Merge branch 'development' of gitee.com:mindspore/graphengine into development

build.sh

@@ -134,11 +134,7 @@ build_graphengine()
   mk_dir "${BUILD_PATH}"
   cd "${BUILD_PATH}"
 
-  if [[ "X$MINDSPORE_MODE" = "Xoff" ]]; then
-    CMAKE_ARGS="${CMAKE_ARGS} -DENABLE_OPEN_SRC=True -DCMAKE_INSTALL_PREFIX=${OUTPUT_PATH} -DPLATFORM=${PLATFORM} -DPRODUCT=${PRODUCT}"
-  else
-    CMAKE_ARGS="${CMAKE_ARGS} -DENABLE_D=ON -DCMAKE_INSTALL_PREFIX=${OUTPUT_PATH}"
-  fi
+  CMAKE_ARGS="-DBUILD_PATH=$BUILD_PATH"
 
   if [[ "X$ENABLE_GE_COV" = "Xon" ]]; then
     CMAKE_ARGS="${CMAKE_ARGS} -DENABLE_GE_COV=ON"
@@ -156,7 +152,13 @@ build_graphengine()
   if [[ "X$ENABLE_GITEE" = "Xon" ]]; then
     CMAKE_ARGS="${CMAKE_ARGS} -DENABLE_GITEE=ON"
   fi
-  CMAKE_ARGS="${CMAKE_ARGS} -DENABLE_OPEN_SRC=True -DCMAKE_INSTALL_PREFIX=${OUTPUT_PATH} -DPLATFORM=${PLATFORM} -DPRODUCT=${PRODUCT}"
+
+  if [[ "X$MINDSPORE_MODE" = "Xoff" ]]; then
+    CMAKE_ARGS="${CMAKE_ARGS} -DENABLE_OPEN_SRC=True -DCMAKE_INSTALL_PREFIX=${OUTPUT_PATH} -DPLATFORM=${PLATFORM} -DPRODUCT=${PRODUCT}"
+  else
+    CMAKE_ARGS="${CMAKE_ARGS} -DENABLE_D=ON -DCMAKE_INSTALL_PREFIX=${OUTPUT_PATH}"
+  fi
+
   echo "${CMAKE_ARGS}"
   cmake ${CMAKE_ARGS} ..
   if [ $? -ne 0 ]
@@ -233,14 +235,14 @@ if [[ "X$ENABLE_GE_UT" = "Xon" || "X$ENABLE_GE_COV" = "Xon" ]]; then
 #     fi
 
 #     if [[ "X$ENABLE_GE_COV" = "Xon" ]]; then
-         echo "Generating coverage statistics, please wait..."
-         cd ${BASEPATH}
-         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 --remove cov/tmp.info '*/output/*' '*/build/opensrc/*' '*/build/proto/*' '*/third_party/*' '*/tests/*' '/usr/local/*' -o cov/coverage.info
-	 cd ${BASEPATH}/cov
-	 genhtml coverage.info
+        echo "Generating coverage statistics, please wait..."
+        cd ${BASEPATH}
+        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
+        cd ${BASEPATH}/cov
+        genhtml coverage.info
 fi
 
 # generate output package in tar form, including ut/st libraries/executables

ge/CMakeLists.txt

@@ -35,6 +35,7 @@ protobuf_generate(ge PROTO_HEADER_SRCS PROTO_HEADER_HDRS ${PROTO_HEADER_LIST})
 if (NOT ENABLE_D AND NOT ENABLE_ACL AND NOT ENABLE_MS_TESTCASES)
 ############ libge_proto_common.a ############
 add_library(ge_proto_common STATIC
+    ${PROTO_HEADER_HDRS}
     ${PROTO_SRCS}
 )
 
@@ -55,6 +56,7 @@ target_link_libraries(ge_proto_common PRIVATE
 
 ############ libge_proto_client.a ############
 add_library(ge_proto_client STATIC
+    ${PROTO_HEADER_HDRS}
     ${PROTO_CLIENT_SRCS}
 )
 
@@ -127,38 +129,38 @@ set(TRAIN_SRC_LIST
     "graph/label/partitioned_call_label_maker.cc"
     "graph/label/while_label_maker.cc"
     "graph/load/graph_loader.cc"
-    "graph/load/new_model_manager/cpu_queue_schedule.cc"
-    "graph/load/new_model_manager/data_dumper.cc"
-    "graph/load/new_model_manager/data_inputer.cc"
-    "graph/load/new_model_manager/davinci_model.cc"
-    "graph/load/new_model_manager/davinci_model_parser.cc"
-    "graph/load/new_model_manager/model_manager.cc"
-    "graph/load/new_model_manager/model_utils.cc"
-    "graph/load/new_model_manager/aipp_utils.cc"
-    "graph/load/new_model_manager/task_info/end_graph_task_info.cc"
-    "graph/load/new_model_manager/task_info/model_exit_task_info.cc"
-    "graph/load/new_model_manager/task_info/event_record_task_info.cc"
-    "graph/load/new_model_manager/task_info/event_wait_task_info.cc"
-    "graph/load/new_model_manager/task_info/fusion_start_task_info.cc"
-    "graph/load/new_model_manager/task_info/fusion_stop_task_info.cc"
-    "graph/load/new_model_manager/task_info/hccl_task_info.cc"
-    "graph/load/new_model_manager/task_info/kernel_ex_task_info.cc"
-    "graph/load/new_model_manager/task_info/kernel_task_info.cc"
-    "graph/load/new_model_manager/task_info/label_set_task_info.cc"
-    "graph/load/new_model_manager/task_info/label_switch_by_index_task_info.cc"
-    "graph/load/new_model_manager/task_info/label_goto_ex_task_info.cc"
-    "graph/load/new_model_manager/task_info/memcpy_addr_async_task_info.cc"
-    "graph/load/new_model_manager/task_info/memcpy_async_task_info.cc"
-    "graph/load/new_model_manager/task_info/profiler_trace_task_info.cc"
-    "graph/load/new_model_manager/task_info/stream_active_task_info.cc"
-    "graph/load/new_model_manager/task_info/stream_switch_task_info.cc"
-    "graph/load/new_model_manager/task_info/stream_switchn_task_info.cc"
-    "graph/load/new_model_manager/task_info/super_kernel/super_kernel.cc"
-    "graph/load/new_model_manager/task_info/super_kernel/super_kernel_factory.cc"
-    "graph/load/new_model_manager/task_info/task_info.cc"
-    "graph/load/new_model_manager/tbe_handle_store.cc"
-    "graph/load/new_model_manager/zero_copy_task.cc"
-    "graph/load/new_model_manager/zero_copy_offset.cc"
+    "graph/load/model_manager/cpu_queue_schedule.cc"
+    "graph/load/model_manager/data_dumper.cc"
+    "graph/load/model_manager/data_inputer.cc"
+    "graph/load/model_manager/davinci_model.cc"
+    "graph/load/model_manager/davinci_model_parser.cc"
+    "graph/load/model_manager/model_manager.cc"
+    "graph/load/model_manager/model_utils.cc"
+    "graph/load/model_manager/aipp_utils.cc"
+    "graph/load/model_manager/task_info/end_graph_task_info.cc"
+    "graph/load/model_manager/task_info/model_exit_task_info.cc"
+    "graph/load/model_manager/task_info/event_record_task_info.cc"
+    "graph/load/model_manager/task_info/event_wait_task_info.cc"
+    "graph/load/model_manager/task_info/fusion_start_task_info.cc"
+    "graph/load/model_manager/task_info/fusion_stop_task_info.cc"
+    "graph/load/model_manager/task_info/hccl_task_info.cc"
+    "graph/load/model_manager/task_info/kernel_ex_task_info.cc"
+    "graph/load/model_manager/task_info/kernel_task_info.cc"
+    "graph/load/model_manager/task_info/label_set_task_info.cc"
+    "graph/load/model_manager/task_info/label_switch_by_index_task_info.cc"
+    "graph/load/model_manager/task_info/label_goto_ex_task_info.cc"
+    "graph/load/model_manager/task_info/memcpy_addr_async_task_info.cc"
+    "graph/load/model_manager/task_info/memcpy_async_task_info.cc"
+    "graph/load/model_manager/task_info/profiler_trace_task_info.cc"
+    "graph/load/model_manager/task_info/stream_active_task_info.cc"
+    "graph/load/model_manager/task_info/stream_switch_task_info.cc"
+    "graph/load/model_manager/task_info/stream_switchn_task_info.cc"
+    "graph/load/model_manager/task_info/super_kernel/super_kernel.cc"
+    "graph/load/model_manager/task_info/super_kernel/super_kernel_factory.cc"
+    "graph/load/model_manager/task_info/task_info.cc"
+    "graph/load/model_manager/tbe_handle_store.cc"
+    "graph/load/model_manager/zero_copy_task.cc"
+    "graph/load/model_manager/zero_copy_offset.cc"
     "graph/manager/graph_context.cc"
     "graph/manager/graph_manager.cc"
     "graph/manager/graph_manager_utils.cc"
@@ -200,6 +202,7 @@ set(TRAIN_SRC_LIST
     "graph/passes/compile_nodes_pass.cc"
     "graph/passes/constant_folding_pass.cc"
     "graph/passes/constant_fuse_same_pass.cc"
+    "graph/passes/fuse_data_nodes_with_common_input_pass.cc"
     "graph/passes/remove_same_const_pass.cc"
     "graph/passes/useless_control_out_remove_pass.cc"
     "graph/passes/control_trigger_pass.cc"
@@ -372,6 +375,7 @@ set(TRAIN_SRC_LIST
     "hybrid/node_executor/host_cpu/kernel/variable_kernel.cc"
     "hybrid/node_executor/host_cpu/kernel/assign_kernel.cc"
     "hybrid/node_executor/host_cpu/kernel/random_uniform_kernel.cc"
+    "hybrid/node_executor/host_cpu/kernel/data_kernel.cc"
     "hybrid/node_executor/controlop/control_op_executor.cc"
     "hybrid/node_executor/partitioned_call/partitioned_call_node_executor.cc"
     "hybrid/node_executor/hccl/hccl_node_executor.cc"
@@ -482,6 +486,7 @@ set(INFER_SRC_LIST
     "graph/passes/net_output_pass.cc"
     "graph/passes/replace_transshape_pass.cc"
     "graph/passes/constant_fuse_same_pass.cc"
+    "graph/passes/fuse_data_nodes_with_common_input_pass.cc"
     "graph/passes/print_op_pass.cc"
     "graph/passes/no_use_reshape_remove_pass.cc"
     "graph/passes/iterator_op_pass.cc"
@@ -601,37 +606,37 @@ set(INFER_SRC_LIST
     "graph/manager/util/rt_context_util.cc"
     "graph/manager/util/variable_accelerate_ctrl.cc"
     "graph/manager/util/debug.cc"
-    "graph/load/new_model_manager/model_manager.cc"
-    "graph/load/new_model_manager/data_inputer.cc"
-    "graph/load/new_model_manager/davinci_model.cc"
-    "graph/load/new_model_manager/davinci_model_parser.cc"
-    "graph/load/new_model_manager/model_utils.cc"
-    "graph/load/new_model_manager/aipp_utils.cc"
-    "graph/load/new_model_manager/tbe_handle_store.cc"
-    "graph/load/new_model_manager/cpu_queue_schedule.cc"
-    "graph/load/new_model_manager/zero_copy_task.cc"
-    "graph/load/new_model_manager/zero_copy_offset.cc"
-    "graph/load/new_model_manager/data_dumper.cc"
-    "graph/load/new_model_manager/task_info/task_info.cc"
-    "graph/load/new_model_manager/task_info/event_record_task_info.cc"
-    "graph/load/new_model_manager/task_info/event_wait_task_info.cc"
-    "graph/load/new_model_manager/task_info/fusion_start_task_info.cc"
-    "graph/load/new_model_manager/task_info/fusion_stop_task_info.cc"
-    "graph/load/new_model_manager/task_info/kernel_ex_task_info.cc"
-    "graph/load/new_model_manager/task_info/kernel_task_info.cc"
-    "graph/load/new_model_manager/task_info/label_set_task_info.cc"
-    "graph/load/new_model_manager/task_info/label_switch_by_index_task_info.cc"
-    "graph/load/new_model_manager/task_info/label_goto_ex_task_info.cc"
-    "graph/load/new_model_manager/task_info/memcpy_async_task_info.cc"
-    "graph/load/new_model_manager/task_info/memcpy_addr_async_task_info.cc"
-    "graph/load/new_model_manager/task_info/profiler_trace_task_info.cc"
-    "graph/load/new_model_manager/task_info/stream_active_task_info.cc"
-    "graph/load/new_model_manager/task_info/stream_switch_task_info.cc"
-    "graph/load/new_model_manager/task_info/stream_switchn_task_info.cc"
-    "graph/load/new_model_manager/task_info/end_graph_task_info.cc"
-    "graph/load/new_model_manager/task_info/model_exit_task_info.cc"
-    "graph/load/new_model_manager/task_info/super_kernel/super_kernel_factory.cc"
-    "graph/load/new_model_manager/task_info/super_kernel/super_kernel.cc"
+    "graph/load/model_manager/model_manager.cc"
+    "graph/load/model_manager/data_inputer.cc"
+    "graph/load/model_manager/davinci_model.cc"
+    "graph/load/model_manager/davinci_model_parser.cc"
+    "graph/load/model_manager/model_utils.cc"
+    "graph/load/model_manager/aipp_utils.cc"
+    "graph/load/model_manager/tbe_handle_store.cc"
+    "graph/load/model_manager/cpu_queue_schedule.cc"
+    "graph/load/model_manager/zero_copy_task.cc"
+    "graph/load/model_manager/zero_copy_offset.cc"
+    "graph/load/model_manager/data_dumper.cc"
+    "graph/load/model_manager/task_info/task_info.cc"
+    "graph/load/model_manager/task_info/event_record_task_info.cc"
+    "graph/load/model_manager/task_info/event_wait_task_info.cc"
+    "graph/load/model_manager/task_info/fusion_start_task_info.cc"
+    "graph/load/model_manager/task_info/fusion_stop_task_info.cc"
+    "graph/load/model_manager/task_info/kernel_ex_task_info.cc"
+    "graph/load/model_manager/task_info/kernel_task_info.cc"
+    "graph/load/model_manager/task_info/label_set_task_info.cc"
+    "graph/load/model_manager/task_info/label_switch_by_index_task_info.cc"
+    "graph/load/model_manager/task_info/label_goto_ex_task_info.cc"
+    "graph/load/model_manager/task_info/memcpy_async_task_info.cc"
+    "graph/load/model_manager/task_info/memcpy_addr_async_task_info.cc"
+    "graph/load/model_manager/task_info/profiler_trace_task_info.cc"
+    "graph/load/model_manager/task_info/stream_active_task_info.cc"
+    "graph/load/model_manager/task_info/stream_switch_task_info.cc"
+    "graph/load/model_manager/task_info/stream_switchn_task_info.cc"
+    "graph/load/model_manager/task_info/end_graph_task_info.cc"
+    "graph/load/model_manager/task_info/model_exit_task_info.cc"
+    "graph/load/model_manager/task_info/super_kernel/super_kernel_factory.cc"
+    "graph/load/model_manager/task_info/super_kernel/super_kernel.cc"
     "single_op/task/op_task.cc"
     "single_op/task/build_task_utils.cc"
     "single_op/task/tbe_task_builder.cc"

ge/common/CMakeLists.txt

@@ -187,6 +187,8 @@ target_compile_options(ge_common PRIVATE
     -fvisibility=hidden
     -O2
     -Werror
+    -Wno-deprecated-declarations
+    -fno-common
 )
 
 target_include_directories(ge_common PRIVATE

ge/common/helper/model_cache_helper.cc

@@ -28,7 +28,7 @@
 #include "framework/common/util.h"
 #include "graph/detail/attributes_holder.h"
 #include "graph/detail/model_serialize_imp.h"
-#include "graph/load/new_model_manager/davinci_model_parser.h"
+#include "graph/load/model_manager/davinci_model_parser.h"
 #include "graph/model.h"
 #include "graph/utils/graph_utils.h"
 #include "graph/utils/tensor_utils.h"

ge/common/helper/model_helper.cc

@@ -23,7 +23,7 @@
 #include "framework/common/debug/ge_log.h"
 #include "framework/omg/version.h"
 #include "graph/debug/ge_attr_define.h"
-#include "graph/load/new_model_manager/davinci_model_parser.h"
+#include "graph/load/model_manager/davinci_model_parser.h"
 #include "graph/utils/attr_utils.h"
 #include "graph/utils/graph_utils.h"
 

ge/common/profiling/profiling_manager.cc

@@ -21,7 +21,7 @@
 #include "framework/common/string_util.h"
 #include "graph/ge_context.h"
 #include "runtime/base.h"
-#include "graph/load/new_model_manager/davinci_model.h"
+#include "graph/load/model_manager/davinci_model.h"
 
 namespace {
 const char *const kTrainingTrace = "training_trace";
@@ -218,6 +218,7 @@ FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY void ProfilingManager::Profilin
     uint32_t stream_id = task.stream_id;
     std::string shape_type = task.shape_type;
     int64_t cur_iter_num = task.cur_iter_num;
+    uint32_t task_type = task.task_type;
     data = model_name.append(" ")
                      .append(op_name).append(" ")
                      .append(std::to_string(block_dim)).append(" ")
@@ -225,7 +226,8 @@ FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY void ProfilingManager::Profilin
                      .append(std::to_string(stream_id)).append(" ")
                      .append(std::to_string(model_id)).append(" ")
                      .append(shape_type).append(" ")
-                     .append(std::to_string(cur_iter_num)).append("\n");
+                     .append(std::to_string(cur_iter_num)).append(" ")
+                     .append(std::to_string(task_type)).append("\n");
 
     ReporterData reporter_data{};
     reporter_data.deviceId = device_id;

ge/common/types.cc

@@ -388,6 +388,7 @@ REGISTER_OPTYPE_DEFINE(HCOMRECEIVE, "HcomReceive");
 REGISTER_OPTYPE_DEFINE(HCOMREMOTEREAD, "HcomRemoteRead");
 REGISTER_OPTYPE_DEFINE(HCOMREMOTEREFREAD, "HcomRemoteRefRead");
 REGISTER_OPTYPE_DEFINE(HCOMREMOTEWRITE, "HcomRemoteWrite");
+REGISTER_OPTYPE_DEFINE(HCOMREMOTESCATTERWRITE, "HcomRemoteScatterWrite");
 
