoBranchMax, args_, stream_);
   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);

ge/graph/load/model_manager/task_info/label_goto_ex_task_info.h

@@ -14,24 +14,26 @@
  * limitations under the License.
  */
 
-#ifndef GE_GRAPH_LOAD_NEW_MODEL_MANAGER_TASK_INFO_LABEL_GOTO_EX_TASK_INFO_H_
-#define GE_GRAPH_LOAD_NEW_MODEL_MANAGER_TASK_INFO_LABEL_GOTO_EX_TASK_INFO_H_
+#ifndef GE_GRAPH_LOAD_MODEL_MANAGER_TASK_INFO_LABEL_GOTO_EX_TASK_INFO_H_
+#define GE_GRAPH_LOAD_MODEL_MANAGER_TASK_INFO_LABEL_GOTO_EX_TASK_INFO_H_
 
 #include "graph/load/model_manager/task_info/task_info.h"
 
 namespace ge {
 class LabelGotoExTaskInfo : public TaskInfo {
  public:
-  LabelGotoExTaskInfo() : label_(nullptr) {}
+  LabelGotoExTaskInfo() = default;
 
-  ~LabelGotoExTaskInfo() override { label_ = nullptr; }
+  ~LabelGotoExTaskInfo() override;
 
   Status Init(const domi::TaskDef &task_def, DavinciModel *davinci_model) override;
 
   Status Distribute() override;
 
  private:
-  void *label_;
+  void *index_value_{nullptr};    // switch index input.
+  void *args_{nullptr};           // label info memory.
+  uint32_t args_size_{0};         // label info length.
 };
 }  // namespace ge
-#endif  // GE_GRAPH_LOAD_NEW_MODEL_MANAGER_TASK_INFO_LABEL_GOTO_EX_TASK_INFO_H_
+#endif  // GE_GRAPH_LOAD_MODEL_MANAGER_TASK_INFO_LABEL_GOTO_EX_TASK_INFO_H_

ge/graph/load/model_manager/task_info/label_set_task_info.h

@@ -14,8 +14,8 @@
  * limitations under the License.
  */
 
-#ifndef GE_GRAPH_LOAD_NEW_MODEL_MANAGER_TASK_INFO_LABEL_SET_TASK_INFO_H_
-#define GE_GRAPH_LOAD_NEW_MODEL_MANAGER_TASK_INFO_LABEL_SET_TASK_INFO_H_
+#ifndef GE_GRAPH_LOAD_MODEL_MANAGER_TASK_INFO_LABEL_SET_TASK_INFO_H_
+#define GE_GRAPH_LOAD_MODEL_MANAGER_TASK_INFO_LABEL_SET_TASK_INFO_H_
 
 #include "graph/load/model_manager/task_info/task_info.h"
 
@@ -34,4 +34,4 @@ class LabelSetTaskInfo : public TaskInfo {
   void *label_;
 };
 }  // namespace ge
-#endif  // GE_GRAPH_LOAD_NEW_MODEL_MANAGER_TASK_INFO_LABEL_SET_TASK_INFO_H_
+#endif  // GE_GRAPH_LOAD_MODEL_MANAGER_TASK_INFO_LABEL_SET_TASK_INFO_H_

ge/graph/load/model_manager/task_info/label_switch_by_index_task_info.cc

@@ -16,20 +16,13 @@
 
 #include "graph/load/model_manager/task_info/label_switch_by_index_task_info.h"
 
-#include "graph/debug/ge_attr_define.h"
 #include "graph/load/model_manager/davinci_model.h"
 
 namespace ge {
 constexpr uint8_t kLabelSwitchIndexNum = 1;
 
 LabelSwitchByIndexTaskInfo::~LabelSwitchByIndexTaskInfo() {
-  if (args_ != nullptr) {
-    rtError_t ret = rtFree(args_);
-    if (ret != RT_ERROR_NONE) {
-      GELOGE(RT_FAILED, "Call rt api failed, ret: 0x%X", ret);
-    }
-  }
-  args_ = nullptr;
+  GE_FREE_RT_LOG(args_);
   index_value_ = nullptr;
 }
 
@@ -37,13 +30,12 @@ Status LabelSwitchByIndexTaskInfo::Init(const domi::TaskDef &task_def, DavinciMo
   GELOGI("LabelSwitchByIndexTaskInfo Init Start.");
   GE_CHECK_NOTNULL(davinci_model);
 
-  const vector<rtLabel_t> &label_list = davinci_model->GetLabelList();
   Status ret = SetStream(task_def.stream_id(), davinci_model->GetStreamList());
   if (ret != SUCCESS) {
     return FAILED;
   }
 
-  // Get LabelSwitch task def
+  // Get LabelSwitchByIndex task def
   const domi::LabelSwitchByIndexDef &label_switch = task_def.label_switch_by_index();
   OpDescPtr op_desc = davinci_model->GetOpByIndex(label_switch.op_index());
   if (op_desc == nullptr) {
@@ -68,7 +60,7 @@ Status LabelSwitchByIndexTaskInfo::Init(const domi::TaskDef &task_def, DavinciMo
 
   davinci_model->DisableZeroCopy(index_value_);
 
-  std::vector<uint32_t> label_idx_list;
+  vector<uint32_t> label_idx_list;
   if (!AttrUtils::GetListInt(op_desc, ATTR_NAME_LABEL_SWITCH_LIST, label_idx_list)) {
     GELOGE(INTERNAL_ERROR, "LabelSwitchByIndexTaskInfo: %s Get attr %s failed.", op_desc->GetName().c_str(),
            ATTR_NAME_LABEL_SWITCH_LIST.c_str());
@@ -81,7 +73,8 @@ Status LabelSwitchByIndexTaskInfo::Init(const domi::TaskDef &task_def, DavinciMo
     return INTERNAL_ERROR;
   }
 
-  label_list_.resize(branch_max_, nullptr);
+  vector<rtLabel_t> label_used(branch_max_, nullptr);
+  const vector<rtLabel_t> &label_list = davinci_model->GetLabelList();
   for (size_t idx = 0; idx < label_idx_list.size(); ++idx) {
     uint32_t label_id = label_idx_list[idx];
     if (label_id >= label_list.size()) {
@@ -90,8 +83,7 @@ Status LabelSwitchByIndexTaskInfo::Init(const domi::TaskDef &task_def, DavinciMo
       return INTERNAL_ERROR;
     }
     GE_CHECK_NOTNULL(label_list[label_id]);
-
-    label_list_[idx] = label_list[label_id];
+    label_used[idx] = label_list[label_id];
   }
 
   rtMemType_t memory_type = op_desc->HasAttr(ATTR_NAME_MEMORY_TYPE_RANGE) ? RT_MEMORY_TS_4G : RT_MEMORY_HBM;
@@ -103,7 +95,7 @@ Status LabelSwitchByIndexTaskInfo::Init(const domi::TaskDef &task_def, DavinciMo
     return RT_ERROR_TO_GE_STATUS(rt_ret);
   }
 
-  rt_ret = rtLabelListCpy(label_list_.data(), label_list_.size(), args_, args_size_);
+  rt_ret = rtLabelListCpy(label_used.data(), label_used.size(), args_, args_size_);
   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);
@@ -125,7 +117,7 @@ Status LabelSwitchByIndexTaskInfo::Distribute() {
   rtError_t rt_ret = rtLabelSwitchByIndex(index_value_, branch_max_, args_, stream_);
   if (rt_ret != RT_ERROR_NONE) {
     GELOGE(RT_FAILED, "Call rt api failed, ret: 0x%X", rt_ret);
-    return RT_FAILED;
+    return RT_ERROR_TO_GE_STATUS(rt_ret);
   }
 
   GELOGI("LabelSwitchByIndexTaskInfo Distribute Success.");

ge/graph/load/model_manager/task_info/label_switch_by_index_task_info.h

@@ -14,16 +14,15 @@
  * limitations under the License.
  */
 
-#ifndef GE_GRAPH_LOAD_NEW_MODEL_MANAGER_TASK_INFO_LABEL_SWITCH_BY_INDEX_TASK_INFO_H_
-#define GE_GRAPH_LOAD_NEW_MODEL_MANAGER_TASK_INFO_LABEL_SWITCH_BY_INDEX_TASK_INFO_H_
+#ifndef GE_GRAPH_LOAD_MODEL_MANAGER_TASK_INFO_LABEL_SWITCH_BY_INDEX_TASK_INFO_H_
+#define GE_GRAPH_LOAD_MODEL_MANAGER_TASK_INFO_LABEL_SWITCH_BY_INDEX_TASK_INFO_H_
 
 #include "graph/load/model_manager/task_info/task_info.h"
 
 namespace ge {
 class LabelSwitchByIndexTaskInfo : public TaskInfo {
  public:
-  LabelSwitchByIndexTaskInfo()
-      : index_value_(nullptr), branch_max_(0), args_(nullptr), args_size_(0), fixed_addr_offset_(0) {}
+  LabelSwitchByIndexTaskInfo() = default;
 
   ~LabelSwitchByIndexTaskInfo() override;
 
@@ -34,12 +33,11 @@ class LabelSwitchByIndexTaskInfo : public TaskInfo {
   Status CalculateArgs(const domi::TaskDef &task_def, DavinciModel *davinci_model) override;
 
  private:
-  void *index_value_;    // switch index input.
-  uint32_t branch_max_;  // max branch count.
-  void *args_;           // label info memory.
-  uint32_t args_size_;   // label info length.
-  std::vector<rtLabel_t> label_list_;
-  int64_t fixed_addr_offset_;
+  void *index_value_{nullptr};    // switch index input.
+  uint32_t branch_max_{0};        // max branch count.
+  void *args_{nullptr};           // label info memory.
+  uint32_t args_size_{0};         // label info length.
+  int64_t fixed_addr_offset_{0};
 };
 }  // namespace ge
-#endif  // GE_GRAPH_LOAD_NEW_MODEL_MANAGER_TASK_INFO_LABEL_SWITCH_BY_INDEX_TASK_INFO_H_
+#endif  // GE_GRAPH_LOAD_MODEL_MANAGER_TASK_INFO_LABEL_SWITCH_BY_INDEX_TASK_INFO_H_

ge/graph/manager/graph_caching_allocator.cc

@@ -40,7 +40,7 @@ static bool BlockComparator(const Block *left, const Block *right) {
 }
 
 bool CanMerge(Block *block) {
-  if (block == nullptr || block->allocated || !block->IsSplit()) {
+  if ((block == nullptr) || block->allocated || !block->IsSplit()) {
     return false;
   }
   return true;
@@ -52,7 +52,7 @@ size_t GetBinIndex(size_t size) {
     if (size <= range) {
       break;
     }
-    ++index;
+    index++;
   }
   if (index > kNumBins - 1) {
     index = kNumBins - 1;
@@ -87,15 +87,15 @@ bool ShouldSplit(const Block *block, size_t size) {
 
 void IncreaseCount(std::map<size_t, size_t> &count, size_t size) {
   auto it = count.find(size);
-  if (it != count.end()) {
-    it->second++;
-  } else  {
+  if (it == count.end()) {
     count.emplace(size, 1);
+  } else  {
+    it->second++;
   }
 }
 
 CachingAllocator::CachingAllocator(rtMemType_t memory_type) : memory_type_(memory_type), memory_allocator_(nullptr) {
-  for (uint32_t i = 0; i < kNumBins; ++i) {
+  for (uint32_t i = 0; i < kNumBins; i++) {
     free_block_bins_[i] = nullptr;
   }
 }
@@ -105,7 +105,7 @@ Status CachingAllocator::Initialize(uint32_t device_id) {
   // when redo Initialize free old memory
   FreeBlocks();
   std::lock_guard<std::recursive_mutex> lock(mutex_);
-  for (uint32_t i = 0; i < kNumBins; ++i) {
+  for (uint32_t i = 0; i < kNumBins; i++) {
     if (free_block_bins_[i] != nullptr) {
       continue;
     }
@@ -132,18 +132,18 @@ void CachingAllocator::Finalize(uint32_t device_id) {
 
 uint8_t *CachingAllocator::Malloc(size_t size, uint8_t *org_ptr, uint32_t device_id) {
   GELOGI("Start malloc pool memory, size = %zu, device id = %u", size, device_id);
-  uint8_t *ptr = nullptr;
   size = GetBlockSize(size);
+  uint8_t *ptr = nullptr;
   Block *block = FindFreeBlock(size, org_ptr, device_id);
-  if (block != nullptr) {
-    ptr = block->ptr;
-  } else {
+  if (block == nullptr) {
     if (ge::SUCCESS == TryExtendCache(size, device_id)) {
       block = FindFreeBlock(size, org_ptr, device_id);
       if (block != nullptr) {
         ptr = block->ptr;
       }
     }
+  } else {
+    ptr = block->ptr;
   }
   if (ptr == nullptr) {
     GELOGE(FAILED, "Malloc failed device id = %u, size= %zu", device_id, size);
@@ -171,7 +171,7 @@ Status CachingAllocator::Free(uint8_t *ptr, uint32_t device_id) {
 }
 
 void CachingAllocator::FreeBlock(Block *block) {
-  if (block == nullptr || !block->allocated) {
+  if ((block == nullptr) || !block->allocated) {
     return;
   }
   GELOGI("Free block size = %zu", block->size);
@@ -187,7 +187,7 @@ void CachingAllocator::FreeBlock(Block *block) {
 }
 
 void CachingAllocator::MergeBlocks(Block *dst, Block *src, BlockBin &bin) {
-  if (!CanMerge(dst) || !CanMerge(src)) {
+  if (!CanMerge(src) || !CanMerge(dst)) {
     return;
   }
 
@@ -316,7 +316,7 @@ size_t CachingAllocator::FreeCachedBlocks() {
   GELOGI("Free cached blocks");
   std::lock_guard<std::recursive_mutex> lock(mutex_);
   size_t free_cached_memory_size = 0;
-  for (uint32_t i = 0; i < kNumBins; ++i) {
+  for (uint32_t i = 0; i < kNumBins; i++) {
     auto pool = free_block_bins_[i];
     if (pool == nullptr) {
       continue;
@@ -324,7 +324,8 @@ size_t CachingAllocator::FreeCachedBlocks() {
     for (auto it = pool->begin(); it != pool->end();) {
       Block *block = *it;
       // free block memory that has not been split
-      if ((block != nullptr) && (block->ptr != nullptr) && (block->prev == nullptr) && (block->next == nullptr) &&
+      if ((block != nullptr) && (block->ptr != nullptr) &&
+          (block->prev == nullptr) && (block->next == nullptr) &&
           (memory_allocator_->FreeMemory(block->ptr) == ge::SUCCESS)) {
         auto itcount = malloced_memory_.find(block->size);
         free_cached_memory_size += block->size;
@@ -345,7 +346,7 @@ size_t CachingAllocator::FreeCachedBlocks() {
 }
 
 void CachingAllocator::FreeBlocks() {
-  GELOGI("Free blocks");
+  GELOGI("Free blocks.");
   std::lock_guard<std::recursive_mutex> lock(mutex_);
   // free allocated blocks and put to cache
   for (auto &it : allocated_blocks_) {
@@ -355,10 +356,18 @@ void CachingAllocator::FreeBlocks() {
   (void) FreeCachedBlocks();
 }
 