 REGISTER_OPTYPE_DEFINE(VARASSIGN, "VarAssign");
 REGISTER_OPTYPE_DEFINE(VARISINITIALIZEDOP, "VarIsInitializedOp");

ge/executor/CMakeLists.txt

@@ -32,37 +32,37 @@ set(SRC_LIST
     "../hybrid/node_executor/aicpu/aicpu_ext_info.cc"
     "../model/ge_model.cc"
     "../model/ge_root_model.cc"
-    "../graph/load/new_model_manager/davinci_model.cc"
-    "../graph/load/new_model_manager/davinci_model_parser.cc"
-    "../graph/load/new_model_manager/model_manager.cc"
-    "../graph/load/new_model_manager/tbe_handle_store.cc"
-    "../graph/load/new_model_manager/cpu_queue_schedule.cc"
-    "../graph/load/new_model_manager/model_utils.cc"
-    "../graph/load/new_model_manager/aipp_utils.cc"
-    "../graph/load/new_model_manager/data_inputer.cc"
-    "../graph/load/new_model_manager/data_dumper.cc"
-    "../graph/load/new_model_manager/zero_copy_task.cc"
-    "../graph/load/new_model_manager/zero_copy_offset.cc"
-    "../graph/load/new_model_manager/task_info/task_info.cc"
-    "../graph/load/new_model_manager/task_info/event_record_task_info.cc"
-    "../graph/load/new_model_manager/task_info/event_wait_task_info.cc"
-    "../graph/load/new_model_manager/task_info/fusion_start_task_info.cc"
-    "../graph/load/new_model_manager/task_info/fusion_stop_task_info.cc"
-    "../graph/load/new_model_manager/task_info/kernel_ex_task_info.cc"
-    "../graph/load/new_model_manager/task_info/kernel_task_info.cc"
-    "../graph/load/new_model_manager/task_info/label_set_task_info.cc"
-    "../graph/load/new_model_manager/task_info/label_switch_by_index_task_info.cc"
-    "../graph/load/new_model_manager/task_info/label_goto_ex_task_info.cc"
-    "../graph/load/new_model_manager/task_info/memcpy_async_task_info.cc"
-    "../graph/load/new_model_manager/task_info/memcpy_addr_async_task_info.cc"
-    "../graph/load/new_model_manager/task_info/profiler_trace_task_info.cc"
-    "../graph/load/new_model_manager/task_info/stream_active_task_info.cc"
-    "../graph/load/new_model_manager/task_info/stream_switch_task_info.cc"
-    "../graph/load/new_model_manager/task_info/stream_switchn_task_info.cc"
-    "../graph/load/new_model_manager/task_info/end_graph_task_info.cc"
-    "../graph/load/new_model_manager/task_info/model_exit_task_info.cc"
-    "../graph/load/new_model_manager/task_info/super_kernel/super_kernel_factory.cc"
-    "../graph/load/new_model_manager/task_info/super_kernel/super_kernel.cc"
+    "../graph/load/model_manager/davinci_model.cc"
+    "../graph/load/model_manager/davinci_model_parser.cc"
+    "../graph/load/model_manager/model_manager.cc"
+    "../graph/load/model_manager/tbe_handle_store.cc"
+    "../graph/load/model_manager/cpu_queue_schedule.cc"
+    "../graph/load/model_manager/model_utils.cc"
+    "../graph/load/model_manager/aipp_utils.cc"
+    "../graph/load/model_manager/data_inputer.cc"
+    "../graph/load/model_manager/data_dumper.cc"
+    "../graph/load/model_manager/zero_copy_task.cc"
+    "../graph/load/model_manager/zero_copy_offset.cc"
+    "../graph/load/model_manager/task_info/task_info.cc"
+    "../graph/load/model_manager/task_info/event_record_task_info.cc"
+    "../graph/load/model_manager/task_info/event_wait_task_info.cc"
+    "../graph/load/model_manager/task_info/fusion_start_task_info.cc"
+    "../graph/load/model_manager/task_info/fusion_stop_task_info.cc"
+    "../graph/load/model_manager/task_info/kernel_ex_task_info.cc"
+    "../graph/load/model_manager/task_info/kernel_task_info.cc"
+    "../graph/load/model_manager/task_info/label_set_task_info.cc"
+    "../graph/load/model_manager/task_info/label_switch_by_index_task_info.cc"
+    "../graph/load/model_manager/task_info/label_goto_ex_task_info.cc"
+    "../graph/load/model_manager/task_info/memcpy_async_task_info.cc"
+    "../graph/load/model_manager/task_info/memcpy_addr_async_task_info.cc"
+    "../graph/load/model_manager/task_info/profiler_trace_task_info.cc"
+    "../graph/load/model_manager/task_info/stream_active_task_info.cc"
+    "../graph/load/model_manager/task_info/stream_switch_task_info.cc"
+    "../graph/load/model_manager/task_info/stream_switchn_task_info.cc"
+    "../graph/load/model_manager/task_info/end_graph_task_info.cc"
+    "../graph/load/model_manager/task_info/model_exit_task_info.cc"
+    "../graph/load/model_manager/task_info/super_kernel/super_kernel_factory.cc"
+    "../graph/load/model_manager/task_info/super_kernel/super_kernel.cc"
     "../graph/common/local_context.cc"
     "../opskernel_manager/ops_kernel_builder_manager.cc"
     "../single_op/single_op_manager.cc"
@@ -104,6 +104,7 @@ set(SRC_LIST
     "../hybrid/node_executor/host_cpu/kernel/variable_kernel.cc"
     "../hybrid/node_executor/host_cpu/kernel/assign_kernel.cc"
     "../hybrid/node_executor/host_cpu/kernel/random_uniform_kernel.cc"
+    "../hybrid/node_executor/host_cpu/kernel/data_kernel.cc"
     "../hybrid/node_executor/controlop/control_op_executor.cc"
     "../hybrid/node_executor/partitioned_call/partitioned_call_node_executor.cc"
     "../hybrid/node_executor/rts/rts_node_executor.cc"

ge/executor/ge_executor.cc

@@ -29,15 +29,15 @@
 #include "framework/common/util.h"
 #include "graph/execute/graph_execute.h"
 #include "graph/load/graph_loader.h"
-#include "graph/load/new_model_manager/davinci_model_parser.h"
-#include "graph/load/new_model_manager/model_manager.h"
+#include "graph/load/model_manager/davinci_model_parser.h"
+#include "graph/load/model_manager/model_manager.h"
 #include "graph/manager/graph_mem_allocator.h"
 #include "graph/model.h"
 #include "graph/utils/graph_utils.h"
 #include "mmpa/mmpa_api.h"
 #include "single_op/single_op_manager.h"
 #include "graph/manager/graph_var_manager.h"
-#include "graph/load/new_model_manager/davinci_model.h"
+#include "graph/load/model_manager/davinci_model.h"
 #include "opskernel_manager/ops_kernel_builder_manager.h"
 
 using std::string;

ge/executor/module.mk

@@ -22,37 +22,37 @@ local_ge_executor_src_files :=  \
     ../graph/manager/util/debug.cc \
     ../model/ge_model.cc \
     ../model/ge_root_model.cc \
-    ../graph/load/new_model_manager/davinci_model.cc \
-    ../graph/load/new_model_manager/davinci_model_parser.cc \
-    ../graph/load/new_model_manager/model_manager.cc \
-    ../graph/load/new_model_manager/tbe_handle_store.cc \
-    ../graph/load/new_model_manager/cpu_queue_schedule.cc \
-    ../graph/load/new_model_manager/model_utils.cc \
-    ../graph/load/new_model_manager/aipp_utils.cc \
-    ../graph/load/new_model_manager/data_inputer.cc \
-    ../graph/load/new_model_manager/data_dumper.cc \
-    ../graph/load/new_model_manager/zero_copy_task.cc \
-    ../graph/load/new_model_manager/zero_copy_offset.cc \
-    ../graph/load/new_model_manager/task_info/task_info.cc                  \
-    ../graph/load/new_model_manager/task_info/event_record_task_info.cc     \
-    ../graph/load/new_model_manager/task_info/event_wait_task_info.cc       \
-    ../graph/load/new_model_manager/task_info/fusion_start_task_info.cc     \
-    ../graph/load/new_model_manager/task_info/fusion_stop_task_info.cc      \
-    ../graph/load/new_model_manager/task_info/kernel_ex_task_info.cc        \
-    ../graph/load/new_model_manager/task_info/kernel_task_info.cc           \
-    ../graph/load/new_model_manager/task_info/label_set_task_info.cc        \
-    ../graph/load/new_model_manager/task_info/label_switch_by_index_task_info.cc \
-    ../graph/load/new_model_manager/task_info/label_goto_ex_task_info.cc    \
-    ../graph/load/new_model_manager/task_info/memcpy_async_task_info.cc     \
-    ../graph/load/new_model_manager/task_info/memcpy_addr_async_task_info.cc \
-    ../graph/load/new_model_manager/task_info/profiler_trace_task_info.cc   \
-    ../graph/load/new_model_manager/task_info/stream_active_task_info.cc    \
-    ../graph/load/new_model_manager/task_info/stream_switch_task_info.cc    \
-    ../graph/load/new_model_manager/task_info/stream_switchn_task_info.cc   \
-    ../graph/load/new_model_manager/task_info/end_graph_task_info.cc        \
-    ../graph/load/new_model_manager/task_info/model_exit_task_info.cc       \
-    ../graph/load/new_model_manager/task_info/super_kernel/super_kernel_factory.cc   \
-    ../graph/load/new_model_manager/task_info/super_kernel/super_kernel.cc  \
+    ../graph/load/model_manager/davinci_model.cc \
+    ../graph/load/model_manager/davinci_model_parser.cc \
+    ../graph/load/model_manager/model_manager.cc \
+    ../graph/load/model_manager/tbe_handle_store.cc \
+    ../graph/load/model_manager/cpu_queue_schedule.cc \
+    ../graph/load/model_manager/model_utils.cc \
+    ../graph/load/model_manager/aipp_utils.cc \
+    ../graph/load/model_manager/data_inputer.cc \
+    ../graph/load/model_manager/data_dumper.cc \
+    ../graph/load/model_manager/zero_copy_task.cc \
+    ../graph/load/model_manager/zero_copy_offset.cc \
+    ../graph/load/model_manager/task_info/task_info.cc                  \
+    ../graph/load/model_manager/task_info/event_record_task_info.cc     \
+    ../graph/load/model_manager/task_info/event_wait_task_info.cc       \
+    ../graph/load/model_manager/task_info/fusion_start_task_info.cc     \
+    ../graph/load/model_manager/task_info/fusion_stop_task_info.cc      \
+    ../graph/load/model_manager/task_info/kernel_ex_task_info.cc        \
+    ../graph/load/model_manager/task_info/kernel_task_info.cc           \
+    ../graph/load/model_manager/task_info/label_set_task_info.cc        \
+    ../graph/load/model_manager/task_info/label_switch_by_index_task_info.cc \
+    ../graph/load/model_manager/task_info/label_goto_ex_task_info.cc    \
+    ../graph/load/model_manager/task_info/memcpy_async_task_info.cc     \
+    ../graph/load/model_manager/task_info/memcpy_addr_async_task_info.cc \
+    ../graph/load/model_manager/task_info/profiler_trace_task_info.cc   \
+    ../graph/load/model_manager/task_info/stream_active_task_info.cc    \
+    ../graph/load/model_manager/task_info/stream_switch_task_info.cc    \
+    ../graph/load/model_manager/task_info/stream_switchn_task_info.cc   \
+    ../graph/load/model_manager/task_info/end_graph_task_info.cc        \
+    ../graph/load/model_manager/task_info/model_exit_task_info.cc       \
+    ../graph/load/model_manager/task_info/super_kernel/super_kernel_factory.cc   \
+    ../graph/load/model_manager/task_info/super_kernel/super_kernel.cc  \
     ../opskernel_manager/ops_kernel_builder_manager.cc \
     ../single_op/single_op_manager.cc \
     ../single_op/single_op_model.cc \
@@ -95,6 +95,7 @@ local_ge_executor_src_files :=  \
     ../hybrid/node_executor/host_cpu/kernel/variable_kernel.cc              \
     ../hybrid/node_executor/host_cpu/kernel/assign_kernel.cc                \
     ../hybrid/node_executor/host_cpu/kernel/random_uniform_kernel.cc        \
+    ../hybrid/node_executor/host_cpu/kernel/data_kernel.cc                  \
     ../hybrid/node_executor/controlop/control_op_executor.cc                \
     ../hybrid/node_executor/partitioned_call/partitioned_call_node_executor.cc \
     ../hybrid/node_executor/rts/rts_node_executor.cc                        \

ge/ge_inference.mk

@@ -103,6 +103,7 @@ OMG_HOST_SRC_FILES := \
     graph/passes/net_output_pass.cc \
     graph/passes/replace_transshape_pass.cc \
     graph/passes/constant_fuse_same_pass.cc \
+    graph/passes/fuse_data_nodes_with_common_input_pass.cc \
     graph/passes/print_op_pass.cc \
     graph/passes/no_use_reshape_remove_pass.cc \
     graph/passes/iterator_op_pass.cc \
@@ -227,37 +228,37 @@ OME_HOST_SRC_FILES := \
     graph/manager/util/rt_context_util.cc               \
     graph/manager/util/variable_accelerate_ctrl.cc       \
     graph/manager/util/debug.cc  \
-    graph/load/new_model_manager/model_manager.cc                        \
-    graph/load/new_model_manager/data_inputer.cc                         \
-    graph/load/new_model_manager/davinci_model.cc                        \
-    graph/load/new_model_manager/davinci_model_parser.cc                 \
-    graph/load/new_model_manager/model_utils.cc                          \
-    graph/load/new_model_manager/aipp_utils.cc                           \
-    graph/load/new_model_manager/tbe_handle_store.cc                     \
-    graph/load/new_model_manager/cpu_queue_schedule.cc                   \
-    graph/load/new_model_manager/zero_copy_task.cc                       \
-    graph/load/new_model_manager/zero_copy_offset.cc                     \
-    graph/load/new_model_manager/data_dumper.cc                          \
-    graph/load/new_model_manager/task_info/task_info.cc                  \
-    graph/load/new_model_manager/task_info/event_record_task_info.cc     \
-    graph/load/new_model_manager/task_info/event_wait_task_info.cc       \
-    graph/load/new_model_manager/task_info/fusion_start_task_info.cc     \
-    graph/load/new_model_manager/task_info/fusion_stop_task_info.cc      \
-    graph/load/new_model_manager/task_info/kernel_ex_task_info.cc        \
-    graph/load/new_model_manager/task_info/kernel_task_info.cc           \
-    graph/load/new_model_manager/task_info/label_set_task_info.cc        \
-    graph/load/new_model_manager/task_info/label_switch_by_index_task_info.cc \
-    graph/load/new_model_manager/task_info/label_goto_ex_task_info.cc    \
-    graph/load/new_model_manager/task_info/memcpy_async_task_info.cc     \
-    graph/load/new_model_manager/task_info/memcpy_addr_async_task_info.cc \
-    graph/load/new_model_manager/task_info/profiler_trace_task_info.cc   \
-    graph/load/new_model_manager/task_info/stream_active_task_info.cc    \
-    graph/load/new_model_manager/task_info/stream_switch_task_info.cc    \
-    graph/load/new_model_manager/task_info/stream_switchn_task_info.cc   \
-    graph/load/new_model_manager/task_info/end_graph_task_info.cc        \
-    graph/load/new_model_manager/task_info/model_exit_task_info.cc       \
-    graph/load/new_model_manager/task_info/super_kernel/super_kernel_factory.cc   \
-    graph/load/new_model_manager/task_info/super_kernel/super_kernel.cc  \
+    graph/load/model_manager/model_manager.cc                        \
+    graph/load/model_manager/data_inputer.cc                         \
+    graph/load/model_manager/davinci_model.cc                        \
+    graph/load/model_manager/davinci_model_parser.cc                 \
+    graph/load/model_manager/model_utils.cc                          \
+    graph/load/model_manager/aipp_utils.cc                           \
+    graph/load/model_manager/tbe_handle_store.cc                     \
+    graph/load/model_manager/cpu_queue_schedule.cc                   \
+    graph/load/model_manager/zero_copy_task.cc                       \
+    graph/load/model_manager/zero_copy_offset.cc                     \
+    graph/load/model_manager/data_dumper.cc                          \
+    graph/load/model_manager/task_info/task_info.cc                  \
+    graph/load/model_manager/task_info/event_record_task_info.cc     \
+    graph/load/model_manager/task_info/event_wait_task_info.cc       \
+    graph/load/model_manager/task_info/fusion_start_task_info.cc     \
+    graph/load/model_manager/task_info/fusion_stop_task_info.cc      \
+    graph/load/model_manager/task_info/kernel_ex_task_info.cc        \
+    graph/load/model_manager/task_info/kernel_task_info.cc           \
+    graph/load/model_manager/task_info/label_set_task_info.cc        \
+    graph/load/model_manager/task_info/label_switch_by_index_task_info.cc \
+    graph/load/model_manager/task_info/label_goto_ex_task_info.cc    \
+    graph/load/model_manager/task_info/memcpy_async_task_info.cc     \
+    graph/load/model_manager/task_info/memcpy_addr_async_task_info.cc \
+    graph/load/model_manager/task_info/profiler_trace_task_info.cc   \
+    graph/load/model_manager/task_info/stream_active_task_info.cc    \
+    graph/load/model_manager/task_info/stream_switch_task_info.cc    \
+    graph/load/model_manager/task_info/stream_switchn_task_info.cc   \
+    graph/load/model_manager/task_info/end_graph_task_info.cc        \
+    graph/load/model_manager/task_info/model_exit_task_info.cc       \
+    graph/load/model_manager/task_info/super_kernel/super_kernel_factory.cc   \
+    graph/load/model_manager/task_info/super_kernel/super_kernel.cc  \
     single_op/task/op_task.cc                                            \
     single_op/task/build_task_utils.cc                                   \
     single_op/task/tbe_task_builder.cc                                   \
@@ -269,7 +270,7 @@ OME_HOST_SRC_FILES := \
     single_op/single_op_manager.cc                                       \
     hybrid/hybrid_davinci_model_stub.cc                                  \
     hybrid/node_executor/aicpu/aicpu_ext_info.cc                         \
-    # graph/load/new_model_manager/task_info/hccl_task_info.cc
+    # graph/load/model_manager/task_info/hccl_task_info.cc
 
 OME_DEVICE_SRC_FILES := $(OME_HOST_SRC_FILES)
 

ge/ge_runner.mk

@@ -54,38 +54,38 @@ LIBGE_LOCAL_SRC_FILES := \
     graph/label/partitioned_call_label_maker.cc \
     graph/label/while_label_maker.cc \
     graph/load/graph_loader.cc \
-    graph/load/new_model_manager/cpu_queue_schedule.cc \
-    graph/load/new_model_manager/data_dumper.cc \
-    graph/load/new_model_manager/data_inputer.cc \
-    graph/load/new_model_manager/davinci_model.cc \
-    graph/load/new_model_manager/davinci_model_parser.cc \
-    graph/load/new_model_manager/model_manager.cc \
-    graph/load/new_model_manager/model_utils.cc \
-    graph/load/new_model_manager/aipp_utils.cc \
-    graph/load/new_model_manager/task_info/end_graph_task_info.cc \
-    graph/load/new_model_manager/task_info/model_exit_task_info.cc \
-    graph/load/new_model_manager/task_info/event_record_task_info.cc \
-    graph/load/new_model_manager/task_info/event_wait_task_info.cc \
-    graph/load/new_model_manager/task_info/fusion_start_task_info.cc \
-    graph/load/new_model_manager/task_info/fusion_stop_task_info.cc \
-    graph/load/new_model_manager/task_info/hccl_task_info.cc \
-    graph/load/new_model_manager/task_info/kernel_ex_task_info.cc \
-    graph/load/new_model_manager/task_info/kernel_task_info.cc \
-    graph/load/new_model_manager/task_info/label_set_task_info.cc \
-    graph/load/new_model_manager/task_info/label_switch_by_index_task_info.cc \
-    graph/load/new_model_manager/task_info/label_goto_ex_task_info.cc \
-    graph/load/new_model_manager/task_info/memcpy_addr_async_task_info.cc \
-    graph/load/new_model_manager/task_info/memcpy_async_task_info.cc \
-    graph/load/new_model_manager/task_info/profiler_trace_task_info.cc \
-    graph/load/new_model_manager/task_info/stream_active_task_info.cc \
-    graph/load/new_model_manager/task_info/stream_switch_task_info.cc \
-    graph/load/new_model_manager/task_info/stream_switchn_task_info.cc \
-    graph/load/new_model_manager/task_info/super_kernel/super_kernel.cc \
-    graph/load/new_model_manager/task_info/super_kernel/super_kernel_factory.cc   \
-    graph/load/new_model_manager/task_info/task_info.cc \
-    graph/load/new_model_manager/tbe_handle_store.cc \
-    graph/load/new_model_manager/zero_copy_task.cc \
-    graph/load/new_model_manager/zero_copy_offset.cc    \
+    graph/load/model_manager/cpu_queue_schedule.cc \
+    graph/load/model_manager/data_dumper.cc \
+    graph/load/model_manager/data_inputer.cc \
+    graph/load/model_manager/davinci_model.cc \
+    graph/load/model_manager/davinci_model_parser.cc \
+    graph/load/model_manager/model_manager.cc \
+    graph/load/model_manager/model_utils.cc \
+    graph/load/model_manager/aipp_utils.cc \
+    graph/load/model_manager/task_info/end_graph_task_info.cc \
+    graph/load/model_manager/task_info/model_exit_task_info.cc \
+    graph/load/model_manager/task_info/event_record_task_info.cc \
+    graph/load/model_manager/task_info/event_wait_task_info.cc \
+    graph/load/model_manager/task_info/fusion_start_task_info.cc \
+    graph/load/model_manager/task_info/fusion_stop_task_info.cc \
+    graph/load/model_manager/task_info/hccl_task_info.cc \
+    graph/load/model_manager/task_info/kernel_ex_task_info.cc \
+    graph/load/model_manager/task_info/kernel_task_info.cc \
+    graph/load/model_manager/task_info/label_set_task_info.cc \
+    graph/load/model_manager/task_info/label_switch_by_index_task_info.cc \
+    graph/load/model_manager/task_info/label_goto_ex_task_info.cc \
+    graph/load/model_manager/task_info/memcpy_addr_async_task_info.cc \
+    graph/load/model_manager/task_info/memcpy_async_task_info.cc \
+    graph/load/model_manager/task_info/profiler_trace_task_info.cc \
+    graph/load/model_manager/task_info/stream_active_task_info.cc \
+    graph/load/model_manager/task_info/stream_switch_task_info.cc \
+    graph/load/model_manager/task_info/stream_switchn_task_info.cc \
+    graph/load/model_manager/task_info/super_kernel/super_kernel.cc \
+    graph/load/model_manager/task_info/super_kernel/super_kernel_factory.cc   \
+    graph/load/model_manager/task_info/task_info.cc \
+    graph/load/model_manager/tbe_handle_store.cc \
+    graph/load/model_manager/zero_copy_task.cc \
+    graph/load/model_manager/zero_copy_offset.cc    \
     graph/manager/graph_context.cc \
     graph/manager/graph_manager.cc \
     graph/manager/graph_manager_utils.cc \
@@ -127,6 +127,7 @@ LIBGE_LOCAL_SRC_FILES := \
     graph/passes/compile_nodes_pass.cc \
     graph/passes/constant_folding_pass.cc \
     graph/passes/constant_fuse_same_pass.cc \
+    graph/passes/fuse_data_nodes_with_common_input_pass.cc \
     graph/passes/remove_same_const_pass.cc \
     graph/passes/useless_control_out_remove_pass.cc \
     graph/passes/control_trigger_pass.cc \
@@ -299,6 +300,7 @@ LIBGE_LOCAL_SRC_FILES := \
     hybrid/node_executor/host_cpu/kernel/variable_kernel.cc              \
     hybrid/node_executor/host_cpu/kernel/assign_kernel.cc                \
     hybrid/node_executor/host_cpu/kernel/random_uniform_kernel.cc        \
+    hybrid/node_executor/host_cpu/kernel/data_kernel.cc                  \
     hybrid/node_executor/controlop/control_op_executor.cc                \
     hybrid/node_executor/partitioned_call/partitioned_call_node_executor.cc \
     hybrid/node_executor/hccl/hccl_node_executor.cc                      \

ge/ge_runtime/CMakeLists.txt

@@ -23,6 +23,8 @@ add_library(ge_runtime SHARED ${GE_SRC_LIST})
 target_compile_options(ge_runtime PRIVATE
     -Werror
     -O2
+    -Wno-deprecated-declarations
+    -fno-common
 )
 
 target_compile_definitions(ge_runtime PRIVATE 

ge/graph/build/graph_builder.cc

@@ -187,8 +187,7 @@ Status GraphBuilder::UpdateParentNodeOutputSize(const ge::ComputeGraphPtr &graph
   return SUCCESS;
 }
 
-Status GraphBuilder::Build(ComputeGraphPtr &comp_graph, std::vector<SubGraphInfoPtr> &subgraph_ptr_list,
-                           GeRootModelPtr &ge_root_model_ptr, uint64_t session_id) {
+Status GraphBuilder::Build(ComputeGraphPtr &comp_graph, GeRootModelPtr &ge_root_model_ptr, uint64_t session_id) {
   if (comp_graph == nullptr) {
     GELOGE(GE_GRAPH_PARAM_NULLPTR, "Graph build comp_graph is null.");
     return GE_GRAPH_PARAM_NULLPTR;
@@ -203,18 +202,18 @@ Status GraphBuilder::Build(ComputeGraphPtr &comp_graph, std::vector<SubGraphInfo
   (void)AttrUtils::GetBool(comp_graph, ATTR_NAME_DYNAMIC_SHAPE_PARTITIONED, is_dynamic_shape);
   if (is_dynamic_shape || comp_graph->GetGraphUnknownFlag()) {
     GE_CHK_STATUS_RET(
-        BuildForDynamicShapeGraph(comp_graph, subgraph_ptr_list, ge_root_model_ptr, ge_model_ptr, session_id),
+        BuildForDynamicShapeGraph(comp_graph, ge_root_model_ptr, ge_model_ptr, session_id),
         "Build for dynamic shape graph failed.");
     return SUCCESS;
   }
 
-  GE_CHK_STATUS_RET(BuildForKnownShapeGraph(comp_graph, subgraph_ptr_list, ge_model_ptr, session_id),
+  GE_CHK_STATUS_RET(BuildForKnownShapeGraph(comp_graph, ge_model_ptr, session_id),
                     "Build for known shape graph failed.");
   ge_root_model_ptr->SetSubgraphInstanceNameToModel(comp_graph->GetName(), ge_model_ptr);
   return SUCCESS;
 }
 
-Status GraphBuilder::BuildForKnownShapeGraph(ComputeGraphPtr &comp_graph, std::vector<SubGraphInfoPtr> &subgraph_list,
+Status GraphBuilder::BuildForKnownShapeGraph(ComputeGraphPtr &comp_graph,
                                              GeModelPtr &ge_model_ptr, uint64_t session_id) {
   if (ge::GetContext().GetHostExecFlag()) {
     GE_CHK_STATUS_RET(BuildForHostCpuGraph(comp_graph, ge_model_ptr, session_id), "Build for host-cpu graph failed.");
@@ -222,7 +221,7 @@ Status GraphBuilder::BuildForKnownShapeGraph(ComputeGraphPtr &comp_graph, std::v
   }
 
   GELOGI("Begin to build known shape graph[%s].", comp_graph->GetName().c_str());
-  Status ret = SecondPartition(comp_graph, subgraph_list);
+  Status ret = SecondPartition(comp_graph);
   GE_CHK_STATUS_RET(ret, "Graph[%s] second partition Failed.", comp_graph->GetName().c_str());
   auto subgraph_map = graph_partitioner_.GetSubGraphMap();
 
@@ -470,7 +469,6 @@ Status GraphBuilder::MarkFpBpProfilingTaskAttr(ComputeGraphPtr &com_graph) {
 }
 
 Status GraphBuilder::BuildForDynamicShapeGraph(ComputeGraphPtr &comp_graph,
-                                               std::vector<SubGraphInfoPtr> &subgraph_ptr_list,
                                                GeRootModelPtr &ge_root_model_ptr, GeModelPtr &ge_model_ptr,
                                                uint64_t session_id) {
   GELOGI("Start to build BuildForDynamicShape for dynamic shape.");
@@ -517,7 +515,7 @@ Status GraphBuilder::BuildForDynamicShapeGraph(ComputeGraphPtr &comp_graph,
         }
       }
       // known shape build flow
-      GE_CHK_STATUS_RET(BuildForKnownShapeGraph(sub_graph, subgraph_ptr_list, ge_model_ptr, session_id),
+      GE_CHK_STATUS_RET(BuildForKnownShapeGraph(sub_graph, ge_model_ptr, session_id),
                         "Build for known shape graph failed.");
     }
     ge_root_model_ptr->SetSubgraphInstanceNameToModel(sub_graph->GetName(), ge_model_ptr);
@@ -719,7 +717,7 @@ Status GraphBuilder::CalcDynShapeRootGraphDataSize(const ge::OpDescPtr &op_desc)
   return SUCCESS;
 }
 
-Status GraphBuilder::SecondPartition(ge::ComputeGraphPtr &comp_graph, vector<ge::SubGraphInfoPtr> &subgraph_ptr_list) {
+Status GraphBuilder::SecondPartition(ge::ComputeGraphPtr &comp_graph) {
   GE_TIMESTAMP_START(GraphPartition2);
   auto ret = graph_partitioner_.Partition(comp_graph, GraphPartitioner::kSecondPartitioning);
   if (ret != SUCCESS) {
@@ -727,10 +725,8 @@ Status GraphBuilder::SecondPartition(ge::ComputeGraphPtr &comp_graph, vector<ge:
     return ret;
   }
   GE_CHK_STATUS_RET(ret, "Graph partition Failed.");
-  auto graph_2_subgraphlist = graph_partitioner_.GetSubGraphMap();
-  if (graph_2_subgraphlist.find(comp_graph) != graph_2_subgraphlist.end()) {
-    subgraph_ptr_list = graph_2_subgraphlist[comp_graph];
-  } else {
+  const auto &graph_2_subgraphlist = graph_partitioner_.GetSubGraphMap();
+  if (graph_2_subgraphlist.find(comp_graph) == graph_2_subgraphlist.end()) {
     GELOGE(FAILED, "Find subgraph failed.");
     return FAILED;
   }

ge/graph/build/graph_builder.h

@@ -47,8 +47,7 @@ class GraphBuilder {
   GraphBuilder(const GraphBuilder &in) = delete;
   GraphBuilder &operator=(const GraphBuilder &in) = delete;
   virtual ~GraphBuilder() = default;
-  Status Build(ComputeGraphPtr &comp_graph, std::vector<SubGraphInfoPtr> &subgraph_ptr_list,
-               GeRootModelPtr &ge_model_ptr, uint64_t session_id = INVALID_SESSION_ID);
+  Status Build(ComputeGraphPtr &comp_graph, GeRootModelPtr &ge_model_ptr, uint64_t session_id = INVALID_SESSION_ID);
   void SetOptions(const GraphManagerOptions &options);
 
  private:
@@ -59,12 +58,12 @@ class GraphBuilder {
   Status UpdateDataInputSize(const ge::NodePtr &node_ptr);
   Status UpdateParentNodeOutputSize(const ge::ComputeGraphPtr &graph, ge::NodePtr &parent_node_ptr);
   Status CalcDynShapeRootGraphDataSize(const ge::OpDescPtr &op_desc);
-  Status SecondPartition(ge::ComputeGraphPtr &comp_graph, vector<ge::SubGraphInfoPtr> &subgraph_ptr_list);
+  Status SecondPartition(ge::ComputeGraphPtr &comp_graph);
   Status MarkFpBpProfilingTaskAttr(ComputeGraphPtr &com_graph);
-  Status BuildForDynamicShapeGraph(ComputeGraphPtr &comp_graph, std::vector<SubGraphInfoPtr> &subgraph_ptr_list,
+  Status BuildForDynamicShapeGraph(ComputeGraphPtr &comp_graph,
                                    GeRootModelPtr &ge_root_model_ptr, GeModelPtr &ge_model_ptr,
                                    uint64_t session_id = INVALID_SESSION_ID);
-  Status BuildForKnownShapeGraph(ComputeGraphPtr &comp_graph, std::vector<SubGraphInfoPtr> &subgraph_list,
+  Status BuildForKnownShapeGraph(ComputeGraphPtr &comp_graph,
                                  GeModelPtr &ge_model_ptr, uint64_t session_id = INVALID_SESSION_ID);
   Status BuildForUnknownShapeGraph(ComputeGraphPtr &comp_graph, GeModelPtr &ge_model_ptr,
                                    uint64_t session_id = INVALID_SESSION_ID);

ge/graph/build/memory/block_mem_assigner.cc

@@ -24,6 +24,7 @@
 #include "graph/buffer.h"
 #include "graph/ge_attr_value.h"
 #include "graph/ge_context.h"
+#include "graph/types.h"
 #include "graph/node.h"
 #include "graph/utils/graph_utils.h"
 #include "graph/utils/node_utils.h"
@@ -1401,6 +1402,7 @@ Status BlockMemAssigner::AssignOutputMemoryWithReuse(const NodePtr &node, vector
     if (output_op_desc != nullptr) {
       GE_IF_BOOL_EXEC(ge::TensorUtils::GetSize(*output_op_desc, size) != SUCCESS, GELOGI("Get size failed"));
     }
+
     // fusion: other type's size not means malloc HBM memory
     bool l1_flag = has_mem_type_attr && memorys_type[i] == RT_MEMORY_L1;
     if (l1_flag) {
@@ -1408,6 +1410,11 @@ Status BlockMemAssigner::AssignOutputMemoryWithReuse(const NodePtr &node, vector
              op_desc->GetName().c_str(), op_desc->GetOutputNameByIndex(i).c_str(), memorys_type[i]);
       size = 0;
     }
+
+    int32_t calc_type = 0;
+    bool ret = ge::AttrUtils::GetInt(output_op_desc, ATTR_NAME_MEMORY_SIZE_CALC_TYPE, calc_type);
+    GE_IF_BOOL_EXEC((ret && (calc_type == static_cast<int32_t>(ge::MemorySizeCalcType::ALWAYS_EMPTY))), size = 0;);
+
     std::string peer_name;
     uint32_t peer_input_index = 0;
     bool out_node_set_continuous_input = false;

ge/graph/build/memory/var_mem_assign_util.cc

@@ -60,9 +60,14 @@ Status VarMemAssignUtil::AssignStaticMemory2Node(ge::ComputeGraphPtr &compute_gr
                     return FAILED);
     ge::ConstGeTensorDescPtr tensor_desc = n->GetOpDesc()->GetOutputDescPtr(0);
     GE_CHECK_NOTNULL(tensor_desc);
+    rtMemType_t memory_type = RT_MEMORY_HBM;
+    uint32_t mem_type = 0;
+    if (AttrUtils::GetInt(n->GetOpDesc(), ATTR_OUTPUT_MEMORY_TYPE, mem_type) && (mem_type == 1)) {
+      memory_type = RT_MEMORY_RDMA_HBM;
+    }
     if (!VarManager::Instance(compute_graph->GetSessionID())->IsVarExist(node_name, *tensor_desc)) {
       GE_CHK_STATUS_RET(
-          VarManager::Instance(compute_graph->GetSessionID())->AssignVarMem(node_name, *tensor_desc, RT_MEMORY_HBM));
+          VarManager::Instance(compute_graph->GetSessionID())->AssignVarMem(node_name, *tensor_desc, memory_type));
       GE_IF_BOOL_EXEC(n->GetType() == VARIABLE,
                       GE_CHK_STATUS_RET(AssignData2Fp32Var(n, compute_graph->GetSessionID())));
       GE_CHK_STATUS_RET(VarManager::Instance(compute_graph->GetSessionID())
@@ -70,7 +75,6 @@ Status VarMemAssignUtil::AssignStaticMemory2Node(ge::ComputeGraphPtr &compute_gr
     }
 
     uint8_t *dev_ptr = nullptr;
-    rtMemType_t memory_type = RT_MEMORY_HBM;
     GE_CHK_STATUS_RET(VarManager::Instance(compute_graph->GetSessionID())
                           ->GetVarAddr(node_name, *tensor_desc, &dev_ptr, memory_type));
     vector<int64_t> output_list = n->GetOpDesc()->GetOutputOffset();

ge/graph/build/stream_allocator.cc

@@ -1013,6 +1013,24 @@ bool StreamAllocator::IsActivated(int64_t stream_id) const {
   return false;
 }
 
+// Iteraotor loop :
+// StreamSwitch  ->  StreamActive
+// FpBp loop:
+// StreamSwitch  ->  AssignAdd  ->  StreamActive
+NodePtr FindSwitchNodeBeforeLoopActiveNode(const NodePtr &active_node) {
+  for (auto pre_node : active_node->GetInControlNodes()) {
+    if (pre_node->GetType() == STREAMSWITCH) {
+      return pre_node;
+    }
+    for (auto pre_pre_node : pre_node->GetInControlNodes()) {
+      if (pre_pre_node->GetType() == STREAMSWITCH) {
+        return pre_pre_node;
+      }
+    }
+  }
+  return nullptr;
+}
+
 Status StreamAllocator::SetActiveStreamsForLoop() {
   vector<uint32_t> loop_active_streams;
   for (int64_t stream_id = 0; stream_id < stream_num_; stream_id++) {
@@ -1038,6 +1056,13 @@ Status StreamAllocator::SetActiveStreamsForLoop() {
     bool is_loop_active = false;
     if (AttrUtils::GetBool(node->GetOpDesc(), ATTR_NAME_IS_LOOP_ACTIVE, is_loop_active) && is_loop_active) {
       vector<string> activated_label_list;
+
+      NodePtr pre_switch_node = FindSwitchNodeBeforeLoopActiveNode(node);
+      if (pre_switch_node == nullptr) {
+        GELOGE(FAILED, "find switch node before loop active node %s failed", node->GetName().c_str());
+        return FAILED;
+      }
+
       if (!AttrUtils::GetListStr(node->GetOpDesc(), ATTR_NAME_ACTIVE_LABEL_LIST, activated_label_list) ||
           activated_label_list.empty()) {
         GE_CHK_BOOL_EXEC(AttrUtils::SetListInt(node->GetOpDesc(), ATTR_NAME_ACTIVE_STREAM_LIST, loop_active_streams),
@@ -1053,7 +1078,7 @@ Status StreamAllocator::SetActiveStreamsForLoop() {
         // it may cause some stream actived by iterator next step when this stream still alive.
         // If above situation happen, active message will lose, cause process block in next iteration.
         // In order to avoid this abnormal happen,
-        // add event between each last node and iterator active node in target active stream
+        // add event between each last node and iterator switch node
         GELOGI("there are %zu next iterator target streams has streamswitch node.", streams_skip_iterator_event.size());
         for (auto iter : stream_id_to_last_node) {
           if (streams_skip_iterator_event.find(iter.first) != streams_skip_iterator_event.end()) {
@@ -1067,7 +1092,7 @@ Status StreamAllocator::SetActiveStreamsForLoop() {
             continue;
           }
           AddSendEventId(iter.second, event_num_);
-          AddRecvEventId(node, event_num_);
+          AddRecvEventId(pre_switch_node, event_num_);
           event_num_++;
         }
 

ge/graph/execute/graph_execute.cc

@@ -21,7 +21,7 @@
 
 #include "common/ge_inner_error_codes.h"
 #include "common/model_parser/base.h"
-#include "graph/load/new_model_manager/model_manager.h"
+#include "graph/load/model_manager/model_manager.h"
 #include "omm/csa_interact.h"
 #include "runtime/dev.h"
 #include "runtime/mem.h"

ge/graph/load/graph_loader.cc

@@ -22,8 +22,8 @@
 #include "common/helper/model_helper.h"
 #include "common/util.h"
 #include "graph/ge_context.h"
-#include "graph/load/new_model_manager/davinci_model_parser.h"
-#include "graph/load/new_model_manager/model_manager.h"
+#include "graph/load/model_manager/davinci_model_parser.h"
+#include "graph/load/model_manager/model_manager.h"
 #include "graph/manager/graph_var_manager.h"
 #include "omm/csa_interact.h"
 #include "runtime/dev.h"

ge/graph/load/new_model_manager/aipp_utils.cc → ge/graph/load/model_manager/aipp_utils.cc

@@ -14,7 +14,7 @@
  * limitations under the License.
  */
 
-#include "graph/load/new_model_manager/aipp_utils.h"
+#include "graph/load/model_manager/aipp_utils.h"
 
 #include <string>
 

ge/graph/load/new_model_manager/cpu_queue_schedule.cc → ge/graph/load/model_manager/cpu_queue_schedule.cc