+void CachingAllocator::TryFreeBlocks() {
+  GELOGI("Try free blocks.");
+  std::lock_guard<std::recursive_mutex> lock(mutex_);
+  if (allocated_blocks_.empty()) {
+    (void) FreeCachedBlocks();
+  }
+}
+
 void CachingAllocator::FreeBlockBins() {
-  GELOGI("Free block bins");
+  GELOGI("Free block bins.");
   std::lock_guard<std::recursive_mutex> lock(mutex_);
-  for (uint32_t i = 0; i < kNumBins; ++i) {
+  for (uint32_t i = 0; i < kNumBins; i++) {
     if (free_block_bins_[i] != nullptr) {
       delete free_block_bins_[i];
       free_block_bins_[i] = nullptr;
@@ -367,9 +376,9 @@ void CachingAllocator::FreeBlockBins() {
 }
 
 void PrintCount(std::map<size_t, size_t> &count, const std::string &name, size_t total_size, size_t total_count) {
-  GELOGI("%6s total[size:%10zu count:%10zu]", name.c_str(), total_size, total_count);
+  GELOGI("%6s total[size:%10zu count:%10zu].", name.c_str(), total_size, total_count);
   for (auto &it : count) {
-    GELOGI("    |- block[size:%10zu count:%10zu]", it.first, it.second);
+    GELOGI("    |- block[size:%10zu count:%10zu].", it.first, it.second);
   }
 }
 
@@ -383,20 +392,20 @@ void CachingAllocator::PrintStatics() {
   size_t total_free_count = 0;
   size_t total_malloc_size = 0;
   size_t total_malloc_count = 0;
-  std::map<size_t, size_t> using_block;
-  std::map<size_t, size_t> free_block;
-  std::map<size_t, size_t> malloc_block;
+  std::map<size_t, size_t> using_block_stat;
+  std::map<size_t, size_t> free_block_stat;
+  std::map<size_t, size_t> malloc_block_stat;
   do {
     std::lock_guard<std::recursive_mutex> lock(mutex_);
-    for (uint32_t i = 0; i < kNumBins; ++i) {
+    for (uint32_t i = 0; i < kNumBins; i++) {
       auto pool = free_block_bins_[i];
       if (pool == nullptr) {
         continue;
       }
-      for (auto it = pool->begin(); it != pool->end(); ++it) {
+      for (auto it = pool->begin(); it != pool->end(); it++) {
         if ((*it) != nullptr) {
           total_free_size += (*it)->size;
-          IncreaseCount(free_block, (*it)->size);
+          IncreaseCount(free_block_stat, (*it)->size);
           total_free_count++;
         }
       }
@@ -405,7 +414,7 @@ void CachingAllocator::PrintStatics() {
     for (auto &it : allocated_blocks_) {
       if (it.second != nullptr) {
         total_using_size += it.second->size;
-        IncreaseCount(using_block, it.second->size);
+        IncreaseCount(using_block_stat, it.second->size);
         total_using_count++;
       }
     }
@@ -413,12 +422,12 @@ void CachingAllocator::PrintStatics() {
     for (auto &it : malloced_memory_) {
       total_malloc_size += it.first * it.second;
       total_malloc_count += it.second;
-      malloc_block[it.first] = it.second;
+      malloc_block_stat[it.first] = it.second;
     }
   } while (0);
 
-  PrintCount(malloc_block, "Malloc", total_malloc_size, total_malloc_count);
-  PrintCount(using_block, "Using", total_using_size, total_using_count);
-  PrintCount(free_block, "Free", total_free_size, total_free_count);
+  PrintCount(malloc_block_stat, "Malloc", total_malloc_size, total_malloc_count);
+  PrintCount(using_block_stat, "Using", total_using_size, total_using_count);
+  PrintCount(free_block_stat, "Free", total_free_size, total_free_count);
 }
 }  // namespace ge

ge/graph/manager/graph_caching_allocator.h

@@ -94,6 +94,13 @@ class CachingAllocator {
   ///
   Status Free(uint8_t *memory_addr, uint32_t device_id = 0);
 
+  ///
+  /// @ingroup ge_graph
+  /// @brief try to free memory when no memory is referenced
+  /// @return void
+  ///
+  void TryFreeBlocks();
+
  private:
 
   ///

ge/graph/manager/graph_manager.cc

@@ -117,6 +117,10 @@ const char *const kAIcoreEngine = "AIcoreEngine";
 const int32_t kDynamicDimsTypeIsGetNext = 0;
 const int32_t kDynamicDimsTypeIsData = 1;
 const char *const kGetNextName = "IteratorV2";
+const uint32_t kInitGraphCount = 1;
+const uint32_t kNotAdded = 0;
+const uint32_t kStartAdd = 1;
+const uint32_t kDoneAdded = 2;
 
 bool IsTailingOptimization() {
   string is_tailing_optimization_option;
@@ -195,6 +199,8 @@ Status GraphManager::Initialize(const std::map<string, string> &options) {
 
   graph_map_.clear();
   cache_helper_map_.clear();
+  graph_id_to_add_graph_cond_.clear();
+  graph_count_.clear();
   init_flag_ = true;
 
   thread_run_flag_ = true;
@@ -204,6 +210,20 @@ Status GraphManager::Initialize(const std::map<string, string> &options) {
   return SUCCESS;
 }
 
+Status GraphManager::UnloadModel(GeRootModelPtr ge_root_model, uint32_t graph_id) {
+  Status ret = SUCCESS;
+  for (size_t i = 0; i < ge_root_model->GetAllModelId().size(); ++i) {
+    uint32_t model_id = ge_root_model->GetAllModelId()[i];
+    GELOGI("Unload model %u.", model_id);
+    ret = GraphLoader::UnloadModel(model_id);
+    if (ret != SUCCESS) {
+      GELOGW("[GraphManager] unload model failed, modelId=%u, graphId=%u.", model_id, graph_id);
+      return ret;
+    }
+  }
+  return ret;
+}
+
 Status GraphManager::Finalize() {
   if (!init_flag_) {
     GELOGW("GraphManager has not been initialized.");
@@ -234,7 +254,6 @@ Status GraphManager::Finalize() {
       unload_model_ret = GE_GRAPH_GRAPH_IS_RUNNING;
       continue;
     }
-
     // unload model
     auto ge_root_model = graph_node->GetGeRootModel();
     if (ge_root_model != nullptr && ge_root_model->GetModelId() != INVALID_MODEL_ID && graph_node->GetLoadFlag()) {
@@ -244,15 +263,14 @@ Status GraphManager::Finalize() {
         unload_model_ret = FAILED;
         continue;
       }
-      ret = GraphLoader::UnloadModel(ge_root_model->GetModelId());
+      ret = UnloadModel(ge_root_model, iter->first);
       if (ret != SUCCESS) {
-        GELOGW("[GraphManager] unload model failed, modelId=%u, graphId=%u.", ge_root_model->GetModelId(), iter->first);
+        GELOGW("[GraphManager] unload model failed, graph_id=%u.", iter->first);
         unload_model_ret = ret;
       }
       rt_ret = rtDeviceReset(GetContext().DeviceId());
       if (rt_ret != RT_ERROR_NONE) {
-        GELOGW("[GraphManager] rtDeviceReset failed, modelId=%u, graphId=%u.", ge_root_model->GetModelId(),
-               iter->first);
+        GELOGW("[GraphManager] rtDeviceReset failed, graphId=%u.", iter->first);
         unload_model_ret = FAILED;
         continue;
       }
@@ -267,6 +285,7 @@ Status GraphManager::Finalize() {
   }
   graph_map_.clear();
   cache_helper_map_.clear();
+  graph_count_.clear();
 
   // graph context
   if (graph_context_ != nullptr) {
@@ -317,30 +336,59 @@ Status GraphManager::InitDynamicParams(ComputeGraphPtr &compute_graph) {
   return SUCCESS;
 }
 
-Status GraphManager::AddGraph(const GraphId &graph_id, const Graph &graph,
-                              const std::map<std::string, std::string> &options,
-                              const OmgContext &omg_context) {
-  if (HasGraphNode(graph_id)) {
-    GELOGE(GE_GRAPH_GRAPH_ALREADY_EXIST, "[GraphManager] graph exists, graph_id = %u.", graph_id);
-    return GE_GRAPH_GRAPH_ALREADY_EXIST;
+void GraphManager::SetAddGraphCondition(GraphId graph_id, uint32_t cond) {
+  std::lock_guard<std::mutex> lock(add_graph_cond_mutex_);
+  graph_id_to_add_graph_cond_[graph_id] = cond;
+  GELOGD("Graph [id:%u] has been added.", graph_id);
+}
+
+uint32_t GraphManager::GetAddGraphCondition(GraphId graph_id) {
+  std::lock_guard<std::mutex> lock(add_graph_cond_mutex_);
+  auto it = graph_id_to_add_graph_cond_.find(graph_id);
+  if (it != graph_id_to_add_graph_cond_.end()) {
+    return it->second;
+  } else {
+    GELOGD("Graph [id:%u] has not been added.", graph_id);
+    return kNotAdded;
   }
+}
 
-  auto compute_graph = GraphUtils::GetComputeGraph(graph);
-  if (compute_graph != nullptr) {
-    compute_graph->SetGraphID(graph_id);
-    bool graph_has_been_added = false;
-    if (AttrUtils::GetBool(*compute_graph, ATTR_NAME_GRAPH_HAS_BEEN_ADDED, graph_has_been_added)
-        && graph_has_been_added) {
-      GELOGE(GE_GRAPH_GRAPH_ALREADY_EXIST,
-             "[GraphManager] same graph object can not be added again, graph_id = %u.", graph_id);
-      return GE_GRAPH_GRAPH_ALREADY_EXIST;
-    }
-    (void)AttrUtils::SetBool(*compute_graph, ATTR_NAME_GRAPH_HAS_BEEN_ADDED, true);
-    compute_graph_ = compute_graph;
+void GraphManager::RemoveAddGraphCondition(GraphId graph_id) {
+  std::lock_guard<std::mutex> lock(add_graph_cond_mutex_);
+  auto it = graph_id_to_add_graph_cond_.find(graph_id);
+  if (it != graph_id_to_add_graph_cond_.end()) {
+    graph_id_to_add_graph_cond_.erase(it);
+    GELOGD("Successfully removed add_graph_cond of graph [id:%u].", graph_id);
   } else {
-    GELOGE(FAILED, "compute graph is null");
-    return FAILED;
+    GELOGD("Graph [id:%u] has not been added. no need to remove.", graph_id);
+  }
+}
+
+Status GraphManager::CheckRepeatAdd(uint32_t graph_id, bool &is_added) {
+  uint32_t count = 0;
+  if (GetGraphCount(graph_id, count) != SUCCESS) {
+    GELOGE(INTERNAL_ERROR, "Get graph [id:%u] count failed, graph might have not been added.", graph_id);
+    return INTERNAL_ERROR;
   }
+  // previous thread owns same graph_id has been in the middle of the AddGraph procession
+  if (count > 1 && GetAddGraphCondition(graph_id) == kStartAdd) {
+    std::unique_lock<std::mutex> lock(add_graph_mutex_);
+    GELOGD("Waitting for build end of previous thread.");
+    while (GetAddGraphCondition(graph_id) != kDoneAdded) {
+      add_graph_cv_.wait(lock);
+    }
+    GraphNodePtr graph_node;
+    Status ret = GetGraphNode(graph_id, graph_node);
+    if (ret != SUCCESS) {
+      GELOGE(ret, "[AddGraph] GetGraphNode failed, graph_id = %u.", graph_id);
+      return ret;
+    }
+    is_added = true;
+  }
+  return SUCCESS;
+}
+
+void GraphManager::SetSessionGraphId(ComputeGraphPtr compute_graph, uint32_t graph_id) {
   std::string session_graph_id;
   if (!AttrUtils::GetStr(*compute_graph, ATTR_NAME_SESSION_GRAPH_ID, session_graph_id) || session_graph_id.empty()) {
     session_graph_id = "-1_" + to_string(graph_id);
@@ -352,17 +400,92 @@ Status GraphManager::AddGraph(const GraphId &graph_id, const Graph &graph,
     }
     GELOGD("Get graph session_graph_id attr failed, set session id to default value: [0]");
   }
+}
+
+Status GraphManager::NotifyWaittingGraph(uint32_t graph_id) {
+  uint32_t count = 0;
+  if (GetGraphCount(graph_id, count) != SUCCESS) {
+    GELOGE(INTERNAL_ERROR, "Get graph [id:%u] count failed, graph might have not been added.", graph_id);
+    return INTERNAL_ERROR;
+  }
+  GELOGD("Add graph finished, graph_id:%u", graph_id);
+  if (count > 1) {
+    GELOGD("Finish addgraph, graph_id:%u, graph_count:%u, start to notify.", graph_id, count);
+    add_graph_cv_.notify_all();
+  }
+  return SUCCESS;
+}
 
+Status GraphManager::CreateGraphNode(uint32_t graph_id, const Graph &graph,
+                                     const std::map<std::string, std::string> &options) {
   GraphNodePtr graph_node = MakeShared<ge::GraphNode>(graph_id);
   GE_IF_BOOL_EXEC(graph_node == nullptr, GELOGE(FAILED, "GraphNode make shared failed");
                   return FAILED);
   std::shared_ptr<Graph> graph_ptr = MakeShared<ge::Graph>(graph);
   GE_IF_BOOL_EXEC(graph_ptr == nullptr, GELOGE(FAILED, "GraphPtr make shared failed");
                   return FAILED);
-
+  // update option about tuning graph
+  ParseOption(options, BUILD_MODE, options_.build_mode);
+  ParseOption(options, BUILD_STEP, options_.build_step);
+  ParseOption(options, TUNING_PATH, options_.tuning_path);
   graph_node->SetGraph(graph_ptr);
   graph_node->SetOptions(options);
+  graph_node->IncreaseLoadCount();
   AddGraphNode(graph_id, graph_node);
+  return SUCCESS;
+}
+
+Status GraphManager::SetStagesOptions(uint32_t graph_id, const GraphManagerOptions &options) {
+  CompilerStages &stages = GetCompilerStages(graph_id);
+  stages.preparer.SetOptions(options_);
+  Status status = stages.optimizer.SetOptions(options_);
+  if (status != SUCCESS) {
+    GELOGE(status, "Graph optimizer set options failed.");
+    return status;
+  }
+  stages.builder.SetOptions(options_);
+  return SUCCESS;
+}
+
+Status GraphManager::AddGraph(const GraphId &graph_id, const Graph &graph,
+                              const std::map<std::string, std::string> &options,
+                              const OmgContext &omg_context) {
+  IncreaseGraphCount(graph_id);
+  // validation for adding graphs of same graph_id in multi-thread secenario
+  // 1.previous thread owns same graph_id has finished the AddGraph procession
+  if (GetAddGraphCondition(graph_id) == kDoneAdded) {
+    GraphNodePtr graph_node;
+    if (GetGraphNode(graph_id, graph_node) != SUCCESS) {
+      GELOGE(GE_GRAPH_GRAPH_NOT_EXIST, "Graph not exist while done adding previously, graph_id = %u.", graph_id);
+      return GE_GRAPH_GRAPH_NOT_EXIST;
+    }
+    graph_node->IncreaseLoadCount();
+    return SUCCESS;
+  }
+  // In multi-thread scenario, former thread owns same graph_id has been
+  // in the middle of the AddGraph procession while following threads have to wait until
+  // done adding graph of the former graph, avoiding repeatively adding same graph.
+  bool is_added = false;
+  if (CheckRepeatAdd(graph_id, is_added) != SUCCESS) {
+    GELOGE(INTERNAL_ERROR, "CheckRepeatAdd for graph[id:%u] failed.", graph_id);
+    return INTERNAL_ERROR;
+  }
+  // The former graph (from different thread) owns same graph id has been successfully added.
+  if (is_added) {
+    return SUCCESS;
+  }
+  // Do add graph
+  SetAddGraphCondition(graph_id, kStartAdd);
+  auto compute_graph = GraphUtils::GetComputeGraph(graph);
+  GE_CHECK_NOTNULL(compute_graph);
+  compute_graph->SetGraphID(graph_id);
+
+  SetSessionGraphId(compute_graph, graph_id);
+
+  if (CreateGraphNode(graph_id, graph, options) != SUCCESS) {
+    GELOGE(FAILED, "Failed to create graph_node.");
+    return FAILED;
+  }
 