@@ -14,7 +14,7 @@
  * limitations under the License.
  */
 
-#include "graph/load/new_model_manager/cpu_queue_schedule.h"
+#include "graph/load/model_manager/cpu_queue_schedule.h"
 #include "common/debug/ge_log.h"
 #include "common/debug/log.h"
 

ge/graph/load/new_model_manager/cpu_queue_schedule.h → ge/graph/load/model_manager/cpu_queue_schedule.h

@@ -20,8 +20,8 @@
 #include <vector>
 
 #include "common/ge_inner_error_codes.h"
-#include "graph/load/new_model_manager/task_info/task_info.h"
-#include "graph/load/new_model_manager/zero_copy_offset.h"
+#include "graph/load/model_manager/task_info/task_info.h"
+#include "graph/load/model_manager/zero_copy_offset.h"
 #include "runtime/kernel.h"
 
 namespace ge {

ge/graph/load/new_model_manager/data_dumper.cc → ge/graph/load/model_manager/data_dumper.cc

@@ -14,7 +14,7 @@
  * limitations under the License.
  */
 
-#include "graph/load/new_model_manager/data_dumper.h"
+#include "graph/load/model_manager/data_dumper.h"
 
 #include <cstdlib>
 #include <ctime>
@@ -29,7 +29,7 @@
 #include "framework/common/util.h"
 #include "graph/anchor.h"
 #include "graph/debug/ge_attr_define.h"
-#include "graph/load/new_model_manager/model_utils.h"
+#include "graph/load/model_manager/model_utils.h"
 #include "graph/manager/util/debug.h"
 #include "graph/utils/attr_utils.h"
 #include "graph/utils/tensor_utils.h"

ge/graph/load/new_model_manager/data_inputer.cc → ge/graph/load/model_manager/data_inputer.cc

@@ -14,7 +14,7 @@
  * limitations under the License.
  */
 
-#include "graph/load/new_model_manager/data_inputer.h"
+#include "graph/load/model_manager/data_inputer.h"
 
 #include <securec.h>
 

ge/graph/load/new_model_manager/davinci_model.cc → ge/graph/load/model_manager/davinci_model.cc

@@ -14,7 +14,7 @@
  * limitations under the License.
  */
 
-#include "graph/load/new_model_manager/davinci_model.h"
+#include "graph/load/model_manager/davinci_model.h"
 
 #include <graph/utils/node_utils.h>
 #include <algorithm>
@@ -36,9 +36,9 @@
 #include "graph/debug/ge_attr_define.h"
 #include "graph/ge_context.h"
 #include "graph/graph.h"
-#include "graph/load/new_model_manager/cpu_queue_schedule.h"
-#include "graph/load/new_model_manager/model_manager.h"
-#include "graph/load/new_model_manager/tbe_handle_store.h"
+#include "graph/load/model_manager/cpu_queue_schedule.h"
+#include "graph/load/model_manager/model_manager.h"
+#include "graph/load/model_manager/tbe_handle_store.h"
 #include "graph/manager/graph_mem_allocator.h"
 #include "graph/manager/graph_var_manager.h"
 #include "graph/manager/trans_var_data_utils.h"
@@ -520,6 +520,8 @@ Status DavinciModel::DoTaskSink() {
 
   GE_CHK_STATUS_RET(InitEntryTask(), "InitEntryTask failed.");
 
+  GE_CHK_STATUS_RET(InitL1DataDumperArgs(), "InitL1DataDumperArgs failed.");
+
   GE_CHK_STATUS_RET(DistributeTask(), "Distribute failed.");
 
   GE_CHK_RT_RET(rtModelLoadComplete(rt_model_handle_));
@@ -716,24 +718,10 @@ Status DavinciModel::Init(void *dev_ptr, size_t mem_size, void *weight_ptr, size
   GE_CHK_STATUS_RET(DoTaskSink(), "Task sink failed");
   GE_TIMESTAMP_END(DoTaskSink, "GraphLoader::DoTaskSink");
 
-  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();
-  bool dump_l1fusion_op = (all_dump_model.find(ge::DUMP_ALL_MODEL) != all_dump_model.end()) ||
-                          findByOmName || findByModelName;
-  if (dump_l1fusion_op) {
-    // malloc 2M for dump l1fusion op
-    GE_CHK_RT_RET(rtMalloc(&l1_fusion_addr_, kDumpL1FusionOpMByteSize, RT_MEMORY_DDR));
-
-    // send l1fusion dump addr to rts
-    GE_CHK_RT_RET(rtDumpAddrSet(rt_model_handle_, l1_fusion_addr_, kDumpL1FusionOpMByteSize, kDumpFlagOfL1Fusion));
-  }
-
   /// In zero copy model, if a aicpu operator is connected to the first or last layer, before model execution,
   /// the aicpu opertor needs to destroy history record, and update operator memory address.
   /// The model with specified aicpu operators is only marked here, and destruction is in ModelManager::ExecuteModel().
   need_destroy_aicpu_kernel_ = IsAicpuKernelConnectSpecifiedLayer();
-  (void)ge::AttrUtils::GetListStr(ge_model_, ATTR_MODEL_OUT_NODES_NAME, out_node_name_);
 
   string fp_ceiling_mode;
   if (ge::AttrUtils::GetStr(ge_model_, ATTR_FP_CEILING_MODE, fp_ceiling_mode)) {
@@ -2079,6 +2067,8 @@ void DavinciModel::CreateOutput(uint32_t index, const OpDescPtr &op_desc, InputO
 
 Status DavinciModel::InitOutputDescInfo(const vector<OpDescPtr> &output_op_list) {
   GELOGD("Output node size: %zu", output_op_list.size());
+  vector<string> out_node_name;
+  (void)ge::AttrUtils::GetListStr(ge_model_, ATTR_MODEL_OUT_NODES_NAME, out_node_name);
   for (const auto &op_desc : output_op_list) {
     uint32_t out_size = static_cast<uint32_t>(op_desc->GetInputsSize());
     for (uint32_t index = 0; index < out_size; index++) {
@@ -2092,11 +2082,11 @@ Status DavinciModel::InitOutputDescInfo(const vector<OpDescPtr> &output_op_list)
       GE_CHK_BOOL_RET_STATUS(src_name.size() > index && src_index.size() > index, INTERNAL_ERROR,
                              "construct output_name failed.");
       // forward compatbility, if old om has no out_node_name, need to return output follow origin way
-      if (out_size == out_node_name_.size()) {
+      if (out_size == out_node_name.size()) {
         // neweast plan, the index will add to name during generate model.
-        bool contains_colon = out_node_name_[index].find(":") != std::string::npos;
+        bool contains_colon = out_node_name[index].find(":") != std::string::npos;
         output_name =
-            contains_colon ? out_node_name_[index] : out_node_name_[index] + ":" + std::to_string(src_index[index]);
+            contains_colon ? out_node_name[index] : out_node_name[index] + ":" + std::to_string(src_index[index]);
       } else {
         output_name = std::string("output_") + std::to_string(index) + "_" + src_name[index] + "_" +
                       std::to_string(src_index[index]);
@@ -3075,6 +3065,64 @@ Status DavinciModel::MallocKnownArgs() {
   return SUCCESS;
 }
 
+void DavinciModel::SaveProfilingTaskDescInfo(const OpDescPtr &op, const TaskInfoPtr &task,
+                                             const domi::TaskDef &task_def, size_t task_index) {
+  bool flag = GetL1FusionEnableOption();
+  char skt_enable_env[MMPA_MAX_PATH] = { 0x00 };
+  INT32 res = mmGetEnv("SKT_ENABLE", skt_enable_env, MMPA_MAX_PATH);
+  int64_t env_flag = (res == EN_OK) ? std::strtol(skt_enable_env, nullptr, kDecimal) : 0;
+  if (env_flag != 0) {
+    flag = true;
+  }
+
+  TaskDescInfo task_desc_info;
+  if (!om_name_.empty()) {
+    task_desc_info.model_name = om_name_;
+  } else {
+    task_desc_info.model_name = name_;
+  }
+  task_desc_info.op_name = op->GetName();
+  task_desc_info.block_dim = task_def.kernel().block_dim();
+  task_desc_info.task_id = task->GetTaskID();
+  task_desc_info.stream_id = task->GetStreamId();
+  task_desc_info.shape_type = "static";
+  task_desc_info.cur_iter_num = 0;
+  // task type
+  task_desc_info.task_type = kTaskTypeInvalid;
+  auto model_task_type = static_cast<rtModelTaskType_t>(task_def.type());
+  if (model_task_type == RT_MODEL_TASK_KERNEL) {
+    const domi::KernelDef &kernel_def = task_def.kernel();
+    const auto &context = kernel_def.context();
+    auto kernel_type = static_cast<ccKernelType>(context.kernel_type());
+    if (kernel_type == ccKernelType::TE) {
+      task_desc_info.task_type = kTaskTypeAicore;
+    } else if (kernel_type == ccKernelType::AI_CPU || kernel_type == ccKernelType::CUST_AI_CPU) {
+      task_desc_info.task_type = kTaskTypeAicpu;
+    } else {
+      GELOGD("Other kernel type: %u", context.kernel_type());
+    }
+  } else if (model_task_type == RT_MODEL_TASK_KERNEL_EX) {
+    task_desc_info.task_type = kTaskTypeAicpu;
+  } else {
+    GELOGD("Skip task type: %d", static_cast<int>(model_task_type));
+  }
+  profiler_report_op_info_[task_desc_info.op_name] =
+    std::pair<uint32_t, uint32_t>(task_desc_info.task_id, task_desc_info.stream_id);
+  task_desc_info_.emplace_back(task_desc_info);
+  if (flag) {
+    if (task->GetSktTaskID() != 0xFFFFFFFF) {
+      TaskDescInfo task_desc_info;
+      string op_name = "super_kernel_" + to_string(task_index);
+      task_desc_info.op_name = op_name;
+      task_desc_info.task_id = task->GetSktTaskID();
+      profiler_report_op_info_[task_desc_info.op_name] =
+        std::pair<uint32_t, uint32_t>(task_desc_info.task_id, task_desc_info.stream_id);
+      task_desc_info_.emplace_back(task_desc_info);
+    }
+  }
+  return;
+}
+
 Status DavinciModel::DistributeTask() {
   GELOGI("do Distribute.");
   for (auto &task : cpu_task_list_) {
@@ -3086,18 +3134,11 @@ Status DavinciModel::DistributeTask() {
   }
 
   task_desc_info_.clear();
-  bool flag = GetL1FusionEnableOption();
-  char skt_enable_env[MMPA_MAX_PATH] = { 0x00 };
-  INT32 res = mmGetEnv("SKT_ENABLE", skt_enable_env, MMPA_MAX_PATH);
-  int64_t env_flag = (res == EN_OK) ? std::strtol(skt_enable_env, nullptr, kDecimal) : 0;
-  if (env_flag != 0) {
-    flag = true;
-  }
-
   const auto &model_task_def = ge_model_->GetModelTaskDefPtr();
   for (size_t task_index = 0; task_index < task_list_.size(); ++task_index) {
     auto &task_def = model_task_def->task(task_index);
     auto &task = task_list_.at(task_index);
+    GE_CHECK_NOTNULL(task);
     GE_CHK_STATUS_RET(task->Distribute(), "Task[%zu] distribute fail", task_index);
     // for data dump
     auto op_index = std::max(task_def.kernel().context().op_index(),
@@ -3117,33 +3158,9 @@ Status DavinciModel::DistributeTask() {
     GE_IF_BOOL_EXEC(no_need_profiling, continue);
 
     SaveDumpOpInfo(runtime_param_, op, task->GetTaskID(), task->GetStreamId());
-    // Load task info for profiling
-    TaskDescInfo task_desc_info;
-    if (!om_name_.empty()) {
-      task_desc_info.model_name = om_name_;
-    } else {
-      task_desc_info.model_name = name_;
-    }
-    task_desc_info.op_name = op->GetName();
-    task_desc_info.block_dim = task_def.kernel().block_dim();
-    task_desc_info.task_id = task->GetTaskID();
-    task_desc_info.stream_id = task->GetStreamId();
-    task_desc_info.shape_type = "static";
-    task_desc_info.cur_iter_num = 0;
-    profiler_report_op_info_[task_desc_info.op_name] =
-      std::pair<uint32_t, uint32_t>(task_desc_info.task_id, task_desc_info.stream_id);
-    task_desc_info_.emplace_back(task_desc_info);
-    if (flag) {
-      if (task->GetSktTaskID() != 0xFFFFFFFF) {
-        TaskDescInfo task_desc_info;
-        string op_name = "super_kernel_" + to_string(task_index);
-        task_desc_info.op_name = op_name;
-        task_desc_info.task_id = task->GetSktTaskID();
-        profiler_report_op_info_[task_desc_info.op_name] =
-          std::pair<uint32_t, uint32_t>(task_desc_info.task_id, task_desc_info.stream_id);
-        task_desc_info_.emplace_back(task_desc_info);
-      }
-    }
+
+    // save task info for profiling
+    SaveProfilingTaskDescInfo(op, task, task_def, task_index);
   }
   // launch dump kernel to aicpu
   GE_CHK_STATUS_RET(data_dumper_.LoadDumpInfo(), "Load dump info failed.");
@@ -3951,7 +3968,6 @@ void DavinciModel::SetDataDumperArgs(const ComputeGraphPtr &graph, const map<str
   data_dumper_.SetOmName(om_name_);
   data_dumper_.SetComputeGraph(graph);
   data_dumper_.SetRefInfo(saved_task_addrs_);
-  data_dumper_.SetL1FusionAddr(l1_fusion_addr_);
 
   int32_t device_id = 0;
   rtError_t rt_ret = rtGetDevice(&device_id);
@@ -4161,4 +4177,28 @@ 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 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) {
+    // malloc 2M for dump l1fusion op
+    GE_CHK_RT_RET(rtMalloc(&l1_fusion_addr_, kDumpL1FusionOpMByteSize, RT_MEMORY_DDR));
+
+    // send l1fusion dump addr to rts
+    if (rtDumpAddrSet(rt_model_handle_, l1_fusion_addr_, kDumpL1FusionOpMByteSize, kDumpFlagOfL1Fusion) !=
+        RT_ERROR_NONE) {
+      // l1_fusion_addr_ will be free when DavinciModel destruct
+      GELOGE(FAILED, "Call rtDumpAddrSet failed");
+      return FAILED;
+    }
+
+    // set addr for l1 data dump
+    data_dumper_.SetL1FusionAddr(l1_fusion_addr_);
+  }
+  return SUCCESS;
+}
+
 }  // namespace ge

ge/graph/load/new_model_manager/davinci_model.h → ge/graph/load/model_manager/davinci_model.h

@@ -32,12 +32,12 @@
 #include "common/types.h"
 #include "framework/common/util.h"
 #include "graph/debug/ge_attr_define.h"
-#include "graph/load/new_model_manager/aipp_utils.h"
-#include "graph/load/new_model_manager/data_dumper.h"
-#include "graph/load/new_model_manager/data_inputer.h"
-#include "graph/load/new_model_manager/model_utils.h"
-#include "graph/load/new_model_manager/zero_copy_offset.h"
-#include "graph/load/new_model_manager/zero_copy_task.h"
+#include "graph/load/model_manager/aipp_utils.h"
+#include "graph/load/model_manager/data_dumper.h"
+#include "graph/load/model_manager/data_inputer.h"
+#include "graph/load/model_manager/model_utils.h"
+#include "graph/load/model_manager/zero_copy_offset.h"
+#include "graph/load/model_manager/zero_copy_task.h"
 #include "graph/model.h"
 #include "graph/node.h"
 #include "graph/op_desc.h"
@@ -623,6 +623,9 @@ class DavinciModel {
 
   Status DistributeTask();
 
+  void SaveProfilingTaskDescInfo(const OpDescPtr &op, const TaskInfoPtr &task,
+                                 const domi::TaskDef &task_def, size_t task_index);
+
   uint8_t *MallocFeatureMapMem(size_t data_size);
 
   uint8_t *MallocWeightsMem(size_t weights_size);
@@ -837,6 +840,8 @@ class DavinciModel {
 
   void SetDataDumperArgs(const ComputeGraphPtr &graph, const map<string, OpDescPtr> &variable_by_name);
 
+  Status InitL1DataDumperArgs();
+
   Status InitModelProfile();
   Status SinkModelProfile();
 
@@ -881,7 +886,6 @@ class DavinciModel {
   GeModelPtr ge_model_;  // release after DavinciModel::Init
 
   bool need_destroy_aicpu_kernel_{false};
-  vector<string> out_node_name_;
 
   map<uint32_t, OpDescPtr> op_list_;  // release after DavinciModel::Init
 

ge/graph/load/new_model_manager/davinci_model_parser.cc → ge/graph/load/model_manager/davinci_model_parser.cc

@@ -14,7 +14,7 @@
  * limitations under the License.
  */
 
-#include "graph/load/new_model_manager/davinci_model_parser.h"
+#include "graph/load/model_manager/davinci_model_parser.h"
 
 namespace ge {
 DavinciModelParser::DavinciModelParser() {}

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

@@ -14,10 +14,11 @@
  * limitations under the License.
  */
 
-#include "graph/load/new_model_manager/model_manager.h"
+#include "graph/load/model_manager/model_manager.h"
 
 #include <string>
 
+#include "mmpa/mmpa_api.h"
 #include "aicpu/aicpu_schedule/aicpu_op_type_list.h"
 #include "common/dump/dump_manager.h"
 #include "common/l2_cache_optimize.h"
@@ -27,8 +28,8 @@
 #include "framework/common/util.h"
 #include "graph/common/ge_call_wrapper.h"
 #include "graph/debug/ge_attr_define.h"
-#include "graph/load/new_model_manager/davinci_model.h"
-#include "graph/load/new_model_manager/davinci_model_parser.h"
+#include "graph/load/model_manager/davinci_model.h"
+#include "graph/load/model_manager/davinci_model_parser.h"
 #include "model/ge_root_model.h"
 #include "graph/common/local_context.h"
 #include "graph/utils/attr_utils.h"
@@ -53,7 +54,6 @@ const char *const kBatchLoadBuf = "batchLoadsoFrombuf";
 const char *const kDeleteCustOp = "deleteCustOp";
 const int kTimeSpecNano = 1000000000;
 const int kTimeSpecMiro = 1000000;
-const int kSessionMaxBias = 100;
 const int kOpNameMaxSize = 100;
 struct CustAicpuSoBuf {
   uint64_t kernelSoBuf;
@@ -1024,6 +1024,12 @@ Status ModelManager::GetAippType(uint32_t model_id, uint32_t index, InputAippTyp
 }
 
 Status ModelManager::GenSessionId(uint64_t &session_id) {
+  const uint64_t kSessionTimeMask = 0xffffffffffff0000;
+  const uint64_t kSessionPidMask  = 0x000000000000ff00;
+  const uint64_t kSessionBiasMask = 0x00000000000000ff;
+
+  const uint64_t kMaskPerOffset = 8;
+
   std::lock_guard<std::mutex> lock(session_id_create_mutex_);
 
   mmTimeval tv;
@@ -1031,12 +1037,14 @@ Status ModelManager::GenSessionId(uint64_t &session_id) {
     GELOGE(INTERNAL_ERROR, "Failed to get current time.");
     return INTERNAL_ERROR;
   }
-  session_id = static_cast<uint64_t>(tv.tv_sec * kTimeSpecMiro + tv.tv_usec);  // 1000000us
+  uint64_t timestamp = static_cast<uint64_t>(tv.tv_sec * kTimeSpecMiro + tv.tv_usec);  // 1000000us
+
+  static uint32_t pid = mmGetPid();
 
   session_id_bias_++;
-  // max bais 100.
-  session_id_bias_ = session_id_bias_ % kSessionMaxBias;
-  session_id = session_id * kSessionMaxBias + session_id_bias_;
+
+  session_id = ((timestamp<<kMaskPerOffset<<kMaskPerOffset) & kSessionTimeMask) +
+               ((pid<<kMaskPerOffset) & kSessionPidMask) + (session_id_bias_ & kSessionBiasMask);
 