   AddLocalOmgContext(graph_id, omg_context);
   if (!options_.output_datatype.empty()) {
@@ -373,16 +496,18 @@ Status GraphManager::AddGraph(const GraphId &graph_id, const Graph &graph,
     return GRAPH_PARAM_INVALID;
   }
 
-  CompilerStages &stages = GetCompilerStages(graph_id);
-  stages.preparer.SetOptions(options_);
-  Status status = stages.optimizer.SetOptions(options_);
-  if (status != SUCCESS) {
-    GELOGE(status, "Graph optimizer set options failed.");
-    return status;
+  if (SetStagesOptions(graph_id, options_) != SUCCESS) {
+    GELOGE(INTERNAL_ERROR, "Set stage options failed.");
+    return INTERNAL_ERROR;
   }
-  stages.builder.SetOptions(options_);
 
   var_acc_ctrl_.AddGraph(graph_id, compute_graph);
+  SetAddGraphCondition(graph_id, kDoneAdded);
+  // There are threads waitting for adding same graph
+  if (NotifyWaittingGraph(graph_id) != SUCCESS) {
+    GELOGE(INTERNAL_ERROR, "NotifyWaittingGraph failed.");
+    return INTERNAL_ERROR;
+  }
   return SUCCESS;
 }
 
@@ -433,6 +558,10 @@ Status GraphManager::AddGraphWithCopy(const GraphId &graph_id, const Graph &grap
     GELOGE(FAILED, "GraphPtr make shared failed");
     return FAILED;
   }
+  // update option about tuning graph
+  ParseOption(options, BUILD_MODE, options_.build_mode);
+  ParseOption(options, BUILD_STEP, options_.build_step);
+  ParseOption(options, TUNING_PATH, options_.tuning_path);
 
   graph_node->SetGraph(graph_ptr);
   graph_node->SetOptions(options);
@@ -888,6 +1017,7 @@ Status GraphManager::StartForRunGraph(const GraphNodePtr &graph_node, const std:
     if (!graph_node->IsAsync()) {
       ret = LoadGraph(ge_root_model, graph_node);
     } else {
+      GE_CHECK_NOTNULL(ge_root_model);
       ret = LoadGraphAsync(ge_root_model, graph_node);
     }
     if (ret != SUCCESS) {
@@ -902,6 +1032,7 @@ Status GraphManager::StartForRunGraph(const GraphNodePtr &graph_node, const std:
     if (!graph_node->IsAsync()) {
       ret = LoadGraph(ge_root_model_ptr, graph_node);
     } else {
+      GE_CHECK_NOTNULL(ge_root_model);
       ret = LoadGraphAsync(ge_root_model_ptr, graph_node);
     }
     if (ret != SUCCESS) {
@@ -914,6 +1045,7 @@ Status GraphManager::StartForRunGraph(const GraphNodePtr &graph_node, const std:
 Status GraphManager::LoadGraph(const GeRootModelPtr &ge_root_model, const GraphNodePtr &graph_node) {
   GELOGI("[LoadGraph] run_graph_flag[%d], graph_id[%u]", options_.run_graph_flag, graph_node->GetGraphId());
   if (options_.run_graph_flag && ge_root_model != nullptr) {
+    ge_root_model->SetTrainFlag(GetTrainFlag());
     // synchronization run graph with model
     std::shared_ptr<GraphModelListener> model_listener = GetModelListener();
     ModelIdInfo model_id_info;
@@ -1308,54 +1440,29 @@ bool GraphManager::CheckModelLoad(const GeRootModelPtr &ge_root_model, bool load
 }
 
 Status GraphManager::RemoveGraph(const GraphId &graph_id) {
+  auto it = to_be_deleted_graphs_.find(graph_id);
+  if (it != to_be_deleted_graphs_.end()) {
+    to_be_deleted_graphs_.erase(it);
+  }
   GraphNodePtr graph_node = nullptr;
   Status ret = GetGraphNode(graph_id, graph_node);
-  if (ret != SUCCESS) {
+  if (ret != SUCCESS || graph_node == nullptr) {
+    REPORT_INNER_ERROR("E19999", "Graph:%u not exist in graph_map, check invalid when GraphManager %s",
+                       graph_id, __FUNCTION__);
     GELOGE(GE_GRAPH_GRAPH_NOT_EXIST, "[GraphManager] Id %u does not exists.", graph_id);
     return GE_GRAPH_GRAPH_NOT_EXIST;
   }
-
-  if ((graph_node == nullptr) || (graph_node->GetRunFlag())) {
-    GELOGE(GE_GRAPH_GRAPH_IS_RUNNING, "[GraphManager] Id %u is running, can't be deleted.", graph_id);
-    return GE_GRAPH_GRAPH_IS_RUNNING;
+  if (graph_node->GetRunFlag()) {
+    // only put graph into to-be-deleted list when exceptional scenario
+    to_be_deleted_graphs_.insert(graph_id);
+    GELOGI("[GraphManager] Trying to remove running graph[Id:%u], added into to_be_deleted_graphs_.", graph_id);
+    return SUCCESS;
   }
 
   std::lock_guard<std::mutex> lock(unload_model_mutex_);
 
   Status middle_ret;
   rtError_t rt_ret;
-  const std::vector<SubGraphInfoPtr> &all_sub_graph = graph_node->GetAllSubGraph();
-  for (size_t i = 0; i < all_sub_graph.size(); ++i) {
-    // must free buffer firstly
-    middle_ret = all_sub_graph[i]->FreeInOutBuffer();
-    if (middle_ret != SUCCESS) {
-      GELOGE(middle_ret, "[GraphManager] RemoveGraph free mem failed, graph_id=%u.", graph_id);
-      ret = middle_ret;
-    }
-    if (all_sub_graph[i]->GeModelIsValid() && all_sub_graph[i]->GetModelIdInfo().model_id != INVALID_MODEL_ID) {
-      // unload model
-      GELOGI("UnloadModel via new ome.");
-      rt_ret = rtSetDevice(GetContext().DeviceId());
-      if (rt_ret != RT_ERROR_NONE) {
-        GELOGE(RT_FAILED, "[GraphManager:] rtSetDevice failed, modelId=%u, graphId=%u.",
-               all_sub_graph[i]->GetModelIdInfo().model_id, graph_id);
-        ret = FAILED;
-        continue;
-      }
-      middle_ret = GraphLoader::UnloadModel(all_sub_graph[i]->GetModelIdInfo().model_id);
-      if (middle_ret != SUCCESS) {
-        GELOGE(middle_ret, "[GraphManager:] unload model failed, modelId=%u, graph_id=%u.",
-               all_sub_graph[i]->GetModelIdInfo().model_id, graph_id);
-        ret = middle_ret;
-      }
-      rt_ret = rtDeviceReset(GetContext().DeviceId());
-      if (rt_ret != RT_ERROR_NONE) {
-        GELOGE(RT_FAILED, "[GraphManager:] unload model failed, modelId=%u, graphId=%u.",
-               all_sub_graph[i]->GetModelIdInfo().model_id, graph_id);
-        ret = FAILED;
-      }
-    }
-  }
   var_acc_ctrl_.RemoveGraph(graph_id);
   RemoveGraphNode(graph_id);
 
@@ -1363,28 +1470,33 @@ Status GraphManager::RemoveGraph(const GraphId &graph_id) {
 
   auto ge_root_model = graph_node->GetGeRootModel();
   if (CheckModelLoad(ge_root_model, graph_node->GetLoadFlag())) {
-    GELOGI("Unload model %u.", ge_root_model->GetModelId());
     rt_ret = rtSetDevice(GetContext().DeviceId());
     if (rt_ret != RT_ERROR_NONE) {
       GELOGE(RT_FAILED, "[GraphManager:] rtSetDevice failed, modelId=%u, graphId=%u.", ge_root_model->GetModelId(),
              graph_id);
       return FAILED;
     }
-    middle_ret = GraphLoader::UnloadModel(ge_root_model->GetModelId());
+    // same graph may be added for several times, different models were created separately,
+    // unload them respectively.
+    middle_ret = UnloadModel(ge_root_model, graph_id);
     if (middle_ret != SUCCESS) {
-      GELOGE(middle_ret, "[GraphManager:] unload model failed, modelId=%u, graph_id=%u.", ge_root_model->GetModelId(),
-             graph_id);
+      REPORT_INNER_ERROR("E19999", "UnloadModel for graph:%u failed, check unload detail in GraphLoader %s",
+                         graph_id, __FUNCTION__);
+      GELOGE(middle_ret, "[GraphManager:] unload model failed, graph_id=%u.", graph_id);
       ret = middle_ret;
     }
     rt_ret = rtDeviceReset(GetContext().DeviceId());
     if (rt_ret != RT_ERROR_NONE) {
-      GELOGE(RT_FAILED, "[GraphManager:] rtDeviceReset failed, modelId=%u, graphId=%u.", ge_root_model->GetModelId(),
-             graph_id);
+      REPORT_CALL_ERROR("E19999", "Call rtDeviceReset failed, device_id:%u, graph_id:%u, when GraphManager %s",
+                          GetContext().DeviceId(), graph_id, __FUNCTION__);
+      GELOGE(RT_FAILED, "[GraphManager:] rtDeviceReset failed, graphId=%u.", graph_id);
       ret = FAILED;
     }
   }
 
   RemoveCompilerStages(graph_id);
+  RemoveGraphCount(graph_id);
+  RemoveAddGraphCondition(graph_id);
 
   GE_CHK_STATUS_RET(ret, "[GraphManager:] Remove graph failed, graph_id=%u.", graph_id);
   GELOGI("[GraphManager] remove graph success, graph_id=%u.", graph_id);
@@ -1466,6 +1578,10 @@ Status GraphManager::ParseOptions(const std::map<std::string, std::string> &opti
   GE_IF_BOOL_EXEC(ret != SUCCESS,
                   GELOGE(GE_GRAPH_OPTIONS_INVALID, "Key:ge.compressFlag value is invalid, must be 0 or 1.");
                   return GE_GRAPH_OPTIONS_INVALID);
+  // Set Build model and step
+  ParseOption(options, BUILD_MODE, options_.build_mode);
+  ParseOption(options, BUILD_STEP, options_.build_step);
+  ParseOption(options, BUILD_STEP, options_.tuning_path);
 
   // ge.graphType.
   options_.run_graph_flag = true;
@@ -1514,10 +1630,6 @@ Status GraphManager::ParseOptions(const std::map<std::string, std::string> &opti
   GELOGD("Dynamic dims params: input shape is %s, dynamic dims is %s, dynamic node type is %d",
          options_.input_shape.c_str(), options_.dynamic_dims.c_str(), options_.dynamic_node_type);
 
-  // Set Build model and step
-  ParseOption(options, BUILD_MODE, options_.build_mode);
-  ParseOption(options, BUILD_STEP, options_.build_step);
-
   return SUCCESS;
 }
 
@@ -1549,6 +1661,7 @@ void GraphManager::ParseOption(const std::map<std::string, std::string> &options
                                std::string &option) {
   auto iter = options.find(key);
   if (iter != options.end()) {
+    GELOGD("Set option %s from value %s to value%s", key.c_str(), option.c_str(), iter->second.c_str());
     option = iter->second;
   }
 }
@@ -2401,6 +2514,7 @@ void GraphManager::ChangeConstTypeWhenTraining(const ComputeGraphPtr &compute_gr
 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());
   if (options_.run_graph_flag && ge_root_model != nullptr) {
+    ge_root_model->SetTrainFlag(GetTrainFlag());
     // synchronization run graph with model
     ModelIdInfo model_id_info;
     bool is_unknown_shape = false;
@@ -2417,9 +2531,9 @@ Status GraphManager::LoadGraphAsync(const GeRootModelPtr &ge_root_model, const G
       }
     }
     GE_TIMESTAMP_START(LoadGraph);
-    GE_CHECK_NOTNULL(graph_node->graph_run_async_listener_);
-    Status ret =
-        GraphLoader::LoadModelOnline(model_id_info.model_id, ge_root_model, graph_node->graph_run_async_listener_);
+    auto listener = MakeShared<RunAsyncListener>();
+    GE_CHECK_NOTNULL(listener);
+    Status ret = GraphLoader::LoadModelOnline(model_id_info.model_id, ge_root_model, listener);
     GE_TIMESTAMP_EVENT_END(LoadGraph, "GraphManager::LoadGraphAsync");
     if (ret != SUCCESS) {
       GELOGE(ret, "[LoadGraphAsync] LoadGraphAsync Failed");
@@ -2433,6 +2547,52 @@ Status GraphManager::LoadGraphAsync(const GeRootModelPtr &ge_root_model, const G
   return SUCCESS;
 }
 
+void GraphManager::ReleaseMemory(const GeModelPtr &ge_model, GraphNodePtr &graph_node,
+                                 const std::vector<uint32_t> &model_ids, uint32_t graph_id, uint64_t session_id) {
+  rtError_t rt_ret = rtSetDevice(GetContext().DeviceId());
+  if (rt_ret != RT_ERROR_NONE) {
+    REPORT_CALL_ERROR("E19999", "Call rtSetDevice failed, device_id:%u, when GraphManager %s",
+                      GetContext().DeviceId(), __FUNCTION__);
+    GELOGE(RT_FAILED, "[GraphManager:] rtSetDevice failed, graphId=%u.", graph_id);
+    return;
+  }
+  for (auto model_id : model_ids) {
+    uint64_t max_memory_size = 0;
+    Status result = GraphLoader::GetMaxUsedMemory(model_id, max_memory_size);
+    if (result != SUCCESS) {
+      continue;
+    }
+    GELOGI("CheckAndReleaseMemory try to UnloadGraph[%u], model[%u] which MaxUsedMemory[%lu].", graph_id, model_id,
+           max_memory_size);
+    if (model_ids.size() > 1) {
+      result = ge_model->GetSessionId(model_id, session_id);
+      if (result != SUCCESS) {
+        GELOGW("[GraphManager:] get session failed when dynamic memory, modelId=%u, graphId=%u.", model_id,
+               graph_id);
+        continue;
+      }
+    }
+    result = GraphLoader::DestroyAicpuKernel(session_id, model_id, 0);
+    if (result != SUCCESS) {
+      GELOGW("[GraphManager:] destroy aicpu kernel failed when dynamic memory, modelId=%u, graphId=%u.", model_id,
+             graph_id);
+    }
+    result = GraphLoader::UnloadModel(model_id);
+    if (result != SUCCESS) {
+      GELOGW("[GraphManager:] unload model failed, modelId=%u, graphId=%u.", model_id, graph_id);
+    }
+    GELOGI("CheckAndReleaseMemory UnloadGraph[%u], model[%u] success.", graph_id, model_id);
+  }
+  graph_node->SetLoadFlag(false);
+  rt_ret = rtDeviceReset(GetContext().DeviceId());
+  if (rt_ret != RT_ERROR_NONE) {
+    REPORT_CALL_ERROR("E19999", "Call rtDeviceReset failed, device_id:%u, when GraphManager %s",
+                      GetContext().DeviceId(), __FUNCTION__);
+    GELOGE(RT_FAILED, "[GraphManager:] rtDeviceReset failed, graphId=%u.", graph_id);
+    return;
+  }
+}
+
 Status GraphManager::CheckAndReleaseMemory(const GeModelPtr &ge_model, const GraphNodePtr &graph_node) {
   GELOGI("CheckAndReleaseMemory graph_id[%u]", graph_node->GetGraphId());
   int64_t value = 0;
@@ -2476,6 +2636,7 @@ Status GraphManager::CheckAndReleaseMemory(const GeModelPtr &ge_model, const Gra
       continue;
     }
     auto model_id = model->GetModelId();
+    auto model_ids = model->GetAllModelId();
     // unload model not release
     bool is_unknown_shape = false;
     GE_CHK_STATUS_RET(model->CheckIsUnknownShape(is_unknown_shape));
@@ -2488,34 +2649,7 @@ Status GraphManager::CheckAndReleaseMemory(const GeModelPtr &ge_model, const Gra
       GELOGI("CheckAndReleaseMemory graph[%u] has not been loaded.", graph_id);
       continue;
     }
-    uint64_t max_memory_size = 0;
-    result = GraphLoader::GetMaxUsedMemory(model_id, max_memory_size);
-    if (result != SUCCESS) {
-      continue;
-    }
-    GELOGI("CheckAndReleaseMemory try to UnloadGraph[%u], model[%u] which MaxUsedMemory[%lu].", graph_id, model_id,
-           max_memory_size);
-    rtError_t rt_ret = rtSetDevice(GetContext().DeviceId());
-    if (rt_ret != RT_ERROR_NONE) {
-      GELOGE(RT_FAILED, "[GraphManager:] rtSetDevice failed, modelId=%u, graphId=%u.", model_id, graph_id);
-      continue;
-    }
-    result = GraphLoader::DestroyAicpuKernel(session_id, model_id, 0);
-    if (result != SUCCESS) {
-      GELOGW("[GraphManager:] destroy aicpu kernel failed when dynamic memory, modelId=%u, graphId=%u.", model_id,
-             graph_id);
-    }
-    result = GraphLoader::UnloadModel(model_id);
-    if (result != SUCCESS) {
-      GELOGW("[GraphManager:] unload model failed, modelId=%u, graphId=%u.", model_id, graph_id);
-    }
-    rt_ret = rtDeviceReset(GetContext().DeviceId());
-    if (rt_ret != RT_ERROR_NONE) {
-      GELOGE(RT_FAILED, "[GraphManager:] rtDeviceReset failed, modelId=%u, graphId=%u.", model_id, graph_id);
-      continue;
-    }
-    it.second->SetLoadFlag(false);
-    GELOGI("CheckAndReleaseMemory UnloadGraph[%u], model[%u] success and set LoadFlag to false.", graph_id, model_id);
+    ReleaseMemory(ge_model, it.second, model_ids, graph_id, session_id);
   }
 
   return SUCCESS;
@@ -2651,6 +2785,38 @@ void GraphManager::ConstructGeInput(const vector<InputTensorInfo> &inputs, vecto
   }
 }
 
+Status GraphManager::CheckIncreBuildAndPreRun(GraphManager *graph_manager, const PreRunArgs &args,
+                                              GraphNodePtr &graph_node, GeRootModelPtr &ge_root_model) {
+  if (!graph_manager->IsGraphNeedBuild(graph_node)) {
+    ge_root_model = graph_node->GetGeRootModel();
+    return SUCCESS;
+  }
+  if (graph_node->GetBuildFlag()) {
+    ReturnError(graph_manager, args.callback, PARAM_INVALID,
+                "The graph " + std::to_string(graph_node->GetGraphId()) +
+                " need to re-build, you should remove it"
+                " from GE first, then AddGraph again and rebuild it.");
+    graph_node->Unlock();
+    return PARAM_INVALID;
+  }
+  // check need incre build.
+  GeModelPtr ge_model = nullptr;
+  if (graph_manager->IncreBuild(graph_node, ge_model) != SUCCESS) {
+    std::vector<GeTensor> ge_inputs;
+    ConstructGeInput(args.input_tensor, ge_inputs);
+    Status ret = graph_manager->PreRun(graph_node, ge_inputs, ge_root_model, args.session_id);
+    // release rts generate context
+    RtContextUtil::GetInstance().DestroyRtContexts(args.session_id, graph_node->GetGraphId());
+    if (ret != SUCCESS) {
+      ReturnError(graph_manager, args.callback, ret, "PreRun Failed.");
+      return ret;
+    }
+  }
+  graph_node->SetBuildFlag(true);
+  graph_manager->var_acc_ctrl_.SetGraphBuildEnd(graph_node->GetGraphId());
+  return SUCCESS;
+}
+
 void GraphManager::PreRunThread(GraphManager *graph_manager) {
   if (prctl(PR_SET_NAME, ("GE_PreRun")) != 0) {
     GELOGW("Set thread name failed.");
@@ -2663,7 +2829,7 @@ void GraphManager::PreRunThread(GraphManager *graph_manager) {
       continue;
     }
 
-    GELOGI("A new loop start.");
+    GELOGI("[PreRunThread] A new loop start, graph_id:%u.", args.graph_id);
 
     ErrorManager::GetInstance().SetErrorContext(args.error_context);
     ErrorManager::GetInstance().SetStage(ErrorMessage::kModelCompile, ErrorMessage::kOther);
@@ -2679,7 +2845,24 @@ void GraphManager::PreRunThread(GraphManager *graph_manager) {
                   "[RunGraph] graph not exist, graph_id=" + std::to_string(args.graph_id));
       return;
     }
-
+    // more than one graph owns same graph_id
+    uint32_t count = 0;
+    if (graph_manager->GetGraphCount(args.graph_id, count) != SUCCESS) {
+      GELOGE(INTERNAL_ERROR, "Get graph [id:%u] count failed.", args.graph_id);
+      return;
+    }
+    // Avoid repeatively prerun for graphs owns same graph_id in online inference concurrency
+    if (count > 1 && graph_node->GetBuildFlag()) {
+      graph_node->Lock();
+      GELOGD("Avoid repeatively prerun, graph_id:%u.", args.graph_id);
+      // In online inference concurrency senario, graph_node is allowed to be locked for 'count' times
+      graph_node->SetSemSize(count);
+      graph_manager->run_args_q_.Push(RunArgs( { graph_node, args.graph_id, args.session_id, args.error_context,
+          args.input_tensor, graph_node->GetGeRootModel(), GetThreadLocalContext(), args.callback }));
+      GELOGI("[PreRunThread] Loop end. Start to run with cached build model.");
+      continue;
+    }
+    // Cannot be put ahead of the repeatively prerun judgement
     graph_node->Lock();
 
     if (graph_node->GetRunFlag()) {
@@ -2711,46 +2894,24 @@ void GraphManager::PreRunThread(GraphManager *graph_manager) {
     // it will not execute graph preprocess, optimize, parition, build if the graph has built successful.
     GELOGI("Start for run graph async.");
     GeRootModelPtr ge_root_model = nullptr;
-    if (graph_manager->IsGraphNeedBuild(graph_node)) {
-      if (graph_node->GetBuildFlag()) {
-        ReturnError(graph_manager, args.callback, PARAM_INVALID,
-                    "The graph " + std::to_string(graph_node->GetGraphId()) +
-                        " need to re-build, you should remove it"
-                        " from GE first, then AddGraph again and rebuild it.");
+
+    ret = CheckIncreBuildAndPreRun(graph_manager, args, graph_node, ge_root_model);
+    if (ret != SUCCESS) {
+      graph_node->SetRunFlag(false);
+      if (!ge::Analyzer::GetInstance()->IsEnableNetAnalyzeDebug()) {
+        ReturnError(graph_manager, args.callback, ret, "CheckIncreBuildAndPreRun Failed, thread exit..");
         graph_node->Unlock();
         return;
+      } else {
+        ReturnError(graph_manager, graph_node, args.callback, ret,
+                    "CheckIncreBuildAndPreRun Failed, keep geop continue!");
+        graph_node->Unlock();
+        continue;
       }
-
-      // check need incre build.
-      GeModelPtr ge_model = nullptr;
-      if (graph_manager->IncreBuild(graph_node, ge_model) != SUCCESS) {
-        std::vector<GeTensor> ge_inputs;
-        ConstructGeInput(args.input_tensor, ge_inputs);
-        ret = graph_manager->PreRun(graph_node, ge_inputs, ge_root_model, args.session_id);
-        // release rts generate context
-        RtContextUtil::GetInstance().DestroyRtContexts(args.session_id, graph_node->GetGraphId());
-        if (ret != SUCCESS) {
-          graph_node->SetRunFlag(false);
-          if (!ge::Analyzer::GetInstance()->IsEnableNetAnalyzeDebug()) {
-            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);
-      graph_manager->var_acc_ctrl_.SetGraphBuildEnd(graph_node->GetGraphId());
-    } else {
-      ge_root_model = graph_node->GetGeRootModel();
     }
-
     graph_manager->run_args_q_.Push(RunArgs( { graph_node, args.graph_id, args.session_id, args.error_context,
         args.input_tensor, ge_root_model, GetThreadLocalContext(), args.callback }));
-    GELOGI("Loop end.");
+    GELOGI("[PreRunThread] Loop end.");
   }
 }
 
@@ -2847,16 +3008,13 @@ void GraphManager::RunThread(GraphManager *graph_manager) {
       continue;
     }
 
-    GELOGI("A new loop start.");
+    GELOGI("[RunThread] A new loop start, graph_id:%u.", args.graph_id);
 
     ErrorManager::GetInstance().SetErrorContext(args.error_context);
     GetContext().SetSessionId(args.session_id);
     GetThreadLocalContext() = args.context;
     graph_manager->UpdateLocalOmgContext(args.graph_id);
 
-    if (args.graph_node->graph_run_async_listener_ != nullptr) {
-      args.graph_node->graph_run_async_listener_->SetCallback(args.callback);
-    }
     Status ret;
     // parse inputs.dims to vector<vector<uint64_t>> dynamic_dims
     ret = graph_manager->ParseInputsDims(args.input_tensor);
@@ -2866,8 +3024,10 @@ void GraphManager::RunThread(GraphManager *graph_manager) {
       return;
     }
 
+    args.graph_node->UpdateLoadFlag();
     if (!args.graph_node->GetLoadFlag()) {
       ErrorManager::GetInstance().SetStage(ErrorMessage::kModelLoad, ErrorMessage::kModelLoad);
+      args.ge_root_model->SetTrainFlag(graph_manager->GetTrainFlag());
       ret = graph_manager->LoadGraphAsync(args.ge_root_model, args.graph_node);
       if (ret != SUCCESS || args.ge_root_model == nullptr) {
         StopQueue(graph_manager);
@@ -2875,6 +3035,10 @@ void GraphManager::RunThread(GraphManager *graph_manager) {
         args.graph_node->Unlock();
         return;
       }
+      // control the times of graph loading in multi-thread scenario
+      args.graph_node->DecreaseLoadCount();
+      args.graph_node->IncreaseLoadRecord();
+
       args.graph_node->SetLoadFlag(true);
       GELOGI("LoadGraph[%u], model[%u] success and set LoadFlag to true.", args.graph_node->GetGraphId(),
              args.ge_root_model->GetModelId());
@@ -2890,7 +3054,7 @@ void GraphManager::RunThread(GraphManager *graph_manager) {
     }
 
     ret = graph_manager->graph_executor_.ExecuteGraphAsync(args.graph_id, args.graph_node->GetGeRootModel(),
-                                                           args.input_tensor);
+                                                           args.input_tensor, args.callback);
     args.graph_node->SetRunFlag(false);
     if (ret != SUCCESS) {
       ReturnError(graph_manager, args.callback, ret, "ExecuteGraphAsync failed, thread exit.");
@@ -3132,6 +3296,21 @@ Status GraphManager::ConvertGraphToFile(ComputeGraphPtr &compute_graph, GraphPar
       non_tuning_subgraphs.push_back(sub_graph_tmp);
     }
   }
+  // for function graphs to tune
+  for (auto &function_graph : compute_graph->GetAllSubgraphs()) {
+    auto subgraph_list = sub_graph_map[function_graph];
+    for (const auto &sub_graph_info_ptr : subgraph_list) {
+      GE_CHECK_NOTNULL(sub_graph_info_ptr);
+      ComputeGraphPtr sub_graph_tmp = sub_graph_info_ptr->GetSubGraph();
+      // need to tuning
+      if (sub_graph_info_ptr->GetEngineName() == kVectorEngine ||
+          sub_graph_info_ptr->GetEngineName() == kAIcoreEngine) {
+        tuning_subgraphs.push_back(sub_graph_tmp);
+      } else {
+        non_tuning_subgraphs.push_back(sub_graph_tmp);
+      }
+    }
+  }
   return TuningUtils::ConvertGraphToFile(tuning_subgraphs, non_tuning_subgraphs, exe_flag, path);
 }
 
@@ -3291,4 +3470,49 @@ void GraphManager::RemoveCompilerStages(GraphId graph_id) {
   std::lock_guard<std::mutex> lock(member_mutex_);
   compiler_stages_.erase(graph_id);
 }
+
+void GraphManager::IncreaseGraphCount(GraphId graph_id) {
+  std::lock_guard<std::mutex> lock(graph_count_mutex_);
+  auto it = graph_count_.find(graph_id);
+  if (it == graph_count_.end()) {
+    graph_count_.insert({graph_id, kInitGraphCount});
+    GELOGD("After increaseGraphCount, graph count of id[%u] is %u.", graph_id, graph_count_[graph_id]);
+  } else {
+    ++graph_count_[graph_id];
+    GELOGD("After increaseGraphCount, graph count of id[%u] is %u.", graph_id, graph_count_[graph_id]);
+  }
+}
+
+void GraphManager::RemoveGraphCount(GraphId graph_id) {
+  std::lock_guard<std::mutex> lock(graph_count_mutex_);
+  auto it = graph_count_.find(graph_id);
+  if (it == graph_count_.end()) {
+    GELOGW("Graph of id: %u has not been added, count cannot be decreased.", graph_id);
+  } else {
+    GELOGD("RemoveGraphCount success, graph count of id[%u] is %u.", graph_id, graph_count_[graph_id]);
+    graph_count_.erase(it);
+  }
+}
+
+void GraphManager::DecreaseGraphCount(GraphId graph_id) {
+  std::lock_guard<std::mutex> lock(graph_count_mutex_);
+  auto it = graph_count_.find(graph_id);
+  if (it == graph_count_.end()) {
+    GELOGW("Graph of id: %u has not been added, count cannot be decreased.", graph_id);
+  } else {
+    --it->second;
+    GELOGD("After DecreaseGraphCount, graph count of id[%u] is %u.", graph_id, graph_count_[graph_id]);
+  }
+}
+
+Status GraphManager::GetGraphCount(GraphId graph_id, uint32_t &count) {
+  std::lock_guard<std::mutex> lock(graph_count_mutex_);
+  auto it = graph_count_.find(graph_id);
+  if (it == graph_count_.end()) {
+    GELOGW("Graph [id:%u] has not been added.", graph_id);
+    return FAILED;
+  }
+  count = it->second;
+  return SUCCESS;
+}
 }  // namespace ge

ge/graph/manager/graph_manager.h

@@ -184,6 +184,20 @@ class GraphManager {
 
   Status SaveCheckPointResult(const Graph &graph, const std::vector<Tensor> &outputs, map<string, Tensor> &var_results);
 