   GELOGD("Generate new session id: %lu.", session_id);
   return SUCCESS;

ge/graph/load/new_model_manager/model_utils.cc → ge/graph/load/model_manager/model_utils.cc

@@ -14,20 +14,13 @@
  * limitations under the License.
  */
 
-#include "graph/load/new_model_manager/model_utils.h"
-
+#include "graph/load/model_manager/model_utils.h"
 #include <string>
-
 #include "common/debug/log.h"
 #include "common/op/ge_op_utils.h"
-#include "graph/debug/ge_attr_define.h"
-#include "graph/utils/attr_utils.h"
 #include "graph/utils/tensor_utils.h"
-#include "runtime/base.h"
-#include "runtime/kernel.h"
-
-#include "framework/common/debug/ge_log.h"
 #include "graph/manager/graph_var_manager.h"
+#include "graph/types.h"
 
 #define VALIDATE_MEM_RANGE(OP, SIZE, OFFSET)                                                                 \
   do {                                                                                                       \
@@ -342,13 +335,13 @@ vector<void *> ModelUtils::GetInputDataAddrs(const RuntimeParam &model_param, Co
     int64_t input_offset = v_input_offset[non_const_index];
     non_const_index++;
     GE_IF_BOOL_EXEC(model_param.var_size != 0 && ge::VarManager::Instance(session_id)->IsVarAddr(input_offset),
-                    VALIDATE_MEM_RANGE(op_desc, model_param.var_size, input_offset - model_param.logic_var_base);
-                    uint8_t *variable_addr = model_param.var_base + input_offset - model_param.logic_var_base;
+                    uint8_t *variable_addr = nullptr;
+                    GE_CHK_STATUS_EXEC(GetVarAddr(model_param, op_desc, input_offset, variable_addr), return {});
                     v_input_data_addr.push_back(variable_addr);
                     GELOGI("[IMAS]GetInputDataAddrs graph_%u type[V] name[%s] input[%lu] memaddr[%p]",
                            model_param.graph_id, op_desc->GetName().c_str(), i, variable_addr);
                     continue);
-    
+
     int64_t mem_type;
     bool tensor_has_mem_type = ge::AttrUtils::GetInt(tensor_desc, ATTR_NAME_TENSOR_MEM_TYPE, mem_type);
     // feature maps
@@ -380,6 +373,34 @@ vector<void *> ModelUtils::GetInputDataAddrs(const RuntimeParam &model_param, Co
   return v_input_data_addr;
 }
 
+///
+/// @ingroup ge
+/// @brief Get variable address.
+/// @return Status
+///
+Status ModelUtils::GetVarAddr(const RuntimeParam &model_param, const ConstOpDescPtr &op_desc, int64_t offset,
+                              uint8_t *&var_addr) {
+  rtMemType_t mem_type = ge::VarManager::Instance(model_param.session_id)->GetVarMemType(offset);
+  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));
+        return PARAM_INVALID;
+      }
+      var_addr = reinterpret_cast<uint8_t *>(offset);
+      break;
+    case RT_MEMORY_HBM:
+      VALIDATE_MEM_RANGE(op_desc, model_param.var_size, offset - model_param.logic_var_base);
+      var_addr = model_param.var_base + offset - model_param.logic_var_base;
+      break;
+    default:
+      GELOGE(PARAM_INVALID, "unsupported memory type %u", mem_type);
+      return PARAM_INVALID;
+  }
+  GE_CHECK_NOTNULL(var_addr);
+  return SUCCESS;
+}
+
 ///
 /// @ingroup ge
 /// @brief Get output data address.
@@ -404,19 +425,26 @@ vector<void *> ModelUtils::GetOutputDataAddrs(const RuntimeParam &model_param, C
     return v_output_data_addr;
   }
   for (size_t i = 0; i < outputs_size; ++i) {
-    GE_IF_BOOL_EXEC(model_param.var_size != 0 && ge::VarManager::Instance(session_id)->IsVarAddr(v_output_offset[i]),
-                    VALIDATE_MEM_RANGE(op_desc, model_param.var_size, v_output_offset[i] - model_param.logic_var_base);
-                    uint8_t *variable_addr = model_param.var_base + v_output_offset[i] - model_param.logic_var_base;
-                    v_output_data_addr.push_back(variable_addr);
-                    GELOGI("[IMAS]GetOutputDataAddrs graph_%u type[V] name[%s] output[%zu] memaddr[%p]",
-                           model_param.graph_id, op_desc->GetName().c_str(), i, variable_addr);
-                    continue);
     const GeTensorDescPtr tensor_desc = op_desc->MutableOutputDesc(i);
     if (tensor_desc == nullptr) {
       GELOGW("Op: %s, Index: %zu, Tensor Desc is null", op_desc->GetName().c_str(), i);
       continue;
     }
 
+    int32_t calc_type = 0;
+    bool ret = ge::AttrUtils::GetInt(tensor_desc, ATTR_NAME_MEMORY_SIZE_CALC_TYPE, calc_type);
+    if (ret && (calc_type == static_cast<int32_t>(ge::MemorySizeCalcType::ALWAYS_EMPTY))) {
+      GELOGD("%s is an optional output, the address don't need to be saved.", tensor_desc->GetName().c_str());
+      continue;
+    }
+    GE_IF_BOOL_EXEC(model_param.var_size != 0 && ge::VarManager::Instance(session_id)->IsVarAddr(v_output_offset[i]),
+                    uint8_t *variable_addr = nullptr;
+                    GE_CHK_STATUS_EXEC(GetVarAddr(model_param, op_desc, v_output_offset[i], variable_addr), return {});
+                    v_output_data_addr.push_back(variable_addr);
+                    GELOGI("[IMAS]GetOutputDataAddrs graph_%u type[V] name[%s] output[%zu] memaddr[%p]",
+                           model_param.graph_id, op_desc->GetName().c_str(), i, variable_addr);
+                    continue);
+
     int64_t mem_type;
     bool tensor_has_mem_type = ge::AttrUtils::GetInt(tensor_desc, ATTR_NAME_TENSOR_MEM_TYPE, mem_type);
     // feature maps

ge/graph/load/new_model_manager/model_utils.h → ge/graph/load/model_manager/model_utils.h

@@ -21,7 +21,7 @@
 
 #include "common/ge_inner_error_codes.h"
 #include "common/types.h"
-#include "graph/load/new_model_manager/task_info/task_info.h"
+#include "graph/load/model_manager/task_info/task_info.h"
 #include "graph/op_desc.h"
 #include "graph/utils/tensor_adapter.h"
 
@@ -107,6 +107,15 @@ class ModelUtils {
   /// @return Status
   ///
   static Status GetRtAddress(const RuntimeParam &model_param, uintptr_t logic_addr, uint8_t *&mem_addr);
+
+ private:
+  ///
+  /// @ingroup ge
+  /// @brief Get variable address.
+  /// @return Status
+  ///
+  static Status GetVarAddr(const RuntimeParam &model_param, const ConstOpDescPtr &op_desc, int64_t offset,
+                           uint8_t *&var_addr);
 };
 }  // namespace ge
 

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

@@ -14,11 +14,11 @@
  * limitations under the License.
  */
 
-#include "graph/load/new_model_manager/task_info/end_graph_task_info.h"
+#include "graph/load/model_manager/task_info/end_graph_task_info.h"
 
 #include "common/properties_manager.h"
 #include "framework/common/debug/ge_log.h"
-#include "graph/load/new_model_manager/davinci_model.h"
+#include "graph/load/model_manager/davinci_model.h"
 
 namespace {
 const uint32_t kDumpFlag = 2;

ge/graph/load/new_model_manager/task_info/end_graph_task_info.h → ge/graph/load/model_manager/task_info/end_graph_task_info.h

@@ -17,7 +17,7 @@
 #ifndef GE_GRAPH_LOAD_NEW_MODEL_MANAGER_TASK_INFO_END_GRAPH_TASK_INFO_H_
 #define GE_GRAPH_LOAD_NEW_MODEL_MANAGER_TASK_INFO_END_GRAPH_TASK_INFO_H_
 
-#include "graph/load/new_model_manager/task_info/task_info.h"
+#include "graph/load/model_manager/task_info/task_info.h"
 
 namespace ge {
 class EndGraphTaskInfo : public TaskInfo {

ge/graph/load/new_model_manager/task_info/event_record_task_info.cc → ge/graph/load/model_manager/task_info/event_record_task_info.cc

@@ -14,10 +14,10 @@
  * limitations under the License.
  */
 
-#include "graph/load/new_model_manager/task_info/event_record_task_info.h"
+#include "graph/load/model_manager/task_info/event_record_task_info.h"
 
 #include "framework/common/debug/ge_log.h"
-#include "graph/load/new_model_manager/davinci_model.h"
+#include "graph/load/model_manager/davinci_model.h"
 
 namespace ge {
 Status EventRecordTaskInfo::Init(const domi::TaskDef &task_def, DavinciModel *davinci_model) {

ge/graph/load/new_model_manager/task_info/event_record_task_info.h → ge/graph/load/model_manager/task_info/event_record_task_info.h

@@ -16,7 +16,7 @@
 
 #ifndef GE_GRAPH_LOAD_NEW_MODEL_MANAGER_TASK_INFO_EVENT_RECORD_TASK_INFO_H_
 #define GE_GRAPH_LOAD_NEW_MODEL_MANAGER_TASK_INFO_EVENT_RECORD_TASK_INFO_H_
-#include "graph/load/new_model_manager/task_info/task_info.h"
+#include "graph/load/model_manager/task_info/task_info.h"
 
 namespace ge {
 class EventRecordTaskInfo : public TaskInfo {

ge/graph/load/new_model_manager/task_info/event_wait_task_info.cc → ge/graph/load/model_manager/task_info/event_wait_task_info.cc

@@ -14,10 +14,10 @@
  * limitations under the License.
  */
 
-#include "graph/load/new_model_manager/task_info/event_wait_task_info.h"
+#include "graph/load/model_manager/task_info/event_wait_task_info.h"
 
 #include "framework/common/debug/ge_log.h"
-#include "graph/load/new_model_manager/davinci_model.h"
+#include "graph/load/model_manager/davinci_model.h"
 
 namespace ge {
 Status EventWaitTaskInfo::Init(const domi::TaskDef &task_def, DavinciModel *davinci_model) {

ge/graph/load/new_model_manager/task_info/event_wait_task_info.h → ge/graph/load/model_manager/task_info/event_wait_task_info.h

@@ -16,7 +16,7 @@
 
 #ifndef GE_GRAPH_LOAD_NEW_MODEL_MANAGER_TASK_INFO_EVENT_WAIT_TASK_INFO_H_
 #define GE_GRAPH_LOAD_NEW_MODEL_MANAGER_TASK_INFO_EVENT_WAIT_TASK_INFO_H_
-#include "graph/load/new_model_manager/task_info/task_info.h"
+#include "graph/load/model_manager/task_info/task_info.h"
 
 namespace ge {
 class EventWaitTaskInfo : public TaskInfo {

ge/graph/load/new_model_manager/task_info/fusion_start_task_info.cc → ge/graph/load/model_manager/task_info/fusion_start_task_info.cc

@@ -14,10 +14,10 @@
  * limitations under the License.
  */
 
-#include "graph/load/new_model_manager/task_info/fusion_start_task_info.h"
+#include "graph/load/model_manager/task_info/fusion_start_task_info.h"
 
 #include "framework/common/debug/ge_log.h"
-#include "graph/load/new_model_manager/davinci_model.h"
+#include "graph/load/model_manager/davinci_model.h"
 
 namespace ge {
 Status FusionStartTaskInfo::Init(const domi::TaskDef &task_def, DavinciModel *davinci_model) {

ge/graph/load/new_model_manager/task_info/fusion_start_task_info.h → ge/graph/load/model_manager/task_info/fusion_start_task_info.h

@@ -16,7 +16,7 @@
 
 #ifndef GE_GRAPH_LOAD_NEW_MODEL_MANAGER_TASK_INFO_FUSION_START_TASK_INFO_H_
 #define GE_GRAPH_LOAD_NEW_MODEL_MANAGER_TASK_INFO_FUSION_START_TASK_INFO_H_
-#include "graph/load/new_model_manager/task_info/task_info.h"
+#include "graph/load/model_manager/task_info/task_info.h"
 
 namespace ge {
 class FusionStartTaskInfo : public TaskInfo {

ge/graph/load/new_model_manager/task_info/fusion_stop_task_info.cc → ge/graph/load/model_manager/task_info/fusion_stop_task_info.cc

@@ -14,10 +14,10 @@
  * limitations under the License.
  */
 
-#include "graph/load/new_model_manager/task_info/fusion_stop_task_info.h"
+#include "graph/load/model_manager/task_info/fusion_stop_task_info.h"
 
 #include "framework/common/debug/ge_log.h"
-#include "graph/load/new_model_manager/davinci_model.h"
+#include "graph/load/model_manager/davinci_model.h"
 
 namespace ge {
 Status FusionStopTaskInfo::Init(const domi::TaskDef &task_def, DavinciModel *davinci_model) {

ge/graph/load/new_model_manager/task_info/fusion_stop_task_info.h → ge/graph/load/model_manager/task_info/fusion_stop_task_info.h

@@ -16,7 +16,7 @@
 
 #ifndef GE_GRAPH_LOAD_NEW_MODEL_MANAGER_TASK_INFO_FUSION_STOP_TASK_INFO_H_
 #define GE_GRAPH_LOAD_NEW_MODEL_MANAGER_TASK_INFO_FUSION_STOP_TASK_INFO_H_
-#include "graph/load/new_model_manager/task_info/task_info.h"
+#include "graph/load/model_manager/task_info/task_info.h"
 
 namespace ge {
 class FusionStopTaskInfo : public TaskInfo {

ge/graph/load/new_model_manager/task_info/hccl_task_info.cc → ge/graph/load/model_manager/task_info/hccl_task_info.cc

@@ -14,14 +14,14 @@
  * limitations under the License.
  */
 
-#include "graph/load/new_model_manager/task_info/hccl_task_info.h"
+#include "graph/load/model_manager/task_info/hccl_task_info.h"
 
 #include <utility>
 
 #include "common/opskernel/ops_kernel_info_store.h"
 #include "framework/common/debug/ge_log.h"
-#include "graph/load/new_model_manager/davinci_model.h"
-#include "graph/load/new_model_manager/model_utils.h"
+#include "graph/load/model_manager/davinci_model.h"
+#include "graph/load/model_manager/model_utils.h"
 
 namespace ge {
 std::mutex HcclTaskInfo::hccl_follow_stream_mutex_;

ge/graph/load/new_model_manager/task_info/hccl_task_info.h → ge/graph/load/model_manager/task_info/hccl_task_info.h

@@ -23,7 +23,7 @@
 #include <vector>
 
 #include "common/opskernel/ge_task_info.h"
-#include "graph/load/new_model_manager/task_info/task_info.h"
+#include "graph/load/model_manager/task_info/task_info.h"
 #include "graph/manager/util/hcom_util.h"
 namespace ge {
 class HcclTaskInfo : public TaskInfo {

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

@@ -14,7 +14,7 @@
  * limitations under the License.
  */
 
-#include "graph/load/new_model_manager/task_info/kernel_ex_task_info.h"
+#include "graph/load/model_manager/task_info/kernel_ex_task_info.h"
 
 #include <vector>
 
@@ -24,8 +24,8 @@
 #include "framework/common/debug/ge_log.h"
 #include "framework/common/fmk_error_codes.h"
 #include "graph/attr_value.h"
-#include "graph/load/new_model_manager/davinci_model.h"
-#include "graph/load/new_model_manager/model_manager.h"
+#include "graph/load/model_manager/davinci_model.h"
+#include "graph/load/model_manager/model_manager.h"
 
 namespace ge {
 Status KernelExTaskInfo::Init(const domi::TaskDef &task_def, DavinciModel *davinci_model) {

ge/graph/load/new_model_manager/task_info/kernel_ex_task_info.h → ge/graph/load/model_manager/task_info/kernel_ex_task_info.h

@@ -17,7 +17,7 @@
 #ifndef GE_GRAPH_LOAD_NEW_MODEL_MANAGER_TASK_INFO_KERNEL_EX_TASK_INFO_H_
 #define GE_GRAPH_LOAD_NEW_MODEL_MANAGER_TASK_INFO_KERNEL_EX_TASK_INFO_H_
 
-#include "graph/load/new_model_manager/task_info/task_info.h"
+#include "graph/load/model_manager/task_info/task_info.h"
 #include "graph/op_desc.h"
 
 namespace ge {

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

@@ -14,7 +14,7 @@
  * limitations under the License.
  */
 
-#include "graph/load/new_model_manager/task_info/kernel_task_info.h"
+#include "graph/load/model_manager/task_info/kernel_task_info.h"
 #include <map>
 #include <memory>
 #include <string>
@@ -25,9 +25,9 @@
 #include "framework/common/debug/ge_log.h"
 #include "framework/common/l2_cache_optimize.h"
 #include "graph/debug/ge_attr_define.h"
-#include "graph/load/new_model_manager/davinci_model.h"
-#include "graph/load/new_model_manager/model_manager.h"
-#include "graph/load/new_model_manager/model_utils.h"
+#include "graph/load/model_manager/davinci_model.h"
+#include "graph/load/model_manager/model_manager.h"
+#include "graph/load/model_manager/model_utils.h"
 #include "runtime/kernel.h"
 #include "super_kernel/super_kernel.h"
 #include "super_kernel/super_kernel_factory.h"

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

@@ -22,7 +22,7 @@
 #include <string>
 #include <vector>
 
-#include "graph/load/new_model_manager/task_info/task_info.h"
+#include "graph/load/model_manager/task_info/task_info.h"
 #include "graph/op_desc.h"
 namespace ge {
 class KernelTaskInfo : public TaskInfo {

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

@@ -14,9 +14,9 @@
  * limitations under the License.
  */
 
-#include "graph/load/new_model_manager/task_info/label_goto_ex_task_info.h"
+#include "graph/load/model_manager/task_info/label_goto_ex_task_info.h"
 
-#include "graph/load/new_model_manager/davinci_model.h"
+#include "graph/load/model_manager/davinci_model.h"
 #include "graph/debug/ge_attr_define.h"
 
 namespace ge {

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

@@ -17,7 +17,7 @@
 #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_
 
-#include "graph/load/new_model_manager/task_info/task_info.h"
+#include "graph/load/model_manager/task_info/task_info.h"
 
 namespace ge {
 class LabelGotoExTaskInfo : public TaskInfo {

ge/graph/load/new_model_manager/task_info/label_set_task_info.cc → ge/graph/load/model_manager/task_info/label_set_task_info.cc

@@ -14,9 +14,9 @@
  * limitations under the License.
  */
 
-#include "graph/load/new_model_manager/task_info/label_set_task_info.h"
+#include "graph/load/model_manager/task_info/label_set_task_info.h"
 
-#include "graph/load/new_model_manager/davinci_model.h"
+#include "graph/load/model_manager/davinci_model.h"
 #include "graph/debug/ge_attr_define.h"
 
 namespace ge {

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

@@ -17,7 +17,7 @@
 #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_
 
-#include "graph/load/new_model_manager/task_info/task_info.h"
+#include "graph/load/model_manager/task_info/task_info.h"
 
 namespace ge {
 class LabelSetTaskInfo : public TaskInfo {

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

@@ -14,10 +14,10 @@
  * limitations under the License.
  */
 
-#include "graph/load/new_model_manager/task_info/label_switch_by_index_task_info.h"
+#include "graph/load/model_manager/task_info/label_switch_by_index_task_info.h"
 
 #include "graph/debug/ge_attr_define.h"
-#include "graph/load/new_model_manager/davinci_model.h"
+#include "graph/load/model_manager/davinci_model.h"
 
 namespace ge {
 constexpr uint8_t kLabelSwitchIndexNum = 1;