+  void RemoveGraphCount(GraphId graph_id);
+
+  void IncreaseGraphCount(GraphId graph_id);
+
+  void DecreaseGraphCount(GraphId graph_id);
+
+  Status GetGraphCount(GraphId graph_id, uint32_t &count);
+
+  void SetAddGraphCondition(GraphId graph_id, uint32_t cond);
+
+  uint32_t GetAddGraphCondition(GraphId graph_id);
+
+  void RemoveAddGraphCondition(GraphId graph_id);
+
  private:
   struct CompilerStages {
     GraphPrepare preparer;
@@ -380,6 +394,24 @@ class GraphManager {
   CompilerStages &GetCompilerStages(GraphId graph_id);
   void RemoveCompilerStages(GraphId graph_id);
 
+  static Status CheckIncreBuildAndPreRun(GraphManager *graph_manager, const PreRunArgs &args, GraphNodePtr &graph_node,
+                                         GeRootModelPtr &ge_root_model);
+
+  void ReleaseMemory(const GeModelPtr &ge_model, GraphNodePtr &graph_node, const std::vector<uint32_t> &model_ids,
+                     uint32_t graph_id, uint64_t session_id);
+
+  Status CheckRepeatAdd(uint32_t graph_id, bool &is_added);
+
+  Status NotifyWaittingGraph(uint32_t graph_id);
+
+  Status CreateGraphNode(uint32_t graph_id, const Graph &graph, const std::map<std::string, std::string> &options);
+
+  Status SetStagesOptions(uint32_t graph_id, const GraphManagerOptions &options);
+
+  Status UnloadModel(GeRootModelPtr ge_root_model, uint32_t graph_id);
+
+  void SetSessionGraphId(ComputeGraphPtr compute_graph, uint32_t graph_id);
+
   std::atomic_bool thread_run_flag_;
   BlockingQueue<PreRunArgs> prerun_args_q_{};
   BlockingQueue<RunArgs> run_args_q_{};
@@ -415,6 +447,16 @@ class GraphManager {
 
   std::mutex member_mutex_;
   std::mutex unload_model_mutex_;
+  // avoid repeatively add same graph (owns same graph id)
+  std::mutex add_graph_mutex_;
+  std::mutex add_graph_cond_mutex_;
+  std::condition_variable add_graph_cv_;
+
+  std::map<GraphId, uint32_t> graph_id_to_add_graph_cond_;
+  // use for multi-thread online-infer scenario
+  std::set<GraphId> to_be_deleted_graphs_;
+  std::map<GraphId, uint32_t> graph_count_;
+  std::mutex graph_count_mutex_;
 };
 }  // namespace ge
 

ge/graph/manager/graph_manager_utils.cc

@@ -60,6 +60,15 @@ void GraphNode::Unlock() {
   sem_.Pop(unused);
 }
 
+void GraphNode::IncreaseLoadCount() {
+  std::unique_lock<std::mutex> lock(load_count_mu_);
+  if (load_record_ == kMaxLoadNum) {
+    GELOGW("Reach the maximum of load_count:%u", kMaxLoadNum);
+    return;
+  }
+  ++load_count_;
+}
+
 SubGraphInfo::SubGraphInfo() : subgraph_ptr_(nullptr), ge_model_ptr_(nullptr), malloc_flag_(false) {}
 
 SubGraphInfo::~SubGraphInfo() {

ge/graph/manager/graph_manager_utils.h

@@ -55,6 +55,7 @@ using ConstGraphPtr = std::shared_ptr<const ge::Graph>;
 using GraphPtr = std::shared_ptr<ge::Graph>;
 
 const uint64_t INVALID_SESSION_ID = 0xffffffffffffffffULL;
+const uint32_t kMaxLoadNum = 8;
 
 struct ModelIdInfo {
   uint32_t model_id{INVALID_MODEL_ID};
@@ -162,6 +163,8 @@ class GraphNode {
   bool GetBuildFlag() const { return build_flag_; }
   void SetBuildFlag(bool buildFlag) { build_flag_ = buildFlag; }
   bool GetLoadFlag() const { return load_flag_; }
+  // allow repeatively load graph owns same graph id
+  void UpdateLoadFlag() { load_flag_ = load_count_ == 0 || load_record_ >= kMaxLoadNum; }
   void SetLoadFlag(bool load_flag) { load_flag_ = load_flag; }
   void SetGeModel(const GeModelPtr &ge_model) { ge_model_ = ge_model; }
   GeModelPtr GetGeModel() const { return ge_model_; }
@@ -172,6 +175,13 @@ class GraphNode {
   void Lock();
   void Unlock();
 
+  void SetSemSize(uint32_t size) { sem_.SetMaxSize(size); }
+
+  uint32_t GetLoadCount() const { return load_count_; }
+  void IncreaseLoadCount();
+  void DecreaseLoadCount() { --load_count_; }
+  void IncreaseLoadRecord() { ++load_record_; }
+
   // run graph asynchronous listener
   std::shared_ptr<RunAsyncListener> graph_run_async_listener_;
 
@@ -184,11 +194,17 @@ class GraphNode {
   GraphPtr graph_;
   ComputeGraphPtr compute_graph_;
   bool build_flag_;
+  // load_flag_ is true if more than 1 model were loaded
   bool load_flag_;
   bool async_;
   GeModelPtr ge_model_;
   GeRootModelPtr ge_root_model_;
   BlockingQueue<uint8_t> sem_;
+  // consist with graph_count of same graph_id in graph_manager
+  uint32_t load_count_ = 0;
+  // total times of loading a graph with same graph_id.
+  uint32_t load_record_ = 0;
+  std::mutex load_count_mu_;
 };
 
 using GraphNodePtr = std::shared_ptr<GraphNode>;
@@ -249,6 +265,7 @@ struct GraphManagerOptions {
   std::string save_original_model;
   std::string build_mode;
   std::string build_step;
+  std::string tuning_path;
   std::string input_shape;
   std::string dynamic_dims;
   int32_t dynamic_node_type = -1;
@@ -275,7 +292,8 @@ struct GraphManagerOptions {
         is_single_op(false),
         save_original_model("false"),
         build_mode(""),
-        build_step("") {}
+        build_step(""),
+        tuning_path(""){}
 };
 }  // namespace ge
 

ge/graph/manager/graph_var_manager.cc

@@ -347,14 +347,18 @@ ge::Status VarManager::Init(const uint32_t &version, const uint64_t &session_id,
                             const uint64_t &job_id) {
   std::lock_guard<std::recursive_mutex> lock(mutex_);
   GELOGI("VarManager::Init, session id = %lu.", session_id);
-  version_ = version;
-  device_id_ = device_id;
-  session_id_ = session_id;
-  job_id_ = job_id;
-  var_resource_ = std::unique_ptr<VarResource>(new (std::nothrow) VarResource(session_id_));
   if (var_resource_ == nullptr) {
-    GELOGW("VarManager has not been init.");
-    return ge::INTERNAL_ERROR;
+    version_ = version;
+    device_id_ = device_id;
+    session_id_ = session_id;
+    job_id_ = job_id;
+    var_resource_ = std::unique_ptr<VarResource>(new (std::nothrow) VarResource(session_id_));
+    if (var_resource_ == nullptr) {
+      GELOGW("VarManager init failed session id = %lu.", session_id);
+      return ge::INTERNAL_ERROR;
+    }
+  } else {
+    GELOGW("VarManager::has been inited, session id = %lu.", session_id);
   }
   return SUCCESS;
 }

ge/graph/passes/atomic_addr_clean_pass.cc

@@ -126,11 +126,11 @@ bool AtomicAddrCleanPass::IsOutputIndexPeerInputAtomic(const NodePtr &node, int6
 
 bool AtomicAddrCleanPass::CheckSkipInsertInLoopGraph(const NodePtr &node) {
   OpDescPtr op_desc = node->GetOpDesc();
-  std::map<string, std::map<int, int>> node_workspace_offset;
+  std::map<string, std::map<int64_t, int64_t>> atomic_workspace_index_size;
   bool has_atomic_input = op_desc->HasAttr(ATOMIC_ATTR_INPUT_INDEX);
   bool has_atomic_output = op_desc->HasAttr(ATOMIC_ATTR_OUTPUT_INDEX);
-  node_workspace_offset = op_desc->TryGetExtAttr(EXT_ATTR_ATOMIC_WORKSPACE_OFFSET, node_workspace_offset);
-  if (!has_atomic_input && has_atomic_output && node_workspace_offset.empty()) {
+  atomic_workspace_index_size = op_desc->TryGetExtAttr(EXT_ATTR_ATOMIC_WORKSPACE_INFO, atomic_workspace_index_size);
+  if (!has_atomic_input && has_atomic_output && atomic_workspace_index_size.empty()) {
     std::vector<int64_t> atomic_output_index;
     (void) ge::AttrUtils::GetListInt(op_desc, ATOMIC_ATTR_OUTPUT_INDEX, atomic_output_index);
     bool is_all_output_peer_also_atomic = true;
@@ -222,6 +222,39 @@ Status AtomicAddrCleanPass::HandleNormalGraph(ComputeGraphPtr &graph, const vect
       }
     }
   }
+  return LinkToPotentialPrecedenceNode(graph, clean_addr_node);
+}
+
+// Add control edges from atomic clean node to all potential precedence nodes which may execute before atomic clean
+// node. We hope that atomic clean node can execute with the highest priority in the entire graph. Because of stream
+// concurrency mechanism, only placing it at the head can not ensure that priority. Therefore, we need to add control
+// edges from atomic clean node to the nodes that may be the first node on each stream. Generally, the first nodes on
+// each stream are successors of Data/Variable, and Data/Variable won't generate task or execute, so we link to the
+// successors of Data/Variable.
+Status AtomicAddrCleanPass::LinkToPotentialPrecedenceNode(ComputeGraphPtr &graph, NodePtr &atomic_clean_node) {
+  GELOGD("Start to add control edges from %s to all second-nodes behind first-nodes which have no input.",
+         atomic_clean_node->GetName().c_str());
+  auto out_ctrl_anchor = atomic_clean_node->GetOutControlAnchor();
+  GE_CHECK_NOTNULL(out_ctrl_anchor);
+
+  for (const auto &node : graph->GetDirectNode()) {
+    GE_CHECK_NOTNULL(node);
+    bool need_handle = (node->GetType() == DATA || node->GetType() == VARIABLE) && node->GetInAllNodes().empty();
+    if (!need_handle) {
+      continue;
+    }
+    auto second_nodes = node->GetOutAllNodes();
+    for (const auto &second_node : second_nodes) {
+      GE_CHECK_NOTNULL(second_node);
+      auto in_ctrl_anchor = second_node->GetInControlAnchor();
+      GE_CHECK_NOTNULL(in_ctrl_anchor);
+      if (!out_ctrl_anchor->IsLinkedWith(in_ctrl_anchor)) {
+        GE_CHK_STATUS_RET(out_ctrl_anchor->LinkTo(in_ctrl_anchor));
+        GELOGD("Add control edge from %s to %s.", atomic_clean_node->GetName().c_str(), second_node->GetName().c_str());
+      }
+    }
+  }
+
   return SUCCESS;
 }
 
@@ -332,11 +365,11 @@ bool AtomicAddrCleanPass::IsAtomicOp(const NodePtr &node) {
   }
 
   // 2.Check atomic attr in node
-  std::map<string, std::map<int, int>> node_workspace_offset;
+  std::map<string, std::map<int64_t, int64_t>> atomic_workspace_index_size;
   bool has_atomic_input = op_desc->HasAttr(ATOMIC_ATTR_INPUT_INDEX);
   bool has_atomic_output = op_desc->HasAttr(ATOMIC_ATTR_OUTPUT_INDEX);
-  node_workspace_offset = op_desc->TryGetExtAttr(EXT_ATTR_ATOMIC_WORKSPACE_OFFSET, node_workspace_offset);
-  if (!has_atomic_input && !has_atomic_output && node_workspace_offset.empty()) {
+  atomic_workspace_index_size = op_desc->TryGetExtAttr(EXT_ATTR_ATOMIC_WORKSPACE_INFO, atomic_workspace_index_size);
+  if (!has_atomic_input && !has_atomic_output && atomic_workspace_index_size.empty()) {
     return false;
   }
 

ge/graph/passes/atomic_addr_clean_pass.h

@@ -68,6 +68,14 @@ class AtomicAddrCleanPass : public GraphPass {
   Status LinkToAtomicNode(const NodePtr &atomic_node, NodePtr &atomic_clean_node);
 
   /**
+   * Link atomic clean node to all potential precedence nodes which may execute before atomic clean node
+   * @param graph
+   * @param atomic_clean_node
+   * @return
+   */
+  Status LinkToPotentialPrecedenceNode(ComputeGraphPtr &graph, NodePtr &atomic_clean_node);
+
+  /**
    * Check if this node is atomic op.
    * @param node
    * @return

ge/graph/passes/attach_stream_label_pass.cc

@@ -137,7 +137,6 @@ Status AttachStreamLabelPass::AttachFlag(const NodePtr &node, std::string &strea
       return INTERNAL_ERROR;
     }
     stream_label = node->GetInDataNodes().at(0)->GetName();
-    GE_CHK_STATUS_RET(SetStreamLabel(node, stream_label), "Set stream label failed.");
     bool value = false;
     OpDescPtr op_desc = node->GetOpDesc();
     GE_CHECK_NOTNULL(op_desc);

ge/graph/passes/base_pass.cc

@@ -30,8 +30,15 @@ constexpr int kMaxRePassTimes = 10000;
 constexpr size_t kMaxOneInNodes = 1000;
 // Each iteration, we take about 0.3k memory on the stack, we should change the recursion to loop later
 constexpr int kMaxRecursiveDepth = 20;
+struct DuringPassNodeSets {
+  std::unordered_set<Node *> nodes_seen;
+  std::unordered_set<NodePtr> nodes_deleted;
+  std::unordered_set<NodePtr> nodes_re_pass;
+  std::unordered_set<NodePtr> nodes_re_pass_immediately;
+  std::unordered_set<NodePtr> nodes_last;
+};
 
-void GetAllNodesNoInputEdge(const ComputeGraphPtr &graph, std::queue<NodePtr> &input_edge_nodes,
+void GetAllNodesNoInputEdge(const ComputeGraphPtr &graph, std::deque<NodePtr> &input_edge_nodes,
                             std::unordered_set<Node *> &nodes_seen, std::unordered_set<NodePtr> &nodes_last) {
   nodes_last.clear();
   for (auto &node : graph->GetDirectNode()) {
@@ -40,7 +47,7 @@ void GetAllNodesNoInputEdge(const ComputeGraphPtr &graph, std::queue<NodePtr> &i
     }
     size_t in_nums = node->GetInNodes().size();
     if (in_nums == 0) {
-      input_edge_nodes.push(node);
+      input_edge_nodes.push_back(node);
       nodes_seen.insert(node.get());
     } else if (in_nums > kMaxOneInNodes) {
       nodes_last.insert(node);
@@ -48,7 +55,7 @@ void GetAllNodesNoInputEdge(const ComputeGraphPtr &graph, std::queue<NodePtr> &i
   }
 }
 
-void AddNextIterNodes(const Node::Vistor<NodePtr> &nodes, std::queue<NodePtr> &nodes_to_pass,
+void AddNextIterNodes(const Node::Vistor<NodePtr> &nodes, std::deque<NodePtr> &nodes_to_pass,
                       std::unordered_set<Node *> &nodes_seen, std::unordered_set<NodePtr> &nodes_last) {
   for (auto &node : nodes) {
     if (node == nullptr) {
@@ -60,13 +67,30 @@ void AddNextIterNodes(const Node::Vistor<NodePtr> &nodes, std::queue<NodePtr> &n
 
     bool all_in_nodes_seen = node->IsAllInNodesSeen(nodes_seen);
     if (all_in_nodes_seen && nodes_seen.insert(node.get()).second) {
-      nodes_to_pass.push(node);
+      nodes_to_pass.push_back(node);
     }
   }
 }
 
-Status RunPasses(NodePtr &node, const NamesToPass &names_to_passes, std::unordered_set<NodePtr> &nodes_re_pass,
-                 std::unordered_set<NodePtr> &nodes_deleted, std::unordered_set<Node *> &nodes_seen) {
+void PushToRePassIfSeen(NodePtr &node, const std::pair<std::string, BaseNodePass *> &name_to_pass,
+                        std::unordered_set<Node *> &nodes_seen, std::unordered_set<NodePtr> &nodes_to_re_pass,
+                        std::unordered_set<NodePtr> &nodes_re_pass) {
+  for (const auto &node_to_re_pass : nodes_to_re_pass) {
+    if (node_to_re_pass == nullptr) {
+      GELOGW("Found null re-pass node when executing %s on node %s type %s", name_to_pass.first.c_str(),
+             node->GetName().c_str(), node->GetType().c_str());
+      continue;
+    }
+    if (nodes_seen.count(node_to_re_pass.get()) > 0 || node_to_re_pass->IsAllInNodesSeen(nodes_seen)) {
+      GELOGD("The node %s will be re-pass.", node_to_re_pass->GetName().c_str());
+      nodes_re_pass.insert(node_to_re_pass);
+    } else {
+      GELOGD("The node %s are not all seen, don't set repass this time", node_to_re_pass->GetName().c_str());
+    }
+  }
+}
+
+Status RunPasses(NodePtr &node, const NamesToPass &names_to_passes, DuringPassNodeSets &during_pass_node_set) {
   if (node == nullptr) {
     GELOGE(FAILED, "parameter is null.");
     return FAILED;
@@ -90,22 +114,15 @@ Status RunPasses(NodePtr &node, const NamesToPass &names_to_passes, std::unorder
     }
 
     auto nodes_to_re_pass = name_to_pass.second->GetNodesNeedRePass();
-    for (const auto &node_to_re_pass : nodes_to_re_pass) {
-      if (node_to_re_pass == nullptr) {
-        GELOGW("Found null re-pass node when executing %s on node %s type %s", name_to_pass.first.c_str(),
-               node->GetName().c_str(), node->GetType().c_str());
-        continue;
-      }
-      if (nodes_seen.count(node_to_re_pass.get()) > 0 || node_to_re_pass->IsAllInNodesSeen(nodes_seen)) {
-        GELOGD("The node %s will be re-pass later", node_to_re_pass->GetName().c_str());
-        nodes_re_pass.insert(node_to_re_pass);
-      } else {
-        GELOGD("The node %s are not all seen, don't set repass this time", node_to_re_pass->GetName().c_str());
-      }
-    }
+    PushToRePassIfSeen(node, name_to_pass, during_pass_node_set.nodes_seen, nodes_to_re_pass,
+                       during_pass_node_set.nodes_re_pass);
+
+    auto nodes_to_re_pass_immediately = name_to_pass.second->GetNodesNeedRePassImmediately();
+    PushToRePassIfSeen(node, name_to_pass, during_pass_node_set.nodes_seen, nodes_to_re_pass_immediately,
+                       during_pass_node_set.nodes_re_pass_immediately);
 
     auto nodes_deleted_by_pass = name_to_pass.second->GetNodesDeleted();
-    nodes_deleted.insert(nodes_deleted_by_pass.begin(), nodes_deleted_by_pass.end());
+    during_pass_node_set.nodes_deleted.insert(nodes_deleted_by_pass.begin(), nodes_deleted_by_pass.end());
     if (nodes_deleted_by_pass.count(node) > 0) {
       GELOGD("The node %s was deleted by pass %s, stop the remain passes", node->GetName().c_str(),
              name_to_pass.first.c_str());
@@ -181,36 +198,33 @@ Status GEPass::Run(const NamesToPass &names_to_passes) {
 
 Status GEPass::RunPassesOneGraph(const NamesToPass &names_to_passes) {
   GELOGD("Begin to run pass on graph, passes count %zu", names_to_passes.size());
-  std::queue<NodePtr> nodes;
-  std::unordered_set<Node *> nodes_seen;
-  std::unordered_set<NodePtr> nodes_deleted;
-  std::unordered_set<NodePtr> nodes_re_pass;
-  std::unordered_set<NodePtr> nodes_last;
-  GetAllNodesNoInputEdge(graph_, nodes, nodes_seen, nodes_last);
+  std::deque<NodePtr> nodes;
+  DuringPassNodeSets during_pass_node_set;
+  GetAllNodesNoInputEdge(graph_, nodes, during_pass_node_set.nodes_seen, during_pass_node_set.nodes_last);
   GELOGD("Start points count %zu", nodes.size());
   int re_pass_times = 0;
 
   do {
-    for (auto &node : nodes_re_pass) {
-      nodes.push(node);
-      nodes_seen.insert(node.get());
+    for (auto &node : during_pass_node_set.nodes_re_pass) {
+      nodes.push_back(node);
+      during_pass_node_set.nodes_seen.insert(node.get());
     }
-    nodes_re_pass.clear();
+    during_pass_node_set.nodes_re_pass.clear();
 
     while (!nodes.empty()) {
       NodePtr node = nodes.front();
-      nodes.pop();
+      nodes.pop_front();
 
-      (void)nodes_re_pass.erase(node);
+      (void)during_pass_node_set.nodes_re_pass.erase(node);
       GE_IF_BOOL_EXEC(node == nullptr, GELOGW("node is null"); continue);
-      if (nodes_deleted.count(node) > 0) {
+      if (during_pass_node_set.nodes_deleted.count(node) > 0) {
         GELOGD("The node %s was deleted before, skip it.", node->GetName().c_str());
         continue;
       }
 
-      AddNextIterNodes(node->GetOutNodes(), nodes, nodes_seen, nodes_last);
+      AddNextIterNodes(node->GetOutNodes(), nodes, during_pass_node_set.nodes_seen, during_pass_node_set.nodes_last);
 
-      auto ret = RunPasses(node, names_to_passes, nodes_re_pass, nodes_deleted, nodes_seen);
+      auto ret = RunPasses(node, names_to_passes, during_pass_node_set);
       if (ret != SUCCESS) {
         GELOGE(ret, "Failed to process passes on node %s type %s, error code: %u",
                node->GetName().c_str(), node->GetType().c_str(), ret);
@@ -227,7 +241,7 @@ Status GEPass::RunPassesOneGraph(const NamesToPass &names_to_passes) {
       if (has_sub_graph) {
         GELOGD("There are subgraphs on node %s, run passes for for the second time", node->GetName().c_str());
         SetFlagOption(kOptimizeAfterSubGraph, names_to_passes);
-        ret = RunPasses(node, names_to_passes, nodes_re_pass, nodes_deleted, nodes_seen);
+        ret = RunPasses(node, names_to_passes, during_pass_node_set);
         if (ret != SUCCESS) {
           GELOGE(ret, "Failed to process passes on node %s type %s, error code: %u",
                  node->GetName().c_str(), node->GetType().c_str(), ret);
@@ -239,16 +253,21 @@ Status GEPass::RunPassesOneGraph(const NamesToPass &names_to_passes) {
         // should be called each time at the begin of the iteration
         ClearOption(names_to_passes);
       }
+      for (const auto &node : during_pass_node_set.nodes_re_pass_immediately) {
+        GELOGD("The node %s will be re-pass immediately.", node->GetName().c_str());
+        nodes.push_front(node);
+      }
+      during_pass_node_set.nodes_re_pass_immediately.clear();
     }
 
-    for (auto &node : nodes_last) {
-      bool all_in_nodes_seen = node->IsAllInNodesSeen(nodes_seen);
-      if (all_in_nodes_seen && nodes_seen.insert(node.get()).second) {
-        nodes.push(node);
+    for (auto &node : during_pass_node_set.nodes_last) {
+      bool all_in_nodes_seen = node->IsAllInNodesSeen(during_pass_node_set.nodes_seen);
+      if (all_in_nodes_seen && during_pass_node_set.nodes_seen.insert(node.get()).second) {
+        nodes.push_back(node);
       }
     }
-    nodes_last.clear();
-  } while ((!nodes_re_pass.empty() || !nodes.empty()) && ++re_pass_times < kMaxRePassTimes);
+    during_pass_node_set.nodes_last.clear();
+  } while ((!during_pass_node_set.nodes_re_pass.empty() || !nodes.empty()) && ++re_pass_times < kMaxRePassTimes);
 
   if (re_pass_times == kMaxRePassTimes) {
     GELOGW("re_pass_times should not come to %d", kMaxRePassTimes);

ge/graph/passes/base_pass.h

@@ -53,6 +53,8 @@ class BaseNodePass {
 
   std::unordered_set<NodePtr> GetNodesNeedRePass() { return nodes_need_re_pass_; }
 
+  std::unordered_set<NodePtr> GetNodesNeedRePassImmediately() { return nodes_need_re_pass_immediately_; }
+
   std::unordered_set<NodePtr> GetNodesDeleted() { return nodes_deleted_; }
 
   void SetOption(NodePassOption option, const std::string &value) { options_[option] = value; }
@@ -62,6 +64,7 @@ class BaseNodePass {
   void init() {
     nodes_need_re_pass_.clear();
     nodes_deleted_.clear();
+    nodes_need_re_pass_immediately_.clear();
   }
 
  protected:
@@ -80,6 +83,14 @@ class BaseNodePass {
   void AddRePassNode(NodePtr &node) { nodes_need_re_pass_.insert(node); }
 
   ///
+  /// Add a node to be optimized immediately again. If you add a new node to the graph, or
+  /// change a node connections, and you want to make sure the node will be
+  /// optimized by other passes, call this function.
+  /// @param node
+  ///
+  void AddImmediateRePassNode(NodePtr &node) { nodes_need_re_pass_immediately_.insert(node); }
+
+  ///
   /// Add a node and it's input/output data nodes to be optimized again.
   /// @param node
   ///
@@ -109,6 +120,7 @@ class BaseNodePass {
 
  private:
   std::unordered_set<NodePtr> nodes_need_re_pass_;
+  std::unordered_set<NodePtr> nodes_need_re_pass_immediately_;
   std::unordered_set<NodePtr> nodes_deleted_;
   std::map<NodePassOption, std::string> options_;
 };

ge/graph/passes/infershape_pass.cc

@@ -25,6 +25,7 @@
 
 namespace ge {
 Status InferShapePass::Run(NodePtr &node) {
+  // kOptimizeAfterSubGraph exist means after subgraph
   auto ret = ShapeRefiner::InferShapeAndType(node, !OptionExists(kOptimizeAfterSubGraph));
   if (ret != GRAPH_SUCCESS) {
     // select INFERSHAPE failed info
@@ -41,6 +42,20 @@ Status InferShapePass::Run(NodePtr &node) {
     GELOGE(GE_GRAPH_INFERSHAPE_FAILED, "infershape failed. node: %s", node->GetName().c_str());
     return GE_GRAPH_INFERSHAPE_FAILED;
   }
+  bool need_repass = false;
+  auto has_attr = AttrUtils::GetBool(node->GetOpDesc(), "need_infer_again_", need_repass);
+  if (has_attr) {
+    if (!OptionExists(kOptimizeAfterSubGraph)) {
+      return SUCCESS;
+    }
+    if (need_repass) {
+      AddImmediateRePassNode(node);
+      GELOGD("Node %s need repass immediately.", node->GetName().c_str());
+    } else {
+      // clear attr on while
+      node->GetOpDesc()->DelAttr("need_infer_again_");
+    }
+  }
   return SUCCESS;
 }
 }  // namespace ge

ge/graph/passes/merge_input_memcpy_pass.cc

@@ -23,7 +23,9 @@ namespace ge {
 Status MergeInputMemcpyPass::Run(ComputeGraphPtr graph) {
   GELOGD("MergeInputMemcpyPass Enter");
   for (const auto &node : graph->GetDirectNode()) {
-    if ((node->GetType() != MERGE) && (node->GetType() != REFMERGE)) {
+    std::string type;
+    GE_CHK_STATUS_RET(GetOriginalType(node, type), "Get node type failed.");
+    if ((type != MERGE) && (type != REFMERGE)) {
       continue;
     }
     GE_CHECK_NOTNULL(node->GetOpDesc());
@@ -95,4 +97,3 @@ NodePtr MergeInputMemcpyPass::CreateMemcpyAsyncNode(const ComputeGraphPtr &graph
   return graph->AddNode(op_desc);
 }
 }  // namespace ge
-

ge/graph/passes/merge_to_stream_merge_pass.cc

@@ -25,7 +25,9 @@ Status MergeToStreamMergePass::Run(ComputeGraphPtr graph) {
 
   bypass_nodes_.clear();
   for (const auto &node : graph->GetDirectNode()) {
-    if ((node->GetType() != MERGE) && (node->GetType() != REFMERGE)) {
+    std::string type;
+    GE_CHK_STATUS_RET(GetOriginalType(node, type), "Get node type failed.");
+    if ((type != MERGE) && (type != REFMERGE)) {
       continue;
     }
 

ge/graph/passes/net_output_pass.cc

@@ -555,6 +555,8 @@ void NetOutputPass::AddInOutForNetOutputOp(const ComputeGraphPtr &graph, OpDescP
       return;
     }
     ge::GeTensorDesc out_desc = src_node->GetOpDesc()->GetOutputDesc(src_index);
+    out_desc.SetFormat(FORMAT_ND);
+    out_desc.SetOriginFormat(FORMAT_ND);
     GE_IF_BOOL_EXEC(net_output_desc->AddInputDesc(out_desc) != SUCCESS, GELOGW("add input desc failed"); return );
     is_input_const.push_back(PassUtils::IsConstant(src_node));
     ++iter;

ge/graph/passes/next_iteration_pass.cc

@@ -101,7 +101,8 @@ Status NextIterationPass::FindWhileGroups() {
     const std::string &frame_name = loop_group_iter.first;
     for (const auto &enter_node : loop_group_iter.second->enter_nodes) {
       for (const auto &out_node : enter_node->GetOutAllNodes()) {
-        const string &type = out_node->GetType();
+        std::string type;
+        GE_CHK_STATUS_RET(GetOriginalType(out_node, type), "Get node type failed.");
         if ((type != MERGE) && (type != REFMERGE)) {
           continue;
         }
@@ -310,7 +311,8 @@ Status NextIterationPass::FindTargetNode(const NodePtr &node, const std::string
   }
 
   for (const auto &tmp_node : nodes) {
-    const std::string type = tmp_node->GetType();
+    std::string type;
+    GE_CHK_STATUS_RET(GetOriginalType(tmp_node, type), "Get node type failed.");
     if ((target_type == LOOPCOND) && (type == target_type)) {
       target_node = tmp_node;
       break;

ge/graph/passes/pass_utils.cc

@@ -35,9 +35,9 @@
 #include "graph/utils/op_desc_utils.h"
 #include "graph/utils/tensor_utils.h"
 #include "graph/utils/type_utils.h"
+#include "utils/node_utils.h"
 
 namespace ge {
-
 Status PassUtils::ConstructTensorDescWithData(const GeTensorDesc &out_desc, std::vector<int64_t> &data,
                                               std::vector<GeTensorPtr> &v_output, const bool scalar_output) {
   Status ret = SUCCESS;
@@ -246,6 +246,12 @@ NodePtr PassUtils::GetInDataNode(const ConstNodePtr &node, int index) {
   return src_node;
 }
 