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

@@ -17,7 +17,7 @@
 #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_
 
-#include "graph/load/new_model_manager/task_info/task_info.h"
+#include "graph/load/model_manager/task_info/task_info.h"
 
 namespace ge {
 class LabelSwitchByIndexTaskInfo : public TaskInfo {

ge/graph/load/new_model_manager/task_info/memcpy_addr_async_task_info.cc → ge/graph/load/model_manager/task_info/memcpy_addr_async_task_info.cc

@@ -14,10 +14,10 @@
  * limitations under the License.
  */
 
-#include "graph/load/new_model_manager/task_info/memcpy_addr_async_task_info.h"
+#include "graph/load/model_manager/task_info/memcpy_addr_async_task_info.h"
 
 #include "framework/common/debug/ge_log.h"
-#include "graph/load/new_model_manager/davinci_model.h"
+#include "graph/load/model_manager/davinci_model.h"
 
 namespace {
 const uint32_t kAlignBytes = 64;

ge/graph/load/new_model_manager/task_info/memcpy_addr_async_task_info.h → ge/graph/load/model_manager/task_info/memcpy_addr_async_task_info.h

@@ -17,7 +17,7 @@
 #ifndef GE_GRAPH_LOAD_NEW_MODEL_MANAGER_TASK_INFO_MEMCPY_ADDR_ASYNC_TASK_INFO_H_
 #define GE_GRAPH_LOAD_NEW_MODEL_MANAGER_TASK_INFO_MEMCPY_ADDR_ASYNC_TASK_INFO_H_
 
-#include "graph/load/new_model_manager/task_info/task_info.h"
+#include "graph/load/model_manager/task_info/task_info.h"
 
 namespace ge {
 class MemcpyAddrAsyncTaskInfo : public TaskInfo {

ge/graph/load/new_model_manager/task_info/memcpy_async_task_info.cc → ge/graph/load/model_manager/task_info/memcpy_async_task_info.cc

@@ -14,10 +14,10 @@
  * limitations under the License.
  */
 
-#include "graph/load/new_model_manager/task_info/memcpy_async_task_info.h"
+#include "graph/load/model_manager/task_info/memcpy_async_task_info.h"
 
 #include "framework/common/debug/ge_log.h"
-#include "graph/load/new_model_manager/davinci_model.h"
+#include "graph/load/model_manager/davinci_model.h"
 
 namespace ge {
 Status MemcpyAsyncTaskInfo::Init(const domi::TaskDef &task_def, DavinciModel *davinci_model) {

ge/graph/load/new_model_manager/task_info/memcpy_async_task_info.h → ge/graph/load/model_manager/task_info/memcpy_async_task_info.h

@@ -17,7 +17,7 @@
 #ifndef GE_GRAPH_LOAD_NEW_MODEL_MANAGER_TASK_INFO_MEMCPY_ASYNC_TASK_INFO_H_
 #define GE_GRAPH_LOAD_NEW_MODEL_MANAGER_TASK_INFO_MEMCPY_ASYNC_TASK_INFO_H_
 
-#include "graph/load/new_model_manager/task_info/task_info.h"
+#include "graph/load/model_manager/task_info/task_info.h"
 #include "graph/op_desc.h"
 
 namespace ge {

ge/graph/load/new_model_manager/task_info/model_exit_task_info.cc → ge/graph/load/model_manager/task_info/model_exit_task_info.cc

@@ -14,11 +14,11 @@
  * limitations under the License.
  */
 
-#include "graph/load/new_model_manager/task_info/model_exit_task_info.h"
+#include "graph/load/model_manager/task_info/model_exit_task_info.h"
 
 #include "common/properties_manager.h"
 #include "framework/common/debug/ge_log.h"
-#include "graph/load/new_model_manager/davinci_model.h"
+#include "graph/load/model_manager/davinci_model.h"
 
 namespace ge {
 Status ModelExitTaskInfo::Init(const domi::TaskDef &task_def, DavinciModel *davinci_model) {

ge/graph/load/new_model_manager/task_info/model_exit_task_info.h → ge/graph/load/model_manager/task_info/model_exit_task_info.h

@@ -17,7 +17,7 @@
 #ifndef GE_GRAPH_LOAD_NEW_MODEL_MANAGER_TASK_INFO_MODEL_EXIT_TASK_INFO_H_
 #define GE_GRAPH_LOAD_NEW_MODEL_MANAGER_TASK_INFO_MODEL_EXIT_TASK_INFO_H_
 
-#include "graph/load/new_model_manager/task_info/task_info.h"
+#include "graph/load/model_manager/task_info/task_info.h"
 
 namespace ge {
 class ModelExitTaskInfo : public TaskInfo {

ge/graph/load/new_model_manager/task_info/profiler_trace_task_info.cc → ge/graph/load/model_manager/task_info/profiler_trace_task_info.cc

@@ -14,10 +14,10 @@
  * limitations under the License.
  */
 
-#include "graph/load/new_model_manager/task_info/profiler_trace_task_info.h"
+#include "graph/load/model_manager/task_info/profiler_trace_task_info.h"
 
 #include "framework/common/debug/ge_log.h"
-#include "graph/load/new_model_manager/davinci_model.h"
+#include "graph/load/model_manager/davinci_model.h"
 
 namespace ge {
 Status ProfilerTraceTaskInfo::Init(const domi::TaskDef &task_def, DavinciModel *davinci_model) {

ge/graph/load/new_model_manager/task_info/profiler_trace_task_info.h → ge/graph/load/model_manager/task_info/profiler_trace_task_info.h

@@ -16,7 +16,7 @@
 
 #ifndef GE_GRAPH_LOAD_NEW_MODEL_MANAGER_TASK_INFO_PROFILER_TRACE_TASK_INFO_H_
 #define GE_GRAPH_LOAD_NEW_MODEL_MANAGER_TASK_INFO_PROFILER_TRACE_TASK_INFO_H_
-#include "graph/load/new_model_manager/task_info/task_info.h"
+#include "graph/load/model_manager/task_info/task_info.h"
 
 namespace ge {
 class ProfilerTraceTaskInfo : public TaskInfo {

ge/graph/load/new_model_manager/task_info/stream_active_task_info.cc → ge/graph/load/model_manager/task_info/stream_active_task_info.cc

@@ -14,12 +14,12 @@
  * limitations under the License.
  */
 
-#include "graph/load/new_model_manager/task_info/stream_active_task_info.h"
+#include "graph/load/model_manager/task_info/stream_active_task_info.h"
 
 #include <vector>
 
 #include "framework/common/debug/ge_log.h"
-#include "graph/load/new_model_manager/davinci_model.h"
+#include "graph/load/model_manager/davinci_model.h"
 #include "graph/debug/ge_attr_define.h"
 
 namespace ge {

ge/graph/load/new_model_manager/task_info/stream_active_task_info.h → ge/graph/load/model_manager/task_info/stream_active_task_info.h

@@ -16,7 +16,7 @@
 
 #ifndef GE_GRAPH_LOAD_NEW_MODEL_MANAGER_TASK_INFO_STREAM_ACTIVE_TASK_INFO_H_
 #define GE_GRAPH_LOAD_NEW_MODEL_MANAGER_TASK_INFO_STREAM_ACTIVE_TASK_INFO_H_
-#include "graph/load/new_model_manager/task_info/task_info.h"
+#include "graph/load/model_manager/task_info/task_info.h"
 
 namespace ge {
 class StreamActiveTaskInfo : public TaskInfo {

ge/graph/load/new_model_manager/task_info/stream_switch_task_info.cc → ge/graph/load/model_manager/task_info/stream_switch_task_info.cc

@@ -14,13 +14,13 @@
  * limitations under the License.
  */
 
-#include "graph/load/new_model_manager/task_info/stream_switch_task_info.h"
+#include "graph/load/model_manager/task_info/stream_switch_task_info.h"
 
 #include <vector>
 
 #include "framework/common/debug/ge_log.h"
-#include "graph/load/new_model_manager/davinci_model.h"
-#include "graph/load/new_model_manager/model_utils.h"
+#include "graph/load/model_manager/davinci_model.h"
+#include "graph/load/model_manager/model_utils.h"
 #include "graph/debug/ge_attr_define.h"
 
 namespace ge {

ge/graph/load/new_model_manager/task_info/stream_switch_task_info.h → ge/graph/load/model_manager/task_info/stream_switch_task_info.h

@@ -16,7 +16,7 @@
 
 #ifndef GE_GRAPH_LOAD_NEW_MODEL_MANAGER_TASK_INFO_STREAM_SWITCH_TASK_INFO_H_
 #define GE_GRAPH_LOAD_NEW_MODEL_MANAGER_TASK_INFO_STREAM_SWITCH_TASK_INFO_H_
-#include "graph/load/new_model_manager/task_info/task_info.h"
+#include "graph/load/model_manager/task_info/task_info.h"
 
 namespace ge {
 class StreamSwitchTaskInfo : public TaskInfo {

ge/graph/load/new_model_manager/task_info/stream_switchn_task_info.cc → ge/graph/load/model_manager/task_info/stream_switchn_task_info.cc

@@ -13,12 +13,12 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  */
-#include "graph/load/new_model_manager/task_info/stream_switchn_task_info.h"
+#include "graph/load/model_manager/task_info/stream_switchn_task_info.h"
 #include <vector>
 #include "framework/common/debug/ge_log.h"
 #include "graph/debug/ge_attr_define.h"
-#include "graph/load/new_model_manager/davinci_model.h"
-#include "graph/load/new_model_manager/model_utils.h"
+#include "graph/load/model_manager/davinci_model.h"
+#include "graph/load/model_manager/model_utils.h"
 
 namespace {
 const uint8_t kStreamSwitchnInputNum = 1;

ge/graph/load/new_model_manager/task_info/stream_switchn_task_info.h → ge/graph/load/model_manager/task_info/stream_switchn_task_info.h

@@ -17,7 +17,7 @@
 #ifndef GE_GRAPH_LOAD_NEW_MODEL_MANAGER_TASK_INFO_STREAM_SWITCHN_TASK_INFO_H_
 #define GE_GRAPH_LOAD_NEW_MODEL_MANAGER_TASK_INFO_STREAM_SWITCHN_TASK_INFO_H_
 
-#include "graph/load/new_model_manager/task_info/task_info.h"
+#include "graph/load/model_manager/task_info/task_info.h"
 #include "graph/op_desc.h"
 
 namespace ge {

ge/graph/load/new_model_manager/task_info/task_info.cc → ge/graph/load/model_manager/task_info/task_info.cc

@@ -14,7 +14,7 @@
  * limitations under the License.
  */
 
-#include "graph/load/new_model_manager/task_info/task_info.h"
+#include "graph/load/model_manager/task_info/task_info.h"
 
 #include <vector>
 

ge/graph/load/new_model_manager/task_info/task_info.h → ge/graph/load/model_manager/task_info/task_info.h

@@ -22,8 +22,8 @@
 #include "cce/customize.h"
 #include "framework/common/taskdown_common.h"
 #include "framework/common/ge_inner_error_codes.h"
-#include "graph/load/new_model_manager/ts_mem_mall.h"
-#include "graph/load/new_model_manager/task_info/task_info_factory.h"
+#include "graph/load/model_manager/ts_mem_mall.h"
+#include "graph/load/model_manager/task_info/task_info_factory.h"
 #include "proto/task.pb.h"
 
 namespace ge {

ge/graph/load/new_model_manager/zero_copy_offset.cc → ge/graph/load/model_manager/zero_copy_offset.cc

@@ -14,12 +14,12 @@
  * limitations under the License.
  */
 
-#include "graph/load/new_model_manager/zero_copy_offset.h"
+#include "graph/load/model_manager/zero_copy_offset.h"
 
 #include "framework/common/debug/ge_log.h"
 #include "framework/common/util.h"
-#include "graph/load/new_model_manager/model_utils.h"
-#include "graph/load/new_model_manager/zero_copy_task.h"
+#include "graph/load/model_manager/model_utils.h"
+#include "graph/load/model_manager/zero_copy_task.h"
 
 namespace ge {
 namespace {

ge/graph/load/new_model_manager/zero_copy_offset.h → ge/graph/load/model_manager/zero_copy_offset.h

@@ -25,7 +25,7 @@
 #include "external/ge/ge_api_error_codes.h"
 #include "framework/common/ge_types.h"
 #include "graph/debug/ge_attr_define.h"
-#include "graph/load/new_model_manager/zero_copy_task.h"
+#include "graph/load/model_manager/zero_copy_task.h"
 #include "graph/utils/attr_utils.h"
 #include "graph/utils/tensor_utils.h"
 #include "runtime/mem.h"
@@ -65,7 +65,7 @@ class ZeroCopyOffset {
   // data_size of Data/Netoutput
   int64_t GetDataSize() const { return data_size_; }
   // value of *outside_addrs_ from davinci_model
-  std::vector<std::map<const void *, std::vector<void *>>> &GetOutsideAddrs() { return outside_addrs_; }
+  const std::vector<std::map<const void *, std::vector<void *>>> &GetOutsideAddrs() { return outside_addrs_; }
   // name of op
   std::string GetOpName() const { return op_name_; }
 

ge/graph/load/new_model_manager/zero_copy_task.cc → ge/graph/load/model_manager/zero_copy_task.cc

@@ -14,11 +14,11 @@
  * limitations under the License.
  */
 
-#include "graph/load/new_model_manager/zero_copy_task.h"
+#include "graph/load/model_manager/zero_copy_task.h"
 
 #include "framework/common/debug/ge_log.h"
 #include "framework/common/util.h"
-#include "graph/load/new_model_manager/model_utils.h"
+#include "graph/load/model_manager/model_utils.h"
 #include "common/ge_compiler_options.h"
 
 namespace ge {

ge/graph/manager/graph_manager.cc

@@ -53,6 +53,7 @@
 #include "graph/passes/dimension_adjust_pass.h"
 #include "graph/passes/dimension_compute_pass.h"
 #include "graph/passes/flow_ctrl_pass.h"
+#include "graph/passes/fuse_data_nodes_with_common_input_pass.h"
 #include "graph/passes/identity_pass.h"
 #include "graph/passes/input_output_connection_identify_pass.h"
 #include "graph/passes/iterator_op_pass.h"
@@ -2104,6 +2105,24 @@ Status GraphManager::OptimizeStage1(ge::ComputeGraphPtr &compute_graph) {
       after_merge_passes.AddPass("OptimizeStage1_1::SwitchDataEdgesBypass", new (std::nothrow) SwitchDataEdgesBypass));
   GE_CHK_STATUS_RET(
       after_merge_passes.AddPass("OptimizeStage1_1::ConstantFuseSamePass", new (std::nothrow) ConstantFuseSamePass));
+  /*
+   * Do CSE before FuseDataNodesWithCommonInputPass to resolve the scene in bertlarge as following:
+   *            const
+   *    /        |        \
+   * cast1      cast2     cast3
+   *    \         |         /
+   *             case
+   * the node `const` is the fused const node after ConstantFuseSamePass
+   * the nodes `cast1`, `cast2` and 'cast3' will be fused by CSE.
+   * in order to eliminate hard code in FuseDataNodesWithCommonInputPass,
+   * we do CSE before FuseDataNodesWithCommonInputPass
+   * But it is a temp solution, this CSE will be deleted after change pass from graph pass to node pass
+   */
+  GE_CHK_STATUS_RET(after_merge_passes.AddPass("OptimizeStage1_1::CSEBeforeFuseDataNodesWithCommonInputPass",
+                                               new (std::nothrow) CommonSubexpressionEliminationPass));
+  // FuseDataNodesWithCommonInputPass: fuse same data with common input in same graph
+  GE_CHK_STATUS_RET(after_merge_passes.AddPass("OptimizeStage1_1::FuseDataNodesWithCommonInputPass",
+                                               new (std::nothrow) FuseDataNodesWithCommonInputPass));
   GE_CHK_STATUS_RET(after_merge_passes.AddPass("OptimizeStage1_1::CommonSubexpressionEliminationPass",
                                                new (std::nothrow) CommonSubexpressionEliminationPass));
   GE_CHK_STATUS_RET(after_merge_passes.AddPass("OptimizeStage1_1::PermutePass", new (std::nothrow) PermutePass))
@@ -2226,12 +2245,12 @@ Status GraphManager::OptimizeStage1(ge::ComputeGraphPtr &compute_graph) {
     GELOGE(ret, "Run passes when OptimizeStage1_3 failed, ret:%u.", ret);
     return ret;
   }
-  NamesToPass identity_remove_pass;
-  GE_TIMESTAMP_START(identity_remove_pass);
+  NamesToPass node_pass;
+  GE_TIMESTAMP_START(node_pass);
   IdentityPass identity_force_pass(false);  // after SwitchToStreamSwitchPass
-  identity_remove_pass.emplace_back("IdentityPass", &identity_force_pass);
-  ret = GEPass(compute_graph).Run(identity_remove_pass);
-  GE_TIMESTAMP_END(identity_remove_pass, "GraphPrepare::IdentityRemovePass");
+  node_pass.emplace_back("IdentityPass", &identity_force_pass);
+  ret = GEPass(compute_graph).Run(node_pass);
+  GE_TIMESTAMP_END(node_pass, "GraphPrepare::node_pass");
   if (ret != SUCCESS) {
     GELOGE(ret, "Run identity remove pass for preprocess failed, ret:%u.", ret);
     return ret;
@@ -3102,9 +3121,8 @@ Status GraphManager::Build(const GraphNodePtr &graph_node, ComputeGraphPtr &comp
     graph_name.append(std::to_string(graph_node->GetGraphId()));
     compute_graph->SetName(graph_name);
   }
-  std::vector<SubGraphInfoPtr> sub_graph_list;
-  auto ret = GetCompilerStages(graph_node->GetGraphId()).builder.Build(compute_graph, sub_graph_list, ge_root_model,
-                                                                      session_id);
+
+  auto ret = GetCompilerStages(graph_node->GetGraphId()).builder.Build(compute_graph, ge_root_model, session_id);
   if (ret != SUCCESS) {
     GELOGE(ret, "SubGraph build Failed.");
     return ret;

ge/graph/manager/graph_var_manager.cc

@@ -16,17 +16,10 @@
 
 #include "graph/manager/graph_var_manager.h"
 
-#include <utility>
-
-#include "common/l2_cache_optimize.h"
-#include "common/types.h"
-#include "framework/common/debug/ge_log.h"
-#include "framework/common/debug/log.h"
-#include "ge/ge_api_types.h"
 #include "graph/debug/ge_attr_define.h"
 #include "graph/manager/graph_mem_allocator.h"
+#include "graph/manager/rdma_pool_allocator.h"
 #include "graph/manager/trans_var_data_utils.h"
-#include "graph/utils/attr_utils.h"
 #include "graph/utils/type_utils.h"
 
 using std::map;
@@ -37,7 +30,7 @@ namespace ge {
 VarResource::VarResource(uint64_t session_id) : session_id_(session_id) {}
 
 VarResource::~VarResource() {
-  var_offset_set_.clear();
+  var_offset_map_.clear();
   var_addr_mgr_map_.clear();
   cur_var_tensor_desc_map_.clear();
   var_broad_cast_info_.clear();
@@ -91,8 +84,10 @@ ge::Status VarResource::SaveVarAddr(const std::string &var_name, const ge::GeTen
   std::string var_key = VarKey(var_name, tensor_desc);
   GELOGD("VarResource::SaveVarAddr, var_key = %s", var_key.c_str());
   if (var_addr_mgr_map_.count(var_key) == 0) {
-    uint64_t logic_address = VarManager::Instance(session_id_)->GetVarMemLogicBase() +
-                             static_cast<uint64_t>(reinterpret_cast<std::uintptr_t>(address));
+    uint64_t logic_address = static_cast<uint64_t>(reinterpret_cast<std::uintptr_t>(address));
+    if (memory_type != RT_MEMORY_RDMA_HBM) {
+      logic_address += VarManager::Instance(session_id_)->GetVarMemLogicBase();
+    }
     GELOGI("SaveVarAddr node_name %s, tensor_desc format %s, type %s.", var_name.c_str(),
            TypeUtils::FormatToSerialString(tensor_desc.GetFormat()).c_str(),
            TypeUtils::DataTypeToSerialString(tensor_desc.GetDataType()).c_str());
@@ -102,7 +97,7 @@ ge::Status VarResource::SaveVarAddr(const std::string &var_name, const ge::GeTen
     var_addr_mgr.tensor_desc = tensor_desc;
     var_addr_mgr.memory_type = memory_type;
     var_addr_mgr_map_[var_key] = var_addr_mgr;
-    var_offset_set_.insert(logic_address);
+    var_offset_map_[logic_address] = memory_type;
 
     return SUCCESS;
   }
@@ -211,7 +206,14 @@ ge::Status VarResource::SyncVarData(uint32_t graph_id, const std::string &var_na
   return SyncVarData2BroadCast(graph_id, var_name, var_tensor_desc, base_ptr);
 }
 