+NodePtr PassUtils::GetInNodeCrossSubgraphByIndex(const ConstNodePtr &node, int index) {
+  auto src_node = GetInDataNode(node, index);
+
+  return NodeUtils::GetInNodeCrossSubgraph(src_node);
+}
+
 bool PassUtils::IsNeedTrainIteFlowCtrl(const ComputeGraphPtr &compute_graph) {
   if (compute_graph == nullptr) {
     return false;

ge/graph/passes/pass_utils.h

@@ -30,6 +30,8 @@ class PassUtils {
 
   static NodePtr GetInDataNode(const ConstNodePtr &node, int index);
 
+  static NodePtr GetInNodeCrossSubgraphByIndex(const ConstNodePtr &node, int index);
+
   static bool IsConstant(const ConstNodePtr &node);
 
   static Status SetOutNodeWeight(const OutDataAnchorPtr &out_data_anchor, const NodePtr &src_node);

ge/graph/passes/subexpression_migration_pass.cc

@@ -279,7 +279,7 @@ Status SubexpressionMigrationPass::GraphNodeMigration(const ComputeGraphPtr &gra
       const auto &in_anchor = in_anchors.at(i);
       const auto &base_node = in_anchor->GetOwnerNode();
       GELOGD("Get Data direct node: %s", base_node->GetName().c_str());
-      if (!base_node->GetHostNode()) {
+      if (!base_node->GetHostNode() || base_node->GetType() == SWITCH) {
         continue;
       }
 

ge/graph/passes/switch_dead_branch_elimination.cc

@@ -94,6 +94,12 @@ Status SwitchDeadBranchElimination::DeleteSwitchNode(NodePtr &node, NodePtr &pre
     GELOGE(FAILED, "parameter is null.");
     return FAILED;
   }
+
+  // If two nodes aren't in same graph, get node's direct in_node instead of pred_node.
+  if (node->GetOwnerComputeGraph() != pred_node->GetOwnerComputeGraph()) {
+    pred_node = PassUtils::GetInDataNode(node, kPredInputIndex);
+  }
+
   // link pred's in control nodes to switch
   if (GraphUtils::CopyInCtrlEdges(pred_node, node) != GRAPH_SUCCESS) {
     return FAILED;
@@ -131,7 +137,7 @@ Status SwitchDeadBranchElimination::Run(NodePtr &node) {
     return SUCCESS;
   }
 
-  auto pred_node = PassUtils::GetInDataNode(node, kPredInputIndex);
+  auto pred_node = PassUtils::GetInNodeCrossSubgraphByIndex(node, kPredInputIndex);
   if (pred_node == nullptr) {
     GELOGD("[%s] Pred input is null.", node->GetName().c_str());
     return SUCCESS;
@@ -143,7 +149,7 @@ Status SwitchDeadBranchElimination::Run(NodePtr &node) {
     return SUCCESS;
   }
 
-  auto input_node = PassUtils::GetInDataNode(node, kDataInputIndex);
+  auto input_node = PassUtils::GetInNodeCrossSubgraphByIndex(node, kDataInputIndex);
   if (input_node == nullptr) {
     GELOGD("[%s] Data input is null.", node->GetName().c_str());
     return SUCCESS;

ge/graph/passes/switch_to_stream_switch_pass.cc

@@ -448,6 +448,8 @@ Status SwitchToStreamSwitchPass::CombineSwitchNode(const ComputeGraphPtr &graph)
 
         // select first stream_switch
         NodePtr stream_switch = switch_list.front();
+        // set stream_label
+        GE_CHK_STATUS_RET(SetStreamLabel(stream_switch, cast_node->GetName()), "Set stream label failed.");
         OpDescPtr switch_desc = stream_switch->GetOpDesc();
         GE_CHECK_NOTNULL(switch_desc);
         switch_desc->SetName(CheckDuplicateName(cond_group + "/" + STREAMSWITCH + (true_branch_flag ? "_t" : "_f")));

ge/graph/preprocess/graph_preprocess.cc

@@ -23,6 +23,7 @@
 #include "common/formats/format_transfers/format_transfer_nhwc_nc1hwc0.h"
 #include "common/formats/format_transfers/format_transfer_transpose.h"
 #include "common/formats/utils/formats_trans_utils.h"
+#include "common/util/error_manager/error_manager.h"
 #include "common/helper/model_helper.h"
 #include "common/math/math_util.h"
 #include "common/op/ge_op_utils.h"
@@ -1304,7 +1305,8 @@ Status GraphPrepare::UpdateInput(const std::vector<GeTensor> &user_input,
       auto format = desc.GetFormat();
       auto origin_format = desc.GetOriginFormat();
       // data maybe internal format [FRACTAL_NZ] at singleop process such as GEMM.
-      bool need_check_internal_format = (!IsTansDataOpData(input_node)) && (!options_.is_single_op);
+      auto tune_flag = (options_.build_mode == BUILD_MODE_TUNING) && (options_.build_step == BUILD_STEP_AFTER_BUILDER);
+      bool need_check_internal_format = (!IsTansDataOpData(input_node)) && (!options_.is_single_op) && (!tune_flag);
       if (need_check_internal_format) {
         bool is_internal = TypeUtils::IsInternalFormat(format) || TypeUtils::IsInternalFormat(origin_format);
         if (is_internal) {
@@ -1346,19 +1348,22 @@ Status GraphPrepare::UpdateInput(const std::vector<GeTensor> &user_input,
         return FAILED;
       }
       ge::TensorUtils::SetSize(desc, shape_size);
-      graphStatus graph_ret = op->UpdateInputDesc(0, desc);
-      if (graph_ret != GRAPH_SUCCESS) {
-        GELOGE(graph_ret, "UpdateInputDesc fail, graph_ret:%u", graph_ret);
-        return graph_ret;
-      }
-      // Size will be recalculated in the build stage
-      ge::TensorUtils::SetSize(desc, 0);
-      graph_ret = op->UpdateOutputDesc(0, desc);
-      if (graph_ret != GRAPH_SUCCESS) {
-        GELOGE(graph_ret, "UpdateOutputDesc fail, graph_ret:%u", graph_ret);
-        return graph_ret;
+      if (!tune_flag) {
+        graphStatus graph_ret = op->UpdateInputDesc(0, desc);
+        if (graph_ret != GRAPH_SUCCESS) {
+          GELOGE(graph_ret, "UpdateInputDesc fail, graph_ret:%u", graph_ret);
+          return graph_ret;
+        }
+        // Size will be recalculated in the build stage
+        ge::TensorUtils::SetSize(desc, 0);
+        graph_ret = op->UpdateOutputDesc(0, desc);
+        if (graph_ret != GRAPH_SUCCESS) {
+          GELOGE(graph_ret, "UpdateOutputDesc fail, graph_ret:%u", graph_ret);
+          return graph_ret;
+        }
+      } else {
+        GELOGI("data %s skip update info in tune mode", op->GetName().c_str());
       }
-
       if (!dynamic_shape_range_vec.empty()) {
         ret = UpdateDynamicInputShapeRange(index, dynamic_shape_range_vec, op, desc);
         GE_CHK_STATUS_RET(ret, "Fail to update dynamic input shape range on %s.", op->GetName().c_str());
@@ -1763,13 +1768,13 @@ Status GraphPrepare::CheckUserInput(const std::vector<GeTensor> &user_input) {
       GeTensorDesc desc(user_input[index].GetTensorDesc());
 
       for (size_t i = 0; i < desc.GetShape().GetDimNum(); ++i) {
-        if (desc.GetShape().GetDim(i) < 0) {
-          std::string situation = "data dim[" + std::to_string(i) + "][" +
-                  std::to_string(desc.GetShape().GetDim(i)) + "]" ;
-          std::string reason = "it need >= 0";
-          ErrorManager::GetInstance().ATCReportErrMessage("E19025", {"situation", "reason"}, {situation, reason});
-          GELOGE(GE_GRAPH_INIT_FAILED, "data dim %zu is not supported, need >= 0, real:%ld.", i,
-                 desc.GetShape().GetDim(i));
+        int64_t dim = desc.GetShape().GetDim(i);
+        if (dim < UNKNOWN_DIM_NUM) {
+          std::string situation = "data dim[" + std::to_string(i) + "][" + std::to_string(dim) + "]" ;
+          std::string reason = "it need >= -2";
+          REPORT_INPUT_ERROR("E19025", std::vector<std::string>({"situation", "reason"}),
+                             std::vector<std::string>({situation, reason}));
+          GELOGE(GE_GRAPH_INIT_FAILED, "[Check][InputDim]data dim %zu is not supported, need >= -2, real:%ld.", i, dim);
           return GE_GRAPH_INIT_FAILED;
         }
       }

ge/graph/preprocess/insert_op/ge_aipp_op.cc

@@ -428,7 +428,8 @@ Status AippOp::ConvertRelatedInputNameToRank() {
   if (!convert_flag) {
     string error_msg = "Top name " + related_input_name + "convert rank failed, Please"
                        " ensure top name in aipp config is the top name of data node.";
-    GE_ERRORLOG_AND_ERRORMSG(PARAM_INVALID, error_msg.c_str());
+    GELOGE(PARAM_INVALID, "[Check][InputParam]%s", error_msg.c_str());
+    REPORT_INPUT_ERROR("E19021", std::vector<std::string>({"reason"}), std::vector<std::string>({error_msg}));
     return PARAM_INVALID;
   }
 

ge/graph/preprocess/insert_op/util_insert_aipp_op.cc

@@ -124,13 +124,15 @@ Status InsertNewOpUtil::CheckInputNamePositionNotRepeat() {
       if (another_item->related_input_name().empty()) {
         string error_msg = "Can not both set related_input_name and related_input_rank!"
                            " Please ensure param is the same with the first aipp config(related_input_name).";
-        GE_ERRORLOG_AND_ERRORMSG(PARAM_INVALID, error_msg.c_str());
+        GELOGE(PARAM_INVALID, "[Check][InputParam]%s", error_msg.c_str());
+        REPORT_INPUT_ERROR("E19021", std::vector<std::string>({"reason"}), std::vector<std::string>({error_msg}));
         return PARAM_INVALID;
       }
       if (item->related_input_name() == another_item->related_input_name()) {
         string error_msg = "Can not insert aipp to the same postion! Please ensure related_input_name"
                            " param is different in different aipp config.";
-        GE_ERRORLOG_AND_ERRORMSG(PARAM_INVALID, error_msg.c_str());
+        GELOGE(PARAM_INVALID, "[Check][InputParam]%s", error_msg.c_str());
+        REPORT_INPUT_ERROR("E19021", std::vector<std::string>({"reason"}), std::vector<std::string>({error_msg}));
         return PARAM_INVALID;
       }
     }
@@ -150,13 +152,15 @@ Status InsertNewOpUtil::CheckInputRankPositionNoRepeat() {
       if (!another_item->related_input_name().empty()) {
         string error_msg = "Can not both set related_input_rank and related_input_name!"
                            " Please ensure param is the same with the first aipp config(related_input_rank).";
-        GE_ERRORLOG_AND_ERRORMSG(PARAM_INVALID, error_msg.c_str());
+        GELOGE(PARAM_INVALID, "[Check][InputParam]%s", error_msg.c_str());
+        REPORT_INPUT_ERROR("E19021", std::vector<std::string>({"reason"}), std::vector<std::string>({error_msg}));
         return PARAM_INVALID;
       }
       if (item->related_input_rank() == another_item->related_input_rank()) {
         string error_msg = "Can not insert aipp to the same postion! Please ensure related_input_rank"
                           " param is different in different aipp config.";
-        GE_ERRORLOG_AND_ERRORMSG(PARAM_INVALID, error_msg.c_str());
+        GELOGE(PARAM_INVALID, "[Check][InputParam]%s", error_msg.c_str());
+        REPORT_INPUT_ERROR("E19021", std::vector<std::string>({"reason"}), std::vector<std::string>({error_msg}));
         return PARAM_INVALID;
       }
     }
@@ -212,7 +216,7 @@ Status InsertNewOpUtil::CheckGraph(const ComputeGraphPtr &graph) {
         }
       }
     }
-    GE_CHK_LOG_AND_ERRORMSG((aippNodes.size() == 0) || (aippNodes.size() == next_nodes_cnt), 
+    GE_CHK_LOG_AND_ERRORMSG((aippNodes.size() == 0) || (aippNodes.size() == next_nodes_cnt),
         PARAM_INVALID,
         "Can not config part of outputs of Data node to support AIPP, config all "
         "of the outputs of Data to support AIPP, or config none of them");

ge/host_cpu_engine/CMakeLists.txt

@@ -3,6 +3,7 @@ set(PROTO_LIST
 )
 
 protobuf_generate(ge PROTO_SRCS PROTO_HDRS ${PROTO_LIST})
+protobuf_generate(ge_atcstub PROTO_ATCSTUB_SRCS PROTO_ATCSTUB_HDRS ${PROTO_LIST})
 
 set(SRC_LIST
     "engine/host_cpu_engine.cc"
@@ -61,7 +62,7 @@ target_link_libraries(host_cpu_engine PRIVATE
 )
 
 ############ atcstub/libhost_cpu_engine.so ############
-add_library(atc_host_cpu_engine SHARED ${SRC_LIST} ${PROTO_HDRS})
+add_library(atc_host_cpu_engine SHARED ${SRC_LIST} ${PROTO_ATCSTUB_HDRS})
 
 target_compile_options(atc_host_cpu_engine PRIVATE
     -Werror
@@ -84,7 +85,7 @@ target_include_directories(atc_host_cpu_engine PRIVATE
     ${METADEF_DIR}/inc/external
     ${METADEF_DIR}/inc/external/graph
     ${CMAKE_BINARY_DIR}
-    ${CMAKE_BINARY_DIR}/proto/ge
+    ${CMAKE_BINARY_DIR}/proto/ge_atcstub
     #### yellow zone ####
     ${GE_CODE_DIR}/../inc
     #### blue zone ####

ge/host_kernels/concat_offset_kernel.cc

@@ -33,7 +33,7 @@ const int kNumOne = 1;
 }  // namespace
 Status ConcatOffsetKernel::Compute(const OpDescPtr op_desc_ptr, const vector<ConstGeTensorPtr> &input,
                                    vector<GeTensorPtr> &v_output) {
-  GELOGI("ConcatOffsetKernel in.");
+  GELOGD("ConcatOffsetKernel in");
   if (op_desc_ptr == nullptr) {
     GELOGE(PARAM_INVALID, "input opdesc is nullptr.");
     return PARAM_INVALID;
@@ -41,7 +41,7 @@ Status ConcatOffsetKernel::Compute(const OpDescPtr op_desc_ptr, const vector<Con
   // validate attrs
   int N = 0;
   if (!(AttrUtils::GetInt(op_desc_ptr, "N", N))) {
-    GELOGW("Attr %s does not exist.", "N");
+    GELOGW("Attr %s does not exist", "N");
     return NOT_CHANGED;
   }
   // follow IR def, the first input is concat_dim
@@ -50,7 +50,7 @@ Status ConcatOffsetKernel::Compute(const OpDescPtr op_desc_ptr, const vector<Con
   int32_t concat_dim = *(const_cast<int32_t *>(reinterpret_cast<const int32_t *>(input_0->GetData().data())));
   // validate inputs
   if ((static_cast<int>(input.size()) != (N + kNumOne)) || (input.size() <= kConcatOffsetInputIndexOne)) {
-    GELOGW("The number of input for concat offset must be equal to %d, and must be more than one.", (N + kNumOne));
+    GELOGW("The number of input for concat offset must be equal to %d, and must be more than one", (N + kNumOne));
     return NOT_CHANGED;
   }
 
@@ -61,7 +61,7 @@ Status ConcatOffsetKernel::Compute(const OpDescPtr op_desc_ptr, const vector<Con
     GELOGW("Concat dim is bigger than the size of output_shape.");
     return NOT_CHANGED;
   }
-  GELOGI("Output shape size is %ld", output_size);
+  GELOGI("Output shape size is %ld.", output_size);
   int32_t offset = 0;
   if (output_size < 0) {
     GELOGE(FAILED, "Index is negative.");
@@ -86,7 +86,7 @@ Status ConcatOffsetKernel::Compute(const OpDescPtr op_desc_ptr, const vector<Con
     output_ptr->MutableTensorDesc().SetShape(output_shape);
     GE_IF_BOOL_EXEC(output_ptr->SetData(reinterpret_cast<uint8_t *>(buf.get()),
                                         static_cast<size_t>(sizeof(DT_INT32) * output_size)) != GRAPH_SUCCESS,
-                    GELOGW("set data failed");
+                    GELOGW("set data failed.");
                     return NOT_CHANGED);
     v_output.push_back(output_ptr);
     // caculate offset
@@ -99,7 +99,7 @@ Status ConcatOffsetKernel::Compute(const OpDescPtr op_desc_ptr, const vector<Con
     }
     offset += input_dim;
   }
-  GELOGI("ConcatOffsetKernel success.");
+  GELOGD("ConcatOffsetKernel success");
   return SUCCESS;
 }
 REGISTER_KERNEL(CONCATOFFSET, ConcatOffsetKernel);

ge/host_kernels/gather_v2_kernel.cc

@@ -278,7 +278,7 @@ Status GatherV2Kernel::SaveIndicesByDataType(ConstGeTensorPtr indices_tensor_ptr
     auto indices_ptr = const_cast<int32_t *>(reinterpret_cast<const int32_t *>(indices_tensor_ptr->GetData().data()));
     for (int64_t i = 0; i < indices_shape.GetShapeSize(); i++) {
       if (*(indices_ptr + i) < 0 || *(indices_ptr + i) >= x_shape.GetDim(axis)) {
-        GELOGW("indices %ld value is not in range [0, %ld)", i, x_shape.GetDim(axis));
+        GELOGW("indices %ld value is not in range [0, %ld).", i, x_shape.GetDim(axis));
         return NOT_CHANGED;
       }
       indicates_.push_back(*(indices_ptr + i));
@@ -288,7 +288,7 @@ Status GatherV2Kernel::SaveIndicesByDataType(ConstGeTensorPtr indices_tensor_ptr
     auto indices_ptr = const_cast<int64_t *>(reinterpret_cast<const int64_t *>(indices_tensor_ptr->GetData().data()));
     for (int64_t i = 0; i < indices_shape.GetShapeSize(); i++) {
       if (*(indices_ptr + i) < 0 || *(indices_ptr + i) >= x_shape.GetDim(axis)) {
-        GELOGW("indices %ld value is not in range [0, %ld)", i, x_shape.GetDim(axis));
+        GELOGW("indices %ld value is not in range [0, %ld).", i, x_shape.GetDim(axis));
         return NOT_CHANGED;
       }
       indicates_.push_back(*(indices_ptr + i));
@@ -344,42 +344,42 @@ Status GatherV2Kernel::Check(const OpDescPtr &op_desc_ptr, const vector<ConstGeT
   auto indices_data_type = tensor1->GetTensorDesc().GetDataType();
   bool is_valid_indices_data_type = indices_data_type == DT_INT32 || indices_data_type == DT_INT64;
   if (!is_valid_indices_data_type) {
-    GELOGW("indices datatype must be DT_INT32 or DT_INT64");
+    GELOGW("indices datatype must be DT_INT32 or DT_INT64.");
     return NOT_CHANGED;
   }
   if (indices_shape.GetDimNum() > kMaxIndicatesDims) {
-    GELOGW("indices input only support 0 or 1 dims");
+    GELOGW("indices input only support 0 or 1 dims.");
     return NOT_CHANGED;
   }
   return SUCCESS;
 }
 void GatherV2Kernel::DebugPrint(int64_t axis, const GeShape &x_shape, const GeShape &indices_shape,
                                 const std::vector<int64_t> &y_shape) {
-  GELOGD("GatherV2Kernel axis:%ld x_shape:%zu indices_shape:%zu y_shape:%zu", axis, x_shape.GetDimNum(),
+  GELOGD("GatherV2Kernel axis:%ld x_shape:%zu indices_shape:%zu y_shape:%zu.", axis, x_shape.GetDimNum(),
          indices_shape.GetDimNum(), y_shape.size());
   for (size_t i = 0; i < x_shape.GetDimNum(); i++) {
-    GELOGD("GatherV2Kernel x_shape[%zu]: %ld", i, x_shape.GetDim(i));
+    GELOGD("GatherV2Kernel x_shape[%zu]: %ld.", i, x_shape.GetDim(i));
   }
   for (size_t i = 0; i < indices_shape.GetDimNum(); i++) {
-    GELOGD("GatherV2Kernel indices_shape[%zu]: %ld", i, indices_shape.GetDim(i));
+    GELOGD("GatherV2Kernel indices_shape[%zu]: %ld.", i, indices_shape.GetDim(i));
   }
   for (size_t i = 0; i < y_shape.size(); i++) {
-    GELOGD("GatherV2Kernel y_shape[%zu]: %ld", i, y_shape[i]);
+    GELOGD("GatherV2Kernel y_shape[%zu]: %ld.", i, y_shape[i]);
   }
   for (auto ele : indicates_) {
-    GELOGD("GatherV2Kernel indices:%ld", ele);
+    GELOGD("GatherV2Kernel indices:%ld.", ele);
   }
 }
 
 Status GatherV2Kernel::Compute(const OpDescPtr op_desc_ptr, const vector<ConstGeTensorPtr> &input,
                                vector<GeTensorPtr> &v_output) {
-  GELOGI("Enter GatherV2Kernel Process.");
+  GELOGI("Enter GatherV2Kernel Process");
   Status ret = Check(op_desc_ptr, input, v_output);
   if (ret != SUCCESS) {
-    GELOGW("param check failed.");
+    GELOGW("param check failed");
     return NOT_CHANGED;
   }
-  GELOGI("GatherV2Kernel[%s] start Process.", op_desc_ptr->GetName().c_str());
+  GELOGI("GatherV2Kernel[%s] start Process", op_desc_ptr->GetName().c_str());
   ConstGeTensorPtr tensor0 = input.at(kGatherV2InputIndexZero);
   ConstGeTensorPtr tensor1 = input.at(kGatherV2InputIndexOne);
   ConstGeTensorPtr tensor2 = input.at(kGatherV2InputIndexTwo);
@@ -394,7 +394,7 @@ Status GatherV2Kernel::Compute(const OpDescPtr op_desc_ptr, const vector<ConstGe
   axis = axis >= 0 ? axis : axis + x_shape.GetDimNum();
   // check axis value
   if (axis < 0 || (axis + 1) > static_cast<int64_t>(x_shape.GetDimNum())) {
-    GELOGW("axis is invalid");
+    GELOGW("axis is invalid!");
     return NOT_CHANGED;
   }
   auto indices_data_type = tensor1->GetTensorDesc().GetDataType();
@@ -407,7 +407,8 @@ Status GatherV2Kernel::Compute(const OpDescPtr op_desc_ptr, const vector<ConstGe
   // check input data type
   auto x_data_type = tensor0->GetTensorDesc().GetDataType();
   if (supported_type.find(x_data_type) == supported_type.end()) {
-    GELOGI("GatherV2Kernel does not support this Data type:%s", TypeUtils::DataTypeToSerialString(x_data_type).c_str());
+    GELOGI("GatherV2Kernel does not support this Data type:%s.",
+           TypeUtils::DataTypeToSerialString(x_data_type).c_str());
     return NOT_CHANGED;
   }
   // calc output shape

ge/host_kernels/identity_kernel.cc

@@ -61,4 +61,5 @@ Status IdentityKernel::Compute(const ge::OpDescPtr op_desc, const std::vector<ge
   return SUCCESS;
 }
 REGISTER_KERNEL(IDENTITY, IdentityKernel);
+REGISTER_KERNEL(PLACEHOLDERWITHDEFAULT, IdentityKernel);
 }  // namespace ge

ge/host_kernels/strided_slice_kernel.cc

@@ -84,14 +84,14 @@ void GetOriginStrideVec(const std::vector<ge::ConstGeTensorPtr> &input, vector<i
 }  // namespace
 Status StridedSliceKernel::Compute(const ge::OpDescPtr attr, const std::vector<ge::ConstGeTensorPtr> &input,
                                    vector<ge::GeTensorPtr> &v_output) {
-  GELOGD("StridedSliceKernel in.");
+  GELOGD("StridedSliceKernel in");
   // 1.Check input and attrs
   if (CheckAndGetAttr(attr) != SUCCESS) {
-    GELOGW("Check and get attrs failed.Ignore kernel.");
+    GELOGW("Check and get attrs failed.Ignore kernel");
     return NOT_CHANGED;
   }
   if (CheckInputParam(input) != SUCCESS) {
-    GELOGW("Check input params failed.Ignore kernel.");
+    GELOGW("Check input params failed.Ignore kernel");
     return NOT_CHANGED;
   }
   // 2.Init param with mask attrs.
@@ -120,7 +120,7 @@ Status StridedSliceKernel::Compute(const ge::OpDescPtr attr, const std::vector<g
   auto ret = OpUtils::SetOutputSliceData(data, static_cast<int64_t>(data_size), data_type, input_dims, begin_vec,
                                          output_dims, output_ptr.get(), stride_vec);
   if (ret != SUCCESS) {
-    GELOGE(INTERNAL_ERROR, "SetOutputSliceData failed.");
+    GELOGE(INTERNAL_ERROR, "SetOutputSliceData failed");
     return NOT_CHANGED;
   }
 
@@ -133,7 +133,7 @@ Status StridedSliceKernel::Compute(const ge::OpDescPtr attr, const std::vector<g
   GetOutputDims(final_dim_size, output_dims, v_dims);
   t_d.SetShape(GeShape(v_dims));
   v_output.push_back(output_ptr);
-  GELOGI("StridedSliceKernel success.");
+  GELOGI("StridedSliceKernel success");
   return SUCCESS;
 }
 Status StridedSliceKernel::CheckAndGetAttr(const OpDescPtr &attr) {
@@ -144,7 +144,7 @@ Status StridedSliceKernel::CheckAndGetAttr(const OpDescPtr &attr) {
   // Get all op attr value of strided_slice
   for (auto &attr_2_value : attr_value_map_) {
     if (!AttrUtils::GetInt(attr, attr_2_value.first, attr_2_value.second)) {
-      GELOGE(PARAM_INVALID, "Get %s attr failed.", attr_2_value.first.c_str());
+      GELOGE(PARAM_INVALID, "Get %s attr failed", attr_2_value.first.c_str());
       return PARAM_INVALID;
     }
   }
@@ -182,7 +182,7 @@ Status StridedSliceKernel::CheckInputParam(const std::vector<ConstGeTensorPtr> &
     return PARAM_INVALID;
   }
   if (kIndexNumberType.find(begin_tensor_desc.GetDataType()) == kIndexNumberType.end()) {
-    GELOGW("Data type of StridedSlice OP(begin,end,strides) must be int32 or int64.");
+    GELOGW("Data type of StridedSlice OP(begin,end,strides) must be int32 or int64");
     return PARAM_INVALID;
   }
 
@@ -250,7 +250,7 @@ Status StridedSliceKernel::InitParamWithAttrs(const std::vector<ConstGeTensorPtr
       end_i = x_dims.at(i);
       stride_i = 1;
     }
-    GELOGD("Before mask calculate. Begin is : %ld\t,end is : %ld\t stride is : %ld\t x_dim_i is : %ld.",
+    GELOGD("Before mask calculate. Begin is : %ld\t,end is : %ld\t stride is : %ld\t x_dim_i is : %ld",
            begin_i, end_i, stride_i, x_dims.at(i));
     auto ret = MaskCal(i, begin_i, end_i, x_dims.at(i));
     if (ret != SUCCESS) {
@@ -258,7 +258,7 @@ Status StridedSliceKernel::InitParamWithAttrs(const std::vector<ConstGeTensorPtr
       return NOT_CHANGED;
     }
     int64_t dim_final;
-    GELOGD("Before stride calculate. Begin is : %ld\t,end is : %ld\t stride is : %ld\t x_dim_i is : %ld.",
+    GELOGD("Before stride calculate. Begin is : %ld\t,end is : %ld\t stride is : %ld\t x_dim_i is : %ld",
            begin_i, end_i, stride_i, x_dims.at(i));
     (void) StrideCal(x_dims.at(i), begin_i, end_i, stride_i, dim_final);
     output_dims.push_back(dim_final);