-bool VarResource::IsVarAddr(const int64_t &offset) { return var_offset_set_.count(offset) > 0; }
+bool VarResource::IsVarAddr(const int64_t &offset) { return var_offset_map_.count(offset) > 0; }
+
+rtMemType_t VarResource::GetVarMemType(const int64_t &offset) {
+  if (var_offset_map_.count(offset) > 0) {
+    return var_offset_map_[offset];
+  }
+  return RT_MEMORY_RESERVED;
+}
 
 VarTransRoad *VarResource::GetTransRoad(const std::string &var_name) {
   auto iter = var_to_trans_road_.find(var_name);
@@ -252,7 +254,19 @@ Status VarResource::SetAllocatedGraphId(const std::string &var_name, uint32_t gr
 
 MemResource::MemResource() : total_size_(0), var_mem_size_(0) {}
 
-Status MemResource::AssignVarMem(const std::string &var_name, uint64_t size, uint64_t session_id, size_t &mem_offset) {
+MemResource *MemResource::BuildMemResourceFromType(rtMemType_t mem_type) {
+  switch (mem_type) {
+    case RT_MEMORY_HBM:
+      return new (std::nothrow) HbmMemResource();
+    case RT_MEMORY_RDMA_HBM:
+      return new (std::nothrow) RdmaMemResource();
+    default:
+      return nullptr;
+  }
+}
+
+Status HbmMemResource::AssignVarMem(const std::string &var_name, uint64_t size, uint64_t session_id,
+                                    size_t &mem_offset) {
   size = (size + kSessionMemAlignSize - 1) / kSessionMemAlignSize * kSessionMemAlignSize;
   uint64_t real_size = size;
   total_size_ = VarManager::Instance(session_id)->GetVarMemMaxSize();
@@ -282,6 +296,19 @@ Status MemResource::AssignVarMem(const std::string &var_name, uint64_t size, uin
   return SUCCESS;
 }
 
+Status RdmaMemResource::AssignVarMem(const std::string &var_name, uint64_t size, uint64_t session_id, size_t &address) {
+  uint8_t *buffer = MemManager::Instance().RdmaPoolInstance(RT_MEMORY_HBM).Malloc(size);
+  if (buffer == nullptr) {
+    GELOGE(MEMALLOC_FAILED, "Failed to malloc rdma memory for node %s, size = %llu", var_name.c_str(), size);
+    return MEMALLOC_FAILED;
+  }
+  address = static_cast<size_t>(reinterpret_cast<uintptr_t>(buffer));
+  var_mem_size_ += size;
+  GELOGI("[IMAS]AssignVarMem Set session_%llu name[%s] output[%d] addr to [%p] size[%llu].",
+         session_id, var_name.c_str(), 0, buffer, size);
+  return SUCCESS;
+}
+
 uint64_t MemResource::GetVarMemSize() const { return var_mem_size_; }
 
 void MemResource::UpdateVarMemSize(int64_t mem_size) { var_mem_size_ = mem_size; };
@@ -428,7 +455,7 @@ Status VarManager::UpdateVarMemSize(rtMemType_t memory_type, int64_t mem_size) {
   MemResource *mem_resource = nullptr;
   auto iter = mem_resource_map_.find(memory_type);
   if (iter == mem_resource_map_.end()) {
-    mem_resource = new (std::nothrow) MemResource();
+    mem_resource = MemResource::BuildMemResourceFromType(memory_type);
     if (mem_resource == nullptr) {
       GELOGE(ge::INTERNAL_ERROR, "Alloc MemResource failed, memory_type = %u.", memory_type);
       return ge::INTERNAL_ERROR;
@@ -465,7 +492,7 @@ ge::Status VarManager::AssignVarMem(const std::string &var_name, const ge::GeTen
   MemResource *mem_resource = nullptr;
   auto it = mem_resource_map_.find(memory_type);
   if (it == mem_resource_map_.end()) {
-    mem_resource = new (std::nothrow) MemResource();
+    mem_resource = MemResource::BuildMemResourceFromType(memory_type);
     if (mem_resource == nullptr) {
       GELOGE(ge::INTERNAL_ERROR, "Alloc MemResource failed, memory_type = %u.", memory_type);
       return ge::INTERNAL_ERROR;
@@ -629,6 +656,15 @@ bool VarManager::IsVarAddr(const int64_t &offset) {
   return var_resource_->IsVarAddr(offset);
 }
 
+rtMemType_t VarManager::GetVarMemType(const int64_t &offset) {
+  std::lock_guard<std::recursive_mutex> lock(mutex_);
+  if (var_resource_ == nullptr) {
+    GELOGW("VarManager has not been init.");
+    return RT_MEMORY_RESERVED;
+  }
+  return var_resource_->GetVarMemType(offset);
+}
+
 ge::Status VarManager::MallocVarMemory(size_t memory_size) {
   std::lock_guard<std::recursive_mutex> lock(mutex_);
   uint8_t *var_mem_base = nullptr;
@@ -654,12 +690,18 @@ ge::Status VarManager::MallocVarMemory(size_t memory_size) {
 
 uint8_t *VarManager::GetVarMemoryBase(rtMemType_t memory_type) {
   std::lock_guard<std::recursive_mutex> lock(mutex_);
+  if (memory_type == RT_MEMORY_RDMA_HBM) {
+    return MemManager::Instance().RdmaPoolInstance(RT_MEMORY_HBM).GetRdmaBaseAddr();
+  }
   string memory_key = std::to_string(session_id_);
   return MemManager::Instance(memory_type)->GetMemoryAddr(memory_key);
 }
 
 uint8_t *VarManager::GetVarMemoryAddr(uint8_t *logic_addr, rtMemType_t memory_type) {
   std::lock_guard<std::recursive_mutex> lock(mutex_);
+  if (memory_type == RT_MEMORY_RDMA_HBM) {
+    return logic_addr;
+  }
   string mem_key = std::to_string(session_id_);
   uint8_t *mem_base = MemManager::Instance(memory_type)->GetMemoryAddr(mem_key);
   if (mem_base == nullptr) {

ge/graph/manager/graph_var_manager.h

@@ -158,13 +158,15 @@ class VarResource {
 
   bool IsVarAddr(const int64_t &offset);
 
+  rtMemType_t GetVarMemType(const int64_t &offset);
+
   std::unordered_map<std::string, ge::GeTensorDesc> GetAllVarDesc() const { return cur_var_tensor_desc_map_; }
 
  private:
   std::string VarKey(const std::string &var_name, const ge::GeTensorDesc &tensor_desc);
 
   uint64_t session_id_;
-  std::unordered_set<uint64_t> var_offset_set_;
+  std::unordered_map<uint64_t, rtMemType_t> var_offset_map_;
   std::unordered_map<std::string, VarAddrMgr> var_addr_mgr_map_;
   std::unordered_map<std::string, ge::GeTensorDesc> cur_var_tensor_desc_map_;
   std::unordered_map<std::string, std::vector<TransNodeInfo>> var_to_trans_road_;
@@ -176,19 +178,36 @@ class VarResource {
 class MemResource {
  public:
   MemResource();
-  ~MemResource() = default;
+  virtual ~MemResource() = default;
+  static MemResource *BuildMemResourceFromType(rtMemType_t mem_type);
 
-  Status AssignVarMem(const std::string &var_name, uint64_t size, uint64_t session_id, size_t &mem_offset);
+  virtual Status AssignVarMem(const std::string &var_name, uint64_t size, uint64_t session_id, size_t &mem_offset) = 0;
 
   uint64_t GetVarMemSize() const;
 
   void UpdateVarMemSize(int64_t mem_size);
 
- private:
+ protected:
   uint64_t total_size_;
   uint64_t var_mem_size_;
 };
 
+class HbmMemResource : public MemResource {
+ public:
+  HbmMemResource() = default;
+  ~HbmMemResource() override = default;
+
+  Status AssignVarMem(const std::string &var_name, uint64_t size, uint64_t session_id, size_t &address) override;
+};
+
+class RdmaMemResource : public MemResource {
+ public:
+  RdmaMemResource() = default;
+  ~RdmaMemResource() override = default;
+
+  Status AssignVarMem(const std::string &var_name, uint64_t size, uint64_t session_id, size_t &address) override;
+};
+
 class FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY VarManager {
  public:
   static VarManager *Instance(uint64_t session_id);
@@ -275,6 +294,8 @@ class FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY VarManager {
 
   bool IsVarAddr(const int64_t &offset);
 
+  rtMemType_t GetVarMemType(const int64_t &offset);
+
   uint8_t *GetVarMemoryBase(rtMemType_t memory_type);
 
   uint8_t *GetVarMemoryAddr(uint8_t *logic_addr, rtMemType_t memory_type);

ge/graph/manager/rdma_pool_allocator.h

@@ -53,6 +53,10 @@ class RdmaPoolAllocator {
 
   Status GetBaseAddr(uint64_t &base_addr, uint64_t &mem_size);
 
+  uint8_t *GetRdmaBaseAddr() { return rdma_base_addr_; }
+
+  size_t GetRdmaMemSize() { return rdma_mem_size_; }
+
  private:
   void MergeBlocks(Block *dst, Block *src);
 

ge/graph/partition/dynamic_shape_partition.cc

@@ -44,18 +44,46 @@
 #define REQUIRE_SUCCESS(cond, ...) REQUIRE(((cond) == SUCCESS), __VA_ARGS__)
 #define REQUIRE_GRAPH_SUCCESS(cond, ...) REQUIRE(((cond) == GRAPH_SUCCESS), __VA_ARGS__)
 
-bool IsExperimental() {
-  const static bool kIsExperimental = (std::getenv("EXPERIMENTAL_DYNAMIC_PARTITION") != nullptr);
-  return kIsExperimental;
-}
-
 namespace ge {
 using Cluster = DynamicShapePartitioner::Cluster;
 using ClusterPtr = std::shared_ptr<Cluster>;
 
+static bool IsInExperimentalMode(const ComputeGraphPtr &root_graph) {
+  for (const auto &node : root_graph->GetAllNodes()) {
+    GE_CHECK_NOTNULL(node->GetOpDesc());
+    for (const auto &input_desc : node->GetOpDesc()->GetAllInputsDesc()) {
+      auto type = input_desc.GetDataType();
+      if (type == DT_STRING || type == DT_RESOURCE || type == DT_STRING_REF) {
+        if (std::getenv("EXPERIMENTAL_DYNAMIC_PARTITION") == nullptr) {
+          return false;
+        } else {
+          GEEVENT("In dynamic shape scene, model contains data type:"
+                 "DT_STRING/DT_RESOURCE/DT_STRING_REF may not be supported well "
+                 "temporarily, please retry with \"unset EXPERIMENTAL_DYNAMIC_PARTITION\".");
+          break;
+        }
+      }
+    }
+    for (const auto &output_desc : node->GetOpDesc()->GetAllOutputsDesc()) {
+      auto type = output_desc.GetDataType();
+      if (type == DT_STRING || type == DT_RESOURCE || type == DT_STRING_REF) {
+        if (std::getenv("EXPERIMENTAL_DYNAMIC_PARTITION") == nullptr) {
+          return false;
+        } else {
+          GEEVENT("In dynamic shape scene, model contains data type:"
+                 "DT_STRING/DT_RESOURCE/DT_STRING_REF may not be supported well "
+                 "temporarily, please retry with \"unset EXPERIMENTAL_DYNAMIC_PARTITION\".");
+          break;
+        }
+      }
+    }
+  }
+  return true;
+}
+
 Status DynamicShapePartitioner::Partition() {
   REQUIRE_NOT_NULL(root_graph_, "Graph is nullptr.");
-  if (!IsExperimental()) {
+  if (!IsInExperimentalMode(root_graph_)) {
     GELOGD("Skip dynamic shape partition as not in experimental mode.");
     REQUIRE(AttrUtils::SetBool(*root_graph_, ATTR_NAME_DYNAMIC_SHAPE_PARTITIONED, false),
             "Failed set dynamic shape partitioned flag on root graph.");
@@ -185,6 +213,7 @@ std::string DynamicShapePartitioner::DebugString() const {
   size_t data = 0;
   size_t netoutput = 0;
   size_t is_inputnode = 0;
+  size_t stage = 0;
   std::stringstream ss;
   ss << "All unknown shape nodes:" << std::endl;
   for (const auto &node : unknown_shape_nodes_) {
@@ -201,10 +230,13 @@ std::string DynamicShapePartitioner::DebugString() const {
       netoutput++;
     } else if (cluster->IsInputNode()) {
       is_inputnode++;
+    } else if (cluster->IsIndependent()) {
+      stage++;
     }
   }
   ss << "All clusters:" << unique_clusters_.size() << ", data:" << data << ", known:" << known
-     << ", unknown:" << unknown << ", netoutput:" << netoutput << ", is_inputnode:" << is_inputnode << std::endl;
+     << ", unknown:" << unknown << ", netoutput:" << netoutput << ", is_inputnode:" << is_inputnode
+     << ", stage:" << stage << std::endl;
   for (const auto &cluster : unique_clusters_) {
     ss << "  " << cluster->DebugString() << std::endl;
   }
@@ -244,12 +276,15 @@ Status DynamicShapePartitioner::InitClusters() {
   for (const auto &node : graph->GetDirectNode()) {
     Cluster::Type type = Cluster::DATA;
     bool is_input = ((node->GetType() == CONSTANT) || (node->GetType() == CONSTANTOP)) && node->GetInNodes().empty();
+    REQUIRE_NOT_NULL(node->GetOpDesc(), "op_desc is null");
     if (node->GetType() == DATA) {
       type = Cluster::DATA;
     } else if (is_input) {
       type = Cluster::INPUT_NODE;
     } else if (node->GetType() == NETOUTPUT) {
       type = Cluster::NETOUTPUT;
+    } else if ((node->GetType() == PARTITIONEDCALL) && (node->GetOpDesc()->HasAttr(ATTR_STAGE_LEVEL))) {
+      type = Cluster::STAGE;
     } else if (unknown_shape_nodes_.count(node) > 0) {
       type = Cluster::UNKNOWN_SHAPE;
     } else {
@@ -332,6 +367,9 @@ static std::string ToString(const std::vector<ClusterPtr> &clusters) {
 void DynamicShapePartitioner::MergeClustersUnknownShape() {
   // Merge unknown shape clusters
   for (const auto &cluster : ordered_cluster_) {
+    if (cluster->IsIndependent()) {
+      continue;
+    }
     for (const auto &in_cluster : cluster->Inputs()) {
       if (!in_cluster->IsUnknownShape()) {
         continue;
@@ -351,6 +389,9 @@ void DynamicShapePartitioner::MergeClustersUnknownShape() {
 void DynamicShapePartitioner::MergeClustersKnownShape() {
   // Merge known shape clusters
   for (const auto &cluster : ordered_cluster_) {
+    if (cluster->IsIndependent()) {
+      continue;
+    }
     if (cluster->IsRefVariable() && cluster->Inputs().size() == 1) {
       auto in_cluster = *(cluster->Inputs().begin());
       in_cluster->Merge(cluster);
@@ -578,6 +619,7 @@ void Cluster::UpdateRank(size_t rank) {
 bool Cluster::IsData() const { return type_ == DATA; };
 bool Cluster::IsKnownShape() const { return type_ == KNOWN_SHAPE; };
 bool Cluster::IsUnknownShape() const { return type_ == UNKNOWN_SHAPE; };
+bool Cluster::IsIndependent() const { return type_ == STAGE; };
 bool Cluster::IsNetOutput() const { return type_ == NETOUTPUT; };
 bool Cluster::IsInputNode() const { return type_ == INPUT_NODE; };
 bool Cluster::IsRefVariable() const {
@@ -613,6 +655,9 @@ void Cluster::RemoveOutput(ClusterPtr out) {
                           out->in_clusters_.end());
 };
 void Cluster::Merge(ClusterPtr other) {
+  if (other->IsIndependent()) {
+    return;
+  }
   nodes_.insert(nodes_.end(), other->nodes_.begin(), other->nodes_.end());
   other->in_clusters_.erase(std::remove(other->in_clusters_.begin(), other->in_clusters_.end(), shared_from_this()),
                             other->in_clusters_.end());
@@ -661,7 +706,9 @@ std::vector<ClusterPtr> Cluster::MergeAllPathFrom(ClusterPtr other) {
   std::unordered_set<ClusterPtr> forward_reached_clusters;
   std::unordered_set<ClusterPtr> backward_reached_clusters;
   std::vector<ClusterPtr> path_clusters;
-
+  if (other->IsIndependent()) {
+    return path_clusters;
+  }
   if (std::find(other->out_clusters_.begin(), other->out_clusters_.end(), shared_from_this()) ==
       other->out_clusters_.end()) {
     return path_clusters;
@@ -744,7 +791,7 @@ Status Cluster::BuildFrame() {
         }
       }
     }
-    if (IsData()) {
+    if (IsData() || IsIndependent()) {
       for (const auto &anchor : node->GetAllOutDataAnchors()) {
         AddFrameOutput(anchor);
       }
@@ -860,7 +907,7 @@ Status Cluster::CombinePartitionFrame() {
 }
 
 Status Cluster::BuildPartitionSubgraph() {
-  if (IsData() || IsNetOutput()) {
+  if (IsData() || IsNetOutput() || IsIndependent()) {
     return SUCCESS;
   }
   int64_t parent_node_index = 0;

ge/graph/partition/dynamic_shape_partition.h

@@ -32,7 +32,7 @@ class DynamicShapePartitioner {
   // DATA:DATA, UNKNOWN_SHAPE:unknowshape, KNOWN_SHAPE:knowshape, NETOUTPUT:NETOUTPUT.
   class Cluster : public std::enable_shared_from_this<Cluster> {
    public:
-    enum Type { DATA, INPUT_NODE, NETOUTPUT, KNOWN_SHAPE, UNKNOWN_SHAPE };
+    enum Type { DATA, INPUT_NODE, NETOUTPUT, STAGE, KNOWN_SHAPE, UNKNOWN_SHAPE };
     Cluster(size_t rank, Type type, NodePtr node, DynamicShapePartitioner *partitioner)
         : id_(rank), min_(rank), max_(rank), type_(type), partitioner_(partitioner) {
       nodes_.push_back(node);
@@ -45,6 +45,7 @@ class DynamicShapePartitioner {
     bool IsData() const;
     bool IsKnownShape() const;
     bool IsUnknownShape() const;
+    bool IsIndependent() const;
     bool IsNetOutput() const;
     std::vector<std::shared_ptr<Cluster>> Inputs() const;
     std::vector<std::shared_ptr<Cluster>> Outputs() const;

ge/graph/partition/stage_partition.cc

@@ -25,6 +25,10 @@
 #include "common/types.h"
 
 namespace ge {
+namespace {
+const std::set<std::string> kSrcNodeTypes = { DATA, AIPPDATA, ANN_DATA };
+}
+
 Status StagePartitioner::Partition() {
   GE_CHECK_NOTNULL(root_graph_);
   if (root_graph_->GetParentGraph() != nullptr) {
@@ -37,6 +41,10 @@ Status StagePartitioner::Partition() {
     if (!AttrUtils::GetInt(op_desc, ATTR_STAGE_LEVEL, level)) {
       continue;
     }
+    if ((kSrcNodeTypes.count(op_desc->GetType()) != 0) && node->GetInAllNodes().empty()) {
+      continue;
+    }
+    GELOGD("original node %s for stage %u", node->GetName().c_str(), level);
     stage_nodes_[level].insert(node);
   }
   if (stage_nodes_.empty()) {
@@ -54,6 +62,13 @@ Status StagePartitioner::Partition() {
     return FAILED;
   }
 
+  root_graph_->TopologicalSorting([](const NodePtr &a, const NodePtr &b) -> bool {
+    uint32_t a_level = UINT32_MAX;
+    (void)AttrUtils::GetInt(a->GetOpDesc(), ATTR_STAGE_LEVEL, a_level);
+    uint32_t b_level = UINT32_MAX;
+    (void)AttrUtils::GetInt(b->GetOpDesc(), ATTR_STAGE_LEVEL, b_level);
+    return a_level < b_level;
+  });
   if (root_graph_->TopologicalSorting() != GRAPH_SUCCESS) {
     GELOGE(FAILED, "Topological sort for graph %s after stage partition failed, "
            "maybe stage_level was not set correctly.", root_graph_->GetName().c_str());
@@ -76,20 +91,26 @@ Status StagePartitioner::SplitStageLevel() {
       auto node = nodes.top();
       nodes.pop();
       GE_CHECK_NOTNULL(node->GetOpDesc());
-      if (node->GetOpDesc()->HasAttr(ATTR_STAGE_LEVEL) && (cur_stage_nodes.count(node) == 0)) {
+      uint32_t tmp_level = cur_stage_level;
+      (void)AttrUtils::GetInt(node->GetOpDesc(), ATTR_STAGE_LEVEL, tmp_level);
+      if (tmp_level != cur_stage_level) {
         continue;
       }
       for (const auto &in_node : node->GetInAllNodes()) {
         if (visited_stage_nodes.count(in_node) != 0) {
           continue;
         }
+        if (!AttrUtils::SetInt(in_node->GetOpDesc(), ATTR_STAGE_LEVEL, cur_stage_level)) {
+          GELOGE(INTERNAL_ERROR, "Set attr ATTR_STAGE_LEVEL on node %s failed.", in_node->GetName().c_str());
+          return INTERNAL_ERROR;
+        }
+        GELOGD("Mark stage_level node %s, stage_level=%u", in_node->GetName().c_str(), cur_stage_level);
+        if ((kSrcNodeTypes.count(in_node->GetType()) != 0) && in_node->GetInAllNodes().empty()) {
+          GELOGD("skip data node %s for stage %u", in_node->GetName().c_str(), cur_stage_level);
+          continue;
+        }
         nodes.push(in_node);
       }
-      if (!AttrUtils::SetInt(node->GetOpDesc(), ATTR_STAGE_LEVEL, cur_stage_level)) {
-        GELOGE(INTERNAL_ERROR, "Set attr ATTR_STAGE_LEVEL on node %s failed.", node->GetName().c_str());
-        return INTERNAL_ERROR;
-      }
-      GELOGD("Mark stage_level node %s, stage_level=%u", node->GetName().c_str(), cur_stage_level);
       visited_stage_nodes.emplace(node);
     }
     for (const auto &node : visited_stage_nodes) {
@@ -219,6 +240,11 @@ NodePtr StagePartitioner::BuildSubgraphNode(const std::string &graph_name, const
   op_desc->AddSubgraphName("f");
   op_desc->SetSubgraphInstanceName(0, graph_name);
 
+  if (!AttrUtils::SetInt(op_desc, ATTR_STAGE_LEVEL, stage_info.stage_level)) {
+    GELOGE(INTERNAL_ERROR, "Set attr ATTR_STAGE_LEVEL on node %s failed", op_desc->GetName().c_str());
+    return nullptr;
+  }
+
   NodePtr subgraph_node = root_graph_->AddNode(op_desc);
   if (subgraph_node == nullptr) {
     GELOGE(FAILED, "Add node %s failed.", graph_name.c_str());

ge/graph/passes/fuse_data_nodes_with_common_input_pass.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 "graph/passes/fuse_data_nodes_with_common_input_pass.h"
+
+#include <map>
+#include <memory>
+#include <string>
+#include <vector>
+#include <set>
+#include "common/ge_inner_error_codes.h"
+#include "graph/utils/op_desc_utils.h"
+#include "graph/utils/type_utils.h"
+#include "graph/utils/node_utils.h"
+
+using std::map;
+using std::vector;
+using std::set;
+using std::string;
+
+namespace ge {
+Status FuseDataNodesWithCommonInputPass::Run(ge::ComputeGraphPtr graph) {
+  if (graph == nullptr) {
+    GELOGE(GE_GRAPH_PARAM_NULLPTR, "Compute graph is null.");
+    return GE_GRAPH_PARAM_NULLPTR;
+  }
+  GELOGD("FuseDataNodesWithCommonInputPass in.");
+  // key: subgraph, value:--key: peer out anchor to parent node, --value: parent indexes to parent node
+  map<ComputeGraphPtr, map<OutDataAnchorPtr, set<uint32_t>>> subgraphs_to_need_fuse_nodes_info;
+  if (InitNeedFuseNodesInfo(graph, subgraphs_to_need_fuse_nodes_info) != SUCCESS) {
+    GELOGE(FAILED, "InitNeedFuseNodesInfo failed.");
+    return FAILED;
+  }
+  return FuseDataNodes(subgraphs_to_need_fuse_nodes_info);
+}
+
+Status FuseDataNodesWithCommonInputPass::InitNeedFuseNodesInfo(ComputeGraphPtr &graph,
+    map<ComputeGraphPtr, map<OutDataAnchorPtr, set<uint32_t>>> &subgraphs_to_need_fuse_nodes_info) {
+  for (const auto &subgraph : graph->GetAllSubgraphs()) {
+    GE_CHECK_NOTNULL(subgraph);
+    auto parent_node = subgraph->GetParentNode();
+    GE_CHECK_NOTNULL(parent_node);
+    if (parent_node->GetType() == CASE || parent_node->GetType() == IF) {
+      auto &peer_out_anchors_to_parent_indexes = subgraphs_to_need_fuse_nodes_info[subgraph];
+      for (const auto &in_data_anchor : parent_node->GetAllInDataAnchors()) {
+        GE_CHECK_NOTNULL(in_data_anchor);
+        OutDataAnchorPtr peer_out_anchor = in_data_anchor->GetPeerOutAnchor();
+        uint32_t parent_index = static_cast<uint32_t>(in_data_anchor->GetIdx());
+        GE_CHECK_NOTNULL(peer_out_anchor);
+        peer_out_anchors_to_parent_indexes[peer_out_anchor].insert(parent_index);
+        GELOGD("Peer node %s is the %d input of parent node %s in %s.",
+               peer_out_anchor->GetOwnerNode()->GetName().c_str(), parent_index, parent_node->GetName().c_str(),
+               subgraph->GetName().c_str());
+      }
+    }
+  }
+  return SUCCESS;
+}
+
+Status FuseDataNodesWithCommonInputPass::FuseDataNodes(
+    const map<ComputeGraphPtr, map<OutDataAnchorPtr, set<uint32_t>>> &subgraphs_to_need_fuse_nodes_info) {
+  for (const auto &subgraph_to_need_fuse_nodes_info : subgraphs_to_need_fuse_nodes_info) {
+    auto subgraph = subgraph_to_need_fuse_nodes_info.first;
+    for (const auto &peer_out_anchors_to_parent_indexes : subgraph_to_need_fuse_nodes_info.second) {
+      if (peer_out_anchors_to_parent_indexes.second.size() <= 1) {
+        continue;
+      }
+      // key: out anchor, value: data nodes with common input will be fused
+      map<OutDataAnchorPtr, vector<NodePtr>> peer_out_anchors_to_need_fuse_nodes;
+      for (const auto &node : subgraph->GetDirectNode()) {
+        if (node->GetType() != DATA) {
+          continue;
+        }
+        GE_CHECK_NOTNULL(node->GetOpDesc());
+        uint32_t parent_index = 0;
+        if (AttrUtils::GetInt(node->GetOpDesc(), ATTR_NAME_PARENT_NODE_INDEX, parent_index)) {
+          if (peer_out_anchors_to_parent_indexes.second.count(parent_index) > 0) {
+            peer_out_anchors_to_need_fuse_nodes[peer_out_anchors_to_parent_indexes.first].emplace_back(node);
+          }
+        }
+      }
+      for (const auto &peer_out_anchor_to_need_fuse_nodes : peer_out_anchors_to_need_fuse_nodes) {
+        auto need_fuse_data_nodes = peer_out_anchor_to_need_fuse_nodes.second;
+        auto first_node = need_fuse_data_nodes.at(0);
+        for (size_t i = 1; i < need_fuse_data_nodes.size(); ++i) {
+          auto node = need_fuse_data_nodes.at(i);
+          GELOGI("Replace redundant data node %s by %s exist in graph: %s.", node->GetName().c_str(),
+                 first_node->GetName().c_str(), subgraph->GetName().c_str());
+          // the data node which can be fused has none input(both data and control in)
+          if (GraphUtils::MoveOutCtrlEdges(node, first_node) != SUCCESS) {
+            return FAILED;
+          }
+          if (GraphUtils::ReplaceNodeDataAnchors(first_node, node, {}, {0}) != SUCCESS) {
+            return FAILED;
+          }
+          if (GraphUtils::RemoveNodeWithoutRelink(subgraph, node) != SUCCESS) {
+            GELOGE(FAILED, "[%s] RemoveNodeWithoutRelink failed.", node->GetName().c_str());
+            return FAILED;
+          }
+        }
+      }
+    }
+  }
+  return SUCCESS;
+}
+}  // namespace ge

ge/graph/passes/fuse_data_nodes_with_common_input_pass.h

@@ -0,0 +1,38 @@
+/**
+ * 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_GRAPH_PASSES_FUSE_DATA_NODES_WITH_COMMON_INPUT_PASS_H_
+#define GE_GRAPH_PASSES_FUSE_DATA_NODES_WITH_COMMON_INPUT_PASS_H_
+
+#include <set>
+#include <map>
+#include <vector>
+#include "graph/types.h"
+#include "inc/graph_pass.h"
+
+namespace ge {
+class FuseDataNodesWithCommonInputPass : public GraphPass {
+ public:
+  Status Run(ge::ComputeGraphPtr graph) override;
+
+ private:
+  Status InitNeedFuseNodesInfo(ComputeGraphPtr &graph,
+      map<ComputeGraphPtr, map<OutDataAnchorPtr, std::set<uint32_t>>> &subgraphs_to_need_fuse_nodes_info);
+  Status FuseDataNodes(
+      const map<ComputeGraphPtr, map<OutDataAnchorPtr, std::set<uint32_t>>> &subgraphs_to_need_fuse_nodes_info);
+};
+} // namespace ge
+#endif // GE_GRAPH_PASSES_FUSE_DATA_NODES_WITH_COMMON_INPUT_PASS_H_

ge/graph/passes/subgraph_const_migration_pass.cc

@@ -145,6 +145,7 @@ Status SubgraphConstMigrationPass::ClassifyGraphNodes(const ComputeGraphPtr &gra
       return GE_GRAPH_EMPTY_SUBGRAPH;
     }
 
+    set<NodePtr> ctrl_only_const_nodes;
     auto &data_nodes = all_data_nodes[subgraph];
     auto &const_nodes = all_const_nodes[subgraph];
     for (auto &node : subgraph->GetDirectNode()) {
@@ -178,15 +179,30 @@ Status SubgraphConstMigrationPass::ClassifyGraphNodes(const ComputeGraphPtr &gra
           peer_name_list.insert(fixed_name + ":" + std::to_string(in_anchor->GetIdx()));
         }
 
+        if (peer_name_list.empty()) {
+          GELOGI("%s, Const: %s, no data output", subgraph->GetName().c_str(), node->GetName().c_str());
+          const auto in_all_nodes = node->GetInAllNodes();
+          if (in_all_nodes.empty() || std::all_of(in_all_nodes.begin(), in_all_nodes.end(),
+                                                  [](const NodePtr &n) { return n->GetType() == DATA; })) {
+            ctrl_only_const_nodes.insert(node);
+          }
+          continue;
+        }
+
         string key_of_const;
         for (const string &name : peer_name_list) {
           key_of_const += (key_of_const.empty() ? name : "_" + name);
         }
 
         const_nodes[key_of_const] = node;
-        GELOGD("%s, Key: %s, Const: %s", subgraph->GetName().c_str(), key_of_const.c_str(), node->GetName().c_str());
+        GELOGD("%s, Const: %s, Key: %s", subgraph->GetName().c_str(), node->GetName().c_str(), key_of_const.c_str());
       }
     }
+
+    for (auto &node : ctrl_only_const_nodes) {
+      GE_CHK_GRAPH_STATUS_RET(GraphUtils::RemoveNodeWithoutRelink(subgraph, node),
+          "Remove node without relink failed, node: %s", node->GetName().c_str());
+    }
   }
 
   return SUCCESS;
@@ -352,7 +368,8 @@ Status SubgraphConstMigrationPass::DetachParallelNode(const ComputeGraphPtr &gra
     const auto owner_node = out_anchor->GetOwnerNode();
     GELOGI("Remove Edge: %s %s", owner_node->GetName().c_str(), const_node->GetName().c_str());
     if (owner_node->GetInAllNodes().empty() && owner_node->GetOutAllNodes().empty() && owner_node != data_node) {
-      graph->RemoveNode(owner_node);
+      GE_CHK_GRAPH_STATUS_RET(GraphUtils::RemoveNodeWithoutRelink(graph, owner_node),
+          "Remove node without relink failed, node: %s", owner_node->GetName().c_str());
     }
   }
 
@@ -414,7 +431,8 @@ Status SubgraphConstMigrationPass::AttachParallelNode(const ComputeGraphPtr &gra
     const auto owner_node = out_anchor->GetOwnerNode();
     GELOGI("Remove Edge: %s %s", owner_node->GetName().c_str(), func_node->GetName().c_str());
     if (owner_node->GetInAllNodes().empty() && owner_node->GetOutAllNodes().empty()) {
-      graph->RemoveNode(owner_node);
+      GE_CHK_GRAPH_STATUS_RET(GraphUtils::RemoveNodeWithoutRelink(graph, owner_node),
+          "Remove node without relink failed, node: %s", owner_node->GetName().c_str());
     }
   }
   GE_CHK_GRAPH_STATUS_RET(GraphUtils::AddEdge(const_node->GetOutDataAnchor(kZeroIndex), in_anchor), "Add edge failed");
@@ -442,7 +460,8 @@ Status SubgraphConstMigrationPass::MoveNodeToParent(const ComputeGraphPtr &graph
                                                     const map<ComputeGraphPtr, map<uint32_t, NodePtr>> &all_data_nodes,
                                                     const string &node_key, uint32_t parent_index) {
   if (node_key.empty() || parent_index == kInvalidParent) {
-    GELOGE(FAILED, "Graph: %s, inputs is empty", graph->GetName().c_str());
+    GELOGE(FAILED, "Graph: %s, node key: %s, parent index: %u invalid",
+           graph->GetName().c_str(), node_key.c_str(), parent_index);
     return FAILED;
   }
 
@@ -472,7 +491,8 @@ Status SubgraphConstMigrationPass::MoveNodeToParent(const ComputeGraphPtr &graph
       return FAILED;
     }
 
-    GE_CHK_GRAPH_STATUS_RET(subgraph->RemoveNode(move_node), "Remove node failed");
+    GE_CHK_GRAPH_STATUS_RET(GraphUtils::RemoveNodeWithoutRelink(subgraph, move_node),
+        "Remove node without relink failed, node: %s", move_node->GetName().c_str());
     GELOGI("Remove Node: %s %s", subgraph->GetName().c_str(), move_node->GetName().c_str());
   }
 

ge/graph/passes/subgraph_pass.cc

@@ -142,17 +142,18 @@ Status SubgraphPass::SubgraphOutputNode(const ComputeGraphPtr &graph, const Node
     GE_CHECK_NOTNULL(in_node);
 
     // Need insert memcpy
-    //   1. Const->NetOutput in subgraph
+    //   1. Const->NetOutput in subgraph & parent graph is known
     //   2. AtomicOp->NetOutput in subgraph
     //   3. OutputContinuesRequiredOp->NetOutput in subgraph
     //   4. Data->NetOutput in subgraph but parent_node is not while
     //   5. While->NetOutput in known subgraph
     std::string op_type;
-    bool insert_flag = NodeUtils::GetConstOpType(in_node, op_type) ||
+    bool insert_flag =
+        (NodeUtils::GetConstOpType(in_node, op_type) && !graph->GetParentGraph()->GetGraphUnknownFlag()) ||
         IsAtomicRequired(in_node, peer_out_anchor->GetIdx()) || IsOutputContinuesRequired(in_node) ||
         ((in_node->GetType() == DATA) && (kWhileOpTypes.count(graph->GetParentNode()->GetType()) == 0)) ||
         (!graph->GetGraphUnknownFlag() && NodeUtils::IsDynamicShape(node) &&
-            (kWhileOpTypes.count(in_node->GetType()) != 0));
+        (kWhileOpTypes.count(in_node->GetType()) != 0));
     if (insert_flag) {
       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";

ge/host_cpu_engine/ops_kernel_store/op/host_op.cc

@@ -32,5 +32,8 @@ REGISTER_OP_CREATOR(Assign, HostOp);
 REGISTER_OP_CREATOR(RandomUniform, HostOp);
 REGISTER_OP_CREATOR(Add, HostOp);
 REGISTER_OP_CREATOR(Mul, HostOp);
+REGISTER_OP_CREATOR(ConcatV2, HostOp);
+REGISTER_OP_CREATOR(Data, HostOp);
+REGISTER_OP_CREATOR(Fill, HostOp);
 }  // namespace host_cpu
 }  // namespace ge

ge/hybrid/executor/hybrid_model_async_executor.cc

@@ -15,7 +15,7 @@
  */
 
 #include "hybrid/executor/hybrid_model_async_executor.h"
-#include "graph/load/new_model_manager/model_utils.h"
+#include "graph/load/model_manager/model_utils.h"
 #include "graph/utils/tensor_utils.h"
 #include "graph/utils/type_utils.h"
 #include "graph/ge_context.h"
@@ -59,6 +59,7 @@ Status HybridModelAsyncExecutor::Start(const std::shared_ptr<ModelListener> &lis
   run_flag_ = true;
   listener_ = listener;
   future_ = std::async(std::launch::async, [&]() -> Status {
+    GetThreadLocalContext() = *executor_->GetContext()->ge_context;
     GetContext().SetSessionId(executor_->GetContext()->session_id);
     return RunInternal();
   });
@@ -229,7 +230,11 @@ Status HybridModelAsyncExecutor::PrepareInputs(const InputData &current_data, Hy
     }
 
     GE_CHECK_GE(tensor_size, 0);
-    auto tensor_buffer = TensorBuffer::Create(allocator, tensor_size);
+    AllocationAttr attr;
+    if (GetContext().GetHostExecFlag()) {
+      attr.SetMemType(HOST_DDR);
+    }
+    auto tensor_buffer = TensorBuffer::Create(allocator, tensor_size, &attr);
     GE_CHECK_NOTNULL(tensor_buffer);
     args.inputs.emplace_back(std::shared_ptr<TensorBuffer>(tensor_buffer.release()));
 

ge/hybrid/executor/hybrid_model_async_executor.h

@@ -21,7 +21,7 @@
 #include <future>
 #include "external/ge/ge_api_error_codes.h"
 #include "external/ge/ge_api_types.h"
-#include "graph/load/new_model_manager/data_inputer.h"
+#include "graph/load/model_manager/data_inputer.h"
 #include "hybrid/executor/hybrid_model_executor.h"
 #include "runtime/stream.h"